text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
A Prolog-based Environment
for Reasoning about Programming Languages
(Extended abstract)
arXiv:0711.0345v1 [cs.PL] 2 Nov 2007
Roberto Bagnara1, Patricia M. Hill2 , and Enea Zaffanella1
1
Department of Mathematics, University of Parma, Italy
{bagnara,zaffanella}@cs.unipr.it
2
School of Computing, University of Leeds, UK
[email protected]
ECLAIR is a Prolog-based prototype system aiming to provide a functionally
complete environment for the study, development and evaluation of programming
language analysis and implementation tools. In this paper, we sketch the overall structure of the system, outlining the main methodologies and technologies
underlying its components. We also discuss the appropriateness of Prolog as the
implementation language for the system: besides highlighting its strengths, we
also point out a few potential weaknesses, hinting at possible solutions.
Motivation for a flexible language resource. Static program analysis aims
to find bugs, verify the absence of errors in software, and ensure the correctness of
given optimizations. The standard and most used theoretical framework that lies
behind static program analysis is abstract interpretation [8, 9]. This framework,
which has been available for more than thirty years, allows us to separate the
abstract domains, for representing the program’s properties of interest, from
the abstract interpreter that should mimic the execution of the programs on
these domains. As far as the abstract domains are concerned, there is now a
good choice of implementations offering a flexible precision/efficiency trade-off.
For the abstract interpreter, one non-commercial interpreter that has been used
for automatically verifying up to one million lines of C code is ASTRÉE [11];
however, ASTRÉE is specially targeted at a particular class of programs and
program properties, so that widening its scope of application is likely to require
significant effort [7]. The interpreters provided by ECLAIR system follow the
methodology described in [2] which handles most of the problematic features of
mainstream, single-threaded languages; the concrete semantics is based on the
structured operational semantics extended to allow for infinite computations [10,
13, 15] and the abstract semantics is based on the work of Schmidt [16–18]. The
ECLAIR system is being designed to have a high degree of flexibility for the
following reasons:
– As there is a constantly evolving and, potentially, huge set of programming
languages for which analyzers are needed, the prototype interpreter needs
to target a variety of language features (including, for example, exception
handling, functions and pointers) and have the capability of being extended
to incorporate further languages and language features, not yet considered.
2
R. Bagnara, P. M. Hill, E. Zaffanella
Thus, the overall environment has to be designed to provide a uniform interface that delegates the details of the implementation of the run time system
to independent components.
– In order to analyze for different types of program properties and error states,
the system has to be able to connect with a wide variety of abstract domains.
Note that the system’s interface with the abstract domains and their operations must be designed so that it can be coupled with the relational domains,
that is, those that can capture the relationships between different data objects, as well as, the more efficient but less precise, non-relational domains.
This flexibility in choice of domains should also be dynamic since, in order
to control the precision/efficiency trade-off, the exact domain used may have
to be changed during the analysis.
– The aim is to have one system that can be used for teaching, for the development of new technologies and their evaluation and for demonstrating
their application. For teaching the basics, we require simple robust systems
that support the core language features such as the assignment, conditional
and while commands found in all imperative languages. On the other hand,
for research and development we need a highly modular and extensible system so that we can plug in new technologies without redefining the whole.
Lastly, for demonstration, we need the system to be highly efficient with a
well-structured user interface that can verify software that is written in real
languages such as C and Java.
From CLAIR to ECLAIR. The ‘Combined Language and Abstract Interpretation Resource’ (CLAIR, http://www.cs.unipr.it/clair/) has been initially
developed for and used in a teaching context, as a prototype system to help in
the study and experimentation with various aspects of programming language
implementation. The CLAIR system consists of the following functional components:
– A parser module. Lexical and syntactic analyses of the source program are
performed according to the concrete grammar of the considered target language, leading to the computation of the corresponding abstract syntax tree
(AST), i.e., a Prolog term.
– A static semantics module. The AST computed in the previous phase is
examined so as to compute the static type of each program fragment and
perform several context-dependent, well-formedness checks.
– A dynamic semantics module. The concrete behavior of a program is specified by adopting the small-step style of the structural operational semantics
(SOS) approach [15]. The dynamic configurations of a running instance of
the program (i.e., the states of the transition system) are appropriately encoded in suitable Prolog terms, which are then evaluated according to the
transition systems’ rules, leading to an interpreter for the language.
The CLAIR system features two simple (though not simplistic) languages: SFL,
a functional language; and SIL, a Pascal-like imperative language. By programming in these languages, students are able to see at first hand how the rather
A Prolog-based environment for reasoning about programming languages
3
dry rules of the structural operational semantics [15] do in fact deliver a working
system capable of executing non-trivial programs, while providing a theoretical
framework for proving interesting program properties. These pedagogic experiences hinted at how the same cleanness and flexibility goals that were driving
the development of a teaching tool could also be pursued in the much more
challenging context of research, leading to the development of ECLAIR (Extended CLAIR), whose overall aim is the analysis of mainstream programming
languages. The newly targeted system, whose specification, implementation and
evaluation is work in progress, has led to both the design of new functional
modules and the restructuring of existing ones:
– The parser module has been extended with several other language instances
including, e.g., the Java bytecode language and an almost complete parser
for standard C.
– The static semantics module has been instrumented to save the collected
type information in a program annotation database, so as to make it readily
available to later phases.
– The dynamic semantics module has been refactored so as to distinguish
the implementation of the concrete semantics rules (now given in the style
of the big-step semantics, relating programs to final configurations) from
the so-called concrete memory structure component that provides both the
concrete memory and some of the control flow management utilities. The
concrete memory structure component is partially implemented in the C++
language so as to greatly improve the efficiency of the generated interpreter,
which is now able to execute non-toy programs using a reasonable amount
of system resources.
– A new abstract semantics module has been designed so as to allow for the
computation of the abstract semantics of the program. As for the concrete
dynamic semantics, a distinction is made between the abstract semantics
rules and the specification of the abstract memory structure component.
This last component is intended as a generic interface for many of the abstract domains and operators that have been proposed for static analysis
applications, often implemented in a foreign language such as C++. The abstract semantics is obtained by means of a post-fixpoint computation: the
intermediate results are saved, using the program annotation database, as
annotations attached to the nodes of the AST; care is taken so as to allow for
the specification of context sensitive analyses, whereby different annotations
can be attached to the same program fragment depending on the calling
context.
Reflections on using Prolog. Prolog has many benefits for the implementation of a system such as ECLAIR, primarily due to the fact that Prolog is based
on Horn clauses which are the basis of the formal specification of all the main
components. Observe that, for the parser, we have the added advantage that the
syntax can be specified by means of the Prolog definite clause grammar rules,
which —being a notational variant of Horn clauses— can be executed directly
4
R. Bagnara, P. M. Hill, E. Zaffanella
[14]. For the other three components (static semantics, dynamic semantics and
abstract semantics), the formal specification is provided in the form of sequents
which map directly to Horn clauses; therefore each component can be regarded
as an executable form of its specification.
It follows that, for each of the supported languages, the implementation of
the formal static and concrete semantics allows us to perform a comprehensive
series of tests so as to evaluate different aspects of its specification and hence
help us build confidence that it does accurately match the intended semantics;
note that, in the case of a real language such as C, the test results can be
compared directly with those obtained using a standard C compiler. Regarding
the abstract semantics rules, these also are almost directly translated to Prolog
code. However, in this case, these rules are incomplete in the sense that they are
parametric on the abstract domain and therefore the code implementing them
must be interfaced with specialized libraries that support a variety of abstract
domains. For ECLAIR, we use the Parma Polyhedra Library so that all the numerical abstractions provided by the PPL such as polyhedra, octagons, intervals
and grids can be used by the analyzer [3–5]. Note that, as the PPL provides
an almost complete interface to a number of Prolog systems, the ECLAIR to
PPL interface can be written entirely in Prolog. We have though had to find
appropriate strategies to overcome some inherent weaknesses in Prolog.
– Some components of the tool require great efficiency and are better expressed
in an imperative style. This is the case for the concrete memory structure
that implements, among other things, destructive updates. As already explained, by partially implementing this structure in the C++ language, we
have been able to significantly improve the performance of the interpreter.
Also, for the analyzer, it is becoming clear that coding the abstract memory
structure in Prolog (using the Prolog interface of the PPL for the numerical abstract domains) is cumbersome. We will thus realize that component
mainly in C++ and interface it to the Prolog analyzer using the generic,
portable Prolog-C++ interface distributed with the PPL (see [1, 5]). The
C++ module of the abstract memory structure will in turn call Prolog for all
the symbolic manipulation tasks.
– The lack of types in Prolog makes simple type errors hard to detect. One
possible solution is to use optional type declarations such as those provided
by the Ciao Prolog system [6]. However, such non-standard annotations,
could lead to a loss of portability of the ECLAIR system. A possibility we
are considering is the use of the TCLP prescriptive type system [12], the
main problem being that it does not currently support SWI-Prolog.
– Well-known deficiencies of Prolog module systems (first of all their flat, not
hierarchical nature) make the development of a modular system like ECLAIR
more difficult than it ought to be.
Conclusion with ongoing and future work. We have sketched the overall
structure of ECLAIR, outlining the main methodologies and technologies underlying its components. We also have justified why we chose Prolog as the imple-
A Prolog-based environment for reasoning about programming languages
5
mentation language. Apart from the continuing development of the concrete and
abstract interpreters, current work includes investigating several applications for
the analyzer, particularly: string cleanness, absence of integer overflows, correctness of array operations, inference of ranking functions and termination analysis.
Further interesting extensions of the system are foreseen. In particular, the development of compiler modules (including code generation and optimizations):
as well as the obvious applications in a teaching context, this will allow to evaluate the usefulness of the information extracted by the abstract semantics module
also from the standpoint of optimized compilation.
References
1. R. Bagnara and M. Carro. Foreign language interfaces for Prolog: A terse survey.
Quaderno 283, Dipartimento di Matematica, Università di Parma, Italy, 2002.
Version 1. Available at http://www.cs.unipr.it/Publications/.
2. R. Bagnara, P. M. Hill, A. Pescetti, and E. Zaffanella.
On the design
of generic static analyzers for modern imperative languages. Technical Report arXiv:cs.PL/0703116, Dipartimento di Matematica, Università di Parma,
Italy, 2007. Available from http://arxiv.org/.
3. R. Bagnara, P. M. Hill, E. Ricci, and E. Zaffanella. Precise widening operators for
convex polyhedra. Science of Computer Programming, 58(1–2):28–56, 2005.
4. R. Bagnara, P. M. Hill, and E. Zaffanella. Not necessarily closed convex polyhedra
and the double description method. Formal Aspects of Computing, 17(2):222–257,
2005.
5. R. Bagnara, P. M. Hill, and E. Zaffanella. The Parma Polyhedra Library: Toward a
complete set of numerical abstractions for the analysis and verification of hardware
and software systems. Quaderno 457, Dipartimento di Matematica, Università di
Parma, Italy, 2006. Available at http://www.cs.unipr.it/Publications/. Also
published as arXiv:cs.MS/0612085, available from http://arxiv.org/.
6. F. Bueno, D. Cabeza, M. Carro, M. Hermenegildo, P. López-Garcı́a, and G. Puebla.
The Ciao Prolog system. Reference manual. Technical Report CLIP3/97.1, School
of Computer Science, Technical University of Madrid (UPM), 1997. Available from
http://www.clip.dia.fi.upm.es/.
7. P. Cousot. The verification grand challenge and abstract interpretation. In Verified
Software: Theories, Tools, Experiments (VSTTE), ETH Zürich, Switzerland, 2005.
Position paper.
8. P. Cousot and R. Cousot. Abstract interpretation: A unified lattice model for static
analysis of programs by construction or approximation of fixpoints. In Proceedings
of the Fourth Annual ACM Symposium on Principles of Programming Languages,
pages 238–252, New York, 1977. ACM Press.
9. P. Cousot and R. Cousot. Abstract interpretation frameworks. Journal of Logic
and Computation, 2(4):511–547, 1992.
10. P. Cousot and R. Cousot. Inductive definitions, semantics and abstract interpretation. In Proceedings of the Nineteenth Annual ACM Symposium on Principles
of Programming Languages, pages 83–94, Albuquerque, New Mexico, USA, 1992.
ACM Press.
11. P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné, D. Monniaux, and X. Rival.
The ASTRÉE analyzer. In M. Sagiv, editor, Programming Languages and Systems,
6
12.
13.
14.
15.
16.
17.
18.
R. Bagnara, P. M. Hill, E. Zaffanella
Proceedings of the 14th European Symposium on Programming, volume 3444 of
Lecture Notes in Computer Science, pages 21–30, Edinburgh, UK, 2005. SpringerVerlag, Berlin.
F. Fages and E. Coquery. Typing constraint logic programs. Theory and Practice
of Logic Programming, 1(6):751–777, 2001.
G. Kahn. Natural semantics. In F.-J. Brandenburg, G. Vidal-Naquet, and M. Wirsing, editors, Proceedings of the 4th Annual Symposium on Theoretical Aspects of
Computer Science, volume 247 of Lecture Notes in Computer Science, pages 22–39,
Passau, Germany, 1987. Springer-Verlag, Berlin.
F. C. N. Pereira and D. H. D. Warren. Definite clause grammars for language
analysis - a survey of the formalism and a comparison with augmented transition
networks. Artificial Intelligence, 13(3):231–278, 1980.
G. D. Plotkin. A structural approach to operational semantics. Journal of Logic
and Algebraic Programming, 60–61:17–139, 2004.
D. A. Schmidt. Natural-semantics-based abstract interpretation (preliminary version). In A. Mycroft, editor, Static Analysis: Proceedings of the 2nd International
Symposium, volume 983 of Lecture Notes in Computer Science, pages 1–18, Glasgow, UK, 1995. Springer-Verlag, Berlin.
D. A. Schmidt. Abstract interpretation of small-step semantics. In M. Dam, editor, Analysis and Verification of Multiple-Agent Languages, volume 1192 of Lecture Notes in Computer Science, pages 76–99. Springer-Verlag, Berlin, 1997. 5th
LOMAPS Workshop Stockholm, Sweden, June 24–26, 1996, Selected Papers.
D. A. Schmidt. Trace-based abstract interpretation of operational semantics. LISP
and Symbolic Computation, 10(3):237–271, 1998.
| 6cs.PL
|
A LAGRANGIAN GAUSS–NEWTON–KRYLOV SOLVER FOR MASS- AND
INTENSITY-PRESERVING DIFFEOMORPHIC IMAGE REGISTRATION
arXiv:1703.04446v2 [cs.CV] 12 Jul 2017
ANDREAS MANG∗ AND LARS RUTHOTTO†
Abstract. We present an efficient solver for diffeomorphic image registration problems in the framework of Large Deformations Diffeomorphic Metric Mappings (LDDMM). We use an optimal control formulation, in which the velocity field of a
hyperbolic PDE needs to be found such that the distance between the final state of the system (the transformed/transported
template image) and the observation (the reference image) is minimized. Our solver supports both stationary and non-stationary
(i.e., transient or time-dependent) velocity fields. As transformation models, we consider both the transport equation (assuming
intensities are preserved during the deformation) and the continuity equation (assuming mass-preservation).
We consider the reduced form of the optimal control problem and solve the resulting unconstrained optimization problem
using a discretize-then-optimize approach. A key contribution is the elimination of the PDE constraint using a Lagrangian
hyperbolic PDE solver. Lagrangian methods rely on the concept of characteristic curves. We approximate these curves using
a fourth-order Runge-Kutta method. We also present an efficient algorithm for computing the derivatives of the final state of
the system with respect to the velocity field. This allows us to use fast Gauss-Newton based methods. We present quickly
converging iterative linear solvers using spectral preconditioners that render the overall optimization efficient and scalable.
Our method is embedded into the image registration framework FAIR and, thus, supports the most commonly used similarity
measures and regularization functionals. We demonstrate the potential of our new approach using several synthetic and real
world test problems with up to 14.7 million degrees of freedom.
Key words. Diffeomorphic Image Registration, Large Deformation Diffeomorphic Metric Mapping, Optimal Control,
PDE-Constrained Optimization, Lagrangian Methods
AMS subject classifications. 68U10, 49J20, 35Q93, 65M32, 76D55, 65K10.
1. Introduction. In this paper, we present efficient numerical methods for diffeomorphic image registration in the framework of Large Deformation Diffeomorphic Metric Mapping (LDDMM) [65, 23, 9]. We
use an optimal control formulation similar to the one in [11, 12, 10, 48, 50, 49]. Here, the task is to find a
smooth velocity field v such that the distance between two images (or densities), T (the template image)
and R (the reference image), is minimized, subject to some regularization norm for v and a transformation
model, given by a hyperbolic partial differential equation (PDE) that models the deformation of T . We
consider the transport equation (assuming intensities are related at corresponding points) and the continuity
equation (assuming that mass is preserved) as constraints. The connection to traditional image registration
formulations [53, 54] is that a sufficiently smooth velocity field v gives rise to a diffeomorphism y via the
method of characteristics. Vice versa, representing diffeomorphisms through velocity fields has been used for
efficient statistical analysis; see, e.g., [3].
Solving the variational problem associated with LDDMM is, in theory, known to yield a diffeomorphic
transformation y if v is sufficiently smooth [23, 65, 9]. Although, the theory of diffeomorphic registration using
LDDMM is well explored [52, 70, 71], efficient numerical optimization is not. Until recently [5, 39, 48, 50, 49]
mostly first-order optimization methods were used; see, e.g., [9, 67, 17, 58]. A key component in LDDMM is
the numerical method for solving the hyperbolic PDE. Hyperbolic PDE solvers can be roughly divided into
Eulerian (in which the density is discretized at the same locations for each time point) and Lagrangian (in
which the grid moves over time along the characteristic curves) solvers; see also [45, 25, 46]. Intermediates
are Semi-Lagrangian (SL) methods, which follow the velocity for a short time step and then estimate the
density at fixed points. In this work, we use a Lagrangian solver.
It is well known that explicit Eulerian methods for hyperbolic PDEs require the size of the time step
to be sufficiently small to ensure numerical stability. The maximal admissible time step size depends on
the accuracy of the spatial discretization and the magnitude of the velocities. In optimal control problems,
like ours, the velocity field v is not known a priori and, thus, it is difficult to come up with an efficient
and stable choice of the time step. SL and Lagrangian solvers are explicit methods that, unlike explicit
Eulerian methods, are stable without a restriction on the maximal admissible time step size. One drawback
∗ Institute for Computational Engineering and Sciences, University of Texas at Austin, Austin, Texas, USA.
([email protected])
† Department of Mathematics and Computer Science, Emory University, Atlanta, Georgia, USA. ([email protected])
1
of SL methods is their memory requirements. For efficient derivative (or sensitivity) computations in Gauss–
Newton type optimization schemes, we have to store intermediate images [51]. Further, SL methods require
a repeated interpolation of the initial image and may therefore introduce severe dissipation if implemented
naively.1 As we will see, Lagrangian methods require only one image interpolation at the final time. Secondly,
derivatives of the Lagrangian solver can be obtained efficiently, without storing intermediate variables. A
feature of SL and Lagrangian methods is that they can be easily modified to solve intensity- and masspreserving problems, since the characteristic curves coincide.
The key idea of our work is to use a discretize-then-optimize strategy based on a Lagrangian hyperbolic
PDE solver to efficiently solve the reduced formulation of the PDE-constrained optimization problem arising
in LDDMM. We show that Lagrangian methods lead to a finite dimensional optimization problem that can
be solved using an inexact Gauss–Newton method. Our PDE solver requires numerical computation of the
characteristic curves. We use a fourth-order Runge–Kutta (RK4) method to numerically approximate the
transformation y that is associated with a, in general non-stationary, velocity field v. As we show, derivatives
of the transformation with respect to v can be derived analytically and computed efficiently. Due to the
hyperbolic nature of the PDE, the derivatives can be represented as sparse matrices; the procedure can
be paralellelized: characteristics starting at different points can be computed independently. Given these
characteristics, the hyperbolic PDEs can be solved by a single interpolation step (for the advection equation)
or the particle-in-cell method (for the continuity equation).
1.1. Contributions.
• We propose a discretize-optimize method for solving LDDMM using a Lagrangian hyperbolic PDE
solver. Our scheme is based on an RK4 method to approximate the characteristic curves and we
derive an efficient algorithm for computing the derivative of the solver with respect to the velocities.
• The storage requirement of our method is independent on the number of time steps used in the
numerical solver. Also, the Hessian of the objective function can be build explicitly at moderate
costs, which is useful, e.g., to accelerate matrix-vector products. In this work, we use and numerically
study the convergence of spectral preconditioners to iteratively solve the Gauss–Newton system.
• We extend the LDDMM framework to mass-preserving registration, which has been proved to be
an adequate model for many relevant biomedical applications involving with density images, e.g.,
in [16, 59, 21, 30, 61].
• We derive a flexible framework supporting both stationary and non-stationary velocity fields. Our
methods are embedded into the FAIR framework [54]. This allows us to consider different regularization norms and distance measures. Our implementation is freely available as an add-on to FAIR
at:
https://github.com/C4IR/FAIR.m/tree/master/add-ons/LagLDDMM
• We provide detailed numerical experiments on four different data sets that demonstrate the flexibility
and effectiveness of our method. We show that our prototype implementation is competitive to stateof-the-art packages for diffeomorphic image registration [15, 60]. We study registration quality for
synthetic benchmark problems and real-world applications leading to optimization problem with up
to 14.7 million degrees of freedom.
1.2. Related Work. We limit this review to work closely related to ours. For a general insight into
the area of image registration, its applications, and its formulation we refer to [53, 54, 63]. Our work
builds upon the LDDMM framework described in [23, 65, 9], which is based on the pioneering work on
velocity-based fluid registration described in [20]. We adopt an optimal control point of view; we also do not
directly invert for the transformation y but for its velocity v. We arrive at a hyperbolic PDE-constrained
optimization problem. We refer to [31, 43, 41, 13] for a general introduction into optimal control theory
and developments in PDE-constrained optimization. Related optimal control formulations for diffeomorphic
image registration can, e.g., be found in [11, 12, 38, 10, 17, 48, 50, 49, 51]. Other formulations for velocitybased diffeomorphic image registration are described in [4, 66, 5]. Our work also shares characteristics with
optical flow formulations [11, 12, 17]. Our formulation for mass-preserving registration problems is related
1 We note, that interpolation errors and numerical diffusion can be minimized, by, e.g., applying high-order interpolation
schemes and/or evaluating the interpolation on a finer grid; this is costly.
2
to the Monge–Kantorovich functional arising in optimal mass transport [10, 35].
Most work on large deformation diffeomorphic image registration still considers first-order methods for
numerical optimization (see, e.g., [6, 7, 8, 11, 9, 38, 44, 17, 67]); the exceptions are [5, 48, 50, 49, 66]. Firstorder schemes for numerical optimization do in general require a larger number of iterations than Newton
type optimization schemes.2 The work in [5] uses geodesic shooting and estimates the initial value of a
non-stationary velocity field that parameterizes the diffeomorphism y. Other approaches that reduce the
size of the optimization problem are based on stationary velocity fields; see, e.g., [40, 47, 50, 49, 51].
PDE-constrained optimization commonly requires a repeated solution of the forward problem. Thus, the
design of an efficient forward solver is critical. The approaches described in [10, 17, 48, 50, 49, 38] are based on
an Eulerian formulation. They employ explicit high-order schemes [11, 12, 48, 50, 38, 58], which suffer from
a restriction on a maximally admissible time step, implicit schemes [10], or explicit SL schemes [49, 51, 17].
The latter were originally proposed in the context of weather prediction [64]. They are a hybrid between
Lagrangian and Eulerian schemes, and unconditionally stable. SL schemes have been used in the context
of Lagrangian formulations for diffeomorphic image registration (to compute the characteristics) [9, 40].
Conditionally stable schemes require small time steps, which can result in a significant amount of memory
that needs to be allocated to store the time-space fields necessary to evaluate the gradient or Hessian. This
makes a direct application of these type of methods to large-scale 3D problems challenging. One remedy is
to turn to parallel architectures and use sophisticated checkpointing schemes in time to reduce the amount
of memory that has to be allocated; see, e.g., [1]. Implicit schemes typically suffer from severe numerical
diffusion. The same is true for straightforward implementations of SL schemes. Pure Lagrangian schemes
for diffeomorphic image registration have been described in [6, 7, 8]. The time integration for computing the
characteristics in [6, 7, 8] is based on a first-order explicit scheme.
What sets our work apart is the numerical solver and the generalization of our formulation to both the
transport and the continuity equation. Most existing works on optimal control formulations for diffeomorphic
image registration consider an optimize-then-discretize approach [17, 48, 50, 49, 51]. We use a discretizethen-optimize strategy instead.3 Similar to [48], we describe a method that can handle stationary and
non-stationary velocity fields. Our numerical scheme is, likewise to [6, 7, 8], based on a purely Lagrangian
approach. We consider a reduced formulation of the PDE-constrained optimization problem arising in
LDDMM, i.e., we eliminate the hyperbolic PDE constraint (state equation) from the variational problem.
We use a Lagrangian solver to parameterize the final state in terms of the velocity. In doing so we avoid many
of the complications we reviewed above: our method is unconditionally stable, limits numerical diffusion,
and does not require the storage of multiple space-time fields for the evaluation of the gradient or Hessian
operators. The work in [6, 7, 8] uses first-order information for numerical optimization and a first-order
accurate explicit time integrator to compute the characteristic. We use a fourth-order, explicit RK scheme
instead. We derive expressions for the exact derivative of the characteristics. We arrive at a Gauss–Newton–
Krylov scheme that—combined with an efficient iterative linear solver—yields an approximate solution within
a few iterations, and has an overall algorithmic complexity of O(n log(n)), where n is the dimension of the
discretized velocity field (i.e., the number of unknowns).
2. Mathematical Formulation. We describe the variational optimal control formulation of the LDDMM problem next. We denote the image domain by Ω ⊂ Rd , where d ∈ {2, 3} represents the spatial
dimension. We assume that the template and the reference image, denoted by T : Ω → R and R : Ω → R,
are compactly supported on Ω and continuously differentiable. Given these two images, the task of image
registration is to find a plausible transformation y : Ω → Rd so that the transformed template image T ◦ y
becomes similar to the reference image R [54]. The definitions of plausibility, similarity, and the transformation model depend on the context; see [54, 53, 63] for examples. Many relevant applications, e.g., in medical
imaging, require that plausible transformations are diffeomorphic, i.e., smooth mappings with a smooth
inverse. One framework that contains the most commonly used definitions of these three terms within the
medical imaging application domain is LDDMM [23, 65, 9]. The variational optimal control formulation of
2 We note that in LDDMM most implementations use a gradient descent scheme in the Sobolev space induced by the
regularization operator (dual space). This leads to a significant speedup compared to standard gradient descent approaches.
However, it has been demonstrated experimentally that Gauss–Newton–Krylov methods are superior [48].
3 Advantages and disadvantages of these two techniques are discussed, e.g., in [31].
3
Table 1.1
Commonly used symbols and abbreviations.
DCT
FAIR
LDDMM
MRI
PDE
PET
PIC
SSD
discrete cosine transform
Flexible Algorithms for Image Registration [54]
large deformation diffeomorphic metric mapping
magnetic resonance imaging
partial differential equation
positron emission tomography
particle-in-cell (method)
sum-of-squared-differences
R(x)
T (x)
C
D
S
α
x
Ω
v(x, t)
u(x, t)
y(x)
N
n
nt
I
∇
∇·
∂t
reference / fixed image
template image (image to be registered)
PDE constraint
distance or similarity measure
regularization model (smoother)
regularization weight
spatial coordinate; x ∈ Ω
spatial domain; Ω ⊂ Rd
velocity field
transported image intensities
transformation / mapping
number of time steps for computing the characteristic
number of unknowns (i.e., the dimension of the discretized velocity field)
number of cells in space-time grid
interpolation operator
gradient operator
divergence operator
time derivative
the LDDMM problem can, in general format, be stated as follows:
min {J (v, u) := D(u(·, 1), R) + αS(v)}
v,u
subject to
C(v, u) = 0,
(2.1)
where D is a distance (or similarity) measure, S is a regularizer (smoother), v : Ω × [0, 1] → Rd is the sought
after velocity field, and u : Ω × [0, 1] → R is a time series of images. In an optimal control context, v is
commonly referred to as the control variable and u as the state variable. Here, α > 0 is a regularization
parameter that balances between minimizing the image distance and the smoothness of the velocity field (and
consequently controls the properties of the resulting transformation). In the current work, we assume that
α is chosen by the user and refer to [36, 34, 68] for some works on automatic selection of the parameters.4
In this work, we assume that the constraint C, which describes the transformation model, is given either
by the advection (also called transport) equation
∂t u(x, t) + v(x, t) · ∇u(x, t) = 0,
C(u, v) =
(2.2)
u(x, 0) = T (x)
or the continuity equation
C(u, v) =
∂t u(x, t) + ∇ · (u(x, t)v(x, t)) = 0,
u(x, 0) = T (x).
(2.3)
The former assumes intensity values are preserved during the transformation; the latter preserves the overall mass of the image. The choice of the transformation model depends on the application. Intensity
preservation is commonly used, e.g., for registration of medical images acquired from different subjects [56].
Mass-preservation has been successfully used, e.g., for motion correction in position emission tomography (PET) [21, 30] or artifact correction of magnetic resonance imaging (MRI) [16, 61].
The models in (2.2) and (2.3) can be used to establish point-to-point correspondences between the
template and the reference image. One way of showing this is the method of characteristics [46, 24]. To
better illustrate this, we consider the advection equation (2.2), for which the intensity u is constant along
4 Examples for an automatic selection of the regularization parameter in the context of image registration can, e.g., be found
in [33, 48].
4
the characteristics. This means that, for all y0 ∈ Ω and t ∈ [0, 1], it holds that u(y(v, y0 , 0, t), t) = T (y0 )
where the characteristic curve t 7→ y(v, y0 , 0, t) satisfies
∂t y(v, y0 , 0, t) = v(y(v, y0 , 0, t), t)
and
y(v, y0 , 0, 0) = y0 .
(2.4)
Similarly, the characteristics can be traced backwards in time to compute the state at some point y1 ∈ Rd at
t = 1. The position of the point is given by t 7→ y(v, y1 , 1, t), which satisfies (2.4) with final time condition
y(v, y1 , 1, 1) = y1 . Clearly, both operations are inverse to one another. That is, for all x ∈ Rn it holds that
y(v, y(v, x, 0, 1), 1, 0) = x, i.e., a composition of both maps yields identity.
Note that (2.2) through (2.4) involve a non-stationary (i.e., time-dependent) velocity field v. This is a
key assumption in the original LDDMM formulation [9]. To make the problem computational tractable it is,
however, often assumed that v is stationary (stationary velocity field based registration) [3, 4, 40, 50, 51, 57].5
The numerical framework we propose in this paper can efficiently handle both stationary and non-stationary
velocities. As demonstrated in our numerical experiments in Sec. 4.2, the stationary model is less flexible
in that we can only invert for a subset of the deformation maps living on the manifold of diffeomorphisms.
Our experiments also suggest that stationary velocity fields are adequate for registration problems involving
two topologically similar images, yielding little to no difference in the recovered deformation map [4, 40, 48].
However, using non-stationary velocity fields may become critical in applications involving large and highly
nonlinear transformations and/or the registration of time series of images with large motion between time
frames (e.g., typically seen in tracking or optical flow problems).
2.1. Regularization functionals. Due to the ill-posedness of the image registration problem, the
literature on regularization in image registration is rich; see, e.g., [53, 27, 54, 63] for extensive overviews.
It is established that the existence and regularity of a diffeomorphic map y depends on the smoothness of
the velocity field v as well as the smoothness of the images R and T [23, 65, 50, 17]. Modeling the images
as functions of bounded variation, H 3 -regularity [17] is required (assuming that v is divergence free). For
continuous images we can relax the H 3 -regularity to an H 2 -regularity [9] or—under additional assumptions
on the divergence of v—even to an H 1 -regularity [17].6 Most implementations for LDDMM consider an
H 2 -norm for S in (2.1) (or an approximation based on a Gaussian kernel within a gradient descent scheme
in the Sobolev space induced by the regularization norm); see, e.g., [5, 9, 40].
In our framework, we regard the regularizer as a modular component that can be replaced or extended.
In practical applications we control the regularization parameter by monitoring the Jacobian in an attempt to
generate transformations that are diffeomorphic (in a discrete setting) and yield high-fidelity (low mismatch)
results. In our numerical examples, we consider H 1 - and H 2 -seminorms as regularization models. These type
of regularization models are also referred to as diffusive or curvature regularizers in the context of traditional
variational image registration formulations [26, 53, 54], respectively. Assuming that v is non-stationary, we
have
S diff (v) =
1
1
2
Z
1
2
Z
0
Z X
d
|∇v k (x, t)|2 dxdt
Ω k=1
and
S curv (v) =
0
1
Z X
d
|∆v k (x, t)|2 dxdt,
Ω k=1
respectively. Regularization of stationary velocity fields is along the same lines, by simply dropping the time
integration in the above equations. Our formulation can also be extended to enforce smoothness in time, as
also done in [11].
5 Another strategy to reduce the computational burden is to invert for an initial momentum [67] that encodes the trajectory
of the diffeomorphism.
6 We use interpolation and padding of the discrete image data to obtain continuously differentiable and compactly supported
functions (see Sec. 3).
5
3. Numerical Methods. In this section, we describe a discretize-then-optimize approach for solving
the variational problem (2.1). We eliminate the hyperbolic PDE constraints (2.2) and (2.3) using Lagrangian
methods. We then describe the discretization of the objective functional itself. Following [54], we consider
a multilevel Gauss–Newton method that allows us to efficiently solve the discrete optimization problem.
Finally, we give some details about our implementation as an extension to the FAIR toolbox [54].
Assume, for simplicity, that the domain Ω = (0, 1)d is divided into a regular mesh of m cells of edge
length h = 1/m along each coordinate direction. We use interpolation and padding of the discrete image
data to obtain continuously differentiable and compactly supported functions. We approximate integrals
in (2.1) by a midpoint quadrature rule, which requires evaluating the final state on the cell-centered points,
d
xc ∈ Rd·m , of the grid. Without loss of generality, we assume that the velocity field v is discretized on the
same mesh. However, the domain size and number of cells can be varied in practice. The discrete velocity
field, denoted by v ∈ Rn , is discretized in time at the nodes of a regular grid with nt cells and in space at
cell-centered grid points. The total number of unknowns is n = (nt + 1) · d · (md ).
3.1. Lagrangian Methods for Hyperbolic PDEs. Lagrangian methods exploit the fact that solutions to hyperbolic PDEs evolve along characteristic curves [46, 24]. These methods are Lagrangian in the
sense that the transport of the density is referred to in the moving coordinate system. Lagrangian methods
typically consist of two steps: First, the characteristics are computed numerically. Then, in a second step,
the final image (or density) is computed. While the characteristic curves are identical for the advection and
the continuity equation, the computation of the second part is not.
Step 1 (Computing the Characteristics). Here, we describe our implementation for computing the characteristic curve passing through a given point. The velocity field v : Ω × [0, 1] → Rd is known and represented
by the coefficients v ∈ Rn . We solve (2.4) using an RK4 method with N equidistant time steps of size
∆t = ±1/N ; the sign of the time step depends on whether the characteristics are computed forward or
backward in time. Note that the number of time steps N for the RK4 method and the number of cells nt in
the space-time grid do not necessarily have to be equal. The former is a parameter of the numerical solver
and controls the accuracy of the characteristics. The latter is a modeling parameter and ultimately controls
the search space for the transformation y.
Our choice of an RK4 scheme is motivated by accuracy considerations for computing the characteristics.
For simplicity, we illustrate the concept of integrating v based on a first-order forward Euler scheme. The
derivation of our RK4 scheme is along the same lines; it is outlined in Algorithm 1.
Let x ∈ Rd·np be the coordinates of the start (or end) points of the characteristics, e.g., the cell-centers
of a regular mesh. Introducing the (time-dependent) transformation y : Rn × Rd·np × [0, 1]2 → Rd·np , and
imposing the initial condition y(v, x, 0, 0) = x, we compute
y(v, x, 0, tk+1 ) = y(v, x, 0, tk ) + ∆t I(v, y(v, x, 0, tk ), tk ),
∀ k = 0, 1, . . . , N − 1,
(3.1)
where tk = k∆t are the time points and I interpolates the velocity field v at the transformed points y. In
our experiments, we use a bi- or trilinear interpolation model in space and a linear interpolation model in
time, applied separately to each component of the velocity field v. We found by experimentation that using
low-order interpolation schemes for the (smoothness regularized) velocity fields is sufficiently accurate for our
numerical scheme to yield high-fidelity results in practical applications. We note that the interpolation is a
modular component; it can be replaced by (computationally more expensive) higher-order methods. Notice,
that the positions at previous time steps, y(v, x, 0, 0), . . . , y(v, x, 0, tk−1 ), do not enter the computation
in (3.1); the memory requirements of our numerical scheme are independent of the number of time steps
N . The end points of the characteristics are then y(v, x, 0, 1), which, if x = xc , can also be interpreted and
visualized as a deformed regular grid.
Step 2 (Solving the hyperbolic PDE). While solutions to both the transport and the continuity equations
evolve along the same characteristics, the steps for computing the transported quantity at t = 1 vary. Thus,
we discuss both cases separately. An illustration of both schemes can be found in Fig. 3.1.
Transport equation: Considering the transport equation (2.2), we compute the intensities of the advected
image T on the deformed cell-centered grid xc by following the characteristics backwards in time. This yields
u(xc , 1) = T (y(v, xc , 1, 0)).
6
(3.2)
Algorithm 1 RK4 method for computing the characteristics y and the derivative dv y.
Input: Discrete (non-stationary) velocity field v ∈ Rn , start points of characteristics x ∈ Rd·np , number
of time steps N
Set: y ← x, dv y ← 0, ∆t ← 1/N .
for k = 0, 1, . . . , N − 1 do
compute v1 ← I v, y, 0, tk (spatio-temporal
vector field interpolation) and set y1 ← y + (∆t/2)v1
compute v2 ← I v, y1 , 0, tk+1/2 and set y2 ← y + (∆t/2)v2
compute v3 ← I v, y2 , 0, tk+1/2
and set y3 ← y + (∆t/2)v3
compute v4 ← I v, y3 , 0, tk+1
if derivative required then
D1 ← dv I v, y, 0, tk + dyI v, y, 0, tk dv y
D2 ← dv I v, y1 , 0, tk+1/2 + dy I v, y1 , 0, tk+1/2 dv y + (∆t/2)D1
D3 ← dv I v, y2 , 0, tk+1/2
2
+ dy I v, y2 , 0, tk+1/2 dv y + (∆t/2)D
D4 ← dv I v, y3 , 0, tk+1 + dy I(v, y3 , 0, tk+1 ) dv y + ∆tD3
dv y ← dv y + (∆t/6) D1 + 2D2 + 2D3 + D4
end if
y ← y + (∆t/6) v1 + 2v2 + 2v3 + v4
end for
Output: end of characteristics, y ∈ Rd·np , and (if required) gradient, dv y ∈ Rn×d·n
In general, y(v, xc , 1, 0) does not coincide with a grid point; the intensity has to be computed by interpolation. In our numerical experiments we use a bi- or tri-linear interpolation model and regularized cubic
approximation methods provided in FAIR to obtain the intensity values of the deformed image; see [54] for
implementation details and other common choices.
Continuity equation: To solve the continuity equation (2.3) we consider the Particle-In-Cell (PIC)
method that pushes mass along the characteristics forward in time; see, e.g., [18]. For a given grid point
y0 ∈ Rd , we introduce a particle with its mass given by the intensity value T (y0 ). Then, we follow the
trajectory of the particle along the characteristic to its final point y1 := y(v, y0 , 0, 1). In general, y1 does not
coincide with a grid point. We obtain the value of the final state in a given cell by integrating the mass of all
particles whose support intersects the cell. Equivalently, we compute the final density by splitting the mass
of each particle among the cells adjacent to its final location. Ideally, the particles are represented by Dirac
delta functions. In practice, we consider bi- or tri-linear hat functions of a certain isotropic width δ > 0 as
proposed in [18].
Using the push-forward matrix F, which has also been used in [29], this process can be written as
u(xc , 1) = F(y(v, xc , 0, 1))T (xc ).
(3.3)
In the following, we construct the push-forward matrix in a way that ensures mass-preservation at the
discrete level for any choice of δ and h. For ease of presentation, we describe the procedure for the one
dimensional case (d = 1). The derivation extends to tensor meshes in higher dimensions in a straightforward
way under the assumption that the basis functions are piecewise polynomials in the coordinate directions.
Let us assume that for j = 1, 2, . . . , m particles are located at the cell-centered points xj = (j − 21 )h and
their respective mass is given by u0,j = T (xj ). The particles are advected to the points yj = y(v, xj , 0, 1).
For some δ > 0 the particles are represented by the shifted basis functions
1 − (yj − x)/δ for yj − δ ≤ x ≤ yj ,
bδ (x, yj ) = 1 + (x − yj )/δ for yj < x ≤ yj + δ,
0
else,
for each j = 1, 2, . . . , np . The mass of u(·, 1) contained in the ith interval [xi , xi+1 ] is given by
np Z xi+1
np
np
X
X
X
δ
δ
δ
uj Fij ,
ui (v, 1) =
uj b (x, yj )dx =
uj B (xi+1 , yj ) − B (xi , yj ) =
j=1
xi
j=1
j=1
7
(3.4)
y(v, x, 0, 1)
x
x
y(v, x, 1, 0)
advection; T (y(v, x, 1, 0)))
continuity; F(y(v, x, 0, 1))T (x)
Fig. 3.1. Illustration of Lagrangian methods for solving linear hyperbolic PDEs. In both cases, the characteristic curves
(indicated by a blue line) are computed starting from a grid point x. Left: The advection problem is solved by traveling
along the characteristics backwards in time to the non-grid point y(v, x, 1, 0). The associated image intensity is computed by
interpolating the intensities of the adjacent cells. Right: The continuity equation is solved by pushing the mass T (x) from x
along the characteristics to the non-grid point y(v, x, 0, 1) and then distributing T (x) among the cells adjacent to y(v, x, 0, 1).
where B δ denotes an anti-derivative of bδ and is given by
0
for x < yj − δ,
x − 1 (2y x − x2 ) for y − δ ≤ x ≤ y ,
j
j
j
2δ
B δ (x, yj ) =
1
2
x
+
(x
−
2y
x)
for
y
<
x
≤
y
+
δ,
j
j
j
2δ
1
for yj + δ < x.
Repeating the process outlined in (3.4) for all i = 1, 2, . . . , m yields the discrete transformed density,
which is summarized in (3.3). Note that our scheme is mass-preserving at the discrete level by design since
exact integration is performed. In other words: the columns in F sum to one regardless of the choices for δ
and h. Also note that F is sparse. The level of sparsity depends on the ration between δ and h. If we choose
np = m, δ = h, and xj = h(j + 1/2), it is easy to see that F is the transpose of the linear interpolation
matrix [29]. Thus, the relation between (3.2) and (3.3) mirrors the adjoint relation between the continuity
and the advection equation.
3.2. Optimization. Using the Lagrangian methods outlined above we parametrize the final state in
terms of the velocities, which we denote by u1 (v). We eliminate the state equation from the variational
problem (2.1) and—upon discretization—obtain the finite-dimensional unconstrained problem
min {J(v) := D(u1 (v), R) + αS(v)} ,
v
(3.5)
where D and S are discrete versions of the distance measure D and regularizer S in (2.1). As an example,
we consider the squared L2 -distance functional. We note, that this is a modular building block of our
formulation; for other choices we refer to [54]. Using a midpoint rule, the discrete distance measure—also
known as the sum-of-squared-differences (SSD)—reads
DSSD (u1 (v), R) =
hd
res(v)> res(v),
2
where
res(v) := u1 (v) − R.
(3.6)
To enable a Gauss–Newton optimization, we compute the derivative of the objective function. Using the
chain rule we obtain
dv J(v) = dv u1 (v)du1 D(u1 (v), R) + αdv S,
where the derivative of the distance measure with respect to the final state u1 is for (3.6) given by
du1 DSSD (u1 (v), R) = hd res(v).
8
We again refer to [54] for the derivatives for other distance measures. The derivative of the regularizer can
be written as
dv S(v) = ∆thd (B> B)v.
Here, B is a discretization of the spatial or spatio-temporal derivative operator. Similarly, the approximated
Hessian is given by
H(v) ≈ d2 J(v) = dv u1 (v) d2 D(u1 (v), R) dv u1 (v)> + α ∆thd B> B + γIn ,
where In ∈ Rn×n denotes the identity matrix and γ > 0 is a small constant to ensure positive semidefiniteness. In our numerical experiments we use γ = 0.01. The Hessian of the distance measure is given
by
d2 DSSD = h3 In .
Next, we compute the derivative of the mapping v 7→ u1 (v). We first consider the advection equation.
Using the chain rule to differentiate (3.2) we obtain
dv u1 (v) = dv T (y(v, xc , 1, 0)) = ∇T (y(v, xc , 1, 0)) dv y(v, xc , 1, 0).
The first term in the product is an image gradient evaluated at the end points of the characteristic curves. It
is computed by differentiating the interpolation model; see [54] for details. The second term is the derivative
of the endpoint of the characteristic curve with respect to v. How we compute this derivative is explained
below. For the continuity equation (3.3) we obtain
dv u1 (v) = dv (F(y(v, xc , 0, 1))T (xc )) = dy (F(y(v, xc , 0, 1))T (xc )) dv y(v, xc , 0, 1).
The first term can be computed by differentiating the terms in (3.4), for which Fij > 0, with respect to the
end points of the characteristic curves. Notice that this also implies that dy (F(y(v, xc , 0, 1))) is at least as
sparse as the push-forward matrix.
We now present an efficient way for computing the derivative of the end point of the characteristics with
respect to the velocity field. Since we use explicit time stepping schemes the derivative can be computed
recursively alongside the computation of the characteristics. For example, if we use the forward Euler method
in (3.1) we have dv y(v, xc , 1, 0) = 0; we obtain
dv y(v, xc , 1, tk+1 ) = dv y(v, xc , 1, tk )+∆t dv I(v, y(v, xc , 1, tk ), tk ) + . . .
. . . +∆t dy I(v, y(v, xc , 1, tk ), tk ) dv y(v, xc , 1, tk ),
for all k = 0, 1, . . . , N − 1. The derivatives of the interpolation scheme, dv I and dy I, are computed as
described in [54]. Notice that we do not need dv y(v, x, 1, tk ) in subsequent time steps. Thus, in practice,
we update it directly. It is straightforward to extend this procedure to other explicit methods, such as the
RK4 scheme used in our experiments; see Algorithm 1 for details. We emphasize that neither intermediate
transformations nor intermediate state variables need to be stored to compute the derivative. Therefore,
the storage requirement is essentially independent of the number of time steps N used to compute the
characteristics. This is different to the methods described in [9, 48, 50, 51], which require storing at least
one time-dependent scalar field to evaluate the gradient or Hessian operator.
We use a standard inexact Gauss–Newton–Krylov method for solving the finite-dimensional optimization
problem. We use the implementation and stopping conditions described in [54]. As to be expected, the
computationally most challenging task is the computation of the search direction. Let vi denote the velocity
field at the ith iteration. Given vi we obtain the search direction δv by solving
H(vi )δv = −dv J(vi ).
(3.7)
The next iterate vi+1 is computed via vi+1 = vi + µδv; an Armijo linesearch is performed to determine the
step size µ (see, e.g., [55]).
9
We solve the symmetric and positive definite linear system in (3.7) via a Cholesky factorization or,
for large-scale problems, via iterative methods such as the conjugate gradient (CG) method [42]. The
convergence of iterative methods depends on the clustering of the eigenvalues of H, which can be improved by
appropriate preconditioning [62]; yielding a preconditioned CG (PCG) method. For the examples considered
in this paper, the Hessian of the regularizer is block diagonal with d blocks corresponding to a discretized
second- or fourth-order differential operator. Since the Hessian of the regularizer is of higher-order as
compared to the Hessian of the distance term, we exploit the structure of A = B> B for preconditioning.
Given that the velocity is discretized on a regular mesh in space and time, A is a structured matrix and can
be written as a sum of Kronecker products of Toeplitz-plus-Hankel matrices. Thus, its pseudo-inverse can be
computed efficiently using the Discrete Cosine Transform (DCT) [37]. In addition to the preconditioners
available in FAIR (such as multigrid or Jacobi) we also provide an option to use A + γIn as preconditioner;
a common choice in PDE-constrained optimization problems [48, 2].
The optimization problem in (3.5) is known to be non-convex. To limit the risk of being trapped in a
local minimum, we use a multilevel strategy similar to the one described in [54]. First, we solve (3.5) with
a coarse discretization for the distance, regularizer, and velocities, and then refine the solution and use it as
a starting guess for the optimization problem obtained on the next level. We continue this procedure until
we reach a sufficiently fine discretization level, which depends on the application at hand. In addition to
improving robustness, the scheme often leads to an overall reduction of computation time.
3.3. Implementation. We have implemented our method in MATLAB as an extension to the 2011
version of the FAIR toolbox described in [54]. This allows us to exploit all distance measures, interpolation
kernels, and numerical schemes provided in FAIR. A pseudocode of the RK4 method used to compute the
characteristics and the derivative of the end point with respect to the velocity field v is given in Algorithm 1.
It can be seen that both the characteristics and the gradient are computed in one sweep over all time points.
The characteristic and the gradient can be updated in each step. This makes the memory requirements
independent of the number of time steps N used to compute the characteristics; the size of dv u1 (v) is
n × md . The gradient matrix is sparse. Its columns will have non-zero entries only in rows associated with
discrete velocities in close proximity to the characteristic curve; we will demonstrate this experimentally; see
Fig. 4.2.
The reduced memory requirement is a significant improvement over existing Eulerian or Semi-Lagrangian
methods. These require storing (or recomputing) the transported images or characteristics for each time
step [9, 48, 50, 51]. The Lagrangian method proposed here, requires only the allocation of the transformed
grid, which is independent of the number of time steps for the forward or adjoint solves. Further, it is
possible to adapt the time step used for computing the characteristics, e.g., depending on the complexity of
the trajectory.
4. Numerical Experiments. In this section, we demonstrate the potential of our solver based on
two- and three-dimensional synthetic and real world problems of varying complexity. We compare our
prototype implementation to tailored and highly optimized state-of-the-art packages for diffeomorphic image
registration. For mass-preserving registration we consider the VAMPIRE package [30]. For large deformation
diffeomorphic registration we consider the hyperelastic registration model originally described in [15, 60].
4.1. General Setup. As stopping criteria for the Gauss–Newton optimization, we use standard settings
provided in FAIR. The maximum number of inner iterations for the PCG method is set to 50; the tolerance
for the relative residual is set to 0.1. We use a spectral preconditioner.7 The benchmark methods employ a
hyperelastic regularization model, for which effective preconditioning is more challenging; see, e.g., [15, 60].
Here, we use a matrix-free implementation of a Jacobi-PCG solver. Problem-specific parameters, such as the
number of time points to represent velocity fields, times steps in the RK4 method, the image domain, the
number of multi-level steps, or the padding of the domain used to represent the velocity field, are described
in the respective subsections. We perform all our experiments on a x68 compute node with 40 Intel(R)
Xeon(R) CPU E5-2660 processors running at 2.60GHz with a total of 256GB of memory.
7 Since the regularization operator corresponds to a block diagonal matrix whose 4 · 2 blocks are discretizations of a 2D
Laplacian, its pseudo inverse can be computed efficiently using DCTs (see Sec. 3.2).
10
We consider H 1 (diffusive) and/or H 2 (curvature) regularization models throughout our experiments.
We emphasize that, as we have already pointed out in Sec. 2.1, theoretical considerations require imposing
H 2 -regularity on v in order to guarantee that v gives rise to a diffeomorphic map y (see, e.g., [9]). Our
argument for also considering an H 1 regularization model is that, in practice, we can control the weight α
by monitoring det ∇y to ensure that the discretized map y is indeed a diffeomorphism. We also note that
the regularization is a modular block of our formulation. If theoretical requirements are of concern, one can
switch to H 2 regularity.
4.2. 2D C-Shape. We consider the classical test case of registering a C-shaped object to a disc as
initially proposed by Christensen [20]. We study registration quality and performance. We compare our
results to the hyperelastic registration method described in [15, 60]. We also study the convergence for
different types of preconditioners for this problem.
Experimental Setup. The test data is taken from FAIR and consists of two binary image data with
128 × 128 pixels on the image domain Ω = (0, 1)2 . To build a continuously differentiable image model from
the binary image data, we use the moments-regularized cubic B-spline interpolation with an experimentally
tuned smoothing parameter of θ = 0.1; see [54] for details. Since the image modality is comparable in both
images, we use the SSD distance measure in (3.6) to assess image similarity.
• LDDMM: We model the velocity field on a padded domain (−0.5, 1.5)2 to reduce boundary effects.
We use the diffusion regularizer with an empirically determined weight of α = 400; we set γ = 0.
We compare results for a stationary velocity model to those obtained for a non-stationary velocity
model with nt = 2 time intervals. We use a three-step multilevel strategy and discretize the domain
for the velocities using regular meshes with 322 , 642 , and 1282 cells, respectively. The characteristics
are computed using an RK4 method with N = 3 time steps. We assess registration quality and the
impact of different preconditioning techniques (no preconditioning, Jacobi preconditioning, Symmetric Gauss Seidel, and spectral) on the convergence of the PCG method used to approximately
solve (3.7). For the convergence study, we only consider the coarsest discretization level (32 × 32
cells). The structure of the Hessian depends on the current velocity estimate. We compare the
convergence of the PCG method at the first and final Gauss–Newton iteration. We consider an H 1
regularization model. We report results for a stationary and a non-stationary velocity field (nt = 2).
In each case, we aim to solve the linear system up to a relative error of 10−10 and set the maximum
number of iterations to 250.
• Hyperelastic Registration [15]: We use the default parameters provided in FAIR to solve this problem. The values for the regularization are empirically chosen and set to α1 = 100 for the length
α2 = 0 for the area, and α3 = 18 for the volume regularizer. We employ a five-step multilevel
strategy, where the transformation is discretized on meshes with 82 , 162 , 322 , 642 , and 1282 cells.
Observations. For the proposed method with a stationary velocity model we require 16, 4, and 4 Gauss–
Newton iterations per level with a total runtime of roughly 6 seconds. Using the non-stationary velocity
model we require 25, 3, and 3 iterations per level and a runtime of about 12 seconds. For the hyperelastic
registration approach 30, 17, 6, 7, and 3 iterations are performed on each resolution level. The total
computational time is about 35 seconds.
We visualize the results in Fig. 4.1. As can be seen in Fig. 4.1, the proposed methods deliver transformed
template images that are qualitatively similar to the reference image (small residual). Both methods result in
diffeomorphic transformations as judged by the values of the determinant of the Jacobian. As to be expected,
the range of the relative volume change is considerably larger for the proposed methods (det ∇y(v, x, 1, 0) ∈
[0.05, 20.58] and det ∇y(v, x, 1, 0) ∈ [0.16, 14.25] for the stationary and non-stationary field, respectively) as
compared to the hyperelastic registration (det ∇y ∈ [0.34, 5.88]). This is due to the fact that the hyperelastic
registration model explicitly controls and penalizes volume change. Comparing the estimated stationary and
non-stationary velocity fields shows that for the latter the estimate changes considerably in time. Also, it
should be noted that the registration quality is slightly better for the non-stationary approach (smaller range
for the Jacobians and a reduction of the distance measure of 99.48% vs. 97.85%, respectively).
We show the results of the experimental evaluation of different preconditioning techniques in Fig. 4.2.
The number of non-zero elements in the Hessian increases from the first to the final iteration for both
regularizers. This is due to the fact that particles travel a longer distance through the domain. The
11
input data
transformed
template
relative volume
change
velocity
H 1 -non-stationary
H 1 -stationary
template image
hyperelastic
reference image
transformation
t=0
t = 1/2
t=1
Fig. 4.1. 2D registration results for an academic benchmark problem also considered in [20]. First column visualizes
test data and the remaining images visualize registration results for hyperelastic registration [15] (first row) and the proposed
method with H 1 -regularization and stationary (second row) and non-stationary (third row) velocity models. It can be seen that
the proposed methods improves the similarity between the reference and the transformed template image without foldings of
the grid. However, the ranges of the relative volume change is considerably larger. It is also evident that the non-stationary
velocity model improves the registration result and comparing the estimated velocity fields (right column) shows substantial
differences of the velocity estimates.
H 1 -regularization; non-stationary velocity
PCG convergence
100
10−5
10−10
transformation
first iteration
relative error
first iteration
sparsity of H(v)
0
100
sparsity of H(v)
PCG convergence
relative error
H 1 -regularization; stationary velocity
transformation
200
100
10−5
10−10
0
10−5
0
100
200
iterations
relative error
100
10−10
100
200
iterations
final iteration
relative error
final iteration
iterations
100
CG
PCG-Jac
PCG-SGS
PCG-Spec
10−5
10−10
0
100
200
iterations
Fig. 4.2. Sparsity pattern and PCG convergence plots at first and final Gauss–Newton iteration for the 2D test problem
on a coarse mesh (m = [32, 32]). We compare the the stationary (left) and non-stationary (right) diffusion regularizer. In
both cases the number of non-zero elements in the Hessian grows between the iterations since particles move farther through
the domain. In all four cases we compare the convergence of different PCG schemes (no preconditioning, Jacobi, Symmetric
Gauss Seidel, and spectral preconditioning). It can be seen that the problems at the final iteration are, in this example, more
difficult to solve, however, the spectral preconditioner outperforms the other choices.
performance of the preconditioner deteriorates in the final iteration for all preconditioners. We observe a
similar behavior for the hyperelastic formulation (see [60]). We can observe that we need fewer iterations for
the stationary case to reach the tolerance; we invert for fewer unknowns, which results in a smaller linear
system that needs to be solved. The proposed spectral preconditioner displays the best rate of convergence
amongst the considered schemes for preconditioning the Hessian; we use it for all our experiments. We note
that we have performed the same study for the curvature regularization (results not included in this study).
We observed a similar behavior.
12
4.3. 2D Mass-Preserving Registration. We consider an academic test problem for mass-preserving
registration. We compare our method against the VAMPIRE toolbox for mass-preserving registration [30].
Experimental Setup. The test data is designed to mimic the contraction of a tissue containing a fixed
amount of tracer. The data is obtained by subtracting two Gaussians with different standard deviations.
The mass is exactly equal, but in the reference image it is concentrated in a smaller region so that the
image overall appears brighter. The image domain is (−5, 5)2 and the full resolution is 256 × 256. For all
experiments we use a four-level multi-level strategy with resolutions 322 , 642 , 1282 , and 2562 , respectively.
A continuous image model is built using bi-linear interpolation and the SSD distance measure is used to
quantify image similarity.
• MP-LDDMM: As in the previous example the velocity field is modeled on a padded spatial domain
(−5.4, 5.4)2 to reduce boundary effects. We use nt = 1 for the spatial discretization of the velocity.
The characteristics are approximated using N = 2 time steps for the RK4 method. The pushforward matrices are build from bilinear basis functions whose width equals the cell size. We use
the diffusion regularizer with weight α = 1000 and γ = 1E−2. We compare results for a stationary
and a non-stationary velocity field.
• VAMPIRE: We use the default parameters for the hyperelastic regularizer (α1 = 10, 000 for the
length-, α2 = 0 for the area, and α3 = 100 for the volume regularizer).
Observations. Registration results are visualized in Fig. 4.3. For the MP-LDDMM using a stationary
velocity model, we perform 4, 2, 1, and 1 iterations per resolution level. The total computation time is
about 4 seconds. Using the non-stationary velocity model we require 5, 2, 2, and 2 iterations and require a
computation time of about 8 seconds. For VAMPIRE we perform 5, 2, 2, and 1 iterations on the respective
levels. The time-to-solution is approximately 12 seconds.
Both methods also yield comparable results in terms of data misfit as well as the final transformation.
This is not only confirmed qualitatively by visual inspection of the transformed template image and the
deformed grids, but also quantitatively: The volume change introduced by the transformation obtained using
the proposed methods (det ∇y(v, x, 1, 0) ∈ [0.89, 2.33] and det ∇y(v, x, 1, 0) ∈ [0.67, 2.54] for a stationary
and a non-stationary velocity model, respectively) is comparable to the one obtained using VAMPIRE
(det ∇y ∈ [0.76, 2.40]). The largest improvement in image similarity (with respect to the SSD) is achieved
for the MP-LDDMM method with a non-stationary velocity (distance reduction of 99.98% vs. 97.43%)
although—in contrast to the previous experiment—it should be noted that the estimated velocities are very
similar for both approaches.
4.4. 3D Cardiac PET. We consider a 3D mass-preserving registration problem of registering systolic
and diastolic cardiac PET data of a mouse heart. The data is provided in FAIR.8 The image domain is
Ω = (0, 32)3 with a resolution of 403 grid points. The results are illustrated in Fig. 4.4. We use a three-level
multi-level strategy with resolutions 103 , 203 , and 403 , respectively, for all approaches. On the finest level,
the number of unknowns is 384 000.
Experimental Setup.
• MP-LDDMM: The velocity field is modeled on the same domain as the image data and the same
number of cells is used for spatial discretization. We will only consider the non-stationary case,
here. We use nt = 1 time intervals for the velocity (which results in two discretization points for
the velocity v). We use an RK4 method to compute the characteristics with N = 2 time steps. The
push-forward matrix is build using tri-linear hat functions, with a width that corresponds to the
voxel size of the image data. We use the diffusion regularizer with regularization weight α = 100
and γ = 1E−2.
• VAMPIRE: We use α1 = 100 for the length regularizer, α2 = 10 for the area regularizer, and
α3 = 100 for the volume regularizer. We use the same number of multi-resolution levels.
Observations. For MP-LDDMM the optimization scheme performs 9, 3, and 3 iterations on the respective
levels. The time-to-solution is about 36 seconds. For VAMPIRE we require 5, 4, and 3 iterations on the
respective levels. The total runtime is about roughly 74 seconds. Both schemes yield qualitatively almost
identical results. The residual differences between the transformed template image and the reference image
8 We thank the European Institute for Molecular Imaging (EIMI) and SFB 656, University of Münster, Germany for
contributing the image data.
13
input data
VAMPIRE
H 1 -stationary
transformed
template
relative volume
change
velocity
v stationary
H 1 -non-stationary
reference image
template image
transformation
v at t = 0
v at t = 1
VAMPIRE
input data
Fig. 4.3. 2D mass-preserving registration for an academic test problem. The image data is generated by subtracting two
Gaussian kernels with different standard deviations; the data is designed to have equal mass. The reference image (top) and
the template image (bottom) are shown in the left column. We compare the VAMPIRE method (first row) [30] to the proposed
mass-preserving LDDMM with stationary (middle row) and non-stationary velocity model (bottom row). For all three methods,
we visualize the transformation, the transformed template, and the relative volume change. For the LDDMM methods we also
visualize the velocity. Comparing the results in the middle and bottom row, it can be seen that the underlying transformation
is rather simple; it can be well represented using a stationary velocity field.
reference image
template image
deformed template image
relative volume change
1.40
proposed
0.54
1.76
0.24
Fig. 4.4. 3D mass-preserving registration of diastolic and systolic PET images of a mouse heart. We display the input
data in the first row (left: reference image; right: template image; from left to right: axial, coronal and sagittal view). The
deformed template images are shown in the left column (middle row: VAMPIRE; bottom row: proposed method). The relative
volume change is shown in the right column (middle row: VAMPIRE; bottom row: proposed method). The color bars to the
right illustrate the color coding and provide the range of the Jacobian fields.
is small. Overall, our current prototype implementation of a Gauss–Newton–Krylov method for LDDMM is
competitive with VAMPIRE in terms of the runtime. We expect to be able to drastically reduce the runtime
in near future. For the hyperelastic registration most time is spent on determining the search direction,
which requires solving an ill-conditioned linear system; see also [60]; a reduction in runtime for this scheme
is much more difficult.
14
4.5. 3D Brain Registration.
Experimental Setup. The data is taken from the NIREP repository [19]. We consider the datasets
na02 (template image) and na01 (reference image) for our experiments. The grid size for these images is
256 × 300 × 256. We downsample these images to a size of 128 × 150 × 128 voxels to make the problem
computationally tractable for our prototype and the reference implementation. The image domain is defined
to be Ω = (0, 20) × (0, 23.4375) × (0, 20). We use SSD as distance measure and a multi-level strategy with
3 resolution levels (32 × 38 × 32, 64 × 75 × 64, and 128 × 150 × 128). The number of unknowns is 7 372 800
for the finest level.
We evaluate registration performance based on overlap measures evaluated for the label maps associated
with the images. The data comes with 32 labels for gray matter regions [19]. We simplify the presentation
of our results by only considering the union of these labels to evaluate the performance of our method.
We use the Dice coefficient as a measure for registration quality, which has an optimal value of 1. We
use a nearest-neighbor interpolation model to transform the label maps with the computed y to avoid any
additional thresholding.
We limit the evaluation of the determinant of the Jacobian to the foreground (i.e., the brain) in the
reference image. We identify this foreground by thresholding; we consider intensities with a value of 0.05
and larger as foreground. We slightly extend this mask by smoothing it with a Gaussian kernel of width 2h.
A second thresholding step defines the final brain mask used for the evaluation of the Jacobians.
• Proposed (LDDMM): The velocity field is modeled on a slightly larger domain than the image
domain to reduce boundary effects; we choose Ωv = (−1, 21) × (−1, 24.4375) × (−1, 21). We consider
stationary and non-stationary velocities v. We use nt = 1 time intervals for the non-stationary
case (which results in two discretization points for the velocity v). We use an RK4 method with
N = 5 time steps to compute the characteristics . The push-forward matrix is build using tri-linear
hat functions, with a width that corresponds to the voxel size of the image data. We consider the
curvature (H 2 ) and the diffusive (H 1 ) regularization model. We study registration performance
(data mismatch and extremal values of the Jacobians det ∇y) as a function of the regularization
weight α. Once we have found the velocity v, we compute the transformation y we use to evaluate
the performance of our method using N = 20 time steps. We experimentally found that a shift
of γ = 0 and γ = 1E−2 yields the optimal rate of convergence for the diffusive and the curvature
regularization model, respectively. We set the tolerance for the optimization to tolJ = 5E−2. We
use a relative tolerance of 1E−1 for the PCG method; we limit the number of Krylov iterations to
50.
• Hyperelastic registration: We experimentally found that regularization weight of α1 = 100 (length
regularizer), α2 = 10 (surface regularizer), and α3 = 100 (volume regularizer) yields high data fidelity
(good mismatch) and well behaved Jacobians. We use this setting throughout our experiments. We
set the tolerance for the optimization to tolJ = 1E − 3.
Observations. We show exemplary results for the registration in Fig. 4.5. We report results for the
quantitative evaluation in Table 4.1. An illustration of the velocities can be found in Fig. 4.6.
All methods yield high fidelity results with diffeomorphic transformations and well behaved Jacobians.
We achieve the best Dice score for a diffusive regularization model for α = 300 (run #13 in Table 4.1) and a
non-stationary velocity field. This is the only run, for which we outperform the hyperelastic approach. The
results for the curvature regularization model do not vary significantly when switching from a stationary to a
non-stationary formulation; we obtain similar extremal values for the Jacobians and similar Dice values. This
is different for the diffusive regularization model. We obtain slightly better values for the Dice coefficient
with similar extremal values for the Jacobian.
If we consider a stationary velocity field we can reduce the time-to-solution by a factor of two compared
to the hyperelastic approach. These findings are consistent for both regularization approaches (runs #2 to
#8 in Table 4.1). If we turn to non-stationary velocity fields, our current implementation of the curvature
regularization model is no longer competitive in terms of time-to-solution. For a diffusive regularization
model we are, however, still slightly faster than the hyperelastic approach (runs #13 to #15 in Table 4.1)
despite an increase of the number of unknowns by a factor of 2 (we use nt = 1, which results in two
discretization points for the velocity).
We need, for instance, 5, 3 , and 3 iterations for the individual levels for the stationary case and a
15
residual before
6.69
0.08
7.65
residual after
0.08
residual after
1.86
residual after
relative
volume change template image
relative
volume change
relative
volume change
original data
hyperelastic
curvature
diffusion
transformed
transformed
transformed
template image template image template image reference image
0.45
Fig. 4.5. Exemplary results for a 3D intensity-preserving registration problem based on MRI datasets of the human brain.
The data is taken from the NIREP repository. We show (from left to right) an axial, coronal, and sagittal view of the reference
image (dataset na01), the template image (dataset na02), and the residual differences between these two images in the top
row. The results correspond to those reported in Table 4.1. We report results for a map based approach with a hyperelastic
regularization model (second row: run #1 in Table 4.1; α1 = 100 (length regularizer), α2 = 10 (surface regularizer), and
α3 = 100 (volume regularizer)) [15], and for the proposed method for a non-stationary velocity field (third row: curvature
regularization model; α = 10; run #10 in Table 4.1; bottom row: diffusive regularization model; α = 300; run #13 in
Table 4.1). For each of these methods we show (from left to right) an axial, a coronal, and a sagittal view of the deformed
template image, a map for the relative volume change, and the residual differences between the transformed template image
and the reference image after registration. We also display the color bar and the maximal and minimal values for the maps
for the relative volume change.
non-stationary
stationary
non-stationary
diffusive
curvature
stationary
4.79E2
2.41E3
2.02E3
5.87E2
5.62E2
5.63E2
Fig. 4.6. Illustration of the obtained velocity fields for the registration of 3D brain imaging data. We show the velocities
for the curvature (left; α = 10; runs #2 and #10 in Table 4.1) and the diffusive regularization model (right; α = 300; runs
#6 and #13 in Table 4.1). We report the `2 -norm of the velocity field below each individual figure.
diffusive regularization model (α = 400; run #6 in Table 4.1). For each iteration we require 22, 23, 24, and
24 PCG iterations (level 1), 22, 25, and 27 PCG iterations (level 2), and 27, 29, and 29 PCG iterations (level
3), respectively. The stationary case and a curvature regularization model (α = 50; run #4 in Table 4.1)
requires 5, 3, and 2 iterations, with 7, 11, 17, 16, and 16 PCG iterations (level 1), 20, 36, 50 PCG iterations
(level 2), and 50, and 50 PCG iterations (level 3), respectively. The hyperelastic regularization approach
converges after 7, 6, and 5 iterations per level.
The results in Fig. 4.5 suggest that all methods yield comparable residual differences after registration.
However, we can, likewise to the former experiments, observe drastic differences in the Jacobians. The
hyperelastic regularization allows us to better control the Jacobians (the values range from 4.51E−1 to
1.86). If this control is indeed of importance in practical applications remains to be seen. Notice, that we
can either add hard constraints on the divergence of the velocity to our formulation [48, 50] or constraints
on det ∇y to enable such control.
The projections of the velocity fields in Fig. 4.6 show significant differences between the stationary and
the non-stationary case for the curvature regularization model. We can also observe large differences in
the appearance of the velocity fields in time for the non-stationary case. This is different for the diffusive
16
Table 4.1
Registration quality. We report registration results as a function of the regularization weight α for the first two datasets of
the NIREP repository. We compare the proposed method considering stationary and non-stationary velocity fields. We report
results for the curvature (H 2 ) regularization model (theoretically required to guarantee the existence of a diffeomorphic deformation map) and a diffusive (H 1 ) regularization model. We compare the proposed method to a formulation for diffeomorphic
image registration based on a hyperelastic regularization method [15]. We use the experimentally determined regularization
weights α1 = 100 (length regularizer), α2 = 10 (surface regularizer), and α3 = 100 (volume regularizer). We report values
(from left to right) for the Dice coefficient after registration and the extremal values for the Jacobian (min and max). The
initial value for the Dice coefficient for the considered datasets is 5.54E−1.
method
run
α
dice
min(det ∇y)
max(det ∇y)
time (speedup)
hyperelastic
#1
100, 10, 100
7.93E−1
4.51E−1
1.86
3.71E+3
(1.00)
curvature
#2
#3
#4
#5
#6
#7
#8
10
25
50
100
300
400
500
7.76E−1
7.55E−1
7.34E−1
7.14E−1
7.86E−1
7.75E−1
7.66E−1
8.79
5.77
4.74
3.13
6.38
6.26
6.17
1.82E+3
1.86E+3
1.45E+3
1.47E+3
1.96E+3
1.94E+3
1.95E+3
(2.04)
(2.00)
(2.56)
(2.51)
(1.89)
(1.91)
(1.90)
#9
#10
#11
#12
#13
#14
#15
5
10
25
50
300
400
500
7.60E−1
7.60E−1
7.47E−1
7.35E−1
8.05E−1
7.93E−1
7.67E−1
6.36
5.73
5.68
3.68
6.48
5.40
6.50
7.64E+3
7.50E+3
6.18E+3
5.98E+3
3.29E+3
3.34E+3
2.59E+3
(0.49)
(0.49)
(0.60)
(0.62)
(1.13)
(1.11)
(1.43)
stationary LDDMM
diffusion
8.88E−2
1.19E−1
1.79E−1
2.37E−1
7.79E−2
1.02E−1
1.18E−1
non-stationary LDDMM
curvature
diffusion
9.55E−2
8.48E−2
1.31E−1
2.12E−1
7.71E−2
1.08E−1
1.53E−1
regularization model. The stationary and non-stationary velocities do not differ significantly. The differences
in time for the non-stationary case are also less pronounced. We can also observe that the energy for both
components of the non-stationary velocity field is quite similar (as judged by the values for the `2 -norm
reported in Fig. 4.6). This is true for both regularization models. As to be expected we obtain much
smoother velocities for the curvature regularization model.
5. Summary and Conclusion. In this paper, we propose efficient numerical algorithms based on
Lagrangian hyperbolic PDE solvers to efficiently solve the reduced formulation of the PDE-constrained
optimization problem arising in LDDMM. Our formulation can be used for classical, intensity-preserving,
registration but also extends the LDDMM framework to mass-preserving registration problems. We consider
an optimal control formulation and propose an efficient discretize-then-optimize approach amendable for
standard Gauss–Newton methods. The key idea of our approach is to eliminate the hyperbolic PDE constraint using a Lagrangian method with an explicit time integration of the characteristics. Our formulation
can handle both stationary and non-stationary velocity fields efficiently. We present economical schemes for
analytically computing its derivatives. A main advantage of our method over existing solvers is that derivatives of the solution to the hyperbolic PDE with respect to the velocity field can be explicitly constructed.
This leads to an overall memory requirement that is independent of the number of time steps used for solving
the PDE. This is a significant advantage over most existing work, which in general require the storage (or
re-computing) of spatio-temporal state and adjoint fields or the transformation.
We studied registration performance considering different synthetic and real-world problems. We made
the following observations:
• Our results are competitive in terms of both time-to-solution and inversion quality (mismatch)
compared to state-of-the-art packages for diffeomorphic image registration across a wide range of
applications, which includes mass-preserving and intensity-preserving registration problems.
• Our spectral preconditioner yields a good performance. However, the rate of convergence deteriorates when switching from stationary to non-stationary velocities. Designing a more effective
preconditioner for these cases is an item of future work.
• We could observe differences between the stationary and non-stationary formulation in terms of the
reconstruction accuracy. This especially becomes apparent for the classical problem of registering a
17
C-shaped object to a disc [20]. In this example a considerable improvement can be achieved using
a small number of time discretization points for the velocity. In general, increasing the number of
time points enriches the space of transformations, however, it also increases the complexity of the
optimization problem.
• Since y appears explicitly in our formulation, we can control det ∇y by adjusting the regularization
weight α (additional comments can be found below). We have considered H 1 - and H 2 -regularization
norms. While theoretical considerations do require (more than) H 2 -regularity on v to guarantee that
a diffeomorphic map y exists [9], we could demonstrate our numerical scheme allows us to ensure
that the final map y is diffeomorphic at a discrete level, even for H 1 -regularity. However, we note
that we consider the regularization as a modular building block. If theoretical requirements are of
concern, one can switch to H 2 -regularity.
In theory, solutions to the variational optimal control problem are guaranteed to be diffeomorphic (under
the assumption of sufficient regularity of v). However, as compared to other diffeomorphic registration
approaches that control and thus guarantee invertibility of the discrete transformation such as [32, 15],
it is more difficult to ensure this for discrete solutions to the optimal control problems. An inaccurate
approximation of the characteristics may cause characteristics to cross and thus lead to non-diffeomorphic
transformations. This problem is also inherent in other numerical implementations of LDDMM. Therefore,
we recommend monitoring volume changes induced by the transformation to adapt the number of time steps
and/or smoothness parameter. In our method, the end points of the characteristics correspond to a deformed
grid, which can be analyzed or even regularized using techniques described in [32, 15]. Monitoring volume
changes via Jacobian determinants can also be done in Eulerian or SL methods, in which the transformation
is generally not computed [48, 50].
In this paper, we optimize over the velocity field v instead of optimizing over the final transformation y,
a strategy that has become predominantly used in many practical applications. Optimizing for the velocity
allows us to use a fairly simple quadratic regularization model while still (in theory) ensuring invertibility of
the resulting transformation. In the discrete setting, the grid might have foldings, depending on the regularization parameter and/or the accuracy of the time integration. This can be seen as a drawback compared
to image registration methods that use invertibility constraints or nonlinear regularizers directly acting on
the transformation. However, these regularizers are very challenging both in theory and in practice; see,
e.g., [60]. Another feature of more complicated regularization models such as, e.g., the elastic regularization proposed in [28, 14, 22, 69, 15] is that they are motivated based on physical principles. The notion of
plausibility of a transformation y is for these type of regularization models not only limited to the prerequisite that y is a diffeomorphism; it is based on bio-mechanical considerations. Thus, while achieving a very
good similarity of the final images, the obtained transformation might not be plausible in all applications.
However, a similar argument can be made for elasticity-based regularizers unless true material properties
are known and incorporated into the regularization. Another approach to integrate bio-physical priors into
diffeomorphic registration is to include more complicated state equations that model the bio-physics of a
system under investigation; an example in the context of large deformation diffeomorphic image registration
is the incorporation of incompressibility constraints [48, 50]. This, likewise to more sophisticated regularization norms, introduces additional parameters, and as such makes an automated calibration of the method
to unseen data more difficult.
Some limitations of the current method will be addressed in future work. First, for mass-preserving
registration, the width of the particle kernels may be adjusted locally depending on the spacing of particles
after transformation [18]. Computing the distance to the closest neighbor is expensive, however, in our
framework the Jacobian determinant is available and can be used to detect relative changes in the density of
particles. Second, we will investigate locally adaptive time stepping schemes for computing the characteristics
that account for the complexity of the velocity fields.
6. Acknowledgements. AM is supported by the U.S. Department of Energy, Office of Science, Office
of Advanced Scientific Computing Research, Applied Mathematics program under Award Numbers desc0010518 and de-sc0009286; and by NIH grant 10042242. LR is supported in part by National Science
Foundation (NSF) award DMS 1522599.
18
REFERENCES
[1] V. Akcelik, G. Biros, and O. Ghattas. Parallel multiscale Gauss-Newton-Krylov methods for inverse wave propagation.
In Proc ACM/IEEE Conference on Supercomputing, pages 1–15, 2002. 3
[2] A. Alexanderian, N. Petra, G. Stadler, and O. Ghattas. A fast and scalable method for A-optimal design of experiments for
infinite-dimensional Bayesian nonlinear inverse problems. SIAM Journal on Scientific Computing, 38(1):A243–A272,
2016. 10
[3] V. Arsigny, O. Commowick, X. Pennec, and N. Ayache. A log-Euclidean framework for statistics on diffeomorphisms.
9(Pt 1):924–931, 2006. 1, 5
[4] J. Ashburner. A fast diffeomorphic image registration algorithm. NeuroImage, 38(1):95–113, Oct. 2007. 2, 5
[5] J. Ashburner and K. J. Friston. Diffeomorphic registration using geodesic shooting and Gauss–Newton optimisation.
NeuroImage, 55(3):954–967, Apr. 2011. 1, 2, 3, 5
[6] B. Avants, P. T. Schoenemann, and J. C. Gee. Lagrangian frame diffeomorphic image registration: Morphometric
comparison of human and chimpanzee cortex. Medical Image Analysis, 10:397–412, 2006. 3
[7] B. B. Avants, C. L. Epstein, M. Brossman, and J. C. Gee. Symmetric diffeomorphic image registration with crosscorrelation: Evaluating automated labeling of elderly and neurodegenerative brain. Medical Image Analysis, 12(1):26–
41, 2008. 3
[8] B. B. Avants, N. J. Tustison, G. Song, P. A. Cook, A. Klein, and J. C. Gee. A reproducible evaluation of ANTs similarity
metric performance in brain image registration. NeuroImage, 54:2033–2044, 2011. 3
[9] M. F. Beg, M. I. Miller, A. Trouvé, and L. Younes. Computing Large Deformation Metric Mappings via Geodesic Flows
of Diffeomorphisms. International journal of computer vision, 61(2):139–157, 2005. 1, 2, 3, 5, 9, 10, 11, 18
[10] M. Benzi, E. Haber, and L. Taralli. A preconditioning technique for a class of PDE-constrained optimization problems.
Advances in Computational Mathematics, 35(2-4):149–173, July 2011. 1, 2, 3
[11] A. Borzi, K. Ito, and K. Kunisch. Optimal control formulation for determining optical flow. SIAM Journal on Scientific
Computing, 24(3):818–847 (electronic), 2002. 1, 2, 3, 5
[12] A. Borzı̀, K. Ito, and K. Kunish. An optimal control approach to optical flow computation. International Journal for
numerical methods in fluids, 40(1–2):231–240, 2002. 1, 2, 3
[13] A. Borzı̀ and V. Schulz. Computational optimization of systems governed by partial differential equations. SIAM, Philadelphia, Pennsylvania, US, 2012. 2
[14] C. Broit. Optimal registration of deformed images. PhD thesis, University of Pennsylvania, 1981. 18
[15] M. Burger, J. Modersitzki, and L. Ruthotto. A hyperelastic regularization energy for image registration. SIAM Journal
on Scientific Computing, 35(1):B132–B148, 2013. 2, 10, 11, 12, 16, 17, 18
[16] H. Chang and J. M. Fitzpatrick. A technique for accurate magnetic-resonance-imaging in the presence of field inhomogeneities. Medical Imaging, IEEE Transactions on, 11(3):319–329, Sept. 1992. 2, 4
[17] K. Chen and D. A. Lorenz. Image Sequence Interpolation Using Optimal Control. Journal of Mathematical Imaging and
Vision, 41(3):222–238, Mar. 2011. 1, 2, 3, 5
[18] A. Chertock and A. Kurganov. On a practical implementation of particle methods. Applied Numerical Mathematics,
56(10-11):1418–1431, 2006. 7, 18
[19] G. E. Christensen, X. Geng, J. G. Kuhl, J. Bruss, T. J. Grabowski, I. A. Pirwani, M. W. Vannier, J. S. Allen, and
H. Damasio. Introduction to the non-rigid image registration evaluation project. In Proc Biomedical Image Registration, volume LNCS 4057, pages 128–135, 2006. 15
[20] G. E. Christensen, R. D. Rabbitt, and M. I. Miller. Deformable templates using large deformation kinematics. Image
Processing, IEEE Transactions on, 5(10):1435–1447, 1996. 2, 11, 12, 18
[21] M. Dawood, C. Brune, X. Jiang, F. Büther, M. Burger, O. Schober, M. Schäfers, and K. P. Schäfers. A Continuity
Equation Based Optical Flow Method for Cardiac Motion Correction in 3D PET Data. In Medical Imaging and
Augmented Reality, pages 88–97. Springer Berlin Heidelberg, Berlin, Heidelberg, Sept. 2010. 2, 4
[22] M. Droske and M. Rumpf. A Variational Approach to Nonrigid Morphological Image Registration. SIAM Journal on
Applied Mathematics, 64(2):668–687, Jan. 2004. 18
[23] P. Dupuis, U. Grenander, and M. I. Miller. Variational problems on flows of diffeomorphisms for image matching. Quarterly
of applied mathematics, 1998. 1, 2, 3, 5
[24] L. C. Evans. Partial Differential Equations. American Mathematical Soc., 2010. 4, 6
[25] R. E. Ewing and H. Wang. A summary of numerical methods for time-dependent advection-dominated partial differential
equations. Journal of Computational and Applied Mathematics, 128(1-2):423–445, 2001. 1
[26] B. Fischer and J. Modersitzki. Curvature Based Image Registration. Journal of Mathematical Imaging and Vision,
18(1):81–85, 2003. 5
[27] B. Fischer and J. Modersitzki. Ill-posed medicine—an introduction to image registration. Inverse Problems, 24(3):034008,
2008. 5
[28] M. A. Fischler and R. A. Elschlager. The representation and matching of pictorial structures. Computers, IEEE Transactions on, 1973. 18
[29] J. Fohring, E. Haber, and L. Ruthotto. Geophysical Imaging of Fluid Flow in Porous Media. SIAM Journal on Scientific
Computing, 36(5):S218–S236, 2014. 7, 8
[30] F. Gigengack, L. Ruthotto, M. Burger, C. H. Wolters, X. Jiang, and K. P. Schafers. Motion Correction in Dual Gated
Cardiac PET Using Mass-Preserving Image Registration. Medical Imaging, IEEE Transactions on, 31(3):698–712,
Mar. 2012. 2, 4, 10, 13, 14
[31] M. D. Gunzburger. Perspectives in flow control and optimization. SIAM, Philadelphia, Pennsylvania, US, 2003. 2, 3
[32] E. Haber and J. Modersitzki. Image Registration with Guaranteed Displacement Regularity. International journal of
19
computer vision, 71(3):361–372, July 2006. 18
[33] E. Haber and J. Modersitzki. A multilevel method for image registration. SIAM Journal on Scientific Computing,
27(5):1594–1607, 2006. 4
[34] E. Haber and D. Oldenburg. A GCV based method for nonlinear ill-posed problems. Computational Geosciences, 4:41–63,
2000. 4
[35] S. Haker, L. Zhu, A. Tannenbaum, and A. Angenent. Optimal mass transport for registration and warping. International
Journal of Computer Vision, 60(3):225–240, 2004. 3
[36] P. C. Hansen. Rank-deficient and discrete ill-posed problems. SIAM Monographs on Mathematical Modeling and Computation. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1998. 4
[37] P. C. Hansen, J. G. Nagy, and D. P. O’Leary. Deblurring Images: Matrices, Spectra and Filtering. Matrices, Spectra,
and Filtering. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2006. 10
[38] G. L. Hart, C. Zach, and M. Niethammer. An optimal control approach for deformable registration. In Proc IEEE
Conference on Computer Vision and Pattern Recognition, pages 9–16, 2009. 2, 3
[39] M. Hernandez. Gauss–Newton inspired preconditioned optimization in large deformation diffeomorphic metric mapping.
Physics in Medicine and Biology, 59(20):6085–6115, 2014. 1
[40] M. Hernandez, M. N. Bossa, and S. Olmos. Registration of anatomical images using paths of diffeomorphisms parameterized with stationary vector field flows. International Journal of Computer Vision, 85(3):291–306, 2009. 3, 5
[41] R. Herzog and K. Kunisch. Algorithms for PDE-constrained optimization. GAMM Mitteilungen, 33(2):163–176, 2010. 2
[42] M. R. Hestenes and E. Stiefel. Methods of Conjugate Gradients for Solving Linear Systems. Journal of Research of the
National Bureau of Standards, 49(6):409–436, 1952. 10
[43] M. Hinze, R. Pinnau, M. Ulbrich, and S. Ulbrich. Optimization with PDE constraints. Springer, Berlin, DE, 2009. 2
[44] E. Lee and M. Gunzburger. An optimal control formulation of an image registration problem. Journal of Mathematical
Imaging and Vision, 36(1):69–80, 2010. 3
[45] R. J. LeVeque. Numerical methods for conservation laws, 1992. 1
[46] R. J. LeVeque. Finite Volume Methods for Hyperbolic Problems. Cambridge University Press, 2002. 1, 4, 6
[47] M. Lorenzi and X. Pennec. Geodesics, parallel transport and one-parameter subgroups for diffeomorphic image registration.
International Journal of Computer Vision, 105(2):111–127, 2013. 3
[48] A. Mang and G. Biros. An inexact Newton–Krylov algorithm for constrained diffeomorphic image registration. SIAM
Journal on Imaging Sciences, 8(2):1030–1069, 2015. 1, 2, 3, 4, 5, 9, 10, 16, 18
[49] A. Mang and G. Biros. A Semi-Lagrangian two-level preconditioned Newton-Krylov solver for constrained diffeomorphic
image registration. Apr. 2016. 1, 2, 3
[50] A. Mang and G. Biros. Constrained H 1 -regularization schemes for diffeomorphic image registration. SIAM Journal on
Imaging Sciences, 9(3):1154–1194, 2016. 1, 2, 3, 5, 9, 10, 16, 18
[51] A. Mang, A. Gholami, and G. Biros. Distributed-memory large-deformation diffeomorphic 3D image registration. In Proc
ACM/IEEE Conference on Supercomputing, number 72, 2016. 2, 3, 5, 9, 10
[52] M. I. Miller and L. Younes. Group actions, homeomorphism, and matching: A general framework. International Journal
of Computer Vision, 41(1/2):61–81, 2001. 1
[53] J. Modersitzki. Numerical methods for image registration. Oxford University Press on Demand, 2004. 1, 2, 3, 5
[54] J. Modersitzki. FAIR: Flexible Algorithms for Image Registration, volume 6 of Fundamentals of Algorithms. Society for
Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2009. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
[55] J. Nocedal and S. J. Wright. Numerical Optimization. Springer, New York, New York, US, 2006. 9
[56] Y. Ou, H. Akbari, M. Bilello, X. Da, and C. Davatzikos. Comparative evaluation of registration algorithms in different
brain databases with varying difficulty: Results and insights. IEEE T Med Imaging, 33(10):2039–2065, 2014. 4
[57] A. Pai, S. Sommer, L. Sorensen, S. Darkner, J. Sporring, and M. Nielsen. Kernel bundle diffeomorphic image registration
using stationary velocity fields and wendland basis functions. Medical Imaging, IEEE Transactions on, 35(6):1369–
1380, 2016. 5
[58] T. Polzin, M. Niethammer, M. P. Heinrich, H. Handels, and J. Modersitzki. Memory efficient LDDMM for lung CT. In
Proc Medical Image Computing and Computer-Assisted Intervention, volume LNCS 9902, pages 28–36, 2016. 1, 3
[59] T. u. Rehman, E. Haber, G. Pryor, J. Melonakos, and A. Tannenbaum. 3D nonrigid registration via optimal mass
transport on the GPU. Medical Image Analysis, 13(6):931–940, Dec. 2009. 2
[60] L. Ruthotto, C. Greif, and J. Modersitzki. A Stabilized Multigrid Solver for Hyperelastic Image Registration. in revision
at Numerical Linear Algebra With Application, pages 1–16, July 2016. 2, 10, 11, 12, 14, 18
[61] L. Ruthotto, H. Kugel, J. Olesch, B. Fischer, J. Modersitzki, M. Burger, and C. H. Wolters. Diffeomorphic susceptibility
artifact correction of diffusion-weighted magnetic resonance images. Physics in Medicine and Biology, 57(18):5715–
5731, Sept. 2012. 2, 4
[62] Y. Saad. Iterative Methods for Sparse Linear Systems. Second Edition. SIAM, Philadelphia, Apr. 2003. 10
[63] A. Sotiras, C. Davatzikos, and N. Paragios. Deformable medical image registration: A survey. Medical Imaging, IEEE
Transactions on, 32(7):1153–1190, 2013. 2, 3, 5
[64] A. Staniforth and J. Côté. Semi-Lagrangian integration schemes for atmospheric models—A review. Montly Weather
Review, 119(9):2206–2223, 1991. 3
[65] A. Trouvé. Diffeomorphisms Groups and Pattern Matching in Image Analysis. International journal of computer vision,
28(3):213–221, 1998. 1, 2, 3, 5
[66] T. Vercauteren, X. Pennec, A. Perchant, and N. Ayache. Diffeomorphic demons: Efficient non-parametric image registration. NeuroImage, 45(1):S61–S72, 2009. 2, 3
[67] F.-X. Vialard, L. Risser, D. Rueckert, and C. J. Cotter. Diffeomorphic 3D Image Registration via Geodesic Shooting
Using an Efficient Adjoint Calculation. International Journal of Computer Vision, 97(2):229–241, Aug. 2011. 1, 3, 5
20
[68] C. R. Vogel. Computational Methods for Inverse Problems. SIAM, Philadelphia, 2002. 4
[69] I. Yanovsky, C. Le Guyader, A. Leow, A. Toga, P. Thompson, and L. Vese. Unbiased volumetric registration via nonlinear
elastic regularization. In 2nd MICCAI Workshop on Mathematical Foundations of Computational Anatomy, 2008.
18
[70] L. Younes. Jacobi fields in groups of diffeomorphisms and applications. Quarterly of Applied Mathematics, 650(1):113–134,
2007. 1
[71] L. Younes, F. Arrate, and M. I. Miller. Evolutions equations in computational anatomy. NeuroImage, 45:S40–S50, 2009.
1
21
| 5cs.CE
|
SAMPLE-LEVEL DEEP CONVOLUTIONAL NEURAL NETWORKS FOR
MUSIC AUTO-TAGGING USING RAW WAVEFORMS
Jongpil Lee
Jiyoung Park
Keunhyoung Luke Kim
Juhan Nam
Korea Advanced Institute of Science and Technology (KAIST)
[richter, jypark527, dilu, juhannam]@kaist.ac.kr
arXiv:1703.01789v2 [cs.SD] 22 May 2017
ABSTRACT
Recently, the end-to-end approach that learns hierarchical representations from raw data using deep convolutional
neural networks has been successfully explored in the image, text and speech domains. This approach was applied
to musical signals as well but has been not fully explored
yet. To this end, we propose sample-level deep convolutional neural networks which learn representations from
very small grains of waveforms (e.g. 2 or 3 samples) beyond typical frame-level input representations. Our experiments show how deep architectures with sample-level filters improve the accuracy in music auto-tagging and they
provide results comparable to previous state-of-the-art performances for the Magnatagatune dataset and Million Song
Dataset. In addition, we visualize filters learned in a samplelevel DCNN in each layer to identify hierarchically learned
features and show that they are sensitive to log-scaled frequency along layer, such as mel-frequency spectrogram
that is widely used in music classification systems.
1. INTRODUCTION
In music information retrieval (MIR) tasks, raw waveforms
of music signals are generally converted to a time-frequency
representation and used as input to the system. The majority of MIR systems use a log-scaled representation in
frequency such as mel-spectrograms and constant-Q transforms and then compress the amplitude with a log scale.
The time-frequency representations are often transformed
further into more compact forms of audio features depending on the task. All of these processes are designed based
on acoustic knowledge or engineering efforts.
Recent advances in deep learning, especially the development of deep convolutional neural networks (DCNN),
made it possible to learn the entire hierarchical representations from the raw input data, thereby minimizing the
input data processing by hands. This end-to-end hierarchical learning was attempted early in the image domain,
particularly since the DCNN achieves break-through results in image classification [1]. These days, the method of
stacking small filters (e.g. 3x3) is widely used after it has
been found to be effective in learning more complex hierarchical filters while conserving receptive fields [2]. In the
Copyright: c 2017 Jongpil Lee et al. This is an open-access article distributed
under the terms of the Creative Commons Attribution 3.0 Unported License, which
permits unrestricted use, distribution, and reproduction in any medium, provided
the original author and source are credited.
text domain, the language model typically consists of two
steps: word embedding and word-level learning. While
word embedding plays a very important role in language
processing [3], it has limitations in that it is learned independently from the system. Recent work using CNNs that
take character-level text as input showed that the end-toend learning approach can yield comparable results to the
word-level learning system [4, 5]. In the audio domain,
learning from raw audio has been explored mainly in the
automatic speech recognition task [6–10]. They reported
that the performance can be similar to or even superior to
that of the models using spectral-based features as input.
This end-to-end learning approach has been applied to
music classification tasks as well [11, 12]. In particular,
Dieleman and Schrauwen used raw waveforms as input of
CNN models for music auto-tagging task and attempted to
achieve comparable results to those using mel-spectrograms
as input [11]. Unfortunately, they failed to do so and attributed the result to three reasons. First, their CNN models were not sufficiently expressive (e.g. a small number of
layers and filters) to learn the complex structure of polyphonic music. Second, they could not find an appropriate
non-linearity function that can replace the log-based amplitude compression in the spectrogram. Third, the bottom
layer in the networks takes raw waveforms in frame-level
which are typically several hundred samples long. The
filters in the bottom layer should learn all possible phase
variations of periodic waveforms which are likely to be
prevalent in musical signals. The phase variations within a
frame (i.e. time shift of periodic waveforms) are actually
removed in the spectrogram.
In this paper, we address these issues with sample-level
DCNN. What we mean by “sample-level” is that the filter size in the bottom layer may go down to several samples long. We assume that this small granularity is analogous to pixel-level in image or character-level in text.
We show the effectiveness of the sample-level DCNN in
music auto-tagging task by decreasing strides of the first
convolutional layer from frame-level to sample-level and
accordingly increasing the depth of layers. Our experiments show that the depth of architecture with samplelevel filters is proportional to the accuracy and also the architecture achieves results comparable to previous state-ofthe-art performances for the MagnaTagATune dataset and
the Million Song Dataset. In addition, we visualize filters
learned in the sample-level DCNN.
Frame-level mel-spectrogram model
Frame-level raw waveform model
Mel-spectrogram extraction
Frame-level strided convolution layer
Sample-level raw waveform model
Sample-level strided convolution layer
Figure 1. Simplified model comparison of frame-level approach using mel-spectrogram (left), frame-level approach using
raw waveforms (middle) and sample-level approach using raw waveforms (right).
2. RELATED WORK
Since audio waveforms are one-dimensional data, previous
work that takes waveforms as input used a CNN that consists of one-dimensional convolution and pooling stages.
While the convolution operation and filter length in upper
layers are usually similar to those used in the image domain, the bottom layer that takes waveform directly conducted a special operation called strided convolution, which
takes a large filter length and strides it as much as the filter
length (or the half). This frame-level approach is comparable to hopping windows with 100% or 50% hop size in a
short-time Fourier transform. In many previous works, the
stride and filter length of the first convolution layer was set
to 10-20 ms (160-320 samples at 16 kHz audio) [8,10–12].
In this paper, we reduce the filter length and stride of
the first convolution layer to the sample-level, which can
be as small as 2 samples. Accordingly, we increase the
depth of layers in the CNN model. There are some works
that use 0.6 ms (10 samples at 16 kHz audio) as a stride
length [6, 7], but they used a CNN model only with three
convolution layers, which is not sufficient to learn the complex structure of musical signals.
3. LEARNING MODELS
Figure 1 illustrates three CNN models in the music autotagging task we compare in our experiments. In this section, we describe the three models in detail.
3.1 Frame-level mel-spectrogram model
This is the most common CNN model used in music autotagging. Since the time-frequency representation is two
dimensional data, previous work regarded it as either twodimensional images or one-dimensional sequence of vectors [11, 13–15]. We only used one-dimensional(1D) CNN
model for experimental comparisons in our work because
the performance gap between 1D and 2D models is not significant and 1D model can be directly compared to models
using raw waveforms.
3.2 Frame-level raw waveform model
In the frame-level raw waveform model, a strided convolution layer is added beneath the bottom layer of the
frame-level mel-spectrogram model. The strided convolution layer is expected to learn a filter-bank representation
that correspond to filter kernels in a time-frequency representation. In this model, once the raw waveforms pass
through the first strided convolution layer, the output feature map has the same dimensions as the mel-spectrogram.
This is because the stride, filter length, and number of filters of the first convolution layer correspond to the hop
size, window size, and number of mel-bands in the melspectrogram, respectively. This configuration was used for
music auto-tagging task in [11, 12] and so we used it as a
baseline model.
3.3 Sample-level raw waveform model
As described in Section 1, the approach using the raw waveforms should be able to address log-scale amplitude compression and phase-invariance. Simply adding a strided
convolution layer is not sufficient to overcome the problems. To improve this, we add multiple layers beneath the
frame-level such that the first convolution layer can handle much smaller length of samples. For example, if the
stride of the first convolution layer is reduced from 729
(= 36 ) to 243 (= 35 ), 3-size convolution layer and maxpooling layer are added to keep the output dimensions in
the subsequent convolution layers unchanged. If we repeatedly reduce the stride of the first convolution layer this
way, six convolution layers (five pairs of 3-size convolution
and max-pooling layer following one 3-size strided convolution layer) will be added (we assume that the temporal
dimensionality reduction occurs only through max-pooling
and striding while zero-padding is used in convolution to
preserve the size). We describe more details on the configuration strategy of sample-level CNN model in the following section.
3.4 Model Design
Since the length of an audio clip is variable in general, the
following issues should be considered when configuring
the temporal CNN architecture:
• Convolution filter length and sub-sampling length
• The temporal length of hidden layer activations on
the last sub-sampling layer
• The segment length of audio that corresponds to the
input size of the network
First, we attempted a very small filter length in convolutional layers and sub-sampling length, following the VGG
net that uses filters of 3 × 3 size and max-pooling of 2 × 2
size [2]. Since we use one-dimensional convolution and
sub-sampling for raw waveforms, however, the filter length
and pooling length need to be varied. We thus constructed
several DCNN models with different filter length and pooling length from 2 to 5, and verified the effect on music
auto-tagging performance. As a sub-sampling method, maxpooling is generally used. Although sub-sampling using
strided convolution has recently been proposed in a generative model [9], our preliminary test showed that maxpooling was superior to the stride sub-sampling method. In
addition, to avoid exhausting model search, a pair of single convolution layer and max-pooling layer with the same
size was used as a basic building module of the DCNN.
Second, the temporal length of hidden layer activations
on the last sub-sampling layer reflects the temporal compression of the input audio by successive sub-sampling.
We set the CNN models such that the temporal length of
hidden layer activation is one. By building the models this
way, we can significantly reduce the number of parameters
between the last sub-sampling layer and the output layer.
Also, we can examine the performance only by the depth
of layers and the stride of first convolution layer.
Third, in music classification tasks, the input size of the
network is an important parameter that determines the classification performance [16, 17]. In the mel-spectrogram
model, one song is generally divided into small segments
with 1 to 4 seconds. The segments are used as the input for
training and the predictions over all segments in one song
are averaged in testing. In the models that use raw waveform, the learning ability according to the segment size has
been not reported yet and thus we need to examine different input sizes when we configure the CNN models.
Considering all of these issues, we construct mn -DCNN
models where m refers to the filter length and pooling length
of intermediate convolution layer modules and n refers to
the number of the modules (or depth). An example of mn DCNN models is shown in Table 1 where m is 3 and n is
9. According to the definition, the filter length and pooling length of the convolution layer are 3 other than the first
strided convolution layer. If the hop size (stride length) of
the first strided convolution layer is 3, the time-wise output
dimension of the convolution layer becomes 19683 when
the input of the network is 59049 samples. We call this “39
model with 19683 frames and 59049 samples as input”.
4. EXPERIMENTAL SETUP
In this section, we introduce the datasets used in our experiments and describe experimental settings.
39 model, 19683 frames
59049 samples (2678 ms) as input
layer
stride
output
# of params
conv 3-128
3
19683 × 128
512
conv 3-128
maxpool 3
1
3
19683 × 128
6561 × 128
49280
conv 3-128
maxpool 3
1
3
6561 × 128
2187 × 128
49280
conv 3-256
maxpool 3
1
3
2187 × 256
729 × 256
98560
conv 3-256
maxpool 3
1
3
729 × 256
243 × 256
196864
conv 3-256
maxpool 3
1
3
243 × 256
81 × 256
196864
conv 3-256
maxpool 3
1
3
81 × 256
27 × 256
196864
conv 3-256
maxpool 3
1
3
27 × 256
9 × 256
196864
conv 3-256
maxpool 3
1
3
9 × 256
3 × 256
196864
conv 3-512
maxpool 3
1
3
3 × 512
1 × 512
393728
conv 1-512
dropout 0.5
1
−
1 × 512
1 × 512
262656
sigmoid
−
50
25650
Total params
1.9 × 106
Table 1. Sample-level CNN configuration. For example,
in the layer column, the first 3 of “conv 3-128” is the filter
length, 128 is the number of filters, and 3 of “maxpool 3”
is the pooling length.
4.1 Datasets
We evaluate the proposed model on two datasets, MagnaTagATune dataset (MTAT) [18] and Million Song Dataset
(MSD) annotated with the Last.FM tags [19]. We primarily examined the proposed model on MTAT and then verified the effectiveness of our model on MSD which is much
larger than MTAT 1 . We filtered out the tags and used most
frequently labeled 50 tags in both datasets, following the
previous work [11], [14, 15] 2 . Also, all songs in the two
datasets were trimmed to 29.1 second long and resampled
to 22050 Hz as needed. We used AUC (Area Under Receiver Operating Characteristic) as a primary evaluation
metric for music auto-tagging.
1 MTAT contains 170 hours long audio and MSD contains 1955 hours
long audio in total
2 https://github.com/keunwoochoi/MSD_split_for_
tagging
2n models
model with 16384 samples (743 ms) as input
model with 32768 samples (1486 ms) as input
model
n
layer
filter length & stride
AUC
model
n
layer
filter length & stride
AUC
64 frames
128 frames
256 frames
512 frames
1024 frames
2048 frames
4096 frames
8192 frames
6
7
8
9
10
11
12
13
1+6+1
1+7+1
1+8+1
1+9+1
1+10+1
1+11+1
1+12+1
1+13+1
256
128
64
32
16
8
4
2
0.8839
0.8899
0.8968
0.8994
0.9011
0.9031
0.9036
0.9032
128 frames
256 frames
512 frames
1024 frames
2048 frames
4096 frames
8192 frames
16384 frames
7
8
9
10
11
12
13
14
1+7+1
1+8+1
1+9+1
1+10+1
1+11+1
1+12+1
1+13+1
1+14+1
256
128
64
32
16
8
4
2
0.8834
0.8872
0.8980
0.8988
0.9017
0.9031
0.9039
0.9040
3n models
model with 19683 samples (893 ms) as input
model with 59049 samples (2678 ms) as input
model
n
layer
filter length & stride
AUC
model
n
layer
filter length & stride
AUC
27 frames
81 frames
243 frames
729 frames
2187 frames
6561 frames
3
4
5
6
7
8
1+3+1
1+4+1
1+5+1
1+6+1
1+7+1
1+8+1
729
243
81
27
9
3
0.8655
0.8753
0.8961
0.9012
0.9033
0.9039
81 frames
243 frames
729 frames
2187 frames
6561 frames
19683 frames
4
5
6
7
8
9
1+4+1
1+5+1
1+6+1
1+7+1
1+8+1
1+9+1
729
243
81
27
9
3
0.8655
0.8823
0.8936
0.9002
0.9030
0.9055
4n models
model with 16384 samples (743 ms) as input
model with 65536 samples (2972 ms) as input
model
n
layer
filter length & stride
AUC
model
n
layer
filter length & stride
AUC
64 frames
256 frames
1024 frames
4096 frames
3
4
5
6
1+3+1
1+4+1
1+5+1
1+6+1
256
64
16
4
0.8828
0.8968
0.9010
0.9021
256 frames
1024 frames
4096 frames
16384 frames
4
5
6
7
1+4+1
1+5+1
1+6+1
1+7+1
256
64
16
4
0.8813
0.8950
0.9001
0.9026
5n models
model with 15625 samples (709 ms) as input
model with 78125 samples (3543 ms) as input
model
n
layer
filter length & stride
AUC
model
n
layer
filter length & stride
AUC
125 frames
625 frames
3125 frames
3
4
5
1+3+1
1+4+1
1+5+1
125
25
5
0.8901
0.9005
0.9024
625 frames
3125 frames
15625 frames
4
5
6
1+4+1
1+5+1
1+6+1
125
25
5
0.8870
0.9004
0.9041
Table 2. Comparison of various mn -DCNN models with different input sizes. m refers to the filter length and pooling length
of intermediate convolution layer modules and n refers to the number of the modules. Filter length & stride indicates the
value of the first convolution layer. In the layer column, the first digit ’1’ of 1+n+1 is the strided convolution layer, and the
last digit ’1’ is convolution layer which actually works as a fully-connected layer.
4.2 Optimization
We used sigmoid activation for the output layer and binary cross entropy loss as the objective function to optimize. For every convolution layer, we used batch normalization [20] and ReLU activation. We should note that, in
our experiments, batch normalization plays a vital role in
training the deep models that takes raw waveforms. We
applied dropout of 0.5 to the output of the last convolution
layer and minimized the objective function using stochastic gradient descent with 0.9 Nesterov momentum. The
learning rate was initially set to 0.01 and decreased by a
factor of 5 when the validation loss did not decrease more
than 3 epochs. A total decrease of 4 times, the learning rate
of the last training was 0.000016. Also, we used batch size
of 23 for MTAT and 50 for MSD, respectively. In the melspectrogram model, we conducted the input normalization
simply by dividing the standard deviation after subtracting
mean value of entire input data. On the other hand, we did
not perform the input normalization for raw waveforms.
5. RESULTS
In this section, we examine the proposed models and compare them to previous state-of-the-art results.
5.1 mn -DCNN models
Table 2 shows the evaluation results for the mn -DCNN
models on MTAT for different input sizes, number of layers, filter length and stride of the first convolution layer. As
3n models,
59049 samples
as input
n
window
(filter length)
hop
(stride)
AUC
Frame-level
(mel-spectrogram)
4
5
5
6
6
729
729
243
243
81
729
243
243
81
81
0.9000
0.9005
0.9047
0.9059
0.9025
Frame-level
(raw waveforms)
4
5
5
6
6
729
729
243
243
81
729
243
243
81
81
0.8655
0.8742
0.8823
0.8906
0.8936
Sample-level
(raw waveforms)
7
8
9
27
9
3
27
9
3
0.9002
0.9030
0.9055
Table 3. Comparison of three CNN models with different window (filter length) and hop (stride) sizes. n represents the number of intermediate convolution and maxpooling layer modules, thus 3n times hop (stride) size of
each model is equal to the number of input samples.
input type
model
MTAT
MSD
Frame-level
(mel-spectrogram)
Persistent CNN [21]
2D CNN [14]
CRNN [15]
0.9013
0.894
-
0.851
0.862
Proposed DCNN
0.9059
-
Frame-level
(raw waveforms)
1D CNN [11]
0.8487
-
Sample-level
(raw waveforms)
Proposed DCNN
0.9055
0.8812
Table 4. Comparison of our works to prior state-of-the-arts
described in Section 3.4, m refers to the filter length and
pooling length of intermediate convolution layer modules
and n refers to the number of the modules. In Table 2,
we can first find that the accuracy is proportional to n for
most m. Increasing n given m and input size indicates that
the filter length and stride of the first convolution layer become closer to the sample-level (e.g. 2 or 3 size). When the
first layer reaches the small granularity, the architecture is
seen as a model constructed with the same filter length and
sub-sampling length in all convolution layers as depicted
in Table 1. The best results were obtained when m was 3
and n was 9. Interestingly, the length of 3 corresponds to
the 3-size spatial filters in the VGG net [2]. In addition, we
can see that 1-3 seconds of audio as an input length to the
network is a reasonable choice in the raw waveform model
as in the mel-spectrogram model.
5.2 Mel-spectrogram and raw waveforms
Considering that the output size of the first convolution
layer in the raw waveform models is equivalent to the melspectrogram size, we further validate the effectiveness of
the proposed sample-level architecture by performing experiments presented in Table 3. The models used in the
experiments follow the configuration strategy described in
Section 3.4. In the mel-spectrogram experiments, 128 melbands are used to match up to the number of filters in the
first convolution layer of the raw waveform model. FFT
size was set to 729 in all comparisons and the magnitude
compression is applied with a nonlinear curve, log(1 +
C|A|) where A is the magnitude and C is set to 10.
The results in Table 3 show that the sample-level raw
waveform model achieves results comparable to the framelevel mel-spectrogram model. Specifically, we found that
using a smaller hop size (81 samples ≈ 4 ms) worked better than those of conventional approaches (about 20 ms or
so) in the frame-level mel-spectrogram model. However,
if the hop size is less than 4 ms, the performance degraded.
An interesting finding from the result of the frame-level
raw waveform model is that when the filter length is larger
than the stride, the accuracy is slightly lower than the models with the same filter length and stride. We interpret that
this result is due to the learning ability of the phase variance. As the filter length decreases, the extent of phase
variance that the filters should learn is reduced.
5.3 MSD result and the number of filters
We investigate the capacity of our sample-level architecture even further by evaluating the performance on MSD
that is ten times larger than MTAT. The result is shown in
Table 4. While training the network on MSD, the number
of filters in the convolution layers has been shown to affect
the performance. According to our preliminary test results,
increasing the number of filters from 16 to 512 along the
layers was sufficient for MTAT. However, the test on MSD
shows that increasing the number of filters in the first convolution layer improves the performance. Therefore, we
increased the number of filters in the first convolution layer
from 16 to 128.
5.4 Comparison to state-of-the-arts
In Table 4, we show the performance of the proposed architecture to previous state-of-the-arts on MTAT and MSD.
They show that our proposed sample-level architecture is
highly effective compared to them.
5.5 Visualization of learned filters
The technique of visualizing the filters learned at each layer
allows better understanding of representation learning in
the hierarchical network. However, many previous works
in music domain are limited to visualizing learned filters
only on the first convolution layer [11, 12].
The gradient ascent method has been proposed for filter
visualization [22] and this technique has provided deeper
understanding of what convolutional neural networks learn
from images [23, 24]. We applied the technique to our
DCNN to observe how each layer hears the raw waveforms. The gradient ascent method is as follows. First, we
generate random noise and back-propagate the errors in the
network. The loss is set to the target filter activation. Then,
Layer 1
Layer 2
Layer 3
Layer 4
Layer 5
Layer 6
Figure 2. Spectrum of the filters in the sample-level convolution layers which are sorted by the frequency of the peak
magnitude. The x-axis represents the index of the filter, and the y-axis represents the frequency. The model used for
visualization is 39 -DCNN with 59049 samples as input. Visualization was performed using the gradient ascent method to
obtain the input waveform that maximizes the activation of a filter in the layers. To effectively find the filter characteristics,
we set the input waveform estimate to 729 samples which is close to a typical frame size.
we add the bottom gradients to the input with gradient normalization. By repeating this process several times, we can
obtain the waveform that maximizes the target filter activation. Examples of learned filters at each layer are in Figure
3. Although we can find the patterns that low-frequency
filters are more visible along the layer, estimated filters are
still noisy. To show the patterns more clearly, we visualized them as spectrum in the frequency domain and sorted
them by the frequency of the peak magnitude.
Note that we set the input waveform estimate to 729 samples in length because, if we initialize and back-propagate
to the whole input size of the networks, the estimated filters
will have large dimensions such as 59049 samples in computing spectrum. Thus, we used the smaller input samples
which can find the filter characteristics more effectively
and also are close to a typical frame size in spectrum.
The layer 1 shows the three distinctive filter bands which
are possible with the filter length with 3 samples (say, a
DFT size of 3). The center frequency of the filter banks
increases linearly in low frequency filter banks but it becomes non-linearly steeper in high frequency filter banks.
This trend becomes stronger as the layer goes up. This
nonlinearity was found in learned filters with a frame-level
end-to-end learning [11] and also in perceptual pitch scales
such as mel or bark.
Figure 3. Examples of learned filters at each layer.
Acknowledgments
This work was supported by Korea Advanced Institute of
Science and Technology (project no. G04140049) and National Research Foundation of Korea (project no. N01160463).
7. REFERENCES
6. CONCLUSION AND FUTURE WORK
In this paper, we proposed sample-level DCNN models
that take raw waveforms as input. Through our experiments, we showed that deeper models (more than 10 layers) with a very small sample-level filter length and subsampling length are more effective in the music auto-tagging
task and the results are comparable to previous state-ofthe-art performances on the two datasets. Finally, we visualized hierarchically learned filters. As future work, we
will analyze why the deep sample-level architecture works
well without input normalization and nonlinear function
that compresses the amplitude and also investigate the hierarchically learned filters more thoroughly.
[1] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in neural information processing
systems, 2012, pp. 1097–1105.
[2] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,”
arXiv preprint arXiv:1409.1556, 2014.
[3] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado,
and J. Dean, “Distributed representations of words
and phrases and their compositionality,” in Advances
in neural information processing systems, 2013, pp.
3111–3119.
[4] X. Zhang, J. Zhao, and Y. LeCun, “Character-level
convolutional networks for text classification,” in Advances in neural information processing systems, 2015,
pp. 649–657.
[17] J. Lee and J. Nam, “Multi-level and multi-scale feature aggregation using pre-trained convolutional neural networks for music auto-tagging,” arXiv preprint
arXiv:1703.01793, 2017.
[5] Y. Kim, Y. Jernite, D. Sontag, and A. M. Rush,
“Character-aware neural language models,” arXiv
preprint arXiv:1508.06615, 2015.
[18] E. Law, K. West, M. I. Mandel, M. Bay, and J. S.
Downie, “Evaluation of algorithms using games: The
case of music tagging,” in ISMIR, 2009, pp. 387–392.
[6] D. Palaz, M. M. Doss, and R. Collobert, “Convolutional neural networks-based continuous speech recognition using raw speech signal,” in IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 4295–4299.
[19] T. Bertin-Mahieux, D. P. Ellis, B. Whitman, and
P. Lamere, “The million song dataset,” in Proceedings
of the 12th International Conference on Music Information Retrieval (ISMIR), vol. 2, no. 9, 2011, pp. 591–
596.
[7] D. Palaz, R. Collobert et al., “Analysis of cnn-based
speech recognition system using raw speech as input,”
Idiap, Tech. Rep., 2015.
[20] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” arXiv preprint arXiv:1502.03167, 2015.
[8] R. Collobert, C. Puhrsch, and G. Synnaeve,
“Wav2letter: an end-to-end convnet-based speech
recognition system,” arXiv preprint arXiv:1609.03193,
2016.
[21] J.-Y. Liu, S.-K. Jeng, and Y.-H. Yang, “Applying topological persistence in convolutional neural
network for music audio signals,” arXiv preprint
arXiv:1608.07373, 2016.
[9] A. van den Oord, S. Dieleman, H. Zen, K. Simonyan,
O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior, and
K. Kavukcuoglu, “Wavenet: A generative model for
raw audio,” CoRR abs/1609.03499, 2016.
[22] D. Erhan, Y. Bengio, A. Courville, and P. Vincent,
“Visualizing higher-layer features of a deep network,”
University of Montreal, vol. 1341, p. 3, 2009.
[10] T. N. Sainath, R. J. Weiss, A. W. Senior, K. W. Wilson,
and O. Vinyals, “Learning the speech front-end with
raw waveform cldnns.” in INTERSPEECH, 2015, pp.
1–5.
[11] S. Dieleman and B. Schrauwen, “End-to-end learning
for music audio,” in IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP),
2014, pp. 6964–6968.
[12] D. Ardila, C. Resnick, A. Roberts, and D. Eck, “Audio
deepdream: Optimizing raw audio with convolutional
networks.”
[13] J. Pons, T. Lidy, and X. Serra, “Experimenting with
musically motivated convolutional neural networks,” in
IEEE International Workshop on Content-Based Multimedia Indexing (CBMI), 2016, pp. 1–6.
[14] K. Choi, G. Fazekas, and M. Sandler, “Automatic tagging using deep convolutional neural networks,” in
Proceedings of the 17th International Conference on
Music Information Retrieval (ISMIR), 2016, pp. 805–
811.
[15] K. Choi, G. Fazekas, M. Sandler, and K. Cho, “Convolutional recurrent neural networks for music classification,” arXiv preprint arXiv:1609.04243, 2016.
[16] P. Hamel, S. Lemieux, Y. Bengio, and D. Eck, “Temporal pooling and multiscale learning for automatic annotation and ranking of music audio,” in Proceedings
of the 12th International Conference on Music Information Retrieval (ISMIR), 2011.
[23] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in European conference on computer vision. Springer, 2014, pp. 818–
833.
[24] A. Nguyen, J. Yosinski, and J. Clune, “Deep neural
networks are easily fooled: High confidence predictions for unrecognizable images,” in Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition, 2015, pp. 427–436.
| 9cs.NE
|
Likelihood of Cyber Data Injection Attacks to
Power Systems
Yingshuai Hao, Student Member, IEEE, Meng Wang, Member, IEEE, Joe Chow, Fellow, IEEE
arXiv:1512.05008v1 [cs.SY] 15 Dec 2015
Department of Electrical, Computer, and Systems Engineering
Rensselaer Polytechnic Institute, Troy, NY, 12180, USA
Email: {haoy2,wangm7,chowj}@rpi.edu
Abstract—Cyber data attacks are the worst-case interacting
bad data to power system state estimation and cannot be
detected by existing bad data detectors. In this paper, we for
the first time analyze the likelihood of cyber data attacks by
characterizing the actions of a malicious intruder. We propose
to use Markov decision process to model an intruder’s strategy,
where the objective is to maximize the cumulative reward across
time. Linear programming method is employed to find the
optimal attack policy from the intruder’s perspective. Numerical
experiments are conducted to study the intruder’s attack strategy
in test power systems.
Index Terms—cyber data attacks, Markov Decision Process,
state estimation, power systems.
I. I NTRODUCTION
State estimation [1] solves for power system states from
obtained measurements. The correct estimation of systems
states is vital for the reliable operation of power systems. Since
bad data can result in significant errors in the outcome of state
estimation and potentially lead to catastrophic consequences,
the detection and identification of bad data has been extensively studied [3], [6], [13], [14], [21] in state estimation.
The integration of cyber infrastructures in future smart grids
inevitably increases the possibility of cyber attacks. Cyber
data attacks, firstly studied in [12], means that a malicious
intruder with system configuration information simultaneously
manipulates multiple measurements and the injected errors
cannot be detected by any bad data detector.
State estimation in the presence of cyber data attacks has
attracted much research attention recently [2], [4], [9], [11],
[12], [16]–[18]. Some work focused on the identification of
a small number of key measurement units such that if those
units are protected from an intruder, the intruder cannot launch
a successful cyber data attack [2], [4], [8]. A few recent
work [11], [17], [19] considered the detection of cyber data
attacks in a power system with Supervisory Control and Data
Acquisition (SCADA) systems or phasor measurement units
(PMUs) and various detection methods have been proposed.
What is missing in the literature of cyber data attacks is the
analysis of the frequency of data attacks in smart grids and
the likelihood of attacks at a given system state. This paper
takes the first step to analyze the likelihood of data attacks
from the intruder’s perspective. We consider a scenario that if
a cyber data attack is detected by the operator, the affected
measurement units will be protected for some time. Thus, the
intruder’s current action affects its future available actions.
To address such challenge in developing an attack strategy,
we propose to use Markov Decision Process (MDP) [15] to
model the intruder’s attack decision across time. The solution
to the resulting MDP is a mapping from system states to the
intruder’s actions (attack or not, which bus to attack and how
much error to inject). Numerical experiments are carried on
PJM 5-bus system to verify the proposed approach and study
the likelihood of cyber attacks in such systems.
The rest of the paper is organized as follows. We motivate
the problem and introduce cyber data attacks and MDP in
Section II. We formulate the intruder’s attack strategy as an
MDP and introduce its solution method in Section III. Section
IV records our numerical study on an example network. We
conclude the paper in Section V.
II. P ROBLEM M OTIVATION
AND
BACKGROUND
We first introduce the definition and existing work on cyber
data attacks in Section II-A. We then motivate and introduce
the problem of likelihood analysis of cyber data attacks in
Section II-B. One main contribution of our paper is to model
this problem as a Markov Decision Process (MDP). MDPs are
introduced in Section II-C.
A. Cyber Data Attacks in Power Systems
In a power system, the state is usually represented by bus
voltage magnitudes V ∈ Rn and angles θ ∈ [−π, π]n , where
n is the number of buses. State estimation [1] solves for
system states from the obtained measurements. Under the AC
measurement model, the measurements z can be expressed as
a nonlinear function of state variables x = (V ,θ):
z = h(x) + ω,
(1)
where ω represents the random measurement noise.
In the AC state estimation, the state variables are determined
from the weighted least square optimization problem:
x̂ = arg min(z − h(x))T · R−1 · (z − h(x)),
(2)
where R is the covariance matrix of measurement noise ω.
Malicious intruders can hack the measuring devices and
inject interacting errors to the measurements. If they have sufficient system information and choose the errors ez satisfying
z + ez = h(x′ ) + ω
= h(V + eV , θ + eθ ) + ω,
(3)
where eV and eθ represent the resulting errors on state
variables V and θ respectively, the manipulated measurements
cannot be detected by existing bad data detectors. In this
case, instead of correctly estimating state variables (V , θ), the
operator would obtain a wrong estimate (V + eV , θ + eθ ).
Since such cyber data attacks cannot be identified by bad
data detectors, many efforts have been devoted to identify and
protect a small number of key measurement units such that
an intruder cannot inject unobservable attacks without hacking
protected units [2], [4], [8]. A few recent work [11], [17], [19]
considered the detection of data attacks and various detection
methods have been proposed.
The potential financial risks of cyber data attacks are studied
in [20] and [7], where the congestion pattern is defined as the
set of congested lines. By injecting false data without being
detected, the intruders could change the congestion pattern and
thus change the locational marginal price (LMP). The intruders
can obtain financial reward from the resulting change in LMP.
In this paper, we restrict our attention to attacks that satisfy
(3) and result in a change of the system congestion pattern.
By launching a data attack, if a line’s real power is wrongly
estimated to exceed its capacity while it actually is not, or the
power is wrongly estimated to below its limit while the line is
actually congested, then the intruder can gain a reward from
the attack.
B. Likelihood of Cyber Data Attacks
One important question that has not been considered before
is the analysis of the likelihood of cyber data attacks at a given
operating state of power systems.
In this paper, we act as an intruder and aim to find the
optimal attack strategy from an intruder’s perspective. We
assume that an intruder can obtain a reward from a change
of the congestion pattern by injecting false data without
being detected. The intruder aims to maximize the cumulative
reward. If a cyber data attack, however, is detected by detection
methods such as [11], [17], [19], we assume the intruded
measuring devices will be protected for some time so that
an intruder cannot change the measurements of these devices
during the period. A detected attack, therefore, can limit the
intruder’s future actions and thus reduce the future reward.
Since the state of a power system changes across time, and
the future state is unknown to the intruder, it needs to decide
when and which buses to attack to maximize the total rewards
based on its current estimate. We employ Markov Chains [15]
to model the evolution of power system states and formulate
the intruder’s decision process as a Markov Decision Process
[15]. The solution of resulting MDP is a mapping from system
states to the intruder’s actions.
C. Markov Decision Processes (MDPs)
An MDP is a mathematical framework employed to model
the decision-making process in stochastic environments. In
this framework, the system is modeled via a series of states
S. Each state s ∈ S has an associated set of actions A(s).
In time step t, a decision is made based on the system’s
current state st ∈ S, and an action at ∈ A(st ) is chosen to
conduct. The cost of taking action at at state st is G(st , at ).
Then following the state transition probability distribution,
the system transits to a new state st+1 with a probability
of P (st+1 |st , at ). A reward R(st+1 |st , at ) is received from
the state transition. As the system evolves, a sequence of
rewards is obtained. The aim for decision-makers is to choose
sequential actions that yield maximal expected rewards over
the total decision-making horizon. The MDP problem can
be solved by numerous methods, like value iteration, policy
iteration and linear programming approaches discussed in [15].
III. P ROBLEM F ORMULATION
AND
S OLUTION
In Section III-A, the problem is formulated from the perspective of attackers and the strategy of cyber data attacks is
modeled as an infinite-horizon MDP. The solution method of
resulting MDP is discussed in Section III-B.
A. Problem Formulation
1) States and Time Steps: Here we employ bus voltage
magnitudes, angles and the states of measuring devices together as system states in an MDP. Because the measurements
contain noise and an intruder may have limited knowledge
of the states of a power system, we use discrete states to
model the intruder’s estimate of actual power system states.
For example, let Vi denote the voltage magnitude of bus i,
Vimax and Vimin denote the upper bound and lower bound of
Vi respectively. ∆Vi = Vimax − Vimin . nV denotes the number
of discrete states in the range. We define the state of the voltage
magnitude of bus i as
V̄i = q/(nV − 1), q ∈ {0, 1, · · · , nV − 1},
(4)
h
i
i
. Sim, Vimin + (q + 1) × n∆V
if Vi ∈ Vimin + q × n∆V
V −1
V −1
max
ilarly, let θi denote the voltage angle of bus i, θi
and
θimin denote its upper bound and lower bound respectively,
nθ denote the number of discrete states. ∆θi = θimax − θimin .
We say the state of the voltage angle of bus i is
θ̄i = q/(nθ − 1), q ∈ {0, 1, · · · , nθ − 1},
(5)
h
i
i
if θi ∈ θimin + q × n∆θ
. The state
, θimin + (q + 1) × n∆θ
θ −1
θ −1
of the jth measuring device is denoted by a variable Ūj :
(
1 jth device is open to attack,
Ūj =
(6)
0 jth device is protected from intrusion .
If a measuring device is protected, then an intruder cannot
change any measurements of that device. Otherwise, an intruder can change partial or all measurements of that device.
We consider a discrete-time system, and the time step is
set as the duration between two consecutive instants of state
estimation. The sampling rate of measuring devices can be
higher than the state estimation frequency, as shown in Fig. 1.
Attacks can happen during either device sampling or data
transmission to control center. We assume if intruders decide
to attack a device during step t, they need to change all
the measurements of that device in the time step. Otherwise,
1st measurement
Device
Sampling
Device State
Update
Data
Transmission
Device
Sampling
nd
Data Arrived at Control
Center
2 measurement
Data
Transmission
...
Data Arrived at
Control Center
kth measurement
Device
Sampling
Time step t
Data
Transmission
Bad Data
Processing,
Cyber Attack Device
State
Detection, Update
State
Estimation
t+1
Fig. 1. Event sequence with cyber data attacks. An intruder changes the
observations of measuring devices to mislead the operator.
change in the congestion pattern, we define the reward as a
function proportional to the gap between the line’s flow limit
and the power bounds with injected errors:
Kij × Pijmin (V¯ ′ , θ¯′ ) − PijM /PijM ,
if Pijmin (V¯ ′ , θ¯′ ) > PijM > Pijmax (V̄ , θ̄);
(9)
Rij =
Kij × PijM − Pijmax (V¯ ′ , θ¯′ ) /PijM ,
if Pijmin (V̄ , θ̄) > PijM > Pijmax (V¯ ′ , θ¯′).
M
the attack can be easily detected by comparing consecutive where Kij is the given reward weight of line ij, Pij is the
′
′
¯
¯
measurements. If a power system has M buses and N devices, power flow limit of line ij, (V , θ ) is the resulting estimate
of system states by injecting errors (eV , eθ ) to (V̄ , θ̄).
the state st at time step t is
The expected immediate reward from action a at state s is:
st = V̄ (t), θ̄(t), Ū (t)
X
R(s, a) =
P (s′ |s, a) × R(s′ |s, a)
= V̄1 (t), · · · , V̄M (t), θ̄1 (t), · · · , θ̄M (t), Ū1 (t), · · · , ŪN (t) ,
s′
(7)
X
(10)
Rij .
= (1 − pd (s, a)) ×
where V̄i (t), θ̄i (t) and Ūj (t) represent the states of voltage
ij∈Φ(s,a)
magnitude and angle of bus i and the state of jth device at
time step t respectively.
where Φ(s, a) is the set of target lines.
2) Actions, Rewards and Costs: Since the reward results
We assume the cost to intrude an accessible measuring
from a change in the congestion pattern, in order to change the device is fixed and known to an intruder, denoted by gu . Let
line power, the intruder needs to inject errors to the estimate f (Φ(s, a)) denote the number of intruded measuring devices
of voltage phasors of incident buses. Let eVi and eθi denote in attack a, the attack cost at state s is:
the injected errors to the voltage magnitude and angle of bus
G(s, a) = gu × f (Φ(s, a)).
(11)
i respectively. To make the problem tractable, we assume eVi
∆Vi
∆θi
and eθi can only be multiples of nV −1 and nθ −1 respectively,
3) State Transition Probabilities: We assume all measuring
and the resulting estimates of system variables still lie in
devices that are currently open to attack will stay open without
individual allowable range. In this case, there are totally
attack. An action a at state s is detected with probability
nV × nθ available ways to inject errors to one bus voltage
pd (s, a), and the intruded devices will be protected as a whole
phasor. Note that in order to pass the bad data detection, given
in the next time step. Each protected device will change
eV and eθ , an intruder needs to choose the injected errors ez
to open in the next time step with a fixed probability pT .
to measurements according to (3).
Intuitively, a smaller pT indicates that once protected, a device
We call a set of buses and lines as target buses and
is more likely to stay inaccessible to intruders for a longer
target lines respectively if the intruder attempts to change the
period of time. When pT = 0, it means the protected devices
congestion pattern of these lines by injecting errors on the
will no longer be accessible to intruders.
target buses. Since an intruder may have limited resources to
To model the dynamics of a power system, we assume
launch attacks, we assume at each time step the intruder can
each load in the system evolves independently as a Markov
manipulate the states ofPat mostd bus voltages. The intruder,
Chain [15] and the system state is determined from economic
therefore, has at most di=0 Mi ways to select target buses.
dispatch. We assume each load has nL states and a load
The launched attacks can be detected by some recently
can transit from state q1 to state q2 with a fixed and known
developed methods, as presented in section II-B. Here we use
probability pq1 ,q2 . In practice, one can learn these transition
pd (s, a) to denote the probability that an action a at state s is
probabilities from historical data. In this case, in a power
detected by the network operator. We suppose it is a function
network with M buses and N devices, if ML loads evolve as
of injected errors on the bus voltage magnitudes and angles:
N
L
Markov
Chains, the total number of system state is nM
L ×2 .
!
M
X
|eVi | |eθi |
, (8) B. MDP Solutions
+
pd (s, a) = 1 − exp −C ×
∆Vi
∆θi
i=1
A linear programming approach [5], [15] is applied to solve
where C is a positive constant. Intuitively, a larger C means MDP. A stationary policy π for an MDP is a mapping π : S 7→
a higher probability with which the launched attack can be A, where π(s) is the action taken in state s. We define Wπ (s)
detected.
as the cumulative expected net reward by starting from state
Since the bus voltage magnitudes and angles are in dis- s and following policy π,
"∞
#
cretized states, instead of computing the power flow of line ij
X
directly from one specific state (V̄ , θ̄), we can obtain the lower
t
Wπ (s) = E
γ (R(st , π(st )) − G(st , π(st ))|s0 = s) ,
and upper bound of its absolute value, denoted as Pijmin (V̄ , θ̄)
t=0
max
(12)
and Pij (V̄ , θ̄) respectively. Since the reward results from a
W ∗ (s) := max Wπ (s),
(13)
π∈Π
where Π is the set of all available policies. The policy that
achieves the maximum in (13) is the optimal policy π ∗ . It
is shown in [5] that W ∗ (s) is the optimal solution to the
following optimization problem:
X
min
Q(s)
Q
s∈S
s.t. Q(s) ≥ R(s, a) − G(s, a) + γ
X
P (s′ |s, a)Q(s′ ) (14)
s′
∀a ∈ A(s), ∀s, s′ ∈ S.
Therefore, we can find W ∗ (s) by solving (14) and compute
π ∗ (s) defined as
X
arg max(R(s, a) − G(s, a) + γ
P (s′ |s, a)W ∗ (s′ )). (15)
a∈A(s)
s′
The optimal attack strategy is a mapping between system state
s and the corresponding optimal action π ∗ (s). An intruder can
solve the MDP to obtain the strategy offline and then inject
attacks accordingly in real-time operations.
determined from the actual values of bus i in eight load states.
For each discretized state, we calculate the lower and upper
bounds of each line power flow. One line can be an available
target line if the upper bound of its power flow is below the
power limit or the lower bound is over the limit.
We solve (14) and (15) to obtain the optimal actions for
23 × 23 = 64 states and compute the static distribution
probabilities of all states. The attack probability of one line
is computed as the sum of the distribution probabilities of
the states under which the optimal action is to change the
congestion pattern of that line. The intrusion probability of
one device is the sum of the distribution probabilities of the
states under which the optimal action requires manipulating
partial or all measurement of that device. The results of the
likelihood of data attacks are shown in Figs. 3 and 4.
Attack Probability of Part of Lines
1
Attack probability
where γ ∈ [0, 1) is the discount factor for future reward. The
value of state s is the maximal cumulative reward,
Line 1
Line 2
Line 4
0.8
0.6
0.4
0.2
0
0
0.5
1
1.5
2
2.5
3
3.5
4
C: Tuning Parameter of Attack Detection Probability
Fig. 3.
IV. S IMULATION
Load Center
D
ļ
$10
600MW
$35
200MW
ĸ
Ļ
$14
40MW
$15
170MW
ĺ
ķ
A
Measuring Device
Fig. 2.
B
300
$30
520MW
C
300
1
Device on bus A
Device on bus C
Device on bus D
0.8
0.6
0.4
0.2
0
0
0.5
1
1.5
2
2.5
3
3.5
4
C: Tuning Parameter of Attack Detection Probability
Fig. 4.
400
Ĺ
Intrusion probability
We test our proposed method on the PJM 5-bus system. The
basic system configuration and the generation bids, generation
MW limits and MW loads are shown in Fig. 2. More details
can be found in [10].
Generation Center
E
Attack probabilities of part of lines
Intrusion Probability of Measuring Devices
Load Unit: MW
The PJM 5-bus system
1) States of loads and transition probability. Each load is
assumed to have 2 states. The loads in Fig. 2 are the base
case loads. Another state for each load is half of its base case.
The transition probability from one state to another is 0.5.
2) Measuring device transition probability. pT is set 0.5. The
attack detection probability is calculated from (8).
3) Costs. The number of target buses at each time step is at
most one. The reward weight of each line is 1. The cost gu is
set as 0.05.
4) Discount factor. The discount factor γ = 0.95.
We solve the economic dispatch in MATPOWER toolbox in
MATLAB. The power flow limit of each line is set 300 MW.
We relax the constraint in economic dispatch from Pij ≤ PijM
to Pij ≤ 1.2PijM. We set V max = 1.1 p.u. (per unit), V min =
1.0 p.u., nV = 5; ∆θ = 5◦ , nθ = 10, θimax and θimin are
Intrusion probabilities of measuring devices
Generally, as C increases, the attack probability of each line
and each measuring device decreases. When C = 0, it means
the launched attack cannot be detected by network operators.
In this case, the intruders always choose the action that brings
about the maximal immediate net reward. When C ≥ 4, the net
expected reward for each action is negative, hence the optimal
action for all states is no attack.
V. C ONCLUSION AND D ISCUSSIONS
We for the first time analyze the likelihood of cyber data
attacks to power systems. We model an intruder’s attack
strategy as a Markov Decision Process (MDP). We compute
the optimal attack action at a given power system state from an
intruder’s perspective. We study the likelihood of cyber data
attacks on a small system through simulation. One ongoing
work is to apply the method to likelihood analysis of cyber
data attacks on larger power systems.
ACKNOWLEDGEMENT
This research is supported in part by the ERC Program of
NSF and DoE under the supplement to NSF Award EEC1041877 and the CURENT Industry Partnership Program, and
in part by NYSERDA Grants #36653 and #28815.
R EFERENCES
[1] A. Abur and A. G. Exposito, Power system state estimation: theory and
implementation. CRC Press, 2004.
[2] R. B. Bobba, K. M. Rogers, Q. Wang, H. Khurana, K. Nahrstedt,
and T. J. Overbye, “Detecting false data injection attacks on DC state
estimation,” in Proc. the First Workshop on Secure Control Systems
(SCS), 2010.
[3] J. Chen and A. Abur, “Placement of PMUs to enable bad data detection
in state estimation,” IEEE Trans. Power Syst., vol. 21, no. 4, pp. 1608–
1615, 2006.
[4] G. Dán and H. Sandberg, “Stealth attacks and protection schemes
for state estimators in power systems,” in Proc. IEEE International
Conference on Smart Grid Communications (SmartGridComm), 2010,
pp. 214–219.
[5] D. P. de Farias and B. Van Roy, “The linear programming approach
to approximate dynamic programming,” Operations Research, vol. 51,
no. 6, pp. 850–865, 2003.
[6] E. Handschin, F. Schweppe, J. Kohlas, and A. Fiechter, “Bad data
analysis for power system state estimation,” IEEE Trans. Power App.
Syst., vol. 94, no. 2, pp. 329–337, 1975.
[7] L. Jia, J. Kim, R. J. Thomas, and L. Tong, “Impact of data quality on
real-time locational marginal price,” Power Systems, IEEE Transactions
on, vol. 29, no. 2, pp. 627–636, 2014.
[8] T. Kim and H. Poor, “Strategic protection against data injection attacks
on power grids,” IEEE Trans. Smart Grid, vol. 2, no. 2, pp. 326–333,
2011.
[9] O. Kosut, L. Jia, R. Thomas, and L. Tong, “Malicious data attacks on
smart grid state estimation: Attack strategies and countermeasures,” in
Proc. IEEE International Conference on Smart Grid Communications
(SmartGridComm), 2010, pp. 220–225.
[10] F. Li and R. Bo, “Small test systems for power system economic studies,”
in Power and Energy Society General Meeting, 2010 IEEE. IEEE, 2010,
pp. 1–4.
[11] L. Liu, M. Esmalifalak, Q. Ding, V. A. Emesih, and Z. Han, “Detecting
false data injection attacks on power grid by sparse optimization,” IEEE
Trans. Smart Grid, vol. 5, no. 2, pp. 612–621, 2014.
[12] Y. Liu, P. Ning, and M. K. Reiter, “False data injection attacks against
state estimation in electric power grids,” ACM Transactions on Information and System Security (TISSEC), vol. 14, no. 1, p. 13, 2011.
[13] H. M. Merrill and F. C. Schweppe, “Bad data suppression in power
system static state estimation,” IEEE Trans. Power App. Syst., no. 6, pp.
2718–2725, 1971.
[14] A. Monticelli and A. Garcia, “Reliable bad data processing for real-time
state estimation,” IEEE Trans. Power App. Syst., no. 5, pp. 1126–1139,
1983.
[15] M. L. Puterman, “Markov decision processes: Discrete stochastic dynamic programming,” 1994.
[16] H. Sandberg, A. Teixeira, and K. H. Johansson, “On security indices
for state estimators in power networks,” in Proc. the First Workshop on
Secure Control Systems (SCS), 2010.
[17] H. Sedghi and E. Jonckheere, “Statistical structure learning of smart grid
for detection of false data injection,” in Proc. IEEE Power and Energy
Society General Meeting (PES), 2013, pp. 1–5.
[18] A. Tajer, S. Kar, H. V. Poor, and S. Cui, “Distributed joint cyber attack
detection and state recovery in smart grids,” in Proc. IEEE International
Conference on Smart Grid Communications (SmartGridComm), 2011,
pp. 202–207.
[19] M. Wang, P. Gao, S. Ghiocel, J. H. Chow, B. Fardanesh, G. Stefopoulos,
and M. P. Razanousky, “Identification of "unobservable" cyber data
attacks on power grids,” in Proc. IEEE SmartGridComm, 2014.
[20] L. Xie, Y. Mo, and B. Sinopoli, “Integrity data attacks in power market
operations,” IEEE Trans. Smart Grid, vol. 2, no. 4, pp. 659–666, 2011.
[21] W. Xu, M. Wang, L. Lai, and A. Tang, “Sparse error correction from
nonlinear measurements with applications in bad data detection for
power networks,” IEEE Trans. Signal Process., vol. 61, no. 24, pp.
6175–6187, 2013.
| 3cs.SY
|
A Novel Framework for Online Amnesic
Trajectory Compression in
Resource-constrained Environments
◦† Jiajun
∗ Shuo
Shang
† Philipp Sommer
Zhao
† Brano Kusy
Jae-Gil Lee
† Raja Jurdak
Liu
† Kun
arXiv:1605.02337v1 [cs.DS] 8 May 2016
◦ Renmin
University of China, Beijing, China {[email protected]}
† Data 61, CSIRO, Pullenvale, Australia
{jiajun.liu, kun.zhao, philipp.sommer, brano.kusy, raja.jurdak}@csiro.au
∗ China University of Petroleum, Beijing, China {[email protected]}
Department of Knowledge Service Engineering, KAIST, Korea {[email protected]}
Abstract— State-of-the-art trajectory compression methods usually involve high space-time complexity or yield unsatisfactory
compression rates, leading to rapid exhaustion of memory, computation, storage and energy resources. Their ability is commonly
limited when operating in a resource-constrained environment especially when the data volume (even when compressed) far exceeds
the storage limit. Hence we propose a novel online framework for error-bounded trajectory compression and ageing called the Amnesic
Bounded Quadrant System (ABQS), whose core is the Bounded Quadrant System (BQS) algorithm family that includes a normal
version (BQS), Fast version (FBQS), and a Progressive version (PBQS). ABQS intelligently manages a given storage and compresses
the trajectories with different error tolerances subject to their ages.
In the experiments, we conduct comprehensive evaluations for the BQS algorithm family and the ABQS framework. Using empirical
GPS traces from flying foxes and cars, and synthetic data from simulation, we demonstrate the effectiveness of the standalone BQS
algorithms in significantly reducing the time and space complexity of trajectory compression, while greatly improving the compression
rates of the state-of-the-art algorithms (up to 45%). We also show that the operational time of the target resource-constrained hardware
platform can be prolonged by up to 41%. We then verify that with ABQS, given data volumes that are far greater than storage space,
ABQS is able to achieve 15 to 400 times smaller errors than the baselines. We also show that the algorithm is robust to extreme
trajectory shapes.
Index Terms—Online trajectory compression; ageing; amnesic; resource-constrained; constant complexity
F
1
I NTRODUCTION
Location tracking is increasingly important for transport, ecology,
and wearable computing. In particular, long-term tracking of
spatially spread assets, such as wildlife [2], augmented reality
glasses [3], or bicycles [4] provides high resolution trajectory
information for better management and services. For smaller
moving entities, such as flying foxes [5] or pigeons [6], the size
and weight of tracking devices are constrained, which presents the
challenge of obtaining detailed trajectory information subject to
memory, processing, energy and storage constraints. However, the
requirements on the tracking precision are barely relaxed, e.g. the
goal of wildlife tracking sometimes is in the form of a guaranteed
tracking error from 10 to a few hundred meters [7].
Consider tracking of flying foxes as a motivating scenario.
The computing platform is constrained in computational resources
and is inaccessible in most occasions once deployed. The position
data is acquired in a streaming fashion. The RAM available on
the platform is 4 KBytes, while the storage space is only 1 MB
to store trajectories over weeks and months before they can be
•
This article is an extended version of [1]
offloaded. The combination of long-term operational requirements
and constrained resources therefore requires an intelligent online
framework that can process the trajectory stream instantaneously
and efficiently, i.e. in constant space and time, and that can achieve
high compression rate.
Current trajectory compression algorithms often fail to operate
under such requirements, as they either require substantial amount
of buffer space or require the entire data stream to perform the
compression [8] [9]. Existing online algorithms, which process
each point exactly once, operate retrospectively on trajectory data
or assume favourable trajectory characteristics, resulting in the
worst-case complexity of their online performance ranging from
O(nlogn) to O(n2 ) [10]. The high complexity of these methods limits their utility in resource-constrained environments. To
address this challenge, we propose the Bounded Quadrant System
(BQS), an online algorithm that can run sustainably on resourceconstrained devices. Its fast version achieves O(n) time and O(1)
space complexity while providing guaranteed error bounds. By
using a convex hull that is formed by a bounding box and two
angular bounds around all points, the algorithm is able to make
quick compression decisions without iterating through the buffer
and calculating the maximum error in most of the cases. Using
empirical GPS traces from the flying fox scenario and from cars,
we evaluate the performance of our algorithms and quantify their
benefits in improving the efficiency of trajectory compression. Its
amnesic version, the Amnesic Bounded Quadrant System (ABQS),
is able to manage a given volume of storage space and intelligently
trade off space for precision, so that no hard data losses (data
overwritten) will be present over a prolonged operation time.
Our contributions are threefold:
1)
2)
3)
We design an online compression algorithm family called
BQS. BQS uses convex hulls to compress streaming trajectories with error guarantees. The fast version in the algorithm
family achieves constant time and space complexity for each
step, or equivalently O(n) time complexity and O(1) space
complexity for the whole data stream.
We formulate the BQS algorithm with an uncertainty vector to support the progressive compression of trajectories,
and subsequently propose Progressive BQS (PBQS). Using
PBQS as the corner stone, we propose a sophisticated online framework called Amnesic Bounded Quadrant System
(ABQS) which not only compresses streaming trajectory
data efficiently, but also manages historical data in an amnesic way. ABQS introduces graceful degradation to the
entire trajectory history without hard data loss (overwriting),
and hence achieves excellent compression performances
when the operation duration is unknown and the data volume
far exceeds the storage limit.
We comprehensively evaluate the BQS family and the ABQS
framework using real-life data collected from wildlife tracking and vehicle tracking applications as well as synthetic
data, and discuss the robustness to various types of trajectories with extreme shapes.
Compared to the original BQS paper [1], this paper makes
significant extensions and technical contributions as it: 1) adds
uncertainty to the BQS formulation and derives the progressive
version of BQS, i.e. PBQS; 2) proposes the ABQS framework so
that BQS is no longer only a family of standalone algorithms but
an actual framework that can effectively and efficiently manage
a storage space and maximize the trajectory information given a
storage limit and an unknown operation duration; 3) significantly
extends the experiments to study the robustness of BQS and the
compression performances of ABQS. In addition, this version also
extends the literature review significantly, and restructures and
improves Sections 3.3 and 4 for better clarity of presentation.
The remainder of this paper is organized as follows. We survey
related work in the next section, and discuss the background and
motivate the need for a new online trajectory compression framework by describing our hardware platform and the data acquisition
process in Section 3. The BQS family is presented subsequently in
Section 4, where a discussion on how to generalize the algorithm
is also provided. We then propose the ABQS framework in Section
5. Finally, we evaluate the proposed BQS family and the ABQS
framework in Section 6, and conclude the paper.
2
R ELATED W ORK
The rapid increase in the number of GPS-enabled devices in
recent years has led to an expansion of location-based services and
applications and our increased reliance on localization technology.
One challenge that location-based applications need to overcome
is the amount of data that continuous location tracking can yield.
Efficient storage and indexing of these datasets is critical to their
success, especially for embedded and mobile devices which have
restricted energy, computational power and storage.
Several trajectory compression algorithms that offer significant
improvements in terms of data storage have been proposed in the
literature. We focus our review on lossy compression algorithms
as they provide better trade-offs between high compression ratios
and an acceptable error of the compressed trajectory.
Douglas and Peucker were amongst the first to propose an
algorithm for reducing the number of points in a digital trajectory [8]. The Douglas-Peucker algorithm starts with the first and
last points of the trajectory and repeatedly adds intermediate points
to the compressed trajectory until the maximum spatial error
of any point falls bellow a predefined tolerance. The algorithm
guarantees that the error of the compressed trajectory is within
the bounds of the target application, but due to its greedy nature,
it achieves relatively low compression ratios. The worst-case
runtime of the algorithm is O(n2 ) with n being the number
of points in the original trajectory which has been improved by
Hershberger et al to O(n log n) [9].
The disadvantage of the Douglas-Peucker algorithm is that it
runs off-line and requires the whole trajectory. This is of limited
use for modern location-aware applications that require online
processing of location data. A generic sliding-window algorithm
(conceptually similar to the one summarized in [11]) is often used
to overcome this limitation and works by compressing the original
trajectory over a moving buffer. The sliding-window algorithm can
run online, but its worst-case runtime is still O(nL) where L is
the maximum buffer size. On the other hand, multiple examples
of fast algorithms exist in the literature [12] [13] [14] [15]. These
algorithms, however, do not apply to our scenario as they only run
off-line and cannot support location-aware applications in-situ.
SQUISH [16] has achieved relatively low runtime, high compression ratios and small trajectory errors. However, its disadvantage was that it could not guarantee trajectory errors to be
within an application-specific bound. A follow up work presented
SQUISH-E [10] that provides options to both minimize trajectory
error given a compression ratio and to maximize compression
ratio given an error bound. The worst-case runtime of SQUISHE algorithms is O(n log n
λ ), where λ is the desired compression
ratio. While the compression ratio-bound flavor of SQUISH-E can
run online, the error-bound version runs offline only.
There are different disadvantages for existing algorithms that
repeatedly iterate through all points in the original trajectory.
SQUISH-E is approaching linear computational complexity for
large compression ratios, however, the compressed trajectory has
unbounded error. The STTrace algorithm [17] and the MBR
algorithm [18] represent more complex algorithms. STTrace [17]
uses estimation of speed and heading to predict the position of
the next location. MBR maintains, divides, and merges bounding
rectangles that represent the original trajectories. However both
algorithsm fall outside of capabilities of our target hardware
platform and do not suit our application scenario well.
In contrast to existing methods, our approach achieves constant
time and space complexity for each point by only considering
the most recent minimal bounding convex-hulls. We show in the
evaluation section that the compression ratios that our approach
achieves are superior to those of the related trajectory compression algorithms. Although simplistic approaches such as Dead
Reckoning [19] [20] achieve comparable runtime performance, we
show that our algorithm significantly outperforms these protocols
in compression ratio while guaranteeing an error bound.
A group of methods have been developed based on the idea of
Piece-wise Linear Regression (PLR) on time-series data, and have
achieved competitive precision and efficiency. For example, using
a bottom-up approach, Keogh et. proposed to use a “segment”based representation of time-series for similarity search and mining [21], [22]. In [23], an algorithm called SWAB successfully
reduced the time complexity of such approximation to O(n) by
combining the power of a sliding-window approach and a bottomup approach. Note that these algorithms are designed to deal
with single dimensional time-series by optimizing the normalized
error residue on the time-series values over a segment (on the y axis). Though such settings can indeed be modified and extended
to handle 2D trajectories, which concern perpendicular errors
instead, they cannot solve our problem becasue they are unable
to guarantee an error bound for the individual points in the original trajectory, though guaranteed error residue for approximated
segments is possible. For moving object tracking, it is important
to keep the “outliers” in the location history, whereas optimizing
a segment-wise error residue does not fulfil such requirement.
Meanwhile, data ageing has become a popular technique to
approximate streaming data in an amnesic fashion. The process is
often referred to as an analogy to amnesia: the older the memory
is, the less its value is and the more likely it will be forgotten. The
idea is popularized for time-series data in [24], [25] as the authors
proposed a generic framework that supports user-specific amnesic
functions. Such technique also received much attention in later
studies such as [26], [27], [28]. Again, we note that such literature
focuses on optimizing the error residue for the segments in a timeseries [24], [25], [26], [27], while in the case of location tracking
it is vital to ensure an error bound for every original location point.
3
BACKGROUND
In this section, we present the background of the study. The
hardware system architecture used in the real-life bat tracking
application is described. We also briefly introduce two existing
solutions [11], [29]. By analyzing these algorithms we provide
insights into a new algorithm is neeeded for our application. Both
algorithms will be evaluated too in the comparative evaluation.
3.1
Motivating Scenario
We employ the Camazotz mobile sensing platform [5], which
has been specifically designed to meet the stringent constraints
for weight and size when being employed for wildlife tracking.
A detailed specification of the nodes is provided in [1]. Sensor
readings and tracking data can be stored locally in external flash
storage (1 MByte) until the data can be uploaded to a base station
deployed at animal congregation areas using the short range radio
transceiver. We also place the same hardware on the dashboard of
a car to capture traces of vehicles in urban road networks.
The GPS traces collected by such platforms are often used to
analyze the mobility and the behavior of the moving object [30],
or to perform spatial queries [31]. Hence, it is most important
to gather information for the object’s major movements. The key
features from the traces are the areas where it often visits, the
route it takes to travel between its places of interests, and the
time windows in which it usually makes the travels. However,
the hardware limitations of the platform constrain its capability
to capture such information in the long term. Motivated by this
application, we propose an online trajectory compression and
management framework on such resource-constrained platforms to
reduce the data size and extend the operational time of the tracking
platform in the wild. The framework will introduce a bounded
error and discard some information for small movements, but
will capture the interesting moments when major traveling occurs.
Moreover, the framework will revisit the compressed trajectories
from time to time when additional storage space is needed. In the
process, older trajectories will be further compressed, subject to
an extended error tolerance to reflect the significance decay of the
data over time.
3.2
Existing Solutions
3.2.1 Buffered Douglas-Peucker
Douglas-Peucker (DP) [29] is a well-known algorithm for trajectory compression. In our scenario, in which the buffer size is
largely constrained due to memory space limit on the platform,
we are still able to apply DP on a smaller buffer of data. We
call it Buffered Douglas-Peucker (BDP) . The incoming points are
accumulated in the buffer until it is full, then the partial trajectory
defined by the buffered points is processed by the DP algorithm.
However, such solution has inferior compression rates mainly due
the extra points taken when the buffer is repeatedly full, preventing
a desirable high compression rate from being achieved.
An implication of BDP is that both the start and end points
in the buffer will be kept in the compressed output every time the
buffer is full, even when they can actually be discarded safely. In
the worst case scenario where the object is moving in a straight
N
line from the start to the end, this solution will use f loor( M
)+
1 points, where N is the number of total points and M is the
buffer size. In contrast, the optimal solution needs to keep only
two points. Although the overhead depends on the shape of the
trajectory and the buffer size, generally BDP takes considerably
more points than necessary, particularly for small buffer sizes.
3.2.2 Buffered Greedy Deviation
Buffered Greedy Deviation (BGD, a variation of the generic
sliding-window algorithm) represents another simplistic approach.
In this strategy whenever a point arrives, we append the point
to the end of the buffer, and do a complete calculation of the
maximum error for the trajectory segment defined by the points in
the buffer to the line defined by the start point and the end point.
If this error already exceeds the tolerance, then we keep the last
point in the compressed trajectory, clear the buffer and start a new
segment at the last point. Otherwise the process continues for the
next incoming point.
The algorithm is easy to implement and guarantees the error
tolerance, however it too has a major weakness. The compression
rate is heavily dependent on the buffer size, as it faces the same
problem as BDP. If we increase the buffer size, because the
time complexity is O(n2 ), the computational complexity would
increase drastically, which is undesirable in our scenario because
of the energy limitations. Therefore, BGD represents a significant
compromise on the performance as it has to make a direct trade-off
between time complexity and compression rate.
Clearly, a more sophisticated algorithm that can guarantee the
bounded error, process the point with low time-space complexity,
and achieve high compression rate is desired. We propose the
Bounded Quadrant System (BQS) algorithm family to address
this problem. Before delving into the details, we present some
notations and definitions to help the reader understand BQS’s
working mechanism.
3.3
Preliminaries
We provide a series of definitions as the necessary foundations for
further discussion.
Definition 3.1 (Location Point). A location point v =<
latitude, longitude, timestamp > is a tuple that records the spatiotemporal information of a location sample.
Definition 3.2 (Segment and Trajectory). A trajectory segment is a set
of location points that are taken consecutively in the temporal domain,
denoted as τ = {v1 , ..., vn }. A trajectory is a set of consecutive
segments, denoted as T = {τ1 , τ2 , ...}.
Given the definitions of segment and trajectory, we introduce
the concept of compressed trajectory with bounded error:
Definition 3.3 (Deviation). Given a trajectory segment τ =
{v1 , ..., vn }, the deviation â(τ ) is defined as the largest point-toline-segment distance from any location vi ∈ {v2 , ..., vn−1 } to
the line segment defined by v1 and vn . The trajectory deviation is
defined as the maximal segment deviation from any of its segments, as
max(â(τi )), τi ∈ T.
Deviation is a distance metric to measure the maximum error
from the compressed trajectory segment to the original trajectory.
Without loss of generality, we use point-to-line-segment distance
in this definition. Note that point-to-line distance can be easily
used within BQS too, after a few minor modifications to the lower
bound and upper bound calculations.
Definition 3.4 (Key Point). Given a buffer τ = {v1 , ..., vk }, a
deviation tolerance S
d , a new location point vk+1 , vk is a key point if
â(τ ) ≤ d and â(τ {vk+1 }) > d , where d is the error tolerance.
In other words, when a new point is sampled, the immediate
previous point is classified as key point if the new point results in
the maximum error of any point in the buffer exceeding d .
Definition 3.5 (Compressed Trajectory). Given a trajectory segment
τ = {v1 , ..., vn }, its compressed trajectory segment is defined by the
start and end location v1 and vn , and is denoted as τ 0 = {v1 , vn }.
The compressed trajectory T0 = {vi , vj , ..., vk } of T is the set of
starting and ending locations of all the segments in T, ordered by the
position of its source segment in the original trajectory.
repeated until the is exceeded again. Such process guarantees
that any trajectory segment has a smaller deviation than .
When a trajectory is turned into a compressed trajectory, the
temporal information it carries changes representation too. Instead
of having a timestamp at every location point in the original
trajectory, the compressed trajectory uses the timestamps of the
key points as the anchors to reconstruct the time sequences.
Given a trajectory segment defined by two key points vs , ve , the
reconstructed location at timestamp t (vs .t ≤ t ≤ ve .t) is:
vt =< hlat (P, vs , ve , t), hlon (P, vs , ve , t), t > .
(1)
where function h is an interpolation function that uses a distribution function P, a start value, an end value, and a timestamp at
which the value should be interpolated. P interpolates the location
at a timestamp according to a distribution. As an example, the h
and P functions for interpolating the latitude can be defined as:
P(t)
=
hlat (P, vs , ve , t)
=
t − tvs
− tvs
vs .lat + P(t) × (ve .lat − vs .lat)
tve
(2)
(3)
where P is set to reconstruct the uniform distribution. However, in
practice this function can be derived online to fit the distribution of
the actual data. For instance, an online algorithm for fitting Gaussian distribution by dynamically updating the variance and mean
can be implemented with semi-numeric algorithms described in
[32], which can be used to derive P.
As we favor an online algorithm where each point is processed
exactly once, the problem is turned into answering the following
question: does the incoming point result in a new compressed
trajectory segment or can it be represented in the previous compressed trajectory segment? To address this question, we first
provide an overview and then the details for the BQS algorithm.
Definition 3.6 (Error-bounded Trajectory). An error-bounded trajectory is a compressed trajectory with the deviation for any of its
compressed segments smaller than or equal to a given error tolerance
d . Formally: given a trajectory T = {τ1 , ..., τk }, and its compressed
trajectory T0 = {vi , vj , ..., vk }, T0 is error-bounded by d if ∀τi0 ∈ T0 ,
â(τi0 ) ≤ d .
Q2
upper bounding line
Q1
candidate
segment
upper bound
bounding
box
c1
e
lower bound
u2
p1
c2
p2
p1
u1
d=0
p1
p2
c4
d=1
s
l1
s
Q3
(b)
p1
p2
p2(e/snew)
p3
s
Q4
4
T HE BQS A LGORITHM
p3
s
(c)
lower bounding line
Fig. 2: An Example of the BQS
d=1
d=3
c3
s
(a)
p1
l2
(d)
Fig. 1: Error-bounded Compression ( = 2)
Figure 1 demonstrates the process of error-bounded trajectory
compression. Assuming that the current trajectory segment starts
from s, when adding the first point p1 (Figure 1(a)), the deviation
is 0. Hence we proceed to the next incoming point p2 (Figure
1(b)). Here the deviation is 1, which lays within the error tolerance, so the current segment can be safely represented by sp2 .
However after p3 arrives, the deviation of the segment reaches
3 > because of p1 (Figure 1(c)). Clearly p3 should not be
included in the current trajectory segment. Instead, the current
segment ends at p2 and a new segment is then started at p2 (Figure
1(d)). The new segment includes p3 and the above process is
The motivation of the framework is that we need a trajectory
management infrastructure to instantaneously and continuously
manage historical trajectory data with minimal storage space while
maintaining maximum compression error bound and capturing the
major movements of the mobile object. First we need an efficient
online trajectory compression algorithm, that yields error-bounded
results in low time and space complexity and that minimizes the
number of points taken, i.e. the Bounded Quadrant System (BQS).
A BQS is a convex hull that bounds a set of points (e.g. Figure
2). A BQS is constructed by the following steps:
1)
For a trajectory segment, we split the space into four quadrants (Q1, Q2, Q3 and Q4), with the origin set at the start
point s of the current segment, and the axes set to the UTM
(Universal Transverse Mercator) projected x and y axes.
2)
3)
4)
5)
For each quadrant, a bounding box is set for the buffered
points in that quadrant, if there are any. There can be at most
four BQS for a trajectory segment. In Figure 2 there is only
one c1 c2 c3 c4 c1 in Q1 as the trajectory has not reached other
quadrants.
We keep two bounding lines that record the smallest and
greatest angles between the x axis and the line from the
origin to any buffered point for each quadrant (su2 and sl2 ).
We have at most eight significant points in every quadrant
systems - four vertices on the bounding box, four intersection
points from the bounding lines intersecting with the bounding box. Some of the points may overlap. In this case, we
have c1 , c2 , c3 , c4, l1 , l2 , u1 , u2 .
Based on the deviations from the significant points to the
current path line, we have a group of lower bound candidates
and upper bound candidates for the maximum deviation.
From these candidates we derive a pair of lower bound and
upper bound for the maximum deviation < dlb , dub >, to
make compression decisions without the full computation of
segment deviation in most of the cases.
Here the lower bound dlb represents the smallest deviation
possible for all the points in the segment buffer to the start point,
while the upper bound dub represents the largest deviation possible
for all the points in the segment buffer to the start point. With dlb
and dub , we have three cases:
1)
2)
3)
If dlb > d , it is unnecessary to perform deviation calculation
because the deviation is guaranteed to break the tolerance, so
a new segment needs to be started.
If dub ≤ d , it is unnecessary to perform deviation calculation because the deviation is guaranteed to be smaller or
equal to the tolerance, so the current point will be included
in the current segment, i.e. no need to start a new segment.
If dlb ≤ d < dub , a deviation calculation is required
to determine whether the actual deviation is breaking the
tolerance.
Hence a pair of bounds is considered “tight” if the difference
between them is small enough that it leaves minimum room for
the error tolerance d to be in between them.
The intuition of the BQS structure is to exploit the favorable
properties of the convex hull formed by the significant points
from the bounding box and bounding lines for the buffered points
(excluding the start point). That is, with the polygons formed by
the bounding box and the bounding lines, we can derive a pair
of tight lower bound and upper bound on the deviation from the
points in the buffer to the line segment se. With such bounds,
most of the deviation calculations on the buffered points are
avoided. Instead we can determine the compression decisions by
only assessing a few significant vertices on the bounding polygon.
The splitting of the space into four quadrants is necessary as it
guarantees a few useful properties to form the error bounds.
To understand how the BQS works, we use the example with a
start point s, a few points in the buffer, and the last incoming point
e as in Figure 2. The goal is to determine whether the deviation
will be greater than the tolerance if we include the last point in the
current segment. First we present two fundamental theorems:
Theorem 4.1. Assume that a point p satisfies d(p, s) ≤ , where s is
the start point, then
dmax (p, se) ≤
(4)
regardless of the location of the end point e.
Note that the proofs for all the theorems provided in this
section are provided in [1]. This theorem supports the quick decision on an incoming point even without assessing the bounding
boxes or lines. Such a point is directly “included” in the current
segment, and the BQS structure will not be affected. It is also
important to separate these points so that BQS need only to
consider points further from the origin, because these points could
potentially result in huge bounding angles and close-to-origin
bounding boxes, rendering the BQS structure ineffective.
Theorem 4.2. Assume that the buffered points {pi } are bounded by
a rectangle in the spatial domain, with the vertices c1 , c2 , c3 , c4 , if we
denote the line defined by the start point s and the end point e as se,
then we always have:
dmax (pi , se) ≥ min{d(ci , se)} = dlb
(5)
dmax (pi , se) ≤ max{d(ci , se)} = dub
(6)
Theorems 4.1 and 4.2 show how some of the points can be
safely discarded and how the basic lower bound and upper bound
properties are derived. However, Theorem 4.2 only provides a pair
of loose bounds that can hardly avoid any deviation computation.
To obtain tighter and useful bounds, we need to introduce a few
advanced theorems. Throughout the theorem definitions we use
the following notations:
•
•
•
Corner Distances: We use dcorner = {d(ci , se)}, i ∈
{1, 2, 3, 4} to denote the distances from each vertex of the
bounding box to the current path line.
Near-far Corner Distances: We use dcorner−near =
{d(cn , se)} and dcorner−f ar = {d(cf , se)}, to denote the
distances from the nearest vertex cn and the farthest vertex cf
of the bounding box (near and far in terms of the distance to
the origin) to the current line. The nearest and farthest corner
points are determined by the quadrant the BQS is in. For
example, in Figure 2 cn = c4 and cf = c2 .
Intersection Distances: We use dintersection
=
{d(p, se)}, p ∈ {l1 , l2 , u1 , u2 } to denote the distances
from each intersection to the current path line, where li are
the intersection points of the lower angle bounding line and
the bounding box, and ui are the intersection points of the
upper angle bounding line and the bounding box.
Some advanced bounds are defined as follows:
Theorem 4.3. Given a BQS, if the line se is in the quadrant, and se
is in between the two bounding lines (θlb ≤ θs,e ≤ θub ), then we have
the following bounds on the segment’s deviation:
dmax (p, se)
max
dmax (p, se)
max
≥ dlb =
(7)
min{d(l
,
se),
d(l
,
se)}
1
2
min{d(u1 , se), d(u2 , se)}
(
dcorner−near , ∀d(s, e) < d(s, cf )
dcorner−f ar , ∀d(s, e) ≥ d(s, cf )
≤ dub =
{d
intersection
(8)
,d
corner−f ar
}
A line l is “in” the quadrant Q if the angle θl between l and
Q
Q
Q
Q
the x axis satisfies θstart ≤ θl < θend , where θstart and θend are
the angle range of the quadrant where the BQS resides. Note that
this definition is distance metric-specific. In future references we
Q
Q
assume θl satisfies θstart ≤ θl < θend if it is “in” the quadrant.
Theorem 4.4. Given a BQS, if se is in the quadrant, and is outside
the two bounding lines (θub < θs,e or θlb > θs,e ), we have the same
bounds on the segment deviation as in Theorem 4.3.
If the path line is not in the same quadrant with the BQS
because the new e moves to another quadrant, we use Theorem
4.5 to derive the bounds:
12
Theorem 4.5. Given a BQS, if the line se is not in the quadrant, the
bounds of the segment deviation are defined as:
≥
max
d(p, se)
4.1
≤
dlb =
min{d(l1 , se), d(l2 , se)}
min{d(u1 , se), d(l2 , se)}
3rd largest({dcorner })
max{dcorner } = dub
8
(9)
Meters
dmax (p, se)
2
(10)
0
0
# start&end, buffer, tolerance
if d(s, e) ≤ d then
# check for trivial decision
Decision : B ← e ∪ B
else
ub
Compute dlb
i , di using Theorems 4.3, 4.4, and 4.5
ub ← max{dub }
dlb ← max{dlb
i } and d
i
if dub ≤ d then
# low upper bound, safe to continue
Decision : B ← e ∪ B
else if dlb > d then
# high lower bound, stop immediately
Decision: Current segment stops and a new segment starts
else if dlb ≤ d < dub then # decision uncertain from bounds
d ← ComputeDeviation(B, se)
Decision: made according to d
end if
end if
Update the corresponding BQS for e
Algorithm 1: The BQS Algorithm
The algorithm starts by checking if there is a trivial decision:
by Theorem 4.1, if the incoming point e lays within the range of
of the start point, no matter where the final end point is, the point
will not result in a greater deviation than , so e is added to buffer
B (Lines 1-2). After e passes this test, it means e may result in a
greater deviation from the buffered points. So we assume that e is
the new end point, and assume the current segment is presented
by se. Now for each quadrant we have a BQS, maintaining their
respective bounding boxes and bounding lines. For each BQS, we
have a few (8 at most) significant points, identified as the four
corner points of the bounding box and the four intersection points
between the bounding lines and the bounding box. According to
the theorems we defined, we can aggregate four sets of lower
bounds and upper bounds for each quadrant, and then a global
lower bound and a global upper bound for all the quadrants (Lines
4-5). According to the global lower bound and upper bound, we
can quickly make a compression decision. If the upper bound is
smaller than , it means no buffered point will have a deviation
greater than or equal to , so the current point e is immediately
included to the current trajectory segment and the current segment
goes on (Lines 6-7). On the contrary, if the lower bound is greater
than , we are guaranteed that at least one buffered point will break
the error tolerance, so we save the current segment (without e) and
start a new segment at e (Lines 8-9). Otherwise if the tolerance
happens to be in between the lower bound and the upper bound,
an actual deviation computation is required, and decision will be
made according to the result (Lines 10-12). Finally, if the current
segment goes on, we put e into its corresponding quadrant and
update the BQS structure in that quadrant (Line 15).
Figure 3 demonstrates the lower and upper bounds as well
as the actual deviations of some randomly-chosen location points
from the real-life flying fox dataset, with d set to 5 m. The x axis
shows the indices of the points, while the solid horizontal line
20
40
60
80
100
Fig. 3: Bounds v.s. Actual Deviation
The BQS algorithm is formally described in Algorithm 1:
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
6
4
The BQS Algorithm
Input: s, e, B , d
Algorithm:
LowerBound
UpperBound
Actual Deviation
Tolerance
10
indicates the error tolerance. It is evident that in most cases the
bounds are very tight and that in more than 90% of the occasions
we can determine if a point is a key point by using only the bounds
and avoid actual deviation calculations.
A technique called data-centric rotation is used to further
tighten the bounds [1], which also shows how to generalize the
algorithm to the 3-D case and to a different error metric.
4.2
Achieving Constant Time and Space Complexity
With the pruning power introduced by the deviation bounds,
Algorithm 1 achieves excellent performance in terms of time
complexity. Its expected time complexity is α × n × c1 + (1 −
α) × n × m × c2 , where α is the pruning power, m is the
maximum buffer size, and c1 , c2 are two constants denoting the
cost of the processing of each point by the BQS structure and
by the full deviation calculation respectively. Empirical study in
Section 6 shows that α is generally greater than 0.9, meaning
the time complexity is approaching O(n) for the whole data
stream. However, the theoretical worst case time complexity is
still O(n2 ). Moreover, because we still keep a buffer for potential
deviation calculation, the worst-case space complexity is O(n).
To further reduce the complexity, we propose a more efficient
version that still utilizes the bounds but completely avoids any full
deviation calculation and any use of buffer, making the time and
space costs constant for processing a point.
The algorithm is nearly identical to Algorithm 1. The only
difference is that whenever the case dlb ≤ d < dub occurs (Line
10), a conservative approach is taken. No deviation calculation is
performed, instead we take the point and start a new trajectory
segment to avoid any computation and to eliminate the necessity
of maintaining a buffer for the points in the current segment. So
Lines 11-12 in Algorithm 1 are changed into making the “stop and
restart” decision (as in Line 9) directly without any full calculation
in Line 11. The maintenance of the buffer is not needed any more.
The Fast BQS (FBQS) algorithm takes slightly more points
than Algorithm 1 in the compression, reducing the compression
rate by a small margin. However, the simplification on the time and
space complexity is significant. The fast BQS algorithm achieves
constant complexity in both time and space for processing a point.
Equivalently, time and space complexity are O(n) and O(1) for
the whole data stream.
The time complexity is only introduced by assessing and
updating a few key variables, i.e. bounding lines, intersection
points and corner points. We can now arrive at the compression
decision by keeping only the significant BQS points of the number
c ≤ 32 ( 4 corner points and 4 intersection points at most for each
quadrant) for the entire algorithm.
When the buffer size is unconstrained, the three algorithms
Buffered Douglas-Peucker (BDP), Buffered Greedy Deviation
(BGD) and Fast BQS (FBQS) have the following worst-case time
and space complexity:
Procedure: abqs()
Input: p
TABLE 1: Worst-case Complexity
FBQS
O(n)
O(1)
Time
Space
5 T RAJECTORY
BQS
BDP
O(n2 )
O(n)
BGD
O(n2 )
O(n)
U NCERTAINTY
AND
A MNESIC
Next we propose an online framework that compresses the historical data in an “amnesic” way to maximize the utilization of
storage capacity. This scenario poses two challenges:
1)
2)
how to compress the already compressed trajectories by
extending and adhering to an increased error tolerance?
how to design the ageing procedure so that trajectories
compressed with different error tolerances could be updated efficiently?
Fig. 4: A compressed trajectory has a bounded uncertainty (error
tolerance )
Q2
upper bounding line
lower bound
upper bound
Q1
candidate
segment
e
p1
c1
u2
c2
bounding
box
p2
u1
l2
c4
c3
l1
lower bounding line
s
Q3
uncertainty
Shared Variables: S, I, k, , m, N
Q4
Fig. 5: BQS on a compressed trajectory that has a bounded uncertainty
To answer the first question, first we observe that after the
first-pass compression with the error tolerance , a trajectory
is transformed into a simplified representation with a bounded
uncertainty, as illustrated in Figure 4. Putting this representation
into the BQS formulation (e.g. Figure 2), we can derive an
uncertain form of the BQS in Figure 5.
In this uncertain form, we use black dots to present trajectory
points (key points with uncertainty from previously compressed
trajectories in this case). The significant points on the bounding
box or the bounding lines are still determined by the key points
themselves, e.g. l1 and u1 . However in addition to the key
points, we also need to consider the uncertainty from the previous
compression. The information for the discarded points from the
original trajectory has been completely lost, and the previous
error tolerance is the only information we could reconstruct such
uncertainty. In this figure we use dash-dotted lines to illustrate
such uncertainty. Note that the uncertainty here is twofold. Firstly,
because the uncertainty around the start point s and the end point
e (for points around both ends may have been discarded from
the previous compression), the current candidate segment se is
uncertain. Secondly, each compressed segment in c1 c2 c3 c4 c1 has
its own bounded uncertainty, thus in the new form every significant
point has an equivalent uncertainty and the bounding polygon
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
if kIk > 0 then
i ← I[0].e + 1
if I[0].a == 0 then
update index(0, −1, i)
else
update index(0, i, i)
end if
else
i←0
update index(0, i, i)
end if
S[i] ← p
amnesic sinking()
# main entry
# index not empty
# find location for new point
# youngest gen has age 0
# update end index only
# otherwise insert age 0 segment
# dealing with completely new storage
# store new point
# check if ageing is needed
Procedure: amnesic sinking()
1: f ← trigger()
2: while f do
3:
if kIk == 1 then
4:
sbuf ← 0
5:
else
6:
sbuf ← I[1].e + 1
7:
end if
# check trigger
# if ageing is triggered
# find dest loc for results from ageing
# dest is 0 when buffer has age 0 only
# otherwise append to next age
# compress youngest generation
8:
compress(I[0].s, I[0].e, I[0].a, sbuf )
9:
f ← trigger()
# check trigger again
10: end while
Algorithm 2: The Amnesic BQS Procedure
u1 u2 l2 l1 c4 u1 is extended too and its corners are turned into round
corners with radius equal to the previous tolerance.
Recall that in the first-pass compression in Figure 2, we
use d(u2 , se) as the upper bound and d(c2 , se) as the lower
bound for the maximum deviation. As we examine the uncertain
form, we find that the lower bound and upper bound can be reused, with simple modifications. In Figure 5, where d(u2 , se)
represents the upper bound without considering uncertainty, it
is easy to prove that dmax (p, se) ≤ d(u2 , se) + 2, and
dmax (p, se) ≥ d(u2 , se) − 2. The uncertainty of se and the
bounding box each introduces uncertainty of to both bounds.
As a generalization of the specific case in Figure 5, we derive
Theorem 5.1 to support the application of BQS on compressed
trajectories with any existing error tolerance (uncertainty).
Theorem 5.1. Given a set of compressed trajectories with a previous
lb , df
ub for
error tolerance , if we obtain the lower and upper bounds df
the maximum deviation from the BQS constructed from the key points
only and without considering the uncertainty, we then have the lower
and upper bounds on the new maximum error dlb , dub as
lb − 2, dub = df
ub + 2
dlb = df
(11)
Proof. The furthest distance possible from the lines u1 u2 and c2 u2
is at the round corner at u2 which will always satisfy d(p, se) ≤
d(u2 , se) + . Then since se has an uncertainty of around itself, we
have Theorem 5.1.
We devise a new Progressive BQS (PBQS) algorithm that supports an existing tolerance, with the main contents in Algorithm
1. With Theorem 5.1, the bounds from Line 5 is now updated by
dlb ← dlb − 2p , dub ← dub + 2p , where p is the previous error
tolerance on the compressed trajectories, passed as an additional
argument for the PBQS algorithm besides the new tolerance d
(d > p ). Interestingly, the algorithm now obtains a progressive
property. That is, once the existing error tolerance is known,
the next compression is independent of any previously applied
Procedure: compress()
Input: ssrc , esrc , a, sdest
# src start&end, age, dest start
1: X ← S[ssrc : esrc ]
2: prev ← ma+1
3: X 0 ← pbqs(X, prev , )
# compute new tolerance
# perform compression
# store results to dest location
4: S[sdest : sdest + kX 0 k − 1] ← X 0
5: update index(a, −1, −1)
# remove youngest gen from index
# update index for ageing results
6: update index(a + 1, sdest , sdest + kX 0 k − 1)
Procedure: trigger()
# ageing triggering
1: if kIk > 0 then
2:
3:
4:
5:
6:
7:
8:
# full with age 0 points
if I[0].a == 0 && I[0].e == N − 1 then
return true
# trigger threshold reached
else if I[0].a > 0 && I[0].e > N − k − I[0].a then
return true
end if
end if
return false
Procedure: update index()
Input: a, s, e
# age, start&end locations
1: if index for age a exists then
2:
update index of age a
3: else
4:
insert index for age a
5: end if
Algorithm 2: The Amnesic sinking Algorithm (Continued)
compression on the trajectory data. The compressed segments
resulted from PBQS can again be the input of the next compression
pass that has a greater error tolerance than the current. The
implication here is that theoretically the tracking node’s storage
space will support operation of an indefinite period without hard
data loss (data overwritten due to space limit). Instead, it will
introduce graceful degradation to the aged data over time.
Now to address the second challenge, the PBQS, which aims
at making compression decisions for individual segments, evolves
to a sophisticated framework called Amnesic BQS (ABQS) that
manages a certain amount of storage space and seeks to maximize
the trajectory information when the number of points far exceeds
the space limit and the exact number of points is unknown at the
begining. The framework is designed to determine when and how
the Progressive BQS is applied on the aged data, and is designed
to optimize the storage arrangements for the co-existence and
handling of trajectories with difference ages and error tolerances.
The framework utilizes a few ideas:
1) we should store as many “younger generations” as possible to improve the precision of more recent data.
2) as the storage space runs out, we always try to compress
the youngest generation as that results in the least precision loss and storage access.
3) mechanisms are needed so that a “younger generation”
can always be turned to an “older generation” with meaningful compressions (to a reduced number of points).
4) in the process of adding new points and making further
compressions, access to the storage should be minimized
when possible for energy saving purposes [28].
In addition, for this framework we still face the limits in memory
and computation power of the platform, so the framework still
needs to be computationally efficient. Hence we propose ABQS
in Algorithm 2.
There are a few shared variables: S is the entire storage space
under ABQS’ management, S[i] means the ith slot of the storage.
I is a list that keeps track of storage segmentations for different
ages, with the entries in ascending order by age (ages without any
point of that age will not appear in the index). Each entry in I
has three properties, start location s, end location e and age a.
For example, I[0] =< 0, 12, 1 > means currently the youngest
generation in the storage is of age 1 (this is the case when all age
0 points in the storage are turned to age 1), and is occupying the
0th to 12th (inclusive) slots in the storage. k is a fixed number to
set the minimum buffer size for the age 0 trajectory points, which
also is used to determine the triggering decision for the ageing
procedure. is the starting error, m is the error multiplier when
data age increases, and N is the storage space limit. The whole
algorithm is described by five procedures in Algorithm 2. Note
that in Algorithm 2 the indices start from 0 and are inclusive.
abqs is the entry point of the procedure, i.e. whenever an initial
data point with the lowest error tolerance (could be 0) is passed
through to the framework, abqs will process it and at the same
time maintain the structure of the entire storage. Thus abqs’ main
responsibilities are to find a proper storage space for the new point,
to update the index structure, and to invoke the amnesic sinking
procedure for further compression when necessary.
amnesic sinking is the main ageing procedure for further
compression and freeing up storage spaces. We call it “amnesic
sinking” because it guarantees that older data that has greater
tolerances will sink to the “bottom” of the storage, while younger
data will remain closer to the top. From bottom to top, the tolerances decrease monotonically. The procedure iteratively checks a
trigger flag f which indicates whether it is necessary to perform
a further compression on aged data, and then invokes compress
when needed.
compress performs progressive compression with an extended
error tolerance from an existing one. The inputs ssrc , esrc specify
the indices of the data segment to be compressed in the storage,
a specifies the current age of the data to be further compressed,
and sdest specifies the location in the storage where the further
compressed data will be stored. Line 2 shows how the tolerance is
updated for each generation. Lines 3 and 4 compress and age the
existing trajectories to an older generation. Lines 5 and 6 remove
the index for the generation being processed, and add the index for
the resulted generation after the processing. Note that here Lines
1-4 are illustrated as a batch/block update but in practice it is
implemented in an online/inline style so that the space complexity
is still constant. That means the procedure will process a point at
a time, similar to the way standalone BQS operates, and store the
results in the space originally taken by the trajectory before the
compression to avoid unnecessary copy and move in the storage.
trigger checks whether compress is needed for the youngest
generation in the storage. The decision is made to fulfil the
requirements on the space arrangement that the data points of
the youngest generation with age a cannot go beyond the limit
of I[0].e > N − k − I[0].a (Line 4) to maintain meaningful
compression results for each generation. k is used to reserve
enough space for incoming age 0 points. For example, if we
assume that the points passed to the ABQS framework have an
initial error of 2m for the age 0 points, we can set this k to be
the average number of points included in an age 1 segment with
5m tolerance from empirical results. The effect of this parameter
is to reduce occurrences in which the points are kept in the aged
results not because the tolerance is reached but because the end of
the youngest generation’s segment is reached.
I[0].e > N − k − I[0].a follows the same rationale. The
reserved space for younger generations is increased when the age
step 0
Adding age 0 trajectory
stream
16,2
15,2
14,2
13,2
12,2
11,2
10,2
9,2
8,2
7,2
6,2
5,2
4,2
3,2
1,2
step 1
16-1
10
step 2
17,2
16-1
10
step 3
31-17
10
16-1
10
step 4
31-17
10
16-1
10
step 5
100-1
50
step 6
101,2
100-1
50
step 7
115-100
10
100-1
50
step 8
2,2
Age 1 ageing procedure
triggered (storage full)
Adding age 0 trajectory
stream
31,2
30,2
29,2
28,2
27,2
26,2
25,2
24,2
23,2
22,2
21,2
20,2
19,2
18,2
Age 1 ageing procedure
triggered (storage full)
...
100-92
10
91-82
10
81-71
10
70-59
10
58-46
10
45-32
10
Age 2 ageing procedure triggered
(threshold reached)
Adding age 0 trajectory
stream
115,2
114,2
113,2
112,2
111,2
110,2
109,2
108,2
107,2
106,2
105,2
104,2
103,2
102,2
Age 1 ageing procedure
triggered (storage full)
Age 0
Age 1
Age 2
k threshold
trigger threshold
Fig. 6: Amnesic sinking: The Inline Ageing Procedure
of the currently youngest generation increases, so that there will
be at least three points for any younger generation when they
trigger an ageing procedure and receive a further compression
(performing compression on two or less points will not result in
any compression). If the youngest generation is of age a, then
ABQS needs to reserve one slot for the immediately younger
generation a − 1, because for age a − 1 there will be at least
2 points from a − 2, so this single served slot guarantees the
meaningful compression for this generation. Similarly, we need 1
slot reserved for a−2. Recursively, we can get the total number of
reserved slots for age a to be a−1 and consequently the maximum
index for age a as N − k − 1 − (a − 1) = N − k − a, because
the k slots are also reserved for the incoming age 0 points.
update index maintains the index structure which records the
segmentation of storage for data points with different ages. The
procedure does a few operations: in Lines 2-4 it attempts to find
an existing index entry for the specified age a, and update its start
and end locations to s and e on success. When a new entry is
needed for an unrecorded age, Lines 6,7, and 10 try to insert the
entry and maintain the ascending order by age of the index entries.
There are several advantages for the ABQS framework. First
there is no theoretical limit on the maximum operational time
(without hard data loss) of the tracking device, as the ABQS
adaptively manages the storage space and trades precision off
storage space. Second the entire procedure only uses little memory
space (memory for a standard BQS plus three integers for each
index entry). Third the compression error for each generation is
still known and guaranteed.
We illustrate this algorithm in Figure 6. Here the initial error
tolerance is set to 2 and the multiplier m to 5. The empty cells
represent available storage spaces, and the cells with color and
numbers mean occupied storage space. The storage is used from
right (bottom) to left (top). Each occupied cell is labeled with two
numbers < data point id range, error tolerance > (ranges
start with 1 and are inclusive), and the background colors of the
cell indicate the age of the data. For example, in the third step,
the right most cell has < 16 − 1, 10 > in it as well as a light
blue background, indicating that the storage block now stores age
1 compressed segments covering the original points 1 − 16, with
an error tolerance of 10. The blue and red dashed lines are the
k threshold and the trigger threshold defined by I[0].e > N −
k − I[0].a (Line 4 in trigger). Step 1 means when the space is
filled with age 0 data then an ageing process is invoked which
compresses data points 1 − 16 into a segment with error tolerance
10 and stores the results to the bottom of the storage. In step 3 the
storage is full again and age 0 data points are compressed again
to the adjacent storage to < 16 − 1, 10 >. Step 5 illustrates that
when the age 1 compressed segments take over a great part of
the storage and leave insufficient space for younger data, they are
further compressed. In step 6 we see that because the youngest
generation becomes age 2, the trigger threshold moves rightwards,
reserving more spaces for age 1 and age 0, whereas in step 8 the
trigger threshold moves back to the k threshold because age 1 is
now the youngest generation.
6
E XPERIMENTS
In this section we evaluate the performance of the proposed BQS
family and ABQS framework.
6.1
Dataset
We use three types of data, namely the flying fox (bat) dataset, the
vehicle dataset, and the synthetic dataset. The two real-life datasets
0.16
BQS
FBQS
BDP
BGD
DP
Compression Rate
0.12
0.12
0.10
0.08
0.06
0.10
0.08
0.06
0.04
0.04
0.02
0.02
0.00
2
4
6
8
10
12
14
Error Tolerance (m)
(a) Bat Tracking Data
16
18
BQS
FBQS
BDP
BGD
DP
0.14
Compression Rate
0.14
20
0.00
5
10
15
20
25
30
35
40
45
Error Tolerance (m)
(b) Vehicle Tracking Data
Fig. 7: Comparison of Compression Rate on Real-life Datasets (The
lower the better)
Note that there are couple of differences between the two
datasets. The vehicle dataset shows larger scales in terms of travel
distances well as moving speed. For instance, the length of a
car trip varies from a few kilometers to 1, 000 km while a trip
for flying-foxes are usually around 10 km. The car can travel
constantly at 100 km/h on a highway or 60 km/h on common
roads, while the common and maximum continuous flying speeds
for a flying-fox are approximately 35 km/h and 50 km/h. In
regard to these differences, the two datasets are evaluated with
different ranges of error tolerance. With a much greater spatial
scale of the movements, the error tolerance used for the vehicle
dataset is generally greater.
The vehicle dataset also shows more consistency in the heading angles due to the physical constraints of the road networks. On
the contrary, the bats’s movements are unconstrained in the 3-D
space, so their turns tend to be more arbitrary. We argue that by
performing extensive experiments on both datasets, the robustness
of BQS and ABQS is demonstrated.
The synthetic dataset consists of data generated by a statistical model that anchors patterns from real-life data, special
trajectories that follow certain shapes. The comparison between
Dead Reckoning [19] and Fast BQS (FBQS) is conducted on
this dataset. This is because continuous high-frequency samples
with speed readings are required to implement DR in an errorbounded setting, while such data is lacking in the real-life datasets.
The model uses an event-based correlated random walk model to
simulate the movement of the object. In the simulation, waiting
events and moving events are executed alternately. The object
stays at its previous location during a waiting event, and it
moves in a randomly selected speed and turning angle for a
randomly selected time. Note that the speed follows the empirical
distribution of speed, the turning angle is drawn from the von Mise
distribution [33], while the move time is exponentially distributed,
corresponding to the Poisson process. The trajectories are bounded
by a rectangular area of 10 km×10 km, and the speed and turning
angle follow approximately the distributions of the bat data. A
total of 30, 000 points are generated by the model. The synthetic
dataset also contains trajectories of special shapes to explicitly test
the robustness of the proposed methods.
50
Experimental Settings
The evaluation is conducted on a desktop computer, however the
extremely low space and time complexity of FBQS makes it plausible to implement the algorithms on the platform aforementioned
in Section 3 (32 KBytes ROM, 4 KBytes RAM). In particular,
if we look into the FBQS algorithm, we only need tiny memory
space to store at most 32 points besides the program image itself
(4 corner points and 4 intersection points for each quadrant).
Two main performance indicators, namely compression rate
and pruning power are tested on the real-life datasets to evaluate
compressed
the BQS family. We define compression rate as NN original
compressed
where N
is the number of points after compression,
and N original is the number of points in the original trajectory.
computed
Pruning power is defined as 1 − N N total , where N computed
total
and N
are the number of full deviation calculations and the
number of total points respectively.
For compression rate, we perform comprehensive comparative
study to show BQS’s superiority over the other three methods,
namely Buffered-DouglasPeucker (BDP), Buffered-Greedy (BGD)
and Douglas-Peucker (DP). DR is compared against FBQS on
the synthetic dataset. For buffer-dependent algorithms, we set the
buffer size to be 32 data points, the same as the memory space
needed by the FBQS algorithm to hold the significant points.
To intuitively demonstrate the advantage of FBQS in compression rate, we provide comparison showing the number of
points taken by the FBQS algorithm and the DR algorithm on
the synthetic dataset. We also show the estimated operational time
of tracking devices based on such compression rate. Finally, we
study comparatively the actual run time efficiency of FBQS.
The ABQS framework is evaluated with extensive experiments, for which the details are given in Section 6.3.7.
Pruning Power
0.16
6.2
1.0
1.0
0.8
0.8
Pruning Power
comprise of 138, 798 GPS samples, collected by 6 Camazotz
nodes (five on bats and one on a vehicle). The total travel
distances for the bat dataset and vehicle dataset are 7, 206 km and
1, 187 km respectively. The tracking periods spanned six months
and two weeks for the bat and vehicle datasets respectively.
0.6
0.4
0.2
0.0
2
0.6
0.4
0.2
4
6
8
10
12
14
16
18
20
Error Tolerance (m)
(a) Bat Tracking Data
0.0
5
10
15
20
25
30
35
40
45
50
Error Tolerance (m)
(b) Vehicle Tracking Data
Fig. 8: Pruning Power of the BQS Algorithm (The higher the better)
6.3
Compression Algorithm
6.3.1 Compression Rate on Real-life Data
Compression rate is a key performance indicator for trajectory
compression algorithms. Here we conduct tests on the two real-life
datasets. We compare the performance of five algorithms, namely
BQS, FBQS, BDP, BGD and DP. All of the algorithms give errorbounded results. The former four are online algorithms, and the
last one is offline. The compression rates are illustrated in Figures
7(a) and 7(b).
Evidently, BQS achieves the highest compression rate among
the five algorithms, while BDP and BGD constantly use approximately 30% to 50% more points than BQS does. FBQS’s
compression rates swing between BQS’s and DP’s, showing the
second best overal performance. BDP has the worst performance
overall as it inherits the weaknesses from both DP and windowbased approaches. BGD’s performance is generally in between DP
1600
7000
1400
6000
1200
5000
1000
No. Points
8000
4000
3000
2000
800
6.3.4
100
BQS
FBQS
DouglasPeucker
6
8
10
12
14
16
18
0.4 0.4 0.4
1.0 1.0 0.6
one-way
commute
(b) No.Points Used on Synthetic Data
Fig. 9: Shape of Synthetic Data and Comparison of Number of Points
Used (The lower the better)
Pruning Power
Pruning power determines how efficient BQS and FBQS are.
In FBQS, if the relation between the bounds and the deviation
tolerance is deterministic, FBQS generates a lossless result as
BQS. If it is uncertain, then FBQS will take a point regardless
of the actual deviation. The pruning power reflects how often the
relation is deterministic, and it indicates how many extra points
could be taken by the approximate algorithm. With high pruning
power, the overhead of FBQS will be small. Here we investigate
the pruning power of BQS in this subsection.
Figures 8(a) and 8(b) show the pruning power achieved by
BQS on both datasets. The sensitivity of the algorithm to the
error tolerance or to the shape of the trajectories appears low,
4.0 4.0 5.8
spiral
-60
-60 -40 -20
0
20
40
60
(b) Spiral Trajectory
Fig. 10: Robustness of BQS
20
Error Tolerance (m)
20
-40
20
(a) Extreme Cases
4
40
-20
zigzag
2
60
0
40
0
600
1000 2000 3000 4000 5000 6000 7000 8000
(a) Synthetic Dataset (unit:m)
98 98 98
60
0
0
Robustness to Trajectory Shapes
80
200
0
6.3.2
6.3.3 Comparison with Dead Reckoning on Synthetic Data
In Figure 9(a) we show the simulated trajectories from our statistical model. Visibly the trajectories show little physical constraint
and considerable variety in heading and turning angles. On this
dataset we study the performance comparison because on this
dataset we are able to simulate the tracking node closely with
high frequency sampling. Hence FBQS is used as a light-weight
setup to fit such online environment.
We show in Figure 9(b) the numbers of points taken after the
compression of 30, 000 points under different error tolerances.
With smaller such as 2 m, DR uses 1, 547 points compared to
1, 122 for FBQS, indicating that DR needs 37% more points. As
grows, DR’s performance tends to slowly approach FBQS in absolute numbers yet the difference ratio becomes more significant.
At 20 m error tolerance, FBQS only takes 341 points while DR
uses 493 points, the difference ratio to FBQS is around 45%.
Evidently, FBQS has achieved excellent compression rate
compared to other existing online algorithms.
400
1000
-1000
-1000
FBQS
DR
as the pruning power generally stays above 90% for most of
the tolerance values on both datasets. This means approximately
only 10% more points will be taken in the Fast BQS algorithm
compared to the original BQS algorithm. The running values in
Figure 3 also support this observation.
BQS shows higher pruning power on the car dataset than on
the bat dataset. The higher pruning power on the vehicle data
is a result of the physical constraints of the road networks, preventing abrupt turning and deviations, and making the trajectories
smoother. Naturally the pruning power will be higher as a result of
the higher regularity in the data’s spatio-temporal characteristics.
Compression Rate(%)
and BDP, but it still suffers from the excessive points taken when
the buffer is full.
Comparing the two figures, it is worth noting that all the
algorithms perform generally better on the bat data. Take the
results at 10 m error tolerance from both figures for example, on
the bat data the best and worst compression rates reach 3.9% and
6.3% respectively, while on the vehicle data the corresponding
figures are 5.4% and 7.7%. This may seem to contradict the
results of the pruning power. However, it is in fact reasonable
because bats perform stays as well as small movement around
certain locations, making those points easily discardable. Hence
the room for compression is larger for the bat tracking data given
the same error tolerance.
On the bat data, with 10 m error tolerance, BQS and FBQS
achieve compression rates of 3.9% and 4.1% respectively. DP,
as an offline algorithm that runs in O(nlogn) time, yields a
worse compression rate than FBQS at 4.6%. Despite having
poorer worst-case complexities, BDP and BGD also obtain worse
compression rates than BQS and FBQS do at 6.3% and 5.8%
respectively. At this tolerance, the offline DP algorithm uses
approximately 20% and 10% more points than the online BQS
and FBQS do, respectively. Furthermore, for online algorithms
with 20 m tolerance, FBQS (2.7%) improves BDP (5.1%) and
BGD (4.9%) by 47% and 45% respectively.
The results on the vehicle data show very similar trends of
the algorithms’ compression rate curves. Interestingly, with this
dataset, because the pruning power is in most of the cases around
and above 0.95 as demonstrated in 8(b), the compression rate of
FBQS is remarkably close to BQS’. For instance, at 20 m, 30 m
and upwards, the difference between the two is smaller than 1%.
This observation supports our aforementioned argument that the
bounds of the original BQS are so effective that the number of
extra points taken by FBQS is insignificant.
In this experiment we examine BQS’ robustness to trajectory
shapes. Though in previous experiments we have used trajectories
with a variety of characteristics such as speed, spatial range and
turning angles, and have verified that BQS achieves competitive
performances on those datasets, here we synthesize a few more
with extreme shapes to further examine the robustness of BQS.
We use four synthetic datasets:
•
•
•
•
zigzag: the trajectory travels in a zigzag fashion.
x =< 1, 0.5, 3, 0.5....2N − 2, 0.5, 2N, 0.5 >, y =<
0.5, 2, 0.5, 4, ....0.5, 2N − 2, 0.5, 2N >.
one-way: a straight line x =< 1, 2, ..., N >, y =<
0.5, 0.5..., 0.5 >.
commute: moving back and forth on a straight line x =<
1, 2, ..., n, n − 1, ...1, 2, ..., n, ... >, y =< 0.5, 0.5..., 0.5 >.
spiral: the Achimedean spiral defined as ρ = 1 + 2θ.
For each synthetic dataset we sample 500 points with the above
strategies and display the compression rates of BQS, FBQS and
Douglas-Peucker (DP) in Figure 10(a), with the tolerance = 10.
The results of BQS and FBQS are almost identical, suggesting
that the conservative decision making in the FBQS introduces
little overhead for even the most extreme shapes. The BQS family
achieves comparable results with DP on all four datasets. On the
first two they are all 98% and 0.4%. On the third DP achieves
slight better result 0.6% versus 1.0% of BQS, but on the fourth
BQS has a much better compression rate of 4.0%, in comparison
with 5.8% from DP. Figure 10(b) demonstrates the Achimedean
Spiral with 500 samples in blue dots. The red stars and the dashed
line illustrate the key points selected by BQS.
6.3.5 Effect on Operational Time of Tracking Device
Next we investigate how different online algorithms affect the
maximum operational time of the targeted device. This operational
time indicates how long the device can keep records of the
locations before offloading to a server, without data loss.
In the real-life application, the nodes also store other sensor
information such as acceleration, heading, temperature, humidity,
energy profiling, sampled at much higher frequencies due to their
relatively low energy cost. We assume that of the 1MBytes storage,
GPS traces can use up to 50KBytes, and that the sampling rate of
GPS is 1 sample per minute. Each GPS sample requires at least
12 bytes storage (latitude, longitude, timestamp). For the error
tolerance, we use 10 meters as it is reasonable for both animal
tracking and vehicle tracking. The average compression rate at
10 meters for both datasets is used for the algorithms. For the
DR algorithm, we assume it uses 39% more points than FBQS as
shown in Figure 9(b) at the same tolerance.
Given the set up, the compression rate and estimated operational time without data loss for each algorithm are listed in Table
2. We can see a maximum 36% improvement from FBQS over the
existing methods (60 v.s. 44), and a maximum 41% improvement
from BQS (62 v.s. 44).
TABLE 2: Estimated Operational Time
Compression rate
Time (days)
BQS
4.8%
62
FBQS
5.0%
60
BDP
6.65%
45
BGD
6.75%
44
DR
6.65%
45
6.3.6 Run Time Efficiency
We compared the run time efficiency of FBQS, BDP and BGD.
87,704 points from the empirical traces are used as the test data.
The error tolerance is set to 10 meters. For BDP and BGD, to
minimize the effect of the buffer size, we report their performances
with different buffer sizes, as in Table 3:
TABLE 3: Performance Comparison with different buffer sizes
Buffer size (points)
Compression rate
Run time (ms)
FBQS
BDP
BGD
FBQS
BDP
BGD
32
3.6%
6.8%
6%
99
76
182
64
—
6.7%
4.8%
—
101
285
128
—
5.4%
4.6%
—
163
446
256
—
4.9%
4.4%
—
292
628
There are two notice-able advantages of FBQS from the
comparison. Firstly, both the compression rate and the run time
efficiency of FBQS algorithm are stable, independent of the buffer
size setting. Secondly, it offers competitive run time efficiency
while providing leading compression rate. The only case when
BDP is able to outperform FBQS in run time efficiency is when
the buffer is set to 32, where BDP has a far worse compression
rate (89% more points).
6.3.7 Evaluation of the ABQS Framework
We evaluate ABQS with the results presented in Figure 11,
in a series of experiments in comparison with FBQS and DP.
Here instead of using the bat and vehicle datasets separately,
we concatenate the data points from both datasets and put them
into a unified timeframe. This introduces greater variations of the
dynamics in the trajectory dataset, while the characteristics of both
bats and vehicle are still inherited. The error multiplier m is set
to 2.5 (as PBQS introduces 2prev uncertainty in both the lower
and the upper bounds, this multiplier leaves 0.5prev margin for
pruning). We mainly use two error metrics: deviation (maximum
perpendicular distance as in Definition 3.3) and time-synchronized
error (as in Figure 5(b) in [34]).
In Figures 11(a) and 11(b), we fix the error tolerance = 20
(m) and the number of points N = 80, 000, and study how the
errors change with the storage/data ratio r. This ratio represent
the situation where the location data far exceeds the storage limit
in volume. The smaller r is, the greater challenge it is to control
the compression error. For ABQS, it will continuously tradeoff
precision for space, while for FBQS and DP, hard data loss
(overwriting) is inevitable after the storage is full. We observe
that in both error metrics, ABQS’s performances are significantly
better than FBQS and DP. For example, as r changes from 0.5% to
5%, the synchronized error and deviation of ABQS decreased from
15km to 1km, and from 15km to 30m. On the other hand, FBQS
and DP both have much worse results which have not shown
clear reduction of compression error even when storage space is
increased. For synchronized error their numbers drop from 40km
merely to 20km, and for deviation the error lingers around 250km
with fluctuation. Note that the standard deviations of the dataset
are 30km and 85km on the x and y axis, hence a synchronized
error of 40km and a deviation of 150km basically render the
compression results useless for FBQS and DP. Such huge error
is mainly introduced by the hard data loss over time. However,
with ABQS, it does not only guarantee a decreasing deviation
on each segment, but also obtains a reasonable and decreasing
synchronized error as storage space grows.
Figures 11(c) and 11(d) demonstrate how the error tolerance
influences the compression results. Here we set N = 80, 000 and
the storage/data ratio r = 3%, and change from 10m to 80m. We
find that under both metrics, none of the algorithms is particularly
sensitive to this parameter. Take Figure 11(c) for example, FBQS
and DP again have almost identical results (45km to 25km), while
ABQS’ number is varying between 1.5km to 2km. Again in both
results we see that ABQS has far better performance. This implies
that ABQS’ performance is rather robust to the starting error
tolerance, as it automatically increases the tolerance on demand.
For FBQS and DP, the sychronized error decreased significantly
when the tolerance changes from 10 to 30, this is because with a
smaller tolerance the number of points after compression is greater
and hence the ratio of the hard data loss is greater. Hence more
information is lost even when the tolerance is smaller. After 30,
the sychronized error stablizes around 25km. It is possible that
when continues to grow, the gap between ABQS, FBQS, and DP
will close (when trajectories size is smaller than storage after first
compression pass), however in practice it is unrealistic to set an
appropriate value for when the operational time is unknown.
Figures 11(e) and 11(f) study the effect of data volume. When
= 20 and r = 3%, we change the number of points to be
compressed from 10,000 to 80,000, and examine the errors. It is
evident that ABQS has very consistent performance regardless of
data size (around 1.5km synchronized error and 200m deviation).
However, with hard data losses, both FBQS and DP perform
poorly, with consistently over 15 times greater synchronized errors
102
0.5 1.0
101
5.0
0.5 1.0
(a)
2.0
3.0
4.0
Storage/data Ratio (%)
ABQS
FBQS
DP
102
20
30
40
50
60
No. Points (thousand)
(e)
70
104
101
20
30
40
50
60
No. Points (thousand)
70
80
30
40
50
60
Error Tolerance (m)
70
103
80
10
104
ABQS
FBQS
DP
103
102
10
20
30
40
50
60
No. Points (thousand)
20
70
80
(g)
30
40
50
60
Error Tolerance (m)
70
80
(d)
106
105
(f)
ABQS
FBQS
DP
104
102
20
(c)
ABQS
FBQS
DP
103
10
103
10
102
80
104
5.0
105
103
ABQS
FBQS
DP
(b)
106
104
105
Deviation (m)
ABQS
FBQS
DP
102
2.0
3.0
4.0
Storage/data Ratio (%)
105
10
103
106
Time-synchronized Error (m)
ABQS
FBQS
DP
104
Time-synchronized Error with Decay (m)
Deviation (m)
103
Time-synchronized Error (m)
105
104
Time-synchronized Error (m)
105
106
Deviation (m)
Time-synchronized Error (m)
105
105
104
103
ABQS, µ =1433
FBQS, µ =49354
DP, µ =46065
102
101
0
10
20
30
40
50
Index of point (thousand)
60
70
(h)
Fig. 11: Performance of ABQS for the Combined Datasets
and over 400 times greater deviation. These results show that
ABQS is robust to the data size.
Figure 11(g) shows the changes of the synchronized error
with a linear significant decay function (similar to the “amnesic
function” in [24]) that reduces the importance of points from 1.0
to 0.2 from younger to older data (eei i=1:N = ei ( 0.8i
N + 0.2) ),
as a simplistic demonstration of the significance function’s effect
in the relative performance for each method. The results verifiy
that though the errors induced by older points (which are likely
overwritten in FBQS and DP) are regarded less important, ABQS
still wins by a large margin.
Finally, Figure 11(h) presents how the synchronized errors are
distributed over time for the 80, 000-point long empirical traces.
Here the error is calculated for each individual point in the original
trajectory, and is smoothed with a sliding window of 1,000 to
improve visual clarity. The mean errors for ABQS, FBQS and
DP are 1,433m, 49km and 46km, showing a clear advantage of
ABQS. We also notice a clear cutoff point for FBQS and DP
around 27,000 points. Before this point FBQS and DP have better
precision than ABQS because the trajectories they store have gone
through one compression. However, after the cutoff point where
hard data loss begins to occur, the error jumped dramatically from
80m to 100km. Contrarily, ABQS’ error remains much more stable
during for the whole course. Indeed changing the initial error
tolerance for FBQS and DP may as well change the location of
the cutoff point slightly for FBQS and ABQS, but when the total
number of points is unknown when deployed, ABQS has a clear
advantage as it does not require a precise estimation of the final
data size, or a carefully chosen error tolerance.
7
C ONCLUSION
We present a family of online trajectory compression algorithms called BQS and an amnesic compression framework called
ABQS for resourced-constrained environments. We first propose
a convex-hull bounding structure and then show tight bounds can
be derived from it so that compression decisions will be efficiently
determined without actual deviation calculations. A light version
of the algorithm is hence proposed for the most constrained
computation environments. The standalone algorithms are then
incorporated in an online framework that manages a given storage
and performs amnesic compression called ABQS.
To evaluate the proposed methods, we have collected empirical
data using a low-energy tracking platform called Camazotz on
both animals and vehicles. We also used synthetic dataset that is
statistically representative of flying foxes’ movement dynamics
to improve data diversity. In the experiments we evaluate the
framework in various aspects from pruning power, compression
rate, run time efficiency, operational time, compression error,
etc. The proposed methods demonstrate significant advantages in
general. In some experiments it even achieves 15 to 400 times
smaller error than its competitors.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
J. Liu, K. Zhao, P. Sommer, S. Shang, B. Kusy, and R. Jurdak, “Bounded
quadrant system: Error-bounded trajectory compression on the go,” in
ICDE, 2015, pp. 987–998.
R. Jurdak, P. Corke, A. Cotillon, D. Dharman, C. Crossman, and
G. Salagnac, “Energy-efficient localisation: Gps duty cycling with radio
ranging,” Transactions on Sensor Networks, vol. 9, no. 2, 2013.
D. Van Krevelen and R. Poelman, “A survey of augmented reality technologies, applications and limitations,” International Journal of Virtual
Reality, vol. 9, no. 2, p. 1, 2010.
S. B. Eisenman, E. Miluzzo, N. D. Lane, R. A. Peterson, G.-S. Ahn, and
A. T. Campbell, “Bikenet: A mobile sensing system for cyclist experience
mapping,” TOSN, vol. 6, no. 1, 2009.
R. Jurdak, P. Sommer, B. Kusy, N. Kottege, C. Crossman, A. Mckeown,
and D. Westcott, “Camazotz: multimodal activity-based gps sampling,”
in IPSN, 2013, pp. 67–78.
M. Nagy, Z. Ákos, D. Biro, and T. Vicsek, “Hierarchical group dynamics
in pigeon flocks,” Nature, vol. 464, no. 7290, pp. 890–893, 2010.
R. Jurdak, P. Corke, D. Dharman, and G. Salagnac, “Adaptive GPS duty
cycling and radio ranging for energy-efficient localization,” in SenSys,
2010, pp. 57–70.
D. H. Douglas and T. K. Peucker, Algorithms for the Reduction of
the Number of Points Required to Represent a Digitized Line or its
Caricature. John Wiley and Sons, Ltd, 2011, pp. 15–28.
J. Hershberger and J. Snoeyink, “Speeding up the douglas-peucker linesimplification algorithm,” in Proc. 5th Intl. Symp. on Spatial Data
Handling, 1992, pp. 134–143.
J. Muckell, J. Olsen, PaulW., J.-H. Hwang, C. Lawson, and S. Ravi,
“Compression of trajectory data: a comprehensive evaluation and new
approach,” GeoInformatica, pp. 1–26, 2013.
E. Keogh, S. Chu, D. Hart, and M. Pazzani, “An online algorithm for
segmenting time series,” in Proc. ICDM 2001, 2001, pp. 289–296.
M. Chen, M. Xu, and P. Fränti, “A fast o(n) multiresolution polygonal
approximation algorithm for gps trajectory simplification.” IEEE Transactions on Image Processing, vol. 21, no. 5, pp. 2770–2785, 2012.
C. Long, R. C.-W. Wong, and H. V. Jagadish, “Direction-preserving
trajectory simplification,” Proc. VLDB Endow., vol. 6, no. 10, pp. 949–
960, Aug. 2013.
R. Song, W. Sun, B. Zheng, and Y. Zheng, “PRESS: A
novel framework of trajectory compression in road networks,”
PVLDB, vol. 7, no. 9, pp. 661–672, 2014. [Online]. Available:
http://www.vldb.org/pvldb/vol7/p661-song.pdf
X. Xie, M. L. Yiu, R. Cheng, and H. Lu, “Scalable evaluation of
trajectory queries over imprecise location data,” IEEE Trans. Knowl.
Data Eng., vol. 26, no. 8, pp. 2029–2044, 2014. [Online]. Available:
http://dx.doi.org/10.1109/TKDE.2013.77
[16] J. Muckell, J.-H. Hwang, V. Patil, C. T. Lawson, F. Ping, and S. S. Ravi,
“Squish: An online approach for gps trajectory compression,” in Proc.
COM.Geo ’11. ACM, 2011, pp. 13:1–13:8.
[17] M. Potamias, K. Patroumpas, and T. Sellis, “Sampling trajectory streams
with spatiotemporal criteria,” in Scientific and Statistical Database Management, 2006. 18th International Conference on, 2006, pp. 275–284.
[18] G. Liu, M. Iwai, and K. Sezaki, “An online method for trajectory
simplification under uncertainty of gps,” in Information and Media
Technologies; VOL.8; NO.3, 2013, pp. 665–674.
[19] G. Trajcevski, H. Cao, P. Scheuermanny, O. Wolfsonz, and D. Vaccaro,
“On-line data reduction and the quality of history in moving objects
databases,” in Proc. MobiDE ’06. ACM, 2006, pp. 19–26.
[20] M. B. Kjærgaard, J. Langdal, T. Godsk, and T. Toftkjær, “Entracked:
Energy-efficient robust position tracking for mobile devices,” in Proc.
MobiSys ’09. ACM, 2009, pp. 221–234.
[21] E. J. Keogh, “Fast similarity search in the presence of longitudinal scaling
in time series databases,” in ICTAI, 1997, pp. 578–584.
[22] E. J. Keogh and M. J. Pazzani, “An enhanced representation of time series
which allows fast and accurate classification, clustering and relevance
feedback,” in KDD, 1998, pp. 239–243.
[23] E. J. Keogh, S. Chu, D. M. Hart, and M. J. Pazzani, “An online algorithm
for segmenting time series,” in ICDM, 2001, pp. 289–296.
[24] T. Palpanas, M. Vlachos, E. J. Keogh, D. Gunopulos, and W. Truppel,
“Online amnesic approximation of streaming time series,” in ICDE, 2004,
pp. 339–349.
[25] T. Palpanas, M. Vlachos, E. J. Keogh, and D. Gunopulos, “Streaming
time series summarization using user-defined amnesic functions,” IEEE
Trans. Knowl. Data Eng., vol. 20, no. 7, pp. 992–1006, 2008.
[26] S. Gandhi, L. Foschini, and S. Suri, “Space-efficient online approximation of time series data: Streams, amnesia, and out-of-order,” in ICDE,
2010, pp. 924–935.
[27] M. Potamias, K. Patroumpas, and T. K. Sellis, “Amnesic online synopses
for moving objects,” in CIKM, P. S. Yu, V. J. Tsotras, E. A. Fox, and
B. Liu, Eds. ACM, 2006, pp. 784–785.
[28] S. Nath, “Energy efficient sensor data logging with amnesic flash storage,” in IPSN, 2009, pp. 157–168.
[29] P. S. Heckbert and M. Garland, “Survey of polygonal surface simplification algorithms,” 1995.
[30] K. Zhao, R. Jurdak, J. Liu, D. Westcott, B. Kusy, H. Parry, P. Sommer,
and A. McKeown, “Optimal levy-flight foraging in a finite landscape,” J.
R. Soc. Interface, vol. 12, no. 104, March 2015.
[31] S. Shang, B. Yuan, K. Deng, K. Xie, K. Zheng, and X. Zhou, “PNN
query processing on compressed trajectories,” GeoInformatica, vol. 16,
no. 3, pp. 467–496, 2012.
[32] D. E. Knuth, The Art of Computer Programming, Volume 2 (3rd Ed.):
Seminumerical Algorithms. Boston, MA, USA: Addison-Wesley Longman Publishing Co., Inc., 1997.
[33] H. Risken, The Fokker-Planck Equation: Methods of Solutions and
Applications, 2nd ed., ser. Springer Series in Synergetics. Springer,
Sep. 1996.
[34] N. Meratnia and R. A. de By, “Spatiotemporal compression techniques
for moving point objects,” in EDBT, 2004, pp. 765–782.
AUTHORS ’ B IOGRAPHIES
Jiajun Liu is an Associate Professor at Renmin
University of China. He received his PhD and
BEng from The University of Queensland, Australia and from Nanjing University, China in 2012
and 2006 respectively. Before joining Renmin
University he has been a Postdoctoral Fellow
at the CSIRO of Australia from 2012 to 2015.
From 2006 to 2008 he also worked as a Researcher/Software Engineer for IBM China Research/Development Labs. His main research
interests are in multimedia and spatio-temporal
data management and mining. He serves as a reviewer for multiple
journals such as VLDBJ, TKDE, TMM, and as a PC member for ACM
MM and CCF Big Data.
Kun Zhao is a Postdoctoral Fellow at CSIRO
Australia. He obtained his PhD from Northeastern University, USA and his research interests
include network science, mobility data analysis
in sensor networks.
Philipp Sommer is a Research Scientist at
ABB Corporate Research in Switzerland. His research interests include a broad range of topics
in the field of wireless sensor networks, distributed computing and embedded systems. He
received MSc and PhD degrees in Electrical
Engineering from the Swiss Federal Institute of
Technology (ETH) in 2007 and 2011 respectively. He has been a postdoctoral fellow at the
Autonomous System lab of the CSIRO, Australia
between Sept. 2011 and Oct. 2014.
Shuo Shang is currently a faculty member
of Computer Science at China University of
Petroleum-Beijing, P.R.China. He was a Research Assistant Professor level Research Fellow with Department of Computer Science, Aalborg University, and was a faculty member of
the Center for Data-intensive Systems (Daisy),
which conducts research and offers education
with a focus on data management for various
data-intensive systems.
Brano Kusy is a Principal Research Scientist
in CSIRO’s Autonomous Systems Program in
Brisbane, Australia. CSIRO is Australia’s leading scientific research organization that operates across more than 50 sites in Australia. Autonomous systems program has a diverse group
of researchers and engineers that investigate
new technology in wireless sensor networks and
field robotics, including autonomous land, sea,
and air vehicles. He is also an adjunct senior
lecturer at the University of Queensland.
Jae-Gil Lee is an Associate Professor at KAIST
and he is leading Data Mining Lab. Before
that, he was a Postdoctoral Researcher at
IBM Almaden Research Center and a Postdoc Research Associate at Illinois at UrbanaChampaign. His research interests encompass
spatio-temporal data mining, social-network and
graph data mining, and big data analysis with
Hadoop/MapReduce.
Raja Jurdak is a Principal Research Scientist and leads the Distributed Sensing Systems
Group at CSIRO. He received the B.E. degree
from the American University of Beirut in 2000,
and the MSc and Ph.D. degrees from the University of California Irvine in 2001 and 2005 respectively. His current research interests are around
energy and mobility in sensor networks.
| 8cs.DS
|
Statistical Modeling of Propagation Channels for
Terahertz Band
Ali Rıza Ekti∗§ , Ali Boyacı‡ , Altan Alparslan∗‡‡ , İlhami Ünal† , Serhan Yarkank ,
Ali Görçin∗¶ , Hüseyin Arslan†† , Murat Uysal∗∗
∗ Informatics
and Information Security Research Center (BİLGEM), TÜBİTAK, Kocaeli, Turkey
of Electrical–Electronics Engineering, Balıkesir University, Balıkesir, Turkey
‡ Department of Electrical–Electronics Engineering, Istanbul Commerce University, İstanbul, Turkey
‡‡ Department of Electrical and Electronics Engineering, Yeditepe University, İstanbul, Turkey
† Millimeter Wave and Terahertz Technologies Research Laboratories (MİLTAL), TÜBİTAK, Kocaeli, Turkey
k Center for Applied Research on Informatics Technologies (CARIT), Istanbul Commerce University, İstanbul, Turkey
¶ Faculty of Electronics and Communications Engineering, Yıldız Technical University, İstanbul, Turkey
†† Department of Electrical–Electronics Engineering, Medipol University, İstanbul, Turkey
∗∗ Department of Electrical–Electronics Engineering, Özyeğin University, İstanbul, Turkey
Emails: [email protected], {aboyaci, syarkan}@ticaret.edu.tr,
[email protected], [email protected],
[email protected], [email protected], [email protected]
arXiv:1707.09740v1 [cs.IT] 31 Jul 2017
§ Department
Abstract—Digital revolution and recent advances in telecommunications technology enable to design communication systems
which operate within the regions close to the theoretical capacity
limits. Ever–increasing demand for wireless communications
and emerging numerous high–capacity services and applications
mandate providers to employ more bandwidth–oriented solutions
to meet the requirements. Trend and predictions point out that
marketplace targets data rates around 10Gbps or even more
within the upcoming decade. It is clear that such rates could only
be achieved by employing more bandwidth with the state–of–the–
art technology. Considering the fact that bands in the range of
275GHz–3000GHz, which are known as Terahertz (THz) bands,
are not allocated yet for specific active services around the globe,
there is an enormous potential to achieve the desired data rates.
Although THz bands look promising to achieve data rates on the
order of several tens of Gbps, realization of fully operational
THz communications systems obliges to carry out a multi–
disciplinary effort including statistical propagation and channel
characterizations, adaptive transceiver designs, reconfigurable
platforms, advanced signal processing algorithms and techniques
along with upper layer protocols equipped with various security
and privacy levels. Therefore, in this study, several important
statistical parameters for line–of–sight (LOS) channels are measured. High resolution frequency domain measurements are
carried out at single–sweep within a span of 60GHz. Impact of
antenna misalignment under LOS conditions is also investigated.
By validating exponential decay of the received power in both
time and frequency domain, path loss exponent is examined for
different frequencies along with the frequency–dependent path
loss phenomenon. Furthermore, impact of humidity is also tested
under LOS scenario. Measurement results are presented along
with relevant discussions and future directions are provided as
well.
This paper is accepted for publication in 2017 IEEE Conference on
Standards for Communications & Networking (CSCN 2017).
I. I NTRODUCTION
Wireless data traffic has grown tremendously in the past
years. According to the data analysis reports, total Internet
traffic will reach 2.3ZB whose 66% will be generated by
wireless devices [1]. It should be noted here that not only the
volume of data increases exponentially, but also the number of
users –and therefore– devices escalate. Short–term predictions
reveal that there will be seven billion people with seven trillion
nodes/devices connected with each other via various forms
of wireless communications technologies [2]. Moreover, it is
believed that marketplace targets data rates approaching multi–
Gbps even Tbps within ten to fifteen years [3, 4]. Evidently,
such rates could only be achieved by either extremely spectral
efficient modulations or expanding the transmission bandwidth
dramatically. Considering the fact that bands in the range
of 275GHz–3000GHz, which are known as Terahertz (THz)
bands, are not allocated yet for specific active services around
the globe, there is an enormous potential to achieve the desired
data rates.
Systems operating at THz frequencies are attracting great
interest and expected to meet the ever–increasing demand for
high–capacity wireless communications as well as consumer
expectations. Technological progress towards designing the
electronic components operating at THz frequencies will lead
to a wide range of applications especially for short–range
communications such as chip–to–chip communications, kiosk
downloading, device–to–device (D2D) communications, and
wireless backhauling [4–6]. Although THz bands look promising to achieve data rates on the order of several tens of Gbps,
realization of fully operational THz communications systems
obliges to carry out a multi–disciplinary effort including
statistical propagation and channel characterizations, adaptive
transceiver designs (including both baseband and radio frequency (RF) front–end portions), reconfigurable platforms,
advanced signal processing algorithms and techniques along
with upper layer protocols equipped with various security
and privacy levels. As in traditional wireless communications
systems design process, realization of high–performance and
reliable THz communications systems should start with obtaining detailed knowledge about the statistical properties of
the propagation channel. Next, these properties are incorporated into various channel characterizations and models. Upon
verification and validation of the characterizations and models
under different scenarios, system design stage is initiated at
the end.
A. Related Work
Studies that focus on modeling the channel for THz bands
in the literature could be categorized in various ways. Measurement methodology; bandwidth; temporal, frequency, and
spatial domain behaviors; and application–specific scenarios
are some of them among others. Each and every measurement
set concentrates on a specific scenario with some parameter
changes. For instance, channel measurement results at 300GHz
are described in [7] for different scenarios focusing on path
loss measurements under line–of–sight (LOS).
Ray tracing is also employed in the literature to characterize
the channel from the perspective of several parameters including regarding path loss, delay, angle of arrival, polarization,
and angle of departure [8]. But ray tracing approach suffers
from the following two problems: (i) A minuscule change in
the propagation environment mandates to redo the calculations.
Each scenario should be treated separately. Therefore, no
insight into fundamental statistics of the channel is provided.
(ii) Calculations are computationally intensive and any increase in the area/volume causes the computation time to
increase exponentially.
Bearing in mind that THz bands allow vast amount of
spectrum to be exploited, certain measurement and processing
strategies should be devised due to the practical concerns.
For instance, in [9], entire THz region measurement data
set is split into relatively smaller chunks and then post–
processed. Note that such a “divide–and–conquer” approach
is not safe, since there always exists a risk of creating
artifacts during compilation. Spatial diversity through the use
of multiple–input multiple–output (MIMO) measurements are
studied in literature as well. In [10], a 2 × 2–MIMO LOS
system operating at THz region with 10GHz bandwidth is
presented. However, there is a significant discrepancy between
measurement data presented and the theoretical rates, which
needs further research, verification, and validation. Apart from
the traditional channel measurements, scenario–specific measurement results are presented in the literature as well. In
[11], a 2D geometrical propagation model for short–range
D2D desktop communication channels and multipath fading
channels are considered.
Having a brief overview of the measurement studies at
hand,1 one could conclude that a systematic and comprehensive channel modeling strategy for THz bands has not been
established yet. Although THz bands manifest many intrinsic
propagation characteristics and mechanisms, LOS state is
preeminent among others because of the following reasons:
First, LOS is desired in THz bands for high–performance
operation. Second, LOS presents the elementary propagation
characteristics. In this regard, a detailed investigation of LOS
measurements should be the first step towards acquiring a
systematic and comprehensive statistical channel model.
B. Contributions
Contribution of this study is three–fold considering the
statistical channel characterization for THz scenarios: (i) To
the best knowledge of authors’, this work provides one of the
first single–sweep THz measurement results within 240GHz–
300GHz band and relevant statistical analysis. (ii) Detailed
statistical analyses of antenna–tilt measurement results under
LOS conditions within large–volume anechoic chamber are
provided. (iii) In addition, impact of humidity is also considered under LOS scenarios and relevant results are given.
C. Paper Organization
The remainder of this paper is structured as follows. System
and signal model are depicted in Section II. The measurement
setup is presented in Section III. Measurement results and relevant discussions are given in Section IV. Finally, conclusions
are drawn in Section V.
II. S TATISTICAL C HARACTERIZATION OF T ERAHERTZ
C HANNEL
A. System and Signal Model
1) System Model: In this study, LOS transmission scenario
for THz channels is investigated. As will be described in
Section III, a LOS transmission scenario is established within
an anechoic chamber along with very well–isolated setup
through the use of absorbers.
In order to characterize the THz channels, several factors
and parameters such as channel geometry, relative motion,
antenna directivity and radiation patterns along with environmental conditions (e. g., humidity, vapor, dust, etc.) and
so on should all be taken into account accordingly. On the
other hand, both analysis and transceiver design at further
stages should be simplified as well. From this point of view,
traditional linear, time–invariant channel model approach will
be adopted.
2) Signal Model: Any signal arriving at the receiver antenna at passband could be represented as:
r(t) = Re (xI (t) + jxQ (t)) ej2πfc t
(1)
√
where j = −1; Re {·} is the real part of its complex input;
xI (t) and xQ (t) represent the in–phase and quadrature components of the complex baseband equivalent of the transmitted
1 Interested readers may refer to [12, and references therein] for further
measurement results and campaigns present in the literature.
signal, respectively; fc denotes the transmission frequency.
Considering a static, general propagation environment, the
received signal is modeled as a superposition of multiple
copies of the transmitted signal with different delays and
attenuation levels. Therefore, the channel at passband could
be modeled with:
hP B (t) =
L−1
X
al δ (t − tl )
l=0
where hP B (t) denotes the channel impulse response at baseband; L is the number of resolvable multipath components;
al and tl represent attenuation factor and delay corresponding
to l–th path, respectively; and δ(·) represents the Dirac delta
function. Complex baseband equivalent of the channel is given
by
L−1
X
h(t) =
al e−j2πfc tl δ (t − tl )
l=0
Static LOS channels are a special sub–class of the multipath
channels where L = 1. Therefore, the complex baseband
equivalent channel can be described as:
h(t) = af ejθ δ(t − t0 )
(2)
where af represents the LOS path amplitude with respect to
the specular power affected by the environment geometry; θ is
the phase; t0 denotes the propagation delay as a deterministic
value of t0 = d/c with d denoting the transmitter–receiver
separation and c is the speed of light (3 × 108 m/s). It is
important to note here that in case directional antennas used,
which is the common practice in THz communications, the
impact of misalignment of the antennae; frequency–dependent
loss; and frequency dispersion index could all be absorbed by
the term, af .
In the literature, stochastic description of multipath components in a static environment is generally considered to be
superposition of both specular and diffused components such
that [13]:
ml = al e−j2πfc tl
= sl + dl
(3)
In (3),
and
sl = σsl e(j2πfc cos (θl )+φl )
(4)
M
1 Xl
bm e(j2πfc cos (θm )+φm )
dl = σdl √
Ml m=1
(5)
where σsl denotes the magnitude; θl is the angle of arrival
(AoA); and φl is the phase of the specular component,
respectively. Similarly, σdl is the magnitude of the diffused
component; Ml represents the number; bm is the amplitude;
θm is the AoA; and φm is the phase of the incoming waves
forming the diffused component, respectively. Without loss of
generality, both σsl and σdl could be assumed to be unity
under idealized conditions. Therefore, al could theoretically
be represented with Rayleigh distributed in the absence of
LOS component under uniform AoA assumption. In case LOS
is present, Rice and Nakagami distributions are employed to
describe the first–order statistics of al .
Because LOS is a special case among all other general
wireless propagation scenarios, both large– and small–scale
fading mechanisms exhibit intrinsic characteristics. As will be
discussed in Section III, fully isolated measurement setup with
absorbers in an anechoic chamber guarantees the LOS transmission. This implies that the losses introduced by propagation
channel are limited to distance–dependent path loss, possible
antenna misalignment(s), and equipment imperfections along
with non–ideal behaviors due to operating around/above ideal
regions. In this regard, signal model could be deduced down
to a direct path experiencing a distant–dependent path loss
without shadowing with a possible misalignment. This implies
that the contribution of path loss to the term af in (2) could
be given by:
P L = P L0 + 10n log10 (d) + M
(6)
where P L0 denotes the path loss at a reference distance
in the antenna far–field; n is the path loss exponent whose
value depends on the propagation environment and conditions
within; d is the transmitter–receiver separation; and M is the
random variable representing the impact of misalignment as a
random antenna gain. In the literature, due to its tractability
and plausibility, misalignment is considered only from the
transmitter or receiver antenna perspective as a zero–mean
Gaussian random variable with certain standard deviation, σM ,
representing the misalignment.
III. T ERAHERTZ M EASUREMENT E XPERIMENT S ETUP
A. Description of Measurement Setup
We constructed an experimental measurement setup in
the Millimeter Wave and Terahertz Technologies Research
Laboratories (MİLTAL) at the Scientific and Technological
Research Council of Turkey (TÜBİTAK) in Gebze, Turkey
and it can be seen in Figure 1. The dimensions of the
anechoic chamber used in the measurements are 7m×3m×4m
(length×width×height).
The measurement setup consists of four major parts: Agilent
performance network analyzer (PNA) vector network analyzer
(VNA) E8361A, Oleson Microwave Labs (OML) V03VNA2–
T and V03VNA2–T/R–A millimeter wave extender modules
and N5260A extender controller. Since the upper limit of the
VNA is 67GHz, the 220GHz to 325GHz extender modules
are attached to the VNA using the controller to measure
channel characteristics at THz frequencies. V03VNA2–T/R–
A extender module contains multipliers (×18) that extends
12.2GHz to 18.1GHz RF input signal to 220GHz to 325GHz
frequency range. Test intermediate frequency (IF) and reference IF signals for VNA are obtained by using downconversion mixers before transmiting. After passing through the
channel, the received signal is downconverted at the receiver
module, V03VNA2–T, by using downconversion mixers and
mitter and the receiver. Full 60GHz band measurements are
recorded with 4096 points averaging and 100Hz IF bandwidth
(BW). These parameters significantly reduce noise floor and
improves dynamic range.
B. Measurement Methodology
Fig. 1. Measurement setup in anechoic chamber from two different angles.
Note that LOS conditions are emulated by well–isolated measurement equipment in the anechoic chamber.
WR-03
RF
WR-03
xNxM
12.2 to 18.1
GHz
220-325 GHz
220-325 GHz
REF IF
TEST IF
5 to 300 MHz
Low Pass Filter
Low Pass Filter
5 to 300 MHz
xM
xM
xM
TEST IF
IV. M EASUREMENT R ESULTS
LO
5 to 300 MHz
xN
Low Pass Filter
LO
12.2 to 18.1 GHz
VNA measurements are performed for the frequency interval of 240GHz to 300GHz. This interval is measured using
standard gain horn antennas which are attached to OML
extenders. Each spectral measurement is represented with
4096 equally spaced frequency points (data points) within the
interval specified by the VNA. Therefore, a spectral resolution
of 14.648MHz is obtained. The measurement system including
connectors and cables is calibrated to remove the impairment
caused by the components. The calibration data is saved both
into the internal memory of the Agilent PNA VNA E8361A
and an external universal serial bus (USB) storage device.
Then, the calibration data is removed from the measurement
by using its internal memory and provided S21 parameter in
complex number format with real and imaginary parts for each
data point.
12.2 to 18.1
GHz
xN
TX
RX
Fig. 2. Block diagram of the measurement setup.
the resulting test IF (5MHz to 300MHz) is fed back to the
VNA. Difference in the transmitted and received signal is
analyzed to find channel characteristics. The corresponding
block diagram of our setup is shown in Figure 2.
The 220GHz to 325GHz source modules include balanced
multipliers, which are driven by an extended band WR–10
multiplier chain with a WR–03 waveguide output interface.
The output source match of the modules is typically 9dB.
The RF drive signal may be either continuous wave (CW)
or swept frequency. The required RF or local oscillator (LO)
signal power to operate OML modules is +10dBm. The
dynamic range is typically 75dB (min. 60dB) and the WR–03
waveguide output power of the V03VNA2–T/R–A is around
−23dBm. The magnitude and phase stability of the extender
modules are ±0.4dB and ±8◦ , respectively.
In this study, 240GHz to 300GHz is specially used to
get better performance from the extender modules, in terms
of magnitude and phase stability. Channel transfer function
is acquired by recording scattering parameters (s–parameter)
using the measurement setup shown in Figure 1 and Figure 2.
For channel modeling, calibration is performed with direct
interconnection of the transmitting and receiving extender
module’s waveguide ports. All the later measurements are
taken with standard gain horn antenna, with the gain of
24.8dBi at the center frequency, attached at both the trans-
As stated before, channel magnitude response for THz
channels is an important qualitative tool. Especially wideband measurements reveal different mechanisms whose effects
are obscured in traditional channel characterizations due to
relatively narrower band measurements. Frequency–dependent
loss is one of them. In this regard, channel magnitude response
could be used to analyze the frequency dependency of the
loss in frequency domain very easily. In Figure 3, averaged
magnitude responses are given. However, first it is appropriate
to evaluate the distance–dependent path loss. Based on the
measurement data, the path loss coefficients for specific frequencies are given in Table I. Overall mean path loss exponent
is found to be n = 1.9704 based on 4096–point resolution with
a variance of ≈ 0.003 by taking into account entire 60GHz
span. This result is in conformity with the LOS argument
and with the measurement results reported in the literature
[14, 15]. Assuming that the path loss exponent distributed
normally, the maximum likelihood estimate of the parameters
of the distribution match with the mean perfectly, whereas
2
σMLE
= 0.0591.
n
Another important aspect of wideband measurements is to
be able to observe the impact of frequency–dependent path
loss. Note that frequency dependency stems from the antenna
not from the transmission path itself in . This dependency is
clearly seen in Figure 3 between 0.27–0.29THz as a slight
collapse.
In order to validate the frequency domain results, time
domain analysis is carried out as well. By applying inverse
fast Fourier Transform (IFFT) operation, time domain data
are obtained. Raw time data are plotted in Figure 4. In
Figure 4, it is seen that the delay t0 is given in terms of
corresponding distance matching with the experimental setup
TABLE I
PATH L OSS E XPONENTS FOR S PECIFIC F REQUENCIES A LONG WITH
OVERALL S TATISTICS
240
250
260
270
280
290
300
PL. Exp. (n)
2.02
2.04
1.96
1.90
1.96
1.94
2.01
Overall Statistics
Mean (n)
1.9704, based on 4096 points
Variance (n)
0.0035, based on 4096 points
−22
−24
−26
Received Power (dB)
Freq. (GHz)
−20
−28
−30
−32
20cm
30cm
40cm
60cm
80cm
80cm tilt
−34
−36
L(λ) =
N
Y
−38
−40
240
250
260
270
Frequency (GHz)
280
290
300
Fig. 3. Averaged channel frequency responses in logarithmic scale for various
LOS scenarios including the impact of antenna misalignment via antenna tilt.
−10
20cm LOS
30cm LOS
40cm LOS
60cm LOS
80cm LOS
−20
−30
−40
Received Power (dB)
values. After validating the post–processing stage by delay–
distance match, raw time data are post–processed further by
removing the propagation delay and normalizing the power
with respect to the reference power. Results of this procedure
are given in Figure 5. As a cross–check for the tabulated
results given in Table I, exponential decay in the received
power is validated through the use of the impulse responses
obtained. Here, it is assumed that the underlying distribution is
of exponential form with f (x; λ) = λe−λx where λ is the parameter to be estimated for candidate exponentially distributed
random variable X where x ∈ R+ . In addition, assuming
that the underlying distribution of the power levels of paths
observed/measured is the same exponential distribution with
the parameter λ and that all N measurements are independent
of each other, the likelihood function, f (x1 , x2 , · · · , xN ; λ),
for the observed/measured data becomes
−50
−60
−70
−80
P
N −λ( N
k=1 xk )
f (xk ; λ) = λ e
k=1
Applying
logarithm on each side reads ln (L(λ)) = N ln (λ)−
PN
λ k=1 xk . Finally, differentiating the log–likelihood function
ln (L(λ))Pwith respect to λ and equating it to zero yields
N
N/λ − k=1 xk = 0. Therefore, the maximum likelihood
estimate of the parameter λ is obtained by
PN
xk
λ̂ = k=1
N
Both the measured data and corresponding maximum likelihood estimated theoretical curve are plotted in Figure 5.
Almost perfect match between the measured and maximum
likelihood estimation (MLE)–driven values could be recognized easily in the figure.
As discussed in Section II-A, antenna misalignment is an
important problem for THz communication channels. In this
regard, how antenna misalignment affects the received power
is evaluated in time domain. The results are given in Figure 6.
As can be seen from the figure, antenna misalignment causes
a significant drop in the peak power corresponding to LOS
path. Measurement results demonstrate that 10◦ tilt leads to
≈ 2.3dB, whereas 20◦ tilt causes ≈ 13dB degradation for
the peak power. Note that degradation in power level due
to antenna misalignment could also be verified in frequency
domain as plotted in Figure 3.
One of the main factors that affects the propagation behavior
of THz band signals is the presence of water vapor within the
−90
−100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Distance (m)
Fig. 4. First arriving paths in temporal domain for different transmitter–
receiver separations under LOS. The horizontal axis is given in terms of
transmitter–separation to validate the experimental setup values.
environment. It is known that water vapor molecules introduce
not only attenuation but also colored noise [16, 17]. In this
regard, the impact of the humidity is also investigated for THz
communication channels. Channel impulse response obtained
under a humid environment is appended into Figure 6 as well.
It is seen that peak power level of the LOS path does not
exhibit a significant drop in parallel with the measurement
results reported in the literature [18].
V. C ONCLUDING R EMARKS AND F UTURE D IRECTIONS
Deploying communication systems operating within THz
bands is considered to be an alternative strategy to meet
the ever–increasing data rate demands along with escalating
number of devices subscribing wireless networks. Due to the
technical limitations and propagation loss concerns, THz bands
have not attracted a significant attention up until the last
couple of years. However, with the technological advances, it
is possible to migrate up to THz region. Nevertheless, a successful and reliable communication system relies heavily on
scenarios.In this regard, non line–of–sight (NLOS) behavior
of the channels especially for indoor applications should be
examined in detail. Even though there are several studies
focusing on LOS and NLOS in the literature, more results are
required in order to have a more comprehensive understanding
of the channels. Finally, how mobility affects the channel
behavior should be studied under different conditions with
various measurement campaigns.
1
Experimental
Theoretical
0.9
Normalized Peak Path Power
0.8
0.7
0.6
0.5
0.4
ACKNOWLEDGMENT
0.3
The authors would like to thank Mr. Mustafa Kılıç from
MİLTAL, TÜBİTAK for his help during the measurements.
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
Tx−RX Separation (m)
0.6
0.7
0.8
Fig. 5. Normalized measured peak LOS path power levels along with
corresponding exponential decay function whose parameter is obtained by
MLE.
−30
80cm LOS
80cm 10° tilt
80cm 20° tilt
80cm with humidifier
−40
Received Power (dB)
−50
−36.8dB
−49.79dB
−60
−39.03dB
−70
−80
−90
−100
−110
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
Distance (m)
Fig. 6. Measured channel impulse responses at 80cm in logarithmic scale
with several antenna tilts along with a separate measurement in the presence
of dense humidity.
well–established propagation channel models and appropriate
transceiver designs.
In this study, single–sweep band measurement data for
240GHz–3000GHz band are collected in frequency domain
with a very high resolution within an anechoic chamber along
with a very well isolated setup to emulate the propagation.
Behavior of the channel within a 60GHz span (i.e., 240GHz–
300GHz interval) is captured at once. In addition, high–
resolution measurement data are collected so that finer temporal details are obtained to help design reliable transceiver systems including antenna misalignment problem. Since scenario
provides a theoretical borderline, collected data could be used
to validate other results obtained in different measurement
campaigns.
Practical applications for THz communications are generally envisioned to operate in short–range such as infotainment
systems. This implies that locations and spatial orientations
of THz devices could be random especially in residential
R EFERENCES
[1] Cisco, “White Paper: The Zettabyte Era-Trends and Analysis,” Available: http://www.cisco.com/c/en/us/solutions/collateral/service-provider/
visual-networking-index-vni/vni-hyperconnectivity-wp.pdf, Accessed:
Apr. 18, 2017.
[2] W. W. R. Forum, “Visions and Research Directions for The Wireless
World: Multi–RAT Network Architecture,” Available: http://www.wwrf.
ch/files/wwrf/content/files/publications/outlook/Outlook9.pdf, Accessed:
Apr. 18, 2017.
[3] T. Schneider, “Ultrahigh-Bitrate Wireless Data Communications via
THz-Links; Possibilities and Challenges,” J. Infrared Milli. Terahz.
Waves, vol. 36, no. 2, pp. 159–179, Feb. 2015.
[4] I. F. Akyıldız, J. M. Jornet, and C. Hana, “Terahertz band: Next frontier
for wireless communications,” Physical Communication, vol. 12, pp.
16–32, Sep. 2014.
[5] T. Kürner and S. Priebe, “Towards THz Communications - Status in
Research, Standardization and Regulation,” J. Infrared Milli. Terahz.
Waves, vol. 35, no. 1, pp. 53–62, Jan. 2014.
[6] T. Tajima, T. Kosugi, H.-J. Song, H. Hamada, A. E. Moutaouakil,
H. Sugiyama, H. Matsuzaki, M. Yaita, and O. Kagami, “Terahertz
MMICs and Antenna-in-Package Technology at 300 GHz for KIOSK
Download System,” J. Infrared Milli. Terahz. Waves, vol. 37, no. 2, pp.
1213–1224, Dec. 2016.
[7] S. Priebe, C. Jastrow, M. Jacob, T. Kleine-Ostmann, T. Schrader, and
T. Kürner, “Channel and Propagation Measurements at 300 GHz,” IEEE
Trans. Antennas Propag., vol. AP-59, no. 5, pp. 1688–1698, May 2011.
[8] S. Priebe and T. Kürner, “Stochastic Modeling of THz Indoor Radio
Channels,” IEEE Trans. Wireless Commun., vol. WC-12, no. 9, pp.
4445–4455, Sep. 2013.
[9] N. Khalid and O. B. Akan, “Wideband THz Communication Channel
Measurements for 5G Indoor Wireless Networks,” in Proc. IEEE 2016
ICC Int. Conf. Commun., Kuala Lumpur, Malaysia, Jul. 2016, pp. 1–6.
[10] ——, “Experimental Throughput Analysis of Low-THz MIMO Communication Channel in 5G Wireless Networks,” IEEE Wireless Commun.
Lett., vol. WCL-5, no. 6, pp. 616–619, Dec. 2016.
[11] S. Kim and A. Zajic, “Statistical Modeling and Simulation of ShortRange Device-to-Device Communication Channels at Sub-THz Frequencies,” IEEE Trans. Wireless Commun., vol. WC-15, no. 9, pp. 6423–
6433, Sep. 2016.
[12] T. Kürner, “Towards future THz communications systems,” Terahertz
Science and Technology, vol. 5, no. 1, pp. 11–17, Mar. 2012.
[13] S. Yarkan and H. Arslan, “Identification of LOS in Time-varying,
Frequency Selective Radio Channels,” EURASIP J. Wirel. Commun.
Netw., vol. 2008, pp. 8:1–8:14, Jan. 2008. [Online]. Available:
http://dx.doi.org/10.1155/2008/195948
[14] S. Kim and A. G. Zajic, “Statistical characterization of 300-ghz propagation on a desktop,” IEEE Transactions on Vehicular Technology, vol. 64,
no. 8, pp. 3330–3338, Aug 2015.
[15] H. Sawada, K. Fujii, A. Kasamatsu, H. Ogawa, K. Ishizu, and F. Kojima,
“Path loss model at 300 ghz for indoor mobile service applications,”
IEICE Communications Express, vol. 5, no. 11, pp. 424–428, 2016.
[16] J. M. Jornet and I. F. Akyildiz, “Channel capacity of electromagnetic
nanonetworks in the terahertz band,” in 2010 IEEE International Conference on Communications, May 2010, pp. 1–6.
[17] ——, “Channel modeling and capacity analysis for electromagnetic
wireless nanonetworks in the terahertz band,” IEEE Transactions on
Wireless Communications, vol. 10, no. 10, pp. 3211–3221, October 2011.
[18] G. A. Siles, J. M. Riera, and P. G. del Pino, “Atmospheric Attenuation in
Wireless Communication Systems at Millimeter and THz Frequencies,”
IEEE Antennas Propag. Mag., vol. 57, no. 1, pp. 48–61, Feb. 2015.
| 7cs.IT
|
A Static Analyzer for Large Safety-Critical Software
(Extended Abstract)
arXiv:cs/0701193v1 [cs.PL] 30 Jan 2007
Bruno Blanchet ∗ §
Laurent Mauborgne §
Patrick Cousot §
Antoine Miné §
Radhia Cousot ∗ ¶
David Monniaux ∗ §
Jérôme Feret §
Xavier Rival §
ABSTRACT
1. INTRODUCTION
We show that abstract interpretation-based static program
analysis can be made efficient and precise enough to formally
verify a class of properties for a family of large programs
with few or no false alarms. This is achieved by refinement
of a general purpose static analyzer and later adaptation to
particular programs of the family by the end-user through
parametrization. This is applied to the proof of soundness
of data manipulation operations at the machine level for
periodic synchronous safety critical embedded software.
The main novelties are the design principle of static analyzers by refinement and adaptation through parametrization (Sect. 3 and 7), the symbolic manipulation of expressions to improve the precision of abstract transfer functions
(Sect. 6.3), the octagon (Sect. 6.2.2), ellipsoid (Sect. 6.2.3),
and decision tree (Sect. 6.2.4) abstract domains, all with
sound handling of rounding errors in floating point computations, widening strategies (with thresholds: Sect. 7.1.2,
delayed: Sect. 7.1.3) and the automatic determination of
the parameters (parametrized packing: Sect. 7.2).
Critical software systems (as found in industrial plants,
automotive, and aerospace applications) should never fail.
Ensuring that such software does not fail is usually done by
testing, which is expensive for complex systems with high reliability requirements, and anyway fails to prove the impossibility of failure. Formal methods, such as model checking,
theorem proving, and static analysis, can help.
The definition of “failure” itself is difficult, in particular
in the absence of a formal specification. In this paper, we
choose to focus on a particular aspect found in all specifications for critical software, that is, ensuring that the critical
software never executes an instruction with “undefined” or
“fatal error” behavior, such as out-of-bounds accesses to arrays or improper arithmetic operations (such as overflows or
division by zero). Such conditions ensure that the program
is written according to its intended semantics, for example
the critical system will never abort its execution. These correctness conditions are automatically extractable from the
source code, thus avoiding the need for a costly formal specification. Our goal is to prove automatically that the software never executes such erroneous instructions or, at least,
to give a very small list of program points that may possibly
behave in undesirable ways.
In this paper, we describe our implementation and experimental studies of static analysis by abstract interpretation
over a family of critical software systems, and we discuss the
main technical choices and possible improvements.
Categories and Subject Descriptors
D.2.4 [Software Engineering]: Program Verification—formal methods, validation, assertion checkers; D.3.1 [Programming Languages]: Formal Definitions and Theory—semantics; F.3.1 [Logics and Meanings of Programs]:
Specifying and Verifying and Reasoning about Programs—
Mechanical verification, assertions, invariants; F.3.2 [Logics
and Meanings of Programs]: Semantics of Programming
Languages—Denotational semantics, Program analysis.
General Terms
Algorithms, Design, Experimentation, Reliability, Theory,
Verification.
Keywords
Abstract Interpretation; Abstract Domains; Static Analysis; Verification; Floating Point; Embedded, Reactive, RealTime, Safety-Critical Software.
∗
CNRS (Centre National de la Recherche Scientifique)
École normale supérieure. [email protected]
¶
École polytechnique. [email protected]
§
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
PLDI’03, June 9–11, 2003, San Diego, California, USA.
Copyright 2003 ACM 1-58113-662-5/03/0006 ...$5.00.
2. REQUIREMENTS
When dealing with undecidable questions on program execution, the verification problem must reconcile correctness
(which excludes non exhaustive methods such as simulation or test), automation (which excludes model checking
with manual production of a program model and deductive
methods where provers must be manually assisted), precision (which excludes general analyzers which would produce
too many false alarms, i.e., spurious warnings about potential errors), scalability (for software of a few hundred thousand lines), and efficiency (with minimal space and time
requirements allowing for rapid verification during the software production process which excludes a costly iterative
refinement process).
Industrialized general-purpose static analyzers satisfy all
criteria but precision and efficiency. Traditionally, static
analysis is made efficient by allowing correct but somewhat
imprecise answers to undecidable questions. In many usage
contexts, imprecision is acceptable provided all answers are
sound and the imprecision rate remains low (e.g. 5 to 15%
of the runtime tests cannot typically be eliminated). This is
the case for program optimization (such as static elimination
of run-time array bound checks), program transformation
(such as partial evaluation), etc.
In the context of program verification, where human interaction must be reduced to a strict minimum, false alarms
are undesirable. A 5% rate of false alarms on a program of a
few hundred thousand lines would require a several personyear effort to manually prove that no error is possible. Fortunately, abstract interpretation theory shows that for any
finite class of programs, it is possible to achieve full precision and great efficiency [7] by discovering an appropriate
abstract domain. The challenge is to show that this theoretical result can be made practical by considering infinite but
specific classes of programs and properties to get efficient
analyzers producing few or no false alarms. A first experiment on smaller programs of a few thousand lines was quite
encouraging [5] and the purpose of this paper is to report on
a real-life application showing that the approach does scale
up.
3.
DESIGN PRINCIPLE
The problem is to find an abstract domain that yields an
efficient and precise static analyzer for the given family of
programs. Our approach is in two phases, an initial design
phase by specialists in charge of designing a parametrizable
analyzer followed by an adaptation phase by end-users in
charge of adapting the analyzer for (existing and future)
programs in the considered family by an appropriate choice
of the parameters of the abstract domain and the iteration
strategy (maybe using some parameter adaptation strategies
provided by the analyser).
3.1 Initial Design by Refinement
Starting from an existing analyzer [5], the initial design
phase is an iterative manual refinement of the analyzer. We
have chosen to start from a program in the considered family that has been running for 10 years without any run-time
error, so that all alarms are, in principle, due to the imprecision of the analysis. The analyzer can thus be iteratively
refined for this example until all alarms are eliminated.
Each refinement step starts with a static analysis of the
program, which yields false alarms. Then a manual backward inspection of the program starting from sample false
alarms leads to the understanding of the origin of the imprecision of the analysis. There can be two different reasons
for the lack of precision:
• Some local invariants are expressible in the current version of the abstract domain but were missed either:
– because some abstract transfer function (Sect. 5.4) was
too coarse, in which case it must be rewritten closer to
the best abstraction of the concrete transfer function [9],
(Sect. 6.3);
– or because a widening (Sect. 5.5) was too coarse, in which
case the iteration strategy must be refined (Sect. 7.1);
• Some local invariants are necessary in the correctness
proof but are not expressible in the current version of the
abstract domain. To express these local invariants, a new
abstract domain has to be designed by specialists and incorporated in the analyzer as an approximation of the reduced
product [9] of this new component with the already existing
domain (Sect. 7.2).
When this new refinement of the analyzer has been implemented, it is tested on typical examples and then on the full
program to verify that some false alarms have been eliminated. In general the same cause of imprecision appears
several times in the program; furthermore, one single cause
of imprecision at some program point often leads later to
many false alarms in the code reachable from that program
point, so a single refinement typically eliminates a few dozen
if not hundreds of false alarms.
This process is to be repeated until there is no or very few
false alarms left.
3.2 Adaptation by Parametrization
The analyzer can then be used by end-users in charge
of proving programs in the family. The necessary adaptation of the analyzer to a particular program in the family
is by appropriate choice of some parameters. An example
provided in the preliminary experience [5] was the widening
with thresholds (Sect. 7.1.2). Another example is relational
domains (such as octagons [30], Sect. 6.2.2) which cannot
be applied to all global variables simultaneously because the
corresponding analysis would be too expensive; it is possible to have the user supply for each program point groups
of variables on which the relational analysis should be independently applied.
In practice we have discovered that the parametrization
can be largely automated (and indeed it is fully automated
for octagons as explained in Sect. 7). This way the effort to
manually adapt the analyzer to a particular program in the
family is reduced to a minimum.
3.3 Analysis of the Alarms
We implemented and used a slicer [34] to help in the
alarm inspection process. If the slicing criterion is an alarm
point, the extracted slice contains the computations that
led to the alarm. However, the classical data and control
dependence-based backward slicing turned out to yield prohibitively large slices.
In practice we are not interested in the computation of
the variables for which the analyzer already provides a value
close to end-user specifications, and we can consider only the
variables we lack information about (integer or floating point
variables that may contain large values or boolean variables
that may take any value according to the invariant). In the
future we plan to design more adapted forms of slicing: an
abstract slice would only contain the computations that lead
to an alarm point wherever the invariant is too weak.
4. THE CONSIDERED FAMILY OF PROGRAMS
The considered programs in the family are automatically
generated using a proprietary tool from a high-level specification familiar to control engineers, such as systems of differential equations or synchronous operator networks (block
diagrams as illustrated in Fig. 1), which is equivalent to
the use of synchronous languages (like Lustre [20]). Such
synchronous data-flow specifications are quite common in
real-world safety-critical control systems ranging from letter sorting machine control to safety control and monitoring
systems for nuclear plants and “fly-by-wire” systems. Periodic synchronous programming perfectly matches the need
for the real-time integration of differential equations by for-
ward, fixed step numerical methods. Periodic synchronous
programs have the form:
declare volatile input, state and output variables;
initialize state variables;
loop forever
– read volatile input variables,
– compute output and state variables,
– write to volatile output variables;
wait for next clock tick;
end loop
Our analysis proves that no exception can be raised (but
the clock tick) and that all data manipulation operations
are sound. The bounded execution time of the loop body
should also be checked by static analysis [16] to prove that
the real-time clock interrupt does occur at idle time.
We operate on the C language source code of those systems, ranging from a few thousand lines to 132,000 lines of
C source code (75 kLOC after preprocessing and simplification as in Sect. 5.1). We take into account all machinedependent aspects of the semantics of C (as described in
[5]) as well as the periodic synchronous programming aspects (for the wait). We use additional specifications to
describe the material environment with which the software
interacts (essentially ranges of values for a few hardware
registers containing volatile input variables and a maximal
execution time to limit the possible number of iterations in
the external loop1 ).
The source codes we consider use only a reduced subset
of C, both in the automatically generated glue code and the
handwritten pieces. As it is often the case with critical systems, there is no dynamic memory allocation and the use
of pointers is restricted to call-by-reference. On the other
hand, an important characteristics of those programs is that
the number of global and static2 variables is roughly linear in the length of the code. Moreover the analysis must
consider the values of all variables and the abstraction cannot ignore any part of the program without generating false
alarms. It was therefore a grand challenge to design an analysis that is precise and does scale up.
5.
STRUCTURE OF THE ANALYZER
The analyzer is implemented in Objective Caml [25]. It
operates in two phases: the preprocessing and parsing phase
followed by the analysis phase.
5.1 Preprocessing Phase
The source code is first preprocessed using a standard
C preprocessor, then parsed using a C99-compatible parser.
Optionally, a simple linker allows programs consisting of several source files to be processed.
The program is then type-checked and compiled to an
intermediate representation, a simplified version of the abstract syntax tree with all types explicit and variables given
unique identifiers. Unsupported constructs are rejected at
this point with an error message.
Syntactically constant expressions are evaluated and re1
Most physical systems cannot run forever and some event
counters in their control programs are bounded because of
this physical limitation.
2
In C, a static variable has limited lexical scope yet is persistent with program lifetime. Semantically, it is the same
as a global variable with a fresh name.
placed by their value. Unused global variables are then
deleted. This phase is important since the analyzed programs use large arrays representing hardware features with
constant subscripts; those arrays are thus optimized away.
Finally the preprocessing phase includes preparatory work
for trace partitioning (Sect. 7.1.5) and parametrized packing
(Sect. 7.2).
5.2 Analysis Phase
The analysis phase computes the reachable states in the
considered abstract domain. This abstraction is formalized
by a concretization function γ [8, 9, 11]. The computation
of the abstraction of the reachable states by the abstract
interpreter is called abstract execution.
The abstract interpreter first creates the global and
static variables of the program (the stack-allocated variables are created and destroyed on-the-fly). Then the abstract execution is performed compositionally, by induction
on the abstract syntax, and driven by the iterator.
5.3 General Structure of the Iterator
The abstract execution starts at a user-supplied entry
point for the program, such as the main function. Each program construct is then interpreted by the iterator according
to the semantics of C as well as some information about the
target environment (some orders of evaluation left unspecified by the C norm, the sizes of the arithmetic types, etc.,
see [5]). The iterator transforms the C instructions into directives for the abstract domain that represents the memory
state of the program (Sect. 6.1), that is, the global, static
and stack-allocated variables.
The iterator operates in two modes: the iteration mode
and the checking mode. The iteration mode is used to generate invariants; no warning is displayed when some possible
errors are detected. When in checking mode, the iterator issues a warning for each operator application that may give
an error on the concrete level (that is to say, the program
may be interrupted, such as when dividing by zero, or the
computed result may not obey the end-user specification for
this operator, such as when integers wrap-around due to an
overflow). In all cases, the analysis goes on with the nonerroneous concrete results (overflowing integers are wiped
out and not considered modulo, thus following the end-user
intended semantics).
Tracing facilities with various degrees of detail are also
available. For example the loop invariants which are generated by the analyzer can be saved for examination.
5.4 Primitives of the Iterator
Whether in iteration or checking mode, the iterator starts
with an abstract environment E ♯ at the beginning of a statement S in the program and outputs an abstract environment JSK♯ (E ♯ ) which is a valid abstraction after execution
of statement S. This means that if a concrete environment
maps variables to their values, JSKs is a standard semantics
of S (mapping an environment ρ before executing S to the
corresponding environment JSKs (ρ) after execution of S),
JSKc is the collecting semantics of S (mapping a set E of
environments before executing S to the corresponding set
JSKc (E) = {JSKs (ρ) | ρ ∈ E} of environments after execution of S), γ(E ♯ ) is the set of concrete environments before S then JSK♯ (E ♯ ) over-approximates the set JSKc (γ(E ♯ ))
of environments after executing S in that JSKc (γ(E ♯ )) ⊆
γ(JSK♯ (E ♯ )). The abstract semantics JSK♯ is defined as follows:
• Tests: let us consider a conditional
S = if (c) { S1 } else { S2 }
(an absent else branch is considered as an empty execution
sequence). The condition c can be assumed to have no side
effect and to contain no function call, both of which can be
handled by first performing a program transformation. The
iterator computes:
JSK♯ (E ♯ )
= JS1 K♯ (guard♯ (E ♯ , c)) ⊔♯ JS2 K♯ (guard♯ (E ♯ , ¬c))
where the abstract domain implements:
– ⊔♯ as the abstract union that is an abstraction of the
union ∪ of sets of environments;
– guard♯ (E ♯ , c) as an approximation of JcKc (γ(E ♯ )) where
the collecting semantics JcKc (E) = {ρ ∈ E | JcKs (ρ) = true}
of the condition c is the set of concrete environments ρ in
E satisfying condition c. In practice, the abstract domain
only implements guard♯ for atomic conditions and compound
ones are handled by structural induction.
• Loops are by far the most delicate construct to analyze.
Let us denote by E0♯ the environment before the loop:
while (c) {body }
The abstract loop invariant to be computed for the head of
the loop is an upper approximation of the least invariant
of F where F (E) = γ(E0♯ ) ∪ JbodyKc (JcKc (E)). The fixpoint
computation F ♯ (E ♯ ) = E0♯ ⊔♯ JbodyK♯ (guard♯ (E ♯ , c)) is always done in iteration mode, requires a widening (Sect. 5.5)
and stops with an abstract invariant E ♯ satisfying F ♯ (E ♯ )
⊑♯ E ♯ (where the abstract partial ordering x ⊑♯ y implies
γ(x) ⊆ γ(y)) [11]. When in checking mode, the abstract
loop invariant has first to be computed in iteration mode
and then, an extra iteration (in checking mode this time),
starting from this abstract invariant is necessary to collect
potential errors.
• Sequences i1 ;i2 : first i1 is analyzed, then i2 , so that:
Ji1 ;i2 K♯ (E ♯ ) = Ji2 K♯ ◦ Ji1 K♯ (E ♯ ) .
• Function calls are analyzed by abstract execution of the
function body in the context of the point of call, creating
temporary variables for the parameters and the return value.
Since the considered programs do not use recursion, this
gives a context-sensitive polyvariant analysis semantically
equivalent to inlining.
• Assignments are passed to the abstract domain.
• Return statement: We implemented the return statement by carrying over an abstract environment representing the accumulated return values (and environments, if the
function has side effects).
5.5 Least Fixpoint Approximation
Widening and Narrowing
with
The analysis of loops involves the iterative computation of
an invariant E ♯ that is such that F ♯ (E ♯ ) ⊑♯ E ♯ where F ♯ is
an abstraction of the concrete monotonic transfer function
F of the test and loop body. In abstract domains with infinite height, this is done by widening iterations
computing
`
♯
a finite sequence E0♯ = ⊥, . . . , En+1
= En♯
F ♯ (En♯ ), . . . ,
♯
EN
of successive abstract elements, until finding an invari`
♯
ant EN
. The widening operator
should be sound (that
`
is the concretization of x
y should overapproximate the
concretizations of x and y) and ensure the termination in
finite time [8, 11] (see an example in Sect. 7.1.2).
In general, this invariant is not the strongest one in the
abstract domain. This invariant is then made more and
♯
♯
more precise by narrowing iterations: EN
, . . . , En+1
=
a ♯
a
♯
♯
En F (En ) where a
the narrowing operator
is sound (the
concretization of x
y is an upper approximation of the
intersection of x and y) and ensures termination [8, 11].
6. ABSTRACT DOMAINS
The elements of an abstract domain abstract concrete
predicates, that is, properties or sets of program states. The
operations of an abstract domain are transfer functions abstracting predicate transformers corresponding to all basic
operations in the program [8]. The analyzer is fully parametric in the abstract domain (this is implemented using an Objective Caml functor). Presently the analyzer uses the memory abstract domain of Sect. 6.1, which abstracts sets of program data states containing data structures such as simple
variables, arrays and records. This abstract domain is itself
parametric in the arithmetic abstract domains (Sect. 6.2)
abstracting properties of sets of (tuples of) boolean, integer
or floating-point values. Finally, the precision of the abstract
transfer functions can be significantly improved thanks to
symbolic manipulations of the program expressions preserving the soundness of their abstract semantics (Sect. 6.3).
6.1 The Memory Abstract Domain
When a C program is executed, all data structures (simple
variables, arrays, records, etc) are mapped to a collection of
memory cells containing concrete values. The memory abstract domain is an abstraction of sets of such concrete memory states. Its elements, called abstract environments, map
variables to abstract cells. The arithmetic abstract domains
operate on the abstract value of one cell for non-relational
ones (Sect. 6.2.1) and on several abstract cells for relational
ones (Sect. 6.2.2, 6.2.3, and 6.2.4). An abstract value in a
abstract cell is therefore the reduction of the abstract values
provided by each different basic abstract domain (that is an
approximation of their reduced product [9]).
6.1.1 Abstract Environments
An abstract environment is a collection of abstract cells,
which can be of the following four types:
• An atomic cell represents a variable of a simple type
(enumeration, integer, or float) by an element of the arithmetic abstract domain. Enumeration types, including the
booleans, are considered to be integers.
• An expanded array cell represents a program array using one
element of the array. Formally, let
` cell for each
´
A = (v1i , . . . , vni ) i∈∆ be the family (indexed by a set ∆)
of values of the array (of size n) to be abstracted. The abstraction is ⊥ (representing non-accessibility of dead code)
when A is empty. Otherwise the abstraction is an abstract
array A♯e of sizeSn such the expanded array cell A♯e [k] is the
abstraction of i∈∆ vki for k = 1, . . . , n. Therefore the abstraction is component-wise, each element of the array being
abstracted separately.
• A shrunk array cell represents a program array using a
single cell. Formally
the abstraction is a shrunk array cell
S
S
i
A♯s abstracting n
k=1
i∈∆ vk . All elements of the array are
thus “shrunk” together. We use this representation for large
arrays where all that matters is the range of the stored data.
• A record cell represents a program record (struct) using one cell for each field of the record. Thus our abstraction
is field-sensitive.
6.1.2 Fast Implementation of Abstract Environments
A naive implementation of abstract environments may use
an array. We experimented with in-place and functional arrays and found this approach very slow. The main reason
is that abstract union ⊔♯ operations are expensive, because
they operate in time linear in the number of abstract cells;
since both the number of global variables (whence of abstract cells) and the number of tests (involving the abstract
union ⊔♯ ) are linear in the length of the code, this yields a
quadratic time behavior.
A simple yet interesting remark is that in most cases,
abstract union operations are applied between abstract environments that are identical on almost all abstract cells:
branches of tests modify a few abstract cells only. It is therefore desirable that those operations should have a complexity proportional to the number of differing cells between
both abstract environments. We chose to implement abstract environments using functional maps implemented as
sharable balanced binary trees, with short-cut evaluation
when computing the abstract union, abstract intersection,
widening or narrowing of physically identical subtrees [5,
§6.2]. An additional benefit of sharing is that it contributes
to the rather light memory consumption of our analyzer.
On a 10,000-line example we tried [5], the execution time
was divided by seven, and we are confident that the execution times would have been prohibitive for the longer examples. The efficiency of functional maps in the context of
sophisticated static analyses has also been observed by [26]
for representing first-order structures.
6.1.3 Operations on Abstract Environments
Operations on a C data structure are translated into operations on cells of the current abstract environments. Most
translations are straightforward.
– Assignments: In general, an assignment lvalue := e is
translated into the assignment of the abstract value of e
into the abstract cell corresponding to lvalue. However, for
array assignments, such as x[i] := e, one has to note that the
array index i may not be fully known, so all cells possibly
corresponding to x[i] may either be assigned the value of
e, or keep their old value. In the analysis, these cells are
assigned the upper bound of their old abstract value and
the abstract value of e. Similarly, for a shrunk array x, after
an assignment x[i] := e, the cell representing x may contain
either its old value (for array elements not modified by the
assignment), or the value of e.
– Guard: The translation of concrete to abstract guards
is not detailed since similar to the above case of assignments.
– Abstract union, widening, narrowing: Performed cellwise between abstract environments.
6.2 Arithmetic Abstract Domains
The non-relational arithmetic abstract domains abstract
sets of numbers while the relational domains abstract sets of
tuples of numbers. The basic abstract domains we started
with [5] are the intervals and the clocked abstract domain
abstracting time. They had to be significantly refined using
octagons (Sect. 6.2.2), ellipsoids (Sect. 6.2.3) and decision
trees (Sect. 6.2.4).
6.2.1 Basic Abstract Domains
• The Interval Abstract Domain. The first, and simplest,
implemented domain is the domain of intervals, for both
integer and floating-point values [8]. Special care has to be
taken in the case of floating-point values and operations to
always perform rounding in the right direction and to handle
special IEEE [23] values such as infinities and NaN s (Not a
Number).
• The Clocked Abstract Domain. A simple analysis using
the intervals gives a large number of false warnings. A great
number of those warnings originate from possible overflows
in counters triggered by external events. Such errors cannot happen in practice, because those events are counted at
most once per clock cycle, and the number of clock cycles
in a single execution is bounded by the maximal continuous
operating time of the system.
We therefore designed a parametric abstract domain. (In
our case, the parameter is the interval domain [5].) Let
X ♯ be an abstract domain for a single scalar variable. The
elements of the clocked domain consist in triples in (X ♯ )3 .
♯
♯
A triple (v ♯ , v−
, v+
) represents the set of values x such that
♯
♯
♯
x ∈ γ(v ), x − clock ∈ γ(v−
) and x + clock ∈ γ(v+
), where
clock is a special, hidden variable incremented each time the
analyzed program waits for the next clock signal.
6.2.2 The Octagon Abstract Domain
Consider the following program fragment:
R := X−Z;
L := X;
if (R>V) L := Z+V;
At the end of this fragment, we have L ≤ X. In order to
prove this, the analyzer must discover that, when the test
is true, we have R = X − Z and R > V, and deduce from this
that Z + V < X (up to rounding). This is possible only with a
relational domain able to capture simple linear inequalities
between variables.
Several such domains have been proposed, such as the
widespread polyhedron domain [13]. In our prototype, we
have chosen the recently developed octagon abstract domain
[28, 30], which is less precise but faster than the polyhedron domain: it can represent sets of constraints of the form
±x±y ≤ c, and its complexity is cubic in time and quadratic
in space (w.r.t. the number of variables), instead of exponential for polyhedra. Even with this reduced cost, the huge
number of live variables prevents us from representing sets
of concrete environments as one big abstract state (as it was
done for polyhedra in [13]). Therefore we partition the set of
variables into small subsets and use one octagon for some of
these subsets (such a group of variables being then called a
pack ). The set of packs is a parameter of the analysis which
can be determined automatically (Sect. 7.2.1).
Another reason for choosing octagons is the lack of support for floating-point arithmetics in the polyhedron domain. Designing relational domains for floating-point variables is indeed a difficult task, not much studied until recently [27]. On one hand, the abstract domain must be
sound with respect to the concrete floating-point semantics
(handling rounding, NaN s, etc.); on the other hand it should
use floating-point numbers internally to manipulate abstract
data for the sake of efficiency. Because invariant manipulations in relational domains rely on some properties of the
real field not true for floating-points (such as x + y ≤ c and
z − y ≤ d implies x + z ≤ c + d), it is natural to consider
that abstract values represent subsets of RN (in the relational invariant x + y ≤ c, the addition + is considered in
R, without rounding, overflow, etc.). Our solution separates
the problem in two. First, we design a sound abstract domain for variables in the real field (our prototype uses the
octagon library [28] which implementation is described in
[29]). This is much easier for octagons than for polyhedra,
as most computations are simple (addition, multiplication
and division by 2). Then, each floating-point expression is
transformed into a sound approximate real expression taking rounding, overflow, etc. into account (we use the linear
forms described in Sect. 6.3) and evaluated by the abstract
domain.
Coming back to our example, it may seem that octagons
are not expressive enough to find the correct invariant as
Z + V < X is not representable in an octagon. However,
our assignment transfer function is smart enough to extract
from the environment the interval [c, d] where V ranges (with
d ≤ RM where RM is an upper bound of R already computed
by the analysis) and synthesize the invariant c ≤ L − Z ≤ d,
which is sufficient to prove that subsequent operations on L
will not overflow. Thus, there was no need for this family
of programs to use a more expressive and costly relational
domain.
Remark that this approach provides a generic way of
implementing relational abstract domains on floating-point
numbers. It is parametrized by:
• a strategy for the determination of packs (Sect. 7.2.1);
• an underlying abstract domain working in the real field.
Aspects specific to floating-point computation (such as
rounding and illegal operations) are automatically taken
care of by our approach.
6.2.3 The Ellipsoid Abstract Domain
To achieve the necessary precision, several new abstract
domains had to be designed. We illustrate the general approach on the case of the ellipsoid abstract domain.
By inspection of the parts of the program on which the
previously described analyses provide no information at all
on the values of some program variables, we identified code
of the form:
if (B) {
Y := i;
X := j;
} else {
X′ := aX − bY + t;
Y := X;
X := X′ ;
}
where a and b are floating-point constants, i, j and t are
floating-point expressions, B is a boolean expression, and X,
X′ , and Y are program variables. The previously described
analyses yield the imprecise result that X and Y may take any
value. This apparently specialized style of code is indeed
quite frequent in control systems since it implements the
simplified second order digital filtering discrete-time system
illustrated in Fig. 1.
The first branch is a reinitialization step, the second
branch consists in an affine transformation Φ. Since this
code is repeated inside loops, the analysis has to find an
invariant preserved by this code. We looked manually for
such an invariant on typical examples, identified the above
a
b
++
+
t
z-1
z-1
Unit delay
Unit delay
Switch
x(n)
Switch
j
Switch
B
i
Figure 1: A simplified second-order digital filtering
system.
generic form (essentially depending on a and b), then designed a generic abstract domain εa,b able to discover such
invariants, implemented the abstract domain lattice and
transfer operations and finally let the analyzer automatically
instantiate the specific analysis to the code (in particular to
parts that may not have been inspected).
To find an interval that contains the values of X and Y
in the specific case where we can compute bounds to the
expression t by the previously described analyses, say |t| ≤
tM , we have designed a new abstract domain εa,b based on
ellipsoids, that can capture the required invariant. More
precisely, we can show that:
Proposition
1 If 0 < b < 1, a2 − 4b < 0, and k ≥
”2
“
tM
√
, then the constraint X2 − aXY + bY2 ≤ k is pre1− b
served by the affine transformation Φ.
The proof of this proposition follows by algebraic manipulations using standard linear algebra techniques. In our
examples, the conditions on a and b required in Prop. 1 are
satisfied. We still have to design the abstract operations to
propagate the invariant in the program, and to take into
account rounding errors that occur in floating-point computations (and are not modeled in the above proposition).
Having fixed two floating-point numbers a and b such that
0 < b < 1 and a2 − 4b < 0, we present a domain εa,b , for
describing sets of ellipsoidal constraints. An element in εa,b
is a function r which maps a pair of variables (X, Y) to a
floating-point number r(X, Y) such that X2 − aXY + bY2 ≤
r(X, Y).
We briefly describe some primitives and transfer functions
of our domain:
• Assignments. Let r ∈ εa,b be the abstract element describing some constraints before a statement X := e, our
goal is to compute the abstract element r ′ describing a set
of constraints satisfied after this statement:
1. in case e is a variable Y, each constraint containing Y
gives a constraint for X. Formally, we take r ′ such that
r ′ (U, V ) = r(σU, σV ) where σ is the substitution of the
variable Y for the variable X;
2. in case e is an expression of the form aY + bZ + t, we first
remove any constraint containing X, then we add a new
constraint for X and Y. We therefore take:
r ′ = r[(X, ) 7→ +∞][( , X) 7→ +∞][(X, Y) 7→ δ(r(Y, Z))] .
We have used the function δ defined as follows:
!2
!!
√
√
√
|a| b + b
b + 4f √
k + (1 + f )tM
δ(k) =
4b − a2
where f is the greatest relative error of a float with respect to a real and t ∈ [−tM , tM ]. Indeed, we can show
that, if Y2 − aYZ + bZ2 ≤ k and X = aY −
√ bZ + t, then in
exact real arithmetic X2 − aXY + bY2 ≤ ( bk + tM )2 , and
taking into account rounding errors, we get the above
formula for δ(k);
3. otherwise, we remove all constraints containing X by taking r ′ = r[(X, ) 7→ +∞][( , X) 7→ +∞]3 .
′
• Guards are ignored, i.e., r = r.
• Abstract union, intersection, widening and narrowing
are computed component-wise. The widening uses thresholds as described in Sect. 7.1.2.
The abstract domain εa,b cannot compute accurate results by itself, mainly because of inaccurate assignments (in
case 3.) and guards. Hence we use an approximate reduced
product with the interval domain. A reduction step consists in substituting in the function r the image of a couple
(X, Y) by the smallest element among r(X, Y) and the floatingpoint number k such that k is the least upper bound to the
evaluation of the expression X2 − aXY + bY2 in the floatingpoint numbers when considering the computed interval constraints. In case the values of the variable X and Y are proved
to be equal, we can be much more precise and take the smallest element among r(X, Y) and the least upper bound to the
evaluation of the expression (1 − a + b)X2 .
These reduction steps are performed:
• before computing the union between two abstract elements r1 and r2 , we reduce each constraint ri (X, Y) such that
ri (X, Y) = +∞ and r3−i (X, Y) 6= +∞ (where i ∈ {1; 2});
• before computing the widening between two abstract
elements r1 and r2 , we reduce each constraint r2 (X, Y) such
that r2 (X, Y) = +∞ and r1 (X, Y) 6= +∞;
• before an assignment of the form X′ := aX − bY + t, we
refine the constraints r(X, Y).
These reduction steps are especially useful in handling a
reinitialization iteration.
Ellipsoidal constraints are then used to reduce the intervals of variables: after each assignment A of q
the form
√
′ (X′ ,X)
′
′
X := aX − bY + t, we use the fact that |X | ≤ 2 b r4b−a
2 ,
where r ′ is the abstract element describing a set of ellipsoidal
constraints just after the assignment A.
This approach is generic and has been applied to handle
the digital filters in the program.
6.2.4 The Decision Tree Abstract Domain
Apart from numerical variables, the code uses also a
great deal of boolean values, and no classical numerical domain deals precisely enough with booleans. In particular,
booleans can be used in the control flow and we need to relate the value of the booleans to some numerical variables.
Here is an example:
B := (X=0);
if (¬ B) Y := 1/X;
We found also more complex examples where a numerical variable could depend on whether a boolean value had
3
This is also the case for initialization.
changed or not. In order to deal precisely with those examples, we implemented a simple relational domain consisting
in a decision tree with leaf an arithmetic abstract domain4 .
The decision trees are reduced by ordering boolean variables
(as in [6]) and by performing some opportunistic sharing of
subtrees.
The only problem with this approach is that the size of
decision trees can be exponential in the number of boolean
variables, and the code contains thousands of global ones.
So we extracted a set of variable packs, and related the
variables in the packs only, as explained in Sect. 7.2.3.
6.3 Symbolic Manipulation of Expressions
We observed, in particular for non-relational abstract domains, that transfer functions proceeding by structural induction on expressions are not precise when the variables
in the expression are not independent. Consider, for instance, the simple assignment X := X − 0.2 ∗ X performed in
the interval domain in the environment X ∈ [0, 1]. Bottomup evaluation will give X − 0.2 ∗ X ⇒ [0, 1] − 0.2 ∗ [0, 1] ⇒
[0, 1] − [0, 0.2] ⇒ [−0.2, 1]. However, because the same X
is used on both sides of the − operator, the precise result
should have been [0, 0.8].
In order to solve this problem, we perform some simple
algebraic simplifications on expressions before feeding them
to the abstract domain. Our approach is to linearize each
expression e, that is to say, transform it into a linear form
ℓJeK on the set of variables v1 , . . . , vN with interval coeffiP
cients: ℓJeK = N
i=1 [αi , βi ]vi +[α, β]. The linear form ℓJeK is
computed by recurrence on the structure of e. Linear operators on linear forms (addition, subtraction, multiplication
and division by a constant interval) are straightforward. For
instance, ℓJX − 0.2 ∗ XK = 0.8 ∗ X, which will be evaluated to
[0, 0.8] in the interval domain. Non-linear operators (multiplication of two linear forms, division by a linear form, nonarithmetic operators) are dealt by evaluating one or both
linear form argument into an interval.
Although the above symbolic manipulation is correct in
the real field, it does not match the semantics of C expressions for two reasons:
• floating-point computations incur rounding;
• errors (division by zero, overflow, etc.) may occur.
Thankfully, the systems we consider conform to the IEEE
754 norm [23] that describes rounding very well (so that,
e.g., the compiler should be prevent from using the multiplyadd-fused instruction on machines for which the result of a
multiply-add computation may be slightly different from the
floating point operation operation A + (B × C) for some input values A, B, C). Thus, it is easy to modify the recursive
construction of linear forms from expressions to add the error contribution for each operator. It can be an absolute
error interval, or a relative error expressed as a linear form.
We have chosen the absolute error which is more easily implemented and turned out to be precise enough.
To address the second problem, we first evaluate the expression in the abstract interval domain and proceed with
the linearization to refine the result only if no possible arithmetic error was reported. We are then guaranteed that the
simplified linear form has the same semantics as the initial
expression.
4
The arithmetic abstract domain is generic. In practice, the
interval domain was sufficient.
7.
ADAPTATION
TION
VIA
PARAMETRIZA-
In order to adapt the analyzer to a particular program
of the considered family, it may be necessary to provide information to help the analysis. A classical idea is to have
users provide assertions (which can be proved to be invariants and therefore ultimately suppressed). Another idea is
to use parametrized abstract domains in the static program
analyzer. Then the static analysis can be adapted to a particular program by an appropriate choice of the parameters.
We provide several examples in this section. Moreover we
show how the analyzer itself can be used in order to help or
even automatize the appropriate choice of these parameters.
7.1 Parametrized Iteration Strategies
7.1.1 Loop Unrolling
In many cases, the analysis of loops is made more precise
by treating the first iteration of the loop separately from
the following ones; this is simply a semantic loop unrolling
transformation: a while loop may be expanded as follows:
if (condition) { body; while (condition) { body } }
The above transformation can be iterated n times, where the
concerned loops and the unrolling factor n are user-defined
parameters. In general, the larger the n, the more precise
the analysis, and the longer the analysis time.
7.1.2 Widening with Thresholds
Compared to normal interval analysis [10, §2.1.2], ours
does not jump straight away to ±∞, but goes through`a
number of thresholds. The widening with thresholds T
for the interval analysis of Sect. 6.2.1 is parametrized by
a threshold set T that is a finite set of numbers containing
−∞ and +∞ and defined such that:
[a, b]
`
T
[a′ , b′ ] = [if a′ < a then max{ℓ ∈ T | ℓ ≤ a′ } else a,
if b′ > b then min{h ∈ T | h ≥ b′ } else b]
In order to illustrate the benefits of this parametrization
(see others in [5]), let x0 be the initial value of a variable X
subject to assignments of the form X := αi ∗ X + βi , i ∈ ∆
in the main loop, where the αi , βi , i ∈ ∆ are floating point
constants such that 0 ≤ αi < 1. Let be any M such that
|βi |
M ≥ max{|x0 |, 1−α
, i ∈ ∆}. We have M ≥ |x0 | and
i
M ≥ αi M + |βi | and so all possible sequences x0 = x0 ,
xn+1 = αi xn + βi , i ∈ ∆ of values of variable X are bounded
since ∀n ≥ 0 : |xn | ≤ M . Discovering M may be difficult in
particular if the constants αi , βi , i ∈ ∆ depend on complex
boolean conditions. As long as the set T of thresholds contains some number greater or equal to the minimum M , the
interval analysis of X with thresholds T will prove that the
value of X is bounded at run-time since some element of T
will be an admissible M .
In practice we have chosen T to be (±αλk )0≤k≤N . The
choice of α and λ mostly did not matter much in the first
experiments. After the analysis had been well refined and
many causes of imprecision removed, we had to choose a
smaller value for λ to remove some false alarms. In any
case, αλN should be large enough; otherwise, many false
alarms for overflow are produced.
7.1.3 Delayed Widening
When widening the previous iterate by the result of the
transfer function on that iterate at each step as in Sect. 5.5,
some values which can become stable after two steps of
widening may not stabilize. Consider the example:
X := Y + γ;
Y := α ∗ X + δ
This should be equivalent to Y := α ∗ Y + β (with β =
δ + αγ), and so a widening with thresholds should find a
stable interval. But if we perform a widening with thresholds at each step, each time we widen Y, X is increased to
a value surpassing the threshold for Y, and so X is widened
to the next stage, which in turn increases Y further and the
next widening stage increases the value of Y. This eventually
results in top abstract values for X and Y.
In practice, we first do N0 iterations with unions on all
abstract domains, then we do widenings unless a variable
which was not stable becomes stable (this is the case of
Y here when the threshold is big enough as described in
Sect. 7.1.2). We add a fairness condition to avoid livelocks in
cases for each iteration there exists a variable that becomes
stable.
7.1.4 Floating Iteration Perturbation
The stabilization check for loops considered in Sect. 5.4
has to be adjusted because of the floating point computations in the abstract. Let us consider that [a, b] is the mathematical interval of values of a variable X on entry of a loop.
We let FC,A ([a, b]) be the mathematical interval of values of
X after a loop iteration. C = R means that the concrete
operations in the loop are considered to be on mathematical
real numbers while C = F means that the concrete operations in the loop are considered to be on machine floating
point numbers. If FR,A ([a, b]) = [a′ , b′ ] then FF,A ([a, b]) =
[a′ − ǫ1 , b′ + ǫ1 ] because of the cumulated concrete rounding errors ǫ1 ≥ 0 when evaluating the loop body5 . The
same way A = R means that the interval abstract domain
is defined ideally using mathematical real numbers while
A = F means that the interval abstract domain is implemented with floating point operations performing rounding
in the right direction. Again, if FC,R ([a, b]) = [a′ , b′ ] then
FC,F ([a, b]) = [a′ − ǫ2 , b′ + ǫ2 ] because of the cumulated abstract rounding errors during the static analysis of the loop
body. The analyzer might use FF,F which is sound since if
FR,R ([a, b]) = [a′ , b′ ] then FF,F ([a, b]) = [a′ −ǫ1 −ǫ2 , b′ +ǫ1 +ǫ2 ]
which takes both the concrete and abstract rounding errors
into account (respectively ǫ1 and ǫ2 ).
Mathematically, a loop invariant for variable X is an interval [a, b] such that FF,R ([a, b]) ⊆ [a, b]. However, the loop
stabilization check is made as FF,F ([a, b]) ⊆ [a, b], which is
sound but incomplete: if FF,R ([a, b]) is very close to [a, b],
e.g. FF,R ([a, b]) = [a, b] then, unfortunately, FF,F ([a, b]) =
[a − ǫ2 , b + ǫ2 ] * [a, b]. This will launch useless additional
iterations whence a loss of time and precision.
The solution we have chosen is to overapproximate FF,F
by FbF,F such that FbF,F ([a, b]) = [a′ − ǫ ∗ |a′ |, b′ + ǫ ∗ |b′ |] where
[a′ , b′ ] = FF,F ([a, b]) and ǫ is a parameter of the analyzer chosen to be an upper bound of the possible abstract rounding
errors in the program loops. Then the loop invariant inter5
We take the rounding error on the lower and upper bound
to be the same for simplicity.
val is computed iteratively with FbF,F , which is simply less
precise than with FF,F , but sound. The loop stabilization
test is performed with FF,F which is sound. It is also more
precise in case ǫ∗(min{|a′ |; |b′ |}) is greater than the absolute
error on the computation of FF,F ([a′ − ǫ ∗ |a′ |, b′ + ǫ ∗ |b′ |]).
We have not investigated about the existence (nor about
the automatic computation) of such a parameter in the general case yet, nevertheless attractiveness of the encountered
fixpoints made the chosen parameter convenient.
7.1.5 Trace Partitioning
In the abstract execution of the program, when a test
is met, both branches are executed and then the abstract
environments computed by each branch are merged. As described in [5] we can get a more precise analysis by delaying
this merging.
This means that:
if (c) { S1 } else { S2 } rest
is analyzed as if it were
if (c) { S1 ; rest } else { S2 ; rest } .
A similar technique holds for the unrolled iterations of loops.
As this process is quite costly, the analyzer performs this
trace partitioning in a few end-user selected functions, and
the traces are merged at the return point of the function.
Informally, in our case, the functions that need partitioning
are those iterating simultaneously over arrays a[] and b[]
such that a[i] and b[i] are linked by an important numerical constraint which does not hold in general for a[i] and
b[j] where i 6= j. This solution was simpler than adding
complex invariants to the abstract domain.
7.2 Parametrized Abstract Domains
Recall that our relational domains (octagons of Sect. 6.2.2,
and decision trees of Sect. 6.2.4) operate on small packs of
variables for efficiency reasons. This packing is determined
syntactically before the analysis. The packing strategy is a
parameter of the analysis; it gives a trade-off between accuracy (more, bigger packs) and speed (fewer, smaller packs).
The strategy must also be adapted to the family of programs
to be analyzed.
7.2.1 Packing for Octagons
We determine a set of packs of variables and use one octagon for each pack. Packs are determined once and for
all, before the analysis starts, by examining variables that
interact in linear assignments within small syntactic blocks
(curly-brace delimited blocks). One variable may appear in
several packs and we could do some information propagation
(i.e. reduction [9]) between octagons at analysis time, using
common variables as pivots; however, this precision gain was
not needed in our experiments. There is a great number of
packs, but each pack is small; it is our guess that our packing
strategy constructs, for our program family, a linear number of constant-sized octagons, effectively resulting in a cost
linear in the size of the program. Moreover, the octagon
packs are efficiently manipulated using functional maps, as
explained in Sect. 6.1.2, to achieve sub-linear time costs via
sharing of unmodified octagons.
Our current strategy is to create one pack for each syntactic block in the source code and put in the pack all variables that appear in a linear assignment or test within the
associated block, ignoring what happens in sub-blocks of
the block. For example, on a program of 75 kLOC, 2,600
octagons were detected, each containing four variables on
average. Larger packs (resulting in increased cost and precision) could be created by considering variables appearing
in one or more levels of nested blocks; however, we found
that, in our program family, it does not improve precision.
7.2.2 Packing Optimization for Octagons
Our analyzer outputs, as part of the result, whether each
octagon actually improved the precision of the analysis. It
is then possible to re-run the analysis using only packs that
were proven useful, thus greatly reducing the cost of the
analysis. (In our 75 kLOC example, only 400 out of the
2,600 original octagons were in fact useful.) Even when the
program or the analysis parameters are modified, it is perfectly safe to use a list of useful packs output by a previous
analysis. We experimented successfully with the following
method: generate at night an up-to-date list of good octagons by a full, lengthy analysis and work the following
day using this list to cut analysis costs.
7.2.3 Packing for Decision Trees
In order to determine useful packs for the decision trees
of Sect. 6.2.4, we used the following strategy: each time a
numerical variable assignment depends on a boolean, or a
boolean assignment depends on a numerical variable, we put
both variables in a tentative pack. If, later, we find a program point where the numerical variable is inside a branch
depending on the boolean, we mark the pack as confirmed.
In order to deal with complex boolean dependences, if we
find an assignment b := expr where expr is a boolean expression, we add b to all packs containing a variable in expr.
In the end, we just keep the confirmed packs.
At first, we restrained the boolean expressions used to extend the packs to simple boolean variables (we just considered b := b’) and the packs contained at most four boolean
variables and dozens of false alarms were removed. But we
discovered that more false alarms could be removed if we extended those assignments to more general expressions. The
problem was that packs could then contain up to 36 boolean
variables, which gave very bad performance. So we added
a parameter to restrict arbitrarily the number of boolean
variables in a pack. Setting this parameter to three yields
an efficient and precise analysis of boolean behavior.
8. EXPERIMENTAL RESULTS
The main program we are interested in is 132,000 lines of
C with macros (75 kLOC after preprocessing and simplification as in Sect. 5.1) and has about 10,000 global/static
variables (over 21,000 after array expansion as in Sect. 6.1).
We had 1,200 false alarms with the analyzer [5] we started
with. The refinements of the analyzer described in this paper reduce the number of alarms down to 11 (and even 3,
depending on the versions of the analyzed program). Fig. 2
gives the total analysis time for a family of related programs
on commodity hardware (2.4 GHz, 1 Gb RAM PC), using a
slow but precise iteration strategy.
The memory consumption of the analyzer is reasonable
(550 Mb for the full-sized program). Several parameters, for
instance the size of the octagon packs (Sect. 7.2.1), allow for
a space-precision trade-off.
The packing optimization strategy of reusing results
from preceding analysis to reduce the number of octagons
8000
7000
6000
time (s)
5000
4000
3000
2000
1000
0
0
10
20
30
40
kLOC
50
60
70
80
Figure 2: Total analysis time for the family of programs without packing optimization (Sect. 7.2.2).
(Sect. 7.2.2) reduces, on the largest example code, memory consumption from 550 Mb to 150 Mb and time from
1 h 40 min to 40 min. Furthermore, the current automatic
tuning of the iteration strategy may be made more efficient,
using fewer iterations and thus reducing analysis time.
9.
RELATED WORK
Let us discuss some other verification methods that could
have been considered. Dynamic checking methods were excluded for a safety critical system (at best data can be collected at runtime and checked offline). Static methods requiring compiler or code instrumentation (such as [15]) were
also excluded in our experiment since the certified compiler
as well as the compiled code, once certified by traditional
methods, cannot be modified without costly re-certification
processes. Therefore we only consider the automated static
proof of software run-time properties, which has been a recurrent research subject since a few decades.
9.1 Software Model Checking
Software model checking [22] has proved very useful to
trace logical design errors, which in our case has already
been performed at earlier stages of the software development, whereas we concentrate on abstruse machine implementation aspects of the software. Building a faithful model
of the program (e.g. in Promela for Spin [22]) would be just
too hard (it can take significantly more time to write a model
than it did to write the code) and error-prone (by checking a
manual abstraction of the code rather than the code itself, it
is easy to miss errors). Moreover the abstract model would
have to be designed with a finite state space small enough
to be fully explored (in the context of verification, not just
debugging), which is very difficult in our case since sharp
data properties must be taken into account. So it seems important to have the abstract model automatically generated
by the verification process, which is the case of the abstract
semantics in static analyzers.
9.2 Dataflow Analysis and Software Abstract
Model Checking
Dataflow analyzers (such as ESP [14]) as well as abstraction based software model checkers (such as a.o. Blast [21],
CMC [31] and Slam [4]) have made large inroads in tackling
programs of comparable size and complexity. Their impressive performance is obtained thanks to coarse abstractions
(e.g. resulting from a program “shrinking” preprocessing
phase [14, 1] or obtained by a globally coarse but locally precise abstraction [31]). In certain cases, the abstract model
is just a finite automaton, whose transitions are triggered
by certain constructions in the source code [15]; this allow
checking at the source code level high-level properties, such
as “allocated blocks of memory are freed only once” or “interrupts are always unmasked after being blocked”, ignoring
dependencies on data.
The benefit of this coarse abstraction is that only a small
part of the program control and/or data have to be considered in the actual verification process. This idea did not
work out in our experiment since merging paths or data inevitably leads to many false alarms. On the contrary we had
to resort to context-sensitive polyvariant analyses (Sect. 5.4)
with loop unrolling (Sect. 7.1.1) so that the size of the (semantically) “expanded” code to analyze is much larger than
that of the original code. Furthermore, the properties we
prove include fine numerical constraints, which excludes simple abstract models.
9.3 Deductive Methods
Proof assistants (such as Coq [33], ESC [17] or PVS [32])
face semantic problems when dealing with real-life programming languages. First, the prover has to take the machinelevel semantics into account (e.g., floating-point arithmetic
with rounding errors as opposed to real numbers, which is
far from being routinely available 6 ). Obviously, any technique for analyzing machine arithmetic will face the same
semantic problems. However, if the task of taking concrete
and rounding errors into account turned out to be feasible
for our automated analyzer, this task is likely to be daunting in the case of complex decision procedures operating on
ideal arithmetic [32]. Furthermore, exposing to the user the
complexity brought by those errors is likely to make assisted
manual proof harrowing.
A second semantic difficulty is that the prover needs to
operate on the C source code, not on some model written
in a prototyping language so that the concrete program semantics must be incorporated in the prover (at least in the
verification condition generator). Theoretically, it is possible to do a “deep embedding” of the analyzed program into
the logic of the proof assistant — that is, providing a mathematical object describing the syntactic structure of the program as well as a formal semantics of the programming language. Proving any interesting property is then likely to be
extremely difficult. “Shallow embeddings” — mapping the
original program to a corresponding “program” in the input
syntax of the prover — are easier to deal with, but may
be difficult to produce in the presence of nondeterministic
inputs, floating-point rounding errors etc. . .
The last and main difficulty with proof assistants is that
they must be assisted, in particular to help providing inductive arguments (e.g. invariants). Of course these provers
could integrate abstract domains in the form of abstraction procedures (to perform online abstractions of arbitrary
predicates into their abstract form) as well as decision procedures (e.g. to check for abstract inclusion ⊑♯ ). The main
problem is to have the user provide program independent
6
For example ESC is simply unsound with respect to modular arithmetics [17].
hints, specifying when and where these abstraction and decision procedures must be applied, as well as how the inductive arguments can be discovered, e.g. by iterative fixpoint
approximation, without ultimately amounting to the implementation of a static program analysis.
Additionally, our analyzer is designed to run on a whole
family of software, requiring minimal adaptation to each
individual program. In most proof assistants, it is difficult
to change the program without having to do a considerable
amount of work to adapt proofs.
9.4 Predicate Abstraction
Predicate abstraction, which consists in specifying an abstraction by providing the atomic elements of the abstract
domain in logical form [19] e.g. by representing sets of states
as boolean formulas over a set of base predicates, would certainly have been the best candidate. Moreover most implementations incorporate an automatic refinement process by
success and failure [2, 21] whereas we successively refined
our abstract domains manually, by experimentation. In addition to the semantic problems shared by proof assistants,
a number of difficulties seem to be insurmountable to automate this design process in the present state of the art of
deductive methods:
9.4.1 State Explosion Problem:
To get an idea of the size of the necessary state space,
we have dumped the main loop invariant (a textual file over
4.5 Mb).
The main loop invariant includes 6,900 boolean interval
assertions (x ∈ [0, 1]), 9,600 interval assertions (x ∈ [a, b]),
25,400 clock assertions (Sect. 6.2.1), 19,100 additive octagonal assertions (a ≤ x + y ≤ b), 19,200 subtractive octagonal
assertions (a ≤ x − y ≤ b, see Sect. 6.2.2), 100 decision trees
(Sect. 6.2.4) and 1,900 ellipsoidal assertions (Sect. 6.2.3)7 .
In order to allow for the reuse of boolean model checkers, the conjunction of true atomic predicates is usually encoded as a boolean vector over boolean variables associated
to each predicate [19] (the disjunctive completion [9] of this
abstract domain can also be used to get more precision [2,
21], but this would introduce an extra exponential factor).
Model checking state graphs corresponding to several tenths
of thousands of boolean variables (not counting hundreds of
thousands of program points) is still a real challenge. Moreover very simple static program analyzes, such as Kildall’s
constant propagation [24], involve an infinite abstract domain which cannot be encoded using finite boolean vectors
thus requiring the user to provide beforehand all predicates
that will be indispensable to the static analysis (for example the above mentioned loop invariant involves, e.g., over
16,000 floating point constants at most 550 of them appearing in the program text).
Obviously some of the atomic predicates automatically
generated by our analysis might be superfluous. On one
hand it is hard to say which ones and on the other hand this
does not count all other predicates that may be indispensable at some program point to be locally precise. Another
approach would consist in trying to verify each potential
7
Figures are rounded to the closest hundred. We get more
assertions than variables because in the 10,000 global variables arrays are counted once whereas the element-wise abstraction yields assertions on each array element. Boolean
assertions are needed since booleans are integers in C.
faulty operation separately (e.g., focus on one instruction
that may overflow at a time) and generate the abstractions
lazily [21]. Even though repeating this analysis over 100,000
times might be tractable, the real difficulty is to automatically refine the abstract predicates (e.g. to discover that
considered in Prop. 1).
9.4.2 Predicate Refinement:
Predicate abstraction per se uses a finite domain and is
therefore provably less powerful than our use of infinite abstract domains (see [12], the intuition is that all inductive assertions have to be provided manually). Therefore predicate
abstraction is often accompanied by a refinement process to
cope with false alarms [2, 21].
Under specific conditions, this refinement can be proved
equivalent to the use of an infinite abstract domain with
widening [3].
Formally this refinement is a fixpoint computation [7, 18]
at the concrete semantics level, whence introduces new elements in the abstract domain state by state without termination guarantee whereas, e.g., when introducing clocks
from intervals or ellipsoids from octagons we exactly look
for an opposite more synthetic point of view. Therefore the
main difficulty of counterexample-based refinement is still
to automate the presently purely intellectual process of designing precise and efficient abstract domains.
10. CONCLUSION
In this experiment, we had to cope with stringent requirements. Industrial constraints prevented us from requiring
any change in the production chain of the code. For instance, it was impossible to suggest changes to the library
functions that would offer the same functionality but would
make the code easier to analyze. Furthermore, the code
was mostly automatically generated from a high-level specification that we could not have access to, following rules of
separation of design and verification meant to prevent the
intrusion of unproved high-level assumptions into the verification assumptions. It was therefore impossible to analyze
the high-level specification instead of analyzing the C code.
That the code was automatically generated had contrary
effects. On the one hand, the code fit into some narrow
subclass of the whole C language. On the other hand, it
used some idioms not commonly found in human-generated
code that may make the analysis more difficult; for instance,
where a human would have written a single test with a
boolean connective, the generated code would make one test,
store the result into a boolean variable, do something else
do the second test and then retrieve the result of the first
test. Also, the code maintains a considerable number of
state variables, a large number of these with local scope but
unlimited lifetime. The interactions between several components are rather complex since the considered program
implement complex feedback loops across many interacting
components.
Despite those difficulties, we developed an analyzer with
a very high precision rate, yet operating with reasonable
computational power and time. Our main effort was to
discover an appropriate abstraction which we did by manual refinement through experimentation of an existing analyzer [5] and can be later adapted by end-users to particular
programs through parametrization (Sect. 6.3 and 7). To
achieve this, we had to develop two specialized abstract domains (Sect. 6.2.3 and 6.2.4) and improve an existing domain
(Sect. 6.2.2).
The central idea in this approach is that once the analyzer
has been developed by specialists, end-users can adapt it to
other programs in the family without much efforts. However
coming up with a tool that is effective in the hands of end
users with minimal expertise in program analysis is hard.
This is why we have left to the user the simpler parametrizations only (such as widening thresholds in Sect. 7.1.2 easily
found in the program documentation) and automated the
more complex ones (such as parametrized packing Sect. 7.2).
Therefore, the approach should be economically viable.
11. REFERENCES
[1] S. Adams, T. Ball, M. Das, S. Lerner, K. Rajamani,
M. Seigle, , and W. Weimer. Speeding up dataflow
analysis using flow-insensitive pointer analysis. SAS
(2002), LNCS 2477, Springer, 117–132.
[2] T. Ball, R. Majumdar, T. Millstein, and S. Rajamani.
Automatic predicate abstraction of C programs.
PLDI. ACM SIGPLAN Not. 36(5) (2001), 203–213.
[3] T. Ball, A. Podelski, and S. Rajamani. Relative
completeness of abstraction refinement for software
model checking. TACAS (2002), LNCS 2280,
Springer, 158–172.
[4] T. Ball and S. Rajamani. The SLAM project:
debugging system software via static analysis. 29th
ACM POPL (2002), 1–3.
[5] B. Blanchet, P. Cousot, R. Cousot, J. Feret,
L. Mauborgne, A. Miné, D. Monniaux, and X. Rival.
Design and implementation of a special-purpose static
program analyzer for safety-critical real-time
embedded software. The Essence of Computation:
Complexity, Analysis, Transformation. (2002), LNCS
2566, Springer, 85–108.
[6] R. Bryant. Graph based algorithms for boolean
function manipulation. IEEE Trans. Comput. C-35
(1986), 677–691.
[7] P. Cousot. Partial completeness of abstract fixpoint
checking. SARA (2000), LNAI 1864, Springer, 1–25.
[8] P. Cousot and R. Cousot. Abstract interpretation: a
unified lattice model for static analysis of programs by
construction or approximation of fixpoints. 4th ACM
POPL (1977), 238–252.
[9] P. Cousot and R. Cousot. Systematic design of
program analysis frameworks. 6th ACM POPL (1979),
269–282.
[10] P. Cousot and R. Cousot. Abstract interpretation and
application to logic programs. J. of Logic Prog. 2–3,
13 (1992), 103–179.
[11] P. Cousot and R. Cousot. Abstract interpretation
frameworks. J. Logic and Comp. 2, 4 (1992), 511–547.
[12] P. Cousot and R. Cousot. Comparing the Galois
connection and widening/narrowing approaches to
abstract interpretation. PLILP (1992), LNCS 631,
Springer, 269–295.
[13] P. Cousot and N. Halbwachs. Automatic discovery of
linear restraints among variables of a program. 5th
ACM POPL (1978), 84–97.
[14] M. Das, S. Lerner, and M. Seigle. ESP: Path-sensitive
program verification in polynomial time. 29th ACM
PLDI (2002), 58–70.
[15] D. Engler, B. Chelf, A. Chou, and S. Hallem.
Checking system rules using system-specific,
programmer-written compiler extensions. Usenix
Association, OSDI 2000, 1–16.
[16] C. Ferdinand, R. Heckmann, M. Langenbach,
F. Martin, M. Schmidt, H. Theiling, S. Thesing, and
R. Wilhelm. Reliable and precise WCET
determination for a real-life processor. ESOP (2001),
LNCS 2211, Springer, 469–485.
[17] C. Flanagan, K. R. M. Leino, M. Lillibridge,
G. Nelson, J. Saxe, and R. Stata. Extended static
checking for Java. PLDI. ACM SIGPLAN Not. 37(5)
(2002), 234–245.
[18] R. Giacobazzi and E. Quintarelli. Incompleteness,
counterexamples and refinements in abstract
model-checking. SAS (2001), LNCS 126, Springer,
356–373.
[19] S. Graf and H. Saı̈di. Construction of abstract state
graphs with PVS. CAV (1997), LNCS 1254, Springer,
72–83.
[20] N. Halbwachs, P. Caspi, P. Raymond, and D. Pilaud.
The synchronous dataflow programming language
Lustre. Proc. of the IEEE 79, 9 (1991), 1305–1320.
[21] T. Henzinger, R. Jhala, R. Majumdar, and G. Sutre.
Lazy abstraction. 29th ACM POPL (2002), 58–70.
[22] G. Holzmann. The model checker Spin. IEEE Trans.
Softw. Eng. 23, 5 (1997), 279–295.
[23] IEEE Computer Society. IEEE standard for binary
floating-point arithmetic. Tech. rep., ANSI/IEEE Std
745-1985, 1985.
[24] G. Kildall. A unified approach to global program
optimization. 1st ACM POPL (1973.), 194–206.
[25] X. Leroy, D. Doligez, J. Garrigue, D. Rémy, and
J. Vouillon. The Objective Caml system,
documentation and user’s manual (release 3.06). Tech.
rep., INRIA, Rocquencourt, France, 2002.
[26] R. Manevich, G. Ramalingam, J. Field, D. Goyal, and
M. Sagiv. Compactly representing first-order
structures for static analysis. SAS (2002), LNCS 2477,
Springer, 196–212.
[27] M. Martel. Static analysis of the numerical stability of
loops. SAS (2002), LNCS 2477, Springer, 133–150.
[28] A. Miné. The octagon abstract domain library.
http://www.di.ens.fr/~ mine/oct/.
[29] A. Miné. A new numerical abstract domain based on
difference-bound matrices. PADO (2001), LNCS 2053,
Springer, 155–172.
[30] A. Miné. The octagon abstract domain. IEEE AST in
WCRE (2001), 310–319.
[31] M. Musuvathi, D. Park, A. Chou, D. Engler, and
D. Dill. Cmc: A pragmatic approach to model
checking real code. Usenix Association, OSDI 2002.
[32] S. Owre, N. Shankar, and D. Stringer-Calvert. PVS:
An experience report. FM-Trends’98 (1999), LNCS
1641, Springer, 117–132.
[33] The Coq Development Team. The Coq proof assistant
reference manual (version 7.4). Tech. rep., INRIA,
Rocquencourt, France, 2003.
[34] M. Weiser. Program slicing. IEEE Transactions on
Software Engineering SE-10, 4 (1984), 352–357.
| 6cs.PL
|
Improving SAT-solving with Machine Learning
Haoze Wu
Department of Mathematics and Computer Science
Davidson College
Davidson, NC, USA
arXiv:1710.11204v1 [cs.AI] 30 Oct 2017
[email protected]
ABSTRACT
In this project, we aimed to improve the runtime of Minisat, a Conflict-Driven Clause Learning (CDCL) solver that
solves the Propositional Boolean Satisfiability (SAT) problem. We first used a logistic regression model to predict
the satisfiability of propositional boolean formulae after fixing the values of a certain fraction of the variables in each
formula. We then applied the logistic model and added a
preprocessing period to Minisat to determine the preferable
initial value (either true or false) of each boolean variable
using a Monte-Carlo approach. Concretely, for each MonteCarlo trial, we fixed the values of a certain ratio of randomly
selected variables, and calculated the confidence that the resulting sub-formula is satisfiable with our logistic regression
model. The initial value of each variable was set based on
the mean confidence scores of the trials that started from the
literals of that variable. We were particularly interested in
setting the initial values of the backbone variables correctly,
which are variables that have the same value in all solutions
of a SAT formula. Our Monte-Carlo method was able to set
78% of the backbones correctly. Excluding the preprocessing
time, compared with the default setting of Minisat, the runtime of Minisat for satisfiable formulae decreased by 23%.
However, our method did not outperform vanilla Minisat in
runtime, as the decrease in the conflicts was outweighed by
the long runtime of the preprocessing period.
Keywords
3SAT; Satisfiability; Logistic; Monte-Carlo; Machine Learning
1.
PROBLEM AND MOTIVATION
In the Propositional Boolean Satisfiability (SAT) problem, one is given a Boolean formula, i.e., an expression that
consists of Boolean variables connected by the fundamental Boolean operators ”and”, ”or” and ”not”. One is then
tasked with determining whether there is an assignment of
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
SIGCSE ’17 March 08-11, 2017, Seattle, WA, USA
c 2017 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-4698-6/17/03.
DOI: http://dx.doi.org/10.1145/3017680.3022464
true/false values to the variables such that the overall formula evaluates to true. For instance, P ∧ (Q ∨ ¬R), is a
Boolean formula with three Boolean variables. This formula evaluates to true when P , Q, and R are all set to true.
Most state-of-the-art complete SAT solvers use the ConflictDriven Clause Learning (CDCL) algorithm. Though able to
solve large industrial formulae, the CDCL algorithm cannot efficiently solve random formulae of even moderate size
(300-500 variables). Our goal is to improve the runtime of
Minisat, a CDCL SAT solver, on random Boolean formulae,
with machine learning.
2. BACKGROUND AND RELATED WORK
The SAT problem is one of the most studied NP-complete
problems because of its theoretical significance and practical
applications[6]. Our work is inspired by the works of Xu, et
al. who successfully applied machine learning to SAT problems. In [5], they used a machine learning algorithm to
create an ensemble SAT-solver, SATzilla, which uses several
SAT-solvers as subroutines and makes a per-instance decision on which solver to pick for a given input formula. In
another work [4], Xu, et al used machine learning to predict
the satisfiability of hard Boolean formulae and obtained 70%
accuracy. In our project, instead of using machine learning to optimize the selection of SAT-solvers, we aimed to
exploit the learnability of satisfiability to improve the performance of a particular solver, Minisat [1]. Minisat uses
the CDCL algorithm, which selects a variable from the formula to ”branch on”, fixes its value to either true or false,
simplifies the formula with this assignment, and then recursively solves the rest of the formula. When a conflict (i.e.,
a contradiction) is detected at a certain branching level, the
algorithm backtracks. Though Minisat uses a heuristic to
optimize the selection of the branching variable[1], it does
not make intelligent choices when assigning values to these
variables. We aim to use machine learning to set the default
branching values of variables before running the CDCL algorithm, in order to decrease the number of conflicts, and
thus decrease the runtime of Minisat.
3. APPROACH AND UNIQUENESS
Our project relied on the hypothesis that if at each branching point, the branching variable is set to the value that is
more likely to lead to a solution, then a solution would be
found relatively quickly.
We first trained a logistic regression model to predict the
satisfiability of random Boolean formulae represented in conjunctive normal form(CNF). A CNF formula is composed of
1
2
3
4
5
6
7
8
9
10
Clause to variable ratio
Fraction of binary clauses
Fraction of horn clauses
POSNEG ratio var max
POSNEG ratio var min
POSNEG ratio var mean
POSNEG ratio var std
POSNEG ratio var variation
LPSLACK mean
LPSLACK coeff variation
Table 1: We refer the reader to [4] for the definitions
of features 4-10
a conjunction of clauses, where each clause is the disjunction of variables (or their negations). A 3-CNF formula is
one in which each clause contains exactly three variables (or
their negations) — for example, (x1 ∨ x2 ∨ x4 ) ∧ (¬x2 ∨ x3 ∨
¬x4 ) ∧ (x1 ∨ ¬x3 ∨ ¬x4 ). There is a well-known polynomialtime procedure for converting arbitrary SAT formulae into
3-CNF form[3], so this restriction in clause length does not
compromise generality.
We created the training data for our learning model by
first generating random 3-CNF formulae with 300 variables.
Since we planned to apply our fitted model to make predictions about the formulae that are generated by the CDCL
process, we generated additional instances by randomly fixing n% of the variables in these formulae, for varying values
of n. We extracted 10 features from each of these formulae to create individual training examples — these features
are listed in Table 1. The selection of the features was inspired by [4]. The target variable is a binary value that
indicated whether the formula in question was satisfiable.
Building this regression model constitutes the offline phase
of our new solver.
In the online phase of the solver, we added a preprocessing
period to Minisat to determine the preferred initial value of
each Boolean variable with a Monte-Carlo approach. Concretely, to determine the preferred initial value of a variable
v, we first set v to true and conducted a Monte-Carlo simulation starting from the remaining (simplified) formula. In
each Monte-Carlo trial, we randomly fixed the values of n%
of the variables, and calculated the likelihood that the resulting formula was satisfiable, with our logistic regression
model. A mean likelihood score was computed by averaging
the scored from the outcomes of many trials. We repeated
the same operations with v set to false. We then set v to the
value (i.e., true or false) that suggested a higher chance of
leading to a satisfiable solution, based on the outcome of the
Monte-Carlo simulations. We were particularly interested to
gauge the effectiveness of our model in correctly setting the
values of so-called ”backbone variables” — those variables
that have the same value in all solutions of a SAT formula
[2].
We evaluated the performance of the preprocessing period by the ratio of backbones it set correctly, the number
of conflicts the subsequent CDCL run yielded, and the actual runtime of Minisat. Since our method was geared towards finding solutions quickly, we focused our attention on
satisfiable Boolean formulae.
n
4%
2%
0%
SAT Prediction Score
77.94%
68.56%
70.38%
Backbone setting Score
77.56%
77.56%
68.00%
Table 2: The performance of SAT prediction and
backbone setting with different depths of MonteCarlo trails
4. RESULTS AND CONTRIBUTIONS
As shown by Table 2, our regression model achieved an
accuracy of 70% in predicting satisfiability when n was 0%,
and an accuracy of 78% when n was set to 4%. The best
performance of the Monte-Carlo method was obtained when
n was set to 4%. On average, it set 78% of the backbones
correctly.
Compared with the default setting of Minisat (which always sets the branching variable to false), our method yielded
an average decrease in conflicts of 23% and outperformed default Minisat in terms of conflicts in 55% of the test cases.
However, it did not outperform vanilla Minisat in runtime,
as the decrease in the conflicts was outweighed by the long
runtime of the preprocessing period.
5. REFERENCES
[1] N. Eén and N. Sörensson. An extensible sat-solver.
pages 502–518, May 2003.
[2] P. Kilby, J. Slaney, S. Thiébaux, T. Walsh, et al.
Backbones and backdoors in satisfiability. 2005.
[3] T. J. Schaefer. The complexity of satisfiability
problems. In Proceedings of the tenth annual ACM
symposium on Theory of computing, pages 216–226.
ACM, 1978.
[4] L. Xu, H. H. Hoos, and K. Leyton-Brown. Predicting
satisfiability at the phase transition. In Proceedings of
the Twenty-Sixth AAAI Conference on Artificial
Intelligence, AAAI’12, pages 584–590. AAAI Press,
2012.
[5] L. Xu, F. Hutter, H. H. Hoos, and K. Leyton-Brown.
Satzilla: portfolio-based algorithm selection for sat.
Journal of Artificial Intelligence Research, 32:565–606,
2008.
[6] L. Zhang, C. F. Madigan, M. H. Moskewicz, and
S. Malik. Efficient conflict driven learning in a boolean
satisfiability solver. In Proceedings of the 2001
IEEE/ACM international conference on
Computer-aided design, pages 279–285. IEEE Press,
2001.
| 2cs.AI
|
arXiv:1304.6274v2 [cs.PL] 29 Jul 2013
Interprocedural Data Flow Analysis
in Soot using Value Contexts
Rohan Padhye
Uday P. Khedker
Indian Institute of Technology Bombay
[email protected]
Indian Institute of Technology Bombay
[email protected]
Abstract
An interprocedural analysis is precise if it is flow sensitive and fully
context-sensitive even in the presence of recursion. Many methods
of interprocedural analysis sacrifice precision for scalability while
some are precise but limited to only a certain class of problems.
Soot currently supports interprocedural analysis of Java programs using graph reachability. However, this approach is restricted
to IFDS/IDE problems, and is not suitable for general data flow
frameworks such as heap reference analysis and points-to analysis
which have non-distributive flow functions.
We describe a general-purpose interprocedural analysis framework for Soot using data flow values for context-sensitivity. This
framework is not restricted to problems with distributive flow functions, although the lattice must be finite. It combines the key ideas
of the tabulation method of the functional approach and the technique of value-based termination of call string construction.
The efficiency and precision of interprocedural analyses is heavily affected by the precision of the underlying call graph. This is
especially important for object-oriented languages like Java where
virtual method invocations cause an explosion of spurious call
edges if the call graph is constructed naively. We have instantiated
our framework with a flow and context-sensitive points-to analysis
in Soot, which enables the construction of call graphs that are far
more precise than those constructed by Soot’s SPARK engine.
Categories and Subject Descriptors F.3.2 [Logics and Meanings
of Programs]: Semantics of Programming Languages—Program
Analysis
General Terms Algorithms, Languages, Theory
Keywords Interprocedural analysis, context-sensitive analysis,
points-to analysis, call graph
1.
Introduction
Interprocedural data flow analysis incorporates the effects of procedure calls on the callers and callees. A context-insensitive analysis
does not distinguish between distinct calls to a procedure. This
causes the propagation of data flow values across interprocedurally
invalid paths (i.e. paths in which calls and returns may not match)
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. To copy otherwise, to republish, to post on servers or to redistribute
to lists, requires prior specific permission and/or a fee.
SOAP ’13 June 20, 2013, Seattle, Washington, USA.
Copyright c 2013 ACM ISBN 978-1-4503-2201-0/13/06. . . $15.00
resulting in a loss of precision. A context-sensitive analysis restricts
the propagation to valid paths and hence is more precise.
Two most general methods of precise flow and context-sensitive
analysis are the Functional approach and the Call Strings approach [11]. The functional approach constructs summary flow
functions for procedures by reducing compositions and meets of
flow functions of individual statements to a single flow function,
which is used directly in call statements. However, constructing
summary flow functions may not be possible in general. The
tabulation method of the functional approach overcomes this
restriction by enumerating the functions as pairs of input-output
data flow values for each procedure, but requires a finite lattice.
The call strings method remembers calling contexts in terms of
unfinished calls as call strings. However, it requires an exponentially large number of call strings. The technique of value based
termination of call string construction [5] uses data flow values to
restrict the combinatorial explosion of contexts and improves the
efficiency significantly without any loss of precision.
Graph reachability based interprocedural analysis [9, 10] is a
special case of the functional approach. Formally, it requires flow
functions 2A 7→ 2A to distribute over the meet operation so that
they can be decomposed into meets of flow functions A 7→ A.
Here A can be either a finite set D (for IFDS problems [9]) or a
mapping D 7→ L (for IDE problems [10]) from a finite set D to a
lattice of values L. Intuitively, A represents a node in the graph
and a function A 7→ A decides the nature of the edge from the
node representing the argument to the node representing the result.
Flow function composition then reduces to a transitive closure of
the edges resulting in paths in the graph.
The efficiency and precision of interprocedural analyses is heavily affected by the precision of the underlying call graph. This is
especially important for object oriented languages like Java where
virtual method invocations cause an explosion of spurious call
edges if the call graph is constructed naively.
Soot [12] has been a stable and popular choice for hundreds
of client analyses for Java programs, though it has traditionally
lacked an interprocedural framework. Bodden [4] has recently
implemented support for interprocedural analysis using graph
reachability. The main limitation of this approach is that it is not
suitable for general data flow frameworks with non-distributive
flow functions such as heap reference analysis or points-to analysis.
For example, consider the Java statement x = y.n to be processed
for points-to analysis. If we have points-to edges y → o1 and
o1 .n → o2 before the statement (where o1 and o2 are heap objects),
then it is not possible to correctly deduce that the edge x → o2
should be generated after the statement if we consider each input
edge independently. The flow function for this statement is a function of the points-to graph as a whole and cannot be decomposed
into independent functions of each edge and then merged to get a
correct result.
We have implemented a generic framework for performing flow
and context-sensitive interprocedural data flow analysis that does
not require flow functions to be distributive. However, the flow
functions must be monotonic and the lattice of data flow values
must be finite. The framework uses value-based contexts and is an
adaptation of the tabulation method of the functional approach and
the modified call strings method.
Our implementation is agnostic to any analysis toolkit or intermediate representation as it is parameterized using generic types.
Since our core classes are similar to the intra-procedural framework
of Soot, it integrates with Soot’s Jimple IR seamlessly.
We have instantiated our framework with a flow and contextsensitive points-to analysis in Soot, which enables the construction
of call graphs that are far more precise than those constructed by
Soot’s SPARK engine.
The rest of the paper is organized as follows: Section 2 describes
our method. Section 3 outlines the API of our implementation
framework. Section 4 presents the results of call graph construction. Finally, Section 5 concludes the paper by describing the current status and future possibilities of our work.
2.
Interprocedural Analysis Using Value Contexts
The tabulation method of the functional approach [11] and the
modified call strings approach [5] both revolve around the same
key idea: if two or more calls to a procedure p have the same the
data flow value (say x) at the entry of p, then all of them will
have an identical data flow value (say y) at the exit of p. The
tabulation method uses this idea to enumerate flow functions in
terms of pairs of input-output values (x, y) whereas the modified
call strings method uses it to partition call strings based on input
values, reducing the number of call strings significantly.
The two methods lead to an important conclusion: Using data
flow values as contexts of analysis can avoid re-analysis of procedure bodies. We make this idea explicit by defining a value context
X = hmethod, entryValuei, where entryValue is the data flow
value at the entry to a procedure method. Additionally, we define a mapping exitValue(X) which gives the data flow value at
the exit of method. As data flow analysis is an iterative process,
this mapping may change over time (although it will follow a
descending chain in the lattice). The new value is propagated to
all callers of method if and when this mapping changes. With this
arrangement, intraprocedural analysis can be performed for each
value context independently, handling flow functions in the usual
way; only procedure calls need special treatment.
Although the number of value contexts created per procedure is
theoretically proportional to the size of the lattice in the worst-case,
we have found that in practice the number of distinct data flow values reaching each procedure is often very small. This is especially
true for heap-based analyses that use bounded abstractions, due to
the locality of references in recursive paths. This claim is validated
is Section 4, in which we present the results of a points-to analysis.
Algorithm
Figure 1 provides the overall algorithm. Line 1 declares three globals: a set of contexts that have been created, a transition table mapping a context and call site of a caller method to a target context at
the called method and a work-list of context-parametrized controlflow graph nodes whose flow function has to be processed.
The procedure INIT C ONTEXT (lines 2-11) initializes a new
context with a given method and entry value. The exit value is
initialized to the > element. IN/OUT values at all nodes in the
method body are also initialized to >, with the exception of the
method’s entry node, whose IN value is initialized to the context’s
entry value. All nodes of this context are added to the work-list.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
50:
51:
global contexts, transitions, worklist
procedure INIT C ONTEXT(X)
ADD (contexts, X)
Set EXIT VALUE(X) ← >
Let m ← METHOD(X)
for all nodes n in the body of m do
ADD (worklist, hX, ni)
Set IN(X, n) ← > and OUT(X, n) ← >
end for
Set IN(X, ENTRY N ODE(m)) ← ENTRY VALUE(X)
end procedure
procedure DOA NALYSIS
INIT C ONTEXT (hmain, BIi)
while worklist is not empty do
Let hX, ni ← REMOVE N EXT(worklist)
if n is not the entry node then
Set IN(X, n) ← >
for all predecessors p of n do
Set IN(X, n) ← IN(X, n) u OUT(X, p)
end for
end if
Let a ← IN(X, n)
if n contains a method call then
Let m ← TARGET M ETHOD(n)
Let x ← CALL E NTRY F LOW F UNCTION(X, m, n, a)
Let X 0 ← hm, xi
. x is the entry value at m
Add an edge hX, ni → X 0 to transitions
if X 0 ∈ contexts then
Let y ← EXIT VALUE(X 0 )
Let b1 ← CALL E XIT F LOW F UNCTION(X, m, n, y)
Let b2 ← CALL L OCAL F LOW F UNCTION(X, n, a)
Set OUT(X, n) ← b1 u b2
else
0
INIT C ONTEXT (X )
end if
else
Set OUT(X, n) ← NORMAL F LOW F UNCTION(X, n, a)
end if
if OUT(X, n) has changed then
for all successors s of n do
ADD (worklist, hX, si)
end for
end if
if n is the exit node then
Set EXIT VALUE(X) ← OUT(X, n)
for all edges hX 0 , ci → X in transitions do
0
ADD (worklist, hX , ci)
end for
end if
end while
end procedure
Figure 1. Algorithm for performing inter-procedural analysis
using value contexts.
The DOA NALYSIS procedure (lines 12-51) first creates a value
context for the main method with some boundary information (BI).
Then, data flow analysis is performed using the traditional work-list
method, but distinguishing between nodes of different contexts.
A node is removed from the work-list and its IN value is set to
the meet of the OUT values of its predecessors (lines 16-21). For
nodes without a method call, the OUT value is computed using the
normal flow function (line 37). For call nodes, parameter passing is
handled by a call-entry flow function that takes as input the IN value
at the node, and the result of which is used as the entry value at the
>
hX1 , a+ b− i
hX3 , a− b+ i
hX0 , >i
n1
p = 5
n2
hX1 , a+ b− i
hX3 , a− b+ i
c1 q = f(p, -3) n3 c = a * b
c2 c = g(10)
hX1 , a+ b− c− i
hX3 , a− b+ c− i
hX0 , p+ q − i
c4 r = g(-q)
exit
hX2 , u+ v − i
n6
return v
hX1 , a+ b− c− i
hX3 , a− b+ c− i
n4
hX1 , a+ b− c− i
hX3 , a− b+ c− i
hX0 , p+ q − r − i
n5
−
0
+
hX2 , u+ i
c3 v = f(-u, u)
if (...)
hX1 , a+ b− i
hX3 , a− b+ i
hX0 , p+ i
n6
g(u)
f(a, b)
main()
⊥
(b) Lattice for a single variable
Context
X0
X1
X2
X3
Proc.
main
f
g
f
Entry
>
a+ b−
u+
− +
a b
Exit
p+ q − r −
a+ b− c−
u+ v −
a− b+ c−
(c) Value contexts for the program
c1
c2
c3
return c
X0
(a) Control flow graphs annotated with context-sensitive data flow values
X1
c4
X2
X3
c2
(d) Context transition diagram
Figure 2. A motivating example of a non-distributive sign-analysis performed on a program with mutually recursive procedures.
callee context (lines 24-26). The transition from caller context and
call-site to callee context is also recorded (line 27). If a context with
the target method and computed entry value has not been previously
created, then it is initialized now (line 34). Otherwise, the exit value
of the target context is used as the input to a call-exit flow function,
to handle returned values. A separate call-local flow function takes
as input the IN value at the call node, and propagates information
about local variables. The results of these two functions are merged
into the OUT value of the call node (lines 29-32).
Once a node is processed, its successors are added to the worklist if its OUT value has changed in this iteration (lines 39-43). If
the node is the exit of its procedure (lines 44-49), then the exit value
of its context is set and all its callers are re-added to the work-list.
The termination of the algorithm follows from the monotonicity
of flow functions and the finiteness of the lattice (which bounds the
descending chain as well as the number of value contexts).
The algorithm can easily be extended to handle multiple entry/exit points per procedure as well as virtual method calls by
merging data flow values across these multiple paths. It can also
be easily adapted for backward data flow analyses.
Example
Consider the program in Figure 2 (a), for which we wish to perform
a simplified sign analysis, to determine whether a scalar local
variable is negative, positive or zero. The call from main to f at c1
will only return when the mutual recursion of f and g terminates,
which happens along the program path n2 n3 n4 n5 . Notice that the
arguments to f at call-site c3 are always of opposite signs, causing
the value of variable c to be negative after every execution of n3 in
this context. Thus, f and hence g always returns a negative value.
To compute this result using the algorithm described above, we
use data flow values that are elements of the lattice in Figure 2 (b),
where > indicates an uninitialized variable and ⊥ is the conservative assumption. We use superscripts to map variables to a sign or
⊥, and omit uninitialized variables.
At the start of the program no variables are initialized and hence
the analysis starts with the initial value context X0 = hmain, >i.
For work-list removal, we will use lexicographical ordering of
contexts (newer first) before nodes (reverse post-order).
The flow function of hX0 , n1 i is processed first, which makes
p positive (written as p+ ). The next node picked from the work-list
is c1 , whose call-entry flow function passes one positive and one
negative argument to parameters a and b of procedure f respectively. Thus, a new value context X1 = hf, a+ b− i is created and
the transition hX0 , c1 i → X1 is recorded.
Analysis proceeds by processing hX1 , n2 i and then hX1 , c2 i,
which creates a new value context X2 = hg, u+ i due to the
positive argument. The transition (X1 , c2 ) → X2 is recorded.
When hX2 , c3 i is processed, the arguments to f are found to be
negative and positive respectively, creating a new value context
X3 = hf, a− b+ i and a transition (X2 , c3 ) → X3 .
The work-list now picks nodes of context X3 , and when
hX3 , c2 i is processed, the entry value at g is u+ , for which a value
context already exists – namely X2 . The transition hX3 , c2 i → X2
is recorded. The exit value of X2 is at the moment > because its
exit node has not been processed. Hence, the call-exit flow function
determines the returned value to be uninitialzed and the OUT of
hX3 , c2 i gets the value a− b+ . The next node to be processed
is hX3 , n3 i, whose flow function computes the sign of c to be
negative as it is the product of a negative and positive value. The
IN value at hX3 , n4 i is (a− b+ c− u a− b+ ) = a− b+ c− . Thus, the
sign of the returned variable c is found to be negative. As n4 is the
exit node of procedure f, the callers of X3 are looked up in the
transition table and added to the work-list.
The only caller hX2 , c3 i is now re-processed, this time resulting
in a hit for an existing target context X3 . The exit value of X3
being a− b+ c− , the returned variable v gets a negative sign, which
propagates to the exit node n6 . The callers of X2 , namely hX1 , c2 i
and hX3 , c2 i, are re-added to the work-list.
hX3 , c2 i is processed next, and this time the correct exit value of
target context X2 , which is u+ v − , is used and the OUT of hX3 , c2 i
is set to a− b+ c− . When its successor hX3 , n4 i is subsequently
processed, the OUT value does not change and hence no more
nodes of X3 are added to the work-list. Analysis continues with
nodes of X1 on the work-list, such as hX1 , c2 i and hX1 , n3 i. The
sign of c is determined to be negative and this propagates to the end
of the procedure. When exit node hX1 , n5 i is processed, the caller
of X1 , namely hX0 , c1 i, is re-added to the work-list. Now, when
this node is processed, q is found to be negative.
Value-based contexts are not only useful in terminating the analysis of recursive procedures, as shown above, but also as a simple
cache table for distinct call sites. For example, when hX0 , c4 i is
InterProceduralAnalysis<M,N,A>
+
+
+
+
+
+
+
+
+
+
+
+
Context<M,N,A>
topValue() : A
boundaryValue(M) : A
copy(A) : A
meet(A,A) : A
normalFlowFunction(Context<M,N,A>, N, A) : A
callEntryFlowFunction(Context<M,N,A>, M, N, A) : A
callExitFlowFunction(Context<M,N,A>, M, N, A) : A
callLocalFlowFunction(Context<M,N,A>, N, A) : A
programRepresentation() : ProgramRepresentation<M,N>
doAnalysis() : void
getContexts() : Map<M,List<Context<M,N,A>>>
getMeetOverPathsSolution() : DataFlowSolution<M,N,A>
ForwardInterProceduralAnalysis<M,N,A>
+ doAnalysis() : void
+
+
+
+
+
getMethod(): M
getEntryValue() : A
getExitValue() : A
getValueBefore(N) : A
getValueAfter(N) : A
ProgramRepresentation<M,N>
+
+
+
+
getEntryPoints() : List<M>
getControlFlowGraph(M) : DirectedGraph<N>
isCall(N) : boolean
resolveTargets(M, N) : List<M>
BackwardInterProceduralAnalysis<M,N,A>
+ doAnalysis() : void
Figure 3. The class diagram of our generic interprocedural analysis framework.
processed, the positive argument results in a hit for X2 , and thus
its exit value is simply re-used to determine that r is negative.
Figure 2 (b) lists the value contexts for the program and Figure 2 (c)
shows the transitions between contexts at call-sites.
A context-insensitive analysis would have merged signs of a
and b across all calls to f and would have resulted in a ⊥ value for
the signs of c, v, q and r. Our context-sensitive method ensures a
precise data flow solution even in the presence of recursion.
Notice that the flow function for n3 is non-distributive since
fn3 (a+ b− ) u fn3 (a− b+ ) = a+ b− c− u a− b+ c− = a⊥ b⊥ c−
but fn3 (a+ b− u a− b+ ) = fn3 (a⊥ b⊥ ) = a⊥ b⊥ c⊥ . Hence this
problem does not fit in the IFDS/IDE framework, but such flow
functions do not pose a problem to our algorithm.
3.
Implementation Framework
The implementation framework consists of a handful of core
classes as shown in Figure 3. The use of generic types makes the
framework agnostic to any particular toolkit or IR. The classes are
parameterized by three types: M represents the type of a method,
N represents a node in the control flow graph and A is the type
of data flow value used by the client analysis. The framework
can be naturally instantiated for Soot using the type parameters
SootMethod and Unit for M and N respectively.
Users would extend ForwardInterProceduralAnalysis or
BackwardInterProceduralAnalysis, which are subclasses of
an abstract class InterProceduralAnalysis. The abstract methods topValue, boundaryValue, copy and meet provide a hook
for client analyses to express initial lattice values and basic operations on them. The major functionality of the client analysis
would be present in the *FlowFunction methods, whose roles
were explained in Section 2. Additionally, clients are expected
to provide a ProgramRepresentation object, which specifies
program entry points (for which boundary values are to be defined)
and resolves virtual calls. Our framework ships with default program representations for Soot’s Jimple IR. The launch point of the
analysis is the doAnalysis method, which is implemented as per
the algorithm from Figure 1 in the directional sub-classes.
The Context class encapsulates information about a value context. Every context is associated with a method, an entry value and
an exit value, each of which can be retrieved using the corresponding getter methods. The getValueBefore and getValueAfter
methods return data flow values for a context just before and after a
node respectively. This is the recommended way for accessing the
results of the analysis in a context-sensitive manner. A mapping
of methods to a list of all its contexts is available through the
getContexts method of the InterProceduralAnalysis class.
Alternatively, getMeetOverValidPathsSolution can be used to
obtain a solution that is computed by merging data flow results
across all contexts of each method. The DataFlowSolution class
(not shown in the figure) simply provides getValueBefore and
getValueAfter methods to access the resulting solution.
4.
The Role of Call Graphs
We initially developed this framework in order to implement heap
reference analysis [6] using Soot, because it could not be encoded as an IFDS/IDE problem. However, even with our general
framework, performing whole-program analysis turned out to be
infeasible due to a large number of interprocedural paths arising
from conservative assumptions for targets of virtual calls.
The SPARK engine [7] in Soot uses a flow and context insensitive pointer analysis on the whole program to build the call graph,
thus making conservative assumptions for the targets of virtual
calls in methods that are commonly used such as those in the Java
library. For example, it is not uncommon to find call sites in library
methods with 5 or more targets, most of which will not be traversed
in a given context. Some call sites can even be found with more than
250 targets! This is common with calls to virtual methods defined
in java.lang.Object, such as hashCode() or equals().
When performing whole-program data flow analysis, the use
of an imprecise call graph hampers both efficiency, due to an
exponential blow-up of spurious paths, and precision, due to the
meet over paths that are actually interprocedurally invalid, thereby
diminishing the gains from context-sensitivity.
Soot provides a context-sensitive call graph builder called PAD DLE [8], but this framework can only perform k-limited call-site or
object-sensitive analysis, and that too in a flow-insensitive manner.
We were unable to use PADDLE with our framework directly because at the moment it not clear to us how the k-suffix contexts of
PADDLE would map to our value-contexts.
Call Graph Construction using Points-To Analysis
We have implemented a flow and context-sensitive points-to analysis using our interprocedural framework to build a call graph onthe-fly. This analysis is both a demonstration of the use of our
framework as well as a proposed solution for better call graphs
intended for use by other interprocedural analyses.
Benchmark
SPEC JVM98
DaCapo 2006
compress
jess
db
mpegaudio
jack
antlr
chart
Time
1.15s
140.8s
2.19s
4.51s
89.33s
697.4s
242.3s
Methods (M )
Total App.
367
54
690
328
420
56
565
245
721
288
1,406
798
1,799
598
Contexts (X)
Total
App.
1,550
70
17,280
9,397
2,456
159
2,397
705
7,534
2,548
30,043 21,599
16,880
4,326
X/M
Total
App.
4.22
1.30
25.04 28.65
5.85
2.84
4.24
2.88
10.45
8.85
21.37 27.07
9.38
7.23
Clean
Total App.
50
47
34
30
62
46
50
47
273
270
769
727
458
423
Table 1. Results of points-to analysis using our framework. “App.” refers to data for application classes only.
The data flow value used in our analysis is a points-to graph
in which nodes are allocation sites of objects. We maintain two
types of edges: x → m indicates that the root variable x may
point to objects allocated at site m, and m.f → n indicates that
objects allocated at site m may reference objects allocated at site
n along the field f . Flow functions add or remove edges when
processing assignment statements involving reference variables.
Nodes that become unreachable from root variables are removed.
Type consistency is maintained by propagating only valid casts.
The points-to graphs at each statement only maintain objects
reachable from variables that are local to the method containing the
statement. At call statements, we simulate assignment of arguments
to locals of the called method, as well as the assignment of returned
values to a local of the caller method. For static fields (and objects
reachable from them) we maintain a global flow-insensitive pointsto graph. For statements involving static loads/stores we operate
on a temporary union of local and global graphs. The call graph
is constructed on-the-fly by resolving virtual method targets using
type information of receiver objects.
Points-to information cannot be precise for objects returned by
native methods, and for objects shared between multiple threads (as
our analysis is flow-sensitive). Thus, we introduce the concept of a
summary node, which represents statically unpredictable points-to
information and is denoted by the symbol ⊥. For soundness, we
must conservatively propagate this effect to variables and fields that
involve assignments to summary nodes. The rules for summarization along different types of assignment statements are as follows:
Statement
x=y
x.f = y
x = y.f
x = p(a1 , a2 , ...)
Rule used in the flow function
If y → ⊥, then set x → ⊥
If y → ⊥, then ∀o : x → o, set o.f → ⊥
If y → ⊥ or ∃o : y → o and o.f → ⊥,
then set x → ⊥
If p is unknown, then set x → ⊥, and
∀o : ai → o, ∀f ∈ f ields(o) set o.f → ⊥
The last rule is drastically conservative; for soundness we must
assume that a call to an unknown procedure may modify the fields
of arguments in any manner, and return any object. An important
discussion would be on what constitutes an unknown procedure.
Native methods primarily fall into this category. In addition, if p is
a virtual method invoked on a reference variable y and if y → ⊥,
then we cannot determine precisely what the target for p will be.
Hence, we consider this call site as a default site, and do not enter
the procedure, assuming worst-case behaviour for its arguments
and returned values. A client analysis using the resulting call graph
with our framework can choose to do one of two things when
encountering a default call site: (1) assume worst case behaviour for
its arguments (eg. in liveness analysis, assume that all arguments
and objects reachable from them are live) and carry on to the next
statement, or (2) fall-back onto Soot’s default call graph and follow
the targets it gives.
A related approach partitions a call graph into calls from application classes and library classes [2]. Our call graph is partitioned
into call sites that we can precisely resolve to one or more valid
targets, and those that cannot due to statically unpredictable factors.
Experimental Results
Table 1 lists the results of points-to analysis performed on seven
benchmarks. The experiments were carried out on an Intel Core i7960 with 19.6 GB of RAM running Ubuntu 12.04 (64-bit) and JDK
version 1.6.0 27. Our single-threaded analysis used only one core.
The first two columns contain the names of the benchmarks;
five of which are the single-threaded programs from the SPEC
JVM98 suite [1], while the last two are from the DaCapo suite [3]
version 2006-10-MR2. The third column contains the time required
to perform our analysis, which ranged from a few seconds to a
few minutes. The fourth and fifth columns contain the number of
methods analyzed (total and application methods respectively). The
next two columns contain the number of value-contexts created,
with the average number of contexts per method in the subsequent
two columns. It can be seen that the number of distinct data flow
values reaching a method is not very large in practice. As our
analysis ignores paths with method invocations on null pointers, it
was inappropriate for other benchmarks in the DaCapo suite when
using stub classes to simulate the suite’s reflective boot process.
The use of default sites in our call graph has two consequences:
(1) the total number of analyzed methods may be less than the
total number of reachable methods and (2) methods reachable from
default call sites (computed using SPARK’s call graph) cannot be
soundly optimized by a client analysis that jumps over these sites.
The last column lists the number of clean methods which are not
reachable from default sites and hence can be soundly optimized.
In all but two cases, the majority of application methods are clean.
In order to highlight the benefits of using the resulting call
graph, just listing the number of edges or call-sites alone is not
appropriate, as our call graph is context-sensitive. We have thus
computed the number distinct paths in the call graph, starting from
the entry point, which are listed in Table 2. As the total number
of call graph paths is possibly infinite (due to recursion), we have
counted paths of a fixed length length k, for 1 ≤ k ≤ 10. For
each benchmark, we have counted these paths using call graphs
constructed by our Flow and Context-sensitive Pointer Analysis
(FCPA) as well as SPARK, and noted the difference as percentage
savings (∆%) from using our context-sensitive call graph. The
option implicit-entry was set to false for SPARK.
The savings can be clearly observed for k > 5. For k = 10,
SPARK ’s call graph contains more than 96% spurious paths for three
of the benchmarks, and 62-92% for the remaining. The gap only
widens for larger values of k (for which the number of paths was
too large to compute in some cases).
Client analyses using our interprocedural framework can be
configured to use our context-sensitive call graphs which avoid
these spurious paths, hence enabling efficient and precise solutions.
Depth k =
compress
FCPA
SPARK
jess
FCPA
SPARK
db
FCPA
SPARK
mpegaudio
FCPA
SPARK
jack
FCPA
SPARK
antlr
FCPA
SPARK
chart
FCPA
SPARK
∆%
∆%
∆%
∆%
∆%
∆%
∆%
1
2
2
0
2
2
0
2
2
0
2
2
0
2
2
0
6
6
0
6
6
0
2
5
5
0
5
5
0
5
5
0
14
16
12
18
18
0
24
24
0
24
24
0
3
7
9
22.2
7
9
22.2
11
13
15.4
42
46
8.7
106
106
0
202
206
1.9
217
219
0.9
4
20
22
9.09
30
32
6.25
46
48
4.17
113
118
4.24
1,560
1,577
1.08
560
569
1.58
696
714
2.52
5
55
57
3.51
127
149
14.77
258
443
41.76
804
834
3.60
22,652
27,201
16.72
1,651
1,669
1.08
2,109
2,199
4.09
6
263
273
3.66
470
924
49.13
1,791
4,726
62.10
11,286
15,844
28.77
235,948
356,867
33.88
4,669
9,337
49.99
9,778
20,171
51.52
7
614
1,237
50.36
4,932
24,224
79.64
21,426
71,907
70.20
129,807
250,096
48.10
2,897,687
5,583,858
48.11
18,953
107,012
82.29
45,010
306,396
85.31
8
2,225
23,426
90.50
75,112
367,690
79.57
215,465
860,851
74.97
1,772,945
4,453,608
60.19
45,480,593
104,211,833
56.36
110,228
1,669,247
93.40
517,682
7,676,266
93.26
9
21,138
545,836
96.13
970,044
8,591,000
88.71
2,687,625
13,231,026
79.69
27,959,747
87,096,135
67.90
835,791,756
2,136,873,586
60.89
975,090
27,670,645
96.48
7,796,424
192,839,216
95.96
10
202,071
12,052,089
98.32
15,052,927
196,801,775
92.35
42,842,761
245,964,733
82.58
496,420,128
1,811,902,298
72.60
17,285,586,592
46,356,206,503
62.71
11,935,918
468,973,725
97.45
164,476,462
4,996,310,985
96.71
Table 2. Number of k-length call graph paths for various benchmarks using SPARK and FCPA (Flow and Context-sensitive Pointer Analysis).
5.
Conclusion and Future Work
We have presented a framework for performing value-based contextsensitive inter-procedural analysis in Soot. This framework does
not require distributivity of flow functions and is thus applicable
to a large class of analyses including those that cannot be encoded
as IFDS/IDE problems. Another advantage of our method is the
context-sensitive nature of the resulting data flow solution, which
can be useful in dynamic optimizations.
In order to deal with the difficulties in whole-program analysis
performed over an imprecise call graph, we constructed call graphs
on-the-fly while performing a flow and context-sensitive pointsto analysis. This analysis also demonstrated a sample use of our
framework and showed that it was practical to use data flow values
as contexts because the number of distinct data flow values reaching
each method is often very small.
The interprocedural framework has been released and is available at https://github.com/rohanpadhye/vasco. However,
our points-to analysis implementation is experimental and makes
heavy use of HashMaps and HashSets, thus running out of memory
for very large programs. We would like to improve this implementation by using bit-vectors or maybe even BDDs for compact
representation of points-to sets.
The precision of our call graphs is still limited in the presence of
default sites, which are prominent due to the liberal use of summary
nodes in the points-to graphs. We would like to reduce the number of summary nodes by simulating some commonly used native
methods in the Java library, and also by preparing a summary of
initialized static fields of library classes. Our hope is that these extensions would enable our call graph to be fully complete, thereby
enabling users to precisely perform whole-program analysis and
optimization for all application methods in an efficient manner. We
believe that improving the precision of program analysis actually
helps to improve its efficiency, rather than hamper it.
Acknowledgments
We would like to thank the anonymous reviewers for suggesting
the inclusion of a non-distributive example and the separation of
call/return flow functions. These changes improved the paper.
References
[1] http://www.spec.org/jvm98. Accessed: April 3, 2013.
[2] K. Ali and O. Lhoták. Application-only call graph construction.
In Proceedings of the 26th European conference on Object-Oriented
Programming, ECOOP’12, 2012.
[3] S. Blackburn et al. The DaCapo benchmarks: Java benchmarking
development and analysis. In OOPSLA ’06: Proceedings of the 21st
annual ACM SIGPLAN conference on Object-Oriented Programing,
Systems, Languages, and Applications, Oct. 2006.
[4] E. Bodden. Inter-procedural data-flow analysis with IFDS/IDE and
Soot. In Proceedings of the ACM SIGPLAN International Workshop
on State of the Art in Java Program analysis, SOAP ’12, 2012.
[5] U. P. Khedker and B. Karkare. Efficiency, precision, simplicity,
and generality in interprocedural data flow analysis: resurrecting the
classical call strings method. In Proceedings of the Joint European
Conferences on Theory and Practice of Software 17th international
conference on Compiler construction, CC’08/ETAPS’08, 2008.
[6] U. P. Khedker, A. Sanyal, and A. Karkare. Heap reference analysis
using access graphs. ACM Trans. Program. Lang. Syst., 30(1), Nov.
2007.
[7] O. Lhoták and L. Hendren. Scaling Java points-to analysis using
SPARK. In Proceedings of the 12th international conference on
Compiler construction, CC’03, 2003.
[8] O. Lhoták and L. Hendren. Evaluating the benefits of context-sensitive
points-to analysis using a BDD-based implementation. ACM Trans.
Softw. Eng. Methodol., 18(1), Oct. 2008.
[9] T. Reps, S. Horwitz, and M. Sagiv. Precise interprocedural dataflow
analysis via graph reachability.
In Proceedings of the 22nd
ACM SIGPLAN-SIGACT symposium on Principles of programming
languages, POPL ’95, 1995.
[10] M. Sagiv, T. Reps, and S. Horwitz. Precise interprocedural dataflow
analysis with applications to constant propagation. Theor. Comput.
Sci., 167(1-2), Oct. 1996.
[11] M. Sharir and A. Pnueli. Two approaches to interprocedural dataflow
analysis. In S. S. Muchnick and N. D. Jones, editors, Program Flow
Analysis: Theory and Applications. Prentice-Hall, Inc., 1981.
[12] R. Vallée-Rai, P. Co, E. Gagnon, L. Hendren, P. Lam, and
V. Sundaresan. Soot - a Java bytecode optimization framework. In
Proceedings of the 1999 conference of the Centre for Advanced Studies
on Collaborative research, CASCON ’99, 1999.
| 6cs.PL
|
Data-Oblivious Stream Productivity
Jörg Endrullis1 , Clemens Grabmayer2, and Dimitri Hendriks1
Vrije Universiteit Amsterdam, Department of Computer Science
De Boelelaan 1081a, 1081 HV Amsterdam, The Netherlands
[email protected] [email protected]
2
Universiteit Utrecht, Department of Philosophy
Heidelberglaan 8, 3584 CS Utrecht, The Netherlands
[email protected]
Abstract. We are concerned with demonstrating productivity of specifications of infinite streams of data, based on orthogonal rewrite rules.
In general, this property is undecidable, but for restricted formats computable sufficient conditions can be obtained. The usual analysis, also
adopted here, disregards the identity of data, thus leading to approaches
that we call data-oblivious. We present a method that is provably optimal among all such data-oblivious approaches. This means that in order
to improve on our algorithm one has to proceed in a data-aware fashion.3
1
Introduction
For programming with infinite structures, productivity is what termination is for
programming with finite structures. Productivity captures the intuitive notion of
unlimited progress, of ‘working’ programs producing defined values indefinitely.
In functional languages, usage of infinite structures is common practice. For
the correctness of programs dealing with such structures one must guarantee
that every finite part of the infinite structure can be evaluated, that is, the
specification of the infinite structure must be productive.
We investigate this notion for
stream specifications, formalized as
orthogonal term rewriting systems.
P = productive
Common to all previous approaches
F = flat
for recognizing productivity is a quanOur contribution:
titative analysis that abstracts away
= automated
from the concrete values of stream
recognition
elements. We formalize this by a
= decision
notion of ‘data-oblivious’ rewriting,
P
F
¬P
¬F
and introduce the concept of dataFig. 1: Map of stream specifications
oblivious productivity. Data-oblivious
(non-)productivity implies (non-)productivity, but neither of the converse implications holds. Fig. 1 shows a Venn diagram of stream specifications, highlighting the subset of ‘data-obliviously recognizable’ specifications where (non-)
pu
re
da
t
r e a -o
co b
gn liv
iz iou
ab s
le ly
arXiv:0806.2680v5 [cs.LO] 19 Jul 2008
1
3
This research has been partially funded by the Netherlands Organisation for Scientific Research (NWO) under FOCUS/BRICKS grant number 642.000.502.
2
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
productivity can be recognized by a data-oblivious analysis.
We identify two syntactical classes of stream specifications: ‘flat’ and ‘pure’
specifications, see the description below. For the first we devise a decision algorithm for data-oblivious (d-o) productivity. This gives rise to a computable, d-o
optimal, criterion for productivity: every flat stream specification that can be established to be productive by whatever d-o argument is recognized as productive
by this criterion (see Fig. 1). For the subclass of pure specifications, we establish
that d-o productivity coincides with productivity, and thereby obtain a decision
algorithm for productivity of this class. Additionally, we extend our criterion
beyond the class of flat stream specifications, allowing for ‘friendly nesting’ in
the specification of stream functions; here d-o optimality is not preserved.
In defining the different formats of stream specifications, we distinguish between rules for stream constants, and rules for stream functions. Only the latter
are subjected to syntactic restrictions. In flat stream specifications the defining
rules for the stream functions do not have nesting of stream function symbols;
however, in defining rules for stream constants nesting of stream function symbols is allowed. This format makes use of exhaustive pattern matching on data
to define stream functions, allowing for multiple defining rules for an individual stream function symbol. Since the quantitative consumption/production behaviour of a symbol f might differ among its defining rules, in a d-o analysis one
has to settle for the use of lower bounds when trying to recognize productivity.
If for all stream function symbols f in a flat specification T the defining rules for
f coincide, disregarding the identity of data-elements, then T is called pure.
Our decision algorithm for d-o productivity determines the tight d-o lower
bound on the production behaviour of every stream function, and uses these
bounds to calculate the d-o production of stream constants. We briefly explain
both aspects. Consider the stream specification A → 0 : f(A) together with the
rules f(0 : σ) → 1 : 0 : 1 : f(σ), and f(1 : σ) → 0 : f(σ), defining the stream
0 : 1 : 0 : 1 : . . . of alternating bits. The tight d-o lower bound for f is the function
id : n 7→ n. Further note that suc: n 7→ n + 1 captures the quantitative behaviour
of the function prepending a data element to a stream term. Therefore the d-o
production of A can be computed as lfp(suc ◦ id ) = ∞, where lfp(f ) is the
least fixed point of f : N → N and N := N ∪ {∞}; hence A is productive. As
a comparison, only a ‘data-aware’ approach is able to establish productivity of
B → 0 : g(B) with g(0 : σ) → 1 : 0 : g(σ), and g(1 : σ) → g(σ). The d-o lower
bound of g is n 7→ 0, due to the latter rule. This makes it impossible for any
conceivable d-o approach to recognize productivity of B.
We obtain the following results:
(i) For the class of flat stream specifications we give a computable, d-o optimal,
sufficient condition for productivity.
(ii) We show decidability of productivity for the class of pure stream specifications, an extension of the format in [2].
(iii) Disregarding d-o optimality, we extend (i) to the bigger class of friendly
nesting stream specifications.
(iv) A tool automating (i), (ii), and (iii), which can be downloaded from, and
used via a web interface at: http://infinity.few.vu.nl/productivity.
Data-Oblivious Stream Productivity
3
Related work. Previous approaches [8, 3, 9, 1] employed d-o reasoning (without
using this name for it) to find sufficient criteria ensuring productivity, but did
not aim at optimality. The d-o production behaviour of a stream function f
is bounded from below by a ‘modulus of production’ ν f : Nk → N with the
property that the first ν f (n1 , . . . , nk ) elements of f(t1 , . . . , tk ) can be computed
whenever the first ni elements of ti are defined. Sijtsma develops an approach
allowing arbitrary production moduli ν : Nk → N, which, while providing an
adequate mathematical description, are less amenable to automation. Telford
and Turner [9] employ production moduli of the form ν(n) = n + a with a ∈ Z.
Hughes, Pareto and Sabry [3] use ν(n) = max{c·x+d | x ∈ N, n ≥ a·x+b}∪{0}
with a, b, c, d ∈ N. Both classes of production moduli are strictly contained in
the class of ‘periodically increasing’ functions which we employ in our analysis.
We show that the set of d-o lower bounds of flat stream function specifications
is exactly the set of periodically increasing functions. Buchholz [1] presents a
type system for productivity, using unrestricted production moduli. To obtain
an automatable method for a restricted class of stream specifications, he devises
a syntactic criterion to ensure productivity. This criterion easily handles all the
examples of [9], but fails to deal with functions that have a negative effect like odd
defined by odd(x : y : σ) → y : odd(σ) with a (periodically increasing) modulus
νodd (n) = ⌊ n2 ⌋.
Overview. In Sec. 2 we define the notion of stream specification, and the syntactic
format of flat and pure specifications. In Sec. 3 we formalize the notion of d-o
rewriting. In Sec. 4 we introduce a production calculus as a means to compute
the production of the data-abstracted stream specifications, based on the set
of periodically increasing functions. A translation of stream specifications into
production terms is defined in Sec. 5. Our main results, mentioned above, are
collected in Sec. 6. We conclude and discuss some future research topics in Sec. 8.
2
Stream Specifications
We introduce the notion of stream specification. An example is given in Fig. 2,
a productive specification of Pascal’s triangle where the rows are separated by
P → 0 : s(0) : f(P)
f(s(x) : y : σ) → a(s(x), y) : f(y : σ)
stream layer
f(0 : σ) → 0 : s(0) : f(σ)
a(s(x), y) → s(a(x, y))
a(0, y) → y
data layer
Fig. 2: Example of a flat stream specification.
4
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
zeros. Indeed, evaluating this specification, we get:
P։
։ 0 :1 : 0 :1 : 1 :0 : 1: 2 :1 : 0 :1 : 3 :3 : 1 :... ,
where, for n ∈ N, n is the numeral for n, defined by n := sn (0).
Stream specifications consist of a stream layer (top) where stream constants
and functions are specified and a data layer (bottom).
The hierarchical setup of stream specifications is motivated as follows. In
order to abstract from the termination problem when investigating productivity,
the data layer is a term rewriting system on its own, and is required to be strongly
normalizing. Moreover, an isolated data layer prevents the use of stream symbols
by data rules. The stream layer may use symbols of the data layer, but not viceversa. Stream dependent data functions, like head(x : σ) → x, might cause the
output of undefined data terms. Consider the following examples from [8]:
S → 0 : S(6) : S
T → 0 : T(7) : T ,
where σ(n) := head(tailn (σ)); here S is well-defined, whereas T is not. Note that
this is not a severe restriction, since stream dependent data functions usually can
be replaced by pattern matching: f(σ, τ ) → (head(σ)+head(τ )):f(tail(σ), tail(τ )),
for example, can be replaced by the better readable f(x:σ, y :τ ) → (x+y):f(σ, τ ).
Stream specifications are formalized as many-sorted, orthogonal, constructor
term rewriting systems [10]. We distinguish between stream terms and data
terms. For the sake of simplicity we consider only one sort S for stream terms
and one sort D for data terms. Without any complication, our results extend to
stream specifications with multiple sorts for data terms and for stream terms.
Let U be a finite set of sorts.
A U-sorted set A is a family of sets {Au }u∈U ;
S
for V ⊆ U we define AV := v∈V Av . A U-sorted signature Σ is a U-sorted set of
function symbols f , each equipped with an arity ar (f ) = hu1 · · · un , ui ∈ U ∗ × U
where u is the sort of f ; we write u1 × . . . × un → u for hu1 · · · un , ui. Let X be
a U-sorted set of variables. The U-sorted set of terms Ter (Σ, X) is inductively
defined by: for all u ∈ U , Xu ⊆ Ter (Σ, X)u , and f (t1 , . . . , tn ) ∈ Ter (Σ, X)u
if f ∈ Σ, ar (f ) = u1 × . . . × un → u, and ti ∈ Ter (Σ, X)ui . Ter ∞ (Σ, X)
denotes the set of (possibly) infinite terms over Σ and X (see [10]). Usually we
keep the set of variables implicit and write Ter (Σ) and Ter ∞ (Σ). A U-sorted
term rewriting system (TRS) is a pair hΣ, Ri consisting of a U -sorted signature
Σ and a U-sorted set R of rules that satisfy well-sortedness, for all u ∈ U :
Ru ⊆ Ter (Σ, X)u × Ter (Σ, X)u , as well as the standard TRS requirements.
Let T = hΣ, Ri be a U-sorted TRS. For a term t ∈ Ter (Σ)u where u ∈ U we
denote the root symbol of t by r oot(t). We say that two occurrences of symbols in
a term are nested if the position [10, p.29] of one is a prefix of the position of the
other. We define D(Σ) := {r oot(l) | l → r ∈ R}, the set of defined symbols, and
C(Σ) := Σ \ D(Σ), the set of constructor symbols. Then T is called a constructor
TRS if for every rewrite rule ρ ∈ R, the left-hand side is of the form f (t1 , . . . , tn )
with ti ∈ Ter (C(Σ)); then ρ is called a defining rule for f . We call T exhaustive
for f ∈ Σ if every term f (t1 , . . . , tn ) with closed constructor terms ti is a redex.
Data-Oblivious Stream Productivity
5
A stream TRS is a finite {S , D }-sorted, orthogonal, constructor TRS hΣ, Ri
such that ‘:’ ∈ ΣS , the stream constructor symbol, with arity D × S → S is the
single constructor symbol in ΣS . Elements of ΣD and ΣS are called the data
symbols and the stream symbols, respectively. We let Σs := ΣS \ {‘:’}, and, for
all f ∈ Σs , we assume w.l.o.g. that the stream arguments are in front: ar (f) ∈
S ars (f) × D ard (f) → S , where ars (f) and ard (f) ∈ N are called the stream arity and
the data arity of f, respectively. By Σscon we denote the set of symbols in Σs with
stream arity 0, called the stream constant symbols, and Σsfun := Σs \ Σscon the
set of symbols in Σs with stream arity unequal to 0, called the stream function
symbols. By Rscon we mean the defining rules for the symbols in Σscon .
We repeat that the restriction to a single data sort D is solely for keeping
the presentation simple; all of the definitions and results in the sequel generalize
to multiple data and stream sorts. For stream TRSs we assume non-emptyness
of data sorts, that is, for every data sort there exists a finite, closed, contructor
term of this sort. In case there is only one data sort D , then the requirement
boils down to the existence of a nullary constructor symbol of this sort.
We come to the definition of stream specifications.
Definition 2.1. A stream specification T is a stream TRS T = hΣ, Ri such
that the following conditions hold:
(i) There is a designated symbol M0 ∈ Σscon with ard (M0 ) = 0, the root of T .
(ii) hΣD , RD i is a terminating, D -sorted TRS; RD is called the data-layer of T .
(iii) T is exhaustive (for all defined symbols in Σ = ΣS ⊎ ΣD ).
In the sequel we restrict to stream specifications in which all stream constants
in Σscon are reachable from the root: M ∈ Σscon is reachable if there is a term s
such that M0 →∗ s and M occurs in s. Note that reachability of stream constants
is decidable, and that unreachable symbols may be neglected for investigating
(non-)productivity.
Note that Def. 2.1 indeed imposes a hierarchical setup; in particular, stream
dependent data functions are excluded by item (ii). Exhaustivity for ΣD together
with strong normalization of RD guarantees that closed data terms rewrite to
constructor normal forms, a property known as sufficient completeness [4].
We are interested in productivity of recursive stream specifications that make
use of a library of ‘manageable’ stream functions. By this we mean a class of
stream functions defined by a syntactic format with the property that their d-o
lower bounds are computable and contained in a set of production moduli that
is effectively closed under composition, pointwise infimum and where least fixed
points can be computed. As such a format we define the class of flat stream
specifications (Def. 2.2) for which d-o lower bounds are precisely the set of ‘periodically increasing’ functions (see Sec. 4). Thus only the stream function rules
are subject to syntactic restrictions. No condition other than well-sortedness is
imposed on the defining rules of stream constant symbols.
In the sequel let T = hΣ, Ri be a stream specification. We define the relation
on rules in RS : for all ρ1 , ρ2 ∈ RS , ρ1
ρ2 (ρ1 depends on ρ2 ) holds if and
only if ρ2 is the defining rule of a stream function symbol on the right-hand
6
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
side of ρ1 . Furthermore, for a binary relation → ⊆ A × A on a set A we define
(a →) := {b ∈ A | a → b} for all a ∈ A, and we denote by →+ and →∗ the
transitive closure and the reflexive and transitive closure of →, respectively.
Definition 2.2. A rule ρ ∈ RS is called nesting if its right-hand side contains
nested occurrences of stream symbols from Σs . We use Rnest to denote the subset
of nesting rules of R and define R¬nest := RS \Rnest , the set of non-nesting rules.
A rule ρ ∈ RS is called flat if all rules in (ρ ∗ ) are non-nesting. A symbol
f ∈ Σs is called flat if all defining rules of f are flat; the set of flat symbols is
denoted Σflat . A stream specification T is called flat if Σs ⊆ Σflat ∪ Σscon , that
is, all symbols in Σs are either flat or stream constant symbols.
The specification given in Fig. 2 is an example of a flat specification. A second
example is given in Fig. 3. This is a productive specification that defines the
Q → a : Q′
Q′ → b : c : f(Q′ )
f(a : σ) → a : b : c : f(σ)
stream layer
f(b : σ) → a : c : f(σ)
f(c : σ) → b : f(σ)
data layer
Fig. 3: Example of a flat stream specification.
ternary Thue–Morse sequence, a square-free word over {a, b, c}; see, e.g., [7].
Indeed, evaluating this specification, we get: Q ։
։ a:b: c:a: c:b: a:b: c:b: a:c: . . ..
As the basis of d-o rewriting (see Def. 3.3) we define the data abstraction of
terms as the results of replacing all data-subterms by the symbol •.
Definition 2.3. Let LΣM := {•} ⊎ ΣS . For stream terms s ∈ Ter (Σ)S , the data
abstraction LsM ∈ Ter (LΣM)S is defined by:
LσM = σ ,
Lu : sM = • : LsM ,
Lf(s1 , . . . , sn , u1 , . . . , um )M = f(Ls1 M, . . . , Lsn M, •, . . . , •) .
Based on this definition of data abstracted terms, we define the class of pure
stream specifications, an extension of the equally named class in [2].
Definition 2.4. A stream specification T is called pure if it is flat and if for
every symbol f ∈ Σs the data abstractions LℓM → LrM of the defining rules ℓ → r
of f coincide (modulo renaming of variables).
Data-Oblivious Stream Productivity
7
Q → diff(M)
M → 0 : zip(inv(M), tail(M))
zip(x : σ, τ ) → x : zip(τ, σ)
inv(x : σ) → i(x) : inv(σ)
stream layer
tail(x : σ) → σ
diff(x : y : σ) → x(x, y) : diff(y : σ)
i(0) → 1
x(0, 0) → b
i(1) → 0
x(0, 1) → a
x(1, 0) → c
x(1, 1) → b
data layer
Fig. 4: Example of a pure stream specification.
Fig. 4 shows an alternative specification of the ternary Thue–Morse sequence,
this time constructed from the binary Thue–Morse sequence specified by M. This
specification belongs to the subclass of pure specifications, which is easily inferred
by the shape of the stream layer: for each symbol in Σsfun = {zip, inv, tail, diff}
there is precisely one defining rule.
Another example of a pure stream specification is given in Fig. 5. Both defin-
M → 0 : M′
M′ → 1 : h(M′ )
h(0 : σ) → 0 : 1 : h(σ)
stream layer
h(1 : σ) → 1 : 0 : h(σ)
data layer
Fig. 5: Example of a pure stream specification.
ing rules for h consume one, and produce two stream elements, that is, their data
abstractions coincide.
Def. 2.4 generalizes the specifications called ‘pure’ in [2] in four ways concerning the defining rules of stream functions: First, the requirement of rightlinearity of stream variables is dropped, allowing for rules like f(σ) → g(σ, σ).
Second, ‘additional supply’ to the stream arguments is allowed. For instance, in
the defining rule for the function diff in Fig. 4, the variable y is ‘supplied’ to
the recursive call of diff. Third, the use of non-productive stream functions is
allowed now, relaxing an earlier requirement on stream function symbols to be
‘weakly guarded’. Finally, defining rules for stream function symbols may use a
restricted form of pattern matching as long as, for every stream function f, the
8
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
d-o consumption/production behaviour (see Sec. 3) of all defining rules for f is
the same.
Next we define friendly nesting stream specifications, an extension of the
class of flat stream specifications.
Definition 2.5. A rule ρ ∈ RS is called friendly if for all rules γ ∈ (ρ ∗ )
we have: (1) γ consumes in each argument at most one stream element, and
(2) it produces at least one. The set of friendly nesting rules Rfnest is the largest
extension of the set of friendly rules by non-nesting rules from RS that is closed
under . A symbol f ∈ Σs is friendly nesting if all defining rules of f are friendly
nesting. A stream specification T is called friendly nesting if Σs ⊆ Σfnest ∪Σscon ,
that is, all symbols in Σs are either friendly nesting or stream constant symbols.
An example of a friendly nesting stream specification is given in Fig. 6. The root
nats → 0 : ×(ones, ones)
ones → s(0) : ones
×(x : σ, y : τ ) → m(x, y) : add(times(τ, x), ×(σ, y : τ ))
stream layer
times(x : σ, y) → m(x, y) : times(σ, y)
add(x : σ, y : τ ) → a(x, y) : add(σ, τ )
a(0, y) → y
a(s(x), y) → s(a(x, y))
m(0, y) → 0
data layer
m(s(x), y) → a(y, m(x, y))
Fig. 6: Example of a friendly nesting stream specification.
of this specification, the constant nats, evaluates to the stream 0 : 1 : 2 : . . .. The
stream layer specifies a parameterized stream function times, which multiplies
every element of a stream with the parameter, and the binary stream functions
add for componentwise addition, and ×, the convolution product of streams
(see [6]), mathematically defined as an operation hσ, τ i 7→ σ × τ :
(σ × τ )(i) =
i
X
σ(j) · τ (i − j)
(for all i ∈ N).
j=0
The rewrite rule for × is nesting, because its right-hand has nested stream function symbols, both times and × are nested within add. Because of the presence
of a nesting rule in the stream layer, which is not a defining rule of a stream constant symbol, this stream specification is not flat. However, the defining rules for
add, times, and × are friendly. For instance for the rule for × we check: in both
arguments it consumes (at most) one stream element (x and y), it produces
Data-Oblivious Stream Productivity
9
(at least) one stream element (m(x, y)), and the defining rules of the stream
function symbols in the right-hand side (add, times, and ×) are again friendly.
Thus the function stream layer consists of one friendly nesting rule, two flat
(and friendly) rules, and one defining rule for a stream constant. Therefore this
stream specification is friendly nesting.
Definition 2.6. Let A = hTer (Σ)S , →i be an abstract reduction system (ARS)
on the set of terms over a stream TRS signature Σ. The production function
ΠA : Ter (Σ)S → N of A is defined for all s ∈ Ter (Σ)S by:
ΠA (s) := sup { n ∈ N | s →∗A u1 : . . . : un : t } .
We call A productive for a stream term s if ΠA (s) = ∞. A stream specification
T is called productive if T is productive for its root M0 .
The following proposition justifies this definition of productivity by stating an
easy consequence: the root of a productive stream specification can be evaluated
continually in such a way that a uniquely determined stream in constructor
normal form is obtained as the limit. This follows easily from the fact that a
stream specification is an orthogonal TRS, and hence has a confluent rewrite
relation that enjoys the property UN∞ (uniqueness of infinite normal form) [5].
Proposition 2.7. A stream specification is productive if and only if its root has
an infinite constructor term of the form u1 : u2 : u3 : . . . as its unique infinite
normal form.
3
Data-Oblivious Analysis
We formalize the notion of d-o rewriting and introduce the concept of d-o productivity. The idea is a quantitative reasoning where all knowledge about the
concrete values of data elements during an evaluation sequence is ignored. For
example, consider the following stream specification:
M → f(0 : 1 : M)
(1) f(0 : x : σ) → 0 : 1 : f(σ)
(2) f(1 : x : σ) → x : f(σ)
The specification of M is productive: M →2 0:1:f(M) →3 0:1:0:1:f(f(M)) →∗ . . . .
During the rewrite sequence (2) is never applied. Disregarding the identity of
data, however, (2) becomes applicable and allows for the rewrite sequence:
M → f(• : • : M) →(2) • : f(M) →∗ • : f(• : f(• : f(. . .))) ,
producing only one element. Hence M is not d-o productive.
D-o term rewriting can be thought of as a two-player game between a rewrite
player R which performs the usual term rewriting, and an opponent G which
before every rewrite step is allowed to arbitrarily exchange data elements for
(sort-respecting) data terms in constructor normal form. The opponent can either handicap or support the rewrite player. Respectively, the d-o lower (upper)
10
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
bound on the production of a stream term s is the infimum (supremum) of the
production of s with respect to all possible strategies for the opponent G. Fig. 7
depicts d-o rewriting of the above stream specification M; by exchanging data
elements, the opponent G enforces the application of (2).
Remark 3.1. At first glance it might appear natural to model the opponent using
a function G : Ter (ΣD ) → Ter (ΣD ) from data terms to data terms. However,
such a per-element deterministic exchange strategy preserves equality of data
elements. Then the following specification of W :
W → g(Z, Z)
Z→0:Z
g(0 : σ, 0 : τ ) → 0 : g(σ, τ )
g(1 : σ, 1 : τ ) → 0 : g(σ, τ )
g(0 : σ, 1 : τ ) → g(σ, τ )
g(1 : σ, 0 : τ ) → g(σ, τ )
would be productive for every such G, which is clearly not what one would expect
of a d-o analysis.
The opponent can be modelled by an operation on stream terms: Ter (Σ)S →
Ter (Σ)S . However, it can be shown that it is sufficient to quantify over strategies
for G for which G(s) is invariant under exchange of data elements in s for all terms
s. Therefore we first abstract from the data elements in favour of symbols • and
then define the opponent G on the abstracted terms, G : Ter (LΣM)S → Ter (Σ)S .
Definition 3.2. Let T = hΣ, Ri be a stream specification. A data-exchange
function on T is a function G : Ter (LΣM)S → Ter (Σ)S such that ∀r ∈ Ter (LΣM)S
it holds: LG(r)M = r, and G(r) is in data-constructor normal form.
Definition 3.3. We define the ARS AT,G ⊆ Ter (LΣM)S × Ter (LΣM)S for every
data-exchange function G, as follows:
AT,G := {s → LtM | s ∈ Ter (LΣM), t ∈ Ter (Σ) with G(s) →T t} .
The d-o production range do T (s) of a data abstracted stream term s ∈ Ter (LΣM)S
is defined as follows:
do T (s) := {ΠAT,G (s) | G a data-exchange function on T } .
G
M
f(0 : 1 : M)
R
G
R
R
0 : f(f(0 : 1 : M))
R
0 : f(0 : f(M))
f(1 : 0 : M)
G
0 : f(M)
G
M
0 : f(M)
0 : f(f(1 : 0 : M))
G
...
Fig. 7: Data-oblivious rewriting
Data-Oblivious Stream Productivity
11
The d-o lower and upper bound on the production of s are defined by
do T (s) := inf(do T (s)) ,
and
do T (s) := sup(do T (s)) ,
respectively. These notions carry over to stream terms s ∈ Ter (Σ)S by taking
their data abstraction LsM.
A stream specification T is d-o productive (d-o non-productive) if do T (M0 ) =
∞ (if do T (M0 ) < ∞) holds.
Proposition 3.4. For T = hΣ, Ri a stream specification and s ∈ Ter (Σ)S :
do T (s) ≤ ΠT (s) ≤ do T (s) .
Hence d-o productivity implies productivity and d-o non-productivity implies nonproductivity.
We define lower and upper bounds on the d-o consumption/production behaviour of stream functions. These bounds are used to reason about d-o (non-)
productivity of stream constants, see Sec. 6.
Definition 3.5. Let T = hΣ, Ri be a stream specification, g ∈ Σs , k = ars (g),
and ℓ = ard (g). The d-o production range do T (g) : Nk → P(N) of g is:
do T (g)(n1 , . . . , nk ) := do T ( g((•n1 : σ), . . . , (•nk : σ), •, . . . , •) ) ,
| {z }
ℓ-times
m times
z }| {
where •m :σ := • : . . . : • : σ. The d-o lower and upper bounds on the production of
g are defined by do T (g) := inf(do T (g)) and do T (g) := sup(do T (g)), respectively.
Even simple stream function specifications can exhibit a complex d-o behaviour. For example, consider the flat function specification:
f(σ) → g(σ, σ)
g(0 : y : σ, x : τ ) → 0 : 0 : g(σ, τ )
g(1 : σ, x1 : x2 : x3 : x4 : τ ) → 0 : 0 : 0 : 0 : 0 : g(σ, τ )
Fig. 8 (left) shows a (small) selection of the possible function-call traces for f. In
particular, it depicts the traces that contribute to the d-o lower bound do T (f).
The lower bound do T (f), shown on the right, is a superposition of multiple traces
of f. In general do T (f) can even be a superposition of infinitely many traces.
First Observations on Data-Oblivious Rewriting
For a stream function symbol g, we define its optimal production modulus ν g ,
the data-aware, quantitative lower bound on the production of g, and compare
it to its d-o lower bound do T (g).
12
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
output
output
15
15
10
10
5
5
5
10
15
input
5
10
15
input
Fig. 8: Traces
Definition 3.6. Let T = hΣ, Ri be a stream specification. We define the set
Sn of n-defined stream terms, i.e. finite stream terms with a constructor-prefix
of length n ∈ N:
Sn := {u1 : . . . : un : σ | u1 , . . . , un ∈ Ter (C(ΣD ), ∅)} ,
Moreover, let g ∈ Σsfun with k = ars (g) and ℓ = ard (g), and define, for all
n = n1 , . . . , nk , the set Gn of applications of g to ni -defined arguments:
Gn := { g(s, v) | ∀i. si ∈ Sni , ∀j. vj ∈ Ter (C(ΣD ), ∅) } ,
where s = s1 , . . . , sk and v = v1 , . . . , vℓ . Then, the optimal production modulus
k
ν g : N → N of g is defined by:
ν g (n) := inf {ΠT (t) | t ∈ Gn } .
To illustrate the difference between the optimal production modulus ν h and
the d-o lower bound do T (h), consider the following stream function specification:
h(0 : x : σ) → x : x : h(0 : σ)
(ρh0 )
h(1 : x : σ) → x : h(0 : σ)
(ρh1 )
. 3 is the optimal production modulus of
with ΣD = {0, 1}. Then ν h (n) = 2n −
the stream function h. To obtain this bound one has to take into account that
the data element 0 is supplied to the recursive call and conclude that (ρh1 ) is
only applicable in the first step of a rewrite sequence h(u1 : . . . : un : σ) → . . ..
. 1, derived from rule (ρ ).
However, the d-o lower bound is do T (h)(n) = n −
h1
The following lemmas state an observation about the role of the opponent
and the rewrite player. Basically, the opponent G can select the rule which is
applicable for each position in the term; the rewrite player R can choose which
position to rewrite. We use subscripts for pebbles •, for example •w , to introduce
‘names’ for referring to these pebbles.
Data-Oblivious Stream Productivity
13
Definition 3.7 (Instantiation with respect to a rule ρ). For s ∈ Ter (LΣM)S
with s ≡ f(s1 , . . . , sars (f) , •1 , . . . , •ard (f) ) and ρ ∈ R a defining rule of f, we define
a data-exchange function Gs,ρ as follows. Note that ρ is of the form:
ρ : f(u1 : σ1 , . . . , un : σn , v1 , . . . , vard (f) ) → r ∈ R .
For i = 1, . . . , ars (f) let ni ∈ N be maximal such that si ≡ •i,1 : . . . •i,ni : s′i .
Let Gs,ρ for all 1 ≤ i ≤ ard (f) instantiate the pebbles •i with closed instances
of the data patters vi , and for all 1 ≤ i ≤ ars (f ), 1 ≤ j ≤ min(ni , |ui |) instantiate
the pebble •i,j with closed instances of the data pattern uij , respectively.
Lemma 3.8. Let s ∈ Ter (LΣM)S , s ≡ f(s1 , . . . , sars (f) , •1 , . . . , •ard (f) ) and ρ ∈ R
a defining rule of f; we use the notation from Def. 3.7. Then Gs,ρ (s) is a redex
if and only if for all 1 ≤ i ≤ ars (f ) we have: ui ≤ ni , that is, there is ‘enough
supply for the application of ρ’. Furthermore if Gs,ρ (s) is a redex, then it is a
redex with respect to ρ (and no other rule from R by orthogonality).
Proof. A consequence of the fact that T is an orthogonal constructor TRS.
⊓
⊔
Lemma 3.9. Let G be a data-exchange function, t ∈ Ter (LΣM) a term and define
P = {p | p ∈ Pos(t), r oot(t|p ) ∈ Σ}. For every ς : P → R such that ς(p) is a
defining rule for r oot(t|p ) for all p ∈ P, there is a data-exchange function G t,ς
such that for all p ∈ P if the term G t,ς (t)|p is a redex, then it is an ς(p)-redex,
and G t,ς (s) ≡ G(s) for all s 6≡ t.
Proof. For all p ∈ P and ς(p) ≡ ℓ → r we alter G to obtain a data-guess function
G t,ς as follows. If q ∈ Pos(ℓ) such that ℓ|q is a data term and none of the t(q · p′ )
with p′ a non empty prefix of p is a symbol from Σsfun , then instantiate the data
term at position q · p in t with an instance of the data pattern ℓ|q . Then if G t,ς (t)
is a redex, then by orthogonality of T it can only be a ℓ → r-redex.
⊓
⊔
History Aware Data-Exchange
Above we have modelled the opponent using history free data-exchange strategies and the rewrite player was omniscient, that is, she always chooses the best
possible rewrite sequence, which produces the maximum possible number of elements.
Now we investigate the robustness of our definition. We strengthen the opponent, allowing for history aware data-exchange strategies, and we weaken the
rewrite player, dropping omniscience, assuming only an outermost-fair rewrite
strategy. However it turns out that these changes do not affect the d-o production range do T (s), in this way providing evidence for the robustness of our first
definition.
Definition 3.10. Let T = hΣ, Ri be a stream specification. A history in T is
a finite list in (Ter (Σ)S × R × N∗ × Ter (Σ)S )∗ of the form:
hs0 , ℓ0 → r0 , p0 , t0 ihs1 , ℓ1 → r1 , p1 , t1 i . . . hsn , ℓn → rn , pn , tn i
14
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
such that ∀0 ≤ i ≤ n: si → ti by an application of rule ℓi → ri at position pi .
We use HT to denote the set of all histories of T .
A function G : HT → (Ter (LΣM)S → Ter (Σ)S ) from histories in T to dataexchange functions on T is called a history-aware data-exchange function on T .
For such a function we write Gh as shorthand for G(h) for all h ∈ HT .
Definition 3.11. Let T = hΣ, Ri be a stream specification. Let G be a historyaware data-exchange function. We define the ARS HT,G ⊆ (Ter (LΣM)S × HT ) ×
(Ter (LΣM)S × HT ) as follows:
HT,G := {hLsM, hi → hLtM, hhs, ℓ → r, p, tii |
s, t ∈ Ter (Σ), h ∈ HT , with Gh (LsM) = s and s →T t} .
For B ⊆ HT,G we define the production function ΠB : Ter (Σ)S → N by:
ΠB (s) := sup { n ∈ N | hs, ǫi →∗ B hu1 : . . . : un : t, hi } ,
that is, the production of s starting with empty history ǫ.
In an ARS HT,G we allow to write s0 →HT,G s1 →HT,G s2 →HT,G . . . when
we will actually mean a rewrite sequence in HT,G of the form hs0 , h0 i →HT,G
hs1 , h1 i →HT,G hs2 , h2 i →HT,G . . . with some histories h0 , h1 , h2 , . . . ∈ HT .
The notion of positions of rewrite steps and redexes extends to the ARS HT,G
in the obvious way, since the steps in HT,G are projections of steps in T . Note that
the opponent can change the data elements in HT,G at any time, therefore the
rules might change with respect to which a certain position is a redex. For this
reason we adapt the notion of outermost-fairness for our purposes. Intiutively,
a rewrite sequence is outermost-fair if every position which is always eventually
an outermost redex position, is rewritten infinitely often. Formally, we define a
rewrite sequence hs1 , h1 i → hs2 , h2 i → . . . in HT,G to be outermost-fair if it is
finite, or it is infinite and whenever p is an outermost redex position for infinitely
many sn , n ∈ N, then there exist infinitely many rewrite steps sm → sm+1 ,
m ∈ N, at position p. Moreover, a strategy B ⊆ HT,G is outermost-fair if every
rewrite sequence with respect to this strategy is outermost-fair.
Proposition 3.12. For every stream specification T = hΣ, Ri there exists a
computable, deterministic om-fair strategy on HT,G .
Definition 3.13. The history aware d-o production range do H
T (s) of s ∈ Ter (LΣM)S
is defined as follows:
do H
T (s) := ΠB (s) | G a history-aware data-exchange function on T ,
B ⊆ HT,G an outermost-fair strategy .
H
For s ∈ Ter (Σ)S define do H
T (s) := do T (LsM).
Note that also the strategy of the rewrite player B ⊆ HT,G is history aware
since the elements of the ARS HT,G have history annotations.
Data-Oblivious Stream Productivity
15
Lemma 3.14. Let T = hΣ, Ri be a stream specification. Let f ∈ Σsfun with
ars (f) = n and ar (f) = n′ , and let s1 , . . . , sn ∈ Ter (Σ)S , u1 , . . . , un′ ∈ Ter (Σ)D .
Then it holds:
m1
H
: σ, . . . , •mn : σ, u1 , . . . , un′ )) , (1)
do H
T (f(s1 , . . . , sn , u1 , . . . , un′ )) = do T (f(•
where mi := do H
T (si ) for all i ∈ {1, . . . , n}.
Proof. Let T = hΣ, Ri be a stream specification. Let f ∈ Σsfun with ars (f) = n,
and s1 , . . . , sn ∈ Ter (Σ)S . We assume that the data arity of f is zero; in the
presence of data arguments the proof proceeds analogously. Furthermore we let
mi := do H
T (si ) for i ∈ {1, . . . , n}. We show (1) by demonstrating the inequalities
“≤” and “≥” in this equation.
≥: Suppose that do H
T (f(s1 , . . . , sn )) = ΠB (Lf(s1 , . . . , sn )M) = N ∈ N for some
om-fair strategy B ⊆ HT,G and some history-aware data-exchange function
G on T . (In case that do H
T (f(s1 , . . . , sn )) = ∞, nothing remains to be shown.)
Then there exists an om-fair rewrite sequence
ξ : Lf(s1 , . . . , sn )M →HT,G u1 →HT,G u2 →HT,G . . .
. . . →HT,G ul →HT,G ul+1 →HT,G . . .
in HT,G with liml→∞ #: (ul ) = N . Due to the definition of the numbers
mi , ah history-aware data-exchange function G ′ from HT to data-exchange
functions can be defined that enables a rewrite sequence
′
=
′
=
ξ ′ : f(•m1 : σ, . . . , •mn : σ) →=
HT,G′ u1 →HT,G′ u2 →HT,G′ . . .
′
=
=
′
. . . →=
HT,G′ ul →HT,G′ ul+1 →HT,G′ . . .
with the properties that #: (ul ) = #: (u′l ) holds for all l ∈ N, and that ξ ′ is
the projection of ξ to a rewrite sequence with source f(•m1 : σ, . . . , •mn : σ).
More precisely, ξ ′ arises as follows: Steps ul →HT,G ul+1 in ξ that contract
redexes in descendants of the outermost symbol f in the source of ξ give
rise to steps u′l →HT,G′ u′l+1 that contract redexes at the same position and
apply the same rule. This is possible because, due to the definition of the
numbers mi , always enough pebble-supply is guaranteed to carry out steps
in ξ ′ corresponding to such steps in ξ. Steps ul →HT,G ul+1 that contract
redexes in descendants of one of the subterms s1 , . . . , sn in the source of
ξ project to empty steps u′l = u′l+1 in ξ ′ . (On histories h ∈ HT and terms
r ∈ Ter (LΣM)S occurring in ξ, G ′ is defined in such a way as to make this
projection possible. On histories h and terms r that do not occur in ξ ′ , G ′ is
defined arbitrarily with the only restriction LGh (r)M ensuring that G ′ behaves
as a history-aware data-exchange function on these terms.)
By its construction, ξ ′ is again om-fair. Now let B ′ be the extension of the
sub-ARS of HT,G induced by ξ ′ to a deterministic om-fair strategy for HT,G ′ .
(Choose an arbitrary deterministic om-fair strategy B̃ on HT,G , which is
16
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
possible by Proposition 3.12. On term-history pairs that do not occur in
ξ ′ , define B ′ according to B̃.) Then ξ ′ is also a rewrite sequence in B ′ that
witnesses ΠB (f(•m1 : σ, . . . , •mn : σ)) = liml→∞ #: (u′l ) = liml→∞ #: (ul ) =
N = ΠB (Lf(s1 , . . . , sn )M). Now it follows:
m1
: σ, . . . , •mn : σ)) ≤ ΠB′ (f(•m1 : σ, . . . , •mn : σ))
do H
T (f(•
= ΠB (Lf(s1 , . . . , sn )M) = do H
T (f(s1 , . . . , sn )) ,
establishing “≥” in (1).
m1
≤: Suppose that do H
:σ, . . . , •mn :σ)) = ΠB (f(•m1 :σ, . . . , •mn :σ)) = N for
T (f(•
some N ∈ N, and an om-fair strategy B ⊆ HT,G and a history-aware datam1
: σ, . . . , •mn : σ)) = ∞,
exchange function G on T . (In case that do H
T (f(•
nothing remains to be shown.) Then there exists an om-fair rewrite sequence
ξ : f(•m1 : σ, . . . , •mn : σ)
→HT,G u1 →HT,G u2 →HT,G . . . →HT,G ul →HT,G ul+1 →HT,G . . .
in HT,G with liml→∞ #: (ul ) = N . Let l0 ∈ N be minimal such that #: (ul0 ) =
N . Furthermore, let, for all i ∈ {1, . . . , n}, m′i ≤ mi be minimal such that
each of the topmost m′i symbols “•” of the subterm •mi : σ of the source
f(•m1 : σ, . . . , •mn : σ) of ξ is part of a redex pattern at some step among the
first l0 steps of ξ.
Since by assumption, we have mi := do H
T (si ) for i ∈ {1, . . . , n}, there exist,
for each i ∈ {1, . . . , n}, history-aware data-guess functions Gi , and om-fair
strategies Bi ⊆ HT,Gi such that ΠBi (Lsi M) = mi ; let all Bi and Gi be chosen
as such. Then there exist, for each i, om-fair rewrite sequences of the form:
′
ξi : Lsi M →HT,Gi ui,1 →HT,Gi . . . →HT,Gi ui,li = •mi : ũi,li →HT,Gi . . .
in HT,G with liml→∞ #: (ui,l ) = mi .
Now it is possible to combine the history-aware data-exchange functions G,
G1 , . . . , Gn into a function G ′ that makes it possible to combine the rewrite
sequences ξ and ξ1 , . . . , ξn into a rewrite sequence of the form:
ξ ′ : Lf(s1 , . . . , sn )M
′
′
→∗ HT,G′ f(•m1 : u1,l1 , . . . , •mn : un,ln )
(carry out the first li steps in ξi in the context f(21 , . . . , 2n ))
→HT,G′ u′1 →HT,G′ u′2 →HT,G′ . . . →HT,G′ u′l0 = •N : ũ′l0
(parallel to the first l0 steps of ξ)
→HT,G′ w1 →HT,G′ w2 →HT,G′ . . .
(fair interleaving of the rests of the ξi put in context
and of steps parallel to the rest of ξ ′ )
By its construction, ξ ′ is again om-fair and it holds: liml→∞ wl = N . Now let
B ′ be the extension of the sub-ARS of HT,G induced by ξ ′ to a deterministic
Data-Oblivious Stream Productivity
17
om-fair strategy for HT,G ′ . (Again choose an arbitrary deterministic om-fair
strategy B̃ on HT,G , which is possible by Proposition 3.12. On term-history
pairs that do not occur in ξ ′ , define B ′ according to B̃.) Then ξ ′ is also a
rewrite sequence in B ′ that witnesses ΠB (Lf(s1 , . . . , sn )M) = liml→∞ #: (wl ) =
N = ΠB (f(•m1 : σ, . . . , •mn : σ)). Now it follows:
do H
T (f(s1 , . . . , sn )) ≤ ΠB (Lf(s1 , . . . , sn )M)
m1
: σ, . . . , •mn : σ)) ,
= ΠB′ (f(•m1 : σ, . . . , •mn : σ)) = do H
T (f(•
⊓
⊔
establishing “≤” in (1).
Lemma 3.15. Let s ∈ Ter (LΣM)S be a stream term and φ : Var S → Ter (LΣM)S
a substitution of stream variables. Then
φ
ϕ
H
do H
T (s ) = do T (s )
where ϕ : Var S → Ter (LΣM)S is defined by ϕ(τ ) = •mτ :σ, and mτ := do H
T (φ(τ )).
Proof. Induction using Lem. 3.14.
⊓
⊔
Corollary 3.16. Let s ∈ Ter (LΣM)S and φ, ϕ : Var S → Ter (LΣM)S substitutions
H
of stream variables such that do H
T (φ(τ )) = do T (ϕ(τ )) for all τ ∈ Var S . Then
φ
ϕ
H
do H
T (s ) = do T (s ) .
Proof. Two applications of Lem. 3.15.
⊓
⊔
We define a history free data-exchange function G T such that no reduction
with respect to G T produces more than do H
T (s) elements.
Definition 3.17. Let a stream specification T = hΣ, Ri be given, and let > be
a well-founded order on R. We define the lower bound data-exchange function
G T : Ter (LΣM)S → Ter (Σ)S as follows.
Let s ∈ Ter (LΣM)S and f(s1 , . . . , sars (f) , •1 , . . . , •ard (f) ) a subterm occurrence
of s. For i = 1, . . . , ars (f) let ni ∈ N be maximal such that si ≡ •i,1 : . . . •i,ni : s′i .
We define mi := do H
T (si ) for 1 ≤ i ≤ n, then we have by Lem. 3.14:
do H
T (t := f(s1 , . . . , sars (f) , •1 , . . . , •ard (f) )) =
′
m1
: σ, . . . , •mars (f) : σ, •1 , . . . , •ard (f) )) .
do H
T (t := f(•
We choose a defining rule ρ of f as follows. In case there exists a rule
γ : f(u1 : σ1 , . . . , un : σn , v1 , . . . , vard (f) ) → t ∈ R
such that mi < |ui | for some i ∈ {1, . . . , n}, then let ρ := γ. In case there are
multiple possible choices for γ, we pick the minimal γ with respect to >.
Otherwise in exists a history-aware data-exchange function G, and an outermost′
fair rewrite sequence σ : t′ →HT,G . . . in HT,G producing only do H
T (t ) elements.
′
From exhaustivity of T we get G(t ) is not a normal form, since all rules have
18
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
enough supply. Moreover, by orthogonality exactly one defining rule γ of f is applicable, let ρ := γ. Again, in case there are multiple possible choices for γ, due
to freedom in the choice of the data-exchange function G, we take the minimal
γ with respect to >.
We define G T to instantiate:
– for 1 ≤ i ≤ ard (f) the occurrences of pebbles •i in s, and
– for 1 ≤ i ≤ ars (f), 1 ≤ j ≤ ni the occurrences of pebbles •i,j in s
with respect to the rule ρ (Def. 3.7).
Rewrite steps with respect to the lower bound data-exchange function G T do
not change the d-o lower bound of the production of a term.
H
Lemma 3.18. For all s, s ∈ Ter (LΣM)S : s →AT,G T t implies do H
T (s) = do S (t).
Proof. We use the notions introduced in Def. 3.17. Note that the lower bound
data-exchange function G T is defined in a way that the pebbles are instantiated
independent of the context above the stream function symbol. Therefore it is
sufficient to consider rewrite steps s →AT,G T t at the root, closure under contexts
follows from Cor. 3.16.
⊓
⊔
Corollary 3.19. For all s ∈ Ter (LΣM)S we have: ΠAT,G T (s) = do H
T (s).
Proof. Direct consequence of Lem. 3.18.
⊓
⊔
Hence, as a consequence of Cor. 3.19, the lower bound of the history-free d-o
production conincides with the lower bound of the history-aware d-o production.
Lemma 3.20. For all s ∈ Ter (Σ)S we have: do T (s) = do H
T (s).
H
Proof. The direction do T (s) ≥ do H
T (s) is trivial, and do T (s) ≤ do T (s) follows
from Cor. 3.19, that is, using the history-free data-exchange function G T no
reduction produces more than do H
⊓
⊔
T (s) elements.
Lemma 3.21. For all s ∈ Ter (Σ)S we have: do T (s) = do H
T (s).
Proof. ⊆ Let n ∈ do T (s). Then there exists a data-exchange function G such
that ΠAT,G (s) = n, and a rewrite sequence σ : s0 →AT,G . . . →AT,G sm with
sm ≡ u1 : . . . : un : t. We define for all h ∈ HT a data-exchange function Gh by
setting Gh (s) := G(s) for all s ∈ Ter (LΣM)S . Moreover, we define the strategy
B ⊆ HT,G to execute σ and continue outermost-fair afterwards. Since sm
does not produce more than n elements, every maximal rewrite sequence
with respect to B will produce exactly n elements. Hence n ∈ do H
T (s).
⊇ Let n ∈ do H
T (s). Then there exist data-exchange functions Gh for every h ∈
HT , and an outermost-fair strategy B ⊆ HT,G such that ΠB (s) = n. There
exists a (possibly infinite) maximal, outermost-fair rewrite sequence
σ : hs, ǫi ≡ hs0 , h0 i →B . . . →B hsm , hm i →B . . .
Data-Oblivious Stream Productivity
19
such that sm ≡ u1 : . . . : un : t. Since the elements of HT,G are annotated with
their history, we do not encounter repetitions during the rewrite sequence.
Hence, w.l.o.g. we can assume that B is a deterministic strategy, admitting
only the maximal rewrite sequence σ. Disregarding the history annotations
there might exist repetitions, that is, i < j ∈ N with si ≡ sj . Nevertheless,
from σ we can construct a history free data-exchange function G as follows.
For all s ∈ Ter (LΣM)S let S := {i ∈ N | si ≡ s} and define:
if j := sup S < ∞
Ghj (sj )
G ′ (s) := Ghk (sk )
if sup S = ∞, k := inf S
arbitrary if ¬∃i ∈ N. si ≡ s
⊓
⊔
4
The Production Calculus
As a means to compute the d-o production behaviour of stream specifications,
we introduce a ‘production calculus’ with periodically increasing functions as
its central ingredient. This calculus is the term equivalent of the calculus of
pebbleflow nets that was introduced in [2], see Sec. 4.3.
4.1
Periodically Increasing Functions
We use N := N ⊎ {∞}, the extended natural numbers, with the usual ≤, +, and
we define ∞ − ∞ := 0. An infinite sequence σ ∈ X ω is eventually periodic if
σ = αβββ . . . for some α ∈ X ∗ and β ∈ X + . A function f : N → N is eventually
periodic if the sequence hf (0), f (1), f (2), . . .i is eventually periodic.
Definition 4.1. A function g : N → N is called periodically increasing if it
is increasing, i.e. ∀n. g(n) ≤ g(n + 1), and the derivative of g, n 7→ g(n +
1) − g(n), is eventually periodic. A function h : N → N is called periodically increasing if its restriction to N is periodically increasing and if h(∞) =
limn→∞ h(n). Finally, a k-ary function i : (N)k → N is called periodically increasing if i(n1 , ..., nk ) = min(i1 (n1 ), . . . , ik (nk )) for some unary periodically
increasing functions i1 , . . . , ik .
Periodically increasing (p-i) functions can be denoted by their value at 0
followed by a representation of their derivative. For example, 0312 denotes the
p-i function f : N → N with values 0, 3, 4, 6, 7, 9, . . .. However, we use a finer and
more flexible notation over the alphabet {+, −} that will be useful for computing
several operations on p-i functions. For instance, we represent f as above by the
‘io-sequence’
−+++−+−++ = −+++ −+−++ −+−++ −+−++ . . . ,
in turn abbreviated by the ‘io-term’ h−+++, −+−++i.
20
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Definition 4.2. Let ± := {+, −}. An io-sequence is a finite (∈ ±∗ ) or infinite
sequence (∈ ±ω ) over ±, the set of io-sequences is denoted by ±∞ := ±∗ ∪ ±ω .
We let ⊥ denote the empty io-sequence, and use it as an explicit end marker.
An io-term is a pair hα, βi of finite io-sequences α, β ∈ ±∗ with β 6= ⊥. The
set of io-terms is denoted by I, and we use ι, κ to range over io-terms. For an
io-sequence σ ∈ ±ω and an io-term ι = hα, βi ∈ I we say that ι denotes σ if
σ = αβ where β stands for the infinite sequence βββ . . .. A sequence σ ∈ ±∞
is rational if it is finite or if it is denoted by an io-term. The set of rational
ω
io-sequences is denoted by ±∞
R . An infinite sequence σ ∈ ± is productive if it
contains infinitely many +’s:
σ ∈ ±ω is productive ⇐⇒ ∀n. ∃m ≥ n. σ(m) = + .
∗
ω
∞
We let ±ω
P denote the set of productive sequences and define ±P := ± ∪ ±P .
The reason for having both io-terms and io-sequences is that, depending on the
purpose at hand, one or the other is more convenient to work with. Operations
are often easier to understand on io-sequences than on io-terms, whereas we need
finite representations to be able to compute these operations.
Definition 4.3. We define [[σ]] : N → N, the interpretation of σ ∈ ±∞ , by:
[[⊥]](n) = 0
[[+σ]](n) = 1 + [[σ]](n)
[[−σ]](0) = 0
[[−σ]](n + 1) = [[σ]](n)
(2)
(3)
(4)
(5)
for all n ∈ N, and extend it to N → N by defining [[σ]](∞) = limn→∞ [[σ]](n). We
say that [[σ]] interprets σ, and, conversely, that σ represents [[σ]]. We overload
notation and define [[ ]] : I → (N → N) by [[hα, βi]] := [[αβ]].
It is easy to verify that, for every σ ∈ ±∞ , the function [[σ]] is increasing, and
that, for every ι ∈ I, the function [[ι]] is periodically increasing. Furthermore,
every increasing function is represented by an io-sequence, and every p-i function
is denoted by an io-term.
Subsequently, for an eventually constant function f : N → N, we write f
for the shortest finite io-sequence representing f (trailing −’s can be removed).
For an always eventually strictly increasing function f : N → N, i.e. ∀n. ∃m >
n. f (m) > f (n), we write f for the unique io-sequence representing f ; note that
then f ∈ ±ω
P follows. For a periodically increasing function f : N → N, we write
f for the io-term ι = hα, βi such that [[ι]] = f and |α|+ |β| is minimal; uniqueness
follows from the following lemma:
Lemma 4.4. For all p-i functions f : N → N, the term f ∈ I is unique.
Data-Oblivious Stream Productivity
21
Proof. Consider the ‘compression’ TRS consisting of the rules:
hαx, βxi → hα, xβi
p
hα, β i → hα, βi
(x ∈ ±)
(6)
(n > 1)
(7)
Clearly, ι → κ implies |ι| > |κ| where, for ι = hα, βi, |ι| := |α| + |β|. Hence,
compression is terminating.
There are two critical pairs due to an overlap of rule (7) with itself, and in
rules (6) and (7): hhα, β n i, hα, β m ii originating from a term of the form hα, β n·m i,
and hhα, (xβ)n i, hαx, βxii originating from a term of the form hαx, (βx)n i, respectively. Both pairs are easily joinable in one step: the first to hα, βi, and the
second to hα, xβi. Hence, the system is locally confluent. Therefore, by Newman’s
Lemma, it is also confluent, and normal forms are unique.
Finally, assume that two io-terms hα1 , β1 i, hα2 , β2 i ∈ I each of minimal
length represent the same p-i function f . It is easy to see that then α1 β1 = α2 β2 .
It follows that there exist hα, βi such that hα, βi →∗ hα1 , β1 i and hα, βi →∗
hα2 , β2 i. By confluence and termination of compression it follows that they have
the same normal form hα0 , β0 i. Since compression reduces the length of io-terms
it must follow that hα1 , β1 i = hα0 , β0 i = hα2 , β2 i.
⊓
⊔
Note that we have that [[f ]] = f for all increasing functions f , and [[f ]] = f
for all p-i functions f . As an example one can check that the interpretation
of the aforementioned io-term f = h−+++, −+−++i indeed has the sequence
0, 3, 4, 6, 7, 9, . . . as its graph.
Remark 4.5. Note that Def. 4.3 is well-defined due to the productivity requirement on infinite io-sequences. This can be seen as follows: starting on [[σ]](n),
after finitely many, say m ∈ N, −’s, we either arrive at rule (3) (if m ≤ n) or at
rule (4) (if m > n). Had we not required infinite sequences σ ∈ ±ω to contain
infinitely many +’s, then, e.g., the computation of [[−−− . . .]](∞) would consist
of applications of rule (5) only, and hence would not lead to an infinite normal
form.
Proposition 4.6. Unary periodically increasing functions are closed under composition and pointwise infimum.
We want to constructively define the operations of composition, pointwise infimum, and least fixed point calculation of increasing functions (p-i functions)
on their io-sequence (io-term) representations. We first define these operations
on io-sequences by means of coinductive clauses. The operations on io-terms
are more involved, because they require ‘loop checking’ on top. We will proceed
by showing that the operations of composition and pointwise infimum on iosequences preserve rationality. This will then give rise to the needed algorithms
for computing the operations on io-terms.
22
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
∞
∞
Definition 4.7. The operation composition ◦ : ±∞
P × ±P → ±P , hσ, τ i 7→ σ ◦ τ
of (finite or productive) io-sequences is defined coinductively as follows:
⊥◦τ =⊥
+σ ◦ τ = +(σ ◦ τ )
(8)
(9)
−σ ◦ ⊥ = ⊥
−σ ◦ +τ = σ ◦ τ
(10)
(11)
−σ ◦ −τ = −(−σ ◦ τ )
(12)
An argument similar to Rem. 4.5 concerning well-definedness applies for Def. 4.7.
Remark 4.8. Note that the defining rules (8)–(12) are orthogonal and exhaust all
cases. As it stands, it is not immediately clear that this definition is well-defined
(productive!), the problematic case being (11). Rule (11) is to be thought of as
an internal communication between components σ and τ , as a silent step in the
black box σ ◦ τ . How to guarantee that always, after finitely many internal steps,
either the process ends or there will be an external step, in the form of output
or a requirement for input? The recursive call in (11) is not guarded. However,
well-definedness of composition can be justified as follows:
∞
Consider arbitrary sequences σ, τ ∈ ±∞
P . By definition of ±P , there exists m ∈ N
m
′
′
∞
m
such that σ = − +σ for some σ ∈ ±P or σ = − ⊥; likewise there exists n ∈ N
n
such that τ = −n +τ ′ for some τ ′ ∈ ±∞
P , or τ = − ⊥. Rules (11) and (12) are
decreasing with respect to the lexicographic order on (m, n). After finitely many
applications of (11) and (12), one of the rules (8)–(10) must be applied. Hence
composition is well-defined and the sequence produced is an element of ±∞
P , i.e.
either it is a finite sequence or it contains an infinite number of +’s.
Lemma 4.9. Composition of io-sequences is associative.
Proof. Let R = {hσ ◦ (τ ◦ υ), (σ ◦ τ ) ◦ υi | σ, τ, υ ∈ ±∞
P }. To show that R is
a bisimulation, we prove that, for all σ, τ, υ, φ1 , φ2 ∈ ±∞
P , if φ1 = σ ◦ (τ ◦ υ)
and φ2 = (σ ◦ τ ) ◦ υ, then either φ1 = ⊥ = φ2 or head(φ1 ) = head(φ2 ) and
htail(φ1 ), tail(φ2 )i ∈ R, by induction on the number n ∈ N of leading −’s of σ 4 ,
and a sub-induction on the number m ∈ N of leading −’s of τ .
If n = 0 and σ = ⊥, then φ1 = ⊥ = φ2 . If σ = +σ ′ , we have φ1 =
+(σ ′ ◦ (τ ◦ υ)), and φ2 = +((σ ′ ◦ τ ) ◦ υ), and so htail(φ1 ), tail(φ2 )i ∈ R.
If n > 0, then σ = −σ ′ , and we proceed by sub-induction on m. If m = 0
and τ = ⊥, then φ1 = ⊥ = φ2 . If τ = +τ ′ , we compute φ1 = σ ′ ◦ (τ ′ ◦ υ), and
φ2 = (σ ′ ◦ τ ′ ) ◦ υ, and conclude by IH.
If m > 0, then τ = −τ ′ , and we proceed by case distinction on υ. If υ = ⊥,
then φ1 = ⊥ = φ2 . If υ = +υ ′ , we compute φ1 = σ◦(τ ′ ◦υ ′ ), and φ2 = (σ◦τ ′ )◦υ ′ ,
and conclude by sub-IH. Finally, if υ = −υ ′ , then φ1 = −(σ ◦ (τ ◦ υ ′ )), and
φ2 = −((σ ◦ τ ) ◦ υ ′ ). Clearly htail(φ1 ), tail(φ2 )i ∈ R.
⊓
⊔
4
∞
By definition of ±∞
P the number of leading −’s of any sequence in ±P is finite: for
n
n
′
all σ ∈ ±∞
there
exists
n
∈
N
such
that
either
σ
=
−
⊥
or
σ
=
−
+σ
.
P
Data-Oblivious Stream Productivity
23
Remark 4.10. If we allow to use extended natural numbers at places where an
io-sequence is expected, using a coercion n 7→ +n , with +0 = ⊥, one observes
that the interpretation of an io-sequence (Def. 4.3) is just a special case of
composition:
[[σ]](n) = σ ◦ n .
Proposition 4.11. For all increasing functions f, g : N → N: [[f ◦g]] = [[f ]]◦[[g]].
Proof. Immediate from Rem. 4.10, and Lem. 4.9.
⊓
⊔
Next, we show that composition of io-sequences preserves rationality.
∞
Lemma 4.12. If σ, τ ∈ ±∞
R , then σ ◦ τ ∈ ±R .
Proof. Let σ, τ ∈ ±∞
R and set σ0 := σ and τ0 := τ . In the rewrite sequence
starting with σ0 ◦ τ0 each of the steps is either of the form:
σn ◦ τn → Cn [σn+1 ◦ τn+1 ] ,
(13)
where σn+1 ∈ {σn , tail(σn )}, τn+1 ∈ {τn , tail(τn )} and Cn ∈ {, + : , − : }, or
the rewrite sequence ends with a step of the form:
σn ◦ τn → ⊥ .
In the latter case, σ ◦ τ results in a finite and hence rational io-sequence.
Otherwise the rewrite sequence is infinite, consisting of steps of form (13)
only. For θ ∈ ±∗ we define the context Cθ inductively: Cθ ≡ , for θ ≡ ⊥;
and C+θ ≡ +Cθ as well as C−θ ≡ −Cθ , for all θ ∈ ±∗ . Because the sequences
σ, τ are rational, the sets {σn | n ∈ N} and {τn | n ∈ N} are finite. Then, the
pigeonhole principle implies the existence of i, j ∈ N such that i < j, σi = σj
and τi = τj . Now let α, γ ∈ ±∗ with γ 6= ⊥ such that Cα ≡ C0 [. . . Ci−1 [Ci ] . . .],
Cγ ≡ Ci+1 [. . . Cj−1 [Cj ] . . .], and:
σ0 ◦ τ0 →∗ Cα [σi ◦ τi ] , and
σi ◦ τi →∗ Cγ [σj ◦ τj ] = Cγ [σi ◦ τi ] .
Then we find an eventually ‘spiralling’ rewrite sequence:
σ0 ◦ τ0 →∗ Cα [σi ◦ τi ] = α(σi ◦ τi )
→∗ Cα [Cγ [σi ◦ τi ]] = αγ(σi ◦ τi )
→∗ Cα [Cγ [Cγ [σi ◦ τi ]]] = αγγ(σi ◦ τi ) ,
and therefore σ0 ◦ τ0 ։
։ αγ. This shows that σ ◦ τ is rational.
⊓
⊔
Next, we define the operation of pointwise infimum of increasing functions
on io-sequences.
24
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Definition 4.13. The operation pointwise infimum ∧ : ±∞ × ±∞ → ±∞ ,
hσ, τ i 7→ σ∧τ of io-sequences is defined coinductively by the following equations:
⊥∧τ =⊥
del(⊥) = ⊥
σ∧⊥=⊥
+σ ∧ +τ = +(σ ∧ τ )
−σ ∧ τ = −(σ ∧ del(τ ))
σ ∧ −τ = −(del(σ) ∧ τ )
del(+σ) = +del(σ)
del(−σ) = σ
(τ 6= ⊥)
(σ 6= ⊥)
where del(σ) removes the first requirement of σ ∈ ±∞ (if any).
We let ◦ bind stronger than ∧.
Requirement removal distributes over pointwise infimum:
Lemma 4.14. For all σ, τ ∈ ±∞
P , it holds that: del(σ ∧ τ ) = del(σ) ∧ del(τ ).
Proof. Check that {hdel(σ ∧ τ ), del(σ) ∧ del(τ )i | σ, τ ∈ ±∞ } ∪ {hσ, σi | σ ∈ ±∞ }
is a bisimulation.
⊓
⊔
Lemma 4.15. Infimum is idempotent, commutative, and associative.
Proof. By coinduction, with the use of Lem. 4.14 in case of associativity.
⊓
⊔
In a composition requirements come from the second component:
Lemma 4.16. For all σ, τ ∈ ±∞
P , it holds that: del(σ ◦ τ ) = σ ◦ del(τ ).
Composition distributes both left and right over pointwise infimum:
Lemma 4.17. For all σ, τ, υ ∈ ±∞
P , it holds that: σ ◦ (τ ∧ υ) = σ ◦ τ ∧ σ ◦ υ.
Proof. By coinduction; let L = {hσ ◦ (τ ∧ υ), σ ◦ τ ∧ σ ◦ υi | σ, τ, υ ∈ ±∞
P }.
To show that L is a bisimulation, we prove that, for all σ, τ, υ, φ1 , φ2 ∈ ±∞
P ,
if φ1 = σ ◦ (τ ∧ υ) and φ2 = σ ◦ τ ∧ σ ◦ υ, then either φ1 = ⊥ = φ2 , or
head(φ1 ) = head(φ2 ) and htail(φ1 ), tail(φ2 )i ∈ L, by induction on the number
n ∈ N of leading −’s of σ.
If n = 0 and σ = ⊥, then φ1 = ⊥ = φ2 .
If n = 0 and σ = +σ ′ , then: φ1 = +(σ ′ ◦ (τ ∧ υ)), and φ2 = +(σ ′ ◦ τ ∧ σ ′ ◦ υ),
and so htail(φ1 ), tail(φ2 )i ∈ L.
If n > 0 and σ = −σ ′ , we proceed by case analysis of τ and υ.
If one of τ , υ is empty, then φ1 = ⊥ = φ2 .
If τ = +τ ′ and υ = +υ ′ , then φ1 = σ ′ ◦ (τ ′ ∧ υ ′ ) and φ2 = σ ◦ τ ∧ σ ◦ υ, and
we conclude by the induction hypothesis.
If τ = +τ ′ and υ = −υ ′ , then φ1 = −(σ ◦ (del(τ ) ∧ υ ′ )) and φ2 = −(del(σ ◦
τ ) ∧ σ ◦ υ ′ ) = −(σ ◦ del(τ ) ∧ σ ◦ υ ′ ) by Lem. 4.16. Thus htail(φ1 ), tail(φ2 )i ∈ L.
The case τ = −τ ′ , υ = +υ ′ is proved similarly.
Finally, if τ = −τ ′ and υ = −υ ′ , we compute φ1 = −(σ ◦ (τ ′ ∧ υ ′ )) and
φ2 = −(σ ◦ τ ′ ∧ σ ◦ υ ′ ), and conclude htail(φ1 ), tail(φ2 )i ∈ L.
⊓
⊔
Data-Oblivious Stream Productivity
25
Lemma 4.18. For all σ, τ, υ ∈ ±∞
P , it holds that: (τ ∧ υ) ◦ σ = τ ◦ σ ∧ υ ◦ σ.
Proof. Analogous to the proof of Lem. 4.17.
⊓
⊔
The operation of pointwise infimum of (periodically) increasing functions is
defined on their io-sequence (io-term) representations.
Proposition 4.19. For all increasing functions f, g : N → N: [[f ∧g]] = [[f ]]∧[[g]].
Proof. Immediate from Rem. 4.10 and Lem. 4.18.
⊓
⊔
We give a coinductive definition of the calculation of the least fixed point of
an io-sequence. The operation del was defined in Def. 4.13.
Definition 4.20. The operation fix : ±∞ → N computing the least fixed point
of a sequence σ ∈ ±∞ , is defined, for all σ ∈ ±∞ , by:
fix(⊥) = 0
fix(+σ) = 1 + fix(del(σ))
fix(−σ) = 0
Removal of a requirement and feeding an input have equal effect:
Lemma 4.21. For all σ, τ ∈ ±∞
P , it holds that: del(σ) ◦ τ = σ ◦ (+τ ).
Proof. We show that R = {hdel(σ) ◦ τ , σ ◦ (+τ )i} ∪ {hσ, σi} is a bisimulation, by
case analysis of σ. For σ = ⊥, and σ = −σ ′ , hdel(σ) ◦ τ , σ ◦ (+τ )i ∈ R follows by
reflexivity. If σ = +σ ′ , then del(σ)◦τ = +(del(σ ′ )◦τ ), and σ◦(+τ ) = +(σ ′ ◦(+τ )).
Hence, htail(del(σ) ◦ τ ), tail(σ ◦ (+τ ))i ∈ R.
⊓
⊔
The following proposition states that fix(σ) is a fixed point of [[σ]]:
Proposition 4.22. For all σ ∈ ±∞
P , it holds that: [[σ]](fix(σ)) = fix(σ).
Proof. We prove fix(σ) = σ ◦ fix(σ) by case analysis and coinduction.
If σ = ⊥, then fix(σ) = ⊥ = σ ◦ fix(σ).
If σ = −σ ′ , then fix(σ) = ⊥, and σ ◦ fix(σ) = (−σ ′ ) ◦ ⊥ = ⊥.
If σ = +σ ′ , then fix(σ) = +fix(del(σ ′ )) and σ ◦ fix(σ) = +(σ ′ ◦(+fix(del(σ ′ )))),
and we have to prove that fix(del(σ ′ )) = σ ′ ◦ (+fix(del(σ ′ ))) which follows by an
instance of Lem. 4.21: σ ′ ◦ (+fix(del(σ ′ ))) = del(σ ′ ) ◦ fix(del(σ ′ )).
⊓
⊔
Lemma 4.23. For all σ ∈ ±∞ , it holds that: lfp([[σ]]) = fix(σ).
Proof.
26
4.2
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Computing Production
We introduce a term syntax for the production calculus and rewrite rules for
evaluating closed terms.
Definition 4.24. Let X be a set of recursion variables. The set of production
terms P is generated by:
p ::= k | x | •(p) | f (p) | µx.p | min(p, p)
where x ∈ X , for k ∈ N, the symbol k is a numeral (a term representation) for
k, and, for a unary p-i function f : N → N, f ∈ I, the io-term representing f .
For every finite set P = {p1 , . . . , pn } ⊆ P, we use min(p1 , . . . , pn ) and min P as
shorthands for the production term min(p1 , min(p2 , . . . , min(pn−1 , pn ))).
The ‘production’ [[p]] ∈ N of a closed production term p ∈ P is defined by
induction on the term structure, interpreting µ as the least fixed point operator,
f as f , k as k, and min as min, as follows.
Definition 4.25. The production [[p]]α ∈ N of a term p ∈ P with respect to an
assignment α : X → N is defined inductively by:
[[k]]α = k
[[x]]α = α(x)
[[•(p)]]α = 1 + [[p]]α
[[ι(p)]]α = [[ι]]([[p]]α )
[[µx.p]]α = lfp(λn.[[p]]α[x7→n] )
[[min(p1 , p2 )]]α = min([[p1 ]]α , [[p2 ]]α )
where α[x 7→ n] denotes an update of α, defined by α[x 7→ n](y) = n if y = x,
and α[x 7→ n](y) = α(y) otherwise.
Finally, we let [[p]] := [[p]]α0 with α0 defined by α0 (x) = 0 for all x ∈ X .
As becomes clear from Def. 4.25, we could have done without the clause •(p) in
the BNF grammar for production terms in Def. 4.24, as • can be abbreviated to
+−+, an io-term that denotes the successor function. However, we take it up as
a primitive constructor here in order to match with pebbleflow nets, see Sec. 4.3.
For faithfully modelling the d-o lower bounds of stream functions with stream
arity r, we employ r-ary p-i functions, which we represent by ‘r-ary gates’.
Definition 4.26. An r-ary gate gatek (σ1 , . . . , σr ) is defined as a production
term context of the form:
gatek (σ1 , . . . , σr ) := min(k, σ1 (21 ), . . . , σr (2r )) ,
where k ∈ N and σ1 , . . . , σr ∈ ±∞ . We use γ as a syntactic variable for gates.
The interpretation of a gate γ = gatek (σ1 , . . . , σr ) is defined by:
[[γ]](n1 , . . . , nr ) := min(k, [[σ1 ]](n1 ), . . . , [[σr ]](nr )) .
In case k = ∞, we simplify gatek (σ1 , . . . , σr ) to
gate(σ1 , . . . , σr ) := min(σ1 (21 ), . . . , σr (2r )) .
Data-Oblivious Stream Productivity
27
It is possible to choose unique gate representations f of p-i functions f that are
efficiently computable from other gate representations, see Section 4.3.
Owing to the restriction to (term representations of) periodically increasing
functions in Def. 4.25 it is possible to calculate the production [[p]] of terms p ∈ P.
For that purpose, we define a rewrite system which reduces any closed term to
a numeral k. This system makes use of the computable operations ◦ and fix on
io-terms mentioned above.
Definition 4.27. The rewrite relation →R on production terms is defined as
the compatible closure of the following rules:
•(p) → +−+(p)
ι1 (ι2 (p)) → ι1 ◦ ι2 (p)
(R1)
(R2)
ι(min(p1 , p2 )) → min(ι(p1 ), ι(p2 ))
µx.min(p1 , p2 ) → min(µx.p1 , µx.p2 )
(R3)
(R4)
µx.p → p
if x 6∈ FV(p)
µx.ι(x) → fix(ι)
(R5)
(R6)
min(k1 , k2 ) → min(k1 , k2 )
ι(k) → [[ι]](k)
µx.x → 0
(R7)
(R8)
(R9)
The following two lemmas establish the usefulness of the rewrite relation →R . In
order to compute the production [[p]] of a production term p it suffices to obtain
a →R -normal form of p.
Lemma 4.28. The rewrite relation →R is production preserving:
p →R p′ =⇒ [[p]] = [[p′ ]] .
Proof. It suffices to prove: C[ℓσ ] →R C[rσ ] =⇒ ∀α. [[C[ℓσ ]]]α = [[C[rσ ]]]α , where
ℓ → r is a rule of the TRS given in Def. 4.27, and C a unary context over P.
We proceed by induction on C. For the base case, C = [], we give the essential
proof steps only: For rule (R1), observe that [[−+]] is the identity function on
N. For rule (R2), we apply Prop. 4.11. For rule (R3) the desired equality follows
from C1 on page 31. For rule (R4) we conclude by C2 ibid. For rule (R6) we
use Lem. 4.23. For the remaining rules the statement trivially holds. For the
induction step, the statement easily follows from the induction hypotheses. ⊓
⊔
Lemma 4.29. The rewrite relation →R is terminating and confluent, and every
closed p ∈ P has a numeral k as its unique →R -normal form.
Proof. To see that →R is terminating, let w : P → N be defined by:
w(x) = 1
w(k) = 1
w(•(p)) = 2 · w(p) + 1
w(σ(p)) = 2 · w(p)
w(µx.p) = 2 · w(p)
w(min(p1 , p2 )) = w(p1 ) + w(p2 ) + 1 ,
28
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
and observe that p →R q implies w(p) > w(q).
Some of the rules of →R overlap; e.g. rule (R2) with itself. For each of the five
critical pairs we can find a common reduct (the critical pair hσ ◦(τ ◦υ), (σ ◦τ )◦ υi
due to an (R2)/(R2)-overlap can be joined by Lem. 4.9), and hence →R is locally
confluent, by the Critical Pairs Lemma (cf. [10]). By Newman’s Lemma, we
obtain that →R is confluent. Thus normal forms are unique.
To show that every closed net normalises to a source, let p be an arbitrary
normal form. Note that the set of free variables of a net is closed under →R , and
hence p is a closed net. Clearly, p does not contain pebbles, otherwise (R1) would
be applicable. To see that p contains no subterms of the form µx.q, suppose it
does and consider the innermost such subterm, viz. q contains no µ. If q ≡ k or
q ≡ x, then (R5), resp. (R9) is applicable. If q ≡ σ(q ′ ), we further distinguish four
cases: if q ′ ≡ k or q ′ ≡ x, then (R8) resp. (R6) is applicable; if the root symbol
of q ′ is one of box, min, then q constitutes a redex w.r.t. (R2), (R3), respectively.
If q ≡ min(q1 , q2 ), we have a redex w.r.t. (R4). Thus, there are no subterms µx.q
in p, and therefore, because p is closed, also no variables x. To see that p has
no subterms of the form σ(q), suppose it does and consider the innermost such
subterm. Then, if q ≡ k or q ≡ min(q1 , q2 ) then (R8) resp. (R3) is applicable;
other cases have been excluded above. Finally, p does not contain subterms of
the form min(p1 , p2 ). For if it does, consider the innermost occurrence and note
that, since the other cases have been excluded already, p1 and p2 have to be
sources, and so we have a redex w.r.t. (R7). We conclude that p ≡ k for some
k ∈ N.
⊓
⊔
4.3
Pebbleflow Nets
Production terms can be visualized by ‘pebbleflow nets’ introduced in [2], and
serve as a means to model the ‘data-oblivious’ consumption/production behaviour of stream specifications. The idea is to abstract from the actual stream
elements (data) in a stream term in favour of occurrences of the symbol •, which
we call ‘pebble’. Thus, a stream term u : s is translated to [u : s] = •([s]), see
Section 5.
We give an operational description of pebbleflow nets, and define the production of a net as the number of pebbles a net is able to produce at its output
port. Then we prove that this definition of production coincides with Def. 4.25.
Pebbleflow nets are networks built of pebble processing units (fans, boxes,
meets, sources) connected by wires. We use the term syntax given in Def. 4.24
for nets and the rules governing the flow of pebbles through a net, and then give
an operational meaning of the units a net is built of.
Data-Oblivious Stream Productivity
29
Definition 4.30. The pebbleflow rewrite relation →P is defined as the compatible closure of the union of the following rules:
min(•(p1 ), •(p2 )) → •(min(p1 , p2 ))
µx.•(p(x)) → •(µx.p(•(x)))
(P1)
(P2)
+σ(p) → •(σ(p))
−σ(•(p)) → σ(p)
(P3)
(P4)
1 + k → •(k)
(P5)
Wires are unidirectional FIFO communication channels. They are idealised
in the sense that there is no upper bound on the number of pebbles they can
store; arbitrarily long queues are allowed. Wires have no counterpart on the term
level; in this sense they are akin to the edges of a term tree. Wires connect boxes,
meets, fans, and sources, that we describe next.
A meet is waiting for a pebble at each of its input ports and only then
produces one pebble at its output port, see Fig. 9. Put differently, the number
of pebbles a meet produces equals the minimum of the numbers of pebbles
available at each of its input ports. Meets enable explicit branching; they are
used to model stream functions of arity > 1, as will be explained below. A meet
with an arbitrary number n ≥ 1 of input ports is implemented by using a single
wire in case n = 1, and if n = k + 1 with k ≥ 1, by connecting the output port
of a ‘k-ary meet’ to one of the input ports of a (binary) meet.
min
min
p
p1
p2
p1
Fig. 9: Rule (P1).
p
p2
Fig. 10: Rule (P2).
The behaviour of a fan is dual to that of a meet: a pebble at its input port
is duplicated along its output ports. A fan can be seen as an explicit sharing
device, and thus enables the construction of cyclic nets. More specifically, we use
fans only to implement feedback when drawing nets; there is no explicit term
representation for the fan in our term calculus. In Fig. 10 a pebble is sent over
the output wire of the net and at the same time is fed back to the ‘recursion
wire(s)’. Turning a cyclic net into a term (tree) means to introduce a notion of
binding; certain nodes need to be labelled by a name (µx) so that a wire pointing
30
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
to that node is replaced by a name (x) referring to the labelled node. In rule (P2)
feedback is accomplished by substituting •(x) for all free occurrences x of p.
A source has an output port only, contains a number k ∈ N of pebbles, and
can fire if k > 0, see Fig. 13. The rewrite relation →R given in Def. 4.27 collapses
any closed net (without input ports that is) into a source.
A box consumes pebbles at its input port and produces pebbles at its output
port, controlled by an io-sequence σ ∈ ±∞ associated with the box. For example,
consider the unary stream function dup, defined as follows, and its corresponding
io-sequence:
dup(x : σ) = x : x : dup(σ)
−++
which is to be thought of as: for dup to produce two outputs, it first has to
consume one input, and this process repeats indefinitely. Intuitively, the symbol −
represents a requirement for an input pebble, and + represents a ready state for
an output pebble. Pebbleflow through boxes is visualised in Figs. 11 and 12.
+σ
σ
−σ
σ
p
p
p
p
Fig. 11: Rule (P3).
1+k
Fig. 12: Rule (P4).
k
Fig. 13: Rule (P5).
The data-oblivious production behaviour of
stream functions f with a stream arity ars (f) = r are
modelled by r-ary gates (defined in Def. 4.26) that
express the contribution of each individual stream argument to the total production of f, see Fig. 14. The
precise translation of stream functions into gates is
given in Sec. 5, in particular in Def. 5.9.
min
σ1
σr
Fig. 14: gate(σ1 , . . . , σr )
Lemma 4.31. The pebbleflow rewrite relation →P is confluent.
Proof. The rules of →P can be viewed as a higher-order rewriting system (HRS)
that is orthogonal. Applying Thm. 11.6.9 in [10] then establishes the lemma. ⊓
⊔
Definition 4.32. The production function Π : P → N of nets is defined by:
Π(p) := sup {n ∈ N | p ։P •n (p′ )} ,
for all p ∈ P. Π(p) is called the production of p. Moreover, for a net p and an
assignment α ∈ X → N, let Π(p, α) := Π(pα ) where pα denotes the net obtained
by replacing each free variable x of p with •α(x) (x).
Data-Oblivious Stream Productivity
31
Note that for closed nets p (where FV(p) = ∅), pα = p and therefore [[p]]α = [[p]],
for all assignments α.
An important property used in the following lemma is that functions of the
form λn ∈ N. Π(p, α[x 7→ n]) are monotonic functions over N. Every monotonic
function f : N → N in the complete chain N has a least fixed point lfp(f ) which
can be computed by lfp(f ) = limn→∞ f n (0). In what follows we employ, for
monotonic f, g : N → N, two basic properties:
∀n, m. f (min(n, m)) = min(f (n), f (m))
(C1 )
lfp(λn. min(f (n), g(n))) = min(lfp(f ), lfp(g))
(C2 )
Lemma 4.33. For all nets q ∈ P and all assignments α, we have that Π(µx.q, α)
is the least fixed point of λn ∈ N. Π(q, α[x 7→ n]).
Proof. Let α : X → N be an arbitrary assignment and q0 := q α[x7→0] . Observe
that Π(µx.q, α) = Π(µx.q0 ) and consider a rewrite sequence of the form
µx.q0 →∗ . . . →∗ •ni (µx.qi ) →∗ •ni (µx.•ℓi (qi′ )) →∗ •ni +ℓi (µx.qi+1 ) →∗ . . .
where ℓi = Π(qi ), n0 = 0, ni+1 = ni + ℓi , and qi+1 := qi′ (•ℓi (x)). Note that
limm→∞ nm = Π(µx.q0 ); ‘≤’ follows from ∀m. µx.q0 →∗ •nm (µx.qm ), and ‘≥’
since if limm→∞ nm < ∞ then ∃m ∈ N such that ℓm := Π(qm ) = 0 and therefore
Π(µx.q0 ) = Π(•nm (µx.qm )) = nm by confluence.
Let fi = λn.[[qi (•n (x))]], and fi′ = λn.[[qi′ (•n (x))]]. We prove
∀k ∈ N. f0 (nm + k) = nm + fm (k)
(∗)
by induction over m. The base case m = 0 is trivial, we consider the induction
′
) and by substituting •k (x) for x we get
step. We have qm →∗ •ℓm (qm
′
∀k ∈ N. fm (k) = ℓm + fm
(k)
(∗∗)
′
′
Moreover, since fm+1 (k) = fm
(ℓm +k), we get nm+1 +fm+1 (k) = nm+1 +fm
(ℓm +
(∗∗)
(∗)
′
k) = nm +ℓm +fm
(ℓm +k) = nm +fm (ℓm +k) = f0 (nm +ℓm +k) = f0 (nm+1 +k).
Let f := f0 . We proceed with showing ∀m. f m (0) = nm by induction over
m ∈ N. For the base case m = 0 we have f 0 (0) = 0 and n0 = 0, and for the
IH
(∗)
induction step we get f m+1 (0) = f (f m (0)) = f (nm ) = nm +fm (0) = nm +ℓm =
nm+1 .
Hence lfp(f ) = limm→∞ f m (0) = limm→∞ nm = Π(µx.q0 ) = Π(µx.q, α).
⊓
⊔
Lemma 4.34. For p ∈ P, σ ∈ ±∞
P , α : X → N: Π(σ(p), α) = [[σ]](Π(p, α)).
Proof. We show that the relation R ⊆ N × N defined as follows is a bisimulation:
R := hΠ(σ(p), α), [[σ]](Π(p, α))i | σ ∈ ±∞
P , p ∈ P, α : X → N ,
32
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
that is, we prove that, for all k1 , k2 ∈ N, σ ∈ ±∞
P , p ∈ P, and α : X → N, if
k1 = Π(σ(p), α) and k2 = [[σ]](Π(p, α)), then either k1 = k2 = 0 or k1 = 1 + k1′ ,
k2 = 1 + k2′ and hk1′ , k2′ i ∈ R.
Let k1 , k2 ∈ N, σ ∈ ±∞
P , p ∈ P, and α : X → N, be such that k1 = Π(σ(p), α)
and k2 = [[σ]](Π(p, α)). By definition of ±∞ , we have that σ ≡ −n +τ for some
n ∈ N and τ ∈ ±∞ . We proceed by induction on n. If n = 0, then k1 = 1 + k1′
with k1′ = Π(τ (p), α) and k2 = 1 + k2′ with k2′ = [[τ ]](Π(p, α)), and hk1′ , k2′ i ∈ R.
If n = n′ + 1, we distinguish cases: If Π(p, α) = 0, then k1 = k2 = 0. If
Π(p, α) = 1 + m, then p ։P •(q) for some q ∈ P with Π(q, α) = m. Thus we get
′
′
k1 = Π(−n +τ (q), α) and k2 = [[−n +τ ]](Π(q, α)), and hk1 , k2 i ∈ R by induction
hypothesis.
⊓
⊔
Next we show that production of a term p ∈ P (Def. 4.25) coincides with the
maximal number of pebbles produced by the net p via the rewrite relation →P
(Def. 4.32).
Lemma 4.35. For all nets p and assignments α, it holds Π(p, α) = [[p]]α .
Proof. The statement of the lemma can be proved by a straightforward induction
on the number of µ-bindings of a net p, with a subinduction on the size of p.
In the cases p ≡ σ(p′ ) and p ≡ µx.p′ Lem. 4.34 and Lem. 4.33 are applied,
respectively.
⊓
⊔
Subsequently, we will use the interpretation of terms in P and the production
of pebbleflow nets, [[ ]] and Π, interchangeably.
5
Translation into Production Terms
In this section we define a translation from stream constants in flat or friendly
nesting specifications to production terms. In particular, the root M0 of a specification T is mapped by the translation to a production term [M0 ] with the
property that if T is flat (friendly nesting), then the d-o lower bound on the
production of M0 in T equals (is bounded from below by) the production of [M0 ].
5.1
Translation of Flat and Friendly Nesting Symbols
As a first step of the translation, we describe how for a flat (or friendly nesting)
stream function symbol f in a stream specification T a periodically increasing
function hfi can be calculated that is (that bounds from below) the d-o lower
bound on the production of f in T .
Let us again consider the rules (i) f(s(x) : y : σ) → a(s(x), y) : f(y : σ), and
(ii) f(0 : σ) → 0 : s(0) : f(σ) from Fig. 2. We model the d-o lower bound on the
production of f by a function from N to N defined as the unique solution for
Xf of the following system of equations. We disregard what the concrete stream
elements are, and therefore we take the infimum over all possible traces:
Xf (n) = inf Xf,(i) (n), Xf,(ii) (n)
Data-Oblivious Stream Productivity
33
where the solutions for Xf,(i) and Xf,(ii) are the d-o lower bounds of f assuming
that the first rule applied in the rewrite sequence is (i) or (ii), respectively. The
rule (i) consumes two elements, produces one element and feeds one element
back to the recursive call. For rule (ii) these numbers are 1, 2, 0 respectively.
Therefore we get:
Xf,(i) (n) = let n′ := n − 2, if n′ < 0 then 0 else 1 + Xf (1 + n′ ) ,
Xf,(ii) (n) = let n′ := n − 1, if n′ < 0 then 0 else 2 + Xf (n′ ) .
. 1, represented by the io-term −−+.
The unique solution for Xf is n 7→ n −
In general, functions may have multiple arguments, which during rewriting
may get permuted, deleted or duplicated. The idea is to trace single arguments,
and to take the infimum over traces in case an argument is duplicated.
In the definition of the translation of stream functions, we need to distinguish
the cases according to whether a symbol is weakly guarded or not: On Σs we
define ; := {hr oot(ℓ), r oot(r)i | ℓ → r ∈ RS , ‘:’ 6= r oot(r) ∈ Σs } . the dependency relation between symbols in Σs . We say that a symbol f ∈ Σs is weakly
guarded if f is strongly normalising with respect to ; and unguarded, otherwise.
Note that since Σs is finite it is (easily) decidable whether a symbol f ∈ Σs is
weakly guarded or unguarded.
The translation of a stream function symbol is defined as the unique solution
of a (usually infinite) system of defining equations where the unknowns are
functions. More precisely, for each symbol f ∈ Σfnest ⊇ Σsfun of a flat or friendly
nesting stream specification, this system has a p-i function hfi as the solution for
Xf , which is unique among the continuous functions. Later we will see (Prop. 5.14
and Lem. 5.15) that the translation hfi of a flat (friendly nesting) stream function
symbol f coincides with (is a lower bound for) the d-o lower bound do T (f) of f.
Definition 5.1. Let T = hΣ, Ri be a stream specification. For each flat or
friendly nesting symbol f ∈ Σfnest ⊇ Σflat with arities k = ars (f) and ℓ = ard (f)
k
we define hfi : N → N, called the (p-i function) translation of f in T , as
the unique solution sXf for Xf of the following system of defining equations,
where the solution of an equation of the form X(n1 , . . . , nk ) = . . . is a function
k
sX : N → N (it is to be understood that sX ∈ N if k = 0): For all n1 , . . . , nk ∈ N,
i ∈ {⋆, 1, . . . , k}, and n ∈ N:
Xf (n1 , . . . , nk ) = min(Xf,⋆ , Xf,1 (n1 ), . . . , Xf,k (nk )) ,
(
inf Xf,⋆,ρ | ρ a defining rule of f
if f is weakly guarded,
Xf,⋆ =
0
if f is unguarded,
(
inf Xf,i,ρ (n) | ρ a defining rule of f
if f is weakly guarded,
Xf,i (n) =
0
if f is unguarded.
We write ui : σi for ui,1 : . . . : ui,p : σi , and |ui | for p. For specifying Xf,i,ρ we
distinguish the possible forms the rule ρ can have. If ρ is nesting, then Xf,⋆,ρ = ∞,
34
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
and Xf,i,ρ (n) = n for all n ∈ N. Otherwise, ρ is non-nesting and of the form:
f((u1 : σ1 ), . . . , (uk : σk ), v1 , . . . , vℓ ) → w1 : . . . : wm : s ,
where either (a) s ≡ σj , or (b) s ≡ g((u′1 : σφ(1) ), . . . , (u′k′ : σφ(k′ ) ), v1′ , . . . , vℓ′ ′ )
with k ′ = ars (g), ℓ′ = ard (g), and φ : {1, . . . , k ′ } → {1, . . . , k}. Let:
(
∞
case (a)
Xf,⋆,ρ =
m + Xg,⋆ case (b)
Xf,i,ρ (n) = let n′ := n − |ui |, if n′ < 0 then 0 else
′
n
m+ ∞
inf Xg,j (n′ + |u′j |) | j ∈ φ−1 (i)
case (a), i = j
case (a), i 6= j
case (b)
where we agree inf ∅ = ∞.
We mention a useful intuition for understanding the fabric of the formal
definition above of the translation hfi of a stream function f, the solution sXf for
Xf of the system of equations in Def. 5.1. For each i ∈ {1, . . . , k} where k = ars (f)
the solution sXf,i for Xf,i (a unary p-i function) in this system describes to what
extent consumption from the i-th component of f ‘delays’ the overal production.
Since in a d-o rewrite sequence from f(•n1 : σ1 , . . . , •nk : σk ) it may happen that
all stream variables σ1 , . . . , σk are erased at some point, and that the sequence
subsequently continues rewriting stream constants, monitoring the delays caused
by individual arguments is not sufficient alone to define the production function
of f. This is the reason for the use, in the definition of hfi, of the solution sXf,⋆
for the variable Xf,⋆ , which defines a ‘glass ceiling’ for the not adequate ‘overall
delay’ function min(sXf,1 (n1 ) , . . . , sXf,k (nk ) ), taking account of situations in which
in d-o rewrite sequences from terms f(•n1 : σ1 , . . . , •nk : σk ) all input components
have been erased.
Concerning non-nesting rules on which defining rules for friendly nesting
symbols depend via , this translation uses the fact that their production is
bounded below by ‘min’. These bounds are not necessarily optimal, but can be
used to show productivity of examples like X → 0 :f(X) with f(x : σ) → x : f(f(σ)).
Def. 5.1 can be used to define, for every flat or friendly nesting symbol f ∈
Σsfun in a stream specification T , a ‘gate translation’ of f: by defining this
translation by choosing a gate that represents the p-i function hfi.
However, it is desirable to define this translation also in an equivalent way
that lends itself better for computation, and where the result directly is a production term (gate) representation of a p-i function: by specifying the io-terms in
a gate translation as denotations of rational io-sequences that are the solutions
of ‘io-sequence specifications’. In particular, the gate translation of a symbol
f ∈ Σsfun will be defined in terms of the solutions, for each argument place of a
stream function symbol f, of a finite io-sequence specification that is extracted
from an infinite one which precisely determines the d-o lower bound of f in that
argument place.
Data-Oblivious Stream Productivity
35
Definition 5.2. Let X be a set of variables. The set of io-sequence specification
expressions over X is defined by the following grammar:
E ::= ⊥ | X | −E | +E | E ∧ E
where X ∈ X . Guardedness is defined by induction: An io-sequence specification
expression is called guarded if it is of one of the forms ⊥, −E0 , or +E0 , or if it
is of the form E1 ∧ E2 for guarded E1 and E2 .
Suppose that X = {Xα | α ∈ A} for some (countable) set A. Then by
an io-sequence specification over (the set of recursion variables) X we mean a
family {Xα = Eα }α∈A of recursion equations, where, for all α ∈ A, Eα is an iosequence specification expression over X . Let E be an io-sequence specification. If
E consists of finitely many recursion equations, then it is called finite. E is called
guarded (weakly guarded ) if the right-hand side of every recursion equation in
E is guarded (or respectively, can be rewritten, using equational logic and the
equations of E, to a guarded io-sequence specification expression).
Let E = {Xα = Eα }α∈A an io-sequence specification. Furthermore let α0 ∈ A
and τ ∈ ±ω . We say that τ is a solution of E for Xα0 if there exist io-sequences
{σα }α∈A such that σα0 = τ , and all of the recursion equations in E are true
statements (under the interpretation of ∧ as defined in Def. 4.13), when, for all
α ∈ A, σα is substituted for Xα , respectively.
It turns out that weakly guarded io-sequence specifications have unique solutions, and that the solutions of finite, weakly guarded io-sequence specifications
are rational io-sequences.
Lemma 5.3. For every weakly guarded io-sequence specification E and recursion
variable X of E, there exists a unique solution σ ∈ ±∞ of E for X. Moreover,
if E is finite then the solution σ of E for X is a rational io-sequence, and an
io-term that denotes σ can be computed on the input of E and X.
Let T be a stream specification, and f ∈ Σsfun . By exhaustiveness of T for
f, there is at least one defining rule for f in T . Since T is a constructor stream
TRS, it follows that every defining rule ρ for f is of the form:
f(p1 , . . . , pars (f) , q1 , . . . , qard (f) ) → u1 : . . . : um : s
(ρ)
with p1 , . . . , pars (f) ∈ Ter (C(Σ))S , q1 , . . . , qard (f) ∈ Ter (C(Σ))D , u1 , . . . , um ∈
Ter (Σ)D and s ∈ Ter (Σ)S where r oot(s) 6= ‘:’. If ρ is non-nesting then either
s ≡ σj or s ≡ g(w 1 : σφ(1) , . . . , wars (g) : σφ(ars (g)) , u′1 , . . . , u′ard (g) ) where wi : σi is
shorthand for wi,1 : . . . : wi,mi : σi , and φ : {1, . . . , ars (g)} → {1, . . . , ars (f)} is a
function that describes how the stream arguments are permuted and replicated.
We now define, for every given stream specification T , an infinite io-sequence
specification ET that will be instrumental for defining gate translations of the
stream function symbols in T .
36
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Definition 5.4. Let T = hΣ, Ri be a stream specification. Based on the set:
A := {hǫ, −i, hǫ, +i, hǫ, −+i}
n
o
f ∈ Σfnest , q ∈ N, ρ defining rule for f,
hf, ⋆ i, hf, ⋆ , ρi,
∪
1 ≤ i ≤ ars (f)
hf, i, qi, hf, i, q, ρi
of tuples we define the infinite io-sequence specification ET = {Xα = Eα }α∈A
by listing the equations of ET . We start with the equations
Xhǫ, −i = ⊥ ,
Xhǫ, +i = +Xhǫ, +i ,
Xhǫ, −+i = −+Xhǫ, −+i .
Then we let, for all friendly nesting (or flat) f ∈ Σsfun with arities k = ars (f)
and ℓ = ard (f):
(
min Xhf, ⋆, ρi | ρ a defining rule of f
if f is weakly guarded,
Xhf, ⋆i =
Xhǫ, −i
if f is unguarded,
and for all 1 ≤ i ≤ ars (f), and q ∈ N:
(V
Xhf, i, q, ρi | ρ a defining rule of f
Xhf, i, qi =
Xhǫ, −i
if f is weakly guarded,
if f is unguarded.
For specifying Xhf, ⋆, ρi and Xhf, i, q, ρi we distinguish the possible forms the rule
ρ can have. In doing so, we abbreviate terms ui,1 : . . . : ui,p : σi with σi a variable
of sort stream by ui : σi and let |ui | := p. If ρ is nesting, then we let
Xhf, ⋆, ρi = Xhǫ, +i ,
Xhf, i, q, ρi = Xhǫ, −+i .
Otherwise, ρ is non-nesting and of the form:
f((u1 : σ1 ), . . . , (uk : σk ), v1 , . . . , vℓ ) → w1 : . . . : wm : s ,
where either (a) s ≡ σj , or (b) s ≡ g((u′1 : σφ(1) ), . . . , (u′k′ : σφ(k′ ) ), v1′ , . . . , vℓ′ ′ )
with k ′ = ars (g), ℓ′ = ard (g), and φ : {1, . . . , k ′ } → {1, . . . , k}. Let:
(
Xhǫ, +i
case (a)
Xhf, ⋆, ρi =
m
+ Xhg, ⋆i case (b)
. q, q ′ := q −
. |u | in
Xhf, i, q, ρi = let p := |ui | −
i
′
q
case (a), i = j
+ Xhǫ, −+i
p m
− +
Xhǫ, +i
case (a), i 6= j
V
Xhg, j, q′ +|u′j |i | j ∈ φ−1 (i)
case (b)
where we agree
V
∅ := Xhǫ, +i .
We formally state an easy observation about the system ET defined in Def. 5.9,
and an immediate consequence due to Lem. 5.3.
Data-Oblivious Stream Productivity
37
Proposition 5.5. Let T be a stream specification. The io-sequence specification ET (defined in Def. 5.9) is weakly guarded. As a consequence, ET has a
unique solution in ±∞ for every variable X in ET .
Another easy observation is that the solutions for the variables Xhf, ⋆i in a
system ET correspond very directly to numbers in N.
Proposition 5.6. Let T be a stream specification and f ∈ Σsfun . The unique
solution of ET (defined in Def. 5.9) for Xhf, ⋆i is of the form +ω or +n ⊥.
Furthermore observe that ET is infinite in case that Σsfun 6= ∅, and hence
typically is infinite. Whereas Lem. 5.3 implies unique solvability of ET for individual variables in view of (the first statement in) Prop. 5.5, it will usually not
guarantee that these unique solutions are rational io-sequences. Nevertheless it
can be shown that all solutions of ET are rational io-sequences. For our purposes
it will suffice to show this only for the solutions of ET for certain of its variables.
We will only be interested in the solutions of ET for variables Xhf, i, 0i and
Xhf, ⋆i . It turns out that from ET a finite weakly guarded specification ET′ can be
extracted that, for each of the variables Xhf, i, 0i and Xhf, ⋆i , has the same solution
as ET , respectively. Lem. 5.7 below states that, for a stream specification T , the
finite io-sequence specification ET′ can always be obtained algorithmically, which
together with Lem. 5.3 implies that the unique solutions of ET for the variables
Xhf, i, 0i and Xhf, ⋆i are rational io-sequences for which representing io-terms can
be computed. As a consequence these solutions, for a stream specification T , of
the recursion system ET , can be viewed to represent p-i functions: the p-i functions that are represented by the io-term denoting the respective solution, a
rational io-sequence.
As an example let us consider an io-sequence specification that corresponds
to the defining rules for f in Fig. 2:
X = −++X ∧ −−+Y
Y = ++X ∧ −+Y .
The unique solution for X of this system is the rational io-sequence (and respectively, io-term) −−+, which is the translation of f (as mentioned earlier).
Lemma 5.7. Let T be a stream specification. There exists an io-sequence specification ET′ = {Xα = Eα′ }α∈A′ such that:
(i) ET′ is finite and weakly guarded;
(ii) {hf, i, 0i, hf, ⋆i | f ∈ Σsfun , i, q ∈ N, i ∈ {1, . . . , ars (f)} } ⊆ A′ ;
for each f ∈ Σsfun ∩ Σfnest and i ∈ {1, . . . , ars (f)}, ET′ has the same solution
for Xhf, i, 0i , and respectively for Xhf, ⋆i , as the io-sequence specifications ET
(see Def. 5.4);
(iii) on the input of T , ET′ can be computed.
Proof (Sketch). An algorithm for obtaining ET′ from ET can be obtained as
follows. On the input of ET , set E := ET and repeat the following step on E as
long as it is applicable:
38
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
(RPC) Detect and remove a reachable non-consuming pseudo-cycle from the iosequence specification E : Suppose that, for a function symbol h, for j, k, l ∈
N, and for a recursion variable Xhh, j, ki that is reachable from Xhf, i, 0i , we
w
have Xhh, j, ki −→ Xhh, j, li (from the recursion variable Xhh, j, ki the variable
Xhh, j, li is reachable via a path in the specification on which the finite iosequence w is encountered as the word formed by consecutive labels), where
k < l and w only contains symbols ‘+’. Then modify E by setting Xhh, j, ki =
Xhǫ, +i .
It is not difficult to show that a step (RPC) preserves weakly guardedness and
the unique solution of ET , and that, on the input of ET , the algorithm terminates
in finitely many steps, having produced an io-sequence specification ET′ with [f]i
as the solution for Xhf, i, 0i and the property that only finitely many recursion
variables are reachable in ET′ from Xhf, i, 0i .
⊓
⊔
As an immediate consequence of Lem. 5.7,
and of Lem. 5.3 we obtain the following lemma.
Lemma 5.8. Let T be a stream specification, and let f ∈ Σsfun .
(i) For each i ∈ {1, . . . , ars (f)} there is precisely one io-sequence σ that solves ET
for Xhf, i, 0i ; furthermore, σ is rational. Moreover there exists an algorithm
that, on the input of T , f, and i, computes the shortest io-term that denotes
the solution of ET for Xhf, i, 0i .
(ii) There is precisely one io-sequence σ that solves ET for Xhf, ⋆i ; this io-sequence
is rational. And there is an algorithm that, on the input of T and f, computes
the shortest io-term that denotes the solution of ET for Xhf, ⋆i .
Lem. 5.8 guarantees the well-definedness in the definition below of the ‘gate
translation’ for flat or friendly nesting stream functions in a stream specification.
Definition 5.9. Let T = hΣ, Ri be a stream definition. For each flat or friendly
nesting symbol f ∈ Σsfun ∩ Σfnest with stream arity k = ars (f) we define the gate
translation [f] of f by:
[f] := gate[f]⋆ ([f]1 , . . . , [f]ars (f) ) ,
where [f]⋆ ∈ N is defined by:
(
n the unique solution of ET for Xhf, ⋆i is +n ⊥
[f]⋆ :=
∞ the unique solution of ET for Xhf, ⋆i is +ω
and, for 1 ≤ i ≤ ars (f), [f]i is the shortest io-term that denotes the unique solution for Xhf, i, 0i of the weakly guarded io-sequence specification ET in Def. 5.4.
From Lem. 5.8 we also obtain that, for every stream specification, the function
which maps stream function symbols to their gate translations is computable.
Data-Oblivious Stream Productivity
39
Lemma 5.10. There is an algorithm that, on the input of a stream specification T , and a flat or friendly nesting symbol f ∈ Σsfun ∩ Σfnest , computes the
gate translation [f] of f.
Example 5.11. Consider a flat stream specification consisting of the rules:
f(x : σ) → x : g(σ, σ, σ) ,
g(x : y : σ, τ, υ) → x : g(y : τ, y : υ, y : σ) .
The translation of f is [f] = gate([f]1 ), where [f]1 is the unique solution for Xhf, 1, 0i
of the io-sequence specification ET :
Xhf, 1, 0i = −+(Xhg, 1, 0i ∧ Xhg, 2, 0i ∧ Xhg, 3, 0i )
Xhg, 1, 0i = −−+Xhg, 3, 1i
Xhg, 1, 1i = −+Xhg, 3, 1i
Xhg, 1, qi = +Xhg, 3, q−1i
(q ≥ 2)
Xhg, 2, qi = +Xhg, 1, q+1i
Xhg, 3, qi = +Xhg, 2, q+1i
(q ∈ N)
(q ∈ N)
By the algorithm referred to in Lem. 5.10 this infinite specification can be
+++
turned into a finite one. The ‘non-consuming pseudocycle’ Xhg, 3, 1i −→ Xhg, 3, 2i
justifies the modification of ET by setting Xhg, 3, 1i = +Xhg, 3, 1i ; likewise we set
Xhg, 3, 0i = +Xhg, 3, 0i . Furthermore all equations not reachable from Xhf, 1, 0i are
removed (garbage collection), and we obtain a finite specification ET′ , which, by
Lem. 5.3, has a periodically increasing solution, with io-term-denotation [f]1 =
h−+−−, +i. The reader may try to calculate the gate corresponding to g, it is:
[g] = gate(−−+, +−+, +).
Second, consider the flat stream specification with data constructor symbols
0 and 1:
f(0 : σ) → g(σ) ,
f(1 : x : σ) → x : g(σ) ,
g(x : y : σ) → x : y : g(σ) ,
denoted ρ f0 , ρ f1 , and ρ g , respectively. Then, [f]1 is the solution for Xhf, 1, 0i of
Xhf, 1, 0i = Xhf, 1, 0, ρ f0 i ∧ Xhf, 1, 0, ρ f1 i
Xhf, 1, 0, ρ f0 i = −Xhg, 1, 0i
Xhf, 1, 0, ρ f1 i = −−+Xhg, 1, 0i
Xhg, 1, 0i = −−++Xhg, 1, 0i .
Example 5.12. Consider a pure stream specification with the function layer:
f(x : σ) → x : g(σ, σ, σ) ,
g(x : y : σ, τ, υ) → x : g(y : τ, y : υ, y : σ) .
40
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
The translation of f is hfi, the unique solution for Xf of the system:
Xf (n) = min(Xf,⋆ (0), Xf,1 (n))
Xf,1 (n) = let n′ := n − 1
if n′ < 0 then 0 else 1 + inf Xg,1 (n′ ), Xg,2 (n′ ), Xg,3 (n′ )
Xf,⋆ (n) = 1 + Xg,⋆ (0)
Xg,1 (n) = let n′ := n − 2, if n′ < 0 then 0 else 1 + Xg,3 (1 + n′ )
Xg,2 (n) = 1 + Xg,1 (1 + n)
Xg,3 (n) = 1 + Xg,2 (1 + n)
Xg,⋆ (n) = 1 + Xf,⋆ (0)
An algorithm for solving such systems of equations is described in (the proof of)
Lemma 5.7; here we solve the system directly. Note that Xg,3 (n) = 1 + Xg,2 (n +
1) = 2 + Xg,1 (n + 2) = 3 + Xg,3 (n), hence ∀n ∈ N. Xg,3 (n) = ∞. Likewise we
obtain Xg,2 (n) = ∞ if n ≥ 1 and 1 for n = 0, and Xg,1 (n) = ∞ if n ≥ 2 and 0
for n ≤ 1. Then we get hfi(0) = 0, hfi(1) = hfi(2) = 1, and hfi(n) = ∞ for all
n ≥ 2, represented by the gate hfi = gate(−+−−+). The gate corresponding to
g is hgi = gate(−−+, +−+, +).
Example 5.13. Consider a flat stream function specification with the following
rules which use pattern matching on the data constructors 0 and 1:
f(0 : σ) → g(σ)
f(1 : x : σ) → x : g(σ)
g(x : y : σ) → x : y : g(σ)
denoted ρ f 0 , ρ f 1 , and ρ g , respectively. Then, hfi is the solution for Xf,1 of:
Xf (n) = min(Xf,⋆ (0), Xf,1 (n))
Xf,1 (n) = inf Xf,1,ρ f 0 (n), Xf,1,ρ f 1 (n)
Xf,1,ρ f 0 (n) = let n′ := n − 1, if n′ < 0 then 0 else Xg,1 (n′ )
Xf,1,ρ f 1 (n) = let n′ := n − 2, if n′ < 0 then 0 else 1 + Xg,1 (n′ )
Xf,⋆ (n) = min(Xg,⋆ (0), 1 + Xg,⋆ (0))
Xg,1 (n) = let n′ := n − 2, if n′ < 0 then 0 else 2 + Xg,1 (n′ )
Xg,⋆ (n) = 2 + Xg,⋆ (0) .
As solution we obtain an overlapping of both traces hfi1,ρ f 0 and hfi1,ρ f 1 , that is,
. 2 represented by the gate hfi = gate(−−−+).
hfi1 (n) = n −
The following proposition explains the correspondence between Def. 5.1 and
Def. 5.9.
Proposition 5.14. Let T be a stream specification. For each f ∈ Σsfun it holds:
hfi(n1 , . . . , nk ) = [[[f ]]](n1 , . . . , nk )
(for all n1 , . . . , nk ∈ N),
where k = ars (f). That is, the p-i function translation hfi of f coincides with the
p-i function that is represented by the gate translation [f].
Data-Oblivious Stream Productivity
41
The following lemma states that the translation hfi of a flat stream function
symbol f (as defined in Def. 5.1) is the d-o lower bound on the production
function of f. For friendly nesting stream symbols f it states that hfi pointwisely
bounds from below the d-o lower bound on the production function of f.
Lemma 5.15. Let T be a stream specification, and let f ∈ Σfnest ⊇ Σflat .
(i) If f is flat, then: hfi = do T (f). Hence, do T (f) is periodically increasing.
(ii) If f is friendly nesting, then it holds: hfi ≤ do T (f) (pointwise inequality).
5.2
Translation of Stream Constants
In the second step, we now define a translation of stream constants in a flat or
friendly nesting stream specification into production terms under the assumption that gate translations for the stream functions are given. Here the idea is
that the recursive definition of a stream constant M is unfolded step by step; the
terms thus arising are translated according to their structure using gate translations of the stream function symbols from a given family of gates; whenever
a stream constant is met that has been unfolded before, the translation stops
after establishing a binding to a µ-binder created earlier.
Definition 5.16. Let T = hΣ, Ri be a stream specification, and F = {γ f }f∈Σsfun
a family of gates that are associated with the symbols in Σsfun .
Let Ter (Σ)0S be the set of terms in T of sort stream that do not contain
variables of sort stream. The translation function [·]F : Ter (Σ)0S → P is defined
F
by s 7→ [s]F := [s]F
∅ based on the following definition of expressions [s]α , where
0
α ⊆ Σscon , by induction on the structure of s ∈ Ter (Σ)S , using the clauses:
(
µM.min {[r]F
α∪{M} | M(v) → r ∈ R} if M 6∈ α
[M(u)]F
α :=
M
if M ∈ α
F
[u : s]F
α := •([s]α )
F
F
[f(s1 , . . . , sars (f) , u1 , . . . , uard (f) )]F
α := γ f ([s1 ]α , . . . , [sars (f) ]α )
where M ∈ Σscon , u = hu1 , . . . , uard (M) i a vector of terms in Ter (Σ)D , f ∈ Σsfun ,
s, s1 , . . . sars (f) ∈ Ter (Σ)0S , and u, u1 , . . . , uard (f) ∈ Ter (Σ)D . Note that the definition of [M(u)]F
α does not depend on the vector u of terms in Ter (Σ)D . Therefore we define, for all M ∈ Σscon , the translation of M with respect to F by
[M]F := [M(x)]F
∅ ∈ P where x = hx1 , . . . , xard (M) i is a vector of data variables.
The following lemma is the basis of our result in Sec. 6, Thm. 6.1, concerning
the decidability of d-o productivity for flat stream specifications. In particular
the lemma states that if we use gates that represent d-o optimal lower bounds
on the production of the stream functions, then the translation of a stream
constant M yields a production term that rewrites to the d-o lower bound of the
production of M.
42
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Lemma 5.17. Let T = hΣ, Ri be a stream specification, and F = {γ f }f∈Σsfun be
a family of gates such that, for all f ∈ Σsfun , the arity of γ f equals the stream
arity of f. Then the following statements hold:
(i) Suppose that [[γ f ]] = do T (f) holds for all f ∈ Σsfun . Then for all M ∈ Σscon
and vectors u = hu1 , . . . , uard (M) i of data terms [[[M]F ]] = do T (M(u)) holds.
And consequently, T is d-o productive if and only if [[[M0 ]F ]] = ∞.
(ii) Suppose that [[γ f ]] = do T (f) holds for all f ∈ Σsfun . Then for all M ∈ Σscon
and vectors u = hu1 , . . . , uard (M) i of data terms [[[M]F ]] = do T (M(u)) holds.
And consequently, T is d-o non-productive if and only if [[[M0 ]F ]] < ∞.
Lem. 5.17 is an immediate consequence of the following lemma, which also is the
basis of our result in Sec. 6 concerning the recognizability of productivity for flat
and friendly nesting stream specifications. In particular the lemma below asserts
that if we use gates that represent p-i functions which are lower bounds on the
production of the stream functions, then the translation of a stream constant M
yields a production term that rewrites to a number in N smaller or equal to the
d-o lower bound of the production of M.
Lemma 5.18. Let T be a stream specification, and let F = {γ f }f∈Σsfun be a
family of gates such that, for all f ∈ Σsfun , the arity of γ f equals the stream
arity of f. Suppose that one of the following statements holds:
(a)
(b)
(c)
(d)
[[γ f ]] ≤ do T (f)
[[γ f ]] ≥ do T (f)
do T (f) ≤ [[γ f ]]
do T (f) ≥ [[γ f ]]
for
for
for
for
all
all
all
all
f
f
f
f
∈ Σsfun ;
∈ Σsfun ;
∈ Σsfun ;
∈ Σsfun .
Then, for all M ∈ Σscon and vectors u = hu1 , . . . , uard (M) i of data terms in T ,
the corresponding one of the following statements holds:
(a)
(b)
(c)
(d)
[[[M]F ]] ≤ do T (M(u)) ;
[[[M]F ]] ≥ do T (M(u)) ;
do T (M(u)) ≤ [[[M]F ]] ;
do T (M(u)) ≥ [[[M]F ]] .
6
Deciding Data-Oblivious Productivity
In this section we assemble our results concerning decision of d-o productivity,
and automatable recognition of productivity. We define methods:
(DOP) for deciding d-o productivity of flat stream specifications,
(DP) for deciding productivity of pure stream specifications, and
(RP) for recognising productivity of friendly nesting stream specifications,
that proceed in the following steps:
Data-Oblivious Stream Productivity
43
(i) Take as input a (DOP) flat, (DP) pure, or (RP) friendly nesting stream specification T = hΣ, Ri.
(ii) Translate the stream function symbols into gates F := {hfi}f∈Σsfun (Def. 5.1).
(iii) Construct the production term [M0 ]F with respect to F (Def. 5.16).
(iv) Compute the production k of [M0 ]F using →R (Def. 4.27).
(v) Give the following output:
(DOP) “T is d-o productive” if k = ∞, else “T is not d-o productive”.
(DP) “T is productive” if k = ∞, else “T is not productive”.
(RP) “T is productive” if k = ∞, else “don’t know”.
Note that all of these steps are automatable (cf. our productivity tool, Sec. 8).
Our main result states that d-o productivity is decidable for flat stream specifications. It rests on the fact that the algorithm DOP obtains the lower bound
do T ([M0 ]) on the production of M0 in T . Since d-o productivity implies productivity (Prop. 3.4), we obtain a computable, d-o optimal, sufficient condition
for productivity of flat stream specifications, which cannot be improved by any
other d-o analysis. Second, since for pure stream specifications d-o productivity
and productivity are the same, we get that productivity is decidable for them.
Theorem 6.1. (i) DOP decides d-o productivity of flat stream specifications,
(ii) DP decides productivity of pure stream specifications.
Proof. Let k be the production of the term [M0 ]F ∈ P in step (iv) of DOP/DP.
(i) By Lem. 5.15 (i), Lem. 5.17 (i), Lem. 4.28, Lem. 4.29 we find: k = do T (M0 ).
⊓
⊔
(ii) For pure specifications we additionally note: ΠT (M0 ) = do T (M0 ).
Third, we obtain a computable, sufficient condition for productivity of friendly
nesting stream specifications.
Theorem 6.2. A friendly nesting (flat) stream specification T is productive if
the algorithm RP( DOP) recognizes T as productive.
Proof. Let k be the production of the term [M0 ]F ∈ P in step (iv) of RP/DOP. By
Lem. 5.15 (ii), Lem. 5.18 (a), Lem. 4.28 and Lem. 4.29: k ≤ do T (M0 ) ≤ ΠT (M0 ).
⊓
⊔
Example 6.3. We illustrate the translation and decision of d-o productivity by
means of Pascal’s triangle, Fig. 2. The translation of the stream function symbols is F = {hfi} with hfi = gate(−−+), see page 32. We calculate [P]F , the
translation of P, and reduce it to normal form with respect to →R :
[P]F = µP.•(•(−−+(P ))) ։R µP.++−−+(P ) →R ∞
Hence do T (P) = ∞, and P is d-o productive and therefore productive.
44
7
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Examples
Productivity of all of the following examples is recognized fully automatically by
a Haskell implementation of our decision algorithm for data-oblivious productivity. The tool and a number of examples can be found at:
http://infinity.few.vu.nl/productivity.
In Subsections 7.1–7.6 below, we give possible input-representations as well as
the tool-output for the examples of stream specifications in Section 2 and for an
example of a stream function specification in Section 3.
7.1
Example in Fig. 2 on Page 3
For applying the automated productivity prover to the flat stream specification
in Fig. 2 on page 2 of the stream of rows in Pascal’s triangle, one can use the
input:
Signature(
P : stream(nat),
0 : nat,
f : stream(nat) -> stream(nat),
a : nat -> nat -> nat,
s : nat -> nat
)
P = 0:s(0):f(P)
f(s(x):y:sigma) = a(s(x),y):f(y:sigma)
f(0:sigma) = 0:s(0):f(sigma)
a(s(x),y) = s(a(x,y))
a(0,y) = y
On this input the automated productivity prover produces the following output:
The automated productivity prover has been applied to:
Signature(
-- stream symbols -P : stream(nat),
f : stream(nat) -> stream(nat),
-- data symbols -0 : nat,
a : nat -> nat -> nat,
s : nat -> nat
)
-- stream layer -P = 0:s(0):f(P)
f(s(x):y:sigma) = a(s(x),y):f(y:sigma)
Data-Oblivious Stream Productivity
45
f(0:sigma) = 0:s(0):f(sigma)
-- data layer -a(s(x),y) = s(a(x,y))
a(0,y) = y
Termination of the data layer has been proven automatically.
The function symbol f is flat, we can compute the precise data-oblivious lower
bound:
[f] = gate([f]⋆,0 (0), [f]1,0 )
n
−+ ∧ f1,1
−−+ ∧ µf1,1 . ∧
n
++ ∧ f1,0
[f]1,0 = µf1,0 . ∧
n
−++ ∧ f1,0
= −−+
[f]⋆,0
(
+f⋆,0
= µf⋆,0 . ∧
++f⋆,0
=+
P depends only on flat stream functions, we can decide data-oblivious productivity.
We translate P into a pebbleflow net and collapse it to a source:5
[P] = µP.•(•([f](P )))
= µP.•(•(−−+(P )))
։R µP.+−+(•(−−+(P )))
։R µP.+−+(+−+(−−+(P )))
։R µP.++−(−−+(P ))
։R µP.++−−+(P )
։R ∞
The specification of P is productive.
7.2
Example in Fig. 3 on Page 6
For applying the automated productivity prover to the flat stream specification
in Fig. 3 on page 6 of the ternary Thue-Morse sequence, one can use the input:
Signature(
Q, Qprime : stream(char),
46
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
f : stream(char) -> stream(char),
a, b, c: char
)
Q = a:Qprime
Qprime = b:c:f(Qprime)
f(a:sigma) = a:b:c:f(sigma)
f(b:sigma) = a:c:f(sigma)
f(c:sigma) = b:f(sigma)
On this input the automated productivity prover produces the following output:
The automated productivity prover has been applied to:
Signature(
-- stream symbols -Q : stream(char),
Qprime : stream(char),
f : stream(char) -> stream(char),
-- data symbols -a : char,
b : char,
c : char
)
-- stream layer -Q = a:Qprime
Qprime = b:c:f(Qprime)
f(a:sigma) = a:b:c:f(sigma)
f(b:sigma) = a:c:f(sigma)
f(c:sigma) = b:f(sigma)
-- data layer --
Data-Oblivious Stream Productivity
47
The function symbol f is flat, we can compute the precise data-oblivious lower
bound:
[f] = gate([f]⋆,0 (0), [f]1,0 )
n
−+++
∧
f1,0
n
[f]1,0 = µf1,0 . ∧ −++ ∧ f1,0
n
−+ ∧ f1,0
= −+
[f]⋆,0
+++f⋆,0
= µf⋆,0 . ∧ ++f⋆,0
+f⋆,0
=+
Q depends only on flat stream functions, we can decide data-oblivious productivity.
We translate Q into a pebbleflow net and collapse it to a source:6
[Q] = µQ.•(µQprime.•(•([f](Qprime))))
= 1 µQ.•(µQprime.•(•(−+(Qprime))))
։R µQ.+−+(µQprime.•(•(−+(Qprime))))
։R µQ.+−+(µQprime.+−+(•(−+(Qprime))))
։R µQ.+−+(µQprime.+−+(+−+(−+(Qprime))))
։R µQ.+−+(µQprime.++−(−+(Qprime)))
։R µQ.+−+(µQprime.++−(Qprime))
։R µQ.+−+(∞)
։R µQ.∞
։R ∞
The specification of Q is productive.
Qprime depends only on flat stream functions, we can decide data-oblivious
productivity.
48
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
We translate Qprime into a pebbleflow net and collapse it to a source:7
[Qprime] = µQprime.•(•([f](Qprime)))
= µQprime.•(•(−+(Qprime)))
։R µQprime.+−+(•(−+(Qprime)))
։R µQprime.+−+(+−+(−+(Qprime)))
։R µQprime.++−(−+(Qprime))
։R µQprime.++−(Qprime)
։R ∞
The specification of Qprime is productive.
7.3
Example in Fig. 4 on Page 7
For applying the automated productivity prover to the pure stream specification
in Fig. 4 on page 7 of the ternary Thue-Morse sequence, one can use the input:
Signature(
Q : stream(char),
M : stream(bit),
zip : stream(x) -> stream(x) -> stream(x),
inv : stream(bit) -> stream(bit),
tail : stream(x) -> stream(x),
diff : stream(bit) -> stream(char),
i : bit -> bit,
X : bit -> bit -> char,
0, 1 : bit,
a,b,c : char
)
Q = diff(M)
M = 0:zip(inv(M),tail(M))
zip(x:s,t) = x:zip(t,s)
inv(x:s) = i(x):inv(s)
tail(x:s) = s
diff(x:y:s) = X(x,y):diff(y:s)
i(0) =
i(1) =
X(0,0)
X(0,1)
X(1,0)
X(1,1)
1
0
=
=
=
=
b
a
c
b
The automated productivity prover then gives the following output:
Data-Oblivious Stream Productivity
The automated productivity prover has been applied to:
Signature(
-- stream symbols -Q : stream(char),
M : stream(bit),
zip : stream(x) -> stream(x) -> stream(x),
inv : stream(bit) -> stream(bit),
tail : stream(x) -> stream(x),
diff : stream(bit) -> stream(char),
-- data symbols -i : bit -> bit,
X : bit -> bit -> char,
0 : bit,
1 : bit,
a : char,
b : char,
c : char
)
-- stream layer -Q = diff(M)
M = 0:zip(inv(M),tail(M))
zip(x:s,t) = x:zip(t,s)
inv(x:s) = i(x):inv(s)
tail(x:s) = s
diff(x:y:s) = X(x,y):diff(y:s)
-- data layer -i(0) = 1
i(1) = 0
X(0,0) = b
X(0,1) = a
X(1,0) = c
X(1,1) = b
Termination of the data layer has been proven automatically.
49
50
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
The function symbol zip is pure, we can compute its precise production modulus:
[zip] = gate([zip]⋆,0 (0), [zip]1,0 , [zip]2,0 )
n
n
n
n
[zip]1,0 = µzip 1,0 . ∧ −+ ∧ µzip 2,0 . ∧ + ∧ zip 1,0
= −++
n
n
n
n
[zip]2,0 = µzip 2,0 . ∧ + ∧ µzip 1,0 . ∧ −+ ∧ zip 2,0
= +−+
n
[zip]⋆,0 = µzip ⋆,0 . ∧ +zip ⋆,0
=+
The function symbol inv is pure, we can compute its precise production modulus:
[inv] = gate([inv]⋆,0 (0), [inv]1,0 )
n
n
[inv]1,0 = µinv 1,0 . ∧ −+ ∧ inv 1,0
= −+
n
[inv]⋆,0 = µinv ⋆,0 . ∧ +inv ⋆,0
=+
The function symbol tail is pure, we can compute its precise production modulus:
[tail] = gate([tail]⋆,0 (0), [tail]1,0 )
n
[tail]1,0 = µtail 1,0 . ∧ −µx.−+x
= −−+
n
[tail]⋆,0 = µtail ⋆,0 . ∧ µx.+x
=+
The function symbol diff is pure, we can compute its precise production modulus:
[diff] = gate([diff]⋆,0 (0), [diff]1,0 )
n
n
n
n
[diff]1,0 = µdiff 1,0 . ∧ −−+ ∧ µdiff 1,1 . ∧ −+ ∧ diff 1,1
= −−+
n
[diff]⋆,0 = µdiff ⋆,0 . ∧ +diff ⋆,0
=+
Data-Oblivious Stream Productivity
51
Q depends only on pure stream functions, we can decide productivity.
We translate Q into a pebbleflow net and collapse it to a source:8
[Q] = µQ.[diff](µM.•([zip]([inv](M ), [tail](M ))))
= µQ.−−+(µM.•(min(−++(−+(M )), +−+(−−+(M )))))
։R µQ.−−+(µM.•(min(−++(M ), +−−++(M ))))
։R µQ.−−+(µM.+−+(min(−++(M ), +−−++(M ))))
։R µQ.−−+(µM.min(+−+(−++(M )), +−+(+−−++(M ))))
։R µQ.−−+(µM.min(+−+(M ), ++−−++(M )))
։R µQ.−−+(min(µM.+−+(M ), µM.++−−++(M )))
։R µQ.min(−−+(µM.+−+(M )), −−+(µM.++−−++(M )))
։R min(µQ.−−+(µM.+−+(M )), µQ.−−+(µM.++−−++(M )))
։R min(µQ.−−+(∞), µQ.−−+(∞))
։R min(µQ.∞, µQ.∞)
։R min(∞, ∞)
։R ∞
The specification of Q is productive.
M depends only on pure stream functions, we can decide productivity.
We translate M into a pebbleflow net and collapse it to a source:8
[M] = µM.•([zip]([inv](M ), [tail](M )))
= µM.•(min(−++(−+(M )), +−+(−−+(M ))))
։R µM.•(min(−++(M ), +−−++(M )))
։R µM.+−+(min(−++(M ), +−−++(M )))
։R µM.min(+−+(−++(M )), +−+(+−−++(M )))
։R µM.min(+−+(M ), ++−−++(M ))
։R min(µM.+−+(M ), µM.++−−++(M ))
։R min(∞, ∞)
։R ∞
The specification of M is productive.
7.4
Example in Fig. 5 on Page 7
For applying the automated productivity prover to the pure stream specification
in Fig. 5 on page 7 of the Thue–Morse sequence using the D0L-system 0 7→ 01,
1 7→ 10, one can use the input:
Signature(
52
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
M : stream(bit),
Mprime : stream(bit),
h : stream(bit) -> stream(bit),
0, 1 : bit
)
M = 0:Mprime
Mprime = 1:h(Mprime)
h(0:sigma) = 0:1:h(sigma)
h(1:sigma) = 1:0:h(sigma)
The automated productivity prover then gives the following output:
The automated productivity prover has been applied to:
Signature(
-- stream symbols -M : stream(bit),
Mprime : stream(bit),
h : stream(bit) -> stream(bit),
-- data symbols -0 : bit,
1 : bit
)
-- stream layer -M = 0:Mprime
Mprime = 1:h(Mprime)
h(0:sigma) = 0:1:h(sigma)
h(1:sigma) = 1:0:h(sigma)
-- data layer --
The function symbol h is pure, we can compute its precise production modulus:
[h] = gate([h]⋆,0 (0), [h]1,0 )
n
−++ ∧ h1,0
n
[h]1,0 = µh1,0 . ∧
−++ ∧ h1,0
= −++
[h]⋆,0 = µh⋆,0 . ∧
(
++h⋆,0
++h⋆,0
=+
M depends only on pure stream functions, we can decide productivity.
Data-Oblivious Stream Productivity
53
We translate M into a pebbleflow net and collapse it to a source:9
[M] = µM.•(µM prime.•([h](M prime)))
= µM.•(µM prime.•(−++(M prime)))
։R µM.+−+(µM prime.•(−++(M prime)))
։R µM.+−+(µM prime.+−+(−++(M prime)))
։R µM.+−+(µM prime.+−+(M prime))
։R µM.+−+(∞)
։R µM.∞
։R ∞
The specification of M is productive.
Mprime depends only on pure stream functions, we can decide productivity.
We translate Mprime into a pebbleflow net and collapse it to a source:9
[Mprime] = µM prime.•([h](M prime))
= µM prime.•(−++(M prime))
։R µM prime.+−+(−++(M prime))
։R µM prime.+−+(M prime)
։R ∞
The specification of Mprime is productive.
7.5
Example in Fig. 6 on Page 8
For applying the automated productivity prover to the pure stream specification
in Fig. 6 on page 8 of the Thue–Morse sequence using the D0L-system 0 7→ 01,
1 7→ 10, one can use the input:
Signature(
nats, ones : stream(nat),
conv, add : stream(nat) -> stream(nat) -> stream(nat),
times : stream(nat) -> nat -> stream(nat),
0 : nat,
s : nat -> nat,
a, m : nat -> nat -> nat
)
nats = 0:conv(ones,ones)
ones = s(0):ones
conv(x:sigma,y:tau) = m(x,y):add(times(tau,x),conv(sigma,y:tau))
times(x:sigma,y) = m(x,y):times(sigma,y)
add(x:sigma,y:tau) = a(x,y):add(sigma,tau)
54
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
a(0,y) = y
a(s(x),y) = s(a(x,y))
m(0,y) = 0
m(s(x),y) = a(y,m(x,y))
The automated productivity gives as output:
The automated productivity prover has been applied to:
Signature(
-- stream symbols -nats : stream(nat),
ones : stream(nat),
conv : stream(nat) -> stream(nat) -> stream(nat),
add : stream(nat) -> stream(nat) -> stream(nat),
times : stream(nat) -> nat -> stream(nat),
-- data symbols -0 : nat,
s : nat -> nat,
a : nat -> nat -> nat,
m : nat -> nat -> nat
)
-- stream layer -nats = 0:conv(ones,ones)
ones = s(0):ones
conv(x:sigma,y:tau) = m(x,y):add(times(tau,x),conv(sigma,y:tau))
times(x:sigma,y) = m(x,y):times(sigma,y)
add(x:sigma,y:tau) = a(x,y):add(sigma,tau)
-- data layer -a(0,y) = y
a(s(x),y) = s(a(x,y))
m(0,y) = 0
m(s(x),y) = a(y,m(x,y))
Termination of the data layer has been proven automatically.
The function symbol conv is friendly nesting, we can compute a data-oblivious
Data-Oblivious Stream Productivity
55
lower bound:
[conv] = gate([conv]⋆,0 (0), [conv]1,0 , [conv]2,0 )
n
[conv]1,0 = µconv 1,0 . ∧ µx.−+x
= −+
n
[conv]2,0 = µconv 2,0 . ∧ µx.−+x
= −+
n
[conv]⋆,0 = µconv ⋆,0 . ∧ µx.−+x
= −+
The function symbol add is pure, we can compute its precise production modulus:
[add] = gate([add]⋆,0 (0), [add]1,0 , [add]2,0 )
n
n
[add]1,0 = µadd 1,0 . ∧ −+ ∧ add 1,0
= −+
n
n
[add]2,0 = µadd 2,0 . ∧ −+ ∧ add 2,0
= −+
n
[add]⋆,0 = µadd ⋆,0 . ∧ +add ⋆,0
=+
The function symbol times is pure, we can compute its precise production modulus:
[times] = gate([times]⋆,0 (0), [times]1,0 )
n
n
[times]1,0 = µtimes 1,0 . ∧ −+ ∧ times 1,0
= −+
n
[times]⋆,0 = µtimes ⋆,0 . ∧ +times ⋆,0
=+
nats depends only on flat stream functions, we try to prove productivity.
56
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
We translate nats into a pebbleflow net and collapse it to a source:10
[nats] = µnats.•([conv](µones.•(ones), µones.•(ones)))
= µnats.•(min(−+(0), min(−+(µones.•(ones)), −+(µones.•(ones)))))
։R µnats.+−+(min(−+(0), min(−+(µones.•(ones)), −+(µones.•(ones)))))
։R µnats.min(+−+(−+(0)), +−+(min(−+(µones.•(ones)), −+(µones.•(ones)))))
։R µnats.min(+−(0), +−+(min(−+(µones.•(ones)), −+(µones.•(ones)))))
։R min(µnats.+−(0), µnats.+−+(min(−+(µones.•(ones)), −+(µones.•(ones)))))
։R min(µnats.1, µnats.min(+−+(−+(µones.•(ones))), +−+(−+(µones.•(ones)))))
։R min(µnats.1, µnats.min(+−(µones.•(ones)), +−(µones.•(ones))))
։R min(1, min(µnats.+−(µones.•(ones)), µnats.+−(µones.•(ones))))
։R min(1, min(µnats.+−(µones.+−+(ones)), µnats.+−(µones.+−+(ones))))
։R min(1, min(µnats.+−(∞), µnats.+−(∞)))
։R min(1, min(µnats.∞, µnats.∞))
։R min(1, min(∞, ∞))
։R min(1, ∞)
։R 1
Failed to prove productivity of nats.
ones depends only on pure stream functions, we can decide productivity.
We translate ones into a pebbleflow net and collapse it to a source:10
[ones] = µones.•(ones)
= µones.•(ones)
։R µones.+−+(ones)
։R ∞
The specification of ones is productive.
7.6
Example on Page 12
Now we consider the stream function specification below Def. 3.5 on page 12,
which is not a stream specification because it does not contain a stream constant
as its root. Still, the automated productivity prover can be used to compute the
precise data-oblivious lower bound of the stream function f. On the input:
Signature(
f : stream(bit) -> stream(bit),
g : stream(bit) -> stream(bit) -> stream(bit),
0, 1 : bit
)
Data-Oblivious Stream Productivity
57
f(s) = g(s,s)
g(0:y:s,x:t) = 0:0:g(s,t)
g(1:s,x1:x2:x3:x4:t) = 0:0:0:0:0:g(s,t)
the productivity prover generates the following output:
The automated productivity prover has been applied to:
Signature(
-- stream symbols -f : stream(bit) -> stream(bit),
g : stream(bit) -> stream(bit) -> stream(bit),
-- data symbols -0 : bit,
1 : bit
)
-- stream layer -f(s) = g(s,s)
g(0:y:s,x:t) = 0:0:g(s,t)
g(1:s,x1:x2:x3:x4:t) = 0:0:0:0:0:g(s,t)
-- data layer --
The function symbol f is flat, we can compute the precise data-oblivious lower
bound:
[f] = gate([f]⋆,0 (0), [f]1,0 )
n
−−++ ∧ g1,0
n
µg1,0 . ∧
−+++++ ∧ g1,0
n
[f]1,0 = µf1,0 . ∧ ∧
−++ ∧ g2,0
n
µg2,0 . ∧
−−−−+++++ ∧ g2,0
[f]⋆,0
= −−−−++−++−+−−++−+−++−
(
(
++g⋆,0
= µf⋆,0 . ∧ µg⋆,0 . ∧
+++++g⋆,0
=+
58
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
The function symbol g is flat, we can compute the precise data-oblivious lower
bound:
[g] = gate([g]⋆,0 (0), [g]1,0 , [g]2,0 )
n
−−++ ∧ g1,0
n
[g]1,0 = µg1,0 . ∧
−+++++ ∧ g1,0
= −−++
[g]2,0 = µg2,0 . ∧
[g]⋆,0
n
−++ ∧ g2,0
n
−−−−+++++ ∧ g2,0
= −−−−++−++−+
(
++g⋆,0
= µg⋆,0 . ∧
+++++g⋆,0
=+
8
Conclusion and Further Work
In order to formalize quantitative approaches for recognizing productivity of
stream specifications, we defined the notion of d-o rewriting and investigated d-o
productivity. For the syntactic class of flat stream specifications (that employ
pattern matching on data), we devised a decision algorithm for d-o productivity. In this way we settled the productivity recognition problem for flat stream
specifications from a d-o perspective. For the even larger class including friendly
nesting stream function rules, we obtained a computable sufficient condition for
productivity. For the subclass of pure stream specifications (a substantial extension of the class given in [2]) we showed that productivity and d-o productivity
coincide, and thereby obtained a decision algorithm for productivity of pure
specifications.
We have implemented in Haskell the decision algorithm for d-o productivity.
This tool, together with more information including a manual, examples, our
related papers, and a comparison of our criteria with those of [3, 9, 1] can be
found at our web page http://infinity.few.vu.nl/productivity. The reader is
invited to experiment with our tool.
It is not possible to obtain a d-o optimal criterion for non-productivity of
flat specifications in an analogous way to how we established such a criterion
for productivity. This is because d-o upper bounds on the production of stream
functions in flat stream specifications are not in general periodically increasing
Data-Oblivious Stream Productivity
59
functions. For example, for the following stream function specification:
f(x : σ, τ ) → x : f(σ, τ ) ,
f(σ, y : τ ) → y : f(σ, τ ) ,
it holds that do (f)(n1 , n2 ) = n1 + n2 , which is not p-i. While this example is not
orthogonal, do (f) is also not p-i for the following similar orthogonal example:
f(0 : x : σ, y : τ ) → x : f(σ, τ ) ,
f(1 : σ, x : y : τ ) → y : f(σ, τ ) .
Currently we are developing a method that goes beyond a d-o analysis, one
that would, e.g., prove productivity of the example B given in the introduction.
Moreover, we study a refined production calculus that accounts for the delay
of evaluation of stream elements, in order to obtain a faithful modelling of lazy
evaluation, needed for example for S on page 4, where the first element depends
on a ‘future’ expansion of S.
Acknowledgement. We thank Jan Willem Klop, Carlos Lombardi, Vincent van
Oostrom, and Roel de Vrijer for useful discussions.
References
1. W. Buchholz. A Term Calculus for (Co-)Recursive Definitions on Streamlike Data
Structures. Annals of Pure and Applied Logic, 136(1-2):75–90, 2005.
2. J. Endrullis, C. Grabmayer, D. Hendriks, A. Isihara, and J.W. Klop. Productivity
of Stream Definitions. In Proc. of FCT 2007, LNCS, pages 274–287. Springer,
2007.
3. J. Hughes, L. Pareto, and A. Sabry. Proving the Correctness of Reactive Systems
Using Sized Types. In POPL ’96, pages 410–423, 1996.
4. D. Kapur, P. Narendran, D.J. Rosenkrantz, and H. Zhang. Sufficient-Completeness,
Ground-Reducibility and their Complexity. Acta Informatica, 28(4):311–350, 1991.
5. J.W. Klop and R. de Vrijer.
Infinitary Normalization.
In S. Artemov, H. Barringer, A.S. d’Avila Garcez, L.C. Lamb, and J. Woods,
editors, We Will Show Them: Essays in Honour of Dov Gabbay,
volume 2, pages 169–192. College Publications, 2005.
Item 95 at
http://web.mac.com/janwillemklop/iWeb/Site/Bibliography.html.
6. J.J.M.M. Rutten. Behavioural Differential Equations: a Coinductive Calculus of
Streams, Automata, and Power Series. Theoretical Computer Science, 308(1-3):1–
53, 2003.
7. A. Salomaa. Jewels of Formal Language Theory. Pitman Publishing, 1981.
8. B.A. Sijtsma. On the Productivity of Recursive List Definitions. ACM Transactions
on Programming Languages and Systems, 11(4):633–649, 1989.
9. A. Telford and D. Turner. Ensuring Streams Flow. In AMAST, pages 509–
523, 1997.
10. Terese. Term Rewriting Systems, volume 55 of Cambridge Tracts in Theoretical
Computer Science. Cambridge University Press, 2003.
60
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Appendix A
Solving Weakly Guarded IO-Term
Specifications (Lemma 5.3)
Let E = {Xα = Eα }α∈A be a weakly guarded finite io-sequence specification,
root ∈ A, and σroot the unique solution of E for Xroot .
Definition A.1. We define a finite, rooted graph GE = hV, L, vroot i that represents traces through the expressions in E. The set of nodes V consists of all pairs
of α ∈ A together with positions in Eα , the right-hand side of the corresponding
equation in E:
V = {hα, pi | α ∈ A, p ∈ Pos(Eα )} .
Note that the set of nodes V is finite. The root node vroot is hroot , hii. The set of
labelled edges L ⊆ V × {−, +, ǫ} × V is the smallest set such that: for all α ∈ A,
and every position p ∈ Pos(Eα ) the following conditions hold:
ǫ
(i) if Eα |p = Xα′ then hα, pi → hα′ , hii ∈ L;
−
(ii) if Eα |p = −E ′ then hα, pi → hα, p · 1i ∈ L;
+
(iii) if Eα |p = +E ′ then hα, pi → hα, p · 1i ∈ L;
ǫ
ǫ
(iv) if Eα |p = E ′ ∧ E ′′ then hα, pi → hα, p · 1i and hα, pi → hα, p · 2i ∈ L.
ℓ
where we have written v → w to denote hv, ℓ, wi ∈ L. Moreover, for ℓ ∈ {−, +, ǫ},
ℓ
ℓ
we let → denote the relation {hv, wi | v → w}.
ǫ
Note that by weak guardedness of E the relation → is terminating.
We proceed by defining the set of traces T (GE , v) through the graph GE
starting at node v ∈ V.
l
l
2
1
. . . in GE , where l1 , l2 , . . . ∈ {−, +, ǫ},
v2 →
Definition A.2. For a path ρ : v1 →
∞
we use trace(ρ) to denote l1 · l2 · · · ∈ ± , the word obtained by concatenating
the symbols ‘−’ and ‘+’ along the path ρ, thus treating the label ‘ǫ’ as the
empty word. Note that trace(ρ) ∈ ±ω whenever the path ρ is infinite by weak
guardedness of E. The set of traces T (GE , v) ⊆ ±ω through GE from v ∈ V is
defined as follows:
T (GE , v) := {trace(ρ) | ρ is an infinite path starting from node v in GE } .
V
. . . ∧ τn , where ∧ is the
For Z ≡ {τ1 , . . . , τn } ⊆ ±ω , let τ ∈Z τ denote τ1 ∧ V
binary operation defined in Def. 4.13. If Z = ∅, we set τ ∈Z τ = +.
The following lemma establishes a correspondence between the specification
E and the graph GE , in particular it implies that the unique solution σroot of E
for Xroot equals the infimum over all traces through GE from the root node.
Lemma A.3. For α ∈ A, p ∈ Pos(Eα ), let σ(α, p) ∈ ±ω denote V
the unique solution for the expression Eα |p in E, and let τ (α, p) be shorthand for τ ∈T (GE ,hα, pi) τ .
Then, for all α ∈ A, p ∈ Pos(Eα ), we have σ(α, p) = τ (α, p).
Data-Oblivious Stream Productivity
61
Proof. By coinduction, i.e., we show that the relation R ∈ ±ω × ±ω defined by:
R := {hσ(α, p), τ (α, p)i | α ∈ A, p ∈ Pos(Eα )}
is a bisimulation. We proceed by well-founded induction on pairs hα, pi with
ǫ
respect to →.
ǫ
(Base) If v := hα, pi is in normal form with respect to →, then v has precisely
−
+
−
one outgoing edge: v → or v →. In case v →, we have that Eα |p = −E ′ , and
therefore σ(α, p) = −σ(α, p · 1). Furthermore, we have τ (α, p) = −τ (α, p · 1), and
+
clearly hσ(α, p · 1), τ (α, p · 1)i ∈ R. The proof for case v → proceeds analogously.
ǫ
(Step) If v := hα, pi is not a normal form with respect to →, then either v
ǫ
ǫ
ǫ
has one outgoing edge v → v ′ or two outgoing edges v → v ′ and v → v ′′ . In
the first case we have Eα |p = Xα′ , and then v ′ = hα′ , hii, σ(α, p) = σ(α′ , hi)
and τ (α, p) = τ (α′ , hi), and we conclude by an application of the induction
hypothesis for hα′ , hii. In the second case we have that Eα |p = E ′ ∧ E ′′ , and then
v ′ = hα, p · 1i and v ′′ = hα, p · 2i. Thus we get σ(α, p) = σ(α, p · 1) ∧ σ(α, p · 2),
and, by associativity of ∧, τ (α, p) = τ (α, p · 1) ∧ τ (α, p · 2). We conclude by a
double application of the induction hypothesis.
⊓
⊔
The set T (GE , v) consists of infinite, possibly non-productive io-sequences.
Therefore we adopt the definition of the production function [[σ]] : N → N to be
applicable to non-productive io-sequences σ by dropping ∞ from the domain.
Definition A.4. The production function π̃σ of a sequence σ ∈ ±ω is corecursively defined by, for all n ∈ N, π̃σ (n) := π̃(σ, n):
π̃(+σ, n) = 1 + π̃(σ, n) ,
π̃(−σ, 0) = 0 ,
π̃(−σ, n + 1) = π̃(σ, n) .
A property of the operation ∧ given in Def. 4.13 is that the mapping [[ ]] :
±∞ → N → N preserves infimum, see Prop. 4.19.
Clearly, a similar property holds for π̃: for all σ1 , σ2 ∈ ±ω , we have: π̃σ1 ∧σ2 =
π̃σ1 ∧ π̃σ2 . Then, the following corollary directly follows from Lemma A.3.
Corollary A.5. For all n ∈ N we have π̃σroot (n) = inf{π̃τ (n) | τ ∈ T (GE , vroot )}.
Moreover, given n ∈ N, we can always pick a trace τ ∈ T (GE , vroot ) such that
π̃τ (n) = π̃σroot (n).
Using GE we define a ‘two-dimensional diagram’ DE ⊆ V × N× N representing
the traces of E starting from Xroot as step functions. For v ∈ V, x, y ∈ N let
vx,y be shorthand for hv, x, yi. The horizontal axis (x-axis) of T corresponds to
input, and the vertical axis (y-axis) to output; that is, a ‘−’ in the trace yields
a step to the right in the diagram and ‘+’ a step upwards.
62
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Definition A.6. For U ⊆ V × N × N we define GE (U ) ⊆ V × N × N, the one-step
closure of U under GE , as follows:
GE (U ) := U ∪ GE,ǫ (U ) ∪ GE,+ (U ) ∪ GE,− (U )
ǫ
GE,ǫ (U ) := {wx,y | vx,y ∈ U, v → w ∈ L}
+
GE,+ (U ) := {wx,y+1 | vx,y ∈ U, v → w ∈ L}
−
GE,− (U ) := {wx+1,y | vx,y ∈ U, v → w ∈ L}
Furthermore we define
GE∗ (U ) :=
[
GEn (U ) ,
n∈N
the many-step closure of U under GE .
Definition A.7. We define DE ⊆ V × N × N as DE := GE∗ ({vroot 0,0 }).
The diagram DE contains the traces through GE starting from vroot as ‘step
functions’.
Lemma A.8. For all v ∈ V, x, y ∈ N we have vx,y ∈ DE if and only if there
exists a path ρ : vroot →∗ v in GE such that x = #− (trace(ρ)), y = #+ (trace(ρ)).
Proof. Follows immediately from the definition of DE .
⊓
⊔
We proceed by showing that the solution σroot of E for Xroot coincides with the
‘lower bound’ of the traces in DE .
Definition A.9. For U ⊆ V × N × N and GE = hV, L, vroot i as above, we define
low U : N → N, the lower bound of U , as follows:
−
low U (x) := inf{y ∈ N | vx,y ∈ U, ∃w. v → w ∈ L} ,
where inf ∅ = ∞.
Lemma A.10. The lower bound of DE is the solution σroot of E for Xroot :
∀n ∈ N. low DE (n) = π̃σroot (n)
Proof. Let n ∈ N, we show π̃σroot (n) ≥ low DE (n) and π̃σroot (n) ≤ low DE (n).
≥ For π̃σroot (n) = ∞ the proof obligation is trivial, thus assume π̃σroot (n) < ∞.
From Corollary A.5 it follows that there exists a sequence τ ∈ T (GE , vroot )
such that π̃τ (n) = π̃σroot (n). Since π̃τ (n) < ∞ and τ ∈ ±ω , there exists m ∈ N
with #− (τ0...m−1 ) = n < n + 1 = #− (τ0...m ), and #+ (τ0...m−1 ) = π̃τ (n).
−
Then there is a path ρ : vroot →∗ v → v ′ in GE such that trace(ρ) = τ0...m . By
−
Lemma A.8 we have vn,π̃σroot (n) ∈ DE , and together with v → v ′ we conclude
π̃σroot (n) ≥ low DE (n) according to Definition A.9.
Data-Oblivious Stream Productivity
63
≤ Assume low DE (n) < ∞. From Definition A.9 it follows that there exist nodes
−
v, v ′ ∈ V with vn,low DE (n) ∈ DE and v → v ′ . Consequently by Lemma A.8
−
there exists a path ρ : vroot →∗ v → v ′ in GE such that #− (trace(ρ)) = n + 1
and #+ (trace(ρ)) = low DE (n). We choose a τ ∈ ±ω that has trace(ρ) as
a prefix, so that π̃τ (n) = low DE (n). Then, by Corollary A.5, we know that
π̃σroot (n) ≤ low DE (n).
⊓
⊔
We conclude π̃σroot (n) = low DE (n).
Although in principle DE could be used to calculate σroot for every n ∈ N, it
remains to be shown how a rational representation of the solution σroot of E
for Xroot can be obtained. For this purpose we construct a ‘simpler’ diagram
DE0 ⊆ DE that has the same lower bound. We construct DE0 such that there are
only finitely many nodes on each vertical line; thereby we employ the idea that
whenever there are two nodes wx,y , wx,y′ with y < y ′ , then we can omit the
upper occurrence wx,y′ (including all traces rising from wx,y′ ) without having
an influence on the lower bound low DE (x).
Definition A.11. We define functions GE0 , omit : (V × N × N) → (V × N × N):
GE0 (U ) := U ∪ GE,ǫ (U ) ∪ GE,+ (U )
omit (U ) := {vx,y | vx,y ∈ U, ¬∃y ′ < y. vx,y′ ∈ U }
Let DE0 (n) ⊆ V × N × N for all n ∈ N be defined as follows:
DE0 (0) := vclosure({vroot 0,0 })
DE0 (n + 1) := DE0 (n) ∪ DE0 |n+1
DE0 |n+1 := vclosure(step right (n))
−
step right (n) := {wn+1,y | vn,y ∈ DE0 (n), v → w ∈ L} ,
where vclosure(U ) ⊆ V × N × N is defined to be vco m (U ) for the smallest m ∈ N
such that vco m+1 (U ) = vco m (U ) where vco := omit ◦ GE0 , the vertical one-step
closure followed by an application of omit . Furthermore we define
S
DE0 := n∈N DE0 (n) .
Note that DE0 (n) is finite for all n ∈ N. The important step in the construction
is the termination of vclosure( ) which is guaranteed by the following lemma.
Lemma A.12. The computation of vclosure(U ) is terminating for finite sets
U ⊆ V × N × N, that is, there exists m ∈ N such that vco m+1 (U ) = vco m (U ).
Proof. For U ⊆ V × N × N we introduce auxiliary definitions:
X(U ) := {x ∈ N | ∃v ∈ V, y ∈ N. vx,y ∈ U } ⊆ N
V(U, x) := {v ∈ V | ∃y ∈ N. vx,y ∈ U } ⊆ V
Y (U, x, v) := {y ∈ N | vx,y ∈ U } ⊆ N
y(U, x, v) := min(Y (U, x, v)) ∈ N .
64
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Note that Y (W, x, v) = {y(W, x, v)} is a singleton set for every W := omit (U ),
x ∈ X(W ) and v ∈ V(W, x) with U ⊆ V × N × N arbitrary.
Let U ⊆ V × N × N. We define U0 := U and for n = 1, 2, . . . let
Un := vco(Un−1 ) .
Note that Un = vco n (U ). From the definition of GE0 and omit it follows that:
(i) ∀n ∈ N. X(Un ) = X(U ),
(ii) ∀n ∈ N. Un ⊆ GE0 (Un ),
(iii) ∀n ∈ N, x ∈ X(U ). V(Un , x) ⊆ V(Un+1 , x), since
V(Un , x) ⊆ V(GE0 (Un ), x) = V(vco(Un ), x) = V(Un+1 , x)
(iv) ∀n ≥ 1, x ∈ X(U ). |V(Un , x)| ≤ |V|.
By (iii) and (iv) we conclude that there exists n0 ∈ N:
∀n ≥ n0 , x ∈ X(U ). V(Un , x) = V(Un0 , x) ,
(1)
that is, the set of nodes V(Un , x) on each vertical strip x becomes fixed after n0
steps. Furthermore by (ii) and the definition of omit we get
∀n ∈ N, x ∈ X(U ), v ∈ V(Un , x). y(Un+1 , x, v) ≤ y(Un , x, v) ,
(2)
that is, the heights of the nodes are non-increasing. Therefore for each x ∈ X(U )
and node v ∈ V(Un0 , v) there exists mx,v ≥ n0 such that:
∀n ≥ mx,v . y(Un , x, v) = y(Umx,v , x, v)
Let m := max{mx,v | x ∈ X(U ), v ∈ V(Un0 , v)}, then the heights of all nodes
are fixed from m onwards and consequently we have ∀n ≥ m. Un = Um .
⊓
⊔
Lemma A.13. DE and DE0 have the same lower bound, that is, low DE = low DE0 .
Proof.
Now we construct a rational representation of the lower bound of DE0 , which
by Lemma A.13 and A.10 is the solution σroot of E for Xroot . The construction
is based on a ‘repetition search’ on DE0 , that is, we search two vertical strips
x1 and x2 that contain similar constellations of nodes allowing to conclude that
from the vertical position x1 onwards the lower bound is ‘quasi periodic’.
Definition A.14. For x ∈ N, v ∈ V let
V(x) := {v ∈ V | ∃y ∈ N. vx,y ∈ DE0 (x)} ⊆ V
(
y
if there exists y ∈ N with vx,y ∈ DE0 (x)
y(x, v) :=
.
∞ otherwise
Note that y(x, v) is well-defined since there exists at most one such y ∈ N.
Data-Oblivious Stream Productivity
65
Definition A.15. For x ∈ N we define
h(x) := min{y(x, v) | v ∈ V} ,
the height of the lowest node on the vertical strip x.
Definition A.16. For x ∈ N, v ∈ V let
yr (x, v) := y(x, v) − h(x) ,
the height of the node v relative to the lowest node on strip x.
Definition A.17. For x1 < x2 ∈ N with V(x1 ) 6= ∅ and V(x2 ) 6= ∅ let
d(x1 , x2 ) := h(x2 ) − h(x1 ) ,
the height difference between the vertical strips x1 and x2 .
Definition A.18. For x1 < x2 ∈ N, v ∈ V with v ∈ V(x1 ) and v ∈ V(x2 ) let
d(x1 , x2 , v) := y(x2 , v) − y(x1 , v) ,
the height difference of the node v between the vertical strips x1 and x2 , and
dr (x1 , x2 , v) := h(x1 , x2 , v) − d(x1 , x2 ) ,
the relative height difference of the node v between the vertical strips x1 and x2 .
Lemma A.19. For x1 , x2 ∈ N, v ∈ V(x1 ) it holds:
yr (x2 , v) = yr (x1 , v) + dr (x1 , x2 , v) .
Proof. Immediately from the definitions.
⊓
⊔
We proceed with the definition of ‘pseudo repetitions’ hx1 , x2 i ∈ N2 . That is,
we require that the vertical strips x1 and x2 possess a similar constellation of
nodes. A pseudo repetition does not yet guarantee ‘quasi periodicity’ of the lower
bound (having found the cyclic part of the io-sequence), as will be explained after
the definition.
Definition A.20. A pseudo repetition is a pair hx1 , x2 i ∈ N2 such that:
(i) The vertical strips x1 and x2 contain the same set of nodes v ∈ V:
V(x1 ) = V(x2 )
(ii) From strip x1 to x2 all nodes have increased their height by at least d(x1 , x2 ):
∀v ∈ V. dr (x1 , x2 , v) ≥ 0
66
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
For a pseudo repetition hx1 , x2 i we can distinguish between nodes v for which
dr (x1 , x2 , v) = 0 and those with dr (x1 , x2 , v) > 0. If for all nodes v ∈ V(x1 ) it
holds that dr (x1 , x2 , v) = 0 then we have an exact repetition, yielding ‘quasi
periodicity’ of the lower bound. On the other hand assume that there is a node
v ∈ V(x1 ) which increases its relative height and contributes to the lower bound
between the vertical strips x1 and x2 . Then we do not yet have a ‘quasi periodic’
lower bound, because due to the increasing relative height of v, the contribution
v to the lower bound will change. For this reason we will later strengthen the
conditions on pseudo repetitions, yielding ‘repetitions’ which then will guarantee
‘quasi periodicity’, see Definition A.27.
Lemma A.21. There exist pseudo repetitions in DE0 .
Proof. For every x ∈ N we have V(x) ∈ P(V) and P(V) is a finite set. Hence it
follows from the Pigeonhole Principle that there exist indices i1 < i2 < i3 < . . .
with V(i1 ) = V(i2 ) = V(i3 ) = . . ., let V ′ := V(i1 ) = {v1 , . . . , vk } with k = |V(i1 )|.
Let ≤ on Nk be defined as follows:
(a1 , . . . , al ) ≤ (b1 , . . . , bl ) ⇔ a1 ≤ b1 ∧ . . . ∧ al ≤ bl .
For x ∈ {i1 , i2 , . . .} we define the tuple tx := (yr (x, v1 ), . . . , yr (x, vk )) ∈ Nk . It is
a well-known fact that ≤ is an almost full relation on Nk , see [10]. Hence there
exist x1 , x2 ∈ {i1 , i2 , . . .} with tx1 ≤ tx2 . Then dr (x1 , x2 , v) ≥ 0 for all v ∈ V ′
which establishes that hx1 , x2 i is a pseudo repetition.
⊓
⊔
The following definitions and lemmas are auxiliary for the purpose of proving
Lemma A.26.
Definition A.22. For U, W ⊆ V × N × N and dx , dy ∈ N we define
U ≤dx ,dy W ⇐⇒ ∀vx,y ∈ W. ∃y ′ < y − d. vx−dx ,y′ ∈ U .
Lemma A.23. For x1 < x2 ∈ N and dy ∈ N we have:
DE0 |x1 ≤x2 −x1 ,dy DE0 |x2 ⇐⇒ ∀v ∈ V(x1 ). y(x2 , v) ≥ dy + y(x1 , v)
Proof. Follows from the definitions.
⊓
⊔
Lemma A.24. If U1 , U2 ⊆ V × N × N, dx , dy ∈ N with U1 ≤dx ,dy U2 then:
(i) vco(U1 ) ≤dx ,dy vco(U2 ), and
(ii) vclosure(U1 ) ≤dx ,dy vclosure(U2 ).
Proof. From U1 ≤dx ,dy U2 it follows that GE0 (U1 ) ≤dx ,dy GE0 (U2 ) holds, and then
we get omit ◦ GE0 (U1 ) ≤dx ,dy omit ◦ GE0 (U2 ) by definition of omit .
By induction we get ∀n ∈ N. vco n (U1 ) ≤dx ,dy vco n (U2 ). Furthermore there
exist m1 , m2 ∈ N such that vclosure(Ui ) = vco mi (Ui ) for i ∈ {1, 2} by definition.
′
From ∀m′i ≥ mi . vco mi (Ui ) = vco mi (Ui ) it follows that for i ∈ {1, 2} we have
max(m1 ,m2 )
⊔
vclosure(Ui ) = vco
(Ui ). Hence vclosure(U1 ) ≤dx ,dy vclosure(U2 ). ⊓
Data-Oblivious Stream Productivity
67
Lemma A.25. If x1 < x2 ∈ N with V(x1 ) = V(x2 ), then V(x1 + 1) = V(x2 + 1).
If moreover DE0 |x1 ≤x2 −x1 ,dy DE0 |x2 for dy ∈ N, then DE0 |x1 +1 ≤x2 −x1 ,dy DE0 |x2 +1 .
Proof. From V(x1 ) = V(x2 ) and the definitions of DE0 (x1 + 1) and DE0 (x2 + 1) it
follows that V(x1 + 1) = V(x2 + 1).
Assume DE0 |x1 ≤x2 −x1 ,dy DE0 |x2 and consider Definition A.11. Then it follows
that step right (x1 ) ≤x2 −x1 ,dy step right (x2 ). We conclude with an application
⊓
⊔
of Lemma A.24 yielding DE0 |x1 +1 ≤x2 −x1 ,dy DE0 |x2 +1 .
Lemma A.26. Let hx1 , x2 i be a pseudo repetition. Then the pair hx1 + m, x2 +
mi is a pseudo repetition for all m ∈ N.
Proof. It suffices to show that hx1 + 1, x2 + 1i is a pseudo repetition, for m > 1
the claim follows by induction.
We get V(x1 + 1) = V(x2 + 1) by Lemma A.25. Furthermore by the definition
of pseudo repetition we have ∀v ∈ V(x1 ). dr (x1 , x2 , v) ≥ 0, which is equivalent
to DE0 |x1 ≤x2 −x1 ,d(x1 ,x2 ) DE0 |x2 since V(x1 ) = V(x2 ). Emloying Lemma A.25 we
get DE0 |x1 +1 ≤x2 −x1 ,d(x1 ,x2 ) DE0 |x2 +1 ,
that is ∀v ∈ V(x1 + 1). dr (x1 + 1, x2 + 1, v) ≥ 0. Hence hx1 + 1, x2 + 1i is a
pseudo repetition.
⊓
⊔
We proceed with the definition of ‘repetitions’. Then we show that repetitions
exist and that these indeed guarantee quasi periodicity of the lower bound.
Definition A.27. A repetition is a pseudo repetition hx1 , x2 i if:
(i) hx1 , x2 i is a pseudo repetition
(ii) Let I(x1 , x2 ) := {vx,y | vx,y ∈ T0 , x = x1 , i(x1 , x2 , p) = 0}. We require that
only nodes from I(x1 , x2 ) contribute to the lower bound between x1 and x2
and these nodes reconstruct themselfs on the strip x2 :
∀x ∈ N, x1 ≤ x ≤ x2 . low GE (I) (x) = low T0 (x)
(3)
∀vx,y ∈ I(x1 , x2 ). vx2 ,y+d(x1 ,x2 ) ∈ GE (I)
(4)
Both conditions can be effectively checked since GE (I)|≤n can be computed
for every n ∈ N. Having such a repetition is is easy to see that the lower bounds
behaves ‘quasi periodic’ from this point onwards, that is the io-sequence is indeed rational. This is because the the nodes with i(x1 , x2 , v) = 0 consitute a
exactly repeating pattern. Furthermore the nodes with increasing relative height
i(x1 , x2 , v) > 0 do not contribute to the lower bound between x1 and x2 , and
since their height cannot decrease they will also not contribute in the future.
It remains to show that we indeed always encounter a repetition hx1 , x2 i. We
know that there exists a pseudo repetition hx1 , x2 i. Then we have for all m ∈ N
we have hx1 + m · (x2 − x1 ), x2 + m · (x2 − x1 )i is a pseudo repetition. For all
m ∈ N and v ∈ V we have i(x1 + m·(x2 − x1 ), x2 + m·(x2 − x1 ), v) ≥ 0, therefore
it follows that there are some nodes for which the relative height increase will
eventually stay at 0 while other nodes continue increasing their relative height.
The latter will from some point on no longer contribute to the lower bound and
the pattern. Hence there exists m0 for which hx1 +m0 ·(x2 −x1 ), x2 +m0 ·(x2 −x1 )i
is a repetition.
68
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Appendix B
Termination of the Function Layer
Translation (Lemma 5.10)
It suffices to show that, for every stream function symbol f and argument position
i of f, the infinite io-sequence specification ET defined in Def. 5.4 that has [f]i as
unique solution for the recursion variable Xhf, i, 0i can be transformed in finitely
many steps into an io-sequence specification E ′ with the same solution for the
recursion variable Xhf, i, 0i and the property that only finitely many recursion
variables are reachable in E ′ from Xhf, i, 0i . (This is because from such an iosequence specification E ′ eventually a finite io-sequence specification E0′ with [f]i
as unique solution for Xhf, i, 0i can be extracted.) An algorithm for this purpose
is obtained as follows. On the input of ET , set E := ET and repeat the following
step on E as long as it is applicable:
(RPC) Detect and remove a reachable non-consuming pseudo-cycle from the iosequence specification E : Suppose that, for a function symbol h, for j, k, l ∈
N, and for a recursion variable Xhh, j, ki that is reachable from Xhf, i, 0i , we
w
have Xhh, j, ki −→ Xhh, j, li (from the recursion variable Xhh, j, ki the variable
Xhh, j, li is reachable via a path in the specification on which the finite iosequence w is encountered as the word formed by consecutive labels), where
l < k and w only contains symbols ‘+’. Then modify E by setting Xhh, j, ki =
+Xhh, j, ki .
It is not difficult to show that a step (RPC) preserves weakly guardedness and
the unique solution of E, and that, on input ET , the algorithm terminates in
finitely many steps, producing an io-sequence specification E ′ with [f]i as the
solution for Xhf, i, 0i and the property that only finitely many recursion variables
are reachable in E ′ from Xhf, i, 0i .
Data-Oblivious Stream Productivity
Appendix C
69
Soundness of the Function Layer Translation
Throughout, let T = hΣ, Ri be a fixed stream specification. First we introduce
a few axillary lemmas and definitions.
Definition C.1. For all stream terms t ∈ Ter (LΣM) we define by
#• (t) := sup{n ∈ N | t = •k (t′ )}
the number of leading stream pebbles of t.
Definition C.2. Let h ∈ Σsfun be a a stream function symbol and j, k ∈ N. We
define hj,k := h(. . . •∞ . . . , •k (σ), . . . •∞ . . .) ∈ Ter (LΣM) where the j-th argument
position is •k (σ) while all other arguments are •∞ .
Definition C.3. For terms
s ≡ g(•n1 (σ), . . . , •nk (σ)) ∈ Ter (LΣM)
t ≡ h(•m1 (σ), . . . , •mk (σ)) ∈ Ter (LΣM)
with n1 , . . . , nk , m1 , . . . , mk ∈ N, we write s ⊳ t if ni ≤ mi for all i.
Lemma C.4. If s, t ∈ Ter (LΣM) as in Def. C.3 with s ⊳ t then do T (s) ≤ do T (t).
Proof. Immediately from the definition.
⊓
⊔
We proof the soundness of the function layer translation, that is:
Lemma C.5. Let T be a stream definition, and let f ∈ Σsfun .
(i) If T is flat, then it holds: [[[f]]] = do T (f).
Hence, do T (f) is a periodically increasing function.
(ii) If T is friendly nesting, then it holds: [[[f]]] ≤ do T (f).
Proof. We start with (i). Let ET be the weakly guarded io-sequence specification
(over variables X ) obtained from the translation. Let V : X → (N → N) be the
unique solution of ET , which exists by Lemma 5.3. Then by definition we have
[[[f]]] = min(V(Xhf, 1, 0i ), . . . , V(Xhf, n, 0i )).
Assume f is not weakly guarded. There exists g ∈ Σsfun and γ : f ;∗ g ;+ g
without production; let Σγ be the set of symbols occurring in γ and ρ1 ,. . . ,ρn ∈ R
the rules applied in γ (all having zero production). We define Tγ as the set of
terms t ∈ Ter (LΣM) with r oot(t) ∈ Σγ . Accoring to Lemma 3.9 we can contruct
a data-exchange function G such that for every term t ∈ Tγ either G(t) is not
a redex or a redex with respect to a ρi . Then every reduct of a term t ∈ Tγ in
AT (G) as again in Tγ and has therefore no production. Hence do T (f) = 0 and
by definition [[[f]]] = 0 since Xhf, i, 0i = Xhǫ, −i for all 1 ≤ i ≤ ars (f).
The remaining proof obligation is [[[f]]] = do T (f) for weakly guarded f. We
start with [[[f]]] ≥ do T (f). Let p1 , . . . , pars (f) ∈ N and define o := [[[f]]](p1 , . . . , pars (f) ).
We have V(Xhf, i, 0i )(pi ) = o for some 1 ≤ i ≤ ars (f). Note that
do T (f)(p1 , . . . , pi , . . . , pars (f) ) ≤ do T (f)(. . . ∞ . . . , pi , . . . ∞ . . .) = do T (fi,pi ) .
70
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Hence it suffices to show
do T (gi,q+p ) ≤ V(Xhg, i, qi )(p)
for all g ∈ Σsfun , argument positions 1 ≤ i ≤ ars (g) ∈ N and q, p ∈ N. For the
case V(Xhg, i, qi )(p) = ∞ there is nothing to be shown, hence let V(Xhg, i, qi )(p) <
∞. Employing well-founded induction with respect to the ordering > over tuples
hg, i, q, pi with V(Xhg, i, qi )(p) < ∞
we construct a data-exchange function G together with an outermost-fair rewrite
sequence in AT (G) starting from gi,p+q as witness of do T (gi,q+p ) ≤ V(Xhg, i, qi )(p).
The ordering > is defined as follows:
hg, i, q, pi > hg′ , i′ , q ′ , p′ i ⇐⇒ g is weakly-guarded
∧ (o > o′ ∨ (o = o′ ∧ g ;+ g′ ))
where o := V(Xhg, i, qi )(p), o′ := V(Xhg′ , i′ , q′ i )(p′ )
Let G be an arbitrary data-exchange function. The induction basis consists
of the tuples hg, i, q, pi where
– g is not weakly guarded, or
– V(Xhg, i, qi )(p) = 0 and g is a normal form with respect to ;+ .
The not weakly guarded symbols have been discussed above, V
therefore let g be
a normal form with respect to ;+ . By definition Xhg, i, qi = ρ∈R g Xhg, i, q, ρi ,
and hence V(Xhg, i, q, ρi )(p) = 0 for some ρ ∈ Rsfun .
Then Xhg, i, q, ρi is either of shape
.
(shape 1) −in(ρ,i)−q + . . .
or
.
(shape 2) −(in(ρ,i)−q)+1 + . . .
To see this consider the case distinction in the definition of Xhg, i, q, ρi :
.
Xhg, i, q, ρi = −in(ρ,i)−q +out(ρ)
V
.
j∈φ−1 (i) Xhh, j, (q−in(ρ,i))+su(ρ,j)i
.
q−in(ρ,i)
+
Xhǫ, −+i
Xhǫ, +i
for (a)
(C1)
for (b) if j = i
(C2)
for (b) if j 6= i
(C3)
We have (shape 1) in case of:
– (C1) since g being a ; normal form and therefore out (ρ) > 0,
. in(ρ, i)) > 0, and
– (C2) if out(ρ) + (q −
– (C3)
. in(ρ, i)) = 0. Let t ≡ g
The (shape 2) occurs in case (C2) if out(ρ) + (q −
i,q+p
′
and G = G(t, ρ), then:
Data-Oblivious Stream Productivity
71
– Assume Xhg, i, q, ρi is of (shape 1).
. > p and hence in(ρ, i) > q+p.
From V(Xhg, i, q, ρi )(p) = 0 it follows in(ρ, i) −q
′
G (t) cannot be a ρ-redex since ρ has at argument position i a consumption
greater than q + p. Hence gi,q+p is a normal form in AT (G ′ ) by Lem. 3.9 and
it follows that do T (gi,q+p ) = 0.
– Assume Xhg, i, q, ρi is of (shape 2).
. in(ρ, i)) = 0, hence out(ρ) = 0 and q ≤ in(ρ, i).
Then we have out(ρ) + (q −
.
From V(Xhg, i, q, ρi )(p) = 0 it follows that (in(ρ, i) −q)+1
= in(ρ, i)−q+1 > p
and therefore in(ρ, i) ≥ q + p. Then gi,q+p is a normal form if in(ρ, i) > q + p
or gi,q+p → σ if in(ρ, i) = q + p in AT (G ′ ); in both cases do T (gi,q+p ) = 0.
This concludes the base case of the induction.
V
Now we consider the induction step. By definition Xhg, i, qi = ρ∈R g Xhg, i, q, ρi ,
and hence V(Xhg, i, qi )(p) = V(Xhg, i, q, ρi )(p) for some ρ ∈ Rsfun . Let t ≡ gi,q+p
and G ′ = G(t, ρ). Again we consider the case distinction (C1),(C2) and (C3) in
the definition of Xhg, i, q, ρi .
. q > p then in all cases V(X
If in(ρ, i) −
hg, i, q, ρi )(p) = 0 and gi,q+p is a normal
. ≤ p.
form in AT (G ′ ), consequently do T (gi,q+p ) = 0. Therefore assume in(ρ, i) −q
V
.
. in(ρ,i))+su(ρ,j)i
(C1) Xhg, i, q, ρi = −in(ρ,i)−q +out(ρ) j∈φ−1 (i) Xhh, j, (q−
There exists j ∈ φ−1 (i) with V(Xhg, i, q, ρi )(p) = eV (p) where
.
e := −in(ρ,i)−q +out(ρ) Xhh, j, q′ i
. in(ρ, i)) + su(ρ, j) .
q ′ := (q −
. q ≤ p, let p′ := p − (in(ρ, i) −
. q). We get
Since in(ρ, i) −
V(Xhg, i, q, ρi )(p) = out(ρ) + V(Xhh, j, q′ i )(p′ ) .
(5)
G ′ (gi,q+p ) is a ρ-redex by Lem. 3.9, hence
gi,q+p →AT (G ′ ) •out (ρ) (h(•n1 (σ), . . . , •nars (h) (σ)))
where by definition of ρ:
nj = su(ρ, j) + (q + p − in(ρ, i)) = q ′ + p′ .
Therefore
h(•n1 (σ), . . . , •nars (h) (σ)) ⊳ hj,q′ +p′ .
We have V(Xhg, i, q, ρi )(p) > V(Xhh, j, q′ , ρi )(p′ ) or g ; h by (5), therefore we
get do T (hj,q′ +p′ ) ≤ V(Xhh, j, q′ i )(p′ ) by induction hypothesis and
do T (gi,q+p ) ≤ out (ρ) + do T (hj,q′ +p′ )
≤ out(ρ) + V(Xhh, j, q′ i )(p′ ) = V(Xhg, i, q, ρi )(p) = V(Xhg, i, qi )(p)
which proves the claim.
72
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
.
.
(C2) Xhg, i, q, ρi = −in(ρ,i)−q +out(ρ) +q−in(ρ,i) Xhǫ, −+i
Then
. in(ρ, i)) + V(X
.
V(Xhg, i, q, ρi )(p) = out(ρ) + (q −
hǫ, −+i )(p − (in(ρ, i) − q))
. in(ρ, i)) + (p − (in(ρ, i) −
. q))
= out(ρ) + (q −
= out(ρ) + q + p − in(ρ, i)
Furthermore in case (C2) ρ is collapsing with i = j, therefore
gi,q+p →AT (G ′ ) •out (ρ)+(q+p−in (ρ,i)) (σ) .
Hence
do T (gi,q+p ) ≤ out(ρ) + q + p − in(ρ, i) = V(Xhg, i, q, ρi )(p) .
.
(C3) Xhg, i, q, ρi = −in(ρ,i)−q +out(ρ) Xhǫ, +i
Then V(Xhg, i, q, ρi )(p) = ∞ = do T (gi,q+p ) since ρ is collapsing with i 6= j.
This concludes the direction [[[f]]] ≥ do T (f).
The remaining proof obligation is [[[f]]] ≤ do T (f). Let p1 , . . . , pars (f) ∈ N
and define o := do T (f)(p1 , . . . , pars (f) ) = do T (f(•p1 , . . . , •pars (f) )). Assume o <
∞, otherwise there is nothing to be shown. Then there exists a data-exchange
function G and a finite reduction sequence
f(•p1 (σ), . . . , •pars (f) (σ)) →nAT (G ′ ) •o (t)
such that do T (t) = 0 and in particular
– t is a normal form in AT (G ′ ), or
– g is not weakly guarded.
W.l.o.g. choose G and the reduction sequence such that n is minimal. We apply
induction on the length of the reduction n to show
– There exists 1 ≤ i ≤ ars (f) such that
do T (f(•p1 (σ), . . . , •pars (f) (σ))) = do T (f(•∞ , . . . , •pi , . . . , •∞ )) ,
– and [[[f]]] ≤ do T (f(•p1 (σ), . . . , •pars (f) (σ))).
Data-Oblivious Stream Productivity
Appendix D
D.1
73
Soundness of the Stream Layer Translation
(Lemma 5.17 and Lemma 5.18)
Proof Sketch of Lemma 5.17
We only sketch a proof of statement (i) of Lem. 5.17 because item (ii) of this
lemma can be demonstrated analogously.
Proof (of Lem. 5.17, (i)). Let T be an SCS, and let F = {γ f }f∈Σsfun be a family
of gates such that, for all f ∈ Σsfun , [[γ f ]] = do T (f). We give the ideas for the
proofs of the inequations “≤” and “≥” that make part of the equation
do T (M0 ) = [[[M0 ]F ]] .
(6)
We assume, without loss of generality, that every f ∈ Σsfun has precisely one
occurrence in the SCS-layer. Note that, by replacing for every f ∈ Σsfun the
individual occurrences of f in the SCS-layer of T by individual occurrences of
the symbols f, f ′ , f ′′ , . . . , respectively, and introducing defining rules for f ′ , f ′′ ,
. . . in the SFS-part analogous to the defining rules for f that are already present
there, an SCS T ′ is obtained that satisfies the restriction, and for which it is
′
easy to establish that it holds: do T (M0 ) = do T ′ (M0 ), [M0 ]F = [M0 ]F , and hence
′
′
′
[[[M0 ]F ]] = [[[M0 ]F ]], where F ′ = {γ f }f∈Σsfun
is the family of gates (Σsfun
is the
′
stream function signature of T ) by adding to F , for all symbols f ∈ Σsfun , the
′
,
corresponding copies γ f ′ , γ f ′′ , . . . of γ f . Note that we also have, for all f ′ ∈ Σsfun
′
[[γ f ′ ]] = do T (f ).
“≤”: We only have to consider the case in which for k := [[[M0 ]F ]] it holds k < ∞.
Hence we assume that k < ∞. Then there exists a finite rewrite sequence
ρ : [M0 ]F ։P •k (p′ ) .
Then with p := [M0 ]F it follows that p′ has the same ‘gate structure’ as
p, but possibly the io-sequences in the gates have changed, and there are
likely to appear many pebbles queuing inside p′ . By the definition of k it
follows that [[p′ ]] = 0. Without loss of generality, we assume that during ρ all
extractable pebbles have been extracted from ‘inside’ the gates in p′ , that
is, there is no occurrence of a redex in p′ with respect to the pebbleflow rule
min(•(p1 ), •(p2 )) → •(min(p1 , p2 )). Note that the only occurrences of the
symbol min in terms of ρ are part of gates. (If this would not be the case for
ρ, then we would extend ρ by finitely many →P -steps to a rewrite sequence
ρ′ which has this property, and base the further argumentation on ρ′ .)
Due to the assumption that function symbols in Σsfun only occur once in the
SCS-layer, to every f ∈ Σsfun there corresponds in p (via the net translation
in Def. 5.16) a unique gate γ f = [f]. We now monitor, for every gate γ f in
p, the consumption/production behaviour during ρ. Let f ∈ Σsfun . Suppose
that during ρ, γ f consumes n1 , . . . , nars (f) pebbles from its input components,
respectively, and that in total it produces n pebbles. By the assumption
74
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
that no extractable pebbles are stuck inside γ f at the end of ρ, in particular, [[γ f ]](n1 , . . . , nars (f) ) = n follows. Moreover, by the assumption that the
data-oblivious lower bound do T (f) of a symbol f ∈ Σsfun coincides with the
production function of the gate γ f , we get do T (f)(n1 , . . . , nars (f) ) = n. As a
consequence, there is a rewrite sequence:
ρ̃ f : f(•n1 : σ1 , . . . , •nars (f) : σars (f) , •, . . . , •) ։AT,G •n : t
| {z }
ard (f)
in an ARS AT,G such that ρ̃ f can be extended to an outermost-fair rewrite
sequence that does not produce further pebbles. Then from ρ̃ f we extract
a rational gate γ̃ f with the properties: (a) γ̃ f precisely describes the consumption/production behaviour that happens during ρ̃ f , (b) thereafter γ̃ f
continues like the gate corresponding to f in p′ . We gather the new gates
in the new family of gates: F̃ := {γ̃ f }f∈Σsfun . Note that, for all f ∈ Σsfun ,
[[γ f ]] ≤ [[γ̃ f ]] holds. Nevertheless, it can be shown that there is actually also a
rewrite sequence:
ρ̃ : [M0 ]F̃ ։P •k (p̃′ ) .
such that [[p̃′ ]] = 0. Using this pebbleflow rewrite sequence ρ̃ as well as the
rewrite sequences ρ̃f on the abstracted versions of T , it is possible to obtain
a rewrite sequence in an ARS AT,G of the form:
k
ρ̄ : T T (M0 ) ։
։mω
AT,G • : s ,
where m is the length of the reduction sequence ρ̃, that can be extended
in AT,G to a rewrite sequence ρ̄¯ which consists of outermost-fair sequences
of rewrite sequences of length ω in which respectively, for some g ∈ Σsfun ,
g-redexes are reduced uniformously in all unfoldings, and such that furthermore ρ̄¯ does not contain terms with a prefix of more than k pebbles. The
existence of the rewrite sequence ρ̄¯ on M0 in an ARS AT,G demonstrates that
do T (M0 ) ≤ k.
“≥”: This direction of the proof is a consequence of three statements that are
mentioned below. First, it holds that:
For all p ∈ P :
[[T (p)]] = [[p]] ,
(7)
where T (p) denotes the infinite unfolding of p into a possibly infinite ‘pebbleflow tree’. Second:
For all stream terms s in T :
ΠT (s) = ΠT (T T (s)) ,
(8)
where by s ∈ Ter (Σ) we mean the result of infinitely unfolding all stream
constants in s. Finally, third, it holds that:
For all stream terms s in T :
do T (T T (s)) ≥ [[T ([s]F )]] .
Putting (7), (8), and (9) together, we obtain:
do T (M0 ) = do T (T T (M0 )) ≥ [[T ([M0 ]F )]] = [[[M0 ]F ]] ,
which shows “≥”.
(9)
Data-Oblivious Stream Productivity
75
This concludes our sketch of the proof of Lem. 5.17 (i).
D.2
Proof of Lemma 5.18
Let T be a stream specification, and let F = {γ f }f∈Σsfun be a family of gates
such that, for all f ∈ Σsfun , the arity of γ f equals the stream arity of f. Suppose
that one of the following statements holds:
(a)
(b)
(c)
(d)
[[γ f ]] ≤ do T (f)
[[γ f ]] ≥ do T (f)
do T (f) ≤ [[γ f ]]
do T (f) ≥ [[γ f ]]
for
for
for
for
all
all
all
all
f
f
f
f
∈ Σsfun ;
∈ Σsfun ;
∈ Σsfun ;
∈ Σsfun .
In this section we show that then, for all M ∈ Σscon , the corresponding one of
the following statements holds:
(a)
(b)
(c)
(d)
[[[M]F ]] ≤ do T (M) ;
[[[M]F ]] ≥ do T (M) ;
do T (M) ≤ [[[M]F ]] ;
do T (M) ≥ [[[M]F ]] .
Multiple Numbered Contexts
Definition D.1. Let T = hΣ, Ri be a stream specification, and k, l ∈ N. A
(multiple numbered) stream context C[] over stream holes 21 , . . . , 2k , and data
holes 2
· 1, . . . , 2
· ℓ is a (stream) term
C[] ∈ Ter ∞ (Σ ⊎ {21 , . . . , 2k } ⊎ {2
· 1, . . . , 2
· ℓ }, X)S
where the 2i :: S and 2
· i :: D are distinct constant symbols not occurring in Σ.
We denote by C[t1 , . . . , tk ; u1 , . . . , uℓ ] the result of replacing the occurrences
of 2i by si and 2
· j by tj in C[]T , respectively.
Remark D.2. Each of the hole symbols 21 , . . . , 2k , and 2
· 1, . . . , 2
· ℓ may have
multiple occurrences in a stream context C[] over 21 , . . . , 2k and 2
· 1, . . . , 2
· ℓ,
and this includes the case that it does not occur in C[] at all.
Example D.3. Let T be a stream specification with data constants b and c, a
unary stream function symbol f, and a binary stream function symbol g. Then
the expression b : f(c : g(2
· 1 : 21 , g(22 , g(21 , 23 )))) is a stream context over 21 ,
22 , 23 and 2
· 1 . Also f(b : 23 ) is a stream context over 21 , 22 , 23 and 2
· 1 . But
b:21 :σ is not a stream context over 21 , because the symbol 21 occurs at a ‘data
position’ (this is excluded by many-sortedness, : :: D → S → S and 21 :: S ).
We extend the d-o production function of stream function symbols f ∈ Σsfun ,
Def. 3.5, to multiple numbered contexts.
76
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
Definition D.4. Let T = hΣ, Ri be a stream specification, and C[] a multiple numbered stream context over stream holes 21 , . . . , 2k , and data holes
2
· 1, . . . , 2
· ℓ . The d-o production range do T (C[]) : Nk → N of the context C[] is:
do T (C[])(n1 , . . . , nk ) := do T ( LC[]M((•n1 : σ), . . . , (•nk : σ)) ) ,
m times
z }| {
where •m : σ := • : . . . : • : σ.
The d-o lower and upper bounds on the production of g are defined by
do T (C[]) := inf(do T (C[])) and do T (C[]) := sup(do T (C[])), respectively.
Note that it is indeed an extension of Def. 3.5, for every g ∈ Σsfun we have
do T (g) = do T (g(21 , . . . , 2ars (g) , 2
· 1, . . . , 2
· ard (g) )) .
The following lemma states that data-oblivious lower and upper bounds are
reachable.
Lemma D.5. Let T = hΣ, Ri be a stream specification, and s ∈ Ter ∞ (Σ)S .
Colouring Terms, Tracing Data Flow For the purpose of tracing the flow of
data-elements during rewrite sequences, we introduce auxiliary ‘tracker symbols’
and partition terms into ‘coloured’ multiple numbered contexts. Usual rewriting
will then be allowed only within contexts. Rewrite steps crossing the border
between contexts are prohibited. In order to circumvent the loss of possible
reduction steps, we introduce rules for exchanging pebbles between neighboring
contexts (recolouring of the pebbles).
We introduce auxiliary, unary function symbols ⋄c and ⋄·c in the pebbleflow
rewrite system as well as in the stream specification T . The symbols ⋄c mark the
top of a context of colour c and at the same time terminate the context above
it. The symbols ⋄·c only close, but do not open a context. We call them tracker
symbols in the sequel, since they are also used to track pebble/data exchange
between neighboring contexts. Therefore we enrich the pebbleflow rewrite system
with the following rules
⋄c (•(p)) → •(⋄c (p))
⋄·c (•(p)) → •(⋄·c (p))
(10)
⋄·p (t : σ) → t : ⋄·c (σ)
(11)
and the stream specification T with rules
⋄c (t : σ) → t : ⋄c (σ)
for every function symbol ⋄c and ⋄·c . We define ⋄c , ⋄·c :: stream → stream, that is,
the tracker symbols cannot be placed at data positions.
Definition D.6. For terms s, t we say that t is an enrichment of s with tracker
symbols, denoted s ∠ t, if s can be obtained from t by removing tracker symbols,
that is, t →∗ s via rules ⋄c (u) → u and ⋄·c (u) → u.
Data-Oblivious Stream Productivity
77
Lemma D.7. Every pebbleflow reduction can be lifted to enrichted terms:
∀s, s′ ∈ P, s ։P s′ , s ∠ t. ∃t′ . t ։P⋄ t′ ∧ s′ ∠ t′
and stream specifications T
Every pebbleflow rewrite sequence and every stream specification rewrite
sequence can be lifted to enriched terms, i.e. ∀s′ և s ∠ t. ∃t′ . t ;∗ t′ ∧ s′ ∠ t′
where ; = → ∪ →(10) (or ; = → ∪ →(11) ). Sometimes the tracker symbols are
in the way of rule application, but, because of the special structure of pebbleflow
(resp. SCS) rules, it is always possible to move them away using rules (10)
(resp. (11)).
Now we sketch the proof for the inequalities “≤” (i) and “≥” (ii).
“≤”: We need to show [[[M ]F ]] ≤ ΠT (M ). For the proof of (i) we map a given finite pebbleflow rewrite sequence on [M ]F producing n pebbles to a T rewrite
sequence on the infinite unfolding of M that produces at least n data elements. (For (ii) the proof proceeds the other way arround.)
Thereby not every rewrite step will be mapped, but we trace ‘essential’ steps
and map them to a many-step (possibly infinite) reduction on ‘the other side’.
For this purpose we partition the pebbleflow term [M ]F as well as infinite
unfolding of M into coloured contexts (using the same colours for the net
and the unfolding of M ). The partitioning is arranged in such a way that
contexts of the same colour have the same production function. In the sequel
we will refer to this property as the context production property (CPP).
Usual rewriting will then be allowed only within contexts. Rewrite steps
crossing the border between contexts are prohibited. In order to circumvent
the loss of possible reduction steps, we introduce additional rules for exchanging pebbles between neighboring contexts (recolouring of the pebbles).
The construction of the mapped reduction sequence will be driven by the goal
to uphold CPP. It can be shown pebbleflow rewriting and SCS rewriting does
not change the production function of a context. Hence in order to maintain
CPP it is not necessary to map internal steps ocurring within contexts. The
‘essential’ steps are the exchange steps between the contexts.
We use CcP to refer to the context of colour c in the pebbleflow net [M ]F .
Likewise we use CcT for the context of colour c in the infinite unfolding of
the stream constant M . Although there may be multiple, typically infinitely
many, occurrences of a context with colour c on the term level, contexts with
the same colour are syntactically equal. Initially all contexts are of one of
the following (simple) shapes:
Clearly [[ ]]CcP = [[ ]]CcT for case 1. An important step in the proof is to show
that the translation of functions symbols into rational gates preserves the
quantitative production behaviour, yielding [[ ]]CcP = [[ ]]CcT for case 2.
The proof now continues as follows. Every pebble exchange step in the pebbleflow net is mapped to an infinite rewrite sequence on the corresponding
coloured SCS term. For the definition of these corresponding rewrite steps
we apply the assumption [[ ]]CcP = [[ ]]CcT guaranteeing that a data element
78
Jörg Endrullis, Clemens Grabmayer, and Dimitri Hendriks
case
CcP
CcT
1
•(21 )
d : 21
2 min(σf,1 (21 ), . . . , σf,r (2r )) f(21 , . . . , 2r )
Fig. 15: Initial context configurations
can also be extracted from the context on SCS term level, and that after the
step we have again [[ ]]CcP = [[ ]]CcT for all colours c. In this way, we define a
rewrite sequence ρT in T by induction on the number of rewrite steps of the
given pebbleflow reduction ρ such that ρT and ρ produce the same amount
of data elements and pebbles, respectively.
“≥”: Here we need to show [[[M ]F ]] ≥ ΠT (M ). The proof of this inequality
proceeds similarly to that of “≤”. Let ρT in T be a reduction sequence
in T . In order to map ρT back to a pebbleflow rewrite sequence ρ it is
crucial that the contexts of one colour are changed synchronously and in
the same way. It cannot be assumed that ρT posesses this property. For this
purpose ρT is transformed into a sequence ρ′T of complete developments.
This is always possible since complete developments in orthogonal TRSs are
a cofinal rewrite strategy. For mapping ρ′T to a rewrite sequence ρ with the
same production on the pebbleflow net, we can then proceed by induction
on the length of ρ′T , in converse direction to the argument for “≤”, above.
| 6cs.PL
|
arXiv:1706.05749v1 [stat.ML] 19 Jun 2017
Dex: Incremental Learning for Complex
Environments in Deep Reinforcement Learning
Nick Erickson, Qi Zhao
Department of Science and Engineering
University of Minnesota, Twin Cities
Minneapolis, MN 55455
[email protected], [email protected]
Abstract
This paper introduces Dex, a reinforcement learning environment toolkit specialized for training and evaluation of continual learning methods as well as general
reinforcement learning problems. We also present the novel continual learning
method of incremental learning, where a challenging environment is solved using
optimal weight initialization learned from first solving a similar easier environment.
We show that incremental learning can produce vastly superior results than standard
methods by providing a strong baseline method across ten Dex environments. We
finally develop a saliency method for qualitative analysis of reinforcement learning,
which shows the impact incremental learning has on network attention.
1
Introduction
Complex environments such as Go, Starcraft, and many modern video-games present profound
challenges in deep reinforcement learning that have yet to be solved. They often require long, precise
sequences of actions and domain knowledge in order to obtain reward, and have yet to be learned from
random weight initialization. Solutions to these problems would mark a significant breakthrough on
the path to artificial general intelligence.
Recent works in reinforcement learning have shown that environments such as Atari games [2]
can be learned from pixel input to superhuman expertise [9]. The agents start with randomly
initialized weights, and learn largely from trial and error, relying on a reward signal to indicate
performance. Despite these successes, complex games, including those where rewards are sparse such
as Montezuma’s Revenge, have been notoriously difficult to learn. While methods such as intrinsic
motivation [3] have been used to partially overcome these challenges, we suspect this becomes
intractable as complexity increases. Additionally, as environments become more complex, they will
become more expensive to simulate. This poses a significant problem, since many Atari games
already require upwards of 100 million steps using state-of-the-art algorithms, representing days of
training on a single machine.
Thus, it appears likely that complex environments will become too costly to learn from randomly
initialized weights, due both to the increased simulation cost as well as the inherent difficulty of the
task. Therefore, some form of prior information must be given to the agent. This can be seen with
AlphaGo [18], where the agent never learned to play the game without first using supervised learning
on human games. While supervised learning certainly has been shown to aid reinforcement learning,
it is very costly to obtain sufficient samples and requires the environment to be a task humans can
play with reasonable skill, and is therefore impractical for a wide variety of important reinforcement
learning problems.
In this paper we introduce Dex, the first continual reinforcement learning toolkit for training and
evaluating continual learning methods. We present and demonstrate a novel continual learning method
we call incremental learning to solve complex environments. In incremental learning, environments
are framed as a task to be learned by an agent. This task can be split into a series of subtasks that are
solved simultaneously.
Similar to how natural language processing and object detection are subtasks of neural image
caption generation [23], reinforcement learning environments also have subtasks relevant to a given
environment. These subtasks often include player detection, player control, obstacle detection,
enemy detection, and player-object interaction, to name a few. These subtasks are common to
many environments, but they are often sufficiently different in function and representation that
reinforcement learning algorithms fail to generalize them across environments, such as in Atari.
These critical subtasks are what expert humans utilize to quickly learn in new environments that share
subtasks with previously learned environments, and are a reason for humans superior data efficiency
in learning complex tasks.
In the case of deliberately similar environments, we can construct the subtasks such that they are
similar in function and representation that an agent trained on the first environment can accelerate
learning on the second environment due to its preconstructed subtask representations, thus partially
avoiding the more complex environment’s increased simulation cost and inherent learning difficulty.
2
Related work
Transfer learning [13] is the method of utilizing data from one domain to enhance learning of another
domain. While sharing significant similarities to continual learning, transfer learning is applicable
across all machine learning domains, rather than being confined to reinforcement learning. For
example, it has had significant use with using networks trained on ImageNet [5] to accelerate or
enhance learning and classification accuracy by finetuning in a variety of vision tasks [12].
While the concept of continual learning, where an agent learns from a variety of experiences to
enhance future learning, has been defined for some time [15], it has remained largely untapped by
recent powerful algorithms that are best fit to benefit from its effects. Recent work has been done
with Progressive Neural Networks [17], where transfer learning was used to apply positive transfer
in a variety of reinforcement learning domains. However, our method differs in that it does not
add additional parameters for each environment or use lateral connections to features that result in
increased memory space and training time.
Recent work related to subtask utilization comes from Kirkpatrick et al. [7], which shows that
expertise can be maintained on multiple environments that have not been experienced for a long time
through elastic weight consolidation (EWC). Viable weights were found that simultaneously achieve
expertise in a variety of Atari games. Incremental learning similarly trains on multiple environments,
but with the goal of achieving enhanced expertise in a single environment, rather than expertise in all
environments. We leave it to future work to overcome this limitation.
3
Dex
Dex is a novel deep reinforcement learning toolkit for training and evaluating agents on continual
learning methods. Dex acts as a wrapper to the game Open Hexagon [16], sending screen pixel
information, reward information and performing actions via an OpenAI Gym like API [4]. Dex
contains hundreds of levels, each acting as their own environment. These environments are collected
into groups of similar environments for the task of continual learning. Dex environments vary greatly
in difficulty, ranging from very simple levels where agents achieve superhuman performance in less
that 4 minutes, to levels we consider far more complex than any previously learned environments.
Refer to videos available at github.com/innixma/dex of the Dex environments shown in Figure 1,
as screenshots do not capture the environment complexity.
Open Hexagon is a game involving navigating a triangle around a center figure to avoid incoming
randomly generated walls. Screenshots of various environments from the game are shown in Figure
1. The game progresses regardless of player action, and thus the player must react to the environment
in real-time. If the player contacts a wall the game is over. At each point in the game, a player has
only three choices for actions: move left, right, or stay put. It is a game of survival, with the score
and thus total reward being the survival time in seconds. Open Hexagon contains hundreds of levels
2
drastically ranging in difficulty, yet they each contain many similar core subtasks. This makes Open
Hexagon an ideal platform for testing continual learning methods.
(a) Distortion
(b) Lanes
(c) Reversal
(d) System
(e) Desync
(f) Stutter
(g) Arithmetic
(h) Stretch
(i) Open
(j) Overcharge
(k) Enneagon
(l) Blink
(m) Tunnel Town
(n) Roulette
(o) Echnoderm (p) Transmission
(q) Triptych
(r) Duality
(s) World Battle
(t) Universe
(u) Apeirogon
(w) Pi
(x) Golden Ratio
(v) Euclidean
Figure 1: Dex environments. The small triangle is the player that must be rotated around the center to
avoid incoming walls. Many of these environments incorporate various distortion effects that are not
evident in screenshots. Reversal periodically flips game controls, and some environments even add
additional actions, such as in Arithmetic, which requires the agent to correctly solve various math
equations during the level through the numpad.
The Dex toolkit along with its source code is available at github.com/innixma/dex.
4
Incremental learning
The novel continual learning method of incremental learning is defined as follows.
In the formal case, an agent must learn from a series of n environments E, each with identical legal
game actions A = {1, ..., K}. Note that any series of environments can be made to have the same
action space by considering A to be the superset of all possible actions in each game, with new
actions performing identically to no action, assuming no action is within the action space.
Each environment Ei has a corresponding cost per step ci > 0 and a step count si ≥ 0, indicating the
number of steps taken in that environment. Typically, more complex environments will have a higher
cost per step. A resource maximum M is given, which indicates the total amount of data that can be
gathered, shown in the following inequality:
M≥
n
X
ci si
i=1
The problem is to maximize the mean total reward R of the agent in the goal environment, En , while
maintaining the above inequality. Steps can be taken in any order from the n environments, and there
3
is no assumed correlation between the environments beyond their data and action dimensions. While
it may appear an optimal solution to only examine and gather data from En , as has been done for
virtually all reinforcement learning algorithms in the past, this is not always the case. For example, if
En−1 and En are highly correlated, and cn−1 << cn , then training with En−1 may be superior due
to its lesser cost. Additionally, an environment En−1 may contain important aspects of En , while
avoiding state spaces and rewards that are not useful for training, potentially allowing training on
En−1 to be optimal, even if cn−1 > cn .
By taking environments E to represent all real environments with their respective costs, the solution
to this problem corresponds to the globally optimal sequence of training steps to achieve maximum
performance with a finite amount of computational resources for a given algorithm. It therefore
necessarily contains the solution of achieving artificial general intelligence with minimal resources.
Unfortunately, this has several drawbacks. Most importantly, the selection of the optimal environments
is not obvious, and their order even less so.
While this formal definition is useful to define incremental learning, the following simplified version
will be what this paper focuses on. In the simplified case, all variables are the same, except that steps
must be taken in environments sequentially, without going back to previous environments. Thus,
once a step has been taken in Ei , no steps may be taken in future environments Ej where j < i.
Furthermore, it is assumed that the environments are correlated, where environment Ei−1 contains a
subset of subtasks in environment Ei , with environment Ei being typically harder than environment
Ei−1 .
The intuition behind this simplified case is that simple environments are both cheap to simulate
and easy to learn, and that the features and strategies learned from the simple environment could
transfer to a more difficult correlated environment. This process can be done repeatedly, producing
a compounding acceleration of learning as additional useful incremental environments are added.
Thus, the general process of incremental environment selection is to use easier subsets of the goal
environment. Furthermore, this means that every environment Ei in a simplified incremental learning
problem can be seen as the goal environment in an easier problem containing E1:i .
5
Baseline learning algorithm and model architecture
To analyse the effectiveness of incremental learning, our agents learned environments from Dex. For
training agents, we use Asynchronous Advantage Actor-Critic (A3C) framework introduced in [10],
coupled with the network architecture described below.
ConvNet implementation details. All Dex experiments were learned by a network with the following architecture. The network takes as input a series of 2 consecutive 42 × 42 grayscale images
of a game state. It is then processed through a series of 3 convolutional layers, of stride 1 × 1 and
size 5 × 5, 4 × 4, and 3 × 3 with filter counts of 32, 64, and 64 respectively. Between each pair of
convolutional layers is a 2 dimensional maxpooling layer, of size 2 × 2. The intermediate output is
then flattened and processed by a dense layer of 512 nodes. The output layers are identical to those
specified for A3C. All convolutional layers and maxpooling layers are zero-padded. This results
in an network with approximately 4,000,000 parameters. All intermediate layers are followed by
rectified linear units (ReLU) [11], and Adam is used for the optimizer, with a learning rate of 10−3 .
Learning rate is not changed over the course of training as opposed to Mnih et al. [10], but instead
stays constant. A gamma of 0.99 is used, along with n-step reward [14] with n = 4 , which is used to
speed up value propagation in the network, at a cost of minor instability in the learning. Training
is done with batches of 128 samples. The architecture was created using TensorFlow 1.1.0 [1], and
Keras 2.0.3.
Since Dex runs in real time, a slightly altered method of learning was utilized to avoid inconsistent time
between steps. Our implemented A3C was modified to work in an offline manner, with experience
replay as done in Deep-Q Networks [8]. This is a naive approximation to the more complex ACER
algorithm [22]. While this does destabilize the algorithm, which bases its computations on data being
representative of the network in its current state, the agents are still able to learn the environments
and significantly outperform Double Deep-Q Networks [20], thus serving as a reasonable baseline.
4
6
Experiments
So far, we have performed experiments on ten different environments in Dex for incremental learning.
These ten environments are split into two incremental learning sets of three and seven environments.
The first set will be referred to as a, and deals with increasingly complex patterns. The second set will
be referred to as b, and deals with increasingly complex task representation. The results show that
incremental learning can have a significant positive impact on learning speed and task performance,
but can also lead to bad initializations when trained on overly simplistic environments.
Code to reproduce the experiments in this paper will be released at a future date.
6.1
Setup
In the following experiments, we naively assume that achieving near optimal performance in Ei with
minimal resources requires as a prerequisite learning Ei−1 with minimal resources. This proceeds
downward to the base case of E1 , which can be considered a trivially solvable environment from
random weight initialization. This assumption is used to simplify training. Furthermore, due to the
exploratory nature of the baseline experiments, the costs per step are ignored, as we seek only to show
that positive feature transfer is occurring, rather than optimizing the feature transfer itself, which we
leave for future work.
All environments are learned with identical architecture, algorithm and hyperparameters. We act and
gather data on the environments 50 times per second. We use an -greedy policy with = 0.05, and a
replay memory size of 40,000. Replay memory is initialized by a random agent. For all experiments,
agents are trained with 75 batches after each episode. Total reward is equal to the number of seconds
survived in the environment each episode. Mean reward is calculated from an hour of testing weights
without training. Each episode scales in difficulty indefinitely the longer the agent survives.
6.2
Models
We evaluate four different types of models in this experiment. The first is a random agent for
comparison. The second is the baseline, which is the standard reinforcement learning training method.
To establish the baseline, each environment Ei is trained on from random initialization for one hour,
equivalent to roughly 150,000 training steps. The weights which achieve the maximum total reward
in a single episode are selected as the output weights from the training, called wi .
A third model, which we shall call initial, is the initial weights to the incremental learning method
before continued training. Thus, for environment Ei this model uses weights wi−1 . This is used to
measure the correlation between the two environments. We would expect uncorrelated environments
to result in near random agent reward for initial.
To establish the incremental learning agents, for each environment Ei we take the weights wi−1
outputted by the baseline, using them as the initial weights to training on Ei for an additional hour.
The weights outputted by this method we call wi0 .
6.3
Results and Observations
The results can be found in Table 1 and Table 2.
As can be seen in Table 1, incremental learning provided superior or roughly equivalent results
on every environment in set b, with some environments such as b3 and b4 experiencing substantial
increases in maximum reward, with the incrementally learned b4 achieving nearly triple the baseline
in both maximum and mean reward. However, on the harder environments, there was not significant
improvement seen. This is likely because the earlier environments were not sufficiently learned along
with the fact that the tasks were generally too difficult to learn in one hour of training, indicated by
the near random performance of b6 and b7 on both the baseline and incremental methods.
The other set, results shown in Table 2 for set a, show that incremental learning was harmful in the
case of a2 . This is likely due to the difference in the wall patterns of a1 and a2 . In a1 , a single wall
is on the screen at a time, requiring a simple avoidance. In a2 , up to four separate walls occupy
the screen at once, requiring a more complex method involving future planning and understanding
of which walls are more important in a given state. We suspect that learning a1 leads the agent to
5
Table 1: Set b rewards
Max
E
Random
Initial
Baseline
Incremental
b1
b2
b3
b4
b5
b6
b7
31.43
8.23
7.73
8.27
8.79
9.33
9.97
—
120.77
58.30
35.10
19.76
11.05
7.13
425.92
259.80
132.31
35.66
52.81
10.79
9.59
—
280.59
221.96
105.48
41.57
13.11
14.54
E
Random
Initial
Baseline
Incremental
b1
b2
b3
b4
b5
b6
b7
9.10
2.23
2.43
2.17
1.88
2.17
2.30
—
36.05
8.22
5.30
4.17
2.89
2.34
169.54
92.24
41.17
8.11
12.23
2.15
2.08
—
85.52
66.32
26.75
11.61
2.25
2.58
Mean
This table shows the max and mean reward for an agent on the given environment E with a given
training method, as described in the experimental setup. Here we observe that incremental learning
provided superior results to the baseline in nearly all environments, particularly b3 and b4 .
Table 2: Set a rewards
Max
E
Random
Initial
Baseline
Incremental
a1
a2
a3
40.73
19.65
10.85
—
53.37
15.69
771.67
445.85
49.50
—
86.57
59.10
E
Random
Initial
Baseline
Incremental
a1
a2
a3
8.93
7.46
6.01
—
18.34
7.32
717.93
87.06
9.81
—
18.31
13.52
Mean
This table shows the max and mean reward for an agent on the given environment E with a given
training method, as described in the experimental setup. Here we observe that incremental learning
provided inferior results to the baseline in the a2 environment, likely due to overfitting.
6
overtrain on a variety of weights, hindering future learning. This indicates that certain environments
may be too simple to include in incremental learning, as a1 can be learned to a reward of over 700 in
less than four minutes.
Additionally, the initial model shows that the environments are correlated, with generally far superior
performance than random, despite never training on the environment explicitly. In the case of b4 , it is
nearly equal to the baseline in the maximum metric, indicating significant correlation. This is likely a
reason for the greatly enhanced performance of incremental learning over the baseline in b4 .
Note that these experimental results are strictly a simple baseline for both Dex and incremental
learning, and suffer from instability due to the short timeline of training and lack of repeated
experiments. This means that agents do not necessarily consistently improve throughout training, but
rather may quickly improve to maximum performance followed by decreased performance for the
remainder of training. We leave more comprehensive experiments to future work.
7
Visualization
To qualitatively analyze the effects of incremental learning on a networks weights, we develop a
saliency visualization method based on Simonyan et al. [19] for reinforcement learning. Heatmaps are
generated with a given networks weights and an input image. The method for gathering the heatmaps
is identical to Simonyan et al. [19], and thus equations and derivations shall not be repeated in this
paper. The most likely action the network will take at a given frame is used for the action to minimize
through the gradient. This is a difference from the supervised learning case, where the ground truth is
used. In reinforcement learning, the ground truth is not known, and thus must be inferred, as done in
Wang et al. [21]. Results of the visualization on the trained weights of environment sets a and b can
be seen in Figure 2 and Figure 3.
Figure 2: Saliency mappings of a state from environment a3 , with weights from set a. The first row
consists of the trained baseline weights w1 to w3 . The second row consists of the incrementally
learned weights w20 to w30 .
Figure 3: Saliency mappings of a state from environment b3 , with weights from set b. The first row
consists of the trained baseline weights w1 to w7 . The second row consists of the incrementally
learned weights w20 to w70 .
7
As can be expected, saliency of a well performing trained agent will focus on the player location,
being very sensitive to changes in a small region surrounding the player. This intuition is confirmed
in the first row of Figure 3. The agents trained on the easier environments, such as b1 , b2 , and b3 ,
focus attention on the player and nearby threats. Interestingly, as the trained environment becomes
more complex, the agent appears less developed, and increasingly focuses on irrelevant locations in
the state, most prominent in the agent trained on b7 , which explicitly avoids attention on the player
and nearby threats, such as the walls. This indicates that the agent did not learn its environment
sufficiently in the time it was given, due to the complexity of its training environment. This is further
confirmed through the experimental results of b6 and b7 , which were near random.
Additionally, the saliency mappings show the correlation between the environments, as indicated by
the similarity in saliency in w1 and w2 to w3 in both Figure 2 and Figure 3, despite never explicitly
being trained on the environment that the state in the visualization is from.
Interestingly, the saliency mappings of the incrementally learned agents more closely resemble
the mappings of the weights it was initialized to than the weights its environment learned without
incremental learning. This suggests that a significant amount of the initialized weights are retained
after incrementally learning.
Further visualizations are included as a video of realtime saliency of an agent episode at
github.com/innixma/dex.
8
Future work
We hope to expand the analysis done in this paper by investigating incremental learning chains
involving more than two environments. This will more effectively explore the effects of incremental
learning in complex problems. We also wish to extend the training time of experiments to allow for
more complex environments such as those shown in Figure 1 to be learned, as well as to compare our
method to Progressive Neural Networks [17]. Incremental learning could be expanded to function as
a feature extractor in reinforcement learning. This would allow a more contained action space for
incremental learning tasks of varying domains, and is similar to the work done in Rusu et al. [17].
Additionally, training could be done separately on multiple environments and then combined to
learn an environment that shares subproblems with all the previous environments. This would be a
natural merging of incremental learning and elastic weight consolidation [7]. It may also provide a
synergistic effect to algorithms such as UNREAL [6] which rely on auxiliary tasks. These tasks could
act to maximize network utilization across multiple environments, potentially leading the network to
better generalizations through incremental learning.
Finally, environments such as Montezuma’s Revenge and Labyrinth [10] which require exploration
with sparse rewards are a natural expansion to the application of incremental learning. Developing
incremental exploration environments could result in far superior performance on exploration tasks.
9
Conclusion
The ability to learn and transfer knowledge across domains is essential to the advancement of agents
that can solve complex tasks. This paper introduced the continual learning toolkit Dex for training and
evaluation of continual learning methods. We proposed the method of incremental learning for deep
reinforcement learning, and demonstrated its ability to accelerate learning and produce drastically
superior results to standard training methods in multiple Dex environments, supporting the notion
of avoiding randomly initialized weights and instead using continual learning techniques to solve
complex tasks.
Source code for both the training methods used in this paper as well as the Dex toolkit source code
can be found at github.com/innixma/dex.
Acknowledgments
We thank Vittorio Romeo for designing Open Hexagon.
8
References
[1] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis,
J. Dean, M. Devin, S. Ghemawat, I. J. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia,
R. Józefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mané, R. Monga, S. Moore, D. G.
Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. A. Tucker,
V. Vanhoucke, V. Vasudevan, F. B. Viégas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke,
Y. Yu, and X. Zheng. Tensorflow: Large-scale machine learning on heterogeneous distributed
systems. CoRR, abs/1603.04467, 2016.
[2] M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An
evaluation platform for general agents. CoRR, abs/1207.4708, 2012.
[3] M. G. Bellemare, S. Srinivasan, G. Ostrovski, T. Schaul, D. Saxton, and R. Munos. Unifying
count-based exploration and intrinsic motivation. CoRR, abs/1606.01868, 2016.
[4] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba.
Openai gym. CoRR, abs/1606.01540, 2016.
[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical
image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE
Conference on, pages 248–255. IEEE, 2009.
[6] M. Jaderberg, V. Mnih, W. M. Czarnecki, T. Schaul, J. Z. Leibo, D. Silver, and K. Kavukcuoglu.
Reinforcement learning with unsupervised auxiliary tasks. CoRR, abs/1611.05397, 2016.
[7] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan,
J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran, and
R. Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National
Academy of Sciences, 114(13):3521–3526, 2017. doi: 10.1073/pnas.1611835114.
[8] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. A. Riedmiller. Playing atari with deep reinforcement learning. CoRR, abs/1312.5602, 2013.
[9] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves,
M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
[10] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. P. Lillicrap, T. Harley, D. Silver,
and K. Kavukcuoglu. Asynchronous methods for deep reinforcement learning. CoRR,
abs/1602.01783, 2016.
[11] V. Nair and G. E. Hinton. Rectified linear units improve restricted boltzmann machines. In
J. Fürnkranz and T. Joachims, editors, Proceedings of the 27th International Conference on
Machine Learning (ICML-10), pages 807–814. Omnipress, 2010.
[12] M. Oquab, L. Bottou, I. Laptev, and J. Sivic. Learning and transferring mid-level image
representations using convolutional neural networks. In The IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), June 2014.
[13] S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Transactions on knowledge and
data engineering, 22(10):1345–1359, 2010.
[14] J. Peng and R. J. Williams. Incremental multi-step q-learning. Machine Learning, 22(1):
283–290, 1996. ISSN 1573-0565. doi: 10.1023/A:1018076709321.
[15] M. B. Ring. Child: A first step towards continual learning. Machine Learning, 28(1):77–104,
1997.
[16] V. Romeo. Open hexagon, 2012.
[17] A. A. Rusu, N. C. Rabinowitz, G. Desjardins, H. Soyer, J. Kirkpatrick, K. Kavukcuoglu,
R. Pascanu, and R. Hadsell. Progressive neural networks. CoRR, abs/1606.04671, 2016.
9
[18] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrittwieser,
I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis.
Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587):
484–489, jan 2016. ISSN 0028-0836. doi: 10.1038/nature16961.
[19] K. Simonyan, A. Vedaldi, and A. Zisserman. Deep inside convolutional networks: Visualising
image classification models and saliency maps. CoRR, abs/1312.6034, 2013.
[20] H. van Hasselt, A. Guez, and D. Silver. Deep reinforcement learning with double q-learning.
CoRR, abs/1509.06461, 2015.
[21] Z. Wang, N. de Freitas, and M. Lanctot. Dueling network architectures for deep reinforcement
learning. CoRR, abs/1511.06581, 2015.
[22] Z. Wang, V. Bapst, N. Heess, V. Mnih, R. Munos, K. Kavukcuoglu, and N. de Freitas. Sample
efficient actor-critic with experience replay. CoRR, abs/1611.01224, 2016.
[23] K. Xu, J. Ba, R. Kiros, K. Cho, A. C. Courville, R. Salakhutdinov, R. S. Zemel, and Y. Bengio. Show, attend and tell: Neural image caption generation with visual attention. CoRR,
abs/1502.03044, 2015.
10
| 2cs.AI
|
arXiv:1803.01209v1 [math.PR] 3 Mar 2018
University of Reading
Department of Mathematics and
Statistics
Stochastic Resonance for a Model with Two Pathways
Tommy Liu
Thesis submitted for the Degree of Doctor of Philosophy
September 2016
Abstract
In this thesis we consider stochastic resonance for a diffusion with drift given by a potential,
which has two metastable states and two pathways between them. Depending on the
direction of the forcing the height of the two barriers, one for each path, will either oscillate
alternating or in synchronisation.
We consider a simplified model given by discrete and continuous time Markov Chains
with two states. This was done for alternating and synchronised wells. The invariant
measures are derived for both cases and shown to be constant for the synchronised case.
A PDF for the escape time from an oscillatory potential is reviewed.
Methods of detecting stochastic resonance are presented, which are linear response,
signal-to-noise ratio, energy, out-of-phase measures, relative entropy and entropy. A new
statistical test called the conditional Kolmogorov-Smirnov test is developed, which can be
used to analyse stochastic resonance.
An explicit two dimensional potential is introduced, the critical point structure derived
and the dynamics, the invariant state and escape time studied numerically.
The six measures are unable to detect the stochastic resonance in the case of synchronised saddles. The distribution of escape times however not only shows a clear sign of
stochastic resonance, but changing the direction of the forcing from alternating to synchronised saddles an additional resonance at double the forcing frequency starts to appear.
The conditional KS test reliably detects the stochastic resonance even for forcing quick
enough and for data so sparse that the stochastic resonance is not obvious directly from
the histogram of escape times.
1
Declaration
I confirm that this is my own work and the use of all material from other sources has been
properly and fully acknowledged.
Tommy Liu
2
Acknowledgement
I would like to thank Tobias Kuna for supervising this thesis; Valerio Lucarini for cosupervising; Tristan Pryer, Horatio Boedihardjo, Martin Kolb and the late Professor Alexei
Likhtman for being on the Monitoring Committee; Jochen Broecker and Ostap Hryniv for
being the examiners on my viva; Peter Imkeller for helpful discussions; Pawel Stasiak for
introducing me to the Meteorology Computer Clusters; Peta-Ann King and Sue Davis for
their pastoral care; the EPSRC for funding and finally to my family and friends for their
support over the years.
Tommy Liu
September 2016
University of Reading
3
Contents
Abstract
1
Declaration
2
Acknowledgement
3
Introduction
Outline of Problem . . . . . . .
Historical Background . . . . .
Physical Background . . .
Mathematical Background
Summary of Research . . . . .
.
.
.
.
.
11
11
12
12
12
13
.
.
.
.
.
.
15
15
20
21
23
25
27
.
.
.
.
.
.
.
.
28
28
28
29
30
31
32
34
34
3 Theoretical Escape Time from a Well of an Oscillatory Potential
3.1 Markov Chain Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Discrete Time Markov Chain . . . . . . . . . . . . . . . . . . . . . . . . . .
36
36
38
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1 Stochastic Resonance
1.1 Laplace Method . . . . . . . . . . . . . . . . . .
1.2 One Dimensional Potential . . . . . . . . . . . .
1.2.1 One Dimensional Potential - Case F = 0
1.2.2 One Dimensional Potential - Case F 6= 0
1.2.3 Conclusion and Resonance Condition res
1.3 Remarks on One Dimensional Potential . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2 Theoretical Escape Time from a Well of a Static Potential
2.1 Freidlin-Wentzell Theory and Large Deviation . . . . . . . . . . .
2.1.1 Stochastic Processes . . . . . . . . . . . . . . . . . . . . .
2.1.2 Deterministic Limit . . . . . . . . . . . . . . . . . . . . . .
2.1.3 Action Functional for Wiener processes . . . . . . . . . . .
2.1.4 Action Functional for General processes . . . . . . . . . .
2.1.5 Main Theorems . . . . . . . . . . . . . . . . . . . . . . . .
2.1.6 Remarks on Freidlin-Wentzell and Large Deviation Theory
2.2 Kramers’ Formula and Potential Theory . . . . . . . . . . . . . .
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Discrete Time Markov Chain - Alternating Saddles p 6= q . . . . . .
Discrete Time Markov Chain - Synchronised Saddles p = q . . . . .
Discrete Time Markov Chain - Invariant Measures, Relaxation Time
and Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . .
Continuous Time Markov Chain . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Continuous Time Markov Chain - Alternating Saddles p 6= q . . . .
3.3.2 Continuous Time Markov Chain - Synchronised Saddles p = q . . .
3.3.3 Continuous Time Markov Chain - Invariant Measures and Fourier
Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Probability Density Function of Escape Times . . . . . . . . . . . . . . . .
3.4.1 Normalised Time Probability Density Function of Escape Times . .
3.4.2 Perfect Phase Approximation of Probability Density Function of Escape Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Adiabatic Large Deviation . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1
3.2.2
3.2.3
3.3
3.4
3.5
4 Theory of Analysis of Stochastic Resonance
4.1 Six Measures of Stochastic Resonance . . . . .
4.2 Statistical Tests . . . . . . . . . . . . . . . . .
4.2.1 Kolmogorov-Smirnov Test . . . . . . .
4.2.2 Conditional Kolmogorov-Smirnov Test
.
.
.
.
5 Mexican Hat Toy Model
5.1 Case Fx = 0 and Fy = 0 . . . . . . . . . . . . .
5.2 Case Fx > 0 and Fy = 0 . . . . . . . . . . . . .
5.2.1 Case Fx > 0, Fy = 0 and b < 21 . . . . . .
5.2.2 Case Fx > 0, Fy = 0 and b ≥ 12 . . . . .
5.3 Case Fx = 0 and Fy > 0 . . . . . . . . . . . . .
5.3.1 Case Fx = 0, Fy > 0 and b < 12 . . . . . .
5.3.2 Case Fx = 0, Fy > 0 and b ≥ 12 . . . . . .
5.4 Case Fx 6= 0 and Fy 6= 0 . . . . . . . . . . . . .
5.5 Remarks on Mexican Hat . . . . . . . . . . . .
5.5.1 Beyond Criticality . . . . . . . . . . . .
5.5.2 Numerical Problems . . . . . . . . . . .
5.5.3 Comparison with One Dimensional Case
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
39
43
44
47
48
50
51
53
54
54
55
.
.
.
.
57
57
60
60
61
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
64
68
69
69
81
82
82
88
91
95
96
96
97
6 Numerical Methods
6.1 Basic Conditions - Estimating tstep ≤ t1 , tstep ≤ t2 and tstep ≤ t3 . .
6.2 Increment Conditions . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 Increment Theory - Developing W (S, l) . . . . . . . . . . . .
6.2.2 Absence of Large Jumps - Estimating tstep ≤ t4 and tstep ≤ t5
6.3 Stability and Radius Conditions . . . . . . . . . . . . . . . . . . . .
6.3.1 Stability Problems . . . . . . . . . . . . . . . . . . . . . . .
6.3.2 Estimating R ≤ R1 and R ≤ R2 . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
99
100
102
103
105
107
111
112
5
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6.4
6.3.3 Estimating tstep ≤ t6 . .
Selection of Parameters . . . . .
6.4.1 Selection of Parameters 6.4.2 Selection of Parameters 6.4.3 Selection of Parameters 6.4.4 Selection of Parameters 6.4.5 Selection of Parameters 6.4.6 Selection of Parameters 6.4.7 Selection of Parameters -
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
Simulation . . . . . . . . . . . . . . .
Validity of Kramers’ Formula . . . . .
Adiabatic Approximation . . . . . . .
Stability of Deterministic Trajectory .
Random Number Generator . . . . . .
Calculating Positions of Critical Points
Higher Precision Numerics . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
113
114
114
115
116
117
118
118
118
7 Simulations, Results and Analysis
121
7.1 Details of the Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.2 Six Measures Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2.1 Interpretation of the Six Measures Analysis . . . . . . . . . . . . . 128
7.3 Escape Time and Conditional KS Test Analysis . . . . . . . . . . . . . . . 130
7.3.1 Interpretation of the Escape Time and Conditional KS Test Analysis 140
7.4 Sparse Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.4.1 Interpretation of the Sparse Data Analysis . . . . . . . . . . . . . . 144
7.5 Remarks on Analysis of Stochastic Resonance . . . . . . . . . . . . . . . . 144
7.5.1 Remarks on Implementing the Conditional KS Test . . . . . . . . . 144
7.5.2 Remarks on Adiabatic Approximation . . . . . . . . . . . . . . . . 145
Conclusion
146
Outline of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Further Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
A Conventions in Defining the SDEs, Potential, Time Dependency and
Forcing
149
B Further Numerical Methods
B.1 Numerical Methods for measuring Escape Times . . . . . . . . . . . . . . .
B.2 Numerical Methods for calculating Fourier Transform and Linear Response
B.3 Numerical Methods for calculating M5 and M6 . . . . . . . . . . . . . . . .
151
151
152
153
C Further Commentary on Sparse Data Analysis
155
C.1 Examples of Oversampling . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
C.2 Empirical CDF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
References
158
6
List of Figures
1.1
1.2
1.3
5.1
5.2
5.3
5.4
5.5
This trajectory is exhibiting quasi-determinism.
nance. . . . . . . . . . . . . . . . . . . . . . . .
Transitions occur irregularly and are rare. . . .
Transitions occur very often. . . . . . . . . . . .
It is
. . .
. . .
. . .
near stochastic reso. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
26
26
27
As Fx increases from 0 to Fxcrit , the x1,2,3 move as shown in the diagram.
As Fx increases from 0 to Fxsad the (xsaddle , ±ysaddle ) meet each
√other on the
x-axis. There are three possible paths for (xsaddle , ±ysaddle). If 1 − 2b ∈ R1 ,
√
√
then the two (xsaddle , ±ysaddle ) would meet in the interval √13 1 + 2a, 1 + 2a
√
and collide into (x2 , 0). If
2b ∈ R
saddle )
1−
2 , then the two (xsaddle , ±y
√
√
would meet in the interval 0, √13 1 + 2a and collide into (x1 , 0). If 1 − 2b =
√
√1
1 + 2a, which is also when Fxsad = Fxcrit , then the two (xsaddle , ±ysaddle )
3
√
would meet at x = √13 1 + 2a and collide simultaneously into (x1 , 0) and
(x2 , 0). . . . . . . . . . .√. . . . . . . . . . . . . . . . . . . . . . . . . . . 78
This is the case for when 1 − 2b ∈ R1 . When F = Fxsad the two saddles
collide into the right well and turns into a new saddle. At Fxcrit , this newly
created saddle collides into the hill and both disappears. When Well 2 turns
into a Saddle here, it is like creating a new path for the particle to transit
to Well 1. . . . . . √
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
This is the case for 1 − 2b ∈ R2 . When F = Fxsad the two saddles collide
into the hill and turns into a new saddle. At Fxcrit , this newly created saddle
collides into the right well and both disappears. This system behaves in a
similar way to a One Dimensional
Potential.
. . . . . . . . . . . . . . . . 80
√
√
This is the case for when 1 − 2b = √13 1 + 2a, which is also when Fxsad =
Fxcrit . At F = Fxsad = Fxcrit the two saddles, hill and right well mutually
collide at the same place and disappears. . . . . . . . . . . . . . . . . . . 81
As Fx increases from 0 to Fxcrit , the saddle collides into the right well and
disappears. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
7
5.6
As Fy increases from 0 to Fycrit the y0 , y1 and y2 move as shown in the
diagram. As Fy increases from 0 to Fysad , the two (±xwell , ysaddle ) meet
each other on the y-axis. There are three possible paths for (±xwell , ywell ).
If Fysad < Fycrit , then the two (±xwell , ywell ) meet between in the interval
√
√
− √23 1 − 2b, − 1 − 2b . If Fysad = Fycrit , then the two (±xwell , ywell ) meet
√
at y = − √23 1 − 2b. If Fysad > Fycrit then the two (±xwell , ywell ) meet in the
√
interval −∞, − √23 1 − 2b . . . . . . . . . . . . . . . . . . . . . . . . . .
At F = Fycrit the top saddle collides into the hill and both then disappears.
At F = Fysad the bottom saddle collides with the two wells and turns into a
new well. These two collisions can occur simultaneously or occur one after
the other, depending on whether we have Fysad < Fycrit , Fysad = Fycrit or
Fysad > Fycrit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.8 At Fy = F sad the saddle collides with the two wells and turns into a new well.
5.9 The critical points move very little here. . . . . . . . . . . . . . . . . . . .
5.10 The critical points have a more extreme trajectory here. . . . . . . . . . .
5.11 Notice that the use of F crit as a critical force is just an educated guess. Here
the system is so close to criticality the saddle is almost colliding with the hill.
5.12 An example of the potential VF (x, y) = 41 r4 − 12 r2 − ax2 + by 2 + Fx x + Fy y
p
where r = x2 + y 2 . Here a = 0.1, b = 0.1, Fx = 0.1 and Fy = 0. Notice
there are two saddles just ahead of the hill. The well on the right is higher
than the well on the left. . . . . . . . . . . . . . . . . . . . . . . . . . . .
87
5.7
6.1
6.2
6.3
6.4
6.5
6.6
6.7
7.1
7.2
7.3
7.4
7.5
7.6
7.7
7.8
The trajectory is so unstable the particle even transits to the other well. .
The trajectory is more stable but the particle now oscillates near the well.
The trajectory is sufficiently stable here. . . . . . . . . . . . . . . . . . . .
Notice that transitions tend to occur near the saddles. This is when Kramers’
formula gives a good approximation for the escape rates and escape times.
For higher noise levels transitions would occur near the hill, which is close
to the origin. Kramers’ formula is not a good approximation here. . . . .
Here 2239 transitions were used. The averaged measured escape time is
0.0977T . The radius used was R = 0.5386. . . . . . . . . . . . . . . . . .
Here 56244 transitions were used. The averaged measured escape time is
0.1064T . The radius used was R = 0.19. . . . . . . . . . . . . . . . . . . .
The
The
The
The
The
The
The
The
measure
measure
measure
measure
measure
measure
measure
measure
M1
M2
M1
M2
M3
M4
M5
M6
for
for
for
for
for
for
for
for
the
the
the
the
the
the
the
the
diffusion case for various angles and noise levels. .
diffusion case for various angles and noise levels. .
Markov Chain for various angles and noise levels.
Markov Chain for various angles and noise levels.
Markov Chain for various angles and noise levels.
Markov Chain for various angles and noise levels.
Markov Chain for various angles and noise levels.
Markov Chain for various angles and noise levels.
8
88
91
94
94
95
96
107
108
108
115
115
119
120
124
124
125
125
126
126
127
127
7.9
7.10
7.11
7.12
7.13
7.14
7.15
7.16
7.17
7.18
7.19
7.20
7.21
7.22
7.23
7.24
7.25
7.26
7.27
7.28
7.29
The blue trajectory is x(t) and the green trajectory is y(t). . . . . . . . . .
The blue trajectory is x(t) and the green trajectory is y(t). . . . . . . . . .
The blue trajectory is x(t) and the green trajectory is y(t). . . . . . . . . .
This is an example of the Single Frequency. . . . . . . . . . . . . . . . . .
This is an example of the Intermediate Frequency. . . . . . . . . . . . . .
This is an example of the Intermediate Frequency tending closer to the
Double Frequency. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
This is an example of the Double frequency. . . . . . . . . . . . . . . . . .
The u is the time of entrance into the well and t is the time of exit from the
well. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The u is the time of entrance into the well and t is the time of exit from the
well. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The u is the time of entrance into the well and t is the time of exit from the
well. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The u is the time of entrance into the well and t is the time of exit from the
well. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The u is the time of entrance into the well and t is the time of exit from the
well. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The u is the time of entrance into the well and t is the time of exit from the
well. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
This is an example of the conditional KS test being implemented
for the
√
data in √
Figure 7.12. Note that = 0.18, φ = 0◦ , n = 200, nSn− = 0.5233
and Q ( nSn− ) = 0.0529. . . . . . . . . . . . . . . . . . . . . . . . . . . .
This is an example of the conditional KS test being implemented
for the
√
data in √
Figure 7.13. Note that = 0.20, φ = 84◦ , n = 200, nSn− = 0.6223
and Q ( nSn− ) = 0.1665. . . . . . . . . . . . . . . . . . . . . . . . . . . .
This is an example of the conditional KS test being implemented
√ − for the
◦
data in √
Figure 7.14. Note that = 0.21, φ = 87 , n = 200, nSn = 1.2587
and Q ( nSn+ ) = 0.9159. . . . . . . . . . . . . . . . . . . . . . . . . . . .
This is an example of the conditional KS test being implemented
√ − for the
◦
data in √
Figure 7.15. Note that = 0.21, φ = 90 , n = 200, nSn = 1.0465
and Q ( nSn− ) = 0.7766. . . . . . . . . . . . . . . . . . . . . . . . . . . .
The ptot ≈ p+ (t, 0) is not a good approximation here. . . . . . . . . . . . .
This is a KS test on the data in Figure 7.26. The conditional null hypothesis
can be reasonably
accepted. Note that = 0.17, φ = 81◦ , n = 20 and
√
Sn+ = 0.2750. Q( nSn+ ) = 0.9029. . . . . . . . . . . . . . . . . . . . . . .
The ptot ≈ p+ (t, 0) is not a good approximation here. . . . . . . . . . . . .
This is a conditional KS test on the data in Figure 7.28. The conditional
null hypothesis can be reasonably accepted.√Note that = 0.27, φ = 78◦ ,
n = 20 and Sn− = 0.1400. Also note that Q( nSn− ) = 0.1720. . . . . . . .
129
129
130
131
132
132
133
134
134
135
135
136
136
138
139
139
140
142
142
143
143
C.1 This is Figure 7.24 redone with 20 transitions. Note that = 0.21, φ = 87◦ ,
n = 20, Sn+ = 0.1960. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
9
C.2 This is Figure 7.25 redone with 20 transitions. Note that = 0.21, φ = 90◦ ,
n = 20, Sn− = 0.1030. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
10
Introduction
Outline of Problem
Consider the following problem. Let Xt be the random variable describing the trajectory
of a diffusion process in Rr where t is the time and 2 is the variance level. More precisely
we consider processes described by the following type of stochastic differential equation
dXt = b (Xt , t) dt + dWt
where b : Rr × R −→ Rr and Wt is a Wiener process in Rr . We suppose that the drift term
b has the form
b(x, t) = −∇V0 (x) + F cos Ωt
where F, x ∈ Rr and V0 : Rr −→ R is called the unperturbed potential. We consider
unperturbed potentials with two or more minimas (wells). Most importantly, we consider
potentials where there are multiple pathways between the wells. To our knowledge systems
with two pathways have not been studied in the context of stochastic resonance.
Consider the case Ω = 0 and where the noise is very small. The particle will stay very
close to one of the wells of the potential and will occasionally escape to the other well. The
time of the actual transition from one well to the other is very short compare to the time
it stays in any particular well.
Now consider the case where Ω > 0. For particular choices of Ω > 0 and > 0, these
transitions between the two wells will become synchronised with the driving frequency
Ω. This is called stochastic resonance. Thus the term noise induced synchronisation was
used for systems where the amplitude of the forcing F was not large [1, 2] (see also the
discussions in [3]). New insights into the exact manner of these synchronised transitions
will be studied in this thesis, which may be more appropriate in light of the results obtained
in this thesis.
For small noise , one would expect that stochastic resonance depends only on the
essential properties of the system, such as the height difference between the wells and the
pathways for escape. We investigate what effects these multiple pathways have on the
appearance of stochastic resonance. Varying F , Ω and should thus reveal the qualitative
structure of the unperturbed potential V0 . In this thesis we test this paradigm by studying
a two dimensional example with two wells and two independent pathways between them,
see Chapter 5.
11
Historical Background
Stochastic resonance has attracted interest among mathematicians and physicist. An
overview of the studies that have occurred in both physics and mathematics are given
here.
Physical Background
Stochastic resonance was first observed in 1981 [4, 5, 6]. The first example [4] considered
transitions between two metastable states to model the cyclic occurrences of ice ages. Since
then many examples of stochastic resonance were found in optics [7, 8, 9, 10], electronics
[11, 12, 13, 14, 15, 16, 17, 18, 19], neuronal systems [20], quantum systems [21, 22] and
paddlefish [23, 24]. Stochastic resonance could be thought of as quasi-deterministically
periodic transition between two metastable states. For example, the climate of the Earth
could be modelled by two states. There is a state corresponding to an Ice Age and another
corresponding to the opposite of an Ice Age, a so-called “Hot Age”. As the Earth’s climate
cyclically changes many times between Cold Ages and Hot Ages, its behaviour could be
modelled by stochastic resonance.
A range of techniques for example linear response [25, 26], signal-to-noise ratio [27,
28] and distribution of escape times [29, 28, 30] were used to define, analyse and study
stochastic resonance. These techniques along with other examples of stochastic resonance
are reviewed in the long overview paper by Gammaitoni, Hänggi, Jung and Marchesoni
[31]. We will evaluate the usefulness of some of these techniques for our problem, see
Chapter 7.
Mathematical Background
There are various mathematical studies of stochastic resonance. These often involve different orders of approximations for small noise levels. The first and second order of approximations are discussed below. Adiabatic large deviation is also presented.
In the first leading order of approximation, a key element of study is to control the
escape times from the wells as given by the so called large deviation theory, see the monograph of Freidlin and Wentzell [32]. The distribution of the exit time was derived by Day
in [33] and by Galves, Kifer, Olivieri and Vares [34, 35, 36]. To go beyond leading order
has been much more difficult for the transition problem between two wells as WKB theory
could up to now not be rigorously applied.
The next order of approximation was rigorously derived by Bovier, Eckhoff, Gayrard,
Klein [37] and Berglund and Gentz [38] using techniques from potential theory. Berglund
and Gentz in a series of papers studied the situation of low, non-quadratic barriers and
drifts not given by autonomous potentials [38, 3]. A review of different techniques used to
derive Kramers’ formula can be found in the review paper [39].
In [40] Friedlin considered stochastic resonance in the adiabatic regime. This means
the diffusion can effectively be described by a Markov process which describes the jumps
12
between wells. This problem was revisited by Hermann, Imkeller and Pavlyukevich, see
Chapter 4 in [41] and references therein, to derive results uniformly for varying time scale
to identify the optimal resonance point asymptotically for small noise even outside the
adiabatic regime leading to different logarithmic corrections including the famous cycling
effect discovered by Day [42], see also [43] for the connection with stochastic resonance.
Escape time outside of adiabatic regime is studied in [44].
As mentioned above in leading order the transitions of the diffusion process Xt between the wells can be approximated by a two state Markov Chain Yt = ±1 which have
been studied [45, 46, 47, 41]. Further comparative studies of the stochastic resonance for
the diffusion case Xt versus the Markov Chain Yt case were done by Hermann, Imkeller,
Pavlyukevich and Peithmann in [48, 49, 50, 51]. A collection of papers on comparative
studies between stochastic resonance in diffusion and Markov Chains can be found in the
monograph [41]. One of the main conclusions in [50, 51, 49, 41] is rigorously showing that
using linear response and signal-to-noise ratio to analyse stochastic resonance in the diffusion case Xt gives a different result to analysing the Markov Chain case Yt = ±1 with the
same techniques even asymptotically in the small noise limit. Other common methods used
to study stochastic resonance include invariant measures and Fourier transforms. We consider six measures of stochastic resonance frequently used and considered by Pavlyukevich
in his thesis [45, 41] which are linear response, signal-to-noise ratio, energy, out-of-phase
measure, relative entropy and entropy.
In this thesis we will study stochastic resonance on a two dimensional toy model, in
both the diffusion and Markov Chain cases, and where there are two independent pathways
between the wells going through two different saddles. The escape times and the six
measures of stochastic resonance introduced above are studied.
Summary of Research
In Chapter 1 we review the first model in which stochastic resonance was observed, that
is, we are considering the unperturbed potential
V0 (x) =
x2
x4
−a
4
2
and the corresponding SDE
dXtε = [−∇V0 + F cos(Ωt)] dt + dWt .
In one dimension the escape time can be explicitly computed as the solution to an ODE
and using Laplace method asymptotic formulas can be derived. In Chapter 2 a review
of large deviation theory and results concerning escape times are given. In Chapter 2.2
further results, based on potential theory, are given and the analogue of Kramers’ formula
for our case is presented. In Chapter 3 discrete and continuous time Markov Chains are
considered. The associated invariant measures and the relaxation time to this invariant
measure is derived for alternating and synchronised wells. The probability density function
13
of escape times is derived as well. In Chapter 4 the six measures used to analyse stochastic
resonance mentioned above are introduced. Furthermore, methods used to study escape
times are given and in particular a new version of the Kolmogorov-Smirnov test suitable for
this problem is discussed. In Chapter 5 the main model under consideration in this thesis
is studied, which has two wells and two saddles. The two wells are connected through to
independent pathways each through one of the saddles. Due to its form, we nicknamed it
the Mexican Hat Toy Model
1
1
V0 (x, y) = r4 − r2 − ax2 + by 2
4
2
where r =
p
x2 + y 2 .
We rigorously derive the qualitative structure of the potential with and without external
forcing. In Chapter 6 the numerical methods used to simulate the associated SDE
∂V0
dx = −
+ Fx cos Ωt dt + dwx
∂x
∂V0
dy = −
+ Fy cos Ωt dt + dwy
∂y
are discussed and non rigorous estimates of all relevant error sources are given necessary to
be confident about the precision of the simulation needed. The dwx and dwy are x and y
components of the two dimensional Wiener processes. The numerical algorithm used is the
Euler method [52] which is sufficiently accurate for our purposes. In Chapter 7 the results
from simulating the SDE are presented and interpreted. The six measures are studied and
the quality of the approximation by the aforementioned Markov chains is tested using the
Kolomogorov-Smirnov test developed. The results were repeated in a sparse data context.
In Chapter 7 the main findings and conclusions of this thesis are presented. It is shown
that the six measures are unable to detect stochastic resonance in the case of synchronised
saddles. The six measures show no sharp signature as the saddles change from alternating
to synchronised saddles. This is due to the fact that the invariant measures are constant
for synchronised saddles. By contrast, not only did the distribution of escape times show
a signature for stochastic resonance with synchronised saddles; the distribution of escape
times did show a clear signature as the saddles change from alternating to synchronised, by
exhibiting signatures which we call the Single, Intermediate and Double Frequency. The
newly developed conditional Kolomogorov-Smirnov test was shown to be a good method
to analyse the statistics of the escape times.
This thesis then finishes with a conclusion of the results obtained. In Appendix A the
conventions used are collected. In Appendix B the methods used to calculate the Fourier
transforms and the escape times are explained.
14
Chapter 1
Stochastic Resonance
The earliest known and simplest example of stochastic resonance is reviewed. This was
done in 1981 [4]. Properties about its escape times are derived. Estimates for the resonance
noise level res are given. The techniques involved include a review of Laplace method. This
study only works for in the small noise approximation. Only one dimensional systems
will be studied in this Chapter. Deducing properties about the underlying potential is
trivial.
1.1
Laplace Method
The main technique used to study exit times in one dimension is the so called Laplace
Method. For completeness and to get a better understanding of the mechanism we are
going to study, a proof will be provided later on.
Theorem 1.1. (Laplace Method) Let f : [a, b] → R be twice differentiable on [a, b]. Let
x0 ∈ (a, b) be unique such that f (x0 ) = maxx∈(a,b) f (x). Assuming f 00 (x) is continuous on
[a, b] with f 0 (x0 ) = 0 and f 00 (x0 ) < 0 then
R b nf (x)
e
dx
a
= 1.
q
lim
n→∞
2π
nf
(x
)
0
e
n(−f 00 (x0 ))
A Corollary follows from Laplace Method as a special case of Theorem 1.1.
Corollary 1.2. Let f : [a, b] → R be twice differentiable on [a, b]. Let x0 = a or x0 = b
be unique such that f (x0 ) = maxx∈[a,b] f (x). Assuming f 00 (x) is continuous on [a, b] with
f 0 (x0 ) = 0 and f 00 (x0 ) < 0 then
R b nf (x)
e
dx
a
= 1.
q
lim
n→∞
1 nf (x0 )
2π
e
2
n(−f 00 (x0 ))
15
We recall Taylor’s Remainder Theorem which is needed in the proofs.
Theorem 1.3. Suppose that f : R → R is (n + 1) times differentiable on R. Let x, a ∈ R,
with x > a then f can be expressed as
f (x) = f (a) +
f 0 (a)
f 00 (a)
f n (a)
(x − a) +
(x − a)2 + · · · +
(x − a)n + Rn+1 (x)
1!
2!
n!
where Rn+1 the remainder can be expressed as
1
Integral Form Rn+1 (x) =
n!
Z
x
(x − t)n f n+1 (t)dt
a
n+1
Lagrange Form Rn+1 (x) =
f (ξ)
(x − a)n+1 , ξ ∈ [a, x].
(n + 1)!
The following simple Lemma is also needed in the proof of Laplace Method.
Lemma 1.4. Let f : [a, b] −→ R be continuous. Let x0 be a unique maximum such that
f (x0 ) = maxx∈[a,b] f (x), then for any fixed δ > 0, there exists an η > 0 such that for any
s∈
/ (x0 − δ, x0 + δ) we have
η ≤ f (x0 ) − f (s).
Now we review proofs of the methods needed.
Proof of Lemma 1.4. We know that x0 is the unique maximum, which means
0 < f (x0 ) − f (s)
for any s ∈
/ (x0 − δ, x0 + δ). This means the infinum of the set is bounded by zero
inf {f (x0 ) − f (s) : s ∈
/ (x0 − δ, x0 + δ)} ≥ 0.
Suppose that the infinum of the set is zero
inf {f (x0 ) − f (s) : s ∈
/ (x0 − δ, x0 + δ)} = 0
and yet all elements of the set are strictly greater than zero. This means some members
would be arbitrarily close to zero,
0 < f (x0 ) − f (s) <
where is arbitrarily small. There exists a sequence
(sn )n≥1 ⊂ [a, b]\(x0 − δ, x0 + δ) = [a, x0 − δ] ∪ [x0 + δ, b]
such that
0 < f (x0 ) − f (sn ) <
1
1
=⇒ f (x0 ) < f (sn ) + .
n
n
16
But this sequence is in a compact set, which must have a subsequence which converges to
a member s0 ∈ [a, x0 − δ] ∪ [x0 + δ, b], that is
f (x0 ) ≤ f (s0 )
which contradicts the fact x0 is the unique maximum. This implies that
inf {f (x0 ) − f (s) : s ∈
/ (x0 − δ, x0 + δ)} > 0
so the η > 0 as in the assertion of the Lemma must exist.
Proof of Theorem 1.1. A differentiable function is also a continuous function. Since f (x0 ) =
maxx∈[a,b] f (x) we can say f 0 (x0 ) = 0. Using the Taylor’s Remainder Theorem we can
rewrite f (x) for x ∈ [x0 , x0 + δ] for some δ > 0 and ξ ∈ [x0 , x] as
f (x) = f (x0 ) +
f 00 (ξ)
(x − x0 )2 .
2
We can obtain an upper and lower bound for f 00 (ξ) by exploiting its continuity on [a, b].
Since x ∈ [x0 , x0 + δ] we must also have ξ ∈ [x0 , x0 + δ]. So
|ξ − x0 | ≤ δ.
For any > 0 and for a sufficiently small δ, we can have
|f 00 (ξ) − f 00 (x0 )| < .
This means we can say
− < f 00 (ξ) − f 00 (x0 ) <
f 00 (x0 ) − <f 00 (ξ) < f 00 (x0 ) +
which gives
1
f (x) ≤ f (x0 ) + (f 00 (x0 ) + )(x − x0 )2
2
1
f (x) ≥ f (x0 ) + (f 00 (x0 ) − )(x − x0 )2 .
2
(1.1)
(1.2)
We start with the lower bound for f (x) as in Equation 1.1
Z b
Z x0 +δ
nf (x)
e
dx ≥
enf (x) dx
x0 −δ
a
≥e
nf (x0 )
Z
x0 +δ
n
e 2 (f
00 (x
2
0 )−)(x−x0 )
dx
√
x0 −δ
=e
nf (x0 )
Z
1
p
n(−f 00 (x0 ) + )
17
+δ
−δ
√
n(−f 00 (x0 )+)
n(−f 00 (x0 )+)
1 2
e− 2 y dy
where we have made a transformation using
p
y = n(−f 00 (x0 ) + )(x − x0 ).
nf (x0 )
Dividing both sides by e
Rb
nf (x)
a
enf (x0 )
e
q
q
2π
n(−f 00 (x0 ))
dx
2π
n(−f 00 (x0 ))
≥ √1
2π
s
gives
√
−f 00 (x0 )
−f 00 (x0 ) +
+δ
Z
−δ
√
n(−f 00 (x0 )+)
1 2
e− 2 y dy.
(1.3)
n(−f 00 (x0 )+)
Using Lemma 1.4 we can say that for any fixed δ, there exists an η > 0 such that for any
s∈
/ (x0 − δ, x0 + δ) we have
η ≤ f (x0 ) − f (s).
So we can proceed with
Z b
Z x0 −δ
Z
nf (x)
nf (x)
e
dx =
e
dx +
a
e
nf (x)
Z
x0 −δ
≤
e
n(f (x0 )−η)
b
enf (x) dx
dx +
x0 −δ
a
Z
x0 +δ
x0 +δ
x0 +δ
Z
nf (x)
dx +
e
Z
x0 −δ
a
n(f (x0 )−η)
b
en(f (x0 )−η) dx
dx +
x0 +δ
n(f (x0 )−η)
= (x0 − δ − a)e
+ (b − x0 − δ)e
Z
x0 +δ
enf (x) dx
+
x0 −δ
= (b − a − 2δ)e
n(f (x0 )−η)
x0 +δ
Z
enf (x) dx.
+
x0 −δ
Now we use the upper bound for f (x) from Equation 1.2. So
b
Z
e
nf (x)
n(f (x0 )−η)
dx ≤ (b − a)e
nf (x0 )
Z
x0 +δ
a
≤ (b − a)en(f (x0 )−η) + enf (x0 )
n
e 2 (f
+e
x −δ
Z 0+∞
n
e 2 (f
00 (x
00 (x
2
0 )+)(x−x0 )
2
0 )+)(x−x0 )
dx
dx
−∞
s
= (b − a)en(f (x0 )−η) + enf (x0 )
2π
n(−f 00 (x
0)
− )
00
where is chosen
q small enough so that (f (x0 ) + ) < 0 is still negative. Now divide both
sides by enf (x0 ) n(−f2π
00 (x )) which gives
0
Rb
a
enf (x0 )
nf (x)
e
q
dx
2π
n(−f 00 (x0 ))
≤
r
(b − a)e−nη
18
n(−f 00 (x0 ))
+
2π
s
−f 00 (x0 )
−f 00 (x0 ) −
!
.
(1.4)
Now using the other bound for
Rb
enf (x) dx
aq
enf (x0 ) n(−f2π
00 (x ))
0
from Equation 1.3 together with Equa-
tion 1.4 gives
s
R b nf (x)
Z +δ√n(−f 00 (x0 )+)
00
e
dx
−f (x0 )
1
− 21 y 2
a
√
q
e
dy
≤
√
2π
2π −f 00 (x0 ) + −δ n(−f 00 (x0 )+)
enf (x0 )
n(−f 00 (x0 ))
r
(b − a)e−nη
≤
Now we can take the limit as n → +∞ which gives
s
R b nf (x)
e
dx
−f 00 (x0 )
a
q
≤
lim
2π
−f 00 (x0 ) + n→∞ enf (x0 )
≤
n(−f 00 (x0 ))
s
n(−f 00 (x0 ))
+
2π
s
−f 00 (x0 )
−f 00 (x0 ) −
−f 00 (x0 )
−f 00 (x0 ) −
√
R +δ√n(−f 00 (x0 )+) − 1 y2
√
2
√
after noting that limn→0 e−nη n = 0 and limn→∞
2π. Since
e
dy
=
00
−δ
n(−f (x0 )+)
can be chosen to be arbitrarily small using the Sandwich Theorem gives
R b nf (x)
e
dx
a
= 1.
q
lim
n→∞
2π
nf
(x
)
0
e
n(−f 00 (x0 ))
Proof of Corollary 1.2. If x0 = a then the proof is the same but with a few adjustments.
In other words, the interval [x0 − δ, a] does not need to be considered as it is outside the
region of integration.
Z x0 +δ
Z x0 +δ
→
x0 −δ
x0 −δ
x0
Z
a
Z
b
→0
Z
→
x0 +δ
b
x0 +δ
and the resulting computation would give the extra factor of 21 after using
Z x0 +δ
Z
n
1 x0 +δ n (f (x0 )±)
(f
(x
)±)
0
e2
e2
dy =
dy.
2 x0 −δ
x0
A similar argument holds for x0 = b. Note that the computation shows that only a small
neighbourhood of x0 is relevant asymptotically and that the average term is exponentially
small in n.
19
!
.
1.2
One Dimensional Potential
The potential we are interested in is
V0 =
x2
x4
−a
4
2
where a > 0. When this is given a driving frequency it is1
Vt = V0 − F x cos Ωt
x4
x2
=
− a − F x cos Ωt
4
2
√
which when the forcing is zero, the potential has two wells at x0 = ± a. The SDE we
want to study is
dx
dw
= −∇Vt +
dt
dt
2
dx = x(a − x ) + F cos Ωt dt + dw
where w is a one dimensional Wiener process. Consider a realisation of the trajectory x(t)
starting at x(0) = y. Its escape time from the left well τ1 and right well τ2 are defined as
τ1 (y) = inf{t : x(t) = 0 and x(0) = y} where y ∈ (−∞, 0)
τ2 (y) = inf{t : x(t) = 0 and x(0) = y} where y ∈ (0, +∞).
(1.5)
(1.6)
Note that the trajectory x(t) is related to the escape times τ1 and τ2 . Define a new quantity
by
fni (y) = hτi (y)n i
with i = 1, 2 for the two wells and n = 1, 2, . . .. Note h·i denotes the mean average over
all realisations. Also note that the nth moment is being used here. In [4] a method by
Gihman and Skorohod [53] was used to derive the following equation
d
1 2 d2 i
i
fn (y) − V 0 fni (y) = −nfn−1
(y)
2
2 dy
dy
(1.7)
where V 0 is a shorthand for V 0 = ∇V0 if we are escaping from the potential described by
V0 . Note that the potential is frozen in the case of V0 . Similarly V 0 is a shorthand for
V 0 = ∇Vt if we are escaping from the potential described by Vt . The following boundary
conditions are
fni (0) = 0,
1
d 1
f (−∞) = 0 and
dy n
d 2
f (+∞) = 0.
dy n
See Appendix A for a full explanation of the notation used for the potentials.
20
Having fni (0) = 0 is appropriate since being at y = 0 means it is in neither well and so has
d 1
d 2
already escaped at t = 0 anyway. We can see how dy
fn (−∞) = 0 and dy
fn (+∞) = 0 make
2
example. If the particle starts at x(0) = y, where y is a very
sense by considering f1 as an √
large positive number y N a (where N 1) then with the equilibrium
point being an
√
attractor, it would more or less deterministically
slide
towards
x
=
+
a.
We
call the time
√
0
further beyond y
it takes for it to travel to x = + a, τ . If the particle starts somewhere
√
say x = y + δ, where δ > 0 , it would also√slide down to x = + a almost deterministically.
We call this new time to get to x = + a, τ 00 . Intuitively, we would expect τ 0 ≈ τ 00 so
d 2
f (−∞) = 0.
dy 1
The aim now is to solve Eqn 1.7 for different cases. These are for F = 0 and F 6= 0, in
the small noise approximation.
1.2.1
One Dimensional Potential - Case F = 0
The potential is stationary and does not depend on time. It is symmetric at x = 0 so we
must have
f11 (−y) = f12 (y).
For simplicity we denote the following
f = f11
and I =
df
dy
which rewrites the differential equation as
1 2 dI
− V00 I = −1
2 dy
where V00 = ∇V0 . Using an integrating factor gives
Z y
Z y
2 0
2 0
2
d
− 2 V0 (s) ds
− 2 V0 (s) ds
I × exp
= − 2 exp
dy
0
0
which gives
d
dy
2
2
2
= − 2 exp − 2 V0 (y) .
I × exp − 2 V0 (y)
We know that I(−∞) = 0 so integrating we have
Z
2
2
2 y
2
I(y) × exp − 2 V0 (y) − I(−∞) × exp − 2 V0 (−∞) = − 2
exp − 2 V0 (s) ds
−∞
and proceeding we have
Z
2
2 y
2
I(y) × exp − 2 V0 (y) = − 2
exp − 2 V0 (s) ds
−∞
21
Z y
2
2
2
I(y) = − 2 exp + 2 V0 (y)
exp − 2 V0 (s) ds.
−∞
We know that f (0) = 0 so integrating we have
Z u
Z
2
2
2 y
exp + 2 V0 (u)
exp − 2 V0 (s) ds du
f (y) − f (0) = − 2
0
−∞
which we rewrite as
Z
2
2 y
exp + 2 V0 (u) g(u) du
f (y) = − 2
0
Z u
2
exp − 2 V0 (s) ds.
where g(u) =
−∞
(1.8)
Up to now the methods we have used for solving f (·) are exact and the boundary conditions
on f (·) have also been kept. There are no approximations to our approach√so far. Recall
that y < 0. We seek an approximate solution for f (y) in the region y ∈ [− a, 0]. Now we
use Laplace Method in the small noise limit (small ) to evaluate the integrals in Equation
1.8. Note that
√
2
2
max
− 2 V0 (u) = − 2 V0 − a .
u∈[−∞,0]
Using the Laplace Method for small gives the approximation
q
n 2o
√
π2
a
exp
if u ∈ (− a, 0]
2a
22
q
g(u) ≈
n o
√
1 π2 exp a22
if u = − a
2
2a
2
q
n 2o
2
a
This means g(u) ≈ π
almost everywhere with respect to the Lebesgue meaexp
2a
22
√
sure on [− a, 0]. This approximates f (·) to
r
2 Z 0
2 π2
a
2
f (y) ≈ + 2
exp
exp + 2 V0 (s) ds
2a
22
y
where we have switched the limits of the integral. We use Laplace Method again after
noting that
2
2
max
+ 2 V0 (s) = + 2 V0 (0)
√
s∈[− a,0]
here the maximum is on the edge on the boundary meaning we would need an extra factor
of 12 . So
r
2 r 2
2 π2
a
1 π
f (y) ≈ + 2
exp
2
2a
2
2
a
2
1 π
a
=√
exp
.
22
2a
22
1.2.2
One Dimensional Potential - Case F 6= 0
The potential is now oscillating. We aim to do a similar calculation to the static potential
case. We make an approximation by assuming that the amplitude of the oscillations F is
small enough such that there will always be two distinct wells. The positions of the critical
points and wells will be very close to the static potential case. We can just focus on one
well x0 (t) on the left of the hill which is dependent on the time t. The method is similar
to what we have used in the F = 0 case. We have
Z u
Z
2
2
2 y
exp + 2 Vt (u)
exp − 2 Vt (s) ds du
(1.9)
f (y) = − 2
0
−∞
Z
2
2 0
exp + 2 Vt (u) gt (u) du
=+ 2
y
Z u
2
exp − 2 Vt (s) ds
where gt (u) =
−∞
since y < 0 the limits of the integral may be switched. Notice we have approximated the
situation by assuming that the hill moves very little away from x = 0 which is what makes
Equation 1.9 valid. We seek a solution for f (y) in the region y ∈ [x0 (t), 0]. For small , we
can use Laplace’s Method to approximate the integrals in Equation 1.9. Note that
2
2
max
− 2 Vt (u) = − 2 Vt (x0 (t)) .
u∈[−∞,0]
Now using the Laplace’s Method gives
q
π2
exp − 22 Vt (x0 (t))
if u ∈ (x0 (t), 0]
Vt00 (x0 (t))
q
gt (u) ≈
π2
1 exp − 2 V (x (t))
if u = x0 (t)
0
2
2 t
V 00 (x0 (t))
t
where Vt00 = ∇2 Vt . In other words gt ≈ exp − 22 Vt (x0 (t)) almost everywhere on [x0 (t), 0]
with respect to the Lebesgue measure. So
s
Z 0
2
2
2
π2
exp + 2 Vt (u) du.
f (y) ≈ + 2 exp − 2 Vt (x0 (t))
Vt00 (x0 (t)) y
Now
max
u∈[x0 (t),0]
2
2
+ 2 Vt (u) = + 2 Vt (0)
where the maximum is on the boundary of [x0 (t), 0] meaning we would need an extra factor
of 12 . So
s
s
π2
1
π2
2
2
f (y) ≈ + 2 exp − 2 Vt (x0 (t))
Vt00 (x0 (t)) 2 −Vt00 (0)
23
which gives
2
exp − 2 Vt (x0 (t)) .
f (y) ≈ p 00
aVt (x0 (t))
π
Finding f (y) is very hard so we consider when the oscillations are very slow, that is for
very small Ω. At the two extremes we have
dx = [x(a − x2 ) + F ]dt + dw when t = 0
π
dx = [x(a − x2 ) − F ]dt + dw when t = .
Ω
(1.10)
(1.11)
We also assume that the oscillations are so small, the now time dependent
equilibrium
√
point does not differ much from the time independent case x0 = − a. We solve f (y) for
the case of Equation 1.10. The case for Equation 1.11 is similar. Let x0 (t) be approximated
and denoted with a new notation by
x0 (t) = z0 + δ = s
√
where z0 = − a. We seek an expression for δ by
[s(a − s2 ) + F ]dt = 0 from Equation 1.10
s(a − s2 ) = −F
δ≈
−F
F
=+
2
a − 3z0
2a
after ignoring terms of higher order than δ 2 . Progressing further gives
2
2 s4
s2
− 2 Vt=0 (s) = − 2
− a − Fs
4
2
≈−
2
3
V
(z
)
+
δ(z
−
az
−
F
)
−
F
z
0
0
0
0
0
2
(1.12)
again after ignoring terms of higher order than δ 2 . Equation 1.12 is now approximated by
2
2
− 2 Vt=0 (s) ≈ − 2 V0 (z0 ) + δ(z03 − az0 − F ) − F z0
2
a
4F 2 4F
= 2 1+ 3 − 3
2
2a
a2
a2
4F
≈ 2 1− 3
2
a2
after assuming F 2 is small. We make another approximation by
p
p
00
aVt=0
(x0 (t)) ≈ aV000 (z0 )
24
q
a[3z02 − a]
√
=a 2
=
√
after noting that z0 = − a. So f (y) for Equation 1.10 and 1.11 are
2
a
π
4F
f (x0 (t)) = √ exp
1− 3
when t = 0
22
a 2
a2
2
π
a
4F
π
f (x0 (t)) = √ exp
1
+
when
t
=
.
3
22
Ω
a 2
a2
We can see how the solution make physical sense because when t = Ωπ the left well is lower,
and so the probability to escape is lower and the time to escape would also be longer. Now
that all of our calculations are done for both the time independent and time dependent
case we can compare them.
1.2.3
Conclusion and Resonance Condition res
We have effectively reviewed f11 in the limit of small noise, which is hτ i the averaged escape
time for small . Comparing them more clearly here gives
2
a
1 π
exp
F = 0 hτ i = √
22
2a
2
π
a
4F
F 6= 0 hτ i = √ exp
when t = 0
1− 3
22
a 2
a2
2
π
a
π
4F
hτ i = √ exp
when t = .
1+ 3
2
2
Ω
a 2
a2
For the F 6= 0 case if we impose
π
Ω
π
hτ i =
Ω
hτ i =
for t = 0
for t =
(1.13)
π
Ω
(1.14)
and solve for the noise in both cases (that is solving Equation 1.13 and 1.14 for ) we have
1 = a
2 = a
1 − 4F/a3/2
√
2 ln(2 2a/Ω)
1/2
1 + 4F/a3/2
√
2 ln(2 2a/Ω)
1/2
25
for t = 0
for t =
π
Ω
then the resonance condition res should be inside the interval res ∈ [1 , 2 ]. For the
example below it just so happen that [1 , 2 ] = [0.18, 0.31] and res ≈ 0.26, which gives the
trajectory
Figure 1.1: This trajectory is exhibiting quasi-determinism. It is near stochastic resonance.
We can increase and decrease the noise away from res ≈ 0.26, and transitions will occur
more frequently or less frequently as we move away from resonance.
Figure 1.2: Transitions occur irregularly and are rare.
26
Figure 1.3: Transitions occur very often.
1.3
Remarks on One Dimensional Potential
Notice that this is a very crude way to study the system. The oscillating potential is being
approximated by a frozen static potential. This is precisely the adiabatic approximation.
Not only does it assume small noise, small forcing and adiabatic time development, the
exact positions of the critical points (the two wells and the hill) were not calculated. All
the calculations assumed that the hill was near x = 0. This means the forcing is assumed
to be small enough such that the hill does not move far away from x = 0. Benzi et al’s
definition of the escape time is so crude it will be used only as a rough guide.
Also note it is hard to tell if Figures 1.1, 1.2 and 1.3 show any regularity or not. As we
shall see, regularity shows itself in the distribution of the escape times.
27
Chapter 2
Theoretical Escape Time from a Well
of a Static Potential
We consider the theoretical escape time of a particle from a well of a static potential. This
is done in two parts. The first part is Freidlin-Wentzell theory or large deviation and the
second part is Kramers’ formula which is derived using potential theory. For small noise
levels large deviation is considered and for higher noise levels potential theory is used.
2.1
Freidlin-Wentzell Theory and Large Deviation
A review of the major results of the Freidlin-Wentzell theory is presented which is found
in [32]. This is done by considering stochastic systems converging to the deterministic
limit for small noise, action functional for Wiener processes, action functional for general
processes and the main theorems concerning the escape time.
2.1.1
Stochastic Processes
Let (Ω, F, P ) be a probability space. Let (Rr , B) be a measure space on Rr . Let T be an
indexing set. For ω ∈ Ω and t ∈ T define a mapping Ω × T −→ Rr by
Xt (ω) : Ω × T −→ Rr
where Xt (ω) ∈ Rr is called a stochastic process on Rr . The probability measure defined on
A ∈ F is denoted by P (A). But if this probability measure can depend on a value x ∈ Rr ,
we will often put this dependence explicitly into the notation
P (A, x) = Px (A).
We will only consider Markov processes in this thesis. There are further technical properties
a Markov process has to fulfil.2 Intuitively this can be understood in the following way;
2
For more details see page 20 in [32].
28
the Ω can be thought of as the set of all trajectories of a stochastic process; the T can be
thought of as the set of time, for example T = [0, ∞); the Rr can be thought as the space
in which the trajectory is in, then Xt (ω) is a trajectory in Rr with continuous time t ≥ 0.
2.1.2
Deterministic Limit
Consider the following system. We have the r-dimensional real space Rr . Let xt ∈ Rr be
a time dependent variable in Rr . Let b(xt ) be a function b : Rr → Rr on xt . We then let
dxt = b(xt )dt
(2.1)
which can be seen as a system of r differential equations for each of the elements of xt .
When we consider random systems we denote the (random) variable by Xt with values
in Rr . The stochastic processes we consider in this thesis are diffusion processes. More
precisely we consider random dynamical systems which are solutions of the following system
of stochastic differential equations
dXt = b(Xt )dt + σ(Xt )dwt
(2.2)
where is the noise level, wt is a l-dimensional Wiener process and σ(Xt ) is a function on
Xt returning a l × r matrix.
The first circle of results in the book of Freidlin and Wentzell are about how the solutions
of the random system Equation 2.2 approximate the solutions of the deterministic system
Equation 2.1. For example we know that
lim Xt = xt .
→0
But the exact manner of this limit and the conditions under which Xt → xt is reached is
documented in Freidlin-Wentzell.3 For example we have4
Theorem 2.1. Suppose that the coefficients of Equation 2.2 satisfy a Lipschitz condition
and a growth condition given by
X
X
[bi (x) − bi (y)]2 +
[σij (x) − σij (y)]2 ≤ K 2 |x − y|2
(2.3)
i
i,j
X
[bi (x)]2 +
i
X
[σij (x)]2 ≤ K 2 (1 + |x|2 )
(2.4)
i,j
then for all t > 0 and δ > 0 we have
E
|Xt
2
2
− xt | ≤ a(t) and
lim P
→0
max
0≤s≤t
|Xs
− xs | > δ
=0
where a(t) is a monotone increasing function, which is expressed in terms of |x| and K.
3
4
See pages 44-59 in [32].
Adapted from Theorem 1.2 page 45 of [32].
29
Theorem 2.1 can be explained in another way. Intuitively as → 0 we would expect to be
back in the deterministic system xt . Note that Equation 2.3 is the Lipschitz condition and
2.4 is the growth condition.
There is also a stochastic analogue of Taylor’s Remainder’s Theorem where it can be
shown that Xt admits the following decomposition5
(0)
(1)
(k)
Xt = Xt + Xt + · · · + k Xt
+ Rk+1
(t)
where the remainder is bounded by new functions
sup Rk+1
(t) < C(ω)k+1
and P {C(ω) < ∞} = 1
0≤t≤T
(k)
and the Xt
2.1.3
are solutions of stochastic differential equations.
Action Functional for Wiener processes
The second part of Freidlin-Wentzell has a new setting.6 Let b(Xt ) = 0, σ(Xt ) = 1 and wt
be a r-dimensional Wiener process, that is to say
dXt = dwt
where we have reduced the system to a r-dimensional Wiener process. Let CT1 T2 =
CT1 T2 (Rr ) denote the set of all continuous paths in Rr starting at time T1 and ending
at T2 . On this set we define a metric by
ρT1 T2 (ψ, ϕ) = sup |ψt − ϕt |.
T1 ≤t≤T2
We define a new functional by
1
S(ϕ) = ST1 T2 (ϕ) =
2
Z
T2
|ϕ̇s |2 ds
T1
for absolutely continuous (and differentiable) ϕt . If ϕt is not absolutely continuous or if
the integral is divergent, we set S(ϕ) = +∞. We define the action functional by
IT 1 T2 (ϕ) = −2 ST1 T2 (ϕ)
and ST1 T2 (ϕ) will be called the normalized action functional. The paths should be interpreted as points, that is elements of the functional space of paths, that is each point is
itself a path. The distance between these points, and hence paths, is given by the metric
just defined. We define a new set
Φ(s) = {ϕ ∈ C0T such that ϕ0 = 0 and S0T (ϕ) ≤ s}
5
6
See Theorem 2.1 page 52 of [32].
See pages 70-79 in [32].
30
which can be shown to be compact in the uniform topology.7 Now the SDE
dXt = dwt
cannot be solved pathwise as in the deterministic case. The solution is a randomly chosen
path out of an infinitude of possible paths. This is described by the probability of the
path having certain properties. Also Xt is self-similar and non-differentiable, but may be
approximated by differentiable functions ϕt . The next major theorems in Freidlin-Wentzell
show that for any δ > 0 and γ > 0 we have8
P {ρ0T (Xt , ϕt ) < δ} ≥ exp −−2 [S0T (ϕ) + γ]
and for any δ > 0, γ > 0, s0 > 0 with s < s0 we have9
P {ρ0T (Xt , Φ(s)) ≥ δ} ≤ exp −−2 (s − γ) .
These two statements may be interpreted as a Laplace type theorem in function spaces.
A physical interpretation is that this gives an asymptotic description (in small ) for the
probability that the path Xt is near to ϕt , that is
P {ρ(Xt , ϕ) < δ} ≈ exp −−2 S(ϕ) .
In the next section we develop the action functional for more general processes.
2.1.4
Action Functional for General processes
So far the action theory was developed for just one particular example of a stochastic process, that is the Wiener process. Now we develop an action theory for a general stochastic
process described by Equation 2.2.10
Before we do that we state a list of properties a functional should have so that we can
consider a suitable action functional. We state these properties in a more general context.
Let (X, ρ) be a metric space with metric ρ. On the σ-algebra of its Borel subsets let µh be
a family of probability measures depending on a parameter h > 0. Let λ(h) be a positive
function going to +∞ as h ↓ 0. Let S(x) be a function such that S : X → [0, ∞]. We say
that λ(h)S(x) is an action function if the following holds.
(0) the set Φ(s) = {x : S(x) ≤ s} is compact for every s ≥ 0.
(I) for any δ > 0, any γ > 0 and any x ∈ X there exists an h0 > 0 such that
µh {y : ρ(x, y) < δ} ≥ exp{−λ(h)[S(x) + γ]}
7
See
See
9
See
10
See
8
Lemma 2.1 page 77 of [32].
Theorem 2.1 page 74 in [32].
Theorem 2.2 page 74 in [32].
pages 79-92 of [32].
31
for all h ≤ h0
(II) for any δ > 0, any γ > 0 and any s > 0 there exists an h0 > 0 such that
µh {y : ρ(y, Φ(s)) ≥ δ} ≤ exp{−λ(h)(s − γ)}
for all h ≤ h0 .
S(x) and λ(h) will be called the normalized action functional and normalizing coefficient.
The results given in Chapter 2.1.3 show that the functional considered there has all
the above properties, where X = CT1 T2 (Rr ), S(ϕ) = ST1 T2 (ϕ), λ(h) = −2 and = h.
Thus we can see how X = CT1 T2 (Rr ) with IT 1 T2 is an action functional since it satisfied all
three properties. Doubtless that there will be many other systems which satisfy all these
three properties as well. This higher level of abstraction would allow us to prove powerful
theorems.
2.1.5
Main Theorems
The action functional was given for a diffusion with drift term zero i.e. b(Xt ) = 0. We
now put this term back in to consider the equation
dXt = b(Xt )dt + dwt
where wt is a r-dimensional Wiener process. It can be shown that letting11
Z
1 T
|ϕ̇s − b(ϕs )|2 ds
S0T (ϕ) =
2 0
with λ(h) = −2 and h = satisfy the three properties of the action functional. The action
functional then allows us to compute asymptotically different probabilities. For example,
let D ⊂ Rr be a region of space in Rr and let
HD (t, x) = {ϕ ∈ C0T (Rr ) : ϕ0 = x, ϕt ∈ D ∪ ∂D}
H D (t, x) = {ϕ ∈ C0T (Rr ) : ϕ0 = x, xs ∈
/ D for some s ∈ [0, t]}
then it can be shown that12
lim 2 ln Px {Xt ∈ D} = −
→0
lim 2 ln Px {τ ≤ t} = −
→0
min
ϕ∈HD (t,x)
min
S0T (ϕ)
(2.5)
S0T (ϕ)
(2.6)
ϕ∈H D (t,x)
where τ = min{t : Xt ∈
/ D} is the escape time from D. This theorem gives us the leading
term for the probabilities leaving this region of space D. The HD is the set of all paths
11
12
See Theorem 1.1 page 104 of [32].
See Theorem 1.2 page 105 of [32].
32
that stay in D and its boundary. The H D is the set of all paths that leave D at some time.
Equation 2.5 is thus the probability of remaining in D and Equation 2.6 is the probability
of the escape time being less than t.
So far the above results hold for a general region D. Now we want to consider the case
where D is the vicinity of a well, that is the region near and around a metastable state.
This means D is attracted to a point inside of D. Without loss of generality we can choose
this point, which is the position of the well, to be zero xwell = 0. But the minimiser of
S0T (ϕ) is very difficult to compute explicitly using the usual differential equations. Let
f (t, x, y) = ϕmin
S (ϕ).
=x 0t
0
ϕt =y
The Hamilton-Jacobi equations are given by
∂f (t, x, y) 1
+ |∇y f (t, x, y)|2 + (b(y), ∇y f (t, x, y)) = 0
∂t
2
(2.7)
where ∇y is the gradient operator in the variable y and note that
min
ϕ∈HD (t,x)
min
ϕ∈H D (t,x)
S0t (ϕ) = min f (t, x, y)
y∈D∪∂D
S0t (ϕ) = min f (s, x, y)
0≤s≤t
y ∈D
/
and the solution to Equation 2.7 would be closely related to Equation 2.5 and 2.6.13 Let
us introduce the so-called quasipotential
Ṽ (x, y) = inf{ST1 T2 (ϕ) : ϕ ∈ CT1 T2 (Rr ), ϕT1 = x, ϕT2 = y}
which is the least action over all paths which starts at x and ends at y. Suppose that the
drift term can be written as the gradient of a potential V
b(x) = −∇V (x)
then it can be shown that14
Ṽ (0, x) = 2V (x).
(2.8)
Note that this only holds for points x ∈ Rr such that V (x) ≤ miny∈∂D V (y), that is for
points lower than the exit point. Now suppose that there exists a unique point y0 ∈ ∂D
for which Ṽ (0, y0 ) = miny∈∂D Ṽ (0, y) then15
lim Px {ρ(Xs , y0 ) < δ} = 1 where s = inf {t : Xt ∈ ∂D}
→0
(2.9)
for every δ > 0 and any x ∈ D. This means Xt will exit near points of least height in the
small noise limit.
13
See pages 105-108 of [32].
See Theorem 3.1 page 118 of [32].
15
See Theorem 2.1 page 108 of [32].
14
33
2.1.6
Remarks on Freidlin-Wentzell and Large Deviation Theory
The main Theorems of Freidlin-Wentzell were developed on a precise and rigorous mathematical setting. It would be appropriate to interpret what they mean in a more physical
setting. Consider Equation 2.8 and 2.9. Equation 2.8 gives an easier way to calculate the
quasipotential, because the quasipotential is related in a very simple way to the height of
the potential. Equation 2.9 means that the particle will escape whilst travelling through
a path which gives the least height, or interpreted in another way, a path of least action.
Thus, one of the main conclusion of the Fredlin-Wentzell theory is that the particle will
tend to escape following close to a path which gives the least distance to climb out of a
well.
2.2
Kramers’ Formula and Potential Theory
Let V : Rr −→ R. Let x ∈ Rr be a well and zi ∈ Rr be saddles labelled by i = 1 . . . n. The
saddles would be gateways providing a passage for escape from the well. Define
∆Vi = V (zi ) − V (x)
which is the height difference between the well and the ith saddle. For small noise , an
approximate expression can be estimated for the escape time of the particle going through
the ith saddle. In the smallest order of the noise the mean exit time, as described in the
previous section, is given by16
2
τi = e+2∆Vi / .
Inverting this gives the escape rate
2
Ri = e−2∆Vi /
and the total escape rate would be to sum over all the saddles
R=
n
X
i=1
Ri =
n
X
2
e−2∆Vi / .
i=1
The order correction is done by adding a coefficient called Kramers’ coefficient and the
resulting corrected rate is called Kramers’ rate
2
Ri = ki e−2∆Vi /
where
p
|∇2 V (x)|
|λ(zi )|
p
ki =
2π
k∇2 V (zi )k
16
See Theorem 4.1 and 4.2 on pages 124-127 of [32].
34
where |∇2 (x)| denotes the determinant of the Hessian of the potential at the well x,
k∇2 V (zi )k denotes the modulus of the determinant of the potential at the saddle zi and
|λ(zi )| denotes the minimum eigenvalue of the Hessian of the potential at the saddle zi .
This gives the escape rate in the next order of approximation to be
R=
n
X
i=1
Ri =
n
X
2
ki e−2∆Vi /
i=1
which is rewritten as
p
n
|∇2 V (x)| X
−2 (V (zi ) − V (x))
|λ(zi )|
p
exp
.
R=
2
2 V (z )k
2π
k∇
i
i=1
The last order of approximation for higher noise is done by bounding the error on Kramers’
coefficient. This is
!
p
|∇2 V (x)|
1
|λ(zi )|
p
ki =
2
2
2π
k∇2 V (zi )k 1 + O 2 ln 2
which is rewritten as
1
2π
=p
ki
|∇2 V (x)|
p
2
k∇2 V (zi )k
2
1+O
ln
.
|λ(zi )|
2
2
(2.10)
We conclude with a few words on the derivation of Kramers’ formula and the bound on its
error. Equation 2.10 was derived rigorously using techniques from potential theory instead
of large deviations. Part of the technique involves the escape time being expressed in terms
of a partial differential equation, similar to Equation 1.7 for example. This derivation was
done in [37] which is beyond the scope of this thesis. This Chapter reviewed the escape
rates of a particle from a static well, which will be relevant when we consider escape rates
from an oscillating well.
35
Chapter 3
Theoretical Escape Time from a Well
of an Oscillatory Potential
Stochastic resonance usually involves studying transitions between two stable states. Studying a stochastic differential equation in multidimensional real space can be complicated. It
would be useful to simplify stochastic resonance down to a Markov Chain with transitions
between two states +1 and −1, then we try to model stochastic resonance with a two state
Markov Chain. This is done for both discrete and continuous time Markov Chains with
two states, for both alternating and synchronised saddles.
3.1
Markov Chain Reduction
Let V : Rr −→ R be a potential with two wells. This potential is subjected to a periodic
forcing F with frequency Ω and perturbed by noise , which is described by the SDE17
Ẋt = −∇V + F cos(Ωt) + Ẇt
(3.1)
where Wt is a Wiener process in Rr and F ∈ Rr . We call Xt the diffusion case. The Xt
can be reduced to a Markov Chain Yt on {−1, +1}
Xt −→ Yt
in the following way. In what follows we will assume that the diffusion Xt is continuous
in time and space. Let wl (t) denote the position of the left well at time t and wr (t) the
position of the right well at time t. Note that wl (t) and wr (t) are also continuous in time.
Let R ∈ R be constant. The reduction from the Xt to the Markov Chain Yt is
−1 if |Xt − wl (t)| ≤ R
+1 if |Xt − wr (t)| ≤ R
Yt =
Z if |Xt − wl (t)| > R and |Xt − wr (t)| > R
17
See Appendix A for how the forcing is denoted.
36
where Z is given by
(
Z=
−1 if s2 < s1
+1 if s1 < s2
where s1 and s2 are given by
s1 = max {u : |Xu − wl (u)| ≤ R}
u<t
s2 = max {u : |Xu − wr (u)| ≤ R} .
u<t
When Yt = −1 we say the particle is in the left well and when Yt = +1 we say the particle
is in the right well. Only when it enters the other well would Yt change sign. When the
condition |Xu − wl (u)| ≤ R is satisfied we say the particle is covered by the left well. When
the condition |Xu − wr (u)| ≤ R is satisfied we say the particle is covered by the right well.
Note that R is chosen small enough such that it is impossible for the particle to be covered
by both wells at any time, that is
{x ∈ Rr : |Xt − wl (t)| < R and
|Xt − wr (t)| < R} = ∅
for all times t ≥ 0. This means that s1 is the most recent time the particle is covered by
the left well and s2 is the most recent time the particle is covered by the right well. Notice
that if initially at t = 0, the particle is covered by neither well then Yt cannot be derived
nor defined by the above definitions. In this case either Y0 = −1 or Y0 = +1 is chosen
depending on what initial conditions are required. In other words if Y0 = −1 is chosen as
the initial condition then the particle is covered by the left well for t < 0. If Y0 = +1 is
chosen as the initial condition then the particle is covered by the right well for t < 0.
The escape time from the left to right well τ−1+1 and from the right to left well τ+1−1
are defined in the following way18
τ−1+1 = µ ({t : Yt = −1})
τ+1−1 = µ ({t : Yt = +1})
where
where
{t : Yt = −1}
{t : Yt = +1}
is an interval
is an interval
where µ denotes the Lebesgue measure. In other words the time spent being in the state
Yt = −1 is τ−1+1 and the time spent being in the state Yt = +1 is τ+1−1 . These intervals
will always be closed intervals. The process Yt has two states, hence each sample is a
piecewise constant function. The length of each piece is the escape time τ−1+1 or τ+1−1 .
Note that τ−1+1 and τ+1−1 are random times and random variables.
For the diffusion the escape time can be explained in the following way. Each well is
surrounded by a circle with a constant radius R which moves with the well. A particle
is said to have entered the left well if it enters the region covered by the radius R over
the left well. The particle is then said to have entered the right well when it enters the
region covered by R in the right well. The time difference between entering the left well
18
See Appendix B.1 for details of the actual use of R in the measurement of the escape times.
37
and entering the right well is defined to be the escape time from left to right τ−1+1 . A
similar argument is said for τ+1−1 . This also means the escape times in the Markov Chain
is the same as the diffusion trajectory Xt by definition. Notice that the diffusion Xt has
to be defined first before the Markov Chain Yt which is a derived quantity.
Notice that all of our reasoning in deriving Yt only assumes that Xt is a continuous time
process in Rr . We did not check whether Yt satisfy the strict definitions of a continuous
time Markov Chain. If Yt is a Markov Chain it should also satisfy the Markov property,
that is
P Yt = i|Yt1 = −1, Yt2 = +1 = P Yt = i|Yt1 = −1
= P Yt = i|Yt2 = +1
for any 0 ≤ t1 < t2 < t. Again we stress that the only assumption we made when deriving
Yt from Xt is that Xt is a continuous time process in Rr , which is not sufficient for Yt to
be a Markov Chain nor for Yt to satisfy the Markov property. But throughout the rest of
this thesis the diffusion Xt will be a Markov process, which means Yt should be a good
approximation to a Markov Chain.19
A discrete time and continuous time Markov Chain model for Equation 3.1 are studied
in the following sections.
3.2
Discrete Time Markov Chain
Let the time be discrete. This to say time t belongs to
t ∈ {0, 1, 2, . . .} .
The Markov Chain is a time dependent stochastic process which can take values +1 or −1
Yt = ±1.
At time t the probability of Yt jumping from −1 to +1 is denoted by p−1+1 (t); the probability of jumping from +1 to −1 is denoted by p+1−1 (t); the probability of staying in −1 is
denoted by p−1−1 (t); and the probability of staying in +1 is denoted by p+1+1 (t). Notice
that they have the following properties for all time t
p−1−1 (t) + p−1+1 (t) = 1
p+1+1 (t) + p+1−1 (t) = 1.
A transition matrix can be defined as
Pt :=
p−1−1 (t) p−1+1 (t)
p+1−1 (t) p+1+1 (t)
19
!
.
Whether Yt is a Markov Chain for a Markov process Xt , or for Xt described by an SDE requires
proof. This is an open question.
38
At every point in time it is possible to define a state probability, that is the probability of
the trajectory being −1 or +1,
P (Yt = −1) = ν− (t)
P (Yt = +1) = ν+ (t).
Notice that the state probability satisfy the following condition for all time t
ν− (t) + ν+ (t) = 1.
The two ν− (t) and ν+ (t) can be written compactly in vector notation
!
ν− (t)
ν(t) =
.
ν+ (t)
The state probability at time t + 1 can be expressed in terms of the last time t, that is
ν(t + 1) = Pt† ν(t)
ν− (t + 1)
ν+ (t + 1)
=
p−1−1 (t) p+1−1 (t)
!
p−1+1 (t) p+1+1 (t)
ν− (t)
!
ν+ (t)
where Pt† denote the transpose of the matrix Pt . This means if the initial value of the state
probability is known at t = 0, then the future behaviour of the state probability can be
described by computing all subsequent values of ν(t), that is
ν(t) =
i=t−1
Y
Pi† ν(0).
i=0
The main aim for the rest of our studies of the Markov Chain is to compute the state
probability for various transition matrices. When the wells of the potential are oscillating
such that one well is higher than the other, we model using p 6= q. When the wells of the
potential are oscillating such that both wells are always at the same height as each other,
we model using p = q.
3.2.1
Discrete Time Markov Chain - Alternating Saddles p 6= q
We want to study a system with periodic elements. The transition matrix would change
periodically in time. Let the period be
T = 2m
where m is an integer. Let the time t be written in the form
t = NT + n
39
where N is an integer number of periods. The transition matrix would vary periodically
according to
!
1−p
p
For n = mod(t, T ) ∈ T1 Pt = P1 =
q
1−q
!
1−q
q
For n = mod(t, T ) ∈ T2 Pt = P2 =
p
1−p
where
T1 = {0, 1, . . . , m − 1}
T2 = {m, m + 1, . . . , 2m − 1} .
We interpret Yt = −1 as being in the left well and Yt = +1 as being in the right well. We
also interpret p as the probability of escape from a shallow well and q as the probability of
escape from a deep well. The transition matrix varying periodically in time can be used to
model the periodic forcing being applied to the potential. The following Theorem derives
the state probabilities.
Theorem 3.1. Let the time be t = N T + n. Let λ = 1 − p − q. The state probability at
time t is
λn
p−q
1
q
−1
×
−
For n ∈ T1 ν =
1
p+q p
p + q 1 + λm
ν+ (0)(p + qλm ) − ν− (0)(q + pλm ) 2mN +n −1
+
λ
1
1 + λm
λn−m
1
p−q
p
−1
×
For n ∈ T2 ν =
+
1
p+q q
p + q 1 + λm
ν+ (0)(p + qλm ) − ν− (0)(q + pλm ) 2mN +n −1
λ
+
1
1 + λm
Proof. Notice that the eigenvectors and eigenvalues of the transpose matrix P1† are
−1
λ1 = 1 − p − q
v1 =
1
q
λ2 = 1
v2 =
p
which also spans the R2 space. For short we call λ = λ1 . This means an arbitrary vector
can be expressed as a linear combination of the eigenvectors of P1† . This is
1
x
−1
q
=
(qy − px)
+ (x + y)
.
y
1
p
p+q
40
Now consider m application of the P1† matrix on the arbitrary vector.
1
−1
q
x
† m
m
(qy − px)λ
+ (x + y)
(P1 )
=
1
p
y
p+q
1
x
q + pλm q(1 − λm )
=
m
m
y
p + q p(1 − λ ) p + qλ
†
1 − p0
p0
x
=
q0
1 − q0
y
where
p0 =
p(1 − λm )
p+q
and q 0 =
q(1 − λm )
p+q
with a similar expression for the other transition matrix
†
x
1 − q0
q0
x
† m
(P2 )
=
.
0
0
y
p
1−p
y
Now denote a new matrix by
Ptot = (P2† )m (P1† )m
1 − p0
q0
1 − q0
p0
=
p0
1 − q0
q0
1 − p0
where Ptot has eigenvalues and eigenvectors
1 − q0
e1 =
1 − p0
−1
e2 =
1
ξ1 = 1
ξ2 = (1 − p0 − q 0 )2
and these eigenvectors span the R2 space
1
x
1 − q0
−1
0
0
(x + y)
+ [y(1 − q ) − x(1 − p )]
=
1 − p0
1
y
1 + λ0
where we have denoted
λ0 = 1 − p0 − q 0 .
Now consider N applications of the matrix Ptot on the initial value of the state probability
1
1 − q0
−1
N
0
0
02N
Ptot ν(0) =
+ [ν+ (0)(1 − q ) − ν− (0)(1 − p )] λ
1 − p0
1
1 + λ0
41
and express the results in terms of the eigenvectors of P1†
1
1
−1
q
N
0
0
0
Ptot ν(0) =
[q(1 − p ) − p(1 − q )]
+ (1 + λ )
×
0
1
p
1+λ
p+q
[ν+ (0)(1 − q 0 ) − ν− (0)(1 − p0 )] λ02N
−1
.
+
0
1
1+λ
If n ∈ T1 consider
N
(P1† )n Ptot
−1
q
0
0
0
n
+ (1 + λ )
[q(1 − p ) − p(1 − q )] λ
1
p
[ν+ (0)(1 − q 0 ) − ν− (0)(1 − p0 )] λ02N n −1
+
λ
1
1 + λ0
1
p−q
λn
q
−1
=
−
×
1
p+q p
p + q 1 + λm
ν+ (0)(p + qλm ) − ν− (0)(q + pλm ) 2mN +n −1
λ
.
+
1
1 + λm
1
1
ν(0) =
×
0
1+λ
p+q
If n ∈ T2 consider
N
(P2† )n−m (P1† )m Ptot
ν(0) =
(P2† )n−m
1
p+q
q
p
p−q
λm
−
×
p + q 1 + λm
ν+ (0)(p + qλm ) − ν− (0)(q + pλm ) 2mN +m
λ
+
1 + λm
−1
1
−1
1
and we express (q, p)† in terms of the eigenvectors of P2†
λm
1
p−q
−1
p
−1
† n−m
† m N
† n−m p − q
(P2 )
×
(P1 ) Ptot ν(0) = (P2 )
+
−
1
1
p+q
p+q q
p + q 1 + λm
ν+ (0)(p + qλm ) − ν− (0)(q + pλm ) 2mN +m −1
+
λ
1
1 + λm
1
p−q
λn−m
p
−1
=
+
×
m
q
1
p+q
p+q 1+λ
ν+ (0)(p + qλm ) − ν− (0)(q + pλm ) 2mN +n −1
+
λ
.
1
1 + λm
This completes the proof.
42
3.2.2
Discrete Time Markov Chain - Synchronised Saddles p = q
Let the period be
T = 4m
where m is an integer. Let the time t be written in the form
t = NT + n
where N is an integer number of periods. The transition matrix would vary periodically
according to
!
1−p
p
For n = mod(t, T ) ∈ T1 P1 =
p
1−p
!
1−q
q
For n = mod(t, T ) ∈ T2 P2 =
q
1−q
!
1−p
p
For n = mod(t, T ) ∈ T3 P3 =
p
1−p
!
1−q
q
For n = mod(t, T ) ∈ T4 P4 =
q
1−q
where
T1
T2
T3
T4
= {0, 1, . . . , m − 1}
= {m, m + 1, . . . , 2m − 1}
= {2m.2m + 1, . . . , 3m − 1}
= {3m, 3m + 1, . . . , 4m − 1}
where again p should be interpreted as the probability of escape from a shallow well and q
from a deep well. The following Theorem derives the state probabilities.
Theorem 3.2. Let the time be t = N T + n. The state probability at time t is
1
1
−1
2mN +n
2mN
+ (1 − 2p)
(1 − 2q)
[ν+ (0) − ν− (0)]
For n ∈ T1 ν =
1
1
2
1
−1
1
2mN +m
2mN +(n−m)
For n ∈ T2 ν =
+ (1 − 2p)
(1 − 2q)
[ν+ (0) − ν− (0)]
1
1
2
1
1
−1
2mN +m+(n−2m)
2mN +m
For n ∈ T3 ν =
+ (1 − 2p)
(1 − 2q)
[ν+ (0) − ν− (0)]
1
1
2
1
1
−1
2mN +2m
2mN +m+(n−3m)
For n ∈ T4 ν =
+ (1 − 2p)
(1 − 2q)
[ν+ (0) − ν− (0)]
1
1
2
43
Proof. Notice that the transpose of the matrix P1† has the following eigenvalues and eigenvectors
−1
λ1 = 1 − 2p v1 =
1
1
λ2 = 1
v2 =
.
1
These eigenvectors span the space, which means any vectors can be expressed as a linear
combination of them
1
−1
1
x
(y − x)
+ (x + y)
.
=
1
1
y
2
This means the initial values of the state probability ν can be expressed in terms of the
eigenvectors of P1† . Denote the matrix
Ptot = (P2† )m (P1† )m (P2† )m (P1† )m
which is the total transition matrix in one period. Proceeding we have
1
1
−1
N
2mN
2mN
Ptot ν(0) =
+ (1 − 2p)
(1 − 2q)
[ν+ (0) − ν− (0)]
1
1
2
with the added condition ν− (0) + ν+ (0) = 1. Now note the following
For n ∈ T1
N
ν(t) = (P1† )n Ptot
ν(0)
For n ∈ T2
N
ν(t) = (P2† )n−m (P1† )m Ptot
ν(0)
For n ∈ T3
N
ν(t) = (P1† )n−2m (P2† )m (P1† )m Ptot
ν(0)
For n ∈ T4
N
ν(t) = (P2† )n−3m (P1† )m (P2† )m (P1† )m Ptot
ν(0)
This completes the proof.
3.2.3
Discrete Time Markov Chain - Invariant Measures, Relaxation Time and Fourier Transform
We consider a discrete Markov Chain on {−1, +1}, that is
Yt = ±1
and the time is
t ∈ {0, 1, 2, . . .}
44
and the probabilities for being in Yt = −1 or Yt = +1 at time t are given by the state
probabilities ν± (t)
P (Yt = −1) = ν− (t) and P (Yt = +1) = ν+ (t).
The probabilities of transitions occurring as given in the transition matrices changes with
period T . After a very long time the state probabilities ν± (·) should not depend on the
initial state probabilities ν± (0). At time infinity ν± (·) should also be cyclic on [0, T ]. Let
the time be given by t = N T + n where N is a discrete number of periods. This leads us
to define the invariant measure as the state probabilities in the limit as N −→ ∞
ν(n) := lim ν(N T + n)
N −→∞
and since ν(·) is periodic on [0, T ] it should also satisfy
ν(t + T ) =
i=t+T
Y−1
Pi† ν(t)
i=t
where Pi are the transition matrices, that is to say ν is invariant over one period of application of the transition matrices. This brings us to the following.
Corollary 3.3. For the state probabilities in Theorem 3.1 the invariant measures are
1
λt
p−q
q
−1
×
For t ∈ T1 ν(t) =
−
1
p+q p
p + q 1 + λm
p−q
λt−m
1
p
−1
+
×
For t ∈ T2 ν(t) =
1
p+q q
p + q 1 + λm
Corollary 3.4. For the state probabilities in Theorem 3.2 the invariant measures are
1 1
ν(t) =
2 1
The proof is easy and omitted. The fact that the ν are invariant over one period of
application of the transition matrices follow from the proof of the Theorems.
The rate of convergence to the invariant measure would depend on the value of p and
q themselves. Define the relaxation time Trelax as the first time t = Trelax such that
|ν (Trelax ) − ν (Trelax )| ≤ e−1
which is a measure of the rate of convergence to the invariant measure.
Consider the averaged Markov Chain over many realisations. This is related to the
invariant measure by
hYt i = ν − (t)(−1) + ν + (t)(+1) = ν + (t) − ν − (t).
45
The Fourier Transform of the averaged Markov Chain hỸω i is often studied (see Chapter
4), that is
hỸω i = F (hYt i) .
We can Fourier Transform both the alternating saddle p 6= q case and the synchronised
saddle p = q case. This brings us to the following.
Corollary 3.5. For the Markov Chain in Theorem 3.1 the Fourier Transform of the averaged trajectory is
1 − e−iπω
1 − λm e−iπω
1 p−q
2
−iπω
1−e
.
−
hỸω i =
T p+q
1 − e−iπω/m 1 + λm 1 − λe−iπω/m
Proof. Notice that
ν − (t + m) = ν + (t) and ν + (t + m) = ν − (t)
so we have
hỸω i = F (hYt i)
2m−1
1 X −2πiωt/2m
=
hY ie
T t=0 t
m−1
2m−1
1 X −2πiωt/2m 1 X −2πiωt/2m
hY ie
hY ie
=
+
T t=0 t
T t=m t
m−1
m−1
1 X
1 X
−2πiωt/2m
[ν + (t) − ν − (t)] e
+
[ν − (t) − ν + (t)] e−2πiω(t+m)/2m
=
T t=0
T t=0
m−1
1 X
=
[ν + (t) − ν − (t)] e−2πiωt/2m − e−2πiω(t+m)/2m
T t=0
m−1
X
1
1 − e−iπω
[ν + (t) − ν − (t)] e−iπωt/m
=
T
t=0
m−1
X
λt
1
−iπω p − q
1−e
1−2
e−iπωt/m
=
T
p + q t=0
1 + λm
1 − e−iπω
2
1 − λm e−iπω
1 p−q
−iπω
−
=
1−e
.
T p+q
1 − e−iπω/m 1 + λm 1 − λe−iπω/m
This completes the proof.
Corollary 3.6. For the Markov Chain in Theorem 3.2 the Fourier Transform of the averaged trajectory is
hỸω i = 0.
46
Again the proof is trivial and omitted. If we study the Fourier Transform at ω = 1, this
would be the same as studying the driving frequency, which is the frequency at which the
transition matrices are changing. The physical intuition is that one has the most significant
response at this frequency.
3.3
Continuous Time Markov Chain
Let the time be continuous. This is to say time t belongs to
t ∈ R.
The Markov Chain is a time dependent stochastic process with values −1 or +1,
Yt = ±1.
Let p and q be real functions
p : R −→ R and q : R −→ R
and p and q are periodic on [0, T ]
p(t + T ) = p(t) and q(t + T ) = q(t).
Let A ⊆ [0, T ] be a subset of the interval [0, T ]. The probability of Yt transiting from
Yt = −1 to Yt = +1 for the times in A, t ∈ A, is denoted by
p−1+1 (A).
Similarly the probability of Yt transiting from Yt = +1 to Yt = −1 for the times in A,
t ∈ A, is denoted by
p+1−1 (A).
The probability of Yt staying at −1 in the time t ∈ A is given by
p−1−1 (A) = 1 − p−1+1 (A).
Similarly the probability of Yt staying at +1 in the time t ∈ A is given by
p+1+1 (A) = 1 − p+1−1 (A).
If A is a small time interval A = [t, t + δt] then the following infinitesimal representation
can be made
p−1+1 ([t, t + δt]) = p(t)δt
p+1−1 ([t, t + δt]) = q(t)δt.
47
Now we consider a small change in the state probabilities at times t and t + δt.
!
!†
!
ν− (t)
p−1−1 ([t, t + δt]) p−1+1 ([t, t + δt])
ν− (t + δt)
=
ν+ (t)
p+1−1 ([t, t + δt]) p+1+1 ([t, t + δt])
ν+ (t + δt)
!
!†
ν− (t)
1 − p(t)δt
p(t)δt
=
ν+ (t)
q(t)δt
1 − q(t)δt
then
ν(t + δt) − ν(t) =
=
ν(t + δt) − ν(t)
=
δt
ν− (t + δt)
!
ν− (t)
−
ν+ (t + δt)
ν+ (t)
!†
−p(t)δt
p(t)δt
q(t)δt
−q(t)δt
!†
p(t)
−p(t)
q(t)
−q(t)
!
ν− (t)
!
ν+ (t)
!
ν− (t)
ν+ (t)
which in the limit of small δt leads to a differential equation describing the behaviour of
ν(t)
dν
= Q† ν
dt
where the infinitesimal generator Q is defined as
Q=
−p(t)
p(t)
q(t)
−q(t)
(3.2)
!
.
Note that the transpose of Q is taken in Equation 3.2. The aim now is to derive the state
probability by solving this differential equation for various forms of p and q. The extra
conditions we use are
ν− (t) + ν+ (t) = 1
ν−0 (t) + ν+0 (t) = 0
for all times t and the initial conditions at t = 0 are ν− (0) and ν+ (0).
3.3.1
Continuous Time Markov Chain - Alternating Saddles p 6=
q
Notice that p may be interpreted as the probability of escape from the left well and q as
the probability of escape from the right well. If p and q are cyclic over [0, T ], then this can
be interpreted as modelling a potential with periodic forcing in continuous time.
48
Theorem 3.7. Let p 6= q and t ≥ 0. The state probabilities are given by
R s
Rt
ν− (0) + 0 q(s) exp 0 p(u) + q(u) du ds
nR
o
ν− (t) =
t
exp 0 p(u) + q(u) du
ν+ (t) =
ν+ (0) +
Rt
R s
p(s) exp 0 p(u) + q(u) du ds
nR
o
t
exp 0 p(u) + q(u) du
0
Proof. The differential equations we want to solve are given by
!
!†
!
ν
(t)
−p(t)
p(t)
ν
(t)
−
−
d
=
dt ν+ (t)
q(t) −q(t)
ν+ (t)
!
!
−p(t) q(t)
ν− (t)
=
p(t) −q(t)
ν+ (t)
which gives
dν−
= −pν− + qν+
dt
dν+
= pν− − qν+
dt
and by using ν− + ν+ = 1 we get
dν−
+ (p + q)ν− = q
dt
dν+
+ (p + q)ν+ = p.
dt
(3.3)
(3.4)
We will only solve for ν− (t). The case for ν+ (t) is similar. Equation 3.3 can easily be
solved with an integrating factor
Z t
Z t
d
p(u) + q(u) du
p(u) + q(u) du
= q(t) exp
ν− (t) exp
dt
0
0
and proceeding we have
Z s
Z t
Z t
q(s) exp
p(u) + q(u) du ds
ν− (t) exp
p(u) + q(u) du − ν− (0) =
0
0
0
which rearranges to give
ν− (t) =
ν− (0) +
Rt
R s
q(s) exp 0 p(u) + q(u) du ds
nR
o
.
t
exp 0 p(u) + q(u) du
0
This completes the proof.
49
3.3.2
Continuous Time Markov Chain - Synchronised Saddles
p=q
If p = q for continuous time, then this can be modelled as both wells of the potential
always being at the same height but moving together.
Theorem 3.8. Let p = q and t ≥ 0. The state probabilities are given by
Z t
1 ν+ (0) − ν− (0)
p(s) ds
exp −2
ν− (t) = −
2
2
0
Z t
1 ν+ (0) − ν− (0)
ν+ (t) = +
p(s) ds
exp −2
2
2
0
Proof. The differential equations we have to solve are given by
!
!
!
ν
(t)
−1
1
ν
(t)
−
−
d
= p(t)
dt ν+ (t)
ν+ (t)
1 −1
which gives
dν−
= p (ν+ − ν− )
dt
dν+
= p (ν− − ν+ )
dt
(3.5)
(3.6)
and subtracting Equation 3.5 away from Equation 3.6 leads to
d
(ν+ − ν− ) = −2p (ν+ − ν− ) .
dt
Denote the difference by
z(t) = ν+ (t) − ν− (t)
which gives the differential equation we need to solve to
dz
= −2pz
dt
Z z(t)
Z t
dz
= −2
p(s) ds
z(0) z
0
Z t
ln(z(t)) − ln(z(0)) = −2
p(s) ds
0
Z t
z(t) = z(0) exp −2
p(s) ds
0
50
and using z = ν+ − ν− and ν− + ν+ = 1 rearranges the solution to
Z t
1 ν+ (0) − ν− (0)
p(s) ds
ν− (t) = −
exp −2
2
2
0
Z t
1 ν+ (0) − ν− (0)
p(s) ds .
ν+ (t) = +
exp −2
2
2
0
This completes the proof.
3.3.3
Continuous Time Markov Chain - Invariant Measures and
Fourier Transform
As in the discrete time case we can compute the corresponding invariant measures.
Corollary 3.9. For the state probabilities in Theorem 3.7 the invariant measures are
Rt
RT
p(s)g(s)
ds
p(s)g(s) ds
ν − (t) = 0
+ 0
g(t)
g(t) (g(T ) − 1)
RT
Rt
q(s)g(s) ds
q(s)g(s) ds
0
ν + (t) =
+ 0
g(t)
g(t) (g(T ) − 1)
where
Z
g(t) = exp
t
p(u) + q(u) du .
0
Proof. We derive the invariant measure for ν − (t). The case for ν + (t) is similar. Consider
the fact that p(·) and q(·) are cyclic on [0, T ] and let i be an integer, then the following
integral can be rewritten as
Z (i+1)T
Z T
p(s)g(s) ds =
p(s)g(iT + s) ds
iT
0
Z T
=
p(s)g(iT )g(s) ds
0
Z T
= g(iT )
p(s)g(s) ds
0
Z T
i
p(s)g(s) ds.
= g(T )
0
Let the time be given by N T + t where N is an integer number of periods. This means the
following integral can be written as
Z N T +t
Z N T +t
N
−1 Z (i+1)T
X
p(s)g(s) ds =
p(s)g(s) ds +
p(s)g(s) ds
0
NT
i=0
51
iT
t
Z
=
T
Z
p(s)g(N T + s) ds +
0
p(s)g(s) ds
0
= g(T )
N
Z
t
Z
T
0
0
g(T )i
i=0
p(s)g(s) ds
p(s)g(s) ds +
N
−1
X
N
−1
X
g(T )i .
i=0
So the state probability is equal to
R s
R N T +t
ν− (0) + 0
q(s) exp 0 p(u) + q(u) du ds
nR
o
ν− (N T + t) =
N T +t
exp 0
p(u) + q(u) du
RT
P −1
i
p(s)g(s) ds + 0 p(s)g(s) ds N
i=0 g(T )
=
g(T )N g(t)
Rt
RT
N −1
p(s)g(s)
ds
p(s)g(s) ds 1 X
ν− (0)
0
0
=
+
+
g(T )i
g(T )N g(t)
g(t)
g(t)
g(T )N i=0
ν− (0) + g(T )N
ν− (0)
+
=
g(T )N g(t)
Rt
0
Rt
0
p(s)g(s) ds
+
g(t)
RT
0
p(s)g(s) ds
1
g(t)
g(T ) − 1
1−
1
g(T )N
.
Letting N −→ ∞ gives the required result.
Corollary 3.10. For the state probabilities in Theorem 3.8 the invariant measures are
1 1
ν(t) =
.
2 1
The proof is trivial and omitted. Similar to the discrete time case we can also study the
Fourier Transform of the averaged Markov Chain.
Corollary 3.11. For the Markov Chain in Theorem 3.7 the Fourier Transform of the
averaged Markov Chain is
!
RT
Z +∞ R t
[q(s)
−
p(s)]
g(s)
ds
[q(s)
−
p(s)]
g(s)
ds
0
F (hYt i) =
+ 0
e−i2πωt dt.
g(t)
g(t)
(g(T
)
−
1)
−∞
Corollary 3.12. For the Markov Chain in Theorem 3.8 the Fourier Transform of the
averaged Markov Chain is
F (hYt i) = 0.
52
3.4
Probability Density Function of Escape Times
The escape rates from the left to right are denoted by R−1+1 (·) and right to left escape
rates are denoted by R+1−1 (·). The PDFs for the escape times are given by the Theorem
below.
Theorem 3.13. Let u be the time of entry into a well, then the PDFs for the escape
occurring at time t > u are
Z t
p− (t, u) = R−1+1 (t) exp −
R−1+1 (s) ds
u
Z t
R+1−1 (s) ds
p+ (t, u) = R+1−1 (t) exp −
u
where p− (t, u) is for left to right and p+ (t, u) is for right to left.
Proof. We consider escaping from the left well. The right well is similar. Divide the time
interval [u, t] into many small time intervals
δt =
t−u
.
N
Similar to how we derived the invariant measures we want to derive the probability of
escape in a very small time interval [t, t + δt]. This is given by
p−1+1 ([t, t + δt]) = p(t)δt
= 1 − e−R−1+1 (t)δt
≈ R−1+1 (t)δt
which is valid for small δt. Large deviations allow us to say even more about the escape
time τ−1+1 and τ+1−1 . Theorem 1 in [35] shows that it is an exponentially distributed
random variable. The probability of staying in the left well is given by
p−1−1 ([t, t + δt]) = 1 − p−1+1 ([t, t + δt])
= 1 − p(t)δt
= 1 − 1 − e−R−1+1 (t)δt
= e−R−1+1 (t)δt .
We want to know the probability of escaping in the time interval [t, t + δt] given that the
particle has entered at u and stayed up to time t. This is given by
p−1−1 ([u, t])p−1+1 ([t, t + δt]) =
N
Y
p−1−1 ([u + (i − 1)δt, u + iδt]) p−1+1 ([t, t + δt])
i=1
53
=
N
Y
exp {−R−1+1 (u + (i − 1)δt) δt} p−1+1 ([t, t + δt])
i=1
= exp
( N
X
)
−R−1+1 (u + (i − 1)δt) δt p−1+1 ([t, t + δt])
i=1
Z t
R−1+1 (s) ds R−1+1 (t)δt.
= exp −
u
This completes the proof.
3.4.1
Normalised Time Probability Density Function of Escape
Times
The period of the forcing is T and we can make a change of variables to normalised time
tnorm =
treal
T
which measures time in how many periods have elapsed. This rearranges the PDFs to
Z t
R−1+1 (T s) ds
p− (t, u) = T R−1+1 (T t) exp −T
u
Z t
p+ (t, u) = T R+1−1 (T t) exp −T
R+1−1 (T s) ds
u
where u, t and s are in normalised time. Note that R−1+1 (·) and R+1−1 (·) always have
their arguments in real time and R−1+1 (·) and R+1−1 (·) always give the averaged number
of transitions per real unit time. Expressing the PDF in normalised time can be found in
[41].
3.4.2
Perfect Phase Approximation of Probability Density Function of Escape Times
The PDF for the escape times derived in Theorem 3.13 had to differentiate between left
and right escapes and are conditioned on the time u of entrance into the well. Suppose
now that t is the escape time from any well, which does not differentiate between left and
right escape. Note that t is the actual time it takes to escape from a well and is not a time
coordinate. The PDF for t is given by
Z
1 T
ptot (t) =
p− (t + u, u)m− (u) + p+ (t + u, u)m+ (u) du.
2 0
This is because after a long time has elapsed we would expect that many transitions would
have occurred between left and right. The number of transitions escaping from the left
54
and right should be roughly the same. The m− (u) is a PDF for the time of entrance into
the left well and the m+ (u) is a PDF for the time of entrance into the right well. We may
not have explicit expressions for m− (u) and m+ (u). We derive an approximate expression
for ptot without an explicit expressions for m− (u) and m+ (u). Let m− (u) and m+ (u) be
approximated by
m− (u) ≈ δ (u − T /2)
1
1
m+ (u) ≈ δ (u) + δ (u − T )
2
2
where δ(·) is the Dirac delta function. This approximation is used because in the SDEs
which we will simulate, the times when transition into the left well is greatest is at half
the period u = T2 and the times when transition into the right well is greatest is at u = 0
and u = T . Due to the fact that m− (u) and m+ (u) are probabilities a factor of 21 is used
in m+ (u). Progressing we have
Z
1 T
ptot (t) =
p− (t + u, u)m− (u) + p+ (t + u, u)m+ (u) du
2 0
Z
1
1
1 T
δ (u) + δ (u − T ) du
p− (t + u, u)δ (u − T /2) + p+ (t + u, u)
≈
2 0
2
2
1
1
1
p− (t + T /2, T /2) + p+ (t, 0) + p+ (t + T, T )
=
2
2
2
1
1
1
=
p− (t + T /2, T /2) + p+ (t, 0) + p+ (t + 0, 0)
2
2
2
=
1
{p− (t + T /2, T /2) + p+ (t, 0)}
2
= p+ (t, 0).
This is because for the simulations which we are going to do, the Kramers’ rate satisfy
R−1+1 (t) = R+1−1 (t + T /2) (see later in Chapter 5 for the geometry of the Mexican Hat
Toy Model which justifies this). Thus the following approximation
ptot ≈ p+ (t, 0)
is only valid for the simulations we do, and not for a general potential. We call this way
of approximating m− (u) and m+ (u) the perfect phase approximation.
3.5
Adiabatic Large Deviation
We have to stress that this thesis is built on three approximations, which form the backbone
of all the research presented. These are small noise approximation, adiabatic approximation
and perfect phase approximation.
55
Perfect phase approximation only works for small noise. This is because the noise is so
small the particle will only escape when the maximum probability to escape has arrived.
When the minimum probability to escape is present it will almost never escape. This is
the idea behind the perfect phase approximation.
Notice one subtlety behind all the theory presented in this Chapter. The derivations
involved probabilities of escape p and q and the escape rates R−1+1 and R+1−1 . But it was
assumed that p, q, R−1+1 and R+1−1 are accurately known no matter how large or small
the noise level is and no matter how fast or slow the driving frequency Ω is. But such
ideal expressions for p, q, R−1+1 and R+1−1 are not known.
When we come to do the analysis in Chapter 7, the ptot is calculated with the approximation ptot ≈ p+ (t, 0). When the rates R−1+1 and R+1−1 are needed they are calculated
using Kramers’ formula as though it is escape from a static potential in the small noise
limit. This means an oscillatory potential is being approximated by a static potential
which is the adiabatic approximation.
In the paper [54] the adiabatic approximation was justified in the small noise, slow
forcing limit using time dependent large deviation theory, that is, it was shown asymptotically the escape times are given by the adiabatic approximation. This result is only for
the leading term, whether the analogue result holds for the Kramers’ rate is unknown.
56
Chapter 4
Theory of Analysis of Stochastic
Resonance
We present different criteria that have been used to define stochastic resonance. This
includes the six measures, which are linear response, signal-to-noise ratio, energy, out-ofphase measures, relative entropy and entropy. A new statistical test called the conditional
Kolmogorov-Smirnov test is introduced.
4.1
Six Measures of Stochastic Resonance
We introduce six possible criteria of measuring how close a process is to exhibiting stochastic resonance [45, 41]. These six criteria are closely related to linear response [25, 26],
signal-to-noise ratio [27, 28] and distribution of escape times [29, 28, 30]. We call them the
six measures denoted by M1 , M2 , M3 , M4 , M5 and M6 . Recall that the SDE we want to
study is
Ẋt = −∇V0 + F cos Ωt + Ẇt
where V0 : R2 −→ R is the unperturbed potential, F is the forcing, Ω is the forcing
frequency, is the noise level and Wt is a Wiener process in two dimensions, which when
rewritten into separate components are
∂V0
+ Fx cos Ωt dt + dwx
dx = −
∂x
∂V0
dy = −
+ Fy cos Ωt dt + dwy
∂y
where wx and wy are two independent Wiener processes. The solution to these equations
is the trajectory in two dimensions
Xt = (xt , yt ) .
57
This diffusion can be reduced to a Markov Chain on {−1, +1} denoted by
Yt = ±1
where by definition of the Markov Chain the escape times are the same as the diffusion
case (see Chapter 3). The probability of the Markov Chain being in one state at time t is
given by the state probabilities
P (Yt = −1) = ν− (t) and P (Yt = +1) = ν+ (t).
In what follows we will consider so large times, that the relaxation time has effectively
elapsed for both the diffusion and Markov Chain, in other words the state probability
would have effectively converged to the invariant measure ν. This means that over one
period T = 2π/Ω of the forcing, the invariant measures will have the properties
ν ± (t) = ν ± (t + T ) and ν ± (t) = ν ∓ (t + T /2).
We obtain the averaged trajectories given by
hXt i = E (Xt )
and
hYt i = E (Yt )
which are the trajectories obtained after averaging over many realisations. Notice that
hYt i is related to the invariant measures by
hYt i = ν + (t) − ν − (t).
We introduce the Out-of-Phase Markov Chain defined by
0 if Yt = −1 and mod(t, T ) ≤ T /2
1 if Yt = −1 and mod(t, T ) > T /2
Yt =
1 if Yt = +1 and mod(t, T ) ≤ T /2
0 if Yt = +1 and mod(t, T ) > T /2
and similarly the averaged Out-of-Phase Markov Chain is defined by
Yt =E Yt .
Define two new functions by
−
1 if mod(t, T ) ≤ T /2
0 if mod(t, T ) > T /2
0 if mod(t, T ) ≤ T /2
1 if mod(t, T ) > T /2.
φ (t) =
+
φ (t) =
The following trajectories are Fourier transformed
x̃(ω) = F (hxt i) = hF (xt )i
58
Ỹ (ω) = F (hYt i) = hF (xt )i.
The linear response is defined as the intensity of the Fourier Transform at the driving
frequency Ω 20
Ω
Ω
and Ylin = Ỹ
.
Xlin = x̃
2π
2π
Now we can define the six measures. For the diffusion case only M1 and M2 are defined.
1
Xlin
F
1
M2 =
Xlin
F
M1 =
where F is the magnitude of the forcing. For the Markov Chain M1 , M2 , M3 , M4 , M5 and
M6 are all defined as
1
Ylin
F
1
=
Ylin
F
Z T
hYt i2 dt
=
0
Z T
=
Y t dt
0
−
+
Z T
φ (t)
φ (t)
−
+
=
φ (t) ln
+ φ (t) ln
dt
ν − (t)
ν + (t)
0
Z T
−ν − (t) ln ν − (t) − ν + (t) ln ν + (t) dt.
=
M1 =
M2
M3
M4
M5
M6
0
Note that in the definition of the six measures it is assumed that the process has relaxed
to equilibrium. We give a few physical interpretation of the six measures M1 , M2 , M3 , M4 ,
M5 and M6 . The M1 is the intensity of the driving frequency Ω in the spectrum of the
Fourier transform. The M2 is sometimes called signal-to-noise ratio as it compares this
intensity to the noise level . The M3 is sometimes called the energy. The M4 is sometimes
called the out-of-phase measure since it measures the amount of time the Markov Chain
spends in the “wrong” well. The M5 and M6 are sometimes called relative entropy and
entropy respectively, since they measure how far away the invariant measures are from
being constant. If the invariant measures are constant then these six measures will also
be constant. Thus it can be understood that these six measures is a measure of how far
away the invariant measures are from being constant. M6 measures how non-constant the
invariant measure is. M5 is extremal if the invariant measure is constant.
20
See Appendix B.2 for how the linear response is calculated numerically.
59
4.2
Statistical Tests
We will measure the escape time for many consecutive transitions. This will result in a
collection of measurements of escape times
τ1 , τ2 , . . . , τn .
A new method for analysing such a collection of measurements is presented.
4.2.1
Kolmogorov-Smirnov Test
First we recall results about the Kolmogorov-Smirnov statistic and the Kolmogorov-Smirnov
test [55]. Let
ξ1 , ξ2 , . . . , ξn
be n independently and identically distributed real random variables. Each ξi is distributed
with PDF f (·) as in
Z
P (ξi ∈ A) =
f (s) ds
A
and distributed with CDF F (·) as in
Z
x
P (ξi ≤ x) = F (x) =
f (s) ds.
−∞
Define a function by Fn (·) by
n
1X
Fn (x) =
1(−∞,x] (ξi )
n i=1
where 1A is the indicator function for a set A. We may think of ξ1 , ξ2 , . . . , ξn as n empirical
or numerical realisations of the same random variable ξ. The Fn (x) is therefore an approximation to the CDF of ξ that is empirically found using ξ1 , ξ2 , . . . , ξn , therefore Fn (·) is
called the empirical CDF. Consider the supremum metric on the space of real continuous
functions. Consider the distance between the real and the empirical CDF in this metric.
Dn = kFn − F k∞
n
1X
1(−∞,x] (ξi ) − F (x)
= sup
x∈R n
i=1
where Dn is called the Kolmogorov-Smirnov statistic or KS statistic. Intuitively we would
expect Dn to tend to zero as n increases, that is
lim Dn = 0
n−→∞
if the ξi are distributed by F (·). There are times when we experimentally obtain n values
of a random variable ξ1 , ξ2 , . . . , ξn , and want to test whether they are distributed by a CDF
F (·). We define what we mean by the null hypothesis.
60
Definition 4.1. Let ξ1 , ξ2 , . . . , ξn be n real random variables. The null hypothesis is that
each ξi is independently distributed with CDF F (x).
We want to know how large or small Dn needs to be before deciding whether to reject the
null hypothesis. The following Theorem offers a remarkable answer to this problem.
Theorem 4.2. Suppose the null hypothesis is true, then the distribution of Dn depends
only on n.
Notice that Dn is in itself a real random variable. The PDF and CDF of Dn is a function of
n only, and will be the same whatever F (·) is. This distribution is called the KS distribution
and tables are available upto n = 100. There is a Theorem which describes the asymptotic
behaviour of the KS distribution [56, 57].21
√
Theorem 4.3. In the limit n −→ ∞, nDn is asymptotically Kolmogorov distributed with
the CDF
∞
X
2 2
Q(x) = 1 − 2
(−1)k−1 e−2k x
k=1
that is to say
√
lim P ( nDn ≤ x) = Q(x).
n−→∞
4.2.2
Conditional Kolmogorov-Smirnov Test
Let ζ1 , ζ2 , . . . , ζn be n iid real random variables. They are n empirical observations of a
random variable ζ. Now suppose that each of the ξ1 , ξ2 , . . . , ξn is conditioned and dependent on the corresponding ζ1 , ζ2 , . . . , ζn . This means a conditional PDF f (·, ·) gives the
probability
Z
f (s, ζi ) ds
P (ξi ∈ A | ζi ) =
A
and the conditional CDF F (·, ·) is
Z
x
P (ξi ≤ x | ζi ) = Fζi (x) =
f (s, ζi ) ds.
−∞
But ξ1 , ξ2 , . . . , ξn are empirical measurements of the same random variable ξ. The PDF for
ξ is given by
Z Z +∞
P (ξ ∈ A) =
f (s, u)m(u) ds du
A
−∞
21
There appears
errors in the literature for the limitingP
function. Some sources
P∞ to be topographical
2 2
2 2
∞
cite Q1 = 1 − 2 k=1 (−1)k−1 e−2k x (see [58, 59, 60]) and some cite Q2 = 1 − 2 k=1 (−1)k−1 e−k x (see
[56, 57]). But the proof of Theorem 1 in [57] shows Q = Q1 . Nevertheless in this thesis we use Q = Q1
which actually gives smaller and more conservative values of the metric Dn that are needed.
61
and the CDF for ξ is
Z
x
Z
+∞
P (ξ ≤ x) = F (x) =
f (s, u)m(u) ds du
−∞
−∞
where m(·) is the PDF for ζ, that is
Z
P (ζ ∈ A) =
m(s) ds.
A
In our context we have the problem that the random variables are not identically distributed
under the null hypothesis. The ξ1 , ξ2 , . . . , ξn and ζ1 , ζ2 , . . . , ζn are obtained experimentally
and Fζi (ξi ) can be calculated but a PDF for ζi , that is m(·), has no easy expression. We
still want to perform a statistics test that is similar to the KS test even in such situations
where the distribution m(·) of ζ is unknown. First we define what we call the total null
hypothesis and the conditional null hypothesis.
Definition 4.4. Let ξ1 , ξ2 , . . . , ξn be n empirical observations of a random variable ξ.
The total null hypothesis is that ξ is distributed with the CDF F (·). The conditional null
hypothesis is that each ξi is distributed with the conditional CDF Fζi (·).
A new statistical test is developed, which is similar to the KS test.
Theorem 4.5. Suppose the conditional null hypothesis is true. Let Fζi (·) be continuous.
Let Sn be the statistic given by
n
Sn = sup
x∈[0,1]
1X
1[0,x] (Fζi (ξi )) − x
n i=1
then Sn is KS distributed.
Proof. Denote
Yi = Fζi (ξi )
which means
P (Yi ≤ x) = P (Fζi (ξi ) ≤ x)
= P ξi ≤ Fζ−1
(x)
i
= Fζi Fζ−1
(x)
i
=x
and 0 ≤ Yi ≤ 1, so Yi is uniformly distributed on [0, 1]. Note that Fζi (·) is a function of
one variable only. Let
n
n
1X
1X
Fn (x) =
1[0,x] (Yi ) =
1[0,x] (Fζi (ξ))
n i=1
n i=1
62
where Fn (·) is the empirical CDF of a uniformly distributed random variable, computed
using n observations. The statistic Sn is the suprenum metric
Sn = kFn − xk∞
n
1X
= sup
1[0,x] (Fζi (ξi )) − x .
x∈[0,1] n i=1
So clearly Sn is KS distributed.
We call Sn the conditional KS statistic. Compare this to the original KS statistic, which
under the assumption of the total null hypothesis can be rewritten as
n
n
1X
1X
Dn = sup
1(−∞,x] (ξi ) − F (x) = sup
1[0,x] (F (ξi )) − x .
x∈R n
x∈[0,1] n i=1
i=1
When both the total and conditional hypothesis are true Dn and Sn are KS distributed,
that is
P (Dn ∈ A) = P (Sn ∈ A)
and P (Dn ≤ x) = P (Sn ≤ x) .
The subtlety here is that Dn and Sn are different objects, yet they have the same distribution. Dn is KS distributed under the total null hypothesis, whereas Sn is KS distributed
under the conditional null hypothesis. This can be explained in another way. We have
n experimental observations of a random variable ξ denoted by ξ1 , ξ2 , . . . , ξn and each are
conditioned on observations of another random variable ζ denoted by ζ1 , ζ2 , . . . , ζn . The
Dn is KS distributed if the random variable ξ is distributed by CDF F (·), but the Sn is
KS distributed if each ξi is conditionally distributed by the CDF Fζi (·).
63
Chapter 5
Mexican Hat Toy Model
The main object of consideration of this project, which is called the Mexican Hat Toy
Model, is now introduced. Let a > 0, b > 0 and V0 : R2 −→ R be a real function from the
plane to the line. The unperturbed potential is defined as
1
1
V0 (x, y) = r4 − r2 − ax2 + by 2
4
2
where r =
p
x2 + y 2 .
Let Fx , Fy ∈ R be the forcing. The potential with forcing VF is defined as
1
1
VF (x, y) = r4 − r2 − ax2 + by 2 + Fx x + Fy y
4
2
1 4 1 2
= r − r − ax2 + by 2 + F · x
4
2
= V0 + F · x
written more compactly in vector notation. When VF is defined using +F · x we say
positive forcing. Alternatively if VF is defined using −F · x we say negative forcing. The
VF is defined with a positive forcing because for the rest of this Chapter we will study the
critical points which are solutions to the simultaneous equations
∂VF
= 0 and
∂x
∂VF
= 0.
∂y
The properties of the critical points will change as F is increased from zero, therefore it
is convenient to define VF with a positive forcing. The behaviour of the critical points
are studied for different cases. The main aim is to find the positions and nature of the
critical points for a range of parameter values. This is complex due to several cases to be
considered and previewed in the following Theorem, which is one of the main conclusions
of this Chapter. Although this Theorem only considers the case for non-negative forcing
Fx ≥ 0 and Fy ≥ 0, the case for Fx < 0 and Fy < 0 is similar by considering the symmetry
of the potential.
64
Theorem 5.1. Let Fx ≥ 0, Fy ≥ 0, a > 0, b > 0. Note that definitions of constants are
at the end. The positions and nature of the critical points of the Mexican Hat Toy Model
VF (·) are for the following range of parameters.
For Fx = 0, Fy = 0 and b <
1
2
(0, 0)
hill
√
(± 1 + 2a, 0) well
√
(0, ± 1 − 2b) saddle
For Fx = 0, Fy = 0 and b ≥
1
2
(0, 0)
saddle
√
(± 1 + 2a, 0) well
For Fx > 0, Fy = 0, b <
1
2
and the following values of Fx
(x0 , 0) well
For any Fx > 0
well
(x1 , 0)
sad
(x
,
0)
hill
and if Fx < Fx
2
(xsaddle , ±ysaddle ) saddle
or
or
Fxsad
(x1 , 0)
(x1 , 0)
< Fx < Fxcrit
(x2 , 0)
(x , 0)
2
(xsaddle , ±ysaddle )
saddle
well
hill
saddle
nonexistent
(x1 , 0)
unidentified
(x
,
0)
well
1
sad
crit
Fx = Fx < Fx
(x2 , 0)
hill
(x , 0)
unidentified
2
(xsaddle , ±ysaddle ) unidentified
(x1 , 0) unidentified
crit
or Fx = Fx
(x2 , 0) unidentified
(x1 , 0) nonexistent
crit
or Fx > Fx
(x2 , 0) nonexistent
65
√
for √1 − 2b ∈ R1
for √1 − 2b ∈ R2
for √1 − 2b ∈ R1
for
1 − 2b ∈ R2
√
for √1 − 2b ∈ R1
for √1 − 2b ∈ R2
for √1 − 2b ∈ R1
1 − 2b ∈ R2
for
For Fx > 0, Fy = 0, b ≥
1
2
For Fx = 0, Fy > 0, b ≥
1
2
For Fx = 0, Fy > 0, b >
1
2
and the following values of Fx
(x0 , 0) well
crit
(x1 , 0) well
Fx < Fx
(x2 , 0) saddle
(x0 , 0) well
crit
(x1 , 0) nonexistent
Fx > Fx
(x2 , 0) nonexistent
(x0 , 0) well
crit
(x1 , 0) unidentified
Fx = Fx
(x2 , 0) unidentified
and the following values of Fy
(0, y1 ) saddle
crit
Fy < Fy
(0, y2 ) hill
(0, y0 )
saddle
sad
Fy < Fy
(±xwell , ywell ) well
(0, y1 ) nonexistent
Fy > Fycrit
(0, y2 ) nonexistent
(0, y0 )
well
Fy > Fysad
(±xwell , ywell ) nonexistent
(0, y1 ) unidentified
Fy = Fycrit
(0, y2 ) unidentified
(0, y0 )
unidentified
Fy = Fysad
(±xwell , ywell ) unidentified
and the following values of Fy
(0, y0 )
saddle
sad
Fy < Fy
(±xwell , ywell ) well
(0, y0 ) = (±xwell , ywell ) unidentified
Fy = Fysad
(0, y0 )
well
sad
Fy > Fy
(±xwell , ywell ) nonexistent
66
where
!
)
p
3 − 27F 2
4(1
+
2a)
2π
1
x
√
+
tan−1
k
3
3
Fx 27
q
3 − 27F 2
4(1
−
2b)
y
1
2 √
+ 2π k
√
tan−1
yk = − √ 1 − 2b cos
3
3
3
Fy 27
2 √
xk = − √ 1 + 2a cos
3
xsaddle =
Fx
2(a + b)
s
s
xwell =
Fx
2(a + b)
2
Fy
2(a + b)
2
(1 − 2b) −
ysaddle =
(1 + 2a) −
(
−Fy
2(a + b)
r
4(1 + 2a)3
Fxcrit =
27√
sad
Fx = 2(a + b) 1 − 2b
r
4(1 − 2b)3
Fycrit =
27√
sad
Fy = 2(a + b) 1 + 2a
√
1 √
R1 = √ 1 + 2a, 1 + 2a
3
1 √
R2 = 0, √ 1 + 2a .
3
ywell =
The proof is given in a series of Lemmas for each of the six different cases. Theorem 5.4
proves the case for Fx = 0 and Fy = 0, Theorem 5.10 proves the case for Fx > 0, Fy = 0
and b < 21 , Theorem 5.11 proves the case for Fx > 0, Fy = 0 and b ≥ 12 , Theorem 5.16
proves the case for Fx = 0, Fy > 0 and b < 12 and Theorem 5.17 proves the case for Fx = 0,
Fy > 0 and b ≥ 12 . All the notation used will be consistent with this current Theorem 5.1.
The following standard result is used.
Theorem 5.2. Let V : R2 −→ R be twice differentiable everywhere. Let H be the Hessian
at a critical point (x0 , y0 ). The nature of the critical point can be determined by
det H < 0 ⇒
(
det H > 0 then
if
∂2V
∂x2
saddle
>0 ⇒
well
if
∂2V
∂x2
<0 ⇒
hill
67
We recall results about the cubic equation.
Theorem 5.3. Let a3 , a2 , a1 , a0 ∈ R where a3 6= 0. Consider the cubic equation
a3 x 3 + a2 x 2 + a1 x + a0 = 0
and its discriminant
∆ = 18a3 a2 a1 a0 − 4a32 a0 + a22 a21 − 4a3 a31 − 27a23 a20
then following statements hold
if ∆ > 0 then the equation has 3 distinct real roots.
if ∆ = 0 then the equation has a multiple real root and all its roots are real.
if ∆ < 0 then the equation has 1 real root and 2 complex conjugate roots.
and the three roots of the equations are
1
−iψk ∆0
iψk
a2 + e C + e
xk = −
3a3
C
where
ψ0 = 0, ψ1 = 2π/3, ψ2 = 4π/3
s
√
−27∆
3 ∆1 +
C=
2
2
∆0 = a2 − 3a3 a1
∆1 = 2a32 − 9a3 a2 a1 + 27a23 a0 .
5.1
Case Fx = 0 and Fy = 0
The case for no forcing F = 0 is considered first.
Theorem 5.4. When Fx = Fy = 0 the critical points of the potential have the following
properties. For b < 12 the critical points and their nature are
b<
For b ≥
1
2
1
2
(0, 0)
hill
√
(± 1 + 2a, 0) well
√
(0, ± 1 − 2b) saddle
the critical points and their nature are
b≥
1
2
(0, 0)
saddle
√
(± 1 + 2a, 0) well
The proof is trivial and omitted.
68
5.2
Case Fx > 0 and Fy = 0
When forcing is only in the x direction two cases are considered separately, that is for
b < 12 and b ≥ 12 . The case for Fx ≤ 0 is similar.
5.2.1
Case Fx > 0, Fy = 0 and b <
1
2
When there is no forcing there are five critical points. Intuitively as forcing is increased the
system could gradually start to deviate away from having five critical points. The critical
points may collide and coincide. The structure of the following proofs are first determining
the bounds on the critical points and then determining their nature. We have the following
consequence which uses the solution and theory of the cubic equation with three real roots.
Theorem 5.5. Let Fx > 0, Fy = 0 and b < 21 . Let Fx be bounded by
Fx ≤ Fxsad
Fx ≤ Fxcrit
and
where
Fxsad
√
= 2(a + b) 1 − 2b
r
and
Fxcrit =
4(1 + 2a)3
27
then there are five critical points given by
(xk , 0) k = 1, 2, 3
(xsaddle , ±ysaddle )
where
xsaddle =
Fx
2(a + b)
s
2
Fx
ysaddle = (1 − 2b) −
2(a + b)
2 √
2π
xk = − √ 1 + 2a cos φ + k
3
3
1
φ = tan−1 (l)
3p
4(1 + 2a)3 − 27Fx2
√
l=
Fx 27
k = 0, k = 1, k = 2.
Proof. The simultaneous equations to be solved are
∂VF
= x(x2 + y 2 ) − (1 + 2a)x + Fx = 0
∂x
69
(5.1)
∂VF
= y(x2 + y 2 ) − (1 − 2b)y = 0.
∂y
(5.2)
Equation 5.2 holds if either (x2 + y 2 ) − (1 − 2b) = 0 or y = 0. The (x2 + y 2 ) − (1 − 2b) = 0
case is considered first, which gives (x2 + y 2 ) = (1 − 2b). Substituting this into Equation
5.1 gives x as
Fx
xsaddle =
2(a + b)
which when substituted back into (x2 + y 2 ) = (1 − 2b) gives y as
s
2
Fx
ysaddle = ± (1 − 2b) −
.
2(a + b)
For the y = 0 case, Equation 5.1 becomes
x3 − (1 + 2a)x + Fx = 0
which is a cubic equation. Solving this cubic equation using the notation in Theorem 5.3
gives
∆0 = 3(1 + 2a), ∆1 = 27Fx , ∆ = 4(1 + 2a)3 − 27Fx2
which gives22
√
27Fx + i 27∆
C=
2
!2 1/6
(
√
2
27∆
27Fx
exp i(1/3) tan−1
+
=
2
2
s
3
√
27∆
27Fx
which simplifies to
27Fx
2
√
2
+
27∆
2
!2
=
1
27 × 27Fx2 + 27∆
4
= 27(1 + 2a)3 ,
after letting
√
φ = (1/3) tan−1
22
27∆
27Fx
!
After noting that (x + iy)1/3 = (x2 + y 2 )1/6 exp i(1/3) tan−1 (y/x) .
70
!)
we have
p
C = 3(1 + 2a) eiφ
∆0
3(1 + 2a) −iφ
=p
e
C
3(1 + 2a)
p
= 3(1 + 2a) e−iφ
which gives the 3 solution as
1p
3(1 + 2a) ei(φ+ψk ) + e−i(φ+ψk )
3
2p
=−
3(1 + 2a) cos(φ + ψk )
3
(
!
)
p
3 − 27F 2
4(1
+
2a)
2p
1
x
√
tan−1
=−
3(1 + 2a) cos
+ ψk
3
3
Fx 27
xk = −
where k = 0, 1, 2, ψ0 = 0, ψ1 = 2π
and ψ2 = 4π
. Now notice that ysaddle requires taking the
3
3
square root of a real number. This means ysaddle will be real if and only if the argument
under the square root is positive
s
2
Fx
ysaddle = ± (1 − 2b) −
2(a + b)
∈R
⇔ 0 ≤ (1 − 2b) −
Fx
2(a + b)
2
⇔ Fx ≤ Fxsad
√
where Fxsad = 2(a + b) 1 − 2b.
Notice also how the argument inside the tan−1 (·) function contains a square root as well,
which is actually the square root of the discriminant. The three cubic roots xk would be
real if and only if the discriminant is positive
p
√
∆ = 4(1 + 2a)3 − 27Fx2
∈R
⇔ 0 ≤ 4(1 + 2a)3 − 27Fx2
⇔ Fx ≤ Fxcrit
r
4(1 + 2a)3
crit
where Fx =
27
which clearly puts bounds on the forces. This completes the proof.
Next there is a simple but useful Lemma.
71
Lemma 5.6. The function l (as in Theorem 5.5) is monotone in Fx for Fx < Fxcrit .
Proof. We differentiate l with respect to Fx .
!
p
4(1 + 2a)3 − 27Fx2
d
dl
√
=
dFx
dFx
Fx 27
− 1
1
1
= √
4(1 + 2a)3 − 27Fx2 2 (−2Fx ) 27
2
Fx 27
p
4(1 + 2a)3 − 27Fx2 −1
√
+
Fx2
Fx 27
≤0
which is always negative since we assumed Fx < Fxcrit for the square roots to be real and
forcing is assumed to be in the positive direction.
Although the monotonicity of l is trivial, it would prove essential for the next series of
reasoning. It is also easy to see that
Fx = 0
then l = +∞
Fx = Fxcrit then l = 0.
Since the derivative of l is negative this means that l would decrease from +∞ to 0 as
Fx increase from 0 to Fxcrit . This function being monotone means it would decrease to 0
without any oscillations. For short this means
l = ∞ ↓ 0 as Fx = 0 ↑ Fxcrit .
But φ = 31 tan−1 (l), with tan−1 is also monotone over [−∞, +∞]. So similarly we can also
say
φ=
π
↓ 0 as Fx = 0 ↑ Fxcrit
6
monotonically for increasing Fx . This means that for 0 ≤ Fx ≤ Fxcrit we would have
0≤φ≤
π
.
6
We note the following values of the cos(·) function.
cos(0) = 1
cos 2π
=
3
4π
cos 3 =
−1
2
−1
2
cos
π
6
cos
2π
3
4π
3
cos
72
√
=
3
2
+
π
6
π
6
+
=
√
− 3
2
= 0.
From this we can bound cos(·) for the three values of k for 0 ≤ Fx ≤ Fxcrit .
√
3
2π
k=0
≤
cos
φ
+
k
≤ 1
2
3
√
k = 1 −2 3 ≤ cos φ + k 2π
≤ −1
3
2
2π
−1
≤ cos φ + k 3
≤ 0.
k=2
2
We also note that the cos(·) function is monotone on [0, π] and [π, 2π]. But φ + k 2π
3
is in the intervals where cos(·) is monotone, therefore we can say that the three critical
points on the x-axis xk are also monotone in Fx . We can now have bounds on the xk for
0 ≤ Fx ≤ Fxcrit .
√
−2 √
√ 1 + 2a ≤ x0 ≤ − 1 + 2a
3
√
1 √
√ 1 + 2a ≤ x1 ≤ 1 + 2a
3
1 √
0 ≤ x2 ≤ √ 1 + 2a.
3
Using the monotonicity of xk we get that
√
x0 : − 1 + 2a −→
√
x1 :
1 + 2a −→
x2 :
0 −→
√
−2
√
1
3
√
√1
1
3
√
√1
1
3
+ 2a as Fx → Fxcrit
+ 2a as Fx → Fxcrit
+ 2a as Fx → Fxcrit .
The monotonicity of xk means the movements of the xk are always in one direction and
they will never oscillate. We obtain the following
Lemma 5.7. Let Fx > 0, Fy = 0 and b < 21 . These three scenarios hold.
If Fx < Fxcrit we have 3 critical points on the x-axis: x0 < x2 < x1 .
If Fx = Fxcrit we have 2 critical points on the x-axis: x0 < x2 =√x1 .
−2
1 + 2a and x0 → −∞
If Fx > Fxcrit we have 1 critical point on the x-axis: x0 < √
3
monotonically with increasing Fx .
Proof. The three xk are solutions to a cubic equation. This cubic equation was derived
assuming y = 0. The other critical points (xsaddle , ±ysaddle ) were derived assuming y 6= 0.
If Fx < Fxcrit the discriminant of this cubic equation dictates that there should be three
distinct real solution. The bounds on x0 , x1 and x2 show that x0 < x2 < x1 . If Fx = Fxcrit
the discriminant of this cubic equation dictates
at least two repeated
√ that there should be
1
crit
√
solution. It was shown that x1 = x2 = 3 1 + 2a when Fx = Fx . The bound on x0 ,
x1 and x2 shows that x0 < x2 = x1 . If Fx > Fxcrit the discriminant of this cubic equation
dictates that there should only be one real solution. If we can show that x0 < 0 is real
then we are done. For Fx > Fxcrit the l function becomes
p
27Fx2 − 4(1 + 2a)3
√
l=i
.
Fx 27
73
Differentiating the imaginary part gives
!
p
2
3
27F
−
4(1
+
2a)
d
d
x
√
(−il) =
dFx
dFx
Fx 27
√
Fx 27
27Fx2 − 4(1 + 2a)3
=p
1−
27Fx
27Fx2 − 4(1 + 2a)3
> 0,
since we have assumed Fx > Fxcrit . This also shows that the imaginary part of l is monotonically increasing as Fx increases. We also note that
0 ≤ (−il) ≤ 1
because (−il) = 0 when Fx = Fxcrit and (−il) = 1 when Fx = ∞ and the increase in (−il)
is monotone. Using tan−1 (·) defined for complex arguments gives
1
tan−1 (l)
3
1 1
= × i{ln(1 − il) − ln(1 + il)}
3 2
1
= i{ln(1 + ) − ln(1 − )} where 0 ≤ ≤ 1
6
φ=
after letting = −il. Notice that φ now has zero real part, which means we can denote φ
with a real γ by writing
φ = iγ.
We also note that γ will be monotonically decreasing as Fx increases because
d
1
1
[ln(1 + ) − ln(1 − )] =
−
d
1+ 1−
−2
=
(1 + )(1 − )
< 0 for Fx > Fxcrit .
Because it was shown that as Fx increases from Fx = Fxcrit to Fx = ∞, would increase
from = 0 to = 1, which means γ would monotonically decrease. So the critical point
may now be written as
2 √
x0 = − √ 1 + 2a cos (φ)
3
2 √
ei(iγ) + e−i(iγ)
= − √ 1 + 2a
2
3
74
2 √
= − √ 1 + 2a cosh(γ)
3
which means x0 <
Fx = Fxcrit .
−2
√
3
√
1 + 2a and monotonically decreasing. Note that γ = 0 when
Now we consider a special case of the forcing when Fx = Fxsad . This gives
(xsaddle , ±ysaddle ) = (xsaddle , 0)
√
= ( 1 − 2b, 0)
√
⇒ xsaddle = 1 − 2b
which seemingly adds a fourth critical point onto the x-axis. This brings us to the next
Lemma.
Lemma 5.8. Fxsad ≤ Fxcrit holds.
Proof. Proof by contradiction. Assume that Fxsad > Fxcrit . Let Fx = Fxsad which means
(xsaddle , ±ysaddle ) = (xsaddle , 0) as a new critical point on the x-axis. Now we have to show
that (xsaddle , 0) is not one of (x0 , 0), (x1 , 0) or (x2 , 0). The expressions for the critical points
mean we would always have xsaddle > 0. But Fx = Fxsad also implies Fx > Fxcrit which by
Lemma 5.7 means the only critical point is x0 < 0 which is a contradiction.
The next Lemma will be useful in avoiding complicated manipulation of trigonometric
identities when it comes to proving properties about the critical points.
Lemma 5.9. Let Fx > 0, Fy = 0 and b < 12 . If Fx = Fxsad we must have either xsaddle = x1
or xsaddle = x2 . If Fx = Fxsad = Fxcrit then xsaddle = x1 = x2 .
Proof. For strictly positive Fx > 0 some of the bounds on the critical points would have
to be made strict inequalities. This means x0 < 0, x1 > 0, x2 > 0 and xsaddle > 0. By the
time Fx = Fxsad , xsaddle would be a critical point on the x-axis. By Lemma 5.8 we must
always have Fxsad ≤ Fxcrit . If Fxsad < Fxcrit then by Lemma 5.7 there must be three distinct
critical points on the x-axis, so we must have either xsaddle = x1 or xsaddle = x2 (as x0 < 0).
If Fxsad = Fxcrit then again by Lemma 5.7 x1 = x2 and there can only be two critical points
on the x-axis, therefore xsaddle = x1 = x2 .
Now we are ready for one of the main Theorems of this Chapter. The ultimate aim is to
find the nature and position of all the critical points under different values of the forcing
Fx .
Theorem 5.10. Let Fx > 0, Fy = 0 and b < 12 . The positions and nature of the critical
points are as follows
(x0 , 0) well
For any Fx > 0
75
well
(x1 , 0)
sad
(x2 , 0)
hill
and if Fx < Fx
(xsaddle , ±ysaddle ) saddle
or
or
Fxsad
(x1 , 0)
(x1 , 0)
< Fx < Fxcrit
(x2 , 0)
(x , 0)
2
(xsaddle , ±ysaddle )
saddle
well
hill
saddle
nonexistent
(x1 , 0)
unidentified
well
(x1 , 0)
Fx = Fxsad < Fxcrit
(x2 , 0)
hill
(x , 0)
neither
2
(xsaddle , ±ysaddle ) unidentified
(x1 , 0) unidentified
crit
or Fx = Fx
(x2 , 0) unidentified
(x1 , 0) nonexistent
crit
or Fx > Fx
(x2 , 0) nonexistent
√
for √1 − 2b ∈ R1
for √1 − 2b ∈ R2
for √1 − 2b ∈ R1
for
1 − 2b ∈ R2
√
for √1 − 2b ∈ R1
for √1 − 2b ∈ R2
for √1 − 2b ∈ R1
for
1 − 2b ∈ R2
Proof. For the three xk we note that for 0 < Fx ≤ Fxcrit they are elements of the intervals
√
−2 √
x0 ∈ √ 1 + 2a, − 1 + 2a
3
√
1 √
x1 ∈ √ 1 + 2a, 1 + 2a
3
1 √
x2 ∈ 0, √ 1 + 2a .
3
√
Note also that at Fx = Fxsad the associated value of xsaddle (a, b, Fxsad ) = 1 − 2b can only
ever be elements of certain intervals.
√
√
1 √
sad
crit
For Fx < Fx
either
1 − 2b ∈ R1 := √ 1 + 2a, 1 + 2a
3
√
1 √
or
1 − 2b ∈ R2 := 0, √ 1 + 2a
3
√
1 √
and if Fxsad = Fxcrit then
1 − 2b = √ 1 + 2a
3
where R1 and R2 are defined as above. These statements above can be justified as follows.
Lemma 5.8 says Fxsad ≤ Fxcrit . If Fx = Fxsad = Fxcrit then Lemma 5.9 says xsaddle = x1 =
76
√
√
x2 = √13 1 + 2a. Also, 1 − 2b must always live in the regions specified, because we must
√
√
always have 0 < 1 − 2b < 1 + 2a for 0 < b < 21 . Or, to justify it in another way, if
√
1 − 2b > 0 live beyond the regions R1 or R2 then we would have four critical points on
the x-axis which is not possible.
√
Now we see how
√ the critical points collide. If 1 − 2b ∈ R1 then we got to have
xsaddle = x1 . If 1 − 2b ∈ R2 then we got to have xsaddle = x2 . This is because for
Fx = Fxsad < Fxcrit there have to be three distinct critical points on the x-axis as argued
by Lemma 5.7 and Lemma 5.9. From this information new bounds on the xk critical
points may be derived. The bounds for the critical points on the x-axis written compactly,
concisely and definitively for Fx > 0 are
√
√
−2
√3 √1 + 2a ≤ x0 < −√1 + 2a
√1
1 + 2a ≤ x1 <
1 + 2a
Fx ≤ Fxcrit
3
√
1
√
0 < x2 ≤
1 + 2a
3
√
√
x1 > √1 − 2b for √1 − 2b ∈ R1
sad
Fx < Fx
x2 < 1 − 2b for
1 − 2b ∈ R2
√
√
x1 < √1 − 2b for √1 − 2b ∈ R1
Fx > Fxsad
x2 > 1 − 2b for
1 − 2b ∈ R2
√
x1 = xsaddle = √1 − 2b
not necessarily x1 = x2
Fx = Fxsad
x2 = xsaddle = 1 − 2b
(
√
−2
x0 = √
1 + 2a
crit
3
√
Fx = Fx
x1 = x2 = √13 1 + 2a
n
√
−2
crit
x0 < √
1 + 2a
Fx > Fx
3
n
√
√
x1 = x2 = xsaddle = 1 − 2b = √13 1 + 2a
Fx = Fxsad = Fxcrit
Now that the bounds on the critical points for various forces are known, we can deduce their
2
nature. The (xsaddle , ±ysaddle ) is the easiest to prove, taking into account of x2saddle +ysaddle
=
(1 − 2b) we have for the determinant of the Hessian at (xsaddle , ±ysaddle )
2
det H = −4ysaddle
(a + b) < 0
which is definitely a saddle. The Hessian for the critical points (xk , 0) is already diagonal
even with forcing. It is
!
∂ 2 VF
0
3x2k − (1 + 2a)
0
2
∂x
=
H(xk , 0) =
∂ 2 VF
0
x2k − (1 − 2b)
0
∂y 2
and so by using the bounds we derived, and by considering whether the eigenvalues are
both positive (well), both negative (hill) or opposite signs (saddle) we can finally deduce
the nature of all five critical points.
77
The situation can be represented graphically as
Figure 5.1: As Fx increases from 0 to Fxcrit , the x1,2,3 move as shown in the diagram. As Fx
increases from 0 to Fxsad the (xsaddle , ±ysaddle
√ ) meet each other on the x-axis. There are three
− 2b ∈ R1 , then the two (xsaddle , ±ysaddle ) would
possible paths for (xsaddle
, ±ysaddle ). If 1
√
√
√
1
meet in the interval √3 1 + 2a, 1 + 2a and collide into (x2 , 0). If 1 − 2b ∈ R2 , then
√
1
√
the two (xsaddle , ±ysaddle ) would meet in the interval 0, 3 1 + 2a and collide into (x1 , 0).
√
√
If 1 − 2b = √13 1 + 2a, which is also when Fxsad = Fxcrit , then the two (xsaddle , ±ysaddle )
√
would meet at x = √13 1 + 2a and collide simultaneously into (x1 , 0) and (x2 , 0).
78
√
Figure 5.2: This is the case for when 1 − 2b ∈ R1 . When F = Fxsad the two saddles
collide into the right well and turns into a new saddle. At Fxcrit , this newly created saddle
collides into the hill and both disappears. When Well 2 turns into a Saddle here, it is like
creating a new path for the particle to transit to Well 1.
79
√
Figure 5.3: This is the case for 1 − 2b ∈ R2 . When F = Fxsad the two saddles collide into
the hill and turns into a new saddle. At Fxcrit , this newly created saddle collides into the
right well and both disappears. This system behaves in a similar way to a One Dimensional
Potential.
80
√
√
Figure 5.4: This is the case for when 1 − 2b = √13 1 + 2a, which is also when Fxsad =
Fxcrit . At F = Fxsad = Fxcrit the two saddles, hill and right well mutually collide at the same
place and disappears.
5.2.2
Case Fx > 0, Fy = 0 and b ≥
For b ≥ 12 the reasoning is similar to the b <
We have the following Theorem.
1
2
1
2
case, but (xsaddle , ±ysaddle ) does not exist.
Theorem 5.11. Let Fx > 0, Fy = 0 and b ≥ 21 . The positions and nature of the critical
81
points are as follows
(x0 , 0)
(x1 , 0)
Fx < Fxcrit
(x2 , 0)
(x0 , 0)
(x1 , 0)
Fx > Fxcrit
(x2 , 0)
(x0 , 0)
(x1 , 0)
Fx = Fxcrit
(x2 , 0)
well
well
saddle
well
nonexistent
nonexistent
well
neither
neither
This can be graphically conveyed as
Figure 5.5: As Fx increases from 0 to Fxcrit , the saddle collides into the right well and
disappears.
5.3
Case Fx = 0 and Fy > 0
Similarly when forcing is only in the y direction, the cases for b <
considered separately. The case for Fy ≤ 0 is similar.
5.3.1
Case Fx = 0, Fy > 0 and b <
1
2
and b ≥
1
2
have to be
1
2
We have some Lemmas and Theorems which are almost analogous to the case for Fx > 0,
Fy = 0 and b < 12 . Their proofs are very similar and are omitted.
Theorem 5.12. Let Fx = 0, F > 0 and b < 12 . Let Fy be bounded by
Fy ≤ Fysad
and
82
Fy ≤ Fycrit
where
Fysad
√
= 2(a + b) 1 + 2a
r
and
Fycrit =
4(1 − 2b)3
27
then there are five critical points given by
(0, yk ) k = 1, 2, 3
(±xwell , ywell )
where
s
xwell =
(1 + 2a) −
Fy
2(a + b)
2
−Fy
2(a + b)
2 √
2π
yk = − √ 1 − 2b cos ψ + k
3
3
1
ψ = tan−1 (p)
3q
ywell =
4(1 − 2b)3 − 27Fy2
√
p=
Fy 27
k = 0, k = 1, k = 2.
Lemma 5.13. The function p (as in Theorem 5.12) is monotone in Fy for Fy < Fycrit .
Lemma 5.14. Let Fx = 0, Fy > 0 and b < 12 . These three scenarios hold.
If Fy < Fycrit we have 3 critical points on the x-axis: y0 < y2 < y1 .
If Fy = Fycrit we have 2 critical points on the y-axis: y0 < y2 = y1 .
√
−2
If Fy > Fycrit we have 1 critical point on the y-axis: y0 < √
1 − 2b and y0 → −∞
3
monotonically with increasing Fy .
Using the same reasoning as for the x-direction case we have bounds on the three yk for
0 ≤ Fy ≤ Fycrit .
√
−2 √
√ 1 − 2b ≤ y0 ≤ − 1 − 2b
3
√
1 √
√ 1 − 2b ≤ y1 ≤ 1 − 2b
3
1 √
0 ≤ y2 ≤ √ 1 − 2b.
3
83
The monotonicity of yk means
√
y0 : − 1 − 2b −→
√
1 − 2b −→
y1 :
y2 :
0 −→
√
−2
√
1
3√
1
√
1
3√
1
√
1
3
− 2b as Fy → Fycrit
− 2b as Fy → Fycrit
− 2b as Fy → Fycrit
without any oscillations. Now consider the special case when Fy = Fysad . This gives
(±xwell , ywell ) = (0, ywell )
√
= (0, − 1 + 2a)
√
⇒ ywell = − 1 + 2a
√
√
which definitely satisfies − 1 + 2a < − 1 − 2b for 0 < b < 12 . This seemingly adds a
fourth critical point onto the y-axis. We have a Lemma whose method of proof is similar
to Lemma 5.8. But now it does not take long to find numerical examples such that
Fysad < Fxcrit , Fycrit < Fysad and Fysad = Fycrit . These would form the separate sub-cases we
would have to consider.
Lemma 5.15. The following statements hold.
√
√
−2
1. If Fysad < Fycrit , then − 1 + 2a > √
1 − 2b
3
√
2. If Fysad > Fycrit , then − 1 + 2a <
−2
√
3
√
3. If Fysad = Fycrit , then − 1 + 2a =
−2
√
3
√
√
1 − 2b
1 − 2b
√
√
−2
1 − 2b. Let Fy = Fysad . But this
Proof. If Fysad < Fycrit , assume that − 1 + 2a ≤ √
3
means Fy < Fycrit and by Lemma 5.14 there must be three critical points on the y-axis.
√
−2
But monotonicity implies y0 > √
1 − 2b for Fy < Fycrit meaning there would be 4 critical
3
points on the y-axis, which is a contradiction.
√
√
−2
If Fysad > Fycrit assume, that − 1 + 2a ≥ √
1 − 2b. Let Fy = Fysad . But this means
3
Fy > Fycrit and Lemma 5.14 implies that there should only be one critical point on the
√
−2
y-axis. We know that y0 = √
1 − 2b at Fy = Fycrit and yet the monotonicity of y0 means
3
√
−2
y0 < √
1 − 2b for Fy > Fycrit . This would mean 2 critical points on the y-axis which is a
3
contradiction.23
If Fysad = Fycrit then let Fy = Fysad = Fycrit . But Lemma 5.14 says there can only be 2
critical points on the y-axis. But y1 = y2 > 0 and y0 < 0. This means ywell must collide
into y0 , hence the statement of the Theorem.
√
−2
Just like in the x-direction case it can be shown that for Fy > Fycrit , y0 = √
1 − 2b cosh(z) where z
3
is a real number which monotonically decreases with increasing Fy , and yet z = 0 when Fy = Fycrit , which
justifies the idea of this proof.
23
84
Again we are ready for another main Theorem of this Chapter. It is finding the positions
and nature of all the critical points for different values of Fy .
Theorem 5.16. Let Fx = 0, Fy > 0 and b < 12 . The positions and nature of the critical
points are as follows
(0, y1 ) saddle
crit
Fy < Fy
(0, y2 ) hill
(0, y0 )
saddle
Fy < Fysad
(±xwell , ywell ) well
(0, y1 ) nonexistent
crit
Fy > Fy
(0, y2 ) nonexistent
(0, y0 )
well
sad
Fy > Fy
(±xwell , ywell ) nonexistent
(0, y1 ) unidentified
Fy = Fycrit
(0, y2 ) unidentified
(0, y0 )
unidentified
Fy = Fysad
(±xwell , ywell ) unidentified
Proof. Just like in the x-direction case, monotonicity of p is essential in justifying the
following bounds on the critical points for Fy > 0. Note that Lemma 5.15 is used to
determine the bounds on the yk critical points
√
√
−2
√3 √1 − 2b ≤ y0 < −√1 − 2b
√1
1 − 2b ≤ y1 <
1 − 2b
Fy ≤ Fycrit
3
√
1
0 < y2 ≤ √3 1 − 2b
√
Fy < Fysad
y0 > − 1 + 2a
√
Fy > Fysad
y0 < − 1 + 2a
n
√
−2
crit
y0 < √
1 − 2b
Fy > Fy
3
√
Fy = Fysad
y0 = ywell = − 1 + 2a
n
√
crit
y1 = y2 = √13 1 − 2b
Fy = Fy
n
√
√
−2
sad
crit
1 − 2b
y0 = ywell = − 1 + 2a = √
Fy = Fy = Fy
3
Now that the bounds on the critical points are found we can determine their nature.
Similarly (±xwell , ywell ) is the one whose nature is easiest to prove. After noting that
2
x2well + ywell
= (1 + 2a), the determinant of the Hessian at (±xwell , ywell ) gives
det H = 4x2well (a + b) > 0
85
and the second partial derivative in x gives
∂ 2 VF
(±xwell , ywell ) = 2x2well > 0
2
∂x
which is a well by Theorem 5.2. The Hessian matrix for the three critical points on the
y-axis (0, yk ) is already diagonal even with forcing
!
∂ 2 VF
0
yk2 − (1 + 2a)
0
2
∂x
H(0, yk ) =
=
.
∂ 2 VF
0
3yk2 − (1 − 2b)
0
∂y 2
Lemma 5.15 has to be used in conjunction with the bounds on y0 , y1 , y3 and ywell (as
derived in the proof of this Theorem) to determine the nature of the critical points.
This can be shown graphically.
86
Figure 5.6: As Fy increases from 0 to Fycrit the y0 , y1 and y2 move as shown in the diagram.
As Fy increases from 0 to Fysad , the two (±xwell , ysaddle ) meet each other on the y-axis. There
are three possible paths for (±xwell , ywell ). If Fysad < Fycrit , then the two (±xwell , ywell )
√
√
2
√
meet between in the interval − 3 1 − 2b, − 1 − 2b . If Fysad = Fycrit , then the two
√
(±xwell , ywell ) meet at y = − √23 1 − 2b. If Fysad > Fycrit then the two (±xwell , ywell ) meet
√
2
√
in the interval −∞, − 3 1 − 2b .
87
Figure 5.7: At F = Fycrit the top saddle collides into the hill and both then disappears. At
F = Fysad the bottom saddle collides with the two wells and turns into a new well. These
two collisions can occur simultaneously or occur one after the other, depending on whether
we have Fysad < Fycrit , Fysad = Fycrit or Fysad > Fycrit .
5.3.2
Case Fx = 0, Fy > 0 and b ≥
1
2
The case for Fy > 0, b ≥ 12 is slightly different in the sense that we have to consider the
discriminant of the cubic equation with one real solution for the potential. We have the
last main Theorem in this Chapter.
88
Theorem 5.17. Let Fx = 0, Fy > 0 and b ≥ 12 . The positions and nature of the critical
points are as follows
(0, y0 )
saddle
sad
Fy < Fy
(±xwell , ywell ) well
(0, y0 ) = (±xwell , ywell ) unidenitified
Fy = Fysad
(0, y0 )
well
sad
Fy > Fy
(±xwell , ywell ) nonexistent
Proof. The simultaneous equations we have to solve are
∂VF
= x(x2 + y 2 ) − (1 + 2a)x = 0
∂x
∂VF
= y(x2 + y 2 ) − (1 − 2b)y + Fy = 0.
∂y
(5.3)
(5.4)
Equation 5.3 holds if either x = 0 or (x2 +y 2 )−(1+2a) = 0. The case for (x2 +y 2 )−(1+2a) =
0 gives (±xwell , ywell ) as a critical point in similar way as before. The case for x = 0 reduces
Equation 5.4 to
y 3 − (1 − 2b)y + Fy = 0.
It is this resulting cubic equation which forms the next series of discussions. The required
expressions in solving this cubic equation are
∆0 = 3(1 − 2b),
∆1 = 27Fy ,
∆ = 4(1 − 2b)3 − 27Fy2 .
Since b ≥ 21 the discriminant of this cubic equation is strictly negative meaning ∆ < 0,
which means there can only be one real solution. All the solutions whether complex or real
are given by
1 iψk
−iψk ∆0
e C +e
where
yk = −
3
C
ψ0 = 0, ψ1 = 2π/3, ψ2 = 4π/3
s
√
−27∆
3 ∆1 +
C=
.
2
Notice that the ∆ < 0, C and ∆0 are all real numbers. Since there can only be one real yk
solution this has to be y0 , as y0 would just be a sum of real numbers. This then means y1
and y2 would be complex conjugate solutions. Now written explicitly we have
v
q
u
u
3 27Fy +
(27Fy2 − 4(1 − 2b)3 ) × 27
t
C=
2
89
which
for
b ≥ 12 is clearly monotonically increasing in Fy for Fy > 0. This is because the
√
√
· and 3 · are both monotone with respect to their own argument. This means we can say
dC
≥0
dFy
dy0
1
∆0 dC
=−
1− 2
dFy
3
C
dFy
≤0
because (−∆0 ) > 0 for b ≥ 12 . This means y0 would always monotonically decrease with
increasing Fy . We also know that at Fy = Fysad the two wells on the sides become
√
(±xwell , ywell ) = (0, − 1 + 2a).
Note that earlier when x = 0 was imposed in Equation 5.3 and 5.4 we were reduced with a
cubic equation that only admits one real solution in y. This means there can only be one
critical point on the y-axis so we must have
√
y0 = − 1 + 2a
at Fy = Fysad . This justifies the following bounds
Fy < Fysad
Fy = Fysad
Fy > Fysad
√
y0 > − 1 + 2a
√
y0 = − 1 + 2a
√
y0 < − 1 + 2a
which can be justified by the monotonicity of y0 . This together with the Hessian can allow
us to identify the nature of the critical points.
Again the situation can be represented graphically.
90
Figure 5.8: At Fy = F sad the saddle collides with the two wells and turns into a new well.
5.4
Case Fx 6= 0 and Fy 6= 0
Consider the forcing
F=
Fx
Fy
=
F cos φ
F sin φ
where F =
q
Fx2 + Fy2
so far we have only studied the case when φ = 0◦ , 90◦ , 180◦ , 360◦ . Now we consider the
case for forcing in a general direction. The critical points are given by solutions to
∂VF
= x(x2 + y 2 ) − (1 + 2a)x + Fx = 0
∂x
∂VF
= y(x2 + y 2 ) − (1 − 2b)y + Fy = 0.
∂y
(5.5)
(5.6)
The arguments presented next can actually apply for any a, b, Fx and Fy regardless of
whether they are positive or negative. Notice that if φ ∈
/ {0◦ , 90◦ , 180◦ , 360◦ } then Fx 6= 0
and Fy 6= 0. This means x = 0 or y = 0 must not appear in any solution. This is the
assumption used. Solving Equation 5.6 for x2 gives
y(x2 + y 2 ) − (1 − 2b)y + Fy = 0
Fy
(x2 + y 2 ) − (1 − 2b) +
=0
y
91
Fy
y
F
y
− y2
x2 = (1 − 2b) −
y
(x2 + y 2 ) = (1 − 2b) −
(5.7)
and substituting Equation 5.7 into Equation 5.5 gives
s
s
Fy
Fy
Fy
− y 2 (1 − 2b) −
− (1 + 2a) (1 − 2b) −
− y 2 + Fx = 0
(1 − 2b) −
y
y
y
s
Fy
Fy
2
(1 − 2b) −
− y (1 − 2b) −
− (1 + 2a) + Fx = 0
y
y
2
Fy
Fy
2
−2b −
(1 − 2b) −
−y
− 2a − Fx2
y
y
2
Fy
Fy
2
(1 − 2b) −
−y
2(a + b) +
− Fx2
y
y
2
Fy
3
− Fx2 y
y(1 − 2b) − Fy − y
2(a + b) +
y
Fy Fy2
3
2
y(1 − 2b) − Fy − y
4(a + b) + 4(a + b) + 2 − Fx2 y
y
y
y(1 − 2b) − Fy − y 3 4(a + b)2 y 2 + 4y(a + b)Fy + Fy2 − Fx2 y 3
=0
=0
=0
=0
=0
which rearranges into a fifth degree polynomial in y
y 5 [−4(a + b)2 ]
+ y 4 [−4(a + b)Fy ]
+ y 3 [(1 − 2b)4(a + b)2 − Fy2 − Fx2 ]
+ y 2 [(1 − 2b)4(a + b)Fy − Fy 4(a + b)2 ]
+ y 1 [(1 − 2b)Fy2 − Fy 4(a + b)Fy ]
+ y 0 [−Fy3 ]
=0
(5.8)
which can only be solved numerically. The five solutions for Equation 5.8 are denoted by
yi
where i = 1, 2, 3, 4, 5.
92
Now consider Equation 5.5
x(x2 + y 2 ) − (1 + 2a)x + Fx = 0
x=
=
x2
+
−Fx
− (1 + 2a)
−Fx
y2
(1 − 2b) −
Fy
y
− (1 + 2a)
where we have used Equation 5.7 for the last line. This means the x part of the final set
of solutions is
xi =
−Fx
(1 − 2b) −
Fy
yi
− (1 + 2a)
where i = 1, 2, 3, 4, 5.
Notice how the calculation is not straightforward if one feeds yi into Equation 5.7 by means
of
s
Fy
− yi2 .
xi 6= (1 − 2b) −
yi
We do not know whether to take the positive or negative solution. Any easy counting
shows that one obtains too many solutions. These five solutions (xi , yi ) are sorted into
wells, hills and saddles. Although an explicit value for the critical forcing cannot be given
analytically, an educated guess can be made
F crit = min Fxsad , Fxcrit , Fysad , Fycrit
(5.9)
that is because a critical force in a general direction must encompass all the other directions.
Here are some real examples of the critical points, as the force is being changed during half
a period of an oscillatory potential.
93
Figure 5.9: The critical points move very little here.
Figure 5.10: The critical points have a more extreme trajectory here.
94
Figure 5.11: Notice that the use of F crit as a critical force is just an educated guess. Here
the system is so close to criticality the saddle is almost colliding with the hill.
5.5
Remarks on Mexican Hat
We give an example of how the Mexican Hat Toy Model look like.
95
Figure 5.12: An example of the potential VF (x, y) = 41 r4 − 21 r2 − ax2 + by 2 + Fx x + Fy y
p
where r = x2 + y 2 . Here a = 0.1, b = 0.1, Fx = 0.1 and Fy = 0. Notice there are two
saddles just ahead of the hill. The well on the right is higher than the well on the left.
5.5.1
Beyond Criticality
At first glance since the critical points can all be found numerically, one may ask why
studied the cubic formula. This actually provided exact analytic information about the
system near and beyond criticality and in the extremal cases as well. Besides, stochastic
resonance is studied when the forcing is small enough such that the topology of the potential
does not change significantly. This is because if the forcing is too large (beyond criticality)
then transitions are almost certain, and there is little point to consider stochastic resonance
in this case.
5.5.2
Numerical Problems
When the critical points are numerically found, they were fed back into Equations 5.5 and
5.6 and were correct to 10−9 . But a few problems remain. In simulations when the angle
of the forcing
Fy
−1
φ = tan
Fx
is changed from φ = 0◦ to φ = 90◦ the potential was continuously changing from a system
needing to solve third order roots to a system needing to solve fifth order roots. This
96
means the numerical algorithms for solving the quintic polynomial (Equation 5.8) became
very unstable when the system is close to solving a cubic equation.24 No further numerical
investigation is necessary as analytic results are available to interpolate the correct solution.
5.5.3
Comparison with One Dimensional Case
The one dimensional potential is
VF =
x2
x4
− a + Fx
4
2
and their critical points are given by solutions to the equation
∂VF
= x3 − ax + F = 0
∂x
which when compared to the solutions in the Mexican Hat yields the critical points as
√ 3
2√
4a − 27F 2 2π
1
−1
√
xk = −
tan
+
k
.
a cos
3
3
3
F 27
Similarly if we want three critical points, the tan−1 (·) must take real arguments, which
means
r
4a3
F < F crit =
27
and the nature of the critical points are
x0
x1
F < F crit
x2
x0
crit
x1
F >F
x2
x0
crit
x1
F =F
x2
well
well
hill
well
nonexistent
nonexistent
well
unidentified
unidentified
which is similar to the Fx > 0, Fy = 0 and b ≥
by
1
2
case. These critical points are bounded
x0 < x2 < x1 .
24
Algorithms used in the roots(·) function in MatLab.
97
This was a calculation not studied in the paper by Benzi et al [4] and other literature. In
the paper [4] it was assumed that the forcing is so small the hill is very near to x2 = 0. In
[4], the escape times were defined by Equations 1.5 and 1.6, where having reached the hill
(which is assumed to be at x2 = 0) is sufficient for escape. Our calculations show that the
hill actually moves as the potential oscillate. Taking this into account may give a better
approximation of the exit times than in [4].
98
Chapter 6
Numerical Methods
Let us remind ourselves of the unperturbed potential of the Mexican Hat Toy Model.
p
1
1
V0 (x, y) = r4 − r2 − ax2 + by 2 where r = x2 + y 2 .
4
2
The SDE we want to study and simulate is
Ẋt = −∇V0 + F cos Ωt + Ẇt
where Wt is a two dimensional Wiener process. When this SDE is expressed for the separate
x and y components we have
∂V0
+ Fx cos Ωt dt + dwx
dx = −
∂x
∂V0
dy = −
+ Fy cos Ωt dt + dwy
∂y
where is the noise level and wx and wy are two independent Wiener processes. When this
SDE is numerically approximated by the Euler method we have
tn = tn−1 + tstep
p
∂V0
xn = xn−1 + −
(xn−1 , yn−1 ) + Fx cos(Ωtn−1 ) tstep + tstep ξx
∂x
p
∂V0
(xn−1 , yn−1 ) + Fy cos(Ωtn−1 ) tstep + tstep ξy
yn = yn−1 + −
∂y
for the iterative scheme, where ξx and ξy are two independent normal random variables.
The level of precision for the numerics are estimated assuming the Euler method is being
used in simulations. More on numerical solutions to SDEs can be found in [61].
When the escape times are measured for an oscillatory potential, a fixed radius R is
defined around the well which moves with the well. The two parameters we need to consider
are
tstep
and R
99
and derive appropriate values for them. The particle is defined as having entered a well
when it enters the area covered by the radius R around the well. The time difference
between entering the first well and the second is measured as the escape time from the first
well. The idea behind the considerations is to identify possible sources of error for tstep
and R and eliminate them. The consequence is that tstep is bounded by six bounds on the
time step
tstep ≤ min {t1 , t2 , t3 , t4 , t5 , t6 }
where the time step tstep has to meet six different conditions. Similarly the radius is
bounded in the following way
R ≤ min {R1 , R2 } .
Although these theories are not rigorous, it gives a fair idea of the level of precision that
is needed. It is assumed in the following that the same precision needed for measuring
escape times is also precise enough for studying the six measures of resonance. Although
the Mexican Hat is a two dimensional system, part of the numerical theory is derived on
a general number of dimensions Rr .
6.1
Basic Conditions - Estimating tstep ≤ t1, tstep ≤ t2
and tstep ≤ t3
When the potential oscillates, one period T is achieved when
⇒
ΩT = 2π
T =
2π
Ω
and the time step tstep has to be precise enough such that the potential is well presented.
Thus it is reasonable to take as the first bound
t1 =
2π
ΩN1
where N1 is an appropriately large number say N1 = 1000. Denote by tend for the end
time. This is the time the simulation is being run for. At least 1000 transitions need to be
detected, which makes the following a reasonable choice
tend = 1000 × max τ + min τ
wj (t)
0≤t≤T
wj (t)
0≤t≤T
where τ is the predicted escape time as given by Kramers’ formula, wj (t) where j = 1, 2, . . .
are the positions of the wells at time t. Thus the tend is 1000 times the minimum and
100
maximum predicted escape times over all wells over one period. Trivially, the time step
has to be smaller than the shortest predicted escape time so the second bound is
t2 =
1
min τ
N2 wj (t)
0≤t≤T
where we use N2 = 1000. Most of the time, the particle is near the bottom of the well,
then one can approximate the iteration scheme to
tn = tn−1 + tstep
p
xn = xn−1 + tstep ξx
p
yn = yn−1 + tstep ξy
which allows the distance travelled by the particle in one increment, in the time of one
time step tstep to be given as
q
∆z = (xn − xn−1 )2 + (yn − yn−1 )2 .
Let the critical points be given by
c1 (t), c2 (t), . . .
If the particle starts at the well and ever reaches a hill or saddle, then transition to the
other well is almost certain. Thus travelling from the well to another critical point should
be almost impossible in a single time step tstep , that is one increment ∆z. The length of
this forbidden jump is
l1 =
min
wi (t)6=cj (t)
0≤t≤T
|wi (t) − cj (t)|
which is the minimal distance from the wells to any other critical points (which are not
wells) over one period. But the normal random variables ξi are normal distributed in
N (0, 1). This means
Z ∞ −x2 /2
e
√
P (ξx > r) =
dx
2π
r
and the joint distribution is given by
Z ∞ Z ∞
P (ξx > r1 , ξy > r2 ) =
n(x, y) dx dy
x=r1
y=r2
where
1
1 2
2
n(x, y) =
exp − (x + y ) .
2π
2
101
This gives
q
P (∆z > l1 ) = P
ZZ
=
l1
ξx2 + ξy2 > √
tstep
√
(x,y):
Z
θ=2π
Z
l
x2 +y 2 > √t1
n(x, y) dx dy
step
∞
1 −r2 /2
e
r dr dθ
2π
θ=0
(
2 )
1
l1
= exp −
√
2 tstep
=
√ l1
tstep
and the number of increments achieving such a direct jump must be almost zero in one
session of the simulation. So
tend
P (∆z > l1 ) < N3
tstep
where N3 needs to be smaller than one for example N3 = 0.1. Rearranging the expression
to
(
2 )
tend
1
l1
exp −
= N3
√
tstep
2 tstep
and solving for the time step tstep we get
tstep = J(, l1 , N3 , tend )
where J is a function which numerically inverts the expressions to give the time step
required. This gives the third bound as
t3 = J(, l1 , N3 , tend ).
6.2
Increment Conditions
The bound t3 on the time step hinges on finding bounds on the length of a single increment
q
∆z = (xn − xn−1 )2 + (yn − yn−1 )2
that is the distance travelled in the time of one time step tstep . This is now considered
again but in a more general setting in Rr a general number of dimensions.
102
6.2.1
Increment Theory - Developing W (S, l)
Let V0 : Rr −→ R be a real function from Rr to R. Its gradient with a periodic forcing
and noise gives rise to the SDE
Ẋt = −∇V0 + F cos(Ωt) + ẇt
where Xt is a trajectory in Rr , F is the force in Rr and wt is a vector of r independent
Wiener processes. Thus
Xt = (x1 (t), x2 (t), . . . , xr (t))
F = (F1 , F2 , . . . , Fr )
wt = (w1 (t), w2 (t), . . . , wr (t)).
This trajectory can be numerically approximated with the Euler scheme
tn+1 = tn + tstep
p
∂V0 n n
n+1
n
n
(x1 , x2 , . . . , xr ) + Fi cos(Ωtn ) tstep + tstep ξi
xi = xi + −
∂xi
(6.1)
where the partial derivative is evaluated at the previous iteration step (xn1 , xn2 , . . . , xnr ) and
ξi is a normal random variable. Rearranging Equation 6.1 gives
p
∂V0 n n
n+1
n
n
xi − xi ≤ −
(x1 , x2 , . . . , xr ) + Fi cos(Ωtn ) tstep + tstep ξi .
(6.2)
∂xi
Notice that we can make the following bound
∂V0 n n
∂V0 n n
n
(x1 , x2 , . . . , xr ) + Fi cos(Ωtn ) tstep ≤ −
(x1 , x2 , . . . , xnr ) + Fi cos(Ωtn ) tstep
−
∂xi
∂xi
∂V0 n n
n
≤ −
(x , x , . . . , xr ) + |Fi cos(Ωtn )| tstep
∂xi 1 2
∂V0 n n
n
≤ −
(x , x , . . . , xr ) + |Fi | tstep
∂xi 1 2
∂V0
≤ max
+ |Fi | tstep
(6.3)
S
∂xi
where S is a set that is large enough such that
(xn1 , xn2 , . . . , xnr ) ∈ S ⊂ Rn .
We define
p
∂V0
∆xi = max
+ |Fi | tstep + tstep ξi = Ai + Bξi .
S
∂xi
103
(6.4)
By using the triangle inequality we can bound Equation 6.2 and 6.4 in the following way
p
∂V0 n n
(x1 , x2 , . . . , xnr ) + Fi cos(Ωtn ) tstep + tstep ξi
xn+1
− xni ≤ −
i
∂xi
p
∂V0
|∆xi | ≤ max
+ |Fi | tstep + tstep ξi
S
∂xi
and by using 6.3 we know that
xn+1
− xni ≤ |∆xi | .
i
This means the total increment in the time of one iteration is bounded by
q
q
2
n+1
n+1
n 2
n 2
n+1
n
x1 − x1 + x2 − x2 + . . . + |xr − xr | ≤ ∆x21 + ∆x22 + . . . + ∆x2r
leading us to define
∆z =
q
∆x21 + ∆x22 + . . . + ∆x2r .
Now introduce a new variable
∆xi
ηi =
∼N
B
Ai
,1
B
which is a normal random variable with mean Ai /B and variance one. Let
η = η12 + η22 + . . . + ηr2
and λ =
1
2
2
2
A
+
A
+
.
.
.
+
A
1
2
r
B2
which means η is a sum of the squares of r normal random variables with variance one
and λ is the sum of their means. This means η is noncentral chi-squared distributed with
r degrees of freedom. Its CDF is
√ √
P (η ≤ x) = 1 − Q r2 ( λ, x)
where Qm is the Marcum Q-function. Thus the total increment is distributed by
P (∆z > l) = P (∆z 2 > l2 )
= P (ηB 2 > l2 )
l2
=P η> 2
B
√
l
λ, √
= Q r2
tstep
where l would be the length of a forbidden increment. The number of such forbidden jumps
must stay below an appropriate number N3 , that is
tend
P (∆z > l) = N3
tstep
104
(6.5)
and Equation 6.5 has to be numerically inverted to give the required time step tstep
tstep = W (S, l)
where a different region S and length l fulfilling different criteria is used to calculate a
different bound on the time step. Note that we choose N3 = 0.1.
6.2.2
Absence of Large Jumps - Estimating tstep ≤ t4 and tstep ≤ t5
Let the set S1 be given by
n √
√
o
√
√
S1 =
− 1 + 2a, 0 , + 1 + 2a, 0 , 0, − 1 − 2b , 0, + 1 − 2b , (0, 0)
which is the positions of all the critical points as they would be when the forcing is zero
F = 0. This means
max
S1
∂V0
= 0 and
∂x
max
S1
∂V0
= 0.
∂y
The set S1 can be used as an approximation for small forcing. Equation 6.4 now becomes
p
∆x = |Fx | tstep + tstep ξx
p
∆y = |Fy | tstep + tstep ξy
and we want a time step tstep small enough such that almost every value of ∆z =
is bounded by
p
∆z = ∆x2 + ∆y 2 ≤ l1
p
∆x2 + ∆y 2
and this time step is given by t4 below
t4 = W (S1 , l1 ).
Now we remind ourselves that if ζ is an exponentially distributed random variable, its
PDF, CDF, mean and variance are given by
Z
P (ζ ∈ A) =
λe−λx dx
ZAx
P (ζ ≤ x) =
λe−λx dx = 1 − e−λx = F (x)
−∞
1
hζi =
λ
1
var(ζ) = 2
λ
105
where λ is the parameter associated with the exponential distribution. Now define the
height
h0 = max Vt (wj (t))
wj (t)
0≤t≤T
which is the maximum height any well can ever reach. Now define the expression
∆Vh = h1 − h0
where h1 is chosen so high that the particle will probably never reach there. Even if it
starts from the highest possible well the chances are still very slim. Now we try to estimate
what this height h1 may be. From Freidlin-Wentzell we know that the escape time from
V = h0 to V = h1 is roughly
2
τh ≈ e2∆Vh /
and we want the time it takes to reach V = h1 to be significantly more than the duration
of the simulation
τh tend
which means a reasonable estimate would be
2
N2 tend = e2∆Vh /
2∆Vh = 2 ln(N2 tend )
1
⇒ h1 = 2 ln(N2 tend ) + h0
2
where as before we choose N2 = 1000. The escape times leaving V = h0 and arriving at
V = h1 is exponentially distributed. The average of them would be
2
hτh i = N2 tend = e2∆Vh /
and so by using the CDF of the exponential distribution we can say
P (τh < tend ) = 1 − e−1/N2
= 0.632 for
= 0.095 for
= 0.01 for
= 0.001 for
N2
N2
N2
N2
=1
= 10
= 100
= 1000.
So if we choose N2 ≥ 1000 the chances of reaching V = h1 are less then one in a thousand.
Define the set
S2 = {x ∈ Rr : [V0 (x) − F x cos(Ωt)] ≤ h1 : 0 ≤ t ≤ T }
106
= {x ∈ Rr : Vt ≤ h1 : 0 ≤ t ≤ T }
which is the set of all the points below Vt ≤ h1 over the time of one period. The forbidden
increment is taken as the same as last time
l2 = l1
so the second bound is
t5 = W (S2 , l2 ).
6.3
Stability and Radius Conditions
Stability of the trajectory in the context of this thesis is for the time step tstep to be small
enough such that the simulated discrete trajectory is a good enough approximation of a
physical continuous trajectory. For example consider the following trajectories for a particle
falling down to the well of the Mexican Hat starting at (xstart , ystart ) = (−0.75, −0.75).
Figure 6.1: The trajectory is so unstable the particle even transits to the other well.
107
Figure 6.2: The trajectory is more stable but the particle now oscillates near the well.
Figure 6.3: The trajectory is sufficiently stable here.
It is the aim of this section to study these stability problems. The Euler method is effectively a discrete iterative map with some operations. We have the following.
108
Lemma 6.1. Let (X, k · k) be a vector space endowed with the norm k · k over the complex
scalar field F. Let M be an operator M : X −→ X with the property kM xk ≤ kM kkxk
and kM k ≥ 0. Let x ∈ X be part of an iterative scheme
√
xn = M xn−1 + t ξn−1
where ξn−1 ∈ X is a term which depends on the iterative step n. The entire term xn is
then bounded by
n−1
√ X
n
kM ki kξn−1−i k
kxn k ≤ kM k kx0 k + t
i=0
where x0 and ξ0 are the starting (first) steps.
Proof. Rewrite the iterative scheme with a new operator
√
xn = M xn−1 + t ξn−1
= Mn0 xn−1
where Mn0 is the total operator which depends on the nth step. In general Mn0 is not
commutative so we write
0
xn = Mn0 Mn−1
. . . M10 x0
and consider just one operation on Mn0
√
Mn0 x = M x + t ξn−1
√
kMn0 xk = M x + t ξn−1
√
≤ kM xk + t kξn−1 k
√
≤ kM k kxk + t kξn−1 k .
So we have an iterative expression
√
kMn0 xk ≤ kM k kxk + t kξn−1 k
and iterating Equation 6.6 gives
0
kxn k = Mn0 Mn−1
. . . M10 x0
n−1
√ X
≤ kM k kx0 k + t
kM ki kξn−1−i k
n
i=0
which completes the proof.
Remark 6.2. There are a lot of remarks to say about this simple Lemma.
109
(6.6)
1. Nowhere was it assumed M is bounded, linear or commutative.
2. Notice that kM k was not defined. It could be the usual operator norm or something
else.
3. Notice that we are only truncating the bound as in Equation 6.6 and NOT truncating
the original iterative Euler scheme.
4. Notice that M 0 = 0 for the zero vector. Notice also that M (λx) = λM (x) where λ
is a scalar was not assumed.
5. Notice that a norm on kMn0 k was not needed. This is because such a norm would
have to satisfy kMn0 xk ≤ kMn0 k kxk which implies Mn0 0 = 0. But with the random
vector ξn−1 in Mn0 this cannot be achieved.
Consider the two dimensional Mexican Hat system with z = (x, y) ∈ R2 . Let the potential
be stationary by freezing it as it would be at a certain fixed point in time t = tf ix . The
gradient then becomes
!
∂V0
− Fx cos Ωtf ix
∂x
∇Vt=tf ix (z) =
∂V0
− Fy cos Ωtf ix
∂y
where tf ix is a constant point in time. The Euler method can now be rewritten as
tn = tn−1 + tstep
p
∂V0
xn = xn−1 + −
+ Fx cos Ωtf ix tstep + tstep ξx
∂x
p
∂V0
yn = yn−1 + −
+ Fy cos Ωtf ix tstep + tstep ξy
∂y
and we recast it into vector notation by writing
tn = tn−1 + tstep
p
zn = zn−1 − tstep ∇Vt=tf ix (zn−1 ) + tstep ξn−1
(6.7)
where
zn =
xn
yn
.
Now we apply Lemma 6.1 to Equation 6.7. Clearly the operator as mentioned in Theorem
6.1 is
M (z) = z − tstep ∇Vt=tf ix (z)
110
and this operator has to satisfy kM zk ≤ kM k kxk. The usual operator norm would suffice
with some restrictions
kM zk
kM kS = sup
z∈S kzk
z6=0
where S ⊂ R2 is a strict and suitable subset of the whole space. This is because in general,
the expression kM zk is unbounded on R2 . For example in our case ∇V is unbounded.
This would give
kM zk = kzk
kM zk
kM zk
≤ kzk sup
≤ kM kS kzk .
kzk
z∈S kzk
z6=0
But we also need the condition M 0 = 0 for the zero vector (see Remark 6.2). This means
we have to shift the coordinates to
znew = zold − zwell
where the well is now the new origin. Now we can apply Lemma 6.1 to Equation 6.7 and
get
kzn k ≤
kM knS
n−1
X
p
kM kiS kξn−1−i k
kz0 k + tstep
i=0
where the starting position is
z0 = (xstart , ystart ) − (xwell (tf ix ), ywell (tf ix ))
where (xstart , ystart ) is the starting position and (xwell (tf ix ), ywell (tf ix )) is the position of
one well at time t = tf ix
6.3.1
Stability Problems
Lemma 6.1 has allowed us to rewrite the Euler method, expressed in Equation 6.7, as
p
|zn | = M (zn−1 ) + tstep ξn−1
n−1
X
p
≤ kM knS kz0 k + tstep
kM kiS kξn−1−i k
(6.8)
i=0
where if = 0 we would reduce back to the deterministic system
|zn | ≤ kM knS kz0 k .
It is known that if the time step tstep is too large then even the deterministic trajectory is
unstable. A stable time step tstep is one which gives
kM kS ≤ (1 − δ)
(6.9)
where 0 < δ < 1 and the particle would settle at the bottom of the well as n −→ ∞. There
are two approaches to the problem here.
111
1. Fix the set S and solve for the time step tstep such that kM kS ≤ (1 − δ) holds. This
is solving for time.
2. Fix the time step tstep and solve for the set S such that kM kS ≤ (1 − δ) holds. This
is solving for space.
and each approach hinges on the assumption that the S and tstep we fixed to begin with is
a good and stable choice. We can only solve for time or space but not both. This is also
complicated by the fact that the following estimate is too rough
z − tstep ∇Vt=tf ix (z)
kzk + tstep ∇Vt=tf ix (z)
kM zk
≤
≤
≤ 1 + δ0
kzk
kzk
kzk
(6.10)
where δ 0 > 0.
6.3.2
Estimating R ≤ R1 and R ≤ R2
Now two bounds on the radius R is derived. The first bound is
1
|wj (t) − ci (t)|
R1 = min
wj (t),ci (t)
2
wj 6=ci
0≤t≤T
which is half the distance from the wells to all critical points, for all wells, over one period.
This is such that there is always exactly one critical point inside the region covered by the
radius R. The second bound comes from considering Equation 6.8
zn ≤
kM knS
n−1
X
p
kz0 k + tstep
kM kiS kξn−1−i k
i=0
and seek a bound on the variance of the random part. The variance of the random part is
!
n−1
n−1
X
X
p
i
2
var tstep
kM k2i
kM kS kξn−1−i k = tstep
S var (kξn−1−i k) .
i=0
i=0
Notice how each of the kξn−1−i k is χ distributed with r = 2 degrees of freedom. Its variance
is 25
σ 2 = var (kξn−1−i k)
=r−
25
!2
√ Γ r+1
2
2
Γ 2r
Let ξx and ξy be independently and normally distributed in N (0, 1). Let ζ =
q
ξx2 + ξy2 and η = ξx2 +ξy2 ,
then ζ is χ distributed and η is χ2 distributed. Notice that we want the variance of ζ and NOT the variance
of η, therefore only the χ distribution is needed.
112
= 0.4292 for r = 2
which means the variance of the random part is bounded by
!
n−1
n−1
X
X
p
i
2
2
var tstep
kM kS kξn−1−i k = tstep
kM k2i
S σ
i=0
i=0
2 2
≤ σ tstep
∞
X
kM k2i
S
i=0
= 2 σ 2 tstep
1
1 − kM k2S
which means the random part has bounded variance, even if one considers infinite time.
The error is constantly cancelling out with itself, and the Euler method can run for a very
long time and still be stable. The radius must be similar to the size of this variance. Define
tosc
step = min {t1 , t2 , t3 , t4 , t5 }
which is the smallest of all the time steps we have derived so far. Assume that the operator
can indeed be bounded by kM kS ≤ (1 − δ), then we can define a second bound R2 on the
radius as
!
n−1
X
p
kM kiS kξn−1−i k
R22 = var tstep
i=0
1
≤ 2 σ 2 tstep
1 − kM k2S
1
= 2 σ 2 tstep
1 − (1 − δ)2
where we have used kM kS ≤ (1 − δ). This gives R2 as
s
2 σ 2 tosc
step
R2 =
2δ − δ 2
after choosing a suitable value for δ say δ = 0.01, then a reasonable choice on the radius
R could be
R = min {R1 , R2 } .
6.3.3
Estimating tstep ≤ t6
Define the set
S3 = {x ∈ Rr : |wj (t) − x| ≤ R : 0 ≤ t ≤ T, ∀wj (t)}
113
which is the set of points where the distance to a well is less than the radius over all times
and all wells. The new jump size we do not want to see during our simulation is l3 = R
l3 = R
which gives another condition on the time step as
t6 = W (S3 , l3 ).
The idea behind this final condition is so that the time step tstep is small and precise enough
such that the region around the well can capture it.
6.4
Selection of Parameters
Six conditions on the time step and two bounds on the radius were developed. These give
the recommended values for R and tstep as
R = min {R1 , R2 }
tstep = min {t1 , t2 , t3 , t4 , t5 , t6 } .
These are not rigorous estimates and hence can only be used as a guideline. We performed
several checks of consistency of our simulations at a different level of precision before
comparing them with the theoretical result. Notice tstep and R are just some of the considerations we have to make when choosing a set of parameters to use for the simulations.
More details are given below.
6.4.1
Selection of Parameters - Simulation
In Chapter 7 we will introduce the simulations which we are going to do. Notice that the
SDE we want to simulate can be rewritten as
∂V0
+ F cos φ cos Ωt dt + dwx
dx = −
∂x
∂V0
+ F sin φ cos Ωt dt + dwy .
dy = −
∂y
The following parameters will be fixed with the values
a = 0.15 b = 0.1 F = 0.7F crit
Ω = 0.001
and and φ will systemically vary by going through all possible combinations of
= 0.15, 0.16, . . . , 0.30 and φ = 0◦ , 75◦ , 78◦ , 81◦ , 84◦ , 87◦ , 90◦
and the following value of the time step and radius is used
tstep = 0.014 R = 0.19.
We give some reasons as to why these parameters were chosen.
114
6.4.2
Selection of Parameters - Validity of Kramers’ Formula
There is the validity of Kramers’ formula. Consider the graphs below.
Figure 6.4: Notice that transitions tend to occur near the saddles. This is when Kramers’
formula gives a good approximation for the escape rates and escape times.
Figure 6.5: For higher noise levels transitions would occur near the hill, which is close to
the origin. Kramers’ formula is not a good approximation here.
115
Note that in Figures 6.4 and 6.5 the position of the hill is near the origin and the positions
of the saddles are near the y-axis. For very high noise levels Kramers’ formula would start
to fail as an approximation to the escape times and rates. This is when the particle tend to
transit through both the saddles and the hill. A level of subjective judgement is required to
gauge how good an approximation Kramers’ formula is. Nevertheless it has been checked
that at = 0.30, that Kramers’ formula is good enough an approximation for all angles.
This checking was also done for an unperturbed static potential, a static potential with
maximal forcing and an oscillating potential.26
6.4.3
Selection of Parameters - Adiabatic Approximation
There is a reason why we need Kramers’ formula to be valid, that is a good approximation to
the escape times. Recall that in Chapter 3 theories about escape times from an oscillatory
potential was developed. One of the main results were the continuous time invariant
measures for an oscillatory potential (see Corollary 3.9) and the PDF for the escape times
(see Theorem 3.13). The invariant measures are
Rt
RT
p(s)g(s)
ds
p(s)g(s) ds
ν − (t) = 0
+ 0
g(t)
g(t) (g(T ) − 1)
RT
Rt
q(s)g(s) ds
q(s)g(s) ds
0
+ 0
ν + (t) =
g(t)
g(t) (g(T ) − 1)
where
t
Z
g(t) = exp
p(u) + q(u) du
0
and the PDFs are
Z t
p− (t, u) = R−1+1 (t) exp −
R−1+1 (s) ds
u
Z t
p+ (t, u) = R+1−1 (t) exp −
R+1−1 (s) ds
u
where we will make the approximation
Z
1 T
ptot (t) =
p− (t + u, u)m− (u) + p+ (t + u, u)m+ (u) du ≈ p+ (t, 0)
2 0
26
This footnote also applies for graphs later in the thesis. Note that Figures 6.4, 6.5, 7.9, 7.10 and 7.11
have use the following parameters in the simulations.
tstart = 0 tstep = 0.014
tend = 100000
although a time step size of tstep = 0.014 was used, only the data for every ten time steps were shown.
This was done to avoid handling a very large graph in MatLab.
116
because we do not have expressions for m− (u) and m+ (u). Notice one subtlety about all
the theories developed in Chapter 3. It was assumed that the probabilities for transit, that
is p(t) and q(t), were known for whatever the driving frequency Ω may be. In the PDFs
it was also assumed that the escape rates R−1+1 (t) and R+1−1 (t) were known for however
fast or slow the driving frequency Ω may be.
But such ideal expressions for p(t), q(t), R−1+1 (t) and R+1−1 (t) are not known. Note
that the escape rates as given by Kramers’ formula in Chapter 2.2 are only valid for
small noise for a static potential. When we do analysis in Chapter 7 the rates R−1+1 (t)
and R+1−1 (t) are calculated by Kramers’ formula as though it is a static potential. This
means an oscillatory potential is being approximated by a static potential. This is the
adiabatic approximation. The following conditions are proposed to decide if the adiabatic
approximation is valid
min
kram
2π
kram
τ−1+1 (t), τ+1−1
(t) ≤
t∈[0,T ]
Ω
(6.11)
kram
2π
kram
(t) ≤
(t), τ+1−1
max τ−1+1
t∈[0,T ]
Ω
(6.12)
that is we consider the minimum and maximum escape times over one period as given
by Kramers’ formula for a static potential. If this is less than the period of the driving
frequency T = 2π/Ω, then the adiabatic approximation may be valid. This was checked
for all the parameters and Equations 6.11 and 6.12 only hold for the following range of
parameters
φ ≥ 75◦
and ≥ 0.28.
This is a compromise we made. Nevertheless Equations 6.11 and 6.12 do not define the
adiabatic approximation, but give an idea of what range the parameters need to be in.
6.4.4
Selection of Parameters - Stability of Deterministic Trajectory
There is also the stability of the deterministic trajectory (when = 0) to be concerned
about. The following starting positions were chosen for four particles
xstart = +2
xstart = +2
xstart = −2
xstart = −2
ystart
ystart
ystart
ystart
= +2
= −2
= +2
= −2
and their trajectories falling through an unperturbed static potential, a static potential
with maximal forcing and an oscillating potential were all shown to be stable for all angles.
Note that these values of xstart and ystart were chosen because they are at a place where
117
the potential is so high the particle will probably never go there. Notice how in Figures 6.4
and 6.5 the trajectory almost never reaches any of the four corners (2, 2), (2, −2), (−2, 2)
and (−2, −2). This is the reason for checking the trajectories there.
6.4.5
Selection of Parameters - Random Number Generator
A few words may also be said about the random number generator we are using. Note
that these are pseudo random numbers. They are deterministic sequences of numbers with
a very long period. We use the randn() function in MatLab. It is very random and will
almost certainly not repeat itself for many years. This is because the period is 21492 . Even
with the computer generating 60 million random numbers per second it would still take
10434 years to reach the end of the cycle [62]. The function rng(’shuffle’) was also used,
which picks a seed for the random number generator according to the time of the computer
clock. When the Parallel ToolBox is used in MatLab, each worker randomly picks a seed
for itself.
6.4.6
Selection of Parameters - Calculating Positions of Critical
Points
As mentioned in Chapter 5.5.2, the numerical algorithms can be very unstable for calculating the positions of the critical points. In the simulations which we are going to conduct,
a table of the positions of all the critical points within one period are calculated first, then
stored in the temporary memory of the computer, and looked up every time the position
of a critical point is needed. This table is calculated in the following way. Define what we
call the pseudo parameters to be
ustart = 0 ustep = 0.001 uend = 2π
Ω=1
and then numerically find the positions of the critical points of the equation
Vt = V0 − Fx cos Ωt − Fy cos Ωt
where
t = 0,
t = ustep ,
2ustep ,
3ustep ,
...
t ≈ 2π.
Due to the way a matrix is define in MatLab, the last value of t is not exactly at t = 2π.
This table was checked for all angles, and the pseudo parameters we have chosen are stable.
6.4.7
Selection of Parameters - Higher Precision Numerics
We also have some remarks about the time step we have chosen. As we shall see in Chapter
7, one of the main effects which we have observed in this thesis is what we call the Single,
118
Intermediate and Double Frequency in the histograms of escape times. These effects were
first observed for the values of a = 0.15, b = 0.1, F = 0.7F crit and Ω = 0.001 when
tstep ≥ 0.0286 R ≥ 0.3218
and was observed again when tstep = 0.014 and R = 0.19. Note that tstep = 0.014 and
R = 0.19 was used for the results of this thesis. Thus we have confidence in believing that
the data we have collected is reliable. Consider the graphs below. They are histograms
of escape times from both the left and right wells combined. They are also normalised to
give an empirical PDF.
Figure 6.6: Here 2239 transitions were used. The averaged measured escape time is
0.0977T . The radius used was R = 0.5386.
119
Figure 6.7: Here 56244 transitions were used. The averaged measured escape time is
0.1064T . The radius used was R = 0.19.
The higher the noise level is the more susceptible to errors would the Euler method be.
This is why the highest level of noise = 0.30 is chosen for these examples. One may
argue that having more transitions would give a better measurement of the escape times
in Figure 6.7. But the difference in real time between the measured averaged escape times
is only 54 seconds out of a period of T = 2π/Ω = 6283 seconds.
120
Chapter 7
Simulations, Results and Analysis
This Chapter presents the main results of this thesis. The six measures M1 , M2 , M3 , M4 ,
M5 , M6 , the distributions of escape times and the newly developed conditional KolmogorovSmirnov test are used to analyse simulations of the SDE with the Mexican Hat Toy Model
used as the potential. The six measures are shown to be insensitive to the saddles changing
from alternating to synchronised. This is shown to be due to the fact that the invariant
measure is constant for synchronised saddles. The distribution of escape times shows new
signatures as the saddles change from alternating to synchronised and the conditional
Kolmogorov-Smirnov test is demonstrated to be an appropriate way to analyse the escape
times collected from many transitions.
We simulate a series of stochastic trajectories for the Mexican Hate Toy Model and
analyse them. We remind ourselves of the unperturbed Mexican Hat potential
1
1
V0 (x, y) = r4 − r2 − ax2 + by 2
4
2
where r =
p
x2 + y 2
and the SDE we want to simulate is
∂V0
+ Fx cos Ωt dt + dwx
dx = −
∂x
∂V0
dy = −
+ Fy cos Ωt dt + dwy
∂y
where Fx and Fy are the x and y components of the forcing, Ω is the forcing frequency,
is the noise level and wx and wy are two independent Wiener processes. We can define the
magnitude and angle of the forcing by
q
Fy
−1
2
2
.
F = Fx + Fy and φ = tan
Fx
This means the SDE can be written alternatively as
∂V0
dx = −
+ F cos φ cos Ωt dt + dwx
∂x
121
∂V0
dy = −
+ F sin φ cos Ωt
∂y
dt + dwy .
The critical forcing is defined by (see Equation 5.9)
F crit = min Fxsad , Fxcrit , Fysad , Fycrit .
7.1
Details of the Simulations
The Euler method was used to simulate this SDE with the following parameters being
fixed at the following values
a = 0.15 b = 0.1 F = 0.7F crit
Ω = 0.001.
The angle of the forcing φ and the noise level were varied. The values used were
= 0.15, 0.16, . . . , 0.30 and φ = 0◦ , 75◦ , 78◦ , 81◦ , 84◦ , 87◦ , 90◦ .
The averaged diffusion trajectories hxt i and hyt i were collected. The averaged Markov
Chain hYt i and the averaged Out-of-Phase Markov Chain hY t i were collected as well.
This would allow for the calculation of the invariant measures ν − (·) and ν + (·). The time
coordinates of the entrance and exit to and from the left and right wells were also collected.
This would allow for the calculation of the escape times. We use the following values for
the time step and the radius around the wells
tstep = 0.014 and R = 0.19.
Note that the period of the forcing is denoted by
T =
2π
.
Ω
The averaged trajectories were simulated by taking the averaged of 200 realisations. Each
realisation was 30 periods long, that is a trajectory over the interval [0, 30T ]. The initial
value of the state probabilities were set at
ν− (0) = ν+ (0) =
1
2
which assists in giving a faster convergence to the invariant measures (see Theorems 3.1,
3.2, and 3.8). We should also stress that a lot of the data and results presented in this
Chapter is just a selection out of a much wider range of results. All 112 combinations of
the parameters were simulated and analysed.
122
7.2
Six Measures Analysis
The six measures are calculated for the diffusion and Markov Chain case for all angles of
the forcing φ and all noise levels used in the simulations. When φ = 90◦ the wells were
moving up and down but they were always at the same height as each other. The distance
from either wells to the saddles, which is a gateway for escape, is the same in both wells at
all times. This means the φ = 90◦ can be modelled by a synchronised Markov Chain with
p = q. The invariant measures for the φ = 90◦ case as predicted by Corollary 3.4 and 3.10
is ν − = ν + = 12 . The Fourier Transform of the averaged Markov Chain is predicted to be
zero by Corollary 3.6 and 3.12. This predicts the six measures at φ = 90◦ to be
M1 = 0
M2 = 0
Z T
Z T
2
(ν+ (t) − ν− (t))2 dt = 0
hYt i dt =
M3 =
0
0
Z T
M4 =
Y t dt
0
Z T /2
Z T
=
0 × ν− (t) + 1 × ν+ (t) dt +
1 × ν− (t) + 0 × ν+ (t) dt
0
T /2
1
= T
2
−
+
Z T
φ (t)
φ (t)
−
+
φ (t) ln
M5 =
+ φ (t) ln
dt
ν − (t)
ν + (t)
0
Z T
Z T /2
1
1
ln
ln
dt +
dt
=
ν − (t)
ν + (t)
T /2
0
= +T ln(2)
Z T
−ν − (t) ln ν − (t) − ν + (t) ln ν + (t) dt
M6 =
0
= +T ln(2).
Note that ln(2) = 0.6931 ≈ 0.7. Notice that for very low noise level ≈ 0 the probabilities
of escape from either well is so small it may be approximately modelled by a synchronised
Markov Chain with p ≈ q. The results below confirm the predictions for the case of
φ = 90◦ .
123
Figure 7.1: The measure M1 for the diffusion case for various angles and noise levels.
Figure 7.2: The measure M2 for the diffusion case for various angles and noise levels.
124
Figure 7.3: The measure M1 for the Markov Chain for various angles and noise levels.
Figure 7.4: The measure M2 for the Markov Chain for various angles and noise levels.
125
Figure 7.5: The measure M3 for the Markov Chain for various angles and noise levels.
Figure 7.6: The measure M4 for the Markov Chain for various angles and noise levels.
126
Figure 7.7: The measure M5 for the Markov Chain for various angles and noise levels.
Figure 7.8: The measure M6 for the Markov Chain for various angles and noise levels.
127
7.2.1
Interpretation of the Six Measures Analysis
The six measures M1 , M2 , M3 , M4 , M5 and M6 were plotted as a function of the noise
level . The six measures show a regular systematic behaviour in the angle φ. The shape of
the graphs of the six measures were very similar for all the angles. As the angle increased
from φ = 0◦ to φ = 90◦ the six measures gradually tended to being nearly constant in .
This effect can be explained with the invariant measures. When φ = 0◦ the probabilities
for escaping from left to right p−1+1 was different to the probabilities for escaping from
right to left p+1−1 . But in the φ = 90◦ case they are the same, that is
φ = 0◦
φ = 90◦
p−1+1 =
6 p+1−1
p−1+1 = p+1−1 .
The is can be understood geometrically. For φ = 0◦ we have Fx 6= 0 and Fy = 0. The two
wells in the Mexican Hat potential move up and down and are alternating with each other.
When one well is high the other is low. For φ = 90◦ we have Fx = 0 and Fy 6= 0. The two
wells are always at the same height as each other and the distance to the saddles (which
is a gateway to escape) is also the same in both wells.
Recall our discussions on the Markov Chain in Chapter 3. The p is related to left to
right escape p−1+1 and q was related to right to left escape p+1−1 . For φ = 0◦ the Markov
Chain can be modelled with p 6= q and for φ = 90◦ the Markov Chain can be modelled
with p = q. In the case of φ = 0◦ the invariant measure was cyclically changing in time. In
the case of φ = 90◦ the invariant measure was constant at ν − (·) = ν + (·) = 12 . This explains
why the six measures M1 , M2 , M3 , M4 , M5 and M6 were nearly constant for angle φ = 90◦ .
As φ changed from φ = 0◦ to φ = 90◦ , the Markov Chain changed from being modelled by
p 6= q to being modelled by p = q. This explains the change in the six measures tending
to being constant in as φ was varied. The six measures can be thought of as a way of
measuring how far away the invariant measures are from being constant. If the invariant
measures are constant then the six measures will also be constant.27
For fixed φ near φ = 90◦ there is no pronounced maximum of any measure for varying
. Hence the six measures indicate the absence of a pronounced stochastic resonance near
φ = 90◦ . But consider the trajectories at a range of angles.
27
See Appendix B.3 for how M5 and M6 were numerically calculated. The ideas were not that trivial.
128
Figure 7.9: The blue trajectory is x(t) and the green trajectory is y(t).
Figure 7.10: The blue trajectory is x(t) and the green trajectory is y(t).
129
Figure 7.11: The blue trajectory is x(t) and the green trajectory is y(t).
When φ = 0◦ the x(t) shows quasi-deterministic behaviour. The transitions are very regular
and y(t) fluctuates around zero. As the angle varies the transitions become less regular
and y(t) starts to oscillate. This suggests that there is some regularity in the behaviour of
the trajectories but the six measures are not detecting it. Further studies with the escape
times would tell us more.
7.3
Escape Time and Conditional KS Test Analysis
We remind ourselves of the PDF of escape times and the way the conditional KS test can
be applied in our context. The conditional PDF of the escape times are
Z t
p− (t, u) = R−1+1 (t) exp −
R−1+1 (s) ds
u
Z t
p+ (t, u) = R+1−1 (t) exp −
R+1−1 (s) ds
u
where R−1+1 and R+1−1 are the Kramers’ escape rate from left to right and right to left. In
the case of p− (t, u), t is the time coordinate of escape from the left well and u is the time
coordinate of entrance into the left well. In the case of p+ (t, u), t is the time coordinate of
escape from the right well and u is the time coordinate of entrance into the right well. If
we do not differentiate between escaping from the left or right then the PDF for an escape
130
time t is (note that t here is an escape time as it is and not a time coordinate)
1
ptot (t) =
2
Z
T
p− (t + u, u)m− (u) + p+ (t + u, u)m+ (u) du
0
where m− (·) and m+ (·) are PDFs of the time of entrance into the left and right well
respectively. We do not have explicit expressions for m− (·) and m+ (·). The ptot (t) is
approximated by
ptot (t) ≈ p+ (t, 0).
The times it took to escape from both the left or right wells are plotted in histograms.
This is an empirical approximation to the PDF ptot (t) ≈ p+ (t, 0). A selection of some of
the results are given below for various angles of the forcing φ and noise level . They are
examples of the Singles, Intermediate and Double Frequencies which we will explain later.
Note that the escape times are given in units of normalised time, which is in the number
of periods T .
Figure 7.12: This is an example of the Single Frequency.
131
Figure 7.13: This is an example of the Intermediate Frequency.
Figure 7.14: This is an example of the Intermediate Frequency tending closer to the
Double Frequency.
132
Figure 7.15: This is an example of the Double frequency.
It is important to note that Figures 7.12, 7.13, 7.14 and 7.15 are histograms of the actual
times it took to escape from either wells without differentiation between wells on the left
or right. The times of entrance into the wells are not shown. The PDF used is ptot (·) which
is being approximated by ptot (t) ≈ p+ (t, 0).
These escape times can be analysed in a different way. Let u be the time of entrance
into a well and t the time of exit from a well. Figures 7.12, 7.13, 7.14 and 7.15 are therefore
histograms of the (t−u) for both left and right escapes combined. Thus 0 ≤ mod(u, T ) ≤ 1
is the phase of entrance into a well and mod(t − u, T ) is the escape time itself in normalised
time. Such an analysis is done for the times in Figures 7.12, 7.13, and 7.15 for both the
left and right wells respectively.
133
Figure 7.16: The u is the time of entrance into the well and t is the time of exit from the
well.
Figure 7.17: The u is the time of entrance into the well and t is the time of exit from the
well.
134
Figure 7.18: The u is the time of entrance into the well and t is the time of exit from the
well.
Figure 7.19: The u is the time of entrance into the well and t is the time of exit from the
well.
135
Figure 7.20: The u is the time of entrance into the well and t is the time of exit from the
well.
Figure 7.21: The u is the time of entrance into the well and t is the time of exit from the
well.
Notice the general behaviour of the data for mod(u, T ) and mod(t − u, T ). For the φ = 0◦
case the wells are alternating and one well is higher than the other. Entrance into the left
136
well tend to occur near u = 0.5 and entrance into the right well tend to occur near u = 0
and u = 1. For φ = 90◦ the wells are synchronised and are always at the same height as
each other. Entrance and exit to and from either well tend to occur at u = 0, u = 0.5 and
u = 1. Notice that the Single, Intermediate and Double Frequencies can be seen in Figures
7.16, 7.17, 7.18, 7.19, 7.20 and 7.21.
Notice also in Figure 7.16 the data points are tiled near 0.5. This seems to suggest
that the use of the Dirac delta function to approximate ptot ≈ p+ (t, 0) (see Chapter 3.4.2)
may not be very good. The main problem here is the fact that we do not have an explicit
formula for a probability measure of the time of entrance into a well, that is we do not
have expressions for m− (u) and m+ (u). This motivates us into developing the conditional
KS test.
We want to test whether the escape times we have measured are really distributed by
the conditional PDFs p− (t, u) and p+ (t, u). This is testing the conditional null hypothesis.
Define the conditional CDFs by
Z t
Z t
−
R−1+1 (s) ds
p− (s, u) ds = 1 − exp −
Fu (t) =
u
u
Fu+ (t) =
Z
Z t
t
p+ (s, u) ds = 1 − exp −
R+1−1 (s) ds .
u
u
The time coordinates of the entrance and exit from the wells are collected. These are
u1 u2 . . . un
t1 t2 . . . tn
where ui is the time coordinate of the ith entrance into a well and ti is the time coordinate
of the ith exit from a well. The conditional KS statistic is calculated by
n
Sn−
= sup
x∈[0,1]
1X
1[0,x] Fu−i (ti ) − x
n i=1
n
Sn+
= sup
x∈[0,1]
1X
1[0,x] Fu+i (ti ) − x
n i=1
where in Sn− we sum over the time coordinates of entrance and exit to and from the left well
and in Sn+ we sum over the time coordinates of entrance and exit to and from the right well.
Recall that if the conditional null hypothesis is true then Sn− and Sn+ are asymptotically
distributed by
∞
X
√
2 2
lim P ( nSn ≤ x) = Q(x) where Q(x) = 1 − 2
(−1)k−1 e−2k x .
n−→∞
k=1
We want 99% confidence. Note that
√
P
nSn ≤ 1.6920 = Q(1.6920) = 0.99.
137
√
√
The Q ( nSn ) is also calculated. The smaller Q ( nSn ) is the more certain we are in
accepting the null hypothesis. A selection of some of the data being implemented with the
conditional KS test are given below for various angles of the forcing φ and noise level .
These are examples of the KS test being implemented for the histograms of escape times
just given in Figures 7.12, 7.13, 7.14 and 7.15
Figure 7.22: This is an example of the conditional KS test
implemented for√the data
√ being
−
◦
in Figure 7.12. Note that = 0.18, φ = 0 , n = 200, nSn = 0.5233 and Q ( nSn− ) =
0.0529.
138
Figure 7.23: This is an example of the conditional KS test
implemented for√the data
√ being
−
◦
in Figure 7.13. Note that = 0.20, φ = 84 , n = 200, nSn = 0.6223 and Q ( nSn− ) =
0.1665.
Figure 7.24: This is an example of the conditional KS test
implemented for√the data
√ being
◦
−
in Figure 7.14. Note that = 0.21, φ = 87 , n = 200, nSn = 1.2587 and Q ( nSn+ ) =
0.9159.
139
Figure 7.25: This is an example of the conditional KS test
implemented for√the data
√ being
−
◦
in Figure 7.15. Note that = 0.21, φ = 90 , n = 200, nSn = 1.0465 and Q ( nSn− ) =
0.7766.
7.3.1
Interpretation of the Escape Time and Conditional KS Test
Analysis
When φ = 0◦ there were peaks in the empirical PDF of the escape times. These occurred
at times 12 T , 32 T , 52 T , . . . . This effect we call the Single frequency. When φ = 90◦ the
peaks occurred at 21 T , 32 T , 52 T , . . . and 0, T , 2T , 3T , 4T , . . . . This effect we call the Double
Frequency. When 0◦ < φ < 90◦ an intermediate effect is seen. There were major peaks at
1
T , 32 T , 25 T , . . . and minor peaks at 0, T , 2T , 3T , 4T .
2
The behaviour of the Single, Intermediate and Double Frequencies can be explained
geometrically. When the height between a well and a saddle is minimum, the optimal
probability of escape has occurred. When φ = 0◦ the frequency of the return of the optimal
probability of escape is the same as the driving frequency Ω. This optimal probability
comes back very T which is once in a period. When φ = 90◦ the frequency of the return
of the optimal probability of escape is double the driving frequency at 2Ω. This optimal
probability comes back very T2 which is twice in a period. This explains why the peaks in
the Single and Double Frequencies are seen where they are.
As the angle changed from φ = 0◦ to φ = 90◦ the Single Frequency gradually changes
into the Double Frequency with the Intermediate Frequency seen in between. Thus the
angle of the forcing is leaving a mark in the PDFs of escape times.
140
When the conditional KS test was implemented, the functions
y0 (x) = x,
y− (x) =
n
X
1[0,x]
Fu−i (ti )
and y+ (x) =
i=1
n
X
1[0,x] Fu+i (ti )
i=1
were used to calculate the following distances which are the conditional KS statistics
Sn− = ky0 − y− k∞
and Sn+ = ky0 − y+ k∞ .
It is reasonable to say that y− (·) and y+ (·) were close enough to y0 (·) that we can accept
the conditional null hypothesis. This can be seen and judged graphically with Sn− and
Sn+ calculated as well. This is an example of the conditional KS test giving a reasonable
result.28
7.4
Sparse Data Analysis
We do the same analysis with the escape time and the conditional KS test. But now
we artificially make the data sparse by only implementing the conditional KS test for 20
transitions. We want 99% confidence. Thus with n = 20 tables for the KS distribution
show that
P (S20 ≤ 0.356) = 0.99
and there are two particular examples we want to focus on. These are when ptot ≈ p+ (t, 0)
is not a good approximation and the conditional KS test is performed in such a situation.
28
See Appendix C.1 for discussions as to how some of our implementation of the conditional KS test
are examples of oversampling.
141
Figure 7.26: The ptot ≈ p+ (t, 0) is not a good approximation here.
Figure 7.27: This is a KS test on the data in Figure 7.26. The conditional null hypothesis
can√be reasonably accepted. Note that = 0.17, φ = 81◦ , n = 20 and Sn+ = 0.2750.
Q( nSn+ ) = 0.9029.
142
Figure 7.28: The ptot ≈ p+ (t, 0) is not a good approximation here.
Figure 7.29: This is a conditional KS test on the data in Figure 7.28. The conditional
null hypothesis can be reasonably
accepted. Note that = 0.27, φ = 78◦ , n = 20 and
√
Sn− = 0.1400. Also note that Q( nSn− ) = 0.1720.
143
7.4.1
Interpretation of the Sparse Data Analysis
The aim of Sparse Data Analysis is to see how the conditional KS test performs even if
less data is available. This is done by looking at two cases where ptot ≈ p+ (t, 0) is not a
good approximation and implementing the conditional KS test on them after artificially
making the data sparse.
Consider the case for the parameters in Figure 7.26. Figure 7.26 is an example of when
the noise is so small there is very little escape times being detected in the range [0, 5T ].
The ptot ≈ p+ (t, 0) is not a good approximation here. In Figure 7.27 the conditional KS
test was performed on the data in 7.26 and the distance between the two functions is small.
This means we can accept the conditional null hypothesis even when there is fewer data
and ptot ≈ p+ (t, 0) is not a good approximation.
Now consider the case for the parameters in Figure 7.28. The noise is so large ptot ≈
p+ (t, 0) is no longer a good approximation. But in Figure 7.29 the conditional KS test was
performed on the data in Figure 7.28. Again this is an example of us being able to accept
the conditional null hypothesis even if ptot ≈ p+ (t, 0) is not a good approximation.
Only 20 escape times were implemented in the conditional KS test and the conditional
null hypothesis can still be accepted with a reasonable degree of certainty. But the ptot ≈
p+ (t, 0) was not a good approximation for the empirical PDF of escape times. These are
examples of the conditional KS test giving reasonable conclusions even when there are
sparse data. It also shows that the conditional KS test can still be used even if there is no
good approximation of the PDF of escape times.
Back in Chapter 3.4.2 we approximated m− (u) and m+ (u) by
m− (u) ≈ δ (u − T /2)
1
1
m+ (u) ≈ δ (u) + δ (u − T ) .
2
2
Although the escape times (represented as dots on a scatter graph) tend to cluster around
u = 0, u = 0.5 and u = 1 there are spread around them. As the noise levels increases the
spread around u = 0, u = 0.5 and u = 1 would increase and ptot ≈ p+ (t, 0) would stop to
be a good approximation. Despite this the conditional KS test still shows sensible results,
in that we can accept the conditional null hypothesis.
7.5
Remarks on Analysis of Stochastic Resonance
There are a few subtleties and setbacks to the analysis which is worth mentioning here.
7.5.1
Remarks on Implementing the Conditional KS Test
Notice that all the theories developed about the KS Test were based on the assumption
that the null hypothesis is true. This means strictly speaking a small KS statistic, that is
a small Sn− or Sn+ does not immediately allow us to accept the null hypothesis but good
144
reasons not
√ is− for large√n, the
√ to− reject it.√Also+ when there were many transitions, that
terms Q( nSn ) and Q( nSn ) were also calculated. The smaller Q( nSn ) and Q( nSn+ )
are the more confidence we have in not rejecting the null hypothesis. This is because for
very large n, we would expect
√
√
lim nSn− = 0 and lim nSn+ = 0
n→∞
n→∞
√
√
so the smaller Q( nSn− ) and Q( nSn+ ) are the more certain we are in not rejecting the
null hypothesis.
7.5.2
Remarks on Adiabatic Approximation
Notice that in the PDFs p− (t, u), p+ (t, u) and ptot (t) expressions for the escape rates
R−1+1 (t) and R+1−1 (t) were required. These rates were also required for the conditional
KS test. Strictly speaking these rates are dependent on the driving frequency Ω, but we
stress that these rates were calculated using Kramers’ formula as though the particle is
escaping from a static potential. This is the adiabatic approximation where an oscillatory
potential is approximated by a static potential. Considerations for whether the adiabatic
approximation would fail in our calculations were done back in Chapter 6.4.3
It is worth summarising all the approximations which the analysis of the data have
been based. There is the small noise approximation and slow forcing approximation from
Kramers’ formula, the adiabatic approximation and the perfect phase approximation where
ptot is approximated by ptot ≈ ptot (t, 0).
145
Conclusion
Outline of Results
In this thesis we have considered the following problem. Let Xt be a stochastic process in
R2 which is described by the SDE
dXt = b (Xt , t) dt + dWt
and the drift term b(·, ·) is expressed by
b(x, t) = −∇V0 (x) + F cos Ωt
where V0 : R2 −→ R is a time independent function, the unperturbed potential, with two
metastable states, and two pathways between these states. The F ∈ Rr is the magnitude
of the forcing and Ω is the driving frequency. Our aim was to see characteristics of the
trajectory Xt which only depends on the qualitative structure of V0 , that is the existence
of two metastable states and two pathways.
For concreteness we considered a model, which we call the Mexican Hat Toy Model
1
1
V0 (x, y) = r4 − r2 − ax2 + by 2
4
2
where r =
p
x2 + y 2 .
The magnitude and angle of the forcing are given by
q
F = Fx2 + Fy2
−1
and φ = tan
Fy
Fx
.
The angle φ and noise level were varied. At φ = 0 the wells were alternating, that is one
well is higher than the other, in the sense that it is easer to jump from one well to the
other than vice versa. At φ = 90◦ the wells are synchronised, that is both wells are always
at the same height but the heights of the barrier for the two paths is alternating.
A potential with two pathways has never been considered before in the context of
stochastic resonance. We studied it using approximation techniques and direct simulations.
In an adiabatic regime the Freidlin-Wentzell theory allows one to give analytical solutions
of the jump type distributions asymptotically in this regime. This theory predicted the
appearance of additional resonance peaks at half the frequency when the angle approaches
φ = 90◦ .
146
We simulated Xt for different values of φ and and computed for the values of angle
increasing from φ = 0 to φ = 90◦ the six measures M1 , M2 , M3 , M4 , M5 and M6 as
functions of the noise level. The first major surprise was that the graphs showed less and
less pronounced minima (or maxima) and hence suggests that the phenomena of stochastic
resonance gets less and less pronounced, see Chapter 7.2. The effect of resonance seems to
disappear overall.
However, considering the path Xt itself, one sees that there may be nevertheless some
synchronisation, see Figure 7.9, 7.10 and 7.11. We carefully controlled our simulation
and checked it for consistency, see Chapter 6.4. To properly quantify synchronisation
we considered the histograms of the escape times, which to our knowledge has been not
considered thoroughly before. The histograms showed a clear periodicity and also the
emergence of peaks at the Double Frequency for increasing angle. For a quantitative
consideration we assume that the entrance time is in perfect phase (this is when m− (u)
and m+ (u) can be approximated by Dirac delta functions). This gives for several cases good
quantitative and in general good qualitative agreement with the combined adiabatic and
small noise approximation. A more sophisticated analysis based on a Kolmogorov-Smirnov
test developed here shows that this approximation works for a larger range of parameters
where the approximation of the perfect phase of the entrance time is not appropriate (this
is when m− (u) and m+ (u) cannot be approximated by Dirac delta functions) see Chapter
7.4.1. Summarizing, the theoretical and the simulation results are in very good agreement.
We want to stress that in the comparison no free parameters were present and so no fitting
took place.
The fact that the six measures are blind can be explained using Markov chain models
approximating the SDE. As one expects from large deviation theory, for small noise and
in an adiabatic regime the SDE can be approximated by a continuous time Markov chain.
In this Markov chain model we showed that the invariant measures are constant when
φ = 90◦ . Hence we expect that the invariant measure gives in the diffusion case equal
weights to the left and the right well. Together, this gives us the following qualitative
picture of the dynamics for any angle. At a fixed time the probability that one sees a jump
from the left to the right well or vice versa has the same probability. However, conditioned
on the phase and the direction of the last jump, for concreteness assume that it was at
phase u and from the left to the right (that is to say the particle entered the well at time
u) the next jump will be at phase which is near to a multiple of T /2 (that is to say the
particle will leave the well near the times t = nT /2 where n is an integer). The jump rates
will be given by the height of the potential barriers.
At φ = 90◦ , the path Xtε and −Xtε will appear with the same probability if one starts in
the invariant measure. This explains why the six measures are all insensitive in this case.
The equilibration happens because the process will skip some of the jump opportunities
and in this way the left-right synchronization will get lost quickly.
This new phenomena we discovered has added an additional motivation to the observation of Hermann, Imkeller, Pavlyukevich, Berglund and Gentz that the appropriate
consideration has to be on the path level. Averaged quantities like the six measures can
be very misleading and masking the real behaviour of the system. The escape time dis147
tribution shows a clear signal of stochastic resonance in accordance with the theoretical
consideration. The presence of a two pathways manifests itself in an appearance of peaks
at the Double Frequency. We showed that adiabatic small noise approximation gives a
good statistical model. We demonstrated that this appearance can be detected also when
only a limited number of transitions is available. Our analysis provides us with a clear
footprint indicating the existence of a second pathway. The angle dependence of our result
should also allow us to predict the orientation of the saddles with respect to the wells.
Further Studies
The invariant measures studied in this thesis are for a simple two state model. One could
try to generalise this to continuous states, that is a space-time phase PDF for the position
of the particle could be derived.
The conditional KS test gives us confidence that one could develop statistical inference,
using maximum likelihood for example, to develop a statistic test to estimate the basic
parameters of the system, if they are unknown to us. Instead of the approximation ptot ≈
p+ (t, 0) used in parts of the consideration, a better approximation may be found by studying
the PDFs of m− (u) and m+ (u) theoretically and statistically.
A theory beyond adiabatic approximation may be developed for very slow to fast frequencies. Higher order of approximation to the escape times than Kramers’ formula could
be studied. Analytic and theoretical developments to go beyond adiabatic approximation
and potential theory may be a real mathematical challenge. But experimental simulations
may provide an idea of what this new theory may be like.
148
Appendix A
Conventions in Defining the SDEs,
Potential, Time Dependency and
Forcing
This is more of a clarification on the notation being used. In this thesis only two toy model
potentials are studied. These in their most unperturbed forms are denoted by
x2
x4
−a
4
2
1 4 1 2
V0 (x, y) = r − r − ax2 + by 2
4
2
V0 (x) =
where r =
p
x2 + y 2 , a > 0 and b > 0 which when perturbed by a force are denoted by
x4
x2
− a + Fx
4
2
= V0 (x) + F x
1
1
VF (x, y) = r4 − r2 − ax2 + by 2 + Fx x + Fy y
4
2
= V0 (x, y) + Fx x + Fy y
= V0 (x, y) + F · x
VF (x) =
and when given a periodic forcing are denoted by
x4
x2
− a − F x cos Ωt
4
2
= V0 (x) − F x cos Ωt
1
1
Vt (x, y, Fx , Fy ) = r4 − r2 − ax2 + by 2 − Fx x cos Ωt − Fy y cos Ωt
4
2
= V0 (x, y) − Fx x cos Ωt − Fy y cos Ωt
= V0 (x, y) − F · x cos Ωt.
Vt (x, F ) =
149
This is so that the SDEs can be written in the form
dXt = −∇Vt dt + dw
which when expanded can be written as
∂V0
dx = −
+ Fx cos Ωt dt + dwx
∂x
∂V0
+ Fy cos Ωt dt + dwy
dy = −
∂y
meaning more details about the system can be quickly seen in the notation. This also
implies that the SDEs are always defined with a negative forcing. When the critical points
of the system are being studied (in Chapter 5 for example) we can study the critical points
with a positive force and VF would be an appropriate notation to use. Using V0 , VF and
Vt may seem like an abuse of notation, but if anything specific is being referred to, we can
denote VF =F crit for example. When the most general expression for a potential V is being
used, it should be deduced from context whether V = V0 , V = VF or V = Vt is being
referred to.
Note also that for a stochastic process in Rr which is described by the SDE
Ẋt = −∇V + F cos(Ωt) + Ẇt
and the magnitude of the forcing is sometimes denoted by
q
F = F12 + F22 + . . . + Fr2 .
Again this may seem like an abuse of notation, but it should be clear from context whether
F is a vector or scalar.
150
Appendix B
Further Numerical Methods
B.1
Numerical Methods for measuring Escape Times
The Markov Chain takes the values Yt = ±1. But in simulations time is discrete with a
time step tstep , that is
0, tstep , 2tstep , . . . , N tstep .
The reduction from the diffusion Xt to the Markov Chain at the (n + 1)th time step is
actually given by
− wl (ntstep ) < R
−1
if
Xnt
step
Y(n+1)t
=
− wr (ntstep ) < R
+1
if
Xnt
step
step
Y
if otherwise
ntstep
which is slightly different from the way Yt was defined in Chapter 3 (see page 37). This is
so that the definition of Yt was easier to write down theoretically, such that the sets
{t : |Xt − wl (t)| ≤ R}
and
{t : |Xt − wr (t)| ≤ R}
are compact sets given the continuity of Xt , wl (t) and wr (t). This meant
−1
if
|Xt − wl (t)| ≤ R
+1
if
|Xt − wr (t)| ≤ R
Yt =
Z if neither
then Yt would be easier to define for t ∈
/ {t : |Xt − wl (t)| ≤ R} ∪ {t : |Xt − wr (t)| ≤ R}.
But alternatively if we had
−1
if
|Xt − wl (t)| < R
+1
if
|Xt − wr (t)| < R
Yt =
Z if neither
151
then
{t : |Xt − wl (t)| < R}
and
{t : |Xt − wr (t)| < R}
would be open sets and Yt would be harder to define for t ∈
/ {t : |Xt − wl (t)| < R} ∪
{t : |Xt − wr (t)| < R} which is the neither case. Nevertheless the simulations should gloss
out all these details.
B.2
Numerical Methods for calculating Fourier Transform and Linear Response
Fourier Transforms are involved in finding the linear response. The trajectory of the particle
is in theory a continuous object, but in practice when simulations are done it is a finite
discrete object. The exact mechanism of obtaining the linear response from a simulated
trajectory is now being discussed.
When the trajectory is being numerically realised it is a finite discrete set. Let the x
(or y) coordinate of the particle at time ntstep where 0 ≤ n ≤ (N − 1)tstep be denoted by
Xntstep . This gives rise to the set
X = X0 , Xtstep , X2tstep , X3tstep , . . . , X(N −1)tstep
= {x0 , x1 , x2 , x3 , . . . , xN −1 }
where xn = Xntstep etc. Notice that time is discrete here. When this is Discrete Fourier
Transformed (being quickly implemented by the Fast Fourier Transform algorithm) it is
denoted by
o
n
X̃ = X̃0 , X̃ωstep , X̃2ωstep , X̃3ωstep , . . . , X̃(N −1)ωstep
= {x̃0 , x̃1 , x̃2 , x̃3 , . . . , x̃N −1 }
where x̃n = X̃nωstep etc and the transform is given by
x̃k =
N
−1
X
xn e−2πikn/N
n=0
and the following relation is used
ωstep =
1
(N − 1)tstep
which is the highest detectable frequency divided by the number of steps. If we want to
find the linear response at driving frequency Ω, then Ω needs to be approximated by a
finite number of ωstep as in
Ω
≈ nωstep
2π
152
and the linear response at this driving frequency is then given by
Ω
Xlin
= 2 × X̃nωstep .
Notice the factor of 2 being used here. Suppose that the trajectory can be approximated
by
Xt ≈ A cos(Ωt + φ)
then a good approximate expression for A and φ would be
Im X̃nωstep
Ω
A ≈ Xlin
and φ ≈ arg X̃nωstep = tan−1
Re X̃
nωstep
where φ is the angle of the complex number X̃nωstep .
B.3
Numerical Methods for calculating M5 and M6
Here we present how we computed M5 and M6 numerically. This is how M5 and M6 are
calculated in theory
−
+
Z T
φ (t)
φ (t)
−
+
M5 =
φ (t) ln
+ φ (t) ln
dt
ν − (t)
ν + (t)
0
Z T
−ν − (t) ln ν − (t) − ν + (t) ln ν + (t) dt
M6 =
0
where
−
1 if mod(t, T ) ≤ T /2
0 if mod(t, T ) > T /2
0 if mod(t, T ) ≤ T /2
1 if mod(t, T ) > T /2.
φ (t) =
+
φ (t) =
When the invariant measures are generated numerically they are finite discrete objects
described by
ν− = ν1− , ν2− , · · · , νN−
ν+ = ν1+ , ν2+ , · · · , νN+ .
The real invariant measure were close to zero sometimes and in the numerical approximation they became actually zero or even negative which lead to numerical artefacts. Note
that
1
lim ln
= ∞ and lim x ln (x) = 0.
x−→0
x−→0
x
153
Define
ν−lim =
ν+lim =
min
ν1− , ν2− , · · · , νN−
min
ν1+ , ν2+ , · · · , νN+ .
i=1,2,...,N
νi− >0
i=1,2,...,N
νi+ >0
The quantities M5 and M6 are computed numerically in the following way
X
X
X
X
1
1
1
1
M5 =
tstep ln
tstep ln
+
tstep ln
tstep ln
+
+
ν−lim
ν+lim
νi−
νi+
N
N
N
N
i≤ 2
νi− >0
M6 =
X
i=1,2,··· ,N
νi− >0
i> 2
νi+ >0
i≤ 2
νi− ≤0
νi− ln(νi− )(−tstep )
+
X
i=1,2,··· ,N
νi+ >0
154
νi+ ln(νi+ )(−tstep ).
i> 2
νi+ ≤0
Appendix C
Further Commentary on Sparse Data
Analysis
C.1
Examples of Oversampling
Subjectively one may think that Figures 7.24 and 7.25 are so bad the conditional null
hypothesis may be rejected. This is actually an example of oversampling, where too many
transitions were used in the implementation of the conditional KS test. We know the PDF
we are fitting is not the real PDF but an approximation in the limit of small noise and
adiabatic forcing. Hence if one has enough data points this should be picked up and the
conditional KS test will refuse the approximate PDF as it will pick up even slight deviation
from the real PDF. When n = 20 are used we have the following.
155
Figure C.1: This is Figure 7.24 redone with 20 transitions. Note that = 0.21, φ = 87◦ ,
n = 20, Sn+ = 0.1960.
Figure C.2: This is Figure 7.25 redone with 20 transitions. Note that = 0.21, φ = 90◦ ,
n = 20, Sn− = 0.1030.
156
C.2
Empirical CDF
Consider Figure 7.27. Notice that the empirical CDF is on top on the y = x line. There
were enough data to give 10 more realisations of the random variable Sn+ . Note that all
ten of these Sn+ with n = 20 were calculated from 200 transitions divided into ten sets for
the ten Sn+ . This meant 10 more versions of the Figure 7.27 were plotted. Out of these 10
plots, one had the empirical CDF to the bottom of the y = x line and one had roughly
half the empirical CDF above and below the y = x line. The noise level was very low at
= 0.17, which meant the escape times were very long with a very large spread, which gave
rise to data looking unreasonable. Only 200 transitions were detected which is significantly
less than other parameters, which meant only 10 realisations of the Sn+ random variable
was possible. No further conclusions are drawn here.
157
References
[1] A. Neiman, A. Silchenko, V. Anishchenko, and L. Schimansky-Geier, “Stochastic resonance: Noise-enhanced phase coherence,” Physical Review E, vol. 58, no. 6, p. 7118,
1998.
[2] B. Shulgin, A. Neiman, and V. Anishchenko, “Mean switching frequency locking
in stochastic bistable systems driven by a periodic force,” Physical Review Letters,
vol. 75, no. 23, p. 4157, 1995.
[3] N. Berglund and B. Gentz, “A sample-paths approach to noise-induced synchronization: Stochastic resonance in a double-well potential,” Annals of Applied Probability,
pp. 1419–1470, 2002.
[4] R. Benzi, A. Sutera, and A. Vulpiani, “The mechanism of stochastic resonance,”
Journal of Physics A: Mathematical and General, vol. 14, no. 11, p. L453, 1981.
[5] C. Nicolis and G. Nicolis, “Stochastic aspects of climatic transitionsadditive fluctuations,” Tellus, vol. 33, no. 3, pp. 225–234, 1981.
[6] R. Benzi, G. Parisi, A. Sutera, and A. Vulpiani, “A theory of stochastic resonance in
climatic change,” SIAM Journal on applied mathematics, vol. 43, no. 3, pp. 565–578,
1983.
[7] B. McNamara, K. Wiesenfeld, and R. Roy, “Observation of stochastic resonance in a
ring laser,” Phys. Rev. Lett., vol. 60, pp. 2626–2629, Jun 1988.
[8] L. Guidoni, R. Mannella, V. Isaia, P. Verkerk, and E. Arimondo, “Stochastic resonance
in a laser with saturable absorber,” Il Nuovo Cimento D, vol. 17, no. 7, pp. 803–810,
1995.
[9] J. Grohs, S. Apanasevich, P. Jung, H. Issler, D. Burak, and C. Klingshirn, “Noiseinduced switching and stochastic resonance in optically nonlinear cds crystals,” Phys.
Rev. A, vol. 49, pp. 2199–2202, Mar 1994.
[10] A. Simon and A. Libchaber, “Escape and synchronization of a brownian particle,”
Phys. Rev. Lett., vol. 68, pp. 3375–3378, Jun 1992.
158
[11] S. Fauve and F. Heslot, “Stochastic resonance in a bistable system,” Physics Letters
A, vol. 97, no. 1, pp. 5 – 7, 1983.
[12] R. N. Mantegna and B. Spagnolo, “Stochastic resonance in a tunnel diode,” Phys.
Rev. E, vol. 49, pp. R1792–R1795, Mar 1994.
[13] R. N. Mantegna and B. Spagnolo, “Stochastic resonance in a tunnel diode in the
presence of white or coloured noise,” Il Nuovo Cimento D, vol. 17, no. 7, pp. 873–881,
1995.
[14] R. N. Mantegna and B. Spagnolo, “Noise enhanced stability in an unstable system,”
Phys. Rev. Lett., vol. 76, pp. 563–566, Jan 1996.
[15] I. Lin and J.-M. Liu, “Experimental observation of stochastic resonance like behavior
of autonomous motion in weakly ionized rf magnetoplasmas,” Physical Review Letters,
vol. 74, no. 16, p. 3161, 1995.
[16] A. N. Grigorenko, P. I. Nikitin, A. N. Slavin, and P. Y. Zhou, “Experimental observation of magnetostochastic resonance,” Journal of Applied Physics, vol. 76, no. 10,
1994.
[17] G. Debnath, T. Zhou, and F. Moss, “Remarks on stochastic resonance,” Phys. Rev.
A, vol. 39, pp. 4323–4326, Apr 1989.
[18] L. Gammaitoni, F. Marchesoni, E. Menichella-Saetta, and S. Santucci, “Multiplicative
stochastic resonance,” Phys. Rev. E, vol. 49, pp. 4878–4881, Jun 1994.
[19] L. Gammaitoni, M. Martinelli, L. Pardi, and S. Santucci, “Observation of stochastic resonance in bistable electron-paramagnetic-resonance systems,” Phys. Rev. Lett.,
vol. 67, pp. 1799–1802, Sep 1991.
[20] A. Longtin, A. Bulsara, and F. Moss, “Time-interval sequences in bistable systems
and the noise-induced transmission of information by sensory neurons,” Phys. Rev.
Lett., vol. 67, pp. 656–659, Jul 1991.
[21] A. D. Hibbs, A. L. Singsaas, E. W. Jacobs, A. R. Bulsara, J. J. Bekkedahl, and
F. Moss, “Stochastic resonance in a superconducting loop with a josephson junction,”
Journal of Applied Physics, vol. 77, no. 6, 1995.
[22] R. Rouse, S. Han, and J. E. Lukens, “Flux amplification using stochastic superconducting quantum interference devices,” Applied Physics Letters, vol. 66, no. 1, 1995.
[23] P. E. Greenwood, L. M. Ward, D. F. Russell, A. Neiman, and F. Moss, “Stochastic
resonance enhances the electrosensory information available to paddlefish for prey
capture,” Phys. Rev. Lett., vol. 84, pp. 4773–4776, May 2000.
159
[24] J. A. Freund, L. Schimansky-Geier, B. Beisner, A. Neiman, D. F. Russel, T. Yakusheva, and F. Moss, “Behavioral stochastic resonance: How the noise from a daphnia
swarm enhances individual prey capture by juvenile paddlefish,” Journal of Theoretical Biology, vol. 214, no. 1, pp. 71 – 83, 2002.
[25] R. Benzi, G. Parisi, A. Sutera, and A. Vulpiani, “Stochastic resonance in climatic
change,” Tellus, vol. 34, no. 1, pp. 10–16, 1982.
[26] R. Benzi, G. Parisi, A. Sutera, and A. Vulpiani, “A theory of stochastic resonance in
climatic change,” SIAM Journal on Applied Mathematics, vol. 43, no. 3, pp. 565–578,
1983.
[27] G. Vemuri and R. Roy, “Stochastic resonance in a bistable ring laser,” Phys. Rev. A,
vol. 39, pp. 4668–4674, May 1989.
[28] T. Zhou and F. Moss, “Analog simulations of stochastic resonance,” Phys. Rev. A,
vol. 41, pp. 4255–4264, Apr 1990.
[29] T. Zhou, F. Moss, and P. Jung, “Escape-time distributions of a periodically modulated
bistable system with noise,” Phys. Rev. A, vol. 42, pp. 3161–3169, Sep 1990.
[30] R. Löfstedt and S. N. Coppersmith, “Stochastic resonance: Nonperturbative calculation of power spectra and residence-time distributions,” Phys. Rev. E, vol. 49,
pp. 4821–4831, Jun 1994.
[31] L. Gammaitoni, P. Hänggi, P. Jung, and F. Marchesoni, “Stochastic resonance,” Rev.
Mod. Phys., vol. 70, pp. 223–287, Jan 1998.
[32] M. I. Freidlin and A. D. Wentzell, Random Perturbations of Dynamical Systems. New
York, NY: Springer US, 1984.
[33] M. V. Day, “On the exponential exit law in the small parameter exit problem,”
Stochastics: An International Journal of Probability and Stochastic Processes, vol. 8,
no. 4, pp. 297–323, 1983.
[34] Y. I. Kifer, “Certain results concerning small random perturbations of dynamical
systems,” Theory of Probability & Its Applications, vol. 19, no. 3, pp. 487–505, 1975.
[35] A. Galves, E. Olivieri, and M. E. Vares, “Metastability for a class of dynamical systems
subject to small random perturbations,” The Annals of Probability, vol. 15, no. 4,
pp. 1288–1305, 1987.
[36] E. Olivieri and M. E. Vares, Large deviations and metastability. Cambridge University
Press, 2005.
[37] A. Bovier, M. Eckhoff, V. Gayrard, and M. Klein, “Metastability in reversible diffusion
processes i: Sharp asymptotics for capacities and exit times,” Journal of the European
Mathematical Society, vol. 6, no. 4, pp. 399–424, 2004.
160
[38] N. Berglund and B. Gentz, “The eyring-kramers law for potentials with nonquadratic
saddles,” Markov Processes and Related Fields, vol. 16, no. 3, pp. 549–598, 2010.
[39] N. Berglund, “Kramers’ law: Validity, derivations and generalisations,” Markov Processes and Related Fields, vol. 19, no. 3, pp. 459–490, 2011.
[40] M. I. Freidlin, “Quasi-deterministic approximation, metastability and stochastic resonance,” Physica D: Nonlinear Phenomena, vol. 137, no. 34, pp. 333 – 352, 2000.
[41] S. Herrmann, P. Imkeller, I. Pavlyukevich, and D. Peithmann, Stochastic Resonance:
A Mathematical Approach in the Small Noise Limit, vol. 194. American Mathematical
Soc., 2013.
[42] M. V. Day, “Some phenomena of the characteristic boundary exit problem,” Diffusion
processes and related problems in analysis, vol. 1, pp. 55–71, 1990.
[43] N. Berglund and B. Gentz, “Universality of first-passage-and residence-time distributions in non-adiabatic stochastic resonance,” EPL (Europhysics Letters), vol. 70,
no. 1, p. 1, 2005.
[44] N. Berglund and B. Gentz, “On the noise-induced passage through an unstable periodic orbit ii: General case,” SIAM Journal on Mathematical Analysis, vol. 46, no. 1,
pp. 310–352, 2014.
[45] I. Pavlyukevich, Stochastic Resonance. PhD thesis, Humboldt University Berlin, 2002.
[46] P. Imkeller and I. Pavlyukevich, “Model reduction and stochastic resonance,” Stochastics and Dynamics, vol. 2, no. 4, pp. 463–506, 2002.
[47] P. Imkeller and I. Pavlyukevich, “Stochastic resonance in two-state markov chains,”
Archiv der Mathematik, vol. 77, no. 1, pp. 107–115, 2001.
[48] P. Imkeller and I. Pavlyukevich, “Stochastic resonance: a comparative study of twostate models,” in Seminar on Stochastic Analysis, Random Fields and Applications
IV, pp. 141–154, Springer, 2004.
[49] S. Herrmann, P. Imkeller, and I. Pavlyukevich, Two Mathematical Approaches to
Stochastic Resonance, pp. 327–351. Berlin, Heidelberg: Springer Berlin Heidelberg,
2005.
[50] S. Herrmann, P. Imkeller, and D. Peithmann, “Large deviations for diffusions with
time periodic drift and stochastic resonance,” HU Berlin and U Nancy, 2005.
[51] S. Herrmann and P. Imkeller, “The exit problem for diffusions with time-periodic drift
and stochastic resonance,” Ann. Appl. Probab., vol. 15, pp. 39–68, 02 2005.
[52] R. Mannella, “Integration of Stochastic Differential Equations on a Computer,” International Journal of Modern Physics C, vol. 13, pp. 1177–1194, 2002.
161
[53] I. I. Gihman and A. V. Skorohod, Stochastic Differential Equations. Berlin: Springer,
1972.
[54] S. Herrmann, P. Imkeller, and D. Peithmann, “Transition times and stochastic resonance for multidimensional diffusions with time periodic drift: A large deviations
approach,” Ann. Appl. Probab., vol. 16, pp. 1851–1892, 11 2006.
[55] A. N. Kolmogorov, “Sulla Determinazione Empirica di una Legge di Distribuzione,”
Giornale dell’Istituto Italiano degli Attuari, vol. 4, pp. 83–91, 1933.
[56] N. Smirnov, “Table for estimating the goodness of fit of empirical distributions,” Ann.
Math. Statist., vol. 19, pp. 279–281, 06 1948.
[57] W. Feller, “On the kolmogorov-smirnov limit theorems for empirical distributions,”
Ann. Math. Statist., vol. 19, pp. 177–189, 06 1948.
[58] J. F. Monahan, “Evaluating the smirnov distribution function,” Center for Research
in Scientific Computation, vol. 89, p. 2, 1989.
[59] G. Marsaglia, W. W. Tsang, and J. Wang, “Evaluating kolmogorov’s distribution,”
Journal of Statistical Software, vol. 8, no. 1, pp. 1–4, 2003.
[60] R. Simard and P. L’Ecuyer, “Computing the two-sided kolmogorov-smirnov distribution,” Journal of Statistical Software, vol. 39, no. 1, pp. 1–18, 2011.
[61] E. Platen and N. Bruti-Liberati, Numerical solution of stochastic differential equations
with jumps in finance, vol. 64. Springer Science & Business Media, 2010.
[62] C. Moler, “Cleve’s corner, random thoughts, 10435 years is a very long time,” Matlab
News & Notes, pp. 12–13, Fall 1995.
162
| 10math.ST
|
InverseNet: Solving Inverse Problems with Splitting Networks
Kai Fan1∗, Qi Wei2∗ , Wenlin Wang1 , Amit Chakraborty2 , Katherine Heller1
1
Duke University, Durham, North Carolina, USA
{kai.fan, kheller}@stat.duke.edu, [email protected]
2
Siemens Corporate Technology, Princeton, New Jersey, USA
arXiv:1712.00202v1 [cs.CV] 1 Dec 2017
{qi.wei, amit.chakraborty}@siemens.com
Abstract
ments. This forward problem, generally relies on a developed physical theory which reveals the link between the
ground-truth and the measurements. Solving inverse problems involves learning the inverse mapping from the measurements to the ground-truth. Specifically, it recovers a
signal from one or a small number of degraded or noisy
measurements, which is usually ill-posed [42]. Mathematically, the goal is to reconstruct a high dimensional groundtruth x ∈ Rn from a low dimensional measurement denoted
as y ∈ Rm , which is reduced from x by a a forward model
A such that y = Ax. This forward model A is constructed
to tie the observed data y to a set of learned model parameters x. For example, in compressive sensing, y is a compressive measurement with random sampled regions and A
is the measurement matrix, e.g., a random Gaussian matrix;
in super-resolution, y is a low-resolution image and the operation A downsamples high resolution images. The main
difficulty of these underdetermined systems comes from the
operator A which has a non-trivial null space leading to an
infinite number of feasible solutions. Though most of the
inverse problems are formulated directly to the setting of
an optimization problem associated with the forward model
[41], a number of learning-based algorithms have been proposed to solve inverse problems by learning a mapping from
the measurement domain of y to the signal space of x, with
the help of large datasets and neural nets [32, 14]. 1
More recently, deep learning techniques have arisen as a
promising framework and gained great popularity for providing state-of-the-art performance on applications include
pattern analysis (unsupervised), classification (supervised),
computer vision, image processing, etc [13]. Exploiting
deep neural networks to solve inverse problems has been
explored recently [14, 40, 1, 24]. In these works, inverse
problems are viewed as a pattern mapping problem and
most existing learning-based methods propose to learn an
We propose a new method that uses deep learning techniques to solve the inverse problems. The inverse problem is cast in the form of learning an end-to-end mapping
from observed data to the ground-truth. Inspired by the
splitting strategy widely used in regularized iterative algorithm to tackle inverse problems, the mapping is decomposed into two networks, with one handling the inversion of
the physical forward model associated with the data term
and one handling the denoising of the output from the former network, i.e., the inverted version, associated with the
prior/regularization term. The two networks are trained
jointly to learn the end-to-end mapping, getting rid of a
two-step training. The training is annealing as the intermediate variable between these two networks bridges the gap
between the input (degraded version of output) and output
and progressively approaches to the ground-truth. The proposed network, referred to as InverseNet, is flexible in the
sense that most of the existing end-to-end network structure
can be leveraged in the first network and most of the existing denoising network structure can be used in the second
one. Extensive experiments on both synthetic data and real
datasets on the tasks, motion deblurring, super-resolution,
and colorization, demonstrate the efficiency and accuracy
of the proposed method compared with other image processing algorithms.
1. Introduction
Over the past decades, inverse problems have been
widely studied in image and signal processing and computer vision, e.g., denoising [15], deconvolution [2], superresolution [48] and compressive sensing [18]. An inverse
problem is resulted from the forward model which maps unknown signals, i.e., the ground-truth, to acquired/observed
information about them, which we call data or measure∗ The
1 These
algorithms refer to directly learning optimum mappings from
the observed data to their high-resolution correspondents and are different
from learning from training datasets some specific priors to be incorporated
in the regularized iterative algorithms[15, 47].
authors contributed equally to this work.
4321
end-to-end mapping from y to x [37, 40]. By leveraging the powerful approximation ability of deep neural networks, these deep learning based data-driven methods have
achieved state-of-the-art performance in many challenging
inverse problems like super-resolution [5, 14, 40], image reconstruction [35], automatic colorization [27]. More specifically, massive datasets currently enables learning end-toend mappings from the measurement domain to the target
image/signal/data domain to help deal with these challenging problems instead of solving the inverse problem by inference. A strong motivation to use neural networks stems
from the universal approximation theorem [11], which
states that a feed-forward network with a single hidden layer
containing a finite number of neurons can approximate any
continuous function on compact subsets of Rn , under mild
assumptions on the activation function. In these recent
works [5, 40, 27, 35], an end-to-end mapping from measurements y to ground-truth x was learned from the training
data and then applied to the testing data. Thus, the complicated inference scheme needed in the conventional inverse
problem solver was replaced by feeding a new measurement
through the pre-trained network, which is much more efficient. However, despite their superior performance, these
specifically-trained solvers are designed for specific inverse
problems and usually cannot be reused to solve other problems without retraining the mapping function - even when
the problems are similar. To improve the scope and generability of deep neural network models, more recently, in [7],
a splitting strategy was proposed to decompose an inverse
problem into two optimization problems, where one subproblem, related to regularization, can be solved efficiently
using trained deep neural networks, leading to an alternating direction method of multipliers (ADMM) framework
[4, 31]. This method involved training a deep convolutional
auto-encoder network for low-level image modeling, which
explicitly imposed regularization that spanned the subspace
that the ground-truth images lived in. For the sub-problem
that required inverting a big matrix, a conventional gradient
descent algorithm was used, leading to an alternating update, iterating between feed-forward propagation through a
network and iterative gradient descent. Thus, an inner loop
for gradient descent is still necessary in this framework.
A similar approach to learn approximate ISTA (Iterative
Shrinkage-Thresholding Algorithm) with neural networks
was illustrated in [21]. A more flexible splitting strategy
of training two reusable neural networks for the two subproblems within ADMM framework leading to an innerloop free update rule has been proposed in [17]. The two
pre-trained deep neural networks are flexible and reusable
in the sense that the trained network for the proximity operator can be used as a plug-and-play prior to regularize other
inverse problems sharing similar statistical characteristics
and the trained network for the inversion operator can be
(a)
(b) Iter = 1, 25, 50, 100
(c) Iter = 200, 400, 800, 1600
(d)
Figure 1. Illustration of Annealing Training. (a) the motion blurred
photos. (b-c) the left column is z the results of U-Nets and the right
column is x̂ the results of DAEs. (d) the ground-truth.
used to solve the same inverse problem for other datasets.
To leverage the advantages of both the end-to-end mapping and the splitting strategy, in this work, we propose to
learn an end-to-end mapping consisting of two networks,
with one handling the inversion associated with the forward
physical model and the other one handling the denoising,
respectively. A degraded signal, i.e., an observed data point
is fed into the inverse network to output an intermediate update and then fed into the denoising network to refine the
result. The intermediate update bridges the information gap
between the input, i.e., the degraded signal and the output, i.e., the ground-truth. The training for the proposed
InverseNet is annealing in the sense that the input of the denoising network, e.g., the denoising autoencoder, referred
to as DAE (as an example explained later) or equivalently,
the output of the inversion networks, e.g., the U-Nets (as
an example explained later), progressively becomes better
(closer to the ground-truth) following by a refined (better)
result output by the denoising network, as displayed in Fig.
1. This training leverages both the data term by the U-Nets
and a generative prior by the DAEs. More specifically (and
heuristically), the inversion network tries to restore the information lost in the forward model, and the denoising network tries to refine the result with learned details, inspired
by the two update steps of ADMM.
Contributions: We propose to learn an end-to-end mapping from the observed data to the ground-truth. Inspired by
the splitting strategy widely used to solve inverse problems
as discussed above, instead of learning one network to solve
them all, we propose to decompose the end-to-end mapping
into two parts with one handling the model inversion part
and the other one handling the denoising part. There are
several benefits to such structure: i) Any existing end-toend learning algorithm can be regarded as the inversion part
4322
For the priors R of special forms, such as kxk1 , a closedform solution for Eq. (4), i.e., a soft thresholding solution is
easily obtained. On the contrary, for some more sophiscated
regularizations, e.g., a patch based prior [15, 47], solving
Eq. (4) is nontrivial, and may require iterative methods. To
solve Eq. (3), a matrix inversion is inevitable, for which
gradient descent (GD) method is usually applied to update
z [7]. Thus, solving Eq. (4) and (3) is in general cumbersome as inner loops are required to solve these two subminimization problems. More specifically, most of computational complexity comes from the proximity operator due
to its intractability and the matrix inversion which is not
easy diagonalized.
Motivation to introduce ADMM: The proposed network
structure is inspired by the two iterative steps in ADMM
updates to solve an inverse problem. More specially, the inverse network imitates the process of solving Eq. (3) and the
denoising network imitates an algorithm to solve Eq. (4). In
this work, the inverse network exploits the U-Nets structure
[33] and the denoising network uses the DAE [43], which
will be elaborated in the following section.
and incorporated with any existing denoising network. ii)
The two network can be trained jointly to learn the mapping
to its best extent, getting rid of any two-step sub-optimal solution. iii) In the testing phase, only one feed-forward propagation is necessary to solve the inverse problem, getting
rid of any further iterations.
2. Inverse Problems
As explained above, the forward model connects the low
dimensional measurement y ∈ Rm to high dimensional
ground-truth x ∈ Rn by a linear operator A as y = Ax.
The fact that n ≥ m makes the number of parameters to
estimate larger than the number of available data points in
hand. Since A is an underdetermined measurement matrix, this imposes an ill-posed problem for finding solution
x on a new observation y. For instance, the matrix A is a
strided Gaussian convolution and not invertible for superresolution tasks in [37, 40]. To address this issue, computational solutions including approximate inference based
on Markov chain Monte Carlo (MCMC) and optimization
based on variable splitting under the ADMM framework,
were proposed and applied to different kinds of priors, e.g.,
the empirical Gaussian prior [45, 48], the Total Variation
prior [38], etc. The ADMM framework is popular due to its
low computational complexity and recent success in solving large scale optimization problems. Mathematically, the
optimization problem is formulated as
x̂ = arg min ky − Azk2 + λR(x),
x,z
s.t. z = x
3. Approach
The proposed method decomposes the end-to-end mapping from data to ground-truth into two mappings, one
corresponding to inversion of the physical model and the
other corresponding to the regularized denoising. From a
Bayesian perspective, the first mapping handles the likelihood associated with data term and the second mapping
handles the prior term. As shown in Fig. 2, the observed
data y is fed into the first network to get an approximation
z of the ground-truth. Note that z is of the same size as x
and can be regarded as a noisy version of it. The noisy one
z is then fed into the second network to get refined to the
clear version x̂. This two-network structure is also echoed
from the recent popular refining technique used in [36, 8].
More details about these two networks will be elaborated in
the following sections.
(1)
where the introduced auxiliary variable z is forced to be
equal to x, and R(x) models the structure promoted by the
prior/regularization. To leverage the befinits of ‘big data’,
the regularization can be imposed in an empirical Bayesian
way, by designing an implicit data dependent prior on x,
i.e., R(x; y) for amortized inference [40]. The augmented
Lagrangian for Eq. (1) by replacing R(x) with R(x; y) is
L(x, z, u) =
ky − Azk2 + λR(x; y) + hu, x − zi + βkx − zk2
(2)
3.1. Inversion network mapping y to z
where u is the Lagrange multiplier and β > 0 is the penalty
parameter. The conventional augmented Lagrange multiplier method that minimizes L w.r.t. x and z simultaneously, is difficult and does not exploit the fact that the objective function is separable. To tackle this problem, ADMM
decomposes the minimization into two subproblems, i.e.,
minimizations w.r.t. x and z, respectively. More specifically, the updates are as follows:
The inversion network tries to learn the inversion of the
degradation which maps the ground-truth x to the observed
data y. Note that while it may be straightforward to write
down the closed-form solution for sub-problem 3 w.r.t. z,
explicitly computing this solution is nontrivial due to the indefeasibly of inverting a big matrix. Similar to the strategy
in [17, 40], we design a deep convolutional neural network
to learn the inversion. More specifically, we have used a recently developed network structure referred to as U-Nets
[33], which was originally developed for medical image
segmentation. The U-Net architecture allows low-level information to shortcut by filtering concatenation across different layers, differing from the element-wise addition in
uk 2
k (3)
z
2β
= arg min βkx − zk + uk /2βk2 + λR(x; y) (4)
zk+1 = arg min ky − Azk2 + βkxk+1 − z +
xk+1
x
uk+1 = uk + 2β(xk+1 − zk+1 ).
(5)
4323
y
z
y
Convolution
Discriminator
Leaky ReLU
ReLU
Tanh
No
Deconvolution
Yes
y
Dropout
Bicubic
z
DAE
x̂
x
Concatenate
BatchNorm
Pixel Shuffle
U-Nets
Comparator
Figure 2. The pipeline architecture of InverseNet. Initially, the degraded input y is pre-processed by bicubic resizing depending on the
task, e.g., for super-resolution. Thus, the input y of U-Nets has the same size as the output z. Next, the output z is fed into the DAEs with
two pixel shuffling blocks. In the end, the intermediate result z and final estimation x̂ are both fed into the discriminator and comparator
with the ground-truth.
3.3. Practical Adversarial Training
ResNet [22]. Extensive experiments have demonstrated the
effectiveness of U-Nets to learn the complex mapping between high-dimensional data, such as images [9, 24].
From either the perspective of the model structure or the
ADMM update rules, z and x̂ are both approximates of
the ground-truth x. Inspired by the idea in autoencoding
beyond pixels using a learned similarity metric [26], standard negative likelihood loss (equivalent to reconstruction
loss) incorporating additional generative adversarial loss
can practically improve the quality of generated samples.
In our work, we also force the outputs of two designed networks to share one discriminator used as a binary classifier,
inducing the standard minimax loss [20] in our model, i.e.,
In general, U-Nets require the input and output to have
the same size, for the height and width at least. However,
in many inverse problems, for example, super-resolution or
compressive sensing, only a low dimensional signal is available as the input. In such cases, we initially apply the bicubic interpolation to obtain the same size input which is suitable for the U-Nets. As a summary, the architecture of our
U-Nets is shown in the left dashed box in Fig. 2.
LD = LGAN :D (z) + LGAN :D (x̂)
(6)
LG = LGAN :G (·) + λl Llikelihood (·),
(7)
3.2. Denoising network mapping z to x̂
where LGAN :D is the negative likelihood loss with respect
to classification problem, LGAN :G is the non-saturated generative loss, and “·” in Eq. (7) should be substituted by z or
x̂ to introduce two different generative loss functions.
Besides using the classifier as the discriminator, many
works [34, 3] argued that a pre-trained regression network
to match the high level features between the real images
and the fake ones can be significantly better than the discriminative loss in practice and theoretically is more robust
from the viewpoint of distribution matching. Therefore, a
regression network referred to as comparator is shared by
the outputs of two proposed nets as well. Since the transfer
learning is well established in the image domain, such as
style learning with feature matching [19] in VGG [39], we
prefer to use the pre-trained VGG or AlexNet [25] as the
comparator. In this case, to leverage both the adversarial
training and the comparator, we do not modify the discriminator loss in Eq. (6), but add an extra feature matching loss
to Eq. (7) as
The denoising network plays a role to learn from the
dataset a signal prior that can deal with any inverse problems. In optimization algorithms for solving inverse problems, signal priors are usually cast in the form of a denoiser,
more specifically, a proximity operator [10]. From the geometrical perspective, the proximity operator projects a noisy
data point into the feasible sets spanned by the signal prior.
Thus, any existing denoising network can be exploited in
the proposed framework. In our work, we propose to a specially designed denoising auto-encoders with pixel shuffling
(or sub-pixel) trick [37]. Similar to the inversion net, pixel
shuffling convolutional network is designed to deal with inputs and outputs of the same size by periodically reordering
the pixels in each channel mapping a high resolution image to to the scale as the same as the low dimensional image. The resulting denoising auto-encoders does not have
a bottle-neck shape structure as each layer shares the same
filter size. This allows us to transform z into a tensor that
has the same size as y, and concatenate this tensor with y
if desired in the regularization of amortized inference. The
detailed architecture of the proposed DAE shown in Fig. 2.
LG = LGAN :G (·) + λl Llikelihood (·) + λf Lf eature (·).
(8)
4324
PSNR
Wiener filter (baseline)
Robust motion deblur [46]
Neural motion deblur [6]
Pix2Pix [23] w. comparator
Ours
SSIM
Wiener filter (baseline)
Robust motion deblur [46]
Neural motion deblur [6]
Pix2Pix [23] w. comparator
Ours
Analogous to the traditional optimization for generative adversarial loss, LD and LG are iteratively minimized with
respect to the parameters of the discriminator, U-Nets, and
DAEs.
3.4. Underlying Annealing Training
The proposed adversarial training leads to an underlying annealing training for DAEs where the input (the output
of U-Nets) at early training stage is extremely noisy, and
then progressively approaches to a clearer version with further training. This incremental tuning nature in adversarial training allows the training to imitate the inner loop of
ADMM updates, where the difference u in Eq. (5) gradually becomes constant and negligible. This is the main reason why we call our model as InverseNet. Unlike the traditional DAEs, in which the input data is contaminated by the
pre-defined noise, e.g., Gaussian noise with fixed variance,
the output of U-Nets or the input of DAEs in the proposed
network is contaminated by non-stationary and time-variant
noise without an explicit analytic form. If we remove the
DAEs part from our model, we found that the performance
gradually becomes worse during the training, which can be
explained by the instability of adversarial training [3]. The
existence of the DAEs make the network be able to purify
the complicated noisy output from the inversion network,
leading to a much stable and efficient training.
Generality A critical point for learning-based methods
is whether the method generalizes to other problems. More
specifically, how does a method that is trained on a specific
dataset perform when applied to another dataset? To what
extent can we reuse the trained network without re-training?
Compared with the reusable matrix inversion neural networks learned with pure noise proposed in [40, 17], this
inversion network is less flexible due to the joint training
of the inversion and denoising network. However, because
the learning is based on problem formulation, though not
fully reusable, the well-trained neural networks have better
generalization ability compared with the other end-to-end
learning networks. For example, the well-trained networks
for image motion deblurring on PASCAL VOC dataset [16]
can be directly applied to the same task on ImageNet [12]
with no necessity of fine-tuning, if the degrade operator A
remains the same, since these two datasets are both natural images. However, if a significant domain changes to the
dataset, such as MRI, all networks-reusable algorithms will
fail and require re-training as well as our approach.
CUB
22.42
25.03
25.73
23.67
28.39
CUB
0.6572
0.7459
0.8853
0.7554
0.9421
CelebA ImageNet
20.92
20.44
25.06
23.37
25.76
24.74
23.59
22.05
34.02
28.87
CelebA ImageNet
0.7020 0.6357
0.8052 0.7283
0.9649 0.9074
0.8553 0.7335
0.9738 0.9446
Table 1. PSNR and SSIM Comparison on motion deblurring
geNet [25], and the learning tasks considered include motion deblurring, ×4 super-resolution, and joint ×2 superresolution and colorization. Note that the degraded operator
A in unknown during training, and is only used to generate
the measurement and ground-truth paired data. Our code
will be available on the repository https://github.
com/.
4.1. Motion Deblurring
The target of motion deblurring is to recover a sharp nonblurred image from a single motion-blurred image, which
has been a fundamental problem in computational imaging
[46]. The motion deblurring is challenging as both the blurring kernel and the latent non-blurred images are unknown
in most cases, leading to an ill-posed inverse problem. Note
that the size of an image after motion deblurring remains the
same. Thus, any resizing of the input images, e.g., bicubic
interpolation or pixel shuffling for the initialization of our
approach is unnecessary and can be simply removed from
the pipeline. We trained our model on all datasets with size
64 × 64 images without any other pre-processing.
Qualitative Results The proposed InverseNet is compared with the baseline optimization methods by Wiener
filtering, the robust non-blind motion deblurring [46], and
another deep learning approach method neural motion deblurring [6] visually in Fig. 3. Note that in [46], the algorithm for a single image motion deblurring requires to feed
the blurring kernel, while other methods including ours are
all blind recoveries. As shown in Fig. 3, the traditional single image processing algorithms including Wiener filtering
and robust motion deblurring suffered heavily from ring artifacts while the neural motion deblurring and the proposed
InverseNet gave much better and high-quality recoveries.
The ring artifact can be explained by the fact that the chosen
convolutional kernel to blur images has larger variance and
is more challenging.
Quantitative Results The PSNR and SSIM were calcu-
4. Experiments and Results
In this section, we provide experimental results and analysis on the proposed InverseNet for solving inverse problems. The datasets used in our experiments include CaltechUCSD Birds-200-2011 (CUB) [44], Large-scale CelebFaces Attributes (CelebA) [29], PASCAL VOC and Ima4325
(a) Blurred image
(b) Baseline: Wiener filter
(c) Robust motion deblurring [46]
(d) Neural motion deblurring [6]
(e) Ours: InverseNet
(f) Ground-truth
Figure 3. Example deblurred results from different methods and three different datasets (zooming in can see details).
Figure 5. Transfer Learning: Testing on the same images from
Fig. 3(a) with well-trained model on VOC dataset.
InverseNet outperformed the other methods with a significant improvement, which is in consistence with the qualitative results in Fig. 3.
Transfer Training We also checked to what extent
the well-trained model could be adapted to other datasets.
Specifically, we trained the InverseNet on PASCAL VOC
and tested it on CUB, celebA and ImageNet, with the results shown in Fig. 5. The quantitative results is shown in
Table 2. As we argued before, the similar image domain
between ImageNet and VOC leads the PSNR does not decrease much, i.e., from 28.87 to 28.52, compared with using the InverseNet trained on ImageNet itself. This demonstrated the generality of the learned InverseNet.
Annealing Training We also validated the importance
of annealing training on the datasets (more intensive results
Figure 4. Annealing training at iteration 1, 25, 50, 100 on celebA
dataset. 1st and 3rd Rows: the error images between z and model
estimation x̂, and the ones between z and ground-truth x. 2nd and
4th Rows: the annealing empirical distribution of pixel level noise
w.r.t x̄ − z and x − z.
lated on the same testing datasets across all compared algorithms and summarized in Table. 1. Clearly, the proposed
4326
Figure 6. Ring Effect Remover on ImageNet for 64 × 64 → 256 × 256. 1st Row: LR images; 2nd Row: The bicubic interpolation results
(having ring artifacts); 3rd Row: Results by InverseNet; 4th Row: HR ground-truth.
Metric CUB CelebA ImageNet
PSNR 25.65 27.90
28.52
SSIM 0.9105 0.9373 0.9303
at the final training stage. This potentially enables the DAEs
to depress noise with variances of different scales added to
the input, which is different from traditional DAEs, and allows more robust training.
Table 2. Performance with well-trained model on VOC
4.2. Super-Resolution
Super-resolution is another popular inverse problem in
image processing, and it aims at the restoration of high frequencies to enrich details in an image based on a set of
prior examples with low resolution (LR) and corresponding high resolution (HR) images. The degraded operator A
can source from quantization error, limitation of the sensor from the capturing camera, the presence of blurriness
and the use of downsampling operators to reduce the image
resolution for storage purposes. It is well known that superresolution is ill-posed, since for each LR image the space of
corresponding HR images can be very large.
In this experiment, we first synthesized a challenge
×4 downsampling operator A, which was a channel-wise
strided convolution and may result in the ring artifact on
low resolution images (see detailed form in supplementary
materials) and trained the super-resolution model on two
are available in the supplementary materials). If we only
trained the model with U-Nets with a discriminator (this
will recover the model pix2pix [23]) and a comparator, the
training was usually very unstable and we have to do cherrypick from results in every iteration. However, with the help
of refining network DAEs, the performance was boosted up
significantly, reflected by both the quantitative results summarized in Table. 1 and the qualitative results displayed in
Fig. 1. Additionally, we also visualized the residual images between the output of U-Nets and the final output of
DAEs or the ground-truth in Fig. 4. The pixel level noise
approximately followed a Gaussian distribution at each iteration, but with varying mean and variance, leading to a
non-stationary and time variant noise. It was observed that
the mean and variance gradually became smaller and stable
4327
datasets, CUB and ImageNet. Our task is trying to superresolve images from 64 × 64 to 256 × 256. In addition, we
also compared with SRGAN [28] in supplementary materials.
Ring Artifact Remover The difficulty of the superresolution is highly affected by the shape and scale of the
convolution kernel in the degradation model. Convolving
an image with a wider and flatter convolution kernel leads
to more aliasing of high frequency information in spectral
domain thus leads to a more challenging problem. Usually, ring artifact can be resulted from this heavy aliasing
on the spectral domain. On the contrary, a narrower and
sharper convolutional kernel results to less information loss
thus is less challenging. In this task, we used a flat kernel
of size 7 × 7 which can cause heavy ring artifact. Fig. 6
shows the super-resolution results on held-out testing ImageNet dataset. In this case, the bicubic interpolation failed
to remove the ring effects produced during dowmsampling.
However, our end-to-end trained InverseNet can well tackle
this problem by greatly depressing the ring effect.
4.3. Jointly Super-Resolution and Colorization
A more challenging task is to enhance both spectral
and spatial resolutions from one single band image. More
specifically, the enhancement is a combination of superresolution and colorization (hallucinating a plausible color
version of a colorless image), which can be considered as
the compound of two degraded operators. Thus, our single channel colorless low resolution image was obtained by
convolving a kernel A of size 9 × 9 × 3 × 1 and downsampling with a stride 2 in both horizontal and vertical directions. Note that this is different from and more challenging
than the conventional multi-band image fusion, e.g., fusing
a multispectral image and a high spatial panchromatic image (also referred to as pansharpening [30]), in the sense
that only one low-spatial low-spectral resolution image is
available. Additionally, our inverse task is also different
from the traditional colorization in the CIE Lab color space.
In this section, we have tested the ability to perform joint
×2 super-resolution and colorization from one single colorless LR image on the dataset celebA and CUB. The celebA
32 × 32 → 64 × 64 mainly includes faces images with less
color variance, so it is an easier dataset compared with CUB
wild bird images 64 × 64 → 128 × 128. The results are
displayed in Fig. 7. Visually, the joint super-resolved and
colorized images are very similar with their ground-truth
with variances for some details. It is especially interesting
to observe that the restored high-resolution images looked
more natural than the ground-truth of some ‘outlier’ images.
This results from the fact that the model was trained from
massive images in which most look normal. For example,
the person in the third column had purple hairs which rarely
Figure 7. Joint Super-resolution and colorization for CelebA and
CUB datasets. (top): colorless LR images; (middle): recovery by
InverseNet; bottom: full color HR images.
appeared in the dataset. Feeding its degraded version, i.e.,
a blurred colorless image to the trained InverseNet gave a
face with most probable, popular and thus ‘natural’ brown
hairs as shown in the second row. More results are available
in the supplementary materials.
5. Conclusion
In this paper we proposed the InverseNet to solve inverse
problems by end-to-end mapping. To take the advantage
of the efficiency (in testing phase) of end-to-end learning
and the flexibility brought by splitting strategy simultaneously, the mapping was decomposed into two neural networks, i.e., the inversion network and the denoising one.
The former one was designed to learn the inversion of the
physical forward model associated with the data term and
the latter one was to learn a proximity operator or a projection onto the ground-truth signal space associated with the
prior term. The two pre-trained deep neural networks were
trained jointly using prepared data pairs (xi , yi ), getting rid
of any two-step separate training. Experiments and analysis
on various datasets demonstrated the efficiency and accuracy of the proposed method. In future work we hope to
further extend the proposed method to tackle the ‘learn to
learn’ problem.
4328
References
[16] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and
A. Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303–
338, 2010.
[17] K. Fan∗ , Q. Wei∗ , L. Carin, and K. Heller. An inner-loop free
solution to inverse problems using deep neural networks. In
Advances in Neural Information Processing Systems, Long
Beach, CA, USA, Dec. 2017.
[18] M. A. Figueiredo, R. D. Nowak, and S. J. Wright. Gradient projection for sparse reconstruction: Application to compressed sensing and other inverse problems. IEEE J. Sel.
Topics Signal Process., 1(4):586–597, 2007.
[19] L. A. Gatys, A. S. Ecker, and M. Bethge. A neural algorithm
of artistic style. arXiv preprint arXiv:1508.06576, 2015.
[20] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in Neural Information
Processing Systems, pages 2672–2680, 2014.
[21] K. Gregor and Y. LeCun. Learning fast approximations of
sparse coding. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pages 399–406,
2010.
[22] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In Proc. IEEE Int. Conf. Comp. Vision
and Pattern Recognition (CVPR), June 2016.
[23] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Imageto-image translation with conditional adversarial networks.
arXiv preprint arXiv:1611.07004, 2016.
[24] K. H. Jin, M. T. McCann, E. Froustey, and M. Unser. Deep
convolutional neural network for inverse problems in imaging. IEEE Trans. Image Process., 26(9):4509–4522, Sept
2017.
[25] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Advances in Neural Information Processing Systems, pages
1097–1105, 2012.
[26] A. B. L. Larsen, S. K. Sønderby, H. Larochelle, and
O. Winther. Autoencoding beyond pixels using a learned
similarity metric. In Proceedings of the 33rd International Conference on International Conference on Machine
Learning-Volume 48, pages 1558–1566. JMLR. org, 2016.
[27] G. Larsson, M. Maire, and G. Shakhnarovich. Learning representations for automatic colorization. In Proc. European
Conf. Comp. Vision (ECCV), pages 577–593. Springer, 2016.
[28] C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham,
A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, et al.
Photo-realistic single image super-resolution using a generative adversarial network. arXiv preprint arXiv:1609.04802,
2016.
[29] Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face
attributes in the wild. In Proc. IEEE Int. Conf. Comp. Vision
(ICCV), pages 3730–3738, 2015.
[30] L. Loncan, L. B. Almeida, J. M. Bioucas-Dias, X. Briottet,
J. Chanussot, N. Dobigeon, S. Fabre, W. Liao, G. Licciardi, M. Simoes, J.-Y. Tourneret, M. Veganzones, G. Vivone,
Q. Wei, and N. Yokoya. Hyperspectral pansharpening: a review. IEEE Geosci. Remote Sens. Mag., 3(3):27–46, Sept.
2015.
[1] J. Adler and O. Öktem. Solving ill-posed inverse problems using iterative deep neural networks. arXiv preprint
arXiv:1704.04058, 2017.
[2] M. V. Afonso, J. M. Bioucas-Dias, and M. A. Figueiredo.
Fast image recovery using variable splitting and constrained
optimization. IEEE Trans. Image Process., 19(9):2345–
2356, 2010.
[3] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan.
arXiv preprint arXiv:1701.07875, 2017.
[4] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein.
Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and
Trends R in Machine Learning, 3(1):1–122, 2011.
[5] J. Bruna, P. Sprechmann, and Y. LeCun. Super-resolution
with deep convolutional sufficient statistics. arXiv preprint
arXiv:1511.05666, 2015.
[6] A. Chakrabarti. A neural approach to blind motion deblurring. In European Conference on Computer Vision, pages
221–235. Springer, 2016.
[7] J. Chang, C.-L. Li, B. Poczos, B. Kumar, and A. C. Sankaranarayanan. One network to solve them all—solving linear inverse problems using deep projection models. arXiv preprint
arXiv:1703.09912, 2017.
[8] Q. Chen and V. Koltun. Photographic image synthesis with cascaded refinement networks. arXiv preprint
arXiv:1707.09405, 2017.
[9] Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and
O. Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In International Conference on Medical Image Computing and Computer-Assisted
Intervention, pages 424–432. Springer, 2016.
[10] P. Combettes and J.-C. Pesquet. Proximal splitting methods
in signal processing. In H. H. Bauschke, R. S. Burachik,
P. L. Combettes, V. Elser, D. R. Luke, and H. Wolkowicz,
editors, Fixed-Point Algorithms for Inverse Problems in Science and Engineering, Springer Optimization and Its Applications, pages 185–212. Springer New York, 2011.
[11] B. C. Csáji. Approximation with artificial neural networks.
Faculty of Sciences, Etvs Lornd University, Hungary, 24:48,
2001.
[12] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. FeiFei. Imagenet: A large-scale hierarchical image database.
In Computer Vision and Pattern Recognition, 2009. CVPR
2009. IEEE Conference on, pages 248–255. IEEE, 2009.
[13] L. Deng, D. Yu, et al. Deep learning: methods and applications. Foundations and Trends R in Signal Processing, 7(3–
4):197–387, 2014.
[14] C. Dong, C. C. Loy, K. He, and X. Tang.
Image
super-resolution using deep convolutional networks. IEEE
Transactions on Pattern Analysis and Machine Intelligence,
38(2):295–307, Feb 2016.
[15] M. Elad and M. Aharon. Image denoising via sparse and
redundant representations over learned dictionaries. IEEE
Trans. Image Process., 15(12):3736–3745, 2006.
4329
[47] J. Yang, J. Wright, T. S. Huang, and Y. Ma. Image superresolution via sparse representation. IEEE Trans. Image Process., 19(11):2861–2873, 2010.
[48] N. Zhao, Q. Wei, A. Basarab, N. Dobigeon, D. Kouamé, and
J. Y. Tourneret. Fast single image super-resolution using a
new analytical solution for `2 − `2 problems. IEEE Trans.
Image Process., 25(8):3683–3697, Aug. 2016.
[31] S. Lu, M. Hong, and Z. Wang. A nonconvex splitting method
for symmetric nonnegative matrix factorization: Convergence analysis and optimality. IEEE Transactions on Signal
Processing, 65(12):3120–3135, June 2017.
[32] Y. Lu, M. Inamura, and M. del Carmen Valdes. Superresolution of the undersampled and subpixel shifted image
sequence by a neural network. International Journal of
Imaging Systems and Technology, 14(1):8–15, 2004.
[33] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and
Computer-Assisted Intervention, pages 234–241. Springer,
2015.
[34] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved techniques for training gans. In
Advances in Neural Information Processing Systems, pages
2234–2242, 2016.
[35] J. Schlemper, J. Caballero, J. V. Hajnal, A. Price, and
D. Rueckert. A deep cascade of convolutional neural
networks for MR image reconstruction. arXiv preprint
arXiv:1703.00555, 2017.
[36] S. Shankar, D. Robertson, Y. Ioannou, A. Criminisi, and
R. Cipolla. Refining architectures of deep convolutional neural networks. In Proc. IEEE Int. Conf. Comp. Vision and Pattern Recognition (CVPR), pages 2212–2220, Las Vegas, NV,
USA, 2016.
[37] W. Shi, J. Caballero, F. Huszár, J. Totz, A. P. Aitken,
R. Bishop, D. Rueckert, and Z. Wang. Real-time single image and video super-resolution using an efficient sub-pixel
convolutional neural network. In Proc. IEEE Int. Conf.
Comp. Vision and Pattern Recognition (CVPR), pages 1874–
1883, 2016.
[38] M. Simoes, J. Bioucas-Dias, L. Almeida, and J. Chanussot.
A convex formulation for hyperspectral image superresolution via subspace-based regularization. IEEE Trans. Geosci.
Remote Sens., 53(6):3373–3388, Jun. 2015.
[39] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556, 2014.
[40] C. K. Sønderby, J. Caballero, L. Theis, W. Shi, and F. Huszár.
Amortised MAP inference for image super-resolution. arXiv
preprint arXiv:1610.04490, 2016.
[41] A. Tarantola. Inverse problem theory and methods for model
parameter estimation. SIAM, 2005.
[42] A. Tikhonov and V. Arsenin. Solutions of ill-posed problems.
Scripta series in mathematics. Winston, 1977.
[43] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol.
Extracting and composing robust features with denoising autoencoders. In Proc. Int. Conf. Machine Learning (ICML),
pages 1096–1103. ACM, 2008.
[44] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie.
The caltech-ucsd birds-200-2011 dataset. 2011.
[45] Q. Wei, N. Dobigeon, and J.-Y. Tourneret. Bayesian fusion
of multi-band images. IEEE J. Sel. Topics Signal Process.,
9(6):1117–1127, Sept. 2015.
[46] L. Xu and J. Jia. Two-phase kernel estimation for robust
motion deblurring. In European Conference on Computer
Vision, pages 157–170. Springer, 2010.
4330
Appendices
A. Training details
A.1. Adversarial training algorithm
We define a soft GAN loss with the following form
t
Lsof
GAN (x, l) = −l · log D(x) − (1 − l) · log(1 − D(x))
(9)
where l ∈ [0, 1] is the soft label for data x. The explicit forms of the Discriminator (D) loss and the generator (G) loss are
t
sof t
sof t
LD = Lsof
GAN (x, 0.99) + LGAN (z, 0.01) + LGAN (x̂, 0.01)
LG (z) =
LG (x̂) =
t
2
2
Lsof
GAN (z, 0.99) + λr · kz − xk2 + λf · kC(z) − C(x)k2
t
2
2
Lsof
GAN (x̂, 0.99) + λr · kx̂ − xk2 + λf · kC(x̂) − C(x)k2
(10)
(11)
(12)
where z is the output of U-Nets, x̂ is the output of DAE and C(·) is the comparator output. In our experiment, we used FC6
layer of the AlexNet. The tunable hyper-parameters λr = λf = 0.5 were used in our experiment. Experimentally, this soft
loss can alleviate the vulnerability of neural networks to adversarial examples.
Instead of pre-training the generator as in SRGAN, we jointly trained the whole model, i.e., the generator and discriminator
simultaneously. The detailed optimization updates for adversarial training are summarized in Algorithm 1. We chose K = 1
in all experiments, and Adam [?] optimizer was applied in the gradient descent for all parameter updates. For Adam, the
learning rate is 10−4 , and β1 = 0.5, β2 = 0.999, = 10−8 . In addition, the gradient is clipped by norm 5 for each net. The
batch size of each iteration is 36 for all experiments. For batch normalization, momentum is 0.9 and = 10−5 .
Algorithm 1 Adversarial training for InverseNet
for t = 1, 2, . . . do
Randomly get batch data pairs (x, y);
for k = 1, 2, . . . , K do
Update parameter of Discriminator with gradient ∇LD ;
end for
Update parameter of U-Nets with gradient ∇LG (z);
Update parameter of DAEs with gradient ∇LG (x̂);
end for
A.2. Model structure
In Table 3, we describe the model structure for U-Nets with input of size 256 × 256. The structure of DAEs is summarized
in Table 4). Beside of the pixel shuffling layer, all the other layers can share the same structure as the sizes, i.e., the height
and width of an input image and an output image are the same.
B. More results for motion deblurring
We tested our motion deblurred model on another natural image as shown in Fig. 8. The model was trained on 128 × 128
patches from ImageNet, and tested on one Wonder Woman poster image with size 256 × 256 and 512 × 512. We also
compared it with the blind neural motion deblurring algorithm [6] on 256 × 256 2 . As shown in Fig. 8, the deblurred image
using InverseNet is much more clear and closer to the original image than the neural deblurred one. In the 512 × 512 size
deblurring, visually, the restored image by InverseNet is almost exactly the same with the original one as displayed in Fig. 9.
C. More results for ring artifact remover super-resolution
We include the results of SRGAN [28] for comparison as shown in Fig. 10. For SRGAN, except downgraded kernel, we
follow the training instruction in the paper by first pretraining the generator for 106 iterations and then fine-tuning the model
2 We
used the codes offered by the authors in https://github.com/ayanc/ndeblur.
4331
Table 3. Network hyper-parameters of U-Nets
Layer Name
data
e1
e2
e3
e4
e5
e6
e7
e8
d1
d2
d3
d4
d5
d6
d7
d8
Dimension
256 × 256 × 3
128 × 128 × 16
64 × 64 × 32
32 × 32 × 64
16 × 16 × 128
8 × 8 × 128
4 × 4 × 128
2 × 2 × 128
1 × 1 × 128
2 × 2 × 256
4 × 4 × 256
8 × 8 × 256
16 × 16 × 256
32 × 232 × 128
64 × 64 × 64
128 × 1282 × 32
256 × 256 × 3
Layer Operations
Conv(4, 4, 3, 16)-‘SAME’
Leaky Relu-Conv(4, 4, 16, 32)-‘SAME’-Batch Norm
Leaky Relu-Conv(4, 4, 32, 64)-‘SAME’-Batch Norm
Leaky Relu-Conv(4, 4, 64, 128)-‘SAME’-Batch Norm
Leaky Relu-Conv(4, 4, 128, 128)-‘SAME’-Batch Norm
Leaky Relu-Conv(4, 4, 128, 128)-‘SAME’-Batch Norm
Leaky Relu-Conv(4, 4, 128, 128)-‘SAME’-Batch Norm
Leaky Relu-Conv(4, 4, 128, 128)-‘SAME’-Batch Norm
Relu-Conv Trans(4, 4, 128, 128)-‘SAME’-Batch Norm-Cancat(e7)
Relu-Conv Trans(4, 4, 256, 128)-‘SAME’-Batch Norm-Cancat(e6)
Relu-Conv Trans(4, 4, 256, 128)-‘SAME’-Batch Norm-Cancat(e5)
Relu-Conv Trans(4, 4, 256, 128)-‘SAME’-Batch Norm-Cancat(e4)
Relu-Conv Trans(4, 4, 256, 64)-‘SAME’-Batch Norm-Cancat(e3)
Relu-Conv Trans(4, 4, 128, 32)-‘SAME’-Batch Norm-Cancat(e2)
Relu-Conv Trans(4, 4, 64, 16)-‘SAME’-Batch Norm-Cancat(e1)
Relu-Conv Trans(4, 4, 32, 3)-‘SAME’-Tanh
Table 4. Network hyper-parameters of DAEs
Input Dimension
256 × 256 × 3
64 × 64 × 48
64 × 64 × 128
64 × 64 × 64
64 × 64 × {32, 3}
64 × 64 × 35
64 × 64 × 64
64 × 64 × 128
64 × 64 × 48
(a) Blurred Image
Layer
periodical pixel shuffling
Conv(4, 4, 48, 128)-‘SAME’-Batch Norm-Relu
Conv(4, 4, 128, 64)-‘SAME’-Batch Norm-Relu
Conv(4, 4, 64, 32)-‘SAME’-Batch Norm-Relu
Concatenate in Channel
Conv(4, 4, 35, 64)-‘SAME’-Batch Norm-Relu
Conv(4, 4, 64, 128)-‘SAME’-Batch Norm-Relu
Conv(4, 4, 128, 48)-‘SAME’-Batch Norm-Relu
periodical pixel shuffling
(b) Neural Deblur
(c) InverseNet
Output Dimension
64 × 64 × 48
64 × 64 × 128
64 × 64 × 64
64 × 64 × 32
64 × 64 × 35
64 × 64 × 64
64 × 64 × 128
64 × 64 × 48
256 × 256 × 3
(d) Original Image
Figure 8. Motion deblurring for a 256 × 256 image with model trained on 128 × 128 images.
with discriminator included for another 106 iterations. Since the author did not release the codes, this implementation was
based on our own and we tried our best to fine tune these parameters to the best performance. As shown in Fig. 10, the
super-resolved result by SRGAN did not remove the ring effect but could sharpen some details of images. On the contrary,
the results of InverseNet successfully removed the ring artifact while keeping the details.
4332
Figure 9. Motion deblurring for 512 × 512 image with model trained on 128 × 128 images. From left to right: blurred image, deblurred
image by InverseNet, and original image.
D. Visualization of degraded kernel A
D.1. Motion deblurring
The degradation matrix A in motion deblurring is a square matrix corresponding to a 2-D convolution. If the convolution
is implemented with periodic boundary conditions, i.e., the pixels out of an image is padded with periodic extension of itself,
the matrix H is a block circulant matrix with circulant blocks (BCCB). The first row of the matrix A is the motion blurring
convolution kernel and the other row vectors are rotated one element to the right relative to the preceding row vector. The
9 × 9 motion blurring convolution kernel is displayed as below.
Figure 11. 9 × 9 motion blurring kernel.
4333
(a) Low Resolution Images
(b) Bicubic
(c) SRGAN
(d) InverseNet
(e) High Resolution Image
Figure 10. Ring Effect Remover on ImageNet for 64 × 64 → 256 × 256.
D.2. Super-resolution
The degradation matrix A in super-resolution can be decomposed as the multiplication of a convolution matrix H (as in
motion blurring) and a down-sampling matrix S, i.e., A = SH. The down-sampling matrix S represents the regular 2-D
decimation by keeping one pixel every d pixels in both horizontal and vertical directions, where d is the sampling ratio. The
matrix A is equivalent to the strided convolution widely used in CNN architecture. The 3 × 3 convolution kernel used in our
experiment is displayed as below. The down-sampling ratio d is fixed to 4 in the experiments, corresponding to shrinking the
4334
size of an image from 256 × 256 to 64 × 64.
Figure 12. 3 × 3 convolution kernel in super-resolution.
D.3. Joint super-resolution and colorization
The degradation matrix A in joint super-resolution and colorization can be decomposed as the product of a convolution
matrix H, a down-sampling matrix S and a spectral degradation matrix L. The matrices H and S are similarly defined as in
Section D.2. The main difference here is that the kernel of H is channel wise in the sense that the convolution kernel for each
channel is different, as shown in Fig. 13. The role of spectral degradation L is making the average of the R, G, B channels to
get a one-channel grayscale image.
Figure 13. Three 9 × 9 convolution kernels for red (left), green (middle) and blue (right) channels.
4335
| 1cs.CV
|
PARAMETER ESTIMATION UNDER MODEL UNCERTAINTIES BY ITERATIVE
COVARIANCE APPROXIMATION
arXiv:1612.04059v2 [math.ST] 23 Nov 2017
O. Lang, M. Lunglmayr and M. Huemer
Institute of Signal Processing
Johannes Kepler University Linz
Altenbergerstraße 69, 4040 Linz, Austria
ABSTRACT
We propose a novel iterative algorithm for estimating a deterministic but unknown parameter vector in the presence of
model uncertainties. This iterative algorithm is based on a
system model where an overall noise term describes both, the
measurement noise and the noise resulting from the model
uncertainties. This overall noise term is a function of the true
parameter vector, allowing for an iterative algorithm. The
proposed algorithm can be applied on structured as well as
unstructured models and it outperforms prior art algorithms
for a broad range of applications.
In contrast to the LS estimator and the BLUE, total least
squares (TLS) estimation techniques incorporate model errors. E.g., for independent and identically distributed (i.i.d.)
model errors with Gaussian PDF, the maximum likelihood
(ML) solution of the TLS problem was analyzed in [4]. However, in many practical applications H has some sort of structure as it is the case for Toeplitz or Hankel matrices. Then, the
model errors are clearly not i.i.d. any more. Structured total
least squares (STLS) techniques have been developed to deal
with these kind of problems [5–7]. An overview of different
TLS and STLS methods can be found in [8–10].
Index Terms— Robust Estimation, Model Uncertainties,
iterative BLUE
In this work we compare our novel approach with two iterative algorithms, which serve as performance reference in
the remainder of this paper. The first one, introduced in [4], is
an approach for solving the maximum likelihood (ML) problem based on classical expectation-maximization (EM) [11].
This algorithm, referred to as ML-EM algorithm, treats the
model errors as random and allows for an incorporation of
the model error variance. By doing so, a uniform variance
for every element in H was assumed. The second one represents an algorithm from the class of STLS approaches and is
introduced in [12]. This iterative algorithm is called the structured total least norm (STLN) algorithm and it is capable of
dealing with structured measurement matrices. This approach
treats the model errors as deterministic but unknown. Hence,
it prevents the usage of model error variances.
1. INTRODUCTION
The linear model
y = Hx + n
(1)
is frequently used in many areas of signal processing. Here,
y ∈ RNy ×1 is the vector of measurements, x ∈ RNx ×1 is a
deterministic but unknown parameter vector, H ∈ RNy ×Nx
is the measurement matrix with Ny > Nx and full rank, and
n ∈ RNy ×1 is zero mean measurement noise with known covariance matrix Cnn . The probability density function (PDF)
of n is otherwise arbitrary. Linear classical estimators such as
the least squares (LS) estimator or the best linear unbiased estimator (BLUE) [1,2] assume that the measurement matrix H
is perfectly known. In practice, this assumption often does not
hold. A prominent case is where H is a convolution matrix
that is itself estimated from an imperfectly measured system
output. The error in H is often neglected since it is unknown.
There exist several ways to account for the errors in H.
Two prominent algorithms that are related to the approach in
this work can be found in [3]. These algorithms were derived for the task of image restoration, where the point-spread
function that distorts the image is considered to be the sum
of a known mean and an unknown zero-mean random part.
It also provides an algorithm in the Bayesian context. In this
work, however, classical estimation is considered. Hence, no
prior distribution about x is assumed.
In this paper, we propose a novel iterative algorithm that
incorporate information about the model error variances.
Moreover, this algorithm can be employed on structured as
well as unstructured problems. In contrast to the ML-EM
algorithm, the algorithm is capable of incorporating different
variances for every element of H. A difference to the STLN
algorithm is that the proposed algorithm treats the model errors as random variables, allowing to incorporate the model
error variances. All three algorithms require solving an inverse linear problem at each iteration. Simulation examples
are presented which show that the proposed algorithm is able
to outperform both competing algorithms in a mean square
error (MSE) sense for a broad range of model error and noise
variances.
The proposed iterative algorithm is based on a system
model where an overall noise term describes both, the measurement noise and the noise resulting from the model uncertainties. The covariance matrix of this overall noise term is
evaluated for different cases. Considering the model errors
as random with known second order statistics (but otherwise arbitrary PDF) is motivated by practical examples such
as multiple-input multiple-output (MIMO) communication
channels or beamforming [13–16].
The remainder of this paper is organized as follows: In
Sec. 2, the underlying system model is introduced. Here we
distinguish between unstructured and structured measurement
matrices. For the structured case, we considered convolution
matrices in this work. However, extensions to other kind of
structured matrices are easily possible. The proposed iterative
algorithm is discussed in Sec. 3. Simulation results demonstrating its performance are given in Sec. 4.
Notation:
Lower-case bold face variables (a, b,...) indicate vectors, and
upper-case bold face variables (A, B,...) indicate matrices.
We further use R and C to denote the set of real and complex
numbers, respectively, (·)T to denote transposition, In×n to
denote the identity matrix of size n × n, and 0m×n to denote
the all-zero matrix of size m × n. If the dimensions are clear
from the context we simply write I and 0, respectively. E[·]
denotes the expectation operator, [·]i the ith element of a vector and [·]i,j the element of a matrix at the ith row and the j th
column.
2. SYSTEM MODEL
This section describes the underlying model used in the remainder of this paper. In a first step, the measurement matrix is assumed to be unstructured and the model uncertainties
are assumed to be independent. Afterwards, H is assumed to
be a structured convolution matrix built from an estimated or
measured impulse response. Hence, H is a special form of a
Toeplitz matrix and, as it will be shown, results in correlated
model uncertainties.
2.1. Unstructured Measurement Matrices
We denote Ĥ as the measured or estimated measurement matrix and assume it comes along with error variances for every entry. The error variances assembled in a matrix of the
same size as Ĥ is denoted as V ∈ RNy ×Nx . Furthermore,
the errors are assumed to be independent zero mean random
variables. The measurements are modeled as
y = Hx + n = (Ĥ + B)x + n,
(2)
where H = Ĥ + B, with Ĥ being the estimated measurement
matrix and B being a zero mean random matrix. In (2), H
and B are unknown while Ĥ is known. We further rewrite (2)
according to
y
=
Ĥx + Bx + n
| {z }
(3)
w
=
Ĥx + w,
(4)
with the new overall noise vector w. This noise vector combines the measurement noise with the noise from the model
uncertainties. Let bTi be the ith row of B, then the ith element
of w is given by
[w]i = bTi x + [n]i
(5)
Since [w]i is evaluated as the scalar product of a vector with
zero mean random elements with an unknown but deterministic vector plus [n]i , [w]i has zero mean and its variance in
dependence of the unknown parameter vector x can be derived as
σi2 =[V]i,1 |[x]1 |2 + [V]i,2 |[x]2 |2 + · · · + [V]i,Nx |[x]Nx |2
+ [Cnn ]i,i .
(6)
All variances assembled in a covariance matrix are combined
in
Cww = diag(V|x|2 ) + Cnn ,
(7)
where the term |x|2 represents a column vector of the
element-wise absolute squares of the vector x.
2.2. Convolution Matrices
We will now assume that H is a linear convolution matrix
constructed from the impulse response h ∈ RNh ×1 of a linear
system such that Hx describes the convolution of the underlying sequences h[n] and x[n]. An extension to other structured
measurement matrices is easily possible. Let H = Ĥ + B
have the dimension Ny × Nx where Ny = Nx + Nh − 1. The
ith column of the convolution matrix is defined as
(i−1)×1
(i−1)×1
0
0
,
,
h
[H]:,i =
[Ĥ]:,i =
ĥ
(Nx −i)×1
(N
−i)×1
x
0
0
(i−1)×1
0
e
[B]:,i =
∀i = 1, . . . , Nx
(8)
0(Nx −i)×1
where ĥ is the estimated impulse response and e is the
unknown error of ĥ with known error covariance matrix
Cee ∈ RNh ×Nh . In this case, the model uncertainties of Ĥ
are clearly not independent anymore, leading to a different
calculation of Cww .
Let b′i = [B]:,i denote the ith column of B. The subsequent column b′i+1 can be derived by shifting down the elements of b′i by one position:
01×(Ny −1)
0
′
bi+1 = (Ny −1)×(Ny −1)
b′ = Db′i . (9)
I
0(Ny −1)×1 i
With that, the product Bx in (3) follows to
Bx =b′1 [x]1 + b′2 [x]2 + . . . + b′Nx [x]Nx
Nx −1
= [x]1 I + [x]2 D + . . . + [x]Nx D
=P(x)b′1 .
(10)
b′1
(11)
(12)
With this result, w can be written as w = P(x)b′1 + n with
the covariance matrix
i
h
H
(13)
Cww =E (P(x)b′1 ) (P(x)b′1 ) + Cnn
=P(x)Cb′1 b′1 P(x)H + Cnn .
(14)
The covariance matrix Cb′1 b′1 follows from (8) and the covariance matrix of the estimation error e according to
Cee
0Nh ×(Nx −1)
′
′
Cb1 b1 = (Nx −1)×Nh
∈ RNy ×Ny .
0
0(Nx −1)×(Nx −1)
(15)
Note that this formulation allows for two sources of correlations. The first source comes from the structure in H.
The second source of correlation comes from Cee , which
describes the errors in ĥ. Hence, the iterative algorithm introduced in the next section is capable of dealing with both kind
of correlations.
3. ITERATIVE ALGORITHM
An ideal but theoretical estimator is the BLUE applied on the
linear model in (2) using the true H according to
x̂ = HH C−1
nn H
−1
HH C−1
nn y.
(16)
This theoretical estimator is referred to as BLUE with perfect
model knowledge. Similarly, the BLUE applied on the linear
model in (4), incorporating the estimated measurement matrix
Ĥ but the true covariance matrix Cww follows as
−1 H −1
x̂ = (ĤH C−1
Ĥ Cww y
ww Ĥ)
(17)
and is referred to as BLUE with perfect knowledge of Cww
[17]. The determination of the true Cww according to (7) or
(14), however, requires the knowledge of the true parameter
vector. To overcome this problem, we propose the iterative
algorithm described below. Its basic idea is to make an initial
guess of the parameter vector termed x̂0 (the index denotes
the algorithm’s iteration number). This first guess could, e.g.,
origin from an LS estimation which does not incorporate any
noise statistics. x̂0 is then used to estimate Ĉww,0 in (7) or
(14). This estimated covariance matrix is then incorporated
by the BLUE in order to yield a better estimate x̂1 and so on.
This procedure is summarized as shown in Algorithm 1.
The proposed algorithm is of similar complexity as the
ML-EM and STLN algorithms. It performs a weighting of
the measurements according to Ĉww,k , which incorporates
the model error variances as well as the measurement noise
Initialization:
LS estimation
−1
T
x̂0 = Ĥ Ĥ
ĤT y;
for k ← 0 to Niter do
estimate Cww,k according to (7) or (14) using x̂k
instead of x ;
−1
x̂k+1 = ĤT Ĉ−1
Ĥ
ĤT Ĉ−1
ww,k
ww,k y ;
end
Algorithm 1: proposed algorithm
variances. In the case of H being a convolution matrix, even
the covariances of the estimated impulse response are considered in order to improve the estimation.
Note that for both cases Ĉww,k is almost surely invertible
since Cnn serves as a regularization term in (7) and (14).
Although convergence cannot be ensured, simulations
showed that divergence is a rare exception for reasonable
values of V.
A stopping criteria can be implemented in several ways.
One possibility is to stop the iterations when x̂ does not significantly change from one iteration to the next. Simulations
showed that the major performance gain is usually achieved
after the first iteration. Hence, a predefined number of iterations may be utilized instead of a stopping criteria.
Naturally, there exists at least one case where the iterations yield no performance gain. If Ĉww,k is a scaled identity
matrix, the proposed algorithm reduces to the ordinary LS
estimator, preventing any performance increase. This is, e.g.,
the case when the following two conditions hold: a) The measurement matrix is unstructured and V has the same variance
at every element. b) the noise covariance matrix Cnn is a
scaled identity matrix.
We note that a similar iterative application of the BLUE
was applied in [17–19] for channel impulse response estimation in wireless communication applications. Compared to
them, the proposed algorithm is applicable to various applications with structured or unstructured model uncertainties.
In [20] investigations of a similar procedure as the presented
algorithm can be found but only for a very simplified model
compared to the investigations in this work. As a result of
that, the algorithms presented in [17–20] are not considered
in the following simulations. Here, we rather compare the
proposed algorithm with the STLS algorithm [12], the MLEM algorithm [4] as well as the estimators in (16) and (17).
4. SIMULATION RESULTS
In this example, H ∈ R7×3 is a convolution matrix and
describes the discrete convolution of the impulse response
h[n] with signal x[n]. The vector notations of h[n] and x[n]
are given by h ∈ R5×1 and x ∈ R3×1 , respectively. For
the simulations, the impulse responses is randomly generated from a Gaussian distribution with mean E[h] = 05×1
ML-EM algorithm
STLN algorithm
proposed algorithm
BLUE with perfect knowledge of Cww
BLUE with perfect model knowledge
Average MSE
10−3
10−5
10−7 −8
10
10−7
10−6
10−5
10−4
10−3
σn2
Fig. 1. Average MSEs of different iterative algorithms plotted over the noise variance σn2 .
ML-EM algorithm
proposed algorithm
STLN algorithm
Average MSE
10−5
10−6
0
2
4
6
Iteration
8
10
gorithm comes with its own termination criterium for which
we choose ǫ = 10−10 [12], the proposed algorithm and the
ML-EM algorithm were executed for Niter = 10 iterations.
However, as we discuss below, Niter could be reduced significantly. The resulting MSE values averaged over the elements
of the MSE vector are presented in Fig. 1. This figure shows
that the proposed algorithm attains the performance bound
given by the BLUE with perfect knowledge of Cww and
outperforms the competing algorithms especially for low σn2 .
The performance gain is more than one order of magnitude
in MSE for small noise variances. For large noise variances
all investigated algorithms perform approximately equal. The
reason for this is that the model uncertainties vanish compared
to the large measurement noise samples in that case. For the
same reason, the gap between all considered algorithms and
the BLUE with perfect model knowledge decreases with
increasing noise variance. Simulations showed that, if one
would have chosen Cee to be a scaled identity matrix, the
STLN algorithm would have similar performance as the proposed algorithm for very low noise variances. Furthermore,
simulations showed that the performance gain approximately
stays the same for other values of x. Fig. 2 shows the convergence behavior of the algorithms for σn2 = 10−6 . First
of all, one recognizes that the ML-EM algorithm is not able
to significantly improve the estimation accuracy compared
to the initial LS estimation in this example. Furthermore, it
shows that the STLN algorithm as well as the proposed algorithm achieve most of their performance gain during the first
iteration. This extremely fast convergence allows to reduce
the number of iterations to one without any significant loss in
performance.
5. CONCLUSIONS
Fig. 2. Average MSE values plotted over the number of iterations.
and covariance matrix Chh = I5×5 . The input signal is
T
chosen to be x = 1 0.5 0.25 . For the first analysis, the noise covariance matrix is a scaled identity matrix
Cnn = σn2 I7×7 , where the scaling factor σn2 is varied between 10−8 and 10−3 . The impulse response estimation step
is assumed to yield zero
mean errors with error covariance
matrix Cee = diag 10−4 10−5 10−6 10−6 10−6 .
For this model, the proposed algorithm in Sec. 3 is compared
with the BLUE with perfect model knowledge in (16), the
BLUE with perfect knowledge of Cww in (17), the ML-EM
algorithm, and the STLN algorithm. For the latter one the l2
norm minimization, a tolerance ǫ = 10−10 and D = I5×5
is chosen. Furthermore, X ( (2.1) in [12]) is identified to be
the first Nh columns of P(x) in (12). For more details on
these parameters we refer to [12]. For the ML-EM algorithm
σh2 is set to the mean value of V [4]. While the STLN al-
In this work, a novel iterative algorithm for estimating an unknown but deterministic parameter vector in the presence of
model errors and measurement noise is presented. This algorithm iteratively estimates the covariance matrix of an overall
noise term, which describes the effects of the measurement
noise as well as the noise resulting from the model uncertainty. This overall noise term was analyzed for unstructured
model errors and for the case where the measurement matrix
is a convolution matrix. For the latter case, simulation results
are presented demonstrating the performance gain compared
to competing algorithms. Convergence curves demonstrate
the extremely fast convergence of the proposed algorithm,
which achieves almost its optimum estimation accuracy after
a single iteration.
6. REFERENCES
[1] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory, Vol. 1. Prentice Hall, 1993.
[2] L. Lin and H. C. So, “Best linear unbiased estimator algorithm for received signal strength based localization,”
In Signal Processing Conference, 2011 19th European,
pp. 1989–1993, Aug 2011.
[3] V. Mesarovic, N. Galatsanos, R. Molina, and A. Katsaggelos, “Hierarchical Bayesian image restoration
from partially-known blurs,” In Acoustics, Speech and
Signal Processing, 1998. Proceedings of the 1998 IEEE
International Conference on, Vol. 5, pp. 2905–2908
vol.5, May 1998.
[14] D. Gesbert, M. Shafi, D. Shiu, P. J. Smith, and
A. Naguib, “From theory to practice: an overview of
MIMO space-time coded wireless systems,” In IEEE
Journal on Selected Areas in Communications, Vol. 21,
No. 3, pp. 281–302, Apr 2003.
[15] S. Shahbazpanahi, A. B. Gershman, Z.-Q. Luo, and
K. M. Wong, “Robust adaptive beamforming for
general-rank signal models,” In IEEE Transactions on
Signal Processing, Vol. 51, No. 9, pp. 2257–2269, Sept
2003.
[4] A. Wiesel, Y. C. Eldar, and A. Yeredor, “Linear Regression With Gaussian Model Uncertainty: Algorithms and
Bounds,” In IEEE Transactions on Signal Processing,
Vol. 56, No. 6, pp. 2194–2205, June 2008.
[16] M. Huemer, C. Hofbauer, and J. B. Huber, “NonSystematic Complex Number RS Coded OFDM by
Unique Word Prefix,” In IEEE Transactions on Signal
Processing, Vol. 60, No. 1, pp. 285–299, Jan 2012.
[5] M. Pilanci, O. Arikan, and M. C. Pinar, “Structured
Least Squares Problems and Robust Estimators,” In
IEEE Transactions on Signal Processing, Vol. 58, No. 5,
pp. 2453–2465, May 2010.
[17] S. Ozen, C. Pladdy, M. J. Fimoff, S. M. Nerayanuru,
and M. D. Zoltowski, “Approximate best linear unbiased channel estimation for frequency selective multipath channels with long delay spreads,” In Proc. Asilomar Conf. Signals, Syst., Comput., Vol. 1, pp. 1122–
1127, Nov 2003.
[6] S. V. Huffel, H. Park, and J. B. Rosen, “Formulation
and solution of structured total least norm problems for
parameter estimation,” In IEEE Transactions on Signal
Processing, Vol. 44, No. 10, pp. 2464–2474, Oct 1996.
[7] A. Kukush, I. Markovsky, and S. V. Huffel, “Consistency of the structured total least squares estimator in
a multivariate errors-in-variables model,” In Journal of
Statistical Planning and Inference, Vol. 133, No. 2, pp.
315 – 358, 2005.
[8] B. D. Moor, “Structured total least squares and L2 approximation problems,” In Linear Algebra and its Applications, Vol. 188, pp. 163 – 205, 1993.
[9] G. H. Golub and C. F. van Loan, “An Analysis of the
Total Least Squares Problem,” In SIAM Journal on Numerical Analysis, Vol. 17, No. 6, pp. 883–893, 1980.
[10] I. Markovsky and S. V. Huffel, “Overview of total leastsquares methods,” In Signal Processing, Vol. 87, No. 10,
pp. 2283 – 2302, 2007.
[11] A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum Likelihood from Incomplete Data via the EM Algorithm,” In Journal of the Royal Statistical Society. Series B (Methodological), Vol. 39, No. 1, pp. 1–38, 1977.
[12] J. B. Rosen, H. Park, and J. Glick, “Total Least Norm
Formulation and Solution for Structured Problems,” In
SIAM Journal on Matrix Analysis and Applications,
Vol. 17, No. 1, pp. 110–126, 1996.
[13] Y. C. Eldar, “Minimax estimation of deterministic parameters in linear models with a random model matrix,”
In IEEE Transactions on Signal Processing, Vol. 54,
No. 2, pp. 601–612, Feb 2006.
[18] Z. Tang, R. C. Cannizzaro, G. Leus, and P. Banelli,
“Pilot-Assisted Time-Varying Channel Estimation for
OFDM Systems,” In IEEE Transactions on Signal Processing, Vol. 55, No. 5, pp. 2226–2238, May 2007.
[19] M. Ghogho and A. Swami, “Improved channel estimation using superimposed training,” In IEEE 5th Workshop on Signal Processing Advances in Wireless Communications, pp. 110–114, July 2004.
[20] L. Lista, “The bias of the unbiased estimator: A study
of the iterative application of the BLUE method,” In
Nuclear Instruments and Methods in Physics Research,
Vol. 764, pp. 82 – 93, 2014.
| 10math.ST
|
Optimal lower bounds for universal relation, and for samplers and
finding duplicates in streams∗
Michael Kapralov†
Jelani Nelson‡
arXiv:1704.00633v1 [cs.CC] 3 Apr 2017
David P. Woodruffk
Jakub Pachocki§
Zhengyu Wang¶
Mobin Yahyazadeh∗∗
April 4, 2017
Abstract
In the communication problem UR (universal relation) [KRW95], Alice and Bob respectively
receive x, y ∈ {0, 1}n with the promise that x 6= y. The last player to receive a message must
output an index i such that xi 6= yi . We prove that the randomized one-way communication
n
)})
complexity of this problem in the public coin model is exactly Θ(min{n, log(1/δ) log2 ( log(1/δ)
for failure probability δ. Our lower bound holds even if promised support(y) ⊂ support(x).
As a corollary, we obtain optimal lower bounds for ℓp -sampling in strict turnstile streams for
0 ≤ p < 2, as well as for the problem of finding duplicates in a stream. Our lower bounds do
not need to use large weights, and hold even if promised x ∈ {0, 1}n at all points in the stream.
We give two different proofs of our main result. The first proof demonstrates that any algorithm A solving sampling problems in turnstile streams in low memory can be used to encode
subsets of [n] of certain sizes into a number of bits below the information theoretic minimum.
Our encoder makes adaptive queries to A throughout its execution, but done carefully so as
to not violate correctness. This is accomplished by injecting random noise into the encoder’s
interactions with A, which is loosely motivated by techniques in differential privacy. Our correctness analysis involves understanding the ability of A to correctly answer adaptive queries
which have positive but bounded mutual information with A’s internal randomness, and may be
of independent interest in the newly emerging area of adaptive data analysis with a theoretical
computer science lens. Our second proof is via a novel randomized reduction from Augmented
Indexing [MNSW98] which needs to interact with A adaptively. To handle the adaptivity we
identify certain likely interaction patterns and union bound over them to guarantee correct interaction on all of them. To guarantee correctness, it is important that the interaction hides
some of its randomness from A in the reduction.
∗
This paper is a merger of [NPW17], and of work of Kapralov, Woodruff, and Yahyazadeh.
EPFL. [email protected].
‡
Harvard University. [email protected]. Supported by NSF grant IIS-1447471 and CAREER award
CCF-1350670, ONR Young Investigator award N00014-15-1-2388, and a Google Faculty Research Award.
§
OpenAI. [email protected]. Work done while affiliated with Harvard University, under the support of ONR
grant N00014-15-1-2388.
¶
Harvard University. [email protected]. Supported by NSF grant CCF-1350670.
k
IBM Research Almaden. [email protected].
∗∗
Sharif University of Technology. [email protected]. Work done while an intern at EPFL.
†
1
Introduction
In turnstile ℓ0 -sampling, a vector z ∈ Rn starts as the zero vector and receives coordinate-wise
updates of the form “zi ← zi + ∆” for ∆ ∈ {−M, −M + 1, . . . , M }. During a query, one must
return a uniformly random element from support(x) = {i : zi 6= 0}. The problem was first defined
in [FIS08], where a data structure (or “sketch”) for solving it was used to estimate the Euclidean
minimum spanning tree, and to provide ε-approximations of a point set P in a geometric space
(that is, one wants to maintain a subset S ⊂ P such that for any set R in a family of bounded VCdimension, such as the set of all axis-parallel rectangles, ||R ∩ S|/|S| − |R ∩ P |/|P || < ε). Sketches
for ℓ0 -sampling were also used to solve various dynamic graph streaming problems in [AGM12a]
and since then have been crucially used in almost all known dynamic graph streaming algorithms1 ,
such as for: connectivity, k-connectivity, bipartiteness, and minimum spanning tree [AGM12a],
subgraph counting, minimum cut, and cut-sparsifier and spanner computation [AGM12b], spectral
sparsifiers [AGM13], maximal matching [CCHM15], maximum matching [AGM12a, BS15, Kon15,
AKLY16, CCE+ 16, AKL17], vertex cover [CCHM15, CCE+ 16], hitting set, b-matching, disjoint
paths, k-colorable subgraph, and several other maximum subgraph problems [CCE+ 16], densest
subgraph [BHNT15, MTVV15, EHW16], vertex and hyperedge connectivity [GMT15], and graph
degeneracy [FT16]. For an introduction to the power of ℓ0 -sketches in designing dynamic graph
stream algorithms, see the recent survey of McGregor [McG14, Section 3]. Such sketches have
also been used outside streaming, such as in distributed algorithms [HPP+ 15, PRS16] and data
structures for dynamic connectivity [KKM13, Wan15, GKKT15].
Given the rising importance of ℓ0 -sampling in algorithm design, a clear task is to understand
the exact complexity of this problem. The work [JST11] gave an Ω(log2 n)-bit space lower bound
for data structures solving even the case M = 1 which fail with constant probability, and otherwise
whose query responses are (1/3)-close to uniform in statistical distance. They also gave an upper
bound for M ≤ poly(n) with failure probability δ, which in fact gave min{kzk0 , Θ(log(1/δ))} uniform samples from the support of z, using space O(log2 n log(1/δ)) (here kzk0 denotes | support(z)|).
Thus we say their data structure actually solves the harder problem of ℓ0 -samplingk for k =
Θ(log(1/δ)) with failure probability δ, where in ℓ0 -samplingk the goal is to recover min{kzk0 , k}
uniformly random elements, without replacement, from support(z). The upper and lower bounds
in [JST11] thus match up to a constant factor for k = 1 and δ a constant. We note though in many
settings, even if the final application desires constant failure probability, ℓ0 -samplingk with either
failure probability o(1) or k > 1 (or both) is needed as a subroutine (see Figure 1).
Universal relation. The work of [JST11] obtains its lower bound for ℓ0 -sampling (and some
other problems) via reductions from universal relation (UR). The problem UR was first defined
in [KRW95] and arose in connection with work of Karchmer and Wigderson on circuit depth lower
bounds [KW90]. For f : {0, 1}n → {0, 1}, D(f ) is the minimum depth of a fan-in 2 circuit over the
basis {¬, ∨, ∧} computing f . Meanwhile, the (deterministic) communication complexity C(f ) is
defined as the minimum number of bits that need to be communicated in a correct protocol for Alice
and Bob to solve the following communication problem: Alice receives x ∈ f −1 (0) and Bob receives
y ∈ f −1 (1) (and hence in particular x 6= y), and they must both agree on an index i ∈ [n] such that
xi 6= yi . It is shown in [KW90] that D(f ) = C(f ), where they then used this correspondence to show
a tight Ω(log2 n) depth lower bound for monotone circuits solving undirected s-t connectivity. The
1
The spectral sparsification algorithm of [KLM+ 14] is a notable exception.
1
reference
[FIS08]
[AGM12a]
[AGM12a]
[AGM12a]
[AGM12a]
[AGM12b]
[AGM12b]
[AGM12b]
[AGM12b]
[AGM12b]
[CCHM15]
[BS15]
[Kon15]
[AKLY16]
[AKL17]
[CCE+ 16]
[BHNT15]
[MTVV15]
[EHW16]
[GMT15]
[FT16]
problem
Euclidean minimum spanning tree
connectivity2
k-connectivity2
bipartiteness2
minimum spanning tree
subgraph counting
minimum cut
cut sparsifiers
spanners
spectral sparsifiers
maximal matching
maximum matching (unweighted)
maximum matching (weighted)
maximum matching
maximum matching
maximum matching
maximum matching
vertex cover
hitting set
b-matching
disjoint paths
k-colorable subgraph
densest subgraph
densest subgraph
densest subgraph
vertex connectivity
hyperedge connectivity
graph degeneracy
distribution
ℓ0
any
any
any
any
ℓ0
any
any
any
any
ℓ0
ℓ0
ℓ0
any
ℓ0
ℓ0
ℓ0
k > 1?
yes
δ = o(1)?
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
ℓ0
ℓ0
ℓ0
any
yes
yes
ℓ0
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
Figure 1: Guarantees needed by various works using samplers as subroutines. The last two columns
indicate whether the work needs to use a sampler that returns k samples at a time when queried
for some k > 1, or for some subconstant failure probability δ even to achieve failure probability
1/3 in the main application. The “distribution” column indicates the output distribution needed
from the sampler for the application (“any” means a support-finding subroutine is sufficient, i.e. it
suffices for a query to return any index i for which zi 6= 0).
2
work of [KRW95] then proposed a strategy to separate the complexity classes NC1 and P: start with
a function f on log n bits requiring depth Ω(log n), then “compose” it with itself k = log n/ log log n
times (see [KW90] for a precise definition of composition). If one could prove a strong enough direct
sum theorem for communication complexity after composition, even for a random f , such a k-fold
composition would yield a function that is provably in P (and in fact, even in NC2 ), but not in
NC1 . Proving such a direct sum theorem is still wide open, and the statement that it is true is
known as the “KRW conjecture”; see for example the recent works [GMWW14, DM16] toward
resolving this conjecture. As a toy problem en route to resolving it, [KRW95] suggested proving
a direct sum theorem for k-fold composition of a particular function UR that they defined. That
task was positively resolved in [EIRS91] (see also [HW90]).
The problem UR abstracts away the function f , and Alice and Bob are simply given x, y ∈
{0, 1}n with the promise that x 6= y. The players must then agree on any index i with xi 6= yi .
The deterministic communication complexity of UR is nearly completely understood, with upper
and lower bounds that match up to an additive 3 bits, even if one imposes an upper bound on the
number of rounds of communication [TZ97]. Henceforth we also consider a generalized problem
URk , where the output must be min{k, kx − yk0 } distinct indices on which x, y differ. We also use
UR⊂ , UR⊂
k to denote the variants when promised support(y) ⊂ support(x), and also Bob knows
kxk0 . Clearly UR, URk can only be harder than UR⊂ , UR⊂
k , respectively.
More than twenty years after its initial introduction in connection with circuit depth lower
bounds, Jowhari et al. in [JST11] demonstrated the relevance of UR in the randomized oneway communication model for obtaining space lower bounds for certain streaming problems, such
as various sampling problems and finding duplicates in streams. In the one-way version, Bob
simply needs to find such an index i after a single message from Alice, and we only charge Alice’s
single message’s length as the communication cost. If R→,pub
(f ) denotes the randomized one-way
δ
communication complexity of f in the public coin model with failure probability δ, [JST11] showed
(UR).
that the space complexity of FindDuplicate(n) with failure probability δ is at least R→,pub
7
+δ
8
8
In FindDuplicate(n), one is given a length-(n + 1) stream of integers in [n], and the algorithm must
output some element i ∈ [n] which appeared at least twice in the stream (note that at least one
such element must exist, by the pigeonhole principle). The work [JST11] then showed a reduction
demonstrating that any solution to ℓ0 -sampling with failure probability δ in turnstile streams
immediately implies a solution to FindDuplicate(n) with failure probability at most (1 + δ)/2 in the
same space (and thus the space must be at least R→,pub
δ (UR)). The same result is shown for ℓp 15
+ 16
16
P
sampling for any p > 0, in which the output index should equal i with probability |xi |p /( j |xj |p ),
and a similar result is shown even if the distribution on i only has to be close to this ℓp -distribution
in variational distance (namely, the distance should be bounded away from 1). It is then shown in
[JST11] that Rδ→,pub (UR) = Ω(log2 n) for any δ bounded away from 1. The approach used though
unfortunately does not provide an improved lower bound for δ ↓ 0.
Seemingly unnoticed in [JST11], we first point out here that the lower bound proof for UR in
that work actually proves the same lower bound for the promise problem UR⊂ . This observation
has several advantages. First, it makes the reductions to the streaming problems trivial (they were
already quite simple when reducing from UR, but now they are even simpler). Second, a simple
2
[AGM12a] writes their algorithm as only needing δ a constant, but for a different definition of support-finding:
when the data structure fails, it should output Fail instead of behaving arbitrarily. They then cite [JST11] as providing
the sampler they use, but unfortunately [JST11] does not solve this variant of this problem. This issue can be avoided
by using [JST11] with δ < 1/ poly(n) so that whp no failures occur throughout their algorithm.
3
reduction from UR⊂ to sampling problems provides space lower bounds even in the strict turnstile
model, and even for the simpler support-finding streaming problem for which when queried is allowed
to return any element of support(z), without any requirement on the distribution of the index
output. Both of these differences are important for the meaningfulness of the lower
bound. This
n
is because in dynamic graph streaming applications, typically z is indexed by 2 for some graph
on n vertices, and ze is the number of copies of edge e in some underlying multigraph. Edges then
are not deleted unless they had previously been inserted, thus only requiring correctness for strict
turnstile streams. Also, for every single application mentioned in the first paragraph of Section 1
(except for the two applications in [FIS08]), the known algorithmic solutions which we cited as using
ℓ0 -sampling as a subroutine actually only need a subroutine for the easier support-finding problem.
Finally, third and most relevant to our current work’s main focus, the straightforward reductions
from UR⊂ to the streaming problems we are considering here do not suffer any increase in failure
probability, allowing us to transfer lower bounds on R→,pub
(UR⊂ ) for small δ to lower bounds on
δ
various streaming problems for small δ. The work [JST11] could not provide lower bounds for the
streaming problems considered there in terms of δ for small δ.
We now show simple reductions from UR⊂ to FindDuplicate(n) and from UR⊂
k to supportfindingk . In support-findingk we must report min{k, kzk0 } elements in support(z). In the claims
below, δ is the failure probability for the considered streaming problem.
Claim 1. Any one-pass streaming algorithm for FindDuplicate(n) must use R→,pub
(UR⊂ ) space.
δ
Proof. We reduce from UR⊂ . Suppose there were a space-S algorithm A for FindDuplicate(n).
Alice creates a stream consisting of all elements of support(x) and runs A on those elements, then
sends the memory contents of A to Bob. Bob then continues running A on n + 1 − kxk0 arbitrarily
chosen elements of [n]\ support(y). Then there must be a duplicate in the resulting concatenated
stream, i satisfies xi 6= yi iff i is a duplicate.
Claim 2. Any one-pass streaming algorithm for support-findingk in the strict turnstile model must
n
use R→,pub
(UR⊂
k ) bits of space, even if promised that z ∈ {0, 1} at all points in the stream.
δ
Proof. This is again via reduction from UR⊂
k . Let A be a space-S algorithm for support-findingk
in the strict turnstile model. For each i ∈ support(x), Alice sends the update zi ← zi + 1 to A.
Alice then sends the memory contents of A to Bob. Bob then for each i ∈ support(y) sends the
update zi ← zi − 1 to A. Now note that z is exactly the indicator vector of the set {i : xi 6= yi }.
Claim 3. Any one-pass streaming algorithm for ℓp -sampling for any p ≥ 0 in the strict turnstile
n
model must use R→,pub
(UR⊂
k ) bits of space, even if promised z ∈ {0, 1} at all points in the stream.
δ
Proof. This is via straightforward reduction from support-findingk , since reporting min{k, kzk0 }
elements of support(z) satisfying some distributional requirements is only a harder problem than
finding any min{k, kzk0 } elements of support(z).
The reductions above thus raise the question: what is the asymptotic behavior of R→,pub
(UR⊂
δ
k )?
Our main contribution: We prove for any δ bounded away from 1 and k ∈ [n], R→,pub
(UR⊂
δ
k)=
Θ(min{n, t log2 (n/t)}) where t = max{k, log(1/δ)}. Given known upper bounds in [JST11], our
lower bounds are optimal for FindDuplicate(n), support-finding, and ℓp -sampling for any 0 ≤ p < 2
.99
for nearly the full range of n, δ (namely, for δ > 2−n ). Also given an upper bound of [JST11],
4
our lower bound is optimal for ℓ0 -samplingk for nearly the full range of parameters n, k, δ (namely,
for t < n.99 ). Previously no lower bounds were known in terms of δ (or k). Our main theorem:
2
Theorem 1. For any δ bounded away from 1 and 1 ≤ k ≤ n, R→,pub
(UR⊂
k ) = Θ(min{n, t log (n/t)}).
δ
We give two different proofs of Theorem 1 (in Sections 3 and 4). Our upper bound is also new,
though follows by minor modifications of the upper bound in [JST11] and thus we describe it in
the appendix. The previous upper bound was O(min{n, t log2 n}). We also mention here that it
is known that the upper bound for both URk and ℓ0 -samplingk in two rounds (respectively, two
passes) is only O(t log n) [JST11]. Thus, one cannot hope to extend our new lower bound to two
or more passes, since it simply is not true.
1.1
Related work
The question of whether ℓ0 -sampling is possible in low memory in turnstile streams was first asked
in [CMR05, FIS08]. The work [FIS08] applied ℓ0 -sampling as a subroutine in approximating the
cost of the Euclidean minimum spanning tree of a subset S of a discrete geometric space subject
to insertions and deletions. The algorithm given there used space O(log3 n) bits to achieve failure
probability 1/poly(n) (though it is likely that the space could be improved to O(log2 n log log n) with
a worse failure probability, by replacing a subroutine used there with a more recent ℓ0 -estimation
algorithm of [KNW10]). As mentioned, the currently best known upper bound solves ℓ0 -samplingk
using O(t log2 n) bits [JST11], which Theorem 1 shows is tight.
For ℓp -sampling, conditioned on not failing, the data structure should output i with probability
(1 ± ε)|xi |p /kxkpp . The first work to realize its importance came even earlier than for ℓ0 -sampling:
[CK04] showed that an ℓ2 -sampler using small memory would lead to a nearly space-optimal streaming algorithm for multiplicatively estimating kxk3 in the turnstile model, but did not know how to
implement such a data structure. The first implementation was given in [MW10], achieving space
poly(ε−1 log n) with δ = 1/poly(n). . For 1 ≤ p ≤ 2 the space was improved to O(ε−p log3 n) bits
for constant δ [AKO11]. In [JST11] this bound was improved to O(ε− max{1,p} log(1/δ) log 2 n) bits
for failure probability δ when 0 < p < 2 and p 6= 1. For p = 1 the space bound achieved by [JST11]
was a log(1/ε) factor worse: O(ε−1 log(1/ε) log(1/δ) log 2 n) bits.
For finding a duplicate item in a stream, the question of whether a space-efficient randomized
algorithm exists was asked in [Mut05, Tar07]. The question was positively resolved in [GR09],
which gave an O(log3 n)-space algorithm with constant failure probability. An improved algorithm
was given in [JST11], using O(log(1/δ) log 2 n) bits of space for failure probability δ.
2
Overview of techniques
We now describe our two proofs of Theorem 1. For the upper bound, [JST11] achieved O(t log2 n),
but in the appendix we show that slight modifications to their approach yield O(min{n, t log2 (n/t)}).
Our main contribution is in proving an improved lower bound. Assume t < cn for some sufficiently
small constant c (since otherwise we already obtain an Ω(n) lower bound). In both our lower bound
proofs in this regime, the proof is split into two parts: we show R→,pub
(UR⊂ ) = Ω(log 1δ log2 logn 1 )
δ
δ
2 n
and R→,pub
(UR⊂
.99
k ) = Ω(k log k ) separately. We give an overview the former here, which is the
more technically challenging half. Our two proofs of the latter are in Sections 3.2 and 4.2.
5
2.1
Lower bound proof via encoding subsets and an adaptivity lemma
Our first proof of the lower bound on Rδ→,pub (UR⊂ ) is via an encoding argument. Fix m. A
randomized encoder is given a set S ⊂ [n] with |S| = m and must output an encoding ENC(S),
and a decoder sharing public randomness with the encoder must be able to recover S given only
ENC(S). We consider such schemes in which the decoder must succeed with probability 1, and the
n
) bits
encoding length is a random variable. Any such encoding must use Ω(log(nm )) = Ω(m log m
in expectation for some S.
There is a natural, but sub-optimal approach to using a public-coin one-way protocol P for
UR⊂ to devise such an encoding/decoding scheme. The encoder pretends to be Alice with input x
being the indicator set of S, then lets ENC(S) be the message M Alice would have sent to Bob. The
decoder attempts to recover S by iteratively pretending to be Bob m times, initially pretending to
have input y = 0 ∈ {0, 1}n , then iteratively adding elements found in S to y’s support. Henceforth
let 1T ∈ {0, 1}n denote the indicator vector of a set T ⊂ [n].
Algorithm 1 Simple Decoder.
1: procedure DEC(M )
2:
T ←∅
3:
for r = 1, . . . , m do
4:
Let i be Bob’s output upon receiving message M from Alice when Bob’s input is 1T
5:
T ← T ∪ {i}
6:
end for
7:
return T
8: end procedure
One might hope to say that if the original failure probability were δ < 1/m, then by a union
bound, with constant probability every iteration succeeds in finding a new element of S (or one
could even first apply some error-correction to x so that the decoder could recover S even if only
a constant fraction of iterations succeeded). The problem with such thinking though is that this
decoder chooses y’s adaptively! To be specific, P being a correct protocol means
∀x, y ∈ {0, 1}n , P(P is correct on inputs x, y) ≥ 1 − δ,
s
(1)
where s is the public random string that both Alice and Bob have access to. The issue is that
even in the second iteration (when r = 2), Bob’s “input” 1T depends on s, since T depends on the
outcome of the first iteration! Thus the guarantee of (1) does not apply.
One way around the above issue is to realize that as long as every iteration succeeds, T is always a
subset of S. Thus it suffices for the following event E to occur: ∀T ⊂ S, P is correct on inputs 1S , 1T .
Then Ps (¬E) ≤ 2m δ by a union bound, which is at most 1/2 for m = ⌊log2 (1/δ)⌋ − 1. We have
n
}).
thus just shown that R→,pub
(UR⊂ ) = Ω(min{n, log(nm )}) = Ω(min{n, log 1δ log log(1/δ)
δ
Our improvement is as follows. Our new decoder again iteratively tries to recover elements of
S as before. We will give up though on having m iterations and hoping for all (or even most) of
them to succeed. Instead, we will only have R = Θ(log 1δ log logn 1 ) iterations, and our aim is for
δ
the decoder to succeed in finding a new element in S for at least a constant fraction of these R
iterations. Simplifying things for a moment, let us pretend for now that all R iterations do succeed
in finding a new element. ENC(S) will then be Alice’s message M , together with the set B ⊂ S
6
n
of size m − R not recovered during the R rounds, explicitly written using ⌈log |B|
⌉ bits. If the
decoder can then recover these R remaining elements, this then implies the decoder has recovered
n
n
n
). The decoder proceeds as
) = Ω(R log m
− log |B|
S, and thus we must have |M | = Ω(log m
follows. Just as before, initially the decoder starts with T = ∅ and lets i be the output of Bob
on 1T and adds it to T . Then in iteration r, before proceeding to the next iteration, the decoder
randomly picks some elements from B and adds them into T , so that the number of elements
left to be uncovered is some fixed number nr . These extra elements being added to T should be
viewed as “random noise” to mask information about the random string s used by P, an idea very
loosely inspired by ideas in differential privacy. For intuition, as an example suppose the iteration
r = 1 succeeds in finding some i ∈ S. If the decoder were then to add i to T , as well as ≈ m/2
random elements from B to T , then the resulting T reveals only ≈ 1 bit of information about i
(and hence about s). This is as opposed to the log m bits T could have revealed if the masking
were not performed. Thus the next query in round r = 2, although correlated with s, has very
weak correlation after masking and we thus might hope for it to succeed. This intuition is captured
in the following lemma, which we prove in Section 3.1:
Lemma 1. Consider f : {0, 1}b × {0, 1}q → {0, 1} and X ∈ {0, 1}b uniformly random. If ∀y ∈
{0, 1}q , P(f (X, y) = 1) ≤ δ where 0 < δ < 1, then for any random variable Y supported on {0, 1}q ,
P(f (X, Y ) = 1) ≤
I(X; Y ) + H2 (δ)
,
log 1δ
(2)
where I(X; Y ) is the mutual information between X and Y , and H2 is the binary entropy function.
Fix some x ∈ {0, 1}n . One should imagine here that f (X, y) is 1 iff P fails when Alice has
input x and Bob has input y in a UR⊂ instance, and the public random string is X = s. Then the
lemma states that if y = Y is not arbitrary, but rather random (and correlated with X), then the
failure probability of the protocol is still bounded as long as the mutual information between X
and Y is bounded. It is also not hard to see that this lemma is sharp up to small additive terms.
Consider the case x, y ∈ [n], and f (x, y) = 1 iff x = y. Then if X is uniform, for all y we have
P(f (X, y) = 1) = 1/n. Now consider the case where Y is random and equal to X with probability
t/ log n and is uniform in [n] with probability 1 − t/ log n. Then in expectation Y reveals t bits of
X, so that I(X; Y ) = t. It is also not hard to see that P(f (X, Y ) = 1) ≈ t/ log n + 1/n.
In light of the strategy stated so far and Lemma 1, the path forward is clear: at each iteration
r, we should add enough random masking elements to T to keep the mutual information between
T and all previously added elements below, say, 21 log 1δ . Then we expect a constant fraction of
iterations to succeed. The encoder knows which iterations do not succeed since it shares public
randomness with the decoder (and can thus simulate it), so it can simply tell the decoder which
rounds are the failed ones, then explicitly include in M correct new elements of S for the decoder
to use in the place of Bob’s wrong output in those rounds. A calculation shows that if one adds
a (1 − 1/K) ≈ 2−1/K fraction of the remaining items in S to T after drawing one more support
element from Bob, the mutual information between the next query to Bob and the randomness
used by P will be O(K) (see Lemma 5). Thus we do this for K a sufficiently small constant times
log 1δ . We will then have nr ≈ (1 − 1/K)r m. Note that we cannot continue in this way once nr < K
(since the number of “random noise” elements we inject should at least be one). Thus
p we are forced
to stop after R = Θ(K log(m/K)) = Θ(log 1δ log logn 1 ) iterations. We then set m = n log(1/δ), so
that R→,pub
(UR⊂ ) = Ω(|R| log
δ
δ
n
m)
= Ω(min{n, log 1δ log2
7
n
})
log δ1
as desired.
The argument for lower bounding Rδ→,pub (UR⊂
k ) is a bit simpler, and in particular does not
need rely on Lemma 1. Both the idea and rigorous argument can be found in Section 3.2, but again
the idea is to use a protocol for this problem to encode appropriately sized subsets of [n].
As mentioned above, our lower bounds use protocols for UR⊂ and UR⊂
k to establish protocols
for encoding subsets of some fixed size m of [n]. These encoders always consist of some message
M Alice would have sent in a UR⊂ or UR⊂
k protocol, together with a random subset B ⊂ S
n
(using ⌈log2 |B|⌉ + ⌈log |B| ⌉ bits, to represent both |B| and the set B itself). Here |B| is a random
variable. These encoders are thus Las Vegas: the length of the encoding is a random variable,
but the encoder/decoder always succeed in compressing and recovering the subset. The final lower
bounds then come from the following simple lemma, which follows from the source coding theorem.
′
Lemma 2. Let s denote the number of bits used by the UR⊂ or UR⊂
k protocol, and let s denote the
expected number of bits to represent B. Then (1+s +s′ ) ≥ log(nm ). In particular, s ≥ log(nm )−s′ −1.
n
) log 1δ }). We
Section 3.1 provides our first proof that R→,pub
(UR⊂ ) = Ω(min{n, log2 ( log(1/δ)
δ
2
extend our results in Section 3.2 to UR⊂
k for k ≥ 1, proving a lower bound of Ω(k log (n/k))
communication even for constant failure probability.
2.2
Lower bound proof via reduction from AugIndexN
Our second lower bound proof for UR⊂ is via a randomized reduction from AugIndexN [MNSW98].
In this problem, Charlie receives z ∈ {0, 1}N and Diane receives j ∗ ∈ [N ] together with zj for j =
j ∗ + 1, . . . , N , and Diane must output zj ∗ . It is shown in [MNSW98] that R→,pub
(AugIndexN ) =
δ
n
).
Ω(N ) for any δ bounded away from 1/2. In our reduction, N = Θ(log(1/δ) log 2 log(1/δ)
⊂
For UR , we can also think of the problem as Alice being given S ⊆ [n] and Bob being given
T ( S, and Bob must output some element of S\T . In AugIndexN , Charlie views his input as
n
L = Θ(log log(1/δ)
) blocks of bits of nearly equal size, where the ith block represents a subset Si
of [ui ] in some collection Sui ,m of sets, for some carefully chosen universe sizes ui per block. Here
Sui ,m is a collection of subsets of [ui ] of size m of maximal size such any two sets in the collection
have intersection size strictly less than m/2. Furthermore, Diane’s index j ∗ is in some particular
block of bits corresponding to some set Si∗ , and Diane also knows Si for i > j.
Now we explain the reduction. We assume some protocol
P for UR⊂ , and we give a protocol
S
L
P ′ for AugIndexN . First, we define the universe A = i=1 ({i} × [ui ] × [100i ]), which has size n.
S
i
Charlie then defines S = L
i=1 ({i}× Si × [100 ]). Charlie and Diane use public randomness to define
a uniformly random permutation πS on [n]. Charlie can compute π(S). Also, since Diane knows
i
Si for i > i∗ , she can define T = L
i=i∗ +1 ({i} × Si × [100 ]) and compute π(T ). Then π(S) and
π(T ) are the inputs to Alice and Bob in the protocol P for UR⊂ . Charlie sends Diane the message
Alice would have sent Bob in P if her input had been π(S), and Diane simulates Bob to recover
an element in π(S)\π(T ). Importantly, Alice and Bob do not know anything about π at this point
other than that π(S) = S and π(T ) = T . Thus, the protocol P for UR⊂ , if it succeeds, outputs
an arbitrary element j ∈ π(S)\π(T ), which is a deterministic function of the labels of elements in
π(S) and π(T ) and the randomness R that Alice and Bob share, which is independent from the
randomness in π. Since π is still a uniformly random map conditioned on π(S) = S and π(t) = t for
each t ∈ T , and j ∈ π(S)\π(T ), it follows that π −1 (t) is a uniformly random element of S \ T . After
receiving π −1 (j), if (i, a, r) = π −1 (j), then Charlie and Diane reveal the pairs ((i, a, z), π((i, a, z)))
for each z ∈ [100i ] to Alice and Bob and Bob updates his set π(T ) to include π(i, a, z) for each
8
z ∈ [100i ]. One can show that at each step in this process, if Alice and Bob succeed in outputting an
arbitrary item j from π(S) \ π(T ), then this is a uniformly random item from π(S) \ π(T ). The fact
that this item is uniformly random is crucial for arguing the number of computation paths of the
protocol of Alice and Bob is o(1/δ) with good probability, over π, so that one can argue (see below)
that with good probability on every such computation path Alice and Bob succeed on that path,
over their randomness R. Although the idea of using a random permutation appeared in [JST11] to
show that any public coin UR protocol can be made into one in which a uniformly random element
of S\T is output, here we must use this idea adaptively, slowly revealing information about π and
arguing that this property is maintained for each of Bob’s successive queries.
Due to geometrically increasing repetitions of items for increasing i, a uniformly random element
in S\T is roughly 100 times more likely to correspond to an item in Si∗ than in Si for i < i∗ . Thus
if Diane simulates Bob to recover a random element in S\T , it is most likely to recover an element
∗
j of Si∗ . She can then tell Bob to include π(j) and its 100i redundant copies to π(T ) and iterate.
There are several obstacles to overcome to make this work. First, iterating means using P
adaptively, which was the same issue that arose in Section 2.1. Second, a constant fraction of the
time (1/100), we expect to obtain an element not in Si∗ , but rather from some Si for i < i∗ . If this
happened too often, then Diane would need to execute many queries to recover a sufficiently large
number of elements from Si∗ in order to solve AugIndexN . This would then require a union bound
over too many possible computation paths, which would not be possible as Alice likely would fail
on one of them (over the choice of R). However, since the random permutation argument above
ensures that at each step we receive a uniformly random item from the current set S \ T , if we
continue for m iterations, we can argue that with large probability, our sequence of inputs T over
the iterations with which Diane invokes Bob’s output are all likely to come from a family T of size
at most 2O(m) . Here we need to carefully construct this family to contain a smaller number of sets
from levels i for which i∗ − i is larger so that the overall number of sets is small. Given this, we
can union bound over all such T , for total failure probability δ|T | ≪ 1. Furthermore, we can also
argue that after m iterations, it is likely that we have recovered at least m/2 of the elements from
Si∗ , which is enough to uniquely identify Si∗ ∈ Sui ,m by the limited intersection property of Sui ,m .
3
Lower bounds via the adaptivity lemma
3.1
Communication Lower Bound for UR⊂
Consider a protocol P for UR⊂ with failure probability δ, operating in the one-way public coin
model. When Alice’s input is x and Bob’s is y, Alice sends Alice(x) to Bob, and Bob outputs
Bob(Alice(x), y), which with probability at least 1−δ is in support(x−y). As
pin Section 2,
mentioned
[n]
we use P as a subroutine in a scheme for encoding/decoding elements of m for m = ⌊ n log(1/δ)⌋.
We assume log 1δ ≤ n/64, since for larger n we have an Ω(n) lower bound.
3.1.1
Encoding/decoding scheme
We now describe our encoding/decoding scheme (ENC, DEC) for elements in [n]
m , which uses P in
a black-box way. The parameters shared by ENC and DEC are given in Algorithm 2.
As discussed in Section 2, on input S ∈ [n]
m , ENC computes M ← Alice(1S ) as part of its
output. Moreover, ENC also outputs a subset B ⊆ S computed as follows. Initially B = S and
S0 = S. ENC proceeds in R rounds. In round r ∈ [R], ENC computes sr ← Bob(M, 1S\Sr−1 ).
9
Let b denote a binary string of length R, where br records whether Bob succeeds in round r. ENC
also outputs b. If sr ∈ Sr−1 , i.e. Bob(M, 1S\Sr−1 ) succeeds, ENC sets br = 1 and removes sr from
B (since the decoder can recover sr from the UR⊂ -protocol, ENC does not need to include it in
B); otherwise
ENC sets br = 0. At the end of round r, ENC picks a uniformly random set Sr
Sr−1 \{sr }
. In particular, ENC uses its shared randomness with DEC to generate Sr in such a
in
nr
way that ENC, DEC agree on the sets Sr (DEC will actually iteratively construct Cr = S\Sr ). We
present ENC in Algorithm 3.
The decoding process is symmetric. Let C0 = ∅ and A = ∅. DEC proceeds in R rounds. On
round r ∈ [R], DEC obtains sr ∈ S\Cr−1 by invoking Bob(M, 1Cr−1 ). By construction of Cr−1
(to be described later), it is guaranteed that Sr−1 = S\Cr−1 . Therefore, DEC recovers exactly the
same sr as ENC. DEC initially assigns Cr ← Cr−1 . If br = 1, DEC adds sr to both A and Cr . At
the end of round r, DEC inserts many random items from B into Cr so that Cr = S\Sr . DEC can
achieve this because of the shared random permutation π when constructing Sr . In the end, DEC
outputs B ∪ A. We present DEC in Algorithm 4.
Algorithm 2 Variables shared by encoder ENC and decoder DEC.
q
1: m ← ⌊ n log 1δ ⌋
2:
3:
4:
5:
6:
7:
1
K ← ⌊ 16
log 1δ ⌋
R ← ⌊K log(m/4K)⌋
for r = 0, . . . , R do
r
nr ← ⌊m · 2− K ⌋
end for
π is a random permutation on [n]
Algorithm 3 Encoder ENC.
1: procedure ENC(S)
2:
M ← Alice(1S )
3:
A←∅
4:
S0 ← S
5:
for r = 1, . . . , R do
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
⊲ |Sr | = nr , and ∀r nr − nr+1 ≥ 2
⊲ Used to generate Sr and Cr
⊲ the set DEC recovers just from M
⊲ at end of round r, DEC still needs to recover Sr
?
sr ← Bob(M, 1S\Sr−1 )
⊲ sr ∈ Sr−1 found in round r
Sr ← Sr−1
if sr ∈ Sr−1 then
⊲ i.e. if sr is a valid sample
br ← 1
⊲ b ∈ {0, 1}R indicating which rounds succeed
A ← A ∪ {sr }
Sr ← Sr \{sr }
else
br ← 0
end if
Remove |Sr | − nr elements from Sr with smallest πa ’s among a ∈ Sr
⊲ now |Sr | = nr
end for
return (M , S\A, b)
end procedure
10
Algorithm 4 Decoder DEC.
1: procedure DEC(M , B, b)
⊲ M is Alice(1S )
⊲ b ∈ {0, 1}R indicates rounds in which Bob succeeds
⊲ B contains all elements of S that DEC doesn’t recover via M
2:
A←∅
⊲ the subset of S DEC recovers just from M
3:
C0 ← ∅
⊲ subset of S we have built up so far
4:
for r = 1, . . . , R do
⊲ each iteration tries to recover 1 element via M
5:
Cr ← Cr−1
6:
if br = 1 then
⊲ this means Bob succeeds in round r
7:
sr ← Bob(M, 1Cr−1 )
⊲ Invariant: Cr = S\Sr (Sr is defined in ENC)
8:
A ← A ∪ {sr }
9:
Cr ← Cr ∪ {sr }
10:
end if
11:
Insert m − nr − |Cr | items into Cr with smallest πa ’s among a ∈ B\Cr
⊲ Random masking “Differential Privacy” step. Still nr elements left to recover.
12:
end for
13:
return B ∪ A
14: end procedure
3.1.2
Analysis
We have two random objects in our encoding/decoding scheme: (1) the random source used by P,
denoted by X, and (2) the random permutation π. These are independent.
First, we can prove that DEC(ENC(S)) = S. That is, for any fixing of the randomness in X
and π, DEC will always decode S successfully. It is because ENC and DEC share X and π, so that
DEC essentially simulates ENC. We formally prove this by induction in Lemma 3.
Now our goal is to prove that by using the UR⊂ -protocol, the number of bits that ENC saves
n
in expectation over the naive ⌈log(nm )⌉-bit encoding is Ω(log 1δ log2 log(1/δ)
) bits. Intuitively, it is
1
n
equivalent to prove the number of elements that ENC saves is Ω(log δ log log(1/δ)
). We formalize
this in Lemma 4. Note that ENC also needs to output b (i.e., whether the Bob succeeds on R
rounds), which takes R bits. By our setting of parameters, we can afford the loss of R bits. Thus
n
it is sufficient to prove E |B| = |S| − Ω(log 1δ log log(1/δ)
).
PR
We have |S| − |B| = r=1 br . In Lemma 1, we prove the probability that Bob fails on round
r−1 )+1
, where I(X; Sr−1 ) is the mutual information between X and
r is upper bounded by I(X;S
log 1
δ
Sr−1 . Furthermore, we will show in Lemma 5 that I(X; Sr−1 ) is upper bounded by O(K). By our
n
).
setting of parameters, we have E br = Ω(1) and thus E(|S| − |B|) = Ω(R) = Ω(log 1δ log log(1/δ)
Lemma 3. DEC(ENC(S)) = S.
Proof. We claim that for r = 0, . . . , R, {Sr , Cr } is a partition of S (Sr is defined in Algorithm 3,
and Cr in Algorithm 4). We prove the claim by induction on r. Our base case is r = 0, for which
the claim holds since S0 = S, C0 = ∅.
Assume the claim holds for r − 1 (1 ≤ r ≤ R), and we consider round r. On round r, by
induction S\Sr−1 = Cr−1 , the index sr obtained by both ENC and DEC are the same. Initially
11
Sr = Sr−1 and Cr = Cr−1 , and so {Sr , Cr } is a partition of S. If sr is a valid sample (i.e. sr ∈ Sr−1 ),
then br = 1, and ENC removes sr from Sr and in the meanwhile DEC inserts sr into Cr , so that
{Sr , Cr } remains a partition of S. Next, ENC repeats removing the a from Sr with the smallest
πa value until |Sr | = nr . Symmetrically, DEC repeats inserting the a into Cr with the smallest
πa value among a ∈ B\Cr , until |Cr | = |S| − nr . In the end we have |Sr | + |Cr | = |S|, so ENC
and DEC execute repetition the same number of times. Moreover, we can prove that during the
same iteration of this repeated insertion, the element removed from Sr is exactly the same element
inserted to Cr . This is because in the beginning of a repetition {Sr , Cr } is a partition of S. We
have B\Cr ⊆ S\Cr = Sr . Let a∗ denote a ∈ Sr that minimizes πa . Then a∗ ∈ B\Cr ⊆ Sr (since a∗
will be removed from Sr , it has no chance to be included in S in ENC, so that B contains a∗ ), and
πa∗ is also the smallest among {πa : a ∈ B\Cr }. Thus both ENC and DEC will take a∗ (for ENC,
to remove from Sr , and for DEC, to insert into Cr ). Therefore, {Sr , Cr } remains a partition of S.
Given the fact that {Sr , Cr } is a partition of S, the sr are the same in ENC and DEC. Furthermore, A = {sr : br = 1, r = 1, . . . , R} are the same in ENC and DEC. We know A ⊆ S. Since ENC
outputs S\A, and DEC outputs (S\A) ∪ A, we have DEC(ENC(S)) = S.
n
Lemma
4. Let W ∈ N be a random variable with W ≤ m and E W ≤ m − d. Then E(log m
−
n
n
log W ) ≥ d log( m − 1).
Proof.
n
n
n!/(m!(n − m)!)
log
− log
= log
m
W
n!/(W !(n − W )!)
=
m−W
X
log
i=1
n−W −i+1
m−i+1
n−W
m
n−m
≥ (m − W ) · log
m
n
n
n
Taking expectation on both sides, we have E(log m
− log W
) ≥ d log( m
− 1).
≥ (m − W ) · log
Lemma 1 (restated). Consider f : {0, 1}b × {0, 1}q → {0, 1} and X ∈ {0, 1}b uniformly random.
If ∀y ∈ {0, 1}q , P(f (X, y) = 1) ≤ δ where 0 < δ < 1, then for any r.v. Y supported on {0, 1}q ,
P(f (X, Y ) = 1) ≤
I(X; Y ) + H2 (δ)
,
log 1δ
where I(X; Y ) is the mutual information between X and Y , and H2 is the binary entropy function.
Proof. It is equivalent to prove
I(X; Y ) ≥ E(f (X, Y )) · log
1
− H2 (δ).
δ
By definition of mutual entropy I(X; Y ) = H(X) − H(X|Y ), where H(X) = b and we must show
1
1
H(X|Y ) ≤ H2 (δ) + (1 − E(f (X, Y ))) · b + E(f (X, Y )) · (b − log ) = b + H2 (δ) − E(f (X, Y )) · log .
δ
δ
12
The upper bound for H(X|Y ) is obtained by considering the following one-way communication
problem: Alice knows both X and Y while Bob only knows Y , and Alice must send a single
message to Bob so that Bob can recover X. The expected message length in an optimal protocol
is exactly H(X|Y ). Thus, any protocol gives an upper bound for H(X|Y ), and we simply take
the following protocol: Alice prepends a 1 bit to her message iff f (X, Y ) = 1 (taking H2 (δ) bits
in expectation). Then if f (X, Y ) = 0, Alice sends X directly (taking b bits). Otherwise, when
f (X, Y ) = 1, Alice sends the index of X in {x|f (x, Y ) = 1} (taking log(δ2b ) = b − log 1δ bits).
Corollary 1. Let X denote the random source used by the UR⊂ -protocol with failure probability
)+H2 (δ)
at most δ. If S is a fixed set and T ⊂ S, P(Bob(Alice(1S ), 1T ) 6∈ S\T ) ≤ I(X;Tlog
.
1
δ
Lemma 5. I(X; Sr ) ≤ 6K, for r = 1, . . . , R.
Proof. Note that I(X; Sr ) = H(Sr ) − H(Sr |X). Since |Sr | = nr and Sr ⊆ S, H(Sr ) ≤ log nmr .
Here is the main idea to lower bound H(Sr |X): By definition of conditional entropy, H(Sr |X) =
P
x px · H(Sr |X = x). We fix an arbitrary x. If we can prove that for any T ⊆ S where |T | = nr ,
P(Sr = T |X = x) ≤ p, then by definition of entropy we have H(Sr |X = x) ≥ log 1p .
First we can prove for any fixed T ,
P(Sr = T |X = x) ≤
r
Y
i=1
ni−1 −nr −1
ni−1 −ni −1
ni−1 −1 .
ni−1 −ni −1
(3)
We have P(Sr = T |X = x) = Πri=1 P(T ⊆ Si |T ⊆ Si−1 ). On round i (1 ≤ i ≤ r), ENC removes ni−1 − ni elements (at least ni−1 − ni − 1 of which are chosen all at random) from Si−1
to obtain S
on the event that T ⊆ Si−1 , the probability that T ⊆ Si is at most
i . Conditioned
ni−1 −nr −1
ni−1 −1
ni−1 −ni −1 / ni−1 −ni −1 , where the equation achieves when si ∈ Si−1 \T , and ENC takes a uniformly random subset of Si−1 \{si } of size ni−1 − ni − 1, so that the subset does not intersect with
T.
Next we can prove
r
Y
i=1
ni−1 −nr −1
ni−1 −ni −1
ni−1 −1
ni−1 −ni −1
≤
26K
m .
(4)
nr
For notational simplicity, let nk denote n · (n − 1) . . . (n − k + 1). We have
!
ni−1 −nr −1
nr
nr
r
r
r
r
Y
Y
Y
Y
ni
ni
(ni−1 − nr − 1)!ni !
ni−1
ni−1 −ni −1
=
=
.
nr ·
ni−1 −1 =
(ni−1 − 1)!(ni − nr )!
(ni−1 − 1)nr
ni−1 ni−1 − nr
n
−n −1
i=1
i−1
i
i=1
(5)
i=1
i=1
By telescoping,
nr
nr
r
Y
ni
nr !(n0 − nr )!
nr
=
nr = nr =
n0 !
n
n0
i=1 i−1
13
1
n0
nr
=
1
m.
nr
(6)
Moreover,
r
Y
i=1
r
Y
ni−1
≤
ni−1 − nr
1−
i=1
1
m·2−r/K
m·2−(i−1)/K −1
By our setting of parameters
≤
r
Y
i=1
1
1−
m·2−r/K +1
=
r
Y
j=1
m·2−(i−1)/K
1
1−
2−j/K
−
2
r−j
K
.
(7)
m
R
r
2K
1
2K
≤
≤
.
m
m
4K
Therefore, for j ∈ {1, . . . , r},
1
j
−K
1−2
By Taylor series 21/K =
P∞
−
(ln 2)n
n=0 n!K n
r−j
2 K
≤
m
1
1 − (1 +
j
1
−K
4K )2
.
1
> 1 + lnK2 > 1 + 4K
, and thus
1
1
1−(1+ 4K
)2−j/K
≤
1
,
1−2(1−j)/K
1
K
for j = 2, . . . , r. For j = 1, we have
1 ≤ 2 .
1
1−(1+ 4K
)2− K
Q
1
5K . Therefore, the right hand side of (7) is upper
By Lemma 6, we have ∞
j=1 1−2−j/K ≤ 2
bounded by 26K . Together with
(6), we prove (4) holds.
Finally, let p = 26K / nmr , we have P(Sr = T |X = x) ≤ p and thus H(Sr |X = x) ≥ log p1 =
log nmr − 6K. Therefore, H(Sr |X) ≥ log nmr − 6K and so I(X; Sr ) = H(Sr ) − H(Sr |X) ≤ 6K.
Q
1
5K .
Lemma 6. Let K ∈ N and K ≥ 1. We have ∞
j=1 1−2−j/K ≤ 2
Proof. First, we bound the product of first 2K terms. Note that
Therefore,
1
1−2−x
≤
8
3x
for 0 < x ≤ 2.
2K
Y
1
K 2K
K 2K
2K
2K
≤
(8/3)
·
≤
(8/3)
·
= (4e/3)2K < 24K .
2K
−j/K
(2K)!
(2K/e)
1−2
j=1
(8)
Then, we bound the product of the rest terms
K
K
∞
∞
Y
Y
1
1
1
1
P
≤
= 2K .
≤
≤
−i
−j/K
−⌊j/K⌋
1 − 2−i
1− ∞
2
1
−
2
1
−
2
i=2
i=2
j=2K+1
j=2K+1
∞
Y
(9)
Multiplying two parts proves the lemma.
Theorem 2. R→,pub
(UR⊂ ) = Ω(log 1δ log2
δ
n
log(1/δ) ),
given that 64 ≤ log 1δ ≤
n
64 .
Proof. By Lemma 3, the success probability of protocol (ENC, DEC) is 1. By Lemma 2, P
we have
s ≥ log(nm ) − s′ − 1, where s′ = log n + R + E(log(n|B| )). The size of B is |B| = |S| − R
r=1 br .
By Corollary 1, conditioned on S, P(br = 0) ≤
I(X;Sr−1 )+1
.
log δ1
By Lemma 5, I(X; Sr−1 ) ≤ 6K
(Note that when r = 1, I(X; S0 ) = 0 ≤ 6K). Therefore, E(br ) ≥ 1 −
39
parameters (see Algorithm 2) we have E(br ) ≥ 64
. Therefore, E(|B|) ≤
n
1
n
39
n
n
). Furthermore,
log(m ) − E(log(|B| )) ≥ 64 R · log( m − 1) ≥ 2 R log( log(1/δ)
we obtain s ≥
R
3
log
n
log(1/δ)
− (log n + 1) = Ω(log 1δ log2
14
n
log(1/δ) ).
6K+1
. By the setting of
log δ1
39
|S| − 64
R. By Lemma 4,
n
1
6 R log log(1/δ) ≥ R. Thus
3.2
Communication Lower Bound for UR⊂
k
→,pub
2 n
(UR⊂
In this section, we prove the lower bound R1/2
k ) = Ω(min{n, k log k }). In fact, our lower
bound holds for any failure probability δ bounded away from 1. Let P denote a UR⊂
k -protocol
where Alice sends Alicek (x) to Bob, and Bob outputs Bob
(Alice
(x),
y).
We
consider
the
following
k
k
encoding/decoding scheme (ENCk , DECk ) for S ∈ [n]
.
ENC
computes
M
←
Alice
(1
)
as part
k
k S
m
n
of its message. In addition, ENCk includes B ⊆ S constructed as follows, spending ⌈log |B| ⌉ bits.
Initially B = S, and ENCk proceeds in R = Θ(log(n/k)) rounds. Let S0 = S ⊇ S1 ⊇ . . . ⊇ SR
where Sr is generated by sub-sampling each element in Sr−1 with probability 21 . In round r
(r = 1, . . . , R), ENCk tries to obtain k elements from Sr−1 by invoking Bobk (M, 1S\Sr−1 ), denoted
by Ak , and removes Ak ∩ (Sr−1 \Sr ) (whose expected size is k2 ) from B. Note that DECk is able to
recover the elements in Ak ∩ (Sr−1 \Sr ). For each round the failure probability of Bobk is at most
δ. Thus we have E(|S| − |B|) ≥ k2 · (1 − δ) · R = Ω(k log nk ). Furthermore, each element contains
Θ(log nk ) bits of information, thus yielding a lower bound of Ω(k log2 nk ) bits.
In this section we assume k ≤ n/210 , since for larger n we have an Ω(n) lower bound.
3.2.1
Encoding/decoding scheme
Algorithm 5 Variables Shared by Encoder ENCk and Decoder DECk .
√
1: m ← ⌊ nk⌋
⊲ Note that R ≥ 3 because k ≤ 2n10
2: R ← ⌊ 12 log(n/k) − 2⌋
3: T0 ← [n]
4: for r = 1, . . . , R do
5:
Tr ← ∅
⊲ We have Sr = S ∩ Tr
6:
For each a ∈ Tr−1 , Tr ← Tr ∪ {a} with probability 12
7: end for
Algorithm 6 Encoder ENCk .
1: procedure ENCk (S)
2:
M ← Alicek (1S )
3:
A←∅
4:
for r = 1, . . . , R do
5:
Ar ← Bobk (M, 1S\(S∩Tr−1 ) )
6:
if Ar ⊆ S ∩ Tr−1 then
⊲ i.e. if Ar is valid
7:
br ← 1
⊲ b is a binary string of length R, indicating if Bobk succeeds in round r
8:
A ← A ∪ (Ar ∩ (Tr−1 \Tr ))
9:
else
10:
br ← 0
11:
end if
12:
end for
13:
return (M , S\A, b)
14: end procedure
15
Algorithm 7 Decoder DECk .
1: procedure DECk (M , B, b)
2:
A←∅
3:
C0 ← ∅
4:
for r = 1, . . . , R do
5:
Cr ← Cr−1
6:
if br = 1 then
7:
Ar ← Bobk (M, 1Cr−1 )
8:
A ← A ∪ (Ar ∩ (Tr−1 \Tr ))
9:
Cr ← Cr ∪ (Ar ∩ (Tr−1 \Tr ))
10:
end if
11:
Cr ← Cr ∪ (B ∩ (Tr−1 \Tr ))
12:
end for
13:
return B ∪ A
14: end procedure
3.2.2
⊲ Invariant: Cr = S\(S ∩ Tr )
Analysis
2 n
Theorem 3. R→,pub
(UR⊂
k ) = Ω((1−δ)k log k ), given that 1 ≤ k ≤
δ
n
210
50 log n
and 0 < δ ≤ 1− k log
.
2
(n/k)
Proof. Let Sr = S ∩ Tr . Let SUCC denote the event that |S ∩ TR | = |SR | ≥ k. Note that
E |SR | = 21R m = 4k. By the Chernoff bound, P(SUCC) ≥ 12 . In the following, we argue conditioned
on SUCC. Namely, in each round r, there are at least k items in Sr .
Similar to Lemma 3, we can prove the protocol (ENCk , DECk ) always succeeds. By Lemma 2,
we have s ≥ log(nm ) − s′ − 2, where s′ = log n + R + E log(n|B| ). The size of B is |B| = |S| −
PR
r=1 (br · |Ar ∩ (Sr−1 \Sr )|). The randomness used by P is independent from S\Sr−1 for every
r ∈ [R]. Therefore, E br ≥ 1 − δ, and br is independent from |Ar ∩ (Sr−1 \Sr )|. We have E |Ar ∩
(Sr−1 \Sr )| = k2 , and thus E(|S| − |B|) ≥ (1−δ)kR
. By Lemma 4, log(nm ) − E log(n|B| ) ≥ (1−δ)kR
·
2
2
n
log( m
− 1) ≥ (1−δ)kR
log( nk ). Moreover, R ≤ log n and log n ≤
5
n
s = Ω((1 − δ)kR log k ) = Ω((1 − δ)k log2 nk ).
4
(1−δ)kR
12
log nk . Thus we have
Lower bounds proofs via augmented indexing
2
Here we show another route to proving R→,pub
(UR⊂
k ) = Ω(min{n, t log (n/t)} via reduction from
δ
augmented indexing. We again separately prove lower bounds for R→,pub
(UR⊂ ) and R→,pub
(UR⊂
1
k ).
δ
5
Both proofs make use of the following standard lemma. The proof can be found in the appendix
(see Section A.2).
Lemma 7. For any integers u ≥ 1 and 1 ≤ m ≤ u/(4e), there exists a collection Su,m ⊂ [u]
m with
log |Su,m | = Θ(m log(u/m)) such that for all S 6= S ′ ∈ Su,m , |S ∩ S ′ | < m/2.
Both our lower bounds in Sections 4.1 and 4.2 reduce from augmented indexing (henceforth
AugIndex) to either UR⊂ with low failure probability, or UR⊂
k with constant failure probability,
in the public coin one-way model of communication. We remind the reader of the setup for the
AugIndexN problem. There are two players, Charlie and Diane. Charlie receives z ∈ {0, 1}N and
16
Diane receives j ∗ ∈ [N ] together with zj ∗ +1 , . . . , zN . Charlie must send a single message to Diane
such that Diane can then output zj ∗ . The following theorem is known.
→,pub
(AugIndexN ) = Θ(N ).
Theorem 4. [MNSW98] R1/3
We show that if there is an s-bit communication protocol P for UR⊂ on n-bit vectors with
failure probability δ (or for URk with constant failure probability), that implies the existence of
an s-bit protocol P ′ for AugIndexN for some N = Θ(log 1δ log2 logn 1 ) (or N = Θ(k log2 (n/k)) for
URk ). The lower bound on s then follows from Theorem 4.
4.1
δ
Communication Lower Bound for UR⊂
Set t = log 1δ . In this section we assume t < n/(4e) and show R→,pub
(UR⊂ ) = Ω(t log2 (n/t)). This
δ
implies a lower bound of Ω(min{n, t log2 (n/t)}) for all δ > 0 bounded away from 1.
As mentioned, we assume we have an s-bit protocol P for UR⊂ with failure probability δ, with
players Alice and Bob.We use P to give an s-bit protocol P ′ for AugIndexN , which has players
Charlie and Diane, for N = Θ(t log2 (n/t)).
The protocol P ′ operates as follows. Without loss of generality we may assume that, using the
notation of Lemma 7, |Su,m | is a power of 2 for u, m as in the lemma statement. This is accomplished
by simply rounding |Su,m | down to the nearest power of 2 by removing elements arbitrarily. Also,
define L = c log(n/t) for some sufficiently small constant c ∈ (0, 1) to be determined later. Now,
Charlie partitions the bits of his input z ∈ {0, 1}N into L consecutive sequences of bits such that the
ith chunk of bits for each i ∈ [L] can be viewed as specifying an element Si ∈ Sui ,m for ui = 100ni ·L
and m = ct. Lemma 7 gives log |Sui ,m | = Θ(m log(ui /m)), which is Θ(t log(n/t)) for c < 1/14.
Thus N = Θ(L · t log(n/t)) = Θ(t log2 (n/t)). Given these sets S1 , . . . , SL , we now discuss how
Charlie generates a vector x ∈ {0, 1}n . Charlie then simulates Alice on x to generate the message
Alice would have send to Bob in protocol P, then sends that same message to Diane.
To generate x ∈ {0, 1}n , assume Charlie and Diane have sampled a bijection from
A=
L
[
({i} × [ui ] × [100i ])
(10)
i=1
to [n] uniformly at random. We denote this bijection by π. This is possible since |A| = n. Then
Charlie defines x to be the indicator vector 1π(S) , where
S=
L
[
i=1
({i} × Si × [100i ]),
then sends a message M to Diane, equal to Alice’s message with input 1π(S) . This completes the
description of Charlie’s behavior in the protocol P ′ .
We describe how Diane uses M to solve AugIndexN . Diane’s input j ∗ ∈ [N ] lies in some
chunk i∗ ∈ [L]. We now show how Diane can use P to recover Si∗ with probability 2/3 (and thus
in particular recover zj ∗ ). Since Diane knows zj for j > j ∗ , she knows Si for i > i∗ . She can then
execute the following algorithm.
17
Algorithm 8 Behavior of Diane in P ′ for UR⊂ .
1: procedure Diane(M )
S
i
2:
T ← L
i=i∗ +1 ({i} × Si × [100 ])
3:
Ti∗ ← ∅
4:
while |Ti∗ | < m
2 do
5:
(i, a, r) ← π −1 (Bob(M, 1π(T ) ))
6:
T ← T ∪ ((i, a) × [100i ])
7:
if i = i∗ then
8:
Ti∗ ← Ti∗ ∪ {a}
9:
end if
10:
end while
11:
if there exists S ∈ Sui∗ ,m with Ti∗ ⊂ S then
12:
return the unique such S
13:
else
14:
return Fail
15:
end if
16: end procedure
In Algorithm 8 Diane is building up a subset Ti∗ of Si∗ . Once |Ti∗ | ≥ |Si∗ |/2 = m/2, Diane
can uniquely recover Si∗ by the limited intersection property of Sui ,m guaranteed by Lemma 7.
Until then, she uses P to recover elements of S\T , which, as we now show, are chosen uniformly
at random from S \ T .
Claim 4. For every protocol for Alice and Bob that uses shared randomness with Bob’s behaviour
given by Bob(·), for every choice of shared random string R of Alice and Bob, for every S, T ⊂ S,
the following conditions hold. If π is a uniformly random permutation, the success or failure of
Bob(M, 1π(T ) ) is determined by {π(j)}j∈T and the image π(S \ T ) of S \ T under π. Conditioned on a choice of R, {π(j)}j∈T and π(S \ T ) such that Bob(M, 1π(T ) ) succeeds, one has that
π −1 (Bob(M, 1π(T ) )) is a uniformly random element of S \ T .
Proof. The first claim follows by noting that the message M that Alice sends to Bob is solely a
function of R and π(S). The behaviour of Bob is determined by M and π(T ) (and the latter is
determined by {π(j)}j∈T ).
Now condition on the values of R, {π(j)}j∈T and π(S \ T ) such that Bob(M, 1π(T ) ) succeeds,
and let j ∗ ∈ [n] denote the output. Note that by our conditioning j ∗ is a fixed quantity. The
only randomness left is the exact mapping of S \ T to π(S \ T ). This mapping is independent of
{π(j)}j∈T and π(S \ T ) and uniformly random, so π −1 (j ∗ ) is a uniformly random element of S \ T ,
as required.
g
Fix any protocol Bob(M,
1π(T ) ) (not necessarily the one that Charlie and Diane use; see analysis
e
of the idealized process P below). Now fix T together with values of R, {π(j)}j∈T and π(S \ T )
g
such that Bob(M,
1π(T ) ) succeeds.
Elements in Sj , j < i∗ , are unlikely to be recovered. Given Claim 4, since the elements of
g
Sj appear with frequency 100j in S\T , they are less likely to be returned by π −1 (Bob(M,
1π(T ) ))
18
when j is small. More precisely, as long as |Si∗ ∩ Ti∗ | ≥ m/2, for any 1 ≤ j < i∗
m · 100j
g
P(i = j|(R, {π(j)}j∈T , π(S \ T )) s.t. Bob(M,
1π(T ) ) succeeds) ≤ m
i∗
2 · 100
≤ 2 · 100−(i
≤ 50−(i
∗ −j)
∗ −j)
(11)
.
Here again the probability is over the choice of π|S\T : (S \ T ) → π(S \ T ) (recall that we condition
on the image π(S \ T ) under π, but not on the actual mapping).
We now define the set T of typical intermediate sets, which plays a crucial role in our
analysis. Let Qi for i ∈ [L] denote {i} × Si × [100i ]. Let T be the collection of all T ⊂ S such that
∗
(1) Qi ⊂ T for all i > i∗ , and (2) for each i < i∗ , |T ∩ Qi | ≤ 100i · m/4i −i . The following claim
will be useful:
Claim 5. For the set T defined above one has |T | = 2O(m) .
Proof.
|T | ≤ 2m ·
∗ −1
iY
i=1
≤2 ·
m
≤2 ·
X m
(the 2m term comes from Si∗ )
r
m
∗
4i −i
r=0
m
4i∗ −i
m
Y m +
i∗ −1
m
i=1
∗ −1
iY
i=1
2i∗ −i
i∗ −i
(2e · 4
P∞
≤ 2O(m) · 2m·O(
j=1
)
m
∗
4i −i
n
(using
≤ (en/k)k )
k
j4−j )
≤ 2O(m)
We will show that for most choices of π and shared random string R Algorithm 8 (a) never
leaves the set T and (b) successfully terminates. Note that Algorithm 8 is a random process whose
sample space is the product of the set of all possible permutations π and shared random strings R.
As before, we denote this process by P ′ . It is useful for analysis purposes to define another process
e which is an idealized version of P ′ . In this process instead of running Bob(M, 1π(T ) ) Alice runs
P,
g
Bob(M,
1π(T ) ), which is guaranteed to output an element of π(S \ T ) for every choice of T ⊂ S,
shared random string R, {π(j)}j∈T , and π(S \ T ). The proof proceeds in three steps.
e succeeds in recovering Ti∗ and never leaves T with high
Step 1: proving that P
probability. Choose π uniformly at random. By (11), as long as |Si∗ ∩ Ti∗ | ≥ m/2, the expected
g from Si for i < i∗ in the first m iterations is at most m/50i∗ −i .
number of items recovered by Bob
∗
∗
Thus the probability of recovering more than m/4i −i items from Si is at most (1/12)i −i by
g is assumed to
Markov’s inequality. Note that the probability is over the choice of π only, as Bob
e Thus
succeed with probability 1 by definition of P.
P(Pe leaves T ) ≤
∗ −1
iX
i=1
19
(1/12)i
∗ −i
< 1/10.
P
i∗ −i < m/2 items
In particular
this
means
that
with
probability
at
least
1
−
1/10
at
most
i<i∗ m/4
S
from i<i∗ Si are recovered in the first m (or fewer, if the algorithm terminates earlier) iterations.
This also implies that with probability at least 1 − 1/10 if the algorithm proceeds for the entire
m iterations, it recovers at least m/2 elements of Ti∗ and hence terminates. We thus get that Pe
succeeds at least with probability 1 − 1/10.
Step 2: coupling Pe to P ′ on most of the probability space. For every T ⊂ S and every π
let ET (π) be the probabilistic event (over the choice of Bob’s random string R) that Bob(M, 1π(T ) )
succeeds in returning an element in π(S\T ). Note that ET (π) is a subset of the probability space
of shared random strings R, and depends on π. We let
ET (π) := ∧T ∈T ET (π)
to simplify notation. Using Claim 5 and the union bound we have for every π
P(¬(ET (π))) ≤ δ · |T | ≤ 1/20
R
as long as for m = c log(1/δ) for c a sufficiently small constant.
g
Now recall that Bob(M,
1π(T ) ) is an idealized protocol, which is guaranteed to output an element
of π(S \ T ) for every choice of T ⊂ S, shared random string R, {π(j)}j∈T , and π(S \ T ). We have
just shown that for every π the event ET (π) occurs with probability at least 1 − 1/20 over the
g
choice of R. Now define Bob(M,
1π(T ) ) as equal to Bob(M, 1π(T ) ) for all T ∈ T (the typical set of
g
intermediate sets) and (π, R) such that R ∈ ET (π), and extend Bob(M,
1π(T ) ) to return an arbitrary
g defined in this way
element of π(S \ T ) for remaining tuples (T, R, π(T ), π(S \ T )). Note that Bob
is a deterministic function once T , R, π(T ) and π(S \ T ) are fixed.
Note that with probability at least 1 − 1/20 over the choice of π and R one has Bob(M, 1π(T ) ) =
g
Bob(M,
1π(T ) ) for all T ∈ T , as required.
Step 3: arguing that P ′ succeeds with high probability. Choose (π, R) uniformly at
random. By Step 2 we have that with probability at least 1−1/20 over this choice Bob(M, 1π(T ) ) =
g
Bob(M,
1π(T ) ) for all T ∈ T . At the same time we have by Step 1 that with probability at
e succeeds in recovering Ti∗ and never
least 1 − 1/10 over the choice of π the idealized process P
leaves T . Putting the two bounds together, we get that P ′ succeeds with probability at least
1 − 1/20 − 1/10 > 2/3, showing the following theorem.
Theorem 5. For any 0 < δ < 1/2 and integer n ≥ 1 with log
2 n
1
R→,pub
1/3 (AugIndexN ) for N = Θ(log δ log log 1 ).
1
δ
< n/(4e), R→,pub
(UR⊂ ) ≥
δ
δ
Corollary 2. For any 0 < δ < 1/2 and integer n ≥ 1, R→,pub
(UR⊂ ) = Ω(min{n, log 1δ log2
δ
4.2
n
}).
log 1δ
Communication Lower Bound for UR⊂
k
The idea for lower bounding R→,pub
(UR⊂
1
k ) is as in Section 4.1, but slightly simpler. That is because
5
for the protocol P ′ for AugIndexN , Diane will not make adaptive queries to Bob in the protocol P
for UR⊂
k . Rather, she will only make one query using Bob and will be able to decide AugIndexN
with good probability from that single query. We make use of the following lemma from [JST11],
whose proof is similar to our analysis in Section 4.1.
20
Lemma 8. [JST11] Any public coin protocol for UR⊂ can be turned into one that outputs every
index i ∈ [n] with xi 6= yi with the same probability. The number of bits sent, failure probability,
and number of rounds do not change. Similarly, any UR⊂
k protocol can be turned into one in which
all subsets of [n] of size min{k, kx − yk0 } on which x, y differ are equally likely to be output.
Henceforth we assume P outputs random differing indices, which is without loss of generality
by Lemma 8.
Again Charlie receives z ∈ {0, 1}N and Diane receives j ∗ and zj ∗ +1 , . . . , zN and they want to
solve AugIndexN . Charlie views his input as consisting of L blocks for L = c log(n/k) for a
sufficiently small constant c ∈ (0, 1), and the ith block for i ∈ [L] specifies a set Si ∈ Sui ,m for
m = ck and ui = n/(100i L). As before, for c sufficiently small we have N = Θ(L · k log(n/k)) =
Θ(k log2 (n/k)). The bijection A and set S are defined exactly as in Section 4.1, and Charlie
simulates Alice to send the message M to Diane that Alice would have sent to Bob on input 1S .
Again, Diane knows Si for i > i∗ , where j ∗ lies in the i∗ th block of bits. Diane’s algorithm to
produce her output is then described in Algorithm 9.
Algorithm 9 Behavior of Diane in P ′ for UR⊂
k.
1: procedure Diane(M )
S
i
2:
T ← L
i=i∗ +1 ({i} × Si × [100 ])
3:
Ti∗ ← ∅
4:
B ← Bob(M, 1T )
5:
for (i, a, r) ∈ B do
6:
if i = i∗ and a ∈
/ T then
7:
Ti∗ ← Ti∗ ∪ {a}
8:
end if
9:
end for
10:
if |Ti∗ | < m
2 then
11:
return Fail
12:
else
13:
return the unique S ∈ Sui∗ ,m with Ti∗ ⊂ S
14:
end if
15: end procedure
Recall Bob, when he succeeds, returns min{k, |S\T |} = k uniformly random elements from
S\T . Meanwhile, Si∗ only has m = ck elements for some small constant c. As in Section 4.1,
almost all of the support of S\T comes from items in block i∗ , and hence we expect almost all our
k samples to come from (and be uniform in) items corresponding to elements of Si∗ .
We now provide a formal analysis. Henceforth we condition on Bob succeeding, which happens
∗
with probability 4/5. The number of elements in S\T corresponding to an element of Si∗ is 100i m,
whereas the number of elements corresponding to an element of Si for i < i∗ is
m·
∗ −1
iX
i=1
100i =
∗
m
∗
m
· (100i − 1) <
· 100i
99
99
Thus, we expect at most k/99 elements in B to correspond to elements in Si for i 6= i∗ , and the
probability that we have at least k/9 such elements in B is less than 1/10 by Markov’s inequality.
21
We henceforth condition on having less than k/9 such elements in B. Now we know B contains
at least 8k/9 elements corresponding to Si∗ , chosen uniformly from Si∗ × [100i ]. For any given
element a ∈ Si∗ , the probability that none of the elements in B from Si∗ correspond to a is
8
(1 − 1/m) 9 k ≤ e−(8/9)k/m < 1/30 for c sufficiently small (where m = ck). Thus the expected
number of a ∈ Si∗ not covered by B is less than m/30. Thus the probability that fewer than m/2
elements are covered by B is a most 1/15 by Markov’s inequality (and otherwise, Diane succeeds).
Thus, the probability that Diane succeeds is at least 4/5 · 9/10 · 14/15 > 2/3. We have thus shown
the following theorem.
→,pub
(UR⊂
(AugIndexN ) for N =
Theorem 6. For any integers 1 ≤ k ≤ n, R→,pub
1
k ) ≥ R1
5
Θ(k log2 (n/k)).
3
2
Corollary 3. For any integers 1 ≤ k ≤ n, R→,pub
(UR⊂
1
k ) = Ω(min{n, k log (n/k)}).
5
Remark 1. One may wish to understand R→,pub
(UR⊂
δ
k ) for δ near 1 (or at least, larger than
1/2). Such a lower bound is given in Theorem 3. The proof given above as written would yield
no lower bound in this regime for δ since AugIndex is in fact easy when the failure probability is
allowed to be least 1/2 (Charlie can send no message at all, and Diane can simply guess zj ∗ via a
coin flip). One can however get a handle on R→,pub
(UR⊂
δ
k ) by instead directly reducing from the
following variant of augmented indexing: Charlie receives D ∈ Su1 ,m ×· · ·×SuL ,m and Diane receives
j ∗ ∈ [L] and Dj ∗ +1 , . . . , DL and must output Dj ∗ , where the ui are as above. One can show that
unless Charlie sends almost his entire input, Diane cannot have success probability significantly
better than random guessing (which has success probability O(maxi∈L 1/|Sui ,m |)). The proof is
nearly identical to the analysis of augmented indexing over large domains [EJS10, JW13]. Indeed,
the problem is even almost identical, except that here we consider Charlie receiving a vector whose
entries come from different alphabet sizes (since the |Sui ,m | are different), whereas in [EJS10, JW13]
all the entries come from the same alphabet.
Acknowledgments
Initially the authors were focused on proving optimal lower bounds for samplers, but we thank
Vasileios Nakos for pointing out that our UR⊂ lower bound immediately implies a tight lower
bound for finding a duplicate in data streams as well. Also, initially our proof of Lemma 1 incurred
an additive 1 in the numerator of the right hand side of (2). This is clearly suboptimal for small
I(X; Y ) (for example, consider I(X; Y ) = 0, in which case the right hand side should be δ and not
1/ log(1/δ))). We thank T.S. Jayram for pointing out that a slight modification of our proof could
actually replace the additive 1 with the binary entropy function (and also for showing us a different
proof of this lemma, which resembles the standard proof of Fano’s inequality).
References
[AGM12a]
Kook Jin Ahn, Sudipto Guha, and Andrew McGregor. Analyzing graph structure via
linear measurements. In Proceedings of the 23rd ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 459–467, 2012.
22
[AGM12b]
Kook Jin Ahn, Sudipto Guha, and Andrew McGregor. Graph sketches: sparsification, spanners, and subgraphs. In Proceedings of the 31st ACM SIGMOD-SIGACTSIGART Symposium on Principles of Database Systems (PODS), pages 5–14, 2012.
[AGM13]
Kook Jin Ahn, Sudipto Guha, and Andrew McGregor. Spectral sparsification in
dynamic graph streams. In Proceedings of the 16th International Workshop on Approximation Algorithms for Combinatorial Optimization Problems (APPROX), pages
1–10, 2013.
[AKL17]
Sepehr Assadi, Sanjeev Khanna, and Yang Li. On estimating maximum matching
size in graph streams. In Proceedings of the 28th Annual ACM-SIAM Symposium on
Discrete Algorithms (SODA), pages 1723–1742, 2017.
[AKLY16]
Sepehr Assadi, Sanjeev Khanna, Yang Li, and Grigory Yaroslavtsev. Maximum
matchings in dynamic graph streams and the simultaneous communication model.
In Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 1345–1364, 2016.
[AKO11]
Alexandr Andoni, Robert Krauthgamer, and Krzysztof Onak. Streaming algorithms
via precision sampling. In Proceedings of the 52nd Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 363–372, 2011.
[BHNT15]
Sayan Bhattacharya, Monika Henzinger, Danupon Nanongkai, and Charalampos E.
Tsourakakis. Space- and time-efficient algorithm for maintaining dense subgraphs on
one-pass dynamic streams. In Proceedings of the 47th Annual ACM on Symposium on
Theory of Computing (STOC), pages 173–182, 2015.
[BS15]
Marc Bury and Chris Schwiegelshohn. Sublinear estimation of weighted matchings in
dynamic data streams. In Proceedings of the 23rd Annual European Symposium on
Algorithms (ESA), pages 263–274, 2015.
[CCE+ 16]
Rajesh Chitnis, Graham Cormode, Hossein Esfandiari, MohammadTaghi Hajiaghayi,
Andrew McGregor, Morteza Monemizadeh, and Sofya Vorotnikova. Kernelization
via sampling with applications to finding matchings and related problems in dynamic
graph streams. In Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 1326–1344, 2016.
[CCHM15]
Rajesh Hemant Chitnis, Graham Cormode, Mohammad Taghi Hajiaghayi, and
Morteza Monemizadeh. Parameterized streaming: Maximal matching and vertex
cover. In Proceedings of the 26th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1234–1251, 2015.
[CF14]
Graham Cormode and Donatella Firmani. A unifying framework for ℓ0 -sampling
algorithms. Distributed and Parallel Databases, 32(3):315–335, 2014. Preliminary
version in ALENEX 2013.
[CK04]
Don Coppersmith and Ravi Kumar. An improved data stream algorithm for frequency
moments. In Proceedings of the 15th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 151–156, 2004.
23
[CMR05]
Graham Cormode, S. Muthukrishnan, and Irina Rozenbaum. Summarizing and mining inverse distributions on data streams via dynamic inverse sampling. In Proceedings
of the 31st International Conference on Very Large Data Bases (VLDB), pages 25–36,
2005.
[DM16]
Irit Dinur and Or Meir. Toward the KRW composition conjecture: Cubic formula
lower bounds via communication complexity. In Proceedings of the 31st Conference
on Computational Complexity (CCC), pages 3:1–3:51, 2016.
[EHW16]
Hossein Esfandiari, MohammadTaghi Hajiaghayi, and David P. Woodruff. Brief announcement: Applications of uniform sampling: Densest subgraph and beyond. In
Proceedings of the 28th ACM Symposium on Parallelism in Algorithms and Architectures (SPAA), pages 397–399, 2016.
[EIRS91]
Jack Edmonds, Russell Impagliazzo, Steven Rudich, and Jiri Sgall. Communication
complexity towards lower bounds on circuit depth. In Proceedings of the 32nd Annual
IEEE Symposium on the Foundations of Computer Science (FOCS), pages 249–257,
1991.
[EJS10]
Funda Ergün, Hossein Jowhari, and Mert Sağlam. Periodicity in streams. In Proceedings of the 14th International Workshop on Randomization and Approximation
Techniques in Computer Science (RANDOM), pages 545–559, 2010.
[FIS08]
Gereon Frahling, Piotr Indyk, and Christian Sohler. Sampling in dynamic data
streams and applications. Int. J. Comput. Geometry Appl., 18(1/2):3–28, 2008. Preliminary version in SOCG 2005.
[FT16]
Martin Farach-Colton and Meng-Tsung Tsai. Tight approximations of degeneracy in
large graphs. In Proceedings of the 12th Latin American Symposium on Theoretical
Informatics (LATIN), pages 429–440, 2016.
[GKKT15]
David Gibb, Bruce M. Kapron, Valerie King, and Nolan Thorn. Dynamic graph
connectivity with improved worst case update time and sublinear space. CoRR,
abs/1509.06464, 2015.
[GMT15]
Sudipto Guha, Andrew McGregor, and David Tench. Vertex and hyperedge connectivity in dynamic graph streams. In Proceedings of the 34th ACM Symposium on
Principles of Database Systems (PODS), pages 241–247, 2015.
[GMWW14] Dmitry Gavinsky, Or Meir, Omri Weinstein, and Avi Wigderson. Toward better
formula lower bounds: an information complexity approach to the KRW composition conjecture. In Proceedings of the 46th Annual ACM Symposium on Theory of
Computing (STOC), pages 213–222, 2014.
[GR09]
Parikshit Gopalan and Jaikumar Radhakrishnan. Finding duplicates in a data stream.
In Proceedings of the 20th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 402–411, 2009.
24
[HPP+ 15]
James W. Hegeman, Gopal Pandurangan, Sriram V. Pemmaraju, Vivek B. Sardeshmukh, and Michele Scquizzato. Toward optimal bounds in the congested clique: Graph
connectivity and MST. In Proceedings of the 34th Annual ACM Symposium on Principles of Distributed Computing (PODC), pages 91–100, 2015.
[HW90]
Johan Håstad and Avi Wigderson. Composition of the universal relation. In Proceedings of a DIMACS Workshop on Advances In Computational Complexity Theory,
pages 119–134, 1990.
[JST11]
Hossein Jowhari, Mert Sağlam, and Gábor Tardos. Tight bounds for Lp samplers,
finding duplicates in streams, and related problems. In Proceedings of the 30th ACM
SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems (PODS),
pages 49–58. ACM, 2011.
[JW13]
T. S. Jayram and David P. Woodruff. Optimal bounds for Johnson-Lindenstrauss
transforms and streaming problems with subconstant error. ACM Trans. Algorithms,
9(3):26:1–26:17, 2013.
[KKM13]
Bruce M. Kapron, Valerie King, and Ben Mountjoy. Dynamic graph connectivity
in polylogarithmic worst case time. In Proceedings of the 24th Annual ACM-SIAM
Symposium on Discrete Algorithms (SODA), pages 1131–1142, 2013.
[KLM+ 14]
Michael Kapralov, Yin Tat Lee, Cameron Musco, Christopher Musco, and Aaron
Sidford. Single pass spectral sparsification in dynamic streams. In Proceedings of the
55th IEEE Annual Symposium on Foundations of Computer Science (FOCS), pages
561–570, 2014.
[KNW10]
Daniel M. Kane, Jelani Nelson, and David P. Woodruff. An optimal algorithm for
the distinct elements problem. In Proceedings of the 29th ACM SIGMOD-SIGACTSIGART Symposium on Principles of Database Systems (PODS), pages 41–52, 2010.
[Kon15]
Christian Konrad. Maximum matching in turnstile streams. In Proceedings of the
23rd Annual European Symposium on Algorithms (ESA), pages 840–852, 2015.
[KRW95]
Mauricio Karchmer, Ran Raz, and Avi Wigderson. Super-logarithmic depth lower
bounds via the direct sum in communication complexity. Computational Complexity,
5(3-4):191–204, 1995.
[KW90]
Mauricio Karchmer and Avi Wigderson. Monotone circuits for connectivity require
super-logarithmic depth. SIAM J. Discrete Math., 3(2):255–265, 1990.
[McG14]
Andrew McGregor. Graph stream algorithms: a survey. SIGMOD Record, 43(1):9–20,
2014.
[MNSW98]
Peter Bro Miltersen, Noam Nisan, Shmuel Safra, and Avi Wigderson. On data structures and asymmetric communication complexity. J. Comput. Syst. Sci., 57(1):37–49,
1998.
[MTVV15]
Andrew McGregor, David Tench, Sofya Vorotnikova, and Hoa T. Vu. Densest subgraph in dynamic graph streams. In Proceedings of the 40th International Symposium
on Mathematical Foundations of Computer Science (MFCS), pages 472–482, 2015.
25
[Mut05]
S. Muthukrishnan. Data Streams: Algorithms and Applications. Foundations and
Trends in Theoretical Computer Science, 1(2):117–236, 2005.
[MW10]
Morteza Monemizadeh and David P. Woodruff. 1-pass relative-error lp -sampling with
applications. In Proceedings of the 21st Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 1143–1160, 2010.
[NPW17]
Jelani Nelson, Jakub Pachocki, and Zhengyu Wang. Optimal lower bounds for universal relation, samplers, and finding duplicates. CoRR, abs/1703.08139, March 2017.
[PRS16]
Gopal Pandurangan, Peter Robinson, and Michele Scquizzato. Fast distributed algorithms for connectivity and MST in large graphs. In Proceedings of the 28th ACM
Symposium on Parallelism in Algorithms and Architectures (SPAA), pages 429–438,
2016.
[Tar07]
Jun Tarui. Finding a duplicate and a missing item in a stream. In Proceedings of
the 4th Annual Conference on Theory and Applications of Models of Computation
(TAMC), pages 128–135, 2007.
[TZ97]
Gábor Tardos and Uri Zwick. The communication complexity of the universal relation.
In Proceedings of the 12th Annual IEEE Conference on Computational Complexity
(CCC), pages 247–259, 1997.
[Wan15]
Zhengyu Wang. An improved randomized data structure for dynamic graph connectivity. CoRR, abs/1510.04590, 2015.
A
A.1
Appendix
A tight upper bound for R→,pub
(URk )
δ
In [JST11, Proposition 1] it is shown that R→,pub
(URk ) = O(min{n, t log2 n}) for t = max{k, log(1/δ)}.
δ
Here we show that a minor modification of their protocol in fact shows the correct complexity
R→,pub
(URk ) = O(min{n, t log2 (n/t)}), which given our new lower bound, is optimal up to a
δ
constant factor for the full range of n, k, δ as long as δ is bounded away from 1.
Recall Alice and Bob receive x, y ∈ {0, 1}n , respectively, and share a public random string.
Alice must send a single message M to Bob, from which Bob must recover min{k, kx − yk0 } indices
i ∈ [n] for which xi 6= yi . Bob is allowed to fail with probability δ. The fact that R→,pub
(URk ) ≤ n
δ
is obvious: Alice can simply send the message M = x, and Bob can then succeed with failure
(URk ) ≤ k log2 (n/k) for some constant c > 0, which
probability 0. We thus now show R→,pub
e−ck
completes the proof of the upper bound. We assume k ≤ n/2 (otherwise, Alice sends x explicitly).
As mentioned, the protocol we describe is nearly identical to one in [JST11] (see also [CF14]).
We will describe the new protocol here, then point out the two minor modifications that improve
the O(k log2 n) bound to O(k log2 (n/k)) in Remark 2. We first need the following lemma.
Lemma 9. Let Fq be a finite field and n > 1 an integer. Then for any 1 ≤ k ≤ n2 , there exists
for m = O(k logq (qn/k)) s.t. for any w 6= w′ ∈ Fnq with kwk0 , kw′ k0 ≤ k, Πk w 6= Πk w′ .
Πk ∈ Fm×n
q
26
Proof. The proof is via the probabilistic method. Πk w = Πk w′ iff Πk (w − w′ ) = 0. Note v = w − w′
has kvk0 ≤ 2k. Thus it suffices to show that such a Πk exists with no (2k)-sparse vector in its
n
· q 2k . For any fixed v,
kernel. The number of vectors v ∈ Fnq with kv0 k ≤ 2k is at most 2k
−m
P(Πk v = 0) = q . Thus
n
P(∃v, kvk0 ≤ 2k : Πk v = 0) ≤
· q 2k · q −m
2k
n
by a union bound. The above is strictly less than 1 for m > 2k + logq 2k
, yielding the claim.
Corollary 4. Let Fq be a finite field and n > 1 an integer. Then for any 1 ≤ k ≤ n2 , there exists
for m = O(k logq (qn/k)) together with an algorithm R such that for any w ∈ Fnq with
Πk ∈ Fm×n
q
kwk0 ≤ k, R(Πk w) = w.
Proof. Given Lemma 9, a simple such R is as follows. Given some y = Πk w∗ with kw∗ k0 ≤ k, R
loops over all w in Fnq with kwk0 ≤ k and outputs the first one it finds for which Πk w = y.
The protocol for URk is now as follows. Alice and Bob use public randomness to pick commonly
known random functions h0 , . . . , hL : [n] → {0, 1} for L = ⌊log2 (n/k)⌋, such that for any i ∈ [n]
and for any j, P(hj (i) = 1) = 2−j . They also agree on a matrix Π16k and R as described in
Corollary 4 for a sufficiently large constant C > 0 to be determined later, with q = 3. Thus
Π16k has m = O(k log(n/k)) rows. Alice then computes vj = Π16k x|h−1 (1) for j = 0, . . . , L where
j
vj ∈ F m
q , and her message to Bob is M = (v0 , . . . , vL ). For S ⊆ [n] and x an n-dimensional vector,
x|S denotes the n-dimensional vector with (x|S )i = xi for i ∈ S, and (x|S )i = 0 for i ∈
/ S. Note
Alice’s message M is O(k log2 (n/k)) bits, as desired. Bob then executes the following algorithm
and outputs the returned values.
Algorithm 10 Bob’s algorithm in the URk protocol.
1: procedure Bob(v0 , . . . , vL )
2:
for j = L, L − 1, . . . , 0 do
3:
vj ← vj − Π16k y|h−1 (1)
j
4:
5:
6:
7:
8:
9:
wj ← R(vj )
if kwj k0 ≥ k or j = 0 then
return an arbitrary min{k, kwj k0 } elements from support(wj )
end if
end for
end procedure
The correctness analysis is then as follows, which is nearly the same as the ℓ0 -sampler of [JST11].
If Alice’s input is x and Bob’s is y, let a = x − y ∈ {−1, 0, 1}n , so that a can be viewed as an
element of Fn3 . Also let aj = a|h−1 (1) . Then E kvj k0 = kak0 · 2−j , and since 0 ≤ kak0 ≤ n, there
j
∗
either (1) exists a unique 0 ≤ j ∗ ≤ L such that 2k ≤ E kaj k0 · 2−j < 4k, or (2) kak0 < 2k (in which
case we define j ∗ = 0). Let E be the event that kaj k0 ≤ 16k simultaneously for all j ≤ j ∗ . Let F
be the event that either we are in case (2), or we are in case (1) and kaj ∗ k0 ≥ k holds. Note that
conditioned on E, F both occurring, Bob succeeds by Corollary 4.
27
We now just need to show P(¬E ∧ ¬F) < e−Ω(k) . We use the union bound. First, consider
F. If j ∗ = 0, then P(¬F) = 0. If j ∗ 6= 0, then P(¬F) ≤ P(kaj ∗ k0 < 12 · E kaj ∗ k0 ), which is
e−Ω(k) by the Chernoff bound since E kaj ∗ k0 = Θ(k). Next we bound P(¬E). For j ≥ j ∗ , we know
∗
E kaj k0 ≤ 4k/2j−j . Thus, letting µ denote E kaj k0 ,
µ
16k
−1
µ
16k −Ω(k)
e
∗
(12)
< (e−Ck )j−j
<
P(kaj k0 > 16k) <
16k
µ
( 16k ) µ
µ
for some constant C > 0 by the Chernoff bound and the fact that 16k/µ ≥ 4 > e. Recall that the
Chernoff bound states that for X a sum of independent Bernoullis,
E X
eδ
∀δ > 0, P(X > (1 + δ) E X) <
.
(1 + δ)1+δ
Then by a union bound over j ≥ j ∗ and applying (12),
∗
P(¬E) = P(∃j ≥ j : kaj k0 > 16k) <
∞
X
∗
(e−Ck )j−j = O(e−Ck ).
j=j ∗
Remark 2. As already mentioned, the protocol given above and the one described in [JST11]
using O(k log2 n) bits differ in minor points. First: the protocol there used ⌊log2 n⌋ different hash
functions hj , but as seen above, only ⌊log2 (n/k)⌋ are needed. This already improves one log n factor
to log(n/k). The other improvement comes from replacing the k-sparse recovery structure with 2k
rows used in [JST11] with our Corollary 4. Note the matrix Πk in our corollary has even more rows,
but the key point is that the bit complexity is improved. Whereas using a k-sparse recovery scheme
as described in [JST11] would use 2k linear measurements of a k-sparse vector w ∈ {−1, 0, 1}n with
log n bits per measurement (for a total of O(k log n) bits), we use O(k log(n/k)) measurements with
only O(1) bits per measurement. The key insight is that we can work over Fn3 instead of Rn when
the entries of w are in {−1, 0, 1}, which leads to our slight improvement.
A.2
Proof of the existence of the desired Su,m
Lemma 7 (restated). For any integers u ≥ 1 and 1 ≤ m ≤ u/(4e), there exists a collection
′
′
Su,m ⊂ [u]
m with log |Su,m | = Θ(m log(u/m)) such that for all S 6= S ∈ Su,m , |S ∩ S | < m/2.
Proof. The proof is via the probabilistic
method. We pick S1 , . . . , SN independently, each one
uniformly at random from [u]
.
Fix
i
=
6
j ∈ [N ]. Imagine Si being fixed and picking the m
m
elements of Sj one by one. Let Xk denote the P
indicator random variable for the event that the kth
2
element picked is also in Si . Then |Si ∩ Sj | = m
k=1 Xk , and we set µ := E |Si ∩ Sj |, which is m /u
by linearity of expectation. We have P(|Si ∩Sj | ≥ m/2) = P(|Si ∩Sj | ≥ (1+δ)µ) for δ = u/(2m)−1.
The Xk are not independent, but they are negatively dependent. Thus the Chernoff bound yields
!m2 /u
µ
u
u − m
e 2m −1
eδ
2
.
≤
≤
P(|Si ∩ Sj | ≥ (1 + δ)µ) ≤
u
1+δ
u
(1 + δ)
2em
( 2m ) 2m
p
Setting N = (u/(2em))m/2 − 1 so that N2 ≤ N 2 = (u/(2em))m/2 − 1, by a union bound with
positive probability |Si ∩ Sj | < m/2 for all i 6= j, simultaneously, as desired. Note for this choice of
N , we have log |Su,m | = log N = Θ(m log(u/m)).
28
| 8cs.DS
|
Derandomized Balanced Allocation
Xue Chen∗
Computer Science Department
University of Texas at Austin
[email protected]
arXiv:1702.03375v2 [cs.DS] 24 Jan 2018
January 26, 2018
Abstract
In this paper, we study the maximum loads of explicit hash families in the d-choice schemes
when allocating sequentially n balls into n bins. We consider the Uniform-Greedy scheme
[ABKU99], which provides d independent bins for each ball and places the ball into the bin
with the least load, and its non-uniform variant — the Always-Go-Left scheme introduced by
Vöcking [Vöc03]. We construct a hash family with O(log n log log n) random bits based on the
previous work of Celis et al. [CRSW13] and show the following results.
1. This hash family has a maximum load of
2. It has a maximum load of
φd > 1.61.
log log n
d log φd
log log n
log d
+ O(1) in the Uniform-Greedy scheme.
+ O(1) in the Always-Go-Left scheme for a constant
The maximum loads of our hash family match the maximum loads of a perfectly random
hash function [ABKU99, Vöc03] in the Uniform-Greedy and Always-Go-Left scheme separately. Previously, the best known hash families that guarantee the same maximum loads as
a perfectly random hash function in any of these schemes were O(log n)-wise independent
functions [Vöc03], which needs Θ(log2 n) random bits.
∗
Supported by NSF Grant CCF-1526952 and a Simons Investigator Award (#409864, David Zuckerman), part of this work was
done while the author was visiting the Simons Institute.
1
Introduction
We investigate explicit constructions of hash functions for the classical problem of placing balls into bins.
The basic model is to hash n balls into n bins independently and uniformly at random, which we call 1choice scheme. A well-known and useful fact of the 1-choice scheme is that with high probability, each bin
contains at most O( logloglogn n ) balls. For convenience, we always use logarithm of base 2 in this work. Here,
by high probability, we mean probability 1 − n−c for an arbitrary constant c.
An alternative variant, which we call Uniform-Greedy, is to provide d ≥ 2 independent random choices
for each ball and place the ball in the bin with the lowest load. In a seminal work, Azar et al. [ABKU99]
showed that the Uniform-Greedy scheme with d independent random choices guarantees a maximum load
of only logloglogd n + O(1) with high probability for n balls. Later, Vöcking [Vöc03] introduced the AlwaysGo-Left scheme to further improve the maximum load to
φdd
φd−1
d .
log log n
d log φd
+ O(1) for d choices where φd > 1.61 is
the constant satisfying
= 1 + φd + · · · +
For convenience, we always use d-choice schemes to
denote the Uniform-Greedy and Always-Go-Left scheme with d ≥ 2 choices.
Traditional analysis of load balancing assumes a perfectly random hash function. A large body of research is dedicated to the removal of this assumption by designing explicit hash families using fewer random
bits. In the 1-choice scheme, it is well known that O( logloglogn n )-wise independent functions guarantee a max2
log n
imum load of O( logloglogn n ) with high probability, which reduces the number of random bits to O( log
log n ).
Recently, Celis et al. [CRSW13] designed a hash family with a description of O(log n log log n) random
bits that achieves the same maximum load of O( logloglogn n ) as a perfectly random hash function.
In this work, we are interested in the explicit constructions of hash families that achieve the same maximum loads as a perfectly random hash function in the d-choice schemes. More precisely, we study how to
derandomize the perfectly random hash function whose maximum loads are logloglogd n + O(1) in the Uniformlog n
Greedy scheme [ABKU99, Vöc03] and log
d log φd + O(1) in the Always-Go-Left scheme [Vöc03]. For these
two schemes, O(log n)-wise independent hash functions achieve the same maximum loads from Vöcking’s
argument [Vöc03], which provides a hash family with Θ(log2 n) random bits. Very recently, Reingold et
al. [RRW14] showed that the hash family designed by Celis et al. [CRSW13] guarantees a maximum load
of O(log log n) in the Uniform-Greedy scheme with O(log n log log n) random bits.
1.1
Our Contributions
For multiple-choice schemes, we strengthen the hash family of Celis et al. [CRSW13] for the 1-choice
scheme — our hash family is O(log log n)-wise independent over n bins and “almost” O(log n)-wise independent over a fraction of poly(log n) bins. Then we prove that our hash family derandomizes Vöcking’s
witness tree argument [Vöc03] such that O(log n log log n) random bits could guarantee the same maximum
loads as a perfectly random hash function in the multiple-choice schemes.
We first show our hash family guarantees a maximum load of logloglogd n + O(1) in the Uniform-Greedy
scheme [ABKU99, Vöc03] with d choices. We use U to denote the pool of balls and consider placing
m = O(n) balls into n bins here. Without loss of generality, we always assume |U | = poly(n) and d is a
constant at least 2 in this work.
Theorem 1.1 (Informal version of Theorem 5.1) For any m = O(n), any constants c and d, there exists
a hash family with O(log n log log n) random bits such that given any m balls in U , with probability at least
1 − n−c , the max-load of the Uniform-Greedy scheme with d independent choices of h is logloglogd n + O(1).
1
Our hash family has an evaluation time O (log log n)4 in the RAM model based on the algorithm designed
by Meka et al. [MRRR14] for the hash family of Celis et al. [CRSW13].
log n
Then we show this hash family guarantees a load balancing of log
d log φd + O(1) in the Always-Go-Left
scheme [Vöc03] with d choices. The Always-Go-Left scheme [Vöc03] is an asymmetric allocation scheme
that partitions the n bins into d groups with equal size and uses an unfair tie-breaking mechanism. Its
allocation process provides d independent choices for each ball from the d groups separately and always
chooses the left-most bin with the least load for each ball. We defer the formal description of the AlwaysGo-Left scheme to Section 6. Notice that the constant φd in equation φdd = 1 + φd + · · · + φdd−1 satisfies
1.61 < φ2 < φ3 < φ4 < · · · < φd < 2. Compared to the Uniform-Greedy scheme, the Always-GoLeft scheme [Vöc03] improves the maximum load exponentially with regard to d. Even for d = 2, the
Always-Go-Left scheme improves the maximum load from log log n + O(1) to 0.7 log log n + O(1).
Theorem 1.2 (Informal version of Theorem 6.3) For any m = O(n), any constants c and d, there exists
a hash family with O(log n log log n) random bits such that given any m balls in U , with probability at least
log n
1 − n−c , the max-load of the Always-Go-Left scheme with d independent choices of h is log
d log φd + O(1).
log n
At the same time, from the lower bound log
d log φd − O(1) on the maximum load of any random d-choice
scheme shown by Vöcking [Vöc03], the maximum load of our hash family is optimal for d-choice schemes
up to the low order term of constants.
Finally, we show our hash family guarantees the same maximum load as a perfectly random hash function in the 1-choice scheme for m = n · poly(log
Given m > n log n balls in U , the maximum
p n) balls.
m
load of the 1-choice scheme becomes m
+
O(
log
n
·
)
n
n from the Chernoff bound. For convenience,
we refer to this case of m ≥ n log n balls as a heavy load. In a recent breakthrough, Gopalan, Kane,
and Meka [GKM15] designed a pseudorandom generator of seed length O(log n(log log n)2 ) that fools the
Chernoff bound within polynomial error. Hence the pseudorandom generator [GKM15] provides a hash
function with O(log n(log log n)2 ) random bits for the heavy load case. Compared to the hash function
of [GKM15], we provide a simplified construction that achieves the same maximum load but only works for
m = n · poly(log n) balls.
Theorem 1.3 (Informal version of Theorem 7.1) For any constants c and a ≥ 1, there exist a hash function generated by O(log n log log n) random bits such that for any m = loga n · n balls,
at
√ withpprobability
m
least 1 − n−c , the max-load of the n bins in the 1-choice scheme with h is m
log
n
·
+
O
.
n
n
1.2
Previous Work
The 1-choice scheme. Natural explicit constructions of hash functions using a few random bits are k-wise
independent functions, small-biased spaces, and k-wise small-biased spaces. For the 1-choice scheme with
m = n balls, Alon et al. [ADM+ 99] showed the existence of a pairwise independent hash family that always
√
has a maximum load of n. On the other hand, it is well known that O( logloglogn n )-wise independent functions
2
log n
achieve a maximum load of O( logloglogn n ) with high probability, which needs Θ( log
log n ) random bits. Using
O(log n)-wise small-biased spaces as milder restrictions, Celis et al. [CRSW13] designed a hash family
with O(log n log log n) random bits achieving the same maximum load with high probability.
For the heavyp
load case in the 1-choice scheme, a perfectly random hash function guarantees a maximum
load of m
+
O(
log n · m
n
n ) from the Chernoff bound. Hence any pseudorandom generator fooling the
Chernoff bound within polynomial small error is a hash family matching this maximum load. Schmidt
et al. [SSS95] showed that O(log n)-wise independence could derandomize the Chernoff bound, which
2
provides a hash function with O(log2 n) random bits. In a recent breakthrough [GKM15], Gopolan, Kane,
and Meka designed a pseudorandom generator with seed length O(log n(log log n)2 ) to fool halfspaces, the
Chernoff bound, and many other classes, which provides a hash family of O(log n(log log n)2 ) bits.
Multiple-choice schemes. For m = n balls in the d-choice schemes, the original argument of [ABKU99]
adopts an inductive proof that relies on the assumption of full randomness. It is folklore (e.g., [RRW14,
DKRT16]) that O(log n)-wise independent functions could derandomize Vöcking’s witness tree argument
[Vöc03] to achieve a maximum load of logloglogd n +O(1) in the Uniform-Greedy scheme, which takes Θ(log2 n)
random bits. Reingold et al. [RRW14] prove the hash family of [CRSW13] derandomizes the following
property of random graphs: for a graph with n vertices and n/C0 random edges, every connected component
has size O(log n) with high probability. Besides applications in Cuckoo hashing, Reingold et al. [RRW14]
show the hash family [CRSW13] guarantees a maximum load of log log n + O(1) in the Uniform-Greedy
scheme for n/C0 balls. However, for m > n balls with d choices, the upper bound of the maximum load
proved by Reingold et al. [RRW14] becomes C0nm · log log n + O(1), while the maximum load of a perfectly
random hash function is logloglogd n + O( m
n ).
Vöcking [Vöc03] introduced Always-Go-Left scheme to further improve the maximum loads of d-choice
log n
schemes to log
d log φd + O(1). In the same work, Vöcking showed a lower bound to illustrate that the load
log log n
d log φd
is optimal for random d-choice schemes. However, much less is known about the derandomization of the Always-Go-Left scheme except O(log n)-wise independent functions for Vöcking’s witness tree
argument [Vöc03], which is pointed out in [RRW14, DKRT16].
Another long line of research on hash families focuses on studying functions with a constant evaluation
time despite the expense of the number of random bits. For O(log n)-wise independence, Siegel [Sie89]
showed how to implement it in constant time. For multiple-choice schemes, Woelfel [Woe06] showed that
the hash family of [DW03], which takes constant evaluation time and nΘ(1) random bits, guarantees the
same maximum loads as a perfectly random hash functions in the multiple-choices schemes. Pǎtraşcu and
Thorup [PT12] introduced simple tabulation hashing, a function with constant evaluation time and nΘ(1)
random bits, that can replace the perfectly random hash functions in various applications. Very recently,
Dahlgaard et al. [DKRT16] proved that the maximum load of the Uniform-Greedy scheme is O(log log n)
with high probability in simple tabulation [PT12] using O(log2 n) random bits. For the hash family in
[CRSW13], Meka et al. [MRRR14] improved its evaluation time to O((log log n)2 ).
scheme
1-choice
1-choice
Uniform-Greedy
Uniform-Greedy
Uniform-Greedy
Uniform-Greedy
Uniform-Greedy
Always-Go-Left
Always-Go-Left
reference
well known
[CRSW13]
[ABKU99]
[Vöc03]
[RRW14]
[DKRT16]
this work
[Vöc03]
this work
maximum load
O( logloglogn n )
O( logloglogn n )
log log n
log d + O(1)
log log n
log d + O(1)
O(log log n)
O(log log n)
log log n
log d + O(1)
log log n
d log φd + O(1)
log log n
d log φd + O(1)
number of random bits
log2 n
Θ( log
log n )
O(log n log log n)
full randomness
Θ(log2 n)
O(log n log log n)
Θ(log2 n)
O(log n log log n)
Θ(log2 n)
O(log n log log n)
Table 1: Summary of previous works about the maximum loads of placing n balls in n bins with d choices
3
We summarize these results in Table 1. Finally, we refer surveys [MRS00, Mit01, Wie17] and the reference therein for various applications of multiple-choice schemes in computer science.
1.3
Discussion
In this work, we provide a hash family with O(log n log log n) random bits that matches the maximum loads
of a perfectly random hash function in multiple-choice schemes. A natural question is to reduce the number
of random bits to O(log n). A starting point would be to improve the hash families in the 1-choice scheme,
where the best construction needs O(log n log log n) random bits from Celis et al. [CRSW13].
For the 1-choice scheme, the load of each bin is the summation of m random indicator variables,
which allows us to use the pseudorandom generators for concentration bounds [SSS95, GKM15] and spacebounded computation [Nis92, NZ96, GKM15]. One interesting direction is to investigate the application of
the techniques in these two problems to the design of hash functions. At the same time, although Alon et
al. [ADM+ 99] proved lower bounds of k-wise independent functions in the 1-choice scheme, it is still interesting to explore natural algebraic constructions of small-biased spaces in [AGHP90] such as the quadratic
characters of modulo p.
Our work is an application of the technique — milder restrictions [CRSW13, GMR+ 12] in the design
of pseudorandom generators. Even though k-wise independence and small-biased spaces fool variants of
classes, these two tools will not provide optimal pseudorandom generators for basic classes such as the 1choice scheme [ADM+ 99] or read-once CNFs [DETT10]. After Celis et al. [CRSW13] introduced milder
restrictions, this technique has been successfully applied to construct almost optimal pseudorandom generators with log n · poly(log log n) random bits for several classes such as 1-choice scheme [CRSW13],
read-once CNFs [GMR+ 12], modulo p functions and halfspaces [GKM15]. To the best of our knowledge,
our work is the first application of milder restrictions in non read-once functions, while all previous applications of milder restrictions are in the read-once case. It would be of great interest to investigate this
technique to the design of pseudorandom generators for broader classes such as AC 0 circuits and spacebounded computation.
1.4
Organization
This paper is organized as follows. In Section 2, we introduce some notations and tools. We define witness
trees and revisit Vöcking’s argument [Vöc03] in Section 3. We show the construction of our hash family
in Section 4 and sketch our derandomization in Section 4.1. Next we prove Theorem 1.1 in Section 5
and Theorem 1.2 in Section 6, which provide upper bounds on the maximum loads of the Uniform-Greedy
scheme and the Always-Go-Left scheme separately. Finally, we prove Theorem 1.3 in Section 7 which shows
a bound of the heavy load case in the 1-choice scheme.
2
Preliminaries
We use U to denote the pool of balls, m to denote the numbers of balls in U , and n to denote the number of
bins. We assume m ≥ n and n is a power of 2 in this work. We use 1E to denote the indicator function of
the event E and Fp to denote the Galois field of size p for a prime power p.
Definition 2.1 Given a prime power p, a distribution D on Fnp is a δ-biased space if for any non-trivial
character function χα in Fnp , E [χα (x)] ≤ δ.
x∼D
4
A distribution D on Fnp is a k-wise δ-biased space if for any non-trivial character function χα in Fnp of
support size at most k, E [χα (x)] ≤ δ.
x∼D
The seminal works [NN90, AGHP90] provide small-biased spaces with optimal seed length.
Lemma 2.2 ( [NN90, AGHP90]) For any prime power p and integer n, there exist explicit constructions of
δ-biased spaces on Fnp with seed length O(log pn
δ ) and explicit constructions of k-wise δ-biased spaces with
kp log n
seed length O(log δ )
n
Given two
Pdistributions D1 and D2 with the same support Fp , we define the statistical distance to be kD1 −
D2 k1 = x∈Fnp |D1 (x)−D2 (x)|. Vazirani [Vaz86] proved that small-biased spaces are close to the uniform
distribution.
Lemma 2.3 ( [Vaz86]) A δ-biased space on Fnp is δ · pn/2 close to the uniform distribution in statistical
distance.
Given a subset S of size k in [n], a k-wise δ-biased space on Fnp is δ·pk/2 close to the uniform distribution
on S in statistical distance.
Given a distribution D on functions from U to [n], D is k-wise independent if for any k elements
x1 , . . . , xk in U , D(x1 ), . . . , D(xk ) is a uniform distribution on [n]k . For small-biased spaces, we choose
|U |
p = n and the space to be Fn in Lemma 2.2 and summarize the discussion above.
Lemma 2.4 Given k and n, a k-wise δ-biased space from U to [n] is δ·nk/2 close to the uniform distribution
n
) random bits.
from U to [n] on any k balls, which needs O(log kn log
δ
Remark 2.5 In this work, we always choose δ ≤ 1/n and k = poly(log n) in the small biased spaces such
that the seed length is O(log 1δ ). At the same time, we only use k-wise small-biased spaces rather than small
biased spaces to improve the evaluation time from O(log n) to O(log log n)4 .
We state the Chernoff bound in k-wise independence by Schmidt et al. in [SSS95].
Lemma 2.6 (Theorem 5 (I) (b) in [SSS95]) If X is the sum of k-wise independent random variables, each
of which is confined to the interval [0, 1] with µ = E[X], then for δ ≤ 1 and k ≥ δ 2 µ · e−1/3 ,
Pr[|X − µ| ≥ δµ] ≤ e−δ
3
2 µ/3
.
Witness Trees
We first provide several notations and definitions in this work. Then we review the witness tree argument of
Vöcking [Vöc03] for the Uniform-Greedy scheme.
Definition 3.1 (Uniform-Greedy with d choices) The process inserts balls in any fixed order. Let h(1) , . . . , h(d)
be d hash functions from U to [n]. The allocation process works as follows: for each ball i, the algorithm considers d bins {h(1) (i), . . . , h(d) (i)} and puts the ball i into the bin with the least load among
{h(1) (i), . . . , h(d) (i)}. When there are several bins with the least load, it picks an arbitrary one.
We define the height of a ball to be the height of it on the bin allocated in the above process.
Next we follow the notations of Vöcking [Vöc03] to define witness trees and pruned witness trees. Given
the balls and d hash functions h(1) , . . . , h(d) in the allocation process, we construct a symmetric witness tree
for each ball in this process.
5
Definition 3.2 (Symmetric witness trees) A symmetric witness tree T with height l for a ball b is a complete d-ary tree of height l. Every node w in this tree corresponds to a ball T (w) ∈ [n]; and the root
corresponds to the ball b. A ball u in T has a ball v as its ith child iff when we allocate
u in the process, ball
v is the top ball in the bin h(i) u . Hence v < u and the bin h(i) (u) is in the subset h(1) (v), . . . , h(d) (v)
of [n] when v is the ith child of u.
Next we trim the repeated nodes in a witness trees such that there is no duplicate edge after the trimming.
Definition 3.3 (Pruned witness trees and collisions) Given a witness tree T where nodes v1 , . . . , vj in T
correspond to the same ball, let v1 be the node among them in the most bottom level of T . Consider the
following process: first remove v2 , . . . , vj and their subtrees; then, redirect the edges of v2 , . . . , vj from
their parents to v1 and call these edges collisions. Given a symmetric witness tree T , we call the new tree
without repeated nodes after the above process as the pruned witness tree of T .
We call different witness trees with the same structure but different balls a configuration. For example, the
configuration of symmetric witness trees with distinct nodes is a full d-ary tree without any collision.
48
24
1
3
4
48
37
42
19
26
8
11
13
23
14
24
18
1
3
48
37
42
19
26
6
1
3
11
37
Trim T
19
6
24
1
1
42
19
3
6
26
11
Figure 1: A witness tree with distinct balls and a pruned witness tree with 3 collisions
Next we define the height and size of pruned witness trees.
Definition 3.4 (Height of witness trees) Given any witness tree T , let the height
of T be the length of the
shortest path from the root of T to its leaves. Because height(u) =
min
height(v) + 1, the height
v∈children(u)
of the pruned witness tree equals the height of the original witness tree. Given a ball b of height h and any
h0 < h, we always consider the pruned witness tree of b with height h0 whose leaves have height h − h0 .
At the same time, let |T | denote the number of vertices in T for any witness tree T and |C| denote the
number of nodes in a configuration C.
Finally we review the argument of Vöcking [Vöc03] for m = n balls. One difference between this proof
and Vöcking’s [Vöc03] proof is an alternate argument for the case of witness trees with many collisions.
Lemma 3.5 ( [Vöc03]) For any constants c ≥ 2 and d, with probability at least 1−n−c , the max-load of the
Always-Go-Left scheme with d independent choices from perfectly random hash functions is logloglogd n +O(1).
Proof. We fix a parameter l = dlogd (2 + 2c) log n + 3c + 5e for the height of witness trees. In this proof,
we bound the probability that any symmetric witness tree of height l with leaves of height at least 4 exists
6
in perfectly random hash functions. From the definition of witness trees, this also bounds the probability of
a ball with height l + 4 in the d-choice Uniform-Greedy scheme.
For symmetric witness trees of height l, it is sufficient to bound the probability that their pruned counterparts appear in perfectly random hash functions. We separate all pruned witness trees into two cases
according to the number of edge collisions: pruned witness trees with at most 3c collisions and pruned
witness trees with at least 3c collisions.
Pruned witness trees with at most 3c collisions. Let us fix a configuration C with at most 3c collisions
and consider the probability any pruned witness trees with configuration C appears in perfectly random
hash functions. Because each node of this configuration C corresponds a distinct ball, there are at most n|C|
possible ways to instantiate balls into C.
Next, we fix one possible pruned witness tree T and bound the probability of the appearance of T in
h(1) , . . . , h(d) . We consider the probability of two events: every edge (u, v) in the tree T appears during the
allocation process; and every leaf of T has height at least 4. For the first event, an edge (u, v) holds during
the process only if the hash functions satisfy
n
o
d
h(i) (u) ∈ h(1) (v), . . . , h(d) (v) , which happens with probability at most .
n
(1)
Secondly, the probability that a fixed leaf ball has height at least 4 is at most 3−d . A leaf ball of height 4
indicates that each bin in his choices has height at least 3. Because at most n/3 bins contain at least 3 balls
at any moment, the probability that a random bin has height at least 3 is ≤ 1/3. Thus the probability that d
random bins have height 3 is at most 3−d .
We apply a union bound on the probability that any witness tree with the configuration C appears in
perfectly random hash functions:
n|C| ·
Y
(u,v)∈C
d
· (3−d )number of leaves
n
(2)
We lower bound the number of edges in C by |C| − 1 because C is connected. Next we lower bound
the number of leaves. Because C is a d-ary tree with at most 3c collisions, the number of leaves is at
. At the same time, C is trimmed from the d-ary symmetric witness tree of height l. Thus
least |C|−3c
2
|C| ≥ (1 + d + · · · + dl−3c ). From all discussion above, we bound (2) by
|C|−3c
d
l−3c
n|C| ·( )|C|−1 ·(3−d ) 2 ≤ n·(d2.5 ·3−d )|C|/2.5 ≤ n·(d2.5 ·3−d )d /2.5 ≤ n·(0.8)10(2+2c) log n ≤ n−2c−1 .
n
Finally, we apply a union
on all possible configurations with at most 3c collisions: the number of
P bound
l+1 )2·i ≤ n such that the probability of any witness tree with height l and
configurations is at most 3c
(d
i=0
at most 3c collisions existing is at most n−c .
Pruned witness trees with at least 3c collisions. We use the extra 3c collisions with equation (1) instead
of the number of leaves in this case.
Given any configuration C with at least 3c collisions, we consider the first 3c collisions e1 , . . . , e3c in the
BFS of C. Let C 0 be the induced subgraph of C that only contains nodes in e1 , . . . , e3c and their ancestors
in C. At the same time, the size |C 0 | ≤ 3c(2l + 1) and the number of edges in C 0 is |C 0 | + 3c − 1.
7
Extract C 0 from C
Figure 2: An example of extracting C 0 from C given two collisions.
Because any pruned witness tree of C exists only if its corresponding counterpart of C 0 exists in perfectly
0
random hash functions, it is suffice to bound the probability of the latter event. There are at most n|C |
instantiations of balls in C 0 . For each instantiation, we bound the probability that all edges survive by (1):
d
d 0
( )number of edges = ( )|C |+3c−1 .
n
n
We bound the probability that any pruned witness tree of configuration C 0 survives in the perfectly
random hash function by
1
d 0
0
( )|C |+3c−1 · n|C | ≤ ( )3c−1 · d(2l+2)·3c ≤ n−2c .
n
n
Finally, we apply a union bound over all possible configurations C 0 : there are at most (1+d+· · ·+dl )2·3c ≤
n configurations of 3c collisions.
t
u
Remark 3.6 Because the sizes of all witness trees are bounded by dl+1 = O(log n), Oc,d (log n)-wise
independent hash functions could adopt the above argument to prove a max-load of logd log n + O(d + c).
4
Hash functions
We construct our hash family and show its properties for the derandomization of d-choice schemes in this
section. We sketch the derandomization of Lemma 3.5 of Vocking’s argument in Section 4.1.
Let ◦ denote the concatenation operation and ⊕ denote the bit-wise XOR operation.
Construction 4.1 Given δ1 > 0, δ2 > 0, and two integers k, kg , let
−i
1. hi : U → [n2 ] denote a function generated by an O(log2 n)-wise δ1 -biased space for each i ∈ [k],
−k
2. hk+1 : U → [n2 ] denote a function generated by an O(log2 n)-wise δ2 -biased space such that
(h1 (x) ◦ h2 (x) ◦ · · · ◦ hk (x) ◦ hk+1 (x)) is a function by U to [n],
3. g : U → [n] denote a function from a kg -wise independent family from U to [n].
We define a random function h : U → [n] in our hash family H with parameters δ1 , δ2 , k and kg to be:
h(x) = h1 (x) ◦ h2 (x) ◦ · · · ◦ hk (x) ◦ hk+1 (x) ⊕ g(x).
8
2
2
|U |
|U |
Hence the seed length of our hash family is O(k log n·log δn·log
+ log n·log δn·log
+ kg log n). We
1
2
−O(log
n) such that the
always choose k ≤ log log n, kg = O(log log n), δ1 = 1/poly(n), and δ2 = (log n)
seed length is O(log n log log n).
Remark 4.2 Our parameters of h1 ◦ · · · ◦ hk+1 are stronger than the parameters in [CRSW13]. While the
last function hk+1 of [CRSW13] is still a δ1 -biased space, we use δ2 = (δ1 )O(k) in hk+1 to provide almost
O(log n)-wise independence on (log n)O(log n) subsets of size O(log n) for our calculations.
Properties of h. We state the properties of h that will be used in the derandomization. Because of the
kg -wise independence in g and the ⊕ operation, we have the same property for h.
Property 4.3 h is kg -wise independent.
Then we fix g and discuss h1 ◦ · · · ◦ hk ◦ hk+1 . For each i ∈ [k], it is natural to think h1 ◦ · · · ◦ hi as
1
1
a function from U to [n1− 2i ], i.e., a hash function maps all balls into n1− 2i bins. Celis et al. [CRSW13]
showed that for every i ∈ [k], the number of balls in every bin of h1 ◦ · · · ◦ hi is close to its expectation
1
1
n 2i · m
n in poly(n) -biased spaces.
Lemma 4.4 ( [CRSW13]) Given k = log2 (log n/3 log log n) and β = (log n)−0.2 , for any constant c > 0,
there exists δ1 = 1/poly(n) such that given m = O(n) balls, with probability at least 1 − n−c , for all
1
1
i ∈ [k], every bin in [n1− 2i ] contains at most (1 + β)i n 2i · m
n balls under h1 ◦ · · · ◦ hi .
For completeness, we provide a proof of Lemma 4.4 in Appendix A. In this work, we use the following
version that after fixing g in the Construction 4.1, h1 ◦ h2 ◦ · · · ◦ hk still allocates the balls evenly.
Corollary 4.5 For any constant c > 0, there exists δ1 = 1/poly(n) such that given m = O(n) balls and any
1−
1
function g0 : U → [n/ log3 n], with probability at least 1 − n−c over h1 , . . . , hk , for any bin j ∈ [n 2k ] =
[n/ log3 n], it contains at most 1.01 · log3 n · m
n balls in the hash function h1 (x) ◦ · · · ◦ hk (x) ⊕ g0 (x).
Next we discuss the last function hk+1 generated from a δ2 -biased space on [log3 n]U . For a subset
S ⊆ U , let h(S) denote the distribution of a random function h on S and U[log3 n] (S) denote the uniform
distribution over all maps from S → [log3 n]. From Lemma 2.3 and the union bound, we have the following
claim.
Claim 4.6 Given δ2 = (log n)−C log n , for a fixed subset S of size
C
3
C
·log n, hk+1 (S) is (log n)− 2 ·log n -close
C
to the uniform distribution on S, i.e., khk+1 (S) − U[log3 n] (S)k1 ≤ (log n)− 2 ·log n .
C
Then for m = (log n) 3 ·log n subsets S1 , . . . , Sm of size C3 · log n, we have
X
C
C
khk+1 (Si ) − U[log3 n] (Si )k1 ≤ m · (log n)− 2 ·log n ≤ (log n)− 6 ·log n .
i∈[m]
C
In another word, hk+1 is close to the uniform distribution on (log n) 3
hk+1 (or h) is not close to log n-wise independence on n balls.
log n
subsets of size
C
3
log n. However,
Remark 4.7 (Evaluation time) Our hash function has an evaluation time O((log log n)4 ) in the RAM
model. Because we use (log n)−O(log log n) -biased spaces in hk+1 , we lose a factor of O(log log n)2 compared to the hash family of [CRSW13]. The reason is as follows.
9
g can be evaluated by a degree O(log log n) polynomial in the Galois field of size poly(n), which takes
O(log log n) time. The first k hash functions h1 , . . . , hk use 1/poly(n)-biased spaces, which have total
evaluation time O(k · log log n) = O(log log n)2 in the RAM model from [MRRR14].
The last function hk+1 in the RAM model is a O(log n)-wise n−O(log log n) -biased space from U to
[log3 n], which needs O(log log n) words in the RAM model. Thus the evaluation time becomes O(log log n)
times the cost of a quadratic operation in the Galois field of size nO(log log n) , which is O((log log n)4 ).
4.1
Proof Overview
We sketch the derandomization of Lemma 3.5 in this section. Similar to the proof of Lemma 3.5, we bound
the probability that any pruned witness tree of height l = logd log n + O(1) exists in h(1) , . . . , h(d) , where
(i)
(i)
each h(i) = h1 (x) ◦ · · · ◦ hk+1 (x) ⊕ g (i) (x). We use the property of h1 ◦ · · · ◦ hk+1 to derandomize the
case of pruned witness trees with at most 3c collisions and the property of g to derandomize the other case.
Pruned witness trees with at most 3c collisions. We show how to derandomize the union bound (2)
for a fixed configuration C with at most 3c collisions. There are two probabilities in (2): the second term
Q
d
−d·number of leaves over all leaves. We focus on the second
(u,v)∈C n over all edges in C and the last term 3
Q
term (u,v)∈C nd in this discussion, because it contributes a smaller probability. Since |C| ∈ [dl−3c , dl+1 ] =
Θ(log n), it needs O(log n)-wise independence over [n] bins for every possible witness trees in (2), which
is impossible to support with o(log2 n) bits [Sti94].
(i)
(i)
(i)
We omit {g (1) , . . . , g (d) } and focus on the other part h1 ◦ · · · ◦ hk ◦ hk+1 i ∈ [d] in this case. Our
(i)
(i)
strategy is to first fix the prefixes in the d hash functions, h1 ◦ · · · ◦ hk i ∈ [d] , then recalculate (2) using
(1)
(d)
the suffixes hk+1 , . . . , hk+1 . Let T be a possible witness tree in the configuration C. For an edge (u, v) in
T to satisfy (1), the prefixes of h(1) (v), . . . , h(d) (v) and h(i) (u) must satisfy
n
o
(i)
(i)
(1)
(1)
(d)
(d)
h1 (u) ◦ · · · ◦ hk (u) ∈ h1 (v) ◦ · · · ◦ hk (v), . . . , h1 (v) ◦ · · · ◦ hk (v) .
(3)
After fixing the prefixes, let FT denote the subset of possible witness trees in the configuration C that
satisfy the prefix condition (3) for every edge. Because each bin of [n/ log3 n] receives at most 1.01 log3 n
(j)
(j)
(j)
balls from every prefix function h1 ◦ h2 ◦ · · · ◦ hk by Corollary 4.5, we could bound
|FT | ≤ n(d · 1.01 log3 n)|C|−1 = n · (1.01d)|C| · (log3 n)|C|−1 = (log n)O(log n)
instead of n|C| in the original argument.
(1)
(d)
Now we consider all possible witness trees in FT under the suffixes hk+1 , . . . , hk+1 . We could treat
(1)
(d)
hk+1 , . . . , hk+1 as O(log n)-wise independent functions for all possible witness trees in FT from Claim 4.6,
because |C| = O(log n) and |FT | = (log n)O(log n) . In the next step, we use O(log n)-wise independence
to rewrite (2) and finish the proof of this case.
Pruned witness trees with at least 3c collisions. In our alternate argument of this case in Lemma 3.5,
the subconfiguration C 0 of C has at most 3c · (2l + 1) nodes and 3c · (2l + 1) + 3c edges. Since l =
logd log n + O(1), the number of edges in C 0 is O(log log n). By choosing kg = Θ(log log n) with a
sufficiently large constant, h with kg -wise independence supports the argument in Lemma 3.5.
10
5
The Uniform Greedy scheme
We prove our main result for the Uniform-Greedy scheme — Theorem 1.1 in this section.
Theorem 5.1 For any m = O(n), any constant c ≥ 2, and integer d, there exists a hash family H from
Construction 4.1 with O(log n log log n) random bits that guarantees the max-load of the Uniform Greedy
−c for
scheme with d independent choices from H is logd log n + O c + m
n with probability at least 1 − n
any m balls in U .
Proof. We specify the parameters of H as follows: kg = 10c(logd log m + logd (2 + 2c) + 5 + 3c), k =
log n
−C log n for a large constant C, and δ = 1/poly(n) such that Corollary 4.5 holds
log2 3 log
1
log n , δ2 = log n
−c−1
(1)
(d)
with probability at least 1 − n
. Let h , . . . , h denote the d independent hash functions from H with
the above parameters, where each
(j)
(j)
(j)
(j)
h(j) (x) = h1 (x) ◦ h2 (x) ◦ · · · ◦ hk (x) ◦ hk+1 (x) ⊕ g (j) (x).
We use the notation g to denote {g (1) , g (2) , . . . , g (d) } in the d choices and hi to denote the group of hash
(1)
(d)
functions {hi , . . . , hi } in this proof.
We bound the probability that any symmetric witness tree of height l = dlogd log m + logd (2 + 2c) +
(1)
(d)
5 + 3ce with leaves of height at least b = 10d · m
n + 1 exists in h , . . . , h . Similar to the proof of
Lemma 3.5, we bound the probability of pruned witness trees of height l in h(1) , . . . , h(d) . We separate all
pruned witness trees into two cases according to the number of edge collisions: pruned witness trees with at
most 3c collisions and pruned witness trees with at least 3c collisions.
Pruned witness trees with at least 3c collisions. We start with a configuration C of pruned witness trees
with height l and at least 3c collisions. Let e1 , . . . , e3c be the first 3c collisions in the BFS of C. Let C 0
be the induced subgraph of C that only contains nodes in these edges e1 , . . . , e3c and their ancestors in C.
Therefore any pruned witness tree T of configuration C exists in h(1) , . . . , h(d) only if the corresponding
counterpart T 0 of T with configuration C 0 exists in h(1) , . . . , h(d) . The existence of T 0 in h(1) , . . . , h(d)
indicates that for every edge (u, v) in T 0 , h(1) , . . . , h(d) satsify
n
o
when v is the ith child of u.
(4)
h(i) T (u) ∈ h(1) T (v) , . . . , h(d) T (v)
Notice that the number of edges in C 0 and T 0 is at most 3c · 2l + 3c = 2l(3c + 1) ≤ kg /2.
Because h(1) , . . . , h(d) are kg -wise independent, We bound the probability that all edges of T 0 satisfy
(4) in h(1) , . . . , h(d) by
Y d
d 0
( ) = ( )|C |+3c−1 .
n
n
0
(u,v)∈T
0
Now we apply a union bound over all choices of balls in C 0 . There are at most m|C | choices of balls in
the nodes of C 0 . Therefore we bound the probability that any witness with at least 3c collisions survives in
kg -wise independent functions by
d 0
d
m
d
m
0
0
( )|C |+3c−1 · m|C | ≤ ( )3c−1 · ( · d)|C | ≤ ( )3c−1 · ( · d)3c·(2l+1) ≤ n−2c .
n
n
n
n
n
Next we apply a union bound over all configurations C 0 . Because there are at most (1+d+· · ·+dl )2·3c ≤
n configurations of 3c collisions, with probability at least 1 − n−c , there is no pruned witness trees with at
least 3c collision and height l exists in h(1) , . . . , h(d) .
11
Pruned witness trees with at most 3c collisions. We fix a configuration C of pruned witness trees with
height l and less than 3c collisions. Next we bound the probability that any pruned witness trees in this
configuration C with leaves of height at least b exists in h(1) , . . . , h(d) .
We extensively use the fact that after fixing g and h1 ◦ · · · ◦ hk , at most d(1.01 log3 n · m
n ) elements in
3
(1)
(d)
h , . . . , h are mapped to any bin of [n/ log n] from Corollary 4.5. Another property is the number of
leaves in C: because there are at most 3c collisions in C, C has at least dl−3c ∈ [d5 (2 + 2c) log m, d6 (2 +
2c) log m] leaves. On the other hand, the number of leaves is at least |C|−3c
.
2
(1)
For a pruned witness tree T with configuration C, T exists in h , . . . , h(d) only if
n
o
∀(u, v) ∈ C, h(i) T (u) ∈ h(1) T (v) , . . . , h(d) T (v)
when v is the ith child of u.
(5)
We restate the above condition on the prefixes and suffixes of h(1) , . . . , h(d) separately. Let gp (x) denote
the first log n − 3 log log n bits of g(x) and gs (x) denote the last 3 log log n bits of g(x), which matches
(i)
(i)
h1 (x) ◦ · · · ◦ hk (x) and hk+1 (x) separately. Since h(i) (x) = h1 (x) ◦ · · · ◦ hk+1 (x) ⊕ g (i) (x), property
(5) indicates that the prefixes of the balls bu = T (u) and bv = T (v) satisfy
n
o
(i)
(i)
(1)
(1)
(d)
(d)
h1 (bu )◦· · ·◦hk (bu ) ⊕gp(i) (bu ) ∈ h1 (bv ) ◦ · · · ◦ hk (bv ) ⊕ gp(1) (bv ), . . . , h1 (bv ) ◦ · · · ◦ hk (bv ) ⊕ gp(d) (bv ) .
(6)
and their suffixes satisfy
n
o
(i)
(1)
(d)
hk+1 (bu ) ⊕ gs(i) (bu ) ∈ hk+1 (bv ) ⊕ gs(1) (bv ), . . . , hk+1 (bv ) ⊕ gs(d) (bv ) .
(7)
Let FT be the subset of witness trees in the configuration C whose edges satisfy the condition (6) in
preffixes h(1) , . . . , h(k) , i.e., FT = {T |configuration(T ) = C and (u, v) satisfies (6) ∀(u, v) ∈ T }. We
show that
m
|FT | ≤ m · (d · 1.01 log3 n · )|C|−1 .
n
The reason is as follows. There are m choices of balls for the root u in C. For the ith child
the root u,
v of
(i)
(i)
(i)
we have to satisfy the condition (6) for (u, v). For a fixed bin h1 (bu ) ◦ · · · ◦ hk (bu ) ⊕ gp (bu ), there
(j) mapped to this bin from Corollary 4.5.
are at most 1.01 · log3 n · m
n elements from each hash function h
Hence there are at most d · 1.01 log3 n · m
n choices for each child of u. Then we repeat this arguments for
all non-leaf nodes in C.
(1)
(d)
Next we consider the suffixes hk+1 , . . . , hk+1 . We first calculate the probability that any possible witness
(1)
(d)
tree in FT survives in hk+1 , . . . , hk+1 from t-wise independence for t = 5b · dl+2 = O(log n). After fixing
(1)
(d)
gs , for a possible witness tree T in FT , hk+1 , . . . , hk+1 satisfy (7) for every edge (u, v) ∈ C with probability
d
in t/2-wise independent distributions because the number of edges in C is less than t/2.
log3 n
2
−3d · ( n )2d
For each leaf v in T , we bound the probability that its height is at least b = 10d · m
n + 1 by 2
m
in (b · d + 1)-wise independence. Given a choice i ∈ [d] of leaf v, we fix the bin to be h(i) (v). Then we
bound the probability that there are at least b − 1 balls w1 , . . . , wb−1 in this bin excluding all balls in the tree
by
X
X
X
···
Pr[h(i) (v) = h(j1 ) (w1 ) = · · · = h(jb−1 ) (wb−1 )]
w1 :w1 <v,w1 ∈T
/ w2 :w1 <w2 <v,w2 ∈T
/
wb−1 :wb−2 <wb−1 <v,wb−1 ∈T
/
≤
12
1.01d·log3 n· m
n
b−1
(log3 n)b−1
≤
b−1
(1.01d · m
3
n)
≤ ( )b−1 .
(b − 1)!
4
2
n 2d
) .
For all d choices of this leaf v, this probability is at most ( 34 )(b−1)·d ≤ 2−3d · ( m
Because w1 , . . . , wb are not in the tree T for every leaf, they are disjoint and independent with the events
of (7) in T , which are over all edges in the tree. Hence we could multiply these two probability together
in t-wise independence given t/2 ≥ (b · d + 1) · number of leaves. Then we apply a union bound over all
possible pruned witness trees in FT to bound the probability (in the t-wise independence) that there is one
witness tree of height l whose leaves have height at least 10d · m
n + 1 by
|C|
|C|−3c
m
d
d |C|−1
3 b·d number of leaves
n 2d 2
3
−3d2
≤m 1.01d · log n ·
·
|FT | · ( 3 )
· ( )
· 2
·( )
4
n log3 n
m
log n
|C|
|C|/3
m
n
2
≤m · 2d2 ·
· 2−3d · ( )2d
≤ m · 2−|C|/3 ≤ n−c−1 .
n
m
Finally we replace the t-wise independence by a δ2 -biased space for δ2 = n−c−1 · (log3 n)−t /|FT | =
(log n)−O(log n) . We apply Claim 4.6 to all possible pruned witness tress in FT : in δ2 -biased spaces, the
probability of the existence of any height-l witness tree with leaves of height at least b = 10d · m
n + 1 is at
most
n−c−1 + |FT | · δ2 · (log3 n)t ≤ 2n−c−1 .
Then we apply a union bound on all possible configurations with at most 3c collisions:
(dl+1 )|3c| · 2n−c−1 ≤ 0.5n−c .
From all discussion above, with probability at least 1 − n−c , there is no ball of height more than l + b =
logd log n + O(1).
t
u
6
The Always-Go-Left Scehme
log n
We show that the hash family in Section 4 with proper parameters also achieves a max-load of log
d log φd +O(1)
in the Always-Go-Left scheme [Vöc03] with d choices, where φd > 1 is the constant satisfying φdd =
1 + φd + · · · + φd−1
d . We define the Always-Go-Left scheme [Vöc03] as follows:
Definition 6.1 (Always-Go-Left with d choices) Our algorithm partition the bins into d groups G1 , . . . , Gd
of the same size n/d. Let h(1) , . . . , h(d) be d functions from U to G1 , . . . , Gd separately. For each ball b,
the algorithm consider d bins {h(1) (b) ∈ G1 , . . . , h(d) (b) ∈ Gd } and chooses the bin with the least number
of balls. If there are several bins with the least number of balls, our algorithm always choose the bin with
the smallest group number.
We define asymmetric witness trees for the Always-Go-Left mechanism such that a ball of height l + C
in the Always-Go-Left scheme indicates that there is an asymmetric witness tree of height l whose leaves
have height at least C. For an asymmetric witness tree T , the height of T is still the shortest distance from
the root to its leaves.
Definition 6.2 (Asymmetric Witness tree) The asymmetric witness tree T of height l in group Gi is a dary tree. The root has d children where the subtree of the jth child is an asymmetric witness tree in group
Gj of height (l − 1j≥i ).
Given d functions h(1) , . . . , h(d) from U to G1 , . . . , Gd separately, a ball b with height more than l + C
in a bin of group Gi indicates an asymmetric witness tree T of height l in Gi whose leaves have height at
13
least C. Each node of T corresponds to a ball, and the root of T corresponds to the ball b. A ball u in T
has a ball v as its jth child iff when we insert the ball u in the Always-Go-Left mechanism, v is the top ball
in the bin h(j) (u). Hence v < u and h(j) (u) = h(j) (v) when the jth child of u is v.
For an asymmetric witness tree T of height l in group Gi , We use the height l and the group index i ∈ [d] to
determine its size. Let f (l, i) be the size of a full asymmetric witness tree of height l in group Gi . From the
definition, we have f (0, i) = 1 and
f (l, i) =
i−1
X
f (l, j) +
j=1
d
X
f (l − 1, j).
j=i
Let g (l − 1) · d + i = f (l, i) such that
g(n) = g(n − 1) + g(n − 2) + · · · + g(n − d).
We know there exist c0 > 0, c1 = O(1), and φd > 1 satisfying
such that g(n) ∈ [c0 · φnd , c1 · φnd ].
φdd = 1 + φd + · · · + φd−1
d
Hence
(l−1)d+i
(l−1)d+i
f (l, i) = g (l − 1) · d + i ∈ [c0 · φd
, c1 · φd
].
Similar to the pruned witness tree of a symmetric witness tree, we use the same process in Definition 3.3 to
obtain the pruned asymmetric witness tree of an asymmetric witness tree.
log n
Vöcking in [Vöc03] showed that in a perfectly random hash function, the maximum load is log
d log φd +
O(1) with high probability given any n balls. We outline Vöcking’s argument for distinct balls here: let b
be a ball of height l + 4 for l = log logdn+log(1+c)
+ 1. Without loss of generality, we assume that b is in the
log φd
first group G1 . By the definition of the asymmetric witness tree, there exists a tree T in G1 with root b and
height l whose leaves have height at least 4. For each ball u and its ith ball v, the hash function h(i) satisfies
h(i) (u) = h(i) (v). Similar to (2), we apply a union bound on all possible witness trees of height l in this
configuration to bound the probability by
d
1
nf (l,1) · ( )f (l,1)−1 · ( d )number of leaves in f (l,1) ,
n
3
(l−1)d+1
which is less than n−c given f (l, 1) = Θ(φd
) = Θ (1 + c) log n .
We prove our derandomization of Vöcking’s argument here.
Theorem 6.3 For any m = O(n), any constants c > 1 and d ≥ 2, there exist a constant φd ∈ (1.61, 2)
and a hash family H in Construction 4.1 with O(log n log log n) random bits such that for any m balls in
U , with probability at least 1 − n−c , the max-load of the Always-Go-Left mechanism with d independent
log n
m
choices from H is log
d log φd + O(c + n ).
m
Proof. Let l be the smallest integer such that c0 φld
d ≥ 10(2 + 2c) log m and b = 10d · n + 1. We bound the
probability of a witness tree of height l+3c+1 whose leaves have height more than b in h(1) , . . . , h(d) during
the Always-Go-Left scheme. Notice that there is a ball of height l + b + 3c + 1 in any bin of G2 , G3 , . . . , Gd
indicates that there is a ball of the same height in G1 .
We choose the parameters of H as follows: kg = 20c · d · (l + b + 1 + 3c) = O(log log n), k =
log2 (log n/3 log log n), δ1 = 1/poly(n) such that Corollary 4.5 happens with probability at most n−c−1 ,
14
and the bias δ2 = log n−O(log n) of hk+1 later. We set hk+1 to be a hash function from U to [log3 n/d] and
g to be a function from U to [n/d] such that
(j)
(j)
(j)
(j)
h(j) = h1 ◦ h1 ◦ · · · ◦ hk ◦ hk+1 ⊕ g (j)
is a map from U to Gj of [n/d] bins for each j ∈ d.
We use h(1) , . . . , h(d) to denote d independent hash functions from H with the above parameters. We
(1)
(d)
use the notation of hi to denote the group of hash functions {hi , . . . , hi } in this proof. We assume
Corollary 4.5 and follow the same argument in the proof of Theorem 5.1. We bound the probability of
witness trees from 2 cases depending on the number of collisions.
Pruned witness trees with at least 3c collisions: Given a configuration C with at least 3c collisions, we
consider the first 3c collisions e1 , . . . , e3c in the BFS of C. Let C 0 be the induced subgraph of C that only
contains all vertices in e1 , . . . , e3c and their ancestors in C. Therefore C survives under h(1) , . . . , h(d) only
if C 0 survives under h(1) , . . . , h(d) .
0
Observe that |C 0 | ≤ 3c · 2 · d · height(T ) . There are at most m|T | possible instantiations of balls in
C 0 . For each instantiation T of C 0 , because kg ≥ 2 · number of edges = 2(|C 0 | + 3c − 1), we bound the
probability that any instantiation of C 0 survives in h by
d
d 0
d
0
0
0
m|C | · ( )number of edges = m|C | · ( )|C |+3c−1 ≤ (dm/n)|C | · ( )3c−1 ≤ n−2c .
n
n
n
At the same time, there are at most (|T |2 )3c = poly(log n) configurations of C 0 . Hence we bound the
probability of any witness with at least 3c collisions surviving by n−c .
Pruned witness tree with less than 3c collisions: We fix a configuration C of witness tree in group G1
with height l + 1 + 3c and less than 3c collisions. Thus |C| ∈ [f (l + 1, 1), f (l + 1 + 3c, 1)].
Let FT be the subset of possible asymmetric witness tree with configuration
C after fixing
the prefixes
h1 , h2 , . . . , hk . For any T ∈ FT , each edge (u, v) has to satisfy h(i) T (u) = h(i) T (v) in the AlwaysGo-Left scheme when v is the ith child of u. This indicates their prefixes are equal:
(i)
(i)
(i)
(i)
h1 T (u) ◦ · · · ◦ hk T (u) = h1 T (v) ◦ · · · ◦ hk T (v) .
From the same argument in the proof of Theorem 5.1, we bound
|FT | ≤ m · (1.01 log3 n ·
m |C|−1
)
n
under h1 , h2 , . . . , hk from Corollary 4.5.
We first consider hk+1 as a t-wise independent distribution from U to [log3 n/d] for t = 5bd · f (l + 3c +
1, 1) = O(log m) then move to δ2 -biased spaces. For each asymmetric witness tree, every edge (u, v) maps
to the same bin w.p. d/ log3 n in hk+1 .
For each leaf, its height is at least b if each bin in its choices has height at least b − 1, which happens
with probability at most
d
1.01·log3 n· m
n
b−1
(log3 n/d)b−1
≤
b−1
(1.01d · m
n)
(b − 1)!
15
!d
2
≤ 2−3d · (
n 2d
)
m
from the proof of Theorem 5.1.
Because these two types of events are on disjoint subsets of balls, the probability that any possible
asymmetric witness tree in FT exists in t-wise independent distributions over the suffixes is at most
|C|−1
(d−1)(|C|−3c)
n 2d
d
m |C| −3d2 n 2d |C|/3
d
−3d2
·
2
·
(
· 2
·( )
|FT | ·
)
≤m
·
1.01d
·
m
n
m
log3 n
≤m · 2−f (l+1,1) ≤ n−c−1 .
We choose δ2 = n−c−1 · (log3 n/d)−t /|FT | = (log n)−O(log n) such that in δ2 -biased spaces, any
possible asymmetric witness tree in FT exists hk+1 is at most happens with probability at most n−c−1 +
|FT | · δ2 · (log3 /d)bd·f (l+3c+1,1) ≤ 2n−c−1 . At the same time, the number of possible configurations is at
most (f (l + 3c + 1, 1)2 )3c ≤ 0.1n.
From all discussion above, with probability at most n−c , there exists a ball in the Always-Go-Left mechn log n
anism with height at least l + b + 3c + 1 = log
t
u
d log φd + O(1).
7
Heavy load
We consider the derandomization of the 1-choice scheme when we have m =√n·poly(log
p nn)balls and n bins.
From the Chernoff bound, w.h.p, the max-load among n bins is m
log
n
·
1
+
O(
n
m ) when we throw
m > n log n balls into n bins independently at random. We modify the hash function from √
[CRSW13]
p nwith
1
+
O(
log
n
·
proper parameters for m = poly(log n) · n balls and prove the max-load is still m
n
m) .
We assume m = loga n · n for a constant a ≥ 1 in the rest of this section.
Theorem 7.1 For any constant c > 0 and a ≥ 1, there exist a constant C and a hash function from U to
[n] generated by O(log n log log n) random bits such that for any m = loga n · n balls, with probability
at least 1 −√n−c , thepmax-load
of the n bins in the 1-choice scheme with the hash function h is at most
m
n
n 1 + C · log n ·
m .
n
We omit g in this section and change h1 , . . . , hk+1 with different parameters. We choose k = log (2a)log
log log n ,
−i
hi to denote a hash function from U to [n2 ] for i ∈ [k], and hk+1 to denote a hash function from U to
−k
[n2 ] = [log2a
p nthat h1 ◦ h2 ◦ · · · ◦ hk ◦ hk+1 constitute a hash function from U to [n]. We set
√ n] such
β = 4(c + 2) log n m
. For convenience, we still think h1 ◦ h2 ◦ · · · ◦ hi as a hash function maps to
−i
1−2
n
bins for any i ≤ k. In this section, we still use δ1 -biased spaces on h1 , . . . , hk and a δ2 -biased space
on hk+1 for δ1 = 1/poly(n) and δ2 = (log n)−O(log n) .
Claim 7.2 For any constant c > 0, there exists δ1 = 1/poly(n) such that given m =Q
loga n · n balls, with
−i
β
−c−1
1−2
probability 1 − n
, for any i ∈ [k] and any bin b ∈ [n
], there are less than j≤i (1 + (k+2−i)
2) ·
m
n
−i
· n2
balls in this bin.
Proof. We still use induction on i. The base case is i = 0. Because there are at most m balls, the hypothesis
is true.
Q
β
m 2−l
Suppose it is true for i = l. Now we fix a bin and assume there are s = j≤l (1 + (k+2−i)
≤
2) · n n
−l
−(l+1)
2
(1 + β) m
balls in this bin from the induction hypothesis. hl+1 maps these s balls to t = n2
bins.
nn
β
We will prove that with high probability, every bin in these t bins of hl+1 contains at most (1 + (k+1−l)2 )s/t
balls.
16
We use Xi ∈ {0, 1} to denote whether ball i is in one fixed bin of [t] or not. Hence Pr[Xi = 1] = 1/t.
Let Yi = Xi − E[Xi ]. Therefore E[Yi ] = 0 and E[|Yi |l ] ≤ 1/t for any l ≥ 2. Let b = β2l for a large
constant β later.
P
X
EDδ1 [( i Yi )b ]
β
Xi > (1 +
Pr [
)s/t] ≤
β
b
Dδ1
(k + 1 − l)2
( (k+1−l)
2 s/t)
i
P
2b
i1 ,...,ib EU [Yi1 · · · Yib ] + δ1 s
≤
β
b
( (k+1−l)
2 s/t)
≤
≤
2b b!(s/t)b/2 + δ1 s2b
β
b
( (k+1−l)
2 s/t)
!b/2
2b(s/t)
β
2
( (k+1−l)
2 s/t)
n
k
We use these bounds k = log (2l)log
log log n < log log n, b < β2 <
(m/n)2 to simplify the above bound by
2 log n
β2
(log log n)4
· s/t
β log n
(2l) log log n
+ δ1 · s2b
−l−1
and n2
k
≥ n2 ≥ log2l n ≥
b/2
+ δ1 s2b
!b/2
2 log2 n
≤
+ δ1 s2b
n
2−l−1 )
(log n · m
) · (m
n
n
b/2
1
+ δ1 s2b
≤
−l−1
0.5·2
n
l
2m 2−l 2β2
−0.5·2−l−1 ·β2l /2
≤n
+ δ1
n
≤ n−β/8 + δ1 · n6β .
n
Hence we choose the two parameters β > 8(c + 2) and δ1 = n−6β−c−2 such that the above probability is
bounded by 2n−c−2 . Finnally, we apply the union bound on i and all bins.
t
u
Proof of Theorem 7.1. We first apply Claim 7.2 to h1 , . . . , hk .
In hk+1 , we first consider it as a b = 16(c + 2)2 log n = O(log n)-wise independent distribution that
Q
−k
β
m 2−k
maps s < j≤k (1 + (k+2−i)
balls to t = n2 bins. From Lemma 2.6 and Theorem 5 (I)
2) · n n
2
in [SSS95], we bound the probability that one bin receives more than (1 + β)s/t by eβ ·E[s/t]/3 ≤ n−c−2
given b ≥ β 2 E[s/t].
2a
Then we choose δ2 = (log n)−b·5a = (log n)−O(log n) such that any δ2 -biased space from [2 m
n log n]
m
2a
2 log n
· (log2a n)b < n−c−2 -close to a b-wise independent distribution. Hence in
to [log2a n] is δ2 · n ≤b
hk+1 , with probability at most 2 · n−c−2 , there is one bin that receives more than (1 + β)s/t balls. Overall,
17
the number of balls in any bin of [n] is at most
Y
i≤k
(1 +
X
β
m
β
m
m
)(1 + β) ≤ (1 +
) ≤ (1 + 2β) .
2
2
(k + 2 − i)
n
(k + 2 − i) n
n
i≤k+1
t
u
Acknowledgement
The author is grateful to David Zuckerman for his constant support and encouragement, as well as for many
fruitful discussions. We thank Eric Price for introducing us to the multiple-choice schemes. We also thank
the anonymous referee for the detailed feedback and comments.
References
[ABKU99] Yossi Azar, Andrei Z. Broder, Anna R. Karlin, and Eli Upfal. Balanced allocations. SIAM J.
Comput., 29(1):180–200, September 1999.
[ADM+ 99] Noga Alon, Martin Dietzfelbinger, Peter Bro Miltersen, Erez Petrank, and Gábor Tardos. Linear hash functions. J. ACM, 46(5):667–683, September 1999.
[AGHP90] N. Alon, O. Goldreich, J. Hastad, and R. Peralta. Simple construction of almost k-wise independent random variables. In Proceedings of the 31st Annual Symposium on Foundations of
Computer Science. IEEE Computer Society, 1990.
[CRSW13] L. Elisa Celis, Omer Reingold, Gil Segev, and Udi Wieder. Balls and bins: Smaller hash
families and faster evaluation. SIAM J. Comput., 42(3):1030–1050, 2013.
[DETT10] Anindya De, Omid Etesami, Luca Trevisan, and Madhur Tulsiani. Improved pseudorandom
generators for depth 2 circuits. In RANDOM 2010, pages 504–517, 2010.
[DKRT16] Søren Dahlgaard, Mathias Bæk Tejs Knudsen, Eva Rotenberg, and Mikkel Thorup. The power
of two choices with simple tabulation. In Proceedings of the Twenty-Seventh Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2016, pages 1631–1642, 2016.
[DW03]
Martin Dietzfelbinger and Philipp Woelfel. Almost random graphs with simple hash functions.
In Proceedings of the Thirty-fifth Annual ACM Symposium on Theory of Computing, STOC ’03,
pages 629–638, 2003.
[GKM15]
Parikshit Gopalan, Daniek Kane, and Raghu Meka. Pseudorandomness via the discrete fourier
transform. In IEEE 56th Annual Symposium on Foundations of Computer Science, FOCS 2015,
pages 903–922, 2015.
[GMR+ 12] Parikshit Gopalan, Raghu Meka, Omer Reingold, Luca Trevisan, and Salil P. Vadhan. Better pseudorandom generators from milder pseudorandom restrictions. In 53rd Annual IEEE
Symposium on Foundations of Computer Science, FOCS 2012, pages 120–129, 2012.
[IMZ12]
Russel Impagliazzo, Raghu Meka, and David Zuckerman. Pseudorandomness from shrinkage.
In Proceedings of the 52st IEEE symposium on Foundations of Computer Science, 2012.
18
[Mit01]
Michael Mitzenmacher. The power of two choices in randomized load balancing. IEEE Trans.
Parallel Distrib. Syst., 12(10):1094–1104, 2001.
[MRRR14] Raghu Meka, Omer Reingold, Guy N. Rothblum, and Ron D. Rothblum. Fast pseudorandomness for independence and load balancing - (extended abstract). In Automata, Languages,
and Programming - 41st International Colloquium, ICALP 2014, Proceedings, pages 859–870,
2014.
[MRS00]
Michael Mitzenmacher, Andrea W. Richa, and Ramesh Sitaraman. The power of two random
choices: A survey of techniques and results. In in Handbook of Randomized Computing, pages
255–312, 2000.
[Nis92]
Noam Nisan. Pseudorandom generators for space-bounded computation. Combinatorica,
12(4):449–461, 1992.
[NN90]
J. Naor and M. Naor. Small-bias probability spaces: Efficient constructions and applications.
In Proceedings of the Twenty-second Annual ACM Symposium on Theory of Computing, STOC
’90, pages 213–223. ACM, 1990.
[NZ96]
Noam Nisan and David Zuckerman. Randomness is linear in space. J. Comput. Syst. Sci.,
52(1):43–52, 1996.
[PT12]
Mihai Pǎtraşcu and Mikkel Thorup.
59(3):14:1–14:50, June 2012.
[RRW14]
Omer Reingold, Ron D. Rothblum, and Udi Wieder. Pseudorandom graphs in data structures.
In Automata, Languages, and Programming - 41st International Colloquium, ICALP 2014,
Proceedings, pages 943–954, 2014.
[RSV13]
Omer Reingold, Thomas Steinke, and Salil P. Vadhan. Pseudorandomness for regular branching programs via fourier analysis. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques - 16th International Workshop, APPROX 2013, and 17th
International Workshop, RANDOM 2013, Berkeley, CA, USA, August 21-23, 2013. Proceedings, pages 655–670, 2013.
[Sie89]
Alan Siegel. On universal classes of fast high performance hash functions, their time-space
tradeoff, and their applications. In Proceedings of the 30th Annual Symposium on Foundations
of Computer Science, FOCS 1989, pages 20–25, 1989.
[SSS95]
Jeanette P. Schmidt, Alan Siegel, and Aravind Srinivasan. Chernoff-hoeffding bounds for applications with limited independence. SIAM J. Discret. Math., 8(2):223–250, May 1995.
[Sti94]
D. R. Stinson. Combinatorial techniques for universal hashing. Journal of Computer and
System Sciences, 48(2):337–346, 1994.
[TX13]
Luca Trevisan and Tongke Xue. A derandomized switching lemma and an improved derandomization of AC0. In Proceedings of the 28th Conference on Computational Complexity, pages
242–247, 2013.
[Vöc03]
Berthold Vöcking. How asymmetry helps load balancing. J. ACM, 50(4):568–589, July 2003.
The power of simple tabulation hashing.
19
J. ACM,
[Vaz86]
Umesh Vazirani. Randomness, adversaries and computation. In Ph.D. Thesis, EECS, UC
Berkeley, pages 458–463. 1986.
[Wie17]
Udi Wieder. Hashing, load balancing and multiple choice. Foundations and Trends in Theoretical Computer Science, 12(3-4):275–379, 2017.
[Woe06]
Philipp Woelfel. Asymmetric balanced allocation with simple hash functions. In Proceedings
of the Seventeenth Annual ACM-SIAM Symposium on Discrete Algorithm, SODA ’06, pages
424–433, 2006.
A
Proof of Lemma 4.4
We apply an induction from i = 0 to i = k. The base case i = 0 is true because there are at most m balls.
1
1− 1
Suppose it is true for i = l < d. For a fixed bin j ∈ [n 2l ], there are at most (1 + β)l n 2l · m
n balls. With
1
out loss of generality, we assume there are exactly s = (1 + β)l n 2l · m
n balls from the induction hypothesis.
1
Under the hash function hl+1 , we allocate these balls into t = n 2l+1 bins.
We fix one bin in hl+1 and prove that this bin receives at most
1
1
(1 + β)s/t = (1 + β) · (1 + β)l n 2l /n 2l+1 ·
1
m
m
= (1 + β)l+1 n 2l+1 ·
n
n
balls with probability ≤ 2n−c−2 in a log3 n-wise δ1 -biased space for δ1 = 1/poly(n) with
P a sufficiently
large polynomial. We use Xi ∈ {0, 1} to denote the ith ball is in the bin or not. Hence E[ i∈[s] Xi ] = s/t.
For convenience, we use Yi = Xi − E[Xi ]. Hence Yi = 1 − 1/t w.p. 1/t, o.w. Yi = −1/t. Notice that
E[Yi ] = 0 and | E[Yil ]| ≤ 1/t for any l ≥ 2.
P
We choose b = 2l · β = O(log n) for a large even number β and compute the bth moment of i∈[s] Yi
as follows.
X
X
Pr[|
Xi | > (1 + β)s/t] ≤ Eδ1 -biased [(
Yi )b ]/(βs/t)b
i
i∈[s]
P
EU [( i Yi )b ] + δ1 · s2b tb
≤
(βs/t)b
P
3b
i1 ,...,ib E[Yi1 · Yi2 · · · · · Yib ] + δ1 · s
≤
(βs/t)b
Pb/2 b−j−1 b!
· 2b/2 · sj (1/t)j + δ1 · s3b
j=1
j−1
≤
(βs/t)b
≤
2b/2 b! · (s/t)b/2 + δ1 · s3b
(βs/t)b
1
β log n
1/3 and β = (log n)−0.2 < (s/t)0.1 , we simplify
Because s/t ≥ n 2k ≥ log3 n, b ≤ β2k ≤ 3 log
log n ≤ (s/t)
it to
!b/2
2
b/2
2b · s/t
2(s/t)2/3 · s/t
3b
+ δ1 · s ≤
+ δ1 · s3b
(βs/t)2
(s/t)1.8
2
l
3
l
=(s/t)(−0.1)·b/2 + δ1 · s3b ≤ (n 2l+1 )−0.1·β2 + δ1 (n 2l )3β·2 = n−c−2 + δ1 n9β ≤ 2n−c−2 .
20
Finally we choose β = 40(c + 2) = O(1) and δ1 = n−9β−c−2 to finish the proof.
21
| 8cs.DS
|
Discovery Radiomics via Deep Multi-Column Radiomic Sequencers for Skin Cancer Detection
Mohammad Javad Shafiee
Alexander Wong
Abstract
arXiv:1709.08248v1 [cs.CV] 24 Sep 2017
While skin cancer is the most diagnosed form of cancer in men
and women, with more cases diagnosed each year than all other
cancers combined, sufficiently early diagnosis results in very good
prognosis and as such makes early detection crucial. While radiomics have shown considerable promise as a powerful diagnostic
tool for significantly improving oncological diagnostic accuracy and
efficiency, current radiomics-driven methods have largely rely on
pre-defined, hand-crafted quantitative features, which can greatly
limit the ability to fully characterize unique cancer phenotype that
distinguish it from healthy tissue. Recently, the notion of discovery
radiomics was introduced, where a large amount of custom, quantitative radiomic features are directly discovered from the wealth of
readily available medical imaging data. In this study, we present
a novel discovery radiomics framework for skin cancer detection,
where we leverage novel deep multi-column radiomic sequencers
for high-throughput discovery and extraction of a large amount of
custom radiomic features tailored for characterizing unique skin
cancer tissue phenotype. The discovered radiomic sequencer was
tested against 9,152 biopsy-proven clinical images comprising of
different skin cancers such as melanoma and basal cell carcinoma,
and demonstrated sensitivity and specificity of 91% and 75%, respectively, thus achieving dermatologist-level performance and
hence can be a powerful tool for assisting general practitioners
and dermatologists alike in improving the efficiency, consistency,
and accuracy of skin cancer diagnosis.
1
Introduction
Skin cancer is the most diagnosed form of cancer, with more new
cases of skin cancer diagnosed each year than all other forms of
cancers combined [14]. Furthermore, the annual cost for the treatment of skin cancer in the U.S. is estimated at $8.1 billion [7]. Fortunately, there are high chance of prognosis for various forms of skin
cancer given sufficiently early diagnosis, making early skin cancer
screening and detection crucial for patient recovery. A powerful diagnostic tool that has shown considerable promise for ushering in
a new era of imaging-driven quantitative personalized cancer decision support and management is the notion of radiomics [10],
which involves the high-throughput extraction and analysis of a
large number of quantitative features to characterize cancer tissue
traits to improve oncological diagnostic efficiency, consistency, and
accuracy.
Despite its considerable promise [11, 12], current radiomicsdriven methods have largely relied on predefined, hand-crafted
imaging-based feature models based on human notions of intensity, texture, and shape, and as such can greatly limit the ability to
fully characterize unique cancer phenotype that distinguish it from
healthy tissue. Recently, to alleviate the limitations of radiomics,
the notion of discovery radiomics was introduced [8, 9, 13, 16],
where a large amount of custom, quantitative radiomic features
are directly discovered from the wealth of readily available medical imaging data.
In this study, we present a novel discovery radiomics framework for skin cancer detection, where we leverage novel multicolumn deep radiomic sequencers for high-throughput discovery
and extraction of a large amount of custom radiomic features tailored for characterizing unique skin cancer tissue phenotype. Deep
neural networks have shown that they can learn effective and accurate feature extraction framework via convolutional layers. This
type of operations decrease the human model bias since they are
trained as an end-to-end systems. Due to this fact, the features extracted from deep neural networks (particularly convolutional neural networks) have showing promising results in different applications such as object classification [1, 2], object segmentation and
detection [3] and super-resolution [4].
The proposed framework has considerable potential for discovering a large amount of quantitative biomarkers beyond what clini-
University of Waterloo, ON, Canada
Elucid Labs, Canada
University of Waterloo, ON, Canada
Elucid Labs, Canada
cians can visually identify, thus making it a powerful tool for assisting general practitioners and dermatologists alike in improving the
efficiency, consistency, and accuracy of skin cancer diagnosis.
This paper is organized as follows; In the next section the notion of discovery radiomics and the deep multi-column neural network for the purpose of generating discovery radiomic sequences
is explained. Then the Results are demonstrated and the conclusion will be drawn at the end.
2
Methodology
The proposed discovery radiomics framework for skin cancer detection can be described as follows (see Figure 1). Given a wealth
of past clinical images and corresponding biopsy-verified diagnostic information from a skin imaging data archive, the radiomic sequencer discovery process discovers a radiomic sequencer that
can perform high-throughput extraction of radiomic sequences comprising of a large amount of highly customized, quantitative features tailored for characterizing unique skin cancer traits that are
particularly effective at differentiating between malignant and benign skin lesions. The discovered radiomic sequencer can then
be applied to clinical images of a new patient to extract the corresponding dermatological radiomic sequence for skin cancer screening and diagnosis purposes.
In this study, the radiomic sequencer being proposed in the discovery radiomics framework is a novel deep multi-column radiomic
sequencer, inspired by the work of Ciresan et al. [6]. More specifically, the proposed deep multi-column radiomic sequencer splits
information into multiple, parallel columns so that parallel streams
of increasingly more abstract skin cancer traits are discovered and
modeled based on a wealth of skin imaging data before being
merged into a single representation, thus allowing for improved
and more complete characterization of the complex physiological
characteristics of skin cancer.
The underlying architecture of the proposed deep multi-column
radiomic sequencer being discovered is shown in Figure 2. In
this architecture, low-level skin tissue characteristics are modeled
in the lower convolutional layers. However the layers are divided
into multiple parallel columns of deep convolutional layers that decompose into unique mid- to high-level skin tissue characteristics.
These parallel columns of deep convolutional layers are then merged
via a fully-connected layer to produce a final radiomic sequence for
the skin lesion being analyzed.
In this study, taking inspiration from [1], the realization of the
deep multi-column network architecture consists of two columns,
with each column consisting of five convolutional layers. The examined network architecture in this study is the combination of
two-column networks with each column consisting of 5 convolutional layers. The number of convolutional filters are increased and
the size of the convolutional filters are decreased as we go deeper
to improve the modeling performance of the deep, multi-column radiomic sequencer. This led the proposed radiomic sequencer to
possess the following architectural configuration for each column:
five convolutional layers with 96@ 7 × 7 filters, 256@ 5 × 5 filters,
384@ 3 × 3 filters, 384@ 3 × 3 filters, 256@ 5 × 5 filters, and one
fully connected layer with 8192 neurons. The output of two columns
are combined together to produce the final radiomic sequence.
2.1
Results and Discussion
In this study, the discovered deep multi-column radiomic sequencer
was tested against 9,152 biopsy-proven clinical images extracted
from ISIC 2017 Challenge dataset [17] (see Figure 3), with the
malignant cases comprising of 90% melanoma cases and 10%
basal cell carcinoma cases. Sequencer validation was performed
by splitting the clinical images into 10% training (473 randomly selected samples for each class label) and 90% testing. The testing
dataset comprises of 8,049 benign and 157 malignant cases. To
Radiomics Sequencer
Discovery
Skin Imaging Data
(Current Patient)
Discovered Radiomics
Sequencer
Radiomic Seqence
Skin Image Data Archives
Fig. 1: Overview of the proposed discovery radiomics framework for skin cancer detection. A custom radiomic sequencer is discovered
via past skin imaging data; for new patients, radiomic sequences of custom radiomic features are generated for skin cancer quantification and analysis. In this study, we leverage a deep multi-column radiomic sequencer to characterize and model unique skin cancer
phenotype.
high-throughput discovery and extraction of a large amount of custom radiomic features tailored for characterizing unique skin cancer
tissue phenotype. The deep multi-column architecture used in the
proposed radiomic sequencer can significantly boost the modeling
power of the sequencer. The discovered deep multi-column radiomic sequencer can then be applied to clinical images of a new
patient to extract the corresponding dermatological radiomic sequence for skin cancer screening and diagnosis purposes.
The proposed framework is examined on ISIC skin cancer chaltrue positive
(1) lenge, with the deep multi-column radiomic sequencer discovered
sensitivity =
positive
using a small balanced dataset of malignant and benign cases
where only 473 cases were utilized for each class label. Quantitative evaluation of the discovered radiomic sequencer was then
true negative
specificity =
.
(2) performed using an unbalanced dataset with 8,049 benign cases
negative
and 157 malignant cases. Experimental results demonstrated that
the proposed discovery radiomics approach for skin cancer modelwhere true positive is the number of malignant cases which are ing and classification is able to achieving sensitivity and specificity
classified correctly by the proposed approach and true negative is of 91% and 75%, respectively. These promising results show the
the number of benign cases classified as negative by the proposed applicability and modeling power of the proposed approach and
approach.
illustrates that it can be a powerful tool for assisting general practiExperimental results show that the proposed radiomic sequencer tioners and dermatologists alike in improving the efficiency, consisachieved sensitivity and specificity of 91% and 75%, respectively. tency, and accuracy of skin cancer diagnosis.
As a point of reference, in the study by Wells et al. [15], it was
found that the dermatologists in the study had a sensitivity and
specificity of 80% and 43%, respectively for melanoma screen- Acknowledgments
ing, while MelaFind, a non-invasive light-based tool for skin cancer
screening, had a sensitivity and specificity of 96% and 8%, respec- This work was supported by Elucid Labs. The authors also thank
tively. Furthermore, in a study by Esteva et al. [5], it was found Nvidia for the GPU hardware used in this study through the Nvidia
that using an Inception v3 convolutional neural network trained for Hardware Grant Program.
distinguishing between benign skin lesions, malignant lesions, and
non-neoplastic lesions achieved an accuracy of 72.1 ± 0.9% on
dermatologist-labeled clinical images. As such, these experimental References
results show that the proposed radiomic sequencer for skin cancer
detection is able to achieve dermatologist-level performance and
[1] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classihence can be a powerful tool for assisting general practitioners
fication with deep convolutional neural networks Advances in
and dermatologists alike in improving the efficiency, consistency,
neural information processing systems (NIPS)(2012).
and accuracy of skin cancer diagnosis.
quantitatively evaluate the efficacy of the discovery radiomic sequencer for cancer detection, the radiomic sequences produced
using the sequencer are then fed into a fully-connected feed-foward
neural network with two fully-connected layers and a softmax layer.
The efficacy of the proposed discovery radiomics framework is
examined quantitatively using two performance metrics: i) sensitivity, and ii) specificity; where
3
Conclusion
In this paper we proposed a new discovery radiomics approach for
the purpose of skin cancer modeling and classification. A deep
multi-column radiomic sequencer is proposed and discovered for
[2] K. Simonyan, and A. Zisserman. Very deep convolutional
networks for large-scale image recognition arXiv preprint
arXiv:1409.1556(2014).
[3] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2015).
Radiomics Sequence
FC
Pooling
Relu Activation
Conv Stack
Fig. 2: Overview of the proposed deep multi-column radiomic sequencer. The network architecture of the radiomic sequencer is
inspired by [1]. The network architecture is a multi-column network comprising of two individual columns of convolutional, activation,
and pooling layers. Each column of the deep multi-column radiomic sequencer is composed of 5 convolutional layers, with the number
of convolutional filters are increased and the size of the convolutional filters are decreased as we go deeper to improve the modeling
performance of the deep multi-column radiomic sequencer.
proven computed tomography lung cancer prediction. International Conference Image Analysis and Recognition, 2017.
[10] P. Lambin, E. Rios-Velazquez, R. Leijenaar, S. Carvalho, R. G.
van Stiphout, P. Granton, C. M. Zegers, R. Gillies, R. Boellard,
and A. D. et al. Radiomics: extracting more information from
medical images using advanced feature analysis. European
Journal of Cancer, 2012.
[11] A. Cameron, F. Khalvati, M. Haider, and A. Wong. MAPS: A
Quantitative Radiomics Approach for Prostate Cancer Detection. IEEE Transactions on Biomedical Engineering, (2016).
[12] R. Amelard, J. Glaister, A. Wong, and D. Clausi. High-level
intuitive features (HLIFs) for intuitive skin lesion description.
IEEE Transactions on Biomedical Engineering, (2015).
[13] M. J. Shafiee, A. G. Chung, D. Kumar, F. Khalvati, M. Haider,
and A. Wong. Discovery radiomics via stochasticnet sequencers for cancer detection. NIPS Workshop on Machine
Learning in Healthcare, (2015).
Fig. 3: Examples of biopsy-proven clinical images of skin lesions
from [17]
[14] R. Stern. Prevalence of a history of skin cancer in 2007: results of an incidence-based model. Arch Dermatol, (2010).
[4] C. Dong, C. Loy, K. He and X. Tang. Learning a deep convo- [15] R. Wells, D. Gutkowicz-Krusin, and e. a. E. Veledar. Comparlutional network for image super-resolution European Conferison of diagnostic and management sensitivity to melanoma
ence on Computer Vision(2014).
between dermatologists and melafind: A pilot study. International Conference Image Analysis and Recognition, (2012).
[5] A. Esteva, and B. Kuprel Dermatologist-level classification of
skin cancer with deep neural networks. Nature, (2016).
[16] A. Wong, A. G. Chung, D. Kumar, M. J. Shafiee, F. Khalvati,
and M. Haider. Discovery radiomics for imaging-driven quanti[6] D. Ciresan, U. Meier, and J. Schmidhuber. Multicolumn deep
tative personalized cancer decision support. Journal of Compneural networks for image classification. IEEE Conference on
tutational Vision and Imaging Systems, (2015).
Computer Vision and Pattern Recognition,(2012).
[17] D. Gutman, N. Codella, E. Celebi, B. Helba, M. Marchetti,N.
[7] G. GP, M. SR, E. DU, and Y. KR. Prevalence and costs of skin
Mishra, Nabin and A. Halpern. Skin lesion analysis toward
cancer treatment in the u.s., 2002-2006 and 2007-2011. Am
melanoma detection: A challenge at the international symJ Prev Med, 2014.
posium on biomedical imaging (ISBI) 2016, hosted by the
international skin imaging collaboration (ISIC)arXiv preprint
[8] A.-H. Karimi, A. G. Chung, M. J. Shafiee, F. Khalvati, M. A.
arXiv:1605.01397(2016).
Haider, A. Ghodsi,and A. Wong. Discovery radiomics via a
mixture of deep convnet sequencers for multi-parametric mri
prostate cancer classification. International Conference Image Analysis and Recognition, (2017).
[9] D. Kumar, A. G. Chung, M. J. Shafiee, F. Khalvati, M. A.
Haider, and A. Wong. Discovery radiomics for pathologically-
| 1cs.CV
|
Intensionality, Intensional Recursion,
and the Gödel-Löb axiom
G. A. Kavvos
arXiv:1703.01288v1 [cs.PL] 3 Mar 2017
Department of Computer Science
University of Oxford
Oxford, United Kingdom
[email protected]
The use of a necessity-like modality in a typed λ -calculus can be used as a device for separating
the calculus in two separate regions. These can be thought of as intensional vs. extensional data:
data in the first region, the modal one, are available as code, and their description can be examined,
whereas data in the second region are only available as values up to ordinary equality. This allows us
to add seemingly non-functional operations at modal types, whilst maintaining consistency. In this
setting the Gödel-Löb axiom acquires a novel constructive reading: it affords the programmer the
possibility of a very strong kind of recursion, by enabling him to write programs that have access to
their own code. This is a type of computational reflection that is strongly reminiscent of Kleene’s
Second Recursion Theorem. We prove that it is consistent with the rest of the system.
1
Introduction
This paper is about putting a logical twist on two old pieces of programming lore:
• First, it is about using modal types to treat programs-as-data in a type-safe manner.
• Second, it is about noticing that—in the context of intensional programming—a constructive reading of the Gödel-Löb axiom, i.e. (A → A) → A, amounts to a strange kind of recursion,
namely intensional recursion.
We will introduce a typed λ -calculus with modal types that supports both of these features. We will call
it Intensional PCF, after the simply-typed λ -calculus with fixed points studied by Scott [21] and Plotkin
[20].
1.1 Intensionality and Programs-as-data
To begin, we want to discuss our notion of programs-as-data. We mean it in a way that is considerably
stronger than the higher-order functional programming with which we are already familiar, i.e. ‘functions
as first-class citizens.’ In addition to that, our notion hints at a kind of homoiconicity, similar to the one
present in the L ISP family of languages. It refers to the ability given to a programmer to quote code,
and carry it around as a datum; see [4] for an instance of that in L ISP. This ability can be used for
metaprogramming, which is the activity of writing programs that write other programs; indeed, this is
what L ISP macros excel at [10], and this is what the metaprogramming community has been studying
for a long time: see [24, 26].
But we want go even further. In L ISP a program is abel to process code by treating it as mere symbols,
thereby disregarding its function and behaviour. This is what we call intensionality: an operation is
intensional if it is finer than equality. Hence, it amounts to a kind of non-functional behaviour.
To our knowledge, this paper presents the first sound, type-safe attempt at intensional programming.
Submitted to:
c G. A. Kavvos
This work is licensed under the
Creative Commons Attribution License.
Intensionality, Intensional Recursion, and the Gödel-Löb axiom
2
1.2 Intensional Recursion
We also want to briefly explain what we mean by intensional recursion—a fuller discussion may be
found in [13, 1]. Most modern programming languages support extensional recursion: in the body of a
function definition, the programmer may freely make a finite number of calls to the definiendum itself.
Operationally, this leads a function to examine its own values at a finite set of points at which it has—
hopefully—already been defined. In the untyped λ -calculus, this is modelled by the First Recursion
Theorem (FRT) [3, §2.1, §6.1]:
Theorem 1 (First Recursion Theorem). ∀ f ∈ Λ. ∃u ∈ Λ. u = f u.
However, as Abramsky [1] notes, in the intensional paradigm we have described above, a stronger
kind of recursion is imaginable. Instead of merely examining the result of a finite number of recursive
calls, the definiendum can recursively have access to a full copy of its own source code. This is embodied
in Kleene’s Second Recursion Theorem (SRT) [16]. Here is a version of the SRT in the untyped λ calculus, where VuW means ‘the Gödel number of the term u’ [3, §6.5].
Theorem 2 (Second Recursion Theorem). ∀ f ∈ Λ. ∃u ∈ Λ. u = f VuW.
Kleene also proved the following, where Λ0 is the set of closed λ -terms:
Theorem 3 (Existence of Interpreter). ∃E ∈ Λ0 . ∀M ∈ Λ0 . E VMW → M
As an interpreter exists, it is not hard to see that the SRT implies the FRT. It is not at all evident whether
we can go the other way around. This is because the SRT is a first-order theorem that is about diagonalization and code, whereas the FRT really is about higher types. See the discussion in [13].
It thus follows that—in the presence of intensional operations—the SRT affords us with a much
stronger kind of recursion. In fact, it allows for a certain kind of computational reflection, or reflective
programming, of the same kind envisaged by Brian Cantwell Smith [22]. But the programme of Smith’s
reflective tower involved a rather mysterious construction with unclear semantics [8, 29, 6], eventually
leading to a theorem that—even in the presence of a mild reflective construct, the so-called fexpr—
observational equivalence of programs collapses to α -conversion [28].
We will use modalities to stop intension from flowing back into extension, so the theorem of [29]
(which involves unrestricted quoting) will not apply. We will achieve reflection by internalising the SRT,
and the key observation for doing so is the following. Suppose that our terms are typed, and that u : A.
Suppose as well that there is a type constructor, , so that A means ‘code of type A.’ Then certainly
VuW : A, and f is forced to have type A → A. A logical reading of the SRT is then the following:
for every f : A → A, there exists a u : A such that u = f V f W. This corresponds to Löb’s rule from
provability logic [5]:
A → A
A
which is equivalent to adding the Gödel-Löb axiom to the logic. In fact, the punchline of this paper is
that the type of the Second Recursion Theorem is the Gödel-Löb axiom of provability logic.
To obtain reflective features, then, we will add a version of Löb’s rule to our modal λ -calculus. It
will be equivalent to the above, but proof-theoretically well-behaved—see [27, 14]:
A → A
A
To our knowledge, this paper presents the first sound, type-safe attempt at reflective programming.
G. A. Kavvos
3
1.3 Prospectus
In §2 we will introduce the syntax of iPCF, and in §3 we will show that it satisfies basic metatheoretic
properties. Follows that, in section §4, we show how to add intensional operations to iPCF. By proving
that the resulting notion of reduction is confluent, we obtain consistency for the system. We then look at
the computational behaviour of some important terms in §5, and conclude with two key examples of the
new powerful features of our language in §6.
2
Introducing Intensional PCF
Intensional PCF (iPCF) is a typed λ -calculus with modal types. As discussed before, the modal types
work in our favour by separating intension from extension, so that the latter does not leak into the former.
Given the logical nature of the categorical constructs used in our previous work on intensionality [15], we
shall model the types of iPCF after the constructive modal logic S4, in the dual-context style pioneered
by Pfenning and Davies [19, 7]. Let us seize this opportunity to remark that (a) there are also other ways
to capture S4, for which see the survey [13], and that (b) dual-context formulations are not by any means
limited to S4: they began in the context of intuitionistic linear logic, but have recently been shown to
also encompass other modal logics; see [14].
iPCF is not related to the language Mini-ML that is introduced by [7]: that is a call-by-value, MLlike language, with ordinary call-by-value fixpoints. In contrast, ours is a call-by-name language, with
new kind of fixpoints, namely intensional fixpoints. These fixpoints will afford the programmer the full
power of intensional recursion. Logically, they correspond to throwing the Gödel-Löb axiom, namely
(A → A) → A into S4. Modal logicians might object to this, as—in conjuction with the T axiom
A → A—it will make every type inhabited. We remind them that a similar situation occurs in PCF,
where the YA : (A → A) → A combinator allows one to write a term YA (λ x:A. x) : A for any A. As in the
study of PCF, we mostly care about the underlying computation, so it is the morphisms that matter, not
the objects.
The syntax and the typing rules of iPCF may be found in Figure 1. These are largely the same
as Pfenning and Davies’ DS4, save the addition of some constants (drawn from PCF), and a rule for
intensional recursion. The introduction rule for the modality restricts terms under a box (−) to those
containing only modal variables, i.e. variables carrying only intensions or code, but never ‘live values:’
∆;· ⊢ M : A
∆ ; Γ ⊢ box M : A
There is also a rule for intensional recursion:
∆ ; z : A ⊢ M : A
∆ ; Γ ⊢ fix z in box M : A
This will be coupled the reduction fix z in box M −→ box M [fix z in box M/z]. This rule is actually a
version of Löb’s rule, and including it in the Hilbert system of a (classical or intuitionistic) modal logic
is equivalent to including the Gödel-Löb axiom: see [5] and [27]. In fact, we have derived this fixpoint
rule through general considerations of sequent calculi for GL; see [14]. Finally, let us record a fact noticed
by Samson Abramsky, which is that erasing the modality from the types appearing in either Löb’s rule or
the Gödel-Löb axiom yields the type of YA : (A → A) → A, either as a rule, or axiomatically internalised
as a constant (both variants exist in the literature, see [11] and [18].)
Intensionality, Intensional Recursion, and the Gödel-Löb axiom
4
Figure 1: Syntax and Typing Rules for Intensional PCF
::=
Nat | Bool
Types A, B
::=
G | A → B | A
Terms M, N
::=
x | λ x:A. M | MN | box M | let box u ⇐ M in N |
Ground Types G
nb | true | false | succ | pred | zero? | ⊃G | fix z in box M
Canonical Forms V
Γ, ∆
Contexts
::=
nb | true | false | λ x:A. M | box M
::=
· | Γ, x : A
(b ∈ {true, false})
∆ ; Γ ⊢ nb : Nat
∆ ; Γ ⊢ b : Bool
∆ ; Γ ⊢ zero? : Nat → Bool
∆ ; Γ ⊢ f : Nat → Nat
( f ∈ {succ, pred})
∆ ; Γ ⊢ ⊃G : Bool → G → G → G
∆ ; Γ, x:A, Γ′ ⊢ x : A
(var)
∆ ; Γ, x:A ⊢ M : B
∆ ; Γ ⊢ λ x:A. M : A → B
∆;· ⊢ M : A
∆ ; Γ ⊢ box M : A
∆, u:A, ∆′ ; Γ ⊢ u : A
(→ I )
(I )
(var)
∆;Γ ⊢ M : A → B ∆;Γ ⊢ N : A
∆ ; Γ ⊢ MN : B
∆ ; Γ ⊢ M : A
∆, u:A ; Γ ⊢ N : C
∆ ; Γ ⊢ let box u ⇐ M in N : C
∆ ; z : A ⊢ M : A
∆ ; Γ ⊢ fix z in box M : A
(→ E )
(E )
G. A. Kavvos
3
5
Metatheory
iPCF satisfies the expected basic metatheoretic properties, namely structural and cut rules are admissible.
This is no surprise given its origin in the well behaved Davies-Pfenning DS4. We assume the typical conventions for λ -calculi: terms are identified up to α -equivalence, for which we write ≡, and substitution
[·/·] is defined in the ordinary, capture-avoiding manner. Bear in mind that we consider occurences of u
in N to be bound in let box u ⇐ M in N. Contexts Γ, ∆ are lists of type assignments x : A. Furthermore,
we shall assume that whenever we write a judgment like ∆ ; Γ ⊢ M : A, then ∆ and Γ are disjoint, in the
def
sense that VARS (∆) ∩ VARS (Γ) = 0,
/ where VARS (x1 : A1 , . . . , xn : An ) = {x1 , . . . , xn }. We write Γ, Γ′ for
the concatenation of disjoint contexts. Finally, we sometimes write ⊢ M : A whenever · ; · ⊢ M : A.
Theorem 4 (Structural & Cut). The following rules are admissible in the typing system of iPCF:
1. (Weakening)
3. (Contraction)
∆ ; Γ, Γ′ ⊢ M : A
∆ ; Γ, x:A, y:A, Γ′ ⊢ M : A
∆ ; Γ, x:A, Γ′ ⊢ M : A
∆ ; Γ, w:A, Γ′ ⊢ M[w, w/x, y] : A
2. (Exchange)
4. (Cut)
∆ ; Γ, x:A, y:B, Γ′ ⊢ M : C
∆ ; Γ, x:A, Γ′ ⊢ M : A
∆;Γ ⊢ N : A
∆ ; Γ, Γ′ ⊢ M[N/x] : A
∆ ; Γ, y:B, x:A, Γ′ ⊢ M : C
Theorem 5 (Modal Structural & Cut). The following rules are admissible:
1. (Modal Weakening)
3. (Modal Contraction)
∆, ∆′ ; Γ ⊢ M : C
∆, x:A, y:A, ∆′ ; Γ ⊢ M : C
∆, u:A, ∆′ ; Γ ⊢ M : C
∆, w:A, ∆′ ; Γ ⊢ M[w, w/x, y] : C
2. (Modal Exchange)
4. (Modal Cut)
∆, x:A, y:B, ∆′ ; Γ ⊢ M : C
∆, y:B, x:A, ∆′ ; Γ ⊢ M : C
4
∆;· ⊢ N : A
∆, u:A, ∆′ ; Γ ⊢ M : C
∆, ∆′ ; Γ ⊢ M[N/u] : C
Consistency of Intensional Operations
In this section we shall prove that the modal types of iPCF allow us to consistently add ‘intensional
operations’ on the modal types. These are non-functional operations on terms, which are not ordinarily
definable, because they violate equality. However, we will show that our use of modal types enables
the separation of intension from extension. We shall do this by adding intensional operations to iPCF,
introducing a notion of reduction, and proving it confluent. A known corollary of confluence is that the
equational theory induced by reduction is consistent, i.e. it does not equate all terms.
There is one very serious caveat, involving extension flowing into intension. That is, we need to
exclude from consideration terms where a variable bound by a λ occurs under the scope of a box (−)
construct. These will never be well-typed, but—since we discuss types and reduction orthogonally—we
also need to explicitly exclude them here too.
Intensionality, Intensional Recursion, and the Gödel-Löb axiom
6
4.1 Adding intensionality
Davies and Pfenning [19] suggested that the modal type can be used to signify intensionality. In fact,
in their previous paper [7] they had prevented reductions from happening under box (−) construct, “
[...] since this would violate its intensional nature.” But the truth is that neither of these presentations
included any genuinely non-functional operations at modal types, and hence their only use was for staged
metaprogramming.
Adding intensional, non-functional operations is a much more difficult task. Intensional operations
are dependent on descriptions and intensions rather than values and extensions. Hence, unlike reduction
and evaluation, they cannot be blind to substitution. This is the most challenging aspect to the whole
endeavour.
The task was recently taken up by Gabbay and Nanevski [9], who attempted to add a construct is-app
to the system of Davies and Pfenning, with the reduction rules
is-app (PQ) −→ true
is-app M −→ false
if M is not an application term
They tried to justify this addition in terms of a denotational semantics for modal types, in which the
semantic domain JAK directly involves the actual closed syntax of type A. But something seems to
have gone wrong with substitution. In fact, we believe that their proof of soundness is wrong: it is not
hard to see that their semantics is not stable under the second of those two reductions: take M to be u,
and let the semantic environment map u to an application box PQ. We can also see this in the fact that
their notion of reduction is not confluent; here is the relevant counterexample: we can reduce
let box u ⇐ box (PQ) in is-app (box u) −→ is-app (box PQ) −→ true
but we we can also reduce as follows:
let box u ⇐ box (PQ) in is-app (box u) −→ let box u ⇐ box (PQ) in false −→ false
This is easily discoverable if one tries to plough through a proof of confluence: it is very clearly not
the case that M → N implies M[P/u] → N[P/u] if u is under a box (−) and operations made of the same
ilk as is-app. Perhaps the following idea has the makings of a workable proposal: let us limit intensional
operations to a chosen set of functions f : T (A) → T (B) from terms of type A to terms of type B,
and then represent them in the language by a constant f˜, such that f˜(box M) → box f (M). This set of
functions would then be chosen so that they satisfy some sanity conditions. Since we want to have a let
construct that allows us to substitute code variables for code, the following general situation will occur:
if N → N ′ , we have
let box u ⇐ box M in N −→ N[M/u]
and
let box u ⇐ box M in N −→ let box u ⇐ box M in N ′ −→ N ′ [M/u]
Thus, in order to have confluence, we need N[M/u] → N ′ [M/u]. This will only be the case for
reductions of the form f˜(box M) → box f (M) if f (N[M/u]) ≡ ( f (N)) [M/u], i.e. if f is substitutive. But
then a simple naturality argument gives that f (N) ≡ f (u[N/u]) ≡ ( f (u)) [N/u], and hence f˜ is already
definable by λ x:A. let box u ⇐ x in box f (u), so such a ‘substitutive’ function is neither intensional nor
non-functional after all.
G. A. Kavvos
7
In fact, the only truly intensional operations we can add to our calculus will be those acting on
closed terms. We will see that this circumvents the problems that arise when intensionality interacts with
substitution. So we will limit intensional operations to the following set:
Definition 1 (Intensional operations). Let T (A) be the set of terms M such that · ; · ⊢ M : A. Let F (A, B)
be the set of all functions f : T (A) → T (B).
We will include all of these intensional operations f : T (A) → T (B) in our calculus, as constants:
∆ ; Γ ⊢ f˜ : A → B
with reduction rule f˜(box M) → box f (M), under the proviso that M is closed. Of course, this also
includes operations on terms that might in some sense not be computable. However, we are interested
in proving consistency in the most general setting. The questions of which intensional operations are
computable, and which primitives can and should be used to express them, are both still open.
4.2 Reduction and Confluence
We introduce a notion of β -reduction for iPCF, which we present in Figure 2. Unlike most studies
of PCF-like languages, we do not consider a reduction strategy at first, but simply ordinary, ‘nondeterministic’ β -reduction. We do so because are trying to show consistency.
The equational theory induced by this notion of reduction is a symmetric version of it, annotated
with types. It is easy to write down, so we omit it. Note that fact that, like in the work of Davies and
Pfenning, we do not include the congruence rule for the modality:
∆;· ⊢ M = N : A
∆ ; Γ ⊢ box M = box N : A
(cong)
In fact, the very absence of this rule is what will allow modal types to become intensional. Otherwise, the
only new rules are (a) intensional recursion, embodied in the rule (fix), and (b) intensional operations,
exemplified by the rule (int).
We can now show that
Theorem 6. The reduction relation −→ is confluent.
The easiest route to that theorem is to use a proof like that in [14], i.e. the method of parallel
reduction. This kind of proof was originally discovered by Tait and Martin-Löf, and is nicely documented
in [25]. Beause of the intensional nature of our box (−) constructs, ours will be more nuanced and fiddly.
We omit the details, for want of space (but see the appendix).
5
Some important terms
Let us look at the kinds of terms we can write in iPCF.
From the axioms of S4 First, we can write a term corresponding to axiom K, the normality axiom:
def
axK = λ f : (A → B). λ x : A. let box g ⇐ f in let box y ⇐ x in box (g y)
It is easy to see that ⊢ axK : (A → B) → (A → B). An intensional reading of this is as follows:
any function given as code can be transformed into an effective operation, mapping code to code.
Intensionality, Intensional Recursion, and the Gödel-Löb axiom
8
Figure 2: Beta reduction for Intensional PCF
(λ x:A. M)N −→ M[N/x]
M −→ N
MP −→ NP
M −→ N
(−→ β )
λ x:A. M −→ λ x:A. N
P −→ Q
(app1 )
MP −→ MQ
let box u ⇐ box M in N −→ N[M/u]
f˜(box M) −→ box f (M)
let box u ⇐ M in P −→ let box u ⇐ N in P
P −→ Q
let box u ⇐ M in P −→ let box u ⇐ M in Q
(zero?1 )
succ nb −→ n[
+1
(succ)
⊃G true M N −→ M
(⊃1 )
(fix)
(int)
M −→ N
zero? b
0 −→ true
(app2 )
(β )
fix z in box M −→ box M[fix z in box M/z]
M closed, f (M) ↓
(congλ )
(let-cong1 )
(let-cong2 )
zero? n[
+ 1 −→ false
·
[
pred nb −→ n − 1
(zero?2 )
(pred)
⊃G false M N −→ N
(⊃2 )
G. A. Kavvos
9
The rest of the axioms correspond to evaluating and quoting, as discussed before. Axiom T takes
code to value, or intension to extension:
def
⊢ evalA = λ x : A. let box y ⇐ x in y : A → A
and axiom 4 quotes code into code-for-code:
def
⊢ quoteA = λ x : A. let box y ⇐ x in box (box y) : A → A
The Gödel-Löb axiom: intensional fixed points Since the (fix) rule is a variant of Löb’s rule, we
expect to already have a term corresponding to the Gödel-Löb axiom of provability logic. We
do—and it is an intensional fixed-point combinator:
def
YA = λ x : (A → A). let box f ⇐ x in (fix z in box ( f z))
and ⊢ YA : ( → A) → A. We observe that
YA (box M) −→ ∗ fix z in box M z −→ box M(fix z in box M z)
Undefined The combination of eval and YA lead to non-termination, in a style reminiscent of the term
(λ x. xx)(λ x. xx) of the untyped λ -calculus. Let
def
ΩA = fix z in box (evalA z)
Then ⊢ ΩA : A, and ΩA −→ box (evalA ΩA ), so that evalA ΩA −→ ∗ evalA ΩA .
Extensional Fixed Points Perhaps surprisingly, the ordinary PCF Y combinator is also definable in the
iPCF. Let
def
UA = fix z in box (λ f : A → A. f (eval z f ))
Then ⊢ UA : ((A → A) → A), and we can define
def
YA = eval(A→A)→A UA
with the result that YA f −→ ∗ f (YA f ).
6
Two intensional examples
No description of an intensional language with intensional recursion would be complete without an
example of both of these features. Our first example is about intensionality, and is drawn from the
study of PCF and issues related to sequential vs. parallel (but not concurrent) computation. Our second
example is slightly more adventurous: it is a computer virus.
Intensionality, Intensional Recursion, and the Gödel-Löb axiom
10
6.1 ‘Parallel or’ by dovetailing
In [20], Gordon Plotkin proved the following theorem: there is no term por : Bool → Bool → Bool of PCF
such that por true M ։β true and por M true ։β true for any ⊢ M : Bool, whereas por false false ։β
false. Intuitively, the problem is that por has to first examine one of its two arguments, and this can be
troublesome if that argument is non-terminating. It follows that the parallel or function is not definable
PCF . In order to regain the property of so-called full abstraction for the Scott model of PCF, a constant
denoting this function has to be manually added to PCF, and endowed with the above rather clunky
operational semantics. See [20, 11, 18, 23].
However, the parallel or function is a perfectly computable partial recursive functional [23, 17].
The way to prove that is informally the following: given two closed terms M, N : Bool, take turns in
β -reducing each one for a few steps. This is called dovetailing. If at any point one of the two terms
reduces to true, then output true. But if at any point both reduce to false then output false.
This procedure is not definable in PCF, because a candidate term por does not have access to a code
for its argument, for it can only inspect its value. But in iPCF, we can use the modality to have access
to code, and intensional operations to implement reduction. Suppose we pick a reduction strategy −→ r .
Then, let us include a constant tick : Bool → Bool that implements one step of this reduction strategy
on closed terms:
M −→r N
tick (box M) −→ box N
Also, let us include a constant done? : Bool → Bool, which tells us if a closed term under a box is a
normal form:
M normal
M not normal
done? (box M) −→ true
done? (box M) −→ false
These two can be subsumed under our previous scheme for introducing intensional operations. The
above argument is now implemented by the following term:
por :≡ Y(λ por. λ x : Bool. λ y : Bool.
⊃Bool (done? x)
(lor (eval x)(eval y))
(⊃Bool (done? y)
(ror (eval x)(eval y))
(por (tick x)(tick y)))
where lor, ror : Bool → Bool → Bool are terms defining the left-strict and right-strict versions of the
‘or’ connective respectively. Notice that the type of this term of Bool → Bool → Bool: we require
intensional access to the terms of boolean type in order to define this computation.
6.2 A computer virus
Abstract computer virology is the study of formalisms that model computer viruses. There are many
ways to formalise viruses. We will use the model of Adleman [2], where files can be interpreted either
as data, or as functions. We introduce a data type F, and two constants
in : (F → F) → F
and
out : F → (F → F)
such that out (in M) → M, making (F → F) a retract of F. This might seem the same as the situation
where F → F is a retract of F, which yields models of the (untyped) λ -calculus and is difficult to
G. A. Kavvos
11
construct [3, §5.4], but it is not nearly as worrying: (F → F) is populated by intensions and codes,
not by actual functions. Under this interpretation, the pair (in, out) corresponds to a kind of Gödel
numbering—especially if F is N.
Now, in Adleman’s model, a virus is a given by its infected form, which either injures, infects, or
imitates other programs. The details are unimportant in the present discussion, save from the fact that
one can construct such a virus from its infection routine by using Kleene’s SRT. Let us model it by a term
infect : (F → F) → F → F, which accepts a piece of viral code and a file, and it returns either the file
itself, or a version infected with the viral code. We can then define a term virus : (F → F):
def
virus = fix z in box infect z
so that virus −→ ∗ box (infect virus). We then have that
eval virus −→ ∗ infect virus
which is a program that is ready to infect its input with its own code.
7
Conclusion
We have achieved the desideratum of an intensional programming language, with intensional recursion.
There are two main questions that result from this development.
Firstly, does there exist a good set of intensional primitives from which all others are definable? Is
there perhaps more than one such set, hence providing us with a choice of programming primitives?
Secondly, what is the exact kind of programming power that we have unleashed? Does it lead to
interesting programs that we have not been able to write before? We have outlined some speculative
applications for intensional recursion in [12]. Is iPCF a useful tool when it comes to attacking these?
Acknowledgments
I would like to thank Mario Alvarez-Picallo for our endless conversations on types and metaprogramming. This work was supported by the EPSRC as part of my doctoral research (award reference 1354534).
References
[1] Samson Abramsky (2014): Intensionality, Definability and Computation. In Alexandru Baltag & Sonja
Smets, editors: Johan van Benthem on Logic and Information Dynamics, Springer-Verlag, pp. 121–142,
doi:10.1007/978-3-319-06025-5 5.
[2] Leonard M Adleman (1990): An Abstract Theory of Computer Viruses. In: Advances in Cryptology CRYPTO’ 88, Lecture Notes in Computer Science 403, Springer New York, New York, NY, pp. 354–374,
doi:10.1007/0-387-34799-2 28.
[3] Henk Barendregt (1984): Lambda Calculus: Its Syntax and Semantics. North-Holland, Amsterdam.
[4] Alan Bawden (1999): Quasiquotation in LISP. In: Proceedings of the 6th ACM SIGPLAN Workshop on Partial Evaluation and Semantics-Based Program Manipulation (PEPM ’99). Available at
http://repository.readscheme.org/ftp/papers/pepm99/bawden.pdf.
[5] George S. Boolos (1994): The Logic of Provability.
Cambridge University Press, Cambridge,
doi:10.1017/CBO9780511625183.
12
Intensionality, Intensional Recursion, and the Gödel-Löb axiom
[6] Olivier Danvy & Karoline Malmkjaer (1988): Intensions and extensions in a reflective tower. In: Proceedings
of the 1988 ACM conference on LISP and functional programming (LFP ’88), ACM Press, New York, New
York, USA, pp. 327–341, doi:10.1145/62678.62725.
[7] Rowan Davies & Frank Pfenning (2001): A modal analysis of staged computation. Journal of the ACM
48(3), pp. 555–604, doi:10.1145/382780.382785.
[8] Daniel P. Friedman & Mitchell Wand (1984): Reification: Reflection without metaphysics. In: Proceedings
of the 1984 ACM Symposium on LISP and functional programming (LFP ’84), ACM Press, New York, New
York, USA, pp. 348–355, doi:10.1145/800055.802051.
[9] Murdoch J. Gabbay & Aleksandar Nanevski (2013): Denotation of contextual modal type theory (CMTT):
Syntax and meta-programming. Journal of Applied Logic 11(1), pp. 1–29, doi:10.1016/j.jal.2012.07.002.
[10] Paul Graham (1993): On LISP: Advanced Techniques for Common LISP. Prentice Hall.
[11] Carl A Gunter (1992): Semantics of programming languages: structures and techniques. Foundations of
Computing, The MIT Press.
[12] G. A. Kavvos (2016):
Kleene’s Two
http://arxiv.org/abs/1602.06220.
Kinds
of
Recursion.
CoRR.
Available
at
[13] G. A. Kavvos (2016): The Many Worlds of Modal λ -calculi: I. Curry-Howard for Necessity, Possibility and
Time. CoRR. Available at http://arxiv.org/abs/1605.08106.
[14] G. A. Kavvos (2017):
Dual-Context Calculi for Modal Logic.
http://arxiv.org/abs/1602.04860.
CoRR.
Available at
[15] G. A. Kavvos (2017): On the Semantics of Intensionality. In: Proceedings of the 20th International
Conference on Foundations of Software Science and Computation Structures (FoSSaCS). Available at
http://arxiv.org/abs/1602.01365.
[16] Stephen Cole Kleene (1938): On notation for ordinal numbers. The Journal of Symbolic Logic 3(04), pp.
150–155, doi:10.2307/2267778.
[17] John R. Longley & Dag Normann (2015): Higher-Order Computability. Theory and Applications of Computability, Springer Berlin Heidelberg, Berlin, Heidelberg, doi:10.1007/978-3-662-47992-6.
[18] John C Mitchell (1996): Foundations for programming languages. Foundations of Computing, The MIT
Press.
[19] Frank Pfenning & Rowan Davies (2001): A judgmental reconstruction of modal logic. Mathematical Structures in Computer Science 11(4), pp. 511–540, doi:10.1017/S0960129501003322.
[20] Gordon D. Plotkin (1977): LCF considered as a programming language. Theoretical Computer Science 5(3),
pp. 223–255, doi:10.1016/0304-3975(77)90044-5.
[21] Dana S. Scott (1993): A type-theoretical alternative to ISWIM, CUCH, OWHY. Theoretical Computer Science 121(1-2), pp. 411–440, doi:10.1016/0304-3975(93)90095-B.
[22] Brian Cantwell Smith (1984): Reflection and Semantics in LISP. In: Proceedings of the 11th ACM SIGACTSIGPLAN Symposium on Principles of Programming Languages (POPL ’84), ACM Press, New York, New
York, USA, pp. 23–35, doi:10.1145/800017.800513.
[23] Thomas Streicher (2006): Domain-theoretic Foundations of Functional Programming. World Scientific.
[24] Walid Taha & Tim Sheard (2000): MetaML and multi-stage programming with explicit annotations. Theoretical Computer Science 248(1-2), pp. 211–242, doi:10.1016/S0304-3975(00)00053-0.
[25] M. Takahashi (1995): Parallel Reductions in λ -Calculus. Information and Computation 118(1), pp. 120–127,
doi:10.1006/inco.1995.1057.
[26] Takeshi Tsukada & Atsushi Igarashi (2010): A logical foundation for environment classifiers. Logical Methods in Computer Science 6(4), pp. 1–43, doi:10.2168/LMCS-6(4:8)2010.
G. A. Kavvos
13
[27] Aldo Ursini (1979): A modal calculus analogous to K4W, based on intuitionistic propositional logic.
Studia Logica 38(3), pp. 297–311, doi:10.1007/BF00405387.
Available at
http://link.springer.com/10.1007/BF00405387.
[28] Mitchell Wand (1998): The Theory of Fexprs is Trivial. LISP and Symbolic Computation 10(3), pp. 189–199,
doi:10.1023/A:1007720632734.
[29] Mitchell Wand & Daniel P. Friedman (1988): The mystery of the tower revealed: A nonreflective description
of the reflective tower. Lisp and Symbolic Computation 1(1), pp. 11–38, doi:10.1007/BF01806174.
A
Appendix: Proof of confluence
We will use a variant of the proof in [14], i.e. the method of parallel reduction. This kind of proof was
originally discovered by Tait and Martin-Löf, and is nicely documented in [25]. Beause of the intensional
nature of our box (−) constructs, ours will be more nuanced and fiddly than any in op. cit. The method
largely follows the following pattern: we will introduce a second notion of reduction,
=⇒ ⊆ Λ × Λ
which we will ‘sandwich’ between reduction proper and its transitive closure:
−→ ⊆ =⇒ ⊆ −→ ∗
We will then show that =⇒ has the diamond property. By the above inclusions, the transitive closure
=⇒∗ of =⇒ is then equal to −→ ∗ , and hence −→ is Church-Rosser.
In fact, we will follow [25] in doing something better: we will define for each term M its complete
development, M ⋆ . The complete development is intuitively defined by ‘unrolling’ all the redexes of M at
once. We will then show that if M =⇒ N, then N =⇒ M ⋆ . M ⋆ will then suffice to close the diamond:
M
P
Q
M⋆
The parallel reduction =⇒ is defined in Figure 3. Ordinarily, instead of the axiom (refl) ensuring
M =⇒ M, we would simply have an axiom for variables, x =⇒ x. Then, M =⇒ M would be derivable.
However, we do not have a congruence rule for box (−), and also our fixpoint construct would block the
possibility of this. We are thus forced to include M =⇒ M, which slightly complicates the lemmas that
follow.
Indeed, the main lemma that usually underpins the confluence proof is this: if M =⇒ N and P =⇒ Q,
M[P/x] =⇒ N[Q/x]. However, this is intuitively wrong: no reductions should happen under boxes, so
this should only hold if we are substituting for a variable not occuring under boxes. Hence, this lemma
splits into three different ones:
• one showing that P =⇒ Q implies M[P/x] =⇒ M[Q/x], if x does not occur under boxes: this is the
price to pay for replacing the variable axiom with (refl);
• one showing that M =⇒ N implies M[P/u] =⇒ N[P/u], even if u is under a box; and
Intensionality, Intensional Recursion, and the Gödel-Löb axiom
14
Figure 3: Parallel Reduction
M =⇒ M
M =⇒ N
(refl)
M =⇒ N
λ x:A. M =⇒ λ x:A. N
P =⇒ Q
(λ x:A. M)P =⇒ N[Q/x]
M =⇒ N
(congλ )
M =⇒ M ′
⊃G true M N =⇒ M ′
P =⇒ Q
MP =⇒ NQ
N =⇒ N ′
(⊃1 )
⊃G false M N =⇒ N ′
M =⇒ N
let box u ⇐ box P in M =⇒ N[P/u]
f˜(box M) =⇒ box f (M)
M =⇒ N
(app)
(⊃2 )
(β )
fix z in box M =⇒ box M[fix z in box M/z]
M closed, f (M) ↓
(→ β )
(fix)
(int)
P =⇒ Q
let box u ⇐ M in P =⇒ let box u ⇐ N in Q
(let-cong)
Remark. In addition to the above, one should also include rules for the constants, but these are merely
restatements of the rules in Figure 2.
G. A. Kavvos
15
• one showing that, if x does not occur under boxes, M =⇒ N and P =⇒ Q indeed imply M[P/x] =⇒
N[Q/x]
But let us proceed with the proof.
Lemma 1. If M =⇒ N then M[P/u] =⇒ N[P/u].
Proof. By induction on the generation of M =⇒ N. Most cases trivially follow, or consist of simple
invocations of the IH. In the case of (→ β ), the known substitution lemma suffices. Let us look at the
cases involving boxes.
C ASE (β ). Then M =⇒ N is let box v ⇐ box R in S =⇒ S′ [R/v] with S =⇒ S′ . By the IH, we
have that S[P/u] =⇒ S′ [P/u], so let box v ⇐ box R[P/u] in S[P/u] =⇒ S[P/u][R[P/u]/v], and this
last is α -equivalent to S[R/v][P/u] by the substitution lemma.
C ASE (fix). A similar application of the substitution lemma.
C ASE (int). Then M =⇒ N is f˜(box Q) =⇒ box f (Q), with Q closed. Hence
f˜(box Q) [P/u] ≡ f˜(box Q) =⇒ box f (Q) ≡ ( f (Q)) [P/u]
simply because Q is closed.
Lemma 2. If P =⇒ Q and x 6∈ FV ≥1 (M), then M[P/x] =⇒ M[Q/x].
Proof. By induction on the term M. The only non-trivial cases are those for M being a variable, or
box M ′ . In the first case, depending on which variable M is, use either (refl), or the assumption P =⇒ Q.
In the second case, (box M ′ )[P/x] ≡ box M ′ ≡ (box M ′ )[Q/x] as x does not occur under a box, so use
(refl) again.
Lemma 3. If M =⇒ N, P =⇒ Q, and x 6∈ FV ≥1 (M), then
M[P/x] =⇒ N[Q/x]
Proof. By induction on the generation of M =⇒ N. The cases for most congruence rules and constants
follow trivially, or from the IH. We prove the rest.
C ASE (refl). Then M =⇒ N is actually M =⇒ M, so we use Lemma 2 to infer M[P/x] =⇒ M[Q/x].
˜
C ASE (int).
Then M =⇒ N is actually f (box M) =⇒ box f (M). But M is then closed, so
f˜(box M) [P/x] ≡ f˜(box M) =⇒ box f (M) ≡ (box f (M)) [Q/x].
C ASE (⊃i ). Then M =⇒ N is ⊃G true M N =⇒ M ′ with M =⇒ M ′ . By the IH, M[P/x] =⇒
M ′ [Q/x], so
⊃G true M[P/x] N[P/x] =⇒ M ′ [Q/x]
by a single use of (⊃1 ). The case for false is similar.
C ASE (→ β ). Then (λ x′ :A. M)N =⇒ N ′ [M ′ /x′ ], where M =⇒ M ′ and N =⇒ N ′ . Then
(λ x′ :A. M)N [P/x] ≡ (λ x′ :A. M[P/x])(N[P/x])
16
Intensionality, Intensional Recursion, and the Gödel-Löb axiom
But, by the IH, M[P/x] =⇒ M ′ [Q/x] and N[P/x] =⇒ N ′ [Q/x]. So, by the rules (congλ ) and (app),
and then rule (→ β ), we have
(λ x′ :A. M[P/x])(N[P/x]) =⇒ M ′ [Q/x] N ′ [Q/x]/x′
But this last is α -equivalent to (M ′ [N ′ /x′ ]) [Q/x] by the substitution lemma.
C ASE (β ). Then let box u′ ⇐ box M in N =⇒ N ′ [M/u′ ] where N =⇒ N ′ . By assumption, we
have that x 6∈ FV (M) and x 6∈ FV ≥1 (N). Hence, we have by the IH that N[P/x] =⇒ N ′ [Q/x], so by
applying (β ) we get
(let box u′ ⇐ box M in N)[P/x] ≡ let box u′ ⇐ box M[P/x] in N[P/x]
≡ let box u′ ⇐ box M in N[P/x]
=⇒ N ′ [Q/x][M/u′ ]
But this last is α -equivalent to N ′ [M/u′ ][Q/x], by the substitution lemma and the fact that x does
not occur in M.
C ASE (fix). As x 6∈ FV≥1 (fix z in box M), we have that x does not occur in M, and thus (fix z in box M)[P/x] ≡
fix z in box M, and also
(box M[fix z in box M/z])[Q/x] ≡ box M[fix z in box M/z]
Thus, a single use of (fix) suffices.
We now pull the following definition out of the hat:
Definition 2 (Complete development). The complete development M ⋆ of a term M is defined by the
following clauses:
x⋆ = x
def
c⋆ = c
def
(c ∈ { f˜, nb, zero?, . . . })
(λ x:A. M)⋆ = λ x:A. M ⋆
⋆ def
f˜(box M) = box f (M)
def
⋆
def
⋆
if M is closed
⋆
((λ x:A. M)N) = M [N /x]
(⊃G true M N)⋆ = M ⋆
def
(⊃G false M N)⋆ = N ⋆
def
(MN)⋆ = M ⋆ N ⋆
def
(box M)⋆ = box M
def
(let box u ⇐ box M in N)⋆ = N ⋆ [M/u]
def
(let box u ⇐ M in N)⋆ = let box u ⇐ M ⋆ in N ⋆
def
(fix z in box M)⋆ = box M[fix z in box M/z]
def
G. A. Kavvos
17
We need the following two technical results as well.
Lemma 4. M =⇒ M ⋆
Proof. By induction on the term M. Most cases follow immediately by (refl), or by the IH and an
application of the relevant rule. The case for box M follows by (refl), the case for fix z in box M follows
by (fix), and the case for f˜(box M) by (int).
Lemma 5 (BFV antimonotonicity). If M =⇒ N then FV ≥1 (N) ⊆ FV ≥1 (M).
Proof. By induction on M =⇒ N.
And here is the main result:
Theorem 7. If M =⇒ P, then P =⇒ M ⋆ .
Proof. By induction on the generation of M =⇒ P. The case of (refl) follows by lemma variable rule is
trivial, and the cases of congruence rules follow from the IH. We show the rest.
C ASE (→ β ). Then we have (λ x:A. M)N =⇒ M ′ [N ′ /x], with M =⇒ M ′ and N =⇒ N ′ . By the IH,
M ′ =⇒ M ⋆ and N ′ =⇒ N ⋆ . We have that x 6∈ FV ≥1 (M), so by Lemma 5 we get that x 6∈ FV ≥1 (M ′ ).
Hence, by Lemma 3 we get M ′ [N ′ /x] =⇒ M ⋆ [N ⋆ /x] ≡ ((λ x:A. M) N)⋆ .
C ASE (β ). Then we have
let box u ⇐ box M in N =⇒ N ′ [M/u]
where N =⇒ N ′ . By the IH, N ′ =⇒ N ⋆ , so it follows that
N ′ [M/u] =⇒ N ⋆ [M/u] ≡ (let box u ⇐ box M in N)⋆
by Lemma 1.
C ASE (fix). Then we have
fix z in box M =⇒ box M[fix z in box M/z]
But (fix z in box M)⋆ ≡ box M[fix z in box M/z], so a single use of (refl) suffices.
C ASE (int). Similar.
| 6cs.PL
|
Constructing Likelihood Functions for Interval-valued
Random Variables
X. Zhang∗†
S. A. Sisson∗
arXiv:1608.00107v1 [stat.ME] 30 Jul 2016
Abstract
There is a growing need for the ability to analyse interval-valued data. However, existing descriptive frameworks to achieve this ignore the process by which
interval-valued data are typically constructed; namely by the aggregation of realvalued data generated from some underlying process. In this article we develop
the foundations of likelihood based statistical inference for random intervals that
directly incorporates the underlying generative procedure into the analysis. That is,
it permits the direct fitting of models for the underlying real-valued data given only
the random interval-valued summaries. This generative approach overcomes several
problems associated with existing methods, including the rarely satisfied assumption of within-interval uniformity. The new methods are illustrated by simulated
and real data analyses.
Keywords: Likelihood theory; Imprecise data; Interval data; Symbolic data analysis.
1
Introduction
As we move inevitably towards a more data-centric society, there is a growing need
for the ability to analyse data that are constructed in non-standard forms, rather than
represented as continuous points in Rp (Billard and Diday, 2003). The simplest and
most popular of these is interval-valued data.
Interval-valued observations can arise naturally through the data recording process,
and essentially result as a way to characterise measurement error or uncertainty of
an observation. Examples include blood pressure, which is typically reported as an
interval due to the inherent continual changes within an individual (Billard and Diday,
2006); data quantisation, such as rounding or truncation, which results in observations
being known to lie within some interval (McLachlan and Jones, 1988; Vardeman and
Lee, 2005); and the expression of expert-elicited intervals that contain some quantity of
interest (Fisher et al., 2015), among others.
The use of intervals as a summary representation of a collection of classical real-valued
data is also rapidly gaining traction. Here the aggregation of a large and complex dataset
∗
†
School of Mathematics and Statistics, University of New South Wales, Sydney 2052 Australia.
Communicating Author: [email protected]
1
into a smaller collection of suitably constructed intervals can enable a statistical analysis
that would otherwise be computationally unviable (Billard and Diday, 2003). Where
interest in the outcome of an analysis exists at the level of a group, rather than at an
individual level, interval-valued data provide a convenient group-level aggregation device
(Neto and Carvalho, 2010; Noirhomme-Fraiture and Brito, 2011). Similarly, aggregation
of individual observations within an interval structure allows for some preservation of
privacy for the individual (Domingues et al., 2010).
The earliest systematic study of interval-valued data is in numerical analysis, where
Moore (1966) used intervals as a description for imprecise data. Random intervals can
be considered as a special case of random sets (Molchanov, 2005), the theory of which
brings together elements of topology, convex geometry, and probability theory to develop a coherent mathematical framework for their analysis. Matheron (1975) gave the
first self-contained development of statistical models for random sets, including central
limit theorems and a law of large numbers, and Beresteanu and Molinari (2008) derived
these limit theorems specifically
for random intervals. In this framework, interval-valued
random
variables
[X]
=
X,
X
are modelled as a bivariate real-valued random vector
X, X , where X ≤ X, using standard inferential techniques. This approach is also used
for partially identified models (Beresteanu et al., 2012; Molchanov and Molinari, 2014),
where the object of economic and statistical interest is a set rather than a point. In
probabilistic modelling, Lyashenko (1983) introduced normal random compact convex
sets in Euclidean space, and showed that a normal random interval is simply a Gaussian
displacement of a fixed closed bounded interval. Sun and Ralescu (2014) subsequently
extended this idea to normal hierarchical models for random intervals.
A more popular framework for the analysis of interval-valued data, and one which
we focus on here, is symbolic data analysis (Diday, 1987). Symbols can be considered as
distributions of real-valued data points in Rp , such as intervals and histograms, or more
general structures including lists. They are typically constructed as the aggregation into
summary form of real-valued data within some group, and so the symbol is interpreted
as taking values as described by the summary distribution. As a result, symbols have
internal variations and structures which do not exist in real-valued data, and methods
for analysing them must account for within-symbol variation in addition to between
symbol variation. In practice, the most common form of symbol is the random interval
or its p-dimensional extension, the p-hypercube. See Billard and Diday (2003, 2006);
Noirhomme-Fraiture and Brito (2011) for a review of recent results.
While many exploratory and descriptive data analysis techniques for symbolic data
have been developed (see e.g. Billard, 2007, for an overview), there is a paucity of results for developing a robust statistical inferential framework for these data. The most
significant of these (Le-Rademacher and Billard, 2011) maps the parameterisation of the
symbol into a real-valued random vector, and then uses the standard likelihood framework to specify a suitable model. In the random interval setting, this
to
is equivalent
the random set theory approach by modelling
the
random
set
[X]
=
X,
X
by
the
con
strained real-valued random vector X, X or, more commonly, a reparameterisation to
the unconstrained interval centre and half-range (Xc , Xr ) = (X + X)/2, (X − X)/2 ,
2
which is then more easily modelled, e.g. (Xc , log Xr ) ∼ N2 (µ, Σ). This likelihood framework has been used for the analysis of variance (Brito and Duarte Silva, 2012), time
series forecasting (Arroyo et al., 2010) and interval-based regression models (Xu, 2010)
among others.
While sensible, by nature the above methods for modelling real-valued random variables only permit descriptive modelling at the level of the real-valued random vector
(X, X) (or its equivalent for p-hypercubes). However this descriptive approach completely ignores the generative procedure commonly assumed and implemented for the
construction of observed intervals; namely that the underlying real-valued data are produced from some generative model X1 , . . . , Xm ∼ f (x1 , . . . , xm |θ), and the interval is
then constructed via some aggregation process, e.g. X = mink {Xk } and X = maxk {Xk }.
If interest is then in fitting the generative model f (x1 , . . . , xm |θ) for inferential or predictive purposes, while only observing the random intervals rather than the underlying
real-valued dataset, or in having the interpretation of model parameters be independent
of the form of the interval construction process, then the above descriptive models will
be inadequate. Further, the existing descriptive models for random intervals typically
assume that the distribution of latent data points within the random interval is uniform.
Under the above generative procedure, except in specific cases this will almost always
be untrue, and this is generally accepted as a false assumption in practice.
In this article we develop the theoretical and methodological foundations of statistical
models for random intervals that are directly constructed from an assumed generative
model f (x1 , . . . , xm |θ) and a data aggregation function ϕ(·) that maps the space of
real-valued data to the space of intervals.
To the best of our knowledge, this represents the first attempt to move beyond
the restrictive descriptive models which are prevalent in the literature, and provide an
inferential framework that aligns with the generative interval-construction process that
is typical in practice. In addition to providing more directly interpretable parameters,
it also provides a natural mechanism for departure from the uncomfortable uniformitywithin-intervals assumption of descriptive models.
After establishing a measurable space for a random interval [X] in Section 2 (of which,
Section 2.1 may be omitted on a first reading), we construct both distribution functions,
F[X] (·), and density functions, f[X] (·), for random intervals on this space based on the
idea of containment functionals (Molchanov, 2005). From this we are able to establish
that the current practice of modelling random intervals, by the bivariate real-valued
random vector (X, X), is a valid procedure under this framework. In Section 3, these
results naturally lead to the construction of likelihood functions for generative models
that are directly constructed from likelihood functions for the underlying real-valued
data. This approach is evidenced by the recovery of existing results on the distribution
of the order statistics of a random sample under certain conditions. We are also able to
show that a limiting case of the derived generative models results in a valid descriptive
model, implying that existing descriptive models in fact have a direct interpretation in
terms of an underlying generative model.
All results are naturally extended from random intervals to random p-hypercubes
3
in Section 4. In Section 5, we contrast the performance of generative and descriptive
models for interval-valued random variables on both simulated data, and for a reanalysis
of a credit card dataset previously examined by Brito and Duarte Silva (2012). Here we
establish that the use of existing descriptive models to analyse interval data constructed
under a generative process (which is typical in practice), can result in misinterpretable
and biased parameter estimates, and poorer overall fits to the observed interval data
than those obtained under generative models. Finally, we conclude with a discussion.
2
Distributions of Random Intervals
We begin by establishing the measurable space of I = {[x, y] : − ∞ < x ≤ y < +∞}
for a random closed1 interval [X] = [X, X] before constructing distribution and density
functions on this space. Throughout we denote Ω as a sample space equipped with a σalgebra F w.r.t. a probability measure P (·). We denote a vector of m < ∞ real-valued
random variables by X1:m = (X1 , . . . , Xm )| , where Xk : Ω 7→ R for k = 1, . . . , m, and
xk is a realisation of Xk . We can then define a data aggregation function ϕ : Rm 7→ I
that maps a vector X1:m to the space of intervals I via [X] = ϕ(X1:m ), so that [X] is a
random (closed) interval. For example, a useful specification for random intervals might
construct the bivariate real-valued random variable (X, X) from the minimum (X) and
maximum (X) of the components of X1:m .
2.1
Constructing a measurable space
In order to construct a measurable space of I, we identify those subsets of I, which are
equivalent to particular subsets of Rm . A subset of interest is {[x0 ] ⊆ [x]} = {[x0 ] : [x0 ] ⊆
[x]}, which corresponds to the collection of all intervals that are a subinterval of or equal
to [x]. This subset is the image of the event {[X] ⊆ [x]} = {ω ∈ Ω : [X](ω) ⊆ [x]} on I.
{[X] ⊆ [x]} may also be written as {ϕ(X1:m ) ⊆ [x]} = {ω ∈ Ω : ϕ(X1:m (ω)) ⊆ [x]}, of
which the image on Rm is {ϕ(x01:m ) ⊆ [x]} = {x01:m : ϕ(x01:m ) ⊆ [x]}, i.e. the subset of Rm
containing those x01:m that can generate an interval which is a subinterval of or equal to
[x]. The two subsets {[x0 ] ⊆ [x]} and {ϕ(x01:m ) ⊆ [x]} are equivalent as their pre-images
on Ω are identical. As a result, given a probability measure on Rm , the probability of
{ϕ(x01:m ) ⊆ [x]}, and hence of {[x0 ] ⊆ [x]}, can be calculated if only if it is measurable.
This implies that in a measurable space of I, {[x0 ] ⊆ [x]} should be measurable.
We construct the metric topology on I, denoted by TI , induced by the Hausdorff
metric, which specifies the distance between elements [a] and [b] as
dH ([a], [b]) = max |a − b|, |a − b| ,
where |·| denotes absolute value.
If we consider the mapping h([x]) = (x, x) from I to R2 ,
then we have d2 (a, a), (b, b) = dH ([a], [b]) for any [a], [b] ∈ I, where d2 (·) is the square
metric on R2 . That is, h is a distance preserving map, or isometry, and hence (I, TI ) is
1
Throughout this paper, we only consider closed intervals (hypercubes). Results for other types of
intervals (hypercubes) can be obtained in a similar way.
4
isometrically embedded into the metric topological space on R2 induced by d2 (·), which
is also known as the standard topology.2 This implies that TI inherits properties of the
standard topology on R2 , such as completeness, local compactness and separability. (See
Appendix A for details)
Let F = {{[x0 ] ⊆ [x]} : [x] ∈ I} be the collection of subsets of interest. We can now
construct a measurable space involving F from the topology TI . Let BI be the smallest
σ-algebra containing all open subsets BI = σ(TI ), i.e. the Borel σ-algebra on I.3 This
Borel σ-algebra contains F, as all elements of F are closures of some elements of TI
(Appendix A). The following theorem provides the stronger result that F is sufficient to
construct BI .
Theorem 1. The Borel σ-algebra on I is the smallest σ-algebra generated by F, i.e.
BI = σ(F). (Proof in Appendix B.1)
This property indicates that BI is rich enough to ensure that all elements in F are
measurable. It also suggests that if we only define a proper non-negative function on
F, we can extend it to a measure on (I, BI ). In particular, if the induced measure is a
probability measure, it would then be the distribution function of [X].
Based on the isometry h([x]) = (x, x) between I and R2 , we now construct a measure on (I, BI ), representing the uniform measure on I, which gives equal weight to all
intervals.
Let the Borel σ-algebra on R2 be BR2 , and µ : BR2 7→ [0, +∞) be the Lebesgue
measure on (R2 , BR2 ). Due to the isometry h([x]) = (x, x), we then have that µI = µ ◦ h
is the uniform measure on (I, BI ). Consequently, the uniform measure of every Borel
subset of I can be calculated via µ(·) and h(·). Specifically, for every element of F, we
have
1
µI ({[x0 ] ⊆ [x]}) = µ(h({[x0 ] ⊆ [x]})) = (x − x)2 ,
2
as h({[x0 ] ⊆ [x]}) = {(x0 , x0 ) : x ≤ x0 ≤ x0 ≤ x} is the region of an isosceles right triangle
on the real plane. From Theorem 1, the uniform measure of all Borel subsets E ∈ BI is
also available.
Theorem 2. Define the infinitesimal neighbourhood of [x] as
d[x] = {[x0 ] ∈ I | x − dx < x0 ≤ x ≤ x ≤ x0 < x + dx},
where dx, dx > 0. Its uniform measure is µI (d [x]) = dx × dx. (Proof in Appendix B.2)
From the above we note that µI (·) is a non-atomic measure, i.e. µI ({[x]}) = 0, where
{[x]} is a set containing a single interval [x]. Further, there is a convenient way to
compute the value of µI (·) for any Borel subsets via the Lebesgue integration on R2 .
Namely, for any subsets E ∈ BI
Z
ZZ
µI (E) =
µI (d [x]) =
dxdx.
E
h(E)
2
The standard topology on R2 is generated by the open rectangles (Munkres, 2000).
The topology TI is the collection of all open subsets of I, and the Borel σ-algebra is the smallest
σ-algebra containing all open subsets (Munkres, 2000).
3
5
Accordingly, through such isometry, the measurable space of intervals (I, BI ) inherits
the convenient structure and properties of the real plane. These results permit the
construction of distribution and density functions of random intervals.
2.2
Containment distribution functions
In random set theory (Molchanov, 2005; Molchanov and Molinari, 2014), two types of
functionals, the capacity functional, T[X] ([x]) = P ([X] ∩ [x]), and the containment func? ([x]) = P ([X] ⊂ [x]), are commonly used to identify a unique probability
tional, C[X]
measure on (I, BI ). In the present setting, we consider a variant of the containment
functional, C[X] ([x]) = P ([X] ⊆ [x]), which is more convenient for theoretical devel? (·) and C
opment and model construction. Due to the similarity of C[X]
[X] (·) in both
functionality and interpretation, we still refer to C[X] (·) as the containment functional
throughout this article.
? (·), a containment functional of a random interval [X] is a functional
Similar to C[X]
C[X] : I 7→ [0, 1] having the following properties:
i) C[X] ([x, x]) → 1, when x → −∞ and x → +∞;
ii) If [x1 ] ⊇ [x2 ] ⊇ · · · ⊇ [xn ] ⊇ · · · and ∩∞
n=1 [xn ] ∈ I, then
lim C[X] ([xn ]) = C[X] (∩∞
n=1 [xn ]);
n→∞
iii) For any [x] ⊆ [y], C[X] ([x]) ≤ C[X] ([y]) and
C[X] ([y]) − C[X] ([y, x]) − C[X] ([x, y]) + C[X] ([x]) ≥ 0.
Based on these properties we can establish the following result.
Theorem 3. Let [X] : Ω 7→ I be a measurable function defined on the sample space
(Ω, F ) and σ([X]) ⊆ F be the σ-algebra generated by [X]. The containment functional
C[X] : I 7→ [0, 1] determines a unique probability measure P : σ([X]) 7→ [0, 1] for [X], such
that P ([X] ⊆ [x]) = C[X] ([x]) for all [x] ∈ I. (Proof in Appendix B.3)
That is, the containment functional C[X] (·) plays the role of the distribution function
of [X]. However, it is more convenient to work with functions defined on the real plane,
so we equivalently define the containment distribution function as F[X] (x, x) = C[X] ([x]).
Definition 1. The containment distribution function F[X] : R2 7→ [0, 1] of the random
interval [X] has the following properties:
i) F[X] (−∞, +∞) = 1 and F[X] (x, x) = 0 when x > x;
ii) F[X] (x, x) is left-continuous in x and right-continuous in x;
iii) F[X] (x, x) is non-increasing in x, and non-decreasing in x;
6
iv) For y ≤ x ≤ x ≤ y, F[X] (y, y) − F[X] (y, x) − F[X] (x, y) + F[X] (x, x) ≥ 0.
From the above, 1 − F[X] (x, +∞) and F[X] (−∞, x) are the marginal cumulative
distribution functions of the lower bound X and the upper bound X, respectively. In
addition, F[X] (x, x) can also be naturally constructed from the generative framework,
where [X] = ϕ(X1:m ), by noting that the two events of Ω, {ϕ(X1:m ) ⊆ [x]} and {[X] ⊆
[x]}, are equal. If ϕ is measurable, we may compute the probability of {[X] ⊆ [x]} via
P (ϕ(X1:m ) ⊆ [x]), given the distribution of latent data points X1:m . Accordingly, the
containment distribution function of [X] can be constructed as
F[X] (x, x) = P (ϕ(X1:m ) ⊆ [x]).
(1)
Intuitively, for m > 2, the distribution of [X] = ϕ(X1:m ) carries less information
than the joint distribution of latent data points X1:m , due to the data aggregation. More
−1
precisely, we may rewrite the σ-algebra generated by [X] as σ(ϕ(X1:m )) = X1:m
(σ(ϕ)),
−1
m
where σ(ϕ) is the σ-algebra on R generated by ϕ and X1:m is the inverse map of X1:m .
−1
−1
As ϕ is measurable, σ(ϕ) ⊆ BRm and σ([X]) = X1:m
(σ(ϕ)) ⊆ X1:m
(BRm ), where the
last term is the σ-algebra generated by X1:m . Hence, we have σ([X]) ⊆ σ(X1:m ).
Note that [X] degenerates to a scalar random variable when it only contains a single
point, i.e. when X = X = X, and so P ([X] ⊆ [x]) = P (X ∈ [x]) identifies the
distribution of a univariate real-valued random variable. In the generative framework,
a univariate real-valued random variable is produced when either m = 1, or when X1 =
· · · = Xm = X for m > 1. Accordingly, this theory for random intervals is consistent
with standard statistical theory.
For the following sections we assume that the data aggregation function ϕ(·) is always
measurable.
2.3
Density functions
The density function of [X] is formally defined as the Radon-Nikodym derivative (Durrett, 2010) of a probability measure on I over the uniform measure µI (·) as the reference
measure.
Theorem 4. Let F[X] : R2 7→ [0, 1] be the containment distribution function of a random
interval [X]. If F[X] (·) is twice differentiable, then the density function of [X] is
f[X] (x, x) = −
∂2
F (x, x).
∂x∂x [X]
(2)
(Proof in Appendix B.4)
From Definition 1, we have that f[X] (x, x) is a non-negative function and f[X] (x, x) =
0 when x > x. Conversely, the containment distribution function of [X] can be obtained
by integration of a valid density function.
7
Theorem 5. A non-negative function f[X] : R2 7→ R is the density function of a random
RxRx
interval [X] with the containment distribution function F[X] (x, x) = x x f[X] (x0 , x0 ) dx0 dx0
R +∞ R +∞
if and only if it satisfies f[X] (x, x) = 0 when x > x, and −∞ −∞ f[X] (x, x) dxdx = 1.
(Proof in Appendix B.5)
Note that a valid density function of [X] is also a density function for a bivariate
real-valued random variable. Being able to express the density function f[X] (x, x) of the
random interval [X] as the joint density of two real-valued random variables, X and
X, justifies those existing (descriptive) methods for modelling random intervals (e.g.
Arroyo et al., 2010; Le-Rademacher and Billard, 2011; Brito and Duarte Silva, 2012,
see Section 3.1) that directly specify a joint distribution for X, X|X ≤ X, or some
reparameterisation that circumvents bounding the parameter space.
3
Likelihood Functions and Models
Theorem 5 implies that one approach for constructing models for [X] is by constructing
models for the two real-valued random variables X and X with X ≤ X. We term
this approach the descriptive model. While it can describe the structure and variation
between intervals, it is unable to model the distribution of latent data points within
an interval, as it is simply a model for the interval endpoints. This approach is almost
universal in the symbolic data analysis literature. As an alternative we develop the
generative model, which is constructed directly at the level of the latent data points
X1:m through the data aggregation function ϕ(·). In the following, we use F[X] (·) and
f[X] (·) for interval-valued random variables, and F (·) and f (·) for real-valued random
variables.
3.1
Descriptive models
A descriptive model treats [X] = [X, X] as a bivariate real-valued random variable
(X, X) with X ≤ X. The results in Section 2, demonstrate that the joint density function
of (X, X) is a valid density function for a random interval. We write f[X] (x, x|α) =
f (x, x|α) as the likelihood function of X, X , where f (x, x|α) satisfies both conditions
in Theorem 5 and α denotes the parameter vector of interest. Rather than construct
models directly on (X, X) with the awkward constraint X ≤ X, a simpler approach is
to remove this constraint through reparameterisation. For example, defining the interval
x+x x−x
1
centre Xc = X+X
and half-range Xr = X−X
2
2 , we obtain f[X] (x, x|α) = 2 g( 2 , 2 |α),
where g(xc , xr |α) is a density function for Xc and Xr .
Most existing methods to model random intervals (e.g. Arroyo et al., 2010; LeRademacher and Billard, 2011; Brito and Duarte Silva, 2012) can be classified as descriptive models. Their interpretation is simple and they are convenient to use. However,
by construction they are only models for interval endpoints, and as a consequence have
limitations in providing information about the distribution of the latent data points
X1:m .
8
3.2
Generative models
A generative model of the random interval may be constructed bottom up from the
distribution of latent data points X1:m and the data aggregation function ϕ(·), based on
(1). Here, the random interval [X] is constructed from X1:m and ϕ(·) via [X] = ϕ(X1:m ).
If f (x1:m |θ) is the likelihood function of the m data points, then from (1) we may form
the containment distribution function of [X] as
Z
F[X] (x, x|θ, m) = f (x1:m |θ) dx1:m ,
(3)
A
where A = {ϕ(x1:m ) ⊆ [x]} denotes the collection of x1:m , for which the corresponding
interval is a subset of or equal to [x]. If ϕ(·) is continuous, the containment distribution
function (3) is twice differentiable, and so from (2) its contribution to the likelihood
function would be
Z
∂2
f[X] (x, x|θ, m) = −
f (x1:m |θ) dx1:m .
(4)
∂x∂x A
Note that containment distribution functions (3) and density functions (4) of generative
models contain a parameter m specifying the number of latent data points within [X].
When m is large, the evaluation of (4) can be challenging as it involves a high
dimensional integration. This integration can be simplified in the case where X1:m is a
sequence of i.i.d. random variables with Xk ∼ f (x|θ) for k = 1, . . . , m. We denote the
likelihood function of [X] with the i.i.d. latent data points by
?
f[X]
(x, x|θ, m)
∂2
=−
∂x∂x
Z Y
m
f (xk |θ) dx1:m ,
(5)
A k=1
and term it the i.i.d. generative model.
In practice, the data aggregation function ϕ(·) will typically depend on the order
statistics of the latent data points, so that ϕl,u (x1:m ) = [x(l) , x(u) ], where x(l) and x(u)
are respectively the l-th (lower) and u-th (upper) order statistics of x1:m . The region
for integration in (3) and (4) then becomes A = {x1:m : x ≤ x(l) ≤ x(u) ≤ x} – the
collection of x1:m for which the l-th order statistic is no less than x and the u-th order
statistic is no greater than x. In this case, and for i.i.d. random variables Xk ∼ f (x|θ)
for k = 1, . . . , m, the likelihood function (5) becomes
?
f[X]
(x, x|θ, m, l, u) =
m!
[F (x|θ)]l−1
(l − 1)!(u − l − 1)!(m − u)!
× [F (x|θ) − F (x|θ)]u−l−1 [1 − F (x|θ)]m−u f (x|θ)f (x|θ), (6)
Rx
where F (x|θ) = −∞ f (z|θ) dz is the cumulative distribution function of Xk . That is, (5)
reduces to (6), which is the joint likelihood function of the l-th and u-th order statistics
of m i.i.d. samples. Consequently, if l/(m + 1) → p and u/(m + 1) → p as m → ∞, the
9
distribution of [X] converges to a point mass at [Q(p; θ), Q(p; θ)], where Q(·; θ) is the
quantile function of f (x|θ).
Further simplification is possible when [X] is constructed from the minimum and
maximum values of X1:m (so that l = 1 and u = m). Here A = {x1:m : x ≤ xk ≤
x, k = 1, . . . , m} is a hypercube in Rm with identical length in each dimension, and the
likelihood function (6) becomes
??
(x, x|θ, m) = m(m − 1)[F (x|θ) − F (x|θ)]m−2 f (x|θ)f (x|θ).
f[X]
(7)
In this case, if the support of f (x|θ) is bounded on [a, b], then as m → ∞, the distribution
of [X] converges to a point mass at [a, b]. However, if f (x|θ) has unbounded support,
the distribution of [X] will diverge to (−∞, +∞).
From the above we may conclude that for i.i.d. generative models, when m is large,
all interval-valued observations will be similar. As in practice we may expect significant variation in interval-valued observations, even for large m, this indicates that the
usefulness of an i.i.d. model may be restricted to specific settings.
3.3
Hierarchical generative models
Evaluating the likelihood function (4) of the generative model for general latent distributions f (x1:m |θ) of latent data points is challenging, except in simplified settings. Here
we consider a special class of the generative model for which the latent data points X1:m
are exchangeable. This exchangeability leads to a hierarchical generative model, which
can capture both inter- and intra- interval structure and variability.
Suppose that X1:m are exchangeable, i.e. their joint distribution is invariant to any
permutations of X1:m . From de Finetti’s Theorem (Aldous, 1985), the distribution of
X1:m may be represented as a mixture, i.e.
Z
(m)
P (X1:m ∈ A) = P? (X1:m ∈ A) µP? (dP? ),
(8)
(m)
where
µP? is the distribution on the space of all probability measures of R, and P? =
Q
P
is the product measure on Rm . In other words, all Xk for k = 1, . . . , m, are i.i.d.
?
m
from P? with P? ∼ µP? . By recalling from (3) and (4) that A = {ϕ(x1:m ) ⊆ [x]}, then the
(m)
(m)
mixture component P? (X1:m ∈ A) equals P? ([X] ⊆ [x]), which is the containment
distribution function for an i.i.d. generative model of [X], with Xk ∼ P? for k = 1, . . . , m
and the same data aggregation function ϕ(·). This means that P ([X] ⊆ [x]), which equals
(m)
P (X1:m ∈ A), may be represented as the mixture of P? ([X] ⊆ [x]) with P? ∼ µP? , i.e.
as a mixture of i.i.d. generative models.
In the following we consider the case when P? belongs to some parametric family, so
that dP? = f (x|θ) dx. From
R Qm(8), the joint density function of X1:m is then given by the
mixture representation
k=1 f (xk |θ)π(θ) dθ, where the mixing distribution π(θ) may
be non-parametric or parametric π(θ|α) with parameter α. The resulting containment
distribution function of [X] is then the mixture of F[X] (x, x|θ, m) given in (3), with
10
Q
f (x1:m |θ) = m
k=1 f (xk |θ), w.r.t. π(θ|α). If ϕ(·) is continuous, we obtain the likelihood
function of such a generative model as
Z
?
f[X] (x, x|α, m) = f[X]
(x, x|θ, m)π(θ|α) dθ,
(9)
? (x, x|θ, m) is the likelihood function of the i.i.d. generative model (5).
where f[X]
In practice, the latent data points X1:m may not be exchangeable. However the data
aggregation functional ϕ(·) may be symmetric. Let Γ be the set of all permutations of
the indices from 1 to m, and Xγ be the latent data points X1:m permuted according
to γ ∈ Γ with density function f (xγ ). As ϕ(·) is symmetric, ϕ(xγ ) = ϕ(x1:m ) and
[Xγ ] = ϕ(Xγ ) has the same containment distribution function
as [X]. As a result, for
1 P
the exchangeable random variables defined as Y1:m ∼ m! γ∈Γ f (Xγ ), [Y ] = ϕ(Y1:m ) has
the same containment distribution function as [X].
The existence of such Y1:m implies that when the latent data points X1:m are aggregated into intervals [X] by symmetric data aggregation methods, information on the
order-related dependence structure will vanish. As a result, it is unnecessary to model
the distribution of X1:m with a more complex dependence structure than exchangeability
– modelling the exchangeable Y1:m will be sufficient.
Accordingly, for random intervals [X1 ], . . . , [Xn ], the generative model (9) can be
directly interpreted as the hierarchical model
[Xi ] = ϕ(Xi,1:m )
Xi,k ∼ f (x|θi ), k = 1, . . . , mi
θi ∼ π(θ|α)
with known mi for i = 1, . . . , n. Thus, we term them hierarchical generative models. The contribution to the integrated likelihood (9) for the first two terms is given
? (x , x |θ , m ) – the likelihood function of the i.i.d. generative model (5) for the
by f[X]
i
i i i
interval-valued observation [xi ], with the density function of each (conditionally) i.i.d.
latent data points Xi,1:mi given by f (xi,k |θi ) – and where π(θ|α) is the mixing distribution for θi given the parameter α. Given such interpretation, f (xi,k |θi ) (or θi ) is the local
density function (or parameter) for [Xi ], while π(θ|α) (or α) is the global density function
(or parameter) among all intervals. Therefore, the intra-interval structure is described
by the local density function and m, while the inter-interval variability is modelled by
the global density function.
As a result, inference on this model permits direct analysis of the underlying distribution of data points X1:m within each interval [Xi ] and its model parameter θi – an
advantageous property of the generative model over the descriptive model. For example
if the global density π(θ|α) works as the prior distribution, in the Bayesian framework,
? (x , x |θ , m )π(θ |α) is the posterior disfor the local parameter θi , π(θi |α, [xi ]) ∝ f[X]
i
i
i i i
tribution of the parameter of the local density f (x|θi ) underlying [xi ]. Similarly, the
posterior predictive
R distribution of latent data points underlying [xi ] is directly available
as π(x|α, [xi ]) ∝ f (x|θi )π(θi |α, [xi ]) dθi .
11
3.4
An asymptotic property of the generative model
Although they are constructed quite distinctly, it is possible to directly relate the descriptive and generative models under specific circumstances. In particular for standard
(descriptive) symbolic analysis techniques, when there is no prior knowledge on the distribution of data within an interval, this distribution is commonly assumed to be uniform
U (a, b) with a ≤ b (e.g. Le-Rademacher and Billard, 2011). Let I(x, x : a ≤ x ≤ x ≤ b)
be an indicator function of x and x, which equals 1 when a ≤ x ≤ x ≤ b, and 0
elsewhere. Defining f (x|θ) so that Xk ∼ U (a, b) for k = 1, . . . , m, and constructing
[X] = ϕ1,m (X1:m ) from the minimum and maximum values of these latent data points,
then the density function of [X] given by (7) becomes
??
(x, x|a, b, m) = m(m − 1)(x − x)m−2 (b − a)−m I(x, x : a ≤ x ≤ x ≤ b),
f[X]
which converges to a point mass at [a, b] as m → ∞ (Section 3.2). Then, by substituting
?? (x, x|a, b, m) into (9), the hierarchical generative model becomes
f[X]
ZZ
m(m − 1)
f[X] (x, x|m) =
{a≤x,b≥x}
(x − x)m−2
π(a, b) dadb.
(b − a)m
(10)
where π(θ|α) = π(a, b) describes the inter-interval parameter variability. When m is
large, the following theorem states that this hierarchical generative model converges to
π(x, x), which is a valid descriptive model.
Theorem 6. Suppose that [X] = ϕ1,m (X1:m ) with Xk ∼ U (a, b) for k = 1, . . . , m, and
the global density function π(a, b) is bounded, continuous and equal to 0 when a > b.
Then as m → ∞, the density function of [X] (10) converges to π(x, x) pointwise, i.e.
lim f[X] (x, x|m) = π(x, x).
m→∞
(Proof in Appendix B.6)
This result is interesting in that it reveals that descriptive models for [X] ∼ f[X] (x, x|θ)
described in Section 3.1 (e.g. Arroyo et al., 2010; Le-Rademacher and Billard, 2011; Brito
and Duarte Silva, 2012) actually possess an underlying and implicit generative structure.
Specifically, the sampling process of the descriptive model [X] ∼ f[X] (x, x) = π(x, x) can
be expressed via the generative process
[X] =
lim ϕ1,m (X1:m ),
m→∞
X1 , X2 . . . ∼ U (X ? , X ? ),
(X ? , X ? ) ∼ π(x, x),
That is, to obtain a sample realisation of [X], values of lower and upper bound parameters, (X ? , X ? ), of local uniform distribution are first drawn from the descriptive model
π(x, x), which in this case is exactly equivalent to the global density for the associated underlying hierarchical generative model. As the resulting infinite collection of latent data
12
points X1 , X2 , . . . ∼ U (X ? , X ? ) fully identifies the local density, and mink {Xk } = X ?
and maxk {Xk } = X ? are sufficient statistics for uniform distributions, the generated interval [X] is then determined as [X] = [X ? , X ? ] with (X ? , X ? ) ∼ π(x? , x? ). As a result,
there is no loss of information from the data aggregation procedure and the variation of
[X] is completely due to the variation permitted in the distribution of local parameters,
which is the global distribution. In this manner, the descriptive model is a special case
of and directly interpretable as a particular generative model.
This idea can be extended to a more general class of hierarchical generative models
in which the local distribution is only governed by location (µ) and scale (τ > 0) parameters, so that Xk ∼ f (x|µ, τ ) for k = 1, . . . , m. Suppose x and x are the l-th and u-th
order statistics, respectively. The associated values of µ and τ are available by solving
(
Q(l/(m + 1); µ, τ ) = x
,
(11)
Q(u/(m + 1); µ, τ ) = x
where Q(·; µ, τ ) is the quantile function of f (x|µ, τ ). If a unique solution exists for (11),
then f (x|µ, τ ) is an interval-identifiable local distribution.
We previously discussed that under the order statistic based data aggregation function, the i.i.d. generative model (6) will converge to a point mass as m → ∞. Similar
to Theorem 6, those hierarchical generative models (9) with interval-identifiable local
density functions f (x|µ, τ ) will also converge to descriptive models.
Theorem 7. Suppose that [X] = ϕl,u (X1:m ) with Xk ∼ f (x|µ, τ ) for k = 1, . . . , m, where
the local density function f (x|µ, τ ) is interval-identifiable with location parameter µ and
scale parameter τ > 0. Further suppose that l/(m + 1) → p > 0 and u/(m + 1) → p < 1
as m → ∞, and
i) the global density function π(µ, τ ) is twice differentiable;
ii) f (x|µ, τ ) is positive and continuous in neighbourhoods of Q(p; µ, τ ) and Q(p; µ, τ );
RR ?
iii)
|f[X] (x, x|µ, τ, m, l, u)|π(µ, τ ) dµdτ < ∞ for any 0 < l < u < m.
Then as m → ∞, the density function of [X] for the hierarchical generative model (9)
converges pointwise to
−1
π? (x, x) = π µ(x, x; p, p), τ (x, x; p, p) × J(µ(x, x; p, p), τ (x, x; p, p); p, p)
,
where µ(x, x; p, p) and τ (x, x; p, p) are the solution of (11) and
J(µ, τ ; p, p) =
∂
∂µ Q(p|µ, τ )
∂
∂µ Q(p|µ, τ )
(Proof in Appendix B.7)
13
∂
∂τ Q(p|µ, τ )
∂
∂τ Q(p|µ, τ )
!
.
In the specific case where f (x|a, b) is a U [a, b] local density function, with quantile
function Q(p|a, b) = (1 − p)a + pb, the hierarchical generative model (9) converges to the
distribution of
[(1 − p)X ? + pX ? , (1 − p)X ? + pX ? ],
where (X ? , X ? ) ∼ π(x, x).
4
Multivariate Models for Hypercubes
The p-dimensional equivalent of the univariate interval-valued random variable [X] is the
random p-hypercube, which corresponds to a p-tuple of random intervals. In specific,
we denote [x] = ([x1 ], . . . , [xp ]) ∈ Ip as a hypercube in the space of p-hypercubes, and
x = (x1 , . . . , xp ) ∈ Rp as one p-dimensional latent data point. It is straightforward to
extend the previous theory on distributions and likelihood functions for random intervals
(Sections 2 and 3) to random hypercubes.
4.1
Distribution and density functions
Let TIp be the metric topology induce by the Hausdorff metric on Ip . Due to the
isometry, hp ([x]) = (x1 , x1 , . . . , xp , xp ), between Ip and R2p , TIp is also the product
topology induced by TI . Accordingly, the product topology on Ip leads to the Borel
σ-algebra BIp = σ(TIp ) which is the smallest σ-algebra generated by the collection of all
{[x0 ] ⊆ [x]}. Suppose that d[x] = (d[x1 ], . . . , d[xp ]) is the infinitesimal neighbourhood of
[x]. On the measurable space (Ip , BIp ), the uniform product measure can be defined as
µIp (d[x]) =
p
Y
µI (d[xj ]),
j=1
which is the product of the uniform measure on I. The existence and uniqueness of µIp
is then guaranteed by Fubini’s theorem (Durrett, 2010). (See Appendix A.3 for more
details)
The containment distribution function of [X], denoted F[X] : R2p 7→ [0, 1], is a function on the real hyperplane, having similar properties to those described in Definition 1
(not stated here for clarity). The following theorems define the density function for [X].
Theorem 8. Let F[X] : R2p 7→ [0, 1] be the containment distribution function of a random
hypercube [X]. If F[X] is 2p-times differentiable, then the density function of [X] is
f[X] (x1 , x1 , . . . , xp , xp ) = (−1)p
∂ 2p
F (x , x1 , . . . , xp , xp ).
∂x1 ∂x1 . . . ∂xp ∂xp [X] 1
(Proof in Appendix B.8)
14
(12)
Theorem 9. A non-negative function f[X] : R2p 7→ R is the density function of a random
p-hypercube [X], with the containment distribution function
Z
xp
Z
x1
···
F[X] (x1 , x1 , . . . , xp , xp ) =
xp
f[X] (x01 , x01 , . . . , x0p , x0p ) dx01 dx01 . . . dx0p dx0p
x1
if and only if it satisfies
1. f[X] (x1 , x1 , . . . , xp , xp ) = 0 when xj > xj for at least one j;
R +∞ R +∞
R +∞ R +∞
2. −∞ −∞ · · · −∞ −∞ f[X] (x1 , x1 , . . . , xp , xp ) dx1 dx1 . . . dxp dxp = 1.
(Proof in Appendix B.8)
Note that f[X] (x1 , x1 , . . . , xp , xp ) is the joint density function of 2p-dimensional realvalued random variables with constraints. Similar to the random-interval case, this
suggests that a random p-hypercube can be equivalently modelled by a 2p-dimensional
real-valued random vector with the constraints xj ≤ xj for j = 1, . . . , p.
4.2
Likelihood functions and models
From Theorem 9, descriptive models for random p-hypercubes may be constructed
through direct specification of the 2p-dimensional density function f[X] (x1 , x1 , . . . , xp , xp ).
These models are easily constructed and simple to use, but have the same limitations as
the descriptive models for random intervals discussed in Section 3.1.
Containment distribution functions and likelihood functions of generative models
may be formulated using the same ideas as in (3) and (4). However, due to the necessity
of calculating 2p-th order mixed derivatives in (12), although intuitive, the structure
of the resulting likelihood functions would be highly complex, even for i.i.d. generative
models of random rectangles. The form of the likelihood function for an i.i.d. generative
model in the bivariate case [X] = [X1 ] × [X2 ] is presented in Appendix B.9.
The complex form of the likelihood function of an i.i.d. generative model accordingly induces a similarly complex hierarchical generative model. One option to produce
more tractable models is to impose a conditional independent
Qp structure within each pdimensional latent data point, so that xk ∼ f (x|θ1:p ) = j=1 f (xj |θj ). Consequently,
each random interval marginal distribution of the p-hypercube
Qp is? conditionally inde? (x , x , . . . , x , x |θ ) =
pendent of the others, i.e. f[X]
1 1
p p 1:p
j=1 f[Xj ] (xj , xj |θj ), where
?
f[Xj ] (xj , xj |θj ) is the likelihood function of the i.i.d. generative model (5) for [Xj ]. Although this choice will result in clear modelling consequences, the resulting likelihood
function for the hierarchical generative model
f[X] (x1 , x1 , . . . , xp , xp |m, α) =
Z Y
p
?
f[X
(xj , xj |θj )π(θ1:p |α) dθ1:p
j]
j=1
will only then require p second-order mixed derivatives.
15
(13)
In this scenario, dependencies between the random interval marginal distributions of
[X], such as temporal or spatial dependencies, are controlled only by the dependence
among local parameters θ1:p as introduced by the global distribution π(θ1:p |α). As a
result, beyond any a priori information on the joint distribution of the p-dimensional
latent data points underlying construction of the random interval [X] being incorporated
within π(θ1:p |α), it will be impossible to identify any further dependence based on the
observed p-hypercubes. If this is inadequate for a given analysis, the full multivariate
likelihood will need to be derived (e.g. Appendix B.9).
5
Applications
We illustrate our proposed approaches by firstly comparing the performance of the generative models to the existing descriptive models for simulated univariate (random interval) data. We then provide a generative model reanalysis of a real dataset of 5,000
credit card customers, as previously analysed by Brito and Duarte Silva (2012) using a
descriptive model.
5.1
Simulated data analysis
In order to provide a direct comparison between descriptive and generative models, we
construct our observed random intervals under the generative model as [xi ] = [xi , xi ],
where xi and xi are respectively the observed minimum and maximum values of xi1 , . . . , ximi
under the mixture model
xi1 , . . . , ximi
ci ∼
N (µc , σc2 )
U (ci − eτi , ci + eτi ),
∼
and τi ∼
(14)
N (µτ , στ2 ),
for i = 1, . . . , n. From Theorem 6, this hierarchical model is asymptotically equivalent
?
?
(as mi → ∞ for each i) to a descriptive model with [x?i ] = [c?i − eτi , c?i + eτi ], where
(c?i , τi? ) follows the same joint distribution as (ci , τi ). While in practice random intervals
will generally be constructed from different numbers of random samples, xi1 , . . . , ximi
(e.g. see Section 5.2), here we specify mi = m for all i = 1, . . . , n. In this analysis we will
compare the maximum likelihood estimators (MLEs) of parameters for both generative
and descriptive models obtained using data simulated from each model.
For each random interval [xi ] under the mixture model, the two-dimensional integration (9), with θ = (ci , τi ), can be reduced to a one-dimensional integration by first
integrating out ci , and then reparameterising to zi = m(τi − log 12 (xi − xi )). This leads
to the likelihood function of a single interval observation [xi ] given by
Z
0
∞
xi − xi
; µτ , στ2 )×
2
xi − xi m−1 zi
xi − xi m−1 zi
{Φ(xi +
e
; µc , σc2 ) − Φ(xi −
e
; µc , σc2 )} dzi , (15)
2
2
(xi − xi )−2 (m − 1)e−zi φ(m−1 zi + log
16
(b) Negative log likelihood
10
5
2
Value
2.4
0.000
2.6
0.005
2.8
3.0
3.2
0.020
0.015
0.010
Integrand
Generative
Descriptive
3.4
0.025
100
50
25
3.6
0.030
(a) Forms of integrands
0
2
4
6
8
10
0
zi
20
40
60
80
100
m
Figure 1: (a) Forms of the integrand in (15), with m = 2, 5, 10, 25, 50, 100, as a function of
zi , and (b) the negative log likelihood function as a function of m, when xi = −1, xi = 1,
µc = µτ = 0 and σc2 = στ2 = 1.
where φ and Φ respectively denote the Gaussian density and distribution function. This
form may be quickly and accurately approximated by Gauss-Laguerre quadrature methods (e.g. Evans and Swartz, 2000). The form of the integrand in (15) for varying m and
the resulting negative log-likelihood function is shown in Figure 1 for xi = −1, xi = 1,
µc = µτ = 0 and σc2 = στ2 = 1. These plots illustrate the convergence of the generative
model to the descriptive model as m gets large (Theorem 6), with only very minor differences observed for m > 30, and also suggest (panel (a)) that quadrature integration
methods will be accurate with around 20 nodes.
We simulate 1,000 replicate datasets, each comprising n = 100 intervals, from the
descriptive model with c?i , τi? ∼ N (0, 1) for i = 1, . . . , n (i.e. µc = µτ = 0 and σc2 =
στ2 = 1). MLEs of the model parameters (µc , µτ , σc2 , στ2 ) are obtained from fitting both
descriptive and generative models, with the latter assuming a specified number of latent
variables, m. Note that in practice, the number of latent variables, m, will typically be
known (and finite). The first column of Figure 2 illustrates the differences between the
(D)
(G)
resulting descriptive and generative model parameter MLEs (e.g. µ̂c − µ̂c , where
the superscripts indicate parameters of the descriptive (D) and generative (G) models),
with the solid line indicating the mean and the dotted lines the central 95% interval,
computed over the 1,000 replicates.
Firstly, we notice that the difference between the estimates is large for small m, and
becomes gradually smaller as m increases. This is not surprising as in this model specification, the generative model approaches the descriptive mode as m → ∞. However, as
both models are identically centred, the mean difference between the location parameter
17
0.05
−0.15
−0.05
^ (D) − µ
^ (G)
µ
c
c
0.05
−0.05
−0.15
^ (D) − µ
^ (G)
µ
c
c
0.15
(b) Generative model data
0.15
(a) Descriptive model data
20
40
60
80
100
20
40
100
60
80
100
60
80
100
60
80
100
0.1
−0.1 0.0
−0.5
40
60
80
100
20
40
0.4
0.3
0.2
0.1
0.0
0.0
0.1
0.2
^ 2(D) − σ
^ 2(G)
σ
c
c
0.3
0.4
0.5
m
0.5
m
^ 2(D) − σ
^ 2(G)
σ
c
c
80
−0.3
^ (D) − µ
^ (G)
µ
τ
τ
−0.1 0.0
−0.3
−0.5
^ (D) − µ
^ (G)
µ
τ
τ
20
20
40
60
80
100
20
40
0.15
0.10
0.05
−0.05 0.00
−0.05 0.00
0.05
0.10
^ 2(D) − σ
^ 2(G)
σ
τ
τ
0.15
0.20
m
0.20
m
^ 2(D) − σ
^ 2(G)
σ
τ
τ
60
m
0.1
m
20
40
60
80
100
20
m
40
m
Figure 2: Differences between MLEs of the descriptive model and the generative hierarchical
model, based on data generated from each model (left column = descriptive model data, right
column = generative model data), as a function of m = 5, . . . , 100, the number of latent data
points xi1 , . . . , xim in the generative model. Lines indicate the MLE means (solid lines) and 2.5%
and 97.5% quantiles (dashed lines) based on 1,000 replicate datasets.
18
(D)
(G)
estimates µ̂c and µ̂c is zero, regardless of the number of latent variables.
An obvious area of difference is that the point estimates of the interval half-range
(modelled by µτ ) are much smaller for the (correct) descriptive model than for the
generative model. This occurs as, the expected range of xi1 , . . . , xim under a generative
model is lower for small m than it is for large m. As a result, the generative model will
determine that µτ should be sufficiently larger for small m than it would be for large
m, given the same observed [xi , xi ]. That is, if the data are truly generated from the
descriptive model, parameters estimated from the generative model are effectively biased
for any finite m, and overestimate the true model parameters, with the magnitude of
the bias determined by the assumed value of m. Of course, this bias can be reduced by
setting m to be large in this case.
The second area of difference is that the estimated variability of the point estimates
of interval location and scale (σ̂c2 and σ̂τ2 ) is higher under the descriptive model than
under the generative model. This occurs as the generative model assumes that the
x +x
variability of e.g. i 2 i comprises both the variability of the latent data xi1 , . . . , xm
within interval i, in addition to the variability of interval locations ci between intervals.
2(D)
Under the descriptive model, this first source of variability is zero, and therefore σ̂c
2(G)
2(D)
will always be greater than σ̂c
for finite m. Similar reasoning explains why σ̂τ
is
2(G)
always greater than σ̂τ .
The second column of Figure 2 shows the same output as the first column, but
based on data simulated from the generative model with the same parameter settings
as before, and for varying (true) numbers of latent data points m = 5, . . . , 100. The
results are similar to before, except critically with the interpretation that the generative
model with fixed m is now correct. This means that, for example, if intervals are
constructed using the generative process (which is the most likely scenario in practice)
but are then analysed with a descriptive model, the point estimates of interval range
(µτ ) can be substantially underestimated by assuming m → ∞ under the descriptive
model, when in fact m is small and finite. Similarly, the estimates of σc2 and στ2 will
always be overestimated when assuming an incorrect descriptive model. These scenarios
will obviously be problematic for data analysis in practice.
The takeaway message of this analysis is that it is important to fit the model (descriptive or generative) that matches the interval (or p-hypercube) construction process.
Failure to do so can result in misinterpretation of model parameters, resulting in severe
biases in parameter estimates, which can then detrimentally impact on an analysis. In
practice, intervals tend to be constructed from underlying classical data (e.g. see Section
5.2), using a known process and where m is also known. This implies that the generative
model is a more natural construction than the descriptive model, and with parameters
that more directly relate to the observed data.
While this analysis has assumed uniformity of the generative process (14) in order that the descriptive model is obtained as m → ∞, and hence that the parameter
estimates between the two models can be directly compared, the same principles of interpretation and bias occur regardless of the generative model. The parameters are simply
less directly comparable with each other.
19
(a): mia = 5 individuals
Credit card customers
Log debt
0
−8
−4
−6
−4
−2
−2
Log debt
0
2
4
2
6
(a)
(b)
(c)
3
4
5
6
7
8
1
2
3
4
Log income
Log income
(b): mib = 28 individuals
(c): mic = 56 individuals
5
6
0
Log debt
−4
−6
−4
−2
−2
Log debt
0
2
2
4
2
4
1
2
3
4
5
6
7
2
Log income
3
4
5
6
Log income
Figure 3: Log income and log credit card debt in thousands of US$ for 5, 000 customers. Panels
illustrate rectangle-valued observations constructed from three groups of customers comprising
(a) mia = 5, (b) mib = 28 and (c) mic = 56 individuals. The contours in the last three panels indicate the predictive distributions of individuals for each group conditional on the corresponding
rectangle-valued observations, based on the generative model.
5.2
Analysis of credit card data
The data (available in the SPSS package customer.dbase) comprise log income and log
credit card debt in thousands of US$ of 5,000 credit card customers. In a previous analysis using descriptive models by Brito and Duarte Silva (2012), these data were aggregated
into random bivariate rectangles by stratifying individuals according to gender, age category (18-24, 25-34, 35-49, 50-64, 65+ years old), level of education (did not complete
high school, high-school degree, some college, college degree, undergraduate degree+),
and designation of primary credit card (none, gold, platinum, other). This leads to 192
non-empty groups, each producing a random rectangle [xi1 ] × [xi2 ] constructed by the
intervals bounded by the minimum and maximum observed values on log income and
log credit card debt.
The data are illustrated in Figure 3, along with the underlying data and constructed
20
random rectangles for three of the 192 groups, containing (a) mia = 5, (b) mib = 28 and
(c) mic =56 individuals. The number of individuals in all groups varies greatly (from 5 to
56), and it is noticeable that the distribution of individuals within each group comes from
a non-uniform distribution. As a result, the usual uniformity assumption of descriptive
models for random rectangles is clearly inappropriate. The generative model is more
suited to dealing with these heterogeneous rectangle-valued data containing complex
intra-rectangle structures.
Given the clear non-uniformity within each group i, we assume that the underlying
data are Gaussian with group-specific means and covariances. That is
(xi1 , xi2 ) ∼ N2 (µi , Σi )
2 , σ 2 ). Note that we choose
for i = 1, . . . , n = 192, where µi = (µi1 , µi2 ) and Σi = diag(σi1
i2
to model log income and log credit card debt as uncorrelated, despite there being some
visual evidence of positive correlation in the data underlying each random rectangle. It
is worth briefly explaining this decision in detail. For a small number of latent data
points mi , it is possible for a single point to determine both upper (or lower) ranges of
the random rectangle, and the probability of this occurring increases as the correlation
of the underlying data increases. So in principle, there is some information about the
correlation structure of the underlying data available through the associated random
rectangle. However, for groups with larger mi , the upper and lower ranges of the random
rectangles are more likely to be determined by four individual data points, in which case
it is not then possible to discern the underlying correlation structure. Although we
have several groups with small numbers of latent data points (e.g. mia = 5), in principle
allowing their correlation to be estimated, note that the same random rectangles will arise
whether the latent data are positively or negatively correlated. That is, the correlation
parameter is non-identifiable from the observed rectangle data. As such, we proceed
without attempting to estimate this parameter, despite information on the magnitude
of the correlation being available in principle for some groups.
We model the group-specific (local) parameters as
(µi1 , µi2 ) ∼ N2 (θ1 , θ2 , λ21 , λ22 , ρµ )
2
log σij
∼ N (ηj , 2j )
(16)
for j = 1, 2 and i = 1, . . . , 192. The integration in the generative model (13) is achieved
using Gauss-Hermite quadratures with 204 nodes to integrate over the four parameters.
Maximum likelihood estimates and 95% confidence intervals for each model parameter are illustrated in Table 1 for both generative and descriptive models. Similar to
the results for the simulated examples, the point estimates of location (θ1 and θ2 ) are
broadly insensitive to the choice of model, however the estimated values for many of
other parameters differ between the two models. Most importantly, the estimated values of ρµ are considerably larger for the generative model (ρ̂µ = 0.9040) compared to
the descriptive model (ρ̂µ = 0.5695). While both of these indicate a positive relationship between income and credit card debt, which is evident in the underlying data in
21
MLE
Generative
95% CI
MLE
Descriptive
95% CI
θ1
λ21
θ2
λ22
ρµ
η1
21
η2
22
3.76
3.70
3.82
0.13
0.10
0.17
-0.36
-0.44
-0.26
0.21
0.13
0.29
0.90
0.83
0.98
-1.20
-1.31
-1.09
0.48
0.35
0.61
0.41
0.34
0.48
0.09
0.04
0.13
3.79
3.74
3.85
0.17
0.13
0.20
-0.42
-0.53
-0.32
0.52
0.41
0.62
0.57
0.47
0.67
0.02
-0.04
0.08
0.20
0.16
0.24
0.82
0.78
0.87
0.09
0.07
0.11
Table 1: Maximum likelihood estimates and 95% asymptotic confidence intervals for the parameters of the generative and descriptive models for the credit card dataset.
Figure 3, there is a clear difference in the strength of that relationship. The descriptive
model results in a weaker estimated value in the correlation because it does not take the
noisy data generating process into account. While we suspect that the generative model
may be the more accurate of the two given the generative procedure used to construct
the random rectangles, in terms of drawing inferential conclusions about the underlying
data, it is critical that we are certain in this regard.
2 ) for each
For the generative models, the distribution of the local parameters (µij , σij
rectangle-valued observation can be computed by empirical Bayesian methods (previously these parameters were integrated out for the optimisation in Table 1). The prior
distribution for the local parameters is the global distribution (16) with its parameter
values given by the estimates in Table 1, and the likelihood function is the local density function of one observed rectangle. The resulting marginal posterior distributions
for the parameters of the observed rectangles (a)–(c) (Figure 3) are shown in Figure 4.
Compared to the prior (solid line) the parameters are well informed, even for rectangle
(a) with mia = 5 observations, with the level of precision increasing with the number of
individuals within each rectangle.
Goodness-of-fit for both descriptive and generative models can be evaluated through
model predictive distributions of random rectangles, in addition to predictive distributions for individual data points for the generative model. In the latter case, based on the
posterior distributions of the local parameters in Figure 4, the predictive distributions of
individual data points within the random intervals (a)–(c), conditional on observing the
associated random interval, are shown in Figure 3. While the predictive distributions
are marginally independent due to the model specification, their coverage describes the
observed data well. For group (a) the predictive distribution covers a wider region than
the observed rectangle, as this rectangle is constructed from only 5 individuals. As the
number of individuals increases in groups (b) and (c), the predictive regions more closely
represent the region of the observed rectangle, indicating that the generative model has
the ability to correctly account for the different numbers of individuals used to construct
each rectangle. The predictive distribution for group (b) individuals also indicates some
robustness to the two outliers that completely define the observed rectangle. This occurs
as the model correctly accounts for the fact that rectangle (b) is constructed from half
22
2.5
2.0
1.5
0.0
0.5
1.0
Density
1.5
0.0
0.5
1.0
Density
2.0
2.5
(a)
(b)
(c)
2
3
4
5
6
−3
−2
−1
1
2
1
2
3
2.5
2.0
Density
0.0
0.5
1.0
1.5
2.0
1.5
0.0
0.5
1.0
Density
0
µi2
2.5
µi1
−4
−3
−2
−1
0
1
2
−2
log(σi21)
−1
0
log(σi22)
2
Figure 4: Estimated marginal posterior distributions of the local parameters µi1 , µi2 , σi1
and
2
σi2
associated with the three groups (a)–(c) shown in Figure 3. Solid lines correspond to the
prior distributions for local parameters
the number of observations used to construct the rectangle of group (c), even though
both rectangles are roughly the same size.
The predictive distributions of random rectangles for groups (a)–(c) are illustrated in
Figure 5 for both descriptive (dashed lines) and generative (solid lines) models. Shown
are the bivariate predictive distributions of interval centre and log half-range, for both
log income (top row) and log debt (bottom row). The dot indicates the observed interval.
Under the generative model, these distributions are obtained directly from the predictive
distributions for individuals (Figure 3).
In all cases, the predictive distributions of the generative model more accurately, and
more precisely identify the location of the observed data. This is particularly the case
in group (a) in which the descriptive model is clearly indicating a lack of model fit. The
predicted interval for log debt in group (b) is not fully centred on the observed interval,
as the model attempts to account for the unlikely (under the model) construction of
the observed interval by outliers (Figure 3). However, the observed data are still well
23
(b) mib = 28: Log income
(c) mic = 56: Log income
0.5
0.0
Log half−range: log(y1r)
−1.0
−0.5
0.5
−0.5
0.0
Log half−range: log(y1r)
0.0
−0.5
−1.0
−1.0
−1.5
Log half−range: log(y1r)
0.5
1.0
1.0
1.0
(a) mia = 5: Log income
3.0
3.5
4.0
4.5
3.0
3.5
4.0
4.5
3.0
3.5
4.0
Interval centre: y1c
Interval centre: y1c
Interval centre: y1c
(a) mia = 5: Log debt
(b) mib = 28: Log debt
(c) mic = 56: Log debt
4.5
1.0
0.8
Log half−range: log(y2r)
0.4
0.6
1.0
Log half−range: log(y2r)
0.5
1.0
0.5
0.0
−2.0
−1.5
−1.0
−0.5
0.0
Interval centre: y2c
0.5
1.0
0.2
0.0
−0.5
Log half−range: log(y2r)
1.2
1.5
1.5
1.4
2.5
−2.0
−1.5
−1.0
−0.5
0.0
Interval centre: y2c
0.5
1.0
−2.0
−1.5
−1.0
−0.5
0.0
0.5
1.0
Interval centre: y2c
Figure 5: Posterior predictive distribution of a random rectangle [y1 ]×[y2 ] for each of the groups
(a)–(c) (left column to right) in Figure 3. Columns illustrate the marginal random intervals of
log income ([y1 ], top row) and log debt ([y2 ], bottom row) with each interval [yj ] = [yj1 , yj2 ]
expressed in interval centre and half-range form (yjc , yjr ) = ((yj1 + yj2 )/2, (yj2 − yj1 )/2) for
j = 1, 2. Solid and dashed lines indicate predictive distributions of generative and descriptive
models, respectively. The dot indicates the observed interval [xi1 ] × [xi2 ] used for model fitting.
predicted under the generative model. The overall fit to the observed data is better under
the generative model than the descriptive model, indicating that it more accurately
describes the complexities of the observed data.
6
Discussion
Current techniques for modelling random intervals (and random p-hypercubes) are based
on constructing models directly at the level of the interval-valued data (e.g. Arroyo
et al., 2010; Le-Rademacher and Billard, 2011; Brito and Duarte Silva, 2012). These
approaches are additionally based on the assumption that the unobserved individual
data points from which the interval is constructed are uniformly distributed within the
interval. As we have demonstrated in Section 5, using these descriptive methods when
the data are constructed from an underlying individual data points, which is typical
in practical applications, can result in misleading and biased parameter estimates and
therefore unreliable inferences.
In this article we have established the distribution theory for interval-valued random
24
variables which are constructed bottom-up from distributions of latent real-valued data
and aggregation functions used to construct the random intervals. These generative
models explicitly permit the fitting of standard statistical models for latent data points
through likelihood-based techniques, while accounting for the manner in which the observed interval-valued data are constructed. This approach directly accounts for the
non-uniformity of latent individual data points within intervals, and provides a natural
way to handle the differing number of latent data points mi within each random interval,
which is again typical in practice.
By deriving a descriptive model as the limiting case of a generative model (i.e. as
mi → ∞ for each i), we have demonstrated that these descriptive models have an explicit
underlying generative model interpretation. In turn this indicates why inferences from
descriptive models may be potentially misleading in practice.
In order to evaluate the integrated generative likelihood function (13) for the unimodal distributions considered in Section 5, we have used Gaussian quadrature methods.
This technique will be less useful when integrating over more than 6 parameters (Evans
and Swartz, 1995), or when there are strong dependencies between local parameters. In
these cases, approximate MLE’s can be obtained using e.g. Monte Carlo maximum likelihood estimation (Geyer and Thompson, 1992) or Monte Carlo expectation maximization
techniques (Wei and Tanner, 1990), or in the Bayesian framework Gibbs sampling (Geman and Geman, 1984) or pseudo-marginal Monte Carlo methods (Andrieu and Roberts,
2009).
In order to construct the likelihood function (13) for p-hypercubes we assumed independence among all margins in local distributions to avoid the 2p-th order mixeddifferentiation of F[X] (x1 , x1 , . . . , xp , xp ). Although this differentiation may be achieved
using symbolic computation software, the resulting likelihood functions are complex even
when p = 2 (see Appendix B.9), and the alternative of numerical differentiation would
be highly computational. However, this independence assumption does not hold if there
is priori information on the dependence structure within each latent data point x. As
pointed out by Billard and Diday (2006), this is often the case because the structure
of symbolic data might determine inherent dependencies such as logical, taxonomic and
hierarchical dependencies, but not statistical dependencies. In the generative model,
those dependencies as well as statistical dependencies can be addressed simultaneously
through the local distribution function f (x|θ). However, without the marginal independence assumption, inference for these models can be challenging.
Finally, while our examples have focused on minimum and maximum based data aggregation functions ϕ(x1:m ), there is clear interest in parameter estimation and inference
for more robust order-based functions ϕl,u (x1:m ), as the resulting intervals will be less
sensitive to outliers. The procedures for constructing the associated likelihood functions
are analogous to those presented here, and Theorem 7 provides their limiting descriptive model counterpart. An additional practical question for inference using order-based
aggregation functions is which order-based statistics to use. As this choice will impact
on the efficiency of the resulting inference, it is an open question to understand what
method of random interval construction would be optimal for any given analysis.
25
Acknowledgements
Xin Zhang is supported by the China Scholarship Council. Scott A. Sisson is supported
by the Australian Research Council (CE140100049).
References
Aldous, D. J. (1985). Exchangeability and related topics. In École d’Été de Probabilités
de Saint-Flour XIII — 1983, Volume 1117, pp. 1–198. Springer Berlin Heidelberg.
Andrieu, C. and G. O. Roberts (2009, 04). The pseudo-marginal approach for efficient
Monte Carlo computations. The Annals of Statistics 37 (2), 697–725.
Arroyo, J., R. Espı́nola, and C. Maté (2010). Different approaches to forecast interval
time series: A comparison in finance. Computational Economics 37 (2), 169–191.
Beresteanu, A., I. Molchanov, and F. Molinari (2012). Partial identification using random
set theory. Journal of Econometrics 166 (1), 17–32.
Beresteanu, A. and F. Molinari (2008). Asymptotic properties for a class of partially
identified models. Econometrica 76 (4), 763–814.
Billard, L. (2007). Dependencies and variation components of symbolic interval-valued
data. In Selected Contributions in Data Analysis and Classification, pp. 3–12. Springer
Berlin Heidelberg.
Billard, L. and E. Diday (2003). From the statistics of data to the statistics of knowledge:
Symbolic data analysis. Journal of the American Statistical Association 98 (462), 470–
487.
Billard, L. and E. Diday (2006). Symbolic data analysis: Conceptual statistics and data
mining. John Wiley & Sons.
Brito, P. and A. Duarte Silva (2012). Modelling interval data with Normal and SkewNormal distributions. Journal of Applied Statistics 39 (1), 3–20.
Diday, E. (1987). Introduction a e’approche symbolique en analyse des donnees. Premiere
Journées Symbolique-Numerique CEREMADE, Universite Paris - Dauphine, 21–56.
Domingues, M. A. O., R. M. C. R. de Souza, and F. J. A. Cysneiros (2010). A robust
method for linear regression of symbolic interval data. Pattern Recognition Letters 31,
1991–1996.
Durrett, R. (2010). Probability: Theory and Examples. Cambridge University Press.
Evans, M. and T. Swartz (1995, 08). Methods for approximating integrals in Statistics
with special emphasis on Bayesian integration problems. Statistical Science 10 (3),
254–272.
26
Evans, M. and T. Swartz (2000). Approximating integrals via Monte Carlo and deterministic methods. Oxford: Oxford University Press.
Fisher, R., R. A. O’Leary, S. Low-Choy, K. Mengersen, N. Knowlton, and R. E. B. amd
M. J. Caley (2015). Species richness on coral reefs and the pursuit of convergent global
estimates. Current Biology 25, 500–505.
Geman, S. and D. Geman (1984, Nov). Stochastic relaxation, Gibbs distributions, and
the Bayesian restoration of images. Pattern Analysis and Machine Intelligence, IEEE
Transactions on 6 (6), 721–741.
Geyer, C. J. and E. A. Thompson (1992). Constrained Monte Carlo maximum likelihood
for dependent data. Journal of the Royal Statistical Society. Series B (Methodological) 54 (3), 657–699.
Le-Rademacher, J. and L. Billard (2011). Likelihood functions and some maximum
likelihood estimators for symbolic data. Journal of Statistical Planning and Inference 141 (4), 1593–1602.
Lyashenko, N. (1983). Statistics of random compacts in Euclidean space. Journal of
Soviet Mathematics 21 (1), 76–92.
Matheron, G. (1975). Random sets and integral geometry. Wiley.
McLachlan, G. J. and P. N. Jones (1988). Fitting mixture models to grouped and
truncated data via the EM algorithm. Biometrics 44, 571–578.
Molchanov, I. (2005). Theory of random sets. Springer Science & Business Media.
Molchanov, I. and F. Molinari (2014). Applications of random set theory in Econometrics. Annual Review of Economics 6 (1), 229–251.
Moore, R. (1966). Interval analysis, Volume 22. Prentice-Hall.
Munkres, J. R. (2000). Topology. Prentice Hall, Incorporated.
Neto, E. A. L. and F. A. T. Carvalho (2010). Constrained linear regression models for
symbolic interval-valued variables. Computational Statistics and Data Analysis 54,
333–347.
Noirhomme-Fraiture, M. and P. Brito (2011). Far beyond the classical data models:
Symbolic data analysis. Statistical Analysis and Data Mining 4 (2), 157–170.
Reiss, R.-D. (1989). Approximate distributions of order statistics.
Sun, Y. and D. Ralescu (2014). A normal hierarchical model and minimum contrast
estimation for random intervals. Annals of the Institute of Statistical Mathematics,
1–21.
27
Vardeman, S. B. and C.-S. Lee (2005). Likelihood-based statistical estimation from
quantised data. IEEE Transactions on Instrumentation and Measurement 54, 409–
414.
Wei, G. C. G. and M. A. Tanner (1990). A Monte Carlo implementation of the EM
algorithm and the poor man’s data augmentation algorithms. Journal of the American
Statistical Association 85 (411), 699–704.
Xu, W. (2010). Symbolic data analysis: Interval-valued data regression. Ph. D. thesis,
University of Georgia, USA.
A
Topology
A.1
The basis
The basis of the standard topology on R2 is the collection of all open rectangles. Its
subspace topology induced by {(x, y) : x ≤ y}, as shown in Figure 6, has the basis of
which each element is the remaining part of a open rectangle on the top-left half plane.
Therefore, the collection of their counterparts on I via the isometry, h([x]) = (x, x), is
the basis of TI .
The open subset of I corresponding to the rectangle (a) in Figure 6 is
B([a], [b]) = {[x] : b < x < a ≤ a < x < b}.
This is the collection of all intervals for which the lower bounds are bounded between
a and b, while the upper bounds are bounded between a and b. The open subset of I
corresponding to the triangle (b) is
W ([c]) = {[x] : c < x ≤ x < c}.
This is the collection of all intervals for which the lower bounds are greater than c, while
the upper bounds are smaller than c.
Lemma 1. Suppose that E is the collection of all B([a], [b]) and W ([c]). Then E is a
basis for TI .
Proof. We first show that E is a basis for a topology. Note that for any [x] ∈ I, there
exists at least one E ∈ E s.t. [x] ∈ E. Then, we show in the following that for any
E1 , E2 ∈ E, if [x] ∈ E1 ∩ E2 , then there exists E3 ∈ E s.t. [x] ∈ E3 and E3 ⊂ E1 ∩ E2 .
Note that ∨ and ∧ take the maximum and the minimum of two operands, respectively.
0
i) Consider [x] ∈ B([a], [b]) ∩ B([a0 ], [b0 ]) 6= ∅. Then b ∨ b0 < a ∧ a0 and a ∨ a0 < b ∧ b .
From [x] ∈ B([a], [b]), we have that b < x < a ≤ a < x < b. From [x] ∈ B([a0 ], [b0 ]),
0
we have that b0 < x < a0 ≤ a0 < x < b . Therefore, b ∨ b0 < x < a ∧ a0 and
0
a ∨ a0 < x < b ∧ b . There exists [a00 ], [b00 ] ∈ I s.t. b ∨ b0 < b00 < x < a00 < a ∧ a0
00
0
and a ∨ a0 < a00 < x < b < b ∧ b . That is [x] ∈ B([a00 ], [b00 ]) and B([a00 ], [b00 ]) ⊂
B([a], [b]) ∩ B([a0 ], [b0 ]).
28
x
x=x
(c)
(a)
(b)
x
O
Figure 6: B([a], [b]) and W {[a, b]} are (a) and (b), respectively. (a), (b) and (c) constitute the
basis of TI .
ii) Consider [x] ∈ W ([c1 ]) ∩ W ([c2 ]) 6= ∅. Then c1 ∨ c2 < c1 ∧ c2 . From [x] ∈ W ([c1 ]),
we have that c1 < x ≤ x < c1 . From [x] ∈ W ([c2 ]), we have that c2 < x ≤ x < c2 .
Therefore, c1 ∨ c2 < x ≤ x < c1 ∧ c2 . There exists [c] ∈ I s.t. c1 ∨ c2 < c < x ≤
x < c < c1 ∧ c2 . That is [x] ∈ W ([c]) and W ([c]) ⊂ W ([c1 ]) ∩ W ([c2 ]).
iii) Consider [x] ∈ B([a], [b]) ∩ W ([c]) 6= ∅. Then c < a and c > a. From [x] ∈
B([a], [b]), we have that b < x < a ≤ a < x < b. From [x] ∈ W ([c]), we have that
c < x ≤ x < c. Therefore, c ∨ b < x < a ≤ a < x < c ∧ b. There exists [a0 ], [b0 ] ∈ I
0
s.t. c ∨ b < b0 < x < a0 < a and a < a0 < x < b < c ∧ b. That is [x] ∈ B([a0 ], [b0 ])
and B([a0 ], [b0 ]) ⊂ B([a], [b]) ∩ W ([c]).
That is, E is a basis for a topology. Next, we show E is a basis for TI . Figure 6 shows
that the basis of TI consists of three types of subsets. As B([a], [b]) is an (a)-type subset
and W {[c]} is a (b)-type subset, the topology generated by E is coarser than TI . On
the other hand, for any [x] in a (c)-type subset, we can find at least one (a)-type subset
or (b)-type subset that contains that [x] and subsets of that (c)-type subset. Therefore,
the topology generated by E is finer than TI . In conclusion, the topology generated by
E is TI .
A.2
Properties of the topology
Lemma 2. B([a], [b]) = W ([b]) \ {[x] ⊆ [a, b]} ∪ {[x] ⊆ [b, a]} .
Proof. For any [x] ∈ B([a], [b]), i.e. b < x < a ≤ a < x < b, we have [x] ∈ W ([b]).
Also [x] * [a, b] and [x] * [b, a], i.e [x] ∈
/ {[x] ⊆ [a, b]} ∪ {[x] ⊆ [b, a]}. Therefore,
B([a], [b]) ⊆ W ([b]) \ {[x] ⊆ [a, b]} ∪ {[x] ⊆ [b, a]} .
On the other hand, for any [x] ∈ W ([b]) \ {[x] ⊆ [a, b]} ∪ {[x] ⊆ [b, a]} , we have
[x] ∈ W ([b]), i.e. b < x ≤ x < b. Also [x] * [a, b] and [x] * [b, a], i.e. x < a
and x > a. Hence b < x < a and
a < x < b, i.e. [x] ∈ B([a], [b]). Therefore,
W ([b]) \ {[x] ⊆ [a, b]} ∪ {[x] ⊆ [b, a]} ⊆ B([a], [b]).
In conclusion, B([a], [b]) = W ([b]) \ {[x] ⊆ [a, b]} ∪ {[x] ⊆ [b, a]} .
29
Lemma 3. TI is the smallest topology containing all W ([c]) and {[x] ⊆ [c]}c .
Proof. {[x] ⊆ [c]} is a closed subset, as it is the closure of W {[c]}. Accordingly its
complement {[x] ⊆ [c]}c is open, and thus {[x] ⊆ [c]}c ∈ TI . From Lemma 2, B([a], [b]) =
W ([b]) ∩ {[x] ⊆ [a, b]}c ∩ {[x] ⊆ [b, a]}c . So every element in E can be generated by set
operations over finite elements of W {[c]} and {[x] ⊆ [c]}c . As E is a basis of TI , every
element in TI can be generated by set operations over finite elements of W {[c]} and {[x] ⊆
[c]}c . Therefore, TI is the smallest topology containing W ([c]) and {[x] ⊆ [c]}c .
A.3
Hypercubes
Similarly, through the property of isometry, hp ([x]) = (x1 , x1 , . . . , xp , xp ), it can be
shown that a basis of the topology TIp is the collection of the following two classes of
subsets:
Bp ([a], [b]) = {[x] : bj < xj < aj ≤ aj < xj < bj , j = 1, . . . , p},
Wp ([c]) = {[x] : cj < xj ≤ xj < cj , j = 1, . . . , p}.
The next lemma shows an analogous result of Lemma 2.
Lemma 4. Bp ([a], [b]) = Wp ([b]) \ ∪pj=1 [{[x] ⊆ [aj1 ]} ∪ {[x] ⊆ [aj2 ]}], where
[aj1 ] =
[aj2 ] =
[a1 ], . . . , [aj , bj ], . . . , [ap ] ,
[a1 ], . . . , [bj , aj ], . . . , [ap ] .
Proof. Similar to the proof of Lemma (2).
Based on the above lemma, the hypercube’s version of Theorem 1 can be proved in
a similar way.
B
B.1
Proofs
Proof of Theorem 1
Proof. As an isometric embedding to the standard topology of the real plane, the topology TI is separable, and thus it has a countable basis. We define rational intervals [q] ∈ I
where q, q are rational numbers. Then, the collection of all rational intervals, IQ , is dense
in I.
We first show that EQ is a countable basis of TI . Let EQ be the collection of all
B([q1 ], [q2 ]) and W ([q]). As rational numbers are countable, EQ is countable. It can be
shown that EQ is a basis of a topology and its generated topology is TI in a similar way
to Lemma 1. As a result, EQ is a countable basis of TI .
Then we show that σ(F) = σ(EQ ). For any {[x0 ] ⊆ [x]} ∈ F, {[x0 ] ⊆ [x]} = W ([x])c
and W ([x]) ∈ TI can be generated by set operations over countable elements from
EQ , as EQ is a countable basis of TI . So, σ(F) ⊆ σ(EQ ). On the other hand, for any
30
0
W ([q]) ∈ EQ , we have W ([q]) = ∪∞
where q−q ≥ 2/k, and for
n=k {[q ] ⊆ [q+1/n, q−1/n]},
any B([q1 ], [q2 ]) ∈ EQ , we have B([q1 ], [q2 ]) = W ([q2 ]) \ {[q] ⊆ [q1 , q2 ]} ∪ {[q] ⊆ [q2 , q1 ]}
(Lemma 2). So, σ(EQ ) ⊆ σ(F).
That is, σ(F) = σ(EQ ) = σ(TI ).
B.2
Proof of Theorem 2
Proof. We let
B ? ([a], [b]) = {[x] : b < x ≤ a ≤ a ≤ x < b}.
(17)
In a way analogous to Lemma 2, we have
B ? ([a], [b]) = W ([b]) \ W ([a, b]) ∪ W ([b, a]) .
(18)
By the continuity of the measure,
∞
µI (W ([x])) = µI ( ∪{[x]0 ⊆ [x + 1/n, x − 1/n]})
= lim µI ({[x]0 ⊆ [x + 1/n, x − 1/n]})
n→∞
= lim
n→∞
1
1
(x − x − 2/n)2 = (x − x)2 .
2
2
Note that W ([a, b]) ∩ W ([b, a]) = W ([a]). We have
µI (B ? ([a], [b])) = µI (W ([b])) − µI (W ([a, b])) − µI (W ([b, a])) + µI (W ([a]))
= (a − b)(b − a).
Therefore, µI (d[x]) = µI (B ? ([x], [x − dx, x + dx])) = dx × dx.
B.3
Proof of Theorem 3
Proof. As BI = σ(F) (Theorem 1), any E ∈ BI can be generated by set operations
over at most countable elements from F. So, it’s probability measure P ([X] ∈ E) will
be available if P ([X] ⊆ [x]) is known for any [x]. Therefore, the uniqueness has been
proved.
Next, we prove the existence of a probability measure P[X] : BI 7→ [0, 1] on the
space of intervals s.t. P[X] ({[x0 ] ⊆ [x]}) = C[X] ([x]). Let G be the collection of all
B 0 ([x], [y]) = {[x0 ] : y ≤ x0 < x ≤ x < x0 ≤ y}. Similar to Lemma 2, we have B 0 ([x], [y]) =
{[x]0 ⊆ [y]}\ {[x]0 ⊆ [x, y]} ∪ {[x]0 ⊆ [y, x]} . Then, define H = F ∪G∪{∅, I}, and extend
C[X] (·) to a function PC (·) on H s.t. PC (∅) = 0, PC (I) = 1, PC ({[x0 ] ⊆ [x]}) = C[X] ([x])
and
PC (B 0 ([x], [y])) = C[X] ([y]) − C[X] ([x, y]) − C[X] ([y, x]) + C[X] ([x]) ≥ 0,
by condition iii) of the definition of C[X] (·) in Section 2.2. That is PC (·) is non-negative.
31
In addition as I is locally compact, for any A ⊂ I and δ > 0, there exists E1 , . . . , EN ∈
H with all µI (Ei ) ≤ δ, such that A ⊂ ∪N
i=1 Ei . Therefore, we can use Carathéodory con? (A) = lim
struction (Durrett, 2010) to define a metric outer measure. Let P[X]
δ→0 Pδ (A),
where
(∞
)
X
Pδ (A) = inf
PC (Ei ) : Ei ∈ H, diam(Ei ) ≤ δ, ∪∞
i=1 Ei ⊇ A ,
i=1
? (·) is a metric outer measure, and thus
where diam(Ei ) is the diameter of Ei . So, P[X]
? (·). That is, there exists a probability
the Borel subsets on I are measurable w.r.t. P[X]
? (E) for any E ∈ B .
measure P[X] : BI 7→ [0, 1], such that P[X] (E) = P[X]
I
Finally, we can check that P[X] ({[x0 ] ⊆ [x]}) = C[X] ([x]). For any n = 1, 2, . . .,
there exits δn → 0 as n → ∞, s.t. Pδn ({[x0 ] ⊆ [x]}) ≤ C[X] ([x − n1 , x + n1 ]). Also
Pδ ({[x0 ] ⊆ [x]}) ≥ C[X] ([x]) by definition for any δ > 0. Therefore
C[X] ([x]) ≤ lim Pδn ({[x0 ] ⊆ [x]}) ≤ lim C[X] ([x − 1/n, x + 1/n]).
n→∞
n→∞
By condition ii) of the definition of C[X] (·) in Section 2.2,
lim C[X] ([x − 1/n, x + 1/n]) = C[X] (∩∞
n=1 [x − 1/n, x + 1/n]) = C[X] ([x]).
n→∞
Therefore,
P[X] ({[x0 ] ⊆ [x]}) = lim Pδn ({[x0 ] ⊆ [x]}) = C[X] ([x]).
n→∞
As a result, given a random interval [X] : Ω 7→ I, we obtain a probability measure
P : σ([X]) 7→ [0, 1], s.t. P ([X] ⊆ [x]) = P[X] ({[x0 ] ⊆ [x]}) = C[X] ([x]).
B.4
Proof of Theorem 4
Proof. Let C[X] ([x]) = F[X] (x, x) be the containment functional. From Theorem 3 and
its proof in Appendix B.3, it determines a unique probability measure P[X] : BI 7→ [0, 1]
on the space of intervals s.t. P[X] ([x]) = F[X] (x, x). As d[x] = B? ([x], [x − dx, x + dx]),
from (17) and (18),
B? ([x], [x − dx, x + dx]) = W ([x − dx, x + dx]) \ [W ([x, x + dx]) ∪ W ([x − dx, x])] .
Therefore,
P[X] (d[x]) = P[X] (W ([x − dx, x + dx])) − P[X] (W ([x, x + dx]))−
P[X] (W ([x − dx, x])) + P[X] (W ([x, x])).
1
1
0
By the continuity of the measure and W ([x]) = ∪∞
n=k {[x] ⊆ [x + n , x − n ]},
P[X] (W ([x])) = 0lim
lim P[X] ([X] ⊆ [x]0 ) = 0lim
x →x+ x0 →x−
lim F[X] (x0 , x0 ).
x →x+ x0 →x−
As F[X] is twice differentiable (thus continuous), P[X] (W ([x])) = F[X] (x, x). Therefore,
P[X] (d[x]) = F[X] (x − dx, x + dx) − F[X] (x, x + dx) − F[X] (x − dx, x) + F[X] (x, x).
32
Substituting second order Taylor expansions for the first three terms in the above equation, we obtain
∂2
P[X] (d[x]) = −
F (x, x)dxdx + o(dxdx).
∂x∂x [X]
Note that µI (d[x]) = dxdx (Theorem 2), and so
P[X] (d[x]) = −
∂2
F (x, x)µI (d[x]) + o(µI (d[x])).
∂x∂x [X]
In addition, P[X] (d[x]) = 0 when µI (d[x]) = 0, i.e. P[X] (·) is absolute continuous w.r.t.
µI (·). Therefore the the Radon-Nikodym derivative exists and
P[X] (d[x])
∂2
=−
F (x, x).
µI (d[x])
∂x∂x [X]
B.5
Proof of Theorem 5
Proof. For any function f[X] (x, x) satisfying the conditions in the theorem, we can construct its containment distribution function F[X] (x, x) as
Z
xZ b
F[X] (x, x) =
Z
xZ x
f[X] (a, b) dbda.
f[X] (a, b) dadb or F[X] (x, x) =
x
x
x
a
It is easy to check that F[X] (x, x) satisfies the conditions in Definition 1.
B.6
Proof of Theorem 6
a+b
2
∈ (−∞, +∞) and r =
b−a
2
≥ 0. We can rewrite (10) as
ZZ
−m
m−2
r−m g(c, r) dcdr,
f[X] (x, x|m) = 2 m(m − 1)(x − x)
Proof. Let c =
A
where g(c, r) = 2π(c − r, c + r) is the density function of (c,Rr) and A = {(c, r) : x −
+∞
r ≤ c ≤ x + r, r ≥ x−x
2 }. As π(·) is bounded continuously, −∞ g(c, r) dc < ∞. Let
R +∞
g(r) = −∞ g(c, r) dc, B0 = {r : g(r) = 0} and B1 = {r : g(r) 6= 0}. When g(r) 6= 0, we
have g(c|r) = g(c,r)
g(r) . The above integration can be decomposed into the following two
cases. In the case that g(r) 6= 0, we replace g(c, r) with g(r)g(c|r) and integrate out c,
Z x+r
ZZ
Z ∞
−m
−m
r g(c, r) dcdr =
r g(r)
g(c|r) dc dr.
A∩B1
x−x
2
In the case that g(r) = 0, we have g(c, r) = 0 and
33
x−r
RR
A∩B0 r
−m g(c, r) dcdr
= 0.
Then, writing z = (m − 1)(log r − log x−x
2 ), we have
1
f[X] (x, x|m) = m(x − x)−1 ×
2
(Z x+ x−x e(m−1)−1 z
)
Z ∞
2
x − x (m−1)−1 z
x − x (m−1)−1 z
−z
e g
e
e
dc dz.
g c|
−1
2
2
x− x−x e(m−1) z
0
2
As π(·) is bounded continuously, g(c, r) = 2π(c − r, c + r) is bounded continuously. Due
to the mean value theorem, the above term can be simplified as
Z
o
1 ∞ n (m−1)−1 z
x − x (m−1)−1 z
f[X] (x, x|m) =
m e
− 1 e−z g(ξ,
e
) dz,
2 0
2
−1
−1
−1 z
x−x (m−1)
(m−1) z ≤ ξ ≤ x − x−x e(m−1) z . Let M (ξ) = sup
where x + x−x
z≥0 g(ξ, 2 e
2 e
2
M (ξ) is bounded as g(c, r) is bounded. When m ≥ 3, we have
Z
o
m
M (ξ) ∞ n (m−1)−1 z
3
f[X] (x, x|m) ≤
m e
− 1 e−z dz =
M (ξ) ≤ M (ξ).
2
2(m − 2)
2
0
).
Therefore, f[X] (x, x|m) is bounded when m → ∞, and thus
Z
n
o
x − x (m−1)−1 z
1 ∞
(m−1)−1 z
−z
lim f[X] (x, x|m) =
lim m e
− 1 e g ξ,
e
dz
m→∞
2 0 m→∞
2
x+x x−x
1
,
= π(x, x).
= g
2
2
2
B.7
Proof of Theorem 7
Proof. Let fµ,τ = f (·|µ, τ ), and denote Fµ,τ = F (·|µ, τ ) and Qµ,τ = Q(·; µ, τ ) as its
cumulative distribution function and quantile function, respectively. As fµ,τ is positive
and continuous in the neighbourhoods of Qµ,τ (p) and Qµ,τ (p) with p > 0 and p < 1, the
joint density function of
(
1
(m + 1) 2 fµ,τ (Qµ,τ (p))(X − Qµ,τ (p))
1
(m + 1) 2 fµ,τ (Qµ,τ (p))(X − Qµ,τ (p))
converges pointwise to a bivariate normal density function, with zero mean and covariance matrix
!
p(1 − p) p(1 − p)
Σ=
p(1 − p) p(1 − p)
when m → ∞ (Reiss, 1989). Thus when m is large, the density function of the i.i.d.
generative model (6) is asymptotically equivalent to
m+1
1
2π|Σ| 2
fµ,τ (Qµ,τ (p))fµ,τ (Qµ,τ (p)) exp{−(m + 1)T (x, x; µ, τ )},
34
where
1
T (x, x; µ, τ ) = (t(x; µ, τ ), t(x; µ, τ ))Σ−1 (t(x; µ, τ ), t(x; µ, τ ))| ,
2
t(x; µ, τ ) =fµ,τ (Qµ,τ (p))(x − Qµ,τ (p)),
t(x; µ, τ ) =fµ,τ (Qµ,τ (p))(x − Qµ,τ (p)).
That is, the density function of the hierarchical generative model (9) is asymptotically
equivalent to
m+1
(19)
1 × H(x, x; p, p, m)
2π|Σ| 2
where
ZZ
H(x, x; p, p, m) =
fµ,τ (Qµ,τ (p))fµ,τ (Qµ,τ (p))π(µ, τ ) exp{−(m + 1)T (x, x; µ, τ )} dµdτ.
Note that Σ is positive definite, and so T (x, x; µ, τ ) ≥ 0. Also T (x, x; p, p, m) reaches its
minimum 0, when Qµ,τ (p) = x and Qµ,τ (p) = x, i.e. µ and τ satisfy (11). As fµ,τ is
interval-identifiable, (11) has a unique solution, and thus T (x, x; p, p, m) is unimodal.
As µ? = µ(x, x; p, p) and τ? = τ (x, x; p, p) are the solution of (11), given conditions
i) and iii) in the theorem, a Laplace approximation can be applied to H(x, x; p, p, m) at
the point (µ? , τ? ), giving
1
H(x, x; p, p, m) ≈ 2π(m + 1)−1 |∇2 T (x, x; µ? , τ? )|− 2 fµ? ,τ? (x)fµ? ,τ? (x)π(µ? , τ? ).
We let T = T (x, x; µ, τ ), t = t(x; µ, τ ), t = t(x; µ, τ ) and Σ−1 =
(20)
!
a11 a12
, so we
a12 a22
2
have T = 21 (a11 t2 + 2a12 tt + a22 t ). The first order partial derivatives of T are
∂T
∂µ
∂T
∂τ
∂t
∂t
∂t
∂t
+ a12 t
+ a12 t
+ a22 t ,
∂µ
∂µ
∂µ
∂µ
∂t
∂t
∂t
∂t
= a11 t
+ a12 t
+ a12 t
+ a22 t .
∂τ
∂τ
∂τ
∂τ
= a11 t
?
Let T ? , t? and t denote the corresponding functions and their derivatives taking values
?
at (µ? , τ? ). As t? = t = 0, the second order partial derivatives at (µ? , τ? ) are
∂t?
∂µ
2
?
∂t ∂t?
+ 2a12
+ a22
∂µ ∂µ
? 2
∂2T ?
∂µ2
= a11
∂2T ?
∂τ 2
∂2T ?
∂µ∂r
? 2
?
∂t? 2
∂t ∂t?
∂t
= a11
+ 2a12
+ a22
,
∂τ
∂τ ∂τ
∂τ
?
?
?
?
∂t? ∂t?
∂t? ∂t
∂t? ∂t
∂t ∂t
= a11
+ a12
+ a12
+ a22
.
∂µ ∂τ
∂µ ∂τ
∂τ ∂µ
∂µ ∂τ
35
∂t
∂µ
,
Therefore, ∇2 T at (µ? , τ? ) is
∇2 T ? =
∂t?
∂µ
?
∂t
∂µ
∂t?
∂µ
∂t?
∂τ
?
∂t
∂µ
?
∂t
∂τ
!|
Σ−1
∂t? ∂t?
∂µ ∂τ
(µ? , τ ? ) are
and its determinant is |∇2 g| = |Σ|−1
The derivatives of t and t at
∂t?
∂τ
?
∂t
∂τ
−
∂t?
∂µ
?
∂t
∂µ
∂t? ∂t?
∂τ ∂µ
2
∂t?
∂τ
?
∂t
∂τ
!
,
.
∂
Qµ? ,τ ? (p),
∂µ
∂
= −fµ? ,τ ? (x) ×
Qµ? ,τ ? (p),
∂τ
∂
= −fµ? ,τ ? (x) ×
Qµ? ,τ ? (p),
∂µ
∂
= −fµ? ,τ ? (x) ×
Qµ? ,τ ? (p).
∂τ
= −fµ? ,τ ? (x) ×
and thus
2
|∇2 T ? | = |Σ|−1 fµ? ,τ ? (x)2 fµ? ,τ ? (x)2 J(µ? , τ ? ; p, p) ,
where
J(µ? , τ ? ; p, p) =
∂
? ?
∂µ Qµ ,τ (p)
∂
? ?
∂µ Qµ ,τ (p)
∂
? ?
∂τ Qµ ,τ (p)
∂
? ?
∂τ Qµ ,τ (p)
(21)
!
.
From (20) and (21), we obtain that the density function of the hierarchical generative
−1
model (19) converges pointwise to π(µ? , τ ? ) J(µ? , τ ? ; p, p) .
B.8
Proof of Theorem 8 and 9
Proof. Similar to the proof of Theorem 4 and 5. Use Lemma 4 and Taylor expansions.
B.9
Likelihood function of two dimension i.i.d. generative model
Let [X] = [X1 ] × [X2 ] be the random rectangle generated from m i.i.d. bivariate latent
data points from f (x1 , x2 |θ) with the data aggregation function taking the minimum
and maximum values at each margin. Let F (x1 , x2 |θ) be the distribution function of
f (x1 , x2 |θ). The distribution function of [X1 ] × [X2 ] is
F[X] (x1 , x1 , x2 , x2 |θ) = [F (x1 , x2 |θ) − F (x1 , x2 |θ) − F (x1 , x2 |θ) + F (x1 , x2 |θ)]m .
This is the probability that all m latent data points fall within the rectangle [x1 ] × [x2 ].
From Theorem 8, the likelihood function is the fourth order mixed derivative as shown
36
below
f[X] (x1 , x1 , x2 , x2 |θ)
=m(m − 1)(m − 2)(m − 3) {F (x1 , x2 |θ) − F (x1 , x2 |θ) − F (x1 , x2 |θ) + F (x1 , x2 |θ)}m−4 ×
Z x1
Z x2
Z x2
Z x1
f (y1 , x2 |θ) dy1
f (y2 , x2 |θ) dy2
f (x1 , y3 |θ) dy3
f (x1 , y4 |θ) dy4 +
x1
x1
x2
x2
m(m − 1)(m − 2) {F (x1 , x2 |θ) − F (x1 , x2 |θ) − F (x1 , x2 |θ) + F (x1 , x2 |θ)}m−3 ×
Z x1
Z x2
f (x1 , x2 |θ)
f (y2 , x2 |θ) dy2
f (x1 , y4 |θ) dy4 +
Z
x1
x1
f (x1 , x2 |θ)
Z
x2
x2
f (x1 , y4 |θ) dy4 +
f (y1 , x2 |θ) dy1
x1
Z x1
x2
Z x2
f (y2 , x2 |θ) dy2
f (x1 , x2 |θ)
x1
Z x1
f (x1 , x2 |θ)
f (x1 , y3 |θ) dy3 +
x2
Z x2
f (y1 , x2 |θ) dy1
x1
f (x1 , y3 |θ) dy3 +
x2
m(m − 1) {F (x1 , x2 |θ) − F (x1 , x2 |θ) − F (x1 , x2 |θ) + F (x1 , x2 |θ)}m−2 ×
f (x1 , x2 |θ)f (x1 , x2 |θ) + f (x1 , x2 |θ)f (x1 , x2 |θ) .
Although it is rather complex, in fact it has a similar intuitive interpretation to (7). The
first term denotes the case that m − 4 points fall within [x1 ] × [x2 ] while the remaining
four points are (y1 , x2 ), (y2 , x2 ), (x1 , y3 ) and (x1 , y4 ), where x1 ≤ y1 , y2 ≤ x1 and
x2 ≤ y3 , y4 ≤ x2 . The second term represents the case that m − 3 points fall within
[x1 ] × [x2 ] while the remaining three points determine the boundary of the rectangle.
The last terms is the case the boundary is formed by only two points.
37
| 10math.ST
|
arXiv:1702.00364v1 [cs.SE] 1 Feb 2017
EasyInterface: A toolkit for rapid
development of GUIs for research prototype tools∗
Jesús Doménech1 , Samir Genaim1 , Einar Broch Johnsen2 , and
Rudolf Schlatte2
1
Complutense University of Madrid, Madrid, Spain
2
University of Oslo, Oslo, Norway
Abstract
In this paper we describe EasyInterface, an open-source toolkit for
rapid development of web-based graphical user interfaces (GUIs). This
toolkit addresses the need of researchers to make their research prototype
tools available to the community, and integrating them in a common environment, rapidly and without being familiar with web programming or
GUI libraries in general. If a tool can be executed from a command-line
and its output goes to the standard output, then in few minutes one can
make it accessible via a web-interface or within Eclipse. Moreover, the
toolkit defines a text-based language that can be used to get more sophisticated GUIs, e.g., syntax highlighting, dialog boxes, user interactions, etc.
EasyInterface was originally developed for building a common frontend
for tools developed in the Envisage [4] project.
1
Introduction
During the lifetime of a research project, research prototype tools are often
developed which share many common aspects. For example, in the Envisage
[4] project, we developed various tools for processing ABS programs: static
analyzers, compilers, simulators, etc. Both as individual researchers and as
groups, we often develop several related tools over time to pursue a specific line
of research.
Providing the community with easy access to research prototype tools is
crucial to promote the research, get feedback, and increase the tools’ lifetime
beyond the duration of a specific project. This can be achieved by building
∗ This work was partially funded by the EU project FP7-ICT-610582 ENVISAGE: Engineering Virtualized Services, the Spanish MINECO projects TIN2012-38137 and TIN201569175-C4-2-R, and the CM project S2013/ICE-3006.
1
GUIs that facilitate trying tools; in particular, tools with web-interfaces can be
tried without the overhead of first downloading and installing the tools.
In practice, we typically avoid developing GUIs until tools are fairly stable. Since prototype tools change continuously, in particular during a research
project, they will often not be available to the community during early development. Both programming plug-ins for sophisticated frameworks such as Eclipse
Scout and building simpler GUIs from scratch are tedious tasks, in particular
for web-interfaces. It typically gets low priority when developing a research prototype. Often we opt for copying the GUI of one tool and modifying it to fit the
needs of a new related tool. Apart from code duplication, these tools will “live”
separately, although we might benefit from having them in a common GUI.
EasyInterface is a toolkit that aims at simplifying the process of building and maintaining GUIs for (but not limited to) research prototype tools.
Avoiding complex programming, it provides an easy, declarative way to make
existing (command-line) tools available via different environments such as a
web-interface, within Eclipse, etc. It also defines a text-based output language
that can be used to improve the way results are presented to the user without
requiring any particular knowledge of GUI/Web programming; e.g., if the output of a tool is (a structured version of) “highlight line number 10 of file ex.c”
and “when the user clicks on line 10, open a dialog box with the text ...”, the
web-interface will interpret this and convert it to corresponding visual effects.
An advantage of using such an output language is that it will be understood
by all the frontend environments of EasyInterface, e.g., the web-interface
and the Eclipse plug-in (which is still under development). EasyInterface is
open source and available at http://github.com/abstools/easyinterface.
A more detailed descreption on how to use EasyInterface is availale in Appendix A, and in the user manual [3]. An online demo is available at https:
//youtu.be/YE7YwR2duzk.
2
General Overview
The overall architecture of EasyInterface is depicted in Fig. 1. Its two main
components are (i) server side: a machine with several tools (the circles Tool1,
etc.) executable from the command-line, and with output going to the standard
output. These are the tools that we want to make available for the outside
world; and (ii) client side: several clients that communicate with the server to
execute a tool. Tools may run on the server machine or on other machines; e.g.,
the web-interface can be installed as a web-page on the server, and accessed
from anywhere with a web browser. Clients can connect to several servers
simultaneously.
The server side addresses the problem of providing a uniform way to remotely execute locally installed tools. This problem is solved by the server, which
consists of PHP programs (on top of an HTTP server). The server supports
declarative specifications of how local tools can be executed and which parameters they take, using simple configuration files. For example, the following XML
2
Server Side
Client Side
(A machine with Linux, Windows or OSX)
Web-Interface
Tool1.cfg
Tool2.cfg
app2.cfg
ToolN.cfg
EasyInterface
Server
The server executes the tools via
command-line and forward their
standard output to the clients
Tool1
Eclipse Plugin
Clients communicate
with the server using
HTTP POST requests
Tool2
Remote Shell
ToolN
Tool3
Figure 1: EasyInterface architecture.
snippet is a configuration file for a tool called
"myapp".
<app id="myapp" visible="true">
...
<execinfo method="cmdline">
<cmdlineapp>
/path-to/myapp _ei_parameters
</cmdlineapp>
</execinfo>
<parameters prefix = "-" check="false">
...
<selectone name="c">
<option value="1" />
<option value="2" />
</selectone>
</parameters>
</app>
The cmdlineapp tag is a template describing how to execute the tool from the
command-line. The template parameter _ei_parameters is replaced by an appropriate value before execution. The server also supports template parameters
for, e.g., passing files, temporal working directories, session identifiers, etc. The
parameters tag includes a list of parameters accepted by the tool. For example,
the parameter “c” above takes one of the values 1 or 2.
Once the configuration file is installed on the server, we can access the tool
using an HTTP POST request that includes JSON-formatted data like the
following one
3
{
command: "execute",
app id: "myapp",
parameters: {
c: ["1"],
...
},
...
}
When receiving such a request, the server generates a shell command according
to the specification in the configuration file (e.g., “/path-to/myapp -c 1”),
executes it and redirects the standard output to the client. The server also
supports (i) tools that generate output in the background, we let clients fetch
output (partially) when it is ready; and (ii) tools that generate files, we let
clients download them later when needed. In all cases, the server can restrict
the resources available to a tool (e.g., the processing time), and guarantees the
safety of the generated command; i.e., clients cannot manipulate the server to
execute other programs installed on the server. In addition to tools, the server
can include example files, so users can easily try the tools.
EasyInterface not only makes the server side execution of a tool easy, it
provides client side GUIs that (1) connect to the server and ask for available
tools; (2) let users select the tool to execute, set its parameters and provide a
source program; (3) generate and send a request to the server; and (4) display
the returned output. EasyInterface provides three generic clients: a webinterface similar to an IDE; an Eclipse IDE plug-in; and a remote commandline shell. The last two clients are under development, so we focus here on the
web-interface.
The web-interface, shown in Fig. 2, is designed like an IDE where users can
edit programs, etc. Next to the Run button there is a drop-down menu with all
available tools obtained from the associated servers. In the settings window, the
user can select values for the different parameters of each tool. These parameters
are specified in the corresponding configuration files on the server side, and
automatically converted to combo-boxes, etc., by the web-interface. When the
user clicks the Run button, the web-interface sends a request to the associated
server to execute the selected tool and prints the received output back in the
console area of the web-interface.
Since the web-interface and Eclipse plug-in are GUI based clients, EasyInterface allows tools to generate output with some graphical effects, such
as opening dialog-boxes, highlighting code lines, adding markers, etc. To use
this feature, tools must support the EasyInterface output language, as in the
following XML snippet
4
Settings
Tools
Menu
Code Editor
File-Manager
Outline
Console
Figure 2: EasyInterface Web-Interface Client
<highlightlines dest="/path-to/sum.c">
<lines> <line from="5" to="10"/> </lines>
</highlightlines>
...
<oncodelineclick dest="/path-to/sum.c" outclass="info">
<lines><line from="17" /></lines>
<eicommands>
<dialogbox boxtitle="Hey!">
<content format="text"> some message </content>
</dialogbox>
</eicommands>
</oncodelineclick>
The tag highlightlines indicates that Lines 5–10 of file /path-to/sum.c should
be highlighted. The tag oncodelineclick indicates that when clicking on Line
17, a dialog-box with a corresponding message should be opened. Note that
a tool is only modified once to produce such output, with similar effect in all
EasyInterface clients (including future ones).
5
3
Concluding Remarks
EasyInterface is a toolkit for the rapid development of GUIs for commandline research prototype tools. The toolkit has been successfully used in the
Envisage project to integrate the tools from the different partners in a common
web-based environment, including parsers, type-checkers, compilers, simulators,
deadlock and worst-case cost analyzers, and a systematic testing framework
(see http://abs-models.org). Our experience suggests that the methodology
implied by EasyInterface for building GUIs is adequate for research prototype tools; as such tools change continuously, the corresponding GUIs can be
modified immediately and with negligible effort. Future work includes plans
to develop more clients, and libraries for different programming languages to
facilitate generation of the output commands/actions instead of printing these
directly.
References
[1] CodeMirror. http://codemirror.net.
[2] DyGraphs. http://dygraphs.com.
[3] Easyinterface User Manual. Available at http://costa.ls.fi.upm.es/
papers/costa/eiusermanual.pdf.
[4] Envisage:
Engineering
http://www.envisage-project.eu.
[5] JQuery. http://jquery.com.
[6] JSTree. http://www.jstree.com.
6
Virtualized
Services.
A
Using
EasyInterface
In this appendix we describe the methodology that EasyInterface implies
for its users by explaining the basics of its different parts and the way they
are supposed to be used. In Sec. A.1 we describe the different options that
are provided by the EasyInterface server, in Sec. A.2 we review the different
features of the web-interface, and in Sec. A.3 we discuss the EasyInterface
output language. Note that the installation of EasyInterface is practically
immediate: It only consists in cloning the github repository and making its root
directory accessible via an HTTP server. Thus, we omit the installation details
here and refer the reader to the installation guide that is available in the github
repository for further information.
A.1
Adding Tools and Examples to the Server
As described in Sec. 2, adding a tool to the server is simply done by providing a
configuration file such as the one in Sec. 2. This file includes two main components: (1) the command-line template that describes how to execute the corresponding tool; and (2) the parameter section that describes which command-line
parameters the tool will take.
EasyInterface supports several types of parameters such as a parameter
with a single value, a parameter with multiple values, a Boolean parameter, etc.
Each parameter has a name, a set of valid values, and a set of default values.
When receiving a request to execute a tool (such as the JSON-formatted data in
Sec. 2), the server generates a sequence of command-line parameters from those
specified in the request and replaces the template parameter _ei_parameters by
this list. For example, if there is a parameter named “c” and its provided
value is “1”, then what is passed to the tool is “-c 1”. The prefix “-” that is
attached to the parameter name can be specified using the prefix attribute in
the parameters section. In addition, the check attribute indicates if the server
should reject requests with invalid parameter values. Apart from _ei_parameters
the command-line template might include other template parameters, all are
first replaced by corresponding values and then the resulting shell command is
executed and its output is redirected back to the client. Next we describe some
of the available template parameters:
• _ei_files : tools typically receive input files (e.g., programs or models) to
process. The execution request (i.e., the JSON-formatted data of Sec. 2) can
include such files, and, in order to pass them to the corresponding tool, the
server first saves them locally in a temporary directory and replaces _ei_files
by a list of corresponding local paths.
• _ei_outline : since EasyInterface was first developed for tools that process
programs (e.g., program analysis tools), the execution request can include
so-called outline entities. These are elements of the input programs such as
method names, class names, etc., and they are used to, e.g., indicate from
where the tool should start the analysis. The server replaces _ei_outline by
the list of the provided outline entities.
7
•
_ei_execid : this corresponds to a unique execution identifier that is assigned
(by the server) to the current execution request, which can be used for future
references to this execution as discussed below.
• _ei_stream : there are tools that do not generate output immediately, such as
simulators. In this case we would like to keep the tools running in the background and fetch their output periodically without maintaining the current
connection to the server. The template parameter _ei_stream corresponds to
a temporary directory where the tool can write its output and where clients
can fetch this output by corresponding requests to the server. These requests
should include the corresponding execution identifier. For example, the tool
could output the following command (in the EasyInterface output language) and terminate, while keeping a process running in the background
(complying with the server’s permissions) writing its output chunks to the
_ei_stream directory:
<content format="text" execid="EI65231" time="60sec">>
The program is running in the background,
the output goes here ...
</content>
This command indicates that the output, in text format, should be fetched
every 60 seconds using the execution identifier specified in execid. Note that it
is the responsibility of the client (e.g., the web-interface) to fetch the output
once it receives the above command.
• _ei_download : some tools generate output in the form of large files, e.g., compiled code. Instead of redirecting this output directly to the client it can
be more convenient to return download links. This template parameter corresponds to a temporary directory where such files can be stored and later
fetched by sending a special request to the server with the file name and the
corresponding execution identifier, similar to the stream mode above. The
EasyInterface output language includes a special command for downloading such files:
<download
execid="EI65231" filename="file.zip" />
Once the client (e.g., the web-interface) receives this command, it sends a
request to download the file file.zip that is associated with the execution
identifier “EI65231”. The command can also be assigned to an “on click”
action instead of downloading it immediately (see Sec. A.3).
• _ei_sessionid : this corresponds to a unique session identifier that is assigned to
the user, and can be used to track the user’s activity among several requests.
It is implemented using PHP sessions.
• _ei_clientid : this corresponds to the client identifier, e.g., webclient, eclipse,
etc. It can be used to generate client directed output which uses the EasyInterface output language for the selected clients and plain text for others.
The server configuration files also include options for controlling security issues,
timeout for tools, etc.
In addition to the tools, one can install example sets on the server side, which
8
are meant to be used by clients (e.g., the web-interface) to provide users with
some examples on which they can apply the available tools. Adding example
sets is done via configuration files such as the following:
<examples>
<exset id="iter">
<folder name="Examples_1">
<folder name="iterative">
<file name="sum.c" url="https://.../sum.c" />
...
</folder>
...
</folder>
</exset>
<exset id="set2">
<folder name="Examples_2">
<github owner="username" repo="examples" branch="master" path="benchmarks"/>
</folder>
</exset>
</examples>
This defines two sets of examples. The first uses a directory structure, while
the second refers to a github repository. Note that in the first case the example
files are not necessarily installed on the server; a link must be provided for each
file.
A.2
Using the Web-Interface Client
The web-interface client of EasyInterface is a JavaScript program that runs
in a web browser, a screenshot is depicted in Fig. 2. It uses JQuery [5] as well
as some other libraries like JSTree [6] and CodeMirror [1]. It is designed like
an IDE, this is because EasyInterface was first developed for integrating the
tools of the Envisage [4] project (static analyzers, simulators, etc.) in a common
GUI. It includes the following components: (1) the code editor, where users can
edit programs; (2) the file-manager that contains a list of predefined examples
and user files; (3) the outline view that includes an outline (e.g., methods and
classes) of one or more files; (4) the console where the results of executing a
tool is printed by default; and (5) the tool bar that includes several buttons to
execute a tool, etc. Next we describe the workflow, and give more details on
these components.
The Workflow. The workflow within the web-interface is as follows: (a) write a
new program or open an existing one from the file-manager; (b) click on the
Refresh Outline button to generate the outline of the currently open program,
and select some entities from this outline; (c) select a tool from the tools menu;
and (d) click on the Settings button to set the values of the different parameters; (e) click on the Run button to execute the selected tool. At this point
the web-interface sends a request to the corresponding server to execute the selected tool (passing the currently opened file, parameter values, selected outline
entities, etc. to the tool), and the output is returned and printed in the console
area of the web-interface. If the tool’s output is in the EasyInterface output
9
language, then it passes through an interpreter that converts it to some corresponding graphical output. The user can apply a tool (and generate an outline)
on several files by selecting the corresponding option from the context menu in
the file-manager (opened with a right click on an entry in the file-manager).
Code Editor. The code editor is based on CodeMirror [1], it can be easily configured to use syntax-highlighting for different languages.
Tools Menu. The tools menu includes a list of tools that can be executed by
the user. This list can be set in the web-interface configuration file, simply by
providing the URLs of the corresponding EasyInterface servers and, for each,
indicating if all available tools should be included or only some selected ones.
The default configuration of the web-interfaces fetches all tools that are installed
on the server running on the machine where the web-interface is installed.
Settings. When clicking the Settings button, a settings window is opened where
the user can choose values for the different parameters of the different tools (see
the top part of Fig. 2). This is automatically generated using the parameters
defined in the server configuration file (the web-interface fetches this information
from the corresponding server).
File-Manager. The predefined examples included in the file-manager can be set in
the web-interface configuration file by providing the URLs of the corresponding
EasyInterface servers and identifiers for the example sets to be included. As
we have seen in Sec. 2, such a set can simply be given as a reference to a github
repository. The file-manager can also allow users to create their own files, upload
files from local storage, and clone public and private github repositories.
Outline. The Outline area is supposed to include an outline of some of the pro-
gram files (available in the file-manager), and thus depends on the structure of
those programs. For example, in the case of the Envisage [4] project it includes
elements (methods, classes, etc.) of ABS programs. Using it for Java programs,
for example, would require changing the way the outline is generated. EasyInterface already provides an easy way to change the outline generator. All we
have to do is (1) to provide a tool (installed on an EasyInterface server, like
any other tool) that takes a set of files and generates an XML structure that
represents an outline, the web-interface will convert this XML to a tree-view;
and (2) to modify the web-interface configuration file in order to use this tool
for generating the outline.
A.3
Using The Output Language
The EasyInterface output language is a text-based language that allows tools
to generate more sophisticated output. It is supported in both the web-interface
and Eclipse clients. In this section we will explain the basics of this language
by example. An output in the EasyInterface output language is an XML of
the following form:
10
<eiout>
<eicommands>
[EICOMMAND]*
</eicommands>
<eiactions>
[EIACTION]*
</eiactions>
</eiout>
where [EICOMMAND]* is a list of commands to be executed; and [EIACTION]* is a
list of actions to be declared. An [EICOMMAND] is an instruction like: print a text
on the console, highlight lines 5-10, etc. An [EIACTION] is an instruction like:
when the user clicks on line 13, highlight lines 20-25, etc. In the rest of this
section we consider some representative commands and actions.
Printing in the Console. Recall that when the EasyInterface output language is
used, the web-interface does not redirect the output to the console area and thus
we need a command to print in the console area. The following is an example
of a command that prints “Hello World” in the console area:
<printonconsole consoleid="1" consoletitle="Welcome">
<content format="text">Hello World</content>
</printonconsole>
The value of the consoleid attribute is the console identifier in which the given
text should be printed (e.g., in the web-interface the console area has several
tabs, so the identifier refers to one of those tabs). If a console with the provided
identifier does not exist yet, a new one is created, with a title as specified in
consoletitle. If consoleid is not given the output goes to the default console.
Inside printonconsole we can have several content tags which include the content
to be printed. The attribute format indicates the format of the content. In the
above example it is plain text, other formats are supported as well, e.g., html,
svg, and graphs. The graphs option refers to diagrams that are drawn using
DyGraphs [2], where the data is provided inside the content tag using a JSON
based format.
Adding Markers. The following is an example of a command that adds a marker
next to a code line in the editor area:
<addmarker dest="path" outclass="info">
<lines> <line from="4" /> </lines>
<content format='text'> some associated text </content>
</addmarker>
The attribute dest indicates the full path of the file in which the marker should
be added. The attribute outclass indicates the nature of the marker, which can
be info, error, or warning. This value typically affects the type/color of the
icon to be used for the marker. The tag lines includes the lines in which markers
should be added, each line is given using the tag line where the from attribute is
the line number (line can be used to define a region in other commands, this is
11
why the attribute is called from). The text inside the content tag is associated to
the marker (as a tooltip, a dialog box, etc., depending on the client).
Highlighting Code Lines. The following command can be used to highlight code
lines:
<highlightlines dest="path" outclass="info" >
<lines> <line from="5" to="10"/> </lines>
</highlightlines>
Attributes dest and outclass are as in the addmarker command. Each line tag
defines a region to be highlighted. In the example above, lines 5–10 get highlighted. We can also use the attributes fromch and toch to indicate the columns
in which the highlight starts and ends respectively.
Opening a Dialog Box. The following command can be used to open a dialog box
with some content:
<dialogbox outclass="info" boxtitle="Some Title!" boxwidth="100" boxheight="100">
<content format="html"> some associated text </content>
</dialogbox>
The dialog box will be titled as specified in boxtitle and it will include the content
as specified in the content environments.
CodeLine Actions. A CodeLine action defines a list of commands to be executed
when the user clicks on a line of code (more precisely, on a marker placed next
to the line). The commands can be any of those seen above. The following is
an example of such an action:
<oncodelineclick dest="path" outclass="info" >
<lines> <line from="17" /> </lines>
<eicommands>
<highlightlines>
<lines> <line from="17" to="19"/> </lines>
</highlightlines>
<dialogbox boxtitle="Hey!">
<content format="html"> some text </content>
</dialogbox>
</eicommands>
</oncodelineclick>
First note that the XML description above should be placed inside the eiactions
tag. When the above action is executed, e.g., by the web-interface client, a
marker will be shown next to line 17 of the file specified in the attribute dest. If
the user clicks on this marker the commands inside the eicommands tag will be
executed, and if the user clicks again the effect of these commands is undone.
In the case above, a click highlights lines 17–19 and opens a dialog box, and
another click removes the highlights and closes the dialog box.
OnClick Actions. OnClick actions are similar to CodeLine actions. The difference
is that instead of being assigned to a line of code, they are assigned to an HTML
tag that we have previously generated. Let us suppose that the tool has already
generated the following content in the console area:
12
<content format="html"/>
<span style="color: red;" id="err1">10 errors</span> were found in sum.c
</content>
Note that the text “10 errors” is wrapped by a span tag with an identifier err1.
The OnClick action can assign a list of commands to a click on this text as
follows:
<onclick>
<elements> <selector value="#err1"/> </elements>
<eicommands>
<dialogbox boxtitle="Errors">
<content format="html"> some text </content>
</dialogbox>
</eicommands>
</onclick>
The selectors used in the tag
selector
are as in JQuery [5].
13
| 6cs.PL
|
Under review as a conference paper at ICLR 2016
D EEP F EATURE L EARNING FOR EEG R ECORDINGS
Sebastian Stober, Avital Sternin, Adrian M. Owen & Jessica A. Grahn
The Brain and Mind Institute
University of Western Ontario
London, ON, Canada
{sstober,asternin,adrian.owen,jgrahn}@uwo.ca
arXiv:1511.04306v4 [cs.NE] 7 Jan 2016
A BSTRACT
We introduce and compare several strategies for learning discriminative features from electroencephalography (EEG) recordings using deep learning techniques. EEG data are generally only available in small quantities, they are highdimensional with a poor signal-to-noise ratio, and there is considerable variability between individual subjects and recording sessions. Our proposed techniques
specifically address these challenges for feature learning. Cross-trial encoding
forces auto-encoders to focus on features that are stable across trials. Similarityconstraint encoders learn features that allow to distinguish between classes by demanding that two trials from the same class are more similar to each other than to
trials from other classes. This tuple-based training approach is especially suitable
for small datasets. Hydra-nets allow for separate processing pathways adapting to
subsets of a dataset and thus combine the advantages of individual feature learning
(better adaptation of early, low-level processing) with group model training (better generalization of higher-level processing in deeper layers). This way, models
can, for instance, adapt to each subject individually to compensate for differences
in spatial patterns due to anatomical differences or variance in electrode positions. The different techniques are evaluated using the publicly available OpenMIIR dataset of EEG recordings taken while participants listened to and imagined
music.
1
I NTRODUCTION
Over the last decade, deep learning techniques have become very popular in various application domains such as computer vision, automatic speech recognition, natural language processing, and bioinformatics where they produce state-of-the-art results on various tasks. At the same
time, there has been very little progress investigating the application of deep learning in cognitive neuroscience research, where these techniques could be used to analyze signals recorded with
electroencephalography (EEG) – a non-invasive brain imaging technique that relies on electrodes
placed on the scalp to measure the electrical activity of the brain. EEG is especially popular for the
development of brain-computer interfaces (BCIs), which work by identifying different brain states
from the EEG signal.
Working with EEG data poses several challenges. Brain waves recorded in the EEG have a very low
signal-to-noise ratio and the noise can come from a variety of sources. For instance, the sensitive
recording equipment can easily pick up electrical line noise from the surroundings. Other unwanted
electrical noise can come from muscle activity, eye movements, or blinks. Usually, only certain brain
activity is of interest, and this signal needs to be separated from background processes. EEG lacks
spatial resolution on the scalp with additional spatial smearing caused by the skull but it has a good
(millisecond) time resolution to record both, slowly and rapidly changing dynamics of brain activity.
Hence, in order to identify the relevant portion of the signal, sophisticated analysis techniques are
required that should also take into account temporal information.
1
Under review as a conference paper at ICLR 2016
This is where deep learning techniques could help. For these techniques, training usually involves
the usage of large corpora such as ImageNet or TIMIT. As EEG data are high-dimensional1 and
complex, this also calls for large datasets to train deep networks for EEG analysis and classification.
Unfortunately, there is no such abundance of EEG data. Unlike photos or texts extracted from the
internet, EEG data are costly to collect and generally unavailable in the public domain. It requires
special equipment and a substantial effort to obtain high quality data. Consequently, EEG datasets
are only rarely shared beyond the boundaries of individual labs and institutes. This makes it hard
for deep learning researchers to develop more sophisticated analysis techniques tailored to this kind
of data.
With this paper, we want to address several of these challenges. We briefly review related work in
Section 2 and introduce our dataset in Section 3. Section 4 describes our proposed techniques and
the experiments conducted to test them. We discuss the results and draw conclusions in Section 5.
Supplementary material in the appendix comprises further illustrations of learned features as well as
details on the implementation of the proposed methods.
2
R ELATED W ORK
A major purpose of the work presented in this paper is to help advance the state of the art of signal
analysis techniques in the field of cognitive neuroscience. In this application domain, the potential of
deep learning techniques for neuroimaging has been demonstrated very recently by Plis et al. (2014)
for functional and structural magnetic resonance imaging (MRI) data. However, applications of deep
learning techniques within cognitive neuroscience and specifically for processing EEG recordings
have been very limited so far.
Mirowski et al. (2009) applied convolutional neural networks (CNNs) for epileptic seizure prediction
in EEG and intercranial EEG. Wulsin et al. (2011) used deep belief nets (DBNs) to detect anomalies
related to epilepsy in EEG recordings by classifying individual “channel-seconds”, i.e., one-second
chunks from a single EEG channel without further information from other channels or about prior
values. Their classifier was first pre-trained layer by layer as an auto-encoder on unlabelled data,
followed by a supervised fine-tuning with backpropagation on a much smaller labeled dataset. They
found that working on raw, unprocessed data (sampled at 256Hz) led to a classification accuracy
comparable to hand-chosen features. Längkvist et al. (2012) similarly employed DBNs combined
with hidden Markov models (HMMs) to classify different sleep stages. Their data for 25 subjects
comprised EEG as well as recordings of eye movements and skeletal muscle activity. Again, the data
was segmented into one-second chunks. Here, a DBN on raw data showed a classification accuracy
close to one using 28 selected features.
Furthermore, there have been some applications of CNNs for BCIs. Cecotti & Gräser (2008) used a
special CNN for classification of steady-state visual evoked potentials (SSVEPs) – i.e., brain oscillation induced by visual stimuli. The network integrated the Fourier transform between convolutional
layers, which transformed the data from the time domain to a time-frequency representation. Another CNN for detecting P300 waves (a well established waveform in EEG research) was described
in Cecotti & Gräser (2011). There has also been early work on emotion recognition from EEG using
deep neural networks such as described by Jirayucharoensak et al. (2014) and Zheng et al. (2014).
In our early work, we used stacked denoising auto-encoders (SDAs) and CNNs to classify EEG
recordings of rhythm perception and identify their ethnic origin – East African or Western – (Stober
et al. (2014b)) as well as to distinguish individual rhythms (Stober et al. (2014a)).
3
DATASET AND P RE -P ROCESSING
The OpenMIIR dataset (Stober et al. (2015)) is a public domain dataset of EEG recordings taken
during music perception and imagination.2 We collected this data from 10 subjects who listened to
and imagined 12 short music fragments – each 7s–16s long – taken from well-known pieces. These
stimuli were selected from different genres and systematically span several musical dimensions such
1
A single trial comprising ten seconds of EEG with 64 channels sampled at 100 Hz has already 64000
dimensions and the number of channels and the sampling rate of EEG recordings can be much higher than this.
2
The dataset is available at https://github.com/sstober/openmiir
2
Under review as a conference paper at ICLR 2016
as meter, tempo and the presence of lyrics as shown in Table 2 in the appendix. This way, various
retrieval and classification scenarios can be addressed.
All stimuli were normalized in volume and kept as similar in length as possible with care taken to
ensure that they all contained complete musical phrases starting from the beginning of the piece.
The pairs of recordings for the same song with and without lyrics were tempo-matched. The stimuli
were presented to the participants in several conditions while we recorded EEG. For the experiments
described in this paper, we only focus on the perception condition, where participants were asked
to just listen to the stimuli. The presentation was divided into 5 blocks that each comprised all 12
stimuli in randomized order. In total, 60 perception trials were recorded per subject.
EEG was recorded with a BioSemi Active-Two system using 64+2 EEG channels at 512 Hz. Horizontal and vertical electrooculography (EOG) channels were used to record eye movements. The
following common-practice pre-processing steps were applied to the raw EEG and EOG data using
the MNE-python toolbox by Gramfort et al. (2013) to remove unwanted artifacts. We removed and
interpolated bad EEG channels (between 0 and 3 per subject) identified by manual visual inspection.3 The data was then filtered with a bandpass keeping a frequency range between 0.5 and 30 Hz.
This also removed any slow signal drift in the EEG. To remove artifacts caused by eye blinks, we
computed independent components using extended Infomax independent component analysis (ICA)
as described by Lee et al. (1999) and semi-automatically removed components that had a high correlation with the EOG channels. Afterwards, the 64 EEG channels were reconstructed from the
remaining independent components without reducing dimensionality. Furthermore, the data of one
participant was excluded at this stage because of a considerable number of trials with movement
artifacts due to coughing. Finally, all trial channels were additionally normalized to zero mean and
range [−1, 1].
4
E XPERIMENTS
Using the EEG dataset described in the previous section, we would like to learn discriminative
features that can be used by a classifier to distinguish between the different music stimuli. Ideally,
these feature should also allow interpretation by cognitive neuroscientists to facilitate findings about
the underlying cognitive processes. In our previous experiments with EEG recordings of rhythm
perception, CNNs showed promising classification performance but the learned features were not
easy to interpret (Stober et al. (2014a)).
For the experiments described here, the following general implementation conventions applied: All
convolutional layers used the sigmoid tanh nonlinearity because its output naturally matches the
value range of the network inputs ([-1,1]) and thus facilitates easier interpretation of the activation
values. Furthermore, bias terms were not used. Convolution was always solely applied along the
time (samples) axis.4 For the classifiers, we used a DLSVM output layer employing the hinge loss
as described by Tang (2013) with an implementation based on the one provided by Kastner.5 This
generally resulted in a better classification performance than the commonly used Softmax in all our
previous experiments. For the convolutional auto-encoders (CAEs), our implementation of the deconvolutional layers has been derived from the code for generative adversarial nets by Goodfellow
et al. (2014).6 Stochastic gradient descent with batches of 128 trials was used for training. During
supervised training, we applied Dropout regularization (Hinton et al. (2012)) and a learning rate
momentum. During unsupervised pre-training, we did not use Dropout as the expected benefit is
much lower here and does not justify the increase in processing time. Generally, the learning rate
was set to decay by a constant factor per epoch. Furthermore, we used a L1 weight regularization
penalty term in the cost function to encourage feature sparsity.
To measure classification accuracy, we used the trials of the third block of each subject as test
set. This set comprised 108 trials (9 subjects x 12 stimuli x 1 trial). The remaining 432 trials (9
subjects x 12 stimuli x 4 trials) were used for training and model selection. For supervised training,
we employed a 9-fold cross-validation scheme by training on the data from 8 subjects (384 trials)
3
The removed and interpolated bad channels are marked in the topographic visualizations shown in Figure 3.
Beyond the scope of this paper, convolution could be applied in the spatial or frequency domain.
5
https://github.com/kastnerkyle/pylearn2/blob/svm_layer/
6
https://github.com/goodfeli/adversarial
4
3
Under review as a conference paper at ICLR 2016
and validating on the remain one (48 trials). This approach allowed us to additionally estimate the
cross-subject performance of a trained model.7 For hyper-parameter selection, we employed the
Bayesian optimization technique described by Snoek et al. (2012) which has been implemented in
the Spearmint library.8 We limited the number of jobs for finding optimal hyper-parameters to 100
except for the baseline at 64 Hz for which we ran 300 jobs. Each job comprised training the 9 fold
models for 50 epochs and selecting the model with the lowest validation error for each fold. We
compared two strategies for aggregating the separate fold models – averaging the model parameters
over all fold models (“avg”) or using a majority vote (“maj”). The accuracy on the test set is reported
for these aggregated models.
In the experiments described in the following, we focused on pre-training the first CNN layer. The
proposed techniques can nevertheless also be applied to train deeper layers and network structures
different from CNNs. Once pre-trained, the first layer was not changed during supervised classifier
training to obtain a measurement of the feature utility for the classification task. Furthermore, this
also resulted in a significant speed-up of training – especially at the high sampling rate of 512 Hz.
Except for the pre-training method described in Section 4.2, which uses full-length trials, all trials
were cut off at 6.9 s, the length of the shortest stimulus, which resulted in an equal input length.
For comparison, we additionally trained a linear support vector machine classifier (SVC) on top
of the learned features using Liblinear (Fan et al. (2008)). We also tested polynomial kernels, but
this did not lead to an increase in classification accuracy. For the SVC, the optimal value for the
parameter C that controls the trade-off between the model complexity and the proportion of nonseparable training instance was determined through a grid search during cross-validation.
4.1
S UPERVISED CNN T RAINING BASELINE
In order to establish a baseline for the OpenMIIR dataset, we first applied plain supervised classifier
training. We considered CNNs with two convolutional layers using raw EEG as input, which was
either down-sampled to 64 Hz or kept at the original sampling rate of 512 Hz, as well as respective
SVCs for comparison. The higher sampling rate offers better timing precision at the expense of
increasing the processing time and the memory requirement. We wanted to know whether using data
at the full rate of 512 Hz could by justified by increasing our classification accuracy. We conducted a
search on the hyper-parameter grid optimizing solely structural network parameters and the learning
rate. Results are shown in Table 1 (B and E). The baseline results give us a starting point against
which we can compare the results obtained through our proposed feature learning approaches.
4.2
L EARNING BASIC C OMMON S IGNAL C OMPONENTS
No matter how much effort one puts into controlling the experimental conditions during EEG recordings, there will always be some individual differences between subjects and between recording sessions. This can make it hard to combine recordings from different subjects to identify general
patterns in the EEG signals. A common way to address this issue is to average over many very short
trials such that differences cancel out each other. When this is not feasible because of the trial length
or a limited number of trials, an alternative strategy is to derive signal components from the raw EEG
data hoping that these will be stable and representative across subjects. Here, principle component
analysis (PCA) and ICA are well-established techniques. Both learn linear spatial filters, whose
channel weights can be visualized by topographic maps as shown in Figure 3. The first two columns
contain the signal components learned with PCA and extended Infomax ICA (Lee et al. (1999)) on
the concatenated perception trials whereas the remaining components have been obtained using a
convolutional auto-encoders (CAEs) with 4 filters.
Using CAEs, allows us to learn individually adapted components that are linked between subjects.
To this end, a CAE is first trained on the combined recordings from the different subjects to identify the most common components. The learned filter weights are then used as initial values for
individually trained auto-encoders for each subject. This leads to adaptations that reflect individual
differences. Ideally, these are small enough such that the relation to the initial common components
7
8
There is still a model selection bias because the cross-subject validation set is used for early stopping.
https://github.com/JasperSnoek/spearmint
4
Under review as a conference paper at ICLR 2016
Figure 1: A spatio-temporal filter over all 64 EEG channels and 3 samples after pre-training using
cross-trial encoding. Left column: Global filter after training on within-subject trial pairs (stage 1).
Remaining columns: Individual filters (hydra-net) after training with cross-subject trials (stage 2).
The output of this filter was used in the classifiers with ID M listed in Table 1.
still remains as can be seen in Figure 3. Optionally, a regularization term can be added to the cost
function to penalize weight changes, but this was not necessary here.
We measured reconstruction error values similar to those obtained with PCA and substantially lower
than for ICA. This demonstrates the suitability for general dimensionality reduction with the additional benefit of obtaining a common data representation across subjects that can accommodate individual differences. Using the filter activation of either the global or the individually adapted CAEs
shown in Figure 3 as common feature representation, we constructed SVC and CNN classifiers. The
latter consisted of another convolutional layer and a fully connected output layer with hinge loss.
The obtained accuracies after hyper-parameter optimization are shown in Table 1 (IDs G and I).
4.3
C ROSS -T RIAL E NCODING
Aiming to find signal components that are stable across trials of the same class, we changed the
training scheme of the CAE to what we call cross-trial encoding. Instead of trying to simply reconstruct the input trial, the CAE now had to reconstruct a different trial belonging to the same class.9
This strategy can be considered as a special case of the generalized framework for auto-encoders
proposed by Wang et al. (2014).10 Given nC trials for a class C, n2C or nC (nC − 1) pairs of input
and target trials can be formed depending on whether pairs with identical trials are included. This
increases the number of training examples by a factor depending on the partitioning of the dataset
with respect to the different classes. As for the basic auto-encoding scheme, the training objective is
to minimize a reconstruction error. In this sense, it is unsupervised training but the trials are paired
for training using knowledge about their class labels. We found that using the distance based on the
dot product worked best as reconstruction error.
We split the training process into two stages. In stage 1, trials were paired within subjects whereas
in stage 2, pairs across subjects were considered as well. To allow the CAE to adapt to individual
differences between the subjects in stage 2, we introduced a modified network structure called hydranet. Hydra-nets allow to have separate processing pathways for subsets of a dataset. This makes it
possible to have different weights – depending on trial meta-data – in selected layers of a deep
network. Such a network can, for instance, adapt to each subject individually. With individual input
layers, the structure can be considered as a network with multiple “heads” – hence the reference
to Hydra. Here, we applied this approach at both ends of the CAE, i.e., the encoder filters were
selected based on the input trial’s subject whereas the decoder filters were chosen to match the
target trial’s subject. Encoder and decoder weights were tied within subjects. Figure 1 shows the
result learned by a CAE with a single filter of width 3 after stage 1 and 2. Further examples are
shown in Appendix C. Classification results were obtained in the same way as in Section 4.2 and
are shown in Table 1 (K–O).
9
Under a strict perspective, trying to reconstruct a different trial may no longer be considered as autoencoding. However, we rather see the term as a reference to the network architecture, which is still the same as
for regular auto-encoders. Only the training data has changed.
10
They similarly define a “reconstruction set” (all instances belonging to the same class) but use a different
(and much simpler) loss function based on linear reconstruction and do not consider convolution.
5
Under review as a conference paper at ICLR 2016
P,Q,R
S,T
mean
t
t+1
t+2
64 Hz
U,V,W
63.9%
64.5%
X,Y
512 Hz
65.7%
+
0
-
65.9%
Figure 2: Filters learned by similarity-constraint encoding for EEG sampled at 64 Hz (top row) and
512 Hz (bottom row). Left: Simple spatial filters with width 1. Right: Spatio-temporal filters with
width 3. The mean over all time steps is shown additionally for comparison. Letters P–Y refer to the
classifier IDs in Table 1 and Table 3 that use the respective filter as their first layer. The percentage
of correctly classified training triples is shown at the lower right of each filter.
4.4
S IMILARITY-C ONSTRAINT E NCODING
Cross-trial encoding as described above aims to identify features that are stable across trials and
subjects. Ideally, however, features should also allow us to distinguish between classes but this is
not captured by the reconstruction error used so far. In order to identify such features, we propose a
pre-training strategy called similarity-constraint encoding in which the network is trained to encode
relative similarity constraints. As introduced by Schultz & Joachims (2004), a relative similarity
constraint (a, b, c) describes a relative comparison of the trials a, b and c in the form “a is more
similar to b than a is to c.” Here, a is the reference trial for the comparison. There exists a vast
literature on using such constraints to learn similarity measures in various application domains.
Here, we use them to define an alternative cost function for learning a feature encoding. To this end,
we combine all pairs of trials (a, b) from the same class (as described in Section 4.3) with all trials
c from other classes demanding that a and b are more similar.
All trials within a triplet that constitutes a similarity constraint are processed using the same encoder pipeline. This results in three internal feature representations. Based on these, the reference
trial is compared with the paired trial and the trial from the other class resulting in two similarity
scores. Here, we used the dot product as similarity measure because this matched the computation
performed later during supervised training in the classifier output layer. The output layer of the similarity constraint encoder finally predicts the trial with the highest similarity score without further
applying any additional transformations. The whole network can be trained like a common binary
classifier, minimizing the error of predicting the wrong trial as belonging to the same class as the
reference. Technical details and a schematic of our similarity-constraint encoding approach can be
found in Section G.4.
This strategy is different from the one proposed by Yang & Shah (2014) to jointly learn features
together with similarity metrics. In particular, they used pairs for training and predicted whether
these were from the same or different classes. Their approach also required balancing two cost
functions (reconstruction vs. discrimination). We hypothesize that our relative comparison approach
leads to smoother learning. Each fulfilled constraint only makes a small local contribution to the
global structure of the similarity space. This way, the network can more gradually adapt compared
to a scenario where it would have to directly recognize the different classes based on a few training
trials. Optionally, the triplets can be extended to tuples of higher order by adding more trials from
other classes. This results in a gradually harder learning task because there are now more other
trials to compare with. At the same time, each single training example comprises multiple similarity
constraints, which might speed up learning. In the context of this paper, we focus only on triplets.
Figure 2 shows filters learned by similarity-constraint encoding for different structural configurations. They look very different from the results obtained earlier. Remarkably, the same pattern
emerged in all tested configurations. Using more than one filter generally did not lead to improvements in the constraint encoding performance. A kernel width of more than 3 samples at 64 Hz
sampling rate did not lead to different patterns as channel weights for further time steps remained
close to zero. The classification accuracies of using these pre-trained features in SVC and CNN
classifiers with optimized hyper-parameters are reported in Table 1 (IDs Q–Y).
6
Under review as a conference paper at ICLR 2016
Table 1: Classification accuracy measured on the test set using the different feature learning techniques described earlier in combination with SVC and CNN classifiers. Convolution parameters for
feature computation are shown as the number of filters times the filter width in samples. Classifier
hyper-parameters were optimized during cross-validation. The ID refers to the respective row in
Table 3, which provides more details on the CNN parameters for the different configurations.
Feature Learning Technique
None (baseline using raw EEG at 64 Hz)
None (baseline using raw EEG at 512 Hz)
Common Components (64 Hz, global filters)
Common Components (64 Hz, individually adapted filters)
Cross-Trial Encoder (64 Hz, individually adapted filters)
Similarity-Constraint Encoder (64 Hz)
Similarity-Constraint Encoder (512 Hz)
5
Filters
SVC
4x1
4x1
1x1
1x3
4x1
1x1
1x3
1x1
1x3
14.8%
14.8%
17.6%
19.4%
16.7%
23.1%
20.4%
20.4%
24.1%
22.2%
23.1%
CNN (avg) CNN (maj) ID
18.5%
18.5%
17.6%
20.4%
20.4%
23.1%
22.2%
25.0%
20.4%
22.2%
27.8%
26.9%
19.4%
18.5%
20.4%
19.4%
22.2%
22.2%
27.8%
18.5%
23.1%
26.9%
B
E
G
I
K
M
O
Q
T
V
Y
D ISCUSSION
Trying to determine which music piece somebody listened to based on the EEG is a challenging
problem. Attempting to do this with a small training set, makes the task even harder. Specifically, in
the experiments described above, we trained classifiers for a 12-class problem (recognizing the 12
music stimuli listed in Table 2) on the combined data from 9 subjects with an input dimensionality
of 28,160 (at 64 Hz) or 225,280 (at 512 Hz) respectively given only 4 training examples per class
and subject.
Remarkably, all CNN classifiers listed in Table 1 had an accuracy that was significantly above
chance. Even for model G, the value of 17.59% was significant at p=0.001. Significance values
were determined by using the cumulative binomial distribution to estimate the likelihood of observing a given classification rate by chance. It is likely that the classification accuracy will slightly
increase when the pre-trained filters of the first layer are allowed to change during supervised training of the full CNN. We did not consider this here because we wanted to analyze the impact of the
pre-trained features. There appears to be a ceiling effect as the median cross-validation accuracy did
not exceed 40%. Examining the cross-validation performance for the individual subjects, however,
indicates that cross-subject accuracies above 50% are possible for some of the subjects when using
individual models. We plan to investigate this possibility in the future. For this paper, we focused on
training with the combined data from all subjects using new techniques to compensate for individual
differences as we expected this to result in more robust features.
For CNN model aggregation, averaging the filters of the 9 cross-validation fold models (“avg”)
worked surprisingly well in general and did not result in a significantly different performance compared to using a majority vote (“maj”). A single average CNN also has the advantage that it is much
easier to analyze than the 9 individual voting models.
Comparing the baseline classifiers (using raw EEG) with those relying on the learned features as
listed in Table 1 using a z-test, we found a significant improvement at p=0.05 for the SVC model
T. For the average CNN model Y, the improvement over the respective baseline was significant at
p=0.061. Both classifiers relied on features learned by similarity-constraint encoding. For the SVC
models M and Y, we obtained a p-value of 0.068. Due to the rather small test set size (n=108), many
performance differences were only significant at higher p-values. There appears to be a small performance advantage in using the full sampling rate of 512 Hz as indicated by the scatter plots of the
model performance during hyper-parameter optimization shown in Appendix H. Using similarityconstraint encoding to pre-train a simple spatial filter for the first CNN layer that only aggregates the
64 EEG channels into a single waveform not only significantly improved the classification accuracy
but also reduced the time needed for hyper-parameter optimization from over a week for the full
CNN to a few hours. This improvement makes working at the full sampling rate feasible.
7
Under review as a conference paper at ICLR 2016
In our experiments, we used similarity-constraint encoding to learn global filters. These filters could
be further adapted to the individual subjects through our hydra-net approach. We believe that this
combination has a high potential to further improve the pre-trained filters and the resulting classification accuracy. However, technical limitations of our current implementation cause a significant
increase in processing time for this setting (details can be found in Appendix G). We are currently
working on a different implementation to resolve this issue.
Amongst the best CNN classifiers, which are listed in Table 3 with more details, models R and
W stand out for their simplicity and accuracy that is on par with much bigger models. Their improvement over the baseline is significant at p=0.1 and p=0.05 respectively. Remarkably, model
R does effectively not even use the second convolutional layer. We interpret this as an indicator
for the quality of the pre-trained features. Both models are simple enough to allow for interpretation of the learned features by domain experts. Visualizations can be found in Appendix E. The
visualizations of the layer 1 filters indicate which recording electrodes are most important to the
classifier. The electrodes within the dark red areas that appear bilaterally towards the back of the
head lie directly over the auditory cortex. These electrodes may be picking up on brain activation
from the auditory cortex that is modulated by the perception of the stimuli. The electrodes within
the blue areas that appear more centrally may be picking up on the cognitive processes that occur as
a result of the brain processing the music at a higher level. In layer 3 there is remarkable similarity
of the learned temporal patterns between stimuli 1–4 and their corresponding stimuli 11–14, which
are tempo-matched recordings of songs 1–4 without lyrics (cf. Table 2 for details). This indicates
that the models have picked up musical features from the stimuli such as down-beats (marking the
beginning of a measure). Further investigations will look at what aspects of the music are causing
these similarities.
Ultimately, we would like to apply the EEG analysis techniques described here in a brain-computer
interface (BCI) setting. To this end, it would be already sufficient to distinguish between two stimuli
with a high accuracy. An analysis of the binary classification performance of model W showed
already promising results with up to perfect accuracy for certain stimulus pairs even though the
model had been only trained for the 12-class problem (cf. Figure 10).
6
C ONCLUSIONS
We have proposed several novel techniques for deep feature learning from EEG recordings that
address specific challenges of this application domain. Cross-trial encoders aim to capture invariance between trials within and across subjects. Hydra-nets can learn specifically adapted versions
of selected layers and switch between them based on trial metadata to, for instance, compensate
for differences between subjects. Finally, similarity-constraint encoders allow us to identify signal
components that help to distinguish trials of the same class from others. First experiments on the
OpenMIIR dataset demonstrate that our techniques are able to learn features that are useful for classification. Most significant improvements over a baseline using raw EEG input were obtained by
similarity-constraint encoding picking up relevant brain regions as well musically meaningful temporal signal characteristics such as the position of down-beats. Furthermore, the learned features are
also still simple enough to allow interpretation by domain experts such as cognitive neuroscientists.
Whilst targeting problems specific to EEG analysis, the proposed techniques are not limited to this
kind of data or even CNNs. Thus, we believe that they will be useful beyond the scope of our
application domain. We also hope that this paper will encourage other researchers to join us in the
challenge of decoding brain signals with deep learning techniques. For future work, we want to
further refine our approach and investigate whether it can be used to identify overlapping features
between music perception and imagery, which would be very useful for BCI settings.
ACKNOWLEDGMENTS
This work has been supported by a fellowship within the Postdoc-Program of the German Academic Exchange Service (DAAD), the Canada Excellence Research Chairs (CERC) Program, an
National Sciences and Engineering Research Council (NSERC) Discovery Grant, an Ontario Early
Researcher Award, and the James S. McDonnell Foundation. The authors would further like to thank
the study participants.
8
Under review as a conference paper at ICLR 2016
References:
Bergstra, J., Breuleux, O., Bastien, F., Lamblin, P., Pascanu, R., Desjardins, G., Turian, J., Warde-Farley, D.,
and Bengio, Y. Theano: a CPU and GPU math expression compiler. In Proc. of the Python for Scientific
Computing Conference (SciPy), 2010.
Bourlard, H. and Kamp, Y. Auto-association by multilayer perceptrons and singular value decomposition.
Biological Cybernetics, 59(4-5):291–294, 1988.
Cecotti, H. and Gräser, A. Convolutional Neural Network with embedded Fourier Transform for EEG classification. In 19th Int. Conf. on Pattern Recognition (ICPR), pp. 1–4, 2008.
Cecotti, H. and Gräser, A. Convolutional Neural Networks for P300 Detection with Application to BrainComputer Interfaces. IEEE Trans. on Pattern Analysis and Machine Intelligence, 33(3):433–445, 2011.
Fan, R.E., Chang, K.W., Hsieh, C.J., Wang, X.R., and Lin, C.J. Liblinear: A library for large linear classification. The Journal of Machine Learning Research, 9:1871–1874, 2008.
Glorot, X., Bordes, A., and Bengio, Y. Deep sparse rectifier networks. In NIPS 2010 Workshop on Deep
Learning and Unsupervised Feature Learning, 2010.
Goodfellow, I.J., Warde-Farley, D., Lamblin, P., Dumoulin, V., Mirza, M., Pascanu, R., Bergstra, J., Bastien, F.,
and Bengio, Y. Pylearn2: a machine learning research library. arXiv: 1308.4214 [cs, stat], 2013.
Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio,
Y. Generative Adversarial Networks. arXiv: 1406.2661 [cs, stat], 2014.
Gramfort, A., Luessi, M., Larson, E., Engemann, D.A., Strohmeier, D., Brodbeck, C., Goj, R., Jas, M., Brooks,
T., Parkkonen, L., and Hämäläinen, M. MEG and EEG data analysis with MNE-Python. Frontiers in
Neuroscience, 7, 2013.
Hinton, G.E., Srivastava, N., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R.R. Improving neural networks
by preventing co-adaptation of feature detectors. arXiv:1207.0580, 2012.
Jirayucharoensak, S., Pan-Ngum, S., and Israsena, P. EEG-Based Emotion Recognition Using Deep Learning
Network with Principal Component Based Covariate Shift Adaptation. The Scientific World Journal, 2014.
Längkvist, M., Karlsson, L., and Loutfi, M. Sleep stage classification using unsupervised feature learning.
Advances in Artificial Neural Systems, 2012.
Lee, T., Girolami, M., and Sejnowski, T. J. Independent Component Analysis Using an Extended Infomax
Algorithm for Mixed Subgaussian and Supergaussian Sources. Neural Computation, 11(2):417–441, 1999.
Masci, J., Meier, U., Cireşan, D., and Schmidhuber, J. Stacked convolutional auto-encoders for hierarchical
feature extraction. In Proc. of the 21th Int. Conf. on Artificial Neural Networks (ICANN), pp. 52–59, 2011.
Mirowski, P., Madhavan, D., LeCun, Y., and Kuzniecky, R. Classification of patterns of EEG synchronization
for seizure prediction. Clinical Neurophysiology, 120(11):1927–1940, 2009.
Plis, S.M., Hjelm, D.R., Salakhutdinov, R., Allen, E.A., Bockholt, H.J., Long, J.D., Johnson, H.J., Paulsen,
J.S., Turner, J.A., and Calhoun, V.D. Deep learning for neuroimaging: a validation study. Frontiers in
Neuroscience, 8, 2014.
Schultz, M. and Joachims, T. Learning a distance metric from relative comparisons. Advances in neural
information processing systems (NIPS), pp. 41–48, 2004.
Snoek, J., Larochelle, H., and Adams, R.P. Practical bayesian optimization of machine learning algorithms. In
Advances in Neural Information Processing Systems (NIPS), pp. 2951–2959, 2012.
Stober, S., Cameron, D. J., and Grahn, J. A. Using convolutional neural networks to recognize rhythm stimuli
from electroencephalography recordings. In Advances in Neural Information Processing Systems (NIPS),
pp. 1449–1457, 2014a.
Stober, S., Cameron, D.J., and Grahn, J.A. Classifying EEG recordings of rhythm perception. In 15th Int.
Society for Music Information Retrieval Conf. (ISMIR), pp. 649–654, 2014b.
Stober, S., Sternin, A, Owen, A.M., and Grahn, J.A. Towards music imagery information retrieval: Introducing
the OpenMIIR dataset of EEG recordings from music perception and imagination. In 16th Int. Society for
Music Information Retrieval Conf. (ISMIR), pp. 763–769, 2015.
Tang, Y. Deep Learning using Linear Support Vector Machines. arXiv: 1306.0239 [cs, stat], 2013.
Wang, W., Huang, Y., Wang, Y., and Wang, L. Generalized Autoencoder: A Neural Network Framework
for Dimensionality Reduction. In 2014 IEEE Conference on Computer Vision and Pattern Recognition
Workshops (CVPRW), pp. 496–503, 2014.
Wulsin, D.F., Gupta, J.R., Mani, R., Blanco, J.A., and Litt, B. Modeling electroencephalography waveforms
with semi-supervised deep belief nets: fast classification and anomaly measurement. Journal of Neural
Engineering, 8(3), 2011.
Yang, Y. and Shah, M. Learning discriminative features and metrics for measuring action similarity. In 2014
IEEE International Conference on Image Processing (ICIP), pp. 1555–1559, 2014.
Zheng, W.-L., Zhu, J.-Y., Peng, Y., and Lu, B.-L. EEG-based emotion classification using deep belief networks.
In 2014 IEEE Int. Conf. on Multimedia and Expo (ICME), pp. 1–6, 2014.
9
Under review as a conference paper at ICLR 2016
A
A DDITIONAL I NFORMATION ABOUT THE O PEN MIIR M USIC S TIMULI
Table 2: Information about the tempo, meter and length of the stimuli (without cue clicks).
ID Name
1
2
3
4
11
12
13
14
21
22
23
24
Meter Length
Chim Chim Cheree (lyrics)
Take Me Out to the Ballgame (lyrics)
Jingle Bells (lyrics)
Mary Had a Little Lamb (lyrics)
Chim Chim Cheree
Take Me Out to the Ballgame
Jingle Bells
Mary Had a Little Lamb
Emperor Waltz
Hedwig’s Theme (Harry Potter)
Imperial March (Star Wars Theme)
Eine Kleine Nachtmusik
mean
10
3/4
3/4
4/4
4/4
3/4
3/4
4/4
4/4
3/4
3/4
4/4
4/4
Tempo
13.3s
7.7s
9.7s
11.6s
13.5s
7.7s
9.0s
12.2s
8.3s
16.0s
9.2s
6.9s
212 BPM
189 BPM
200 BPM
160 BPM
212 BPM
189 BPM
200 BPM
160 BPM
178 BPM
166 BPM
104 BPM
140 BPM
10.4s
176 BPM
Under review as a conference paper at ICLR 2016
B
V ISUALIZATIONS OF C OMMON C OMPONENTS
Figure 3: Topographic map visualization of the EEG channel weights for the top-4 signal components derived from the perception trials. The leftmost two columns show the top-4 principal components (explaining over 80% of the variance in total; individual values are shown next to the components) and the corresponding independent components computed using extended Infomax ICA. The
remaining columns show the weights learned by a tied-weights convolutional auto-encoder for all
subjects combined (third column) and all individual subjects labeled P01 to P14 accordingly using
4 filters of width 1 (samples) with tanh nonlinearity and without biases. Noisy EEG channels that
were removed and interpolated are marked by red X’s. Values at the bottom of each column refer to
the mean squared reconstruction error (MRSE) per sample, which was used as cost function.
Figure 4: Topographic map visualization of the EEG channel weights for the top-4 signal components derived from the cued imagination trials. They look very similar to the ones shown in Figure 3
for the perception trials. The leftmost two columns show the top-4 principal components (explaining
over 80% of the variance in total; individual values are shown next to the components) and the corresponding independent components computed using extended Infomax ICA. The remaining columns
show the weights learned by a tied-weights convolutional auto-encoder for all subjects combined
(third column) and all individual subjects labeled P01 to P14 accordingly using 4 filters of width 1
(samples) with tanh nonlinearity and without biases. Noisy EEG channels that were removed and
interpolated are marked by red X’s. Values at the bottom of each column refer to the mean squared
reconstruction error (MRSE) per sample, which was used as cost function.
11
Under review as a conference paper at ICLR 2016
C
F URTHER V ISUALIZATIONS OF C ROSS -T RIAL E NCODERS
Figure 5: A single spatial filter after pre-training using cross-trial encoding. Left column: Global
filter after training on within-subject trial pairs (stage 1). Remaining columns: Individual filters
(hydra-net) after training with cross-subject trials (stage 2). This filter was used as fixed first layer
in the CNNs classifiers J and K listed in Table 1 and Table 3.
Figure 6: Four spatial filters after pre-training using cross-trial encoding. Left column: Global filters
after training on within-subject trial pairs (stage 1). Remaining columns: Individual filters (hydranet) after training with cross-subject trials (stage 2). This filter was used as fixed first layer in the
CNNs classifiers N and O listed in Table 1 and Table 3.
12
Under review as a conference paper at ICLR 2016
D
D ETAILED CNN C LASSIFIER P ERFORMANCE
Table 3: Overview of best classifier CNNs identified by hyper-parameter optimization. Layer parameters refer to output channels x [width x height] x input channels / sub-sampling factor.
ID description
valid. test (avg) test (maj)
1st layer
2nd layer
3rd layer
4x[9x1]x4
4x[10x1]x1
2x[12x1]x4
1708x12+12
1708x12+12
854x12+12
22188
20868
11124
3x[1x1]x64 1x[256x1]x3/7 466x12+12
3x[1x1]x64 1x[256x1]x3/19 171x12+12
6564
3024
Baseline (64 Hz, no pre-training, 300 jobs for optimization):
A best test value
33.3% 33.3%
27.8% 4x[6x1]x64
B best validation value 39.6% 18.5%
26.9% 1x[5x1]x64
C smallest amongst best 35.4% 27.7%
25.9% 4x[3x1]x64
Baseline (512 Hz, no pre-training):
D best test value
29.2% 23.1%
E best validation value 31.2% 18.5%
29.6%
19.4%
Common Components (64 Hz, global filters):
F best test value
29.2% 27.8%
25.9%
G best validation value 33.3% 17.6%
18.5%
#params
4x[1x1]x64
4x[1x1]x64
4x[16x1]x4
1x[16x1]x4
1700x12+12
425x12+12
20924
5432
Common Components (64 Hz, individually adapted filters):
H best test value
29.2% 28.7%
25.0% 4x[1x1]x64
I best validation value 31.2% 20.4%
20.4% 4x[1x1]x64
4x[16x1]x4
8x[3x1]x4
1700x12+12
3504x12+12
20924
42412
6944x12+12
435x12+12
438x12+12
435x12+12
2044x12+12
4400x12+12
83516
5302
5461
5428
24852
53108
36490
25986
5346
83612
17464
Cross-Trial Encoder (64 Hz, individually adapted filters):
J best test value
33.3% 25.0%
23.1% 1x[1x1]x64 16x[7x1]x1
K best validation value 35.4% 20.4%
19.4% 1x[1x1]x64
1x[6x1]x1
L best test value
31.2% 25.0%
24.1% 1x[3x1]x64
1x[1x1]x1
M best validation value 39.6% 23.1%
22.2% 1x[3x1]x64
1x[4x1]x1
N best test value
31.2% 27.8%
25.0% 4x[1x1]x64 14x[1x1]x4/3
O best validation value 39.6% 22.2%
22.2% 4x[1x1]x64 10x[1x1]x4
Similarity-Constraint Encoder (64 Hz):
P best test value
33.3% 28.7%
Q best validation value 35.4% 25.0%
R smallest amongst best 33.3% 25.9%
S best test value
31.2% 26.9%
T best validation value 35.4% 20.4%
25.9%
27.8%
25.9%
27.8%
18.5%
1x[1x1]x64 14x[9x1]x1/2 3024x12+12
1x[1x1]x64 10x[11x1]x1/2 2150x12+12
1x[1x1]x64
1x[1x1]x1
440x12+12
1x[3x1]x64 16x[5x1]x1
6944x12+12
1x[3x1]x64 10x[10x1]x1/3 1430x12+12
Similarity-Constraint Encoder (512 Hz):
U best test value
29.2% 33.3%
V best validation value 35.4% 22.2%
W smallest amongst best 33.3% 28.7%
X best test value
29.2% 31.5%
Y best validation value 35.4% 27.8%
34.3%
23.1%
27.8%
31.5%
26.9%
1x[1x1]x64
1x[1x1]x64
1x[1x1]x64
1x[3x1]x64
1x[3x1]x64
13
13x[85x1]x1/10
7x[69x1]x1/14
1x[37x1]x1/11
16x[128x1]x1
5x[40x1]x1/7
4459x12+12
54689
1722x12+12
21223
316x12+12
3905
54224x12+12 652940
2480x12+12
30164
Under review as a conference paper at ICLR 2016
E
M ODEL V ISUALIZATIONS
Layer 2
+
0
-
Layer 3 (classes)
Layer 1
1
2
3
4
11
12
13
14
21
22
23
240
100
200
300
time (in samples)
400
Figure 7: Visualization of CNN R (average of the 9 cross-validation fold models), which processes
raw EEG at a sampling rate of 64 Hz. Layer 1 was pre-trained using similarity-constraint encoding.
+
0
-
Layer 2
0
1
2
3
4
11
12
13
14
21
22
23
240
5
10
15
20
25
30
35
Layer 3 (classes)
Layer 1
50
100
150
200
time (in samples, down-sampled by factor 11)
250
300
Figure 8: Visualization of CNN W (average of the 9 cross-validation fold models), which processes
raw EEG at a sampling rate of 512 Hz. Layer 1 was pre-trained using similarity-constraint encoding.
14
Under review as a conference paper at ICLR 2016
F
C ONFUSION
A NALYSIS FOR
CNN M ODEL
W ( AVG )
12-Class
Stimuli
Confusion
Chim Chim Cheree (lyrics)
Take Me Out to the Ballgame (lyrics)
Jingle Bells (lyrics)
Mary Had a Little Lamb (lyrics)
Chim Chim Cheree
Take Me Out to the Ballgame
Jingle Bells
Mary Had a Little Lamb
Emperor Waltz
Hedwig’s Theme (Harry Potter)
Imperial March (Star Wars Theme)
Eine Kleine Nachtmusik
Figure 9: 12-class confusion matrix
for CNN W (average of the 9 cross-validation
fold models).
Sebastian Stober - Deep Feature Learning for EEG
2015-12-01 51
Stimulus Class B
1
2
2
3
3
4
4
Stimulus Class A
11
11
12
12
binomial p-value
13
1
13
2
14
3
14
Stimulus Class A
4
21
11
21
12
13
22
14
5e-05
0.0001
0.0005
0.001
0.005
0.01
0.05
21
22
23
2
3
22
23
23
4
11
12 13 14 21
Stimulus Class B
22
23
24
24
Figure 10: Binary confusion matrices for CNN W (average of the 9 cross-validation fold models).
For each binary classification, only the 18 test trials belonging to either stimulus class A or stimulus
class B were considered. The inset at the lower bottom visualizes the p-values determined by using
the cumulative binomial distribution to estimate the likelihood of observing the resepective binary
classification rate by chance. Note that the classifier was only trained for the 12-class problem.
15
Under review as a conference paper at ICLR 2016
G
I MPLEMENTATION D ETAILS
For reproducibility and to encourage further developments and research in this direction, all code necessary to
build the proposed deep network structures and to run the experiments described in the following is shared as
open source within the deepthought library.11 The implementation is based on the libraries Pylearn2 (Goodfellow et al. (2013)) and Theano (Bergstra et al. (2010)) and comprises various custom Layer and Dataset classes
– such as for on-the-fly generation of trial tuples and the respective classification targets during iteration.
The following subsections provide details about our implementation.
G.1
D ETAILS ON C ONVOLUTIONAL AUTO -E NCODERS FOR EEG
Convolutional autoencoders are a special variant of CNNs that encode their input using convolution into a
compressed internal representation which is then decoded using de-convolution into the original space trying to
minimize the reconstruction error. Such encoder/decoder pairs can optionally be stacked as described in Masci
et al. (2011) to increase the complexity of the internal representations.
We considered two measures of the reconstruction quality. The mean square reconstruction error (MSRE) is the
commonly used error measure for auto-encoding and computed as the squared Euclidean distance between the
input and the reconstruction averaged over all samples. We further define the mean channel correlation (MCC)
as the mean correlation of the individual input channels and their reconstructions. The correlation is computed
as Pearson’s r, which is in the range [-1,1] where 1 is total positive correlation, 0 is no correlation, and -1 is
total negative correlation.
The convolutional filters applied on multi-channel raw EEG data are 2-dimensional where the width is time
(samples) and the height is optionally used for different frequency bands in a time-frequency representation,
which is not considered in the context of this paper. In the simplest case, the filter width in the time dimension is
only 1. A CNNs with k filters then behaves like a conventional feed-forward network without convolution that
aggregates EEG data for a single time point across all channels into k output values – with the only difference
that the CNN processes the whole recording as one input.
Such a convolutional auto-encoder with filter width 1 and linear activation functions is strongly related to PCA
(Bourlard & Kamp (1988)) and consequently obtains similar results. Using an auto-encoder, however, opens
up further possibilities such as non-linear activation functions, a filter width greater than 1, regularization of
channel weights and activations, and stacking multiple encoder/decoder layer pairs. This allows to identify
more complex components, which can also cover the time dimension.
Here, we used all training trials from the dataset for training which were preprocessed as described in Section 3.
The training trials have different lengths due to the different duration of the stimuli. As the network expects
inputs of equal length, shorter trials were zero-padded at the end to match the duration of the longest stimulus.12
Zero-padded trials should not be used for supervised training because this will result in the trial length being
used as a distinctive feature, which is not desirable. However, for unsupervised pre-training using basic autoencoding, the variable trial length is not problematic as the objective here is only to reconstruct.
The de-convolution filters were set to share the weights with the corresponding convolution filters. We experimented with different activation functions – linear, tanh and rectified linear (Glorot et al. (2010)) – and obtained
very similar topographies that only had small variations in intensity. We finally selected the tanh nonlinearity
because its output matches the value range of the network inputs ([-1,1]). Furthermore, we found that using
bias terms did not lead to improvements and therefore chose network units without bias.
Comparing the commonly used MSRE to the MCC measure showed that MSRE was the superior cost function
in this setting. The training process was generally very stable such that a high learning rate with linear decay
over epochs in combination with momentum could be used to quickly obtain results. As termination criterion,
we tracked the error on the training set for early stopping and specified a (rather defensive) maximum of 1000
epochs.
We first trained an auto-encoder for all trials of the perception condition. The learned filter weights were
then used as initialization values for individual auto-encoders for each subject. Minor differences between
subjects can be recognized in the example shown in Figure 3 but the component relations between subjects
are still obvious. This was generally the case for different network structures that we tested. The individual
components seem to be very stable. For instance, the results obtained with the same network on the imagination
trials with cue (condition 2; otherwise not considered in the context of this paper) only differ slightly as shown
in Figure 4.
11
https://github.com/sstober/deepthought (code will be updated paper publication)
This is a limitation of our current Pylearn2-based implementation. In principle, convolutional autoencoders can operate on variable-length input.
12
16
Under review as a conference paper at ICLR 2016
G.2
I MPLEMENTING A H YDRA -N ET
Hydra-nets allow to have separate processing pathways for subsets of a dataset – such as all trials of the same
subject or all trials belonging to the same stimulus. We call the respective meta-data property that controls the
choice of the individual pathway the selector. The selector data needs to be provided additionally to the regular
network input and, if necessary, has to be propagated (without being processed) up to the point where it is used
to choose the individual processing pathway. Individualization can range from a single layer with different
weights for each value of the selector up to multi-layer sub-networks that may also differ in their structure as
long as their input and output format is identical.
In our implementation, we wrapped the part of the network that we wanted to individualize with a hydra-net
layer. Within this wrapper layer, a copy of the internal network was created for each possible selector value. For
each input, the respective individualized version of the internal network was chosen based on the additionally
provided selector input. For mini-batch processing, we implemented two different strategies:
a) Compute the output of all individual pathways and apply a mask based on the selector values for the
whole mini-batch.
b) Iterate through the input instances within the mini-batch (using Theano’s scan operation) and process
each one individually solely with the selected individual pathway. (For selection, Theano’s ifelse
operation was used that implements a lazy evaluation.)
Option a) is highly inefficient as it results in an overhead of unnecessary computations. This effect gets worse
with an increasing number of different selector values. For option b), this is not the case. However, there is a
different performance penalty from using the scan function. In the experiments described in Section 4.3, we
found that option a) still performed slightly better. We plan to optimize our code and investigate different ways
to implement hydra-net layers for future experiments.
G.3
T RAINING A C ROSS - TRIAL E NCODER WITH C ROSS -S UBJECT T RIAL PAIRS
Figure 11 shows the network structure of the auto-encoder that learned the spatial filters shown in Figure 5.
Here, we used two separate hydra-net layers – one as encoder and the other one as decoder. Both shared their
weights, i.e., for the same subject, the convolution and deconvolution filter would be identical. The encoder
hydra-net layer used the subject of the input trial as selector (subject A) whereas the decoder hydra-net layer
used the subject of the paired trial (subject B). This general structure was identical for all cross-trial encoders
considered here except for the shape and number of the filters used.
For initialization, we first trained a basic auto-encoder on all 1296 (= 432x3) within-subject trial pairs derived
from the training set. This way, we obtained the general filters labeled “all” in Figures 1, 5 and 6. We then
continued trained a copy of the general cross-trial encoder for each individual subject. Finally, we used the
resulting individual filters as initialization for the hydra-net. This network was trained on 15120 (= 432x(4x9
- 1)) cross-subject trial pairs. In order to avoid excessive memory usage of the dataset, we implemented a
dataset wrapper that only stored an index data structure and generated mini-batches of paired trials and selector
meta-data on-the-fly from a base dataset when they were requested.
G.4
C ONSTRUCTING AND T RAINING A S IMILARITY-C ONSTRAINT E NCODER
Starting from a basic auto-encoder, a similarity-constraint encoder as shown schematically in Figure 12 can
be constructed as follows: The decoder part of the original auto-encoder can be removed as we are now only
interested in the internal feature representations and their similarity.13 Next, we redesign the encoder part such
that it can process multiple trials at the same time as a single input tuple. To ensure that all trials within a tuple
are processed in the exactly same way, weights and biases need to be shared between the parallel processing
pipelines. One easy way to achieve such a weight sharing for basic CNN-based architectures is by adding a third
dimension for the trials (additionally to the channels and time/samples axes) and using filters of size 1 along this
dimension. Unfortunately, this approach of convolution along the trials axis does not work together with hydranet layers. Here, different individualized weights for the different trials of the input tuple might be required
within the processing pipeline. In our implementation, we therefore wrap the processing pipeline within a
custom layer that takes care of processing each trial separately (using individualized weights if necessary) and
then computes the pair-wise similarities between the reference trial and the other trials. For k-tuples, this
results in k − 1 similarity scores. The final output layer only has to predict the trial with the highest value.
13
Note that simply maximizing the feature similarity for the paired trials used for cross-trial encoding would
be an ill-posed learning problem. The optimum could easily be achieved by transforming all inputs into the
same representation. This way, all trials would have maximum similarity – including those from different
classes. It would also be impractical to train the network with pre-defined target similarity values for all possible
pairs of trials.
17
Under review as a conference paper at ICLR 2016
Trial A
Encoder
Subject A
select
Decoder
select
(same weights as encoder)
Subject B
Trial B
minimize difference
Reconstruction
Figure 11: Cross-trial encoder training with cross-subject trial pairs (A,B). Both trials always belong
to the same class. Two separate hydra-net layers with shared weights (within subjects) were used
as encoder (convolution) and decoder (deconvolution). Encoder weights were selected based on the
subject meta-data of the input trial (A) whereas the decoder weights were selected based on the
subject meta-data of the paired trial (B). In this example, a trial from subject P06 is paired with a
trial from subject P11. The respective processing pathway for these selector values is highlighted
in blue. During backpropagation of the reconstruction error, only the weights of the filter weights
along the selected pathway are adapted.
This can, for instance, be accomplished with a Softmax (without affine transformation). A variety of similarity
measures could be used to compute the pair-wise similarities. We choose the dot product because this matches
the computation in the classifier output layer – independent of whether it is a Softmax or hinge loss (DLSVM)
layer. More complex approaches are possible as well, as long as they allow training through backpropagation –
such as first performing a channel-wise comparison and then aggregating the individual scores using trainable
weights. (virtual network structure)
Similarity-Constraint Encoder
former encoder
Input Triplet
Feature Extraction Pairwise Similarity
(dot product)
Reference
Input
Convolution
Deconvolution
Paired
InputTrial
Convolution
Similarity
Prediction
(binary)
Output
Classifier
Other
Input
Trial
Convolution
Similarity
(shared weights)
minimize constraint violations
Figure 12: Processing scheme of a similarity-constraint encoder as used in the experiments described
Stober - Deep
Learning
for EEG
in this paper. In general, theSebastian
feature extraction
partFeature
can also
be more
complex 2015-12-01
than shown46
here and
does not necessarily involve convolution.
The dataset for training comprised 57024 within-subject trial triplets. This number results from pairing each of
the 432 trials from the training set with the 3 other trials from the same class and subject and further combining
18
Under review as a conference paper at ICLR 2016
these pairs with the 11x4 trials from the same subject that belong to a different class. As for cross-trial encoder
training, we again implemented a dataset wrapper that only requires storing an index data structure for the
tuples from a base dataset.
G.5
S IMILARITY-C ONSTRAINT E NCODERS WITH H YDRA -N ET L AYERS
We believe that the combination of similarity-constraint encoding with the hydra-net approach to account for
individual differences between subjects has a high potential to further improve pre-trained global filters and
the resulting classification accuracy. An experiment with synthesized data has already shown the feasibility
of this combined approach. For this test, we generated random single-channel signals for 12 stimuli. These
signals were then added to 64-channel Gaussian noise, randomly choosing a different channel for each subject
to contain the relevant signal. As in our real dataset, we had 9 different subjects. Using these synthetic data,
a basic similarity-constraint encoder was able to identify the combined set of channels that contained relevant
signals for the group of all subject. Turning the first layer into a hydra-net layer allowed the spatial filters to
become more specific and the network correctly learned the individual channel masks.
Training on the real dataset with less defined signals and much more structured background noise is of course
much harder. However, the biggest obstacle is currently the increase in processing time caused by the added
computational complexity of the hydra-net layer (cf. Section G.2) with the amount of triplets available for
training. From the training set of 432 trials, 57024 within-subject trial triplets can be derived as described
earlier. Considering also cross-subject trial triplets increases the number to 5987520 (= 432x(4x9-1)x(11x4x9)).
Using training sets of this size is still feasible. However, the hydra-net processing overhead should be addressed
before attempting this.
19
Under review as a conference paper at ICLR 2016
H
C LASSIFIER P ERFORMANCE DURING H YPER -PARAMETER O PTIMIZATION
The following scatter plots show the estimated error (computed as the median over the 9-fold cross-validation
sets) and the actual test set error obtained for the averaged model (left) and the majority vote from the 9 fold
models (right) for all models trained during hyper-parameter optimization. Selected models listed in Table 1
are highlighted and labeled with their ID used in the table.
30
25
20
15
15
5
5
0
0
5
10 15 20 25 30 35 40
median validation set error (%)
average model
0
5
40
10 15 20 25 30 35 40
median validation set error (%)
majority vote
D
E
35
105
30
test set error (%)
30
25
20
15
25
20
10
5
5
0
5
10 15 20 25 30 35 40
median validation set error (%)
104
15
10
0
103
Baseline (512 Hz)
D
E
35
104
20
10
40
test set error (%)
25
10
0
A
B
C
35
test set error (%)
30
majority vote
0
0
20
5
number of model parameters (log scale)
A
B
C
35
test set error (%)
40
10 15 20 25 30 35 40
median validation set error (%)
number of model parameters (log scale)
average model
40
Baseline (64 Hz)
Under review as a conference paper at ICLR 2016
Common Components (64 Hz, global filters)
F
G
35
test set error (%)
30
25
20
15
25
15
10
5
5
0
5
0
10 15 20 25 30 35 40
median validation set error (%)
104
20
10
0
F
G
35
30
test set error (%)
majority vote
40
103
0
5
number of model parameters (log scale)
average model
40
10 15 20 25 30 35 40
median validation set error (%)
Common Components (64 Hz, individually adapted filters)
H
I
35
H
I
35
30
test set error (%)
30
test set error (%)
majority vote
40
25
20
15
25
15
10
10
5
5
0
0
0
5
10 15 20 25 30 35 40
median validation set error (%)
104
20
103
0
21
5
10 15 20 25 30 35 40
median validation set error (%)
number of model parameters (log scale)
average model
40
Under review as a conference paper at ICLR 2016
Cross-Trial Encoder (64 Hz, individually adapted filters, shape 1x[1x1]x64)
J
K
35
test set error (%)
30
25
20
15
25
15
10
5
5
0
5
0
10 15 20 25 30 35 40
median validation set error (%)
104
20
10
0
J
K
35
30
test set error (%)
majority vote
40
0
103
5
number of model parameters (log scale)
average model
40
10 15 20 25 30 35 40
median validation set error (%)
Cross-Trial Encoder (64 Hz, individually adapted filters, shape 1x[3x1]x64)
35
L
M
35
30
test set error (%)
30
test set error (%)
majority vote
40
L
M
25
20
15
25
15
10
10
5
5
0
0
0
5
10 15 20 25 30 35 40
median validation set error (%)
104
20
0
103
5
number of model parameters (log scale)
average model
40
10 15 20 25 30 35 40
median validation set error (%)
Cross-Trial Encoder (64 Hz, individually adapted filters, shape 4x[1x1]x64)
35
N
O
35
30
test set error (%)
30
test set error (%)
majority vote
40
N
O
25
20
15
25
15
10
10
5
5
0
0
0
5
10 15 20 25 30 35 40
median validation set error (%)
104
20
0
22
103
5
10 15 20 25 30 35 40
median validation set error (%)
number of model parameters (log scale)
average model
40
Under review as a conference paper at ICLR 2016
Similarity-Constraint Encoder (64 Hz, shape 1x[1x1]x64)
P
Q
R
35
30
25
20
15
25
15
10
5
5
0
5
0
10 15 20 25 30 35 40
median validation set error (%)
104
20
10
0
P
Q
R
35
test set error (%)
test set error (%)
30
majority vote
40
103
0
5
number of model parameters (log scale)
average model
40
10 15 20 25 30 35 40
median validation set error (%)
Similarity-Constraint Encoder (64 Hz, shape 1x[3x1]x64)
S
T
35
test set error (%)
30
25
20
15
25
15
10
5
5
0
5
10 15 20 25 30 35 40
median validation set error (%)
104
20
10
0
S
T
35
30
test set error (%)
majority vote
40
0
103
0
23
5
10 15 20 25 30 35 40
median validation set error (%)
number of model parameters (log scale)
average model
40
Under review as a conference paper at ICLR 2016
Similarity-Constraint Encoder (512 Hz, shape 1x[1x1]x64)
U
V
W
35
30
25
20
15
20
15
10
5
5
0
5
0
10 15 20 25 30 35 40
median validation set error (%)
105
25
10
0
U
V
W
35
test set error (%)
test set error (%)
30
majority vote
40
104
0
5
number of model parameters (log scale)
average model
40
10 15 20 25 30 35 40
median validation set error (%)
Similarity-Constraint Encoder (512 Hz, shape 1x[3x1]x64)
X
Y
35
test set error (%)
30
25
20
15
20
15
10
5
5
0
5
10 15 20 25 30 35 40
median validation set error (%)
105
25
10
0
X
Y
35
30
test set error (%)
majority vote
40
0
104
0
24
5
10 15 20 25 30 35 40
median validation set error (%)
number of model parameters (log scale)
average model
40
| 9cs.NE
|
The degrees of a system of parameters of the ring
of invariants of a binary form
arXiv:1404.5722v1 [math.AC] 23 Apr 2014
Andries E. Brouwer, Jan Draisma & Mihaela Popoviciu
2009-08-18, 2014-04-18
Abstract
We consider the degrees of the elements of a homogeneous system of
parameters for the ring of invariants of a binary form, give a divisibility
condition, and a complete classification for forms of degree at most 8.
1
The degrees of a system of parameters
Let R be a graded C-algebra. A homogeneous system of parameters (hsop) of
R is an algebraically independent set S of homogeneous elements of R such
that R is module-finite over the subalgebra generated by S. By the Noether
normalization lemma, a hsop always exists. The size |S| of S equals the Krull
dimension of R.
In this note we consider the special case where R is the ring I of invariants
of binary forms of degree n under the action of SL(2, C). This ring is CohenMacaulay, that is, I is free over the subring generated by any hsop S. Its Krull
dimension is n − 2.
One cannot expect to classify all hsops of I. Indeed, any generic subset with
the right degrees will be a hsop (cf. Dixmier’s criterion below). But one can
expect to classify the sets of degrees of hsops. In this note we give a divisibility
restriction on the set of degrees for the elements of a hsop, and conjecture that
when all degrees are large this restriction also suffices for the existence of a hsop
with these given degrees. For small degrees there are further restrictions. We
give a complete classification for n ≤ 8.
2
Hilbert’s criterion
Hilbert’s criterion gives a characterization of homogeneous systems of parameters as sets that define the nullcone.
Denote by Vn the set of binary forms of degree n. The nullcone of Vn ,
denoted N (Vn ), is the set of binary forms of degree n on which all invariants
vanish. By the Hilbert-Mumford numerical criterion (see [6] and [7, Chapter 2])
this is precisely the set of binary forms of degree n with a root of multiplicity
> n2 . Moreover, the binary forms with no root of multiplicity ≥ n2 have closed
SL(2, C)-orbits. The elements of N (Vn ) are called nullforms. Another result
from [6] that we will use is the following.
1
Proposition 2.1. For n ≥ 3, consider i1 , . . . , in−2 homogeneous invariants of
Vn . The following two conditions are equivalent:
(i) N (Vn ) = V(i1 , . . . , in−2 ),
(ii) {i1 , . . . , in−2 } is a hsop of the invariant ring of Vn .
3
A divisibility condition
Assume n ≥ 3.
Lemma 3.1.
integers j, t with t > 0. If an invariant of degree d is nonzero
P Fixn−i
on a form
ai x y i with the property that all nonzero ai have i ≡ j (mod t),
then d(n − 2j)/2 ≡ 0 (mod t).
Q mi
an invariant of degree d with nonzero term
ai we have
P Proof For P
m
=
d
and
im
=
nd/2.
If
i
≡
j
(mod
t)
when
a
=
6
0,
then nd/2 =
i
i
i
P
P
imi ≡ j mi = jd (mod t).
For odd n we recover the well-known fact that all degrees are even (take t = 1).
Lemma 3.2. Fix integers j, t with t > 1 and 0 ≤ j ≤ n. Among the degrees d
of a hsop, at least ⌊(n − j)/t⌋ satisfy d(n − 2j)/2 ≡ 0 (mod t).
Proof Subtracting a multiple of t from j results in a stronger statement,
so it suffices to prove the lemma for 0 ≤ j < t. There are 1+⌊(n−j)/t⌋ =: 1+N
coefficients ai with i ≡ j (mod t), so the subpace U of Vn defined by ai = 0 for
i 6≡ j (mod t) has dimension 1 + N . If N = 0 there is nothing to prove, so we
assume that N > 0. We claim that a general form f ∈ U has only zeroes of
multiplicity strictly less than n/2. Indeed, write
f = aj xn−j y j + aj+t xn−j−t y j+t + . . . + aj+mt xn−j−mt y j+mt
where j + (m + 1)t > n and m > 0. So f has a factor y of multiplicity j and a
factor x of multiplicity n − j − mt. If j were at least n/2, then j + mt ≥ j + t >
2j ≥ n, a contradiction. If n − j − mt were at least n/2, then j + mt ≤ n/2
and hence t ≤ n/2 and hence j + (m + 1)t ≤ n, a contradiction. The remaining
roots of f are roots of
aj xmt + aj+t x(m−1)t y t + . . . + aj+mt y mt ,
which is a general binary form of degree m in xt , y t and hence has mt distinct
roots.
Let π : Vn → Vn //SL(2, C) be the quotient map; so the right-hand side is
the spectrum of the invariant ring I. Set X := π(U ). We claim that X has
dimension N . It certainly cannot have dimension larger than N , since acting
with the one-dimensional torus of diagonal matrices on an element of U gives
another element of U . To show that dim X = N we need to show that for
general f ∈ U the fibre π −1 (π(f )) intersects U in a one-dimensional variety. By
the above and the Hilbert-Mumford criterion, the SL(2, C)-orbit of f is closed.
Moreover, its stabiliser is zero-dimensional. So by properties of the quotient
map we have π −1 (π(f )) = SL(2, C) · f . Hence it suffices that the intersection of
this orbit with U is one-dimensional. For this a Lie algebra argument suffices,
2
∂
∂
+ cy ∂x
)f lies in U ,
in which we may ignore the Lie algebra of the torus: if (bx ∂y
then we find that b = c = 0 if t > 2 (so that the contribution of one term from
f cannot cancel the contribution from the next term); and b = 0 if j > 0 (look
at the first term), and then also c = 0; and c = 0 if j + mt < n (look at the last
term), and then also b = 0. Hence the only case that remains is t = 2, j = 0,
and n ≥ 4 even. Then the equations ca0 n + ba2 2 = 0 and ca2 (n − 2) + ba4 4 = 0
are independent and force b = c = 0.
This concludes the proof that dim X = N . Intersecting X with the hypersurfaces corresponding to elements of an hsop reduces X to the single point in
X representing the null-cone. In the process, dim X drops by N . But the only
invariants that contribute to this dimension drop, i.e., the only invariants that
do not vanish identically on X (hence on U ) are those considered in Lemma 3.1.
Hence there must be at least N of these among the hsop.
Lemma 3.3. Let t be an integer with t > 1.
(i) If n is odd, and j is minimal such that 0 ≤ j ≤ n and (n − 2j, t) = 1,
then among the degrees of any hsop at least ⌊(n − j)/t⌋ are divisible by 2t.
(ii) If n is even, and j is minimal with 0 ≤ j ≤ 21 n and ( 12 n − j, t) = 1, then
among the degrees of any hsop at least ⌊(n − j)/t⌋ are divisible by t.
Theorem 3.4. Let t be an integer with t > 1.
(i) If n is odd, then among the degrees of any hsop at least ⌊(n − 1)/t⌋ are
divisible by 2t (and all degrees are even).
(ii) If n is even, then among the degrees of any hsop at least ⌊(n − 1)/t⌋ are
divisible by t, and if n ≡ 2 (mod 4) then at least n/2 by 2.
Proof (i) By part (i) of Lemma 3.3 we find a lower bound ⌊(n − j)/t⌋ for
a j as described there. If that is smaller than ⌊(n − 1)/t⌋, then there is some
multple at of t with n − j + 1 ≤ at ≤ n − 1. Put n = at + b, where 1 ≤ b ≤ j − 1.
By definition of j we have (b − 2i, t) > 1 for i = 0, 1, ..., j − 1. If b is odd, say
b = 2i + 1, we find a contradiction. If b is even, say b = 2i + 2, then t is even
and n is even, contradiction.
(ii) By part (ii) of Lemma 3.3 we find a lower bound ⌊(n − j)/t⌋ for a j as
described there. For t = 2 our claim follows. Now let t > 2. If ⌊(n − j)/t⌋ is
smaller than ⌊(n − 1)/t⌋, then there is some multple at of t with n − j + 1 ≤
at ≤ n − 1. Put n = at + b, where 1 ≤ b ≤ j − 1. By definition of j we have
(b − 2i, 2t) > 2 for i = 0, 1, ..., j − 1, impossible.
For example, it is known that there exist homogeneous systems of parameters
with degree sequences 4 (n = 3); 2, 3 (n = 4); 4, 8, 12 (n = 5); 2, 4, 6, 10 (n = 6);
4, 8, 12, 12, 20 and 4, 8, 8, 12, 30 (n = 7) [3]; 2, 3, 4, 5, 6, 7 (n = 8) [10]; 4, 8,
10, 12, 12, 14, 16 and 4, 4, 10, 12, 14, 16, 24 and 4, 4, 8, 12, 14, 16, 30 and 4,
4, 8, 10, 12, 16, 42 and 4, 4, 8, 10, 12, 14, 48 (n = 9) [1]; 2, 4, 6, 6, 8, 9, 10, 14
(n = 10) [2].
Conjecture 3.5. Any sequence d1 , ..., dn−2 of sufficiently large integers satisfying the divisibility conditions of Theorem 3.4 is the sequence of degrees of a
hsop.
This can be compared to the conjecture
3
Conjecture 3.6. (Dixmier[4])
(i) If n is odd, n ≥ 15, then 4, 6, 8, ..., 2n − 2 is the sequence of degrees of a
hsop.
(ii) If n ≡ 2 (mod 4), n ≥ 18, then 2, 4, 5, 6, 6, 7, 8, 9, ..., n − 1 is the sequence
of degrees of a hsop.
(iii) If n ≡ 0 (mod 4), then 2, 3, 4, ..., n − 1 is the sequence of degrees of a
hsop.
4
Poincaré series
If there exists a hsop with degrees d1Q
, . . . , dn−2 , then the Poincaré series can
be written as a quotient P (t) = a(t)/ (tdi − 1) for some polynomial a(t) with
nonnegative coefficients. If one does not have a hsop, but only a sequence of degrees, the conditions of Theorem 3.4 above are strong enough to guarantee that
P (t) can be written in this way, but without the condition that the numerator
has nonnegative coefficients.
Proposition 4.1. Let d1 , . . . , dn−2 be a sequence
of positive integers satisfying
Q
the conditions of Theorem 3.4. Then P (t) (tdi − 1) is a polynomial.
Proof Dixmier [4] proves that P (t)B(t) is a polynomial, where B(t) is
defined by
Qn−1
2i
if n is odd
Qi=2 (1 − t )
n−1
i
B(t) =
(1
−
t
).(1
+
t)
if n ≡ 2 (mod 4)
i=2
Qn−3
i
(n−2)/2
n−1
)(1 − t
) if n ≡ 0 (mod 4)
i=2 (1 − t ).(1 + t)(1 − t
Consider a primitive t-th root of unity ζ. We have to show that if B(t) has
root ζ with multiplicity m, then at least m of the di are divisible by t, but this
follows immediately from Theorem 3.4. Note that in case n ≡ 0 (mod 4) the
factor (1 + t)(1 − t(n−2)/2 ) divides (1 − tn−2 ).
We see that if n ≡ 0 (mod 4), n > 4, then P (t) can be written with a smaller
denominator than corresponds to the degrees of a hsop.
We shall need the first few coefficients of P (t). Messy details arise for small
n because there are too few invariants of certain small degrees. Let I be the
ring of invariants of a binary form of degree (order) n, let Im be the
P graded part
of I of degree m, and put hm = hnm = dimC Im , so that P (t) = m hm tm .
The coefficients hnm can be computed by the Cayley-Sylvester formula: The
dimension of the space of covariants of degree m and order a is zero when mn−a
is odd, and equals N (n, m, t) − N (n, m, t − 1) if nm − a = 2t, where N (n, m, t)
is the number of ways t can be written as sum of m integers in the range 0..n,
that is, the number of Ferrers diagrams of size t that fit into a m × n rectangle.
We have Hermite reciprocity hnm = hm
n , as follows immediately since reflection in the main diagonal shows N (n, m, t) = N (m, n, t). That means that
Table 1 is symmetric.
Dixmier [4] gives the cases in which hm = 0. Since his statement is not
precisely accurate, we repeat his proof.
4
hn
m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
.
1
.
1
.
1
.
1
.
1
.
1
.
1
.
1
.
1
3
.
.
.
1
.
.
.
1
.
.
.
1
.
.
.
1
.
.
4
.
1
1
1
1
2
1
2
2
2
2
3
2
3
3
3
3
4
5
.
.
.
1
.
.
.
2
.
.
.
3
.
.
.
4
.
1
6
.
1
.
2
.
3
.
4
.
6
.
8
.
10
1
13
1
16
7
.
.
.
1
.
.
.
4
.
.
.
10
.
4
.
18
.
13
8
.
1
1
2
2
4
4
7
8
12
13
20
22
31
36
47
54
71
9
.
.
.
2
.
.
.
8
.
5
.
28
.
27
.
84
.
99
10
.
1
.
2
.
6
.
12
5
24
13
52
33
97
80
177
160
319
11
.
.
.
2
.
.
.
13
.
13
.
73
.
110
.
320
.
529
12
.
1
1
3
3
8
10
20
28
52
73
127
181
291
418
639
902
1330
13
.
.
.
2
.
.
.
22
.
33
.
181
.
375
.
1120
.
2342
14
.
1
.
3
.
10
4
31
27
97
110
291
375
802
1111
2077
2930
5034
15
.
.
.
3
.
1
.
36
.
80
.
418
.
1111
.
3581
.
8899
Table 1: Values of hnm = dimC Im with I the ring of invariantsP
of a binary form
n m
of degree n. Here . denotes 0. One has hnm = hm
and
P
(t)
=
n
m hm t .
Proposition 4.2. Let m, n ≥ 1. One has hm = hnm = 0 precisely in the
following cases:
(i) if mn is odd,
(ii) if m = 1; if n = 1,
(iii) if m = 2 and n is odd; if n = 2 and m is odd,
(iv) if m = 3 and n ≡ 2 (mod 4); if n = 3 and m ≡ 2 (mod 4),
(v) if m = 5 and n = 6, 10, 14; if n = 5 and m = 6, 10, 14,
(vi) if m = 6 and n = 7, 9, 11, 13; if n = 6 and m = 7, 9, 11, 13,
(vii) if m = 7 and n = 10; if n = 7 and m = 10.
Proof (i) If n is odd, then all degrees are even. (ii) For n = 1 we have
P (t) = 1. (iii) For n = 2 we have P (t) = 1/(1 − t2 ). (iv) For n = 3 we have
P (t) = 1/(1 − t4 ). Now let m, n ≥ 4. For n = 4 we have invariants of degrees 2,
3 and hence of all degrees m 6= 1. That means that hn4 6= 0. For n = 6 we have
invariants of degrees 2, 15 and hence of all degrees m ≥ 14. That means that
hn6 6= 0 for n ≥ 14. If n is odd this shows the presence of invariants of degrees
4, 6 and hence of all even degrees m > 2, provided n ≥ 15. For n = 5 we have
invariants of degrees 4, 18 and hence of all even degrees m ≥ 16. That means
that hn5 6= 0 for even n ≥ 16. If n is even this shows the presence of invariants
of degrees 2, 5 and hence of all degrees m ≥ 4, provided n ≥ 16. It remains only
to inspect the table for 4 ≤ m, n ≤ 14.
5
5
Dixmier’s criterion
Dividing out the ideal spanned by p elements of a hsop diminishes the dimension
by precisely (and hence at least) p. This means that the below gives a necessary
and sufficient condition for a sequence of degrees to be the degree sequence of
a hsop.
Proposition 5.1. (Dixmier [4]) Let G be a reductive group over C, with a
rational representation in a vector space R of finite dimension over C. Let
C[R] be the algebra of complex polynomials on R, C[R]G the subalgebra of Ginvariants, and C[R]G
d the subset of homogeneous polynomials of degree d in
C[R]G . Let V be the affine variety such that C[V ] = C[R]G . Let r = dim V . Let
(d1 , . . . , dr ) be a sequence of positive integers. Assume that for each subsequence
(j1 , . . . , jp ) of (d1 , . . . , dr ) the subset of points of V where all elements of all
C[R]G
j with j ∈ {j1 , . . . , jp } vanish has codimension not less than p in V . Then
C[R]G has a system of parameters of degrees d1 , . . . , dr .
This criterion is very convenient, it means that one can work with degrees
only, without worrying about individual elements of a hsop.
6
Minimal degree sequences
If y1 , ..., yr is a hsop, then also y1e1 , ..., yrer for any sequence of positive integers
e1 , ..., er , not all 1. This means that if the degree sequence d1 , ..., dr occurs,
also the sequence d1 e1 , ..., dr er occurs. We would like to describe the minimal
sequences, where such multiples are discarded.
There are further reasons for non-minimality.
Lemma 6.1. If there exist hsops with degree sequences d1 , ..., dr−1 , d′ and d1 , ...,
dr−1 , d′′ , then there also exists a hsop with degree sequence d1 , ..., dr−1 , d′ + d′′ .
Proof We verify Dixmier’s criterion. Consider a finite basis f1 , ..., fs for
the space of invariants of degree d′ . Split the variety V in the s pieces defined
by fi 6= 0 (1 ≤ i ≤ s) together with the single piece defined by f1 = ... = fs = 0.
Given p elements of the sequence d1 , ..., dr−1 , d′ + d′′ we have to show that the
codimension in V obtained by requiring all invariants of such degrees to vanish
is at least p, that is, that the dimension is at most r − p. This is true by
assumption if d′ + d′′ is not among these p elements. Otherwise, consider the
s + 1 pieces separately. We wish to show that each has dimension at most r − p,
then the same will hold for their union. For the last piece, where all invariants
of degree d′ vanish, this is true by assumption. But if some invariant of degree
d′ does not vanish, and all invariants of degree d′ + d′′ vanish, then all invariants
of degree d′′ vanish, and we are done.
Note that taking multiples is a special case of (repeated application of) this
lemma, used with d′ = d′′ .
Let us call a sequence minimal if it occurs (as the degree sequence of the
elements of a hsop), and its occurrence is not a consequence, via the above
lemma or via taking multiples, of the occurrence of smaller sequences. We
might try to classify all minimal sequences, at least in small cases.
6
Is it perhaps true that a hsop exists for any degree sequence that satisfies
the conditions of Theorem 3.4Qwhen there are sufficiently many invariants? E.g.
when the coefficients of P (t) (1 − tdi ) are nonnegative?
Example Some caution is required. For example, look at n = 6. The conditions
of Theorem 3.4 are: at least three factors 2, at least one factor of each of 3, 4,
5. The sequence 6, 6, 6, 20 satisfies this restriction. Moreover, P (t)(1 − t6 )3 (1 −
t20 ) = 1 + t2 + 2t4 + t8 + 2t12 + t14 + t15 + t16 + t17 + 2t19 + t23 + 2t27 + t29 + t31
has only nonnegative coefficients. But no hsop with these degrees exists: since
h2 = 1, h4 = 2, h6 = 3 it follows that there are invariants i2 , i4 , i6 of degrees 2,
4, 6, and we have I4 = hi22 , i4 i and I6 = hi32 , i2 i4 , i6 i. Requiring all invariants of
degree 6 to vanish is equivalent to the two conditions i2 = i6 = 0, and a hsop
cannot contain three elements of degree 6.
Still, the above conditions almost suffice. And for n < 6 they actually do
suffice.
6.1
n=3
For n = 3 we only have simple multiples of the minimal degree.
Proposition 6.2. A positive integer d is the degree of a hsop in case n = 3 if
and only if it is divisible by 4.
If i4 is an invariant of degree 4, then {i4 } is a hsop.
6.2
n=4
For n = 4 one has the sequence 2, 3, but for example also 5, 6.
Proposition 6.3. A sequence d1 , d2 of two positive integers is the sequence of
degrees of a hsop for the quartic if and only if neither of them equals 1, at least
one is divisible by 2, and at least one is divisible by 3.
Proof Clearly the conditions are necessary. In order to show that they
suffice apply induction and the known existence of a hsop with degrees 2, 3.
If d2 > 7, then apply Lemma 6.1 to the two sequences d1 , 6 and d1 , d2 − 6 to
conclude the existence of a hsop with degrees d1 , d2 . If 2 ≤ d1 , d2 ≤ 7 and one
is divisible by 2, the other by 3, then we have a multiple of the sequence 2, 3.
Otherwise, one equals 6 and the other is 5 or 7. But 5, 6 is obtained from 2, 6
and 3, 6, and 7, 6 is obtained from 2, 6 and 5, 6.
If i2 and i3 are invariants of degrees 2 and 3, then {i2 , i3 } is a hsop.
Proposition 6.4. There is precisely one minimal degree sequence of hsops in
case n = 4, namely 2, 3.
6.3
n=5
Proposition 6.5. A sequence d1 , d2 , d3 of three positive integers is the sequence
of degrees of a hsop for the quintic if and only if all di are even, and distinct
from 2, 6, 10, 14, and no two are 4, 4 or 4, 22 and at least two are divisible by
4, at least one is divisible by 6, and at least one is divisible by 8.
7
Proof For n = 5 the Poincaré series is P (t) = 1 + t4 + 2t8 + 3t12 + 4t16 +
t18 + 5t20 + t22 + 7t24 + 2t26 + 8t28 + 3t30 + .... The stated conditions are
necessary: the divisibility conditions are seen from Theorem 3.4, and there are
no invariants of degrees 2, 6, 10, 14. Finally, we have h4 = 1 and h18 = h22 = 1,
so that there are unique invariants i4 and i18 of degrees 4 and 18, respectively,
and I22 = hi4 i18 i, so that all invariants of degree 22 will vanish as soon as i4
vanishes.
The stated conditions suffice: We use (and verify below) that there are hsops
with degrees 4, 8, 12 and with degrees 4, 8, 18. If all di are divisible by 4, and
we do not have a multiple of 4, 8, 12, then we have 4a, 4b, 24c where a and b
have no factor 2 or 3, and not both are 1. It suffices to find 4, 4b, 24. Since 4, 8,
24 exists, we can decrease b by 2, and it suffices to find 4, 12, 24, which exists.
So, some di , is not divisible by 4. We have one of the three cases 24a, 4b, 2c
and 8a, 12b, 2c and 8a, 4b, 6c, where c is odd. In the middle case we have c ≥ 9
and it suffices to make 8, 12, 2c. Since 8, 12, 4 exists, we can reduce c by 2, and
it suffices to make 8, 12, 18, which exists since 4, 8, 18 exists.
In the first case we have c ≥ 9 and it suffices to make 24, 4, 2c. Since 12, 4,
8 exists, we can reduce c by 4, and it suffices to make 24, 4, 18 and 24, 4, 30.
The former is a multiple of 4, 8, 18 and the latter follows from 24, 4, 18 and 24,
4, 12. Since 24, 4, 22 does not exist we still have to consider 24a, 4b, 22. Since
8, 12, 22 exists we can reduce b by 2, and it suffices to make 24, 12, 22. But
that is a multiple of 8, 12, 22.
Finally in the last case we have c ≥ 3, and since 8, 4, 12 exists we can reduce
c by 2. So it suffices to do 4, 8, 18, and that exists.
Proposition 6.6. There are precisely two minimal degree sequences of hsops
in case n = 5, namely 4, 8, 12 and 4, 8, 18.
Proof By the proof of the previous proposition, all we have to do is show
the existence of hsops with the indicated degree sequences. It is well-known
(see, e.g., Schur [9], p.86) that the quintic has four invariants i4 , i8 , i12 , i18
(with degrees as indicated by the index) that generate the ring of invariants,
and every invariant of degree divisible by 4 (in particular i218 ) is a polynomial
in the first three. Thus, when i4 , i8 , i12 vanish, all invariants vanish, and
{i4 , i8 , i12 } is a hsop. Knowing this, it is easy to see that also {i4 , i8 , i18 } is
a hsop: a simple Groebner computation shows that i312 ∈ (i4 , i8 , i18 ), hence
N (V5 ) = V(i4 , i8 , i18 ).
6.4
n=6
Similarly, we find for n = 6:
Proposition 6.7. A sequence d1 , d2 , d3 , d4 of four positive integers is the sequence of degrees of a hsop for the sextic if and only if all di are distinct
from 1, 3, 5, 7, 9, 11, 13, and no two are in {2, 17}, and no three are in
{2, 4, 8, 14, 17, 19, 23, 29}, and no three are in {2, 6, 17, 21}, and at least three
are divisible by 2, at least one is divisible by 3, at least one by 4, and at least
one by 5.
8
Proof For n = 6 the Poincaré series is
P (t)
= 1 + t2 + 2t4 + 3t6 + 4t8 + 6t10 + 8t12 + 10t14 + t15 + 13t16 + t17 +
16t18 + 2t19 + 20t20 + 3t21 + 24t22 + 4t23 + 29t24 + 6t25 + 34t26 +
8t27 + 40t28 + 10t29 + 47t30 + · · · .
We have
I2 = hi2 i, I4 = hi22 , i4 i, I6 = hi32 , i2 i4 , i6 i, I8 = hi42 , i22 i4 , i2 i6 , i24 i,
I10 = hi52 , i32 i4 , i22 i6 , i2 i24 , i4 i6 , i10 i, I12 = hi62 , i42 i4 , i32 i6 , i22 i24 , i2 i4 i6 , i2 i10 , i34 , i26 i,
I14 = hi72 , i52 i4 , i42 i6 , i32 i24 , i22 i4 i6 , i22 i10 , i2 i34 , i2 i26 , i24 i6 , i4 i10 i, I15 = hi15 i,
and the invariants in degrees 17, 19, 23, 29 are i15 times the invariants in
degrees 2, 4, 8, 14, respectively. Let us denote by [i1 , ..., it ] the condition that all
invariants of degrees i1 , ..., it vanish. Then [2] = [2, 17] and hence a hsop cannot
have two element degrees among 2, 17. Also [4] = [2, 4, 8, 14, 17, 19, 23, 29] and
hence a hsop cannot have three element degrees among 2, 4, 8, 14, 17, 19, 23, 29.
And [6] = [2, 6, 17, 21] is the condition i2 = i6 = 0 so that a hsop cannot have
three element degrees among 2, 6, 17, 21. It follows that the stated conditions
are necessary.
The stated conditions suffice: We use (and verify below) that there are hsops
with each of the degree sequences 2, 4, 6, 10 and 2, 4, 6, 15 and 2, 4, 10, 15.
Prove by induction that any 4-tuple of degrees that satisfies the given conditions
occurs as the degree sequence of a hsop. Given d1 , d2 , d3 , d4 , if di ≥ 90 then by
induction we already have the 4-tuples obtained by replacing di by 60 and by
di − 60. It remains to check the finitely many cases where all di are less than
90. A small computer check settles this.
Proposition 6.8. There are precisely three minimal degree sequences of hsops
in case n = 6, namely 2, 4, 6, 10 and 2, 4, 6, 15 and 2, 4, 10, 15.
Proof By the proof of the previous proposition, all we have to do is show
the existence of hsops with the indicated degree sequences. It is well-known (see,
e.g., Schur [9], p.90) that the sextic has five invariants i2 , i4 , i6 , i10 , i15 (with
degrees as indicated by the index) that generate the ring of invariants, where
i215 is a polynomial in the first four. This implies that N (V6 ) = V(i2 , i4 , i6 , i10 ),
so that {i2 , i4 , i6 , i10 } is a hsop. Now {i2 , i4 , i6 , i15 } and {i2 , i4 , i10 , i15 } are also
hsops: we verified by computer that i310 ∈ (i2 , i4 , i6 , i15 ) and i56 ∈ (i2 , i4 , i10 , i15 ),
so that N (V6 ) = V(i2 , i4 , i6 , i15 ) = V(i2 , i4 , i10 , i15 ).
6.5
n=7
For n = 7 we have to consider the invariants a bit more closely in order to decide
which degree sequences are admissable for hsops.
Let f be our septimic and let ψ be the covariant ψ = (f, f )6 . There are
thirty basic invariants, of degrees 4, 8 (3×), 12 (6×), 14 (4×), 16 (2×), 18 (9×),
20, 22 (2×), 26, 30. These can all be taken to be transvectants with a power
of ψ except for three basic invariants of degrees 12, 20 and 30 (that von Gall
[5] calls R, A, B and Dixmier [3] q12 , p20 , p30 ). This means that all invariants
9
of degrees not of the form 12a + 20b + 30c vanish on the set defined by ψ = 0.
But ψ is a covariant of order 2, i.e., ψ = Ax2 + Bxy + Cy 2 for certain A, B,
C. It follows that no hsop degree sequence can have four elements in the set
{4, 8, 14, 16, 18, 22, 26, 28, 34, 38, 46, 58}.
Proposition 6.9. A sequence of five positive even integers is the sequence of
degrees of a hsop for the septimic if and only if all are distinct from 2, 6, 10, no
two equal 4, no four are in {4, 8, 14, 16, 18, 22, 26, 28, 34, 38, 46, 58} and at least
three are divisible by 4, at least two by 6, at least one by 8, at least one by 10
and at least one by 12.
Proof We already saw that these conditions are necessary. For sufficiency,
use induction. The divisibility conditions concern moduli with l.c.m. 120, and
the restrictions concern numbers smaller than 60, so if one of the degrees is not
less than 180, we are done by induction. A small computer program checks all
degree sequences with degrees at most 180, and finds that all can be reduced to
the 23 sequences given in the following proposition.
Proposition 6.10. There are precisely 23 minimal degree sequences of hsops
in case n = 7, namely
4, 8, 8, 12, 30
4, 8, 12, 12, 20
4, 8, 12, 12, 30
4, 8, 12, 14, 30
4, 8, 12, 18, 20
4, 8, 12, 18, 30
4, 12, 12, 12, 40
4, 12, 12, 14, 40
4, 12, 12, 18, 40
4, 12, 14, 14, 120
4, 12, 14, 18, 40
4, 12, 14, 20, 24
4, 12, 18, 18, 40
4, 14, 14, 24, 60
4, 14, 18, 20, 24
4, 14, 18, 32, 60
4, 18, 18, 20, 24
4, 18, 18, 32, 60
8, 12, 12, 14, 20
8, 12, 14, 14, 60
8, 12, 14, 18, 20
12, 12, 14, 14, 40
12, 14, 14, 20, 24
Proof We only have to show existence. Apply Dixmier’s criterion. Denote
by [d1 , ..., dp ] the codimension in V of the subset of points of V where all elements
of all C[R]G
dj vanish (1 ≤ j ≤ p). We have to show that for all p and each of
these 23 sequences (di ) the inequality [d1 , ..., dp ] ≥ p holds.
For p = 1 that means that we need [m] ≥ 1 for m = 4, 8, 12, 14, 18, 20, 24,
30, 32, 40, 60, 120, and that is true, for example by inspection of Table 1.
We can save some work by observing that Dixmier [3] already showed the
existence of hsops with degree sequences 4, 8, 8, 12, 30 and 4, 8, 12, 12, 20.
It follows that [8] ≥ 3 and [12] ≥ 3 and [24] ≥ [8, 12] ≥ 4 and [20] ≥ 2 and
[60] ≥ [12, 20] ≥ 4 and [4, 30] ≥ 2 and [8, 30] ≥ 4. Since there are several basic
invariants of degree 14 or 18, no two of which can have a common factor, it
follows that [14] ≥ 2 and [18] ≥ 2. This suffices to settle p = 2.
For p = 3 we must look at triples [d, d′ , d′′ ] without element 8 or 12 or
multiple. First check that [4, 14] ≥ 3 and [4, 18] ≥ 3. We’ll do this below. Now
all the rest needed for p = 3 follows.
Below we shall show that [12] ≥ 4. For p = 4 we must look at quadruples [d, d′ , d′′ , d′′′ ] without element 12 or 8, 30 or multiple. The minimal of
these are (omitting implied elements) [18, 20] and [18, 32]. However, [18, 32] ≥
min([18, 12], [18, 20]) and [18, 20] ≥ min([18, 20, 8], [18, 20, 12]).
Finally for p = 5 we have to show that each of these 23 sets determines
the nullcone. But that follows immediately, since it is known already that
[8, 12, 20] = [8, 12, 30] = 5.
10
Altogether, our obligations are: show that [4, 14] ≥ 3, [4, 18] ≥ 3, [12] ≥ 4
and [8, 18, 20] ≥ 4.
Consider the part of V defined by ψ = 0. Dixmier shows that if ψ = q12 =
p20 = 0 (for certain invariants q12 and p20 of degrees 12 and 20, respectively),
then f is a nullform. It follows that the subsets of V defined by ψ = q12 = 0 or
by ψ = p20 = 0 have codimension at least 4 in V .
Now we have to do some actual computations. With f = ax7 + 71 bx6 y +
· · · + 71 gxy 6 + hy 7 (the two meanings of f , as form and as coefficient will not
cause confusion), we find ψ = (ag − 6bf + 15ce − 10d2 )x2 + (ah − 5bg + 9cf −
5de)xy + (bh − 6cg + 15df − 10e2 )y 2 .
Assume that the invariant of degree 4 vanishes, as it does in all cases we still
have to consider. Then ψ has zero discriminant. If ψ 6= 0, then w.l.o.g. ψ ∼ x2 ,
and ah−5bg +9cf −5de = bh−6cg +15df −10e2 = 0, ag −6bf +15ce−10d2 6= 0.
Distinguish the four cases (i) h 6= 0, (ii) h = 0, g 6= 0, (iii) h = g = 0,
f 6= 0, (iv) h = g = f = 0, e 6= 0. W.l.o.g. these become (i) h = 1, g = 0,
a + 9cf − 5de = 0, b + 15df − 10e2 = 0, (ii) h = 0, g = 1, f = 0, b + de = 0,
3c + 5e2 = 0, (iii) h = g = 0, f = 1, e = 0, c = 0, d = 0, b 6= 0, (iv)
h = g = f = 0, e = 1, d = 0, contradiction.
Let us first show that [12] ≥ 4. We may suppose ψ 6= 0. One of the
invariants of degree 12 is (ψ1 , ψ 5 )10 ∼ (ψ1 , x10 )10 = f h − g 2 , where ψ1 = (f, f )2 .
If all invariants of degree 12 vanish, then in case (i) f = 0, and in case (ii)
contradiction. Look at case (iii). The only invariant of degree 12 that does not
vanish identically is a2 b2 f 8 , and we find a = 0, a 1-dimensional set. Finally, in
case (i), if all invariants of degree 12 vanish, but ag − 6bf + 15ce − 10d2 6= 0,
then the remaining conditions define an ideal (18e3 − cd, 12de2 − c2 , 2cd2 − 3c2 e)
in the three variables c, d, e and the quotient is 1-dimensional. This shows that
[12] ≥ 4.
Let us show next that [8, 18] ≥ 4. We may suppose ψ 6= 0. One of the
invariants of degree 8 is (ψ2 , ψ 3 )6 ∼ (ψ2 , x6 )6 = dh − 4eg + 3f 2 where ψ2 =
(f, f )4 . This gives a contradiction in case (iii). In case (ii) it gives e = b = c = 0,
leaving only variables a, d. In case (i) it gives d + 3f 2 = 0, leaving only variables
c, e, f .
An invariant of degree 18 is ((ψ1 , ψ2 )1 , ψ 7 )14 ∼ ((ψ1 , ψ2 )1 , x14 )14 = −cf h2 +
2
cg h + deh2 + 2df gh − 3dg 3 − 4e2 gh + ef 2 h + 6ef g 2 − 3f 3 g. In case (ii) this says
d = 0, leaving only variable a. In case (i) this says f (2ef + c) = 0. This gives
us two subcases: (ia) with f = 0 and variables c, e, and (ib) with c + 2ef = 0
and variables e, f .
Another invariant of degree 8 is (ψ3 , ψ 2 )4 ∼ (ψ3 , x4 )4 , where ψ3 = (ψ2 , ψ2 )4 ,
which vanishes in case (ii) and says c2 f + 4cef 2 + 76e2 f 3 + 9e4 + 144f 6 = 0 in
case (i). In case (ia) this means e = 0 leaving only variable c. In case (ib) this
means (4f 3 + e2 )2 = 0, leaving the dimension 1. This proves [8, 18] ≥ 4.
Let us show next that [4, 14] ≥ 3. First consider the case ψ = 0. Now all
invariants of degrees 4 or 14 (or 18) vanish, but the condition ψ = 0 itself yields
the three equations A = B = C = 0 where ψ = Ax2 + Bxy + Cy 2 . Earlier, the
choice ψ ∼ x2 used up some of the freedom given by the group, but here we are
free to choose a zero for the form, and assume h = 0. Again consider the four
cases, this time with ag − 6bf + 15ce − 10d2 zero instead of nonzero. We have
(iii) f = 1, h = g = e = d = c = b = 0, only variables a, f left. And (ii) g = 1,
11
h = f = 0, b + de = 0, 3c + 5e2 = 0, a + 15ce − 10d2 = 0, only variables d, e left.
And by assumption h = 0 we are not in case (i). That settles the case ψ = 0.
Now assume ψ 6= 0 and take ψ ∼ x2 . In case (iii) only variables a, b are
left, and we are done. In case (ii) only variables a, d, e are left. In case (i)
only variables c, d, e, f are left. An invariant of degree 14 is (f.(f, ψ2 )5 , ψ 5 )10 ∼
(f.(f, ψ2 )5 , x10 )10 = −2af h2 +2ag 2h+7beh2 −7bf gh−5cdh2 −22cegh+27cf 2h+
25d2 gh − 45def h + 20e3 h. In case (ii) this vanishes. In case (i) this becomes
(up to a constant) 18e3 − 32def + 9cf 2 − cd. Another invariant of degree 14 is
((ψ2 , ψ3 )1 , ψ 4 )8 ∼ ((ψ2 , ψ3 )1 , x8 )8 . In case (ii) this becomes de(26e3 −35d2 −10a)
and we are reduced to three pieces, each with only two variables. In case (i) this
becomes (up to a constant) 70e3 f 4 −120def 5 +27cf 6 +36e5f −60de3 f 2 +6ce2f 3 +
3cdf 4 + 6d2 e3 + 18ce4 − 8d3 ef − 54cde2f + 33cd2f 2 + 3c2ef 2 + cd3 − 3c2de + 2c3 f .
Both polynomials found are irreducible and hence have no common factor, and
we are reduced to a 2-dimensional situation. This proves [4, 14] ≥ 3.
Finally, let us show that [4, 18] ≥ 3. The subcase ψ = 0 was handled already,
so we can assume that ψ 6= 0 and take ψ ∼ x2 . Again only cases (i) and (ii) need
to be considered. Above we already considered the invariant ((ψ1 , ψ2 )1 , ψ 7 )14
of degree 18. In case (ii) this yields d = 0, leaving only the two variables
a, e. In case (i) we find ef 2 + de − cf = 0. Another invariant of degree 18
is (f.((f, ψ2 )5 , ψ2 )2 , ψ 6 )12 . In case (i) this yields 70e3 f 3 − 120def 4 + 27cf 5 −
54e5 + 210de3f − 200d2ef 2 − 15ce2f 2 + 30cdf 3 + 15cde2 − 25cd2 f − c3 = 0. Both
polynomials found are irreducible and hence have no common factor, and we
are reduced to a 2-dimensional situation. This proves [4, 18] ≥ 3.
6.6
n=8
For the octavic there there are nine basic invariants id (2 ≤ d ≤ 10). There is a
hsop with degrees 2, 3, 4, 5, 6, 7. The Poincaré series is
P (t)
= 1 + t2 + t3 + 2t4 + 2t5 + 4t6 + 4t7 + 7t8 + 8t9 +
12t10 + 13t11 + 20t12 + 22t13 + 31t14 + · · · =
7
Y
(1 − td ).
= (1 + t8 + t9 + t10 + t18 )/
d=2
Given a finite sequence (diQ
), the numerator of P (t) corresponding to this
sequence is by definition P (t) (1 − tdi ). If (di ) is a subsequence of the sequence of degrees of a hsop, then the corresponding numerator has nonnegative
coefficients. This rules out, e.g., the following sequences (di ).
2, 2
3, 3
2, 4, 4
2, 5, 5
3, 5, 5
4, 4, 4
5, 5, 5
2, 3, 7, 7
What is wrong with these sequences is that there just aren’t enough invariants of these degrees. More interesting are the cases where there are enough
invariants, but they cannot be chosen algebraically independent.
Proposition 6.11. A sequence of six integers larger than 1 is the sequence of
degrees of a hsop for the octavic if and only if
(i) (‘divisibility’) at least three of them are even, at least two are divisible by
3, at least one has a factor 4, at least one a factor 5, at least one a factor 6,
and at least one a factor 7, and moreover
12
(ii) (‘nonnegativity’) none of the eight sequences in the above table occur as
a subsequence, and moreover
(iii) (‘algebraic independence’) there are no four elements in any of {2, 3, 6},
{2, 4, 5}, {2, 4, 7}, and no five elements in any of {2, 3, 4, 5, 11}, {2, 3, 4, 6, 11},
{2, 3, 4, 7}, {2, 3, 4, 8}, {2, 3, 4, 9}, {2, 3, 5, 6}, {2, 3, 6, 7, 11}.
Proof We have
I2 = hi2 i, I3 = hi3 i, I4 = hi22 , i4 i, I5 = hi2 i3 , i5 i, I6 = hi32 , i2 i4 , i23 , i6 i,
I7 = hi22 i3 , i2 i5 , i3 i4 , i7 i, I8 = hi42 , i22 i4 , i2 i23 , i2 i6 , i3 i5 , i24 , i8 i,
I9 = hi32 i3 , i22 i5 , i2 i3 i4 , i2 i7 , i33 , i3 i6 , i4 i5 , i9 i,
I11 = hi42 i3 , i32 i5 , i22 i3 i4 , i22 i7 , i2 i33 , i2 i3 i6 , i2 i4 i5 , i2 i9 , i23 i5 , i3 i24 , i3 i8 , i4 i7 , i5 i6 i.
We see that V (∪a∈A Ia ) = V ({ib | b ∈ B}) for A and B as in the table below.
A
2,3,6
2,4,5
2,4,7
2,3,4,5,11
B
2,3,6
2,4,5
2,4,7
2,3,4,5
A
2,3,4,6,11
2,3,4,7
2,3,4,8
2,3,4,9
B
2,3,4,6
2,3,4,7
2,3,4,8
2,3,4,9
A
2,3,5,6
2,3,6,7,11
B
2,3,5,6
2,3,6,7
This shows that the given conditions are necessary. For sufficiency, use
induction. The basis of the induction is provided by the 13 hsops constructed in
the next proposition. Given a sequence of six numbers satisfying the conditions,
order the numbers in such a way that the last is divisible by 7 and at least one
of the last two is divisible by 5. All restrictions concern numbers at most 11,
so if we split a number from the sequence into two parts each at least 12,
such that the divisibility conditions remain true for the two resulting sequences,
then by Lemma 6.1 and induction there exists a hsop with the given sequence
as degree sequence. This means that one can reduce the first four numbers
modulo 12, the fifth modulo 60, and the last modulo 420. It remains to check
a 24 × 24 × 24 × 24 × 72 × 432 box, and this is done by a small computer
program.
Proposition 6.12. There are precisely 13 minimal degree sequences of hsops
in case n = 8, namely
2, 3, 4, 5, 6, 7
2, 3, 4, 5, 8, 42
2, 3, 4, 5, 9, 42
2, 3, 4, 5, 10, 42
2, 3, 4, 6, 8, 35
2, 3, 4, 6, 9, 35 2, 3, 5, 6, 10, 28
2, 3, 4, 7, 8, 30 2, 3, 5, 9, 12, 14
2, 3, 4, 7, 9, 30 2, 4, 5, 6, 8, 21
2, 3, 4, 8, 9, 210
2, 3, 5, 6, 9, 28
Proof Minimality is immediately clear, so we only have to show existence.
Apply Dixmier’s criterion. As before we have to show that for all p and each
subsequence d1 , ..., dp of one of these 13 sequences the inequality [d1 , ..., dp ] ≥ p
holds.
We can save some work by observing that Shioda [10] already showed the
existence of a hsop with degree sequence 2, 3, 4, 5, 6, 7. It follows that
[d1 , ..., dp ] ≥ p when (at least) p of the numbers 2, 3, 4, 5, 6, 7 divide some
of the di .
13
For p = 1, nothing remains to check.
For p = 2, there only remains to show [9] ≥ 2, and this follows since there
are two invariants of degree 9 without common factor, for example i3 i6 and i4 i5 .
For p = 3, we have to show [8] ≥ 3, [2, 9] ≥ 3, [5, 9] ≥ 3, [7, 9] ≥ 3, [10] ≥ 3.
For p = 4, we have to show [3, 8] ≥ 4, [5, 8] ≥ 4, [7, 8] ≥ 4, [4, 9] ≥ 4,
[2, 5, 9] ≥ 4, [6, 9] ≥ 4, [2, 7, 9] ≥ 4, [8, 9] ≥ 4, [3, 10] ≥ 4, [4, 10] ≥ 4, [9, 14] ≥ 4.
For p = 5, we have to show [3, 5, 8] ≥ 5, [6, 8] ≥ 5, [3, 7, 8] ≥ 5, [4, 5, 9] ≥ 5,
[4, 6, 9] ≥ 5, [5, 6, 9] ≥ 5, [4, 7, 9] ≥ 5, [8, 9] ≥ 5, [3, 4, 10] ≥ 5, [6, 10] ≥ 5,
[5, 9, 14] ≥ 5.
There are no conditions left to check for p = 6.
Remain 27 conditions to check. Let V [d1 , ..., dp ] denote the variety defined by all invariants of degrees di . Split V [9] into two parts depending on
whether i2 vanishes or not. Where it does not vanish, all invariants of degrees 3, 5, 7 must vanish. Hence [5, 9], [7, 9] ≥ [9] ≥ min([2, 9], [3, 5, 7, 9]).
Split [2, 9] into two parts depending on whether i4 vanishes or not. The first
part has [2, 3, 4, 9] ≥ 3, the second [2, 3, 5, 9] ≥ 3. Hence [9] ≥ 3. Similarly,
[8] = [2, 4, 8] ≥ min([2, 3, 4, 8], [2, 4, 5, 8]) ≥ 3. Finally, [10] = [2, 5, 10] ≥
min([2, 3, 5, 10], [2, 3, 7, 10]) ≥ 3. This settles p = 3.
The same argument shows that [7, 8], [2, 7, 9], [6, 9], [3, 10], [4, 10], [9, 14] ≥ 4
and [5, 9, 14] ≥ 5.
Since adding a single condition diminishes the dimension by at most one,
[3, 8] ≥ 4 follows from [3, 5, 8] ≥ 5. (Given that i2 vanishes since i42 has degree 8,
the condition that all invariants of degree 5 vanish is equivalent to the requirement that i5 vanishes.) Similarly [5, 8] ≥ 4 and [4, 9] ≥ 4 and [2, 5, 9] ≥ 4 follow
from [3, 5, 8] ≥ 5 and [4, 5, 9] ≥ 5. Trivially, [8, 9] ≥ 4 follows from [8, 9] ≥ 5.
This settles p = 4, assuming the inequalities for p = 5.
Remain 10 conditions to check: [3, 5, 8] ≥ 5, [6, 8] ≥ 5, [3, 7, 8] ≥ 5, [4, 5, 9] ≥
5, [4, 6, 9] ≥ 5, [5, 6, 9] ≥ 5, [4, 7, 9] ≥ 5, [8, 9] ≥ 5, [3, 4, 10] ≥ 5, [6, 10] ≥ 5.
Equivalently, for each of the sets A, where A is one of
{2, 3, 4, 5, 8}, {2, 3, 4, 6, 8}, {2, 3, 4, 7, 8}, {2, 3, 4, 5, 9}, {2, 3, 4, 6, 9},
{2, 3, 5, 6, 9}, {2, 3, 4, 7, 9}, {2, 3, 4, 8, 9}, {2, 3, 4, 5, 10}, {2, 3, 5, 6, 10},
we must have dim V ({ia | a ∈ A}) = 1.
For example, we want dim V (i2 , i3 , i4 , i5 , i8 ) = 1. Now i2 , i3 , i4 , i5 form part
of a hsop, so V (i2 , i3 , i4 , i5 ) is irreducible and has dimension 2. Moreover i8
does not vanish identically on V (i2 , i3 , i4 , i5 ) as we shall see, and it follows that
dim V (i2 , i3 , i4 , i5 , i8 ) = 1.
This argument works in all cases except that of V (i2 , i3 , i4 , i8 , i9 ) and shows
that each of the claimed sequences of degrees with the possible exception of 2, 3,
4, 8, 9, 210, is that of a hsop. In particular, e.g. 2, 3, 4, 5, 8, 42 is the sequences
of degrees of a hsop. But now this argument also applies to V (i2 , i3 , i4 , i8 , i9 ):
V (i2 , i3 , i4 , i8 ) is irreducible of dimension 2 and i9 does not vanish identically
on it, and it follows that V (i2 , i3 , i4 , i8 , i9 ) has dimension 1.
It remains to check the ten conditions that say that i8 does not vanish on any
of V (i2 , i3 , i4 , i5 ), V (i2 , i3 , i4 , i6 ), V (i2 , i3 , i4 , i7 ), that i9 does not vanish on any
of V (i2 , i3 , i4 , i5 ), V (i2 , i3 , i4 , i6 ), V (i2 , i3 , i5 , i6 ), V (i2 , i3 , i4 , i7 ), V (i2 , i3 , i4 , i8 ),
and that i10 does not vanish on V (i2 , i3 , i4 , i5 ) or V (i2 , i3 , i5 , i6 ). Using Singular
14
we computed the radical of the ideals (i2 , i3 , i4 , i5 ), (i2 , i3 , i4 , i6 ), (i2 , i3 , i4 , i7 ),
(i2 , i3 , i5 , i6 ) and (i2 , i3 , i4 , i8 ) and checked the required facts.
(This shows that i8 , i9 and i10 do not vanish on the 2-dimensional pieces
mentioned. Note that these invariants do vanish on various 1-dimensional pieces.
For example, i28 ∈ (i2 , i3 , i4 , i6 , i7 ), so that i8 vanishes on V (i2 , i3 , i4 , i6 , i7 ),
and i58 ∈ (i2 , i3 , i4 , i5 , i6 ), and i210 ∈ (i2 , i3 , i4 , i5 , i6 ) and i39 ∈ (i2 , i3 , i4 , i5 , i6 ) ∩
(i2 , i3 , i4 , i6 , i7 ) ∩ (i2 , i3 , i5 , i6 , i7 ).)
References
[1] A. E. Brouwer & M. Popoviciu, The invariants of the binary nonic, J. Symb.
Comp. 45 (2010) 709–720.
[2] A. E. Brouwer & M. Popoviciu, The invariants of the binary decimic, J.
Symb. Comp. 45 (2010) 837–843.
[3] J. Dixmier, Série de Poincaré et systèmes de paramètres pour les invariants
des formes binaires de degré 7, Bull. SMF 110 (1982) 303–318.
[4] J. Dixmier, Quelques résultats et conjectures concernant les séries de
Poincaré des invariants de formes binaires. pp. 127–160 in: Séminaire
d’algèbre Paul Dubreil et Marie-Paule Malliavin 1983–1984, Springer LNM
1146, Berlin 1985.
[5] A. von Gall, Das vollständige Formensystem der binären Form 7ter Ordnung,
Math. Ann. 31 (1888) 318–336.
[6] D. Hilbert, Über die vollen Invariantensysteme, Math. Ann. 42 (1893) 313–
373.
[7] D. Mumford, J. Fogarty, and F. Kirwan, Geometric invariant theory. 3rd
enl. ed., Springer, 1993.
[8] P. J. Olver, Classical Invariant Theory, Cambridge, 1999.
[9] I. Schur, Vorlesungen über Invariantentheorie, Springer, 1968.
[10] T. Shioda, On the graded ring of invariants of binary octavics, Amer. J.
Math. 89 (1967) 1022–1046.
15
| 0math.AC
|
1
Optimal Slotted ALOHA under Delivery Deadline
Constraint for Multiple-Packet Reception
arXiv:1706.00350v2 [cs.IT] 25 Jul 2017
Yijin Zhang, Yuan-Hsun Lo, Feng Shu, and Jun Li
Abstract—This paper considers the slotted ALOHA protocol in a communication channel shared by N users. It is assumed that the
channel has the multiple-packet reception (MPR) capability that allows the correct reception of up to M (1 ≤ M < N ) time-overlapping
packets. To evaluate the reliability in the scenario that a packet needs to be transmitted within a strict delivery deadline D (D ≥ 1) (in
unit of slot) since its arrival at the head of queue, we consider the successful delivery probability (SDP) of a packet as performance
metric of interest. We derive the optimal transmission probability that maximizes the SDP for any 1 ≤ M < N and any D ≥ 1, and
show it can be computed by a fixed-point iteration. In particular, the case for D = 1 (i.e., throughput maximization) is first completely
addressed in this paper. Based on these theoretical results, for real-life scenarios where N may be unknown and changing, we develop
a distributed algorithm that enables each user to tune its transmission probability at runtime according to the estimate of N . Simulation
results show that the proposed algorithm is effective in dynamic scenarios, with near-optimal performance.
Index Terms—Slotted ALOHA, multiple-packet reception, optimal transmission probability, successful delivery probability
✦
1
I NTRODUCTION
1.1 Motivation
Since Abramson’s seminal work [1] in 1970, ALOHA-type
protocols have been widely used for initial terminal access
or short packet transmissions in a variety of distributed
wireless communication systems due to their simplicity.
There were extensive studies on the slotted ALOHA under
the traditional model of a single-packet reception (SPR) channel: a packet can be correctly received if there is no other
packet transmission during its transmission. However, the
SPR has become restrictive due to the advent of multiplepacket reception (MPR) techniques that allow the correct
reception of time-overlapping packets. Hence, there is a
natural interest in gaining a clear insight into the fundamental impact of MPR on the behavior of the slotted ALOHA
protocol.
Differently from previous studies that dealt with the
stability, throughput or delay issue of slotted ALOHA under
MPR [2]–[10], we in this paper concentrate on achieving
maximum reliability in the scenario that a packet needs
to be transmitted within a strict delivery deadline D (in
unit of slot) since its arrival at the head of queue. Such a
scenario can be safety message dissemination in vehicular
networks [11] or machine to-machine communications in
Internet of Things [12]. Some recent work on a similar issue
can be found in [13]–[17] for an SPR channel, and [18]–[20]
for a multichannel system.
•
Y. Zhang, F. Shu and J. Li are with the School of Electronic and Optical Engineering, Nanjing University of Science and Technology, Nanjing, China,
and also with National Mobile Communications Research Laboratory,
Southeast University, Nanjing, China. E-mails: [email protected];
{shufeng, jun.li}@njust.edu.cn.
•
Y.-H. Lo is with the School of Mathematical Sciences, Xiamen University,
Xiamen, China. E-mail: [email protected].
1.2 Contribution
Along the lines of [6]–[10], this paper considers a specific
MPR channel, namely the M -user MPR channel, in which
up to M time-overlapping packets can be received correctly.
Our key contributions are summarized as follows:
•
•
•
We derive the optimal transmission probability for
the reliability maximization when N users contend
for the channel access. Our work can be seen as a
generalization of the work in [13] that only focused
on the SPR channel. Moreover, we show that the
optimal transmission probability can be computed
by a fixed-point iteration.
As explained in Section 3, the saturation throughput maximization of finite-user slotted ALOHA [8],
[10] can be studied as a special case D = 1 of
reliability maximization that we investigate here. It
should be pointed out that Bae et al. [10] obtained
the optimal transmission probability for saturation
throughput maximization necessarily relying on an
E[X 2 ]
unproved technical condition ddτ E[X] < N − 1 (X
is the number of users involved in each successful
transmission and τ is the transmission probability).
In this paper, we present analysis to prove that this
technical condition always holds for an arbitrary
1 ≤ M < N . Hence, the issue of saturation throughput maximization under an M -user MPR channel is
first completely addressed in this paper.
Clearly, deriving the optimal transmission probability requires priori knowledge of N . To achieve
maximum reliability in real-life scenarios where N is
unknown and changing over time, built on the theoretical results derived above, we propose a tuning
algorithm that allows each user to estimate N without requiring any access parameter input, and adjust
its transmission probability accordingly at runtime.
Through an extensive performance study we show
2
that the tuning algorithm is effective in a variety of
dynamic network configurations considered in the
paper, with near-optimal performance.
1.3 Related work
The first attempt to study slotted ALOHA under MPR
was made by Ghez et al. [2], [3], in which they proposed
the symmetric MPR channel model and analyzed stability
properties under an infinite-user assumption. Naware et
al. [4] extended the stability study to finite-user systems
without posing any limitation on the MPR model, and in
addition investigated the average delay in capture channels.
Luo et al. [5] further established the throughput and stability
regions for finite population over a standard MPR channel
in which simultaneous packet transmissions are not helpful
for the reception of any particular group of packets.
After the aforementioned studies for various generalized MPR channels, the throughput performance of slotted ALOHA over an M -user MPR channel has received
much attention recently. Gau [6], [7] derived the saturation and non-saturation throughput for finite-user cases.
To demonstrate the capacity-enhancement, Zhang et al.
in [8] proved that the maximum achievable throughput
increases superlinearly with M for both finite-user case
with saturation traffic and infinite-user case with random
traffic, and in [9] further showed that superlinear scaling
also holds under bounded delay-moment requirements. Following [8], to fully utilize the M -user MPR channel, Bae et
al. [10] derived the optimal transmission probability that
maximizes the saturation throughput in the finite-user case
under some unproved technical conditions. To the best of
our knowledge, little work has been done to investigate the
reliability issue over an M -user MPR channel. Our paper
here is an attempt along this direction.
Under unknown and time-varying operating conditions,
developing an estimation algorithm to acquire knowledge
of the number of users N is of significant importance. Many
approaches have been proposed for the SPR channel. The
method in [21] estimates N based on the channel state in the
previous slot, and the schemes in [22], [23] estimate N with
statistics of consecutive idle and collision slots. However,
all of them require that N follows a Poisson distribution.
To remove the assumption on the distribution of N , the
algorithm in [24] applies an ARMA filter to estimate N
relying on the measured collision probability, the algorithm
in [25] estimates N by the number of idle slots between two
successful transmissions, and the algorithm in [26] estimates
N from the knowledge of number of consecutive idle slots.
The extension of the estimation algorithm to the MPR case
is rarely reported. We are aware of only one previously
proposed algorithm in [10] that estimates N according to
the collected information on the number of users involved
in each successful transmission. However, we find it is
ineffective in some dynamic scenarios, which will be shown
in Section 5.
The remainder of this paper is organized as follows.
In Section 2, we describe the considered system model. In
Section 3, we derive the optimal transmission probability
that maximizes the reliability for any 1 ≤ M < N and
any D ≥ 1, and show it can be computed by a fixedpoint iteration. In section 4, we propose a tuning algorithm
by which each user can achieve a reliability level close to
the theoretical limit at runtime. In Section 5, simulation
results verify the accuracy of our analytical results and
the effectiveness of the proposed tuning algorithm. Finally,
Section 6 concludes this paper.
2
S YSTEM MODEL
As adopted in [8], [10], [13], we develop our analytical
model based on the following assumptions:
(i) There are N (N ≥ 2) users with saturated traffic in the
network, and all of them are within the transmission
range of each other.
(ii) The system is limited by user interference and the channel has an M -MPR capability, which means a packet
can be correctly decoded by the receiver if at most M −1
other packet transmissions overlap with it at any time,
and is unrecoverable otherwise. To avoid some trivial
cases, we assume 1 ≤ M < N . Specially, M = 1
corresponds to the SPR channel.
(iii) The channel time is divided into time slots of an equal
length, and every packet exactly occupies the duration
of one time slot.
(iv) Each user knows the slot boundaries, and attempts to
transmit a packet with a common transmission probability τ at the beginning of a time slot, 0 ≤ τ ≤ 1.
(v) Every packet is neither acknowledged nor retransmitted, since that an acknowledgement mechanism would
incur extra overhead, waiting time and energy cost for
short packets, and meanwhile, for some periodic traffic,
the content can simply be replaced, and hence there is
no need to retransmit an outdated packet.
(vi) Every packet should be delivered within a strict delivery deadline D (D ≥ 1) (in unit of slot), which
is defined as the duration from the moment of its
arrival at the head of the queue to the completion of
its transmission.
3
O PTIMAL T RANSMISSION P ROBABILITY
Given any real number τ ∈ [0, 1] and integer D ≥ 1,
let PD (τ ), called successful delivery probability (SDP), be the
probability that a packet will be successfully received within
the delivery deadline D under the common transmission
probability τ . Consider a tagged user. Let Y denote the number of packets transmitted by the other N − 1 users in a time
slot. It is easy to see Y follows a binomial distribution with
parameters N − 1 and τ , and then for i = 0, 1, . . . , N − 1,
we have
P(Y = i) =
!
N −1 i
τ (1 − τ )N −1−i .
i
(1)
3
Furthermore, the value PD (τ ) can be obtained as:
PD (τ ) =
D
X
By adopting the notation f1 and f2 , the derivative of
PD (τ ) with respect to τ can be written as
τ (1 − τ )k−1 P(Y ≤ M − 1)
d
PD (τ ) = D(1 − τ )D−1 f1 (τ )
dτ
f (τ )
(N − 1)f1 (τ )
2
+
−
1 − (1 − τ )D
τ (1 − τ )
1−τ
N − 1
= (N + D − 1)(1 − τ )D−1 −
f1 (τ )
1−τ
1 − (1 − τ )D
f2 (τ )
(5)
+
τ (1 − τ )
1
=
1 − (1 − τ )D f2 (τ )
τ (1 − τ )
− N − 1 − (N + D − 1)(1 − τ )D τ f1 (τ ) . (6)
k=1
D
X
!
M−1
X N −1
k−1
τ i (1 − τ )N −1−i
τ (1 − τ )
=
i
i=0
k=1
!
M−1
X N −1
D
τ i (1 − τ )N −1−i .
= 1 − (1 − τ )
i
i=0
(2)
In this section, we aim to obtain the optimal transmission
probability for maximizing PD (τ ).
max
For a given integer D ≥ 1, let PD
denote the maximum SDP going through all possible τ ∈ [0, 1], that is,
max
:= max PD (τ ).
PD
τ ∈[0,1]
Then, define the optimal transmission probability, denoted by
opt
τD
, to be the transmission probability such that the SDP
max
achieves PD
, i.e.,
opt
:= arg max PD (τ ).
τD
τ ∈[0,1]
opt
τD
Note that
may not be unique by definition.
Remark 1: As P1 (τ ) refers to the individual saturation
throughput defined as the time average of the number of
packets successfully transmitted by a user provided that all
opt
users have saturated traffic, τ1 is indeed the optimal transmission probability maximizing the saturation throughput
under MPR, which has been investigated in [10].
opt
Remark 2: When M = 1, τD is the optimal transmission
probability maximizing the SDP within the delivery deadline D under SPR, which has been derived in [13].
It is easy to see from (2) that, when D is fixed, PD (τ )
is a continuous function of τ on the closed interval [0, 1].
opt
Hence, by The Extreme Value Theorem, τD exists. In the
remainder of this subsection, we shall show the uniqueness
opt
of τD , and present how to obtain it.
Define the following semi open interval
h
N −1
1
I := 1 − (
)D , 1 .
N −1+D
opt
We first provide some properties of τD
.
(i)
(ii)
is a solution of
must lie in I .
d
dτ PD (τ )
= 0, and
Proof. We prove these two statements by investigating the
monotonicity of PD (τ ). Define
!
M−1
X N −1
τ i (1 − τ )N −1−i
(3)
f1 (τ ) :=
i
i=0
and
f2 (τ ) :=
M−1
X
i=0
!
N −1 i
i
τ (1 − τ )N −1−i .
i
Let
H1 (τ ) :=
f2 (τ )
f1 (τ )
and
H2 (τ ) := τ N + D − 1 −
D
.
D
1 − (1 − τ )
Following the proof of Lemma 1, in (6), τ ∗ is a solution of
equation ddτ PD (τ ) = 0 if and only if it is a solution of the
following equation:
H1 (τ ) − H2 (τ ) = 0.
(7)
In what follows, we will show that the equation (7) has a
unique solution in the interval τ ∈ (0, 1) by investigating
the monotonicity of H1 (τ ) and H2 (τ ) separately.
Lemma 1. For any integers D ≥ 1 and 1 ≤ M < N ,
opt
τD
opt
τD
It is easy to see that ddτ PD (τ ) is continuous on the interval
(0, 1).
Since that PD (τ ) > PD (0) = PD (1) = 0 if τ ∈ (0, 1), we
know the continuous function PD (τ ) has a local maximum
opt
at τD , which lies in (0, 1). As ddτ PD (τ ) always exists on
opt
the interval τ ∈ (0, 1), by the Fermat’s Theorem, τD is a
d
solution of dτ PD (τ ) = 0.
Furthermore, we have from (5) that ddτ PD (τ ) > 0 for τ ∈
(0, 1) \ I , and from (6) that ddτ PD (τ ) < 0 as τ → 1− . By The
Intermediate Value Theorem, the solutions of ddτ PD (τ ) = 0
opt
must be in I , i.e., τD must lie in I .
(4)
Obviously, f1 (τ ) > 0 and f2 (τ ) ≥ 0 for 0 < τ < 1, 1 ≤
M < N and D ≥ 1.
Lemma 2. For τ ∈ (0, 1), we have ddτ H1 (τ ) = 0 if M = 1, and
otherwise
d
H1 (τ ) < N − 1.
0<
dτ
Proof. The case for M = 1 obviously holds as H1 (τ ) = 0. In
the following, we only consider 2 ≤ M < N .
We first show that ddτ H1 (τ ) > 0 by a known result
in [10]. Let
T (τ ) :=
PM
2
i=1 i
PM
i=1 i
N i
N −i
i τ (1 − τ )
.
N
i
N −i
i τ (1 − τ )
(8)
4
By letting j = i − 1, after some algebraic manipulations, we
have
PM 2
(i − i) Ni τ i (1 − τ )N −i
T (τ ) − 1 = i=1
PM
N
i
N −i
i=1 i i τ (1 − τ )
PM−1
N j+1
(1 − τ )N −1−j
j=0 j(j + 1) j+1 τ
= PM−1
N j+1
(1 − τ )N −1−j
j=0 (j + 1) j+1 τ
PM−1 N −1 j
τ N j=0 j j τ (1 − τ )N −1−j
=
P
N −1 j
τ (1 − τ )N −1−j
τ N M−1
j=0
j
= H1 (τ ).
(9)
Since it has been proven in [10] that ddτ T (τ ) ≥ 0 for τ ∈
(0, 1) and M > 1, we have ddτ H1 (τ ) = ddτ T (τ ) > 0 by (9).
Now, we will show that ddτ H1 (τ ) < N − 1. By the
binomial theorem, H1 (τ ) can be rewritten as
P −1 N −1 i
τ (1 − τ )N −1−i
(N − 1)τ − N
i=M i
i
H1 (τ ) =
PN −1 N −1 i
1 − i=M i τ (1 − τ )N −1−i
= (N − 1)τ
N −1 i
PN −1
τ (1 − τ )N −1−i
i=M i − (N − 1)τ
i
−
PN −1 N −1 i
1 − i=M i τ (1 − τ )N −1−i
N −1 M
τ (1 − τ )N −M
(N − M ) M−1
(∗)
= (N − 1)τ − PM−1 N −1
,
τ i (1 − τ )N −1−i
i=0
i
!
N −1
= (N − 1)τ − (N − M )
R(τ ),
(10)
M −1
where
τ M (1 − τ )N −M
.
R(τ ) := PM−1 N −1
τ i (1 − τ )N −1−i
i=0
i
The proof of (∗) is as follows.
For m = 2, 3, . . . , N − 1, we have
!
N −1
(N − m)τ m (1 − τ )N −m
m−1
!
N − 1 m−1
τ
(1 − τ )N −m
+ m − 1 − (N − 1)τ
m−1
!
N −1
=
(N − m)τ m (1 − τ )N −m
m−1
!
N − 1 m−1
− (N − m)
τ
(1 − τ )N −m
m−1
!
N − 1 m−1
+ (N − 1)
τ
(1 − τ )N −m+1
m−1
!
N −1
=−
(N − m)τ m−1 (1 − τ )N −m+1
m−1
!
N − 1 m−1
+ (N − 1)
τ
(1 − τ )N −m+1
m−1
!
N − 1 m−1
=(m − 1)
τ
(1 − τ )N −m+1
m−1
!
N −1
=
(N − m + 1)τ m−1 (1 − τ )N −m+1 .
m−2
Then by recursively using the above equation, we have
!
N
−1
X
N −1 i
τ (1 − τ )N −1−i
i − (N − 1)τ
i
i=M
!
N − 1 N −1
=
τ
(1 − τ )
N −2
!
N
−2
X
N −1 i
τ (1 − τ )N −1−i
i − (N − 1)τ
+
i
i=M
!
N −1
=
(N − M )τ M (1 − τ )N −M .
M −1
To prove ddτ H1 (τ ) < N − 1, by (10), it suffices to show
that ddτ R(τ ) > 0. Let
By plugging x =
xM−1
Q(x) := PM−1 N −1 .
xi
i=0
i
τ
1−τ ,
we have
R(τ ) = τ Q(x).
(11)
Note that as τ increases from 0 to 1, x increases from 0 to
∞, and hence Q(x) > 0. By taking the derivative of Q(x)
with respect to x, we have, for x > 0,
PM−1
N −1 M+i−2
x
d
i=0 (M − 1 − i)
i
> 0. (12)
Q(x) =
2
P
dx
M−1 N −1 i
x
i=0
i
Then,
d
d
d
τ
·
R(τ ) =
Q(x) > 0.
τ Q(x) = Q(x) +
dτ
dτ
(1 − τ )2 dx
(13)
Hence the result follows.
Lemma 3. For τ ∈ (0, 1), we have
d
H2 (τ ) > N − 1.
dτ
Proof. Taking the derivative of H2 (τ ) with respect to τ
derives that
d
1 − (1 − τ )D − Dτ (1 − τ )D−1
H2 (τ ) = N + D − 1 − D
2
dτ
1 − (1 − τ )D
1 − (1 − τ )D − Dτ (1 − τ )D−1
=N −1+D 1−
2
1 − (1 − τ )D
So we have
d
H2 (τ ) > N − 1
dτ
2
⇔ 1 − (1 − τ )D > 1 − (1 − τ )D − Dτ (1 − τ )D−1
2
⇔ 1 − (1 − τ )D + (1 − τ )D + Dτ (1 − τ )D−1 − 1 > 0
⇔(1 − τ )2D − (1 − τ )D + Dτ (1 − τ )D−1 > 0
⇔(1 − τ )D+1 + (D + 1)τ − 1 > 0
Let G(τ ) := (1 − τ )D+1 + (D + 1)τ − 1. We have
d
G(τ ) = −(D + 1)(1 − τ )D + D + 1
dτ
= (D + 1) 1 − (1 − τ )D ,
(14)
5
For the case M = 1, which implies H1 (τ ) = 0, it is easy
to see from (7) that
N − 1 D1
opt
.
τD
=1−
N −1+D
For the case 1 < M < N , as H1 (τ ) > H2 (τ ) = 0 when
1
−1
D
τ = 1 − ( NN
−1+D ) , we by Lemma 1 know that
N − 1 D1
opt
,1 .
τD
∈ 1−
N −1+D
Define a fixed-point iteration:
9
H1(τ)
8
H2(τ)
M=9
7
M=8
6
5
4
3
2
M=3
D=50
1
D=5
0
−1
M=1
0
0.2
0.4
0.6
The transmission probability, τ
0.8
H1 (xn ) + 1
H2 (xn ) + 1
for n = 0, 1, 2, . . .. With the following theorem we guarantee
opt
that τD for M > 1 can be obtained by this fixed-point
iteration.
xn+1 = xn ·
M=2
D=1
1
1
Fig. 1. H1 (τ ) for the varying M and τ , H2 (τ ) for the varying D and τ
when N = 10.
which is larger than 0 for τ ∈ (0, 1). Therefore, G(τ ) is
strictly increasing for τ ∈ (0, 1), which implies that
G(τ ) > lim+ G(τ ) = 0.
τ →0
Hence we complete the proof by (14).
To illustrate that 0 ≤ ddτ H1 (τ ) < N − 1 < ddτ H2 (τ ) on
the interval τ ∈ (0, 1) for any D ≥ 1 and 1 ≤ M < N
obtained by Lemma 1 and Lemma 2, a numerical example
is presented in Fig. 1, which plots H1 (τ ) and H2 (τ ) for the
varying τ , M and D when N = 10.
opt
Now, we are ready to derive the uniqueness of τD .
Theorem 4. For any integers D ≥ 1 and 1 ≤ M < N , the
equation (7) has a unique solution on τ in the interval 0 < τ < 1,
opt
denoted by τ ∗ , and τD = τ ∗ .
Proof. Suppose there are two distinct solutions to the equation (7) in (0, 1). By The Mean Value Theorem, there exists
a solution of ddτ H1 (τ ) = ddτ H2 (τ ). However, by Lemma 2
and Lemma 3, we have
d
d
H1 (τ ) < N − 1 <
H2 (τ ),
dτ
dτ
∀τ ∈ (0, 1).
This implies a contradiction to ddτ H1 (τ ) = ddτ H2 (τ ). Hence
we conclude that the equation (7) has a unique solution on
τ in the interval 0 < τ < 1, which by Lemma 1 promises the
opt
opt
uniqueness of τD , and yields τD = τ ∗ .
Remark 3: In the context of saturation throughput maxopt
imization, Bae et al. in [10] derived the τ1 under the
d
assumption dτ
T (τ ) < N in the interval τ ∈ (0, 1). They
claimed that they proved ddτ T (τ ) < N for M = 1, 2, 3,
but could not prove it for any arbitrary M due to extremely complex algebraic manipulations. Here, the proof
in Lemma 2 has addressed this unsolved question, as
d
T (τ ) = ddτ H1 (τ ) < N − 1. In other words, the issue
dτ
of saturation throughput maximization under an M -user
MPR channel is first completely addressed in this paper.
Moreover, we in Theorem 4 obtained the existence and
opt
uniqueness of τD for any D ≥ 1 and any 1 ≤ M < N
without any assumption.
−1
D
Theorem 5. For any initial guess x0 ∈ (1 − ( NN
−1+D ) , 1),
opt
the sequence x0 , x1 , x2 , . . . converges to the fixed point τD for
1 < M < N.
Proof. Define
g(x) :=
x(H1 (x) + 1)
H2 (x) + 1
on the domain (0, 1). By (7) and Theorem 4, the equation
opt
g(x) = x has a unique solution at x = τD
in (0, 1).
opt
Therefore, the case that x0 = τD obtains the fixed point
opt
opt
opt
opt
τD
since g(τD ) = τD . So we consider x0 6= τD in what
follows.
opt
Since g(x) = x has a unique solution at τD ∈ (0, 1), to
opt
∞
prove the sequence {xn }0 converges to τD , it suffices to
show that
(
1
opt
opt
−1
D
x < g(x) < τD
, for x ∈ (1 − ( NN
−1+D ) , τD ); (15)
opt
opt
τD < g(x) < x, for x ∈ (τD , 1).
As proved in Appendix that
d
g(x) > 0 for x ∈ (0, 1),
dx
g(x) is increasing on (0, 1). This implies that
(
1
opt
opt
−1
D
g(x) < τD
, for x ∈ (1 − ( NN
−1+D ) , τD );
opt
opt
τD
< g(x), for x ∈ (τD
, 1).
(16)
(17)
Since, in the interval (0, 1), H1 (x) = H2 (x) only when x =
opt
τD
and ddx H1 (x) < ddx H2 (x) by Lemma 2 and Lemma 3,
opt
we have 0 < H1 (x) < H2 (x) for x ∈ (τD , 1). Then,
g(x) = x
H1 (x) + 1
< x,
H2 (x) + 1
opt
for x ∈ (τD , 1).
(18)
Following the same argument, we have H1 (x) > H2 (x) for
1
opt
−1
D
x ∈ (1 − ( NN
−1+D ) , τD ). By the result in Lemma 3 that
d
H (x) > 0 and the L’Hospital’s Rule that
dx 2
lim H2 (x) = −1,
x→0+
we further have H1 (x) > H2 (x) > −1 for x ∈ (1 −
1
opt
−1
D
( NN
−1+D ) , τD ). Then,
H1 (x) + 1
> x,
H2 (x) + 1
N −1
1
opt
for x ∈ (1 − (
) D , τD
).
N −1+D
g(x) =x
(19)
6
P(Y =i1 )P(Y =i2 −1)
P(Y =i2 )P(Y =i1 −1)
Let µn be the measure of
during the
nth update interval, and its value is calculated by:
Therefore, (15) can be derived combining (17)–(19), and thus
the result follows.
4
µn =
RUNTIME O PTIMIZATION
In the previous section, the optimal transmission probability
has been derived, however, the analysis requires knowing in
advance the number of users, N , which may be unavailable
in some practical scenarios of distributed networks. To cope
with this restriction, we in this section develop a distributed
algorithm to estimate N for M > 1, by which each user can
tune the transmission probability to obtain an SDP close to
the theoretical limit.
Consider a tagged user. Recall the variable Y that denotes the number of packets transmitted by the other N − 1
users in a time slot. By the distribution of Y in (1), for
1 ≤ i1 < i2 ≤ M , we obtain the following equation:
P(Y = i1 )P(Y = i2 − 1)
P(Y = i2 )P(Y = i1 − 1)
N −1 N −1 i1 +i2 −1
(1 − τ )2N −1−i1 −i2
i −1 τ
1
= Ni−1
N2 −1
i1 +i2 −1 (1 − τ )2N −1−i1 −i2
i2
i1 −1 τ
N −1 N −1
i2 (N − i1 )
i −1
1
.
= Ni−1
N2 −1 =
i1 (N − i2 )
i
i −1
2
(20)
1
It then directly follows that:
N=
i2 (i2 − i1 )
P(Y =i1 )P(Y =i2 −1)
i1 P(Y =i2 )P(Y =i1 −1)
− i2
+ i2 .
(21)
P(Y =i )P(Y =i −1)
As P(Y =i21 )P(Y =i12 −1) is locally measurable if a user is
equipped with an array with at least i2 antennas [27], [28],
we find that (21) provides a linear function for the tagged
user to estimate N without priori knowledge of other access
parameters.
We assume that the tagged user knows there are at most
Nmax users in the network, but the actual number of users is
changing and unknown. An update interval is defined as a
block of consecutive of L time slots. We require the tagged
user to update the transmission probability at the beginning
of the n + 1th update interval, according to the following
two necessary estimates of the network status.
P(Y =i )P(Y =i −1)
(i) µn : the estimated value of P(Y =i21 )P(Y =i12 −1) at the end
of the nth update interval;
(ii) Nn : the estimated value of N at the end of the nth
update interval.
The tagged user initially guesses that there are N0 =
0 −i1 )
Nmax users, and sets µ0 = ii21 (N
(N0 −i2 ) . The transmission
probability for the first update interval is obtained with
given N0 , by Theorem 4 and a fixed-point iteration. During
the nth interval for n = 1, 2, . . ., we describe the proposed
algorithm as follows:
(i) Update µn at the end of the nth update interval: To evaluate
P(Y =i1 )P(Y =i2 −1)
at runtime, the tagged user is required
P(Y =i2 )P(Y =i1 −1)
to record Ai,n , the number of the slots during the nth
update interval in which i users are simultaneously
transmitting and the tagged user is not transmitting,
for i = i1 − 1, i1 , i2 − 1 and i2 .
Ai1 ,n · Ai2 −1,n
Ai2 ,n · Ai1 −1,n
To avoid sharp changes in the estimated value, the
tagged user further applies an Exponential Moving
Average filter as follows:
µn = δ · µn−1 + (1 − δ) · µn
where δ ∈ [0, 1] is a memory factor.
(ii) Update Nn at the end of the nth update interval: By (21),
the tagged user obtains
E
D i (i − i )
2 2
1
+ i2
Nn =
i 1 µn − i 2
where hxi is the integer closest to x.
(iii) Update the transmission probability at the beginning of
the n + 1th update interval: Update the transmission
probability with given Nn , by Theorem 4 and a fixedpoint iteration.
To avoid harmful measure of µn due to occasional fluctuations of the network status, (i) if µn is found to be smaller
than µ0 , a contradiction to the fact that (20) must be equal
to or larger than µ0 as M < N ≤ N0 , the tagged user
i (M+1−i )
sets µn = µ0 ; (ii) if µn is found larger than i12 (M+1−i21 ) ,
a contradiction to (20) as N > M , the tagged user sets
(M+1−i1 )
; and (iii) if µn has an invalid value as
µn = ii12 (M+1−i
2)
Ai2 ,n · Ai1 −1,n = 0, the tagged user sets µn = µn−1 .
Remark 4: It should be noted that, although the proposed algorithm for estimating N is devised for the slotted ALOHA in saturation assumption, it can apply also
to unsaturated traffic, provided that the estimation target
becomes the average number of competing users rather than
the total number of users; and can also apply to τ -persistent
CSMA, provided that the length of a slot may vary due to
different channel status: idle, collision or success.
5
S IMULATION R ESULTS
In this section, to demonstrate the accuracy of the derived
analytical results and the effectiveness of our proposed
tuning algorithm, we present simulation results with respect
to different network parameters obtained from Monte Carlo
simulation developed in Matlab. In the following, we will
analyze the performance under stationary scenarios when
N is known, and then under dynamic scenarios when N is
unknown. Each simulation point in stationary conditions
represents the average value over 10 independent runs,
each of which consists of 106 slots. The results in dynamic
conditions are from a single representative simulation run
for 500 updated intervals.
5.1 Stationary Scenarios when N is known
Fig. 2 and Fig. 3 show the optimal transmission probability
and the maximum SDP as a function of N for different
values of M and D, respectively. We see a good agreement between analytical and simulation results in all the
scenarios. Notice that the results for D = 1 can be seen as
the throughput maximization issue investigated in [10]. In
7
0.18
0.4
M=5, analysis
M=5, simulation
0.12
D=10
0.1
D=20
0.08
0.6
D=1
Optimal transmission probability
D=1
D=5
0.06
M=8, analysis
M=8, simulation
0.35
Optimal transmission probability
Optimal transmission probability
0.14
0.7
M=2, analysis
M=2, simulation
0.16
0.3
D=5
0.25
D=10
0.2
D=20
0.15
0.5
D=1
D=5
0.4 D=10
0.3
D=20
0.2
D=50
D=50
D=50
0.1
0.04
0.02
10
15
20
25
30
The number of users, N
35
40
0.05
10
0.1
15
(a) M = 2
20
25
30
The number of users, N
35
0
10
40
15
(b) M = 5
20
25
30
The number of users, N
35
40
(c) M = 8
Fig. 2. The optimal transmission probability as a function of N for different values of M and D .
0.9
1
D=50
0.7
0.6
D=20
0.5
0.4
D=10
0.3
D=5
0.2
0.1
0
10
D=20
0.7
D=10
0.6
0.5
0.4
M=5, analysis
M=5, simulation
20
25
30
The number of users, N
35
40
D=5
0.3
0.2
D=1
0.1
D=1
15
0.8
Maximum successful delivery probability
0.9
Maximum successful delivery probability
Maximum successful delivery probability
0.8
1
D=50
D=50
M=2, analysis
M=2, simulation
0
10
(a) M = 2
15
20
25
30
The number of users, N
(b) M = 5
0.9
D=20
0.8
D=10
0.7
0.6
D=5
0.5
0.4
M=8, analysis
M=8, simulation
0.3
0.2
35
40
0.1
10
D=1
15
20
25
30
The number of users, N
35
40
(c) M = 8
Fig. 3. The maximum SDP as a function of N for different values of M and D .
Fig. 2, as expected, we observe that the optimal transmission probability becomes smaller when N increases. This
is because that more users attempt to access the channel
and the contention level becomes severer. We also see the
optimal transmission probability becomes larger when M
increases or D decreases. The reason is that a user needs
to be more aggressive in accessing the channel if more
concurrent packets can be successfully received or the user
wants to successfully send out a packet within a shorter
delivery deadline. In Fig. 3, as expected, the curves show
that a smaller N , a larger M or a larger D leads to a larger
value of the maximum SDP. In particular, we note that the
optimal transmission probability for D > 1 is smaller than
that for D = 1 for given N and M . This phenomenon
indicates that the throughput maximization degrades the
SDP for D > 1, and in turn the maximization of SDP for
D > 1 degrades the throughput performance.
5.2 Dynamic Scenarios when N is unknown
By setting i2 = M and i1 = 2 in (21), we analyze the
performance of the proposed tuning algorithm when the
number of users sharply changes. In detail, 20 users are
always active from 1st to 500th updated interval, and 20
new users are active only from the 101st updated interval
to the 400th updated interval. Each user knows there are
at most Nmax = 100 users, and initially guesses there are
N0 = 100 users.
5.2.1 The cases with D = 1
We start by discussing the performance for D = 1. Fig. 4
and Fig. 5 show the estimated number of users and SDP for
M = 5 with different L and δ , respectively. To characterize
individual behavior and improve the readability, we only
present the real trajectories of two representative users in
the same run.
We observe from Fig. 4 that, in all the cases, when the
actual number of users is changed to 20, the estimate rapidly
adapts to changes and keeps a high level of accuracy; when
the actual number of users is changed to 40, the estimate is
still able to capture changes within a few update intervals,
but the slop increases, i.e, the accuracy of the estimation
degrades. This phenomenon is because that as the actual
number of users increases, the fluctuation in measuring µn
is amplified in estimation by using (21).
As expected, we observe from Fig. 5 that the users
achieve SDPs close to the theoretical limit at runtime by tuning the transmission probability according to the estimated
number of users as shown in Fig. 4. Moreover, even when
the oscillation of estimation is high, we find that the SDP still
keeps relatively stable. This phenomenon is due to the fact
that the SDP is less sensitive to the change in transmission
probability as the number of users increases.
To evaluate the fairness of the proposed algorithm, we
further study mean and variance of SDP among all active
users at different stages. As shown in Table 1, we find
8
100
100
The estimated number of users
90
80
70
60
50
40
30
20
10
One of initially active users
One of new users
One of initially active users [2]
One of new users [2]
Actual value
90
The estimated number of users
One of initially active users
One of new users
One of initially active users [2]
One of new users [2]
Actual value
80
70
60
50
40
30
20
0
50
100
150
200
250
300
Interval index
350
400
450
10
500
0
50
100
(a) L = 50000, δ = 0.7
150
200
250
300
Interval index
350
400
450
500
400
450
500
(b) L = 20000, δ = 0.9
Fig. 4. The estimated number of users for M = 5 and D = 1 when the number of users sharply changes.
0.16
0.16
One of initially active users
One of new users
One of initially active users [2]
One of new users [2]
Theoretical maximum value
0.12
0.1
0.08
0.06
0.04
0.02
0
One of initially active users
One of new users
One of initially active users [2]
One of new users [2]
Theoretical maximum value
0.14
Successful delivery probability
Successful delivery probability
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
50
100
150
200
250
300
Interval index
350
400
450
500
0
0
50
(a) L = 50000, δ = 0.7
100
150
200
250
300
Interval index
350
(b) L = 20000, δ = 0.9
Fig. 5. Individual SDP for M = 5 and D = 1 when the number of users sharply changes.
the mean value is very close to the theoretical maximum
value for each stage with a tolerance of at most 5%, and
the variance value is also very small. This result indicates
that the proposed algorithm can enable every active user
to achieve near-optimal performance in dynamic scenarios.
Notice that the transient period is a dominant factor to
degrade the performance.
become active at different time instants, they may adopt
different transmission probabilities, and then have biased
estimate of N . Whereas, our algorithm does not require any
access parameter input, and hence would not lead to biased
estimate in such a scenario.
As the case with D = 1 can be viewed as runtime
optimization of throughput, to show the proposed algorithm is more effective in a variety of dynamic scenarios,
we consider the approach in [10] for comparison purposes.
The method therein recursively updates the transmission
probability with an estimated N , and then uses the measure
E[X 2 ]
of E[X] (X is the number of users involved in each successful transmission slot) to estimate new N necessarily relying
on the adopted transmission probability. As shown in Fig. 4
and Fig. 5, one sees their estimate keeps a very high level of
accuracy when 20 new users are inactive, but is not able to
keep track of the change well when 20 new users are active.
This phenomenon is because when two groups of users
Fig. 6 shows the real trajectories of the estimated number of
users of two representative users for M = 5 and D = 20.
Similar to the case of D = 1 as shown in Fig. 4, the estimate
is able to capture changes within a few update intervals,
but the accuracy degrades when the actual number of users
increases.
Fig. 7 shows the real trajectories of individual SDPs of
two representative users for M = 5 and D = 20. We observe
that both users are able to adapt to different changes in the
number of users, and achieve SDPs close to the theoretical
limit. In addition, to evaluate the fairness issue, we in Table 2
compare the mean and variance of individual SDP among
all active users against the theoretical maximum SDP for
5.2.2 The cases with D > 1
9
Tuning parameters
L = 50000,
δ = 0.7
L = 20000,
δ = 0.9
Interval index
1 − 100
101 − 400
401 − 500
1 − 100
101 − 400
401 − 500
Theoretical maximum value
0.1357
0.0656
0.1357
0.1357
0.0656
0.1357
Mean in our simulation
0.1328
0.0646
0.1344
0.1290
0.0647
0.1337
Variance in our simulation
0.0012
5.54 · 10−4
6.80 · 10−4
0.0027
0.0016
0.0010
TABLE 1
Mean and variance of individual SDP among all active users at different stages for M = 5 and D = 1.
D = 20. We confirm that the average performance of each
user is near-optimal at every stage.
6
Since
1 − Dx − (1 − Dx + x)(1 − x)D
= 1 − Dx − (1 − Dx + x)(1 − x)D
< 1 − Dx − (1 − Dx + x)
= − x < 0,
C ONCLUSION
In this paper, we investigated the impact of MPR capability
M and delivery deadline D on the optimal transmission
probability that maximizes the SDP of the slotted ALOHA
protocol in a communication channel shared by N users
with saturated traffic. We have obtained the optimal transmission probability for any 1 ≤ M < N and any D ≥ 1, and
shown it can be computed by a fixed-point iteration. Then,
we developed an adaptive tuning algorithm for maximizing
the SDP when N is unknown and time-varying. Simulation
results show the proposed algorithm is effective in a variety
of dynamic scenarios.
As a special case of our work, the maximization of the
SDP when D = 1 can be viewed as the maximization of
individual throughput. It should be pointed out that [10] is
a first attempt to deal with this issue, however, the optimal
transmission probability therein was obtained by necessarily assuming ddτ T (τ ) < N − 1 for any 1 ≤ M < N in the
interval 0 < τ < 1. Therefore, the throughput maximization
for an M -user MPR channel is first completely addressed in
this paper.
(23)
where (23) is due to 0 < (1 − x)D < 1, we have ddx W (x) < 0
for x ∈ (0, 1), i.e., W (x) is decreasing. Moreover, by the
L’Hospital’s Rule we have
lim W (x) = 1.
x→0+
This concludes that W (x) < 1 for x ∈ (0, 1), which implies
(22). Hence we complete the proof.
ACKNOWLEDGEMENTS
This work was partially supported by the National Natural Science Foundation of China (No. 11601454, 61472190,
61501238), the Open Research Fund of National Mobile
Communications Research Laboratory, Southeast University
(No. 2017D09, 2017D04), and the Natural Science Foundation of Fujian Province of China (No. 2016J05021).
R EFERENCES
A PPENDIX
[1]
Proof of (16). We first have
[2]
d
dH1 (x)
H1 (x) + 1
g(x) = x
+
dx
dx
(H2 (x) + 1)2
dH2 (x)
· 1 + H2 (x) + x
.
dx
[3]
[4]
Since, for x ∈ (0, 1), ddx H1 (x) > 0 by Lemma 2 and H1 (x) >
0 by definition, to prove ddx g(x) > 0 it suffices to show that
1 + H2 (x) + x
dH2 (x)
D2 x2 (1 − x)D−1
=1 −
dx
(1 − (1 − x)D )2
>0,
[5]
[6]
(22)
[7]
for x ∈ (0, 1).
Let W (x) :=
D2 x2 (1−x)D−1
(1−(1−x)D )2 .
Observe that
2D2 x(1 − x)D−2
d
W (x) =
(1 − Dx) 1 − (1 − x)D
D
3
dx
(1 − (1 − x) )
− x(1 − x)D .
[8]
[9]
N. Abramson, “The ALOHA system: Another alternative for computer communications,” in Proc. Fall Joint Computer Conf., AFIPS
Conf. Proc., vol. 44, Montvale, NJ., 1970, pp. 281–285.
S. Ghez, S. Verdu, and S. Schwartz, “Stability properties of slotted ALOHA with multipacket reception capability,” IEEE Trans.
Automatic Control, vol. 33, no. 7, pp. 640–649, July 1988.
S. Ghez, S. Verdu, and S. Schwartz, “Optimal decentralized control
in the random access multipacket channel,” IEEE Trans. Automatic
Control, vol. 34, no. 11, pp. 1153–1163, Nov. 1989.
V. Naware, G. Mergen, and L. Tong, “Stability and delay of finiteuser slotted ALOHA with multipacket reception,” IEEE Trans.
Information Theory, vol. 51, no. 7, pp. 2636–2656, July 2005.
J. Luo and A. Ephremides, “On the throughput, capacity, and
stability regions of random multiple access,” IEEE Trans. Inf.
Theory, vol. 52, no. 6, pp. 2593–2607, Jun. 2006.
R.-H. Gau, “Performance analysis of slotted ALOHA in interferencedominating wireless ad-hoc networks, IEEE Commun. Lett.,
vol. 10, no. 5, pp. 402–404, May 2006.
R.-H. Gau, “Performance analysis of finite-user slotted ALOHA
in wireless networks with multiple packet reception and random
traffic,” IEEE Commun. Lett., vol. 12, no. 2, pp. 140–142, Feb. 2008.
Y. J. Zhang, P. Zheng, and S. C. Liew, “How does multiple-packet
reception capability scale the performance of wireless local area
networks?” IEEE Trans. Mobile Comput., vol. 8, no. 7, pp. 923–935,
Jul. 2009.
Y. J. Zhang, S.C. Liew, and D.R. Chen, “Sustainable throughput
of wireless LANs with multipacket reception capability under
bounded delay-moment requirements, IEEE Trans. Mobile Comput.,
vol. 9, no. 9, pp. 1226–1241, Sept. 2010.
10
100
100
One of initially active users
One of new users
Actual value
80
70
60
50
40
30
20
10
One of initially active users
One of new users
Actual value
90
The estimated number of users
The estimated number of users
90
80
70
60
50
40
30
20
0
50
100
150
200
250
300
Interval index
350
400
450
10
500
0
50
100
(a) L = 50000, δ = 0.7
150
200
250
300
Interval index
350
400
450
500
400
450
500
(b) L = 20000, δ = 0.9
Fig. 6. The estimated number of users for M = 5 and D = 20 when the number of users sharply changes.
1
1
One of initially active users
One of new users
Theoretical maximum value
One of initially active users
One of new users
Theoretical maximum value
0.9
Successful delivery probability
Successful delivery probability
0.9
0.8
0.7
0.6
0.5
0.8
0.7
0.6
0.5
0.4
0.4
0
50
100
150
200
250
300
Interval index
350
400
450
500
0
50
(a) L = 50000, δ = 0.7
100
150
200
250
300
Interval index
350
(b) L = 20000, δ = 0.9
Fig. 7. Individual SDP for M = 5 and D = 20 when the number of users sharply changes.
[10] Y. H. Bae, B. D. Choi, and A. S. Alfa, “Achieving maximum
throughput in random access protocols with multipacket reception,” IEEE Trans. Mobile Comput., vol. 13, no. 3, pp. 497–511, Mar.
2014.
[11] R. Stanica, E. Chaput, and A.-L. Beylot, “Properties of the MAC
layer in safety vehicular ad hoc networks,” IEEE Commun. Magazine, vol. 50, no. 5, pp. 192–200, May 2012.
[12] M. R. Palattella, M. Dohler, A. Grieco, G. Rizzo, J. Torsner, T.
Engel, and L. Ladid, “Internet of Things in the 5G era: Enablers,
architecture, and business models,” IEEE J. Sel. Areas Commun.,
vol. 34, no. 3, pp. 510–527, Mar. 2016.
[13] Y. H. Bae, “Analysis of optimal random access for broadcasting
with deadline in cognitive radio networks,” IEEE Commun. Lett.,
vol. 17, no. 3, pp. 573–575, Mar. 2013.
[14] Y. H. Bae, “Optimal retransmission-based broadcasting under
delivery deadline constraint,” IEEE Commun. Lett., vol. 19, no. 6,
pp. 1041–1044, Jun. 2015.
[15] Y. H. Bae, “Queueing analysis of deadline-constrained broadcasting in wireless networks,” IEEE Commun. Lett., vol. 19, no. 10, pp.
1782–1785, Oct. 2015.
[16] A. Vinel, “3GPP LTE versus IEEE 802.11p/WAVE: which technology is able to support cooperative vehicular safety applications?”
IEEE Wireless Commun. Lett., vol. 1, no. 1, pp. 125–128, 2012.
[17] C. Campolo, A. Molinaro, A. Vinel, and Y. Zhang, “Modeling prioritized broadcasting in multichannel vehicular networks,” IEEE
Trans. Veh. Technol., vol. 61, no. 2, pp. 687–701 Feb. 2012.
[18] Y. Birk and Y. Keren, “Judicious use of redundant transmissions in
multichannel ALOHA networks with deadlines,” IEEE J. Sel. Areas
Commun., vol. 17, no. 2, pp. 257–269, Feb. 1999.
[19] Dror Baron and Y. Birk, “Coding schemes for multislot messages
in multichannel ALOHA with deadlines,” IEEE Trans. Wireless
Commun., vol. 1, no. 2, pp. 292–301, Apr. 2002.
[20] Y. Birk and U. Tal, “Maximizing delay-constrained throughput
in multi-channel DS-CDMA ALOHA networks through power
diversity and successive decoding,” Wireless Netw., vol. 15, no. 8,
pp. 1126-1139, Nov. 2009.
[21] R. L. Rivest, “Network control by Bayesian broadcast,” IEEE Trans.
Inf. Theory, vol. IT-33, no. 3, pp. 323–328, May 1987.
[22] B. Hajek and T. van Loon, “Decentralized dynamic control of a
multiaccess broadcast channel,” IEEE Trans. Autom. Control, vol.
AC-27, no. 3, pp. 559–569, Jun. 1982.
[23] H. Wu, C. Zhu, R. La, X. Liu, and Y. Zhang, “FASA: Accelerated
SALOHA using access history for event-driven M2M communications,” IEEE/ACM Trans. Networking, vol. 21, no. 6, pp. 1904–1917,
Dec. 2013.
[24] G. Bianchi, I. Tinnirello, “Kalman filter estimation of the number
of competing terminals in an IEEE 802.11 network,” IEEE Infocom
2003, vol. 2, pp. 844–852, March 2003.
[25] F. Cali, M. Conti, and E. Gregori, “Dynamic tuning of the
IEEE 802.11 protocol to achieve a theoretical throughput limit,”
IEEE/ACM Trans. Networking, vol. 8, no. 6, pp. 785-799, Dec. 2000.
[26] F. Cali, M. Conti, and E. Gregori, “IEEE 802.11 protocol: Design
11
Tuning parameters
L = 50000,
δ = 0.7
L = 20000,
δ = 0.9
Interval index
1 − 100
101 − 400
401 − 500
1 − 100
101 − 400
401 − 500
Theoretical maximum value
0.8595
0.6628
0.8595
0.8595
0.6628
0.8595
Mean in our simulation
0.8525
0.6578
0.8558
0.8547
0.6579
0.8545
Variance in our simulation
0.0017
0.0028
0.0011
0.0053
0.0054
0.0022
TABLE 2
Mean and variance of individual SDP among all active users at different stages for M = 5 and D = 20.
and performance evaluation of an adaptive backoff mechanism,
IEEE J. Selected Areas Comm., vol. 18, no. 9, pp. 1774–1786, Sept.
2000.
[27] L. C. Godara, “Application of antenna arrays to mobile communicationsCpart II: beam forming and direction-of-arrival considerations,” Proc. IEEE, vol. 85, pp. 1193–1245, Aug. 1997.
[28] A. Mukhopadhyay, N. Mehta, and V. Srinivasan, “Design and
analysis of an acknowledgment-aware asynchronous MPR MAC
protocol for distributed WLANs,” IEEE Trans. Wireless Commun.,
vol. 12, no. 5, pp. 2068–2079, May 2013.
| 7cs.IT
|
arXiv:1605.03017v3 [math.ST] 27 Jul 2016
Marginalized Particle Filtering and Related
Filtering Techniques as Message Passing
July 29, 2016
Abstract
In this manuscript a factor graph approach is employed to investigate the recursive filtering problem for mixed linear/nonlinear state-space
models. Our approach allows us to show that: a) the factor graph characterizing the considered filtering problem is not cycle free; b) in the case
of conditionally linear Gaussian systems, applying the sum-product rule,
together with different scheduling procedures for message passing, to this
graph results in both known and novel filtering techniques. In particular,
it is proved that, on the one hand, adopting a specific message scheduling
for forward only message passing leads to marginalized particle filtering
in a natural fashion; on the other hand, if iterative strategies for message passing are employed, a novel filtering method, dubbed turbo filter
for its conceptual resemblance to the turbo decoding methods devised for
concatenated channel codes, can be developed.
Giorgio M. Vitetta, Emilio Sirignano, Francesco Montorsi and Matteo Sola
University of Modena and Reggio Emilia
Department of Engineering ”Enzo Ferrari”
Via P. Vivarelli 10/1, 41125 Modena - Italy
email: [email protected], [email protected],
[email protected], [email protected]
Keywords: State Space Representation, Hidden Markov Model, Particle
Filter, Belief Propagation, Turbo Processing.
1
G. M. Vitetta et al., Marginalized Particle Filtering . . .
1
2
Introduction
The nonlinear filtering problem consists of inferring the posterior distribution of
the hidden state of a nonlinear dynamic system from a set of past and present
measurements [1]. It is well known that, if a nonlinear dynamic system can be
described by a state-space model (SSM), a general sequential procedure, based
on the Bayes’ rule and known as Bayesian filtering, can be easily derived for
recursively computing the posterior distribution of the system current state
[1]. Unluckily, Bayesian filtering is analytically tractable in few cases for the
following two reasons [2]: a) one of the two steps it consists of requires multidimensional integration which, in most cases, does not admit a closed form
solution; b) the functional form of the required posterior distribution may not
be preserved over successive recursions. For this reason, sequential techniques
employed in practice are based on various analytical approximations and, consequently, generate a functional approximation of the desired distribution. Such
techniques are commonly divided into local and global methods on the basis of
the way posterior distributions are approximated [3, 4, 5]. Local methods, like
extended Kalman filtering [6] and unscented filtering [7], are computationally
efficient, but may suffer from the problem of error accumulation over time. On
the contrary, global methods, like sequential Monte Carlo methods [8, 9] (also
known as particle filtering, PF, methods [10, 11, 12]) and point mass filtering
[5, 13] may achieve high accuracy at the price, however, of an unmanageable
complexity and numerical problems in the presence of a large dimension of system state [14]. These considerations have motivated various research activities
focused on the development of novel Bayesian filters able to achieve high accuracy under given computational constraints. Significant results in this research
area concern the use of the new representations for complex distributions, like
belief condensation filtering [3], and the development of novel filtering techniques combining local and global methods, like marginalized particle filtering
(MPF) [15, 16], and other methods originating from it [4, 17, 18]. Note that
the last class of methods applies to mixed nonlinear/nonlinear models [19], that
is to models whose state can be partitioned in a conditionally linear portion
(usually called linear state variable) and in a nonlinear portion (representing
the remaining part of system state and called nonlinear state variable). This
partitioning of system state allows to combine a global method (e.g., particle
filtering) operating on the nonlinear state variable with a local technique (e.g.,
Kalman filtering) involving the linear state variable only.
In this manuscript the factor graph (FG) approach illustrated by Loeliger et
G. M. Vitetta et al., Marginalized Particle Filtering . . .
3
al. in [20] is employed to revisit the problem of recursive Bayesian filtering for
mixed linear/nonlinear models from a perspective substantially different from
that adopted in MPF [15]. This allows us to shed new light on the problem of
filtering for mixed linear/nonlinear models, providing a new interpretation of
MPF and paving the way for the development of new filtering techniques. In
particular, based on this approach, we are able to show that: a) the considered
filtering problem can be formulated as a message massing problem over a specific
FG, which, unluckily, is not cycle free; b) in the case of a conditionally linear
Gaussian (CLG) SSM [19], MPF results from the application of the sum-product
algorithm (SPA) [20, 21], together with a specific scheduling procedure for forward only message passing, to this graph; c) our graphical representation leads,
in a natural fashion, to the development of novel filtering methods simplifying
and/or generalising it. As far as the last point is concerned, in our work specific
attention is paid to the development of a novel iterative filtering technique that
exploits the exchange of probabilistic (i.e., soft) messages to progressively refine
the posteriors of the linear and nonlinear state variables within each recursion
and is dubbed turbo filtering (TF) for its conceptual resemblance to the iterative
(i.e., turbo) decoding of concatenated channel codes.
It is important to point that our approach has been inspired by various
ideas and results already available in the technical literature concerning different
research areas; here, we limit to mention the following relevant facts:
• A mixed linear/nonlinear Markov system can be represented as the concatenation of two interacting subsystems, one governed by linear dynamics, the other one accounting for a nonlinear behavior; conceptually related
(finite state) Markov models can be found in data communications and,
in particular, in concatenated channel coding (e.g., turbo coding [22]) and
in coded transmissions over inter-symbol interference channels for which
turbo decoding methods [22, 23] and turbo equalization techniques [24] have
been developed, respectively1 .
• Factor graphs play an essential role in the derivation and interpretation
of turbo decoding and equalization [20, 26] (for instance, turbo decoding
techniques emerge in a natural fashion from graphical models of codes
[27]).
• Both Kalman filtering and particle methods can be viewed as message
1 Note that these classes of algorithms can be seen as specific applications of the so called
turbo principle [25], [35, Par. 10.5.1]
4
G. M. Vitetta et al., Marginalized Particle Filtering . . .
passing procedures on factor graphs, as shown in [20, 21] and in [28], respectively.
• Various methods to progressively refine distributional approximations through
multiple iterations have been developed in the field of Bayesian inference
on dynamic systems (even if implementations substantially different from
that we devise have been proposed), and, in particular, in expectation
propagation in Bayesian networks [29, 30] and in variational Bayesian
filtering [4]. Consequently, various links to previous work on Bayesian
inference on graphical models and variational Bayes methods [31] can be
also established.
The remaining part of this manuscript is organized as follows. The mathematical model of the considered class of mixed linear/nonlinear systems is illustrated in Section 2, whereas a representation of the filtering problem for these
systems through a proper FG is provided in Section 3. Then, it is shown that
applying the SPA and proper message scheduling strategies to this FG leads
to MPF in Section 4. This approach paves the way, in a natural fashion, for
the development of simplifications and generalizations of MPF, that are devised
in Sections 5 and 6, respectively. The novel filtering methods proposed in this
manuscript are compared, in terms of accuracy and computational effort, with
MPF in Section 7. Finally, some conclusions are offered in Section 8.
Notations: The probability density function (pdf) of a random vector R
evaluated at point r is denoted f (r); N (r; ηr , Cr ) represents the pdf of a Gaussian random vector R characterized by the mean ηr and covariance matrix Cr
evaluated at point r; the precision (or weight ) matrix associated with the covariance matrix Cr is denoted Wr , whereas the transformed mean vector Wr ηr
is denoted wr ; xi denotes the i-th element of the vector x.
2
System Model
In the following we focus on a discrete-time mixed linear/nonlinear SSM [15],
whose hidden state in the l-th interval is represented by a D-dimensional real
vector xl , [x0,l , x1,l , ..., xD−1,l ]T . We assume that this vector can be partitioned as
T
T T
(L)
(N )
xl = xl
, xl
,
(1)
(L)
(L)
(L)
(L)
(N )
(N )
(N )
(L)
where xl , [x0,l , x1,l , ..., xDL −1,l ]T (xl , [x0,l , x1,l , ..., xDN −1,l ]T ) is the
so called linear (nonlinear ) component of xl (1), with DL < D (DN = D − DL ).
5
G. M. Vitetta et al., Marginalized Particle Filtering . . .
(L)
This partitioning of xl is accomplished as follows. First, xl is identified as
that portion of xl (1) characterized by the following two properties:
1. Conditionally linear dynamics - This means that its update equation, con(N )
ditioned on xl , is linear, so that
(L)
(L)
(N )
(L)
(L)
(N )
(L)
xl+1 = Al
xl
xl + fl
xl
+ wl ,
(2)
(L)
(L)
(N )
where fl (x) is a time-varying DL -dimensional real function, Al (xl )
(L)
is a time-varying DL × DL real matrix and wl is the l-th element of
(L)
the process noise sequence {wk }, which consists of DL - dimensional
independent and identically distributed (iid) noise vectors.
2. Conditionally linear (or almost linear ) dependence of all the available
measurements on it - In other words, these quantities, conditioned on
(N )
(L)
xl , exhibit a linear dependence on xl (additional details about this
feature are provided below).
(N )
Then, xl
is generated by putting together all the components of xl that
(L)
do not belong to xl . For this reason, generally speaking, this vector is characterized by at least one of the following two properties:
a) Nonlinear dynamics - The update equation
(N )
(N )
(N )
(N )
(N )
(L)
(N )
xl+1 = fl
xl
+ Al
xl
xl + wl
(3)
is assumed in the following for the nonlinear component of system state, where
(N ) (N )
(N )
Al (xl ) is a time-varying DN × DL real matrix, fl (x) is a time-varying
(N )
DN -dimensional real function and wl is the l-th element of the process noise
(N )
sequence {wk }, which consists of DN -dimensional iid noise vectors and is
(L)
statistically independent of {wk }.
b) A nonlinear dependence of all the available measurements on it (further
details are provided below).
In the following Section we focus on the so-called filtering problem, which
concerns the evaluation of the posterior pdf f (xt |y1:t ) at an instant t > 1, given
a) the initial pdf f (x1 ) and b) the t · P -dimensional measurement vector
T
y1:t = y1T , y2T , ..., ytT ,
(4)
where yl , [y0,l , y1,l , ..., yP −1,l ]T denotes the P -dimensional real vector collecting all the noisy measurements available at time l. As already mentioned
G. M. Vitetta et al., Marginalized Particle Filtering . . .
6
above, the measurement vector yl exhibits a linear (nonlinear ) dependence on
(L)
(N )
xl (xl ), so that the model [18]
(N )
(L)
(N )
xl + el
(5)
+ Bl xl
yl = hl xl
(N )
(N )
can be adopted, where Bl (xl ) is a time-varying P × DL real matrix, hl (xl )
is a time-varying P -dimensional real function and el the l-th element of the
measurement noise sequence {ek } consisting of P -dimensional iid noise vectors
(N )
(L)
and independent of both {wk } and {wk }.
3
Representation of the Filtering Problem via
Factor Graphs
Generally speaking, the filtering problem for a SSM described by the Markov
model f (xl+1 |xl ) and the observation model f (yl |xl ) for any l concerns the
computation of the posterior pdf f (xt |y1:t ) for t ≥ 1 by means of a recursive
procedure [1]. It is well known that, if the pdf f (x1 ) is known, a general Bayesian
recursive procedure, consisting of a measurement update step followed by a time
update step, can be employed. In practice, in the first step of the l-th recursion
(with l = 1, 2, ..., t) the conditional pdf
f (xl |y1:l ) = f xl y1:(l−1) f (yl |xl )
1
f yl y1:(l−1)
(6)
is computed on the basis of pdf f (xl |y1:(l−1) ) (evaluated in the last step of the
previous recursion2 ), the present measurement vector yl and the pdf
Z
f yl y1:(l−1) = f (yl |xl ) f xl y1:(l−1) dxl .
(7)
In the second step f (xl |y1:l ) (6) is exploited to compute the pdf
Z
f (xl+1 |y1:l ) = f (xl+1 |xl ) f (xl |y1:l ) dxl ,
(8)
which represents a prediction about the future state xl+1 . It is important to
point out that: 1) the term 1/f (yl |y1:(l−1) ) appearing in the right hand side
(RHS) of (6) represents a normalization factor ; 2) both (7) and (8) require
integration with respect to xl and this may represent a formidable task when
the dimensionality of xl is large and/or the pdfs appearing in the integrands are
2 Note that in the first recursion (i.e., for l = 1) f (x |y
l 1:(l−1) ) = f (x1 |y1:0 ) = f (x1 ) and
f (yl |y1:(l−1) ) = f (y1 |y1:0 ) = f (y1 ), so that f (x1 |y1 ) = f (x1 )f (y1 |x1 )/f (y1 ).
G. M. Vitetta et al., Marginalized Particle Filtering . . .
7
not Gaussian; 3) this recursive procedure lends itself to be efficiently represented
by a message passing algorithm over a proper FG3 [28], in which each factor
of a product of functions is represented by a distinct node (a rectangle in our
diagrams), whereas each variable is associated with a specific (and, usually,
unoriented) edge or half edge. As far as the last point is concerned, we also
note that the derivation of this FG relies on the fact that the a posteriori pdf
f (xt |y1:t ) has the same FG as the joint pdf f (xt , y1:t ) (see [20, Sec. II, p. 1297])
and the last pdf can be computed recursively through a procedure similar to
that illustrated above, but in which the measurement update (6) and the time
update (8) are replaced by
and
f (xl , y1:l ) = f xl , y1:(l−1) f (yl |xl ) ,
f (xl+1 , y1:l ) =
Z
f (xl+1 |xl ) f (xl , y1:l ) dxl ,
(9)
(10)
respectively, so that the evaluation of the above mentioned normalization factor
is no more required. In fact, eqs. (9) and (10) involve only products of pdfs
and a sum (i.e., integration) of products, so that they can be represented by
means of the FG shown in Fig. 1 (where, following [20], a simplified notation
is used for the involved pdfs and the equality constraint node, which represents
an equality constraint “function”, that is a Dirac delta function). Since this
FG is cycle free, the pdf f (xl , y1:l ) can be evaluated applying the well known
sum-product rule 4 (i.e., the SPA) to it, i.e. developing a proper mechanism for
passing probabilistic messages along this FG (the flow of messages is indicated
by red arrows in Fig. 1). In fact, if the input message m
~ in (xl ) = f (xl , y1:(l−1) )
enters this FG, the message going out of the equality node is given by
m
~ e (xl ) = m
~ in (xl ) f (yl |xl ) ,
(11)
so that m
~ e (xl ) = f (xl , y1:l ) (see (9)); then, the message emerging from the
function node referring to the pdf f (xl+1 |xl ) is expressed by
Z
m
~ out (xl+1 ) = f (xl+1 |xl ) m
~ e (xl ) dxl ,
(12)
so that m
~ out (xl+1 ) = f (xl+1 , y1:l ) = m
~ in (xl+1 ) (see (10)). From this result it
can be easily inferred that the pdf f (xt , y1:t ) (and, up to a scale factor, the pdf
3 Forney-style
factor graphs are always considered in the following [20].
a Forney-style FG, such a rule can be formulated as follows [20]: the message emerging
from a node f along some edge x is formed as the product of f and all the incoming messages
along all the edges that enter the node f except x, summed over all the involved variables
except x.
4 In
G. M. Vitetta et al., Marginalized Particle Filtering . . .
8
xl
min ( xl )
=
f y l / xl
f xl +1 / xl
yl
me ( xl )
x l +1
mout ( xl +1 ) = min ( xl +1 )
Figure 1: Factor graph representing (9) and (10). The SPA message flow characterizing the l-th recursion of Bayesian filtering is indicated by red arrows.
f (xt |y1:t )) results from the application of the SPA to the overall FG originating
from the ordered concatenation of multiple subgraphs, each structured like the
one shown in Fig. 1 and associated with l = 1, 2, ..., t. In this graph the flow of
messages produced by the SPA proceeds from left to right, i.e. the pdf f (xt , y1:t )
is generated by a forward only message passing. Note also that, in principle, the
desired pdf f (xt , y1:t ) is computed as the product between two messages, one
for each direction, reaching the rightmost half edge of the overall FG, but one of
− (x ) = 1.
the two incoming messages for that edge is the constant function ←
m
he t
Unluckily, as the size D of xl (1) gets large, the computational burden associated with (6)-(8) (or, equivalently, (9) and (10)) becomes unmanageable. In
principle, a substantial complexity reduction can be achieved decoupling 5 the
(L)
(N )
(L)
filtering problem for xl from that for xl , i.e. the evaluation of f (xl |y1:l )
(N )
from that of f (xl |y1:l ). In fact, this approach potentially provides the following two benefits: a) a given filtering problem is turned into a couple of filtering
problems of smaller dimensionality and b) some form of computationally efficient
standard filtering (e.g., Kalman or extended Kalman filtering) can be hopefully
(L)
exploited for the linear portion xl of the state vector xl (1). As a matter of
fact, these principles have been exploited in devising MPF [10, 15] and, as it will
become clearer in the following, they must be always kept into account in the
derivation of our FG representation. Before illustrating this derivation, however, the measurement and state models on which such a representation relies
need to be clearly defined; for this reason, these models are analysed in detail
in the following part of this Section. To begin, let us concentrate on the models
(L)
(L)
involved in the filtering problem for xl , i.e. on the evaluation of f (xl |y1:l ),
5 Note
that, generally speaking, in the considered problem the coupling of the filtering
(L)
(N)
problem for xl with that for xl
is due not only to the structure of the update equations
(2) and (3), but also to the measurement vector yl (5), since this exhibits a mixed dependence
on the two components of the state vector xl (1).
9
G. M. Vitetta et al., Marginalized Particle Filtering . . .
(N )
under the assumption that the nonlinear portion xl
of the system state is
known for any l. In this case, the evaluation of this pdf can benefit not only
from the knowledge of yl (5), but also from that of the quantity (see (3))
(L)
(N )
(N )
(N )
(N )
(N )
(L)
(N )
zl , xl+1 − fl
xl
= Al
xl
xl + wl ,
(13)
which can be interpreted as a pseudo-measurement [15], since it does not originate from real measurements, but from the constraints expressed by the state
equation (3). This leads to considering the overall observation model
(L)
(L)
(N )
f yl , zl xl , xl
(14)
(L)
(N )
(L)
(L)
(N )
= f yl xl , xl
f zl xl , xl
(L)
for xl , where
(L)
(N )
f yl xl , xl
= f (el )|e =y −B x(N ) x(L) −h x(N ) ,
l
l
l
l
l
l
(15)
l
and
(L)
(L)
(N )
(N )
f zl xl , xl
= f wl
(N )
wl
(L)
=zl
(N )
−Al
(L)
(N )
xl
xl
.
(16)
If the observation model (14) and the state model (see (2))
(L)
(L)
(N )
f xl+1 xl , xl
(L)
(L)
(N )
(L)
(N )
(L)
xl
− Al
xl
xl
= fw(L) xl+1 − fl
(17)
(L)
are adopted for xl , the graph identified by the blue lines and rectangles appearing in Fig. 1 can be drawn. Then, in principle, if the sum-product rule is
(N )
(N )
applied to it under the assumption that the couple (xl , xl+1 ) is known for any
l, the expressions of the messages flowing in the overall graph for the evaluation
(L)
(L)
of f (xt , y1:t , z1:t ) can be easily derived. It is important to point out that:
• This graph contains a node which does not refer to the above mentioned
density factorizations6 , but represents the transformation from the couple
(N )
(N )
(L)
(xl , xl+1 ) to zl (see (13)); this feature of the graph has to be carefully
kept into account when deriving message passing algorithms.
(L)
(L)
(N )
• Generally speaking, the evaluation of the conditional pdf f (zl |xl , xl )
(N )
(N )
requires the knowledge of the joint pdf of xl+1 and xl
conditioned on
(L)
xl (see (13)).
6 This peculiarity is also evidenced by the presence of an arrow on all the edges connected
to such a node.
10
G. M. Vitetta et al., Marginalized Particle Filtering . . .
The same line of reasoning can be followed for the filtering problem concern(N )
(L)
ing xl . Consequently, in this case the linear portion xl of the system state
is assumed to be known for any l and the pseudo-measurement (see (2))
(N )
(L)
(L)
(N )
(L)
(L)
(N )
(L)
zl , xl+1 − Al
xl
xl = fl
xl
+ wl
(18)
is defined. This leads to the overall observation model
(N )
(N )
(L)
f yl , zl
xl , xl
=
(N )
(L)
(N )
(N )
f yl xl , xl
f zl
xl
(N )
(N )
(N )
(19)
(L)
(L)
(N )
for xl , where f (zl |xl ) can be expressed similarly as f (zl |xl , xl )
(see (16)). Then, if the observation model (19) and the state model
(N )
(N )
(L)
f xl+1 xl , xl
(20)
(N )
(N )
(N )
(N )
(N )
(L)
xl
− Al
xl
xl
= fwN xl+1 − fl
(N )
are adopted for xl , the red graph of Fig. 2 can be drawn and exploited in
(N )
(N )
a similar way as the blue graph for the evaluation of f (xt , y1:t , z1:t ), under
(L)
(L)
the assumption that the couple (xl , xl+1 ) is known for any l. Note also
that, similarly to what has been mentioned earlier about the conditional pdf
(L) (L)
(N )
(N ) (N )
f (zl |xl , xl ), the evaluation of f (zl |xl ) requires the knowledge of the
(L)
(L)
(N )
joint pdf of xl+1 and xl conditioned on xl .
Finally, merging the blue graph with the red one (i.e., adding four equality
(L)
(N )
(L)
(N )
constraint nodes for the variables xl , xl , xl+1 and xl+1 shared by the red
graph and the blue one) produces the overall FG illustrated in Fig. 2. Given this
FG, we would like to follow the same line of reasoning as that adopted for the FG
(L)
(L)
of Fig. 1. In other words, given the input messages m
~ in (xl ) = f (xl , y1:(l−1) )
(N )
(N )
and m
~ in (xl ) = f (xl , y1:(l−1) ) (entering the FG along the half edges asso(L)
(N )
ciated with xl and xl , respectively), we would like to derive a forward
only message passing algorithm based on this FG and generating the output
(L)
(L)
(N )
(N )
messages m
~ out (xl+1 ) = f (xl+1 , y1:l ) and m
~ out (xl+1 ) = f (xl+1 , y1:l ) (emerging
(L)
(N )
from the FG along the half edges associated with xl+1 and xl+1 , respectively)
on the basis of the available a priori information and the noisy measurement
yl . Unluckily, the new FG, unlike that shown in Fig. 1, is not cycle-free, so
that any application of the SPA to it unavoidably leads to approximate results
[21], whatever message scheduling procedure [21, 27] is adopted. This consideration must be carefully kept into account in both the derivation of MPF as
a message passing algorithm and in the development of possible modifications
and generalizations of this technique, as it will become clearer in Sections 4-6.
11
G. M. Vitetta et al., Marginalized Particle Filtering . . .
mout ( xl(+L1) )
xl( L )
min ( xl( L ) )
( L)
xl+
1
f x( L ) / x( L ) , x ( N )
l +1
=
=
l
l
=
=
f z ( L ) / x ( L ) ,x ( N )
l
z
l
l
( L)
l
xl( N ) , xl(+N1) → z l( L )
=
f y / x ( L ) ,x ( N )
l
l
l
yl
=
xl( L ) , xl(+L1) → z l( N )
z l( N )
f z ( N ) / x( N )
l
x
(N )
l
=
min ( xl( N ) )
=
l
=
=
f x( N ) / x( N ) , x( L )
l +1
l
l
mout ( xl(+N1) )
(N )
xl+
1
Figure 2: Overall factor graph resulting from the merge of two subgraphs, one
(L)
(N )
referring to filtering for xl (in blue), the other one to that for xl (in red).
The equality constraint nodes introduced to connect these subgraphs are identi(L)
(N )
fied by black lines. The flow of the messages along the half edges xl and xl
(L)
(N )
(input) and that of the messages along the half edges xl+1 and xl+1 (output)
are indicated by green arrows.
G. M. Vitetta et al., Marginalized Particle Filtering . . .
4
12
Message Passing in Marginalized Particle Filtering
In the following Section we show how the equations describing the l-th recursion of MPF result from the application of the SPA to the FG shown in Fig. 2.
However, before illustrating the detailed derivation of such equations, it is important to discuss the following relevant issues. First of all, the MPF technique
has been developed for the specific class of GLG SSMs [15, 19], to which we always refer in the following discussion. In particular, in the following we assume
(L)
(N )
that: a) {wk } ({wk }) is a Gaussian random process and all its elements
(L)
(N )
(L)
have zero mean and covariance Cw (Cw ) for any l; b) {ek } is a Gaussian
random process having zero mean and covariance matrix Ce for any l; c) all the
above mentioned Gaussian processes are statistically independent. Under these
(L)
(N
(L) (L)
(L) (L)
(N )
assumptions, the pdfs f (yl |xl , xl ), f (zl |xl ) and f (xl+1 |xl , xl ) (see
(N ) (L)
(N )
(15)-(17)) are Gaussian with mean (covariance matrix) Bl (xl )xl +hl (xl ),
(N ) (N )
(L)
(L) (N )
(L) (N )
(L)
(N )
Al (xl ) xl and fl (xl ) + Al (xl ) xl , respectively (Ce , Cw and
(N ) (N )
(N ) (N )
(L)
(L)
Cw , respectively). Similarly, the pdfs f (zl |xl ) and f (xl+1 |xl , xl )
(L) (N )
(see (19) and (20)) are Gaussian with mean (covariance matrix) fl (xl ) and
(N )
(N ) (N )
(N ) (N )
(L)
(L)
fl (xl ) + Al (xl ) xl , respectively (Cw and Cw , respectively).
Secondly, as explained below in detail, the MPF can be interpreted as a
forward only message passing algorithm operating over the FG shown in Fig.
2. The scheduling procedure adopted for MPF unavoidably leads to ignoring
(N )
the evaluation of the pseudo-measurement zl (18). For this reason, in the following we refer to the simplified FG shown in Fig. 3, which has been obtained
from that illustrated in Fig. 2 removing the block representing the transfor(L)
(L)
(N )
mation from (xl , xl+1 ) to zl
and the edges referring to the evaluation of
the last vector. Note that: a) in the new graph the block referring to the pdf
(L)
(N
f (yl |xl , xl ) appears twice, since this pdf is involved in the two subgraphs
shown in Fig. 2; b) some brown edges and equality nodes have been added to
(N )
(L)
feed such blocks with min (xl ) and min (xl ), since these represents the only a
(N )
(L)
priori information available about xl and xl , respectively, at the beginning
of the l-th recursion.
Thirdly, in MPF a particle-based model and a Gaussian model are adopted
(N )
(L)
for the input and the output messages referring to xl and xl , respectively,
and the functional structure of the generated messages is preserved in each
recursion. More specifically, on the one hand, the a priori information avail(N )
able about xl
at the beginning of the l-th recursion is represented by a set
(N )
(N )
of Np particles Sl/(l−1) = {xl/(l−1),j , j = 0, 1, ..., Np − 1} and their weights
G. M. Vitetta et al., Marginalized Particle Filtering . . .
13
{wl/(l−1),j , j = 0, 1, ..., Np − 1)}; following [15], we assume that such weights
are uniform (in other words, wl/(l−1),j = 1/Np for j = 0, 1, ..., Np − 1), so
that they can be ignored in the following derivation. On the other hand,
(L)
the a priori information available about xl is represented by a set of Gaussian pdfs, each associated with a specific particle; in particular, the Gaussian
(L)
(L) (L)
model N (xl ; ηl/(l.−1),j , Cl/(l−1),j ) is associated with the j-th particle (with
j = 0, 1, ..., Np − 1) at the beginning of the same recursion. From the last point
it can be inferred that, in developing a message passing algorithm that represents the MPF technique, we can focus on: a) a single particle contained in the
(N )
(N )
input message min (xl ) and, in particular, on the j-th particle xl/(l−1),j ; b)
(L)
(L)
(L)
on the Gaussian model N (xl ; ηl/(l.−1),j , Cl/(l−1),j ) associated with that particle. For this reason, we assume that, at the beginning of the l-th recursion,
(L)
(N )
our knowledge about xl and xl is condensed in the message
(L)
(L) (L)
(L)
(21)
= N xl ; ηl/(l.−1),j , Cl/(l−1),j
m
~ in,j xl
and in the message
(N )
(N )
(N )
= δ xl − xl/(l−1),j ,
m
~ in,j xl
(22)
respectively, with j = 0, 1, ..., Np − 1; these are processed to generate the corre(L)
(N )
sponding output messages m
~ out,j (xl+1 ) and m
~ out,j (xl+1 ), which are required to
(L)
(N )
have the same functional form as m
~ in,j (xl ) (21) and m
~ in,j (xl ) (22), respec(N )
tively. For this reason, the algorithm for computing m
~ out,j (xl+1 ) is expected
(N )
to generate a new particle x(l+1)/l,j with a (uniform) weight w(l+1)/l,j = 1/Np ;
(L)
similarly, that for evaluating m
~ out,j (xl+1 ) is expected to produce a new Gaus(N )
(L)
(L)
(L)
sian pdf N (xl+1 ; η(l+1)/l,j , C(l+1)/l,j ) associated with the particle x(l+1)/l,j (note
that, in deriving this pdf, possible scale factors are unrelevant and, consequently,
can be dropped).
(L)
(N )
Given m
~ in,j (xl ) (21) and m
~ in,j (xl ) (22), if the SPA is applied to the
considered graph and the message scheduling illustrated in Fig. 3 (and, as a
matter of fact, adopted in MPF) is employed, the steps described below are
(L)
(N )
carried out to evaluate m
~ out,j (xl+1 ) and m
~ out,j (xl+1 ) in the l-th recursion of
MPF7 .
(N )
1. Measurement update for xl
- This step aims at updating the weight
(N )
of the j-th particle xl/(l−1),j on the basis of the new measurements yl (this
7 In the following derivations some mathematical results about Gassian random variables
(e.g., see [32, Par. 2.3.3]) and Gaussian message passing in linear models (e.g., see [20, Table
2, p. 1303]) are exploited. As far as the MPF formulation is concerned, we always refer to
that given by algorithm 1 in [15, Sec. II].
14
G. M. Vitetta et al., Marginalized Particle Filtering . . .
min , j ( x (l L ) )
x (l L )
yl
min , j ( x l( L ) )
m2, j ( x (l L ) )
m2, j ( x l( L ) )
m4, j ( x l( L ) )
=
f x ( L ) / x ( L ) ,x ( N )
l +1
=
fy
m1, j ( x l( L ) )
l
,/ x l( L ) ,x l( N )
=
m3, j ( x (l L ) )
l
m j ( z l( L ) )
z
l
x (l +L1)
m5, j ( x l(+L1) )
= mout , j ( x (l +L1) )
f z( L ) / x ( L ) ,x ( N )
l
min , j ( x l( N ) )
l
=
l
m2, j ( x l( N ) )
( L)
l
x (l N ) , x l(+N1) → z l( L )
m5, j ( x (l +N1) )
m2, j ( x l( N ) )
=
m2, j ( x (l L ) )
m2, j ( x (l L ) )
min , j ( x l( L ) )
fy
l
=
x (l N )
m2, j ( x l( N ) )
m1, j ( x l( N ) )
,/ x l( L ) ,x l( N )
=
min , j ( x l( N ) )
m2, j ( x (l N ) )
yl
m2, j ( x (l N ) )
m5, j ( x (l +N1) )
=
=
f x ( N ) / x ( N ) ,x ( L )
l +1
l
l
mout , j ( x l(+N1) )
x (l +N1)
Figure 3: Overall factor graph for the representation of MPF processing; this
graph is obtained from the one shown in Fig. 3 removing the part referring to
(N )
the evaluation of zl and inserting two new (brown) equality constraints and
(L)
(N )
some (brown) edges referring to xl and xl . The message flow characterizing
MPF and referring to the j-th particle is also shown.
15
G. M. Vitetta et al., Marginalized Particle Filtering . . .
corresponds to step 2) of algorithm 1 in [15, Sec. II]). It involves the computation
of the messages
Z
(N )
(L)
(N )
(L)
(L)
m
~ 1,j xl
m
~ in,j xl
= f yl xl , xl
dxl
(23)
and
(N )
(N )
(N )
,
m
~ 1,j xl
=m
~ in,j xl
m
~ 2,j xl
(24)
(N )
(N )
(L)
(N )
(N )
ηl/(l.−1),j + hl xl
η1,l,j xl
, Bl xl
(26)
which provides the new importance weight for the considered particle (see Fig.
(N )
(L)
3). Substituting the expression of f (yl xl , xl ) (see (15)) and (21) in (23)
produces, after some manipulation (see the Appendix)
(N )
(N )
(N )
(N )
(N )
= N yl ; η1,l,j xl
, C1,l,j xl
,
(25)
m
~ 1,j xl
where
and
T
(N )
(N )
(L)
(N )
(N )
+ Ce .
Cl/(l−1),j Bl xl
C1,l,j xl
, Bl xl
(N )
(27)
(N )
Then, substituting m
~ in,j (xl ) (22) and m
~ 1,j (xl ) (25) in (24) yields
(N )
(N )
(N )
= wl,j δ xl − xl/(l−1),j ,
m
~ 2,j xl
(28)
where8
(N )
(N )
wl,j , N yl ; η1,l,j , C1,l,j
(29)
(N )
is the new particle weight combining the a priori information about xl with
the information provided by the new measurements; here (see (26) and (27))
(L)
(N )
(N )
(N )
(30)
η1,l,j , η1,l,j xl/(l−1),j = Bl,j ηl/(l.−1),j + hl,j
and
(L)
(N )
(N )
(N )
C1,l,j , C1,l,j xl/(l−1),j = Bl,j Cl/(l−1),j BTl,j + Ce ,
(31)
(N )
(N )
with hl,j , hl (xl/(l−1),j ) and Bl,j , Bl (xl/(l−1),j ). In MPF, after normal(w)
ization9 of the particle weights {wl,j } (i.e., after dividing them by Pl
(N)
,
8 In evaluating this weight, the factor [det(C
−P /2 appearing in the expression of the
1,l,j )]
involved Gaussian pdf is usually neglected, since this entails a negligible loss in estimation
accuracy.
9 Note that normalization requires the knowledge of all the weights {w } (29) and that,
l,j
unlike it, all the previous and following tasks can be carried out in parallel (i.e., on a particleby-particle basis).
16
G. M. Vitetta et al., Marginalized Particle Filtering . . .
Np −1
X
wl,j ), particle resampling with replacement is accomplished (this corre-
j=0
sponds to step 3) of algorithm 1 in [15, Sec. II]). Note that, even if resampling
does not emerge from the application of SPA to the considered graph, its use,
as it will become clearer at the end of this Section, plays an important role in
(N )
the generation of the new particles for xl+1 . Moreover, it can be easily incorporated in our message passing; in fact, resampling simply entails that Np parti(N )
cles {xl/(l−1),j } and their associated weights {wl,j } (29) are replaced by the new
(N )
(N )
particles {xl/l,j } (forming the new set Sl/l ) and their weights {wl/l,j = 1/Np },
(N )
respectively. Consequently, m
~ 2 (xl ) (28) is replaced by
(N )
(N )
(N )
= δ xl − xl/l,j ,
m
~ 2,j xl
(32)
since the particle weight does not depend on the index j. It is also worth
(L)
mentioning that, after resampling, the set of Gaussian messages {m
~ in,j (xl )}
(21) needs to be properly reordered and that the messages associated with all
the discarded particles are not propagated to the next steps.
(L)
2. Measurement update for xl - This step aims at updating our statistical
(L)
knowlege about xl on the basis of the new measurement yl (and corresponds
to step 3-a) of algorithm 1 in [15, Sec. II]). It involves the computation of the
messages
Z
(N )
(N )
(L)
(N )
(L)
dxl
(33)
m
~ in,j xl
= f yl xl , xl
m
~ 1,j xl
and
(L)
(L)
(L)
,
m
~ 1,j xl
=m
~ in,j xl
m
~ 2,j xl
(34)
(L)
which represents the output of the measurement update for xl (see Fig. 3).
Substituting (15) and (22) in (33) produces (see [32, Par. 2.3.3, eq. (2.115)])
(L)
(L)
(35)
= N yl ; Bl,j xl + hl,j , Ce
m
~ 1,j xl
which, after some manipulation (in which unrelevant scale factors are dropped),
can be put in the Gaussian form
(L)
(L) (L)
(L)
= N xl ; η1,l,j , C1,l,j ,
(36)
m
~ 1,j xl
with
(L)
(L)
(L)
w1,l,j , W1,l,j η1,l,j = BTl,j We (yl − hl,j ) ,
−1
(L)
(L)
W1,l,j , C1,l,j
= BTl,j We Bl,j
(37)
(38)
17
G. M. Vitetta et al., Marginalized Particle Filtering . . .
and We , C−1
e . Then, substituting (21) and (36) in (34) yields
(L)
(L)
(L) (L)
= N xl ; ηl/(l−1),j , Cl/(l−1),j
m
~ 2,j xl
(L) (L)
(L)
·N xl ; η1,l,j , C1,l,j ,
which can be reformulated as
(L)
(L) (L)
(L)
= N xl ; η2,l,j , C2,l,j ,
m
~ 2,j xl
(39)
(40)
if scale factors are ignored; here,
(L)
(L)
(L)
(L)
(41)
−1
(L)
(L)
(L)
(L)
W2,l,j , C2,l,j
= Wl/(l−1),j + W1,l,j ,
(42)
(L)
(L)
(L)
(L)
(L)
(L)
w2,l,j , W2,l,j η2,l,j = wl/(l−1),j + w1,l,j ,
Wl/(l−1),j , (Cl/(l−1),j )−1 and wl/(l−1),j , Wl/(l−1),j ηl/(l−1),j .
(N )
3. Time update for xl - This step aims at generating the j-th particle for
and its associated weight (this corresponds to step 3-b) of algorithm 1 in
(N )
[15, Sec. II]); these information are conveyed by the message m
~ 5,j (xl+1 ), which
can be expressed as (see Fig. 3)
RR
(N )
(N )
(L)
(N )
f xl+1 xl , xl
m
~ 5,j xl+1 =
(43)
(N )
(L)
(L)
(N )
·m
~ 2,j xl
m
~ 2,j xl
dxl dxl .
(N )
xl+1
The double integral appearing in the RHS of the last equation can be evaluated
(N )
as follows. First of all, substituting m
~ 2,j (xl ) (32) in (43) yields
Z
(L)
(L)
(N )
(N )
(L)
(N )
dxl .
(44)
~ 2,j xl
m
~ 5,j xl+1 = f xl+1 xl , xl/l,j m
(N )
(N )
(L)
(L)
Then, substituting the expression of f (xl+1 |xl , xl ) (see (20)) and m
~ 2,j (xl )
(40) in (44) yields, after some manipulation, the Gaussian message (see [32, Par.
2.3.3, eq. (2.115)])
(N )
(N ) (N )
(N )
(45)
m
~ 5,j xl+1 = N xl+1 ; η5,l,j , C5,l,j ,
where
(N )
(N )
C5,l,j
(N )
(N )
Al,j , Al
(N )
(N ) (L)
(N )
η5,l,j , Al,j η2,l,j + fl,j ,
T
(N ) (L)
(N )
)
, C(N
+
A
C
A
,
w
l,j
2,l,j
l,j
(N )
(N )
(46)
(47)
(N )
(xl/l,j ) and fl,j , fl (xl/l,j ). Note that, in principle,
(N )
(N )
~ 5,j xl+1 ,
m
~ out,j xl+1 = m
(48)
18
G. M. Vitetta et al., Marginalized Particle Filtering . . .
as it can be easily inferred from Fig. 3. However, as already mentioned above,
(j)
(N )
in MPF the output message m
~ out (xl+1 ) is required to have the same functional
(j) (N )
form as m
~ in (xl ) (22). This result can be achieved a) sampling the Gaussian
(N ) (N )
(N )
(N )
function N (xl+1 ; η5,l,j , C5,l,j ) (see (45)), that is drawing a sample x(l+1)/l,j from
(N )
it and b) assigning to the sample x(l+1)/l,j a probability w(l+1)/l,j equal to the
weight wl/l,j = 1/Np (originating from resampling). It is worth pointing out
(N )
that: 1) the particles {x(l+1)/l,j } form the new set S(l+1)/l ; 2) in accomplishing
step a) of this procedure, it may be useful to introduce artificial noise (this can
be simply done adding the same positive quantity to the diagonal elements of
(N )
the matrix Cw appearing in the RHS of (47)) in order to mitigate the so called
(N )
degeneracy problem [1, 34]. If this approach is adopted, the message m
~ 5,j (xl+1 )
(45) is replaced by
(N )
(N )
(N )
(49)
m
~ 5,j xl+1 = δ xl+1 − x(l+1)/l,j ,
(N )
which emerges from the graph as m
~ out,j (xl+1 ). This message is also used in the
(L)
time update for xl , as illustrated in the next step.
(L)
4. Time update for xl - This step aims at generating a new Gaussian
(L)
(N )
~ 5,j (xl+1 ) =
pdf associated with the j-th particle x(l+1)/l,j and conveyed by m
(L)
m
~ out,j (xl+1 ) (this corresponds to step 3-c) of algorithm 1 in [15, Sec. II]).
However, before doing that, a further measurement update is accomplished on
(L)
the basis of the pseudo-measurement zl (13). This involves the evaluation of
(L)
the messages m
~ j (zl ),
Z
(L)
(L)
(L)
(N )
(L)
(L)
f zl xl , xl
dzl
(50)
= m
~ j zl
m
~ 3,j xl
and
(L)
(L)
(L)
m
~ 4,j xl
=m
~ 2,j xl
m
~ 3,j xl
,
(51)
(L)
as shown in Fig. 3. Generally speaking, the message m
~ j (zl ) can be expressed
as
RR
(L)
(L)
(N )
(N )
=
f zl xl , xl+1
m
~ j zl
(52)
(N )
(N )
(N )
(N )
(N )
·f xl+1 xl
f xl
dxl dxl+1 .
(N )
(N )
However, since in this case f (xl
) = δ(xl
(N )
x(l+1)/l,j )
(N )
m
~ 2,j (xl )
(N )
(N )
(L)
(N +1)
−xl/l,j ), f (xl+1 |xl ) = δ(xl
(N )
m
~ 5,j (xl+1 )
−
can be assumed (see
(32) and
(49), respectively), eq. (52) easily leads to the expression
(L)
(L)
(N )
(L)
(L)
(N )
(53)
= f zl xl/l,j , x(l+1)/l,j = δ zl − zl,j ,
m
~ j zl
19
G. M. Vitetta et al., Marginalized Particle Filtering . . .
where
(L)
(N )
(N )
zl,j , x(l+1)/l,j − fl,j .
(54)
(L)
(L)
(N )
Then, substituting (53) and the expression of f (zl |xl , xl
(50) yields
(L)
(L)
(N ) (L)
)
.
= N zl,j ; Al,j xl , C(N
m
~ 3,j xl
w
) (see (16)) in
Finally, substituting the last expression and (40) in (51) produces
(L)
(L) (L)
(L)
= N xl ; η2,l,j , C2,l,j
m
~ 4,j xl
(L)
(N ) (L)
(N )
·N zl,j ; Al,j xl , Cw
,
(55)
(56)
which, after some manipulation (in which unrelevant scale factors are dropped),
can be rewritten as
(L)
(L) (L)
(L)
= N xl ; η4,l,j , C4,l,j ,
(57)
m
~ 4,j xl
where
T
(L)
(L) (L)
(L)
(N )
(N ) (L)
w4,l,j , W4,l,j η4,l,j = w2,l,j + Al,j
Ww
zl,j ,
(N )
T
−1
(L)
(N )
(L)
(L)
(N ) (N )
Ww
Al,j
= W2,l,j + Al,j
W4,l,j , C4,l,j
(58)
(59)
(N )
and Ww , [Cw ]−1 .
(L)
The last part of the time update step for xl requires the evaluation of the
output message
RR
(L)
(L)
(L)
(N )
f xl+1 xl , xl
m
~ 5,j xl+1 =
(60)
(N )
(L)
(N )
(L)
dxl dxl ,
m
~ 2,j xl
·m
~ 4,j xl
(N )
which, similarly as m
~ 5,j (xl+1 ) (43), requires double integration. Substituting
(N )
m
~ 2,j (xl ) (32) in the RHS of the last expression yields
Z
(L)
(L)
(L)
(L)
(L)
(N )
dxl .
(61)
~ 4,j xl
f xl+1 xl , xl/l,j m
m
~ 5,j xl+1 =
(L)
(L)
(N )
Then, substituting the expression of f (xl+1 |xl , xl/(l−1),j ) (see (17)) and (57)
in the last equation gives (see [32, Par. 2.3.3, eq. (2.115)])
(L)
(L)
(L)
(L)
m
~ 5,j xl+1 = N xl+1 ; η5,l,j , C5,l,j
(62)
(L)
=m
~ out,j xl+1 ,
where
(L)
(L) (L)
(L)
(L)
η5,l,j , Al,j η4,l,j + fl,j = η(l+1)/l,j ,
(63)
20
G. M. Vitetta et al., Marginalized Particle Filtering . . .
(L)
(L)
T
(L)
(L) (L)
(L)
(L)
C5,l,j , C(L)
+
A
C
A
= C(l+1)/l,j ,
w
l,j
4,l,j
l,j
(N )
(L)
(L)
(64)
(L)
(N )
~ out,j (xl+1 ) (62)
fl,j , fl (xl/l,j ) and Al,j , Al (xl/l,j ). The evaluation of m
concludes the MPF message passing procedure, which needs to be carried out
for each of the Np particles available at the beginning of the l-th recursion.
Note that this procedure needs a proper inizialization (this corresponds to step
1) of algorithm 1 in [15, Sec. II]). In practice, before starting the first recursion
(N )
(N )
(corresponding to l = 1), the set S1/0 = {x1/0,j , j = 0, 1, ..., Np − 1}, consisting
(N )
of Np particles, is generated for x1
sampling the pdf
Z
(L)
(N )
= f (x1 ) dx1 ,
f x1
(65)
(L)
(L)
(L)
and the same weight w1/0 = 1/Np and Gaussian model N (x1 ; η1/0 , C1/0 ) for
(L)
are assigned to each of them.
Finally, it is worth pointing out that: 1) the processing accomplished in the
(L)
measurement and time update for xl can be interpreted as a form of Kalman
filtering, in which both the real measurement yl and the pseudo-measurement
(L)
(N )
(L)
zl are processed [15]; 2) in the l-th recursion estimates of xl and xl can
P
Np −1
(w)
(L)
(N )
(N )
(see (28)) and as x̂l =
be evaluated as x̂l
=
j=0 wl,j xl/(l−1),j /Pl
PNp −1 (L)
j=0 η4,l,j /Np (see (58)), respectively; 3) the result expressed by eq. (45)
shows that, generally speaking, the statistical representation generated by the
(N )
SPA for the state xl+1 is a Gaussian mixture (GM), whose Np components have
the same weight (equal to 1/Np ) because resampling is always used in step 1.
The last point leads to the conclusion that, if resampling was not accomplished
in the l-th recursion, the weight of the j-th component of this GM would be
proportional by wl,j (29); this would unavoidably raise the problem of sampling
(N )
a GM with unequally weighted components in generating the particle set Sl/(l+1)
x1
(L)
and that of properly handling the resulting pseudo-measurements {zl,j }. These
considerations motivate the use of resampling in each recursion, indipendently
(N )
of the effective sample size [1] characterizing the particle set Sl/(l−1) .
5
Simplifying Marginalized Particle Filtering
The MPF derivation illustrated in the last two Sections unveils the real nature
of MPF and its limitations, and shows the inner structure of the processing
accomplished within each step. For these reasons, it paves the way for the development of new filtering methods related to MPF. In this Section we exploit
our FG-based representation of Bayesian filtering to develop reduced complexity
G. M. Vitetta et al., Marginalized Particle Filtering . . .
21
alternatives to MPF by simplifying the message passing derived in the previous
Section. It is worth mentioning that some methods for reducing MPF computational complexity [16] have been already proposed in the technical literature
[4], [17], [18]. In particular, the method proposed in [4] and [17] is based on rep(N )
resenting the particle set for xl as a single particle (that corresponds to the
center of mass of the set itself), so that a single Kalman filter is employed in up(L)
dating the statistics of the linear component xl ; consequently, the statistical
(L)
knowledge about xl is condensed in the single message
(L)
(L)
(L) (L)
(66)
= N xl ; ηl/(l.−1) , Cl/(l−1) ,
m
~ in xl
(L)
instead of the Np messages {m
~ in,j (xl )} (21). Unluckily, this simplified MPF
(N )
algorithm works well only if the posterior distribution of xl is unimodal. Its
(N )
generalization to the case in which the posterior distribution of xl
is multimodal has been illustrated later in [18]. In the proposed technique the particles
are partitioned into Kl groups or clusters (the parameter Kl is required to equal
(N )
the number of modes of the posterior density of xl ) and each group is represented by a single particle that corresponds to its center of mass; this allows
to reduce the overall number of Kalman filters from Np to Kl . The implementation of this technique requires, however, solving the following two specific
problems: a) identifying the number of modes of the posterior distribution of
(N )
xl ; b) partitioning the particles into clusters according to a grouping method
in each recursion. Unluckily, practical solutions for suche problems have not
been proposed in [18].
Our derivation of simplified algorithms has been only partly inspired by the
manuscripts cited above. In fact, first of all, it relies on the following specific
methods: a) a set of Np equal weight particles {xj ; j = 0, 1, ..., Np − 1} is
represented through its center of mass
x̄ ,
Np −1
1 X
xj ,
Np j=0
(67)
as already suggested in [17] and [18], when the computation of a message refer(L)
(N )
ring to xl involves the particle-based representation of xl ; b) a set of Np
Gaussian messages {N (x; ηj , Cj ); j = 0, 1, ..., Np − 1}, that refer to a set of Np
equal weight particles, is represented as the Np −component GM
Np −1
1 X
N (x; ηj , Cj ) ,
fGM (x) ,
Np j=0
(68)
22
G. M. Vitetta et al., Marginalized Particle Filtering . . .
and this GM is approximated through its projection onto the Gaussian pdf
fG (x) = N (x; ηG , CG ), where ηG and CG are selected as
ηG ,
Np −1
1 X
ηj
Np j=0
(69)
and
PNp −1
CG = (1/Np ) j=0
Cj
PNp −1
T
T
−ηG (ηG ) + (1/Np ) j=0 ηj (ηj )
(70)
respectively, so that the mean and covariance of fGM (x) (68) are preserved
(L)
(e.g., see [33, Sec. IV]). Secondly, as far as the messages {m
~ in,j (xl )} (21)
are concerned, we do not adopt the approximations proposed in [4], [17] and
[18]. In fact, we focus on the following two cases: case #1 - the messages
(L)
{m
~ in,j (xl )} are all different but, when needed in message passing, are con(L)
(L)
densed in the single message (66), where ηl/(l.−1) and Cl/(l−1) are given by (69)
(L)
(L)
and (70), respectively, with ηj = ηl/(l.−1),j and Cj = Cl/(l−1),j ; b) case #2 (L)
the messages {m
~ in,j (xl )} have different means, but their covariance matrices
(L)
(L)
{Cl/(l−1),j } are all equal (their common value is denoted C̃l/(l−1) in the following). In both cases our simplifications aim at minimizing the overall number of
a) Cholesky decompositions for the generation of the new particle set S(l+1)/l
(L)
(such decompositions are required for the Np matrices {C5,l,j } (47)) and b)
matrix inversions; such inversions are needed to compute: a) the Np matrices
(N )
(N )
{W1,l,j , (C1,l,j )−1 } (required in the evaluation of the weights {wl,j } on the
(L)
basis of (29)); b) the Np matrices {Wl/(l−1),j } (required to evaluate the vec(L)
(L)
(L)
tors {w2,l,j } (41) and the matrices {W2,l,j } (42)); c) the Np matrices {C2,l,j }
(L)
(employed in (47)); c) the Np matrices {C4,l,j } (employed in (64)).
Based on the methods illustrated above, our simplified versions of MPF are
(N )
derived as follows. First of all, in the measurement update for xl , we use a
single covariance matrix in the Gaussian pdf appearing in the RHS of (29); in
other words, the j-th weigth wl,j is computed as
(N )
(N )
wl,j , N yl ; η1,l,j , C1,l ,
(71)
(N )
(N )
where C1,l is evaluated on the basis of (70) (see also (69)), setting ηj = η1,l,j
(N )
(30) and Cj = C1,l,j (31) for any j.
(L)
(N )
Second, in the measurement update for xl , the particle set Sl/(l−1) is con(N )
densed in its center of mass x̄l/(l−1) (see (67)). Consequently, the message
23
G. M. Vitetta et al., Marginalized Particle Filtering . . .
(L)
m
~ 1,j (xl ) (35) is replaced by its particle-independent form
(L)
(L)
= N yl ; B̄l xl + h̄l , Ce
m
~ 1 xl
(72)
(N )
(N )
where B̄l , Bl (x̄l/(l−1) ) and h̄l , hl (x̄l/(l−1) ). This message, similarly as (35),
can be put in the Gaussian form (see (36)-(38))
(L)
(L) (L)
(L)
= N xl ; η1,l , C1,l ,
(73)
m
~ 1 xl
where
(L)
(L) (L)
w1,l , W1,l η1,l = B̄Tl We yl − h̄l
and
(74)
−1
(L)
(L)
W1,l , C1,l
= B̄Tl We B̄.
(75)
(L)
This allows us to replace the message m
~ 2,j (xl ) (40) with its particle-independent
counterpart
(L)
(L) (L)
(L)
= N xl ; η2,l , C2,l ,
(76)
m
~ 2 xl
(L)
(L) (L)
(L)
(L)
(L)
(L)
where w2,l , W2,l η2,l and W2,l , (C2,l )−1 are easily obtained from (41) and
(L)
(L)
(42) replacing a) w1,l,j and W1,l,j with w1,l (74) and W1,l (75), respectively;
(L)
(L)
(L)
(L)
(L)
(L)
b) wl/(l−1),j and Wl/(l−1),j with wl/(l−1) , Wl/(l−1) ηl/(l−1) and Wl/(l−1) ,
(L)
(L)
(L)
(Cl/(l−1) )−1 (ηl/(l.−1) and Cl/(l−1) are given by (69) and (70), respectively, with
(L)
(L)
(L)
ηj = ηl/(l.−1),j and Cj = Cl/(l−1),j ). Note that, since the precision matrix W2,l
(L)
is particle-independent, a single matrix C2,l has to be computed for the next
step.
(N )
(L)
Thirdly, in the time update for xl , the message m
~ 2 (xl ) (76) can be used
(L)
(N )
in place of m
~ 2,j (xl ) (40) in the evaluation of m
~ 5,j (xl+1 ) (see (44) and Fig. 3).
This leads to the particle-dependent message
(N )
(N ) (N )
(N )
m
~ 5,j xl+1 = N xl+1 ; η5,l,j , C5,l,j ,
(77)
(N )
(N )
where η5,l,j and C5,l,j are obtained from (46) and (47), respectively, replac(L)
(L)
(L)
(L)
ing η2,l,j and C2,l,j with η2,l and C2,l , respectively. Then, to simplify the
(N )
(N )
generation of the new particle set S(l+1)/l , the covariance matrices {C5,l,j } are
(N )
(N )
condensed in a single matrix C5,l using (70) (see also (69)) with ηj = η5,l,j =
(N ) (L)
(N )
Al,j η2,l + fl,j
(N )
(N )
and Cj = C5,l,j = Cw
(N )
(N )
(L)
(N )
+ Al,j C2,l (Al,j )T ; consequently,
the particle generation mechanism for xl+1 requires computing the Cholesky de(N )
composition of a single matrix (namely, C5,l ), since it is based on the modified
message
(N )
(N ) (N )
(N )
(78)
m
~ 5 xl+1 = N xl+1 ; η5,l,j , C5,l ,
24
G. M. Vitetta et al., Marginalized Particle Filtering . . .
(N )
which depends on the particle index j through the mean vector η5,l,j only.
(L)
Finally, in the time update for xl , the pdf
(L)
(L)
(L)
(N ) (L)
)
f zl xl
= N zl ; Āl xl , C(N
w
(L)
(79)
(N )
is employed in the evaluation of m
~ 3,j (xl ) through (50), where Āl
(N )
= Al
(N )
(x̄l/l )
(N )
(N )
and x̄l/l denotes the center of mass of the particle set Sl/l (see (67)). Then,
(L)
(L)
the messages m
~ 3,j (xl ) (55) and m
~ 4,j (xl ) (57) can be replaced by
(L)
(L)
(N ) (L)
)
= N zl,j ; Āl xl , C(N
m
~ 3,j xl
w
(80)
and
(L)
(L) (L)
(L)
= N xl ; η4,l,j , C4,l ,
m
~ 4,j xl
(81)
respectively, with (see (58) and (59))
and
T
(L)
(L) (L)
(L)
(N )
(N ) (L)
w4,l,j , W4,l η4,l,j = w2,l + Āl
Ww
zl,j
(82)
−1
T
(L)
(L)
(L)
(N )
(N ) (N )
W4,l , C4,l
= W2,l + Āl
Ww
Āl .
(83)
(L)
(L)
(N )
Substituting now m
~ 4,j (xl ) (81) (in place of m
~ 4,j (xl ) (57)) and m
~ 2,j (xl )
(32) in (60), produces, after some manipulation
(L)
(L)
(L)
(L)
(84)
m
~ 5,j xl+1 = N xl+1 ; η5,l,j , C5,l,j ,
where
(L)
(L) (L)
(L)
η5,l,j , Al,j η4,l,j + fl,j
and
(L)
(L)
(L)
C5,l,j , C(L)
w + Al,j C4,l
(L)
(L)
Al,j
(85)
T
.
(86)
Note that: a) since the precision matrix W4,l (83) is particle-independent, a
(L)
single matrix inversion is needed to evaluate the matrix C4,l appearing in (86);
(L)
(L)
b) in case # 2 the matrix set {C5,l,j } is condensed in a single matrix C5,l (this
(L)
(L)
represents the common value C̃(l+1)/l taken on by all the matrices {C(l+1)/l,j }
(L)
processed in the next recursion); c) the matrix C5,l is evaluated on the basis of
(L)
(L)
(69) and (70), setting ηj = η5,l,j and Cj = C5,l,j for any j.
The new filtering techniques, based on MPF and on the simplified messages
derived above, are called simplified MPF (SMPF) in the following; in particular
the acronyms SMPF1 and SMPF2 are used to refer to case #1 and case #2,
respectively.
25
G. M. Vitetta et al., Marginalized Particle Filtering . . .
6
Message Passing in Iterative Filtering Techniques Inspired by Marginalized Particle Filtering
As already mentioned above, the suboptimality of MPF can related to the fact
that the FG underlying the considered filtering problem is not cycle free. It is
well known that the SPA can also be applied to a factor graph with cycles simply
by following the same message propagation rules; however, generally speaking,
this leads to an “iterative” algorithm with no natural termination (and known
as loopy belief propagation), since its messages are passed multiple times on a
given edge [20], [21]. Despite this, some of the most relevant applications of
the SPA have been developed for systems in which the underlying FG does
have cycles, like the one shown in Fig. 1. In the following we show how a
novel iterative technique can be developed for our filtering problem following
this approach. To begin, we note that our interest in iterative methods is also
(N )
motivated by the possibility of exploiting the pseudo-measurement zl (18); in
(N )
fact, the message mj (zl ) referring to this random vector cannot be computed
in MPF because of the adopted scheduling, but can certainly provide additional
information to refine our statistical knowledge about the nonlinear component
(N )
(L)
xl . This message, similarly as the one referring to10 zl , can be put in a
Gaussian form, that is
(N )
(N ) (N )
(N )
= N zl ; ηz,l,j , Cz,l,j ,
(87)
m
~ j zl
(L)
(L)
(N )
since xl+1 and xl , conditioned on xl , are modelled as jointly Gaussian
random vectors. Let us show now how this message can be computed in an
iterative filtering procedure generalising MPF and how it can exploited in such
(L)
a procedure. First of all, we assume that the message m
~ 5,j (xl+1 ) (62), repre(L)
(N )
(N )
senting the pdf of xl+1 conditioned on xl = xl/l,j , is already available when
(L)
the time update for xl (i.e., the first step of MPF) is accomplished. Then,
(L)
(L)
(N )
given m
~ 2,j (xl ) (40) and m
~ 5,j (xl+1 ) (62), the mean and covariance of zl can
be evaluated as (see (18))
(N )
(L)
(L) (L)
ηz,l,j = η5,l,j − Al,j η2,l,j
10 If
(N)
xl
(N)
(L)
= xl/l,j , the random vector zl
(N)
(N)
(88)
(N)
(13) becomes xl+1 − fl
(N)
(xl/l,j ); consequently,
(L)
adopting the Gaussian model (45) for xl+1 results in a Gaussian model for zl
too.
26
G. M. Vitetta et al., Marginalized Particle Filtering . . .
and
T
(N )
(L)
(L) (L)
(L)
Cz,l,j = C5,l,j + Al,j C2,l,j Al,j
T
T
(L) (L)
(L)
(L)
,
−Al,j Cx,l,j − Cx,l,j
Al,j
(89)
(L)
respectively, where Cx,l,j denotes the cross covariance matrix for the vectors
(L)
xl
(L)
(N )
and xl+1 (conditioned on xl
(L)
(L)
(L)
(N )
~ 2,j (xl ) (40) and the
= xl/l,j ). Given m
(L)
(N )
(L)
(L) (L)
(L)
conditional pdf f (xl+1 |xl , xl/l,j ) = N (xl+1 ; fl,j + Al,j xl , Cw ) (see (17)),
(L)
(L)
(L)
it is easy to show that Cx,l,j = C2,l,j (Al,j )T (e.g., see [32, Par. 2.3.3, eq.
(2.104)]); consequently, eq. (89) can be rewritten as
T
(N )
(L)
(L) (L)
(L)
Cz,l,j = C5,l,j − Al,j C2,l,j Al,j
.
(90)
Equations (88) and (90) represent the desired result, since they provide a com(N )
plete statistical characterization of the message m
~ j (zl ) (87). In principle,
(L)
this message could be exploited in a similar way as that adopted for m
~ j (zl )
(N )
(53); this approach would lead to draw a set of Np samples {zl,j } from the
Gaussian function appearing in the RHS of (87) and to process the resulting
pseudo-measurements to generate a new weight for each particle of the set Sl/l .
However, our computer simulations have shown that this approach is outperformed by more refined method illustrated in the following. In practice, in the
(N )
iterative filtering method we propose the message m
~ j (zl ) (87), similarly as
(L)
m
~ j (zl ) (53), is employed to evaluate the new message11
Z
(N )
(N )
(N )
(N )
(N )
(91)
f zl
xl/l,j dzl ,
= m
~ j zl
m
~ 3,j xl
(N )
(L)
which represents for xl the counterpart of the message m
~ 3,j (xl ) (50). Sub(N ) (N )
(N )
(N )
stituting (87) and the expression of f (zl |xl ) (given xl
= xl/l,j ) in the
RHS of the last expression gives12 (see the Appendix)
(N )
m
~ 3,j xl
T
(N )
(N )
(N ) (N )
= D3,l,j · exp 21
η3,l,j W3,l,j η3,l,j
(92)
T
T
(N )
(N ) (N )
(L)
(L) (L)
− ηz,l,j Wz,l,j ηz,l,j − fl,j
Ww fl,j
, pl,j ,
11 Note
(N)
that the following message represents the correlation between the pdf m
~ j (zl
evaluated on the basis of the definition of
(N)
zl
)
(18) and the pdf originating from the fact that
(L)
(N)
(L)
this quantity is expected to equal the random vector fl (xl/l,j ) + wl . For this reason, it
expresses the degree of similarity between these two functions.
12 In our computer simulations the factor D (N) appearing in this weight has been always
3,l,j
neglected, since it negligibly influences estimation accuracy.
27
G. M. Vitetta et al., Marginalized Particle Filtering . . .
where
−1
(N )
(N )
(N )
(L)
W3,l,j , C3,l,j
= Wz,l,j + Ww
,
(N )
(N )
(N )
(N )
(93)
(L)
(L)
w3,l,j , W3,l,j η3,l,j = wz,l,j + Ww
fl,j ,
(N )
(N )
(N )
(N )
(N )
(L)
(94)
(L)
Wz,l,j , (Cz,l,j )−1 , wz,l,j , Wz,l,j ηz,l,j , Ww , [Cw ]−1 ,
(N )
(N )
(L)
h
i−DL /2
(N )
(N )
D3,l,j , det C̃l,j
(95)
(N )
and C̃l,j , Cz,l,j + Cw . Then, the new message m
~ 3,j (xl
similarly as
) (92) is exploited,
(L)
m
~ 3,j (xl )
(N )
(55), to generate the message (see (51))
(N )
(N )
(N )
,
m
~ 3,j xl
=m
~ 2,j xl
m
~ 4,j xl
(96)
where m
~ 2,j (xl ) is expressed by (28) (i.e., it is the message emerging from the
(N )
measurement update for xl in the absence of resampling). This produces (see
(92))
(N )
(N )
(N )
(97)
= Wl,j δ xl − xl/(l−1),j ,
m
~ 4,j xl
where
Wl,j , wl,j · pl,j
(98)
represents the new weight for the j-th particle; such a weight accounts for both
(N )
the (real) measurement yl and the pseudo-measurement zl,j . Resampling with
(N )
(N )
replacement can now be accomplished for the set Sl/(l−1) , {xl/(l−1),j } on
the basis of the more refined weights {Wl,j } (98); if this is done, the message
(N )
(N )
m
~ 4,j (xl ) takes on the same form as m
~ 2,j (xl ) (32), i.e. it can be expressed
as
(N )
(N )
(N )
(99)
= δ xl − x̃l/l,j ,
m
~ 4,j xl
(N )
with j = 0, 1, ..., Np − 1. Note that, generally speaking, the particle set S̃l/l ,
(N )
(N )
(N )
{x̃l/l,j } produced by resampling in this case is different from Sl/l , {xl/l,j }
(i.e., from the one obtained with MPF), even if both of them originate from the
(N )
same set Sl/(l−1) ; this is due to the fact that the weights {Wl,j } (98) may be
substantially different from the MPF weights {wl,j } because of the factor pl,j .
(N )
(N )
Finally, the new message m
~ 4,j (xl ) (99) is used in place of m
~ 2,j (xl ) in the
(L)
RHS of (43) for the evaluation of the message m
~ 5,j (xl+1 ).
As already state above, our previous derivations rely on the assumption that
(L)
(L)
the message m
~ 5,j (xl+1 ) (62) is available when the time update for xl is carried
out; unluckily, this message becomes available only in the last step of MPF.
However, if MPF is generalised in a way that Nit > 1 iterations (i.e., message
28
G. M. Vitetta et al., Marginalized Particle Filtering . . .
passes) are carried out within the same recursion, in the k-th iteration (with k =
(N )
2, 3, ..., Nit ) the message mj (zl ) (87) can be really evaluated exploiting the
(L)
message m
~ 5,j (xl+1 ) computed in the previous iteration. These considerations
lead, in a natural fashion, to the development of the message passing illustrated
in Fig. 4, which describes the message flow occurring in the k-th iteration
of a new filtering technique, generalising MPF and called turbo filtering (TF)
in the following; note that the superscripts (k) or (k − 1) have been added
to all the messages flowing in the considered graph to identify the iteration
in which they are generated and that the grey circle appearing in the figure
represents a unit delay cell. The processing tasks accomplished by TF can be
summarized as follows. The first part of this technique can be considered as a
(N )
(L)
form of initialization, in which the messages {m
~ 2,j (xl )} and {m
~ 2,j (xl )} are
computed; however, unlike MPF, resampling is not accomplished in the last part
(N )
(N )
of the time update for xl
(so that the messages {m
~ 2,j (xl )} are expressed
(N )
by (28) instead of (32) and refer to the particle set Sl/(l−1) ). Moreover, as
(N )
(L)
shown in Fig. 4, the messages {m
~ 2,j (xl )} and {m
~ 2,j (xl )} emerging from
the first part of TF remain unchanged within the l-th recursion, since they
(N )
(L)
represent the a priori information available about xl and xl , respectively;
consequently, like in any turbo processing method, these information are made
available to all the iterations carried out within each recursion. In the second
part of TF Nit -iterations are accomplished with the aim of progressively refining
(N )
the set S(l+1)/l (the version of this set generated in the k-th iteration is denoted
(k)
(N )
(L)
~ 5,j (xl+1 )}. To achieve
S(l+1)/l [k]) and the associated Gaussian messages {m
these result, in the k-th iteration (with k = 1, 2, ..., Nit ) the ordered computation
(k) (N )
(k) (N )
~ 3,j (xl ) (92)
of the following messages is accomplished: m
~ j (zl ) (87), m
(k)
(N )
(conveying the weight pl,j [k]), m
~ 4,j (xl
(k)
(N )
(L)
(k)
(L)
) (97) (conveying the weight Wl,j [k]),
(k)
(L)
(k)
(L)
~ 5,j (xl+1 )
~ 4,j (xl ) (57) and m
~ j (zl ) (53), m
~ 3,j (xl ) (55), m
m
~ 5,j (xl+1 ) (45), m
13
(62). Moreover, in the k-th iteration resampling is accomplished on the basis of
the particle weights {Wl,j [k]}; generally speaking, this results in a new particle
(N )
(N )
set denoted Sl/l [k] = {xl/(l−1),j [k], j = 0, 1, ..., Np − 1} (which is always a
(N )
subset of Sl/(l−1) ). It is also important to point out that:
(0)
(N )
• In the first iteration (i.e, for k = 1) the messages {m
~ 5,j (xl+1 )} are un(L)
13 Note that, after carrying out resampling, the set of messages {m
~ 2,j (xl )} needs to be
properly reordered, since the messages associated with the discarded particles are not preserved. This modifies the set of particles available in the next iteration and, consequently,
the set of weights {wl,j } associated with them (these weights need to be renormalized after
any change). In the following the notation {wl,j [k]} is adopted to denote the set of weights
employed in the k-th iteration for the evaluation of the overall weights Wl,j [k] according to
(98).
29
G. M. Vitetta et al., Marginalized Particle Filtering . . .
(1)
(N )
(1)
(N )
~ 3,j (xl ) =
defined, so that {m
~ j (zl ) = 1} (and, consequently, {m
1}) must be assumed; moreover, resampling is not accomplished (i.e.,
(N )
(N )
(N )
Sl/l [1] = Sl/l = Sl/(l−1) ), since the weights pl,j appearing in the overall
weight Wl,j (98) become available in the following iterations.
− (x(L) ) = ←
− (x(N ) ) = 1 is assumed for
• In each iteration the equality ←
m
m
he l+1
he l+1
(L)
the two messages entering the FG along the half edges associated with xl+1
(N )
and xl+1 (see Fig. 4), since no information comes from the next recursion.
For this reason, at the end of the last iteration (i.e., for k = Nit ), the
output messages (i.e., the input messages feeding the (l + 1)-th recursion)
are evaluated as (see (48)-(49) and (62))
(N )
(N )
(Nit )
(N )
(N )
− (x(N ) ) = m
x
(100)
m
~
~ 5,jit xl+1 ←
m
~ out,j xl+1 = m
he l+1
5,j
l+1
and
(L)
(L)
(N )
(L)
(N )
− (x(L) ) = m
m
~ 5,jit xl+1 ,
~ 5,jit xl+1 ←
m
~ out,j xl+1 = m
he l+1
(101)
with j = 0, 1, ..., Np − 1.
Another relevant issue concerns the interpretation of the processing tasks
accomplished in the TF technique. In fact, our derivations show that, at the
end of the k-th iteration, the a posteriori statistical information about the j-th
(k) (N )
(N )
(N )
~ 4,j (xl ) (97), which
particle xl/l,j [k] of Sl/l [k] is provided by the message m
conveys the weight (see (98))
(a)
Wl,j [k] = wl,j [k] · pl,j [k] · wl,j ,
(102)
(a)
where wl,j denotes the a priori information available at the beginning of the l(a)
th recursion for the j-th particle ( in our derivation wl,j = 1 has been assumed,
(a)
in place of wl,j = 1/Np , to simplify the notation; see (22)), pl,j [k] is the weight
(N )
(k)
(N )
originating from m
~ j (zl ) (87) and conveyed by m
~ 3,j (xl ) (92), and wl,j [k] is
the weight computed on the basis of the available measurement yl . Taking the
natural logarithm of both sides of (102) produces
(a)
(y)
(z)
Ll,j [k] = Ll,j + Ll,j [k] + Ll,j [k],
(y)
(z)
(103)
(a)
where Ll,j [k] , ln(Wl,j [k]), Ll,j [k] , ln(wl,j [k]), Ll,j , ln(pl,j [k]) and Ll,j ,
(a)
ln(wl,j ). The last equation has exactly the same structure as the well known
formula (see [35, Sec. 10.5, p. 450, eq. (19.15)] or [36, Par. II.C, p. 432, eq.
(20)])
L (uj |y) = L (uj ) + Lc (yj ) + Le (uj )
(104)
30
G. M. Vitetta et al., Marginalized Particle Filtering . . .
min , j ( x (l L ) )
x (l L )
yl
m1, j ( x (l L ) )
f y ,/ x( L ) ,x( N )
l +1
=
=
l
m2, j ( x l( L ) )
=
l
m4,( k j) ( x (l L ) )
mhe ( xl(+L1) ) = 1
f x ( L ) / x ( L ) ,x ( N )
m2,( k j) ( x (l L ) )
min , j ( x (l L ) )
=
l
m2, j ( x l( L ) )
m5,( kj) ( x (l +L1) )
f z( L ) / x ( L )
m (jk ) ( z l( L ) )
mout , j ( x l(+L1) )
l
=
m3,( kj) ( x (l L ) )
l
min , j ( x (l N ) )
l
x (l +L1)
l
m2, j ( x l( N ) )
z (l L )
x (l N ) , x l( +N1) → z l( L )
m5,( kj) ( x (l +N1) )
m2, j ( x (l N ) )
m2, j ( x l( N ) )
=
m2, j ( x l( L ) )
=
m2, j ( x
(N )
l
)
m (jk ) ( z l( N ) )
f y ,/ x( L ) ,x( N )
l
m3,( kj) ( x l( N ) )
l
=
=
min , j ( x (l N ) )
=
( k −1) ( L )
z (l N ) m
( xl +1 )
5, j
f z( N ) / x ( N )
l
l
yl
min , j ( x (l N ) )
m2, j ( x l( N ) )
m2,( k j) ( x l( L ) )
m5,( kj) ( x l(+N1) )
=
=
m2, j ( x l( N ) )
x (l N )
m5,( kj) ( x (l +L1) )
x (l L ) , x (l +L1) → z (l N )
min , j ( x (l L ) )
l
m2, j ( x l( L ) )
m4,( k j) ( x l( N ) )
f x ( N ) / x ( N ) ,x ( L )
l +1
l
l
mhe ( x l(+N1) ) = 1
mout , j ( x (l +N1) )
x
(N )
l +1
Figure 4: Message passing over the FG of Fig. 1 for the proposed TF technique.
All the quantities appearing in this figure refer to the k-th iteration of the l-th
recursion; the messages available at the beginning of the considered iteration are
indicated by red arrows, those entering the graph by blue arrows, those leaving
it at the end of the last iteration by black arrows and those computed within
the considered iteration by green arrows.
31
G. M. Vitetta et al., Marginalized Particle Filtering . . .
expressing of the log-likelihood ratio (LLR) available for the j-th information
bit uj at the output of a soft-input soft-output channel decoder operating over
an additive white Gaussian noise (AWGN) channel and fed by: a) the channel
output vector y (whose j-th element yj is generated by the communications
channel in response to a channel symbol conveying uj and is processed to produce the so-called channel LLR Lc (yj )); b) the a priori LLR L (uj ) about uj ;
c) the extrinsic LLR Le (uj ), i.e. a form of soft information available about
uj , but intrinsically not influenced by such a bit (in turbo decoding of concatenated channel codes extrinsic infomation is generated by another channel
decoder with which soft information is exchanged with the aim of progressively
refining data estimates). This correspondence is not only formal, since in eqs.
(103) and (104) terms playing similar roles can be easily identified. For instance,
(y)
(a)
the term Ll,j [k] (Ll,j ) in (103) provides the same kind of information as Lc (yj )
(L (uj )), since these are both related to the noisy data (a priori information)
available about the quantities to be estimated (the system state in one case, an
(z)
information bit the in the other one). What about the term Ll,j [k] appearing
in the RHS of (103)? The link we have established between (103) and (104)
unavoidably leads to the conclusion that such a term should represent the counterpart of the quantity Le (uj ) appearing in (104), i.e. the so called extrinsic
(N )
information (in other words, that part of the information available about xl
(N )
and not intrinsically influenced by xl itself). This interpretation is confirmed
(z)
by the fact that Ll,j [k] is computed on the basis of the statistical knowledge
(L)
(L)
available about xl and xl+1 (see (88) and (90)), which, thanks to (2), does
(N )
provide useful information about xl . The theory of turbo decoding of channel
codes shows that, generally speaking, extrinsic information originates from code
constraints. In our scenario, a similar interpretation can be also provided for
(z)
(L)
Ll,j [k], since xl+1 can be seen as the noisy output of a communication channel,
(L)
(L)
affected by the bias fl,j and the additive noise wl , and over which the code(L) (L)
word Al,j xl
of a rate-1 block code is transmitted in response to the message
(L)
xl .
(z)
These considerations show that, in evaluating Ll,j [k], we are actually
exploiting a sort of ‘code’ constraints, which are mathematically expressed by
(2). The reader can easily verify that a similar interpretation can be provided
(L)
for m
~ j (zl ) (53), which represents the extrinsic information component14 con(L)
(L)
tained in m
~ 4,j (xl ) (57) (conveying our a posteriori information about xl );
(L)
the other two components are represented by the message m
~ 2,j (xl ) (40) (repre14 In
(L)
practice, the mechanism employed to generate zl,j is based on the state update equation (3).
G. M. Vitetta et al., Marginalized Particle Filtering . . .
(L)
32
(L)
senting the measurement information about xl ) and the message m
~ in,j (xl )
(L)
(21) (corresponding to our a priori information about xl ). Consequently, TF
can be seen, in the domain of Bayesian filtering techniques, as the counterpart
of turbo decoding of concatenated codes; this parallelism can be exploited to
provide further insights into iterative filtering techniques. For instance, it is
well known that, in turbo decoding of concatenated channel codes, the extrinsic
information generated by soft decoders become more and more correlated as iterations evolve; this entails that diminishing benefits are provided by additional
iterations. This phenomenon should be observed in TF too for similar reasons
(N )
(N )
and can be motivated by rewriting ηz,l,j (88) and Cz,l,j (90) as
(N )
(L)
(L)
ηz,l,j = fl,j + Al,j
and
(N )
(L)
Cz,l,j = C(L)
w + Al,j
h
h
i
(L)
(L)
η4,l,j − η2,l,j
(L)
(L)
C4,l,j − C2,l,j
i
T
(L)
,
Al,j
(105)
(106)
respectively (thanks to (85) and (86), respectively). In fact, the last two equa(N )
(N )
tions show that the vector ηz,l,j and the matrix Cz,l,j are influenced by the
difference between the statistical information (expressed by a mean vector and
(L)
a covariance matrix) available about xl before processing the pseudomeasure(L)
ment zl and those available after this task has been carried out. In other
(L)
(N )
words, the extrinsic information provided by zl influences zl and viceversa.
Finally, it is important to point out that the proposed analogy between turbo
filtering and turbo decoding suggests the potential limits of the TF technique
(and of any other iterative filtering method relying on the developed FG). In
fact, it is well known that turbo decoding methods do not provide real benefits
below a certain signal-to-noise ratio, i.e. when the quality of the received signal
is so poor that the transmitted coded sequence cannot be recovered. A similar
phenomenon is expected occur with TF too; consequently, this filtering method
could not outperform MPF in the presence of strong measurement noise and/or
fast dynamics affecting the considered SSM.
7
Numerical Results
In this Section MPF and the related filtering methods developed in this manuscript
are compared in terms of accuracy and computational load for a specific CLG
system, characterized by DL = 3, DN = 1 (so that D = 4) and P = 2. The
structure of the considered system has been partly inspired by the example
G. M. Vitetta et al., Marginalized Particle Filtering . . .
proposed by Schön
0.8
(L)
xl+1 = 0
0
and
(L)
in [37] and is characterized by: a) the state models
(N )
cos(xl )
0.2
0
(L)
(L)
)
0.7 −0.2 xl + − sin(x(N
) + wl
l
(N )
0.2 0.7
0.5 sin(2xl )
(N )
(N )
(L)
(N )
xl+1 = arctan xl
+ (0.9 0 0) xl + wl
(L)
(N )
33
(107)
(108)
(N )
∼ N (0, (σw )2 I3 ), wl ∼ N (0, (σw )2 ; b) the measurement model
!
(N ) 2
(N )
0 0 0
0.1 xl
· sgn xl
(L)
yl =
+
xl + el
(109)
1 −1 1
0
with wl
with el ∼ N (0, (σe )2 I2 ). Note that the state equation (107), unlike itscounter
(N )
(L)
(N )
part proposed in [37], depends on xl (i.e., it contains a function fl
xl+1 6=
(N )
03 in its RHS), so that TF, which relies on the availability of the vector zl
(18), can be employed for this system.
In our computer simulations the root mean square error (RMSE) has been
evaluated to compare the accuracy of the state estimates generated by different
filtering techniques. More specifically, for each technique two RMSEs have been
computed, one (denoted RM SEL (alg), where ‘alg’ denotes the algorithm this
parameter refers to) representing the square root of the average mean square
(L)
error (MSE) evaluated for the three elements of xl , the other one (denoted
(N )
RM SEN (alg)) referring to the (monodimensional) nonlinear component xl ;
this distinction is important since, as shown by our simulation results, the es(L)
timation accuracy for xl can be quite different from (and is usually smaller
(N )
than) that referring to xl .
As far the assessment of the computational requirements of the investigated
filtering techniques is concerned, MPF (for which an accurate analysis of its
computational complexity is available in [16]) has been taken as a baseline. For
this reason, our comparisons between the considered filtering techniques are
based on the evaluation of a single parameter, denoted ∆c (alg) and representing the percentage variation in computation time of the considered algorithm
(denoted ‘alg’) with respect to MPF (operating with the same parameters and,
in particular, with the same Np as alg).
Moreover, in our computer simulations, the following choices have been
(N )
(L)
made: a) σw = σw = 5 · 10−3 has been selected for the standard devia(L)
(N )
tions of the process noises {wk } and {wk }, unless differently stated; b) the
so called jittering technique [34] has been employed to mitigate the so called
depletion problem in the generation of new particles.
34
G. M. Vitetta et al., Marginalized Particle Filtering . . .
Some results illustrating the dependence of RM SEL and RM SEN on the
number of particles (Np ) for MPF and SMPF#1 are illustrated in Fig. 5 (in this
and in the following figures simulation results are identified by markers, whereas
continuous lines are drawn to ease reading); σe = 10−2 and Np ∈ [50, 300] have
been selected in this case. From these results the following conclusions can be
easily inferred for the considered system:
1. A negligible improvement in the estimation accuracy of both MPF and
SMPF#1 is achieved if the value of Np exceeds 200.
2. A significant gap between RM SEL (MPF) and RM SEN (MPF) exists; this
(L)
is motivated by the fact that, in MPF, the estimation of xl relies on both
(L)
the real measurement yl (109) and the pseudo-measurements {zl,j } (54),
(N )
whereas the estimation of xl
benefits from yl only.
3. The gap between RM SEL (SMPF#1) and RM SEN (SMPF#1) is much
smaller than that observed for MPF, even if the pseudo-measurements
(L)
{zl,j } are also exploited by SMPF#1. This reduction in the RMSE gap
(L)
can be related to the degradation in the estimation accuracy of xl ; for instance, for Np = 200, RM SEL (SMPF#1) is about twice RM SEL (MPF)
(on the contrary, RM SEN (SMPF#1) ∼
= 1.09· RM SEN (MPF)).
Our numerical results have also evidenced that: a) SMPF#1 requires a substantially smaller computational effort than MPF, since ∆c (SMPF#1) approximately ranges in the interval [−62%, −58%] for the considered values of Np ;
b) despite the above mentioned degradation in estimation accuracy (see point
3.), SMPF#1 does not suffer from tracking losses in the considered scenario;
c) SMPF#2 accuracy is almost indentical that of SMPF#1; d) ∆c (SMPF#2)
is approximately lower than ∆c (SMPF#1) by 6% in the considered range for
Np and, consequently, achieves a better complexity-performance tradeoff than
SMPF#1. All this suggests that simplified MPF techniques can be really developed without incurring the serious technical problems that affect the techniques
proposed in [17] (tracking losses and poor RMSE performance) and in [18] (partitioning and update of the particle set into groups within each recursion; see
Section 5).
The dependence of RM SEL and RM SEN on σe (i.e., on the intensity of
the noise affecting the available measurements) has been also analysed for MPF
and SMPF#1. Some results are shown in Fig. 6; Np = 500 and σe ∈ [10−3 , 5 ·
10−2 ] have been selected in this case. These results show that the gap between RM SEL (MPF) and RM SEL (SMPF#1) slightly increases as σe becomes
35
G. M. Vitetta et al., Marginalized Particle Filtering . . .
5
4
SMPF1
RMSE
3
-2
2x10
RMSEL
RMSEN
MPF
0.01
50
100
150
200
250
300
Np
Figure 5: RMSE performance versus Np for the linear component (RM SEL )
and the nonlinear component (RM SEN ) of the state xl for the system described
(L)
by eqs. (107)-(109). MPF and SMPF#1 are considered; in both cases σw =
(N )
σw = 5 · 10−3 and σe = 10−2 have been selected.
36
G. M. Vitetta et al., Marginalized Particle Filtering . . .
6
SMPF1
5
4
3
RMSE
2
MPF
0.01
9
8
7
RMSEL
RMSEN
6
5
10
20
30
40
-3
50x10
σe
Figure 6: RMSE performance versus σe for the linear component (RM SEL ) and
the nonlinear component (RM SEN ) of the state xl for the system described by
(L)
eqs. (107)-(109). MPF and SMPF#1 are considered; in both cases σw =
(N )
σw = 5 · 10−3 and Np = 200 have been selected.
smaller; the opposite occurs for RM SEN (MPF) and RM SEN (SMPF#1). These
(L)
results can be related again to the fact that the pseudo-measurements {zl,j }
(L)
(54) really play a role in the MPF estimation of xl and that the quality of the
information conveyed by these pseudo-measurements indirectly improves as the
real measurements become less noisy.
Some results illustrating the dependence of RM SEL and RM SEN on the
number of particles (Np ) for MPF and TF are illustrated in Fig. 7; σe = 10−2 ,
Np ∈ [1, 300] and Nit = 2 for TF, and Np ∈ [30, 300] for MPF have been chosen
in this case. From these results it is easily inferred that:
1. TF outperforms MPF in tems of both RM SEL and RM SEN ; for instance, RM SEL (MPF) ∼
= 1.71· RM SEL (TF) and RM SEN (MPF) ∼
=
2.86· RM SEN (TF) if Np = 200 is selected.
G. M. Vitetta et al., Marginalized Particle Filtering . . .
37
2. The gap between RM SEL (MPF) and RM SEN (MPF) is substantially
larger than the corresponding gap for TF (in particular, RM SEN (MPF)−
RM SEL (MPF) ∼
= 13.4·[RM SEN (TF)−RM SEL (TF)]; this can be easily
(L)
related to the fact that, unlike MPF, in TF the estimation of both xl
(N )
and xl benefits from the availability of pseudo-measurements.
3. The performance gap between TF and MPF increases as Np gets smaller.
Moreover, TF accuracy starts quickly degrading when Np drops below
11, whereas the same phenomenon starts for Np ∼
= 30 with MPF. Note,
however, that different phenomena occur with MPF and TF when the
particle set is small. In fact, our computer simulations have shown that
MPF suffers from frequent tracking losses for Np < 30 (since it is unable
(N )
to generate a reliable representation of xl
when a limited number of
particles is available); on the contrary, TF does not suffer from the same
problem even if a very small particle set is used.
We believe that last result is really important and can be motivated as
follows. The TF technique, through its feedback mechanism, makes a substantially more efficient use of the available particles than MPF; this results in an
appreciable improvement of both stability and accuracy of state estimation.
Our simulation results have also shown that: a) in the considered scenario a
negligible improvement in estimation accuracy is obtained if Nit > 2 is selected;
b) ∆c (TF) approximately ranges in the interval [70%, 80%] for the considered
values of Np and, consequently, requires a substantially larger computational
effort than MPF if these algorithms operate with the same number of particles. In practice, however, as evidenced by the results shown in Fig. 7, TF
can reliably operate with a very small particle set and, consequently, it outperforms MPF in terms of both performance and complexity if Np is properly
selected. For instance, in the considered scenario TF with Np = 20 particles
achieves a substally better accuracy than MPF with Np = 40 particles, even
if, as evidenced by our computer simulations, they approximately require the
same computation time. It is not difficult to show that similar considerations
hold if SMPF#1 and SMPF#2 are considered in place of MPF. Therefore, our
results suggest that the real key to the complexity reduction of the filtering techniques relying on the FG shown in Fig. 2 is not provided by the approximations
adopted for the MPF processing tasks in Section 5, but by the exploitation of
(N )
the new pseudo-measurement zl (18). We should never forget, however, that
TF cannot
be adopted for all CLG systems; in fact, it cannot be employed if
(L)
(N )
fl
xl+1 = 0DL in the RHS of (2).
38
G. M. Vitetta et al., Marginalized Particle Filtering . . .
5
4
RMSEL
RMSEN
MPF
3
RMSE
2
0.01
9
8
7
TF
6
5
0
50
100
150
Np
200
250
300
Figure 7: RMSE performance versus Np for the linear component (RM SEL )
and the nonlinear component (RM SEN ) of the state xl for the system described
(L)
(N )
by eqs. (107)-(109). MPF and TF are considered; in both cases σw = σw =
−3
−2
5 · 10 and σe = 10 have been selected.
39
G. M. Vitetta et al., Marginalized Particle Filtering . . .
8
7
6
5
RMSEL
RMSEN
σw = 5 10-3
σw = 10-3
MPF
4
3
RMSE
2
0.01
8
7
6
5
TF
4
3
-2
3
2x10
4
σe
Figure 8: RMSE performance versus σe for the linear component (RM SEL ) and
the nonlinear component (RM SEN ) of the state xl for the system described by
eqs. (107)-(109). MPF and TF are considered; σe = 10−2 and Np = 200 are
(L)
assumed in both cases. As far system noise is concerned, the cases σw =
(N )
(L)
(N )
σw = 5 · 10−3 and σw = σw = 10−1 are taken into consideration.
Finally, the dependence of RM SEL and RM SEN on σe has been assessed for
TF and compared with that characterizing MPF. Some numerical are illustrated
in Fig. 8. In this case, we have selected σe ∈ [1.5 · 10−2 , 5 · 10−2 ] , Np = 200
(L)
(N )
and Nit = 2 for TF, and Np = 200 for MPF; moreover, σw = σw = 5 · 10−3
(L)
(N )
and σw = σw = 10−3 have been considered to analyse the dependence of
the performance gap between TF and MPF on the intensity of process noise
(and, consequently, on system dynamics). These results show that: a) the
performance gap between TF and MPF undergoes small changes if σe varies
in the considered interval; b) on the contrary, a substantial change in this gap
(L)
(N )
is obtained if σw and σw are reduced from 5 · 10−3 to 10−3 . This suggests
that measurement noise and process noise can have different impacts on the
performance gain provided by TF over MPF.
5
G. M. Vitetta et al., Marginalized Particle Filtering . . .
8
40
Conclusions
In this manuscript a FG approach has been employed to analyse the filtering
problem for mixed linear/nolinear models. This has allowed us to: a) prove
that this problem involves a FG which is not cycle free; b) provide a new interpretation of MPF as a forward only message passing algorithm over a specific
FG; c) develop novel filtering algorithms for simplifying or generalising it. In
particular, an important iterative filtering technique, dubbed turbo filtering, has
been devised and its relation with the turbo decoding techniques for concatenated channel codes has been analysed in detail. All the considered filtering
techniques have been compared in terms of both accuracy and computational
requirements for a specific CLG system. The most interesting result emerging from our computer simulations is represented by the clear superiority of
turbo filtering over marginalized particle filtering. In fact, the former technique, through the exploitation of new pseudo-measurements, can achieve a
better accuracy than the latter one at an appreciably smaller computational
load. Our ongoing research activities in this area include the development of
other related filtering techniques and the application of turbo filtering to specific
state estimation problems.
A
Appendix
Given the pdfs f1 (y) , N (y; η1 , C1 ) and f2 (y) , N (y; η1 , C2 ) for the N dimensional vector y, we are interested in evaluating the correlation between
these two functions, i.e. the quantity
Z
c1,2 , f1 (y) · f2 (y) dy
(110)
Substituting the expressions of f1 (y) and f2 (y) in the RHS of the last equation
produces, after some manipulation,
1 T
T
T
(111)
η Wη − η1 W1 η1 − η2 W2 η2
c1,2 = D exp
2
−1
where W1 , C−1
1 , W 2 , C2 ,
W = W1 + W2 ,
(112)
Wη = W1 η1 + W2 η2
(113)
D = (2π det [C1 + C2 ])−N/2 .
(114)
and
G. M. Vitetta et al., Marginalized Particle Filtering . . .
41
References
[1] M. S. Arulampalam, S. Maskell, N. Gordon and T. Clapp, “A Tutorial on
Particle Filters for Online Nonlinear/Non-Gaussian Bayesian Tracking”,
IEEE Trans. Sig. Proc., vol. 50, no. 2, pp. 174-188, Feb. 2002.
[2] F. E. Daum, “Exact Finite-Dimensional Nonlinear Filters”, IEEE Tran.
Aut. Contr., vol. 31, no. 7, pp. 616-622, July 1986.
[3] S. Mazuelas, Y. Shen and M. Z. Win, “Belief Condensation Filtering”,
IEEE Trans. Sig. Proc., vol. 61, no. 18, pp. 4403-4415, Sept. 2013.
[4] V. Smidl and A. Quinn, “Variational Bayesian Filtering”, IEEE Trans. Sig.
Proc., vol. 56, no. 10, pp. 5020-5030, Oct. 2008.
[5] M. S̆imandl, J. Královeca and T. Söderströmc, “Advanced Point-Mass
Method for Nonlinear State Estimation”, Automatica, vol. 42, pp. 11331145, 2006.
[6] B. Anderson and J. Moore, Optimal Filtering, Englewood Cliffs, NJ,
Prentice-Hall, 1979.
[7] S. J. Julier and J. K. Uhlmann, “Unscented Filtering and Nonlinear Estimation”, IEEE Proc., vol. 92, no. 3, pp. 401-422, Mar. 2004.
[8] A. Doucet, J. F. G. de Freitas and N. J. Gordon, “An Introduction to Sequential Monte Carlo methods,” in Sequential Monte Carlo Methods
in Practice, A. Doucet, J. F. G. de Freitas, and N. J. Gordon, Eds. New
York: Springer-Verlag, 2001.
[9] A. Doucet, S. Godsill and C. Andrieu, “On Sequential Monte Carlo Sampling Methods for Bayesian Filtering”, Statist. Comput., vol. 10, no. 3, pp.
197-208, 2000.
[10] F. Gustafsson, “Particle Filter Theory and Practice with Positioning Applications”, IEEE Aerosp. and Electr. Syst. Mag., vol. 25, no. 7, pp. 53-82,
July 2010.
[11] F. Gustafsson, F. Gunnarsson, N. Bergman, U. Forssell, J. Jansson, R.
Karlsson and P. Nordlund, “Particle Filters for Positioning, Navigation,
and Tracking”, IEEE Trans. Sig. Proc., vol. 50, 425-435, 2002.
G. M. Vitetta et al., Marginalized Particle Filtering . . .
42
[12] P. J. Nordlund and F. Gustafsson, “Marginalized Particle Filter for Accurate and Reliable Terrain-Aided Navigation”, IEEE Trans. on Aerosp. and
Elec. Syst., vol. 45, no. 4, pp. 1385-1399, Oct. 2009.
[13] R. Bucy and K. Senne, “Digital Synthesis of Non-Linear Filters”, Automatica, vol. 7, no. 3, pp. 287-298, 1971.
[14] F. Daum and J. Huang, “Curse of Dimensionality and Particle Filters”,
Proc. IEEE Aerosp. Conf., vol. 4, pp. 1979-1993, March 2003.
[15] T. Schön, F. Gustafsson, P.-J. Nordlund, “Marginalized Particle Filters for
Mixed Linear/Nonlinear State-Space Models”, IEEE Trans. Sig. Proc., vol.
53, no. 7, pp. 2279-2289, July 2005.
[16] R. Karlsson, T. Schön, F. Gustafsson, “Complexity Analysis of the
Marginalized Particle Filter”, IEEE Trans. Sig. Proc., vol. 53, no. 11, pp.
4408-4411, Nov. 2005.
[17] F. Mustiere, M. Bolic and M. Bouchard, “A Modified Rao-Blackwellised
Particle Filter”, Proc. of the 2006 IEEE Int. Conf. on Ac., Sp. and Sig.
Proc. (ICASSP 2006), vol. 3, 14-19 May 2006.
[18] T. Lu, M. F. Bugallo and P. M. Djuric, “Simplified Marginalized Particle
Filtering for Tracking Multimodal Posteriors”, Proc. IEEE/SP 14th Workshop on Stat. Sig. Proc. (SSP ’07), pp. 269-273, Madison, WI (USA), 2007.
[19] F. Lindsten, P. Bunch, S. Särkkä, T. B. Schön and S. J. Godsill, “RaoBlackwellized Particle Smoothers for Conditionally Linear Gaussian Models”, IEEE J. Sel. Topics in Sig. Proc., vol. 10, no. 2, pp. 353-365, March
2016.
[20] H.-A. Loeliger, J. Dauwels, Junli Hu, S. Korl, Li Ping, F. R. Kschischang,
“The Factor Graph Approach to Model-Based Signal Processing”, IEEE
Proc., vol. 95, no. 6, pp. 1295-1322, June 2007.
[21] F. R. Kschischang, B. Frey, and H. Loeliger, “Factor Graphs and the SumProduct Algorithm”, IEEE Trans. Inf. Theory, vol. 41, no. 2, pp. 498-519,
Feb. 2001.
[22] C. Berrou and A. Glavieux, “Near Optimum Error Correcting Coding and
Decoding: Turbo-Codes”, IEEE Trans. Commun., vol. 44, no. 10, pp. 1261
- 1271, Oct. 1996.
G. M. Vitetta et al., Marginalized Particle Filtering . . .
43
[23] S. Benedetto, D. Divsalar, G. Montorsi and F. Pollara, “Serial Concatenation of Interleaved Codes: Performance Analysis, Design, and Iterative
Decoding”, IEEE Trans. Inf. Theory, vol. 44, no. 3, pp. 909-926, May 1998.
[24] R. Koetter, A. C. Singer and M. Tüchler, “Turbo Equalization”, IEEE Sig.
Proc. Mag., vol. 21, no. 1, pp. 67-80, Jan. 2004.
[25] J. Hagenauer, “The Turbo Principle: Tutorial Introduction & State of the
Art”, Proc. Int. Symp. Turbo Codes & Related Topics, Brest, France, Sep.
1997, pp. 1-11.
[26] A. P. Worthen and W. E. Stark, “Unified Design of Iterative Receivers using
Factor Graphs”, IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 843–849, Feb.
2001.
[27] F. R. Kschischang and B. J. Frey, “Iterative Decoding of Compound Codes
by Probability Propagation in Graphical Models”, IEEE J. Sel. Areas Commun., vol. 16, no. 2, pp. 219-230, Feb. 1998.
[28] J. Dauwels, S. Korl and H.-A. Loeliger, “Particle Methods as Message
Passing”, Proc. 2006 IEEE Int. Symp. on Inf. Theory, pp. 2052-2056, 9-14
July 2006.
[29] T. P. Minka, “Expectation Propagation for Approximate Bayesian Inference”, Proc. 17th Annual Conf. Uncertainty in Artif. Intell., pp. 362-369,
Seattle, WA (USA), Aug. 2001.
[30] O. Zoeter and T. Heskes, “Deterministic Approximate Inference Techniques
for Conditionally Gaussian State Space Models”, Statistics and Computing,
vol. 16, no. 3, pp. 279-292, Sep. 2006.
[31] V. Smı́dl and A. Quinn, The Variational Bayes Method in Signal
Processing, Berlin, Germany, Springer, 2005.
[32] C. M. Bishop, Pattern Recognition and Machine Learning, Springer,
2006.
[33] A. R. Runnalls, “Kullback-Leibler Approach to Gaussian Mixture Reduction”, IEEE Trans. on Aer. and Elec. Syst., vol. 43, no. 3, pp. 989-999,
July 2007.
[34] T. Li, M. Bolic, P. Djuric, “Resampling Methods for Particle Filtering:
Classification, Implementation, and Strategies”, IEEE Sig. Proc. Mag., vol.
32, no. 3, pp.70-86, May 2015.
G. M. Vitetta et al., Marginalized Particle Filtering . . .
44
[35] G. M. Vitetta, D. P. Taylor, G. Colavolpe, F. Pancaldi and P. A. Martin,
Wireless Communications: Algorithmic Techniques, John Wiley &
Sons, 2013.
[36] J. Hagenauer, E. Offer and L. Papke, “Iterative decoding of binary block
and convolutional codes”, IEEE Trans. Inf. Theory, vol. 42, no. 2, pp.
429-445, Mar 1996.
[37] T. Schön, “Example Used in Exemplifying the Marginalized
(Rao-Blackwellized) Particle Filter”,
Nov. 2010 (available at
http://users.isy.liu.se/en/rt/schon/Code/RBPF/Document/MPFexample.pdf).
| 3cs.SY
|
Global Model Interpretation via Recursive Partitioning
Chengliang Yang
Anand Rangarajan
Sanjay Ranka
Department of Computer &
Information Science & Engineering
University of Florida
Gainesville, Florida 32611
[email protected]
Department of Computer &
Information Science & Engineering
University of Florida
Gainesville, Florida 32611
[email protected]
Department of Computer &
Information Science & Engineering
University of Florida
Gainesville, Florida 32611
[email protected]
arXiv:1802.04253v1 [cs.LG] 11 Feb 2018
ABSTRACT
In this work, we propose a simple but effective method to interpret
black-box machine learning models globally. That is, we use a compact binary tree, the interpretation tree, to explicitly represent the
most important decision rules that are implicitly contained in the
black-box machine learning models. This tree is learned from the
contribution matrix which consists of the contributions of input
variables to predicted scores for each single prediction. To generate
the interpretation tree, a unified process recursively partitions the
input variable space by maximizing the difference in the average
contribution of the split variable between the divided spaces. We
demonstrate the effectiveness of our method in diagnosing machine
learning models on multiple tasks. Also, it is useful for new knowledge discovery as such insights are not easily identifiable when
only looking at single predictions. In general, our work makes it
easier and more efficient for human beings to understand machine
learning models.
KEYWORDS
Interpretable machine learning, Model understanding, Knowledge
discovery
1
INTRODUCTION
Though machine learning advances greatly in many areas in recent
years such as computer vision and natural language processing,
limited interpretability hinders it from impacting areas that require clearer evidence for decision making such as health care and
economy. In these domains, most widely used machine learning
models are linear regression or decision trees that people can easily understand. To deploy cutting-edge machine learning in such
domains, some transparent mechanisms are needed to explain the
sophisticated models to users. Limited interpretability also harms
improving machine learning models. The black-box behavior makes
it difficult to diagnose the models. Without good understanding of
how the model works, lots of effort are wasted in model parameters
tuning.
Thus, machine learning researchers have been trying to better interpret machine learning models. Recent progresses include
designing specific neural network structure that imposes linear
constraints on weights of input variables in the decision function
[8], using model structure based heuristics to decompose prediction
scores [28], approximating any model linearly in a local area [21],
and track predictions using influence functions back to training
data [14]. However, all these work can only provide local interpretation. That is, the interpretation is generated for a particular
sample of data. This is not desired when we want to know the
general picture of the model. Modern machine models are usually
trained and tested on millions of data samples. It is not practical
for a researcher to review the interpretation of each of them for
model diagnostics. What’s more, when machine learning models
are used to inform population level decisions, such as an economic
policy change, a global effect estimate would be more helpful than
thousands of local explanations. Thus there is a gap between local
and global machine learning model interpretation, which this paper
is going to address.
By ”interpreting a machine learning model globally”, we mean
representing a trained machine learning model in an aggregated
and human understandable way. This is done by extracting the
most important rules that the model learned from training data
and would apply to testing data. These rules affect a substantial
portion of data from the model perspective and thus useful to inform decision impacting globally for all data samples. The simplest
example of such rules are the coefficients we could learn in a linear
regression model. These coefficients represent the magnitude of
changes in the output due to one unit of change in the input variables. From the assumption of the linear model, the coefficients are
identical for any data sample. Thus they are used widely for effect
estimation in observational studies and randomized experiments.
Another globally interpretable machine learning model is decision
tree, which presents the decision rules in a straightforward tree
structure. However, both linear regression and decision tree lack
high predictive power. which means people have to tradeoff between predictability and interpretability when both properties are
desired.
In this work, we propose a new method, Global Interpretation via
Recursive Partitioning (GIRP), to build a global interpretation tree
for a wide range of machine learning models based on their local
explanations. That is, we recursively partition the input variable
space by maximizing the difference in the contribution of input
variables averaged from local explanations between the divided
spaces. By doing so, we end up with a binary tree that we call
the interpretation tree describing a set of decision rules that is an
approximation of the original machine learning model. Figure 1
describes the work flow of building a global model interpretation.
With a trained machine learning model and the data you want to
use to explain it, we first generate a contribution matrix from local
explanations either using model specific heuristics or some local
model interpreter [21]. Then we send contribution matrix to our
Global Interpretation via Recursive Partitioning (GIRP) algorithm.
The algorithm returns with an interpretation tree that generally
describes the machine learning model and is fully comprehensible
by human beings. The key contributions of our paper are as follows:
Figure 1: Work flow of global model interpretation
• We propose an efficient and effective method to address the
gap in the literature of globally interpreting many machine
learning models. The interpretation is in the form of an
easily understandable binary tree that could be used to
diagnose the interpreted model or inform population level
decisions.
• The CART [6] like algorithm that we use to build the interpretation tree could model interactions between input
variables. Thus, we are able to find the heterogeneity in
variable importances among different subgroups of data.
• In our experiments, we showcase that our method can
discover whether a particular machine learning model is
behaving in a reasonable way or overfit to some unreasonable pattern.
The rest of the paper is organized as follows. Section 2 describes
some works that are closely connected to our work. Section 3
presents the Global Interpretation via Recursive Partitioning (GIRP)
algorithm. Section 4 applies GIRP to computer vision, natural language processing, and health care predictive models from structured
tabular data. Section 5 concludes the paper.
2
RELATED WORK
There are four parts of existing work that are closely related to
our method, that is, local model interpretation, global model interpretation, recursive partitioning for effects estimation, and feature
selections.
There are several ways to achieve local model interpretation.
First, people structure the model in a way that the output is linear
in terms of input variables so that the weights could be used as a
measure of importance. For example, [8] uses the neural attention
mechanism [4] to generate interpretable attention weights in recurrent neural networks. However, due to the stochastic training
process used, these attentions are shown to be unstable [29]. The
second way uses model specific heuristics to decompose the predicted scores by input variables. [28] described such methods for
regularized regression and gradient boosted machine. Third, locally
approximating sophisticated models using simple interpretable
model could explain individual predictions. Gradient vector and
sparse linear methods are tried as the local explainer [3, 15, 21]. Finally, influence functions from robust statistics can be used to track
a particular prediction back to training data that are responsible
for it [14]. In conclusion, all local model interpretation methods
work at the single data sample level, generating the contributions
of input variables to the final predicted score for a specific data
sample.
People also try to directly build a globally interpretable model,
including additive models for predicting pneumonia risk [7] and
rule sets generated from sparse Bayesian generative model [16].
However, these models are usually specifically structured thus limited in predictability to preserve interpretability. [9] uses queries to
build a tree to approximate neural networks. [2] generally discusses
presenting machine learning models in different levels.
Recursive partitioning and its resulting tree structure is an intuitive way to present rule sets and model interactions between input
variables. It has been used for a long time to analyze heterogeneity
for subgroup analysis in survey data [18]. Recently it is applied
to study heterogeneous causal or treatment effects [1, 23]. It is a
good fit for our global model interpretation task because we want
to extract the rules that machine learning model finds and these
rules are affected by the interactions between input variables.
Feature selection methods select a subset of important features
from the input variables when the machine learning model is
trained. The model interpretability could be benefited from this
process because it reduces the dimension of input variables, making
the model compact and easier to be presented [24]. This is very
useful when the input is very high dimensional [19]. The feature
selection process could either be conducted before the model fitting
[11] or embedded into it [24, 27]. Though feature selection and
global model interpretation tasks both aim to extract the most important variables or their combinations, they are different because
global model interpretation is a post model fitting process. We
represent the trained model in a compact and comprehensible way
with good fidelity to the original model. The goal is not to make
predictions using this representation but understand how it predicts. In contrast, feature selection discards unimportant variables
and predictions will be solely based on selected ones.
3
GLOBAL INTERPRETATION VIA
RECURSIVE PARTITIONING
We follow the CART [6] work flow to build the interpretation
tree, including growing a large initial tree, pruning, and using the
validation set for best tree size selection. But before we describe
the tree building process in detail, we introduce the contribution
matrix first, which our method takes as input.
3.1
Contribution Matrix
As mentioned, local model interpretation methods [3, 8, 15, 21, 28]
can generate the contribution of each single input variable to the
final predicted score for a specific data sample. In detail, for a
machine learning model that take N input variables, given a new
data sample, it generates a quantity c i for the i-th variable vi to
measure the importance of this variable in the prediction made.
We call this quantity the contribution of variable vi . If there are
M data samples in total, we could generate a contribution matrix
using local model interpretation methods as shown in Table 1. c ij is
the contribution of variable vi to predicted score p j of sample s j .
Thus, each row of the contribution matrix represents how the model
thinks of variable importances in the corresponding prediction.
Contribution
Sample 1
Sample 2
Sample 3
…
Sample j
…
Sample M
Var 1
c 11
c 21
c 31
…
c 1j
…
1
cM
Var 2
c 12
c 22
c 32
…
c 2j
…
2
cM
Var 3
c 13
c 23
c 33
…
c 3j
…
3
cM
Var i
c 1i
c 2i
c 3i
…
c ij
…
i
cM
…
…
…
…
…
…
…
…
…
…
…
…
…
…
…
…
Var N
c 1N
c 2N
c 3N
…
c jN
…
N
cM
Predicted Score
p1
p2
p3
…
pj
…
pM
Table 1: Contribution matrix generated from local model interpretations for every single data sample. c ij is the contribution
of variable v i to predicted score p j of sample s j .
It is straightforward to obtain the contribution matrix when
features are explicit and individual contributions could be generated along with predictions like in linear regression and [8, 28].
However, in other cases we need some workaround to identify
the variables that the contributions could be attributed to. When
analyzing convolutional nerual networks in [21], segmentation
of images is generated first to carry contributions. In our experiment diagnosing the scene classification deep learning method, a
semantic segmentation algorithm is applied to the scene images
to break them into semantic meaningful segments as well. These
workarounds are problem specific and affect the formation of the
contribution matrix.
3.2
Growing a Large Initial Tree
Now we can move forward to the first step to build the interpretation tree, growing a large initial tree. The same greedy process as
CART is adopted. For any input variable i, we could apply a split
based on values of variable i to divide all the data samples into two
subgroups. Note that the split is based on the input variable value
but not contribution c i . We use v i to denote the input values to
discriminate it from contribution c i . The type of split depends on
type of variable v i . If it is binary, the split criteria could be ”v i = 1?”.
If v i is ordinal, we could apply the criteria ”if v i < d” where d is
some constant value. If v i is categorical, let D denote a subset of all
possible values of variable v i , we could apply ”v i ∈ D?” as the split
criteria. For convenience, assume that all data samples meet the
split criteria go to the right subset S R and the others go to the left
subset S L . For the two subsets of data samples S R and S L . Consider
the quantity below:
i
s j ∈S L c j
Í
G(spliti ) =
|S L |
i
s j ∈S R c j
Í
−
|S R |
!
(1)
spliti means the split is over variable v i . The first term quantifies
the average contribution of variable vi in the left subset S L . So does
the second term for the right subset S R . The difference between
these two terms measures how differently variable v i contributes
to the predicted score in S R and S L . The larger this difference is,
the more discriminative the model think variable v i is. Thus, by
finding the maximum |G(spliti )|, we could get to know the most
import variable from the model perspective. So |G(spliti )| is used
as a measure of split strength in terms of variable importance.
We search all possible splits for all variables to find the best
initial split. After dividing the data sample into S R and S L , we
follow CART’s greedy approach to recursively partition S R and S L
and their child nodes until we reach to some pre-set threshold for
maximum tree depth or minimum number of samples in a node. As
a result of this step, we would get a large initial tree that explicitly
represents the most discriminative rules that the model implicitly
contains. We denote this large initial tree T0 .
3.3
Pruning
Due to the greedy approach to grow the initial tree, the rules contained in initial tree T0 are overly optimistic about the real world
problem and may not generalize well. Thus we need a procedure
to prune T0 to improve generalizability. Consider all the internal
nodes (non-leaf nodes) in T0 . All these nodes contain a split, say t.
Each split corresponds to a split value G(t) defined by Equation (1).
Suppose T is any interpretation tree and t is an internal node of T ,
we have
G(T ) =
Õ
|G(t)|
(2)
t ∈T
as a measure of the total split strength of the tree T that we generally want to maximize. To control the complexity of T , we add a
penalizing term to G(T ) to punish for more nodes in the tree.
G λ (T ) =
Õ
|G(t)| − λ|T |
(3)
t ∈T
Here |T | stands for the number of internal nodes in T . To maximize
G λ (T ), some of the internal nodes need to be removed from T if
G(t) for these nodes are less than λ. For larger λ, more nodes would
be removed so the resulting tree would be simpler and vice versa.
But how can we decide which internal nodes to remove? We first
define a new quantity for each internal nodes t for this purpose.
We use Tt to denote the subtree of T0 that has t as root.
|G(Tt )|
(4)
|Tt |
The above quantity intuitively defines the average split strength of
internal nodes in subtree Tt . With д(Tt ) defined, we iteratively remove the subtree with the smallest д(Tt ) from the initial full tree T0 .
Due to the greedy process to grow the initial tree, this process would
result in a series of nested tree, {TK ,TK −1 , ...,Tk ,Tk−1 , ...,T0 }. TK is
the null tree that only contains one node. [6] has proved that these
д(Tt ) =
Algorithm:
Step 1:
Step 2:
Step 3:
Step 4:
Step 5:
Global Interpretation via Recursive Partitioning
(GIRP)
Randomly split out a held-out validation dataset.
The rest data are fed to the trained machine learning model to get the contribution matrix;
Use Equation (1) to split the initial node;
Recursively partition the left and right child
nodes S L and S R to get the full Tree T0 until reaching to maximum tree depth or minimum number
of data samples in one node;
Use Equation (4) to calculate the average split
strength of each internal node of T0 . Iteratively
remove internal nodes from the one with smallest split strength to get a series of nested tree,
{TK ,TK −1 , ...,Tk ,Tk −1 , ...,T0 };
Use the held-out validation set and Equation
(5) and (6) to calculate Gval idat ion (Tk ) for each
of {TK ,TK −1 , ...,Tk ,Tk −1 , ...,T0 }. The one with
largest Gval idat ion (Tk ) is selected as the best
sized interpretation tree;
Table 2: The complete algorithm to generate the interpretation tree.
nested trees created by the iterative pruning process correspond to
a series of λ values, with λ K > λ K −1 > ... > λk > ... > λ 0 = 0.
3.4
Select Best Sized Tree
But how can we decide which Tk is the best sized tree for the final
interpretation tree, i.e., which value of λk is the best? Here we use
a held-out validation set for making this decision. We feed these
new validation data into each of {TK ,TK −1 , ...,Tk ,Tk −1 , ...,T0 } and
calculate for each internal node t
i
s j ∈S L c j
Í
Gval idat ion (t) = sдn(G(t))
|S L |
i
s j ∈S R c j
Í
−
|S R |
!
(5)
where sдn() is the sign function. Then we select the tree Tk as the
best sized tree with the largest Gval idat ion (Tk ):
Gval idat ion (Tk ) =
Õ
(Gval idat ion (t))
(6)
t ∈Tk
3.5
Choice of Hyperparameters
The only two hyperparameters we have in our approach are the
maximum depth of the interpretation tree and the minimum number
of data samples within a leaf or internal node. These are mostly
chosen empirically depending on the problem setting.
The full algorithm to generate the interpretation tree is described
in Table 2. Now we will move forward to demonstrate it on multiple
datasets using various machine learning models.
4
EXPERIMENT
In this section, we will try to interpret different types of machine
learning model on different tasks by representing them using the
interpretation tree. First, We will apply the proposed Global Interpretation via Recursive Partitioning (GIRP) algorithm to a scene
understanding deep learning model in computer vision. Second, we
will try a text classification task and see what words are important
to the random forest classifier. Finally, intensive care unit (ICU)
mortality prediction using recurrent neural network from medical
records demonstrates our approach on tabular data. Each of these
cases is different in obtaining the contribution matrix. We will
explain it in detail for each of them.
4.1
Scene Understanding
As many computer vision tasks greatly advanced by deep learning,
scene understanding has breakthroughs in accuracy with the help
of multi-million item labeled dataset and large scale deep neural
networks [32]. However, as most successful neural network architectures for computer vision, scene understanding neural networks
are not easily understandable because they are trained end-to-end
and act in a black-box way. What’s more, [20] shows that many
popular network architectures are easily fooled. Though some
workarounds are proposed to examine the evidence of neural predictions [5], they are at the single prediction level that could not
be efficiently used when there are millions of training and testing
data samples. Thus, people need a tool to extract the general rules
contained in the model from the whole set of data. If these rules
make sense to humans, we could trust the models more that they
will generalize well in real world.
In our demonstration, we will try to understand a deep residual
network trained for scene understanding on the MIT Place 365
dataset [12, 32]. To be specific, we will send images with ground
truth label in ”kitchen”, ”living room”, ”bedroom” and ”bathroom”
in the validation dataset, 100 images per category, to the trained
model and collect the predicted probabilities for these four categories. To obtain the contribution matrix, we apply a scene parsing
algorithm, dilated convolutional network [30, 33], to segment each
image into semantically meaningful parts. Then we perturb each
part with noise and re-evaluate the perturbed image in the scene
understanding neural network for new prediction scores for the
four categories. Using the varying scores, the contribution of each
semantic part of the image can be calculated via a sparse linear
regression model as the local model interpreter does [21]. Therefore, we could obtain the contribution of each semantic category
to prediction scores of the four scene categories, which form the
contribution matrix. Figure 2 describes this process more clearly.
After getting the contribution matrix that measures the importance of each semantic category for the scene categories for each
image, we could run our Global Interpretation via Recursive Partitioning (GIRP) algorithm to generate the interpretation tree for
each category, that is, ”kitchen”, ”living room”, ”bedroom”, and
”bathroom”. We set the maximum depth of the resulting tree to 100
and each node contains at least 20 images. The results are shown
in Figure 3. Only the first four levels of the resulting trees are
presented due to space limit. The actual best-sized tree is usually 5
to 10 levels in height. For each node in the interpretation tree, the
numbers of images in the nodes are shown. The accuracy number
measures what proportion of images are correctly identified as the
ground truth category for each tree. The split variable for each
Figure 2: From row 1 to row 4, each row have one example image from the four categories of ”bedroom”, ”living room”,
”kitchen”, and ”bathroom” in the MIT Place 365 scene understanding dataset [32]. The first column is the raw image. The
second column shows that semantic categories that found by the semantic segmentation algorithm, dilated convolutional
network [30] in the image. Column 3 shows actual semantic segmentation, which contains several superpixels for each image.
Using the local prediction interpreter [21], we could get the contribution of each superpixel, i.e., semantic category, to the
predicted scores. Column 4 presents the important semantic superpixel (with highest contribution scores) that are highlighted
green for corresponding ground truth category score respectively. For the ”bedroom” image, the ”bed” and ”floor” superpixels
are important. For the ”living room” image, ”sofa”, ”window pane”, and ”fireplace” are important. For the ”kitchen” image,
”cabinet” is the most important. Finally for the ”bathroom” image, ”toilet”, ”screen door” play the most important role. All
these explanations seem to be reasonable to us human being.
node is also shown. The contribution number is the average contribution of the split variable in the left and right child node. From
the trees, we can see that for ”kitchen”, ”living room”, ”bedroom”,
and ”bathroom” scenes, the model finds ”cabinet”, ”sofa”, ”bed”, and
”toilet” are the most discriminative semantic categories, which does
match our common sense. Besides, our approach also reveals some
useful rules that the model is following. For example, the ”sofa”,
”cushion”, and ”fireplace” combination achieve 0.958 in accuracy
for identifying ”living room”, while the ”cabinet”, ”stove”, and ”dishwasher” combination gets a perfect accuracy of 1 for ”kitchen”. All
these findings increase our confidence in the black-box residual
network based scene understanding deep learning model because it
is picking the right important object in the scene to make decisions.
4.2
Text Classification
We now turn our attention to the text classification task. [21]
reports that the text classifiers are picking up unreasonable words
to discriminate articles related to ”Christianity” from ones related to
”Atheism” using a subset of the 20-newsgroups corpus. While they
are showcasing this phenomenon by some randomly picked articles,
we want to check if at the corpus level the model does use words
unrelated to both concepts to classify articles. For this purpose, we
use our proposed approach to generate an interpretation tree using
words in the articles as features.
We train a random forest classifier [17] with 500 trees that
achieves 0.92 in accuracy on the test set to classify ”Christianity”
and ”Atheism” articles. We use TF-IDF [22] vectorizer to transfer
the article into vectors and then send them to the classifier. To get
the contribution matrix for building the interpretation tree. we
(a) Living room
(b) Kitchen
(d) Bathroom
(c) Bedroom
Figure 3: Interaction trees learned for scene categories ”living room”, ”kitchen”, ”bedroom”, and ”bathroom”.
use the local interpreter [21] that removing each word from the
articles one by one and monitoring the change in predicted score
to do a regression for evaluating the contribution of each word.
After running our Global Interpretation via Recursive Partitioning
algorithm, we obtain an interpretation tree as shown in Figure 4a.
Maximum tree depth is set to 100 and minimum number of data
samples in each node is 50. The results show that most words found
in the tree are not very related to concepts either of ”Christianity”
or fiAtheismfi, except ”God” and ”Christians” in the lower levels.
The most important words pulled, ”Posting”, ”Rutgers”, and ”com”,
look like just coincidental fake correlations captured by the model.
[21] reports an imbalanced word frequency of these words in the
two classes in the corpus. The model definitely overfits to these
patterns and would not generalize well in classifying new articles.
This finding implies that it would be a better practice to train robust
text classification machine learning models from multiple corpora
so that they are less likely overfitting to corpus specific features. In
this text classification example, we show that GIRP and interpretation tree could be used to diagnose models that overfit to the data
and reveal the incorrectly learned pattern.
4.3
Tabular Data: Predicting ICU Mortality
Structured tabular data widely exist in all kinds of relational databases
to represent various types of events or transactions. Hospitals use
standardized codes to record medical diagnosis, procedure and
pharmacy codes. The MIMIC database [13] is this kind of medical
database that contains intensive care unit (ICU) medical records
and is publicly available. We apply the RETAIN algorithm [8] to the
MIMIC database to predict mortality in the intensive care unit (ICU).
RETAIN is a specifically designed interpretable recurrent neural
network that can produce the contributions of past medical events
to a predicted new event using the neural attention mechanism
(b) ICU Mortality
(a) Text Classification
Figure 4: Left: Text Classification, The interpretation tree to explain the random forest model classifying ”Christianity” and
”Atheism” related articles. Unfortunately, we can see from the tree that the model is picking up unreasonable words such as
”Posting”, ”rutgers”, and ”com” as important features. We could expect bad generalizability of this model. Right: ICU Mortality,
The interpretation tree to explain the recurrent neural network model predicting mortality in ICU using past medical records.
The codes found by the algorithm are relevant to high or low risk of mortality.
[4]. However, due to the stochastic optimization process, [29] has
shown that these contributions are not stable when the recurrent
neural network model is re-trained by re-sampling training data.
We apply our proposed Global Interpretation via Recursive Partitioning (GIRP) to see if the global interpretation of the RETAIN
model is making sense when the local interpretations are unstable.
Past diagnosis codes are used to predict whether a patient will
die in the ICU. The contribution of each diagnosis code is generated by RETAIN along with the prediction. For the convenience of
interpretation, we aggregate the diagnosis codes to different time
frames, though RETAIN predicts on continuous time series. We
collect these contributions and organize them as the contribution
matrix. Then it is sent to GIRP to generate the interpretation tree,
which is shown in Figure 4b. Maximum tree depth is set to 100 and
minimum number of data samples in each node is 100. We can see
the most relevant diagnosis found by the algorithm is ”convalescence and palliative care in 1 month” that indicates a mortality of
85.3%. This is making sense because this diagnosis probably means
most medical treatments are tried and the doctors can do nothing
about the patients’ situation. On the other hand, ”other perinatal
jaundice in 1 month” seems to be a big protective factor of death
in the ICU. This is also reasonable because mostly jaundice is not
life threatening but needs emergent care. For other codes we may
not comment on the rationality because of the lack of health care
knowledge. However, this figure may help doctors if it finds some
relations between medical conditions and death in the ICU that are
not well investigated before in the medical practice. In this way, our
proposed method could potentially help discovering new important
factors or interactions related to some outcome in complicated situations such as health care, which enables the black-box predictive
models for knowledge discovery.
5
CONCLUSION AND DISCUSSION
In this paper, we propose a simple but effective method to interpret
black-box machine learning models globally from local explanations of single data samples. Global interpretation is more refined
than local explanations thus more efficient when used to diagnose
the trained model or extract knowledge from it. We show that our
Global Interpretation via Recursive Partitioning (GIRP) algorithm
can represent many types of machine learning models in a compact manner. We demonstrate our algorithms using various kinds
of machine learning models on different tasks. We have shown
that the deep residual network is looking for the right object when
classifying scenes. In contrast, in text classification, the interpretation tree indicates that the random forest classifier is focusing on
wrong words to discriminate texts with different topics. Besides,
the proposed method is also useful to extract decision rules from
sophisticated models. Such rules are hidden in black-box models
but are critical to know if we want to impact the outcome. We
showcase this usage by extracting disease comorbidities leading to
high mortality in intensive care unit. In conclusion, our method
helps people understand machine learning models efficiently, making it easier to check if the model is behaving reasonably and make
use of the knowledge it discovers.
However, the proposed method is limited in several ways. First,
we lack a quantitative measure of the fidelity of the interpretation
tree to the original explained machine learning model. Though
the interpretation tree is directly developed from contributions
generated from the original model, we lose some details when we
extract the general rules. We don’t know how important these
details are to the high predictability. Second, though we present
the split strength as a measure of variable importance in the interpretation tree, the confidence for this measure is unknown. Linear
methods are popular in evidence based studies partially because it
is easier for confidence interval estimation. Due to the complexity
of underlying probabilistic distributions for sophisticated machine
learning methods, it is difficult to estimate confidence intervals
for the split strengths in the interpretation tree. Finally, the proposed method needs a contribution matrix as an input which is
difficult to obtain when feature representation from input variables
is not well established such as speech recognition. This difficulty
is closely connected to the broader problem and theories such as
the information bottleneck [25, 26] of understanding how machine
learning models identify important high level features and ignore
the noises. Sometimes, even the explicit feature representation is
available to form the contribution matrix, group effects of features
are not well captured in the current method. Mechanisms similar
to group LASSO [31] could be added to solve this problem.
Each of the limitations mentioned points to a good direction for
future work. We want to quantify the fidelity of the interpretation
tree to the explained original model. We are considering setting up
bootstrapping methods [10] for confidence interval estimation as
direct probabilistic distribution estimation is difficult. At last, some
representation learning methods could be incorporated into the
algorithm when the contribution matrix is difficult to obtain. All
these pose exciting challenges for making machine learning more
transparent for human beings.
REFERENCES
[1] Susan Athey and Guido Imbens. 2016. Recursive partitioning for heterogeneous
causal effects. Proceedings of the National Academy of Sciences 113, 27 (2016),
7353–7360.
[2] Martin Atzmueller and Thomas Roth-Berghofer. 2011. The mining and analysis
continuum of explaining uncovered. In Research and Development in Intelligent
Systems XXVII. Springer, 273–278.
[3] David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja
Hansen, and Klaus-Robert MÞller. 2010. How to explain individual classification
decisions. Journal of Machine Learning Research 11, Jun (2010), 1803–1831.
[4] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint
arXiv:1409.0473 (2014).
[5] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. 2017.
Network Dissection: Quantifying Interpretability of Deep Visual Representations.
arXiv preprint arXiv:1704.05796 (2017).
[6] L. Breiman, J. Friedman, R. Olshen, and C. Stone. 1984. Classification and Regression Trees. Wadsworth and Brooks, Monterey, CA.
[7] Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie
Elhadad. 2015. Intelligible models for healthcare: Predicting pneumonia risk and
hospital 30-day readmission. In Proceedings of the 21th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining. ACM, 1721–1730.
[8] Edward Choi, Mohammad Taha Bahadori, Jimeng Sun, Joshua Kulas, Andy
Schuetz, and Walter Stewart. 2016. RETAIN: An Interpretable Predictive Model
for Healthcare using Reverse Time Attention Mechanism. In Advances in Neural
Information Processing Systems. 3504–3512.
[9] Mark Craven and Jude W Shavlik. 1996. Extracting tree-structured representations of trained networks. In Advances in neural information processing systems.
24–30.
[10] Bradley Efron. 1979. Bootstrap methods: another look at the jackknife. The
annals of Statistics (1979), 1–26.
[11] Mark A Hall. 1999. Correlation-based feature selection for machine learning. Ph.D.
Dissertation. The University of Waikato.
[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual
learning for image recognition. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition. 770–778.
[13] Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng,
Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi,
and Roger G Mark. 2016. MIMIC-III, a freely accessible critical care database.
Scientific data 3 (2016).
[14] Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via
influence functions. arXiv preprint arXiv:1703.04730 (2017).
[15] Igor Kononenko et al. 2010. An efficient explanation of individual classifications
using game theory. Journal of Machine Learning Research 11, Jan (2010), 1–18.
[16] Benjamin Letham, Cynthia Rudin, Tyler H McCormick, David Madigan, et al.
2015. Interpretable classifiers using rules and Bayesian analysis: Building a better
stroke prediction model. The Annals of Applied Statistics 9, 3 (2015), 1350–1371.
[17] Andy Liaw and Matthew Wiener. 2002. Classification and regression by randomForest. R news 2, 3 (2002), 18–22.
[18] James N Morgan and John A Sonquist. 1963. Problems in the analysis of survey
data, and a proposal. Journal of the American statistical association 58, 302 (1963),
415–434.
[19] Zahra Mungloo-Dilmohamud, Yasmina Jaufeerally-Fakim, and Carlos PeñaReyes. 2017. A Meta-Review of Feature Selection Techniques in the Context of
Microarray Data. In International Conference on Bioinformatics and Biomedical
Engineering. Springer, 33–49.
[20] Anh Nguyen, Jason Yosinski, and Jeff Clune. 2015. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition. 427–436.
[21] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. Why Should I
Trust You?: Explaining the Predictions of Any Classifier. In Proceedings of the
22nd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining. ACM, 1135–1144.
[22] Karen Sparck Jones. 1972. A statistical interpretation of term specificity and its
application in retrieval. Journal of documentation 28, 1 (1972), 11–21.
[23] Xiaogang Su, Chih-Ling Tsai, Hansheng Wang, David M Nickerson, and Bogong
Li. 2009. Subgroup analysis via recursive partitioning. Journal of Machine
Learning Research 10, Feb (2009), 141–158.
[24] Robert Tibshirani. 1996. Regression shrinkage and selection via the lasso. Journal
of the Royal Statistical Society. Series B (Methodological) (1996), 267–288.
[25] Naftali Tishby, Fernando C Pereira, and William Bialek. 2000. The information
bottleneck method. arXiv preprint physics/0004057 (2000).
[26] Naftali Tishby and Noga Zaslavsky. 2015. Deep learning and the information
bottleneck principle. In Information Theory Workshop (ITW), 2015 IEEE. IEEE,
1–5.
[27] Zhixiang Xu, Gao Huang, Kilian Q Weinberger, and Alice X Zheng. 2014. Gradient
boosted feature selection. In Proceedings of the 20th ACM SIGKDD international
conference on Knowledge discovery and data mining. ACM, 522–531.
[28] Chengliang Yang, Chris Delcher, Elizabeth Shenkman, and Sanjay Ranka. 2016.
Predicting 30-day all-cause readmissions from hospital inpatient discharge data.
In e-Health Networking, Applications and Services (Healthcom), 2016 IEEE 18th
International Conference on. IEEE, 1–6.
[29] Chengliang Yang, Chris Delcher, Elizabeth Shenkman, and Sanjay Ranka. 2017.
Machine Learning Approaches for Predicting High Utilizers in Health Care. In
International Conference on Bioinformatics and Biomedical Engineering. Springer,
382–395.
[30] Fisher Yu and Vladlen Koltun. 2015. Multi-scale context aggregation by dilated
convolutions. arXiv preprint arXiv:1511.07122 (2015).
[31] Ming Yuan and Yi Lin. 2006. Model selection and estimation in regression with
grouped variables. Journal of the Royal Statistical Society: Series B (Statistical
Methodology) 68, 1 (2006), 49–67.
[32] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Antonio Torralba, and Aude Oliva.
2016. Places: An image database for deep scene understanding. arXiv preprint
arXiv:1610.02055 (2016).
[33] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio
Torralba. 2017. Scene parsing through ade20k dataset. In Proc. CVPR.
| 2cs.AI
|
A parallel algorithm for the constrained shortest
path problem on lattice graphs
arXiv:1511.06441v2 [math.OC] 13 Dec 2017
Ivan Matic
Published in: Adamatzky, A (Ed.) Shortest path solvers. From software to wetware.
Springer, 2018.
Abstract The edges of a graph are assigned weights and passage times which are
assumed to be positive integers. We present a parallel algorithm for finding the shortest path whose total weight is smaller than a pre-determined value. In each step the
processing elements are not analyzing the entire graph. Instead they are focusing on
a subset of vertices called active vertices. The set of active vertices at time t is related to the boundary of the ball Bt of radius t in the first passage percolation metric.
Although it is believed that the number of active vertices is an order of magnitude
smaller than the size of the graph, we prove that this need not be the case with an
example of a graph for which the active vertices form a large fractal. We analyze an
OpenCL implementation of the algorithm on GPU for cubes in Zd .
1 Definition of the problem
The graph G(V, E) is undirected and the function f : E → Z2+ is defined on the set
of its edges. The first component f1 (e) of the ordered pair f (e) = ( f1 (e), f2 (e)) for
a given edge e ∈ E represents the time for traveling over the edge e. The second
component f2 (e) represents the weight of e.
A path in the graph G is a sequence of vertices (v1 , v2 , . . . , vk ) such that for each
i ∈ {1, 2, . . . , k − 1} there is an edge between vi and vi+1 , i.e. (vi , vi+1 ) ∈ E. For each
path π = (v1 , . . . , vk ) we define F1 (π) as the total time it takes to travel over π and
F2 (π) as the sum of the weights of all edges in π. Formally,
k−1
F1 (π) =
∑ f1 (vi , vi+1 )
i=1
k−1
and
F2 (π) =
∑ f2 (vi , vi+1 ) .
i=1
Ivan Matic
Department of Mathematics, Baruch College, CUNY, One Bernard Baruch Way, New York, NY
10010, USA e-mail: [email protected]
1
2
Ivan Matic
Let A, B ⊆ V be two fixed disjoint subsets of V and let M ∈ R+ be a fixed positive
real number. Among all paths that connect sets A and B let us denote by π̂ the one
(or one of) for which F1 (π) is minimal under the constraint F2 (π) < M. We will
describe an algorithm whose output will be F1 (π̂) for a given graph G.
The algorithm belongs to a class of label correcting algorithms [11, 20]. The
construction of labels will aim to minimize the memory consumption on SIMD
devices such as graphic cards. Consequently, the output will not be sufficient to
determine the exact minimizing path. The reconstruction of the minimizing path is
possible with subsequent applications of the method, because the output can include
the vertex X ∈ B that is the endpoint of π̂, the last edge x on the path π̂, and the value
F2 (π̂). Once X and x are found, the entire process can be repeated for the graph
G0 (V 0 , E 0 ) with
V 0 = V \ B,
A0 = A,
B0 = {X},
and
M 0 = F2 (π̂) − f2 (x).
The result will be second to last vertex on the minimizing path π̂. All other vertices
on π̂ can be found in the same way.
Although the algorithm works for general graphs and integer-valued functions f ,
its implementation on SIMD hardware requires the vertices to have bounded degree.
This requirement is satisfied by subgraphs of Zd .
Finding the length of the shortest path in graph is equivalent to finding the shortest passage time in first passage percolation. Each of the vertices in A can be thought
of as a source of water. The value f1 (e) of each edge e is the time it takes the water
to travel over e. Each drop of water has its quality and each drop that travels through
edge e looses f2 (e) of its quality. Each vertex P of the graph has a label Label(P)
that corresponds to the quality of water that is at the vertex P. Initially all vertices in
A have label M while all other vertices have label 0. The drops that get their quality
reduced to 0 cannot travel any further. The time at which a vertex from B receives
its first drop of water is exactly the minimal F1 (π) under the constraint F2 (π) < M.
Some vertices and edges in the graph are considered active. Initially, the vertices
in A are active. All edges adjacent to them are also called active. Each cycle in algorithm corresponds to one unit of time. During one cycle the water flows through
active edges and decrease their time components by 1. Once an edge gets its time
component reduced to 0, the edge becomes used and we look at the source S and
the destination D of this water flow through the edge e. The destination D becomes
triggered, and its label will be corrected. The label correction is straight-forward if
the edge D was inactive. We simply check whether Label(S) − f2 (e) > Label(D),
and if this is true then the vertex D gets its label updated to Label(S) − f2 (e) and
its status changed to active. If the vertex D was active, the situation is more complicated, since the water has already started flowing from the vertex D. The existing
water flows correspond to water of quality worse than the new water that has just arrived to D. We resolve this issue by introducing phantom edges to the graph that are
parallel to the existing edges. The phantom edges will carry this new high quality
water, while old edges will continue carrying their old water flows. A vertex stops
A parallel algorithm for the constrained shortest path problem on lattice graphs
3
being active if all of its edges become used, but it may get activated again in the
future.
2 Related problems in the literature
The assignment of phantom edges to the vertices of the graph and their removal is
considered a label correcting approach in solving the problem. Our particular choice
of label correction is designed for large graphs in which the vertices have bounded
degree. Several existing serial computation algorithms can find the shortest path by
maintaining labels for all vertices. The labels are used to store the information on the
shortest path from the source to the vertex and additional preprocessing of vertices
is used to achieve faster implementations [5, 9]. The ideas of first passage percolation and label correction have naturally appeared in the design of pulse algorithms
for constrained shortest paths [17]. All of the mentioned algorithms can also be parallelized but this task would require a different approach in designing a memory
management that would handle the label sets in programming environments where
dynamical data structures need to be avoided.
The method of aggressive edge elimination [22] can be parallelized to solve the
Lagrange dual problems. In the case of road and railroad networks a substantial
speedup can be achieved by using a preprocessing of the network data and applying
a generalized versions of Dijkstra’s algorithm [12].
The parallel algorithm that is most similar in nature to the one discussed in this
paper is developed for wireless networks [16]. There are two features of wireless
networks that are not available to our model. The first feature is that the communication time between the vertices can be assumed to be constant. The other feature is
that wireless networks have a processing element available to each vertex. Namely,
routers are usually equipped with processors. Our algorithm is build for the situations where the number of processing cores is large but not at the same scale as
the number of vertices. On the other hand our algorithm may not be effective for
the wireless networks since the underlying graph structure does not imply that the
vertices are of bounded degree. The increase of efficiency of wireless networks can
be achieved by solving other related optimization problems. One such solution is
based on constrained node placement [21].
The execution time of the algorithm is influenced by the sizes of the sets of active vertices, active edges, and phantom edges. The sizes of these sets are order of
magnitude smaller than the size of the graph. Although this cannot be proved at the
moment, we will provide a justification on how existing conjectures and theorems
from the percolation theory provide some estimates on the sizes of these sets. The
set of active vertices is related to the limit shape in the model of first passage percolation introduced by Hammersley and Welsh [10]. The first passage percolation
corresponds to the case M = ∞, i.e. the case when there are no constraints. If we
assume that A = {0}, for each time t we can define the ball of radius t in the first
passage percolation metric as:
4
Ivan Matic
Bt = {x : τ(0, x) ≤ t} ,
where τ(0, x) is the first passage time, i.e. the first time at which the vertex x is
reached.
The active vertices at time t are located near the boundary of the ball Bt . It is
known that for large t the set 1t Bt will be approximately convex. More precisely, it
is known [7] that there is a convex set B such that
1
P (1 − ε)B ⊆ Bt ⊆ (1 + ε)B for large t = 1.
t
However, the previous theorem does not guarantee that the boundary of Bt has
to be of zero volume. In fact the boundary can be non-polygonal as was previously
shown [8].
The set of active vertices does not coincide with the boundary of Bt , but it is
expected that if ∂ Bt is of small volume then the number of active vertices is small
in most typical configurations of random graphs. We provide an example for which
the set of active vertices is a large fractal, but simulations suggest that this does not
happen in average scenario.
The fluctuations of the shape of Bt are expected to be of order t 2/3 in the case of
2
Z and the first passage time τ(0, n) is proven to have fluctuations of order at least
log n [23]. The fluctuations are of order at most n/ log n [4, 3] and are conjectured
to be of order t 2/3 . They can be larger and of order n for modifications of Z2 known
as thin cylinders [6].
The scaling of t 2/3 for the variance is conjectured for many additional interface
growth models and is related to the Kardar-Parisi-Zhang equation [1, 15, 25].
The constrained first passage percolation problem is a discrete analog to HamiltonJacobi equation. The large time behaviors of its solutions are extensively studied
and homogenization results are obtained for a class of Hamiltonians [2, 13, 14, 26].
Fluctuations in dimension one are of order t [24] while in higher dimensions they
are of lower order although only the logarithmic improvement to the bound has been
achieved so far [19].
3 Example
Before providing a more formal description of the algorithm we will illustrate the
main ideas on one concrete example of a graph. Consider the graph shown in Figure
1 that has 12 vertices labeled as 1, 2, . . ., 12. The set A contains the vertices 1, 2,
and 3, and the set B contains only the vertex 12. The goal is to find the length of the
shortest path from A to B whose total weight is smaller than 19.
The vertices are drawn with circles around them. The circles corresponding to
the vertices in A are painted in blue and have the labels 19. The picture contains the
time and weight values for each of the edges. The time parameter of each edge is
written in the empty oval, while the weight parameter is in the shaded oval. Since
A parallel algorithm for the constrained shortest path problem on lattice graphs
5
Fig. 1 The initial state of the graph.
the number of edges in this graph is relatively small it is not difficult to identify the
the minimizing path (3, 6, 10, 11, 12). The time required to travel over this path is
16 and the total weight is 15.
Initially, the vertices in set A are called active. Active vertices are of blue color
and edges adjacent to them are painted in blue. These edges are considered active.
Numbers written near their centers represent the sources of water. For example, the
vertex 2 is the source of the flow that goes through the edge (2, 5).
Notice that the smallest time component of all active edges is 2. The first cycle
of the algorithm begins by decreasing the time component of each active edge by
2. The edge (1, 4) becomes just used because its time component is decreased to 0.
The water now flows from the vertex 1 to the vertex 4 and its quality decreases by
5, since the weight of the edge (1, 4) is equal to 5. The vertex 4 becomes active and
its label is set to
Label(4) = Label(1) − f2 (1, 4) = 19 − 5 = 14.
The edge (1, 4) becomes used, and the vertex 1 turns into inactive since there are no
active edges originating from it. Hence, after two seconds the graph turns into the
one shown in Figure 2.
The same procedure is repeated until the end of the 5th second and the obtained
graph is the top one in Figure 3. In the 6th second the edge (2, 5) gets its time
parameter decreased to 0 and the vertex 5 gets activated. Its label becomes
6
Ivan Matic
Fig. 2 The configuration after the second 2.
Label(5) = Label(2) − f2 (2, 5) = 19 − 2 = 17.
However, the edges (4, 5), (5, 9), and (5, 6) were already active and the water was
flowing through them towards the vertex 5.
The old flow of water through the edge (4, 5) will complete in additional 5 seconds. However, when it completes the quality of the water that will reach the vertex
5 will be
Label(4) − f2 (4, 5) = 14 − 2 = 12 < Label(5)
because the label of the vertex 5 is 17. Thus there is no point in keeping track of this
water flow. On the other hand, the water flow that starts from 5 and goes towards 4
will have quality
Label(5) − f2 (4, 5) = 17 − 2 = 15
which is higher than the label of the vertex 4. Thus the edge (4, 5) will change its
source from 4 to 5 and the time parameter has to be restored to the old value 7. At
this point the vertex 4 becomes inactive as there is no more flow originating from it.
The same reversal of the direction of the flow happens with the edge (5, 9). On
the other hand, something different happens to the edge (5, 6): it stops being active.
The reason is that the old flow of water from 6 to 5 will not be able to increase the
label of the vertex 5. Also, the new flow of water from 5 to 6 would not be able to
change the label of vertex 6.
A special care has to be taken when a water flow reaches a vertex that is already
active. In the case of the graph G such situation happens after the 11th second.
A parallel algorithm for the constrained shortest path problem on lattice graphs
7
Fig. 3 The configurations after the seconds 5 and 6.
The configuration is shown in the top picture of Figure 4. The edge (7, 11) has the
smallest time parameter 2. The time will progress immediately to 13 and all active
edges get their time parameters decreased by 2. In the 13th second the water from
the edge (7, 11) reaches the vertex 11. The label of vertex 7 is Label(7) = 11, while
Label(11) = 6. The weight of the flow over the edge between these two vertices is
2, hence this new water is of higher quality than the one present at the vertex 11.
In this situation we consider every active edge originating from 11 and create a
phantom edge through which this new water will flow. We will create a new vertex
110 with label
Label(110 ) = Label(7) − f2 (7, 11) = 9
8
Ivan Matic
Fig. 4 The configurations after the seconds 11 and 13.
and connect it with each of the neighbors of 11 that can get their label increased
with the new flow. The only one such neighbor is 12 and we obtain the graph as
shown in the lower part of Figure 4.
It can be now easily verified that after additional 3 seconds, i.e. in the end of
the second 16 the vertex 12 becomes active with the label 4. Thus we conclude that
it takes water to travel 16 seconds over the shortest path. The total weight of the
shortest path is 19 − 4 = 15. The minimizing path is (3, 6, 10, 11, 12).
A parallel algorithm for the constrained shortest path problem on lattice graphs
9
4 Pseudo-code of the algorithm
We will organize the algorithm by dividing it into smaller components. The first
component is the initialization, and the others are performed in the main loop that
consists of 9 major steps. The parallelizations will happen only in these individual
steps of the main loop. The pseudo-code for the main function is presented in Algorithm 1. Each step will be described in full details and accompanied by a pseudocode that outlines the main ideas. For the sake of brevity, some data structures in
pseudo-code will be modeled with sets. However, the usage of sets is avoided as they
cannot support insertion and deletion of elements in parallel. The sets are replaced
by indicator sequences for which appropriate operations are easier to parallelize.
For the full source code the reader is referred to [18].
Algorithm 1 Main function
Input: Graph G = (V, E); A, B ⊂ G such that A ∩ B = 0,
/ M ∈ R, two functions f1 , f2 : E → R.
f1 (e) is the time to travel over the edge e and f2 (e) is the weight of the edge e.
Output: The shortest time to travel from A to B over a path whose weight is less than M.
1: function MAIN
2:
Initialization
3:
L =ShortestTravelTimeAndTerminalConditionCheck
4:
while L = 0 do
5:
TriggerVertices
6:
AnalyzeTriggeredVertices
7:
GetInputFromPhantoms
8:
TriggerEdges
9:
TreatTriggeredEdges
10:
L =ShortestTravelTimeAndTerminalConditionCheck
11:
FinalTreatmentOfPhantoms
12:
FinalTreatmentOfVertices
13:
FinalTreatmentOfActiveEdges
14:
return L
5 Memory management and initialization
5.1 Labels for vertices and edges
In this section we will describe the memory management of variables necessary for
the implementation of the algorithm. Before providing the precise set of variables let
us describe the information that has to be carried throughout the execution process.
As we have seen before, the vertices will have labels assigned to them. Initially we
label each vertex of G with 0 except for vertices in A which are labeled by M.
10
Ivan Matic
To each vertex and edge in G we assign a State. The vertices have states in the set
{active, inactive}. Initially all vertices in A are active, while the other vertices are
inactive. The states of the edges belong to the set {active, passive, used, just used}.
Initially the edges adjacent to the vertices in A are set to active while all other are
passive.
To each edge we associate a pointer to one of its endpoints and call it Source.
This variable is used at times when the water is flowing through the edge and it
records the source of the current water flow. Initially, to each edge that originates
from a vertex in A we set the source to be the pointer to the vertex in A. All other
edges have their source initially set to 0.
There is additional variable Time that represents the time and is initially set to 0.
5.2 Termination
The algorithm terminates if one of the following two conditions is satisfied:
1◦ A vertex from B becomes active. The variable Time contains the time it takes to
reach this vertex along the shortest path π̂, i.e.
Time = F1 (π̂) .
The label of the last vertex B̂ on the path allows us to determine the value F2 (π̂).
Namely,
F2 (π̂) = M − Label(B̂).
We will not go into details on how to recover the exact shortest path. Instead we
will just outline how this can be done. We need to identify the used edge f (or
one of the used edges, if there are more than one) that is adjacent to B̂. This edge
can help us in finding the second to last point of the path π̂. Let us denote by F
the other endpoint of f . It could happen that F is a phantom vertex (i.e. a copy of
another vertex), and we first check whether F ∈ PhantomVertices. If this is not the
case, then F is the second to last element of the path π̂. If F ∈ PhantomVertices
then the vertex F is a copy of some other vertex in the graph and the phantom
vertex F has the pointer to the original based on which it is created. This original
vertex is the second to last point on the path π̂.
2◦ There is no active edge in the graph. In this case there is no path that satisfies the
constraint F2 ≤ M.
A parallel algorithm for the constrained shortest path problem on lattice graphs
11
Algorithm 2 Function that checks whether the algorithm has finished and returns
the time for travel over the shortest path
1: function S HORTEST PATH L ENGTH A ND T ERMINAL C ONDITION C HECK
2:
// Returns 0 if the path is not found yet.
3:
// Returns −1 if there is no path with weight smaller than M.
4:
// Returns the weight of the shortest path if it is found.
5:
// A non-zero return value is the indication that the algorithm is over.
6:
#Performed in parallel
7:
if ∃B0 ∈ B such that B0 = active then
8:
result ← Time
9:
else
10:
if there are no active vertices then
11:
result ← −1
12:
else
13:
result ← 0
14:
#barrier
15:
return result
5.3 Sequences accessible to all processing elements
It is convenient to store the vertices and edges in sequences accessible to all processing elements. We will assume here that the degree of each vertex is bounded
above by d.
5.3.1 Vertices
Each vertex takes 5 integers in the sequence of vertices. The first four are name,
label, status, and the location of the first edge in the sequence of edges. The fifth
element is be used to store a temporary replacement label. Initially, and between
algorithm steps, this label is set to −1.
When a first drop of water reaches an inactive vertex V , we say that the vertex
is triggered, and that state exists only temporarily during an algorithm cycle. In the
end of the algorithm cycle some triggered vertices become active. However it could
happen that a triggered vertex does not get a water flow of higher quality than the
one already present at the vertex. The particular triggered vertex with this property
does not get activated.
5.3.2 Edges
Each edge e takes 8 integers in the sequence of edges. Although the graph is undirected, each edge is stored twice in the memory. The 8 integers are the start point,
the end point, remaining time for water to travel over the edge (if the edge is active),
the weight of the travel f2 (e), the initial passage time f1 (e), the label of the vertex
12
Ivan Matic
that is the source of the current flow through the edge (if there is a flow), status, and
the location of the same edge in the opposite direction.
5.3.3 Active vertices
The sequence contains the locations of the vertices that are active. This sequence
removes the need of going over all vertices in every algorithm step. The locations
are sorted in decreasing order. In the end of the sequence we will add triggered
vertices that will be joined to the active vertices in the end of the cycle.
5.3.4 Active edges
The role of the sequence is similar to the one of active vertices. The sequence maintains the location of the active edges. Each edge is represented twice in this sequence. The second appearance is the one in which the endpoints are reversed. The
locations are sorted in decreasing order. During the algorithm cycle we will append
the sequence with triggered edges. In the end of each cycle the triggered edges will
be merged to the main sequence of active edges.
5.3.5 Sequence of phantom edges
The phantom edges appear when an active vertex is triggered with a new drop of water. Since the vertex is active we cannot relabel the vertex. Instead each of the edges
going from this active triggered vertex need to be doubled with the new source of
water flowing through these new edges that are called phantoms. They will disappear once the water finishes flowing through them.
5.3.6 Sequence of elements in B
Elements in B have to be easily accessible for quick check whether the algorithm
has finished. For this reason the sequence should be in global memory.
Listing 3 summarizes the initializing procedures.
6 Graph update
The algorithm updates the graph in a loop until one vertex from B becomes active.
Each cycle consists of the following nine steps.
A parallel algorithm for the constrained shortest path problem on lattice graphs
13
Algorithm 3 Initialization procedure
1: procedure I NITIALIZATION
2:
for e ∈ E do
3:
State(e) ← passive
4:
Source(e) ← 0
5:
TimeRemaining(e) ← 0
6:
for v ∈ V \ A do
7:
State(v) ← inactive
8:
Label(v) ← 0
9:
for v ∈ A do
10:
State(v) ← active
11:
Label(v) ← M
12:
for e ∈ Edges(v) do
13:
State(e) ← active
14:
Source(e) ← v
15:
TimeRemaining(e) ← f1 (e)
16:
TriggeredVertices ← 0/
17:
PhantomVertices ← 0/
18:
PhantomEdges ← 0/
19:
Time ← 0
6.1 Step 1: Initial triggering of vertices
In this step we go over all active edges and decrease their time parameters by m,
where m is the smallest remaining time of all active edges. If for any edge the time
parameter becomes 0, the edge becomes just used and its destination triggered.
To avoid the danger of two processing elements writing in the same location of
the sequence of active vertices, we have to make sure that each processing element
that runs concurrently has pre-specified location to write. This is accomplished by
first specifying the number of threads in the separate variable nThreads. Whenever
kernels are executed in parallel we are using only nThreads processing elements.
Each processing element has its id number which is used to determine the memory
location to which it is allowed to write. The sequence of triggered vertices has to be
cleaned after each parallel execution and at that point we take an additional step to
ensure we don’t list any of the vertices as triggered twice.
6.2 Step 2: Analyzing triggered vertices
For each triggered vertex Q we look at all of its edges that are just used. We identify
the largest possible label that can result from one of just used edges that starts from
Q. That label will be stored in the sequence of vertices at the position reserved for
temporary replacement label. The vertex is labeled as just triggered. If the vertex
Q is not active, this label will replace the current label of the vertex in one of the
14
Ivan Matic
Algorithm 4 Procedure TriggerVertices
1: procedure T RIGGERV ERTICES
2:
TriggeredEdges ← 0/
3:
#Performed in parallel
4:
m ← min {TimeRemaining(e) : e ∈ ActiveEdges}
5:
#barrier
6:
Time ← Time + m
7:
#Performed in parallel
8:
for e ∈ ActiveEdges do
9:
TimeRemaining(e) ← TimeRemaining(e) − m
10:
if TimeRemaining(e) = 0 then
11:
State(e) = just used
12:
Se ← Source(e)
13:
De ← TheTwoEndpoints(e) \ {Se }
14:
TriggeredVertices ← TriggeredVertices ∪ {De }
15:
#barrier
later steps. If the vertex Q is active, then this temporary label will be used later to
construct an appropriate phantom edge.
We are sure that different processing elements are not accessing the same vertex
at the same time, because before this step we achieved the state in which there are
no repetitions in the sequence of triggered vertices.
Algorithm 5 Analysis of triggered vertices
1: procedure A NALYZE T RIGGEREDV ERTICES
2:
TempLabel ← 0/
3:
#Performed in parallel
4:
for Q ∈ TriggeredVertices do
5:
TempLabel(Q) ← max {Label(P) − f2 (P, Q) : State(P, Q) = just used}
6:
#barrier
6.3 Step 3: Gathering input from phantoms
The need to have this step separated from the previous ones is the current architecture of graphic cards that creates difficulties with dynamic memory locations. It
is more efficient to keep phantom edges separate from the regular edges. The task
is to look for all phantom edges and decrease their time parameters. If a phantom
edge gets its time parameter equal to 0, its destination is studied to see whether it
should be added to the sequence of triggered vertices. We calculate the new label
that the vertex would receive through this phantom. We check whether this new label is higher than the currently known label and the temporary label from possibly
previous triggering of the vertex. The phantoms will not result in the concurrent
A parallel algorithm for the constrained shortest path problem on lattice graphs
15
writing to memory locations because each possible destination of a phantom could
have only one edge that has time component equal to 0.
Algorithm 6 Input from phantoms
1: procedure G ET I NPUT F ROM P HANTOMS
2:
#Performed in parallel
3:
Decrease time parameters of fantom edges (as in Listing 4)
4:
Trigger the destinations of phantom edges (as in Listing 4)
5:
#barrier
6:
#Performed in parallel
7:
Analyze newly triggered vertices, in a way similar to Listing 5
8:
#barrier
6.4 Step 4: Triggering edges
In this step we will analyze the triggered vertices and see whether each of their
neighboring edges needs to change the state. Triggered vertices are analyzed using
separate processing elements. A processing element analyzes the vertex Q in the
following way.
Each edge j of Q will be considered triggered if it can cause the other endpoint to
get better label in future through Q. The edge j is placed in the end of the sequence
of active edges.
Algorithm 7 Procedure that triggers the edges
1: procedure T RIGGER E DGES
2:
#Performed in parallel
3:
for Q ∈ TriggeredVertices do
4:
for P ∈ Neighbors(Q) do
5:
if TempLabel(Q) − f2 (P, Q) > Label(P) then
6:
State(P, Q) ← active
7:
TriggeredEdges ← TriggeredEdges ∪ {(P, Q)}
8:
#barrier
6.5 Step 5: Treatment of triggered edges
Consider a triggered edge j. We first identify its two endpoints. For the purposes of
this step we will identify the endpoint with the larger label, call it the source, and
denote by S. The other will be called the destination and denoted by D. In the end
of the cycle, this vertex S will become the source of the flow through j.
16
Ivan Matic
Notice that at least one of the endpoints is triggered. If only one endpoint is
triggered, then we are sure that this triggered endpoint is the one that we designated
as the source S.
We then look whether the source S was active or inactive before it was triggered.
6.5.1 Case in which the source S was inactive before triggering
There are several cases based on the prior status of j. If j was passive, then it should
become active and no further analysis is necessary. If it was used or just used, then
it should become active and the time component should be restored to the original
one. Assume now that the edge j was active. Based on the knowledge that S was
inactive vertex we can conclude that the source of j was D. However we know that
the source of j should be S and hence the time component of j should be restored to
the backup value.
Consequently, in the case that S was inactive, regardless of what the status of j
was, we are sure its new status must be active and its time component can be restored
to the original value. This restoration is not necessary in the case that j was passive,
although there is no harm in doing it.
If the edge j was not active before, then the edge j should be added to the list of
active edges. If the edge j was active before, then it should be removed from the list
of triggered edges because all triggered edges will be merged into active edges. The
edge j already appears in the list of active edges and need not be added again.
6.5.2 Case in which the source S was active before triggering
In this case we create phantom edges. Each such triggered edge generates four entries in the phantom sequence. The first one is the source, the second is the destination, the third is the label of the source (or the label stored in the temporary label
slot, if higher), and the fourth is the original passage time through the edge j.
6.6 Step 6: Checking terminal conditions
In this step we take a look whether a vertex from B became active or if there are
no active edges. These would be the indications of the completion of the algorithm.
The function that checks the terminal conditions is presented earlier in Listing 2.
A parallel algorithm for the constrained shortest path problem on lattice graphs
17
Algorithm 8 Treatment of triggered edges
1: procedure T REAT T RIGGERED E DGES
2:
#Performed in parallel
3:
for j ∈ TriggeredEdges do
4:
S ← The endpoint of j with larger label
5:
D ← The endpoint of j with smaller label
6:
OldStateOfS ← State(S)
7:
OldStateOfJ ← State( j)
8:
if OldStateOfS = inactive then
9:
State( j) ← active
10:
Source( j) ← S
11:
TimeRemaining( j) ← f1 ( j)
12:
if OldStateOfS = active then
13:
Create a phantom vertex S0 and connect it to D
14:
TimeRemaining(S0 , D) ← f1 (S, D)
15:
#barrier
6.7 Step 7: Final treatment of phantoms
In this step we go once again over the sequence of phantoms and remove each one
that has its time parameter equal to 0.
Algorithm 9 Final treatment of phantoms
1: procedure F INALT REATMENT O F P HANTOMS
2:
#Performed in parallel
3:
for j ∈ PhantomEdges do
4:
if TimeRemaining( j) = 0 then
5:
Remove j and its source from the sequence of phantoms
6:
#barrier
6.8 Step 8: Final treatment of vertices
In this step of the program the sequence of active vertices is updated so it contains
new active vertices and looses the vertices that may cease to be active.
6.8.1 Preparation of triggered vertices
For each triggered vertex Q we first check whether it was inactive before. If it was
inactive then its label becomes equal to the label stored at the temporary storing
location in the sequence of vertices. If it was active, its label remains unchanged.
18
Ivan Matic
The phantoms were created and their labels are keeping track of the improved water
quality that has reached the vertex Q.
We may now clean the temporary storing location in the sequence of vertices so
it now contains the symbol for emptiness (some pre-define negative number).
6.8.2 Merging triggered with active vertices
Triggered vertices are now merged to the sequence of active vertices.
6.8.3 Check active vertices for potential loss of activity
For each active vertex Q look at all edges from Q. If there is no active edge whose
source is Q, then Q should not be active any longer.
6.8.4 Condensing the sequence of active vertices
After previous few steps some vertices may stop being active in which case they
should be removed from the sequence.
Algorithm 10 Final treatment of vertices
1: procedure F INALT REATMENT O F V ERTICES
2:
#Performed in parallel
3:
for Q ∈ TriggeredVertices do
4:
if State(Q) = inactive then
5:
State(Q) ← active
6:
Label(Q) ← TempLabel(Q)
7:
#barrier
8:
TempLabel ← 0/
9:
#Performed in parallel
10:
Merge triggered vertices to active vertices
11:
#barrier
12:
#Performed in parallel
13:
for Q ∈ TriggeredVertices do
14:
if there are no active edges starting from Q then
15:
State(Q) ← inactive
16:
#barrier
A parallel algorithm for the constrained shortest path problem on lattice graphs
19
6.9 Step 9: Final treatment of active edges
We first need to merge the triggered edges with active edges. Then all just used
edges have to become used and their source has to be re-set so it is not equal to any
of the endpoints. Those used edges should be removed from the sequence of active
edges.
The remaining final step is to condense the obtained sequence so there are no
used edges in the sequence of active edges.
Algorithm 11 Final treatment of active edges
1: procedure F INALT REATMENT O FACTIVE E DGES
2:
#Performed in parallel
3:
Merge triggered edges to active edges
4:
#barrier
5:
#Performed in parallel
6:
Transform all just used into used and erase their Source components
7:
#barrier
7 Large sets of active vertices
In this section we will prove that it is possible for the set of active vertices in dimension 2 to contain more than O(n) elements. We will construct examples in the case
when the time to travel over each vertex is from the set {1, 2} and when M = +∞.
We will consider the subgraph Vn = [−n, n] × [0, n] of Z2 . At time 0 the water
is located in all vertices of the x axis. For sufficiently large n we will provide an
example of configuration ω of passage times for the edges of the graph Vn such that
the number of active vertices at time n is of order n log n. This would establish a
lower bound on the probability that the number of active vertices at time t is large.
Let us assume that each edge of the graph has the time component assigned from
the set {1, 2} independently from each other. Assume that the probability that 1 is
assigned to each edge is equal to p, where 0 < p < 1.
Theorem 1. There exists t0 ≥ 0, µ > 0, and α > 0 such that for each t > t0 there
exists n such that the number At of active vertices at time t in the graph Vn satisfies
2
P (At ≥ αt logt) ≥ e−µt .
To prepare for the proof of the theorem we first study the evolution of the set of
active edges in a special case of a graph. Then we will construct a more complicated
graph where the set of active edges will form a fractal of length t logt.
Lemma 1. If all edges on the y-axis have time parameter equal to 1 and all other
edges have their time parameter equal to 2, then at time T the set of active vertices
20
Ivan Matic
is given by
AT = {(0, T )} ∪ {(0, T − 1)} ∪
+1
b T[
4 c
{(−k, T − 2k) , (k, T − 2k)}
k=1
∪
[
T +1
z∈Z\{−b T +1
4 c,...,b 4 c}
T
z,
.
2
Proof. After T −2k units of time the water can travel over the path γk that consists of
vertices (0, 0), (0, 1), . . ., (0, T − 2k). In additional 2k units of time the water travels
over the path γk0 that consists of vertices (0, T − 2k), (1, T − 2k), . . ., (k, T − 2k).
Consider any other path that goes from x axis to the point (k, T − 2k) for some fixed
Fig. 5 The active edges at time T .
k ≤ T +1
. If the path takes some steps over edges that belong to y axis then it
4
would have to go over at least k horizontal edges to reach y axis, which would take
2k units of time. The path would have to take at least T − 2k vertical edges, which
would take at least T − 2k units of time. Thus the travel would be longer than or
equal to T .
However, if the path does not take steps over the edges along y axis then it would
have to take at least T − 2k steps over edges that have passage time equal to 2. This
would take 2(T − 2k) = 2T − 4k units of time. If T + 1 is not divisible by 4, then
k < T +1
4 and
2T − 4k > 2T − T − 1 = T − 1,
which
mean that the travel time is at least T . If T + 1 is divisible by 4 and
would
k = T +1
then the vertical path would reach (k, T − 2k) at time T − 1. However,
4
A parallel algorithm for the constrained shortest path problem on lattice graphs
21
the vertex (k, T − 2k) would still be active because the water would not reach (k +
1, T − 2k) which is a neighbor of (k, T − 2k).
Let us denote by Nt the number of active vertices at time t whose x coordinate is
between −t and t,
Nt = (x, y) ∈ {−t, −t + 1, . . . ,t − 1,t} × Z+
0 : (x, y) is active at time t .
Theorem 2. There exist real numbers α and t ≥ 0 and an environment ω for which
Nt (ω) ≥ αt logt.
Proof. Assume that t = 2k for some k ∈ N. Let us define the following points with
their coordinates T = (0,t), L = − 2t , 0 , and O = 0, 2t . We will recursively construct the sequence of pairs (ω1 , I1 ), (ω2 , I2 ), . . ., (ωk , Ik ) where ω j is an assignment of passage times to the edges and I j is a subgraph of Z2 . This subgraph will
be modified recursively. All edges in I j have passage times equal to 2 in the assignment ω j . Having defined the pair (ω j , I j ) we will improve passage times over
some edges in the set I j by changing them from 2 to 1. This way we will obtain a
new environment ω j+1 and we will define a new set I j+1 to be a subset of I j . The
new environment ω j+1 will satisfy
Nt (ω j+1 ) ≥ Nt (ω j ) + βt,
for some β > 0.
Let us first construct the pair (ω1 , I1 ). We will only construct the configuration
to the left of the y axis and then reflect it across the y axis to obtain the remaining
configuration.
All edges on the y axis have the passage times equal to 1, and all edges on the
segment LO have the passage times equal to 1. All other edges have the passage
times equal to 2. Define I1 = 4LOT . Then the polygonal line LY T contains the
active vertices whose x coordinate is between −t and 0.
The environment ω2 is constructed in the following way. Let us denote by L0
and T0 the midpoints of LO and T O. Let X be the midpoint of LT . We change
all vertices on L0 X and T0 X to have the passage time equal to 1. We define I2 =
4LL0 X ∪ 4XT0 T .
Let L1 and L2 be the midpoints of LL0 and L0 O and let L0 and L00 be the intersections of XL1 and XL2 with LY . The points T 0 and T 00 are defined in an analogous
way: first T1 and T2 are defined to be the midpoints of T T0 and OT0 and T 0 and T 00
are the intersections of XT1 and XT2 with TY .
The polygonal line LL0 XL00Y T 00 XT 0 T is the set of active edges that are inside the
triangle LOT . The following lemma will allow us to calculate Nt (ω2 ) − Nt (ω1 ).
Lemma 2. Let Λ and λ denote the lengths of the polygonal lines LL0 XL00Y T 00 XT 0 T
and LY T respectively. If t is the length of OT then
4
Λ = λ + √ t.
3 5
22
Ivan Matic
Fig. 6 The set of active edges in configuration ω2 .
Proof. It suffices to prove that LL0 + L0 X + XL00 + L00Y = LY + 3√2 5 t. From the similarities 4LL0 X ∼ 4LOT and 4LL0 L0 ∼ LOX we have that L0 L0 kOX. Therefore
L0 is the midpoint of LY and LY = LL0 + L0Y = LL0 + L0 X. It remains to prove that
XL00 + L00Y = 3√2 5 t. From
∠L0 XL00 = ∠L0 XL0 = ∠L0 LL00
we conclude that the quadrilateral LL0 L00 X is inscribed in a circle. The segment LX
is a diameter of the circle hence ∠LL00 X = ∠LL0 X = 90◦ . We also have ∠L00 XY =
∠L0 XY − ∠L0 XL00 = 45◦ − ∠OLT0 = 45◦ − arctan 12 .
The point Y is the centroid of 4LOT hence XY = 13 XO = 3√1 2 t. Therefore
1
1
◦
◦
XL + L Y = XY cos 45 − arctan
+ XY sin 45 − arctan
2
2
1
1
◦
◦
cos 45 − arctan 2 + sin 45 − arctan 2
√
t
=
3 2
cos 45◦ − arctan 12 cos 45◦ + sin 45◦ − arctan 21 sin 45◦
=
t
3
cos 45◦ − arctan 12 − 45◦
cos arctan 12
=
t=
t
3
3
2
= √ t.
3 5
00
00
The number of edges on each of the segments of the polygonal lines we obtained is
equal to √u5 , where u is the length of the segment. Using this fact with the previous
lemma applied to both 4LOT and its reflection along OT gives us
A parallel algorithm for the constrained shortest path problem on lattice graphs
23
1
4
4
Nt (ω2 ) − Nt (ω1 ) = √ t · √ = t.
3 5
5 15
We now continue in the same way and in each of the triangles LL0 X and XT0 T we
Fig. 7 The set of active edges in configuration ω3 .
perform the same operation to obtain ω3 and I3 . Since the side length of LL0 X is
t
2 t
2 , the increase in the number of elements in the new set of active vertices is 15 · 2 .
However, this number has to be now multiplied by 4 because there are 4 triangles
to which the lemma is applied: 4LL0 X, 4XT0 T , and the reflections of these two
triangles with respect to OT . Therefore the increase in the number of active vertices
4
2 t
· 2 = 15
t.
is Nt (ω3 ) − Nt (ω2 ) = 4 · 15
This operation can be repeated k times and we finally get that
Nt (ωk ) = Nt (ω1 ) + (k − 1) ·
Thus the theorem holds if we set α =
4
4
t ≥ k · t.
15
15
4
15 log 2 .
Proof (Proof of Theorem 1). Recall that p is the probability that the time 1 is assigned to each edge. Let ρ = min {p, 1 − p}. The configuration provided in the proof
2
of Theorem 2 has its probability greater than or equal to ρ t . Therefore
2
P (At ≥ αt logt) ≥ ρ t = et
Therefore we may take µ = − ln ρ.
2 ln ρ
.
24
Ivan Matic
8 Performance analysis
The algorithm was implemented in C++ and OpenCL. The hardware used has a quad
core Intel i5 processor with clock speed of 3.5GHz and AMD Radeon R9 M290X
graphic card with 2 gigabytes of memory. The graphic card has 2816 processing
elements.
The table provides a comparison of the performance of the algorithm on 4 samples of three dimensional cubes with edges of lengths 50, 75, 100, and 125. The
initial configuration for each of the graphs assumes that there is water on the boundary of the cube, while the set B is defined to be the center of the cube. The same
program was executed on graphic card and on CPU.
Graph
GPU time (s) CPU time (s)
50 × 50 × 50
3
10
75 × 75 × 75
8
61
100 × 100 × 100 21
275
125 × 125 × 125 117
1540
The graph that corresponds to the cube 100 × 100 × 100 has 1000000 vertices
and 2970000 edges, while the graph corresponding to the cube 125 × 125 × 125 has
1953125 vertices and 5812500 edges.
9 Conclusion
The algorithm described in this chapter solves the constrained shortest path problem
using parallel computing. It is suitable to implement on graphic cards and CPUs that
have large number of processing elements. The algorithm is implemented in C++
and OpenCL and the parallelization improves the speed tenfold.
The main idea is to follow the percolation of water through the graph and assign
different qualities to drops that travel over different edges. Each step of the algorithm corresponds to a unit of time. It suffices to analyze only those vertices and
edges through which the water flows. We call them active vertices and active edges.
Therefore, the performance of the algorithm is tied to the sizes of these active sets.
Theorem 1 proves that it is possible to have at time t an active set of size
O(t logt). The proof of the theorem relied on constructing one such set. It is an
open problem to find the average size of the active set at time t.
Problem 1. If the weights and travel times of the edges are chosen independently at
random, what is the average size of the active set at time t?
At some stages of the execution, the program needs additional memory to store
phantom edges in the graph. It would be interesting to know how many phantom
edges are allocated during a typical execution. This can be formally phrased as an
open problem.
A parallel algorithm for the constrained shortest path problem on lattice graphs
25
Problem 2. If the weights and travel times of the edges are chosen independently at
random, what is the average number of phantoms that need to be created during the
execution of the algorithm?
Acknowledgements
The author was supported by PSC-CUNY grants #68387 − 0046, #69723 − 0047
and Eugene M. Lang Foundation.
References
1. G. Amir, I. Corwin, and J. Quastel. Probability distribution of the free energy of the continuum directed random polymer in 1+1 dimensions. Comm. Pure Appl. Math., 64:466–537,
2011.
2. S. Armstrong, H. Tran, and Y. Yu. Stochastic homogenization of a nonconvex Hamilton–
Jacobi equation. Calc. Var. Partial Differential Equations, 54:1507–1524, 2015. (Submitted)
arXiv:1311.2029.
3. M. Benaim and R. Rossignol. Exponential concentration for first passage percolation through
modified poincaré inequalities. Ann. Inst. Henri Poincaré Probab. Stat., 44(3):544–573, 2008.
4. I. Benjamini, G. Kalai, and O. Schramm. First passage percolation has sublinear distance
variance. The Annals of Probability, 31(4):1970–1978, 2003.
5. N. Boland, J. Dethridge, and I. Dumitrescu. Accelerated label setting algorithms for the
elementary resource constrained shortest path problem. Operations research letters, 34:58–
68, 2006.
6. S. Chatterjee and P. S. Dey. Central limit theorem for first-passage percolation time across
thin cylinders. Probability Theory and Related Fields, 156(3):613–663, 2013.
7. J. T. Cox and R. Durrett. Some limit theorems for percolation processes with necessary and
sufficient conditions. The Annals of Probability, 9(4):583–603, 1981.
8. M. Damron and M. Hochman. Examples of nonpolygonal limit shapes in i.i.d. first-passage
percolation and infinite coexistence in spatial growth models. The Annals of Applied Probability, 23(3):1074–1085, 2013.
9. M. Desrochers, J. Desrosiers, and M. Solomon. A new optimization algorithm for the vehicle
routing problem with time windows. Operations research, 40(2):342–354, 1992.
10. J. Hammersley and D. Welsh. First-passage percolation, subadditive processes, stochastic
networks, and generalized renewal theory. Bernoulli-Bayes-Laplace Anniversary Volume,
1965.
11. S. Irnich and G. Desaulniers. Shortest path problems with resource constraints. In G. Desaulniers, J. Desrosiers, and M. M. Solomon, editors, Column Generation, GERAD 25th Anniversary Series, pages 33–65. Springer, 2005.
12. E. Köhler, R. H. Möhring, and H. Schilling. Acceleration of shortest path and constrained
shortest path computation. Lecture notes in computer science, 3503:126–138, 2005.
13. E. Kosygina, F. Rezakhanlou, and S. R. S. Varadhan. Stochastic homogenization of
Hamilton–Jacobi–Bellman equations. Comm. Pure Appl. Math., 59(10):1489–1521, 2006.
14. E. Kosygina, F. Yilmaz, and O. Zeitouni. Nonconvex homogenization of a class of onedimensional stochastic viscous Hamilton-Jacobi equations. in preparation, 2017.
15. J. Krug and H. Spohn. Kinetic roughening of growing surfaces. Solids far from equilibrium,
pages 412–525, 1991.
26
Ivan Matic
16. X.-Y. Li, P.-J. Wan, Y. Wang, and O. Frieder. Constrained shortest paths in wireless networks.
IEEE MilCom, pages 884–893, 2001.
17. L. Lozano and A. L. Medaglia. On an exact method for the constrained shortest path problem.
Computers and Operations Research, 40(1):378–384, 2013.
18. I. Matic. Parallel algorithm for constrained shortest path problem in C++/OpenCL.
github.com/maticivan/parallel_constrained_shortest_path
19. I. Matic and J. Nolen. A sublinear variance bound for solutions of a random Hamilton–Jacobi
equation. J. Stat. Phys., 149:342–361, 2012.
20. K. Mehlhorn and M. Ziegelmann. Resource constrained shortest paths. In Lecture Notes in
Computer Science, volume 1879, pages 326–337, 2000.
21. S. Misra, N. E. Majd, and H. Huang. Approximation algorithms for constrained relay node
placement in energy harvesting wireless sensor networks. IEEE Transactions on Computers,
63(12):2933–2947, 2014.
22. R. Muhandiramge and N. Boland. Simultaneous solution of lagrangean dual problems interleaved with preprocessing for the weight constrained shortest path problem. Networks,
53:358–381, 2009.
23. C. M. Newman and M. S. T. Piza. Divergence of shape fluctuations in two dimensions. The
Annals of Probability, 23(3):977–1005, 1995.
24. F. Rezakhanlou. Central limit theorem for stochastic Hamilton–Jacobi equations. Commun.
Math. Phys., 211:413–438, 2000.
25. T. Sasamoto and H. Spohn. One-dimensional kardar-parisi-zhang equation: An exact solution
and its universality. Phys. Rev. Lett., 104, 2010.
26. P. E. Souganidis. Stochastic homogenization of Hamilton–Jacobi equations and some applications. Asymptot. Anal., 20(1):1–11, 1999.
| 8cs.DS
|
IJICIS, Vol.9, No. 2 JULY 2009
TO PARALLELIZE OR NOT TO PARALLELIZE, CONTROL AND DATA
FLOW ISSUE
A. I. El-Nashar
Computer Science Department, Faculty of Science, Minia University,
Minia - Egypt
[email protected]
Abstract: New trends towards multiple core processors imply using standard programming models to
develop efficient, reliable and portable programs for distributed memory multiprocessors and
workstation PC clusters. Message passing using MPI is widely used to write efficient , reliable and
portable applications. Control and data flow analysis concepts, techniques and tools are needed to
understand and analyze MPI programs. If our point of interest is the program control and data flow
analysis, to decide to parallelize or not to parallelize our applications, there is a question to be
answered, " Can the existing concepts, techniques and tools used to analyze sequential programs also
be used to analyze parallel ones written in MPI?". In this paper we'll try to answer this question.
Keywords: Parallel programming, Message Passing Interface, Data flow, Control flow
1. Introduction
A concurrent program contains two or more threads that execute concurrently and work together to
perform some task [17]. Using multiple threads can increase the efficiency. The increase in the use
of parallel computing is being accelerated by the development of various parallel hardware
architectures and also by constructing new programming models to achieve the best use of these
architectures. Message Passing Interface or MPI is one of these models which is widely used for
programming distributed memory systems. While concurrent programs offer some advantages, they
also exhibit nondeterministic behavior, making them difficult to test. One significant challenge in
bringing the power of parallel machines to application programmers is providing them with a suite
of software tools similar to the tools that sequential programmers currently utilize as a practical way
of obtaining increased confidence in software and also to guarantee that the program is correct. In
particular, automatic or semi automatic testing tools for parallel programs are lacking compared
with that tools for sequential programs [1] .
Control flow and data flow analysis of serial programs differ from those of MPI parallel programs
as a result of complex interaction between concurrent processes and also due to the SPMD nature.
The overhead of analyzing this type of programs may make application programmers to think a lot
to decide to parallelize or not to parallelize and as a result they may prefer using the ordinary serial
programming style to be not faced with the extra analysis effort. This paper presents an
implemented technique to analyze MPI programs. The analysis report serves as a guide to the
programmer to compromise between the advantage of parallelism and the cost of analysis effort.
161
El-Nashar : To Parallelize Or Not To Parallelize, Control And Data Flow Issue
The paper is organized as follows: Section 2 describes the MPI Programming Model. In section 3,
MPI Program Analysis is presented. Section 4, concerns with MPI Program Representation. MPI
Static Data Flow Analysis technique is described in section 5.
2. MPI Programming Model
MPI programs are coded in a special manner, in which each process executes the same program
with unique data. All parallelism is explicit; the programmer is responsible for correctly identifying
parallelism and implementing parallel algorithms using MPI constructs.
1
1
1
1
1
1
1
2
3
1
2
3
4
5
6
7
8
9
4
10
5
5
5
11
12
12 5
$
$
$
6
7
8
8
9
10
10
11
12
13
14
15
16
17
17
18
19
20
21
22
23
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
include 'mpif.h'
integer ierr, myid, status(MPI_STATUS_SIZE)
integer count,x,sum,received,sender,process
SUM=3
call MPI_Init(ierr)
call MPI_Comm_size(MPI_COMM_WORLD, count, ierr)
call MPI_Comm_rank(MPI_COMM_WORLD, myid, ierr)
IF( MYID.EQ.0) THEN
call MPI_Recv(received,1,MPI_INTEGER, MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, status, ierr)
call MPI_Recv(sender, 1, MPI_INTEGER,MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, status, ierr)
SUM=SUM+ RECEIV
WRITE(*,5) RECEIVD , SENDER, SUM
format (' The value: ', I3 , 2x , 'is received from process:',
I3 , 'The sum:' , I3)
ENDIF
IF( MYID.EQ.1 ) THEN
X=5
IF(X.LT.0) THEN
X=X+1
ELSE
X=X-1
END IF
call MPI_Send(X,1,MPI_INTEGER,0,0,MPI_COMM_WORLD,ierr)
PROCESS=MYID
call MPI_Send(process,1,MPI_INTEGER,0,0,MPI_COMM_WORLD,ierr)
ENDIF
IF( MYID.EQ.2 ) THEN
X=7
X=X * 2
call MPI_Send(X,1,MPI_INTEGER,0,0,MPI_COMM_WORLD,ierr)
PROCESS=MYID
call MPI_Send(process,1,MPI_INTEGER,0,0,MPI_COMM_WORLD,ierr)
ENDIF
call MPI_Finalize(ierr)
END
Figure 1. A typical MPI program
MPI is available as an open sources implementations on a wide range of parallel platform
[6,9,15,16]. MPICH2 [4,10] is a recent MPI open sources implementation, in which the MPI source
program is compiled and then linked with the MPI libraries to obtain the executable. The user issues
a directive to the operating system that places a copy of the executable program on each processor,
162
IJICIS, Vol.9, No. 2 JULY 2009
the number of processes is provided within the user directive. Each processor begins execution of its
copy of executable. Each process can execute different statements by branching within the program
based on a unique rank "process identifier". This form of MIMD programming is frequently called
Single-program multiple-data SPMD. Each process has its own local memory address space; there
are no shared global variables among processes. All communications are performed through special
calls to MPI message passing routines.
MPI uses objects [11] called communicators and groups to define which collection of processes may
communicate with each other. A communicator must be specified as an argument for most MPI
routines. The predefined communicator MPI_COMM_WORLD is used whenever a communicator
is required, it includes all of MPI processes. Within a communicator, every process has its own
unique, integer identifier "rank" assigned by the system when the process initializes. Ranks are
contiguous and begin at zero, used by the programmer to specify the source and destination of
messages, and also used conditionally by the application to control program execution.
An MPI program consists of four parts, a typical MPI program is shown in figure 1. The first one is
the MPI include file which is required for all programs/routines which make MPI library calls
(line1). The second part is responsible for initializing MPI environment, MPI environment
management routines are used for an initializing and terminating the MPI environment, querying the
environment and identity. MPI_Init [8], (line 5), initializes the MPI execution environment. This
function must be called in every MPI program, must be called before any other MPI functions and
must be called only once in an MPI program.
MPI_Comm_size, (line 6), determines the number of processes in the group associated with a
communicator. Generally used within the communicator MPI_COMM_WORLD to determine the
number of processes being used by the application.
MPI_Comm_rank (line 7), determines the rank of the calling process within the communicator.
Initially, each process will be assigned a unique integer rank between 0 and number of processes.
The third part is the body of program instructions, calculations, and message passing calls. The last
one is terminating MPI environment (line 32). MPI provides several routines used to manage the
inter-process communications via send / receive operations ( lines 9,10,21,23,28,30 ), wait for a
message's arrival or probe to find out if a message has arrived.
3. MPI Program Analysis
Running the executable of the source code listed in figure 1 several times using three process may
yields one of two outputs, one of them indicates that the value 4 is sent from process 1 to process 0
and the sum value is 7, the other one indicates that the value 14 is sent from process 2 to process 0
and the sum value is 17.
The order of these outputs is unpredictable. This situation reflects the non-deterministic behavior of
program execution.
163
El-Nashar : To Parallelize Or Not To Parallelize, Control And Data Flow Issue
Applying ordinary data flow analysis that does not consider the SPMD nature of MPI programs on the
source program listed in figure 1 to identify the statements influenced by the definitions of "sum" in
lines 4 and 11 and also definition of "x" in lines 15,17,19,26 and 27 is shown in table 1 .
Table 1
Case
definition of sum in line 4
definition of sum in line 11
definition of x in line 15
definition of x in line 17
definition of x in line 19
definition of x in line 26
definition of x in line 27
Affected statements
sum=sum + received
write(*,5) received , sender, sum
IF(x.LT.0) THEN
x = x +1
x=x-1
call MPI_Send (x,…)
call MPI_Send (x,…)
x=x*2
call MPI_Send (x,…)
11
12
16
17
19
21
21
27
28
The execution behavior demonstrates that the affected statements shown in table1 are not the only
affected ones but there are some other statements that should be encountered, it can be noticed that
this analysis fails to detect the effect of the definition of "x" on the computation of "sum" in line
11, as shown in table 2.
Table 2
Case
definition of x in lines 17, 19
definition of x in line 27
Statements should be encountered
9. call MPI_Recv(received,……
9. call MPI_Recv(received,……
Variable definitions like " sum = 3 " in line 4, are shared in SPMD programs without
communication channels so they are considered as global variables. On the other hand, variables
defined within each process section , enclosed between " IF( myid.eq.process_id) " and " ENDIF ",
like " sum=sum + received" in line 11 can't be shared outside this section otherwise appropriate
communication channels are used. These variables are considered as local variables.
4. MPI Program Representation
As shown above, ordinary data flow analysis techniques fail to demonstrate a correct analysis for
MPI, the SPMD nature needs to be modeled correctly in the program representation to be
considered during static program analysis, this can be achieved by using a special data structure that
can represent sequential flow, parallel flow and synchronization in explicitly MPI parallel programs.
4.1 MPI-CFG Construction Challenges
Data flow analysis techniques represent a program by its control flow graph, CFG, which consists
of a set of nodes and edges. Each node represents a basic block which is a set of consecutive
statements of the program, and each edge represents the control flow between these blocks. The goal
of analysis techniques is to identify which definitions of program variables can affect which uses.
To build such CFG's, the analyzed program is divide into a set of basic blocks and the set of edges
164
IJICIS, Vol.9, No. 2 JULY 2009
connecting these blocks according to the flow of control are generated. The constructed CFG is then
used by the static analyzer to identify the def-use associations among the blocks.
Extra effort has to be done in building CFG representing MPI programs (MPI-CFG) according to
the following challenges:
1. Processes in MPI programs are declared by using the ordinary conditional IF statement
depending on a unique identifier. This will make a confusion during dividing the
program into basic blocks, and also during the process of generating edges which will
badly affect the operations of static analyzer. So, the IF statements that used to declare
processes must be treated in a special manner rather than that is used in treating the IF
statements used within the body of each process as shown in figure 1, lines 14 and 16 .
2. MPI program is executed by more than one process, each process has its local memory
address space; there is no shared global variables among these processes except that are
defined before the initialization of the MPI execution environment. This requires to
identify both local and global variables.
3. def-use association among process can be achieved only by calling inter-process
communication message passing routines (send, receive,…). This implies constructing
extra edges that represent these constructs.
4.2 Implementation of MPI-CFG Construction
Now we present our algorithm to build the MPI-CFG. This flow graph resembles the synchronized
flow graph [3], program execution graph [14] , parallel flow graph [7], and parallel program flow
graph PPFG [2]. The algorithm works as follows:
1. MPI program statements identification.
In this phase, each program statement is assigned a unique number "type" to be identified from
the other statements of the program. The algorithm must check for the following:
If the statement " Call MPI_Comm_rank( XX,YY,ZZ) " is encountered, it is assigned its type
and the second parameter YY which indicates the variable name that will be used to identify
the parallel processes is recorded as "special_id".
The assigned type of conditional IF statements depends on the recorded "special_id"; if the
value of "special_id" appears in the condition, this means that the encountered IF statement is
used to declare a process, otherwise, it is an ordinary conditional statement.
The output of this phase is a numbered statements of the input program associated with their
types and the recorded "special_id".
2. Building Basic Blocks
This phase uses the output of the previous phase to build the program basic blocks as in the case
of ordinary CFG. We construct two extra special types of basic blocks called "message block"
and "finalize block". A message block is either "receive block" or "send block". A basic block
that has at most one communication statement at start of the block is said to be "receive block".
This block is constructed if the statement call MPI_Recv( ) is encountered. The "send block"
has at most one communication statement at its end. This block is constructed if the statement
call MPI_Send( ) is encountered. The "finalize block" is constructed if call MPI_Finalize( )
statement is encountered.
165
El-Nashar : To Parallelize Or Not To Parallelize, Control And Data Flow Issue
During building basic blocks the program variables, their block numbers and their status (def, cuse, or p-use) are also recorded. Extra effort has to be done in case of call MPI_Recv( ) and call
MPI_Send( ) to record and classify the parameters of these two calls. At the termination of this
phase another version of the input MPI program is generated. This version contains the
statement and block number for each program statement as shown in figure 1. All the required
information about the variable names and the parameters of send/receive constructs are also
recorded.
3. Generating Edges.
This phase connects the basic blocks generated in the previous phase with the appropriate edges.
The edges are classified into three categories, sequential, parallel, and synchronization edges.
Sequential edges indicate a possible flow from a block to another one. This type of edges is
used to connect the basic blocks within each process as the ordinary sequential flow edges.
Parallel edges indicate the parallel control flow as at process declaration and termination points.
Synchronization edges represent the inter-process communication via send/receive operations.
Synchronization edges are generated by matching the parameters of call MPI_Recv( ) and call
MPI_Send( ) recorded in the second phase. The output of this phase is the MPI-CFG shown in
figure 2.
Figure 2. MPI Control Flow Graph for the MPI program
166
IJICIS, Vol.9, No. 2 JULY 2009
5. MPI Static Data Flow Analysis
Static data flow analysis is a technique for gathering information about the possible set of values
calculated at various points in a sequential program. CFG is used to determine those parts of a
program to which a particular value assigned to a variable might propagate. This can be done by
generating two sets, dcu (i ) and dpu (i, j ) [13] for program variables.
MPI PROGRAM ANALYSIS REPORT
(A) DEF-USE ASSOCIATION WITHIN PROCESSES
1- THE VARIABLE "SUM" DEFINED AT BLOCK NO.:1 HAS THE FOLLOWING USES:
1.1 C-USE AT BLOCK NO. 5
2- THE VARIABLE "RECEIV" DEFINED AT BLOCK NO.:3 HAS THE FOLLOWING USES:
2.1 C-USE AT BLOCK NO. 5
3- THE VARIABLE "SENDER" DEFINED AT BLOCK NO.:4 HAS THE FOLLOWING USES:
3.1 C-USE AT BLOCK NO. 5
4- THE VARIABLE "SUM" DEFINED AT BLOCK NO.:5 HAS THE FOLLOWING USES:
NO GLOBAL USE
5- THE VARIABLE "X" DEFINED AT BLOCK NO.:8 HAS THE FOLLOWING USES:
5.1 C-USE AT BLOCK NO. 9
5.2 C-USE AT BLOCK NO. 10
5.3 P-USE AT EDGE (8- 9)
5.4 P-USE AT EDGE (8- 10)
6- THE VARIABLE "X" DEFINED AT BLOCK NO.:9 HAS THE FOLLOWING USES:
6.1 C-USE AT BLOCK NO. 12
7- THE VARIABLE "X" DEFINED AT BLOCK NO.:10 HAS THE FOLLOWING USES:
7.1 C-USE AT BLOCK NO. 12
8- THE VARIABLE "PROCESS" DEFINED AT BLOCK NO.:13 HAS THE FOLLOWING USES:
8.1 C-USE AT BLOCK NO. 14
9- THE VARIABLE "X" DEFINED AT BLOCK NO.:17 HAS THE FOLLOWING USES:
9.1 C-USE AT BLOCK NO. 18
10- THE VARIABLE "PROCESS" DEFINED ATBLOCK NO.:19 HAS THE FOLLOWING USES:
10.1 C-USE AT BLOCK NO. 20
(B) INTER-PROCESS COMMUNICATION VARIABLES
1- THE VARIABLE "RECEIVD" OF RECEIVE BLOCK NO.:3
WILL BE ASSIGNED ONE OF THE FOLLOWING VARIABLE(S):
1.1 "X" FROM SEND BLOCK NO.:12 VIA SYNCHRONIZATION EDGE(12- 3)
1.2 "X" FROM SEND BLOCK NO.:18 VIA SYNCHRONIZATION EDGE(18- 3)
2- THE VARIABLE "SENDER" OF RECEIVE BLOCK NO.:4
WILL BE ASSIGNED ONE OF THE FOLLOWING VARIABLE(S):
2.1 "PROCESS" FROM SEND BLOCK NO.:14 VIA SYNCHRONIZATION EDGE(14- 4)
2.2 "PROCESS" FROM SEND BLOCK NO.:20 VIA SYNCHRONIZATION EDGE(20- 4)
Figure 3. Data Flow Analysis Report
167
El-Nashar : To Parallelize Or Not To Parallelize, Control And Data Flow Issue
These two sets are necessary to determine the definitions of every variable in the program and the
uses that might be affected by these definitions. The set dcu (i ) is the set of all variable definitions
for which there are def-clear paths to their c-uses at node i
.
The dpu (i, j ) is the set of all variable definitions for which there are def-clear paths to their p-uses
at edge (i, j ) [12]. Using information concerning the location of variable definitions and references,
together with the “basic static reach algorithm” [5], the two sets can be determined. The basic static
reach algorithm is used to determine the sets reach(i) and avail(i). The set reach(i) is the set of all
variable definitions that reach node i . The set avail(i) is the set of all available variables at node i.
This set is the union of the set of global definitions at node i together with the set of all definitions
that reach this node and are preserved through it. Using these two sets, the sets dcu (i ) and
dpu (i, j ) are constructed from the formula :
dcu (i ) reach(i ) c use(i )
dpu (i, j ) avail (i ) p use(i, j )
We applied this technique on the constructed MPI-CFG with some modifications to handle the
nature of MPI programs. The output of the implemented technique is shown in figure 3.
6. Conclusion and Future Work
Unlike sequential programs, data flow analysis of MPI programs requires extra effort. Applying existing
concepts, techniques and tools used to analyze sequential programs on MPI programs fails to report a
correct program analysis. These techniques require some modifications to handle the SPMD nature of
MPI programs. We have implemented a technique to extend the program CFG to represent the MPI
programs MPI-CFG. The static analyzer uses the constructed graph to generate the program analysis
report. We have implemented the techniques of building message basic blocks, constructing parallel
edges, and also constructing synchronization edges represent send/receive constructs to produce a
correct MPI program representation. In future, we hope to implement the construction of
synchronization edges for all MPI inter-process communication constructs.
168
IJICIS, Vol.9, No. 2 JULY 2009
References
1.
C.Yang , A. L. Souter and L. L. Pollock “ All-du-path coverage for Parallel Programs” , the
Sixth Asian Test Symposium, 1998.
2. C. Yang and L. L. Pollock “ The Challenges in Automated Testing of Multithreaded Programs”,
the 14th International Conference on Testing Computer Software, 1997.
3. D. Callahan, K. Kennedy, and J. Subholk " Analysis of event synchronization in a parallel
programming tool" 2nd ACM SIGPLAN Symposium on principles and practice, Mach 1990.
4. Download the Win32IA32 version of MPICH2 from: http://www-unix.mcs.anl.gov/mpi/mpich2/
5. F. E. Allen and J. Cocke “ A Program Data Flow Analysis Procedure,” Communications of the
ACM, vol. 9, p.137-147, 1976.
6. G. Burns, R. Daoud, and J. Vaigl " LAM: An open cluster environment for MPI", Proceedings of
the Supercomputing Symposium, 1994.
7. H. Srinivasan and Dirk Grunwald " An Efficient Construction of Parallel Static Assignment Form
for Structured Parallel Programs", Technical report CU-CS-564-91, University of Colorado at
Boulder., December 1991.
8. J.M. Squyres " MPI Mechanic", vol. 2 , Feb. 2004.
9. J.M. Squyres and A. Lumsadaine " A component Architecture for LAM/MPI", Proceedings of
the 10th European PVM/MPI User's Group Meeting, 2003.
10. Main MPICH homepage: http://www-unix.mcs.anl.gov/mpi/
11. MPI tutorial: http://www.llnl.gov/computing/tutorials/mpi/
12. M. R. Girgis and M. R. Woodward “ An Integrated System for Program Testing Using Weak
Mutation and Data Flow Analysis”, Proceedings of Eights International Conference on Software
Engineering , IEEE
Computer Society, p. 313-319, 1985.
13. M. R. Girgis “ Using Symbolic Execution and Data Flow Criteria to Aid Test Data Selection” ,
software testing, verification and reliability, v. 3, p.101-113, 1993.
14. V. Balasundaram and K. Kennedy " Compile-time detection of race conditions in a parallel
program" 3rd International conference on supercomputing, June 1989.
15. W . Gropp, E. Lusk, N. Doss, and A. Skjellum " A high-performance, portable implementation of
the MPI message passing interface standard", parallel computing, v. 22, p.789-828, Sept. 1996.
16. W. D. Gropp and E. Lusk " User's guide for mpich, a portable implementation of MPI", technical
report ANL-96/6, Math. And Computer science Division, Argonne National Laboratory, 1996.
17. Y. Lei and R. H. Carver " Reachability testing of concurrent programs", IEEE transactions on
software engineering, v. 32, no. 6, June 2006.
169
| 6cs.PL
|
Logical Methods in Computer Science
Vol. 13(4:7)2017, pp. 1–46
https://lmcs.episciences.org/
Submitted
Published
Nov. 09, 2016
Nov. 13, 2017
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
CARSTEN LUTZ AND FRANK WOLTER
University of Bremen, Germany
e-mail address: [email protected]
University of Liverpool
e-mail address: [email protected]
Abstract. We analyze the data complexity of ontology-mediated querying where the
ontologies are formulated in a description logic (DL) of the ALC family and queries
are conjunctive queries, positive existential queries, or acyclic conjunctive queries. Our
approach is non-uniform in the sense that we aim to understand the complexity of each
single ontology instead of for all ontologies formulated in a certain language. While doing
so, we quantify over the queries and are interested, for example, in the question whether all
queries can be evaluated in polynomial time w.r.t. a given ontology. Our results include a
PTime/coNP-dichotomy for ontologies of depth one in the description logic ALCFI, the
same dichotomy for ALC- and ALCI-ontologies of unrestricted depth, and the non-existence
of such a dichotomy for ALCF-ontologies. For the latter DL, we additionally show that it
is undecidable whether a given ontology admits PTime query evaluation. We also consider
the connection between PTime query evaluation and rewritability into (monadic) Datalog.
1. Introduction
In recent years, the use of ontologies to access instance data has become increasingly popular
[PLC+ 08, KZ14, BO15]. The general idea is that an ontology provides domain knowledge
and an enriched vocabulary for querying, thus serving as an interface between the query
and the data, and enabling the derivation of additional facts. In this emerging area, called
ontology-mediated querying, it is a central research goal to identify ontology languages
for which query evaluation scales to large amounts of instance data. Since the size of the
data typically dominates the size of the ontology and the size of the query by orders of
magnitude, the central measure for such scalability is data complexity—the complexity of
query evaluation where only the data is considered to be an input, but both the query and
the ontology are fixed.
In description logic (DL), ontologies take the form of a TBox, data is stored in an ABox,
and the most important classes of queries are conjunctive queries (CQs) and variations
thereof, such as positive existential queries (PEQs). A fundamental observation regarding
this setup is that, for expressive DLs such as ALC and SHIQ, the complexity of query
evaluation is coNP-complete and thus intractable [Sch93, HMS07, GLHS08].1 The classical
Key words and phrases: Description Logic, Ontology-Mediated Querying, Data Complexity.
1When speaking of complexity, we always mean data complexity
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.23638/LMCS-13(4:7)2017
CC
Carsten Lutz and Frank Wolter
Creative Commons
2
CARSTEN LUTZ AND FRANK WOLTER
approach to avoiding this problem is to replace ALC and SHIQ with less expressive DLs
that are ‘Horn’ in the sense that they can be embedded into the Horn fragment of first-order
(FO) logic. Horn DLs typicall admit query evaluation in PTime, examples include a variety
of logics from the EL [BBL05] and DL-Lite families [CDGL+ 07] as well as Horn-SHIQ, a
large fragment of SHIQ with PTime query evaluation [HMS07].
It may thus seem that the data complexity of query evaluation in the presence of DL
ontologies is understood rather well. However, all results discussed above are at the level
of logics, i.e., traditional results about data complexity concern a class of TBoxes that is
defined in a syntactic way in terms of expressibility in a certain DL language, but no attempt
is made to identify more structure inside these classes. Such a more fine-grained study,
however, seems very natural both from a theoretical and from a practical perspective; in
particular, it is well-known that ontologies which emerge in practice tend to use ‘expensive’
language constructs that can result in coNP-hardness of data complexity, but they typically
do so in an extremely restricted and intuitively ‘harmless’ way. This distinction between
hard and harmless cases cannot be analyzed on the level of logics. The aim of this paper is
to initiate a more fine-grained study of data complexity that is non-uniform in the sense
that it does not treat all TBoxes formulated in the same DL in a uniform way.
When taking a non-uniform perspective, there is an important choice regarding the level
of granularity. First, one can analyze the complexity on the level of TBoxes, quantifying over
the actual query. Then, query evaluation for a TBox T is in PTime if every query (from
the class under consideration) can be evaluated in PTime w.r.t. T and it is coNP-hard if
there is at least one query that is coNP-hard to evaluate w.r.t. T . And second, one might
take an even more fine-grained approach where the query is not quantified away and the
aim is to classify the complexity on the level of ontology-mediated queries (OMQs), that is,
combinations of a TBox and an actual query. From a practical perspective, both setups
make sense; when the actual queries are fixed at the design time of the application, one
would probably prefer to work on the level of OMQs whereas the level of TBoxes seems
more appropriate when the queries can be freely formulated at application running time.
A non-uniform analysis on the level of OMQs has been carried out in [BtCLW14]. In this
paper, we concentrate on the level of TBoxes. The ultimate goal of our approach is as
follows:
For a fixed DL L and query language Q, classify all TBoxes T in L according to the
complexity of evaluating queries from Q w.r.t. T .
We consider the basic expressive DL ALC, its extensions ALCI with inverse roles and
ALCF with functional roles, and their union ALCF I. As query languages, we cover CQs,
acyclic CQs, and PEQs (which have the same expressive power as unions of conjunctive
queries, UCQs, which are thus implicitly also covered). In the current paper, we mainly
concentrate on understanding the boundary between PTime and coNP-hardness of query
evaluation w.r.t. DL TBoxes, mostly neglecting other relevant classes such as AC0 , LogSpace,
and NLogSpace.
Our main results are as follows (they apply to all query languages mentioned above).
1. There is a PTime/coNP-dichotomy for query evaluation w.r.t. ALCF I-TBoxes of depth
one, i.e., TBoxes in which no existential or universal restriction is in the scope of another
existential or universal restriction.
The proof rests on interesting model-theoretic characterizations of polynomial time CQevaluation which are discussed below. Note that this is a relevant case since most TBoxes
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
3
from practical applications have depth one. In particular, all TBoxes formulated in DL-Lite
and its extensions proposed in [CDGL+ 07, ACKZ09] have depth one, and the same is true for
more than 80 percent of the 429 TBoxes in the BioPortal ontology repository. In connection
with Point 1 above, we also show that PTime query evaluation coincides with rewritability
into monadic Datalog (with inequalities, to capture functional roles). As in the case of data
complexity, what we mean here is that all queries are rewritable into monadic Datalog w.r.t.
the TBox T under consideration.
2. There is a PTime/coNP-dichotomy for query evaluation w.r.t. ALCI-TBoxes.
This is proved by showing that there is a PTime/coNP-dichotomy for query evaluation w.r.t.
ALCI-TBoxes if and only if there is a PTime/NP-dichotomy for non-uniform constraint
satisfaction problems with finite templates (CSPs). The latter is known as the FederVardi conjecture that was recently proved in [Bul17, Zhu17], as the culmination of a major
research programme that combined complexity theory, graph theory, logic, and algebra
[BJK05, KS09, Bul11, Bar14]. Our equivalence proof establishes a close link between query
evaluation in ALC and ALCI and CSP that is relevant for DL research also beyond the stated
dichotomy problem. Note that, in contrast to the proof of the Feder-Vardi conjecture, the
dichotomy proof for TBoxes of depth one (stated as Point 1 above) is much more elementary.
Also, it covers functional roles and establishes equivalence between PTime query evaluation
and rewritability into monadic Datalog, which fails for ALCI-TBoxes of unrestricted depth
even when monadic Datalog is replaced with Datalog; this is a consequence of the link to
CSPs establishes in this paper.
3. There is no PTime/coNP-dichotomy for query evaluation w.r.t. ALCF-TBoxes (unless
PTime = NP).
This is proved by showing that, for every problem in coNP, there is an ALCF-TBox for
which query evaluation has the same complexity (up to polynomial time reductions); it then
remains to apply Ladner’s Theorem, which guarantees the existence of NP-intermediate
problems. Consequently, we cannot expect an exhaustive classification of the complexity
of query evaluation w.r.t. ALCF-TBoxes. Variations of the proof of Point 3 allow us to
establish also the following:
4. For ALCF-TBoxes, the following problems are undecidable: PTime-hardness of query
evaluation, coNP-hardness of query evaluation, and rewritability into monadic Datalog and
into Datalog (with inequalities).
To prove the results listed above, we introduce two new notions that are of independent
interest and general utility. The first one is materializability of a TBox T , which means
that evaluating a query over an ABox A w.r.t. T can be reduced to query evaluation in a
single model of A and T (a materialization). Note that such models play a crucial role in
the context of Horn DLs, where they are often called canonical models or universal models.
In contrast to the Horn DL case, however, we only require the existence of such a model
without making any assumptions about its form or construction.
5. If an ALCF I-TBox T is not materializable, then CQ-evaluation w.r.t. T is coNP-hard.
We also investigate the nature of materializations. It turns out that if a TBox is materializable
for one of the considered query languages, then it is materializable also for all others. The
concrete materializations, however, need not agree. To obtain these results, we characterize
CQ-materializations in terms of homomorphisms and ELIQ-materializations in terms of
4
CARSTEN LUTZ AND FRANK WOLTER
simulations (an ELIQ is an ELI-instance query, thus the DL version of an acyclic CQ, with
a single answer variable).
Perhaps in contrary to the intuitions that arise from the experience with Horn DLs,
materializability of a TBox T is not a sufficient condition for query evaluation w.r.t. T to
be in PTime (unless PTime = NP) since the existing materialization might be hard to
compute. This leads us to study the notion of unraveling tolerance of a TBox T , meaning
that answers to acyclic CQs over an ABox A w.r.t. T are preserved under unraveling the
ABox A. In CSP, unraveling tolerance corresponds to the existence of tree obstructions,
a notion that characterizes the well known arc consistency condition and rewritability
into monadic Datalog [FV98, Kro10a]. It can be shown that every TBox formulated in
Horn-ALCF I (the intersection of ALCF I and Horn-SHIQ) is unraveling tolerant and that
there are unraveling tolerant TBoxes which are not equivalent to any Horn-ALCF I-TBox.
Thus, the following result yields a rather general (and uniform!) PTime upper bound for
CQ-evaluation.
6. If an ALCF I-TBox T is unraveling tolerant, then query evaluation w.r.t. T is in PTime.
Although the above result is rather general, unraveling tolerance of a TBox T is not a
necessary condition for CQ-evaluation w.r.t. T to be in PTime (unless PTime = NP).
However, for ALCF I-TBoxes T of depth one, being materializable and being unraveling
tolerant turns out to be equivalent. For such TBoxes, we thus obtain that CQ-evalutation
w.r.t. T is in PTime iff T is materializable iff T is unraveling tolerant while, otherwise,
CQ-evaluation w.r.t. T is coNP-hard. This establishes the first main result above.
Our framework also allows one to formally capture some intuitions and beliefs commonly
held in the context of CQ-answering in DLs. For example, we show that for every ALCF ITBox T , CQ-evaluation is in PTime iff PEQ-evaluation is in PTime iff ELIQ-evaluation
is in PTime, and the same is true for coNP-hardness and for rewritability into Datalog
and into monadic Datalog. In fact, the use of multiple query languages and in particular
of ELI-instance queries does not only yield additional results, but is at the heart of our
proof strategies. Another interesting observation in this spirit is that an ALCF I-TBox is
materializable iff it is convex, a condition that is also called the disjunction property and
plays a central role in attaining PTime complexity for standard reasoning in Horn DLs such
as EL, DL-Lite, and Horn-SHIQ; see for example [BBL05, KL07] for more details.
This paper is a significantly extended and revised version of the conference publication [LW12].
Related Work. An early reference on data complexity in DLs is [Sch93], showing coNPhardness of ELQs in the fragment ALE of ALC (an ELQ is an ELIQ in which all edges
are directed away from the answer variable). A coNP upper bound for ELIQs in the
much more expressive DL SHIQ was obtained in [HMS07] and generalized to CQs in
[GLHS08]. Horn-SHIQ was first defined in [HMS07], where also a PTime upper bound
for ELIQs is established; the generalization to CQs can be found in [EGOS08]. See also
[KL07, Ros07, OCE08, CDL+ 13] and references therein for the data complexity in DLs and
[BGO10, BMRT11] for related work on the guarded fragment and on existential rules.
To the best of our knowledge, the conference version of this paper was first to initiate the
study of data complexity in ontology-mediated querying at the level of individual TBoxes and
the first to observe a link between this area and CSP. There is, however, a certain technical
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
5
similarity to the link between view-based query processing for regular path queries (RPQs)
and CSP found in [CGLV00, CGLV03b, CGLV03a]. In this case, the recognition problem
for perfect rewritings for RPQs can be polynomially reduced to non-uniform CSP and vice
versa. On the level of OMQs, the data complexity of ontology-mediated querying with DLs
has been studied in [BtCLW14], see also [FKL17]; also here, a connection to CSP plays a
central role. In [LSW13, LSW15], the non-uniform data complexity of ontology-mediated
query answering is studied in the case where the TBox is formulated in an inexpressive DL
of the DL-Lite or EL family and where individual predicates in the data can be given a
closed-world reading, which also gives rise to coNP-hardness of query evaluation; while
[LSW13] is considering the level of TBoxes, [LSW15] treats the level of OMQs, establishing a
connection to surjective CSPs. Rewritability into Datalog for atomic queries and at the level
of OMQs has also been studied in [KNC16]. Finally, we mention [LS17] where a complete
classification of the data complexity of OMQs (also within PTime) is achieved when the
TBox is formulated in EL and the actual queries are atomic queries.
Recently, the data complexity at the level of TBoxes has been studied also in the guarded
fragment and in the two-variable guarded fragment of FO with counting [HLPW17a]. This
involves a generalization of the notions of materializability and unraveling tolerance and
leads to a variety of PTime/coNP-dichotomy results. In particular, our dichotomy between
Datalog-rewritability and coNP is extended from ALCIF-TBoxes of depth one to ALCHIFTBoxes of depth two. Using a variant of Ladner’s Theorem, several non-dichotomy results
for weak fragments of the two-variable guarded fragment with counting of depth two are
established and it is shown that PTime data complexity of query evaluation is undecidable.
For ALCHIQ-TBoxes of depth one, though, PTime data complexity of query evaluation
and, equivalently, rewritability into Datalog (with inequalities) is proved to be decidable. In
[HLPW17b], the results presented in this paper have been used to show that whenever an
ALCIF-TBox of depth one enjoys PTime query evaluation, then it can be rewritten into a
Horn-ALCIF-TBox that gives the same answers to CQs (the converse is trivial). It is also
proved that this result does not hold in other cases such as for ALCHIF-TBoxes of depth
one.
The work on CSP dichotomies started with Schaefer’s PTime/NP-dichotomy theorem,
stating that every CSP defined by a two element template is in PTime or NP-hard [Sch78].
Schaefer’s theorem was followed by dichotomy results for CSPs with (undirected) graph
templates [HN90] and several other special cases, leading to the widely known Feder-Vardi
conjecture which postulates a PTime/NP-dichotomy for all CSPs, independently of the size
of the template [FV98]. The conjecture has recently been confirmed [Bul17, Zhu17] using
an approach to studying the complexity of CSPs via universal algebra [BJK05]. Interesting
results have also been obtained for other complexity classes such as AC0 [ABI+ 05, LLT07].
2. Preliminaries
We introduce the relevant description logics and query languages, define the fundamental
notions studied in this paper, and illustrate them with suitable examples.
We shall be concerned with the DL ALC and its extensions ALCI, ALCF, and ALCF I.
Let NC , NR , and NI denote countably infinite sets of concept names, role names, and individual
names, respectively. ALC-concepts are constructed according to the rule
C, D := > | ⊥ | A | C u D | C t D | ¬C | ∃r.C | ∀r.C
6
CARSTEN LUTZ AND FRANK WOLTER
where A ranges over NC and r ranges over NR . ALCI-concepts admit, in addition, inverse
−
roles from the set N−
R = {r | r ∈ NR }, which can be used in place of role names. Thus,
−
A u ∃r .∀s.B is an example of an ALCI-concept. To avoid heavy notation, we set r− := s
if r = s− for a role name s; in particular, we thus have (r− )− = r.
In DLs, ontologies are formalized as TBoxes. An ALC-TBox is a finite set of concept
inclusions (CIs) C v D, where C, D are ALC concepts, and ALCI-TBoxes are defined
analogously. An ALCF-TBox (resp. ALCF I-TBox) is an ALC-TBox (resp. ALCI-TBox)
that additionally admits functionality assertions func(r), where r ∈ NR (resp. r ∈ NR ∪ N−
R ),
declaring that r is interpreted as a partial function. Note that there is no such thing as
an ALCF-concept or an ALCF I-concept, as the extension with functional roles does not
change the concept language.
An ABox A is a non-empty finite set of assertions of the form A(a) and r(a, b) with
A ∈ NC , r ∈ NR , and a, b ∈ NI . In some cases, we drop the finiteness condition on ABoxes and
then explicitly speak about infinite ABoxes. We use Ind(A) to denote the set of individual
names used in the ABox A and sometimes write r− (a, b) ∈ A instead of r(b, a) ∈ A.
The semantics of DLs is given by interpretations I = (∆I , ·I ), where ∆I is a non-empty
set and ·I maps each concept name A ∈ NC to a subset AI of ∆I and each role name r ∈ NR
to a binary relation rI on ∆I . The extension (r− )I of r− under the interpretation I is
defined as the converse relation (rI )−1 of rI and the extension C I ⊆ ∆I of concepts under
the interpretation I is defined inductively as follows:
>I
⊥I
(¬C)I
= ∆I
= ∅
= ∆I \ C I
(C u D)I
= C I ∩ DI
(C t D)I
= C I ∪ DI
(∃r.C)I
= {d ∈ ∆I | ∃d0 ∈ ∆I : (d, d0 ) ∈ rI and d0 ∈ C I }
(∀r.C)I
= {d ∈ ∆I | ∀d0 ∈ ∆I : (d, d0 ) ∈ rI implies d0 ∈ C I }
An interpretation I satisfies a CI C v D if C I ⊆ DI , an assertion A(a) if a ∈ AI ,
an assertion r(a, b) if (a, b) ∈ rI , and a functionality assertion func(r) if rI is a partial
function. Note that we make the standard name assumption, that is, individual names are
not interpreted as domain elements (like first-order constants), but as themselves. This
assumption is common both in DLs and in database theory. The results in this paper do
not depend on it.
An interpretation I is a model of a TBox T if it satisfies all CIs in T and I is a model
of an ABox A if all individual names from A are in in ∆I and I satisfies all assertions in A.
We call an ABox A consistent w.r.t. a TBox T if A and T have a joint model.
We consider several query languages. A positive existential query (PEQ) q(~x) is a
first-order formula with free variables ~x = x1 , . . . , xn constructed from atoms A(x) and
r(x, y) using conjunction, disjunction, and existential quantification, where A ∈ NC , r ∈ NR ,
and x, y are variables. The variables in ~x are the answer variables of q(~x). A PEQ without
answer variables is Boolean. An assignment π for q(~x) in an interpretation I is a mapping
from the variables that occur in q(~x) to ∆I . A tuple ~a = a1 , . . . , an in Ind(I) is an answer to
q(~x) in I if there exists an assigment π for q(~x) in I such that I |=π q(~x) (in the standard
first-order sense) and π(xi ) = ai for 1 ≤ i ≤ n. In this case, we write I |= q(~a). A tuple
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
7
~a ∈ Ind(A), A an ABox, is a certain answer to q(~x) in A w.r.t. a TBox T , in symbols
T , A |= q(~a), if I |= q(~a) for all models I of T and A. Computing certain answers to a query
in the sense just defined is the main querying problem we are interested in. Although this
paper focusses on the theoretical aspects of query answering, we given a concrete example
that illustrates the usefulness of query answering with DL ontologies.
Example 2.1. Let
T = {Professer v Academic, Professor v ∃gives.Course}
A = {Student(john), supervisedBy(john, mark), Professor(mark)}
q(x, y) = ∃z Student(x) ∧ supervisedBy(x, y) ∧ Academic(y) ∧ gives(y, z) ∧ Course(z)
Thus the query asks to return all pairs that consist of a student x and an academic y such
that x is supervised by y and y gives a course. Although this information is not directly
present in the ABox, because of the TBox it is easy to see that (john, mark) is a certain
answer.
Apart from PEQs, we also study several fragments thereof. A conjunctive query (CQ) is
a PEQ without disjunction. We generally assume that a CQ q(~x) takes the form ∃~y ϕ(~x, ~y ),
where ϕ(~x, ~y ) is a conjunction of atoms of the W
form A(x) and r(x, y). It is easy to see that
every PEQ q(~x) is equivalent to a disjunction i∈I qi (~x), where each qi (~x) is a CQ (such a
disjunction is often called a union of conjunctive queries, or UCQ).
To introduce simple forms of CQs that play a crucial role in this paper, we recall two
further DLs that we use here for mainly querying purposes. EL-concepts are constructed
from NC and NR according to the syntax rule
C, D := > | A | C u D | ∃r.C
and ELI-concepts additionally admit inverse roles. An EL-TBox is a finite set of concept
inclusions C v D with C and D EL-concepts, and likewise for ELI-TBoxes.
We now use EL and ELI to define restricted classes of CQs. If C is an ELI-concept
and x a variable, then C(x) is called an ELI query (ELIQ); if C is an EL-concept, then
C(x) is called an EL query (ELQ). Note that every ELIQ can be regarded as an acyclic CQ
with one answer variable, and indeed this is an equivalent definition of ELIQs; in the case of
ELQs, it is additionally the case that all edges are directed away from the answer variable.
For example, the ELIQ ∃r.(A u ∃s− .B)(x) is equivalent to the acyclic CQ
∃y1 ∃y2 (r(x, y1 ) ∧ A(y1 ) ∧ s(y2 , y1 ) ∧ B(y2 )).
In what follows, we will not distinguish between an ELIQ and its translation into an acyclic
CQ with one answer variable and freely apply notions introduced for PEQs also to ELIQs
and ELQs. We also sometimes slightly abuse notation and use PEQ to denote the set of all
positive existential queries, and likewise for CQ, ELIQ, and ELQ.
Example 2.2.
(1) Let T∃,r = {A v ∃r.A} and q(x) = ∃r.A(x). Then we have for any ABox A, T∃,r , A |= q(a)
iff A(a) ∈ A or there are r(a, b), A(b) ∈ A.
(2) Let T∃,l = {∃r.A v A} and q(x) = A(x). For any ABox A, T∃,l , A |= q(a) iff there is an
r-path in A from a to some b with A(b) ∈ A; that is, there are r(a0 , a1 ), . . . , r(an−1 , an ) ∈ A,
n ≥ 0, with a0 = a, an = b, and A(b) ∈ A.
8
CARSTEN LUTZ AND FRANK WOLTER
(3) Consider an undirected graph G represented as an ABox A with assertions r(a, b), r(b, a) ∈
A iff there is an edge between a and b. Let A1 , . . . , Ak , M be concept names. Then G is
k-colorable iff Tk , A 6|= ∃x M (x), where
Tk = {Ai u Aj v M | 1 ≤ i < j ≤ k} ∪
{Ai u ∃r.Ai v M | 1 ≤ i ≤ k} ∪
{> v t Ai }.
1≤i≤k
Instead of actually computing certain answers to queries, we concentrate on the query
evaluation problem, which is the decision problem version of query answering. We next
introduce this problem along with associated notions of complexity. An ontology-mediated
query (OMQ) is a pair (T , q(~x)) with T a TBox T and q(~x) a query. The query evaluation
problem for (T , q(~x)) is to decide, given an ABox A and ~a in Ind(A), whether T , A |= q(~a). We
shall typically be interested in joint complexity bounds for evaluating all OMQs formulated
in a query language Q of interest w.r.t. a given TBox T .
Definition 2.3. Let T be an ALCF I-TBox and let Q ∈ {CQ, PEQ, ELIQ, ELQ}. Then
• Q-evaluation w.r.t. T is in PTime if for every q(~x) ∈ Q, the query evaluation problem
for (T , q(~x)) is in PTime.
• Q-evaluation w.r.t. T is coNP-hard if there exists q(~x) ∈ Q such that the query evaluation
problem for (T , q(~x)) is coNP-hard.
Note that one should not think of ‘Q-evaluation w.r.t. T ’ as a decision problem since,
informally, this is a collection of infinitely many decision problems, one for each query in Q.
Instead, one should think of ‘Q-evaluation w.r.t. T to be in PTime’ (or coNP-hard) as a
property of T .
Example 2.4.
(1) PEQ-evaluation w.r.t. the TBoxes T∃,r and T∃,l from Example 2.2 is in PTime. This
follows from the fact that these TBoxes are EL-TBoxes (TBoxes using only EL-concepts)
and it is well known that PEQ-evaluation w.r.t. EL-TBoxes is in PTime [KL07].
(2) Consider the TBoxes Tk from Example 2.2 that express k-colorability using the query
∃x M (x). For k ≥ 3, CQ-evaluation w.r.t. Tk is coNP-hard since k-colorability is NP-hard.
However, in contrast to the tractability of 2-colorability, CQ-evaluation w.r.t. T2 is still
coNP-hard. This follows from Theorem 3.11 below and, intuitively, is the case because
T2 ‘entails a disjunction’: for A = {B(a)}, we have T2 , A |= A1 (a) ∨ A2 (a), but neither
T2 , A |= A1 (a) nor T2 , A |= A2 (a).
In addition to the classification of TBoxes according to whether query evaluation is in
PTime or coNP-hard, we are also interested in whether OMQs based on the TBox are
rewritable into more classical database querying languages, in particular into Datalog and
into monadic Datalog.
A Datalog rule ρ has the form S(~x) ← R1 (~y1 ) ∧ · · · ∧ Rn (~yn ) where n > 0, S is a relation
symbol, and R1 , . . . , Rn are relation symbols, that is, concept names and role names. We
refer to S(~x) as the head of ρ and R1 (~y1 ) ∧ · · · ∧ Rn (~yn ) as its body. Every variable in the
head of ρ is required to occur also in its body. A Datalog program Π is a finite set of Datalog
rules with a selected goal relation goal that does not occur in rule bodies. Relation symbols
that occur in the head of at least one rule are called intensional relation symbols (IDBs), the
remaining symbols are called extensional relation symbols (EDBs). Note that, by definition,
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
9
goal is an IDB. The arity of the program is the arity of the goal relation. Programs of arity
zero are called Boolean. A Datalog program that uses only IDBs of arity one, with the
possible exception of the goal relation, is called monadic.
For an ABox A, a Datalog program Π, and ~a from Ind(A) of the same length as the
arity of goal, we write A |= Π(~a) if Π returns ~a as an answer on A, defined in the usual way
[CGT89]. A (monadic) Datalog program Π is a (monadic) Datalog-rewriting of an OMQ
(T , q(~x)) if for all ABoxes A and ~a from Ind(A), T , A |= q(~a) iff A |= Π(~a). In this case
the OMQ (T , q(~x)) is called (monadic) Datalog-rewritable. When working with DLs such
as ALCF I that include functional roles, it is more natural to admit the use of inequalities
in the bodies of Datalog rules instead of working with ‘pure’ programs. We refer to such
extended programs as (monadic) Datalog6= programs and accordingly speak of (monadic)
Datalog6= -rewritability.
Example 2.5.
(1) The OMQ (T∃,l , A(x)) from Example 2.2 expressing a form of reachability is rewritable
into the monadic Datalog program
goal(x) ← P (x),
P (x) ← A(x),
P (x) ← r(x, y) ∧ P (y).
(2) The OMQ (Tk , ∃x M (x)) from Example 2.2 is Datalog-rewritable when k = 2 since
non-2-colorability can be expressed by a Datalog program (but not as a monadic one). For
k ≥ 3, non-k-colorability cannot be expressed by a Datalog program (in fact, not even by a
Datalog6= program) [ACY91].
(3) The OMQ ({func(r)}, ∃x M (x)) is rewritable into the monadic Datalog6= program
goal() ← r(x, y1 ) ∧ r(x, y2 ) ∧ y1 6= y2 ,
goal() ← M (x)
but is not rewritable into pure Datalog.
Definition 2.6. Let T be an ALCF I-TBox and let Q ∈ {CQ, PEQ, ELIQ, ELQ}. Then T
is (monadic) Datalog6= -rewritable for Q if (T , q(~x)) is (monadic) Datalog6= -rewritable for
every q(~x) ∈ Q.
We would like to stress that the extension of Datalog to Datalog6= makes sense only in
the presence of functional roles. In fact, it follows from the CSP connection established in
Section 6 and the results in [FV03] that for ALCI-TBoxes, Datalog6= -rewritability for Q
agrees with Datalog-rewritability for Q, for all query classes Q considered in this paper.
Example 2.7. It is folklore that every EL-TBox is monadic Datalog-rewritable for ELQ,
ELIQ, CQs, and PEQs. Thus, this applies in particular to the EL-TBoxes T∃,l and T∃,r from
Example 2.2. A concrete construction of Datalog-rewritings for ELIQs can be found in the
proof of Theorem 4.6 below. In contrast, the ALC-TBox Tk from Example 2.2 is not Datalog6= rewritable for ELQ when k ≥ 3 since the OMQ (Tk , ∃x M (x)) is not Datalog-rewritable, by
Example 2.5 (2).
Datalog6= -programs can be evaluated in PTime [CGT89] in data complexity, and thus
Datalog6= -rewritability for Q of a TBox T implies that Q-evaluation w.r.t. T is in PTime in
data complexity. We shall see later that the converse direction does not hold in general.
We will often be concerned with homomorphisms between ABoxes and between interpretations, defined next. Let A and B be ABoxes. A function h : Ind(A) → Ind(B) is a
homomorphism from A to B if it satisfies the following conditions:
10
CARSTEN LUTZ AND FRANK WOLTER
(1) A(a) ∈ A implies A(h(a)) ∈ B and
(2) r(a, b) ∈ A implies r(h(a), h(b)) ∈ B.
We say that h preserves I ⊆ NI if h(a) = a for all a ∈ I. Homomorphisms from an
interpretation I to an interpretation J are defined analogously as functions h : ∆I → ∆J .
Note that these two notions are in fact identical since, up to presentation, ABoxes and finite
interpretations are the same thing. In what follows we will not always distinguish between
the two presentations.
3. Materializability
We introduce materializability as a central notion for analyzing the complexity and rewritability of TBoxes. A materialization of a TBox T and ABox A for a class of queries Q is a
(potentially infinite) model of T and A that gives the same answers to queries in Q as T and
A do. It is not difficult to see that a materialization for ELIQs is not necessarily a materialization for CQs and that a materialization for ELQs is not necessarily a materialization for
ELIQs. We shall call a TBox T materializable for a query language Q if for every ABox A
that is consistent w.r.t. T , there is a materialization of T and A for Q. Interestingly, we show
that materializability of ALCF I-TBoxes does not depend on whether one considers ELIQs,
CQs, or PEQs. This result allows us to simply talk about materializable TBoxes, independently of the query language considered. The fundamental result linking materializability of
a TBox to the complexity of query evaluation is that ELIQ-evaluation is coNP-hard w.r.t.
non-materializable ALCF I-TBoxes. As another application of materializability, we show
that for ALCF I-TBoxes, PTime query evaluation, coNP-hardness of query evaluation,
and Datalog6= -rewritability also do not depend on the query language. In the case of ALCF,
materializability for ELIQs additionally coincides with materializability for ELQs.
Definition 3.1. Let T be an ALCF I-TBox and Q ∈ {CQ, PEQ, ELIQ, ELQ}. Then
(1) a model I of T and an ABox A is a Q-materialization of T and A if for all queries
q(~x) ∈ Q and ~a ⊆ Ind(A), we have I |= q(~a) iff T , A |= q(~a);
(2) T is Q-materializable if for every ABox A that is consistent w.r.t. T , there exists a
Q-materialization of T and A.
In Point (1) of Definition 3.1, it is important that the materialization I of T and A is a
model of T and A. In fact, for an ABox A that is consistent w.r.t. T , we can always find an
interpretation I such that for every CQ q(~x) and ~a ⊆ Ind(A), I |= q(~a) iff T , A |= q(~a). In
particular, the direct product of all (up to isomorphisms) countable models of T and A can
serve as such an I. However, the interpretation is, in general, not a model of T .
Note that a Q-materialization can be viewed as a more abstract version of the canonical
or minimal or universal model as often used in the context of ‘Horn DLs’ such as EL
and DL-Lite [LTW09, KLT+ 10, BO15] and more expressive ontology languages based on
tuple-generating dependencies (tgds) [CGK13] as well as in data exchange [FKMP05]. In
fact, the ELQ-materialization in the next example is exactly the ‘compact canonical model’
from [LTW09].
Example 3.2.
(1) Let T∃,l = {∃r.A v A} be as in Example 2.2 and let A be an ABox. Let I be the
interpretation obtained from A by adding to AI all a ∈ Ind(A) such that there exists an
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
11
r-path from a to some b with A(b) ∈ A. Then I is a PEQ-materialization of T and A and
so T is PEQ-materializable.
(2) Let T∃,r = {A v ∃r.A} be as in Example 2.2 and let A be an ABox with at least one
assertion of the form A(a). To obtain an ELQ-materialization I of T and A, start with A
as an interpretation, add a fresh domain element dr to ∆I and to AI , and extend rI with
(a, dr ) and (dr , dr ) for all A(a) ∈ A. Thus T∃,r is ELQ-materializable.
(3) The TBox T = {A v A1 t A2 } is not ELQ-materializable. To see this let A = {A(a)}.
Then no model I of T and A is an ELQ-materialization of T and A as it satisfies a ∈ AI1
or a ∈ AI2 but neither T , A |= A1 (a) nor T , A |= A2 (a).
Trivially, every PEQ-materialization is a CQ-materialization, every CQ-materialization
is an ELIQ-materialization and every ELIQ-materialization is an ELQ-materialization.
Conversely, it follows directly from the fact that each PEQ is equivalent to a disjunction of
CQs that every CQ-materialization is also a PEQ-materialization. In contrast, the following
example demonstrates that ELQ-materializations are different from ELIQ-materializations.
A similar argument separates ELIQ-materializations from CQ-materializations.
Example 3.3. Let T∃,r be as in Example 3.2,
A = {B1 (a), B2 (b), A(a), A(b)} and
q(x) = (B1 u ∃r.∃r− .B2 )(x),
Then the ELQ-materialization I from Example 3.2 (2) is not a Q-materialization for any
Q from the set of query languages ELIQ, CQ, PEQ. For example, we have I |= q(a), but
T , A 6|= q(a). An ELIQ/CQ/PEQ-materialization of T and A is obtained by unfolding I
(see below): instead of using only one additional domain element dr as a witness for ∃r.A,
we attach to both a and b an infinite r-path of elements that satisfy A. Note that every
CQ/PEQ-materialization of T∃,r and A must be infinite.
We will sometimes restrict our attention to materializations I that are countable and
generated, i.e, every d ∈ ∆I is reachable from some a ∈ ∆I ∩ NI in the undirected graph
[
GI = (∆I , {{d, d0 } | (d, d0 ) ∈
rI }).
r∈NR
The following lemma shows that we can make that assumption without loss of generality.
Lemma 3.4. Let T be an ALCF I-TBox, A an ABox, and Q ∈ {CQ, PEQ, ELIQ, ELQ}.
If I is a Q-materialization of T and A, then there exists a subinterpretation J of I that is
a countable and generated Q-materialization of T and A.
Proof. Let I be a Q-materialization of T and A. To construct J we apply a standard
selective filtration procedure to I. More precisely, we identify
S a sequence Ind(A) = S0 ⊆
S1 ⊆ · · · ⊆ ∆I and then define J to be the restriction of I to i Si . Let C be the set of all
concepts of the form ∃r.C that occur in T and of all concepts ∃r.¬C such that ∀r.C occurs
in T . Assume Si has already been defined. Then define Si+1 as the union of Si and, for
every d ∈ Si and concept ∃r.C ∈ C with d ∈ (∃r.C)I , an arbitrary d0 ∈ ∆I with (d, d0 ) ∈ rI
and d0 ∈ C I (unless such a d0 exists already in Si ). It is easy to see that J is a countable
and generated Q-materialization of T and A.
12
CARSTEN LUTZ AND FRANK WOLTER
3.1. Model-Theoretic Characterizations of Materializability. We characterize materializations using simulations and homomorphisms. This sheds light on the nature of materializations and establishes a close connection between materializations and initial models as studied in model theory, algebraic specification, and logic programming [Mal71, MG85, Mak87].
A simulation from an interpretation I1 to an interpretation I2 is a relation S ⊆ ∆I1 ×∆I2
such that
(1) for all A ∈ NC : if d1 ∈ AI1 and (d1 , d2 ) ∈ S, then d2 ∈ AI2 ;
(2) for all r ∈ NR : if (d1 , d2 ) ∈ S and (d1 , d01 ) ∈ rI1 , then there exists d02 ∈ ∆I2 such that
(d01 , d02 ) ∈ S and (d2 , d02 ) ∈ rI2 ;
(3) for all a ∈ ∆I1 ∩ NI : a ∈ ∆I2 and (a, a) ∈ S.
Note that, by Condition (3), domain elements that are individual names need to be respected
by simulations while other domain elements need not. In database parlance, the latter are
thus treated as labeled nulls, that is, while their existence is important, their identity is not.
We call a simulation S an i-simulation if Condition (2) is satisfied also for inverse roles.
Note that S is a homomorphism preserving ∆I1 ∩ NI if S is a function with domain ∆I . We
remind the reader of the following characterizations of ELQs using simulations, ELIQs using
i-simulations, and CQs using homomorphisms (see e.g. [LW10]). An interpretation I has
finite outdegree if the undirected graph GI has finite outdegree.
Lemma 3.5. Let I and J be interpretations such that ∆I ∩ NI is finite, I is countable and
generated, and J has finite outdegree. Then the following conditions are equivalent (where
none of the assumed conditions on I and J is required for (2) ⇒ (1)).
(1) For all ELIQs C(x) and a ∈ ∆I ∩ NI : if I |= C(a), then J |= C(a);
(2) There is an i-simulation from I to J .
The same equivalence holds when ELIQs and i-simulations are replaced by ELQs and
simulations, respectively. Moreover, the following conditions are equivalent (where none of
the assumed conditions on I and J is required for (5) ⇒ (3)).
(3) For all PEQs q(~x) and ~a ⊆ ∆I ∩ NI : if I |= q(~a), then J |= q(~a);
(4) For all CQs q(~x) and ~a ⊆ ∆I ∩ NI : if I |= q(~a), then J |= q(~a);
(5) There is a homomorphism from I to J preserving ∆I ∩ NI .
Proof. We prove the equivalence of (3)-(5). The equivalence of (1) and (2) is similar (both
for ELIQs and ELQs) but simpler and left to the reader. The implication (3) ⇒ (4) is
trivial. For the proof of (4) ⇒ (5), assume that I is countable and generated and let J have
finite outdegree. We first assume that only a finite set Σ of concept and role names have a
non-empty interpretation in I and then generalize the result to arbitrary I. Assume that
(4) holds. First observe that for every finite subset X of ∆I there is a homomorphism hX
preserving X ∩ NI from the subinterpretation IX of I induced by X into J : associate with
every d ∈ X a variable xd and regard IX as the CQ
^
^
qX (~x) = ∃~y
A(xd ) ∧
r(xd , xd0 ),
d∈X∩AI
(d,d0 )∈(X×X)∩rI
where ~x comprises the variables in {xa | a ∈ X ∩ NI } and ~y comprises the variables xd with
d ∈ X \ NI (qX is a CQ by our assumption that only finitely many concept and role names
have non-empty interpretation). For the assignment π(xd ) = d, we have I |=π ϕ(~x, ~y ). Thus
I |=π qX (~x) and so, by (2), J |=π qX (~x). Consequently, there exists an assignment π 0 for
qX (~x) in J which coincides with π on {xa | a ∈ X ∩ NI } such that J |=π0 ϕ(~x, ~y ). Let
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
13
hX (d) = π 0 (d) for d ∈ X. Then hX is a homomorphism from IX to J preserving X ∩ NI ,
as required.
We now lift the homomorphisms hX to a homomorphism h from I to J preserving
I
∆ ∩ NI . Since I is countable and generated,
there exists a sequence X0 ⊆ X1 ⊆ · · · of finite
S
subsets of ∆I such that X0 = ∆I ∩ NI , i≥0 Xi = ∆I , and for all d ∈ Xi there exists a path
in Xi from some a ∈ X0 to d.
By the observation above, we find homomorphisms hXi from IXi to J preserving Xi ∩NI ,
for i ≥ 0. Let d0 , d1 . . . be an enumeration of ∆I . We define the required homomorphism
h as the limit of a sequence h0 ⊆ h1 ⊆ · · · , where each hn has domain {d0 , . . . , dn } and
where we ensure for each hn and all d ∈ {d0 , . . . , dn } that there are infinitely many j with
hn (d) = hXj (d). Observe that since J has finite outdegree and since for all d ∈ Xi , there
exists a path in Xi from some a ∈ X0 to d, for each d ∈ ∆I there exist only finitely many
distinct values in {hXi (d) | i ≥ 0}. By the pigeonhole principle, there thus exist infinitely
many j with the same value hXj (d). For h0 (d0 ) we take such a value for d0 . Assume hn has
been defined and assume that the set I = {j | hn (d) = hXj (d) for all d ∈ {d0 , . . . , dn }} is
infinite. Again by the pigeonhole principle, we find a value e ∈ ∆J such that hX
Sj (dn+1 ) = e
for infinitely many j ∈ I. We set hn+1 (dn+1 ) = e. The function h = i≥0 h0 is a
homomorphism from I to J preserving ∆I ∩ NI , as required.
To lift this result to arbitrary interpretations I, it is sufficient to prove that the
homomorphisms hX still exist. This can be shown using again the pigeonhole principle. Let
X ⊆ ∆I be finite. We may assume that for each d ∈ X, there exists a path in X from
some a ∈ X ∩ NI , to d. We have shown that for each finite set Σ of concept and role names,
Σ
Σ
there exists a homomorphism hΣ
X from the Σ-reduct IX of IX to J (IX interprets only the
symbols in Σ as non-empty). Since J has finite outdegree, infinitely many hΣ
X coincide. A
straightforward modification of the pigeonhole argument above can now be used to construct
the required homomorphism hX .
For the proof of (5) ⇒ (3), assume I |= q(~a) and let h be a homomorphism from I to
J preserving ∆I ∩ NI . Let π be an assignment for q(~x) in I witnessing I |= q(~a). Then the
composition h ◦ π is an assignment for q(~x) in J witnessing J |= q(~a).
For the next steps, we need some observations regarding the unfolding of interpretations
into forest-shaped interpretations. Let us first make precise what we mean by unfolding. The
i-unfolding of an interpretation I is an interpretation J defined as follows. The domain ∆J
of J consists of all words d0 r1 . . . rn dn with n ≥ 0, each di from ∆I and each ri a (possibly
inverse) role such that
(a) di ∈ NI iff i = 0;
I
(b) (di , di+1 ) ∈ ri+1
for 0 ≤ i < n;
−
(c) if ri = ri+1 , then di−1 6= di+1 for 0 < i < n.
For d0 · · · dn ∈ ∆J , we set tail(d0 · · · dn ) = dn . Now set
AJ
rJ
= {w ∈ ∆J | tail(w) ∈ AI }
= (rI ∩ (NI × NI )) ∪
{(σ, σrd) | σ, σrd ∈ ∆J } ∪ {(σr− d, σ) | σ, σr− d ∈ ∆J }
for all A ∈ NC
for all r ∈ NR .
We say that an interpretation I is i-unfolded if it is isomorphic to its own i-unfolding. Clearly,
every i-unfolding of an interpretation is i-unfolded.
14
CARSTEN LUTZ AND FRANK WOLTER
For ALCF-TBoxes, it is not required to unfold along inverse roles. This is reflected
in the unfolding of an interpretation I, where in contrast to the i-unfolding we use as the
domain the set of all words d0 r1 . . . rn dn with n ≥ 0, each di from ∆I , and each ri a role
name such that Conditions (a) and (b) above are satisfied. The interpretation of concept
and role names remains the same. We call an interpretation I unfolded if it is isomorphic to
its own unfolding. The following lemma summarizes the main properties of unfoldings. Its
proof is straightforward and left to the reader.
Lemma 3.6. Let I be an interpretation, I i its i-unfolding, and I u its unfolding. Then for
every interpretation J , the following conditions are satisfied:
i
(1) the function f (w) := tail(w), w ∈ ∆I , is a homomorphism from I i to I preserving
∆ I ∩ NI ;
u
(2) the function f (w) := tail(w), w ∈ ∆I , is a homomorphism from I u to I preserving
∆ I ∩ NI ;
(3) if there is an i-simulation from I to J , then there is a homomorphism from I i to J
preserving ∆I ∩ NI ;
(4) if there is a simulation from I to J , then there is a homomorphism from I u to J
preserving ∆I ∩ NI ;
(5) if I is a model of T and A with T an ALCF I-TBox, then I i is a model of T and A;
(6) if I is a model of T and A with T an ALCF-TBox, then I u is a model of T and A.
An interpretation I is called hom-initial in a class K of interpretations if for every
J ∈ K, there exists a homomorphism from I to J preserving ∆I ∩ NI . I is called sim-initial
(i-sim-initial) in a class K of interpretations if for every J ∈ K, there exists a simulation
(i-simulation) from I to J . The following theorem provides the announced characterization
of materializations in terms of simulations and homomorphisms. In the following, the class
of all models of T and A is denoted by Mod(T , A).
Theorem 3.7. Let T be an ALCF I-TBox, A an ABox, and let I ∈ Mod(T , A) be countable
and generated. Then I is
(1) an ELIQ-materialization of T and A iff it is i-sim-initial in Mod(T , A);
(2) a CQ-materialization of T and A iff it is a PEQ-materialization of T and A iff it is
hom-initial in Mod(T , A);
(3) an ELQ-materialization of T and A iff it is sim-initial in Mod(T , A), provided that T
is an ALCF-TBox.
The ‘only if ’ directions of all three points hold without any of the assumed conditions on I.
Proof. We show that (1) follows from Lemma 3.5 and Lemma 3.6; (2) and (3) can be proved
similarly.
We start with the direction from right to left. Assume that I is i-sim-initial in Mod(T , A).
Since I is a model of T and A, we have I |= C(a) whenever T , A |= C(a) for any ELIQ
C(x) and a ∈ Ind(A). Conversely, if T , A 6|= C(a) then there exists a model J of T and A
such that J 6|= C(a). There is an i-simulation from I to J . Thus, by the implication (2) ⇒
(1) from Lemma 3.5, we have I 6|= C(a) as required.
For the direction from left to right, assume that I is a materialization of T and A and
take a model J of T and A. We have to construct an i-simulation from I to J . It actually
suffices to construct an i-simulation from I to the i-unfolding J i of J : by Point (3) of
Lemma 3.6, there is a homomorphism from J i to J and the composition of an i-simulation
with a homomorphism is again an i-simulation.
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
15
To obtain an i-simulation from I to J i , we first identify a subinterpretation J 0 of J i
that has finite outdegree and is still a model of T and A. By the implication (1) ⇒ (2) from
Lemma 3.5 and since I is a materialization, there must then be an i-simulation from I to
J 0 . Clearly this is also an i-simulation from I to J i and we are done.
It thus remains to construct J 0 , which is done by applying selective filtration to J i in
exactly the same way as in the proof of Lemma 3.4. It can be verified that the outdegree of
the resulting subinterpretation J 0 of J i is bounded by |T | + |A| and, therefore, finite. By
construction, J 0 ∈ Mod(T , A).
The following example shows that the generatedness condition in Theorem 3.7 cannot
be dropped. We leave it open whether the same is true for countability.
Example 3.8. Let T = {A v ∃r.A, B v A} and A = {B(a)} and consider the interpretation
I defined by
∆I
AI
BI
rI
=
=
=
=
{a} ∪ {0, 1, 2 . . .} ∪ {. . . , −20 , −10 , 00 , 10 , 20 , . . . }
∆I
{a}
{(a, 0)} ∪ {(n, n + 1) | n ∈ } ∪ {(n0 , n0 + 1) | n ∈ }.
N
Z
Then I is a PEQ-materialization of T and A, but it is not hom-initial (and in fact not even
sim-initial) since the restriction of I to domain {a} ∪ {0, 1, 2 . . .} is also a model of T and
A, but there is no homomorphism (and no simulation) from I to this restriction preserving
{a}.
As an application of Theorem 3.7, we now show that materializability coincides for the
query languages PEQ, CQ, and ELIQ (and that for ALCF-TBoxes, all these also coincide
with ELQ-materializability).
Theorem 3.9. Let T be an ALCF I-TBox. Then the following conditions are equivalent:
(1) T is PEQ-materializable;
(2) T is CQ-materializable;
(3) T is ELIQ-materializable;
(4) Mod(T , A) contains an i-sim-initial I, for every ABox A that is consistent w.r.t. T ;
(5) Mod(T , A) contains a hom-initial I, for every ABox A that is consistent w.r.t. T .
If T is an ALCF-TBox, then the above is the case iff T is ELQ-materializable iff Mod(T , A)
contains a sim-initial I, for every ABox A that is consistent w.r.t. T .
Proof. The implications (1) ⇒ (2) and (2) ⇒ (3) are trivial. For (3) ⇒ (4), let I be an ELIQmaterialization of T and an ABox A. By Lemma 3.4, we may assume that I is countable
and generated. By Lemma 3.7, Mod(T , A) contains an i-sim-initial interpretation. For (4)
⇒ (5), assume that I ∈ Mod(T , A) is i-sim-initial. By Points (3) and (5) of Lemma 3.6,
the i-unfolding of I is hom-initial in Mod(T , A) and (5) follows. (5) ⇒ (1) follows from
Theorem 3.7. The implications for ALCF-TBoxes are proved similarly.
Because of Theorem 3.9, we sometimes speak of materializability without reference to a
query language and of materializations instead of PEQ-materializations.
16
CARSTEN LUTZ AND FRANK WOLTER
3.2. Materializability and coNP-hardness. We show that non-materializability of a
TBox T implies coNP-hardness of ELIQ-evaluation w.r.t. T . To this end, we first establish
that materializability is equivalent to the disjunction property, which is sometimes also
called convexity and plays a central role in attaining PTime complexity for subsumption
in DLs [BBL05], and for attaining PTime data complexity for query answering with DL
TBoxes [KL07].
Let T be a TBox. For an ABox A, individual names a0 , . . . , ak ∈ Ind(A), and ELIQs
C0 (x), . . . , Ck (x), we write T , A |= C0 (a0 ) ∨ · · · ∨ Ck (ak ) if for every model I of T and A,
I |= Ci (ai ) holds for some i ≤ k. We say that T has the ABox disjunction property for
ELIQ (resp. ELQ) if for all ABoxes A, individual names a0 , . . . , ak ∈ Ind(A), and ELIQs
(resp. ELQs) C0 (x), . . . , Ck (x), T , A |= C0 (a0 ) ∨ · · · ∨ Ck (ak ) implies T , A |= Ci (ai ) for some
i ≤ k.
Theorem 3.10. An ALCF I- (ALCF-)TBox T is materializable iff it has the ABox disjunction property for ELIQs (ELQs).
Proof. For the nontrivial “if” direction, let A be an ABox that is consistent w.r.t. T and
such that there is no ELIQ-materialization of T and A. Then T ∪ A ∪ Γ is not satisfiable,
where
Γ = {¬C(a) | T , A 6|= C(a), a ∈ Ind(A), C(x) ELIQ}.
In fact, any satisfying interpretation would be an ELIQ-materialization. By
W compactness,
there is a finite subset Γ0 of Γ such that T ∪A∪Γ0 is not satisfiable, i.e. T , A |= ¬C(a)∈Γ0 C(a).
Since Γ0 ⊆ Γ, we have T , A 6|= C(a) for all ¬C(a) ∈ Γ0 . Thus, T lacks the ABox disjunction
property.
Based on Theorems 3.9 and 3.10, we now establish that materializability is a necessary
condition for query evaluation to be it PTime.
Theorem 3.11. If an ALCF I-TBox T (ALCF-TBox T ) is not materializable, then ELIQevaluation (ELQ-evaluation) w.r.t. T is coNP-hard.
Proof. The proof is by reduction of 2+2-SAT, a variant of propositional satisfiability that was
first introduced by Schaerf as a tool for establishing lower bounds for the data complexity of
query answering in a DL context [Sch93]. A 2+2 clause is of the form (p1 ∨ p2 ∨ ¬n1 ∨ ¬n2 ),
where each of p1 , p2 , n1 , n2 is a propositional letter or a truth constant 0, 1. A 2+2 formula
is a finite conjunction of 2+2 clauses. Now, 2+2-SAT is the problem of deciding whether a
given 2+2 formula is satisfiable. It is shown in [Sch93] that 2+2-SAT is NP-complete.
We first show that if an ALCF I-TBox T is not materializable, then UELIQ-evaluation
w.r.t. T is coNP-hard, where a UELIQ is a disjunction C0 (x) ∨ · · · ∨ Ck (x), with each Ci (x)
an ELIQ. We then sketch the modifications necessary to lift the result to ELIQ-evaluation
w.r.t. T .
Since T is not materializable, by Theorem 3.10 it does not have the ABox disjunction
property. Thus, there is an ABox A∨ , individual names a0 , . . . , ak ∈ Ind(A), and ELIQs
C0 (x), . . . , Ck (x), k ≥ 1, such that T , A∨ |= C0 (a0 ) ∨ · · · ∨ Ck (ak ), but T , A∨ 6|= Ci (ai )
for all i ≤ k. Assume w.l.o.g. that this sequence is minimal, i.e., T , A∨ 6|= C0 (a0 ) ∨ · · · ∨
Ci−1 (ai−1 ) ∨ Ci+1 (ai+1 ) ∨ · · · ∨ Ck (ak ) for all i ≤ k. This clearly implies that for all i ≤ k,
(∗) there is a model Ii of T and A∨ with I |= Ci (ai ) and I 6|= Cj (aj ) for all j 6= i.
We will use A∨ , the individual names a1 , . . . , ak , and the ELIQs C0 (x), . . . , Ck (x) to generate
truth values for variables in the input 2+2 formula.
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
17
Let ϕ = c0 ∧ · · · ∧ cn be a 2+2 formula in propositional letters z0 , . . . , zm , and let
ci = pi,1 ∨pi,2 ∨¬ni,1 ∨¬ni,2 for all i ≤ n. Our aim is to define an ABox Aϕ with a distinguished
individual name f and a UELIQ q(x) such that ϕ is unsatisfiable iff T , Aϕ |= q(f ). To start,
we represent the formula ϕ in the ABox Aϕ as follows:
• the individual name f represents the formula ϕ;
• the individual names c0 , . . . , cn represent the clauses of ϕ;
• the assertions c(f, c0 ), . . . , c(f, cn ), associate f with its clauses, where c is a role name
that does not occur in T ;
• the individual names z0 , . . . , zm represent variables, and the individual names 0, 1 represent
truth constants;
• the assertions
[
{p1 (ci , pi,1 ), p2 (ci , pi,2 ), n1 (ci , ni,1 ), n2 (ci , ni,2 )}
i≤n
associate each clause with the variables/truth constants that occur in it, where p1 , p2 , n1 , n2
are role names that do not occur in T .
We further extend Aϕ to enforce a truth value for each of the variables zi . To this end,
add to Aϕ copies A0 , . . . , Am of A∨ obtained by renaming individual names such that
Ind(Ai ) ∩ Ind(Aj ) = ∅ whenever i 6= j. As a notational convention, let aij be the name used
for the individual name aj ∈ Ind(A∨ ) in Ai for all i ≤ m and j ≤ k. Intuitively, the copy Ai
of A is used to generate a truth value for the variable zi . To actually connect each individual
name zi to the associated ABox Ai , we use role names r0 , . . . , rk that do not occur in T .
More specifically, we extend Aϕ as follows:
• link variables zi to the ABoxes Ai by adding assertions rj (zi , aij ) for all i ≤ m and j ≤ k;
thus, truth of zi means that the concept ∃r0 .C0 is true at zi and falsity means that one of
the concepts ∃rj .Cj , j ≤ k, is true at zi ;
• to ensure that 0 and 1 have the expected truth values, add a copy of C0 (x) viewed as an
ABox with root 00 and a copy of C1 (x) viewed as an ABox with root 10 ; add r0 (0, 00 ) and
r1 (1, 10 ).
Consider the query q0 (x) = C(x) with
C = ∃c.(∃p1 .ff u ∃p2 .ff u ∃n1 .tt u ∃n2 .tt)
which describes the existence of a clause with only false literals and thus captures falsity
of ϕ, where tt is an abbreviation for ∃r0 .C0 and ff an abbreviation for the ELU-concept
∃r1 .C1 t · · · t ∃rk .Ck . It is straightforward to show that ϕ is unsatisfiable iff T , A |= q0 (f ).
To obtain the desired UELIQ q(x), it remains to take q0 (x) and distribute disjunction to
the outside.
We now show how to improve the result from UELIQ-evaluation to ELIQ-evaluation. Our
aim is to change the encoding of falsity of a variable zi from satisfaction of ∃r1 .C1 t· · ·t∃rk .Ck
at zi to satisfaction of ∃h.(∃r1 .C1 u · · · u ∃rk .Ck ), at zi , where h is an additional role
that does not occur in T . We can then replace the concept ff in the query q0 with
∃h.(∃r1 .C1 u · · · u ∃rk .Ck ), which gives the desired ELIQ q(x).
It remains to modify Aϕ to support the new encoding of falsity. The basic idea is that
each zi has k successors bi1 , . . . , bik reachable via h such that for 1 ≤ j ≤ k,
• ∃r` .C` is satisfied at bij for all ` = 1, . . . , j − 1, j + 1, . . . , k and
• the assertion rj (bij , aij ) is in Aϕ .
18
CARSTEN LUTZ AND FRANK WOLTER
Thus, ∃r1 .C1 u · · · u ∃rk .Ck is satisfied at bij iff Cj is satisfied at aij , for all j with 1 ≤ j ≤ k.
In detail, the modification of Aϕ is as follows:
• for 1 ≤ j ≤ k, add to Aϕ a copy of Cj viewed as an ABox, where the root individual name
is dj ;
• for all i ≤ m, replace the assertions rj (zi , aij ), 1 ≤ j ≤ k, with the following:
– h(zi , bi1 ), . . . , h(zi , bik ) for all i ≤ m;
– rj (bij , aij ), r1 (bij , d1 ), . . . , rj−1 (bij , dj−1 ),
rj+1 (bij , dj+1 ), . . . , rk (bij , dk ) for all i ≤ m and 1 ≤ j ≤ k.
This finishes the modified construction. Again, it is not hard to prove correctness.
It remains to note that, when T is an ALCF-TBox, then the above construction of q
yields an ELQ instead of an ELIQ.
The converse of Theorem 3.11 fails, i.e., there are TBoxes that are materializable, but
for which ELIQ-evaluation is coNP-hard. In fact, materializations of such a TBox T and
ABox A are guaranteed to exist, but cannot always be computed in PTime (unless PTime
= coNP). Technically, this follows from Theorem 6.5 later on which states that for every
non-uniform CSP, there is a materializable ALC-TBox for which Boolean CQ-answering has
the same complexity, up to complementation of the complexity class.
3.3. Complexity of TBoxes for Different Query Languages. We make use of our
results on materializability to show that PTime query evaluation w.r.t. ALCF I-TBoxes
does not depend on whether we consider PEQs, CQs, or ELIQs, and the same is true
for coNP-hardness, for Datalog6= -rewritability, and for monadic Datalog6= -rewritability.
For ALCF-TBoxes, we can add ELQs to the list. Theorem 3.12 below gives a uniform
explanation for the fact that, in the traditional approach to data complexity in OBDA, the
complexity of evaluating PEQs, CQs, and ELIQs has turned out to be identical for almost
all DLs. It allows us to (sometimes) speak of the ‘complexity of query evaluation’ without
reference to a concrete query language.
Theorem 3.12. For all ALCF I-TBoxes T ,
(1) PEQ-evaluation w.r.t. T is in PTime iff CQ-evaluation w.r.t. T is in PTime iff
ELIQ-evaluation w.r.t. T is in PTime;
(2) T is (monadic) Datalog6= -rewritable for PEQ iff it is Datalog6= -rewritable for CQ iff it
is (monadic) Datalog6= -rewritable for ELIQ (unless PTime = coNP);
(3) PEQ-evaluation w.r.t. T is coNP-hard iff CQ-evaluation w.r.t. T is coNP-hard iff
ELIQ-evaluation w.r.t. T is coNP-hard.
If T is an ALCF-TBox, then ELIQ can be replaced by ELQ in (1), (2), and (3). If T is an
ALCI-TBox, then Datalog6= -rewritability can be replaced by Datalog-rewritability in (2).
Proof. We start with Points (1) and (2), for which the “only if” directions are trivial. For the
converse directions, we may assume by Theorem 3.11 that the TBox T is materializable. The
implications from CQ to PEQ in Points (1) and (2) follow immediately from this assumption:
W
one can first transform a given PEQ q(~x) into an equivalent disjunction of CQs i∈I qi (~x).
CQ-materializability of T implies that, for any ABox A and ~a in Ind(A), T , A |= q(~a) iff
there exists i ∈ I such that T , A |= qi (~a). Thus if CQ-evaluation w.r.t. T is in PTime,
evaluation of (T , q(~x)) is in PTime. The same holds for (monadic) Datalog6= -rewritability
because the class of Datalog6= -queries is closed under finite union.
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
19
We now consider the implications from ELIQ to CQ (and from ELQ to CQ if T is a
ALCF-TBox) in Points (1) and (2). The following claim is the main step of the proof. It
states that for any CQ q(~x), we can reduce the evaluation of q(~x) w.r.t. T on an ABox A to
evaluating quantifier-free CQs and ELIQs C(x) w.r.t. T (ELQs if T is an ALCF-TBox),
both on A.
Claim 1. For any materializable TBox T and CQ q(~x) with ~x = x1 · · · xn , one can construct
a finite set Q of pairs (ϕ(~x, ~y ), C), where
• ϕ(~x, ~y ) is a (possibly empty) conjunction of atoms of the form x = y or r(x, y), where r is
a role name in q(~x) and
• C is a finite set of ELIQs
such that the following conditions are equivalent for any ABox A and ~a = a1 · · · an from
Ind(A):
(i) T , A |= q(~a);
(ii) there exists (ϕ(~x, ~y ), C) ∈ Q and an assignment π in Ind(A) with π(xi ) = ai for 1 ≤ i ≤ n,
A |=π ϕ(~x, ~y ), and T , A |= C(π(x)) for all C(x) ∈ C.
If T is an ALCF-TBox, then one can choose ELQs instead of ELIQs in each C in Q.
Before we prove Claim 1, we show how the desired results follow from it. Let a CQ q(~x) be
given and let Q be the set of pairs from Claim 1.
• Assume that ELIQ-evaluation w.r.t. T is in PTime. Then T , A |= q(~a) can be decided
in polynomial time since there are only polynomially many assignments π and for any
such π, A |=π ϕ(~x, ~y ) can be checked in polynomial time (using a naive algorithm) and
T , A |= C(π(x)) can be checked in polynomial time for each ELIQ C(x) ∈ C.
• Assume that T is (monadic) Datalog6= -rewritable for ELIQ. Let p = (ϕ(~x, ~y ), C) ∈ Q. For
each C(x) ∈ C, choose a (monadic) Datalog6= -rewriting ΠC (x) of (T , C(x)), assume w.l.o.g
that none of the chosen programs share any IDB relations, and that the goal relation of
ΠC (x) is goalC . Let Πp be the (monadic) Datalog6= program that consists of the rules of
all the chosen programs, plus the following rule:
^
goal(~x) ← ϕ(~x, ~y ) ∧
goalC (x).
C(x)∈C
Datalog6= -rewriting
The desired (monadic)
of (T , q(~x)) is obtained by taking the union of
all the constructed (monadic) Datalog6= queries.
The implications from ELQs to CQs for ALCF-TBoxes in Points (1) and (2) follow in the
same way since, then, each C in Q consists of ELQs only.
For the proof of Claim 1, we first require a technical observation that allows us to deal
with subqueries that are not connected to an answer variable in the CQ q(~x). To illustrate,
consider the query q0 = ∃x B(x). To prove Claim 1 for q0 , we have to find a set Q of pairs
(ϕ(~y ), C) satisfying Conditions (i) and (ii). Clearly, in this case the components ϕ(~y ) will
be empty and so we have to construct a finite set C of ELIQs such that for any ABox A,
T , A |= ∃x B(x) iff there exists an ELIQ C(x) ∈ C and an assignment π in Ind(A) such that
T , A |= C(π(x)). An infinite set C with this property is given by the set of all ELIQs ∃~r.B(x),
where ~r is a sequence r1 · · · rn of roles ri in T and ∃~r.B stands for ∃r1 · · · ∃rn .B—this follows
immediately from the assumption that T is materializable and that, by Lemma 3.4, for any
ABox A that is consistent w.r.t. T , there exists a generated CQ-initial model of T and ABox
20
CARSTEN LUTZ AND FRANK WOLTER
A. The following result states that it is sufficient to include in C the set of all ∃~r.B(x) with
~r of length bounded by n0 := 2(2(|T |+|C|) · 2|T | + 1.
Claim 2. Let C be an ELI-concept and assume that T , A |= ∃x C(x). If T is materializable,
then there exists a sequence of roles ~r = r1 · · · rn with ri in T and of length n ≤ n0 and an
a ∈ Ind(A) such that T , A |= ∃~r.C(a). If C is an EL-concept and T an ALCF-TBox, then
the sequence ~r consists of role names in T .
Proof of Claim 2. Let I be a CQ-materialization of T and A. By Points (3) and (5) of
Lemma 3.6, we may assume that I is i-unfolded. From T , A |= ∃x C(x), we obtain C I 6= ∅.
Let n be minimal such that there are a ∈ Ind(A) and d ∈ C I with n = distI (d, a) where
distI (d, a) denotes the length of the shortest path from d to a in the undirected graph GI .
If n ≤ n0 , we are done. Otherwise fix an a ∈ Ind(A) and denote by M the set all e ∈ C I
I
with n = distI (e, a). Let d0 , . . . , dn with a = d0 , dn = d, and (di , di+1 ) ∈ ri+1
for i < n
be the shortest path from a to d. Observe that T , A |= ∃~r.C(a) for ~r = r0 · · · rn−1 since I
is a materialization of T and A. We now employ a pumping argument to show that this
leads to a contradiction. Let cl(T , C) denote the closure under single negation of the set of
subconcepts of concepts in T and C. Set
tIT ,C (e) = {D ∈ cl(T , C) | e ∈ DI }.
As n > n0 , then there exist di and di+j with j > 1 and i + j < n such that
tIT ,C (di ) = tIT ,C (di+j ),
tIT ,C (di+1 ) = tIT ,C (di+j+1 ),
ri+1 = ri+j+1
Now replace in I the interpretation induced by the subtree rooted at di+j+1 by the interpretation induced by the subtree rooted at di+1 . We do the same construction for all elements
of M and denote the resulting interpretation by J . One can easily show by induction
that J is still a model of T and A, but now J 6|= ∃~r.C(a) and so T , A 6|= ∃~r.C(a). This
contradiction finishes the proof of Claim 2. For EL-concepts and ALCF-TBoxes, Claim 2
can be proved similarly by using an unfolded (rather than i-unfolded) materialization, which
exists by Points (4) and (6) of Lemma 3.6.
Proof of Claim 1. Assume that T and q(~x) = ∃~x ψ(~x, ~y ) are given. We have to construct
a set Q such that Conditions (i) and (ii) are satisfied. Let A be an ABox with T , A |= q(~a),
I a materialization of T and A that is i-unfolded, and π an assignment in I such that
I |=π ψ(~x, ~y ). We define a corresponding pair p = (ϕ(~x, ~z), C) to be included in Q (and
these are the only pairs in Q).
For identifying ϕ(~x, ~z), set x ∼ y if π(x) = π(y) and denote by [x] the equivalence class
of x w.r.t. “∼”. Let ϕ0 be the set of all atoms r([x], [y]) such that π(x), π(y) ∈ Ind(A) and
there are x0 ∈ [x] and y 0 ∈ [y] with r(x0 , y 0 ) in ψ. We obtain ϕ(~x, ~y ) by selecting an answer
variable y ∈ [x] for every [x] that contains such a variable and then replacing [x] by y in ϕ0 ,
adding xi = xj to ϕ(~x, ~z) for any two (selected) answer variables xi , xj with xi ∼ xj , and
by regarding the remaining equivalences classes [y] that do not contain answer variables as
variables in ~z.
We now identify C. Assume w.l.o.g. that I uses the naming scheme of i-unravelings. Let
a ∈ Ind(A). By Ia , we denote the subinterpretation of I induced by the set of all elements
ar1 d1 · · · rn dn ∈ ∆I . Let M be a maximal connected component of ∆Ia ∩ {π(y) | y ∈ var(ψ)}.
We associate with M an ELIQ to be included in C (and these are the only ELIQs in C).
The conjunctive query ϕM consists of all atoms r([x], [y]) such that π(x), π(y) ∈ M
and there are x0 ∈ [x], y 0 ∈ [y] with r(x0 , y 0 ) in ψ and all atoms A([x]) such that π(x) ∈ M
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
21
and there is x0 ∈ [x] with A(x0 ) in ψ. We again assume that equivalence classes [x] that
contain an answer variable (there is at most one such class in ϕM ) are replaced with an
answer variable from [x] and regard the remaining equivalences classes as variables. Note
that ϕM is tree-shaped since Ia is. We can thus pick a variable x0 with π(x0 ) ∈ M such that
distI (a, π(x0 )) is minimal. Let x be [x0 ] if [x0 ] contains no answer variable and, otherise, let
x be the answer variable that [x0 ] has been replaced with. Let [y1 ], . . . , [ym ] be the variables
in ϕM that are distinct from x and consider the ELIQ ∃[y1 ] · · · ∃[ym ]ϕM (x, [y1 ], . . . , [ym ]),
which we write as CM (x) where CM is an appropriate ELI-concept. We now distinguish
the following cases:
• π(x) = a. In this case, we include CM (x) in C;
• otherwise, we still know that T , A |= ∃x C(x). Thus, by Claim 2 there is a sequence
of roles ~r = r1 · · · rn with ri in T and n ≤ n0 such that T , A |= ∃~r.CM (a) for some
a ∈ Ind(A). In this case, we include ∃~r.CM (y) in C for some fresh variable y.
This finishes the construction of C and thus of Q. Clearly, Q is finite. The stated properties
of Q follow directly from its construction. For ALCF-TBoxes and ELQs, Claim 1 can be
proved similarly using an unfolded materialization (instead of an i-unfolded one) and the
observation that in this case all CM ([x]) and ∃~r.CM (y) are ELQs (~r uses role names only by
Claim 2).
We now turn our attention to Point (3). Here, the “if” directions are trivial and we
prove the “only if” part. It suffices to show that if PEQ-evaluation w.r.t. a TBox T is
coNP-hard, then so is ELIQ-evaluation. We start with showing the slightly simpler result
that coNP-hardness of CQ-evaluation w.r.t. T implies coNP-hardness of UELIQ-evaluation,
and then sketch the modifications required to strengthen the proof to attain the original
statement.
Thus assume that evaluating the CQ q(~x) w.r.t. T is coNP-hard. We shall exhibit
an UELIQ q 0 (x) such that for every ABox A and all ~a ∈ Ind(A), one can produce in
polynomial time an ABox A0 with a distinguished individual name a0 such that T , A |= q(~a)
iff T , A0 |= q 0 (a0 ). Instead of constructing q 0 (~x) right away, we will start with describing the
translation of A to A0 . Afterwards, it will be clear how to construct q 0 (~x).
Thus, let A be an ABox and ~a from Ind(A). The construction of A0 builds on Claim 1
above. Let Q be the set of pairs from that claim and reserve a fresh individual name a0 .
To obtain the desired ABox A0 , we extend A for every pair p = (ϕ(~x, ~y ), C) in Q. Let
C = Cp,1 (x1 ), . . . , Cp,kp (xkp ). Then
• introduce a fresh individual name ap and fresh role names rp , r, rp,1 , . . . , rp,kp ;
• add the assertion rp (a0 , ap );
• for every assignment π in Ind(A) with A |=π ϕ(~x, ~y ) and π(~x) = ~a, introduce
– a fresh individual name ap,π and the assertion r(ap , ap,π );
– the assertion rp,i (ap , π(xi )) for 1 ≤ i ≤ kp .
From Claim 1, it is immediate that T , A |= q(~a) iff T , A0 |= q 0 (x) where q 0 (x) is the UELIQ
t q0 (x) with Cp = ∃rp .∃r. u ∃rp,i .Cp,i . Thus, evaluating q0 (x) w.r.t. T is coNP-hard,
p∈Q
1≤i≤kp
as required. It remains to modify the reduction by replacing CQs with PEQs and UELIQs
with ELIQs. The former is straightforward: every PEQ is equivalent to a finite disjunction of
CQs, and thus we can construct A0 and q 0 (x) in essentially the same way as before; instead of
considering all pairs from Q for a single CQ, we now use the union of all sets Q for the finitely
many CQs in question. Finally, we can replace UELIQs with ELIQs by using the same
22
CARSTEN LUTZ AND FRANK WOLTER
construction as in the proof of Theorem 3.11: after adding some straightforward auxiliary
structure to A0 , one can replace a disjunction of ELIQs by (essentially) their conjunction,
which is again an ELIQ. Details are left to the reader.
We remark that Theorem 3.12 can be extended to also cover rewritability into first-order
(FO) queries, and that the proof is almost identical to that for Datalog6= -rewritability.
4. Unraveling Tolerance
We develop a condition on TBoxes, called unraveling tolerance, that is sufficient for the TBox
to be monadic Datalog6= -rewritable for PEQ, and thus also sufficient for PEQ-evaluation
w.r.t. the TBox being in PTime. Unraveling tolerance strictly generalizes syntactic ‘Horn
conditions’ such as the ones used to define the DL Horn-SHIQ, which was designed as a
(syntactically) maximal DL with PTime query evaluation [HMS07, EGOS08].
Unraveling tolerance is based on an unraveling operation on ABoxes, in the same spirit
as the unfolding of an interpretation into a tree interpretation we discussed above. Formally,
the unraveling Au of an ABox A is the following (possibly infinite) ABox:
• Ind(Au ) is the set of sequences b0 r0 b1 · · · rn−1 bn , n ≥ 0, with b0 , . . . , bn ∈ Ind(A) and
−
r0 , . . . , rn−1 ∈ NR ∪ N−
R such that for all i < n, we have ri (bi , bi+1 ) ∈ A and (bi−1 , ri−1 ) 6=
(bi+1 , ri ) when i > 0;
• for each C(b) ∈ A and α = b0 r0 · · · rn−1 bn ∈ Ind(Au ) with bn = b, C(α) ∈ Au ;
• for each α = b0 r0 · · · rn−1 bn ∈ Ind(Au ) with n > 0, rn−1 (b0 r0 · · · rn−2 bn−1 , α) ∈ Au .
For all α = b0 · · · bn ∈ Ind(Au ), we write tail(α) to denote bn . Note that the condition
−
(bi−1 , ri−1
) 6= (bi+1 , ri ) is needed to ensure that functional roles can still be interpreted in a
functional way after unraveling.
Definition 4.1 (Unraveling Tolerance). A TBox T is unraveling tolerant if for all ABoxes
A and ELIQs q, we have that T , A |= q implies (T , Au ) |= q.
It is not hard to prove that the converse direction ‘T , Au |= q implies T , A |= q’ is true
for all ALCF I-TBoxes. Note that it is pointless to define unraveling tolerance for queries
that are not necessarily tree shaped, such as CQs.
Example 4.2.
(1) The ALC-TBox T1 = {A v ∀r.B} is unraveling tolerant. This can be proved by showing
+
that (i) for any (finite or infinite) ABox A, the interpretation IA
that is obtained from A by
+
I
−
extending B A with all a ∈ Ind(A) that satisfy ∃r .A in A (when viewed as an interpretation)
+
+
is an ELIQ-materialization of T1 and A; and (ii) IA
|= C(a) iff IA
u |= C(a) for all ELIQs
C(x) and a ∈ Ind(A). The proof of (ii) is based on a straightforward induction on the
structure of the ELI-concept C. As illustrated by the ABox A = {r(a, b), A(a)} and the fact
that Au , T |= B(b), the use of inverse roles in the definition of Au is crucial here despite the
fact that T1 does not use inverse roles.
(2) A simple example for an ALC-TBox that is not unraveling tolerant is
T2 = {A u ∃r.A v B, ¬A u ∃r.¬A v B}.
For A = {r(a, a)}, it is easy to see that we have T2 , A |= B(a) (use a case distinction on the
truth value of A at a), but T2 , Au 6|= B(a).
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
23
Before we show that unraveling tolerance indeed implies PTime query evaluation,
we first demonstrate the generality of this property by relating it to Horn-ALCF I, the
ALCFI-fragment of Horn-SHIQ. Different versions of Horn-SHIQ have been proposed in
the literature, giving rise to different versions of Horn-ALCF I [HMS07, KRH07, EGOS08,
Kaz09]. As the original definition from [HMS07] based on polarity is rather technical, we
prefer to work with the following equivalent and less cumbersome definition. A HornALCFI-TBox T is a finite set of concept inclusions L v R and functionality assertions
where L and R are built according to the following syntax rules:
R, R0 ::= > | ⊥ | A | ¬A | R u R0 | L → R | ∃r.R | ∀r.R
L, L0 ::= > | ⊥ | A | L u L0 | L t L0 | ∃r.L
with r ranging over NR ∪ N−
R and L → R abbreviating ¬L t R. Whenever convenient, we
may assume w.l.o.g. that T contains only a single concept inclusion > v CT where CT is
built according to the topmost rule above.
By applying some simple transformations, it is not hard to show that every Horn-ALCF ITBox according to the original polarity-based definition is equivalent to a Horn-ALCF I-TBox
of the form introduced here. Although not important in our context, we note that even a
polynomial time transformation is possible.
Theorem 4.3. Every Horn-ALCF I-TBox is unraveling tolerant.
Proof. As a preliminary, we give a characterization of the entailment of ELIQs in the presence
of Horn-ALCF I-TBoxes which is in the spirit of the chase procedure as used in database
theory [FKMP05, CGK13] and of consequence-driven algorithms as used for reasoning in
Horn description logics such as EL++ and Horn-SHIQ [BBL05, Kaz09, Krö10b].
We use extended ABoxes, i.e., finite sets of assertions C(a) and r(a, b) with C a potentially
compound concept. An ELIU ⊥ -concept is a concept that is formed according to the second
syntax rule in the definition of Horn-ALCF I. For an extended ABox A0 and an assertion
C(a), C an ELIU ⊥ -concept, we write A0 ` C(a) if C(a) has a syntactic match in A0 ,
formally:
• A0 ` >(a) is unconditionally true;
• A0 ` ⊥(a) if ⊥(b) ∈ A0 for some b ∈ Ind(A);
• A0 ` A(a) if A(a) ∈ A0 ;
• A0 ` C u D(a) if A0 ` C(a) and A0 ` D(a);
• A0 ` C t D(a) if A0 ` C(a) or A0 ` D(a);
• A0 ` ∃r.C(a) if there is an r(a, b) ∈ A0 such that A0 ` C(b).
Now for the announced characterization. Let T = {> v CT } be a Horn-ALCF I-TBox and
A a potentially infinite ABox (so that the characterization also applies to unravelings of
ABoxes). We produce a sequence of extended ABoxes A0 , A1 , . . . , starting with A0 = A. In
what follows, we use additional individual names of the form ar1 C1 · · · rk Ck with a ∈ Ind(A0 ),
r1 , . . . , rk roles that occur in T , and C1 , . . . , Ck subconcepts of concepts in T . We assume
that NI contains such names as needed and use the symbols a, b, . . . also to refer to individual
names of this compound form. Each extended ABox Ai+1 is obtained from Ai by applying
the following rules in a fair way:
R1 if a ∈ Ind(Ai ), then add CT (a).
R2 if C u D(a) ∈ Ai , then add C(a) and D(a);
R3 if C → D(a) ∈ Ai and Ai ` C(a), then add D(a);
R4 if ∃r.C(a) ∈ Ai and func(r) ∈
/ T , then add r(a, arC) and C(arC);
24
CARSTEN LUTZ AND FRANK WOLTER
R5 if ∃r.C(a) ∈ Ai , func(r) ∈ T , and r(a, b) ∈ Ai , then add C(b);
R6 if ∃r.C(a) ∈ Ai , func(r) ∈ T , and there is no r(a, b) ∈ Ai , then add r(a, arC) and
C(arC);
R7 if ∀r.C(a) ∈ Ai and r(a, b) ∈ Ai , then add C(b).
S
Let Ac = i≥0 Ai be the completion of the original ABox A.2 Note that Ac may be infinite
even if A is finite, and that none of the above rules is applicable in Ac . We write ‘Ac ` ⊥’
instead of ‘Ac ` ⊥(a) for some a ∈ NC ’. If A 6` ⊥, then Ac corresponds to an interpretation
Ic in the standard way, i.e.,
∆Ic
AIc
rIc
= Ind(Ac )
= {a | A(a) ∈ Ac }
for all A ∈ NC
= {r(a, b) | r(a, b) ∈ Ac } for all r ∈ NR
where in Ic we assume that only the individual names in Ind(A) are elements of NI .
Claim 1. If Ac 6` ⊥, then Ic is a PEQ-materialization of T and A.
To prove Claim 1 it suffices to show that there is a homomorphism h preserving Ind(A)
from Ic into every model J of T and A and that Ic is a model of T and A. The latter is
immediate by construction of Ac . Regarding the former, the desired homomorphism h can
can be constructed inductively starting with the ABox A0 and then extending to A1 , A2 , . . . .
Using Claim 1 and the easily proved fact that Ac 6` ⊥ iff A is consistent w.r.t. T one can
now show the following.
Claim 2. T , A |= C(a) iff Ac ` C(a) or Ac ` ⊥, for all ELIQs C(x) and a ∈ Ind(A).
We now turn to the actual proof of Theorem 4.3. Consider the application of the above
completion construction to both the original ABox A and its unraveling Au . Recall that
individual names in Au are of the form a0 r0 a1 · · · rn−1 an . Consequently, individual names
in Auc take the form a0 r0 a1 · · · rn−1 an s1 C1 · · · sk Ck . For a ∈ Ind(Ac ) and α ∈ Ind(Auc ), we
write a ∼ α if a and α are of the form an s1 C1 · · · sk Ck and a0 r0 a1 · · · rn−1 an s1 C1 · · · sk Ck ,
respectively, with k ≥ 0. Note that, in particular, a ∼ a for all a ∈ Ind(A). The following
claim can be shown by induction on i.
Claim 3. For all a ∈ Ind(Ai ) and α ∈ Ind(Aui ) with a ∼ α, we have
(1) Ai ` C(a) iff Aui ` C(α) for all ELI-concepts C;
(2) C(a) ∈ Ai iff C(α) ∈ Aui for all subconcepts C of concepts in T .
Now, unraveling tolerance of T follows from Claims 2 and 3.
Theorem 4.3 shows that unraveling tolerance and Horn logic are closely related. Yet,
the next example demonstrates that there are unraveling tolerant ALCF I-TBoxes that are
not equivalent to any Horn sentence of FO. Since any Horn-ALCF I-TBox is equivalent
to such a sentence, it follows that unraveling tolerant ALCF I-TBoxes strictly generalize
Horn-ALCF I-TBoxes. This increased generality will pay off in Section 5 when we establish
a dichotomy result for TBoxes of depth one.
Example 4.4. Take the ALC-TBox
T = {∃r.(A u ¬B1 u ¬B2 ) v ∃r.(¬A u ¬B1 u ¬B2 )}.
2Order of rule application has an impact on the shape of A , but is irrelevant for the remainder of the
c
proof.
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
25
One can show as in Example 4.2 (1) that T is unraveling tolerant; here, the materialization
is actually A itself rather than some extension thereof, i.e., as far as ELIQ (and even PEQ)
evaluation is concerned, T cannot be distinguished from the empty TBox.
It is well-known that FO Horn sentences are preserved under direct products of interpretations [CK90]. To show that T is not equivalent to any such sentence, it thus suffices to show
that T is not preserved under direct products. This is simple: let I1 and I2 consist of a single
r-edge between elements d and e, and let e ∈ (A u B1 u ¬B2 )I1 and e ∈ (A u ¬B1 u B2 )I2 ;
then the direct product I of I1 and I2 still has the r-edge between (d, d) and (e, e) and
satisfies (e, e) ∈ (A u ¬B1 u ¬B2 )I , thus is not a model of T .
We next show that unraveling tolerance is indeed a sufficient condition for monadic
Datalog6= -rewritability (and thus for PTime query evaluation). In Section 6, we will establish
a connection between query evaluation under DL TBoxes and constraint satisfaction problems
(CSPs). The monadic Datalog6= programs that we construct resemble the canonical monadic
Datalog programs for CSPs [FV98].
Let T be an unraveling tolerant ALCF I-TBox and q = C0 (x) an ELIQ. We show how
to construct a Datalog6= -rewriting of the OMQ (T , q(x)). Using the construction from the
proof of Theorem 3.12, one can extend this construction from ELIQs to PEQs. Recall from
the proof of Theorem 3.12 that cl(T , C0 ) denotes the closure under single negation of the set
of subconcepts of T and C0 . For an interpretation I and d ∈ ∆I , we use tIT ,q (d) to denote
the set of concepts C ∈ cl(T , C0 ) such that d ∈ C I . A T , q-type is a subset t ⊆ cl(T , C0 )
such that for some model I of T , we have t = tIT ,q (d). We use tp(T , q) to denote the set of
all T ,q-types. Observe that one can construct the set tp(T , q) in exponential time as the set
of all t ⊆ cl(T , C0 ) such that for any concept ¬C ∈ cl(T , C0 ) either C ∈ t or ¬C ∈ t and the
concept
C is satisfiable in a model of T .
u
C∈t
t, t0
For
∈ tp(T , q) and r a role, we write t r t0 if there are a model I of T and
I
∈ ∆ such that tIT ,q (d) = t, tIT ,q (d0 ) = t0 , and (d, d0 ) ∈ rI . One can construct the set
of all (t, t0 , r) such that t r t0 in exponential time by checking for each candidate tuple
(t, t0 , r) whether the concept
C u ∃r.
C
0
d, d0
u
C∈t
u
C∈t
is satisfiable in a model of T .
Introduce, for every set T ⊆ tp(T , C0 ) a unary IDB relation PT . Let Π be the monadic
Datalog6= program that contains the following rules:
(1) PT (x) ← A(x) for all concept names A ∈ cl(T , C0 ) and T = {t ∈ tp(T , q) | A ∈ t};
(2) PT (x) ← PT0 (x) ∧ r(x, y) ∧ PT1 (y) for all T0 , T1 ⊆ tp(T , q) and all role names r that
occur in cl(T , C0 ) and their inverses, where T = {t ∈ T0 | ∃t0 ∈ T1 : t r t0 };
(3) PT0 ∩T1 (x) ← PT0 (x) ∧ PT1 (x) for all T0 , T1 ⊆ tp(T , q);
(4) goal(x) ← PT (x) for all T ⊆ tp(T , q) such that t ∈ T implies C0 ∈ T ;
(5) goal(x) ← P∅ (y);
(6) goal(x) ← r(y, z1 ) ∧ r(y, z2 ) ∧ z1 6= z2 for all func(r) ∈ T .
To show that Π is a rewriting of the OMQ (T , C0 (x)), it suffices to establish the following
lemma.
Lemma 4.5. A |= Π(a0 ) iff T , A |= C0 (a0 ), for all ABoxes A and a0 ∈ Ind(A).
Proof. The “if” direction is straightforward: by induction on the number of rule applications,
one can show that whenever Π derives PT (a), then every model of T and A satisfies
26
CARSTEN LUTZ AND FRANK WOLTER
tIT ,q (a) ∈ T . By definition of the goal rules of Π, A |= Π(a0 ) thus implies that every model of
T and A makes C0 (a0 ) true or that A is inconsistent w.r.t. T . Consequently, T , A |= C0 (a0 ).
For the “only if” direction, it suffices to show that A 6|= Π(a0 ) implies T , Au 6|= C0 (a0 )
since T is unraveling tolerant. Because of the rules in Π of the form (3), for every a ∈ Ind(A)
we can find a unique minimal Ta such that PTa (a) is derived by Π. Observe that, A(α) ∈ Au ,
tail(α) = a, and t ∈ Ta implies A ∈ t because of the rules of the form (1) in Π and by
construction of Au .
We first associate with every α ∈ Ind(Au ) a concrete T , q-type tα ∈ Ttail(α) . To start, we
choose ta ∈ Ta arbitrarily for all a ∈ Ind(A). Now assume that tα has already been chosen
and that β = αrb ∈ Ind(Au ). Then r(tail(α), b) ∈ A. Because of the rules in Π of the form
(2) and (5), we can thus choose tβ ∈ Tb such that tα r tβ . In this way, all types tα will
eventually be chosen. We now construct an interpretation I, starting with
∆I
AI
rI
= Ind(Au )
= {α | A ∈ tα } for all concept names A
= {(α, β) | r(α, β) ∈ Au } for all role names r.
Next, consider every α ∈ Ind(Au ) and every ∃r.C ∈ tα such that Au does not contain an
assertion r(α, β) with C ∈ tβ . First assume that func(r) 6∈ T . There must be a T , q-type
t such that tα r t and C ∈ t. Choose a model Jα,∃r.C of T and D = u ta u ∃r. u t, a
d ∈ DJα,∃r.C , and an e ∈ (u t)Jα,∃r.C with (d, e) ∈ rJα,∃r.C . W.l.o.g., we can assume that
−
Jα,∃r.C is tree-shaped with root d. Let Jα,∃r.C
be obtained from Jα,∃r.C by dropping the
−
subtree rooted at e. Now disjointly add Jα,∃r.C to I, additionally including (a, d) in rI .
Now assume that func(r) ∈ T . Then, if there exists r(α, β) ∈ Au , then C ∈ tβ as otherwise
we do not have tα r tβ . Thus, assume there is no r(α, β) ∈ Au . There must be a T , q-type
t such that tα r t and C ∈ t. We then have D ∈ t for all ∃r.D ∈ tα and so construct only
−
−
a single Jα,∃r.C
for the role r and disjointly add Jα,∃r.C
to I, additionally including (a, d)
I
in r . This finishes the construction of I. The following claim can be proved by induction
on C, details are omitted.
Claim. For all C ∈ cl(T , C0 ) :
(a) α ∈ C I iff C ∈ tα for all α ∈ Ind(Au ) and
−
(b) d ∈ C Jα,∃r.D iff d ∈ C I for all Jα,∃r.D and all d ∈ ∆Jα,∃r.D .
By construction of I and since A(α) ∈ Au implies A ∈ tα , I is a model of A. Due to the
rules in Π that are of the form (4), Point (a) of the claim yields I 6|= C0 (a0 ). Finally, we
observe that I is a model of T . The concept inclusions in T are satisfied by the above
claim, since C v D ∈ T means that C ∈ t implies D ∈ t for every T , q-type t, and since
each Jα,∃r.C is a model of T . Due to the rules in Π that are of the form (6) and since each
Jα,∃r.C is a model of T , all functionality assertions in T are satisfied as well. Summing up,
we have shown that T , Au 6|= C0 (a0 ), as required.
Together with Theorem 3.12, we have established the following result.
Theorem 4.6. Every unraveling tolerant ALCF I-TBox is monadic Datalog6= -rewritable for
PEQ.
Together with Theorems 3.12 and 4.3, Theorem 4.6 also reproves the known PTime
upper bound for the data complexity of CQ-evaluation in Horn-ALCF I [EGOS08]. Note
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
27
that it is not clear how to attain a proof of Theorem 4.6 via the CSP connection established
in Section 6 since functional roles break this connection.
By Theorems 3.11 and 4.6, unraveling tolerance implies materializability unless PTime =
NP. Based on the disjunction property, this implication can also be proved without the side
condition.
Theorem 4.7. Every unraveling tolerant ALCF I-TBox is materializable.
Proof. We show the contrapositive using a proof strategy that is very similar to the second
step in the proof of Theorem 3.11. Thus, take an ALCF I-TBox T that is not materializable.
By Theorem 3.9, T does not have the disjunction property. Thus, there are an ABox A∨ ,
ELIQs C0 (x0 ), . . . , Ck (xk ), and a1 , . . . , ak ∈ Ind(A∨ ) such that T , A∨ |= C0 (a0 )∨· · ·∨Ck (ak ),
but T , A∨ 6|= Ci (ai ) for all i ≤ k. Let Ai be Ci viewed as a tree-shaped ABox with root bi ,
for all i ≤ k. Assume w.l.o.g. that none of the ABoxes A∨ , A0 , . . . , Ak share any individual
names and reserve fresh individual names c0 , . . . , ck and fresh role names r, r0 , . . . , rk . Let
the ABox A be the union of
A∨ ∪ A0 ∪ · · · ∪ Ak ∪ {r(c, c0 ), . . . , r(c, ck )}
and
{r0 (cj , b0 ), . . . , rj−1 (cj , bj−1 ), rj (cj , aj ), rj+1 (cj , bj+1 ), . . . , rk (cj , bk )}
for 1 ≤ j ≤ k. Consider the ELIQ
q = ∃r.(∃r0 .C0 u · · · u ∃rk .Ck )(x).
By the following claim, A and q witness that T is not unraveling tolerant.
Claim. T , A |= q(c), but T , Au 6|= q(c).
Proof. “T , A |= q(c)”. Take a model I of T and A. By construction of A, we have
aIi ∈ (∃rj .Cj )I whenever i 6= j. Due to the edges r0 (c0 , a0 ), . . . , rk (ck , ak ) and since T , A∨ |=
C0 (a0 ) ∨ · · · ∨ Ck (ak ), we thus find at least one ci such that cIi ∈ (∃ri .Ci )I . Consequently,
I |= q(c).
“T , Au 6|= q(c)” (sketch). Consider the elements crci ri ai in Au . Each such element is
the root of a copy of the unraveling Au∨ of A∨ , restricted to those individual names in A∨
that are reachable from ai . Since T , A∨ 6|= Ci (ai ), we find a model Ii of T and A∨ with
Iu
ai ∈
/ CiIi . By unraveling Ii , we obtain a model Iiu of T and Au∨ with ai ∈
/ Ci i . Combining
the models I0u , . . . , Iku in a suitable way, one can craft a model I of T and Au∨ such that
crci ri ai ∈
/ CiI for all i ≤ k and the ‘role edges of I’ that concern the roles r, r0 , . . . , rk are
exactly those in A. This implies I 6|= q(c) as desired.
In some more detail, I is obtained as follows. We can assume w.l.o.g. that the domains
of I0u , . . . , Iku are disjoint. Take the disjoint union of I0u , . . . , Iku , renaming ai in Iiu to crci ri ai
for all i. Now take copies J , J0 , . . . , Jk of any model of T , make sure that their domains
are disjoint and that they are also disjoint from the domain of the model constructed so
far. Additionally make sure that c ∈ ∆J and ci ∈ ∆Ji for all i. Disjointly add these models
to the model constructed so far. It can be verified that the model constructed up to this
point is a model of T . Add all role edges from A that concern the roles r, r0 , . . . , rk to the
resulting model, which has no impact on the satisfaction of T since r, r0 , . . . , rk do not occur
in T . It can be verified that I is as required.
28
CARSTEN LUTZ AND FRANK WOLTER
5. Dichotomy for ALCF I-TBoxes of Depth One
In practical applications, the concepts used in TBoxes are often of very limited quantifier
depth. Motivated by this observation, we consider TBoxes of depth one which are sets of
CIs C v D such that no restriction ∃r.E or ∀r.E in C and D is in the scope of another
restriction of the form ∃r.E or ∀r.E. To confirm that this is indeed a practically relevant
case, we have analyzed the 429 ontologies in the BioPortal repository,3 finding that after
removing all constructors that are not part of ALCF I, more than 80% of them are of depth
one. The main result of this section is a dichotomy between PTime and coNP for TBoxes
of depth one which is established by proving a converse of Theorem 4.7, that is, showing that
materializability implies unraveling tolerance (and thus PTime query evaluation and even
monadic Datalog6= -rewritability by Theorem 4.6) for TBoxes of depth one. Together with
Theorem 3.11, which says that non-materializability implies coNP-hardness, this yields the
dichotomy.
We remark that the same strategy cannot be used to obtain a dichotomy in the case of
unrestricted depth. In particular, the well-known technique of rewriting a TBox into depth
one by introducing fresh concept names can change its complexity because it enables querying
for concepts such as ¬A or ∀r.A which are otherwise ‘invisible’ to (positive existential) queries.
For TBoxes of unrestricted depth (and even in ALC) it is in fact neither the case that PTime
query evaluation implies unraveling tolerance (or even Datalog6= -rewritability) nor that
materializability implies PTime query evaluation. This is formally established in Section 6.
Theorem 5.1. Every materializable ALCF I-TBox of depth one is unraveling tolerant.
Proof. Let T be a materializable TBox of depth one, A an ABox, C0 (x) an ELIQ, and
a0 ∈ Ind(A) such that T , Au 6|= C0 (a0 ). We have to show that T , A 6|= C0 (a0 ). It follows
from T , Au 6|= C0 (a0 ) that Au is consistent w.r.t. T . There must thus be a materialization
I u for T and Au , despite the fact that Au is infinite: by Theorem 5.1, T has the disjunction
property and the argument from the proof of Theorem 5.1 that the disjunction property
implies materializability goes through without modification also for infinite ABoxes. Our
aim is to turn I u into a model I of A and T such that I 6|= C0 (a0 ). To achieve this, we first
uniformize I u in a suitable way.
We assume w.l.o.g. that I u has forest-shape, i.e., that I u can be constructed by selecting
a tree-shaped interpretation Iα with root α for each α ∈ Ind(Au ), then taking the disjoint
u
union of all these interpretations, and finally adding role edges (α, β) to rI whenever
r(α, β) ∈ Au . In fact, to achieve the desired shape we can take the i-unfolding of I u defined
and analysed in Lemmas 3.6 and 3.7, where we start the i-unfolding from the elements of
u
Ind(Au ) ⊆ ∆I .
We start with exhibiting a self-similarity inside the unraveled ABox Au and inside I u .
Claim 1. For all α, β ∈ Ind(Au ) with tail(α) = tail(β),
(1) Au |= C(α) iff Au |= C(β) for all ELIQs C(x);
u
u
(2) α ∈ C I iff β ∈ C I for all concepts C built only from concept names, ¬, and u.
To establish Point (1), take α, β ∈ Ind(Au ) such that tail(α) = tail(β) and Au 6|= C(α).
Then there is a model I of Au and T such that I 6|= C(α). One can find a model J of Au
and T such that J 6|= C(β), as follows. By construction of Au , there is an isomorphism
ι : Ind(Au ) → Ind(Au ) with ι(α) = β such that A(γ) ∈ Au iff A(ι(γ)) ∈ Au and r(γ, γ 0 ) ∈ Au
3The ontologies are available at https://bioportal.bioontology.org/ontologies.
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
29
iff r(ι(γ), ι(γ 0 )) ∈ Au for all γ ∈ Ind(Au ), all concept names A, and all role names r. We
obtain J from I by renaming each γ ∈ Ind(Au ) with ι(γ). Point (2) can be proved by a
straightforward induction on C. The base case uses Point (1) and the fact that I u is a
materialization of T and A. This finishes the proof of Claim 1.
Now for the announced uniformization of I u . What we want to achieve is that for all
α, β ∈ Ind(Au ), tail(α) = tail(β) implies Iα = Iβ (recall that Iα is the tree component of I u
rooted at α, and likewise for Iβ ). Construct the interpretation J u as follows:
• for each α ∈ Ind(Au ) with tail(α) = a, take a copy Jα of Ia with the root a renamed to α;
• then J u is the disjoint union of all interpretations Jα , α ∈ Ind(Au ), extended with a role
u
edge (α, β) ∈ rJ whenever r(α, β) ∈ Au .
Our next aim is to show that J u is as required, that is, it is a model of T and Au and
satisfies J u 6|= C0 (a0 ).
It is indeed straightforward to verify that J u is a model of Au : all role assertions are
satisfied by construction; moreover, A(α) ∈ Au implies A(a) ∈ Au where a = tail(α) , thus
a ∈ AIu and α ∈ AJu .
u
u
Next, we show that J u is a model of T . Let f : ∆J → ∆I be a mapping that assigns
to each domain element of J u the original element in I u of which it is a copy.
u
u
u
Claim 2. d ∈ C J iff f (d) ∈ C I for all d ∈ ∆J and ALCI-concepts C of depth one.
The proof of claim 2 is by induction on the structure of C. We assume w.l.o.g. that C
is built only from the constructors ¬, u, and ∃r.C. The base case, where C is a concept
name, is an immediate consequence of the definition of J u . The case where C = ¬D and
C = D1 u D2 is routine. It thus remains to consider the case C = ∃r.D, where D is built
from ¬ and u only.
u
u
u
First let d ∈ C J . Then there is a (d, e) ∈ rJ with e ∈ DJ . First assume that
u
the edge (d, e) was added to rJ because d = α and e = β for some α, β ∈ Ind(Au ) with
r(α, β) ∈ Au . Let tail(α) = a and tail(β) = b. Then we have f (α) = a and f (β) = b. By
construction of Au , r(α, β) ∈ Au implies that β = αrb or α = βr− a. In both cases we
u
u
have r(a, b) ∈ A, thus r(a, arb) ∈ Au , thus (a, arb) ∈ rI . Since β = e ∈ DJ , induction
u
u
hypothesis yields that b ∈ DI . From Point (2) of Claim 1, we obtain arb ∈ DI and we are
done. Now assume that there is an α ∈ Ind(Au ) such that (d, e) ∈ Jα . By construction of
u
u
J u , we then have (f (d), f (e)) ∈ rI and induction hypothesis yields f (e) ∈ DI .
u
u
u
Now let f (d) ∈ C I . Then there is an (f (d), e) ∈ rI with e ∈ DI . First assume that
f (d) = α and e = β for some α, β ∈ Ind(Au ) with r(α, β) ∈ Au . Since f (d) ∈ Ind(Au ), we
must have d = γ ∈ Ind(Au ) and f (d) = a ∈ Ind(A) with tail(γ) = a. By construction of Au ,
r(a, β) ∈ Au implies that β = arb, thus r(a, b) ∈ A, thus r(γ, δ) ∈ Au with (i) δ = γrb or
u
(ii) γ = δr− a and tail(δ) = b. Since arb = e ∈ DI , Point (2) of Claim 1 yields b ∈ DIu .
u
Since tail(δ) = b implies f (δ) = b, induction hypothesis yields δ ∈ DJ and we are done.
Now assume that there is an α ∈ Ind(Au ) such that (f (d), e) ∈ Iα . By construction of J u ,
f (d) being in Iα implies that α = a for some a ∈ Ind(A) and that there is an α0 ∈ Ind(Au )
such that d is in Jα0 and tail(α0 ) = a. Again by construction of J u , we thus find an e0 in
u
u
Jα0 with f (e0 ) = e and (d, e0 ) ∈ rJα0 ⊆ rJ . Induction hypothesis yields e0 ∈ DJ . This
finishes the proof of Claim 2.
It follows from Claim 2 that J u satisfies all CIs in T . To show that J u is a model of T ,
it remains to show that J u satisfies all functionality assertions. Thus, let func(r) ∈ T and
u
d ∈ ∆J . If d ∈
/ Ind(Au ), then it is straightforward to verify that, by construction of J u , d
30
CARSTEN LUTZ AND FRANK WOLTER
has at most one r-successor in J u . Now assume d = α ∈ Ind(Au ) and let tail(α) = a. By
construction of J u and Au , α has the same number of r-successors in J u as a in I u . Since
I u satisfies func(r), α can have at most one r-successor in J u .
The final condition that J u should satisfy is J u 6|= C0 (a0 ). Assume to the contrary.
We view C0 (x0 ) as a (tree-shaped) CQ. Take a homomorphism h from C0 (x0 ) to J u with
h(x0 ) = a0 . (In this proof we consider homomorphisms that do not have to preserve any
individual names.) Let the CQ q(x) be obtained from C0 (x0 ) by identifying variables y1 , y2
whenever h(y1 ) = h(y2 ). To achieve a contradiction, it suffices to exhibit a homomorphism
h0 from q(x0 ) to I u with h0 (x0 ) = a0 . We start with setting h0 (x) = h(x) whenever
h(x) ∈ Ind(Au ). Let q 0 be obtained from q(x0 ) by dropping all role atoms r(x, y) with h0 (x)
and h0 (y) already defined (which are satisfied under h0 by construction of J u and since I u
is a model of A). Because of the forest shape of J u and by construction, q 0 is a disjoint
union of ELIQs such that, in each ELIQ C(x) contained in q 0 , h0 is defined for the root x of
C(x), but not for any other variable in it. Consequently, it suffices to show that whenever
Jα |= C(α) for some ELIQ C(x) and α ∈ Ind(Au ), then I u |= C(α); the remaining part of
h0 can then be constructed in a straightforward way. Now Jα |= C(α) implies Ia |= C(a)
where a = tail(α) by choice of Jα , which yields I u |= C(a) and thus I u |= C(α) by Point (1)
of Claim 1.
This finishes the construction and analysis of the uniform model J u . It remains to
convert J u into a model I of T and the original ABox A such that I 6|= C0 (a0 ):
• take the disjoint union of the components Ja of J u , for each a ∈ Ind(A);
• add the edge (a, b) to rI whenever r(a, b) ∈ A.
It is straightforward to verify that I is a model of A: all role assertions are satisfied by
u
construction of I; moreover, A(a) ∈ A implies A(a) ∈ Au implies a ∈ AJ implies a ∈ AI .
To show that I is a model of T and that I 6|= C0 (a0 ), we first observe the following. A
bisimulation between interpretations I1 and I2 is a relation S ⊆ ∆I1 × ∆I2 such that
(1) for all A ∈ NC and (d1 , d2 ) ∈ S: d1 ∈ AI1 iff d2 ∈ AI2 ;
(2) for all r ∈ NR ∪ {s− | s ∈ NR }: if (d1 , d2 ) ∈ S and (d1 , d01 ) ∈ rI1 , then there exists
d02 ∈ ∆I2 such that (d01 , d02 ) ∈ S and (d2 , d02 ) ∈ rI2 ;
(3) for all r ∈ NR ∪ {s− | s ∈ NR }: if (d1 , d2 ) ∈ S and (d2 , d02 ) ∈ rI2 , then there exists
d01 ∈ ∆I1 such that (d01 , d02 ) ∈ S and (d1 , d01 ) ∈ rI2 .
Recall that, whenever there is a bisimulation S between I1 and I2 with (d, e) ∈ S, then
d ∈ C I1 iff e ∈ C I2 for all ALCI-concepts C [GO07, LPW11].
Claim 3. There is a bisimulation S between J u and I such that (a, a) ∈ S for all a ∈ Ind(A).
Since J u is uniform in the sense that Jα is isomorphic to Jβ whenever tail(α) = tail(β), we
find a bisimulation between Jα and Ja whenever tail(α) = a. It can be verified that the
union of all these bisimulations qualifies as the desired bisimulation S for Claim 3. Thus,
Claim 3 is proved.
It follows from Claim 3 that I satisfies all concept inclusions in T , and that I 6|= C0 (a0 ).
It thus remains to verify that I also satisfies all functionality assertions in T . This can be
done in the same way in which we have verified that J u satisfies all those assertions.
The desired dichotomy follows: If an ALCF I-TBox T of depth one is materializable,
then PEQ-evaluation w.r.t. T is in PTime and monadic Datalog6= -rewritable by Theorems 5.1
and 4.6. Otherwise, ELIQ-evaluation w.r.t. T is coNP-complete by Theorem 3.11.
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
31
Theorem 5.2 (Dichotomy). For every ALCF I-TBox T of depth one, one of the following
is true:
• Q-evaluation w.r.t. T is in PTime for any Q ∈ {PEQ,CQ,ELIQ} (and monadic Datalog6= rewritable);
• Q-evaluation w.r.t. T is coNP-complete for any Q ∈ {PEQ,CQ,ELIQ}.
For example of depth one TBoxes for which query evaluation is in PTime and for which
it is coNP-hard, please see Example 3.2; there, cases (1) and (2) are materializable and
thus in PTime while case (3) is not materializable and thus coNP-hard.
6. Query Evaluation in ALC/ALCI = CSP
We drop functional roles and consider TBoxes formulated in ALC and in ALCI showing
that query evaluation w.r.t. TBoxes from these classes has the same computational power as
non-uniform CSPs, in the following sense:
(1) for every OMQ (T , q) with T an ALCI-TBox and q an ELIQ, there is a CSP such
that the complement of the CSP and the query evaluation problem for the (T , q) are
reducible to each other in polynomial time;
(2) for every CSP, there is an ALC-TBox T such that the CSP is equivalent to the complement
of evaluating an OMQ (T , ∃x M (x)) and, conversely, for every OMQ (T , q) query
evaluation can be reduced in polynomial time to the CSP’s complement.
This result has many interesting consequences. In particular, the PTime/NP-dichotomy
for non-uniform CSPs [Bul17, Zhu17], formerly also known as the Feder-Vardi conjecture,
yields a PTime/coNP-dichotomy for query evaluation w.r.t. ALC-TBoxes (equivalently:
w.r.t. ALCI-TBoxes). Remarkably, all this is true already for materializable TBoxes. By
Theorem 3.12 and since we carefully choose the appropriate query language in each technical
result below, it furthermore holds for any of the query languages ELIQ, CQ, and PEQ (and
ELQ for ALC-TBoxes).
We begin by introducing CSPs. Since every CSP is equivalent to a CSP with a single
predicate that is binary, up to polynomial time reductions [FV98], we consider CSPs over
unary and binary predicates (concept names and role names) only. A signature Σ is a finite
set of concept and role names. We use sig(T ) to denote the set of all concept and role names
that occur in the TBox T . An ABox A is a Σ-ABox if all concept and role names in A
are in Σ. Moreover, we write A|Σ to denote the restriction of an ABox A to the assertions
that use a symbol from Σ. For two finite Σ-ABoxes A and B, we write A → B if there is
a homomorphism from A to B that does not have to preserve any individual names. A
Σ-ABox B gives rise to the following (non-uniform) constraint satisfaction problem CSP(B):
given a finite Σ-ABox A, decide whether A → B. B is called the template of CSP(B). Many
problems in NP can be given in the form CSP(B). For example, k-colorability is CSP(Ck ),
where Ck is the {r}-ABox that contains r(i, j) for all 1 ≤ i 6= j ≤ k.
We now formulate and prove Points (1) and (2) from above, starting with (1). The
following is proved in [BtCLW14].
Theorem 6.1. For every ALCI-TBox T and ELIQ C(x), one can compute a template B
in exponential time such that the query evaluation problem for the OMQ (T , C(x)) and the
complement of CSP(B) are polynomial time reducible to each other.
32
CARSTEN LUTZ AND FRANK WOLTER
The proof of Theorem 6.1 given in [BtCLW14] proceeds in two steps. To deal with
answer variables, it uses generalized CSPs with constants, defined by a finite set of templates
(instead of a single one) and admitting the inclusion of constant symbols in the signature of
the CSP (instead of only relation symbols). One shows that (i) for every OMQ (T , C(x)),
one can construct a generalized CSP with a single constant whose complement is mutually
reducible in polynomial time with the query evaluation problem for (T , C(x)) and (ii) every
generalized CSP with constants is mutually reducible in polynomial time with some standard
CSP. For the reader’s convenience, we illustrate the construction of the template from a
given OMQ, concentrating on Boolean ELIQs which are of the form ∃x C(x) with C(x) an
ELIQ. In this special case, one can avoid the use of generalized CSPs with constants.
Theorem 6.2. Let T be an ALCI-TBox, q = ∃x C(x) with C(x) an ELIQ, and Σ the
signature of T and q. Then one can construct (in time single exponential in |T | + |C|) a
Σ-template BT ,q such that for all ABoxes A:
T , A |= q iff A|Σ 6→ BT ,q
(HomDual)
Proof. Assume T and q = ∃x C(x) are given. We use the notation from the proof of Theorem 4.6. Thus, cl(T , C) denotes the closure under single negation of the set of subconcepts
of T and C, tp(T , q) denotes the set of T , q-types and for t, t0 ∈ tp(T , q) we write t r t0 if
t and t0 can be satisfied in domain elements of a model of T that are related by r. Now, a
T , q-type t omits q if it is satisfiable in a model I of T with C I = ∅. Let BT ,q be the set of
assertions A(t) such t omits q and A ∈ t and r(t, t0 ) such that t and t0 omit q and t r t0 .
It is not difficult to show that condition (HomDual) holds for all ABoxes A. Observe that
BT ,q can be constructed in exponential time since the set of T , q-types omitting q can be
constructed in exponential time.
Example 6.3. Let T = {A v ∀r.B} and define q = ∃x B(x). Then up to isomorphism,
BT ,q is {r(a, a), r(a, b), A(b), r(a, c)}.
As a consequence of Theorem 6.1, we obtain the following dichotomy result.
Theorem 6.4 (Dichotomy). For every ALCI-TBox T , one of the following is true:
• Q-evaluation w.r.t. T is in PTime for any Q ∈ {PEQ,CQ,ELIQ};
• Q-evaluation w.r.t. T is coNP-complete for any Q ∈ {PEQ,CQ,ELIQ}.
For ALC-TBoxes, this dichotomy additionally holds for ELQs.
Proof. Assume to the contrary of what is to be shown that there is an ALCI-TBox T
such that Q-evaluation w.r.t. T is neither in PTime nor coNP-hard, for some Q ∈
{PEQ,CQ,ELIQ}. Then by Theorem 3.12, the same holds for ELIQ-evaluation w.r.t. T .
It follows that there is a concrete ELIQ q such that query evaluation for (T , q) is coNPintermediate. By Theorem 6.1, there is a template B such that evaluating (T , q) is mutually reducible in polynomial time with the complement of CSP(B). Thus CSP(B) is
NP-intermediate, a contradiction to the fact that there are no such CSPs [Bul17, Zhu17].
We now establish Point (2) from the beginning of the section. In a sense, the following
provides a converse to Theorem 6.1.
Theorem 6.5. For every template B over signature Σ, one can construct in polynomial
time a materializable ALC-TBox TB such that, for a distinguished concept name M , the
following hold:
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
33
(1) CSP(B) is equivalent to the complement of the OMQ (TB , ∃x M (x)) in the sense that
for every Σ-ABox A, A → B iff TB , A 6|= ∃x M (x);
(2) the query evaluation problem for (TB , q) is polynomial time reducible to the complement
of CSP(B), for all PEQs q.
Note that the equivalence formulated in Point (1) implies polynomial time reducibility
of CSP(B) to the complement of (TB , ∃x M (x)) and vice versa, but is much stronger than
that.
Our approach to proving Theorem 6.5 is to generalize the reduction of k-colorability
to query evaluation w.r.t. ALC-TBoxes discussed in Examples 2.2 and 2.4, where the main
challenge is to overcome the observation from Example 2.4 that PTime CSPs such as
2-colorability may be translated into coNP-hard TBoxes. Note that this is due to the
disjunction in the TBox Tk of Example 2.2, which causes non-materializability. Our solution
is to replace the concept names A1 , . . . , Ak in Tk with suitable compound concepts that are
‘invisible’ to the (positive existential) query. Unlike the original depth one TBox Tk , the
resulting TBox is of depth three. This ‘hiding’ of concept names also plays a crucial role in
the proofs of non-dichotomy and undecidability presented in Section 7.4 We now formally
develop this idea and establish some crucial properties of the TBoxes that are obtained by
hiding concept names (which are called enriched abstractions below). We return to the proof
of Theorem 6.5 afterwards.
Let T be an ALCI-TBox and Σ ⊆ sig(T ) a signature that contains all role names
in T . Our aim is to hide all concept names that are not in Σ. For B ∈ NC \ Σ, let ZB
be a fresh concept name and let rB and sB be fresh role names. The abstraction of B is
the ALC-concept HB := ∀rB .∃sB .¬ZB . The Σ-abstraction C 0 of a (potentially compound)
concept C is obtained from C by replacing every B ∈ NC \ Σ with HB . The Σ-abstraction of
a TBox T is obtained from T by replacing all concepts in T with their Σ-abstractions. We
associate with T and Σ an auxiliary TBox
T ∃ = {> v ∃rB .>, > v ∃sB .ZB | B ∈ Σ}
Finally, T 0 ∪ T ∃ is called the enriched Σ-abstraction of T and Σ. To hide the concept names
that are not in Σ, we can replace a TBox T with its enriched abstraction. The following
example shows that the TBox T ∃ is crucial for this: without T ∃ , disjunctions in T over
concept names from Σ can still induce disjunctions in the Σ-abstraction.
Example 6.6. Let T = {A v ¬B1 t ¬B2 } and Σ = {A}. Then T 0 = {A v ¬HB1 t ¬HB2 }
is the Σ-abstraction of T . For A = {A(a)}, we derive T 0 , A |= ∃rB1 .>(a) ∨ ∃rB2 .>(a) but
T 0 , A 6|= ∃rB1 .>(a) and T 0 , A 6|= ∃rB2 .>(a). Thus T 0 does not have the ABox disjunction
property and is not materializable. In contrast, it follows from Lemma 6.7 below that T 0 ∪ T ∃
is materializable and, in fact, T 0 ∪ T ∃ , A |= q(a) iff T ∃ , A |= q(a) holds for all PEQs q.
In the proof of Theorem 6.5 and in Section 7, we work with TBoxes that enjoy two
crucial properties which ensure a good behaviour of enriched Σ-abstractions. We introduce
these properties next.
A TBox T admits trivial models if the singleton interpretation I with X I = ∅ for all
X ∈ NC ∪ NR is a model of T . It is Σ-extensional if for every Σ-ABox A consistent w.r.t. T ,
there exists a model I of T and A such that ∆I = Ind(A), AI = {a | A(a) ∈ A} for all
concept names A ∈ Σ, and rI = {(a, b) | r(a, b) ∈ A} for all role names r ∈ Σ.
4The ‘hiding technique’ introduced here has been adopted in [BLR+ 16] in the context of query inseparability.
34
CARSTEN LUTZ AND FRANK WOLTER
The following lemma summarizes the main properties of abstractions. Point (1) relates
consistency of ABoxes w.r.t. a TBox T to consistency w.r.t. their enriched Σ-abstraction
T 0 ∪ T ∃ . Note that the ABox A might contain the fresh symbols from T 0 but these have no
impact on consistency (as witnessed by the use of A|Σ rather than A on the left-hand side
of the equivalence). Point (2) is similar to Point (1) but concerns the evaluation of OMQs
based on T and based on T 0 ∪ T ∃ ; we only consider a restricted form of actual queries that
are sufficient for the proofs in Section 7. Points (3) and (4) together state that evaluating
OMQs (T 0 ∪ T ∃ , q) with q a PEQ is tractable on ABoxes whose Σ-part is consistent w.r.t. T .
Lemma 6.7. Let T be an ALCI-TBox, Σ ⊆ sig(T ) contain all role names in T , and assume
that T is Σ-entensional and admits trivial models. Let T 0 ∪ T ∃ be the enriched Σ-abstraction
of T . Then for every ABox A and all concept names A (that are not among the fresh symbols
in T 0 ):
(1) A|Σ is consistent w.r.t. T iff A is consistent w.r.t. T 0 ∪ T ∃ ;
(2) for all a ∈ Ind(A) and the Σ-abstraction A0 of A:
T , A|Σ |= A(a)
iff
T 0 ∪ T ∃ , A |= A0 (a)
and
T , A|Σ |= ∃x A(x) iff T 0 ∪ T ∃ , A |= ∃x A0 (x);
(3) T ∃ is monadic Datalog6= -rewritable for PEQs;
(4) if A|Σ is consistent w.r.t. T , then
T 0 ∪ T ∃ , A |= q(~a)
iff
T ∃ , A |= q(~a)
for all PEQs q and all ~a.
Proof. (1) Assume first that A is consistent w.r.t. T 0 ∪ T ∃ . We show that A|Σ is consistent
w.r.t. T . Take a model I of T 0 ∪ T ∃ and A. Define an interpretation J in the same way as I
I for all B ∈ N \ Σ. It is straightforward to show by induction for all
except that B J := HB
C
ALCI-concepts D not using the fresh symbols from Σ-abstractions and their Σ-abstractions
D0 : d ∈ DJ iff d ∈ D0I , for all d ∈ ∆I . Thus J is a model of T and A|Σ and it follows that
A|Σ is consistent w.r.t. T .
Now assume that A|Σ is consistent w.r.t. T . We show that A is consistent w.r.t. T 0 ∪ T ∃ .
Take a model I of T and A|Σ . Construct a model J of T 0 ∪ T ∃ and A as follows: ∆J is
the set of words w = dv1 · · · vn such that d ∈ ∆I and vi ∈ {rB , sB , s̄B | B ∈ NC \ Σ} where
I or v 6= r ). Now let
vi 6= s̄B if (i) i > 2 or (ii) i = 2 and (d 6∈ HB
1
B
AJ
BJ
= AI for all A ∈ NC ∩ Σ
= {d ∈ Ind(A) | B(d) ∈ A} for all B ∈ NC \ Σ
J
ZB
I
= ZB
∪ {w | tail(w) = sB } for all B ∈ NC \ Σ
rJ
J
rB
= rI for all r ∈ NC ∩ Σ
I
= rB
∪ {(w, wrB ) | wrB ∈ ∆J } for all B ∈ NC \ Σ
sJ
B
= sIB ∪ {(w, wsB ) | wrB ∈ ∆J } ∪ {(w, ws̄B ) | ws̄B ∈ ∆J } for all B ∈ NC \ Σ
J
It follows directly from the construction of J that HB
= B I , for all B ∈ NC \ Σ. Thus, for
all concepts D (not using fresh symbols from Σ-abstractions) and their Σ-abstractions D0
and all d ∈ ∆I : d ∈ D0J iff d ∈ DI . Thus, the CIs of T 0 hold in all d ∈ ∆I since the CIs
of T hold in all d ∈ ∆I . The CIs of T 0 also hold in all d ∈ ∆J \ ∆I since T admits trivial
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
35
models. Thus, J is a model of T 0 . Since J is a model of T ∃ by construction, it follows that
J is a model of T 0 ∪ T ∃ .
(2) can be proved using the models constructed in the proof of (1).
(3) is a consequence of the fact that T ∃ can be viewed as a TBox formulated in the
description logic DL-LiteR and that any OMQ (T , q) with T a DL-LiteR -TBox and q a
PEQ is known to be rewritable into a union of CQs [CDGL+ 07, ACKZ09].
(4) Assume that A|Σ is consistent w.r.t. T and that T ∃ , A 6|= q(~a). We show T 0 ∪T ∃ , A 6|=
∃ of T ∃ and A in the same way
q(~a). Note first that one can construct a hom-initial model IA
as J was constructed from I in the proof of Point (2) (by replacing I with the interpretation
∃
IA corresponding to A and not using the symbols s̄B in the construction). Thus, ∆IA
is the set of words w = av1 · · · vn such that a ∈ Ind(A) and vi ∈ {rB , sB , | B ∈ NC \ Σ}.
∃ 6|= q(~
We have IA
a). Now, as T is Σ-extensional, there is a model I of T and A|Σ with
I
∆ = Ind(A) and AI = {a | A(a) ∈ A} for all A ∈ Σ, and rI = {(a, b) | r(a, b) ∈ A} for all
role names r ∈ Σ. Construct the model J of T 0 ∪ T ∃ and A in the same way as in the proof
∃ by setting h(w) = w 0 , where w 0 is obtained from
of Point (2). Define a mapping h : J → IA
w by replacing every s̄B by sB . One can show that h is a homomorphism. Thus J 6|= q(~a)
and so T 0 ∪ T ∃ , A 6|= q(~a), as required. The converse direction is trivial.
We are now ready to prove Theorem 6.5.
Proof of Theorem 6.5. Assume a Σ-template B is given. We construct the TBox TB in two
steps. First take for any d ∈ Ind(B) a concept name Ad and define a TBox HB with the
following CIs:
t
dom v
d∈Ind(B)
Ad u Ae v ⊥
Ad u ∃r.Ae v ⊥
Ad u A v ⊥
Here dom v
t
d∈Ind(B)
Ad
for all d, e ∈ Ind(B), d 6= e
for all d, e ∈ Ind(B), r ∈ Σ, r(d, e) 6∈ B
for all d ∈ Ind(B), A ∈ Σ, A(d) 6∈ B.
Ad stands for the set of CIs
∃r.> v
t
d∈Ind(B)
Ad ,
Av
t
d∈Ind(B)
Ad ,
> v ∀r.(
t
d∈Ind(B)
Ad )
where r and A range over all role and concept names in Σ, respectively. We use a CI of the
form dom v C rather than > v C to ensure that the TBox HB admits trivial models. It
should also be clear that T is Σ-extensional. Now let M be a fresh concept name. Then the
following can be proved in a straightforward way.
Claim 1. For any ABox A the following conditions are equivalent:
(1) HB , A|Σ 6|= ∃x M (x);
(2) A|Σ is consistent w.r.t. HB ;
(3) A|Σ → B.
Thus, CSP(B) and the complement of the query evaluation problem for (HB , ∃x M (x)) are
reducible to each other in polynomial time. Because of the disjunctions, however, the query
evaluation problem w.r.t HB is typically coNP-hard even if CSP(B) is in PTime.
In the second step, we thus ‘hide’ the concept names Ad by replacing them with their
abstractions HAd . Let HB0 ∪ T ∃ be the enriched Σ-abstraction of HB . From Claim 1 and
36
CARSTEN LUTZ AND FRANK WOLTER
Lemma 6.7 (1) according to which A|Σ is consistent w.r.t. HB iff A is consistent w.r.t. HB0 ∪T ∃ ,
we obtain
Claim 2. For any ABox A not containing the concept name M , the following conditions
are equivalent:
(1) HB0 ∪ T ∃ , A 6|= ∃x M (x);
(2) A is consistent w.r.t. HB0 ∪ T ∃ ;
(3) A|Σ → B.
Let TB = HB0 ∪ T ∃ be the enriched Σ-abstraction of HB . We show that TB is as required to
prove Theorem 6.5. The theorem comprises two points:
(1) We have to show that CSP(B) is equivalent to the complement of the OMQ (TB , ∃x M (x)).
This is an immediate consequence of Claim 2.
(2) For the converse reduction, let q be a PEQ. We have to show that the query evaluation
problem for (TB , q) is reducible in polynomial time to the complement of CSP(B). Let A be
an ABox and ~a from Ind(A). We show that the following are equivalent:
(a) TB , A |= q(~a);
(b) A|Σ 6→ B or T ∃ , A |= q(~a).
Regarding (b), we remark that checking whether T ∃ , A |= q(~a) can be part of the reduction
since, by Lemma 6.7 (3), it needs only polynomial time. First assume that (a) holds. If
A|Σ → B, then by Claim 1 the ABox A|Σ is consistent w.r.t. HB . By Lemma 6.7 (4), we
obtain TB , A |= q(~a) iff T ∃ , A |= q(~a) for all PEQs q and all ~a, as required.
Conversely, assume (b) holds. If A|Σ 6→ B, then by Claim 2 A is not consistent w.r.t. TB
and so TB , A |= q(~a). If T ∃ , A |= q(~a), then TB , A |= q(~a) since T ∃ ⊆ TB .
We close this section by illustrating an example consequence of Theorem 6.5. It was
proved in [FV98] that there are CSPs that are in PTime yet not rewritable into Datalog,
and in fact also not into Datalog6= due to the results in [FV03]. This was strengthened to
CSPs that contain no relations of arity larger than two in [Ats08]. It was also observed in
[FV98] that there are CSPs that are rewritable into Datalog, but not into monadic Datalog
(such as the CSP expressing 2-colorability). This again extends to Datalog6= and applies to
CSPs with relations of arity at most two. With this in mind, the following is a consequence
of Theorems 6.5 and 3.12.
Theorem 6.8.
(1) There are ALC-TBoxes T such that PEQ-evaluation w.r.t. T is in PTime, but T is
not Datalog-rewritable for ELIQs;
(2) there are ALC-TBoxes that are Datalog-rewritable for PEQs, but not monadic Datalogrewritable for ELIQs.
7. Non-Dichotomy and Undecidability in ALCF
We show that the complexity landscape of query evaluation w.r.t. ALCF-TBoxes is much
richer than for ALCI, and in fact too rich to be fully manageable. In particular, we prove that
for CQ-evaluation, there is no dichotomy between PTime and coNP (unless PTime = NP).
We also establish that materializability, (monadic) Datalog6= -rewritability, PTime query
evaluation, and coNP-hardness of query evaluation are undecidable. We start with the
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
37
undecidability proofs, which are by reduction of an undecidable rectangle tiling problem and
reuse the ‘hidden concepts’ introduced in the previous section. Next, the TBox from that
reduction is adapted to prove the non-dichotomy result by an encoding of the computations
of nondeterministic polynomial time Turing machines (again using hidden concepts). The
basis for the technical development in this section is a TBox constructed in [BBLW16] to
prove the undecidability of query emptiness in ALCF.
An instance of the finite rectangle tiling problem is given by a triple P = (T, H, V )
with T a finite set of tile types including an initial tile Tinit to be placed on the lower left
corner and a final tile Tfinal to be placed on the upper right corner, H ⊆ T × T a horizontal
matching relation, and V ⊆ T × T a vertical matching relation. A tiling for (T, H, V ) is a
map f : {0, . . . , n} × {0, . . . , m} → T such that n, m ≥ 0, f (0, 0) = Tinit , f (n, m) = Tfinal ,
(f (i, j), f (i + 1, j)) ∈ H for all i < n, and (f (i, j), f (i, j + 1)) ∈ V for all i < m. We say that
P admits a tiling if there exists a map f that is a tiling for P. It is undecidable whether an
instance of the finite rectangle tiling problem admits a tiling.
Now let P = (T, H, V ) be a finite rectangle tiling problem with T = {T1 , . . . , Tp }. We
regard the tile types in T as concept names and set Σg = {T1 , . . . , Tp , x, y, x̂, ŷ}, where x, y,
x̂, and ŷ are functional role names. The TBox TP is defined as the following set of CIs, where
(Ti , Tj , T` ) range over all triples from T such that (Ti , Tj ) ∈ H and (Ti , T` ) ∈ V and where
for e ∈ {c, x, y} the concept Be ranges over all conjunctions L1 u L2 with Li ∈ {Ze,i , ¬Ze,i },
for concept names Ze,i (i = 1, 2):
Tfinal v Y u U u R
∃x.(U u Y u Tj ) u Ix u Ti v U u Y
∃y.(R u Y u T` ) u Iy u Ti v R u Y
∃x.(Tj u Y u ∃y.Y ) u ∃y.(T` u Y u ∃x.Y ) u Ix u Iy u C u Ti v Y
Y u Tinit v A
Bx u ∃x.∃x̂.Bx v Ix
By u ∃y.∃ŷ.By v Iy
∃x.∃y.Bc u ∃y.∃x.Bc v C
t
1≤s<t≤p
Ts u Tt v ⊥
U v ∀y.⊥ R v ∀x.⊥ U v ∀x.U R v ∀y.R
Y u Tinit v D u L D v ∀ŷ.⊥ L v ∀x̂.⊥ D v ∀x.D u ∀x̂.D L v ∀y.L u ∀ŷ.L
With the exception of the CIs in the last line, the TBox TP has been defined and analyzed in
[BBLW16]. Here, we briefly give intuition and discuss its main properties. The role names
x and y are used to represent horizontal and vertical adjacency of points in a rectangle. The
role names x̂ and ŷ are used to simulate the inverses of x and y. The concept names in TP
serve the following puroposes:
• U , R, L, and D mark the upper, right, left, and lower (‘down’) border of the rectangle.
• In the Bc concepts, the concept names Zc,1 and Zc,2 serve as second-order variables and
ensure that a flag C is set at positions where the grid cell is closed.
• In the concepts Bx and By , the concept names Zx,1 , Zx,2 , Zy,1 , Zy,2 also serve as secondorder variables and ensure that flags Ix and Iy are set at positions where x and x̂ as well
as y and ŷ are inverse to each other.
38
CARSTEN LUTZ AND FRANK WOLTER
• The concept name Y is propagated through the grid from the upper right corner to the
lower left one, ensuring that these flags are set everywhere, that every position of the
grid is labeled with at least one tile type, and that the horizontal and vertical matching
conditions are satisfied.
• Finally, when the lower left corner of the grid is reached, the concept name A is set as a
flag.
Because of the use of the concepts Be , CQ evaluation w.r.t. TP is coNP-hard: we leave it as
an exercise to the reader to verify that TP does not have the ABox disjunction property. TP
without the three CIs involving the concepts Be , however, is (equivalent to) a Horn-ALCF
TBox and thus enjoys PTime CQ-evaluation. Call a Σg -ABox A an grid ABox (with
initial individual a) if A represents a grid along with a proper tiling for P. In detail, we
require that there is a tiling f for P with domain {0, . . . , n} × {0, . . . , m} and a bijection
g : {0, . . . , n} × {0, . . . , m} → Ind(A) with g(0, 0) = a such that
• for all j < n, k ≤ m: T (g(j, k)) ∈ A iff T = f (j, k);
• for all b1 , b2 ∈ Ind(A): x(b1 , b2 ) ∈ A iff x̂(b2 , b1 ) ∈ A iff there are j < n, k ≤ m such that
(b1 , b2 ) = (g(j, k), g(j + 1, k));
• for all b1 , b2 ∈ Ind(A): y(b1 , b2 ) ∈ A iff ŷ(b2 , b1 ) ∈ A iff there are j ≤ n, k < m such that
(b1 , b2 ) = (g(j, k), g(j, k + 1)).
Clearly, if P admits a tiling then a grid ABox exists and for any grid ABox A, TP , A |= A(a)
for the (uniquely determined) initial individual a of A. The following summarizes relevant
properties of Σg -ABoxes that follow almost directly from the analysis of TP in [BBLW16]. We
say that an ABox A contains an ABox A0 if A0 ⊆ A and that A contains a closed ABox A if,
additionally, r(a, b) ∈ A and a ∈ Ind(A0 ) implies r(a, b) ∈ A0 for r ∈ {x, y, x̂, ŷ}. Moreover,
we say that inconsistency of (Σ-)ABoxes w.r.t. a TBox T is monadic Datalog6= -rewritable if
there is a Boolean monadic Datalog6= -program Π such that for any (Σ-)ABox A, A |= Π()
iff A is inconsistent w.r.t. T .
Lemma 7.1. Let P be a finite rectangle tiling problem. Then the following holds.
(1) TP admits trivial models and is Σg -extensional.
(2) Inconsistency of Σg -ABoxes w.r.t. TP is monadic Datalog6= -rewritable.
(3) If a Σg -ABox A is consistent w.r.t. TP , then A contains
• closed grid ABoxes A1 , . . . , An , n ≥ 0, with mutually disjoint sets Ind(Ai ) and
• a (possibly empty) Σg -ABox A0 disjoint from A1 ∪ · · · ∪ An
such that A = A1 ∪ · · · ∪ An ∪ A0 and TP , A |= A(a) iff a is the initial node of some Ai .
Moreover, there is a model I of A witnessing Σg -extensionality of TP that satisfies
a ∈ AI iff a is the initial node of some Ai .
Proof. (1) is a straightforward consequence of the definition of TP . (2) Assume a Σg -ABox
A is given. Apply the following rules exhaustively to A:
(a) add Ix (a) to A if there exists b with x(a, b), x̂(b, a) ∈ A;
(b) add Iy (a) to A if there exists b with y(a, b), ŷ(b, a) ∈ A;
(c) add C(a) to A if there exist a1 , a2 , b with x(a, a1 ), y(a, a2 ), y(a1 , b), x(a2 , b) ∈ A.
Denote the resulting ABox by A† . Now remove the three CIs involving the concepts Be
from TP and denote by TP† the resulting TBox. Using the analysis of the CIs involving the
concepts Be in [BBLW16], one can show that A is consistent w.r.t. TP iff A† is consistent
w.r.t. TP† . Since the latter is a Horn-ALCF-TBox, it is unraveling tolerant and one can build
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
39
a monadic Datalog6= -rewriting of the inconsistency of Σg -ABoxes w.r.t. TP† , essentially as in
the proof of Theorem 4.6. Finally, the obtained program can be modified so as to behave as
if started on A† when started on A, by implementing rules (a) to (c) as monadic Datalog
rules.
(3) This is almost a direct consequence of the properties established in [BBLW16]. In
particular, one finds the desired model I from the ‘moreover’ part by applying the three rules
from the proof of Point (2) and then applying the CIs in TP† as rules. The only condition on
the decomposition of the ABox A into A = A1 ∪ · · · ∪ An ∪ A0 that does not follow from
[BBLW16] is that the containment of A1 , . . . , An in A is closed also for the role names x̂ and
ŷ. To ensure this condition, we use the CIs that mention L and D that were not present in
the TBox used in that paper. In fact, the following two properties follow directly from these
CIs: (i) the individuals c reachable along an x-path in A from some a with TP , A |= A(a) all
satisfy TP , A |= D(c) and so do not have an x̂-successor; and (ii) the individuals c reachable
along a y-path in A from some a with TP , A |= A(a) all satisfy TP , A |= L(c) and so do
not have a ŷ-successor. (i) and (ii) together with the properties established in [BBLW16]
entail that the containment of A1 , . . . , An in A is closed also for the role names x̂ and ŷ, as
required.
Note that it follows from Lemma 7.1 (3) that if A is consistent w.r.t. T , then the
sequence A1 , . . . , An is empty iff TP , A 6|= ∃x A(x) (and A0 is non-empty since ABoxes are
non-empty). In particular this must be the case when P does not admit a tiling. In the proof
Lemma 7.2 below, this is actually all we need from Lemma 7.1 (3). In full generality, it will
only be used in the proof of non-dichotomy later on. We also remark that the decomposition
A1 ∪ · · · ∪ An ∪ A0 of A in Lemma 7.1 (3) is unique.
Let T = TP ∪ {A v B1 t B2 }, where B1 and B2 are fresh concept names. Set
Σ = Σg ∪ {B1 , B2 } and let T 0 ∪ T ∃ be the enriched Σ-abstraction of T .
Lemma 7.2.
(1) If P admits a tiling, then CQ-evaluation w.r.t. T 0 ∪ T ∃ is coNP-hard and T 0 ∪ T ∃ is
not materializable.
(2) If P does not admit a tiling, then CQ-evaluation w.r.t. T 0 ∪ T ∃ is monadic Datalog6= rewritable and T 0 ∪ T ∃ is materializable.
Proof. (1) If P admits a tiling, then there is a grid ABox A with initial node a. A uses
symbols from Σg , only. We have TP , A |= A(a) and A is consistent w.r.t. TP . By Lemma 6.7
(2), T 0 ∪ T ∃ , A |= A0 (a) where A0 is the Σ-abstraction of A. Since T 0 contains A0 v B1 t B2
and B1 and B2 do not occur elsewhere in T 0 ∪ T ∃ , it is clear that T 0 ∪ T ∃ , A |= B1 (a) ∨ B2 (a)
but T 0 ∪ T ∃ , A 6|= Bi (a) for i = 1, 2. Thus T 0 ∪ T ∃ does not have the ABox disjunction
property. It follows that T 0 ∪ T ∃ is not materializable and that CQ-evaluation w.r.t. T 0 ∪ T ∃
is coNP-hard.
(2) Assume that P does not admit a tiling. Let q be a PEQ. We show how to construct
a monadic Datalog6= -rewriting Π of (T 0 ∪ T ∃ , q). On ABoxes A that are inconsistent w.r.t.
T 0 ∪ T ∃ , Π is supposed to return all tuples over Ind(A) of the same arity as q. By Lemma 6.7
(1), an ABox A is consistent w.r.t. T 0 ∪ T ∃ iff A|Σ is consistent w.r.t. T . It thus follows from
Lemma 7.1 (2) that inconsistency of ABoxes w.r.t. T 0 ∪ T ∃ is monadic Datalog6= -rewritable.
From a concrete rewriting, we can build a monadic Datalog6= -program Π0 that checks
inconsistency and, if successful, returns all answers.
40
CARSTEN LUTZ AND FRANK WOLTER
Now for ABoxes A that are consistent w.r.t. T 0 ∪T ∃ . By Lemma 6.7 (1), A|Σ is consistent
w.r.t. TP . Since P does not admit a tiling and by Lemma 7.1 (2), TP , A|Σ 6|= ∃x A(x).
Thus, by Lemma 7.1 (1) we find a model I of TP and A|Σ such that ∆I = Ind(A),
B I = {a | B(a) ∈ A} for all B ∈ Σ, rI = {(a, b) | r(a, b) ∈ A} for all r ∈ Σ, and AI = ∅.
Since AI = ∅, I is a model of T . From Lemma 6.7 (4), we thus obtain that T 0 ∪ T ∃ , A |= q(~a)
iff T ∃ , A |= q(~a), for all ~a. By Lemma 6.7 (3), (T ∃ , q) is monadic Datalog6= -rewritable into a
program Π1 .
The desired program Π is simply the union of Π0 and Π1 , assuming disjointness of IDB
relations.
Lemma 7.2 implies the announced undecidability results.
Theorem 7.3. For ALCF-TBoxes T , the following problems are undecidable (Points 1 to 4
are subject to the side condition that PTime 6= NP):
(1) CQ-evaluation w.r.t. T is in PTime;
(2) CQ-evaluation w.r.t. T is coNP-hard;
(3) T is monadic Datalog6= -rewritable;
(4) T is Datalog6= -rewritable;
(5) T is materializable.
We now come to the proof of non-dichotomy.
Theorem 7.4 (Non-Dichotomy). For every language L ∈ coNP, there exists an ALCFTBox T such that, for a distinguished concept name M0 , the following holds:
(1) L is polynomial time reducible to the evaluation of (T , ∃x M0 (x));
(2) the evaluation of (T , q) is polynomial time reducible to L, for all PEQs q.
To prove Theorem 7.4 let L ∈ coNP. Take a non-deterministic polynomial time Turing
Machine M that recognizes the complement of L. Let M = (Q, Γ0 , Γ1 , ∆, q0 , qa , qr ) with Q a
finite set of states, Γ0 and Γ1 finite input and tape alphabets such that Γ0 ⊆ Γ1 and Γ1 \ Γ0
contains the blank symbol β, q0 ∈ Q the starting state, ∆ ⊆ Q × Γ1 × Q × Γ1 × {L, R} the
transition relation, and qa , qr ∈ Q the accepting and rejecting states. Denote by L(M ) the
language recognized by M . We can assume w.l.o.g. that there is a fixed symbol γ0 ∈ Γ0 such
that all words accepted by M are of the form γ0 v with v ∈ (Γ0 \ {γ0 })∗ ; in fact, it is easy
to modify M to satisfy this property without changing the complexity of its word problem.
We also assume that for any input v ∈ Γ∗0 , M uses exactly |v|k1 cells for the computation,
halts after exactly |v|k2 steps in the accepting or rejecting state, and does not move to the
left of the starting cell.
To represent inputs to M and to provide the space for simulating computations, we use
grid ABoxes as in the proof of Theorem 7.3, where the tiling of the bottom row represents the
input word followed by blank symbols. As the set of tile types, we use T = Γ0 ∪ {β, T, Tfinal }
where T is a ‘default tile’ that labels every position except those in the bottom row and
the upper right corner. Identify Tinit with γ0 and let Σg = Γ0 ∪ {β, T, Tfinal } ∪ {x, y, x̂, ŷ}.
Consider the TBox TPM defined above, for PM = (T, H, V ) with
H = {(γ0 , γ), (γ, γ 0 ), (γ, β), (β, β), (T, T ), (T, Tfinal ) | γ, γ 0 ∈ Γ0 \ {γ0 }}
V = {(γ, T ), (β, T ), (T, T ), (T, Tfinal ), (γ, Tfinal ), (β, Tfinal ) | γ ∈ Γ0 }.
Recall that TPM checks whether a given Σg -ABox contains a grid structure with a tiling that
respects H, V , Tinit , and Tfinal , and derives the concept name A at the lower left corner of
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
41
such grids. We now construct a TBox TM that, after the verification has finished, initiates a
computation of M on the grid. In addition to the concept names in TPM , TM uses concept
names Aγ and Aq,γ for all γ ∈ Γ1 and q ∈ Q to represent symbols written during the
computation (in contrast to the elements of Γ1 as concept names, used to encode the input
word) and to represent the state and head position. In detail, TM contains the following CIs:
• When the verification of the grid has finished, A floods the ABox:
A v ∀r.A
for all r ∈ {x, y, x̂, ŷ}.
• The initial configuration is as required:
γ0 u A v Aq0 ,γ
γ u A v Aγ
for all γ ∈ (Γ0 ∪ {β}) \ {γ0 }.
• For every (q, γ) ∈ Q × Γ1 , the transition relation of M is satisfied:
Aq,γ u A v
t
∃y.(Aγ 0 u
t
∃y.(Aγ 0 u
(q,γ,q 0 ,γ 0 ,L)∈∆
(q,γ,q 0 ,γ 0 ,R)∈∆
t
∃x̂.Aq0 ,γ 00 ) t
t
∃x.Aq0 ,γ 00 ).
γ 00 ∈Γ1
γ 00 ∈Γ1
• The representations provided by Aq,γ and Aγ for symbols in Γ1 coincide:
Aq,γ u Aγ 0 v Aq,γ 0 u Aγ ,
for all γ, γ 0 ∈ Γ1
• The symbol written on a cell does not change if the head is not on the cell:
Aγ u A v ∀y.Aγ
for all γ ∈ Γ1
• The rejecting state is never reached:
Aqr ,γ u A v ⊥
for all γ ∈ Γ1 .
Let T = TPM ∪ TM . We are going to show that an appropriate extended abstraction of T
satisfies Conditions (1) and (2) of Theorem 7.4. We start with the following lemma which
summarizes two important properties of T .
Lemma 7.5.
(1) T admits trivial models and is Σg -extensional.
(2) For any Σg -ABox A, A is consistent w.r.t. T iff the following two conditions hold:
(a) A is consistent w.r.t. TPM ;
(b) let A = A1 ∪ · · · ∪ An ∪ A0 be the decomposition of A given in Lemma 7.1 (2) and
assume that Ai is the ni × mi -grid ABox with input vi for 1 ≤ i ≤ n. Then the
following hold for 1 ≤ i ≤ n:
(i) ni ≥ |vi |k1 and mi ≥ |vi |k2 and
(ii) vi ∈ L(M ).
Proof. Since (1) follows directly from the construction of T , we concentrate on (2). Assume
first that A is consistent w.r.t. T . Then A is consistent w.r.t. TPM and so we can assume
that there is a decomposition A1 ∪ · · · ∪ An ∪ A0 of A as in Lemma 7.1 (3). By definition,
each Ai is an ni × mi -grid ABox with input vi . Since A is consistent w.r.t. T , there is a
model I of A and T . By the first CIs of TM and since the initial node a of each Ai must
be in AI by Lemma 7.1, Ind(Ai ) ⊆ AI for each i. Thus the restriction of I to Ind(Ai )
simulates an accepting computation of M starting with vi . But since every computatation
of M starting with a word of length n requires at least nk1 space and mk2 time and the
containment of Ai in A is closed for the role names x and y, this is impossible if ni < |vi |k1
or mi < |vi |k2 and also if vi 6∈ L(M ). Thus (i) and (ii) hold, as required.
42
CARSTEN LUTZ AND FRANK WOLTER
For the converse direction, assume that (a) and (b) hold. Since A is consistent w.r.t. TPM ,
there is a decomposition A1 ∪ · · · ∪ An ∪ A0 of A as in Lemma 7.1 (3). Also by Lemma 7.1
(3), there is a model I of A that witnesses Σg -extensionality of TPM such that a ∈ AI iff a is
the initial node of some Ai . We construct a model I 0 of T by modifying I as follows: with
the exception of A, the symbols of TPM are interpreted in the same way as in I and thus
0
I 0 is a model of TPM . To satisfy the first CI of TM , we set AI = Ind(A1 ∪ · · · ∪ An ). Note
that this suffices since the containment of each Ai in A is closed for the role names x and y.
The remaining symbols from TM are now interpreted in such a way that they describe on
each Ai an accepting computation for vi . This is possible since vi ∈ L(M ), ni ≥ |vi |k1 and
mi ≥ |vi |k2 , and each computation of M starting with a word v of length n requires at most
nk1 space and mk2 time. It can be verified that I 0 is a model of TM ; note that since A is a
conjunct of the left hand side of every CI in TM , the CIs in TM are trivially satisfied in every
node d ∈ ∆I \ Ind(A1 ∪ · · · ∪ An ). Thus I 0 satisfies T and A and we have proved consistency
of A w.r.t. T , as required.
We are now in the position to prove Theorem 7.4.
Proof of Theorem 7.4. Let L ∈ coNP and let M and T be the TM and TBox from above.
Set Σ = Σg ∪ {M0 } where M0 is a fresh concept name and let T 0 ∪ T ∃ be the enriched
Σ-abstraction of T . We show that T satisfies Points (1) and (2) from Theorem 7.4.
(1) It suffices to give a polynomial time reduction from L(M ) to the complement of evaluating
(T 0 ∪ T ∃ , ∃x M0 (x)) (note that M0 does not occur in any of the involved TBoxes). Assume
that an input word v for M is given. If v is not from γ0 (Γ0 \ {γ0 })∗ , then reject. Otherwise,
construct in polynomial time the |v|k1 × |v|k2 -grid ABox A with input v. Observe that A is
consistent w.r.t. TPM and has the trivial decomposition A = A1 in Lemma 7.1 (3). Thus
Lemma 7.5 (2) implies that v ∈ L(M ) iff A is consistent w.r.t. T . The latter condition is
equivalent to T , A 6|= ∃x M0 (x) since M0 does not occur in A or T . Since T admits trivial
models, Lemma 6.7 (2) thus yields v ∈ L(M ) iff T 0 ∪ T ∃ , A 6|= ∃x M0 (x).
(2) We first make the following observation.
Claim 1. T 0 ∪ T ∃ , A |= q(~a) iff
• A|Σ is not consistent w.r.t. T or
• T ∃ , A |= q(~a).
For the ‘only if’ direction, observe that if A|Σ is consistent w.r.t. T , then by Lemma 6.7
(4), T 0 ∪ T ∃ , A |= q(~a) iff T ∃ , A |= q(~a). For the ‘if’ direction, observe that if A|Σ is not
consistent w.r.t. T , then by Lemma 6.7 (1) A is not consistent w.r.t. T 0 ∪ T ∃ . This finishes
the proof of the claim.
By Lemma 6.7 (3), T ∃ , A |= q(~a) can be decided in polynomial time. Thus, Claim 1
implies that it suffices to give a polynomial time reduction of ABox consistency w.r.t. T to
L(M ). But Lemma 7.5 (2) provides a polynomial reduction of ABox consistency w.r.t. T to
L(M ) since
• Condition (a) of Lemma 7.5 (2) can be checked in polynomial time (by Lemma 7.1 (1));
• the decomposition of A in Condition (b) of Lemma 7.5 (2) as well as the words vi ,
1 ≤ i ≤ n, can be computed in polynomial time;
• and Point (i) of Condition (b) can be checked in polynomial time.
It thus remains to check whether vi ∈ L(M ) for 1 ≤ i ≤ n. This finishes the proof of
Theorem 7.4.
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
43
Theorems 7.4 and 3.12 imply that that there is no PTime/coNP-dichotomy for query
evaluation w.r.t. ALCF-TBoxes, unless PTime = NP.
Observe that the TBoxes constructed in the undecidability and the non-dichotomy proof
are both of depth four. This can be easily reduced to depth three: recall that the TBoxes of
depth four are obtained from TBoxes T of depth two by taking their enriched Σ-abstractions.
One can obtain a TBox of depth three (for which query evaluation has the same complexity
up to polynomial time reductions) by first replacing in T compound concepts C in the scope
of a single value or existential restriction by fresh concept names AC and adding AC ≡ C
to T . Then the fresh concept names are added to the signature Σ and one constructs the
enriched abstraction of the resulting TBox for the extended signature. This TBox is as
required. Thus, our undecidability and non-dichotomy results hold for ALCF-TBoxes of
depth three already.
8. Discussion
We have studied the complexity of query evaluation in the presence of an ontology formulated
in a DL between ALC and ALCF I, focussing on the boundary between PTime and coNP.
For ALCF I-TBoxes of depth one, we have established a dichotomy between PTime and
coNP and shown that it can be precisely characterized in terms of unraveling tolerance and
materializability. Moreover and unlike in the general case, PTime complexity coincides with
rewitability into Datalog6= . The case of higher or unrestricted depth is harder to analyze. We
have shown that for arbitrary ALC- and ALCI-TBoxes there is a dichotomy between PTime
and coNP. The proof is by a reduction to the recently confirmed PTime/NP-dichotomy for
CSPs. For ALCF TBoxes of depth three we have shown that there is no dichotomy unless
PTime = NP and that deciding whether a given TBox admits PTime query evaluation is
undecidable, and so are related questions.
Several interesting research questions remain. We briefly discuss three possible directions.
(1) Is it decidable whether a given ALC- or ALCI-TBox admits PTime query evaluation
and, closely related, whether it is unraveling tolerant and whether it is materializable?
First results for TBoxes of depth one have been obtained in [HLPW17a], but the general
problem remains open. It is interesting to point out that unraveling tolerance is decidable
for OMQs whose TBox is formulated in ALCI (where a concrete query is given, unlike in
the case of unraveling tolerance of TBoxes); in that case, unraveling tolerance is equivalent
to rewritability into monadic Datalog [FKL17]. It would also be interesting to study more
general notions of unraveling tolerance based on unravelings into structures of bounded
treewidth rather than into real trees.
(2) It would be interesting to study additional complexity classes such as LogSpace,
NLogSpace, and AC0 . It is known that all these classes occur even for ALC-TBoxes of
depth one, see e.g. [CDL+ 13] and the recent [LS17] which establishes a full complexity
complexity classification of OMQs that are based on an EL-TBox and an ELIQ. For example,
CQ-evaluation w.r.t. the depth one EL-TBox {∃r.A v A}, which encodes reachability in
directed graphs, is NLogSpace-complete. It would thus be interesting to identify further
dichotomies such as between NLogSpace and PTime. We conjecture that for ALCF ITBoxes of depth one, it is decidable whether query evaluation is in PTime, NLogSpace,
LogSpace, and AC0 .
44
CARSTEN LUTZ AND FRANK WOLTER
(3) Apart from Datalog, rewritability into FO queries is also of interest. In the context
of OMQs where the actual query is fixed rather than quantified, several results have been
obtained, see e.g. [BLW13, HLSW15, BHLW16] for FO-rewritability of OMQs whose TBox
is formulated in a Horn DL and [BtCLW14, FKL17] for FO- and Datalog-rewritability of
OMQs whose TBox is formulated in ALC or an extension thereof. When the query is
quantified (as in the current paper), a first relevant result has been established in [LW11]
where it is shown that that FO-rewritability is decidable for materializable ALCF I-TBoxes
of depth one. This underlines the importance of deciding materializability, which would
allow to lift this result to (otherwise unrestricted) ALCF I-TBoxes of depth one.
Acknowledgments. Carsten Lutz was supported by ERC consolidator grant 647289. Frank
Wolter was supported by EPSRC grant EP/M012646/1.
References
+
[ABI 05]
Eric Allender, Michael Bauland, Neil Immerman, Henning Schnoor, and Heribert Vollmer. The
complexity of satisfiability problems: Refining Schaefer’s theorem. In MFCS, pages 71–82, 2005.
[ACKZ09] Alessandro Artale, Diego Calvanese, Roman Kontchakov, and Michael Zakharyaschev. The
DL-Lite family and relations. J. Artif. Intell. Res. (JAIR), 36:1–69, 2009.
Foto N. Afrati, Stavros S. Cosmadakis, and Mihalis Yannakakis. On datalog vs. polynomial
[ACY91]
time. In PODS, pages 13–25, 1991.
[Ats08]
Albert Atserias. On digraph coloring problems and treewidth duality. Eur. J. Comb., 29(4):796–
820, 2008.
[Bar14]
Libor Barto. Constraint satisfaction problem and universal algebra. SIGLOG News, 1(2):14–24,
2014.
[BBLW16] Franz Baader, Meghyn Bienvenu, Carsten Lutz, and Frank Wolter. Query and predicate emptiness
in ontology-based data access. J. Artif. Intell. Res. (JAIR), 56:1–59, 2016.
[BBL05]
Franz Baader, Sebastian Brandt, and Carsten Lutz. Pushing the EL envelope. In IJCAI, pages
364–369. Professional Book Center, 2005.
Vince Barany, Georg Gottlob, and Martin Otto. Querying the guarded fragment. In LICS, pages
[BGO10]
1–10, 2010.
[BHLW16] Meghyn Bienvenu, Peter Hansen, Carsten Lutz, and Frank Wolter. First order-rewritability and
containment of conjunctive queries in Horn description logics. In IJCAI, pages 965–971, 2016.
[Bul17]
Andrei A. Bulatov. A dichotomy theorem for nonuniform CSPs. In FOCS, 2017.
[BJK05]
Andrei A. Bulatov, Peter Jeavons, and Andrei A. Krokhin. Classifying the complexity of
constraints using finite algebras. SIAM J. Comput., 34(3):720–742, 2005.
[BLW13]
Meghyn Bienvenu, Carsten Lutz, and Frank Wolter. First-order rewritability of atomic queries
in horn description logics. In IJCAI, pages 754–760, 2013.
[BMRT11] Jean-Francois Baget, Marie-Laure Mugnier, Sebastian Rudolph, and Michael Thomazo. Walking
the complexity lines for generalized guarded existential rules. In IJCAI, pages 712–717, 2011.
[BO15]
Meghyn Bienvenu and Magdalena Ortiz. Ontology-mediated query answering with data-tractable
description logics. In Reasoning Web, volume 9203 of LNCS, pages 218–307. Springer, 2015.
[BtCLW14] Meghyn Bienvenu, Balder ten Cate, Carsten Lutz, and Frank Wolter. Ontology-based data
access: A study through disjunctive datalog, CSP, and MMSNP. ACM Trans. Database Syst.,
39(4):33:1–33:44, 2014.
[BLR+ 16] Elena Botoeva, Carsten Lutz, Vladislav Ryzhikov, Frank Wolter and Michael Zakharyaschev.
Query-Based Entailment and Inseparability for ALC Ontologies In IJCAI, pages 1001–1007,
2016.
[Bul02]
Andrei A. Bulatov. A dichotomy theorem for constraints on a three-element set. In FOCS, pages
649–658, 2002.
[Bul11]
Andrei A. Bulatov. On the CSP dichotomy conjecture. In CSR, pages 331–344, 2011.
THE DATA COMPLEXITY OF DESCRIPTION LOGIC ONTOLOGIES
45
[CDGL+ 07] Diego Calvanese, Giuseppe De Giacomo, Domenico Lembo, Maurizio Lenzerini, and Riccardo
Rosati. Tractable reasoning and efficient query answering in description logics: The DL-Lite
family. J. of Autom. Reasoning, 39(3):385–429, 2007.
[CDL+ 13] Diego Calvanese, Giuseppe De Giacomo, Domenico Lembo, Maurizio Lenzerini, and Riccardo
Rosati. Data complexity of query answering in description logics. Artificial Intelligence, 195:335–
360, 2013.
[CGK13]
Andrea Calı̀, Georg Gottlob, and Michael Kifer. Taming the infinite chase: Query answering
under expressive relational constraints. J. Artif. Intell. Res. (JAIR), 48:115–174, 2013.
[CGLV00] Diego Calvanese, Giuseppe De Giacomo, Maurizio Lenzerini, and Moshe Y. Vardi. View-based
query processing and constraint satisfaction. In LICS, pages 361–371, 2000.
[CGLV03a] Diego Calvanese, Giuseppe De Giacomo, Maurizio Lenzerini, and Moshe Y. Vardi. Reasoning on
regular path queries. SIGMOD Record, 32(4):83–92, 2003.
[CGLV03b] Diego Calvanese, Giuseppe De Giacomo, Maurizio Lenzerini, and Moshe Y. Vardi. View-based
query containment. In PODS, pages 56–67, 2003.
[CGT89]
Stefano Ceri, Georg Gottlob, and Letizia Tanca. What you always wanted to know about datalog
(and never dared to ask). IEEE Trans. Knowl. Data Eng., 1(1):146–166, 1989.
[CK90]
C. C. Chang and H. Jerome Keisler. Model Theory, volume 73 of Studies in Logic and the
Foundations of Mathematics. Elsevier, 1990.
[EGOS08] Thomas Eiter, Georg Gottlob, Magdalena Ortiz, and Mantas Simkus. Query answering in the
description logic Horn-SHIQ. In JELIA, pages 166–179, 2008.
[FKL17]
Cristina Feier, Antti Kuusisto, and Carsten Lutz. Rewritability in monadic disjunctive datalog,
MMSNP, and expressive description logics. In ICDT, pages 1–17, 2017.
[FKMP05] Ronald Fagin, Phokion G. Kolaitis, Renée J. Miller, and Lucian Popa. Data exchange: semantics
and query answering. Theor. Comput. Sci., 336(1):89–124, 2005.
[FV98]
Tomás Feder and Moshe Y. Vardi. The Computational Structure of Monotone Monadic SNP
and Constraint Satisfaction: A Study through Datalog and Group Theory SIAM J. Comput.,
28(1):57–104, 1998.
[FV03]
Tomás Feder and Moshe Y. Vardi. Homomorphism closed vs. existential positive. In LICS, pages
311–320, 2003.
[GLHS08] Birte Glimm, Carsten Lutz, Ian Horrocks, and Ulrike Sattler. Conjunctive query answering for
the description logic SHIQ. JAIR, 31:157–204, 2008.
[GO07]
Valentin Goranko and Martin Otto. Model theory of modal logic. In Handbook of Modal Logic,
pages 249–329. Elsevier, 2007.
[HLSW15] Peter Hansen, Carsten Lutz, Inanç Seylan, and Frank Wolter. Efficient query rewriting in the
description logic EL and beyond. In IJCAI, pages 3034–3040, 2015.
[HLPW17a] André Hernich, Carsten Lutz, Fabio Papacchini, Frank Wolter. Dichotomies in OntologyMediated Querying with the Guarded Fragment. In PODS, pages 185–199, 2017.
[HLPW17b] André Hernich, Carsten Lutz, Fabio Papacchini, Frank Wolter. Horn Rewritability vs PTime
Query Answering for Description Logic TBoxes. In Description Logics, 2017.
[HMS07]
Ullrich Hustadt, Boris Motik, and Ulrike Sattler. Reasoning in description logics by a reduction
to disjunctive datalog. J. Autom. Reasoning, 39(3):351–384, 2007.
[HN90]
Pavol Hell and Jaroslav Nesetril. On the complexity of h-coloring. J. Comb. Theory, Ser. B,
48(1):92–110, 1990.
[KNC16]
Mark Kaminski, Yavor Nenov, and Bernardo Cuenca Grau, Datalog rewritability of Disjunctive
Datalog programs and non-Horn ontologies. Artificial Intelligence, 236: 90–118, 2016.
[Kaz09]
Yevgeny Kazakov. Consequence-driven reasoning for Horn-SHIQ ontologies. In Craig Boutilier,
editor, IJCAI, pages 2040–2045, 2009.
[KL07]
Adila Krisnadhi and Carsten Lutz. Data complexity in the EL family of description logics. In
LPAR, pages 333–347, 2007.
[KLT+ 10] Roman Kontchakov, Carsten Lutz, David Toman, Frank Wolter, and Michael Zakharyaschev.
The combined approach to query answering in DL-Lite. In KR, 2010.
[KRH07]
Markus Krötzsch, Sebastian Rudolph, and Pascal Hitzler. Complexity boundaries for Horn
description logics. In AAAI, pages 452–457, 2007.
[Kro10a]
Andrei A. Krokhin. Tree dualities for constraint satisfaction. In CSL, pages 32–33, 2010.
[Krö10b]
Markus Krötzsch. Efficient inferencing for OWL EL. In JELIA, pages 234–246, 2010.
46
[KS09]
[KZ14]
[LLT07]
[LPW11]
[LSW13]
[LS17]
[LSW15]
[LTW09]
[LW10]
[LW11]
[LW12]
[Mak87]
[Mal71]
[MG85]
[OCE08]
[PLC+ 08]
[Ros07]
[Sch78]
[Sch93]
[Zhu17]
CARSTEN LUTZ AND FRANK WOLTER
Gábor Kun and Mario Szegedy. A new line of attack on the dichotomy conjecture. In STOC,
pages 725–734, 2009.
Roman Kontchakov and Michael Zakharyaschev. An introduction to description logics and query
rewriting. In Reasoning Web, volume 8714 of LNCS, pages 195–244. Springer, 2014.
Benoit Larose, Cynthia Loten, and Claude Tardif. A characterisation of first-order constraint
satisfaction problems. Logical Methods in Computer Science, 3(4), 2007.
Carsten Lutz, Robert Piro, and Frank Wolter. Description logic tboxes: Model-theoretic
characterizations and rewritability. In IJCAI, 2011.
Carsten Lutz, Inanç Seylan, and Frank Wolter. Ontology-based data access with closed predicates
is inherently intractable (sometimes). In IJCAI, pages 1024–1030. IJCAI/AAAI, 2013.
Carsten Lutz and Leif Sabellek. Ontology-Mediated Querying with the Description Logic EL:
Trichotomy and Linear Datalog Rewritability. In IJCAI, pages 1181–1187. IJCAI/AAAI, 2017.
Carsten Lutz, Inanç Seylan, and Frank Wolter. Ontology-mediated queries with closed predicates.
In IJCAI, pages 3120–3126. AAAI Press, 2015.
Carsten Lutz, David Toman, and Frank Wolter. Conjunctive query answering in the description
logic EL using a relational database system. In IJCAI, pages 2070–2075, 2009.
Carsten Lutz and Frank Wolter. Deciding inseparability and conservative extensions in the
description logic EL. J. Symb. Comput., 45(2):194–228, 2010.
Carsten Lutz and Frank Wolter. Non-uniform data complexity of query answering in description
logics. In Description Logics, 2011.
Carsten Lutz and Frank Wolter. Non-uniform data complexity of query answering in description
logics. In KR. AAAI Press, 2012.
Johann A. Makowsky. Why Horn formulas matter in computer science: Initial structures and
generic examples. J. Comput. Syst. Sci., 34(2/3):266–292, 1987.
Anatoli I. Malcev. The metamathematics of algebraic systems, collected papers:1936-1967. NorthHolland, 1971.
Jose Meseguer and Joseph A. Goguen. Initiality, induction, and computability. In Algebraic
Methods in Semantics, pages 459–541. Cambridge University Press, 1985.
Magdalena Ortiz, Diego Calvanese, and Thomas Eiter. Data complexity of query answering in
expressive description logics via tableaux. J. of Autom. Reasoning, 41(1):61–98, 2008.
Antonella Poggi, Domenico Lembo, Diego Calvanese, Giuseppe De Giacomo, Maurizio Lenzerini,
and Riccardo Rosati. Linking data to ontologies. J. Data Semantics, 10:133–173, 2008.
Riccardo Rosati. The limits of querying ontologies. In ICDT, volume 4353 of LNCS, pages
164–178. Springer, 2007.
Thomas J. Schaefer. The complexity of satisfiability problems. In STOC, pages 216–226, 1978.
Andrea Schaerf. On the complexity of the instance checking problem in concept languages with
existential quantification. J. of Intel. Inf. Systems, 2:265–278, 1993.
Dmitriy Zhuk. The Proof of CSP Dichotomy Conjecture. In FOCS, 2017.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view a
copy of this license, visit https://creativecommons.org/licenses/by-nd/4.0/ or send a
letter to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105, USA, or
Eisenacher Strasse 2, 10777 Berlin, Germany
| 2cs.AI
|
arXiv:1803.07074v1 [math.ST] 19 Mar 2018
On Time-Varying Amplitude HGARCH Model
Ferdous Mohammadi Basatini and Saeid Rezakhah1
Amirkabir University of Technology
424 Hafez Avenue, Tehran 15914, Iran.
1
email: [email protected]
Abstract
The HGARCH model allows long-memory impact in volatilities. A new HGARCH model
with time-varying amplitude is considered in this paper. We show the stability of the model
as well. A score test is introduced to check the time-varying behavior in amplitude. Some
value-at-risk tests are applied to evaluate the forecastings. Simulations are provided which
provide further support to the proposed model. We have also have shown the competative
performance of our model in forecasting, by compairing it with HGARH and FIGARCH
models for some period of SP500 indices.
Keyword: HGARCH, long-memory, time-varying, amplitude.
JEL: 13, 22, 58
Mathematics Subject Classification: 91B84, 91B30, 62F03
1
Introduction
Determining the volatility structure is the main step in measuring risk in financial time series.
The GARCH models (Engle, 1982; Bollerslev,1986) are widely used for modeling volatility. Two
kinds of structure are recognized for GARCH models as geometric and hyperbolic decaying that
can be described as some kinds of short-memory and long-memory respectively. Long-memory
property is present in the volatility of many financial data (Kwan et al., 2011). As a hyperbolicmemory model, HYGARCH (Davidson, 2004) is the most popular one and has shown good
performance in modeling long-memory behavior for many financial time series (Davidson, 2004;
Tang and Shieh, 2006). The conditional variance of HYGARCH model is a convex combination
of the conditional variances of GARCH (Bollerslev, 1986) and FIGARCH (Baillie, 1996). The
FIGARCH also shows hyperbolic-memory but has infinite variance. Li et al. (2015) argued that
the conditional variance of the HYGARCH model has an unnecessarily complicated form. This
motivated them to propose a new hyperbolic GARCH (HGARCH) model which is as simple as
FIGARCH but has finite variance.
Financial time series often have time-varying volatilities which in many cases follow long
memory in effect of exogenous and endogenous shocks. Thus models with time-varying structure are more appropriate for many financial time series. We consider a HGARCH model with
logistic time-varying amplitude to impose a more flexible behavior which we call TV-HGARCH.
This time-varying amplitude allows the conditional variance to be more sensitive to the last
observation. So when a sudden shock influences the volatilities the TV-HGARCH permits the
magnitude of variations in the conditional variance changes and so make more dynamical behavior. We show under some regularity conditions the moments of the model are bounded.
1
Maximum likelihood estimators (MLEs) of the parameters are derived. We develop a score test
to check the presence of the time-varying amplitude in the proposed TV-HGARCH structure.
The asymptotic behavior of MLEs and score test is verified by simulation. Value-at-risk (VaR) is
a useful measure for quantifying the risk which depends directly on the volatility. The forecasts
from various volatility models are evaluated and compared on the basis of how well they forecast VaR. Hence, we perform some statistical hypothesis testing to compare the VaR forecasts
of competing models. We consider S &P 500 indices from 17th February 2009 to 30th January
2015 to show the competitive behavior of TV-HGARCH model in compare to HGARCH and
FIGARCH. The paper organized as follows. The TV-HGARCH model and the moment properties are given in section 2. Maximum likelihood estimation is proposed in section 3. A score
test is developed in Section 4 for checking time-varying amplitude. Section 5 reports the simulation studies. The VaR forecasting and its statistical testings are provided in Section 6. The
performance of the model for the empirical data of S &P 500 indices is reported in Section 7.
Conclusions are presented in the last section.
2
The model
Let {yt } follows a HGARCH(q, d, p) model as
yt = t
ht =
p
ht
γ
δ(B)
+ w[1 −
(1 − B)d ]yt2 ,
β(1)
β(B)
(2.1)
where {t } are identically and independently (i.i.d.) random variables with mean 0 and variance
P
P
1, γ > 0, 0 < w < 1, B is the back-shift operator, β(x) = 1 − pi=1 βi xi , δ(x) = 1 − qi=1 δi xi
dΓ(i − d)
i
and p, q are known positive integers; also (1 − B)d = 1 − Σ∞
i=1 gi B where gi =
Γ(1 − d)Γ(i + 1)
in which 0 < d < 1. Let Υt−1 be the information up to t-1 then ht is the conditional variance
as, V ar(yt |Υt−1 ) = ht . The parameter w is called the amplitude parameter that determines the
magnitude of variations in the conditional variance (Kwan et al., 2012). For w = 1 the model
will reduce to the FIGARCH. In this model the ht has fixed form by enriching the HGARCH
model with a time-varying amplitude we provide a more dynamical model for describing the
volatilities.
2.1
The Time-Varying HGARCH Model
Let {yt } follows the TV-HGARCH(q, d, p) model as
yt = t
2
p
ht
ht =
γ
δ(B)
+ wt [1 −
(1 − B)d ]yt2 ,
β(1)
β(B)
(2.2)
where {t }, γ, B, β(x), δ(x), p, q, (1 − B)d , are defined as in (2.1). Here wt is a logistic timevarying function defined as
wt =
exp(η y˜t )
.
1 + exp(η y˜t )
(2.3)
It is clear that wt bounded between 0,1. η > 0 is called the smoothness parameter which determines the speed of transition between high and low volatility. In financial time series several
possible choices for the transition variable, y˜t are proposed (Dijk et al., 2002; McAleer, 2008).
2
We consider y˜t = yt−1
so the amplitude changes with the size of the last observation and hence
the magnitude of the last shock cause of the smooth changes of the conditional variance.
2.2
Moment properties
Now we study the moments of the {yt }. Let ϕ = (γ, β1 , ..., βp , δ1 , ..., δq , d)0 , we can rewrite model
(2.2) into the form:
ht = φ0 + wt φ(B)yt2 = φ0 + wt
∞
X
2
φi yt−i
,
i=1
where the φi ’s for i = 0, 1, 2, ... are functions of ϕ. Denote E|yt2 |m = Mm , E|2t |m = µm and
P∞
m
S =
i=1 φi . Note that Mm = µm E(ht ), assuming that φi ≥ 0 and using the fact that
0 ≤ wt ≤ 1 it holds that
hm
t
= (φ0 + wt
∞
X
2
φi yt−i
)m
(2.4)
i=1
m
∞
m−r
X
m r m−r X
2
=
φ 0 wt
φi yt−i
r
r=0
i=1
m
∞
m−r
X m X
2
≤
φr0
φi yt−i
r
r=0
i=1
m
∞ X
∞
∞
X m X
X
r
2
2
=
φ0
...
φi1 φi2 ... φim−r yt−i
y 2 ... yt−i
.
m−r
1 t−i2
r
r=0
i1 =1 i2 =1
im−r =1
By the law of iterated expectations,
E(hm
t )
≤
m
X
m
r=0
r
φr0
∞ X
∞
X
...
i1 =1 i2 =1
∞
X
2
2
2
φi1 φi2 ... φim−r E yt−i
y
...
y
.
t−i
t−i
m−r
1
2
im−r =1
Using Holder’s inequality, it holds that
3
(2.5)
(2.6)
(2.7)
Mm
m
X
m r m−r
φ0 S
Mm−r
≤ µm
r
r=0
and therefore
µm
m r m−r
φ0 S
Mm−r
r=1
r
.
1 − S m µm
Pm
Mm ≤
(2.8)
The right hand side of (2.8) is a recursive relation, so if the M1 , M2 , ...Mm−1 are exist the
condition S m µm < 1 is sufficient condition for the existence of the Mm . We find that for
m = 6, 8 the condition (2.8) is the same as the conditions for ARFIMA-HYGARCH model
presented by Kwan et al. (2012). As an example, we calculate the second-order moment for the
TV-HGARCH(1,d,1) model
yt = t
ht =
p
ht
γ
1 − δB
+ wt [1 −
(1 − B)d ]yt2 ,
1−β
1 − βB
wt =
exp(η y˜t )
.
1 + exp(η y˜t )
After some calculations it holds that
2
ht = γ + βht−1 + wt [(δ − β + g1 )yt−1
+
∞
X
2
(gi − δgi−1 )yt−i
].
(2.9)
i=2
Also if (δ − β + g1 ) ≥ 0 and (gi − δgi−1 ) ≥ 0 for i = 2, 3, ... we have that
ht ≤ γ + βht−1 + [(δ − β +
2
g1 )yt−1
∞
X
2
+
(gi − δgi−1 )yt−i
],
i=2
and therefore
M≤
Thus the (δ +g1 )+
P∞
1 − (δ + g1 ) −
i=2 (gi −δgi−1 )
γ
P∞
i=2 (gi
− δgi−1 )
.
< 1 is sufficient for the existence of the second-order moment
of the yt ; i.e. M < ∞.
3
Estimation
Let θ = (ϕ0 , η)0 denotes the parameter vector of the TV-HGARCH model defined in relations
(2.2) - (2.3) and ht (θ) refers to the conditional variance of the yt when the true parameters
in TV-HGARCH model are replaced by the corresponding unknown parameters. Suppose the
y1 , ..., yT are a sample from the TV-HGARCH model. By assuming the normality on t , the
P
conditional log likelihood function is L(θ) = −0.5 Tt=1 lt (θ) where
lt (θ) = ln 2π + ln ht (θ) +
4
yt2
.
ht (θ)
The derivatives of L(θ) with respect to the parameters are given as follows:
T
∂L(θ) X 1
y2
∂ht (θ)
=
( t − 1)
∂θ(i)
2ht (θ) ht (θ)
∂θ(i)
t=1
where θ(i) refers to the i − th element of the θ. The partial derivatives of ht (θ) are obtained as:
p
X ∂ht−i
∂ht (θ)
=1+
βi
,
∂γ
∂γ
i=1
p
X ∂ht−i
∂ht (θ)
2
= ht−k +
βi
− wt yt−k
∂βk
∂βk
k = 1, ..., p,
i=1
p
∂ht (θ) X ∂ht−i
2
=
βi
+ wt (1 − B)d yt−j
∂δj
∂δj
∂ht (θ)
=
∂d
∂ht (θ)
=
∂η
i=1
p
X
i=1
p
X
i=1
j = 1, ..., q,
βi
∂ht−i
− wt δ(B)(1 − B)d log(1 − B)yt2 ,
∂d
βi
∂ht−i ∂wt
δ(B)
−
1−
(1 − B)d yt2 ,
∂η
∂η
β(B)
∂wt
y˜t exp(η y˜t )
=
.
∂η
(1 + exp(η y˜t ))2
Here we need some numerical approaches such as quasi-Newton algorithms to find the maximum
likelihood estimator of the θ (Chong and Zak, 2001).
4
Testing Time-Varying Amplitude
For fitted HGARCH a score test is developed to check the presence of the time-varying amplitude
in the model. It is very proper test because only requires the constrained estimator under
H0 . The null hypothesis of testing time-varying amplitude corresponds to testing H0 : η = 0
against H1 : η > 0 in the TV-HGARCH model defined by relations (2.2) - (2.3). Under null
1
hypothesis wt = . The null hypothesis implies the absence of the time-varying amplitude and
2
we obtain standard HGARCH model (Amado and Teräsvirta, 2008). Consider the conditional
P
log-likelihood function L(ϕ, η) = −0.5 Tt=1 lt (ϕ, η) where
lt (ϕ, η) = ln 2π + ln ht (ϕ, η) +
yt2
.
ht (ϕ, η)
At following the ∼ indicates the maximum likelihood estimator under H0 .
1 PT ∂lt (θ)
Let ξT (θ) = √
is the average score test vector and I(θ) is the population informat=1
∂θ
T
tion matrix. Consider θ0 = (ϕ00 , 0)0 as true parameter vector under H0 . The score test statistic
5
is defined as follows:
λs = ξT (θ̃)0 I −1 (θ0 )ξT (θ̃) ∼ χ2(1) .
(4.1)
1 PT ∂lt (ϕ, η)
Also, let ξT (θ) = (ξ1T (ϕ0 ), ξ2T (η))0 where ξ1T (ϕ) = √
and
t=1
∂ϕ
T
1 PT ∂lt (ϕ, η)
ξ2T (η) = √
. So
t=1
∂η
T
0
ξT (θ̃) = (0, ξ2T (0)) ,
T
1 X ∂lt (ϕ̃, 0)
ξ2T (0) = √
∂η
T t=1
(4.2)
and
∂ht (ϕ̃, 0)
∂lt (ϕ̃, 0)
yt2
1
= (1 −
)
.
∂η
ht (ϕ̃, 0) ht (ϕ̃, 0)
∂η
Under normality, the population information matrix equals to negative expected value of the
average Hessian matrix:
"
#
"
#
T
T
1 X ∂ 2 lt (θ)
1 X ∂lt (θ) ∂lt (θ)
∂ 2 log f (yt |Υt−1 , θ)
= −E
=E
.
I(θ) = E
∂θ∂θ0
T
∂θ∂θ0
T
∂θ
∂θ0
t=1
t=1
Note (4.1) depend on the unknown parameter value θ0 so it is useless. It is common to
evaluate the I −1 (θ0 ) at the θ̃ to get a usable statistic. Hence
I˜11 I˜12
I(θ̃) =
I˜21 I˜22
(4.3)
where
I˜11 = κ̃J,
I˜12 = I˜21 = κ̃R,
I˜22 = κ̃Q,
(4.4)
T
2
1 X yt2
−1
κ̃ =
T
ht (ϕ̃, 0)
J=
R=
Q=
Denote
1
T
1
T
1
T
t=1
T
X
t=1
T
X
t=1
T
X
t=1
∂h (ϕ̃, 0) ∂h (ϕ̃, 0)
1
t
t
∂ϕ
∂ϕ0
h2t (ϕ̃, 0)
∂h (ϕ̃, 0) ∂h (ϕ̃, 0)
1
t
t
∂η
∂ϕ
h2t (ϕ̃, 0)
∂h (ϕ̃, 0) 2
1
t
.
∂η
h2t (ϕ̃, 0)
T
1 X ∂lt (ϕ̃, 0)
S(ϕ̃) = ξ2T (0) = √
,
∂η
T t=1
6
(4.5)
then by substituting (4.2) - (4.5) in (4.2) the score test statistic can be obtained as
λs =
S 2 (ϕ̃)
.
κ̃(Q − R0 J −1 R)
(4.6)
Hence if θ̃ = (ϕ̃0 , 0)0 is asymptotically normal then under H0 : η = 0 the λs will asymptotically
follows the chi-squared distribution with 1 degree of freedom under some regularity conditions
(Li et al., 2011).
5
Simulation Study
This section conducts two simulation experiments to investigate the consistency of the MLEs
(section 3) and the asymptotic behavior of the score test (section 4). We consider three sample
sizes, n=300, 500 and 1000 in two experiments, and there are 1000 replications for each sample
size. In each generated sequence the first 1000 observations have been discarded to avoid the
initialization effects, so there are 1000+n observations generated each time. We simulate the
data from a TV-HGARCH(1,d,1) model as follows:
yt = t
ht =
p
ht
1 − δB
γ
+ wt [1 −
(1 − B)d ]yt2 ,
1−β
1 − βB
wt =
2 )
exp(ηyt−1
2 ).
1 + exp(ηyt−1
(5.1)
where {t } are iid standard normal variables.
In the first experiment the value of the parameter vector is θ = (γ, β, δ, d, η)0 = (.3, .4, .2, .7, 1)0 .
The MLE values in section 3 are calculated, the biases (Bias) and the root mean squared error
(RMSE) are summarized in Table 1. It is observed that both Bias and RMSE are generally
small and decrease as the sample size increases.
The second experiment is conducted to evaluate the empirical sizes and powers of the
score test statistic λs in section 4. The value of the parameter vector is θ = (γ, β, δ, d, η)0 =
(.3, .4, .2, .7, η)0 , when η = 0 corresponds to the size and η > 0 correspond to the power of the
test. We consider three different values η=0.4, 1 and 3 and two significance values .05 and .10.
The empirical rejection rates are reported in Table 2. It can be seen that the empirical sizes
are all close to the nominal values and this closeness increases as the sample size increases also
empirical powers are increasing function of the sample size and of the η.
6
VaR Forecasting
In order to investigate the ability of the TV-HGARCH model in forecasting the future behavior
of the volatilities, we study the VaR forecasts. The one-day-ahead VaR with probability ρ,
7
Table 1: Estimation results of the TV-HGARCH model based on 1000 replications.
n=300
n=500
n=1000
parameter
Real value
Bias
RMSE
Bias
RMSE
Bias
RMSE
γ
0.3
0.031
0.162
0.017
0.130
0.005
0.030
β
0.4
0.030
0.014
0.008
0.012
0.001
0.006
δ
0.2
0.048
0.003
0.044
0.002
0.038
0.001
d
0.7
0.055
0.003
0.030
0.001
0.026
0.0008
η
1
0.084
0.019
0.057
0.019
0.025
0.018
Table 2: Empirical rejection rates of the score test for the TV-HGARCH model based on 1000
replications for two significance level 0.05 and 0.10. Also η = 0 corresponds to the size and
η > 0 correspond to the power of the test.
n=300
n=500
n=1000
η
0.05
0.10
0.05
0.10
0.05
0.10
0
0.069
0.112
0.057
0.108
0.049
0.095
0.4
0.247
0.421
0.547
0.731
0.838
0.913
1.5
0.290
0.474
0.578
0.759
0.889
0.958
3
0.294
0.492
0.613
0.791
0.915
0.966
V aR(ρ), is calculated by V aRt (ρ) = F −1 (ρ)σt , where F −1 (ρ) is the inverse distribution of
p
standardized observation (yt /σt ) and σt = V (yt |Υt−1 ). Due to the importance of VaR in
management risk, the accuracy of the VaR forecasts from different models is evaluated based on
some likelihood ratio (LR) tests (Ardia, 2009; Brooks and Persand, 2000).
Unconditional Coverage test
The Kupiec test (Kupiec, 1995), also known as the unconditional coverage (UC) test, is designed
to test whether VaR forecasts cover the pre-specified probability. If the actual loss exceeds
the VaR forecasts, this is termed an “exception,” which is a Bernoulli random variable with
probability ξ. The null hypothesis of the UC test is H0 : ξ = ψ. Then the LR statistic of the
unconditional coverage (LRU C ) is defined as
LRU C = −2 log(
8
ρn (1 − ρ)T −n
).
ˆ T −n
ξˆn (1 − ξ)
n
Where T is the number of the forecasting samples, n is the number of the exceptions and ξˆ =
T
is the MLE of the ξ under H1 . Then under H0 the LRU C is asymptotically distributed as a χ2
random variable with one degree of freedom.
Independent Test
If the volatilities are low in some periods and high in others, the forecasts should respond to this
clustering event. It means that, the exceptions should be spread over the entire sample period
independently and do not appear in clusters (Sarma et al., 2003). Christoffersen (1998) designed
an independent (IND) test to check the clustering of the exceptions. The null hypothesis of the
IND test assumes that the probability of an exception on a given day t is not influenced by what
happened the day before. Formally, H0 : ξ10 = ξ00 , where ξij denotes that the probability of an
i event on day t − 1 must be followed by a j event on day t where i, j = 0, 1. The LR statistic
of the IND test (LRIN D ) can be obtained as
LRIN D = −2 log(
ξˆn (1 − ˆξ)T −n
).
n11
ξˆn01 (1 − ξˆ01 )n00 ξˆ11 (1 − ξˆ11 )n10
01
Where nij is the number of observations with value i followed by value j (i, j = 0, 1), ξ01 =
n11
n01
and ξ11 =
. Under H0 , the LRU C is asymptotically distributed as a χ2
n00 + n01
n10 + n11
random variable with one degree of freedom.
Conditional Coverage test
Also Christoffersen (1998) proposed a joint test: the conditional coverage (CC) test, which
combines the properties of both the UC and IND tests. The null hypothesis of the CC test
checks both the exception cluster and consistency of the exceptions with VaR confidence level.
The null hypothesis of the test is H0 : ξ01 = ξ11 = ρ. The LR statistic of the CC test (LRCC )
is obtained as
LRCC = −2 log(
ρn (1 − ρ)T −n
).
n11
ξˆn01 (1 − ξˆ01 )n00 ξˆ11 (1 − ξˆ11 )n10
01
Under H0 , LRCC is asymptotically distributed as a χ2 random variable with two degrees of
freedom. It is a summation of two separate statistics, LRU C and LRIN D .
7
Empirical Data
In this section, we apply the TV-HGARCH(1,d,1), HGARCH(1,d,1) and FIGARCH(1,d,1) models on the daily percentage log-returns of the S &P 500 indices from February 17, 2009 to January
9
30, 2015 (1500 observations). Figure 1 presents the time plot of data. Some descriptive statistics
of the S &P 500 indices are listed in Table 3. We observe the negative skewness and excess kurtosis of these returns. To compare the empirical performance of the models from both fitting and
forecasting the whole sample is divided into two parts. The first part contains 1,000 observations
and is used as in-sample data to conduct fitting and the second part is used as out-of-sample
data to evaluate model forecasting. The models are then applied to the first part of data. The
MLE values are reported in Table 4. Also the score test in section 4 is performed and the value
λs = 6.44 is obtained, so the critical value 3.84 shows that at 5% significance level the data
possesses a time-varying amplitude. To evaluate the performance of the models in computing
true conditional variances that are measured by squared returns, we considered the root mean
squared error (RMSE) and the log likelihood value (LLV) for in-sample and out-of-sample data.
As out-of-sample performance, the one-day-ahead forecasts are computed using estimated models. The results are given in Table 5. It is observed that the TV-HGARCH model has the best
performance. The HGARCH model outperforms the FIGARCH model, and has a lower RMSE
and a higher LLV. To clarify the out-performance of the TV-HGARCH model, we plot the forecasting conditional variances and true conditional variances for some of the data in Figure 2. It
can be seen that the TV-HGARCH follows the shocks very well. Figure 3 shows the absolute
forecasting errors between different models and the true conditional variances for some of the
data, it can be observed that the TV-HGARCH model has the smallest absolute error. Based
on the out-of-sample data, one-day-ahead VaR forecasts at a level risk of ρ = 0.05, 0.10 for the
models are calculated and the accuracy tests are performed. The results are reported in Table 6.
The first and second rows show the number of expected exceptions (Ex.e) and empirical exceptions (Em.e) respectively. It can be seen that the Em.e for the TV-HGARCH model is closer to
the Ex.e than HGARCH and FIGARCH models; in this respect also the HGARCH make better
results than the FIGARCH. For VaR(0.05) at 5% significance level, the TV-HGARCH model
passes all the tests while the HGARCH model passes IND and CC tests and FIGARCH model
passes only IND test. Also for VaR(0.10), all models pass only IND test but the TV-HGARCH
model has the smallest LRU C and LRCC . Hence, the results indicate that the TV-HGARCH
model produces the most accurate VaR forecasts. Also the HGARCH model outperforms the
FIGARCH model.
Conclusion:
10
Figure 1: Percentage log returns of S &P 500 daily log-returns.
Figure 2: Squared returns and forecasting conditional variances with TV-HGARCH, HGARCH
and FIGARCH models for some of S &P 500 daily log-returns.
11
Table 3: Descriptive statistics of S &P 500 daily log-returns
series
Mean
Std.dev
Minimum
Maximum
Skewness
Kurtosis
S &P
0.062
1.114
-6.896
6.837
-0.148
4.564
Figure 3: Absolute forecasting errors between squared returns and forecasting conditional variances with TV-HGARCH, HGARCH and FIGARCH models for some of S &P 500 daily logreturns.
Table 4: Maximum likelihood estimates of TV-HGARCH, HGARCH and FIGARCH models on
S &P 500 daily log-returns.
TV-HYGARCH
HGARCH
FIGARCH
γ
0.179
0.237
0.237
β
0.340
0.455
0.505
δ
0.315
0.315
0.315
d
0.550
0.567
0.505
η
2.444
-
-
w
-
0.972
-
HGARCH is a hyperbolic-memory process. In this study we have generalized it by introducing
TV-HGARCH model to have a better description of the dynamic volatilities. Our proposed
model exploits a time-varying amplitude to update the structure of the volatility using logistic
function of last observation. We show under some conditions the moments of model are bounded.
12
Table 5: Measures of performance of TV-HGARCH, HGARCH and FIGARCH models on
S &P 500 daily log- returns
In-Sample
Out-of-Sample
Model
RMSE
LLV
RMSE
LLV
TV-HGARCH
1.406
-1187.4
0.431
-447.9
HGARCH
1.808
-1294.3
0.637
-514.6
FIGARCH
1.988
-1325.6
0.710
-530.5
Table 6: VaR forecasting for TV-HGARCH, HGARCH and FIGARCH models on S &P 500 daily
log-returns at level ρ = 0.05, 0.10.
TV-HGARCH
HGARCH
FIGARCH
VaR(0.05)
VaR(0.10)
VaR(0.05)
VaR(0.10)
VaR(0.05)
VaR(0.10)
Ex.e
25
50
25
50
25
50
Em.e
21
33
16
29
13
28
0.711
∗
7.210
3.890
11.371
7.298
12.588
LRIN D
1.932
∗
∗
∗
∗
∗
0.379∗
LRCC
2.642∗
8.047
12.967
LRU C
0.954
8.164
1.125
5.013∗
0.481
11.852
0.748
Notes: 1. At the 5% significance level the critical value of the LRU C and LRIN D is 3.84 and for LRCC
is 5.99. 2. * indicates that the model passes the test at 5% significance level.
One of the privilege of this work is implying of score test to check existence of the time-varying
structure. Simulation evidences showed that empirical performance of test is competitive. The
empirical example of some periods of S &P 500 indices showed that the TV-HGARCH model
gives better forecasting of volatilities and more accurate VaR than HGARCH and FIGARCH.
References
[1] Amado C. and Teräsvirta T. (2008). Modelling conditional and unconditional heteroskedasticity with smoothly time-varying structure, CREATES; Research Paper,
http://ssrn.com/abstract=1148141.
[2] Ardia D. (2009). Bayesian estimation of a Markov-switching threshold asymmetric GARCH
model with student t innovations, Econometrics Journal, 12(1), 105-126.
13
[3] Baillie R. T., Bollerslev T. and Mikkelsen H. O. (1996). Fractionally integrated generalized
autoregressive conditional heteroscedasticity, Journal of Econometrics, 74(1), 3-30.
[4] Bollerslev T. (1986). Generalized autoregressive conditional heteroscedasticity, Journal of
Econometrics, 31(3), 307-327.
[5] Brooks C. and Persand G. (2000). Value at risk and market crashes, Journal of Risk, 2(4),
526.
[6] Christoffersen P. (1998). Evaluating interval forecasts. International Economic Review, 39,
841-862.
[7] Davidson J. (2004). Moment and memory properties of linear conditional heteroscedasticity
models, and a new model, Journal of Business and Economic Statistics, 22(1), 16-19.
[8] Dijk D., Teräsvirta T. and Franses H. (2002). Smooth transition autoregressive models: A
survey of recent developments, Econometric Reviews, 21(1), 1-47.
[9] Engle R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the
variance of United Kingdom inflation, Econometrica, 50(4), 987-1007.
[10] Gerlach R. and Chen C. (2008). Bayesian inference and model comparison for asymmetrics
smooth transition heteroskedastic models, Statistics and Computing, 18(4), 391-408.
[11] Kupiec P. (1995). Techniques for verifying the accuracy of risk measurement models, Journal
of Derivatives, 2, 73-84.
[12] Kwan W., Li W. K. and Li G. (2011). On the threshold hyperbolic GARCH models, Statistics and its Interface, 4(2), 159-166.
[13] Kwan W., Li W. K. and Li G. (2012). On the estimation and diagnostic checking of the
AFRIMA-HYGARCH model, Computational statistics and data Analysis, 56(11), 36323644.
[14] Li M., Li G. and Li W. K. (2011). Score tests for hyperbolic GARCH models, Journal of
Business and Economic Statistics, 29(4), 579-586.
[15] Li M., Li W. K. and Li G. (2015). A new hyperbolic GARCH model, Journal of Econometrics, 189(2), 428-436.
14
[16] McAleer M. and Medeiros M. C. (2008). A multiple regime smooth transition heterogeneous
autoregressive model for long memory and asymmetries, Journal of Econometrics, 147(1),
104-119.
[17] Sarma M., Thomas S. and Shah A.(2003). Selection of value-at-risk models, Journal of
Forecasting, 22(4), 337-358.
[18] Tang T. L. and Shieh S. J. (2006). Long memory in stock index futures markets: A valueat-risk approach, Physica A:Statistical Mechanics and its Applications, 366, 437-448.
[19] Chong E. K. P. and Zak S. H. (2001). An Introduction to Optimization, New York: John
Wiley.
15
| 10math.ST
|
Estimation in the Spiked Wigner Model:
A Short Proof of the Replica Formula
Ahmed El Alaoui∗
Florent Krzakala†
arXiv:1801.01593v1 [cs.IT] 5 Jan 2018
Abstract
We consider the problem of estimating the rank-one perturbation of a Wigner matrix in
a setting of low signal-to-noise ratio. This serves as a simple model for principal component
analysis in high dimensions. The mutual information per variable between the spike and
the observed matrix, or equivalently, the normalized Kullback-Leibler divergence between
the planted and null models are known to converge to the so-called replica-symmetric
formula, the properties of which determine the fundamental limits of estimation in this
model. We provide in this note a short and transparent proof of this formula, based
on simple executions of Gaussian interpolations and standard concentration-of-measure
arguments. The Franz-Parisi potential, that is, the free entropy at a fixed overlap, plays
an important role in our proof. Our proof can be generalized straightforwardly to spiked
tensor models of even order.
1
Introduction
Extracting low-rank information from a data matrix corrupted with noise is a fundamental
statistical task. Spiked random matrix models have attracted considerable attention in statistics, probability and machine learning as rich testbeds for theoretical investigation on this
problem [Joh01, BAP05, Péc06, Péc14]. A basic such model is the spiked Wigner model in
which one observes a rank-one deformation of a Wigner matrix W :
r
λ ∗ ∗⊤
x x + W,
(1)
Y =
N
where Wij = Wji ∼ N (0, 1) and Wii ∼ N (0, σ 2 ) are independent for all 1 ≤ i ≤ j ≤ N .
The spike vector x∗ ∈ RN represents the signal to be recovered, and λ ≥ 0 plays the role of
a Signal-to-Noise Ratio (SNR) parameter. The entries of x∗ come i.i.d. from a (Borel) prior
Px on R with bounded support, so that the scaling in the above model puts the problem in
a high-noise regime where only partial recovery of the spike is possible. A basic statistical
question about this model is for what values of the SNR λ is it possible to estimate the spike
x∗ with non-trivial accuracy? Spectral methods, or more precisely, estimation using the top
eigenvector of Y , are know to succeed above a spectral threshold and fail below [BGN11].
Since the posterior mean is the estimator with minimal mean squared error, this question
boils down to the study of the posterior distribution of x∗ given Y , which by Bayes’ rule, can
be written as
e−H(x) dPx⊗N (x)
,
(2)
d Pλ (x|Y ) = R
e−H(x) dPx⊗N (x)
∗
Department of EECS, UC Berkeley, CA. Email: [email protected]
Laboratoire de Physique Statistique, CNRS, PSL Universités & Ecole Normale Supérieure, Sorbonne Universités et Université Pierre & Marie Curie, Paris, France.
†
1
where H is the Hamiltonian
−H(x) :=
=
X
i<j
X
i<j
r
r
λ 2 2
λ
Yij xi xj −
x x
N
2N i j
(3)
λ
λ
λ 2 2
Wij xi xj + xi xj x∗i x∗j −
x x .
N
N
2N i j
Let us define the free entropy1 of the model as the expected log-partition function (i.e.,
normalizing constant) of the posterior Pλ (·|Y ):
Z
1
E log
e−H(x) dPx⊗N (x),
(4)
FN =
N
The above quantity is known to converge to the Replica-Symmetric (RS) formula φRS defined
as follows: For r ∈ R+ , let
Z
√
r
ψ(r) := Ex∗ ,z log exp
rzx + rxx∗ − x2 dPx (x),
2
where z ∼ N (0, 1), and x∗ ∼ Px . Moreover, define the RS potential
F (λ, q) := ψ(λq) −
λq 2
,
4
and the RS formula
φRS (λ) := sup F (λ, q).
q≥0
Theorem 1 ([KM09, KXZ16, BDM+ 16, DAM16, LM16]). For all λ ≥ 0,
lim FN = φRS (λ).
N →∞
The above statement contains precious statistical information. It can be written in at
least two other equivalent ways, in terms of the mutual information between x∗ and Y :
lim
N →∞
2
1
λ
I(Y , x∗ ) =
EPx [X 2 ] − φRS (λ),
N
4
or, denoting by Pλ the probability distribution of the matrix Y as per (1), in terms of the
Kullback-Liebler divergence between Pλ and P0 :
lim
N →∞
1
DKL (Pλ , P0 ) = φRS (λ).
N
Furthermore, the point q ∗ (λ) achieving the maximum in the RS formula (which can be shown
to be unique and finite for almost every λ) can be interpreted as the best overlap any estimator
b ) can have with the spike x∗ . Indeed, the overlap of a draw x from the posterior Pλ (·|Y )
θ(Y
with x∗ concentrates about q ∗ (λ). See [BDM+ 16, LM16, EKJ17] for various forms of this
statement.
1
The term free energy is also used, although the physics convention requires to put a minus sign in front of
the expression in this case.
2
Comment on existing proofs. The proof of the lower bound lim inf FN ≥ φRS (λ) relies on
an application of Guerra’s interpolation method, and is fairly short and transparent [KXZ16].
Available proofs of the converse bound lim sup FN ≤ φRS (λ) (and overlap concentration) are
on the other hand highly involved. Barbier et al. [BDM+ 16] and Deshpande et al. [DAM16]
adopt an algorithmic approach: they analyze an Approximate Message Passing (AMP) procedure and show that the produced estimator asymptotically achieves an overlap of q ∗ (λ)
with the spike. Thus the posterior mean, being the optimal estimator, must also achieve the
same overlap. This allows to prove overlap convergence and thus show the converse bound.
A difficulty one has to overcome with this method is that AMP (and supposedly any other
algorithm) may fail to achieve the optimal overlap in the presence of first-order phase transitions, which trap the algorithm in a bad local optimum of the RS potential. Spatial coupling,
an idea from coding theory, is used in [BDM+ 16] to overcome this problem. Lelarge and
Miolane [LM16] on the other hand use the Aizenman-Sims-Starr scheme [ASS03], a relative of
the cavity method developed within spin-glass theory, to prove the upper bound. Barbier and
Macris [BM17] prove the upper bound via a “stochastic” version of the interpolation method.
Recently, the optimal rate of convergence and constant order corrections to the RS formula
were proved in [EKJ17] using a variant of the cavity method. All the current approaches
(perhaps to a lesser extent for [BM17]) require the execution of long and technical arguments.
In this note, we show that the upper bound in Theorem 1 admits a fairly simple proof based
on the same interpolation idea that yielded the lower bound, combined with an application of
the Laplace method and concentration of measure. The main idea is to consider a version of
the free entropy (4) of a subsystem of configurations x having a fixed overlap with the spike
x∗ . We then proceed by applying the Guerra bound and optimize over this free parameter
to obtain an upper bound in the form of a saddle (max-min) formula. A small extra effort
is needed to show that the latter is another representation of the RS formula. The idea of
restricting the overlap dates back to Franz and Parisi [FP95]. The free entropy at fixed overlap
bears the name of the Franz-Parisi potential. Our proof thus hinges on a upper bound on
this potential, which is may be of independent interest. We first start by presenting the proof
of the lower bound, which is a starting point for our argument. We present the proof in the
case σ = 0. This is only done to keep the displays concise; recovering the general case is
straightforward since the diagonal has vanishing contribution. Finally, the method can be
easily generalized to all spiked tensor models of even order [RM14], thus recovering the main
results of [LML+ 17].
2
Proof of Theorem 1
Let t ∈ [0, 1] and consider an interpolating Hamiltonian
r
X tλ
tλ
tλ 2 2
Wij xi xj + xi x∗i xj x∗j −
x x
−Ht(x) :=
N
N
2N i j
(5)
i<j
+
N p
X
(1 − t)r 2
(1 − t)rzi xi + (1 − t)rxi x∗i −
xi ,
2
i=1
where the zi ’s are i.i.d. standard Gaussian r.v.’s independent of everything else. For f :
(RN )n+1 7→ R, we define the Gibbs average of f as
R
Q
(l)
D
E
f (x(1) , · · · , x(n) , x∗ ) nl=1 e−Ht (x ) dPx⊗N (x(l) )
(1)
(n)
∗
.
(6)
f (x , · · · , x , x ) :=
R Qn
−Ht (x(l) ) dP ⊗N (x(l) )
t
x
l=1 e
3
This is the average of f with respect to the posterior distribution of n copies x(1) , · · · , x(n) of
x∗ given the augmented set of observations
q
(
x∗ x∗ + Wij , 1 ≤ i ≤ j ≤ N,
Yij = tλ
(7)
pN i j ∗
yi = (1 − t)rxi + zi , 1 ≤ i ≤ N.
The variables x(l) , l = 1 · · · , n are called replicas, and are interpreted as random variables
independently drawn from the posterior. When n = 1 we simply write f (x, x∗ ) instead of
f (x(1) , x∗ ). We shall denote the overlaps between two replicas as follows: for l, l′ = 1, · · · , n, ∗,
we let
N
′
1 X (l) (l′ )
Rl,l′ := x(l) · x(l ) =
xi xi .
N
i=1
A simple consequence of Bayes’ rule is that the n+1-tuples (x(1) , · · · , x(n+1) ) and (x(1) , · · · ,
have the same law under Eh·it (see Proposition 16 in [LM16]). This bears the name
of the Nishimori property in the spin glass literature [Nis01].
x(n) , x∗ )
2.1
The lower bound
Reproducing the argument of [KXZ16], we prove using Guerra’s interpolation argument [Gue01]
and the Nishimori property that
1
.
FN ≥ φRS (λ) − O
N
We let r = λq in the definition of Ht and let
Z
1
E log e−Ht (x) dPx⊗N (x).
ϕ(t) :=
N
A short calculation based on Gaussian integration by parts shows that
N
λ
λ 2
λ X
(1) 2 (2) 2
′
2
E xi xi
ϕ (t) = − E (R1,2 − q) t + q +
4
4
4N 2
t
i=1
λ
+ E (R1,∗ − q)2
2
N
λ X D 2 ∗2E
λ 2
E xi xi
,
− q −
t
2
2N 2
t
i=1
By the Nishimori property, the expressions involving the pairs (x, x∗ ) on the one hand and
(x(1) , x(2) ) on the other in the brackets are equal. We then obtain
E
D
λ
λ
λ
ϕ′ (t) = E (R1,∗ − q)2 t − q 2 −
E xN 2 x∗N 2 .
4
4
4N
t
Observe that the last term is O(1/N ) since the variables xN are bounded. Moreover, the first
term is always non-negative so we obtain
K
λ
ϕ′ (t) ≥ − q 2 − .
4
N
Since ϕ(1) = FN and ϕ(0) = ψ(λq), integrating over t, we obtain for all q ≥ 0
FN ≥ F (λ, q) −
and this yields the lower bound.
4
K
,
N
2.2
The upper bound
We prove the converse bound
FN ≤ φRS (λ) + O
log N
√
.
N
We introduce the Franz-Parisi potential [FP95]. For x∗ ∈ RN fixed, m ∈ R and ǫ > 0 we
define
Z
1
∗
Φǫ (m, x ) :=
E log 1{R1,∗ ∈ [m, m + ǫ)}e−H(x) dPx⊗N (x),
(8)
N
where the expectation is over W . This is the free entropy of a subsystem of configurations having an overlap close to a fixed value m with a planted signal x∗ . It is clear that
Ex∗ Φǫ (m, x∗ ) ≤ FN . We will argue via the Laplace method and concentration of measure
that supm∈R Ex∗ Φǫ (m, x∗ ) ≈ FN , then use Guerra’s interpolation to upper bound Φǫ (m, x∗ )
(notice that this method yielded a lower bound on FN due to the Nishimori property). Let
us define a bit of more notation. For r ∈ R+ , s ∈ R, let
Z
√
r
b
rzx + sx − x2 dPx (x),
ψ(r, s) := Ez log exp
2
b sx∗ ) where x∗ ∼ Px . Moreover, let
where z ∼ N (0, 1), and ψ(r, s) = Ex∗ ψ(r,
N
1 Xb
λm2 λq 2
∗
b
F (λ, m, q, x ) :=
ψ(λq, λmx∗i ) −
+
,
N
2
4
i=1
and similarly define F (λ, m, q) = Ex∗ Fb(λ, m, q, x∗ ) = ψ(λq, λm) −
λm2
2
+
λq 2
4 .
Proposition 2 (Laplace method). There exist K > 0 such that for all ǫ > 0, we have
h
i log(K/ǫ)
max Φǫ (lǫ, x∗ ) + √
FN ≤ Ex∗
.
l∈Z,|l|≤K/ǫ
N
Now we upper bound Φǫ in terms of Fb:
Proposition 3 (Interpolation upper bound). There exist K > 0 depending on λ ≥ 0 such
that for all m ∈ R and ǫ > 0 we have
λ
K
Φǫ (m, x∗ ) ≤ inf Fb(λ, m, q, x∗ ) + ǫ2 + .
q≥0
2
N
Remark: This simple upper bound on the Franz-Parisi potential—which may be of independent interest—can be straightforwardly generalized to spiked tensor models of even order.
Indeed, as will be apparent from the proof in the present matrix case, a crucial step in obtaining the inequality is the positivity of a certain hard-to-control remainder term. Tensor
models of even order enjoy a convexity property that ensures the positivity of this remainder.
From Propositions 2 and 3, an upper bound on FN in the form of a saddle formula begins
to emerge. For a fixed m ∈ R let q̄ = q̄(λ, m) be any minimizer of q 7→ F (λ, m, q) on R+ . (By
differentiating F , we can check that q̄ is bounded uniformly in m.) Then we have
i λ
h
log(K/ǫ)
∗
b
∗
FN ≤ Ex
max F (λ, m, q̄(λ, m), x ) + ǫ2 + √
.
(9)
m=lǫ
2
N
|l|≤K/ǫ
At this point we need to push the expectation inside the supremum. This will be done using
a concentration argument.
5
Lemma 4. There exists K > 0 such that for all λ ≥ 0, m ∈ R, q ≥ 0 and t ≥ 0,
Pr∗
x
Nt2
−
Fb(λ, m, q, x∗ ) − F (λ, m, q) ≥ t ≤ 2e λ2 K(|m|+q)2 .
It is a routine computation to deduce from Lemma 4 (and boundedness of both m and
q) that the expected supremum is bounded by the supremum of the expectation plus a small
entropy term (the full details of a similar argument are given in the proof of Proposition 2):
K log(K/ǫ)
√
.
E sup Fb (λ, m, q̄(λ, m), x∗ ) ≤ sup F (λ, m, q̄(λ, m)) +
m=lǫ
m=lǫ
N
|l|≤K/ǫ
|l|≤K/ǫ
Since q̄ is a minimizer of F , it follows from (9) that
FN ≤ sup inf F (λ, m, q) +
m∈R q≥0
λ 2 K log(K/ǫ)
√
ǫ +
.
2
N
(10)
We now let ǫ = N −1/4 , and conclude by noticing that the above saddle formula is another
expression for φRS :
Proposition 5.
φRS (λ) = sup inf F (λ, m, q).
m∈R q≥0
Proof. One inequality follows from (10) and the lower bound FN ≥ φRS (λ) − oN (1). For the
converse inequality, we notice that for all m ∈ R
inf F (λ, m, q) ≤ F (λ, m, |m|) = ψ(λ|m|, λm) −
q≥0
λ
|m|2 .
4
Now we use the fact that the function ψ is largest when its second arguments is positive:
Lemma 6 (Proposition 23 [EKJ17]). For all r ≥ 0 we have
ψ(r, −r) ≤ ψ(r, r).
This implies
inf F (λ, m, q) ≤ F (λ, |m|) −
q≥0
Taking the supremum over m yields the converse bound.
λ
|m|2 .
4
Proof of Proposition 2. Let ǫ > 0. Since the prior Px has bounded support, we can grid
the set of the overlap values R1,∗ by 2K/ǫ many intervals of size ǫ for some K > 0. This
allows the following discretization, where l runs over the finite range {−K/ǫ, · · · , K/ǫ}:
XZ
1
E log
FN =
1{R1,∗ ∈ [lǫ, (l + 1)ǫ)}e−H(x) dPx⊗N (x)
N
l
Z
2K
1
E log
max 1{R1,∗ ∈ [lǫ, (l + 1)ǫ)}e−H(x) dPx⊗N (x)
≤
l
N
ǫ
Z
1
log(2K/ǫ)
=
E max log 1{R1,∗ ∈ [lǫ, (l + 1)ǫ)}e−H(x) dPx⊗N (x) +
.
(11)
l
N
N
6
In the above, E is w.r.t. both W and x∗ . We use concentration of measure to push the
expectation over W to the left of the maximum. Let
Z
Zl := 1{R1,∗ ∈ [lǫ, (l + 1)ǫ)}e−H(x) dPx⊗N (x).
We show that each term Xl = N1 log Zl concentrates about its expectation (in the randomness
of W ). Let E′ denote the expectation w.r.t. W .
Lemma 7. There exists a constant K > 0 such that for all γ ≥ 0 and all l,
′
Kγ
2
E′ eγ(Xl −E [Xl ]) ≤ √ eKγ /N .
N
Therefore, the expectation of the maximum concentrates as well:
1
′
′
′
′
E max(Xl − E [Xl ]) ≤ log E exp γ max(Xl − E [Xl ])
l
l
γ
′
1
= log E′ max eγ(Xl −E [Xl ])
l
γ
X
′
1
eγ(Xl −E [Xl ])
≤ log E′
γ
l
1
2K γK γ 2 K/N
√ e
≤ log
γ
ǫ
N
log(2K/ǫ)
1
γK
γK
=
+ log √ +
.
γ
γ
N
N
√
We set γ = N and obtain
E′ max(Xl − E′ [Xl ]) ≤
l
log(K/ǫ)
√
.
N
Therefore, plugging the above estimates into (11), we obtain
log(K/ǫ) log(K/ǫ)
√
+
l
N
N
log(K/ǫ)
≤ Ex∗ max Φǫ (lǫ, x∗ ) + 2 √
.
l
N
FN ≤ Ex∗ max E′ Xl +
Proof of Proposition 3. Let t ∈ [0, 1] and consider a slightly modified interpolating Hamiltonian that has two parameters r = λq ≥ 0 and s = λm ∈ R:
r
X tλ
tλ
tλ 2 2
Wij xi xj + xi x∗i xj x∗j −
x x
(12)
−Ht (x) :=
N
N
2N i j
i<j
+
N p
X
i=1
(1 − t)rzi xi + (1 − t)sxi x∗i −
(1 − t)r 2
xi ,
2
where the zi ’s are i.i.d. standard Gaussian r.v.’s independent of everything else. Let
Z
1
E log 1{R1,∗ ∈ [m, m + ǫ)}e−Ht (x) dPx⊗N (x),
ϕ(t) :=
N
7
where E is over the Gaussian disorder W and z (x∗ is fixed). Let h·it be the corresponding
Gibbs average, similarly to (6). By differentiation and Gaussian integration by parts,
λ
ϕ (t) = − E (R1,2 − q)2
4
′
+
λ
E (R1,∗ − m)2
2
N
λ 2
λ X D (1) (2) 2 E
E (xi xi )
+ q +
t
4
4N 2
t
t
−
i=1
N
X
λ
λ 2
m −
2
2N 2
m)2
E (xi x∗i )2
t
,
i=1
ǫ2 .
Notice that by the overlap restriction, E (R1,∗ −
≤
Moreover, the last terms in the
t
first and second lines in the above are of order 1/N since the variables xi are bounded. Next,
since E (R1,2 − q)2 t has non-negative sign, we can ignore it and obtain an upper bound:
λ
λ
λ
K
ϕ′ (t) ≤ − m2 + q 2 + ǫ2 + .
2
4
2
N
Integrating over t, we obtain
λ
λ
K
λ
Φǫ (m, x∗ ) ≤ − m2 + q 2 + ǫ2 + ϕ(0) + .
2
4
2
N
Now we use a trivial upper bound on ϕ(0):
Z
1
ϕ(0) =
E log 1{R1,∗ ∈ [m, m + ǫ)}e−H0 (x) dPx⊗N (x)
N
Z
1
≤
E log e−H0 (x) dPx⊗N (x)
N
N
1 Xb
ψ(λq, λmx∗i ).
=
N
i=1
Hence,
λ
K
Φǫ (m, x∗ ) ≤ Fb(λ, m, q, x∗ ) + ǫ2 + .
2
N
Proof of lemma 4. The random part of Fb(λ, m, q̄, x∗ ) is the average of i.i.d. terms
b 0) = 0, where K
b sx∗ ) ≤ K 2 /2 and ψ(0,
b sx∗ ) ≤ K 2 , ∂r ψ(r,
b
ψ(λq,
λmx∗i ). Since ∂s ψ(r,
b sx∗ ) ≤ K 2 (r/2 + |s|). For bounded r and s,
is a bound on the support of Px , we have ψ(r,
the claim follows from concentration of the average of i.i.d. bounded r.v.’s.
Proof
q of Lemma 7. We notice that Xl seen as a function of W is Lipschitz with constant
K Nλ . By Gaussian concentration of Lipschitz functions, there exist a constant K depending
only on λ such that for all t ≥ 0,
2
Pr Xl − E′ Xl ≥ t ≤ e−N t /K .
Then we conclude by means of the identity
Z +∞
′
Pr(Xl − E′ [Xl ] ≥ t) eγt dt,
E′ eγ(Xl −E [Xl ]) = γ
−∞
and integrate the tail.
Acknowledgments. Florent Krzakala acknowledges funding from the ERC under the European Union 7th Framework Programme Grant Agreement 307087-SPARCS.
8
References
[ASS03]
Michael Aizenman, Robert Sims, and Shannon L Starr. Extended variational principle for
the Sherrington-Kirkpatrick spin-glass model. Physical Review B, 68(21):214403, 2003.
[BAP05]
Jinho Baik, Gérard Ben Arous, and Sandrine Péché. Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices. Annals of Probability, 33(5):1643–
1697, 2005.
[BDM+ 16] Jean Barbier, Mohamad Dia, Nicolas Macris, Florent Krzakala, Thibault Lesieur, and
Lenka Zdeborová. Mutual information for symmetric rank-one matrix estimation: A proof
of the replica formula. In Advances in Neural Information Processing Systems (NIPS),
pages 424–432, 2016.
[BGN11]
Florent Benaych-Georges and Raj Rao Nadakuditi. The eigenvalues and eigenvectors
of finite, low rank perturbations of large random matrices. Advances in Mathematics,
227(1):494–521, 2011.
[BM17]
Jean Barbier and Nicolas Macris. The stochastic interpolation method: A simple scheme
to prove replica formulas in bayesian inference. arXiv preprint arXiv:1705.02780, 2017.
[DAM16]
Yash Deshpande, Emmanuel Abbé, and Andrea Montanari. Asymptotic mutual information for the binary stochastic block model. In IEEE International Symposium on Information Theory (ISIT), pages 185–189. IEEE, 2016.
[EKJ17]
Ahmed El Alaoui, Florent Krzakala, and Michael I Jordan. Finite size corrections and
likelihood ratio fluctuations in the spiked Wigner model. arXiv preprint arXiv:1710.02903,
2017.
[FP95]
Silvio Franz and Giorgio Parisi. Recipes for metastable states in spin glasses. Journal de
Physique I, 5(11):1401–1415, 1995.
[Gue01]
Francesco Guerra. Sum rules for the free energy in the mean field spin glass model. Fields
Institute Communications, 30:161–170, 2001.
[Joh01]
Iain M Johnstone. On the distribution of the largest eigenvalue in principal components
analysis. Annals of Statistics, pages 295–327, 2001.
[KM09]
Satish Babu Korada and Nicolas Macris. Exact solution of the gauge symmetric p-spin
glass model on a complete graph. Journal of Statistical Physics, 136(2):205–230, 2009.
[KXZ16]
Florent Krzakala, Jiaming Xu, and Lenka Zdeborová. Mutual information in rank-one
matrix estimation. In Information Theory Workshop (ITW), pages 71–75. IEEE, 2016.
[LM16]
Marc Lelarge and Léo Miolane. Fundamental limits of symmetric low-rank matrix estimation. arXiv preprint arXiv:1611.03888, 2016.
[LML+ 17] Thibault Lesieur, Léo Miolane, Marc Lelarge, Florent Krzakala, and Lenka Zdeborová.
Statistical and computational phase transitions in spiked tensor estimation. In IEEE
International Symposium on Information Theory (ISIT), pages 511–515, June 2017.
[Nis01]
Hidetoshi Nishimori. Statistical physics of spin glasses and information processing: an
introduction, volume 111. Clarendon Press, 2001.
[Péc06]
Sandrine Péché. The largest eigenvalue of small rank perturbations of Hermitian random
matrices. Probability Theory and Related Fields, 134(1):127–173, 2006.
[Péc14]
Sandrine Péché. Deformed ensembles of random matrices. In Proceedings of the International Congress of Mathematicians, Seoul, pages 1059–1174. ICM, 2014.
[RM14]
Emile Richard and Andrea Montanari. A statistical model for tensor PCA. In Advances
in Neural Information Processing Systems (NIPS), pages 2897–2905, 2014.
9
| 10math.ST
|
Published as a conference paper at ICLR 2018
E SPRESSO : E FFICIENT F ORWARD P ROPAGATION FOR
B INARY D EEP N EURAL N ETWORKS
arXiv:1705.07175v2 [cs.DC] 7 Mar 2018
Fabrizio Pedersoli
University of Victoria
[email protected]
George Tzanetakis
University of Victoria
[email protected]
Andrea Tagliasacchi
University of Victoria
[email protected]
A BSTRACT
There are many applications scenarios for which the computational performance
and memory footprint of the prediction phase of Deep Neural Networks (DNNs)
need to be optimized. Binary Deep Neural Networks (BDNNs) have been shown
to be an effective way of achieving this objective. In this paper, we show how
Convolutional Neural Networks (CNNs) can be implemented using binary representations. Espresso is a compact, yet powerful library written in C/CUDA that
features all the functionalities required for the forward propagation of CNNs, in
a binary file less than 400KB, without any external dependencies. Although it
is mainly designed to take advantage of massive GPU parallelism, Espresso also
provides an equivalent CPU implementation for CNNs. Espresso provides special convolutional and dense layers for BCNNs, leveraging bit-packing and bitwise computations for efficient execution. These techniques provide a speed-up
of matrix-multiplication routines, and at the same time, reduce memory usage
when storing parameters and activations. We experimentally show that Espresso
is significantly faster than existing implementations of optimized binary neural
networks (≈ 2 orders of magnitude). Espresso is released under the Apache 2.0
license and is available at http://github.com/fpeder/espresso.
1
I NTRODUCTION
Convolutional Neural Networks have revolutionized computer vision, pushing the task of object
recognition beyond human capabilities (Krizhevsky et al., 2012; Simonyan & Zisserman, 2014;
Szegedy et al., 2015). Deep Neural Networks (DNN), have also been successfully applied in other
fields, such as speech recognition (Graves et al., 2013; Hinton et al., 2012) and automated translation (Bahdanau et al., 2014; Sutskever et al., 2014). Despite achieving impressive classification
accuracy results, DNNs require too much memory and power to be used effectively on embedded or
low-power devices. Many networks consume a considerable amount of memory. Memory remains
a very limited resource on mobile platforms making harder the usage of trained DNNs 1 . Even
when memory is not an issue, DNNs remain very computationally intensive, and can quickly drain
the battery. Reducing the computational load does not only improve energy efficiency, but can also
enable further applications. For example, when processing real-time object classification on mobile,
being able to perform faster predictions frees up computational resources that can be spent on tasks
such as speech recognition and analysis. Therefore, there is a substantial interest in reducing the
computational and memory requirements of DNNs.
Efficient deep neural networks One way to achieve this target is to use specialized hardware for
DNNs. Another strategy is to reduce the network’s memory footprint and associated computation,
hence increasing its efficiency. Such solutions are preferable as they can be implemented in software without requiring specialized hardware. In our research we follow the software approach, and
focus our attention to quantized networks. In this case, the parameters are stored as “small” integers
1
for example, the popular AlexNet (Krizhevsky et al., 2012) and VGG (Simonyan & Zisserman, 2014) architectures consume respectively ≈ 250 MB and ≈ 520 MB
1
Published as a conference paper at ICLR 2018
(typically less than 8-bit) instead of single precision floating point numbers (32-bit). In particular,
we consider the binary deep neural networks (BDNN) proposed by Hubara et al. (2016) where parameters and activations are 1-bit integers: {−1, +1}. At the expense of a relatively small decrease
in accuracy, BDNNs can considerably reduce memory usage, and result in faster execution time (i.e.
forward propagation). Further, note that potential hardware implementation of BDNNs would also
be cheaper due to the reduced number of required FPUs. While these results are highly promising,
currently only proof-of-concept implementations of BinaryNets have been published (Hubara et al.,
2016). Therefore, the availability of a flexible end-to-end framework, with particular emphasis
placed on computational efficiency, can enable further research on BDNNs, as well as its application to practical scenarios.
Contributions With Espresso we provide an optimized framework for BDNNs capable of achieving state-of-the-art run-time performance with minimal memory footprint while being numerical
equivalent to their non-optimized binary counterpart. Espresso provides a complete optimized
framework for BDNNs supporting both the dense and the convolutional layer. Current state-ofthe-art optimized BDNNs implementations are limited to the fully connected layer, with the serious
drawback of not being able to run optimized state-of-art convolutional BDNNs (BCNNs). While
our work is a necessary stepping stone towards optimization of training routines, in this paper we
focus on the optimization of forward-propagation (i.e. testing), rather than back-propagation (i.e.
training). Espresso is designed to have no external dependencies. This not only results in a highly
optimized implementation of BDNNs, but also substantially simplifies its deployment in practical
applications, such as those executing on mobile or embedded devices.
2
R ELATED WORK
Improving the performance of DNNs can be achieved at either the hardware or software level. At
the hardware level, chipsets that are dedicated to DNN execution can outperform general-purpose
CPUs/GPUs (Jouppi, 2017; Han et al., 2016). At software level one approach is to design simpler
architectures, in terms of overall floating point operations, that can offer the same accuracy as the
original model (Iandola et al., 2016). Another approach is to prune the weights (Guo et al., 2016), or
even entire filters (Li et al., 2016), that have low impact on the activations such that a simpler model
can be derived. These simplified models can be further compressed by weight sharing (Han et al.,
2015). Finally instead of removing connections, another approach is to quantize the network weights
(Courbariaux et al., 2014) such that computations can be executed more efficiently.
Quantized networks In quantized networks, the objective is to train DNNs whose (quantized) weights do not significantly impact the network’s classification accuracy. For example,
Courbariaux et al. (2014) show that 10-bits are enough for Maxout Networks, and how more efficient multiplications can be performed with fixed-point arithmetic. Continuing this trend, more
aggressive quantization schemes, up to ternary (Zhu et al., 2016), have also been studied.
Binary Deep Neural Networks (BDNN) Recently, Courbariaux et al. (2015) showed that
a network with binary {−1, +1} weights can achieve near state-of-the-art results on several
standard datasets. Binary DNNs (BDNNs) were shown to perform effectively on datasets
with relatively small images, such as the permutation-invariant MNIST (LeCun et al., 1998),
CIFAR-10 (Krizhevsky et al., 2009) and SVHN (Netzer et al., 2011). Recently, Rastegari et al.
(2016) show that binarized CNNs can perform well even on massive datasets such as ImageNet (Deng et al., 2009) using binarized versions of well-known DNN architectures such as
AlexNet (Krizhevsky et al., 2012), ResNet-18 (He et al., 2016), and GoogLenet (Szegedy et al.,
2015). Similarly interesting results can be achieved by binarizing both DNN weights and activations as showed by Hubara et al. (2016). In this work, the authors introduce BinaryNet, a technique
to effectively train DNNs where both weights and activations are constrained to {−1, +1}. BinaryNet achieves nearly state-of-the-art accuracy for MLP training on MNIST and CNN training on
CIFAR-10. The authors also propose a binary optimized implementation of matrix multiplication
which result in 7× faster performance than the base-line non optimized implementation, and, almost
2× faster than Theano (Bergstra et al., 2010). Their core contributions, namely to replace Floatingpoint Multiply and Add operations (FMAs) with XNORs and bit-counts, represent the cornerstone
over which we build our research.
2
Published as a conference paper at ICLR 2018
3
T HE E SPRESSO F RAMEWORK
Espresso provides the user with the necessary tools for executing forward-propagation of DNNs,
with particular emphasis placed on convolutional neural networks due to their ubiquitousness in
computer vision applications. As the complexity of these networks is cubic to the size of the problem, they are less memory efficient and more computationally intensive than traditional machinelearning algorithms. Identifying the memory and computational bottlenecks of DNNs is therefore
essential to enable their practical application. In particular, our primary focus is GPU-optimized
BDNN architectures, which we refer to as GPU opt , but we also support the equivalent floatingpoint counterparts on heterogeneous architectures, which we refer to as CPU and GPU . The CPU
and GPU implementations of Espresso do not feature binary optimizations because the data is encoded as single precision floating point numbers. However they still utilize an optimized library for
matrix multiplication.
Hybrid DNNs The Espresso’s implementations of tensors and layers come in three variants
{CPU, GPU, GPU opt }. A CPU-tensor is allocated in CPU memory, and is processed on the CPU
using sequential code. A GPU-tensor is allocated on GPU main memory and is processed by CUDA
kernels. Espresso provides functions for converting tensors and layers from one variant to the other,
and different variants can also be interconnected with each other. Consequently, Espresso enables
the design of hybrid DNNs consisting of a combination of {CPU, GPU, GPU opt } layers.
The computational bottleneck: dot products Dense linear algebra is at the heart of deeplearning as deep networks can be viewed as a composition of matrix-matrix, matrix-vector and
elementwise matrix-matrix or vector-vector multiplications. The implementation of these dense linear algebra operations relies heavily on the efficient computation of the dot-product. The execution
of this operator consists of (single precision) Floating-point Multiply and Add (FMA) operations. In
modern architectures, floating-point multiplications executing on the FPU dominate the complexity
of FMAs, and BDNNs address these concerns by replacing FMAs with simpler bitwise operations;
see Section 4.
Technical highlights The superior computational performance of Espresso derives from three
main technical contributions: (1) the use of bit-packing in network layers, (2) better memory layout
and management, and (3) the use of custom optimized CUDA kernels. Through the use of bitpacked layers, Espresso can execute a forward operation without the need for expensive memory
re-arrangements employed by existing implementations. As dynamic memory allocation on GPUs
is a performance bottleneck, Espresso implements a custom memory allocator that pre-allocates
memory at start-up, and replaces the traditional malloc and free system calls. Finally, matrix multiplications are performed with CUDA kernels that have been adapted to bit-packing, and only resort
to XNORs and bit-counts.
4
B INARY D EEP N EURAL N ETWORKS (BDNN S )
In this section, we overview the fundamental characteristics of BDNNs (Hubara et al., 2016) that
inform the basics of Espresso’s design. In BDNNs, computationally intensive FMA operations
are replaced by XNOR (for multiplications) and bit-count (for additions), enabling significant computational speed-ups. In particular, XNOR is a simpler machine instruction compared to floating
point multiplication, and therefore achieves much higher throughput on many architectures. More
importantly, a single XNOR step can execute multiple 64-bit wide blocks of dot-products, further
increasing the overall computational efficiency. In what follows, we describe how a network is binarized, detail a compressed memory layout enabling efficient execution of dot-products, show how to
re-interpret input data to allow execution on fixed-precision input (e.g. images), and provide a few
notes regarding the training procedure.
4.1
N ETWORK
BINARIZATION
A BDNN is composed of a sequence of k = 1, . . . , L layers whose weights Wkb and activations abk
are binarized to the values {−1, +1}. The superscript b in the notation indicates binary quantities.
3
Published as a conference paper at ICLR 2018
Weights and activations are {−1, +1}, but at the hardware level they must be encoded as {0, 1}.
Our convention is to encode −1 → 0 and +1 → 1. Amongst many possible choices, e.g. stochastic binarization (Courbariaux et al., 2015), we employ the following activation function due to its
efficient implementation:
+1 x ≥ 0
xb = sign(x) =
(1)
−1 otherwise
4.2
B IT- PACKING
The weights of a BDNN can be stored in the bits of a 64-bit word. One immediate advantage of
bit-packing is to drastically reduce the memory usage by a 32× factor. An even more significant
advantage is the ability to process multiple values at the same time using registers. This is particularly useful for dot-products: with bit-packing we can compute a dot-product of 64 element vectors
by using just one XNOR and one bit-count. Furthermore, modern computer architectures provide
a hardware instruction for counting the number of bits set to 1 in a given word. Assuming binary
vectors a, b ∈ B1×N where N is a multiple of 64, the dot-product is then equivalent to:
N/64
X
bitcount(XNOR(ai , bi )) ≪ 1 , a ⊙ b
a·b ≡ N −
(2)
i=1
where ≪ represents the bit-shift operator. This simple computation becomes the building block of
optimized BDNNs as binary matrix-matrix or matrix-vector operations are computed in this fashion.
4.3
I NPUT
DATA BINARIZATION
BDNNs require binary input data, which is not typically available at the first layer of the network.
However, the input data usually comes in a fixed precision format (e.g. 8-bit/channel in RGB images). Therefore, the optimized computation of dot-products can still be applied if we split the input
data according to bit-planes, and then sum back each contribution according to the corresponding
weight. For instance, if with hain we indicate the n-th bit of a fixed precision vector, and with i the
corresponding bit-plane, we obtain:
a·b≡
n−1
X
2i ha ⊙ bii
(3)
i=0
4.4
T RAINING
When training a BDNN, it is important to note that the gradient is computed with the binary weights,
but is accumulated with floating point precision (Hubara et al., 2016). That is because the optimizer
needs sufficient precision to make a reliable update. In addition, the derivative of the sign function,
which is zero almost everywhere, cannot be used for back-propagation. To overcome these issues,
the straight-through estimator (Bengio et al., 2013) is employed, where 1 is back-propagated if the
floating point argument |x| ≤ 1, and 0 otherwise. Finally, during training weights are clipped to
[−1, 1] to avoid a large growth of the floating point weights that would not have an impact on the
binary weights.
5
E SPRESSO
ARCHITECTURE
The principal components of our framework are tensors, layers and the network. These components
are organized as a hierarchy. Tensors are n dimensional matrices used for storing inputs, weights
and activations (outputs). A layer processes an input tensor and produces an output tensor, while a
network consists of a concatenation of layers.
5.1
T ENSORS
In Espresso, each element of a tensor A ∈ RM×N ×L is identified by the triplet m, n, l, where
m ∈ [0, M ) indicates the row, n ∈ [0, N ) indicates the column, and l ∈ [0, L) indicates the channel.
4
Published as a conference paper at ICLR 2018
A tensor is stored in memory using row-major order with interleaved channels. Therefore, according
to this layout, the element Am,n,l is found at position (mN + n)L + l in linear memory.
A0, 0, :
A0, :, :
A1, :, :
We use the notation Am,n,: , to indicate all the channels of the (m, n)-th element. Using the same
storing scheme Espresso also defines bit-packed tensors for GPU opt implementations but with the
following changes to further increase its performance. Bit-packing is performed according to the
number of channels: when L > 1 bit-packing is done along the l dimension; when L = 1 bitpacking is done along the n dimension. For convolutional layers this packing direction enables
n
l
m
convolutional
n
m
M ×N ×1
dense
M ×N ×L
M ×N ×
M×
N
64
×1
L
64
efficient memory access when unrolling/lifting a tensor, which would have not been possible if
either m or n had been chosen instead. More specifically, this layout is optimal for retrieving a
pixel neighborhood as needed by convolution without requiring the layout to be changed. Further,
typically a large number of filters are used resulting in an increase of tensor dimension in the l
direction, while the m and n dimensions are progressively shrunk by pooling layers. For other layer
types, n is the most efficient packing direction, as neurons are stored along rows and their number
decreases as we move toward later stages in the network.
5.2
L AYERS
Espresso provides the following layer types: Input, Convolutional, Pooling, Dense (i.e. fully connected) and Batch-normalization. Each layer is characterized by its size, tensor parameters and
output. The Espresso API defines for each layer a forward function that computes the output of a
layer given an input tensor, and a function for applying non-linearity to the outputs of convolutional
and dense layers. Moreover, the convolutional layer features additional functions for pooling and
unrolling.
Input tensor
Lowered matrix
Output matrix
Filter matrix
Output tensor
(m,n)
M1
mN2 + n
h
w
Am:m+h , n:n+w , :
F: ,: ,l
×
=
Om, n , :
M2
(m,n)
L2
L1
N2
N1
L2
unrolling M N
2 2
lifting
whL1
Figure 1: unrolling and lifting operations for CNN layers
5
Published as a conference paper at ICLR 2018
Convolutional layers In our framework, 2D convolutions are computed through matrix multiplications – an operation involving a very high reuse of data. For both CPU and GPU , this computation
is performed by sectioning data in amounts that are cache-friendly (Dongarra et al., 1990), resulting
in implementations attaining close to peak computational performance. However, in order to express convolution as matrix multiplication we need to re-organize the input memory appropriately.
This is achieved through the unrolling procedure; see Figure 1. It consists of transforming a tensor
into a matrix where each row is formed by unrolling the tensor data contained in each convolution
sliding volume. The unrolled matrix is then multiplied by the filter matrix. Finally, the result of the
convolution is reordered back to tensor by using the lifting procedure. In Espresso we do need to
manually lift the convolution result in order to undo the unrolling: thanks to our tensor representation this happens automatically and at zero cost. Espresso provides CUDA kernels for the unrolling
and pooling of tensors for both GPU and GPU opt implementations.
Efficient Matrix multiplication Matrix-vector multiplications are fundamental operations of both
dense and CNN layers. For the CPU architecture, we use the OpenBLAS library (Xianyi et al.) to
implement these operations. For GPU and GPU opt architectures, the CUDA kernels are based on
MAGMA(sgemm) (Abdelfattah et al., 2016), modified to make it compatible with our binary data
representation. These kernels for matrix multiplication feature register blocking optimization: since
the introduction of Fermi architectures the number of registers have been increased, while register
access latency has been substantially reduced compared to shared-memory; hence caching at the
register-memory level results in considerably faster throughput (Nath et al., 2010). Espresso first
fetches the tiles of the two matrices into shared-memory and then process sub-tiles using registers.
In the GPU opt variant, we modify the code by replacing blocks of 64 (or blocks of 32 for GPU opt
32) single precision multiply and add (FMA) operations with XNOR and bit-count using packed
tensors. We also re-tune the kernel block size parameters for improving the performance on reduced
size matrices.
Zero-padding for convolutions Typical CNN implementations apply a tensor convolution in a
“same” configuration, where the sizes of input and output tensors matches. This is achieved by zeropadding input tensors, but in convolutional GPU opt layers the zero-padding of the input introduces
the side-effect of making the data ternary {−1, 0, +1}. We deal with this problem by treating the
data as if it was binary (zero is considered a minus one) and fix the results of the convolution
at these corner-cases in post-processing. This allows us to leave the convolution kernel code –
the computational bottleneck of the code – untouched. The corner-cases are fixed using a highly
efficient kernel which executes an element-wise sum between the results of the convolution and the
correction matrix. The correction matrix is computed once, when the GPU opt layer is loaded, and
it simply consists of the convolution of the layer’s weights with a (+1)-padded zero-tensor.
Converting a network to Espresso A DNN in Espresso is defined as a combination of layers,
which is loaded at run-time by reading its parameters file. The parameters file specifies the storage
format of all the layers, as well as their weights. Therefore, it completely specifies a DNN as layers
are stored sequentially. Training of the network is done by BinaryNet (Hubara et al., 2016); the
resulting parameters are converted to the Espresso format by utility script distributed together with
our sources.
6
E VALUATION
The performance of our framework is evaluated in terms of average computational time needed
to perform a particular task. The execution times, averaged over 100 experiments, are obtained
on a machine equipped with an NVIDIA GeForce GTX 960 with 2GB of RAM, and a Intel R
dual-Xeon R X5660 @ 2.80 GHz. In CPU mode, we configure the OpenBLAS library for matrix
multiplication to use all the 24 available cores.
Experimental design We perform three quantitative evaluations: (Section 6.1) matrix multiplications of two dense square matrices of size 8192 × 8192; (Section 6.2) forward-propagations
of a Multi-Layer Perceptron (MLP) trained on the MNIST dataset (LeCun et al., 1998); (Section 6.3) forward-propagations of a Convolutional Neural Network (CNN) trained on the CIFAR-10
dataset (Krizhevsky et al., 2009). By Using the same models and datasets, we compare Espresso
6
Published as a conference paper at ICLR 2018
with: (1) the author provided optimized implementation of BinaryNet (Courbariaux et al., 2015);
(2) the optimized BDNN implemented in the Intel Nervana neon framework (NervanaSystems);
(3) a self-comparison across {CPU, GPU, GPU opt } as no binary-optimized implementations of convolutional layers are publicly available. Espresso is numerically equivalent to BinaryNet in terms of
classification accuracy. Therefore our evaluation focuses on computation speed.
Public datasets The MNIST dataset (LeCun et al., 1998) consists of 60K instances for training
and, 10K instances for testing. Each instance is a 28 × 28 grayscale image that depicts digits ranging
from 0 to 9. The CIFAR-10 dataset (Krizhevsky et al., 2009), consists of 50K training instances and
10K testing instances of 32 × 32 × 3 color images. Images are subdivided into 10 classes (airplanes,
automobiles, birds, cats, deers, dogs, frogs, horses, ships and trucks). Since our interest is to asses
the real-time performance of binary optimized DNNs, in those experiment we use a batch-size of
one, and measure the averaged forward time for each image of the testing-sets for each dataset.
6.1
B INARY
DENSE MATRIX MULTIPLICATION
Table 1: Averaged time of binary optimized matrix multiplication.
BinaryNet
Espresso GPU opt 32-bit
Espresso GPU opt 64-bit
88 ms
16 ms (5.5×)
11 ms (8×)
In computing dense matrix multiplication, Espresso outperforms BinaryNet by a ≈ 8× factor. Much
of the gain can be attributed to our optimized kernels, and the use of register blocking: by fetching
bigger data from main memory and shared memory, our kernel increases the bandwidth utilization
by decreasing the number of memory fetch instructions. The use of 64-bit packing instead of the
32-bit (such as that of BinaryNet), introduces an additional performance improvement. The 64-bit
kernel achieves a memory DRAM throughput of 40 GB s−1 for reads and 5 GB s−1 for writes, while
the 32-bit kernel obtain 29.6 GB s−1 for reads and 3.6 GB s−1 for writes. This translates into the
resulting ≈ 25% speed improvement.
6.2
M ULTI - LAYER PERCEPTRON
ON
MNIST
Table 2: Average prediction time of the BMLP.
BinaryNet
Nervana/Neon
Espresso CPU
Espresso GPU
Espresso GPU opt
18 ms
17 ms
37.4 ms
3.2 ms (5.6×)
0.26 ms (68×)
We evaluate the average classification execution time over the MNIST dataset, where we trained
the MLP architecture from (Courbariaux et al., 2016, Sec 2.1) with author-provided sources, and
then converted it to Espresso’s format. In Table 2, Espresso achieves a consistent speed-up of ≈
68× when compared to BinaryNet. As the Nervana/neon implementation of binary network is
a BinaryNet derivative, it is affected by the same drawbacks of BinaryNet, and hence achieves
comparable performance. Both alternatives have the additional cost of running CUDA through
Python/Theano which may introduce further latency in the process. In Table 2, the evaluation over
the three variants of Espresso shows the expected outcome, with the GPU opt implementation leading
the ranking. Note that we are able to achieve a speedup of ≈ 12× on an NVIDIA GTX 960 (≈
2.5 TFLOPs), although this device has only roughly four times more throughput than the Xeon
X5660 (≈ 500 GFLOPs without turbo-boost). Through binary optimization, we are able to further
increase the performance to ≈ 15× with respect to the GPU implementation. We attribute our
computational gains to (1) the use of binary-optimized layers, (2) our use of optimized kernels for
matrix multiplication and (3) Espresso’s ability to perform binary optimization of the first layer.
Binary optimized layers An evident drawback of Binary-Net is the need for binarizing/packing
the layer’s parameters every time a forward method is called. In the case of binary optimized networks, the cost of packing the parameters is closely related to the cost of multiplication itself. There7
Published as a conference paper at ICLR 2018
fore, the reduction of bit-packing function calls leads to a consistent improvement. This motivates
our choice of designing specific layers, where bit-packing is done once during network loading.
Optimized kernels BinaryNet employs two bit-packing kernels: one for row-packing, the other
for column-packing. Although BinaryNet’s pack-by-rows kernel is slightly slower than ours (8%),
the pack-by-columns kernel is significantly slower (≈ 4×) due to non-coalesced accesses to global
memory. An additional performance gain of ≈ 15% is achieved by swapping matrix-vector in favour
of matrix-matrix multiplication kernels when appropriate (i.e. Dense layers with batch size equal to
1); for this reason, Espresso also includes the binary-optimized MAGMA(sgemv) kernel.
First-layer binary optimization Another important advantage offered by Espresso is the ability
to leverage binary optimization in the first layer. Since the first stage of a network processes nonbinary data, BinaryNet does not feature binary optimization for this layer. However if the input data
is split into its constituent bit-planes, binary optimization can still be applied. In particular, we split
the input vector in a matrix of 8 rows, and recombine the result after multiplication by a weighted
sum. Our experimental results report an overall ≈ 3× performance boost when comparing the full
binary optimized network with one in which the first layer is not binary optimized.
Finally, in terms of memory the GPU opt implementation requires 4.57 MB instead of 140.6 MB as
in the case of non binary optimized implementation, resulting in a saving ≈ 31× amount of memory.
6.3
C ONVOLUTIONAL N EURAL N ETWORK
ON
CIFAR-10
Table 3: Average prediction time of the BCNN.
Espresso CPU
Espresso GPU
Espresso GPU opt
85.2 ms
5.2 ms (16×)
1.0 ms (85×)
To the best of our knowledge, no BDNN implementation of binary-optimized CNN layers is publicly available. Our self-evaluation implements the VGGNet-like CNN architecture from Hubara
et al. (Hubara et al., 2016, Sec. 2.3), and evaluates it across our three modalities: as expected the
GPU opt implementation achieves significantly better performance.
Unrolling and pooling Note how the GPU implementation offers a slightly better improvement
over CPU with respect to the MLP test, with an ≈ 16× speed-up. In this experiment, the inherent
parallelism of unrolling and pooling, and the GPU higher memory throughput explain the behavior.
Gains are marginal as FMA still represents the computational bottleneck.
Bit-packing The GPU opt implementation results in a ≈ 5× performance gain with the respect
to GPU . These gains, to binary optimizations, are slightly smaller than those discussed for MLP in
Section 6.2. The output of convolutional layers is significantly larger than those of MLP’s dense
layers, therefore, the computation of bit-packing sign-activation requires more computational effort.
Finally, in terms of memory the GPU opt implementation requires 1.73 MB instead of 53.54 MB as
in the case of non binary optimized implementation, resulting in a saving ≈ 31× amount of memory.
7
C ONCLUSIONS
In this paper we presented Espresso, a highly optimized forward-propagation framework for both traditional DNNs as well as BCNNs, that supports heterogeneous deployment on CPU and GPU. While
BinaryNet and Nervana/neon BDNN implementations are limited to MLP networks, our framework
also supports the popular CNN while simultaneously outperforming state-of-the-art implementations of MLP networks. Espresso is highly-efficient, light-weight and self-contained. Computation
on the GPU side is done though specifically designed CUDA kernels, which, combined with a
more careful handling of memory allocation and bit-packing, allows us to obtain considerable performance improvements. In future work we would like to add training capabilities, and perform
additional performance comparisons on larger standard datasets.
8
Published as a conference paper at ICLR 2018
R EFERENCES
Ahmad Abdelfattah, Azzam Haidar, Stanimire Tomov, and Jack Dongarra. Performance, design,
and autotuning of batched GEMM for GPUs. In International Conference on High Performance
Computing, pp. 21–38. Springer, 2016. http://icl.cs.utk.edu/magma/ (Link verified
on May 11th, 2017).
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly
learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.
Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients
through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.
James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume
Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: A CPU and GPU
math compiler in python. In Proc. 9th Python in Science Conf, pp. 1–7, 2010.
Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Training deep neural networks with
low precision multiplications. arXiv preprint arXiv:1412.7024, 2014.
Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural
networks with binary weights during propagations. In Advances in Neural Information Processing
Systems, pp. 3123–3131, 2015.
Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized
neural networks: Training deep neural networks with weights and activations constrained to+ 1
or-1. arXiv preprint arXiv:1602.02830, 2016.
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale
hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009.
IEEE Conference on, pp. 248–255. IEEE, 2009.
Jack J Dongarra, Jeremy Du Croz, Sven Hammarling, and Iain S Duff. A set of level 3 basic linear
algebra subprograms. ACM Transactions on Mathematical Software (TOMS), 16(1):1–17, 1990.
Alex Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. Speech recognition with deep recurrent neural networks. In Acoustics, speech and signal processing (icassp), 2013 ieee international
conference on, pp. 6645–6649. IEEE, 2013.
Yiwen Guo, Anbang Yao, and Yurong Chen. Dynamic network surgery for efficient dnns. In
Advances In Neural Information Processing Systems, pp. 1379–1387, 2016.
Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks
with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.
Song Han, Xingyu Liu, Huizi Mao, Jing Pu, Ardavan Pedram, Mark A Horowitz, and William J
Dally. EIE: efficient inference engine on compressed deep neural network. In Proceedings of the
43rd International Symposium on Computer Architecture, pp. 243–254. IEEE Press, 2016.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.
770–778, 2016.
Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly,
Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks
for acoustic modeling in speech recognition: The shared views of four research groups. IEEE
Signal Processing Magazine, 29(6):82–97, 2012.
Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized
neural networks. In Advances in Neural Information Processing Systems, pp. 4107–4115, 2016.
https://github.com/MatthieuCourbariaux/BinaryNet.
Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt
Keutzer. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 mb model
size. arXiv preprint arXiv:1602.07360, 2016.
9
Published as a conference paper at ICLR 2018
Norm Jouppi. Google supercharges machine learning tasks with TPU custom chip, 2017.
Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton.
The
https://www.cs.toronto.edu/~kriz/cifar.html, 2009.
May 11th, 2017.
CIFAR-10 dataset.
Link verified on
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,
2012.
Yann LeCun, Corinna Cortes, and Christopher J.C. Burges. The MNIST database of handwritten
digits. http://yann.lecun.com/exdb/mnist, 1998. Link verified on May 11th, 2017.
Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for
efficient convnets. arXiv preprint arXiv:1608.08710, 2016.
Rajib Nath, Stanimire Tomov, and Jack Dongarra. An improved magma gemm for fermi graphics
processing units. The International Journal of High Performance Computing Applications, 24(4):
511–515, 2010.
Intel
NervanaSystems.
The
Neon
deep
learning
framework.
https://github.com/NervanaSystems/neon. Link verified on May 11th, 2017.
Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading
digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning
and unsupervised feature learning, volume 2011, pp. 5, 2011.
Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet
classification using binary convolutional neural networks. In European Conference on Computer
Vision, pp. 525–542. Springer, 2016.
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556, 2014.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks.
In Advances in neural information processing systems, pp. 3104–3112, 2014.
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1–9, 2015.
Zhang Xianyi, Wang Qian, and Werner Saar. OpenBLAS: An optimized BLAS library.
http://www.openblas.net. Link verified on May 11th, 2017.
Chenzhuo Zhu, Song Han, Huizi Mao, and William J Dally. Trained ternary quantization. arXiv
preprint arXiv:1612.01064, 2016.
10
| 9cs.NE
|
arXiv:1208.5785v2 [math.RA] 6 Sep 2012
THE NEGATIVE SIDE OF COHOMOLOGY FOR CALABI-YAU
CATEGORIES
PETTER ANDREAS BERGH, DAVID A. JORGENSEN & STEFFEN OPPERMANN
Abstract. We study Z-graded cohomology rings defined over Calabi-Yau categories. We show that the cohomology in negative degree is a trivial extension
of the cohomology ring in non-negative degree, provided the latter admits a
regular sequence of central elements of length two. In particular, the product
of elements of negative degrees are zero. As corollaries we apply this to TateHochschild cohomology rings of symmetric algebras, and to Tate cohomology
rings over group algebras. We also prove similar results for Tate cohomology
rings over commutative local Gorenstein rings.
1. Introduction
Let T be an n-Calabi-Yau triangulated category, and X an object in T . We
investigate the structure of the Z-graded cohomology ring
Hom∗T (X, X) = ⊕i∈Z HomT (X, Σi X).
One of our main results states that when the subalgebra ⊕i≥0 HomT (X, Σi X)
possesses a regular sequence of length 2 of central elements of positive degree,
then n ≤ −1 and most products in negative degree are trivial. More precisely, if
x, y ∈ Hom∗T (X, X) with |x| ≤ n and |y| < 0, then xy = 0. In particular, when
n = −1, which is a common case in applications, then one can remove the word
“most” from the previous sentence, and in fact, the cohomology in negative degree
is a trivial extension of the cohomology ring in non-negative degree.
The main situation to which our results apply is when T is the stable module
category of a finite dimensional symmetric algebra Λ over a field. Here our results
∗
d Λ (M, M ) of a finitely generated (left) Λtranslate to the Tate cohomology ring Ext
module M having trivial products in negative degree. As a special case we recover
d ∗ (k, k),
a result of Benson and Carlson [BenCa] on the Tate cohomology ring Ext
kG
for G a finite group whose order is divisible by the characteristic of the ground field
k. Indeed, their paper served as a major inspiration for the present one. Another
∗
d (Λ) of a symmetric alcorollary is that the Tate-Hochschild cohomology ring HH
gebra Λ, defined in [BerJo], has trivial products in negative degree. We also give
various results on the structure of the cohomology rings when ⊕i≥0 HomT (X, Σi X)
is only assumed to admit a single central regular element of positive degree.
In Section 2 we derive properties in the abstract for products in Z-graded associative algebras which possess a natural duality. The key point here is that when
the nonnegative side of such algebras admit a regular sequence of length 1 or 2
Date: January 6, 2018.
2010 Mathematics Subject Classification. 13H10, 16E40, 16W50.
Key words and phrases. Graded algebras, negative products, Calabi-Yau categories,
1
2
PETTER ANDREAS BERGH, DAVID A. JORGENSEN & STEFFEN OPPERMANN
contained in the positive side of the algebra, then products in negative degrees are
often zero. These results lay the foundation for the rest of the paper.
In Section 3 we apply the results of Section 2 to cohomology rings of objects
in n-Calabi-Yau categories. This includes the specific results on Tate and TateHochschild cohomology mentioned above.
Finally, in Section 4 we prove similar results for Tate cohomology rings over
commutative Noetherian local rings, in particular Gorenstein rings. For example,
we show that when (R, m) is a zero-dimensional local Gorenstein ring, M is a finitely
generated R-module, and ⊕i≥0 ExtiR (M, M ) contains a regular sequence of length
∗
d R (M, M ) the products of negative degree are zero (cf. [AvVe]
two, then in Ext
d ∗ (R/ m, R/ m) is treated in any dimension.)
where the structure of Ext
R
2. Products in Z-graded algebras
L
i
that is, a
Throughout this section we let A =
i∈Z A be a Z-graded ring,
L
≥n
unital ring graded by the integers. For an integer n we write A = i≥n Ai (and
similarly for A>n , A≤n , . . . ). We will be particularly interested in the non-negative
subalgebra A≥0 = ⊕i≥0 Ai of A. For any graded A-module M and any integer t,
the shifted graded module M [t] is defined by M [t]i = M i+t .
Definition. The Z-graded ring A has non-degenerate products in degree n if for
any integer i and any a ∈ Ai \ {0} there are b, c ∈ An−i such that
ab 6= 0 6= ca.
If A is also a k-algebra, for k a field (always assumed to be concentrated in
degree zero) we let D A denote the graded dual of A, that is the bimodule given by
(D A)i = Homk (A−i , k).
Definition. Assume A is a Z-graded k-algebra over a field k. Then A is n-shifted
selfdual if
A[n] ∼
= DA
as left (or equivalently as right) A-modules.
Bilinear forms. For a graded ring A, consider a graded bilinear form
h−, −i : A × A → X,
where X is some abelian group, concentrated in degree n. (If A additionally is a
k-algebra, we assume X to be a k-module and the form to be k-bilinear.) This
means we have a family of bilinear maps
h−, −ii : Ai × An−i → X.
Definition. The graded bilinear form h−, −i : A × A → X is called
• non-degenerate if for i ∈ Z and a ∈ Ai \ {0} there are b, c ∈ An−i such that
ha, bi 6= 0 6= hc, ai.
• associative if for any a ∈ Ai , b ∈ Aj , and c ∈ An−i−j we have hab, ci =
ha, bci.
Proposition 2.1. Let A be a Z-graded ring. Then A has non-degenerate products
in degree n if and only if A admits a non-degenerate associative graded bilinear
form to an abelian group concentrated in degree n.
THE NEGATIVE SIDE OF COHOMOLOGY FOR CALABI-YAU CATEGORIES
3
Proof. Giving an associative graded bilinear form A×A → X is equivalent to giving
a graded map A ⊗A A → X of abelian groups (where the tensor product A ⊗A A is
the graded tensor product). Note that A ⊗A A ∼
= A as graded rings, and thus we
obtain a map An → X of abelian groups. The claim now follows.
Proposition 2.2. Let A be a Z-graded k-algebra. Then the following are equivalent
(1) A is n-shifted selfdual, and
(2) all graded pieces Ai are finite dimensional, and A admits a non-degenerate
associative graded bilinear form A × A → k[−n].
Proof. (1) ⇒ (2): Fix an isomorphism Φ : A[n] → D A of graded left A-modules.
Then Φ consists of isomorphisms Φi : An+i → Homk (A−i , k) for all i. In particular
we obtain isomorphisms
Homk ((Φ−n−i )−1 , k) ◦ Φi : An+i → Homk (Homk (An+i , k), k).
It follows that An+i is a finite dimensional k-vector space for all i.
Now we can define a non-degenerate associative graded bilinear form by
1×Φ[−n]
ev.
A × A −−−−−−→ A × D A[−n] −−→ k[−n]
(a, ϕ) 7→ ϕ(a)
(2) ⇒ (1): Since the graded pieces are now assumed to be finite dimensional,
having a non-degenerate bilinear form Ai × An−i → k gives an isomorphism Ai →
Homk (An−i , k). Together these isomorphisms form an isomorphism A → D(A[n]) =
D A[−n] of graded k-vector spaces.
The associativity of the bilinear form implies that this isomorphism respects the
A-module structure.
Remark. We note that the two propositions above in particular show that any
algebra being n-shifted selfdual also has non-degenerate products in degree n.
Regular elements. Let A be a Z-graded ring. A homogeneous central element
r ∈ A is called a regular element of A if the implication
ra = 0
=⇒
a=0
holds for all a ∈ A.
Throughout this subsection, we assume that r is a central element of A of positive
degree, and a regular element of A≥0 . We assume additionally that A has nondegenerate products in degree n.
We consider the following two subbimodules of A:
Torr A = {a ∈ A | ri a = 0 for some i ≥ 0}
≤n
I = (A
)
the r-torsion part of A
the ideal of A generated by elements
of degree at most n
Theorem 2.3. Let A be a Z-graded ring which has non-degenerate products in
degree n. If r is a homogeneous central element of A of positive degree, and a
regular element of A≥0 , then we have the following:
(1) I · Torr A = 0 = Torr A · I;
(2) either n < 0, or r acts regularly on A.
4
PETTER ANDREAS BERGH, DAVID A. JORGENSEN & STEFFEN OPPERMANN
Proof. For (1) it is sufficient to show that A≤n · Torr A = 0 = Torr A · A≤n . We
show the first equality, the second one can be shown similarly.
Let a ∈ Ai for some i ≤ n, and b ∈ (Torr A)j . Assume ab 6= 0. Then there exists
c ∈ An−i−j such that abc 6= 0. Since Torr A is a subbimodule of A we have
bc ∈ (Torr A)j+(n−i−j) = (Torr A)n−i ⊆ (Torr A)≥0 = 0,
where the last equality follows from the assumption that r acts regularly on A≥0 .
This is clearly a contradiction, so ab = 0.
(2): Note that if n ≥ 0 then 1A ∈ I, so I = A. It follows from (1) that
Torr A = 0, so r acts regularly on A.
We observe that if A is an algebra over a field, which is n-shifted selfdual, we
obtain the following more explicit description of I, and of what it means that r acts
regularly on A.
Lemma 2.4. Assume A is a Z-graded algebra over a field k, which is n-shifted
selfdual, and r is a central element of A of positive degree and regular on A≥0 .
Then
DI ∼
= (A/ Torr A)[n]
as right and as left A-modules.
Proof. By definition I is the smallest subbimodule of A containing A≤n . Applying
D −[−n] we obtain that D I[−n] is the smallest factor bimodule of A, in which no
elements of A≥0 become zero. It is clear that this is a factor bimodule of A/ Torr A.
On the other hand r acts regularly on A/ Torr A, so any non-zero element a of
A/ Torr A has a non-zero multiple of the form ri a ∈ A≥0 , and thus cannot be
mapped to zero. It follows that D I[−n] ∼
= A/ Torr A.
Lemma 2.5. Assume A is an algebra over a field k, which is n-shifted selfdual. If
r is a central element of A of positive degree which acts regularly on A, then A is
periodic.
Proof. It follows from duality that r acting regularly is equivalent to every element
being divisible by r. Thus r acts bijectively on A, and thus induces isomorphisms
Ai → Ai+|r| for all i.
It follows from Theorem 2.3 and Lemma 2.5 that the case n ≥ 0 is very special.
In practice it seems the most typical case is n = −1, which we therefore record here
explicitly.
Theorem 2.6. Assume A is an algebra over a field k, which is (−1)-shifted selfdual,
and r is a central element of A of positive degree which is regular on A≥0 . Then A
has a filtration of subbimodules
0 ⊆ Torr A ⊆ I ⊆ A,
such that
(1) Torr A is concentrated in negative degrees, A/I is concentrated in nonnegative degrees, and these two are dual as left and as right A-modules.
(2) I/ Torr A is periodic.
THE NEGATIVE SIDE OF COHOMOLOGY FOR CALABI-YAU CATEGORIES
5
Proof. We first note that Torr A is concentraded in negative degrees by assumption,
and I contains all elements of negative degree by definition. Thus we have the
filtration of the theorem, and the first two statements of (1).
For the duality statement in (1) note that
D(A/I) = Ker[D A ։ D I] = (Torr A)[−1]
by Lemma 2.4.
For (2) note that r acts regularly on I/ Torr A, and that I/ Torr A is selfdual.
The proof now follows the proof of Lemma 2.5.
Depth ≥ 2. In this subsection we assume that the non-negative part A≥0 of a
Z-graded ring A admits a regular sequence (r, r̃) of length 2 of central elements of
positive degree. That means that r acts regularly on A≥0 , and r̃ acts regularly on
A≥0 /rA≥0 .
In this situation we obtain the following:
Lemma 2.7. Let A be a Z-graded ring, such that A≥0 admits a regular sequence
(r, r̃) of central elements of A of positive degree. Then
Torr A = A<0 .
Proof. Assume there is a ∈ A<0 \ Torr A. Then for any i we have ri a 6= 0. Choose
i minimal such that ri a ∈ A≥0 . Clearly ri a 6∈ rA≥0 , but for j ≫ 0 we have
r̃j ri a = rr̃j ri−1 a ∈ rA≥0 . This contradicts the assumption that r̃ acts regularly on
A≥0 /rA≥0 .
Remark. We know that Torr A is an A-subbimodule of A. Lemma 2.7 shows that
it is also a quotient A≥0 -bimodule of A, and thus
A∼
= A≥0 ⊕ A<0
as A≥0 -bimodules.
We obtain the following result immediately from Theorem 2.3 and Lemma 2.7.
Theorem 2.8. Let A be a Z-graded ring having non-degenerate products in degree
n, such that A≥0 admits a regular sequence (r, r̃) of central elements of A of positive
degree. Then n < 0, and A≤n ⊆ A<0 are ideals of A annihilating each other.
Proof. By Lemma 2.7 we know that A<0 = Torr A is an ideal of A. It follows from
Theorem 2.3(2) that n < 0. (Note that Torr A = A<0 6= 0, since otherwise the
non-degenerate products assumption would imply A>n = 0, clearly contradicting
the existence of a regular element.)
As before we denote by I the ideal generated by A≤n . By Theorem 2.3(1) we
know that I and A<0 annihilate each other.
It remains to show that I = A≤n . Assume a ∈ I i \ {0} for some i > n. By
assumption there is b ∈ An−i ⊆ A<0 such that ab 6= 0, and this contradicts the fact
that I · A<0 = 0.
If we restrict to the special case that A is (−1)-shifted selfdual we otain the
following more explicite description of A.
Corollary 2.9. Let A be a Z-graded ring having non-degenerate products in degree
−1. If A≥0 admits a regular sequence (r, r̃) of central elements of A of positive
degree, then
A∼
= A≥0 ⋉ A<0 .
6
PETTER ANDREAS BERGH, DAVID A. JORGENSEN & STEFFEN OPPERMANN
If A is moreover a k-algebra which is (−1)-shifted selfdual, then D A<0 ∼
= A≥0 [−1]
≥0
both as left and as right A -modules.
Proof. The first statement holds since, by Theorem 2.8 A<0 is an ideal which
squares to zero. The second claim is a restatement of Lemma 2.4, using Torr A =
A<0 (Lemma 2.7) and I = A<0 (proof of Theorem 2.8).
3. Calabi-Yau Categories
In this section we apply the general results from Section 2 to endomorphism rings
in Calabi-Yau triangulated categories. We fix a field k, and start with a general
lemma, which guarantees that a certain family of bilinear forms is associative and
non-degenerate. Recall first that a category T is a k-category if for all objects
X, Y, Z ∈ T , the set HomT (X, Y ) is a k-vector space, and composition
HomT (Y, Z) × HomT (X, Y ) → HomT (X, Z)
is k-bilinear.
Lemma 3.1. Let k be a field, T a k-category, and X and Z objects in T such that
there is a natural isomorphism
η : HomT (X, −) → D HomT (−, Z).
Then the bilinear forms
h−, −iY : HomT (Y, Z) ⊗k HomT (X, Y ) → k
f ⊗ g 7→ ηY (g)(f )
are non-degenerate. Moreover, they are associative in the sense that hf, g ◦ hiY =
f
g
h
→ Y and X −
→Y′
→ Z, Y ′ −
hf ◦ g, hiY ′ for all objects Y ′ ∈ T and all morphisms Y −
in T .
Proof. The bilinear forms are non-degenerate by definition, so we only have to
check associativity. The isomorphism η being natural means that for any morphism
g
Y′ −
→ Y there is an equality
D HomT (g, Z) ◦ ηY = ηY ′ ◦ HomT (X, g).
h
Applying this to a morphism X −
→ Y ′ we obtain
ηY ′ (h)(− ◦ g) = ηY (g ◦ h)(−),
f
and inserting a morphism Y −
→ Z we obtain the formula of the lemma.
Having established this elementary lemma, we turn now to Calabi-Yau triangulated categories. Let T be a Hom-finite (i.e. HomT (X, Y ) is a finite dimensional
k-vector space for all objects X, Y ) triangulated k-category with suspension functor
Σ : T → T . A Serre functor on T is an equivalence S : T → T of triangulated kcategories, together with functorial isomorphisms HomT (X, Y ) ∼
= D HomT (Y, SX)
of vector spaces for all objects X, Y ∈ T . By [BoKa], such a functor is unique if
it exists. For an integer n ∈ Z, the category T is called n-Calabi-Yau if it admits
a Serre functor which is isomorphic as a triangle functor to Σn (cf. [Ke]). In this
case there are functorial isomorphisms HomT (X, Y ) ∼
= D HomT (Y, Σn X) of vector
spaces for all objects X, Y ∈ T .
We show next that Lemma 3.1 implies these functorial isomorphisms induce a
non-degenerate associative graded bilinear form on the graded endomorphism ring
THE NEGATIVE SIDE OF COHOMOLOGY FOR CALABI-YAU CATEGORIES
7
of an object in T . We first recall briefly some facts about such graded endomorphism
rings. For an object X ∈ T , the graded endomorphism ring of X is the graded
vector space
A = ⊕i∈Z Ai with Ai = HomT (X, Σi X),
together with a product defined as follows. For f ∈ Ai and g ∈ Aj , we identify
f with its image under the isomorphism Ai ∼
= HomT (Σj X, Σi+j X) and define
i+j
fg ∈ A
by composition.
Proposition 3.2. Let k be a field, T an n-Calabi-Yau triangulated k-category,
and X an object in T . Let A = ⊕i Ai denote the graded endomorphism ring of
X, namely, the Z-graded k-algebra with Ai = HomT (X, Σi X). Then there exists a
non-degenerate associative graded bilinear form
h−, −i : A ⊗k A → k[−n].
Proof. Since T is n-Calabi-Yau, there are functorial isomorphisms
HomT (X, Σn−i X) ∼
= D HomT (Σn−i X, Σn X)
for all i. Therefore from Lemma 3.1 we have a family of non-degenerate bilinear
forms
h−, −ii : HomT (Σn−i X, Σn X) ⊗k HomT (X, Σn−i X) → k.
f
Moreover, this family of bilinear forms is associative in the sense that for Σn−i X −
→
g
h
n
n−j
n−i
n−j
i
Σ X, Σ
X −
→Σ
X and X −
→Σ
in T , we have hf, g ◦ hi = hf ◦ g, hij .
i ∼
Using the functorial isomorphisms A = HomT (Σn−i X, Σn X) for all i, we thus
obtain a non-degenerate associative graded bilinear form
A ⊗k A → k[−n],
We may therefore apply the results of Section 2 to the graded endomorphism
rings of objects in n-Calabi-Yau categories. For future reference we summarize
these in the following.
As in the previous section we let Torr A = {a ∈ A | ri a = 0 for some i ≥ 0}
denote the r-torsion part of A, and I = (A≤n ) denote the ideal of A generated by
elements of degree at most n.
Theorem 3.3. Let k be a field, and T an n-Calabi-Yau triangulated k-category.
Furthermore, let X be an object in T , and denote by A the Z-graded k-algebra with
Ai = HomT (X, Σi X).
If r is a homogeneous central element of A of positive degree which is regular on
A≥0 , then
(1) I · Torr A = 0 = Torr A · I;
(2) either n < 0, or r acts regularly on A, and A is periodic;
(3) D I ∼
= (A/ Torr A)[n] as right and as left A-modules.
If in addition A≥0 admits a regular sequence (r, r̃) of central elements of A of
positive degree, then
(4) Torr A = A<0 ;
(5) A ∼
= A≥0 ⋉ A<0 ;
≤n
(6) A ⊆ A<0 are ideals of A annihilating one another.
8
PETTER ANDREAS BERGH, DAVID A. JORGENSEN & STEFFEN OPPERMANN
We can further apply Theorem 3.3 and the results of Section 2 to Tate cohomology rings of objects in the stable module category of a symmetric algebra. Recall
that a finite dimensional k-algebra Λ is symmetric if there exists an isomorphism
Λ ∼
= D(Λ) of Λ-bimodules. Such an algebra is in particular selfinjective, and its
stable module category mod Λ (we consider only finitely generated left modules) is
a Hom-finite triangulated k-category with suspension functor Σ = Ω−1
Λ . Given an
i
d (M, N )
integer i ∈ Z and two Λ-modules M, N , the ith Tate cohomology group Ext
Λ
is defined as
i
d Λ (M, N ) def
Ext
= HomΛ (M, Ω−i
Λ (N )),
that is, the vector space Hommod Λ (M, Σi N ). When M = N , the graded k-vector
∗
i
d (M, M ) = ⊕i∈Z Ext
d (M, M ) becomes a ring, the Tate cohomology ring
space Ext
Λ
Λ
of M .
Proposition 3.4. Let k be a field, Λ a finite dimensional symmetric k-algebra.
Then the stable module category mod Λ is (−1)-Calabi-Yau.
Proof. Given finitely generated Λ-modules M and N , there is a functorial isomorphism
HomΛ (M, N ) ∼
= D HomΛ (N, τ Ω−1
Λ (M )),
where τ is the Auslander-Reiten translate (cf. [KrLe] or [Bu]). Also, by the result
[AuReSm, Proposition IV.3.8], the Auslander-Reiten translate is naturally isomorphic to Ω2Λ , hence we obtain a functorial isomorphism
Hom (M, N ) ∼
= D Hom (N, Ω1 (M )).
Λ
Λ
Λ
Consequently, the stable module category mod Λ is (−1)-Calabi-Yau.
Theorem 3.5. Let k be a field, Λ a finite dimensional symmetric k-algebra and
M a finitely generated left module. Denote by A the Z-graded k-algebra with Ai =
i
d (M, M ).
Ext
Λ
If r is a homogeneous central element of A of positive degree which is regular on
A≥0 , then
(1) I · Torr A = 0 = Torr A · I;
(2) A has a filtration of subbimodules 0 ⊆ Torr A ⊆ I ⊆ A such that
• Torr A is concentrated in negative degrees, A/I is concentrated in nonnegative degrees, and these two are dual as left and as right A-modules;
• I/ Torr A is periodic.
If in addition A≥0 admits a regular sequence (r, r̃) of central elements of A of
positive degree, then
(3) Torr A = A<0 ;
(4) (A<0 )2 = 0 and A ∼
= A≥0 ⋉ A<0 ;
<0 ∼ ≥0
(5) D A = A [−1] both as left and as right A≥0 -modules.
The group algebra of a finite group is symmetric, and so Theorem 3.5 applies to
Tate cohomology rings of modules over such an algebra. In particular, the theorem
applies to the Tate group cohomology ring, which is just the Tate cohomology ring
of the trivial module. We thus recover [BenCa, Theorem 3.1]. Note that the Tate
group cohomology ring is always graded commutative, hence homogeneous elements
THE NEGATIVE SIDE OF COHOMOLOGY FOR CALABI-YAU CATEGORIES
9
in odd degree square to zero when the characteristic of the underlying field is not
two. Therefore, when the characteristic of the underlying field is not two, a homogeneous regular sequence must consist of elements in even degrees. Consequently, in
any characteristic, a homogeneous regular sequence must automatically be central.
Corollary 3.6. Let k be a field, G a finite group whose order is divisible by the
∗
d kG (k, k).
characteristic of k, and denote by A the Tate group cohomology ring Ext
Then the conclusions of Theorem 3.5 hold.
As a second application of Theorem 3.5, we consider the Tate analogue of
Hochschild cohomology rings. Let Λ be a finite dimensional symmetric k-algebra.
Then by [BerJo, Corollary 3.3], its enveloping algebra Λe = Λ ⊗k Λop is also symi
d (Λ)
metric. Given an integer i ∈ Z, the ith Tate-Hochschild cohomology group HH
is defined as
i
i
d (Λ) def
d Λe (Λ, Λ),
HH
= Ext
d ∗ (Λ) is the Tate-Hochschild cohomology ring of Λ. In
and the graded ring HH
positive degree, this ring coincides with the ordinary Hochschild cohomology ring,
≥1
d (Λ) = HH≥1 (Λ). As with group cohomology rings of finite groups, Tatei.e. HH
Hochschild cohomology rings are graded commutative, hence homogeneous regular
sequences must be central.
Corollary 3.7. Let k be a field, Λ a finite dimensional symmetric k-algebra, and
∗
d (Λ). Then the conclusions
denote by A the Tate-Hochschild cohomology ring HH
of Theorem 3.5 hold.
We include an example illustrating Corollary 3.7.
Example. Let c be a positive integer and Λ the commutative local complete intersection
k[x1 , . . . , xc ]/(xa1 1 , . . . , xac c ),
where a1 , . . . , ac are integers, all at least 2. This is a finite dimensional symmetric
k-algebra. We denote the characteristic of the ground field k by p.
Suppose that c = 1, and denote x1 and a1 by just x and a, so that Λ = k[x]/(xa ).
Then it follows from [BerJo, Theorem 4.4] that
n
a − 1 if p ∤ a
d (Λ) =
dim HH
a
if p | a
for every integer n ∈ Z. It is well known that Λ is periodic as a bimodule over
itself, with period at most 2. Also, by [Ho, Theorem 7.1], its ordinary Hochschild
cohomology ring HH∗ (Λ) is given by
k[u, v, w]/(ua , wua−1 , vua−1 , v 2 ) if p ∤ a
k[u, v, w]/(ua , v 2 )
if p | a and p 6= 2
HH∗ (Λ) ∼
=
a 2
a−2
k[u,
v,
w]/(u
,
v
−
wu
)
if p | a, p = 2 and a ≡ 2(mod 4)
k[u, v, w]/(ua , v 2 )
if p | a, p = 2 and a ≡ 0(mod 4)
where |u| = 0, |v| = 1 and |w| = 2. The element w is the “periodicity element”, and the element u corresponds to the element x ∈ Λ. When passing to
∗
d (Λ), the socle element xa−1 , i.e. the elethe Tate-Hochschild cohomology ring HH
≥0
d (Λ) of the
ment ua−1 , vanishes when p ∤ a. Therefore the non-negative part HH
10
PETTER ANDREAS BERGH, DAVID A. JORGENSEN & STEFFEN OPPERMANN
Tate-Hochschild cohomology ring is given by
k[u, v, w]/(ua−1 , v 2 )
≥0
k[u, v, w]/(ua , v 2 )
d (Λ) ∼
HH
=
k[u, v, w]/(ua , v 2 − wua−2 )
k[u, v, w]/(ua , v 2 )
if
if
if
if
p∤a
p | a and p 6= 2
p | a, p = 2 and a ≡ 2(mod 4)
p | a, p = 2 and a ≡ 0(mod 4).
d≥0 (Λ) of length 2
Consequently, there does not exist a regular sequence in HH
of homogeneous elements of positive degree; the periodicity element w forms a
maximal homogeneous regular sequence. Of course, this is not unexpected; since Λ
is periodic as a bimodule, there are lots of negative degree homogeneous elements
d∗ (Λ) with ηθ 6= 0.
η, θ ∈ HH
Next, suppose that c ≥ 2. Since Λ is isomorphic to the tensor product
k[x1 ]/(xa1 1 ) ⊗k · · · ⊗k k[xc ]/(xac c ),
its Hochschild cohomology ring HH∗ (Λ) is isomorphic to a twisted tensor product
bk · · · ⊗
b k HH∗ (k[xc ]/(xac c )).
HH∗ (k[x1 ]/(xa1 1 ))⊗
This is almost the same as the usual tensor product, the only difference is that
elements of odd degrees anticommute. From the proof of [BerJo, Theorem 4.4] it
0
d (Λ) is given by
follows that the dimension of HH
Qc
0
d (Λ) =
Qci=1 ai − 1 if p ∤ ai for all i
dim HH
if p | ai for one i.
i=1 ai
Consequently, when p divides one of the ai , then when passing from HH∗ (Λ) to
∗
d (Λ), the socle element xa1 · · · xac vanishes. At any rate, regardless of the
HH
c
1
≥0
d
characteristic p, we see that HH (Λ) contains a regular sequence of length c
of homogeneous elements of positive degrees. For instance, we could take the
sequence w1 , . . . , wc , where each wi corresponds to the periodicity element in
∗
∗
d (Λ) are negative degree homoged (k[xi ]/(xai )). By Corollary 3.7, if η, θ ∈ HH
HH
i
neous elements, then ηθ = 0.
4. Commutative local rings
In this final section we consider negative products in Tate cohomology over
commutative local (meaning also Noetherian) rings. Let R be such a ring, and
denote the functor HomR (−, R) by (−)∗ . Recall that a finitely generated R-module
X is called totally reflexive if the following three conditions hold:
(1) X is reflexive, that is, the canonical homomorphism X → X ∗∗ is an isomorphism.
(2) ExtiR (X, R) = 0 for i ≥ 1.
(3) ExtiR (X ∗ , R) = 0 for i ≥ 1.
Such a module is also called a module of Gorenstein dimension zero. Now recall
that a complex
T :
∂T
∂T
∂T
0
1
2
F−1 → F−2 → · · ·
F0 −−→
F1 −−→
· · · → F2 −−→
of R-modules is called a complete resolution if the modules Fi are finitely generated
free and both T and T ∗ are acyclic. It is well known that a module X is totally
reflexive if and only if it is the image Im ∂0T of a map in such a complete resolution
THE NEGATIVE SIDE OF COHOMOLOGY FOR CALABI-YAU CATEGORIES
11
d i (X, Y )
T . In this case, given any R-module Y , the ith Tate cohomology group Ext
R
is defined as
i
d (X, Y ) def
Ext
= Hi (HomR (T, Y )) .
R
i
d Λ (X, Y ) is the ith cohomology of the complex HomR (T, Y ).
That is, the group Ext
i
∗
d R (X, X) becomes
d R (X, X) = ⊕i∈Z Ext
When Y = X, then the graded R-module Ext
an R-algebra called the Tate cohomology ring of X, where the multiplication is given
by composition of chain maps on the complete resolution of X.
It should be noted that Tate cohomology can be defined more generally when
X has finite Gorenstein dimension, not only when X is totally reflexive. However,
we shall only consider modules that are totally reflexive. Note also that the ring
R is Gorenstein if and only every maximal Cohen-Macaulay R-module is totally
reflexive [AuBr]. In particular, if R is a zero-dimensional Gorenstein ring (i.e. if R
is selfinjective), then every R-module is totally reflexive.
Commutative local Gorenstein rings. Let R be a commutative local Gorenstein ring with maximal ideal m and Krull dimension d. As mentioned above, in
this case all maximal Cohen-Macaulay R-modules are totally reflexive. Recall that
a finitely generated R-module X is said to be free on the punctured spectrum of
R if Xp is a free Rp -module for all non-maximal primes ideals p of R. By [Bu,
7.7.5(i)], when X is free on the punctured spectrum of R we have for all i natural
isomorphisms
i
d−1−i
d R (X, −) → HomR (Ext
dR
η i : Ext
(−, X), E)
where E denotes the injective hull of the residue field R/ m of R. It follows that
we have a graded bilinear form
(1)
h−, −i : A ⊗R A → E[1 − d]
d i (X, X),
= ⊕i∈Z Ext
R
d−1−i
where A
and h−, −ii : Ai ⊗R Ad−1−i → E[1 − d] is defined by
i
hf, gi = η
(g)(f ).
The following result is a special case of [Bu, 7.7.5(iii)].
Proposition 4.1. Let R be a commutative local Gorenstein ring of dimension
d, and X be a maximal Cohen-Macualay R-module which is free on the punctured
spectrum of R. Then the graded bilinear form (1) is non-degenerate and associative.
We now make statements in terms of regular sequences as in the previous section.
However, we encounter a dichotomy due to (2) of Theorem 2.8, since the index of
duality n = d − 1 may be non-negative. We treat the d = 0 and d > 0 cases
separately, starting with the former. In this case all R-modules are free on the
punctured spectrum vacuously.
Theorem 4.2. Let (R, m) be a zero-dimensional commutative local Gorenstein
d ∗ (X, X) of the finitely generated
ring, and denote by A the Tate cohomology ring Ext
R
R-module X. If A≥0 contains a regular sequence of length two of central elements
of positive degree, then
A∼
= A≥0 ⋉ A<0 .
Remark. In [AvVe] additional structure is specified in the special case of M = R/ m,
but for any dimension.
12
PETTER ANDREAS BERGH, DAVID A. JORGENSEN & STEFFEN OPPERMANN
Theorem 4.3. Let (R, m) be a commutative local Gorenstein ring of dimension
∗
d R (X, X) of the finitely generd > 0, and denote by A the Tate cohomology ring Ext
ated maximal Cohen-Macaulay R-module X which is free on the punctured spectrum
of R. Then A≥0 contains at most a regular sequence of length one of central homogeneous elements of positive degree.
Proof. The result follows from Proposition 4.1 and (3) of Theorem 2.8.
0
d (X, X)
Ext
R
i
d (X, X)
Ext
R
Since ExtiR (X, X) ∼
for all i > 0, and
is a quotient of
=
Ext0R (X, X) when X is a maximal Cohen-Macaulay R-module, we get the following
corollary involving the absolute Ext algebra.
Corollary 4.4. Let (R, m) be a commutative local Gorenstein ring of Krull dimension d > 0, and X a finitely generated maximal Cohen-Macaulay R-module which is
free on the punctured spectrum of R. Then Ext∗R (X, X) contains at most a regular
sequence of length one of central homogeneous elements of positive degree.
When (R, m) is a commutative local Gorenstein ring of positive Krull dimension,
then the residue field k = R/ m is never a maximal Cohen-Macaulay R-module. In
contrast to the above corollary, Lemma 8.3 of [AvVe] shows that if R is moreover
a complete intersection, then the absolute Ext algebra, Ext∗R (k, k), does have a
regular sequence of central elements of length equal to the codimension of R. Hence
when the codimension is at least two (i.e. when R is not a hypersurface) then the
∗
d R (k, k) is zero. We illustrate this
product of two negative degree elements in Ext
(and Theorem 4.2) with an example, using the same ring as in the example in
Section 3.
Example. Let c be a positive integer and R the commutative local zero-dimensional
complete intersection
k[x1 , . . . , xc ]/(xa1 1 , . . . , xac c ),
where a1 , . . . , ac are integers, all at least 2. It follows from [BerOp, Theorem 5.3]
that the Ext-algebra of k is given by
Ext∗R (k, k) = khy1 , . . . , yc , z1 , . . . , zc i/I,
where I is the ideal in khy1 , . . . , yc , z1 , . . . , zc i generated by the following relations:
zi zj − zj zi
zi yj − yj zi
yi yj + yj yi
yi2 − zi
yi2
for i 6= j
if ai = 2
if ai 6= 2.
The degrees of the homogeneous generators are given by |yi | = 1 and |zi | = 2. For
example, when all the exponents ai are at least 3, then
Ext∗R (k, k) = ∧∗ (y1 , . . . , yc ) ⊗k k[z1 , . . . , zc ],
where ∧∗ (y1 , . . . , yc ) is the exterior algebra on the generators y1 , . . . , yc .
The depth of Ext∗R (k, k) is c, and the sequence z1 , . . . , zc is a maximal regular
∗
d R (k, k)
sequence in Ext∗R (k, k). When c = 1, then k is a periodic module, and so Ext
contains lots of nonzero products of elements of negative degrees. When c ≥ 2, then
∗
d R (k, k)
Theorem 4.2 shows that the product of two negative degree elements in Ext
is always zero.
THE NEGATIVE SIDE OF COHOMOLOGY FOR CALABI-YAU CATEGORIES
13
General non-degeneracy. Finally, we include a discussion and result which explains in a transparent manner the non-degeneracy of the graded bilinear form
∗
∗
d (R/ m, R/ m) ⊗R Ext
d (R/ m, R/ m) → R[1].
h−, −i : Ext
R
R
in the case where R is a zero-dimensional commutative Gorenstein ring.
Let I be an ideal of the commutative local ring (R, m), and assume that R/I
has a complete resolution of the form
∂
∂
s
2
1
· · · → Rb2 −→
Rb1 −→
R−
→ R → ···
T :
so that s factors through the natural map R → R/I. Furthermore, assume that M
is a totally reflexive R-module whose complete resolution U satisfies Im ∂iU ⊆ IUi−1
for all i. We define a non-degenerate pairing
−i−1
d
Ext
R
i
d (R/I, X) → R/I
(X, R/I) ⊗R Ext
R
d i (R/I, X). Assume that η is repreas follows. Choose a nonzero element η ∈ Ext
R
sented by the top vertical map in the following diagram:
s
R
/R
a1
U
∂−i
U−i
..
.
an
/ U−i−1
ej
R
s
/R
Since η is assumed to be nonzero, aj ∈
/ I for some j. Letting ej denote the unit row
matrix with 1 in the jth component, we see that the composition of the vertical
maps is just multiplication by aj . From the assumption on the differentials of
d −i−1 (X, R/I).
U , it follows that this map ej corresponds to an element θ ∈ Ext
R
Thus we obtain a non-degenerate pairing by setting hθ, ηi = aj ∈ R/I. A similar
−i−1
d
argument holds if we take η ∈ Ext
(X, R/I) nonzero. Moreover, this pairing is
R
d i (R/I, X), φ ∈ Ext
d j (X, X) and
obviously associative in the sense that, if η ∈ Ext
R
R
−i−j−1
dR
θ ∈ Ext
(X, R/I), then hθφ, ηi = hθ, φηi. From Section 2 we therefore obtain
the following result.
Theorem 4.5. Let R be a commutative local ring, I an ideal, and assume that R/I
has a complete resolution of the form
∂
∂
∂
0
2
1
· · · → Rb2 −→
R → ···
Rb1 −→
R −→
with Im ∂i ⊆ IRbi−1 for all i. Furthermore, denote by A the Tate cohomology ring
d ∗ (R/I, R/I), and assume that A≥0 contains a regular sequence of length two of
Ext
R
central homogeneous elements of positive degree. Then
A∼
= A≥0 ⋉ A<0 .
14
PETTER ANDREAS BERGH, DAVID A. JORGENSEN & STEFFEN OPPERMANN
References
[Au]
M. Auslander, Isolated singularities and the existence of almost split sequences, Proc.
ICRA IV, Springer Lect. Notes in Math. 1178 (1986), 194-241.
[AuBr]
M. Auslander, M. Bridger, Stable module theory, Memoirs of the A.M.S. 94 (1969),
American Mathematical Society, Providence, RI.
[AuReSm] M. Auslander, I. Reiten, S.O. Smalø, Representation theory of Artin algebras, Cambridge Studies in Advanced Mathematics, 36, Cambridge University Press, Cambridge,
1995, xiv+423 pp.
[AvVe]
L.L. Avramov, O. Veliche, Stable cohomology over local rings, Adv. Math. 213 (2007),
no. 1, 93-139.
[BenCa]
D.J. Benson, J.F. Carlson, Products in negative cohomology, J. Pure Appl. Algebra 82
(1992), no. 2, 107-129.
[BerJo]
P.A. Bergh, D.A. Jorgensen, Tate-Hochschild homology and cohomology of Frobenius
algebras, to appear in J. Noncommut. Geom.
[BerOp]
P.A. Bergh, S. Oppermann, Cohomology of twisted tensor products, J. Algebra 320
(2008), no. 8, 3327-3338.
[BoKa]
A.I. Bondal, M.M Kapranov, Representable functors, Serre functors, and reconstructions, (Russian) Izv. Akad. Nauk SSSR Ser. Mat. 53 (1989), no. 6, 1183-1205, 1337;
translation in Math. USSR-Izv. 35 (1990), no. 3, 519-541.
[Bu]
R.-O. Buchweitz, Maximal Cohen-Macaulay modules and Tate-cohomology over
Gorenstein rings, 1987, 155 pp; see
https://tspace.library.utoronto.ca/handle/1807/16682.
[Ho]
T. Holm, Hochschild cohomology rings of algebras k[X]/(f ), Beiträge Algebra Geom.
41 (2000), no. 1, 291-301.
[HuJo]
C. Huneke, D.A. Jorgensen, Symmetry in the vanishing of Ext over Gorenstein rings,
Math. Scand. 93 (2003), 161–184.
[Ke]
B. Keller, Calabi-Yau triangulated categories, in Trends in representation theory of
algebras and related topics, 467-489, EMS Ser. Congr. Rep., Eur. Math. Soc., Zürich,
2008.
[KrLe]
H. Krause, J. Le, The Auslander-Reiten formula for complexes of modules, Adv. Math.
207 (2006), no. 1, 133-148.
Petter Andreas Bergh, Institutt for matematiske fag, NTNU, N-7491 Trondheim,
Norway
E-mail address: [email protected]
David A. Jorgensen, Department of mathematics, University of Texas at Arlington,
Arlington, TX 76019, USA
E-mail address: [email protected]
Steffen Oppermann, Institutt for matematiske fag, NTNU, N-7491 Trondheim, Norway
E-mail address: [email protected]
| 0math.AC
|
Subgroups of 3-factor direct products
arXiv:1607.03444v1 [math.GR] 12 Jul 2016
Daniel Neuen and Pascal Schweitzer
RWTH Aachen University
{neuen,schweitzer}@informatik.rwth-aachen.de
July 13, 2016
Extending Goursat’s Lemma we investigate the structure of subdirect products of
3-factor direct products. We give several example constructions and then provide a
structure theorem showing that every such group is essentially obtained by a combination of the constructions. The central observation in this structure theorem is that
the dependencies among the group elements in the subdirect product that involve all
three factors are of Abelian nature. In the spirit of Goursat’s Lemma, for two special
cases, we derive correspondence theorems between data obtained from the subgroup
lattices of the three factors (as well as isomorphism between arising factor groups)
and the subdirect products. Using our results we derive an explicit formula to count
the number of subdirect products of the direct product of three symmetric groups.
1 Introduction
The lemma of Goursat [7] is a classic result of group theory that characterizes the subgroups of
a direct product of two groups G1 × G2 . A version of the lemma also provides means to describe
the subgroups of G1 × G2 by inspecting the subgroup lattices of G1 and G2 and considering
isomorphisms between arising factor groups.
In an expository article, Anderson and Camillo [1] demonstrate for example the applicability of
Goursat’s lemma to determine normal subgroups of G1 × G2 , to count the number of subgroups
of S3 × S3 , and to prove the Zassenhaus Lemma. They also describe how Goursat’s Lemma can
be stated in the context of rings, ideals, subrings and in modules. The Lemma itself can also be
found in various introductory algebra and group theory texts (e.g., [8, 10]).
While Goursat’s Lemma applies to subgroups of the direct product of two groups, in this work
we are concerned with subgroups of the direct product of three groups.
It seems that there is no straightforward generalization to three factors. Indeed, Bauer, Sen,
and Zvengrowski [2] develop a generalization to an arbitrary finite number of factors by devising
a non-symmetric version of Goursat’s lemma for two factors that can then be applied recursively.
A more category theory focused approach is taken by Gekas. However no simple correspondence
theorem between the subdirect products of 3-factor direct products and data depending on the
sublattice of the subgroups of the factors and isomorphisms between them is at hand. In fact,
in [2] the authors exhibit two Abelian examples that stand in the way of such a correspondence
theorem by sharing the various characteristic subgroups and isomorphisms between them and yet
being distinct. Both these papers are able to recover several identities provided by Remak [13]
who is explicitly concerned with 3-factor subdirect products.
1
In this paper we analyze the structure of subdirect products of 3-factor direct products. To this
end we give several example constructions of such groups and then provide a structure theorem
showing that every such group is essentially obtained by a combination of the constructions. The
central observation in this structure theorem is that the dependencies among the group elements
in the subdirect product that involve all three factors are of Abelian nature. We call a subdirect
product of G1 × G2 × G3 2-factor injective if each of the three projections onto two factors is
injective. By taking suitable quotients, it is possible to restrict our investigations to 2-factor
injective subdirect products, for which we obtain the following theorem.
Theorem 1.1 (Characterization of 2-factor injective subdirect products of 3-factor products).
Let ∆ ≤ G1 ×G2 ×G3 be a 2-factor injective subdirect product. Then there is a normal subgroup
of H E ∆ with [πi (∆) : πi (H)] = [∆ : H] for i ∈ {1, 2, 3} and H is isomorphic to a group of the
following form: there are three finite groups H1 , H2 , H3 that all have an Abelian subgroup M
contained in their the center such that H is isomorphic to the factor group of
{((h2 , h′3
−1
), (h3 , h′1
−1
), (h1 , h′2
−1
)) | hi , h′i ∈ Hi , hi h′i
−1
∈ M, h1 h′1
−1
h2 h′2
−1
h3 h′3
−1
= 1},
by the normal subgroup {((m1 , m1 ), (m2 , m2 ), (m3 , m3 )) | mi ∈ M }.
In the spirit of Goursat’s Lemma we then investigate correspondence theorems between data
obtained from the subgroup lattices of the Gi (as well as isomorphism between arising factor
groups) and the subdirect products of G1 × G2 × G3 . For two special cases, namely the cases H =
∆, and M = {1}, we obtain such a correspondence theorem for three factors. Here the second
case is a particular special case hinted at in [2], which is indeed describable by a symmetric
version of a generalized Goursat’s Lemma. In a third special case, where one of the Gi is the
semi-direct product of the projection of H onto the i-th component and some other group, we
also obtain a partial correspondence theorem.
As demonstrated by Petrillo [12], Goursat Lemma can readily be applied to count subgroups
of the product of two Abelian groups. For a direct product of an arbitrary number of Abelian
groups the number of subgroups has been extensively studied. We refer to the monograph of
Butler [5]. In fact there are also explicit formulas for the counts of subgroups of Zp × Zq × Zr
(see for example [9]). In line with the papers and as an application of our characterization, we
derive an explicit formula to count the number of subdirect products of the direct product of
three symmetric groups Sn1 × Sn2 × Sn3 . It is also possible for example to count the normal
subgroups of such direct products. In fact, the normal subgroups can be also characterized for
arbitrary finite products of symmetric groups [11]. Let us finally also point to some literature
concerning finiteness properties of groups [3, 4] which also contains some structural results on
3-factor direct products (in particular on the case we call 3-factor surjective).
2 Goursat’s Lemma
Let G = G1 × G2 × · · · × Gt be a direct product of groups. We define for i ∈ {1, . . . , t} the
map πi as the projection to the i-th coordinate and we define the homomorphism ψi : ∆ →
G1 × · · · × Gi−1 × Gi+1 × · · · × Gt : (g1 , g2 , . . . , gt ) 7→ (g1 , . . . , gi−1 , gi+1 , . . . , gt ).
A subgroup ∆ ≤ G of the direct product is said to be a subdirect product if πi (∆) = Gi for
all 1 ≤ i ≤ t. Goursat’s Lemma is a classic statement concerned with the structure of subdirect
products of direct products of two factors.
Theorem 2.1 (Goursat’s Lemma). Let ∆ ≤ G1 × G2 = G be a subdirect product and define N1 = {g1 ∈ G1 | (g1 , 1) ∈ ∆} as well as N2 = {g2 ∈ G1 | (1, g2 ) ∈ ∆}. Then G1 /N1 is
isomorphic to G/N2 via an isomorphism ϕ for which (g1 , g2 ) ∈ ∆ if and only if ϕ(g1 ) = g2 .
2
This gives a natural homomorphism ∆ → G1 /N1 × G2 /N2 defined as (g1 , g2 ) 7→ (g1 N1 , g2 N2 )
with image {(g1 , g2 ) | ϕ(g1 ) = g2 }. Thus we can view ∆ as a fiber product (or pull back) of G1
and G2 over Gi /Ni .
A typical application of the lemma is a proof of the fact that subdirect products of non-Abelian
finite simple groups are direct products of diagonal subgroups. Furthermore, the lemma can be
applied to count subgroups of direct products. For this, the following correspondence version of
the lemma is more convenient.
Theorem 2.2. There is a natural one-to-one correspondence between the subgroups of G1 × G2
and the tuples (P1 , P2 , N1 , N2 , ϕ) for which for i ∈ {1, 2} we have
1. Ni E Pi ≤ Gi and
ϕ
2. P1 /N1 ∼
= P2 /N2 .
ϕ
Here, we write G1 ∼
= G2 to denote that G1 and G2 are isomorphic via an isomorphism ϕ.
The subdirect products correspond to those tuples for which P1 = G1 and P2 = G2 . Diagonal
subgroups are those subdirect products that also satisfy N1 = N2 = 1. Subproducts are those
for which N1 = P1 and N2 = P2 .
3 Three factors
We now focus on 3-factor subdirect products. Before we investigate the general case, we consider
four examples of subdirect products.
In our first examples, we consider groups that are 2-factor surjective. We say ∆ ≤ G1 ×G2 ×G3
is 2-factor surjective if ψi is surjective for all 1 ≤ i ≤ 3. Note that the analogous definition of 1factor surjectivity (i.e., all πi are surjective) means then the same as being subdirect.
Similarly, we say ∆ is 2-factor injective if ψi is injective for all 1 ≤ i ≤ 3. Note that this
assumption is equivalent to saying that two components of an element of ∆ determine the third.
Analogously 1-factor injective then means that one component determines the other two.
3.1 Examples of 3-factor direct products
Example 3.1. The subgroup of (G1 )3 comprised of the set {(g, g, g) | g ∈ G1 )} is called the
diagonal subgroup.
It is not difficult to see that the only 1-factor injective subdirect products are diagonal subgroups. As a second example, let G1 be an Abelian group. Then the group
Example 3.2.
∆ := {(a, b, c) ∈ (G1 )3 | abc = 1}.
(3.1)
is a subdirect product of (G1 )3 that is 2-factor surjective and 2-factor injective.
It turns out that this is the only type of group with these properties.
Lemma 3.3. Let G = G1 × G2 × G3 be a group and ∆ a subdirect product of G that is 2-factor
surjective and 2-factor injective. Then G1 , G2 and G3 are isomorphic Abelian groups and ∆
is isomorphic to the subgroup of G31 given by {(a, b, c) ∈ (G1 )3 | abc = 1}, which in turn is
isomorphic to (G1 )2 as an abstract group.
3
Proof. Let g1 and g1′ be elements of G1 . Then by 2-factor surjectivity there are elements g2 and g3
′
g′
such that (g1 , g2 , 1) ∈ ∆ and (g1′ , 1, g3 ) ∈ ∆. We thus have that (g1 , g2 , 1)(g1 ,1,g3 ) = (g11 , g2 , 1).
g′
By 2-factor injectivity it follows that g11 = g1 and thus G1 is Abelian.
For every g1 ∈ G1 there is exactly one element of the form (g1 , g2 , 1) in ∆. The map ϕ that
sends every g1 to the corresponding g2 provides us with a map from G1 to G2 . By 2-factor
surjectivity and 2-factor injectivity, this map is an isomorphism from G1 to G2 . Similarly, G1
and G3 are isomorphic.
Finally, note that the map that sends (g1 , g2 , g3 ) to (g1 , ϕ−1 (g2 )−1 , g1−1 ϕ−1 (g2 )) is an isomorphism from ∆ to {(a, b, c) ∈ (G1 )3 | abc = 1}.
We now drop the requirement for the group to be 2-factor surjective. Our next examples of
2-factor injective subdirect products will be non-Abelian.
Example 3.4. Let G1 = H ⋊ K be a semidirect product with an Abelian normal subgroup H.
Then
∆ = {(ak, bk, ck) ∈ (G1 )3 | a, b, c ∈ H, k ∈ K, abc = 1}.
(3.2)
is a 2-factor injective subdirect product of (G1 )3 . To see this we verify that ∆ is closed under
multiplication. Let d = (ak, bk, ck), d′ = (a′ k ′ , b′ k ′ , c′ k ′ ) ∈ ∆. Then
dd′ = (ak, bk, ck)(a′ k ′ , b′ k ′ , c′ k ′ )
= (aka′ k ′ , bkb′ k ′ , ckc′ k ′ )
= (a(ka′ k −1 )kk ′ , b(kb′ k −1 )kk ′ , c(kc′ k −1 )kk ′ )
and a(ka′ k −1 )b(kb′ k −1 )c(kc′ k −1 ) = (abc)k(a′ b′ c′ )k −1 = 1 implying dd′ ∈ ∆. So ∆ ≤ (G1 )3 . The
fact that ∆ is a subdirect product and 2-factor injective follows directly from the definition.
Example 3.5. As a next example suppose G1 = H2 × H3 , G2 = H1 × H3 and G3 = H1 × H2
with arbitrary finite groups Hi . Then the group consisting of the set
{((h2 , h3 ), (h1 , h3 ), (h1 , h2 )) | hi ∈ Hi }
is a 2-factor injective subdirect product of G1 × G2 × G3 .
Finally, it is not difficult to construct subdirect products that are not 2-factor injective by
considering extensions of the factors.
f1 be a surjective
Example 3.6. Let ∆ ≤ G1 × G2 × G3 be a subdirect product and let G
f
f
homomorphism κ : G1 → G1 . Then {(g1 , g2 , g3 ) ∈ G1 × G2 × G3 | (κ(g1 ), g2 , g3 ) ∈ ∆} is a
f1 × G2 × G3 that is not 2-factor injective if κ is not injective.
subdirect product of G
3.2 The structure of subgroups of 3-factor direct products
We now analyze the general case, showing that it must essentially be a combination of the
examples presented above. We first argue that we can focus our attention on 2-factor injective
subdirect products.
Lemma 3.7. Let ∆ ≤ G1 ×G2 ×G3 be a subdirect product. Let Ni = πi (ker(ψi )) for i ∈ {1, 2, 3}.
Then ∆′ = ∆/(N1 × N2 × N3 ) is a 2-factor injective subdirect product and ∆ = {(g1 , g2 , g3 ) |
(g1 N1 , g2 N2 , g3 N3 ) ∈ ∆′ }.
4
Thus in the following suppose ∆ is a 2-factor injective subdirect product of G1 × G2 × G3 .
Let Hi = ker(πi ) ∩ ∆ = {(g1 , g2 , g3 ) ∈ ∆ | gi = 1}. Then H = hH1 , H2 , H3 i is a normal
subgroup of ∆.
Lemma 3.8. For i, j ∈ {1, 2, 3} with i 6= j we have [Hi , Hj ] = 1 that is, all elements in Hi
commute with all elements in Hj .
Proof. Without loss of generality assume i = 1 and j = 2. For (1, g2 , g3 ) ∈ H1 and (h1 , 1, h3 )
in H2 we get that (1, g2 , g3 )(h1 ,1,h3 ) = (1, g2 , g3h3 ). By 2-factor injectivity we conclude that g3h3 =
g3 and thus the two elements commute.
Define Mi := πi (Hk ) ∩ πi (Hj ), where j and k are chosen so that {i, j, k} = {1, 2, 3}.
Lemma 3.9. Let i, j, k be integers such that {i, j, k} = {1, 2, 3}. Then there is a canonical
isomorphism ϕ := ϕij,k from πj (Hi ) to πk (Hi ) that maps Mj to Mk .
Proof. Assume without loss of generality that i = 1, j = 2 and k = 3. Define a map ϕ : π2 (H1 ) →
π3 (H1 ) such that (1, g2 , ϕ(g2 )−1 ) ∈ ∆ for all g2 ∈ π2 (H1 ). Such a map exists and is well defined
since ∆ is a 2-factor injective subdirect product. Suppose g2 ∈ M2 then (1, g2 , ϕ(g2 )−1 ) ∈ ∆
and there is a g1 such that (g1 , g2 , 1) ∈ ∆. Then (1, g2 , ϕ(g2 )−1 )(g1 , g2 , 1)−1 = (g1−1 , 1, ϕ(g2 )−1 )
so ϕ(g2 ) ∈ M3 . It follows by symmetry that all Mi are isomorphic and that ϕ|M2 is an isomorphism from M2 to M3 .
Note that the canonical isomorphisms behave well with respect to composition. In particular
we have ϕij,k = (ϕik,j )−1 and ϕij,k |Mj ◦ ϕjk,i |Mk = ϕkj,i |Mj . This implies for example that the composition of the canonical isomorphism from M1 to M2 and the canonical isomorphism from M2
to M3 is exactly the canonical isomorphism from M1 to M3 . We can thus canonically identify
the subgroups M1 , M2 and M3 with a fixed subgroup M .
Moreover, we can canonically associate the elements in H1 with the elements in π2 (H1 ) and
with the elements in π3 (H1 ) by associating (1, g2 , ϕ(g2 )−1 ) ∈ H1 with the element g2 ∈ G2
and ϕ(g2 ) in G3 . Similarly we can associate elements (ϕ(g3 )−1 , 1, g3 ) ∈ H2 with g3 ∈ G3
and ϕ(g3 ) ∈ G1 and also associate (g1 , ϕ(g1 )−1 , 1) ∈ H3 with g1 ∈ G1 and ϕ(g1 ) ∈ G2 .
Theorem 1.1 (restated). Let ∆ ≤ G1 × G2 × G3 be a 2-factor injective subdirect product.
Then there is a normal subgroup of H E ∆ with [πi (∆) : πi (H)] = [∆ : H] for i ∈ {1, 2, 3} and H
is isomorphic to a group of the following form: there are three finite groups H1 , H2 , H3 that all
have an Abelian subgroup M contained in their the center such that H is isomorphic to the
factor group of
{((h2 , h′3
−1
), (h3 , h′1
−1
), (h1 , h′2
−1
)) | hi , h′i ∈ Hi , hi h′i
−1
∈ M, h1 h′1
−1
h2 h′2
−1
h3 h′3
−1
= 1},
by the normal subgroup {((m1 , m1 ), (m2 , m2 ), (m3 , m3 )) | mi ∈ M }.
Proof. As before define Hi = ker(πi ) = {(g1 , g2 , g3 ) ∈ ∆ | gi = 1} and H = hH1 , H2 , H3 i. By
Lemma 3.9 and the comment about the compatibility of the isomorphism between the groups
we can canonically associate the elements of Mi with those of Mj . Moreover we can assume that
there is an Abelian group M that is isomorphic to the intersection of every pair of {H1 , H2 , H3 }.
All elements of H commute with all elements in such an intersection.
If (g1 , g2 , g3 ) is an element of H then gi can be written as ci · b−1
with ci ∈ Ci := πi (Hi+1 )
i
and bi ∈ Bi := πi (Hi+2 ), where indices will always be taken modulo 3. We can set (g1 , g2 , g3 ) =
−1
−1
−1
2
(c1 b−1
1 , c2 b2 , c3 b3 ). Since c1 ∈ π1 (H2 ) we conclude that (c1 , 1, ϕ1,3 (c1 )) ∈ H. We also
3
−1
have (b1 , ϕ1,2 (b1 ) , 1) ∈ H. This implies that
5
−1
2
3
(g1 , g2 , g3 )(c−1
), 1) = (1, c2 b2 −1 ϕ31,2 (b1 −1 ), c3 b3 −1 ϕ21,3 (c1 )) ∈ H.
1 , 1, ϕ1,3 (c1 ))(b1 , ϕ1,2 (b1
We thus see by looking at the third component that c3 b3 −1 ϕ21,3 (c1 ) ∈ π3 (H1 ) which implies
that b3 −1 ϕ21,3 (c1 ) ∈ π3 (H1 ) and thus b3 −1 ϕ21,3 (c1 ) ∈ M3 . By symmetry we conclude that
bi+1 −1 ϕii−1,i+1 (ci−1 ) ∈ Mi+1 for i ∈ {1, 2, 3}.
(3.3)
Looking at the second component, we also see that c2 b2 −1 ϕ31,2 (b1 −1 ) ∈ π2 (H1 ) and conclude
that ϕ12,3 (c2 b2 −1 ϕ31,2 (b1 −1 ))−1 = c3 b3 −1 ϕ21,3 (c1 ). Recalling that Hi and Hj commute for i 6= j,
we thus conclude that
−1
c3 ϕ12,3 (b2 )
· ϕ21,3 (c1 )b3 −1 · ϕ12,3 (c2 ϕ31,2 (b1 −1 )) = 1.
(3.4)
Thus, since all involved isomorphisms are compatible we can reinterpret this equation in M and
−1
we see that c3 (b2 ) · c1 (b3 )−1 · c2 (b1 )−1 = 1.
−1
−1
−1
Suppose that (g1 , g2 , g3 ) = (cb1 bb1 , cb2 bb2 , cb3 bb3 ) is a second representation of the ele−1
= c1 b1 −1 and
ment (g1 , g2 , g3 ) of H. Then we have for the first component that cb1 bb1
−1
−1 b
for some eleb1 ∈ M1 or, in other words, cb1 m1 = c1 and bb1 m1 = b−1
thus c−1
1
1 cb1 = b1
ment m1 ∈ M . Similarly there are elements m2 and m3 for the other components. We conclude
that the map sending (g1 , g2 , g3 ) to ((c1 , b1 −1 ), (c2 , b2 −1 ), (c3 , b3 −1 )) is a homomorphism from H
to the group described in the theorem.
It remains to show injectivity of this homomorphism. Suppose (g1 , g2 , g3 ) is mapped to the
trivial element. This implies for the first component of the image ((c1 , b1 −1 ), (c2 , b2 −1 ), (c3 , b3 −1 ))
that (c1 , b1 −1 ) = (m1 , m1 ) for some m1 implying that g1 = m1 m1 −1 = 1. Repeating the same
argument for the other components we see that the homomorphism is an isomorphism.
3.3 Correspondence theorems
We now investigate the possibility of having a correspondence theorem in the style of Theorem 2.2
for 3 factors. As before, we can readily reduce to the case of 2-factor injective subdirect products.
Lemma 3.10. There is a natural one-to-one correspondence between subdirect products of G1 ×
G2 × G3 and the tuples (N1 , N2 , N3 , ∆′ ), where Ni E Gi for every i ∈ {1, 2, 3} and ∆′ is a 2-factor
injective subdirect product of G1 /N1 × G2 /N2 × G3 /N3 .
Proof. Let ∆ ≤ G1 × G2 × G3 be a subdirect product. Choose Ni = πi (ker(ψi )) for i ∈ {1, 2, 3}
and let ∆′ = ∆/(N1 × N2 × N3 ). Then Ni E Gi , because ∆ is a subdirect product, and ∆′ is
2-factor injective by Lemma 3.7.
Conversely, let Ni EGi and let ∆′ be a 2-factor injective subdirect product of G1 /N1 ×G2 /N2 ×
G3 /N3 . Then ∆ = {(g1 , g2 , g3 ) | (g1 N1 , g2 N2 , g3 N3 ) ∈ ∆′ } is subdirect product of G1 × G2 ×
G3 .
Suppose ∆ is a 2-factor injective subdirect product. Then we can define for i ∈ {1, 2, 3}
the groups Hi = {(g1 , g2 , g3 ) ∈ ∆ | gi = 1} and with them the groups Bi = πi (Hi+2 ) and Ci =
πi (Hi+1 ). As we will see, the canonical isomorphism ϕi := ϕi+2
i,i+1 that exists by Lemma 3.9 can be
extended to an isomorphism from Gi /Ci to Gi+1 /Bi+1 . We would like to have a correspondence
theorem in the style of Theorem 2.2 for 3 factors. For this, in principle, we would like to relate
the 2-factor injective subdirect products of G1 , G2 , G3 to the tuples
(B1 , B2 , B3 , C1 , C2 , C3 , ϕ1 , ϕ2 , ϕ3 )
6
that satisfy certain consistency properties. However, in general, neither is it clear which consistency properties to choose so that every tuple corresponds to a subdirect product, nor do distinct
subdirect products always correspond to distinct tuples. Indeed, in [2] the authors describe two
distinct Abelian subdirect products of the same group G1 × G2 × G3 for which the corresponding
tuples agree.
In the light of that we content ourselves with study two special cases, namely those where ∆ =
H and those where Mi = Bi ∩ Ci = 1.
Theorem 3.11. There is a natural one-to-one correspondence between subdirect products
of ∆ ≤ G1 × G2 × G3 which are 2-factor injective satisfying ∆ = H, and tuples of the
form (B1 , B2 , B3 , C1 , C2 , C3 , ϕ1 , ϕ2 , ϕ3 ) for which for all i ∈ {1, 2, 3} (indices taken modulo 3)
we have
1. Bi , Ci E Gi ,
2. Bi Ci = Gi ,
ϕi
3. Bi ∼
= Ci+1 ,
4. [Bi , Ci ] = 1,
5. ϕi (Bi ∩ Ci ) = Bi+1 ∩ Ci+1 ,
6. ϕ3 |B3 ∩C3 ◦ ϕ2 |B2 ∩C2 ◦ ϕ1 |B1 ∩C1 = id.
Proof. Let ∆ ≤ G1 × G2 × G3 be a subdirect product. Define Hi := {(g1 , g2 , g3 ) ∈ ∆ | gi = 1}
for i ∈ {1, 2, 3} and H = hH1 , H2 , H3 i. Suppose that H = ∆. For i ∈ {1, 2, 3}, define Bi =
πi (Hi+2 ) and Ci = πi (Hi+1 ). Clearly Bi , Ci E Gi . By Lemma 3.8 we get that [Bi , Ci ] = 1
and Bi Ci = πi (H). The assumption ∆ = H implies Bi Ci = Gi . By Lemma 3.9 the groups
Bi and Ci+1 are isomorphic via an isomorphism ϕi = ϕi+2
i,i+1 , which maps Mi = Bi ∩ Ci to
Mi+1 = Bi+1 ∩ Ci+1 . Finally Property 6 follows directly from the comment below Lemma 3.9.
This gives us the tuple (B1 , B2 , B3 , C1 , C2 , C3 , ϕ1 , ϕ2 , ϕ3 ) with the desired properties.
On the other hand suppose we are given a tuple (B1 , B2 , B3 , C1 , C2 , C3 , ϕ1 , ϕ2 , ϕ3 ) with the
desired properties. Let Mi = Bi ∩ Ci . Define
)
(
gi = ci bi−1 for bi ∈ Bi , ci ∈ Ci , ci+1 Mi+1 = ϕi (bi Mi )
.
∆ = (g1 , g2 , g3 ) ∈ G1 × G2 × G3
−1
−1
· ϕ2 (c2 ϕ1 (b1 −1 )) = 1
and c3 ϕ2 (b2 ) · ϕ−1
3 (c1 )b3
′
For i ∈ {1, 2, 3} suppose gi = ci b−1
= c′i b′−1
i
i . Then there is some mi ∈ Mi with bi = bi mi and
′
′
′
ci = ci mi . Hence, bi Mi = bi Mi and ci Mi = ci Mi . Furthermore, we have
c′3 ϕ2 (b′2 )
=
=
−1
′ ′
· ϕ−1
3 (c1 )b3
−1
· ϕ2 (c′2 ϕ1 (b′1
−1
))
−1 −1
−1
−1 −1
c3 m−1
· ϕ−1
3 ϕ2 (b2 m2 )
3 (c1 m1 )(b3 m3 )
−1
· ϕ2 (c2 ϕ1 (b1 −1 )),
c3 ϕ2 (b2 )−1 · ϕ−1
3 (c1 )b3
−1
−1
· ϕ2 (c2 m−1
2 ϕ1 ((b1 m1 )
))
so the membership in ∆ is independent from the representation of gi ∈ Gi . Also, it is easy to
check that ∆ is closed under multiplication because [Bi , Ci ] = 1. The group ∆ is a subdirect
−1
product, since for g1 = c1 b−1
we have (c1 b1−1 , ϕ1 (b1−1 )−1 b−1
1
2 , ϕ3 (c1 )ϕ2 (b2 )) ∈ ∆, and it can
be checked that the group is 2-factor injective. Define Hi = {(g1 , g2 , g3 ) ∈ ∆ | gi = 1} for
i ∈ {1, 2, 3} and H = hH1 , H2 , H3 i. Then Bi = πi (Hi+2 ) and Ci = πi (Hi+1 ), which means that
H = ∆ by Property 2 and Theorem 1.1. Finally, it can be checked that ϕi = ϕi+2
i,i+1 .
7
It remains to show that for each 2-factor injective subdirect product ∆ ≤ G1 × G2 × G3
with ∆ = H, the group ∆ is of the form described above. But this follows from Theorem 1.1,
Equation (3.3) and (3.4).
The previous theorem shows that for the subdirect products with ∆ = H we can devise a
correspondence theorem. As a second case, on the other end of the spectrum, we can also devise
a correspondence theorem if the Abelian part that interlinks the three components is trivial. In
fact this case corresponds to the case discussed by Bauer, Sen, and Zvengrowski [2, 5.1 Remark],
who already suspect that a theorem like the previous can be obtained. We remark that Example
3.5 from the previous section is of this form. In fact, we can already conclude from Theorem
3.11 that for every group ∆, where the Abelian part interlinking the components is trivial, the
group H essentially has the form of Example 3.5.
Definition 3.12. Let ∆ be a subdirect product of G1 × G2 × G3 . We say ∆ is degenerate if
πi (ker(πi+1 )) ∩ πi (ker(πi+2 )) = πi (ker(ψi )) (i.e. Mi = 1) for some, and thus every, i ∈ {1, 2, 3}.
Lemma 3.13. For i ∈ {1, 2, 3} let Bi , Ci E Gi , such that Bi ∩ Ci = 1 and [Bi , Ci ] = 1. Furtherϕi
∼ Gi+1 /Bi+1 and suppose ϕi (Bi Ci ) = Ci+1 Bi+1 and ϕ3 (ϕ2 (ϕ1 (g1 B1 C1 ))) =
more let Gi /Ci =
g1 B1 C1 for all g1 ∈ G1 . Define
∆ = {(g1 , g2 , g3 ) ∈ G1 × G2 × G3 | ϕi (gi Ci ) = gi+1 Bi+1 }.
(3.5)
Then πi (∆) = Gi and ∆ is a degenerate 2-factor injective subdirect product.
Proof. We first show, that ∆ is closed under multiplication. Let (g1 , g2 , g3 ), (g1′ , g2′ , g3′ ) ∈ ∆. Then
′
ϕi (gi gi′ Ci ) = ϕi (gi Ci )ϕi (gi′ Ci ) = gi+1 gi+1
Bi+1 for all i ∈ {1, 2, 3}, so (g1 g1′ , g2 g2′ , g3 g3′ ) ∈ ∆. Let
E1 = hB1 , C1 i and pick e1 ∈ E1 . The element e1 can uniquely be written as e1 = b1 c1 with
b1 ∈ B1 , c1 ∈ C1 . For each i ∈ {1, 2, 3} define ϕ∗i : Bi → Ci+1 with ϕ∗i (bi ) = ci+1 for the
−1
unique ci+1 ∈ Ci+1 with ϕi (bi Ci ) = ci+1 Bi+1 . Then (b1 c1 , b2 ϕ∗1 (b1 ), (ϕ∗3 ) (c1 )ϕ∗2 (b2 )) ∈ ∆. So
E1 ≤ π1 (∆). The argument for the other components is analogous. Now let n11 , . . . , n1t be a
1
transversal of E1 in G1 . Let n2i B2 = ϕ1 (n1i C1 ) and n3i C3 = ϕ−1
3 (ni B1 ) for i ∈ {1, . . . , t}. Then
2
3
2
2 2
ϕ2 (ni C2 ) ⊆ ni E3 and hence there is some bi ∈ B2 with ϕ2 (ni bi C2 ) = n31 B3 . So (n1i , n2i b2i , n3i ) ∈
∆ and G1 ≤ π1 (∆).
It remains to prove that ∆ is 2-factor injective. Let (g1 , g2 , g3 ) ∈ ∆ with g2 = g3 = 1. Then
g1 ∈ B1 , because ϕ3 (C3 ) = B1 = g1 B1 , and g1 ∈ C1 , because ϕ1 (g1 C1 ) = g2 B2 = B2 . So g1 = 1.
Again, the argument for the other components is analogous.
Lemma 3.14. Let ∆ be a 2-factor injective subdirect product of G1 × G2 × G3 . Furthermore,
let Hi = {(g1 , g2 , g3 ) ∈ ∆ | gi = 1} for i ∈ {1, 2, 3} and H = hH1 , H2 , H3 i. Define Bi = πi (Hi+2 )
and Ci = πi (Hi+1 ). Suppose Bi ∩ Ci = 1 for all i ∈ {1, 2, 3}.
ϕi
∼ Gi+1 /Bi+1 and ϕi (Bi Ci ) =
Then there are canonical isomorphisms ϕ1 , ϕ2 , ϕ3 with Gi /Ci =
Ci+1 Bi+1 such that ϕ3 (ϕ2 (ϕ1 (g1 B1 C1 ))) = g1 B1 C1 for all g1 ∈ G1 . Furthermore ∆ is given by
Equation (3.5).
Proof. For i ∈ {1, 2, 3} define a homomorphism ϕi : Gi /Ci → Gi+1 /Bi+1 by setting ϕi (gi Ci ) =
gi+1 Bi+1 if (g1 , g2 , g3 ) ∈ ∆ for some gi ∈ Gi . We first have to show that ϕi is well-defined.
Without loss of generality consider i = 1 and let (g1 , g2 , g3 ), (g1′ , g2′ , g3′ ) ∈ ∆ with g1 C1 = g1′ C1 .
Then there is a (c, 1, h2 ) ∈ ∆ with g1′ c = g1 . We obtain (g1′ , g2′ , g3′ )(c, 1, h2 )(g1 , g2 , g3 )−1 =
(1, g2′ g2−1 , g3′′ ) for some g3′′ ∈ G3 and hence, g2 B2 = g2′ B2 . So ϕi is well-defined. Since ∆
is a subdirect product, ϕi is a surjective homomorphism. Suppose ϕ1 (g1 C1 ) = B2 . Then
8
(g1 c1 , b2 , g3 ) ∈ ∆ for some c1 ∈ C1 , b2 ∈ B2 and g3 ∈ G3 . Also there is h3 ∈ G3 with (1, b2 , h3 ) ∈
ϕi
∼
∆ and hence, (g1 c1 , 1, g3 h−1
3 ) ∈ ∆ implying that g1 ∈ C1 . So Gi /Ci = Gi+1 /Bi+1 .
For every b1 ∈ B1 there is a c2 ∈ C2 with (b1 , c2 , 1) ∈ ∆ and ϕ1 (b1 C1 ) = c2 B2 ∈ C2 B2 . By
symmetry it follows that ϕi (Bi Ci ) = Ci+1 Bi+1 for all i ∈ {1, 2, 3}. Now let ∆′ be the group
defined in Equation (3.5). Clearly ∆ ≤ ∆′ by the definition of ϕi for i ∈ {1, 2, 3}. So let
(g1′ , g2′ , g3′ ) ∈ ∆′ . Since ∆ is subdirect there is a (g1′ , g2 , g3 ) ∈ ∆ with g2 B2 = g2′ B2 . So we can
assume that g2 = g2′ . But then, by 2-factor injectivity of ∆′ , we get that g3 = g3′ .
Finally for (g1 , g2 , g3 ) ∈ ∆ we have that ϕi (gi Bi Ci ) = ϕi (gi Ci )ϕi (Bi Ci ) = gi+1 Ci+1 Bi+1 =
gi+1 Bi+1 Ci+1 . So ϕ2 (ϕ1 (g1 B1 C1 )) = ϕ−1
3 (g1 B1 C1 ) for all g1 ∈ G1 .
Theorem 3.15. There is a natural one-to-one correspondence between degenerate 2-factor injective subdirect products of G1 × G2 × G3 and the tuples (B1 , B2 , B3 , C1 , C2 , C3 , ϕ1 , ϕ2 , ϕ3 ) for
which for all i ∈ {1, 2, 3} (indices taken modulo 3) we have
1. Bi , Ci E Gi ,
2. Bi ∩ Ci = 1,
3. [Bi , Ci ] = 1,
ϕi
4. Gi /Ci ∼
= Gi+1 /Bi+1 ,
5. ϕi (Bi Ci ) = Ci+1 Bi+1 ,
6. ϕ3 (ϕ2 (ϕ1 (g1 B1 C1 ))) = g1 B1 C1 for all g1 ∈ G1 .
Proof. The statement follows from Theorem 1.1, Lemma 3.13 and 3.14.
By combining Theorem 3.15 with Lemma 3.10 we obtain the following correspondence result
for degenerate subdirect products.
Corollary 3.16. There is a natural one-to-one correspondence between degenerate subdirect
products of G1 × G2 × G3 and the tuples (N1 , N2 , N3 , B1 , B2 , B3 , C1 , C2 , C3 , ϕ1 , ϕ2 , ϕ3 ) for which
for all i ∈ {1, 2, 3} (indices taken modulo 3) we have
1. Ni E Gi ,
2. Bi , Ci E Gi /Ni ,
3. Bi ∩ Ci = 1,
4. [Bi , Ci ] = 1,
ϕi
5. (Gi /Ni )/Ci ∼
= (Gi+1 /Ni+1 )/Bi+1 ,
6. ϕi (Bi Ci ) = Ci+1 Bi+1 ,
7. ϕ2 (ϕ1 (g1 B1 C1 )) = ϕ−1
3 (g1 B1 C1 ) for all g1 ∈ G1 /N1 .
We conclude with the particular case in which πi (∆) has a complement in Gi for some i ∈
{1, 2, 3}. Example 3.4 described in the previous section is of this form. For this case we only
obtain an injection to tuples, rather than a one-to-one correspondence. We will exploit having
this injection in the next section for small special cases in our analysis of subdirect products of
symmetric groups.
9
Theorem 3.17. Suppose G1 = E1 ⋊ K is a semidirect product. There is an injective mapping
from the set of 2-factor injective subdirect products ∆ of G1 × G2 × G3 with π1 (H) = E1
κ
∼ G3 and ι is an automorphism of G1 that fixes K as
and B1 = C1 to the tuples (κ, ι) where G2 =
a set. Moreover if (κ, ι) is in the image of this mapping then (κ, ι′ ) is also in the image for every
automorphism ι′ of G1 that fixes K as a set.
Proof. Let ∆ be a 2-factor injective subdirect product of G1 × G2 × G3 satisfying π1 (H) = E1
and B1 = C1 .
For every element g2 ∈ G2 there is exactly one element (k1 , g2 , g3 ) ∈ ∆ with k1 ∈ K. We
obtain a well defined isomorphism κ from G2 to G3 .
Suppose now that ∆′ is a second 2-factor injective subdirect product G1 × G2 × G3 satisfying π1 (H) = E1 and B1 = C1 for which we obtain the same isomorphism κ. Then we can construct an automorphism ι of G1 as follows. For every g2 ∈ G2 there is exactly one (k, g2 , κ(g2 )) ∈
∆. There is also an element of the form (k ′ , g2 , κ(g2 )) ∈ ∆′ . We define ιK : K → K so that
it maps k to k ′ , this gives us an automorphism ιK of K. For every e ∈ E1 there is an element (e, g2 , 1) ∈ ∆. There is also an element (e′ , g2 , 1) ∈ ∆′ and define the map ιE : E1 → E1 by
mapping e to e′ . Then the map ιE is an automorphism of E.
We claim that the map that sends e · k to ιE (e) · ιK (k) is an automorphism of G1 . To
see this suppose a = (e1 , h, 1)(k, g, κ(g)) and a = (e1 , h, 1)(k, g, κ(g)) are two elements in ∆.
Then ι(a) = (ιE (e1 ), h, 1)(ιK (k), g, κ(g)) and ι(a) = (ιE (e1 ), h, 1)(ιK (k), g, κ(g)) are elements
of ∆′ .
For the products we obtain that
g
aa = (e1 e1 k , hh , 1)(kk, gg, κ(g)κ(g))
and
g
ι(a)ι(a) = (ιE (e1 )ιE (e1 )ιK (k) , hh , 1)(ιK (k)ιK (k), gg, κ(g)κ(g)).
To conclude that ι is an isomorphism we now only need to argue that ιE (e1 )ιK (k) is equal
to ιE (e1 k ). However, this is the case since (e1 k , hg , 1) ∈ ∆ and (ιE (e1 )ιK (k) , hg , 1) ∈ ∆′ .
Now suppose that ∆ is a subdirect product with π1 (H) = B1 = C1 and let κ : G2 → G3
be defined as above. Let ι be an automorphism of G1 that fixes K then ∆′ = {(ι(g1 ), g2 , g3 ) |
(g1 , g2 , g3 ) ∈ ∆} is a subdirect product of G1 × G2 × G3 . If we apply the above construction for
the automorphism of G1 we reobtain ι. This shows that the construction of ι from ∆′ and the
construction of ∆′ from ι are inverses to one another.
Note that in the theorem, the isomorphism κ associated with a subdirect product is canonical
(it only depends on the choice of K) but the choice of ι is not.
4 Subdirect products of two or three symmetric groups
In this section we apply the correspondence theorems to count the subdirect products of the direct
product of three symmetric groups. We first reduce this problem to counting the number of 2factor injective subdirect products. For finite groups G1 , . . . , Gk let ℓ(G1 , . . . , Gk ) be the number
of subdirect products of G1 × · · · × Gk . Furthermore, for k = 3, we denote by ℓ2-inj (G1 , . . . , G3 )
the number of 2-factor injective subdirect products.
10
Lemma 4.1. Let G1 , G2 , G3 be finite non-trivial groups. Then
X
ℓ2-inj (G1 /N1 , G2 /N2 , G3 /N3 )
ℓ(G1 , G2 , G3 ) =
Ni EGi
= ℓ(G1 , G2 ) + ℓ(G2 , G3 ) + ℓ(G1 , G3 ) − 2 +
X
ℓ2-inj (G1 /N1 , G2 /N2 , G3 /N3 ).
Ni ⊳Gi
Proof. The first equality follows from the correspondence described in Lemma 3.10. The second
equality follows by noting that the direct product is counted three times, so 2 has to be subtracted.
We are now interested in the number ℓ(n1 , n2 , n3 ) := ℓ(Sn1 , Sn2 , Sn3 ), where Sni is the symmetric group of a set with ni elements. Recall, that every factor group of a symmetric group is
isomorphic to a symmetric group over another set. Thus, by the previous lemma, it suffices to
compute the numbers ℓ(n1 , n2 ) and ℓ2-inj (n1 , n2 , n3 ) := ℓ2-inj (Sn1 , Sn2 , Sn3 ).
We start by analyzing the situation for two factors.
Lemma 4.2. Let n1 , n2 ≥ 2. For the number ℓ(n1 , n2 ) of subdirect products of Sn1 × Sn2 we
have
2
if n1 6= n2 and {n1 , n2 } 6= {3, 4},
8
if {n1 , n2 } = {3, 4},
n ! + 2
if
n1 = n2 ∈
/ {2, 4, 6},
1
ℓ(n1 , n2 ) =
2
if n1 = n2 = 2,
n
!
+
8
if n1 = n2 = 4,
1
2n1 ! + 2 if n1 = n2 = 6.
Proof. We assume for our considerations that n1 ≥ n2 . Let (Sn1 , Sn2 , N1 , N2 , ϕ) be a tuple
corresponding to a subdirect product via the correspondence of Theorem 2.2.
• If N1 = 1 then N2 = 1 and n1 = n2 . The number of isomorphisms from Sn1 to Sn1 is
if n1 = 2
1
i(n1 ) = 2n! if n1 = 6 .
n!
otherwise
This corresponds to the number of possible choices for ϕ. (These are the diagonal subgroups.)
• If N1 = Sn1 then N2 = Sn2 . There is only one subgroup of this type. (This is the direct
product).
• If N1 = An1 (n1 ≥ 3) then N2 = An2 , since the only index 2 subgroup that a symmetric
group can have is the alternating group. There is only one subgroup of this type.
If n1 6= 4 then N1 ∈ {1, An1 , Sn1 }, and we already considered all these cases. Suppose now
that n1 = 4 and n2 ≤ 4. Then N1 ∈ {1, V, An1 , Sn1 }, where V is the Klein-four-group. Three of
cases are considered above.
• If N1 = V then N2 = V and n2 = 4 or N2 = 1 and n2 = 3. In either case there are 6
options for ϕ.
11
In the following we use “{{” and “}}” to denote multisets.
Lemma 4.3. Let n1 , n2 , n3 ≥ 2. For the number ℓ2-inj (n1 , n2 , n3 ) of 2-factor injective subdirect
products of Sn1 × Sn2 × Sn3 we have
(n1 !)2
2
2
(n1 !) + 2n1 !
2
(n1 !) + 6n1 !
ℓ2-inj (n1 , n2 , n3 ) = (2n1 !)2
1440
n!
144
0
if n1 = n2 = n3 ∈
/ {2, 3, 4, 6},
if n1 = n2 = n3 = 2,
if n1 = n2 = n3 = 3,
if n1 = n2 = n3 = 4,
if n1 = n2 = n3 = 6,
if {{n1 , n2 , n3 }} = {{2, 6, 6}}
if {{n1 , n2 , n3 }} = {{2, n, n}} for n ∈
/ {2, 6},
if {{n1 , n2 , n3 }} = {{3, 4, 4}}
otherwise.
Proof. Suppose without loss of generality that n1 ≥ n2 ≥ n3 . Let ∆ ≤ Sn1 × Sn2 × Sn3 be a
2-factor injective subdirect product. Define H1 , H2 , H3 , H as in Section 3.2. Let Ei = πi (H).
Then H E ∆ and Ei E Sni for i ∈ {1, 2, 3}. By Theorem 1.1, it holds that Sn1 /E1 ∼
=
= Sn2 /E2 ∼
Sn3 /E3 . Also, by Theorem 3.11, for H we get a canonical tuple (B1 , B2 , B3 , C1 , C2 , C3 , ϕ1 , ϕ2 , ϕ3 )
with subgroups Bi , Ci E Ei , such that Bi ∩ Ci is Abelian and Bi Ci = Ei . We obtain the following
options.
• If E1 = 1 then n1 = n2 = n3 and E2 = E3 = 1. We have Bi = Ci = 1 for i ∈ {1, 2, 3}
In this case H = 1 and ∆ is degenerate. By Theorem 3.15 there are i(n1 )2 groups of this
type, where i(n1 ) is the number of isomorphisms from Sn1 to Sn1 . This corresponds to the
choices for ϕ1 and ϕ2 . For ϕ3 we get ϕ−1
3 = ϕ1 ◦ ϕ2 .
∼ S3 . In this case {n1 , n2 , n3 } ⊆ {3, 4}. Let us
• If E1 = V then n1 = 4 and Sn1 /E1 =
first consider the case that n3 = 3. Then B3 = C3 = E3 = 1 and we can thus apply
Theorem 3.15. By Lemma 3.9 we conclude that C1 = B2 = 1 and E1 = B1 = C2 =
V . Using the correspondence given in Theorem 3.15 the number of such groups equals
the number of pairs (ϕ1 , ϕ2 ), where ϕ1 is an isomorphism from S4 to S4 and ϕ2 is an
isomorphism from S3 to S3 . There are 144 such pairs.
Next let us consider the case n1 = n2 = n3 = 4. This implies that Bi = Ci = Ei = V for
all i ∈ {1, 2, 3}. Since S4 is the semidirect product V ⋊ S3 , we can apply Theorem 3.17.
For every isomorphism κ from G2 to G3 we can find a subdirect product realizing κ by
setting ∆ = h{k, g1 , κ(g1 ) | g1 ∈ G1 , k ∈ S3 ∩ g1 V } ∪ {(a, a−1 , 1) | a ∈ V }i. Thus, by
Theorem 3.17 the number of such subdirect products is equal to the number pairs (κ, ι)
where κ is an isomorphism from S4 to S4 and ι is an automorphism of S4 that fixes V as
a set. There are 6n1 ! = 144 such pairs.
• If E1 = An1 (and n1 ≥ 3) then E2 = An2 and E3 = An3 . By applying Theorem 3.11
to H it follows that either n1 = n2 = n3 = 3 or n1 = n2 > n3 = 2. In the former
case Bi = Ci = A3 = Z3 for i ∈ {1, 2, 3} and in total there are 2n1 ! = 12 groups of this
type by Theorem 3.17 (by the same arguments as in the previous case). In the latter case
B1 = C2 = An1 and C1 = B2 = B3 = C3 = 1. So ∆ is degenerate and by Theorem 3.15
there are in total i(n1 ) · i(2) = i(n1 ) groups of this type, where again i(n1 ) is the number
of isomorphisms from Sn1 to Sn1 .
12
n1 = 4
4
3
2
1
4
1386
282
66
32
3
282
90
18
8
2
66
18
6
2
1
32
8
2
1
n1 = 3
3
2
1
3
90
18
8
2
18
6
2
1
8
2
1
n1 = 2
2
1
2
6
2
1
2
1
Table 1: The numbers ℓ(n1 , n2 , n3 ) for n1 , n2 , n3 ∈ {1, 2, 3, 4}
• If E1 = Sn1 then E2 = Sn2 and E3 = Sn3 . In this case H = ∆. Using the correspondence
described in Theorem 3.11 we get that n1 = n2 = n3 = 2 and Bi = Ci = S2 for i ∈ {1, 2, 3}.
Since there is only one isomorphism from S2 to S2 there is exactly one option in this case,
namely the group given in Example 3.2 for G1 = S2 .
Corollary 4.4. Let n1 ≥ n2 ≥ n3 ≥ 2, n1 ≥ 5. For the number ℓ(n1 , n2 , n3 ) of subdirect
products of Sn1 × Sn2 × Sn3 we have
(n1 !)2 + 6n1 ! + 6
2082246
66
ℓ(n1 , n2 , n3 ) = 18
2886
2m1 ! + 6
6
if n1 = n2 = n3 ∈
/ {6},
if n1 = n2 = n3 = 6,
if n2 = n3 = 4,
if n2 ∈ {3, 4}, n3 = 3,
if {{n1 , n2 , n3 }} = {{6, 6, m2}}, m2 6= 6,
if {{n1 , n2 , n3 }} = {{m1 , m1 , m2 }}, m1 6= m2 , 6 6= m1 ≥ 5,
otherwise.
Proof. Using Lemma 4.1 and 4.3 we get
ℓ2-inj (n1 , n1 , n1 ) + 3ℓ2-inj(2, n1 , n1 ) if n1 = n2 = n3 ,
ℓ2-inj (2, 4, 4) + ℓ2-inj (2, 3, 3)
if n2 = n3 = 4,
if n2 ∈ {3, 4}, n3 = 3,
ℓ2-inj (2, 3, 3)
X
ℓ(ni , nj ) +
ℓ(n1 , n2 , n3 ) =
if {{n1 , n2 , n3 }} =
i<j
ℓ2-inj (2, m1 , m1 )
{{m1 , m1 , m2 }} for
m1 6= m2 , m1 ≥ 5,
0
otherwise.
Then apply Lemma 4.2 and 4.3.
The finitely many cases not covered by the previous corollary are listed in Table 1. These
numbers were calculated using the Lemmas 4.1, 4.2 and 4.3. However, these numbers were also
double-checked with the computer algebra system gap [6].
Acknowledgments. Funded by the Excellence Initiative of the German federal and state governments.
13
References
[1] D. D. Anderson and V. Camillo. Subgroups of direct products of groups, ideals and subrings
of direct products of rings, and Goursat’s lemma. In Rings, modules and representations,
volume 480 of Contemp. Math., pages 1–12. Amer. Math. Soc., Providence, RI, 2009.
[2] K. Bauer, D. Sen, and P. Zvengrowski. A generalized Goursat lemma. Tatra Mt. Math.
Publ., 64, 2015.
[3] M. R. Bridson, J. Howie, C. F. Miller, III, and H. Short. On the finite presentation of
subdirect products and the nature of residually free groups. Amer. J. Math., 135(4):891–
933, 2013.
[4] M. R. Bridson and C. F. Miller, III. Structure and finiteness properties of subdirect products
of groups. Proc. Lond. Math. Soc. (3), 98(3):631–651, 2009.
[5] L. M. Butler. Subgroup lattices and symmetric functions.
112(539):vi+160, 1994.
Mem. Amer. Math. Soc.,
[6] The GAP Group. GAP – Groups, Algorithms, and Programming, Version 4.8.3, 2016.
[7] E. Goursat. Sur les substitutions orthogonales et les divisions régulières de l’espace. Ann.
Sci. École Norm. Sup. (3), 6:9–102, 1889.
[8] M. Hall, Jr. The theory of groups. Chelsea Publishing Co., New York, 1976. Reprinting of
the 1968 edition.
[9] M. Hampejs and L. Tóth. On the subgroups of finite abelian groups of rank three. Ann.
Univ. Sci. Budapest. Sect. Comput., 39:111–124, 2013.
[10] S. Lang. Algebra, volume 211 of Graduate Texts in Mathematics. Springer-Verlag, New
York, third edition, 2002.
[11] M. D. Miller. On the lattice of normal subgroups of a direct product. Pacific J. Math.,
60(2):153–158, 1975.
[12] J. Petrillo. Counting subgroups in a direct product of finite cyclic groups. College Math. J.,
42(3):215–222, 2011.
[13] R. Remak. Über Unterguppen direkter Produkte von drei Faktoren. J. Reine Angew. Math.,
166:65–100, 1932.
14
| 4math.GR
|
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND
APPLICATIONS
arXiv:1406.4947v4 [math.AT] 18 Jul 2016
AKHIL MATHEW
Abstract. Let A be an E∞ -ring over the rational numbers. If A satisfies a noetherian condition
on its homotopy groups π∗ (A), we construct a collection of E∞ -A-algebras that realize on homotopy the residue fields of π∗ (A). We prove an analog of the nilpotence theorem for these residue
fields. As a result, we are able to give a complete algebraic description of the Galois theory of A
and of the thick subcategories of perfect A-modules. We also obtain partial information on the
Picard group of A.
Contents
1. Introduction
2. Degree zero elements of rational E∞ -rings
3. Degree −1 elements
4. Residue fields
5. A thick subcategory theorem
6. Galois groups
7. The Picard group
8. Non-noetherian counterexamples
References
1
4
10
16
23
27
32
36
40
1. Introduction
1.1. Motivation. The goal of this paper is to describe certain invariants of structured ring spectra
in characteristic zero. We start by first reviewing the motivation from stable homotopy theory.
The chromatic picture of stable homotopy theory identifies a class of “residue fields” which play
an important role in global phenomena. Consider the following ring spectra:
(1) HQ: rational homology.
(2) For each prime p, mod p homology HFp .
(3) For each prime p and height n, the nth Morava K-theory K(n).
These all define multiplicative homology theories on the category of spectra satisfying perfect
Künneth isomorphisms: they behave like fields. Moreover, as a consequence of the deep nilpotence
technology of [DHS88, HS98], they are powerful enough to describe much of the structure of the
stable homotopy category. For example, one has the following result:
Date: March 26, 2018.
1
2
AKHIL MATHEW
Theorem 1.1 (Hopkins-Smith [HS98]). Let R be a ring spectrum and let α ∈ π∗ (R). Then α is
nilpotent if and only if the Hurewicz image of α in π∗ (F ⊗ R) is nilpotent, as F ranges over all the
ring spectra above.
This fundamental result was used in [HS98] to classify the thick subcategories of the category of
finite p-local spectra for a fixed prime p: all thick subcategories are defined by vanishing conditions
for the various residue fields. One can attempt to ask such questions not only for spectra but
for general symmetric monoidal, stable ∞-categories, as Hovey, Palmieri, and Strickland have
considered in [HPS97]; whenever one has an analog of Theorem 1.1, it is usually possible to prove
results along these lines.
For instance, let A be an E∞ -ring. Then one can try to study such questions in the ∞-category
Mod(A) of A-modules. If π∗ (A) is concentrated in even degrees and is regular noetherian, then it is
possible to construct residue fields, prove an analog of Theorem 1.1, and obtain a purely algebraic
description of the thick subcategories of perfect A-modules. This has been observed independently
by a number of authors. For E∞ -rings (such as the E∞ -ring TMF of periodic topological modular
forms) which are “built up” appropriately from such nice E∞ -rings, it is sometimes possible to
construct residue fields as well. We used this to classify thick subcategories for perfect modules
over E∞ -rings such as TMF in [Mat15].
1.2. Statement of results. In this paper, we will study such questions over the rational numbers.
Let A be a rational E∞ -ring such that the even homotopy groups πeven (A) form a noetherian ring
and such that the odd homotopy groups πodd (A) form a finitely generated πeven (A)-module. We
will call such rational E∞ -rings noetherian.
For the statement of our first result, we work with E∞ -rings containing a unit in degree two. In
this case, we will produce, for every prime ideal p ⊂ π0 (A), a “residue field” of A, which will be an
E∞ -A-algebra whose homotopy groups form a graded field.
Theorem 1.2 (Existence of residue fields). Let A be a rational, noetherian E∞ -ring containing a
unit in degree two. Given a prime ideal p ⊂ π0 (A), there exists an E∞ -A-algebra κ(p) such that
κ(p) is even periodic and the map π0 (A) → π0 (κ(p)) induces the reduction π0 A → π0 (A)p /pπ0 (A)p .
κ(p) is unique up to homotopy as an object of the ∞-category CAlgA/ of E∞ -rings under A.
We will prove an analog of Theorem 1.1 in Mod(A) for these residue fields.
Theorem 1.3 (Nilpotence). Suppose A is as above, and let B be an A-ring spectrum; that is, an
algebra object in the homotopy category Ho(Mod(A)). Let x ∈ π∗ (B). Then x is nilpotent if and
only if for every prime ideal p ⊂ π0 (A), the image of x in π∗ (B ⊗A κ(p)) is nilpotent.
The proof of Theorem 1.3 uses entirely different (and much simpler) techniques than Theorem 1.1.
However, the conclusion is similar, and we thus find Mod(A) as an interesting new example of an
“axiomatic stable homotopy theory” ([HPS97]) where many familiar techniques can be applied.
In particular, from Theorem 1.3, we will deduce a classification of thick subcategories of the
∞-category Modω (A) of perfect A-modules,
for A rational noetherian (not necessarily containing
L
a unit in degree two). Let πeven (A) = i∈2Z πi (A); this is a graded ring, so Specπeven (A) inherits
a Gm -action.
Theorem 1.4 (Thick subcategory theorem). Let A be a rational, noetherian E∞ -ring. The thick
subcategories of Modω (A) are in natural correspondence with the subsets of the collection of homogeneous prime ideals of πeven (A) which are closed under specialization or, equivalently, specializationclosed subsets of the topological space associated to the stack (Specπeven (A))/Gm .
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
3
In particular, we determine the spectrum in the sense of Balmer [Bal05] of Modω (A) as the topological space associated to the stack (Specπeven (A))/Gm . It shows that the spectrum is determined
in terms of π∗ (A), i.e., the map of [Bal10] is an isomorphism.
We will then apply these ideas to the computation of Galois groups, which we introduced in
[Mat16] as an extension of Rognes’s work [Rog08]. The use of residue fields in Galois theory goes
back to Baker-Richter’s work in [BR08], which studied the Galois groups of Morava E-theories at
odd primes. We will show that the Galois theory of a noetherian rational E∞ -ring is “almost”
entirely algebraic. (The “almost” comes from, e.g., the possibility of adjoining roots of periodicity
generators in degrees 2n, n > 1.) We prove:
Theorem 1.5. If A is a noetherian rational E∞ -ring, then the Galois group of A is the étale
fundamental group of the stack (Specπeven (A)) /Gm .
We note that the Galois group depends on the choice of a basepoint is not truly “canonical”;
however, there is a canonical equivalence of the associated Galois theories, or equivalently of the
Galois groupoids.
Finally, we will study the Picard groups of noetherian rational E∞ -rings. Here our results are
much less conclusive, but we prove:
Theorem 1.6. If A is a noetherian rational E∞ -ring, then the cokernel of the natural map
Pic(π∗ (A)) → Pic(A) (see Construction 7.4) is a torsion-free abelian group.
Usually, results such as Theorem 1.4 and Theorem 1.5 are proved using strong homological
assumptions on π∗ (A), e.g., that it is a regular ring. We will be able to get away with much weaker
hypotheses on π∗ (A) (i.e., nothing close to regularity) because, over characteristic zero, E∞ -rings
are simpler. They have a more algebraic feel which gives one a wider range of techniques, and they
have been studied in detail starting with Quillen’s work on rational homotopy theory [Qui69]. In
particular, there are two basic coincidences that will be used in this paper.
(1) The free E∞ -ring on a generator in degree zero is equivalent to the suspension spectrum
Σ∞
+ Z≥0 . In particular, as a result, it is possible to quotient an E∞ -ring by an element in
degree zero to get a new E∞ -ring.
(2) The free E∞ -ring on a generator in degree −1 is equivalent to cochains on S 1 . This has
important descent-theoretic consequences and enables one to compare modules over this
E∞ -ring with local systems on the circle S 1 .
Both these conditions are specific to the rational numbers. They fail away from characteristic
zero, because of the existence of power operations [CLM76, BMMS86].
The above theorems rely crucially on the noetherianness hypotheses. We will discuss various
counterexamples in §8. These counterexamples are related to classical purity questions for Picard
groups and étale fundamental groups for local rings. As a consequence, we produce new examples
of non-algebraic Galois extensions of ring spectra in the sense of Rognes [Rog08], which seem to be
of interest in itself.
Theorem 1.7. Given n > 1, there exists a rational E∞ -ring A with π0 (A) = C[xn , y n−1 x, y n−2 x2 , . . . , y n ],
πi (A) = 0 for i > 0, and such that the Galois group of A is Z/n.
As the étale fundamental group of SpecC[xn , y n−1 x, y n−2 x2 , . . . , y n ] is trivial, this Galois group
is not algebraic. To our knowledge, this is the first non-algebraic example of a Galois extension of
rational ring spectra to be observed. We will describe explicitly the Z/n-Galois extension in §8.
4
AKHIL MATHEW
Organization. This paper is organized as follows. In §2, we analyze the operation of attaching
a 1-cell in a rational E∞ -ring in detail. In §3, we do the same for the operation of attaching a
0-cell, via a comparison between modules over the cochain E∞ -ring C ∗ (S 1 ; Q) and local systems
on S 1 . By reducing to the case where one has a unit in degree 2, we will be able to obtain all our
constructions by only attaching cells in these degrees. The main technical results (existence of the
residue fields and the nilpotence theorem) are proved in §4. §5, §6, and §7 contain the applications
to thick subcategories, Galois groups, and Picard groups, respectively. Finally, §8 discusses various
non-noetherian counterexamples.
Notation. In this paper, we will adopt the following notational conventions. We will frequently
identify abelian groups with their Eilenberg-MacLane spectra without additional notation. The
letters R, S, T, ... will refer to ordinary (discrete) rings. The subscript ∗ will refer to a grading. We
will let CAlg denote the ∞-category of E∞ -rings. The letters A, B, C will refer to E∞ -rings. Given
an E∞ -ring A, we let Mod(A) denote the ∞-category of A-modules and Modω (A) ⊂ Mod(A) the full
subcategory spanned by the perfect A-modules. If X is a space and A an E∞ -ring, we let C ∗ (X; A)
denote the cochain E∞ -algebra on X with values in A, often also denoted AX or Fun(X+ , A).
Finally, if A ∈ CAlg is a rational E∞ -ring, we will write A[t±1
2 ] for the free E∞ -A-algebra on an
invertible degree two generator.
Although we use the language of ∞-categories and higher algebra [Lur14] in this paper, we note
that everything can be carried out in the world of model categories. The theory of E∞ -ring spectra
and modules over them was originally developed [EKMM97] in the world of model categories. In
characteristic zero, the issue simplifies further as one can work with commutative differential graded
algebras.
Acknowledgments. I am grateful to Bhargav Bhatt and Jacob Lurie for helpful discussions related
to the subject of this paper, and the referee for many detailed comments. The author was supported
by the NSF Graduate Fellowship under grant DGE-114415.
2. Degree zero elements of rational E∞ -rings
In this section, we describe the first set of the basic characteristic zero techniques needed for
this paper. In particular, we discuss the “coincidence” of E∞ -rings that Σ∞
+ Z≥0 is free on a degree
zero class and thus analyze the operation of attaching cells in degree one. The main result of the
section (Theorem 2.21) controls the behavior on homotopy rings of attaching cells in degree one.
We also prove a version (Proposition 2.14) of the classical result that a complete local ring with
residue field of characteristic zero contains a copy of its residue field.
Some of the more refined results require the noetherianness hypothesis that will be crucial for
most of the main results of this paper.
Definition 2.1. We say that a rational E∞ -ring A is noetherian if:
(1) The commutative ring πeven (A) is noetherian.
(2) The πeven (A)-module πodd (A) is finitely generated.
The “noetherian” hypothesis ensures that certain categorical constructions one may perform on
A affect the homotopy groups of A in a reasonable manner and, as such, will be indispensable to
this paper.
Warning 2.2. The noetherian condition is not a purely categorical one. For instance, a finitely
presented E∞ -algebra over a noetherian E∞ -ring (even Q) need not be noetherian, i.e., the analog
of Hilbert’s basis theorem fails. See Section 8.2 for an example.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
5
2.1. Cofibers of degree zero elements. Let A be an E∞ -ring and let x ∈ πk (A), defining a map
x
of A-modules Σk A → A.
Definition 2.3. We will write A/x for the cofiber of this map x : Σk A → A.
One wants to think of A/x as a homotopy-theoretic “quotient” of A by the “ideal” generated by
x and, as in algebra, turn this into an E∞ -ring under A. There is, in general, no reason to expect
this to be possible (or canonical in any way).
Example 2.4. The sphere S 0 is the most basic example of an E∞ -ring, but the Moore spectrum
S 0 /2 is not even a ring spectrum up to homotopy.
The obstructions to multiplicative structures have been discussed, for example, in [Oka79, Str99].
Some further obstructions to structured multiplications, via the theory of power operations, are
discussed in [MNN15].
Suppose first that k = 0. To understand the failure of such quotients to be ring spectra, recall how
the quotient is constructed in classical commutative algebra. Let R be a (classical) commutative
ring, and fix x ∈ R. The (classical) quotient R/(x) is the pushout of the diagram of commutative
rings
t7→0
// Z .
Z[t]
x7→t
// R/(x)
R
Here Z[t] is the free commutative ring on a generator t, and forming the pushout R/(x) as above
amounts to setting x = 0.
In homotopy theory, one can make a similar construction, which has been discussed in, for
example, [Szy14].
Definition 2.5. There is a free E∞ -ring on a single generator, denoted S 0 {t}, whose underlying
spectrum is given by
M
S 0 {t} ≃
Σ∞
+ BΣn ,
n≥0
as {Σn }n≥0 ranges over the symmetric groups. Given an E∞ -ring A and an element x ∈ π0 A,
we obtain a map (defined up to homotopy) of E∞ -rings S 0 {t} → A sending the tautological class
t ∈ π0 S 0 {t} to x, by the universal property: the space of maps of E∞ -rings S 0 {t} → A is precisely
Ω∞ A. In particular, we can form a pushout square in CAlg,
S 0 {t}
t7→0
// S 0 ,
t7→x
A
// A′
where A′ is called the free A-algebra with x = 0. Following Szymik [Szy14], we will write A′ = A//x.
Given an E∞ -A-algebra A′′ , the space1 HomA/ (A′ , A′′ ) is the space of nullhomotopies of x in A′′
(which is empty unless x maps to zero in π0 A′′ , and in this case is Ω∞+1 A′′ ).
For future reference, it will be convenient to have the following more general definition.
1We write CAlg
A/ for the ∞-category of E∞ -A-algebras and HomA/ (·, ·) for mapping spaces here.
6
AKHIL MATHEW
Definition 2.6.
X a spectrum, we write Sym∗ (X) for the free E∞ -ring on X, so that
L For ⊗n
∗
Sym (X) ≃
)hΣn . If A is an E∞ -ring and x ∈ πk A is an element, we denote by
n≥0 (X
A//x the pushout
Sym∗ (S k )
x
// A
,
0
S0
// A//x
where the map Sym∗ (S k ) → A is determined by the map x : S k → A, and where Sym∗ (S k ) → S 0
is determined by 0 : S k → S 0 . We will call A//x the free E∞ -A-algebra with x = 0. Given a
sequence of elements x1 , . . . , xn ∈ π∗ (A), we will write A//(x1 , . . . , xn ) for the iterated quotient
(. . . (A//x1 )//x2 ) . . . //xn−1 ) . . . )) //xn .
We return to the case k = 0. In general, if x ∈ π0 A is fixed, then Definition 2.5 gives
A//x ≃ A ⊗S 0 {t} S 0 ,
since pushouts of E∞ -rings are relative tensor products. This is usually very different, as an Amodule, from A/x. For example, the free E∞ -ring with pn = 0 is not S 0 /pn . From the “chromatic”
point of view, it is actually invisible: its Er -localization vanishes for each r, by the main result of
[MNN15].
Remark 2.7. In fact, the S 0 {t}-module S 0 is quite complicated, and is not, for example, perfect.
For instance, if we worked over F2 rather than S 0 , then F2 {t} has homotopy groups given by a
polynomial ring on the tautological class t and certain admissible monomials in the Dyer-Lashof
algebra applied to t ([CLM76]), so that F2 is an infinite quotient of F2 {t} by a regular sequence of
polynomial generators.
However, there is another E∞ -ring which is better behaved in this regard, and which will enable
us to place E∞ -structures on quotients in certain cases. Recall that the E∞ -ring S 0 {t} is obtained
from the free E∞ -space on a single generator by applying Σ∞
+ . This E∞ -space is the free symmetric
monoidal category
on
one
object:
the
groupoid
of
finite
sets
and isomorphisms between them, or
F
topologically n≥0 BΣn .
Definition 2.8. We can apply Σ∞
+ instead to the symmetric monoidal groupoid Z≥0 , which has
objects given by the natural numbers (under addition) and no nontrivial isomorphisms. The resulting E∞ -ring Σ∞
+ Z≥0 , the “monoid algebra” of the natural numbers (as studied, for example, in
[ABG+ 14]), will be written S 0 [t] since its homotopy groups actually are given by (π∗ S 0 )[t]. More
generally, we will write A[t] for the E∞ -ring A ⊗ Σ∞
+ Z≥0 , if A is any E∞ -ring.
Construction 2.9. Now let A be an E∞ -ring, and let S 0 {t} → A be a map classifying an element
x ∈ π0 (A). Suppose that we have a factorization in the ∞-category CAlg
S 0 {t}
x
// A .
④==
④
④
④
④
S 0 [t]
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
7
In this case, we can form the relative tensor product A ⊗S 0 [t] S 0 (using the map S 0 [t] → S 0 which
sends t 7→ 0), as an E∞ -ring. The cofiber sequence of S 0 [t]-modules
t
S 0 [t] → S 0 [t] → S 0 ,
shows that this relative tensor product, as an A-module spectrum, is actually A/x.
In other words, by the universal property of the monoid algebra, if there exists a factorization
in the diagram of E∞ -spaces
F
x
// Ω∞ A ,
n≥0 BΣn
s99
s
s
s
s s
Z≥0
where Ω∞ A is given the multiplicative E∞ -structure, then we can place a natural E∞ -structure on
A/x.
Remark 2.10. Let X be an E∞ -space and let x ∈ π0 X, classified by a map of E∞ -spaces
F
n≥0 BΣn → X. If this map admits a factorization over Z≥0 , then x has been called by Lurie
a “strictly commutative” element of X. Construction 2.9 shows that, while arbitrary cofibers A/x
need not admit E∞ -structures, one can find an E∞ -structure if x is strictly commutative.
Unfortunately, in general, describing maps out of S 0 [t] is difficult, since Z≥0 does not admit a
simple presentation as an E∞ -space. In characteristic zero, i.e., over Q, the natural map
Q {t} → Q[t],
becomes an equivalence of E∞ -rings, because the maps (BΣn )+ → S 0 are rational equivalences. In
particular, given any rational E∞ -ring A, and an element x ∈ π0 (A), we can obtain a map
Q[t] → A,
t 7→ x,
and we can form the relative tensor product A/x ≃ A ⊗Q[t] Q as an E∞ -ring. This process can
equivalently be described as attaching a 1-cell to kill the element x ∈ π0 A, i.e., as forming A//x.
We may summarize these observations in the following proposition.
Proposition 2.11. Let A be a rational E∞ -ring and let x ∈ π0 (A). Then A//x ∈ CAlgA/ has as
underlying A-module A/x. In particular, we may make A/x into an E∞ -A-algebra.
It is similarly possible to quotient by even degree elements of a rational E∞ -ring, as we show
below. In fact, we did not strictly need the discussion of “strict commutativity” for the present
paper, but included it for its intrinsic interest (as it becomes more relevant away from characteristic
zero).
Proposition 2.12. Let A be a rational E∞ -ring and let x ∈ πn (A). Suppose n is an even integer.
Then A//x ∈ CAlgA/ has underlying A-module A/x.
Proof. Recall that π∗ Sym∗ (Q[n]) is a polynomial ring on a class in degree n. Thus, the result
follows from A//x ≃ A ⊗Sym∗ (Q[n]) Q and the cofiber sequence Σn Sym∗ Q[n] → Sym∗ Q[n] → Q of
Sym∗ Q[n]-modules.
8
AKHIL MATHEW
2.2. The Cohen structure theorem. Let (R, m) be a complete local noetherian ring with residue
field k of characteristic zero. In this case, a basic piece of the Cohen structure theorem (see for
instance [Eis95, Ch. 8]) implies that R contains a copy of its residue field:
Theorem 2.13 (Cohen). Hypotheses as above, the projection R → R/m = k admits a section.
We refer to [Mat80, Theorem 60, §28.J] for a proof of a more general result than Theorem 2.13.
It is closely related to the fact that, in characteristic zero, all field extensions can be obtained as
an inductive limit of smooth morphisms; this argument implies an analogous result in the world of
E∞ -rings, and it is the purpose of this subsection to describe that. In particular, we prove:
Proposition 2.14. Let A be a noetherian, rational E∞ -ring such that π0 A is a complete local
ring with residue field k. Then there exists a morphism of E∞ -rings k → A such that on π0 , the
composite map k → π0 A → k is the identity.
Proof. By Theorem 2.13, there is a section φ : k → π0 A of the reduction map. We want to realize
this topologically. To start with, we obtain a map Q → A as A is rational. Let {tα }α∈Γ be a
transcendence basis of k/Q (cf. [Lan02, Ch. VIII, sec. 1] for a textbook reference), so that we have
field extensions
Q ⊂ Q({tα }) ⊂ k,
where the first extension is purely transcendental and the second extension is algebraic. For each
α ∈ Γ, choose uα ∈ π0 A be defined by uα = φ(tα ), so that uα projects to tα in the residue field.
We obtain a map of E∞ -rings
O
Q[tα ] → A, tα 7→ uα ,
Γ
where the left-hand-side is a free E∞ -ring on |Γ| variables (i.e., a discrete polynomial ring on the
{tα }). It necessarily factors over the localization Q({tα }), so we obtain a map Q({tα }) → A. This
realizes on π0 the restriction φ|Q({tα }) .
Finally, we want to find an extension over k such that the diagram
//
①k ,
Q({tα })
①
①
{{①
A
①
such that the composite k → π0 A → k is the identity. Since k is a colimit of finite étale Q({tα })algebras (i.e., finite separable extensions; recall that we are in characteristic zero), it is equivalent
to doing this at the level of π0 ([Lur14, §7.5]), and the map φ : k → π0 (A) enables us to do that.
Remark 2.15. The argument shows that the set of homotopy classes of maps of E∞ -rings k → A
is in bijection with the set of ring-homomorphisms k → π0 (A).
2.3. Properties of the quotient E∞ -ring. We will need a few more preliminaries on the construction of Definition 2.5 and its behavior on homotopy. If A is a rational E∞ -ring and x ∈ π0 A,
then the homotopy groups of A//x ≃ A/x are determined additively by the short exact sequence
(1)
0 → πj (A)/xπj (A) → πj (A/x) → (ker x)|πj−1 (A) → 0,
but this fails to determine the precise multiplicative structure. In this section, we will show (Theorem 2.21) that the multiplicative structure is not so different from that of the subring π∗ (A)/xπ∗ (A)
under noetherian hypotheses. We will not need the full strength of these results in the sequel.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
9
We begin by reviewing the theory of finite universal homeomorphisms [Sta13, Tag 04DC]. Recall
that a map of rings R → R′ is called a universal homeomorphism if, for every R-algebra R′′ , the
map R′′ → R′′ ⊗R R′ induces a homeomorphism upon applying Spec.
Proposition 2.16. A morphism R → R′ , such that R′ is a finitely generated R-module, is a
universal homeomorphism if and only if, for every morphism R → k where k is a field, the basechange R′ ⊗R k is a local artinian k-algebra (in particular, nonzero).
Proof. Suppose that for every map from R to a field k, the base-change R′ ⊗R k is a local artinian
k-algebra. It follows that the residue field of R′ ⊗R k is necessarily a purely inseparable extension
of k, since otherwise we can replace k by k and R′ ⊗R k would have nontrivial idempotents. It
then follows that the map SpecR′ → SpecR is a radicial morphism [Sta13, Tag 01S2]. Given any
R-algebra R′′ , the map R′′ → R′ ⊗R R′′ is finite, so induces a closed map on Spec which is also
injective since R → R′ is radicial. The map is surjective since all the fibers are nonempty by
assumption, and thus a homeomorphism.
Conversely, suppose R → R′ is a finite universal homeomorphism. Then all the base-changes k →
′
R ⊗R k, for an R-field k, are universal homeomorphisms themselves. In particular, Spec(R′ ⊗R k)
is connected. But R′ ⊗R k is a finite-dimensional k-algebra, so if its spectrum is connected, then
R′ ⊗R k must be local artinian.
Corollary 2.17. A finite map R → R′ of Q-algebras is a universal homeomorphism if and only if
for every residue field R → k, the tensor product R′ ⊗R k is local with residue field k.
Proof. This follows from the fact that all finite extensions in characteristic zero are separable, so
that if A is a finite-dimensional local k-algebra with residue field strictly containing k, then A ⊗k k
necessarily has nontrivial idempotents.
We will now begin working towards the proof of Theorem 2.21. We will first need a preliminary
lemma on idempotents in these quotients.
Definition 2.18. If A is an E∞ -ring, we let Idem(A) denote the set of idempotents in π0 A. The
construction A 7→ Idem(A) sends homotopy limits of E∞ -rings to inverse limits of sets, since the
set Idem(A) is homotopy equivalent to the space of maps of E∞ -rings S 0 × S 0 → A in view of the
theory of étale algebras [Lur14, §7.5]. If R is a discrete ring, we will also write Idem(R) for the set
of idempotents in R.
Lemma 2.19. Let A be a rational noetherian E∞ -ring with π0 (A) local and let x1 , . . . , xr ∈ π0 (A).
Then the map π0 (A)/(x1 , . . . , xr ) → π0 (A//(x1 , . . . , xr )) of discrete rings induces an isomorphism
on Idem.
Proof. In fact, we have a map of E∞ -rings A → A//(x1 , . . . , xr ), and if we form the cobar construction on this map, we obtain a coaugmented cosimplicial object
→
A//(x1 , . . . , xr ) →
...,
→ A//(x1 , . . . , xr ) ⊗A A//(x1 , . . . , xr ) →
→
whose homotopy limit is the (x1 , . . . , xr )-adic completion of A. We refer to [Lur11b, §4] for preliminaries on completions of ring spectra. In particular, the idempotents in the totalization are the
same as the idempotents in the (x1 , . . . , xr )-adic completion of π0 A, or equivalently, by the lifting
idempotents theorem [Eis95, Cor. 7.5], in π0 (A)/(x1 , . . . , xr ).
Since the operation of taking idempotents commutes with homotopy limits, we conclude that
the set of idempotents in π0 (A)/(x1 , . . . , xr ) is the reflexive equalizer
Idem(A//(x1 , . . . , xr )) →
→ Idem(A//(x1 , . . . , xr ) ⊗A A//(x1 , . . . , xr )).
10
AKHIL MATHEW
However, we claim that the two maps in the reflexive equalizers are isomorphisms (and thus
equal). In fact, A//(x1 , . . . , xr ) ⊗A A//(x1 , . . . , xr ) is obtained by attaching 1-cells to kill the
classes x1 , . . . , xr in A//(x1 , . . . , xr ) which are already zero; in particular, as an E∞ -ring, we have
A//(x1 , . . . , xr ) ⊗A A//(x1 , . . . , xr ) ≃ A//(x1 , . . . , xr ) ⊗ Sym∗ [y1 , . . . , yr ],
which has the same idempotents as A//(x1 , . . . , xr ).
|yi | = 1,
Proposition 2.20. Let A be a rational noetherian E∞ -ring and let x ∈ π0 (A). The map π0 (A)/(x) →
π0 (A//x) is a finite universal homeomorphism.
Proof. We already know that π0 (A//x) is a finitely generated π0 (A)/(x)-module, by the short exact
sequence (1). We will check that the map is a finite universal homeomorphism fiberwise at each
prime.
Fix a prime ideal p of π0 (A)/(x). Let x1 , . . . , xn ∈ π0 (A) project to generators of p. Localizing
A at p, we may assume that π0 (A) is local with maximal ideal p. By completing A, we may assume
that that A admits the structure of an E∞ -k-algebra for k the residue field of π0 (A), in view of
Proposition 2.14.
We need to show that the map of commutative rings
(2)
π0 (A)/(x, x1 , . . . , xn ) → π0 (A//x)/(x1 , . . . , xn )
is a finite universal homeomorphism. The left-hand-side of (2) is the residue field k of π0 (A)
at the maximal ideal, and the right-hand-side is a finite module over the left-hand-side and is
in particular a product of local artinian k-algebras. By replacing A with A ⊗k k ′ for a finite
extension k ′ /k, we may assume that each of the residue fields of the left-hand-side of (2) is k
itself. Thus, it suffices to show π0 (A//x)/(x1 , . . . , xn ) has no nontrivial idempotents in this case.
As a result, our claim follows from Lemma 2.19, which implies that the connected components of
Specπ0 (A//x)/(x1 , . . . , xn ) are in bijection with those of Specπ0 (A//(x, x1 , . . . , xn )), and in turn
with those of Specπ0 (A)/(x, x1 , . . . , xn ), while the latter is just a point.
By induction (and transitivity), one obtains an analogous result for any finite sequence of elements in π0 A. Moreover, by replacing A with A[t2±1 ], we can thus obtain a result for quotients by
even degree elements. We find:
Theorem 2.21. Let A be a rational noetherian E∞ -ring and let x1 , . . . , xn ∈ πeven (A) be a sequence
of elements. Then the map
πeven (A)/(x1 , . . . , xn ) → πeven (A//(x1 , . . . , xn )),
is a finite universal homeomorphism.
3. Degree −1 elements
We will also encounter odd degree elements in homotopy, and thus, in this section, we consider
the free E∞ -Q-algebra Sym∗ Q[−1] on a generator in degree −1. It is the purpose of this section to
use the coincidence (Proposition 3.2) that Sym∗ Q[−1] ≃ C ∗ (S 1 ; Q) to prove certain basic facts (in
Section 3.3) about Sym∗ Q[−1] ≃ C ∗ (S 1 ; Q)-modules, and ultimately about the construction A//y
where A is a rational E∞ -ring and y ∈ π−1 (A).
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
11
3.1. The free E∞ -ring on k[−1]. Let k be a field of characteristic zero, which we will work over.
Recall that the free E∞ -k-algebra on k[−1] is
M
⊗n
Sym∗ k[−1] ≃
(k[−1])hΣn .
n≥0
⊗n
Here k[−1] ≃ k[−n], and the Σn -action is via the sign representation. For n ≥ 2, this action is
nontrivial, and it follows that the homotopy coinvariants are zero. In particular, we find:
Corollary 3.1. For char k = 0, the homotopy groups
k
πi (Sym∗ k[−1]) ≃ k
0
of Sym∗ k[−1] are given by:
if i = 0
if i = −1 ,
otherwise
and the multiplication is determined (“square zero” in degree −1).
There are two other E∞ -rings which have a similar multiplication law on their homotopy groups:
(1) The cochain E∞ -ring on S 1 , C ∗ (S 1 ; k).
(2) The square-zero E∞ -ring k ⊕ k[−1].
Proposition 3.2. Let k be a field of characteristic zero. Then there are equivalences of E∞ -rings
Sym∗ k[−1] ≃ C ∗ (S 1 ; k) ≃ k ⊕ k[−1].
Proof. In fact, we can produce maps
Sym∗ k[−1] → C ∗ (S 1 ; k),
Sym∗ k[−1] → k ⊕ k[−1],
such that they are isomorphisms on π−1 (using the universal property of Sym∗ ), and therefore are
equivalences of E∞ -rings by inspection of π∗ . So, all three are equivalent.
Remark 3.3. If one worked over Fp , the symmetric algebra Sym∗ (Fp [−1]) is definitely much too
large to be either C ∗ (S 1 ; Fp ) or Fp ⊕ Fp [−1], but C ∗ (S 1 ; Fp ) and Fp ⊕ Fp [−1] have the same squarezero multiplication on homotopy groups. They are not equivalent as E∞ -rings under Fp because
the zeroth reduced power P 0 acts as the identity on π−1 of the former and zero on the latter.
More generally, if n is any odd integer, we can repeat the above reasoning:
Corollary 3.4. If n is odd, then we have equivalences of E∞ -rings
(3)
Sym∗ k[−n] ≃ k ⊕ k[−n],
and, if n > 0, then these are equivalent to cochains on the n-sphere, C ∗ (S n ; k).
3.2. Descent properties and A//x. If A is a rational E∞ -ring and x ∈ π∗ (A) is an element in an
even degree, then we saw in Proposition 2.12 that the construction A//x was reasonably hands-on:
it gave us the underlying A-module A/x. If x is in odd degree, however, A//x may be much bigger
than A.
Example 3.5. Let x = 0 ∈ π−1 (A). Then A//x ≃ A[t].
In fact, for x in an odd degree, it is not even a priori evident that if A is nonzero, then A//x is
also nonzero, even as x2 = 0. We will prove this (and more) using descent theory. We recall first a
definition.
Definition 3.6. Let φ : A → A′ be a morphism of E∞ -rings. We say that φ admits descent if the
thick tensor-ideal that A′ generates, in Mod(A), is all of Mod(A).
12
AKHIL MATHEW
We refer to [Mat16, §3-4] for preliminaries on the notion of “admitting descent.” Here is a simple
example.
Proposition 3.7. Let A be a rational E∞ -ring and let x ∈ π0 A be nilpotent. Then the E∞ -Aalgebra A//x admits descent over A.
Proof. In fact, thanks to the octahedral axiom, the thick subcategory of Mod(A) generated by the
A-module A/x (which is equivalent to the underlying A-module of A//x) contains A/x2 , A/x3 , . . . ,,
and eventually A/xN where N is so large that xN = 0. But A/xN ≃ A ⊕ ΣA for such N , and
therefore the thick subcategory generated by A//x actually contains A.
Let A be an E∞ -ring and let y ∈ πn (A), with n odd. Then y 2 = 0, so that one would hope that the
analog of Proposition 3.7 would be automatic. That is, one would hope that A//y ≃ A⊗Sym∗ (Q[n]) Q
admits descent over A. A priori, it is harder to control this tensor product, because Q is no longer
a perfect Sym∗ (Q[n])-module for n odd, so one cannot imitate the above argument. However, we
can still prove the statement.
Proposition 3.8. If A is nonzero and y ∈ πn A (for n odd), then A//y ∈ CAlgA/ admits descent.
In particular, A//y =
6 0.
Proof. The map Sym∗ Q[n] → Q admits descent, in view of the equivalence Sym∗ (Q[n]) ≃ Q ⊕ Q[n]
of Corollary 3.4. It follows that the map A → A//y admits descent as well by base-change.
Example 3.9. Proposition 3.8 definitely fails for E∞ -rings under Fp . For example, if p = 2, then
tZ/2
odd degree elements can be invertible (take the Tate spectrum F2 ). If p > 2, odd degree elements
tZ/p
square to zero but can still be “resilient.” In the Tate spectrum Fp , we have
π∗ (FtZ/p
) ≃ Fp [t2±1 ] ⊗ E(α−1 ),
p
where the exterior generator α−1 has the property that βP 0 α−1 = t−1
is invertible. Thus, even
2
though α−1 squares to zero, a basic power operation goes from it to an invertible element. It
tZ/p
follows that in any E∞ -ring under Fp , if α maps to zero, the whole E∞ -ring has to be zero. Such
phenomena can never happen in characteristic zero.
3.3. Comparison with local systems. In this subsection, we give the most important (for this
paper) application of Proposition 3.2. We will be able to describe the ∞-category of modules over
the free algebra Sym∗ k[−1] ≃ C ∗ (S 1 ; k) for chark = 0. In fact, let k be any field, not necessarily of
characteristic zero. We will describe modules over the cochain algebra C ∗ (S 1 ; k). For example, we
will be able to give a complete classification of all perfect modules.
We first recall a basic construction from [Mat16, §7.2].
Definition 3.10. Let X be a space and let C be an ∞-category. We define LocX (C) = Fun(X, C)
and refer to it as the ∞-category of C-valued local systems on X.
Construction 3.11. Let A be an E∞ -ring and let C = Mod(A). Let X be a finite complex. We
have a natural fully faithful, symmetric monoidal imbedding
Mod(C ∗ (X; A)) ⊂ LocX (Mod(A)),
from modules over the cochain E∞ -ring C ∗ (X; A) into local systems of A-modules on X, whose
image in LocX (Mod(A)) is the localizing subcategory generated by the unit.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
13
Informally, the functor sends a C ∗ (X; A)-module M to the A-module M ⊗C ∗ (X;A) A, which lives
as a local system over X (since there is an X’s worth of evaluation maps C ∗ (X; A) → A).
In general, it is somewhat subtle to test whether an object in LocX (Mod(A)) belongs to the
essential image of Mod(C ∗ (X; A)). However, if X = S 1 then things simplify considerably. Given a
local system N ∈ LocS 1 (Mod(A)), the evaluation Nq (for a fixed basepoint q ∈ S 1 ) is an A-module,
and Nq acquires a monodromy automorphism φ
φ : Nq ≃ Nq ,
1
coming from a choice of generator of π1 (S ; q).
Proposition 3.12 ([Mat16, Remark 7.9]). Given N as above, then N belongs to the image of
Mod(C ∗ (S 1 ; A)) if and only if the action of φ−1 on the homotopy groups π∗ (Nq ) is locally nilpotent.
It will be important for us to have the correspondence in Proposition 3.12 in as clear terms as
possible. Thus, we state the following construction.
Construction 3.13. The right adjoint LocS 1 (Mod(A)) → Mod(C ∗ (S 1 ; A)) is given, for a local
system N , by taking its global sections limS 1 N . Explicitly, if q ∈ S 1 and φ : Nq → Nq is as above,
←−
we have
φ−1
(4)
lim N = fib Nq → Nq ,
←−
1
S
and in particular, we can determine the homotopy groups of limS 1 N via a long exact sequence.
←−
1
Remark 3.14. This discussion is special to the case of S . For any finite complex X and any E∞ ring A, we have an inclusion Mod(C ∗ (X; A)) ⊂ LocX (Mod(A)), and the image always is contained
in the subcategory of local systems satisfying an ind-unipotence property on homotopy groups, but
the precise identification of the image relies on the 1-dimensionality of the circle.
Let C be an arbitrary ∞-category. To give a local system on S 1 in some ∞-category C is equivalent
to giving an object of that ∞-category and an automorphism, via S 1 ≃ K(Z, 1). Fix two C-valued
local systems on S 1 , (x, φx ), (y, φy ), where x, y ∈ C and φx : x ≃ x, φy : y ≃ y are automorphisms in
C. Given a map f : x → y such that the diagram
(5)
φx
x
// x ,
f
y
f
φy
// y
commutes up to homotopy, then we can produce a map of local systems (x, φx ) → (y, φy ) extending
the map f : x → y.
Remark 3.15. Specifying such a map amounts in addition to choosing a homotopy to make the
diagram commute, and there may be many homotopy classes of such.
We can state this formally:
Proposition 3.16. Every object in LocS 1 (C) is represented by a pair (x, φx ) where x ∈ C and
φx : x → x is an automorphism, and two pairs (x, φx ), (y, φy ) are isomorphic if and only if there
exists an isomorphism f : x → y such that the diagram (5) is homotopy commutative.
We now specialize to the case where A = k is a field. We will use Construction 3.11, Proposition 3.12, and Proposition 3.16 to classify C ∗ (S 1 ; k)-modules.
14
AKHIL MATHEW
Proposition
3.17. Any local system L ∈ LocS 1 (Mod(k)) decomposes uniquely as a direct sum
L
L ≃ n∈Z Ln [n] where the fiber of Ln at a point of S 1 is discrete.
Proof. L
Let L = (M, φ). The k-module M decomposes as a sum of its homotopy groups, i.e.,
M ≃
n∈Z (πn M )[n], and the automorphism φ : M → M is determined by its behavior on its
homotopy groups. It follows that, for each n, we can produce squares
(πn M )[n]
M
φ∗
// (πn M )[n] ,
φ
// M
which commute up to homotopy. Putting this together, we can produce a square
L
L
φ∗
//
n∈Z (πn M )[n]
n∈Z (πn M )[n] ,
M
φ
// M
which commutes up to homotopy, and where the vertical maps are now equivalences. It follows from Proposition 3.16 that the pair (M, φ) is equivalent to the direct sum of the pairs
{((πn M )[n], φ∗ )}n∈Z , where each of these is concentrated in a single degree.
Using the imbedding Mod(C ∗ (S 1 ; k)) ⊂ LocS 1 (Mod(k)), we find from this:
L
Corollary 3.18. Any C ∗ (S 1 ; k)-module M admits a unique decomposition M ≃
n≥0 Mn [n],
∗
1
where Mn ∈ Mod(C (S ; k)) has the property that M ⊗C ∗ (S 1 ;k) k is a discrete k-module.
In order to give a discrete local system on S 1 , it suffices simply to give a k-vector space with an
automorphism. In the case we are interested, i.e., local systems coming from C ∗ (S 1 ; k)-modules, it
follows that to give an equivalence class of C ∗ (S 1 ; k)-modules M equates to giving, for each n ∈ Z,
a discrete k[x]-module on which x acts locally nilpotently. The nth such object corresponds to
πn (M ⊗C ∗ (S 1 ;k) k) and x is the monodromy automorphism minus the identity.
In general, the classification of torsion modules over a PID is nontrivial, but in the finitely
generated case, we have a simple complete classification. This leads to:
Construction 3.19. Fix i ∈ Z>0 . We consider the n-dimensional k-vector space Vi = k[x]/xn and
the nilpotent endomorphism given by multiplication by x. Let Vi be the associated local system on
S 1 with fiber Vi and monodromy automorphism 1 + x. Then, by Proposition 3.12, Vi corresponds to
a C ∗ (S 1 ; k)-module that we will denote by Ni . Thus, Ni ⊗C ∗ (S 1 ;k) k is discrete, and i-dimensional,
and the monodromy automorphism is unipotent with a single Jordan block.
In order to determine the homotopy groups of Ni , we have to form the associated local system,
and take global sections over S 1 , as in Construction 3.13. We have
k if j = 0
πj (Ni ) = k if j = −1 .
0 otherwise
For example, N1 = C ∗ (S 1 ; k).
Proposition 3.20. Ni is a perfect C ∗ (S 1 ; k)-module.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
15
Proof. By construction, Ni ⊗C ∗ (S 1 ;k) k is a perfect k-module. Moreover, C ∗ (S 1 ; k) → k admits
descent [Mat16, Prop. 3.35]. Therefore, Ni is a perfect C ∗ (S 1 ; k)-module in view of [Mat16, Prop.
3.27].
Proposition 3.21. Any perfect C ∗ (S 1 ; k)-module M decomposes uniquely as a sum of copies of
shifts of the Ni .
Proof. Given a perfect C ∗ (S 1 ; k)-module, the associated local system (which is a local system of
perfect k-modules) splits as a direct sum of shifts of local systems of discrete k-modules, by Proposition 3.17. Each of these is determined by a finite-dimensional k-vector space with an unipotent
automorphism, and these are classified as a direct sum of indecomposable ones determined by their
Jordan type, corresponding to the decomposition of a finitely generated k[x]-module as a direct
sum of cyclic ones. These summands correspond to the C ∗ (S 1 ; k)-modules Ni .
3.4. Evenly graded C ∗ (S 1 ; k)-modules. Let k be a field. We will have to work with certain
non-perfect C ∗ (S 1 ; k)-modules in the sequel, and here we will prove a basic technical result (Proposition 3.23) concerning them. We need the following classical algebraic fact with R = k[x](x) .
Theorem 3.22. Let R be a discrete valuation ring with quotient field K. Then any torsion divisible
R-module is a direct sum of copies of K/R.
We refer to [KT08, Theorem 6.3] for a proof of Theorem 3.22. We will translate it into our
setting and prove the following.
Proposition 3.23. Let M be a C ∗ (S 1 ; k)-module. Suppose that πi M = 0 if i is odd. Then M
is a direct sum of copies of even shifts of the C ∗ (S 1 ; k)-module k (under the map of E∞ -rings
C ∗ (S 1 ; k) → k given by evaluation at a point).
Proof. We know, first, that M is a direct sum of copies of C ∗ (S 1 ; k)-modules whose base-change
to k is shifted discrete (Proposition 3.17), so we may assume this to begin with. That is, we may
assume that M ⊗C ∗ (S 1 ;k) k is concentrated in one degree, say n, in homotopy. In particular, under
the correspondence between C ∗ (S 1 ; k)-modules and local systems with the unipotence property,
M comes from a discrete k[x]-module P0 , on which x is locally nilpotent. Moreover, at most
one of ker x, coker x can nonzero because of the hypothesis on the homotopy groups on M , and
Construction 3.13, which describes how to get from P0 to M .
Since x is locally nilpotent, ker x is always nonzero (if P0 6= 0), so the conclusion must be that
coker x = 0, and we have an even shift of a discrete local system. In other words, P0 is a x-torsion
divisible k[x]-module. Any such is a direct sum of copies of k[x±1 ]/k[x] by Theorem 3.22.
f is the C ∗ (S 1 ; k)-module corresponding to the k[x]-module k[x±1 ]/k, then M
It follows that if M
f. It remains to argue that k ≃ M
f. In fact, our reasoning
is a direct sum of even shifts of copies of M
f
f.
shows that k must be a direct sum of copies of M , but clearly k is indecomposable, so k ≃ M
Corollary 3.24. Let M be a C ∗ (S 1 ; k)-module such that πi M = 0 if i is odd. Then the k-module
M ⊗C ∗ (S 1 ;k) k has the same property.
Proof. This follows from Proposition 3.23, but it could also have been seen directly.
16
AKHIL MATHEW
4. Residue fields
In this section, we will prove Theorems 1.2 and 1.3 on the existence of residue fields and the
detection of nilpotence. It will be convenient to work throughout with an extra assumption of
a degree two unit. In this case, the attachment of even cells can always be replaced with the
attachment of degree zero cells.
4.1. Definitions. Let A be a rational, noetherian E∞ -ring such that π2 (A) contains a unit. Fix a
prime ideal p ⊂ π0 (A).
Definition 4.1. A residue field for A is an object κ(p) ∈ CAlgA/ such that:
(1) The map π0 (A) → π0 (κ(p)) exhibits π0 (κ(p)) as the residue field of π0 (A) at p.
(2) π1 (κ(p)) = 0.
In particular, if k(p) is the residue field of π0 A at p, then π∗ (κ(p)) is a Laurent series ring on k(p)
on a generator in degree two.
In this section, we will show that residue fields for such E∞ -rings exist uniquely, and are sufficient
to detect nilpotence in Mod(A). The rest of the paper will use these residue fields to describe certain
invariants of Mod(A).
Remark 4.2. The name “residue field” is appropriate because of the perfect Künneth isomorphism
κ(p)∗ (M ) ⊗κ(p)∗ κ(p)∗ (N ) ≃ κ(p)∗ (M ⊗A N ),
M, N ∈ Mod(A);
indeed, there is a map from left to right which is an isomorphism for M = N = A, and both sides
define two-variable homology theories on Mod(A), so the natural map must be an isomorphism in
general. Alternatively, any κ(p)-module is a sum of shifts of free ones.
We describe the connection with the use of residue fields as in [BR08]. Given an even periodic
E∞ -ring A (not necessarily over Q) with π0 (A) regular noetherian, it is possible to form “residue
fields” of A as E1 -algebras in Mod(R), by successively quotienting by a regular sequence. These
residue fields have analogous properties of detecting nilpotence [Mat15, Cor. 2.6] and are quite
useful for describing invariants of Mod(A) (e.g., [BR05, BR08, Mat16]). These residue fields are
usually not E∞ -algebras in Mod(A). For example, in the “chromatic” setting, the associated residue
fields (such as the Morava K-theories K(n) for the E∞ -ring En ) are almost never E∞ .
Over the rational numbers, we are able to produce residue fields without such regularity hypotheses, and as E∞ -algebras. However, we will have to work a bit harder: the residue fields of such an
A will no longer in general be perfect as A-modules (or as E∞ -A-algebras), and we will have to use
a countable limiting procedure, together with the techniques from the previous sections.
Let A be as above. In order to construct a residue field for A for the prime ideal p ∈ Specπ0 A,
we may first localize at p, and assume that π0 A is local and that p is the maximal ideal. Then,
given generators x1 , . . . , xn ∈ π0 A for p, we will need to set them equal to zero by attaching 1-cells.
That of course will introduce new elements (in both π0 , π1 ) and we will have to kill them in turn.
This process will be greatly facilitated by the analysis in Section 3.
4.2. Detection of nilpotence. Given an E∞ -ring A, we start by reviewing what it means for a
collection of A-algebras to detect nilpotence, following ideas of [DHS88, HS98].
Definition 4.3. Let A be an E∞ -ring, and let A′ be an A-ring spectrum: that is, an associative
algebra object in the homotopy category of Mod(A). We say that A → A′ detects nilpotence if,
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
17
whenever T is an A-ring spectrum, then the map of associative rings
π∗ (T ) → π∗ (A′ ⊗A T )
has the property that any u ∈ π∗ (T ) which maps to a nilpotent element is nilpotent. More generally,
a collection of A-ring spectra {A′α }α∈S is said to detect nilpotence if any u ∈ π∗ (T ) which maps to
nilpotent elements under each map π∗ (T ) → π∗ (A′α ⊗A T ) is itself nilpotent.
For example, the nilpotence theorem (Theorem 1.1) states that the Morava K-theories and
homology (rational and mod p) detect nilpotence for A = S 0 . The original form (in [DHS88])
states that the E∞ -ring M U of complex bordism detects nilpotence by itself, again over S 0 .
As in [DHS88, §1], one has the following consequences of detecting nilpotence:
Proposition 4.4. Let {A′α }α∈S be a collection of A-ring spectra that detect nilpotence.
(1) Given a map of perfect A-modules φ : T → T ′ such that each 1A′α ⊗A φ : A′α ⊗A T → A′α ⊗A T ′
is nullhomotopic as a map of A-modules, then φ is smash nilpotent: φ⊗N : T ⊗N → T ′⊗N
is nullhomotopic for N ≫ 0.
(2) Given a self-map of perfect A-modules v : Σk T → T , if each 1A′α ⊗A v : Σk (A′α ⊗A T ) →
A′α ⊗A T is nilpotent in Mod(A), then v itself is nilpotent.
Example 4.5. Suppose A′ is an A-ring spectrum that detects nilpotence. Then A′ cannot annihilate any nonzero perfect A-module M ; in fact, that would force the identity M → M to be
nilpotent.
Moreover, one sees:
Proposition 4.6. Let A be an E∞ -ring.
(1) Let A1 → A2 → A3 → . . . be a diagram of A-ring spectra, such that the colimit A∞ = lim Ai
−→
has a compatible structure of an A-ring spectrum. If each Ai detects nilpotence over A, then
A∞ detects nilpotence over A.
(2) Let {A′α }α∈S be a collection of E∞ -A-algebras that detect nilpotence over A. For each
α ∈ S, let {A′′αβ }β∈Tα be a collection of E∞ -A′α -algebras that detect nilpotence over A′α .
o
n
Then the collection A′′αβ
of A-algebras detects nilpotence over A.
α∈S,β∈Tα
(3) Suppose the {A′α }α∈S are a collection of E∞ -A-algebras detecting nilpotence over A such
that each π∗ (A′α ) is a graded field. Then an A-ring spectrum A′′ detects nilpotence over R
if and only if A′′ ⊗A A′α 6= 0 for each α ∈ S.
Proof. By a graded field, we mean a graded ring which is either a field (concentrated in degree zero)
or k[t±1 ] for |t| > 0 and k a field. The third assertion then follows from the second, since any
nonzero ring spectrum over each A′α detects nilpotence over A′α . The proofs of the first and second
assertions are straightforward.
Finally, we need an important example of a pair that detects nilpotence.
Example 4.7. Let A be a rational E∞ -ring, and let x ∈ π0 (A). As before, the cofiber A/x
inherits the canonical structure of an E∞ -algebra under A, as A//x. The localization
A[x−1 ] always
inherits a natural E∞ -ring structure. The claim is that the pair of A-algebras A/x, A[x−1 ] detects
nilpotence.
To see this, let T be an A-ring spectrum, and let α ∈ πj (A). Suppose α maps to zero in
A[x−1 ] = T ⊗A A[x−1 ]. This means that xN α = 0 for N chosen large enough. Suppose also that α
18
AKHIL MATHEW
maps to zero in πj (T /x) = πj (T ⊗A A/x). This means that α = xβ for some β ∈ πj (T ). We then
have
α2N = αN αN = (xβ)N αN = β N xN αN = 0,
N
since x α = 0. In other words, α is nilpotent.
This example will be extremely important to us in making induction arguments on the Krull
dimension.
Example 4.8. Let A → A′ be a map of E∞ -rings. Suppose that A → A′ admits descent (Definition 3.6). Then A → A′ detects nilpotence; see for instance [Mat16, Prop. 3.26].
4.3. The main result. In this subsection, we prove the main technical result of this paper (Theorem 4.14): the existence of residue fields and the detection of nilpotence. We begin with a preliminary technical result.
Proposition 4.9. Let B be a rational E∞ -ring such that:
(1) π0 (B) is a field k.
(2) π2 (B) contains a unit u.
(3) π−1 (B) is a countably dimensional k-vector space.
Then there exists a sequence of E∞ -rings
(6)
B = B (0) → B (1) → B (2) → . . .
such that:
(1) Each B (i) satisfies the three hypotheses above on B.
(2) There exists an element yi ∈ π−1 (B (i−1) ) such that B (i) ≃ B (i−1) //yi .
(3) Given any element y ∈ π−1 (B), there exists N such that y maps to zero under the map
B → B (N ) .
Proof. Note first that, by Proposition 2.14, B naturally admits the structure of an E∞ -k-algebra.
Let u1 , u2 , · · · ∈ π−1 (B) be a k-basis. We define B (1) ≃ B//u1 = B ⊗Sym∗ k[−1] k via the
map Sym∗ k[−1] → B classifying u1 . Next, we define the E∞ -B (1) -algebra B (2) ≃ B (1) //u2 ≃
B (1) ⊗Sym∗ k[−1] k where the map Sym∗ k[−1] → B (1) classifies u2 . Inductively, we obtain a sequence
of E∞ -rings B (i) . We need to verify the above three conclusions on the B (i) . The second and third
conclusions are immediate from the construction.
Consider the Sym∗ k[−1]-module B under the map Sym∗ k[−1] → B classifying u1 . We have a
map
∗
k[t±1
2 ] ⊗k Sym k[−1] → B,
∗
of Sym k[−1]-modules. The cofiber C, by hypothesis, is a Sym∗ k[−1]-module whose homotopy
groups are concentrated entirely in odd degrees. In particular, we find by Proposition 3.23 that C
is a direct sum of odd shifts of the Sym∗ k[−1]-module k, so the cofiber sequence
k[t2±1 ] → B ⊗Sym∗ k[−1] k → C ⊗Sym∗ k[−1] k,
(which has to induce split exact sequences on the level of homotopy groups) implies that, at the level
of homotopy groups, B (1) = B ⊗Sym∗ k[−1] k has the same property as did B: the even homotopy
groups are given by the Laurent series ring. Moreover, the map π∗ (C) → π∗ (C ⊗Sym∗ k[−1] k) is
injective, so the {u2 , u3 , . . . , } remain nonzero and linearly independent in π−1 (B (1) ). We find
inductively that all the B (i) have homotopy groups entirely in odd degrees except for the Laurent
series over k.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
19
Lemma 4.10. Let k be a field of characteristic zero and let A be any E∞ -ring. Then the map
±1
HomCAlg (k[t±1
2 ], A) → HomRing∗ (π∗ (k[t2 ]), π∗ (A))
(7)
is a bijection.
Proof. This follows as in the proof Proposition 2.14. That is, using a transcendence basis for k
over Q, one sees that there exists a free E∞ -ring on a discrete Q-module V such that k is a filtered
colimit of étale Sym∗ (V )-algebras. The analog of (7) is easily seen to be true for Sym∗ (V ) and, by
the theory of étale extensions [Lur14, §7.5], it must hold for k. In other words, the map
HomCAlg (k, A) → HomRing∗ (k, π0 (A))
(8)
is a bijection. It is similarly easy to see that the analog holds for k replaced by Q[t±1
2 ]. Since
±1
k[t±1
]
≃
k
⊗
Q[t
],
we
may
conclude.
Q
2
2
Proposition 4.11. Let B satisfy the hypotheses of Proposition 4.9.
(1) Then there exists a map of E∞ -rings B → k[t±1
2 ] which detects nilpotence.
(2) If L is any field of characteristic zero, then the map
±1
π0 HomCAlg (B, L[t±1
2 ]) → HomRing∗ (π∗ (B), π∗ (L[t2 ]))
(9)
is a bijection.
Proof. Given B and the basis u1 , u2 , . . . , as above, we let B1 be the colimit lim B (i) of the sequence
−→
(6) obtained by applying Proposition 4.9. Then B1 satisfies the hypotheses of this proposition as
well. Moreover, the map π−1 (B) → π−1 (B1 ) is the zero map. Thus, we can repeat the above
sequential construction of Proposition 4.9 to B1 to produce a new sequence
(1)
B1 → B1
(2)
→ B1 → . . . ,
obtained by iteratively coning off the degree −1 elements in π∗ (B1 ). Let the colimit be the E∞ -B1 algebra B2 . Then B2 satisfies the hypotheses of Proposition 4.9, but π−1 (B1 ) → π−1 (B2 ) is zero.
Repeating the process, we get a sequence
(10)
B → B1 → B2 → . . . ,
whose colimit, finally, is the E∞ -ring k[t±1
2 ], since each of the maps is zero on π−1 .
We need to see that the map B → k[t±1
2 ] thus constructed detects nilpotence. For this, it suffices
to argue that each Bi → Bi+1 in the above sequence detects nilpotence, since detecting nilpotence is
preserved in filtered colimits. But Bi → Bi+1 is a filtered colimit of maps each of which is obtained
by coning off a degree −1 element, and these maps detect nilpotence by Proposition 3.8.
Finally, we need to analyze homotopy classes of maps B → L[t±1
2 ] where L is a field of characteristic zero. We have already shown that (9) is a surjection, so we only need to prove injectivity. For this,
we observe that any map B → L[t±1
2 ] extends over B → B1 . Indeed, B → B1 is a filtered colimit of
(i−1)
(i−1)
maps B
→B
//yi where |yi | = −1. Since π−1 L[t2±1 ] = 0, any map B (i−1) → L[t±1
2 ] can be
(i)
(i−1)
extended over B . In particular, π0 HomCAlg (B (i) , L[t±1
, L[t±1
2 ]) → π0 HomCAlg (B
2 ]) is a sur±1
jection. Taking inverse limits, it follows easily that π0 HomCAlg (B1 , L[t±1
2 ]) → π0 HomCAlg (B, L[t2 ])
±1
is a surjection too, as claimed. Applying this to Bi , we find that each map π0 HomCAlg (Bi , L[t2 ]) →
π0 HomCAlg (Bi−1 , L[t±1
2 ]), is a surjection, and taking limits, that the map
±1
±1
π0 HomCAlg (k[t±1
2 ], L[t2 ]) → π0 HomCAlg (B, L[t2 ])
20
AKHIL MATHEW
is a surjection, too. However, maps k[t2±1 ] → L[t±1
2 ] of E∞ -rings are determined by their action
on homotopy groups in view of Lemma 4.10. This proves uniqueness and completes the proof of
Proposition 4.11.
Proposition 4.12. Let A0 be a noetherian, rational E∞ -ring containing a unit in degree two.
Suppose π0 (A0 ) is a local artinian ring with residue field k. Let L be any field of characteristic zero.
Then:
(1) The natural map
±1
±1
π0 HomCAlg (A0 , L[t2±1 ]) → HomRing∗ (π∗ A0 , L[t±1
2 ]) ≃ HomRing∗ (k[t2 ], L[t2 ])
is a bijection.
(2) Any map A0 → L[t±1
2 ] of E∞ -rings detects nilpotence.
Proof. We first treat existence in case L = k. Our goal is to produce a map of E∞ -rings from A0
to k[t±1
2 ]. For this, we will need to kill the degree zero elements, and the degree −1 elements. We
will first kill the degree 0 elements by adding cells in dimension one, to reduce to the case where
there is nothing (except for k) in odd dimensions. Then, we will use a separate argument to kill
the odd homotopy.
We first make A0 into an E∞ -k-algebra, using Proposition 2.14. Given A0 , let m ⊂ π0 (A0 ) be
the maximal ideal, which is a finite-dimensional k-vector space. Consider the map
Sym∗ (m) → A0 ,
of E∞ -k-algebras, and form the pushout
def
A1 = A0 ⊗Sym∗ (m) k,
which has the same property as A0 : π∗ (A1 ) satisfies the desired noetherianness assumptions (in
fact, all the homotopy groups are finite-dimensional k-vector spaces), and π0 (A1 ) is local artinian
with residue field k by Theorem 2.21. Note that A0 → A1 admits descent in view of Proposition 3.7.
Let m1 ⊂ π0 (A1 ) be the maximal ideal, and continue the process with
def
A2 = A1 ⊗Sym∗ (m2 ) k,
and repeating this, we find a sequence of 2-periodic, noetherian E∞ -rings
A0 → A1 → A2 → . . . ,
where each Ai has the following properties:
(1) π0 (Ai ) is a local artinian ring with residue field k and π1 (Ai ) is a finite-dimensional k-vector
space.
(2) Ai+1 is obtained from Ai by attaching a 1-cell for each element in a k-basis of the maximal
ideal of π0 (Ai ), and thus Ai → Ai+1 admits descent.
(3) In particular, the map π0 (Ai ) → π0 (Ai+1 ) annihilates the maximal ideal of the former.
If we take the colimit A∞ = limi Ai , we find an E∞ -A-algebra A∞ such that π0 (A∞ ) = k. This
−→
process of iteratively attaching 1-cells has likely introduced elements in π−1 , but we know that
π−1 (A∞ ) is a countably dimensional k-vector space. Note that A0 → A∞ detects nilpotence, as it
is the sequential colimit of a sequence of objects in CAlgA0 / that admit descent. But now we can
appeal to Proposition 4.11 to obtain a map A∞ → k[t2±1 ] which detects nilpotence. This completes
the proof of existence.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
21
We may handle the first assertion (of which we now only need to prove injectivity of the map)
in a similar manner as in the proof of Proposition 4.11. That is, we observe that
HomCAlg (Ai , L[t2±1 ]) → HomCAlg (Ai−1 , L[t2±1 ])
(1)
(n)
(j)
is surjective for each i, because Ai ≃ Ai−1 //(xi , . . . , xi ) where the xi
limit, we find that the map
are nilpotent. In the
±1
HomCAlg (A∞ , L[t±1
2 ]) → HomCAlg (A0 , L[t2 ])
is surjective. But now we can apply the uniqueness statement of Proposition 4.11 to complete the
proof. That is, maps A∞ → L[t±1
2 ] are determined by their behavior on homotopy groups. So, if
we had two different maps A0 → L[t±1
2 ] inducing the same behavior on homotopy groups, we would
get two different maps A∞ → L[t±1
2 ] inducing the same behavior on homotopy groups, and this is
a contradiction.
Proposition 4.13. Let A be a noetherian, rational E∞ -ring containing a unit in degree two.
Suppose k is a field of characteristic zero. Then the map
±1
π0 HomCAlg (A, k[t±1
2 ]) → HomRing∗ (π∗ (A), k[t2 ])
is a bijection.
Proof. We begin with surjectivity. Suppose given a map π0 (A) → k. We want to realize this at the
level of E∞ -rings. By localizing, we may assume that π0 (A) is a local ring and that p ⊂ π0 (A) is its
maximal ideal; let k ′ be the residue field π0 (A)/p. By hypothesis, we have a map k ′ → k. Choose
ideal generators x1 , . . . , xn ∈ p and use them to construct a map
Q[u1 , . . . , un ] → A,
ui 7→ xi .
We let A0 be the E∞ -ring A ⊗Q[u1 ,...,un ] Q ≃ A//(u1 , . . . , un ). Then we get a map A → A0 →
k ′ [t±1
2 ] by Proposition 4.12 realizing the map on π∗ desired. We can compose this with the map
±1
k[t±1
2 ] → k[t2 ].
Suppose now we have two distinct E∞ -maps A → k[t±1
2 ] realizing the same map on π∗ . The
same argument as before shows that both maps extend to (necessarily distinct) maps A0 → k[t±1
2 ]
realizing the same map on homotopy groups, but this contradicts Proposition 4.12.
We can now prove our main result.
Theorem 4.14. Let A be a noetherian, rational E∞ -ring containing a unit in degree two.
(1) For each prime ideal p ⊂ π0 (A), a residue field κ(p) ∈ CAlgA/ for A at p exists and is
unique up to homotopy.
(2) The E∞ -A-algebras κ(p) detect nilpotence.
Proof. The existence and uniqueness of residue fields is a consequence of Proposition 4.13, since we
can of course construct them uniquely in the setting of commutative algebra (i.e., at the level of
homotopy groups).
Finally, we need to show that the residue fields detect nilpotence. If π0 A is local artinian, then we
have already seen this (Proposition 4.12). Now, assume the result on detection of nilpotence proved
for all noetherian A with the Krull dimension of π0 (A) at most n − 1. We will then prove it for
dimension ≤ n. In fact, we may assume that π0 A is noetherian local of Krull dimension ≤ n. Choose
x ∈ π0 A such that π0 A/(x) has Krull dimension ≤ n−1. As we saw in Example 4.7, the pair of E∞ A-algebras A//x, A[x−1 ] detect nilpotence over A, and each of these is noetherian with π0 having
22
AKHIL MATHEW
Krull dimension ≤ n − 1. Therefore, the residue fields of the E∞ -rings A/x, A[x−1 ] are sufficient to
detect nilpotence over each of them, and thus detect nilpotence over A by Proposition 4.6.
Given any rational noetherian E∞ -ring A, in order to prove that the residue fields {κ(p)}p∈Specπ0 A
detect nilpotence over A, it suffices to reduce to the case where π0 A is local, and thus of finite Krull
dimension, so that what we have already done suffices to show that the residue fields detect nilpotence.
Proposition 4.15. Let A be a rational noetherian E∞ -ring containing a unit in π2 . Suppose
B ∈ CAlgA/ is noetherian as well. Let p ∈ Specπ0 A. Let k(p) be the associated algebraic residue
field and let κ(p) ∈ CAlgA/ be the topological one. Then the following are equivalent:
(1) π0 (B) ⊗π0 (A) k(p) 6= 0.
(2) B ⊗A κ(p) 6= 0.
Proof. There is a map of commutative rings π0 (B) ⊗π0 (A) k(p) → π0 (B ⊗A κ(p)), so if the latter is
nonzero, clearly the former is as well. Suppose the former is nonzero now. Without loss of generality,
we may assume that π0 (A) is local and p is its maximal ideal. Let (x1 , . . . , xn ) be a system of
generators for p and let A0 = A//(x1 , . . . , xn ). Then B ⊗A A0 6= 0 as π0 (B)/(x1 , . . . , xn ) 6= 0, in
view of Theorem 2.21; it is here that we use that B is noetherian. However, the map A0 → κ(p)
detects nilpotence as π0 (A0 ) is local artinian, so that if B ⊗A A0 6= 0, then B ⊗A κ(p) 6= 0 too.
Remark 4.16. Proposition 4.15 is false if we do not assume B is noetherian. We refer to Section 8.2
for a counterexample: the map Q[x, y] → A constructed is an isomorphism on π0 , but B//(x, y) = 0.
Corollary 4.17. Let A be a rational noetherian E∞ -ring containing a unit in π2 . Given two
different prime ideals p, q ⊂ π0 A, the tensor product κ(p) ⊗A κ(q) is contractible.
Corollary 4.18. Let A be a rational noetherian E∞ -ring containing a unit in degree two. Let
A′ , A′′ ∈ CAlgA/ be noetherian. Then the following are equivalent:
(1) π0 (A′ ) ⊗π0 (A) π0 (A′′ ) 6= 0.
(2) A′ ⊗A A′′ 6= 0.
Proof. Consider the π0 (A)-algebra R = π0 (A′ ) ⊗π0 (A) π0 (A′′ ). Since it is a nonzero algebra, there
exists p ∈ Specπ0 (A) such that R ⊗π0 (A) k(p) 6= 0, where k(p) is the residue field of π0 A at p. Thus,
π0 (A′ ) ⊗π0 (A) k(p) 6= 0 and π0 (A′′ ) ⊗π0 (A) k(p) 6= 0. By Proposition 4.15, the E∞ -rings A′ ⊗A κ(p)
and A′′ ⊗A κ(p) are nonzero, and thus their relative tensor product over κ(p) is nonzero. Thus,
(A′ ⊗A A′′ ) ⊗A κ(p) 6= 0, so that A′ ⊗A A′′ 6= 0 as well.
Replacing A by A[t±1
2 ], we also get the following result, which will be important for the next
section.
Corollary 4.19. Let A be a rational noetherian E∞ -ring (not necessarily containing a unit in
degree two). Let A′ , A′′ ∈ CAlgA/ be noetherian. Then the following are equivalent:
(1) πeven (A′ ) ⊗πeven(A) πeven (A′′ ) 6= 0.
(2) A′ ⊗A A′′ 6= 0.
Remark 4.20. Given a rational, noetherian E∞ -ring A containing a unit in π2 , we constructed a
family of residue fields {κ(p)}p∈SpecA that detect nilpotence in the ∞-category Mod(A). One could
ask if one has in fact a Bousfield decomposition. That is, if an A-module M (not necessarily perfect)
has the property that κ(p)∗ (M ) = 0 for all p ∈ SpecA, does that force M to be contractible? In the
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
23
regular and even periodic case, this is known (e.g., [Mat15, Prop. 2.5]). The analog over the sphere
fails: there are noncontractible spectra that smash to zero with every residue field, for instance the
Brown-Comenentz dual I of the sphere ([HS99, Appendix B]). We do not know what happens in
Mod(A).
5. A thick subcategory theorem
In this section, we use Theorem 4.14 to obtain the classification of thick subcategories of perfect
modules over a rational, noetherian E∞ -ring.
5.1. Review of the axiomatic argument. Let A be an E∞ -ring together with a collection
{κ(p)∗ } of multiplicative homology theories on Mod(A), satisfying perfect Künneth isomorphisms,
such that together they detect nilpotence over A. In this case, it is well-known that they are sufficient
to detect thick subcategories as well. We note that these are the same as thick tensor-ideals, since
the unit generates Modω (A) as a thick subcategory. In particular, consider M, N ∈ Modω (A). We
recall:
Theorem 5.1 (Hopkins-Smith-Hovey-Palmieri-Strickland [HPS97, Th. 5.2.2 and Cor. 5.2.3]).
Suppose M, N ∈ Modω (A) are such that whenever κ(p)∗ (M ) 6= 0, then κ(p)∗ (N ) 6= 0 too. Then the
thick subcategory that N generates in Modω (A) contains M .
At least under noetherian hypotheses, every thick subcategory C ⊂ Modω (A) is then determined
by a subset of the {κ(p)∗ } (the κ(p)∗ such that there exists X ∈ C with κ(p)∗ (X) 6= 0), and the
classification of thick subcategories reduces to the determination of which subsets arise from thick
subcategories; or equivalently, which subsets of the {κ(p)} arise as {p : κ(p)∗ (M ) 6= 0} for some
M ∈ Modω (A).
5.2. The even periodic case. The main subtleties of the present section will revolve around the
grading. As a result, we start with the simplest case, where the ring contains a unit in degree
two. This is a direct consequence of the work of the previous section and the axiomatic argument,
Theorem 5.1.
Construction 5.2. Let A be a rational, noetherian E∞ -ring containing a unit in degree two.
Let Z ⊂ Specπ0 A be a specialization-closed subset. We define a thick subcategory Modω
Z (A) ⊂
Modω (A) consisting of modules M such that π0 (M ) ⊕ π1 (M ) is set-theoretically supported on a
closed subset of Z.
Construction 5.2 clearly defines thick subcategories of Modω (A). We start by noting that they
can also be defined in terms of the residue fields of A.
Proposition 5.3. Let A be a rational, noetherian 2-periodic E∞ -ring. Let M be a perfect A-module.
Then the following are equivalent, for p ∈ Specπ0 A:
(1) Mp 6= 0.
(2) κ(p)∗ (M ) 6= 0 (where κ(p) ∈ CAlgA/ is the residue field of Theorem 4.14).
Proof. Without loss of generality, we can assume that π0 (A) is local and that p is the maximal
ideal of π0 (A). Then we need to show that if κ(p)∗ (M ) = 0, then M itself is contractible, a form
of Nakayama’s lemma. Without loss of generality, we can assume that π0 (A) is complete local,
because the completion is faithfully flat over A [Mat80, Theorem 56, §24].
Let x1 , . . . , xn ∈ π0 (A) be generators for the maximal ideal p. Then it suffices to show that
M/(x1 , . . . , xn ), which is the base-change of M to A//(x1 , . . . , xn ), is contractible, because M is
24
AKHIL MATHEW
(x1 , . . . , xn )-adically complete. In particular, we may replace A with A//(x1 , . . . , xn ) and thus
assume that π0 (A) is actually local artinian. We thus reduce to this case.
But if π0 (A) is local artinian, we know that the map A → κ(p) actually detects nilpotence: in
particular, it cannot annihilate a nonzero perfect A-module (Example 4.5). So, if κ(p)∗ (M ) = 0,
then M is contractible.
Proposition 5.4. Let A be a rational, noetherian E∞ -ring containing a unit in degree two. Then
the thick subcategories of Modω (A) are in natural bijection with the specialization-closed subsets of
Specπ0 A, via the correspondence given in Construction 5.2.
Proof. By Theorem 5.1 and Proposition 5.3, it follows that if M, N ∈ Modω (A) and the set-theoretic
support of π0 (M )⊕π1 (M ) contains that of π0 (N )⊕π1 (N ), then N belongs to the thick subcategory
generated by M .
Next, we argue that any closed subset of Specπ0 A is realized as the support of π0 M for some
M ∈ Modω (A). If the closed subset is defined by the ideal (x1 , . . . , xn ) ∈ π0 (A), then we can
take M = A/(x1 , . . . , xn ), thanks to Theorem 2.21. This implies that if Z, Z ′ ⊂ Specπ0 A are two
distinct specialization-closed subsets, say Z \ Z ′ 6= ∅ then there exists a module M ∈ Modω (A)
ω
which belongs to Modω
Z (A) \ ModZ ′ (A). In other words, the map of Construction 5.2 from subsets
to thick subcategories is injective.
Finally, if C ⊂ Modω (A) is a thick subcategory, we let Z be the specialization-closed subset of
those p ∈ Specπ0 (A) such that there there exists M ∈ C with κ(p)∗ (M ) 6= 0. Clearly, C ⊂ Modω
Z (A).
To see equality, fix an arbitrary M ∈ Modω
Z (A). Then there exists a set {Mα }α∈S of objects in C
such that the support of π0 M ⊕π1 M is contained in the union of the supports of the {π0 Mα ⊕π1 Mα }.
Therefore, since π0 (A) is noetherian, there exists a finite subcollection S ′ ⊂ S such that the same
conclusion
holds, and Theorem 5.1 implies that M belongs to the thick subcategory generated by
L
α∈S ′ Mα ∈ C.
5.3. Graded rings. In the rest of the section, we will explain how to adapt the argument of
Proposition 5.4 to the general case, where we do not assume the existence of a unit in degree two.
We begin with a review of some facts about graded rings. We will work with graded rings which
are commutative (in the ungraded sense) such as πeven of an E∞ -ring.
Definition 5.5. Let R∗ be a commutative, graded ring. The topological space GrSpec(R∗ ) consists
of the homogeneous prime ideals of R∗ . The topology on GrSpec(R∗ ) is defined by taking as a basis
def
of open sets the subsets V (a) = {p ∈ GrSpec(R∗ ) : a ∈
/ p} for each homogeneous element p. Note
that GrSpec(R∗ ) ⊂ SpecR∗ and the inclusion map is continuous (for the usual Zariski topology on
the latter).
Given a graded ring R∗ , the space GrSpec(R∗ ) is also the underlying topological space [LMB00,
Ch. 5] of the algebraic stack (SpecR∗ )/Gm , where the Gm -action on SpecR∗ is given by the grading
of R∗ . Given a point of (SpecR∗ )/Gm , represented a map Speck → (SpecR∗ )/Gm where k is a
field, one obtains a map of graded rings R∗ → k[t±1 ] where |t| = 1. The kernel of this map is
a homogeneous prime ideal of GrSpec(R∗ ), which gives the correspondence between points of the
stack and GrSpec(R∗ ).
Example 5.6. Suppose R∗ is nonnegatively graded, i.e., Ri = 0 for i < 0, and suppose R0 is a
field. Then GrSpec(R
∗ ) is the union of Proj(R∗ ) and one additional point, corresponding to the
L
irrelevant ideal i>0 Ri .
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
25
Definition 5.7. Let R∗ be a commutative, graded ring. A collection C ⊂ GrSpec(R∗ ) of homogeneous prime ideals of R∗ is closed under specialization if, whenever p ∈ C and q ⊃ p is a larger
homogeneous prime ideal, then q ∈ C too. C ⊂ GrSpec(R∗ ) is closed under specialization if and
only if it is a union of closed subsets.
We next observe that there is a notion of “support” in the graded setting.
Definition 5.8. Suppose R∗ is noetherian and M∗ is a finitely generated graded R∗ -module. Let
Supp(M∗ ) denote the collection of all p ∈ GrSpec(R∗ ) such that the localization (M∗ )p does not
vanish. Then Supp(M∗ ) ⊂ GrSpec(R∗ ) is closed as it is the intersection of the usual support of M∗
in Spec(R∗ ) with GrSpec(R∗ ) ⊂ Spec(R∗ ).
A priori, the construction of the localization (M∗ )p (which is not a graded R∗ -module) is somewhat unnatural. However, it is easy to see that the condition that (M∗ )p 6= 0 is equivalent to the
condition that the localization of M∗ at the multiplicative subset S = {r ∈ R∗ homogeneous : r ∈
/ p}
should not vanish, and this latter localization is naturally a graded R∗ -module.
We will next need the notion of a graded-local ring.
Definition 5.9. R∗ is graded-local if it has a unique maximal homogeneous ideal. R∗ is a graded
field if either:
(1) R∗ = k, concentrated in degree zero, where k is a field.
(2) R∗ = k[u, u−1 ] where |u| > 0 and k is a field.
It is easy to see that a graded ring is a graded field if and only if the zero ideal is a maximal
homogeneous ideal. In fact, this condition implies that any homogeneous element is either zero or
a unit. As a result, given any p ∈ GrSpec(R∗ ), we can form the graded R∗ -algebra R∗[p] /p, where
R∗[p] is the localization of R∗ at the set of homogeneous elements not in p. This is a graded field.
We will next need to understand a little about the interaction between homogeneous and inhomogeneous prime ideals.
Construction 5.10. Let q ∈ Spec(R∗ ) be a prime ideal (not assumed homogeneous). We define
a homogeneous prime ideal p ∈ GrSpec(R∗ ) such that x ∈ p if and only if each homogeneous
component of x belongs to q. Clearly, p ⊂ q is the maximal homogeneous ideal contained in q.
In the language of stacks, we have a quotient map SpecR∗ → (SpecR∗ )/Gm , which induces a
map on points SpecR∗ → GrSpec(R∗ ). This map sends q 7→ p as above.
Finally, to use both graded and ungraded techniques, we need the following construction.
Construction 5.11. Let R∗ be a commutative, graded noetherian ring. Let R∗′ be the graded ring
R∗ [u±1 ] where |u| = 1. Then graded R∗′ -modules are canonically in correspondence with ungraded
R∗ -modules. For example, let q ∈ Spec(R∗ ) be a prime ideal, not assumed homogeneous. We let
k(q)∗ denote the graded R∗′ -module corresponding to ungraded R∗ -module which is the residue field
of R∗ at q.
Given an ungraded R∗ -module M , we can form a graded R∗′ -module M∗′ as in Construction 5.11
and restrict to get a graded R∗ -module (still denoted M∗′ ), such that Mn′ = M for every n. The
following elementary lemma will be crucial in the next subsection.
Lemma 5.12. Let R∗ be a graded, commutative noetherian ring. Let R∗′ = R∗ [u±1 ] where |u| = 1.
Let q ∈ SpecR∗ and let p ∈ GrSpec(R∗ ) be the homogeneous part. Define the graded R∗′ -algebras
k(p)∗ , k(q)∗ as in Construction 5.11. Then k(p)∗ ⊗R∗ k(q)∗ 6= 0.
26
AKHIL MATHEW
Proof. We can assume without loss of generality that R∗ is graded-local with maximal homogeneous
ideal p. After replacing R∗ with R∗ /p, we can assume p = 0. In this case, R∗ is a graded field so
that the tensor product of any two nonzero graded R∗ -modules (e.g., k(p)∗ , k(q)∗ , under pull-back
from R∗ → R∗′ ) is nonzero.
5.4. The thick subcategory theorem. Let A be a rational, noetherian E∞ -ring. The purpose of
this subsection is to give the proof that thick subcategories of Modω (A) correspond to specializationclosed subsets of GrSpec(πeven (A)), without assuming the existence of a unit in degree two. We
first state formally the map that realizes the correspondence. We note that our results prove that
the map from the spectrum (cf. [Bal05]) of the ⊗-triangulated category associated to Modω (A) to
the homogeneous spectrum of π∗ (A), as constructed by Balmer [Bal10], is an isomorphism.
Definition 5.13. Let M ∈ Modω (A). We define the support Supp(M ) to be the support (in the
sense of Definition 5.8) of the graded πeven (A)-module π∗ (M ). We will denote this by SuppM ; it
is a subset of GrSpec(πeven (A)). Given a specialization-closed subset Z ⊂ GrSpec(πeven (A)), we
ω
define Modω
Z (A) ⊂ Mod (A) to be the full subcategory spanned by those perfect A-modules M
with SuppM ⊂ Z.
Theorem 5.14. Let A be a rational, noetherian E∞ -ring. Then the construction Z 7→ Modω
Z (A)
defines a correspondence between the thick subcategories of Modω (A) and specialization-closed subsets of GrSpec(πeven (A)).
The primary goal of this section is to give a proof of Theorem 5.14, which we already did (in
Proposition 5.4) in case π2 (A) contains a unit.
To begin with, we will need to discuss residue fields for A. Let A be a noetherian rational E∞ ring and p ⊂ πeven (A) a homogeneous prime ideal. We form the E∞ -ring A′ = A[t±1
2 ] and we then
have
π0 A′ ≃ πeven A.
In particular, p becomes a prime ideal of π0 A′ . As a result, in view of Theorem 4.14, we can
construct a residue field κ(p) of A′ at p and we obtain maps A → A′ → κ(p). Rather than
considering the {κ(p)} as E∞ -A′ -algebras, we consider them as E∞ -A-algebras. They satisfy a
perfect Künneth isomorphism as homology theories on Mod(A), as before.
Lemma 5.15. Let q ⊂ πeven (A) be an inhomogeneous prime ideal and let p ⊂ πeven (A) be the
homogeneous part. Let κ(q), κ(p) ∈ CAlgA′ denote the respective residue fields, which we regard as
E∞ -A-algebras under A → A′ . Then κ(q) ⊗A κ(p) 6= 0.
Proof. This follows from Corollary 4.19 and Lemma 5.12.
As a result, we can now show that the residue fields κ(p) ∈ CAlgA/ for p ∈ GrSpec(πeven (A))
suffice to detect nilpotence.
Theorem 5.16. The {κ(p)}, as p ranges over the homogeneous prime ideals in Specπeven (A),
detect nilpotence over A.
This is not an immediate consequence of Theorem 4.14 applied to A′ , because the homogeneous
p do not exhaust all the prime ideals of π0 (A′ ). In other words, the κ(p) in question do not detect
nilpotence over A′ .
Proof. We know that the {κ(q)} ⊂ CAlgA′ for q ∈ Specπeven A = π0 A′ detect nilpotence over A′
(Theorem 4.14), and thus over A. Thus, in order to prove that the {κ(p)} for p ranging over the
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
27
homogeneous prime ideals detect nilpotence over A, we appeal to the third part of Proposition 4.6
and Lemma 5.15.
We note now that if M ∈ Modω (A), then the support of M in GrSpec(πeven (A)) is equivalently
the set of p ∈ GrSpec(πeven (A)) such that M ⊗A κ(p) 6= 0; this is a consequence of Proposition 5.3.
Proof of Theorem 5.14. In particular, it now follows formally (via the axiomatic argument given
in Theorem 5.1, together with the noetherianness of GrSpec(πeven (A)) from Theorem 5.16 that a
thick subcategory of Modω (A) is determined by a subcollection of the {κ(p)} as p ranges over the
homogeneous prime ideals of πeven (A). It remains to determine what subsets are allowed to arise.
We will show that those subsets are precisely those which are closed under specialization. To
see this, we need to show that every closed subset of GrSpec(πeven (A)) (associated to a homogeneous ideal I ⊂ πeven (A)) can be realized as the support of some M , but this follows by forming
A/(x1 , . . . , xn ) where x1 , . . . , xn ∈ πeven (A) generate I. In particular, this completes the proof of
Theorem 5.14.
6. Galois groups
Let A be an E∞ -ring such that π0 (A) has no nontrivial idempotents. In [Mat16], we introduced
the Galois group π1 Mod(A) of A, a profinite group defined “up to conjugacy” (canonically as a
profinite groupoid ), by developing a version of the étale fundamental group formalism. The Galois
group π1 Mod(A) has the property that if G is a finite group, then to give a continuous group
homomorphism π1 Mod(A) → G is equivalent to giving a faithful G-Galois extension of A in the
sense of Rognes [Rog08]. More generally, we introduced ([Mat16, Def. 6.1]) the notion of a finite
cover of an E∞ -ring A, as a homotopy-theoretic version of the classical notion of a finite étale
algebra over a commutative ring. A continuous action of π1 Mod(A) on a finite set is equivalent to a
finite cover of the E∞ -ring A. The Galois group can be a fairly sensitive invariant of E∞ -rings; for
instance ([Mat16, Ex. 7.21]) two different E∞ -structures on the same E1 -ring can yield different
Galois groups, and computing it appears to be a subtle problem in general. Here, we will show that
the Galois group is much less sensitive over the rational numbers, under noetherian hypotheses.
The Galois group comes with a surjection
(11)
π1 Mod(A) ։ π1et Specπ0 (A),
since every algebraic Galois cover of Specπ0 (A) can be realized topologically. More generally, to
every finite étale π0 (A)-algebra A′0 one can canonically associate A′ ∈ CAlgA/ such that π0 A′ ≃ A′0
and such that πk A′ ≃ A′0 ⊗π0 A πk A [Lur14, §7.5]. This yields a full subcategory of the category
of finite covers which corresponds to the above surjection. In general, however, it is an insight of
Rognes that the above surjection has a nontrivial kernel: that is, there exist finite covers that do
not arise algebraically in this fashion. A basic example is the complexification map KO → KU .
In [Mat16], we computed Galois groups in certain instances. Our basic ingredient ([Mat16, Th.
6.30]) was a strengthening of work of Baker-Richter [BR08] to show that the Galois theory is entirely
algebraic for even periodic E∞ -rings with regular π0 , using the theory of residue fields. Over the
rational numbers, the methods of the present paper enable one to construct these “residue fields”
without regularity assumptions. In particular, we will show in this section that, for noetherian
rational E∞ -rings, the computation of the Galois group can be reduced to a problem of pure
algebra. (For instance, we will show that (11) is an isomorphism if A contains a unit in degree
two.) In general, (11) will not be an isomorphism, because over Q, it is permissible to adjoin square
28
AKHIL MATHEW
roots of invertible elements in homotopy in degrees divisible by four, for instance. But we will see
that such issues of grading are the only failure of (11) to be an isomorphism.
6.1. Review of invariance properties. To obtain the results of the present section, we will need
some basic tools for working with Galois groups, which will take the form of the “invariance results”
of [Mat16]. For example, we will need to know that killing a nilpotent degree zero class does not
affect the Galois group. For convenience, we will assume that all E∞ -rings A considered in this
section have no nontrivial idempotents in π0 , so that we can speak about a Galois group.
Theorem 6.1. Let A be a rational E∞ -ring and let x ∈ π0 A be a nilpotent element. Then the map
A → A//x induces an isomorphism on Galois groupoids.
Proof. This is [Mat16, Theorem 8.13], for the map Q[[t]] → A sending t 7→ x.
Proposition 6.2. Let A be a rational E∞ -ring and let x ∈ π−1 A be a class. Then the map
A → A//x ≃ A ⊗Sym∗ Q[−1] Q,
obtained by coning off x, induces a surjection on Galois groups.
Proof. By [Mat16, §8.1], it suffices to show that the map C ∗ (S 1 ; Q) ≃ Sym∗ Q[−1] → Q is universally connected : that is, for any A ∈ CAlgC ∗ (S 1 ;Q)/ , the natural map A → A ⊗C ∗ (S 1 ;Q) Q induces
an isomorphism on Idem.
Since C ∗ (S 1 ; Q) → Q admits descent, the set Idem(A) of idempotents in A is the equalizer of
the two maps
→
A ⊗C ∗ (S 1 ;Q) Q →
A ⊗C ∗ (S 1 ;Q) Q ⊗C ∗ (S 1 ;Q) Q ≃ A ⊗C ∗ (S 1 ;Q) Q [t].
This is a reflexive equalizer, and one of the maps is the natural inclusion
A ⊗C ∗(S 1 ;Q) Q → A ⊗C ∗ (S 1 ;Q) Q [t],
which induces an isomorphism on idempotents. It follows that all the maps in the reflexive equalizer
are isomorphisms and thus the two forward maps are equal, proving that A → A⊗C ∗ (S 1 ;Q) Q induces
an isomorphism on idempotents.
6.2. The periodic case. We are now ready to show (Theorem 6.4 below) that the Galois theory
of a noetherian rational E∞ -ring containing a degree two unit is algebraic.
Lemma 6.3. Let A be a rational, noetherian E∞ -ring containing a unit in π2 such that π0 A is
local artinian. Then the Galois theory of A is algebraic.
Proof. The strategy is to imitate the proof of Theorem 4.14, while cognizant of the invariance
results for Galois groups reviewed in the previous subsection. Namely, we not only showed that A
had a residue field, but we constructed it via a specific recipe. Let k be the residue field of π0 A.
In proving Theorem 4.14 (that is, in the course of the proof of Proposition 4.12), we first formed
a sequence of rational, noetherian, E∞ -rings with artinian π0 ,
A = A0 → A1 → A2 → · · · → A∞ = lim Ai ,
−→
such that:
(1) Ai+1 is obtained from Ai by attaching 1-cells along a finite number of nilpotent elements
in π0 (Ai ).
(2) All the π0 (Ai ) are local artinian rings with residue field k, and each map π0 (Ai ) → π0 (Ai+1 )
annihilates the maximal ideal.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
29
By Theorem 6.1, at no finite stage do we change the Galois group; each map A → Ai induces
an isomorphism on Galois groups. Now, by [Mat16, Th. 6.21], the Galois group is compatible with
filtered colimits and therefore A → A∞ induces an isomorphism on Galois groups.2
Now, the E∞ -ring A∞ has the properties of Proposition 4.9: it has a unit in degree two, its π0
is isomorphic to k, and π1 is countably dimensional. We showed in the proof of Proposition 4.9
that by killing degree −1 cells repeatedly and forming countable colimits, and repeating countably
many times, we could start with A∞ and reach k[t±1
2 ]. It follows by Proposition 6.2 (along with
the compatibility of Galois groups and filtered colimits, again) that the map
A∞ → k[t±1
2 ],
induces a surjection on Galois groups. But the Galois group of k[t±1
2 ] is algebraic (i.e., Gal(k/k)) in
view of the Künneth isomorphism [Mat16, Prop. 6.28], so the Galois group of A∞ must be bounded
by Gal(k/k), and therefore that of A must be, too.
We can now prove the main result of the present subsection.
Theorem 6.4. Let A be a rational, noetherian E∞ -ring containing a unit in degree two. Then the
Galois theory of A is algebraic, i.e., π1 Mod(A) ≃ π1et Specπ0 (A).
Proof. Fix a finite cover A → A′ of E∞ -rings. We need to show that A′ is flat over A: that is, the
natural map π0 (A) → π0 (A′ ) is flat, and the map π∗ (A) ⊗π0 (A) π0 (A′ ) → π∗ (A′ ) is an isomorphism.
This is a local question, so we may assume that π0 (A) is a local noetherian ring. Moreover, by
completing A at the maximal ideal m ⊂ π0 A, we may assume that π0 A is complete; we may do this
because the completion of any noetherian local ring is faithfully flat over it.
Let k be the residue field of the (discrete) commutative ring π0 A. The étale fundamental group
of Specπ0 A is naturally isomorphic to that of Speck, via the inclusion Speck ֒→ Specπ0 A as the
closed point, since π0 A is a complete local ring. Let x1 , . . . , xn ∈ π0 A be generators for the maximal
ideal. Consider the tower of E∞ -A-algebras
· · · → A//(x31 , . . . , x3n ) → A//(x21 , . . . , x2n ) → A//(x1 , . . . , xn ),
whose inverse limit is given by A itself (by completeness). Observe that the maps at each stage are
not uniquely determined. For instance, to give a map A//(x21 , . . . , x2n ) → A//(x1 , . . . , xn ) amounts
to giving nullhomotopies of each of x21 , . . . , x2n in A//(x1 , . . . , xn ), and there are many possible
choices of nullhomotopies. One has to make choices at each stage.
Denote the E∞ -algebras in this tower by {Am }. As a result, the equivalence A ≃ lim Am leads
←−
to a fully faithful imbedding
Modω (A) ⊂ lim Modω (Am ),
←−
from the ∞-category Modω (A) of perfect A-modules into the homotopy limit of the ∞-categories
Modω (Am ) of perfect Am -modules. As discussed in [Mat16, §7.1], this implies that if we show that
the Galois group of each Am is equivalent to that of k (i.e., is algebraic), then the Galois group of
A itself is forced to be algebraic. This, however, is precisely what we proved in Lemma 6.3 above.
2In fact, all we need for the proof of this lemma is that A → A
∞ induces a surjection on Galois groups. This
does not require the obstruction theory used in proving [Mat16, Th. 6.21], and is purely formal.
30
AKHIL MATHEW
6.3. The general case. In the previous parts of this section, we showed that the Galois theory of
a rational noetherian E∞ -ring A containing a unit in π2 was entirely algebraic. In this subsection,
we will explain the modifications needed to handle the case where we do not have a unit in π2 ;
in this case, the structure of the entire homotopy ring π∗ A (rather than simply π0 A) intervenes.
We will begin with some generalities from [BR07] which, incidentally, shed further light on Galois
groups of general E∞ -rings.
Let R∗ be a commutative, Z-graded ring (not graded-commutative!), such as πeven (A) for A ∈
CAlg. We start by setting up a Galois formalism for R∗ that takes into account the grading.
Definition 6.5. A graded finite étale R∗ -algebra is a commutative, graded R∗ -algebra R∗′ such
that, as underlying commutative rings, the map R∗ → R∗′ is finite étale.
We list two fundamental examples:
Example 6.6. Given a finite étale R0 -algebra R0′ , then one can build from this a graded finite
def
étale R∗ -algebra via R∗′ = R0′ ⊗R0 R∗ .
±1
′
Example 6.7. Let R∗ = Z[1/n, x±1
n ] where |xn | = n. Then the map R∗ → R∗ = Z[1/n, y1 ], xn 7→
n
y1 is graded finite étale. In other words, one can adjoin nth roots of invertible generators in degrees
divisible by n, over a Z[1/n]-algebra.
Consider the category CR∗ of graded finite étale R∗ -algebras and graded R∗ -algebra homomorphisms. We start by observing that it is opposite to a Galois category. One can formulate this in
the following manner. The grading on R∗ determines an action of the multiplicative group Gm on
SpecR∗ , in such a manner that to give a quasi-coherent sheaf on the quotient stack (SpecR∗ )/Gm
is equivalent to giving a graded R∗ -module. To give a finite étale cover of the quotient stack
(SpecR∗ )/Gm is equivalent to giving a graded R∗ -algebra which is finite étale over R∗ . In other
words, graded finite étale R∗ -algebras are equivalent to finite étale covers of the stack SpecR∗ /Gm .
Definition 6.8. We define the graded étale fundamental group π1et,gr SpecR∗ to be the étale fundamental group of the stack (SpecR∗ )/Gm .
Example 6.9. Suppose R∗ contains a unit in degree 1. In this case, R∗ ≃ R0 ⊗Z Z[t±1 ] where
|t| = 1. In particular, the quotient stack (SpecR∗ )/Gm is simply SpecR0 , so the graded étale
fundamental group of R∗ is the fundamental group of SpecR0 .
Now, let Aeven = πeven (A), constructed with the degree halved (so that (Aeven )1 = π2 (A)) and
fix a graded, finite étale Aeven -algebra A′∗ . We will construct an E∞ -A-algebra A′ equipped with
an isomorphism πeven A′ ≃ A′∗ which is a finite cover of A, generalizing the construction that starts
with a finite étale π0 A-algebra and obtains a finite cover of A. As a result, we will obtain:
Theorem 6.10 (Baker-Richter [BR07]). There is a natural fully faithful imbedding from the category of graded, finite étale Aeven -algebras into the category of finite covers of A in E∞ -rings.
Dually, we obtain surjections of profinite groups
(12)
π1 Mod(A) ։ π1et,gr (SpecAeven ) ։ π1et (Specπ0 A),
refining the surjection (11). Since our formulation of the (essentially same) result is slightly different
from that of Baker-Richter, we give the deduction from their statement below.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
31
Proof. Start with a G-Galois object A′∗ in the category of graded, finite étale Aeven -algebras (for G
e′ = A′ ⊗Aeven π∗ (A) be the gradeda finite group). Then A′∗ is a projective Aeven -module. Let A
∗
∗
commutative algebra obtained by tensoring up; it acquires a G-action. Moreover, it is a finitely
generated, projective π∗ (A)-module, and the map
Y
e′ ⊗π (A) A
e′ →
e′ ,
A
A
∗
∗
∗
∗
G
given by all the twisted multiplications a1 ⊗ a2 7→ a1 g(a2 ) for g ∈ G, is an isomorphism. By [BR07,
Theorem 2.1.1], we can construct an object A′ ∈ CAlgA/ with a G-action such that A′ is a G-Galois
e′∗ on homotopy groups.
extension of A and realizes the above map π∗ (A) → A
′
We can now describe the universal property of A as an E∞ -A-algebra. In fact, we claim that
for any E∞ -A-algebra B, we have a natural homotopy equivalence
(13)
HomCAlgA/ (A′ , B) = HomAeven (A′∗ , πeven (B)),
so in particular the left-hand-side is discrete. But by Galois descent, this is also the G-fixed points
of the set of maps
Y
HomCAlgA′ (A′ ⊗A A′ , B ⊗A A′ ) ≃
Idem(B ⊗A A′ )
G
= HomA′∗ (A′∗ ⊗πeven(A) A′∗ , πeven (B) ⊗πeven(A) A′∗ ),
Q
since A′ has homotopy groups which are flat over π∗ (A) and since A′∗ ⊗πeven(A) A′∗ ≃ G A′∗ . But
using the algebraic form of Galois descent, we get that
HomA′∗ (A′∗ ⊗πeven(A) A′∗ , πeven (B) ⊗πeven(A) A′∗ )G = HomAeven (A′∗ , πeven (B)),
so we get (13) as claimed.
This imbeds the Galois objects in the category of graded, finite étale πeven (A)-algebras fully
faithfully (by (13)) in the category of finite covers of the E∞ -ring. To associate a finite cover to
any graded, finite étale πeven (A)-algebra, one now uses Galois descent: the Galois objects can be
used to split any finite étale algebra object. Full faithfulness on these more general covers can now
be checked locally, using descent.
With this in mind, we can state and prove our main result.
Theorem 6.11. Let A be a noetherian, rational E∞ -ring. Then the natural map π1 Mod(A) →
π1et,gr (Specπeven (A)) is an isomorphism of profinite group(oid)s.
Proof. We have already proved this result if A has a unit in degree two, thanks to Theorem 6.4
(see also Example 6.9). We want to claim that for any A satisfying the conditions of this result, the
functor from graded, finite étale πeven (A)-algebras to finite covers of the E∞ -ring A is an equivalence
of categories C1 ≃ C2 . We already know that the functor C1 → C2 is fully faithful.
Both categories depend functorially on A and have a good theory of descent via the base-change
±1
of E∞ -rings Q → Q[t±1
2 ], where Q[t2 ] is the free rational E∞ -ring on an invertible degree two
class. By descent theory, we can thus reduce to the case of a Q[t±1
2 ]-algebra, for which we have
proved the result in Theorem 6.4.
32
AKHIL MATHEW
7. The Picard group
7.1. Generalities. Another natural invariant that one might attempt to study using the theory
of residue fields is the Picard group. Recall that the Picard group of an E∞ -ring A, denoted
Pic(A), is the group of isomorphism classes of ⊗-invertible A-modules. In fact, these techniques
were originally introduced in [HMS94] in the study of the K(n)-local Picard group. It is known
that if A is an E∞ -ring which is even periodic and whose π0 is regular, then the Picard group is
algebraic ([BR05]).
In this paper, we have given global descriptions (in terms of algebra) of both the Galois theory
and the thick subcategories of Mod(A), for A a noetherian rational E∞ -ring. We do not know if it
is possible to describe the Picard group of rational E∞ -rings in a similar global manner. However,
the following example shows that any answer will be necessarily more complicated.
Example 7.1. Consider the E∞ -ring A = Sym∗ Q[−1] ⊗ Q[ǫ]/ǫ2 , which is obtained from the ring
of “dual numbers” by freely adding a generator in degree −1. By the results of Section 3.3, we have
Mod(A) ⊂ LocS 1 (Mod(Q[ǫ]/ǫ2 )),
that is, to give an A-module is equivalent to giving a Q[ǫ]/ǫ2 -module together with an automorphism
whose action on homotopy groups is ind-unipotent.
For example, we might consider the Q[ǫ]/ǫ2 -module Q[ǫ]/ǫ2 and equip it with the automorphism
given by 1 + rǫ, for any r ∈ Q. For any r ∈ Q, this defines an A-module Mr . Since the underlying
Q[ǫ]/ǫ2 -module is invertible, it follows that Mr ∈ Pic(Mod(A)). Moreover, Mr ⊗ Mr′ ≃ Mr+r′ (by
composing automorphisms). This shows that there is a copy of Q inside the Picard group of Mod(A)
that one does not see from the homotopy groups of A. In fact, one sees easily that Pic(A) ≃ Z ⊕ Q.
Nonetheless, we will be able to obtain a weak partial result about the Picard groups of noetherian rational E∞ -rings in comparison to the algebraic analog. We will first need some algebraic
preliminaries.
Definition 7.2. Given a graded-commutative ring R∗ , the category of graded R∗ -modules has a
symmetric monoidal structure (the graded tensor product). We let Pic(R∗ ) denote the Picard group
of the category of graded R∗ -modules, i.e., the group of isomorphism classes of invertible graded
R∗ -modules.
Note that any invertible graded R∗ -module must be flat, since tensoring with it is an autoequivalence (hence exact).
Proposition 7.3. Suppose R∗ is a graded-commutative ring such that R2 contains a unit and
R0 is a noetherian ring, and R1 is a finitely generated R0 -module. Suppose R0 has no nontrivial
idempotents. Then Pic(R∗ ) = Z/2 ⊕ Pic(R0 ), where the Z/2 comes from the shift of R∗ .
Proof. It suffices to show that if R0 is local, then Pic(R∗ ) = Z/2. We prove this first if R0 is a field
k, so there is a map R∗ → k[t±1
2 ] of graded rings with nilpotent kernel. Let M∗ be an invertible
R∗ -module. Then M ⊗R∗ k[t2±1 ] is either k[t2±1 ] or its shift since Pic(k[t±1
2 ]) ≃ Z/2; assume the
former without loss of generality. Choose a homogeneous x ∈ (M∗ ⊗R∗ k[t±1
2 ])0 which is a generator
and lift it to a homogeneous element x ∈ M0 . We then obtain a map R∗ → M∗ which induces an
isomorphism after tensoring with k[t±1
2 ]. By Nakayama’s lemma, one sees that it is surjective. Let
K∗ be the kernel. Then the short exact sequence
0 → K∗ → R∗ → M∗ → 0
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
33
has the property that
±1
±1
0 → K∗ ⊗R∗ k[t±1
2 ] → k[t2 ] → k[t2 ] → 0
is still exact, by flatness of M∗ , and it shows that K∗ = 0 by Nakayama’s lemma.
If R0 is not a field k, then we can consider the maximal ideal m ⊂ R0 and consider the map
R∗ → R∗ /R∗ m to reduce to this case. Using a similar argument with Nakayama’s lemma, and
the fact that the Picard group of R∗ /R∗ m is Z/2, we find that the Picard group of R∗ is Z/2 as
well.
Recall that if A is an E∞ -ring, one has the following basic construction.
Construction 7.4 ([BR05]). There is an inclusion Pic(π∗ (A)) → Pic(A) which sends an invertible graded π∗ (A)-module M∗ to an invertible A-module M (which is uniquely determined) with
π∗ (M ) ≃ M∗ . The image consists of those invertible A-modules M such that π∗ (M ) is a flat
π∗ (A)-module.
Elements in the image of the map Pic(π∗ (A)) → Pic(A) are said to be algebraic; if every element
is algebraic, then the Picard group itself is said to be algebraic. There are many cases in which
the Picard group of an E∞ -ring can be shown to be algebraic. For instance, if A is even periodic
with regular noetherian π0 , or if A is connective, then it is known [BR05] that the Picard group
of A is algebraic (see also [MS14, §2.4.6]). Example 7.1 shows that the Picard group of a rational
noetherian E∞ -ring need not be algebraic. Our main result (Theorem 7.7), however, implies that
any torsion in the Picard group is necessarily algebraic.
To prove this result, we will need to use some techniques from [MS14] which we review briefly here.
Recall ([MS14, Def. 2.2.1]) that the Picard group Pic(A) is the group of connected components of
a connective spectrum pic(A) called the Picard spectrum of A. The infinite loop space Ω∞ pic(A) is
associated to the symmetric monoidal ∞-groupoid of invertible A-modules (under tensor product).
The use of pic(A) (as opposed to simply Pic(A)) is critical when one wishes to appeal to descenttheoretic techniques.
We now outline a basic descent-theoretic technique in the study of Picard groups of ring spectra.
Construction 7.5 (Compare [MS14, §3]). Let A → B be a morphism of E∞ -rings which admits
descent. In this case, we can obtain an expression for the ∞-category Mod(A) as the totalization
→
Mod(A) ≃ Tot Mod(B) →
.
.
.
,
→ Mod(B ⊗A B) →
→
by descent theory [Mat16, §3]. As a result, one obtains an expression for the spectrum pic(A),
(14)
pic(A) = τ≥0 Tot pic(B (⊗•+1) ) ,
and a resulting homotopy spectral sequence
(15)
E2s,t = H s πt pic(B (⊗•+1) ) =⇒ πt−s pic(A),
t − s ≥ 0.
We recall, moreover, that for any E∞ -ring A, we have natural isomorphisms π1 (pic(A)) ≃ (π0 A)×
and πt (pic(A)) ≃ πt−1 (A) for t ≥ 2. In particular, the descent spectral sequence (15), for t ≥ 2, has
the same E2 -page as the usual Adams spectral sequence. If A → B admits descent, it follows from
[MS14, Comparison Tool 5.2.4] together with the analogous result for the A → B Adams spectral
sequence [Mat16, Cor. 4.4] that (15) degenerates (for t − s ≥ 0) after a finite stage with a horizontal
vanishing line.
34
AKHIL MATHEW
Remark 7.6. We refer to [MS14] for several computational examples and applications of this
spectral sequence. In this paper, we will only use the existence of this spectral sequence and its
degeneration at a finite stage with a horizontal vanishing line.
7.2. The main result.
Theorem 7.7. Let A be a rational, noetherian E∞ -ring. Then the cokernel of the map Pic(π∗ (A)) →
Pic(A) is torsion-free.
The main goal of this subsection is to prove Theorem 7.7, which while not entirely satisfying
still provides significant information. In other settings, most of the interesting information in such
Picard groups is precisely the torsion. We will prove this in several steps, following the construction
of residue fields in Theorem 4.14.
Lemma 7.8. Let A be a rational E∞ -ring and let y ∈ π−1 (A). Then the kernel of Pic(A) →
Pic(A//y) is a Q-vector space.
Proof. In fact, A → A//y admits descent (Proposition 3.8), so we use the homotopy spectral
sequence associated to the expression (14). We observe that, for i ≥ 2, the cosimplicial abelian
group πi (pic(B ⊗•+1 )) consists of rational vector spaces. Therefore, to prove the lemma, it suffices
to show that there is no contribution in filtration one. However, we note that π0 ((A//y)⊗k ) ≃
π0 (A//y)[u1 , . . . , uk−1 ]. As a result, the cosimplicial abelian group π0 ((A//y)⊗•+1 )× is actually
constant and has no nontrivial cohomology. This proves the claim.
Lemma 7.9. Let A be a rational E∞ -ring such that:
(1) π2 (A) contains a unit.
(2) π0 (A) is a field k.
(3) π−1 (A) is countably dimensional vector space over k.
Then the torsion subgroup of Pic(A) is {A, ΣA}.
Proof. We will imitate the argument of Proposition 4.11. Consider the sequence of E∞ -rings A ≃
A(0) → A(1) → A(2) → . . . of Proposition 4.9. By Lemma 7.8, the maps
Pic(A(i) )tors → Pic(A(i+1) )tors
are injections.
Let A1 = lim A(i) . It follows that the map A → A1 induces an injection Pic(A)tors → Pic(A1 )tors ,
−→
because the Picard functor commutes with filtered colimits [MS14, Prop. 2.4.1]. Moreover, A1
(0)
(1)
satisfies the same hypotheses, and we can construct a sequence of E∞ -rings A1 ≃ A1 → A1 → . . .
(i)
(i+1)
(i)
from Proposition 4.9. Similarly, each of the maps A1 → A1
induces an injection Pic(A1 )tors →
(i+1)
(i)
Pic(A1
)tors . If we set A2 = lim A1 , we get that the map A1 → A2 induces an injection
−→
Pic(A1 )tors → Pic(A2 )tors . Inductively, we follow the proof of Proposition 4.11 and construct the
sequence
A = A0 → A1 → A2 → . . . ,
such that each Ai satisfies the same hypotheses as A did, and such that Ai → Ai+1 is a colimit of a
sequence constructed in Proposition 4.9, whose colimit is k[t±1
2 ]. Our reasoning shows that we get
a sequence of injections
Pic(A)tors → Pic(A1 )tors → Pic(A2 )tors → · · · → Pic(k[t±1
2 ])tors ≃ Z/2
because, again, the Picard functor commutes with filtered colimits. This completes the proof.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
35
Lemma 7.10. Let A be a rational, noetherian E∞ -ring. Suppose that π2 (A) contains a unit and
that π0 (A) is a local artinian ring with residue field. Let x ∈ π0 (A) belong to the maximal ideal.
Then the map A → A//x induces an injection Pic(A)tors → Pic(A//x)tors .
Proof. The map A → A//x admits descent since x is nilpotent (Proposition 3.7). Therefore, we
can apply the expression (14) and the associated spectral sequence. As in Lemma 7.9, to run
the argument, it suffices to see that there are no torsion contributions in filtration one. All the
rings π0 ((A//x)⊗(n+1) ) are local artinian with the same residue field k, in view of Theorem 2.21.
Given any local artinian ring R with residue field k, the group R× of units has a natural splitting
k × ⊕ R×,unip where R×,unip is a Q-vector space. From this, it follows easily that the contribution
in filtration one in the spectral sequence is a Q-vector space, which proves the lemma.
Lemma 7.11. Let A be a rational, noetherian E∞ -ring. Suppose that π2 (A) contains a unit and
that π0 (A) is a local artinian ring. Then the only nontrivial element in the torsion subgroup of
Pic(A) is ΣA.
Proof. As in the proof of Proposition 4.12, we can construct a sequence of E∞ -rings
A = A0 → A1 → A2 → . . . ,
such that:
(i)
(i)
(i)
(1) Ai+1 ≃ Ai //(x1 , . . . , xni ) for some finite sequence of nilpotent elements xj ∈ π0 (Ai ).
(2) Each ring π0 (Ai ) is local artinian with residue field k.
(3) Each map Ai → Ai+1 induces a map of local artinian rings π0 (Ai ) → π0 (Ai+1 ) which
annihilates the maximal ideal of π0 (Ai ).
By Lemma 7.10, it follows that each of the maps Ai → Ai+1 induces an injection Pic(Ai )tors →
Pic(Ai+1 )tors . The colimit A∞ = limi Ai has Pic(A∞ )tors ≃ Z/2 in view of Lemma 7.9, and
−→
Pic(A)tors ֒→ Pic(A∞ )tors . Putting everything together, we get that Pic(A)tors ≃ Z/2 as desired.
Proof of Theorem 7.7. Let M be an invertible A-module. Suppose that n > 0 and that the tensor
power M ⊗n ∈ Pic(A) has the property that π∗ (M ) is a flat π∗ (A)-module, i.e., M ⊗n belongs to
the image of the map Pic(π∗ (A)) → Pic(A). We need to show that M itself has this property.
For this, we may make the base change A → A[t2±1 ], as π∗ (A) → π∗ (A[t±1
2 ]) is faithfully flat,
and thus assume that π2 (A) contains a unit. Since flatness is a local property, it also suffices to
work under the assumption that π0 (A) is a local ring. Similarly, by completing, we can assume
that π0 (A) is complete local with maximal ideal m = (x1 , . . . , xn ) ⊂ π0 (A). Here we have to use
the fact that any invertible A-module is necessarily perfect [MS14, Prop. 2.1.2]. In this case, the
algebraic Picard group is trivial (Proposition 7.3). So it suffices to show that Pic(A)tors = Z/2.
def
m
Consider the tower of E∞ -A-algebras Am = A//(xm
1 , . . . , xn ). We have Pic(Am )tors ≃ Z/2 by
Lemma 7.11. One sees that Pic(A) ⊂ π0 (limm pic(Am )) because Modω (A) ⊂ limm Modω (Am ). Note
←−
←−
that the tower of abelian groups {π1 (pic(Am ))} = {π0 (Am )× } satisfies the Mittag-Leffler condition
since the tower {π0 (Am )} of finite-dimensional k-vector spaces clearly satisfies the Mittag-Leffler
condition. In particular, π0 (limm pic(Am )) = limm Pic(Am ) by the Milnor exact sequence. Finally,
←−
←−
!
Pic(A)tors ⊂
⊂ lim Pic(Am )tors = lim Z/2 = Z/2.
←−
←−
lim Pic(Am )
←−
m
tors
m
m
36
AKHIL MATHEW
8. Non-noetherian counterexamples
The purpose of this section is to describe certain counterexamples that can arise from nonnoetherian rational E∞ -rings. In particular, we obtain as a result new constructions of Galois
extensions of ring spectra (Theorem 8.15) and of elements in Picard groups (Example 8.17). The
main point of this section is that we can obtain such examples for quasi-affine schemes (such as
punctured spectra) which fail to satisfy a form of “purity.”
8.1. Quasi-affineness. Let X be a noetherian scheme. Recall the presentable, stable ∞-category
QCoh(X) of quasi-coherent sheaves of OX -complexes on X [Lur11a]. QCoh(X) is a symmetric
monoidal ∞-category with unit the structure sheaf OX , whose endomorphisms are given by the
E∞ -ring RΓ(X, OX ) of (derived) global sections of OX . In this generality, we obtain an adjunction
(16)
(· ⊗RΓ(X,OX ) OX , RΓ) : Mod(RΓ(X, OX )) ⇄ QCoh(X),
which sends the RΓ(X, OX )-module RΓ(X, OX ) to the structure sheaf OX . The right adjoint takes
the derived global sections. We need the following basic fact.
Theorem 8.1 ( [Lur11a, Prop. 2.4.4]). Suppose X is quasi-affine. Then the adjunction (16) is a
pair of inverse equivalences of symmetric monoidal ∞-categories.
E∞ -rings of the form RΓ(X, OX ) for quasi-affine schemes will be our primary source of counterexamples, because questions about RΓ(X, OX ) can often be reduced to questions (in ordinary
algebraic geometry) about the scheme X. For instance, we obtain immediately:
Corollary 8.2. Suppose X is a quasi-affine scheme and L is a line bundle on X. Then the
RΓ(X, OX )-module RΓ(X, L) is invertible.
We can also obtain a comparison for Galois theory.
Corollary 8.3. Suppose X is a quasi-affine scheme. Suppose Y → X is a finite étale cover. Then
the map RΓ(X, OX ) → RΓ(Y, OY ) of E∞ -rings is a finite cover. If Y → X is a G-torsor for a finite
group G, then the map (together with the natural G-action on the target) exhibits RΓ(Y, OY ) as a
faithful G-Galois extension of RΓ(X, OX ). In fact, the Galois group of the E∞ -ring RΓ(X, OX ) is
naturally identified with the étale fundamental group of X.
Proof. If f : Y → X is a finite étale cover, then Y can be constructed as the relative Spec of a sheaf
of commutative OX -algebras f∗ (OY ). On any affine open SpecR ⊂ X, f∗ (OY )|SpecR is obtained
from a finite étale algebra. In other words, if we write QCoh(X) = limSpecR⊂X Mod(R), as the
←−
inverse limit ranges over all Zariski open subsets SpecR ⊂ X, then f∗ (OY ) ∈ QCoh(X) defines a
family of finite covers in each of these symmetric monoidal ∞-categories. It follows from [Mat16,
Prop. 7.1] (and [Mat16, Th. 6.5] for the comparison between weak finite covers and finite covers)
that f∗ (OY ) defines a finite cover of the unit in the ∞-category QCoh(X). Thus, applying the
equivalence RΓ then completes the proof.
In general, there is no reason for the Galois theory (resp. Picard group) of the quasi-affine scheme
to be determined by that of π0 RΓ(X, OX ), and this will be a source of counterexamples. This is
related to subtle purity questions. However, we can obtain immediately counterexamples to the
thick subcategory theorem without noetherian hypotheses in view of the following result.
Theorem 8.4 (Thomason [Tho97, Th. 3.15]). Let X be a noetherian scheme. Let QCohω (X)
denote the ∞-category of quasi-coherent complexes F ∈ QCoh(X) such that for every open affine
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
37
SpecR ⊂ X, F (SpecR) is a perfect R-module (equivalently, QCohω (X) consists of the dualizable
objects in QCoh(X)). Then the thick tensor-ideals in QCohω (X) are in natural bijection with the
specialization-closed subsets of X.
Suppose X is quasi-affine. In this case, QCohω (X) corresponds under the equivalence of Theorem 8.1 to the RΓ(X, OX )-modules which are dualizable, i.e., perfect, and thick tensor-ideals are
the same as thick subcategories. Thus, we get:
Corollary 8.5. If X is a quasi-affine scheme, then the thick subcategories of Modω (RΓ(X, OX ))
are in natural bijection with the specialization-closed subsets of X.
Construction 8.6. Let X be a quasi-affine, noetherian scheme which is not affine. In this case,
the natural map X → Specπ0 (RΓ(X, OX )) is an open immersion of schemes [Sta13, Tag 01P5]. By
Corollary 8.5, the thick subcategories of Mod(RΓ(X, OX )) are in bijection not with specializationclosed subsets of the scheme Specπ0 (RΓ(X, OX )), but rather specialization-closed subsets of an
open subset (i.e., X) of this scheme.
8.2. The punctured affine plane. Not every compact E∞ -Q-algebra has the noetherianness
properties used in this paper. In this subsection, we explain how E∞ -rings of the form RΓ(X, OX )
can give counterexamples, and work out the simplest nontrivial case.
Construction 8.7. Consider the E∞ -ring A of functions on the punctured affine plane A2 \{(0, 0)},
which fits into a homotopy pullback
A
Q[x, y ±1 ]
// Q[x±1 , y] .
// Q[x±1 , y ±1 ]
The homotopy groups π∗ (A) are given by
i=0
Q[x, y]
∞
∞
πi (A) = Q[x, y]/(x , y ) i = −1 ,
0
otherwise
where Q[x, y]/(x∞ , y ∞ ) denotes the cokernel of the map Q[x±1 , y] ⊕ Q[x, y ±1 ] → Q[x±1 , y ±1 ]. In
particular, π−1 (A) is not a finitely generated π0 (A)-module (though π0 (A) is noetherian).
Proposition 8.8. The E∞ -ring A is compact in CAlgQ/ .
We are grateful to J. Lurie for explaining this to us.
Proof. In fact, to give a morphism A → B, for B a rational E∞ -ring, is equivalent to giving two
elements u, v ∈ Ω∞ B which have the property that B/(x, y) is contractible. This follows from the
fact that A is the finite localization ([Mil92]) of Q[x, y] away from the Q[x, y]-module Q[x, y]/(x, y)
(which is supported at the origin). In particular, to give a morphism of E∞ -rings A → B is
equivalent to giving a map Q[x, y] → B such that B//(x, y) = B/(x, y) is contractible; note that
this condition is detected in a finite stage of a filtered colimit.
In [BHL15, Ex. 2.8], B. Bhatt and D. Halpern-Leistner give in fact an explicit presentation of A
as an E∞ -ring under Q[x, y]. Consider the Q[x, y]-module M = Q[x, y]/(x, y) and the natural map
φ : Q[x, y] → M . The dual gives a map ψ : DM → Q[x, y], where DM is the Spanier-Whitehead
dual of M . Then one has:
38
AKHIL MATHEW
Proposition 8.9 (Bhatt, Halpern-Leinster). The E∞ -Q[x, y]-algebra is the pushout
Sym∗Q[x,y] (DM )
// Q[x, y] ,
Q[x, y]
// A
where:
(1) Sym∗Q[x,y] (DM ) is the free E∞ -Q[x, y]-algebra on the Q[x, y]-module DM .
(2) The two maps Sym∗Q[x,y] (DM ) → Q[x, y] are adjoint to two maps of Q[x, y]-modules DM →
Q[x, y] which are given by ψ and the zero map.
Proof. Indeed, let A′ ∈ CAlgQ[x,y]/ . Then
HomCAlgQ[x,y]/ (A, A′ ) ≃ ∗ ×HomMod(Q[x,y]) (DM,A′ ) ∗.
ψ
Here the two maps ∗ → HomMod(Q[x,y]) (DM, A′ ) send, respectively, ∗ to 0 and to the map DM →
A → A′ . If A′ //(x, y) = 0, then DM = 0 and the mapping space HomCAlgQ[x,y]/ (A, A′ ) is therefore
contractible. If not, then DM → A → A′ is not the zero map, so that mapping space is empty.
This is precisely the universal property of A′ ∈ CAlgA′ / .
Remark 8.10. The ∞-category of A-modules is equivalent to the ∞-category of quasi-coherent
sheaves on the scheme A2Q \ {(0, 0)}, since this scheme is quasi-affine. In particular, it follows
Corollary 8.5 that the thick subcategories of Modω (A) correspond to the subsets of A2Q \ {(0, 0)}
which are closed under specialization. In particular, Theorem 1.4 fails for A, as there is no thick
subcategory corresponding to the origin in Specπ0 A.
Remark 8.11. A also illustrates the failure of Theorem 2.21 in the non-noetherian case. In fact,
π0 (A)/(x, y) ≃ Q while A//(x, y) is contractible.
8.3. Punctured spectra and counterexamples. We will now describe counterexamples to our
theorems on Galois groups and Picard groups in the non-noetherian case, arising from quasi-affine
schemes in a similar way. We start by recalling the context.
Construction 8.12. Let (R, m) be a noetherian local ring. We define the punctured spectrum
Spec◦ R = SpecR \ {m}.
The punctured spectrum Spec◦ R is a quasi-affine scheme, and many “purity” results in algebraic
geometry and commutative algebra relate invariants of Spec◦ R to those of SpecR.
Theorem 8.13 (Zariski-Nagata [Gro05, Exp. X, Th. 3.4]). Let R be a regular local ring of
dimension ≥ 2. Then the inclusion Spec◦ R → SpecR induces an isomorphism on étale fundamental
groups.
From our point of view, we can restate “purity” results such as the Zariski-Nagata theorem in
terms of ring spectra, by passage to the E∞ -ring RΓ(Spec◦ R, OSpec◦ R ). Let (R, m) be a regular local
ring of dimension ≥ 2. Then π0 (RΓ(Spec◦ R, OSpec◦ R )) = R and π−i RΓ(Spec◦ R, OSpec◦ R )) = 0 if
i∈
/ {0, dim(R) − 1} by general results on local cohomology and depth [Gro05, Exp. III, Ex. 3.4].
For example, we obtain:
(1) Theorem 8.13 is thus equivalent to the statement that the Galois group of the E∞ -ring
RΓ(Spec◦ R, OSpec◦ R ) is algebraic.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
39
(2) Similarly, on a much more elementary level, let (R, m) be a regular local ring of dimension ≥ 2. Then R is factorial, so that it has trivial Picard group. Since the Picard
group is isomorphic to the class group, it follows that the inclusion Spec◦ R → SpecR
induces an isomorphism on Picard groups. In particular, it follows that the Picard group
of RΓ(Spec◦ R, OSpec◦ R ) is algebraic. (More subtle purity results of the Picard group in
non-regular cases can be phrased in this form too.)
The main point of this subsection is that non-regular rings for which purity fails can be used
to give interesting examples of Galois extensions and invertible modules over non-noetherian ring
spectra. Our example (which is not local) follows [Fos73, Example 16.5].
Let K be a field of characteristic zero containing a primitive nth root ζn of unity. Let m ≥ 2.
Consider the Z/n-action on the ring R′ = K[x1 , . . . , xm ] sending xi 7→ ζn xi . Then R = R′Z/n
is the subring generated by all the homogeneous degree n monomials. Geometrically, the map
SpecR′ → SpecR = SpecR′Z/n corresponds to the quotient of the affine space Am by rotation by
the angle 2π/n in each direction. In particular, this map is étale away from the origin, the only
place where the action fails to be free.
Construction 8.14. Let X = SpecR and let Y = SpecR′ . We have a Z/n-action on Y and a
map Y → X which exhibits X as the quotient Y /(Z/n). If y ∈ Y is the point corresponding to the
prime ideal (x1 , . . . , xm ) and x ∈ X its image, then the point y is Z/n-invariant, and the induced
map Y \ {y} → X \ {x} is a Z/n-torsor. We write X ◦ = X \ {x} , Y ◦ = Y \ {y}.
We define E∞ -rings A = RΓ(X ◦ , OX ◦ ) and B = RΓ(Y ◦ , OY ◦ ). Note that B ∈ CAlgA/ has a
natural Z/n-action.
Theorem 8.15. We have π0 (A) = R, π0 (B) = R′ . The map A → B, together with the Z/n-action
on B exhibits B as a faithful Z/n-Galois extension of A.
Proof. The natural map R′ = Γ(Y, OY ) → Γ(Y ◦ , OY ◦ ) is an isomorphism since Y is normal and the
missing locus is codimension ≥ 2 by [Mat80, §17, Th. 35]. Moreover, the Z/n-torsor Y ◦ → X ◦ shows
Z/n
=
that RΓ(X ◦ , OX ◦ ) ≃ RΓ(Y ◦ , OY ◦ )hZ/n . Taking π0 , we find that π0 (A) ≃ (π0 (Γ(Y ◦ , OY ◦ )))
′Z/n
R
= R. The assertion that A → B is a faithful Z/n-Galois extension comes from Corollary 8.3.
The map of commutative rings R → R′ is not étale: in fact, R′ is not regular (at zero) while R
is. In particular, the Galois extension of Theorem 8.15 does not come from algebra.
Example 8.16. Suppose K = C is the field of complex numbers. In this case, the topological
realization of SpecR (i.e., the topological space Cm /(Z/n)) is easily seen to be contractible: one
can scale down to the image of the origin. Therefore, the étale fundamental group of R = π0 (A) is
trivial. However, the Z/n-Galois extension A → B (and the fact that B has trivial Galois group
by Corollary 8.3 applied to Cm \ {(0, . . . , 0)}) shows that the Galois group of the E∞ -ring A is
precisely Z/n.
We can also obtain elements of the Picard group.
Example 8.17. We compute the (classical) Picard group of the scheme X ◦ again with K = C.
First, we observe that the Picard group of Y ◦ is trivial since that of affine space Y is and Y ◦ =
Y \ {y}. Moreover, Γ(Y ◦ , OY◦ ) = C× . We have a Z/n-torsor Y ◦ → X ◦ , and we can use Galois
descent to compute the Picard group as Pic(X ◦ ) = H 1 (Z/n; H 0 (Y ◦ , OY×◦ )) = Z/n since the action
is trivial. Therefore, the Picard group of the E∞ -ring A is given by Z⊕Z/n where the Z comes from
40
AKHIL MATHEW
suspensions. However, we claim that the Picard group of π0 (A) is trivial. Indeed, by Lemma 8.18
below, we have Pic(X) ⊂ Pic(X ◦ ) = Z/n. But the Picard group of X can have no torsion as
X is topologically contractible, in view of the Kümmer sequence, and therefore Pic(X) = 0. In
particular, by Corollary 8.2, we find that Pic(A) ≃ Z ⊕ Z/n though the Picard group of π0 (A) is
trivial.
Lemma 8.18. Let X be a noetherian, normal, integral scheme. Let Z ⊂ X have codimension ≥ 2.
Then the map Pic(X) → Pic(X \ Z) is injective.
×
×
Proof. Let j : X\Z → X be the open imbedding. Then the map OX
→ j∗ (OX\Z
) is an isomorphism.
×
×
The Leray spectral sequence now shows that the natural map H 1 (X, OX ) → H 1 (X \ Z, OX\Z
) is
an injection.
References
[ABG+ 14]
Matthew Ando, Andrew J. Blumberg, David Gepner, Michael J. Hopkins, and Charles Rezk. An ∞categorical approach to R-line bundles, R-module Thom spectra, and twisted R-homology. J. Topol.,
7(3):869–893, 2014.
[Bal05]
Paul Balmer. The spectrum of prime ideals in tensor triangulated categories. J. Reine Angew. Math.,
588:149–168, 2005.
[Bal10]
Paul Balmer. Spectra, spectra, spectra—tensor triangular spectra versus Zariski spectra of endomorphism
rings. Algebr. Geom. Topol., 10(3):1521–1563, 2010.
[BHL15]
Bhargav Bhatt and Daniel Halpern-Leistner. Tannaka duality revisited. 2015. Available at
http://arxiv.org/abs/1507.01925.
[BMMS86] R. R. Bruner, J. P. May, J. E. McClure, and M. Steinberger. H∞ ring spectra and their applications,
volume 1176 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1986.
[BR05]
Andrew Baker and Birgit Richter. Invertible modules for commutative S-algebras with residue fields.
Manuscripta Math., 118(1):99–119, 2005.
[BR07]
Andrew Baker and Birgit Richter. Realizability of algebraic Galois extensions by strictly commutative
ring spectra. Trans. Amer. Math. Soc., 359(2):827–857 (electronic), 2007.
[BR08]
Andrew Baker and Birgit Richter. Galois extensions of Lubin-Tate spectra. Homology, Homotopy Appl.,
10(3):27–43, 2008.
[CLM76]
Frederick R. Cohen, Thomas J. Lada, and J. Peter May. The homology of iterated loop spaces. Lecture
Notes in Mathematics, Vol. 533. Springer-Verlag, Berlin-New York, 1976.
[DHS88]
Ethan S. Devinatz, Michael J. Hopkins, and Jeffrey H. Smith. Nilpotence and stable homotopy theory.
I. Ann. of Math. (2), 128(2):207–241, 1988.
[Eis95]
David Eisenbud. Commutative algebra: with a view towards algebraic geometry, volume 150 of Graduate
Texts in Mathematics. Springer-Verlag, New York, 1995.
[EKMM97] A. D. Elmendorf, I. Kriz, M. A. Mandell, and J. P. May. Rings, modules, and algebras in stable homotopy theory, volume 47 of Mathematical Surveys and Monographs. American Mathematical Society,
Providence, RI, 1997. With an appendix by M. Cole.
[Fos73]
Robert M. Fossum. The divisor class group of a Krull domain. Springer-Verlag, New York-Heidelberg,
1973. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 74.
[Gro05]
Alexander Grothendieck. Cohomologie locale des faisceaux cohérents et théorèmes de Lefschetz locaux
et globaux (SGA 2). Documents Mathématiques (Paris) [Mathematical Documents (Paris)], 4. Société
Mathématique de France, Paris, 2005. Séminaire de Géométrie Algébrique du Bois Marie, 1962, Augmenté d’un exposé de Michèle Raynaud. [With an exposé by Michèle Raynaud], With a preface and
edited by Yves Laszlo, Revised reprint of the 1968 French original.
[HMS94]
Michael J. Hopkins, Mark Mahowald, and Hal Sadofsky. Constructions of elements in Picard groups. In
Topology and representation theory (Evanston, IL, 1992), volume 158 of Contemp. Math., pages 89–126.
Amer. Math. Soc., Providence, RI, 1994.
[HPS97]
Mark Hovey, John H. Palmieri, and Neil P. Strickland. Axiomatic stable homotopy theory. Mem. Amer.
Math. Soc., 128(610):x+114, 1997.
[HS98]
Michael J. Hopkins and Jeffrey H. Smith. Nilpotence and stable homotopy theory. II. Ann. of Math. (2),
148(1):1–49, 1998.
RESIDUE FIELDS FOR A CLASS OF RATIONAL E∞ -RINGS AND APPLICATIONS
[HS99]
[KT08]
[Lan02]
[LMB00]
[Lur11a]
[Lur11b]
[Lur14]
[Mat80]
[Mat15]
[Mat16]
[Mil92]
[MNN15]
[MS14]
[Oka79]
[Qui69]
[Rog08]
[Sta13]
[Str99]
[Szy14]
[Tho97]
41
Mark Hovey and Neil P. Strickland. Morava K-theories and localisation. Mem. Amer. Math. Soc.,
139(666):viii+100, 1999.
Piotr A. Krylov and Askar A. Tuganbaev. Modules over discrete valuation domains, volume 43 of de
Gruyter Expositions in Mathematics. Walter de Gruyter GmbH & Co. KG, Berlin, 2008.
Serge Lang. Algebra, volume 211 of Graduate Texts in Mathematics. Springer-Verlag, New York, third
edition, 2002.
Gérard Laumon and Laurent Moret-Bailly. Champs algébriques, volume 39 of Ergebnisse der Mathematik
und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics
and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Springer-Verlag, Berlin,
2000.
Jacob Lurie. DAG VIII: Quasi-coherent sheaves and Tannaka duality theorems. 2011. Available at
http://math.harvard.edu/~ lurie.
Jacob Lurie. DAG XII: Proper morphisms, completions, and the Grothendieck existence theorem. 2011.
Available at http://math.harvard.edu/~ lurie.
Jacob Lurie. Higher algebra. 2014. Available at http://math.harvard.edu/~ lurie/higheralgebra.pdf.
Hideyuki Matsumura. Commutative algebra, volume 56 of Mathematics Lecture Note Series. Benjamin/Cummings Publishing Co., Inc., Reading, Mass., second edition, 1980.
Akhil Mathew. A thick subcategory theorem for modules over certain ring spectra. Geom. Topol.,
19(4):2359–2392, 2015.
Akhil Mathew. The Galois group of a stable homotopy theory. Adv. Math., 291:403–541, 2016.
Haynes Miller. Finite localizations. Bol. Soc. Mat. Mexicana (2), 37(1-2):383–389, 1992. Papers in honor
of José Adem (Spanish).
Akhil Mathew, Niko Naumann, and Justin Noel. On a nilpotence conjecture of J. P. May. J. Topol.,
8(4):917–932, 2015.
Akhil Mathew and Vesna Stojanoska. The Picard group of topological modular forms via descent theory.
2014. Available at http://arxiv.org/abs/1409.7702.
Shichirô Oka. Ring spectra with few cells. Japan. J. Math. (N.S.), 5(1):81–100, 1979.
Daniel Quillen. Rational homotopy theory. Ann. of Math. (2), 90:205–295, 1969.
John Rognes. Galois extensions of structured ring spectra. Stably dualizable groups. Mem. Amer. Math.
Soc., 192(898):viii+137, 2008.
The Stacks Project Authors. Stacks Project. http://stacks.math.columbia.edu , 2013.
N. P. Strickland. Products on MU-modules. Trans. Amer. Math. Soc., 351(7):2569–2606, 1999.
Markus Szymik. Commutative S-algebras of prime characteristics and applications to unoriented bordism.
2014. Available at http://arxiv.org/pdf/1211.3239.pdf.
R. W. Thomason. The classification of triangulated subcategories. Compositio Math., 105(1):1–27, 1997.
Harvard University, Cambridge, MA 02138
E-mail address: [email protected]
URL: http://math.harvard.edu/~amathew
| 0math.AC
|
PAGERANK BEYOND THE WEB
DAVID F. GLEICH
Abstract. Google’s PageRank method was developed to evaluate the importance of web-pages
via their link structure. The mathematics of PageRank, however, are entirely general and apply to
any graph or network in any domain. Thus, PageRank is now regularly used in bibliometrics, social
and information network analysis, and for link prediction and recommendation. It’s even used for
systems analysis of road networks, as well as biology, chemistry, neuroscience, and physics. We’ll see
the mathematics and ideas that unite these diverse applications.
arXiv:1407.5107v1 [cs.SI] 18 Jul 2014
Key words. PageRank, Markov chain
1. Google’s PageRank. Google created PageRank to address a problem they
encountered with their search engine for the world wide web [Brin and Page, 1998;
Page et al., 1999]. Given a search query from a user, they could immediately find an
immense set of web pages that contained virtually the exact same words as the user
entered. Yet, they wanted to incorporate a measure of a page’s importance into these
results to distinguish highly recognizable and relevant pages from those that were less
well known. To do this, Google designed a system of scores called PageRank that used
the link structure of the web to determine which pages are important. While there are
many derivations of the PageRank equations [Langville and Meyer, 2006; Pan et al.,
2004; Higham, 2005], we will derive it based on a hypothetical random web surfer.
Upon visiting a page on the web, our random surfer tosses a coin. If it comes up heads,
the surfer randomly clicks a link on the current page and transitions to the new page.
If it comes up tails, the surfer teleports to a – possibly random – page independent of
the current page’s identity. Pages where the random surfer is more likely to appear
based on the web’s structure are more important in a PageRank sense.
More generally, we can consider random surfer models on a graph with an arbitrary
set of nodes, instead of pages, and transition probabilities, instead of randomly clicked
links. The teleporting step is designed to model an external influence on the importance
of each node and can be far more nuanced than a simple random choice. Teleporting
is the essential distinguishing feature of the PageRank random walk that had not
appeared in the literature before [Vigna, 2009]. It ensures that the resulting importance
scores always exist and are unique. It also makes the PageRank importance scores
easy to compute.
These features: simplicity, generality, guaranteed existence, uniqueness, and fast
computation are the reasons that PageRank is used in applications far beyond its
origins in Google’s web-search. (Although, the success that Google achieved no
doubt contributed to additional interest in PageRank!) In biology, for instance,
new microarray experiments churn out thousands of genes relevant to a particular
experimental condition. Models such as GeneRank [Morrison et al., 2005] deploy the
exact same motivation as Google, and almost identical mathematics in order to assist
biologists in finding and ordering genes related to a microarray experiment or related
to a disease. Throughout our review, we will see applications of PageRank to biology,
chemistry, ecology, neuroscience, physics, sports, and computer systems.
Two uses underlie the majority of PageRank applications. In the first, PageRank
is used as a network centrality measure [Koschützki et al., 2005]. A network centrality
score yields the importance of each node in light of the entire graph structure. And
the goal is to use PageRank to help understand the graph better by focusing on what
1
2
D. F. Gleich
PageRank reveals as important. It is often compared or contrasted with a host of
other centrality or graph theoretic measures. These applications tend to use global,
near-uniform teleportation behaviors.
In the second type of use, PageRank is used to illuminate a region of a large
graph around a target set of interest; for this reason, we call the second use a localized
measure. It is also called personalized PageRank based on PageRank’s origins in the
web. Consider a random surfer in a large graph that periodically teleports back to
a single start node. If the teleportation is sufficiently frequent, the surfer will never
move far from the start node, but the frequency with which the surfer visits nodes
before teleporting reveals interesting properties of this localized region of the network.
Because of this power, teleportation behaviors are much more varied for these localized
applications.
2. The mathematics of PageRank. There are many slight variations on the
PageRank problem, yet there is a core definition that applies to the almost all of them.
It arises from a generalization of the random surfer idea. Pages where the random
surfer is likely to appear have large values in the stationary distribution of a Markov
chain that, with probability α, randomly transitions according to the link structure of
the web, and with probability 1 − α teleports according to a teleportation distribution
vector v, where v is usually a uniform distribution over all pages. In the generalization,
we replace the notion of “transitioning according to the link structure of the web”
with “transitioning according to a stochastic matrix P.” This simple change divorces
the mathematics of PageRank from the web and forms the basis for the applications
we discuss. Thus, it abstracts the random surfer model from the introduction in a
relatively seamless way. Furthermore, the vector v is a critical modeling tool that
distinguishes between the two typical uses of PageRank. For centrality uses, v will
resemble a uniform distribution over all possibilities; for localized uses, v will focus
the attention of the random surfer on a region of the graph.
Before stating the definition formally, let us fix some notation. Matrices and
vectors are written in bold, Roman letters (A, x), scalars are Greek or indexed, unbold
Roman (α, Ai,j ). The vector e is the column vector of all ones, and all vectors are
column vectors.
Let Pi,j be the probability of transitioning from page j to page i. (Or more
generally, from “thing j” to “thing i”.) The stationary distribution of the PageRank
Markov chain is called the PageRank vector x. It is the solution of the eigenvalue
problem:
(αP + (1 − α)veT )x = x.
(2.1)
Many take this eigensystem as the definition of PageRank [Langville and Meyer, 2006].
We prefer the following definition instead:
Definition 2.1 (The PageRank Problem). Let P be a column-stochastic matrix
where all entries are non-negative and the sum of entries in each column is 1. Let
v be a column stochastic vector (eT v = 1), and let 0 < α < 1 be the teleportation
parameter. Then the PageRank problem is to find the solution of the linear system
(I − αP)x = (1 − α)v,
(2.2)
where the solution x is called the PageRank vector.
The eigenvector and linear system formulations are equivalent if we seek an
eigenvector x of (2.1) with x ≥ 0 and eT x = 1, in which case:
x = αPx + (1 − α)veT x = αPx + (1 − α)v
⇔
(I − αP)x = (1 − α)v.
PAGERANK BEYOND THE WEB
3
We prefer the linear system because of the following reasons. In the linear system
setup, the existence and uniqueness of the solution is immediate: the matrix I − αP is
a diagonally dominant M-matrix. The solution x is non-negative for the same reason.
Also, there is only one possible normalization of the solution: x ≥ 0 and eT x = 1.
Anecdotally, we note that, among the strategies to solve PageRank problems, those
based on the linear system setup are both more straightforward and more effective
than those based on the eigensystem approach. And in closing, Page et al. [1999]
describe an iteration more akin to a linear system than an eigenvector.
Computing the PageRank vector x is simple. The humble iteration
x(k+1) = αPx(k) + (1 − α)v
where
x(0) = v or x(0) = 0
is equivalent both to the power method on (2.1) and the Richardson method on (2.2),
and more importantly, it has excellent convergence properties when α is not too close
to 1. To see this fact, note that the true solution x = αPx + (1 − α)v and consider
the error after a single iteration:
x − x(k+1) = [αPx + (1 − α)v] − [αPx(k) + (1 − α)v] = αP(x − x(k) ).
|
{z
} |
{z
}
the true solution x
the updated iterate x(k+1)
Thus, the following theorem characterizes the error after k iterations from two different
starting conditions:
Theorem 2.2. Let α, P, v be the data for a PageRank problem to compute
a PageRank vector x. Then the error after k iterations of the update x(k+1) =
αPx(k) + (1 − α)v is:
1. if x(0) = v, then kx − x(k) k1 ≤ kx − vk1 αk ≤ 2αk ; or
2. if x(0) = 0, then the error vector x − x(k) ≥ 0 for all k and kx − x(k) k1 =
eT (x − x(k) ) = αk .
Common values of α range between 0.1 and 0.99; hence, in the worst case,
this method needs at most 3656 iterations to converge to a global 1-norm error of
2−52 ≈ 10−16 (because α3656 ≤ 2−53 to account for the possible factor of 2 if starting
from x(0) = v). For the majority of applications we will see, the matrix P is sparse
with fewer than 10, 000, 000 non-zeros; and thus, these solutions can be computed
efficiently on a modern laptop computer.
Aside 2.3. Although this theorem seems to suggest that x(0) = 0 is a superior
choice, practical experience suggests that starting with x(0) = v results in a faster
method. This may be confirmed by using a computable bound on the error based on
the residual. Let r(k) = (1 − α)v − (I − αP)x(k) = x(k+1) − x(k) be the residual after
1
k iterations. We can use kx − x(k) k1 = k(I − αP)−1 r(k) k1 ≤ 1−α
kr(k) k1 in order to
check for early convergence.
This setup for PageRank, where the choice of P, v, and α vary by application,
applies broadly as the subsequent sections show. However, in many descriptions,
authors are not always careful to describe their contributions in terms of a column
stochastic matrix P and distribution vector v. Rather, they use the following pseudoPageRank system instead:
Definition 2.4 (The pseudo-PageRank problem). Let P̄ be a column substochastic matrix where P̄i,j ≥ 0 and eT P̄ ≤ eT element-wise. Let f be a non-negative
vector, and let 0 < α < 1 be a teleportation parameter. Then the pseudo-PageRank
problem is to find the solution of the linear system
(I − αP̄)y = f
(2.3)
4
D. F. Gleich
where the solution y is called the pseudo-PageRank vector.
Again, the pseudo-PageRank vector always exists and is unique because I − αP̄ is
also a diagonally dominant M-matrix. Boldi et al. [2007] was the first to formalize
this definition and distinction between PageRank and pseudo-PageRank, although
they used the term PseudoRank and the normalization (I − αP̄)y = (1 − α)f ; some
advantages of this alternative form are discussed in Section 5.2. The two problems
are equivalent in the following formal sense (which has an intuitive understanding
explained in Section 3.1, Strongly Preferential PageRank):
Theorem 2.5. Let y be the solution of a pseudo-PageRank system with α, P̄
and f . Let v = f /(eT f ). Then if y is renormalized to sum to 1, that is x = y/(eT y),
then x is the solution of a PageRank system with α, P = P̄ + vcT , and v, where
cT = eT − eT P̄ ≥ 0 is a correction vector to make P̄ stochastic.
Proof. First note that α, P, and v is a valid PageRank problem. This is because f
is non-negative and thus v is column stochastic by definition, and also P is column
stochastic because c ≥ 0 (hence P ≥ 0) and eT P = eT P̄ + cT = eT . Next, note that
the solution of the PageRank problem for x satisfies:
x = αP̄x + αvcT x + (1 − α)v = αP̄x + γf
where
γ=
αcT x + (1 − α)
.
eT f
Hence (I − αP̄)x = γf and so x = γy. But, we know that eT x = 1 because x is a
solution of a PageRank problem, and the theorem follows.
The importance of this theorem is it shows that underlying any pseudo-PageRank
system is a true PageRank system in the sense of Definition 2.1. The difference is
entirely in terms of the normalization of the solution – which was demonstrated by Del
Corso et al. [2004]; Berkhin [2005]; Del Corso et al. [2005]. The result of Theorem 2.2
also applies to solving the pseudo-PageRank system, albeit with the following revisions:
Theorem 2.6. Let α, P̄, f be the data for a pseudo-PageRank problem to compute
a pseudo-PageRank vector y. Then the error after k iterations of the update y(k+1) =
αP̄y(k) + f is:
T
1
f k
1. if y(0) = 1−α
f , then ky − y(k) k1 ≤ ky − f k1 αk ≤ 2e
1−α α ; or
(0)
(k)
2. if y = 0, then the error vector y − y ≥ 0 for all k and ky − y(k) k1 =
eT (y − y(k) ) ≤ αk .
Aside 2.7. The error progression proceeds at the same rate for both PageRank
and pseudo-PageRank. This can be improved for pseudo-PageRank if the vector
cT = eT − eT P̄ > 0 (element-wise). In such cases, then we can derive an equivalent
system with a smaller value of α and a suitably rescaled matrix P̄.
These formal results represent the mathematical foundations of all of the PageRank
systems that arise in the literature (with a few technical exceptions that we will study
in Section 5). The results depend only on the construction of a stochastic matrix or
sub-stochastic matrix, a teleportation distribution, and a parameter α. Thus, they
apply generally and have no intrinsic relationship back to the original motivation
of PageRank for the web. Each type of PageRank problem has a unique solution
that always exists, and the two convergence theorems justify that simple algorithms
for PageRank converge to the unique solutions quickly. These are two of the most
attractive features of PageRank.
One final set of mathematical results is important to understand the behavior of
localized PageRank; however, the precise statement of these results requires a lengthy
and complicated diversion into graph partitioning, graph cuts, and spectral graph
PAGERANK BEYOND THE WEB
5
Fig. 2.1. An illustration of the empirical properties of localized PageRank vectors with teleportation to a single node in an isolated region. In the graph at left, the teleportation vector is the single
circled node. The PageRank vector is shown as the node color in the right figure. PageRank values
remain high within this region and are nearly zero in the rest of the graph. Theory from Andersen
et al. [2006] explains when this property occurs.
theory. Instead, we’ll state this a bit informally. Suppose that we solve a localized
PageRank problem in a large graph, but the nodes we select for teleportation lie in
a region that is somehow isolated, yet connected to the rest of the graph. Then the
final PageRank vector is large only in this isolated region and has small values on the
remainder of the graph. This behavior is exactly what most uses of localized PageRank
want: they want to find out what is nearby the selected nodes and far from the rest
of the graph. Proving this result involves spectral graph theory, Cheeger inequalities,
and localized random walks – see Andersen et al. [2006] for more detail. Instead, we
illustrate this theory with Figure 2.1.
Next, we will see some of the common constructions of the matrices P and P̄ that
arise when computing PageRank on a graph. These justify that PageRank is also a
simple construction.
3. PageRank constructions. When a PageRank method is used within an
application, there are two common motivations. In the centrality case, the input is
a graph representing relationships or flows between a set of things – they may be
documents, people, genes, proteins, roads, or pieces of software – and the goal is to
determine the expected importance of each piece in light of the full set of relationships
and the teleporting behavior. This motivation was Google’s original goal in crafting
PageRank. In the localized case, the input is also the same type of graph, but the
goal is to determine the importance relative to a small subset of the objects. In
either case, we need to build a stochastic or sub-stochastic matrix from a graph. In
this section, we review some of the common constructions that produce a PageRank
or pseudo-PageRank system. For a visual overview of some of the possibilities, see
Figures 3.1 and 3.2.
Notation for graphs and matrices. Let A be the adjacency matrix for a graph
where we assume that the vertex set is V = {1, . . . , n}. The graph could be directed,
in which case A is non-symmetric, or undirected, in which case A is symmetric. The
graph could also be weighted, in which case Ai,j gives the positive weight of edge (i, j).
Edges with zero weight are assumed to be irrelevant and equivalent to edges that are
not present. For such a graph, let d be the vector of node out-degrees, or equivalently,
the vector of row-sums: d = Ae. The matrix D is simply the diagonal matrix with d
on the diagonal. Weighted graphs are extremely common in applications when the
6
D. F. Gleich
Applications &
Constructions
Strongly preferential
PageRank
PseudoPageRank
Random Substochastic
walk
matrix
Weakly preferential
PageRank
PageRank
Weighted
Sink preferential
PageRank
Reverse
Directed or
undirected
graph
Theory &
Algorithms
Dirichlet
Eigensystems
Linear systems
Other transformations
Fig. 3.1. An overview of PageRank constructions and how they relate. The vast majority of
PageRank applications fall somewhere on the red path.
weights reflect a measure of the strength of the relationships between two nodes.
3.1. The standard random walk. In the standard construction of PageRank,
the matrix P represents a uniform random walk operation on the graph A. When
the graph is weighted, the simple generalization is to model a non-uniform walk that
chooses subsequent nodes with probability proportional to the connecting edge’s weight.
The elements of P̄ are rather similar between the two cases:
Ai,j
Ai,j
probability of taking the transition
=
P̄j,i = P
=
from i to j via a random walk step.
d
A
i
i,k
k
Notice two features of this construction. First, we transpose between j, i and i, j.
This is because Ai,j indicates an edge from node i to node j, whereas the probability
transition matrix element i, j indicates that node i can be reached via node j. Second,
we have written P̄ and P̄j,i here because there may be nodes of the graph with no
outlinks. These nodes are called dangling nodes. Dangling nodes complicate the
construction of stochastic matrices P in a few ways because we must specify a behavior
for the random walk at these nodes in order to fully specify the stochastic matrix.
As a matrix formula, the standard random walk construction is:
P̄ = AT D+ .
Here, we have used the pseudo-inverse of the degree matrix to “invert” the diagonal
matrix in light of the dangling nodes with 0 out-degrees. Let cT be the sub-stochastic
correction vector. For the standard random walk construction, cT is just an indicator
vector for the dangling nodes:
(
X
1 node i is dangling
ci = 1 −
P̄k,i =
0 otherwise.
k
We shall now see a few ideas that turn these sub-stochastic matrices into fully
stochastic PageRank problems.
Strongly Preferential PageRank. Given a directed graph with dangling nodes, the
standard random walk construction produces the sub-stochastic matrix P̄ described
above. If we had just used this matrix to solve a pseudo-PageRank problem with
PAGERANK BEYOND THE WEB
7
a stochastic teleportation vector f = (1 − α)v, then, by Theorem 2.5, the result is
equivalent up to normalization to computing PageRank on the matrix:
P = P̄ + cvT .
This construction models a random walk that transitions according to the distribution
v when visiting a dangling node. This behavior reinforces the effect of the teleportation
vector v, or preference vector as it is sometimes called. Because of this reinforcement,
Boldi et al. [2007] called the construction P = P̄+cvT a strongly preferential PageRank
problem. Again, many authors are not careful to explicitly choose a correction to
turn the sub-stochastic matrix into a stochastic matrix. Their lack of choice, then,
implicitly chooses the strongly preferential PageRank system.
Weakly Preferential PageRank & Sink Preferential PageRank. Boldi et al. [2007]
also proposed the weakly preferential PageRank system. In this case, the behavior
of the random walk at dangling nodes is adjusted independently of the choice of
teleportation vector. For instance, Langville and Meyer [2004] advocates transitioning
uniformly from dangling nodes. In such a case, let u = e/n be the uniform distribution
vector, then a weakly preferential PageRank system is:
P = P̄ + cuT .
We note that another choice of behavior is for the random walk to remain at dangling
nodes until it moves away via a teleportation step:
P = P̄ + diag(c).
We call this final method sink preferential PageRank. These systems are less common.
These choices should be used when the matrix P models some type of information or
material flow that must be decoupled from the teleporting behavior.
3.2. Reverse PageRank. In reverse PageRank, we compute PageRank on the
transposed graph AT . This corresponds to reversing the direction of each edge (i, j) to
be an edge (j, i). Reverse PageRank is often used to determine why a particular node
is important rather than which nodes are important [Fogaras, 2003; Gyöngyi et al.,
2004; Bar-Yossef and Mashiach, 2008]. Intuitively speaking, in reverse PageRank, we
model a random surfer that follows in-links instead of out-links. Thus, large reverse
PageRank values suggest nodes that can reach many nodes in the graph. When these
are localized, they then provide evidence for why a node has large PageRank.
3.3. Dirichlet PageRank. Consider a PageRank problem where we wish to fix
the importance score of a subset of nodes [Chung et al., 2011]. Let S be a subset of
nodes such that i ∈ S implies than vi = 0. A Dirichlet PageRank problem seeks a
solution of PageRank where each node i in S is fixed to a boundary value bi . Formally,
the goal is to find x:
(I − αP)x = (1 − α)v
where
xi = bi for i ∈ S.
These problems reduce to solving a pseudo-PageRank system. Consider a block
partitioning of P based on the set S and the complement set of vertices S̄:
PS,S PS,S̄
P=
.
PS̄,S PS̄,S̄
8
D. F. Gleich
Then the Dirichlet PageRank problem is
I
0
b
0
= (1 − α)
.
−αPS̄,S I − αPS̄,S̄ xS̄
vS̄
This system is equivalent to a pseudo-PageRank problem with P̄ = PS̄,S̄ and f =
(1 − α)vS̄ + αPS̄,S b.
3.4. Weighted PageRank. In the standard random walk construction for PageRank on an unweighted graph, the probability of transitioning from node i to any of
it’s neighbors j is the same: 1/di . Weighted PageRank [Xing and Ghorbani, 2004;
Jiang, 2009] alters this assumption such that the walk preferentially visits high-degree
nodes. Thus, the probability of transitioning from node i to node j depends on the
degree of j relative to the total sum of degrees of all i’s neighbors. In our notation, if
the input is adjacency matrix A with degree matrix D, then the sub-stochastic matrix
P̄ is given by the non-uniform random walk construction on the weighted graph with
adjacency matrix W = AD, that is, P̄ = DAT diag(ADe)+ . More generally, let DW
be a non-negative weighting matrix. It could be derived from the graph itself based
on the out-degree, in-degree, or total-degree (the sum of in- and out-degree), or from
some external source. Then P̄ = DW AT diag(ADW e)−1 . Let us note that weighted
PageRank uses a specific choice of weights for the prior importance of each node; the
setting here already adapts seamlessly to edge-weighted graphs.
3.5. PageRank on an undirected graph. One final construction is to use
PageRank on an undirected graph. Those familiar with Markov chain theory often find
this idea puzzling at first. A uniform random walk on a connected, undirected graph
has a well-known, unique stationary distribution [Stewart, 1994, is a good numerical
treatment of such issues]:
T −1
A
D } x = x is solved by x = De/(eT d).
| {z
P
This works because both the row and column sums of A and AT are identical, and the
resulting construction is a reversible Markov chain [Aldous and Fill, 2002, is a good
reference on this topic]. If α < 1, then the PageRank Markov chain is not a reversible
Markov chain even on an undirected graph, and hence, has no simple stationary
distribution. PageRank vectors of undirected graphs, when combined with carefully
constructed teleportation vectors v, yield important information about the presence
of small isolated regions in the graph [Andersen et al., 2006; Gleich and Mahoney,
2014]; formally these results involve graph cuts and small conductance sets. These
vectors are most useful when the teleportation vector is far away from the uniform
distribution, such as the case in Figure 2.1 where the graph is undirected.
Aside 3.1. Of course, if the teleportation distribution v = De/(eT d), then the
resulting chain is reversible. The PageRank vector is then equal to v itself. There
are also specialized PageRank-style constructions that preserve reversibility with more
interesting stationary distributions [Avrachenkov et al., 2010].
4. PageRank applications. When PageRank is used within applications, it
tends to acquire a new name. We will see:
9
PAGERANK BEYOND THE WEB
3
2
0
1
0
A=
0
0
0
5
4
6
1
A directed graph
P̄ =
0
0
0
0
1
0
P̄ =
0 1/2 0 0 0
0 0 0 1/3 0
1/3 1/2 0 1/3 0
1/3 0 0 0 0
1/3 0 1 1/3 0
0
0 0 0 1
P=
0
0
1
1
0
1
0
0
0
0
1
0
0
1
2
0
1
0
d= c=
3
0
1
0
1
0
Weakly preferential
0
0
0
0
1
0
P=
P = P̄ + vcT
Reverse
0
1
0
0
0
0
0
0
0
0
0
0
Strongly preferential
0 1/2 0 0 0
0 0 0 1/3 0
0 1/2 0 1/3 0
0 0 0 0 0
0 0 1 1/3 0
0 0 0 0 1
P̄ = AT D+
0
1
0
1
0
0
The adjacency matrix, degree vector, and correction vector
Random walk
0
0
0
1
0
0
0
1/2
0
0
0
P̄ =
0
0
0
0
1
0
Weighted
0 1/3 0
0 1/3 0
0 0 0
1 1/3 0
0 0 1
0
0
0
1
0
P̄ =
S = {2, 3, 4, 5, 6}
P̄ = A diag(AT e)+
0
0
0
0
1
0
P = P̄ + ucT
u 6= v
Dirichlet
0 0 0 0 0
0 1/2 0 0 0
0 0 0 1/3 0
1 1/2 0 1/3 0
0 0 0 0 1
0 0 0 1/3 0
1/6 1/2 0 0 0
1/6 0 0 1/3 0
1/6 1/2 0 1/3 0
1/6 0 0 0 0
1/6 0 1 1/3 0
1/6 0 0 0 1
0 1/4 0
0
0
0 0 0 3/10 0
0 3/4 0 3/10 0
0 0 0
0
0
0 0 1 4/10 0
0 0 0
0
1
P̄ = (DW AT ) diag(ADW e)+
DW is a diagonal weighting
matrix, e.g. total degree here
P̄ = P̄S̄,S̄
S⊂V
Fig. 3.2. A directed graph and some of the different PageRank constructions on that graph.
For the stochastic constructions, we have vT = [ 0 0 13 31 13 0 ] and u = e/n. Note that node 4 is
dangling in the reverse PageRank construction. For the weighted construction, the total degrees are
[ 1 3 3 3 4 2 ].
GeneRank
TimedPageRank
ObjectRank
HostRank
ProteinRank
CiteRank
FolkRank
DirRank
IsoRank
AuthorRank
ItemRank
TrustRank
MonitorRank
PopRank
BuddyRank
BadRank
BookRank
FactRank
TwitterRank
VisualRank
The remainder of this section explores the uses of PageRank within different
domains. It is devoted to the most interesting and diverse uses and should not,
necessarily, be read linearly. Our intention is not to cover the full details, but to survey
the diversity of applications of PageRank. We recommend returning to the primary
sources for additional detail.
Chemistry · §4.1
Literature · §4.7
Biology · §4.2
Bibliometrics · §4.8
Neuroscience · §4.3
Databases & Knowledge systems · §4.9
Engineered systems · §4.4
Recommender systems · §4.10
Mathematical systems · §4.5
Social networks · §4.11
Sports · §4.6
The web, redux · §4.12
10
D. F. Gleich
4.1. PageRank in chemistry. The term “graph” arose from “chemico-graph”
or a picture of a chemical structure [Sylvester, 1878]. Much of this chemical terminology
remains with us today. For instance, the valence of a molecule is the number of potential
bonds it can make. The valence of a vertex is synonymous with its degree, or the
number of connections it makes in the graph. It is fitting, then, that recent work by
Mooney et al. [2012] uses PageRank to study molecules in chemistry. In particular, they
use PageRank to assess the change in a network of molecules linked by hydrogen bonds
among water molecules. Given the output of a molecular dynamics simulation that
provides geometric locations for a solute in water, the graph contains edges between
the water molecules if they have a potential hydrogen bond to a solute molecule. The
goal is to assess the hydrogen bond potential of a solvent. The PageRank centrality
scores using uniform teleportation with α = 0.85 are strongly correlated with the
degree of the node – which is expected – but the deviance of the PageRank score from
the degree identifies important outlier molecules with smaller degree than many in
their local regions. The authors compare the networks based the PageRank values
with and without a solute to find structural differences.
4.2. PageRank in biology & bioinformatics: GeneRank, ProteinRank,
IsoRank. Biology and bioinformatics are currently awash in network data. Some of
the most interesting applications of PageRank arise when it is used to study these
networks. Most of these applications use PageRank to reveal localized information
about the graph based on some form of external data.
GeneRank. Microarray experiments are a measurement of whether or not a gene’s
expression is promoted or repressed in an experimental condition. Microarrays estimate
the outcomes for thousands of genes simultaneously in a few experimental conditions.
The results are extremely noisy. GeneRank [Morrison et al., 2005] is a PageRankinspired idea to help to denoise them. The essence of the idea is to use a graph of known
relationships between genes to find genes that are highly related to those promoted or
repressed in the experiment, but were not themselves promoted or repressed. Thus,
they use the microarray expression results as the teleportation distribution vector for
a PageRank problem on a network of known relationships between genes. The network
of relationships between genes is undirected, unweighted with a few thousand nodes.
This problem uses a localized teleportation behaviour and, experimentally, the best
choice of α ranges between 0.75 and 0.85. Teleporting is used to focus the search.
Finding correlated genes. This same idea of using a network of known relationships
in concert with an experiment encapsulates many of the other uses of PageRank in
biology. Jiang et al. [2009] use a combination of PageRank and BlockRank [Kamvar
et al., 2003; Kamvar, 2010] on tissue-specific protein-protein interaction networks in
order to find genes related to type 2 diabetes. The teleportation is provided by 34
proteins known to be related to that disease with α = 0.92.
Winter et al. [2012] use PageRank to study pancreatic ductal adenocarcinoma,
a type of cancer responsible for 130,000 deaths each year, with a particularly poor
prognosis (2% mortality after five years). They identified seven genes that better
predicted patient survival than all existing tools, and validated this in a clinical trial.
One curious feature is that their teleportation parameter was small, α = 0.3. This
was chosen based on a cross-validation strategy in a statistically rigorous way. The
particular type of teleportation they used was based on the correlation between the
expression level of a gene and the survival time of the patient.
ProteinRank. The goal of ProteinRank [Freschi, 2007] is similar, in spirit, to
GeneRank. Given an undirected network of protein-protein interactions and human-
PAGERANK BEYOND THE WEB
11
curated functional annotations about what these proteins do, the goal is to find proteins
that may share a functional annotation. Thus, the PageRank problem is, again, a
localized use. The teleportation distribution is given by a random choice of nodes with
a specific functional annotation. The PageRank vector reveals proteins that are highly
related to those with this function, but do not themselves have that function labeled.
Protein distance. Recall that the solution of a PageRank problem for a given
teleportation vector v involves solving (I − αP)x = (1 − α)v. The resolvent matrix
X = (1 − α)(I − αP)−1 corresponds to computing PageRank vectors that teleport to
every individual node. The entry Xi,j is the value of the ith node when the PageRank
problem is localized on node j. One interpretation for this score is the PageRank that
node j contributes to node i, which has the flavor of a similarity score between node
i and j. Voevodski et al. [2009] base an affinity measure between proteins on this
idea. Formally, consider an undirected, unweighted protein-protein interaction network.
Compute the matrix X for α = 0.85, and the affinity matrix S = min(X, XT ). (For
an undirected graph, a quick calculation shows that XT = D−1 XD.) For each vertex
i in the graph, form links to the k vertices with the largest values in row of i of S.
These PageRank affinity scores show a much larger correlation with known protein
relationships than do other affinity or similarity metrics between vertices.
IsoRank. Consider the problem of deciding if the vertices of two networks can
be mapped to each other. The relationship between this problem and PageRank is
surprising and unexpected; although precursor literature existed [Jeh and Widom, 2002;
Blondel et al., 2004]. Singh et al. [2007] proposes a PageRank problem to estimate
how much of a match the two nodes are in a diffusion sense. They call it IsoRank
based on the idea of ranking graph isomorphisms. Let P be the Markov chain for
one network and let Q be the Markov chain for the second network. Then IsoRank
solves a PageRank problem on Q ⊗ P. The solution vector x is a vectorized form of a
matrix X where Xij indicates a likelihood that vertex i in the network underlying P
will match to vertex j in the network underlying Q. See Figure 4.1 for an example. If
we have an apriori measure of similarity between the vertices of the two networks, we
can add this as a teleportation distribution term. IsoRank problems are some of the
largest PageRank problems around due to the Kronecker product (e.g. Gleich et al.
[2010] has a problem with 4 billion nodes and 100 billion edges). But there are quite a
few good algorithmic approaches to tackle them by using properties of the Kronecker
product [Bayati et al., 2013] and low-rank matrices [Kollias et al., 2011].
The IsoRank authors consider the problem of matching protein-protein interaction
networks between distinct species. The goal is to leverage insight about the proteins
from a species such as a mouse in concert with a matching between mouse proteins and
human proteins, based on their interactions, in order to hypothesize about possible
functions for proteins in a human. For these problems, each protein is coded by a gene
sequence. The authors construct a teleportation distribution by comparing the gene
sequences of each protein using a tool called BLAST. They found that using α around
0.9 gave the highest structural similarity between the two networks.
4.3. PageRank in neuroscience. The human brain connectome is one of the
most important networks, about which we understand surprisingly little. Applied
network theory is one of a variety of tools currently used to study it [Sporns, 2011].
Thus, it is likely not surprising that PageRank has been used to study the properties of
networks related to the connectome [Zuo et al., 2011]. Most recently, PageRank helped
evaluate the importance of brain regions given observed correlations of brain activity.
In the resulting graph, two voxels of an MRI scan are connected if the correlation
12
D. F. Gleich
1
2
C
B
4
A
(1) Two graphs
P=
D
3
E
0 1/3 1/2 0
1/2 0 1/2 1
1/2 1/3 0 0
0 1/3 0 0
0 0
0 1/4 0
0 0 1/2 1/4 0
0 1/2 0 1/4 0
1 1/2 1/2 0 1
0 0
0 1/4 0
"
Q=
#
(b) Their stochastic matrices
1
2
3
4
A
0.03
0.04
0.03
0.02
B
0.05
0.07
0.05
0.03
C
0.05
0.07
0.05
0.03
D
0.09
0.15
0.09
0.05
E
0.03
0.04
0.03
0.02
(c) The IsoRank solution
Fig. 4.1. An illustration of the IsoRank problem. The solution, written here as a matrix, gives
the similarity between pairs of nodes of the graph. For instance, node 2 is most similar to node
D. Removing this match, then nodes 1 and 3 are indistinguishable from B and C. Removing these
leaves node 4 equally similar to A and E. In this example we solved (I − αQ ⊗ P)x = (1 − α)e/20
with α = 0.85.
between their functional MRI time-series is high. Edges with weak correlation are
deleted and the remainder are retained with either binary weights or the correlation
weights. The resulting graph is also undirected, and they use PageRank, combined
with community detection and known brain regions, in order to understand changes in
brain structure across a population of 1000 individuals that correlate with age.
Connectome networks are widely hypothesized to be hierarchically organized.
Given a directed network that should express a hierarchical structure, how can we
recover the order of the nodes that minimizes the discrepancy with a hierarchical
hypothesis? Crofts and Higham [2011] consider PageRank for this application on
networks of neural connections from C. Elegans. They find that this gives poor
results compared with other network metrics such as the Katz score [Katz, 1953], and
communicability [Estrada et al., 2008]. In their discussion, the authors note that this
result may have been a mismatch of models, and conjecture that the flow of influence
in PageRank was incorrect. Literature involving Reverse PageRank (Section 3.2)
strengthens this conjecture. Let us reiterate that although PageRank models are easy
to apply, they must be employed with some care in order to get the best results.
4.4. PageRank in complex engineered systems: MonitorRank. The applications of PageRank to networks in chemistry, biology, and neuroscience are part
of the process of investigating and analyzing something we do not fully understand.
PageRank methods are also used to study systems that we explicitly engineered. As
these engineered systems grow, they become increasingly complex, with networks and
submodules interacting in unpredictable, nonlinear ways. Network analysis methods
like PageRank, then, help reveal these details. We’ll see two examples: software
systems and city systems.
MonitorRank. Diagnosing root causes of issues in a modern distributed system
is painstaking work. It involves repeatedly searching through error logs and tracing
debugging information. MonitorRank [Kim et al., 2013] is a system to provide guidance
to a systems administrator or developer as they perform these activities. It returns a
ranked list of systems based on the likelihood that they contributed to, or participated
in, an anomalous situation. Consider the systems underlying the LinkedIn website:
each service provides one or more APIs that allow other services to utilize its resources.
For instance, the web-page generator uses the database and photo store. The photo
store in turn uses the database, and so on. Each combination of a service and a
programming interface becomes a node in the MonitorRank graph. Edges are directed
and indicate the direction of function calls – e.g. web-page to photo store. Given that
an anomaly was detected in a system, MonitorRank solves a personalized PageRank
PAGERANK BEYOND THE WEB
13
problem on a weighted, augmented version of the call graph, where the weights and
augmentation depend on the anomaly detected. (The construction is interesting, albeit
tangential, and we refer readers to that paper for the details.) The localized PageRank
scores help determine the anomaly. The graphs involved are fairly small: a few hundred
to a few thousand nodes.
PageRank of the Linux kernel. The Linux kernel is the foundation for an open
source operating system. It has evolved over the past 20 years with contributions from
nearly 2000 individuals in an effort with an estimated value of $3 billion. As of July
2013, the Linux kernel comprised 15.8 million lines of code containing around 300,000
functions. The kernel call graph is a network that represents dependencies between
functions and both PageRank and reverse PageRank, as centrality scores, produce an
ordering of the most important functions in Linux [Chepelianskii, 2010]. The graphs
were directed with a few million edges. Teleportation was typical: α = 0.85 with
a global, uniform v = e/n. They find that utility functions such as printk, which
prints messages from the kernel, and memset, a routine that initializes a region of
memory, have the highest PageRank, whereas routines that initialize the system such
as start kernel have the highest reverse PageRank. Chepelianskii [2010] further
uses the distribution of PageRank and reverse PageRank scores to characterize the
properties of a software system. (This same idea is later used for Wikipedia too, Zhirov
et al. 2010, Section 4.12.)
Roads and Urban Spaces. Another surprising use of PageRank is with road and
urban space networks. PageRank helps to predict both traffic flow and human
movement in these systems. The natural road construction employed is an interesting
graph. A natural road is more or less what it means: it’s a continuous path, built
from road segments by joining adjacent segments together if the angle is sufficiently
small and there isn’t a better alternative. (For help visualizing this idea, consider
traffic directions that state: “Continue straight from High street onto Main street.”
This would mean that there is one natural road joining High street and Main street.)
Using PageRank with α = 0.95, Jiang et al. [2008] finds that PageRank is the best
network measure in terms of predicting traffic on the individual roads. These graphs
have around 15,000 nodes and around 50,000 edges. Another group used PageRank
to study Markov chain models based on the line-graph of roads [Schlote et al., 2012].
That is, given a graph of intersections (nodes) and roads (edges), the line graph, or
dual graph, changes the role of roads to the nodes and intersections to the edges. In
this context, PageRank’s teleportation mirrors the behavior of starting or ending a
journey on each street. This produces a different value of α for each node that reflects
the tendency of individuals to park, or end their journey, on each street. Note that
this is slightly different setup where each node has a separate teleportation parameter
α, rather than a different entry in the teleportation vector. Assuming that each street
has some probability of a journey ending there, then this system is equivalent to a
more general PageRank construction (Section 5.5). These Markov chains are used
to study road planning and optimal routing in light of new constraints imposed by
electric vehicles.
An urban space is the largest space of a city observable from a single vantage
point. For instance, the Mission district of San Francisco is too large, but the area
surrounding Dolores Park is sufficiently small to be appreciated as a whole. For the
study by Jiang [2009], an urban space is best considered as a city neighborhood or block.
The urban space network connects adjacent spaces, or blocks, if they are physically
adjacent. The networks of urban spaces in London, for instance, have up to 20,000
14
D. F. Gleich
nodes and 100,000 links. In these networks, weighted PageRank (Section 3.4) best
predicts human mobility in a case study of movement within London. It outperforms
PageRank, and in fact, they find that weighted PageRank with α = 1 accounts for up
to 60% of the observed movement. Both using weighted PageRank and α = 1 make
sense for these problems – individuals and businesses are likely to co-locate places with
high connectivity, and individuals cannot teleport over the short time-frames used for
the human mobility measurements. Based on the evidence here, we would hypothesize
that using α < 1 would better generalize over longer time-spans.
4.5. PageRank in mathematical systems. Graphs and networks arise in
mathematics to abstract the properties of systems of equations and processes to
relationships between simple sets. We present one example of what PageRank reveals
about a dynamical system by abstracting the phase-space to a discrete set of points and
modeling transitions among them. Curiously, PageRank and its localization properties
has not yet been used to study properties of Cayley graphs from large, finite groups,
although closely related structures have been examined [Frahm et al., 2012].
PageRank of symbolic images and Ulam networks. Let f be a discrete-time
dynamical system on a compact state space M . For instance, M will be the subset of
R2 formed by [0, 2π] × [0, 2π] for our example below. Consider a covering of M by cells
C. In our forthcoming example, this covering will just be a set of non-overlapping cells
that form a regular, discrete partition into cells of size 2π/N × 2π/N . The symbolic
image [Osipenko, 2007] of f with respect to C is a graph where the vertices are the
cells and Ci ∈ C links to Cj ∈ C if x ∈ Ci and f (x) ∈ Cj . The Ulam network is a
weighted approximation to this graph that is constructed by simulating s starting
points within cell Ci and forming weighted links to their destinations Cj [Shepelyansky
and Zhirov, 2010]. The example studied by those authors, and the example we will
consider here, is the Chirikov typical map.
yt+1 = ηyt + k sin(xt + θt )
xt+1 = xt + yt+1 .
It models a kicked oscillator. We generate T random phases θt and look at the map:
f (x, y) = (xT +1 , yT +1 ) mod 2π
where
x1 = x, y1 = y.
That is, we iterate the map for T steps for each of the T random phase shifts θ1 , . . . , θT .
Applying the construction above with s = 1000 random samples from each cell yields a
directed weighted graph G with N 2 nodes and at most N 2 s edges. PageRank on this
graph, with uniform teleportation, yields beautiful pictures of the transient behaviors
of this chaotic dynamical system; these are easy to highlight with modest teleportation
parameters such as α = 0.85 because this regime inhibits the dynamical system from
converging to its stable attractors. This application is particularly useful for modeling
the effects of different PageRank constructions as we illustrate in Figure 4.2. For that
figure, the graph has 262, 144 nodes and 4, 106, 079 edges, η = 0.99, k = 0.22, T = 10.
4.6. PageRank in sports. Stochastic matrices and eigenvector ranking methods
are nothing new in the realm of sports ranking [Keener, 1993; Callaghan et al., 2007;
Langville and Meyer, 2012]. One of the natural network constructions for sports is
the winner network. Each team is a node in the network, and node i points to node
j if j won in the match between i and j. These networks are often weighted by the
score by which team j beat team i. Govan et al. [2008] used the centrality sense of
PageRank with uniform teleportation and α = 0.85 to rank football teams with these
PAGERANK BEYOND THE WEB
15
Fig. 4.2. PageRank vectors of the symbolic image, or Ulam network, of the Chirikov typical map
with α = 0.9 and uniform teleportation. From left to right, we show the standard PageRank vector,
the weighted PageRank vector using the unweighted cell in-degree count as the weighting term, and
the reverse PageRank vector. Each node in the graph is a point (x, y), and it links to all other points
(x, y) reachable via the map f (see the text). The graph is weighted by the likelihood of the transition.
PageRank, itself, highlights both the attractors (the bright regions), and the contours of the transient
manifold that leads to the attractor. The weighted vector looks almost identical, but it exhibits an
interesting stippling effect. The reverse PageRank highlights regions of the phase-space that are exited
quickly, and thus, these regions are dark or black in the PageRank vector. The solution vectors were
scaled by the cube-root for visualization purposes. These figures are incredibly beautiful and show
important transient regions of these dynamical systems.
winner networks. The intuitive idea underlying these rankings is that of a random
fan that follows a team until another team beats them, at which point they pick
up the new team, and periodically restarts with an arbitrary team. In the Govan
et al. [2008] construction, they corrected dangling nodes using a strongly preferential
modification, although, we note that a sink preferential modification may have been
more appropriate given the intuitive idea of a random fan. Radicchi [2011] used
PageRank on a network of tennis players with the same construction. Again, this was
a weighted network. PageRank with α = 0.85 and uniform teleportation on the tennis
network placed Jimmy Conors in the best player position.
4.7. PageRank in literature: BookRank. PageRank methods help with three
problems in literature. What are the most important books? Which story paths in
hypertextual literature are most likely? And what should I read next?
For the first question, Jockers [2012] defines a complicated distance metric between
books using topic modeling ideas from latent Dirichlet allocation [Blei et al., 2003].
Using PageRank as a centrality measure on this graph, in concert with other graph
analytic tools, allows Jockers to argue that Jane Austin and Walter Scott are the most
original authors of the 19th century.
Hypertextual literature contains multiple possible story paths for a single novel.
Among American children of similar age to me, the most familiar would be the Choose
your own adventure series. Each of these books consists of a set of storylets; at the
conclusion of a storylet, the story either ends, or presents a set of possibilities for
the next story. Kontopoulou et al. [2012] argue that the random surfer model for
PageRank maps perfectly to how users read these books. Thus, they look for the most
probable storylets in a book. For this problem, the graphs are directed and acyclic,
the stochastic matrix is normalized by outdegree, and we have a standard PageRank
problem. They are careful to model a weakly preferential PageRank system that
deterministically transitions from a terminal (or dangling) storylet back to the start of
16
D. F. Gleich
the book. Teleporting is uniform in their experiments. They find that both PageRank
and a ranking system they derive give useful information about the properties of these
stories.
Books & tags: BookRank. Traditional library catalogs use a carefully curated set
of index terms to indicate the contents of books. These enabled content-based search
prior to the existence of fast full-text search engines. Social cataloging sites such as
LibraryThing and Shelfari allow their users to curate their own set of index terms for
books that they read, and easily share this information among the user sites. The
data on these websites consists of books and tags that indicate the topics of books.
BookRank, which is localized PageRank on the bipartite book-tag graph [Meng, 2009],
produces eerily accurate suggestions for what to read next. For instance, if we use
teleportation to localize on Golub and van Loan’s text “Matrix Computations”, Boyd
and Vandenberghe’s “Convex Optimization”, and Hastie, Tibshirani, and Friedman’s
“Elements of Statistical Learning”, then the top suggestion is a book on Combinatorial
Optimization by Papadimitriou and Steiglitz. A similar idea underlies the general
FolkRank system [Hotho et al., 2006] that we’ll see shortly (Section 4.9).
4.8. PageRank in bibliometrics: TimedPageRank, CiteRank, AuthorRank. The field of bibliometrics is another big producer and consumer of network
ranking methods, starting with seminal work by Garfield on aggregating data into
a citation network between journals [Garfield, 1955; Garfield and Sher, 1963] and
proceeding through Pinski and Narin [1976], who defined a close analogue of PageRank.
In almost all of these usages, PageRank is used as a centrality measure to reveal the
most important journals, papers, and authors.
Citations among journals. The citation network Garfield originally collected and
analyzed is the journal-journal citation network. It is a weighted network where each
node is a journal and each edge is the number of citations between articles of the
journals. ISI’s impact factor is a more refined analysis of these citation patterns.
Bollen et al. [2006] takes ISI’s methods a step further and finds that a combination of
the impact factor with the PageRank value in the journal citation produces a ranked
list of journals that better correlates with experts’ judgements. PageRank is used as a
centrality measure here with uniform teleportation and weights that correspond to
the weighted citation network. The graph had around 6000 journals. The Eigenfactor
system [West et al., 2010] uses a PageRank vector on the journal co-citation network
with uniform teleportation and α = 0.85 to measure the influence of a journals. It also
shows these rankings on easy-to-browse website.
Citations among papers: TimedPageRank, CiteRank. Moving beyond individual
journals, we can also study the citation network among individual papers using
PageRank. In a paper citation network, each node is an individual article and the
edges are directed based on the citation. Modern bibliographic and citation databases
such as arXiv and DBLP make these networks easy to construct. They tend to have
hundreds of thousands of nodes and a few million edges. TimedPageRank is an idea to
weight the edges of the stochastic matrix in PageRank such that more recent citations
are more important. Formally, it is the solution of
(I − αAT D−1 W)x = (1 − α)e
where W is a diagonal matrix with weights between 0 and 1 that reflects the age of
the paper (1 is recent and 0 is old). The matrix AT D−1 W is column sub-stochastic
and so this is a pseudo-PageRank problem. CiteRank is a subsequent idea that uses
the teleportation in PageRank to increase the rank of recent articles [Walker et al.,
PAGERANK BEYOND THE WEB
17
2007]. Thus, vi is smaller if paper i is older and vi is larger if paper i is more recent.
The goal of both methods is to produce temporally relevant orderings that remove the
bias of older articles to acquire citations.
While the previous two papers focused on how to make article importance more
accurate, Chen et al. [2007] attempts to use PageRank in concert with the number of
citations to find hidden gems. One notable contribution is the study of α in citation
analysis: based on a heuristic argument about how we build references for an article,
they recommend α = 0.5. Moreover, they find papers with higher PageRank scores
than would be expected given their citation count. These are the hidden gems of the
literature. Ma et al. [2008] uses the same idea in a larger study and find a similar
effect.
Citations among authors: AuthorRank. Another type of bibliographic network is
the co-authorship graph. For each paper, insert edges among all co-authors. Thus,
each paper becomes a clique in the co-authorship network. The weights on each edge
are either uniform (and set to 1), based on the number of papers co-authored, or based
on another weighting construction defined in that paper. All of these constructions
produce an undirected network. PageRank on this network gives a practical ranking
of the most important authors [Liu et al., 2005]. The teleportation is uniform with
α = 0.85, or can be focused on a subset of authors to generate an area-specific ranking.
Their data have a few thousand authors. These graphs are constructions based on an
underlying bipartite matrix B that relates authors and papers. More specifically, the
T
weighted co-authorship network
0 Bisthe matrix BB . Many such constructions can be
related back to the matrix BT 0 [Dhillon, 2001]. We are not aware of any analysis
that makes a relationship between PageRank in the bipartite graph B0T B
and the
0
weighted matrix BBT .
Author, paper, citation networks. Citation analysis and co-authorship analysis
can, of course, be combined, and that is exactly what Fiala et al. [2008] and Jezek
et al. [2008] do. Whereas Liu et al. [2005] study the co-authorship network, here,
they study a particular construction that joins the bipartite author-paper network to
the citation network to produce an author-citation network. This is a network where
author i links to author j if i has a paper that cites j where j is not a co-author on
that particular paper. Using α = 0.9 and uniform teleportation produces another
helpful list of the most important authors. In the notation of the previous paragraph,
a related construction is the network with adjacency matrix
0 B
A=
,
BT C
where B is the bipartite author-paper matrix and C is the citation matrix among
papers. PageRank on these networks takes into account both the co-authorship and
directed citation information, and it rewards authors that have many, highly cited
papers. The graphs studied have a few hundred thousand authors and author-author
citations.
4.9. PageRank in databases and knowledge information systems: PopRank, FactRank, ObjectRank, FolkRank. Knowledge information systems store
codified forms of information, typically as a relational database. For instance, a
knowledge system about movies consists of relationships between actors, characters,
movies, directors, and so own. Contemporary information systems also often contain
large bodies of user-generated content through tags, ratings, and such. Ratings are a
sufficiently special case that we review them in a forthcoming section (Section 4.10),
18
D. F. Gleich
but we will study PageRank and tags here. PageRank serves important roles as both
a centrality measure and localized measure in networks derived from a knowledge
system. We’ll also present slightly more detail on four interesting applications.
Centrality scores: PopRank, FactRank. PageRank’s role as a centrality measure in
a knowledge information system is akin to its role on the web as an importance measure.
For instance, the authors of PopRank [Nie et al., 2005] consider searching through
large databases of objects – think of academic papers – that have their own internal
set of relationships within the knowledge system – think of co-author relationships.
But these papers are also linked to by websites. PopRank uses web-importance as a
teleportation vector for a PageRank vector defined on the set of object relationships.
The result is a measure of object popularity biased by its web popularity. One of
the challenges in using such a system is that collecting good databases of relational
information is hard. FactRank helps with this process [Jain and Pantel, 2010]. It is a
measure designed to evaluate the importance and accuracy of a fact network. A fact
is just a sentence that connects two objects, such as “David-Gleich wrote the-paper
PageRank-Beyond-The-Web.” These sentences come from textual analysis of large
web crawls. In a fact network, facts are connected if they involve the same set of
objects. Variations on PageRank with uniform teleportation provide lists of important
facts. The authors of FactRank found that weighting relationships between facts and
using PageRank scores of this weighted network gave higher performance than both a
baseline and standard PageRank method in the task of finding correct facts. The fact
networks are undirected and have a few million nodes.
Localized scores: Random-walk with restart, Semi-Supervised Learning. Prediction
tasks akin to the bioinformatics usages of PageRank are standard within knowledge
information systems: networks contain noisy relationships, and the task is inferring,
or predicting, missing data based on these relationships. Zhou et al. [2003] used a
localized PageRank computation to infer the identity of handwritten digits from only a
few examples. These problems were called semi-supervised learning on graphs because
they model the case of finding a vector over vertices (or learning a function) based
on a few values of the function (supervised). It differs from the standard supervised
learning problem because the graph setup implies that only predictions on the vertices
are required, instead of the general prediction problem with arbitrary future inputs. In
the particular study, the graph among these images is based on a radial basis function
construction. For this task α = 0.99 in the pseudo-PageRank system (I − αP)Ỹ = S,
where S is a binary matrix indicating known samples Sij = 1 if image i is known to
be digit j. The largest value in each row of Y = DỸ gives the predicted digit for any
unknown image. While these graphs were undirected, later work [Zhou et al., 2005]
showed how to use PageRank with global teleportation, in concert with symmetric
Laplacian structure defined on a directed graph [Chung, 2005], to enable the same
methodology on a general directed graph.
Pan et al. [2004] define a random walk with restart, which is exactly a personalized
PageRank system, to infer captions for a database of images. Given a set of images
labeled by captions, define a graph where each image is connected to its regions,
each region is connected to other regions via a similarity function, and each image is
connected to the terms in its caption. A query image is a distribution over regions,
and we find terms by solving a PageRank problem with this as the teleportation vector.
These graphs are weighted, undirected graphs. Curiously, the authors chose α based
on experimentation and found that α = 0.1 or α = 0.2 works best. They attribute the
difference to the incredibly small diameter of their network. Subsequent work in the
PAGERANK BEYOND THE WEB
19
same vein showed some of the relationships with the normalized Laplacian matrix of a
graph [Tong et al., 2006] and returned to a larger value of α around 0.9.
Application 1 – Database queries: ObjectRank. ObjectRank is an interesting type
of database query [Balmin et al., 2004]. A typical query to a database will retrieve all
of the rows of a specified set of tables matching a precise criteria, such as, “find all
students with a GPA of 3.5 that were born in Minnesota.” These tables often have
internal relationships – the database schema – that would help determine which are
the most important returned results. In the ObjectRank model, a user queries the
database with a textual term. The authors describe a means to turn the database
objects and schema into a sub-stochastic transition matrix and define ObjectRank as
the query-dependent solution of the PageRank linear system where the teleportation
vector reflects textual matches. They suggest a great deal of flexibility with defining the
weights of this matrix. For instance, there may be no natural direction for many of these
links and the authors suggest differently weighting forward edges and backward edges
– their intuition is that a paper cited by many important papers is itself important,
but that citing important paper papers does not transfer any importance. They use
α = 0.85 and the graphs have a few million edges.
Application 2 – Folksonomy search: FolkRank. A more specific situation is folksonomy search. A folksonomy is a collection of objects, users, and tags. Each entry is
a triplet of these three items. A user such as myself may have tagged a picture on the
flickr network with the term “sunset” if it contained a sunset, thus creating the triplet
(picture,user,“sunset”). FolkRank scores [Hotho et al., 2006] are designed to measure
the importance of an object, tag, or user with respect to a small set of objects, tags,
or users that define a topic. (This idea is akin to topic-sensitive PageRank, Haveliwala
2002.) These scores then help reveal important objects related to a given search, as
well as the tags that relate them. The scores are based on localized PageRank scores
from an undirected, tripartite weighted network. There is a wrinkle, however. The
FolkRank scores are taken as the difference between a PageRank vector computed
with α = 1 and α = 1/2. The graph is undirected, so the solution with α = 1 is
just the weighted degree distribution. Thus, FolkRank downweights items that are
important for everyone.
Application 3 – Semantic relatedness. The Open Directory Project, or odp, is a
hierarchical, categorical index of web-pages that organizes them into related groups.
Bar-Yossef and Mashiach [2008] suggests a way of defining the relatedness of two
categories on odp using their localized PageRank scores. The goal is to generalize
the idea of the least-common ancestor to random walks to give a different sense of
the distance between categories. To do so, create a graph from the directed hierarchy
in the odp. Let x be the reverse PageRank vector that teleports back to a single
category, and let y be the reverse PageRank vector that teleports back to another
(single) category. Then the relatedness of these categories is the cosine of the angle
between x and y. Let x be the localized PageRank vector (Note the use of reverse
PageRank here so that edges go from child to parent.) They show evidence that this
is a useful measure of relationship in ODP.
Application 4 – Logic programming. A fundamental challenge with scaling logic
programming systems like Prolog is that there is an exponential explosion of potential
combinations and rules to evaluate and, unless the system is extremely well-designed,
these cannot be pruned away. This limits applications to almost trivial problems.
Internally, Prolog-type systems resolve, or prove, logical statements using a search
procedure over an implicitly defined graph that may be infinite. At each node of the
20
D. F. Gleich
graph, the proof system generates all potential neighbors of the node by applying a
rule set given by the logic system. Thus, given one node in the graph, the search
procedure eventually visits all nodes. Localized PageRank provides a natural way
to restrict the search space to only “short” and “likely” proofs [Wang et al., 2013].
Formally, they use PageRank’s random teleportation to control the expansion of the
search procedure. However, there is an intuitive explanation for the random restarts
in such a problem: periodically we all abandon our current line of attack in a proof
and start out fresh. Their system with localized PageRank allows them to realize this
behavior in a rigorous way.
4.10. PageRank in recommender systems: ItemRank. A recommender
system attempts to predict what its users will do based on their past behavior. Netflix
and Amazon have some of the most famous recommendation systems that predict
movies and products, respectively, their users will enjoy. Localized PageRank helps to
score potential predictions in many research studies on recommender systems.
Query reformulation. A key component of modern web-search systems is predicting
future queries. Boldi et al. [2008] run localized PageRank on a query reformulationgraph that describes how users rewrite queries with α = 0.85. Two queries, q1 and q2 ,
are connected in this graph if a user searched for q1 before q2 within a close time-frame
and both q1 and q2 have some non-trivial textual relationships. This graph is directed
and weighted. The teleportation vector is localized on the current query, or a small set
of previously used terms. PageRank has since had great success for many tasks related
to query suggestion and often performs among the best methods evaluated [Song et al.,
2012].
Item recommendation: ItemRank. Both Netflix and Amazon’s recommender systems are called item recommendation problems. Users rate items – typically with
a 5-star scale – and we wish to recommend items that a user will rate highly. The
ratings matrix is an items-by-users matrix where Rij is the numeric rating given to
item i by user j. These ratings form a bipartite network between the two groups and
we collapse this to a graph over items as follows. Let G be a weighted graph where
the weights on an edge (i, j) are the number of users that rated both items i and j.
(These weights are equivalent to the number of paths of length 2 between each pair of
items in terms of the bipartite graph.) Let P be the standard weighted random walk
construction on G. Then the ItemRank scores [Gori and Pucci, 2007] are the solutions
of:
(I − αP)S = (1 − α)RD−1
R
where DR are column sums of the rating matrix. Each column of S is a set of
recommendations for user j, and Sij is a proxy for the interest of user j in item i.
Note that any construction of the transition matrix P based on correlations between
items based on user ratings would work in this application as well.
Link prediction. Given the current state of a network, link prediction tries to
predict which edges will come into existence in the future. Liben-Nowell and Kleinberg
[2006] evaluated the localized PageRank score of an unknown edge in terms of its
predictive power. These PageRank values were entries in the matrix (I − αP)−1 for
edges that currently do not exist in the graph. PageRank with α between 0.5 and 0.99
was not one of their best predictors, but the Katz matrix (I−αA)−1 was one of the best
with α = 0.0005. Note that Katz’s matrix is, implicitly, a pseudo-PageRank problem
1
if α < dmax
where dmax is the largest degree in the graph. The co-authorship graphs
tested seem to have had degrees less than 2000, making this hidden pseudo-PageRank
PAGERANK BEYOND THE WEB
21
problem one of the best predictors of future co-authorship. More recent work using
PageRank for predicting links on the Facebook social network includes a training
phase to estimate weights of the matrix P to achieve higher prediction [Backstrom
and Leskovec, 2011]. Localized PageRank is believed to be part of Twitter’s follower
suggestion scheme too [Bahmani et al., 2010].
4.11. PageRank in social networks: BuddyRank, TwitterRank. PageRank serves three purposes in a social network, where the nodes are people and the
edges are some type of social relationship. First, as we discussed in the previous
section, it can help solve link prediction problems to find individuals that will become
friends soon. Second, it serves a classic role in evaluating the centrality of the people
involved to estimate their social status and power. Third, it helps evaluate the potential
influence of a node on the opinions of the network.
Centrality: BuddyRank. Centrality methods have a long history in social networks
– see Katz [1953] and Vigna [2009] for a good discussion. The following claim is difficult
to verify, but we suspect that the first use of PageRank in a large-scale social network
was the BuddyRank measure employed by BuddyZoo in 2003.1 BuddyZoo collected
contact lists from users of the AOL Instant Messenger service and assembled them
into one of the first large-scale social networks studied via graph theoretic methods.
Since then, PageRank has been used to rank individuals in the Twitter network by
their importance [Java, 2007] and to help characterize properties of the Twitter social
network by the PageRank values of their users [Kwak et al., 2010]. These are standard
applications of PageRank with global teleportation and α ≈ 0.85.
Influence. Finding influential individuals is one of the important questions in social
network analysis. This amounts to finding nodes that can spread their influence widely.
More formalizations of this question result in NP-hard optimization problems [Kempe
et al., 2003] and thus, heuristics and approximation algorithms abound [Kempe
et al., 2003, 2005]. Using Reverse PageRank with global teleportation as a heuristic
outperforms out-degree for this task, as shown by Java et al. [2006] for web-blog
influence and Bar-Yossef and Mashiach [2008] for the social network LiveJournal.
Reverse PageRank, instead of traditional PageRank, is the correct model to understand
the origins of influence – the distinction is much like the treatment of hubs and
authorities in other ranking models on networks [Kleinberg, 1999; Blondel et al., 2004].
These ideas also extend to finding topical authorities in social networks by using the
teleportation vector and topic-specific transition probabilities to localize the PageRank
vector in TwitterRank [Weng et al., 2010].
4.12. PageRank in the web, redux: HostRank, DirRank, TrustRank,
BadRank, VisualRank. At the conclusion of our survey of applications, we return
to uses of PageRank on the web itself. Before we begin, let us address the elephant
in the room, so to speak. Does Google still use PageRank? Google reportedly uses a
basket of ranking metrics to determine the final order that results are returned. These
evolve continuously and vary depending on where and when you are searching. It is
unclear to what extent PageRank, or more generally, link analysis measures play a
role in Google’s search ordering, and this is a closely guarded secret unlikely to be
known outside of an inner-circle at Google. One the one hand, in perhaps the only
large-scale published study on PageRank’s effectiveness in a search engine, Najork et al.
[2007] found that it underperformed in-degree. On the other hand, PageRank is still
widely believed to still play some role based on statements from Google. For instance,
1 http://web.archive.org/web/20050724231459/http://buddyzoo.com/
22
D. F. Gleich
Matt Cutts, a Google engineer, wrote about how Google uses PageRank to determine
crawling behavior [Cutts, 2006], and later wrote about how Google moved to a full
substochastic matrix in terms of their PageRank vector [Cutts, 2009]. The latter case
was designed to handle a new class of link on the web called rel=nofollow. This was
an optional HTML parameter that would tell a crawler that the following link is not
useful for relevance judgements. All the major web companies created this parameter
to combat links created in the comment sections of extremely high quality pages such
as the Washington Post. These links are created by users of the Washington Post,
not the staff themselves, and shouldn’t constitute an endorsement on a page. Cutts
described how Google’s new PageRank equation would count these rel=nofollow
links in the degree of a node when it was computing a stochastic normalization, but
would remove the links when computing relevance. For instance, if my page had three
true links and two rel=nofollow links, then my true links would have probabilities
1/5 instead of 1/3, and the sum of my outgoing probability would be 3/5 instead of 1.
Thus, Google’s PageRank computation is a pseudo-PageRank problem now.
Outside of Google’s usage, PageRank is also used to evaluate the web at coarser
levels of granularity through HostRank and DirRank. Reverse PageRank provides
a good measure of a page’s similarity to a hub, according to both Fogaras [2003]
and Bar-Yossef and Mashiach [2008]. PageRank and reverse PageRank also provide
information on the “spaminess” of particular pages through metrics such as TrustRank
and BadRank. PageRank-based information also helped to identify spam directly in a
study by Becchetti et al. [2008]. Finally, PageRank helps identify canonical images to
place on a web-search result (VisualRank).
Coarse PageRank: HostRank, DirRank. Arasu et al. [2002] was an important early
paper that defined HostRank, where the web is aggregated at the level of hostnames.
In this case, all links to and from a hostname, such as www.cs.purdue.edu, become
equivalent. This particular construction models a random surfer that, when visiting a
page, makes a censored, or silent, transition within all pages on the same host, and then
follows a random link. The HostRank scores are the sums of these modified PageRank
scores on the pages within each host [Gleich and Polito, 2007]. Later work included
BlockRank [Kamvar et al., 2003], which used HostRank to initialize PageRank, and
DirRank [Eiron et al., 2004], which forms an aggregation at the level of directories of
websites.
Trust, Reputation, & Spam: TrustRank, BadRank. PageRank typically provides
authority scores to estimate the importance of a page on the web. As the commercial
value of websites grew, it became highly profitable to create spam sites that contain no
new information content but attempt to capture Google search results by appearing
to contain information. BadRank [Sobek, 2003] and TrustRank [Gyöngyi et al., 2004]
emerged as new, link analysis tools to combat the problem. Essentially, these ideas
solve localized, reverse PageRank problems. The results are either used directly, or
as a “safe teleportation” vector for PageRank, as in TrustRank, or in concert with
other techniques, as likely done in BadRank. Kolda and Procopio [2009] generalizes
these models and includes the idea of adding self-links to fix the dangling nodes, like
in sink preferential PageRank, but they add them everywhere, not just at dangling
nodes. For spam-link applications, this way of handling dangling nodes is superior –
in a modeling sense – to the alternatives.
Wikipedia. Wikipedia is often used as a subset of the web for studying ranking. It
is easy to download the data for the entire website, which makes building the web-graph
convenient. (A crawl from a few years ago is in the sparse matrix repository, Davis and
PAGERANK BEYOND THE WEB
23
Hu 2010, as the matrix Gleich/wikipedia-20070206.) Current graphs of the English
language pages have around 100,000,000 links and 10,000,000 articles. The nature
of the pages on Wikipedia also makes it easy to evaluate results anecdotally. For
instance, we would all raise an eyebrow and demand explanation if “Gene Golub” was
the page with highest global PageRank in Wikipedia. On the other hand, this result
might be expected if we solve a localized PageRank problem around the Wikipedia
article for “numerical linear algebra.” Wissner-Gross [2006] used Wikipedia as a test
set to build reading lists using a combination of localized and global PageRank scores.
Later, Zhirov et al. [2010] computed a 2d ranking on Wikpedia by combining global
PageRank and reverse PageRank. Finally, this 2d ranking showed that Frank Sinatra
was one of the most important people [Eom et al., 2014].
Image search: VisualRank. PageRank also helps to identify “canonical” images to
display as a visual summary of a larger set of images returned from an image search
engine. In the VisualRank system, Jing and Baluja [2008] compute PageRank of
an image similarity graph generated from an image search result. The graphs are
small – around 1000 nodes – which reflects the standard textual query results, and
they are also symmetric and weighted. They solve a global PageRank problem with
uniform teleportation or high-result biased teleportation. The highest ranked images
are canonical images of Mona Lisa amid a diverse collection of views.
5. PageRank generalizations. Beyond the applications discussed so far, there
is an extremely wide set of PageRank-like models that do not fit into the canonical
definition and constructions from Section 3. These support a wide range of additional
applications with mathematics that differs slightly, and some of them are formal
mathematical generalizations of the PageRank vectors. For instance, in prior work, we
studied PageRank with a random teleportation parameter [Constantine and Gleich,
2010]. The standard deviation of these vectors resulted in increased accuracy in
detecting spam pages on the web. We now survey some of these formal generalizations.
5.1. Diffusions, damped sums, & heat kernels. Recall that the pseudoPageRank vector is the solution of (2.3),
(I − αP̄)y = f .
Since all of the eigenvalues of P̄ are bounded by 1 in magnitude, the solution y has an
expansion in terms of the Neumann series:
y=
∞
X
k
αk P̄ f .
k=0
This expressions gives the pseudo-PageRank vector as a damped sum of powers of P̄
k
where each power, P̄ , has the geometrically decaying weight αk . These are often called
damped diffusions because this equation models how the quantities in f probabilistically
diffuse through the graph where the probability of a path of length k is damped by αk .
Many other sequences serve the same purpose as pointed out by a variety of authors.
Generalized damping. Perhaps the most general setting for these ideas is the
generalized damped PageRank vector:
z=
∞
X
k=0
k
γk P̄ f
(5.1)
24
D. F. Gleich
P
where γk is a non-negative `1 -sequence (that is, k γk < ∞ and γk ≥ 0). This reduces
to PageRank if γk = αk . Huberman et al. [1998] suggested using such a construction
where γk arises from real-world path following behaviors on the web, which they found
to resemble inverse Gaussian functions. Later results from Baeza-Yates et al. [2006]
proposed essentially the same formula in (5.1). They suggested a variety of interesting
functions γk , including some with only a finite number of non-zero terms. These
authors drew their motivation from the earlier work of TotalRank [Boldi, 2005], which
1
1
suggested γk = k+1
− k+2
in order to evaluate the TotalRank vector:
Z
1
z=
(I − αP̄)−1 (1 − α)v dα.
0
This integrates over all possible values of α. (As an aside, this integral is well defined
because a unique limiting PageRank value exists at α = 1, see Section 5.2. This
sidesteps a technical issue with the singular matrix at α = 1.) Our work with making the
value of α in PageRank a random variable is really a further generalization [Constantine
and Gleich, 2010]. Let x(α) be a parameterized form for the PageRank vector for a
fixed graph and teleportation vector. Let A be a random variable supported on [0, 1]
with an infinite number of finite moments, that is, E[Ak ] < ∞ for all k. Intuitively,
A is the probability that a random user of the web follows a link. Our idea was to
use the expected value of PageRank E[x(A)] to produce a ranking that reflected the
distribution of path-following behaviors in the random surfers. We showed:
E[x(A)] =
∞
X
(E[Ak ] − E[Ak+1 ])Pk v.
k=0
This results in a family of sequences of γk that depend on the random variable A.
Recent work by Kollias et al. [2013] shows how to evaluate these generalized damped
vectors as a polynomial combination of PageRank vectors in the sense of (2.2).
Heat kernels & matrix exponentials. Another specific case of generalized damping
arises from the matrix exponential, or heat kernel:
z = eβ P̄ f =
∞
X
βk
k=0
k!
k
P̄ f .
Such functions arose in a wide variety of domains that would be tangential to review
here [Estrada, 2000; Miller et al., 2001; Kondor and Lafferty, 2002; Farahat et al.,
2006; Chung, 2007; Kunegis and Lommatzsch, 2009; Estrada and Higham, 2010]. In
terms of a specific relationship with PageRank, Yang et al. [2007] noted that the
pseudo-PageRank vector itself was a single-term approximation to these heat kernel
diffusions. Consider
z = eβ P̄ f
⇔
e−β P̄ z = f
⇔
(I − β P̄ + . . .)z = f .
If we truncate the heat kernel expansion after just the first two terms (I − βP), then we
arise at the pseudo-PageRank vector. (A similar result holds for the formal PageRank
vector too.)
5.2. PageRank limits & eigenvector centrality. In the definition of PageRank used in this paper, we assume that α < 1. PageRank, however, has a unique
well-defined limit as α → 1 [Serra-Capizzano, 2005; Boldi et al., 2005, 2009b]. This
25
PAGERANK BEYOND THE WEB
is easy to prove using the Jordan canonical form for the case of PageRank (2.2),
but extensions to pseudo-PageRank are slightly more nuanced. As in the previous
section, let x(α) be the PageRank vector as a function of α for a fixed stochastic P:
(I − αP)x(α) = (1 − α)v. Let XJX−1 be the Jordan canonical form of P. Because P
is stochastic, it’s eigenvalues on the unit circle are all semi-simple [Meyer, 2000, page
696]. Thus:
J=
hI
i
D1
J2
,
where D1 is a diagonal matrix of the eigenvalues on the unit circle and J2 is a Jordan
block for all eigenvalues with |λ| < 1. We now substitute this into the PageRank
equation:
(I − αP)x(α) = (1 − α)v ⇔ (I − αJ)−1 x̂(α) = (1 − α) |{z}
v̂ .
| {z }
−1
X
X−1 x(α)
v
Using the structure of J decouples these equations:
i
h I
I
I
−α
hI
i x̂(α)0
D1
J2
x̂(α)1
x̂(α)2
= (1 − α)
v̂0
v̂1
v̂2
.
As α → 1, both x̂(α)1 and x̂(α)2 go to 0 because these linear systems remain nonsingular. Also, note that x̂(α)0 = v̂1 for all α =
6 1, so this point is a removable
singularity. Thus, x̂ can be uniquely defined at α = 1, and hence, so can x. Vigna
[2005] uses the structure of this limit to argue that taking α → 1 in practical applications
is not useful unless the underlying graph is strongly connected, and they propose a new
PageRank construction to ensure this property. Subsequent work by Vigna [2009] does
a nice job of showing how limiting cases of PageRank vectors converge to traditional
eigenvector centrality measures from bibliometrics [Pinski and Narin, 1976] and social
network analysis [Katz, 1953].
The pseudo-PageRank problem does not have nice limiting properties in our
formulation. Let y(α) be a parametric form for the solution of the pseudo-PageRank
system (I − αP̄)y = f . As α → 1, then y → ∞, unless the non-zero support of f
lies outside of a recurrent class, in which case y → 0. Boldi et al. [2005] defines the
PseudoRank system as:
(I − αP̄)y = (1 − α)f
instead. This system always has a non-infinite limit as α → 1. It could, however, have
zero as a limit if P̄ has all eigenvalues less than 1.
5.3. Over-teleportation, negative teleportation, & the Fiedler vector.
The next generalization of PageRank is to values of α > 1. These arose in our prior
work to understand the convergence of quadrature formulas for approximating the
expected value of PageRank with random teleportation parameters [Constantine and
Gleich, 2010]. Mahoney et al. [2012] subsequently showed an amazing relationship
among (i) the Fiedler vector of a graph [Fiedler, 1973; Anderson and Morley, 1985;
Pothen et al., 1990], (ii) a particular generalization of the PageRank vector, which we
call MOV, and (iii) values of α > 1.
26
D. F. Gleich
The Fiedler vector. In contrast to the remainder of this paper, the constructions
and statements in this section are specific to connected, undirected graphs with
symmetric adjacency matrices. The conductance of a set of vertices in a graph is
defined as the number of edges leaving that set, divided by the sum of the degrees of
the vertices within the set. Conductance and its relatives are often used as numeric
quality scores for graph partitioning in parallel processing [Pothen et al., 1990] and
for community detection in graphs [Schaeffer, 2007]. It is NP-hard to find the set of
smallest conductance, but Fiedler’s vector reveals information about the presence of
small conductance sets in a graph through the Cheeger inequality [Chung, 1992]. Let
G be a connected, undirected graph with symmetric adjacency matrix A and diagonal
degree matrix D. The Fiedler vector is the generalized eigenvector of (D−A)q = λ∗ Dq,
with the smallest positive eigenvalue λ∗ > 0. All of the generalized eigenvalues are
non-negative, the smallest is 0, and the largest is bounded above by 1. Cheeger’s
inequality bounds the relationship between λ∗ and the set of smallest conductance in
the graph.
MOV. The MOV vector is defined as the pseudo-inverse solution r in the consistent
linear system of equations:
[(D − A) − γD]r = ρ(γ)Ds,
(5.2)
where γ < λ∗ , s is a “seed” vector such that sT De = 0, and ρ(γ) is a scaling constant
such that r has a fixed norm. When γ = 0, this system is singular but consistent,
and thus, we take the pseudo-inverse solution. Note that this is equivalent to the
pseudo-PageRank problem:
(I − αP)z = αρ(γ)f̂
where α =
1
1−γ ,
z = Dr, and f̂ = Ds. The properties of s in MOV imply that
T
f̂ e = 0, and thus, f̂ must have negative elements, which generalizes the standard
pseudo-PageRank.
In a small surprise, allowing f to take on negative values results in no additional
modeling power in the case of symmetric A. To establish this result, we first observe
that:
σ
(I − αAD−1 ) 1−α
d = σd.
This preliminary fact shows that the pseudo-PageRank vector of an undirected graph
with teleportation according to the degree vector d simply results in a rescaling. We
can use this property to shift any f with negative values in a controlled manner:
σ
(I − αP) (z + 1−α
d) = αf̂ + σd,
|
{z
} | {z }
y
f
where σ is chosen such that f ≥ 0 element-wise. Solving these shifted pseudo-PageRank
systems, then, effectively computes the solution z with a well-understood bias term θd.
σ
This is easy to remove afterwards: z = y − 1−α
d, at which point we can normalize z
to account for ρ(γ) if desired.
Values of α > 1. While this generalization with negative entries in f gives
no additional mathematical power, it does permit a seamless limit from PageRank
1
> 1. The formal result is that the
vectors to the Fiedler vector. Let α∗ = 1−λ
∗
1
limit limα→α∗ ρ(α) z(α) = q, the Fiedler vector. Note that for the construction of
PAGERANK BEYOND THE WEB
27
P = AD−1 on an undirected, connected graph, we have that Pk → eT1 d deT as
k → ∞. Thus, when α = 1, the MOV solution z is equivalent to the solution of
(I − (P − eT1 d deT ))z = f because the right hand side of f is orthogonal to the left
eigenvector eT . As all of the eigenvalues of (P − eT1 d deT ) are distinct from 1, this is
a non-singular system. And this fact allows the limit construction to pass through
α = 1 seamlessly. If we additionally assume that f T q 6= 0, then
lim
α→α∗
1
z(α) = q,
ρ(α)
and the limiting value of PageRank with over-teleportation is the Fiedler vector. The
analysis in Mahoney et al. [2012], then, interpolates many of the arguments in Vigna
[2009] beyond α = 1 to yield important relationships between spectral graph theory
and PageRank vectors.
5.4. Complex-valued teleportation parameters and a time-dependent
generalization. Again, let x(α) be the PageRank vector (in the sense of (2.2)) as a
function of α for a fixed graph and teleportation vector. Mathematically, the PageRank
vector is a rational function of α. This simple insight produces a host of possibilities,
one of which is evaluating the derivative of the PageRank vector [Boldi et al., 2005;
Golub and Greif, 2006; Gleich et al., 2007]. Another is that PageRank with complexvalued α is a reasonable mathematical generalization [Horn and Serra-Capizzano, 2007].
Let α ∈ C with |α| < 1, then x(α) has some interesting properties and usages. In
Constantine and Gleich [2010], we needed to bound kx(α)k1 when α was complex. If
α is real and 0 < α < 1, then kx(α)k1 = 1 independent of the choice of α. However, if
|1−α|
α is complex we have: kxk1 ≤ 1−|α|
. Later, in Gleich and Rossi [2014], we found that
complex values of α arise in computing closed form solutions to PageRank dynamical
systems where the teleportation vector is a function of time, but the graph remains
fixed. Specifically, the PageRank vector with complex teleportation arises in the
steady-state time-dependent solution of
x0 (t) = (1 − α)v(t) − (I − αP)x(t),
when v(t) oscillates between a fixed set of vectors. Thus, PageRank with complex teleportation is both an interesting mathematical problem and has practical applications
in a time-dependent generalization of PageRank.
5.5. Censored node constructions. The final generalized PageRank construction we wish to discuss is, in fact, a PageRank system hiding inside a Markov chain
construction with a different type of teleportation. In order to motivate the particular
form of this construction, we first review an alternative derivation of the PageRank
vector.
A censored node in a Markov chain is one that exhibits a virtual influence on
the chain in the sense that walks proceed through it as if it were not present. Let
us illustrate this idea by crafting teleportation behavior into a Markov chain in a
different way and computing the PageRank vector itself by censoring that Markov
chain. Suppose that we want to find the stationary distribution of a walk where, if a
surfer wants to teleport, they first transition to a teleport state, and then move from
the teleport state according to the teleportation distribution. The transition matrix of
the Markov chain is:
αP
v
0
P =
.
(1 − α)eT 0
28
D. F. Gleich
And the stationary distribution of this Markov chain is:
0
αP
v x0
x
=
,
eT x0 + γ = 1.
γ
(1 − α)eT 0 γ
Censoring the final teleportation state amounts to modeling its influence on the
stationary distribution, but leaving it with no final contribution. Put more formally,
the stationary distribution of the censored chain is just x0 renormalized to be a
probability distribution: x = x0 /eT x0 . In other words, censoring that state models
pretending that it wasn’t there when determining the stationary distribution, but the
transitions through it still took place; this is equivalent to the standard teleporting
behavior. The vector x is also the PageRank vector of α, P, v, which follows from
x=
1−α 0
γ x
=
1−α
γ
[αPx0 + γv] = αPx + (1 − α)v.
Tomlin [2003], Eiron et al. [2004] and Lee et al. [2007, written in 2003] were some
of the first to observe this property in the context of PageRank; although censoring
Markov chains goes back much further.
There is a more general class of PageRank-style methods that craft transitions
akin to non-uniform teleportation through a censored node construction. Consider,
for example, adding a teleportation node c that connects to all nodes of a network
as in Figure 5.1. This construction gives rise to an implicit PageRank problem with
dmax
α = dmax
+1 as we now show. Let
A0 =
A
vT
e
0
be the adjacency matrix for the modified graph, where v is the teleportation destination
vector. A uniform random walk on this adjacency structure has a single recurrent class,
and thus, a unique stationary distribution [Berman and Plemmons, 1994, Theorem
3.23]. The stationary distribution satisfies:
T
0
x
A (D + I)−1 v/eT v x0
0
Px=x ⇔
=
.
γ
γ
e(D + I)−1
0
0
Let P̄ = AT (D + I)−1 . The censored distribution x = x0 /eT x0 is a normalized
solution of the linear system:
0
(I − P̄ )x = v.
0
(5.3)
Note that cT = eT − eT P̄ > 0, and so all columns are substochastic. This means that
0
all of the nodes “leak probability” in a semi-formal sense. Scaling P̄ by 1−c1max > 1
adjusts the probabilities such that there is at least one column that is stochastic.
0
0
Consequently, we can write P̄ = αP̄ where α = (1 − cmax ) and P̄ = 1−c1max P̄ . By
substituting this form into (5.3), we have that x is the normalized solution of a
pseudo-PageRank problem where α = 1−c1max . Assuming that A is an unweighted
dmax
graph, then α = dmax
+1 .
This idea frequently reappears; for instance, Bini et al. [2010], Lü et al. [2011] and
Schlote et al. [2012] all use it in different contexts.
In a different context, this same type of analysis shows that the Colley matrix
for ranking sports teams is a diagonally perturbed, generalized pseudo-PageRank
29
PAGERANK BEYOND THE WEB
3
2
0
0
0
P̄ =
0
0
0
5
4
6
1
1/3
0
1/3
0
0
0
0
0
0
0
1/2
0
0
1/4
1/4
0
1/4
0
0
0
0
0
0
1/2
0
0
0
0
1/2
0
1
1/3
1/2
c=
1/4
1/2
1/2
c
(a) A directed graph with a
censored node c
(b) The substochastic matrix and correction vector for the Markov
chain construction after node c is censored.
Fig. 5.1. In this teleportation construction we add a node c to the original graph as in subfigure
(a). The probability of transitioning to c, or teleporting after we censor node c, then depends on the
degree of each node. A random surfer teleports from node 2 with probability 1/3 and from node 4
with probability 1/4. This construction yields a substochastic matrix P̄ where all the elements of
the correction vector c are positive. This means it’s equivalent to a PageRank construction with
α = 1 − min c, or α = 3/4 for this problem.
system [Colley, 2002; Langville and Meyer, 2012]. Let the symmetric, weighted graph
G represent the network of times team i played team j. And let f be a vector of the
accumulated scores differences over all of those games. It could have negative entries,
rendering it outside of our traditional framework, however, as we saw in Section 5.3,
this is a technical detail that is avoidable. The vector of Colley scores r is the solution
of:
(D + 2I − A)r = f .
Let y = (D + 2I)−1 r. Then,
(I − αP̄)y = f
dmax
where α = dmax
+2 . This analysis establishes a formal relationship between Markov
style ranking metrics [Langville and Meyer, 2012] and the least-squares style ranking
metrics employed by Colley. It also enables us to use fast PageRank solvers for these
Colley systems.
6. Discussion & a positive outlook on PageRank’s wide usage. PageRank
has gone from being used to evaluate the importance of web pages to a much broader
set of applications. The method is easy to understand, is robust and reliable, and
converges quickly. Most applications solve PageRank problems of only a modest size,
with fewer than 100,000,000 vertices; this regime permits a much wider variety of
algorithmic possibilities than those that must only work on the web.
We have avoided the discussion of PageRank algorithms entirely in this manuscript
because, by and large, simple iterations suffice for fast convergence in this regime.
Values of α tend to be less than 0.99, which requires fewer than 2000 iterations to
converge to machine precision. Nevertheless, there is ample opportunity to accelerate
PageRank computations in this regime as there are ideas that involve computing
multiple PageRank vectors for a single task. One example is PerturbationRank [Du
et al., 2008], which uses the perturbation induced in a PageRank vector by removing a
node to compute a new ranking of all nodes. Thus, innovations in PageRank algorithms
are still relevant, but must be made within the context of these small-scale uses.
There are also a great number of PageRank-like ideas outside of our specific canon.
For instance, none of the following models fit our PageRank framework:
30
D. F. Gleich
BrowseRank Liu et al. [2008] define a continuous time Markov chain to model a
random surfer that remains on a specified node for some period of time before
transitioning away. This model handles sites like Facebook, where users spend
significant time interacting within a single page.
Voting Boldi et al. [2009a] and Boldi et al. [2011] define a voting model on a social
network inspired by computing Katz or PageRank on a random network where
each node picks a single outlink.
SimRank This problem is another way to use PageRank-like ideas to evaluate similarity between the nodes of a graph (like the IsoRank problem) [Jeh and
Widom, 2002]. SimRank, however, involves solving a linear system on a row
sub-stochastic matrix.
Food webs The food web is a network where species are linked by the feeding
relationships. Allesina and Pascual [2009] point out a few modifications to
PageRank to make it more appropriate. First, they use teleportation to
model a constant loss of nutrients from higher-level species and reinject these
nutrients through primary producers (such as bacteria). Second, they note
that the flow of importance ought to be reversed so that species i points
to species j if i is important for j’s survival. The result is an eigenvector
computation on a fully stochastic matrix.
Opinion dynamics Models of opinion formation on social network posit strikingly
similar dynamics to a PageRank iteration [Friedkin and Johnsen, 1990, 1999].
The essential difference is that a node’s opinion is the average of its in-links,
instead of propagating its value to its out-links. Like SimRank, this results in
a row sub-stochastic iteration.
The details and implications of these models are fascinating, and this manuscript
would double in size if we were to treat them.
In most successful studies, PageRank is used as a type of baseline measure. Its
widespread success above extremely simple baselines suggests that its modified random
walk is a generally useful alternative worth investigating. In this sense, it resembles a
form of regularization. And this is how we feel that PageRank should be used. Note
that studies must use care when determining the type of PageRank construction –
weighted, reverse, Dirichlet, etc. – as this can make a large difference in the quality
of the results. Consider, for instance, the use of weighted PageRank in Jiang [2009].
In their application, they wanted to model where people move, and it makes good
sense that businesses would locate in places with many connections and therefore, that
people would preferentially move to these same locations. Given the generality of the
idea and its intuitive appeal, we anticipate continued widespread use of the PageRank
idea over the next 20 years in new and exciting applications as network data continues
to proliferate.
REFERENCES
D. Aldous and J. A. Fill. Reversible Markov chains and random walks on graphs. 2002. Unfinished
monograph, recompiled 2014, available at http://www.stat.berkeley.edu/$\sim$aldous/RWG/
book.html. Cited on page 8.
S. Allesina and M. Pascual. Googling food webs: Can an eigenvector measure species’ importance
for coextinctions? PLoS Comput Biol, 5 (9), p. e1000494, 2009. doi:10.1371/journal.pcbi.
1000494. Cited on page 30.
R. Andersen, F. Chung, and K. Lang. Local graph partitioning using PageRank vectors. In
Proceedings of the 47th Annual IEEE Symposium on Foundations of Computer Science. 2006.
Cited on pages 5 and 8.
PAGERANK BEYOND THE WEB
31
W. N. J. Anderson and T. D. Morley. Eigenvalues of the Laplacian of a graph. Linear and
Multilinear Algebra, 18 (2), pp. 141–145, 1985. doi:10.1080/03081088508817681. Cited on
page 25.
A. Arasu, J. Novak, A. Tomkins, and J. Tomlin. PageRank computation and the structure of the
web: Experiments and algorithms. In Proceedings of the 11th international conference on the
World Wide Web. 2002. Poster session. Cited on page 22.
K. Avrachenkov, B. Ribeiro, and D. Towsley. Improving random walk estimation accuracy with
uniform restarts. In Algorithms and Models for the Web-Graph, pp. 98–109. Springer Berlin
Heidelberg, 2010. doi:10.1007/978-3-642-18009-5_10. Cited on page 8.
L. Backstrom and J. Leskovec. Supervised random walks: Predicting and recommending links in
social networks. In Proceedings of the Fourth ACM International Conference on Web Search
and Data Mining, pp. 635–644. 2011. doi:10.1145/1935826.1935914. Cited on page 21.
R. Baeza-Yates, P. Boldi, and C. Castillo. Generalizing PageRank: Damping functions for
link-based ranking algorithms. In Proceedings of the 29th annual international ACM SIGIR
conference on Research and development in information retrieval (SIGIR2006), pp. 308–315.
2006. doi:10.1145/1148170.1148225. Cited on page 24.
B. Bahmani, A. Chowdhury, and A. Goel. Fast incremental and personalized pagerank . Proc.
VLDB Endow., 4 (3), pp. 173–184, 2010. Cited on page 21.
A. Balmin, V. Hristidis, and Y. Papakonstantinou. ObjectRank: authority-based keyword search
in databases. In Proceedings of the Thirtieth international conference on Very large data bases
- Volume 30, pp. 564–575. 2004. Cited on page 19.
Z. Bar-Yossef and L.-T. Mashiach. Local approximation of PageRank and reverse PageRank . In
CIKM ’08: Proceeding of the 17th ACM conference on Information and knowledge management,
pp. 279–288. 2008. doi:10.1145/1458082.1458122. Cited on pages 7, 19, 21, and 22.
M. Bayati, D. F. Gleich, A. Saberi, and Y. Wang. Message-passing algorithms for sparse network
alignment. ACM Trans. Knowl. Discov. Data, 7 (1), pp. 3:1–3:31, 2013. doi:10.1145/2435209.
2435212. Cited on page 11.
L. Becchetti, C. Castillo, D. Donato, R. Baeza-Yates, and S. Leonardi. Link analysis for
web spam detection. ACM Trans. Web, 2 (1), pp. 1–42, 2008. doi:10.1145/1326561.1326563.
Cited on page 22.
P. Berkhin. A survey on PageRank computing. Internet Mathematics, 2 (1), pp. 73–120, 2005.
Cited on page 4.
A. Berman and R. J. Plemmons. Nonnegative Matrices in the Mathematical Sciences, SIAM, 1994.
Cited on page 28.
D. A. Bini, G. M. D. Corso, and F. Romani. A combined approach for evaluating papers,
authors and scientific journals. Journal of Computational and Applied Mathematics, 234 (11),
pp. 3104 – 3121, 2010. Numerical Linear Algebra, Internet and Large Scale Applications.
doi:10.1016/j.cam.2010.02.003. Cited on page 28.
D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent Dirichlet allocation. Journal of Machine Learning
Research, 3, pp. 993–1022, 2003. Cited on page 15.
V. D. Blondel, A. Gajardo, M. Heymans, P. Senellart, and P. V. Dooren. A measure of
similarity between graph vertices: Applications to synonym extraction and web searching. SIAM
Review, 46 (4), pp. 647–666, 2004. doi:10.1137/S0036144502415960. Cited on pages 11 and 21.
P. Boldi. TotalRank: Ranking without damping. In Poster Proceedings of the 14th international
conference on the World Wide Web (WWW2005), pp. 898–899. 2005. Cited on page 24.
P. Boldi, F. Bonchi, C. Castillo, D. Donato, A. Gionis, and S. Vigna. The query-flow graph:
Model and applications. In Proceedings of the 17th ACM Conference on Information and
Knowledge Management, pp. 609–618. 2008. doi:10.1145/1458082.1458163. Cited on page 20.
P. Boldi, F. Bonchi, C. Castillo, and S. Vigna. Voting in social networks. In CIKM ’09:
Proceeding of the 18th ACM conference on Information and knowledge management, pp.
777–786. 2009a. doi:10.1145/1645953.1646052. Cited on page 30.
———. Viscous democracy for social networks. Commun. ACM, 54 (6), pp. 129–137, 2011.
doi:10.1145/1953122.1953154. Cited on page 30.
P. Boldi, R. Posenato, M. Santini, and S. Vigna. Traps and pitfalls of topic-biased PageRank .
In WAW2006, Fourth International Workshop on Algorithms and Models for the Web-Graph,
pp. 107–116. 2007. doi:10.1007/978-3-540-78808-9_10. Cited on pages 4 and 7.
P. Boldi, M. Santini, and S. Vigna. PageRank as a function of the damping factor . In Proceedings
of the 14th international conference on the World Wide Web (WWW2005). 2005. doi:
10.1145/1060745.1060827. Cited on pages 24, 25, and 27.
———. PageRank: Functional dependencies. ACM Trans. Inf. Syst., 27 (4), pp. 1–23, 2009b.
doi:10.1145/1629096.1629097. Cited on page 24.
J. Bollen, M. A. Rodriquez, and H. Van de Sompel. Journal status. Scientometrics, 69 (3), pp.
32
D. F. Gleich
669–687, 2006. doi:10.1007/s11192-006-0176-z. Cited on page 16.
S. Brin and L. Page. The anatomy of a large-scale hypertextual web search engine. Comput. Netw.
ISDN Syst., 30 (1-7), pp. 107–117, 1998. doi:10.1016/S0169-7552(98)00110-X. Cited on page
1.
T. Callaghan, P. J. Mucha, and M. A. Porter. Random walker ranking for NCAA division I-A
football. The American Mathematical Monthly, 114 (9), pp. 761–777, 2007. Cited on page 14.
P. Chen, H. Xie, S. Maslov, and S. Redner. Finding scientific gems with Google’s PageRank
algorithm. Journal of Informetrics, 1 (1), pp. 8–15, 2007. doi:10.1016/j.joi.2006.06.001.
Cited on page 17.
A. Chepelianskii. Towards physical laws for software architecture. arXiv, cs.SE, p. 1003.5455,
2010. Cited on page 13.
F. Chung. Laplacians and the Cheeger inequality for directed graphs. Annals of Combinatorics,
9 (1), pp. 1–19, 2005. 10.1007/s00026-005-0237-z. doi:10.1007/s00026-005-0237-z. Cited on
page 18.
———. The heat kernel as the PageRank of a graph. Proceedings of the National Academy of
Sciences, 104 (50), pp. 19735–19740, 2007. doi:10.1073/pnas.0708838104. Cited on page 24.
F. Chung, A. Tsiatas, and W. Xu. Dirichlet pagerank and trust-based ranking algorithms. In
Algorithms and Models for the Web Graph, pp. 103–114. Springer Berlin Heidelberg, 2011.
doi:10.1007/978-3-642-21286-4_9. Cited on page 7.
F. R. L. Chung. Spectral Graph Theory, American Mathematical Society, 1992. Cited on page 26.
W. N. Colley. Colley’s bias free college football ranking method: The Colley matrix explained.
Technical report, Princeton University, 2002. Cited on page 29.
P. G. Constantine and D. F. Gleich. Random alpha PageRank . Internet Mathematics, 6 (2), pp.
189–236, 2010. doi:10.1080/15427951.2009.10129185. Cited on pages 23, 24, 25, and 27.
J. J. Crofts and D. J. Higham. Googling the brain: Discovering hierarchical and asymmetric
network structures, with applications in neuroscience. Internet Mathematics, 7, pp. 233–254,
2011. doi:10.1080/15427951.2011.604284. Cited on page 12.
M. Cutts. Q&A march 27, 2006 . Matt Cutts: Gadgets, Google, and SEO. Available online
http://www.mattcutts.com/blog/q-a-thread-march-27-2006/, 2006. Cited on page 22.
———. PageRank scupting. Matt Cutts: Gadgets, Google, and SEO blog. Available online
http://www.mattcutts.com/blog/pagerank-sculpting/, 2009. Cited on page 22.
T. A. Davis and Y. Hu. The University of Florida sparse matrix collection. ACM Transactions on
Mathematical Software, 2010. To appear. Cited on page 22.
G. Del Corso, A. Gullı́, and F. Romani. Fast PageRank computation via a sparse linear system
(extended abstract). In Algorithms and Models for the Web-Graph, pp. 118–130. Springer Berlin
Heidelberg, 2004. doi:10.1007/978-3-540-30216-2_10. Cited on page 4.
G. M. Del Corso, A. Gullı́, and F. Romani. Fast PageRank computation via a sparse linear
system. Internet Mathematics, 2 (3), pp. 251–273, 2005. Cited on page 4.
I. S. Dhillon. Co-clustering documents and words using bipartite spectral graph partitioning. In
Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery
and data mining, pp. 269–274. 2001. doi:10.1145/502512.502550. Cited on page 17.
Y. Du, J. Leung, and Y. Shi. PerturbationRank: A non-monotone ranking algorithm. Technical
report, University of Michigan, 2008. Cited on page 29.
N. Eiron, K. S. McCurley, and J. A. Tomlin. Ranking the web frontier . In Proceedings of
the 13th international conference on the World Wide Web (WWW2004), pp. 309–318. 2004.
doi:10.1145/988672.988714. Cited on pages 22 and 28.
Y.-H. Eom, P. Aragón, D. Laniado, A. Kaltenbrunner, S. Vigna, and D. L. Shepelyansky.
Interactions of cultures and top people of Wikipedia from ranking of 24 language editions. arXiv,
cs.SI, p. 1405.7183, 2014. Cited on page 23.
E. Estrada. Characterization of 3d molecular structure. Chemical Physics Letters, 319 (5-6), pp.
713–718, 2000. doi:10.1016/S0009-2614(00)00158-5. Cited on page 24.
E. Estrada and D. J. Higham. Network properties revealed through matrix functions. SIAM
Review, 52 (4), pp. 696–714, 2010. doi:10.1137/090761070. Cited on page 24.
E. Estrada, D. J. Higham, and N. Hatano. Communicability and multipartite structures in
complex networks at negative absolute temperatures. Phys. Rev. E, 78, p. 026102, 2008.
doi:10.1103/PhysRevE.78.026102. Cited on page 12.
A. Farahat, T. LoFaro, J. C. Miller, G. Rae, and L. A. Ward. Authority rankings from HITS,
PageRank, and SALSA: Existence, uniqueness, and effect of initialization. SIAM Journal on
Scientific Computing, 27 (4), pp. 1181–1201, 2006. doi:10.1137/S1064827502412875. Cited on
page 24.
D. Fiala, F. Rousselot, and K. Jeek. PageRank for bibliographic networks. Scientometrics, 76 (1),
pp. 135–158, 2008. doi:10.1007/s11192-007-1908-4. Cited on page 17.
PAGERANK BEYOND THE WEB
33
M. Fiedler. Algebraic connectivity of graphs. Czechoslovak Mathematical Journal, 23 (98), pp.
298–305, 1973. Cited on page 25.
D. Fogaras. Where to start browsing the web? In Innovative Internet Community Systems, pp.
65–79. Springer Berlin Heidelberg, 2003. doi:10.1007/978-3-540-39884-4_6. Cited on pages
7 and 22.
K. M. Frahm, A. D. Chepelianskii, and D. L. Shepelyansky. PageRank of integers. Journal of
Physics A: Mathematical and Theoretical, 45 (40), p. 405101, 2012. doi:10.1088/1751-8113/
45/40/405101. Cited on page 14.
V. Freschi. Protein function prediction from interaction networks using a random walk ranking
algorithm. In Proceedings of the 7th IEEE International Conference on Bioinformatics and
Bioengineering (BIBE 2007), pp. 42–48. 2007. doi:10.1109/BIBE.2007.4375543. Cited on page
10.
N. E. Friedkin and E. C. Johnsen. Social influence and opinions. The Journal of Mathematical
Sociology, 15 (3-4), pp. 193–206, 1990. doi:10.1080/0022250X.1990.9990069. Cited on page
30.
———. Social influence networks and opinion change. Advances in Group Processes, 16 (1), pp.
1–29, 1999. Cited on page 30.
E. Garfield. Citation indexes for science: A new dimension in documentation through association
of ideas. Science, 122 (3159), pp. 108–111, 1955. doi:10.1126/science.122.3159.108. Cited
on page 16.
E. Garfield and I. H. Sher. New factors in the evaluation of scientific literature through citation
indexing. American Documentation, 14 (3), pp. 195–201, 1963. doi:10.1002/asi.5090140304.
Cited on page 16.
D. F. Gleich, P. Glynn, G. H. Golub, and C. Greif. Three results on the PageRank vector:
eigenstructure, sensitivity, and the derivative. In Web Information Retrieval and Linear Algebra
Algorithms. 2007. Cited on page 27.
D. F. Gleich, A. P. Gray, C. Greif, and T. Lau. An inner-outer iteration for PageRank . SIAM
Journal of Scientific Computing, 32 (1), pp. 349–371, 2010. doi:10.1137/080727397. Cited on
page 11.
D. F. Gleich and M. M. Mahoney. Algorithmic anti-differentiation: A case study with min-cuts,
spectral, and flow. In Proceedings of the International Conference on Machine Learning (ICML).
2014. Cited on page 8.
D. F. Gleich and M. Polito. Approximating personalized PageRank with minimal use of webgraph
data. Internet Mathematics, 3 (3), pp. 257–294, 2007. doi:10.1080/15427951.2006.10129128.
Cited on page 22.
D. F. Gleich and R. A. Rossi. A dynamical system for PageRank with time-dependent teleportation.
Internet Mathematics, 10 (1–2), pp. 188–217, 2014. doi:10.1080/15427951.2013.814092. Cited
on page 27.
G. Golub and C. Greif. An Arnoldi-type algorithm for computing PageRank . BIT Numerical
Mathematics, 46 (4), pp. 759–771, 2006. doi:10.1007/s10543-006-0091-y. Cited on page 27.
M. Gori and A. Pucci. ItemRank: a random-walk based scoring algorithm for recommender engines.
In IJCAI’07: Proceedings of the 20th international joint conference on Artifical intelligence,
pp. 2766–2771. 2007. Cited on page 20.
A. Y. Govan, C. D. Meyer, and R. Albright. Generalizing Google’s PageRank to rank national
football league teams. In SAS Global Forum 2008. 2008. Cited on pages 14 and 15.
Z. Gyöngyi, H. Garcia-Molina, and J. Pedersen. Combating web spam with TrustRank . In
Proceedings of the 30th International Very Large Database Conference. 2004. Cited on pages 7
and 22.
T. H. Haveliwala. Topic-sensitive PageRank . In Proceedings of the 11th international conference
on the World Wide Web. 2002. Cited on page 19.
D. J. Higham. Google PageRank as mean playing time for pinball on the reverse web. Applied
Mathematics Letters, 18 (12), pp. 1359 – 1362, 2005. doi:10.1016/j.aml.2005.02.020. Cited
on page 1.
R. A. Horn and S. Serra-Capizzano. A general setting for the parametric Google matrix . Internet
Mathematics, 3 (4), pp. 385–411, 2007. Cited on page 27.
A. Hotho, R. Jäschke, C. Schmitz, and G. Stumme. Information retrieval in folksonomies: Search
and ranking. In Proceedings of the 3rd European Semantic Web Conference, pp. 411–426. 2006.
doi:10.1007/11762256_31. Cited on pages 16 and 19.
B. A. Huberman, P. L. T. Pirolli, J. E. Pitkow, and R. M. Lukose. Strong regularities in World
Wide Web surfing. Science, 280 (5360), pp. 95–97, 1998. doi:10.1126/science.280.5360.95.
Cited on page 24.
A. Jain and P. Pantel. FactRank: Random walks on a web of facts. In Proceedings of the 23rd
34
D. F. Gleich
International Conference on Computational Linguistics, pp. 501–509. 2010. Cited on page 18.
A. Java. Twitter social network analysis. UMBC ebquity blog, http://ebiquity.umbc.edu/
blogger/2007/04/19/twitter-social-network-analysis/, 2007. Cited on page 21.
A. Java, P. Kolari, T. Finin, , and T. Oates. Modeling the spread of influence on the blogosphere.
Technical Report UMBC TR-CS-06-03, University of Maryland, Baltimore, 2006. Cited on page
21.
G. Jeh and J. Widom. SimRank: a measure of structural-context similarity. In KDD ’02: Proceedings
of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining,
pp. 538–543. 2002. doi:10.1145/775047.775126. Cited on pages 11 and 30.
K. Jezek, D. Fiala, and J. Steinberger. Exploration and evaluation of citation networks. In
Proceedings of the 12th International Conference on Electronic Publishing, pp. 351–362. 2008.
Cited on page 17.
B. Jiang. Ranking spaces for predicting human movement in an urban environment. Int. J. Geogr.
Inf. Sci., 23 (7), pp. 823–837, 2009. doi:10.1080/13658810802022822. Cited on pages 8, 13,
and 30.
B. Jiang, S. Zhao, and J. Yin. Self-organized natural roads for predicting traffic flow: a sensitivity
study. Journal of Statistical Mechanics: Theory and Experiment, 2008 (07), p. P07008, 2008.
doi:10.1088/1742-5468/2008/07/P07008. Cited on page 13.
B.-B. Jiang, J.-G. Wang, J.-F. Xiao, and Y. Wang. Gene prioritization for type 2 diabetes in
tissue-specic protein interaction networks. In Third International Symposium on Optimization
and Systems Biology. 2009. Cited on page 10.
Y. Jing and S. Baluja. VisualRank: Applying pagerank to large-scale image search. Pattern
Analysis and Machine Intelligence, IEEE Transactions on, 30 (11), pp. 1877–1890, 2008. doi:
10.1109/TPAMI.2008.121. Cited on page 23.
M. Jockers. Computing and visualizing the 19th-century literary genome. In Digital Humanities.
2012. Cited on page 15.
S. Kamvar. Numerical Algorithms for Personalized Search in Self-organizing Information Networks,
Princeton University Press, 2010. Cited on page 10.
S. Kamvar, T. Haveliwala, C. Manning, and G. Golub. Exploiting the block structure of the web
for computing PageRank . Technical Report 2003-17, Stanford InfoLab, 2003. Cited on pages 10
and 22.
L. Katz. A new status index derived from sociometric analysis. Psychometrika, 18 (1), pp. 39–43,
1953. doi:10.1007/BF02289026. Cited on pages 12, 21, and 25.
J. P. Keener. The Perron-Frobenius theorem and the ranking of football teams. SIAM Review,
35 (1), pp. 80–93, 1993. doi:10.1137/1035004. Cited on page 14.
D. Kempe, J. Kleinberg, and E. Tardos. Maximizing the spread of influence through a social
network . In Proceedings of the ninth ACM SIGKDD international conference on Knowledge
discovery and data mining, pp. 137–146. 2003. doi:10.1145/956750.956769. Cited on page 21.
———. Influential nodes in a diffusion model for social networks. In Proceedings of the 32nd
International Conference on Automata, Languages and Programming, pp. 1127–1138. 2005.
doi:10.1007/11523468_91. Cited on page 21.
M. Kim, R. Sumbaly, and S. Shah. Root cause detection in a service-oriented architecture. In
Proceedings of the ACM SIGMETRICS/international conference on Measurement and modeling
of computer systems, pp. 93–104. 2013. doi:10.1145/2465529.2465753. Cited on page 12.
J. M. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM, 46 (5),
pp. 604–632, 1999. Cited on page 21.
T. G. Kolda and M. J. Procopio. Generalized BadRank with graduated trust. Technical Report
SAND2009-6670, Sandia National Laboratories, 2009. Cited on page 22.
G. Kollias, E. Gallopoulos, and A. Grama. Surfing the network for ranking by multidamping.
Knowledge and Data Engineering, IEEE Transactions on, PP (99), pp. 1–1, 2013. doi:
10.1109/TKDE.2013.15. Cited on page 24.
G. Kollias, S. Mohammadi, and A. Grama. Network similarity decomposition (NSD): A fast and
scalable approach to network alignment. Knowledge and Data Engineering, IEEE Transactions
on, PP (99), p. 1, 2011. doi:10.1109/TKDE.2011.174. Cited on page 11.
R. I. Kondor and J. D. Lafferty. Diffusion kernels on graphs and other discrete input spaces.
In Proceedings of the Nineteenth International Conference on Machine Learning, pp. 315–322.
2002. Cited on page 24.
E.-M. Kontopoulou, M. Predari, T. Kostakis, and E. Gallopoulos. Graph and matrix metrics
to analyze ergodic literature for children. In Proceedings of the 23rd ACM conference on
Hypertext and social media, pp. 133–142. 2012. doi:10.1145/2309996.2310018. Cited on page
15.
D. Koschützki, K. A. Lehmann, L. Peeters, S. Richter, D. Tenfelde-Podehl, , and O. Zlo-
PAGERANK BEYOND THE WEB
35
towski. Centrality indicies. In Network Analysis: Methodological Foundations, chapter 3, pp.
16–61. Springer, 2005. doi:10.1007/b106453. Cited on page 1.
J. Kunegis and A. Lommatzsch. Learning spectral graph transformations for link prediction. In
Proceedings of the 26th Annual International Conference on Machine Learning, pp. 561–568.
2009. doi:10.1145/1553374.1553447. Cited on page 24.
H. Kwak, C. Lee, H. Park, and S. Moon. What is Twitter, a social network or a news media? In
WWW ’10: Proceedings of the 19th international conference on World wide web, pp. 591–600.
2010. doi:10.1145/1772690.1772751. Cited on page 21.
A. N. Langville and C. D. Meyer. Deeper inside PageRank . Internet Mathematics, 1 (3), pp.
335–380, 2004. Cited on page 7.
———. Google’s PageRank and Beyond: The Science of Search Engine Rankings, Princeton
University Press, 2006. Cited on pages 1 and 2.
———. Who’s #1? The Science of Rating and Ranking, Princeton University Press, 2012. Cited
on pages 14 and 29.
C. P. Lee, G. H. Golub, and S. A. Zenios. A two-stage algorithm for computing PageRank and
multistage generalizations. Internet Mathematics, 4 (4), pp. 299–327, 2007. Cited on page 28.
D. Liben-Nowell and J. Kleinberg. The link-prediction problem for social networks. Journal
of the American Society of Information Science and Technology, 58 (7), pp. 1019–1031, 2006.
doi:10.1002/asi.20591. Cited on page 20.
X. Liu, J. Bollen, M. L. Nelson, and H. Van de Sompel. Co-authorship networks in the
digital library research community. Inf. Process. Manage., 41 (6), pp. 1462–1480, 2005. doi:
10.1016/j.ipm.2005.03.012. Cited on page 17.
Y. Liu, B. Gao, T.-Y. Liu, Y. Zhang, Z. Ma, S. He, and H. Li. BrowseRank: letting web users
vote for page importance. In SIGIR ’08: Proceedings of the 31st annual international ACM
SIGIR conference on Research and development in information retrieval, pp. 451–458. 2008.
doi:10.1145/1390334.1390412. Cited on page 30.
L. Lü, Y.-C. Zhang, C. H. Yeung, and T. Zhou. Leaders in social networks, the Delicious case.
PLoS ONE, 6 (6), p. e21202, 2011. doi:10.1371/journal.pone.0021202. Cited on page 28.
N. Ma, J. Guan, and Y. Zhao. Bringing PageRank to the citation analysis. Information Processing & Management, 44 (2), pp. 800–810, 2008. ¡ce:title¿Evaluating Exploratory Search
Systems¡/ce:title¿ ¡ce:title¿Digital Libraries in the Context of Users Broader Activities¡/ce:title¿.
doi:http://dx.doi.org/10.1016/j.ipm.2007.06.006. Cited on page 17.
M. W. Mahoney, L. Orecchia, and N. K. Vishnoi. A local spectral method for graphs: With
applications to improving graph partitions and exploring data graphs locally. Journal of Machine
Learning Research, 13, p. 23392365, 2012. Cited on pages 25 and 27.
X. Meng. Computing BookRank via social cataloging. 2009. http://cads.stanford.edu/projects/
presentations/2009visit/bookrank.pdf. Cited on page 16.
C. D. Meyer. Matrix analysis and applied linear algebra, Society for Industrial and Applied
Mathematics, Philadelphia, PA, USA, 2000. Cited on page 25.
J. C. Miller, G. Rae, F. Schaefer, L. A. Ward, T. LoFaro, and A. Farahat. Modifications
of Kleinberg’s HITS algorithm using matrix exponentiation and web log records. In SIGIR
’01: Proceedings of the 24th annual international ACM SIGIR conference on Research and
development in information retrieval, pp. 444–445. 2001. doi:10.1145/383952.384086. Cited
on page 24.
B. L. Mooney, L. R. Corrales, and A. E. Clark. Molecularnetworks: An integrated graph
theoretic and data mining tool to explore solvent organization in molecular simulation. Journal
of Computational Chemistry, 33 (8), pp. 853–860, 2012. doi:10.1002/jcc.22917. Cited on
page 10.
J. L. Morrison, R. Breitling, D. J. Higham, and D. R. Gilbert. GeneRank: using search engine
technology for the analysis of microarray experiments. BMC Bioinformatics, 6 (1), p. 233, 2005.
doi:10.1186/1471-2105-6-233. Cited on pages 1 and 10.
M. A. Najork, H. Zaragoza, and M. J. Taylor. HITS on the web: How does it compare?
In Proceedings of the 30th annual international ACM SIGIR conference on Research and
Development in information retrieval (SIGIR2007), pp. 471–478. 2007. doi:10.1145/1277741.
1277823. Cited on page 21.
Z. Nie, Y. Zhang, J.-R. Wen, and W.-Y. Ma. Object-level ranking: Bringing order to web objects.
In Proceedings of the 14th International Conference on World Wide Web, pp. 567–574. 2005.
doi:10.1145/1060745.1060828. Cited on page 18.
G. Osipenko. Dynamical systems, graphs, and algorithms, Springer, 2007. doi:10.1007/
3-540-35593-6. Cited on page 14.
L. Page, S. Brin, R. Motwani, and T. Winograd. The PageRank citation ranking: Bringing order
to the web. Technical Report 1999-66, Stanford University, 1999. Cited on pages 1 and 3.
36
D. F. Gleich
J.-Y. Pan, H.-J. Yang, C. Faloutsos, and P. Duygulu. Automatic multimedia cross-modal
correlation discovery. In KDD ’04: Proceedings of the tenth ACM SIGKDD international
conference on Knowledge discovery and data mining, pp. 653–658. 2004. doi:10.1145/1014052.
1014135. Cited on pages 1 and 18.
G. Pinski and F. Narin. Citation influence for journal aggregates of scientific publications: Theory,
with application to the literature of physics. Information Processing & Management, 12 (5), pp.
297–312, 1976. doi:10.1016/0306-4573(76)90048-0. Cited on pages 16 and 25.
A. Pothen, H. D. Simon, and K.-P. Liou. Partitioning sparse matrices with eigenvectors of graphs.
SIAM J. Matrix Anal. Appl., 11, pp. 430–452, 1990. doi:10.1137/0611030. Cited on pages 25
and 26.
F. Radicchi. Who is the best player ever? a complex network analysis of the history of professional
tennis. PLoS ONE, 6 (2), p. e17249, 2011. doi:10.1371/journal.pone.0017249. Cited on page
15.
S. E. Schaeffer. Graph clustering. Computer Science Review, 1 (1), pp. 27–64, 2007. doi:
10.1016/j.cosrev.2007.05.001. Cited on page 26.
A. Schlote, E. Crisostomi, S. Kirkland, and R. Shorten. Traffic modelling framework for
electric vehicles. International Journal of Control, 85 (7), pp. 880–897, 2012. doi:10.1080/
00207179.2012.668716. Cited on pages 13 and 28.
S. Serra-Capizzano. Jordan canonical form of the Google matrix: A potential contribution to
the PageRank computation. SIAM Journal on Matrix Analysis and Applications, 27 (2), pp.
305–312, 2005. doi:10.1137/S0895479804441407. Cited on page 24.
D. L. Shepelyansky and O. V. Zhirov. Google matrix, dynamical attractors, and ulam networks.
Phys. Rev. E, 81 (3), p. 036213, 2010. doi:10.1103/PhysRevE.81.036213. Cited on page 14.
R. Singh, J. Xu, and B. Berger. Pairwise global alignment of protein interaction networks by
matching neighborhood topology. In Proceedings of the 11th Annual International Conference
on Research in Computational Molecular Biology (RECOMB), pp. 16–31. 2007. doi:10.1007/
978-3-540-71681-5_2. Cited on page 11.
M. Sobek. PR0 - Google’s PageRank 0 penalty. Online., 2003. http://pr.efactory.de/e-pr0.
shtml. Accessed 2013-09-19. Cited on page 22.
Y. Song, D. Zhou, and L.-w. He. Query suggestion by constructing term-transition graphs. In
Proceedings of the Fifth ACM International Conference on Web Search and Data Mining, pp.
353–362. 2012. doi:10.1145/2124295.2124339. Cited on page 20.
O. Sporns. Networks of the Brain, The MIT Press, 2011. Cited on page 11.
W. J. Stewart. Introduction to the Numerical Solution of Markov Chains, Princeton University
Press, 1994. Cited on page 8.
J. J. Sylvester. Chemistry and algebra. Nature, 17, p. 284, 1878. doi:10.1038/017284a0. Cited
on page 10.
J. A. Tomlin. A new paradigm for ranking pages on the world wide web. In Proceedings of the 12th
International Conference on World Wide Web, pp. 350–355. 2003. doi:10.1145/775152.775202.
Cited on page 28.
H. Tong, C. Faloutsos, and J.-Y. Pan. Fast random walk with restart and its applications. In
ICDM ’06: Proceedings of the Sixth International Conference on Data Mining, pp. 613–622.
2006. doi:10.1109/ICDM.2006.70. Cited on page 19.
S. Vigna. TruRank: Taking PageRank to the limit. In WWW ’05: Special interest tracks
and posters of the 14th international conference on World Wide Web, pp. 976–977. 2005.
doi:10.1145/1062745.1062826. Cited on page 25.
———. Spectral ranking. arXiv, cs.IR, p. 0912.0238, 2009. Cited on pages 1, 21, 25, and 27.
K. Voevodski, S.-H. Teng, and Y. Xia. Spectral affinity in protein networks. BMC Systems
Biology, 3 (1), p. 112, 2009. doi:10.1186/1752-0509-3-112. Cited on page 11.
D. Walker, H. Xie, K.-K. Yan, and S. Maslov. Ranking scientific publications using a model of
network traffic. Journal of Statistical Mechanics: Theory and Experiment, 2007 (06), p. P06010,
2007. doi:10.1088/1742-5468/2007/06/P06010. Cited on page 16.
W. Y. Wang, K. Mazaitis, and W. W. Cohen. Programming with personalized PageRank: A
locally groundable first-order probabilistic logic. In Proceedings of the 22Nd ACM International
Conference on Conference on Information and Knowledge Management, pp. 2129–2138. 2013.
doi:10.1145/2505515.2505573. Cited on page 20.
J. Weng, E.-P. Lim, J. Jiang, and Q. He. TwitterRank: finding topic-sensitive influential twitterers.
In WSDM ’10: Proceedings of the third ACM international conference on Web search and data
mining, pp. 261–270. 2010. doi:10.1145/1718487.1718520. Cited on page 21.
J. D. West, T. C. Bergstrom, and C. T. Bergstrom. The Eigenfactor metrics: A network
approach to assessing scholarly journals. College & Research Libraries, 71 (3), pp. 236–244,
2010. arXiv:http://crl.acrl.org/content/71/3/236.full.pdf+html. Cited on page 16.
PAGERANK BEYOND THE WEB
37
C. Winter, G. Kristiansen, S. Kersting, J. Roy, D. Aust, T. Knösel, P. Rmmele, B. Jahnke,
V. Hentrich, F. Rückert, M. Niedergethmann, W. Weichert, M. Bahra, H. J. Schlitt,
U. Settmacher, H. Friess, M. Büchler, H.-D. Saeger, M. Schroeder, C. Pilarsky, and
R. Grtzmann. Google goes cancer: Improving outcome prediction for cancer patients by
network-based ranking of marker genes. PLoS Comput Biol, 8 (5), p. e1002511, 2012. doi:
10.1371/journal.pcbi.1002511. Cited on page 10.
A. D. Wissner-Gross. Preparation of topical reading lists from the link structure of Wikipedia. In
ICALT ’06: Proceedings of the Sixth IEEE International Conference on Advanced Learning
Technologies, pp. 825–829. 2006. Cited on page 23.
W. Xing and A. Ghorbani. Weighted PageRank algorithm. In Communication Networks and
Services Research, 2004. Proceedings. Second Annual Conference on, pp. 305–314. 2004. doi:
10.1109/DNSR.2004.1344743. Cited on page 8.
H. Yang, I. King, and M. R. Lyu. DiffusionRank: a possible penicillin for web spamming.
In Proceedings of the 30th annual international ACM SIGIR conference on Research and
development in information retrieval, pp. 431–438. 2007. doi:10.1145/1277741.1277815. Cited
on page 24.
A. O. Zhirov, O. V. Zhirov, and D. L. Shepelyansky. Two-dimensional ranking of Wikipedia
articles. The European Physical Journal B, 77 (4), pp. 523–531, 2010. doi:10.1140/epjb/
e2010-10500-7. Cited on pages 13 and 23.
D. Zhou, O. Bousquet, T. N. Lal, J. Weston, and B. Schölkopf. Learning with local and global
consistency. In NIPS. 2003. Cited on page 18.
D. Zhou, J. Huang, and B. Schölkopf. Learning from labeled and unlabeled data on a directed
graph. In ICML ’05: Proceedings of the 22nd International Conference on Machine Learning,
pp. 1036–1043. 2005. doi:10.1145/1102351.1102482. Cited on page 18.
X.-N. Zuo, R. Ehmke, M. Mennes, D. Imperati, F. X. Castellanos, O. Sporns, and M. P.
Milham. Network centrality in the human functional connectome. Cerebral Cortex, 2011.
doi:10.1093/cercor/bhr269. Cited on page 11.
Acknowledgments. We acknowledge the following individuals for their discussions about this manuscript: Sebastiano Vigna, Amy Langville, Michael Saunders,
Chen Greif, Des Higham, and Stratis Gallopoulos, as well as Kyle Kloster for carefully
reading several early drafts. This work was supported in part by NSF CAREER award
CCF-1149756.
| 5cs.CE
|
A computational perspective of the role of Thalamus in cognition
Nima Dehghani1, 2, a) and Ralf D. Wimmer3
1)
Department of Physics, Massachusetts Institute of Technology
Center for Brains, Minds and Machines (CBMM), Massachusetts Institute of Technology
3)
Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology
2)
arXiv:1803.00997v1 [q-bio.NC] 2 Mar 2018
(Dated: March 5, 2018)
Thalamus has traditionally been considered as only a relay source of cortical inputs, with hierarchically
organized cortical circuits serially transforming thalamic signals to cognitively-relevant representations. Given
the absence of local excitatory connections within the thalamus, the notion of thalamic ‘relay’ seemed like a
reasonable description over the last several decades. Recent advances in experimental approaches and theory
provide a broader perspective on the role of the thalamus in cognitively-relevant cortical computations, and
suggest that only a subset of thalamic circuit motifs fit the relay description. Here, we discuss this perspective
and highlight the potential role for the thalamus in dynamic selection of cortical representations through a
combination of intrinsic thalamic computations and output signals that change cortical network functional
parameters. We suggest that through the contextual modulation of cortical computation, thalamus and
cortex jointly optimize the information/cost tradeoff in an emergent fashion. We emphasize that coordinated
experimental and theoretical efforts will provide a path to understanding the role of the thalamus in cognition,
along with an understanding to augment cognitive capacity in health and disease.
Keywords: Thalamo-cortical system, Recurrent Neural Network, Reservoir Computing, Multi-objective Optimization, Cognitive Computing, Artificial Intelligence
Cortico-centric view of perceptual and cognitive processing
Until recently, cognition [in mammalian, bird and reptilian nervous system] has been viewed as a corticocentric process, with thalamus considered to only play
the mere role of a relay system. This classic view, much
driven by the visual hierarchical model of the mammalian
cortex24 , puts thalamus at the beginning of a feedforward
hierarchy. The transmission of information from thalamus to early sensory cortex (V1 in the visual system for
example), and the gradual increasing complex representations from V2 to MT/IT and eventually prefrontal cortex (PFC), constitute the core of the perceptual representation under the hierarchical model. A recent comparative study of biologically-plausible Convolutional Neural Networks (CNN) and the visual ventral stream, emphasizes on the feature enrichment through the network
of hierarchically interconnected computational modules
(layers in the neural network and areas in the ventral
stream)110 .
The strictly static feedforward model has since morphed to a dynamic hierarchical model due to the discoveries of the role of feedback from higher cortical areas
to lower cortical areas35 . These dynamic hierarchies are
even considered to be favorable for recursive optimizations, where the overall optimization can be achieved by
breaking the problem into smaller ones in order to find
the optimum for each of these smaller problems. However, it is possible for the recursive optimization not to
be confined to just one area of the cortex. It may also be
that these smaller optimizations may be solved differently
a) Electronic
mail: [email protected]
at different cortical areas58 . This view of cortical computation is also paralleled with the growing use of recurrent neural networks (RNN) that can capture the dynamics of single neurons or neural population in a variety of
tasks. As such, RNNs can mimic (a) context-dependent
prefrontal response57 or (b) can reproduce the temporal scaling of neural responses in medial frontal cortex
and caudate nucleus103 . Nonetheless, the main attribute
of perception/cognition remains cortico-centric under the
umbrella of dynamic hierarchical models or RNN embodiment of cortical cognitive functions. Since the computation that is carried by the system should match the computing elements at the appropriate scale17 a mismatch
between these presumed computational systems and the
underlying circuitry becomes vividly apparent. Specifically, let’s note that a) associative cortex (and not just
sensory cortex) receives thalamic input, b) certain thalamic territories receive their primary (driving) input from
the cortex, rather than sensory periphery, some of which
are likely to be highly convergent on a single thalamic
cell level and c) cortex is demarcated by local excitatory
connectivity (while the thalamus is devoid of this feature). These points should prompt us to reconsider this
cortico-centric view of computation and indicates that it
need to be extended to a thalamocortical one. As such,
the thalamic contribution to cognitive computations can
be twofold: First, they can modulate the ongoing cortical
activity and second, they might be pure thalamic computations. The nature of such intrinsic thalamic computations would be determined by both the types of inputs
a thalamic neuron receives as well as how these inputs
interact (both linearly and non-linearly). As we will discuss in the next several sections, the combination of these
features allows the thalamus to compute a variety of signals associated with ongoing cortical states, which can be
2
more easily used to dynamically adjust cortical computations through real-time modulation of functional cortical
architecture. Further, we propose that thalamus may
provide a contextual signal that reorganizes functional
connectivity in frontal cortices in response to contextual
changes. We suggest that the unique cognitive capability of the thalamo-cortical system is tightly bound to
parallel processing and contextual modulation that are
enabled by the diversity of computing nodes (including
thalamic and cortical structures) and complexity of the
computing architecture (Fig. 1). We will start with a
brief overview of the thalamic architecture, followed by
experimental evidence and a computational perspective
of the thalamic role in contextual cognitive computation
.
Reservoir
Computing
P
Physarium
Machine
CNN
C
NN
NN
FPGA
P
Logic
gicc Gates
gi
Gate
at
(D
Genetic
ive Com Logic Gate
rsi pu
ty te
&
Co nod
m e
pl
ex
ity
)
DNA
A
Computing
Computing
ting
tin
ting
g
Cognitive capability
cortical
system
g
ocessin
r
p
l
e
l
l
Para
Figure 1. Cognitive computing morphospace: Morphospace of a few example biological and synthetic computing engines
in a multidimensional layout. The thalamo-cortical system standout as a unique system with high cognitive capacity, massive
parallel processing and extreme diversity of the computing nodes. Other computing systems occupy less desirable domains
of this morphospace. Logic Gates: NAND, NOR; Genetic Logic Gate: Synthetic biology adaptation of logic gate; FPGA:
field-programmable gate array (configurable integrated circuit); CNN : Convolutional Neural Network; Physarium Machine:
programmable amorphous biological computer experimentally implemented in slime mould. Reservoir Computing: A reservoir
of recurrent neurons that dynamically change their activity to nonlinearly map the input to a new space; DNA computing: A
computing paradigm where many different DNA molecules are used to perform large number of logical computations in parallel;
Connection Machine: the first commercial supercomputer designed for problems of Artificial Intelligence (AI) using hardware
enabled parallel processing.
Thalamic architecture: anatomical and functional features
Among all brain structures, the forebrain is likely to be
most related to cognitive capacity, given that forebrain
expansion positively correlates (and perhaps defines) cognitive expansion throughout evolution46,74,75,87 . Forebrain is composed of diencephalon (hypothalamus, epithalamus and subthalamus) along with the telencephalon
3
ling the first as observed in sensory systems80 , and ETI
controlling the latter as observed in motor systems102 .
For example, basal ganglia control of thalamic responses,
1 2 3 P
O
F
T
AD
TRN
(cortex and basal ganglia)10 . Analogous structures exist throughout the vertebrate lineage10,15 . Both the cortex and basal ganglia are demarcated by local recurrent
connections, with the main difference being that cortical
excitatory recurrence gives rise to long-lasting attractor
dynamics thought to underlie many cognitive processes
such as memory and decision making104,107 . The basal
ganglia are inhibitory structures, with local recurrence
thought to implement selection through lateral inhibition
rather than long lasting attractor state63,105,106 . Within
this framework, what does the thalamus do? The answer depends on ‘the thalamus’ in question. Traditionally, thalamic nuclei (see Fig. 2) are defined as collections of neurons that are segregated by gross features
such as white matter tracts and various types of tissue
staining procedures42 . This gross anatomical classification has been equated with a functional one, where individual thalamic nuclei giving rise to a set of defined
functions42,43 . More recent fine anatomical studies challenge this notion, showing that within individual nuclei,
single cell input/output connectivity patterns are quite
variable (see below).
This thalamic input diversity sets the bases for nonunitary computational role of thalamus in integrating information. Thalamus lacks lateral excitatory connections
and rather receives inputs from other subcortical structures and/or cortex. In fact, a major feature of forebrain
expansion across evolution is the invasion of the thalamus by cortical inputs30,85 . Most (90-95%) afferents to
the relay nuclei are not from the sensory organs43,93 . The
majority of these inputs originate from intrinsic (local
GABAergic neurons, projections from reticular nucleus,
thalamic interneurons–mostly absent in non-LGN relay
nuclei) and extrinsic sources (feedback from layer 6 of
the cortex as well as from brainstem reticular formation)
(see Fig. 3 for an example view of cell and network architecture diversity). Recent anatomical studies have shown
a great diversity of cortical input type, strength and inferred degree of convergence, even within individual thalamic nuclei11 .
In addition to the diversity of excitatory inputs, thalamic circuits receive a diversity of inhibitory inputs. The
two major systems of inhibitory control are the thalamic reticular nucleus (TRN), a shell of inhibitory nucleus
surrounding thalamic excitatory nuclei, and the extrathalamic inhibitory system; a group of inhibitory projects
across the fore-, mid- and hindbrain (see32 for a review
on thalamic inhibition). Perhaps a major differentiating
feature of these two systems (TRN and ETI) is temporal precision. One of the key characteristics of thalamus
is lack of direct local loops. Only a very small group
of inhibitory neurons with local connections exists in
thalamus98 . A mechanistic consequence of this architecture is the differential control of thalamic response gain
and selectivity (Fig. 3 F and G), with the TRN control-
CN
VL
AV
LD
LD
TRN
MD
AM
H
PU
VPL
IL
VA
IL
VPM
CM
LGN (D)
(V)
MGN
TRN
VPI
1
2
3
Figure 2. Schematic layout of thalamic nuclei: Three
cross sections of monkey thalamus. AD: anterodorsal nucleus; AM : anteromedial nucleus; AV : anteroventral nucleus;
CM : centromedian nucleus; CN : caudate nucleus; H : habenular nucleus; IL: intralaminar nuclei;LD: lateral dorsal nucleus; LGN : lateral geniculate nucleus;MD: mediodorsal nucleus; MGN(D): medial geniculate nucleus (dorsal); MGN(V):
medial geniculate nucleus (ventral); PU : pulvinar; TRN : thalamic reticular nucleus (not a relay nucleus); VA: ventral
anterior nucleus; VL: ventral lateral nucleus; VPI : ventral
posterior nucleus (inferior); VPL: ventral posterior nucleus
(lateral); VPM : ventral posterior nucleus medial. Redrawn
from93 .
a form of ETI control, would be implemented through
thalamic disinhibition, which is not only dependent on
ETI input, but also a special type of thalamic conductance that enables high frequency ‘bursting’ upon release
from inhibition19,29 .
Overall, the variety of thalamic inputs (both excitatory and inhibitory), combined with intrinsic thalamic
features such as excitability and morphology will determine the type of intrinsic computations the thalamus
performs. For example, if a thalamic neuron within a
particular circuit motif received temporally offset cortical inputs, one direct and another through basal ganglia
stations, then that neuron may be perfect for computing
a prediction error based on a raw cortical signal (or state)
at t0 and a processed one at t0 + ∆t (basal ganglia operation). This view appears to be consistent with recent observations of confidence encoding in both sensory47 and
motor systems41,54 .
4
A
B
IIIb
IIIc
IVa
IVb
C
V
100 µm
100 µm
E
D
TRN
VA
VL
100 µm
100 µm
100 µm
F
G
Th-cx
RE
Th-cx
RE
Th-cx
Th-cx
L-circ
Th-cx
RE
L-circ
Figure 3. Diversity and complexity of thalamo-cortical architectures: (A) Comparative size of a single RGC (retinal
ganglion cell) terminal (red) and an LGN neuron (black); Redrawn from25 . (B) Complete terminal arbor of a single LGN
neuron projecting to V1; Redrawn from82 . (C) Projection of a TC (thalamo-cortical) neuron to cat motor cortex (with only
23 terminals in TRN versus 1632 terminals in VA/VL); Redrawn from45,76 (D,E) Axonal arborization of a single MD neuron
(E) and a single POm neuron (E). Principal target layers are: layer I (green), layer II-IV (green-yellow), layer V (blue),
layer VI (purple); Redrawn from50 . Note the comparative size of panels A-E (scale bar at 100 µm). (F,G) Synaptic network
of thalamo-cortical (Th-cx), Reticular thalamic cell (RE) and local circuit (L-circ) thalamic interneuron in rodents (F) and
feline/primates (G). Note that rodents do not have L-circ. Afferent axon excites Th-cx, which in return sends the signal to
cortex. RE inhibitory effect on Th-cx cells varies depending on the excitatory drive to each Th-cx cell (F : compare the two
neurons on the right versus the one on the left). Axonal collaterals of an RE cell could inhibit another RE cell (G: top RE
inhibits the bottom RE), which releases the activity of L-circ leading to inhibition of weakly excited Th-cx (bottom) adjacent
to the active Th-cx (top). Panels F and G are redrawn from99 based on experiments from97,100 .
Thalamic output diversity sets thalamus to be poised
as the relay as well as the modulator of cortical activity. Thalamic relay nuclei mostly project to the cortical
middle layers in a topographic fashion. However, the
majority of thalamic structures project more diffusely to
the cortical superficial layers, such as mediodorsal (MD),
posteriodmedial complex (POm) and pulvinar for example (see Fig. 3 for an example of thalamic cell and circuit
diversity). These diffuse projections seem poorly suited
to relay information in a piecewise manner. Rather,
5
they might have a modulatory role of cortical function.
Further, a great degree of diversity can be observed at
the level of thalamic axonal terminals within the cortex.
While the idea of a thalamic relay was consolidated by
observing that the main LGN neurons thought to be associated with form vision (M and P pathways) exhibit
spatially compact cortical terminals, recent anatomical
studies of individual neurons across the thalamus show
a variety of terminal sizes and degree of spatial spread
and intricate computational architecture (Fig. 3). This
complexity of the architecture and diversity of the computing nodes are among the key factors that set apart
the thalamo-cortical system from other conventional and
unconventional computing engines (Fig. 1). Part of
the complication in understanding how these anatomical
types give rise to different functions is their potential for
contacting different sets of excitatory and inhibitory cortical neurons. For example, activating the mediodorsal
thalamus (MD) does not generate spikes across a population of prefrontal cortical neurons they project to, while
activating LGN generates spikes in striate cortex88 . Instead, MD activation results in overall enhancement of
inhibitory tone, coupled with enhanced local recurrent
connectivity within the PFC. This finding also argues
against a relay function, because in this case the MD is
changing the mode by which PFC neurons interact with
one another, initiating and updating different attractor
dynamics underlying distinct cognitive states. This idea
of the thalamus controlling cortical state parameters is
highlighted in Figs. 4, 5 and the next section.
Many facets of thalamic computation
It is commonly thought that processes like attention,
decision making and working memory are implemented
through distributed computations across multiple cortical regions12,62,89 . However, it is unclear how these computations are coordinated to drive relevant behavioral
outputs. From an anatomical standpoint, the thalamus is
strategically positioned to perform this function, but relatively little is known about its broad functional engagement in cognition. The thalamic cellular composition
and network structure constrain how cortex receives and
processes information. The thalamus is the major input
to the cortex and interactions between the two structures
are critical for sensation, action and cognition44,69,92 . Despite recent studies showing that the mammalian thalamus contains several circuit motifs, each with specific input/output characteristics, the thalamus is traditionally
viewed as a relay to or between cortical regions93 .
It is worth mentioning that this view of bona fide
thalamic computations is quite distinct from the one in
which thalamic responses reflect their inputs, with only
linear changes in response size. This property of reflecting an input (with only slight modification of amplitude)
was initially observed in the lateral geniculate nucleus
(LGN), which receives inputs from the retina. LGN re-
sponses to specific sensory inputs (their receptive fields,
RF) are very similar to those in the retina itself, arguing
that there is little intrinsic computation happening in the
LGN itself outside of gain control. Success in early vision
studies36,37 might have inadvertently given rise to the
LGN relay function being generalized across the thalamus. The strictly feedforward thalamic role in cognition
requires reconsideration33 ; only a few thalamic territories
receive peripheral sensory inputs and project cortically in
a localized manner, as the LGN does25,42,45,82,92 . Below
we will discuss how the distinctive anatomical architecture and computational role of pulvinar and MD differ
from the one just described for LGN.
The largest thalamic structures in mammals, the MD
and pulvinar contain many neurons that receive convergent cortical inputs and project diffusely across multiple
cortical layers and regions11,86 . For example, the primate
pulvinar has both territories that receive topographical,
non-convergent inputs from the striate cortex86 and others that receive convergent inputs from non-striate visual
cortical (and frontal) areas60 . This same thalamic nucleus also receives inputs from the superior colliculus79 , a
subcortical region receiving retinal inputs. This suggests
that the pulvinar contains multiple input ‘motifs’ solely
based on the diversity of excitatory input. This input diversity is not limited to the pulvinar, but is seen within
many thalamic nuclei across the mammalian forebrain8 .
Local inactivation of pulvinar neurons results in reduced
neural activity in primary visual cortex81 suggesting a
feedforward role. In contrast, recent findings show that
during perceptual decision making pulvinar neurons encode choice confidence, rather than stimulus category47 ,
strongly arguing for more pulvinar functions beyond relaying information.
In the case of MD, direct sensory input is limited64 and
the diffuse, distributed projections to cortex50 are poorly
suited for information relay; this input/output connectivity suggests different functions. Recent studies9,88 have
begun to shed light on the type of computation that MD
performs. Studying attentional control in the mouse88
has revealed that MD coordinates task-relevant activity in the prefrontal cortex (PFC) in a manner analogous to a contextual signal regulating distinct attractor
states within a computing reservoir. Specifically, in a
task where animals had to keep a rule in mind over a brief
delay period (Fig. 4A), PFC neurons show populationlevel persistent activity following the rule presentation, a
sensory cue that instructs the animal to direct its attention to either vision or audition (Fig. 4B,C). MD neurons
show responses that are devoid of categorical selectivity
(Fig. 4D), yet are critical for selective PFC activity; optogenetic MD inhibition diminishes this activity, while MD
activation augments it. The conclusion is that MD inputs
enhance synaptic connectivity among PFC neurons or
may adjust the activity of PFC neurons through selective
filtering of the thalamic inputs. In other words, delayperiod MD activity maintains rule-selective PFC representations by augmenting local excitatory recurrence88 .
6
ings indicate that PFC might have to recursively pull in
MD to sustain cortical representations as working memory weakens with time. Together these studies indicate
that PFC cognitive computation can not be dissociated
from MD activity. Further evidence for the critical role
of the MD-PFC interaction for cognition is the disrupted
fronto-thalamic anatomical and functional connectivity
seen in neurodevelopmental disorders59,65,68,78,108 .
In a related study, a delayed nonmatching-to-sample
T-maze working memory task9 , it was shown that MD
amplification and maintenance of higher PFC activity indicated correct performance during the subsequent choice
phase of the task. Interestingly, MD-dependent increased
PFC activity was much more pronounced during the later
(in delay) rather than earlier part of the task. These find-
A
Ignore sound
Delay period
ATTEND TO VISION
Low-pass
High-pass
Cueing
100 ms
Trial available Start initiation
attend to vision
15
attend to vision
5
Firing rate (z-score)
Firing rate (z-score)
4
10
2
0
-2
0
15 attend to audition
10
attend to audition
4
2
0
5
0
Presentation
Response
100 ms
Median: 1521 ms
D
attend to vision
Firing rate (Hz)
Cue-free delay
400 ms
C
B
Auditory selection
Ignore light
ATTEND TO AUDITION
Broadband
Visual selection
Time (s)
0
0.6
0
-2
attend to audition
2
0
-2
0
2
0.5
Time (s)
0
Time (s)
0.5
-2
0
Time (s)
0.5
E
Firing rate (Hz)
MD
12
10
8
6
4
2
0
15
outside
12
10
8
6
4
2
0
15
PFC
inside
12
10
8
6
4
2
0
15
outside
12
10
8
6
4
2
0
15
inside
10
10
10
10
5
5
5
5
0
−.2 0 .2 .4 .6
0
−.2 0 .2 .4 .6
0
−.2 0 .2 .4 .6
0
−.2 0 .2 .4 .6
Figure 4. MD-PFC interactions during sustained rule representations. (A) attentional control task design. (B)
Example peri-stimulus time histogram (PSTH) for a neuron tuned to attend to vision rule signaled through low pass noise
cue (C) Examples showing that rule-specificity is maintained across distinct PFC rule-tuned populations. (D) PSTHs of four
MD neurons showing consistent lack of rule specificity. (E) Example rasters and PSTHs of an MD and PFC neuron when
the animal is engaged in the task and outside of the behavioral arena. In contrast to PFC, MD neurons show the contextual
difference in a change in firing rate. Figure is redrawn from88 .
7
Can MD select cortical subnetworks based on contextual
modulation?
Why would a recurrent network (PFC) computation
depend on its interaction with a non-recurrent (MD) nonrelay network? What computational advantage such system would have? Using a chemogenetic approach, a recent study suggested that information flow in the MDPFC network can be unidirectional. While both inactivating PFC-to-thalamus and MD-to-cortex pathways impaired recognition of a change in reward value in rats
performing a decision making task, only the inactivation
of MD-to-cortex pathway had an impact on the behavioral response to a change in action-reward relationship3 .
Given that a sensory stimulus may require a different action depending on the context in which it occurs, the
ability to flexibly re-route the active PFC subnetwork
to a different output may be crucial. In an architecture like the PFC-MD network, where MD can modulate
PFC functional connectivity, MD might well be suited
to re-route the ongoing activity in a context dependent
manner. In fact, in the mouse cognitive task described
above (Fig. 4A), a subset of MD neurons showed substantial spike rate modulation during task engagement
compared to when the animals is in its home in cage (see
Fig. 4E)88 . In contrast, PFC neurons show very little
difference in spike rates when the animal gets engaged
in the task. This suggests that perhaps different subsets
of MD neurons are capable of encoding task ‘contexts’.
Subsequently, each given subset could unlock a distinct
cortical association. These MD subsets have to be able
to shift the cortical states dynamically while maintaining
the selectivity based on the subset of cortical connections
they target. This idea would also fit with the paradigm
shift indicating that thalamic neurons exert dynamical
control over information relay to cortex6,77 .
Overall, the anatomical and neurophysiological data
show that the thalamic structure and cortico-thalamic
network circuitry, and the interplay between thalamus
and cortex, shape the frame within which thalamus plays
the dual role of relay and modulator. Under this framework, different thalamic nuclei carry out multitude of
computation including but not limited to information relay. A suggestion of this comparative computational role
of LGN, pulvinar and MD is depicted in Fig 8.
Is thalamus a read-write medium for cortical parallel
processing?
If the brain were to function as a simple pattern matching system without wiring and metabolic constrains, evolution would just expand the size and depth of the network to the point that it could potentially memorize a
large number of possible patterns. Possibly, evolution
would have achieved this approximation of arbitrary patterns by evolving a deep network. This would be a desirable solution since any system can be defined as a poly-
nomial Hamiltonians of low order, which can be accurately approximated by neural networks53 . But cognition is much more than template matching and classification achieved by a neural network. The limits of template matching methods in dealing with (rotation, translation and scale) invariance in object recognition quickly
became known to neuroscientists and in early works on
computer vision. One of the early pioneers of AI, Oliver
Selfridge, proposed Pandemonium architecture to overcome this issue90 . Selfridge envisioned serially connected
distinct demons (an image demon, followed by a set of
parallel feature demons, followed by a set of parallel cognitive demons and eventually a decision demon), that independently perceive parts of the input before reaching
a consensus together through a mixture of serial culmination of evidence from parallel processing. This simple
feedforward computational pattern recognition model is
(in some ways) a predecessor to modern day connectionist feedforward neural networks, much like what we discussed earlier in the text. However, despite its simplicity,
Pandemonium was a leap forward in understanding that
the intensity of (independent parallel) activity along with
a need to a summation inference are the keys to move
from simple template matching to a system that has a
concept about the processed input. A later extension of
this idea was proposed by Allen Newell as the Blackboard
model: “Metaphorically, we can think of a set of workers, all looking at the same blackboard: each is able to
read everything that is on it and to judge when he has
something worthwhile to add to it. This conception is
just that of Selfridge’s Pandemonium: a set of demons
independently looking at the total situation and shrieking
in proportion to what they see that fits their natures” 71 .
Blackboard AI systems, adapted based on this model,
have a common knowledge base (blackboard) that is iteratively updated (written to and read from) by a group
of knowledge sources (specialist modules), and a control
shell (organizing the updates by knowledge sources)72 .
Interestingly, this computational metaphor can also be
extended to the interaction between thalamus and cortex, though thalamic blackboard is not a passive one as
in the blackboard systems34,66,67 . Although, initially the
active blackboard was used as an analogy for LGN computation, the nature of MD connectivity and its communication with cortex seem much more suitable to the
type of computations that is enabled by an active blackboard. Starting with an input, thalamus as the common
blackboard visible to processing (cortical) modules, initially presents the problem (input) for parallel processing by modules. Here by module, we refer to a group of
cortical neurons that form a functional assembly which
may or may not be clustered together (in a column for
example). By iteratively reading (via thalamo-cortical
projections) from and writing (via cortico-thalamic projections) to this active blackboard, expert pattern recognition modules, gradually refine their initial guess based
on their internal processing and the updates of the common knowledge. This process continues until the problem
8
is solved (Fig. 5). However, since we are dealing with
a biological system with finite resources, this back and
forth communication needs to have some characteristics
to provide a viable computational solution. First and
foremost, the control of interaction and its scheduling
has to have a plausible biological component and should
bind solutions as time evolves. Second, to avoid turning into an NP-hard (non-deterministic polynomial-time
hardness) problem, there must exist a mechanism that
stops this iterative computation once an approximation
has been reached (Fig. 5). In this paper, we propose specific solution to the first problem and a plausible one for
the later issue. We suggest that phase-dependent contextual modulation serves to deal with the first issue and a
multi-objective optimization of efficiency (computational
information gain) and economy (computational cost, i.e.
metabolic needs and the required time for computation)
handles the second issue (Fig. 6). In both cases, we suggest that thalamus plays an integral role in conjunction
with cortex.
Core attributes of cognitive processing, attention and
binding in time
network models84 . The most widely used ANNs (Feedforward nets , i.e. multilayer perceptrons/Deep Learning
algorithms) face fundamental deficiencies: the ubiquitous
training algorithms (such as back-propagation), i) have
no biological plausibility, ii) have high computational cost
(number of operations and speed), and iii) require millions of examples for proper adjustment of NN weights.
These features renders feedforward NNs not suitable for
temporal information processing. In contrast, recurrent neural networks (RNNs) can universally approximate the state of dynamical systems26 , and because of
their dynamical memory are well suited for contextual
computation. It has been suggested that RNNs are able
to perform real-time computation at the edge of chaos,
where the network harbors optimal readout and memory
capacity7 . The functional similarity of gated-recurrent
neural networks and cortical microcircuitry continues to
be at the forefront of AI and neuroscience interface13 . If
higher cortical areas were to show some features of RNNlike networks, as manifested by the dynamical response of
single neurons57 , then we anticipate that the local computation (interaction between neighboring neurons) to be
mostly driven by external biases. The thalamic projections could then play the role of bias where they seed
the state of the network. While this may be a feasible
portrait, the picture misses the existence of large-scale
feedback loops which are neither feedforward nor locally
recurrent2,31 . To reconcile, we need our system of interest to show phase-sensitive detection that binds the
locally recurrent activity in cortex, with large-scale feedback loops.
To expand the idea of read-write further, let’s revisit
some core attributes of cognitive processing. First, we
wish to point to two key features of cognitive processing, namely “searchlight of attention” and “binding in
time”. Then we examine how these attributes match
the computational constrains that are met by thalamic
circuitry and functions. The “searchlight of attention”,
first proposed by Crick as a function of thalamic reticular formation14 , proposes that through change in firing pattern (bursting vs tonic firing), thalamic nuclei
dynamically switch between detection and perception.
The bursting nonlinear response signals a change in the
environment, while the tonic mode underlies perceptual
processing. One biophysical mechanism responsible for
this change between bursting and tonic modes is implemented by cortico-thalamic activation of glutamate
metabotropic receptors61 . This mechanism puts thalamus as the mediator between peripheral and cortical input, creating a closed-loop computation1 . A necessary
property of thalamic-driven change of perceptual processing, and hence cognition, is the existence of cortical
re-entry. From both anatomical studies and electrophysiological investigations31 , we know that thalamus is at a
prime position to modify the relay signal based on the
cognitive processing that is happening in the cortex9,88 .
This thalamic-driven attention entails “binding in time”
since how thalamus modifies its relay at a given time is
itself influenced by what is perceived by the cortex in
time prior.
But how can the “binding in time” avoid locking-in
the thalamic function to a set of inputs at a given time?
How can thalamus constantly be both ahead of cortex
and yet keep track of the past information? The secret
may be embedded in the non-recurrent intrinsic structure of thalamus and the recurrent structure of the higher
cortical areas. As mentioned earlier, we know that hierarchical convolutional neural networks (HCNN), which
can recapitulate certain properties of static hierarchical
forward models, can not capture any processes that need
to store prior states109 . As a result, context-dependent
processing can be extremely hard to implement in neural
Computational constrains and the role of thalamus in
phase-dependent contextual modulation
Based on the observations of behavior, higher cognition requires “efficient computation”, “time delay feedback”, the capacity to “retain information” and “contextual” computational properties. Such computational cognitive process surpass the computational capacity of simple RNN-like networks. The essential required properties
of a complex cognitive system of such kind are: 1) input
should be nonlinearly mapped onto the high-dimensional
9
A
Static Data
1 1 0 1 1 0 1 0 1
0 1 1 0 1 1 0 1 1
1 0 1 1 1 0 1 0 0
Function 0 1 1 0 1 1 0 0 0
Weight .3 .1 .5 .4 .6 .7 0 .2 .1
.......
.......
Classification
1 0 1 1 1 0 1 1 1
0 1 1 0 1 1 0 1 1
1 1 0 1 1 0 1 0 1
.......
.......
B
Weight .2 .3 .9 .6 .1 ...
Pointer a b c d e ...
e
Tim
(a)
Function 0 1 1
Weight .3 .1 .5
Dynamic Data
1 1 0 1 1 0 1 0 1
0 1 1 0 1 1 0 1 1
1 0 1 1 1 0 1 0 0
tn+3dt
C
Non-recurrent
(c)
1 1
.2 .8
(d)
0 1 1 0 1
.3 .1 .5 .4 .6
(e)
1 0 0 0
.7 0 .2 .1
Weight .3 .7 .5 .6 .2 ...
Pointer a b c d e ...
(a)
Function 1 0 1
Weight .7 .3 .6
.......
.......
1 0 1 1 1 0 1 1 1
0 1 1 0 1 1 0 1 1
1 1 0 1 1 0 1 0 1
(b)
0 1 1 0
.4 .6 .7 0
tn
tn+dt
tn+2dt
(b)
0 1
.3 .2
(d)
(c)
(e)
1 0 1 1 0 0 1 1 0 1 1 0 1
.2 .5 .5 .2 .6 .8 .6 0 .3 .3 .5 .3 .4
Weight .5 .9 .6 0 0
Pointer a b c d e
(b)
(a)
1 0 1 1 0
Function 1 1 1 0 0
.5 .1 .6 0 .9
Weight .8 .6 .5 0 .1
...
...
1 0
.8 .1
(c)
1 1 0
.7 .4 .2
.......
.......
Reservoir
Readout
Figure 5. Schematic representation of thalamic cognitive contextual computation. (A): In the case of static data,
a set of function/weight modules can yield good classification. Function represents a polynomial (since any system that is
known to be a polynomial Hamiltonians of low order can be accurately approximated by neural networks53 ) and the weights
exemplify the connection matrix of an artificial neural network encapsulating this polynomial. Stacking multiple of such module
can increase the accuracy of polynomial approximation (such as in the case of CNN). (B): Thalamo-cortical computation for
contextual processing of dynamic data. Each dataframe is processed by a weight/pointer module (thalamus MD-like structure)
which like a blackboard is writable by different sets of neuronal assemblies in cortex. Thalamic pointers assign the assemblies;
modules’ weights adjust the influence of each assembly in further computational step [inset C shows a non-recurrent thalamic
nuclei (MD-like) modulating the weights in the PFC (reservoir and readout). Here, depending on the context (blue or red),
the interactions between MD and Reservoir, between Reservoir and Readout, and between Readout and MD could pursue
one of the two possible outcomes. Specifically, MD changes the weights in the Reservoir to differentially set assemblies that
produce two different attractor states, each leading to one of the two possible network outputs]. In (B, C), each operation
of the thalamic module is itself influenced, not only by the current frame (t), but also by the computation carried by cortex
module on the prior frame (t − 1). Cortical module is composed of multiple assemblies where each operate similar to the
function/weight module of the static case. These assemblies are locally recurrent and each cell may be recruited to a different
assembly during each operation. This mechanism could explain why prefrontal cells show mixed selectivity in their responses
to stimuli (as reported in27,83 ). Through this recursive interaction between thalamus and cortex, cognition emerges not as just
a pattern matching computation, but through contextual computation of dynamic data (bottom right schematic drawing).
state, while different inputs map onto different states, 2)
slightly different states should map onto identical targets,
3) only recent past should influence the state and network
is essentially unaware of remote past, 4) a phase-locked
10
Biological constrains and the role of thalamus in
computational optimization
Computation and optimization are two sides of the
same coin. But how does the brain optimize the computations that would match its required objective, i.e.
cognitive processing? There is a current trend of thinking
that brain optimizes some arbitrary functions, with the
hopes that the future discovery of these unknown functions may guide us to establish a link between brain’s operations and deep learning58 . This line of approach to optimizational (and computational) operations of the brain
Information
η3
A1
A2
Cost
η4
η2
ζ4
η1
ζ1
B
η3
ζ3
ζ2
η4
η2
η1
ζ1
ζ3
ζ4
ζ2
C
ζ4
ζ3
Cost
loop should decode information that is already encoded
in time and 5) the combination of 1-4, should optimize
sensory processing based on the context. The first three
attributes of such system have close relevance to constrains and computational properties of higher cortical
areas (prefrontal). The same three are also the main features of reservoir computing, namely “separation property”, “approximation property” and “fading memory”.
Reservoir Computing (RC), with Echo State Networks
(ESN)39,40 and Liquid State Machines (LSM)55,56 as two
early variants of what is now unified as RC, have emerged
as powerful RNNs. Although a large enough random
network could essentially represent all combinations of
cortical network84 , the training of such system would be
hard and changing the network from task to task will not
be easily achievable. An advantage of a RC systems is
to “non-linearly” map a lower dimensional system to a
high-dimensional space facilitating classification of the elements of the low-dimensional space. The last two properties match the structure and computational constraints
of non-relay thalamic system as a contextual modulator
that is phasically changing the input to the RC system.
In fact, in an RC model of prefrontal cortex, addition
of a phase neuron significantly improved the networks
performance in complex cognitive tasks. The phase neuron improves the performance by generating input driven
attractor dynamics that best matched the input23 . In
a recent study, electronic implementation and numerical studies of a limited RC system of a single nonlinear node with delayed feedback showed efficient information processing4 . Such reservoir’s transient dynamical response follows delay-dynamical systems, and only a limited set of parameters set the rich dynamical properties
of delay systems38 . This system was able to effectively
process time-dependent signals. The phase neuron23 and
delayed dynamical RC4 both show properties that resemble the thalamic functions as discussed here. The collective system (thalamus and cortex together) is neither
feedforward, nor locally recurrent, but it has a mixture of
non-recurrent phase encoder that keeps copies of the past
processing and modulates the sensory input relay and its
next step processing (Fig. 5). These features emphasize
that the perceptual and cognitive processing can not be
solely cortico-centric operations.
ζ2
ζ1
η1
η2
η3
η4
Information
Figure 6. Dynamic role of thalamo-cortical system in
the information/cost optimization. (A) Iso-maps of information (A1) and cost (A2) in the domain specified by ω
and λ (functions of cortical and thalamic activity). Information across each Iso-quant curve (η1 for example) is constant
and is achieved at a certain mixture of ω and λ. Optimal information can be obtained by moving outward (arrows, A1 ).
Cost optimization can be achieved by moving inward (A2 ).
(B) since information and cost are both defined in the domain
of ω and λ, thalamus and cortex jointly contribute to information and cost optimization. The points where the iso-quant
curves’ tangents are equal (black dashed line), provide the optimal combination of information/cost (green curve). In any
cycle of cognitive operation, depending on the prior state of
the system (ω and λ), the nearest points on the green curve
are the optimal solutions for ending that cycle. (C) mapping of the optima curve to information/cost domain shows
all pareto efficient allocations (cyan curve). The slope of the
parto frontier shows how the system trades cost versus information: along the pareto curve, efficiency is constant but
the exchange between information and cost is not. All allocations inside of this curve could be improved as thalamus
and cortex interact. The grey zone shows the biophysically
not-permissible allocation of computation and resource.
has few flaws. First, it avoids specifying what function
the brain is supposed to optimize (and as a result it remains vague). Second, it refrains from addressing certain
11
limitations that brain has to cope with due to biological
constrains. First of these limitations is the importance of
using just enough resources to solve the current perceptual problem. Second is the necessity to come up with
a solution just in (the needed) time. The importance of
“just-enough” and “just-in-time” computation in cortical computation should not be overlooked20 . If the first
condition is not met, the organism can not sustain continued activity since the metabolic demand surpasses the
dedicated energetic expenditure and the animal can not
survive. In fact, the communication in neural networks
are highly constrained by number of factors, specifically
the energetic demands of the network operations51 . From
estimates of the cost of cortical computation52 , we know
that the high cost of spiking forces the brain to rely on
sparse communication and using only a small fraction of
the available neurons5,94 . While, theoretically, cortex can
dedicate a large number of neurons (and very high dynamical space) to solve any cognitive task, metabolic demand of such high-energetic neural activity renders such
mechanism highly inefficient. As a result, the “law of diminishing returns” dictates that increased energetic cost
causing excessive pooling of active neurons to an assembly would be penalized73 . The penalization for unnecessary high-energetic neural activity, in itself, should be
driven by the nature of computation rather than being
formulated as a fixed arbitrary threshold imposed by an
external observer. On the other hand, a system can resort to low-cost computation at any given time but dedicate long enough time to solve the task on hand. Naturally, such system would not be very relevant to the
biological systems since time is of essence. If an animal
dedicates a long instance of its computational capacity
to solve a problem, the environment has changed before
it reaches a solution and the solution becomes obsolete.
A deer would never have an advantage for its brain to
have fully analyzed the visual scene instead of spotting
the approaching wolf and shifting resources to the mostneeded task, i.e. escape. As a result, many of the optimization techniques and concepts that may be relevant
to artificial neural networks are irrelevant to embodied
computational cognition of the brain. The optimization
that the brain requires is not aiming for the best possible
performance, but rather needs to reach a good mixture
of economy and efficiency.
Not surprisingly, these constrains, i.e. efficiency and
economy, are cornerstones of homeostasis and are observed across many scales in living systems101 . The
simple “Integral feedback” acts as the mainstay of control feedback in such homeostatic systems (such as E
Coli heat-shock or DNA repair after exposure to gamma
radiation)18,21,22,48 . Change in input leads to change in
the output and the proportional change in the controller
aiming to reset the output to the desired regime. The
constrains that we discussed above, remind us of a similar scenario. Instead of just trying to deal with one fitness function at a time (where the minima of the landscape would be deemed as “the” optima), the brain has to
perform a multi-objective optimization, finding solutions
to both metabolic cost (economy) and just-in-time (efficiency) computation. Thus we can infer that a unique
solution does not exist for such a problem. Rather,
any optimization for computational efficiency will cost
us economy and any optimization for economy will cost
us efficiency. In such case, a multi-objective optimization
pareto frontier is desirable. Pareto frontier of information/cost will be the set of solutions where any other
point in the space is objectively worse for both of the
objectives28,49,101 . As a result, the optimization mechanism should push the system to this frontier. The iterative dynamical interaction between thalamus and cortex
seems to provide an elegant solution for this problem.
We discuss this in more details below.
Consider a set of functions, ω and λ of fT h (firing rate
of thalamic cell) and fCx (firing rate of cortical cells).
Uncertainty (or its opposite, information) and computational cost (a mixture of time and metabolic expense) can
both be mapped to this functional space of ω (fT h , fCx )
and λ (fT h , fCx ) (Fig. 6A1,2). Let’s define computational
cost and information as product and linear sum of cordn
dn
n
, θ fdTt h + ψ fdCx
tical and thalamic activity ( αfTnh .βfCx
;
t
with α, β, θ, ψ as coefficients) to reflect the logarithmic
nature of information (entropy) and the fact that biological cost is an accelerating function of the cost-inducing
variables18 . The hypothetical space of cost/information
is depicted in Fig. 6, where top panels show indifference
maps of information (A1) and cost (A2). The example
simulations and parametric plots of the cost and information functions defined as above are shown in Fig. 7.
In each indifference map, along each iso-quant curve,
the total functional attribute is the same. For example,
anywhere on the η1 curve, the uncertainty (or information) in our computational engine is the same. However, different iso-quant curves represent different levels of the functional attribute. For example, moving
outward increases information (reduces uncertainty) as
η1 < η2 < η3 < η4 and thus if computational cost was
not a constrain, the optimal solution would have existed
on η4 or further away (Fig.6 A1). In contrast, moving
inward would preserve the cost (ζ1 < ζ2 < ζ3 < ζ4 )
if the computational engine did not have the objective
of reducing uncertainty (Fig. 6A2). Since information
and cost are interdependent and both depend on the interaction between thalamus and cortex, we suggest that
information/cost optimization happens through an iterative interaction between thalamus and cortex (note
the blackboard analogy and contextual modulation discussed above). Since we defined both information and
cost as a set of iso-quant curves in the functional space of
ω (fT h , fCx ) and λ (fT h , fCx ), they can be co-represented
in the same space (Fig. 6B). Optimal solutions for information/cost optimizations are simply the solutions to
where the tangents of the iso-quant curves are equal (see
the tangents [black dashed lines] and points A, B, C and
D, in Fig. 6B). These points create a set of optimal solutions for the tradeoff between information and cost (green
12
curve). Mapping of the optimal solutions to the computational efficiency space E, gives us the pareto efficient
curve (cyan curve, Fig. 6C). Anywhere inside the curve
is not pareto efficient (i.e. information gain and computational cost can change in such a way that, collectively, the system can be in a better state (on the pareto
curve). Points outside of the pareto efficient curve are
not available to the current state of the system due to
the coefficients of ω and λ. A change in these coefficients
can potentially shape a different co-representation of information and cost (see Fig 7, top row for 3 different
instances of ω and λ based on different α, β, θ, ψ values), and thus a different pareto efficient curve (see Fig 7,
bottom row). These different possible pareto frontiers
can be set based on the prior state of the system and
the complexity of the computational problem on hand.
Nonetheless, the computational efficiency of the system
can not be infinitely pushed outward because of the system’s intrinsic biophysical constrains (neurons and their
wiring). The shaded region in Fig 7, bottom row, shows
this non-permissible zone.
In the defined computational efficiency space E, composed of the two variables information and cost (as the
objective functions, shown in bottom panels of Fig. 6 and
Fig. 7), solving a computational problem is represented
by a decrease in uncertainty. However, any change in
uncertainty has an associated cost. First derivative of
the pareto frontier shows “marginal rate of substitution”
∆
o
as ∆inf
. This ratio varies among different points on
cost
the pareto efficient curve. If we take two points on the
pareto curve in the computational efficiency space, such
as A and C for example, computational efficiency of these
two points are equal EA(η1 ,λ4 ) = EC(η3 ,λ2 ) . The change
in efficiency of point A with respect to information and
∂EA
∂EA
cost, are the partial derivatives ∂inf
o and ∂ cos t , respec∂EA
A
tively. As a result, ∂∂E
inf o dinf o + ∂ cos t dcos t = 0, meaning
that there is constant efficiency along the pareto curve,
the tradeoff between information and cost is not constant. The optimization in this space is not based on
some fixed built-in algorithm or arbitrary thresholds by
an external observer. Rather, information/cost optimization is the result of back and forth interaction between
thalamus and cortex. Based on the computational perspective that we have portrayed, thalamus seems to be
poised to operate as an optimizer. Thalamus receives a
copy of (sensory) input while relaying it, and receives an
efferent copy from the processor (cortex), while trying to
efficiently bind the information from past and present
and sending it back to cortex. The outcome of such
emergent optimization, is a pareto front in the economyefficiency landscape (Fig. 6,7). If the cortex were to be
the sole conductor of cognitive processing, the dynamics
of the relay and cortical processing would meander in the
parameter space and not yielding any optimization that
can provide a feasible solution to economic and just-intime computation. Such system is doomed to fail, either
due to metabolic costs or due to computational freeze
over time ; thus more or less be a useless cognitive en-
gine. In contrast, with the help of an optimizer that
acts as a contextual modulator, the acceptable parameters will be confined to a manifold within the parameter
space. Such regime would be a sustainable and favorable
domain for cognitive computing. This property shows
another important facet of a thalamo-cortical computational cognitive system and the need to move passed the
cortico-centric view of cognition.
Figure 7.
Dynamic parameter space of thalamocortical joint optimization of information/cost. Three
different realization of information/cost interaction as a function of thalamic and cortical activity (ω, λ) and the corresponding pareto curves (see Fig. 6 for details of this optimization construct). Pareto curve shows the optimal set of
both cost and information that can be obtained given the biophysical constrains of neurons and networks connecting them.
Every point on the pareto frontier shows technically efficient
levels for a given parameter set of ω, λ (see text for more details). All the points inside the curve are feasible but are not
maximally efficient. The slope (marginal rate of transformation between cost and information) shows how in order to
increase information, cost has to change. The dynamic nature of interaction between thalamus and cortex enables an
emergent optimization of information/cost depending on the
computational problem on hand and the prior state of the
system.
Concluding remarks: Reframing Thalamic function above
and beyond information relay
Lately, new evidence about the possible role of thalamus has started to challenged the cortico-centric view
of perception/cognition. Anatomical studies and physiological measurements have begun to unravel the importance of the Cortico-Thalamo-Cortical loops in cognitive
processes6,77 . Under this emerging paradigm, thalamus
plays two distinctive roles: a) information relay, b) modulation of cortical function93 , where the neocortex does
not work in isolation but is largely dependent on tha-
13
lamus. In contrast to cortical networks which operate
as specialized memory devices via their local recurrent
excitatory connections, the thalamus is devoid of local
connections, and is instead optimized for capturing state
information that is distributed across multiple cortical
nodes while animals are engaged in context-dependent
task switching88 . This allows the thalamus to explicitly represent task context (corresponding to different
combinations of cortical states), and through its unique
projection patterns to the cortex, different thalamic inputs modify the effective connections between cortical
neurons9,88 .
Figure 8. The emergent view of thalamic role in cognition. (Top) In the traditional view, serial processing of information
confines the role of thalamus to only a relay station. (Bottom) the view that is discussed in this manuscript considers thalamus
as a key player in cognition, above and beyond relay to sensory cortices. Through combining the efferent readout from cortex
with sensory afferent, MD-like thalamic nuclei modulate further activity of the higher cortex. The contextual modulation
enabled by MD is composed of distinctively parallel operations (individual circles represent the non-recurrent nature of these
processes due to lack of local excitatory connections). Under this view, and the computational operatives discussed here, the
thalamo-cortical system (and not just cortex) is in charge of contextual cognitive computing. The computation enabled by
Pulvinar/PO like nuclei is different from LGN and also from MD-like nuclei.
Here, we started with a brief overview of the architecture of thalamus, the back and forth communication between thalamus and cortex, then we provided the electrophysiological evidence of thalamic modulatory function,
and concluded with a computational frame that encapsulates the architectural and functional attributes of the
thalamic role in cognition. In such frame, the computational efficiency of the cognitive computing machinery is
achieved through iterative interactions between thalamus
and cortex embedded in the hierarchical organization
(Fig. 4, 5). Under this emergent view, thalamus serves
not only as relay, but also as a read/write medium for
cortical processing , playing a crucial role in contextual
modulation of cognition (Fig. 8). Such multiscale organization of computational processes is a necessary requirement for design of the intelligent systems16,95,96 . Distributed computing in biological systems in most cases
operates without central control70 . This is well reflected
in the computational perspective that we discussed here.
We suggest that through the continuous contextual modulation of cortical activity, thalamus (along with cortex)
plays a significant role in emergent optimization of computational efficiency and computational cost. This phenomenon has a deep relation with phase transitions in
complex networks. Different states (phases) of the network are associated with the connectivity of the com-
14
puting elements (see thalamic weight/pointer and cortical function/weight modules in Fig. 5). Interestingly,
intrinsic properties of the complex networks do not define the phase transitions in system. Rather, the interplay of the system with its external environment shapes
the landscape where phase transitions occur91 . This parallel in well-studied physical systems and neuronal networks of thalamo-cortical system show the importance of
the interplay between thalamus and cortex in cognitive
computation and optimization. The proposed frame for
contextual cognitive computation and the emergent information/cost optimization in thalamo-cortical system
can guide us in designing novel AI architecture.
ACKNOWLEDGMENTS
We wish to thank Michael Halassa for helpful discussions.
REFERENCES
1 Ahissar,
E., and Kleinfeld, D. Closed-loop Neuronal Computations: Focus on Vibrissa Somatosensation in Rat. Cerebral
Cortex 13, 1 (jan 2003), 53–62.
2 Ahissar, E., and Oram, T.
Thalamic Relay or CorticoThalamic Processing? Old Question New Answers. Cerebral
Cortex 25, 4 (nov 2015), 845–848.
3 Alcaraz, F., Fresno, V., Marchand, A. R., Kremer, E. J.,
Coutureau, E., and Wolff, M. Thalamocortical and corticothalamic pathways differentially contribute to goal-directed
behaviors in the rat. eLife 7 (feb 2018).
4 Appeltant, L., Soriano, M. C., Van der Sande, G., Danckaert, J., Massar, S., Dambre, j., Schrauwen, B., Mirasso,
C. R., and Fischer, I. Information processing using a single
dynamical node as complex system. Nature Communications 2
(sep 2011), 468.
5 Baddeley, R., Abbott, L. F., Booth, M. C. A., Sengpiel,
F., Freeman, T., Wakeman, E. A., and Rolls, E. T. Responses of neurons in primary and inferior temporal visual cortices to natural scenes. Proceedings of the Royal Society B:
Biological Sciences 264, 1389 (dec 1997), 1775–1783.
6 Basso, M. A., Uhlrich, D., and Bickford, M. E. Cortical
Function: A View from the Thalamus. Neuron 45, 4 (feb 2005),
485–488.
7 Bertschinger, N., and Natschläger, T. Real-time computation at the edge of chaos in recurrent neural networks. Neural
Comput 16 (Jul 2004), 1413–36.
8 Bickford, M. E. Thalamic Circuit Diversity: Modulation of
the Driver/Modulator Framework. Frontiers in Neural Circuits
9 (jan 2016).
9 Bolkan, S. S., Stujenske, J. M., Parnaudeau, S., Spellman,
T. J., Rauffenbart, C., Abbas, A. I., Harris, A. Z., Gordon, J. A., and Kellendonk, C. Thalamic projections sustain
prefrontal activity during working memory maintenance. Nature
Neuroscience 20, 7 (may 2017), 987–996.
10 Butler, A. B., and Hodos, W. Comparative Vertebrate Neuroanatomy. John Wiley & Sons Inc., aug 2005.
11 Clasca, F., Rubio-Garrido, P., and Jabaudon, D. Unveiling
the diversity of thalamocortical neuron subtypes. Eur J Neurosci 35 (May 2012), 1524–32.
12 Corbetta, M. Frontoparietal cortical networks for directing
attention and the eye to visual locations: identical, indepen-
dent, or overlapping neural systems? Proceedings of National
Academy of Science 95 (Feb 1998), 831–8.
13 Costa, R. P., Assael, Y. M., Shillingford, B., de Freitas,
N., and Vogels, T. P. Cortical microcircuits as gated-recurrent
neural networks. arXiv (2017).
14 Crick, F. Function of the thalamic reticular complex: the
searchlight hypothesis. Proceedings of the National Academy
of Sciences 81, 14 (jul 1984), 4586–4590.
15 Deacon, T. W.
Rethinking Mammalian Brain Evolution.
American Zoologist 30, 3 (aug 1990), 629–705.
16 Dehghani, N. Design of the Artificial: lessons from the biological roots of general intelligence. arXiv (2017).
17 Dehghani, N. Theoretical principles of multiscale spatiotemporal control of neuronal networks: a complex systems perspective.
BiorXiv (jan 2017).
18 Dekel, E., and Alon, U. Optimality and evolutionary tuning
of the expression level of a protein. Nature 436, 7050 (jul 2005),
588–592.
19 Deniau, J. M., and Chevalier, G. Disinhibition as a basic process in the expression of striatal functions. II. The striato-nigral
influence on thalamocortical cells of the ventromedial thalamic
nucleus. Brain Research 334, 2 (may 1985), 227–233.
20 Douglas, R. J., and Martin, K. A. Mapping the Matrix: The
Ways of Neocortex. Neuron 56, 2 (oct 2007), 226–238.
21 El-Samad, H. J., Goff, J. P., and Khamash, M. H. Calcium
Homeostasis and Parturient Hypocalcemia: An Integral Feedback Perspective. Journal of Theoretical Biology 214, 1 (jan
2002), 17–29.
22 El-Samad, H. J., Kurata, H., Doyle, J. C., Gross, C. A.,
and Khamash, M. H. Surviving heat shock: Control strategies
for robustness and performance. Proceedings of the National
Academy of Sciences 102, 8 (jan 2005), 2736–2741.
23 Enel, P., Procyk, E., Quilodran, R., and Dominey, P. F.
Reservoir Computing Properties of Neural Dynamics in Prefrontal Cortex. PLOS Computational Biology 12, 6 (jun 2016),
e1004967.
24 Felleman, D. J., and Essen, D. C. V. Distributed Hierarchical
Processing in the Primate Cerebral Cortex. Cerebral Cortex 1,
1 (jan 1991), 1–47.
25 FitzGibbon, T., Eriköz, B., Grünert, U., and Martin, P. R.
Analysis of the lateral geniculate nucleus in dichromatic and
trichromatic marmosets. Journal of Comparative Neurology
523, 13 (jul 2015), 1948–1966.
26 Funahashi, K., and Nakamura, Y. Approximation of dynamical systems by continuous time recurrent neural networks. Neural Networks 6, 6 (jan 1993), 801–806.
27 Fusi, S., Miller, E. K., and Rigotti, M. Why neurons mix:
high dimensionality for higher cognition. Current Opinion in
Neurobiology 37 (apr 2016), 66–74.
28 Godfrey, P., Shipley, R., and Gryz, J. Algorithms and analyses for maximal vector computation. The VLDB Journal 16,
1 (sep 2006), 5–28.
29 Goldberg, J. H., Farries, M. A., and Fee, M. S. Basal
ganglia output to the thalamus: still a paradox. Trends in Neurosciences 36, 12 (dec 2013), 695–705.
30 Grant, E., Hoerder-Suabedissen, A., and Molnár, Z. Development of the Corticothalamic Projections. Frontiers in Neuroscience 6 (2012).
31 Groh, A., Bokor, H., Mease, R. A., Plattner, V. M.,
Hangya, B., Stroh, A., Deschenes, M., and Acsády, L.
Convergence of Cortical and Sensory Driver Inputs on Single
Thalamocortical Cells. Cerebral Cortex 24, 12 (jul 2014), 3167–
3179.
32 Halassa, M. M., and Acsády, L. Thalamic Inhibition: Diverse
Sources Diverse Scales. Trends in Neurosciences 39, 10 (oct
2016), 680–693.
33 Halassa, M. M., and Kastner, S. Thalamic functions in distributed cognitive control. Nature Neuroscience 20, 12 (nov
2017), 1669–1679.
34 Harth, E. M., Unnikrishnan, K. P., and Pandya, A. S. The
15
inversion of sensory processing by feedback pathways: a model
of visual cognitive functions. Science 237, 4811 (jul 1987), 184–
187.
35 Heeger, D. J. Theory of cortical function. Proceedings of the
National Academy of Sciences 114, 8 (feb 2017), 1773–1782.
36 Hubel, D. H., and Wiesel, T. N. Receptive fields of single
neurones in the cat's striate cortex. The Journal of Physiology
148, 3 (oct 1959), 574–591.
37 Hubel, D. H., and Wiesel, T. N. Receptive fields binocular
interaction and functional architecture in the cat's visual cortex.
The Journal of Physiology 160, 1 (jan 1962), 106–154.
38 Ikeda, K., and Matsumoto, K. High-dimensional chaotic behavior in systems with time-delayed feedback. Physica D: Nonlinear Phenomena 29, 1-2 (nov 1987), 223–235.
39 Jaeger, H. The echo state approach to analysing and training
recurrent neural networks. Tech. rep., 2001.
40 Jaeger, H. Echo state network. Scholarpedia 2, 9 (2007), 2330.
41 Jazayeri, M., and Shadlen, M. N. A Neural Mechanism for
Sensing and Reproducing a Time Interval. Current Biology 25,
20 (oct 2015), 2599–2609.
42 Jones, E. G. Functional subdivision and synaptic organization
of the mammalian thalamus. Int Rev Physiol 25 (1981), 173–
245.
43 Jones, E. G. Principles of Thalamic Organization. In The
Thalamus. Springer US, 1985, pp. 85–149.
44 Jones, E. G. Viewpoint: the core and matrix of thalamic organization. Neuroscience 85, 2 (apr 1998), 331–345.
45 Kakei, S., Na, J., and Shinoda, Y. Thalamic terminal morphology and distribution of single corticothalamic axons originating from layers 5 and 6 of the cat motor cortex. The Journal
of Comparative Neurology 437, 2 (Aug 2001), 170–185.
46 Karten, H. J. Evolutionary developmental biology meets the
brain: The origins of mammalian cortex. Proceedings of the
National Academy of Sciences 94, 7 (apr 1997), 2800–2804.
47 Komura, Y., Nikkuni, A., Hirashima, N., Uetake, T., and
Miyamoto, A. Responses of pulvinar neurons reflect a subject's
confidence in visual categorization. Nature Neuroscience 16, 6
(may 2013), 749–755.
48 Krishna, S., Maslov, S., and Sneppen, K. UV-Induced Mutagenesis in Escherichia coli SOS Response: A Quantitative
Model. PLoS Computational Biology 3, 3 (2007), e41.
49 Kung, H.-T., Luccio, F., and Preparata, F. P. On Finding
the Maxima of a Set of Vectors. Journal of the ACM 22, 4 (oct
1975), 469–476.
50 Kuramoto, E., Pan, S., Furuta, T., Tanaka, Y. R., Iwai, H.,
Yamanaka, A., Ohno, S., Kaneko, T., Goto, T., and Hioki,
H. Individual mediodorsal thalamic neurons project to multiple
areas of the rat prefrontal cortex: A single neuron-tracing study
using virus vectors. Journal of Comparative Neurology 525, 1
(jul 2017), 166–185.
51 Laughlin, S. B., and Sejnowski, T. J. Communication in
neuronal networks. Science 301 (Sep 2003), 1870–4.
52 Lennie, P. The Cost of Cortical Computation. Current Biology
13, 6 (mar 2003), 493–497.
53 Lin, H. W., Tegmark, M., and Rolnick, D. Why Does Deep
and Cheap Learning Work So Well?
Journal of Statistical
Physics 168, 6 (jul 2017), 1223–1247.
54 Ma, W. J., and Jazayeri, M. Neural Coding of Uncertainty
and Probability. Annual Review of Neuroscience 37, 1 (jul
2014), 205–220.
55 Maass, W., Natschlaeger, T., and Markram, H. Computational Models for Generic Cortical Microcircuits. In Computational Neuroscience. Chapman and Hall/CRC, oct 2003.
56 Maass, W., Natschläger, T., and Markram, H. Real-Time
Computing Without Stable States: A New Framework for Neural Computation Based on Perturbations. Neural Computation
14, 11 (nov 2002), 2531–2560.
57 Mante, V., Sussillo, D., Shenoy, K. V., and Newsome,
W. T. Context-dependent computation by recurrent dynamics in prefrontal cortex. Nature 503, 7474 (nov 2013), 78–84.
58 Marblestone,
A. H., Wayne, G., and Kording, K. P. Toward
an Integration of Deep Learning and Neuroscience. Frontiers in
Computational Neuroscience 10 (sep 2016).
59 Marenco, S., Stein, J. L., Savostyanova, A. A., Sambataro,
F., Tan, H.-Y., Goldman, A. L., Verchinski, B. A., Barnett, A. S., Dickinson, D., Apud, J. A., Callicott, J. H.,
Meyer-Lindenberg, A., and Weinberger, D. R. Investigation of Anatomical Thalamo-Cortical Connectivity and fMRI
Activation in Schizophrenia. Neuropsychopharmacology 37, 2
(sep 2012), 499–507.
60 Mathers, L. H. The synaptic organization of the cortical projection to the pulvinar of the squirrel monkey. The Journal of
Comparative Neurology 146, 1 (sep 1972), 43–59.
61 McCormick, D. A., and von Krosigk, M.
Corticothalamic activation modulates thalamic firing through glutamate
metabotropic receptors. Proceedings of the National Academy
of Sciences 89, 7 (apr 1992), 2774–2778.
62 Mesulam, M. M. Large-scale neurocognitive networks and distributed processing for attention, language, and memory. Ann
Neurol 28 (Nov 1990), 597–613.
63 Middleton, F., and Strick, P. L. Anatomical evidence for
cerebellar and basal ganglia involvement in higher cognitive
function. Science 266, 5184 (oct 1994), 458–461.
64 Mitchell, A. The mediodorsal thalamus as a higher order thalamic relay nucleus important for learning and decision-making.
Neurosci Biobehav Rev 54 (Jul 2015), 76–88.
65 Mitelman, S. A., Byne, W., Kemether, E. M., Hazlett,
E. A., and Buchsbaum, M. S. Metabolic Disconnection Between the Mediodorsal Nucleus of the Thalamus and Cortical
Brodmann’s Areas of the Left Hemisphere in Schizophrenia.
American Journal of Psychiatry 162, 9 (sep 2005), 1733–1735.
66 Mumford, D. On the computational architecture of the neocortex. I: The role of the thalamo-cortical loop. Biological Cybernetics 65, 2 (jun 1991), 135–145.
67 Mumford, D. On the computational architecture of the neocortex. II The role of cortico-cortical loops. Biological Cybernetics
66, 3 (jan 1992), 241–251.
68 Nair, A., Treiber, J. M., Shukla, D. K., Shih, P., and
Müller, R.-A. Impaired thalamocortical connectivity in autism
spectrum disorder: a study of functional and anatomical connectivity. Brain 136, 6 (jun 2013), 1942–1955.
69 Nakajima, M., and Halassa, M. M. Thalamic control of functional cortical connectivity. Current Opinion in Neurobiology
44 (jun 2017), 127–131.
70 Navlakha, S., and Bar-Joseph, Z. Distributed information
processing in biological and computational systems. Communications of the ACM 58, 1 (dec 2014), 94–102.
71 Newell, A. Some problems of basic organization in problem
solving programs, 1962.
72 Nii, P. The Blackboard Model of Problem Solving and the
Evolution of Blackboard Architectures. The AI Magazine 7, 2
(1986), 38–53.
73 Niven, J. E., and Laughlin, S. B. Energy limitation as a
selective pressure on the evolution of sensory systems. Journal
of Experimental Biology 211, 11 (jun 2008), 1792–1804.
74 Northcutt, R. G. Evolution of the Telencephalon in Nonmammals. Annual Review of Neuroscience 4, 1 (mar 1981), 301–350.
75 Northcutt, R. G., and Kaas, J. H. The emergence and evolution of mammalian neocortex. Trends in Neurosciences 18, 9
(sep 1995), 373–379.
76 Ohno, S., Kuramoto, E., Furuta, T., Hioki, H., Tanaka,
Y., Fujiyama, F., Sonomura, T., Uemura, M., Sugiyama, K.,
and Kaneko, T. A morphological analysis of thalamocortical
axon fibers of rat posterior thalamic nuclei: a single neuron
tracing study with viral vectors. Cereb Cortex 22 (Dec 2012),
2840–57.
77 Parnaudeau, S., Bolkan, S. S., and Kellendonk, C. The
Mediodorsal Thalamus: An Essential Partner of the Prefrontal
Cortex for Cognition. Biological Psychiatry (nov 2017).
78 Parnaudeau, S., Taylor, K., Bolkan, S. S., Ward, R. D.,
16
Balsam, P. D., and Kellendonk, C. Mediodorsal Thalamus
Hypofunction Impairs Flexible Goal-Directed Behavior. Biological Psychiatry 77, 5 (mar 2015), 445–453.
79 Partlow, G. D., Colonnier, M., and Szabo, J. Thalamic projections of the superior colliculus in the rhesus monkey,Macaca
mulatta. A light and electron microscopic study. The Journal
of Comparative Neurology 171, 3 (feb 1977), 285–317.
80 Pinault, D. The thalamic reticular nucleus: structure function
and concept. Brain Research Reviews 46, 1 (aug 2004), 1–31.
81 Purushothaman, G., Marion, R., Li, K., and Casagrande,
V. Gating and control of primary visual cortex by pulvinar. Nat
Neurosci 15 (Jun 2012), 905–12.
82 Raczkowski, D., and Fitzpatrick, D. Terminal arbors of individual physiologically identified geniculocortical axons in the
tree shrew's striate cortex. The Journal of Comparative Neurology 302, 3 (dec 1990), 500–514.
83 Rigotti, M., Barak, O., Warden, M. R., Wang, X.-J., Daw,
N. D., Miller, E. K., and Fusi, S. The importance of mixed
selectivity in complex cognitive tasks. Nature 497, 7451 (may
2013), 585–590.
84 Rigotti, M., Rubin, D. B. D., Wang, X.-J., and Fusi, S.
Internal representation of task rules by recurrent dynamics: the
importance of the diversity of neural responses. Frontiers in
Computational Neuroscience 4 (2010).
85 Rouiller, E. M., and Welker, E. A comparative analysis
of the morphology of corticothalamic projections in mammals.
Brain Research Bulletin 53, 6 (dec 2000), 727–741.
86 Rovo, Z., Ulbert, I., and Acsády, L. Drivers of the primate
thalamus. J Neurosci 32 (Dec 2012), 17894–908.
87 Salas, C., Broglio, C., and Rodrı́guez, F. Evolution of Forebrain and Spatial Cognition in Vertebrates: Conservation across
Diversity. Brain Behavior and Evolution 62, 2 (2003), 72–82.
88 Schmitt, L. I., Wimmer, R. D., Nakajima, M., Happ, M.,
Mofakham, S., and Halassa, M. M. Thalamic amplification
of cortical connectivity sustains attentional control. Nature 545,
7653 (may 2017), 219–223.
89 Scott, B. B., Constantinople, C., Akrami, A., Hanks,
T. D., Brody, C. D., and Tank, D. W. Fronto-parietal Cortical Circuits Encode Accumulated Evidence with a Diversity of
Timescales. Neuron 95 (Jul 2017), 385–398.e5.
90 Selfridge, O. G. Pandemonium: a paradigm for learning in .
In Proceedings of the Symposium on Mechanisation of Thought
Processes (London, nov 1959), D. V. Blake and A. M. Uttley,
Eds., National Physical Laboratory, pp. 513–526.
91 Seoane, L. F., and Solé, R. Phase transitions in Pareto optimal complex networks. Physical Review E 92, 3 (sep 2015).
92 Sherman, S. M. Thalamus plays a central role in ongoing cortical functioning. Nature Neuroscience 16, 4 (mar 2016), 533–541.
93 Sherman, S. M., and Guillery, R. W. Functional Connections of Cortical Areas. The MIT Press, aug 2013.
94 Shoham, S., O’Connor, D. H., and Segev, R. How silent is
the brain: is there a “dark matter” problem in neuroscience?
Journal of Comparative Physiology A 192, 8 (mar 2006), 777–
784.
95 Simon,
H. A. The Architecture of Complexity. Proceedings of
the American Philosophical Society 106, 6 (1962), 467–482.
96 Simon, H. A. The Sciences of the Artificial. MIT Press, 1969,
ch. The Architecture of Complexity.
97 Steriade, M., Domich, L., and Oakson, G. Reticularis thalami neurons revisited: activity changes during shifts in states
of vigilance. J Neurosci 6 (Jan 1986), 68–81.
98 Steriade, M., and Llinás, R. R. The functional states of the
thalamus and the associated neuronal interplay. Physiological
Reviews 68, 3 (jul 1988), 649–742.
99 Steriade, M., and Pare, D. Morphology and electroresponsive
properties of thalamic neurons. In Gating in Cerebral Networks.
Cambridge University Press, 2007, pp. 1–26.
100 Steriade, M., Parent, A., and Hada, J. Thalamic projections
of nucleus reticularis thalami of cat: A study using retrograde
transport of horseradish peroxidase and fluorescent tracers. The
Journal of Comparative Neurology 229, 4 (nov 1984), 531–547.
101 Szekely, P., Sheftel, H., Mayo, A., and Alon, U. Evolutionary Tradeoffs between Economy and Effectiveness in Biological
Homeostasis Systems. PLoS Computational Biology 9, 8 (aug
2013), e1003163.
102 Urbain, N., and Deschênes, M. Motor Cortex Gates Vibrissal
Responses in a Thalamocortical Projection Pathway. Neuron
56, 4 (nov 2007), 714–725.
103 Wang, J., Narain, D., Hosseini, E. A., and Jazayeri, M.
Flexible timing by temporal scaling of cortical responses. Nature
Neuroscience (dec 2018).
104 Wang, X.-J.
Neural dynamics and circuit mechanisms of
decision-making. Current Opinion in Neurobiology 22, 6 (dec
2012), 1039–1046.
105 Wei, W., and Wang, X.-J. Inhibitory Control in the CorticoBasal Ganglia-Thalamocortical Loop: Complex Regulation and
Interplay with Memory and Decision Processes. Neuron 92, 5
(dec 2016), 1093–1105.
106 Wiecki, T. V., and Frank, M. J. A computational model of
inhibitory control in frontal cortex and basal ganglia. Psychological Review 120, 2 (2013), 329–355.
107 Wong, K.-F., and Wang, X.-J. A Recurrent Network Mechanism of Time Integration in Perceptual Decisions. Journal of
Neuroscience 26, 4 (jan 2006), 1314–1328.
108 Woodward, N. D., Giraldo-Chica, M., Rogers, B., and
Cascio, C. J. Thalamocortical Dysconnectivity in Autism Spectrum Disorder: An Analysis of the Autism Brain Imaging Data
Exchange. Biological Psychiatry: Cognitive Neuroscience and
Neuroimaging 2, 1 (jan 2017), 76–84.
109 Yamins, D. L. K., and DiCarlo, J. J. Using goal-driven deep
learning models to understand sensory cortex. Nature Neuroscience 19, 3 (feb 2016), 356–365.
110 Yamins, D. L. K., Hong, H., Cadieu, C. F., Solomon, E. A.,
Seibert, D., and DiCarlo, J. J. Performance-optimized hierarchical models predict neural responses in higher visual cortex.
Proceedings of the National Academy of Sciences 111, 23 (may
2014), 8619–8624.
| 9cs.NE
|
1
On the Design of Multi-Dimensional Irregular
Repeat-Accumulate Lattice Codes
arXiv:1710.01475v1 [cs.IT] 4 Oct 2017
Min Qiu, Student Member, IEEE, Lei Yang, Member, IEEE, Yixuan Xie, Member, IEEE,
and Jinhong Yuan Fellow, IEEE,
Abstract—Most multi-dimensional (more than two dimensions)
lattice partitions only form additive quotient groups and lack
multiplication operations. This prevents us from constructing
lattice codes based on multi-dimensional lattice partitions directly
from non-binary linear codes over finite fields. In this paper,
we design lattice codes from Construction A lattices where
the underlying linear codes are non-binary irregular repeataccumulate (IRA) codes. Most importantly, our codes are based
on multi-dimensional lattice partitions with finite constellations.
We propose a novel encoding structure that adds randomly
generated lattice sequences to the encoder’s messages, instead of
multiplying lattice sequences to the encoder’s messages. We prove
that our approach can ensure that the decoder’s messages exhibit
permutation-invariance and symmetry properties. With these two
properties, the densities of the messages in the iterative decoder
can be modeled by Gaussian distributions described by a single
parameter. With Gaussian approximation, extrinsic information
transfer (EXIT) charts for our multi-dimensional IRA lattice
codes are developed and used for analyzing the convergence
behavior and optimizing the decoding thresholds. Simulation
results show that our codes can approach the unrestricted
Shannon limit within 0.46 dB and outperform the previously
designed lattice codes with two-dimensional lattice partitions and
existing lattice coding schemes for large codeword length.
Index Terms—Lattice codes, multi-dimensional lattices, nonbinary irregular repeat-accumulate (IRA) codes, Hurwitz integers, extrinsic information transfer (EXIT) charts.
I. I NTRODUCTION
ATTICES are effective arrangements of equally spaced
points in Euclidean space. They have attracted considerable attentions in the coding community because their
appealing algebraic structures can be efficiently exploited for
encoding and decoding. Although Shannon has shown that the
optimal coding strategy to achieve Gaussian channel capacity
is random coding with Gaussian distribution [2], these random
codes are more or less prohibited in practice. Lattice codes can
be deemed as a natural alternative to random Gaussian codes.
The remarkable work [3] has proved the existence of lattice
codes achieving the capacity of additive white Gaussian noise
(AWGN) channels by using a lattice decoder. This decoder is
suboptimal compared with the optimal maximum-likelihood
(ML) decoder but has a lower decoding complexity. Apart
from point-to-point communications, lattice codes have also
been proved to be useful in a wide range of applications
L
M. Qiu, L. Yang, Y. Xie and J. Yuan are with the School of Electrical Engineering and Telecommunications, University of New South
Wales, Sydney, NSW, 2052 Australia (e-mail: [email protected];
[email protected]; [email protected]; [email protected]).
This paper was presented in part at the IEEE International Symposium on
Information Theory (ISIT) in 2017 [1].
such as index coding [4], cooperative communications [5],
multiple access [6], multiple antenna systems [7] and so on. It
is believed that lattice codes will play a crucial role in future
communication systems.
A. Literatures and Motivations
There is tremendous work on lattice codes which mainly include information theoretical analysis of their capacity achieving properties in different communication systems and lattice
codes construction for practical systems. We focus on the code
construction.
According to the literature, there are two main approaches
to construct lattice codes. The first one is to construct lattice
codes directly in the Euclidean space. There are two wellknown examples: low-density lattice codes (LDLC) [8] and
convolutional lattice codes (CLC) [9]. Another approach is
to adapt modern capacity approaching error correction codes
to construct lattices, i.e., low-density parity-check (LDPC)
lattices [10]–[12] and polar lattices [13]. Their construction
methods involve some well-known methods such as Construction A [14] (constructing lattices based on a linear
code), Construction D [14] (constructing lattices based on
the generator matrices of a series of nested linear codes),
and Construction D’ [14] (constructing lattices based on the
parity check matrices of a series of nested linear codes).
These methods allow one to construct lattice codes not only
with good error performance inherited from capacity-achieving
linear codes, but also having relatively lower construction
complexity compared with LDLCs and CLCs. To sum up, most
of the aforementioned designs have been shown to approach
the Poltyrev limit [15] (i.e., the channel capacity without either
power limit or restrictions on signal constellations) within 1
dB when the codeword length is long enough. In addition,
all of these lattices can be decoded with efficient decoding
algorithms.
However, for LDLCs, in order to attain the best possible
decoding performance, the decoder would have to take the
whole probability density functions (pdf) for processing. This
would require a significant amount of memory. As reported in
[9], the symbol error rate (SER) of the CLCs is higher than that
of LDLCs. Both of these two lattice coding schemes are still
difficult to implement in practice due to the use of non-integer
lattice constellations. The LDPC lattices in [10] and the polar
lattices [13] involve multilevel coding and multistage decoding
due to their construction methods. This poses a much higher
complexity in encoding and decoding than that of Construct
A lattices in [11] and [12].
2
Since most of the available designs are based on infinite
lattice constellations, their error performances are compared
against Poltyrev limit. To put these lattice codes into practice,
a power constraint must be satisfied. Moreover, most lattice
codes have high complexity encoding structures due to the
sparseness of their parity-check matrices which in general can
lead to high-density generator matrices. Furthermore, most
of the Construction A, Construction D and Construction D′
lattice codes are designed based on one or two-dimensional
(real dimension) lattice partitions. It is understood that this can
result in a shaping loss in error performance compared with
using higher-dimensional lattice partitions [16]. Constructing
codes over multi-dimensional lattices have been investigated
in [17]–[20]. In [17] and [19], the authors proposed a method
for constructing lattices over number fields and have studied
their application in wiretap block fading channels. In [18], the
authors have proposed a lattice construction method to allow
Construction A lattices equipped with multiplication, which
has potential application in nonlinear distributed computing
over a wireless network. In [20], the authors have designed
lattices to obtain diversity orders in block fading channels.
However, [17]–[20] mainly focused on constructing lattices
over algebraic number fields with applications to block fading channels while designing lattice codes to approach the
unrestricted Shannon limit (i.e., when transmission is power
limited but not restricted to any signal constellation) was not
taken into account.
Recently, we have designed irregular repeat-accumulate
(IRA) lattice network codes with finite constellations for twoway relay channels (TWRC) in [21]. The lattice codes are
constructed via Construction A on non-binary IRA codes. We
have used the extrinsic information transfer (EXIT) charts
to optimize the degree distribution in a bid to minimize the
required decoding signal-to-noise ratio (SNR). However, this
scheme is based on two-dimensional lattice partitions and thus
still has a performance gap to the unrestricted Shannon limit.
B. Problem Statement
In light of the previous work, we aim to design multidimensional lattice codes to further approach the channel
capacity. That being said, directly extending the design in
[21] to multi-dimensional lattice partitions is very challenging.
There are two fundamental reasons why this is the case.
First, in the previous setting, we employed a two-dimensional
lattice partition to form a quotient ring which is isomorphic
to a finite field. However, most multi-dimensional lattice
partitions form additive quotient groups where addition is
the only group operation. If we use multi-dimensional lattice
partitions in our previous design, the multiplication between
two lattice points cannot be performed on additive groups.
Second, simply removing the multiplication in the encoding structure will prevent us from analysing and optimizing
the multi-dimensional IRA lattice codes effectively. In the
previous design, the encoder’s messages are multiplied by
some randomly generated sequences so that the permutationinvariant property [22] can be obtained. Under this property,
the analysis and optimization of our lattice codes can be
significantly simplified. It is possible to remove all the operations of multiplying random sequences to allow the use of
multi-dimensional lattice partitions. However, the permutationinvariance property will not hold in this case. As a result, the
densities of the messages in the iterative decoder can only
be represented by a multivariate Gaussian distribution. This
will lead to an extremely high complexity for our design and
analysis.
C. Main Contributions
In this paper, we aim to design multi-dimensional IRA
lattice codes with finite constellations to further approach
the unrestricted Shannon limit. This is different from most
lattice codes which are based on infinite constellations in
the literature. Even though these codes have been shown to
approach the Poltyrev limit within 1 dB, it is still unclear
whether these codes with power constraint can approach the
unrestricted Shannon limit within 1 dB. In order to practically
approach the unrestricted Shannon limit, we must optimize
the degree distribution of our codes based on constellations,
detection methods and decoding algorithms. Furthermore, we
continue to use Construction A as it has been proved to be a
simple and powerful tool for constructing capacity-achieving
lattice codes [23]. The main contributions of our work are
summarized as below:
• We designed practical lattice codes with finite constellations based on multi-dimensional lattice partitions. More
specifically, we proposed a novel encoding structure that
adds random lattice sequences to the encoder’s messages
(output of the interleaver, combiner and accumulator). In
addition, we introduced a constraint on the random lattice
sequences in our encoder and proved that the constraint
can lead to linearity of our codes. Since no multiplication
is required in our encoder, our design can be directly
applied to any lattices of any dimensions.
• We investigated the optimal degree distributions of our
lattice codes, aiming at approaching the unrestricted
Shannon limit. We proved and showed that our encoding
structure can produce permutation-invariant and symmetric effects in the densities of the decoder’s messages
(soft information propagated in the iterative docoder).
These two properties enable to use a Gaussian distribution characterised by a single parameter to model the
soft information propagated inside the iterative decoder.
Under this condition, we used a two-dimensional EXIT
chart to analyse the convergence behaviour of the iterative
decoder. With EXIT charts, we designed a set of lattice
codes for different target code rates with the minimum
decoding threshold.
• Numerical results are provided and show that our designed and optimised lattice codes can approach the unrestricted Shannon limit within 0.46 dB. We demonstrate
that our lattice codes not only outperforms previously designed lattice codes in [21] with two-dimensional lattice
partitions, but also have less coding loss compared with
the existing lattice coding schemes in [23]–[27] for large
codeword length, i.e., a codeword has more than 10,000
symbols.
3
The modulo-lattice operation is represented as:
D. Structure of the Paper
The rest of the paper is organised as follows. Section II
provides some background knowledge of lattices and lattice
codes. In Section III we present our lattice coding design
including the construction of our lattice codes, the design of
the encoder and decoder, as well as a design example of
employing the D4 lattice partition. Next in Section IV we
explain how to model the soft information in the decoder. Most
notably, we prove and show that our proposed lattice codes
can achieve permutation-invariance and symmetry properties
in the densities of the decoder’s messages. The complete
proof of all the theorems and lemmas is in Appendix. We
also provide the convergence analysis by using EXIT chart
and explain the use of EXIT chart curve fitting techniques
to design our codes with optimal degree distributions in this
section. The simulation results in Section V show the goodness
of our proposed codes compared with the codes with twodimensional lattice partitions. Finally, this paper finishes with
Section VI summarizing our main achievements from this
work.
II. BACKGROUND
ON
L ATTICES AND L ATTICE C ODES
In this section, we provide some essential definitions in
relation to lattices [14] and lattice codes [28]. All of these
will be used throughout the rest of the paper. Note that all
the concepts below are introduced based on real-dimensional
lattices. Complex lattices can be defined in a similar manner
as real lattices and thus will not be explicitly introduced here.
An n-dimensional lattice Λ is a discrete set of points λ in
Rn . It can be generated from an n × n full rank generator
matrix GΛ with real entries which can be either integers or
non-integers:
Λ = {λ = bGΛ, b ∈ Zn },
(1)
Note that we have restricted our definition to full-rank lattices
because we do not need to treat lower-rank lattices for the
purposes of this work. Here, λ is a lattice point with dimension
n or it can be deemed as a lattice codeword with length n. All
the lattices must contain the origin 0.
Lattices are groups that are closed under addition:
∀λ1, λ2 ∈ Λ, λ1 + λ2 ∈ Λ.
(2)
The Voronoi region associated with the lattice point λ is
defined as:
VΛ (λ) = {x ∈ Rn, kx − λk ≤ kx − λ′ k, ∀λ′ ∈ Λ}.
(3)
The fundamental Voronoi region VΛ (0) is the Voronoi region
associated with the all-zero lattice point.
A lattice quantizer or a lattice decoder with respect to the
lattice Λ is denoted by QΛ (x). It maps a point x in Rn to its
closest lattice point:
QΛ (x) = arg min kx − λk.
(4)
λ ∈Λ
Recall the definition of the Voronoi region from above, if
x ∈ VΛ (λ), then we have the following:
QΛ (x) = λ ∈ Λ.
(5)
x mod Λ = x − QΛ (x).
(6)
It is the difference between a vector and its closest lattice
point. So the output of this operation is always a point in the
Voronoi region VΛ (λ).
We denote the modulo-lattice addition with respect to Λ by
“⊕” where
λ1 ⊕ λ2 = (λ1 + λ2 ) mod Λ′, λ1, λ2 ∈ Λ.
(7)
Similarly, we define the modulo-lattice subtraction “⊖” as
follows:
λ1 ⊖ λ2 = (λ1 − λ2 ) mod Λ′, λ1, λ2 ∈ Λ.
(8)
A sublattice Λ′ of a lattice Λ is a subset of the lattice Λ
that is a lattice itself. We say Λ′ is nested in Λ if Λ′ ⊆ Λ.
The lattice Λ is called fine lattice while its subset Λ′ is called
coarse lattice. The lattice partition is formed by:
Λ/Λ′ = {λ + Λ′, λ ∈ Λ}.
(9)
Note that for each λ ∈ Λ, the set λ + Λ′ is a coset of Λ′ in
Λ. The point λ mod Λ′ is called the coset leader of λ + Λ′ .
The number of cosets or the cardinality of Λ/Λ′ is denoted
by M and calculated as:
M = |Λ/Λ′ | = Vol(Λ′)/Vol(Λ),
(10)
where Vol(Λ) is the volume of the lattice Λ and can be
calculated as Vol(Λ) = |det(GΛ )|. We denote the set of coset
leaders by Ψ = {ψ0, ψ1, . . . , ψ M−1 }.
A nested lattice code L is defined as the set of all coset
leaders in the lattice partition Λ/Λ′ . In other words, it takes
all the lattice points inside the fundamental Voronoi region of
the coarse lattice Λ′ :
L = Λ ∩ VΛ′ (0).
(11)
Due to this geometry property, the fundamental Voronoi
region VΛ′ (0) is also called the shaping region. Shaping is
essential in designing practical lattice codes because a finite
section of the lattice points must be selected to satisfy a
transmission power constraint for a communication system.
Denote the code rate of the nested lattice code by R.
The code rate is measured in bit per dimension and can be
calculated as:
1
(12)
R = log2 (M),
n
where n is the dimension of the lattice and M is the cardinality
defined in (10).
We now look at some figures of merit that used to measure
the goodness of the lattices. Particularly, we focus on the
shaping performance of the lattices. First of all, we define
the second moment P(Λ) as the average energy per dimension
of a uniform distribution over the fundamental Voronoi region
VΛ (0):
∫
1
kxk 2 dx.
(13)
P(Λ) =
nVol(Λ) VΛ (0)
4
The normalised second moment (NSM) of lattice Λ is
defined as:
P(Λ)
G(Λ) =
.
(14)
2
Vol(Λ) n
The shaping gain γs (Λ) is defined as the energy gain by
achieving the reduction of the average energy of a lattice
constellation compared with the constellation points that form
an n-dimensional cube. It can be calculated as:
1/12
,
(15)
γs (Λ) =
G(Λ)
1
where 12
is the NSM of an n-dimensional cubic lattice which
is deemed as the baseline. A lattice with a smaller normalised
second moment is always desirable as its shaping gain is
higher. When the dimension approaches infinite, there exist
a sequence of lattices that can achieves the optimal shaping
gain:
πe
.
(16)
lim γs (Λn ) =
n→∞
6
III. M ULTI -D IMENSIONAL IRA L ATTICE C ODES
In this section we present the proposed multi-dimensional
IRA lattice codes. We consider the channel to be a complex
AWGN channel where the input is non-binary, which means
asymmetric-output in general. For this channel, different transmitted symbols have different error resistance to the nonbinary AWGN noise. Thus the decoding errors for different
symbols are different.
A. IRA Lattices Construction
We begin with the construction of our lattice codes. The
lattice codes are constructed via Construction A [14]. The error
performance of Construction A lattices heavily depends on
the underlying error correction codes. Thus we choose IRA
codes as they have been shown to be capacity approaching in
AWGN channels and has lower encoding complexity than that
of general LDPC codes [29]–[33].
In this work, we extend the conventional Construction A
method to a more generic case which is not merely limited
to two-dimensional lattices. Denote a non-binary IRA codes
over GF(p M ) by C, where p is a prime number and M is
a positive integer. The IRA encoder takes length K input
messages and produces length N codewords. Here K ≤ N
and all the encoding operations are over GF(p M ). We denote
the Construction A lattice by Λ C . It can be generated via:
N
Λ C = {λ = φ(C) + ξR },
(17)
K 1
· log2 (p M ),
(19)
N n
where n is the dimension of the R-lattice.
We now present a specific design example of using the D4
lattice via Construction A. According to [14], the D4 lattice
is a four-dimensional lattice which has the highest sphere
packing density in the four-dimensional space. It is defined
as:
Õ4
4
xi ∈ 2Z .
(20)
D4 = (x1, x2, x3, x4 ) ∈ Z :
R=
i=1
It has the generator matrix in the integer lattice form:
0
0
0
0
.
(21)
G D4
−1 0
1 −1
As explained in Section II, we use the NSM as the goodness
to measure the shaping performance of the lattices. By (14),
we calculate the NSM for D4 is about 0.0766. Then using (15)
we can see that D4 can provide a shaping gain about 0.3657
dB over the four dimensional cubic lattice.
According to [4]. the D4 lattice can be identified as Hurwitz
quaternion integers:
1
H = a + bi + c j + dk|a, b, c, d ∈ Z or a, b, c, d ∈ Z + , (22)
2
−1 −1
1 −1
=
1
0
0
0
where {1, i, j, k} is the basis of the number system for representing Hurwitz integers. Addition in H is component wise
whereas multiplication is non-commutative and defined based
on the following relations:
i 2 = j 2 = k 2 = i j k = −1.
(23)
Given A = a + bi + c j + dk, the norm of A is:
N(A) = a2 + b2 + c2 + d 2 ∈ Z.
(24)
Consider the following example. In (18), if we let ξ = 1+2i,
then the homomorphism mapping function becomes:
φ : F25 → H/(1 + 2i)H.
(25)
Note that this lattice partition can be further expressed as:
where ξ ∈ R and R is a lattice; φ(.) is a homomorphism
mapping function that maps each codeword component to the
elements in the lattice partition:
φ : F pM → R/ξR.
field. In most cases where R is a multi-dimensional lattice,
the lattice partition forms a quotient group [18].
In (18), the R-lattice is partitioned into p M numbers of
cosets where each coset has a coset leader. For designing
finite constellations, only coset leaders are used in transmission
to satisfy the power constraint requirement. Therefore, using
(12), the information rate R for this Construction A lattice is
(18)
Note that N in (17) should be a multiple of M in (18).
It is also noteworthy that in conventional Construction A,
R can be any principal ideal domains (PID) such as rational
integers Z and Gaussian integers Z[i]. In that case, the lattice
partition forms a quotient ring that is isomorphic to a finite
H/(1 + 2i)H = λ/(1 + 2i)D4, (λ ∈ D4 )
(6)
= λ − Q(1+2i)D4 (λ)
λ
(a)
,
= λ − (1 + 2i)Q D4
(1 + 2i)
(26)
where (a) follows [28, Eq. (2.43)]. The multiplication and
division here should follow quaternion arithmetic [34]. For
the quantizer Q D4 , we follow the approach in [35] to develop
the quantization algorithm of finding the closest D4 lattice
5
B. IRA Lattice Encoder and Its Linearity
1) IRA Lattice Encoder: Here we show our proposed
encoder design. The block diagram of the IRA lattice encoder
is depicted in Fig. 2. First of all, the input to the encoder is a
z1
u1
u2
z2
Repeater
s2
gL
Combiner
⊕
⊕ ⊕
g 2′ ⊕
⊕ ⊕
⊕
g1′′
g 2′′
⊕
⊕ ⊕
⊕
⊕
⊕
g1′
sN
⊕
⊕ ⊕
g′N
c0
r1
c1
⊕
x1
⊕
x2
r2
c2
⊕
cN −1
g′′N
Time-varying
Accumulator
rN −1
⊕
xN −1
⊕
xN
rN
cN
Random-coset
Fig. 2. Block diagram of the IRA lattice encoder.
⊕
⊕
zL
⊕
s1
⊕
uK
! " # $%
g1
g2
Random Interleaver
point to an arbitrary point in R4 . The quantization algorithm
has a lower computational complexity compared with ML
decoding. It is very useful in the scenario where we perform
the D4 lattice partitions. The cardinality of this partition can
be calculated by using (24) as N(1 + 2i)2 = 25. In this way,
the D4 lattice is partitioned into 25 cosets. Even though H is
a PID [36], we only have the group homomorphism as the
multiplication for H is non-commutative.
Now we compare the mutual information of the D4 lattice
with that of a two-dimensional lattice to see the performance
gain introduced by the multi-dimensional lattices. In this paper,
the two-dimensional square lattice Z2 is set to be a benchmark
for performance comparison. Note that a finite portion of the
Z2 lattice is known as a quadrature amplitude modulation
(QAM). The Z2 lattice can be identified as Gaussian integers
Z[i] = {a + bi : a, b ∈ Z}. For fair comparison, we partition
both lattices in a way such that the information rates for both
lattice partitions are the same.
Fig. 1. Capacities of H/(1 + 2i)H and Z[i]/(1 + 2i)Z[i].
We consider the examples of lattice partitions H/(1 + 2i)H
and Z[i]/(1 + 2i)Z[i], where both partitions yield the same
information rate. This is because using (12) we can obtain
the information rates for D4 and Z2 as 21 log2 (25) and log2 (5),
respectively. Here the D4 lattice can be deemed as a twodimensional complex lattice while the Z2 lattice is a onedimensional complex lattice. Therefore the dimensions n in
(12) for both lattices are 2 and 1, respectively. In other words,
the Z[i] lattice requires one time slot to transmit its lattice
point where the D4 lattice requires two time slots to transmit
a D4 lattice point.
Given SNR values, the unrestricted Shannon limit for the
AWGN channel is plotted in Fig. 1 along with the capacities of
the D4 lattice and the Z2 lattice. As observed from Fig. 1, the
curve for the D4 lattice always lies above that for the Z2 lattice.
Therefore, under the same information rate, we can construct
D4 lattice partition based IRA lattice codes that require lower
decoding SNR than any IRA lattice codes based on the Z2
lattice partitions. This is due to the advantage of shaping gain.
length K message u = [u1, u2, . . . , u K ]T , where each element
uk for k = 1, 2, . . . , K is taken from the set of coset leaders
Ψ = {ψ0, ψ1, . . . , ψp M −1 }. This message u is then fed into a
repeater and repeated according to a discrete distribution
of
Í
f1, f2, . . . , fI , where fi ≥ 0 for i = 1, 2, . . . , I and i fi = 1.
The number fi represents the fraction of message symbols are
repeated by i times. The maximum repeating times is I times,
where I ≥ 2, thus f1 = 0. After repeating, the total number of
Í
symbols becomes L = K i i fi .
Next, the repeated symbols are passed into a random
interleaver. We denote the interleaved sequence by z =
[z1, z2, . . . , z L ]T . A randomly generated sequence with the
same length g = [g1, g2, . . . , g L ]T is added to the interleaved
sequence z via z ⊕ g in an element-wise manner, where “⊕”
is the modulo-lattice addition defined in (7). Note that each
element of g is randomly and uniformly chosen from the set
of coset leaders Ψ such that a linear code constraint is met,
which will be introduced later.
The resultant symbols are combined according to a discrete
distribution of b1, b2, . . . , bJ , where b j ≥ 0 for j = 1, 2, . . . , J
Í
and j b j = 1. Here the number b j represents the fraction of
message symbols that are obtained from combining j symbols
from the output of the interleaver and the corresponding j
addition factors in g. After combining, the message sequence
becomes a length N sequence denoted by s = [s1, s2, . . . , s N ]T ,
Í
where N = L j jb j . For n = 1, ..., N, each symbol sn is
calculated as:
sn = (zan ⊕ gan ) ⊕ . . . ⊕ (zan +jn −1 ⊕ gan +jn −1 ),
(27)
where zan and zan +jn −1 represent the first and last interleaved
symbols input to the n-th combiner, respectively; gan and
gan +jn −1 are the addition factors with respect to zan and
zan +jn −1 ; jn ∈ {1, 2, . . . , J} represents the number of symbols
6
cn = (sn ⊕ (cn−1 ⊕ gn′ )) ⊕ gn′′,
C. Tanner Graph
Similar to conventional binary IRA codes in [29], our multidimensional IRA lattice codes can be represented by a Tanner
graph as shown in Fig. 3.
Elements of r are uniformly distributed over the set of coset
leaders Ψ. Before transmission, the average energy of codeword symbols is normalised to 1.
Note that although the four lattice sequences g, g ′, g′′
and r are random, they are assumed to be known at both
transmitters and receivers prior to transmission. Furthermore,
the underlying linear codes for our Construction A lattices can
be either systematic or nonsystematic non-binary IRA codes.
2) The Linearity of IRA Lattice Codes: It can be noticed that our proposed lattice encoding structure is different
from previous designs. More specifically, instead of using
the modulo-lattice multiplication between encoder messages
and random lattice sequences in [21], we use a different
approach by introducing the “⊕” operation in the encoding
process. However, this difference introduced non-linearity to
our codes if g, g′ and g ′′ are totally independent, which is
not appealing for low complexity decoding. To address this
issue, we introduce a constraint on these random sequences to
ensure the codes are linear.
Proposition 1. The multi-dimensional IRA lattice codes are
linear if the n-th output element from the encoder satisfies the
following conditions:
(30)
Proof: See Appendix A.
Note that this equation has jn + 2 elements. We randomly
choose any jn + 1 elements out of these jn + 2 elements to be
random and uniformly distributed over the set of coset leaders
Ψ. The last element is then determined by Eq. (30). One can
also notice that the linearity condition excludes the randomcoset vector r. This is because the random-coset vector is
g1′
g1′′
g1
u1
b1
c0
c1
u2
(28)
where the initial condition is given as c0 = 0. Here c0 is a
dummy parity that is fixed to 0 and will not be transmitted.
It is also noteworthy that the random vectors g, g′ and g′′ in
the encoding structure introduce and realize the permutationinvariance property on all edges of a Tanner graph as shown
in Fig. 3 and will be discussed in Section IV-A.
Finally, the output of the accumulator c adds a random-coset
vector r with length N and become the coded lattice sequence
x:
x = c ⊕ r.
(29)
gan ⊕ . . . ⊕ gan +jn −1 ⊕ gn′ ⊕ gn′′ = 0.
independent of the encoder’s messages and is always removed
before decoding. If the random-coset vector is included in
the condition, the output-symmetric effect in the non-binary
AWGN channel will vanish.
f2
f3
Random Interleaver
to be combined at the n-th combiner; an is the index of the
first interleaved symbol input to the n-th combiner. Note that
the combiner is to combine the interleaved messages in order
to satisfy the code rate requirement.
The combined message sequence s is passed into a timevarying accumulator which features a time-varying transfer
function determined by two randomly generated lattice se′ ]T and g ′′ = [g ′′, g ′′, . . . , g ′′ ]T .
quences g′ = [g1′ , g2′ , . . . , g N
N
1
2
All the elements in both sequences are uniformly distributed
over the set of coset leaders Ψ such that a linear code
constraint is met, which will be introduced later. The output
message of the time-varying accumulator is denoted by c =
[c1, c2, . . . , c N ]T . The n-th symbol cn , where n = 1, 2, . . . , N,
is generated by
b2
bJ
fI
uK
Information
Nodes
g′N
gL
Check
Nodes
g ′′N
cN −1
cN
Parity
Nodes
Fig. 3. Tanner graph of the IRA lattice codes.
The Tanner graph is a bipartite graph with variable nodes
and check nodes. In the figure, variable nodes are represented
by circles while check nodes are represented by squares. There
are N + K variable nodes on the Tanner graph. The K variable
nodes that placed on the left, are called information nodes.
They represent the K repeaters in the encoder. The degree
distribution of information nodes with degree i is denoted by
fi in the figure. This means that the fraction of information
nodes are connected to i check nodes. Note that the random
interleaver here introduces randomness in the edges between
information nodes and check nodes. This randomness can
prevent short cycles in the Tanner graph which leads to a better
decoding performance [37]. On the right of the Tanner graph,
there are N variable nodes which are called parity nodes,
representing the output c from the time-vary accumulator. In
the middle of the Tanner graph, there are N check nodes,
representing N combiners. The degree distribution of check
nodes with degree j + 2 is denoted by b j which represents the
fraction of check nodes connected to j information nodes and
2 parity nodes. Note that the random-coset vector r is removed
before performing decoding, thus it is not shown in the Tanner
graph.
Now consider the n-th check node with degree j + 2,
according to (27), (28) and the Tanner graph in Fig. 3, the
parity-check equation at the n-th check node is
(zan ⊕ gan ) ⊕ · · · ⊕ (zan +jn −1 ⊕ gan +jn −1 )⊕
(cn−1 ⊕ gn′ ) ⊕ (cn−1 ⊕ gn′′ ) = 0,
(31)
7
where cn−1 ⊕ cn = 0. Note that in the Tanner graph, c0 is a
dummy bit and will not be transmitted.
We decompose the elements on the left hand side of
Equation (31) into two vectors:
tn = [zan , . . . , zan +jn −1, cn−1, cn−1 ].
(32)
hn = [gan , . . . , gan +jn −1, gn′ , gn′′ ].
(33)
The first vector tn represents the symbols coming from the
variable nodes connected to the n-th check node. More specifically, zan , . . . , zan +jn −1 are from information nodes while cn−1
and cn−1 are from parity nodes. The second vector hn represents
the addition factors on the corresponding edges of the n-th
check nodes as shown in Fig. 3.
D. IRA Lattice Decoder
As shown in the previous section, the multi-dimensional
IRA lattice codes have a Tanner graph representation. Therefore, we can employ a modified belief prorogation (BP)
decoding algorithm to decode our lattice codes.
The decoder attempts to recover the source message u from
the noisy observation of the AWGN channel output y = x+nz ,
2 ) denotes the complex AWGN noise.
where nz ∼ CN(0, σch
Before decoding, we first need to calculate the symbol-wise a
posterior probability (APP) of each coset leader and for each
lattice codeword component xn , which is written as:
P(xn |yn ) =
p(yn |xn )p(xn )
, for n = 1, 2, . . . , N.
p(yn )
(34)
before adding the random-coset vector r. We denote the APP
vector after removing coset by P′ [n]:
P′[n] = P(cn |yn )
= [Pψ0 ⊖rn [n], Pψ1 ⊖rn [n], . . . , Pψp M −1 ⊖rn [n]]T .
(39)
where ⊖ is defined in (8). The resultant APP vector P′ [n] is
then passed into a BP decoder.
The decoder updates the information between check nodes
and variable nodes in an iterative manner. We denote the
message from the m-th variable node to the n-th check node
by r(m, n). The message passed from the n-th check node
to the m-th variable node is denoted by l(n, m). Both vectors
are probability vectors with dimension p M . Use the Tanner
graph in Fig. 3, we let A(m) and B(n) represent the set of
check nodes connected to the m-th variable node and the set of
variable nodes adjacent to the n-th check node, respectively.
Without the loss of generality, let the index of information
nodes be from 1 to K and the index of parity nodes be from
(K + 1) to (K + N) of the variable nodes. The decoding steps
can be summarized in the following.
1) Initialisation step: According to the Tanner graph in Fig.
3, the channel output must go through the parity nodes first.
Thus for all edges (m, n) between the parity nodes and the
check nodes in the Tanner graph, the initial message r(m, n)
is the channel APP in (39):
r(m, n) = P′ [m − K], for m = K + 1, . . . , K + N
n = 1, 2, . . . , N.
(40)
For the sake of simplicity, We let
Pψk [n] = P(xn = ψk |yn ),
(35)
where k = 0, 1, . . . , p M − 1 and ψk is the k-th coset leader.
Since the transmitted codeword symbol is xn = cn + rn , where
rn is uniformly distributed over Ψ, thus the distribution for xn
is also uniform over Ψ. Therefore, Eq. (35) can be written as
and
P(yn |xn = ψk )
,
Pψk [n] = Í M
p −1
P(y
|x
=
ψ
)
n
n
k
k=0
P(yn |xn = ψk ) = q
1
2
2πσch
(36)
!
√
kyn − SNRψk k 2
, (37)
exp −
2
2σch
Í p M −1
In this way, we have k=0 Pψk [n] = 1.
In (37), ψk and yn both are vectors with length equal to
the dimension of the lattice. In our design example, ψk is a
D4 lattice point with four dimensions. We perform the symbolwise maximum-likelihood detection. Considering that practical
systems can only transmit and receive one two-dimensional
signal at each time slot, the detection is a joint detection for
two two-dimensional signals.
We denote the APP vector by P[n] where
P[n] = [Pψ0 [n], Pψ1 [n], . . . , Pψp M −1 [n]]T .
(38)
Then the above APP vectors are fed into a coset remover to
obtain the APP vectors with respect to c in (29) as the message
For all edges (m, n) between the information nodes and the
check nodes in the Tanner graph, we let
rk (m, n) =
1
, for k = 0, 1, . . . , p M − 1
pM
m = 1, 2, . . . , K.
(41)
2) Update the check nodes to variable nodes messages:
For all edges (m, n) that connected to the n-th check node,
generate the probability vector l(n, m) with its k-th element
given by
Ö jn −1
Õ
rt(i)
,
(42)
lk (n, m) =
i
É j n −1
i=1
t1,...,t j n −1 ∈Ψ
i=1
(ti ⊕hi )⊕ψk ⊕(h j n )=0
É
where
is the summation performed by ⊕; jn is the degree
of the n-th check node; r (1), . . . , r (jn −1) are the incoming
messages from all the connected variable nodes except the mth variable node, i.e., {r(m ′, n) : m ′ ∈ B(n) \ m}; t1, . . . , t jn −1
are the lattice symbols from the associated variable nodes;
h j1 , h j2 , . . . , h jn −1 are the addition factors on the corresponding
edges and h jn denotes the addition factor for the edge (m, n).
Note that the calculations of the check node messages are
different from that in conventional IRA decoding as the paritycheck equations and the associated arithmetic are different.
3) Update the variable nodes to check nodes messages:
For all edges (m, n) between the variable nodes and the check
8
nodes in the Tanner graph, generate the probability vector
r(n, m) with the k-th element given by
Î jm −1 (i)
lk
γk(n) i=1
,
(43)
rk (m, n) = Í M
p −1 (n) Î jm −1 (i)
γ
l
′
′
′
i=1
k =0
k
k
where jm denotes the degree of the m-th variable node;
l(1), . . . , l(jm −1) denote the incoming messages from all the
connected check nodes except the n-th check node, i.e.,
{l(n ′, m) : n ′ ∈ A(m)/n}; γk(n) = rk (m, n) in (40) for
m = K + 1, . . . , K + N when the messages are from parity
(n)
nodes to the n-th check node and γk = rk (m, n) in (41) for
m = 1, . . . , K when the messages are from information nodes
to the n-th check node.
4) Stopping condition: For each iteration, make the hard
decision on the m-th variable node by calculating
Î jm (i)
lk
γ (n) i=1
δ̂n = arg max Í M k
,
(44)
Î
p −1 (n)
jm (i)
k
γ
l
′
′
′
k =0
k
i=1 k
for n = 1, 2, . . . , K + N. It contains information from all the
connected edges. If the hard decision results δ̂1, δ̂2, . . . , δ̂K+N
satisfy the parity-check equations in (31) or a predetermined
maximum number of iterations is reached, then stop; otherwise
go to Step 2).
The calculation in (42) has a very high computational
complexity if the cardinality of the lattice partition p M is very
large. We follow [38] to employ DFT and IDFT in our lattice
decoding process to reduce the complexity.
First we need to introduce some important notations which
will be used in the rest of the paper. Define a probability vector
as ρ = [ρψ0 , ρψ1 , . . . , ρψp M −1 ] representing the probability
of a lattice point being ψ0, ψ1, . . . , ψp M −1 . In addition, the
Í p M −1
ρψk = 1.
probability vector must satisfy ρψk ≥ 0 and k=0
Given a probability vector ρ and χ ∈ Ψ, we define the ⊕ χ
operation as the following
ρ ⊕ χ = [ρψ0 ⊕ χ, ρψ1 ⊕ χ, . . . , ρψp M −1 ⊕ χ ].
(45)
Now consider the expression in (42), an equivalent expression can be written as
⊖h j n
Ì
jn −1 (i) ⊖hi
l=
r
,
(46)
i=1
where l is the vector that contains
elements lk , k =
Ë
0, 1, · · · , p M − 1 in (42) and the “ ” operator performs the
modulo-lattice convolution between two vectors. It produces
a vector whose k-th component is:
Õ
[r (1) ⊗ r (2) ]k =
r χ(1) · rψ(2)k ⊖ χ, for k = 0, 1, . . . , p M − 1. (47)
χ ∈Ψ
This convolution can be evaluated by using M-dimensional
DFT and IDFT [39]. In this way, (46) can be evaluated as
"
Ö
# ⊕h j n
⊖hi
jn −1
,
(48)
l = IDFT
DFT r (i)
i=1
where the multiplication of the DFT vectors is performed in a
component-wise manner. A further reduction in complexity
of implementation can be obtained by using fast Fourier
transform and inverse fast Fourier transform algorithms.
E. Complexity of IRA lattice codes
In this section, the complexity of our multi-dimensional IRA
lattice codes will be investigated and compared to that of the
IRA lattice codes with two-dimensional lattice partitions.Note
that both lattice codes are built from Construction A. The
underlying linear code for our design is over F2p while the
linear codes for the design with two-dimensional lattices is
over F p [21].
First, we focus on the complexity of symbol-wise detection.
For an ML detector, the detection is based on the entire
constellation. Thus, for a two-dimensional constellation with
size q, the computational complexity is in the order of O(q).
In our design, we have a four-dimensional constellation with
size q2 , the computational complexity is O(2q2 ). The “2” here
is due to the joint detection for two two-dimensional symbols.
The computational complexity of the nonbinary BP decoding
is in the order of O(q log2 q) when FFT is employed for
check node calculations [40]. For our decoder to decode lattice
codes with four-dimensional lattice partitions, the complexity
is O(q2 log2 q2 ). Compared with our previous coding scheme
with two-dimensional lattice partitions, the complexity of the
code design in this work is 2q times higher.
For Construction A lattices, it has been shown in [3] that the
finite field size of the underlying linear code has to be large
enough to achieve the capacity. Therefore, we have traded the
complexity to attain better performance by introducing multidimensional lattice partitioned in our design.
IV. D ESIGN
AND
A NALYSIS OF M ULTI - DIMENSIONAL IRA
L ATTICE C ODES
In this paper, the analysis of our multi-dimensional IRA
lattice codes focus on the average behaviour of randomly
selected codes from an ensemble of codes. First, let αi
be the fraction of interleaver’s edges that connected to the
information nodes with degree i and let β j be the fraction of
interleaver’s edges that are connected to the check nodes with
degree j + 2. Recall in Section III-B that i = 2, 3, . . . , I and
j = 1, 2, . . . , J. The additional “2” here means every check
node has two deterministic connections from the connected
parity nodes as shown in Fig. 3. Following [29], the edge
degree distributions of our multi-dimensional IRA lattice codes
can be written as
ÕI
αi x i−1 .
(49)
α(x) =
i=2
β(x) =
ÕJ
j=1
β j x j−1 .
(50)
Given α, β, the type of lattice R and the scaling factor ξ
in (18), we define an (α, β, ξ, R) ensemble as the set of our
multi-dimensional IRA lattice codes obtained via Construction
A.
A. Modeling the Decoder’s Message Distributions
In our multi-dimensional IRA lattice codes, the soft information propagated in the iterative decoder can be modeled by
a multi-dimensional LLR vector. Even though APP is used in
our iterative decoder, it is common to use LLR in EXIT chart
9
analysis. Note that APP and LLR are different but equivalent
representations of the decoder’s soft information. In order to
track the convergence behaviour of the iterative decoding,
multi-dimensional EXIT charts may be required. However,
developing these EXIT chart functions can be very difficult.
To deal with this challenge, the new encoding structure is
proposed. We will prove that using this structure, the densities of the messages in BP decoder can attain permutationinvariance and symmetry properties. With these two properties,
the densities of the decoder’s messages can be represented as
a single parameter. In this way, our method only needs to track
one-dimensional variables rather than the true densities of
the multi-dimensional LLR vectors. In addition, the symmetry
property enables to use all-zero lattice codeword assumption
in the EXIT chart analysis. As such, the expression of mutual
information in the EXIT chart analysis can be simplified.
We first introduce some useful definitions and notations in
the following.
1) Preliminaries: Following the definition in [41], we define the LLR values for a given probability vector ρ as
ρψ0
ωψk = log
, for k = 0, 1, . . . , p M − 1.
(51)
ρψk
It is intuitive that ωψ0 = 0.
The p M -dimensional LLR vector is then defined as ω =
[ωψ0 , ωψ1 , . . . , ωψp M −1 ]T . Note that unlike most LLR definitions, we include the element ωψ0 in the LLR vectors as it is
associated with our analysis of permutation-invariance which
will be introduced shortly. When we apply the ⊕ χ operation
defined in (45) on the LLR value ωψk , we have:
ρψ0 ⊕ χ
⊕χ
= ωψk ⊕ χ − ωψ0 ⊕ χ .
(52)
ωψk = log
ρψk ⊕ χ
A p M -dimensional probability-vector random variable
is defined as P = [Pψ0, Pψ1 , . . . , Pψp M −1 ]T that only
takes valid probability values. The associated p M dimensional LLR-vector random variable is defined as
W = [Wψ0 , Wψ1 , . . . , Wψp M −1 ]T .
Now we introduce the definitions of the symmetry and
permutation-invariance properties and explain how we can
achieve these properties.
2) Symmetry: Recall in Section III-B, we add a randomcoset vector r at the end of the encoder. The random-coset
elements are randomly chosen and uniformly distributed over
the set of coset leaders Ψ. Thus we have the following
theorem.
Theorem 1. Adding a random-coset vector r to the encoder
output c, where r is uniformly distributed over Ψ, can produce the output-symmetric effect in non-binary input AWGN
channels.
Proof: See Appendix B.
Similar to the non-binary LDPC codes in [22], the LLR
random vectors are symmetric under the output-symmetric
effect. The symmetry property of an LLR random vector is
defined as follows.
Definition 1. Given an LLR random vector W and an r ∈ Ψ,
W is symmetric if and only if W satisfies
Pr[W = ω] = eωψk Pr[W = ω ⊕r ]
(53)
for all LLR vectors ω and all r ∈ Ψ.
With this property, the probability of decoding error is
equal for any transmitted codeword [22]. In other words, the
symmetry property removes the dependence of the decoder’s
LLRs on transmitted codewords [38]. Therefore, we can use
all-zero lattice codewords in our EXIT chart analysis.
3) Permutation-Invariance: We start with the definition
of permutation-invariance [42, Section 2.6] on a probabilityvector random variable. Then we will show that our approach
can achieve this property under our proposed structure.
Definition 2. A probability-vector random variable X =
[X0, X1, X2 . . .] is permutation-invariant if for any permutation
̟ of the indices such that the random vector ̟(X) =
[X̟(0), X̟(1), X̟(2), . . .] is distributed identically with X.
Under this property, all the random variables in X are identically distributed (but may not be independent). Therefore,
changing the order of the elements in X will not change the
distribution of X .
Recall in Section III-B, our codes have three randomly
generated sequences added to the encoder’s messages. This
leads to a symbol level permutation (the permutation from
a coset leader to another coset leader) on the messages.
The densities of these messages can be shown to have the
permutation-invariance property. Now, we have the following
theorem:
Theorem 2. Given a p M -dimensional probability-vector random variable P and a χ ∈ Ψ, the random vector P ⊕ χ =
[Pψ0 ⊕ χ, Pψ1 ⊕ χ, . . . , Pψp M −1 ⊕ χ ] is identically distributed with P.
Therefore P is permutation-invariant.
Proof: See Appendix C-A.
This theorem can be carried over straightforwardly to LLR
representation. Thus we have the following lemma:
Lemma 1. Let W = [Wψ0 , Wψ1 , . . . , Wψp M −1 ]T be an LLRP
ψ
vector random variable such that Wψk = log Pψ 0 , for k =
k
0, 1, . . . , p M − 1. If P is permutation-invariant, then W is also
permutation-invariant.
Proof: See Appendix C-B.
Therefore, under the BP decoding, the messages passed
within the Tanner graph of our codes satisfy all the symmetry
and permutation-invariance properties.
4) Gaussian Approximation: With the symmetry and
permutation-invariance properties, the p M -dimensional LLR
can be modeled using a multivariate Gaussian distribution
[22]:
fW (ω) =
1
(2π)
pM
2
1
exp − (ω − m)T Σ−1 (ω − m) , (54)
1
2
|Σ| 2
10
σ2
2
σ2
2
m = . and Σ =
..
σ2
2
2
σ 2
σ2
2
.
.
.
σ2
2
σ2
2
σ2
..
.
···
···
···
..
.
···
σ2
2
σ2
2
.. .
.
σ 2
(55)
More specifically, mi = σ2 for i = 1, 2, . . . , p M , and Σi, j =
2
σ 2 if i = j and σ2 otherwise. As a result, the density of
the p M -dimensional LLR is completely described by a single
parameter σ. It is worth mentioning that our definition of LLR
random vector is p M -dimensional rather than p M − 1 in the
literature. This is because the ⊕ χ operation will change the
position of Wψ0 . Thus we need to use a p M -variate Gaussian
distribution to model the p M -dimensional LLR.
B. Convergence Analysis
EXIT charts track the mutual information between the
transmit lattice symbol u and the LLR random vector W.
With the all-zero lattice codeword assumption, the mutual
information can be evaluated according to [22]
#
"
Õ M
p −1 −w
i
e
u=0 ,
(56)
I(u; W) = 1 − E logp M
i=0
where W is modeled by (54) and (55). Thus, the mutual information is a function of the single parameter σ. For simplicity,
we let J(σ) = I(u; W) as every value of σ corresponds to
a value of I(u; W). Since the mapping is bijective, we can
also define the inverse function J(.)−1 to obtain σ when given
I(u; W).
In the EXIT chart analysis, variable nodes are treated as a
component decoder while the combiners and the time-varying
accumulator together is treated as another decoder. As such,
we compute the variable-node decoder (VND) curve and the
check-node decoder (CND) curve. The argument of each curve
is denoted as I A and the value of the curve is denoted as
IE , representing a priori input and the extrinsic output of
each component decoder. The details of obtaining the transfer
functions will be explained next.
1) EXIT Function for VND: For a variable node with im
degrees, the output mutual information of the VND for this
type of variable nodes is given by [43]:
p
IE,V N D (I A, im ) ≈ J (im − 1)J −1 (I A) .
(57)
For a given VN degree distribution (i, αi ), the EXIT function
for the VND of the entire IRA code is:
ÕI
αi IE,V N D (I A; i).
(58)
IE,V N D (I A) =
i=2
2) EXIT Function for CND: For a check node with degree
jn , we use a numerical method to obtain the approximated
EXIT functions as there is no closed-form expression in the
literature.
For a given I A, we obtain the corresponding parameter using
σ = J −1 (I A). Then the input a priori LLR vectors are generated
according to (54) and (55). For a given SNR, generate the allzero lattice codeword, three random sequences g, g′ , g′′ , a
random-coset vector r and an AWGN channel noise sequence
2
with variance of σch
. We calculate the channel APPs by
following (34) to (39) and then substitute the results into (51)
to obtain the channel input LLR Wch . Given g, g ′, g′′ , r,
jn and Wch , we perform BP decoding with one iteration to
produce the output LLR. The IE,C N D (I A) associated with the
check node degree jn is obtained by substituting the output
LLR into (56).
For a given CN degree distribution ( j, β j ), the EXIT function for the CND of the entire IRA code can be obtained by:
ÕJ
β j IE,C N D (I A; j, σch ).
(59)
IE,C N D (I A, σch ) =
j=1
C. Design Examples
Based on our EXIT functions, we now employ the EXIT
chart curve fitting technique [43] to find the optimal CN and
VN degree distributions such that the area between the CN
curve and the VN curve is minimized. First, we carefully
select an appropriate CN degree distribution. Then, we fit the
EXIT curve of VND to CND by using linear programming
to optimize the degree distribution for VN. Next, we update
the CN degree distribution based on the optimized VN degree
distribution. The optimization for the degree distribution of
CN and VN are carried out in an iterative manner. Note that
we have set the minimum gap between the VND curve and
the CND curve to be greater than zero but not too large, e.g.,
0.0001. In this way, the produced VND curve do not intersect
with the CND curve and both curves create a narrow tunnel.
The number of optimization iteration is set to 10 as more
iterations does not improve the optimization results further.
An example of an EXIT chart for our multi-dimensional
IRA lattice codes over H/(1 + 2i)H with code rate of 32 is
illustrated in Fig. 7. In our design, the portion of degree 1 CN
with mean vector m and covariance matrix Σ given by
Fig. 4. EXIT Chart of optimized degree distributions for the rate
dimensional IRA lattice code.
2
3
multi-
must not be too small in order to ensure the decoder works
in the first few iterations because our codes are nonsystematic
[43]. From Fig. 4, we can see that the VND curve literally
touches the CND curve for the range [0, 1], which guarantees
successful convergence and accurate decoding threshold.
11
We have adopted the proposed approach in designing the
(α, β, 1 + 2i, H)-lattice ensemble with three code rates 43 , 32 and
1
2 . The degree distributions and the decoding thresholds are
shown in Table I.
As shown in the table, the optimized CN distributions are
degree 1 and degree 3 because this pair of CN distributions
have the lowest optimization complexity and the minimum
decoding threshold for the three code rates. We have also
designed our codes with other pairs of CN distributions, but
their performance is not much better than the code with only
degree 1 and degree 3 CNs.
Fig. 6. Symbol error rate performance of rate
2
3
codes.
! !
"# $ #%& '%' $
"#( ' )###
Fig. 7. Symbol error rate performance of rate
Fig. 5. Symbol error rate performance of rate
!
!
!
" " !
#$ % $&' (&( %
#$) ( *$$$
In this section, we present our simulation results for our
multi-dimensional IRA lattice codes over H/(1+2i)H. In order
to evaluate the average behavior of our codes, we randomly
generated a codeword from the (α, β, 1 + 2i, H) ensemble and
randomly select the values for g, g′, g′′ and r in every
channel realization. Since our coding scheme is based on finite
constellations with power constraint, the performance for three
designed code rates 43 , 32 and 12 is measured in terms of symbol
error rate (SER) versus SNR, which are depicted in Fig. 5,
Fig. 6 and Fig. 7, respectively. Based on these designed code
rates, the corresponding information rates are calculated by
using (19) as R1 = 1.741 bits/s/Hz, R2 = 1.548 bits/s/Hz
and R3 = 1.161 bits/s/Hz, respectively. The corresponding
unrestricted Shannon limit and uniform input capacity for each
information rate are plotted in each figure. Additionally, we
also show the SER performance for the previously designed
IRA lattice codes over Z[i]/(1 + 2i)Z[i] in all the figures
for comparison because both partitions result in the same
information rate. In our simulations, we set the codeword
length to be 1,000, 10,000 and 100,000 symbols whereas the
corresponding step sizes for SNR are 0.1 dB, 0.05 dB and 0.01
dB, respectively. The maximum number of decoding iterations
was set to be 200.
V. S IMULATION R ESULTS
! !
"# $ #%& '%' $
"#( ' )###
3
4
codes.
1
2
codes.
In Fig. 5, the unrestricted Shannon limit for R1 is 3.70 dB. In
this case, we observe that the gap to the unrestricted Shannon
limit at the SER of 10−5 is 0.90 dB for our rate 43 D4 -partitionbased lattice code and 1.28 dB for the code in [21]. Thus,
our newly designed four-dimensional IRA lattice code is 0.38
dB better than the lattice code with two-dimensional lattice
partitions. The unrestricted Shannon limit for R2 is 2.84 dB.
As shown in Fig. 6, the gap between our lattice code and the
unrestricted Shannon limit is 0.62 dB. For the code in [21],
the gap is 0.88 dB. Therefore, the proposed lattice code is
0.26 dB better. Fig. 7 shows that the gap to the unrestricted
Shannon limit is further reduced to 0.46 dB for our rate 21 fourdimensional IRA lattice code. Our code is 0.1 dB better than
the rate 21 two-dimensional lattice code in [21]. To this end,
our proposed codes have lower decoding thresholds than that
12
TABLE I
O PTIMAL DEGREE DISTRIBUTIONS AND DECODING THRESHOLDS OF (α, β, 1 + 2i, H)- LATTICE ENSEMBLE WITH VARIOUS CODE RATES
Rates
Thresholds
3
4
4.47 dB
2
3
3.31 dB
1
2
1.26 dB
Degree Distributions (i, αi ) for variable nodes, (j, β j ) for check nodes
α: (2,0.288274), (3,0.265333), (7,0.188119), (13,0.123885), (15,0.134389)
β: (1,0.055556), (3,0.944444)
α: (2,0.240605), (3,0.231215), (7,0.081754), (8,0.190942), (19,0.175951), (20,0.079534)
β: (1,0.053861), (3,0.946139)
α: (2,0.163689), (3,0.170788), (8,0.120858), (9,0.148837), (19,0.038618), (20,0.088323), (34,0.268886)
β: (1,0.054328), (3,0.945672)
TABLE II
C OMPARISONS OF CODING SCHEMES
Coding schemes
GLD lattices [25]
LDA lattices [23]
LDA lattices [24]
LDLCs [26]
QC-LDPC lattices [27]
IRA lattices
n [symbols]
1,000
1,000
10,000
10,008
100,008
1,000,008
1,000
10,000
100,000
1,190
30,000
1,000
10,000
100,000
Coding loss [dB]
1.3
1.36
0.7
0.55
0.36
0.3
1.5
0.8
0.6
2
1.5
1.5
0.6
0.3
of the codes in [21] but with higher encoding and decoding
complexities.
Now we compare our designed lattice codes with the lattice
coding schemes from [23]–[27] for the same codeword length.
Since these schemes are based on infinite constellations, their
performances are measured in terms of gap to the Poltyrev
limit which can be considered as coding loss [24, Section VIB]. To obtain the coding loss in our lattice coding scheme, we
measure the gap to uniform input capacity. The comparisons
are listed in Table II, showing the simulation results which
are reported for each scheme in the appropriate reference,
including codeword length and coding loss when SER is at
10−5 .
From Figs. 5-7, one can observe that our code with rate 12
have the smallest coding loss. To be more specific, the coding
loss for our lattice codes with N = 100, 000, N = 10, 000
and N = 1, 000 when SER is at 10−5 is about 0.3 dB, 0.6
dB and 1.5 dB. From Table II, it can be seen that our coding
scheme outperforms all of these schemes for large codeword
length, i.e., N ≥ 10, 000. When the codeword length is 1,000,
our code is about 0.2 dB worse compared with LDA lattices
[23] and GLD lattices [25] because of the probability of
short cycles are higher when the codeword length is small.
Since our goal is to design capacity-approaching lattice codes,
thus we mainly focus on the codes with large codeword
length, i.e., N ≥ 10, 000. Note that the direct comparison
of encoding and decoding complexities for lattice codes with
infinite constellations and our codes with finite constellations
may not be fair and thus is omitted.
It is also worth noting that the waterfall regions of our multidimensional lattice codes are within 0.14 dB to the predicted
decoding thresholds as shown in Table I for various code rates.
Therefore, it is evident that the proposed EXIT chart analysis
Gap to unrestricted Shannon limit [dB]
N/A
N/A
N/A
1.05
0.9
0.8
N/A
N/A
N/A
N/A
N/A
1.7
0.8
0.46
for our multi-dimensional lattice codes is effective.
VI. C ONCLUSION
In this paper, we designed new multi-dimensional IRA
lattice codes with finite constellations. Most compellingly,
we proposed a novel encoding structure and proved that our
codes can attain the permutation-invariance and symmetry
properties in the densities of the decoder’s messages. Under
these properties, we used two-dimensional EXIT charts to analyze the convergence behavior of our codes and to minimize
the decoding threshold. Our design can employ any higherdimensional lattice partitions. Numerical results show that our
designed and optimized lattice codes can achieve within 0.46
dB of the unrestricted Shannon limit and outperform existing
lattice coding schemes for large codeword length.
A PPENDIX A
P ROOF OF P ROPOSITION 1
We divide our encoder into two parts: the first part is from
the input of the repeater to the output of the interleaver; the
second part is from the input of the combiner to the output
of the accumulator. To prove that our codes are linear codes,
we only need to show that the second part is a linear system.
This is because the first part is already linear.
A linear code has the linear property such that the linear
combination of two codewords is still a valid codeword. Now
suppose we have two different codewords Xτ and Xυ with
length N. The linear combination of these two codewords is
Xτ ⊕ Xυ =[x1 [τ], x2 [τ], · · · , x N [τ]]⊕
[x1 [υ], x2 [υ], · · · , x N [υ]]
=[x1 [τ] ⊕ x1 [υ], x2 [τ]⊕
x2 [υ], · · · , x N [τ] ⊕ x N [υ]],
(60)
13
Ê jn −1
Ê jn −1
zan +i [τ] ⊕ cn−1 [τ] ⊕
zan +i [υ] ⊕ cn−1 [υ] ⊕ Cgn ⊕ Cgn
i=0
i=0
Ê jn −1
zan +i [τ] ⊕ zan +i [υ] ⊕ (cn−1 [τ] ⊕ cn−1 [υ]) ⊕ Cgn ⊕ Cgn
=
xn [τ] ⊕ xn [υ] =
i=0
where ⊕ is the modulo lattice addition. Now, we focus on
the n-th component of the codeword xn for 1 ≤ n ≤ N. The
encoding function for the n-th component of the codeword is
Ê jn −1
(61)
zan +i ⊕ gan +i ⊕ cn−1 ⊕ gn′ ⊕ gn′′ = xn,
For the left term in Eq. (67), we have
Pr[Yn < U(Xn )|Cn = ψi ]
Õ
Pr[Yn < U(Xn )|Rn = ri, Cn = ψi ]×
=
ri
i=0
where zan and zan +jn −1 represent the first and last interleaved
symbols to the n-th combiner; cn−1 is the n − 1-th output of
the time-varying accumulator. Note that the random-coset is
removed before iterative decoding, thus it is not considered as
part of the codebook information.
We can then rewrite the above equation as
Ê jn −1
zan +i ⊕ cn−1 ⊕ Cgn = xn,
(62)
i=0
É jn −1
′
′′
where
C is the
i=0 ga n +i ⊕ gn ⊕ gn = Cgn ∈ Ψ and
É jgn
n −1
constant associated with xn . Note that the term
i=0 ga n +i
can be extracted by using the associative law on the addition
of Hurwitz integers.
Now for the n-th codeword component in Xτ and Xυ , we
have
Ê jn −1
zan +i [τ] ⊕ cn−1 [τ] ⊕ Cgn = xn [τ].
(63)
i=0
Ê jn −1
i=0
zan +i [υ] ⊕ cn−1 [υ] ⊕ Cgn = xn [υ].
(64)
Here Cgn is deterministic for a particular codeword position.
The linear combination in Eq. (60) becomes Eq. (65) for
1 ≤ n ≤ N, which is shown at the top of the page. The
deterministic part Cgn ⊕ Cgn can contribute to non-linearity
when Cgn ⊕ Cgn , Cgn . Therefore, when we let Cgn = 0, our
codes are linear.
Pr[Rn = ri |Cn = ψi ].
Pr[Yn < U(Xn )|Cn = ψi ]
Õ
Pr[Yn < U(Xn )|Rn = ri, Cn = ψi ] · Pr[Rn = ri ]
=
r
=
i
Õ
xi
Pr[Yn < U(Xn )|Xn = xi = ri ⊕ ψi ] · Pr[Rn = ri ]
1 Õ
Pr(Yn < U(Xn )|Xn = xi ).
pM x
=
(b)
Similarly, for a different realisation of Cn and Rn , we have
Pr[Yn < U(Xn )|Cn = ψ j ]
Õ
=
Pr[Yn < U(Xn )|Rn = r j , Cn = ψ j ] · Pr[Rn = r j ]
rj
=
Õ
xj
=
Pr[Yn < U(Xn )|Xn = x j = r j ⊕ ψ j ] · Pr[Rn = r j ]
1 Õ
Pr(Yn < U(Xn )|Xn = x j ).
pM x
Since the ranges of xi and x j are Ψ, therefore we can obtain
that
Õ
Pr(Yn < U(Xn )|Xn = xi )
Pr[Yn < U(Xn )|Cn = ψi ] = Pr[Yn < U(Xn )|Cn = ψ j ],
Õ
xj
2 ) is the noise of the AWGN channel; (b)
where Nn ∼ N(0, σch
follows Eq. (29); Cn is the n-th random variable of intended
codeword before adding the random-coset and Rn is the n-th
random variable of the random-coset.
To prove that adding the random-coset can produce the
output-symmetric effect, we must have
(67)
where U(.) outputs the maximum-likelihood decision region;
ψi, ψ j ∈ Ψ and ψi , ψ j . In other words, the decoding error
probability is the same for any transmitted codeword.
(70)
j
=
(66)
(69)
i
A PPENDIX B
P ROOF OF T HEOREM 1
Yn = Xn + Nn = Cn ⊕ Rn + Nn,
(68)
Since Rn is independent of Cn and Rn is uniformly distributed
over Ψ, we then have
xi
Consider the n-th symbol. Let Xn be the channel input
random variable. Let Yn be the n-th received signal with the
input-output relationship given by
(65)
Pr(Yn < U(Xn )|Xn = x j ).
(71)
Plugging Eq. (71) into Eq. (69) and Eq. (70), we obtain Eq.
(67).
P ROOF
OF
A PPENDIX C
P ERMUTATION -I NVARIANCE
A. Proof of Theorem 2
First, we define a probability-vector random variable X =
[Xψ0 , Xψ1 , . . . , Xψp M −1 ] and let P = X+θ where θ is a random
variable and uniformly chosen from Ψ. For the m-th random
variable in X, we denote a probability event by
Pr[Xψm ∈ ε].
(72)
Then for the i-th random variable in P, we have
Pr[Pψi ∈ ε] = Pr[Xψm ∈ ε] · Pr[ψm ⊕ θ = ψi ],
because θ is independent of X.
(73)
14
Similarly, for the j-th random variable in P, where ψ j , ψi ,
we can obtain that:
Pr[Pψ j ∈ ε] = Pr[Xψm ∈ ε] · Pr[ψm ⊕ θ = ψ j ],
(74)
We know θ is a random variable and uniformly chosen from
Ψ. Thus we have:
Pr[ψm ⊕ θ = ψi ] = Pr[ψm ⊕ θ = ψ j ] =
1
.
pM
(75)
Therefore, the distribution of any two random variables in
P is the same. If we let ψ j = ψi ⊕ χ for any fixed χ ∈ Ψ, we
obtain that:
⊕χ
Pr[Pψi ∈ ε] = Pr[Pψ j ∈ ε] = Pr[Pψi ⊕ χ ∈ ε] = Pr[Pψi ∈ ε].
(76)
It can be seen that every random variable in P is identically
distributed. Therefore, we can conclude that P is identically
distributed with P ⊕ χ so P is permutation-invariant.
B. Proof of Lemma 1
For the m-th LLR random variable in W, we denote a
probability event by
Pr[Wψm ∈ δ],
(77)
where
a random event. From (52), we know that Wψm =
P δ is
ψ0
log Pψ , thus we can obtain that
m
Pr[Wψm ∈ δ]
"
#
Pψ0
= Pr log
∈δ
Pψm
= Pr[Pψ0 ∈ e δ Pψm ]
∫
∫
fPψ0 , Pψm (pψ0 , pψm )dpψ0 dpψm ,
=
Pψm
(78)
e δ Pψm
where fPψ0 , Pψm (pψ0 , pψm ) denotes the joint pdf of Pψ0 and
Pψm .
Similarly, for the n-th LLR random variable in W where
n , m, we have
∫
∫
fPψ0 , Pψn (pψ0 , pψn )dpψ0 dpψn .
Pr[Wψn ∈ δ] =
Pψn
e δ Pψn
(79)
We know Pψm and Pψn have the same distribution as
because P is permutation-invariant. Thus, the joint distribution
of Pψ0 and Pψm is the same as that of Pψ0 and Pψn . As a result,
we can obtain that:
Pr[Wψn ∈ δ] = Pr[Wψm ∈ δ].
(80)
This indicates that Wψn and Wψm have the same distribution
for any n , m. Therefore, W is permutation-invariant.
R EFERENCES
[1] M. Qiu, L. Yang, Y. Xie, and J. Yuan, “On the design of multidimensional irregular repeat-accumulate lattice codes,” in Proc. IEEE
ISIT, Jun. 2017, pp. 2598 – 2602.
[2] C. E. Shannon, “A mathematical theory of communication,” Bell System
Technical Journal, vol. 27, no. 3, pp. 379–423, Mar. 1948.
[3] U. Erez and R. Zamir, “Achieving 12 log(1+SNR) on the AWGN channel
with lattice encoding and decoding,” IEEE Trans. Inf. Theory, vol. 50,
no. 10, pp. 2293–2314, Oct. 2004.
[4] L. Natarajan, Y. Hong, and E. Viterbo, “Lattice index coding,” IEEE
Trans. Inf. Theory, vol. 61, no. 12, pp. 6505–6525, Dec. 2015.
[5] Y. C. Huang, N. E. Tunali, and K. R. Narayanan, “A compute-andforward scheme for Gaussian bi-directional relaying with inter-symbol
interference,” IEEE Trans. Commun., vol. 61, no. 3, pp. 1011–1019,
Mar. 2013.
[6] D. Fang, Y.-C. Huang, Z. Ding, G. Geraci, S. L. Shieh, and H. Claussen,
“Lattice partition multiple access: A new method of downlink nonorthogonal multiuser transmissions,” in Proc. IEEE Globecom, Dec.
2016, pp. 1–6.
[7] A. Hindy and A. Nosratinia, “Lattice coding and decoding for multipleantenna ergodic fading channels,” IEEE Trans. Commun., vol. 65, no. 5,
pp. 1873–1885, May 2017.
[8] N. Sommer, M. Feder, and O. Shalvi, “Low-density lattice codes,” IEEE
Trans. Inf. Theory, vol. 54, no. 4, pp. 1561–1585, Apr. 2008.
[9] O. Shalvi, N. Sommer, and M. Feder, “Signal codes: Convolutional
lattice codes,” IEEE Trans. Inf. Theory, vol. 57, no. 8, pp. 5203–5226,
Aug. 2011.
[10] M. R. Sadeghi, A. H. Banihashemi, and D. Panario, “Low-density paritycheck lattices: Construction and decoding analysis,” IEEE Trans. Inf.
Theory, vol. 52, no. 10, pp. 4481–4495, Oct. 2006.
[11] N. di Pietro, J. J. Boutros, G. Zémor, and L. Brunel, “Integer low-density
lattices based on Construction A,” in Inf. Theory Workshop, Sep. 2012,
pp. 422–426.
[12] N. E. Tunali, Y. C. Huang, J. J. Boutros, and K. R. Narayanan, “Lattices
over Eisenstein integers for compute-and-forward,” IEEE Trans. Inf.
Theory, vol. 61, no. 10, pp. 5306–5321, Oct. 2015.
[13] Y. Yan, C. Ling, and X. Wu, “Polar lattices: Where Arikan meets
Forney,” in Proc. IEEE ISIT, Jul. 2013, pp. 1292–1296.
[14] J. H. Conway and N. J. A. Sloane, Sphere Packings, Lattices, and
Groups. Springer Verlag, 1999.
[15] G. Poltyrev, “On coding without restrictions for the AWGN channel,”
IEEE Trans. Inf. Theory, vol. 40, no. 2, pp. 409–417, Mar. 1994.
[16] K. W. Shum and Q. T. Sun, “Lattice network codes over optimal lattices
in low dimensions,” in Proc. 7th Int. Workshop on Signal Design and
its App. in Comm., Sep. 2015, pp. 113–117.
[17] W. Kositwattanarerk, S. S. Ong, and F. Oggier, “Construction A of
lattices over number fields and block fading (wiretap) coding,” IEEE
Trans. Inf. Theory, vol. 61, no. 5, pp. 2273–2282, May 2015.
[18] F. Oggier and J. C. Belfiore, “Enabling multiplication in lattice codes
via Construction A,” in Inf. Theory Workshop, Sep. 2013, pp. 1–5.
[19] W. Kositwattanarerk, S. S. Ong, and F. Oggier, “Wiretap encoding of
lattices from number fields using codes over Fp,” in Proc. IEEE ISIT,
Jul. 2013, pp. 2612–2616.
[20] H. Khodaiemehr, M. Sadeghi, and D. Panario, “Construction of fulldiversity LDPC lattices for block-fading channels,” arXiv: 1612.04039
[cs.IT], Jun. 2016.
[21] M. Qiu, L. Yang, and J. Yuan, “Irregular repeat-accumulate lattice
network codes for two-way relay channels,” in Proc. IEEE Globecom,
Dec. 2016, pp. 1–6.
[22] A. Bennatan and D. Burshtein, “Design and analysis of nonbinary LDPC
codes for arbitrary discrete-memoryless channels,” IEEE Trans. Inf.
Theory, vol. 52, no. 2, pp. 549–583, Feb. 2006.
[23] N. di Pietro, G. Zémor, and J. J. Boutros, “LDA lattices without dithering
achieve capacity on the gaussian channel,” arXiv: 1603.02863 [cs.IT],
Apr. 2016.
[24] N. di Pietro and J. J. Boutros, “Leech constellations of Construction-A
lattices,” IEEE Trans. Commun., vol. PP, no. 99, pp. 1–1, Aug. 2017.
[25] J. J. Boutros, N. di Pietro, and N. Basha, “Generalized low-density
(GLD) lattices,” in Inf. Theory Workshop, Nov. 2014, pp. 15–19.
[26] N. Sommer, M. Feder, and O. Shalvi, “Low-density lattice codes,” IEEE
Trans. Inf. Theory, vol. 54, no. 4, pp. 1561–1585, Apr. 2008.
[27] H. Khodaiemehr, M. R. Sadeghi, and A. Sakzad, “Practical encoder and
decoder for power constrained QC LDPC-lattice codes,” IEEE Trans.
Commun., vol. 65, no. 2, pp. 486–500, Feb. 2017.
[28] R. Zamir, Lattice Coding for Signals and Networks.
Cambridge
University Press, 2015.
15
[29] H. Jin, A. Khandekar, and R. McEliece, “Irregular repeat-accumulate
codes,” in Int. Symp. Turbo Codes and Related Topics, Sep. 2000, pp.
1–8.
[30] M. C. Chiu, “Bandwidth-efficient modulation codes based on nonbinary
irregular repeat-accumulate codes,” IEEE Trans. Inf. Theory, vol. 56,
no. 1, pp. 152–167, Jan. 2010.
[31] L. Yang, T. Yang, J. Yuan, and J. An, “Irregular repeat-accumulate coded
physical-layer network coding design for two-way relay channels,” in
Proc. IEEE ICCC, Oct. 2014, pp. 91–96.
[32] J. Yuan, L. Yang, T. Yang, and J. An, “Design of non-binary irregular
repeat-accumulate codes for reliable physical-layer network coding,” in
Proc. IEEE ICT, Apr. 2015, pp. 265–271.
[33] L. Yang, T. Yang, J. Yuan, and J. An, “Achieving the near-capacity of
two-way relay channels with modulation-coded physical-layer network
coding,” IEEE Trans. Wireless Commun., vol. 14, no. 9, pp. 5225–5239,
Sep. 2015.
[34] J. H. Conway and D. Smith, On Quaternions and Octonions: Their
Geometry, Arithmetic, and Symmetry. Natick, MA, USA: A K Peters,
2003.
[35] J. Conway and N. Sloane, “Fast quantizing and decoding and algorithms
for lattice quantizers and codes,” IEEE Trans. Inf. Theory, vol. 28, no. 2,
pp. 227–232, Mar. 1982.
[36] Y. C. Huang, “Lattice index codes from algebraic number fields,” IEEE
Trans. Inf. Theory, vol. 63, no. 4, pp. 2098–2112, Apr. 2017.
[37] S. J. Johnson and S. R. Weller, “Constructions for irregular repeataccumulate codes,” in Proc. IEEE ISIT, Sep. 2005, pp. 2157–8095.
[38] T. J. Richardson and R. L. Urbanke, “The capacity of low-density paritycheck codes under message-passing decoding,” IEEE Trans. Inf. Theory,
vol. 47, no. 2, pp. 599–618, Feb. 2001.
[39] D. E. Dudgeon and R. M. Mersereau, Multidimensional Digital Signal
Processing. Englewood Cliffs, NJ: Prentice-Hall, 1984.
[40] V. S. Ganepola, R. A. Carrasco, I. J. Wassell, and S. L. Goff, “Performance study of non-binary LDPC codes over GF(q),” in CSNDSP, Jul.
2008, pp. 585–589.
[41] G. Li, I. J. Fair, and W. A. Krzymieri, “Analysis of nonbinary ldpc
codes using Gaussian approximation,” in Proc. IEEE ISIT, Jun. 2003,
pp. 234–234.
[42] T. A. Severini, Elements of Distribution Theory. Cambridge University
Press, 2005.
[43] S. ten Brink and G. Kramer, “Design of repeat-accumulate codes for
iterative detection and decoding,” IEEE Trans. Signal Process., vol. 51,
no. 11, pp. 2764–2772, Nov. 2003.
| 7cs.IT
|
Fast Meta-Learning for Adaptive Hierarchical Classifier
Design
Gerrit J.J. van den Burg1
Alfred O. Hero2
1
2
Econometric Institute, Erasmus University Rotterdam, The Netherlands
Electrical Engineering and Computer Science, University of Michigan, USA
arXiv:1711.03512v1 [cs.LG] 9 Nov 2017
November 10, 2017
Abstract
We propose a new splitting criterion for a meta-learning approach to multiclass
classifier design that adaptively merges the classes into a tree-structured hierarchy of increasingly difficult binary classification problems. The classification tree is
constructed from empirical estimates of the Henze-Penrose bounds on the pairwise
Bayes misclassification rates that rank the binary subproblems in terms of difficulty of classification. The proposed empirical estimates of the Bayes error rate are
computed from the minimal spanning tree (MST) of the samples from each pair of
classes. Moreover, a meta-learning technique is presented for quantifying the onevs-rest Bayes error rate for each individual class from a single MST on the entire
dataset. Extensive simulations on benchmark datasets show that the proposed hierarchical method can often be learned much faster than competing methods, while
achieving competitive accuracy.
Keywords: classifier performance, meta-learning, multiclass classification, hierarchical classifier, classifier comparison
1
Introduction
The Bayes error rate (BER) is a central concept in the statistical theory of classification.
It represents the error rate of the Bayes classifier, which assigns a label to an object
corresponding to the class with the highest posterior probability. By definition, the
Bayes error represents the smallest possible average error rate that can be achieved by
any decision rule (Wald, 1947). Because of these properties, the BER is of great interest
both for benchmarking classification algorithms as well as for the practical design of
classification algorithms. For example, an accurate approximation of the BER can be
used for classifier parameter selection, data dimensionality reduction, or variable selection. However, accurate BER approximation is difficult, especially in high dimension,
and thus much attention has focused on tight and tractable BER bounds. This paper
proposes a model-free approach to designing multiclass classifiers using a bias-corrected
BER bound estimated directly from the multiclass data.
There exists several useful bounds on the BER that are functions of the classdependent feature distributions. These include information theoretic divergence measures such as the Chernoff α-divergence (Chernoff, 1952), the Bhattacharyya divergence
(Kailath, 1967), or the Jensen-Shannon divergence (Lin, 1991). Alternatively, arbitrarily tight bounds on performance can be constructed using sinusoidal or hyperbolic
1
approximations (Avi-Itzhak and Diep, 1996, Hashlamoun et al., 1994). These bounds
are functions of the unknown class-dependent feature distributions.
Recently, Berisha et al. (2016) introduced a divergence measure belonging to the
family of f -divergences which tightly bounds the Bayes error rate in the binary classification problem. The bounds on the BER obtained with this measure are tighter than
bounds derived from the Bhattacharyya or Chernoff bounds. Moreover, this divergence
measure can be estimated nonparametrically from the data without resorting to density
estimates of the distribution functions. Inspired by the Friedman-Rafsky multivariate
runs test (Friedman and Rafsky, 1979), estimation is based on computing the Euclidean
minimal spanning tree (MST) of the data, which can be done in approximately O(n log n)
time. In this paper we propose improvements to this estimator for problems when there
are unequal class priors and apply the improved estimator to the adaptive design of a
hierarchical multiclass classifier. Furthermore, a fast method is proposed for bounding
the Bayes error rate of individual classes which only requires computing a single minimal spanning tree over the entire set of samples. Thus our proposed method is faster
than competing methods that use density plug-in estimation of divergence or observed
misclassification rates of algorithms, such as SVM or logistic regression, which involve
expensive parameter tuning.
Quantifying the complexity of a classification problem has been of significant interest (Ho and Basu, 2002) and it is clear that a fast and accurate estimate of this
complexity has many practical applications. For instance, an accurate complexity estimator allows the researcher to assess a priori whether a given classification problem is
difficult to classify or not. In a multiclass problem, a pair of classes which are difficult
to disambiguate could potentially be merged or could be designated for additional data
collection. Moreover, an accurate estimate of the BER could be used for variable selection, an application that was explored previously in Berisha et al. (2016). In Section 3
further applications of the BER estimates to multiclass classification are presented and
evaluated.
There are many methods available for the design of multiclass classification algorithms, including: logistic regression (Cox, 1958); support vector machines (Cortes and
Vapnik, 1995); and neural networks (McCulloch and Pitts, 1943). It is often the case
that classifier performance will be better for some classes than for others, for instance
due to sample imbalance in the training set. Often classifier designs apply weights to
the different classes in order to reduce the effect of such imbalances on average classifier
accuracy (Lu et al., 1998, Qiao and Liu, 2009). We take a different and more general approach that incorporates an empirical determination of the relative difficulties of
classifying between different classes. Accurate empirical estimates of the BER are used
for this purpose. A multiclass classifier is presented in Section 4 that uses MST-based
BER estimates to create a hierarchy of binary subproblems that increase in difficulty
as the algorithm progresses. This way, the classifier initially works on easily decidable
subproblems before moving on to more difficult multiclass classification problems.
The paper is organized as follows. The theory of the nonparametric Bayes error
estimator of Berisha et al. (2016) will be reviewed in Section 2. We will introduce a bias
correction for this estimator, motivate the use of the estimator for multiclass classification, and discuss computational complexity. Section 3 will introduce applications of the
estimator to meta-learning in multiclass classification. A novel hierarchical classification
method will be introduced and evaluated in Section 4. Section 5 provides concluding
remarks.
2
2
An Improved BER Estimator
Here the motivation and theory for the estimator of the Bayes error rate is reviewed,
as introduced by Berisha et al. (2016). An improvement on this estimator is proposed
for the case where class prior probabilities are unequal. Next, the application of the
estimator in multiclass classification problems is considered. Finally, computational
considerations and robustness analyses are presented.
2.1
Estimating the Bayes Error Rate
Consider the binary classification problem on a dataset D = {(xi , yi )}i=1,...,n , where
xi ∈ X ⊆ Rd and yi ∈ {1, 2}. Denote the multivariate density functions for the two
classes by f1 (x) and f2 (x) and the prior probabilities by p1 and p2 = 1−p1 , respectively.
The Bayes error rate of this binary classification problem can be expressed as (Fukunaga,
1990)
Z
Pe (f1 , f2 ) =
min {p1 f1 (x), p2 f2 (x)} dx.
(1)
Recently, Berisha et al. (2016) derived a tight bound on the BER which can be
estimated directly from the data without a parametric model for the density or density
estimation. This bound is based on a divergence measure introduced by Berisha and
Hero (2015), defined as
"Z
#
1
(p1 f1 (x) − p2 f2 (x))2
DHP (f1 , f2 ) =
dx − (p1 − p2 )2 ,
(2)
4p1 p2
p1 f1 (x) + p2 f2 (x)
and called the Henze-Penrose divergence, as it is motivated by an affinity measure defined
by Henze and Penrose (1999). In Berisha and Hero (2015) it was shown that (2) is a
proper f -divergence as defined by Csiszár (1975).
Estimation of the Henze-Penrose (HP) divergence is based on the multivariate runs
test proposed by Friedman and Rafsky (1979) and convergence of this test was studied
by Henze and Penrose (1999). Let Xk = {xi ∈ X ⊆ Rd : yi = k} with k ∈ {1, 2} denote
multidimensional features from two classes. Define the class sample sizes nk = |Xk |,
k ∈ {1, 2}. Let the combined sample be denoted by X = X1 ∪X2 where n = |X| = n1 +n2
is the total number of samples from both classes. Define the complete graph G over X
as the graph connecting all n nodes {xi }ni=1 with edge weights |ekj | = kxk − xj k equal
to Euclidean distances. The Euclidean minimal spanning tree that spans X, denoted
by T , is defined as the subgraph of G that is both connected and whose sum of edge
weights is the smallest possible. The Friedman-Rafsky test statistic equals the number
of edges in T that connect an observation from class 1 to an observation from class 2
and is denoted by R1,2 .
Building on the work by Henze and Penrose (1999), Berisha et al. (2016) show that
if n1 → ∞ and n2 → ∞ in a linked manner such that n1 /(n1 + n2 ) → δ ∈ (0, 1) then,
1−
nR1,2
→ DHP (f1 , f2 )
2n1 n2
a.s.
(3)
Thus, the number of cross connections between the classes in the Euclidean MST is
inversely proportional to the divergence between the respective probability density functions of these classes.
Finally, the HP-divergence can be used to bound the Bayes error rate, Pe (f1 , f2 ),
following Theorem 2 of Berisha et al. (2016)
p
1
1
uHP (f1 , f2 ) ≤ Pe (f1 , f2 ) ≤ 21 − 12 uHP (f1 , f2 ),
(4)
2 − 2
3
0.6
0.6
BER, p = 0.15
P̂e , p = 0.33
PBC , p = 0.33
0.4
0.3
0.2
0.1
0
P̂e , p = 0.15
PBC , p = 0.15
BER, p = 0.33
0.5
Error Probability
0.5
Error Probability
BER, p = 0.15
P̂e , p = 0.15
PBC , p = 0.15
BER, p = 0.33
P̂e , p = 0.33
PBC , p = 0.33
0.4
0.3
0.2
0.1
0
1
2
3
Distance
4
0
5
(a) without bias correction
0
1
2
3
Distance
4
5
(b) with bias correction
Figure 1: Illustration of the estimates of the Bayes error rate for two different values
of the prior probability p and for spherical bivariate Gaussian distributions with increasing separation between the classes as measured by Euclidean distance between the
class means, kµ1 − µ2 k. We compare the true BER, the Bhattacharyya bound (PBC )
requiring the true class distributions, and the HP-estimator (P̂e ) based on an empirical
sample. Figures (a) and (b) show the estimates for p = 0.15 and p = 1/3, respectively
with and without the proposed bias correction of the HP-estimator. The HP-estimator
was implemented using n = 1000 samples (n1 = pn and n2 = n − n1 ) and averaged over
200 trials.
where
uHP (f1 , f2 ) = 4p1 p2 DHP (f1 , f2 ) + (p1 − p2 )2 .
Averaging these bounds yields an estimate of the BER given by
p
P̂e (f1 , f2 ) = 21 − 41 uHP (f1 , f2 ) − 14 uHP (f1 , f2 ).
(5)
(6)
In the following, this estimator will be referred to as the HP-estimator of the BER.
2.2
A modified HP-estimator for unequal class priors
To illustrate the performance of the HP-estimator, and to motivate the proposed modification, consider a binary classification problem where the samples are drawn from two
independent bivariate Gaussian distributions with equal covariance matrices. For this
example the BER and associated bounds can be computed exactly (Fukunaga, 1990).
In Figure 1a we compare the BER, the HP-estimator of the BER (6) and the popular
Bhattacharyya bound on the BER (Bhattacharyya, 1946, Kailath, 1967). Figure 1a
shows that the HP-estimator is closer to the true BER than the Bhattacharyya bound.
This result was illustrated by Berisha et al. (2016) for the case where p1 = p2 and is
confirmed here for the case where p1 6= p2 .
However, Figure 1a also shows that a significant bias occurs in the HP-estimate of
the BER when the distance between classes is small. Considering that Pe (f1 , f2 ) =
4
min{p1 , p2 } if f1 → f2 , the solution of the equation P̂e (f1 , f2 ) = min{p̂1 , p̂2 } for R1,2
suggests a bias corrected version of the HP-estimate:1
0
R1,2
= min{γ, R1,2 },
(7)
with
γ = 2n min{p̂1 , p̂2 } − 43 n + 14 n
p
9 − 16 min{p̂1 , p̂2 },
(8)
where p̂1 = n1 /n and p̂2 = n2 /n are estimates of the true prior probabilities. Figure 1b
shows the effect of this bias correction on the accuracy of the HP-estimator. As can
be seen, the bias correction significantly improves the accuracy of the HP-estimator for
when the class distributions are not well separated.
2.3
Multiclass classification
Here we apply the HP-estimate to multiclass classification problems by extending the
bias corrected HP-estimator to a multiclass Bayes error rate. The original multiclass
HP-estimator has been defined by Wisler et al. (2016) and we show how the framework
can be applied to hierarchical multiclassifier design.
Consider a multiclass problem with K classes with xi ∈ X ⊆ Rd and yi ∈ {1, . . . , K},
with
P prior probabilities pk and density functions fk (x) for k = 1, . . . , K such that
k pk = 1. Then, the BER can be estimated for each pair of classes using the biascorrected HP-estimator P̂e (fk , fj ) using (7). The binary classification problem with the
largest BER estimate is defined as most difficult.
Recall that the largest BER that can be achieved in a binary classification problem
with unequal class priors is equal to the value of the smallest prior probability. This
makes it difficult to compare empirical estimates of the BER when class sizes are imbalanced. To correct for this, the HP-estimator for pairwise classification BERs can be
normalized for class sizes using
P̂e0 (fk , fl ) =
P̂e (fk , fl )
.
min{p̂k , p̂l }
(9)
This normalization places the HP-estimate in the interval [0, 1] and makes it possible to
more accurately compare the BER estimates of different binary problems.
In practice it can also be of interest to understand how difficult it is to discriminate
each individual class. By reducing the multiclass problem to a One-vs-Rest classification
problem, it is straightforward to define a confusion rate for a given class k. This represents the fraction of instances that are erroneously assigned to class k and the fraction
of instances which are truly from class k that are assigned to a different class. Formally,
define the confusion rate for class k as
|{i : ŷi = k, yi 6= k}| + |{i : ŷi 6= k, yi = k}|
Ck (y, ŷ) =
,
(10)
n
with ŷi the predicted class for instance i. Recall that the Bayes error rate is the error
rate of the Bayes classifier, which assigns an instance x to class k = arg maxl pl fl (x).
Hence, the BER for a single class k equals the error of assigning to a class l 6= k when
the true class is k and the total error of assigning to class k when the true class is c 6= k,
thus
Z
Z
X
Pe,k =
pk fk (x) dx +
pc fc (x) dx
(11)
c6=k
max{pl fl (x)}≥pk fk (x)
l6=k
max{pl fl (x)}<pk fk (x)
l6=k
1
This correction can readily be derived by using the fact that 1 −
under the same conditions as for (3).
5
2R1,2
n
→ uHP (f1 , f2 ), which holds
·10−4
1
2
3
5
7
11
2.5
Variance
2
1.5
1
0.5
0
0
1
2
3
4
5
Distance
Figure 2: Variance of the HP-estimator of the BER for varying number of orthogonal
MSTs as a function of the separation between the classes. Results are based on 500
repetitions of a binary classification problem with n = 1000, d = 2, p1 = 0.33, where
class distributions are bivariate spherical Gaussians with different means.
We make two observations about this One-vs-Rest Bayes error rate (OvR-BER).
First, the OvR-BER for class k is smaller than the sum of the binary BERs for the
problems involving class k (see Appendix A). Second,
the OvR-BER can be estimated
S
using the Henze-Penrose divergence with R(Xk , l6=k Xl ), which yields the estimate P̂e,k .
A computational advantage of using the OvR-BER in multiclass problems is that the
MST only has to be computed only
S once on the set X, since the union of Xk and ∪l6=k Xl
is equal to X. Therefore, R(Xk , l6=k Xl ) can be computed for all k from the single MST
on X by keeping track of the labels of each instance.
2.4
Computational Considerations
The construction of the minimal spanning tree lies at the heart of the HP-estimator of
the BER, so it is important to use a fast algorithm for the MST construction. Since
the HP-estimator is based on the Euclidean MST the dual-tree algorithm by March
et al. (2010) can be applied. This algorithm is based on the construction of Borůvka
(1926) and implements the Euclidean MST in approximately O(n log n) time. For larger
datasets it can be beneficial to partition the space into hypercubes and construct the
MST in each partition.
A simple way to improve the robustness of the HP-estimator is to use multiple
orthogonal MSTs and average the number of cross-connections (Friedman and Rafsky,
1979). Computing orthogonal MSTs is not straightforward in the dual-tree algorithm
of March et al. (2010), but is easy to implement in MST algorithms that use a pairwise
distance matrix such as that of Whitney (1972). Figure 2 shows the empirical variance
of the HP-estimator for different numbers of orthogonal MSTs as a function of the
separation between the classes. As expected, the variance decreases as the number of
orthogonal MSTs increases, although the benefit of including more orthogonal MSTs
also decreases when adding more MSTs. Therefore, 3 orthogonal MSTs are typically
used in practice.
6
0
1
2
3
4
5
6
7
8
9
0
0
0
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
9
9
(a)
1
2
3
4
5
6
7
8
9
(b)
Figure 3: Heat maps illustrating the difficulty of distinguishing between different handwritten digits in the MNIST dataset of LeCun et al. (1998). Brighter squares correspond
to higher values. In (a) the heat map of the BER estimates on the training data is shown.
In (b) the heat map is shown of the 82 misclassifications made by LeNet-5 on the test
data (LeCun et al., 1998). From (a) it can be readily identified that the numbers 3 and
5 are difficult to distinguish, as well as 4 and 9. Easy to distinguish number pairs are
for instance 6 and 7, and 0 and 1. This pattern is reflected in the results on the test
data shown in (b).
3
Meta-Learning of optimal classifier accuracy
Applying the HP-estimator to meta-learning problems creates a number of opportunities
to assess the difficulty of a classification problem before training a classifier. For example, given a multiclass classification problem it may be useful to know which classes
are difficult to distinguish from each other and which classes are easy to distinguish.
Figure 3a shows an illustration of this for handwritten digits in the well-known MNIST
dataset (LeCun et al., 1998). This figure shows a heat map where each square corresponds to an estimate of the BER for a binary problem in the training set. From this
figure it can be seen that the digits 4 and 9 are difficult to distinguish, as well as the
digits 3 and 5. This information can be useful for the design of a classifier, to ensure for
instance that higher weights are placed on misclassifications of the more difficult number
pairs if correct classification of these pairs is of importance to the end-task. In Figure 3b
a similar heat map is shown based on misclassified instances of LeNet-5 (LeCun et al.,
1998) on the test set. This figure shows the symmetric confusion matrix based on the
82 misclassified instances. As can be seen, this figure closely corresponds to the heat
map on the training data, which confirms the predictive accuracy of the HP-estimator
for real data.
Another example of the accuracy of the BER estimates for multiclass classification
problems is given in Figure 4. In this figure, OvR-BER estimates P̂e,k and class accuracy scores Ck (y, ŷ) are shown for the Chess dataset (n = 28056, d = 34, K = 18)
obtained from the UCI repository (Bache and Lichman, 2013). This dataset was split
into a training dataset (70%) and a test dataset (30%) and the OvR-BER estimates
were computed on the training dataset. These estimates are compared with the class
error rates obtained from out-of-sample predictions of the test dataset using GenSVM
(Van den Burg and Groenen, 2016). This figure shows that the OvR-BER estimates are
accurate predictors of classification performance. The classes that are relatively difficult
7
class error rate
0.2
0.15
0.1
0.05
0
1
2
3
4
5
6
7
8 9 10 11 12 13 14 15 16 17 18
class index
Ck (y, ŷ)
P̂e,k
Figure 4: OvR-BER estimates and test set error rates for each class in the Chess
dataset. It can be seen that for most classes the OvR-BER estimate is an accurate
predictor for test set performance. For classes where there is a difference between the
BER and the test set error the BER generally gives a lower bound on the test set
performance. Test set results were obtained with the GenSVM classifier (Van den Burg
and Groenen, 2016).
to classify may benefit from increasing misclassification weights.
The BER estimates can also be applied to feature selection and, in particular, to
the identification of useful feature transformations of the data. A feature selection
strategy based on forward selection was outlined in Berisha et al. (2016). At each
feature selection stage, this algorithm adds the feature which gives the smallest increase
in the BER estimate. Berisha et al. (2016) show that this feature selection strategy
quickly yields a subset of useful features for the classification problem.
Because the BER estimate is a fast and asymptotically consistent estimate of a
bound on classification performance, it is easy to try a number of potential feature
transformations and use the one with the smallest BER estimate in the classifier. This
can be useful both for traditional feature transformations such as PCA (Pearson, 1901)
and Laplacian Eigenmaps (Belkin and Niyogi, 2003), but also for commonly used kernel
transformations in SVMs. For a researcher this can significantly reduce the time needed
to train a classifier on different transformations of the data. In a multiclass setting where
the One-vs-One strategy is used, one can even consider a different feature transformation
for each binary subproblem. When using a unified classification method one can consider
feature transformations which reduce the average BER estimate or the worst-case BER
estimate.
Note that a feature transformation which reduces the dimensionality of the dataset
without increasing the BER estimate can be considered beneficial, as many classification
methods are faster for low-dimensional datasets. For instance, applying PCA with 2
components on the Chess dataset only slightly increases the BER estimates for two
classes, while remaining the same for the other classes. Thus, a classifier will likely
achieve comparable accuracy with this transformed dataset, but will be much faster to
train since the dimensionality can be reduced from d = 34 to d = 2.
4
Hierarchical Multiclass Classification
In this section a novel hierarchical multiclass SVM classifier is introduced which is based
on uncertainty clustering. The BER estimate can be considered a measure of the irre-
8
1
1
cut
7
2
7
2
cut
6
3
5
4
6
V
V0
(a)
3
cut
5
4
V1
(b)
Figure 5: Illustration of the min-cut procedure for a multiclass classification problem
with K = 7 classes. The first cut in (a) splits the vertices in the set V into two sets, V0
and V1 , which are used in a binary classifier. Figure (b) illustrates the next step where
both V0 and V1 are split further by cuts in the complete graph for each subset.
ducible uncertainty of a classification problem, as a high BER indicates an intrinsically
difficult problem. This can be used to construct a tree of binary classification problems
that increase in difficulty along the depth of the tree. By fitting a binary classifier (such
as an SVM) at each internal node of the tree, a classification method is obtained which
proceeds from the easier binary subproblems to the more difficult binary problems.
Similar divide-and-conquer algorithms have been proposed (Frank and Kramer, 2004,
Schwenker and Palm, 2001, Takahashi and Abe, 2002, Tibshirani and Hastie, 2007, Vural
and Dy, 2004, among others). See Lorena et al. (2008) for a review. These approaches
often apply a clustering method to create a grouping of the dataset into two clusters,
repeating this process recursively to form a binary tree of classification problems. In
Lorena and De Carvalho (2010) several empirical distance measures are used as indicators of separation difficulty between classes, which are applied in a bottom-up procedure
to construct a classification tree. Finally, in El-Yaniv and Etzion-Rosenberg (2010) the
Jensen-Shannon divergence is used to bound the BER with inequalities from Lin (1991)
and a classification tree is constructed using a randomized heuristic procedure. Unfortunately, the Jensen-Shannon divergence implementation requires parametric estimation
of distribution functions. Moreover, for the equiprobable case the upper bound on the
BER obtained with the Jensen-Shannon divergence can be shown to be less tight than
that obtained with the HP-divergence (see Appendix B). Because of this, these estimates
of the BER may be less accurate than those obtained with the proposed HP-estimator.
To construct the hierarchical classification tree a complete weighted graph G =
(V, E, w) is created where the vertices correspond to the classes and the weight of the
edges equals the HP-estimate for that binary problem. Formally, let V = {1, . . . , K},
E = {(i, j) : i, j ∈ V, i 6= j} and define the edge weight w(e) for e ∈ E as w(e) =
w(i, j) = P̂e0 (fi , fj ). In the HP-estimator P̂e0 (fi , fj ) the bias correction (7) and the
normalization (9) are used. By recursively applying min-cuts to this graph a tree of
binary classification problems is obtained which increase in difficulty along the depth
of the tree. Min-cuts on this weighted graph can be computed using for instance the
method of Stoer and Wagner (1997). Figure 5 illustrates this process for a multiclass
classification problem with K = 7.
9
V
{1, 2, 3, 4, 5, 6, 7}
V0
{2, 3, 4}
V00
{2}
V1
{1, 5, 6, 7}
V01
{3, 4}
V010
{3}
V10
{1, 7}
V011
{4}
V100
{1}
V11
{5, 6}
V101
{7}
V110
{5}
V111
{6}
Figure 6: Illustration of the tree induced by the graph cutting procedure illustrated in
Figure 5, for a hypothetical problem with K = 7 classes. Each dashed line in the tree
indicates a point where a binary classifier will be trained.
The tree construction can be described formally as follows. Starting with the complete weighted graph with vertices V , apply a min-cut algorithm to obtain the disjoint vertex sets V0 and V1 such that V0 ∪ V1 = V . This pair of vertex sets then
forms a binary classification problem with datasets X0 = {xi ∈ X : yi ∈ V0 } and
X1 = {xi ∈ X : yi ∈ V1 }. Recursively applying this procedure to the sets V0 and
V1 until no further splits are possible yields a tree of binary classification problems, as
illustrated in Figure 6.
In the remainder of this section the results of an extensive simulation study are presented, which aims to evaluate the performance of this hierarchical classifier on multiclass
classification problems. The classifier that will be used in each binary problem in the
tree will be a linear support vector machine, but in practice any binary classifier could
be used in the algorithm. The implementation of the hierarchical classifier based on the
linear binary SVM will be called SmartSVM.2
The experimental setup is comparable to that used in Van den Burg and Groenen
(2016), where a nested cross-validation (CV) approach is used to reduce bias in classifier
performance (Stone, 1974). From each original dataset 5 independent training and test
datasets are generated. Subsequently, each classification method is trained using 10
fold CV on each of the training datasets. Finally, the model is retrained on the entire
training dataset using the optimal hyperparameters and this model is used to predict
the test set. In the experiments 16 datasets are used of varying dimensions from which
80 independent test sets are constructed. The train and test datasets were generated
using a stratified split, such that the proportions of the classes correspond to those in
the full dataset. Table 1 shows the descriptive statistics of each of the datasets used.
Datasets are collected from the UCI repository (Bache and Lichman, 2013) and the
KEEL repository (Alcalá et al., 2010).
SmartSVM will be compared to five other linear multiclass SVMs in these experiments. Three of these alternatives are heuristic methods which use the binary SVM
as underlying classifier, while two others are single-machine multiclass SVMs. One of
the most commonly used heuristic approaches to multiclass SVMs is the One vs. One
(OvO) method (Kreßel, 1999) which solves a binary SVM for each of the K(K − 1) pairs
of classes. An alternative is the One vs. Rest (OvR) method (Vapnik, 1998) in which
2
The SmartSVM classifier and the meta-learning and BER estimation techniques presented in the
previous sections have been implemented in the smartsvm Python package, available at: https://
github.com/HeroResearchGroup/SmartSVM.
10
Table 1: Dataset summary statistics for the datasets used in the experimental study.
The final two columns denote the size of the smallest and the largest class, respectively.
Datasets marked with an asterisk are collected from the KEEL dataset repository, all
others are from the UCI repository.
Dataset
abalone
fars*
flare
kr-vs-k
letter
nursery
optdigits
pageblocks
pendigits
satimage
segment
shuttle
texture*
wine red
wine white
yeast
Instances
4177
100959
1066
28056
20000
12960
5620
5473
10992
6435
2310
58000
5500
1599
4898
1479
Features
8
338
19
34
16
19
64
10
16
36
19
9
40
12
12
8
Classes
20
7
6
18
26
4
10
5
10
6
7
6
11
5
6
9
min nk
14
299
43
27
734
330
554
28
1055
626
330
23
500
18
20
20
max nk
689
42116
331
4553
813
4320
572
4913
1144
1533
330
45586
500
681
2198
463
a binary SVM is solved for each of the K − 1 binary problems obtained by separating
a single class from the others. The directed acyclic graph (DAG) SVM was proposed
by Platt et al. (2000) as an extension of the OvO approach. It has a similar training
procedure as OvO, but uses a different prediction strategy. In the OvO method a voting
scheme is used where the class with the most votes from each binary classifier becomes
the predicted label. In contrast, the DAGSVM method uses a voting scheme where
the least likely class is voted away until only one remains. Finally, two single-machine
multiclass SVMs are also compared: the method by Crammer and Singer (2002) and
GenSVM (Van den Burg and Groenen, 2016).
All methods are implemented in either C or C++, to ensure that speed of the methods
can be accurately compared. The methods that use a binary SVM internally are implemented with LibLinear (Fan et al., 2008). LibLinear also implements a fast solver for the
method by Crammer and Singer (2002) using the algorithm proposed by Keerthi et al.
(2008). For SmartSVM the Bayes error rates and the corresponding classification tree
were calculated once for each training dataset as a preprocessing step. For most datasets
the BERs were computed based on 3 orthogonal MSTs using the algorithm of Whitney
(1972). For the two largest datasets (fars and shuttle) the BER was computed based on
a single MST using the algorithm of March et al. (2010). Computing these MSTs was
done in parallel using at most 10 cores. In the results on training time presented below
the training time of SmartSVM is augmented with the preprocessing time.
The binary SVM has a cost parameter for the regularization term, which is optimized
using cross validation. The range considered for this parameter is C ∈ {2−18 , 2−16 , . . . , 218 }.
The GenSVM method has additional hyperparameters which were varied in the same
way as in the experiments of Van den Burg and Groenen (2016). All experiments were
11
Table 2: Total training time per dataset in seconds, averaged over the 5 nested CV folds.
Minimal values per dataset are underlined. As can be seen, SmartSVM is the fastest
method on 10 out of 16 datasets.
Dataset
abalone
fars
flare
krkopt
letter
nursery
optdigits
pageblocks
pendigits
satimage
segment
shuttle
texture
winered
winewhite
yeast
C&S
13332
158242
467.5
86000
31684
2267
88.0
523.4
2499
5184
377.1
7658
588.1
706.4
2570
1209
DAG
146.5
3108
7.0
728.4
407.3
74.7
20.0
26.8
30.2
74.7
7.1
570.8
34.5
12.3
59.2
16.7
GenSVM
86787
205540
501.7
56554
44960
1363
31615
2001
6591
2563
1542
14618
14774
542.7
2570
1138
OvO
134.3
2866
6.5
680.4
381.5
68.7
18.5
25.2
28.5
70.0
7.0
561.9
32.8
11.1
55.1
15.3
OvR
214.4
5630
8.2
1388
1805
94.9
24.3
48.6
151.3
188.4
18.7
1996
69.6
18.3
70.9
24.2
SmartSVM
75.0
3158
5.0
664.0
652.1
56.8
10.9
27.6
53.7
74.5
5.4
832.0
19.3
8.8
36.7
13.2
performed on the Dutch National LISA Compute Cluster using the abed utility.3
The experiments are compared on training time and out-of-sample predictive performance. Table 2 shows the results for training time, averaged over the 5 nested cross
validation folds for each dataset. As can be seen SmartSVM is the fastest method on
10 out of 16 datasets. This can be attributed to the smaller number of binary problems
that SmartSVM needs to solve compared to OvO and the fact that the binary problems
are smaller than those solved by OvR. The OvO method is the fastest classification
method on the remaining 6 datasets. The single-machine multiclass SVMs by Crammer
and Singer (2002) and Van den Burg and Groenen (2016) both have larger computation
times than the heuristic methods. Since GenSVM has a larger number of hyperparameters, it is interesting to look at the average time per hyperparameter configuration as
well. In this case, GenSVM is on average faster than Crammer and Singer (2002) due
to the use of warm starts (see Appendix C for additional simulation results).
Classification performance of the methods is reported using the adjusted Rand index
(ARI) which corrects for chance (Hubert and Arabie, 1985). Use of this index as a classification metric has been proposed previously by Santos and Embrechts (2009). Table 3
shows the predictive performance as measured with the ARI. As can be seen, SmartSVM
obtains the maximum performance on two of the sixteen datasets. However, SmartSVM
outperforms One vs. One on 3 datasets and outperforms One vs. Rest on 10 out of 16
datasets. The OvO and OvR methods are often used as default heuristic approaches
for multiclass SVMs and are respectively the default strategies in the popular LibSVM
(Chang and Lin, 2011) and LibLinear (Fan et al., 2008) libraries. Since SmartSVM is
often faster than these methods, our results indicate a clear practical benefit to using
SmartSVM for multiclass classification.
3
See https://github.com/GjjvdBurg/abed.
12
Table 3: Out-of-sample predictive performance as measured by the adjusted Rand index.
Although SmartSVM doesn’t often achieve the maximum performance, there are several
datasets for which SmartSVM outperforms One vs. One, or differs from the maximum
performance in only the second or third decimal.
Dataset
abalone
fars
flare
krkopt
letter
nursery
optdigits
pageblocks
pendigits
satimage
segment
shuttle
texture
winered
winewhite
yeast
5
C&S
0.0788
0.8127
0.4274
0.1300
0.6173
0.8279
0.9721
0.7681
0.9068
0.7420
0.9013
0.9275
0.9936
1.0000
1.0000
0.2595
DAG
0.0898
0.8143
0.6480
0.1988
0.6991
0.8175
0.9854
0.7335
0.9566
0.7652
0.9000
0.8543
0.9950
1.0000
1.0000
0.2540
GenSVM
0.0895
0.8085
0.6687
0.1779
0.5909
0.8303
0.9732
0.6847
0.8971
0.7403
0.8812
0.8887
0.9888
0.9985
0.9998
0.2521
OvO
0.0898
0.8146
0.6496
0.2022
0.7118
0.8157
0.9850
0.7353
0.9597
0.7672
0.8982
0.8543
0.9947
1.0000
1.0000
0.2587
OvR
0.0791
0.8134
0.4084
0.1512
0.5155
0.8095
0.9686
0.6696
0.8607
0.7107
0.8354
0.6925
0.9865
0.8369
0.8131
0.2433
SmartSVM
0.0595
0.8115
0.6544
0.1585
0.4533
0.8138
0.9640
0.7214
0.8817
0.7607
0.8986
0.8038
0.8977
1.0000
1.0000
0.2088
Discussion
In this work the practical applicability of nonparametric Bayes error estimates to metalearning and hierarchical classifier design has been investigated. For the BER estimate
introduced by Berisha et al. (2016) a bias correction was derived which improves the
accuracy of the estimator for classification problems with unequal class priors. Furthermore, a normalization term was proposed which makes the BER estimates comparable
in multiclass problems. An expression of the OvR-BER was given which represents the
exact Bayes error for a single class in the multiclass problem and it was shown that this
error can be efficiently estimated using the HP-estimator as well. A robustness analysis
of the HP-estimator was performed which showed the benefit of using orthogonal MSTs
in the estimator.
There are many potential applications of the BER estimates to meta-learning problems. Above, several possibilities were explored including the prediction of which pairs of
classes are most difficult to distinguish and which individual classes will yield the highest
error rate. Preliminary experiments with feature transformations were also performed,
which showed that the BER estimates can be a useful tool in determining beneficial
transformations before a classifier is trained.
Based on the weighted graph of pairwise BER estimates, a hierarchical multiclass
classification method was proposed. The classifier uses a top-down splitting approach
to create a tree of binary classification problems which increase in difficulty along the
depth of the tree. By using a linear SVM for each classification problem, a hierarchical
multiclass SVM was obtained which was named SmartSVM. Extensive simulation studies
showed that SmartSVM is often faster than existing approaches and yields competitive
predictive performance on several datasets.
Note that the SmartSVM classifier is only one example of how the BER estimates
13
can be used to construct better classification methods. As discussed in Section 3, BER
estimates could also be used to define class weights in a multiclass classifier. Moreover,
the min-cut strategy used for SmartSVM may not be the optimal way to construct the
classification tree. Evaluating different approaches to constructing classification hierarchies and other applications of the BER estimates to multiclass classification problems
are topics for further research.
Acknowledgements
The computational experiments of this work were performed on the Dutch National LISA
Compute Cluster, and supported by the Dutch National Science Foundation (NWO).
The authors thank SURFsara (www.surfsara.nl) for the support in using the LISA
cluster. This research was partially supported by the US Army Research Office, grant
W911NF-15-1-0479 and US Dept. of Energy grant DE-NA0002534.
A
One vs. Rest Bayes Error Rate
In this section bounds for the One vs. Rest Bayes error rate will be derived, which
measures the error of the Bayes classifier in correctly identifying an individual class.
Definition 1 (OvR-BER). Let f1 , f2 , . . . , fK and p1 , p2 , . . . , pK denoteP
density functions
and prior probabilities for the classes 1 through K respectively, with c pc = 1. Then,
the Bayes error rate between a class k and the remaining classes is given by
Z
Z
X
Pe,k =
pk fk (x) dx +
pc fc (x) dx.
(12)
c6=k
max{pl fl (x)}≥pk fk (x)
l6=k
max{pl fl (x)}<pk fk (x)
l6=k
Below it will be shown that the OvR-BER can
S be bounded using the FriedmanRafsky statistic in the One-vs-Rest setting, R(Xk , l6=k Xl ). Let the mixture distribution of the classes l 6= k be given by
P
l6=k pl fl (x)
gk (x) = P
,
(13)
l6=k pl
P
S
with prior probability pg = l6=k pl . Then Xgk = l6=k Xl can be seen as a draw from
this mixture distribution. By Theorem 2 of Berisha et al. (2016) it holds that
Z
p
1
1
u(fk , gk ) ≤ min{pk fk (x), pg gk (x)} dx ≤ 12 − 12 u(fk , gk ).
(14)
2 − 2
The following theorem relates this error to the OvR-BER defined above.
Theorem 2. The error rate between class k and the mixture distribution without class
k is bounded from above by the OvR-BER,
Z
Qe,k = min{pk fk (x), pg gk (x)} dx ≤ Pe,k .
(15)
Proof. Note that
Z
Qe,k =
Z
pk fk (x) dx
pk fk (x)≤pg gk (x)
+
pg gk (x) dx.
pk fk (x)>pg gk (x)
14
(16)
To simplify the notation, introduce the sets
n
o
T = x ∈ Rd : pk fk (x) ≤ pg gk (x)
d
S = x ∈ R : pk fk (x) ≤ max{pl fl (x)}
(17)
and denote their respective complements by T 0 and S 0 . Then,
Z
Z
pg gk (x) dx
pk fk (x) dx +
Qe,k =
0
T
T
Z
Z
pg gk (x) dx.
pk fk (x) dx +
Pe,k =
(19)
l6=k
P
l6=k
pl fl (x) ≤ maxl6=k pl fl (x) it holds that S ⊆ T and T 0 ⊆ S 0 . Hence,
Z
Z
pk fk (x) dx =
Z
pk fk (x) dx +
T
pk fk (x) dx
Z
Z
pg gk (x) dx =
T0
(21)
T \S
S
Z
S0
(20)
S0
S
Since pg gk (x) =
(18)
pg gk (x) dx +
S 0 \T 0
pg gk (x) dx
However, the sets T \ S and S 0 \ T 0 both equal
d
U = x ∈ R : max{pl fl (x)} < pk fk (x) ≤ pg gk (x) ,
l6=k
(22)
(23)
so it follows that
Z
Qe,k = Pe,k +
U
pk fk (x) dx −
Z
U
pg gk (x) dx ≤ Pe,k
(24)
by definition of the set U .
This provides a lower bound for Pe,k in terms of u(fk , gk ). What remains to be
shown is that Pe,k has an upper bound in terms of u(fk , gk ). No such bound has yet
been found. However, the following result can be presented which does bound Pe,k from
above.
Theorem 3. For a single class k the OvR-BER is smaller than or equal to the sum of
the pairwise BER estimates involving class k.
Proof. Recall that the OvR-BER for class k is given by
Z
X
Pe,k =
pk fk (x) dx +
Z
pc fc (x) dx,
c6=k
pk fk (x)>max{pl fl (x)}
pk fk (x)≤max{pl fl (x)}
l6=k
(25)
l6=k
and denote the sum of the pairwise BERs involving k as Fe,k given by,
XZ
Fe,k =
min{pk fk (x), pc fc (x)} dx
(26)
c6=k
=
X
Z
pk fk (x) dx
c6=k
pk fk (x)<pc fc (x)
15
+
X
Z
pc fc (x) dx.
c6=k
pk fk (x)>pc fc (x)
(27)
Then comparing the first term of Fe,k with that of Pe,k shows
Z
X Z
pk fk (x) dx,
pk fk (x) dx ≥
c6=k
pk fk (x)<pc fc (x)
(28)
pk fk (x)≤max{pl fl (x)}
l6=k
since the area of integration on the left is larger than on the right. Similarly,
X Z
X Z
pc fc (x) dx
pc fc (x) dx ≥
(29)
c6=k
pk fk (x)>max{pl fl (x)}
c6=k
pk fk (x)>pc fc (x)
l6=k
for the same reason. This completes the proof.
B
Jensen-Shannon Bound Inequality
In this section a proof is given for the statement that the Henze-Penrose upper bound
on the Bayes error rate is tighter than the Jensen-Shannon upper bound derived by Lin
(1991). Before presenting the proof, the following lemma is presented.
Lemma 4. For x, y > 0 it holds that
y
x
4 log(2)xy
x log 1 +
+ y log 1 +
≥
.
x
y
x+y
Proof. Let t =
y
x
(30)
and multiply both sides by 1t + 1 > 0, then the inequality reduces to
1
1
+ 1 log (1 + t) + (1 + t) log 1 +
≥ 4 log(2).
(31)
t
t
Denote the left hand side by f (t). The proof will now proceed by showing that f (t) ≥
f (1) = 4 log(2) for all t > 0. The derivatives of f (t) are given by
1
log(1 + t)
2 log(1 + t) − t
f 0 (t) = log 1 +
−
and
f 00 (t) =
.
(32)
t
t2
t3
Write the numerator of f 00 (t) as g(t) such that
g(t) = 2 log(1 + t) − t,
and
g 0 (t) =
1−t
.
1+t
(33)
Then it is clear that g 0 (t) > 0 for 0 < t < 1 and g 0 (t) < 0 for t > 1. Furthermore
limt→0+ g(t) = 0 and limt→∞ g(t) = −∞. Thus, it follows that g(t) increases on 0 <
t < 1 and decreases for t > 1. Let t = α > 1 be such that g(α) = 0, then g(t) > 0 for
0 < t < α and g(t) < 0 for t > α.
From this it follows that f 00 (t) > 0 for 0 < t < α and f 00 (t) < 0 for t > α. Hence,
f 0 (t) is increasing on 0 < t < α and decreasing for t > α. Moreover, limt→0+ f 0 (t) = −∞
and f 0 (1) = 0. Thus, it follows that f 0 (t) is negative on 0 < t < 1, positive for t > 1,
and attains a maximum at t = α after which it decreases to limt→∞ f 0 (t) = 0. Since
limt→0+ f (t) = ∞ it follows that f (t) is decreasing on 0 < t < 1 and increasing for
t > 1.
16
Definition 5 (Kullback-Leibler Divergence). For probability density functions f1 (x)
and f2 (x) the Kullback-Leibler divergence is given by
Z
f1 (x)
DKL (f1 kf2 ) = f1 (x) log2
dx,
(34)
f2 (x)
Kullback and Leibler (1951).
Definition 6 (Jensen-Shannon Divergence). According to El-Yaniv et al. (1998) the
Jensen-Shannon divergence for two probability density functions f1 (x) and f2 (x) with
prior probabilities p1 and p2 , can be stated in terms of the Kullback-Leibler divergence
as
JS(f1 , f2 ) = p1 DKL (f1 kM ) + p2 DKL (f2 kM )
(35)
with M (x) = p1 f1 (x) + p2 f2 (x) the mixture distribution and p2 = 1 − p1 .
Theorem 7. For p1 = p2 = 21 the Henze-Penrose upper bound on the BER is tighter
than the Jensen-Shannon upper bound of Lin (1991),
Pe (f1 , f2 ) ≤
1
2
− 12 uHP ≤ 12 J,
(36)
where J = H(p1 ) − JS(f1 , f2 ) with H(p1 ) the binary entropy and JS(f1 , f2 ) the JensenShannon divergence.
Proof. First, note that with p1 = p2 = 21 the binary entropy H(p1 ) = H( 12 ) = 1. Second,
for the equiprobable case it holds that
Z
f1 (x)f2 (x)
1
1
dx.
(37)
2 − 2 uHP (f1 , f2 ) =
f1 (x) + f2 (x)
The Jensen-Shannon upper bound can be written as
Z
Z
2f1 (x)
2f2 (x)
1
1
1
1
f1 (x) log2
dx − 4 f2 (x) log2
dx
2J = 2 − 4
f1 (x) + f2 (x)
f1 (x) + f2 (x)
Z
Z
2f1 (x)
dx
= 41 f1 (x) + f2 (x) dx − 14 f1 (x) log2
f1 (x) + f2 (x)
Z
2f2 (x)
1
− 4 f2 (x) log2
dx
f1 (x) + f2 (x)
Z
2f1 (x)
1
= 4 f1 (x) 1 − log2
dx
f1 (x) + f2 (x)
Z
2f2 (x)
1
+ 4 f2 (x) 1 − log2
dx
f1 (x) + f2 (x)
Z
f2 (x)
f1 (x)
1
= 4 f1 (x) log2 1 +
+ f2 (x) log2 1 +
dx
f1 (x
f2 (x)
By Lemma 4 it follows that
f2 (x)
f1 (x)
4f1 (x)f2 (x)
f1 (x) log2 1 +
+ f2 (x) log2 1 +
≥
,
f1 (x
f2 (x)
f1 (x) + f2 (x)
and therefore
1
2J
≥
1
4
Z
4f1 (x)f2 (x)
dx =
f1 (x) + f2 (x)
17
1
2
− 12 uHP (f1 , f2 ).
(38)
(39)
(40)
(41)
(42)
(43)
(44)
Table 4: Average training time per hyperparameter configuration in seconds, averaged
over the 5 nested CV folds. Minimal values per dataset are underlined.
Dataset
C&S
DAG GenSVM
OvO
OvR SmartSVM
abalone
702
7.712
254
7.067 11.282
3.960
fars
8329
164
601
151
296
184
flare
24.606
0.369
1.467
0.344
0.431
0.263
krkopt
4526 38.337
165 35.813 73.053
35.627
letter
1668 21.438
131 20.076 94.977
34.454
nursery
119
3.932
3.985
3.617
4.994
3.133
optdigits
4.630
1.053
92.441
0.971
1.278
0.590
pageblocks 27.550
1.408
5.852
1.326
2.559
1.504
pendigits
132
1.589
19.273
1.498
7.965
2.872
satimage
273
3.931
7.494
3.685
9.918
3.942
segment
19.846
0.376
4.509
0.370
0.985
0.285
shuttle
403 30.042
42.744 29.571
105
43.928
texture
30.954
1.816
43.200
1.727
3.665
1.028
winered
37.179
0.646
1.587
0.586
0.966
0.465
winewhite
135
3.115
7.514
2.903
3.732
1.954
yeast
63.655
0.877
3.327
0.807
1.272
0.699
C
Additional Simulation Results
In this section some additional simulation results are presented for the SmartSVM experiments presented in Section 4. Table 4 shows the average time per hyperparameter
configuration for each of the methods. This is especially useful for comparing GenSVM
(Van den Burg and Groenen, 2016) with the other methods, as it has a larger set of
hyperparameters to consider.
A commonly used tool to summarize results of simulation experiments is to use rank
plots (Demšar, 2006). For each dataset the methods are ranked, with the best method
receiving rank 1 and the worst method receiving rank 6 (since there are 6 methods in
this experiment). In case of ties fractional ranks are used. By averaging the ranks over
all datasets, a visual summary of the results can be obtained. Figures 7a, 7b and 7c
show these average ranks for predictive performance, total training time, and average
training time respectively.
The ordering of OvO and SmartSVM in the rank plots for training time may seem
counterintuitive, considering that SmartSVM is more often the fastest method. This
can be explained by the fact that in the cases where SmartSVM is slower than OvO it
is usually also slower than DAG. In contrast, where SmartSVM is the fastest method
OvO is usually the second fastest method. Because of this, SmartSVM obtains a slightly
higher average rank than OvO.
18
OvO
1.0
1.5
DAG
2.0
C&S
2.5
GenSVM
3.0
3.5
SmartSVM
4.0
4.5
OvR
5.0
5.5
6.0
(a) Predictive performance
OvO
1.0
SmartSVM
1.5
2.0
DAG
2.5
OvR
3.0
3.5
C&S
4.0
4.5
5.0
GenSVM
5.5
6.0
(b) Total training time
OvO
1.0
1.5
SmartSVM
2.0
DAG
2.5
OvR
3.0
3.5
4.0
GenSVM
4.5
5.0
C&S
5.5
6.0
(c) Average training time
Figure 7: Rank plots of classifier performance in the simulation study. Figure (a) shows
the average ranks for out-of-sample predictive performance as measured by the ARI.
Figures (b) and (c) respectively show the average ranks for total training time and
average training time per hyperparameter configuration.
References
Alcalá, J., Fernández, A., Luengo, J., Derrac, J., Garcı́a, S., Sánchez, L., and Herrera, F.
Keel data-mining software tool: data set repository, integration of algorithms and experimental analysis framework. Journal of Multiple-Valued Logic and Soft Computing,
17(2-3):255–287, 2010.
Avi-Itzhak, H. and Diep, T. Arbitrarily tight upper and lower bounds on the bayesian
probability of error. IEEE Transactions on Pattern Analysis and Machine Intelligence,
18(1):89–91, 1996.
Bache, K. and Lichman, M. UCI machine learning repository. 2013.
Belkin, M. and Niyogi, P. Laplacian eigenmaps for dimensionality reduction and data
representation. Neural computation, 15(6):1373–1396, 2003.
Berisha, V. and Hero, A. Empirical non-parametric estimation of the Fisher information.
Signal Processing Letters, IEEE, 22(7):988–992, 2015.
Berisha, V., Wisler, A., Hero, A., and Spanias, A. Empirically estimable classification
bounds based on a nonparametric divergence measure. IEEE Transactions on Signal
Processing, 64(3):580–591, 2016.
Bhattacharyya, A. On a measure of divergence between two multinomial populations.
Sankhyā: The Indian Journal of Statistics, 7(4):401–406, 1946.
Borůvka, O. O jistém problému minimálnı́m. Práce Moravské Pridovedecké Spolecnosti,
3:37–58, 1926.
Chang, C. and Lin, C. LIBSVM: A library for support vector machines. ACM Transactions on Intelligent Systems and Technology, 2(3):27:1–27:27, 2011.
19
Chernoff, H. A measure of asymptotic efficiency for tests of a hypothesis based on the
sum of observations. The Annals of Mathematical Statistics, 23(4):493–507, 1952.
Cortes, C. and Vapnik, V. Support-vector networks. Machine Learning, 20(3):273–297,
1995.
Cox, D. The regression analysis of binary sequences. Journal of the Royal Statistical
Society. Series B (Methodological), 20(2):215–242, 1958.
Crammer, K. and Singer, Y. On the algorithmic implementation of multiclass kernelbased vector machines. The Journal of Machine Learning Research, 2(Dec):265–292,
2002.
Csiszár, I. I-divergence geometry of probability distributions and minimization problems.
The Annals of Probability, 3(1):146–158, 1975.
Demšar, J. Statistical comparisons of classifiers over multiple data sets. The Journal of
Machine Learning Research, 7(Jan):1–30, 2006.
El-Yaniv, R. and Etzion-Rosenberg, N. Hierarchical multiclass decompositions with
application to authorship determination. arXiv preprint arXiv:1010.2102, 2010.
El-Yaniv, R., Fine, S., and Tishby, N. Agnostic classification of markovian sequences.
In: Advances of Neural Information Processing Systems 10, pp. 465–471. MIT Press,
1998.
Fan, R.E., Chang, K.W., Hsieh, C.J., Wang, X.R., and Lin, C.J. LIBLINEAR: A
library for large linear classification. The Journal of Machine Learning Research,
9(Aug):1871–1874, 2008.
Frank, E. and Kramer, S. Ensembles of nested dichotomies for multi-class problems.
In: Proceedings of the 21st International Conference on Machine Learning, pp. 39–46.
ACM, 2004.
Friedman, J. and Rafsky, L. Multivariate generalizations of the Wald-Wolfowitz and
Smirnov two-sample tests. The Annals of Statistics, 7(4):697–717, 1979.
Fukunaga, K. Introduction to Statistical Pattern Recognition. Academic Press, 1990.
Hashlamoun, W., Varshney, P., and Samarasooriya, V. A tight upper bound on the
bayesian probability of error. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 16(2):220–224, 1994.
Henze, N. and Penrose, M. On the multivariate runs test. Annals of statistics, 27(1):290–
298, 1999.
Ho, T. and Basu, M. Complexity measures of supervised classification problems. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 24(3):289–300, 2002.
Hubert, L. and Arabie, P. Comparing partitions. Journal of Classification, 2(1):193–218,
1985.
Kailath, T. The divergence and Bhattacharyya distance measures in signal selection.
IEEE Transactions on Communication Technology, 15(1):52–60, 1967.
20
Keerthi, S., Sundararajan, S., Chang, K.W., Hsieh, C.J., and Lin, C.J. A sequential
dual method for large scale multi-class linear SVMs. In: Proceedings of the 14th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, pp.
408–416. 2008.
Kreßel, U. Pairwise classification and support vector machines. In: B. Schölkopf, C.J.C.
Burges, and A.J. Smola, editors, Advances in Kernel Methods, pp. 255–268. MIT
Press, 1999.
Kullback, S. and Leibler, R. On information and sufficiency. The Annals of Mathematical
Statistics, 22(1):79–86, 1951.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
Lin, J. Divergence measures based on the shannon entropy. IEEE Transactions on
Information Theory, 37(1):145–151, 1991.
Lorena, A. and De Carvalho, A. Building binary-tree-based multiclass classifiers using
separability measures. Neurocomputing, 73(16):2837–2845, 2010.
Lorena, A., De Carvalho, A., and Gama, J. A review on the combination of binary
classifiers in multiclass problems. Artificial Intelligence Review, 30(1):19–37, 2008.
Lu, Y., Guo, H., and Feldkamp, L. Robust neural learning from unbalanced data samples. In: Proceedings of the IEEE International Joint Conference on Neural Networks,
volume 3, pp. 1816–1821. IEEE, 1998.
March, W., Ram, P., and Gray, A. Fast Euclidean minimum spanning tree: algorithm,
analysis, and applications. In: Proceedings of the 16th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, pp. 603–612. ACM, 2010.
McCulloch, W. and Pitts, W. A logical calculus of the ideas immanent in nervous
activity. The Bulletin of Mathematical Biophysics, 5(4):115–133, 1943.
Pearson, K. On lines and planes of closest fit to systems of points in space. The London,
Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11):559–572,
1901.
Platt, J., Cristianini, N., and Shawe-Taylor, J. Large margin DAGs for multiclass
classification. In: S.A. Solla, T.K. Leen, and K. Müller, editors, Advances in Neural
Information Processing Systems 12, pp. 547–553. MIT Press, 2000.
Qiao, X. and Liu, Y. Adaptive weighted learning for unbalanced multicategory classification. Biometrics, 65(1):159–168, 2009.
Santos, J. and Embrechts, M. On the use of the adjusted rand index as a metric
for evaluating supervised classification. In: Proceedings of the 19th International
Conference on Artificial Neural Networks: Part II, pp. 175–184. Springer-Verlag, 2009.
Schwenker, F. and Palm, G. Tree-structured support vector machines for multi-class
pattern recognition. In: International Workshop on Multiple Classifier Systems, pp.
409–417. Springer, 2001.
Stoer, M. and Wagner, F. A simple min-cut algorithm. Journal of the ACM, 44(4):585–
591, 1997.
21
Stone, M. Cross-validatory choice and assessment of statistical predictions. Journal of
the Royal Statistical Society. Series B (Methodological), 36(2):111–147, 1974.
Takahashi, F. and Abe, S. Decision-tree-based multiclass support vector machines. In:
Neural Information Processing, 2002. ICONIP’02. Proceedings of the 9th International Conference on, volume 3, pp. 1418–1422. IEEE, 2002.
Tibshirani, R. and Hastie, T. Margin trees for high-dimensional classification. Journal
of Machine Learning Research, 8(Mar):637–652, 2007.
Van den Burg, G. and Groenen, P. GenSVM: A generalized multiclass support vector
machine. Journal of Machine Learning Research, 17(225):1–42, 2016.
Vapnik, V. Statistical learning theory. Wiley, New York, 1998.
Vural, V. and Dy, J. A hierarchical method for multi-class support vector machines. In:
Proceedings of the 21st International Conference on Machine Learning, pp. 105–112.
ACM, 2004.
Wald, A. Foundations of a general theory of sequential decision functions. Econometrica,
15(4):279–313, 1947.
Whitney, V. Algorithm 422: minimal spanning tree. Communications of the ACM,
15(4):273–274, 1972.
Wisler, A., Berisha, V., Wei, D., Ramamurthy, K., and Spanias, A. Empiricallyestimable multi-class classification bounds. In: 2016 IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), pp. 2594–2598. IEEE, 2016.
22
| 7cs.IT
|
1
Decentralized Clustering based on Robust
Estimation and Hypothesis Testing
Dominique Pastor, Elsa Dupraz, François-Xavier Socheleau
arXiv:1706.03537v1 [math.ST] 12 Jun 2017
IMT Atlantique, Lab-STICC, Univ. Bretagne Loire, Brest, France
Abstract
This paper considers a network of sensors without fusion center that may be difficult to set up in
applications involving sensors embedded on autonomous drones or robots. In this context, this paper
considers that the sensors must perform a given clustering task in a fully decentralized setup. Standard
clustering algorithms usually need to know the number of clusters and are very sensitive to initialization,
which makes them difficult to use in a fully decentralized setup. In this respect, this paper proposes a
decentralized model-based clustering algorithm that overcomes these issues. The proposed algorithm is
based on a novel theoretical framework that relies on hypothesis testing and robust M-estimation. More
particularly, the problem of deciding whether two data belong to the same cluster can be optimally solved
via Wald’s hypothesis test on the mean of a Gaussian random vector. The p-value of this test makes it
possible to define a new type of score function, particularly suitable for devising an M-estimation of the
centroids. The resulting decentralized algorithm efficiently performs clustering without prior knowledge
of the number of clusters. It also turns out to be less sensitive to initialization than the already existing
clustering algorithms, which makes it appropriate for use in a network of sensors without fusion center.
I. I NTRODUCTION
Networks of sensors are now used in a wide range of applications in medicine, in telecommunications,
or in environmental domains [1]. They are employed, for example, for human health monitoring [2],
activity recognition on home environments [3], spectrum sensing in cognitive radio [4], and so forth.
In these applications, a fusion center can collect all the data from all the sensors and perform a given
estimation or learning task over the collected data. However, it is not always practical to set up a fusion
center, especially in recent applications involving autonomous drones or robots [5]. In such applications
in which no fusion center is available, the sensors should perform the learning task by themselves in a
fully decentralized setup.
In this paper, we assume that the sensors have to perform decentralized clustering on the data measured
within the network. The aim of clustering is to divide the data into clusters such that the data inside a
June 13, 2017
DRAFT
2
cluster are similar with each other and different from the data that belong to other clusters. Clustering is
considered in various applications of sensor networks, such as parking map construction [6] or controller
placement in telecommunication networks [7]. One of the most popular clustering algorithms is Kmeans [8], due to its simplicity and its efficiency.
The K-means algorithm groups N measurement vectors into K clusters with a two-step iterative procedure. It was proved that this iterative procedure always converges to a local minimum [8]. However, in
order to get a chance to reach the global minimum, the K-means algorithm needs to be initialized properly.
Proper initialization can be obtained with the K-means++ procedure [9], which requires computing all the
two by two distances between all the measurement vectors in the dataset. As another issue, the K-means
algorithm need to know the number K of clusters. When K unknown, it is possible to apply a penalized
method that requires applying the K-means algorithm several times with different numbers of cluster [10].
It is worth mentioning that the variants of K-means such as Fuzzy K-means [11] suffer from the same
two issues.
The K-means algorithm was initially introduced for non-distributed setups, but decentralized versions
of the algorithm have also been proposed [12]–[14]. In a decentralized setup, each of the N sensors
initially observes one single vector that correspond to its own measurements. Then, in order to apply
the decentralized K-means algorithm, the sensors are allowed to exchange some data with each other.
The objective of decentralized clustering is thus to allow each sensor to perform the clustering with only
partial observations of the available data, while minimizing the amount of data exchanged in the network.
As a result, in this context, it is not desirable to initialize the algorithm with the K-means++ procedure
that would require exchanging all the two by two distances between all the measurement vectors. It is
not desirable either to perform the distributed algorithm several times in order to determine the number
of clusters.
The above limitations have not been addressed in the previous works [12]–[14], and the objective of
this paper is thus to propose a decentralized clustering algorithm that overcomes them. At first sight, the
algorithms DB-SCAN [15] and OPTICS [16] may appear as good candidates for decentralized clustering
since they do need the number of clusters and since they do not have any initialization issues. However,
they require setting two parameters that are the maximum distance between two points in a cluster and
the minimum number of points per cluster. These parameters can hardly be estimated and they must be
chosen empirically, which we would like to avoid. This is why we do not consider these solutions here.
The K-means algorithm makes no assumption on the signal model of the measurement vectors that
belong to a cluster, which is relevant for applications such as document classification [17] information
retrieval, or categorical data clustering [18]. On the other hand, signal processing methods usually assume
June 13, 2017
DRAFT
3
a statistical model on the measurements. This model can be derived, for example, from physical characteristics of the sensors. In this respect, centralized and decentralized model-based clustering algorithms
were proposed in [19], [20], although they suffer from the same two issues as the K-means algorithms.
In the model-based clustering algorithms proposed in [19], [20], the measurement vectors that belong
to a given cluster are modeled as the cluster centroid plus Gaussian noise, and it is assumed that both
the cluster centroid and the noise variance are unknown. However, the noise variance may be estimated,
possibly from preliminary measurements, via a bunch of parametric, non-parametric, and robust methods
(see [21]–[23], among others). Therefore, here, we will consider the same Gaussian model as in [19], [20],
but we will assume that the noise variance is known. This assumption was already made for clustering
in [6] and [11] in order to choose the parameters for the functions that compute the cluster centroids.
In what follows, under the assumption that the noise variance is known, we propose a novel clustering
algorithm which does not require prior knowledge of the number of clusters and which is much less
sensitive to initialization than the K-means algorithm. The centralized and decentralized versions of the
algorithm we propose are both based on the robust estimation of the cluster centroids and on the testing
of whether a given measurement vector belongs to a given cluster. Whence the names CENTREx — for
CENtroids Testing and Robust Estimation — and DeCENTREx, respectively given to the centralized and
the decentralized algorithm.
In both algorithms, the cluster centroids are estimated one after the other via robust M-estimation [24],
assuming that the measurement vectors from the other clusters are outliers. In order to estimate the
centroids, M-estimation looks for the fixed points of a function whose expression depend on a score
function applied to all the measurement vectors of the database. The score function we choose is the
p-value of the Wald hypothesis test for testing the mean of a Gaussian [25] and evaluates the plausibility
that a measurement vector belongs to a given cluster. M-estimation was already used in [11] to estimate
the cluster centroids, with a different score function. In [11], the robustness of the centroid estimation
was evaluated from the standard M-estimation approach, which shows that an outlier of infinite amplitude
gives only a finite estimation error. Here, we propose an alternative analysis that, unlike [11], takes into
account the fact that the outliers are measurement vectors from other clusters. Our asymptotic analysis
shows that the only fixed points of our M-estimation function are the true cluster centroids, which validates
our approach. We also derive the statistics of the estimation error for a finite number of measurement
vectors.
In our algorithm, for each centroid to be estimated, the iterative computation of one of the fixed points
of the M-estimation function is simply initialized with one of the measurement vectors of the dataset. The
iterative computation then retrieves the centroid of the cluster to which the initialization point belongs.
June 13, 2017
DRAFT
4
After each centroid estimation, a Wald hypothesis test is applied to mark all the measurement vectors that
must not be used later for initializing the estimation of any other centroid because they are already close
enough to the newly estimated one. This very simple marking operation avoids using the K-means++
solution. Further, the estimation process stops when all the measurement vectors have been marked,
which permits to determine the number of clusters. The final clustering is standardly performed by
seeking the estimated centroid that is the closest to a given observation. Our simulation results show that
both CENTREx and DeCENTREx achieve performance close to the K-means algorithm initialized with
the proper number of clusters.
The outline of the paper is as follows. Section II describes the signal model we consider for the
measurement vectors. Section III details the theoretical framework, which involves the derivation of all
the functions and hypothesis tests used in the successive steps of the algorithm. Section IV describes
the centralized and decentralized versions of the clustering algorithm we propose. Section V shows the
simulation results and Section VI gives our conclusions and perspectives for future works.
II. S IGNAL MODEL
This section describes the signal model and the notation used throughout the paper. Consider a network
of N sensors in which each sensor n ∈ {1, · · · , N } observes a measurement vector Yn . The vectors
Y1 , . . . , YN are assumed to be N independent d-dimensional random Gaussian vectors. The individual
random components of each random vector Yn are assumed to be independent and identically distributed
(i.i.d.). Note that this i.i.d. simplyfing assumption is considered here as a fist step to introduce the
analysis and more accurate models will be considered in future works (for example, observations with
different variances, see conclusion for more details). To alleviate notation in upcoming computations,
we conveniently assume that the covariance matrices of the observations Yn are normalized so as to all
equal the d × d identity matrix Id . This normalization requires prior knowledge of the noise variance,
which may be known or estimated by various parametric and nonparametric methods, as mentioned in
the introduction.
We further assume that the measurement vectors are split into K clusters defined by K centroids
θ1 , . . . , θK , with θk ∈ Rd for each k = 1, 2, . . . , K . Therefore, according to the foregoing assumption
on the noise measurement, we assume that for each n ∈ {1, . . . , N }, there exists k ∈ {1, 2, . . . , K} such
that Yn ∼ N (θk , Id ) and we say that Yn belongs to cluster k . For each k , we denote by Nk the number
of measurement vectors in cluster k . We also use the notation Yk,1 , Yk,2 , . . . , Yk,Nk to designate the Nk
observations Yn that belong to a given cluster k . In the following, we assume that the number of clusters
K , the centroids θ1 , . . . , θK , and the number of measurement vectors in each cluster N1 , · · · , NK are
June 13, 2017
DRAFT
5
unknown. The objective of the clustering algorithm we propose is to estimate these quantities and also
to determine to which cluster each measurement vector Yn belongs.
III. T HEORETICAL FRAMEWORK
In our algorithm, the centroids are estimated one after each other by a robust M-estimation approach.
As described in [24] and references therein, the M-estimation involves searching the fixed points of a
given function. When the algorithm has estimated a new centroid, it identifies and marks the observations
Yn that are too close to the newly estimated cluster to be used later for estimating other centroids. For
this, we apply a Wald hypothesis test [25] that decides whether a given vector must be marked or not.
In this section, we introduce the theoretical framework on which rely the different parts of the algorithm.
In particular, we present the M-estimation of the centroids and the hypothesis test used to build up the
clusters. As described in the following, the function we use for M-estimation is deduced from Wald’s
test. We further analytically show that the centroids are the only fixed-points of this function, which
justifies that our algorithm can successfully recover the clusters.
A. The Wald test and its p-value for testing the mean of a Gaussian
Wald tests are optimal in a specific sense defined in [25, Definition III, p. 450] to test the mean of a
Gaussian [25, Proposition III, p. 450]. With respect to the model assumptions introduced in Section II,
Wald tests are hereafter exposed where the Gaussian has identity scale covariance matrix. In the sequel,
Wald tests will serve: (i) to define the function we use for M-estimation of the centroids, (ii) to mark
the measurement vectors that are close to the estimated centroids. Here, we first describe the test in a
generic way and we will apply it later in the section for the two purposes recalled above.
Let Z be a real d-dimensional random vector such that Z ∼ N (ξ, σ02 Id ) with σ0 6= 0. Consider the
problem of testing the mean of the Gaussian vector Z , that is the problem of testing the null hypothesis
H0 : ξ = 0 against its alternative H1 : ξ 6= 0. This problem is summarized as:
Observation: Z ∼ N (ξ, σ02 Id ),
H : ξ = 0,
0
Hypotheses:
H : ξ 6= 0.
1
(1)
Recall that a non-randomized test is any (measurable) map of Rd to {0, 1} and that, given some test T
and z ∈ Rd , the value T(z) is the index of the hypothesis accepted by T at z . For instance, if T(z) = 0
(resp. T(z) = 1), T accepts H1 (resp. H0 ) when z is the realization of Z . We can then devise easily
June 13, 2017
DRAFT
6
a non-randomized test that guarantees a given false alarm probability γ ∈ (0, 1) for testing H0 against
H1 . Indeed, for any given λ ∈ [0, ∞), let Tλ be the test defined for any z ∈ Rd by setting:
0 if kzk 6 λ
Tλ (z) =
1 if kzk > λ.
(2)
where k • k denotes the usual Euclidean norm in Rd . This test accepts H0 (resp. H1 ) if kzk 6 λ (resp.
kzk > λ). According to the definition of the Generalized Marcum Function Qd/2 (•, •) [26, Eq. (8)],
the false alarm probability of this test is P kZk2 > λ2 = 1 − Fχ2d (0) (λ2 /σ02 ) = Qd/2 (0, λ/σ0 ) where
Z ∼ N (0, σ02 Id ) and Fχ2d (0) is the cumulative distribution function (CDF) of the centered χ2d distribution
with d degrees of freedom. Throughout the paper, we denote by µ(γ) the unique real value such that:
Qd/2 (0, µ(γ)) = γ
(3)
It then follows that the false alarm probability of the test Tσ0 µ(γ) equates the desired value γ .
Although there is no Uniformly Most Powerful (UMP) test for the composite binary hypothesis testing
problem (1) [27, Sec. 3.7], Tσ0 µ(γ) turns out to be optimal with respect to several optimality criteria and
within several classes of tests with level γ [28, Proposition 2]. In particular, Tσ0 µ(γ) is UMP among all
spherically invariant tests since it has Uniformly Best Constant Power (UBCP) on the spheres centered
at the origin of Rd [25, Definition III & Proposition III, p. 450]. In the sequel, any test Tσ0 µ(γ) will be
called a Wald test, without recalling explicitly the level γ at which the testing is performed.
It turns out that a notion of p-value can be defined for the family of Wald tests Tσ0 µ(γ) when γ ranges
in (0, 1). This p-value is calculated in Appendix A and, for the testing problem (1), it is given for any
z ∈ Rd by:
γ
bσ0 (z) = Qd/2 (0, kzk/σ0 ).
(4)
The p-value can be seen as a measure of the plausibility of the null hypothesis H0 . In particular, when
kzk tends to +∞, γ
bσ0 (z) tends to 0 since lim Qd/2 (0, t) = lim (1− Fχ2d (0) (t2 )) = 0. It is then natural to
t→∞
t→∞
consider that the plausibility of H0 vanishes as kzk grows to +∞. Similarly, γ
bσ0 (z) tends to 1 when kzk
tends to 0, so that the plausibility of H0 is rather high, close to 1, for small values of kzk. Accordingly,
we describe the M-estimation of the centroids and show how the Wald test and its p-value help us choose
the score function that will be used in the M estimation.
B. M-estimation of the centroids
As in [11], we want to estimate the centroid θk of a given cluster with an M-estimator, which amounts
to considering that the measurement vectors from other clusters are outliers. More specifically, if the
number K of clusters were known, robust estimation theory [21], [24], [29]–[31] applied to the problem
June 13, 2017
DRAFT
7
of estimating the centroids θ1 , · · · , θK would lead to calculating the solutions θb1 , . . . , θbK of the K
equations:
(θb1 , . . . , θbK ) = arg min J(θ1 , . . . , θK ),
(5)
(θ1 ,...,θK )
P
PNk
2 and the loss function ρ : R → R is an increasing
where J(θ1 , . . . , θK ) = K
k=1
n=1 ρ kYk,n − θk k
function. If the loss function ρ is differentiable, the solution to (5) can be found by solving the K
equations
∂k J(θ1 , . . . , θK ) =
Nk
X
(Yk,n − θk )w(kYk,n − θk k2 )
n=1
=2
Nk
X
Ψ(Yk,n − θk ) = 0
(6)
n=1
for k = 1, . . . , K , where ∂k J is the partial derivate of J with respect to its k th argument, w = ρ0 and
Ψ : Rd → Rd is defined for every x ∈ Rd by Ψ(x) = x w(kxk2 ). The function Ψ is called the score
function, and it is non negative since ρ increases. Rewriting (6) for each given k ∈ J1, KK gives that the
solution θbk of this equation must verify:
PNk
b 2
n=1 w(kYk,n − θk k )Yk,n
θbk = P
,
Nk
bk k2 )
w(kY
−
θ
k,n
n=1
(7)
where w is henceforth called the weight function. In other words, the estimate θbk is a fixed point of the
function
PNk
gk (x) =
2
n=1 w(kYk,n − xk )Yk,n
, x ∈ Rd
P
Nk
2)
w(kY
−
xk
k,n
n=1
(8)
The computation of the fixed points of this function can be carried out by iterative algorithms described
in [24].
Unfortunately, in our context, we cannot perform the M-estimation of the centroids by seeking the fixed
points of each gk defined by (8) since the number K of clusters is unknown. However, M-estimation
is supposed to be robust to outliers and, when estimating a centroid θk , the measurement vectors from
other clusters may be seen as outliers. Consequently, we can expect that if θbk is a fixed point of gk , then
it should also be a fixed point of the function:
PN
2
n=1 w(kYn − xk )Yn
hN (x) = P
, x ∈ Rd ,
N
2
n=1 w(kYn − xk )
(9)
where the sums involve now all the measurement vectors Yn . The rationale is that the contributions in
hN of the measurement vectors sufficiently remote from a fixed point θbk should significantly be lowered
by the weight function w, given that this weight function is robust to outliers as discussed later in the
paper. Note that the function hN was also considered in [11] for the estimation of the centroids, even
though the number of clusters was assumed to be known.
June 13, 2017
DRAFT
8
At the end, the key-point for robustness to outliers of an M-estimator is the choice of the weight
function w. “Usually, for robust M-estimators, the weights are chosen close to one for the bulk of the
data, while outliers are increasingly downweighted.” [24]. This standard rationale leads the authors in
[11] to choose the Gaussian kernel w(x) = exp −βkxk2 , in which the parameter β is homogeneous
to a noise variance. When d = 2, it is experimentally shown in [11] that the parameter β should be set
to 1/σ 2 . However, when β > 2, no such result exist and the value of β must be chosen empirically (the
best value of β is usually different from 1/σ 2 ).
C. Wald p-value kernel for M-estimation
Here, in order to avoid the empirical choice of the parameter β , we alternatively derive our weight
function from a Wald hypothesis test on the measurement vectors.
Consider the problem of testing whether two random vectors Yi and Yj (i 6= j ) belong to the same
cluster. This problem can be formulated as testing whether θi = θj or not, given Yi ∼ N (θi , Id ) and
Yj ∼ N (θj , Id ). The independence of Yi and Yj implies that Yi − Yj ∼ N (θi − θj , 2Id ). Therefore,
testing θi = θj amounts to testing whether the mean of the Gaussian random vector Yi − Yj is 0. This
√
is Problem (1) with σ0 = 2.
The p-value (4) of the Wald test Tµ(γ) basically measures the plausibility that Yi and Yj belong
to the same cluster. For testing the mean of Yi − Yj ∼ N (θi − θj , 2Id ), the family of Wald tests is
T√2µ(γ) : γ ∈ (0, 1) . The p-value associated with this family of tests follows from (4) and is equal to
√
√
γ
b√2 (z) = Qd/2 (0, kzk/ 2) for any z ∈ Rd . When the p-value γ
b√2 (Yi −Yj ) = Qd/2 (0, kYi −Yj k/ 2) is
low, the null hypothesis should be rejected, that is, the two observations should be regarded as emanating
from two different clusters. In contrast, when this p-value is large, the null hypothesis should be accepted,
that is, the two observations should be considered as elements of the same clusters. This p-value thus
basically satisfies the fundamental requirement for a weight function in robust M-estimation. Thence the
√
idea to put w(kYn − xk2 ) = γ
b√2 (Yn − x) = Qd/2 (0, kYn − xk/ 2) in the expression of hN given by
Eq. (9). Accordingly, the weight function for the M-estimation of the centroids is thus defined as:
w(u) = Qd/2 (0,
p
u/2)
(10)
√
for any u ∈ R, so that w(kzk2 ) = Qd/2 (0, kzk/ 2) for all z ∈ Rd . Because of its derivation,
this weight function is hereafter called Wald p-value kernel. It follows from this choice that ρ(u) =
p
Ru
t/2)dt + C where C is a constant and, as a consequence, the score function is given for
0 Qd/2 (τ,
any x ∈ Rd by:
June 13, 2017
√
Ψ(x) = x Qd/2 (0, kxk/ 2).
(11)
DRAFT
9
From an M-estimation point of view, the relevance of the Wald p-value kernel w can be further
emphasized by calculating and analyzing the influence function of the M-estimator (7). This analysis
is carried out in Appendix B and shows that the weight function given by (10) is robust to outliers.
However, this analysis does not seem to be sufficient for the clustering problem in which outliers may
be numerous since they are in fact measurement vectors belonging to other clusters. At this stage, the
question (which is not answered in [11], even for the Gaussian kernel) is thus whether the centroids
can still be expected to be fixed-points of the function hN defined in (9). The next section brings an
asymptotic answer to this question.
D. Fixed points analysis
The objective of this section is to support our choice of the Wald p-value kernel w and to show
that the centroids can be estimated by seeking the fixed-points of hN . For this, the following proposition
proves that the centroids θ1 , · · · , θK are the fixed-points of hN when the number of measurement vectors
asymptotically increases and when the centroids are asymptotically far away from each other.
Proposition 1. Let θ1 , . . . , θK be K pairwise different elements of Rd . For each k ∈ J1, KK, suppose that
P
iid
Yk,1 , . . . , Yk,Nk ∼ N (θk , Id ) are Nk independent random vectors and set N = K
k=1 Nk . If there exist
α1 , . . . , αK ∈ (0, 1) such that lim Nk /N = αk , then, for any i ∈ J1, KK and any θ in a neighborhood
N →∞
of θi ,
lim
∀k6=i,kθk −θi k→∞
lim
N →∞
hN (θ) − θ
=0
iff θ = θi
iid
Proof: For any k ∈ J1, KK and any n ∈ J1, Nk K, set Xk,n = Yk,n − θk , so that: Xk,1 , . . . , Xk,Nk ∼
P
N (0, Id ). For each k ∈ J1, KK, we also set αk,N = Nk /N . We therefore have K
k=1 αk = 1. The random
function (9) can then be rewritten as:
hN (θ) =
with
UN (θ)
VN (θ)
(θ ∈ Rd )
Nk
K X
X
U
(θ)
=
w(kYk,n − θk2 )Yk,n
N
k=1 n=1
Nk
K X
X
V
(θ)
=
w(kYk,n − θk2 ).
N
(12)
(13)
k=1 n=1
We then have:
hN (θ) − θ =
WN (θ)
VN (θ)
with WN (θ) = UN (θ) − VN (θ)θ . By setting ∆k = θk − θ for k ∈ J1, Nk K, we can write:
June 13, 2017
DRAFT
10
Nk
K
X
1
1 X
w(k∆k + Xk,n k2 ) (∆k + Xk,n )
WN (θ) =
αk,N
N
Nk
(14)
Nk
K
X
1
1 X
w(k∆k + Xk,n k2 )
VN (θ) =
αk,N
N
Nk
(15)
n=1
k=1
In the same way:
n=1
k=1
By the strong law of large numbers, it follows from (14) and (15) that:
αi E w(kZ(∆i
)k2 )Z(∆
i)
+
K
X
αk E w(kZ(∆k )k2 )Z(∆k )
k=1,k6=i
lim (hN (θ) − θ) =
N →∞
αi E [ w(kZ(∆i )k2 ) ] +
K
X
(a-s)
αk E w(kZ(∆k )k2 )
k=1,k6=i
with Z(∆k ) = ∆k + X for each k ∈ J1, KK and X ∼ N (0, Id ). If θ is in a neighborhood of θi , there
exists some positive real number ε > 0 such that k∆i k 6 ε. According to Lemma 2 stated and proved
in Appendix C,
lim
∀k6=i,kθk −θi k→∞
lim
N →∞
hN (θ) − θ
E w(kZ(∆i )k2 )Z(∆i )
=
E [ w(kZ(∆i )k2 ) ]
(a-s)
Since E w(kZ(∆i )k2 ) > 0, the left hand side (lhs) to the equality above is 0 if and only if
E w(kZ(∆i )k2 )Z(∆i ) = 0. The conclusion follows from Lemma 3 in Appendix D.
The results of Proposition 1 state that the centroids are the unique fixed points of the function hN ,
when the sample size and the distances between centroids tend to infinity. In CENTREx, an iterative
procedure [24] is used to seek the fixed points of the function hN . The fixed points are determined
one after the other, and in order to find one fixed point, the iterative procedure is initialized with a
measurement vector that has not been marked yet. The marking operation is applied after each centroid
estimation and consists of applying a Wald test aimed at finding the measurement vectors that have the
same mean as the newly estimated centroid, in the sense of Section III-A. In order to define the Wald
test that will be used to mark the measurement vectors, we need the statistical model of the estimated
centroids. This statistical model is derived in the next section.
E. Fixed point statistical model and fusion
In order to apply the Wald test for the marking operation, we need a model of the statistical behavior
of the fixed points of hN . In practice, the test will be applied when the sample size and the distances
between centroids may not be large enough to consider the asymptotic conditions of Proposition 1. That
is why we determine the statistical model of the fixed points from a rough estimate of their convergence
rate to the centroids.
June 13, 2017
DRAFT
11
A fixed point of hN provides us with an estimated centroid θbk = hN (θbk ) for some unknown centroid
θk . In order to model the estimation error, we can start by writing hN (θbk ) = hN (θk ) + Wk,1 . Of
course, Wk,1 will be all the more small than θbk approximates accurately θk . We can then write that
hN (θk ) = gk (θk ) + Wk,2 , where Wk,2 is likely to be small since the weighted averaging performed by
hN downweights data from clusters other than k . In absence of noise, we would directly have gk (θk ) = θk .
The presence of noise induces that gk (θk ) = θk +Wk,3 . At the end, we have θbk = θk +Wk,3 +Wk,2 +Wk,1 .
According to Proposition 1, Wk,1 and Wk,2 can be expected to remain small if centroids are far from
each other. Unfortunately, we cannot say more about these two terms and we do not know yet how to
model them. In contrast, it turns out that the term Wk,3 can be modeled as follows.
For any k ∈ J1, KK and any n ∈ J1, Nk K, set Xk,n = Yk,n − θk . With the same notation as above, it
follows from (8) that:
N1
X
Wk,3 = gk (θk ) − θk =
w(kXk,n k2 )Xk,n
n=1
Nk
X
w(kXk,n k2 )
n=1
By the central limit theorem and from Appendix E, we get
Nk
h
2 i
1 X
√
w(kXk,n k2 )Xk,n ⇒ N 0, E w kXk2
Id
Nk →∞
Nk n=1
where X ∼ N (0, Id ). On the other hand, the weak law of large numbers yields:
Nk
1 X
P
w(kXk,n k2 ) → E w kXk2
Nk
n=1
Slutsky’s theorem [32, Sec. 1.5.4, p. 19] then implies:
Nk
√ X
w(kXk,n k2 )Xk,n
Nk
n=1
Nk
X
⇒ N 0, r2 Id
w(kXk,n k2 )
Nk →∞
n=1
with
r2 =
h
2 i
E w kXk2
E [ w (kXk2 ) ]2
(16)
Therefore, Wk,3 is asymptotically Gaussian so that: gk (θk ) = θk + Wk,3 ∼ AN θk , (r2 /Nk ) Id Since
we do not know how to characterize Wk,1 and Wk,2 , whose contributions can however be expected to
be small, we opt for a model that does not take the influence of these terms into account. Therefore, we
model the statistical behavior of θbk by setting:
θbk ∼ N (θk , (r2 /Nk ) Id ).
June 13, 2017
(17)
DRAFT
12
This model discards the influence of Wk,1 and Wk,2 , but the experimental results reported in Section V
support the approach.
F. Marking operation
We now define the hypothesis test that permits to mark the measurement vectors given the estimated
centroids. Let θbk be an estimate of the unknown centroid θk . We assume that θbk follows the model of
Eq. (17). Consider a measurement vector Yn and let θ` be the centroid of the unknown cluster to which
Yn belongs, so that Yn ∼ N (θ` , Id ). The clustering problem can then be posed as the problem of testing
whether θk = θ` or not. According to our model for Y and θbk ,
Yn − θb` ∼ N (θ` − θk , (1 + (r2 /Nk ))Id )
Therefore, the problem of testing whether θk = θ` amounts to testing the mean of the Gaussian random
vector Yn − θb` . According to Subsection III-A, the optimal spherically invariant and UBCP test for
this problem is the Wald test T√(1+(r2 /N
k
))µ(γ)
. However, here, the value of Nk is unknown since the
objective of this step is to determine the measurement vectors that belong to the cluster. That is why we
assume that Nk is high enough and simply apply the Wald test Tµ(γ) in order to perform the marking.
G. Estimated centroid fusion
Despite the marking operation, it may occur that some centroids are estimated several times with
different initializations. In order to merge these centroids that are very close to each other, we now
introduce the fusion step that is applied when all the centroids have been estimated.
Consider two estimates θbk and θb` of two unknown centroids θk and θ` , respectively. The fusion
between θbk and θb` can then be posed as an hypothesis testing problem where the null hypothesis is
H0 : θk = θ` and the alternative is H1 : θk 6= θ` . In order to derive a solution to this binary hypothesis
testing problem, we resort to the probabilistic model (17) for the estimated centroids. In this respect, we
assume that θbk ∼ N (θk , (r2 /Nk ) Id ) and that θb` ∼ N (θ` , (r2 /N` ) Id ). In this model,
2
θbk − θb` ∼ N θk − θ` , σk,`
Id
with
s
σk,` = r
1
1
+
Nk
N`
(18)
The testing of H0 against H1 then amounts to testing the mean of the Gaussian random vector θbk − θb` .
According to Section III-A, this testing problem can optimally be solved by the Wald test Tσk,` µ(γ) .
Note that in (18), the values Nk and N` are assumed to be known. In practice, since after estimation of
June 13, 2017
DRAFT
13
θk (resp. θ` ), the measurement vectors close enough to θk (resp. θ` ) are marked, we consider that the
number of these marked vectors approximates sufficiently well Nk (resp. N` ).
Finally, it is worth emphasizing that the model considered above for centroid fusion is more restrictive
than a model that would involve the distribution of the whole sum Wk,1 + Wk,2 + Wk,3 , if this distribution
were known. By writing that, we mean that the Wald test Tσk,` µ(γ) , constructed for Wk,3 only, is likely
to yield more false alarms than a test exploiting the yet unknown whole distribution. In other words,
Tσk,` µ(γ) may not decide to merge estimated centroids that should be. However, the experimental results
of Section V support the idea that the Wald test Tσk,` µ(γ) is actually sufficient and efficient for the fusion.
IV. C LUSTERING A LGORITHMS
The objective of this section is to gather all the estimation functions and hypothesis tests defined in the
previous section in order to build two clustering algorithms. We first describe the centralized algorithm
(CENTREx), and then derive the decentralized version (DeCENTREx) of the algorithm.
A. Centralized clustering algorithm (CENTREx)
The objective of the algorithm is to divide the set of measurement vectors Y = {Y1 , . . . , YN } into
clusters. The centralized algorithm performs the following steps.
1) Initialization: Let us denote by M the set of vectors Yk that are considered as marked, where
marked vectors cannot be used anymore to initialize the estimation of a new centroid. The set M is
initialized as M = {∅}. Also, let Φ be the set of centroids estimated by the algorithm, where Φ is
initialized as Φ = {∅}. Fix a parameter that corresponds to a stopping criterion in the estimation of
the centroids.
2) Estimation of the centroids: The centroids are estimated one after the other, until M = Y . When
the algorithm has already estimated k centroids denoted θb1 , · · · , θbk , we have that Φ = {θb1 , · · · , θbk }. In
order to estimate the k + 1-th centroid, the algorithm picks a measurement vector Y? at random in the
(0)
set ∈ Y \ M and initializes the estimation process with θbk+1 = Y? . It then produces an estimate of the
(`+1)
(`)
centroid by computing recursively θbk+1 = hN (θbk+1 ), where the function hN was defined in (9) and
(`+1)
(`)
the recursion stops when kθbk+1 − θbk+1 k2 ≤ . Once the stopping condition is reached after, say, L
(L)
iterations, the newly estimated centroid is given by θbk+1 = θbk+1 , and the set of estimated centroids is
updated as Φ = Φ ∪ {θbk+1 }.
Once the centroid θbk+1 is estimated, the algorithm marks and stores in set Mk+1 all the vectors that
the Wald test Tµ(γ) defined in (2) accepts as elements of cluster k + 1. Therefore, Mk+1 = {Yi ∈ Y :
Tµ(γ) (Yi − θbk+1 ) = 0}. Note that a vector Yi may belong to several sets Mk , which is not an issue since a
June 13, 2017
DRAFT
14
set Mk of marked vectors is not the final set of vectors assigned to the cluster (see the classification step of
the algorithm). The algorithm finally updates the set of marked vectors as M ← M∪{Y? }∪Mk+1 . Note
that the measurement vector Y? that serves for initialization is also marked in order to avoid initializing
again with the same vectors. If M 6= Y , the algorithm estimates the next centroid θbk+2 . Otherwise, the
algorithm moves to the fusion step.
3) Fusion: Once M = Y and, say, K 0 centroids have been estimated, the algorithm applies the
hypothesis test Tσk,` µ(γ) defined in Section III-E to all pairs (θbk1 , θbk2 ) ∈ Φ × Φ such that k1 6= k2 .
We assume without loss of generality that the indices k1 and k2 are chosen such that k1 < k2 . When
Tσk,` µ(γ) (θbk1 − θbk2 ) = 0, the algorithm sets θbk1 =
θbk1 +θbk2
2
and removes θbk2 from Φ.
The fusion step ends when Tσk,` µ(γ) (θbk1 − θbk2 ) = 1 for all the (θbk1 , θbk2 ) ∈ Φ × Φ such that k1 6= k2 .
At this stage, the algorithm sets the number of centroids K as the cardinal of Φ and re-indexes the
elements of Φ in order to get Φ = {θb1 , · · · θbK }. It then moves to the final classification step.
4) Classification: Denote by Ck the set of measurement vectors assigned to cluster k . At the classification step, Ck is initialized as Ck = {∅}. Each vector Yi ∈ Y is then assigned to the cluster Ck0 whose
b . Note that Ck can be different
centroid θbk0 ∈ Φ is the closest to Yi , that is θbk0 = arg minθ∈Φ
kYi − θk
b
from Mk , due to the fusion step, but also to the closest centroid condition used during the classification.
In particular, each vector Yi can belong to only one single Ck .
In this version of the algorithm, we classify the measurement vectors by using the minimum distance
condition. This ends up to assigning each and every measurement vector to a cluster. However, it is worth
noticing that the algorithm could easily be modified by classifying the measurement vectors via the Wald
hypothesis test (as for the marking process). By so proceeding, a measurement vector would be assigned
to a centroid only if it is sufficiently close to this one. This would permit to detect outliers as vectors
that have not been assigned to any cluster.
B. Decentralized clustering algorithm (DeCENTREx)
In the decentralized algorithm, the operations required by the algorithm are performed by the sensors
themselves over the data transmitted by the other sensors. Each of the N sensors has access to one single
measurement vector Yn , only. We assume that the transmission link between two sensors is perfect, in
the sense that no error is introduced during information transmission. We now describe the decentralized
version of the algorithm, and point out the differences with the centralized algorithm.
1) Initialization of the algorithm: In the distributed algorithm, each sensor n ∈ {1, · · · , N } produces
its own set of centroids, denoted Φn and initialized as Φn = {∅}. Since each sensor only has its own
observation Yn , each sensor n now has its own marking variable Mn initialized as Mn = 0. Denote by
June 13, 2017
DRAFT
15
T the number of time slots available for the estimation of each centroid, and by 0 ≤ L ≤ N a stopping
condition for the centroid estimation.
2) Estimation of a centroid: As for the centralized version, the centroids are estimated one after the
other, until Mn = 1 for all n ∈ {1, · · · , N }. When sensor n has already estimated k centroids denoted
θbn,1 , · · · , θbn,k , we have that Φn = {θbn,1 , · · · , θbn,k }. All the sensors produce their k + 1-th estimates
θbn,k+1 at the same time as follows.
For the initialization of the centroid estimation, one sensor n0 is selected at random among the set
of sensors n for which Mn = 0. The vector Yn0 observed by this sensor is broadcasted to all the other
(0)
sensors. Each sensor n initializes its estimated centroid as θbn,k+1 = Yn0 , as well as two partial sums
(0)
(0)
Pn = w(kYn − θbn,k+1 k2 )Yn , Qn = w(kYn − θbn,k+1 k2 ). It also initializes a counter of the number of
partial sums cn = 0 received by sensor n.
At each time slot t = 0, · · · , T , sensor n receives J partial sums from J other sensors (J can vary
(t)
(t)
from time slot to time slot). We denote the partial sums received by sensor n as P1→n , · · · , PJ→n
(t)
(t)
and Q1→n , · · · , QJ→n . We assume that the sensor also receives the counters c1 , · · · , cJ of partial sums
calculated by the J other sensors. The sensor then updates its partial sums as
Pn(t) = Pn(t−1) +
J
X
(t)
Pj→n ,
(t−1)
Q(t)
+
n = Qn
j=1
and its counter as cn ← cn +
as θbn,k+1 =
Pn(t)
(t)
Qn
PJ
j=1 cj .
J
X
(t)
Qj→n
j=1
Afterwards, if cn ≥ L, the sensor updates its estimate θbn,k+1
(t)
(t)
and reinitializes its partial sums as Pn ← w(kYn − θbn,k+1 k2 )Yn , Qn ← w(kYn −
θbn,k+1 k2 ), and its counter as cn ← 0. As long as cn ≤ L, the sensor has not received enough partial
sums from the other sensors: it does not update its estimated centroid and waits for the next time slot. In
the above process, the centroids are estimated from
Pn(t)
(t) ,
Qn
which correspond to an approximation of the
function hN defined in (9).
The estimation process stops when time slot T is reached. At this time, sensor n updates its set of
centroids as Φn ← Φn ∪{θbn,k+1 }. If then verifies whether its observation Yn belongs to the newly created
cluster by applying the same hypothesis test as in CENTREx. In this case, if Tµ(γ) (Yi − θbk+1 ) = 0, then
Mn = 1. If Mn = 1 for all n ∈ {1, · · · , N }, the algorithm moves to the next step.
3) Fusion: The fusion step is almost the same as in the centralized algorithm, except that each sensor
performs its own fusion over the set Φn . Sensor n applies the hypothesis test Tσk,` µ(γ) (θbn,k1 − θbn,k2 )
defined in Section III-E to all the (θbn,k1 , θbn,k2 ) ∈ Φn ×Φn such that k1 6= k2 . It then fusions the centroids
θbn,k1 , θbn,k2 for which Tσk,` µ(γ) (θbn,k1 − θbn,k2 ) = 0 and denotes by Kn the final number of centroids in
Φn .
June 13, 2017
DRAFT
16
4) Classification: The classification step is exactly the same as in the centralized algorithm, except
that sensor n only classifies its own observation Yn . The sensor identifies the centroid θbn,k0 ∈ Φn that
b
is the closest to Yn , that is θbn,k0 = arg minθ∈Φ
b n kYn − θk.
At the end, the proposed decentralized algorithm induces more latency in the clustering compared to
K-means, since in K-means the centroids are estimated in parallel. However, our algorithm should reduce
the overall number of messages exchanged between sensors, since it does not need to be initialized by
the K-means++ procedure and since it does not need to estimate the number of clusters.
V. E XPERIMENTAL RESULTS
This section evaluates the performance of CENTREx and DeCENTREx through Monte Carlo simulations. In all our simulations, the observation vectors Yn that belong to cluster k are generated according
to the model Yn ∼ N (θk , σ 2 Id ), where σ 2 is the noise variance. Depending on the considered setup,
the centroids θk will be generated differently. In order to evaluate the performance of our algorithm, we
consider two figures of merit:
- The classification error probability Pe estimates the probability that a data point has been assigned to
the wrong cluster. It is determined from the confusion matrix which is a 2D matrix with true clusters
in line and estimated clusters in columns. In cell (i, j) of the matrix is indicated the percentage of
data from estimated cluster j that actually belong to estimated cluster i.
- The estimation distortion D is defined as the average distance between the data points and the
estimated centroids of the clusters to which they have been assigned.
In all our simulations, the parameter r2 (16) used for the fusion is evaluated numerically from Monte
Carlo simulations by averaging over 10000 realizations of X. Note that this parameter depends on the
dimension but depends on neither σ nor the considered data. It is thus computed once for all for every
set of simulations. The probability of false alarm γ is always set to 10−3 and the stopping criterion is
always set to 10−2 . However, these empirical parameters do not influence much the performance of the
algorithm as long as they belong to a reasonable range (roughly, from 10−6 to 10−2 ).
A. Centralized algorithm
We start by evaluating the performance of the centralized version of the algorithm. For comparison, we
evaluate the performance of the K-means algorithm, and we consider two methods in order to alleviate
the sensitivity of K-means to initialization. We first use the method of replicates, that involves running
the algorithm R times with random initializations and choosing the solution that minimizes the average
distance between the measurement vectors and the estimated centroids. In all our simulations, we consider
June 13, 2017
DRAFT
17
1e+1
1e+0
15
1e-1
1e+0
D (estimation)
Pe (classif)
1e-2
1e-3
1e-4
1e-5
K-means++
K-means10
Our algorithm
K-means100
1e-6
1e-7
1
1.5
2
sigma
2.5
3
10
1e-1
1e-2
1
(a)
K-means++
K-means10
Our algorithm
K-means100
1.5
2
sigma
2.5
5
3
0
2
(b)
4
6
8
10
12
14
(c)
Fig. 1: Performance evaluation of CENTREx, with d = 2, K = 4, N = 400: (a) Classification error probability (b) Estimation
distortion. In both Figures, Kmeans10 and Kmeans100 correspond to the K-means algorithm with 10 and 100 replicates,
respectively. (c) Example of clustering when σ > σlim . Triangles give the centroids estimated by K-means10 and squares give
centroids estimated by our algorithm. The circles correspond to the decision thresholds.
R = 10 and R = 100. Second, we consider the K-means++ algorithm without any replicates. As discussed
earlier in the paper, these two methods may not be appropriate for a distributed treatment. However, our
purpose here is to assess the performance of our centralized algorithm compared to existing algorithms,
before considering a distributed context.
In our experiments, we also assume that the number K of clusters is provided to the K-means algorithm.
This choice places K-means in very favorable conditions, but permits to avoid running the algorithm
several times with various values of K . Note that all the K-means algorithms we use in our experiments
are obtained from the Matlab function “kmeans”. We now evaluate the performance of our centralized
algorithm against K-means for various setups described in the following.
1) Parameters d = 2, K = 4, N = 400: In this part, the observations are generated as follows. The
vector dimension is given by d = 2, the number of clusters is K = 4, and the number of observations
is N = 400. The 4 centroids are given by θ1 = [A, 2A], θ2 = [2A, A], θ3 = [A, A], θ4 = [2A, 2A],
where A = 10. The number Nk of observations in cluster k does not depend on k and is given by
Nk = N/K = 100. We consider various values of σ , and evaluate the performance of the centralized
algorithms over N t = 10000 realizations for each considered value of σ .
The results are presented in Figure 1 (a) for the classification error probability and in Figure 1 (b)
for the estimation distortion. First, it is worth noticing than K-means++ without any replicates does
not perform well in this setup. In particular, for K-means++, low values of σ unexpectedly give higher
classification error probability and estimation distortion. This is probably due to the fact that when σ is
low, the clusters are so far from each other that a bad initialization cannot be handled by the algorithm.
June 13, 2017
DRAFT
18
As a result, K-means++ alone may not be sufficient by itself and may need replicates as well.
Further, according to the two figures, CENTREx exhibits the same performance as K-means algorithms
10 and 100 replicates, when σ ranges from 1.2 to 2.5. Below σ = 1.2, K-means with 10 replicates,
henceforth denoted by K-means10, performs worse than CENTREx, probably for the same initialization
issues as those incurred by K-means++. This issue does not appear anymore for K-means with 100
replicates, since the higher number of replicates increases the chances to find a good initialization.
Therefore, even when K is known, K-means requires a large number of replicates in order to increase
the probability of initializing correctly the algorithm.
On the other hand, for σ > 2.5, the K-means algorithm with 100 replicates, hereafter denoted by
K-means100, outperforms our algorithm. In fact, when σ becomes too large, two clusters can be so
entangled that discriminating between them is hardly feasible without prior knowledge of the number of
clusters. In this case, K-means may still be able to separate between two clusters since it already knows
that there are two clusters, while our algorithm may tend to merge the two centroids due to the value of
the variance.
It turns out that we can predict the value σlim above which two clusters may hardly be distinguished
by CENTREx. Basically, if the balls B(θk , σµ(γ)) and B(θ` , σµ(γ)) — with same radius σµ(γ) and
respective centers θk and θ` — actually intersect, there might be some ambiguity in classifying elements
of this intersection. Classifying data belonging to this intersection will be hardly feasible as soon as σ is
such that kθk − θ` k 6 σµ(γ). We conclude from the foregoing that our algorithm should perform well
for σ 6 σlim and may severely degrade for σ > σlim , where
σlim = min kθk − θ` k/µ(γ).
k,`
(19)
According to this rationale, it follows from the values chosen for A and γ that σlim = 2.7, which is close
to the value σ = 2.5 found by considering the experimental results of Figure 1.
In order to illustrate the behavior of our algorithm for σ > σlim , we consider the following setup. We
keep the parameters d = 2, K = 4, N = 400, but we modify the centroids as θ1 = [13, 20], θ2 = [20, 10],
θ3 = [10, 10], θ4 = [17, 20]. For these parameters, σlim = 1.1. We set σ = 2 > σlim and apply both
CENTREx and K-means10 to the set of data. The results are represented in Figure 1 (c). We see that
K-means retrieves four clusters, which is in accordance with the ground truth, while CENTRExfinds three
clusters only. However, by taking a look at the generated data, it does not seem such a bad choice in this
situation to consider three clusters instead of four, which cannot actually be assessed by the classification
error probability since this one performs a comparison to the true generated clusters.
June 13, 2017
DRAFT
19
1e+1
1e+0
1e+0
1e-2
D (est)
Pe (classif)
1e-1
1e-3
1e-1
1e-4
sigma=2.5
sigma=2
sigma=1
1e-5
1e-6
0
sigma=2.5
sigma=2
sigma=1
2
4
1/beta
6
(a)
8
10
1e-2
0
2
4
1/beta
6
8
10
(b)
Fig. 2: Performance of CENTREx with Gaussian kernel, with respect to parameter β and for various values of σ: (a) Classification
error probability (b) Estimation distortion. For 1/β > 3 and σ = 1, no value for the classification error probability is reported
in subfigure (a), because CENTREx with Gaussian kernel commited no classification error in this range.
2) Comparison with Gaussian kernel: Here, we consider the same parameters d = 2, K = 4, N = 400
and the same data generation as in the previous experiment. We want to compare the performance of
CENTREx when, for M-estimation of the centroids, the Wald p-value kernel is replaced by the Gaussian
kernel w(x) = exp − σβ2 kxk2 . This form for the Gaussian kernel is conveniently chosen here, instead
of w(x) = exp −βkxk2 as in [11], so as to hereafter consider parameter values that are directly
proportional to 1/σ 2 .
As explained in Section III-B, the main drawback of the Gaussian kernel resides in the choice of
the value of β . In order to select this value, we first evaluated the performance of CENTREx with the
Gaussian kernel for various values of β and σ . Figures 2 (a) and (b) represent the obtained classification
error probability and estimation distortion. We first see that the choice of the value of β does not depend
on whether we want to optimize the classification error criterion or the estimation distortion criterion.
For example, for σ = 2.5, the value β = 1/2 yields good performance with respect to both criteria. On
the other hand, the optimal value of β turns out to depend on the value of σ . For instance, for σ = 1,
we would select β as small as possible, whereas we should choose a large value of β for σ = 2.5. From
these observations, we can conclude that it is difficult to optimize the parameter β once for all, for the
whole range of possible values of σ .
We also evaluated the performance of CENTREx with Gaussian kernel for fixed values β = 1 (as
June 13, 2017
DRAFT
20
1e+1
1e+0
1e-1
1e+0
1e-3
D (esti)
Pe (classif)
1e-2
1e-4
1e-5
1e-6
1e-7
1
1e-1
Exp, Beta=1/5
p-value
Exp, Beta=1/3
Exp, Beta=1
1.5
2
sigma
2.5
3
1e-2
1
Exp, Beta=1/5
p-value
Exp, Beta=1/3
Exp, Beta=1
1.5
(a)
2
sigma
2.5
3
(b)
Fig. 3: Performance of CENTRExwith Gaussian kernel, with respect to σ and for various values of β: (a) Classification error
probability (b) Estimation distortion.
recommended in [11]), β = 1/3, β = 1/5. The results are presented in Figures 3 (a) and (b). The
parameter β = 1 shows the best performance for high values of σ , but the worst performance for low
values of σ , and the opposite observation can be made for β = 1/5. The value β = 1/3 seems to represent
the best tradeoff since it it close to the best possible performance for any value of σ . Compared to these
results, CENTREx with Wald p-value kernel yields good performance unless the value of σ becomes
too large. This result is obtained without having to optimize any parameter, as we have to do for the
Gaussian kernel. This feature of CENTREx with Wald p-value is useful since it induces no parameter
change when considering various dimensions, number of clusters, etc.
3) Parameters d = 100, K = 10, N = 100 : In this part, we consider a higher vector dimension
d = 100, as well as an increased number of clusters K = 10. The number of observations is set to
N = 100, which gives only 10 vectors per cluster. We still consider various values of σ , and evaluate
the centralized algorithms over N t = 10000 realizations for each considered value of σ . The ten new
centroids are generated once for all as θk ∼ N (0, A2 Id ) with A = 2.
As for d = 2, CENTREx is benchmarked against K-means++ without replicates, K-means10 and
K-means100. The results are presented in Figures 4 (a) and (b). We first observe that K-means++ and
K-means10 replicates perform very poorly with this set of parameters. The relatively high number of
clusters (K = 10) makes it more difficult to obtain a correct initialization with these two solutions,
which explains these poor results. Regarding our algorithm, it is worth mentioning that for σ < 1.6,
June 13, 2017
DRAFT
21
1e+0
1e+0
D (esti)
Pe (classif)
1e-1
1e-2
K-means++
K-means10
Our algorithm
K-means100
1e-3
1e-4
1
K-means++
K-means10
Our algorithm
K-means100
1.5
2
sigma
2.5
1e-1
3
1
1.2
(a)
1.4
sigma
1.6
1.8
2
(b)
Fig. 4: Performance evaluation of CENTREx, K-means10 and K-means100, with d = 100, K = 10, N = 100: (a) Classification
error probability (b) Estimation distortion. For σ < 1.6, no value for the classification error probability is reported in subfigure
(a) for CENTREx, since this one commited no classification error over the considered 1000 realizations.
the simulations did not return any error in terms of classification over the considered 1000 realizations.
As a result, CENTREx outperforms K-means100 for low values of σ . This is of noticeable importance
since the number of data per cluster is relatively small (N/K = 10), whereas the theoretical results of
Section III where proved under the conditions that N and the distances between centroids goes to infinity.
This shows that our algorithm still performs well in non-asymptotic conditions.
On the other hand, our algorithm severely degrades for high values of σ . The value σlim for which
our algorithm does not perform good anymore can be determined theoretically as in Section V-A1 as
σlim = 2. The region where σ > σlim corresponds again to cases where the clusters are too close to each
other for CENTREx to be capable of separating them, since it ignores the number of clusters.
In our experiments, K-means was evaluated with replicates and known K . In contrast, CENTREx uses
no replicates and was not provided with the value of K . It turns out that, despite these conditions favorable
to K-means, CENTREx does not incur a too significant performance loss in comparison to K-means and
even outperforms this one as long as the noise variance is not too big. In addition, CENTREx can
perform so without the need to be repeated several times for proper initialization. For all these reasons,
CENTREx appears as a good candidate for decentralized clustering.
B. Tests on images
Additional tests on images were also performed so as to study further the behavior of CENTREx in
very high dimension. We considered the nine images of Figure 5. Each of these images has size 256×256.
June 13, 2017
DRAFT
22
After computing numerically the minimum distance dmin between these images, application of (19) with
γ = 10−3 returns σlim = 57.
The next experiments are based on the following remark. The a priori known noise standard deviation
σ can easily be a posteriori estimated after clustering. It can then be expected that there should be little
difference between the true value of σ and its estimate after clustering by CENTREx, provided that σ
remains small enough, typically less than σlim . On the opposite, this difference should increase once the
noise standard deviation exceeds some value that must be around σlim .
To verify this claim, we performed clustering by CENTREx on a set of noisy images generated as
follows. For each σ ∈ {40, 41, . . . , 51, 51, . . . , 69}, we generated N = 10 noisy versions of each noiseless
image displayed in Figure 5. By so proceeding, we obtained a set of noisy images. Instances of these
noisy images are presented in Figures 6 and 7 for σ = 50 and σ = 70, respectively. After clustering by
CENTREx, we estimated σ for comparison to its true value. The value of σ was estimated as
N
σ
b2 =
1 X
kYn − θbk (Yn )k2
Nd
(20)
n=1
where θbk (Yn ) is the closest estimated centroid to Yn . For each value of σ , we reiterated 100 times the
above process so as to average the noise standard deviation estimates. The obtained average values are
displayed in Figure 9 with respect to σ .
In this figure, for reasons not yet identified, we observe that σ is underestimated (resp. overrestimated)
after clustering by CENTREx, when σ is below (resp. above) 59. Thus, the value σlim also characterizes
rather well the change in the behavior of the noise standard deviation estimation after clustering by
CENTREx.
Figure 9 also shows that CENTREx can itself assess the quality of its clustering and warns the user
that the clustering may be failing. Indeed, as soon as the noise standard deviation is overestimated
after clustering by CENTREx, it can be deemed that CENTREx is performing out its optimal operating
range. To experimentally verify this assertion, we proceeded as follows. For any given tested σ and
any clustering algorithm, we can always calculate the PSNR of each estimated centroid, with reference
to the closest image among the nine of Figure 5. By so proceeding, we can evaluate the quality of the
centroid estimation performed by the clustering. We recall that the PSNR of a given image I with respect
to a reference one Iref , both with size M × M , is defined by setting PSNR = 10 log10 d(I)2 /QEM
where d(I) is the maximum pixel value of I and QEM is the quadratic error mean given by QEM =
2
1 PM PM
i=1
j=1 (I(i, j) − Iref (i, j)) .
M2
Instances of such PSNRs are given in Table I. For each σ ∈ {50, 60, 70}, the PSNRs were calculated by
generating N = 10 noisy versions of each clean image of Figure 5, shuffling the resulting noisy images
June 13, 2017
DRAFT
23
to form the dataset presented to CENTREx, K-means10 and K-means100. No averaging were performed
to get these values so as to better emphasize the following facts. Clearly, CENTREx may fail to find
out the correct number of centroids when σ becomes too large and, more precisely, above σlim = 59.
However, the centroids estimated by CENTREx are well estimated, with a PSNR equal to that returned
by K-means100. In contrast, K-means10 sometimes fails to estimate correctly the centroids. Indeed,
for each σ in Table I, K-means10 returned a centroid mixing several images (see Figure 8). In short,
CENTREx may fail to retrieve all the existing centroids; nevertheless, those yielded by CENTREx are
correctly estimated, whereas K-means may require quite a lot of replicates to perform a correct estimation
of these same centroids.
Barbara
Cameraman
Einstein
House
Jetplane
Lake
Lena
Mandrill
Peppers
Fig. 5
C. Decentralized algorithm
We now evaluate the performance of the decentralized version DeCENTREx of our algorithm. In
this part, we consider again the parameters d = 2, K = 4, N = 400, and the data are generated
June 13, 2017
DRAFT
24
Barbara
Cameraman
Einstein
(σ = 50)
(σ = 50)
(σ = 50)
House
Jetplane
Lake
(σ = 50)
(σ = 50)
(σ = 50)
Lena
Mandrill
Peppers
(σ = 50)
(σ = 50)
(σ = 50)
Fig. 6
as in Section V-A1. We address the influence of the number of the time slots T allocated to the
estimation of each centroid. We proceed by comparing, when this parameter varies, the performance
of DeCENTREx to that of CENTREx. In all the considered cases, the parameter L, which represents
the number of received partial sums before updating the centroid is chosen as L = T /10. Of course, in
all the subsequent experiments, M-estimation of the centroids is performed by using the Wald p-value
kernel. The results are presented in Figures 10 (a) and (b). We see that it is possible to choose a value
of T such as the decentralized algorithm undergoes only a limited performance degradation compared to
the centralized algorithm. A small value T = 100 induces a rather important performance loss, whereas
higher values T = 300 and T = 500 provide almost the same performance as CENTREx, both in terms of
classification error probability and estimation distortion. These two figures hence permit to conclude that
DeCENTREx performs almost as well as CENTREx, which means that DeCENTREx is also competive
June 13, 2017
DRAFT
25
Barbara
Cameraman
Einstein
(σ = 70)
(σ = 70)
(σ = 70)
House
Jetplane
Lake
(σ = 70)
(σ = 70)
(σ = 70)
Lena
Mandrill
Peppers
(σ = 70)
(σ = 70)
(σ = 70)
Fig. 7
σ = 50
σ = 60
σ = 70
Fig. 8: Examples of centroids estimated by K-means10 and resulting from the mixing of several images. The PSNRs of such
centroids are respectively 3.39, 5.54 and 6.61 (see Table I)
June 13, 2017
DRAFT
26
Fig. 9: Estimate of the noise standard deviation calculated after clustering vs. true value of this noise standard deviation. For
σ ∈ {40, 41, . . . , 51, 51, . . . , 69}, N = 10 noisy versions of each clean image of Figure 5 are generated. The resulting images,
after shuffling, are presented to CENTREx for clustering and estimation of the noise standard deviation. Below σlim , the estimate
is always less than the actual value. Beyond σlim , the estimate becomes to increase significantly. This increase could be used by
the algorithm to assess itself the relevance of its clustering.
TABLE I: PSNRs of the centroids estimated by CENTREx, K-mean10 and K-means100, for various noise standard deviations.
A cross at a junction between a colum and a row indicates that the image to the left is not the closest one to any of the centroids
estimated by the algorithm specified by the column. We then say that this algorithm did not retrieve the image. Two values in the
same case indicate that the image naming the row was the closest one to two estimated centroids. For instance, when σ = 50,
K-means10 returned 2 estimated centroids, to which the closest clean image was ‘Lake’. K-means10 also yields 2 other estimates
for which the closest clean image was ‘Lena’. Consequently, two clean images, namely ‘Barbara’ and ‘Mandrill’, were never
found to be the closest ones to centroids estimated by K-means10. Regarding CENTREx, it may fail to retrieve all the centroids.
For instance, ‘Barbara’ was not retrieved by CENTREx among the estimated centroids. However, the PSNRs of the images
actually retrieved by CENTREx are close to those yielded by K-means100 for σ = 50. Finally, values in boldface correspond to
estimated centroids that appear as noisy mixtures of several other images (see Figure 8). These PSNRs are significantly smaller
than the other ones.
PSNR (σ = 50)
PSNR (σ = 60)
PSNR (σ = 70)
CENTREx
K-means10
K-means100
CENTREx
K-means10
K-means100
CENTREx
K-means10
Barbara
10
X
10.01
X
5.54
9.96
X
6.61
10
Cameraman
9.98
9.98
9.98
9.98
9.98
9.98
10
10
9.99
Einstein
10.04
3.39
10.04
X
X
10.01
X
X
10
House
10.01
10.02
10.02
10.01
5.99/7.77
10.01
10.04
6.97/7.01
10.03
Jetplane
10.02
10.02
10.02
9.96
9.96
9.96
10.04
10.04
10.04
Lake
9.98
8.47/4.76
9.98
9.98
9.98
9.98
10.00
10.00
10.00
Lena
10.03
6/7.78
10.03
9.98
9.98
9.98
X
10.00
10.00
Mandrill
10.01
X
10.01
10.03
10.03
10.03
X
9.96
9.96
Peppers
10.02
10.02
10.02
9.96
9.96
9.96
X
10.02
10.00
June 13, 2017
K-means100
DRAFT
27
1e+1
1e+0
1e+0
1e-2
D (esti)
Pe (classif)
1e-1
1e-3
1e-4
1e-5
1e-6
1
1e-1
Distributed, T=100
Distributed, T=200
Distributed, T=500
Centralized
1.5
2
sigma
2.5
3
1e-2
1
Distributed, T=100
Distributed, T=200
Distributed, T=500
Centralized
1.5
(a)
2
sigma
2.5
3
(b)
Fig. 10: Performance evaluation of DeCENTREx for various values of T , with d = 2, K = 4, N = 400: (a) Classification
error probability (b) Estimation distortion.
compared to the centralized versions of K-means, without suffering from the same drawbacks (no need
to estimate the number of clusters, no need for replicates).
VI. C ONCLUSION & PERSPECTIVES
In this paper, we have introduced a clustering algorithm for statistical signal processing applications.
This algorithm does not need to know the number of clusters and is less sensitive to initialization that the
K-means algorithm. These properties make our algorithm suitable for decentralized clustering and this is
why we also proposed a distributed version of the algorithm. It is worth emphasizing that all the steps of
our algorithm were theoretically derived and analyzed. Both the theoretical analysis and the simulations
assess the efficiency of the centralized and decentralized algorithms.
From a more general point of view, we have introduced a new methodology for clustering. This
methodology relies on a statistical model of the measurements. In this methodology, clustering is performed via M-estimation with score function derived from the p-value of a Wald test that is optimal
for the considered model. This methodology can be adapted to other signal models, which may allow
for adressing more general clustering problems that standard algorithms such as K-means can hardly
handle. For instance, we could consider heterogeneous sensors that would collect measurement vectors
with different variances. We might also introduce a more general notion of cluster, of which centroids
would be random with bounded variations less than a value τ > 0 in norm (this definition corresponds
to τ > 0 in the computations in Appendices).
June 13, 2017
DRAFT
28
A PPENDICES
Given τ ∈ [0, ∞), let λγ (τ ) be the unique real value such that Qd/2 (τ, λγ (τ )) = γ . In particular,
µ(γ) = λγ (0). The results stated in Appendices A, B and C below involve λγ (τ ) and w = Qd/2 (τ, •)
instead of merely µ(γ) and w = Qd/2 (0, •), respectively. The application of these results in the main
core of the paper thus concerns the particular case τ = 0. The reason why we present these results for
any τ ∈ [0, ∞) is twofold. First, the proof of these results in the particular case where τ = 0 would not
be significantly simpler than that for the general case. Second, having results that hold for any τ ∈ [0, ∞)
opens prospects described in the concluding section of the paper.
A PPENDIX A
P - VALUE OF
WALD TEST
For reasons evoked just above, we consider a more general case than that actually needed in the paper.
As a preliminary result, we need the following lemma.
Lemma 1. Given τ ∈ [ 0 , ∞ ), the map γ ∈ ( 0 , 1 ] 7→ λγ (τ ) ∈ [ 0 , ∞ ) is strictly decreasing.
Proof. Let ρ be some element of [0, ∞) and consider two elements γ and γ 0 of (0, 1]. We have
Qd/2 (ρ, λγ (ρ)) = γ and Qd/2 (ρ, λγ 0 (ρ)) = γ 0 . If γ < γ 0 , we thus have Qd/2 (ρ, λγ (ρ)) < Qd/2 (ρ, λγ 0 (ρ)),
which implies that λγ (ρ) > λγ 0 (ρ) since Qd/2 (ρ, •) is strictly decreasing.
Let τ ∈ [0, ∞). Given y ∈ Rd , set γ 0 = Qd/2 (τ /σ0 , kyk/σ0 ). Since γ 0 ∈ (0, 1), we have
Qd/2 (τ /σ0 , λγ 0 (τ /σ0 )) = γ 0
by definition of λγ 0 (τ /σ0 ). It then follows from the bijectivity of Qd/2 (τ /σ0 , •) and the foregoing
equalities that λγ 0 (τ /σ0 ) = kyk/σ0 . According to Lemma 1, we have:
γ ∈ (0, 1) : λγ (τ /σ0 ) < kyk/σ0
= γ ∈ (0, 1) : λγ (τ /σ0 ) < λγ 0 (τ /σ0 )
= (γ 0 , 1)
Therefore, γ 0 = inf γ ∈ (0, 1) : Tσ0 λγ (τ /σ0 ) (y) = 1 , where Tσ0 λγ (τ /σ0 ) is defined according to (2) for
all y ∈ Rd by setting:
0
Tσ0 λγ (τ /σ0 ) (y) =
1
if
kyk 6 σ0 λγ (τ /σ0 )
if
kyk > σ0 λγ (τ /σ0 ).
According to [28], Tσ0 λγ (τ /σ0 ) satisfies several optimality criteria for testing whether kθk 6 τ or not
when we observe Y ∼ N (θ, Id ). Therefore,
def
γ
bσ0 (y) = Qd/2 (τ /σ0 , kyk/σ0 )
June 13, 2017
(21)
DRAFT
29
can be regarded as the p-value of Tσ0 λγ (τ /σ0 ) for testing the hypothesis kθk 6 τ . The Wald test of
Section III-C for testing θ = 0 or not, when the observation is Y ∼ N (θ, Id ), then corresponds to the
particular case τ = 0, for which we have λγ (τ ) = σ0 µ(γ) and γ
bσ0 (y) = Qd/2 (0, kyk/σ0 ).
A PPENDIX B
ROBUSTNESS OF THE M- ESTIMATION FUNCTION
In CENTREx and DeCENTREx, centroids are calculated by M-estimation since, given a cluster, data
from other clusters can be regarded as outliers. Our claim is then that the weight function w, specified
by (10), is particularly suitable because it is the p-value associated with an optimal test, namely the
Wald test, aimed at deciding whether two observations lie within nearby clusters or not. As such, w can
be expected to be discriminating enough between data from different clusters. In this section, we thus
analyze to what extent the M-estimator based on this weight function is actually robust to outliers. This
analysis can be carried out by studying the influence function of the estimator.
As announced at the beginning of this appendices, we carry out the computation in the more general
case where the weight function is w = Qd/2 (τ, •), which requires handling λγ (τ ) instead of merely µ(γ).
The presence of τ does not complexify the analysis and opens more general prospects.
Following [30, Definition 5, p. 230], the general case of an M -estimator of some parameter θ ∈ Θ ⊂ Rd
given cumulative distribution function (cdf) F, is the solution T (F) in t to the equation:
Z
Ψ(y, t)dF(y) = 0
(22)
iid
with Ψ : Rd × Θ → Rd . In particular, given d-dimensional vectors Y1 , . . . , YN ∼ G where G is a given
P
cdf, we can consider the empirical probability distribution G∗N = N1 N
n=1 δYn , which puts mass 1/N at
each Y1 , . . . , YN . The solution T (G∗n ) in t to Eq. (22) with F = G∗n is the standard M -estimator for the
sample Y1 , · · · , YN of distribution G.
Now, choose Ψ(y, t) = Ψ(y − t), where Ψ is given by Eq. (11). Set t = (t1 , . . . , td )T , y =
(y1 , . . . , yd )T and Ψ = (Ψ1 , . . . , Ψd ). For the general M -estimator T (F) obtained by solving Eq. (22),
the differentiability of Ψ induces that we can define the d × d matrix B = (Bj,k )16j,k6d with:
#
Z "
∂Ψj (Y , t)
Bj,k = −
dF(y)
∂tk
t=T (F)
(23)
Since we have:
(yj − tj )(yk − tk ) 0
−
Qd/2 (τ, kY − tk),
if j =
6 k
kY − tk
∂Ψj (Y , t)
=
∂tk
(yk − tk )2 0
Q (τ, kY − tk), if j = k
− Qd/2 (τ, kY − tk) −
kY − tk d/2
June 13, 2017
DRAFT
30
the matrix B can be expressed as:
Z
Z
B = Qd/2 (τ, kY − T (F)k)dF(y)Id + Q0d/2 (τ, kY − T (F)k)(y − T (F))(y − T (F))T dF(y)
The first integral in the right hand side (rhs) of the second equality above is positive. Therefore, the
eigenvalues of the symmetric matrix B are all positive and B is invertible. According to [30, Eq. (4.2.9),
p. 230], the invertibility of B makes it possible to calculate the influence function as:
IF(Y ; T, F ) = B −1 Ψ(Y , T (F))
= B −1 (Y − θ)Qd/2 (τ, kY − θk)
Since the Marcum function is continuous, the influence function IF(Y ; T, F ) is also continuous. The
norm of the influence function can now be bounded by:
kIF(Y ; T, F )k ≤ |||B −1 ||| × kY − θk × Qd/2 (τ, kY − θk)
where ||| · ||| refers to the norm of a matrix. Let us consider the function f : t ∈ [0, +∞[→ tQd/2 (τ, t).
This function is continuous. If τ 6= 0, then Qd/2 (τ, t) ∼ (t/τ )(d−1)/2 Q(t − τ ) from [26, p. 1167, Eq.
(4)], where Q is the Gaussian Q – function. It follows from [33, Corollary 1] that Q(x) 6 21 e−x
2
/2
and
thus, that limt→∞ f (t) = 0. We have the same result if τ = 0. Indeed, if τ = 0, it follows from [26, p.
1168, Eq. (11)] and a straightforward change of variable that:
Qd/2 (0, x) =
1
2d/2 Γ(d/2)
Z
∞
2
td/2−1 e−t/2 dt 6 e−x
/4 d/2
2
x2
Since f (0) = 0 and f is continuous, we derive from the foregoing that f is upper-bounded and so is the
norm of the influence function. Since the influence function is continuous and bounded, the estimator
gross error sensitivity γ ? = supY kIF(Y ; T, F )k is finite. This shows that our estimator is robust to
outliers.
A PPENDIX C
A SYMPTOTIC BEHAVIORS
Lemma 2. If Z(ξ) ∼ N (ξ, Id ) with ξ ∈ Rd , then:
(i) E w(kZ(0)k2 )Z(0) = 0
(ii)
(iii)
lim E w(kZ(ξ)k2 ) = 0
kξk→∞
lim E w(kZ(ξ)k2 )Z(ξ) = 0
kξk→∞
Proof: As a preliminary result, we recall that:
lim Qd/2 (τ, t) = 0,
t→∞
June 13, 2017
(24)
DRAFT
31
which derives from the fact that Qd/2 (τ, t) = 1 − Fχ2 (τ 2 ) (t2 ) [26, Eq. (8)], where Fχ2 (τ 2 ) is the noncentered χ2 distribution with d degrees of freedom and non-centrality parameter τ 2 .
Proof of statement (i): The vector E w(kZ(0)k2 )Z(0) is d-dimensional. Its first component is:
Z
d
X
2
E w(kZ(0)k )Z1 (0) =
w
zk2 z1 ϕ(z)dz
Rd
k=1
where Z1 (0) is the first component of Z(0), z = (z1 , z2 , . . . , zd )T and ϕ is the probability density function
P
d
2
(pdf) of the standard distribution N (0, Id ). Since w
k=1 zk is bounded by 1 and E [ |Z1 (0)| ] is finite,
Fubini’s theorem applies and we have:
E w(kZ(0)k2 )Z1 (0) =
!
! !
Z
Z Z X
d
1
2
2
2
...
w
zi2 z1 e−z1 /2 dz1 e−z2 /2 dz2 . . . e−zd /2 dzd .
N/2
(2π)
i=1
Z
d
X
2
zi2 z1 e−z1 /2 dz1 = 0 because the integrand is odd, it follows that E w(kZ(0)k2 )Z1 (0) =
i=1
0. The same type of computation holds for any component of E w(kZ(0)k2 )Z(0) . Thence the result.
Since
w
Proof of statement (ii): By definition of w and Z(ξ), we have:
w(kZ(ξ)k2 ) = Qd/2 (τ, kZ(ξ)k) = Qd/2 (τ, kξ + Xk)
(25)
Since kξ + Xk > kξk − kXk and Qd/2 (τ, •) is decreasing, we derive from (24) and (25) that:
lim w(kZ(ξ)k2 ) = 0
kξk→∞
(a-s)
(26)
The result then derives from Lebesgue’s dominated convergence theorem.
Proof of statement (iii): We begin by writing that
E w(kZ(ξ)k2 )Z(ξ) = E w(kZ(ξ)k2 ) ξ + E w(kZ(ξ)k2 )X .
(27)
Set ξ = (ξ1 , ξ2 , . . . , ξd )T . The first component of the first term to the rhs of the equality above can be
rewritten as:
1
E w(kZ(ξ)k ) ξ1 =
(2π)d/2
2
Z
1
2
w kyk2 ξ1 e− 2 ky−ξk dy
(28)
Rd
The inequality −kξk 6 ξ1 6 kξk induces that:
1
2
1
2
1
2
−kξke− 2 ky−ξk 6 ξ1 e− 2 ky−ξk 6 kξke− 2 ky−ξk
For any given y ∈ Rd , the left and right bounds in the inequality above tend to 0 when kξk tends to ∞.
Lebesgue’s dominated convergence theorem applied to (28) yields that
lim E w(kZ(ξ)k2 ) ξ1 = 0
kξk→∞
June 13, 2017
DRAFT
32
The same reasoning holds for any component of ξ . Therefore,
lim E w(kZ(ξ)k2 ) ξ = 0
(29)
kξk→∞
As far as the second term to the rhs of (27) is concerned, we have w(kZ(ξ)k2 )X 6 kXk. Since
E [ kXk ] < ∞, we derive from Lebesgue’s dominated convergence theorem and (26) that:
lim E w(kZ(ξ)k2 )X = 0
(30)
kξk→∞
Thence the result as a consequence of (27), (29) and (30).
A PPENDIX D
Lemma 3. Let Z be a d-dimensional Gaussian vector with covariance matrix Id . If f : Rd → [0, ∞) is
non-null, continuous and even in each coordinate of x = (x1 , . . . , xd )T ∈ Rd so that
f (x1 , . . . , xi−1 , xi , xi+1 , . . . , xd ) = f (x1 , . . . , xi−1 , −xi , xi+1 , . . . , xd ),
then:
E [ f (Z)Z ] = 0 if and only if E [ Z ] = 0.
Proof: If d = 1, Z is a random variable Z ∼ N (ξ, 1) and f is a nonnegative real function f : R →
R∞
1
2
[0, ∞). We have E [ f (Z)Z =
] √12π −∞ f (z)ze− 2 (z−ξ) dz . By splitting this integral in two, symmetrically
with respect to the origin, and after the change of variable t = −z in the integral from −∞ to 0 resulting
R∞
1
2
2
from the splitting, some routine algebra leads to E [ f (Z)Z ] = √12π 0 f (−t)te− 2 (t +ξ ) eξt − e−ξt dt.
The integrand in this integral is non-negative and continuous. Therefore, E [ f (Z)Z ] = 0 implies that
1
2
2
f (−t)te− 2 (t +ξ ) eξt − e−ξt = 0 for any t, which induces ξ = 0. The converse is straightforward.
In the d-dimensional case, set Z = (Z1 , Z2 , . . . , Zd ) and denote the expectation E [ Z ] of Z by
ξ = (ξ1 , ξ2 , . . . , ξd ). Let f : Rd → [0, ∞) be a nonnull and continuous function. Clearly, E [ f (Z)Z ] = 0
if and only if E [ f (Z)Zi ] = 0 for each coordinate Zi , i ∈ {1, 2, . . . , d}. For the first coordinate Z1 of
Z , it follows from Fubini’s theorem that:
Z
E [ f (Z)Z1 ] =
1
2
f (z1 ) z1 e− 2 (z1 −ξ1 ) dz1
with
1
f (z1 ) =
(2π)d/2
Z
1
f (z1 , . . . , zd )e− 2
Pd
k=2
(zk −ξk )2
dz2 . . . dzd .
The function f is defined on R, continuous and non-negative too. The result then follows from the
monodimensional case treated above.
June 13, 2017
DRAFT
33
h
P ROOF OF E w kXk
A PPENDIX E
i
h
2 i
XX T = E w kXk2
Id
2 2
WHEN
X ∼ N (0, Id )
Set X = (X1 , X2 , . . . , Xd )T and compute the term ci,j located at the ith line and j th colum of the
h
i
2
matrix E w kXk2 XX T with i 6= j . We have:
h
i Z
2
2
ci,j = E w kXk2 Xi Xj = w kxk2 xi xj ϕ(x)dx
where ϕ is the pdf of X ∼ N (0, Id ). By independence of the components of X and Fubini’s theorem,
the foregoing implies:
ci,j
Z
Since
1
=
(2π)d/2
w kXk2
2
Z
d
Y
−x2k /2
e
Z Z
2 2
w kxk
dxk
xi ϕ(xi )dxi
xj ϕ(xj )dxj
k=1,k6=i,j
xi ϕ(xi )dxi = 0 because the integrand in this integral is odd, we conclude that
ci,j = 0 for i 6= j .
We now compute ci,i for i = 1, . . . , d. Similarly to above, Fubini’s theorem implies that:
ci,i
h
i
2
= E w kXk2 Xi2
Y
Z Z
d
x2
x2
1
2 2 2 − 2i
− 2k
=
w
kxk
x
e
dx
e
dxk
i
i
(2π)d/2
k=1,k6=i
Z h
d
d
X
x2
i Y
1
2
2 2
− 2k
=
E
w
X
+
x
e
dxk
i
j
(2π)d/2
j=1,j6=i
k=1,k6=i
h
2 i
= E w kXk2
the last equality being obtained by iterating the integration over all the components of X .
R EFERENCES
[1] J. Yick, B. Mukherjee, and D. Ghosal, “Wireless sensor network survey,” Computer networks, vol. 52, no. 12, pp. 2292–
2330, 2008.
[2] O. Omeni, A. C. W. Wong, A. J. Burdett, and C. Toumazou, “Energy efficient medium access protocol for wireless medical
body area sensor networks,” IEEE Transactions on biomedical circuits and systems, vol. 2, no. 4, pp. 251–259, 2008.
[3] F. J. Ordóñez, P. de Toledo, and A. Sanchis, “Activity recognition using hybrid generative/discriminative models on home
environments using binary sensors,” Sensors, vol. 13, no. 5, pp. 5460–5477, 2013.
[4] K. Sahasranand and V. Sharma, “Distributed nonparametric sequential spectrum sensing under electromagnetic interference,”
in IEEE International Conference on Communications (ICC).
IEEE, 2015, pp. 7521–7527.
[5] G. Tuna, V. C. Gungor, and K. Gulez, “An autonomous wireless sensor network deployment system using mobile robots
for human existence detection in case of disasters,” Ad Hoc Networks, vol. 13, pp. 54–68, 2014.
[6] Q. Zhou, F. Ye, X. Wang, and Y. Yang, “Automatic construction of garage maps for future vehicle navigation service,” in
IEEE International Conference on Communications (ICC).
IEEE, 2016, pp. 1–7.
[7] G. Wang, Y. Zhao, J. Huang, Q. Duan, and J. Li, “A k-means-based network partition algorithm for controller placement
in software defined network,” in IEEE International Conference on Communications (ICC).
June 13, 2017
IEEE, 2016, pp. 1–6.
DRAFT
34
[8] A. K. Jain, “Data clustering: 50 years beyond K-means,” Pattern recognition letters, vol. 31, no. 8, pp. 651–666, 2010.
[9] D. Arthur and S. Vassilvitskii, “k-means++: The advantages of careful seeding,” in Proceedings of the eighteenth annual
ACM-SIAM symposium on Discrete algorithms.
Society for Industrial and Applied Mathematics, 2007, pp. 1027–1035.
[10] D. Pelleg, A. W. Moore et al., “X-means: Extending k-means with efficient estimation of the number of clusters.” in ICML,
vol. 1, 2000, pp. 727–734.
[11] K.-L. Wu and M.-S. Yang, “Alternative c-means clustering algorithms,” Pattern recognition, vol. 35, no. 10, pp. 2267–2278,
2002.
[12] S. Datta, C. Giannella, and H. Kargupta, “Approximate distributed k-means clustering over a peer-to-peer network,” IEEE
Transactions on Knowledge and Data Engineering, vol. 21, no. 10, pp. 1372–1388, 2009.
[13] G. Di Fatta, F. Blasa, S. Cafiero, and G. Fortino, “Epidemic k-means clustering,” in IEEE 11th international conference
on data mining workshops.
IEEE, 2011, pp. 151–158.
[14] J. Fellus, D. Picard, and P.-H. Gosselin, “Decentralized k-means using randomized gossip protocols for clustering large
datasets,” in IEEE 13th International Conference on Data Mining Workshops.
IEEE, 2013, pp. 599–606.
[15] M. Ester, H.-P. Kriegel, J. Sander, X. Xu et al., “A density-based algorithm for discovering clusters in large spatial databases
with noise.” in KDD, 1996.
[16] M. Ankerst, M. M. Breunig, H.-P. Kriegel, and J. Sander, “Optics: ordering points to identify the clustering structure,” in
ACM Sigmod record, vol. 28, no. 2.
ACM, 1999, pp. 49–60.
[17] M. Steinbach, G. Karypis, V. Kumar et al., “A comparison of document clustering techniques,” in KDD workshop on text
mining, vol. 400, no. 1.
Boston, 2000, pp. 525–526.
[18] Z. Huang, “Extensions to the k-means algorithm for clustering large data sets with categorical values,” Data mining and
knowledge discovery, vol. 2, no. 3, pp. 283–304, 1998.
[19] C. Bouveyron and C. Brunet-Saumard, “Model-based clustering of high-dimensional data: A review,” Computational
Statistics & Data Analysis, vol. 71, pp. 52–78, 2014.
[20] P. A. Forero, A. Cano, and G. B. Giannakis, “Distributed clustering using wireless sensor networks,” IEEE Journal of
Selected Topics in Signal Processing, vol. 5, no. 4, pp. 707–724, 2011.
[21] P. Huber and E. Ronchetti, Robust Statistics, second edition.
John Wiley and Sons, 2009.
[22] P. Rousseeuw and C. Croux, “Alternatives to the median absolute deviation,” Journal of the American Statistical Association,
vol. 88, no. 424, pp. 1273 – 1283, December 1993.
[23] D. Pastor and F. Socheleau, “Robust estimation of noise standard deviation in presence of signals with unknown distributions
and occurrences,” IEEE Transactions on Signal Processing, vol. 60, no. 4, 2012.
[24] A. M. Zoubir, V. Koivunen, Y. Chakhchoukh, and M. Muma, “Robust estimation in signal processing: A tutorial-style
treatment of fundamental concepts,” IEEE Signal Processing Magazine, vol. 29, no. 4, pp. 61–80, 2012.
[25] A. Wald, “Tests of statistical hypotheses concerning several parameters when the number of observations is large,”
Transactions of the American Mathematical Society, vol. 54, no. 3, pp. 426 – 482, Nov. 1943.
[26] Y. Sun, A. Baricz, and S. Zhou, “On the Monotonicity, Log-Concavity, and Tight Bounds of the Generalized Marcum and
Nuttall Q-Functions,” IEEE Transactions on Information Theory, vol. 56, no. 3, pp. 1166 – 1186, Mar. 2010.
[27] E. L. Lehmann and J. P. Romano, Testing Statistical Hypotheses, 3rd edition.
Springer, 2005.
[28] D. Pastor and Q.-T. Nguyen, “Random Distortion Testing and Optimality of Thresholding Tests,” IEEE Transactions on
Signal Processing, vol. 61, no. 16, pp. 4161 – 4171, Aug. 2013.
[29] F. Hampel, “The influence curve and its role in robust estimation,” Journal of the American Statistical Association, vol. 69,
no. 346, pp. 383 – 393, June 1974.
June 13, 2017
DRAFT
35
[30] F. Hampel, E. Ronchetti, P. Rousseeuw, and W. Stahel, Robust Statistics: the Approach based on Influence Functions.
John Wiley and Sons, New York, 1986.
[31] A. van der Vaart, Asymptotic statistics.
Cambridge Unversity Press, 1998.
[32] R. J. Serfling, Approximations theorems of mathematical statistics.
Wiley, 1980.
[33] S.-H. Chang, P. C. Cosman, and L. B. Milstein, “Chernoff-type bounds for the gaussian error function,” IEEE Transactions
on communications, vol. 59, no. 11, Nov. 2011.
June 13, 2017
DRAFT
| 10math.ST
|
RANDOM SAMPLING IN COMPUTATIONAL ALGEBRA:
HELLY NUMBERS AND VIOLATOR SPACES
arXiv:1503.08804v3 [cs.DM] 23 Dec 2015
JESÚS A. DE LOERA, SONJA PETROVIĆ, AND DESPINA STASI
Abstract. This paper transfers a randomized algorithm, originally used in geometric optimization,
to computational problems in commutative algebra. We show that Clarkson’s sampling algorithm
can be applied to two problems in computational algebra: solving large-scale polynomial systems and
finding small generating sets of graded ideals. The cornerstone of our work is showing that the theory
of violator spaces of Gärtner et al. applies to polynomial ideal problems. To show this, one utilizes
a Helly-type result for algebraic varieties. The resulting algorithms have expected runtime linear in
the number of input polynomials, making the ideas interesting for handling systems with very large
numbers of polynomials, but whose rank in the vector space of polynomials is small (e.g., when the
number of variables and degree is constant).
Keywords: Violator spaces, ideal generators, solving polynomial systems, randomized algorithm
in algebra, large sparse systems.
AMS subject classification: 68W20, 68R05, 12Y05, 13P10, 14Q10, 08A40.
1. Introduction
Many computer algebra systems offer excellent algorithms for manipulation of polynomials. But
despite great success in the field, many algebraic problems have bad worst-case complexity. For
example, Buchberger’s [13, 14, 18] groundbreaking algorithm, key to symbolic computational algebra today, computes a Gröbner basis of any ideal, but it has a worst-case runtime that is doubly
exponential in the number of variables [22]. This presents the following problem: what should one do
about computations whose input is a very large, overdetermined system of polynomials? In this paper,
we propose to use randomized sampling algorithms to ease the computational cost in such cases.
One can argue that much of the success in computation with polynomials (of non-trivial size) often
relies heavily on finding specialized structures. Examples include Faugère’s et al. fast computation
of Gröbner bases of zero-dimensional ideals [25, 26, 27, 29, 35], specialized software for computing
generating sets of toric ideals [1], several packages in [32] built specifically to handle monomial
ideals, and the study of sparse systems of polynomials (i.e., systems with fixed support sets of
monomials) and the associated homotopy methods [45]. A more recent example of the need to find
good structures is in [16], where Cifuentes and Parrilo began exploiting chordal graph structure in
computational commutative algebra, and in particular, for solving polynomial systems. Our paper
exploits combinatorial structure implicit in the input polynomials, but this time akin to Helly-type
results from convex discrete geometry [36].
At the same time, significant improvements in efficiency have been obtained by algorithms that
involve randomization, rather than deterministic ones (e.g. [10, 43]); it is also widely recognized
that there exist hard problems for which pathological examples requiring exponential runtimes occur
only rarely, implying an obvious advantage of considering average behavior analysis of many algorithms. For example, some forms of the simplex method for solving linear programming problems
have worst-case complexity that is exponential, yet [44] has recently shown that in the smoothed
analysis of algorithms sense, the simplex method is a rather robust and fast algorithm. Smoothed
analysis combines the worst-case and average-case algorithmic analyses by measuring the expected
1
2
JESÚS A. DE LOERA, SONJA PETROVIĆ, AND DESPINA STASI
performance of algorithms under slight random perturbations of worst-case inputs. Of course, probabilistic analysis, and smoothed analysis in particular, has been used in computational algebraic
geometry for some time now, see e.g., the elegant work in [8, 9, 15]. The aim of this paper is to
import a randomized sampling framework from geometric optimization to applied computational
algebra, and demonstrate its usefulness on two problems.
Our contributions. We apply the theory of violator spaces [31] to polynomial ideals and adapt
Clarkson’s sampling algorithms [17] to provide efficient randomized algorithms for the following
concrete problems:
(1) solving large (overdetermined) systems of multivariate polynomials equations,
(2) finding small, possibly minimal, generating sets of homogeneous ideals.
Our method is based on using the notion of a violator space. Violator spaces were introduced
in 2008 by Gärtner, Matoušek, Rüst, and Škovroň [31] in a different context. Our approach allows
us to adapt Clarkson’s sampling techniques [17] for computation with polynomials. Clarkson-style
algorithms rely on computing with small-size subsystems, embedded in an iterative biased sampling
scheme. In the end, the local information is used to make a global decision about the entire system.
The expected runtime is linear in the number of input elements, which is the number of polynomials
in our case (see [12] for a more recent simplified version of Clarkson’s algorithm for violator spaces).
Violator spaces naturally appear in problems that have a natural linearization and a sampling size
given by a combinatorial Helly number of the problem. While violator spaces and Clarkson’s algorithm have already a huge range of applications, to our knowledge, this is the first time such sampling
algorithms are being used in computational algebraic geometry. For an intuitive reformulation of
Helly’s theorem for algebraic geometers, see Example 2.1. Main ingredients of violator spaces are
illustrated through Examples 2.2, 3.2 and 3.6. A typical setup where problem (1) can be difficult
and a randomized algorithm appropriate can be found in Example 4.8.
Before stating the main results, let us fix the notation used throughout the paper. We assume the
reader is acquainted with the basics of computational algebraic geometry as in the award-winning
undergraduate textbook [18]. Denote by K a (algebraically closed) field; the reader may keep K = C
in mind as a running example. Let f1 = 0, . . . , fm = 0 be a system of m polynomials in n+1 variables
with coefficients in K. We usually assume that m ≫ n. As is customary in the algebra literature,
we write f1 , . . . , fm ∈ R = K[x0 . . . , xn ] and often denote the polynomial ring R by a shorthand
notation K[x]. We will denote by (f1 , . . . , fm ) ⊂ R the ideal generated by these polynomials; that
is, the set of all polynomial combinations of the fi ’s. Note that if F = {f1 , . . . , fm } is a set of
polynomials, the ideal (f1 , . . . , fm ) will equivalently be denoted by (F ).
A polynomial is said to be homogeneous if all of its terms are of same degree; an ideal generated
by such polynomials is a homogeneous ideal. In this paper, the ideals we consider need not be
homogeneous; if they are, that will be explicitly stated. In that case, the set of all homogeneous
polynomials of total degree d will be denoted by [R]d . Finally, denote by V(S) the (affine) variety
defined by the set of polynomials S ∈ R, that is, the Zariski closure of the set of common zeros of
the polynomials in the system S. Therefore, the concrete problem (1) stated above simply asks for
the explicit description of the variety (solution set) corresponding to an ideal (system of polynomial
equations). The concrete problem (2) asks to find a smaller (e.g., minimal with respect to inclusion)
set of polynomial equations that generate the same ideal - and thus have the exact same solution
set.
Solving large polynomial systems. Suppose we would like to solve a system of m polynomials
in n + 1 variables over the field K, and suppose that m is large. We are interested in the coefficients
of the polynomials as a way to linearize the system. To that end, recall first that the d-th Veronese
HELLY NUMBERS AND VIOLATOR SPACES IN COMPUTATIONAL ALGEBRA
n+d
d
embedding of Pn is the following map νd : Pn → P(
3
)−1 :
d
(x0 : · · · : xn ) 7→ (xd0 : xd−1
0 x1 : · · · : xn ).
The map νd induces a coefficient-gathering map for homogeneous polynomials in fixed degree d:
n+d
coeff d : [R]d → K ( d )
X
α
cα x 7→ cα1 , . . . , cα n+d ,
( d )
α:|α|=d
xαi
where
corresponds
to the i-th coordinate of the d-th Veronese embedding. We follow the usual
P
notation |α| = i αi . Therefore, if f is a homogeneous polynomial of deg(f ) = d, coeff(f ) is a vector
n+d
in the K-vector space K ( d ) . This construction can be extended to non-homogeneous polynomials
in the following natural way. Consider all distinct total degrees d1 , . . . , ds of monomials that appear
in a non-homogenous polynomial f . For each di , compute the image under coeff di of all monomials
of f of degree di . Finally, concatenate all these vectors
into the total coefficient vector of f , which
we will call coeff(f ) and which is of size n+d+1
,
the
number
of monomials in n + 1 variables of
n+1
(total) degree ranging from 0 to d. In this way, a system f1 , . . . , fm of polynomials
in n variables
of degree at most d can be represented by its coefficient matrix of size n+d+1
×
m.
Each column
n+1
of this matrix corresponds to the vector produced by the map coeff d above. This map allows us to
think of polynomials as points in a linear affine space, where Helly’s theorem applies.
We utilize this construction to import Clarkson’s method [17] for solving linear problems to algebraic geometry and, in particular, we make use of Helly-type theorems for varieties. Helly-type
theorems allow one to reduce the problem of solving the system to repeated solution of smaller
subsystems, whose size is a Helly number of intersecting linear spaces. As a result, our algorithms
achieve expected linear runtime in the number of input equations.
Theorem 1.1. Let F = {f1 , . . . , fm } ⊂ R be a system of polynomials, and let δ be the dimension of
the vector subspace generated by the coefficient vectors of the fi ’s, as described above.
Then there exists a sampling algorithm
that outputs F ′ = {fi1 , . . . , fiδ } ⊂ F such that V(F ) = V(F ′ )
in an expected number O δm + δO(δ) of calls to the primitive query that solves a small radical ideal
membership problem. F and F ′ generate the same ideal up to radicals.
It is important to point out that our sampling algorithm will find a small subsystem of the input
system that, when solved with whatever tools one has at their disposal, will give the same solution
set as the original (input) system. Here by ‘small’ we mean a system of size δ, where δ is polynomially
bounded or constant when the number of variables is constant or when the degree d is small.
That the rank δ of the coefficient matrix of the system of polynomials gives the combinatorial
dimension for this problem is shown in Theorem 4.11. There are several interesting special cases of
this result. For example, we obtain [21, Corollary 2] as a corollary: if f1 , . . . , fm ∈ K[x0 , . . . , xn ] are
homogeneous and
of degree at most d each, then the dimension δ of the vector subspace they generate
(see Lemma 4.6). Of course, in many situations in practice, this bound is not sharp,
is at most n+d
d
as many systems are of low rank. For example, this situation can arise if the monomial support
of the system is much smaller than the total number of monomials of degree d. In light of this,
Theorem 4.11 gives a better bound for low-rank systems. Note that we measure system complexity
by its rank, that is, the vector space dimension δ, and not the usual sparsity considerations such as
the structure of the monomial support of the system. Further, our result applies to non-homogeneous
systems as well. Its proof is presented in Section 4, along with the proof of Theorem 1.1.
4
JESÚS A. DE LOERA, SONJA PETROVIĆ, AND DESPINA STASI
Computing small generating sets of ideals. The problem of finding “nice” generating sets of
ideals has numerous applications in statistics, optimization, and other fields of science and engineering. Current methods of calculating minimal generating sets of ideals with an a priori large number
of generators are inefficient and rely mostly on Gröbner bases computations, since they usually involve ideal membership tests. Of course there are exceptional special cases, such as ideals of points
in projective space [40] or binomial systems [1]. Our second main result shows how to efficiently
extract a small or close to minimal generating set for any ideal from a given large generating set and
a bound on the size of a minimal generating set.
Theorem 1.2. Let I = (H) be an ideal generated by a (large) finite set of homogeneous polynomials
H, and suppose that γ is a known upper bound for the 0-th total Betti number β(R/I).
Then there exists a randomized algorithm that computes a generating set of I of size γ in expected
number of O(γ|H| + γ γ ) calls to the primitive query that solves a small ideal membership problem.
In particular, if γ = β(R/I), the algorithm computes a minimal generating set of I.
The proof is presented in Section 5.
2. A Warm-Up: Algebraic Helly-type theorems and the size of a meaningful sample
A Helly-type theorem has the following form: Given a family of objects F , a property P , and
a Helly number δ such that every subfamily of F with δ elements has property P , then the entire
family has property P. (See [19, 23, 47, 4].) In the original theorem of E. Helly, F is a finite family
of convex sets in Rn , the constant δ is n + 1, and the property P is to have a non-empty intersection
[34]. Here we are looking for non-linear algebraic versions of the same concept, where the objects in
F are algebraic varieties (hypersurfaces) or polynomials; the property desired is to have a common
point, or to generate the same ideal; and the Helly constant δ will be determined from the structure
of the problem at hand. To better understand the algorithms that we present, it is instructive to
consider two intuitive easy examples that highlight the fundamental combinatorial framework. The
first one is an obvious reformulation of Helly’s theorem for algebraic geometers.
Example 2.1. Let H = {L1 , L2 , . . . , Ls } be a family of affine linear subspaces in Rn . Consider the
case when s is much larger than n. One would like to answer the following question: when do all of
the linear varieties have a nonempty intersection? It is enough to check whether each subfamily of
H with n + 1 elements has a non-empty intersection, as that would imply, by Helly’s Theorem, that
H also has a non-empty intersection. Thus, in practice, one can reduce the task of deciding whether
∩si=1 Li 6= ∅ to the collection of smaller queries ∩n+1
j=1 Lij . However, instead of testing all possible
s
n+1 many (n + 1)-tuples, we may choose to randomly sample multiple (n + 1)-tuples. Each time
we sample, we either verify that one more (n + 1)-tuple has a non-empty intersection thus increasing
the certainty that the property holds for all (n + 1)-tuples, or else find a counterexample, a subfamily
without a common point, the existence of which trivially implies that ∩si=1 Li = ∅. This simple idea
is the foundation of a randomized approach. For now we ask the reader to observe that n + 1 is the
dimension of the vector space of (non-homogeneous) linear polynomials in n variables.
Example 2.2. The next example is just slightly more complicated, but illustrates well some key
concepts. Consider next H = {f1 (x1 , x2 ), f2 (x1 , x2 ), . . . , fs (x1 , x2 )}, a large family of affine real plane
curves of degree at most d. Imagine that H is huge, with millions of constraints fi , but the curves
are of small degree, say d = 2. Nevertheless, suppose that we are in charge of deciding whether the
curves in H have a common real point. Clearly, if the pair of polynomials f, g ∈ H intersect, they
do so in finitely many points, and, in particular, Bezout’s theorem guarantees that no more than d2
intersections occur. One can observe that if the system H has a solution, it must pass through some
of the (at most d2 ) points defined by the pair f, g alone. In fact, if we take triples f, g, h ∈ H, the
same bound of d2 holds, as well as the fact that the solutions for the entire H must also be part of
HELLY NUMBERS AND VIOLATOR SPACES IN COMPUTATIONAL ALGEBRA
5
the solutions for the triplet f, g, h. Same conclusions hold for quadruples, quintuples, and in general
δ-tuples. But how large does an integer δ have to be in order to function as a Helly number? We
seek a number δ such that if all δ-tuples of plane curves in H intersect, then all of the curves in H
must intersect. The reader can easily find examples where δ = d does not work, e.g., for d ≥ 2.
To answer the question posed in Example 2.2, we refer to Theorem 4.11 in Section 4. Without
re-stating the theorem here, we state the following Corollary and note that it gives a nice bound on
δ. Corollary 2.3 is implied by the observation that there are only d+2
monomials in two variables
2
of degree ≤ d (which says they span a linear subspace of that dimension inside the vector space of
all polynomials) and Theorem 4.11.
Corollary 2.3. Let H = {f1 (x, y), f2 (x, y), . . . , fs (x, y)} be a family of affine real plane curves of
degree at most d. If every δ = d+2
of the curves have a real intersection point, then all the curves
2
in H have a real intersection point. If we consider the same problem over the complex numbers, then
the same bound holds.
Thus, it suffices to check all δ-tuples of curves for a common real point of intersection, and if all of
those instances do intersect, then we are sure all |H| polynomials must have a common intersection
point, too. The result suggests a brute-force process to verify real feasibility of the system, which
of course is not a pretty proposition, given that |H| is assumed to be very large. Instead, Section 3
explains how to sample the set of δ-tuples in order to obtain a solution to the problem more efficiently.
Notice that it is important to find a small Helly number δ, as a way to find the smallest sampling
size necessary to detect common intersections. It turns out that in this example and in the case
when all fi are homogeneous, the Helly number is best possible [28].
3. Violator spaces and Clarkson’s sampling algorithms
The key observation in the previous section was that the existence of Helly-type theorems indicates
that there is a natural notion of sampling size to test for a property of varieties. Our goal is to import
to computational algebra an efficient randomized sampling algorithm by Clarkson. To import this
algorithm, we use the notion of violator spaces which we outline in the remainder of this section. We
illustrate the definitions using Example 2.2 as a running example in this section.
In 1992, Sharir and Welzl [41] identified special kinds of geometric optimization problems that
lend themselves to solution via repeated sampling of smaller subproblems: they called these LP-type
problems. Over the years, many other problems were identified as LP-type problems and several
abstractions and methods were proposed [2, 3, 11, 33, 37]. A powerful sampling scheme, devised by
Clarkson [17] for linear programming, works particularly well for geometric optimization problems in
small number of variables. Examples of applications include convex and linear programming, integer
linear programming, the problem of computing the minimum-volume ball or ellipsoid enclosing a
given point set in Rn , and the problem of finding the distance of two convex polytopes in Rn . In
2008, Gärtner, Matoušek, Rüst and Škovroň [31] invented violator spaces and showed they give a
much more general framework to work with LP-type problems. In fact, violator spaces include all
prior abstractions and were proven in [42] to be the most general framework in which Clarkson’s
sampling converges to a solution. Let us begin with the key definition of a violator space.
Definition 3.1 ([31]). A violator space is a pair (H, V), where H is a finite set and V a mapping
2H → 2H , such that the following two axioms hold:
Consistency: G ∩ V(G) = ∅ holds for all G ⊆ H, and
Locality:
V(G) = V(F ) holds for all F ⊆ G ⊆ H such that G ∩ V(F ) = ∅.
Example 3.2 (Example 2.2, continued). To illustrate our definition, we consider Example 2.2 of s
real plane curves {f1 , . . . , fs } = H.
6
JESÚS A. DE LOERA, SONJA PETROVIĆ, AND DESPINA STASI
A violator operator for testing the existence of a real point of intersection of a subset F ⊂ H of the
curves should capture the real intersection property. One possible way to define it is the following
map Vreal : 2H → 2H :
Vreal (F ) = {h ∈ H : V R (F ) ) V R (F ∪ {h})} ,
where V R (F ) is the set of common real intersection points of F , in other words, the real algebraic
variety of F . Note that, by definition, Vreal (F ) = ∅ if the curves in F have no real points of intersection. Before explaining why Vreal captures the real intersection property correctly (see Example 3.6),
let us show that the set (H, Vreal ) is a violator space according to Definition 3.1.
Consistency holds by definition of Vreal : for any h ∈ F , V R (F ) = V R (F ∪{h}), and so h 6∈ Vreal (F ).
To show locality, we begin by showing the auxiliary fact that V R (F ) = V R (G). The inclusion
V R (G) ⊆ V R (F ) is direct. This is because for any S ′ ⊆ S it is always the case that V R (S ′ ) ⊇ V R (S).
We need only show V R (F ) ⊆ V R (G). Recall F ⊆ G ⊆ H and G ∩ Vreal (F ) = ∅. Consider g ∈ G.
By the assumption, g does not violate F , and the proper containment V R (F ) ) V R (F ∪ {g}) does
not hold. Therefore the equality V R (F ) = V R (F ∪ {g}) must hold. By iteratively adding elements
of G r F to F and repeating the argument, we conclude that indeed V R (F ) = V R (G).
Finally, we argue that V R (F ) = V R (G) implies Vreal (F ) = Vreal (G). We first show Vreal (F ) ⊂
Vreal (G). Take h ∈ Vreal (F ). Then V R (G) = V R (F ) ) V R (F ∪ {h}) = V R (F ) ∩ V R ({h}) =
V R (G)∩V R ({h}) = V R (G∪{h}). The last and third-to-last equalities follow from the fact that for any
two sets S1 , S2 , one always has V R (S1 ∪S2 ) = V R (S1 )∩V R (S2 ). The containment Vreal (F ) ⊃ Vreal (G)
follows in a similar argument. Thus (H, Vreal ) is a violator space.
Every violator space comes equipped with three important components: a notion of basis, its
combinatorial dimension, and a primitive test procedure. We begin with the definition of a basis of
a violator space, analogous to the definition of a basis of a linear programming problem: a minimal
set of constraints that defines a solution space.
Definition 3.3 ([31, Definition 7]). Consider a violator space (H, V ). B ⊆ H is a basis if B ∩V(F ) 6=
∅ holds for all proper subsets F ⊂ B. For G ⊆ H, a basis of G is a minimal subset B of G with
V(B) = V(G).
It is very important to note that a violator operator can capture algebraic problems of interest
as long as the basis for that violator space corresponds to a basis of the algebraic object we study.
Violator space bases come with a natural combinatorial invariant, related to Helly numbers we
discussed earlier.
Definition 3.4 ([31, Definition 19]). The size of a largest basis of a violator space (H, V ) is called
the combinatorial dimension of the violator space and denoted by δ = δ(H, V ).
A crucial property was proved in [31]: knowing the violations V(G) for all G ⊆ H is enough to
compute the largest bases. To do so, one can utilize Clarkson’s randomized algorithm to compute
a basis of a violator space (H, V) with m = |H|. The results about the runtime and the size of the
sets involved are summarized below. The primitive operation, used as black box in all stages of the
algorithm, is the violation test primitive.
Definition 3.5. Given a violator space (H, V), some set G ( H, and some element h ∈ H \ G, the
primitive test decides whether h ∈ V(G).
The running example illustrates these three key ingredients.
Example 3.6 (Example 3.2, continued). In the example of s real plane curves, the violator operator
we defined detects whether the polynomials have a real point of intersection. Note that a basis would
be a (minimal) set of curves B = {fi1 , . . . , fiδ }, for some δ < s, such that either the curves in B
have no real point of intersection, or the real points of intersection of the curves in B are the real
HELLY NUMBERS AND VIOLATOR SPACES IN COMPUTATIONAL ALGEBRA
7
intersection of all of H = {f1 , . . . , fs }. If the set F has no real intersection point, then Vreal (F ) = ∅ by
definition, so that set F could be a basis in the sense that it is a certificate of infeasibility for this realintersection problem. If, on the other hand, F does have a real intersection point, and Vreal (F ) = ∅,
then this means that F is a basis in the sense that the curves in F capture the intersections of all of
H. The combinatorial dimension for general H is provided by Corollary 2.3, and it equals δ = d+2
2 .
However, special structure of the curves in H may imply a smaller combinatorial dimension.
The primitive query simply checks, given fi ∈ H and a candidate subset G ⊆ H, whether the
set of real points of intersection of G ∪ {fi } is smaller than the set of real points of intersection of
the curves in G alone. The role of the primitive query is therefore not to find a basis directly, but
to check, instead, whether a given candidate subset G can be a basis of H. This can be done by
checking whether fi ∈ Vreal (G) for all fi ∈ H \ G. Clearly, given the primitive test, a basis for H can
be found by simply testing all sets of size at most δ, but that would be a waste because the number
of times one would need to call the primitive would be O(|H|δ+1 ).
As we will see, this brute-force approach can be avoided. Namely, in our current
example, the
d+2
randomized algorithm from Theorem 3.7 below will only sample subsets of δ = 2 curves from the
set {f1 , . . . , fs }, and find a basis of the violator space of size δ in the sense explained above.
The sampling method in [17] avoids a full brute-force approach. It is presented in two stages,
referred to as Clarkson’s first and second algorithm. We outline these below.
Clarkson’s first algorithm, in the first iteration, draws a small random sample R ⊂ G, calls the
second stage to calculate the basis C of R, and returns C if it is already a basis for the larger subset
G. If C is not already a basis, but the elements of G \ C violating R are few, it adds those elements
to a growing set of violators W , and repeats the process with C being calculated as the basis of the
set W ∪ R for a new randomly chosen small R ⊂ G \ W . The crucial point here is that |R| is much
smaller than |G| and, consequently, it acts as a Helly number of sorts.
Clarkson’s second algorithm (Basis2) iteratively picks a random small (6δ2 elements) subset R of
G, finds a basis C for R by exhaustively testing each possible subset (BruteForce;) taking advantage
of the fact that the sample R is very small, and then calculates the violators of G \ C. At each
iteration, elements that appear in bases with small violator sets get a higher probability of being
selected.
This idea is very important: we are biasing the sampling process, so that some constraints will
be more likely to be chosen. This is accomplished by considering every element h of the set G as
having a multiplicity m(h); the multiplicity of a set is the sum of the multiplicities of its elements.
The process is repeated until a basis of G is found, i.e. until V(G \ C) is empty.
Again, as described above, all one needs is to be able to answer the Primitive query: Given
G ⊂ H and h ∈ H \ G, decide whether h ∈ V (G). The runtime is given in terms of the combinatorial
dimension δ(H, V ) and the size of H. The key result we will use in the rest of the paper concerns
the complexity of finding a basis:
Theorem 3.7. [31, Theorem 27] Using Clarkson’s algorithms, a basis of H of a violator space (H, V)
can be found by answering the primitive query an expected O δ |H| + δO(δ) times.
It is very important to note that, in both stages of Clarkson’s method, the query h ∈ V(C)
is answered via calls to the primitive as a black box. In our algebraic applications, the primitive
computation requires solving a small-size subsystem (e.g., via Gröbner bases or numerical algebraic
geometry methods), or an ideal membership query applied to the ideal generated by a small subset
of the given polynomials. On the other hand, the combinatorial dimension relates to the Helly
number of the problem which is usually a number that is problem-dependent and requires non-trivial
mathematical results.
8
JESÚS A. DE LOERA, SONJA PETROVIĆ, AND DESPINA STASI
Algorithm 1: Clarkson’s first algorithm
input : G ⊆ H, δ: combinatorial complexity of H
output: B, a basis for G
1
2
3
4
5
6
7
8
9
10
11
12
13
14
if |G| ≤ 9δ2 then
return Basis2(G)
else
W ←∅
repeat
p
R ← random subset of G \ W with ⌊δ |G|⌋ elements.
C ← Basis2(W ∪ R)
V ← {h ∈ G \ C s.t. h ∈ V(C)}
p
if |V | ≤ 2 |G| then
W ←W ∪V
end
until V = ∅
end
return C.
Algorithm 2: Clarkson’s second algorithm: Basis2(G)
input : G ⊆ H; δ: combinatorial complexity of H.
output: B: a basis of G
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
if |G| ≤ 6δ2 then
return BruteForce(G)
else
repeat
R ← random subset of G with 6δ2 elements.
C ← BruteForce(R)
V ← {h ∈ G \ C s.t. h ∈ V(C)}
if m(V ) ≤ m(G)/3δ then
for h ∈ V do
m(h) ← 2m(h)
end
end
until V = ∅
end
return C.
In the two sections that follow we show how violator spaces naturally arise in non-linear algebra
of polynomials.
4. A violator space for solving overdetermined systems
We discuss our random sampling approach to solve large-size (non-linear) polynomial systems by
applying Clarkson’s algorithm. In particular, we prove Theorem 1.1 as a corollary of Theorem 4.11.
This result is motivated by, and extends, Helly-type theorems for varieties from [21] and [28], which
HELLY NUMBERS AND VIOLATOR SPACES IN COMPUTATIONAL ALGEBRA
9
we use to show that the above algorithms apply to large dense homogeneous systems as well (Corollary 4.7).
First, we define a violator space that captures (in the sense explained in the previous section)
solvability of a polynomial system.
Definition 4.1. [Violator Space for solvability of polynomial systems] Let S ⊂ H be finite subsets
of polynomials in R. Define the violator operator Vsolve : 2H → 2H to record the set of polynomials
in H which do not vanish on the variety V(S). Formally,
Vsolve (S) = {f ∈ H : V(S) is not contained in V(f )}.
Lemma 4.2. The pair (H, Vsolve ) is a violator space.
Proof. Note that Vsolve (S) ∩ S = ∅ by definition of Vsolve (S), and thus the operator satisfies the
consistency axiom.
To show locality, suppose that F ( G ⊂ H and G ∩ Vsolve (F ) = ∅. Since F ( G we know that
V(G) ⊆ V(F ). On the other hand, T
by definition, G ∩ Vsolve (F ) = ∅ implies that V(F ) ⊆ V(g) for all
g ∈ G. Thus V(F ) is contained in g∈G V(g) = V(G). But then the two varieties are equal.
To complete the argument we show that V(F ) = V(G) implies Vsolve (F ) = Vsolve (G). If h ∈
Vsolve (F ) then V(h) cannot contain V(F ) = V(G), thus h ∈ Vsolve (G) too. The argument is symmetric, hence Vsolve (F ) = Vsolve (G).
It follows from the definition that the operator Vsolve gives rise to a violator space for which a
basis B of G ⊂ H is a set of polynomials such that V(B) = V(G). Therefore, a basis B ⊂ G will
either be a subset of polynomials that has no solution and as such be a certificate of infeasibility of
the whole system G, or it will provide a set of polynomials that are sufficient to find all common
solutions of G, i.e., the variety V(G).
Next, we need a violation primitive test that decides whether h ∈ Vsolve (F ), as in Definition 3.5.
By the definition above, this is equivalent to asking whether h vanishes on all irreducible components
of the algebraic variety V(F ). As is well known, the points of V(F ) where the polynomial h does
not vanish correspond to the variety associated with the saturation ideal ((F ) : h∞ ). Thus, we may
use ideal saturations for the violation primitive. For completeness, we recall the following standard
definitions. The saturation of the ideal (F ) with respect to f , denoted by ((F ) : f ∞ ), is defined to
be the ideal of polynomials g ∈ R with f m g ∈ I for some m > 0. This operation removes from
the variety V(F ) the irreducible components on which the polynomial f vanishes. Recall that every
variety can be decomposed into irreducible components (cf. [18, Section 4.6] for example). The
corresponding algebraic operation is the primary decomposition of the ideal defining this variety.
Lemma 4.3 (e.g. [5, Chapter 4]). Let ∩m
i=1 Qi be a minimal primary decomposition for the ideal I.
The saturation ideal (I : f ∞ ) equals ∩f ∈/ √Qi Qi .
m
∞
m
∞
∞
Proof. It is known that
√ (∩i=1 Qi : f )∞= ∩i=1 (Qi : f ). We observe further that (Qi : f ) = Qi if
f does not belong to Qi and (Qi : f ) = (1) otherwise.
This allows us to set up the primitive query for (H, Vsolve ). However we do not need to calculate
the decomposition explicitly, but can instead carry it out using elimination ideals via Gröbner bases,
as explained for example in [18, Exercise 4.4.9].
Observation 4.4. The primitive query for (H, Vsolve ) is simply the saturation test explained above.
Remark 4.5. There is an obvious reformulation of these two ingredients that is worth stating
explicitly. Namely, since a basis B for the violator space (H,p
Vsolve ) ispa set of polynomials such
that V(B) = V(H), the strong Nullstellensatz implies that (B) = (H). Thus a basis determines the ideal of the input system up to radicals, and we could have named the violator operator
10
JESÚS A. DE LOERA, SONJA PETROVIĆ, AND DESPINA STASI
Vsolve ≡ Vradical instead. Furthermore, a polynomial
p h vanishing on all irreducible components of
the algebraic variety V(F ) is equivalent to h ∈ (F ), i.e., h belonging to the radical of the ideal
(F ). In particular, the primitive query for Vsolve can also be stated as the radical ideal membership
test. This test
p can be implemented using Gröbner bases, as explained for example in [18, Proposition
4.2.8]: h ∈ (F ) if and only if 1 ∈ (F, 1 − yh) ⊆ K[x0 , . . . , xn , y]. Therefore, computation of one
Gröbner basis of the ideal (F, 1 − yh) suffices to carry out this test.
Finally, we solve the problem of finding a combinatorial dimension for Vsolve . For this, consider,
as a warm up, the simpler situation where we have a Helly-type theorem for hypersurfaces defined
by homogeneous polynomials. This was proved by Motzkin [39] and then later reproved by Deza
and Frankl [21], and it provides us with a combinatorial dimension for guaranteeing that a largescale homogeneous system has a solution. Its proof relies on thinking of the polynomial ring R as a
K-vector space (see also the discussion before Definition 4.9).
Lemma 4.6 ([21], Corollary 2). Let f1 , . . . , fm ⊂ R be a system of homogeneous polynomials,
that is, fi ∈ [R]di , and define d = max{di }. Suppose that every subset of p = n+d
polynomials
d
{fi1 , . . . , fip } ⊂ {f1 , . . . , fm } has a solution. Then the entire system {f1 , . . . , fm } does as well.
Lemma 4.6 provides the combinatorial dimension that, along with the variety membership primitive from Observation 4.4, allows us to apply Clarkson’s algorithms to the violator space (H, Vsolve ).
Corollary 4.7. Let (f1 , . . . , fm ) ⊂ R be an ideal generated by m homogeneous polynomials in n + 1
variables of degree at most d; fi ∈ [R]di and d = max{di }. Let δ = n+d
d . Then there is an
adaptation of Clarkson’s sampling algorithm that, in an expected O δm + δO(δ) number of calls to
the primitive query 4.4, computes {fi1 , . . . , fiδ } such that V(f1 , . . . , fm ) = V(fi1 , . . . , fiδ ).
In particular, this algorithm is linear in the number of input equations m, and a randomized
polynomial time algorithm when the number of variables n + 1 and the largest degree d are fixed.
Furthermore, we can extend it to actually solve a large system: once a basis B = {fi1 , . . . , fiδ } for the
space ({f1 , . . . , fm }, Vsolve ) is found, then we can use any computer algebra software (e.g. [6, 32, 46])
to solve fi1 = · · · = fiδ = 0.
Note that Lemma 4.6 can be thought of as a statement about the complexity of Hilbert’s Null-
stellensatz. If (f1 , . . . , fm ) = R (i.e., V(f1 , . . . , fm ) = ∅), then there exists a subset of size δ = n+d
d
polynomials {fi1 , . . . , fiδ } such that V(fi1 , . . . , fiδ ) = ∅ as well. In particular, there is a Nullstel
is, in fact, only an upper bound,
lensatz certificate with that many elements. The dimension n+d
d
attainable only by dense systems. However, in practice, many systems are very large but sparse, and
possibly non-homogeneous. Let us highlight again that the notion of ‘sparsity’ we consider is captured by a low-rank property of the system of polynomial equations, made explicit below in terms of
the coefficient matrix. This is crucially different from the usual considerations of monomial supports
(Newton polytopes) of the system; instead, we look at the coefficients of the input polynomials - that
is, we linearize the problem and consider the related vector spaces, as illustrated in the following
example.
Example 4.8. Consider the following system consisting of two types of polynomials: polynomials of
the form x2i − 1 for i = 1, . . . , n, and polynomials of the form xi + xj for the pairs {i, j : i 6≡ j mod 2}
along with the additional pair i = 1, j = 3. This system has m = n2 + n + 1 equations, and the
interesting situation is when the number of variables is a large even number, that is, n = 2k for any
large integer k. This system of polynomials generates the 2-coloring ideal of a particular n-vertex
non-chordal graph. (See [20] and references therein for our motivation to consider this particular
system.)
HELLY NUMBERS AND VIOLATOR SPACES IN COMPUTATIONAL ALGEBRA
11
Consider a concrete graph. Take the n-cycle with all possible even chords, and one extra edge
{1, 3}. Thus the pairs {i, j} are indexed by edges of the graph G on n nodes where all odd-numbered
vertices are connected to all even-numbered vertices, and with one additional edge {1, 3}.
We wish to decide if the system has a solution, but since there are n2 + n + 1 many polynomials,
we would like to try to avoid computing a Gröbner basis of this ideal. Instead, we search for a
subsystem of some specific size that determines the same variety. It turns out that the system
actually has no solution. Indeed, a certificate for infeasibility is a random subsystem consisting of
the first n quadratic equations, n − 1 of the edge equations xi + xj with {i, j : i 6≡ j mod 2}, and
the additional equation x1 + x3 . For example, the first n − 1 edge polynomials will do to construct
a 2n-sized certificate of this form. Why is the number n + (n − 1) + 1 = 2n so special?
To answer this question, let us linearize the problem: to each of the polynomials f associate
a coefficient (column) vector coeff(f ) ∈ C2n+1 whose coordinates are indexed by the monomials
appearing in the system x21 , . . . , x2n , 1, x1 , . . . , xn . Putting all these column vectors in one matrix
produces the coefficient matrix of the system of the form
In
0
− 1 0,
0
E
where In is the n × n identity, −1 is the row vector with all entries −1, and E is the vertex-edge
incidence matrix of the graph G. Since it is known that the rank of an edge-incidence matrix of an
n-vertex connected graph is n − 1, the rank of this matrix is δ = n + (n − 1) + 1 = 2n.
Remarkably, the magic size of the infeasibility certificate equals the rank of this coefficient matrix.
This motivating example suggests that the desired Helly-type number of this problem is captured
by a natural low-rank property of the system. To define it precisely, let us revisit the extension of
the Veronese embedding to non-homogeneous polynomials explained in the Introduction. Here we
adopt the notation from [7, Section 2] and consider polynomials in R of
degree up to d as a K-vector
d+n+1
n+1
n+1
has dimension n+1 , which, of course, equals the
space denoted by Cd . The vector space Cd
number of monomials in n + 1 variables of (total) degree from 0 to d. In this way, any polynomial
f ∈ R is represented by a (column) vector, coeff(f ) ∈ Cdn+1 , whose entries are the coefficients of f .
Thus, any system S ⊂ R defines a matrix with |S| columns, each of which is an element of Cdn+1 .
Definition 4.9. A system S ⊂ R is said to have rank D if dimK hSi = D, where hSi is the vector
subspace of Cdn+1 generated by the coefficients of the polynomials in S.
We need to also make the notion of Helly-type theorems more precise in the setting of varieties.
Definition 4.10 (Adapted from Definition 1.1. in [28]). A set S ⊂ R is said to have the D-Helly property if for every nonempty subset S0 ⊂ S, one can find p1 , . . . , pD ∈ S0 with V(S0 ) = V(p1 , . . . , pD ).
The following result, which implies Theorem 1.1, is an extension of [28] to non-homogeneous
systems. It also implies (the contrapositive of) Lemma 4.6 when restricted to homogeneous systems,
in the case when the system has no solution. The proof follows that of [28], although we remove the
homogeneity assumption. We include it here for completeness.
Theorem 4.11. Any polynomial system S ⊂ R of rank D has the D-Helly property.
In other words, for all subsets P ⊂ S, there exist p1 , . . . , pD ∈ P such that V(P) = V(p1 , . . . , pD ).
Proof. Let P ⊂ S be an arbitrary subset of polynomials, and denote by hPi ⊂ Cdn+1 the vector
subspace it generates. Let d0 = dimK hPi. We need to find polynomials p1 , . . . , pD such that
V(p1 , . . . , pD ) = V(P). Note that d0 ≤ D, of course, so it is sufficient to consider the case P = S.
Choose a vector space basis hp1 , . . . , pD i = hPi = hSi. It suffices to show V(p1 , . . . , pD ) ⊆ V(S);
indeed, the inclusions V(S) ⊆ V(P) ⊆ V(p1 , . . . , pD ) already hold.
12
JESÚS A. DE LOERA, SONJA PETROVIĆ, AND DESPINA STASI
Suppose, on the contrary, there exists x = (x0 , . . . , xn ) ∈ Cn+1 and p ∈ S such that p(x) 6= 0 but
pi (x) = 0 P
for all i = 1, . . . , D. Since pi ’s generate S as a vector space, there exist constants γi ∈ K
with p = i γi pi , implying that p(x) = 0, a contradiction.
The proof above is constructive: to find a subset p1 , . . . , pD ∈ S0 , one only needs to compute a
vector space basis for hS0 i. Thus, linear algebra (i.e., Gaussian elimination) can construct this subset
in time O(|S0 |3 ). The sampling algorithm based on violator spaces is more efficient.
Proof of Theorem 1.1. From Lemma 4.2, we know that ({f1 , . . . , fm }, Vsolve ) is a violator space.
Theorem 4.11 shows that it has a combinatorial dimension, and Observation 4.4 shows that there
exists a way to answer the primitive test. Having these ingredients, Theorem 3.7 holds and it is
possible for us to apply Clarkson’s Algorithm again.
Remark 4.5 provides the following interpretation of Theorem 4.11:
Corollary 4.12. Let I = (f1 , . . . , fm ) ⊂ R and let D = dimK hf1√
, . . . , fm
pi. Then, for all subsets P
of the generators f1 , . . . , fm , there exist p1 , . . . , pD ∈ P such that P = (p1 , . . . , pD ).
5. A violator space for finding generating sets of small cardinality
In this section, we apply the violator space approach to obtain a version of Clarkson’s algorithm
for calculating small generating sets of general homogeneous ideals as defined on page 3. As in
Section 4, this task rests upon three ingredients: the appropriate violator operator, understanding
the combinatorial dimension for this problem, and a suitable primitive query which we will use as a
black box. As before, fixing the definition of the violator operator induces the meaning of the word
‘basis’, as well as the construction of the black-box primitive.
To determine the natural violator space for the ideal generation problem, let I ⊂ R be a homogeneous ideal, H some initial generating set of I, and define the operator VSmallGen as follows.
Definition 5.1. [Violator Space for Homogeneous Ideal Generators] Let S ⊂ H be finite subsets of
R. We define the operator VSmallGen : 2H → 2H to record the set of polynomials in H that are not
in the ideal generated by the polynomials in S. Formally,
VSmallGen (S) = {f ∈ H : (S, f ) ) (S)}.
Equivalently, the operator can be viewed as VSmallGen (S) = {f ∈ H : f 6∈ (S)}.
Lemma 5.2. The pair (H, VSmallGen ) is a violator space.
Proof. Note that VSmallGen (S) ∩ S = ∅ by definition of VSmallGen (S), and thus the operator satisfies
the consistency axiom.
To show locality, suppose that F ( G ⊂ H and G ∩ VSmallGen (F ) = ∅. Since F ( G, (F ) ⊆ (G).
On the other hand G ∩ VSmallGen (F ) = ∅ implies that G ⊆ (F ) which in turn implies that (G) ⊆ (F ).
Then the ideals are equal. Then, because (G) = (F ) we can prove that VSmallGen (F ) = VSmallGen (G).
Note first that (G) = (F ), holds if and only if (G, h) = (F, h) for all polynomials h ∈ H.
Finally, to show VSmallGen (F ) = VSmallGen (G), we note that h ∈ VSmallGen (F ) if and only if
(G, h) = (F, h) ) (F ) = (G), and this chain of equations and containment holds if and only if
h ∈ VSmallGen (G). Therefore, locality holds as well and VSmallGen is a violator space operator.
It is clear from the definition that (H, VSmallGen ) is a violator space for which the basis of G ⊂ H
is a minimal generating set of the ideal (G).
The next ingredient in this problem is the combinatorial dimension: the size of the largest minimal
generating set. This natural combinatorial dimension already exists in commutative algebra, namely,
it equals a certain Betti number. (Recall that Betti numbers are the ranks βi,j of modules in the
minimal (graded) free resolution of the ring R/I; see, for example, [24, Section 1B (pages 5-9)].)
HELLY NUMBERS AND VIOLATOR SPACES IN COMPUTATIONAL ALGEBRA
13
Specifically, the number β0,j is defined as the number of elements of degree j required among any set
of minimal generators
of I. The (0-th) total Betti number of R/I, which we will denote by β(R/I),
P
simply equals j β0,j , and is then the total number of minimal generators of the ideal I. It is well
known that while I has many generating sets, every minimal generating set has the same cardinality,
namely β(R/I). In conclusion, it is known that
Observation 5.3. The combinatorial dimension for (H, VSmallGen ) is the (0-th) total Betti number
of the ideal I = (H); in symbols, β(R/I) = δ(H, VSmallGen ).
Although it may be difficult to exactly compute β(R/I) in general, a natural upper bound for
β(R/I) is the Betti number for any of its initial ideals (the standard inequality holds by uppersemicontinuity; see e.g. [38, Theorem 8.29]). In particular, if H is known to contain a Gröbner
basis with respect to some monomial order, then the combinatorial dimension can be estimated by
computing the minimal generators of an initial ideal of (H), which is a monomial ideal problem and
therefore easy. In general, however, we only need β(R/I) < |H| for the proposed algorithms to be
efficient.
The last necessary ingredient is the primitive query for VSmallGen .
Observation 5.4. The primitive query for VSmallGen , deciding if h ∈ VSmallGen (G) given h ∈ H
and G ⊂ H, is an ideal membership test.
Of course, the answer to the query is usually Gröbner-based, but, as before, the size of the
subsystems G on which we call the primitive query is small: (O(δ2 )). In fact, it is easy to see that
many small Gröbner computations for ideal membership cost less than the state-of-the-art, which
includes at least one large Gröbner computation.
Proof of Theorem 1.2. From Lemma 5.2 we know (H, VSmallGen ) is a violator space and we have
shown it has a combinatorial dimension and a way to answer the primitive test. Having these
ingredients, Theorem 3.7 holds and it is possible for us to apply Clarkson’s Algorithm.
Remark 5.5. Intuitively, the standard algorithm for finding minimal generators needs to at least
compute a Gröbner basis for an ideal generated by |H| polynomials, and in fact it is much worse
than that. One can simplify this by skipping the computation of useless S-pairs (e.g. as in [30]), but
improvement is not by an order of magnitude, overall. The algorithm remains doubly exponential in
the size of H for general input. In contrast, our randomized algorithm distributes the computation
into many small Gröbner basis calculations, where “many” means no more than O(β|H| + β β ), and
“small” means the ideal is generated by only O(β 2 ) polynomials.
To conclude, in a forthcoming paper we will study further the structure of our violator spaces
and discuss the use of more LP-type methods for the same algorithmic problems. We also intend to
present some experimental results for the sampling techniques we discussed here. We expect better
performance of the randomized methods versus the traditional deterministic algorithms.
Acknowledgements
We are grateful to Nina Amenta, Robert Ellis, Diane Maclagan, Pablo Soberón, Cynthia Vinzant,
and three anonymous referees for many useful comments and references. The first author is grateful
to IMA, the Institute of Mathematics and its Applications of the Univ. of Minnesota for the support
received when this paper was being written. He was also supported by a UC MEXUS grant. This
work was supported by the NSF grant DMS-1522662.
14
JESÚS A. DE LOERA, SONJA PETROVIĆ, AND DESPINA STASI
References
[1] 4ti2 team. 4ti2-a software package for algebraic, geometric and combinatorial problems on linear spaces combinatorial problems on linear spaces.
[2] N. Amenta. Bounded boxes, Hausdorff distance, and a new proof of an interesting Helly-type theorem. In Proceedings of the 10th Annual Symposium on Computational Geometry (SCG), pages 340–347. ACM Press, 1994.
[3] N. Amenta. Helly theorems and generalized linear programming. Discrete and Computational Geometry, 12:241–
261, 1994.
[4] Nina Amenta, Jesús A. De Loera, and Pablo Soberón. Helly’s theorem: New variations and applications. Preprint.
arXiv:1508.07606. math.MG (math.CO).
[5] M. F. Atiyah and I. G. Macdonald. Introduction to commutative algebra. Addison-Wesley Publishing Co., Reading,
Mass.-London-Don Mills, Ont., 1969.
[6] D. J. Bates, J. D. Hauenstein, A. J. Sommese, and C. W. Wampler. Bertini: Software for numerical algebraic
geometry. Available at bertini.nd.edu with permanent doi: dx.doi.org/10.7274/R0H41PB5.
[7] K. Batselier, P. Dreesen, and B. De Moor. The geometry of multivariate polynomial division and elimination.
SIAM Journal on Matrix Analysis and Applications, 34(1):102–125, 2013.
[8] C. Beltrán and L.M. Pardo. On Smale’s 17th problem: a probabilistic positive solution. Foundations of Computational Mathematics, 8(1):1–43, 2008.
[9] C. Beltrán and L.M Pardo. Smale’s 17th problem: average polynomial time to compute affine and projective
solutions. J. Amer. Math. Soc., 22(2):363–385, 2009.
[10] E. R. Berlekamp. Factoring polynomials over large finite fields. Mathematics of Computation, 24(111):713–735,
1970.
[11] H. Björklund, S. Sandberg, and S. Vorobyov. A discrete subexponential algorithm for parity games. In Proc. 20th
Annual Symposium on Theoretical Aspects of Computer Science (STACS), pages 663–674. Springer-Verlag, 2003.
[12] Y. Brise and B. Gärtner. Clarkson’s algorithm for violator spaces. In 21st Canadian Conference on Computational
Geometry (CCCG2009), pages 9–12, 2009.
[13] B. Buchberger. Ein Algorithmus zum Auffinden der Basiselemente des Restklassenringes nach einem nulldimensionalen Polynomideal. PhD thesis, Leopold-Franzens University, 1965.
[14] B. Buchberger. Bruno Buchberger’s PhD thesis 1965: An algorithm for finding the basis elements of the residue
class ring of a zero dimensional polynomial ideal (English translation). Journal of Symbolic Computation, 41(34):475–511, 2006.
[15] P. Bürgisser and F. Cucker. On a problem posed by Steve Smale. Annals of Mathematics, 174(3):1785–1836, 2011.
[16] D. Cifuentes and P. Parrilo. Exploiting chordal structure in polynomial ideals: a Gröbner bases approach. Preprint:
arXiv:1411.1745v1 [cs.SC].
[17] K. L. Clarkson. Las Vegas algorithms for linear and integer programming. Journal of the ACM, 42(2):488–499,
1995.
[18] D. Cox, J.B. Little, and D. O’Shea. Ideals, Varieties, and Algorithms: An Introduction to Computational Algebraic
Geometry and Commutative Algebra. Springer, 2007.
[19] L. Danzer, B. Grünbaum, and V. Klee. Helly’s theorem and its relatives. In Proc. Sympos. Pure Math., Vol. VII,
pages 101–180. Amer. Math. Soc., Providence, R.I., 1963.
[20] J. A. De Loera, S. Margulies, M. Pernpeintner, E. Riedl, D. Rolnick, G. Spencer, D. Stasi, and J. Swenson. Graphcoloring ideals: Nullstellensatz certificates, Gröbner bases for chordal graphs, and hardness of Gröbner bases. In
International Symposium on Symbolic and Algebraic Computation (ISSAC), 2015.
[21] M. Deza and P. Frankl. A Helly type theorem for hypersurfaces. J. Combin. Theory Ser. A, 45(1):27–30, 1987.
[22] T.W. Dube. The structure of polynomial ideals and Gröbner bases. SIAM Journal of Computing, 19(4):750–773,
1990.
[23] J. Eckhoff. Helly, Radon, and Carathéodory type theorems. In Handbook of convex geometry, Vol. A, B, pages
389–448. North-Holland, Amsterdam, 1993.
[24] D. Eisenbud. The geometry of syzygies, volume 229 of Graduate Texts in Mathematics. Springer-Verlag, New York,
2005. A second course in commutative algebra and algebraic geometry.
[25] J.-C. Faugére, P. Gaudry, L. Huot, and G. Renault. Polynomial systems solving by fast linear algebra. arXiv
preprint arXiv:1304.6039, 2013.
[26] J.-C. Faugére, P. Gianni, D. Lazard, and T. Mora. Efficient computation of zero dimensional Gröbner bases by
change of ordering. Journal of Symbolic Computation, 16(4):329–344, 1993.
[27] J.-C. Faugére, P.-J. Spaenlehauer, and J. Svartz. Sparse Gröbner bases: the unmixed case. arXiv preprint
arXiv:1402.7205, 2014.
[28] P. Frankl. Helly-type theorems for varieties. European J. Combin., 10(3):243–245, 1989.
HELLY NUMBERS AND VIOLATOR SPACES IN COMPUTATIONAL ALGEBRA
15
[29] S. Gao. Counting Zeros over Finite Fields Using Gröbner Bases. PhD thesis, MS Thesis in Logic and Computation,
2009.
[30] S. Gao, V. M Rodrigues, and J. Stroomer. Gröbner basis structure of finite sets of points. Preprint. Available on
http://www.math.clemson.edu/~ sgao/papers/GBstr.pdf. Last accesed: November 2015., 2003.
[31] B. Gärtner, J. Matoušek, L. Rüst, and P. Škovroň. Violator spaces: structure and algorithms. Discrete Appl.
Math., 156(11):2124–2141, 2008.
[32] D. R. Grayson and M. E. Stillman. Macaulay2, a software system for research in algebraic geometry. Available
at http://www.math.uiuc.edu/Macaulay2/.
[33] N. Halman. Discrete and lexicographic Helly theorems and their relations to LP-type problems. PhD thesis, Tel-Aviv
University, 2004.
[34] E. Helly. Über mengen konvexer Körper mit gemeinschaftlichen Punkten. Jahresbericht der Deutschen
Mathematiker-Vereinigung, 32(175–176), 1923.
[35] Y. N. Lakshman. On the complexity of computing a Gröbner basis for the radical of a zero dimensional ideal. In
Proceedings of the twenty-second annual ACM Symposium on Theory of Computing, pages 555–563. ACM, 1990.
[36] J. Matoušek. Lectures on discrete geometry, volume 212 of Graduate Texts in Mathematics. Springer-Verlag, New
York, 2002.
[37] J. Matoušek, M. Sharir, and E. Welzl. A subexponential bound for linear programming. Algorithmica, 16(4–5):498–
516, 1996.
[38] E. Miller and B. Sturmfels. Combinatorial commutative algebra. Graduate Texts in Mathematics. Springer, 2005.
[39] T. S. Motzkin. A proof of Hilbert’s Nullstellensatz. Math. Z., 63:341–344, 1955.
[40] I. Ramella. Algoritmi di Computer Algebra relativi agli ideali di punti dello spazio proiettivo. PhD thesis, Universita
degli studi di Napoli “Federico II”, 1990.
[41] M. Sharir and E. Welzl. A combinatorial bound for linear programming and related problems. In Proc. 9th Symposium on Theoretical Aspects of Computer Science (STACS), volume 577 of Lecture Notes in Computer Science,
pages 569–579. Springer-Verlag, 1992.
[42] P. Škovroň. Abstract models of optimization problems. PhD thesis, Charles University, Prague, 2007.
[43] R. Solovay and V. Strasse. A fast Monte-Carlo test for primality. SIAM Journal of Computing, 6(1), 1977.
[44] D.A. Spielman and S.H. Teng. Smoothed analysis: an attempt to explain the behavior of algorithms in practice.
Communications of the ACM, 52(10):76–84, 2009.
[45] B. Sturmfels. Sparse elimination theory. Mathematical Sciences Institute, Cornell University, 1991.
[46] J. Verschelde. PHCpack: a general-purpose solver for polynomial systems by homotopy continuation. Available at
http://homepages.math.uic.edu/∼jan/download.html.
[47] R. Wenger. Helly-type theorems and geometric transversals. In Handbook of discrete and computational geometry,
CRC Press Ser. Discrete Math. Appl., pages 63–82. CRC, Boca Raton, FL, 1997.
J. De Loera, Department of Mathematics, University of California, Davis.
E-mail address: [email protected]
S. Petrović, Applied Mathematics Department, Illinois Institute of Technology.
E-mail address: [email protected]
D. Stasi, Applied Mathematics Department, Illinois Institute of Technology.
E-mail address: [email protected]
| 0math.AC
|
arXiv:1712.04604v2 [cs.NE] 30 Jan 2018
Deep Quaternion Networks
Chase J. Gaudet
Anthony S. Maida
School of Computing & Informatics
University of Lousiana at Lafayette
Lafayette, USA
[email protected]
School of Computing & Informatics
University of Lousiana at Lafayette
Lafayette, USA
[email protected]
Abstract—The field of deep learning has seen significant
advancement in recent years. However, much of the existing work
has been focused on real-valued numbers. Recent work has shown
that a deep learning system using the complex numbers can be
deeper for a fixed parameter budget compared to its real-valued
counterpart. In this work, we explore the benefits of generalizing
one step further into the hyper-complex numbers, quaternions
specifically, and provide the architecture components needed to
build deep quaternion networks. We go over quaternion convolutions, present a quaternion weight initialization scheme, and
present algorithms for quaternion batch-normalization. These
pieces are tested in a classification model by end-to-end training
on the CIFAR-10 and CIFAR-100 data sets and a segmentation
model by end-to-end training on the KITTI Road Segmentation
data set. The quaternion networks show improved convergence
compared to real-valued and complex-valued networks, especially
on the segmentation task.
Index Terms—quaternion, complex, neural networks, deep
learning
I. I NTRODUCTION
There have been many advances in deep neural network
architectures in the past few years. One such improvement
is a normalization technique called batch normalization [1]
that standardizes the activations of layers inside a network
using minibatch statistics. It has been shown to regularize the
network as well as provide faster and more stable training.
Another improvement comes from architectures that add so
called shortcut paths to the network. These shortcut paths
connect later layers to earlier layers typically, which allows for
the stronger gradients to propagate to the earlier layers. This
method can be seen in Highway Networks [2] and Residual
Networks [3]. Other work has been done to find new activation
functions with more desirable properties. One example is
the exponential linear unit (ELU) [4], which attempts to
keep activations standardized. All of the above methods are
combating the vanishing gradient problem [5] that plagues
deep architectures. With solutions to this problem appearing
it is only natural to move to a system that will allow one to
construct deeper architectures with as low a parameter cost as
possible.
Other work in this area has explored the use of complex
and hyper-complex numbers, which are a generalization of
the complex, such as quaternions. Using complex numbers
in recurrent neural networks (RNNs) has been shown to
increase learning speed and provide a more noise robust
memory retrieval mechanism [6]–[8]. The first formulation of
complex batch normalization and complex weight initialization
is presented by [9] where they achieve some state of the art
results on the MusicNet data set. Hyper-complex numbers
are less explored in neural networks, but have seen use in
manual image and signal processing techniques [10]–[12].
Examples of using quaternion values in networks is mostly
limited to architectures that take in quaternion inputs or predict
quaternion outputs, but do not have quaternion weight values
[13], [14]. There are some more recent examples of building
models that use quaternions represented as real-values. In [15]
they used a quaternion multi-layer perceptron (QMLP) for
document understanding and [16] uses a similar approach in
processing multi-dimensional signals.
Building on [9] our contribution in this paper is to formulate
and implement quaternion convolution, batch normalization,
and weight initialization 1 . There arises some difficulty over
complex batch normalization that we had to overcome as their
is no analytic form for our inverse square root matrix.
II. M OTIVATION AND R ELATED W ORK
The ability of quaternions to effectively represent spatial
transformations and analyze multi-dimensional signals makes
them promising for applications in artificial intelligence.
One common use of quaternions is for representing rotation
into a more compact form. PoseNet [14] used a quaternion
as the target output in their model where the goal was to
recover the 6−DOF camera pose from a single RGB image.
The ability to encode rotations may make a quaternion network
more robust to rotational variance.
Quaternion representation has also been used in signal
processing. The amount of information in the phase of an
image has been shown to be sufficient to recover the majority
of information encoded in its magnitude by Oppenheim and
Lin [17]. The phase also encodes information such as shapes,
edges, and orientations. Quaternions can be represented as a
2 x 2 matrix of complex numbers, which gives them a group
of phases potentially holding more information compared to a
single phase.
Bulow and Sommer [12] used the higher complexity representation of quaternions by extending Gabor’s complex signal
to a quaternion one which was then used for texture segmentation. Another use of quaternion filters is shown in [11] where
1 Source
code located at https://github.com/gaudetcj/DeepQuaternionNetworks
they introduce a new class of filter based on convolution with
hyper-complex masks, and present three color edge detecting
filters. These filters rely on a three-space rotation about the
grey line of RGB space and when applied to a color image
produce an almost greyscale image with color edges where the
original image had a sharp change of color. More quaternion
filter use is shown in [18] where they show that it is effective in
the context of segmenting color images into regions of similar
color texture. They state the advantage of using quaternion
arithmetic is that a color can be represented and analyzed
as a single entity, which we will see holds for quaternion
convolution in a convolutional neural network architecture as
well in Section III-C.
A quaternionic extension of a feed forward neural network, for processing multi-dimensional signals, is shown in
[16]. They expect that quaternion neurons operate on multidimensional signals as single entities, rather than real-valued
neurons that deal with each element of signals independently.
A convolutional neural network (CNN) should be able to learn
a powerful set of quaternion filters for more impressive tasks.
Another large motivation is discussed in [9], which is that
complex numbers are more efficient and provide more robust
memory mechanisms compared to the reals [10]–[12]. They
continue that residual networks have a similar architecture to
associative memories since the residual shortcut paths compute
their residual and then sum it into the memory provided by
the identity connection. Again, given that quaternions can be
represented as a complex group, they may provide an even
more efficient and robust memory mechanisms.
III. Q UATERNION N ETWORK C OMPONENTS
This section will include the work done to obtain a working
deep quaternion network. Some of the longer derivations are
given in the Appendix.
A. Quaternion Representation
This representation of quaternions is not unique, but we will
stick to the above in this paper. It is also possible to represent
H as M (2, C) where M (2, C) is a 2 x 2 complex matrix.
With our real-valued representation a quaternion real-valued
2D convolution layer can be expressed as follows. Say that the
layer has N feature maps such that N is divisible by 4. We let
the first N/4 feature maps represent the real components, the
second N/4 represent the i imaginary components, the third
N/4 represent the j imaginary components, and the last N/4
represent the k imaginary components.
B. Quaternion Differentiability
In 1833 Hamilton proposed complex numbers C be defined
as the set R2 of ordered pairs (a, b) of real numbers. He
then began working to see if triplets (a, b, c) could extend
multiplication of complex numbers. In 1843 he discovered a
way to multiply in four dimensions instead of three, but the
multiplication lost commutativity. This construction is now
known as quaternions. Quaternions are composed of four
components, one real part, and three imaginary parts. Typically
denoted as
H = {a + bi + cj + dk : a, b, c, d ∈ R}
Since we will be performing quaternion arithmetic using
reals it is useful to embed H into a real-valued representation.
There exists an injective homomorphism from H to the matrix
ring M (4, R) where M (4, R) is a 4x4 real matrix. The 4 x 4
matrix can be written as
a −b −c −d
1 0 0 0
b a −d c
= a 0 1 0 0
c d
0 0 1 0
a −b
d −c b
a
0 0 0 1
0 −1 0 0
1 0 0 0
+ b
0 0 0 −1
0 0 1 0
0 0 −1 0
0 0
0 1
+ c
1 0
0 0
0 −1 0 0
0 0 0 −1
0 0 −1 0
. (4)
+ d
0 1 0
0
1 0 0
0
(1)
In order for the network to perform backpropagation the cost
function and activation functions used must be differentiable
with respect to the real, i, j, and k components of each
quaternion parameter of the network. As the complex chain
rule is shown in [9], we provide the quaternion chain rule
which is given in the Appendix section VII-A.
C. Quaternion Convolution
where a is the real part, (i, j, k) denotes the three imaginary
axis, and (b, c, d) denotes the three imaginary components.
Quaternions are governed by the following arithmetic:
Convolution in the quaternion domain is done by convolving
a quaternion filter matrix W = A+iB+jC+kD by a quaternion
vector h = w + ix + jy + kz. Performing the convolution by
using the distributive property and grouping terms one gets
i2 = j 2 = k 2 = ijk = −1
(2)
W ∗ h = (A ∗ w − B ∗ x − C ∗ y − D ∗ z)+
which, by enforcing distributivity, leads to the noncommutative
multiplication rules
i(A ∗ x + B ∗ w + C ∗ z − D ∗ y)+
j(A ∗ y − B ∗ z + C ∗ w + D ∗ x)+
ij = k, jk = i, ki = j, ji = −k, kj = −i, ik = −j (3)
k(A ∗ z + B ∗ y − C ∗ x + D ∗ w).
(5)
Using a matrix to represent the components of the convolution
we have:
R(W ∗ h)
A
I (W ∗ h) B
J (W ∗ h) = C
K (W ∗ h)
D
−B
A
D
−C
−C −D
w
x
−D C
∗
A −B y
B
A
z
(6)
An example is shown in Fig. 1 where one can see how quaternion convolution forces a linear depthwise mixture of the
channels. This is similar to a mixture of standard convolution
and depthwise separable convolution from [19]. This reuse of
filters on every layer and combination may help extract texture
information across channels as seen in [18].
sition of V−1 where V is the covariance matrix given by:
Vrr Vri Vrj Vrk
Vir Vii Vij Vik
V =
Vjr Vji Vjj Vjk
Vkr Vki Vkj Vkk
Cov(R{x}, R{x})
Cov(I {x}, R{x})
=
Cov(J {x}, R{x})
Cov(K {x}, R{x})
Cov(R{x}, I {x})
Cov(I {x}, I {x})
Cov(J {x}, I {x})
Cov(K {x}, I {x})
Cov(R{x}, J {x})
Cov(I {x}, J {x})
Cov(J {x}, J {x})
Cov(K {x}, J {x})
Cov(R{x}, K {x})
Cov(I {x}, K {x})
Cov(J {x}, K {x})
Cov(K {x}, K {x})
D. Quaternion Batch-Normalization
Batch-normalization [1] is used by the vast majority of all
deep networks to stabilize and speed up training. It works
by keeping the activations of the network at zero mean and
unit variance. The original formulation of batch-normalization
only works for real-values. Applying batch normalization to
complex or hyper-complex numbers is more difficult, one can
not simply translate and scale them such that their mean is
0 and their variance is 1. This would not give equal variance
in the multiple components of a complex or hyper-complex
number. To overcome this for complex numbers a whitening
approach is used [9], which scales the data by the square root
of their variances along each of the two principle components.
We use the same approach, but must whiten 4D vectors.
However, an issue arises in that there is no nice way to
calculate the inverse square root of a 4 x 4 matrix. It turns out
that the square root is not necessary and we can instead use the
Cholesky decomposition on our covariance matrix. The details
of why this works for whitening given in the Appendix section
VII-B. Now our whitening is accomplished by multiplying the
0-centered data (x − E[x]) by W:
x̃ = W(x − E[x])
(7)
where Cov(·) is the covariance and R{x}, I {x}, J {x}, and
K {x} are the real, i, j, and k components of x respectively.
Real-valued batch normalization also uses two learned
parameters, β and γ. Our shift parameter β must shift a
quaternion value so it is a quaternion value itself with real,
i, j, and k as learnable components. The scaling parameter γ
is a symmetric matrix of size matching V given by:
Vrr Vri Vrj Vrk
Vri Vii Vij Vik
γ=
(9)
Vrj Vij Vjj Vjk
Vrk Vik Vjk Vkk
Because of its symmetry it has only ten learnable parameters.
The variance of the components of input
√ x̃ are variance 1 so
the diagonal of γ is initialized to 1/ 16 in order to obtain a
modulus of 1 for the variance of the normalized value. The off
diagonal terms of γ and all components of β are initialized
to 0. The quaternion batch normalization is defined as:
BN(x̃) = γx̃ + β
(10)
E. Quaternion Weight Initialization
The proper initialization of weights is vital to convergence
of deep networks. In this work we derive our quaternion
weight initialization using the same procedure as Glorot and
Bengio [20] and He et al. [21].
To begin we find the variance of a quaternion weight:
W = |W |e(cosφ1 i+cosφ2 j+cosφ3 k)θ
= R{W } + I {W } + J {W } + K {W }.
where W is one of the matrices from the Cholesky decompo-
(8)
(11)
where |W | is the magnitude, θ and φ are angle arguments,
and cos2 φ1 + cos2 φ2 + cos2 φ3 = 1 [22].
Fig. 1. An illustration of quaternion convolution.
Variance is defined as
Var(W ) = E[|W |2 ] − (E[W ])2 ,
(12)
but since W is symmetric around 0 the term (E[W ])2 is 0.
We do not have a way to calculate Var(W ) = E[|W |2 ] so
we make use of the magnitude of quaternion normal values
|W |, which follows an independent normal distribution with
four degrees of freedom (DOFs). We can then calculate the
expected value of |W |2 to find our variance
Z ∞
E[|W |2 ] =
x2 f (x) dx = 4σ 2
(13)
∞
where f (x) is the four DOF distribution given in the Appendix.
And since Var(W ) = E[|W |2 ], we now have the variance
of W expressed in terms of a single parameter σ:
Var(W ) = 4σ 2 .
(14)
To follow the Glorot and Bengio [20] initialization we have
Var(W ) = 2/(nin +nout ), where nin and nout are the number
of input and output units respectivly.
p Setting this equal to (14)
and solving for σ gives σ = 1/ 2(nin + nout ). To follow
He et al. [21] initialization that is specialized for rectified
linear units (ReLUs) [23], then we have Var(W ) = 2/nin ,
which √
again setting equal to (14) and solving for σ gives
σ = 1/ 2nin .
As shown in (11) the weight has components |W |, θ, and
φ. We can initialize the magnitude |W | using our four DOF
distribution defined with the appropriate σ based on which
initialization scheme we are following. The angle components
are initialized using the uniform distribution between −π and
π where we ensure the constraint on φ.
IV. E XPERIMENTAL R ESULTS
Our experiments covered image classification using both the
CIFAR-10 and CIFAR-100 benchmarks and image segmentation using the KITTI Road Estimation benchmark.
A. Classification
We use the same architecture as the large model in [9],
which is a 110 layer Residual model similar to the one in
[3]. There is one difference between the real-valued network
and the ones used for both the complex and hyper-complex
valued networks. Because the datasets are all real-valued the
network must learn the imaginary or quaternion components.
We use the same technique as [9] where there is an additional
block immediately after the input which will learn the hypercomplex components
BN → ReLU → Conv → BN → ReLU → Conv.
Since all datasets that we work with are real-valued, we
present a way to learn their imaginary components to let the
rest of the network operate in the complex plane. We learn the
initial imaginary component of our input by performing the
operations present within a single real-valued residual block
This means that to maintain the same approximate parameter count the number of convolution kernels for the complex
network was increased. We however did not increase the
number of convolution kernels for the quaternion trials so
any increase in model performance comes from the quaternion
filters and at a lower hardware budget.
The architecture consists of 3 stages of repeating residual
blocks where at the end of each stage the images are downsized by strided convolutions. Each stage also doubles the
previous stage’s number of convolution kernels. The last layers
are a global average pooling layer followed by a single fully
connected layer with a softmax function used to classify the
input as either one of the 10 classes in CIFAR-10 or one of
the 100 classes in CIFAR-100.
We also followed their training procedure of using the
backpropagation algorithm with Stochastic Gradient Descent
with Nesterov momentum [24] set at 0.9. The norm of the
gradients are clipped to 1 and a custom learning rate scheduler
is used. The learning scheduler is the same used in [9] for a
direct comparison in performance. The learning rate is initially
set to 0.01 for the first 10 epochs and then set it to 0.1 from
epoch 10-100 and then cut by a factor of 10 at epochs 120
and 150. Table I presents our results along side the real and
complex valued networks. Our quaternion model outperforms
the real and complex networks on both datasets on a smaller
parameter budget.
Architecture
[3] Real
[9] Complex
Quaternion
CIFAR-10
CIFAR-100
6.37
5.60
27.09
5.44
26.01
TABLE I
C LASSIFICATION ERROR ON CIFAR-10 AND CIFAR-100. N OTE THAT [3]
IS A 110 LAYER RESIDUAL NETWORK , [9] IS 118 LAYER COMPLEX
NETWORK WITH THE SAME DESIGN AS THE PRIOR EXCEPT WITH
ADDITIONAL INITIAL LAYERS TO EXTRACT COMPLEX MAPPINGS .
B. Segmentation
For this experiment we used the same model as the above,
but cut the number of residual blocks out of the model for
memory reasons given that the KITTI data is large color
images. The 110 layer model has 10, 9, and 9 residual blocks
in the 3 stages talked about above, while this model has 2,
1, and 1 and does not perform any strided convolutions. This
gives a total of 38 layers.
The last layer is a 1 × 1 convolution with a sigmoid output
so we are getting a heatmap prediction the same size as the
input. The training procedure is also as above, but the learning
rate is scheduled differently. Here we begin at 0.01 for the first
10 epochs and then set it to 0.1 from epoch 10-50 and then cut
by a factor of 10 at 100 and 150. Table II presents our results
along side the real and complex valued networks where we
used Intersection over Union (IOU) for performance measure.
Quaternion outperformed the other two by a larger margin
compared to the classification tasks.
Architecture
KITTI
Real
0.747
Complex
0.769
Quaternion
0.827
TABLE II
IOU ON KITTI ROAD E STIMATION BENCHMARK .
V. C ONCLUSIONS
We have extended upon work looking into complex valued
networks by exploring quaternion values. We presented the
building blocks required to build and train deep quaternion
networks and used them to test residual architectures on
two common image classification benchmarks. We show that
they have competitive performance by beating both the real
and complex valued networks with less parameters. Future
work will be needed to test quaternion networks for more
segmentation datasets and for audio processing tasks.
VI. ACKNOWLEDGMENT
[15] T. Parcollet, M. Morchid, P.-M. Bousquet, R. Dufour, G. Linarès,
and R. De Mori, “Quaternion neural networks for spoken language
understanding,” in Spoken Language Technology Workshop (SLT), 2016
IEEE, pp. 362–368, IEEE, 2016.
[16] T. Minemoto, T. Isokawa, H. Nishimura, and N. Matsui, “Feed forward
neural network with random quaternionic neurons,” Signal Processing,
vol. 136, pp. 59–68, 2017.
[17] A. V. Oppenheim and J. S. Lim, “The importance of phase in signals,”
Proceedings of the IEEE, vol. 69, no. 5, pp. 529–541, 1981.
[18] L. Shi and B. Funt, “Quaternion color texture segmentation,” Computer
Vision and image understanding, vol. 107, no. 1, pp. 88–96, 2007.
[19] F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” arXiv preprint arXiv:1610.02357, 2016.
[20] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep
feedforward neural networks,” in Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pp. 249–256,
2010.
[21] K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers:
Surpassing human-level performance on imagenet classification,” in
Proceedings of the IEEE international conference on computer vision,
pp. 1026–1034, 2015.
[22] R. Piziak and D. Turner, “The polar form of a quaternion,” 2002.
We would like to thank James Dent of the University
of Louisiana at Lafayette Physics Department for helpful
discussions. We also thank Fugro for research time on this
project.
[23] V. Nair and G. E. Hinton, “Rectified linear units improve restricted boltzmann machines,” in Proceedings of the 27th international conference on
machine learning (ICML-10), pp. 807–814, 2010.
R EFERENCES
[25] A. Kessy, A. Lewin, and K. Strimmer, “Optimal whitening and decorrelation,” The American Statistician, no. just-accepted, 2017.
[1] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” in International
Conference on Machine Learning, pp. 448–456, 2015.
[2] R. K. Srivastava, K. Greff, and J. Schmidhuber, “Training very deep networks,” in Advances in neural information processing systems, pp. 2377–
2385, 2015.
[3] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, pp. 770–778, 2016.
[4] D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate
deep network learning by exponential linear units (elus),” arXiv preprint
arXiv:1511.07289, 2015.
[5] S. Hochreiter, “Untersuchungen zu dynamischen neuronalen netzen,”
Diploma, Technische Universität München, vol. 91, 1991.
[6] M. Arjovsky, A. Shah, and Y. Bengio, “Unitary evolution recurrent
neural networks,” in International Conference on Machine Learning,
pp. 1120–1128, 2016.
[7] I. Danihelka, G. Wayne, B. Uria, N. Kalchbrenner, and A. Graves, “Associative long short-term memory,” arXiv preprint arXiv:1602.03032,
2016.
[8] S. Wisdom, T. Powers, J. Hershey, J. Le Roux, and L. Atlas, “Fullcapacity unitary recurrent neural networks,” in Advances in Neural
Information Processing Systems, pp. 4880–4888, 2016.
[9] C. Trabelsi, O. Bilaniuk, D. Serdyuk, S. Subramanian, J. F. Santos,
S. Mehri, N. Rostamzadeh, Y. Bengio, and C. J. Pal, “Deep complex
networks,” arXiv preprint arXiv:1705.09792, 2017.
[10] T. Bülow, Hypercomplex spectral signal representations for the processing and analysis of images. Universität Kiel. Institut für Informatik und
Praktische Mathematik, 1999.
[11] S. J. Sangwine and T. A. Ell, “Colour image filters based on hypercomplex convolution,” IEE Proceedings-Vision, Image and Signal
Processing, vol. 147, no. 2, pp. 89–93, 2000.
[12] T. Bulow and G. Sommer, “Hypercomplex signals-a novel extension of
the analytic signal to the multidimensional case,” IEEE Transactions on
signal processing, vol. 49, no. 11, pp. 2844–2852, 2001.
[13] A. Rishiyur, “Neural networks with complex and quaternion inputs,”
arXiv preprint cs/0607090, 2006.
[14] A. Kendall, M. Grimes, and R. Cipolla, “Posenet: A convolutional
network for real-time 6-dof camera relocalization,” in Proceedings of
the IEEE international conference on computer vision, pp. 2938–2946,
2015.
[24] Y. Nesterov, “A method of solving a convex programming problem with
convergence rate o (1/k2),”
VII. A PPENDIX
A. The Generalized Quaternion Chain Rule for a Real-Valued
Function
We start by specifying the Jacobian. Let L be a real valued
loss function and q be a quaternion variable such that q =
a + i b + j c + k d where a, b, c, d ∈ R then,
∂L
∂L
∂L
∂L
∂L
=
+i
+j
+k
(15)
∂q
∂a
∂b
∂c
∂d
∂L
∂L
∂L
∂L
=
+i
+j
+k
∂R(q)
∂I(q)
∂J(q)
∂K(q)
= R(∇L (q)) + I(∇L (q)) + J(∇L (q)) + K(∇L (q))
∇L (q) =
Now let g = m + i n + j o + k p be another quaternion variable
where q can be expressed in terms of g and m, n, o, p ∈ R we
where W is an n x n ‘whitening’ matrix. Since cov(Z) = I it
follows that:
then have,
∂L
∂L
∂L
∂L
∂L
=
+i
+j
+k
(16)
∂g
∂m
∂n
∂o
∂p
∂L ∂a
∂L ∂b
∂L ∂c
∂L ∂d
=
+
+
+
∂a ∂m
∂b ∂m
∂c ∂m
∂d ∂m
∂L ∂a ∂L ∂b
∂L ∂c
∂L ∂d
+i
+
+
+
∂a ∂n
∂b ∂n
∂c ∂n
∂d ∂n
∂L ∂a ∂L ∂b ∂L ∂c ∂L ∂d
+j
+
+
+
∂a ∂o
∂b ∂o
∂c ∂o
∂d ∂o
∂L ∂a ∂L ∂b ∂L ∂c ∂L ∂d
+k
+
+
+
∂a ∂p
∂b ∂p
∂c ∂p
∂d ∂p
∂L ∂a
∂a
∂a
∂a
=
+i
+j
+k
∂a ∂m
∂n
∂o
∂p
∂L ∂b
∂b
∂b
∂b
+
+i
+j
+k
∂b ∂m
∂n
∂o
∂p
∂c
∂c
∂c
∂L ∂c
+i
+j
+k
+
∂c ∂m
∂n
∂o
∂p
∂L ∂d
∂d
∂d
∂d
+
+i
+j
+k
∂d ∂m
∂n
∂o
∂p
∂L
∂a
∂a
∂a
∂a
=
+i
+j
+k
∂R(q) ∂m
∂n
∂o
∂p
∂b
∂b
∂b
∂L
∂b
+i
+j
+k
+
∂I(q) ∂m
∂n
∂o
∂p
∂c
∂c
∂c
∂L
∂c
+i
+j
+k
+
∂J(q) ∂m
∂n
∂o
∂p
∂d
∂d
∂d
∂d
∂L
+i
+j
+k
+
∂K(q) ∂m
∂n
∂o
∂p
∂a
∂a
∂a
∂a
= R(∇L (q))
+i
+j
+k
∂m
∂n
∂o
∂p
∂b
∂b
∂b
∂b
+i
+j
+k
+ I(∇L (q))
∂m
∂n
∂o
∂p
∂c
∂c
∂c
∂c
+ J(∇L (q))
+i
+j
+k
∂m
∂n
∂o
∂p
∂d
∂d
∂d
∂d
+ K(∇L (q))
+i
+j
+k
∂m
∂n
∂o
∂p
∇L (q) =
B. Whitening a Matrix
E[ZZT ] = I
E[W(X − µ)(W(X − µ))T ] = I
E[W(X − µ)(X − µ)T WT ] = I
WΣWT = I
WΣWT W = W
WT W = Σ−1
(18)
From (18) it is clear that the Cholesky decomposition provides
a suitable (but not unique) method of finding W.
C. Cholesky Decomposition
Cholesky decomposition is an efficient way to implement
LU decomposition for symmetric matrices, which allows us
to find the square root. Consider AX = b, A = [aij ]n×n , and
aij = aji , then the Cholesky decomposition of A is given by
A = LL0 where
l11 0 . . . 0
..
l21 l22 . . .
.
(19)
L= .
.
.
.
.
.
.
.
0
.
ln1 ln2 ... lnn
Let lki be the k th row and ith column entry of L, then
k<i
0,
q
Pi−1 2
lki =
aii − j=1 lkj ,
k=i
1 (a − Pi−1 l l ), i < k
ki
j=1 ij kj
lii
D. 4 DOF Independent Normal Distribution
Consider the four-dimensional vector Y = (S, T, U, V )
which has components that are normally distributed, centered
at zero, and independent. Then S, T , U , and V all have density
functions
2
Hx
Let X be an n x n matrix and cov(X) = Σ is the
symmetric covariance matrix of the same size. Whitening a
matrix linearly decorrelates the input dimensions, meaning that
whitening transforms X into Z such that cov(Z) = I where I
is the identity matrix [25]. The matrix Z can be written as:
Z = W(X − µ)
(17)
2
e−x /(2σ )
fS (x; σ) = fT (x; σ) = fU (x; σ) = fV (x; σ) = √
.
2πσ 2
(20)
Let
X
be
the
length
of
Y,
which
means
X
=
√
S 2 + T 2 + U 2 + V 2 . Then X has the cumulative distribution function
ZZZZ
FX (x; σ) =
fS (µ; σ)fT (µ; σ)fU (µ; σ)fV (µ; σ) dA,
where Hx is the four-dimensional sphere
n
o
p
Hx = (s, t, u, v) :
s2 + t2 + u2 + v 2 < x .
(21)
(22)
We then can write the integral in polar representation
Z πZ πZ 2π
Z x
2
1
3 −r
2σ 2 sin(θ)sin(φ)cos(ψ) drdθdφdψ
FX (x; σ) =
r
e
4π 2 σ 4 0 0 0 0
Z x
2
2
1
=
r3 e−r /(2σ ) dr.
(23)
4
2σ 0
The probability density function of X is the derivative of its
cumulative distribution function so we use the funamental
theorem of calculus on (23) to finally arrive at
fX (x; σ) =
d
1 3 −x2 /(2σ2 )
x e
.
FX (x; σ) =
dx
2σ 4
(24)
| 9cs.NE
|
arXiv:1712.09938v1 [math.AC] 28 Dec 2017
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
CLAUDIU RAICU
Abstract. The study of homological invariants such as Tor, Ext and local cohomology modules constitutes an
important direction in commutative algebra. Explicit descriptions of these invariants are notoriously difficult
to find and often involve combining an array of techniques from other fields of mathematics. In recent years
tools from algebraic geometry and representation theory have been successfully employed in order to shed some
light on the structure of homological invariants associated with determinantal rings. The goal of this notes
is to survey some of these results, focusing on examples in an attempt to clarify some of the more technical
statements.
1. Introduction
Consider a polynomial ring S = C[X1 , · · · , XN ] and a group homomorphism G −→ GLN (C) giving rise
to an action of G on S by linear changes of coordinates. It is natural to study the following problem:
Problem 1.1. Classify the homogeneous ideals I ⊆ S which are invariant under the action of G.
The difficulty of this problem is inversely correlated with the size of the group G: when G = {1} is the
trivial group we get all the ideals in S, while for G = GLN (C) the only invariant ideals are the powers md ,
d ≥ 0, of the maximal homogeneous ideal m = (X1 , · · · , XN ). An important intermediate case arises by
taking G = (C∗ )N , thought of as the subgroup of diagonal matrices in GLN (C), in which case the invariant
ideals are precisely the ones generated by monomials.
In this article we are concerned with the situation when N = m · n for some positive integers m ≥ n,
in which case we write S = C[X11 , · · · , Xmn ], we think of the variables as the entries of the generic matrix
X = (Xij ), and we consider G = GLm (C) × GLn (C) acting via row and column operations on X. In this
case Problem 1.1 has been completely resolved in [DCEP80], and we recall its solution in Section 3.
Once the classification Problem 1.1 is resolved, or perhaps after we restrict our focus to a subclass of
G-invariant ideals that is of interest, it is natural to study homological invariants associated with these
ideals, and understand the symmetries that these invariants inherit from the action of G. Some of the most
fundamental homological invariants are listed in the following problem.
Problem 1.2. Given a G-invariant ideal I ⊂ S, describe the G-module structure of
• The syzygy modules TorS• (I, C).
• The Ext modules Ext•S (S/I, S).
• The local cohomology modules HI• (S).
Knowledge of the basic invariants listed above allows us to compute further invariants such as depth
and projective dimension, Castelnuovo–Mumford regularity, Hilbert polynomials etc. In the case when
G = GLm (C) × GLn (C) the Ext modules have been computed in [Rai16b], while the local cohomology
modules have been described in [RW14]. As far as syzygy modules are concerned, Problem 1.2 is still
Date: December 29, 2017.
2010 Mathematics Subject Classification. Primary 13D07, 14M12, 13D45.
Key words and phrases. Determinantal varieties, local cohomology, syzygies, Ext and Tor groups, thickenings.
1
2
CLAUDIU RAICU
unresolved, the only class of examples for which it has been answered being the primary G-invariant ideals
which are generated by a single irreducible representation of G [RW17].
Despite the elementary nature of Problems 1.1 and 1.2, their resolution depends on some deep and
beautiful (both commutative and noncommutative) mathematics. For spaces of matrices the representation
theory of the general linear group plays an essential role at every stage. In addition to that, the classification
problem is resolved using the theory of algebras with straightening law. Moreover, as far as the homological
invariants are concerned:
• The calculation of Ext modules uses Grothendieck duality, as well as the Borel–Weil–Bott theorem
describing the cohomology of simple homogeneous vector bundles on Grassmannian varieties.
• The local cohomology modules are best understood through their D-module structure, i.e. their
structure as modules over the Weyl algebra of differential operators.
• The structure of the syzygy modules depends heavily on the representation theory of the general
linear Lie superalgebras.
Without going into the details of the tools involved in the proofs, we summarize the existing results along
with a number of examples that the reader is encouraged to work out in detail. The organization of the
paper is as follows. In Section 2 we explain the solution to Problems 1.1 and 1.2 in the case when the group
G is the full group of linear automorphisms of S, which is equivalent to considering the action by (row
and) column operations on matrices with one row, and as such is a special case of the analysis performed
in the remaining part of the paper. In Section 3 we recall the solution to Problem 1.1 in the case when
G = GLm (C) × GLn (C) acts on S = C[Xij ], following [DCEP80]. In Sections 4, 5, 6 we study Problem 1.2
for each of the three fundamental homological invariants, following [Rai16b, RW14, RWW14, RW17]. We end
with Section 7 where we give a short list of open problems.
2. Ideals invariant under all coordinate changes
Let S = C[X1 , · · · , XN ] and consider the action of G = GLN (C) by linear changes of coordinates. If we
regard X1 , · · · , XN as the entries of the generic N × 1 matrix, then the results of this section are special
cases (m = N and n = 1) of the results in the rest of the paper. They should be regarded as a warm-up for
things to come. If we write
m = hX1 , · · · , XN i
for the maximal homogeneous ideal of S then the answer to Problem 1.1 is given by the following.
Theorem 2.1. The GLN (C)-invariant ideals in S are md , for d ≥ 0.
Proof. The polynomial ring S decomposes as a direct sum of irreducible GLN (C)-representations
M
S=
Symd (CN ),
d≥0
and md is precisely the ideal generated by Symd (CN ), which is therefore invariant. Any GLN (C)-invariant
ideal is generated by a (finite dimensional) subrepresentation of S, which necessarily has the form
Symd1 (CN ) ⊕ · · · ⊕ Symdr (CN ).
This means that I = md1 + · · · + mdr , which implies I = md for d = min{d1 , · · · , dr }.
Having classified the invariant ideals, we now turn to the homological invariants from Problem 1.2. It is
useful to note (see [Eis95, Exercise A.2.17]) that md (d > 0) is the ideal generated by the maximal minors
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
of the (d + N − 1) × d matrix
X1 X2 X3 · · · XN
0
0 X1 X2 · · · XN −1 XN
.. ..
.
.
..
..
.
.
0 ··· ··· 0
X1
X2
3
0 ··· 0
0 ··· 0
..
.
..
.
0
· · · · · · XN
Since the ideal md has grade N = (N + d − 1) − d + 1, we can apply [Eis05, Theorem A.2.60] to conclude
that it has a linear resolution given by the Eagon–Northcott complex. To describe the GLN (C)-structure of
the syzygy modules, we first introduce some notation.
A partition x = (x1 , x2 , · · · ) is a finite collection of non-negative integers, with x1 ≥ x2 ≥ · · · . We call
each xi a part of x, and define the size of x to be |x| = x1 + x2 + · · · . Most of the time we suppress the
parts of size zero from the notation, for instance the partitions (4, 2, 1, 0, 0) and (4, 2, 1) are considered to be
the same; their size is 7 = 4 + 2 + 1. When x has repeated parts, we often use the abbreviation (ba ) for the
sequence (b, b, · · · , b) of length a. For instance, (4, 4, 4, 3, 3, 3, 3, 3, 2, 1) would be abbreviated as (43 , 35 , 2, 1).
We denote by Pn the collection of partitions with at most n non-zero parts. It is often convenient to identify
a partition x with the associated Young diagram:
x = (4, 2, 1, 0, 0)
←→
Of particular interest in this section are the hook partitions, i.e. those x for which x2 ≤ 1, or equivalently
x = (a, 1b ) for some a, b ≥ 0. The name is illustrative of the shape of the associated Young diagram, as seen
for instance in the following example of a hook partition:
←→
x = (4, 1, 1)
Each partition x ∈ PN determines an irreducible GLN (C)-representation, denoted Sx CN (one usually
refers to Sx as the Schur functor associated to the partition x): our conventions are such that when x = (d)
V
we have that Sx CN = Symd CN is a symmetric power, while for x = (1k ) we have that Sx CN = k CN is an
exterior power. The syzygies of powers of the maximal homogeneous ideals can then be described in terms
of hook partitions as follows (unless specified, all tensor products are considered over the base field C).
Theorem 2.2 ([Gre84, (1.a.10)], [BE75, Cor. 3.2]). We have GLN (C)-equivariant isomorphisms
TorSp (md , C)d+p ≃ Sd,1p CN , for p = 0, · · · , N − 1,
and TorSp (md , C)q = 0 for all other values of p, q. Equivalently, md has a linear minimal free resolution
0 ←− md ←− Sd CN ⊗ S(−d) ←− · · · ←− Sd,1p CN ⊗ S(−d − p) ←− · · · ←− Sd,1N−1 CN ⊗ S(−d − N + 1) ←− 0
In order to describe the graded components of Ext and local cohomology modules, we need to consider a
generalization of partitions where we allow some parts to be negative: we define a dominant weight to be an
element λ ∈ ZN with λ1 ≥ · · · ≥ λN . The associated Schur functor Sλ has the property that
Sλ+(1N ) CN ≃ Sλ CN ⊗
N
^
CN for all dominant weights λ,
4
CLAUDIU RAICU
and in particular
S(−1N ) CN ≃
N
^
CN
!∨
= HomC
N
^
!
CN , C .
More generally, if we let λ∨ = (−λN , −λN −1 , · · · , −λ1 ) then we obtain a GLN (C)-equivariant isomorphism
(2.1)
Sλ∨ CN ≃ HomC Sλ CN , C .
With the above conventions, the GLN (C)-equivariant structure on Ext modules is given by the following.
Theorem 2.3. For d ≥ 0 we have a graded GLN (C)-equivariant isomorphism
d
ExtN
S (S/m , S)
≃
d−1
M
S(−1N−1 ,−i−1) CN ,
(2.2)
i=0
where S(−1N−1 ,−i−1) CN is placed in degree −i − N . Moreover, for j 6= N we have ExtjS (S/md , S) = 0.
Proof. We note that A = S/md is a graded Artinian C-algebra which is a quotient of the polynomial ring,
so duality theory (see [Eis95, Chapter 21] or [BH93, Section I.3]) provides canonical isomorphisms
HomC (A, C) ≃ ωA ≃ ExtN
S (A, ωS ),
where ω denotes the canonical module. Using the fact that
ωS ≃ S(1N ) CN ⊗ S(−N )
and A =
Ld−1
i=0
Si CN , which in turn implies using (2.1)
HomC (A, C) ≃
d−1
M
S(0N−1 ,−i) CN ,
i=0
we obtain
N
ExtN
S (A, S) ≃ HomC (A, C) ⊗ S(−1N ) C ≃
d−1
M
S(−1N−1 ,−i−1) CN
i=0
where S(−1N )
components.
CN
is placed in degree −N , which determines the desired grading on the remaining homogeneous
Using the fact that local cohomology is a direct limit of Ext modules (see [Eis95, Appendix 4])
HIj (S) = lim ExtjS (S/I d , S)
−→
(2.3)
d
one easily obtains the following.
Theorem 2.4. We have a graded GLN (C)-equivariant isomorphism
M
HmN (S) ≃
S(−1N−1 ,−i−1) CN ,
i≥0
where S(−1N−1 ,−i−1) CN is placed in degree −i − N . Moreover, for j 6= N we have Hmj (S) = 0.
(2.4)
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
5
Proof. The formula (2.4) follows from Theorem 2.3 and (2.3) once we show that the maps in the directed
system from (2.3) are injections. Using the long exact sequence obtained by applying HomS (•, S) to the
short exact sequence
0 −→ md /md+1 −→ S/md+1 −→ S/md −→ 0
we see that it is sufficient to verify that ExtSN −1 (md /md+1 , S) = 0. Since md /md+1 is annihilated by m we
get in fact that ExtjS (md /md+1 , S) = 0 for all j < N , concluding the proof.
An alternative way of computing local cohomology is using the Čech complex, which yields the more
familiar isomorphism
SX1 ···XN
.
HmN (S) ≃ PN
i=1 SX1 ···X̂i ···XN
Nevertheless, the GLN (C)-action is harder to see from this description, since the Čech complex is not
GLN (C)-equivariant. The advantage of this representation is that it manifestly expresses HmN (S) as a
module over the Weyl algebra D = Sh∂1 , · · · , ∂N i, where ∂i = ∂/∂Xi is the partial derivative with respect
to Xi . Even more remarkable, the D-module HmN (S) is simple, i.e. it has no non-trivial submodules, despite
the fact that as an S-module it is not even finitely generated! Other ways of regarding HmN (S) are as the
injective hull of the residue field C = S/m, or as the graded dual of the polynomial ring S, but they will not
concern us any further.
3. The classification of GL-invariant ideals
Consider positive integers m ≥ n ≥ 1 and let S = C[Xij ], where i = 1, · · · , m and j = 1, · · · , n. We
identify S with the symmetric algebra Sym(Cm ⊗ Cn ) and consider the group GL = GLm (C) × GLn (C)
together with its natural action on S. The goal of this section is to recall from [DCEP80] the classification
of GL-invariant ideals I ⊆ S, which will be used throughout the rest of the paper.
Given our notation and conventions regarding partitions and the associated Young diagrams (see Section 2), whenever we refer to a row/column of a partition x, we always mean a row/column of the associated
Young diagram. Given a partition x, we can then construct the conjugate partition x′ by transposing the
associated Young diagrams: x′i counts the number of boxes in the i-th column of x, e.g. (4, 2, 1)′ = (3, 2, 1, 1).
We define for each l = 1, · · · , n the polynomial
detl = det(Xij )1≤i,j≤l .
For x ∈ Pn we define
detx =
x1
Y
detx′i ,
(3.1)
(3.2)
i=1
and let
Ix = hGL · detx i
be the ideal generated by the GL-orbit of the polynomial detx . More generally, if X ⊆ Pn we let
X
Ix .
IX =
(3.3)
(3.4)
x∈X
Example 3.1. Consider x = (1p ) for some p ≤ n. Since x has a single column, it follows from (3.2) that
detx = detp = det(Xij )1≤i,j≤p .
Using the multilinear property of the determinant, it is easy to check that the C-span of GL · detx is the
same as that of the p × p minors of the generic matrix (Xij ). We write Ip instead of I(1p ) for the ideal
6
CLAUDIU RAICU
generated by these minors, and recall that Ip is a prime ideal corresponding to the affine algebraic variety
of matrices of rank < p.
Example 3.2. When n = 1 (and m = N ) we have that every x ∈ P1 has the form x = (d) for some
non-negative integer d. If we write Xi instead of Xi1 then we get that S = C[X1 , · · · , XN ], det1 = X1 and
detx = X1d .
One can show that the C-span of GL ·X1d coincides with the vector space of homogeneous polynomials of
degree d, i.e. Ix = md using the notation from Section 2. The acute reader might have noticed that here we
are working with the product of groups GLN (C) × GL1 (C) instead of simply GLN (C), but in fact the action
of GL1 (C) is subsumed by that of GLN (C), so the case n = 1 is precisely the one studied in Section 2.
Given two partitions x, y ∈ Pn , we write x ≤ y if xi ≤ yi for all i. We say that x and y are incomparable
if neither x ≤ y nor y ≤ x. It is shown in [DCEP80] that
x ≤ y ⇐⇒ Ix ⊇ Iy ,
(3.5)
which in the case when n = 1 is simply the statement that d ≤ e if and only if md ⊇ me . It follows that when
considering ideals of the form (3.4) there is no harm in assuming that the partitions in X are incomparable.
The following theorem completely answers Problem 1.1 for GL-invariant ideals.
Theorem 3.3 ([DCEP80]). The association X −→ IX establishes a bijective correspondence between
{(finite) subsets X ⊂ Pn consisting of incomparable partitions} ←→ {GL −invariant ideals I ⊆ S}.
Perhaps the most interesting examples of GL-invariant ideals are the powers (usual, symbolic, or saturated)
of the determinantal ideals Ip . Each such power corresponds via Theorem 3.3 to a finite set of partitions,
which can be described as follows. For the usual powers, it is shown in [DCEP80] that Ipd = IXpd where
Xpd = {x ∈ Pn : |x| = p · d, x1 ≤ d}
(3.6)
Example 3.4. Suppose that m = n = 3 (or that m ≥ n = 3) and let p = 2. The following table records
the partitions corresponding to small powers of the ideal of 2 × 2 minors.
d
X2d
1
2
3
4
5
The symbolic and saturated powers of ideals of minors are obtained via saturation as follows. Recall that
the saturation of an ideal I with respect to J is defined via
I : J ∞ = {f ∈ S : f · J r ⊆ I for r ≫ 0}.
(3.7)
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
7
(d)
If we let Ip denote the d-th symbolic power of Ip , and let (Ipd )sat denote the saturation of Ipd with respect
to the maximal homogeneous ideal, then we have
∞
Ip(d) = Ipd : Ip−1
and
(Ipd )sat = Ipd : I1∞ .
It is therefore important to understand how the ideals IX transform under saturation. To do so, we need to
introduce one more piece of notation: given a positive integer c, we write x(c) for the partition defined by
x(c)i = min(xi , c), so that the non-zero columns of x(c) are precisely the first c columns of x. We then have
the following.
Lemma 3.5 ([Rai16b, Lemma 2.3]). For a subset X ⊂ Pn we have IX : Ip∞ = IX :p where
X :p = {x(c) : x ∈ X , c ∈ Z≥0 , x′c > p if c > 0, and x′c+1 ≤ p}
(3.8)
Thinking more concretely in terms of the corresponding Young diagrams, (3.8) asserts that the partitions
in X :p are obtained from those of X by eliminating columns of size ≤ p. Based on this observation, one can
(d)
show that Ip = IX (d) where
p
Xp(d) = {x ∈ Pn : x1 = · · · = xp , xp + xp+1 + · · · + xn = d}
(3.9)
Furthermore, we have that (Ipd )sat = I(Xpd )sat where
(Xpd )sat = {x ∈ Pn : x1 = x2 , x2 + x3 + · · · + xn = (p − 1) · d}
(3.10)
(d)
It is important to note that both Xp and (Xpd )sat consist of incomparable partitions, whereas the set X :p
defined in (3.8) may contain comparable partitions even if X does not (see the example below)!
Example 3.6. Continuing with the notation in Example 3.4 we get the following table recording the
partitions corresponding to small symbolic powers of the ideal of 2 × 2 minors (note that for 2 × 2 minors
the symbolic powers coincide with the saturated powers).
d
(X2d ):1
(d)
X2
= (X2d )sat
1
2
3
4
5
(2)
To put into words some of the information in this table, consider for instance X2 = {(1, 1, 1), (2, 2)}. Recall
(2)
that I2 is the defining ideal of the Segre variety Z consisting of matrices of rank at most 1, and that I2
is the ideal of functions vanishing to order at least two along Z (see [Eis95, Section 3.9]). The fact that
(2)
I2 = IX (2) can then be interpreted as follows:
2
8
CLAUDIU RAICU
• The function f1 = det3 = det(Xij )1≤i,j≤3 vanishes to order two along Z, as well as every other
function in its GL-orbit.
• The function f2 = det22 = (det(Xij )1≤i,j≤2 )2 vanishes to order two along Z, as well as every other
function in its GL-orbit.
• Every function vanishing to order two along Z is contained in the ideal generated by the GL-orbits
of f1 and f2 .
Example 3.7. For an example where usual, symbolic and saturated powers are all distinct, consider the
case when m = n = 4, p = 3, and d = 3. We have
,
,
X33 =
(X33 )sat =
(3)
X3
=
,
,
,
(3)
and note that (based on (3.5)) we get strict inclusions I33 ( (I33 )sat ( I3 .
4. Ext modules
Having described the solution to the classification Problem 1.1 for GL-invariant ideals in S = C[Xij ],
we now turn to the analysis of the basic homological invariants. In this section we explain the solution to
Problem 1.2 for the Ext groups following [Rai16b]. The main result asserts that for the purpose of computing
Ext•S (S/I, S) it is enough to replace S/I with the associated graded gr(S/I) for a natural finite filtration of
S/I where the quotients have explicitly computable Ext modules. In more technical terms, the main result
is about the degeneration of the spectral sequence for computing Ext, associated to the said filtration of
S/I, but as far as we understand it this degeneration occurs for highly non-trivial reasons. For instance the
proofs that we give in [Rai16b] require an analysis of all (or at least a large class) of GL-invariant ideals,
and would not be applicable on a case to case basis: for example we do not know how to compute directly
Ext•S (S/I, S) when I = Ipd is a power of a determinantal ideal, without doing so for arbitrary (or at least
sufficiently general) GL-invariant ideals I.
Theorem 4.1. To any GL-invariant ideal I ⊆ S we can associate a finite set M(I) of GL-equivariant
S-modules, arising as successive quotients in a natural filtration of S/I, and having the property that for
each j ≥ 0 we have a GL-equivariant degree preserving isomorphism (but not an S-module isomorphism!)
M
ExtjS (S/I, S) ≃
ExtjS (M, S),
M ∈M(I)
The sets M(I) and the modules ExtjS (M, S) for M ∈ M(I) can be computed explicitly. Furthermore, the
association I 7→ M(I) has the property that whenever I ⊇ J are GL-invariant ideals, the (co)kernels and
images of the induced maps ExtjS (S/I, S) −→ ExtjS (S/J, S) can be computed as follows.
M
ker ExtjS (S/I, S) −→ ExtjS (S/J, S) =
ExtjS (M, S),
M ∈M(I)\M(J)
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
Im ExtjS (S/I, S) −→ ExtjS (S/J, S) =
M
9
ExtjS (M, S),
M ∈M(I)∩M(J)
coker ExtjS (S/I, S) −→ ExtjS (S/J, S) =
M
ExtjS (M, S).
M ∈M(J)\M(I)
Finally, the Ext modules get smaller under saturation in a very precise sense: we have that
M(I : Ip∞ ) = {M ∈ M(I) : Ann(M ) 6⊆ Ip }
(4.1)
The equivariant modules M appearing in the sets M(I) are indexed by pairs (z, l) where z ∈ Pn is a
partition and l is a non-negative integer (we write M = Jz,l for the module corresponding to the pair (z, l)).
We think of z as combinatorial data, and of l as geometric data: l indicates the fact that the scheme theoretic
support of Jz,l is precisely the variety of matrices of rank ≤ l, or equivalently it says that the annihilator of
Jz,l is the ideal Il+1 . To define Jz,l as a quotient of ideals we consider
succ(z, l) = {x ∈ Pn : x ≥ z and xi > zi for some i > l}.
(4.2)
and define (using notation (3.3) and (3.4))
Jz,l = Iz /Isucc(z,l) .
(4.3)
Example 4.2. When z = (0) is the empty partition we get that Iz = S and Isucc(z,l) = Il+1 so that
J(0),l = S/Il+1 .
Example 4.3. When l = 0 we have that Isucc(z,0) = Iz · I1 and Jz,0 is identified with the vector space
of minimal generators of Iz (since I1 = m is the maximal homogeneous ideal of S). The vector space of
minimal generators of Iz is the C-span of the GL-orbit GL · detz and it is isomorphic to the irreducible
GL-representation Sz Cm ⊗ Sz Cn . If we further assume that n = 1 (and m = N ) then z = (d) for some d ≥ 0
and Jz,0 = md /md+1 ≃ Symd CN .
To explain which of the modules M = Jz,l appear in M(I) we need the following key definition. Recall
that x′ denotes the conjugate partition to x, and that x(c) is the partition obtained from the first c columns
of x, namely x(c)i = min(xi , c) for all i.
Definition 4.4. For X ⊂ Pn a finite subset we define Z(X ) to be the set consisting of pairs (z, l) where
z ∈ Pn and l ≥ 0 are such that if we write c = z1 then the following hold:
(1) There exists a partition x ∈ X such that x(c) ≤ z and x′c+1 ≤ l + 1.
(2) For every partition x ∈ X satisfying (1) we have x′c+1 = l + 1.
With this definition, we can make explicit the sets M(I) in Theorem 4.1: if I = IX for X ⊂ Pn then
M(IX ) = {Jz,l : (z, l) ∈ Z(X )}.
(4.4)
Notice that Definition 4.4 does not require that the set X consist of incomparable partitions, so implicit in
equation (4.4) is the fact that Z(X ) = Z(X ′ ) whenever IX = IX ′ , i.e. whenever X and X ′ have the same
set of minimal partitions (with respect to ≤).
Example 4.5. If X = {(1l+1 )} (so that IX = Il+1 ) then one can check using Definition 4.4 that
Z(X ) = {((0), l)}
and therefore M(Il+1 ) = {S/Il+1 } consists of a single module.
10
CLAUDIU RAICU
Example 4.6. Suppose that n = 1 and let X = {(d)}, so that IX = md . We have that
Z(X ) = {((i), 0) : i = 0, · · · , d − 1}
and using the calculation from Example 4.3 we get
M(md ) = {mi /mi+1 : i = 0, · · · , d − 1}.
The conclusions of Theorem 4.1 are then easy to verify in this situation (see also Theorem 2.3).
Just as we did in Section 3, we would like to understand better the powers (usual, symbolic, saturated)
(d)
of the determinantal ideals. To do so we need to describe the sets Z(Xpd ), Z(Xp ), and Z((Xpd )sat ). In view
of Lemma 3.5 and (4.4), we can rewrite (4.1) more explicitly as
Z(X :p ) = {(z, l) ∈ Z(X ) : l ≥ p} ⊆ Z(X ),
(4.5)
(d)
so knowing Z(Xpd ) immediately determines Z(Xp ) and Z((Xpd )sat ). We have using [Rai16b, Lemma 5.3]
0 ≤ l ≤ p − 1, z ∈ Pn , z1 = · · · = zl+1 ≤ d − 1,
d
Z(Xp ) = (z, l) :
(4.6)
|z| + (d − z1 ) · l + 1 ≤ p · d ≤ |z| + (d − z1 ) · (l + 1)
which in turn based on (4.5) implies
Z((Xpd )sat ) = {(z, l) ∈ Z(Xpd ) : l ≥ 1}
and
Z(Xp(d) ) = {(z, l) ∈ Z(Xpd ) : l ≥ p − 1}
= {(z, p − 1) : z ∈ Pn , z1 = · · · = zp , zp + zp+1 + · · · + zn ≤ d − 1}
(4.7)
(4.8)
Example 4.7. We continue with the situation from Example 3.7: m = n = 4 and p = d = 3. One can
check either directly from Definition 4.4, or based on (4.6–4.8) that (if we write ∅ for the Young diagram of
the (0) partition)
Z(X33 ) =
,2 ,
,2 ,
, 2 , (∅, 2) ,
, 1 ,
, 0
,2 ,
,2 ,
, 2 , (∅, 2) ,
, 1
Z((X33 )sat ) =
(3)
Z(X3 ) =
,2 ,
,2 ,
, 2 , (∅, 2)
(d)
It is worthwhile to observe that the sets Z(Xp ) in (4.8) get larger as d grows, which in view of (4.4) and
Theorem 4.1 implies that for every d ≥ 1 and every j ≥ 0 the induced maps
ExtjS (S/Ip(d−1) , S) −→ ExtjS (S/Ip(d) , S)
(4.9)
are injective. This is certainly not the case if we replace symbolic powers with the usual powers, as seen for
instance in the next example.
Example 4.8. Assume that we are in the situation from Examples 3.4 and 3.6: m = n = 3 and p = 2. Since
(d−1)
(d)
(d)
Z(Xp
) ⊆ Z(Xp ) and Z(Xp ) ⊆ Z(Xpd ), we will record for the sake of compactness only the difference
between these sets in the following table.
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
(d)
1
(d)
Z(X2d ) \ Z(X2 )
)
(∅,1)
2
3
4
5
(d−1)
Z(X2 ) \ Z(X2
d
,1
,1
–
,1
,1
11
,1
,1
,1
,1
,0
,0
,0
,0
,0
,0
,0
,0
,0
The last piece of mystery in Theorem 4.1 is the explicit calculation of Ext•S (M, S) when M = Jz,l . This
is the content of the following (slightly weaker) version of [Rai16b, Thm 2.5] and [RW14, Thm 3.3].
Theorem 4.9. Fix an integer 0 ≤ l < n and assume that z ∈ Pn is a partition with z1 = z2 = · · · = zl = zl+1 .
For 0 ≤ s ≤ t1 ≤ · · · ≤ tn−l ≤ l we consider the set W (z, l; t, s) of dominant weights λ ∈ Zn satisfying
λn = l − zl − m,
(4.10)
λti +i = ti − zn+1−i − m for i = 1, · · · , n − l,
λs ≥ s − n and λs+1 ≤ s − m.
Letting λ(s) = (λ1 , · · · , λs , (s − n)m−n , λs+1 + (m − n), · · · , λn + (m − n)) ∈ Zm , we have
M
Sλ(s) Cm ⊗ Sλ Cn ,
ExtjS (Jz,l , S) =
(4.11)
0≤s≤t1 ≤···≤tn−l ≤l
P
m·n−l2 −s·(m−n)−2·( n−l
i=1 ti )=j
λ∈W (z,l;t,s)
where Sλ(s) Cm ⊗ Sλ Cn appears in degree |λ| = λ1 + · · · + λn .
The formulas in Theorem 4.9 are quite involved, but nevertheless they are completely explicit. In particular
they allow us to determine which graded components of the modules ExtjS (Jz,l , S) are non-zero, and as
a consequence determine a formula for the Castelnuovo–Mumford regularity of Jz,l . Combining this with
Theorem 4.1, we get formulas for reg(S/I) for an arbitrary GL-invariant ideal I. Unfortunately these are not
closed formulas, but rather they involve an often difficult linear integer optimization problem (see [Rai16b,
Theorem 2.6] and [Rai16b, Section 4]). Here we will content ourselves with showing that Theorem 4.9
combined with Theorem 4.1 recovers the description of the Ext modules in Theorem 2.3. For another
example of the concrete calculation of Ext modules see [Rai16b, Section 7].
Example 4.10. Assume that n = 1, m = N , and write Xi = Xi1 so that S = C[X1 , · · · , XN ]. Consider
X = {(d)} for some d ≥ 0, so that IX = md . We have seen in Example 4.6 that Z(X ) = {((i), 0) : i =
0, · · · , d − 1} and in Example 4.3 that J(i),0 = mi /mi+1 . Theorem 4.1 implies that
ExtjS (S/md , S) ≃
d−1
M
i=0
ExtjS (mi /mi+1 , S) for all j,
(4.12)
12
CLAUDIU RAICU
so it is enough to compute ExtjS (mi /mi+1 , S) based on Theorem 4.9. Fix (z, l) = ((i), 0) for some 0 ≤ i ≤ d−1
and observe that since l = 0, (4.11) forces s = t1 = 0 and therefore the only potentially non-zero Ext module
occurs for
!
n−l
X
ti = N.
j = m · n − l2 − s · (m − n) − 2 ·
i=1
Moreover, we get that W (z, l; t, s) = W ((i), 0; (0), 0) consist of a single dominant weight λ ∈ Z1 , namely
λ = (−i − N ), and for that weight we have
λ(s) = λ(0) = (−1N −1 , −i − N + N − 1) = (−1N −1 , −i − 1).
This shows that
i
i+1
ExtN
, S) ≃ S(−1N−1 ,−i−1) CN ⊗ S−i−N C1 ≃ S(−1N−1 ,−i−1) CN
S (m /m
(4.13)
where the last isomorphism simply disregards the GL1 (C)-action on the second factor and is only GLN (C)equivariant. Combining (4.12) with (4.13) yields the conclusion of Theorem 2.3.
5. Local cohomology modules
Having computed the Ext modules in the previous section, as well as the induced maps between them, the
description of local cohomology is a consequence of (2.3). We begin by recalling that the local cohomology
modules HI• (S) depend on I only up to radical. Moreover, any GL-invariant radical ideal I ⊆ S = C[Xij ]
corresponds to a GL-invariant algebraic subset of the space Cm×n of m × n matrices; since any such algebraic
set is the set of matrices of rank < p for some value of p, it follows that every GL-invariant radical ideal is
of the form Ip for some p. It is then sufficient to study
HIjp (S) = lim ExtjS (S/Ipd , S) for j ≥ 0.
−→
d
Based on Theorem 4.1, it follows that
HIjp (S) =
M
ExtjS (M, S),
(5.1)
M ∈Mp
S
where Mp ⊆ d M(Ipd ) is the subset consisting of those M for which M ∈ M(Ipd ) for all d ≫ 0. We
encourage the reader to verify directly that
[
(4.8)
Mp =
M(Ip(d) ) = {Jz,p−1 : z ∈ Pn , and z1 = · · · = zp },
d
which has an alternative explanation as follows. One can tweak the formula (2.3) and get
HIj (S) = lim ExtjS (S/Jd , S)
−→
d
(d)
where (Jd )d is any sequence of ideals which is cofinal with the sequence of powers (I d )d . If we take Jd = Ip
and use the injectivity of the maps (4.9) we get that
M
HIjp (S) =
ExtjS (M, S),
S
(d)
M ∈ d M(Ip )
as desired. In [RW14] we have used yet another sequence of ideals to compute local cohomology, namely
Jd = Id×p , where x = d × p denotes the partition whose Young diagram is the d × p rectangle, i.e. x1 =
· · · = xp = d and xi = 0 for i > p. Coincidentally, the ideals Id×p are precisely the ones for which we can
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
13
compute all the syzygy modules, as explained in the next section. Just as for symbolic powers, it is the case
that the sets M(Id×p ) increase with d, and their union is precisely the set Mp .
To give a cleaner description of the modules HI•p (S) we introduce some notation. We consider the generalized binomial coefficients, also known as q-binomial coefficients or Gauss polynomials, to be the following
polynomials in the indeterminate q, depending on non-negative integers a, b:
a
(1 − q a )(1 − q a−1 ) · · · (1 − q a−b+1 )
.
(5.2)
=
(1 − q b )(1 − q b−1 ) · · · (1 − q)
b q
If we set q = 1 then we recover the usual binomial coefficients:
a
a
a!
=
=
.
b 1
b
b! · (a − b)!
As another example, if we let a = 4 and b = 2 then
4
= 1 + q + 2q 2 + q 3 + q 4 .
2 q
P
L
In what follows we consider formal linear combinations j Aj · q j , where Aj = i∈Z Aji is a graded GLrepresentation. We will interpret an equality of the form
X
X
Aj · q j =
B j · qj
j
j
to mean that we have graded GL-equivariant isomorphisms Aj ≃ B j for all j, i.e. that Aji ≃ Bij for every i
and j. With these conventions, the formula (5.1) can then be written more explicitly as follows.
Theorem 5.1 ([RW14, Thm. 6.1], [RW16, Main Theorem(1)]). If we think of HIjp (S) as a graded GLrepresentation (with the natural grading inherited from that on S) then we have an equality
p−1
X
X j
n−s−1
(n−p+1)2 +(n−s)·(m−n)
j
Ds · q
·
(5.3)
HIp (S) · q =
p − 1 − s q2
j≥0
s=0
where Ds is a graded GL-representation which decomposes as
M
Sλ(s) Cm ⊗ Sλ Cn
Ds =
λ=(λ1 ≥···≥λn )∈Zn
λs ≥s−n
λs+1 ≤s−m
with Sλ(s) Cm ⊗ Sλ Cn living in degree |λ| = λ1 + · · · + λn .
The formula (5.3) is just a shadow of the deeper D-module structure of the local cohomology modules.
More precisely, the graded representations Ds in Theorem 5.1 are the underlying vector spaces of the
simple GL-equivariant D-modules on Cm×n , where Ds corresponds to the module supported on rank ≤ s
matrices [Rai16a]. The local cohomology modules have finite length when regarded as D-modules, and the
multiplicities of the composition factors Ds in their Jordan–Hölder filtrations can be read off by equating
the two sides of (5.3).
Example 5.2. Consider n = 1 (and m = N ), and take p = 1 (so that Ip = m). The equation (5.3) becomes
X
Hmj (S) · q j = D0 · q N .
j≥0
14
CLAUDIU RAICU
This means that Hmj (S) = 0 for j < N , and HmN (S) ≃ D0 as graded representations. Note that the formula
for D0 in Theorem 5.1 specializes to (2.4).
Example 5.3. For a slightly bigger examples which shows that multiplicities bigger than 1 typically occur
in local cohomology modules, consider m = n = 5 and p = 3. We get
!
X j
4
3
2
HI3 (S) · q j = q 9 · D0 ·
+ D1 ·
+ D2 ·
2 q
1 q
0 q
j≥0
= q 9 · (D0 + D1 + D2 ) + q 10 · (D0 + D1 ) + q 11 · (2D0 + D1 ) + q 12 · D0 + q 13 · D0 .
This formula shows that HIj3 (S) = 0 for j < 9 or j > 13. It also says for instance that HI93 (S) ≃ D0 ⊕D1 ⊕D2
as graded GL-representations, or that the D-module HI93 (S) has length three, with composition factors
D0 , D1 , D2 , each occurring with multiplicity one. The D-module HI11
(S) also has length three, but only two
3
distinct composition factors: D0 occurring with multiplicity two, and D1 with multiplicity one. Forgetting
the D-module structure we have an isomorphism of graded GL-representations HI11
(S) ≃ D0⊕2 ⊕ D1 .
3
We close the section by remarking that analogous versions of Theorem 5.1 hold for ideals of minors of
a generic symmetric matrix, and ideals of Pfaffians of a generic skew-symmetric matrix (see [RW16] and
[RWW14]).
6. Syzygy modules
We keep the notation from the previous sections. Given a GL-invariant ideal I ⊂ S = C[Xij ], we are
interested in studying the vector spaces of i-syzygies of I
Bi (I) = TorSi (I, C)
(6.1)
which are naturally graded GL-representations. We encode these syzygies into the equivariant Betti polynomial
X
Bi (I) · q i ,
(6.2)
BI (q) =
i∈Z
where q is an indeterminate, and the coefficients of BI (q) are graded GL-representations (and they are finite
dimensional, unlike in the previous section when we studied local cohomology modules). In the case when
I = Ip , the problem of describing the syzygies of determinantal ideals was solved in [Las78] (see also [PW85]
and [Wey03, Chapter 6]). For the powers of the ideal of maximal minors (I = Ind ) a description of the syzygy
modules was given in [ABW81]. Perhaps the most well-known result in this area refers to the case I = In
(which is a special case of both of the results quoted above), where the syzygies of the ideal of maximal
minors are given by the Eagon–Northcott resolution [Eis95, Appendix A.2.6].
A description of the equivariant Betti polynomials in (6.2) is still unknown for an arbitrary GL-invariant
ideal I. It is known however for an important class ideals which we describe next. Consider positive integers
a, b with a ≤ n and consider the partition x = a × b defined by
x1 = · · · = xa = b, xi = 0 for i > a.
It is the ideals Ia×b (see (3.3) for the general definition of the ideals Ix ) for which we will be able to
describe the syzygy modules in Theorem 6.1 below. One quick way to define Ia×b is as the smallest GLinvariant ideal which contains the b-th powers of the a × a minors of the generic matrix (Xij ). We note that
Ip×1 = Ip and that In×d = Ind , so that Theorem 6.1 will recover as special cases the aforementioned results
of [Las78, ABW81].
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
15
Before stating the main result we introduce some notation. If r, s are positive integers, α is a partition
with at most r parts (i.e. αi = 0 for i > r) and β is a partition with parts of size at most s (i.e. β1 ≤ s), we
construct the partition
λ(r, s; α, β) = (s + α1 , · · · , s + αr , β1 , β2 , · · · ).
(6.3)
This is easiest to visualize in terms of Young diagrams: one starts with an r × s rectangle, and attach α to
the right and β to the bottom of the rectangle. If r = 4, s = 5, α = (4, 2, 1), β = (3, 2), then
α α α α
α α
α
λ(r, s; α, β) = (9, 7, 6, 5, 3, 2) ←→
(6.4)
β β β
β β
Recall that µ′ denotes the conjugate partition to µ and consider the polynomials hr×s (q) given by
X
hr×s (q) =
(Sλ(r,s;α,β) Cm ⊗ Sλ(r,s;β ′ ,α′ ) Cn ) · q |α|+|β| ,
(6.5)
α,β
where the sum is taken over partitions α, β such that α is contained in the min(r, s) × (n − r) rectangle (i.e.
α1 ≤ n − r, α′1 ≤ min(r, s)) and β is contained in the (m − r) × min(r, s) rectangle (i.e. β1 ≤ min(r, s) and
β1′ ≤ m − r), and the representation Sλ(r,s;α,β) Cm ⊗ Sλ(r,s;β ′ ,α′ ) Cn is placed in degree
|λ(r, s; α, β)| = |λ(r, s; β ′ , α′ )| = r · s + |α| + |β|.
With this notation we have the following.
Theorem 6.1 ([RW17, Thm. 3.1]). The equivariant Betti polynomial of the ideal Ia×b is
n−a
X
t + min(a, b) − 1
t2 +2t
h(a+t)×(b+t) (q) · q
·
BIa×b (q) =
t
q2
t=0
Example 6.2. Suppose that n = 1 (and m = N ) and let a = 1, b = d (so that Ia×b = md ). We get that
Bmd (q) = h1×d (q) =
N
−1
X
(Sd,1p CN ⊗ Sd+p C1 ) · q p
p=0
which, if we forget the GL1 (C)-action on the second factor, recovers the conclusion of Theorem 2.2.
Example 6.3. Suppose that m = n = 3 and let a = b = 2. We get that
2
3
BI2×2 (q) = h2×2 (q) + h3×3 (q) · q ·
= h2×2 (q) + q 3 · h3×3 (q) + q 5 · h3×3 (q).
1 q2
Before analyzing the terms further, it is useful to record the Betti table of I2×2 that Macaulay2 [GS] computes
(recall the convention that the Betti number βi,i+j = dimC TorSi (I2×2 , S)i+j is placed in row j, column i):
0 1 2 3 4 5
total: 36 90 84 37 9 1
4:
36 90 84 36 9 1
5:
.
.
.
. . .
6:
.
.
.
1 . .
(6.6)
16
CLAUDIU RAICU
We claim that h2×2 (q) corresponds to the strand [36 90 84 36 9] in the Betti table, q 3 · h3×3 (q) corresponds to the entry 1 in the Betti table in row 6 and column 3, and q 5 · h3×3 (q) corresponds to the entry 1
in the Betti table in row 4 and column 5. Indeed, it follows from (6.5) that
h3×3 (q) = S3,3,3 C3 ⊗ S3,3,3 C3 .
Note that S3,3,3 C3 ⊗ S3,3,3 C3 is a one-dimensional representation of GL, concentrated in degree 9 = 3 + 3 + 3.
It follows from (6.2) that q 3 · h3×3 (q) contributes a one-dimensional subspace to TorS3 (I2×2 , C)9 , so it must
be the case that
TorS3 (I2×2 , C)9 ≃ S3,3,3 C3 ⊗ S3,3,3 C3
since the Betti number β3,9 is equal to 1. Similarly, q 5 · h3×3 (q) contributes to TorS5 (I2×2 , C)9 , so
TorS5 (I2×2 , C)9 ≃ S3,3,3 C3 ⊗ S3,3,3 C3 .
The remaining entries of the Betti table are accounted for by (for compactness we write Sλ instead of Sλ C3 )
h2×2 (q) = S2,2 ⊗ S2,2
+ (S3,2 ⊗ S2,2,1 + S2,2,1 ⊗ S3,2 ) · q
+ (S3,3 ⊗ S2,2,2 + S3,2,1 ⊗ S3,2,1 + S2,2,2 ⊗ S3,3 ) · q 2
+ (S3,3,1 ⊗ S3,2,2 + S3,2,2 ⊗ S3,3,1 ) · q 3
+ (S3,3,2 ⊗ S3,3,2 ) · q 4
We conclude that
TorS0 (I2×2 , C)4 ≃ S2,2 ⊗ S2,2
TorS1 (I2×2 , C)5 ≃ S3,2 ⊗ S2,2,1 ⊕ S2,2,1 ⊗ S3,2
TorS2 (I2×2 , C)6 ≃ S3,3 ⊗ S2,2,2 ⊕ S3,2,1 ⊗ S3,2,1 ⊕ S2,2,2 ⊗ S3,3
TorS3 (I2×2 , C)7 ≃ S3,3,1 ⊗ S3,2,2 ⊕ S3,2,2 ⊗ S3,3,1
TorS4 (I2×2 , C)8 ≃ S3,3,2 ⊗ S3,3,2
To reconcile this with the Betti numbers in (6.6) it suffices to use the following dimension calculations:
λ
(2, 2) (3, 2) (2, 2, 1) (3, 3) (2, 2, 2) (3, 2, 1) (3, 2, 2) (3, 3, 1) (3, 3, 2)
dim(Sλ )
6
15
3
10
1
8
3
6
3
We have for instance
β2,6 (I2×2 ) = dim(TorS2 (I2×2 , C)6 ) = 10 · 1 + 8 · 8 + 1 · 10 = 84.
Just as the graded GL-representations in (5.3) were shadows of a higher structure (they were underlying
representations of simple GL-equivariant D-modules), the polynomials hr×s (q) in (6.5) encode the underlying
GL-representations of certain simple modules over the general linear Lie superalgebra gl(m|n). The key to
determining (6.2) for arbitrary GL-invariant ideals I is perhaps to get a better grasp on the interplay between
the representation theory of gl(m|n) and the syzygies of the GL-invariant ideals.
7. Open questions
As alluded to in the text, Theorem 4.1 in conjunction with Theorem 4.9 gives rise to formulas (which can
be quite involved) for the Castelnuovo–Mumford regularity of arbitrary GL-invariant ideals. If we consider
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
17
high powers (usual, symbolic, saturated) of determinantal ideals then these formulas simplify quite a bit.
We show in [Rai16b] that for d ≥ n − 1 we have
p−1
p−1
·
, and reg(Ip(d) ) = p · d.
reg(Ipd ) = reg((Ipd )sat ) = p · d +
2
2
As far as low powers are concerned, we obtain a closed formula only when p = 2 (besides the cases when
p = 1 and p = n which are easy):
(d)
reg(I2d ) = reg((I2d )sat ) = reg(I2 ) = d + n − 1 for d = 1, · · · , n − 1.
Problem 7.1. Give a closed formula for the regularity of small powers (usual, symbolic, saturated) of
determinantal ideals when 2 < p < n and 1 < d < n − 1.
The local cohomology modules HI•p (S) are finite lenght D-modules, and we explained how formula (5.3)
encodes their composition factors in a Jordan–Hölder filtration. It would be interesting to understand better
how these composition factors fit together to form HI•p (S).
Problem 7.2. Describe the extension data required to build up the local cohomology modules HI•p (S) from
their composition factors.
The D-modules HI•p (S) are not only of finite length, but they are GL-equivariant, regular, and holonomic.
We have learnt from private communication with András Lőrincz that in the case when m > n the category
of such modules is semisimple, and thus Theorem 5.1 yields the decomposition of the modules HI•p (S) as
a direct sum of simples. This conclusion however fails for square matrices (when m = n), as can be seen
for instance by localizing S at the determinant of the generic n × n matrix. Nevertheless, in this case the
category of regular holonomic D-modules has been given a quiver-type description in [BG99], so one would
have to identify the subcategory of GL-equivariant modules and within that to locate the modules HI•p (S).
Given the calculation of Ext modules for GL-invariant ideals, we can determine the regularity and projective dimension of such ideals, so we have a first approximation of the shape of their minimal free resolution.
It would be interesting to carry this further and describe the complete Betti tables.
Problem 7.3. Determine the syzygies TorS• (I, C) of an arbitrary GL-invariant ideal I ⊂ S = C[Xij ].
Following the ideas of [DCEP80], the classification Problem 1.1 was solved for symmetric matrices in
[Abe80] and for skew-symmetric matrices in [ADF80]. It is then interesting to study the homological invariants associated to invariant ideals in these cases.
Problem 7.4. Solve Problem 1.2 for the natural action of G = GLn (C) on the ring of polynomial functions
on n × n symmetric (resp. skew-symmetric) matrices.
In the case of skew-symmetric matrices, the problem of computing Ext modules has been recently resolved
by Michael Perlman [Per17]. The local cohomology modules have been computed for both symmetric and
skew-symmetric matrices in [RW16]. As far as syzygy modules are concerned there is only little progress, but
just as in the case of general m × n matrices it is expected that the structure of syzygy modules is controlled
by the representation theory of certain Lie superalgebras, specifically the periplectic superalgebras [Sam14].
Much of the theory of determinantal rings can be developed over fields of arbitrary characteristic, or
over more general base rings, as it is done for instance in the monograph [BV88]. In contrast, the methods
employed in proving the results surveyed in this article are heavily dependent on the characteristic of the field
being equal to 0, and in fact the structure of the basic homological invariants can change drastically from
characteristic zero to positive characteristic. For instance, since the rings S/Ip are Cohen-Macaulay (see
18
CLAUDIU RAICU
[HE71, DCEP82] or [BV88, Section 12.C]) it follows from [PS73, Prop. III.4.1] that in positive characteristic
the local cohomology groups HI•p (S) are non-zero in only one cohomological degree, which is in stark contrast
with the conclusion of Theorem 5.1. Nevertheless, it is natural to consider the following.
Problem 7.5. Solve Problems 1.1 and 1.2 for matrix spaces over a field of positive characteristic.
Acknowledgements
The author acknowledges the support of the Alfred P. Sloan Foundation, and of the National Science
Foundation Grant No. 1600765.
References
[Abe80] Silvana Abeasis, The GL(V )-invariant ideals in S(S 2 V ), Rend. Mat. (6) 13 (1980), no. 2, 235–262 (Italian, with
English summary). MR602662
[ADF80] S. Abeasis and A. Del Fra, Young diagrams and ideals of Pfaffians, Adv. in Math. 35 (1980), no. 2, 158–178, DOI
10.1016/0001-8708(80)90046-8. MR560133
[ABW81] Kaan Akin, David A. Buchsbaum, and Jerzy Weyman, Resolutions of determinantal ideals: the submaximal minors,
Adv. in Math. 39 (1981), no. 1, 1–30, DOI 10.1016/0001-8708(81)90055-4. MR605350
[BG99] Tom Braden and Mikhail Grinberg, Perverse sheaves on rank stratifications, Duke Math. J. 96 (1999), no. 2, 317–362,
DOI 10.1215/S0012-7094-99-09609-6. MR1666554
[BH93] Winfried Bruns and Jürgen Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics, vol. 39,
Cambridge University Press, Cambridge, 1993. MR1251956
[BV88] Winfried Bruns and Udo Vetter, Determinantal rings, Lecture Notes in Mathematics, vol. 1327, Springer-Verlag,
Berlin, 1988. MR953963
[BE75] David A. Buchsbaum and David Eisenbud, Generic free resolutions and a family of generically perfect ideals, Advances
in Math. 18 (1975), no. 3, 245–301, DOI 10.1016/0001-8708(75)90046-8. MR0396528
[DCEP80] Corrado De Concini, David Eisenbud, and Claudio Procesi, Young diagrams and determinantal varieties, Invent.
Math. 56 (1980), no. 2, 129–165, DOI 10.1007/BF01392548. MR558865
[DCEP82]
, Hodge algebras, Astérisque, vol. 91, Société Mathématique de France, Paris, 1982. With a French summary.
MR680936
[Eis95] David Eisenbud, Commutative algebra, Graduate Texts in Mathematics, vol. 150, Springer-Verlag, New York, 1995.
With a view toward algebraic geometry. MR1322960 (97a:13001)
[Eis05]
, The geometry of syzygies, Graduate Texts in Mathematics, vol. 229, Springer-Verlag, New York, 2005. A
second course in commutative algebra and algebraic geometry. MR2103875
[GS] Daniel R. Grayson and Michael E. Stillman, Macaulay 2, a software system for research in algebraic geometry,
Available at http://www.math.uiuc.edu/Macaulay2/.
[Gre84] Mark L. Green, Koszul cohomology and the geometry of projective varieties. II, J. Differential Geom. 20 (1984), no. 1,
279–289. MR772134
[HE71] M. Hochster and John A. Eagon, Cohen-Macaulay rings, invariant theory, and the generic perfection of determinantal
loci, Amer. J. Math. 93 (1971), 1020–1058, DOI 10.2307/2373744. MR0302643
[Las78] Alain Lascoux, Syzygies des variétés déterminantales, Adv. in Math. 30 (1978), no. 3, 202–237, DOI 10.1016/00018708(78)90037-3 (French). MR520233 (80j:14043)
[Per17] Michael Perlman, Regularity of Pfaffian Thickenings, arXiv 1711.02777 (2017).
[PS73] C. Peskine and L. Szpiro, Dimension projective finie et cohomologie locale. Applications à la démonstration de
conjectures de M. Auslander, H. Bass et A. Grothendieck, Inst. Hautes Études Sci. Publ. Math. 42 (1973), 47–119
(French). MR0374130
[PW85] Piotr Pragacz and Jerzy Weyman, Complexes associated with trace and evaluation. Another approach to Lascoux’s
resolution, Adv. in Math. 57 (1985), no. 2, 163–207, DOI 10.1016/0001-8708(85)90052-0. MR803010
[Rai16a] Claudiu Raicu, Characters of equivariant D-modules on spaces of matrices, Compos. Math. 152 (2016), no. 9, 1935–
1965, DOI 10.1112/S0010437X16007521. MR3568944
[Rai16b]
, Regularity and cohomology of determinantal thickenings, arXiv 1611.00415 (2016). To appear in Proc.
Lond. Math. Soc.
HOMOLOGICAL INVARIANTS OF DETERMINANTAL THICKENINGS
19
[RW14] Claudiu Raicu and Jerzy Weyman, Local cohomology with support in generic determinantal ideals, Algebra & Number
Theory 8 (2014), no. 5, 1231–1257, DOI 10.2140/ant.2014.8.1231. MR3263142
, Local cohomology with support in ideals of symmetric minors and Pfaffians, J. Lond. Math. Soc. (2) 94
[RW16]
(2016), no. 3, 709–725, DOI 10.1112/jlms/jdw056. MR3614925
[RW17]
, The syzygies of some thickenings of determinantal varieties, Proc. Amer. Math. Soc. 145 (2017), no. 1,
49–59, DOI 10.1090/proc/13197. MR3565359
[RWW14] Claudiu Raicu, Jerzy Weyman, and Emily E. Witt, Local cohomology with support in ideals of maximal minors and
sub-maximal Pfaffians, Adv. Math. 250 (2014), 596–610, DOI 10.1016/j.aim.2013.10.005. MR3122178
[Sam14] Steven V Sam, Derived supersymmetries of determinantal varieties, J. Commut. Algebra 6 (2014), no. 2, 261–286,
DOI 10.1216/JCA-2014-6-2-261. MR3249839
[Wey03] Jerzy Weyman, Cohomology of vector bundles and syzygies, Cambridge Tracts in Mathematics, vol. 149, Cambridge
University Press, Cambridge, 2003. MR1988690
Department of Mathematics, University of Notre Dame, 255 Hurley, Notre Dame, IN 46556
Institute of Mathematics “Simion Stoilow” of the Romanian Academy
E-mail address: [email protected]
| 0math.AC
|
A NEW REGISTRATION APPROACH FOR DYNAMIC ANALYSIS OF CALCIUM SIGNALS
IN ORGANS
Peixian Liang1? , Jianxu Chen1? , Pavel A. Brodskiy2 , Qinfeng Wu2
Yejia Zhang3 , Yizhe Zhang1 , Lin Yang1 , Jeremiah J. Zartman2 and Danny Z. Chen1
1
Department of Computer Science and Engineering, University of Notre Dame, USA
Department of Chemical and Biomolecular Engineering, University of Notre Dame, USA
3
Department of Electrical and Computer Engineering, University of California San Diego, USA
Index Terms— Non-rigid registration, Deep neural networks,
Biomedical image segmentation, Calcium imaging
1. INTRODUCTION
Ca2+ is a ubiquitous second messenger in organisms [1]. Quantitatively analyzing spatial-temporal patterns of intercellular calcium
(Ca2+ ) signaling in tissues is important for understanding biological functions. Wing disc pouches of fruit flies are a commonly used
genetic model system of organ development and have recently been
used to study the decoding of Ca2+ signaling in epithelial tissues
[1, 2]. However, wing discs can undergo considerable movements
and deformations during live imaging experiments. Therefore, an
effective automatic image registration approach is needed to align
pouches across image sequences.
Registering tissues that undergo deformations in time-lapse image sequences is a challenging problem. For example, experimental
image data of wing disc pouches at different time points are moving
or deforming due to a general feature of tissue growth, morphogenesis, and due to general movement during the live imaging process.
Furthermore, a time-lapse movie can contain many frames, and a
number of movies are needed to obtain reliable measurements of
Ca2+ activity due to the noisy and stochastic nature of the signals,
which make the processing more complicated and costly. Common
? These authors contribute equally to this work.
Segmentation
stage
Mapping
stage
FCN model
Rigid transformation
Boundary refinement
Non-rigid transformation
Optimal
transformation:
T1 , T2 , ..., Tn
Registration results
Wing disc pouches of fruit flies are a powerful genetic model for
studying physiological intercellular calcium (Ca2+ ) signals for dynamic analysis of cell signaling in organ development and disease
studies. A key to analyzing spatial-temporal patterns of Ca2+ signal waves is to accurately align the pouches across image sequences.
However, pouches in different image frames may exhibit extensive
intensity oscillations due to Ca2+ signaling dynamics, and commonly used multimodal non-rigid registration methods may fail to
achieve satisfactory results. In this paper, we develop a new twophase non-rigid registration approach to register pouches in image
sequences. First, we conduct segmentation of the region of interest. (i.e., pouches) using a deep neural network model. Second, we
use a B-spline based registration to obtain an optimal transformation
and align pouches across the image sequences. Evaluated using both
synthetic data and real pouch data, our method considerably outperforms the state-of-the-art non-rigid registration methods.
Reference
ABSTRACT
Source images
arXiv:1802.00491v1 [cs.CV] 1 Feb 2018
2
Fig. 1. An overview of our proposed registration approach.
approaches suffer considerably due to cumbersome intensity distortions of tissues caused by Ca2+ oscillations. A method for minimizing the error residual between the local phase-coherence representations of two images was proposed to deal with non-homogeneity
in images [3], which relies heavily on structural information. But,
in our problem, the intensity inside the pouches may change a lot,
thus causing such methods to fail. A Markov-Gibbs random field
model with pairwise interaction was used to learn prior appearance
of a given prototype, which makes it possible to align complex images [4]. Incorporation of spatial and geometric information was
proposed to address the limitations of the static local intensity relationship [5]. But, the computational complexity of these methods is
high. In our problem, a single time-lapse movie may have hundreds
of frames, and hundreds of movies are analyzed. Hence, these methods do not work well for our problem. In general, known methods
do not address well the kind of complex intensity distortion in our
images with a reasonable computation cost.
To address the difficulty of spatial intensity distortion along with
various ROI deformations and movements, we propose a new nonrigid registration approach to effectively align pouches at different
time points in image sequences. Our approach consists of two main
stages: (1) a segmentation stage and (2) a mapping stage (see Fig. 1).
The segmentation stage is based on a deep neural network (FCN).
Because the accuracy of the ROI boundary is a key factor influencing the registration results, we apply a graph search algorithm to
refine the segmented pouch boundaries. The mapping stage uses the
segmentation results from the first stage to characterize an optimal
transformation. We first apply a rigid transformation to modulate
the movement of pouches, and then design an object pre-detection
B-spline based non-rigid algorithm to produce the final mapping.
2. METHODOLOGY
In this section, we present a complete pipeline, which takes timelapse image sequences of pouches as input and produces the regis-
Fig. 2. The structure of our FCN model.
tered sequences (see Fig. 1). We first discuss the segmentation stage
(i.e., FCN model and boundary refinement), and then the mapping
stage (i.e., rigid transformation and non-rigid transformation).
2.1. Segmentation Stage
Our registration approach is based on accurate pouch segmentation.
Pouches in our images are commonly surrounded and touched by
extra tissues with similar intensity and texture, and the separation
boundary between pouches and extra tissues is usually of poor visibility. Meanwhile, the noise induced by the live imaging process
makes the segmentation task more challenging. Thus, it is important
for our segmentation algorithm to leverage the morphological and
topological contexts, in order to correctly segment the shape of each
actual pouch, especially its boundary, from the noisy background.
For this, we employ an FCN model to exploit the semantic context
for accurate segmentation, and a graph search algorithm to further
refine the boundaries.
FCN module Recently, deep learning methods have emerged
as powerful image segmentation tools. Fully convolutional networks (FCN) are widely used in general semantic segmentation and
biomedical image segmentation [6, 7].
It is worth mentioning that in our images, the separation boundary between a pouch and other tissues is usually quite subtle (as thin
as 3 to 5 pixels wide) and obscure, while the whole contextual region
(including both the pouch and extra tissues) can be of a relatively
large scale (more than 200 × 200 pixels). Therefore, the FCN model
must fully exploit both the fine details and a very large context. For
this purpose, we carefully design the FCN architecture following the
model in [8] to leverage a large receptive field without sacrificing
model simplicity and neglecting fine details. The exact structure of
our FCN model is depicted in Fig. 2.
Boundary refinement
We observed that the boundaries of the
output pouches from our FCN model may be fuzzy or of irregular shape in difficult cases (see Fig. 4(F)). To improve the boundary accuracy of the segmented pouches, we first apply a Gaussian
smooth filter to reduce the influence of intensity variations inside the
pouches and then employ a graph search algorithm with the node
weights being the negatives of the gradients to further refine the
shape boundaries. This allows the subsequent mapping stage to be
built upon more accurate segmentation and produce better registration results. Details of this process are omitted due to the page limit.
2.2. Mapping Stage
The goal of the registration process is: For every point in the source
image, obtain an optimal corresponding point in the reference image.
A key observation is that the intensity profile of the same pouch may
incur substantial changes in different frames of an image sequence,
due to undergoing Ca2+ signal waves. Hence, intensity is not a
reliable cue for finding optimal correspondence between points in
different frames of a sequence. Here, we utilize the results from the
segmentation stage.
Rigid transformation
Since there are lots of movements (i.e.,
rotation and translation) of pouches in image sequences, we first
compute an optimal rigid transformation to reduce their influence.
This optimization step uses a regular-step gradient descent algorithm. Note that here, local optimum traps could be an issue in practice. Specifically, since the pouches are often of oval shapes, a local
optimum may yield incorrect results where the object is aligned with
the opposite orientation. To resolve this issue, we always initialize
the optimization process using the optimum parameters computed
from the preceding frame in the sequence, since the pouch movement in consecutive frames is usually not very big.
Non-rigid transformation The non-rigid registration seeks an optimal transformation T : (x, y, t) → (x0 , y 0 , t0 ), which maps any
points in the source image at time t to time t0 (i.e., in the reference
image). We use the free-form deformation (FFD) transformation
model, based on B-splines [9], to find an optimal transformation.
For our problem, not too many control points are needed outside
a pouch, because we need to focus only on the ROI. Thus, detecting
ROIs first can save computation time. Pouches will be near the same
position in the frames after the rigid registration, making it possible
to do non-rigid transformation only around the ROI area. Based on
this observation, we can crop an area in the first frame, and apply a
lattice Φ to this area in the following changing frames. We define the
lattice Φ as an (m + 3) × (n + 3) grid in the domain Ω (see Fig. 3).
We define the registration model as follows. Let Ω = {(x, y) |
Xl ≤ x < Xr , Yl ≤ y < Yr } be a rectangular domain in the
xy plane, where the X and Y values specify the boundary of the
detection area. To approximate the intensity of scattered points,
I(x, y), in a pouch, we formulate a function f as a uniform cubic
3 P
3
P
B-spline function: f (x, y) =
Bi (s)Bj (t)Φ(i+k,j+l) , where
j=0 i=0
s = x−bxc, t = y −byc, k = bxc−1, and l = byc−1. In addition,
Bi represents the i-th basis function of the cubic B-spline: B0 (t) =
(1−t)3 /6, B1 (t) = (3t3 −6t2 +4)/6, B2 (t) = (−3t3 −6t2 +4)/6,
and B3 (t) = (t3 )/6.
Since the resolution of the control points determines the nonrigid degree and has a big impact on the computation time, a higher
resolution of control points gives more freedom to do deformation
while also increasing the computation time. To optimize this tradeoff, we use a multi-level mesh grid approach [10] to devise a computationally efficient algorithm. Let Φ1 , Φ2 , . . . , Φg denote a hierarchy of meshes of control points with increasing resolutions and
T1 , T2 , . . . , Tg be the deformation functions with respect to each
mesh. We first apply a coarse deformation, and then refine the deformation gradually. The size of the mesh grid is increased by a factor
of 2 with the same spacing, so that the raw image is down-sampled
to the corresponding size at different levels. The final deformation is
the composition of these functions: T (Ω) = Tg (...T2 (T1 (Ω))...).
To obtain an optimal transformation Φ, we minimize the following energy function: E = Fs + QS, where the first term is the
similarity measure and the second term is for regularization. Q is the
curvature penalization and S is the displacement of control points.
The similarity
use is the sum of squared distance (SSD):
Pmeasure we
SSD = N1
(A − B T )2 , where A is the reference image intensity function, B T is the intensity function of the transformed image
of a source image B under the current transformation T and N is the
Table 1. Comparison results of segmentation.
Mean IU F1 score
Level set
0.8235
0.8294
CUMedNet
0.9394
0.9454
U-Net
0.9479
0.9542
Our FCN
0.9586
0.9643
Our FCN + BR 0.9617
0.9682
number of pixels.
This process iteratively updates the transformation parameters,
T , using a gradient descent algorithm. When a local optimum of the
cost function is smaller than or the number of iterations is bigger
than Nmax , the algorithm will terminate.
deformation to perturb the points according to the grid deformation and a rigid transformation. For intensity distortion, we first
add Gaussian noise in random disk-shaped regions within the pouch
to simulate the calcium signal waves and then rescale the intensity
to [0, 1]. Our target is to find the optimal transformation Ti from
every source image Si2 to the corresponding reference image (i.e.,
T
i
Si2 −→
R). To quantify the performance, we compare the intensity root mean square error (RMSE) between the reference image R
and the clean registered images Ci , which are obtained by applying
Ti to the geometrically distorted images without intensity distortion
T
i
(i.e., Si1 −→
Ci ). The idea is to evaluate whether the registration algorithm is able to find an optimal geometric transformation without
damaging the texture. Fig. 5.I shows some visual results of different
methods, and Table 2 gives quantitative results. The results of our
approach are considerably more accurate.
3. EXPERIMENTS AND EVALUATIONS
We evaluate our registration approach from two perspectives. First,
we evaluate the accuracy of the segmentation method, because the
segmentation accuracy is crucial to our overall approach. Second,
we conduct experiments using both synthetic data and real biological
data to assess the registration performance of different approaches.
3.1. Segmentation Evaluations
We conduct experiments on 100 images of 512×512 pixels selected
from 10 randomly chosen control videos. Our method is compared
with both traditional method (e.g., level set [11]) and state-of-theart FCN models (i.e., U-Net [7] and CUMedNet [12]). The FCN
models are trained using the Adam optimizer [13] with a learning
rate of 0.0005. Data augmentation (random rotation and flip) is used
during training. We use the mean IU (intersection over union) and
F1 as the metrics. Table 1 shows the quantitative results of different
methods, and Fig. 4 shows some segmentation examples of pouches
using different methods. It is evident that our FCN model works
better in segmenting difficult ROIs, and our boundary refinement can
help obtain accurate ROI boundaries.
3.2. Registration Evaluations
We compare the registration performance with the Demon algorithm
[14] and a B-spline method based on Residual Complexity (RC)
[15], which are two state-of-the-art non-rigid registration methods
for images with spatially-varying intensity.
Synthetic data For synthetic data, we choose 8 pouch images as
reference images. Specifically, for each reference image R, we generate 20 source images by adding geometric distortion GT , and intensity distortion IT , to simulate an image sequence with undergoG
Table 2. Registration results of synthetic data.
Movie
Our method
RC
Demon
No.
RMSE (pixel)
RMSE (pixel)
RMSE (pixel)
1
0.097 ± 0.019 0.137 ± 0.039 0.150 ± 0.023
2
0.097 ± 0.021 0.133 ± 0.025 0.184 ± 0.043
3
0.091 ± 0.012 0.137 ± 0.033 0.162 ± 0.021
4
0.098 ± 0.009 0.119 ± 0.024 0.128 ± 0.018
5
0.086 ± 0.011 0.094 ± 0.020 0.124 ± 0.010
6
0.088 ± 0.013 0.132 ± 0.059 0.149 ± 0.015
7
0.095 ± 0.018 0.096 ± 0.013 0.133 ± 0.021
8
0.099 ± 0.019 0.109 ± 0.019 0.175 ± 0.034
Average 0.094 ± 0.015 0.120 ± 0.029 0.151 ± 0.023
Table 3. Registration results of real pouch data.
Movie
Our method
RC
Demon
No.
HD (pixel)
HD (pixel)
HD (pixel)
1
5.199 ± 0.484 5.490 ± 0.876 5.633 ± 0.517
2
5.556 ± 1.250 6.189 ± 0.833 6.819 ± 1.193
3
5.568 ± 0.702 6.403 ± 0.812 6.164 ± 0.514
4
4.342 ± 1.347 6.160 ± 1.311 6.761 ± 0.787
5
4.622 ± 0.684 5.461 ± 0.824 5.072 ± 0.330
6
4.913 ± 0.527 5.328 ± 1.394 6.696 ± 0.703
7
6.882 ± 1.162 7.413 ± 0.574 7.192 ± 0.836
8
5.344 ± 0.443 6.173 ± 0.805 5.845 ± 0.576
Average 5.303 ± 0.825 6.077 ± 0.929 6.273 ± 0.682
I
T
T
ing movements and Ca2+ signaling (i.e., R −→
Si1 −→
Si2 , i =
1, 2, . . . , 20). For geometric distortions, we apply an elastic spline
(A) Raw image
(E) U-net
Fig. 3. Configuration of lattice grid Φ for non-rigid transformation.
(B) Ground truth (C) Level set
(F) Our model
(D) CUMedNet
(G) Our model+BR
Fig. 4. Segmentation results of different methods on a somewhat
difficult pouch image (BR = Boundary Refinement). The red arrow
indicates a location that demonstrates the effect of BR.
Project, the Notre Dame Advanced Diagnostics and Therapeutics
Berry Fellowship, the Notre Dame Integrated Imaging Facility, the
Bloomington Stock Center for fly stocks.
(A)
(B)
(A)
(B)
6. REFERENCES
(C)
(D)
(E)
(F)
I
(C)
(D)
(E)
(F)
II
Fig. 5. Registration results. I: Synthetic data. (A) A reference image;
(B) A source image; (C) The expected registration result; (D) Our
method; (E) RC; (F) Demon. II: Real pouches. (C) Intermediate
segmentation; (D) Our method; (E) RC; (F) Demon. The red arrows
point at some areas with unsatisfactory registration.
Wing disc pouch data We randomly choose 8 movies from 150
control videos. In each movie, we take the first frame as the reference image and all the other frames as source images. To validate
the registration results, we apply transformation T to the annotation
of source images to obtain the registered annotation boundary and
compare the results using the Hausdorff distance (HD) error metric.
Table 3 gives the quantitative results. Our approach achieves accurate boundary shapes. Also, our method obtains clear texture inside
ROIs, as shown by the examples in Fig. 5.II.
4. DISCUSSIONS AND CONCLUSIONS
In this paper, we propose a new two-stage non-rigid image registration approach and apply it to analyze live imaging data of wing disc
pouches for Ca2+ signaling study. Comparing to the state-of-the-art
non-rigid methods for biomedical image registration, our approach
achieves higher accuracy in aligning images with non-negligible texture distortions. Our approach lays a foundation for quantitative
analysis of pouch image sequences in whole-organ studies of Ca2+
signaling related diseases. Our approach may be extended to solving
other biomedical image registration problems, especially when the
intensity profiles and texture patterns of the target objects incur significant changes. The mapping stage of our approach is applicationdependent, while our segmentation method is general and can be
applied to many problems by modifying only the graph search based
boundary refinement procedure.
5. ACKNOWLEDGMENT
This research was supported in part by the Nanoelectronics Research
Corporation (NERC), a wholly-owned subsidiary of the Semiconductor Research Corporation (SRC), through Extremely Energy Efficient Collective Electronics (EXCEL), an SRC-NRI Nanoelectronics
Research Initiative under Research Task ID 2698.005, and by NSF
grants CCF-1640081, CCF-1217906, CNS-1629914, CCF-1617735,
CBET-1403887, and CBET-1553826, NIH R35GM124935, Harper
Cancer Research Institute Research like a Champion awards,
Walther Cancer Foundation Interdisciplinary Interface Training
[1] Q. Wu, P. Brodskiy, C. Narciso, M. Levis, J. Chen, P. Liang,
N. Arredondo-Walsh, D. Z. Chen, and J. J. Zartman, “Intercellular calcium waves are controlled by morphogen signaling
during organ development,” bioRxiv, no. 104745, 2017.
[2] E. L. Ardiel, A. Kumar, J. Marbach, R. Christensen, R. Gupta,
W. Duncan, J. S. Daniels, N. Stuurman, D. Colón-Ramos, and
H. Shroff, “Visualizing calcium flux in freely moving nematode embryos,” Biophysical Journal, vol. 112, no. 9, pp. 1975–
1983, 2017.
[3] A. Wong and J. Orchard, “Robust multimodal registration using local phase-coherence representations,” Journal of Signal
Processing Systems, vol. 54, no. 1-3, p. 89, 2009.
[4] A. El-Baz, A. Farag, G. Gimel’farb, and A. E. Abdel-Hakim,
“Image alignment using learning prior appearance model,” in
ICIP, pp. 341–344, 2006.
[5] J. Woo, M. Stone, and J. L. Prince, “Multimodal registration via mutual information incorporating geometric and spatial context,” IEEE Transactions on Image Processing, vol. 24,
no. 2, pp. 757–769, 2015.
[6] J. Chen, L. Yang, Y. Zhang, M. Alber, and D. Z. Chen, “Combining fully convolutional and recurrent neural networks for
3D biomedical image segmentation,” in NIPS, pp. 3036–3044,
2016.
[7] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in MICCAI, pp. 234–241, 2015.
[8] L. Yang, Y. Zhang, J. Chen, S. Zhang, and D. Z. Chen, “Suggestive annotation: A deep active learning framework for
biomedical image segmentation,” in MICCAI, pp. 399–407,
2017.
[9] S. Lee, G. Wolberg, K.-Y. Chwa, and S. Y. Shin, “Image metamorphosis with scattered feature constraints,” IEEE Transactions on Visualization and Computer Graphics, vol. 2, no. 4,
pp. 337–354, 1996.
[10] D. Rueckert, L. Sonoda, E. Denton, S. Rankin, C. Hayes, M. O.
Leach, D. Hill, and D. J. Hawkes, “Comparison and evaluation
of rigid and non-rigid registration of breast MR images,” in
SPIE, vol. 3661, pp. 78–88, 1999.
[11] C. Li, C.-Y. Kao, J. C. Gore, and Z. Ding, “Minimization of
region-scalable fitting energy for image segmentation,” IEEE
Transactions on Image Processing, vol. 17, no. 10, pp. 1940–
1949, 2008.
[12] H. Chen, X. Qi, J.-Z. Cheng, P.-A. Heng, et al., “Deep contextual networks for neuronal structure segmentation.,” in AAAI,
pp. 1167–1173, 2016.
[13] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014.
[14] D.-J. Kroon and C. H. Slump, “MRI modality transformation
in demon registration,” in ISBI, pp. 963–966, 2009.
[15] A. Myronenko and X. Song, “Image registration by minimization of residual complexity,” in CVPR, pp. 49–56, 2009.
| 1cs.CV
|
Journal of Machine Learning Research 18 (2017) 1-37
Submitted 10/16; Revised 6/17; Published 8/17
Learning Scalable Deep Kernels with Recurrent Structure
Maruan Al-Shedivat
[email protected]
Carnegie Mellon University
Andrew Gordon Wilson
[email protected]
arXiv:1610.08936v3 [cs.LG] 5 Oct 2017
Cornell University
Yunus Saatchi
[email protected]
Zhiting Hu
[email protected]
Carnegie Mellon University
Eric P. Xing
[email protected]
Carnegie Mellon University
Editor: Neil Lawrence
Abstract
Many applications in speech, robotics, finance, and biology deal with sequential data, where
ordering matters and recurrent structures are common. However, this structure cannot
be easily captured by standard kernel functions. To model such structure, we propose
expressive closed-form kernel functions for Gaussian processes. The resulting model, GPLSTM, fully encapsulates the inductive biases of long short-term memory (LSTM) recurrent
networks, while retaining the non-parametric probabilistic advantages of Gaussian processes.
We learn the properties of the proposed kernels by optimizing the Gaussian process marginal
likelihood using a new provably convergent semi-stochastic gradient procedure, and exploit
the structure of these kernels for scalable training and prediction. This approach provides a
practical representation for Bayesian LSTMs. We demonstrate state-of-the-art performance
on several benchmarks, and thoroughly investigate a consequential autonomous driving
application, where the predictive uncertainties provided by GP-LSTM are uniquely valuable.
1. Introduction
There exists a vast array of machine learning applications where the underlying datasets
are sequential. Applications range from the entirety of robotics, to speech, audio and video
processing. While neural network based approaches have dealt with the issue of representation
learning for sequential data, the important question of modeling and propagating uncertainty
across time has rarely been addressed by these models. For a robotics application such as a
self-driving car, however, it is not just desirable, but essential to have complete predictive
densities for variables of interest. When trying to stay in lane and keep a safe following
distance from the vehicle front, knowing the uncertainty associated with lanes and lead
vehicles is as important as the point estimates.
c 2017 Maruan Al-Shedivat, Andrew Gordon Wilson, Yunus Saatchi, Zhiting Hu, Eric P. Xing.
License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/. Attribution requirements are provided
at http://jmlr.org/papers/v18/16-498.html.
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Recurrent models with long short-term memory (LSTM) (Hochreiter and Schmidhuber,
1997) have recently emerged as the leading approach to modeling sequential structure. The
LSTM is an efficient gradient-based method for training recurrent networks. LSTMs use
a memory cell inside each hidden unit and a special gating mechanism that stabilizes the
flow of the back-propagated errors, improving the learning process of the model. While
the LSTM provides state-of-the-art results on speech and text data (Graves et al., 2013;
Sutskever et al., 2014), quantifying uncertainty or extracting full predictive distributions
from deep models is still an area of active research (Gal and Ghahramani, 2016a).
In this paper, we quantify the predictive uncertainty of deep models by following a
Bayesian nonparametric approach. In particular, we propose kernel functions which fully
encapsulate the structural properties of LSTMs, for use with Gaussian processes. The
resulting model enables Gaussian processes to achieve state-of-the-art performance on sequential regression tasks, while also allowing for a principled representation of uncertainty
and non-parametric flexibility. Further, we develop a provably convergent semi-stochastic optimization algorithm that allows mini-batch updates of the recurrent kernels. We empirically
demonstrate that this semi-stochastic approach significantly improves upon the standard
non-stochastic first-order methods in runtime and in the quality of the converged solution.
For additional scalability, we exploit the algebraic structure of these kernels, decomposing
the relevant covariance matrices into Kronecker products of circulant matrices, for O(n)
training time and O(1) test predictions (Wilson et al., 2015; Wilson and Nickisch, 2015).
Our model not only can be interpreted as a Gaussian process with a recurrent kernel, but
also as a deep recurrent network with probabilistic outputs, infinitely many hidden units,
and a utility function robust to overfitting.
Throughout this paper, we assume basic familiarity with Gaussian processes (GPs).
We provide a brief introduction to GPs in the background section; for a comprehensive
reference, see, e.g., Rasmussen and Williams (2006). In the following sections, we formalize
the problem of learning from sequential data, provide background on recurrent networks
and the LSTM, and present an extensive empirical evaluation of our model. Specifically, we
apply our model to a number of tasks, including system identification, energy forecasting,
and self-driving car applications. Quantitatively, the model is assessed on the data ranging
in size from hundreds of points to almost a million with various signal-to-noise ratios
demonstrating state-of-the-art performance and linear scaling of our approach. Qualitatively,
the model is tested on consequential self-driving applications: lane estimation and lead
vehicle position prediction. Indeed, the main focus of this paper is on achieving stateof-the-art performance on consequential applications involving sequential data, following
straightforward and scalable approaches to building highly flexible Gaussian process.
We release our code as a library at: http://github.com/alshedivat/keras-gp. This
library implements the ideas in this paper as well as deep kernel learning (Wilson et al.,
2016a) via a Gaussian process layer that can be added to arbitrary deep architectures
and deep learning frameworks, following the Keras API specification. More tutorials and
resources can be found at https://people.orie.cornell.edu/andrew/code.
2
Learning Scalable Deep Kernels with Recurrent Structure
2. Background
We consider the problem of learning a regression function that maps sequences to real-valued
target vectors. Formally, let X = {xi }ni=1 be a collection of sequences, xi = [x1i , x2i , · · · , xli ],
each with corresponding length, li , where xji ∈ X , and X is an arbitrary domain. Let
y = {yi }ni=1 , yi ∈ Rd , be a collection of the corresponding real-valued target vectors.
Assuming that only the most recent L steps of a sequence are predictive of the targets, the
goal is to learn a function, f : X L 7→ Rd , from some family, F, based on the available data.
As a working example, consider the problem of estimating position of the lead vehicle at
the next time step from LIDAR, GPS, and gyroscopic measurements of a self-driving car
available for a number of previous steps. This task is a classical instance of the sequenceto-reals regression, where a temporal sequence of measurements is regressed to the future
position estimates. In our notation, the sequences of inputs are vectors of measurements,
x1 = [x1 ], x2 = [x1 , x2 ], . . . , xn = [x1 , x2 , · · · xn ], are indexed by time and would be of
growing lengths. Typically, input sequences are considered up to a finite-time horizon, L,
that is assumed to be predictive for the future targets of interest. The targets, y1 , y2 , . . . , yn ,
are two-dimensional vectors that encode positions of the lead vehicle in the ego-centric
coordinate system of the self-driving car.
Note that the problem of learning a mapping, f : X L 7→ Rd , is challenging. While
considering whole sequences of observations as input features is necessary for capturing
long-term temporal correlations, it virtually blows up the dimensionality of the problem. If
we assume that each measurement is p-dimensional, i.e., X ⊆ Rp , and consider L previous
steps as distinct features, the regression problem will become (L × p)-dimensional. Therefore,
to avoid overfitting and be able to extract meaningful signal from a finite amount of data, it
is crucial to exploit the sequential nature of observations.
Recurrent models. One of the most successful ways to exploit sequential structure of
the data is by using a class of recurrent models. In the sequence-to-reals regression scenario,
such a model expresses the mapping f : X L 7→ Rd in the following general recurrent form:
y = ψ(hL ) + t , ht = φ(ht−1 , xt ) + δ t , t = 1, . . . , L,
(1)
where xt is an input observation at time t, ht is a corresponding latent representation,
and y is a target vector. Functions φ(·) and ψ(·) specify model transitions and emissions,
respectively, and δ t and t are additive noises. While φ(·) and ψ(·) can be arbitrary, they
are typically time-invariant. This strong but realistic assumption incorporated into the
structure of the recurrent mapping significantly reduces the complexity of the family of
functions, F, regularizes the problem, and helps to avoid severe overfitting.
Recurrent models can account for various patterns in sequences by memorizing internal
representations of their dynamics via adjusting φ and ψ. Recurrent neural networks (RNNs)
model recurrent processes by using linear parametric maps followed by nonlinear activations:
> t−1
> t−1
> t−1
y = ψ(Why
h ), ht = φ(Whh
h , Wxh
x ), t = 1, . . . , L,
3
(2)
Al-Shedivat, Wilson, Saatchi, Hu, Xing
where Why , Whh , Wxh are weight matrices to be learned1 and φ(·) and ψ(·) here are some
fixed element-wise functions. Importantly and contrary to the standard hidden Markov
models (HMMs), the state of an RNN at any time t is distributed and effectively represented
by an entire hidden sequence, [h1 , · · · , ht−1 , ht ]. A major disadvantage of the vanilla RNNs
is that their training is nontrivial due to the so-called vanishing gradient problem (Bengio
et al., 1994): the error back-propagated through t time steps diminishes exponentially which
makes learning long-term relationships nearly impossible.
LSTM. To overcome vanishing gradients, Hochreiter and Schmidhuber (1997) proposed
a long short-term memory (LSTM) mechanism that places a memory cell into each hidden
unit and uses differentiable gating variables. The update rules for the hidden representation
at time t have the following form (here σ (·) and tanh (·) are element-wise sigmoid and
hyperbolic tangent functions, respectively):
Output
t
> t
> t−1
i
=
tanh
W
x
+
W
h
+
b
c ,
xc
hc
gh
Output gate
> t
> t−1
h
+ Wci> ct−1 + bi ,
git = σ Wxi
x + Whi
h
Forget gate
gc
c
gi
Input gate
i
LSTM
ct = gct ct−1 + git it ,
> t−1
> t−1
> t
c
+ bf ,
h
+ Wcf
x + Whf
gct = σ Wxf
ot = tanh ct ,
> t
> t
> t−1
h
+ Wco
c + bo ,
got = σ Wxo
x + Who
(3)
ht = got ot .
As illustrated above, git , gct , and got correspond to the input, forget, and output gates,
respectively. These variables take their values in [0, 1] and when combined with the internal
states, ct , and inputs, xt , in a multiplicative fashion, they play the role of soft gating. The
gating mechanism not only improves the flow of errors through time, but also, allows the
the network to decide whether to keep, erase, or overwrite certain memorized information
based on the forward flow of inputs and the backward flow of errors. This mechanism adds
stability to the network’s memory.
Gaussian processes. The Gaussian process (GP) is a Bayesian nonparametric model
that generalizes the Gaussian distributions to functions. We say that a random function
f is drawn from a GP with a mean function µ and a covariance kernel k, f ∼ GP (µ, k),
if for any vector of inputs, [x1 , x2 , . . . , xn ], the corresponding vector of function values is
Gaussian:
[f (x1 ), f (x2 ), . . . , f (xn )] ∼ N (µ, KX,X ) ,
Input
with mean µ, such that µi = µ(xi ), and covariance matrix KX,X that satisfies (KX,X )ij =
k(xi , xj ). GPs can be seen as distributions over the reproducing kernel Hilbert space (RKHS)
of functions which is uniquely defined by the kernel function, k (Schölkopf and Smola, 2002).
1. The bias terms are omitted for clarity of presentation.
4
Learning Scalable Deep Kernels with Recurrent Structure
GPs with RBF kernels are known to be universal approximators with prior support to within
an arbitrarily small epsilon band of any continuous function
(Micchelli et al., 2006).
2
Assuming additive Gaussian noise, y | x ∼ N f (x), σ , and a GP prior on f (x), given
training inputs x and training targets y, the predictive distribution of the GP evaluated at
an arbitrary test point x∗ is:
f∗ | x∗ , x, y, σ 2 ∼ N (E[f∗ ], Cov[f∗ ]) ,
where
E[f∗ ] = µX∗ + KX∗ ,X [KX,X + σ 2 I]−1 y,
Cov[f∗ ] = KX∗ ,X∗ − KX∗ ,X [KX,X + σ 2 I]−1 KX,X∗ .
(4)
(5)
Here, KX∗ ,X , KX,X∗ , KX,X , and KX∗ ,X∗ are matrices that consist of the covariance function,
k, evaluated at the corresponding points, x ∈ X and x∗ ∈ X∗ , and µX∗ is the mean function
evaluated at x∗ ∈ X∗ . GPs are fit to the data by optimizing the evidence—the marginal
probability of the data given the model—with respect to kernel hyperparameters. The
evidence has the form:
h
i
log P (y | x) = − y> (K + σ 2 I)−1 y + log det(K + σ 2 I) + const,
(6)
where we use a shorthand K for KX,X , and K implicitly depends on the kernel hyperparameters. This objective function consists of a model fit and a complexity penalty term that
results in an automatic Occam’s razor for realizable functions (Rasmussen and Ghahramani,
2001). By optimizing the evidence with respect to the kernel hyperparameters, we effectively
learn the the structure of the space of functional relationships between the inputs and the
targets. For further details on Gaussian processes and relevant literature we refer interested
readers to the classical book by Rasmussen and Williams (2006).
Turning back to the problem of learning from sequential data, it seems natural to apply
the powerful GP machinery to modeling complicated relationships. However, GPs are
limited to learning only pairwise correlations between the inputs and are unable to account
for long-term dependencies, often dismissing complex temporal structures. Combining GPs
with recurrent models has potential to addresses this issue.
3. Related work
The problem of learning from sequential data, especially from temporal sequences, is well
known in the control and dynamical systems literature. Stochastic temporal processes
are usually described either with generative autoregressive models (AM) or with statespace models (SSM) (Van Overschee and De Moor, 2012). The former approach includes
nonlinear auto-regressive models with exogenous inputs (NARX) that are constructed by
using, e.g., neural networks (Lin et al., 1996) or Gaussian processes (Kocijan et al., 2005).
The latter approach additionally introduces unobservable variables, the state, and constructs
5
Al-Shedivat, Wilson, Saatchi, Hu, Xing
autoregressive dynamics in the latent space. This construction allows to represent and
propagate uncertainty through time by explicitly modeling the signal (via the state evolution)
and the noise. Generative SSMs can be also used in conjunction with discriminative models
via the Fisher kernel (Jaakkola and Haussler, 1999).
Modeling time series with GPs is equivalent to using linear-Gaussian autoregressive or
SSM models (Box et al., 1994). Learning and inference are efficient in such models, but they
are not designed to capture long-term dependencies or correlations beyond pairwise. Wang
et al. (2005) introduced GP-based state-space models (GP-SSM) that use GPs for transition
and/or observation functions. These models appear to be more general and flexible as they
account for uncertainty in the state dynamics, though require complicated approximate
training and inference, which are hard to scale (Turner et al., 2010; Frigola et al., 2014).
Perhaps the most recent relevant work to our approach is recurrent Gaussian processes
(RGP) (Mattos et al., 2015). RGP extends the GP-SSM framework to regression on sequences
by using a recurrent architecture with GP-based activation functions. The structure of
the RGP model mimics the standard RNN, where every parametric layer is substituted
with a Gaussian process. This procedure allows one to propagate uncertainty throughout
the network for an additional cost. Inference is intractable in RGP, and efficient training
requires a sophisticated approximation procedure, the so-called recurrent variational Bayes.
In addition, the authors have to turn to RNN-based approximation of the variational mean
functions to battle the growth of the number of variational parameters with the size of
data. While technically promising, RGP seems problematic from the application perspective,
especially in its implementation and scalability aspects.
Our model has several distinctions with prior work aiming to regress sequences to
reals. Firstly, one of our goals is to keep the model as simple as possible while being
able to represent and quantify predictive uncertainty. We maintain an analytical objective
function and refrain from complicated and difficult-to-diagnose inference schemes. This
simplicity is achieved by giving up the idea of propagating signal through a chain GPs
connected in a recurrent fashion. Instead, we propose to directly learn kernels with recurrent
structure via joint optimization of a simple functional composition of a standard GP with
a recurrent model (e.g., LSTM), as described in detail in the following section. Similar
approaches have recently been explored and proved to be fruitful for non-recurrent deep
networks (Wilson et al., 2016a,b; Calandra et al., 2016). We remark that combinations of
GPs with nonlinear functions have also been considered in the past in a slightly different
setting of warped regression targets (Snelson et al., 2003; Wilson and Ghahramani, 2010;
Lázaro-Gredilla, 2012). Additionally, uncertainty over the recurrent parts of our model is
represented via dropout, which is computationally cheap and turns out to be equivalent
to approximate Bayesian inference in a deep Gaussian process (Damianou and Lawrence,
2013) with particular intermediate kernels (Gal and Ghahramani, 2016a,b). Finally, one
can also view our model as a standalone flexible Gaussian process, which leverages learning
techniques that scale to massive datasets (Wilson and Nickisch, 2015; Wilson et al., 2015).
6
Learning Scalable Deep Kernels with Recurrent Structure
y5
y4
y3
h31
h
x3
x2
x1
x4
x3
x2
g
x3
x2
x2
x1
(a)
y
x5
h13
x4
h12
x3
h11
h23
h22
h21
2
x1
h33
h32
h3
h
g2
g1
y3
1
g3
x1
(b)
φ
x
(c)
Figure 1: (a) Graphical representation of a recurrent model (RNN/LSTM) that maps an input
sequence to a target value in one-step-ahead prediction manner. Shaded variables are observable,
diamond variables denote deterministic dependence on the inputs. (b) Graphical model for GPLSTM with a time lag, L = 3, two training time points, t = 3 and t = 4, and a testing time point,
t = 5. Latent representations are mapped to the outputs through a Gaussian field (denoted in red)
that globally correlates predictions. Dashed variables represent data instances unused at the given
time step. (c) Graphical representation of a GP with a kernel structured with a parametric map, φ.
4. Learning recurrent kernels
Gaussian processes with different kernel functions correspond to different structured probabilistic models. For example, some special cases of the Matérn class of kernels induce models
with Markovian structure (Stein, 1999). To construct deep kernels with recurrent structure
we transform the original input space with an LSTM network and build a kernel directly in
the transformed space, as shown in Figure 1b.
In particular, let L φ : X L 7→ H be an arbitrary deterministic transformation of the
input sequences into some latent space, H. Next, let k : H2 7→ R be a real-valued kernel
defined on H. The decomposition of the kernel and the transformation, k̃ = k ◦ φ, is defined
as
2
k̃(x, x0 ) = k(φ(x), φ(x0 )), where x, x0 ∈ X L , and k̃ : X L 7→ R.
(7)
It is trivial to show that k̃(x, x0 ) is a valid kernel defined on X L (MacKay, 1998, Ch. 5.4.3).
In addition, if φ(·) is represented by a neural network, the resulting model can be viewed as
the same network, but with an additional GP-layer and the negative log marginal likelihood
(NLML) objective function used instead of the standard mean squared error (MSE).
7
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Input embedding is well-known in the Gaussian process literature (e.g. MacKay, 1998;
Hinton and Salakhutdinov, 2008). Recently, Wilson et al. (2016a,b) have successfully scaled
the approach and demonstrated strong results in regression and classification tasks for
kernels based on feedforward and convolutional architectures. In this paper, we apply the
same technique to learn kernels with recurrent structure by transforming input sequences
with a recurrent neural network that acts as φ(·). In particular, a (multi-layer) LSTM
architecture is used to embed L steps of the input time series into a single vector in the
hidden space, H. For the embedding, as common, we use the last hidden vector produced by
the recurrent network. Note that however, any variations to the embedding (e.g., using other
types of recurrent units, adding 1-dimensional pooling layers, or attention mechanisms) are
all fairly straightforward2 . More generally, the recurrent transformation can be random itself
(Figure 1), which would enable direct modeling of uncertainty within the recurrent dynamics,
but would also require inference for φ (e.g., as in Mattos et al., 2015). In this study, we
limit our consideration of random recurrent maps to only those induced by dropout.
Unfortunately, once the the MSE objective is substituted with NLML, it no longer factorizes over the data. This prevents us from using the well-established stochastic optimization
techniques for training our recurrent model. In the case of feedforward and convolutional
networks, Wilson et al. (2016a) proposed to pre-train the input transformations and then
fine-tune them by jointly optimizing the GP marginal likelihood with respect to hyperparameters and the network weights using full-batch algorithms. When the transformation
is recurrent, stochastic updates play a key role. Therefore, we propose a semi-stochastic
block-gradient optimization procedure which allows mini-batching weight updates and fully
joint training of the model from scratch.
4.1 Optimization
The negative log marginal likelihood of the Gaussian process has the following form:
L(K) = y> (Ky + σ 2 I)−1 y + log det(Ky + σ 2 I) + const,
(8)
∆
where Ky + σ 2 I (= K) is the Gram kernel matrix, Ky , is computed on {φ(xi )}N
i=1 and
implicitly depends on the base kernel hyperparameters, θ, and the parameters of the
recurrent neural transformation, φ(·), denoted W and further referred as the transformation
hyperparameters. Our goal is to optimize L with respect to both θ and W .
The derivative of the NLML objective with respect to θ is standard and takes the
following form (Rasmussen and Williams, 2006):
h
i
∂L
1
−1
> −1
−1 ∂K
= tr K yy K − K
,
∂θ
2
∂θ
(9)
2. Details on the particular architectures used in our empirical study are discussed in the next section.
8
Learning Scalable Deep Kernels with Recurrent Structure
where ∂K/∂θ is depends on the kernel function, k(·, ·), and usually has an analytic form.
The derivative with respect to the l-th transformation hyperparameter, Wl , is as follows:3
(
)
>
>
∂k(h
,
h
)
∂h
∂k(h
,
h
)
∂L
1 X −1 > −1
∂h
i
j
j
i
j
i
=
K yy K − K −1
+
,
∂Wl
2
∂hi
∂Wl
∂hj
∂Wl
ij
i,j
(10)
where hi = φ(xi ) corresponds to the latent representation of the the i-th data instance. Once
the derivatives are computed, the model can be trained with any first-order or quasi-Newton
optimization routine. However, application of the stochastic gradient method—the de facto
standard optimization routine for deep recurrent networks—is not straightforward: neither
the objective, nor its derivatives factorize over the data4 due to the kernel matrix inverses,
and hence convergence is not guaranteed.
Semi-stochastic alternating gradient descent.
Observe that once the kernel matrix,
K, is fixed, the expression, K −1 yy> K −1 − K −1 , can be precomputed on the full data
and fixed. Subsequently, Eq. (10) turns into a weighted sum of independent functions of
each data point. This observation suggests that, given a fixed kernel matrix, one could
compute a stochastic update for W on a mini-batch of training points by only using the
corresponding sub-matrix of K. Hence, we propose to optimize GPs with recurrent kernels
in a semi-stochastic fashion, alternating between updating the kernel hyperparameters, θ,
on the full data first, and then updating the weights of the recurrent network, W , using
stochastic steps. The procedure is given in Algorithm 1.
Semi-stochastic alternating gradient descent is a special case of block-stochastic gradient
iteration (Xu and Yin, 2015). While the latter splits the variables into arbitrary blocks
and applies Gauss–Seidel type stochastic gradient updates to each of them, our procedure
alternates between applying deterministic updates to θ and stochastic updates to W of the
(t) (t)
(t) (t)
form θ (t+1) ← θ (t) + λθ gθ and W (t+1) ← W (t) + λW gW 5 . The corresponding Algorithm 1
is provably convergent for convex and non-convex problems under certain conditions. The
following theorem adapts results of Xu and Yin (2015) to our optimization scheme.
Theorem 1 (informal) Semi-stochastic alternating gradient descent converges to a fixed
1+δ
point when the learning rate, λt , decays as Θ(1/t 2 ) for any δ ∈ (0, 1].
Applying alternating gradient to our case has a catch: the kernel matrix (and its inverse)
has to be updated each time W and θ are changed, i.e., on every mini-batch iteration
(marked red in Algorithm 1). Computationally, this updating strategy defeats the purpose
of stochastic gradients because we have to use the entire data on each step. To deal with
the issue of computational efficiency, we use ideas from asynchronous optimization.
3. Step-by-step derivations are given in Appendix B.
4. Cannot be represented as sums of independent functions of each data point.
5. In principle, stochastic updates of θ are also possible. As we will see next, we choose in practice to keep
the kernel matrix fixed while performing stochastic updates. Due to sensitivity of the kernel to even
small changes in θ, convergence of the fully stochastic scheme is fragile.
9
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Algorithm 1 Semi-stochastic alternating
gradient descent.
Algorithm 2 Semi-stochastic asynchronous
gradient descent.
input Data – (X, y), kernel – kθ (·, ·),
recurrent transformation – φw (·).
1: Initialize θ and w; compute initial K.
2: repeat
3:
for all mini-batches Xb in X do
4:
θ ← θ + updateθ (X, θ, K). and
w ← w + updatew (Xb , w, K).
5:
Update the kernel matrix, K.
6:
end for
7: until Convergence
output Optimal θ ∗ and w∗
input Data – (X, y), kernel – kθ (·, ·),
recurrent transformation – φw (·).
1: Initialize θ and w; compute initial K.
2: repeat
3:
θ ← θ + updateθ (X, w, K).
4:
for all mini-batches Xb in X do
5:
w ← w + updatew (Xb , w, K stale ).
6:
end for
7:
Update the kernel matrix, K.
8: until Convergence
output Optimal θ ∗ and w∗
Asynchronous techniques. One of the recent trends in parallel and distributed
optimization is applying updates in an asynchronous fashion (Agarwal and Duchi, 2011).
Such strategies naturally require some tolerance to delays in parameter updates (Langford
et al., 2009). In our case, we modify Algorithm 1 to allow delayed kernel matrix updates.
The key observation is very intuitive: when the stochastic updates of W are small
enough, K does not change much between mini-batches, and hence we can perform multiple
stochastic steps for W before re-computing the kernel matrix, K, and still converge. For
example, K may be updated once at the end of each pass through the entire data (see
Algorithm 2). To ensure convergence of the algorithm, it is important to strike the balance
between (a) the learning rate for W and (b) the frequency of the kernel matrix updates.
The following theorem provides convergence results under certain conditions.
Theorem 2 (informal) Semi-stochastic gradient descent with τ -delayed kernel updates
1+δ
converges to a fixed point when the learning rate, λt , decays as Θ(1/τ t 2 ) for any δ ∈ (0, 1].
Formal statements, conditions, and proofs for Theorems 1 and 2 are given in Appendix C.2.
Why stochastic optimization? GPs with recurrent kernels can be also trained with
full-batch gradient descent, as proposed by Wilson et al. (2016a). However, stochastic
gradient methods have been proved to attain better generalization (Hardt et al., 2016)
and often demonstrate superior performance in deep and recurrent architectures (Wilson
and Martinez, 2003). Moreover, stochastic methods are ‘online’, i.e., they update model
parameters based on subsets of an incoming data stream, and hence can scale to very large
datasets. In our experiments, we demonstrate that GPs with recurrent kernels trained with
Algorithm 2 converge faster (i.e., require fewer passes through the data) and attain better
performance than if trained with full-batch techniques.
10
Learning Scalable Deep Kernels with Recurrent Structure
Stochastic variational inference. Stochastic variational inference (SVI) in Gaussian
processes (Hensman et al., 2013) is another viable approach to enabling stochastic optimization for GPs with recurrent kernels. Such method would optimize a variational lower
bound on the original objective that factorizes over the data by construction. Recently,
Wilson et al. (2016b) developed such a stochastic variational approach in the context of deep
kernel learning. Note that unlike all previous existing work, our proposed approach does
not require a variational approximation to the marginal likelihood to perform mini-batch
training of Gaussian processes.
4.2 Scalability
Learning and inference with Gaussian processes requires solving a linear system involving
an n × n kernel matrix, K −1 y, and computing a log determinant over K. These operations
typically require O(n3 ) computations for n training data points, and O(n2 ) storage. In our
approach, scalability is achieved through semi-stochastic training and structure-exploiting
inference. In particular, asynchronous semi-stochastic gradient descent reduces both the
total number of passes through the data required for the model to converge and the number
of calls to the linear system solver; exploiting the structure of the kernels significantly
reduces the time and memory complexities of the linear algebraic operations.
More precisely, we replace all instances of the covariance matrix Ky with W KU,U W > ,
where W is a sparse interpolation matrix, and KU,U is the covariance matrix evaluated
over m latent inducing points, which decomposes into a Kronecker product of circulant
matrices (Wilson and Nickisch, 2015; Wilson et al., 2015). This construction makes inference
and learning scale as O(n) and test predictions be O(1), while preserving model structure.
For the sake of completeness, we provide an overview of the underlying algebraic machinery
in Appendix A.
At a high level, because W is sparse and KU,U is structured it is possible to take extremely
fast matrix vector multiplications (MVMs) with the approximate covariance matrix KX,X .
One can then use methods such as linear conjugate gradients, which only use MVMs, to
efficiently solve linear systems. MVM or scaled eigenvalue approaches (Wilson and Nickisch,
2015; Wilson et al., 2015) can also be used to efficiently compute the log determinant and
its derivatives. Kernel interpolation (Wilson et al., 2015) also enables fast predictions, as we
describe further in the Appendix.
5. Experiments
We compare the proposed Gaussian processes with recurrent kernels based on RNN and
LSTM architectures (GP-RNN/LSTM) with a number of baselines on datasets of various
complexity and ranging in size from hundreds to almost a million of time points. For the
datasets with more than a few thousand points, we use a massively scalable version of GPs
(see Section 4.2) and demonstrate its scalability during inference and learning. We carry
11
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Table 1: Statistics for the data used in experiments. SNR was determined by assuming a certain
degree of smoothness of the signal, fitting kernel ridge regression with RBF kernel to predict the
targets from the input time series, and regarding the residuals as the noise. Tasks with low average
correlation between inputs and targets and lower SNR are harder prediction problems.
Dataset
Task
Drives
Actuator
system ident.
GEF
power load
wind power
Car
speed
gyro yaw
lanes
lead vehicle
# time steps
# dim
# outputs
Abs. corr.
SNR
500
1,024
1
1
1
1
0.7994
0.0938
25.72
12.47
38,064
130,963
11
16
1
1
0.5147
0.1731
89.93
4.06
932,939
6
6
26
9
1
1
16
2
0.1196
0.0764
0.0816
0.1099
159.33
3.19
—
—
out a number of experiments that help to gain empirical insights about the convergence
properties of the proposed optimization procedure with delayed kernel updates. Additionally,
we analyze the regularization properties of GP-RNN/LSTM and compare them with other
techniques, such as dropout. Finally, we apply the model to the problem of lane estimation
and lead vehicle position prediction, both critical in autonomous driving applications.
5.1 Data and the setup
Below, we describe each dataset we used in our experiments and the associated prediction
tasks. The essential statistics for the datasets are summarized in Table 1.
System identification. In the first set of experiments, we used publicly available nonlinear system identification datasets: Actuator 6 (Sjöberg et al., 1995) and Drives 7 (Wigren,
2010). Both datasets had one dimensional input and output time series. Actuator had the
size of the valve opening as the input and the resulting change in oil pressure as the output.
Drives was from a system with motors that drive a pulley using a flexible belt; the input
was the sum of voltages applied to the motors and the output was the speed of the belt.
Smart grid data8 . We considered the problem of forecasting for the smart grid that
consisted of two tasks (Figure 2). The first task was to predict power load from the historical
temperature data. The data had 11 input time series coming from hourly measurements of
temperature on 11 zones and an output time series that represented the cumulative hourly
power load on a U.S. utility. The second task was to predict power generated by wind
farms from the wind forecasts. The data consisted of 4 different hourly forecasts of the
wind and hourly values of the generated power by a wind farm. Each wind forecast was a
6. http://www.iau.dtu.dk/nnbook/systems.html
7. http://www.it.uu.se/research/publications/reports/2010-020/NonlinearData.zip.
8. The smart grid data were taken from Global Energy Forecasting Kaggle competitions organized in 2012.
12
80
60
40
20
0
140
120
100
80
60
40
Direction, deg Speed, m/s
80
60
40
20
0
Pearson correlation
between the temperature
measurements and the
cumulative power load
P
P
2
4
6
8
10
0
200
400
600
800
1000
1200
1400
2
4
6
8 10
1.0
0.8
0.6
0.4
0.2
0.0
0.2
0.4
Power, nu
Power, MWh
Temperature, F
Learning Scalable Deep Kernels with Recurrent Structure
1600
Days
10
Pearson correlation
between wind parameters
and the cumulative power
generated per hour
5
0
300
P
200
P
100
zc
0
1.0
mc
ws
0.5
0.0
wd
0
100
200
300
400
500
600
700
zc mc ws wd
1.0
0.8
0.6
0.4
0.2
0.0
0.2
0.4
0.6
800
Days
Figure 2: Left: Visualization of the GEF-power time series for two zones and the cumulative load
with the time resolution of 1 day. Cumulative power load is generally negatively correlated with
the temperature measurements on all the zones. Right: Visualization of the GEF-wind time series
with the time resolution of 1 day.
4-element vector that corresponded to zonal component, meridional component, wind speed
and wind angle. In our experiments, we concatenated the 4 different 4-element forecasts,
which resulted in a 16-dimensional input time series.
Self-driving car dataset9 . One of the main target applications of the proposed model
is prediction for autonomous driving. We considered a large dataset coming from sensors
of a self-driving car that was recorded on two car trips with discretization of 10 ms. The
data featured two sets of GPS ECEF locations, ECEF velocities, measurements from a
fiber-optic gyro compass, LIDAR, and a few more time series from a variety of IMU sensors.
Additionally, locations of the left and right lanes were extracted from a video stream for
each time step as well as the position of the lead vehicle from the LIDAR measurements. We
considered the data from the first trip for training and from the second trip for validation
and testing. A visualization of the car routes with 25 second discretization in the ENU
coordinates are given in Figure 3. We consider four tasks, the first two of which are more of
proof-of-concept type variety, while the final two are fundamental to good performance for
a self-driving car:
1. Speed prediction from noisy GPS velocity estimates and gyroscopic inputs.
2. Prediction of the angular acceleration of the car from the estimates of its speed and
steering angle.
3. Point-wise prediction of the lanes from the estimates at the previous time steps, and
estimates of speed, gyroscopic and compass measurements.
4. Prediction of the lead vehicle location from its location at the previous time steps,
and estimates of speed, gyroscopic and compass measurements.
We provide more specific details on the smart grid data and self-driving data in Appendix D.
9. The dataset is proprietary. It was released in part for public use under the Creative Commons
Attribution 3.0 license: http://archive.org/details/comma-dataset. More about the self-driving car:
http://www.bloomberg.com/features/2015-george-hotz-self-driving-car/.
13
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Time
Train
Test
42.8 min
46.5 min
Speed of the self-driving car
Min
0.0 mph
0.0 mph
Max
80.3 mph
70.1 mph
Average
42.4 mph
38.4 mph
Median
58.9 mph
44.8 mph
Distance to the lead vehicle
Min
23.7 m
29.7 m
Max
178.7 m
184.7 m
Average
85.4 m
72.6 m
Median
54.9 m
53.0 m
Figure 3 & Table 2: Left: Train and test routes of the self-driving car in the ENU coordinates
with the origin at the starting location. Arrows point in the direction of motion; color encodes the
speed. Insets zoom selected regions of the routes. Best viewed in color. Right: Summary of the
data collected on the train and test routes.
Models and metrics. We used a number of classical baselines: NARX (Lin et al.,
1996), GP-NARX models (Kocijan et al., 2005), and classical RNN and LSTM architectures.
The kernels of our models, GP-NARX/RNN/LSTM, used the ARD base kernel function and
were structured by the corresponding baselines.10 As the primary metric, we used root mean
squared error (RMSE) on a held out set and additionally negative log marginal likelihood
(NLML) on the training set for the GP-based models.
We train all the models to perform one-step-ahead prediction in an autoregressive setting,
where targets at the future time steps are predicted from the input and target values at a
fixed number of past time steps. For the system identification task, we additionally consider
the non-autoregressive scenario (i.e., mapping only input sequences to the future targets),
where we are performing prediction in the free simulation mode, and included recurrent
Gaussian processes (Mattos et al., 2015) in our comparison. In this case, none of the future
targets are available and the models have to re-use their own past predictions to produce
future forecasts).
A note on implementation. Recurrent parts of each model were implemented using
Keras 11 library. We extended Keras with the Gaussian process layer and developed a backed
engine based on the GPML library12 . Our approach allows us to take full advantage of
the functionality available in Keras and GPML, e.g., use automatic differentiation for the
recurrent part of the model. Our code is available at http://github.com/alshedivat/kgp/.
10. GP-NARX is a special instance of our more general framework and we trained it using the proposed
semi-stochastic algorithm.
11. http://www.keras.io
12. http://www.gaussianprocess.org/gpml/code/matlab/doc/
14
Learning Scalable Deep Kernels with Recurrent Structure
10−2
mini, 16
mini, 64
10000
5000
0
0
10
20
30
40
Epoch number
50
0
10
20
30
40
Epoch number
full
mini
0.5
0.4
0.3
0.2
0.1
0.0 1
2
50
Test RMSE
10−1
full, 16
full, 64
Test RMSE
100
0.6
15000
full, 16
full, 64
mini, 16
mini, 64
Train NLML
Test RMSE
101
22
23
24
25
Number of hidden units
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0 0
2
lr=0.1
lr=0.01
lr=0.001
21
22
23
24
25
Number of batches per epoch
Figure 4: Two charts on the left: Convergence of the optimization in terms of RMSE on test and
NLML on train. The inset zooms the region of the plot right beneath it using log scale for the
vertical axis. full and mini denote full-batch and mini-batch optimization procedures, respectively,
while 16 and 64 refer to models with the respective number of units per hidden layer. Two charts on
the right: Test RMSE for a given architecture trained with a specified method and/or learning rate.
5.2 Analysis
This section discusses quantitative and qualitative experimental results. We only briefly
introduce the model architectures and the training schemes used in each of the experiments.
We provide a comprehensive summary of these details in Appendix E.
5.2.1 Convergence of the optimization
To address the question of whether stochastic optimization of recurrent kernels is necessary
and to assess the behavior of the proposed optimization scheme with delayed kernel updates,
we conducted a number of experiments on the Actuator dataset (Figure 4).
First, we constructed two GP-LSTM models with 1 recurrent hidden layer and 16 or 64
hidden units and trained them with (non-stochastic) full-batch iterative procedure (similar
to the proposal of Wilson et al. (2016a)) and with our semi-stochastic optimizer with delayed
kernel updates (Algorithm 2). The convergence results are given on the first two charts.
Both in terms of the error on a held out set and the NLML on the training set, the models
trained with mini-batches converged faster and demonstrated better final performance.
Next, we compared the two optimization schemes on the same GP-LSTM architecture
with different sizes of the hidden layer ranging from 2 to 32. It is clear from the third chart
that, even though full-batch approach seemed to find a better optimum when the number of
hidden units was small, the stochastic approach was clearly superior for larger hidden layers.
Finally, we compared the behavior of Algorithm 2 with different number of mini-batches
used for each epoch (equivalently, the number of steps between the kernel matrix updates)
and different learning rates. The results are give on the last chart. As expected, there is a
fine balance between the number of mini-batches and the learning rate: if the number of
mini-batches is large (i.e., the delay between the kernel updates becomes too long) while the
learning rate is high enough, optimization does not converge; at the same time, an appropriate
combination of the learning rate and the mini-batch size leads better generalization than
the default batch approach of Wilson et al. (2016a).
15
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Table 3: Average performance of the models in terms of RMSE on the system identification tasks.
The averages were computed over 5 runs; the standard deviation is given in the parenthesis. Results
for the RGP model are as reported by Mattos et al. (2015), available only for the free simulation.
regression
Drives
auto-regression
free simulation
regression
Actuator
auto-regression
free simulation
NARX
RNN
LSTM
0.33 (0.02)
0.53 (0.02)
0.29 (0.02)
0.19 (0.03)
0.17 (0.04)
0.14 (0.02)
0.38 (0.03)
0.56 (0.03)
0.40 (0.02)
0.49 (0.05)
0.56 (0.03)
0.40 (0.03)
0.18 (0.01)
0.17 (0.01)
0.19 (0.01)
0.57 (0.04)
0.68 (0.05)
0.44 (0.03)
GP-NARX
GP-RNN
GP-LSTM
0.28 (0.02)
0.37 (0.04)
0.25 (0.02)
0.16 (0.04)
0.16 (0.03)
0.13 (0.02)
0.28 (0.02)
0.45 (0.03)
0.32 (0.03)
0.46 (0.03)
0.49 (0.02)
0.36 (0.01)
0.14 (0.01)
0.15 (0.01)
0.14 (0.01)
0.63 (0.04)
0.55 (0.04)
0.43 (0.03)
—
—
0.249
—
—
0.368
RGP
5.2.2 Regression, Auto-regression, and Free Simulation
In this set of experiments, our main goal is to provide a comparison between three different
modes of one-step-ahead prediction, referred to as (i) regression, (ii) autoregression, and
(iii) free simulation, and compare performance of our models with RGP—a classical RNN
with every parametric layer substituted with a Gaussian process (Mattos et al., 2015)—on
the Actuator and Drives datasets. The difference between the prediction modes consists in
whether and how the information about the past targets is used. In the regression scenario,
inputs and targets are separate time series and the model learns to map input values at
a number of past time points to a target value at a future point in time. Autoregression,
additionally, uses the true past target values as inputs; in the free simulation mode, the
model learns to map past inputs and its own past predictions to a future target.
In the experiments in autoregression and free simulation modes, we used short time lags,
L = 10, as suggested by Mattos et al. (2015). In the regression mode, since the model does
not build the recurrent relationships based on the information about the targets (or their
estimates), it generally requires larger time lags that can capture the state of the dynamics.
Hence we increased the time lag to 32 in the regression mode. More details are given in
Appendix E.
We present the results in Table 3. We note that GP-based architectures consistently
yielded improved predictive performance compared to their vanilla deep learning counterparts
on both of the datasets, in each mode. Given the small size of the datasets, we attribute
such behavior to better regularization properties of the negative log marginal likelihood loss
function. We also found out that when GP-based models were initialized with weights of
pre-trained neural networks, they tended to overfit and give overly confident predictions
on these tasks. The best performance was achieved when the models were trained from a
random initialization (contrary to the findings of Wilson et al., 2016a). In free simulation
mode RGP performs best of the compared models. This result is expected—RGP was
particularly designed to represent and propagate uncertainty through a recurrent process.
16
Learning Scalable Deep Kernels with Recurrent Structure
Table 4: Average performance of the best models in terms of RMSE on the GEF and Car tasks.
The averages were computed over 5 runs; the standard deviation is given in the parenthesis.
GEF
power load
wind power
speed
Car
gyro yaw
lanes
lead vehicle
NARX
RNN
LSTM
0.54 (0.02)
0.61 (0.02)
0.45 (0.01)
0.84 (0.01)
0.81 (0.01)
0.77 (0.01)
0.114 (0.010)
0.152 (0.012)
0.027 (0.008)
0.19 (0.01)
0.22 (0.01)
0.13 (0.01)
0.13 (0.01)
0.33 (0.02)
0.08 (0.01)
0.41 (0.02)
0.44 (0.03)
0.40 (0.01)
GP-NARX
GP-RNN
GP-LSTM
0.78 (0.03)
0.24 (0.02)
0.17 (0.02)
0.83 (0.02)
0.79 (0.01)
0.76 (0.01)
0.125 (0.015)
0.089 (0.013)
0.019 (0.006)
0.23 (0.02)
0.24 (0.01)
0.08 (0.01)
0.10 (0.01)
0.46 (0.08)
0.06 (0.01)
0.34 (0.02)
0.41 (0.02)
0.32 (0.02)
Our framework focuses on using recurrence to build expressive kernels for regression on
sequences.
The suitability of each prediction mode depends on the task at hand. In many applications
where the future targets become readily available as the time passes (e.g., power estimation
or stock market prediction), the autoregression mode is preferable. We particularly consider
autoregressive prediction in the further experiments.
5.2.3 Prediction for smart grid and self-driving car applications
For both smart grid prediction tasks we used LSTM and GP-LSTM models with 48 hour
time lags and were predicting the target values one hour ahead. LSTM and GP-LSTM
were trained with one or two layers and 32 to 256 hidden units. The best models were
selected on 25% of the training data used for validation. For autonomous driving prediction
tasks, we used the same architectures but with 128 time steps of lag (1.28 s). All models
were regularized with dropout (Srivastava et al., 2014; Gal and Ghahramani, 2016b). On
both GEF and self-driving car datasets, we used the scalable version of Gaussian process
(MSGP) (Wilson et al., 2015). Given the scale of the data and the challenge of nonlinear
optimization of the recurrent models, we initialized the recurrent parts of GP-RNN and
GP-LSTM with pre-trained weights of the corresponding neural networks. Fine-tuning of the
models was performed with Algorithm 2. The quantitative results are provided in Table 4
and demonstrate that GPs with recurrent kernels attain the state-of-the-art performance.
Additionally, we investigated convergence and regularization properties of LSTM and
GP-LSTM models on the GEF-power dataset. The first two charts of Figure 6 demonstrate
that GP-based models are less prone to overfitting, even when the data is not enough. The
third panel shows that architectures with a particular number of hidden units per layer
attain the best performance on the power prediction task. An additional advantage of
the GP-layers over the standard recurrent networks is that the best architecture could be
identified based on the negative log likelihood of the model as shown on the last chart.
17
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Front distance, m
50
40
30
20
10
0
−5
0
5
−5
0
5
−5
0
Side distance, m
5
−5
0
5
−5
0
5
Front distance, m
50
40
30
20
10
0
−5
0
−5
5
0
−5
5
−5
0
5
Side distance, m
0
−5
5
0
5
(a) Point-wise predictions of the lanes made by LSTM (upper) and by GP-LSTM (lower). Dashed
lines correspond to the ground truth extracted from video sequences and used for training.
Front distance, m
100
80
60
40
20
0
−5
0
5
−5
0
5
−5
0
Side distance, m
5
−5
0
5
−5
0
5
−5
0
5
−5
0
5
−5
0
Side distance, m
5
−5
0
5
−5
0
5
Front distance, m
100
80
60
40
20
0
(b) LSTM (upper) and by GP-LSTM (lower) position predictions of the lead vehicle. Black markers
and dashed lines are the ground truth; blue and red markers with solid lines correspond to predictions.
Figure 5: Qualitative comparison of the LSTM and GP-LSTM predictions on self-driving tasks.
Predictive uncertainty of the GP-LSTM model is showed by contour plots and error-bars; the latter
denote one standard deviation of the predictive distributions.
Finally, Figure 5 qualitatively demonstrates the difference between the predictions given
by LSTM vs. GP-LSTM on point-wise lane estimation (Figure 5a) and the front vehicle
tracking (Figure 5b) tasks. We note that GP-LSTM not only provides a more robust fit,
but also estimates the uncertainty of its predictions. Such information can be further used
in downstream prediction-based decision making, e.g., such as whether a self-driving car
should slow down and switch to a more cautious driving style when the uncertainty is high.
18
100
10−1
10−2
0
20
40
60
80
Epoch number
100
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
LSTM-1H
GP-LSTM-1H
0
5
10
15
0.14
0.12
0.10
0.08
0.06
0.04
0.02
LSTM-1H
LSTM-2H
GP-LSTM-1H
GP-LSTM-2H
7500
5000
2500
100
20
101
102
103
100
Number of hidden units
Number of training pts, 103
GP-LSTM-1H
GP-LSTM-2H
10000
Train NLML
LSTM-1H
LSTM-2H
GP-LSTM-1H
GP-LSTM-2H
Test RMSE
Test RMSE
101
Test RMSE
Learning Scalable Deep Kernels with Recurrent Structure
101
102
103
Number of hidden units
400
300
200
100
0
0
20
40
60
80
100 120
Number of training pts, 103
700
600
500
400
300
200
100
0
10
20
100
40
80
200
300
400
Number of inducing pts
8.0
7.5
7.0
6.5
6.0
5.5
5.0
4.5
10
Time per test pt, ms
100 pts
200 pts
400 pts
Time per test pt, ms
500
Time per epoch, s
Time per epoch, s
Figure 6: Left to right: RMSE vs. the number of training points; RMSE vs. the number model
parameters per layer; NLML vs. the number model parameters per layer for GP-based models. All
metrics are averages over 5 runs with different random initializations, computed on a held-out set.
100 pts
200 pts
400 pts
20
30
40
50
60
Number of training pts, 103
8.0
7.5
7.0
6.5
6.0
5.5
5.0
4.5
10
20
100
40
80
200
300
400
Number of inducing pts
Figure 7: The charts demonstrate scalability of learning and inference of MSGP with an LSTM-based
recurrent kernel. Legends with points denote the number of inducing points used. Legends with
percentages denote the percentage of the training dataset used learning the model.
5.2.4 Scalability of the model
Following Wilson et al. (2015), we performed a generic scalability analysis of the MSGPLSTM model on the car sensors data. The LSTM architecture was the same as described
in the previous section: it was transforming multi-dimensional sequences of inputs to a
two-dimensional representation. We trained the model for 10 epochs on 10%, 20%, 40%,
and 80% of the training set with 100, 200, and 400 inducing points per dimension and
measured the average training time per epoch and the average prediction time per testing
point. The measured time was the total time spent on both LSTM optimization and MSGP
computations. The results are presented in Figure 7.
The training time per epoch (one full pass through the entire training data) grows linearly
with the number of training examples and depends linearly on the number of inducing points
(Figure 7, two left charts). Thus, given a fixed number of inducing points per dimension, the
time complexity of MSGP-LSTM learning and inference procedures is linear in the number
of training examples. The prediction time per testing data point is virtually constant and
does not depend on neither on the number of training points, nor on the number of inducing
points (Figure 7, two right charts).
19
Al-Shedivat, Wilson, Saatchi, Hu, Xing
6. Discussion
We proposed a method for learning kernels with recurrent long short-term memory structure
on sequences. Gaussian processes with such kernels, termed the GP-LSTM, have the structure
and learning biases of LSTMs, while retaining a probabilistic Bayesian nonparametric
representation. The GP-LSTM outperforms a range of alternatives on several sequence-toreals regression tasks. The GP-LSTM also works on data with low and high signal-to-noise
ratios, and can be scaled to very large datasets, all with a straightforward, practical, and
generally applicable model specification. Moreover, the semi-stochastic scheme proposed
in our paper is provably convergent and efficient in practical settings, in conjunction with
structure exploiting algebra. In short, the GP-LSTM provides a natural mechanism for
Bayesian LSTMs, quantifying predictive uncertainty while harmonizing with the standard
deep learning toolbox. Predictive uncertainty is of high value in robotics applications, such
as autonomous driving, and could also be applied to other areas such as financial modeling
and computational biology.
There are several exciting directions for future research. The GP-LSTM quantifies
predictive uncertainty but does not model the propagation of uncertainty in the inputs
through a recurrent structure. Treating free simulation as a structured prediction problem
and using online corrective algorithms, e.g., DAGGER (Ross et al., 2011), are likely to
improve performance of GP-LSTM in the free prediction mode. This approach would not
require explicitly modeling and propagating uncertainty through the recurrence and would
maintain the high computational efficiency of our method.
Alternatively, it would be exciting to have a probabilistic treatment of all parameters of
the GP-LSTM kernel, including all LSTM weights. Such an extension could be combined with
stochastic variational inference, to enable both classification and non-Gaussian likelihoods
as in Wilson et al. (2016b), but also open the doors to stochastic gradient Hamiltonian
Monte Carlo (Chen et al., 2014) (SG-HMC) for efficient inference over kernel parameters.
Indeed, SG-HMC has recently been used for efficient inference over network parameters in
the Bayesian GAN (Saatchi and Wilson, 2017). A Bayesian approach to marginalizing the
weights of the GP-LSTM kernel would also provide a principled probabilistic mechanism for
learning model hyperparameters.
One could relax several additional assumptions. We modeled each output dimension
with independent GPs that shared a recurrent transformation. To capture the correlations
between output dimensions, it would be promising to move to a multi-task formulation. In
the future, one could also learn the time horizon in the recurrent transformation, which
could lead to major additional performance gains.
Finally, the semi-stochastic learning procedure naturally complements research in asynchronous optimization (e.g., Deisenroth and Ng, 2015). In combination with stochastic
variational inference, the semi-stochastic approach could be used for parallel kernel learning,
side-stepping the independence assumptions in prior work. We envision that such efforts for
Gaussian processes will harmonize with current progress in Bayesian deep learning.
20
Learning Scalable Deep Kernels with Recurrent Structure
7. Acknowledgements
The authors thank Yifei Ma for helpful discussions and the anonymous reviewers for the
valuable comments that helped to improve the paper. This work was supported in part by
NIH R01GM114311, AFRL/DARPA FA87501220324, and NSF IIS-1563887.
21
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Appendix A. Massively scalable Gaussian processes
Massively scalable Gaussian processes (MSGP) (Wilson et al., 2015) is a significant extension
of the kernel interpolation framework originally proposed by Wilson and Nickisch (2015). The
core idea of the framework is to improve scalability of the inducing point methods (QuinoneroCandela and Rasmussen, 2005) by (1) placing the virtual points on a regular grid, (2)
exploiting the resulting Kronecker and Toeplitz structures of the relevant covariance matrices,
and (3) do local cubic interpolation to go back to the kernel evaluated at the original points.
This combination of techniques brings the complexity down to O(n) for training and O(1)
for each test prediction. Below, we overview the methodology. We remark that a major
difference in philosophy between MSGP and many classical inducing point methods is that
the points are selected and fixed rather than optimized over. This allows to use significantly
more virtual points which typically results in a better approximation of the true kernel.
A.1 Structured kernel interpolation
Given a set of m inducing points, the n × m cross-covariance matrix, KX,U , between the
training inputs, X, and the inducing points, U, can be approximated as K̃X,U = WX KU,U
using a (potentially sparse) n × m matrix of interpolation weights, WX . This allows to
approximate KX,Z for an arbitrary set of inputs Z as KX,Z ≈ K̃X,U WZ> . For any given
kernel function, K, and a set of inducing points, U, structured kernel interpolation (SKI)
procedure (Wilson and Nickisch, 2015) gives rise to the following approximate kernel:
KSKI (x, z) = WX KU,U Wz> ,
(A.1)
which allows to approximate KX,X ≈ WX KU,U WX> . Wilson and Nickisch (2015) note that
standard inducing point approaches, such as subset of regression (SoR) or fully independent
training conditional (FITC), can be reinterpreted from the SKI perspective. Importantly,
the efficiency of SKI-based MSGP methods comes from, first, a clever choice of a set of
inducing points that allows to exploit algebraic structure of KU,U , and second, from using
very sparse local interpolation matrices. In practice, local cubic interpolation is used (Keys,
1981).
A.2 Kernel approximations
If inducing points, U , form a regularly spaced P -dimensional grid, and we use a stationary
product kernel (e.g., the RBF kernel), then KU,U decomposes as a Kronecker product of
Toeplitz matrices:
KU,U = T1 ⊗ T2 ⊗ · · · ⊗ TP .
(A.2)
The Kronecker structure allows to compute the eigendecomposition of KU,U by separately
decomposing T1 , . . . , TP , each of which is much smaller than KU,U . Further, to efficiently
22
Learning Scalable Deep Kernels with Recurrent Structure
eigendecompose a Toeplitz matrix, it can be approximated by a circulant matrix13 which
eigendecomposes by simply applying discrete Fourier transform (DFT) to its first column.
Therefore, an approximate eigendecomposition of each T1 , . . . , TP is computed via the fast
Fourier transform (FFT) and requires only O(m log m) time.
A.3 Structure exploiting inference
To perform inference, we need to solve (KSKI + σ 2 I)−1 y; kernel learning requires evaluating
log det(KSKI + σ 2 I). The first task can be accomplished by using an iterative scheme—linear
conjugate gradients—which depends only on matrix vector multiplications with (KSKI + σ 2 I).
The second is done by exploiting the Kronecker and Toeplitz structure of KU,U for computing
an approximate eigendecomposition, as described above.
A.4 Fast Test Predictions
To achieve constant time prediction, we approximate the latent mean and variance of f∗ by
applying the same SKI technique. In particular, for a set of n∗ testing points, X∗ , we have
−1
E[f∗ ] = µX∗ + KX∗ ,X KX,X + σ 2 I
y
h
i−1
≈ µX∗ + K̃X∗ ,X K̃X,X + σ 2 I
y,
(A.3)
where K̃X,X = W KU,U W > and K̃X∗ ,X = W∗ KU,U W > , and W and W∗ are n×m and n∗ ×m
sparse interpolation matrices, respectively. Since KU,U W > [K̃X,X + σ 2 I]−1 y is precomputed
at training time, at test time, we only multiply the latter with W∗ matrix which results which
costs O(n∗ ) operations leading to O(1) operations per test point. Similarly, approximate
predictive variance can be also estimated in O(1) operations (Wilson et al., 2015).
Note that the fast prediction methodology can be readily applied to any trained Gaussian
process model as it is agnostic to the way inference and learning were performed.
Appendix B. Gradients for GPs with recurrent kernels
GPs with deep recurrent kernels are trained by minimizing the negative log marginal
likelihood objective function. Below we derive the update rules.
By applying the chain rule, we get the following first order derivatives:
∂L
∂L ∂K
=
·
,
∂γ
∂K ∂γ
∂L
∂L ∂K ∂φ
=
·
·
.
∂W
∂K ∂φ ∂W
(B.4)
The derivative of the log marginal likelihood w.r.t. to the kernel hyperparameters, θ, and
the parameters of the recurrent map, W , are generic and take the following form (Rasmussen
13. Wilson et al. (2015) explored 5 different approximation methods known in the numerical analysis
literature.
23
Al-Shedivat, Wilson, Saatchi, Hu, Xing
and Williams, 2006, Ch. 5, Eq. 5.9):
∂L
∂θ
∂L
∂W
= 12 tr
−1 > −1
K yy K − K −1 ∂K
∂θ ,
= 21 tr
K −1 yy> K −1 − K −1
∂K
∂W
.
(B.5)
(B.6)
The derivative ∂K/∂θ is also standard and depends on the form of a particular chosen
kernel function, k(·, ·). However, computing each part of ∂L/∂W is a bit subtle, and hence
we elaborate these derivations below.
Consider the ij-th entry of the kernel matrix, Kij . We can think of K as a matrix-valued
function of all the data vectors in d-dimensional transformed space which we denote by
H ∈ RN ×d . Then Kij is a scalar-valued function of H and its derivative w.r.t. the l-th
parameter of the recurrent map, Wl , can be written as follows:
!
∂Kij
∂Kij > ∂H
= tr
.
(B.7)
∂Wl
∂H
∂Wl
Notice that ∂Kij /∂H is a derivative of a scalar w.r.t. to a matrix and hence is a matrix;
∂H/∂Wl is a derivative of a matrix w.r.t. to a scalar which is taken element-wise and also
gives a matrix. Also notice that Kij is a function of H, but it only depends the i-th and
j-th elements for which the kernel is being computed. This means that ∂Kij /∂H will have
only non-zero i-th row and j-th column and allows us to re-write (B.7) as follows:
∂Kij > ∂hj
∂hi
+
∂Wl
∂hj
∂Wl
>
∂K(hi , hj )
∂K(hi , hj ) > ∂hj
∂hi
=
+
.
∂hi
∂Wl
∂hj
∂Wl
∂Kij
=
∂Wl
∂Kij
∂hi
>
(B.8)
Since the kernel function has two arguments, the derivatives must be taken with respect of
each of them and evaluated at the corresponding points in the hidden space, hi = φ(xi )
and hj = φ(xj ). When we plug this into (B.6), we arrive at the following expression:
)
(
>
>
∂K(h
,
h
)
∂K(h
,
h
)
∂h
∂L
1 X −1 > −1
∂h
i
j
i
j
j
i
.
=
K yy K − K −1
+
∂Wl
2
∂hi
∂Wl
∂hj
∂Wl
ij
i,j
(B.9)
The same expression can be written in a more compact form using the Einstein notation:
!
j ∂K jd ∂K id ∂h dl
∂L
1 −1 > −1
=
K yy K − K −1
+
(B.10)
∂Wl
2
∂h i
∂h j
∂W i
i
where d indexes the dimensions of the h and l indexes the dimensions of W .
24
Learning Scalable Deep Kernels with Recurrent Structure
In practice, deriving a computationally efficient analytical form of ∂K/∂h might be too
complicated for some kernels (e.g., the spectral mixture kernels (Wilson and Adams, 2013)),
especially if the grid-based approximations of the kernel are enabled. In such cases, we can
simply use a finite difference approximation of this derivative. As we remark in the following
section, numerical errors that result from this approximation do not affect convergence of
the algorithm.
Appendix C. Convergence results
Convergence results for the semi-stochastic alternating gradient schemes with and without
delayed kernel matrix updates are based on (Xu and Yin, 2015). There are a few notable
differences between the original setting and the one considered in this paper:
1. Xu and Yin (2015) consider a stochastic program that minimizes the expectation of
the objective w.r.t. some distribution underlying the data:
min f (x) := Eξ F (x; ξ),
x∈X
(C.11)
where every iteration a new ξ is sampled from the underlying distribution. In our
case, the goal is to minimize the negative log marginal likelihood on a particular given
dataset. This is equivalent to the original formulation (C.11), but with the expectation
taken w.r.t. the empirical distribution that corresponds to the given dataset.
2. The optimization procedure of Xu and Yin (2015) has access to only a single random
point generated from the data distribution at each step. Our algorithm requires having
access to the entire training data each time the kernel matrix is computed.
3. For a given sample, Xu and Yin (2015) propose to loop over a number of coordinate
blocks and apply Gauss–Seidel type gradient updates to each block. Our semi-stochastic
scheme has only two parameter blocks, θ and W , where θ is updated deterministically
on the entire dataset while W is updated with stochastic gradient on samples from
the empirical distribution.
Noting these differences, we first adapt convergence results for the smooth non-convex
case (Xu and Yin, 2015, Theorem 2.10) to our scenario, and then consider the variant with
delaying kernel matrix updates.
C.1 Semi-stochastic alternating gradient
As shown in Algorithm 1, we alternate between updating θ and W . At step t, we get a
mini-batch of size Nt , xt ≡ {x̄i }i∈It , which is just a selection of points from the full set, x.
Define the gradient errors for θ and W at step t as follows:
δθt := g̃θt − gθt ,
t
t
t
:= g̃W
− gW
,
δW
25
(C.12)
Al-Shedivat, Wilson, Saatchi, Hu, Xing
T are the true gradients and g̃T and g̃T are estimates14 . We first update θ
where gθT and gW
W
θ
and then W , and hence the expressions for the gradients take the following form:
∂K
1
1 X −1 > −1
t
−1
t
t
t
t
g̃θ ≡ gθ =
∇θ L(θ , W ) =
Kt yy Kt − Kt
(C.13)
N
2N
∂θ ij
ij
i,j
∂K
1
1 X −1 > −1
t+1
−1
t
t+1
t
gW =
∇W L(θ , W ) =
Kt+1 yy Kt+1 − Kt+1
(C.14)
N
2N
∂W
ij
ij
i,j
1 X −1 > −1
∂Kt+1
−1
t
g̃W
=
Kt+1 yy Kt+1 − Kt+1
(C.15)
2Nt
∂W
ij
ij
i,j∈It
Note that as we have shown in Appendix B, when the kernel matrix is fixed, gW and g̃W
factorize over x and xt , respectively. Further, we denote all the mini-batches sampled before
t as x[t−1] .
t |x
t
Lemma 1 For any step t, E[δW
[t−1] ] = E[δθ | x[t−1] ] = 0.
t , we
Proof First, δθt ≡ 0, and hence E[δθt | x[t−1] ] = 0 is trivial. Next, by definition of δW
t |x
t
t
have E[δW
[t−1] ] = E[g̃W − gW | x[t−1] ]. Consider the following:
t+1
t | x
t
and W t . θ t+1 is
• Consider E[gW
[t−1] ]: gW is a deterministic function of θ
being updated deterministically using gθt−1 , and hence only depends on W t and θ t .
t
t |x
Therefore, it is independent of xt , which means that E[gW
[t−1] ] ≡ gW .
t |x
• Now, consider E[g̃W
[t−1] ]: Noting that the expectation is taken w.r.t. the empirical
distribution and Kt+1 does not depend on the current mini-batch, we can write:
X
∂Kt+1
1
−1
−1
−1
t
− Kt+1
yy> Kt+1
E[g̃W
| x[t−1] ] = E
Kt+1
2Nt
∂W
ij
ij
i,j∈It
∂K
X
(C.16)
Nt
t+1
−1
−1
> −1
=
Kt+1 yy Kt+1 − Kt+1
2N Nt
∂W
ij
ij
i,j
≡
t
gW
.
t |x
t
t
Finally, E[δW
[t−1] ] = gW − gW = 0.
In other words, semi-stochastic gradient descent that alternates between updating θ and
W computes unbiased estimates of the gradients on each step.
14. Here, we consider the gradients and their estimates scaled by the number of full data points, N , and the
mini-batch size, Nt , respectively. These constant scaling is introduced for sake of having cleaner proofs.
26
Learning Scalable Deep Kernels with Recurrent Structure
Remark 1 Note that in case if (∂Kt+1 /∂W ) is computed approximately, Lemma 1 still
t and E[g̃t | x
holds since both gW
[t−1] ] will contain exactly the same numerical errors.
W
t k2 ≤ σ 2 .
Assumption 1 For any step t, Ekδθt k2 ≤ σt2 and EkδW
t
Lemma 1 and Assumption 1 result into a stronger condition than the original assumption
given by Xu and Yin (2015). This is due to the semi-stochastic nature of the algorithm, it
simplifies the analysis, though it is not critical. Assumptions 2 and 3 are straightforwardly
adapted from the original paper.
Assumption 2 The objective function, L, is lower bounded and its partial derivatives w.r.t.
θ and W are uniformly Lipschitz with constant L > 0:
k∇θ L(θ, W ) − ∇θ L(θ̃, W )k ≤ Lkθ − θ̃k,
k∇W L(θ, W ) − ∇W L(θ, W̃ )k ≤ LkW − W̃ k.
(C.17)
Assumption 3 There exists a constant ρ such that kθ t k2 + EkW t k2 ≤ ρ2 for all t.
Lemma 2 Under Assumptions 2 and 3,
EkW t k2 ≤ ρ2 ,
Ekθ t k2 ≤ ρ2 ,
t 2
EkgW
k ≤ Mρ2 ,
Ekgθt k2 ≤ Mρ2
∀t,
where Mρ2 = 4L2 ρ2 + 2 max{∇θ L(θ 0 , W 0 ), ∇W L(θ 0 , W 0 )}.
Proof The inequalities for Ekθk2 and EkW k2 are trivial, and the ones for Ekgθt k2 and
t k2 are merely corollaries from Assumption 2.
EkgW
Negative log marginal likelihood of a Gaussian process with a structured kernel is a
nonconvex function of its arguments. Therefore, we can only show that the algorithm
converges to a stationary point, i.e., a point at which the gradient of the objective is zero.
Theorem 1 Let {(θ t , W t )} be a sequence generated from Algorithm 1 with learning rates
for θ and W being λtθ = ctθ αt and λtW = ctW αt , respectively, where ctθ , ctW , αt are positive
scalars, such that
0<
inf ct{θ,W }
t
≤
sup ct{θ,W }
t
< ∞,
αt ≥ αt+1 ,
∞
X
αt = +∞,
t=1
∞
X
t=1
αt2 < +∞,
∀t.
(C.18)
Then, under Assumptions 1 through 3 and if σ = supt σt < ∞, we have
lim Ek∇L(θ t , W t )k = 0.
t→∞
27
(C.19)
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Proof The proof is an adaptation of the one given by Xu and Yin (2015) with the following
three adjustments: (1) we have only two blocks of coordinates and (2) updates for θ are
deterministic which zeros out their variance terms, (3) the stochastic gradients are unbiased.
• From the Lipschitz continuity of ∇W L and ∇θ L (Assumption 2), we have:
≤
=
=
=
≤
L(θ t+1 , W t+1 ) − L(θ t+1 , W t )
L
t
hgW
, W t+1 − W t i + kW t+1 − W t k2
2
L t 2 t 2
t
t
t
−λW hgW , g̃W i + (λW ) kg̃W k
2
t
L
L
t 2
t 2
t
− λtW − (λtW )2 kgW
k + (λtW )2 kδW
k − λtW − L(λtW )2 hgW
, δW
i
2
2
L
L t 2
t 2
t 2
t
k + (λtW )2 kδW
k
− λW − (λW ) kgW
2
2
t
t
t
− λtW − L(λtW )2 hgW
− ∇W L(θ t , W t ), δW
i + h∇W L(θ t , W t ), δW
i
L t 2
L
t
t 2
t 2
t
− λW − (λW ) kgW
k + (λtW )2 kδW
k − λtW − L(λtW )2 h∇W L(θ t , W t ), δW
i
2
2
L
t 2
k + kgθt k2 ),
+ λtθ λtW + L(λtW )2 (kδW
2
t := ∇ L(θ t+1 , W t )):
where the last inequality comes from the following (note that gW
W
≤
t
− λtW − L(λtW )2 h∇W L(θ t+1 , W t ) − ∇W L(θ t , W t ), δW
i
t
λtW − L(λtW )2 kδW
kk∇W L(θ t+1 , W t ) − ∇W L(θ t , W t )k
t
≤ Lλtθ λtW − L(λtW )2 kδW
kkgθt k
L t t
t 2
k + kgθt k2 ).
≤
λθ λW + L(λtW )2 (kδW
2
Analogously, we derive a bound on L(θ t+1 , W t ) − L(θ t , W t ):
L(θ t+1 , W t ) − L(θ t , W t )
L t 2
L
t
≤ − λθ − (λθ ) kgθt k2 + (λtθ )2 kδθt k2 − λtθ − L(λtθ )2 h∇θ L(θ t , W t ), δθt i
2
2
• Since (θ t , W t ) is independent from the current mini-batch, xt , using Lemma 1, we
t i = 0.
have that Eh∇θ L(θ t , W t ), δθt i = 0 and Eh∇W L(θ t , W t ), δW
28
Learning Scalable Deep Kernels with Recurrent Structure
Summing the two obtained inequalities and applying Assumption 1, we get:
E[L(θ t+1 , W t+1 ) − L(θ t , W t )]
L
L t 2
L
t 2
t
k + (λtW )2 σ + λtθ λtW + L(λtW )2 (σ + Ekgθt k2 )
≤ − λW − (λW ) EkgW
2
2
2
L
L
− λtθ − (λtθ )2 Ekgθt k2 + (λtθ )2 σ
2
2
L
L
t 2
≤ − cαt − C 2 αt2 EkgW
k + LC 2 σαt + C 2 αt2 (1 + LCαt ) (σ + Ekgθt k2 )
2
2
L
L
− cαt − C 2 αt2 Ekgθt k2 + C 2 αt2 σ,
2
2
where we denoted c = min{inf t ctW , inf t ctθ }, C = max{supt ctW , supt ctθ }. Note that
t k2 ≤ M 2 . Therefore, summing
from Lemma 2, we also have Ekgθt k2 ≤ Mρ2 and EkgW
ρ
the right hand side of the final inequality over t and using (C.18), we have:
lim EL(θ t+1 , W t+1 ) − EL(θ 0 , W 0 ) ≤ −c
t→∞
∞
X
t=1
t 2
αt Ekgθt k2 + EkgW
k .
(C.20)
Since the objective function is lower bounded, this effectively means:
∞
X
t=1
αt Ekgθt k2
< ∞,
∞
X
t=1
t 2
αt EkgW
k < ∞.
• Finally, using Lemma 2, our assumptions and Jensen’s inequality, it follows that
t+1 2
t 2
EkgW
k − EkgW
k ≤ 2LMρ Cαt
q
2(Mρ2 + σ 2 ).
t k2 → 0
According to Proposition 1.2.4 of (Bertsekas, 1999), we have Ekgθt k2 → 0 and EkgW
as t → ∞, and hence
t
Ek∇L(θ t , W t )k ≤ Ek∇W L(θ t , W t ) − gθt k + Ekgθt k + Ek∇W L(θ t , W t ) − gθt k + EkgW
k
q
t
≤ 2LC 2(Mρ2 + σ 2 )α + Ekgθt k + EkgW
k → 0 as t → ∞,
where the first term of the last inequality follows from Lemma 2 and Jensen’s inequality.
29
Al-Shedivat, Wilson, Saatchi, Hu, Xing
C.2 Semi-stochastic gradient with delayed kernel matrix updates
We show that given a bounded delay on the kernel matrix updates, the algorithm is still
t and applying the
convergent. Our analysis is based on computing the change in δθt and δW
same argument as in Theorem 1. The only difference is that we need to take into account
the perturbations of the kernel matrix due to the introduced delays, and hence we have to
impose certain assumptions on its spectrum.
Assumption 4 Recurrent transformations, φW (x̄), is L-Lipschitz w.r.t. W for all x̄ ∈ X L :
kφW̃ (x̄) − φW (x̄)k ≤ LkW̃ − W k.
Assumption 5 The kernel function, k(·, ·), is uniformly G-Lipschitz and its first partial
derivatives are uniformly J-Lipschitz:
k(h̃1 , h2 ) − k(h1 , h2 ) ≤ Gkh̃1 − h1 k, k(h1 , h̃2 ) − k(h1 , h2 ) ≤ Gkh̃2 − h2 k,
∂1 k(h1 , h2 ) − ∂1 k(h̃1 , h2 ) ≤ Jkh̃1 − h1 k, ∂2 k(h1 , h2 ) − ∂2 k(h1 , h̃2 ) ≤ Jkh̃2 − h2 k.
Assumption 6 For any collection of data representations, {hi }N
i=1 , the smallest eigenvalue
of the corresponding kernel matrix, K, is lower bounded by a positive constant γ > 0.
Note that not only the assumptions are relevant to practice, Assumptions 5 and 6 can be
also controlled by choosing the class of kernel functions used in the model. For example, the
smallest eigenvalue of the kernel matrix, K, can be controlled by the smoothing properties
of the kernel (Rasmussen and Williams, 2006).
Consider a particular stochastic step of Algorithm 2 at time t for a given mini-batch, xt ,
assuming that the kernel was last updated τ steps ago. The stochastic gradient will take
the following form:
∂K
1
1 X −1
t−τ
−1
> −1
t
t+1
t
ĝW =
Kt−τ yy Kt−τ − Kt−τ
.
∇W L(θ , W , Kt−τ ) =
Nt
2Nt
∂W
ij
ij
i,j∈It
(C.21)
t = ĝt − gt and uniformly bound kδ̂ t − δ t k in order to enable the same
We can define δ̂W
W
W
W
W
argument as in Theorem 1. To do that, we simply need to understand effect of perturbation
t .
of the kernel matrix on g̃W
Lemma 3 Under the given assumptions, the following bound holds for all i, j = 1, . . . , N :
−1
−1
−1
Kt−τ
yy> Kt−τ
− Kt−τ
ij
− Kt−1 yy> Kt−1 − Kt−1
√ −1
where D = γ −1 + γ − 2GLτ λσ N
.
30
ij
≤ D2 kyk2 + D,
(C.22)
Learning Scalable Deep Kernels with Recurrent Structure
−1
and Kt−1 is simply due to that the former has been
Proof The difference between Kt−τ
t−τ
computed for W
and the latter for W t . To prove the bound, we need multiple steps.
• First, we need to bound the element-wise difference between Kt−τ and Kt . This is
done by using Assumptions 4 and 5 and the triangular inequality:
|(Kt )ij − (Kt−τ )ij | ≤ G (kφW t (x̄i ) − φW t−τ (x̄i )k + kφW t (x̄j ) − φW t−τ (x̄j )k)
≤ 2GLkW t − W t−τ k
τ
X
+s t−τ +s
= 2GLk
λt−τ
ĝW
k
W
s=1
≤ 2GLτ λσ
• Next, since each element of the perturbed matrix is bounded by 2GLτ λσ, we can
bound its spectral norm as follows:
√
kKt − Kt−τ k ≤ kKt − Kt−τ kF ≤ 2GLτ λσ N ,
which means that√the minimal singular value of the perturbed matrix is at least
σ1 = γ − 2GLτ λσ N due to Assumption 6.
• The spectral norm of the expression of interest can be bounded (quite pessimistically!)
by summing up together the largest eigenvalues of the matrix inverses:
−1
−1
−1
Kt−τ
yy> Kt−τ
− Kt−τ
− Kt−1 yy> Kt−1 − Kt−1
ij
ij
−1
−1
yy> Kt−τ
− Kt
− Kt−τ
− Kt
√ −1 2
√ −1
≤ γ −1 + γ − 2GLτ λσ N
kyk2 + γ −1 + γ − 2GLτ λσ N
.
≤
−1
−1
−1
Kt−τ
− Kt
−1
Each element of a matrix is bounded by the largest eigenvalue.
Using Lemma 3, it is straightforward to extend Theorem 1 to Algorithm 2.
Theorem 2 Let {(θ t , W t )} be a sequence generated from Algorithm 2 with learning rates
for θ and W being λtθ = ctθ αt and λtW = ctW αt , respectively, where ctθ , ctW , αt are positive
scalars, such that
0 < inf ct{θ,W } ≤ sup ct{θ,W } < ∞,
t
t
αt ≥ αt+1 ,
∞
X
αt = +∞,
t=1
∞
X
t=1
αt2 < +∞,
∀t.
(C.23)
Then, under Assumptions 1 through 6 and if σ = supt σt < ∞, we have
lim Ek∇L(θ t , W t )k = 0.
t→∞
31
(C.24)
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Proof The proof is identical to the proof of Theorem 1. The only difference is in the
t k ≤ σ + 2N D(Dkyk2 + 1)Jρ,
following upper bound on the expected gradient error: EkδW
where D is as given in Lemma 3.
Remark 1 Even though the provided bounds are crude due to pessimistic estimates of the
perturbed kernel matrix spectrum15 , we still see a fine balance between the delay, τ , and the
learning rate, λ, as given in the expression for the D constant.
Appendix D. Details on the datasets
The datasets varied in the number of time steps (from hundreds to a million), input and
output dimensionality, and the nature of the estimation tasks.
D.1 Self-driving car
The following is description of the input and target time series used in each of the autonomous
driving tasks (dimensionality is given in parenthesis).
• Car speed estimation:
– Features: GPS velocity (3), fiber gyroscope (3).
– Targets: speed measurements from the car speedometer.
• Car yaw estimation:
– Features: acceleration (3), compass measurements (3).
– Targets: yaw in the car-centric frame.
• Lane sequence prediction: Each lane was represented by 8 cubic polynomial
coefficients [4 coefficients for x (front) and 4 coefficients for y (left) axes in the carcentric frame]. Instead of predicting the coefficients (which turned out to lead to
overall less stable results), we discretized the lane curves using 7 points (initial, final
and 5 equidistant intermediate points).
– Features: lanes at a previous time point (16), GPS velocity (3), fiber gyroscope
(3), compass (3), steering angle (1).
– Targets: coordinates of the lane discretization points (7 points per lane resulting
in 28 total output dimensions).
15. Tighter bounds can be derived by inspecting the effects of perturbations on specific kernels as well as
using more specific assumptions about the data distribution.
32
Learning Scalable Deep Kernels with Recurrent Structure
Table 5: Summary of the feedforward and recurrent neural architectures and the corresponding
hyperparameters used in the experiments. GP-based models used the same architectures as their
non-GP counterparts. Activations are given for the hidden units; vanilla neural nets used linear
output activations.
Name
NARX
RNN
LSTM
∗
Time lag
Layers
Units∗
Actuator
Drives
GEF-power
GEF-wind
Car
32
16
48
48
128
1
1
1
1
1
256
128
256
16
128
Actuator
Drives
GEF-power
GEF-wind
Car
32
16
48
48
128
1
1
1
1
1
64
64
16
32
128
Actuator
Drives
GEF-power
GEF-wind
Car
32
16
48
48
128
1
1
2
1
2
256
128
256
64
64
Data
Type
Regularizer∗∗
Optimizer
ReLU
dropout(0.5)
Adam(0.01)
tanh
dropout(0.25),
rec_dropout(0.05)
Adam(0.01)
tanh
dropout(0.25),
rec_dropout(0.05)
Adam(0.01)
Each layer consisted of the same number of units given in the table.
rec_dropout denotes the dropout rate of the recurrent weights (Gal and Ghahramani, 2016b).
∗∗
• Estimation of the nearest front vehicle position: (x, y) coordinates in the
car-centric frame.
– Features: x and y at a previous time point (2), GPS velocity (3), fiber gyroscope
(3), compass (3), steering angle (1).
– Targets: x and y coordinates.
Appendix E. Neural architectures
Details on the best neural architectures used for each of the datasets are given in Table 5.
33
Al-Shedivat, Wilson, Saatchi, Hu, Xing
References
Alekh Agarwal and John C Duchi. Distributed delayed stochastic optimization. In Advances
in Neural Information Processing Systems, pages 873–881, 2011.
Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with
gradient descent is difficult. Neural Networks, IEEE Transactions on, 5(2):157–166, 1994.
Dimitri P Bertsekas. Nonlinear programming. Athena scientific Belmont, 1999.
George EP Box, Gwilym M Jenkins, Gregory C Reinsel, and Greta M Ljung. Time series
analysis: forecasting and control. John Wiley & Sons, 1994.
Roberto Calandra, Jan Peters, Carl Edward Rasmussen, and Marc Peter Deisenroth. Manifold gaussian processes for regression. In Neural Networks (IJCNN), 2016 International
Joint Conference on, pages 3338–3345. IEEE, 2016.
Tianqi Chen, Emily Fox, and Carlos Guestrin. Stochastic gradient hamiltonian monte carlo.
In International Conference on Machine Learning, pages 1683–1691, 2014.
Andreas C Damianou and Neil D Lawrence. Deep gaussian processes. In AISTATS, pages
207–215, 2013.
Marc Peter Deisenroth and Jun Wei Ng. Distributed gaussian processes. In International
Conference on Machine Learning (ICML), volume 2, page 5, 2015.
Roger Frigola, Yutian Chen, and Carl Rasmussen. Variational gaussian process state-space
models. In Advances in Neural Information Processing Systems, pages 3680–3688, 2014.
Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing
model uncertainty in deep learning. In Proceedings of the 33rd International Conference
on Machine Learning, pages 1050–1059, 2016a.
Yarin Gal and Zoubin Ghahramani. A theoretically grounded application of dropout in
recurrent neural networks. In Advances in Neural Information Processing Systems, pages
1019–1027, 2016b.
Alan Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. Speech recognition with deep
recurrent neural networks. In Acoustics, Speech and Signal Processing (ICASSP), 2013
IEEE International Conference on, pages 6645–6649. IEEE, 2013.
Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability
of stochastic gradient descent. In Proceedings of The 33rd International Conference on
Machine Learning, pages 1225–1234, 2016.
34
Learning Scalable Deep Kernels with Recurrent Structure
James Hensman, Nicolò Fusi, and Neil D Lawrence. Gaussian processes for big data. In
Proceedings of the Twenty-Ninth Conference on Uncertainty in Artificial Intelligence,
pages 282–290. AUAI Press, 2013.
Geoffrey E Hinton and Ruslan R Salakhutdinov. Using deep belief nets to learn covariance
kernels for gaussian processes. In Advances in neural information processing systems,
pages 1249–1256, 2008.
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9
(8):1735–1780, 1997.
Tommi S Jaakkola and David Haussler. Exploiting generative models in discriminative
classifiers. Advances in neural information processing systems, pages 487–493, 1999.
Robert Keys. Cubic convolution interpolation for digital image processing. IEEE transactions
on acoustics, speech, and signal processing, 29(6):1153–1160, 1981.
Juš Kocijan, Agathe Girard, Blaž Banko, and Roderick Murray-Smith. Dynamic systems
identification with gaussian processes. Mathematical and Computer Modelling of Dynamical
Systems, 11(4):411–424, 2005.
John Langford, Alex J Smola, and Martin Zinkevich. Slow learners are fast. Advances in
Neural Information Processing Systems, 22:2331–2339, 2009.
Miguel Lázaro-Gredilla. Bayesian warped gaussian processes. In Advances in Neural
Information Processing Systems, pages 1619–1627, 2012.
Tsungnam Lin, Bil G Horne, Peter Tiňo, and C Lee Giles. Learning long-term dependencies in
narx recurrent neural networks. Neural Networks, IEEE Transactions on, 7(6):1329–1338,
1996.
David JC MacKay. Introduction to gaussian processes. NATO ASI Series F Computer and
Systems Sciences, 168:133–166, 1998.
César Lincoln C Mattos, Zhenwen Dai, Andreas Damianou, Jeremy Forth, Guilherme A Barreto, and Neil D Lawrence. Recurrent gaussian processes. arXiv preprint arXiv:1511.06644,
2015.
Charles A Micchelli, Yuesheng Xu, and Haizhang Zhang. Universal kernels. The Journal of
Machine Learning Research, 7:2651–2667, 2006.
Joaquin Quinonero-Candela and Carl Edward Rasmussen. A unifying view of sparse
approximate gaussian process regression. The Journal of Machine Learning Research, 6:
1939–1959, 2005.
35
Al-Shedivat, Wilson, Saatchi, Hu, Xing
Carl Edward Rasmussen and Zoubin Ghahramani. Occam’s razor. Advances in neural
information processing systems, pages 294–300, 2001.
Carl Edward Rasmussen and Christopher KI Williams. Gaussian processes for machine
learning. The MIT Press, 2006.
Stéphane Ross, Geoffrey J Gordon, and Drew Bagnell. A reduction of imitation learning
and structured prediction to no-regret online learning. In AISTATS, volume 1, page 6,
2011.
Yunus Saatchi and Andrew Gordon Wilson. Bayesian gan. arXiv preprint arXiv:1705.09558,
2017.
Bernhard Schölkopf and Alexander J Smola. Learning with kernels: support vector machines,
regularization, optimization, and beyond. MIT press, 2002.
Jonas Sjöberg, Qinghua Zhang, Lennart Ljung, Albert Benveniste, Bernard Delyon, PierreYves Glorennec, Håkan Hjalmarsson, and Anatoli Juditsky. Nonlinear black-box modeling
in system identification: a unified overview. Automatica, 31(12):1691–1724, 1995.
Edward Snelson, Carl Edward Rasmussen, and Zoubin Ghahramani. Warped gaussian
processes. In NIPS, pages 337–344, 2003.
Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. The Journal of
Machine Learning Research, 15(1):1929–1958, 2014.
ML Stein. Interpolation of spatial data, 1999.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural
networks. In Advances in neural information processing systems, pages 3104–3112, 2014.
Ryan D Turner, Marc P Deisenroth, and Carl E Rasmussen. State-space inference and
learning with gaussian processes. In International Conference on Artificial Intelligence
and Statistics, pages 868–875, 2010.
Peter Van Overschee and Bart De Moor. Subspace identification for linear systems: Theory
— Implementation — Applications. Springer Science & Business Media, 2012.
Jack Wang, Aaron Hertzmann, and David M Blei. Gaussian process dynamical models. In
Advances in neural information processing systems, pages 1441–1448, 2005.
Torbjörn Wigren. Input-output data sets for development and benchmarking in nonlinear
identification. Technical Reports from the department of Information Technology, 20:
2010–020, 2010.
36
Learning Scalable Deep Kernels with Recurrent Structure
Andrew Wilson and Zoubin Ghahramani. Copula processes. In Advances in Neural Information Processing Systems, pages 2460–2468, 2010.
Andrew Gordon Wilson and Ryan Prescott Adams. Gaussian process kernels for pattern
discovery and extrapolation. In Proceedings of the 30th International Conference on
Machine Learning (ICML-13), pages 1067–1075, 2013.
Andrew Gordon Wilson and Hannes Nickisch. Kernel interpolation for scalable structured
gaussian processes (kiss-gp). In Proceedings of The 32nd International Conference on
Machine Learning, pages 1775–1784, 2015.
Andrew Gordon Wilson, Christoph Dann, and Hannes Nickisch. Thoughts on massively
scalable gaussian processes. arXiv preprint arXiv:1511.01870, 2015.
Andrew Gordon Wilson, Zhiting Hu, Ruslan Salakhutdinov, and Eric P Xing. Deep kernel
learning. In Proceedings of the 19th International Conference on Artificial Intelligence
and Statistics, 2016a.
Andrew Gordon Wilson, Zhiting Hu, Ruslan R Salakhutdinov, and Eric P Xing. Stochastic
variational deep kernel learning. In Advances in Neural Information Processing Systems,
pages 2586–2594, 2016b.
D Randall Wilson and Tony R Martinez. The general inefficiency of batch training for
gradient descent learning. Neural Networks, 16(10):1429–1451, 2003.
Yangyang Xu and Wotao Yin. Block stochastic gradient iteration for convex and nonconvex
optimization. SIAM Journal on Optimization, 25(3):1686–1716, 2015.
37
| 2cs.AI
|
LightDP: Towards Automating Differential Privacy Proofs
Danfeng Zhang
Daniel Kifer
arXiv:1607.08228v3 [cs.PL] 9 Dec 2016
Department of Computer Science and Engineering
Penn State University
University Park, PA United States
{zhang,[email protected]}
Abstract
The growing popularity and adoption of differential privacy in
academic and industrial settings has resulted in the development
of increasingly sophisticated algorithms for releasing information
while preserving privacy. Accompanying this phenomenon is the
natural rise in the development and publication of incorrect algorithms, thus demonstrating the necessity of formal verification
tools. However, existing formal methods for differential privacy
face a dilemma: methods based on customized logics can verify
sophisticated algorithms but come with a steep learning curve and
significant annotation burden on the programmers, while existing
programming platforms lack expressive power for some sophisticated algorithms.
In this paper, we present LightDP, a simple imperative language
that strikes a better balance between expressive power and usability.
The core of LightDP is a novel relational type system that separates
relational reasoning from privacy budget calculations. With dependent types, the type system is powerful enough to verify sophisticated algorithms where the composition theorem falls short. In
addition, the inference engine of LightDP infers most of the proof
details, and even searches for the proof with minimal privacy cost
bound when multiple proofs exist. We show that LightDP verifies
sophisticated algorithms with little manual effort.
Categories and Subject Descriptors D.3.1 [Programming Languages]: Formal Definitions and Theory; D.2.4 [Software Engineering]: Software/Program Verification; F.3.1 [Logics and
Meanings of Programs]: Specifying and Verifying and Reasoning
about Programs.
Keywords Differential privacy; dependent types; type inference;
1.
Introduction
Companies, government agencies, and academics are interested in
analyzing and modeling datasets containing sensitive information
about individuals (e.g., medical records, customer behavior, etc.).
Privacy concerns can often be mitigated if the algorithms used to
manipulate the data, answer queries, and build statistical models
satisfy differential privacy (Dwork et al. 2006b) — a set of restrictions on their probabilistic behavior that provably limit the ability
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without
fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice
and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be
honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to
lists, requires prior specific permission and/or a fee. Request permissions from [email protected] or Publications
Dept., ACM, Inc., fax +1 (212) 869-0481.
POPL ’17, January 18-20, 2017, Paris, France
Copyright c 2017 ACM 978-1-4503-4660-3/17/01. . . $15.00
DOI: http://dx.doi.org/10.1145/3009837.3009884
of attackers to infer individual-level sensitive information (Dwork
et al. 2006b; Kifer and Machanavajjhala 2014).
Since 2006, differential privacy has seen explosive growth in
many areas, including theoretical computer science, databases, machine learning, and statistics. This technology has been deployed in
practice, starting with the U.S. Census Bureau LEHD OnTheMap
tool (Machanavajjhala et al. 2008), the Google Chrome Browser
(Erlingsson et al. 2014), and Apple’s new data collection efforts
(Greenberg 2016). However, the increase in popularity and usage
of differential privacy has also been accompanied by a corresponding increase in the development and implementation of algorithms
with flawed proofs of privacy; for example, Chen and Machanavajjhala (2015) and Lyu et al. (2016) catalog some recent cases about
variations of the Sparse Vector method (Dwork and Roth 2014) .
Currently, there are two strategies for combating this trend. The
first is the use of programming platforms (McSherry 2009; Mohan
et al. 2012; Roy et al. 2010) that have privacy primitives that restrict
the privacy-preserving algorithms that can be implemented and often add more noise than is necessary to the computation. The second strategy is the development of languages and formal verification tools for differential privacy (Reed and Pierce 2010; Gaboardi
et al. 2013; Barthe et al. 2012, 2014, 2016c,b). These languages enable the development of much more sophisticated algorithms that
use less noise and hence provide more accurate outputs. However,
the increased power of the formal methods comes with a considerable cost — a programmer has to heavily annotate code and generate proofs using complicated logics such as a customized relational
Hoare logic proposed by Barthe et al. (2012). Moreover, intricate
proof details have to be provided by a programmer, which makes
exploring variations of an algorithm difficult since small variations
in code can cause significant changes to a proof.
In this paper, we present LightDP, a language for developing
provably privacy-preserving algorithms. The goal of LightDP is to
minimize the burden on the programmer while retaining most of
the capabilities of the state-of-the-art, such as verifying the Sparse
Vector method (Dwork and Roth 2014) (an algorithm which, until
very recently (Barthe et al. 2016c,b), was beyond the capabilities
of verification tools). For example, we show that the Sparse Vector
method can be verified in LightDP with little manual effort: just
two lines of annotation from the programmer.
LightDP is equipped with a novel light-weight relational type
system that clearly separates relational reasoning from privacy budget calculation. In particular, it transforms the original probabilistic program into an equivalent nonprobabilistic program, where all
privacy costs become explicit. With dependent types, the explicitly calculated privacy cost in the target language may depend on
program states, hence enabling the verification of sophisticated algorithms (e.g., the Sparse Vector method) that are beyond the capability of many existing methods (Reed and Pierce 2010; Gaboardi
et al. 2013; Barthe et al. 2012, 2014) based on the composition the-
orem (McSherry 2009). Moreover, the transformed nonprobabilistic program is ready for off-the-shelf formal verification methods,
such as Hoare logic, to provide an upper bound of the privacy cost.
On the usability end, LightDP has an inference engine that reduces the already low annotation burden on the programmers. Although the inference engine does not yet automate the privacy budget calculation part of a proof, it does fill in missing details in the relational reasoning part of a proof; furthermore, based on MaxSMT
theory, it even searches for the optimal proof that minimizes privacy
cost with minimal human involvement. For example, with only one
postcondition annotation and one loop invariant annotation from a
programmer, LightDP confirms that the proof in (Dwork and Roth
2014) indeed provides the minimal privacy cost.
To summarize, this paper makes the following contributions:
1. LightDP, a new imperative language for verifying sophisticated
privacy-preserving algorithms (Section 3.1),
2. expressive static annotations incorporating dependent types, enabling precise tracking of privacy costs (Section 3.3),
3. a formal proof that the LightDP type system soundly tracks
differential privacy costs, and new proof techniques involved
in the soundness proof (Section 4),
4. an inference engine that automatically fills in missing details
involved in the relational reasoning part of a proof, and further,
minimizes provable privacy cost bound when multiple proofs
exist, with little manual effort (Section 5),
5. case studies on complex algorithms showing that formal verification of privacy-preserving algorithms are viable with little
programmer annotation burden (Section 6).
2.
Preliminaries and Illustrating Example
2.1
Distributions
We define the set of sub-distributions over a discrete set A, written Dist(A), as the set of functions µ : A → [0, 1], such that
P
a ≤ 1. When applied to an event E ⊆ A, we define
a∈A µ P
P
µ(E) , e∈E µ(e). Notice that we do not require a∈A µ a =
1, a special case when µ is a distribution, since sub-distribution
gives rise to an elegant semantics for programs that may not terminate (Kozen 1981).
Given a distribution µ ∈ Dist(A), its support is defined as
support(µ) , {a | µ(a) > 0}. We use 1a to represent the
degenerate distribution µ that µ(a) = 1 and µ(a0 ) = 0 if a0 6= a.
Moreover, sub-distributions can be given a structure of a monad.
Formally, we define the unit and bind functions as follows:
unit : A → Dist(A) , λa. 1a
bind : Dist(A) → (A → Dist(B)) → Dist(B)
X
, λµ. λf. (λb.
(f a b) × µ(a))
a∈A
That is, unit takes an element in A and returns the Dirac distribution where all mass is assigned to a; bind takes µ, a distribution
on A, and f , a mapping from A to distributions on B (e.g., a conditional distribution of B given A), and returns the corresponding
marginal distribution on B. This monadic view will avoid cluttered
definitions and proofs when probabilistic programs are involved
(Section 3.2).
2.2
Differential Privacy
Differential privacy has two major variants: pure (Dwork et al.
2006b) (obtained by setting δ = 0 in the following definition) and
approximate (Dwork et al. 2006a) (obtained by choosing a δ > 0).
function S PARSE V ECTOR (T, N, : num0 ; q : list num∗ )
returns (out : list bool0 )
precondition ∀i. −1 ≤ (b
q [i]) ≤ 1
c1, c2, i : num0 ; T̃ , η1 : num1 ; η2 : numq[i]+η2 ≥T̃ ?2:0
1
2
3
4
5
6
7
8
9
10
11
12
η1 := Lap (2/);
T̃ := T + η1 ;
c1 := 0; c2 := 0; i := 0;
while (c1 < N )
η2 := Lap (4N/);
if (q[i] + η2 ≥ T̃ ) then
out:= true::out;
c1 := c1 + 1;
else
out:= false::out;
c2 := c2 + 1;
i := i+1;
Figure 1. The Sparse Vector method. Type annotations are shown
in grey. The precondition specifies the adjacency assumption.
Definition 1 (Differential privacy). Let , δ ≥ 0. A probabilistic
computation M : A → Dist(B) is (, δ)-differentially private
with respect to an adjacency relation Ψ ⊆ A × A if for every pair
of inputs a1 , a2 ∈ A such that a1 Ψa2 , and every output subset
E ⊆ B, we have
P (M a1 ∈ E) ≤ exp()P (M a2 ∈ E) + δ
Intuitively, a probabilistic computation satisfies differential privacy if it produces similar distributions for any pair of inputs related
by Ψ. In the most common applications of differential privacy, A is
the set of possible databases and the adjacency relation Ψ is chosen
so that a1 Ψa2 whenever a1 can be obtained from a2 by adding or
removing data belonging to a single individual.
In this paper, we focus on the verification of algorithms that satisfy pure differential privacy. An algorithm is -differentially private iff it is (, 0)-differentially private according to Definition 1.
2.3
The Sparse Vector Method
The goal of formal methods for verifying -differential privacy is
to provide an upper bound on the privacy cost of a program.
Typically, users will have a fixed privacy budget 0 and can only
run programs whose provable privacy cost does not exceed the
budget: ≤ 0 . For this reason, it is important that formal methods
are able to prove a tight upper bound on the privacy cost.
With the exception of (Barthe et al. 2016c,b), most existing formal methods rely on the composition theorem (McSherry 2009).
That is, in the case of -differential privacy, those methods essentially treat a program as a series of modules, each with a provable
upper bound i on its privacy cost. Then by the compositionP
theorem (McSherry 2009), the total privacy cost is bounded by i i .
However, for sophisticated advanced algorithms, the composition
theorem often falls short — it can provide upper bounds that are
arbitrarily larger than the true privacy cost. Providing the tightest
privacy cost for intricate algorithms requires formal methods that
are more powerful but avoid over-burdening the programmers with
annotation requirements. To illustrate these challenges, we consider the Sparse Vector method (Dwork and Roth 2014). It is a
prime example of the need for formal methods because many of
its published variants have been shown to be incorrect (Chen and
Machanavajjhala 2015; Lyu et al. 2016).
The Sparse Vector method also has many correct variants, one
of which is shown in Figure 1. For now, safely ignore the type annotations in grey and the precondition. Here, the input list q repre-
sents a sequence of results of counting queries q1 , q2 , q3 , . . . (e.g.,
how many patients in the data have cancer, how many patients contracted an infection in the hospital, etc.) running on a database. The
goal is to answer as accurately as possible the following question:
which queries, when evaluated on the true database, return an answer greater than the threshold T (a program input unrelated to the
sensitive data)?
To achieve differential privacy, the algorithm adds appropriate
Laplace noise to the threshold and to each query. Here, Lap (4N/)
draws one sample from the Laplace distribution with mean zero
and a scale factor (4N/). If the noisy query answer (q[i] + η2 )
is above the noisy threshold T̃ , it adds true to the output list out
(in the slot reserved for that query) and otherwise, it adds false.
The key to this algorithm is the deep observation that once noise
has been added to the threshold, queries for which we output true
have a privacy cost (so we can answer at most N of them, where
N is a parameter); however, outputting false for a query does
not introduce any new privacy costs (Dwork and Roth 2014). The
algorithm ensures that the total privacy cost is bounded by the input
, the parameter used in Figure 1. This remarkable property makes
the Sparse Vector method ideal in situations where the vast majority
of query counts are expected to be below the threshold.
Failure of the composition theorem If we just use the composition theorem, we would have a privacy cost of /4N for each loop
iteration (i.e., every time Lap (4N/) noise is added to a query
answer) due to the property of the Laplace distribution. Since the
number of iterations are not a priori bounded, the composition theorem could not prove that the algorithm satisfies 0 -differential privacy for any finite 0 ; more advanced methods are needed.
Informal proof and sample runthrough Proofs of correctness
(of this and other variants) can be found in (Dwork and Roth
2014; Chen and Machanavajjhala 2015; Lyu et al. 2016). Here we
provide an informal correctness argument by example to illustrate
the subtleties involved both in proving it and inferring a tight bound
for the algorithm.
Suppose we set the parameters T = 4 (we want to know which
queries have a value at least 4) and N = 1 (we stop the algorithm
after the first time it outputs true). Consider the following two
databases D1 , D2 that differ on one record, and their corresponding
query answers:
D1 :
q[0] = 2,
q[1] = 3,
q[2] = 5
D2 :
q[0] = 3,
q[1] = 3,
q[2] = 4
Suppose in one execution on D1 , the noise added to T is α(1) = 1
(1)
(1)
(1)
and the noise added to q[0], q[1], q[2] is β0 = 2, β1 = 0, β2 =
0, respectively. Thus the noisy threshold is T̃ = 5 and the noisy
(1)
(1)
(1)
query answers are q[0]+β0 = 4, q[1]+β1 = 3, q[2]+β2 = 5
and so the algorithm outputs the sequence: (false, false, true).
According to Definition 1, for any output sequence ω, we
need to show P (M (D1 ) = ω) ≤ e P (M (D2 ) = ω) for
all possible outputs ω and databases D1 , D2 that differ on one
record. For the databases D1 , D2 described above, we will show
that P (M (D1 ) = (false, false, true)) ≤ e P (M (D2 ) =
(false, false, true)). We proceed in two steps.
Aligning randomness We first create an injective (but not necessarily bijective) function from the randomness in the execution under D1 into the randomness in the execution under D2 , so that both
executions generate the same output. For an execution under D2 , let
(2)
(2)
(2)
α(2) be the noise added to the threshold and let β0 , β1 , β2 be
the noise added to the queries q[0], q[1], q[2], respectively. Consider
an injective function candidate that adds 1 to the threshold noise
(i.e., α(2) = α(1) + 1), keeps the noise of queries for which D1 re-
(2)
(1)
(2)
(1)
ported false (i.e., β0 = β0 and β1 = β1 ) and adds 2 to the
(2)
(1)
noise of queries for which D1 reported true (i.e., β2 = β2 +2).
In our running example, execution under D2 with this function
would result in the noisy threshold T̃ = 6 and noise query answers
(2)
(2)
(2)
q[0] + β0 = 5, q[1] + β1 = 3, q[2] + β2 = 6. Hence, the
output once again is (false, false, true). In fact, it is easy to see
that under this injective function, every execution under D1 would
result in an execution under D2 that produces the same answer.
Counting privacy cost For each output ω, let f be the injective
function; let A be the set of random variable assignments that
cause execution under D1 to produce ω; let B be the possible
assignments we can get by applying f to A; and let C be the
set of random variable assignments that are not in the range of f ,
but nevertheless cause execution under D2 to produce the output
ω as well1 . Then we can rewrite P (M (D1 ) = ω) = P (A) and
P (M (D2 ) = ω) = P (B) + P (C).
Once we have done the alignment of randomness as above, and
recalling that N = 1 in our example, the proof finishes by showing:
X
X
P (A) =
P (a) ≤ e 2 e2 4N
P (f (a))
a∈A
a∈A
= e P (B) ≤ e (P (B) + P (C))
where the e 2 factor results from using a threshold value that is 1
larger, while the e2 4N factor results from adding 2 to the noise for
query q[2]. Notice that no privacy cost is paid for queries q[0] and
q[1], since the same noise is added
P under D1 and D2 . Moreover,
due to the injective assumption, a∈A P (f (a)) = P (B).
Challenges The Sparse Vector method is a prime example of the
need for formal methods, since paper-and-pencil proof is shown
to be error-prone for its variants. The intricacy in its proof brings
major challenges for formal methods:
1. Precision: a crucial observation in the proof is that once noise
has been added to the threshold, different privacy costs are only
paid for outputting true. Hence, the cost calculation needs to
consider program states.
2. Aligning randomness: finding an injective function from the
randomness under D1 to that under D2 such that outputting ω
under D1 entails outputting ω under D2 is the most intriguing
piece in the proof. However, coming up with a correct function,
as we described informally above, is non-trivial.
3. Finding the tightest bound: in fact, an infinite number of proofs
exist for the Sparse Vector method, though with various provable privacy costs2 . Since a tighter privacy cost bound allows a
privacy-preserving algorithm to produce more accurate outputs,
a formal method should produce the tightest one when possible.
Except the very recent work by Barthe et al. (2016c,b), existing
formal methods (e.g., (Reed and Pierce 2010; Gaboardi et al. 2013;
Barthe et al. 2012, 2014)) rely on the composition theorem, hence
fail to prove that the Sparse Vector method satisfies 0 -privacy for
any 0 . Recent work by Barthe et al. (2016c,b) verifies variants of
the Sparse Vector method, but its customized relational logic incurs heavy annotation burden, including the randomness alignment.
Moreover, those work cannot search for the tightest cost bound.
1 This
is possible since we do not assume the function to be a bijection.
example, another injective function adds 2 to the threshold noise (i.e.,
α(2) = α(1) + 2), keeps the noise of queries for which D1 reported false
(2)
(1)
(2)
(1)
(i.e., β0 = β0 and β1 = β1 ) and adds 3 to the noise of queries for
2 For
(2)
(1)
which D1 reported true (i.e., β2 = β2 + 3). It is easy to check that
this mapping has the desired property, but the privacy cost is (7)/4.
2.4
Our Approach
To tackle the challenges above, we propose LightDP, an imperative
language that enables verification and even inference of the tightest
privacy cost for sophisticated privacy-preserving algorithms. We
illustrate the key components of LightDP in this section, and detail
all components in the rest of this paper.
Relational reasoning The core of LightDP is a novel light-weight
dependent type system that explicitly captures the exact difference of a variable’s values in two executions under two adjacent
databases. Let v (1) (v (2) ) be the value of a variable x in an execution under D1 (D2 ). The type τ for x in LightDP has the form
of Bd , meaning that x holds a value of basic type B (e.g., int,
real, bool), and v (1) + d = v (2) . Required type annotations for
the Sparse Vector method are shown in grey in Figure 1. Note that
for brevity, we write num for numeric base types (e.g., int, real).
Hereafter, we refer to the d counterpart as the distance.
In the simplest case, the distance is a constant. For example, the
input threshold T has a type num0 , meaning that its value remains
the same in two executions (since T is a parameter unrelated to
private information about individuals). The distance of variable η1
captures one randomness alignment in the informal proof above:
we enforce a distance of 1 for η1 (i.e., we map noise v to v + 1 for
any value v sampled in the execution under D1 ).
The distance may also depend on program states. Hence,
LightDP supports dependent types. For instance, consider the distance of η2 : q[i] + η2 ≥ T̃ ?2 : 0. This annotation specifies an
injective function that maps noise v to v + 2 when the value of
q[i] + η2 ≥ T̃ is true (i.e., the output is true) in an execution
under D1 , and maps noise v to v otherwise. As we will see shortly,
dependent types allow a precise privacy cost to be calculated under
various program states, hence enabling bounding privacy cost in a
tighter way than mechanisms based on the composition theorem.
Moreover, LightDP uses a distinguished distance ∗ as a shorthand for the standard Sigma type (e.g., num∗ , Σx:num0 numx ). In
other words, each value of type num∗ (e.g., each query in the list q)
can be interpreted as a pair of the form (x : num0 , y : numx ), where
the first component specifies the distance of the second component.
Note that by language design, the first component is invisible in the
privacy-preserving algorithm; however, the type system may reason about and manipulate it via a distinguished operationb. For
instance, the precondition in the running example states the adjacent assumption on databases: for each query answer q[i] in q, its
distance (b
q [i]) is bounded by ±1. The star type is also useful for distances that cannot be easily captured at compile time (Section 3.1).
With type annotations, a type system statically verifies that the
distances are maintained as an invariant throughout the execution.
For example, the output out in Figure 1 has type list bool0 ,
meaning that each element in the list has type bool0 . Hence, the
invariant maintained by the type system ensures that two related
executions always generate the same output.
Calculating privacy cost When type-checking succeeds, the type
system transforms the original program to a non-probabilistic, nonrelational program where privacy cost is explicitly calculated. The
transformed program for the Sparse Vector method is shown in
Figure 2.
The transformed program is almost identical to the original
one, except that: 1) the privacy cost is explicitly calculated via a
new variable v , and 2) probabilistic instructions are replaced by a
new nondeterministic instruction havoc η, which semantically sets
variable η to an arbitrary value upon execution.
The fundamental soundness theorem of the type system states
that, informally, if 1) the original program type-checks, and 2) v
is always bounded by some constant 0 in the transformed program,
then the original program being verified is 0 -differentially private.
function TS PARSE V EC (T, N, : num; q : list num; qb : list num)
returns (out : list bool)
precondition ∀i. −1 ≤ (b
q [i]) ≤ 1
2
v := 0;
havoc η1 ; v := v + /2;
3
4
5
T̃ := T + η1 ;
c1 := 0; c2 := 0; i := 0;
while (c1 < N)
1
Invariant : c1 ≤ N ∧ v = /2 + c1 ×
6
7
8
9
10
11
12
13
2N
havoc η2 ; v := v + (q[i] + η2 ≥ T̃ ?2 : 0) × /4N ;
if (q[i] + η2 ≥ T̃ ) then
out:= true::out;
c1 := c1+1;
else
out:= false::out;
c2 := c2+1;
i := i+1;
Figure 2. The transformed program. The instrumented statements
are underlined. The loop invariant to prove the postcondition v ≤
is shown in grey.
Notice that for the second property (the problem of bounding
v ), any off-the-shelf verification tool for functional correctness
can be utilized. For instance, the program above with the desired
postcondition v ≤ can be verified by Hoare logic, with one loop
invariant provided by a programmer (the grey box in Figure 2).
2.5
Type Inference
The mechanisms sketched so far provide a light-weight yet powerful formal method for differential privacy. The annotation burden
is much reduced compared with (Barthe et al. 2016c). However,
providing the correct and optimal type annotations (especially for
random variables η1 and η2 ) is still subtle for a programmer.
Although it is folklore that type inference in face of dependent
types can be daunting, LightDP is equipped with an inference engine that, at least for many algorithms, not only infers correct annotations, but also enables finding annotations that minimize privacy cost when multiple annotations exist. For example, given the
function signature in Figure 1, the inference algorithm in Section 5
automatically infers types for all variables. The inferred types are
identical to the ones in Figure 1 except that T̃ , η1 are assigned
with a distance α and η2 is assigned with a distance expression
q[i]+η2 ≥ T̃ ?β : γ, where α, β, γ are variables to be inferred, subject to constraints generated during type checking. For the Sparse
Vector method, multiple correct type annotations exist. For example, α = 1, β = 2, γ = 0 corresponds to the annotation in Figure 1.
Moreover, α = 0, β = 2, γ = −2 and α = 2, β = 3, γ = 0 are
both correct annotations.
Given type annotations with distance variables, the type-guided
transformation as sketched above generates target program where
variables to be inferred (i.e., α, β, γ) are used in the calculation of
privacy cost. The difference is that v increments by (α)/2 at line
2 and increments by (q[i] + η2 ≥ T̃ ?β : γ) × /4N at line 6 in
Figure 2. By Hoare logic, we can easily bound the privacy cost to
be α/2 + β/4 + c2 × γ/4N .
Putting it all together, finding the optimal proof is equivalent
to a MaxSMT problem: min (2α + β + c2 × γ/N ), given that
constraints generated in type checking are satisfiable. Using an
existing MaxSMT solver, µZ (Bjørner and Phan 2014; Bjørner
et al. 2015), the optimal proof for the Sparse Vector method is
successfully inferred: α = 1, β = 2, γ = 0. This is exactly the
randomness alignment used in its proof (Dwork and Roth 2014).
Reals
r ∈
Booleans
b ∈
Vars
x ∈
Rand Vars η ∈
Linear Ops ⊕ ::=
Other Ops ⊗ ::=
Comparators ::=
Rand Exps g ::=
Expressions e ::=
R
{true, false}
Var
H
+|−
×|/
<|>|=|≤|≥
Lap r
r | b | x | η | e1 ⊕ e2 | e1 ⊗ e2 | e1 e2 |
¬e | e1 :: e2 | e1 [e2 ] | e1 ?e2 : e3
Commands c ::= skip | x := e | η := g | c1 ; c2 | return e |
if e then c1 else c2 | while e do c
Distances
d ::= r | x | η | d1 ⊕ d2 | d1 ⊗ d2 | d1 d2 ?d3 : d4
Types
τ ::= numd | num∗ | bool | list τ | τ1 → τ2
function PARTIAL S UM (, b, size : num0 ; q : list num∗ )
returns (out : num0 )
precondition ∀i ≥ 0. qb[i] ≤ b∧
∀i ≥ 0. qb[i] > 0 ⇒ (∀j > i. qb[j] = 0)
1
2
3
4
5
6
7
sum : num∗ ; i : num0 ; η : num−sum
c
sum := 0; i := 0;
while (i < size)
sum := sum+q[i];
i := i+1;
η = Lap b/;
out := sum + η;
Figure 4. An -differentially private algorithm for summing over
a query list.
Figure 3. LightDP0 : language syntax.
3.
LightDP: A Language for Algorithm Design
We first introduce a simple imperative language, LightDP0 , for designing and verifying privacy-preserving algorithms. This language
is equipped with a dependent type system that enables formal verification of sophisticated algorithms where the composition theorem
falls short. In this section, we assume all type annotations are provided by a programmer. We will remove this restriction and enable
type inference in Section 5.
3.1
Syntax
The language syntax is given in Figure 3. LightDP0 is mostly a
standard imperative language except for the following features.
Random expressions Probabilistic reasoning is essential in privacypreserving algorithms. We use g to represent a random expression.
Since LightDP0 follows a modular design where new randomness
expression can be added easily, we only consider the most interesting random expression, Lap r, for now. Semantically, Lap r
draws one sample from the Laplace distribution, with mean zero
and a scale factor r. We will discuss other random expressions in
Section 6.3.
Each random expression g can be assigned to a random variable η, written as η := g. We distinguish random variables (H)
from normal variables (Var) for technical reasons explained in Section 3.3. Notice that although the syntax restricts the distribution
scale parameters to be a constant, its mean can be an arbitrary expression e, via the legit expression e + η, where η is sampled from
a distribution with mean zero.
List operations Sophisticated algorithms usually make multiple
queries to a database and produce multiple outputs during that
process. Rather than reasoning about the privacy cost associated
with each query in isolation and total the privacy costs using the
composition theorem, LightDP0 enables more precise reasoning
via built-in list type operations: e1 :: e2 appends the element e1 to a
list e2 ; e1 [e2 ] gets the e2 -th element in list e1 , assuming e2 is bound
by the length of e1 . We also assume a list variable is initialized to
an empty list.
Types with distances Each type τ has the form of Bd . Here, B
is a base type, such as num (numeric type), bool (Boolean), or an
application of a type constructor (e.g., list) to another type, or a
function (τ1 → τ2 ). d is a numeric expression that (semantically)
specifies the exact distance of the values stored in a variable in
two related executions. In particular, a distance expression is a
numeric expression in the language, as specified in Figure 3, where
d1 d2 ?d3 : d4 evaluates to d3 when the comparison evaluates to
true, and d4 otherwise.
Since non-numeric types bool, list τ and τ1 → τ2 cannot be
associated with any numeric distance, those types are syntactic sugars for bool0 , (list τ )0 and (τ1 → τ2 )0 respectively. Notice that
elements in a list of type (list τ )0 (e.g., parameter q in Figure 1)
may still have different elements in two related executions, since
the difference of elements is specified by type τ . The subscript 0
here simply restricts the list size in two related executions.
Star type LightDP0 also supports sum types, written as B∗ , a
syntactic sugar of a Sigma type. More specifically, a variable x
with type num∗ is desugared as x : Σ(bx:num0 ) numxb , where x
b is
a distinguished variable invisible in the source code, but can be
reasoned about and manipulated by the type system. Hiding the
first component of a Sigma type simplifies verification (Section 4).
The parameter q in Figure 1 is one example where the star type
is useful. Moreover, the star type enables reasoning about dependencies that cannot be captured otherwise by a distance expression.
Consider the -differentially private Partial Sum algorithm in Figure 4. It implements an immediate solution to answering the sum
of a query list in a privacy preserving manner3 : it aggregates the
accurate partial sum in a loop, and releases a noisy sum using the
Laplace mechanism. The precondition specifies the adjacency assumption: at most one query answer may differ by at most b.
In this algorithm, the distance of variable sum changes in each iteration. Hence, the accurate type for sum is numPi qb[j] . However,
j=0
with the goal of keeping type system as light-weight as possible,
we assign sum to a star type. The type system will reason about and
manipulate distance component sd
um in a sound way (Section 3.3).
3.2
Semantics
The denotational semantics of the probabilistic language is defined
as a mapping from initial memory to a distribution on (possible)
final outputs. Formally, let M be a set of memory states where each
memory state m ∈ M is an assignment of all (normal and random)
variables (Var ∪ H) to values. First, an expression e of base type B
is interpreted as a function JeK : m → JBK, where JBK represents
the set of values belonging to the base type B. We omit expression
semantics since it is mostly standard4 .
A random expression g is interpreted as a distribution on real
values. Hence, JgK : Dist(JnumK). Moreover, a command c is interpreted as a function JcK : M → Dist(M). For brevity, we
write JeKm and JcKm instead of JeK(m) and JcK(m) hereafter. Figure 5 provides the semantics of commands, where functions unit
3 We use this trivial algorithm here for its simplicity. We will analyze a more
sophisticated version in Section 6.
4 The reals in LightDP only come from sampling (or, the havoc command,
which mimics sampling). We assume the sample space is either finite or
countable.
JskipKm = unit m
Jx := eKm = unit (m{JeKm /x})
Jη := gKm = bind JgK (λv. unit m{v/η})
Jc1 ; c2 Km = bind (Jc1 Km ) Jc2 K
(
Jc1 Km if JeKm = true
Jif e then c1 else c2 Km =
Jc2 Km if JeKm = false
Jwhile e do cKm = w∗ m
where w∗ = f ix(λf. λm.if JeKm = true
then (bind JcKm f ) else (unit m))
Jc; return eKm = bind (JcKm ) (λm0 . unit JeKm0 )
Figure 5. LightDP0 : language semantics.
and bind are defined in Section 2.1. This semantics corresponds
directly to a semantics given by Kozen (1981), which interprets
programs as continuous linear operators on measures.
Finally, we assume all programs have the form (c; return e)
where c does not contain return statements. A LightDP0 program
is interpreted as a function m → DistJBK, defined in Figure 5,
where B is the type of expression returned (e).
3.3
Typing Rules and Target Language
We assume a typing environment Γ that tracks the type of each
variable (including random variable). For now, we assume a type
annotation is provided for each variable (i.e., dom(Γ) = Var ∪ H),
but we will remove this restriction in Section 5. The typing rules
are formalized in Figure 6. Since all typing rules share a global
invariant Ψ (e.g., the precondition in Figure 1), typing rules do not
propagate Ψ for brevity. We also write Γ(x) = d for ∃B. Γ(x) =
Bd when the context is clear.
Expressions For expressions, each rule has the form of Γ ` e : τ ,
meaning that the expression e has type τ under the environment Γ.
Rule (T-OP LUS ) precisely tracks the distance of linear operations
(e.g., + and −), while rule (T-OT IMES ) makes a conservative assumption that other numerical operations take identical parameters.
It is completely possible to refine rule (T-OT IMES )) (e.g., by following the sensitivity analysis proposed by Reed and Pierce (2010);
Gaboardi et al. (2013)) to improve precision, however, we leave that
as future work since it is largely orthogonal.
Rule (T-VAR S TAR ) applies when variable x has a star type.
This rule unpacks the corresponding pair with Sigma type and
makes x
b explicit in the type system.
The most interesting and novel rule is (T-OD OT ). It type-checks
a comparison of two real expressions by generating a constraint:
Ψ ⇒ (e1
e2 ⇔ (e1 +d1 )
(e2 +d2 ))
Intuitively, this constraint requires that in two related executions, the Boolean value of e1
e2 must be identical since the
distances of e1 and e2 are specified by d1 and d2 respectively. For
example, consider the branch condition q[i] + η2 ≥ T̃ in Figure 1.
Rule (T-OD OT ) first checks types for subexpressions:
Γ ` q[i] + η2 : numqb[i]+(q[i]+η2 ≥T̃ ?2:0) and Γ ` T̃ : num1
Then the following constraint is generated (free variables in the
generated constraint are universally quantified):
∀qi , qbi , η2 , T̃ ∈ R. (−1 ≤ qbi ≤ 1) ⇒
qi + η2 ≥ T̃ ⇔ qi + η2 + qbi + (qi + η2 ≥ T̃ ?2 : 0) ≥ T̃ + 1
Typing rules for expressions.
Γ ` r : num0
(T-N UM )
Γ, x : Bd ` x : Bd
Γ ` b : bool
(T-VAR )
(T-B OOLEAN )
Γ, x : B∗ ` x : Bxb
(T-VAR S TAR )
Γ ` e1 : numd1 Γ ` e2 : numd2
(T-OP LUS )
Γ ` e1 ⊕ e2 : numd1 ⊕d2
Γ ` e1 : num0 Γ ` e2 : num0
(T-OT IMES )
Γ ` e1 ⊗ e2 : num0
Γ ` e1 : numd1
Ψ ⇒ (e1 e2
Γ ` e2 : numd2
⇔ (e1 +d1 ) (e2 +d2 ))
(T-OD OT )
Γ ` e1 e2 : bool
Γ ` e : bool
(T-N EG )
Γ ` ¬e : bool
Γ ` e1 : τ Γ ` e2 : list τ
(T-C ONS )
Γ ` e1 :: e2 : list τ
Γ ` e1 : list τ Γ ` e2 : num0
(T-I NDEX )
Γ ` e1 [e2 ] : τ
Γ ` e1 : bool Γ ` e2 : τ Γ ` e3 : τ
(T-S ELECT )
Γ ` e1 ?e2 : e3 : τ
Typing rules for commands
Γ ` skip * skip
(T-S KIP )
Γ ` e : τ Γ ` x : Bd τ = Bd
(T-A SGN )
Γ ` x := e * x := e
Γ ` x : Bxb
Γ ` e : Bd
(T-A SGN S TAR )
Γ ` x := e * x := e; x
b := d;
Γ ` c1 * c01 Γ ` c2 * c02
(T-S EQ )
Γ ` c1 ; c2 * c01 ; c02
Γ ` e : num0 or Γ ` e : bool0
(T-R ETURN )
Γ ` return e * return e
Γ ` e : bool Γ ` ci * c0i where i ∈ {1, 2}
(T-I F )
Γ ` if e then c1 else c2 * if e then c01 else c02
Γ ` e : bool
Γ ` c * c0
(T-W HILE )
Γ ` while e do c * while e do c0
Typing rules for random assignments
Γ(η) = numd
(T-L APLACE )
Γ ` η := Lap r * havoc η; v = v + |d|/r;
Figure 6. Typing rules. Ψ is an invariant that holds throughout
program execution.
Vars
x ∈ Var ∪ H ∪ {v }
Statements c ::= skip | x := e | havoc x | c1 ; c2 | return e
if e then c1 else c2 | while e do c
Informally, the transformation says that by paying a privacy cost,
η has a distance of −d
sum in two related executions. Hence, we can
cancel out the distance of sum in line 7.
Figure 7. Target language syntax. The omitted parts are identical
to the source language defined in Figure 3.
Dependent types and imperative programming Mutable states in
imperative programming brings subtleties that are not foreseen in
the standard theory of dependent types (Martin-Löf 1984). Consider a variable x with type numy where y is initialized to zero. The
type of x establishes an invariant on its values v1 , v2 under two executions: v (1) = v (2) . However, if we update the value of y to 1,
the invariant changes to v (1) + 1 = v (2) , but the values of x in two
executions remains unchanged. Hence, the type invariant is broken.
To address this issue, we assume the following assumptions
are checked before type checking. First, for each normal variable
x ∈ Var such that Γ(x) = Bd , all free variables in d are immutable.
For example, each normal variables in Figure 1 has a constant
distance in its type. Note that by language syntax, this restriction
does not apply to variables with a star type. Second, a random
variable η ∈ H may depend on mutable variables. However, we
assume that it has only one use other than the definition, and the
definition of η is adjacent to its use. Hence, each variable that η
depends on appears immutable between η’s definition and use 5 .
This proof obligation captures a subtle yet important property
of the Sparse Vector method: given the randomness alignment as
specified by Γ, two related executions must take the same branch
(hence, produce the same output). This proof obligation can easily
be discharged by an external SMT solver, such as Z3 (de Moura
and Bjørner 2008).
Target language The typing rules for a command have the form
of Γ ` c * c0 , where c is the original program being verified,
and c0 is the transformed program in the target language defined
in Figure 7. The target language is mostly identical to the original
one, except for two significant differences: 1) the target language
involves a distinguished variables v to explicitly track the privacy
cost in the original program; 2) the target language removes probabilistic expressions, and introduces a new nondeterministic command (havoc x), which sets variable x to an arbitrary value upon
execution. Hence, the target language is nonprobabilistic.
Due to nondeterminism, the denotational semantics interprets a
command c in the target language as a function JcK : M → P(M).
For example, the semantics of the havoc command is defined as
follows:
Jhavoc xKm = ∪r∈R {m{r/x}}
Other commands have a standard semantics, hence their semantics
are included in the appendix. Note that for simplicity, we abuse the
notation JcK to denote the semantics of both the source language
and target language. However, its meaning is unambiguous in the
context of a memory m: when v 6∈ dom(m), JcKm denotes a
distribution; otherwise, JcKm denotes a set.
Commands Informally, if Γ ` c * c0 , and the distinguished
variable v in c0 is bounded by some constant 0 in all possible
executions, then program c is 0 -differentially private. Here, we
discuss important typing rules to enforce this property. We will
formalize this soundness property and sketch a proof in Section 4.
For an assignment x := e, rule (T-A SGN ) synthesizes the
types of x and e, and checks that their types are equivalent (i.e.,
both the base type and distance are equivalent). However, rule (TA SGN S TAR ) instruments the original program so that the typing
invariant (i.e., the distance of x is exactly x
b) is maintained after the
assignment. Consider line 4 in Figure 4. Rule (T-A SGN S TAR ) first
checks subexpressions: Γ ` sum + q[i] : sd
um + qb[i]. Hence, the
transformed program is (sum := sum + q[i]; sd
um := sd
um + qb[i]),
which correctly maintains the typing invariant after the assignment.
(T-R ETURN ) checks that the returned value is indistinguishable
in two related executions. Both (T-I F ) and (T-W HILE ) check that
two related executions must follow the same control flow.
Laplace mechanism Intuitively, (T-L APLACE ) assigns a polymorphic type to the random source Lap r. In other words, for any
distance d of random variable η, we can instantiate the type of Lap r
to be numd , though with a privacy cost of |d|/r. Moreover, the transformed program sets η to a nondeterministic value (havoc η), since
any real value can be sampled from the Laplace distribution.
Consider line 1 in Figure 1. (T-L APLACE ) transforms this line
to (havoc η1 ; v := v +/2) since Γ(η1 ) = num1 . Informally, the
transformation says that by paying a privacy cost of /2, Lap (2/)
ensures that η1 has a distance of one in two related executions.
Moreover, consider line 6 in Figure 4. (T-L APLACE ) transforms
this line to (havoc η; v := v + |d
sum|/b) since Γ(η) = −d
sum.
4.
Soundness
The type system in Section 3.3 enforces a fundamental property: if
Γ ` c * c0 and v in c0 is bounded by some constant , then the
original program being verified is -differentially private.
To formalize and prove this soundness property, we first notice
that a typing environment Γ defines a relation on two memories,
since Γ specifies the exact distance of each variable:
Definition 2 (Γ-Relation). Two memories m1 and m2 are related
by a typing environment Γ, written m1 Γ m2 , iff
∀x ∈ Var. m1 (x) + Jdx Km1 = m2 (x), where Γ ` x : Bdx
Note that since Γ(x) might be a dependent type, the definition
needs to evaluate the distance of x (dx ) under m1 .
By the definition above, Γ is a function since for any memory
m, the distance for each variable in the related memory of m is a
constant. Hence, we also write Γ(m) to represent the unique m0
such that m Γ m0 . Moreover, given a set of distinct memories
S ⊆ M, we define Γ S , {Γ(m) | m ∈ S}. Note that by
definition, Γ S is also a set of distinct memories (hence, not a
multiset). Furthermore, we assume that Γ is an injective function.
We make this assumption explicit by the following definition.
Definition 3 (Well-Formed Γ-relation). A typing environment Γ is
well-formed, written ` Γ, iff Γ is an injective function.
Checking well-formedness of the Γ-relation is straightforward.
Intuitively, Γ is well-formed when there is no “circular” dependency, while more careful analysis is needed for circular dependencies. Consider the Sparse Vector method in Figure 1 and any
m1 and m2 such that Γ m1 = Γ m2 = m for some m. For a
variable y with a constant distance v (e.g., T̃ , η1 , qb[i]), we have
m1 (y) = m(y) − v = m2 (y). So m1 and m2 must agree
on those variables. Then for any variable z that depends on variables that m1 and m2 already agree on, the distance of z must be
identical in m1 and m2 ; hence, m1 (z) = m2 (z). For the circular dependency on variable η2 (whose distance depends η2 ), con5 The
one-use assumption may appear restrictive at first glance, but when
η’s type has no dependency on mutable variables, we can always store η to
a normal variable to circumvent this restriction. When η’s type depends on a
mutable variable and multiple uses of η are needed, we can store η’s value to
a normal variable with star type, whose distance counterpart is manipulated
and reasoned about by the type system.
sider t = JT̃ − q[i]Km1 = JT̃ − q[i]Km2 . The mapping for η2 is
v 7→ v + 2 when v > t and v 7→ v otherwise. Since this mapping
is strictly monotonic, it is injective.
For differential privacy, we are interested in the relationship
between two memory distributions. Given a typing environment Γ
and constant , we define the (Γ, ) distance, written ∆Γ , of two
memory distributions:
Definition 4 (Γ -distance). The Γ -distance of two distributions
µ1 , µ2 ∈ Dist(M), written ∆Γ (µ1 , µ2 ), is defined as:
∆Γ (µ1 , µ2 ) , max (µ1 (S) − exp()µ2 (Γ S))
S⊆M
Note that when S = ∅, the distance is 0 by definition. So
∆Γ (µ1 , µ2 ) ≥ 0 for any , µ1 , µ2 .
The soundness theorem connects the “privacy cost” of the probabilistic program to the distinguished variable v in the transformed
nonprobabilistic program. In order to formalize the connection, we
first extend memory in the source language to include v :
Definition 5. For any memory m and constant , there is an
extension of m, written m ] (), so that
∀x ∈ dom(m). m ] ()(x) = m(x)
∧ m ] ()(v ) =
Next, we introduce useful lemmas and theorems. First, we show
that the type-directed transformation Γ ` c * c0 is faithful. In
other words, for any initial memory m and program c, memory m0
is a possible final memory iff for initial extended memory m ] (0)
and c0 , one final memory is an extension of m0 .
Lemma 1 (Faithfulness).
∀m, m0 , c, c0 , Γ. Γ ` c * c0 ⇒
JcKm (m0 ) 6= 0 ⇔ ∃. m0 ] () ∈ Jc0 Km](0)
Distance Vars α, β, γ ∈ DVar
Distances
d
::= · · · | α
Figure 8. LightDP: language syntax extension.
et al. 2013; Barthe et al. 2012, 2014)) have to (conservatively) provide an unique cost bound for all possible executions, rendering a
cost of 2/N .
The point-wise soundness lemma provides a precise privacy
bound per initial and final memory. However, differential privacy
by definition (Definition 1) bounds the worst-case cost. To close
the gap, we define the worst-case cost of the transformed program.
Definition 7. For any program c in the target language, we say c’s
execution cost is bounded by some constants , written c , iff for
any m ] (0),
m0 ] (0 ) ∈ JcKm](0) ⇒ 0 ≤
Note that this safety property can be verified by an external
mechanism such as Hoare logic and model checking. Off-the-shelf
tools can be used to verify that c holds for some . For example,
we have formally proved that the transformed program in Figure 1
satisfies a postcondition v ≤ by providing one line of annotation
(the grey line in Figure 1) using the Dafny tool (Leino 2010).
Theorem 1 (Soundness).
∀c, c0 , m1 , m2 , Γ, . ` Γ∧Γ ` c * c0 ∧m1 Γ m2 ∧c0 , we have
∆Γ (JcKm1 , JcKm2 ) ≤ 0
Proof. By definition, (max(c m
m1 )) ≤ for all m, m1 . Hence by
Lemma 2, ∀m. JcKm1 (m) ≤ exp()JcKm2 (Γ(m)). Hence,
max (JcKm1 (S) − exp()JcKm2 (Γ(S)))
X
= max
(JcKm1 (m) − exp()JcKm2 (Γ(m))) ≤ 0
S⊆M
Proof. By structural induction on c.
S⊆M
0
m∈S
For a pair of initial and final memories m0 and m when executing the original program, we identify a set of possible v values,
so that in the corresponding executions of c0 , the initial and final
memories are extensions of m and m0 respectively:
We note that the equality in the proof above holds due to the
injective assumption (` Γ), which allows us to derive the set-based
privacy from the point-wise privacy (Lemma 2).
Definition 6. Given a target program c0 , an initial memory m0
and a final memory m0 , the consistent costs of executing c0 w.r.t.
0
m0 and m0 , written c0 m
m0 , is defined as follows
We now connect the soundness theorem to differential privacy:
0
c
0
m
m0 ,
0
0
0
0
Theorem 2 (Privacy).
∀Γ, c, c0 , x, . ` Γ∧Γ ` (c; return e) * (c0 ; return e) then
0
{ | m ] ( ) ∈ Jc Km0 ](0) ∧ m = m }
c0 ⇒ c is -differentially private
0
where m = m iff ∀x ∈ dom(m). m (x) = m(x)
Since (c0 m
m1 ) by definition is a set of values of v , we write
max(c0 m
m1 ) for the maximum cost. The next lemma enables
precise reasoning of privacy cost w.r.t. a pair of initial and final
memories when Γ is injective:
Lemma 2 (Point-Wise soundness).
∀c, c0 , m1 , m2 , m, Γ. ` Γ ∧ Γ ` c * c0 ∧ m1 Γ m2 , we have
JcKm1 (m) ≤ exp(max(c0 m
m1 ))JcKm2 (Γ(m))
The full proof of Lemma 2 is available in the appendix. We
comment that this point-wise result enables precise reasoning of
privacy cost where the composition theorem falls short. Consider
the transformed Sparse Vector method in Figure 2. This point-wise
result allows various cost bounds to be provided for various memories: v increments by 2/N when the branch condition is true, but
it remains the same otherwise. On the other hand, methods based
on the composition theorem (e.g., (Reed and Pierce 2010; Gaboardi
Proof. Proof is available in the appendix.
5.
Differential-Privacy Proof Inference
We have so far presented an explicitly typed language LightDP0 .
However, writing down types (especially those dependent types)
for variables is still a non-trivial task. Moreover, when multiple
proofs exist, writing down types accompanied with the minimum
privacy cost is even more challenging. We extend LightDP0 to
automatically infer a proof and even search for the optimal one.
5.1
Type Inference
Since each type has two orthogonal components (base type and distance), inference is needed for both. The former is mostly standard
(e.g., for Hindley/Milner system (Wand 1987; Aiken and Wimmers
1993; Zhang and Myers 2014)), hence omitted in this paper.
Next, we assume all base types are available, and focus on
the inference of the distance counterpart. For brevity, we write
Γ(x) = d instead of ∃B. Γ(x) = Bd . We use DefVars to represent
the set of variables whose distances are given by the programmer.
To enable type inference, we extend LightDP0 with distance
variables such as α, β, γ (shown in Figure 8). Initially, the typing
environment associates each variable in DefVars with its annotated
distance. It associates each other variable with a distinguished
distance variable to be inferred.
Following the idea of modeling type inference as constraint
solving (e.g., (Wand 1987; Aiken and Wimmers 1993; Haack and
Wells 2004)), it is straightforward to interpret the typing rules in
Figure 6 as a (naive) inference algorithm. To see how, consider two
assignments (x := 0; y := x), where Γ(x) = α, Γ(y) = β. With
distance variables, the typing rules now collect constraints (instead
of checking their validity) during type checking. For example, two
constraints are collected for those two assignments: α = 0 and
β = α. Hence, inferring types is equivalent to finding a solution for
those two constraints (i.e., the satisfiability problem of ∃α, β. α =
0 ∧ β = α). It is easy to check that α = 0 ∧ β = 0 is a solution.
Hence, the inferred distances are Γ(x) = 0, Γ(y) = 0. However,
this naive inference algorithm falls short in face of dependent types.
Next, we first explore the main challenges in inferring dependent
types, and then propose our inference algorithm.
Inferring star types Consider the example in Figure 4. If we
follow the naive inference algorithm above, two constraints are
generated from lines 2 and 4: α = 0 and α = α + qb[i], where
α = Γ(sum). These constraints are unsatisfiable, since the value
of qb[i] is an arbitrary value between −1 and 1. Nevertheless, the
powerful type system of LightDP0 still allows formal verification
of this example by assigning sum to the star type, meaning that its
distance is dynamically tracked.
We observe that starting from the initial typing environment, we
can refine it by processing each assignment x := e in the following
way. We first synthesize the type of e from its subexpressions, in
the same fashion as the original typing rules in Figure 6. Then, if
x ∈ DefVars (i.e., given by the programmer), there is nothing
to be refined. Otherwise, we can refine the typing environment by
updating the type of x to a more precise one:
if Γ(x) = α ∈ DVar
Γ{d/α}
refine(Γ, x, d) , Γ
if Γ(x) 6∈ DVar ∧ (Γ(x) = d)
Γ[x 7→ ∗] otherwise
Here, the auxiliary function refine takes an initial environment Γ,
a variable x and a distance expression d. This function replaces all
occurrences of α in Γ to d when Γ(x) is a variable to be inferred
(α ∈ DVar). Otherwise, it statically checks whether the old and new
distance expressions are equivalent. When the equivalence cannot
be determined at static time, it assigns the ∗ type to x.
Our inference algorithm refines the typing environment as it
proceeds. Consider Figure 4 again. At line 4, sum’s distance is
refined to 0. Then at line 6, its distance is refined to ∗, since we
cannot statically check that 0 = 0 + qb[i] is valid.
Inferring dependency on program state Consider Figure 1 where
only the type of η2 is to be inferred. The naive inference algorithm
will generate one constraint for the branch condition in line 6:
∀qi , qbi , η2 , T̃ . (−1 ≤ qbi ≤ 1) ⇒
(qi + η2 ≥ T̃ ⇔ (qi + η2 + qbi + α ≥ T̃ + 1)
which is unsatisfiable, since there is no single value α that can hide
the difference of qi in both directions. We need a more precise type
for η2 (as provided in Figure 1) so that the “if” and “else” branches
can be aligned in different ways.
To infer dependent types, our inference algorithm propagates
context information to subexpressions. In particular, we observe
that only rule (T-OD OT ) generates a constraint that may benefit
Refinement rules for expressions
Γ; P n
or:Γ
b ∈ {true, false}
(R-B OOLEAN )
Γ; P n
ob:Γ
(R-N UM )
Γ; P n
ox:Γ
P=∅
(R-R AND )
Γ; P n
oη:Γ
(R-VAR )
P 6= ∅
αt , αf fresh variables
(R-R AND -R EFINE )
Γ; P n
o η : refine(Γ, η, P?αt : αf )
Γ; P n
o e1 : Γ1 Γ1 ; P n
o e2 : Γ2 op ∈ ⊕ ∪ ⊗
(R-O PS )
Γ; P n
o e1 op e2 : Γ2
Γ; P ∧ (e1
e2 ) n
o e1 : Γ1 Γ1 ; P ∧ (e1
Γ; P n
o e1 e2 : Γ2
e2 ) n
o e2 : Γ2
(R-OD OT )
Γ; P n
o e1 : Γ1 Γ1 ; P n
o e2 : Γ2
(R-C ONS )
Γ; P n
o e1 :: e2 : Γ2
Γ; P n
o e : Γ0
(R-N EG )
Γ; P n
o ¬e : Γ0
Γ; P n
o e1 : Γ1 Γ1 ; P n
o e2 : Γ2
(R-I DX )
Γ; P n
o e1 [e2 ] : Γ2
Refinement rules for commands
Γn
o skip : Γ
(R-S KIP )
Γn
o return e : Γ
(R-R ETURN )
x 6∈ DefVars
Γ, ∅ n
o e : Γ0
Γ0 ` e : d
(R-A SGN -R EF )
Γn
o x := e : refine(Γ0 , x, d)
x ∈ DefVars Γ, ∅ n
o e : Γ0
(R-A SGN )
Γn
o x := e : Γ0
Γn
o c1 : Γ0 Γ0 n
o c2 : Γ00
(R-S EQ )
Γn
o c1 ; c2 : Γ00
Γ; ∅ n
o e : Γ1 Γ1 n
o c1 : Γ2 Γ2 n
o c2 : Γ 3
(R-I F )
Γn
o if e then c1 else c2 : Γ3
Γ; ∅ n
o e : Γ1
Γ1 ≤ Γ2
Γ2 n
o c : Γ2
(R-W HILE )
Γn
o while e do c : Γ2
Refinement for random assignments
Γn
o η := Lap r : Γ
(R-L APLACE )
Figure 9. The refinement algorithm.
from dependency on program states. Hence, our inference algorithm propagates the comparison result to its subexpressions, and
refine subexpressions (e.g., η2 ) for the needed dependency.
Inference algorithm We now present our inference algorithm,
which is still based on the typing rules in Figure 6. However, to
tackle the challenges above, we run a refinement algorithm before
type inference. The algorithm is shown in Figure 9.
For expressions, the refinement algorithm propagates context
information P to subexpressions. Hence, each rule for expression
has the form of Γ, P n
o e : Γ0 , where P is a predicate that may
appear in a dependent type, Γ is the typing environment to be re-
fined, and Γ0 is the refined environment. The context information P
is used to refine distance of a random variable η in rule (R-R AND R EFINE ). Note that the refinement is not needed for a normal variable x (rule (R-VAR )). Intuitively, the reason is that the “shape” of
x is either provided or has been refined when x is initialized. However, this is not true for a random variable: η can have any distance
expression according to rule (T-L APLACE ).
The refinement rules for commands have the form of Γ n
o c : Γ0 .
As we described informally above, rule (R-A SGN -R EF ) refines
the distance of x using the refine function when its distance is
not given. The rule (R-W HILE ) assumes that a fixed point exists.
Based on the definition of the refine function, a fixed point can
be computed as follows. We define ≤ as the lifted relation based on
a point-wise lattice (for each variable) where: ∀α, β ∈ DVar. α ≤
β ∧ β ≤ α and α ≤ d ≤ ∗ if d is not a distance variable. We
can compute a fixed point by Γ1 = Γ0 ` c : Γ1 , Γ1 ` c :
Γ2 , · · · until Γi ` c : Γi for some i. Based on the definition
of the refine function, it is easy to check that Γi ≤ Γi+1 and
the computation terminates since whenever Γi 6= Γi+1 , either
the number of distance variables is reduced by one, or one more
variable has a star type.
Example We consider type inference for our running example in
Figure 1 where all local variables are to be inferred. We first run
the refinement algorithm. The first refinement happens at line 2,
where the distance of T̃ is refined to α, the distance variable of
η1 . At line 3, c1 , c2 and i are refined to distance 0. In the loop
body, η2 is refined to q[i] + η2 ≥ T̃ ?β : γ at line 6, using
rule (R-R AND -R EFINE ). At line 8, refine(Γ, c1, 0) returns Γ
since 0 = 0 is always true. Similar for the “else” branch and line
12. Hence, the environment after line 12 is already a fixed point
for the loop body. Hence, the typing environment after refinement
is: Γ(c1) = Γ(c2) = Γ(i) = 0, Γ(T̃ ) = Γ(η1 ) = α and
Γ(η2 ) = (q[i] + η2 ≥ T̃ )?β : γ.
Type checking with distance variables With type variables in the
refined environment Γ, the type system collects constraints during
type checking, and tries to solve the collected constraints where
the type variables are existentially qualified. For example, with
type refinement, type checking the partial sum example in Figure 4
yields a unique solution, which is identical to the type annotation
in the figure. In general, collected constraint may have multiple
solutions. For example, type checking the Sparse Vector method
generates only one (nontrivial) constraint from the rule (T-OD OT ):
∀qi , qbi , η2 , T̃ ∈ R. (−1 ≤ qbi ≤ 1) ⇒
qi + η2 ≥ T̃ ⇔ qi + qbi + η2 + (qi + η2 ≥ T̃ ?β : γ) ≥ T̃ + α
It is easy to check that the type annotation in Figure 1 (i.e.,
α = 1, β = 2, γ = 0) is a solution of the constraint. But in fact,
other solutions exist. For example, α = 0, β = 2, γ = −2 and
α = 2, β = 3, γ = 0 are both valid solutions. The type system
can either pick a solution, or defer the inference by transforming
the original program to a target program where type variables are
treated as unknown program inputs (as shown in Figure 10).
5.2
Minimizing Privacy Cost
With type variables captured explicitly in the transformed program,
γ
we can verify that the postcondition v = α
+ β
+c2× 2N
holds
2
2
by providing the loop invariant shown in grey. Hence, combined
with the remaining unsolved constraints on those type variables,
finding the optimal proof is equivalent to the following MaxSMT
function MS PARSE V EC (T, N : num; q : list num; qb : list num;
α, β, γ : num)
returns out : list num
precondition ∀i. −1 ≤ (b
q [i]) ≤ 1
1
2
3
4
5
v := 0;
havoc η1 ; v := v + (α/2);
T̃ := T + η1 ;
c1 := 0; c2 := 0; i := 0;
while (c1 < N)
Invariant : c1 ≤ N ∧ v =
6
α
2
+ c1 ×
β
2N
+ c2 ×
γ
2N
havoc η2 ; v := v + (q[i] + η2 ≥ T̃ ?β : γ) × /4c;
7
if (q[i] + η2 ≥ T̃ ) then
out:= true::out;
c1 := c1+1;
else
out:= false::out;
c2 := c2+1;
i := i+1;
8
9
10
11
12
13
14
Figure 10. The target program with unknown type variables. The
instrumented statements are underlined.
problem, where M is a large number since c2 is not bounded:
min(
β
α
+ + M × γ) such that
2
2
∀qi , qbi , η2 , T̃ ∈ R. (−1 ≤ qbi ≤ 1) ⇒
qi + η2 ≥ T̃ ⇔ qi + qbi + η2 + (qi + η2 ≥ T̃ ?β : γ) ≥ T̃ + α
Using a MaxSMT solver µZ (Bjørner and Phan 2014; Bjørner
et al. 2015), we successfully find the optimal solution for the type
variables: α = 1, β = 2, γ = 0. This is exactly the randomness
alignment used in its formal proof (Dwork and Roth 2014).
We note that the translation to the MaxSMT problem at this
stage still requires programmer efforts (e.g., identifying the cost
bound involving type variables and converting the cost bound to an
equivalent formula suitable for a MaxSMT solver). However, this
example clearly demonstrates the potential benefits of explicitly
calculating the privacy cost in the target language.
5.3
Proof Automation
In general, a LightDP-based proof consists four steps involving
manual efforts: 1) writing down the program specification (i.e., the
function signature that specifies private and non-private parameters
and return values), 2) writing down the type annotations for local
variables, 3) verifying that the privacy cost in the transformed
program is bounded by either a known budget, or (MaxSMT only) a
formula involving unsolved type variables, and 4) (MaxSMT only)
solving the MaxSMT problem of “min(upper bound formula) such
that constraints from step 2 are satisfiable”.
As most verification tools, LightDP requires a programmer to
write down specification (step 1). For step 2, we find that the inference algorithm in Section 5.1 is powerful enough to automatically
infer the types for the nontrivial algorithms considered in this paper6 . For step 3 and step 4, LightDP relies on the automation in
existing verification tools. We note that though LightDP currently
adds no automation in step 3 and 4, separating relational reasoning
from counting privacy cost and automating task 2 greatly simplifies those steps for all examples that we have seen so far. We leave
systematic research in automating the entire proof as future work.
6 The
only exception is the algorithm in Section 6.3, since the algorithm
uses a uniformly distributed random source which is currently absent in the
inference algorithm.
function N UM S PARSE V ECTOR (T, N, : num0 ; q : list num∗ )
returns (out : list num0 )
precondition ∀i. −1 ≤ (b
q [i]) ≤ 1
c1, c2, i : num0 ; T̃ , η1 : num1 ; η2 : numq[i]+η2 ≥T̃ ?2:0 ; η3 : num−b
q [i]
η1 := Lap (3/);
2 T̃ := T + η1 ;
3 c1 := 0; c2 := 0; i := 0;
4 while (c1 < N)
5
η2 := Lap (6N/);
6
if (q[i] + η2 ≥ T̃ ) then
7
η3 := Lap (3N/);
8
out:= (q[i]+η3 )::out;
9
c1 := c1 + 1;
10
else
11
out:= 0::out;
12
c2 := c2 + 1;
13
i := i+1;
The transformed program, where underlined commands are added by the
type system. Only one annotation (loop invariant) is needed from the pro2
.
grammer to verify the postcondition v ≤ : c1 ≤ N ∧v = 3 +c1× 3N
1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
v := 0;
havoc η1 ; v := v + /3;
T̃ := T + η1 ;
c1 := 0; c2 := 0; i := 0;
while (c1 < N)
havoc η2 ; v := v + (q[i] + η2 ≥ T̃ ?2 : 0) × /6N ;
if (q[i] + η2 ≥ T̃ ) then
havoc η3 ; v := v + |b
q [i]| × /3N ;
out:= (q[i]+η3 )::out;
c1 := c1 + 1;
else
out:= 0::out;
c2 := c2 + 1;
i := i+1
Figure 11. The Numerical Sparse Vector method.
6.
Case Studies
6.1
Sparse Vector with Numerical Answers
We first study a numerical variant of the Sparse Vector method. The
previous version (Figure 1), produces only two types of outputs for
each query: true, meaning the query answer is probably above
the threshold; and false, meaning that it is probably below. The
numerical variant, shown in Figure 11, replaces the output true
with a noisy query answer. It does this by drawing fresh Laplace
noise and adding it to the query (Line 8).
Verification using LightDP LightDP can easily verify this numerical variant from scratch, in a very similar way as verifying
the Sparse Vector method. However, here we focus on another
interesting scenario of using LightDP: the programmer (or algorithm designer) has already verified the Sparse Vector method using
LightDP, and she is now exploring its variations. This is a common
scenario for algorithm designers. We show that since LightDP automatically fills in most proof details, exploring variations of an
algorithm requires little effort.
In particular, we assume the programmer has already obtained
the (optimal) types for all local variables except η3 , and the loop
invariant shown in Figure 2 from the verification of the Sparse
Vector method. Hence, the type inference engine only needs to infer
a type for η3 , which is trivially solved to be num−bq[i] . Moreover,
LightDP transforms the original program to the one on the bottom
of Figure 11. To finish the proof, according to Theorem 2, it is
sufficient to verify the postcondition that v ≤ . In fact, only one
annotation (shown in Figure 11) that is very close to the one in
Figure 2 is needed to finish the proof. Hence, we just proved the
numerical Sparse Vector variant for (almost) free using LightDP.
Incorrect variants The numerical variant is also historically interesting since it fixes a bug in a very influential set of lecture notes
(Roth 2011); these lecture notes inadverantly re-used the same
noise used for the “if” test (Line 7) instead of drawing new noise
when outputting the noisy query answer. In other words, Lines 5-8
in Figure 1 are replaced with:
η2 := Lap (2N/);
q̃ = q[i] + η2
if (q̃ ≥ T̃ ) then
out:= (q̃)::out;
For this incorrect variant, the refinement algorithm refines the type
of q̃ to be qb[i] + α when q̃ is defined, where Γ(η2 ) = α. Moreover,
during type checking, ((q̃) :: out) generates a constraint (b
q [i]+α =
0) by rule (T-C ONS ). Hence, it must be true that Γ(q̃) = num0
and Γ(η2 ) = num−bq[i] after type inference. Moreover, after type
checking, η2 := Lap (2N/) is transformed to
(havoc η2 ; v := v + |b
q[i]|(/2N))
However, we cannot prove that the incorrect variant is 0 -private for
any 0 . The reason is that v in the transformed program is clearly
not bounded by any constant 0 : v increments by /2N in the
worst case in each loop iteration, but the number of iterations is
unbounded (when most iterations take the “else” branch).
The failure of a formal proof of the incorrect variant also sheds
lights on how to fix it. For example, if we bound the number of
iterations to be N , then the incorrect variant is fixed (though with a
different privacy cost).
6.2
Smart Summation
We next study a smart summation algorithm verified previously
(with heavy annotations) in (Barthe et al. 2012, 2014). The pseudo
code, shown in Figure 12, is adapted from (Barthe et al. 2014). The
goal of this smart sum algorithm is to take a finite sequence of bits
q[0], q[1], . . . , q[T ] and output a noisy
PT version of their partial sum
sequence: q[0], q[0] + q[1], . . . ,
i=0 q[i]. One naive approach
is to add Laplace noise to each partial sum (partial implementation
is shown in Figure 4). An alternative naive algorithm is to compute
a noisy bit q̃[i]
PT= q[i]+Lap(1/) for each i and output q̃[0], q̃[0]+
q̃[1], . . . ,
i=0 q̃[i]. However, in both approaches, the noise will
swamp the true counts.
A much smarter approach was proposed by Chan et al. (2011).
Intuitively, their algorithm groups q into nonoverlapping blocks
of size M . So block G1 = {q[0], q[1], . . . , q[M − 1]}, G2 =
{q[M ], q[M +1], q[M +2], . . . , q[2M −1]}, etc. Then it maintains
2 levels of noisy counts: (1) the noisy bits q̃[i]
P= q[i]+Lap(1/) for
each i, and (2) the noisy block sums G̃j = i∈Gj q[i] + Lap(1/)
for each block. The partial sums are computed from these noisy
counts
bits:
P` in the following way. Consider the sum of the first `+1
`+1
q[i].
We
can
represent
`+1
=
xM
+c
where
x
=
b
c
and
i=0
M
c = ` + 1 mod M . Hence, the noisy partial sum can be computed
from the noisy sum of the first x blocks plus the remaining c noisy
c−1
P
bits: G̃1 + G̃2 + · · · + G̃x +
q̃[xB + j]. This algorithm is shown
j=0
in Figure 12. The “if” branch keeps track of block boundaries and is
responsible for summing up the noisy blocks. The “else” branch is
responsible for adding in the remaining loose noisy bits (once there
are enough loose bits to form a new block Gj , we use its noisy sum
G̃j rather than the sum of its noisy bits).
function S MART S UM (, M, T: num0 ; q: list num∗ )
returns (out : list num0 )
precondition ∀i. −1 ≤ (b
q [i]) ≤ 1 ∧
(∀i. qb[i] 6= 0 ⇒ (∀j 6= i. qb[j] = 0))
next, n, i : num0 ; sum : num∗ ; η1 : num−sum−b
c q [i] ; η2 : num−b
q [i]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
next:=0; n:=0; i:=0; sum := 0;
while i ≤ T
if (i + 1) mod M = 0 then
η1 := Lap 1/;
n := n + sum + q[i] + η1 ;
next:= n;
sum := 0;
out := next::out;
else
η2 := Lap 1/;
next:= next + q[i] + η2 ;
sum := sum + q[i];
out := next::out;
i := i+1;
The transformed program, where underlined commands are added by the
type system. Only one annotation (loop invariant) is needed from the programmer to verify the postcondition v ≤ 2: (v + sd
um > 0 ⇒ ∀j ≥
i. qb[j] = 0) ∧ (d
sum > 0 ⇒ v ≤ ) ∧ (d
sum ≤ 1.0) ∧ (v ≤ 2)
1 v = 0;
d := 0;
2 next:=0, n:=0, i:= 0, sum:=0;sum
3 while i ≤ T
4
if (i + 1) mod M = 0 then
5
havoc η1 ; v := v + |d
sum + qb[i]|;
6
n := n + sum + q[i] + η1 ;
7
next:= n;
8
sum := 0;
d := 0;
9
sum
10
out := next::out;
11
else
12
havoc η2 ; v := v + |b
q [i]|;
13
next:= next + q[i] + η2 ;
14
sum := sum + q[i];
d := sum
d + |b
15
sum
q [i]|;
16
out := next::out;
17
i := i+1;
Figure 12. The SmartSum algorithm.
Assume for two adjacent databases, at most one query answer
differs, and for that query, its distance is at most one (this adjacency assumption is provided as the precondition in function signature). Hence, for queries that generate the same answer on adjacent
databases, no privacy cost is paid. However, privacy cost is paid
twice to hide the query answers that differ: when the noisy sum for
the block containing that query is computed, and when the noisy
version of that query is used. Hence informally, the SmartSum algorithm satisfies 2-privacy where is a function parameter.
Verification using LightDP LightDP successfully infers the type
annotations shown in the box under function signature in Figure 12.
Since all type variables are only involved in equality constraints,
only one solution exists. The transformed program is shown at the
bottom of Figure 12.
By Theorem 2, to prove SmartSum is 2-private, it is sufficient
to verify that the postcondition v ≤ 2 holds for the transformed
program. We notice that this program maintains the loop invariant
shown in Figure 12. One observation is that once the privacy cost or
the distance of variable sum gets positive, the query that generates
different answers must have been handled already. Hence, rest
queries must have identical answers on adjacent databases ((v +
function P RIV B ERNOULLI (t : num∗ )
returns b : bool
precondition 0 ≤ t ≤ 1 ∧ 0 ≤ t + b
t≤1
η := Uniform[0,1] ;
if (η ≤ t) then
b := true;
else
b := false;
1
2
3
4
5
The transformed program where underlined commands are added by the
type system:
1 v :=0;
2 havoc η; v = v − log(1 + (η ≤ t)?(b
t ≥ 0?0 : (b
t/t))
: (b
t ≤ 0?0 : (b
t/t)));
if (η ≤ t) then
b := true;
else
b := false;
3
4
5
6
Figure 13. The PrivBernoulli algorithm.
sd
um > 0 ⇒ ∀j ≥ i. qb[j] = 0)). Using the loop invariant, we
formally verified the desired postcondition v ≤ 2 using Dafny.
6.3
Categorical Outputs
Until now, we have used the Laplace mechanism, which generates numerical outputs, as the primary randomization tool for ensuring differential privacy. It might seem that categorical attributes
would require completely different techniques, but indeed, they can
be cleanly incorporated into LightDP with a new typing rule. We
briefly show how this can be done by considering a simple mechanism that takes a private-data-dependent probability t and outputs
true with probability t and false with probability 1 − t. The algorithm shown in Figure 13.
The standard trick of generating an output true with probability
t can be done by generating an uniform [0,1] random variable x and
returning true if x ≤ t, and false otherwise. This trick converts
numerical randomness into categorical randomness with a notion
of distance that can be aligned between executions under related
databases. Generalizations to a larger output domain are routine
and, in this way, can allow some instantiations of the exponential
mechanism (McSherry and Talwar 2007).
To calculate the privacy cost of aligning the binary output, we
need to add a single typing rule to capture the property of uniform
[0,1] distribution:
Γ(η) = numη·d
−1<d≤0
Γ ` η := Uniform[0,1] * havoc[0,1] η; v = v − log(d + 1)
This rule requires that the random sample is aligned by a distance of η · d for some d (i.e., we map η to (d + 1)η in the randomness alignment). Easy to check this mapping is injective. By
property of uniform distribution, the privacy cost of any such assignment is − log(d + 1) where −1 ≤ d ≤ 0.
To integrate this typing rule and uniform distribution into
LightDP, we need to establish that: 1) the faithfulness of the transformation, and 2) the uniform distribution satisfies Lemma 2. The
former is easy to check, and we establish the latter in the appendix.
With this new typing rule for uniform distribution, we can
precisely compute the privacy cost of the algorithm in Figure 13
by providing the following type for η: Γ(η) = η · d where
d = (η ≤ t)?(b
t ≥ 0?0 : (b
t/t))
b
b
:(t ≤ 0?0 : (t/t))
During type checking, rule (T-OD OT ) checks the following constraint for the branch condition η ≤ t ⇔ η + η · d ≤ t + b
t, which
can be discharged by a SMT solver. Hence, the algorithm is transformed to the program at the bottom of Figure 13. By the fact that
the newly added random source and typing rules satisfies Lemma 2,
the privacy cost of this subtle example is provably bounded by the
transformed cost formula in the transformed program 7 .
7.
Related Work
Type systems for differential privacy Fuzz (Reed and Pierce
2010) and its successor DFuzz (Gaboardi et al. 2013) reason about
the sensitivity (i.e., how much does a function magnify distances
between inputs) of a program. DFuzz combines linear indexed
types and lightweight dependent types to allow rich sensitivity
analysis. However, those systems rely on (without verify) external
mechanisms (e.g., Laplace mechanism, Sparse Vector method) as
trusted black boxes to release final query answers, without verifying those black boxes. LightDP, on the other hand, verifies sophisticated privacy-preserving mechanisms that releases those final
answers. Sensitivity inference (D’Antoni et al. 2013) was proposed
in the context of Fuzz. While sensitivity inference shares the same
goal of minimizing type annotation and it also uses SMT solvers,
the very different type system in LightDP brings unique challenges
(Section 5.1) that do not present in Fuzz.
HOARe2 (Barthe et al. 2015) and its extension PrivInfer (Barthe
et al. 2016a) have the ability to relate a pair of expressions via
relational assertions that appear as refinements in types. Hence,
they can verify mechanisms that privately release final query answers as well as private Bayesian inference algorithms. However,
HOARe2 and PrivInfer incur heavy annotation burden on programmers. Moreover, they can not deal with privacy-preserving algorithms that go beyond the composition theorem (e.g., the Sparse
Vector method).
Program logic for differential privacy Probabilistic relational
program logic (Barthe et al. 2012, 2013; Barthe and Olmedo 2013;
Barthe et al. 2016c,b) use custom relational logics to verify differential privacy. These systems have successfully verified privacy
for many advanced examples. However, only the very recent work
by Barthe et al. (2016c,b) can verify the Sparse Vector method.
While these logics are expressive enough to prove (, δ) privacy,
the main difficulty with these approaches is that they use custom
and complex logics that incurs steep learning curve and heavy annotation burden. Moreover, ad hoc rules for loops are needed for
many advanced examples.
The work by Barthe et al. (2014) transforms a probabilistic relational program to a nondeterministic program, where standard
Hoare logic can be used to reason about privacy. However, the fundamental difference between that work and LightDP is that the former cannot verify sophisticated algorithms where the composition
theorem falls short, since it lacks the power to express subtle dependency between privacy cost and memory state. Moreover, beneath the surface, that work and LightDP are built on very different
principals and proof techniques. Further, their approach requires
7 We
note that without LightDP, the precise calculation of privacy cost
is very difficult and error-prone. To show that the randomness alignment
cancels out the difference in the private-data-dependent probability t, we
need to analyze four cases. When outputting true and b
t ≥ 0, the related
execution must output true as well (η ≤ t ∧ b
t ≥ 0 ⇒ η ≤ t+b
t).
When outputting true and b
t < 0, this alignment maps η to η(d + 1) =
η((t + b
t)/t). Hence, true is the output in the related execution (η ≤ t ⇒
η(t + b
t)/t ≤ t + b
t). Similar reasoning applies to the case outputting false
too. Moreover, connecting this alignment to -privacy require is even more
daunting by a paper-and-pencil proof.
heavier annotation burden since both relational and functional (e.g.,
bounding privacy cost) properties are reasoned about in the transformed program, while the former is completely and automatically
handled by the type system of LightDP.
The notion of aligning randomness has been used in the recent coupling method (Barthe et al. 2016c,b). While the coupling
method is capable of proving (, δ) privacy and it does not require
the injective assumption on the alignment, the cost of doing so is
the steep learning curve and heavy annotation burden. Technically,
the coupling method reasons about privacy for each possible output
(or a set of outputs), while the alignment-based theory used in this
paper aligns two program executions that will produce the same results. The theory in this paper gives a simple proof, a light-weight
type system, and clear insight behind the type system.
Other language-based methods for differential privacy Several
dynamic tools exist for enforcing differential privacy. PINQ (McSherry 2009) tracks (at runtime) the privacy budget consumption,
and terminates the computation when the privacy budget is exhausted. Airavat (Roy et al. 2010) is a MapReduce-based system
with a runtime monitor that enforces privacy policies controlled
by data providers. Recent work by Ebadi et al. (2015) proposed
Personalised Differential Privacy (PDP), where each individual has
its own personal privacy level and a dynamic system that implements PDP. There are also methods based on computing bisimulations families for probabilistic automata (Tschantz et al. 2011;
Xu et al. 2014). However, none of these techniques has the expressive power to provide a tight privacy cost bound for sophisticated
privacy-preserving algorithms.
8.
Conclusions and Future Work
The increased usage and deployment of differentially private algorithms underscores the need for formal verification methods to
ensure that personal information is not leaked due to mistakes or
carelessness. The ability to verify subtle algorithms should be coupled with the ability to infer most of the proofs of correctness to
reduce the programmer burden during the development and subsequent maintenance of a privacy-preserving code base.
In this paper, we present a language with a lightweight type
system that allows us to separate privacy computation from the
alignment of random variables in hypothetical executions under
related databases. Thus enabling inference and search for proofs
with the minimal privacy costs.
These techniques allow us to verify (with much fewer annotations) algorithms that were out of reach of the state of the art until recently. However, additional extensions are possible. The first
challenge is to extend these methods to algorithms that use hidden private state to reduce privacy costs. One example is the noisy
max algorithm that adds noise to each query and returns the index
of the query with the largest noisy answer (although all noisy answers are used in this computation, the fact that their values are
kept secret allows more refined reasoning to replace the composition theorem). The second challenge is verifying subtle algorithms
such as PrivTree (Zhang et al. 2016), in which intermediate privacy costs depend on the data (hence cannot be released) but their
sum can be bounded in a data-independent way. This is another
case where the composition theorem can fail since it requires dataindependent privacy costs. Lastly, LightDP currently only verifies
-privacy, which has a nice point-wise property. We leave extending
LightDP to (, δ)-privacy as future work.
Acknowledgments
We thank Adam Smith, our shepherd Marco Gaboardi and anonymous reviewers for their helpful suggestions. This work was supported by NSF grants CNS-1228669 and CCF-1566411.
References
A. Aiken and E. L. Wimmers. Type inclusion constraints and type inference.
In FPLCA, pages 31–41, 1993.
G. Barthe and F. Olmedo. Beyond differential privacy: Composition theorems and relational logic for f-divergences between probabilistic programs. In ICALP, pages 49–60, 2013.
G. Barthe, B. Köpf, F. Olmedo, and S. Zanella Béguelin. Probabilistic relational reasoning for differential privacy. In Proceedings of the 39th
Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, pages 97–110, 2012.
G. Barthe, G. Danezis, B. Grégoire, C. Kunz, and S. Zanella-Béguelin. Verified computational differential privacy with applications to smart metering. In 2013 IEEE 26th Computer Security Foundations Symposium,
pages 287–301, 2013.
G. Barthe, M. Gaboardi, E. J. G. Arias, J. Hsu, C. Kunz, and P. Y. Strub.
Proving differential privacy in hoare logic. In 2014 IEEE 27th Computer
Security Foundations Symposium, pages 411–424, 2014.
G. Barthe, M. Gaboardi, E. J. G. Arias, J. Hsu, A. Roth, and P. Strub.
Higher-order approximate relational refinement types for mechanism
design and differential privacy. In POPL, 2015.
G. Barthe, G. P. Farina, M. Gaboardi, E. J. G. Arias, A. Gordon, J. Hsu, and
P.-Y. Strub. Differentially private bayesian programming. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, pages 68–79, 2016a.
G. Barthe, N. Fong, M. Gaboardi, B. Grégoire, J. Hsu, and P.-Y. Strub. Advanced probabilistic couplings for differential privacy. In Proceedings of
the 2016 ACM SIGSAC Conference on Computer and Communications
Security, pages 55–67, 2016b.
G. Barthe, M. Gaboardi, B. Gregoire, J. Hsu, and P.-Y. Strub. Proving
differential privacy via probabilistic couplings. In IEEE Symposium on
Logic in Computer Science (LICS), 2016c. To apprear.
N. Bjørner and A.-D. Phan. νZ — maximal satisfaction with Z3. In
T. Kutsia and A. Voronkov, editors, 6th International Symposium on
Symbolic Computation in Software Science (SCSS), volume 30 of EPiC
Series in Computing, pages 1–9, 2014.
N. Bjørner, A.-D. Phan, and L. Fleckenstein. νZ — An Optimizing SMT
Solver, pages 194–199. 2015.
H. Chan, E. Shi, and D. Song. Private and continual release of statistics.
ACM Transactions on Information and System Security, 14(3), 2011.
Y. Chen and A. Machanavajjhala. On the privacy properties of variants on
the sparse vector technique. http://arxiv.org/abs/1508.07306, 2015.
L. D’Antoni, M. Gaboardi, E. J. Gallego Arias, A. Haeberlen, and B. Pierce.
Sensitivity analysis using type-based constraints. In Proceedings of the
1st Annual Workshop on Functional Programming Concepts in Domainspecific Languages, pages 43–50, 2013.
L. M. de Moura and N. Bjørner. Z3: An efficient SMT solver. In Conf.
on Tools and Algorithms for the Construction and Analysis of Systems
(TACAS), 2008.
C. Dwork and A. Roth. The algorithmic foundations of differential privacy.
Foundations and Trends in Theoretical Computer Science, 9(3–4):211–
407, 2014. ISSN 1551-305X. doi: 10.1561/0400000042.
C. Dwork, K. Kenthapadi, F. McSherry, I. Mironov, and M. Naor. Our data,
ourselves: Privacy via distributed noise generation. In EUROCRYPT,
pages 486–503, 2006a.
C. Dwork, F. McSherry, K. Nissim, and A. Smith. Calibrating noise to
sensitivity in private data analysis. In TCC, 2006b.
H. Ebadi, D. Sands, and G. Schneider. Differential privacy: Now it’s getting
personal. In POPL, 2015.
U. Erlingsson, V. Pihur, and A. Korolova. Rappor: Randomized aggregatable privacy-preserving ordinal response. In Proceedings of the 2014
ACM SIGSAC Conference on Computer and Communications Security,
CCS ’14, 2014.
M. Gaboardi, A. Haeberlen, J. Hsu, A. Narayan, and B. C. Pierce. Linear dependent types for differential privacy. In Proceedings of the 40th
Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’13, pages 357–370, 2013.
A. Greenberg. Apple’s ‘differential privacy’ is about collecting your data
– but not Your data. Wired, https://www.wired.com/2016/
06/apples-differential-privacy-collecting-data/,
2016.
C. Haack and J. B. Wells. Type error slicing in implicitly typed higher-order
languages. Science of Computer Programming, 50(1–3):189–224, 2004.
D. Kifer and A. Machanavajjhala. Pufferfish: A framework for mathematical privacy definitions. ACM Trans. Database Syst., 39(1):3:1–3:36,
2014.
D. Kozen. Semantics of probabilistic programs. Journal of Computer and
System Sciences, 22(3):328 – 350, 1981.
K. R. M. Leino. Dafny: An automatic program verifier for functional
correctness. In Proceedings of the 16th International Conference on
Logic for Programming, Artificial Intelligence, and Reasoning, pages
348–370, 2010.
M. Lyu, D. Su, and N. Li. Understanding the sparse vector technique for
differential privacy. https://arxiv.org/abs/1603.01699, 2016.
A. Machanavajjhala, D. Kifer, J. Abowd, J. Gehrke, and L. Vilhuber. Privacy: From theory to practice on the map. In Proceedings of the IEEE
International Conference on Data Engineering (ICDE), pages 277–286,
2008.
P. Martin-Löf. Intuitionistic type theory. Naples: Bibliopolis, 76, 1984.
F. McSherry and K. Talwar. Mechanism design via differential privacy.
In Proceedings of the 48th Annual IEEE Symposium on Foundations of
Computer Science, pages 94–103, 2007.
F. D. McSherry. Privacy integrated queries: An extensible platform for
privacy-preserving data analysis. In Proceedings of the 2009 ACM
SIGMOD International Conference on Management of Data, pages 19–
30, 2009.
P. Mohan, A. Thakurta, E. Shi, D. Song, and D. Culler. Gupt: Privacy preserving data analysis made easy. In Proceedings of the ACM SIGMOD
International Conference on Management of Data, 2012.
J. Reed and B. C. Pierce. Distance makes the types grow stronger: A calculus for differential privacy. In Proceedings of the 15th ACM SIGPLAN
International Conference on Functional Programming, ICFP ’10, pages
157–168, 2010.
A. Roth. The sparse vector technique. http://www.cis.upenn.
edu/˜aaroth/courses/slides/Lecture11.pdf, 2011.
I. Roy, S. Setty, A. Kilzer, V. Shmatikov, and E. Witchel. Airavat: Security
and privacy for MapReduce. In NSDI, 2010.
M. C. Tschantz, D. Kaynar, and A. Datta. Formal verification of differential
privacy for interactive systems (extended abstract). Electron. Notes
Theor. Comput. Sci., 276:61–79, Sept. 2011.
M. Wand. A simple algorithm and proof for type inference. Fundamenta
Informaticae, 10:115–122, 1987.
L. Xu, K. Chatzikokolakis, and H. Lin. Metrics for Differential Privacy in
Concurrent Systems, pages 199–215. 2014.
D. Zhang and A. C. Myers. Toward general diagnosis of static errors. In
ACM Symposium on Principles of Programming Languages (POPL),
pages 569–581, Jan. 2014.
J. Zhang, X. Xiao, and X. Xie. Privtree: A differentially private algorithm
for hierarchical decompositions. In SIGMOD, 2016.
Appendix
A.
Soundness Proof
In the source language semantics (Figure 5), the variable that a
star-typed variable depends on (e.g., sd
um in Figure 4) is invisible.
We first extend the semantics to make the manipulation of such
invisible variables explicit, by the following rule for assignments:
Jx := eKm = unit (m{JeKm /x}{JdKm /b
x}) when Γ(x) = num∗ ,
where Γ(e) = d.
It is straightforward to check that the extended semantics (parameterized on the type system) is consistent with the original semantics in Figure 5, as it does not change the distribution on the
variables that are visible in the source program. The extended semantics is needed to close the gap between the source language and
the one that formal reasoning is applied on.
Next, we prove a few auxiliary lemmas.
When Γ(x) = B∗ , Γ ` x : Bxb . Hence, m01 (x) + m01 (b
x) =
JeKm1 + JdKm1 , where Γ(e) = d, by the extended semantics. By Lemma 4, this is identical to JeKm2 , which is m02 (x)
by the semantics.
Second, we show m01 (y) + JdKm01 = m02 (y), where Γ ` y : d
for y ∈ dom(Γ) ∧ y 6= x. When y ∈ Var, its type cannot depend
on x, which is mutable. So the desired result is true. For η ∈ H,
its type only depends on the memory state when η is used. So
the desired result is true as well.
• Case if e then c1 else c2 : by typing rule, Γ ` e : bool0 . By
Lemma 4, JeKm1 = JeKm2 . Hence, the same branch is taken in
m1 and m2 . Desired result is true by induction hypothesis.
• Case c1 ; c2 : For any m such that Jc1 ; c2 Km1 (m) 6= 0, there
exists some m0 such that
Jc1 Km1 (m0 ) 6= 0 ∧ Jc2 Km0 (m) 6= 0
By induction hypothesis, we have
Lemma 3.
∀m1 , m2 , Γ s.t. m1 Γ m2 , we have
∀m. unit m1 (m) = unit m2 (Γ(m))
Jc1 Km1 (m0 ) ≤ exp(1 )Jc1 KΓ(m1 ) (Γ(m0 ))
Jc2 Km0 (m) ≤ exp(2 )Jc2 KΓ(m0 ) (Γ(m))
0
Proof. By the fact that Γ is a function.
m
where 1 = max(c01 m
m1 ) and 2 = max(c2 m0 ). Hence,
Jc1 Km1 (m0 ) · Jc2 Km0 (m) ≤
Lemma 4 (Expression).
∀e, m1 , m2 , Γ s.t. m1 Γ m2 ∧ Γ ` e : Bd , we have
JeKm1 + JdKm1 = JeKm2
Proof. Induction on the structure of e. Interesting cases are follows.
When e is x or η, result is true by the definition of m1 Γ m2 .
When e is e1 ⊕ e2 , let Jei Km1 = vi , Jei Km2 = vi0 and Γ ` ei :
i
Bdi for i ∈ {1, 2}. Then by typing rule, we have d = d1 + d2 . By
induction hypothesis, we have vi0 = vi + di , where di = Jdi Km1 .
Hence, JeKm2 = v10 ⊕ v20 = (v1 + d1 ) ⊕ (v2 + d2 ) = (v1 ⊕ v2 ) +
(d1 ⊕ d2 ) = Je1 ⊕ e2 Km1 + Jd1 ⊕ d2 Km1 = JeKm1 + JdKm1 .
When e = e1
e2 , let Γ ` ei : Bdi i for i ∈ {1, 2}.
Then by induction hypothesis, we have Jei Km1 + Jdi Km1 =
Jei Km2 for i ∈ {1, 2}. By rule (T-OD OT ), for any memory m,
Je1 e2 Km = J(e1 + d1 ) (e2 + d2 )Km . Hence, Je1 e2 Km1 =
J(e1 + d1 ) (e2 + d2 )Km1 = Je1 + d1 Km1
Je2 + d2 Km1 =
Je1 Km2 Je2 Km2 = Je1 e2 Km2 .
Proof of Lemma 2
∀c, c0 , m1 , m2 , m, Γ. ` Γ ∧ Γ ` c * c0 ∧ m1 Γ m2 , we have
JcKm1 (m) ≤ exp(max(c0 m
m1 ))JcKm2 (Γ(m))
Proof. By structural induction on c.
• Case skip: c0 = skip by typing rule. Hence, max(c0 m
m1 ) =
0. Desired result is true by Lemma 3 and the semantics of skip.
• Case x := e: by the transformation, we have max(c0 m
m1 ) = 0.
Hence, by the semantics and Lemma 3, it is sufficient to show
that the memories after the assignment are related by Γ.
We first show m01 (x) + JdKm01 = m02 (x).
When Γ(x) = Bd , we need to show that m01 Γ m02 where
m01 = m1 {JeKm1 /x} and m02 = m2 {JeKm2 /x} by the
semantics. By typing rule, we have Γ ` e : Bd as well.
By Lemma 4, JeKm1 + JdKm1 = JeKm2 . Hence, we have
m01 (x) + JdKm1 = m02 (x). Since d may only depend on immutable variables in this case, JdKm1 = JdKm01 . So m01 (x)+
JdKm01 = m02 (x) as desired.
exp(1 + 2 )Jc1 KΓ(m1 ) (Γ(m0 )) · Jc2 KΓ(m0 ) (Γ(m))
Notice that m0 ] (1 ) ∈ Jc1 Km1 ](0) and m ] (2 ) ∈ Jc2 Km0 ](0)
since 1 and 2 maximize privacy costs among consistent executions by definition. Hence, m ] (1 + 2 ) ∈ Jc01 ; c02 Km1 ](0) .
Therefore, 1 + 2 ≤ max(c1 ; c2 m
m1 ).
So for any m,
X
Jc1 ; c2 Km1 (m) =
Jc1 Km1 (m0 ) · Jc2 Km0 (m)
m0
≤ exp(0 )
X
Jc1 KΓ(m1 ) (Γ(m0 )) · Jc2 KΓ(m0 ) (Γ(m))
≤ exp(0 )
X
Jc1 KΓ(m1 ) (m0 ) · Jc2 Km0 (Γ(m))
m0
m0
≤ exp(0 )Jc1 ; c2 Km2 (Γ(m))
where 0 = max(c1 ; c2 m
m1 ). Notice that the change of variable in the second to last inequality only holds when Γ is an
injective (but not necessarily onto) mapping, which is true due
to the assumption ` Γ.
• Case while e do c: let W = while e do c. By typing rule,
Γ ` e : bool0 . Hence, JeKm01 = JeKm02 for any m01 Γ m02 .
We proceed by by natural induction on the number of loop
iterations (denoted by i) under m1 .
When i = 0, JbKm1 = false. So JbKm2 = false since
m1 Γ m2 . By semantics, JW Km1 = unit m1 and JW Km2 =
unit m2 , and max(W m
m1 ) = 0. Desired result is true by
Lemma 3.
Consider i = j + 1. JbKm1 = true. So JbKm2 = true since
m1 Γ m2 . By semantics, JW Kmi = Jc; W Kmi for i ∈ {1, 2},
and the latter W iterates for j times. By induction hypothesis
and a similar argument as the sequential case, JW Km1 (m) ≤
exp(max(W m
m1 )JW Km2 (Γ(m)).
• Case η := Lap r: let µr = Lap r. Since µr is the Laplace
distribution with a scale factor of r, we have
∀v, d ∈ R. µr (v) ≤ exp(|d| × r)µr (v + d)
When @v. m = m1 {v/η} , Jη := Lap rKm1 (m) = 0 by the
semantics. Hence, desired inequality is trivial.
When m = m1 {v/η} for some constant v, we have for any
d ∈ R,
semantics of commands are formalized as follows.
JskipKm = {m}
Jη := Lap rKm1 (m)
= µr (v)
Jx := eKm = {m{JeKm /x}}
Jhavoc xKm = ∪r∈R {m{r/x}}
≤ exp(|d| · r)µr (v + d)
Let Γ(η) = numd and JdKm = d for some constant d. Since
m1 Γ m2 , m1 {v/η} Γ m2 {v + d/η}. That is, Γ(m) =
m2 {v + d/η}. By the semantics,
Jc1 ; c2 Km = ∪m0 ∈Jc1 Km Jc2 Km0
(
Jc1 Km if JeKm = true
Jif e then c1 else c2 Km =
Jc2 Km if JeKm = false
Jwhile e do cKm = w∗ m
Jη := Lap rKm2 (Γ(m)) = Jη := Lap rKm2 (m2 {v + d/η})
where w∗ = f ix(λf. λm.if JeKm = true
= µr (v + d)
then (∪m0 ∈JcKm f m0 ) else {m})
Hence, we have
Jη := Lap rKm1 (m) ≤ exp(|d| · r)Jη := Lap rKm2 (Γ(m))
when m = m1 {v/η} for some constant v too. By the typing
rule (T-L APLACE ), the transformed program is (havoc η; v :=
v + |d| · r). Hence, max(c01 m
m0 ) = J|d| · rKm = |JdKm | · r =
|d| · r. Therefore, we showed that
Jη := Lap rKm1 (m) ≤
exp(max(c01 m
m0 ))Jη := Lap rKm2 (Γ(m))
Proof of Theorem 2
∀Γ, c, c0 , x, . ` Γ∧Γ ` (c; return e) * (c0 ; return e) then
c0 ⇒ c is -private
Proof. By the soundness theorem (Theorem 1), we have for any injective Γ, m1 Γ m2 , ∀S ⊆ M, JcKm1 (S) ≤ exp()JcKm2 (Γ(S)).
For clarity, we stress that all sets are over distinct elements (as we
have assumed throughout this paper). Let P = (c; return e). By
typing rule (T-R ETURN ), the return type B must be either num or
bool, and its distance must be zero. By semantics, for any value
set V ⊆ B,
JP Km1 (V ) = JcKm1 ({m | JeKm ∈ V })
≤ exp()JcKm2 ({Γ(m) | JeKm ∈ V })
≤ exp()JcKm2 ({m | JeKm ∈ V })
= exp()JP Km2 (V )
Accordingly, the Hoare logic rules for the target language is
mostly standard, summarized in Figure 14.
C.
Uniform Distribution
Lemma 5 (UniformDist). The following typing rule is sound w.r.t.
Lemma 2:
Γ(η) = numη·d
−1<d≤0
Γ ` η := Uniform[0,1] * havoc[0,1] η; v = v − log(d + 1)
Proof. When @v. m = m1 {v/η} or m(η) < −1 or m(η) > 1,
Jη := Uniform[0,1] Km1 (m) = 0 by the semantics. Hence, desired
inequality is trivial.
When m = m1 {v/η} for some −1 ≤ v ≤ 1. Let µ =
Uniform[0,1] , d = JdKm . Notice that by typing rule d ≤ 0. So
d ≤ 0. We have
=
=
=
(1)
(2)
=
(3)
(4)
≤
where inequality (2) is true due to Theorem 1 (the application
of which requires the injective assumption). For inequality (3),
consider any m0 ∈ {Γ(m) | JeKm ∈ V }. It must be true that
m0 = Γ(m) ∧ JeKm ∈ V for some m ∈ M. Due to Lemma 4,
JeKm = JeKΓ(m) (the distance of e must be 0). That is, JeKm =
v ⇔ JeKΓ(m) = v for any v. Hence,
m0 = Γ(m) ∧ JeKm ∈ V for some m ∈ M
is the same as
m0 = m00 ∧ JeKm00 ∈ V for some m ∈ M, where m00 = Γ(m)
Since m00 ∈ M, m0 ∈ {m | JeKm ∈ V }. Hence, the inequality (3)
holds. We note that (3) is not an equality in general since Γ might
not be a surjection.
Therefore, by definition of differential privacy, c is -private.
B.
Jc; return eKm = ∪m0 ∈JcKm {JeKm0 }
Formal Semantics for the Target Language
The denotational semantics interprets a command c in the target
language (Figure 7) as a function JcK : M → P(M). The
=
Jη := Uniform[0,1] Km1 (m)
µ(v)
Z 1
1{x≤v} dx
0
Z 1
1{(d+1)x≤(d+1)v} dx
0
Z 1+d
1
1{x≤(d+1)v} dx
1+d
0
Z 1
1
1{x≤(d+1)v} dx
1+d 0
exp(− log(d + 1))µ(v + v · d)
Since m1 Γ m2 , we have m1 {v/η} Γ m2 {(v + v · d)/η}.Hence,
Jη := Uniform[0,1] Km1 (m) ≤
exp(− log(d + 1))Jη := Uniform[0,1] Km2 (Γ(m))
By the transformation, Uniform[0,1] * havoc[0,1] η; v =
v − log(d + 1), where Γ(η) = η · d. Hence, max(c01 m
m0 ) =
J− log(d + 1)Km1 {v/η} = − log(d + 1). Therefore,
Jη := Uniform[0,1] Km1 (m) ≤
exp(max(c01 m
m0 ))Jη := Uniform[0,1] Km2 (Γ(m))
{Φ}skip{Φ}
(H-S KIP )
{Φ{e/x}}x := e{Φ}
{∀x. Φ}havoc (x){Φ}
(H-A SGN )
(H AVOC )
{Ψ}c1 {Φ0 } {Φ0 }c2 {Φ}
(H-S EQ )
{Ψ}c1 ; c2 {Φ}
{Φ}return e{Φ}
(H-R ETURN )
{e ∧ Ψ}c1 {Φ}
{¬e ∧ Ψ}c2 {Φ}
(H-I F )
{Ψ}if e then c1 else c2 {Φ}
Ψ⇒I
{I ∧ e}c{I}
I⇒Φ
(H-W HILE )
{Ψ}while e do c{Φ}
Figure 14. Hoare logic rules for the target language.
| 6cs.PL
|
Individuals, Institutions, and Innovation in the
Debates of the French Revolution
Alexander T. J. Barrona , Jenny Huanga,d , Rebecca L. Spangb , Simon DeDeoc,d,∗
a School
arXiv:1710.06867v1 [physics.soc-ph] 18 Oct 2017
of Informatics, Computing, and Engineering, Indiana University, 919 E 10th Street, Bloomington, IN 47408, USA
b Department of History, Indiana University, 1020 E Kirkwood Ave, Bloomington, IN 47405, USA
c Department of Social and Decision Sciences, Carnegie Mellon University, 5000 Forbes Avenue, BP 208, Pittsburgh, PA 15213, USA
d Santa Fe Institute, 1399 Hyde Park Road, Santa Fe, NM 87501, USA
Abstract
The French Revolution brought principles of “liberty, equality, and brotherhood” to bear on the day-to-day challenges
of governing what was then the largest country in Europe. Its experiments provided a model for future revolutions and
democracies across the globe, but this first modern revolution had no model to follow. Using reconstructed transcripts
of debates held in the Revolution’s first parliament, we present a quantitative analysis of how this system managed
innovation. We use information theory to track the creation, transmission, and destruction of patterns of word-use across
over 40,000 speeches and more than one thousand speakers. The parliament as a whole was biased toward the adoption of
new patterns, but speakers’ individual qualities could break these overall trends. Speakers on the left innovated at higher
rates while speakers on the right acted, often successfully, to preserve prior patterns. Key players such as Robespierre (on
the left) and Abbé Maury (on the right) played information-processing roles emblematic of their politics. Newly-created
organizational functions—such as the Assembly’s President and committee chairs—had significant effects on debate
outcomes, and a distinct transition appears mid-way through the parliament when committees, external to the debate
process, gain new powers to “propose and dispose” to the body as a whole. Taken together, these quantitative results
align with existing qualitative interpretations but also reveal crucial information-processing dynamics that have hitherto
been overlooked. Great orators had the public’s attention, but deputies (mostly on the political left) who mastered the
committee system gained new powers to shape revolutionary legislation.
Keywords: cultural evolution, political science, cognitive science, digital history, rhetoric, computational social science
∗ To
whom correspondence should be addressed.
Email address: [email protected] (Simon DeDeo)
Preprint submitted to —
October 20, 2017
The French Revolution was a turning point in European
history. Revolutionary commitments to individual liberty
collided with ideals of social equality, while the rejection
of Divine-Right monarchy and the embrace of laws based
on reason opened a host of practical questions about how
to govern the most populous state in Europe. The first
parliament of the Revolution, the National Constituent
Assembly (NCA), was, not surprisingly, a picture of upheaval from its outset.
Over the course of two years, the thousand or more
individuals in that Assembly took it upon themselves to
propose and argue the previously unimaginable: the revocation of Old-Regime privilege and the reinvention of the
relationship between individual and state. But this parliament was more than a debate society for ambitious young
men. It was also the origin of a system of rule. In the years
that followed, successive legislative bodies declared war on
most of Europe, dissolved the French monarchy, declared
a Republic, and sentenced the former king to death—all
while simultaneously writing constitutions and passing ordinary legislation. Many of their procedures, and some
of their personnel, were drawn from the experience of the
NCA.
As a parliament, the NCA faced the problems that
come with managing massive flows of information from
the outside—problems that face the modern deliberative
political bodies that, in many cases, are the NCA’s direct descendants [1]. But as the first parliament they
also shared the challenges of the knowledge-seekers of the
Enlightenment and their descendants: speakers faced the
risks and rewards of innovation as they introduced ideas
and attempted, under unusual conditions, to achieve lasting influence over the future arguments of their contemporaries [2].
Key questions arise from seeing the NCA as simultaneously a site of political and epistemic activity. On the
epistemic side: How did new ideas enter that parliamentroom, and how were they taken up, or discarded, by those
who heard them? What kinds of roles did individuals play
in this process? On the political side: What institutions
did the parliament evolve to manage the onslaught of novelty and reaction, optimism and grievance, philosophical
argument and organizational minutiae that characterized
the day-to-day tasks of governance and nation-building?
The digitization of historical archives allows us to answer these questions in a fundamentally new way. Using
latent Dirichlet allocation [3] and new techniques in information theory drawn from the cognitive sciences [4], we
track the emergence and persistence of word-usage patterns in over 40,000 speeches made in the NCA and later
reconstructed in the Archives Parliamentaires (AP) from
detailed records kept at the time of the Revolution. Two
critical measures, novelty (how unexpected a speech’s patterns are, given past speeches) and transience (the extent
to which those patterns fade away, or, conversely, persist
in future speeches), allow us to trace both emergent political institutions and the generation and propagation of
new ideas and manners of speech. Viewing the turbulent
early days of the French Revolution through the lens of
pattern creation, sharing, and destruction provides a complementary viewpoint to the study of specific ideas and
particular historical events. This distinction can also be
found in evolutionary biology, where one can study either
(on the one hand) the mechanisms of transmission and
selection or (on the other) the particular phenotypes for
which an environment selects.
We find, at high significance, a bias in favor of the propagation of novel patterns from prior speeches into the future: in the framework of cultural evolution, the flow of
ideas through NCA is out of equilibrium and the system
preferentially selects for what violates prior expectations.
This effect is driven in part by charismatic political radicals such as Robespierre and Pétion de Villeneuve, who
not only introduced new patterns at higher rates than their
peers, but managed to do so in a way that led others to
copy those patterns forward. By contrast, influential conservative figures such as Abbé Maury and Cazalès acted as
inertial dampeners, successfully preserving prior patterns
in the face of an overall revolutionary bias towards innovation. Conservatives “conserve”: not just by referring
to past traditions, but also in the day-to-day dynamics of
discussion, where they act to keep conversations on track.
In parallel with these individual-level differences, our
methods reveal a major transition in how the parliament
as a whole processed novelty, occurring about half-way
through the parliament’s lifetime. This shift is associated
with committees that reported to, but met outside of, the
parliament itself, and that had gained new power both to
raise and resolve questions. While orators on the left and
right continued to confront each other in public speeches
from the floor, it was the left that captured this new institutional mechanism and used it to their own advantage.
The consolidation of this structural shift was accompanied
by radicalization of the left and an accelerating flight of
conservatives from both the parliament and the country
itself.
Results
Social systems are characterized by heteroglossia: the
coexistence, sharing, and competition of different patterns
of speech. Heteroglossia makes linguistics and rhetoric,
both concerned with the study of the reception, influence,
and propagation of language within a community [5, 6],
core components in the quantitative study of culture.
Tracking changes in speech patterns within a social body
allows us to examine cultural evolution: the circulation,
selection, and differential propagation of speech patterns
in the group as a whole (rather than, say, tracking the ideas
of a single individual). Patterns of heteroglossia demonstrate existing power relations, create new ones, and are
a key method for the definition of both institutions and
genres [7, 8, 9]. Our methods here quantify a key aspect
2
of cultural evolution: the extent to which one agent’s language patterns are used and copied by another [10, 11].
To study the flow of rhetorical influence and attention in
the NCA over time, we characterize how patterns of language use, uncovered by topic modeling, are propagated
from speech to speech. We do so using Kullback-Leibler
Divergence [12, KLD]: KLD, or “surprise”, from one pattern to another measures the extent to which the expectations of an optimal learner, trained on the first pattern,
are violated by the second. Other work has demonstrated
that surprise (in the Kullback-Leibler sense of the word)
is a cognitive as well as an information-theoretic quantity.
It predicts, for example, what a subject will look at in a
dynamically-evolving visual scene [13]. KLD can be used
to map an individual’s higher-level activities as well, detecting, for example, biographically-significant transitions
in a subject’s intellectual life [4].
Here, we use surprise to analyze a corpus of speeches by
many different individuals. Surprise then measures both
the extent to which a particular speech deviates from the
patterns of prior speeches (novelty), and from the patterns
that will appear in the future (transience). High surprise
compared to the past indicates the topic mixture is new
compared to previous speeches, hence the term “novelty”;
high surprise compared to the future indicates that later
speeches do not retain that pattern very strongly, hence
the term “transience”.
deliberately turn to new topics, but fades away on longer
timescales.
Organizational roles and individual strategies
In choosing when to speak and what to say, speakers had
some control over the relative novelty of their speeches. A
speaker’s attitude toward the Assembly as an institution,
as well as other political or philosophical commitments,
would contribute to his willingness to create new word patterns or copy those of earlier speakers. Conversely, speakers had much less control over the reception of what they
said. Idiosyncratic properties of the speaker, ranging from
the political (e.g., party membership) and the social (e.g.,
demeanor or prestige) to the rhetorical (everything from
word choice to pitch and volume of speech) would have altered the reception of their words, raising or lowering the
transience of the patterns they create.
Differences in strategy and reception at the individual
level, in other words, could break the system-level trends
in favor of innovation. An unskilled but risk-tolerant delegate might have succeeded only rarely, so that his speeches,
while novel, tended to have low, or even negative, resonance. Conversely, the prestige or social power of a
speaker who wished to maintain a line of argument might
have allowed him to simultaneously maintain low novelty
and high resonance, effectively keeping the conversation
on track.
Individual strategies were not the only source of deviation from system-level innovation bias. From the first
days of its meeting, the Assembly organized itself in such
a way as to assign explicit roles to particular speakers.
These information-processing functions over-ruled an individual’s personal characteristics. An example in the NCA,
and common to many later parliaments today, is the role
of president, who served as both a point of contact for the
King and an enforcer of the daily agenda [14]. The president was tasked with a code of conduct and functional
role, and over the course of the NCA 49 delegates served
in this position.
The NCA also created another specialized entity: the
committee. Committees, whose members were notionally
selected on the basis of relevant expertise, deliberated in
private. They developed content outside the debate process and then presented it to the full body for public review. While a speaker on the floor of the Assembly might
play to the audience in the visitor’s galleries, committee
members addressed only each other.
Lexical markers in our data identify when a committee
proxy was speaking, and allow us to classify their speech
into two categories: “new-item” speech and “in-debate”
committee speech (see SI). New-item speeches introduced
official content to the Assembly floor, typically articles and
draft decrees to be reviewed by the legislative body, and
mark transitions in attention from one topic to another.
In-debate speeches occurred when a committee member or
reporter engaged with other delegates following the introduction of the item.
Innovation Bias
Speeches in the NCA span a wide range in both novelty
and transience; Fig. 1 summarizes the system at the level
of individual speeches, in this case at the relatively rapid
timescale (window width, w) of seven speeches. While the
majority of speeches concentrate near the symmetry line—
speeches with high novelty are likely to have similarly high
transience—two results stand out. First, the scatter is
large: there are many speeches that lie far off the novelty–
transience line of equality, and it is easy (for example) to
find speeches with top-quartile novelty that have bottomquartile transience. “What is new is quickly forgotten” is
a useful heuristic, but holds only in the average. Below, we
consider two potential drivers of this diversity of reception:
the speaker, and the context in which the speech was made.
Second, novel speeches are unexpectedly influential. We
quantify this with “resonance,” the imbalance between
novelty and transience (see Materials and Methods). Resonance, the quality of at once breaking from the past and
influencing the future, increases with novelty, as shown
in the rightmost plot of Fig. 1. We refer to this positive
relationship as innovation bias: penalties to high novelty
speeches are lower than expected in a system at equilibrium. This bias is measured by Γ, the slope of the novelty–
resonance line; positive Γ indicates innovation bias. We
find this bias in place from the most rapid timescales (one
speech to the next, w = 1), all the way up to w ≈ 100,
timescales equivalent to a day or so (see SI). This innovation bias lasts for at least the course of a day, as speakers
3
Transience v. Novelty
sample
Resonant sample KLDs
x=y
Transience T
KLD
6
4
10 2
10 1
10 0
counts
2
0
2
4
6
8 10
Novelty N
Resonance R
8
8
0
Resonance v. Novelty
6
4
2
sample
5
fit
0
10 2
10 1
10 0
counts
5
7 5 3 11 3 5 7
Relative speech position
0
2
4
6
Novelty N
8 10
Figure 1: Novelty, transience, and resonance in the French Revolution. Left: a density plot of transience vs. novelty at scale w = 7. Speeches
close to the identity (x = y) line have equal surprise from the past and to the future. Resonant speeches, whose transience is anomalously low
compared to their novelty, break this symmetry and fall below the identity line; the middle plot shows this asymmetry for a marked sample
of high-resonance speeches. Right: resonance vs. novelty, with regression line. Although novelty is tied to transience, and therefore risk of
irrelevance, it is also directly proportional to resonance.
Institutional roles—speaking on behalf of a committee,
presiding over the Assembly—did not follow these systemlevel trends. High novelty and high resonance together
characterize committees as gatherers of new information
that they injected into debate in ways that defined downstream discussion. In contrast, the president’s role as
agenda-enforcer led to lower than average resonance: he
acted, at best, to summarize what had come before, while
having less influence on the patterns of speech that followed. Though he might break from conversation to further the agenda, the content he introduced to do this
tended not to persist. The overall novelty-bias can not
be explained by the taking-in of new information from
committees: while committees did achieve above-average
resonance (z(R) greater than zero), they achieved lower
resonance than expected (∆z(R) less than zero).
Individual strategies are often as defining as top-down
roles. Of the top forty speakers in the assembly, thirty
show significant deviations from aggregate patterns in either novelty or resonance at the p < 0.05 level, with
twenty-three speakers showing deviations at p < 10−3 .
Speakers deviate in both directions, with some showing
anomalously high tendencies to break with past patterns,
and others showing similarly strong tendencies to preserve
them. High-novelty speakers are overwhelmingly associated with the left wing and the bourgeoisie, while all of
our right wing speakers, and the vast majority of nobility,
are low-novelty.
More than half the speakers in our data pursued strategies that distinguished them from the aggregate; significant differences in resonance, by contrast, are less common. The speakers that do show differences in reception,
however, were among the key players of the Revolution.
The celebrated radicals Robespierre and Pétion not only
achieved the highest average resonance in the system, but
To understand how both individual-level strategies and
system-imposed roles affected the production and reception of new patterns, we consider the novelty and resonance of speeches given by those holding three distinct
roles in the Assembly: the 40 most common orators, the
President (regardless of who held the position at the time),
and committee proxies. We calculate both the average
novelty and resonance for each category (scaled by zscore), z(N ) and z(R), tracking both potentially idiosyncratic novelty-seeking (or avoidant) strategies, and how
they are received by the system as a whole. We do this
at scales of w from one (one speech compared to either
the next, or prior, speech) to 5,000 (speech compared to,
roughly, three months of speeches before or after).
The system’s overall bias in favor of innovation would
predict that the category of speakers with the highest
novelty would also have the highest resonance. To determine if this expectation is validated in our sources,
we compare the overall novelty-resonance relationship of
the system (depicted by the fit line in Fig. 1) to the
measured resonance for speeches from each of the three
categories (orator, President, committee proxy). Specifically, for each category we report ∆z(R), defined as
z(R) − E[z(R)|z(N )], the difference between the measured mean resonance of speeches and the expected mean
resonance under the OLS model z(R) ∼ z(N ). While
z(R) measures the effect speakers had on later discourse,
∆z(R) measures the extent to which they broke systemlevel trends in achieving those effects. For example, a
speaker with high novelty may have high resonance but
negative ∆z(R), indicating that his adventurousness was
rewarded less than expected. Full results for w = 36
(roughly half a day) are shown in Table 1; results for novelty are stable on all timescales, while resonance at w = 36
is strongly correlated from w = 3 to w = 100 (see SI).
4
Name
z(N )
z(R)
∆z(R)
Type
+0.25***
+0.14**
+0.15***
+0.09
+0.04
-0.27***
-0.03
-0.17**
-0.10
-0.07***
0.00
-0.02*
3g
3g
3g
2g
3g
–
3g
3g
2g
–
(g)
–
-0.05
-0.13**
-0.11
-0.13***
-0.08***
-0.31***
3g
33g
3g
–
3g
+0.20***
+0.16***
+0.10*
+0.21***
+0.16***
+0.20***
+0.13**
+0.05***
+0.06
+0.10***
3g
3g
3g
2d
3d
1d
3g
–
3g
(d)
+0.04
+0.12*
+0.05
+0.17*
0.0
+0.11
-0.02
-0.07*
-0.11*
3g
2d
2g
2d
3g
2d
2g
3g
3g
of the right-wing overall: while the novelty-biased left was
composed of both high and low resonance speakers, the
right-wing was able to achieve system influence (positive
z(R) and ∆z(R)) despite their anomalously low novelty.
High novelty, high resonance
Jérôme Pétion de Villeneuve
Maximilien Robespierre
Jean-Denis Lanjuinais
Alexandre Lameth
Charles Antoine Chasset
Committee (new item)
Philippe-Antoine Merlin
Pierre-François Gossin
Jacques François Menou
Committee (in debate)
Left Wing
3rd Estate
0.10
0.11**
0.06
0.17**
0.31***
1.31***
0.27***
0.65***
0.40***
0.29***
0.07***
0.06***
0.28***
0.18***
0.16***
0.14*
0.13*
0.12***
0.05
0.03
0.02
0.02
0.02*
0.01
The Emergence and Evolution of the Committee
Committees were a key innovation of the NCA, and allowed the system to manage large amounts of information
without overwhelming the discussions in the parliamentroom itself. They did not, however, appear overnight.
While the previous section establishes their unusual functional role, the comprehensive coverage of the AP makes
it possible to study how their role emerged. In this section, we show how returns to novelty, Γ, are modulated by
committee roles over time. We fit, separately, two terms
that quantify the additional boost (or decrement) to the
novelty-resonance relationship when speeches either introduce new committee items (Γn ), or advocate on behalf of
a committee during debate (Γd ). A speech of novelty N ,
for example, achieves on average a resonance R equal to
(Γ + Γn )(N − N0 ) when made by a committee member
introducing a new item, compared to Γ(N − N0 ) when the
speaker acts on his own behalf.
We look for discrete shifts in committee function, doing
change-point detection with a maximum-likelihood model
of the novelty-resonance relationship where Γ, Γn , and Γd
are allowed to vary in time. Following Ref. [4], we consider a two-epoch model, where all three quantities are
fixed to constant values in each epoch, with a single discrete change at a particular time-point whose position is
a free parameter. The two-epoch model is preferred to a
single-epoch model, as well as to a linear (secular shift)
model under AIC. We find strong evidence for a changepoint in the nature of committee functions occurring at
the end of 1790; the modal best-fit date across all scales is
31 October 1790. Allowing the intercepts of the new-item
and in-debate speeches, as well as their slopes, to vary
produces nearly identical results (see SI). Fig. 2 displays
the novelty-transience relationships for the two epochs,
demonstrating the emergence of a new role for committee
speech over time. In the first epoch, the resonance of new
items was indistinguishable from speech of similar novelty
in the system as a whole. However, speech received a resonance bonus when delivered by a committee member as
part of a debate. In other words, committee representatives injected new information in a fashion similar to other
delegates (new-item speech), but had privileged abilities in
guiding the subsequent debate (in-debate speech).
The pattern is different in the second epoch, where newitem speech has anomalously low resonance at high novelty. Inspection of the speeches themselves suggests that
this high-novelty/low-resonance tail is associated not with
(as in the case of individual speakers) failure to alter the
course of debate, but rather with the increasing power
committees had to “propose and dispose”: once the committee presented their findings to the parliament, they
were increasingly accepted with minimal discussion. In
High novelty, low resonance
Jacques Guillaume Thouret
Jac. Jo. Def. de Chapeliéres
François Denis Tronchet
Armand-Gaston Camus
President
Théodore Vernier
0.16***
0.35***
0.24***
0.29***
0.02
0.55***
0.00
-0.03
-0.04
-0.04
-0.07***
-0.14**
Low novelty, high resonance
Gu. Fr. Ch. Goupil-Préfelne
Jean-François Reubell
Louis Simon Martineau
Jacques An. Ma. de Cazalès
Pierre Victor Malouet
Jean-Sifrein Maury
Pierre-Louis Prieur
1st & 2nd Estates
Jean-Fr. Gaultier de Biauzat
Right Wing
-0.21***
-0.18***
-0.05
-0.44***
-0.27***
-0.46***
-0.27***
-0.10***
-0.13**
-0.32***
0.13**
0.11*
0.08*
0.08*
0.08*
0.07
0.05
0.03***
0.03
0.03*
Low novelty, low resonance
Dominique Garat
Antoine Ch. Ga. Folleville
Lo.-Mi. Le Pel. de Saint-Fargeau
Fr. Do. de Reynaud Montlosier
Pierre-Louis Roederer
Louis Foucauld de Lardimalie
Charles Malo Lameth
Pierre Fran¸ois Bouche
Antoine Barnave
-0.13**
-0.44***
-0.20***
-0.61***
-0.10*
-0.53***
-0.15***
-0.09**
-0.04
0.00
-0.01
-0.01
-0.02
-0.03
-0.05
-0.06
-0.10*
-0.12*
Table 1: Mean novelty and resonance per speaker at scale 36, for
role and type (in bold) and the top forty orators. z(N ): novelty
compared to system average; z(R): resonance compared to system
average; ∆z(R): resonance relative to predicted resonance given novelty. “Type” codes for estate (3: bourgeoisie; 2: nobility; 1: clergy)
and political affiliation (g: gauche, left-; d: droit, right-wing).
significantly higher resonance than even that due to the
system-wide novelty bias (positive ∆z(R)). Conversely,
speakers such as Armand-Gaston Camus and Théodore
Vernier, called on primarily for their specialized knowledge
in canon law and taxation, respectively, show high-novelty,
but low resonance: they presented information, but either
lacked the lasting influence of more prominent speakers, or
were able to settle questions so conclusively that the room
moved on. Finally, prominent political conservatives such
as Jean-Sifrein Maury and Jacques de Cazalès appear in
the low-novelty, high-resonance quadrant. They were able
to break the system-level novelty bias, and are notable not
only for keeping the conversation on track (low novelty),
but for being able to influence others to do the same (high
resonance). In this, Maury and Cazalés are characteristic
5
Γn & Γd
First epoch
0.4
0.2
0.0
0.2
0.4 0
10
Second epoch
in-debate
new-item
10 1
10 2
Scale
10 3
scale=27
10 0
non-committee
in-debate committee
10 1
10 2
Scale
10 3
new-item committee
z(R)
5
0
5
scale=27
4 2 0 2 4 6
z(N)
scale=27
4 2 0 2 4 6
z(N)
Figure 2: Information-processing functions of NCA committees before (1st column) and after (2nd column) the late-1790 change-point. Top
row: the shift in the novelty-resonance relationship for new-item and in-debate committee speech, with 99% confidence intervals. Bottom
row: scatter plots and fit lines at scale 27 for these speech types, compared to all other speeches. The “undebated tail” is seen in the second
epoch, as committees gain new powers to “propose and dispose”.
some cases, committee reports are noted as “passed without debate”; for a sample of these cases, we find the mean
z(N ) equal to 1.57±0.16 and z(R) equal to 0.14±0.13, i.e.,
committee speeches with these hand-annotated outcomes
follow a similarly high-novelty, low-resonance pattern to
other new-item speeches, further validating the interpretation of the “undebated tail” as a signal of emergent committee power. This power can also be seen when substantive debate on committee matters does occur: “in debate”
speeches made by committee members retain a privileged
role in fixing and propagating the patterns that define that
debate.
set of patterns to another than by maintaining the alreadyestablished patterns of the conversations in which they
participated. Indeed, the spatial metaphor of “left vs.
right” is itself misleading: from the point of view of the
debates themselves, the right wing acts as an inertial center, holding off conversation drift, while left-wing speakers produce a wide spectrum of innovations on a much
larger periphery, only some of which survive. These roles
become visible without reference to the ideas that are
propagated: Robespierre emerges as a high-novelty, highresonance speaker even before we consider the content of
the speeches that made him an icon of revolutionary politics.
Discussion
We also find an emergent distinction between orators
and institution players. Talented speakers like Robespierre, Pétion, and Maury could achieve power through
rhetoric and debate, appealing to the crowds in the galleries as much as their colleagues. Yet not every ambitious
delegate could take this road. Quite apart from a delegate’s rhetorical capacities or political stature, the physical venue itself was demanding: loud and (literally) resonant voices were required to be heard. While the NCA
provided a place for rhetorical masters to thrive, in other
words, it also ended up producing a second class of delegate who concentrated on committee work. Deliberating outside the parliament-room, committees contributed
specialized knowledge and abilities and became indispensable workhorses for the emerging government [14]. The
information-processing functions that committees took on
The turbulent months at the beginning of the French
Revolution led to a durable transformation in the very nature of European government [15]. Our study considers
the elite debates at the center of these events, focusing on
pattern transmission rather than the differential propagation of the ideas these patterns may communicate. This
provides a new window into qualitative analyses, which
usually focus on analyzing the logic of particular ideas and
arguments over time [16, 17].
Our analysis reveals clear differences between how the
left- and right-wing figures created and transmitted patterns of language use. Conservatives may indeed “stand
athwart history, yelling Stop” [18]; our results show that
here they did so less by redirecting conversation from one
6
distinguish them clearly from other forms of speech, while
Fig. 2 shows how these functions emerged over time; once
in place, committees are both strong sources of new patterns of speech and new sources of extra-rhetorical power.
Both individuals and institutions, in other words, mattered, playing distinct roles in the development of the parliament over time.
Many delegates welcomed these extra-parliamentary
functions: committees, wrote the assembly member
Jacques Dinochau, “regulate the order of debate, classify
questions, and maintain a continuity of principles, thus
preventing an incoherence which might otherwise have
menaced our decrees” [14]. Yet the private nature of these
committees was in dramatic contrast to the public debates
in the hall itself; their increasing influence a testimony to
the emergent distinction between the spectacle of democracy and the actual ways functioned in a body and polity
too large for direct participation. The dramatic and early
appearance of a specialized information-processing role for
committees provides a new view on their place in modern
democracies, where they emerge endogenously [19] to serve
as essential information-management systems [20].
The power of these committees continued to grow.
While early committees were devoted to technical matters such as monetary and fiscal policy, they were also
instrumental at key moments such as the dismantlement
of feudal and religious privileges. When the revolution collapsed into chaos in 1793, it was committees, such as the
famous “Committee of Public Safety” with Robespierre at
its head, that effectively replaced the republican government and executed thousands of “enemies of the revolution” including many of the speakers in Table 1.
Materials and Methods
Conclusion
Averaging this measure further backwards in the debate
allows us to consider longer-timescale trends, beyond the
back-and-forth of a single exchange, to study the extent
to which the current speaker has taken the debate in an
unexpected direction given, say, the last twenty speeches.
We refer to this quantity as novelty N at time j on scale
w,
The AP is the definitive source for parliamentary transcripts of the Revolution, reconstructed from primary
sources including transcripts, minutes, and newspaper reports. The French Revolution Digital Archive (https:
//frda.stanford.edu) is a digital version, with full-text
speeches and TEI markup from the beginning of the NCA
on 9 July 1789 through its end on 30 September 1791.
After stop-word removal, each speech is represented as
a count vector over vocabulary of 10, 000 most common
words. The resulting corpus contains 4, 765, 773 words
in 44, 913 speeches. Each speech is matched to a named
speaker, and may be tagged with a “role”: in some cases,
a speech made by a delegate acting as the assembly president; in other cases, on behalf of a committee, either to
introduce a report, or to comment on it. See Supplementary Information (SI; linked as ancillary file on the arXiv)
for details on corpus preparation. All of the delegates to
the Assembly, and all of the speakers in our data, were
male.
We use LDA [3] to model speeches as probability distributions over K = 100 topics; the jth speech in the data is
(j)
si , 1 ≤ i ≤ K.
Novelty, Transience, Resonance
Novelty at the smallest timescales (a speech compared
to the one just previous) is measured by the KLD of s(j)
relative to s(j−1)
!
K
(j)
X
si
(j)
(j) (j−1)
.
(1)
KLD s |s
=
si log2
(j−1)
si
i=1
The history of human culture is more than just the rise
and fall of particular ideas. It is also the emergence of new
information-processing mechanisms and media, and roles
that individuals and institutions could play in creating and
selectively propagating these ideas through time. In the
language of biological evolution, we must understand not
only the particular characteristics for which an environment selects, but the strength of that selection over time,
and the shifting and heterogenous nature of the transmission mechanisms.
By quantifying the flow of patterns between speakers,
we can see revolutionary debates as not just a battle of
ideas, but as struggle for the direction of the conversation
itself. At the earliest stages of the revolution, political
players adopted not just different ideological positions, but
different roles in the propagation of patterns, with varying
levels of innovation and fidelity to the past. And together,
both cooperatively and agonistically, they invented new
mechanisms for the collective management of information.
w
1 X
Nw (j) =
KLD s(j) |s(j−d) ,
w
(2)
d=1
illustrated in Fig. 3.
Any speech can break abruptly from its past—but the
new patterns it introduces may not persist. Consider an
interjection that other speakers ignore in order to return to
the matter at hand. That interjection would be surprising
given its past, but equally surprising given its future, symmetric surprise under time reversal, because all the new information is lost, or transient. In contrast, a rhetorically
effective interjection might shift the conversation to a new
issue with differing vocabulary. This shift would appear
as a surprise asymmetry about the interjection. We define
7
[2] J. G. Foster, A. Rzhetsky, J. A. Evans, Tradition and innovation
in scientists’ research strategies, American Sociological Review
80 (5) (2015) 875–908.
[3] D. M. Blei, A. Y. Ng, M. I. Jordan, Latent Dirichlet allocation,
Journal of Machine Learning Research (JMLR) 3 (2003) 993–
1022.
[4] J. Murdock, C. Allen, S. DeDeo, Exploration and exploitation
of Victorian science in Darwin’s reading notebooks, Cognition
159 (2017) 117–126.
[5] W. L. Benoit, M. J. Smythe, Rhetorical theory as message reception: A cognitive response approach to rhetorical theory and
criticism, Communication Studies 54 (1) (2003) 96–114.
[6] R. A. Harris, Reception studies in the rhetoric of science, Technical Communication Quarterly 14 (3) (2005) 249–255.
[7] M. Foucault, Power/Knowledge: Selected interviews and other
writings, 1972–1977, Pantheon, New York, NY, USA, 1980.
[8] C. Danescu-Niculescu-Mizil, L. Lee, B. Pang, J. Kleinberg,
Echoes of power: Language effects and power differences in social interaction, in: Proceedings of the 21st international conference on World Wide Web, ACM, 2012, pp. 699–708.
[9] S. Klingenstein, T. Hitchcock, S. DeDeo, The civilizing process
in London’s Old Bailey, Proceedings of the National Academy
of Sciences 111 (26) (2014) 9419–9424.
[10] M. M. Bakhtin, The dialogic imagination: Four essays, University of Texas Press, Austin, TX, USA, 2010.
[11] R. A. Blythe, W. Croft, S-curves and the mechanisms of propagation in language change, Language 88 (2) (2012) 269–304.
[12] S. Kullback, R. Leibler, On information and sufficiency, Annals
of Mathematical Statistics 22 (1) (1951) 79–86.
[13] L. Itti, P. Baldi, Bayesian surprise attracts human attention,
Vision Research 49 (10) (2009) 1295–1306.
[14] T. Tackett, Becoming a Revolutionary, Princeton University
Press, Princeton, NJ, USA, 2014.
[15] W. H. Sewell, Historical events as transformations of structures:
Inventing revolution at the Bastille, Theory and society 25 (6)
(1996) 841–881.
[16] K. M. Baker, Inventing the French Revolution: essays on French
political culture in the eighteenth century, Cambridge University Press, Cambridge, UK, 1990.
[17] K. M. Baker, D. Edelstein, Scripting revolution: a historical approach to the comparative study of revolutions, Stanford University Press, Stanford, CA, USA, 2015.
[18] W. F. Buckley Jr., Our mission statement, National Review 1.
[19] D. P. Baron, Legislative organization with informational committees, American Journal of Political Science 44 (3) (2000)
485–505.
[20] K. A. Shepsle, B. R. Weingast, Positive theories of Congressional institutions, Legislative Studies Quarterly 19 (2) (1994)
149–179.
center
speech
time
... ...
... ...
speeches
j-w
j-d
j-1
j
j+1
j+d
j+w
⎫
⎬
⎭
⎫
⎬
⎭
Novelty
Transience
Figure 3: Diagram of novelty and transience. Each speech is represented by an LDA topic
mixture. The Kullback-Leibler Diver
gence KLD s(j) |s(j−d) measures surprise of speech at time j given
a speech d steps away. Novelty at scale w is the mean surprise of a
central speech given speeches in a past window of w steps. Transience
is novelty under time reversal, measuring the mean surprise from the
future. A rhetorically effective speech shifts current conversation to
a new issue with a differing vocabulary, manifesting itself as a surprise asymmetry where novelty exceeds transience. We define this
positive asymmetry as resonance (Eq. 3): the ability to break from
the past and influence the future. An illustration of this asymmetry
is shown in the middle panel of Figure 1.
this asymmetry as “resonance”, R:
Rw (j) =
w
i
1 Xh
KLD s(j) |s(j−d) − KLD s(j) |s(j+d)
w
d=1
≡ Nw (j) − Tw (j)
(3)
Resonance is novelty minus transience, T , where the latter is novelty in Eq. 2 under time reversal. Novel speeches
which also influence future discourse are pivot points in
conversation. For a system in equilibrium, deviations
are time-symmetric in the expectation—i.e., the measured
novelty and transience at any point in time are, on average,
equal. For a non-equilibrium system significant deviations
may appear both for individual samples and for the overall average. Novelty’s effectiveness, Γ, is the rate at which
]
; approximated
resonance increases with novelty, dE[R|N
dN
here using a linear model.
Acknowledgments
J.H. acknowledges a Research Experience for Undergraduates National Science Foundation Grant #ACI-1358567
at the Santa Fe Institute.
References
[1] B. D. Jones, F. R. Baumgartner, The politics of attention: How
government prioritizes problems, University of Chicago Press,
Chicago, IL, USA, 2005.
8
| 7cs.IT
|
arXiv:1401.5097v1 [cs.PL] 20 Jan 2014
Distributed call-by-value machines
Olle Fredriksson
University of Birmingham, UK
Abstract
We present a new abstract machine, called DCESH, which describes the execution of higher-order programs running in distributed architectures. DCESH
implements a generalised form of Remote Procedure Call that supports calling
higher-order functions across node boundaries, without sending actual code.
Our starting point is a variant of the SECD machine that we call the CES
machine, which implements reduction for untyped call-by-value PCF. We successively add the features that we need for distributed execution and show the
correctness of each addition. First we add heaps, forming the CESH machine,
which provides features necessary for more efficient execution, and show that
there is a bisimulation between the CES and the CESH machine. Then we construct a two-level operational semantics, where the high level is a network of
communicating machines, and the low level is given by local machine transitions. Using these networks, we arrive at our final system, the distributed CESH
machine (DCESH). We show that there is a bisimulation relation also between
the CESH machine and the DCESH machine. All the technical results have
been formalised and proved correct in Agda, and a prototype compiler has been
developed.
1
Contents
1
Seamless computing
3
2 The CES machine
5
3 CESH: A heap machine
10
3.1
An interface for heaps . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
3.2
Correctness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
4 Synchronous and asynchronous networks
16
5 DCESH1 : A degenerate distributed machine
19
6 DCESH: The distributed CESH machine
22
6.1
7
Correctness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Related work
28
32
8 Conclusion and further work
33
2
1
Seamless computing
Suppose we need to program a system in which the function F runs on node A in a
distributed system, for instance because F depends on a local resource residing on
node A. Suppose further that we need to write a program G, running on node B, that
uses F. How to achieve this depends on what programming language or library for
distributed computing we choose. One of the most prominent ways to do it is using
message passing, for instance with the Message-Passing Interface [1]. This involves
writing F and G as separate processes, and explicitly constructing messages that are
sent between them.
Suppose now that our specification changes: A part F’ of F actually needs to run
on a third node C. Using conventional languages or libraries, this means that we
have to rewrite big parts of the program since a substantial part of it deals with
the architecture-specific details of the problem. Languages with support for Remote
Procedure Calls [2] can help mitigate this, since such a call has the same syntax as a
local procedure call, but will not work if F’ is a higher-order function that is invoked
with a function as its argument. In previous papers [3, 4] we suggest the following
alternative way to express the two programs above:
...} @ A in {G} @ B
let F = {... F’
let F = {... {F’} @ C ...} @ A in {G} @ B
Here we write the whole program as if it was running on a single computer, and
use pragma-like annotations, written {x} @ A, to indicate the node of execution. We
call such annotations locus specifiers. The compiler uses the annotations to automatically handle architecture-specific details like communication. We call this seamless
computing. A key feature is full support for higher-order functions, even across node
boundaries, without sending actual code (in contrast to e.g. Remote Evaluation [5]).
This is important for full generality, since it is not always the case that all code
is meaningful on all nodes (for example because of resource locality or platform
differences).
Our previous work enables writing these programs but uses an execution model
based on game semantics that is vastly different from conventional compilation
techniques. In this paper we instead develop an approach which is a conservative extension of existing abstract machines. This means that the vast literature on
compiler optimisation more readily applies, and makes it possible to interface with
legacy code. The key idea in this work, like in our previous work, is that computational phenomena like function calls can be subsumed by simple communication
protocols. We assume that a run-time infrastructure can handle system-level aspects
associated with distribution such as failure, load balancing, global reset, and so on.
Technical outline To achieve the goal of an abstract machine for seamless computing, we make gradual refinements to a machine, based on Landin’s SECD machine [6], that we call the CES machine (Sec. 2). The first change is to add heaps
(Sec. 3.1) for dynamically allocating closures, forming the CESH machine (Sec. 3),
3
which provides features necessary for more efficient execution, and we show the
CES and CESH machines to be bisimilar (Sec. 3.2). We then add communication
primitives (synchronous and asynchronous) by defining a general form of networks
of nodes that run an instance of an underlying abstract machine (Sec. 4). Using
these networks, we illustrate the idea of subsuming function calls by communication
protocols by constructing a degenerate distributed machine, DCESH1 (Sec. 5), that
decomposes some machine instructions into message passing, but only runs on one
node. Finally, the main contribution is the fully distributed CESH machine (DCESH,
Sec. 6), which is shown to be bisimilar to the CESH machine (Sec. 6.1).
Formalisation in Agda The theorems that we present in this paper have been
proved correct in Agda [7], an interactive proof assistant and programming language
based on intuitionistic type theory. The definitions and proofs in this paper are
intricate and often consist of many cases, so carrying them out manually would be
error-prone and arduous. Agda has been a helpful tool in producing these proofs,
and also allows us to easily play with alternative definitions (even wrong ones). To
eliminate another source of error, we do not adopt the usual practice of writing up
the results in informal mathematics; in fact, the paper is built from a collection of
literate Agda source files and the code blocks come directly from the formalisation.
Although our work is not about Agda per se, we believe that this presentation is
beneficial also to you, the reader, since you can trust that the propositions do not
contain mistakes. Since Agda builds on a constructive foundation, it also means
that the formalisation of an abstract machine in Agda can act as a verified prototype
implementation.
Syntax and notation for code We assume a certain familiarity with the syntax
of Agda, but since it is close to that of several popular functional programming
languages we believe that this will not cause much difficulty for the audience. We
will use ⋆ for the type of types. We will use implicit parameters, written e.g. f :
{A : ⋆ } → ... which means that f takes, as its first argument, a type A that does not
need to be explicitly spelled out when it can be inferred from other arguments. We
will sometimes use the same name for constructors of different types, and rely on
context for disambiguation. Constructors will be written in bold face and keywords
underlined. We make liberal use of Agda’s ability to define mixfix operators like
if0_then_else_ which is a constructor that accepts arguments in the positions of the
underscores, as in if0 b then t else f.
This paper is organised as follows, where the arrows denote dependence, the lines
with ∼ symbols bisimulations, and the parenthesised numerals section numbers:
4
CES
(2)
∼
(3.2)
∼
(6.1)
CESH
(3)
Heaps
DCESH
(6)
Networks
(4)
(3.1)
DCESH1
(5)
2 The CES machine
Our goal is to make a compiler for a programming language with locus specifiers
that is based on conventional compilation techniques. A very common technique
is the usage of abstract machines to describe the evaluation at a level low enough
to be used as a basis for compilation. The starting point for our work is based
on a variation of Landin’s well-studied SECD machine [6] called Modern SECD [8].
Modern SECD itself can be traced back to the SECD machine of Henderson [9], in
that both use bytecode for the control component of the machine (and so use explicit
return instructions); and to the CEK machine of Felleisen [10], in that they both place
the continuations that originally resided in the dump (the D component) directly on
the stack (the S component), simplifying the machine configurations.
We choose to call this variation the CES machine because of its three configuration
constituents. This machine is important for us since it will be used as the specification for the elaborated machines that we later construct. We will show that their
termination and divergence behaviour is the same as that of CES by constructing
bisimulation relations.
A CES configuration (Config) is a tuple consisting of a fragment of code (Code), an
environment (Env), and a stack (Stack). Evaluation begins with an empty stack and
environment, and then follows a stack discipline. Sub-terms push their result on the
stack so that their super-terms can consume them. When (and if) the evaluation
terminates, the program’s result is the sole stack element.
Source language We show how to compile untyped call-by-value PCF [11]. The
source language has constructors for lambda abstractions (λ t), applications (t $ t’),
and variables (var n). Our representation uses De Bruijn indices [12], so a variable is
simply a natural number.
data Term : ⋆ where
λ_ : Term → Term
_$_ : (t t’ : Term) → Term
var : N → Term
We also have natural number literals, binary operations on them, and conditionals:
lit : N → Term
op : (f : N → N → N) (t t’ : Term) → Term
5
if0_then_else_ : (b t f : Term) → Term
The language can be thought of as an intermediate representation for a compiler
which may expose a more sugary front-end language. Because it is untyped, we can
express fixed-point combinators without adding additional constructors.
We define the bytecode, Code, that the machine will operate on. A fragment of Code
is a list of instructions, Instr, terminated by END, RET, or a conditional COND which
has code fragments for its two branches:
mutual
data Instr : ⋆ where
VAR : N
→ Instr
CLOS : Code → Instr
APPL : Instr
LIT
:N
→ Instr
OP
: (N → N → N) → Instr
data Code : ⋆ where
_;_
: Instr → Code → Code
COND : Code → Code → Code
END : Code
RET : Code
The main work of compilation is done by the function compile’, which takes a term
t to be compiled and a fragment of code c that is placed after the instructions that
the compilation emits.
compile’ : Term → Code → Code
compile’ (λ t)
c = CLOS (compile’ t RET) ; c
compile’ (t $ t’) c = compile’ t (compile’ t’ (APPL ; c))
compile’ (var x) c = VAR x ; c
compile’ (lit n)
c = LIT n ; c
compile’ (op f t t’) c = compile’ t’ (compile’ t (OP f ; c))
compile’ (if0 b then t else f) c =
compile’ b (COND (compile’ t c) (compile’ f c))
It should be apparent that the instructions correspond closely to the constructs of
the source language but are sequentialised. Compilation of a term is simply a call to
compile’, terminated by END:
compile : Term → Code
compile t = compile’ t END
Example 2.1 (codeExample). The term (λx. x) (λx y. x) is compiled as follows:
compile ((λ var 0) $ (λ (λ var 1))) =
CLOS (VAR 0 ; RET) ;
CLOS (CLOS (VAR 1 ; RET) ; RET) ; APPL ; END
Compilation first emits two CLOS instructions containing the code of the function
and its argument. The APPL instruction is then used to perform the actual application.
6
We mutually define values, closures and environments. A closure is a code fragment
paired with an environment. A value is either a natural number literal or a closure.
Since we are working in a call-by-value setting an environment is a list of values.
mutual
Closure = Code × Env
data Value : ⋆ where
nat : N
→ Value
clos : Closure → Value
Env = List Value
A stack is a list of stack elements, defined to be either values or continuations
(represented by closures):
data StackElem : ⋆ where
val : Value → StackElem
cont : Closure → StackElem
Stack = List StackElem
A configuration is, as stated, a tuple consisting of a code fragment, an environment
and a stack:
Config = Code × Env × Stack
Fig. 1 shows the definition of the transition relation for configurations of the CES
machine. The Agda syntax may require some further explanation: The instructions’
constructor names are overloaded to also act as constructors for the relation; their
usage will be disambiguated by context. We use implicit arguments, written in curly
braces, for arguments that can automatically be inferred and do not need to be
spelled out explicitly. The type of propositional equality is written _≡_.
The stack discipline becomes apparent in the definition of the transition relation.
When e.g. VAR is executed, the CES machine looks up the value of the variable
in the environment and pushes it on the stack. A somewhat subtle part of the
relation is the interplay between the APPL instruction and the RET instruction. When
performing an application, two values are required on the stack, one of which has to
be a closure. The machine enters the closure, adding the value to the environment,
and pushes a return continuation on the stack. Looking at the compile function, we
see that the code inside a closure will be terminated by a RET instruction, so once
the machine has finished executing the closure (and thus produced a value on the
stack), that value is returned to the continuation.
Example 2.2. We trace the execution of codeExample defined above, which exemplifies how returning from an application works. Here we write a −−−−→ h x i b
meaning that the machine uses rule x to transition from a to b.
let c1 = VAR 0 ; RET
c2 = CLOS (VAR 1 ; RET) ; RET
cl1 = val (clos (c1 , [ ])); cl2 = val (clos (c2 , [ ]))
7
CES
data _ −−−−→ _ : Rel Config Config where
CES
8
VAR
: ∀ {n c e s v} → lookup n e ≡ just v → (VAR n ; c, e, s)
−−−−→ (c, e, val v :: s)
CLOS
: ∀ {c’ c e s} →
−−−−→ (c, e, val (clos (c’, e)) :: s)
APPL
: ∀ {c e v c’ e’ s} → (APPL ; c, e, val v :: val (clos (c’, e’)) :: s)
−−−−→ (c’, v :: e’, cont (c, e) :: s)
RET
: ∀ {e v c e’ s} →
(RET, e, val v :: cont (c, e’) :: s)
−−−−→ (c, e’, val v :: s)
LIT
: ∀ {n c e s} →
(LIT n ; c, e, s)
−−−−→ (c, e, val (nat n) :: s)
OP
: ∀ {f c e n1 n2 s} → (OP f ; c, e, val (nat n1 ) :: val (nat n2 ) :: s) −−−−→ (c, e, val (nat (f n1 n2 )) :: s)
COND-0
: ∀ {c c’ e s} →
(CLOS c’ ; c, e, s)
CES
CES
CES
CES
CES
CES
COND-1+n : ∀ {c c’ e n s} →
(COND c c’, e, val (nat 0) :: s)
−−−−→ (c, e, s)
(COND c c’, e, val (nat (1 + n)) :: s)
−−−−→ (c’, e, s)
CES
CES
Figure 1: The definition of the transition relation of the CES machine.
in (CLOS c1 ; CLOS c2 ; APPL ; END, [ ], [ ])
−−−−→ h CLOS i (CLOS c2 ; APPL ; END, [ ], [cl1 ])
CES
−−−−→ h CLOS i
(APPL ; END, [ ], [cl2 , cl1 ])
−−−−→ h APPL i
(VAR 0 ; RET, [cl2 ], [cont (END, [ ])])
CES
CES
−−−−→ h VAR refl i (RET, [cl2 ], [cl2 , cont (END, [ ])])
CES
−−−−→ h RET i
CES
(END, [ ], [cl2 ])
The final result is therefore the second closure, cl2 .
Lemma 2.3 (determinismCES ). −−−−→ is deterministic. In the formalisation this means
CES
that we can construct the following term:
determinismCES : _ −−−−→ _ is-deterministic
CES
where the type _is-deterministic is defined as follows:
_is-deterministic : {A B : ⋆ } → Rel A B → ⋆
R is-deterministic =
∀ {a b b’} → R a b → R a b’ → b ≡ b’
This is a key property that is useful when constructing a compiler implementation
of the CES machine. Note that we write the name of the definition containing this
proof in parentheses.
The CES machine terminates with a value v, written cfg ↓CES v if it, through the
reflexive transitive closure of −−−−→, reaches the end of its code fragment with an
CES
empty environment, and v as its sole stack element.
_ ↓CES _ : Config → Value → ⋆
cfg ↓CES v = cfg −−−−→∗ (END, [ ], val v :: [ ])
CES
where the reflexive transitive closure of a relation is defined as follows:
data _* {A : ⋆ } (R : Rel A A) (a : A) : A → ⋆ where
[ ] : (R *) a a
_::_ : {b c : A} → R a b → (R *) b c → (R *) a c
It terminates, written cfg ↓CES if there exists a value v such that it terminates with the
value v. The Agda syntax for the existential quantifier normally written as ∃x.P(x)
is ∃ λ x → P x. Using this syntax, the definition of termination with value v is:
_ ↓CES : Config → ⋆
cfg ↓CES = ∃ λ v → cfg ↓CES v
It diverges, written cfg ↑CES if it is possible to take another step from any configuration reachable from the reflexive transitive closure of −−−−→.
CES
9
_ ↑CES : Config → ⋆
_ ↑CES = ↑ _ −−−−→ _
CES
where
↑ : {A : ⋆ } (R : Rel A A) → A → ⋆
↑ R a = ∀ b → (R *) a b → ∃ λ c → R b c
3 CESH: A heap machine
In a compiler implementation of the CES machine targeting a low-level language,
closures have to be dynamically allocated in a heap. However, the CES machine
does not make this dynamic allocation explicit. In this section, we try to make it
explicit by constructing the CESH machine, which is a CES machine with an extra
heap component in its configuration.
While heaps are not strictly necessary for a presentation of the CES machine, they are
of great importance to us. The distributed machine that we will later define needs
heaps for persistent storage of data, and the CESH machine forms an intermediate
step between that and the CES machine. Another thing that can be done with
heaps is to implement general recursion, without using fix-point combinators, by
constructing circular closures.
A CESH configuration is defined as Config = Code × Env × Stack × Heap Closure,
where Heap is a type constructor for heaps parameterised by the type of its content.
Closures, values and environments are again mutually defined. Now a closure value
is simply represented by a pointer:
ClosPtr = Ptr
mutual
Closure = Code × Env
data Value : ⋆ where
nat : N
→ Value
clos : ClosPtr → Value
Env = List Value
We leave the stack as in the CES machine (even though we could, in principle, change
the continuations, currently represented by closures, to pointers as well – we do not
do this since it is not necessary for our purposes).
data StackElem : ⋆ where
val : Value → StackElem
cont : Closure → StackElem
Stack = List StackElem
Fig. 2 shows the definition of the transition relation of the CESH machine. It is
largely the same as that of the CES machine, but with the added heap component.
10
data _ −−−−−→ _ : Rel Config Config where
CESH
CLOS
: ∀ {c’ c e s h} → let (h’, ptrcl ) = h ◮ (c’, e) in
(CLOS c’ ; c, e, s, h) −−−−−→ (c, e, val (clos ptrcl ) :: s, h’)
APPL
: ∀ {c e v ptrcl c’ e’ s h} → h ! ptrcl ≡ just (c’, e’) →
(APPL ; c, e, val v :: val (clos ptrcl ) :: s, h) −−−−−→ (c’, v :: e’, cont (c, e) :: s, h)
VAR
: ∀ {n c e s h v} → lookup n e ≡ just v →
CESH
CESH
(VAR n ; c, e, s, h) −−−−−→ (c, e, val v :: s, h)
11
CESH
RET
: ∀ {e v c e’ s h} →
(RET, e, val v :: cont (c, e’) :: s, h) −−−−−→ (c, e’, val v :: s, h)
LIT
: ∀ {l c e s h} →
OP
: ∀ {f c e l1 l2 s h} → (OP f ; c, e, val (nat l1 ) :: val (nat l2 ) :: s, h) −−−−−→ (c, e, val (nat (f l1 l2 )) :: s, h)
COND-0
: ∀ {c c’ e s h} →
CESH
(LIT l ; c, e, s, h) −−−−−→ (c, e, val (nat l) :: s, h)
CESH
CESH
COND-1+n : ∀ {c c’ e n s h} →
(COND c c’, e, val (nat 0) :: s, h) −−−−−→ (c, e, s, h)
CESH
(COND c c’, e, val (nat (1 + n)) :: s, h) −−−−−→ (c’, e, s, h)
CESH
Figure 2: The definition of the transition relation of the CESH machine.
The difference appears in the CLOS and APPL instructions. To build a closure, the
machine allocates it in the heap using the _◮_ function, which, given a heap and an
element, gives back an updated heap and a pointer to the element. When performing
an application, the machine has a pointer to a closure, so it looks it up in the heap
using the _!_ function, which, given a heap and a pointer, gives back the element
that the pointer points to (if it exists).
Lemma 3.1 (determinismCESH ). −−−−−→ is deterministic, i.e. we can define the followCESH
ing term:
determinismCESH : _ −−−−−→ _ is-deterministic
CESH
We define what it means for a CESH configuration cfg to terminate with a value v
(cfg ↓CESH v), terminate (cfg ↓CESH ), and diverge (cfg ↑CESH ). These are analogous
to the definitions for the CES machine, with the difference that the CESH machine
is allowed to terminate with any heap since it never deallocates anything:
_ ↓CESH _ : Config → Value → ⋆
cfg ↓CESH v = ∃ λ h → cfg −−−−−→∗ (END, [ ], [val v], h)
CESH
_ ↓CESH : Config → ⋆
cfg ↓CESH = ∃ λ v → cfg ↓CESH v
_ ↑CESH : Config → ⋆
_ ↑CESH = ↑ _ −−−−−→ _
CESH
3.1
An interface for heaps
In this section we formally define an abstract interface for the type constructor Heap
and its associated functions that we use to model the dynamic memory allocation
that we will need in our system. We will leave the details unspecified, requiring
instead that it captures certain algebraic properties that should be fulfilled by any
reasonable implementation.
The type Heap A models heaps with memory cells of type A, and Ptr pointers into
some heap. We require the existence of a heap ∅ without any other requirements.
abstract
Heap : ⋆ → ⋆
Ptr
:⋆
∅
: {A : ⋆ } → Heap A
We need to be able to attempt to lookup (or dereference) pointers, and allocate new
items. Allocating gives back a new heap and a pointer.
_!_
_◮_
: {A : ⋆ } → Heap A → Ptr → Maybe A
: {A : ⋆ } → Heap A → A → Heap A × Ptr
12
We require the following relationship between dereferencing and allocation: if we
dereference a pointer that was obtained from allocating a memory cell with value x,
we get x back:
!-◮
: {A : ⋆ } (h : Heap A) (x : A) →
let (h’, ptr) = h ◮ x in h’ ! ptr ≡ just x
We define a preorder ⊆ for sub-heaps. The intuitive reading for h ⊆ h’ is that h’ can
be used where h can, i.e. that h’ contains at least the allocations of h.
_⊆_
: {A : ⋆ } → Heap A → Heap A → ⋆
h1 ⊆ h2 = ∀ ptr {x} → h1 ! ptr ≡ just x → h2 ! ptr ≡ just x
⊆-refl : {A : ⋆ } (h : Heap A) → h ⊆ h
⊆-refl h ptr eq = eq
⊆-trans : {A : ⋆ } {h1 h2 h3 : Heap A}
→ h1 ⊆ h2 → h2 ⊆ h3 → h1 ⊆ h3
⊆-trans h1 ⊆h2 h2 ⊆h3 ptr eq = h2 ⊆h3 ptr (h1 ⊆h2 ptr eq)
Our last requirement is that allocation does not overwrite any memory cells that
were previously allocated (proj1 means first projection):
abstract
h⊆h◮x : {A : ⋆ } (h : Heap A) {x : A} → h ⊆ proj1 (h ◮ x)
3.2
Correctness
To show that our definition of the machine is correct, we construct a bisimulation
between the CES and CESH machines. Since the configurations of the machines
are very similar, the intuition for the relation that we construct is simply that it
is almost equality. The only place where it is not equality is for closure values,
where the CESH machine stores pointers instead of closures directly. To construct
a relation for closure values we need to to parameterise it by the heap of the CESH
configuration, and then make sure that the closure pointer points to a closure related
to the CES closure.
Formally, the relation is constructed separately for the different components of the
machine configurations. Since they run the same bytecode, we let the relation for
code be equality:
RCode : Rel Code Code
RCode c1 c2 = c1 ≡ c2
We forward declare the relation for environments and define the relation for closures,
which is simply that the components of the closures are related. Since we have used
the same names for some of the components of the CES and CESH machines, we
qualify them, using Agda’s qualified imports, by prepending CES. and CESH. to their
names. These components may contain values, so we have to parameterise the
relations by a closure heap (here ClosHeap = Heap CESH.Closure).
13
REnv : ClosHeap → Rel CES.Env CESH.Env
RClos : ClosHeap → Rel CES.Closure CESH.Closure
RClos h (c1 , e1 ) (c2 , e2 ) = RCode c1 c2 × REnv h e1 e2
Two values are unrelated if they do not start with the same constructor. When they
do start with the same constructor, there are two cases: If the two values are number
literals, they are related if they are equal. If the two values are a CES closure and a
pointer, we require that the pointer points to a CESH closure that is related to the
CES closure.
RVal : ClosHeap → Rel CES.Value CESH.Value
RVal h (nat n1 ) (nat n2 ) = n1 ≡ n2
RVal h (nat ) (clos ) = ⊥
RVal h (clos ) (nat )
= ⊥
RVal h (clos c1 ) (clos ptr) = ∃ λ c2 →
h ! ptr ≡ just c2 × RClos h c1 c2
Two environments are related if they have the same list spine and their values are
pointwise related.
REnv h [ ]
[]
REnv h [ ]
(x2 :: e2 )
REnv h (x1 :: e1 ) [ ]
REnv h (x1 :: e1 ) (x2 :: e2 )
=
=
=
=
⊤
⊥
⊥
RVal h x1 x2 × REnv h e1 e2
Note that we use ⊤ and ⊥ to represent true and false, represented in Agda by the unit
type and the uninhabited type. The relation on stacks, RStack is defined similarly,
using RVal and RClos for values and continuations.
RStackElem : ClosHeap → Rel CES.StackElem CESH.StackElem
RStackElem h (val v1 ) (val v2 ) = RVal h v1 v2
RStackElem h (val ) (cont ) = ⊥
RStackElem h (cont ) (val )
= ⊥
RStackElem h (cont c1 ) (cont c2 ) = RClos h c1 c2
RStack : ClosHeap → Rel CES.Stack CESH.Stack
RStack h [ ]
[]
= ⊤
RStack h [ ]
(x2 :: s2 ) = ⊥
RStack h (x1 :: s1 ) [ ]
= ⊥
RStack h (x1 :: s1 ) (x2 :: s2 ) = RStackElem h x1 x2 × RStack h s1 s2
Finally, two configurations are related if their components are related. Here we pass
the heap of the CESH configuration as an argument to the environment and stack
relations.
RCfg : Rel CES.Config CESH.Config
RCfg (c1 , e1 , s1 ) (c2 , e2 , s2 , h2 ) =
RCode c1 c2 × REnv h2 e1 e2 × RStack h2 s1 s2
14
Lemma 3.2 (HeapUpdate.config). Given two heaps h and h’ such that h ⊆ h’, if
RCfg cfg (c, e, s, h), then RCfg cfg (c, e, s, h’). 1
config : ∀ cfg c e s → RCfg cfg (c, e, s, h) → RCfg cfg (c, e, s, h’)
Theorem 3.3 (simulation). RCfg is a simulation relation.
simulation : Simulation _ −−−−→ _ _ −−−−−→ _ RCfg
CES
CESH
where
Simulation _−→_ _−→’_ _R_ = ∀ a a’ b →
a −→ a’ → a R b → ∃ λ b’ → b −→’ b’ × a’ R b’
Proof. By cases on the CES transition. In each case, the CESH machine can make
analogous transitions. Use HeapUpdate.config to show that RCfg is preserved.
We call a relation a presimulation if it is almost, but not quite, a simulation:
Presimulation _−→_ _−→’_ _R_ = ∀ a a’ b →
a −→ a’ → a R b → ∃ λ b’ → b −→’ b’
Theorem 3.4 (presimulation). The inverse of RCfg is a presimulation.
presimulation : Presimulation _ −−−−−→ _ _ −−−−→ _
CESH
CES
(RCfg -1 )
Lemma 3.5 (presimulation-to-simulation). If R is a simulation between relations −→
and −→’, R -1 is a presimulation, and −→’ is deterministic at states b related to some
a, then R -1 is a simulation.
presimulation-to-simulation : (_R_ : Rel A B)
→ Simulation −→ −→’ _R_
→ Presimulation −→’ −→ (_R_ -1 )
→ (∀ a b → a R b → −→’ is-deterministic-at b)
→ Simulation −→’ −→ (_R_ -1 )
where _is-deterministic-at is a weaker form of determinism:
_is-deterministic-at_ : (R : Rel A B) → A → ⋆
_R_ is-deterministic-at a = ∀ {b b’} → a R b → a R b’ → b ≡ b’
1 In the actual implementation, this is inside a local module HeapUpdate, parameterised by h and h’
and their relation, together with similar lemmas for the constituents of the machine configurations.
15
Theorem 3.6 (bisimulation). RCfg is a bisimulation.
bisimulation : Bisimulation _ −−−−→ _ _ −−−−−→ _ RCfg
CES
CESH
where
Bisimulation −→ −→’ R = Simulation −→ −→’ R × Simulation −→’ −→ (R -1 )
Proof. Theorem presimulation-to-simulation applied to determinismCESH and simulation
implies that RCfg -1 is a simulation, which together with simulation shows that RCfg
is a bisimulation.
Corollary 3.7 (termination-agrees, divergence-agrees). In particular, a CES configuration terminates with a natural number n (diverges) if and only if a related CESH
configuration terminates with the same number (diverges):
termination-agrees : ∀ cfg1 cfg2 n →
RCfg cfg1 cfg2 → cfg1 ↓CES nat n ↔ cfg2 ↓CESH nat n
divergence-agrees : ∀ cfg1 cfg2 →
RCfg cfg1 cfg2 → cfg1 ↑CES ↔ cfg2 ↑CESH
These results are of course not useful until we can show that there are configurations in RCfg . One such example is the “initial” (mostly empty) configuration for a
fragment of code:
initial-related : ∀ c → RCfg (c, [ ], [ ]) (c, [ ], [ ], ∅)
initial-related c = refl, tt, tt
4 Synchronous and asynchronous networks
Since we are later going to define two distributed abstract machines, it would save
us some work if we could make a network model that is general enough to be used
for both. In this section we will define models for synchronous and asynchronous
networks, that are parameterised by an underlying labelled transition system. Both
kinds of networks are modelled by two-level transition systems, which is common in
operational semantics for concurrent and parallel languages. The idea is that the
global level describes the transitions of the system as a whole, and the low level
the local transitions of the nodes in the system. Synchronous communication is
modelled by rendezvous, i.e. that two nodes have to be ready to send and receive a
message at a single point in time. Asynchronous communication is modelled using a
“message soup”, representing messages currently in transit, that nodes can add and
remove messages from, reminiscent of the Chemical Abstract Machine [13].
16
We construct an Agda module Network, parameterised by the underlying transition
_
relation, _ ⊢ _ −−−−−−−−→ _ : Node → Machine → Tagged Msg → Machine → ⋆ . The
Machine
sets Node, Machine, and Msg are additional parameters. Elements of Node will
act as node identifiers, and we assume that these enjoy decidable equality. If we
were using MPI, they would correspond to the so called node ranks, which are just
machine integers. The type Machine is the type of the nodes’ configurations, and
Msg the type of messages that the machines can send. The Node in the type of
_
_ ⊢ _ −−−−−−−−→ _ means, intuitively, that the configuration of a node knows about and
Machine
can depend on its own identifier. The type constructor Tagged is used to separate
different kinds of local transitions: A Tagged Msg can be silent (i.e. a τ transition),
send msg, or receive msg (for msg : Msg).
module Network
(Node : ⋆)
?
(_=_ : (n n’ : Node) → Dec (n ≡ n’))
{Machine Msg : ⋆ }
_
(_ ⊢ _ −−−−−−−−→ _ : Node → Machine →
Machine
Tagged Msg → Machine → ⋆)
where
A synchronous network (SyncNetwork) is an indexed family of machines, Node → Machine,
representing the nodes of the system. An asynchronous network (AsyncNetwork) is
an indexed family of machines together with a list of messages representing the
messages currently in transit, (Node → Machine) × List Msg.
The following function updates an element in a set indexed by node identifiers, and
will be used in defining the transition relations for networks:
update : {A : ⋆ } → (Node → A) → Node → A → Node → A
?
update nodes n m n’ with n’ = n
update nodes n m n’ | yes p = m
update nodes n m n’ | no ¬p = nodes n’
Fig. 3 shows the definition of the transition relation for synchronous and asynchronous networks.
There are two ways for a synchronous network to make a transition. The first,
silent-step, occurs when a machine in the network makes a transition tagged with
silent, and is allowed at any time. The second, comm-step , is the aforementioned
rendezvous. A node s first takes a step sending a message, and afterwards a node
r takes a step receiving the same message. Note that s and r are not necessarily
different, i.e. nodes can send messages to themselves. Asynchronous networks only
have one rule, step, which can be used if a node steps with a tagged message that
“agrees” with the list of messages in transit. The definition is fairly involved, but
the intuition is that if the node receives a message, the message has to be in the list
before the transition. If the node sends a message, it has to be there after. If the node
17
data _ −−−−−→ _ (nodes : SyncNetwork) : SyncNetwork → ⋆ where
Sync
silent
silent-step : ∀ {i m’} → i ⊢ nodes i −−−−−−−−→ m’ → nodes −−−−−→ update nodes i m’
Machine
Sync
comm-step : ∀ {s r msg sender’ receiver’} → let nodes’ = update nodes s sender’ in
send msg
receive msg
Machine
Machine
s ⊢ nodes s −−−−−−−−−−→ sender’ → r ⊢ nodes’ r −−−−−−−−−−−−→ receiver’ →
nodes −−−−−→ update nodes’ r receiver’
Sync
18
data _ −−−−−−→ _ : AsyncNetwork → AsyncNetwork → ⋆ where
Async
step : ∀ {nodes} msgsl msgsr {tmsg m’ i} → let (msgsin , msgsout ) = detag tmsg in
tmsg
i ⊢ nodes i −−−−−−−−→ m’ →
Machine
(nodes, msgsl ++ msgsin ++ msgsr ) −−−−−−→ (update nodes i m’, msgsl ++ msgsout ++ msgsr )
Async
Figure 3: The definition of the transition relations for synchronous and asynchronous networks.
takes a silent step, the list stays the same before and after. This is what the usage of
the detag function, which creates lists of input and output messages from a tagged
message, achieves:
detag : {A : ⋆ } → Tagged A → List A × List A
detag silent
= [ ] ,[ ]
detag (send x)
= [ ] , [x]
detag (receive x) = [x], [ ]
Lemma 4.1. If we have a synchronous transition from a to b, then we have one or
more asynchronous transitions from (a, [ ]) to (b, [ ]), as follows:
−−−−−→ -to-−−−−−−→ + : ∀ {a b} → a −−−−−→ b → (a, [ ]) −−−−−−→ + (b, [ ])
Sync
Async
Sync
−−−−−→ -to-−−−−−−→ + (silent-step s)
Sync
Async
= [step [ ] [ ] s]
Async
−−−−−→ -to-−−−−−−→ + (comm-step s1 s2 ) = step [ ] [ ] s1 :: [step [ ] [ ] s2 ]
Sync
Async
where _+ is defined as follows:
data _+ {A : ⋆ } (R : Rel A A) (a : A) : A → ⋆ where
[ ] : {b : A} → R a b → (R + ) a b
_::_ : {b c : A} → R a b → (R + ) b c → (R + ) a c
We can thus say that asynchronous networks subsume synchronous networks. Going
in the other direction is not possible in general, but for some specific instances of
the underlying transition relation it is, as we will see later.
5 DCESH1 : A degenerate distributed machine
In higher-order distributed programs containing locus specifiers, we will sometimes
encounter situations where a function is not available locally. For example, when
evaluating the function f in the term (f @ A) (g @ B), we may need to apply the
remotely available function g. As stated in the introduction, our general idea is to
do this by decomposing some instructions into communication. In the example, the
function f may send a message requesting the evaluation of g, meaning that the APPL
instruction is split into a pair of instructions: APPL-send and APPL-receive.
This section outlines an abstract machine, called DCESH1 , which decomposes all
application and return instructions into communication. The machine is degenerate,
because it runs as the sole node in a network and sends messages to itself, but
illustrates this decomposition, which will be used in the fully distributed system.
A configuration of the DCESH1 machine (Machine) is a tuple consisting of a possibly
running thread (Maybe Thread), a closure heap (Heap Closure), and a “continuation
heap” (Heap (Closure × Stack)). Since the current work does not support parallelism,
19
we have at most one thread running at once. The thread resembles a CES configuration, Thread = Code × Env × Stack, but stacks are defined differently. A stack is now
a list of values paired with an optional pointer (pointing into the continuation heap),
Stack = List Val × Maybe ContPtr (where ContPtr is a more descriptive synonym for
Ptr). The intuition here is that when performing an application, when CES would
push a continuation on the stack, the DCESH1 machine is going to stop the current
thread and send a message, which means that it has to save the continuation and
the remainder of the stack in the heap for them to persist the thread’s lifetime.
The optional pointer in Stack is to be thought of as being an element at the bottom
of the list of values. Comparing it to the definition of the CES machine, where stacks
are lists of either values or continuations (which were just closures), we can picture
their relation: Whereas the CES machine stores the values and continuations in a
single, contiguous stack, the DCESH1 machine stores first a contiguous block of
values until reaching a continuation, at which point it stores (just) a pointer to the
continuation closure and the rest of the stack.
The definition of closures, values, and environments are otherwise just like in the
CESH machine.
ClosPtr = Ptr
mutual
Closure = Code × Env
data Val : ⋆ where
nat : N
→ Val
clos : ClosPtr → Val
Env = List Val
ClosHeap = Heap Closure
ContPtr = Ptr
Stack
= List Val × Maybe ContPtr
ContHeap = Heap (Closure × Stack)
Thread
= Code × Env × Stack
Machine = Maybe Thread × ClosHeap × ContHeap
The machine communicates with itself using two kinds of messages, APPL and RET,
corresponding to the instructions that we are replacing with communication.
data Msg : ⋆ where
APPL : ClosPtr → Val → ContPtr → Msg
RET : ContPtr → Val → Msg
tmsg
Fig. 4 defines the transition relation for the DCESH1 machine, written m −−−−−→ m’
for a tagged message tmsg and machine configurations m and m’. Most transitions
are the same as in the CESH machine, just framed with the additional heaps and the
just meaning that the thread is running. The interesting rules are the decomposed
application and return rules. When an application is performed, an APPL message
containing a pointer to the closure to apply, the argument value and a pointer to a
return continuation (which is first allocated) is sent, and the thread is stopped (represented by the nothing). The machine can receive an application message if the
20
_
data _ −
→ _ : Machine → Tagged Msg → Machine → ⋆ where
VAR
: ∀ {n c e s v r hcl hcnt } → lookup n e ≡ just v →
CLOS
(just (VAR n ; c, e, s, r), hcl , hcnt )
−−−−−−→
: ∀ {c’ c e s r hcl hcnt } → let (h’cl , ptrcl ) = hcl ◮ (c’, e) in
silent
(just (CLOS c’ ; c, e, s, r), hcl , hcnt )
APPL-send
silent
−−−−−−→
(just (c, e, v :: s, r), hcl , hcnt )
(just (c, e, clos ptrcl :: s, r), h’cl , hcnt )
: ∀ {c e v ptrcl s r hcl hcnt } → let (h’cnt , ptrcnt ) = hcnt ◮ ((c, e), s, r) in
send (APPL ptrcl v ptrcnt )
(just (APPL ; c, e, v :: clos ptrcl :: s, r), hcl , hcnt ) −−−−−−−−−−−−−−−−−−−−−−−−→
APPL-receive : ∀ {hcl hcnt ptrcl v ptrcnt c e} → hcl ! ptrcl ≡ just (c, e) →
(nothing, hcl , h’cnt )
receive (APPL ptrcl v ptrcnt )
RET-send
(nothing, hcl , hcnt ) −−−−−−−−−−−−−−−−−−−−−−−−−−−→ (just (c, v :: e, [ ], just ptrcnt ), hcl , hcnt )
: ∀ {e v ptrcnt hcl hcnt } →
send (RET ptrcnt v)
21
(just (RET, e, v :: [ ], just ptrcnt ), hcl , hcnt )
−−−−−−−−−−−−−−−−−−→
(nothing, hcl , hcnt )
RET-receive : ∀ {hcl hcnt ptrcnt v c e s r} → hcnt ! ptrcnt ≡ just ((c, e), s, r) →
(nothing, hcl , hcnt )
COND-0
receive (RET ptrcnt v)
−−−−−−−−−−−−−−−−−−−−−→
: ∀ {c c’ e s r hcl hcnt } →
silent
(just (COND c c’, e, nat 0 :: s, r), hcl , hcnt )
COND-1+n : ∀ {c c’ e n s r hcl hcnt } →
−−−−−−→
(just (COND c c’, e, nat (1 + n) :: s, r), hcl , hcnt )
LIT
: ∀ {l c e s r hcl hcnt } →
−−−−−−→
(just (LIT l ; c, e, s, r), hcl , hcnt )
: ∀ {f c e l1 l2 s r hcl hcnt } →
−−−−−−→
(just (OP f ; c, e, nat l1 :: nat l2 :: s, r), hcl , hcnt )
−−−−−−→
OP
(just (c, e, v :: s, r), hcl , hcnt )
silent
silent
silent
(just (c, e, s, r), hcl , hcnt )
(just (c’, e, s, r), hcl , hcnt )
(just (c, e, nat l :: s, r), hcl , hcnt )
(just (c, e, nat (f l1 l2 ) :: s, r), hcl , hcnt )
Figure 4: The definition of the transition relation of the DCESH1 machine.
thread is not running. When that happens, the closure pointer is dereferenced and
entered, adding the received argument to the environment. The stack is left empty
apart from the continuation pointer of the received message. When returning from
a function application, the machine sends a return message containing the continuation pointer and the value to return. On the receiving end of that communication,
it dereferences the continuation pointer and enters it, putting the result value on top
of the stack.
Example 5.1. We show what happens when we have instantiated the asynchronous
networks of the Network module with this transition relation, using the unit (oneelement) type for the Node set. Once again we trace the execution of our running
example, codeExample. For readability, we write heaps with pointer mappings like
{ptr 7→ element}. The last list shown in each step is the message list of the asynchronous network.
let hcl = {ptr1 7→ (c1 , [ ])}
h’cl = {ptr1 7→ (c1 , [ ]), ptr2 7→ (c2 , [ ])}
hcnt = {ptrcnt 7→ ((END, [ ]), [ ], nothing)}
in (just (CLOS c1 ; CLOS c2 ; APPL ; END, [ ], [ ],
nothing), ∅, ∅), [ ]
−→h step CLOS i
(just (CLOS c2 ; APPL ; END, [ ], [clos ptr1 ],
nothing), hcl , ∅), [ ]
−→h step CLOS i
(just (APPL ; END, [ ], [clos ptr2 , clos ptr1 ],
nothing), h’cl , ∅), [ ]
−→h step APPL-send i
(nothing, h’cl , hcnt ), [APPL ptr1 (clos ptr2 ) ptrcnt ]
−→h step APPL-receive i
(just (VAR 0 ; RET, [clos ptr2 ], [ ],
just ptrcnt ), h’cl , hcnt ), [ ]
−→h step (VAR refl) i
(just (RET, [clos ptr2 ], [clos ptr2 ],
just ptrcnt ), h’cl , hcnt ), [ ]
−→h step RET-send i
(nothing, h’cl , hcnt ), [RET ptrcnt (clos ptr2 )]
−→h step RET-receive i
(just (END, [ ], [clos ptr2 ], nothing), h’cl , hcnt ), [ ]
Comparing this to Example 2.2 we can see that an APPL-send followed by an
APPL-receive amounts to the same thing as the APPL rule in the CES machine,
and similarly for the RET instruction.
6 DCESH: The distributed CESH machine
We have so far seen two extensions of the CES machine. We have seen CESH, that
adds heaps, and DCESH1 , that decomposes instructions into communication in a
22
degenerate network of only one node. Our final extension is a machine, DCESH,
that supports multiple nodes. The main problem that we now face is that there
is no centralised heap, but each node has its own local heap. This means that, for
supporting higher-order functions across node boundaries, we have to somehow keep
references to closures in the heaps of other nodes. Another problem is efficiency; we
would like a system where we do not pay the higher price of communication for
locally running code. The main idea for solving these two problems is to use remote
pointers, RPtr = Ptr × Node, pointers paired with node identifiers signifying on what
node’s heap the pointer is located. This solves the heap problem because we always
know where a pointer comes from. It can also be used to solve the efficiency problem
since we can choose what instructions to run based on whether a pointer is local or
remote. If it is local, we run the rules of the CESH machine. If it is remote, we run
the decomposed rules of the DCESH1 machine.
The final extension to the term language and bytecode will add support for locus
specifiers.
data Term : ⋆ where
...
_@_ : Term → Node → Term
data Instr : ⋆ where
...
REMOTE : Code → Node → Instr
The locus specifiers, t @ i, are taken to mean that the term t should be evaluated on
node i. For simplicity, we assume that the terms t in all locus specification sub-terms
t @ i are closed. This is a reasonable assumption, since a term where this does not
hold can be transformed into one where it does with roughly similar behaviour, using
e.g. lambda lifting [14] 2 . The REMOTE c i instruction will be used to start running a
code fragment c on node i in the network. We also extend the compile’ function to
handle the new term construct:
compile’ : Term → Code → Code
...
compile’ (t @ i) c = REMOTE (compile’ t RET) i ; c
Note that we reuse the RET instruction to return from a remote computation.
Once again we assume that we are given a set Node with decidable equality:
module DCESH
(Node : ⋆)
?
(_=_ : (n n’ : Node) → Dec (n ≡ n’))
where
The intended meaning of a remote pointer RPtr is that it is a pointer located in the
heap of the given node. We assume once again that the set Node has decidable
2 Transform every sub-term t @ i to t’ = ((λ fv t. t) @ i) (fv t). These have “roughly similar behaviour” in
that the semantics of t and t’ are identical under the assumption that locus specifiers do not change the
meaning of a program.
23
equality meaning that we can, for instance, determine if an RPtr is remote or local.
This generalises the DCESH1 machine, since we can now hold pointers pointing to
something in the heap of another node’s machine.
RPtr = Ptr × Node
ClosPtr = RPtr
The definition of closures, values, environments and closure heaps are the same as
in the CESH machine, but using RPtr instead of Ptr for closure pointers.
mutual
Closure = Code × Env
data Value : ⋆ where
nat : N
→ Value
clos : ClosPtr → Value
Env = List Value
ClosHeap = Heap Closure
The stack combines the functionality of the CES(H) machine, permitting local continuations, with that of the DCESH1 machine, making it possible for a stack to end
with a continuation on another node. A stack element is a value or a (local) continuation signified by the val and cont constructors. A stack (Stack) is a list of stack
elements, possibly ending with a (remote) pointer to a continuation.
data StackElem : ⋆ where
val : Value → StackElem
cont : Closure → StackElem
ContPtr = RPtr
Stack
= List StackElem × Maybe ContPtr
ContHeap = Heap (Closure × Stack)
Threads and machines are defined like in the DCESH1 machine.
Thread = Code × Env × Stack
Machine = Maybe Thread × ClosHeap × ContHeap
The messages that DCESH can send are those of the DCESH1 machine but using remote pointers instead of plain pointers, plus a message for starting a remote
computation, REMOTE c i rptrcnt .
data Msg : ⋆ where
REMOTE : Code → Node → ContPtr → Msg
RET
: ContPtr → Value → Msg
APPL
: ClosPtr → Value → ContPtr → Msg
Note that sending a REMOTE message amounts to sending code in our formalisation,
which is something that we said that it would not do. However, because no code is
generated at run-time, every machine can be “pre-loaded” with all the bytecode it
needs, and the message only needs to contain a reference to a fragment of code.
24
_
data _⊢_ −
→ _ (i : Node) : Machine → Tagged Msg → Machine → ⋆ where
VAR
: ∀ {n c e s v r hcl hcnt } → lookup n e ≡ just v →
i⊢
CLOS
(just (VAR n ; c, e, s, r), hcl , hcnt )
−−−−−−→
: ∀ {c’ c e s r hcl hcnt } → let (h’cl , rptrcl ) = i ⊢ hcl ◮ (c’, e) in
i⊢
APPL
(just (CLOS c’ ; c, e, s, r), hcl , hcnt )
−−−−−−→
: ∀ {c e v c’ e’ s r ptrcl hcl hcnt } → hcl ! ptrcl ≡ just (c’, e’) →
silent
silent
25
−−−−−−→
i⊢
LIT
(just (RET, e, val v :: cont (c, e’) :: s, r), hcl , hcnt )
: ∀ {n c e s r hcl hcnt } →
−−−−−−→
i⊢
OP
(just (LIT n ; c, e, s, r), hcl , hcnt )
: ∀ {f c e n1 n2 s r hcl hcnt } →
−−−−−−→
i ⊢ (just (OP f ; c, e, val (nat n1 ) :: val (nat n2 ) :: s, r), hcl , hcnt )
COND-0
: ∀ {c c’ e s r hcl hcnt } →
−−−−−−→
i⊢
COND-1+n
(just (COND c c’, e, val (nat 0) :: s, r), hcl , hcnt )
: ∀ {c c’ e n s r hcl hcnt } →
−−−−−−→
(just (COND c c’, e, val (nat (1 + n)) :: s, r), hcl , hcnt )
−−−−−−→
REMOTE-send
(just (c, e, val (clos rptrcl ) :: s, r), h’cl , hcnt )
silent
i ⊢ (just (APPL ; c, e, val v :: val (clos (ptrcl , i)) :: s, r), hcl , hcnt )
RET
: ∀ {e v c e’ s r hcl hcnt } →
i⊢
(just (c, e, val v :: s, r), hcl , hcnt )
(just (c’, v :: e’, cont (c, e) :: s, r), hcl , hcnt )
silent
(just (c, e’, val v :: s, r), hcl , hcnt )
silent
(just (c, e, val (nat n) :: s, r), hcl , hcnt )
silent
(just (c, e, val (nat (f n1 n2 )) :: s, r), hcl , hcnt )
silent
(just (c, e, s, r), hcl , hcnt )
silent
(just (c’, e, s, r), hcl , hcnt )
: ∀ {c’ i’ c e s r hcl hcnt } → let (h’cnt , rptr) = i ⊢ hcnt ◮ ((c, e), s, r) in
i⊢
(just (REMOTE c’ i’ ; c, e, s, r), hcl , hcnt )
REMOTE-receive : ∀ {hcl hcnt c rptrcnt } →
send (REMOTE c’ i’ rptr)
−−−−−−−−−−−−−−−−−−−−−−−→
(nothing, hcl , h’cnt )
receive (REMOTE c i rptrcnt )
i⊢
APPL-send
(nothing, hcl , hcnt ) −−−−−−−−−−−−−−−−−−−−−−−−−−−→ (just (c, [ ], [ ], just rptrcnt ), hcl , hcnt )
: ∀ {c e v ptrcl j s r hcl hcnt } → i . j → let (h’cnt , rptrcnt ) = i ⊢ hcnt ◮ ((c, e), s, r) in
i⊢
APPL-receive
(just (APPL ; c, e, val v :: val (clos (ptrcl , j)) :: s, r), hcl , hcnt ) −−−−−−−−−−−−−−−−−−−−−−−−−−−−→ (nothing, hcl , h’cnt )
: ∀ {hcl hcnt ptrcl v rptrcnt c e} → hcl ! ptrcl ≡ just (c, e) →
i⊢
RET-send
(nothing, hcl , hcnt ) −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→ (just (c, v :: e, [ ], just rptrcnt ), hcl , hcnt )
: ∀ {e v rptrcnt hcl hcnt } →
i⊢
RET-receive
(just (RET, e, val v :: [ ], just rptrcnt ), hcl , hcnt )
−−−−−−−−−−−−−−−−−−−→
: ∀ {hcl hcnt ptrcnt v c e s r} → hcnt ! ptrcnt ≡ just ((c, e), s, r) →
i⊢
send (APPL (ptrcl ,j) v rptrcnt )
receive (APPL (ptrcl ,i) v rptrcnt )
send (RET rptrcnt v)
(nothing, hcl , hcnt )
receive (RET (ptrcnt ,i) v)
−−−−−−−−−−−−−−−−−−−−−−−→
(nothing, hcl , hcnt )
(just (c, e, val v :: s, r), hcl , hcnt )
Figure 5: The definition of the transition relation of the DCESH machine.
tmsg
Fig. 5 defines the transition relation of the DCESH machine, written i ⊢ m −−−−−→ m’
for a node identifier i, a tagged message tmsg and machine configurations m and
m’. The parameter i is taken to be the identifier of the node on which the transition
is taking place. Most instructions are similar to those of the CESH machine but
adapted to this new setting, using remote pointers. The following function is used to
allocate a pointer in a heap on a node i, yielding a new heap and a remote pointer
(pointing to the node i):
_⊢_◮_ : {A : ⋆ } → Node → Heap A → A → Heap A × RPtr
i ⊢ h ◮ x = let (h’, ptr) = h ◮ x in h’, (ptr, i)
When an application occurs and the closure pointer is on the current node, i, the
machine dereferences the pointer and enters it locally. If there is a local continuation
on the stack and the machine is to run the return instruction, it also works just
like the original CES machine. When starting a remote computation, the machine
allocates a continuation in the heap and sends a message containing the code and
continuation pointer to the remote node in question. Afterwards the current thread
is stopped. On the receiving end of such a communication, a new thread is started,
placing the continuation pointer at the bottom of the stack for the later return to the
caller node. To run the apply instruction when the function closure is remote, i.e. its
location is not equal to the current node, the machine sends a message containing
the closure pointer, argument value, and continuation, like in the DCESH1 machine.
On the other end of such a communication, the machine dereferences the pointer
and enters the closure with the received value. The bottom remote continuation
pointer is set to the received continuation pointer. After either a remote invocation
or a remote application, the machine can return if it has produced a value on the
stack and has a remote continuation at the bottom of the stack. To do this, a message
containing the continuation pointer and the return value is sent to the location of
the continuation pointer. When receiving a return message, the continuation pointer
is dereferenced and entered with the received value.
Now that we have defined the transition relation for machines we instantiate the
Network module with the −→Machine relation.
?
open import Network Node _=_ −→Machine public
From here on SyncNetwork and AsyncNetwork and their transition relations will thus
refer to the instantiated versions.
An initial network configuration, given a code fragment c and a node identifier i, is
a network where only node i is active, ready to run the code fragment:
initial-networkSync : Code → Node → SyncNetwork
initial-networkSync c i = update (λ i’ → (nothing, ∅, ∅))
i (just (c, [ ], [ ], nothing), ∅, ∅)
An initial asynchronous network configuration is one where there are no messages
in message list:
26
initial-networkAsync : Code → Node →
AsyncNetwork
initial-networkAsync c i = initial-networkSync c i, [ ]
Lemma 6.1. The communication that the local step relation enables is point-to-point:
if two nodes can receive the same message, then they are the same:
point-to-point :
∀ i1 i2 (ms : SyncNetwork) msg {m1 m2 } →
receive msg
i1 ⊢ ms i1 −−−−−−−−−−−−→ m1 →
receive msg
i2 ⊢ ms i2 −−−−−−−−−−−−→ m2 →
i1 ≡ i2
We call a machine inactive if its thread is not running, i.e. it is equal to nothing.
inactive : Machine → ⋆
inactive (t, ) = t ≡ nothing
Lemma 6.2 (determinismSync ). If all nodes in a synchronous network except one
are inactive, then the next step is deterministic.
determinismSync : ∀ nodes i →
all nodes except i are inactive → _ −−−−−→ _ is-deterministic-at nodes
Sync
where
all_except_are_ : {A B : ⋆ } → (A → B) → A → (B → ⋆) → ⋆
all f except i are P = ∀ i’ → i’ . i → P (f i’)
Lemma 6.3 (−−−−−−→ + -to- −−−−−→ + ). If all nodes in a synchronous network except one
Sync
Async
are inactive and the network takes one or more steps asynchronously from and to
configurations without any messages in the air, then that transition can also be done
synchronously.
−−−−−−→ + -to- −−−−−→ + : ∀ {nodes nodes’} i → all nodes except i are inactive →
Async
Sync
(nodes, [ ]) −−−−−−→ + (nodes’, [ ]) → nodes −−−−−→ + nodes’
Sync
Async
This is a key result because it means that it does not matter whether we choose
to look at synchronous or asynchronous networks for single threaded computations.
With this result in place, we will from now on focus on the simpler synchronous
networks.
We define what it means for a synchronous DCESH network nodes to terminate with
a value v (nodes ↓Sync v), terminate (nodes ↓Sync ), and diverge (nodes ↑Sync ). A
27
network terminates with a value v if it can step to a network where only one node is
active, and that node has reached the END instruction with the value v on top of its
stack. The other definitions are analogous to those of the CES(H) machine.
_ ↓Sync _ : SyncNetwork → Value → ⋆
nodes ↓Sync v = ∃ λ nodes’ → nodes −−−−−→∗ nodes’ ×
Sync
∃ λ i → all nodes’ except i are inactive × ∃ λ heaps →
nodes’ i ≡ (just (END, [ ], val v :: [ ], nothing), heaps)
_ ↓Sync : ∀ nodes → ⋆
nodes ↓Sync = ∃ λ v → nodes ↓Sync v
_ ↑Sync : ∀ nodes → ⋆
_ ↑Sync = ↑ _ −−−−−→ _
Sync
6.1 Correctness
To prove the correctness of the machine, we will now establish a bisimulation between the CESH and the DCESH machines.
To simplify this development, we extend the CESH machine with a rule for the
REMOTE c i instruction so that both machines run the same bytecode. This rule
is almost a no-op, but since we are assuming that the code we run remotely is
closed, the environment is emptied, and since the compiled code c will end in a RET
instruction a return continuation is pushed on the stack.
data _ −−−−−→ _ : Rel Config Config where
CESH
...
REMOTE : ∀ {c’ i c e s h} →
(REMOTE c’ i ; c, e, s, h) −−−−−→ (c’, [ ], cont (c, e) :: s, h)
CESH
The intuition behind the relation that we are to construct should be similar to the
intuition for the relation between CES and CESH configurations, i.e. that it is almost
equality, but since values may be pointers to closures, we need to parameterise it
by heaps. The problem now is that both machines use pointers, and the DCESH
machine even uses remote pointers and has two heaps for each node. This means
that we have to parameterise the relations by all the heaps in the system.
As before, two fragments of code are related if they are equal.
RCode : Rel Code Code
RCode c1 c2 = c1 ≡ c2
We define the type of the extra parameter that we need as a synonym for an
indexed family of the closure and continuation heaps (here DCESH.ContHeap =
Heap (DCESH.Closure × DCESH.Stack)):
28
Heaps = Node → DCESH.ClosHeap × DCESH.ContHeap
Simply following the recipe that we used for the relation between the CES and the
CESH machines would not prove effective this time around. When we constructed
that, we could be sure that there would be no circularity, since it was constructed
inductively on the structure of the CES configuration. But now both systems, CESH
and DCESH, have heaps where there is a potential for circular references (e.g. a
closure, residing in a heap, whose environment contains a pointer to itself), so a
direct structural induction cannot work. This is perhaps the most mathematically
(and formally) challenging point of the paper. To fix this we parameterise the affected relation definitions by a natural number rank, which records how many times
pointers are allowed to be dereferenced, in addition to the heap parameters.
The relation for environments and closures is as before, but with the additional
parameters.
REnv : N → CESH.ClosHeap → Heaps → Rel CESH.Env DCESH.Env
RClos : N → CESH.ClosHeap → Heaps → Rel CESH.Closure DCESH.Closure
RClos rank h hs (c1 , e1 ) (c2 , e2 ) = RCode c1 c2 × REnv rank h hs e1 e2
The relation for closure pointers is where the rank is used. If the rank is zero, the
relation is trivially fulfilled. If the rank is non-zero, it makes sure that the CESH
pointer points to a closure in the CESH heap, that the remote pointer of the DCESH
network points to a closure in the heap of the location that the pointer refers to, and
that the two closures are related:
Rrptrcl : N → CESH.ClosHeap → Heaps →
Rel CESH.ClosPtr DCESH.ClosPtr
Rrptrcl 0
= ⊤
Rrptrcl (1 + rank) h hs ptr1 (ptr2 , loc) =
∃2 λ cl1 cl2 → h ! ptr1 ≡ just cl1 ×
proj1 (hs loc) ! ptr2 ≡ just cl2 ×
RClos rank h hs cl1 cl2
The relation for values is also as before, but with the extra parameters.
RVal : N → CESH.ClosHeap → Heaps → Rel CESH.Value DCESH.Value
RVal rank h hs (nat n1 ) (nat n2 ) = n1 ≡ n2
= ⊥
RVal rank h hs (nat ) (clos )
RVal rank h hs (clos ) (nat )
= ⊥
RVal rank h hs (clos ptr) (clos rptr) = Rrptrcl rank h hs ptr rptr
The relation for environments is also as before, but included for completeness:
REnv rank h hs [ ]
[]
= ⊤
REnv rank h hs [ ]
(x :: e2 ) = ⊥
REnv rank h hs (x1 :: e1 ) [ ]
= ⊥
REnv rank h hs (x1 :: e1 ) (x2 :: e2 ) = RVal rank h hs x1 x2 × REnv rank h hs e1 e2
The relation for stack elements is almost as before, but now requires that for any natural number rank, i.e. for any finite number of pointer dereferencings, the relations
hold:
29
RStackElem : CESH.ClosHeap → Heaps →
Rel CESH.StackElem DCESH.StackElem
RStackElem h hs (val v1 ) (val v2 ) =
∀ rank → RVal rank h hs v1 v2
RStackElem h hs (val )
(cont ) = ⊥
RStackElem h hs (cont ) (val )
= ⊥
RStackElem h hs (cont cl1 ) (cont cl2 ) =
∀ rank → RClos rank h hs cl1 cl2
The relation for stacks now takes into account that the DCESH stacks may end in
a pointer representing a remote continuation. It makes sure that the pointer points
to something in the continuation heap of the location of the pointer, related to the
CESH stack element.
RStack : CESH.ClosHeap → Heaps →
Rel CESH.Stack DCESH.Stack
RStack h hs [ ]
([ ], nothing) = ⊤
RStack h hs [ ]
(x :: stack2 , r) = ⊥
RStack h hs (x1 :: stack1 ) (x2 :: stack2 , r) = RStackElem h hs x1 x2 ×
RStack h hs stack1 (stack2 , r)
RStack h hs (x :: stack1 ) ([ ], nothing) = ⊥
RStack h hs [ ]
([ ], just )
= ⊥
RStack h hs (cont1 :: s1 ) ([ ], just (ptr, loc)) =
∃2 λ cont2 s2 → proj2 (hs loc) ! ptr ≡ just (cont2 , s2 ) ×
RStackElem h hs cont1 (cont cont2 ) ×
RStack h hs s1 s2
Finally, a CESH configuration and a DCESH thread are related if the thread is
running and the constituents are pointwise related:
RThread : Heaps → Rel Config (Maybe Thread)
nothing
= ⊥
RThread hs
RThread hs (c1 , e1 , s1 , h1 ) (just (c2 , e2 , s2 )) =
RCode c1 c2 × (∀ rank → REnv rank h1 hs e1 e2 ) ×
RStack h1 hs s1 s2
A configuration is related to an asynchronous DCESH network if the network has
exactly one running node, i, that is related to the configuration, and there are no
messages in the message soup:
RAsync : Rel Config AsyncNetwork
RAsync cfg (nodes, [ ])
= ∃λi→
all nodes except i are inactive ×
RThread (proj2 ◦ nodes) cfg (proj1 (nodes i))
RAsync cfg (nodes, msgs) = ⊥
A configuration is related to a synchronous DCESH network if it is related to the
asynchronous network gotten by pairing the synchronous network with an empty list
of messages:
30
RSync : Rel Config SyncNetwork
RSync cfg nodes = RAsync cfg (nodes, [ ])
We order heaps of a DCESH network pointwise as follows (called ⊆s since it is the
“plural” of ⊆):
_⊆s_ : (hs hs’ : Heaps) → ⋆
hs ⊆s hs’ = ∀ i → let (hcl , hcnt ) = hs i
(h’cl , h’cnt ) = hs’ i
in hcl ⊆ h’cl × hcnt ⊆ h’cnt
⊆s-refl : (hs : Heaps) → hs ⊆s hs
⊆s-refl hs node = let (hcl , hcnt ) = hs node
in ⊆-refl hcl , ⊆-refl hcnt
⊆s-trans : {hs1 hs2 hs3 : Heaps} →
hs1 ⊆s hs2 → hs2 ⊆s hs3 → hs1 ⊆s hs3
⊆s-trans hs1 ⊆shs2 hs2 ⊆shs3 node
= let (clh1 ⊆clh2 , conth1 ⊆conth2 ) = hs1 ⊆shs2 node
(clh2 ⊆clh3 , conth2 ⊆conth3 ) = hs2 ⊆shs3 node
in ⊆-trans clh1 ⊆clh2 clh2 ⊆clh3 ,
⊆-trans conth1 ⊆conth2 conth2 ⊆conth3
Lemma 6.4 (HeapUpdate.env, HeapUpdate.stack). Given CESH closure heaps h and
h’ such that h ⊆ h’ and families of DCESH heaps hs and hs’ such that hs ⊆s hs’, then
we can prove the following:
env : ∀ {n} e1 e2 → REnv n h hs e1 e2 → REnv n h’ hs’ e1 e2
stack : ∀ s1 s2 → RStack h hs s1 s2 → RStack h’ hs’ s1 s2
Theorem 6.5 (simulationSync ). RSync is a simulation relation.
simulationSync : Simulation _ −−−−−→ _ _ −−−−−→ _ RSync
CESH
Sync
Proof. By cases on the CESH transition. In each case, the DCESH network can
make analogous transitions. Use the HeapUpdate lemmas to show that RSync is
preserved.
Theorem 6.6 (presimulationSync ). The inverse of RSync is a presimulation.
presimulationSync : Presimulation _ −−−−−→ _ _ −−−−−→ _ (RSync -1 )
Sync
CESH
Theorem 6.7 (bisimulationSync ). RSync is a bisimulation.
bisimulationSync : Bisimulation _ −−−−−→ _ _ −−−−−→ _ RSync
CESH
31
Sync
Proof. Theorem presimulation-to-simulation applied to determinismSync and simulationSync
implies that RSync -1 is a simulation, which together with simulationSync shows that
RSync is a bisimulation.
Corollary 6.8 (termination-agreesSync , divergence-agreesSync ). In particular, a CESH
configuration terminates with a natural number n (diverges) if and only if a related
synchronous DCESH network terminates with a natural number n (diverges).
termination-agreesSync : ∀ cfg nodes n → RSync cfg nodes →
cfg ↓CESH nat n ↔ nodes ↓Sync nat n
divergence-agreesSync : ∀ cfg1 cfg2 → RSync cfg1 cfg2 →
cfg1 ↑CESH ↔ cfg2 ↑Sync
We also have that initial configurations are in RSync :
initial-relatedSync : ∀ c i → RSync (c, [ ], [ ], ∅)
(initial-networkSync c i)
These final results complete the picture for the DCESH machine. We have established that we get the same final result regardless of whether we choose to run a
fragment of code using the CES, the CESH, or the DCESH machine.
7
Related work
There is a multitude of programming languages and libraries for distributed computing. We focus mostly on those with a functional flavour. For surveys, see [15, 16].
Broadly speaking, we can divide them into those that use some form of explicit
message passing, and those that have more implicit mechanisms for distribution and
communication.
Explicit A prime example of a language for distributed computing that uses explicit message passing is Erlang [17]. Erlang is a very successful language used
prominently in the telecommunication industry. Conceptually similar solutions include MPI [1] and Cloud Haskell [18]. The theoretically advanced projects Nomadic
Pict [19] and the distributed join calculus [20] both support a notion of mobility for
distributed agents, which enables more expressivity for the distribution of a program
than the fairly static networks that our work uses. In general, explicit languages are
well-proven, but far away in the language design-space from the seamless distributed
computing that we envision because they place the burden of explicit communication
on the programmer.
32
Implicit Our work can be seen as a generalised Remote Procedure Call (RPC) [2].
In loc. cit. it is argued that emulating a shared address space is infeasible since it
requires each pointer to also contain location information, and that it is questionable
whether acceptable efficiency can be achieved. These arguments certainly apply to
our work, where we do just this. With the goal of expressivity in mind, however,
we believe that we should enable the programmer to write the potentially inefficient
programs that (internally) use remote pointers, because not all programs are performance critical. Furthermore, using a tagged pointer representation [21] for closure
pointers means that we can tag pointers that are remote, and pay a very low, if any,
performance penalty for local pointers.
Remote Evaluation (REV) [5] is another generalisation of RPC, siding with us on
enabling the use of higher-order functions across node boundaries. The main differences between REV and our work is that REV relies on sending code and that it has
a more general distribution mechanism.
The well-researched project Eden [22], which builds on Haskell, is a semi-implicit
language. Eden allows expressing distributed algorithms at a high level of abstraction, and is mostly implicit about communication, but explicit about process creation. Eden is specified operationally using a two-level semantics similar to ours.
Hop [23], Links [24], and ML5 [25] are examples of so called tierless languages that
allow writing (for instance) the client and server code of web applications in unified
languages with more or less seamless interoperability between them. We believe
that our work shows how a principled back-end and semantics can work for such
languages.
8
Conclusion and further work
We have seen the definition and correctness proofs of DCESH, a distributed abstract machine. Previously we have argued that distributed and heterogeneous programming would benefit from languages that are architecture-independent, using
compilation based on the idea of seamless computing [3]. This would allow the programmer to focus on solving algorithmic problems without having to worry about
the low-level details of the underlying computational system. Our previous work
shows how to achieve this, but is very different from conventional compilation techniques, relying on game semantics. This means that the vast literature on compiler
optimisation does not generally apply to it, and that it is difficult to interface with
legacy code. We believe that the current work alleviates these issues, since it shows
a way to do distributed execution as a conservative extension of existing abstract
machines. Additionally, DCESH adds very little overhead, if any, for local execution,
while permitting any sub-terms to be seamlessly distributed.
Implementation An implementation of the DCESH machine can be constructed
by either a bytecode interpreter or compiling the bytecode into a low-level language
33
by macro expansion. We have a prototype implementation that does the latter,
illustrating the potential for using DCESH as a basis for a usable compiler.
Outstanding questions
• Do the proofs generalise to a language with parallelism?
• Can we efficiently do distributed garbage collection [26]? This is necessary,
since DCESH, in contrast to our previous work, never reclaims heap garbage.
It would also be interesting to find out if parts of programs can use local
garbage collection for better performance.
• Can we find a way to express more complicated distribution patterns than
those made possible by locus specifiers? From our experience, locus specifiers
are excellent for simple programs (especially those with client-server disciplines), but due to the static nature of the specifiers, it is hard to express
dynamic distributed algorithms. We believe that our work can be extended
with dynamic locus specifiers to handle this. A simple first step would be to
add support for compiling parts of a program for more than one node at a
time, making it possible to pass (references to) functions already existing on
some remote node to it.
• Can we add support for sending code code (like REV [5]) when the code is
location-independent?
Two other language features that our abstract machines currently do not handle, but
that we would like to implement are abstract data types and mutable references.
Acknowledgements
The author would like to thank Martín Escardó for assistance with Agda, Fredrik
Nordvall Forsberg for rubber ducking, Dan Ghica for fruitful discussions and supervision, and Paul Blain Levy for simplifying some of the definitions.
This work was supported by Microsoft Research through its PhD Scholarship Programme.
References
[1] W. D. Gropp, E. L. Lusk, and A. Skjellum, Using MPI: portable parallel programming with the message-passing interface. MIT Press, 1999, vol. 1.
[2] A. Birrell and B. J. Nelson, “Implementing Remote Procedure Calls,” ACM Trans.
Comput. Syst., vol. 2, no. 1, pp. 39–59, 1984.
34
[3] O. Fredriksson and D. R. Ghica, “Abstract Machines for Game Semantics, Revisited,” in 28th Annual ACM/IEEE Symposium on Logic in Computer Science, LICS
2013, New Orleans, LA, USA, June 25-28, 2013. IEEE Computer Society, 2013,
pp. 560–569.
[4] ——, “Seamless Distributed Computing from the Geometry of Interaction,” in
Trustworthy Global Computing - 7th International Symposium, TGC 2012, Newcastle
upon Tyne, UK, September 7-8, 2012, Revised Selected Papers. Springer, 2012, pp.
34–48.
[5] J. W. Stamos and D. K. Gifford, “Remote evaluation,” ACM Trans. Program. Lang.
Syst., vol. 12, no. 4, pp. 537–565, 1990.
[6] P. J. Landin, “The mechanical evaluation of expressions,” Computer Journal,
vol. 6, no. 4, pp. 308–320, Jan. 1964.
[7] U. Norell, “Towards a practical programming language based on dependent type
theory,” Ph.D. dissertation, Chalmers Uni. of Tech., 2007.
[8] X. Leroy, “MPRI course 2-4-2, part II: abstract machines,” 2013-2014. [Online].
Available: http://gallium.inria.fr/~xleroy/mpri/progfunc/
[9] P. Henderson, Functional programming - application and implementation, ser.
Prentice Hall International Series in Computer Science. Prentice Hall, 1980.
[10] M. Felleisen and D. P. Friedman, “Control operators, the SECD-machine, and
the lambda-calculus,” in IFIP TC 2/WG 2.2, Aug. 1986.
[11] G. D. Plotkin, “LCF Considered as a Programming Language,” Theor. Comput.
Sci., vol. 5, no. 3, pp. 223–255, 1977.
[12] N. G. de Bruijn, “Lambda calculus notation with nameless dummies, a tool for
automatic formula manipulation, with application to the church-rosser theorem,” Indagationes Mathematicae, pp. 381–392, 1972.
[13] G. Berry and G. Boudol, “The Chemical Abstract Machine,” in Conference Record
of the Seventeenth Annual ACM Symposium on Principles of Programming Languages, San Francisco, California, USA, January 1990. ACM Press, 1990, pp.
81–94.
[14] T. Johnsson, “Lambda Lifting: Treansforming Programs to Recursive Equations,” in FPCA, 1985, pp. 190–203.
[15] P. W. Trinder, H.-W. Loidl, and R. F. Pointon, “Parallel and Distributed Haskells,”
J. Funct. Program., vol. 12, no. 4&5, pp. 469–510, 2002.
[16] H.-W. Loidl, F. Rubio, N. Scaife, K. Hammond, S. Horiguchi, U. Klusik,
R. Loogen, G. Michaelson, R. Pena, S. Priebe, Á. J. R. Portillo, and P. W. Trinder,
“Comparing Parallel Functional Languages: Programming and Performance,”
Higher-Order and Symbolic Computation, vol. 16, no. 3, pp. 203–251, 2003.
35
[17] J. Armstrong, R. Virding, and M. Williams, Concurrent programming in ERLANG.
Prentice Hall, 1993.
[18] J. Epstein, A. P. Black, and S. L. P. Jones, “Towards Haskell in the cloud,” in
Proceedings of the 4th ACM SIGPLAN Symposium on Haskell, Haskell 2011, Tokyo,
Japan, 22 September 2011. ACM, 2011, pp. 118–129.
[19] P. T. Wojciechowski and P. Sewell, “Nomadic pict: language and infrastructure
design for mobile agents,” IEEE Concurrency, vol. 8, no. 2, pp. 42–52, 2000.
[20] C. Fournet, G. Gonthier, J.-J. Lévy, L. Maranget, and D. Rémy, “A calculus of
mobile agents,” in CONCUR, ser. Lecture Notes in Computer Science, U. Montanari and V. Sassone, Eds., vol. 1119. Springer, 1996, pp. 406–421.
[21] S. Marlow, A. R. Yakushev, and S. L. P. Jones, “Faster laziness using dynamic
pointer tagging,” in Proceedings of the 12th ACM SIGPLAN International Conference on Functional Programming, ICFP 2007, Freiburg, Germany, October 1-3, 2007.
ACM, 2007, pp. 277–288.
[22] R. Loogen, Y. Ortega-Mallén, and R. Peña-Marí, “Parallel functional programming in Eden,” J. Funct. Program., vol. 15, no. 3, pp. 431–475, 2005.
[23] M. Serrano, E. Gallesio, and F. Loitsch, “Hop: a language for programming the
web 2.0,” in Companion to the 21th Annual ACM SIGPLAN Conference on ObjectOriented Programming, Systems, Languages, and Applications, OOPSLA 2006, October 22-26, 2006, Portland, Oregon, USA. ACM, 2006, pp. 975–985.
[24] E. Cooper, S. Lindley, P. Wadler, and J. Yallop, “Links: Web Programming Without Tiers,” in Formal Methods for Components and Objects, 5th International Symposium, FMCO 2006, Amsterdam, The Netherlands, November 7-10, 2006, Revised
Lectures. Springer, 2006, pp. 266–296.
[25] T. M. VII, K. Crary, and R. Harper, “Type-Safe Distributed Programming with
ML5,” in Trustworthy Global Computing, Third Symposium, TGC 2007, SophiaAntipolis, France, November 5-6, 2007, Revised Selected Papers. Springer, 2007,
pp. 108–123.
[26] D. Plainfossé and M. Shapiro, “A Survey of Distributed Garbage Collection
Techniques,” in Memory Management, International Workshop IWMM 95, Kinross,
UK, September 27-29, 1995, Proceedings. Springer, 1995, pp. 211–249.
36
| 6cs.PL
|
From One Point to A Manifold:
Knowledge Graph Embedding For Precise Link Prediction
arXiv:1512.04792v5 [cs.AI] 17 Jun 2017
Han Xiao∗ , Minlie Huang∗ , Xiaoyan Zhu
State Key Lab. of Intelligent Technology and Systems,
National Lab. for Information Science and Technology,
Dept. of Computer Science and Technology, Tsinghua University, Beijing 100084, PR China
[email protected]; {aihuang,zxy-dcs}@tsinghua.edu.cn
∗
Corresponding Authors: http://www.ibookman.net, http://www.aihuang.org
Abstract
Knowledge graph embedding aims at offering a
numerical knowledge representation paradigm by
transforming the entities and relations into continuous vector space. However, existing methods
could not characterize the knowledge graph in a
fine degree to make a precise link prediction. There
are two reasons for this issue: being an ill-posed
algebraic system and adopting an overstrict geometric form. As precise link prediction is critical for knowledge graph embedding, we propose a
manifold-based embedding principle (ManifoldE)
which could be treated as a well-posed algebraic
system that expands point-wise modeling in current models to manifold-wise modeling. Extensive experiments show that the proposed models achieve substantial improvements against the
state-of-the-art baselines, particularly for the precise prediction task, and yet maintain high efficiency. All of the related poster, slides, datasets
and codes have been published in http://www.
ibookman.net/conference.html.
1
Introduction
Knowledge is critical to artificial intelligence, and the embedded representation of knowledge offers an efficient basis of
computing over symbolic knowledge facts. More specifically,
knowledge graph embedding projects entities and relations
into a continuous high-dimension vector space by optimizing well-defined objective functions. A variety of methods
have been proposed for this task, including TransE [Bordes et
al., 2013], PTransE [Lin et al., 2015a] and KG2E [He et al.,
2015].
A fact of knowledge graph is usually represented by a triple
(h, r, t), where h, r, t indicate the head entity, the relation and
the tail entity, respectively. The goal of knowledge graph embedding is to obtain the vectorial representations of triples,
i.e., (h, r, t), with some well-defined objective functions. As
a key branch of embedding methods, translation-based methods, such as TransE [Bordes et al., 2013], PTransE [Lin et
al., 2015a] and KG2E [He et al., 2015], treat the triple as
a relation-specific translation from the head entity to the tail
entity, or formally as h + r = t.
Despite the success of previous methods, none of previous studies has addressed the issue of precise link prediction,
which finds the exact entity given another entity and the relation. For a specific query fact, most existing methods would
extract a few candidate entities that may contain correct answers, but there is no mechanism to ensure that the correct
answers rank ahead the candidate list.
Generally speaking, precise link prediction would improve the feasibility of knowledge completion, the effectiveness of knowledge reasoning, and the performance of many
knowledge-related tasks. Taking knowledge completion as
example, when we want to know about the birth place of Martin R.R., what we expect is the exact answer “U.S.”, while a
few other candidates do not make any sense.
The issue of precise link prediction is caused by two reasons: the ill-posed algebraic system and the over-strict geometric form.
First, from the algebraic perspective, each fact can be
treated as an equation of hr + r = tr 1 if following the
translation-based principle, embedding could be treated as a
solution to the equation group. In current embedding methods, the number of equations is more than that of free variables, which is called an ill-posed algebraic problem as defined in [Tikhonov and Arsenin, 1978]. More specifically,
hr + r = tr indicates d equations as hi + ri = ti where d is
the dimension of embedding vector and i denotes each dimension. Therefore, there are T ∗ d equations where T is the
number of facts, while the number of variables are (E+R)∗d,
where E, R are the number of entities and relations, respectively. As is the case that triples are much more than the sum
of entities and relations, the number of variables are much
less than the number of equations, which is typically an illposed algebraic system. Mathematically, an ill-posed algebraic system would commonly make the solutions imprecise
and unstable. In this paper, we propose to address this issue
by replacing the translation-based principle hr + r = tr by a
manifold-based principle M(h, r, t) = Dr2 where M is the
manifold function. With the manifold-based principle, our
model can make a nearly well-posed algebraic system by takT
so that the number of equations (T ) is no more
ing d ≥ E+R
than that of the free parameters ((E + R) ∗ d).
1
More generally speaking, hr and tr are embedding vectors projected w.r.t the relation space, and r is the relation embedding vector.
Figure 1: The visualization comparison of TransE and ManifoldE (Sphere). For ManifoldE, the manifold collapses to a solid
circle by dimension reduction. The data are selected from Wordnet and Freebase. The blue crosses mean the correctly matched
entities while the red crosses indicate the unmatched ones. The upper block corresponds to TransE where more near to the
center , more plausible the triple is. It is clear that the true and false triples near the golden position is chaotically distributed.
The below block is ManifoldE (Sphere) where the triples inside the solid circle are matched and those outside are unmatched.
We can see that there are relatively less errors in ManifoldE than TransE.
Second, from the geometric perspective, the position of the
golden facts in existing methods is almost one point, which is
too strict for all relations and is more insufficient for complex
relations such as many-to-many relations. For example, for
entity American Revolution, there exist many triples such as
(American Revolution, Has Part, Battle Bunker Hill), (American Revolution, Has Part, Battle Cowpens). When many tail
entities compete for only one point, there would be a major loss of objective function. Some previous work such as
TransH [Wang et al., 2014] and TransR [Lin et al., 2015b]
address this problem by projecting entities and relations into
some relation-specific subspaces. However, in each subspace,
the golden position is also one point and the over-strict geometric form is still existing. As can be seen from Fig.1,
the translation-based geometric principle involves too much
noise. However, ManifoldE alleviates this issue by expanding the position of golden triples from one point to a manifold
such as a high-dimensional sphere. By this mean, ManifoldE
avoids much noise to distinguish the true facts from the most
possible false ones, and improves the precision of knowledge
embedding as Fig.1 shows.
2
2.1
Translation-Based Methods
As a pioneering work of knowledge graph embedding,
TransE [Bordes et al., 2013] opens a line of translation-based
methods. TransE treats a triple (h, r, t) as a relation-specific
translation from a head entity to a tail entity, say h + r = t
and the score function has a form of ||h + r − t||22 . Following this principle, a number of models have been proposed.
For instance, TransH [Wang et al., 2014] adopts a projection
transformation, say hr = h − wr> hwr , tr = t − wr> twr ,
while TransR [Lin et al., 2015b] applies a rotation transformation, say hr = Mr h,tr = Mr t. Similar work also includes TransD [Ji et al., ] and TransM [Fan et al., 2014].
Other approaches take into consideration extra information
such as relation-type [Wang et al., 2015], paths with different
confidence levels (PTransE) [Lin et al., 2015a], and semantic smoothness of the embedding space [Guo et al., 2015].
KG2E [He et al., 2015] is a probabilistic embedding method
for modeling the uncertainty in knowledge base. Notably,
translation-based models demonstrate the state-of-the-art performance.
2.2
To summarize, our contributions are two-fold: (1)We have
addressed the issue of precise link prediction and uncover the
two reasons: the ill-posed algebraic system and the over-strict
geometric form. To our best knowledge, this is the first time
to address this issue formally. (2)We propose a manifoldbased principle to alleviate this issue and design a new model,
ManifoldE, which achieves remarkable improvements over
the state-of-the-art baselines in experiments, particularly for
precise link prediction. Besides, our methods are also very
efficient.
Related Work
Other Methods
The Unstructured Model (UM) [Bordes et al., 2012] is
a simplified version of TransE, which ignores relation information and the score function is reduced to fr (h, t) =
||h − t||22 . The Structured Embedding (SE) model [Bordes
et al., 2011] transforms the entity space with the head-specific
and tail-specific matrices and the score function is defined
as fr (h, t) = ||Mh,r h − Mt,r t||. The Semantic Matching Energy (SME) model [Bordes et al., 2012] [Bordes et
al., 2014] can enhance SE by considering the correlations between entities and relations with different matrix operators,
as follows:
fr (h, t) = (M1 h + M2 r + b1 )> (M3 t + M4 r + b2 )
>
fr (h, t) = (M1 h ⊗ M2 r + b1 ) (M3 t ⊗ M4 r + b2 )
where M1 , M2 , M3 and M4 are weight matrices, ⊗ is the
Hadamard product, b1 and b2 are bias vectors. The Single Layer Model (SLM) applies neural network to knowledge graph embedding and the score function is defined
as fr (h, t) = u>
r g(Mr,1 h + Mr,2 t) where Mr,1 , Mr,2
are relation-specific weight matrices. The Latent Factor Model (LFM) [Jenatton et al., 2012], [Sutskever et
al., 2009] makes use of the second-order correlations between entities by a quadratic form and the score function is as fr (h, t) = h> Wr t. The Neural Tensor Network (NTN) model [Socher et al., 2013] defines a very
expressive score function to combine the SLM and LFM:
>
fr (h, t) = u>
r g(h W··r t + Mr,1 h + Mr,2 t + br ), where
ur is a relation-specific linear layer, g(·) is the tanh function, W ∈ Rd×d×k is a 3-way tensor. Besides, RESCAL is a
collective matrix factorization model which is also a common
method in knowledge graph embedding [Nickel et al., 2011],
[Nickel et al., 2012].
3
Methods
In this section, we introduce the novel manifold-based principle and then we analyze these methods from the algebraic
and geometric perspectives.
3.1
ManifoldE : a Manifold-Based Embedding
Model
Instead of adopting the translation-based principle h + r = t,
we apply the manifold-based principle M(h, r, t) = Dr2 for
a specific triple (h, r, t). When a head entity and a relation
are given, the tail entities lay in a high-dimensional manifold.
Intuitively, our score function is designed by measuring the
distance of the triple away from the manifold:
fr (h, t) = ||M(h, r, t) −
Dr2 ||2
(1)
where Dr is a relation-specific manifold parameter. M : E ×
L×E →
− R is the manifold function, where E, L are the entity
set and relation set and R is the real number field.
Sphere. Sphere is a very typical manifold. In this setting,
all the tail (or head) entities for a specific fact such as (h, r, ∗)
are supposed to lay in a high-dimensional sphere where h + r
is the center and Dr is the radius, formally stated as below:
M(h, r, t) = ||h + r − t||22
Obviously, this is a straight-forward extension of
translation-based models in which Dr is zero. From the geometric perspective, the manifold collapses into a point when
applying the translation-based principle.
Reproducing Kernel Hilbert Space (RKHS) usually provides a more expressive approach to represent the manifolds,
which motivates us to apply the manifold-based principle
with kernels. To this point, kernels are involved to lay the
sphere in a Hilbert space (an implicit high-dimensional space)
Data
Table 1: Statistics of datasets
WN18
FB15K
WN11
#Rel
#Ent
#Train
#Valid
#Test
18
40,943
141,442
5,000
5,000
1,345
14,951
483,142
50,000
59,071
11
38,696
112,581
2,609
10,544
FB13
13
75,043
316,232
5,908
23,733
Figure 2: Visualization of embedding for Manifold-based
models. (a) corresponds to the Sphere setting where all the
tail entities are supposed to lay in the sphere. As Clock Dial
is matched by the two facts, it should lay in both spheres.
(b) corresponds to the Hyperplane setting where Clock Dial
should lay and does lay in both hyperplanes, making embedding more precise.
as below:
M(h, r, t)
= ||ϕ(h) + ϕ(r) − ϕ(t)||2
= K(h, h) + K(t, t) + K(r, r)
(2)
−2K(h, t) − 2K(r, t) + 2K(r, h)
where ϕ is the mapping from the original space to the Hilbert
space, and K is the induced kernel by ϕ. Commonly, K
could be Linear kernel (K(a, b) = a> b), Gaussian ker||a−b||2
2
nel (K(a, b) = e− σ2 ), Polynomial Kernel (K(a, b) =
(a> b + d)p , and so on. Obviously, if applying the linear
kernel, the above function is reduced to the original sphere
manifold.
Hyperplane. As shown in Fig.2, we could see that when
two manifolds are not intersected, there may be a loss in embedding. Two spheres would intersect only under some strict
conditions, while two hyperplanes would intersect if their
normal vectors are not in parallel. Motivated by this fact, we
apply a hyperplane to enhance our model as below:
M(h, r, t) = (h + rhead )> (t + rtail )
where rhead and rtail are two specific relation embeddings.
From the geometric perspective, given the head entity and the
relation, the tail entities lay in the hyperplane whose direction is h + rhead and the bias corresponds to Dr2 . In practical
cases, since the two vectors e1 + r1,head and e2 + r2,head
are not likely to be parallel, there would be more chance
Datasets
Metric
SE[Bordes et al., 2011]
TransE [Bordes et al., 2013]
TransH [Wang et al., 2014]
TransR [Lin et al., 2015b]
KG2E [He et al., 2015]
ManifoldE Sphere
ManifoldE Hyperplane
Table 2: Evaluation results on link prediction
WN18
HITS@10(%) HITS@1(%)
Time(s)
HITS@10(%)
Raw Filter
Filter
One Epos Raw Filter
68.5
80.5
28.8
39.8
75.4
89.2
29.5
0.4
34.9
47.1
73.0
82.3
31.3
1.4
48.2
64.4
79.8
92.0
33.5
9.8
48.4
68.7
80.2
92.8
54.1
10.7
48.9
74.0
81.1
94.4
26.4
0.4
52.0
79.5
81.4
93.7
90.8
0.5
52.6
78.2
to lead two intersected hyperplanes than two intersected
spheres. Therefore, there would be more solutions provided
by the intersection of hyperplanes.
Motivated by enlarging the number of precisely predicted
tail entities for the same head and relation, we apply the absolute operators as M(h, r, t) = |h + rhead |> |t + rtail | where
.
|w| = (|w1 |, |w2 |, |w3 |, ..., |wn |). For an instance of onedimensional case that |h+rhead ||t+rtail | = Dr2 , the absolute
operator would double the solution number of t, meaning that
two tail entities rather than one could be matched precisely to
this head for this relation. For this reason, the absolute operator would promote the flexibility of embedding.
We also apply the kernel trick to the Hyperplane setting, as
below:
M(h, r, t) = K(h + rhead , t + rtail )
3.2
Algebraic Perspective
The ill-posed equation system that posses more equations
than free variables always leads to some undesired properties such as instability, which may be the reason why
the translation-based principle performs not so well in precise link prediction. To alleviate this issue, manifold-based
methods model embedding within a nearly well-posed algebraic framework, since our principle indicates only one
equation for one fact triple. Taking an example of sphere
Pd
as i=1 (hi + ri − ti )2 = Dr2 , we could conclude that if
T
d ≥ #Equation
= E+R
, our embedding system would be
E+R
more algebraically stable and this condition is easy to achieve
by just enlarging the embedding dimension to a suitable degree. In theory, larger embedding dimension provides more
solutions to embedding equations, which makes embedding
more flexible. When the suitable condition is satisfied, the
stable algebraic solution would lead the embedding to a fine
characterization, therefore the precise link prediction would
be promoted.
3.3
Geometric Perspective
The translation-based principle allocates just one position for
a golden triple. We extend one point to a whole manifold such
as a high dimensional sphere. For instance, all tail entities for
a 1-N relation could lay on a sphere, which applies h + r as
the center and Dr as the radius. Obversely, it would be more
suitable in a manifold setting than in a point setting.
3.4
FB15K
HITS@1(%)
Filter
24.4
24.8
20.0
40.4
50.2
54.0
Time(s)
One Epos
0.7
4.8
29.1
44.2
0.7
0.8
Training
We train our model with the rank-based hinge loss, which
means to maximize the discriminative margin between the
golden triples and the false ones.
X
X
L=
[fr0 (h0 , t0 ) − fr (h, t) + γ]+ (3)
(h,r,t)∈∆ (h0 ,r 0 ,t0 )∈∆0
where L is the loss function which should be minimized, γ is
.
the margin, and [·]+ = max(0, ·) is the hinge loss. The false
triples are sampled with the Bernoulli Sampling Method as
introduced in [Wang et al., 2014]. We initialize the embedding vectors by the similar methods used in deep neural network [Glorot and Bengio, 2010]. Stochastic gradient descent
(SGD) is applied to solve this problem.
In theory, our computation complexity relative to TransE
is bounded by a very small constant, as O(λ × O(TransE))
where λ ≥ 1. This small constant λ is caused by manifoldbased operations and kernelization. Commonly, TransE is
the most efficient among all the translation-based methods,
while ManifoldE could be comparable to TransE in efficiency, hence faster than other translation-based methods.
4
Experiments
Our experiments are conducted on four public benchmark
datasets that are the subsets of Wordnet [Miller, 1995] and
Freebase [Bollacker et al., 2008]. The statistics of these
datasets are listed in Tab.1. Experiments are conducted on
two tasks : Link Prediction and Triple Classification. To
further demonstrate how the proposed model performs the
manifold-based principle, we present the visualization comparison between translation-based and manifold-based models in the section 4.3. Finally, we conduct error analysis to
further understand the benefit and limits of our models.
4.1
Link Prediction
Reasoning is the focus of knowledge computation. To verify
the reasoning performance of embedding, link prediction task
is conducted. This task aims to predict the missing entities.
An alternative of the entities and the relation are given when
the embedding methods infer the other missing entity. More
specifically, in this task, we predict t given (h, r, ∗), or predict h given (∗, r, t) . The WN18 and FB15K are two benchmark datasets for this task. Notably, many AI tasks could
Table 3: Evaluation results on FB15K by mapping properties of relations(%)
Tasks
Predicting Head(HITS@10) Predicting Tail(HITS@10)
Relation Category
1-1
1-N N-1
N-N
1-1
1-N N-1
N-N
TransE [Bordes et al., 2013] 43.7 65.7 18.2
47.2
43.7 19.7 66.7
50.0
TransH [Wang et al., 2014] 66.8 87.6 28.7
64.5
65.5 39.8 83.3
67.2
TransR [Lin et al., 2015b]
78.8 89.2 34.1
69.2
79.2 37.4 90.4
72.1
ManifoldE Sphere
80.9 94.5 36.7
82.1
80.0 50.7 93.6
82.5
ManifoldE Hyperplane
79.0 94.9 38.4
80.2
76.7 49.5 94.2
83.0
Table 4: Evaluation results for precise prediction on FB15K by mapping properties of relations(%)
Tasks
Predicting Head(HITS@1) Predicting Tail(HITS@1)
Relation Category
1-1
1-N N-1
N-N
1-1
1-N N-1 N-N
TransE [Bordes et al., 2013] 35.4 50.7 8.6
18.1 34.5 10.6 56.1 20.3
TransH [Wang et al., 2014]
35.3 48.7 8.4
16.9 35.5 10.4 57.5 19.3
TransR [Lin et al., 2015b]
29.5 42.8 6.1
14.5 28.0 7.7 44.1 16.2
ManifoldE Sphere
33.3 60.0 13.3 53.8 34.5 19.5 58.5 56.2
ManifoldE Hyperplane
62.6 78.6 21.5 53.0 58.2 32.1 77.8 55.9
be enhanced by “Link Prediction”, such as relation extraction
[Hoffmann et al., 2011].
Evaluation Protocol. We adopt the same protocol used
in previous studies. Firstly, for each testing triple (h, r, t),
we corrupt it by replacing the tail t (or the head h) with every entity e in the knowledge graph. Secondly, we calculate a
probabilistic score of this corrupted triple with the score function fr (h, t). By ranking these scores in descending order, we
then obtain the rank of the original triple. The evaluation metric is the proportion of testing triple whose rank is not larger
than N (HITS@N). HITS@10 is applied for common reasoning ability and HITS@1 concerns the precise embedding
performance. This is called “Raw” setting. When we filter
out the corrupted triples that exist in the training, validation,
or test datasets, this is the“Filter” setting. If a corrupted triple
exists in the knowledge graph, ranking it ahead the original
triple is also acceptable. To eliminate this case, the “Filter”
setting is more preferred. In both settings, a higher HITS@N
means better performance. Note that we do not report the results of “raw” setting for HITS@1, because they are too small
to make a sense. Notably, we actually run each baseline in the
same setting for five times, and average the running time as
the results.
Implementation. As the datasets are the same, we directly reproduce the experimental results of several baselines
from the literature for HITS@10. As to HITS@1, we request the results from the authors of PTransE and KG2E. We
acknowledge these authors Yankai Lin and Shizhu He. We
have attempted several settings on the validation dataset to
get the best configuration. Under the “bern.” sampling strategy, the optimal configurations of ManifoldE are as follows.
For sphere, α = 0.001, k = 100, γ = 3.0, Linear kernel,
on WN18; α = 0.0005, k = 800, γ = 1.0, Polynomial
kernel(p = 2, d = 2) on FB15K. For hyperplane, learning rate α = 0.01, embedding dimension k = 800, margin
γ = 0.2, Linear kernel, on WN18; α = 0.01, k = 1000,
γ = 0.2, Linear kernel, on FB15K. The experimental environment is a common PC with i7-4790 CPU, 16G Memory
and Windows 10. Note that all the symbols are introduced
in “Methods”. Notably, we train the model until convergence
about 10,000 rounds in previous version. But in this version,
we adopt no trick and train the model until 2,000 rounds.
Results. Evaluation results on WN18 and FB15K are reported in Tab.2 and Tab.3. We observe that:
1. ManifoldE beats all the baselines in all the sub-tasks,
yielding the effectiveness and efficiency of the manifoldbased principle.
2. From the algebraic perspective, it’s reasonable to measure the algebraic ill-posed degree with the radio
T /(E + R) because with the translation-based principle, T d is the number of equations and (E + R)d is
the number of free variables, a larger radio means more
ill-posed. Since the manifold-based principle alleviates
this issue, ManifoldE(Sphere) would make a more promotion relatively to the comparable baselines (TransE)
under a larger radio. As to the metric HITS@1, on
WN18, the radio is 3.5 while TransE achieves 29.5%
and ManifoldE(Sphere) achieves 55%, leading to a relative improvement of 85.1%. On FB15K the radio is 30.2
while TransE achieves 24.4% and ManifoldE(Sphere)
achieves 64.1%, leading to a relative improvement of
162.7%. This comparison illustrates manifold-based
methods could stabilize the algebraic property of embedding system, by which means, the precise embedding
could be approached much better.
3. From the geometric perspective, traditional models attempt to express all the matched entities into one position, which leads to unsatisfactory performance on complex relations. Meanwhile, manifold-based model could
perform much better for these complex relations as we
discussed. As to the metric HITS@1, the simple relation
1-1 improves relatively by 87.8% by ManifoldE(Sphere)
Table 5: Triple classification: accuracy(%) for different embedding methods.
Methods
WN11 FB13 AVG.
SE
53.0
75.2
64.1
NTN
70.4
87.1
78.8
TransE
75.9
81.5
78.7
TransH
78.8
83.3
81.1
TransR
85.9
82.5
84.2
KG2E
85.4
85.3
85.4
ManifoldE Sphere
87.5
87.2
87.4
ManifoldE Hyperplane
86.9
87.3
87.1
than TransE while the complex relations such as 1-N,
N-1, N-N improve relatively by 266.5%, 36.2% and
215.7% respectively. This comparison demonstrates
manifold-based method that extends the golden position
from one point to a manifold could better characterize
the true facts, especially for complex relations.
4.2
4.3
As the Fig.1 shows, the translation-based principle involves
too much noise near the center where is supposed to lay the
true facts. We attribute such issues to the precise link prediction issue as introduced previously, However, manifold-based
principle alleviates this issue to enhance precise knowledge
embedding, which could be seen from the visualization results.
4.4
1. Overall, ManifoldE yields the best performance. This
illustrates our method could improve the embedding.
2. More specifically, on WN11, the relation “Type Of” that
is a complex one, improves from 71.4% of TransE to
86.3% of ManifoldE(Sphere) while on FB13, the relation “Gender” that is an extreme N-1 relation, improves from 95.1% to 99.5%. This comparison shows
manifold-based methods could handle complex relations
better.
Error Analysis
To analyze the errors in Link Prediction, we randomly sample 100 testing triples that could not rank at the top positions
by ManifoldE (Hyperplane) and three categories of errors are
summarized. Notably, we call the predicted top rank triple,
which is not the golden one, as “top rank triple”.
1. True Facts (29%): The top rank triple is correct though it is not contained in the knowledge
graph, thus ranking it before the golden one is acceptable. This category is caused by incompleteness of the dataset. For example, reflexive semantics as (Hot, Similar To, Hot), general expression as
(Animal Scientist, Type Of, Individual), professional
knowledge as (Crescentia, Type Of, Plant) and so on.
Triple Classification
In order to present the discriminative capability of our method
between true and false facts, triple classification is conducted.
This is a classical task in knowledge base embedding, which
aims at predicting whether a given triple (h, r, t) is correct or
not. WN11 and FB13 are the benchmark datasets for this task.
Note that evaluation of classification needs negative samples,
and the datasets have already been built with negative triples.
Evaluation Protocol. The decision process is very simple
as follows: for a triple (h, r, t), if fr (h, t) is below a threshold
σr , then positive; otherwise negative. The thresholds {σr }
are determined on the validation dataset. This task is somehow a triple binary classification.
Implementation. As all methods use the same datasets,
we directly re-use the results of different methods from the
literature. We have attempted several settings on the validation dataset to find the best configuration. The optimal configurations of ManifoldE are as follows with “bern” sampling.
For sphere, α = 0.001, k = 100, γ = 10.0, Linear kernel on
WN18; α = 0.00005, k = 1000, γ = 0.3, Gaussian kernel
(σ = 1.0) on FB13. For hyperplane, learning rate α = 0.01,
embedding dimension k = 500, margin γ = 1.0, Linear kernel, on WN18; α = 0.001, k = 1000, γ = 3.0, Polynomial
kernel (p = 2, d = 2), on FB13.
Results. Accuracies are reported in Tab.5. We observe
that:
Visualization Comparison between
Translation-Based and Manifold-Based
Principle
2. Related Concepts (63%): The top rank triple is
a related concept, but the corresponding fact is
not exactly correct.
This category is caused by
the relatively simple manifolds applied by ManifoldE. For example, puzzled place membership as
(Afghanistan, Has Part, Hyderabad), similar mentions as (New Haven, Part Of, One Star State), similar concepts as (Sea, Has Instance, Aleutian Islands),
possible knowledge as (Accent, Type Of, Individual)
and so on. We could further exploit complex manifolds
to enhance the discriminative ability.
3. Others (8%): There are always some top rank triples
that are difficult to interpret.
5
Conclusions
In this paper, we study the precise link prediction problem and reveal two reasons to the problem: the ill-posed
algebraic system and the over-restricted geometric form.
To alleviate these issues, we propose a novel manifoldbased principle and the corresponding ManifoldE models (Sphere/Hyperplane) inspired by the principle. From
the algebraic perspective, ManifoldE is a nearly well-posed
equation system and from the geometric perspective, it expands point-wise modeling in the translation-based principle
to manifold-wise modeling. Extensive experiments show our
method achieves substantial improvements against the stateof-the-art baselines.
This work was partly supported by the National
Basic Research Program (973 Program) under grant
No.2012CB316301 / 2013CB329403, and the National Science Foundation of China under grant No.61272227 /
61332007.
References
[Bollacker et al., 2008] Kurt Bollacker, Colin Evans,
Praveen Paritosh, Tim Sturge, and Jamie Taylor. Freebase:
a collaboratively created graph database for structuring
human knowledge. In Proceedings of the 2008 ACM
SIGMOD international conference on Management of
data, pages 1247–1250. ACM, 2008.
[Bordes et al., 2011] Antoine Bordes, Jason Weston, Ronan
Collobert, Yoshua Bengio, et al. Learning structured
embeddings of knowledge bases. In Proceedings of the
Twenty-fifth AAAI Conference on Artificial Intelligence,
2011.
[Bordes et al., 2012] Antoine Bordes, Xavier Glorot, Jason
Weston, and Yoshua Bengio. Joint learning of words and
meaning representations for open-text semantic parsing.
In International Conference on Artificial Intelligence and
Statistics, pages 127–135, 2012.
[Bordes et al., 2013] Antoine Bordes, Nicolas Usunier,
Alberto Garcia-Duran, Jason Weston, and Oksana
Yakhnenko.
Translating embeddings for modeling
multi-relational data. In Advances in Neural Information
Processing Systems, pages 2787–2795, 2013.
[Bordes et al., 2014] Antoine Bordes, Xavier Glorot, Jason
Weston, and Yoshua Bengio. A semantic matching energy
function for learning with multi-relational data. Machine
Learning, 94(2):233–259, 2014.
[Fan et al., 2014] Miao Fan, Qiang Zhou, Emily Chang, and
Thomas Fang Zheng. Transition-based knowledge graph
embedding with relational mapping properties. In Proceedings of the 28th Pacific Asia Conference on Language,
Information, and Computation, pages 328–337, 2014.
[Glorot and Bengio, 2010] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In International conference on
artificial intelligence and statistics, pages 249–256, 2010.
[Guo et al., 2015] Shu Guo, Quan Wang, Bin Wang, Lihong
Wang, and Li Guo. Semantically smooth knowledge graph
embedding. In Proceedings of ACL, 2015.
[He et al., 2015] Shizhu He, Kang Liu, Guoliang Ji, and Jun
Zhao. Learning to represent knowledge graphs with gaussian embedding. In Proceedings of the 24th ACM International on Conference on Information and Knowledge Management, pages 623–632. ACM, 2015.
[Hoffmann et al., 2011] Raphael Hoffmann, Congle Zhang,
Xiao Ling, Luke Zettlemoyer, and Daniel S Weld.
Knowledge-based weak supervision for information extraction of overlapping relations. In Proceedings of the
49th Annual Meeting of the Association for Computational
Linguistics: Human Language Technologies-Volume 1,
pages 541–550. Association for Computational Linguistics, 2011.
[Jenatton et al., 2012] Rodolphe Jenatton, Nicolas L Roux,
Antoine Bordes, and Guillaume R Obozinski. A latent
factor model for highly multi-relational data. In Advances
in Neural Information Processing Systems, pages 3167–
3175, 2012.
[Ji et al., ] Guoliang Ji, Shizhu He, Liheng Xu, Kang Liu,
and Jun Zhao. Knowledge graph embedding via dynamic
mapping matrix.
[Lin et al., 2015a] Yankai Lin, Zhiyuan Liu, and Maosong
Sun. Modeling relation paths for representation learning
of knowledge bases. Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics,
2015.
[Lin et al., 2015b] Yankai Lin, Zhiyuan Liu, Maosong Sun,
Yang Liu, and Xuan Zhu. Learning entity and relation embeddings for knowledge graph completion. In Proceedings
of the Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015.
[Miller, 1995] George A Miller. Wordnet: a lexical database
for english. Communications of the ACM, 38(11):39–41,
1995.
[Nickel et al., 2011] Maximilian Nickel, Volker Tresp, and
Hans-Peter Kriegel. A three-way model for collective
learning on multi-relational data. In Proceedings of
the 28th international conference on machine learning
(ICML-11), pages 809–816, 2011.
[Nickel et al., 2012] Maximilian Nickel, Volker Tresp, and
Hans-Peter Kriegel. Factorizing yago: scalable machine
learning for linked data. In Proceedings of the 21st international conference on World Wide Web, pages 271–280.
ACM, 2012.
[Socher et al., 2013] Richard Socher, Danqi Chen, Christopher D Manning, and Andrew Ng. Reasoning with neural
tensor networks for knowledge base completion. In Advances in Neural Information Processing Systems, pages
926–934, 2013.
[Sutskever et al., 2009] Ilya Sutskever, Joshua B Tenenbaum, and Ruslan Salakhutdinov. Modelling relational
data using bayesian clustered tensor factorization. In Advances in neural information processing systems, pages
1821–1828, 2009.
[Tikhonov and Arsenin, 1978] A. N. Tikhonov and V. Y. Arsenin. Solutions of ill-posed problems. Mathematics of
Computation, 32(144):491–491, 1978.
[Wang et al., 2014] Zhen Wang, Jianwen Zhang, Jianlin
Feng, and Zheng Chen. Knowledge graph embedding by
translating on hyperplanes. In Proceedings of the TwentyEighth AAAI Conference on Artificial Intelligence, pages
1112–1119, 2014.
[Wang et al., 2015] Quan Wang, Bin Wang, and Li Guo.
Knowledge base completion using embeddings and rules.
In Proceedings of the 24th International Joint Conference
on Artificial Intelligence, 2015.
| 2cs.AI
|
Optimal Sensing and Data Estimation in a Large
Sensor Network
arXiv:1707.08074v2 [cs.IT] 11 Sep 2017
Arpan Chattopadhyay, Urbashi Mitra
Abstract—An energy efficient use of large scale sensor networks necessitates activating a subset of possible sensors for
estimation at a fusion center. The problem is inherently combinatorial; to this end, a set of iterative, randomized algorithms
are developed for sensor subset selection by exploiting the
underlying statistics. Gibbs sampling-based methods are designed
to optimize the estimation error and the mean number of
activated sensors. The optimality of the proposed strategy is
proven, along with guarantees on their convergence speeds. Also,
another new algorithm exploiting stochastic approximation in
conjunction with Gibbs sampling is derived for a constrained
version of the sensor selection problem. The methodology is
extended to the scenario where the fusion center has access
to only a parametric form of the joint statistics, but not the
true underlying distribution. Therein, expectation-maximization
is effectively employed to learn the distribution. Strategies for
iid time-varying data are also outlined. Numerical results show
that the proposed methods converge very fast to the respective
optimal solutions, and therefore can be employed for optimal
sensor subset selection in practical sensor networks.
Index Terms—Wireless sensor networks, active sensing, data
estimation, Gibbs sampling, stochastic approximation, expectation maximization.
I. I NTRODUCTION
A wireless sensor network typically consists of a number of
sensor nodes deployed to monitor some physical process. The
sensor data is often delivered to a fusion center via wireless
links. The fusion center, based on the gathered data from the
sensors, infers the state of the physical process and makes
control decisions if necessary.
Sensor networks have widespread applications in various
domains such as environmental monitoring, industrial process monitoring and control, localization, tracking of mobile
objects, system parameter estimation, and even in disaster
management. However, severe resource constraints in such
networks necessitates careful design and control strategies in
order to attain a reasonable compromise between resource
usage and network performance. One major restriction is that
the nodes are battery constrained, which limits the lifetime of
the network. Also, low capacity of wireless channels due to
transmit power constraint, heavy interference and unreliable
link behaviour restricts the amount of data that can be sent to
the fusion center per unit time. In some special cases, such
Arpan Chattopadhyay and Urbashi Mitra are with Ming Hsieh Department
of Electrical Engineering, University of Southern California, Los Angeles,
USA. Email: {achattop,ubli}@usc.edu
This work was funded by the following grants: ONR N00014-15-1-2550,
NSF CNS-1213128, NSF CCF-1410009, AFOSR FA9550-12-1-0215, and
NSF CPS-1446901, and also by the UK Leverhulme Trust, the UK Royal
Society of Engineers and the Fulbright Foundation.
as mobile crowdsensing applications (see [1]), a certain cost
might be necessary in order to engage a sensor owned by a
third party. All these constraints lead us to the fundamental
question: how to select a small subset of sensors so that the
observations made by these sensors are most informative for
the effiicient inference of the state of the physical process
under measurement?
Recent results have focused on optimal sequential sensor
subset selection in order to monitor a random process modeled as Markov chain or linear dynamical system; see e.g.
[2]–[7]. Sensor subset selection using these control-theoretic
resullts are typically computationally expensive, and the lowcomplexity approximation schemes proposed in some of these
papers (such as [3] and [7]) are not optimal. On the other
hand, there appears to be limited work on optimal subset
selection when sensor data is static and its distribution is
known either absolutely or in parametric form; the major
challenge in this problem is computational ( [8]), where the
computational burden arises for two reasons: (i) finding the
optimal subset requires a search operation over all possible subsets of sensors, thereby requiring exponentially many
number of computations, and (ii) for each subset of active
sensors, computing the estimation error conditioned on the
observation made by active sensors requires exponentially
many computations. In [8], the problem of minimizing the
minimum mean squared error (MMSE) of a vector signal using
samples collected from a given number of sensors chosen by
the network operator is considered; a tractable lower bound
to the MMSE is employed in a certain greedy algorithm
to obviate the complexity in MMSE computation and the
combinatorial problem of searching over all possible subsets
of sensors. In contrast, our paper deals with a general error
metric (which could potentially be the MMSE or even the
lower bound to MMSE as in [8]), and proposes Gibbs sampling
based techniques for the optimal subset selection problem, in
order to minimize a linear combination of the estimation error
and the expected number of activated sensors. To the best of
our knowledge, ours is the first paper to use Gibbs sampling
for optimal sensor subset selection with low complexity in the
context of active sensing. We also provide an algorithm based
on Gibbs sampling and stochastic approximation, which is
provably optimal and which minimizes the expected estimation
error subject to a constraint on the expected number of
activated sensors; this technique can be employed to solve
many other constrained combinatorial optimization problems.1
A. Organization and our contribution
The paper is organized as follows. The system model is
described in Section II. In Section III, we propose Gibbs
sampling based algorithms to minimize a linear combination
of data estimation error and the number of active sensors.
We prove convergence of these algorithms, and also provide a
bound on the convergence speed of one algorithm. Section IV
provides algorithm for minimizing the estimation error subject
to a constraint on the mean number of active sensors. We
propose a novel algorithm based on Gibbs sampling and
stochastic approximation, and prove its convergence to the
desired solution. To the best of our knowledge, this is a novel
technique that can be used for other constrained combinatorial
optimization problems as well. We also discuss how the
Gibbs sampling algorithm can be used when we have a hard
constraint on the number of activated sensors. Section V
discusses expectation maximization (EM) based algorithms
when data comes from a parameterized distribution with
unknown parameters. Numerical results on computational gain
and performance improvement by using some of the proposed
algorithms are presented in Section VI. Finally, we conclude
in Section VII. All proofs are provided in the appendices.
We have also discussed in various sections how the proposed
algorithms with minor modifications can be used for data
varying in time in an i.i.d. fashion.
II. S YSTEM M ODEL AND N OTATION
A. The network and data model
We consider a large connected single or multi-hop sensor
network, whose sensor nodes are denoted by the set N =
{1, 2, · · · , N }. Each node k ∈ N is associated with a (possibly
vector-valued) data X k , and we denote by X = {X k }k∈N
the set of data which has to be reconstructed. A fusion center
determines the set of activated sensors, and estimates the data
in each node given the limited observations only from the
activated sensors.
While our methods assume static data from the sensors;
these methods can be employed with good performance for
data that varies in an iid fashion with respect to time.
B. Reconstruction of sensor data
We denote the activation state of a sensor by 1 if it is
active, and by 0 otherwise. We call B := {0, 1}N the set of
all possible configurations in the network, and denote a generic
configuration by B. Specifying a configuration is equivalent
to selecting a subset S of active sensors. We denote by B−j ∈
{0, 1}N −1 the configuration B with its j-th entry removed.
1 In this connection, we would like to mention that Gibbs sampling based
algorithms were used in wireless caching [9], but to solve an unconstrained
problem. In the current paper, we combine Gibbs sampling and stochastic
approximation to solve a constrained optimization problem; this technique
is general and can iteratively solve many other constrained combinatorial
optimization problems optimally with very small computation per iteration,
while the approximation algorithms are not guaranteed to achieve optimality.
The estimate of X at the fusion center is denoted by X̂. The
corresponding expected error
PNunder configuration B ∈ B is
denoted by EdB (X, X̂) = k=1 EdB (X k , Xˆk ). Specifically,
the mean squared
EdB (X, X̂) = E(||X −
PN error (MSE)ˆ yields
2
X̂||2 ) =
E(||X
−
X
||
).
Let
us denote the cost
k
k
k=1
of activating a sensor by λ. Heterogeneous sensor classes
with different priorities or weights can be straightforwardly
accommodated and thus are not presented herein.
1) The unconstrained problem: Given a configuration B ∈
B, the associated network cost is given by:
h(B) := EdB (X, X̂) + λ||B||1
(1)
In the context of stastistical physics, one can view h(B) as the
potential under configuration B. Our goal herein is to solve
the following optimization problem:
min h(B)
B∈B
(UP)
2) The constrained problem: Problem (UP) is a relaxed
version of the constrained problem below:
min EdB (X, X̂) s.t. E||B||1 ≤ N̄
B∈B
(CP)
Here the expectation in the constraint is over any possible
randomization in choosing the configuration B. The cost of
activating a sensor, λ, can be viewed as a Lagrange multiplier
used to relax this constrained problem.
Theorem 1 relates solution of (UP) to (CP).
Theorem 1: Consider problems (CP) and (UP). If there
exists a Lagrange multiplier λ∗ ≥ 0 and a B ∗ ∈ B, such
that an optimal configuration for (UP) under λ = λ∗ is B ∗ ,
and the constraint in (CP) is satisfied with equality under the
pair (B ∗ , λ∗ ), then B ∗ is an optimal configuration for (CP).
∗
,
In case there exist multiple configurations B1∗ , B2∗ , · · · , Bm
∗
a multiplier λ ≥ 0, and a probability mass function
∗
is opti(p1 , p2 , · · · , pm ) such that (i) each of B1∗P
, B2∗ , · · · , Bm
m
∗
∗
mal for problem (UP) under λ , and (ii) i=1 pi ||Bi ||1 = N̄ ,
then an optimal solution for (CP) is to choose one configu∗
ration from B1∗ , B2∗ , · · · , Bm
with probability mass function
(p1 , p2 , · · · , pm ).
Proof: See Appendix A.
Remark 1: Theorem 1 allows us to obtain a solution for
(CP) from the solution of (UP) by choosing an appropriate
λ∗ ; this will be elaborated upon in Section IV.
III. G IBBS SAMPLING APPROACH TO SOLVE THE
UNCONSTRAINED PROBLEM
In this section, we will provide algorithms based on Gibbs
sampling to compute the optimal solution for (UP).
A. Basic Gibbs sampling
Let us denote the distribution πβ (·) over B as follows:
πβ (B) := P
e−βh(B)
e−βh(B)
:=
−βh(B)
Zβ
B∈B e
Choose any initial B(0) ∈ {0, 1}N . At each discrete time
instant t = 0, 1, 2, · · · , pick a random sensor jt ∈ N
independently and uniformly. For sensor jt , choose
Bjt (t) = 1 with probability
−βh(B−j (t−1),1)
p := −βh(B−jte(t−1),1) t −βh(B−jt (t−1),0) and choose
e
+e
Bjt (t) = 0 with probability (1 − p). Choose
Bk (t) = Bk (t − 1) for all k 6= jt .
Algorithm 1: BASICGIBBS algorithm
. Motivated by the theory of statistical physics, we call the
parameter β the inverse temperature,
and Zβ the partition
P
function. Clearly, limβ↑∞ B∈arg minA∈B h(A) πβ (B) = 1.
Hence, if we can choose a configuration B with probability
πβ (B) for a large β > 0, we can approximately solve (UP).
Computing Zβ will require 2N addition operations, and
hence it is computationally prohibitive for large N . As an
alternative, we provide an iterative algorithm based on Gibbs
sampling, which requires many fewer computations in each
iteration. Gibbs sampling runs a discrete-time Markov chain
{B(t)}t≥0 whose stationary distribution is πβ (·).
The BASICGIBBS algorithm (Algorithm 1) simulates the
Markov chain {B(t)}t≥0 for any β > 0. The fusion center
runs the algorithm to determine the activation set; as such, the
fusion center must create a virtual network graph.
Theorem 2: The Markov chain {B(t)}t≥0 has a stationary
distribution πβ (·) under the BASICGIBBS algorithm.
Proof: Follows from the theory in [10, Chapter 7]).
Remark 2: Theorem 2 tells us that if the fusion center runs
BASICGIBBS algorithm and reaches the steady state distribution of the Markov chain {B(t)}t≥0 , then the configuration
chosen by the algorithm will have distribution πβ (·). For
very large β > 0, if one runs {B(t)}t≥0 for a sufficiently
long, finite time T , then the terminal state BT will belong to
arg minB∈B h(B) with high probability.
B. The exact solution
BASICGIBBS is operated with a fixed β; but, in practice,
the optimal soultion of the unconstrained problem (UP) is
obtained with β ↑ ∞; this is done by updating β at a
slower time-scale than the iterates of BASICGIBBS. This new
algorithm is called MODIFIEDGIBBS (Algorithm 2).
Theorem 3: Under
MODIFIEDGIBBS
algorithm,
the Markov chain {B(t)}t≥0 is strongly ergodic,
and the
P limiting probability distribution satisfies
limt→∞ A∈arg minC∈B h(C) P(B(t) = A) = 1.
Proof: See Appendix C. We have used the notion of
weak and strong ergodicity of time-inhomogeneous Markov
chains from [10, Chapter 6, Section 8]), which is provided in
Appendix B. The proof is similar to the proof of one theorem
in [9], but is given here for completeness.
Remark 3: Theorem 3 shows that we can solve (UP) exactly
if we run MODIFIEDGIBBS algorithm for infinite time, in
contrast with BASICGIBBS algorithm which provides an
This algorithm is same as BASICGIBBS algorithm
except that at time t, we use β(t) := β(0) log(1 + t) to
compute the update probabilities, where β(0) > 0,
β(0)N ∆ < 1, and ∆ := maxB∈B,A∈B |h(B) − h(A)|.
Algorithm 2: MODIFIEDGIBBS algorithm
approximate solution.
Remark 4: For i.i.d. time varying {X(t)}t≥0 with known
joint distribution, we can either: (i) find the optimal configuration B ∗ using MODIFIEDGIBBS and use B ∗ for ever, or
(ii) run MODIFIEDGIBBS at the same timescale as t, and
use the running configuration B(t) for sensor activation; both
schemes will minimize the time-average expected cost.
C. Convergence rate of BASICGIBBS
Let µt denote the probability distribution of B(t) under
BASICGIBBS. Let us consider the transition probability matrix P of the Markov chain {X(l)}l≥0 with X(l) = B(lN ),
under the BASICGIBBS algorithm. Let us recall the definition of the Dobrushin’s ergodic coefficient δ(P ) from [10,
Chapter 6, Section 7] for the matrix P ; using a method
similar to that of the proof of Theorem 3, we can show that
−βN ∆
δ(P ) ≤ (1 − e N N ). Then, by [10, Chapter 6, Theorem 7.2],
we can say that under BASICGIBBS
algorithm,
we have
dV (µlN , πβ ) ≤ dV (µ0 , πβ ) 1 −
e−βN ∆
NN
l
. We can prove
similar bounds for any t = lN + k, where 0 ≤ k ≤ N − 1.
Unfortunately, we are not aware of such a bound for
MODIFIEDGIBBS.
Remark 5: Clearly, under BASICGIBBS algorithm, the
convergence rate decreases as β increases. Hence, there is
a trade-off between convergence rate and accuracy of the
solution in this case. Also, the rate of convergence decreases
with N . For MODIFIEDGIBBS algorithm, the convergence
rate is expected to decrease with time.
IV. G IBBS SAMPLING AND STOCHASTIC APPROXIMATION
BASED APPROACH TO SOLVE THE CONSTRAINED PROBLEM
In Section III, we presented Gibbs sampling based algorithms for (UP). In this section, we provide an algorithm that
updates λ with time in order to meet the constraint in (CP)
with equality, and thereby solves (CP) (via Theorem 1).
Lemma 1: The optimal mean number of active sensors,
E|B|1 , for the unconstrained problem (UP), decreases with
λ. Similarly, the optimal error, EdB (X, X̂), increases with λ.
Proof: See Appendix D.
Lemma 1 provides an intuition about how to update λ in
BASICGIBBS or in MODIFIEDGIBBS in order to solve (CP).
We seek to provide one algorithm which updates λ(t) in
each iteration, based on the number of active sensors in the
previous iteration. In order to maintain the necessary timescale
difference between the {B(t)}t≥0 process and the λ(t) update
process, we use stochastic approximation ( [11]) based update
rules for λ(t).
Choose any initial B(0) ∈ {0, 1}N and λ(0) ≥ 0. At
each discrete time instant t = 0, 1, 2, · · · , pick a random
sensor jt ∈ N independently and uniformly. For sensor
jt , choose Bjt (t) = 1 with probability
−βhλ(t) (B−j (t−1),1)
t
e
p := −βhλ(t) (B−j
(t−1),1)
−βhλ(t) (B−j (t−1),0) and choose
t
t
e
+e
Bjt (t) = 0 with probability (1 − p). For k 6= jt , we
choose Bk (t) = Bk (t − 1).
After this operation, before the (t + 1) decision instant,
update λ(t) at each node as follows.
Choose any initial estimate
θ1 . Sample the sensor
j1 = arg minj∈N E dB:Bj =1,||B||1 =1 (X, X̂) θ1 . In
general, after sampling nodes j1 , j2 , · · · , jk and
observing the partial data X j1 = xj1 , · · · , X jk = xjk ,
obtain a new estimate θk+1 by completely running the
EM algorithm using the available partial data and
starting from the initial estimate θk . Once θk+1 is
obtained, sample
jk+1
λ(t + 1) = [λ(t) + a(t)(||B(t − 1)||1 −
N̄ )]cb
The stepsize
P∞{a(t)}t≥1 constitutes
P∞a positive sequence
such that t=1 a(t) = ∞ and t=1 a2 (t) = ∞. The
nonnegative projection boundaries b and c for the λ(t)
iterates are such that λ∗ ∈ (b, c) where λ∗ is defined in
Assumption 1.
=
arg
min
EB
j∈N ,j ∈{j
/ 1 ,··· ,jk }
dB (X, X̂) X j1 = xj1 , · · · , X jk = xjk ; θk+1
where B is such that Bj = Bj1 = · · · = Bjk = 1, and
||B||1 = k + 1. Continue this process until the N -th
sensor is sampled.
Algorithm 4: EMSTATIC algorithm
Algorithm 3: GIBBSLEARNING algorithm
Remark 6: The optimal mean number of active sensors,
E||B||1 , for the unconstrained problem (UP) is a decreasing
staircase function of λ, where each point of discontinuity is
associated with a change in the optimizer B ∗ (λ).
The above remark tells us that the optimal solution of the
constrained problem (CP) requires us to randomize between
two values of λ in case the optimal λ∗ as in Theorem 1 belongs
to the set of such discontinuities. However, this randomization
will require us to update a randomization probability at another
timescale; having stochastic approximations running in multiple timescales leads to very slow convergence and hence is
not a very practical solution for (CP). Hence, instead of using
a varying β(t), we use a fixed, but large β and update λ(t) in
an iterative fashion using stochastic approximation.
Before proposing the algorithm, we provide a result analogous to that in Lemma 1.
Lemma 2: Under BASICGIBBS algorithm for any given
β > 0, the mean number of active sensors E||B||1 is a
continuous and decreasing function of λ.
Proof: See Appendix E.
Let us fix any β > 0. We make the following feasibility
assumption for (CP), under the chosen β > 0.
Assumption 1: There exists λ∗ ≥ 0 such that the constraint
in (CP) under λ∗ and BASICGIBBS is met with equality.
Remark 7: By Lemma 2, E||B||1 continuously decreases in
λ. Hence, if N̄ is feasible, then such a λ∗ must exist by the
intermediate value theorem.
Let us define: hλ(t) (B) := EdB (X, X̂) + λ(t)||B||1 .
Our proposed algorithm GIBBSLEARNING (Algorithm 3)
updates λ(t) iteratively in order to solve (CP).
Discussion of GIBBSLEARNING algorithm:
•
If ||B(t−1)||1 is more than N̄ , then λ(t) is increased with
the hope that this will reduce the number of active sensors
in subsequent iterations, as suggested by Lemma 2.
•
The B(t) and λ(t) processes run on two different
timescales; B(t) runs in the faster timescale whereas λ(t)
runs in a slower timescale. This can be understood from
the fact that the stepsize in the λ(t) update process decreases with time t. Here the faster timescale iterate will
view the slower timescale iterate as quasi-static, while the
slower timescale iterate will view the faster timescale as
almost equilibriated. This is reminiscent of two-timescale
stochastic approximation (see [11, Chapter 6]).
Let πβ|λ∗ (·) denote πβ (·) under λ = λ∗ .
Theorem 4: Under GIBBSLEARNING algorithm and Assumption 1, we have λ(t) → λ∗ almost surely, and the limiting
distribution of {B(t)}t≥0 is πβ|λ∗ (·).
Proof: See Appendix F.
This theorem says that GIBBSLEARNING produces a configuration from the distribution πβ|λ∗ (·) under steady state.
A. A hard constraint on the number of activated sensors
Let us consider the following modified constrained problem:
min EdB (X, X̂) s.t. ||B||1 ≤ N̄
B∈B
(MCP)
It is easy to see that (MCP) can be easily solved using similar
Gibbs sampling algorithms as in Section III, where the Gibbs
sampling algorithm runs only on the set of configurations
which activate N̄ number of sensors. Thus, as a by-product,
we have also proposed a methodology for the problem in [8],
though our framework is more general than [8].
Remark 8: The constraint in (CP) is weaker than (MCP).
Remark 9: If we choose β very large, then the number of
sensors activated by GIBBSLEARNING will have very small
variance. This allows us to solve (MCP) with high probability.
V. E XPECTATION MAXIMIZATION BASED ALGORITHM FOR
PARAMETERIZED DISTRIBUTION OF DATA
In previous sections, we assumed that the joint distribution
of X is completely known to the fusion center. In case this
joint distribution is not known but a parametric form p(x|θ) of
the distribution is known with unknown parameter θ, selecting
all active sensors at once might be highly suboptimal, and a
better approach would be to sample sensor nodes sequentially
and refine the estimate of θ using the data collected from a
newly sampled sensor. We use standard expectation maximization (EM) algorithm (see [12, Section 5.2]) to refine the estimate of θ. Hence, we present a greedy algorithm EMSTATIC
(Algorithm 4) to solve (MCP):
Remark 10: This algorithm is based on heuristics, and
it does not have any optimality guarantee because (i) EM
algorithm yields a parameter value which corresponds to
only a local maximum of the log-likelihood function of the
observed data, and (ii) the greedy algorithm to pick the nodes
is suboptimal.
The performance of EMSTATIC algorithm depends on the
initial value θ1 , since θ1 will determine {θk }k=1,2,··· ,N̄ and
the chosen subset of activated sensors. If θ1 happens to be
initialized at a favourable value, then EMSTATIC algorithm
might even yield the same optimal subset of sensors as in
MCP with θ known apriori. One trivial example for this case
would be when N = 1 and we set θ1 = θ.
In case X(t) ∼ p(x|θ) varies in time t in an i.i.d.
fashion, we can employ the EMSEQUENTIAL algorithm
(Algorithm 5) to find the optimal subset of sensors at each
discrete time slot t.
Remark 11: The
performance
of
EMSEQUENTIAL algorithm depends on the initial estimate
θ1 . Also, the maximization
operation B(1)
=
arg minB(1)∈B:||B(1)||1 =N E dB(1) (X(1), X̂(1)) θ1
Choose any initial estimate θ1 . In slot t = 1, choose the
configuration B(1) of sensors
B(1) =
arg minB(1)∈B:||B(1)||1 =N E dB(1) (X(1), X̂(1)) θ1 .
Then update the parameter to θ2 using EM algorithm
with the partial observation X B(1) (1) = xB(1) (1) and
with initial estimate θ1 . Useθ2 to choose B(2) =
arg minB(2)∈B:||B(2)||1 =N E dB(2) (X(2), X̂(2)) θ2
slot t = 2. Continue this procedure for all t.
Algorithm 5: EMSEQUENTIAL algorithm
at each node is perfect,2 and that the fusion center estimates X̂
from the observation {Xi }i∈S =: X S as E(X|X S ), where S
is the set of active sensors. Under such an estimation scheme,
the conditional distribution of XS c is still a jointly Gaussian
random vector with mean E(X S c |X S ) and the covariance
matrix M (S c , S c ) − M (S c , S)M (S, S)−1 M (S, S c ) (see [12,
Proposition 3.4.4]), where M (S, S c ) is the restriction of M
to the rows indexed by S and the columns indexed by S c .
The trace of this covariance matrix gives the MMSE when
the subset S of sensors are active.
In this scenario, in Figure 1, we compare the cost for the
following four algorithms:
•
•
•
can be
efficiently done by employing Gibbs sampling algorithms as
in Section III; this shows the potential use of Gibbs sampling
in solving sensor subset selection problem for parameterized
distribution of data with unknown parameters. However,
since this is not the main focus of our paper, we will only
consider known data distribution from now on.
VI. N UMERICAL R ESULTS
A. Performance of BASICGIBBS algorithm
For the sake of illustration, we consider N = 18 sensors
which are supposed to sense X = {X1 , X2 , · · · , X18 }, where
X is a jointly Gaussian random vector with covariance matrix
M . Sensor k has access only to Xk . The matrix M is chosen
as follows. We generated a random N × N matrix A whose
elements are uniformly and independently distributed over the
interval [−1, 1], and set M = AT A as the covariance matrix
of X. We set sensor activation cost λ = 2.3, and seek to solve
(UP) with MMSE as the error metric. We assume that sensing
in
•
OPTIMAL: Here we consider the minimum possible cost
for (UP).
BASICGIBBS under steady state: Here we assume that
the configuration B ∈ B is chosen according to the
distribution πβ (·) defined in Section III. This is done for
several values of β.
BASICGIBBS with finite iteration: Here we run BASICGIBBS algorithm for 100 iterations. This is done
independently for several values of β, where for each
β the iteration starts from an independent random configuration. Note that, we have simulated only one sample
path of BASICGIBBS for each β; if the algorithm is run
again independently, the results will be different.
GREEDY: Start with an empty set S, and find the cost
if this subset of sensors are activated. Then compare this
cost with the cost in case sensor 1 is added to this set. If
it turns out that adding sensor 1 to this set S reduces the
cost, then add sensor 1 to the set S; otherwise, remove
sensor 1 from set S. Do this operation serially for all
sensors, and activate the sensors given by the final set S.
It turns out that, under the optimal configuration, 12 sensors
are activated and the optimal cost is 32.3647. On the other
hand, GREEDY activates 14 sensors and incurred a cost of
35.9663. However, we are not aware of any monotonicity or
supermodularity property of the objective function in (UP);
hence, we cannot provide any constant approximation ratio
guarantee for the problem (UP). On the other hand, we have
already proved that BASICGIBBS performs near optimally
2 However, our analysis can be extended where there is sensing error, but
the distribution of sensing error is known to the fusion center.
50
45
48
Estimation error
44
Cost
40
OPTIMAL
GREEDY
BASICGIBBS with finite iteration
BASICGIBBS under steady state
46
42
40
38
36
35
OPTIMAL
NEWGREEDY
BASICGIBBS under steady state
30
25
20
15
34
10
32
0
2
4
6
8
Fig. 1: Comparison among OPTIMAL, BASICGIBBS under
steady state, GREEDY, and BASICGIBBS with finite iterations, for solving problem (UP). For each β, BASICGIBBS
with finite iterations stops after 100 iterations. Details are
provided in Section VI-A.
for large β. Hence, we choose to investigate the performance
of BASICGIBBS, though it might require more number of
iterations compared to N = 18 iterations for GREEDY. It is
important to note that, (UP) is NP-hard, and BASICGIBBS
allows us to avoid searching over 2N possible configurations.
In Figure 1, we can see that for β ≥ 2, the steady state
distribution πβ (·) of BASICGIBBS achieves better expected
cost than GREEDY, and the cost becomes closer to the optimal
cost as β increases. On the other hand, for each β ≥ 2,
BASICGIBBS after 100 iterations yielded a configuration
that achieves near-optimal cost. Hence, BASICGIBBS with
reasonably small number of iterations can be used to find the
optimal subset of active sensors when N is large.
B. Performance of Gibbs sampling applied to problem (MCP)
Here we seek to solve problem (MCP) with N̄ = 10 under
the same setting as in Section VI-A except that a new sample
of the covariance matrix M is chosen. Here we compare the
estimation error for the following three cases:
•
•
•
0
2
4
6
8
10
10
OPTIMAL: Here we choose an optimal subset for (MCP).
BASICGIBBS under steady state: Here we assume that the
configuration B is chosen according to the steady-state
distribution πβ (·) defined in Section III, but restricted
only to the set {B ∈ B : ||B||1 = N̄ }. This is done by
putting h(B) = EdB (X, X̂) if ||B||1 = N̄ and h(B) =
∞ otherwise. This is done for several values of β.
NEWGREEDY: Start with an empty set S, and find the
estimation error if this subset of sensors are activated.
Then find the sensor j1 which, when added to S, will
result in the minimum estimation error. If the estimation
error for S ∪ {j1 } is less than that of S, then do S =
Fig. 2: Comparison among OPTIMAL, BASICGIBBS under
steady state, and NEWGREEDY, for solving problem (MCP).
Details are provided in Section VI-B.
S ∪ {j1 }. Now find the sensor j2 which, when added
to S, will result in the minimum estimation error. If the
estimation error for S ∪ {j2 } is less than that of S, then
do S = S ∪ {j2 }. Repeat this operation until we have
|S| = N̄ , and activate the set of N̄ sensors given by the
final set S. A similar greedy algorithm is used in [8].
The performances for these three cases are shown in Figure 2. It turns out that, the estimation error for OPTIMAL
and NEWGREEDY are 12.9741 and 15.4343, respectively.
BASICGIBBS outperforms NEWGREEDY for β ≥ 2, and
becomes very close to OPTIMAL performance for β ≥ 5.
C. Convergence speed of GIBBSLEARNING algorithm
We first demonstrate the convergence speed of GIBBSLEARNING algorithm, for one specific sample path.
We consider a setting similar to that of Section VI-A,
except that we fix β = 5. The covariance matrix M is
generated using the same method, but the realization of
M here is different from that in Section VI-A. Under this
setting, for λ∗ = 2, BASICGIBBS algorithm yields the
MMSE 3.5680, and the expected number of sensors activated
by BASICGIBBS algorithm becomes 12.7758. Now, let us
consider problem (CP) with the constraint value N̄ = 12.7758.
Clearly, if GIBBSLEARNING algorithm is employed to find
out the solution of problem (CP) with N̄ = 12.7758, then λ(t)
should converge to λ∗ = 2.
The evolution of λ(t) against the iteration index t is shown
in the top plot in Figure 3. We can see that, starting from
λ(0) = 4 and and using the stepsize sequence a(t) = 1t ,
the iterate λ(t) becomes very close to λ∗ = 2 within
100 iterations. At t = 200, we found that λ(200) = 2.0318.
The resulting configuration yielded by GIBBSLEARNING algorithm at t = 200 achieves MMSE 5.0308 and activates
12 sensors; it is important to remember that these results are
ING algorithm has reasonably fast convergence rate for practical active sensing.
4
3.5
VII. C ONCLUSION
3
(t)
2.5
2
1.5
1
0.5
0
0
50
100
150
200
t
In this paper, we have presented Gibbs sampling, stochastic
approximation and expectation maximization based algorithms
for efficient data estimation in the context of active sensing. We first proposed Gibbs sampling based algorithms for
unconstrained optimization of the estimation error and the
mean number of active sensors, proved convergence of these
algorithms, and provided a bound on the convergence speed.
Next, we proposed an algorithm based on Gibbs sampling
and stochastic approximation, in order to solve a constrained
version of the above unconstrained problem, and proved its
convergence. Finally, we proposed expectation maximization
based algorithms for the scenario where the sensor data is
coming from a distribution with known parametric distribution
but unknown parameter value. Numerical results demonstrate
the near-optimal performance of some of these algorithms with
small number of computations.
As our future research endeavours, we seek to develop
distributed sensor subset selection algorithms to efficiently
track the data varying in time according to a stochastic process.
R EFERENCES
Fig. 3: Illustration for convergence speed of λ(t) in the
GIBBSLEARNING algorithm. Top plot: Result for a single
sample path. Details can be found in Section VI-C. Bottom
plot: Average result over 1000 independent sample paths.
Details can be found in Section VI-C.
for one specific realization of the sample path. On the other
hand, under λ(200) = 2.0318, the steady state distribution
of BASICGIBBS, πβ (·), yields MMSE 3.6524 and mean
number of active sensors 12.7354, which are very close to
the respective target values 3.5680 and N̄ = 12.7758.
However, the top plot in Figure 3 is only for a specific
sample path of GIBBSLEARNING algorithm. In the bottom
plot in Figure 3, we demonstrate convergence speed of λ(t)
averaged over multiple independent sample paths of GIBBSLEARNING algorithm. Here we generate a different covariance matrix M , set λ∗ = 2, and follow the same procedure as
before to set N̄ . Then we run GIBBSLEARNING algorithm
independently 1000 times, starting from λ(0) = 4. The bottom
plot of Figure 3 shows the variation of λ(t) (averaged over
1000 sample paths) with t. We can again see that the average
λ(t) is very close to λ∗ = 2 for t ≥ 100.
Thus, our numerical illustration shows that GIBBSLEARN-
[1] F. Schnitzler, J.Y. Yu, and S. Mannor. Sensor selection for crowdsensing
dynamical systems. In International Conference on Artificial Intelligence
and Statistics (AISTATS), pages 829–837, 2015.
[2] D.S. Zois, M. Levorato, and U. Mitra. Active classification for pomdps:
A kalman-like state estimator. IEEE Transactions on Signal Processing,
62(23):6209–6224, 2014.
[3] D.S. Zois, M. Levorato, and U. Mitra. Energy-efficient, heterogeneous
sensor selection for physical activity detection in wireless body area
networks. IEEE Transactions on Signal Processing, 61(7):1581–1594,
2013.
[4] V. Krishnamurthy and D.V. Djonin. Structured threshold policies for
dynamic sensor schedulinga partially observed markov decision process
approach. IEEE Transactions on Signal Processing, 55(10):4938–4957,
2007.
[5] W. Wu and A. Arapostathis. Optimal sensor querying: General markovian and lqg models with controlled observations. IEEE Transactions
on Automatic Control, 53(6):1392–1405, 2008.
[6] V. Gupta, T.H. Chung, B. Hassibi, and R.M. Murray. On a stochastic
sensor selection algorithm with applications in sensor scheduling and
sensor coverage. Automatica, 42:251–260, 2006.
[7] A. Bertrand and M. Moonen. Efficient sensor subset selection and link
failure response for linear mmse signal estimation in wireless sensor
networks. In European Signal Processing Conference (EUSIPCO), pages
1092–1096. EURASIP, 2010.
[8] D. Wang, J. Fisher III, and Q. Liu. Efficient observation selection
in probabilistic graphical models using bayesian lower bounds. In
Proceedings of the Thirty-Second Conference on Uncertainty in Artificial
Intelligence (UAI), pages 755–764. ACM, 2016.
[9] A. Chattopadhyay and B. Baszczyszyn. Gibbsian on-line distributed
content caching strategy for cellular networks. https:// arxiv.org/ abs/
1610.02318, 2016.
[10] P. Bremaud. Markov Chains, Gibbs Fields, Monte Carlo Simulation,
and Queues. Springer, 1999.
[11] Vivek S. Borkar. Stochastic approximation: a dynamical systems
viewpoint. Cambridge University Press, 2008.
[12] B. Hajek. An Exploration of Random Processes for Engineers. Lecture
Notes for ECE 534, 2011.
A PPENDIX A
P ROOF OF T HEOREM 1
We will prove only the first part of the theorem where there
exists a unique B ∗ . The second part of the theorem can be
proved similarly.
Let us denote the optimizer for (CP) by B, which is
possibly different from B ∗ . Then, by the definition of B ∗ ,
we have EdB∗ (X, X̂) + λ∗ ||B ∗ ||1 ≤ EdB (X, X̂) + λ∗ ||B||1 .
But ||B||1 ≤ K and ||B ∗ ||1 = K. Hence, EdB∗ (X, X̂) ≤
EdB (X, X̂). This completes the proof.
Hence,
∞
X
(1 − δ(Pl ))
l=0
=
≥
≥
=
A PPENDIX B
W EAK AND S TRONG E RGODICITY
Consider a discrete-time Markov chain (possibly not timehomogeneous) {B(t)}t≥0 with transition probability matrix
(t.p.m.) P (m; n) between t = m and t = n. We denote
by D the collection of all possible probasbility distributions
on the state space. Let dV (·, ·) denote the total variation
distance between two distributions in D. Then {B(t)}t≥0
is called weakly ergodic if, for all m ≥ 0, we have
limn↑∞ supµ,ν∈D dV (µP (m; n), νP (m; n)) = 0.
The Markov chain {B(t)}t≥0 is called strongly
ergodic if there exists π
∈
D such that,
limn↑∞ supµ∈D dV (µT P (m; n), π) = 0 for all m ≥ 0.
A PPENDIX C
P ROOF OF T HEOREM 3
We will first show that the Markov chain {B(t)}t≥0 in
weakly ergodic.
Let us define ∆ := maxB∈B,A∈B |h(B) − h(A)|.
Consider the transition probability matrix (t.p.m.) Pl for
the inhomogeneous Markov chain {X(l)}l≥0 (where X(l) :=
B(lN )). The Dobrushin’s ergodic coefficient δ(Pl ) is given
by (see [10, Chapter
6, Section 7] for definition) δ(Pl ) =
P
0
00
1 − inf B 0 ,B 00 ∈B B∈B min{Pl (B , B), Pl (B , B)}. A sufficient condition
P∞for the Markov chain {B(t)}t≥0 to be weakly
ergodic is
l=1 (1 − δ(Pl )) = ∞ (by [10, Chapter 6,
Theorem 8.2]).
Now, with positive probability, activation states for all nodes
0
are updated over a period of N slots. Hence, Pl (B , B) > 0 for
0
all B , B ∈ B. Also, once a node jt for t = lN + k is chosen
in MODIFIEDGIBBS algorithm, the sampling probability for
−β(lN +k)∆
any activation state in a slot is greater than e
. Hence,
2
for independent sampling over N slots, we have, for all pairs
0
B , B:
N
−1 −β(lN +k)∆
Y
0
e
Pl (B , B) >
>0
2N
k=0
≥
∞
X
0
l=0
∞
X
X
inf00
B ,B ∈B
2N
0
00
min{Pl (B , B), Pl (B , B)}
B∈B
N
−1
−β(0) log(1+lN +k)×∆
Y
e
2N
l=0
k=0
∞ N
−1 −β(0) log(1+lN +N )×∆
X
Y
e
l=0 k=0
∞
X
1
NN
l=1
1
N N +1
N
1
(1 + lN )β(0)N ∆
∞
X
1
β(0)N ∆
(1
+
i)
i=N +1
= ∞
(2)
Here the first inequality uses the fact that the cardinality of
B is 2N . The second inequality follows from replacing k by
N in the numerator. The thirdP
inequality follows from lowerlN +N −1
1
bounding (1+lN )1β(0)N ∆ by N1 i=lN
β(0)N ∆ . The last
P∞ (1+i)
equality follows from the fact that i=1 i1a diverges for 0 <
a < 1.
Hence, the Markov chain {B(t)}t≥0 is weakly ergodic.
In order to prove strong ergodicity of {B(t)}t≥0 , we
invoke [10, Chapter 6, Theorem 8.3]. We denote the t.p.m.
of {B(t)}t≥0 at a specific time t = T0 by Q(T0 ) , which is
a given specific matrix. If {B(t)}t≥0 evolves up to infinite
time with fixed t.p.m. Q(T0 ) , then it will reach the stationary
−βT h(B)
distribution πβT0 (B) = e Zβ0
. Hence, we can claim that
T0
Condition 8.9 of [10, Chapter 6, Theorem 8.3] is satisfied.
Next, we check Condition 8.10 of [10, Chapter 6, Theo0
rem 8.3]. For any B ∈ arg minB 0 ∈B h(B ), we can argue
that πβT0 (B) increases with T0 for sufficiently large T0 ; this
can be verified by considering the derivative of πβ (B) w.r.t.
0
β. For B ∈
/ arg minB 0 ∈B h(B ), the probability πβT0 (B)
decreases with T0 for large T0 . Now, using the fact that
any
bounded sequence converges, we can write
P∞ monotone,
P
T0 =0
B∈B |πβT0 +1 (B) − πβT0 (B)| < ∞.
Hence, by [10, Chapter 6, Theorem 8.3], the Markov chain
{B(t)}t≥0 is strongly ergodic. It is straightforward to verify
the claim regarding the limiting distribution.
A PPENDIX D
P ROOF OF L EMMA 1
Let λ1 > λ2 > 0, and the corresponding optimal error and
mean number of active sensors under these multiplier values
be (d1 , n1 ) and (d2 , n2 ), respectively. Then, by definition, d1 +
λ1 n1 ≤ d2 + λ1 n2 and d2 + λ2 n2 ≤ d1 + λ2 n1 . Adding these
two inequalities, we obtain λ1 n1 + λ2 n2 ≤ λ1 n2 + λ2 n1 ,
i.e., (λ1 − λ2 )n1 ≤ (λ1 − λ2 )n2 . Since λ1 > λ2 , we obtain
n1 ≤ n2 . This completes the first part of the proof. The second
part of the proof follows using similar arguments.
A PPENDIX E
P ROOF OF L EMMA 2
||B|| e−βh(B)
P
Let us denote E||B||1 =: f (λ) = B∈B Zβ1
. It is
straightforward to see that E||B||1 is continuously differentiable in λ.
Let us denote Zβ by Z for simplicity, and let h(B) = dB +
λ||B||1 be the linear combination of the error (here we have
written EdB (·, ·) as dB for simplicity in notation) and number
of active sensors under configuration B. Then the derivative
of f (λ) w.r.t. λ is given by:
f 0 (λ)
=
−Zβ
P
B∈B
||B||21 e−β(dB +λ||B||1 ) −
Z2
P
Now, it is straightforward to verify that
Hence,
B∈B
dZ
dλ
||B||1 e−β(dB +λ||B||1 ) dZ
dλ
= −βZf (λ).
0
f (λ)
−Zβ
=
P
−β(dB +λ||B||1 ) βZf (λ)
2 −β(dB +λ||B||1 ) + P
B∈B ||B||1 e
B∈B ||B||1 e
Z2
Now, f 0 (λ) ≤ 0 is equivalent to
P
||B||21 e−β(dB +λ||B||1 )
f (λ) ≤ PB∈B
−β(dB +λ||B||1 )
B∈B ||B||1 e
Noting that E||B||1 =: f (λ) and dividing the numerator and
denominator of R.H.S. by Z, the condition is reduced to
E||B||2
E||B||1 ≤ E||B||11 , which is true since E||B||21 ≥ (E||B||1 )2 .
Hence, E||B||1 is decreasing in λ for any β > 0.
A PPENDIX F
P ROOF OF T HEOREM 4
Let the distribution of B(t) under GIBBSLEARNING algorithm be µt (·). Since limt→∞ a(t) = 0, it follows
that limt→∞ dV (µt−1 , πβ|λ(t−1) ) = 0 (where dV (·, ·) is
the total variation distance), and limt→∞ (Eµt−1 ||B||1 −
Eπβ |λ(t−1) ||B||1 ) := limt→∞ e(t) = 0. Now, we can rewrite
the λ(t) update equation as follows:
λ(t + 1) = [λ(t) + a(t)(Eπβ |λ(t−1) ||B||1 − N̄ + Mt + et )]cb
(3)
Here Mt := ||B(t−1)||1 −Eµt−1 ||B(t−1)||1 is a Martingale
difference noise sequence, and limt→∞ et = 0. It is easy to
see that the derivative of Eπβ |λ ||B||1 w.r.t. λ is bouned for λ ∈
[b, c]; hence, Eπβ |λ ||B||1 is a Lipschitz continuous function of
λ. It is also easy to see that the sequence {Mt }t≥0 is bounded.
Hence, by the theory presented in [11, Chapter 2] and [11,
Chapter 5, Section 5.4], λ(t) converges to the unique zero
of Eπβ |λ ||B||1 − N̄ almost surely. Hence, λ(t) → λ∗ almost
surely. Since limt→∞ dV (µt−1 , πβ|λ(t−1) ) = 0 and πβ|λ is
continuous in λ, the limiting distribution of B(t) becomes
πβ|λ∗ .
| 3cs.SY
|
Event excitation for event-driven control and optimization of
multi-agent systems
Yasaman Khazaeni and Christos G. Cassandras
Division of Systems Engineering
and Center for Information and Systems Engineering
Boston University, MA 02446
arXiv:1604.00691v1 [math.OC] 3 Apr 2016
[email protected],[email protected]
Abstract— We consider event-driven methods in a general
framework for the control and optimization of multi-agent
systems, viewing them as stochastic hybrid systems. Such
systems often have feasible realizations in which the events
needed to excite an on-line event-driven controller cannot occur,
rendering the use of such controllers ineffective. We show that
this commonly happens in environments which contain discrete
points of interest which the agents must visit. To address this
problem in event-driven gradient-based optimization problems,
we propose a new metric for the objective function which
creates a potential field guaranteeing that gradient values
are non-zero when no events are present and which results
in eventual event excitation. We apply this approach to the
class of cooperative multi-agent data collection problems using the event-driven Infinitesimal Perturbation Analysis (IPA)
methodology and include numerical examples illustrating its
effectiveness.
I. I NTRODUCTION
The modeling and analysis of dynamic systems has historically been founded on the time-driven paradigm provided
by a theoretical framework based on differential (or difference) equations: we postulate the existence of an underlying
“clock” and with every “clock tick” a state update is performed which synchronizes all components of the system. As
systems have become increasingly networked, wireless, and
distributed, the universal value of this paradigm has come
to question, since it may not be feasible to guarantee the
synchronization of all components of a distributed system,
nor is it efficient to trigger actions with every time step when
such actions may be unnecessary. The event-driven paradigm
offers an alternative to the modeling, control, communication, and optimization of dynamic systems. The main idea
in event-driven methods is that actions affecting the system
state need not be taken at each clock tick. Instead, one can
identify appropriate events that trigger control actions. This
approach includes the traditional time-driven view if a clocktick is considered a system “event”. Defining the right events
is a crucial modeling step and has to be carried out with a
good understanding of the system dynamics.
The importance of event-driven behavior in dynamic systems was recognized in the development of Discrete Event
The authors work is supported in part by NSF under grants CNS1239021, ECCS-1509084, and IIP-1430145, by AFOSR under grant
FA9550-15-1-0471, by ONR under grant N00014-09-1-1051, and by the
Cyprus Research Promotion Foundation under Grant New Infrastructure
Project/Strategic/0308/26.
Systems (DES) and later Hybrid Systems (HS) [1]. More
recently there have been significant advances in applying
event-driven methods (also referred to as “event-based” and
“event-triggered”) to classical feedback control systems; e.g.,
see [2], [3], [4], as well as [5] and [6] and references
therein. Event-driven approaches are also attractive in receding horizon control, where it is computationally inefficient
to re-evaluate an optimal control value over small time
increments as opposed to event occurrences defining appropriate planning horizons for the controller (e.g., see [7]).
In distributed networked systems, event-driven mechanisms
have the advantage of significantly reducing communication
among networked components which cooperate to optimize
a given objective. Maintaining such cooperation normally
requires frequent communication among them; it was shown
in [8] that we can limit ourselves to event-driven communication and still achieve optimization objectives while
drastically reducing communication costs (hence, prolonging
the lifetime of a wireless network), even when delays are
present (as long as they are bounded).
Clearly, the premise of these methods is that the events
involved are observable so as to “excite” the underlying
event-driven controller. However, it is not always obvious
that these events actually take place under every feasible
control: it is possible that under some control no such events
are excited, in which case the controller may be useless.
In such cases, one can resort to artificial “timeout events”
so as to eventually take actions, but this is obviously inefficient. Moreover, in event-driven optimization mechanisms
this problem results in very slow convergence to an optimum
or in an algorithm failing to generate any improvement in the
decision variables being updated.
In this work, we address this issue of event excitation in
the context of multi-agent systems. In this case, the events
required are often defined by an agent “visiting” a region
or a single point in a mission space S ⊂ R2 . Clearly, it is
possible that such events never occur for a large number
of feasible agent trajectories. This is a serious problem
in trajectory planning and optimization tasks which are
common in multi-agent systems seeking to optimize different
objectives associated with these tasks, including coverage,
persistent monitoring or formation control [9], [10], [11],
[12], [13], [14], [15], [16]. At the heart of this problem is
the fact that objective functions for such tasks rely on a non-
zero reward (or cost) metric associated with a subset S + ⊂ S
of points, while all other points in S have a reward (or cost)
which is zero since they are not “points of interest” in the
mission space. We propose a novel metric which allows all
points in S to acquire generally non-zero reward (or cost),
thus ensuring that all events are ultimately excited. This leads
to a new method allowing us to apply event-based control
and optimization to a large class of multi-agent problems. We
will illustrate the use of this method by considering a general trajectory optimization problem in which Infinitesimal
Perturbation Analysis (IPA) [1] is used as an event-driven
gradient estimation method to seek optimal trajectories for
a class of multi-agent problems where the agents must
cooperatively visit a set of target points to collect associated
rewards (e.g., to collect data that are buffered at these points.)
This defines a family within the class of Traveling Salesman
Problems (TSPs) [17] for which most solutions are based on
techniques typically seeking a shortest path in the underlying
graph. These methods have several drawbacks: (i) they are
generally combinatorially complex, (ii) they treat agents as
particles (hence, not accounting for limitations in motion
dynamics which should not, for instance, allow an agent to
form a trajectory consisting of straight lines), and (iii) they
become computationally infeasible as on-line methods in the
presence of stochastic effects such as random target rewards
or failing agents. As an alternative we seek solutions in terms
of parameterized agent trajectories which can be adjusted on
line as a result of random effects and which are scalable,
hence computationally efficient, especially in problems with
large numbers of targets and/or agents. This approach was
successfully used in [18], [19].
In section II we present the general framework for multiagent problems and address the event excitation issue. In
section III we overview the event-driven IPA methodology
and how it is applied to a general hybrid system optimization
problem. In section IV we introduce a data collection problem as an application of the general framework introduced
in section II and will show simulation results of applying the
new methodology to this example in section V.
II. E VENT-D RIVEN O PTIMIZATION IN M ULTI -AGENT
S YSTEMS
Multi-agent systems are commonly modeled as hybrid
systems with time-driven dynamics describing the motion
of the agents or the evolution of physical processes in a
given environment, while event-driven behavior characterizes
events that may occur randomly (e.g., an agent failure) or
in accordance to control policies (e.g., an agent stopping
to sense the environment or to change directions). In some
cases, the solution of a multi-agent dynamic optimization
problem is reduced to a policy that is naturally parametric. As
such, a multi-agent system can be studied with parameterized
controllers aiming to meet certain specifications or to optimize a given performance metric. Moreover, in cases where
such a dynamic optimization problem cannot be shown to be
reduced to a parametric policy, using such a policy is still
near-optimal or at least offers an alternative.
Fig. 1.
Multi-agent system in a dynamic setting, blue areas are obstacles
In order to build a general framework for multi-agent
optimization problems, assuming S as the mission space,
we introduce the function R(w) : S → R as a “property” of
point w ∈ S. For instance, R(w) could be a weight that gives
relative importance to one point in S compared to another.
Setting R(w) > 0 for only a finite number of points implies
that we limit ourselves to a finite set of points of interest
while the rest of S has no significant value.
Assuming F to be the set of all feasible agent states, We
define P (w, s) : S × F → R to capture the cost/reward
resulting from how agents with state s ∈ F interact with
w ∈ S. For instance, in coverage problems if an “event”
occurs at w, then P (w, s) is the probability of agents jointly
detecting such events based on the relative distance of each
agent from w.
In general settings, the objective is to find the best state
vector s1 , · · · , sN so that N agents achieve a maximal
reward (minimal cost) from interacting with the mission
space S:
Z
min J =
P (w, s)R(w)dw
(1)
s∈F
S
This static problem can be extended to a dynamic version
where the agents determine optimal trajectories si (t), t ∈
[0, T ], rather than static states:
Z TZ
min J =
P (w, s(u(t)))R(w, t)dwdt
(2)
u(t)∈U
0
S
subject to motion dynamics:
ṡj (t) = fj (sj , uj , t), j = 1, · · · , N
(3)
In Fig. 1, such a dynamic multi agent system is illustrated.
As an example, consensus problems are just a special case
of (1). Suppose that we consider a finite set of points w ∈ S
which coincide with the agents states s1 , ..., sN (which are
not necessarily their locations). Then we can set P (w, s) =
ksi − sj k2 and, therefore, replace the integral in (1) by a
sum. In this case, R(w) = Ri is just the weight that an
agent carries in the consensus algorithm. An optimum occurs
when ksi − sj k2 = 0 for all i, j, i.e., all agents “agree”
and consensus is reached. This is a special case because
of the simplicity in P (w, s) making the problem convex so
that a global optimum can be achieved, in contrast to most
problems we are interested in.
As for the formulation in (2), consider a trajectory planning problem where N mobile agents are tasked to visit
M stationary targets in the mission space S. Target behavior is described through state variables xi (t) which may
model reward functions, the amount of data present at i, or
other problem-dependent target properties. More formally, let
(Ω, F, P) be an appropriately defined probability space and
ω ∈ Ω a realization of the system where target dynamics are
subject to random effects:
ẋi (t) = gi (xi (t), ω)
(4)
gi (·) is as such that xi (t) is monotonically increasing by t
and it resets to zero each time a target is completely emptied
by an agent. In the context of (2), we assume the M targets
are located at pointswi , i = 1, · · · , M and define
R(xi (t), w) if w ∈ C(wi )
R(w, t) =
(5)
0
otherwise
to be the value of point w, where C(wi ) is a compact 2manifold in R2 containing wi which can be considered to be
a region defined by the sensing range of that target relative
to agents (e.g., a disk centered at wi ). Note that R(w, t) is
also a random variable defined on the same probability space
above. Given that only points w ∈ C(wi ) have value for the
agents, there is an infinite number of points w ∈
/ C(wi ) such
that R(w, t) = 0 provided the following condition holds:
Condition 1: If ∃i such that w ∈ C(wi ) then w ∈
/ C(wj )
holds ∀j 6= i.
This condition is to ensure that two targets do not share
any point w in their respective sensing ranges. Also it ensures
that the set {C(wi ) | i = 1 : · · · , M } does not create a
compact partitioning of the mission space and there exist
points w which do not belong to any of the C(wi ).
Viewed as a stochastic hybrid system, we may define different modes depending on the states of agents or targets and
events that cause transitions between these modes. Relative
to a target i, any agent has at least two modes: being at a
point w ∈ C(wi ), i.e., visiting this target or not visiting it.
Within each mode, agent j’s dynamics, dictated by (3), and
target i’s dynamics in (4) may vary. Accordingly, there are
0
at least two types of events in such a system: (i) δij
events
+
occur when agent j initiates a visit at target i, and (ii) δij
events occur when agent j ends a visit at target i. Additional
event types may be included depending on the specifics of
a problem, e.g., mode switches in the target dynamics or
agents encountering obstacles.
An example is shown in Fig. 2, where target sensing ranges
are shown with green circles and agent trajectories are shown
in dashed lines starting at a base shown by a red triangle.
In the blue trajectory, agent 1 moves along the trajectory
that passes through points A → B → C → D. It is easy
to see that when passing through points A and C we have
0
δi1
and δi00 1 events, while passing through B and D we
+
have δi1
and δi+0 1 events. The red trajectory is an example
where none of the events is excited. Suppose we consider
an on-line trajectory adjustment process in which the agent
improves its trajectory based on its performance measured
through (5). In this case, R(w, t) = 0 over all t, as long
as the agent keeps using the red trajectory, i.e., no event
ever occurs. Therefore, if an event-driven approach is used
to control the trajectory adjustment process, no action is ever
triggered and the approach is ineffective. In contrast, in the
blue trajectory the controller can extract useful information
from every observed event; such information (e.g., a gradient
of J with respect to controllable parameters as described in
the next section) can be used to adjust the current trajectory
Fig. 2.
Sample trajectories
so as to improve the objective function J in (1) or (2).
Therefore, if we are to build an optimization framework
for this class of stochastic hybrid systems to allow the application of event-driven methods by calculating a performance
measure gradient, then a fundamental property required is the
occurrence of at least some events in a sample realization. In
particular, the IPA method [20] is based on a single sample
realization of the system over which events are observed
along with their occurrence times and associated system
states. Suppose that the trajectories can be controlled through
a set of parameters forming a vector θ. Then, IPA provides
an unbiased estimate of the gradient of a performance metric
J(θ) with respect to θ. This gradient is then used to improve
the trajectory and ultimately seek an optimal one when
appropriate conditions hold.
As in the example of Fig. 2, it is possible to encounter
trajectory realizations where no events occur in the system.
In the above example, this can easily happen if the trajectory
does not pass through any target. The existence of such
undesirable trajectories is the direct consequence of Condition 1. This lack of event excitation results in event-based
controllers being unsuitable.
New Metric: In order to overcome this issue we propose
a new definition for R(w, t) in (5) as follows:
M
X
R(w, t) =
hi (xi (t), di (w))
(6)
i=1
where w ∈ S, hi (·) is a function of the target’s state xi (t)
and di (w) = kwi −wk. Note that, if hi (·) is properly defined,
(6) yields R(w, t) > 0 at all points.
While the exact form of hi (·) depends on the problem, we
impose the condition that hi (·) is monotonically decreasing
in di (w). We can think of hi (·) as a value function associated
with point wi . Using the definition of R(w, t), this value
is spread out over all points w ∈ S rather than being
concentrated at the single point wi . This creates a continuous
potential field for the agents leading to a non-zero gradient of
the performance measure even when the trajectories do not
excite any events. This non-zero gradient will then induce
trajectory adjustments that naturally bring them toward ones
with observable events.
Finally, recalling the definition in (2), we also define:
N
X
P (w, s) =
ksj (t) − wk2
(7)
j=1
the total quadratic travel cost for agents to visit point w.
In Section IV, we will show how to apply R(w, t) and
P (w, s) defined as above in order to determine optimal agent
trajectories for a class of multi-agent problems of the form
(2). First, however, we review in the next section the eventdriven IPA calculus which allows us to estimate performance
gradients with respect to controllable parameters.
III. E VENT-D RIVEN IPA C ALCULUS
Let us fix a particular value of the parameter θ ∈ Θ and
study a resulting sample path of a general SHS. Over such a
sample path, let τk (θ), k = 1, 2, · · · denote the occurrence
times of the discrete events in increasing order, and define
τ0 (θ) = 0 for convenience. We will use the notation τk
instead of τk (θ) when no confusion arises. The continuous
state is also generally a function of θ, as well as of t, and is
thus denoted by x(θ, t). Over an interval [τk (θ), τk+1 (θ)),
the system is at some mode during which the time-driven
state satisfies ẋ = fk (x, θ, t), in which x is any of the
continuous state variables of the system and ẋ denotes ∂x
∂t .
Note that we suppress the dependence of fk on the inputs
u ∈ U and d ∈ D and stress instead its dependence on
the parameter θ which may generally affect either u or d
or both. The purpose of perturbation analysis is to study
how changes in θ influence the state x(θ, t) and the event
times τk (θ) and, ultimately, how they influence interesting
performance metrics that are generally expressed in terms of
these variables.
An event occurring at time τk+1 (θ) triggers a change
in the mode of the system, which may also result in new
dynamics represented by fk+1 . The event times τk (θ) play
an important role in defining the interactions between the
time-driven and event-driven dynamics of the system.
Following the framework in [20], consider a general
performance function J of the control parameter θ:
J(θ; x(θ, 0), T ) = E[L(θ; x(θ, 0), T )]
(8)
where L(θ; x(θ, 0), T ) is a sample function of interest evaluated in the interval [0, T ] with initial conditions x(θ, 0).
For simplicity, we write J(θ) and L(θ). Suppose that there
are K events, with occurrence times generally dependent on
θ, during the time interval [0, T ] and define τ0 = 0 and
τN +1 = T . Let Lk : Rn × Θ × R+ → R be a function and
define L(θ) by
K Z τk+1
X
L(θ) =
Lk (x, θ, t)dt
(9)
k=0
τk
where we reiterate that x = x(θ, t) is a function of θ and
t. We also point out that the restriction of the definition of
J(θ) to a finite horizon T which is independent of θ is made
merely for the sake of simplicity. Returning to the stochastic
setting, the ultimate goal of the iterative process shown is to
maximize Eω [L(θ, ω)], where we use ω to emphasize dependence on a sample path ω of a SHS (clearly, this is reduced to
L(θ) in the deterministic case). Achieving such optimality is
possible under standard ergodicity conditions imposed on the
underlying stochastic processes, as well as the assumption
that a single global optimum exists; otherwise, the gradientbased approach is simply continuously attempting to improve
the observed performance L(θ, ω). Thus, we are interested
in estimating the gradient
dEω [L(θ, ω)]
dJ(θ)
=
(10)
dθ
dθ
based on directly observed data. We
by evaluating dL(θ,ω)
dθ
obtain θ ∗ by optimizing J(θ) through an iterative scheme
of the form
θn+1 = θn − ηn Hn (θn ; x(θ, 0), T, ωn ), n = 0, 1, · · · (11)
where ηn is a step size sequence and Hn (θn ; x(θ, 0), T, ωn )
at θ = θn . In using IPA,
is the estimate of dJ(θ)
dθ
,
Hn (θn ; x(θ, 0), T, ωn ) is the sample derivative dL(θ,ω)
dθ
which is an unbiased estimate of dJ(θ)
if
the
condition
dθ
(dropping the symbol ω for simplicity)
dL(θ) dE[L(θ)]
dJ(θ)
E
=
=
(12)
dθ
dθ
dθ
is satisfied, which turns out to be the case under mild
technical conditions. The conditions under which algorithms
of the form (11) converge are well-known (e.g., see [21]).
Moreover, in addition to being unbiased, it can be shown that
such gradient estimates are independent of the probability
laws of the stochastic processes involved and require minimal
information from the observed sample path. The process
is based on analyzing
through which IPA evaluates dL(θ)
dθ
how changes in θ influence the state x(θ, t) and the event
times τk (θ). In turn, this provides information on how L(θ)
is affected, because it is generally expressed in terms of these
variables. Given θ = [θ1 , ..., θl ]T , we use the Jacobian matrix
notation:
∂x(θ, t)
∂τk (θ)
x0 (θ, t) =
, τk 0 =
, k = 1, · · · , K (13)
∂θ
∂θ
for all state and event time derivatives. For simplicity of
notation, we omit θ from the arguments of the functions
above unless it is essential to stress this dependence. It is
shown in [20] that x0 (t) satisfies:
dx0 (t)
∂fk (t) 0
∂fk (t)
=
x (t) +
(14)
dt
∂x
∂θ
for t ∈ [τk (θ), τk+1| (θ)) with boundary condition
x0 (τk+ ) = x0 (τk− ) + [fk−1 (τk− ) − fk (τk+ )]τk0
(15)
for k = 0, · · · , K. We note that whereas x(t) is often
continuous in t, x0 (t) may be discontinuous in t at the event
times τk ; hence, the left and right limits above are generally
different. If x(t) is not continuous in t at t = τk (θ), the value
of x(τk+ ) is determined by the reset function r(q, q 0 , x, ν, δ)
and
dr(q, q 0 , x, ν, δ)
x0 (τk+ ) =
(16)
dθ
0 +
Furthermore, once the initial condition x (τk ) is given,
the linearized state trajectory x0 (t) can be computed in the
interval t ∈ [τk (θ), τk+1 (θ)) by solving (14) to obtain
R t ∂fk (u)
h Z t ∂f (v) R t ∂fk (u)
i
du
du
−
k
0
∂x
τk
x (t) = e
e τk ∂x
dv + ξk
∂θ
τk
(17)
with the constant ξk determined from x0 (τk+ ). In order to
complete the evaluation of x0 (τk+ ) we need to also determine
τk0 . If the event at τk (θ) is exogenous τk0 = 0 and if the event
at τk (θ) is endogenous:
h ∂g
i ∂g
∂gk 0 −
k
k
τk0 = −
fk (τk− )
+
x (τk )
(18)
∂x
∂θ
∂x
∂gk
+
where gk (x, θ) = 0 and it is defined as long as ∂x fk (τk ) 6=
0 (details may be found in [20].)
The derivative evaluation process involves using the IPA
calculus in order to evaluate the IPA derivative dL
dθ . This is
accomplished by taking derivatives in (9) with respect to θ:
Z τk+1
K
dL(θ) X d
Lk (x, θ, t)dt
(19)
=
dθ
dθ τk
k=0
Applying the Leibnitz rule, we obtain, for every k =
0, · · · , ZK,
τk+1
d
Lk (x, θ, t)dt
dθ τk
Z τk+1 h
∂Lk (x, θ, t) i
∂Lk (x, θ, t) 0
dt
x (t) +
=
∂x
∂θ
τk
0
+ Lk (x(τk+1 ), θ, τk+1 )τk+1
− Lk (x(τk ), θ, τk )τk0
(20)
In summary the three equations (15), (17) and (18) form
the basis of the IPA calculus and allow us to calculate the
final derivative in (20). In the next section IPA is applied to
a data collection problem in a multi-agent system.
connected to a target i even if there are other agents l with
pil (t) > 0; this is not the only possible model, but we adopt
it based on the premise that simultaneous downloading of
packets from a common source creates problems of proper
data reconstruction. This means that j in (22) is the index
of the agent that is connected to target i at time t.
The dynamics of xi (t) in (22) results in two new event
types added to what was defined earlier, (i) ξi0 events occur
when xi (t) reaches zero, and (ii) ξi+ events occur when xi (t)
leaves zero.
The performance measure is the total content of data left
at targets at the end of a finite mission time T . Thus, we
define J1 (t) to be the following (recalling that {σi (t)} are
random processes):
M
X
J1 (t) =
αi E[xi (t)]
(23)
i=1
IV. T HE DATA C OLLECTION P ROBLEM
We consider a class of multi-agent problems where the
agents must cooperatively visit a set of target points to collect
associated rewards (e.g., to collect data that are buffered at
these points.). The mission space is S ⊂ R2 . This class of
problems falls within the general formulation introduced in
(2). The state of the system is the position of agent j time
t, sj (t) = [sxj (t), syj (t)] and the state of the target i, xi (t).
The agent’s dynamics (3) follow a single integrator:
ṡxj (t) = uj (t) cos θj (t),
ṡyj (t) = uj (t) sin θj (t) (21)
where uj (t) is the scalar speed of the agent (normalized so
that 0 ≤ uj (t) ≤ 1) and θj (t) is the angle relative to the
positive direction, 0 ≤ θj (t) < 2π. Thus, we assume that
each agent controls its speed and heading.
We assume the state of the target xi (t) represents the
amount of data that is currently available at target i (this can
be modified to different state interpretations). The dynamics
of xi (t) in (4) for this problem are:
0
if xi (t) = 0 and σi (t) ≤ µij p(sj (t), wi )
ẋi (t) =
σi (t) − µij p(sj (t), wi )
otherwise
(22)
i.e., we model the data at the target as satisfying simple flow
dynamics with an exogenous (generally stochastic) inflow
σi (t) and a controllable rate with which an agent empties
the data queue given by µij p(sj (t), wi ). For brevity we set
p(sj (t), wi ) = pij (t) which is the normalized data collection
rate from target i by agent j and µij is a nominal rate
corresponding to target i and agent j.
Assuming M targets are located at wi ∈ S, i = 1, . . . , M,
and have a finite range of ri , then agent j can collect data
from wi only if dij (t) = kwi − sj (t)k ≤ ri . We then
assume that: (A1) pij (t) ∈ [0, 1] is monotonically nonincreasing in the value of dij (t) = kwi − sj (t)k, and (A2)
it satisfies pij (t) = 0 if dij (t) > ri . Thus, pij (t) can
model communication power constraints which depend on
the distance between a data source and an agent equipped
with a receiver (similar to the model used in [22]) or sensing
range constraints if an agent collects data using on-board
sensors. For simplicity, we will also assume that: (A3) pij (t)
is continuous in dij (t) and (A4) only one agent at a time is
where αi is a weight factor for target i. We can now
formulate a stochastic optimization problem P1 where the
control variables are the agent speeds and headings denoted
by the vectors u(t) = [u1 (t), . . . , uN (t)] and θ(t) =
[θ1 (t), . . . , θN (t)] respectively (omitting their dependence on
the full system state at t).
Z
1 T
J1 (t)dt
(24)
P1 :
min J(T ) =
T 0
u(t),θ(t)
where 0 ≤ uj (t) ≤ 1, 0 ≤ θj (t) < 2π, and T is a given
finite mission time. This problem can be readily placed into
the general framework (2). In particular, the right hand side
of (24) is:"
#
Z T XZ
1
αi
E
2 xi (t)dwdt
T
0
C(wi ) πri
i
"Z Z
# (25)
T
X αi 1{w ∈ C(wi )}
1
xi (t)dwdt
= E
T
πri2
0
S i
This is now in the form of the general framework in (2) with
X αi 1{w ∈ C(wi )}
R(w, t) =
xi (t)
(26)
πri2
i
and
P (sj (t), w) = 1
(27)
Recalling the definition in (5), only points within the sensing
range of each target have non-zero values, while all other
point value are zero, which is the case in (26) above. In
addition, (27) simply shows that there is no meaningful
dynamic interaction between an agent and the environment.
Problem P1 is a finite time optimal control problem. In
order to solve this, following previous work in [19] we
proceed with a standard Hamiltonian analysis leading to a
Two Point Boundary Value Problem (TPBVP) [23]. We omit
this, since the details are the same as the analysis in [19]. The
main result of the Hamiltonian analysis is that the optimal
speed is always the maximum value, i.e.,
u∗j (t) = 1
(28)
Hence, we only need to calculate the optimal θj (t). This
TPBVP is computationally expensive and easily becomes
intractable when problem size grows. The ultimate solution
of the TPBVP is a set of agent trajectories that can be
put in a parametric form defined by a parameter vector
θ and then optimized over θ. If the parametric trajectory
family is broad enough, we can recover the true optimal
trajectories; otherwise, we can approximate them within
some acceptable accuracy. Moreover, adopting a parametric
family of trajectories and seeking an optimal one within it has
additional benefits: it allows trajectories to be periodic, often
a desirable property, and it allows one to restrict solutions to
trajectories with desired features that the true optimal may
not have, e.g., smoothness properties to achieve physically
feasible agent motion.
Parameterizing the trajectories and using gradient based
optimization methods, in light of the discussions from the
previous sections, enables us to make use of Infinitesimal
Perturbation Analysis (IPA) [20] to carry out the trajectory
optimization process. We represent each agent’s trajectory
through general parametric equations
sxj (t) = fx (θj , ρj (t)),
syj (t) = fy (θj , ρj (t))
(29)
where the function ρj (t) controls the position of the agent
on its trajectory at time t and θj is a vector of parameters
controlling the shape and location of the trajectory. Let θ =
[θ1 , . . . , θN ]. We now revisit problem P1 in (24):
Z
1 T
(30)
J1 (θ, t)dt
min J(θ, T ) =
θ∈Θ
T 0
and will bring in the equations that were introduced in the
previous section in order to calculate an estimate of dJ(θ)
dθ
as in (10). For this problem due to the continuity of xi (t)
the last two terms in (20) vanish. From (23) we have:
Z τk+1 X
Z τk+1 X
M
M
d
αi xi (θ, t)dt =
αi x0i (θ, t)dt (31)
dθ τk
τk
i=1
i=1
In summary, the evaluation of (31) requires the state
derivatives x0i (t) explicitly and s0j (t) implicitly, (dropping the
dependence on θ for brevity). The latter are easily obtained
for any specific choice of f and g in (29). The former require
a rather laborious use of (15),(17),(18) which, reduces to a
simple set of state derivative dynamics as shown next.
Proposition 1. After an event occurrence at t = τk ,
the state derivatives x0i (τk+ ) with respect to the controllable
parameter θ satisfy the following:
if e(τk ) = ξi0
0
0
−
+
0 +
xi (τk ) =
x0 (τ ) − µil (t)pil (τk )τk if e(τk ) = δij
i0 k−
xi (τk )
otherwise
where l 6= j with pil (τk ) > 0 if such l exists and
0
τk =
∂dij (sj ) 0
∂sj sj
∂dij (sj )
∂sj ṡj (τk )
−1
.
Proof: The proof is omitted due to space limitations,
but it is very similar to the proofs of Propositions 1-3 in
[24].
As is obvious from Proposition 1, the evaluation of x0i (t)
is entirely dependent on the occurrence of events ξi0 and
+
+
δij
in a sample realization, i.e., ξi0 and δij
cause jumps in
this derivative which carry useful information. Otherwise,
x0i (τk+ ) = x0i (τk− ) is in effect and these gradients remain
unchanged. However, we can easily have realizations where
0
no events occur in the system (specifically, events of type δij
+
and δij ) if the trajectory of agents in the sample realization
does not pass through any target. This lack of event excitation
results in the algorithm in (11) to stall.
In the next section we overcome the problem of no event
excitation using the definitions in (6) and (7). We accomplish
this by adding a new metric to the objective function that
generates a non-zero sensitivity with respect to θ.
A. Event Excitation
Our goal here is to select a function hi (·) in (6) with the
property of “spreading” the value of xi (t) over all w ∈ S.
We begin by determining the convex hull produced by the
targets, since the trajectories need not go outside this convex
hull. Let T = {w1 , w2 , · · · , wM } be the set of all target
points. Then, the convex hull of these points is
X
M
X
C=
βi w i |
βi = 1, ∀i, βi ≥ 0
(32)
i=1
i
Given that C ⊂ S, we seek some R(w, t) that satisfies the
following property for constants ci > 0:
Z
M
X
R(w, t)dw =
ci xi (t)
(33)
C
i=1
so that R(w, t) can be viewed as a continuous density defined
for all points w ∈ C which results in a total value equivalent
to a weighted sum of the target states xi (t), i = 1, . . . , M . In
order to select an appropriate h(xi (t), di (w)) in (6), we first
define d+
i (w) = max(kw − wi k, ri ) where ri is the target’s
sensing range. We then define:
M
X
αi xi (t)
(34)
R(w, t) =
d+ (w)
i=1 i
Here, we are spreading a target’s reward (numerator) over
all w so as to obtain the “total weighted reward density” at
w. Note that d+
i (w) = max(kw − wi k, ri ) > 0 to ensure
that the target reward remains positive and fixed for points
w ∈ C(wi ). Moreover, following (7),
N
X
P (w, s(t)) =
ksj (t) − wk2
(35)
j=1
Using these definitions we introduce a new objective function
metric which is added to
the objective function in (24):
hZ
i
J2 (t) = E
P (w, s(t))R(w, t)dw
(36)
C
The expectation is a result of P (w, s(t)) and R(w, t) being
random variables defined on the same probability space as
xi (t).
Proposition 2. For R(w, t) in (34), there exist ci > 0,
i = 1, . . . , M , such that:
Z
M
X
R(w, t)dw =
ci xi (t)
(37)
C
i=1
Proof: We have
Z
Z X
M
αi xi (t)
R(w, t) =
dw
+
C
C i=1 di (w)
(38)
Z
M
X
xi (t)
=
αi
dw
+
C di (w)
i=1
R
(t)
We now need to find the value of C dx+i(w)
for each target
i
i. To do this we first look at the case of one target in a 2D
space and for now we assume C is just a disk with radius Λ
around the target (black circle with radius Λ in Fig. 3). We
can now calculate the above integral for this target using the
polar coordinates:
Z
Z 2π Z Λ
xi (t)
xi (t)
dw =
drdθ
+
max(r
i , r)
0
0
C di (w)
Z 2π Z ri
Z 2π Z Λ
xi (t)
xi (t)
(39)
=
drdθ +
drdθ
ri
r
0
ri
0
0
Λ
= xi (t) 2π 1 + log( )
ri
In our case C is the convex
hull of all targets. We will use
the
idea to calculate the
R xsame
i (t)
dw
for the actual con+
C di (w)
vex hull. We do this for an interior target i.e., a target inside
the convex hull. Extending the
same to targets on the edge is
straightforward. Using the same Fig. 3. One Target R(w, t)
polar coordinate for each θ we Calculation
define Λ(θ) to be the distance of
the target to the edge of C in the direction of θ. (C shown
by a red polygon in Fig. 3).
Z 2π Z Λ
Z
xi (t)
xi (t)
dw
=
drdθ
+
+
0
0 di (r, θ)
C di (w)
Z 2π Z ri
Z 2π Z Λ(θ)
xi (t)
xi (t)
=
drdθ (40)
drdθ +
r
r
i
0
0
0
ri
Z 2π
Λ(θ)
= xi (t) 2π +
log(
)dθ
r
i
0
The second part in (40) has to be calculated knowing Λ(θ)
but since we assumed the target is inside the convex hull we
know Λ(θ) ≥ ri . This means log( Λ(θ)
ri ) > 0 and the xi (t)’s
multiplier is a positive value. We can define ci in (37) as:
Z 2π
Λ(θ)
ci = αi 2π +
)dθ
log(
(41)
r
i
0
The significance of J2 (t) is that it accounts for the
movement of agents through P (w, s(t)) and captures the
target state values through R(w, t). Introducing this term in
the objective function in the following creates a non-zero
gradient even if the agent trajectories are not passing through
any targets. We now combine the two metrics in (24) and
(36) and define problem P2:
Z
1 T
P2 :
min J(T ) =
J1 (t) + J2 (t) dt (42)
T 0
u(t),θ(t)
In this problem, the second term is responsible for adjusting
the trajectories towards the targets by creating a potential
field, while the first term is the original performance metric
which is responsible for adjusting the trajectories so as to
maximize the data collected once an agent is within a target’s
sensing range. It can be easily shown that the results in (28)
hold for problem P2 as well, through the same Hamiltonian
analysis presented in [19]. When sj (t) follows the parametric
functions in (29), the new metric simply becomes a function
of the parameter vector θ and we have:
Z
1 T
J1 (θ, t) + J2 (θ, t) dt
(43)
θ∈Θ
T 0
The new objective function’s derivative follows the same
procedure that was described previously. The first part’s
derivative can be calculated from (31). For the second part
we have:
Z τk+1 Z
min J(θ, T ) =
d
dθ
P (w, θ, t)R(w, θ, t)dw
τk
Z
C
τk+1 Z
=
τk
C
h dP (w, θ, t)
dθ
R(w, θ, t) + P (w, θ, t)
dR(w, θ, t) i
dw
dθ
(44)
In the previous section, we raised the problem of no events
being excited in a sample realization, in which case the total
derivative in (31) is zero and the algorithm in (11) stalls.
Now, looking at (44) we can see that if no events occur
the second part in the integration which involves dR(w,θ,t)
dθ
PM
will be zero, since i=1 x0i (t) = 0 at all t. However, the
first part in the integral does not depend on the events, but
calculates the sensitivity of P (w, s(t)) in (35) with respect
to the parameter θ. Note that the dependence on θ comes
through the parametric description of s(t) through (29). This
term ensures that the algorithm in (11) does not stall and
adjusts trajectories so as to excite the desired events.
V. S IMULATION R ESULTS
We provide some simulation results based on an elliptical parametric description for the trajectories in (29). The
elliptical trajectory formulation is:
sxj (t) = Aj + aj cos ρj (t) cos φj − bj sin ρj (t) sin φj
syj (t) = Bj + aj cos ρj (t) sin φj + bj sin ρj (t) cos φj
(45)
Here, θj = [Aj , Bj , aj , bj , φj ] where Aj , Bj are the coordinates of the center, aj and bj are the major and minor axis
respectively while φj ∈ [0, π) is the ellipse orientation which
is defined as the angle between the x axis and the major axis
of the ellipse. The time-dependent parameter ρj (t) is the
eccentric anomaly of the ellipse. Since an agent is moving
with constant speed of 1 on this trajectory, based on (28),
we have ṡxj (t)2 + ṡyj (t)2 = 1, which gives
h
2
ρ̇j (t) = a sin ρj (t) cos φj + bj cos ρj (t) sin φj
2 i− 21
+ a sin ρj (t) sin φj − bj cos ρj (t) cos φj
(46)
The first case we consider is a problem with one agent and
seven targets located on a circle, as shown in Fig. 4. We
consider a deterministic case with σi (t) = 0.5 for all i. The
other problem parameters are T = 50, µij = 100, ri = 0.2
and αi = 1. A target’s sensing range is denoted with solid
black circles with the target location at the center. The blue
polygon indicates the convex hull produced by the targets.
The direction of motion on a trajectory is shown with the
small arrow. Starting with an initial trajectory shown in light
blue, the on-line trajectory optimization process converges
to the trajectory passing through all targets in an efficient
manner (shown in dark solid blue). In contrast, starting with
this trajectory - which does not pass through any targets
using the event-based IPA calculus to estimate the objective
function gradient.
R EFERENCES
Fig. 4.
One agent and seven target scenario
Fig. 5.
Two agent and seven targets scenario
- problem P1 does not converge and the initial trajectory
remains unchanged. At the final trajectory, J1∗ = 0.0859
and J ∗ = 0.2128. Using the obvious shortest path solution,
the actual optimal value for J1 is 0.0739 that results from
moving on the edges of the convex hull (which allows for
shorter agent travel times).
In the second case, 7 targets are randomly distributed
and two agents are cooperatively collecting the data. The
problem parameters are σi = 0.5, µij = 10, ri = 0.5, αi =
1, T = 50. The initial trajectories for both agents are shown
in light green and blue respectively. We can see that both
agent trajectories converge so as to cover all targets, shown
in dark green and blue ellipses. At the final trajectories,
J1∗ = 0.1004 and J ∗ = 0.2979. Note that we may use these
trajectories to initialize the corresponding TPBVP, another
potential benefit of this approach. This is a much slower
process which ultimately converges to J1∗ = 0.0991 and
J ∗ = 0.2776.
VI. C ONCLUSIONS
We have addressed the issue of event excitation in a
class of multi-agent systems with discrete points of interest.
We proposed a new metric for such systems that spreads
the point-wise values throughout the mission space and
generates a potential field. This metric allows us to use eventdriven trajectory optimization for multi-agent systems. The
methodology is applied to a class of data collection problems
[1] C. G. Cassandras and S. Lafortune, Introduction to Discrete Event
Systems. Secaucus, NJ, USA: Springer-Verlag New York, Inc., 2006.
[2] W. Heemels, J. Sandee, and P. P. J. van den Bosch, “Analysis of eventdriven controllers for linear systems,” International journal of control,
vol. 81, no. 4, pp. 571–590, 2008.
[3] A. Anta and P. Tabuada, “To sample or not to sample: Self-triggered
control for nonlinear systems,” Automatic Control, IEEE Trans. on,
vol. 55, pp. 2030–2042, Sept 2010.
[4] S. Trimpe and R. D’Andrea, “Event-based state estimation with
variance-based triggering,” Automatic Control, IEEE Trans. on,
vol. 59, no. 12, pp. 3266–3281, 2014.
[5] M. Miskowicz, Event-Based Control and Signal Processing. CRC
Press, 2015.
[6] C. G. Cassandras, “The event-driven paradigm for control, communication and optimization,” Journal of Control and Decision, vol. 1,
no. 1, pp. 3–17, 2014.
[7] Y. Khazaeni and C. G. Cassandras, “A new event-driven cooperative receding horizon controller for multi-agent systems in uncertain
environments,” In Proceedings of IEEE 53rd Annual Conference on
Decision and Control, pp. 2770–2775, Dec 2014.
[8] M. Zhong and C. G. Cassandras, “Asynchronous distributed optimization with event-driven communication,” Automatic Control, IEEE
Trans. on, vol. 55, no. 12, pp. 2735–2750, 2010.
[9] M. Schwager, D. Rus, and J.-J. Slotine, “Decentralized, adaptive
coverage control for networked robots,” The International Journal of
Robotics Research, vol. 28, no. 3, pp. 357–375, 2009.
[10] C. G. Cassandras, X. Lin, and X. Ding, “An optimal control approach
to the multi-agent persistent monitoring problem,” IEEE Trans. on Aut.
Cont., vol. 58, pp. 947–961, April 2013.
[11] M. Cao, A. Morse, C. Yu, B. Anderson, and S. Dasgupta, “Maintaining
a directed, triangular formation of mobile autonomous agents,” Communications in Information and Systems, vol. 11, no. 1, p. 1, 2011.
[12] K.-K. Oh and H.-S. Ahn, “Formation control and network localization
via orientation alignment,” IEEE Trans. on Automatic Control, vol. 59,
pp. 540–545, Feb 2014.
[13] H. Yamaguchi and T. Arai, “Distributed and autonomous control
method for generating shape of multiple mobile robot group,” in Proc.
of the IEEE International Conf. on Intelligent Robots and Systems,
vol. 2, pp. 800–807 vol.2, Sep 1994.
[14] J. Desai, V. Kumar, and J. Ostrowski, “Control of changes in formation
for a team of mobile robots,” in Proc. of the IEEE International Conf.
on Robotics and Automation, vol. 2, pp. 1556–1561, 1999.
[15] M. Ji and M. B. Egerstedt, “Distributed coordination control of
multi-agent systems while preserving connectedness.,” IEEE Trans.
on Robotics, vol. 23, no. 4, pp. 693–703, 2007.
[16] J. Wang and M. Xin, “Integrated optimal formation control of multiple
unmanned aerial vehicles,” IEEE Trans. on Control Systems Technology, vol. 21, pp. 1731–1744, Sept 2013.
[17] D. L. Applegate, R. E. Bixby, V. Chvatal, and W. J. Cook, The traveling salesman problem: a computational study. Princeton University
Press, 2011.
[18] X. Lin and C. G. Cassandras, “An optimal control approach to the
multi-agent persistent monitoring problem in two-dimensional spaces,”
IEEE Trans. on Automatic Control, vol. 60, pp. 1659–1664, June 2015.
[19] Y. Khazaeni and C. G. Cassandras, “An optimal control approach for
the data harvesting problem,” in 54th IEEE Conf. on Decision and
Cont., pp. 5136–5141, 2015.
[20] C. G. Cassandras, Y. Wardi, C. G. Panayiotou, and C. Yao, “Perturbation analysis and optimization of stochastic hybrid systems,” European
Journal of Cont., vol. 16, no. 6, pp. 642 – 661, 2010.
[21] H. Kushner and G. Yin, Stochastic Approximation and Recursive
Algorithms and Applications. Springer, 2003.
[22] J. L. Ny, M. a. Dahleh, E. Feron, and E. Frazzoli, “Continuous path
planning for a data harvesting mobile server,” Proc. of the IEEE Conf.
on Decision and Cont., pp. 1489–1494, 2008.
[23] A. E. Bryson and Y. C. Ho, Applied optimal control: optimization,
estimation and control. CRC Press, 1975.
[24] Y. Khazaeni and C. G. Cassandras, “An optimal control approach for
the data harvesting problem,” arXiv:1503.06133.
| 3cs.SY
|
WEAKLY AMENABLE GROUPS
arXiv:1609.03634v1 [math.GR] 12 Sep 2016
DENIS V. OSIN
Abstract. We construct the first examples of finitely generated non–amenable
groups whose left regular representations are not uniformly isolated from the trivial
representation.
1. Introduction.
Recall that a locally compact group G is called amenable if there exists a finitely
additive measure µ on the set of all Borel subsets of G which is invariant under the left
action of the group G on itself and satisfies µ(G) = 1. The class of amenable groups,
AG, has been introduced by von Neumann [16] in order to explain the Hausdorff–
Banach–Tarski paradox and was investigated by a number of authors.
One of the most interesting characterizations of amenable groups was obtained by
Hulaniski [11] in terms of L2 –representations.
Definition 1. One says that the left regular representation LG of a locally compact
group G on the Hilbert space L2 (G) weakly contains the trivial representation, if for
any ε > 0 and any compact subset S ⊆ G, there exists v ∈ L2 (G) such that kvk = 1
and
(1)
for any s ∈ S.
|hv, svi − 1| < ε
Theorem 2 (Hulaniski). A locally compact group G is amenable if and only if the
left regular representation of G weakly contains the trivial representation.
Given a locally compact group G and a compact subset S ⊆ G, we define α(G, S)
as the supremum of all ε ≥ 0 such that for any vector v ∈ L2 (G) of norm ||v|| = 1,
there exists an element s ∈ S satisfying the inequality
||sv − v|| ≥ ε.
2000 Mathematics Subject Classification. Primary 20F05; Secondary 20F5, 22D10.
Key words and phrases. Left regular representation, amenable group, finitely generated group,
hyperbolic group, torsion group.
The work has been supported by the RFFR grant 99-01-00894 and by the Swiss National Science
Foundation.
1
2
DENIS V. OSIN
In case the group G is discrete and finitely generated, the existence of a finite generating set S such that α(G, S) > 0, implies the inequality α(G, S ′ ) > 0 for any other
generating set S ′ of G. Thus it is natural to consider the quantity
α(G) = inf α(G, S),
S
where S ranges over all finite generating sets of G. The following definition can be
found in [22]
Definition 3. The left regular representation of a finitely generated group G is said
to be uniformly isolated from the trivial representation if α(G) > 0.
Obviously one has
(2)
α(G) = 0
for any finitely
√ generated amenable group. Indeed, it is easy to check that (1) implies
ksv − vk < 2ε. Thus (2) follows from Theorem 1.2. On the other hand, it is not
clear whether the equality (2) is equivalent to the amenability of the group G. The
following problem was suggested by Shalom in [22].
Problem 4. Is the left regular representation of any non–amenable finitely generated
group uniformly isolated from the trivial representation?
In [22], the positive answer was obtained in the particular case of residually finite
hyperbolic groups. However, the question remained open in general. The main purpose of the present note is to show that the answer is negative and can be obtained
by using the methods developed in [20]
The main part of this paper was written during the author’s visit to University of
Geneva. I am grateful to Pierre de la Harpe for invitation and constant attention
to this work. Also I would like to express my gratitude to Rostislav I. Grigorchuk,
Anna G. Erschler, Alexander Yu. Ol’shanskii, and the referee for useful comments
and remarks.
2. Main results
The main results of the paper are gathered in this section. We call a finitely
generated group G weakly amenable if it satisfies (2) and denote by W A the class of
all weakly amenable groups.
Two families of non–amenable weakly amenable groups are constructed in the
present paper. The idea of the first construction is similar to one from [5].
Theorem 5. Let A be a finitely generated abelian group. Suppose that there exist two
monomorphisms λ, µ : A → A with the following properties.
WEAKLY AMENABLE GROUPS
3
1) λ ◦ µ ≡ µ ◦ λ.
2) The subgroup generated by λ(A) ∪ µ(A) coincides with A.
3) λ(A) ∪ µ(A) 6= A.
Then the HNN–extension
(3)
G = hA, t : t−1 λ(a)t = µ(a), a ∈ Ai
is a finitely generated weakly amenable non–amenable group.
Example 6. Suppose that A = Z and λ, µ are defined by λ(1) = m, µ(1) = n. If
m, n are relatively prime, and |m| =
6 1, |n| =
6 1, one can easily verify the conditions of
Theorem 2.1. Taking the HNN–extension, we obtain the Baumslag–Solitar group
BS(m, n) = ha, t : t−1 am t = an i.
Using the Britton lemma on HNN–extensions [15, Ch. IV, Sec. 2], one can prove
that the elements t and a−1 ta generates a free subgroup of rank 2. This shows that
the class W A is not closed under the taking of subgroups.
In the last section of the present paper we give another way to construct a weakly
amenable non–amenable group using limits of hyperbolic groups. The proof involves
the tools of hyperbolic group theory developed in [18] and certain results from [20].
Recall that a locally compact group G is said to have property (T) of Kazhdan if the
one–dimensional trivial representation is an isolated point of the set of all irreducible
unitary representations of G endowed with the Fell topology (we refer to [12], [14]
and [9] for more details). It follows easily from the definition and Hulaniski’s theorem
that every discrete finitely generated amenable group having property (T) is finite.
In contrast, we obtain the following unexpected result in the case of weakly amenable
groups.
Theorem 7. There exists a 2–generated infinite periodic weakly amenable group Q
having property (T) of Kazhdan. In particular, Q is non–amenable.
We also consider a variant of Day’s question which goes back to the papers [2],
[16] and known as the so called ”von Neumann problem”. Let NF denote the class
of all groups containing no non–abelian free subgroups, and AG denote the class of
all amenable groups. Obviously AG ⊆ NF since any non–abelian free group is non–
amenable and the class AG is closed under the taking of subgroups [16]. The question
is whether NF = AG.
Ol’shanskii [17] shown that certain groups constructed by him earlier (torsion
groups with unbounded orders of elements in which all proper subgroups are cyclic)
are non–amenable and thus the answer is negative. Further, in [1] Adian proved
that the free Burnside groups B(m, n) of sufficiently large odd exponent n and rank
4
DENIS V. OSIN
m > 1 are non–amenable. It is a natural stronger version of Day’s question, whether
the inclusion
W A ∩ NF ⊂ AG
is true. We note that all groups constructed in Theorem 2.1 contain non–abelian free
subgroups (see Lemma 3.10 below). Furthermore, B(m, n) ∈
/ W A for any m > 1 and
any n odd and large enough, as follows from the main result of [21]. Thus these groups
do not provide an answer. On the other hand the negative answer is an immediate
consequence of Theorem 2.3.
Corollary 8. There exists a finitely generated weakly amenable non–amenable group
which contains no non–abelian free subgroups.
In conclusion we note that our construction of the group Q from Theorem 2.3 is
closely related to the question whether any finitely generated group of exponential
growth is of uniform exponential growth (see Section 4 for definitions). Originally,
this problem was formulated in [8] and studied intensively during the last few years
(we refer to [10] for survey). In [13], Koubi proved that the exponential growth rate
ω(G) of every non–elementary hyperbolic group G satisfies the inequality ω(G) > 1.
On the other hand, the following question is still open.
Problem 9. Is the set
ΩH = {ω(G) : G is non − elementary hyperbolic}
bounded away from the identity?
In Section 4, we observe that the negative answer would imply the existence of a
finitely generated group having non–uniform exponential growth.
3. Non–Hopfian weakly amenable groups
Let Fm be the free group of rank m, X = {x1 , x2 , . . . , xm } a free generating set of
Fm . We begin this section by describing the Grigorchuk’s construction of a topology
on Gm , the set of all normal subgroups of Fm (or, equivalently, on the set of all group
presentations with the same generating set).
Definition 10. The Cayley graph Γ = Γ(G, S) of a group G generated by a set S
is an oriented labeled 1–complex with the vertex set V (Γ) = G and the edge set
E(Γ) = G × S. An edge e = (g, s) ∈ E(Γ) goes from the vertex g to the vertex gs
and has the label φ(e) = s. As usual, we denote the origin and the terminus of the
edge e, i.e., the vertices g and gs, by α(e) and ω(e) respectively. One can endow
the group G (and, therefore, the vertex set of Γ) with a length function by assuming
kgkS , the length of an element g ∈ G, to be equal to the length of a shortest word in
the alphabet S ∪ S −1 representing g.
WEAKLY AMENABLE GROUPS
5
Let N ∈ Gm . To simplify our notation we will identify the set X with the generating
set of the quotient group Fm /N naturally obtained from X. Now let N1 , N2 be two
normal subgroups of Fm and G1 = Fm /N1 , G2 = Fm /N2 . By Bi (r), i = 1, 2, we
denote the ball of radius r around the identity in the Cayley graph Γi = Γ(Gi , X),
i.e., the oriented labeled subgraph with the vertex set
and the edge set
V (Bi (r)) = {g ∈ Gi : kgkXi ≤ r}
E(Bi (r)) = {e ∈ E(Γi ) : α(e) ∈ V (Bi (r)) and ω(e) ∈ V (Bi (r))}.
One says that the groups G1 and G2 are locally r–isomorphic (being considered
quotients of Fm ) and writes G1 ∼r G2 if there exists a graph isomorphism
ι : B1 (r) → B2 (r)
that preserves labels and orientation.
Definition 11. For every N ∈ Gm and r ∈ N, we consider the set
Wr (N) = {L ∈ Gm : Fm /N ∼r Fm /L}.
One defines the topology on Gm by taking the collection of the sets Wr (N) as the
base of neighborhoods.
Example 12. Suppose that {Ni } is a sequence of normal subgroups of Fm such that
∞
T
N1 ≥ N2 ≥ . . .. Then the limit of the sequence coincides with
Ni . Symmetrically
if N1 ≤ N2 ≤ . . ., then the limit is the union
is left as an exercise to the reader.
∞
S
i=1
Ni . The proof is straightforward and
i=1
We need the following result, which is proved in [20] (up to notation).
Theorem 13. Suppose that {Ni }i∈N is a sequence of elements of Gm which converges
to an element N ∈ Gm . If the group G = Fm /N is amenable, then
(4)
lim α(Fm /Ni , X) = 0.
i→∞
Remark 14. Let AG m denote the subset of all elements N ∈ Gm such that the quotient
group Fm /N is amenable. Essentially the theorem says that the map α : Gm →
[0, +∞) which takes each N ∈ Gm to α(Fm /N, X) is continuous at any point N ∈
AG m . It is not hard to see that α is not continuous at arbitrary point of Gm . Indeed,
consider the sequence of subgroups N1 ≥ N2 ≥ . . . of finite index in Fm such that
(5)
∞
\
i=1
Ni = {1}
6
DENIS V. OSIN
(such a sequence exists since any free group is residually finite). One can easily check
that (5) implies
lim Ni = {1}
i→∞
(see Example 3.3). Since the group Fm is non–amenable whenever m > 1, we have
α({1}) > 0. However, α(Fm /Ni , X) = 0 for any i, as the quotient groups Fm /Ni are
finite (and, therefore, amenable).
Now suppose that G is the group defined by (3). The following four lemmas are
proved under the assumptions of Theorem 2.1. Consider the homomorphism φ : G →
G induced by φ(t) = t and φ(a) = λ(a) for every a ∈ A.
Lemma 15. The homomorphism φ is well–defined.
Proof. We have to check that for any relation R = 1 of the group G one has φ(R) = 1
in G. There are two possibilities for R.
1) First suppose that R = 1 is a relation of the group A. Since the restriction of φ
to A coincides with the monomorphism λ, we have φ(R) = λ(R) = 1.
2) Assume that R has the form (λ(a))t (µ(a))−1 . Taking into account the first
condition of Theorem 2.1, we obtain
φ (λ(a))t (µ(a))−1 = (λ ◦ λ(a))t (λ ◦ µ(a))−1 = µ ◦ λ(a)(µ ◦ λ(a))−1 = 1.
Lemma 16. The map φ is surjective.
Proof. Observe that G is generated by t and A. As t ∈ φ(G), it suffices to prove that
A ≤ φ(G). Clearly we have λ(A) = φ(A) ∈ φ(G) and µ(A) = (λ(A))t ∈ φ(G). It
remains to refer to the second condition of Theorem 2.1.
Let us denote by φi the i–th power of φ and by Ni its kernel. Put N =
∞
S
Ni .
i=1
Obviously the group G = G/N is generated by the images of a and t under the natural
homomorphism G → G. To simplify our notation we will denote these images by a
and t as well.
Lemma 17. The group G is an extension of an abelian group by a cyclic one.
Proof. We denote by B the kernel of the natural homomorphism G → hti. Let us
i
show that B is abelian. It is clear that B is generated by the set {at : a ∈ A, i ∈ Z}.
i
j
Therefore, it is sufficient to show that [at , at ] = 1 for any a ∈ A, i, j ∈ Z. Without
WEAKLY AMENABLE GROUPS
7
loss of generality we can assume that i ≥ j. Moreover, conjugating by a suitable
power of t, we can assume j = 0. In these settings, we have
i
i
φi ([at , a]) = [(λi (a))t , λi (a)] = [µi (a), λi (a)] = 1
i
as A is abelian. Therefore, the element [at , a] belongs to Ni and thus its image in G
is trivial.
We note that in certain particular cases (including, for example, non–Hopfian
Baumslag–Solitar groups) Lemma 3.8 follows from a result of Hirshon [6]. As any
abelian group is amenable and the class of amenable groups is closed under group
extensions, Lemma 3.8 yields
Corollary 18. The group G is amenable.
Lemma 19. The group G contains a non–abelian free subgroup.
Proof. According to the third condition of Theorem 2.1 there exists an element a ∈
A \ (λ(A) ∪ µ(A)). The elements t and a−1 ta generate the free group of rank 2 by the
Britton lemma on HNN–extensions.
Proof of Theorem 2.1. Let us note that the sequence {Ni } converges to N. Applying
Corollary 3.9 and Theorem 3.4, we obtain lim α(F/Ni , X) = 0. On the other hand,
i→∞
F/Ni ∼
= G, this means that α(G) = 0, i.e., G is weakly amenable. Finally, G is
non–amenable according to Lemma 3.10.
4. Common quotient groups of all non–elementary hyperbolic
groups.
Let us recall just one of a number of equivalent definitions of hyperbolicity. A
group G with a finite generating set X is hyperbolic (in the sense of Gromov) if its
Cayley graph Γ = Γ(G, X) is a hyperbolic space with respect to the natural metric.
This means that any geodesic triangle in Γ is δ–thin for a fixed constant δ, i.e., each
of its sides belongs to the closed δ–neighborhood of the union of other two sides.
It has been mentioned by Gromov [7] (see also [9]), that an element g of infinite
order in a hyperbolic group G is contained in a unique maximal elementary subgroup
EG (g) (elementary closure of g). For a subgroup H of a hyperbolic group G, its
elementarizer EG (H) is defined as ∩EG (h), where h ranges over all elements of infinite
order in H. If the subgroup H is non–elementary, EG (H) is the unique maximal finite
subgroup of G normalized by H [18, Proposition 1]; notice also that EG (G) is the
kernel of the action of G on the hyperbolic boundary ∂G induced by left multiplication
on G.
8
DENIS V. OSIN
The following is the simplification of Theorem 2 from [18] (see also [19, Lemma
5.1]).
Lemma 20. Let H1 , . . . , Hk be non–elementary subgroups of a hyperbolic group G
such that EG (H1 ) = . . . = EG (Hk ) = 1. Then there exists a non–elementary hyperbolic quotient K of G such that the image of each subgroup H1 , . . . , Hk under the
natural epimorphism G → K coincides with K.
Corollary 21. Let P1 , . . . , Pk be non–elementary hyperbolic groups. Then there exists
a non–elementary hyperbolic group Q that is a homomorphic image of Pi for every
i = 1, . . . , k.
Proof. The proof can be extracted from the one of Theorem 2 in [19]. Here we provide
it for convenience of the reader. Let us set Hi = Pi /EPi (Pi ). Clearly EHi (Hi ) = 1, as
EPi (Pi ) is the maximal normal finite subgroup of Pi . Moreover, since any quotient of
a non–elementary hyperbolic group modulo a finite normal subgroup is also a non–
elementary hyperbolic group [4, Corollary 23(ii)], it follows that Hi is non–elementary
hyperbolic. Now we take the free product
G = H1 ∗ . . . ∗ Hk .
It is easy to check that EG (Hi ) = 1 for every i as there are no finite subgroups of G
normalized by Hi . It remains to apply Lemma 4.1.
We need one more lemma (the proof can be found in [18]).
Lemma 22. Let G be a non–elementary hyperbolic group, g an element of G. Then
there exists N ∈ N such that the quotient group of G modulo the normal closure of
g N is non–elementary and hyperbolic.
Now we are going to describe the main construction of the present section.
Theorem 23. There exists a 2–generated infinite periodic group Q such that for every
non–elementary hyperbolic group H, there is an epimorphism ρ : H → Q.
Proof. Since any hyperbolic group is finitely presented, the set of all non–elementary
hyperbolic groups is countable. Let us enumerate this set G1 , G2 , . . . and elements
of the first group G1 = {g1 , g2 , . . .}. Consider the following diagram, which is constructed by induction.
G
G
...
G
...
k
2
1
π
π2
π1
y k
y
y
ψ1
φ2
ψ2
φ
ψ
φk+1
k
k
Qk −→
Rk −→ . . .
Q1 −→ R1 −→ Q2 −→ . . . Rk−1 −→
Suppose G1 = Q1 and let π1 denote the corresponding natural isomorphism. Assume
that we have already defined the groups Qi , Ri−1 and homomorphisms φi : Ri−1 →
WEAKLY AMENABLE GROUPS
9
Qi , ψi−1 : Qi−1 → Ri−1 for all i ≤ k. Denote by τk : G1 → Qk the composition
φk ψk−1 . . . φ2 ψ1 π1 and by ḡi the image of gi in Qk under τk . According to Lemma 4.3,
there exists Ni ∈ N such that the quotient Qk /hḡiNi iQk is a non–elementary hyperbolic
group. We set
Rk = Qk /hḡiNi iQk
and denote by ψk the natural homomorphism from Qk to Rk . Further, by Corollary
4.2, there is a non–elementary hyperbolic group Qk+1 such that there exist epimorphisms
φk+1 : Rk → Qk+1 and πk+1 : Gk+1 → Qk+1 .
The inductive step is completed.
Let us denote by Uk the kernel of τk . Evidently we have {1} = U1 ≤ U2 ≤ . . .. Set
∞
S
Ui and consider the quotient group Q = G1 /U. Note that one can assume G1
U=
i=1
to be 2–generated without loss of generality. Further, Q is a quotient group of Qi for
all i, hence Q is a quotient of Gi for all i. The periodicity of Q follows directly from
our construction. It remains to show that Q is infinite. To do this, let us suppose
that Q is finite. Then Q is finitely presented. Therefore, Qi is a quotient group of Q
for all i big enough. In particular, Qi is elementary whenever i is sufficiently big and
we get a contradiction. The theorem is proved.
Let us denote by Hm the subset of all N ∈ Gm such that Fm /N is non-elementary
and hyperbolic. Recall also that AG m denotes the subset of all N ∈ Gm such that
Fm /N is amenable. The following two observations from [20] plays the crucial role in
the studying of the group Q.
Theorem 24. For every m ≥ 2, the intersection of the closure of Hm (with respect
to the Cayley topology on Gm ) and AG m is non–empty.
Lemma 25. Suppose that G is a finitely generated group and φ : G → P is a
surjective homomorphism onto a group P . Then α(G) ≥ α(P ).
Now we want to show that the group Q from Theorem 4.4 has all properties listed
at Theorem 2.3.
Proof of Theorem 2.3. . By Theorem 3.4 and Theorem 4.5, there is a sequence of
elements Ni ∈ H2 , i ∈ N, such that
(6)
lim α(F2 /Ni ) = 0.
i→∞
Let us denote by Gi the quotient group F2 /Ni . According to Theorem 4.4, there
exists an epimorphism ρi : Gi → Q for every Gi . Combining Lemma 4.6 and (6),
we obtain α(Q) = 0. As is well known, there are non–elementary hyperbolic groups
having property (T ) of Kazhdan (for instance, uniform lattices in Sp(n, 1)). Since
10
DENIS V. OSIN
the class of Kazhdan groups is closed under the taking of quotients, the group Q has
the property T . Recall that any discrete amenable Kazhdan group is finite; taking
into account the infiniteness of Q, we conclude that Q is non–amenable.
In conclusion we discuss certain relations with growth functions of hyperbolic
X
groups. The growth function γG
: N −→ N of a group G generated by a finite
set X is defined by the formula
X
γG
(n) = card {g ∈ G : ||g||X ≤ n},
where ||g||X denotes the word length of g relative to X. The exponential growth rate
of G with respect to X is the number
q
X
ω(G, X) = lim n γG
(n).
n→∞
X
The above limit exists by submultiplicativity of γG
. The group G is said to be of exponential growth (respectively of subexponential growth ) if ω(G, X) > 1 (respectively
ω(G, X) = 1) for some generating set X.
It is easy to see that above definitions are independent of the choice of a generating
set in G. Let us consider the quantity
ω(G) = inf ω(G, X),
X
where the infimum is taken over all finite generating sets of G. One says that G has
uniform exponential growth if
(7)
ω(G) > 1.
It is an open question whether any group of exponential growth satisfies (7). We
observe that Theorem 4.4 provides an approach to the solution of this problem.
Lemma 26. Let G be a group generated by a finite set X and φ : G → P be an
epimorphism. Then ω(G, X) ≥ ω(P, φ(X)).
Proof. This observation is well known and quite trivial. The proof follows easily from
the inequality kgkX ≥ kφ(g)kφ(X) . We leave details to the reader.
Obviously Lemma 4.7 and Theorem 4.2 yield the following.
Corollary 27. Suppose that for every ε > 0, there exists a non–elementary hyperbolic
group H such that ω(H) < 1 + ε. Then the group Q from Theorem 4.1 has non–
uniform exponential growth, i.e., ω(Q, X) > 1 for any finite generating set X of Q
but ω(Q) = 1.
WEAKLY AMENABLE GROUPS
11
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
S.I. Adian, Random walks on free periodic groups, Math. USSR Izvestiya, 21 (1983), 425–434
M.M. Day, Amenable semigroups, Illinois J. Math., 1 (1957), 509–544.
E. Følner, On groups with full Banach mean value, Math. Scand., 3 (1955), 243–254.
E. Ghys, P. de la Harpe (Eds.), Sur les gruppes hyperboliques d’apres Mikhael Gromov, in:
Swiss Seminar on Hyperbolic groups, Berne, 1988, Birkhäuser, 1990.
R.I. Grigorchuk, M.J. Mamaghani, On use of iterates of endomorphisms for constructing groups
with specific properties, preprint, Univ. of Teheran, 1996
R. Hirshon, The intersection of the subgroups of finite index in some finitely presented groups,
Proc. Amer. Math. Soc., 53 (1975), 1, 32–36.
M. Gromov, Hyperbolic groups, in: Essays in group Theory (S.M. Gersted, Ed.), MSRI Publ.,
8, 75–263, Springer–Verlag, 1987
M. Gromov, Metric structures for Riemannian and non–Riemannian spaces, Progress in Math.,
152, Birkhäuser Verlag, 1998
P. de la Harpe, A. Valette, La propriété (T) de Kazhdan pour les groupes localement compacts,
Asterisque, 175, Société Mathematique de France, 1989.
P. de la Harpe, Topics in geometric group theory, Univ. of Chicago Press, 2000.
A. Hulaniski, Means and Følner conditions on locally compact groups, Studia Math., 27 (1966),
87–104.
D.A. Kazhdan, On the connection of the dual space of a group with the structure of its closed
subgroups (Russian), Funkcional. Anal. i Priložen., 1 (1967), 71–74.
M. Koubi, Croissance uniforme dans les groupes hyperboliques, Ann. Inst. Fourier, 48 (1998),
1441–1453.
A. Lubotzky, Discrete groups, expanding graphs and invariant measures, Progress in Math.,
125, Birkhäuser Verlag, 1994.
R.C. Lyndon, P.E. Shupp, Combinatorial Group Theory, Ergebnisse der Mathematik und ihrer
Grenzgebiete 89, Springer–Verlag, 1977
J. von Neumann, Zur allgemeinen Theorie des Masses, Fund. Math., 13 (1929), 73–116.
A.Yu. Ol’shanskii, On the problem of the existence of an invariant mean on a group, Russian
Math. Surveys, 35 (1980), 4, 180–181.
A.Yu. Ol’shanskii, On residualing homomorphisms and G–subgroups of hyperbolic groups, Int.
J. Alg. Comput., 3 (1993), 365–409.
A.Yu. Ol’shanskii, On the Bass–Lubotzkii question about quotients of hyperbolic groups, J. Alg.,
226 (2000), 807–817.
D.V. Osin, Kazhdan constants of hyperbolic groups, Funct. Anal. Appl., 36 (2002), no. 4, 290–
297.
D.V. Osin, Uniform non–amenability of free Burnside groups, Arch. Math., 88 (2007), no. 5,
403–412.
Y. Shalom, Explicit Kazhdan constant for representations of semisimple and arithmetic groups,
Ann. Inst. Fourier (Grenoble), 50 (2000), 3, 833-863.
E-mail address: [email protected]
| 4math.GR
|
arXiv:1609.01361v1 [cs.DS] 6 Sep 2016
Fourier-sparse interpolation without a frequency gap
Xue Chen∗
[email protected]
The University of Texas at Austin
Daniel M. Kane
[email protected]
University of California, San Diego
Eric Price
[email protected]
The University of Texas at Austin
Zhao Song
[email protected]
The University of Texas at Austin
September 7, 2016
Abstract
We consider the problem of estimating a Fourier-sparse signal from noisy samples, where
the sampling is done over some interval [0, T ] and the frequencies can be “off-grid”. Previous
methods for this problem required the gap between frequencies to be above 1/T , the threshold
required to robustly identify individual frequencies. We show the frequency gap is not necessary
to estimate the signal as a whole: for arbitrary k-Fourier-sparse signals under `2 bounded noise,
we show how to estimate the signal with a constant factor growth of the noise and sample
complexity polynomial in k and logarithmic in the bandwidth and signal-to-noise ratio.
As a special case, we get an algorithm to interpolate degree d polynomials from noisy measurements, using O(d) samples and increasing the noise by a constant factor in `2 .
∗
Supported by NSF Grant CCF-1526952.
Contents
1 Introduction
1.1 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Our techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
4
6
9
2 Proof Sketch
9
3 Preliminaries
3.1 Notation . . . . . . . . . . . . . . .
3.2 Facts about the Fourier transform .
3.3 Tools and inequalities . . . . . . .
3.4 Legendre polynomials . . . . . . .
3.5 Gram matrix and its determinant .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
12
12
13
14
15
15
4 Robust Polynomial Interpolation Algorithm
16
4.1 Constant success probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2 Boosting success probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5 Bounding the Magnitude of a Fourier-sparse Signal in Terms of Its Average Norm 20
5.1 Bounding the maximum inside the interval . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2 Bounding growth outside the interval . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
6 Hash Functions and Filter Functions
23
6.1 Permutation function and hash function . . . . . . . . . . . . . . . . . . . . . . . . . 23
6.2 Filter function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6.3 HashToBins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
7 Frequency Recovery
7.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Analysis of GetLegal1Sample and GetEmpirical1Energy
. . . . . . .
7.3 A cluster of frequencies, times H, is a one-cluster signal per Definition 7.1 . . .
7.4 Frequency recovery of one-cluster signals . . . . . . . . . . . . . . . . . . . . . .
7.5 The full signal, after multiplying by H and convolving with G, is one-clustered.
7.6 Frequency recovery of k-clustered signals . . . . . . . . . . . . . . . . . . . . . .
7.7 Time and sample complexity of frequency recovery of k-clustered signals . . . .
8 One-cluster Signal Recovery
8.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2 Bounding the Gram matrix determinant . . . . . . . . . . . . . . . . . .
8.3 Perturbing the frequencies does not change the subspace much . . . . .
8.4 Existence of nearby k-Fourier-sparse signal with frequency gap bounded
zero . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.5 Approximating k-Fourier-sparse signals by polynomials . . . . . . . . . .
8.6 Transferring degree-d polynomial to (d+1)-Fourier-sparse signal . . . . .
1
. . .
. . .
. . .
away
. . .
. . .
. . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . .
. . . .
. . . .
from
. . . .
. . . .
. . . .
25
26
27
31
32
35
39
40
41
41
42
43
45
46
47
9 k-cluster Signal Recovery
9.1 Overview . . . . . . . . . . . . . . . . . . . .
9.2 Heavy clusters separation . . . . . . . . . . .
9.3 Approximating clusters by polynomials . . . .
9.4 Main result, with constant success probability
9.5 Boosting the success probability . . . . . . . .
A Technical Proofs
A.1 Proof of Theorem 8.3 . . . . . . . . .
A.2 Proofs of Lemma 5.3 and Lemma 5.4
A.3 Proof of Lemma 4.3 . . . . . . . . .
A.4 Proof of Lemma 6.2 . . . . . . . . .
A.5 Proof of Lemma 3.5 . . . . . . . . .
A.6 Proof of Lemma 3.10 . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
49
49
50
51
52
54
.
.
.
.
.
.
58
58
61
64
64
65
66
B Known Facts
67
B.1 Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
B.2 Linear regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
B.3 Multipoint evaluation of a polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . 68
C Analysis of Hash Functions and Filter Functions
b )) . . . . . . .
C.1 Analysis of filter function (H(t), H(f
b
C.2 Analysis of filter function (G(t), G(f )) . . . . . . .
C.3 Parameters setting for filters . . . . . . . . . . . . .
C.4 Analysis of HashToBins . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
68
68
78
79
80
D Acknowledgments
84
E Algorithm
84
2
1
Introduction
In an interpolation problem, one can observe x(t) = x∗ (t) + g(t), where x∗ (t) is a structured signal
and g(t) denotes noise, at points ti of one’s choice in some interval [0, T ]. The goal is to recover
an estimate x
e of x∗ (or of x). Because we can sample over a particular interval, we would like our
approximation to be good on that interval, so for any function y(t) we define
kyk2T =
1
T
Z
T
0
|y(t)|2 dt.
to be the `2 error on the sample interval. For some parameters C and δ, we would then like to get
ke
x − x∗ kT ≤ C kgkT + δ kx∗ kT
(1)
while minimizing the number of samples and running time. Typically, we would like C to be O(1)
and to have δ be very small (either zero, or exponentially small). Note that, if we do not care about
changing C by O(1), then by the triangle inequality it doesn’t matter whether we want to estimate
x∗ or x (i.e. we could replace the LHS of (1) by ke
x − xkT ).
Of course, to solve an interpolation problem one also needs x∗ to have structure. One common
form of structure is that x∗ have a sparse Fourier representation. We say that a function x∗ is
k-Fourier-sparse if it can be expressed as a sum of k complex exponentials:
∗
x (t) =
k
X
vj e2πifj t .
j=1
for some vj ∈ C and fj ∈ [−F, F ], where F is the “bandlimit”. Given F , T , and k, how many
samples must we take for the interpolation (1)?
If we ignore sparsity and just use the bandlimit, then Nyquist sampling and Shannon-Whittaker
interpolation uses F T + 1/δ samples to achieve (1). Alternatively, in the absence of noise, x∗
can be found from O(k) samples by a variety of methods, including Prony’s method from 1795 or
Reed-Solomon syndrome decoding [Mas69], but these methods are not robust to noise.
If the signal is periodic with period T —i.e., the frequencies are multiples of 1/T —then we can
use sparse discrete Fourier transform methods, which take O(k logc (F T /δ)) time and samples (e.g.
[GGI+ 02, HIKP12a, IKP14]). If the frequencies are not multiples of 1/T (are “off the grid”), then
the discrete approximation is only k/δ sparse, making the interpolation less efficient; and even this
requires that the frequencies be well separated.
A variety of algorithms have been designed to recover off-grid frequencies directly, but they
require the minimum gap among the frequencies to be above some threshold. With frequency gap
at least 1/T , we can achieve a k c approximation factor using O(F T ) samples [Moi15], and with
gap above O(log2 k)/T we can get a constant approximation using O(k logc (F T /δ)) samples and
time [PS15].
Having a dependence on the frequency gap is natural. If two frequencies are very close together—
significantly below 1/T —then the corresponding complex exponentials will be close on [0, T ], and
hard to distinguish in the presence of noise. In fact, from a lower bound in [Moi15], below 1/T
frequency gap one cannot recover the frequencies in the presence of noise as small as 2−Ω(k) . The
lower bound proceeds by constructing two signals using significantly different frequencies that are
exponentially close over [0, T ].
But if two signals are so close, do we need to distinguish them? Such a lower bound doesn’t
apply to the interpolation problem, it just says that you can’t solve it by finding the frequencies.
3
Our question becomes: can we benefit from Fourier sparsity in a regime where we can’t recover the
individual frequencies?
We answer in the affirmative, giving an algorithm for the interpolation using O(poly(k log(F T /δ))
samples. Our main theorem is the following:
Theorem 1.1. Let x(t) = x∗ (t) + g(t), where x∗ is k-Fourier-sparse signal with frequencies in
[−F, F ]. Given samples of x over [0, T ] we can output x
e(t) such that with probability at least
−Ω(k)
1−2
,
ke
x − x∗ kT . kgkT + δ kx∗ kT .
Our algorithm uses poly(k, log(1/δ)) · log(F T ) samples and poly(k, log(1/δ)) · log2 (F T ) time. The
output x
e is poly(k, log(1/δ))-Fourier-sparse signal.
Relative to previous work, this result avoids the need for a frequency gap, but loses a polynomial
factor in the sample complexity and time. We lose polynomial factors in a number of places; some
of these are for ease of exposition, but others are challenging to avoid.
Degree d polynomials are the special case of d-Fourier-sparse functions in the limit of fj → 0, by
a Taylor expansion. This is a regime with no frequency gap, so previous sparse Fourier results would
not apply but Theorem 1.1 shows that poly(d log(1/δ)) samples suffices. In fact, in this special case
we can get a better polynomial bound:
Theorem 1.2. For any degree d polynomial P (t) and an arbitrary function g(t), Procedure RobustPolynomialLearning in Algorithm 5 takes O(d) samples from x(t) = P (t) + g(t) over [0, T ]
and reports a degree d polynomial Q(t) in time O(dω ) such that, with probability at least 99/100,
kP (t) − Q(t)k2T . kg(t)k2T .
where ω < 2.373 is matrix multiplication exponent [Str69],[CW87],[Wil12].
We also show how to reduce the failure probability to an arbitrary p > 0 with O(log(1/p))
independent repetitions, in Theorem 4.5.
Although we have not seen such a result stated in the literature, our method is quite similar to
one used in [CDL13]. Since d samples are necessary to interpolate a polynomial without noise, the
result is within constant factors of optimal.
One could apply Theorem 1.2 to approximate other functions that are well approximated by
polynomials or piecewise polynomials. For example, a Gaussian of standard deviation at least σ
2
can be approximated by a polynomial of degree O( Tσ + log(1/δ)); hence the same bound applies
as the sample complexity of improper interpolation of a positive mixture of Gaussians.
1.1
Related work
Sparse discrete Fourier transforms. There is a large literature on sparse discrete Fourier
transforms. Results generally are divided into two categories: one category of results that carefully
choose measurements that allow for sublinear recovery time, including [GGI+ 02, GMS05, HIKP12b,
Iwe13, HIKP12a, IK14, IKP14, Kap16]. The other category of results expect randomly chosen
measurements and show that a generic recovery algorithm such as `1 minimization will work with
high probability; these results often focus on proving the Restricted Isometry Property [CRT06,
RV08, Bou14, HR15]. At the moment, the first category of results have better theoretical sample
complexity and running time, while results in the second category have better failure probabilities
and empirical performance. Our result falls in the first category. The best results here can achieve
O(k log n) samples [IK14], O(k log2 n) time [HIKP12b], or within log log n factors of both [IKP14].
4
For signals that are not periodic, the discrete Fourier transform will not be sparse: it takes k/δ
frequencies to capture a 1 − δ fraction of the energy. To get a better dependence on δ, one has to
consider frequencies “off the grid”, i.e. that are not multiples of 1/T .
Off the grid. Finding the frequencies of a signal with sparse Fourier transform off the grid has been
a question of extensive study. The first algorithm was by Prony in 1795, which worked in the noiseless
setting. This was refined by classical algorithms like MUSIC [Sch81] and ESPRIT [RPK86], which
empirically work better with noise. Matrix pencil [BM86] is a method for computing the maximum
likelihood signal under Gaussian noise and evenly spaced samples. The question remained how
accurate the maximum likelihood estimate is; [Moi15] showed that it has an O(k c ) approximation
factor if the frequency gap is at least 1/T .
Now, the above results all use F T samples, which is analogous to n in the discrete setting. This
can be decreased down till O(k) by only looking at a subset of time, i.e. decreasing T ; but doing so
increases the frequency gap needed for decent robustness results.
A variety of works have studied how to adapt sparse Fourier techniques from the discrete setting
to get sublinear sample complexity; they all rely on the minimum separation among the frequencies
to be at least c/T for c ≥ 1. [TBSR13] showed that a convex program can recover the frequencies
exactly in the noiseless setting, for c ≥ 4. This was improved in [CF14] to c ≥ 2 for complex signals
and c ≥ 1.87 for real signals. [CF14] also gave a result for c ≥ 2 that was stable to noise, but this
required the signal frequencies to be placed on a finely spaced grid. [YX15] gave a different convex
relaxation that empirically requires smaller c in the noiseless setting. [DB13] used model-based
compressed sensing when c = Ω(1), again without theoretical noise stability. Note that, in the
noiseless setting, exact recovery can be achieved without any frequency separation using Prony’s
method or Berlekamp-Massey syndrome decoding [Mas69]; the benefit of the above results is that
a convex program might be robust to noise, even if it has not been proven to be so.
In the noisy setting, [FL12] gave an extension of Orthogonal Matching Pursuit (OMP) that can
recover signals when c = Ω(k), with an approximation factor O(k), and a few other assumptions.
Similarly, [BCG+ 14] gave a method that required c = Ω(k) and was robust to certain kinds of noise.
[HK15] got the threshold down to c = O(1), in multiple dimensions, but with approximation factor
O(F T k O(1) ).
[TBR15] shows that, under Gaussian noise and with separation c ≥ 4, a semidefinite program
can optimally estimate x∗ (ti ) at evenly spaced sample points ti from observations x∗ (ti )+g(ti ). This
is somewhat analogous to our setting, the differences being that (a) we want to estimate the signal
over the entire interval, not just the sampled points, (b) our noise g is adversarial, so we cannot
hope to reduce it—if g is also k-Fourier-sparse, we cannot distinguish x∗ and g, and of course (c)
we want to avoid requiring frequency separation.
In [PS15], we gave the first algorithm with O(1) approximation factor, finding the frequencies
when c & log(1/δ), and the signal when c & log(1/δ) + log2 k.
Now, all of the above results algorithms are designed to recover the frequencies; some of the ones
in the noisy setting then show that this yields a good approximation to the overall signal (in the
noiseless setting this is trivial). Such an approach necessitates c ≥ 1: [Moi15] gave a lower bound,
showing that any algorithm finding the frequencies with approximation factor 2o(k) must require
c ≥ 1.
Thus, in the current literature, we go from not knowing how to get any approximation for c < 1,
to getting a polynomial approximation at c = 1 and a constant approximation at c & log2 k. In this
work, we show how to get a constant factor approximation to the signal regardless of c.
5
Polynomial interpolation. Our result is a generalization of robust polynomial interpolation,
and in Theorem 1.2 we construct an optimal method for polynomial interpolation as a first step
toward interpolating Fourier-sparse signals.
Our result here can be seen as essentially an extension of a technique shown in [CDL13]. The
focus of [CDL13] is on the setting where sample points xi are chosen independently, so Θ(d log d)
samples are necessary. One of their examples, however, shows essentially the same thing as our
Corollary 4.2. From this, getting our theorem is not difficult.
The recent work [GZ16] looks at robust polynomial interpolation in a different noise model,
featuring `∞ bounded noise with some outliers. In this setting they can get a stronger `∞ guarantee
on the output than is possible in our setting.
Nyquist sampling. The classical method for learning bandlimited signals uses Nyquist sampling—
i.e., samples at rate 1/F , for F T points—and interpolates them using Shannon-Nyquist interpolation. This doesn’t require any frequency gap, but also doesn’t benefit from sparsity like sparse
Fourier transform-based techniques. As discussed in [PS15], on the signal x(t) = 1 it takes
F T + O(1/δ) samples to get δ error on average. Our dependence is logarithmic on both those
terms.
1.2
Our techniques
Previous results on sparse Fourier transforms with robust recovery all required a frequency gap. So
consider the opposite situation, where all the frequencies converge to zero and the coefficients are
adjusted to keep the overall energy fixed. If we take a Taylor expansion of each complex exponential,
then the signal will converge to a degree k polynomial. So robust polynomial interpolation is a
necessary subproblem for our algorithm.
Polynomial interpolation. Let P (x) be a degree d polynomial, and suppose that we can query
f (x) = P (x) + g(x) over the interval [−1, 1], where g represents adversarial noise. We would like to
query f at O(d) points
and output a degree d polynomial Q(x) such that kP − Qk . kgk, where
R1
we define khk2 := −1 |h(x)|2 dx.
One way to do this would be to sample points S ⊂ [−1, 1] uniformly, then output the degree d
polynomial Q with the smallest empirical error
kP + g − Qk2S :=
1 X
|(P + g − Q)(x)|2
|S|
x∈S
on the observed points. If kRkS ≈ kRk for all degree d polynomials R, in particular for P − Q, then
since usually kgkS . kgk by Markov’s inequality, the result follows.
This has two problems: first, uniform sampling is poor because polynomials like Chebyshev polynomials can have most of their energy within O(1/d2 ) of the edges of the interval. This necessitates
Ω(d2 ) uniform samples before kRkS ≈ kRk with good probability on a single polynomial. Second,
the easiest method to extend from approximating one polynomial to approximating all polynomials
uses a union bound over a net exponential in d, which would give an O(d3 ) bound.
To fix this, we need to bias our sampling toward the edges of the interval and we need our
sampling to not be iid. We partition √
[−1, 1] into O(d) intervals I1 , . . . , In so that the interval
containing each x has width at most O( 1 − x2 ), except for the O(1/d2 ) size regions at the edges.
For any degree d polynomial R and any choice of n points xi ∈ Ii , the appropriately weighted
empirical energy is close to kRk. This takes care of both issues with uniform sampling. If the points
6
are chosen uniformly at random from within their intervals, then kgk is probably bounded as well,
and the empirically closest degree d polynomial Q will satisfy our requirements.
This result is shown in Section 4.
Clusters. Many previous sparse Fourier transform algorithms start with a one-sparse recovery
algorithm, then show how to separate frequencies to get a k-sparse algorithm by reducing to the
one-sparse case. Without a frequency gap, we cannot hope to reduce to the one-sparse case; instead,
we reduce to individual clusters of nearby frequencies.
Essentially the problem is that one cannot determine all of the high-energy frequencies of a
function x only by sampling it on a bounded interval, as some of the frequencies might cancel
each other out on this interval. We also cannot afford to work merely with the frequencies of the
truncation of x to the interval [0, T ], as the truncation operation will spread the frequencies of x
over too wide a range. To fix this problem, we must do something in between the two. In particular,
we instead study x·H for a judiciously chosen function H. We want H to approximate the indicator
b ⊂ [−k c /T, k c /T ]. By using
function of the interval [0, T ] and have small Fourier-support, supp(H)
∗
some non-trivial lemmas about the growth rate of x , we can show that the difference between
x · H on R and the truncation of x to [0, T ] has small L2 mass, so that we can use the former as a
substitute for the latter.
b which has most
On the other hand, the Fourier transform of x · H is the convolution x
b ∗ H,
∗
of its mass within poly(k)/T of the frequencies of x . Although it is impossible to determine the
individual frequencies of x∗ , we can hope to identify O(k) intervals each of length poly(k)/T so that
all but a small fraction of the energy of x
b is contained within these intervals.
Note that many of these intervals will represent not individual frequencies of x∗ , but small
clusters of such frequencies. Furthermore, some frequencies of x∗ might not show up in these
intervals either because they are too small, or because they cancel out other frequencies when
b
convolved with H.
One-cluster recovery. Given our notion of clusters, we start looking at Fourier-sparse interpolation in the special case of one-cluster recovery. This is a generalization of one-sparse recovery where
we can have multiple frequencies, but they all lie in [f − ∆, f + ∆] for some base frequency f and
bandwidth ∆ = k c /T . Because all the frequencies are close to each other, values x(a) and x(a + β)
will tend to have ratio close to e2πif β when β is small enough. We find that β < ∆√1T ∆ is sufficient,
√
which lets us figure out a frequency fe with |fe − f | ≤ ∆ T ∆ = k O(1) /T .
e
Once we have the frequency fe, we can consider x0 (t) = x(t)e−2πif . This signal is k-Fouriersparse with frequencies bounded by k O(1) /T . By taking a Taylor approximation to each complex
e
exponential1 , can show x∗ is δ-close to P (t)e2πif for a degree d = O(k c + k log(1/δ)) polynomial P .
Thus we could apply our polynomial interpolation algorithm to recover the signal.
k-cluster frequency estimation. Reminiscent of algorithms such as [HIKP12a, PS15], we choose
c
random variables σ ≈ T /k c , a ∈ [0, 1], and b ∈ [0, 1/σ] and look at v ∈ Ck given by
vi = (x · H)(σ(i − a))e−2πiσbi G(i)
1
There is a catch here, that the coefficients of the exponentials are potentially unbounded, if the frequencies are
arbitrarily close together. We first use Gram determinants to show that the signal is δ-close to one with frequency
gap δ2−k , and coefficients at most 2k /δ.
7
b approxwhere G is a filter function. That is, G has compact support (supp(G) ⊂ [−k c , k c ]), and G
2π
b
imates an interval of length Θ( k ). In other words, G is the same as H with different parameters:
an interval convolved with itself k c times, multiplied by a sinc function.
We alias v down to O(k) dimensions and take the discrete Fourier transform, getting u
b. It has
been implicit in previous work—and we make it explicit—that u
bj is equal to zσa for a vector z
defined by
b ·G
b (j)
zb = (b
x ∗ H)
σ,b
b (j) is a particular permutation of G.
b In particular, G
b (j) has period 1/σ, and approximates
where G
σ,b
σ,b
1
an interval of size σB
within each period.
In previous work, when σ and b were chosen randomly, each individual frequency would have
a good chance of being the only frequency preserved in zb, and we could apply one-sparse recovery
by choosing a variety of a. Without a frequency gap we can’t quite say that: we pick 1/σ ∆ so
that the entire cluster usually lands in the same bin, but then nearby clusters can also often land
in the same bin. Fortunately, it is still usually true that only nearby clusters will collide. Since our
1-cluster algorithm works
√ when the signal frequencies are nearby, we apply it to find a frequency
T /σ
approximation within σ = k O(1) /T of the cluster.
The above algorithm recovers each individual frequency with constant probability. By repeating
it O(log k) times, with high probability we find a list L of O(k) frequencies within k O(1) /T of each
significant cluster.
k-sparse recovery. Because different clusters aren’t anywhere close to orthogonal, we can’t simply
approximate each cluster separately and add them up. Instead, given the list L of candidate
frequencies, we consider the O(kd)-dimensional space of functions
x
e(t) :=
d
XX
αfe,i ti e2πif t
e
fe∈L i=0
where d = O(k O(1) + log(1/δ)). We then take a bunch of random samples of x, and choose the x
e(t)
minimizing the empirical error using linear regression. This regression can be made slightly faster
using oblivious subspace embeddings [CW13], [NN13], [Woo14],[CNW15].
Our argument to show this works is analogous to the naive method we considered for polynomial
recovery. Similarly to the one-cluster setting, using Taylor approximations and Gram determinants,
we can show that this space includes a sufficiently close approximation to x. Since polynomials
are the limit of sparse Fourier as frequencies tend to zero, these functions are arbitrarily close to
O(kd)-Fourier-sparse functions. Hence we know that the maximum of |e
x(t)| is at most a poly(kd)
factor larger than its average over [0, T ]. Using a net argument, this shows poly(kd) samples are
sufficient to find a good approximation to the nearest function in our space.
Growth rate of Fourier-sparse signals. We need that
√1
T
kx∗ · Hk2 ≈ kx∗ kT , where H approx0
imates the interval 1[0,T ] . Because H has support size k c /T , it has a transition region of size T /k c
c00
at the edges, and it decays as (t/T )−k for t T . The difference between √1T kx∗ · Hk2 and kx∗ kT
involves two main components: mass in the transition region that is lost, and mass outside the
e 2 ) kx∗ k
sampling interval that is gained. To show the approximation, we need that |x∗ (t)| . O(k
T
∗
O(k)
∗
within the interval and |x (t)| . (kt/T )
kx kT outside.
8
We outline the bound of max |x∗ (t)| in terms of its average kx∗ kT to bound |x∗ (t)| within the
t∈[0,T ]
interval. Notice that we can assume |x∗ (0)| = max |x∗ (t)|: if t∗ = arg max|x∗ (t)|2 is not 0 or T , we
t∈[0,T ]
[0, t∗ ]
t∈[0,T ]
can rescale the two intervals
and
to [0, T ] separately.P
Then we show that for any t0 , there
2
∗
∗
0
e ) and constants C1 , · · · , Cm such that x (0) =
exist m = O(k
j∈[m] Cj · x (j · t ). Then we take
the integration of t0 over [0, T /m]
to bound |x∗ (0)|2 by its average. For any outside t > T , we follow
P
this approach to show x∗ (t) = j∈[k] Cj · x∗ (tj ) where tj ∈ [0, T ] and |Cj | ≤ poly(k) · (kt/T )O(k)
for each j ∈ [k]. These results are shown in Section 5.
1.3
[t∗ , T ]
Organization
This paper is organized as follows. We provide a brief overview about signal recovery in Section 2.
We introduce some notations and tools in Section 3. Then we show our main Theorem 1.2 about
polynomial interpolation in Section 4. For signals with k-sparse Fourier transform, we show two
bounds on their growth rate in Section 5 and describe the hash functions and filter functions in
Section 6. We provide the algorithm for frequency estimation and its proof in Section 7. In Section 8,
we describe the algorithm for one-cluster recovery. In Section 9, we show the proof of Theorem 1.1.
We defer several technical proofs in Appendix A. Appendix B gives a summary of several wellknown facts are existing in literature. We provide the analysis of hash functions and filter functions
in Appendix C.
2
Proof Sketch
We first consider one-cluster recovery centered at zero, i.e., x∗ (t) =
k
P
vj · e2πifj t where every fj is
j=1
in [−∆, ∆] for some small ∆ > 0. The road map is to replace x∗ by a low degree polynomial P
such that kx∗ (t) − P (t)k2T . δkx∗ k2T then recover a polynomial Q
to approximate P through the
0
0
∗
observation x(t) = P (t) + g (t) where g (t) = g(t) + x (t) − P (t) .
k
P
A natural way to replace x∗ (t) =
vj e2πifj t by a low degree polynomial P (t) is the Taylor
j=1
expansion. To bound the error after taking the low degree terms in the expansion by δkx∗ kT , we
k
P
0
show the existence of x0 (t) =
vj0 e2πifj t approximating x∗ on [0, T ] with an extra property—any
j=1
RT
0
0
coefficient vj in x (t) has an upper bound in terms of kx0 k2T = T1 0 |x0 (t)|2 dt. We prove the existence
of x0 (t) via two more steps, both of which rely on the estimation of some Gram matrix constituted
by these k signals.
The first step is to show the existence of a k-Fourier-sparse signal x0 (t) with frequency gap
η ≥ exp(− poly(k))·δ
that is sufficiently close to x∗ (t).
T
Lemma 2.1. There is a universal constant C1 > 0 such that, for any x∗ (t) =
k
P
vj e2πifj t and any
j=1
δ > 0 , there always exist η ≥
δ
T
· k −C1
k2
and x0 (t) =
k
P
vj0 e
2πifj0 t
j=1
kx0 (t) − x∗ (t)kT ≤ δkx∗ (t)kT
with min|fi0 − fj0 | ≥ η and max{|fj0 − fj |} ≤ kη.
i6=j
j∈[k]
9
satisfying
We outline our approach
and defer the proof to Section 8. We focus on the replacement of one
P
∗
frequency fk in x = j∈[k] vj e2πifj t by a new frequency fk+1 6= fk and its error. The idea is to
consider every signal e2πifj t as a vector and prove that for any vector x∗ in the linear subspace
span{e2πifj t |j ∈ [k]}, there exists a vector in the linear subspace span{e2πifk+1 t , e2πifj t |j ∈ [k − 1]}
with distance at most exp(k 2 ) · (|fk − fk+1 |T ) · kx∗ kT to x∗ .
The second step is to lower bound kx0 k2T by its coefficients through the frequency gap η in x0 .
Lemma 2.2. There exists a universal constant c > 0 such that for any x(t) =
k
P
vj e2πifj t with
j=1
frequency gap η = min|fi − fj |,
i6=j
k
X
2
kx(t)k2T ≥ k −ck min (ηT )2k , 1
|vj |2 .
j=1
Combining Lemma 2.1 and Lemma 2.2, we bound |vj0 | by exp(poly(k)) · δ −O(k) · kx0 kT for any
coefficient vj0 in x0 . Now we apply the Taylor expansion on x0 (t) and keep the first d = O(∆T +
0
poly(k) + k log 1δ ) terms of every signal vj0 · e2πifj t in the expansion to obtain a polynomial P (t) of
degree at most d. To bound the distance
between P (t) and x0 (t), we observe that the error of every
2π∆·T d P
0
0
point t ∈ [0, T ] is at most ( d )
j |vj |, which can be upper bounded by δkx (t)kT via the above
connection. We summarize all discussion above as follows.
P
Lemma 2.3. For any ∆ > 0 and any δ > 0, let x∗ (t) = j∈[k] vj e2πifj t where |fj | ≤ ∆ for each
j ∈ [k]. There exists a polynomial P (t) of degree at most
d = O(T ∆ + k 3 log k + k log 1/δ)
such that
kP (t) − x∗ (t)k2T ≤ δkx∗ k2T .
To recover x∗ (t), we observe x(t) as a degree d polynomial P (t) with noise. We use properties
of the Legendre polynomials to design a method of random sampling such that we only need O(d)
random samples to find a polynomial Q(t) approximating P (t).
Theorem 1.2. For any degree d polynomial P (t) and an arbitrary function g(t), Procedure RobustPolynomialLearning in Algorithm 5 takes O(d) samples from x(t) = P (t) + g(t) over [0, T ]
and reports a degree d polynomial Q(t) in time O(dω ) such that, with probability at least 99/100,
kP (t) − Q(t)k2T . kg(t)k2T .
where ω < 2.373 is matrix multiplication exponent [Str69],[CW87],[Wil12].
We can either report the polynomial Q(t) or transfer Q(t) to a signal with d-sparse Fourier
transform. We defer the technical proofs and the formal statements to Section 8 and discuss the
recovery of k clusters from now on.
∗ · H has at
b )) on x∗ such that x\
As mentioned before, we apply the filter function (H(t), H(f
∗
c with k-sparse Fourier transform. First, we show that all frequencies in the
most k clusters given x
∗
\
“heavy” clusters of x · H constitute a good approximation of x∗ in Section 9.
10
k
P
Definition 2.4. Given x∗ (t) =
b with bounded
vj e2πifj t , any N > 0, and a filter function (H, H)
j=1
\
j t · H) for each j ∈ [k].
support in frequency domain. Let Lj denote the interval of supp(e2πif
Define an equivalence relation ∼ on the frequencies fi by the transitive closure of the relation
fi ∼ fj if Li ∩ Lj 6= ∅. Let S1 , . . . , Sn be the equivalence classes under this relation.
R
\
Define Ci = ∪ Li for each i ∈ [n]. We say Ci is a “heavy” cluster iff Ci |H
· x∗ (f )|2 df ≥
f ∈Si
T · N 2 /k.
Claim 2.5. Given x∗ (t) =
k
P
vj e2πifj t and any N > 0, let H be the filter function defined in
j=1
Appendix C.1 and C1 , · · · , Cl be the heavy clusters from Definition 2.4. For
S = j ∈ [k] fj ∈ C1 ∪ · · · Cl ,
we have x(S) (t) =
P
j∈S
vj e2πifj t approximating x∗ within distance kx(S) (t) − x∗ (t)k2T . N 2 .
b
Hence it is enough to recover x(S) for the recovery of x∗ . Let ∆h denote the bandwidth of H.
R fj +∆
2
2
\
In Section 7, we choose ∆ > k · ∆h such that for any j ∈ S, fj −∆ |H
· x∗ (f )| df ≥ T · N /k from
the fact |Ci | ≤ k · ∆h . Then we prove Theorem 2.6 in Section 7, which finds O(k) frequencies to
∗ · H.
cover all heavy clusters of x\
Theorem 2.6. Let x∗ (t) =
k
P
vj e2πifj t and x(t) = x∗ (t) + g(t) be our observable signal where
j=1
kg(t)k2T ≤ ckx∗ (t)k2T for a sufficiently small constant c. Then Procedure FrequencyRecoveryKCluster returns a set L of O(k) frequencies that covers all heavy clusters of x∗ , which
uses poly(k, log(1/δ)) log(F T ) samples and poly(k, log(1/δ)) log2 (F T ) time. In particular, for ∆ =
poly(k, log(1/δ))/T and N 2 := kg(t)k2T + δkx∗ (t)k2T , with probability 1 − 2−Ω(k) , for any f ∗ with
Z
f ∗ +∆
|x[
· H(f )|2 df ≥ T N 2 /k,
f ∗ −∆
there exists an fe ∈ L satisfying
√
|f ∗ − fe| . ∆ ∆T .
(2)
Let L = {fe1 , · · · , fel } be the list of frequencies from the output of Procedure FrequencyRecoveryKCluster in Theorem 2.6. The guarantee is that, for any fj in x(S) , there exists some
√
pj ∈ [l] such that |fepj − fj | . ∆ ∆T for ∆ = poly(k, log(1/δ))/T . Hence we rewrite x(S) (t) =
P
P
2πifei t (
2πi(fj −fei )t ). For each i ∈ [l], we apply Lemma 2.3 of one-cluster recovery on
j∈S:pj =i e
i∈[l] e
P
2πi(fj −fei )t to approximate it by a degree d polynomial P (t).
i
j∈S:pj =i e
P
e
Now we consider x(t) = i∈[l] e2πifi t · Pi (t) + g 00 (t) where kg 00 (t)kT . kg(t)kT + δkx∗ (t)kT . To
P
e
recover i∈[l] e2πifi t · Pi (t), we treat it as a vector in the linear subspace
V = span e
2πifei t
· t j ∈ {0, · · · , d}, i ∈ [l]
j
with dimension at most l(d + 1) and find a vector in this linear subspace approximating it.
11
We show that for any v ∈ V , the average of poly(kd) random samples on v is enough to estimate
In particular, any vector in this linear subspace satisfies that the maximum of it in [0, T ] has
an upper bound in terms of its average in [0, T ]. Then we apply the Chernoff bound to prove that
poly(kd) random samples are enough for the estimation of one vector v ∈ V .
ei t
2πi
f
j
Claim 2.7. For any ~u ∈ span e
· t j ∈ {0, · · · , d}, i ∈ [l] , there exists some universal conkvk2T .
stants C1 ≤ 4 and C2 ≤ 3 such that
max {|~u(t)|2 } . (ld)C1 logC2 (ld) · k~uk2T
t∈[0,T ]
At last we use an -net to argue that poly(kd) random samples from [0, T ] are enough to
interpolate x(t) by a vector v ∈ V . Because the dimension of this linear subspace is at most
l(d + 1) = O(kd), there exists an -net in this linear subspace for unit vectors with size at most
exp(kd). Combining the Chernoff bound on all vectors in the -net and Claim 2.7, we know that
poly(kd) samples are sufficient to estimate kvk2T for any vector v ∈ V . In Section 9, we show that
a vector v ∈ V minimizing the distance on poly(kd) random samples is a good approximation for
P
2πifei t · P (t), which is a good approximation for x∗ (t) from all discussion above.
i
i∈[l] e
Theorem 1.1. Let x(t) = x∗ (t) + g(t), where x∗ is k-Fourier-sparse signal with frequencies in
[−F, F ]. Given samples of x over [0, T ] we can output x
e(t) such that with probability at least
1 − 2−Ω(k) ,
ke
x − x∗ kT . kgkT + δ kx∗ kT .
Our algorithm uses poly(k, log(1/δ)) · log(F T ) samples and poly(k, log(1/δ)) · log2 (F T ) time. The
output x
e is poly(k, log(1/δ))-Fourier-sparse signal.
3
Preliminaries
We first provide some notations in Section 3.1 and basic Fourier facts in Section 3.2. Then we
review some probability inequalities in Section 3.3. At last, we introduce Legendre polynomials in
Section 3.4 and review some basic properties of Gram matrix and its determinant in Section 3.5.
3.1
Notation
e ) to be f · logO(1) (f ). We use [n] to denote {1, 2, · · · , n}. Let i
For any√function f , we define O(f
z = a + ib ∈ C, where a, b ∈ R. We define z to be a − ib
denote −1.
√ For any Complex number
2
2
2
and |z| = a + b such that |z| = zz. For any function f (t) : R → C, we use supp(f ) to denote
the support of f .
For convenience, we define the sinc function and the Gaussian distribution Gaussianµ,σ on R
with expectation µ and variance σ 2 as follows:
sinc(t) =
sin(πt)
,
πt
(t−µ)2
1
Gaussianµ,σ (t) = √ e− 2σ2 .
σ 2π
For a fixed T > 0, we define the inner product of two functions x, y : [0, T ] → C as
1
hx, yiT =
T
Z
0
12
T
x(t)y(t)dt.
Combs
−2s
−s
s
0
2s
rects
s
2
− 2s
sincs
− 1s
1
s
Gaussianµ,σ
µ
Figure 1: A picture of a Combs , rects , sincs , Gaussianµ,σ .
We define the k · kT norm as
p
kx(t)kT = hx(t), x(t)iT =
3.2
s
1
T
Z
T
0
|x(t)|2 dt.
Facts about the Fourier transform
In this work, we always use x(t) to denote a signal from R → C. The Fourier transform x
b(f ) of an
integrable function x : R → C is defined as
Z +∞
x(t)e−2πif t dt, for any real number f.
x
b(f ) =
−∞
Similarly, x(t) is determined from x
b(f ) by the inverse transform:
Z +∞
x(t) =
x
b(f )e2πif t df, for any real number t.
−∞
Let CFT denote the continuous Fourier transform, DTFT denote the discrete-time Fourier
transform, DFT denote the discrete Fourier transform, and FFT denote the fast Fourier transform.
For any signal x(t) and n ∈ N+ , we define x∗n (t) = x(t) ∗ · · · ∗ x(t) and x
b·n (f ) = x
b(f ) · · · · · x
b(f ).
|
{z
}
|
{z
}
n
Fact 3.1. Let δ∆ (f ) denote the Dirac delta at ∆. Then
Z +∞
b
δ∆ (t) =
δ∆ (f )e2πif t df = e2πit∆ .
−∞
13
n
Fact 3.2. For any s > 0, let Combs (t) =
P
δjs (t). Then the Fourier transform of Combs (t) is
j∈Z
\ s (f ) = 1 Comb1/s (f ).
Comb
s
The following fact says the the Fourier transform of a rectangle function is a sinc function.
Fact 3.3. We use
(
1
rects (t) =
0
if |t| ≤ 2s ,
otherwise.
Then the Fourier transform of rects (t) is r[
ects (f ) =
sin(πf s)
πf s
= sinc(f s).
The Fourier transform of a Gaussian function is another Gaussian function.
Fact 3.4. For Gaussianµ,σ (t) =
√1 e−
σ 2π
(t−µ)2
2σ 2
. Then the Fourier transform is
\ µ,σ (f ) = e−2πif u √1 Gaussian0,σ0 (f ) for σ 0 = 1/(2πσ).
Gaussian
σ 2π
Proof. From the definition of the Fourier transform,
Z +∞
(t−µ)2
1
\
√ e− 2σ2 e−2πif t dt
Gaussianµ,σ (f ) =
−∞ σ 2π
Z +∞
t2
1
−2πif u
√ e− 2σ2 e−2πif t dt
= e
−∞ σ 2π
Z +∞
(t+2πiσ 2 f )2
1
−2π 2 f 2 σ 2
√ e− 2σ2
= e−2πif u
dt
−∞ σ 2π
f2
= e−2πif u e− 2σ02
√
where σ 0 = 1/(2σπ), which is e−2πif u · σ 0 2π · Gaussian0,σ0 (f ).
3.3
Tools and inequalities
From the Chernoff Bound (Lemma B.2), we show that if the maximum of a signal is bounded by d
times its energy over some fixed interval, then taking more than d samples (each sample is drawn
i.i.d. over that interval) suffices to approximate the energy of the signal on the interval with high
probability.
Lemma 3.5. Given any function x(t) : R → C with max |x(t)|2 ≤ dkx(t)k2T . Let S denote a set of
t∈[0,T ]
points from 0 to T . If each point of S is chosen uniformly at random from [0, T ], we have
"
#
1 X
2
Pr
|x(ti )|2 − kx(t)k2T ] ≥ kx(t)k2T
≤ e−Ω( |S|/d)
|S|
i∈S
We provide a proof in Appendix A.5.
1
1
Because d · 2d
+ 12 · (1 − 2d
) ≤ 1, we have the following inequality when the maximum of |x(t)|2
is at most d times its average.
Lemma 3.6. Given any function x(t) : R → C with max |x(t)|2 ≤ dkx(t)k2T . Let S denote a set of
t∈[0,T ]
points from 0 to T . For any point a is sampled uniformly at random from [0, T ], we have,
1
1
2
2
.
Pr
|x(a)| ≥ kx(t)kT ≥
2
2d
a∼[0,T ]
14
3.4
Legendre polynomials
We provide an brief introduction to Legendre polynomials
(please see [Dun10] for a complete introR
1 1
2
2
duction). For convenience, we fix kf (t)kT = 2 −1 |f (t)| dt in this section.
Definition 3.7. Let Ln (x) denote the Legendre polynomials of degree n, the solution to Legendre’s
differential equation:
d
2 d
(1 − x ) Ln (x) + n(n + 1)Ln (x) = 0
(3)
dx
dx
We will the following two facts about the Legendre polynomials in this work.
Fact 3.8. Ln (1) = 1 for any n ≥ 0 in the Legendre polynomials.
Fact 3.9. The Legendre polynomials constitute an orthogonal basis with respect to the inner product
on interval [−1, 1]:
Z 1
2
Lm (x)Ln (x)dx =
δmn
2n
+1
−1
where δmn denotes the Kronecker delta, i.e., it equals to 1 if m = n and to 0 otherwise.
For any polynomial P (x) of degree at most d with complex coefficients, there exists a set of
coefficients from the above properties such that
P (x) =
d
X
i=0
αi · Li (x), where αi ∈ C, ∀i ∈ {0, 1, 2, · · · , d}.
Lemma 3.10. For any polynomial P (t) of degree at most d from R to C, for any interval [S, T ],
Z T
1
2
2
max |P (t)| ≤ (d + 1) ·
|P (t)|2 dx.
T −S S
t∈[S,T ]
We provide a proof in Appendix A.6.
3.5
Gram matrix and its determinant
We provide an brief introduction to Gramian matrices (please see [Haz01] for a complete introduction). We use hx, yi to denote the inner product between vector x and vector y.
Let ~v1 , · · · , ~vn be n vectors in an inner product space
( and span{~v1 , · · · , ~vn } be
) the linear subspace
P
spanned by these n vectors with coefficients in C, i.e.,
αi~vi |∀i ∈ [n], αi ∈ C . The Gram matrix
i∈[n]
Gramn of ~v1 , · · · , ~vn is an n × n matrix defined as Gramn (i, j) = h~vi , ~vj i for any i ∈ [n] and j ∈ [n].
Fact 3.11. det(Gramn ) is the square of the volume of the parallelotope formed by ~v1 , · · · , ~vn .
k
Let Gramn−1 be the Gram matrix of ~v1 , · · · , ~vn−1 . Let ~vn be the projection of vn onto the linear
k
⊥
subspace span{~v1 , · · · , ~vn−1
p } and ~vn = ~vn − ~vn . We use k~v k to denote the length of ~v in the inner
product space, which is h~v , ~v i.
Claim 3.12.
k~vn⊥ k2 =
det(Gramn−1 )
.
det(Gramn )
Proof.
det(Gramn ) = volume2 (~v1 , · · · , ~vn ) = volume2 (~v1 , · · · , ~vn−1 ) · k~vn⊥ k2 = det(Gramn ) · k~vn⊥ k2 .
15
4
Robust Polynomial Interpolation Algorithm
In Section 4.1, we show how to learn a low degree polynomial by using linear number of samples,
running polynomial time, and achieving constant success probability. In Section 4.2, we show to
how boost the success probability by rerunning previous algorithm several times.
4.1
Constant success probability
We show how to learn a degree-d polynomial P with n = O(d) samples and prove Theorem
1.2 in
R1
this section. For convenience, we first fix the interval to be [−1, 1] and use kf k2[−1,1] = 21 −1 |f (t)|2 dt.
Lemma 4.1. Let d ∈ N and ∈ R+ , there exists an efficient algorithm to compute a partition of
[−1, 1] to n = O(d/) intervals I1 , · · · , In such that for any degree d polynomial P (t) : R → C and
any n points x1 , · · · , xn in the intervals I1 , · · · , In respectively, the function Q(t) defined by
Q(t) = P (xj )
if
t ∈ Ij
approximates P by
kQ − P k[−1,1] ≤ kP k[−1,1] .
(4)
One direct corollary from the above lemma is that observing n = O(d/) points each from
I1 , · · · , In provides a good approximation for all degree d polynomials. For any set S = {t1 , · · · , tm }
m
P
where each ti ∈ [−1, 1] and a distribution with support {w1 , · · · , wm } on S where
wi = 1 and
i=1
Pm
wi ≥ 0 for each i ∈ [m], we define kxkS,w = ( i=1 wi · |x(ti )|2 )1/2 .
Corollary 4.2. Let I1 , · · · , In be the intervals in the above lemma and wj = |Ij |/2 for each j ∈ [n].
For any x1 , · · · , xn in the intervals I1 , · · · , In respectively, we consider S = {x1 , · · · , xn } with the
distribution w1 , · · · , wn . Then for any degree d polynomial P , we have
kP kS,w ∈ (1 − )kP k[−1,1] , (1 + )kP k[−1,1] .
We first state the main technical lemma and finish the proof of the above lemma (we defer the
proof of Lemma 4.3 to Appendix A.3).
Lemma 4.3. For any degree d polynomial P (t) : R → C with derivative P 0 (t), we have,
Z 1
Z 1
2
0
2
2
(1 − t )|P (t)| dt ≤ 2d
|P (t)|2 dt.
−1
(5)
−1
Proof of Lemma 4.1.
We set m = 10d/ and show a partition of [−1, 1] into n ≤ 20m intervals.
√
1−t2
We define g(t) = m and y0 = 0. Then we choose yi = yi−1 + g(yi−1 ) for i ∈ N+ . Let l be the
first index of y such that yl ≥ 1 − m92 . We show l . m.
Let jk be the first index in the sequence such that yjk ≥ 1 − 2−k . Notice that
j2 ≤ √
3/4
1−(3/4)2
m
and
yi − yi−1 = g(yi−1 ) =
≤ 1.5m
q
2
1 − yi−1
m
16
≥
√
1 − yi−1
.
m
Then for all k > 2, we have
jk − jk−1 ≤ √
2−k
1−y(jk −1)
m
≤ 2−k/2 m.
Therefore jk ≤ 1.5 + (2−3/2 + · · · 2−k/2 ) m and l ≤ 10m.
Because yl−1 ≤ 1 − m92 , for any j ∈ [l] and any x ∈ [yi−1 , yi ], we have the following property:
2 )
1 − x2
1 (1 − yi−1
≥
·
= (yi − yi−1 )2 /2.
m2
2
m2
(6)
Now we set n and partition [−1, 1] into I1 , · · · , In as follows:
1. n = 2(l + 1).
2. For j ∈ [l], I2j−1 = [yj−1 , yj ] and I2j = [−yj , −yj−1 ].
3. I2l+1 = [yl , 1] and I2l+2 = [−1, −yl ].
For any x1 , · · · , xn where xj ∈ Ij for each j ∈ [n], we rewrite the LHS of (4) as follows:
n−2
XZ
Ij
j=1
|
2
|P (xj ) − P (t)| dt +
{z
}
A
Z
|
2
|P (xn−1 ) − P (t)| dt +
{z
In−1
Z
In
B
For A in Equation (7), from the Cauchy-Schwarz inequality, we have
n−2
XZ
j=1
Ij
2
|P (xj ) − P (t)| dt =
Z
n−2
XZ
j=1
2
t
0
P (y)dy dt ≤
xj
Ij
n−2
XZ
j=1
Ij
|P (xn ) − P (t)|2 dt .
}
|t − xj |
Z
t
xj
(7)
|P 0 (y)|2 dydt.
Then we swap dt with dy and use Equation (6):
n−2
XZ
j=1
Ij
0
2
|P (y)|
Z
t∈(x
/ j ,y)
|t − xj |dtdy ≤
n−2
XZ
j=1
0
Ij
2
2
|P (t)| · |Ij | dt ≤
n−2
XZ
Ij
j=1
|P 0 (t)|2
2(1 − t2 )
dt.
m2
We use Lemma 4.3 to simplify it by
n−2
XZ
j=1
Ij
2
|P (xj ) − P (t)| dt ≤
Z
1
−1
0
− t2 )
2d2
dt
≤
m2
m2
2 2(1
|P (t)|
Z
1
−1
|P (t)|2 dt.
For B in Equation (7), notice that |In−1 | = |In | = 1 − yl ≤ 9m−2 and for j ∈ {n − 1, n}
|P (t) − P (xj )|2 ≤ 4 max |P (t)|2 ≤ 4(d + 1)2 kP k2[−1,1]
t∈[−1,1]
from the properties of degree-d polynomials, i.e., Lemma 3.10. Therefore B in Equation (7) is upper
bounded by 2 · 4(d + 1)2 (9m−2 )kP (t)k2[−1,1] .
From all discussion above, kQ(t) − P (t)k2[−1,1] ≤
99d2
m2
≤ 2 .
Now we use the above lemma to provide a faster learning algorithm for polynomials on interval
[−1, 1] with noise instead of using the -nets argument. Algorithm RobustPolynomialLearningFixedInterval works as follows:
17
1. Let = 1/20 and I1 , · · · , In be the intervals for d and in Lemma 4.1.
2. Random choose xj ∈ Ij for every j ∈ [n] and define S = {x1 , · · · , xn } with weight w1 =
|In |
|I1 |
2 , · · · , wn = 2 .
3. Find the degree d polynomial Q(t) that minimizes kP (t) − Q(t)kS,w using Fact B.3.
Lemma 4.4. For any degree d polynomial P (t) and an arbitrary function g(t), Algorithm RobustPolynomialLearningFixedInterval takes O(d) samples from x(t) = P (t) + g(t) over [−1, 1]
and reports a degree d polynomial Q(t) in time O(dω ) such that, with probability at least 99/100,
kP (t) − Q(t)k2[−1,1] . kg(t)k2[−1,1] .
Proof. Notice that n = O(d/) = O(d) and the running time depends on solving a linear regression
problem( Fact B.3 ), which takes O(dω ) time. It is enough to bound the distance between P and
Q:
kP − Qk[−1,1]
≤ 1.09kP − QkS,w
by Corollary 4.2
= 1.09kx − g − QkS,w
by x = P + g
≤ 1.09kgkS,w + 1.09kx − QkS,w
by triangle inequality
≤ 1.09kgkS,w + 1.09kx − P kS,w
Q = arg min kR − xkS,w
degree-d R
≤ 2.2kgkS,w
Because E[kgk2S,w ] = kgk2[−1,1] , we know that kP − Qk[−1,1] ≤ 2200kgk[−1,1] with probability ≥ .999
S
by using Markov’s inequality.
e
For any function f : [0, T ] → C, let fe(t) = f ( 2t−T
T ). Then kf k[−1,1] = kf kT from the definition.
Hence we can switch any interval [0, T ] to [−1, 1] and use Lemma 4.4.
Theorem 1.2. For any degree d polynomial P (t) and an arbitrary function g(t), Procedure RobustPolynomialLearning in Algorithm 5 takes O(d) samples from x(t) = P (t) + g(t) over [0, T ]
and reports a degree d polynomial Q(t) in time O(dω ) such that, with probability at least 99/100,
kP (t) − Q(t)k2T . kg(t)k2T .
where ω < 2.373 is matrix multiplication exponent [Str69],[CW87],[Wil12].
4.2
Boosting success probability
Notice that the success probability of Theorem 1.2 is only constant, and the proof technique of
obtaining that result cannot be modified to 1 − 1/ poly(d) or 1 − 2−Ω(d) success probability due
to using Markov’s inequality. However, we can use that algorithm as a black box, and rerun it
O(log(1/p)) (for any p > 0) times on fresh samples. Using the careful median analysis from [MP14]
gives
Theorem 4.5. For any degree d polynomial P (t), an arbitrary function g(t), and any p > 0, Procedure RobustPolynomialLearning+ in Algorithm 5 takes O(d log(1/p)) samples from x(t) =
P (t) + g(t) over [0, T ] and reports a degree d polynomial Q(t) in time O(dω log(1/p)) such that, with
probability at least 1 − p,
kP (t) − Q(t)k2T . kg(t)k2T .
where ω < 2.373 is matrix multiplication exponent.
18
Proof. We run algorithm RobustPolynomialLearning R rounds with O(d) independent and
fresh samples per round. We will obtain R degree-d polynomials Q1 (t), Q2 (t), · · · , QR (t). We say a
polynomial Qi (t) is good if kQi (t) − P (t)k2T . kg(t)k2T . Using the Chernoff bound, with probability
at least 1 − 2−Ω(R) , at least a 3/4 fraction of the polynomials are “good”. We output polynomial
Q(t) = Qj ∗ (t) such that
j ∗ = arg min(median{kQj (t) − Q1 (t)k2T , kQj (t) − Q2 (t)k2T , · · · , kQj (t) − QR (t)k2T })
(8)
j∈[R]
The Equation (8) can be solved in following straightforward way. For i 6= j, it takes O(d) time to
compute kQj (t) − Qi (t)k2T . Because of the number of pairs is O(R2 ), thus it takes O(R2 d) time
write down a R × R matrix. For each column, we run linear time 1-median algorithm. This step
takes O(R2 ) time. At the end, j ∗ is index of the column that has the smallest median value. Thus,
polynomial Q(t) = Qj ∗ (t) 1 the 0 with probability at least 1 − p by choosing R = O(log(1/p)). The
running time is not optimized yet.
To improve the dependence on R for running time, we replace the step of solving Equation (8)
by an approach that is similar to [MP14]. We choose a new set of samples S, say S = {t1 , t2 , · · · , tn }
and n = O(d). Using Fact B.4, we can compute Qi (tj ) for all i, j ∈ [R] × [n] in O(Rd poly(log(d)))
time. Define
e j = median Qi (tj ), ∀j ∈ [n].
Q
(9)
i∈[R]
Our algorithm will output a degree-d polynomial Q which is the optimal solution of this problem,
e S,w .2 In the rest of the proof, we will show that kQ−P kT . kgkT with probability
min 0 kQ0 − Qk
degree-d Q
at least 1 − 2−Ω(R) .
e j − P (tj ) = median(Qi (tj ) − P (tj )). Fix a coordinate j
Notice that Equation (9) implies that Q
i∈[R]
and applying the proof argument of Lemma 6.1 in [MP14], we have
e j − P (tj ))2 . mean(Qi (tj ) − P (tj ))2
(Q
good i
Taking the weighted summation over all the coordinates j, we have
e − P k2S,w . meankQi − P k2S,w
kQ
good i
Using Corollary 4.2, for each good i,
kQi − P k2S,w . kQi − P k2T
Combining the above two inequalities gives
e − P k2S,w . meankQi − P k2T . kgk2T
kQ
(10)
e − Qk2 ≤ kQ
e − P k2 . kgk2
kQ
S,w
S,w
T
(11)
e − Qi0 k2S,w . kgk2T
kQ
(12)
good i
e then
Because Q is the optimal solution for Q,
Using Corollary 4.2 and for any good i, i0 , kQi − Qi0 kT . kgkT , we can replace P by Qi0 in the
Equation (10). Thus, for any Qi0 where i0 is good,
2
Outputting Q = arg min kQ0 − xkS,w is not good enough, because it only gives constant success probability.
degree-d Q0
19
For any good i0 ,
kQi0 − QkT
. kQi0 − QkS,w
e S,w + kQ
e − QkS,w
≤ kQi0 − Qk
by Corollary 4.2
by triangle inequality
. kgkT
by Equation (11) and (12)
Thus, our algorithm takes O(dR) samples from x(t) = P (t)+g(t) over [0, T ] and reports a polynomial
Q(t) in time O(Rdω ) such that, with probability at least 1 − 2−Ω(R) , kP (t) − Q(t)k2T . kg(t)k2T .
Choosing R = O(log(1/p)) completes the proof.
5
Bounding the Magnitude of a Fourier-sparse Signal in Terms of
Its Average Norm
The main results in this section are two upper bounds, Lemma 5.1 on max |x(t)|2 and Lemma 5.5
t∈[0,T ]
R
1 T
2
2
on |x(t)| for t > T , in terms of the typical signal value kxkT = T 0 |x(t)|2 dt. We prove Lemma
5.1 in Section 5.1 and Lemma 5.5 in Section 5.2
5.1
Bounding the maximum inside the interval
The goal of this section is to prove Lemma 5.1.
Lemma 5.1. For any k-Fourier-sparse signal x(t) : R → C and any duration T , we have
max |x(t)|2 . k 4 log3 k · kxk2T
t∈[0,T ]
R1
Proof. Without loss of generality, we fix T = 1. Then kxk2T = 0 |x(t)|2 dt. Because kxk2T is the
average over the interval [0, T ], if t∗ = arg max|x(t)|2 is not 0 or T = 1, we can rescale the two
t∈[0,T ]
intervals [0, t∗ ] and [t∗ , T ] to [0, 1] and prove the desired property separately. Hence we assume
|x(0)|2 = max |x(t)|2 in this proof.
t∈[0,T ]
Claim 5.2. For any k, there exists m = O(k 2 log k) such that for any k-Fourier-sparse signal x(t),
any t0 ≥ 0 and τ > 0, there always exist C1 , · · · , Cm ∈ C such that the following properties hold,
Property I
Property II
|Cj | ≤ 11 for all j ∈ [m],
X
x(t0 ) =
Cj · x(t0 + j · τ ).
j∈[m]
We first use this claim to finish the proof of Lemma 5.1. We choose t0 = 0 such that ∀τ > 0,
there always exist C1 , · · · , Cm ∈ C, and
X
x(0) =
Cj · x(j · τ ).
j∈[m]
By the Cauchy-Schwarz inequality, it implies that for any τ ,
X
|Cj |2 |x(j · τ )|2
|x(0)|2 ≤ m
j∈[m]
. m
X
j∈[m]
20
|x(j · τ )|2 .
(13)
At last, we obtain
2
|x(0)| = m
Z
1/m
|x(0)|2 dτ
0
. m·
Z
= m2 ·
2
= m ·
2
≤ m ·
1/m
(m
0
m
X
j=1
Z
m
X 1/m
j=1
m
X
j=1
m
X
j=1
|x(j · τ )|2 dτ
0
1
j
Z
1
·
j
|x(j · τ )|2 )dτ
j/m
|x(τ )|2 dτ
0
Z
1
0
|x(τ )|2 dτ
. m2 log m · kxk2T
where the first inequality follows
Equation (13), the second inequality follows by j/m ≤ 1
Pm by
1
and the last step follows by
=
O(log m). From m = O(k 2 log k), we obtain |x(0)|2 =
i=1 i
O(k 4 log3 kkxk2T ).
To prove Claim 5.2, we use the following lemmas about polynomials. We defer their proofs to
Appendix A.2.
Lemma 5.3. Let Q(z) be a degree k polynomial, all of whose roots are complex numbers with
Pk−1 (l)
rn,k · z l denote the residual polynomial of
absolute value 1. For any integer n, let rn,k (z) = l=0
rn,k (z) ≡ z n
(mod Q(z)).
(l)
Then, each coefficient of rn,k is bounded: |rn,k | ≤ 2k nk−1 for any l.
Lemma 5.4. For any k ∈ Z and any z1 , · · · , zk on the unit circle of C, there always exists a degree
m
P
m = O(k 2 log k) polynomial P (z) =
cj z j with the following properties:
j=0
Property I
Property II
Property III
Proof of Claim 5.2.
For x(t) =
k
P
P (zi ) = 0, ∀i ∈ {1, · · · , k},
c0 = 1,
|cj | ≤ 11, ∀j ∈ {1, · · · , m}.
vi e2πifi t , we fix t0 and τ then rewrite x(t0 +j ·τ ) as a polynomial
i=1
of bi = vi · e2πifi t0 and zi = e2πifi τ for each i ∈ [k].
x(t0 + j · τ ) =
=
=
k
X
i=1
k
X
i=1
k
X
i=1
21
vi e2πifi ·(t0 +j·τ )
vi e2πifi t0 · e2πifi ·jτ
bi · zij .
Given k and z1 , · · · , zk , let P (z) =
m
X
Pm
j=0 cj z
j
be the degree m polynomial in Lemma 5.4.
cj x(t0 + jτ ) =
j=0
m
X
k
X
cj
j=0
=
k
X
i=1
m
X
bi
i=1
=
k
X
j=0
bi · zij
cj · zij
bi P (zi )
i=1
= 0,
(14)
where the last step followsP
by Property I of P (z) in Lemma 5.4. From the Property II and III of
P (z), we obtain x(t0 ) = − m
j=1 cj x(t0 + jτ ).
5.2
Bounding growth outside the interval
Here we show signals with sparse Fourier transform cannot grow too quickly outside the interval.
Lemma 5.5. Let x(t) be a k-Fourier-sparse signal. For any T > 0 and any t > T ,
|x(t)|2 ≤ k 7 · (2kt/T )2.5k · kxk2T .
Proof. For any t > T , let t = t0 + n · τ such that t0 ∈ [0,P
T /k], τ ∈ [0, T /k] and n ≤
bi = vi e2πifi t0 , and zi = e2πifi τ such that x(t0 + n · τ ) = j∈[k] bj zjn .
By Lemma 5.3, we have for any z1 , z2 , · · · , zk and any n,
zn ≡
k−1
X
ai z i
(mod
i=0
2kt
T .
We define
k
Y
(z − zi )),
i=1
where |ai | ≤ 2k · nk , ∀i ∈ {0, 1, · · · , k − 1}. Thus, we obtain
x(t0 + nτ ) =
k
X
bj zjn
=
j=1
From the fact that x(t0 + i · τ ) =
P
j=1
i
j∈[k] bj zj ,
x(t0 + nτ ) =
k−1
X
i=0
ai
k
X
bj (
k−1
X
ai zji ).
i=0
we simplify it to be
k
X
bj zji =
j=1
k−1
X
i=0
ai x(t0 + i · τ ).
Because (t0 +i·τ ) ∈ [0, T ] for any i = 0, · · · , k−1, we have |x(t0 +iτ )|2 ≤ max |x(t)|2 . k 4 log3 kkxk2T
t∈[0,T ]
from Lemma 5.1. Hence
|x(t0 + n · τ )|2 ≤ k
≤ k
k−1
X
i=0
k−1
X
i=0
7
|ai |2 · |x(t0 + i · τ )|2
n2.2k · max |x(t)|2
t∈[0,T ]
≤ k · (2kt/T )2.2k kxk2T .
Thus, we complete the proof.
22
6
Hash Functions and Filter Functions
6.1
Permutation function and hash function
We first review the permutation function Pσ,a,b and the hash function hσ,b in [PS15], which translates
discrete settings to the continuous setting.
Definition 6.1. For any signal x(t) : R → C and a, b, σ ∈ R, let (Pσ,a,b x)(t) = x σ(t − a) e−2πiσbt .
1 −2πiσaf
1 −2πiσa(f /σ+b)
x
b(f ) and P\
x
b(f /σ + b)
Lemma 6.2. P\
σ,a,b x(σ(f − b)) = σ e
σ,a,b x(f ) = σ e
For completeness, we provide a proof of Lemma 6.2 in Appendix A.4.
Definition 6.3. [PS15] Let πσ,b (f ) = 2πσ(f − b) (mod 2π) and hσ,b (f ) = round(πσ,b (f ) ·
the hash function that maps frequency f ∈ [−F, F ] into bins {0, · · · , B − 1}.
B
2π )
be
1
2
Claim 6.4. [PS15] For any ∆ > 0, let σ be a sample uniformly at random from [ B∆
, B∆
].
(B−1)∆
+
−
+
−
(I) If ∆ ≤ |f − f | <
, then Pr[hσ,b (f ) = hσ,b (f )] = 0
2
(B−1)∆
+
−
(II) If
≤ |f − f |, then Pr[hσ,b (f + ) = hσ,b (f − )] . B1
2
From previous work [HIKP12b, HIKP12a, PS15], uniformly sampling from [A, 2A] for some large
A ≥ Te provides an almost uniform sample on [0, Te] when taken modulo over Te.
Lemma 6.5. For any Te, and 0 ≤ e
, δe ≤ Te, if we sample σ
e uniformly at random from [A, 2A], then
h
i 2e
4e
2e
2e
≤ Pr σ
e (mod Te) ∈ [δe − e
, δe + e
] ≤
(15)
−
+ .
A
A
Te
Te
6.2
Filter function
b )) and (G(t), G(f
b )), the details of proofs are
We state the properties of filter function (H(t), H(f
presented in Appendix C.1 and C.2.
Lemma 6.6. Given s0 , s1 , 0 < s3 < 1, ` > 1, 0 < δ < 1, where ` = Θ(k log(k/δ)).The filter function
b )) has the following properties,
(H(t), H(f
1
2
H(t) ∈ [1 − δ, 1], when |t| ≤ ( − )s3 .
2 s1
2
1
1
Property II : H(t) ∈ [0, 1], when ( − )s3 ≤ |t| ≤ s3 .
2 s1
2
|t| 1
1
Property III : H(t) ≤ s0 · (s1 ( − ) + 2)−` , ∀|t| > s3 .
s3 2
2
s
`
s
`
b )) ⊆ [− 1 , 1 ].
Property IV : supp(H(f
2s3 2s3
Property I :
For any exact k-Fourier-sparse signal x∗ (t), we shift the interval from [0, T ] to [−1/2, 1/2] and
consider x∗ (t) for t ∈ [−1/2, 1/2] to be our observation, which is also x∗ (t) · rect1 (t).
Z +∞
Z +∞
2
∗
Property V :
x (t) · H(t) · (1 − rect1 (t)) dt < δ
|x∗ (t) · rect1 (t)|2 dt.
−∞
+∞
Property VI :
Z
−∞
−∞
|x∗ (t) · H(t) · rect1 (t)|2 dt ∈ [1 − , 1] ·
for arbitrarily small constant .
23
Z
+∞
−∞
|x∗ (t) · rect1 (t)|2 dt.
0.06
0.05
0.04
0.03
0.02
0.01
0.00
0.0140
Given signal (Frequency domain)
c(f)
H
x ∗ (f)
f = ± s1 `/(2s3 )
c
30
1.0
0.8
0.6
0.4
0.2
0.0
0.2 4
20
10
0
10
20
Fourier transform (Time domain)
30
40
H(t)
x ∗ (t)
t = ± 0.5
t = ± 0.5s3
t = ± (0.5−2/s1 )s3
3
1
2
0
1
2
3
4
b )) with a k-Fourier-sparse signal. The property I, II and III
Figure 2: The filter function (H(t), H(f
are presented in the bottom one, the property IV is presented in the top one.
b ))[B, δ, α, l]
Lemma 6.7. Given B > 1, δ > 0, α > 0, we set l = Ω(log(δ/k)). The filter function (G(t), G(f
satisfies the following properties,
b ) ∈ [1 − δ/k, 1], if |f | ≤ (1 − α) 2π .
G(f
2B
2π
2π
b ) ∈ [0, 1], if (1 − α)
Property II : G(f
≤ |f | ≤
.
2B
2B
b ) ∈ [−δ/k, δ/k], if |f | > 2π .
Property III : G(f
2B
l −B l B
Property IV : supp(G(t)) ⊂ [ ·
, ·
].
2 πα 2 πα
Property V : max|G(t)| . poly(B, l).
Property I :
t
6.3
HashToBins
(j)
b (j) (f ), then show the result returned by Procedure
We first define two functions Gσ,b (t) and G
σ,b
HashToBins in Algorithm 6 satisfying some nice properties. The details of proofs are presented
in Appendix C.4.
24
π
−B
− (1−α)π
B
(1−α)π
B
π
B
b [PS15]
Figure 3: G and G.
Definition 6.8. ∀σ > 0, b and j ∈ [B]. Define,
1
G(t/σ)e2πit(j/B−σb)/σ
σ
X
b (j) (f ) = G
b dis ( j − σf − σb) =
b + j − σf − σb)
G
G(i
σ,b
B
B
(j)
Gσ,b (t) =
i∈Z
Lemma 6.9. Let u ∈ CB be the result of HashToBins under permutation Pσ,a,b , and let j ∈ [B].
Define
b (j) ,
zb = x[
·H ·G
σ,b
so
(j)
z = (x · H) ∗ Gσ,b .
Let vector u
b ∈ CB denote the B-dimensional DFT of u, then ∀j ∈ [B],
7
u
b[j] = zσa .
Frequency Recovery
The goal of this section is to prove Theorem 2.6, which is able to recover the frequencies of a signal
x∗ has k-sparse Fourier transform under noise.
Theorem 2.6. Let x∗ (t) =
k
P
vj e2πifj t and x(t) = x∗ (t) + g(t) be our observable signal where
j=1
kg(t)k2T ≤ ckx∗ (t)k2T for a sufficiently small constant c. Then Procedure FrequencyRecoveryKCluster returns a set L of O(k) frequencies that covers all heavy clusters of x∗ , which
uses poly(k, log(1/δ)) log(F T ) samples and poly(k, log(1/δ)) log2 (F T ) time. In particular, for ∆ =
poly(k, log(1/δ))/T and N 2 := kg(t)k2T + δkx∗ (t)k2T , with probability 1 − 2−Ω(k) , for any f ∗ with
Z
f ∗ +∆
f ∗ −∆
there exists an fe ∈ L satisfying
|x[
· H(f )|2 df ≥ T N 2 /k,
√
|f ∗ − fe| . ∆ ∆T .
25
(16)
7.1
Overview
We give an overview of proving Theorem 2.6. Instead of starting with k-cluster recovery, we first
show how to achieve one-cluster recovery.
P
One-cluster recovery. we start with x∗ (t) = kj=1 vj e2πifj t where there exists f0 and ∆ such
that fj is in [f0 − ∆, f0 + ∆] for each j ∈ [k] and consider its properties for frequency recovery.
Definition 7.1 ((, ∆)-one-cluster signal). We say that a signal z(t) is an (, ∆)-one-cluster signal
around f0 iff z(t) and zb(f ) satisfy the following two properties:
Z +∞
Z f0 +∆
2
|b
z (f )|2 df
|b
z (f )| df ≥ (1 − )
Property I :
f0 −∆
T
Property II :
Z
0
|z(t)|2 dt ≥ (1 − )
Z
−∞
+∞
−∞
|z(t)|2 dt.
The main result of one-cluster recovery is to prove that the two properties in Definition 7.1 with
a sufficiently small constant are sufficient to return fe0 close to f0 with high probability, which
provides a black-box for k-cluster recovery algorithm.
We first prove that the pair of conditions, Property I and Property II in Definition 7.1, are
sufficient to obtain an estimation of e2πif0 in Section 7.2. We also provide the proof of the correctness
of Procedures GetLegal1Sample and GetEmpirical1Engergy in Section 7.2.
Lemma 7.2. For a sufficiently small constant > 0, any f0 ∈ [−F, F ], and ∆ > 0, given βb h √1
∆ ∆T
and an (, ∆)-one-cluster signal z(t) around f0 , Procedure GetLegal1Sample in Algorithm 3 with
any β ≤ 2βb takes O((T ∆)3 ) samples to output α ∈ R satisfying
|z(α + β) − z(α)e2πif0 β | ≤ 0.08(|z(α)| + |z(α + β)|),
with probability at least 0.6.
The following lemma shows that for any (, ∆)-one-cluster signal z(t) around f0 , we could use
the above procedure to find a frequency fe0 approximating f0 with high probability.
Lemma 7.3. For a sufficiently small constant > 0, any f0 ∈ [−F, F ], and ∆ > 0, given an (, ∆)one-cluster signal z(t) around √f0 , Procedure FrequencyRecovery1Cluster in Algorithm 4
returns fe0 with |fe0 − f0 | . ∆ · ∆T with probability at least 1 − 2−Ω(k) .
We provide a proof of Lemma 7.3 in Section 7.4. We show z(t) = (x∗ (t) + g(t)) · H(t) satisfy
Properties I and II (Definition 7.1) when all frequencies in x
b∗ are in a small range in Section 7.3.
P
Lemma 7.4. For any f0 ∈ [−F, F ], ∆0 > 0, and x∗ (t) = kj=1 vj e2πift with |fj − f0 | ≤ ∆0 for all
j ∈ [k], let x(t) = x∗ (t) + g(t) be our observable signal whose noise kgk2T ≤ ckx∗ k2T for a sufficiently
b = ∆h . Then
small constant c and H(t) be the filter function defined in Section 6 with | supp(H)|
√
0
z = H · x is an (O( c), ∆h + ∆ )-one-cluster signal around f0 .
From all discussion above, we summarize the result of frequency recovery when x
b∗ is in one
cluster.
Pk
2πifj t with |f − f | ≤ ∆0
Theorem 7.5. For any f0 ∈ [−F, F ], ∆0 > 0, and x∗ (t) =
j
0
j=1 vj e
for all j ∈ [k], let x(t) = x∗ (t) + g(t) be our observable signal whose noise kgk2T ≤ ckx∗ k2T for a
b =
sufficiently small constant c and H(t) be the filter function defined in Section 6 with | supp(H)|
0
∆h . Then Procedure FrequencyRecovery1Cluster in Algorithm 4 with ∆ = ∆ + ∆h takes
poly(k, log(1/δ)) · log(F T )√samples, runs in poly(k, log(1/δ)) · log2 (F T ) time, returns a frequency
fe0 satisfying |fe0 − f0 | . ∆ ∆T with probability at least 1 − 2−Ω(k) .
26
P
k-cluster recovery. Given any x∗ (t) = kj=1 vj e2πifj t , we plan to convolve the filter function
G(t) on x(t) · H(t) and use Lemma 7.3 as a black box to find a list of frequencies that covers
{f1 , · · · , fk }.
1
2
We fix ∆ = poly(k, log(1/δ))/T , B = Θ(k) and sample σ uniformly at random from [ B∆
, B∆
]
∗
for k-cluster recovery. We will cover all f ∈ [−F, F ] with the following property :
Z f ∗ +∆
|x[
· H(f )|2 df ≥ T N 2 /k,
(17)
f ∗ −∆
We consider one frequency f ∗ ∈ [−F, F ] satisfying (17) and use j = hσ,b (f ∗ ) to denote its index in
[B] after hashing (σ, b). Recall that for j ∈ [B], any σ > 0 and any b,
(j)
Gσ,b (t) =
X
1
b (j) (f ) =
b + j − σf − σb).
G(t/σ)e2πit(j/B−σb)/σ such that G
G(i
σ,b
σ
B
i∈Z
b (j) and z = (x · H) ∗ G(j) for f ∗ and j = hσ,b (f ∗ ). In Section 7.5, we show that
We set zb = x[
·H ·G
σ,b
σ,b
with high probability over the hashing (σ, b), (z, zb) satisfies Property I with [f ∗ − ∆, f ∗ + ∆] and
Property II in Definition 7.1 such that we could use Lemma 7.3 on z to recover f ∗ .
Lemma 7.6. Let f ∗ ∈ [−F, F ] satisfy (17). For a random hashing (σ, b), let j = hσ,b (f ∗ ) be the
(j)
b (j) . With
bucket that f ∗ maps to under the hash such that z = (x · H) ∗ Gσ,b and zb = x[
·H ·G
σ,b
∗
probability at least 0.9, z(t) is an (, ∆)-one-cluster signal around f .
Combining Lemma 7.6 and Lemma 7.3, we could recover any heavy frequency f ∗ satisfying (17)
with probability at least 0.8. Then we repeat this procedure to guarantee that we cover all heavy
frequencies and finish the proof of the main frequency recovery Theorem 2.6 in Section 7.6.
7.2
Analysis of GetLegal1Sample and GetEmpirical1Energy
Let I = [f0 − ∆, f0 + ∆] and I = (−∞, +∞) \ I in this proof. We define z I (t), zbI (f ) and
z I (t), zbI (f ) as follows:
(
(
z
b
(f
)
if
f
∈
I
0
if f ∈ I
, zbI (f ) =
zbI (f ) =
0
if f ∈ I
zb(f )
if f ∈ I
We consider z I (t) as the “signal” to recover f0 and treat z I (t) as the “noise”. We first show some
basic properties of z I (t).
RT
R +∞
Claim 7.7. For z I (t), we have 0 |z I (t)|2 dt ≤ −∞ |z(t)|2 dt. For z I (t), we have
Z
0
T
√
|z I (t)|2 dt ≥ (1 − 5 )
Z
+∞
−∞
|z(t)|2 dt and
Z
0
T
√
|z I (t)|2 dt ≥ (1 − 6 )
Z
+∞
−∞
|z I (t)|2 dt.
Proof. From the definition and Property I in Definition 7.1, we know
Z +∞
Z +∞
I
I
2
I
z(t) = z (t) + z (t) and
|b
z (f )| df ≤
|b
z (f )|2 df.
−∞
−∞
Notice that Property I(in Definition 7.1) indicates that
Z T
Z +∞
Z +∞
Z
I
2
I
2
I
2
|z (t)| dt =
|b
z (f )| df ≤
|z (t)| dt ≤
0
−∞
−∞
27
+∞
−∞
|b
z (f )|2 df.
On the other hand, from Property II(in Definition 7.1), we know
(1−)
Z
+∞
2
−∞
|z(t)| dt ≤
RT
Z
T
I
2
|z (t)+z (t)| dt ≤
0
We have 0 |z I (t)|2 dt ≤ 2
we bound
I
R +∞
Z
T
I
0
2
|z (t)| dt+2
Z
0
T
I
|z (t)|·|z (t)|dt+
|z(t)|2 dt from the above inequality. From
−∞
Z
√ Z
|z (t)| · |z (t)|dt ≤ 2
T
I
+∞
|z(t)|2 dt
I
−∞
0
I
RT
0
Z
T
0
|z I (t)|2 dt ≤
|z I (t)|2 dt.
R +∞
−∞
|b
z (f )|2 df ,
by the Cauchy-Schwartz inequality and have
Z
T
I
|z (t)| dt ≥ (1 − 5 )
0
Because
R +∞
−∞
R +∞
|z I (t)|2 dt ≤
−∞
Z
√
2
+∞
−∞
|z(t)|2 dt.
(18)
|z(t)|2 dt, inequality (18) also indicates that
T
0
Z
I
√
2
|z (t)| dt ≥ (1 − 6 )
Z
+∞
−∞
|z I (t)|2 dt.
One useful property of z I (t) is that its maximum can be bounded by its average on [0, T ].
√
Claim 7.8. ∀t ∈ [0, T ], |z I (t)| ≤ 2 ∆T · kz I kT .
R f +∆
Proof. From the definition |z I (t)|, it is upper bounded by f00−∆ |zbI (f )|df for any t ∈ [0, T ]. On
the other hand,
Z
f0 +∆
f0 −∆
I
|b
z (f )|df ≤
=
√
2∆(
√
2∆(
Z
f0 +∆
f0 −∆
Z +∞
−∞
Z T
|b
z I (f )|2 df )1/2
|z I (t)|2 dt)1/2
√
≤ 2 ∆(
|z I (t)|2 dt)1/2
0
√
= 2 ∆T kz I kT .
Claim 7.9. Given βb =
in z I (t), we have that
Cβ
√
∆· ∆T
b
with a sufficiently small constant Cβ , for any two β-close
samples
b 2β],
b
∀α ∈ [0, T ], ∀β ∈ [β,
|z I (α)e2πif0 β − z I (α + β)| ≤ 0.01 · kz I kT .
28
Proof. From the definition of the Fourier transform, we have
I
I
|z (a + β) − z (a)e
2πif0 β
|=
Z
f0 +∆
f0 −∆
zbI (f )e2πi(f a+f0 β) (e2πi(f −f0 )β − 1)df
≤ 2 · (2π∆β) ·
≤ 4πβ∆ ·
√
Z
2∆
f0 +∆
f0 −∆
Z
|zbI (f )|df
f0 +∆
f0 −∆
b ·
≤ 10π β∆
√
2∆
Z
T
I
2
|z (t)| dt
0
≤ 10−2 kz I kT .
|zbI (f )|2 df
21
12
by Taylor expansion
by Hölder inequality
by inequality (18)
We consider how to output an α such that e2πif0 β ≈ z(α + β)/z(α) with high probability in the
rest of this section.
If we can sample from z I (t), we already know |z I (α)e2πif0 β − z I (α + β)| ≤ 0.01kz I kT from
Claim 7.9.
to find any α such that |z I (α)| ≥ 0.5kz I kT . From Claim 7.8, we can
√ Then it is enough
I
take O( ∆T ) samples z (α), z I (α + β) where each α is uniformly sampled from [0, T ] such that
with high probability, the sample z I (α) with the largest norm |z I (α)| satisfies |z I (α)| ≥ 0.5kz I kT .
Then we have e2πif0 β ≈ z I (α + β)/z I (α).
Next, we move to z(t) = z I (t) + z I (t) and plan to output α ∈ [0, T ] with probability at least
0.5 such that |z I (α)| ≤ 0.1|z I (α)| and |z I (α + β)| ≤ 0.1|z I (α + β)|. Because the “noise” z I (t) has
√
kz I (t)k2T ≥ kz I (t)k2T for a constant and the bound ∆T in Claim 7.8 is a polynomial in k, the
approach for z I (t) cannot guarantee that z(α + β)/z(α) ≈ e2πif0 β with probability more than 1/2.
The key observation is as follows:
Observation 7.10. For a sufficiently small and kz I k2T ≤ kzk2T , let DT be the weighted distribution
2
on [0, T ] according to |z(t)|2 , i.e., DT (t) = T|z(t)|
. If we sample α ∈ [0, T ] from the distribution DT
kzk2
T
instead of the uniform distribution on [0, T ], |z I (α)| ≤ 0.01|z I (α)| with probability 0.9.
It follows from the fact that
E
α∼DT
|z I (α)|2
=
|z(α)|2
Z
0
T
|z I (α)|2 |z(α)|2
·
dα =
|z(α)|2 T kzk2T
RT
0
|z I (α)|2 dα
≤ .
T kzk2T
In Procedure GetLegal1Sample, we collect (∆T )2 samples (in expectation) z(α), z(α + β)
in Sheavy with |z(α)| ≥ 0.49kzkT and resample one α from these samples according to their norm
|z(α)|2 + |z(α + β)|2 . We show its correctness as follows.
Because we do not know 0.5kzkT , we use zemp to approximate it.
Claim 7.11. Procedure GetEmpirical1Energy in Algorithm 3 takes O((T ∆)2 ) samples to output
zemp such that zemp ∈ [0.8kzkT , 1.2kzkT ] with prob. 0.9.
2
Proof. We know zemp
= Ei∈[Rest ] [|z(αi )|2 ] = Ei∈[Rest ] [|z I (αi ) + z I (αi )|2 ].
Notice that Ei∈[Rest ] [|z I (αi )|2 ] is in [0.99kz I kT , 1.01kz I kT ] with prob. 0.99 from the Chernoff
bound and Claim 7.8.
29
At the same time, Eαi [|z I (αi )|2 ] = kz I k2T . With prob. 0.92, Ei∈[Rest ] [|z I (αi )|2 ] ≤ 13kz I k2T . For
a sufficiently small and kz I k2T ≤ kz I k2T , Ei∈[Rest ] [|z I (αi )|2 ] ≤ 13kz I k2T .
At last, we bound the cross terms of |z I (αi ) + z I (αi )|2 by the Cauchy-Schwartz inequality,
E [|z I (αi )z I (αi )| + |z I (αi )z I (αi )|]
i∈Rest
≤ 2 E [|z I (αi )| · |z I (αi )|]
i∈Rest
≤2
E
i∈[Rest ]
[|z I (αi )|2 ] ·
√
≤ 10 kz I k2T .
E
i∈[Rest ]
1/2
[|z I (αi )|2 ]
For a sufficiently small , we have Ei∈[Rest ] [|z(αi )|2 ]1/2 is in [0.9kz I kT , 1.1kz I kT ], which is also in
[0.8kzkT , 1.2kzkT ] because of Property II.
We assume zemp ∈ [0.8kzkT , 1.2kzkT ] and focus on U = {t ∈ [0, T ] |z(t)| ≥ 0.5zemp }. Notice
that
Z
Z T
Z
Z T
2
2
2
2
|z(t)| dt =
|z(t)| dt −
|z(t)| dt ≥ (1 − 0.6 )
|z(t)|2 dt.
U
0
[0,T ]\U
0
Let Rheavy = |Sheavy |. From Claim 7.8 and , E[Rheavy ] ≥ Rrepeat /(T ∆). So we assume Rheavy ≥
0.01Rrepeat /(T ∆) = 0.01(T ∆)2 in the rest of this section and think each αi ∈ Sheavy is a uniform
sample from U over the randomness on Sheavy .
P
P
Claim 7.12. With probability 0.95, i∈Sheavy (|z I (αi )|2 + |z I (αi + β)|2 ) ≤ 10−4 i∈Sheavy (|z(αi )|2 +
|z(αi + β)|2 ) for a sufficiently small and kz I k2T ≤ kzk2T .
Proof. At first,
E
Sheavy
X
i∈Sheavy
(|z(αi )| + |z(αi + β)| ) ≥ Rheavy · E [|z(t)|2 ] = Rheavy ·
2
2
t∼U
R
U
|z(t)|2 dt
.
|U |
At the same time,
R0
X
2 T |z I (t)|2 dt
I
2
I
2
I
2
I
2
E
.
[|z (αi )| + |z (αi + β)| ] = Rheavy · E [|z (t)| + |z (t + β)| ] ≤
Sheavy
t∼U
|U |
i∈Sheavy
From
R
U
|z(t)|2 dt ≥ 0.64
RT
0
|z(t)|2 dt and
R0
T
|z I (t)|2 dt ≤
RT
0
|z(t)|2 dt, we get the conclusion.
We assume all results in the above claims hold and prove that the sample from Sheavy is a good
sample such that z I (α) is small.
Claim 7.13. If we sample i ∈ Sheavy according to the weight |z(αi )|2 + |z(αi + β)|2 , with prob. at
least 0.9, |z I (αi )| + |z I (αi + β)| ≤ 0.05(|z(αi )| + |z(αi + β)|).
30
Proof. Similar to the proof of the key observation, we compute the expectation of
over the sampling in Sheavy :
X
|z I (αi )|2 + |z I (αi + β)|2
|z(α )|2 + |z(αi + β)|2
P i
·
|z(αj )|2 + |z(αj + β)|2 |z(αi )|2 + |z(αi + β)|2
i∈Sheavy
P
j∈Sheavy
i∈Sheavy
P
=
|z I (αi )|2 +|z I (αi +β)|2
|z(αi )|2 +|z(αi +β)|2
|z I (αi )|2 + |z I (αi + β)|2
i∈Sheavy
|z(αi )|2 + |z(αi + β)|2
≤ 10−4 .
By Markov’s inequality, when we sample i ∈ Sheavy according to the weight |z(αi )|2 + |z(αi + β)|2 ,
|z I (αi )|2 +|z I (αi +β)|2
≤ 10−3 with probability
|z(αi )|2 +|z(αi +β)|2
|z I (αi + β)| ≤ 0.05(|z(αi )| + |z(αi + β)|).
0.9. We have that with prob. at least 0.9, |z I (αi )| +
We assume all above claims hold and finish the proof by setting α = αi . From Claim 7.9, we
know that
|z I (α)e2πif0 β − z I (α + β)| ≤ 0.01 · E [|z I (α)|2 ]1/2 ≤ 0.03|z I (α)|.
t∈[0,T ]
Now we add back the noise z I (α) and z I (α + β) to get
|z(α)e2πif0 β −z(α+β)| ≤ |z I (α)e2πif0 β −z I (α+β)|+|z I (α)|+|z I (α+β)| ≤ 0.08(|z(α)|+|z(α+β)|).
7.3
A cluster of frequencies, times H, is a one-cluster signal per Definition 7.1
The goal of this section is to prove Lemma 7.4. Without loss of generality, we assume g(t) = 0 for
b ∗x
any t ∈
/ [0, T ] and notice that supp(H
b∗ ) ⊆ f0 + [−∆, ∆] for ∆ = ∆0 + ∆h from the definition of
b From the Property VI (presented in Lemma 6.6) of (H, H),
b
H.
Z T
Z +∞
∗
2
|x (t)| dt = (1 ± c)
|H(t) · x∗ (t)|2 dt.
−∞
0
b we bound the energy of g · H:
From the first two properties of (H, H),
Z +∞
Z T
|H(t) · g(t)|2 dt ≤ (1 + c)
|g(t)|2 dt.
−∞
0
Let z(t) = (x∗ (t) + g(t))H(t). We use the triangle inequality on the above two inequalities:
Z T
|z(t)|2 dt
0
Z T
Z T
Z T
∗
2
2
≥
|H(t) · x (t)| dt −
|H(t) · g(t)| dt − 2
|H(t) · x∗ (t)| · |H(t) · g(t)|dt
0
0
0
s
Z T
Z T
Z T
Z T
∗
2
2
2
2
|g(t)| dt
|x∗ (t)|2 dt·
≥ (1 − c)
|x (t)| dt − (1 + c)
|g(t)| dt − 2 (1 + c)
0
√
≥ 1−5 c
Z
0
0
0
T
|x∗ (t)|2 dt,
31
0
where we use the Cauchy-Schwarz inequality and
Similarly,
Z +∞
|z(t)|2 dt
RT
0
|g(t)|2 dt ≤ c
−∞
≤ (1 + c)
Z
T
0
√
≤ (1 + 5 c)
∗
2
|x (t)| dt + (1 + c)
Z
0
T
Z
0
T
2
|g(t)| dt + 2
s
(1 +
c)2
Z
0
RT
0
|x∗ (t)|2 dt in the last step.
T
|x∗ (t)|2 dt
Z
0
T
|g(t)|2 dt
|x∗ (t)|2 dt.
Hence we obtain Property II(in Definition 7.1) when c is sufficiently small.
Then we observe that
Z f0 +∆h
|b
z (f )|2 df
f0 −∆h
f0 +∆h
≥
≥
≥
=
Z
f0 −∆h
Z f0 +∆h
f0 −∆h
Z
f0 +∆h
f0 −∆h
Z
+∞
−∞
|H ·\
(x∗ + g)|2 df
\
[
\
[
|H
· x∗ |2 − | H
· g|2 − 2|H
· x∗ | · | H
· g|df
sZ
Z
+∞
\
|H
· x∗ |2 df −
|H · x∗ |2 dt −
Z
−∞
[
|H
· g|2 df − 2
+∞
−∞
|H · g|2 dt − 2
f0 +∆h
sZ
√ Z
(1 − c) − c(1 + c) − 3 c +∞
√
≥
|z(t)|2 dt.
1+5 c
−∞
f0 −∆h
f0 +∆h
f0 −∆h
\
|H
· x∗ |2 df
|H · x∗ |2 dt
Z
Z
+∞
−∞
[
|H
· g|2 df
+∞
−∞
|H · g|2 dt
Thus we have Property I(in Definition 7.1) for z.
7.4
Frequency recovery of one-cluster signals
The goal of this section is prove Theorem 7.5. We first show the correctness of Procedure Locate1Inner. Second, we analyze the Procedure Locate1Signal. At end, we rerun Procedure
Locate1Signal and use median analysis to boost the constant success probability.3
st
st
Lemma 7.14. Let f0 ∈ region(q 0 ). Let β is sampled from [ 4∆
, 2∆l
] and let γ denote the output of
Procedure GetLegal1Sample in Algorithm 4. Then using the pair of samples z(γ + β) and z(γ),
we have
I. for the q 0 with probability at least 1 − s, vq0 will increase by one.
II. for any q such that |q − q 0 | > 3, with probability at least 1 − 15s, vq will not increase.
b
Proof. We replace f0 by θ in the rest of the proof. By Lemma 7.2, we have that for any βb ≤ β ≤ 2β,
Procedure GetLegal1Sample outputs a γ ∈ [0, T ] satisfying
|z(γ + β) − z(γ)e2πif0 β | ≤ 0.1(|z(γ)| + |z(γ + β)|)
with probability at least 0.6.
3
The proofs in this section are identical to [HIKP12b] and [PS15].
32
Furthermore, there exists such some constant g ∈ (0, 1) such that with probability 1 − g,
1
. sin−1 ( ),
g
kφ(z(γ + β)) − (φ(z(γ)) − 2πβθ)k
where kx − yk
= min|x − y + 2πz| denote the “circular distance” between x and y. We can set
z∈Z
s = Θ(g −1 ). There exists some constant p = Θ(s), with probability at least 1 − p,
ko − 2πβθk
< sπ/2
where o := φ(z(γ + β)/z(γ)). The above equation shows that o is a good estimate for 2πβθ with
good probability. We will now show that this means the true region Qq0 gets a vote with large
probability.
q 0 −1
q0
q 0 −0.5
∆l
∆l
0
For each q 0 with θ ∈ [l− ∆l
2 + t ∆l, l− 2 + t ∆l] ⊂ [−F, F ], we have that θq = l− 2 +
t ∆l
satisfies that
∆l
θ − θq 0 ≤
.
2t
b 2β],
b then 2βb = st ≤ cT 3 (Note that A is
Note that we sample β uniformly at random from [β,
2∆l
some constant > 1), which implies that 2πβ ∆l
2t ≤
to the true region in the following sense,
sπ
2 .
10A 2
Thus, we can show the observation o is close
ko − 2πβθq0 k
≤ ko − 2πβθk + k2πβθ − 2πβθq0 k
sπ
≤
+ 2πkβθ − βθq0 k
2
≤ sπ.
by triangle inequality
Thus, vq0 will increase in each round with probability at least 1 − s.
On the other side, consider q with |q − q 0 | > 3. Then |θ − θq | ≥ 7∆l
2t , and (assuming β ≥
have
st
sπt
7sπ
3sπ
2πβ|θ − θq | ≥ 2π
|θ − θq | =
|θ − θq | ≥
>
.
4∆l
2∆l
4
2
∆l
There are two cases: |θ − θq | ≤ ∆l
st and |θ − θq | > st .
First, if |θ − θq | ≤ ∆l
st . In this case, from the definition of β it follows that
2πβ|θ − θq | ≤
st
4∆l )
we
sπt
|θ − θq | ≤ π
∆l
Combining the above equations implies that
3s
3s
Pr 2πβ(θ − θq ) (mod 2π) ∈ [− 2π, 2π] = 0
4
4
3s
3s
Second, if |θ −θq | > ∆l
st . We show this claim is true : Pr[2πβ(θ −θq ) (mod 2π) ∈ [− 4 2π, 4 2π]] . s.
b
To prove it, we apply Lemma 6.5 by setting Te = 2π, σ
e = 2πβ, δe = 0, = 3s
4 2π, A = 2π β,
∆f = |θ − θq |. By upper bound of Lemma 6.5, the probability is at most
2e
4e
3s
3s
3s
+
=
+
≤
+
b
A∆f
2
2
Te
β∆f
3s
st ∆l
4∆l st
Then in either case, with probability at least 1 − 15s, we have
k2πβθq − 2πβθk
which implies that vq will not increase.
33
>
3s
2π
4
< 15s
Lemma 7.15. Procedure Locate1Inner in Algorithm 4 uses Rloc “legal” samples, and then after Procedure Locate1Signal in Algorithm 4 running Procedure Locate1Inner Dmax times, it
outputs a frequency fe0 such that
√
|fe0 − f0 | . ∆ · T ∆
with arbitrarily large constant probability.
Proof. For each observation, vq0 incremented with probability at least 1 − p and vq is incremented
with probability at most 15s + p for |q − q 0 | > 3. The probabilities corresponding to different
observations are independent. Then after Rloc observations, there exists some constant c < 21 , for
any q such that |q − q 0 | > 3,
Pr[False region gets more than half votes]
= Pr[vj,q > Rloc /2]
Rloc
≤
(15s + p)Rloc /2
Rloc /2
≤ cΩ(Rloc )
Similarly, on the other side,
Pr[True region gets less than half votes]
= Pr[vj,q0 < Rloc /2]
Rloc
≤
(p)Rloc /2
Rloc /2
≤ cΩ(Rloc )
Taking the union bound over all the t regions, it gives with probability at least 1 − tf Ω(Rloc ) we can
find some region q such that |q − q 0 | < 3.
If we repeat the above procedure Dmax rounds, each round we choose the “False” region with
probability at most 1 − tcΩ(R√loc ) . Thus, taking the union bound over all the Dmax rounds, we will
report a region has size h ∆ ∆T and contains f0 with probability at least 1 − Dmax tcΩ(Rloc ) .
The reason for not ending up with region that has size h ∆ is, the upper bound of the sample
range of β force us to choose β is at most . T 3 by Claim 7.9
(∆T ) 2
It remains to explain how to set Dmax , t, and Rloc . At the beginning of the first round, we start
with frequency interval√of length 2F , at the beginning of the last round, we start with frequency
interval of length t · ∆ T ∆. Each round we do a t-ary search, thus
Dmax = logt (
2F
√
) ≤ logt (F/∆).
t∆ T ∆
We can set Rloc h log1/c (t/c) and t > Dmax , e.g. t = log(F/∆). Thus, the probability becomes,
1 − Dmax tcΩ(Rloc ) ≥ 1 − t2 cΩ(Rloc ) ≥ 1 − poly(1/t, c)
which is larger than any constant probability.
Using the same parameters setting in the proof of Lemma 7.15, we show the running time and
sample complexity of Procedure Locate1Signal,
34
Lemma 7.16. Procedure Locate1Signal in Algorithm 4 uses
O(poly(k, log(1/δ))) · log(F T ) samples and runs in O(poly(k, log(1/δ))) · log2 (F T ) time.
Proof. The number of “legal” observations is
Dmax Rloc = O(logt (F/∆) log1/c (t/c)) = O(log(F/∆))
The total number of samples is
Rest + Rrepeat Dmax Rloc = O(T ∆h )2 + (T ∆h )3 · log(F T ) = poly(k, log(1/δ)) · log(F T )
where the first step follows by Claim 7.11 and Lemma 7.2 and the last step follows by the setting
of ∆h in Appendix C.3.
The running time includes two parts, one is approximately computing H(t) for all the samples,
each sample takes poly(k, log(1/δ)) time according to Lemma C.8; the other is for each legal sample
we need to assign vote to some regions.
poly(k, log(1/δ)) · (Rest + Rrepeat Dmax Rloc ) + Dmax Rloc t = poly(k, log(1/δ)) log2 (F T )
Lemma 7.17 only achieves constant success probability, using median analysis we can boost the
success probability,
Lemma 7.17. Let fe0 denote the frequency output by Procedure FrequencyRecovery1Cluster
in Algorithm 5, then with probability at least 1 − 2−Ω(k) ,
√
|fe0 − f0 | . ∆ T ∆
Proof. Because of Procedure FrequencyRecovery1Cluster taking the median of O(k) independent results by repeating algorithm Locate1Signal O(k) times. Each sample Lr is close to fe0
with sufficiently large probability. Thus, using the Chernoff bound will output fe0 with probability
1 − 2−Ω(k) such that
√
|fe0 − f0 | . ∆ T ∆.
Combining Lemma 7.17 with the sample complexity and running time in Lemma 7.15, we are
able to finish the proof of Theorem 7.5.
7.5
The full signal, after multiplying by H and convolving with G, is oneclustered.
The goal of this section is to prove Lemma 7.6. We fix f ∗ ∈ [−F, F ] satisfying (17) in this section.
We first define a good hashing (σ, b) of f ∗ as follows.
Definition 7.18. We say that a frequency f ∗ is well-isolated under the hashing (σ, b) if, for j =
hσ,b (f ∗ ), we have that the signal
b (j)
zb(j) = x[
·H ·G
σ,b
satisfies, over the interval If ∗ = (−∞, ∞) \ (f ∗ − ∆, f ∗ + ∆),
Z
|b
z (j) (f )|2 df . · T N 2 /k.
If ∗
35
For convenience, we simplify z (j) by using z in the rest of this section.
Lemma 7.19. Let f ∗ be any frequency. Then f ∗ is well-isolated by a hashing (σ, b) with probability
1
2
≥ 0.9 given B = Θ(k) and σ ∈ [ B∆
, B∆
] chosen uniformly at random.
Proof. For any other frequency f 0 in x∗ , its contribution in zb depends on how far it is from f ∗ .
Either it is:
• Within ∆ of f ∗ , f 0 and f ∗ will be mapped into the same bucket with probability at least 0.99.
• Between ∆ and 1/σ far, from Claim 6.4, f 0 and f ∗ will always mapped into different buckets.
Hence f 0 always contributes in the δ
Property III in Lemma 6.7 about filter function
k region of
R f 0 +∆
δ
b
(G(t), G(f )), i.e., it contributes at most · 0
|x[
· H|2 df . Overall it will contribute
f −∆
k
δ
·
k
Z
δ
|x[
· H| df =
k
2
Z
|x · H|2 dt.
• More than 1/σ far, in which case they contribute in the same region with probability at most
3/B. By a union bound, it is at most 3k/B ≤ 0.01
∗ · H) = ∅, otherwise we treat it as
Without loss of generality, we assume supp(g[
· H) ∩ supp(x\
∗ · H under G(j) .
a part of x∗ · H. We first consider frequency f ∗ ∈ x\
σ,b
R f ∗ +∆
∗ · H(f )|2 df ≥ T N 2 /k and z
∗·H ·G
b (j) where j =
b = x\
Lemma 7.20. Let f ∗ satisfying f ∗ −∆ |x\
σ,b
hσ,b (f ∗ ). If f ∗ is well-isolated, then z and zb satisfying Property I(in Definition 7.1), i.e.,
Z
T
2
0
|z(t)| dt ≥ (1 − )
Z
+∞
−∞
|z(t)|2 dt.
(j)
Proof. We first notice that z(t) = x∗ (t) · H(t) ∗ Gσ,b (t) and lower bound
Z
+∞
−∞
+∞
=
≥
Z
(j)
|x∗ (t) · H(t) ∗ Gσ,b (t)|2 dt
(j)
∗ · H(f ) · G
b (f )|2 df
|x\
σ,b
−∞
Z f0 +∆
f0 −∆
≥ (1 − δ)
2
2
R0
−∞
|z(t)|2 dt as follows :
by FT
(j)
∗ · H(f ) · G
b (f )|2 df
|x\
σ,b
Z
f0 +∆
f0 −∆
2
∗ · H(f ) |2 df
|x\
≥ (1 − δ) T N /k
Z
δ T ∗ 2
|x (t)| dt
≥ 0.9
k 0
We give an upper bound
proof.
R +∞
2
−∞ |z(t)| dt
+
(19)
R +∞
T
|z(t)|2 dt . kδ
36
RT
0
|x∗ (t)H(t)|2 dt in the rest of this
Consider the case t < 0, by definition of Convolution,
Z +∞
(j)
(j)
(j)
∗
Gσ,b (t − τ ) · (x∗ · H)(τ )dτ
z (t) = x (t) · H(t) ∗ Gσ,b (t) =
−∞
Without loss of generality, we can shift the original signal and H(t) from [0, T ] to [−T /2, T /2],
by Property of H(t), we know that if s3 T /2 ≤ |t| ≤ T /2, then H(t) ≤ 2−OΘ(`) . Note that G(t) is
compact and has support DB, we also assume its compact region is [−DB/2, DB/2] (Recall that
l
D = απ
).
Thus, by definition of convolution,
=
=
≤
z(t)
Z DBσ/2
1
σ
1
σ
≤
(j)
Gσ,b (s) · (x · H)(t − τ )dτ
−DBσ/2
Z DBσ/2
−DBσ/2
DBσ/2
Z
−DBσ/2
1
σ
Z
G(s/σ)e2πis(j/B−σb)/σ · (x · H)(t − τ )dτ
|G(τ /σ)| · |(x · H)(t − τ )|dτ
!
DBσ/2
−DBσ/2
|G(τ /σ)|dτ
·
max
|τ |≤DBσ/2
|(x · H)(t − τ )|
So, if t ∈
/ [−T /2, T /2], then t − s ∈
/ [−T /2 + DBσ/2, T /2 − DBσ/2]. By Property V of G(t),
|G(t)| ≤ poly(k, log(1/δ)). Because of the parameter setting4 , we have the fact [−T s3 /2, T s3 /2] ⊆
[−T /2 + DBσ/2, T /2 − DBσ/2] ⊆ [−T /2, T /2]. Thus, we know T (1 − s3 )/2 > DBσ/2, then for
any t − τ ∈ [−T /2, −T /2 + DBσ/2] ∪ [T /2 − DBσ/2, T /2] = S, then
|z(t)|2 . DBσ ·
2
1
· poly(k, log(1/δ)) · 2−Θ(`) · k 4 · kx∗ (t)k2T . poly(k, log(1/δ)) · 2−Θ(`) · kx∗ (t)k2T .
σ
Thus, taking the integral over S,
Z
|z(t)|2 dt . |S| · 2−Θ(`) poly(k, log(1/δ)) · kx∗ (t)k2 . 2−Θ(`) T kx∗ (t) · H(t)k2T
S
b ), we have
By property of filter function H(t), H(f
t
|(x · H)(t)|2 ≤ ( )−` kx∗ (t) · H(t)k2T if t ≥ 3T
T
Thus for any constant ,
Z
−T /2
−∞
2
|z(t)| dt +
Z
+∞
T /2
2
−`
∗
|z(t)| dt . 2 T kx (t) ·
H(t)k2T
δ
≤ 0.9 ·
k
Z
T /2
−T /2
|x∗ (t)|2 dt
(20)
where the last inequality follows by ` & k log(k/δ). Shifting the interval from [−T /2, T /2] to [0, T ],
the same result is still holding. Combining Equation (19) and (20) completes the proof of Property
II.
4
We will set B to be O(k), D to be poly(k) and σ to be T / poly(k).
37
(j)
We consider frequency f ∗ ∈ g[
· H under Gσ,b and show the energy of noise g(t) is evenly distributed over B bins on expectation.
Lemma 7.21. Given any noise g(t) : [0, T ] → C and g(t) = 0, ∀t ∈
/ [0, T ]. We have, ∀j ∈ [B],
Z +∞
Z +∞
1
(j)
|g(t)H(t) ∗ Gσ,b (t)|2 dt .
E
|g(t)H(t)|2 dt
σ,b
B
−∞
−∞
Proof. Because of Fourier Transform preserves `2 norm, it suffices to prove
Z +∞
Z
1 +∞ [
(j)
2
b
[
|g · H(f ) · Gσ,b (f )| df .
|g · H(f )|2 df
E
σ,b
B −∞
−∞
(j)
b (f ) is a periodic function and outputs at most 1 on O(1/B) fraction of the period, and
Since G
σ,b
outputs ≤ δ on other part. Thus, for any frequency f , we have
h
i
b (j) (f )|2 . 1
E |G
σ,b
σ,b
B
Thus, we have
Z +∞
(j)
2
b (f )| df
E
|g[
· H(f ) · G
σ,b
σ,b
−∞
Z +∞
(j)
2
2
b
[
≤ E
|g · H(f )| · |Gσ,b (f )| df
σ,b
Z
−∞
+∞
b (j) (f )|2 ]df
|g[
· H(f )|2 · E [|G
σ,b
σ,b
−∞
Z +∞
2
(j)
2
b
[
≤
|g · H(f )| df · max E Gσ,b (f )
=
f
−∞
.
1
B
which completes the proof.
Proof of Lemma 7.6.
Z
σ,b
+∞
−∞
|g[
· H(f )|2 df,
Let j = hσ,b (f ∗ ), signal
(j)
b ,
zb = x[
·H ·G
σ,b
(21)
and region If ∗ = (f ∗ − ∆, f ∗ + ∆) with complement If ∗ = (−∞, ∞) \ If ∗ . From Property I of G in
Lemma 6.7, we have that
b (l) (f ) & 1
G
σ,b
for all f ∈ If ∗ , so by (17)
Z
If ∗
|b
z (f )|2 df ≥ T N 2 /k.
On the other hand, f ∗ is will-isolated with probability 0.9:
Z
|b
z (f )|2 df . T N 2 /k.
If ∗
Hence, zb satisfies the Property I(in Definition 7.1) of one-mountain recovery. Combining Lemma 7.20
(j)
and Lemma 7.21, we know that (x∗ · H) ∗ Gσ,b always satisfies Property II(in Definition 7.1) and
R +∞
(j)
2
2
2
−∞ |g(t)H(t) ∗ Gσ,b (t)| dt is less than 20T N /B ≤ T N /k with probability at least 0.95, which
(j)
indicates that z = (x∗ +g)·H ∗Gσ,b satisfies Property II(in Definition 7.1) with probability 0.95.
38
7.6
Frequency recovery of k-clustered signals
The goal of this section is to prove that the frequencies found by Procedure FrequencyRecoveryKCluster in Algorithm 8 have some reasonable guarantee.
We first notice that Lemma 7.6 and Lemma 7.3 imply the following lemma by a union bound.
Lemma 7.22. Let x∗ (t) =
k
P
vj e2πifj t . We observe x(t) = x∗ (t) + g(t), where kg(t)k2T ≤ ckx∗ (t)k2T
j=1
for a sufficiently small constant c and define N 2 := kg(t)k2T +δkx∗ (t)k2T . Then Procedure OneStage
returns a set L of O(k) frequencies that covers the heavy frequencies of x∗ . In particular, for any
f ∗ with
Z f ∗ +∆
|x[
· H(f )|2 df ≥ T N 2 /k,
(22)
f ∗ −∆
there will exist an fe ∈ L satisfying |f ∗ − fe| .
Lemma 7.23. Let x∗ (t) =
k
P
√
T ∆ · ∆T with probability 0.99.
vj e2πifj t and R = O(k). We observe x(t) = x∗ (t) + g(t), where
j=1
kg(t)k2T ≤ ckx∗ (t)k2T for a sufficiently small constant c and choose N 2 := kg(t)k2T + δkx∗ (t)k2T .
Then Algorithm MultipleStages returns a set L of O(k) frequencies that approximates the heavy
frequencies of x∗ . In particular, with probability 1 − 2−Ω(k) , for any f ∗ such that
Z f ∗ +∆
|x[
· H(f )|2 df ≥ T N 2 /k,
(23)
f ∗ −∆
there will exist an fe ∈ L satisfying |f ∗ − fe| .
√
T ∆∆.
Proof. Let A ⊂ [−F, F ] denote the set of frequencies f ∗ satisfying Equation (22). Let A0 ⊂ [−F, F ]
denote a net of A of distance 2∆, so the intervals used in Equation (22) for each f ∗ ∈ A0 are disjoint.
Then
|A0 | ≤ 2k + k = 3k
because each frequency in x∗ contributes to at most two of the intervals, and the total mass of gb is
at most k times the threshold T N 2 .
Let L1 , . . . , LR be the results of R rounds of Algorithm OneStage. We say that a frequency
f ∈ A0 is successfully recovered in round r if there exists an fe ∈ Lr such that |f − fe| ≤ ∆a , where
√
√
∆a = ∆ T ∆ . T ∆∆.
By Lemma 7.22, each frequency is successfully recovered with 0.8 probability in each round. Then
by the Chernoff bound, with 1 − 2−Ω(k) probability, every f ∈ A0 will be successfully recovered in
at least 0.6R rounds.
Then, by Lemma 7.24, we output a set L of O(B) frequencies such that every f ∈ A0 is within
∆a of some fe ∈ L. Hence every f ∈ A is within 2∆a of some fe ∈ L.
Lemma 7.24. Let L1P
, . . . , LR by sets of frequencies and f ∗ be any frequency. Then L = MergedStages(L1 ,
. . . , LR ) is a set of 2 R|Lr | frequencies satisfying
min |f ∗ − fe| ≤ median min |f ∗ − f |.
fe∈L
r∈[R]
f ∈Lr
Proof. The algorithm is to take the union, sort, and take every R2 th entry of the sorted
S list.
∗
Let ∆ = medianr∈[R] minf ∈Lr |f − f |. We have that at least R/2 different f ∈ r Lr lie within
∆ of f ∗ . This set forms a sequential subsequence of the sorted list of frequencies, so our output will
include one.
39
7.7
Time and sample complexity of frequency recovery of k-clustered signals
The goal of this section is to show that Procedure FrequencyRecoveryKCluster takes
poly(k, log(1/δ)) log(F T ) samples, and runs in poly(k, log(1/δ)) log2 (F T ) time.
In order to analyze the running time and sample complexity. We need to extend the one-cluster
version Procedure GetLegal1Sample and GetEmpirical1Energy (in Algorithm 3) to k-cluster
version GetLegalKSample and GetEmpiricalKEnergy(in Algorithm 7)5 ,
Lemma 7.25. Procedure GetLegalKSample in Algorithm 7 runs Procedure HashToBins Rrepeat =
O((T ∆)3 ) times to output two vectors vb, vb0 ∈ CB such that, for each j ∈ [B],
|b
vj − vbj0 e2πifj β | ≤ 0.08(|b
vj | + |b
vj0 |),
holds with probability at least 0.6.
Using the definition of z in Definition 7.18.
Claim 7.26. Procedure GetEmpiricalKEnergy in Algorithm 7 runs Procedure HashTobins
Rest O((T ∆)2 ) times to output a vector zemp ∈ RB such that, for each j ∈ [B],
j
zemp
∈ [0.8kz (j) kT , 1.2kz (j) kT ],
holds with probability at least 0.9.
Claim 7.27. Algorithm LocateKSignal in Algorithm 6 uses O(poly(k, log(1/δ)) · log(F T )), and
runs in O(poly(k, log(1/δ)) · log2 (F T )).
Proof. We first calculate the number of samples. All the samples is basically all the Fourier samples,
each time needs B log(k/δ). In total it calls HashToBins O(Rest + Rrepeat Dmax Rloc ) times where
Dmax Rloc = Θ(log(F T )) by similar analysis as one-cluster frequency recovery. Thus, the total
number of samples is
(Rest + Rrepeat Dmax Rloc )B log(k/δ) = poly(k, log(1/δ)) · log(F T ).
Then, we analyze the running time.
The expected running time includes the following parts: the first part is running Procedure
HashToBins O(Rest + Rrepeat Dmax Rloc ) times, each run takes O(B log(k/δ) + B log B) samples.
For each such sample we need poly(k, log(1/δ)) time to compute H(t) according to Lemma C.8 and
there are poly(k, log(1/δ)) log(F T )) many samples; the second part is updating the counter v,which
takes O(Dmax Rloc Bt) time. Thus, in total
poly(k, log(1/δ)) · O(Rest + Rrepeat Dmax Rloc ) · O(B log(k/δ) + B log B) + O(Dmax Rloc Bt)
= poly(k, log(1/δ)) · log2 (F T ),
where by similar analysis as one-cluster recovery, t = Θ(log(F T )) and Dmax Rloc = Θ(log(F T )).
To boost the success probability, Procedure MultipleStages reruns Procedure LocateKSignal O(k) times. At the end, Procedure FrequencyRecoveryKCluster combining Procedure
MultipleStages and MergedStages directly, and the running time and sample complexity of
MultipleStages are dominating MergedStages. Thus we have
Lemma 7.28. Procedure FrequencyRecoveryKCluster in Algorithm 8 uses O(poly(k, log(1/δ))·
log(F T )), and runs in O(poly(k, log(1/δ)) · log2 (F T )).
5
We omitted the proofs here, because the proofs are identical to the one-cluster situation.
40
8
One-cluster Signal Recovery
8.1
Overview
In this section, we consider x∗ whose frequencies in x
b∗ are in the range [f0 − ∆0 , f0 + ∆0 ] for some
0
frequency f0 and ∆ > 0 and provide an algorithm to approximate
R T it by a polynomial.
We fix T in this section and recall that hf (t), g(t)iT := T1 0 f (t)g(t)dt such that ke2πifi t kT =
k
p
P
vj e2πifj t , we say the frequency gap of this signal
he2πifi t , e2πifi t iT = 1. For convenience, given
j=1
is min|fi − fj |.
i6=j
For simplicity, we first consider frequencies clustered around 0. The main technical lemma in this
section is that any signal x∗ with bounded frequencies in x
b∗ can be approximated by a low-degree
polynomial on [0, T ].
P
Lemma 2.3. For any ∆ > 0 and any δ > 0, let x∗ (t) = j∈[k] vj e2πifj t where |fj | ≤ ∆ for each
j ∈ [k]. There exists a polynomial P (t) of degree at most
d = O(T ∆ + k 3 log k + k log 1/δ)
such that
kP (t) − x∗ (t)k2T ≤ δkx∗ k2T .
One direct corollary is that when x
b∗ are in the range [f0 + ∆0 , f0 + ∆0 ], we can approximate x∗
2πif
t
by P (t) · e 0 for some low degree polynomial P .
We give an overview of this section first. We first show some technical tools in Section 8.2, 8.3.
In Section 8.4, using those tools, we can show for any k-Fourier-sparse signal, there exists another
k-Fourier-sparse signal with bounded frequency gap close to the original signal. In Section 8.5,
we show that for any k-Fourier-sparse signal with bounded frequency gap, then there exists a low
degree polynomial close to it. In Section 8.6, we show how to transfer low degree polynomial back
to a Fourier-sparse signal. Combining all the above steps finishes the proof of Lemma 2.3.
We apply Theorem 7.5 of frequency estimation on x∗ to obtain an estimation fe0 of f0 and use
e
Theorem 4.5 on the approximation Q(t)e2πif0 t of x∗ to recover the signal. We summarize this result
as follows.
Theorem 8.1 (One-cluster Signal Recovery). Let x∗ (t) =
k
P
vj e2πifj t where ∀j ∈ [k], |fj − f0 | ≤ ∆
j=1
and x(t) = x∗ (t) + g(t) be our observable signal. For any δ > 0 and any T > 0, let N 2 :=
kgk2T + δkx∗ k2T . Procedure CFT1Culster in Algorithm 5 finds a polynomial P (t) of degree at most
d = O (T ∆h + T ∆)1.5 + k 3 log k + k log 1/δ and a frequency fe0 such that
kP (t) · e2πif0 t − x∗ (t)k2T . N 2
e
(24)
The algorithm uses O(kd)+poly(k, log(1/δ)) log(F T ) samples, run in O(kdω )+poly(k, log(1/δ)) log2 (F T )
time, and succeeds with probability at least 1 − 2−Ω(k) .
Proof. We apply the algorithm in Theorem 7.5 to obtain an estimation p
fe0 with poly(k) log(F T )
2
e
samples and poly(k) log (F T ) running time such that |f0 − f0 | . (∆h + ∆) T (∆h + ∆) holds with
probability at least 1 − 2−Ω(k) . Notice that |fj − fe0 | ≤ |fj − f0 | + |fe0 − f0 | . (T (∆h + ∆))1.5 .
41
We consider x0 (t) = e−2πif0 t x(t) =
e
k
P
vj e2πi(fj −f0 )t . By Lemma 2.3, there exists a polynomial
e
j=1
P (t) of degree at most
d = O (T ∆h + T ∆)1.5 + k 3 log k + k log 1/δ
such that it approximates x0 by
δ
δ
kP (t) − x0 (t)kT ≤ kx0 (t)kT = kx∗ (t)kT .
4
4
which indicates kQ(t) − e−2πif0 t · x∗ (t)kT ≤ 4δ kx∗ (t)kT .
e
Because we can sample x(t), we can also sample e−2πif0 t · x(t) = Q(t) + g 0 (t) for g 0 (t) =
e
e
e−2πif0 t · g(t) + (e−2πif0 t · x∗ (t) − Q(t)). Hence we apply the algorithm in Theorem 4.5 and choose
R = O(k) in that proof. Then Procedure RobustPolynomialLearning+ takes O(kd) samples
and O(kdω ) time to find a degree d polynomial P (t) approximating Q(t) such that
e
kP (t) − Q(t)kT . kg 0 (t)kT ,
holds with probability at least 1 − 2−Ω(k) . It indicates
kP (t) − e−2πif0 t · x∗ (t)kT . kP (t) − Q(t)kT + kQ(t) − x∗ (t)k . δkx∗ (t)kT + kg(t)kT h N .
e
Therefore we know ke2πif0 t · P (t) − x∗ (t)k2T . N 2 .
e
8.2
Bounding the Gram matrix determinant
We define Gram matrix for e2πif1 t , e2πif2 t , · · · , e2πifk t and provide lower/upper bounds for its determinant.
Definition 8.2 (Gram matrix). We define Gramf1 ,··· ,fk to be
he2πif1 t , e2πif1 t iT
he2πif2 t , e2πif1 t iT
···
he2πifk t , e2πif1 t iT
he2πif1 t , e2πif2 t iT
he2πif2 t , e2πif2 t iT
···
he2πifk t , e2πif2 t iT
···
···
···
···
he2πif1 t , e2πifk t iT
he2πif2 t , e2πifk t iT
···
2πif
t
2πif
t
k
k
he
,e
iT
Note that the above matrix is a Hermitian matrix with complex entries, thus both its determinant
and all eigenvalues are in R.
We defer the proof of the following Theorem to Appendix A.1.
Theorem 8.3. For real numbers ξ1 , . . . , ξk , let Gξ1 ,...,ξk be the matrix whose (i, j)-entry is
Z
1
e2πi(ξi −ξj )t dt.
−1
Then
det(Gξ1 ,...,ξk ) = 2Õ(k
2)
Y
i<j
We use the following corollary in this section.
42
min(|ξi − ξj |2 , 1).
Corollary 8.4. There exists a universal constant α > 0 such that, for any T > 0 and real numbers
f1 , · · · , fk , the k × k Gram matrix of e2πif1 t , e2πif2 t , · · · , e2πifk t whose (i, j)-entry is
Z
1 T 2πi(fi −fj )t
2πifi t 2πifj t
e
dt.
Gramf1 ,··· ,fk (i, j) = he
,e
iT =
T 0
satisfies
k −αk
2
Y
i<j
min((|fi − fj |T )2 , 1) ≤ det (Gramf1 ,··· ,fk ) ≤ k αk
2
Y
i<j
min((|fi − fj |T )2 , 1).
Based on Corollary 8.4, we show the coefficients of a k-Fourier-sparse signal can be upper
bounded by the energy kxk2T .
Lemma 2.2. There exists a universal constant c > 0 such that for any x(t) =
k
P
vj e2πifj t with
j=1
frequency gap η = min|fi − fj |,
i6=j
k
X
2
kx(t)k2T ≥ k −ck min (ηT )2k , 1
|vj |2 .
j=1
vi , v~i i =
Proof. Let v~i denote the vector e2πifi t and V = {v~1 , · · · , v~k }. Notice that k~
vi k2T = h~
k
1. For each v~i , we define ~vi to be the projection of v~i into the linear subspace span{V \ v~i } =
k
span{~v1 , · · · , ~vi−1 , ~vi+1 , · · · , ~vk } and ~vi⊥ = ~vi − ~vi which is orthogonal to span{V \ v~i } by the
definition.
Therefore from the orthogonality,
k
kx(t)k2T
2
≥ max{|vj | ·
j∈[k]
k~vj⊥ k2T }
1X 2
≥
|vj | · k~vj⊥ k2T .
k
j=1
It is enough to estimate k~vj⊥ k2T from Claim 3.12:
k~vj⊥ k2T =
Y
det(Gram(V ))
2
2
min ((fj − fi )T, 1)2 ≥ k −2αk (ηT )2k−2 ,
≥ k −2αk
det(Gram(V \ ~vi ))
j6=i
where we use Corollary 8.4 to lower bound it in the last step.
8.3
Perturbing the frequencies does not change the subspace much
We show that for a k-Fourier-sparse signal with unboundedly close frequency gap, there always
exists another k-Fourier-sparse signal with slightly separated gap.
Lemma 8.5 (Slightly Shifting one Frequency). There is a universal constant C0 > 0 such that for
k
P
any x(t) =
vj e2πifj t and any frequency fk+1 , there always exists
j=1
x0 (t) =
k−1
X
0
vj0 e2πifj t + vk+1
e2πifk+1 t
j=1
0
0
with k coefficients v10 , v20 , · · · , vk−1
, vk+1
satisfying
2
kx0 (t) − x(t)kT ≤ k C0 k · (|fk − fk+1 |T ) · kx(t)kT
43
Proof. We abuse the notation e2πifj t to denote a vector in the linear subspace. We plan to shift fk
to fk+1 and define
V
= {e2πif1 t , · · · , e2πifk−1 t , e2πifk t }
V 0 = {e2πif1 t , · · · , e2πifk−1 t , e2πifk+1 t }
U
W
= {e2πif1 t , · · · , e2πifk−1 t }
= {e2πif1 t , · · · , e2πifk−1 t , e2πifk t , e2πifk+1 t }
where f1 , f2 , · · · , fk are original frequencies in x. The idea is to show that any vector in the linear
subspace span{V } is close to some vector in the linear subspace span{V 0 }.
For convenience, we use ~uk to denote the projection of vector e2πifk t to the linear subspace
span{U } = span{e2πif1 t , · · · , e2πifk−1 t } and w
~ k denote the projection of vector e2πifk+1 t to this
linear subspace span{U }. Let ~u⊥ = e2πifk t − ~uk and w
~ ⊥ = e2πifk+1 t − w
~ k be their orthogonal part
to span{U }.
From the definition e2πifk t = ~uk + ~u⊥ and ~uk ∈ span{U } = span{e2πif1 t , · · · , e2πifk−1 t }, we
rewrite the linear combination
x(t) =
k
X
vj e2πifj t =
j=1
k−1
X
j=1
αj e2πifj t + vk · ~u⊥
for some scalars α1 , · · · , αk−1 .
We will substitute ~u⊥ by w
~ ⊥ in the above linear combination and find a set of new coefficients.
⊥
⊥
~ ⊥ to ~u⊥ . Therefore w
~ 2 is the
Let w
~⊥ = w
~1 + w
~ 2 where w
~ 1 = h~uk~u⊥,w~k2 i ~u⊥ is the projection of w
T
orthogonal part of the vector e2πifk+1 t to span{V } = span{e2πif1 t , · · · , e2πifk−1 t , e2πifk t }. We use
δ = kkw~w~⊥2 kkT for convenience.
T
Notice that the min k~u
β∈C
x0 (t) =
k−1
X
j=1
⊥ −β·w
~ ⊥ kT
k~
u⊥ kT
= δ and β ∗ =
h~
u⊥ , w
~ ⊥i
kw
~ ⊥ k2T
is the optimal choice. Therefore we set
βj e2πifj t + vk · β ∗ · w
~ ⊥ ∈ span{e2πif1 t , · · · , e2πifk−1 t , e2πifk+1 t }
where the coefficients β1 , · · · , βk−1 guarantee that the projection of x0 onto span{U } is as same as
the projection of x onto span{U }. From the choice of β ∗ and the definition of x0 ,
kx(t) − x0 (t)k2T = δ 2 · |vk |2 · k~u⊥ k2T ≤ δ 2 · kx(t)k2T .
44
Eventually, we show an upper bound for δ 2 from Claim 3.12.
δ2 =
=
=
kw
~ 2 k2T
kw
~ ⊥ k2T
det(GramW ) det(GramV 0 )
/
by Claim 3.12
det(GramV ) det(GramU )
det(GramW ) det(GramU )
·
by Corollary 8.4
det(GramV ) det(GramV 0 )
k+1
Q k+1
Q
min(|fi − fj |T, 1)
2
≤ k 4αk ·
i=1 j=1
j6=i
k Q
k
Q
i=1j=1
j6=i
min(|fi − fj |T, 1)
·
k−1
Q k−1
Q
i=1 j=1
j6=i
k−1
Q k−1
Q
i=1 j=1
j6=i
min(|fi − fj |T, 1)
min(|fi − fj |T, 1) ·
k−1
Q
i=1
min(|fi − fk+1 |2 T 2 , 1)
2
= k 4αk |fk − fk+1 |2 T 2
Lemma 8.6. For any k frequencies f1 < f2 < · · · < fk , there exists k frequencies f10 , · · · , fk0 such
0
that min fi+1
− fi0 ≥ η and for all i ∈ [k], |fi0 − fi | ≤ kη.
i∈[k−1]
0
Proof. We define the new frequencies fi0 as follows: f10 = f1 and fi0 = max{fi−1
+ η, fi } for i ∈
{2, 3, · · · , k}.
8.4
Existence of nearby k-Fourier-sparse signal with frequency gap bounded
away from zero
We combine the results in the above section to finish the proof of Lemma 2.3. We first prove
k
P
that for any x∗ (t) =
vj e2πifj t , there always exists another k-Fourier-sparse signal x0 close to
j=1
x∗ (t) =
k
P
vj e2πifj t such that the frequency gap in x0 is at least η ≥ 2− poly(k) . Then we show how
j=1
to find a low degree polynomial P (t) approximating x0 (t).
Lemma 2.1. There is a universal constant C1 > 0 such that, for any x∗ (t) =
k
P
vj e2πifj t and any
j=1
δ > 0 , there always exist η ≥
δ
T
· k −C1
k2
and x0 (t) =
k
P
vj0 e
2πifj0 t
satisfying
j=1
kx0 (t) − x∗ (t)kT ≤ δkx∗ (t)kT
with min|fi0 − fj0 | ≥ η and max{|fj0 − fj |} ≤ kη.
i6=j
j∈[k]
Proof. Using Lemma 8.6 on frequencies f1 , · · · , fk , we obtain k new frequencies f10 , · · · , fk0 such that
their gap is at least η and maxi |fi − fi0 | ≤ kη. Next we use the hybrid argument to find x0 .
45
Let x(0) (t) = x∗ (t). For i = 1, · · · , t, we apply Lemma 8.5 to shift fi to fi0 and obtain
k
X
x(i) (t) =
(i)
vj e2πifj t +
j=i+1
i
X
0
(i)
vj e2πifj t .
j=1
2
From Lemma 8.5, we know kx(i) (t) − x(i−1) (t)kT ≤ k C0 k (|fi − fi0 |T )kx(i−1) kT . Thus we obtain
i
i
2
2
1 − k C0 k (kηT ) kx(0) (t)kT ≤ kx(i) (t)kT ≤ 1 + k C0 k (kηT ) kx(0) (t)kT ,
which is between
k2
h
i
2
2
1 − i · k C0 k (kηT ) kx(0) (t)kT , 1 + 2i · k C0 k (kηT ) kx(0) (t)kT for η ≤
k −C1 with some C1 > C0 .
At last, we set x0 (t) = x(k) (t) and bound the distance between x0 (t) and x∗ (t) by
kx
(k)
(0)
(t) − x
(t)kT ≤
≤
≤
k
X
i=1
k
X
i=1
k
X
kx(i) (t) − x(i−1) (t)kT
1
5T
·
by triangle inequality
2
k C0 k (|fi − fi0 |T )kx(i−1) (t)kT
by Lemma 8.5
2
by max|fi − fi0 | ≤ kη
2k C0 k (kηT )kx(i−1) (t)kT
i
i=1
k2
≤ k · 2k C0 (kηT )kx∗ (t)kT
≤ δkx∗ (t)kT
where the last inequality follows by the sufficiently small η.
8.5
Approximating k-Fourier-sparse signals by polynomials
For any k-Fourier-sparse signal with frequency gap bounded away from zero, we show that there
exists a low degree polynomial which is close to the original k-Fourier-sparse signal in k·kT distance.
Lemma 8.7 (Existence of low degree polynomial). Let x∗ (t) =
k
P
vj e2πifj t , where ∀j ∈ [k], |fj | ≤ ∆
j=1
and min|fi − fj | ≥ η. There exists a polynomial Q(t) of degree
i6=j
such that,
d = O T ∆ + k log 1/(ηT ) + k 2 log k + k log(1/δ)
kQ(t) − x∗ (t)k2T ≤ δkx∗ (t)k2T
Proof. For each frequency fj , let Qj (t) =
d−1
P
k=0
(2πifj t)k
k!
be the first d terms in the Taylor Expansion
of e2πifj t . For any t ∈ [0, T ], we know the difference between Qj (t) and e2πifj t is at most
|Qj (t) − e2πifj t | ≤ |
(2πifj T )d
2πT ∆ · e d
|≤(
) .
d!
d
46
(25)
We define
Q(t) =
k
X
vj Qj (t)
j=1
and bound the distance between Q and x∗ from the above estimation:
kQ(t) − x
∗
(t)k2T
Z
1
=
T
T
0
1
=
T
Z
≤ 2k
k
X
≤k
|Q(t) − x∗ (t)|2 dt
T
|
0
j=1
k
X
j=1
k
X
j=1
1
T
Z
vj (Qj (t) − e2πifj t )|2 dt
T
0
|vj |2 · (
|vj |2 · |Qj (t) − e2πifj t |2 dt
2πT ∆ · e 2d
)
d
by triangle inequality
by Taylor expansion
On the other hand, from Lemma 2.2, we know
kx∗ (t)k2T ≥ (ηT )2k · k −ck
2
X
j
|vj |2 .
Because d = 10·πe(T ∆+k log 1/(ηT )+k 2 log k +k log(1/δ)) is large enough, we have k( 2πTd∆·e )2d ≤
2
δ(ηT )2k · k −ck , which indicates that kQ(t) − x∗ (t)k2T ≤ δkx∗ k2T from all discussion above.
8.6
Transferring degree-d polynomial to (d+1)-Fourier-sparse signal
In this section, we show how to transfer a degree-d polynomial to (d+1)-Fourier-sparse signal.
Lemma 8.8. For any degree-d polynomial Q(t) =
d
P
cj tj , any T > 0 and any > 0, there always
j=0
exist γ > 0 and
x∗ (t) =
d+1
X
αi e2πi(γi)t
i=1
with some coefficients α0 , · · · , αd such that
∀t ∈ [0, T ], |x∗ (t) − Q(t)| ≤ .
47
Proof. We can rewrite x∗ (t),
x∗ (t) =
=
d+1
X
αi e2πiγit
i=1
d+1
X
αi
i=1
=
∞
X
(2πiγit)j
j!
j=0
∞
d+1
X
(2πiγt)j X
j!
j=0
i=1
αi · ij
d+1
∞
X
(2πiγt)j X
=
αi · i +
αi · ij
j!
j!
i=1
i=1
j=0
j=d+1
d
∞
d+1
d+1
j X
j X
X
X
(2πiγt)
(2πiγt)
= Q(t) +
αi · ij − Q(t) +
αi · ij .
j!
j!
j=0
i=1
i=1
j=d+1
{z
} |
{z
}
|
d+1
(2πiγt)j X
d
X
j
C1
C2
Our goal is to show there exists some parameter γ and coefficients {α0 , α1 , · · · , αd } such that the
term C1 = 0 and |C2 | ≤ . Let’s consider C1 ,
!
d
d+1
X
t j (2πiγT )j X j
C1 =
( )
αi i − cj
T
j!
i=1
j=0
To guarantee C1 = 0, we need to solve a linear system with d + 1 unknown variables and d + 1
constraints,
Find α1 , α2 , · · · αd+1
s.t.
d+1
(2πiγT )j X j
αi i − cj = 0, ∀j ∈ {0, 1, · · · , d}
j!
i=1
Define c0j = cj j!/(2πiγ)j , let α and c0 be the length-(d + 1) column vectors with αi and c0j . Let
A ∈ Rd+1×d+1 denote the Vandermonde matrix where Ai,j = ij , ∀i, j ∈ [d + 1] × {0, 1, · · · , d}. Then
Q
2
we need to guarantee Aα = c0 . Using the definition of determinant, det(A) =
|i − j| ≤ 2O(d log d) .
i<j
2
Thus σmax (A) ≤ 2O(d
log d)
and then
det(A)
3
σmin (A) = Qd−1 ≥ 2−O(d log d) .
i=1 σi
We show how to upper bound |αi |,
max |αi | ≤ kαk2 = kA† c0 k2 ≤ kA† k2 · kc0 k2 ≤
i∈[d+1]
48
√
|cj |j!
1
d + 1 max
0≤j≤d (2πγT )j
σmin (A)
Plugging the above equation into C2 , we have
|C2 | =
≤
≤
≤
d+1
∞
X
(2πiγt)j X
αi · ij
j!
i=1
j=d+1
∞
X
j=d+1
∞
X
j=d+1
∞
X
j=d+1
d+1
(2πγt)j X
j!
(2πγt)j
j!
i=1
|αi | · ij
(d + 1)d+1 max |αi |
i∈[d+1]
d!
1
(2πγt)j
max |cj |
(d + 1)d+2
j!
σmin (A) (2πγT )d 0≤j≤d
≤
where the last step follows by choosing sufficiently small
Θ(d3 log d)
γ . / T 2
max |cj | .
0≤j≤d
k-cluster Signal Recovery
9
9.1
Overview
In this section, we prove Lemma 9.1 as the main
P technical lemma to finish the proof of main
Theorem 1.1, which shows how to learn x∗ (t) = kj=1 vj e2πifj t with noise.
P
Lemma 9.1. Let x∗ (t) = kj=1 vj e2πifj t and x(t) = x∗ (t) + g(t) be our observation. For any δ > 0
RT
RT
and T > 0, let N 2 := T1 0 |g(t)|2 dt + δ · T1 0 |x∗ (t)|2 dt. For ∆ = poly(k, log(1/δ))/T , Procedure
SignalRecoveryKCluster+ in Algorithm 8 takes l = O(k) frequencies fe1 , · · · , fel as input and
finds l polynomials Q1 , · · · , Ql of degree d = O((T ∆)1.5 + k 3 log k + k log 1/δ) such that
x
e(t) =
X
j∈[l]
Qj (t)e2πifj t satisfies ke
x(t) − x∗ (t)k2T . N 2 .
e
(26)
The procedure succeeds with probability at least 1 − 2−Ω(k) , uses poly(k, log(1/δ)) · log(F T ) samples,
and runs in poly(k, log(1/δ)) · log2 (F T ) time.
For any set W = {t1 , · · · , tm } where each ti ∈ [0, T ], we use
sP
v (ti )|2
i∈W |~
k~v kW =
for any ~v : [0, T ] → C
|W |
in this section. We first show that Procedure SignalRecoveryKCluster succeeds with constant
probability, then prove that Procedure SignalRecoveryKCluster+ succeeds with probability
at least 1 − 2−Ω(k) .
49
9.2
Heavy clusters separation
Recall the definition of “heavy” clusters.
k
P
Definition 2.4. Given x∗ (t) =
b with bounded
vj e2πifj t , any N > 0, and a filter function (H, H)
j=1
\
j t · H) for each j ∈ [k].
support in frequency domain. Let Lj denote the interval of supp(e2πif
Define an equivalence relation ∼ on the frequencies fi by the transitive closure of the relation
fi ∼ fj if Li ∩ Lj 6= ∅. Let S1 , . . . , Sn be the equivalence classes under this relation.
R
\
Define Ci = ∪ Li for each i ∈ [n]. We say Ci is a “heavy” cluster iff Ci |H
· x∗ (f )|2 df ≥
f ∈Si
T · N 2 /k.
By reordering Ci , we can assume {C1 , C2 , · · · , Cl } are heavy clusters, where l ≤ n ≤ k.
Claim 2.5. Given x∗ (t) =
k
P
vj e2πifj t and any N > 0, let H be the filter function defined in
j=1
Appendix C.1 and C1 , · · · , Cl be the heavy clusters from Definition 2.4. For
S = j ∈ [k] fj ∈ C1 ∪ · · · Cl ,
we have x(S) (t) =
P
j∈S
Proof. Let x(S) (t) =
vj e2πifj t approximating x∗ within distance kx(S) (t) − x∗ (t)k2T . N 2 .
P
j∈[k]\S
vj e2πifj t . Notice that kx∗ − x(S) k2T = kx(S) k2T .
b in Appendix C.1, we have
From the property VI of filter function (H, H)
Z
+∞
−∞
|x
(S)
2
(t) · H(t)| dt ≥ 0.9
Z
0
T
|x(S) (t)|2 dt = 0.9 · T kxS k2T .
From Definition 2.4, we have
Z +∞
Z +∞
2
(S)
(S) · H(f )|2 df
|x\
|x (t) · H(t)| dt =
−∞
−∞
Z
∗ · H(f )|2 df
=
|x\
[−∞,+∞]\C1 ∪···∪Cl
≤ k · T N 2 /k.
Overall, we have kx(S) k2T . N 2 .
√
From the guarantee of Theorem 2.6, for any j ∈ S, min|fj − fei | ≤ ∆ ∆T . From now on, we
i∈[l]
focus on the recovery of
which is enough to approximate x∗ from the above claim. Because
we are looking for x̃ approximating x(S) within distance O(N 2 ), from Lemma 2.1, we can assume
2
δ
there is a frequency gap η ≥ 10T
k −O(k ) among x(S) .
x(S) ,
50
9.3
Approximating clusters by polynomials
P
e
In this section, we show how to approximate x(S) by x0 (t) = i∈[l] e2πifi t Pi (t) where P1 , · · · , Pl are
low degree polynomials.
P
Claim 9.2. For any x(S) (t) = j∈S vj e2πifj t with a frequency gap η = min|fi −fj | and l frequencies
i6=j
√
fe1 , · · · , fel with the property ∀j ∈ S, mini∈[l] |fj − fei | ≤ ∆ ∆T , let
n
o
e
d = 5π (T ∆)1.5 + k 3 log k + log 1/δ and V = tj e2πifi t |i ∈ [l], j ∈ {0, · · · , d} .
There exists x0 (t) ∈ span{V } that approximates x(S) (t) as follows:
∀t ∈ [0, T ], |x0 (t) − x(S) (t)| ≤ δkx(S) kT .
Proof. From Lemma 2.2, we know
kx(S) k2T ≥ (ηT )2k · k −ck
2
X
j∈S
|vj |2 .
√
For each frequency fj , we use pj to denote the index in [l] such that |fj − fepj | ≤ ∆ ∆T . We rewrite
l
X
X
e
e
x(S) (t) =
e2πifi
vj e2πi(fj −fi )t .
i=1
j∈S:pj =i
For d = 5π((T ∆)1.5 + k 3 log k + log 1/δ) and each e2πi(fj −fpj )t , let Qj (t) =
e
2πi(fj −fepj )t
the first d terms in the Taylor Expansion of e
between Qj (t) and e
2πi(fj −fepj )t
is at most
∀t ∈ [0, T ], |Qj (t) − e2πi(fj −fpj )t | ≤ |
e
Let x0 =
Pl
2πifi t
i=1 e
e
Pd−1 (2πi(fj −fepj )t)i
i=0
i!
be
. For any t ∈ [0, T ], we know the difference
(2πi(fj − fepj )T )d
8π(∆T )1.5 d
|≤(
) .
d!
d
v
Q
(t)
. From all discussion above, we know for any t ∈ [0, T ],
j∈S:pj =i j j
P
2
1.5
X
8π(T ∆)
|x0 (t) − x(S) (t)|2 ≤
|vj |(
)d
d
j∈S
≤ k(
8π(T ∆)1.5 2d X
)
|vj |2
d
j
8π(T ∆)1.5
≤
k(
)2d (S) 2
d
kx kT
(ηT )2k · k −ck2
≤ δ 2 kx(S) k2T .
We provide a property of functions in span{V } such that we can use the Chernoff bound and
the -net argument on vectors in span{V }.
51
e
Claim 2.7. For any ~u ∈ span e2πifi t · tj j ∈ {0, · · · , d}, i ∈ [l] , there exists some universal constants C1 ≤ 4 and C2 ≤ 3 such that
max {|~u(t)|2 } . (ld)C1 logC2 (ld) · k~uk2T
t∈[0,T ]
Proof. From Lemma 8.8, we can approximate each polynomial
in ~u by a linear combination of
n
o
˜
{1, e2πi·γt , · · · , e2πi·(γd)t } such that we obtain u∗ ∈ span e2πi·(γj)t · e2πifi t |i ∈ [l], j ∈ {0, · · · , d + 1}
for some small γ such that ∀t ∈ [0, T ], |~u(t) − u∗ (t)| ≤ 0.01k~ukT .
From Lemma 5.1, we know
max |u∗ (t)|2 ≤ C · (ld + 1)4 · log3 (ld + 1) ku∗ k2T .
t∈[0,T ]
For some constant C 0 , we have
max |~u(t)|2 ≤ C 0 (kd)C1 logC2 d k~uk2T .
t∈[0,T ]
9.4
Main result, with constant success probability
In this section, we show that the output x
e is close to x0 with high probability using the -net
argument, which is enough to prove ke
x − xkT . N 2 from all discussion above. Because we can
prove Lemma 9.6(which is the main goal of this section), then combining kx0 −x∗ kT ≤ kx0 −x(S) kT +
kx(S) − x∗ kT . δkx∗ kT and Lemma 9.6, we have kx∗ − x
ekT . kgkT + δkx∗ kT , which finishes the
proof of Procedure SignalRecoveryKCluster in Algorithm 8 achieving the Equation (26) with
constant success probability but not 1 − 2−Ω(k) . We will boost the success probability in Section
9.5.
We first provide an -netnP for the unit vectors Q = {~u ∈ span{V
} k~uk2T = 1} in the linear subo
space span{V } where V = tj · e2πifi t j ∈ {0, 1, · · · , d}, i ∈ [l]
that the dimension of span{V } is at most l(d + 1).
e
from the above discussion. Notice
Claim 9.3. There exists an -net P ⊂ span{V } such that
1. ∀~u ∈ Q, ∃w
~ ∈ P, k~u − wk
~ T ≤ .
2l(d+1)
2. |P| ≤ 5 l(d+1)
.
Proof. Let P 0 be an l(d+1)
-net in the unit circle of C with size at most (4 l(d+1)
+ 1)2 , i.e.,
2l(d + 1)
2l(d + 1)
0
P =
j1 + i
j2 j1 , j2 ∈ Z, |j1 | ≤
, |j2 | ≤
.
2l(d + 1)
2l(d + 1)
Observe that the dimension of span{V } is at most l(d + 1). Then we take an orthogonal basis
w
~ 1, · · · , w
~ l(d+1) in span{V } and set
l(d+1)
P ={
X
i=1
αi w
~ i ∀i ∈ [l(d + 1)], αi ∈ P 0 }.
2l(d+1)
Therefore P is an -net for Q and |P| ≤ 5 l(d+1)
.
52
We first prove that W is a good estimation for all functions in the -net P.
Claim 9.4. For any > 0, there exists a universal constant C3 ≤ 5 such that for a set S of i.i.d.
C3
C3
d/
,then with probability at
samples chosen uniformly at random over [0, T ] of size |S| ≥ 3(kd) log
2
least 1 − k −k , for all w
~ ∈ P, we have
kwk
~ W ∈ [(1 − )kwk
~ T , (1 + )kwk
~ T].
Proof. From Claim 2.7 and Lemma 3.5, for each w
~ ∈ P,
|W |2
−
1.5 d
3(kd)C1 logC2 +0.5 d ≤ 2−kd log
.
Pr kw(t)k
~
∈
/
(1
−
)k
wk
~
,
(1
+
)k
wk
~
≤
2
W
T
T
From the union bound, kwk
~ W ∈ [(1 − )kwk
~ T , (1 + )kwk
~ T ] for any w
~ ∈ P with probability at
d −kd log0.5 d
−d
least 1 − ( )
· |P| ≥ 1 − d .
Then We prove that W is a good estimation for all functions in span{V } using the property of
-nets.
Claim 9.5. For any > 0, there exists a universal constant C3 ≤ 5 such that for a set W of i.i.d.
C3
C3
d/
samples chosen uniformly at random over [0, T ] of size |W | ≥ 3(kd) log
,then with probability
2
−d
at least 1 − d , for all u ∈ span{V }, we have
k~ukW ∈ [(1 − 3)k~ukT , (1 + 3)k~ukT ]
Proof. We assume that the above claim is true for any w
~ ∈ P. Without loss of generality, we
consider ~u ∈ Q such that k~ukT = 1.
Let w
~ 0 be the vector in P that minimizes kw
~ − ~ukT for all w
~ ∈ P, i.e., w
~ 0 = arg minkw
~ − ~ukT .
w∈P
~
Define ~u1 = ~u − w
~ 0 and notice that k~u1 kT ≤ because P is a -net. If k~u1 kT = 0, then we skip the
rest of this procedure. Otherwise, we define α1 = k~u1 kT and normalize u
e1 = ~u1 /α1 .
Then we choose w
~ 1 to be the vector in P that minimizes kw
~ −u
e1 kT for all w
~ ∈ P. Similarly,
we set ~u2 = u
e1 − w
~ 1 and α2 = k~u2 kT . Next we repeat this process for u
e2 = ~u2 /α2 and so on. The
recursive definition can be summarized in the following sense,
initial :
For i ∈ {0, 1, 2, · · · , m} :
u
e0 = ~u and m = 10 log1/ (ld) + 1,
w
~ i = arg minkw
~ −u
ei kT ,
w∈P
~
~ui+1 = u
ei − w
~ i and αi+1 = k~ui+1 kT ,
if αi+1 = 0, stop.
if αi+1 6= 0, u
ei+1 = ~ui+1 /αi+1 and continue,
Q
Eventually, we have ~u = w
~ 0 + α1 w
~ 1 + α1 α2 w
~2 + · · · + m
~ m + ~um+1 ) where each |αi | ≤
j=1 αj (w
and each w
~ i is in the -net P. Notice that k~um+1 kT ≤ 1 and k~um+1 kW ≤ (ld + 1)3 · k~um+1 kT from
Claim 2.7. We prove a lower bound for k~ukW ,
k~ukW
= kw
~ 0 + α1 w
~ 1 + α1 α2 w
~2 + · · · +
m
Y
αj (w
~ m + ~um+1 )kW
j=1
≥ kw
~ 0 kW − kα1 w
~ 1 kW − kα1 α2 w
~ 2 kW − · · · − k
m
Y
j=1
αj w
~ m kW − k
m
Y
j=1
≥ (1 − ) − (1 + ) − 2 (1 + ) − · · · − m (1 + ) − m k~um+1 kW
(1 + )
− m · (ld + 1)3 ≥ 1 − 3.
≥ 1−−
1−
53
αj ~um+1 kW
Similarly, we have k~ukW ≤ 1 + 3.
Lemma 9.6. With probability at least 0.99 over the m i.i.d samples in W ,
0
(S)
0
kx (t) − x
e(t)kT ≤ 2200 kg(t)kT + kx (t) − x (t)kT .
Proof. Let g 0 (t) = g(t) + x∗ (t) − x0 (t) such that x(t) = x0 (t) + g 0 (t). Then we choose = 0.03 and
bound:
kx0 (t) − x
e(t)kT
≤ (1 + 3)kx0 (t) − x
e(t)kW
with prob. 1 − 2−Ω(d log d) by Claim 9.5
= 1.09kx0 (t) − x
e(t)kW
by = 0.03
0
= 1.09kx(t) − g (t) − x
e(t)kW
0
by x (t) = x(t) − g 0 (t)
≤ 1.09kx(t) − x
e(t)kW + 1.09kg 0 (t)kW
0
by triangle inequality
0
≤ 1.09kx(t) − x (t)kW + 1.09kg (t)kW
by x
e = arg min kx − ykW
y∈span{V }
0
by x(t) − x0 (t) = g(t)
= 2.18kg (t)kW .
From the fact that EW [kg 0 kW ] = kg 0 kT , kg 0 kW ≤ 1000kg 0 kT with probability at least .999. It
indicates kx0 (t) − x
e(t)kT ≤ 2200kg 0 kT with probability at least 0.99 from all discussion above.
9.5
Boosting the success probability
In order to achieve 1 − 2−Ω(k) for the main theorem, we cannot combine Procedure SignalRecoveryKCluster with FrequencyRecoveryKCluster directly. However, using the similar
proof technique in Theorem 4.5, we are able to boost the success probability by using Procedure
SignalRecoveryKCluster+ in Algorithm 8. It runs Procedure SignalRecoveryKCluster
R = O(k) times in parallel for independent fresh samples and report R different d-Fourier-sparse
e as before
signals x
ei (t). Then, taking m = poly(k) new locations {t1 , t2 , · · · , tm }, and computing A
e
and bj by taking the median of {e
x1 (tj ), · · · , x
eR (tj )}. At the end, solving the linear regression for
e
e
matrix A and vector b. Thus, we complete the proof of Lemma 9.1.
Because we can transfer a degree-d polynomial to a d-Fourier-sparse signal by Lemma 8.8, the
output of Procedure CFTKCluster in Algorithm 8 matches the main theorem,
Theorem 1.1. Let x(t) = x∗ (t) + g(t), where x∗ is k-Fourier-sparse signal with frequencies in
[−F, F ]. Given samples of x over [0, T ] we can output x
e(t) such that with probability at least
−Ω(k)
1−2
,
ke
x − x∗ kT . kgkT + δ kx∗ kT .
Our algorithm uses poly(k, log(1/δ)) · log(F T ) samples and poly(k, log(1/δ)) · log2 (F T ) time. The
output x
e is poly(k, log(1/δ))-Fourier-sparse signal.
54
References
[BCG+ 14] Petros Boufounos, Volkan Cevher, Anna C Gilbert, Yi Li, and Martin J Strauss. What’s
the frequency, Kenneth?: Sublinear Fourier sampling off the grid. In Algorithmica(A
preliminary version of this paper appeared in the Proceedings of RANDOM/APPROX
2012, LNCS 7408, pp. 61-72), pages 1–28. Springer, 2014.
[BM86]
Y. Bresler and A. Macovski. Exact maximum likelihood parameter estimation of superimposed exponential signals in noise. IEEE Transactions on Acoustics, Speech, and
Signal Processing, 34(5):1081–1089, Oct 1986.
[Bou14]
Jean Bourgain. An improved estimate in the restricted isometry problem. In Geometric
Aspects of Functional Analysis, pages 65–70. Springer, 2014.
[BS12]
Markus Blaser and Chandan Saha. Lecture 6: Multipoint evaluation of a polynomial.
Max-Planck-Institut für Informatik Class Notes, Computational Number Theory and
Algebra, pages 1–4, 2012.
[CDL13]
Albert Cohen, Mark A Davenport, and Dany Leviatan. On the stability and accuracy
of least squares approximations. Foundations of computational mathematics, 13(5):819–
834, 2013.
[CF14]
Emmanuel J Candès and Carlos Fernandez-Granda. Towards a mathematical theory of
super-resolution. Communications on Pure and Applied Mathematics, 67(6):906–956,
2014.
[Che52]
Herman Chernoff. A measure of asymptotic efficiency for tests of a hypothesis based on
the sum of observations. The Annals of Mathematical Statistics, 23:493–507, 1952.
[CNW15]
Michael B Cohen, Jelani Nelson, and David P Woodruff. Optimal approximate matrix
product in terms of stable rank. arXiv preprint arXiv:1507.02268, 2015.
[CRT06]
Emmanuel J Candes, Justin K Romberg, and Terence Tao. Stable signal recovery
from incomplete and inaccurate measurements. Communications on pure and applied
mathematics, 59(8):1207–1223, 2006.
[CW87]
Don Coppersmith and Shmuel Winograd. Matrix multiplication via arithmetic progressions. In Proceedings of the nineteenth annual ACM symposium on Theory of computing,
pages 1–6. ACM, 1987.
[CW13]
Kenneth L. Clarkson and David P. Woodruff. Low rank approximation and regression
in input sparsity time. In Symposium on Theory of Computing Conference, STOC’13,
Palo Alto, CA, USA, June 1-4, 2013, pages 81–90, 2013.
[DB13]
Marco F Duarte and Richard G Baraniuk. Spectral compressive sensing. Applied and
Computational Harmonic Analysis, 35(1):111–129, 2013.
[Dun10]
Mark Dunster. Legendre and Related Functions. Handbook of Mathematical Functions,
Cambridge University Press, 2010.
[FL12]
Albert Fannjiang and Wenjing Liao. Coherence pattern-guided compressive sensing with
unresolved grids. SIAM Journal on Imaging Sciences, 5(1):179–202, 2012.
55
[GGI+ 02]
Anna C Gilbert, Sudipto Guha, Piotr Indyk, S Muthukrishnan, and Martin Strauss.
Near-optimal sparse Fourier representations via sampling. In Proceedings of the thiryfourth annual ACM symposium on Theory of computing, pages 152–161. ACM, 2002.
[GMS05]
Anna C Gilbert, S Muthukrishnan, and Martin Strauss. Improved time bounds for
near-optimal sparse Fourier representations. In Optics & Photonics 2005, pages 59141A–
59141A. International Society for Optics and Photonics, 2005.
[GZ16]
Venkatesan Guruswami and David Zuckerman. Robust Fourier and polynomial curve
fitting. In Foundations of Computer Science (FOCS), 2016 IEEE 57th Annual Symposium on. IEEE, 2016.
[Haz01]
Michiel Hazewinkel. Gram matrix. Encyclopedia of Mathematics, Springer, 2001.
[HIKP12a] Haitham Hassanieh, Piotr Indyk, Dina Katabi, and Eric Price. Nearly optimal sparse
Fourier transform. In Proceedings of the forty-fourth annual ACM symposium on Theory
of computing. ACM, 2012.
[HIKP12b] Haitham Hassanieh, Piotr Indyk, Dina Katabi, and Eric Price. Simple and practical
algorithm for sparse Fourier transform. In Proceedings of the twenty-third annual ACMSIAM symposium on Discrete Algorithms, pages 1183–1194. SIAM, 2012.
[HK15]
Qingqing Huang and Sham M Kakade. Super-resolution off the grid. In Advances in
Neural Information Processing Systems, pages 2647–2655, 2015.
[HR15]
Ishay Haviv and Oded Regev. The restricted isometry property of subsampled Fourier
matrices. arXiv preprint arXiv:1507.01768, 2015.
[IK14]
Piotr Indyk and Michael Kapralov. Sample-optimal Fourier sampling in any constant
dimension. In Foundations of Computer Science (FOCS), 2014 IEEE 55th Annual
Symposium on, pages 514–523. IEEE, 2014.
[IKP14]
Piotr Indyk, Michael Kapralov, and Eric Price. (Nearly) Sample-optimal sparse Fourier
transform. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 480–499. SIAM, 2014.
[Iwe13]
Mark A Iwen. Improved approximation guarantees for sublinear-time Fourier algorithms. Applied And Computational Harmonic Analysis, 34(1):57–82, 2013.
[Kap16]
Michael Kapralov. Sparse Fourier transform in any constant dimension with nearlyoptimal sample complexity in sublinear time. In Symposium on Theory of Computing
Conference, STOC’16, Cambridge, MA, USA, June 19-21, 2016, 2016.
[Mas69]
James L Massey. Shift-register synthesis and BCH decoding. Information Theory, IEEE
Transactions on, 15(1):122–127, 1969.
[Moi15]
Ankur Moitra. The threshold for super-resolution via extremal functions. In STOC,
2015.
[MP14]
Gregory T Minton and Eric Price. Improved concentration bounds for count-sketch.
In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 669–686. Society for Industrial and Applied Mathematics, 2014.
56
[NN13]
Jelani Nelson and Huy L Nguyên. OSNAP: Faster numerical linear algebra algorithms
via sparser subspace embeddings. In Foundations of Computer Science (FOCS), 2013
IEEE 54th Annual Symposium on, pages 117–126. IEEE, 2013.
[PS15]
Eric Price and Zhao Song. A robust sparse Fourier transform in the continuous setting.
In Foundations of Computer Science (FOCS), 2015 IEEE 56th Annual Symposium on,
pages 583–600. IEEE, 2015.
[RPK86]
Robert Roy, Arogyaswami Paulraj, and Thomas Kailath. Esprit–a subspace rotation
approach to estimation of parameters of cisoids in noise. Acoustics, Speech and Signal
Processing, IEEE Transactions on, 34(5):1340–1342, 1986.
[RV08]
Mark Rudelson and Roman Vershynin. On sparse reconstruction from fourier and gaussian measurements. Communications on Pure and Applied Mathematics, 61(8):1025–
1045, 2008.
[Sch81]
Ralph Otto Schmidt. A signal subspace approach to multiple emitter location spectral
estimation. Ph. D. Thesis, Stanford University, 1981.
[Str69]
Volker Strassen. Gaussian elimination is not optimal.
13(4):354–356, 1969.
[Tar09]
Robert E. Tarjan. Lecture 10: More chernoff bounds, sampling, and the chernoff +
union bound. Princeton Class Notes, Probability and Computing, pages 1–9, 2009.
[TBR15]
Gongguo Tang, Badri Narayan Bhaskar, and Benjamin Recht. Near minimax line spectral estimation. Information Theory, IEEE Transactions on, 61(1):499–512, 2015.
Numerische Mathematik,
[TBSR13] Gongguo Tang, Badri Narayan Bhaskar, Parikshit Shah, and Benjamin Recht. Compressed sensing off the grid. Information Theory, IEEE Transactions on, 59(11):7465–
7490, 2013.
[Wil12]
Virginia Vassilevska Williams. Multiplying matrices faster than coppersmith-winograd.
In STOC, pages 887–898. ACM, 2012.
[Woo14]
David P Woodruff. Sketching as a tool for numerical linear algebra. arXiv preprint
arXiv:1411.4357, 2014.
[YX15]
Zai Yang and Lihua Xie. Achieving high resolution for super-resolution via reweighted
atomic norm minimization. In Acoustics, Speech and Signal Processing (ICASSP), 2015
IEEE International Conference on, pages 3646–3650. IEEE, 2015.
57
A
Technical Proofs
A.1
Proof of Theorem 8.3
We prove the following Theorem
Theorem 8.3. For real numbers ξ1 , . . . , ξk , let Gξ1 ,...,ξk be the matrix whose (i, j)-entry is
Z
1
e2πi(ξi −ξj )t dt.
−1
Then
det(Gξ1 ,...,ξk ) = 2Õ(k
2)
Y
i<j
min(|ξi − ξj |2 , 1).
First, we note by the Cauchy-Binet formula that the determinant in question is equal to
Z 1Z 1
Z 1
2
...
det([e2πiξi tj ]i,j ) dt1 dt2 . . . dtk .
−1
−1
(27)
−1
P
We next need to consider the integrand in the special case when
|ξi | ≤ 1/8.
P
Lemma A.1. If ξi ∈ R and tj ∈ R, i |ξi |(maxi |ti |) ≤ 1/8 then
k Q
)
(
2
(2π)
i<j |ti − tj ||ξi − ξj |
| det([e2πiξi tj ]i,j )| = Θ
.
1!2! · · · k!
Proof. Firstly, by adding a constant to all the tj we can P
make them non-negative. This multiplies
the determinant by a root of unity, and at most doubles i |ξi |(maxi |ti |).
By continuity, it suffices to consider the ti to all be multiples of 1/N for some large integer N .
By multiplying all the tj by N and all ξi by 1/N , we may assume that all of the tj are non-negative
integers with t1 ≤ t2 ≤ . . . ≤ tk .
Let zi = exp(2πiξi ). Then our determinant is
h i
t
det zi j
,
i,j
which is equal to the Vandermonde determinant times the Schur polynomial sλ (zi ) where λ is the
partition λj = tj − (j − 1).
Therefore, this determinant equals
Y
(zi − zj )sλ (z1 , z2 , . . . , zk ).
i<j
The absolute value of
Y
i<j
(zi − zj )
k Q
Q
is approximately i<j (2πi)(ξi − ξj ), which has absolute value (2π)(2) i<j |ξi − ξj |. We have left
to evaluate the size of the Schur polynomial.
58
By standard results, sλ is a polynomial in the zi with non-negative coefficients, and all exponents
at most maxj |tj | in each variable. Therefore, the monomials with non-zero coefficients will all have
real part at least 1/2 and absolute value 1 when evaluated at the zi . Therefore,
|sλ (z1 , . . . , zk )| = Θ(|sλ (1, 1, . . . , 1)|).
On the other hand, by the Weyl character formula
Y tj − ti
sλ (1, 1, . . . , 1) =
=
j−i
i<j
Q
i<j
|ti − tj |
1!2! . . . k!
.
This completes the proof.
Next we prove our Theorem when the ξ have small total variation.
P
Lemma A.2. If there exists a ξ0 so that
|ξi − ξ0 | < 1/8, then
!
Q
23k(k−1)/2 π k(k−1) i<j |ξi − ξj |2
det(Gξ1 ,...,ξk ) = Θ
.
Q
(k!)3 k−1
n=0 (2n)!
Proof. By translating the ξi we can assume that ξ0 = 0.
By the above we have
Q
Z 1
Z 1Y
(2π)k(k−1) i<j |ξi − ξj |2
Θ(
)
...
|ti − tj |2 dt1 . . . dtk .
(1!2! · · · k!)2
−1
−1
i<j
We noteR that by the Cauchy-Binet formula the latter term is the determinant of the matrix M with
1
Mi,j = −1 ti+j dt. This is the Graham matrix associated to the polynomials ti for 0 ≤ i ≤ k − 1.
Applying Graham-Schmidt (without the renormalization step) to this set yields the basis Pn αn
n (n!)2
where αn = 2 (2n)!
is the inverse of the leading term of Pn . This polynomial has norm αn2 2/(2n + 1).
Therefore, the integral over the ti yields
k−1
Y
n=0
2n+1 (n!)2
.
(n + 1)(2n)!
This completes the proof.
Next we extend this result to the case that all the ξ are within poly(k) of each other.
Proposition A.3. If there exists a ξ0 so that |ξi − ξ0 | = poly(k) for all i, then
Y
2
det(Gξ1 ,...,ξk ) = 2Õ(k )
min(|ξi − ξj |2 , 1).
i<j
Proof. We begin by proving the lower bound. We note that for 0 < x < 1,
Z xZ x
Z 1
2
det(Gξ1 ,...,ξk ) ≥
...
det([e2πiξi tj ]i,j ) dt1 dt2 . . . dtk = xk det(Gξ1 /x,ξ2 /x,...,ξk /k ).
−x
−x
−1
Taking x = 1/ poly(k), we may apply the above Lemma to compute the determinant on the right
hand side, yielding an appropriate lower bound.
59
To prove the lower bound, we note that we can divide our ξi into clusters, Ci , where for any i, j
in the same cluster |ξi − ξj | < 1/k and for i and j in different clusters |ξi − ξj | ≥ 1/k 2 . We then
note as a property of Graham matrices that
Y
Y
Y
2
2
det(Gξ1 ,...,ξk ) ≤
det(G{ξj ∈Ci } ) = 2Õ(k )
|ξi − ξj |2 = 2Õ(k )
|ξi − ξj |2 .
Ci
i<j, in same cluster
i<j
This completes the proof.
Finally, we are ready to prove our Theorem.
Proof. Let I(t) be the indicator function of the interval [−1, 1].
Recall that there is a function h(t) so that for any function f that is a linear combination of
at most k complex exponentials that |h(t)f (t)|2 = Θ(|I(t)f (t)|2 ) and so that ĥ is supported on an
interval of length poly(k) < k C about the origin.
Note that we can divide our ξi into clusters, Ci , so that for i and j in a cluster |ξi − ξj | < k C+1
and for i and j in different clusters |ξi − ξj | > k C .R
Let G̃ξ1 ,ξ2 ,...,ξk0 be the matrix with (i, j)-entry R |h(t)|2 e(2πi)(ξi −ξj )t dt.
We claim that for any k 0 ≤ k that
0
det(G̃ξ1 ,ξ2 ,...,ξk0 ) = 2O(k ) det(Gξ1 ,ξ2 ,...,ξk0 ).
This is because both are Graham determinants, one for the set of functions I(t) exp((2πi)ξj t) and
the other for h(t) exp((2πi)ξj t). However since any linear combination of the former has L2 norm a
constant multiple of that the same linear combination of the latter, we have that
G̃ξ1 ,ξ2 ,...,ξk0 = Θ(Gξ1 ,ξ2 ,...,ξk0 )
as self-adjoint matrices. This implies the appropriate bound.
Therefore, we have that
det(Gξ1 ,...,ξk ) = 2O(k) det(G̃ξ1 ,...,ξk ).
However, note that by the Fourier support of h that
Z
|h(t)|2 e(2πi)(ξi −ξj )t dt = 0
R
if |ξi − ξj | > k C , which happens if i and j are in different clusters. Therefore G̃ is block diagonal
and hence its determinant equals
Y
Y
det(G̃ξ1 ,...,ξk ) =
det(G̃{ξj ∈Ci } ) = 2O(k)
det(G{ξj ∈Ci } ).
Ci
Ci
However the Proposition above shows that
Y
Y
2
det(G{ξj ∈Ci } ) = 2Õ(k )
min(1, |ξi − ξj |2 ).
Ci
i<j
This completes the proof.
60
A.2
Proofs of Lemma 5.3 and Lemma 5.4
We fix z1 , · · · , zk to be complex numbers on the unit circle and use Q(z) to denote the degree-k
k
Q
polynomial
(z − zi ).
i=1
Lemma 5.3. Let Q(z) be a degree k polynomial, all of whose roots are complex numbers with
Pk−1 (l)
rn,k · z l denote the residual polynomial of
absolute value 1. For any integer n, let rn,k (z) = l=0
rn,k (z) ≡ z n
(mod Q(z)).
(l)
Then, each coefficient of rn,k is bounded: |rn,k | ≤ 2k nk−1 for any l.
Proof. By definition, rn,k (zi ) = zin . From the polynomial interpolation, we have
Q
(z − zj )zin
k
X j∈[k]\i
Q
rn,k (z) =
.
(zi − zj )
i=1
j∈[k]\i
Let SymS,i P
be the
Q symmetry polynomial of z1 , · · · , zk with degree i among subset S ⊆ [k], i.e.,
zj . Then
SymS,i =
0
S 0 ⊆(Si )j∈S
(l)
rn,k = (−1)k−1−l
k
X
Sym[k]\i,k−1−l ·zin
Q
.
(zi − zj )
i=1
j∈[k]\i
(l)
We omit (−1)k−1−l in the rest of proof and use induction on n, k, and l to prove |rn,k | ≤
zn
(l)
|rn,k |
k−1
l
n
k−1
.
Base Case of n: For any n < k, from the definition, r(z) =
and
≤ 1.
l
Suppose it is true for any n < n0 . We consider rn0 ,k from now on. When k = 1, rn,0 = z1n is
bounded by 1 because z1 is on the unit circle of C.
Given n0 , suppose the induction hypothesis is true for any k < k0 and any l < k. For k = k0 ,
61
(k −1)
(k −1)
rn00,k0
=
k0
X
i=1
=
kX
0 −1
i=1
=
kX
0 −1
i=1
=
=
(k −1)
zin0
(zi − zj )
zin0
+
(zi − zj )
Q
=
kX
0 −1
i=1
=
kX
0 −1
i=1
for l = k0 − 2, · · · , 0.
(zk0 − zj )
kX
0 −1
i=1
z n0 −1
Q i
j∈[k0 −1]\i
z n0 −1
Q i
j∈[k0 −1]\i
(k −2)
rn00−1,k0 −1
(zi − zj )
(zi − zj )
+ zk0 ·
+
zkn00
Q
j∈k0 \k0
j∈[k0 ]\i
z n0 −1 zk0
Qi
j∈k0 \i
+ zk0
n0 −2
k0 −2
+
(zk0 − zj )
(zi − zj )
k0
X
i=1
n0 −2
k0 −1
+
Q
j∈k0 \i
=
(zi − zj )
n0 −1
k0 −1
zkn00
j∈k0 \k0
z n0 −1
Q i
(k −1)
rn00−1,k0
(k −1)
k0
X
Sym[k0 ]\i,k0 −1−l ·zin0
Q
=
(zi − zj )
j∈[k0 ]\k0
zin0 − zin0 −1 zk0 + zin0 −1 zk0
Q
+
(zi − zj )
(k −2)
i=1
n0
k0 −1
zkn00
Q
j∈[k0 ]\i
Hence |rn00,k0 | ≤ |rn00−1,[k0 −1] | + |rn00−1,k0 | ≤
(l)
rn0 ,k0
k0 −1
l
then prove that |rn0 ,k0 | ≤
j∈[k0 ]\i
kX
0 −1
i=1
=
Q
(l)
n0
k0 −1
we first prove that |rn00,k0 | ≤
(zk0 − zj )
. For l < k0 − 1, we have
let l0 = k0 − 1 − l
j∈[k0 ]\i
Sym[k0 −1]\i,l0 + Sym[k0 −1]\i,l0 −1 ·zk0 zin0
Sym[k0 −1],l0 ·zkn00
Q
+ Q
(zi − zj )
(zk0 − zj )
j<k0
j∈[k0 ]\i
Sym[k0 −1]\i,l0 ·(zi − zk0 )zin0 −1 + Sym[k0 −1]\i,l0 ·zk0 zin0 −1 + Sym[k0 −1]\i,l0 −1 ·zk0 zin0
Q
(zi − zj )
j∈[k0 ]\i
Sym[k0 −1],l0 ·zkn00
+ Q
j<k0
=
kX
0 −1
i=1
=
kX
0 −1
i=1
=
(zk0 − zj )
Sym[k0 −1]\i,l0 ·(zi − zk0 )zin0 −1 + Sym[k0 −1],l0 ·zk0 zin0 −1 Sym[k0 −1],l0 ·zkn00
Q
+ Q
(zk0 − zj )
(zi − zj )
j<k0
j∈[k0 ]\i
0 −1
Sym[k0 −1]\i,l0 zin0 −1 kX
Sym[k0 −1],l0 ·zk0 zin0 −1 Sym[k0 −1],l0 ·zkn00
Q
Q
+
+ Q
(zi − zj )
(zi − zj )
(zk0 − zj )
i=1
j∈[k0 −1]\i
(l−1)
rn0 −1,k0 −1
+ Sym[k0 −1],k0 −1−l ·zk0 ·
(l)
By induction hypothesis, |rn0 ,k0 | ≤
k0 −2
l−1
Now we finish the proof of Lemma 5.4.
j<k0
j∈[k0 ]\i
(k −1)
rn00−1,k0
n0 −1
k0 −2
+
62
k0 −1
l
n0 −1
k0 −1
≤
k0 −1
l
n0
k0 −1
.
Lemma 5.4. For any k ∈ Z and any z1 , · · · , zk on the unit circle of C, there always exists a degree
m
P
m = O(k 2 log k) polynomial P (z) =
cj z j with the following properties:
j=0
P (zi ) = 0, ∀i ∈ {1, · · · , k},
Property I
Property II
c0 = 1,
|cj | ≤ 11, ∀j ∈ {1, · · · , m}.
Property III
Let m = 10k 2 log k and P denote a set of polynomials that has degree at most m, and all the
coefficients are integers chosen from {−5, · · · , −1, 0, 1, · · · , 5}, i.e.,
(
)
m
X
P := P (z) =
αi z i | ∀i ∈ {0, 1, · · · , m}, |αi | ≤ 2 .
i=0
Claim A.4. There exists P ∗ (z) =
m
P
αi z i with coefficient |αi | ≤ 10 for any i ∈ {0, 1, · · · , m}, such
i=0
that every coefficient of P ∗ (z) mod Q(z) is bounded by 2−m .
Proof. For P (z) =
m
P
αi z i ∈ P, P (z) mod Q(z) ≡
i=0
P (z)
mod Q(z) =
m
X
i=0
αi
m
P
αi rn,k (z) from the definition rn,k (z). Hence
i=0
k−1
X
(l)
ri,k z l =
l=0
k−1
X
zl
l=0
m
P
(l)
m
X
(l)
αi ri,k .
i=0
m
P
2k ik−1 ≤ 2k mk .
i=0
i=0
k k k
m
m
< 11m , there exists
At the same time, |P| = 11 . From the pigeonhole principle and (22−m
)2
Each coefficient in P (z) mod Q(z) is bounded by |
αi ri,k | ≤ 5
P1 , P2 ∈ P such that for P ∗ (z) = P1 (z) − P2 (z), P ∗ (z) mod Q(z) =
|γi | ≤ 2−m .
Let r(z) =
k−1
P
γi z i = P ∗ (z)
i=0
|P ∗ (0) − r(0)| ≥
k−1
P
γi z i where each coefficient
i=0
mod Q(z) for convenience. If P ∗ (0) (the constant term of P ∗ ) is
nonzero, then
0.99 from the above lemma. Therefore the polynomial
satisfies the three properties in Lemma 5.4.
Otherwise, we assume z l is the first term in P ∗ (z) with a non-zero coefficient. Let
r−l,k (z) = z −l
Q
(z − zj )zi−l
k
X
j∈[k]\i
Q
mod Q(z) =
.
(zi − zj )
i=1
For convenience, we use zS =
Q
i∈S
P ∗ (z)−r(z)
P ∗ (0)−r(0)
j∈[k]\i
zi for any subset S ⊆ [k]. Notice that zi−l =
l
z[k]\i
l
z[k]
. Hence r−l,k (z) =
0 (z)/z l where r 0 is the polynomial for k units roots z
rl,k
[k]\1 , · · · , z[k]\k . So each coefficients of r is
[k]
still bounded by 2k lk , which is less than 2−m/2 .
Eventually we choose P ∗ (z)/z l − r(z) · r−l,k (z) and renormalize it to satisfy the three properties
in Lemma 5.4.
63
A.3
Proof of Lemma 4.3
Lemma 4.3. For any degree d polynomial P (t) : R → C with derivative P 0 (t), we have,
Z
1
−1
(1 − t2 )|P 0 (t)|2 dt ≤ 2d2
Z
1
−1
|P (t)|2 dt.
(28)
Given a degree d polynomial P (x), we rewrite P (x) as a linear combination of the Legendre
polynomials:
d
X
P (x) =
αi Li (x).
i=0
We use Fi (x) = (1 − x2 )L0i (x) for convenience. From the definition of the Legendre polynomials in
the Equation (3), Fi0 (x) = −i(i + 1) · Li (x) and Fi00 (x) = −i(i + 1) · L0i (x).
Hence we have
Z −1
Z −1
2
0
2
(1 − x )|P (x)| dx =
(1 − x2 )P 0 (x) · P 0 (x)dx
1
1
Z −1 X
X −F 00 (x)
i
dx
=
αi Fi (x) ·
αi
i(i + 1)
1
i∈[d]
i∈[d]
X
X −F 0 (x) 1
i
=
αi
αi Fi (x) ·
i(i + 1) −1
i∈[d]
i∈[d]
Z −1 X
0
X
F (x)
αi i
dx
+
αi Fi0 (x) ·
i(i
+ 1)
1
i∈[d]
i∈[d]
Z −1 X
X i(i + 1) · Li (x)
dx
=
αi · i(i + 1) · Li (x) ·
αi
i(i + 1)
1
i∈[d]
i∈[d]
X
2
2
=
|αi | i(i + 1)kLi kT
i∈[d]
≤ d(d + 1)kP k2T
A.4
Proof of Lemma 6.2
1 −2πiσaf
1 −2πiσa(f /σ+b)
Lemma 6.2. P\
x
b(f ) and P\
x
b(f /σ + b)
σ,a,b x(σ(f − b)) = σ e
σ,a,b x(f ) = σ e
64
Proof. Let’s compute the Fourier Transform of (Pσ,a,b x)(t),
P\
σ,a,b x(f )
Z +∞
(Pσ,a,b (x))(t)e−2πif t dt
=
−∞
+∞
=
Z
−∞
x(σ(t − a))e−2πiσbt e−2πif t dt
−2πi(σab+f a)
=e
=e−2πi(σab+f a)
Z
+∞
−∞
Z +∞
x(σ(t − a))e−2πiσb(t−a) e−2πif (t−a) dt
x(σt)e−2πiσbt e−2πif t dt
by shifting t by a
by replacing t − a by t
−∞
Z +∞
−2πi(σab+f a)
1
= e
x(σt)e−2πibσt e−2πif σt/σ dσt
σ
−∞
Z +∞
1
= e−2πi(σab+f a)
x(t)e−2πi(b+f /σ)t dt
σ
−∞
1 −2πiaσ(f /σ+b)
= e
x
b(f /σ + b)
σ
by replacing tσ by t
by definition of FT
The first result follows immediately by replacing f /σ + b by f 0 , which gives
1 −2πiaσf 0
0
x
b(f 0 ).
e
P\
σ,a,b x(σ(f − b)) =
σ
Thus, we complete the proof of this Lemma.
A.5
Proof of Lemma 3.5
Lemma 3.5. Given any function x(t) : R → C with max |x(t)|2 ≤ dkx(t)k2T . Let S denote a set of
t∈[0,T ]
points from 0 to T . If each point of S is chosen uniformly at random from [0, T ], we have
"
#
1 X
2
|x(ti )|2 − kx(t)k2T ] ≥ kx(t)k2T
≤ e−Ω( |S|/d)
Pr
|S|
i∈S
Proof. Let M denote max |x(t)|2 . Replacing Xi by
t∈[0,T ]
|x(ti )|2
M
and n by |S| in Lemma B.2, we obtain
that
=⇒
=⇒
=⇒
2
Pr[|X − µ| > µ] ≤ 2 exp(− µ)
3
"
#
X |x(ti )|2
kx(t)k2T
kx(t)k2T
2
> |S|
Pr
− |S|
≤ 2 exp(− µ)
M
M
M
3
i∈S
"
#
1 X
2
Pr
|x(ti )|2 − kx(t)k2T ≥ kx(t)k2T ≤ 2 exp(− µ)
|S|
3
i∈S
"
#
1 X
2 kx(t)k2T
Pr
|x(ti )|2 − kx(t)k2T ≥ kx(t)k2T ≤ 2 exp(− |S|
)
|S|
3
M
i∈S
2
which is less than 2 exp(− 3 |S|/d), thus completes the proof.
65
A.6
Proof of Lemma 3.10
Lemma 3.10. For any polynomial P (t) of degree at most d from R to C, for any interval [S, T ],
Z T
1
2
2
|P (t)|2 dx.
max |P (t)| ≤ (d + 1) ·
T −S S
t∈[S,T ]
Proof. Let t∗ = arg max|P (t)|2 . If t∗ ∈ (S, T ), then it is enough to prove that
t∈[S,T ]
1
|P (t )| ≤ (d + 1) ∗
t −S
∗ 2
2
Z
t∗
1
|P (x)| dx and |P (t )| ≤ (d + 1)
T − t∗
∗ 2
2
S
2
Z
T
t∗
|P (x)|2 dx
on the two intervals [S, t∗ ] and [t∗ , T ] separately.
Without loss of generality, we will prove the inequality for S = −1 and t∗ = T = 1. We find the
minimum kP (x)k2T assuming |P (1)|2 = 1. Because the first (d + 1) Legendre polynomials provide a
basis of polynomials of degree at most d and their evaluation Ln (1) = 1 for any n, we consider:
Z 1
|P (x)|2 dx
min
α0 ,α1 ,··· ,αd ∈C
−1
d
X
P (x) =
s.t.
αi Li (x)
i=0
|P (1)| = |
d
X
i=0
αi | = 1.
We simplify the integration of P (x)2 over [−1, 1] by the orthogonality of Legendre polynomials:
! d
Z 1
Z 1 X
d
X
αj Lj (x) dx
|P (x)|2 dx =
αi Li (x) ·
−1
−1
=
Z
1
i=0
d
X
−1 i=0
=
d
X
i=0
Using
R1
2
−1 |P (x)| dx
=
d
P
j=0
|αi |2 Li (x)2 +
|αi |2
X
αi αj Li (x)Lj (x)dx
i6=j
2
by Fact 3.9
2i + 1
2
|αi |2 2i+1
, we simplify the optimization problem to
i=0
d
X
min
α0 ,α1 ,··· ,αd ∈C
|αi |2
i=0
d
X
s.t.
2
2i + 1
αi = 1
i=0
From the Cauchy-Schwarz inequality, we have
d
X
i=0
Therefore
d
P
2
|αi |2 2i+1
≥
i=0
2
(d+1)2
2
αi
≤
d
X
i=0
2
|αi |
2i + 1
2
!
and |P (1)|2 ≤ (d + 1)2 ·
66
1
2
d
X
2i + 1
i=0
R1
2
−1 |P (x)|
!
.
2 dx.
Algorithm 1 Linear regression algorithms
procedure LinearRegression(A, b,) — Fact B.3
2:
x0 ← arg minkAx − bk2 .
1:
x
return x0
4: end procedure
5: procedure LinearRegressionW(A, b, w) — Fact B.3
d
P
6:
x0 ← arg min wi |(Ax)i − bi |2 .
3:
x
i=1
return x0
8: end procedure
7:
B
Known Facts
This section provides a list of well-known facts existing in literature.
B.1
Inequalities
We state the Hölder’s inequality for complex numbers. We will use the corresponding version
p = q = 2 of Cauchy-Schwarz inequality for complex numbers.
Lemma B.1 (Hölder’s inequality). If S is a measurable subset of Rn with the Lebesgue measure,
and f and g are measurable complex-valued functions on S, then
Z
Z
Z
1
1
|f (x)g(x)|dx ≤ ( |f (x)|p dx) p ( |g(x)|q dx) q
S
S
S
Lemma B.2 (Chernoff Bound [Tar09],[Che52] ). Let X1 , X2 , · · · , Xn be independent random variables. Assume that 0 ≤ Xi ≤ 1 always, for each i ∈ [n]. Let X = X1 + X2 + · · · + Xn and
n
P
µ = E[X] =
E[Xi ]. Then for any > 0,
i=1
Pr[X ≥ (1 + )µ] ≤ exp(−
B.2
2
2
µ) and Pr[X ≥ (1 − )µ] ≤ exp(− µ).
2+
2
Linear regression
Given a linear subspace span{~v1 , · · · , ~vd } and n points, we always use `2 -regression to find a vector
as the linear combination of ~v1 , · · · , ~vd that minimizes the distance of this vector to those n points.
Fact B.3. Given an n × d matrix A and an n × 1 column vector b , it takes O(ndω−1 ) time to
output an x0 such that
x0 = arg minkAx − bk2 .
x
where ω is the exponent of matrix multiplication[Wil12].
Notice that weighted linear regression can be solved by linear regression solver as a black-box.
67
Algorithm 2 Multipoint evaluation of a polynomial
procedure MultipointEvaluation(P, {t1 , t2 , · · · , td }) — Fact B.4
2:
return P (t1 ), P (t2 ), · · · , P (td )
3: end procedure
1:
B.3
Multipoint evaluation of a polynomial
Given a degree-d polynomial, and n locations. The naive algorithm of computing the evaluations
at those n locations takes O(nd). However, the running time can be improved to O(n poly(log d))
by using this well-known result,
Fact B.4 ([BS12]). Given a degree-d polynomial P (t), and a set of d locations {t1 , t2 , · · · , td }. There
exists an algorithm that takes O(d logc d) time to output the evaluations {P (t1 ), P (t2 ), · · · , P (td )},
for some constant c.
C
Analysis of Hash Functions and Filter Functions
C.1
b ))
Analysis of filter function (H(t), H(f
b )) in this section.
We construct the Filter function (H(t), H(f
We fix the interval to be supp(rect1 ) = [−1/2, 1/2] instead of [0, T ] for convenience. We first define the filter function H1 (t) which preserves the energy of a k-Fourier-sparse signal x∗ on [−1/2, 1/2]
to the signal H1 · x∗ on [−∞, +∞].
√
Definition C.1. Let s1 = Θ(k 4 log4 k), ` = Ω(k log k/δ) be a even number, and s0 = C0 s1 `
for some constant C0 that will normalize H1 (0) = 1. Recall that rects (t) = 1 iff |t| ≤ s/2 and
sin(πf s)
\
rect
s (f ) = sinc(f s) =
πf s .
b 1 (f ) as follows:
We define the filter function H1 (t) and its Fourier transform H
b 1 (f ) = s0 · (rects (f ))∗` · sinc (f s2 ) ,
H
1
1
= s0 ·
−s1 /2
∗`
s1 /2
·`
·
1
−1/s2
1/s2
H1 (t) = s0 · (sinc(s1 t)) ∗ rects2 (t)
1
= s0 ·
−1/s1
1/s1
·`
1
∗
−s2 /2
s2 /2
where s0 is a fixed parameter s.t. H1 (0) = 1.
We provide some basic properties about our filter function. Notice that sinc(t) =
is defined to be 1 ) has the following properties (shown in Figure 4):
1. ∀t ∈ R, 1 −
(πt)2
3!
≤ sinc(t) ≤ 1.
2. ∀|t| ≤ 1.2/π, sinc(t) ≤ 1 −
3. ∀|t| > 1.2/π, | sinc(t)| ≤
t2
8.
1
π|t| .
68
sin(πt)
πt
( sinc(0)
sinc(t)
1−
t2
8
1
π|t|
1
−1
1
Figure 4: The Property of sinc(t).
Claim C.2.
R
1.2
πs1
1.2
− πs
1
(sinc(s1 t))` dt h
1√
.
s1 `
Proof. We use the above properties for the sinc function to prove the upper bound:
Z
1.2
+ πs
1
1.2
− πs
1
1
(sinc(s1 t)) dt =
s1
`
2
=
s1
Z
+1.2/π
(sinc(t))` dt
−1.2/π
Z √8/`
0
Z
(sinc t)` dt + √
1.2/π
√ −1
8/`
(sinc(t))` dt
8/`
√
X Z (i+1)
p
2
≤
8/` +
√
s1
i 8/`
8/`
i=1
≤
.
!
1.2/π
(1 − x2 /8)` dx
1.2/π
√ −1
8/`
X p
p
2
−i2
8/` +
8/`
·
2
s1
i=1
1
√ .
s1 `
We prove the lower bound:
Z
1.2
πs1
2
(sinc(s1 t)) dt =
1.2
s
1
−
`
πs1
2
≥
s1
Z √8/`
`
Z
1.2/π
0
(sinc t) dt + √
0
π 2 t2 `
(1 −
) dt
6
Z √8/`
!
`
!
(sinc(t)) dt
8/`
1
& √ .
s1 `
1.2 1.2
We bound the integration outside [− πs
,
] from the last property of the sinc function.
1 πs1
R +∞
Claim C.3. 1.2 (sinc(s1 t))` dt = O(1.2−` ).
πs1
From these two claims, we have the existence of s0 .
√
Claim C.4. There exists a universal constant C0 and s0 = C0 s1 ` such that H1 (0) = 1.
69
1.0
0.8
0.6
0.4
0.2
0.0
0.2 6
1.0
0.8
0.6
0.4
0.2
0.0
0.2 6
(sinc(t))`
4
2
0
4
2
6
(sinc(t))`
4
2
0
4
2
6
R (1/2−2/s )−t
Figure 5: The light red area represents (1/2−2/s11) s0 · sinc (s1 (τ ))` dτ and the light green area
R 1/2−2/s +t
represents 1/2−2/s11 s0 · sinc (s1 (τ ))` dτ .
Proof. Because ` is a large even number,
R
rect1−2/s1
sinc(s1 t)` dt h
1√
s1 `
from all discussion above.
c1 (f ) .
We show several useful properties about the Filter functions H1 (t), H
√
Lemma C.5. Given s0 , s1 , s2 , `, where s22 + s11 ≤ 1/2 and s0 = C0 s1 ` for some constant C0 . The
b 1 (f ))[s0 , s1 , s2 , `] has the following properties,
filter function (H1 (t), H
s0 2π −`
s2
1
·
, 1], if |t| ≤
− .
s1 ` − 1
2
s1
s2
1
1
Property II : H1 (t) ∈ [0, 1], if
−
≤ |t| ≤
2
s1
2
−`
1
Property III : H1 (t) ≤ s0 s2 (s1 |t| − s1 + 2)2 + 1
, ∀|t| >
2
s
`
s
`
1
1
b 1 (f )) ⊆ [−
,
]
Property IV : supp(H
2 2
Property I :
H1 (t) ∈ [1 −
Proof of Property I. First, H1 (0) = 1 follows by definition of s0 , then we can prove the upper
bound for H1 (t) by showing for any t > 0, H1 (0) − H1 (t) > 0 always holds ,
70
By definition of sinc function, we know that sinc(s1 t)` reaches 0 at all the points { s11 +i s21 |i ∈ N}.
By definition of s1 , we know that s11 21 − s11 . For any t > 0,
H1 (0) − H1 (t)
Z
Z
=
s0 · sinc (s1 (τ ))` · rect1−2/s1 (0 − τ )dτ − s0 · sinc (s1 (τ ))` · rect1−2/s1 (t − τ )dτ
Z 1/2−2/s1 +t
Z 1/2−2/s1
`
s0 · sinc (s1 (τ ))` dτ
s0 · sinc (s1 (τ )) dτ −
=
−(1/2−2/s1 )
−(1/2−2/s1 )+t
=
Z
−(1/2−2/s1 )
−(1/2−2/s1 )+t
Z 1/2−2/s1 +t
s0 · sinc (s1 (τ ))` dτ −
≥ 0,
1/2−2/s1
s0 · sinc (s1 (τ ))` dτ
shown in Figure 5
where the last inequality follows by choosing s1 to be an integer. Thus, we prove an upper bound
for H1 (t). Third, we show the lower bound for H1 (t),
Z +∞
H1 (t) =
s0 · sinc(s1 τ )·` rects2 (t − τ )dτ
=
−∞
s
t+ 22
Z
s
t− 22
= 1−
Z
|
s0 · sinc(s1 τ )·` dτ
+∞
s
t+ 22
·`
s0 · sinc(s1 τ ) dτ −
{z
}
A
Z
t−
| −∞
s2
2
s0 · sinc(s1 τ )·` dτ
{z
}
B
Thus, as long as we can upper bound the term A and B, then we will have a lower bound for the
H1 (t), for any |t| ≤ s22 − s11 .
A =
≤
≤
Z
+∞
s
t+ 22
Z
Z
+∞
1
s1
+∞
1
s1
s0 · sinc(s1 τ )·` dτ
s0 · sinc(s1 τ )·` dτ
s0 · (s1 πτ )−` dτ
1
(1/s1 )−`+1
`−1
s0
1
· (π)−`
s1
`−1
= s0 · (s1 π)−`
=
Similarly, we can bound the term B in the same way.
Proof of Property II. In the proof of Property I, we already show that ∀t, H1 (t) ≤ 1. Thus, the
upper bound of Property II is also holding. The lower bound follows by both sinc(s1 t)·` and rects2 (t)
are always nonnegative, thus the convolution of these two functions has to be nonnegative.
71
Proof of Property III. Let’s prove the case when t > 1, since H1 (t) is symmetric, then the case
b 1 (f )), we have
t < −1 will also hold. By definition of (H1 (t), H
Z +∞
H1 (t) = s0 ·
sinc(s1 (t − τ ))·` rects2 (τ )dτ
−∞
= s0 ·
= s0 ·
Z
s2
2
s2
2
s2
2
−
Z
s
− 22
sinc(s1 (t − τ ))·` dτ
sinc(s1 (τ − t))·` dτ
We’d like to choose a middle point τ0 , and then separated the interval into two parts, one is [− s22 , τ0 ]
and the other is [τ0 , s22 ]. To choose a reasonable τ0 , we need to use the following simple facts,
sin(x) `
) ≤ x−` if x ≥ 1.2
x
sin(x) `
x2
(
) ≤ (1 − )` if x < 1.2
x
8
(
1.2
1.2
or τ0− = t − πs
. By relationship between
Thus, |πs1 (τ0 − t)| = 1.2, which implies that τ0+ = t + πs
1
1
−
s2
1
1.2
1
1
s1 and s2 , we know τ0 > 2 − πs1 > 2 − s1 ≥ 2 . Thus, we can use the case x < 1.2 to upper bound
the H1 (t),
H1 (t)
≤ s0 · s2 ·
≤ s0 · s2
max
τ ∈[−s2 /2,s2 /2]
max
(1 −
τ ∈[−s2 /2,s2 /2]
(s1 π( s22
= s0 · s2 (1 −
−
≤ s0 · s2 · (e
sinc(s1 (τ − t))·`
8
(s1 π(τ − t))2 `
)
8
− t))2
s
(s1 π( 22 −t))2
8
)`
by 1 − x ≤ e−x
)`
2
≤ s0 · s2 · (e(s1 (t−s2 /2)) )−`
by 1 < π 2 /8
≤ s0 · s2 · (1 + (s1 (t − s2 /2))2 )−`
by 1 + x ≤ ex
≤ s0 · (1 + (s1 (t − s2 /2))2 )−`
by s1 ≤ 1.
Thus, we complete the proof.
Proof of Property IV. Because of the support of rects1 (f ) is s1 , then the support of (rects1 (f ))∗` =
b 1 (f ) is defined to be the (rects (f ))∗` multiplied by sinc(f s2 ), thus supp(H
b 1 (f )) ⊆
s1 `. Since H
1
s1 ` s1 `
[− 2 , 2 ].
b )) to be the filter function
Definition C.6. Given any 0 < s3 < 1, 0 < δ < 1, we define (H(t), H(f
c
(H1 (t), H1 (f )) by doing the following operations
• Setting ` = Θ(k log(k/δ)),
72
• Setting s2 = 1 −
2
s1 ,
• Shrinking by a factor s3 in time domain,
H(t) = H1 (t/s3 )
b ) = s3 H
c1 (s3 f )
H(f
(29)
(30)
b ) in the frequency
We call the “heavy cluster" around a frequency f0 to be the support of δf0 (f ) ∗ H(f
domain and use
b ))| = s1 · `
∆h = | supp(H(f
(31)
s3
to denote the width of the cluster.
b ) .
We show several useful properties about the Filter functions H(t), H(f
Lemma 6.6. Given s0 , s1 , 0 < s3 < 1, ` > 1, 0 < δ < 1, where ` = Θ(k log(k/δ)).The filter function
b )) has the following properties,
(H(t), H(f
1
2
H(t) ∈ [1 − δ, 1], when |t| ≤ ( − )s3 .
2 s1
2
1
1
Property II : H(t) ∈ [0, 1], when ( − )s3 ≤ |t| ≤ s3 .
2 s1
2
|t| 1
1
Property III : H(t) ≤ s0 · (s1 ( − ) + 2)−` , ∀|t| > s3 .
s3 2
2
s
`
s
`
b )) ⊆ [− 1 , 1 ].
Property IV : supp(H(f
2s3 2s3
Property I :
For any exact k-Fourier-sparse signal x∗ (t), we shift the interval from [0, T ] to [−1/2, 1/2] and
consider x∗ (t) for t ∈ [−1/2, 1/2] to be our observation, which is also x∗ (t) · rect1 (t).
Z +∞
Z +∞
2
∗
Property V :
x (t) · H(t) · (1 − rect1 (t)) dt < δ
|x∗ (t) · rect1 (t)|2 dt.
−∞
+∞
Property VI :
Z
−∞
−∞
|x∗ (t) · H(t) · rect1 (t)|2 dt ∈ [1 − , 1] ·
Z
+∞
−∞
|x∗ (t) · rect1 (t)|2 dt.
for arbitrarily small constant .
b ) inheriting H1 (t), H
b 1 (f ).
The Property I, II, III and IV follow by filter function H(t), H(f
Proof of Property V.
∀t ∈
/ [−1/2, 1/2], we have,
|x∗ (t) · H(t)|2
≤ |x∗ (t)|2 · |H(t)|2
≤ |x∗ (t)|2 · (s1 (|t| − 1/2) + 2)−`
Z +∞
|t| 1
≤ k 7 · (2kt)2.5k ·
|x∗ (t) · rect1 (t)|2 dt · (s1 ( − ) + 2)−`
s3 2
−∞
Z +∞
|t| 1
≤ tO(k log k) ·
|x∗ (t) · rect1 (t)|2 dt · (s1 ( − ) + 2)−`/2 .
s3 2
−∞
73
by Property III of H1 (t)
by Lemma 5.5
(32)
0.06
0.05
0.04
0.03
0.02
0.01
0.00
0.0140
1.0
0.8
0.6
0.4
0.2
0.0
0.2 4
Given signal (Frequency domain)
c(f)
H
x ∗ (f)
H
·x ∗ (f)
d
f = ± s1 `/(2s3 )
c
30
20
0
10
10
20
Fourier transform (Time domain)
30
40
H(t)
x ∗ (t)
H ·x ∗ (t)
t = ± 0.5
t = ± 0.5s3
t = ± (0.5−2/s1 )s3
3
1
2
1
0
2
3
4
\
Figure 6: H
· x∗ (f ) and H · x∗ (t).
Thus taking the integral finishes the proof because ` & k log(k/δ).
Proof of Property VI. First, because of for any t, |H1 (t)| ≤ 1, thus we prove the upper bound for
LHS,
Z +∞
Z +∞
∗
2
|x (t) · H(t) · rect1 (t)| dt ≤
|x∗ (t) · 1 · rect1 (t)|2 dt.
−∞
−∞
Second, as mentioned early, we need to prove the general case when s3 = 1 − 1/ poly(k). Define
interval S = [−s3 ( 21 − s11 ), s3 ( 12 − s11 )], by definition, S ⊂ [−1/2, 1/2]. Then define S = [−1/2, 1/2]\S,
which is [−1/2, −s3 ( 12 − s11 )) ∪ (s3 ( 12 − s11 ), 1/2]. By Property I, we have
Z
S
∗
2
2
|x (t) · H(t)| dt ≥ (1 − δ)
74
Z
S
|x∗ (t)|2 dt
(33)
Then we can show
Z
|x∗ (t)|2 dt
S
≤ |S| ·
max
t∈[−1/2,1/2]
|x∗ (t)|2
2
e 4)
≤ (1 − s3 (1 − )) · O(k
s1
Z 1
2
.
|x∗ (t)|2 dt
Z
1
2
− 12
|x∗ (t)|2 dt
by Lemma 5.1
by min(
− 12
Combining Equations (33) and (34) gives a lower bound for LHS,
Z +∞
|x∗ (t) · H(t) · rect1 (t)|2 dt
−∞
Z
≥
|x∗ (t)H(t)|2 dt
S
Z
≥ (1 − 2δ) |x∗ (t)|2 dt
ZS
Z
∗
2
≥ (1 − 2δ)
|x (t)| dt − (1 − 2δ) |x∗ (t)|2 dt
ZS∪S
ZS
≥ (1 − 2δ)
|x∗ (t)|2 dt − (1 − 2δ)
|x∗ (t)|2 dt
S∪S
≥ (1 − 2δ − )
≥ (1 − 2)
Z
Z
(34)
by Equation (33)
by Equation (34)
S∪S
1
2
− 12
+∞
−∞
1
e 4)
, s1 ) ≥ O(k
1 − s3
|x∗ (t)|2 dt
|x∗ (t) · rect1 (t)|2 dt
by δ
b )) on [−1/2, 1/2] with signal x(t) on [0, T ], we will scale the
Remark C.7. To match (H(t), H(f
time domain from [−1/2, 1/2] to [−T /2, T /2] and shift it to [0, T ]. For example, the rectangle
function in Property V and VI will be replaced by rectT (t − T /2). For the parameters s0 , s1 , s3 , δ, `
in the definition of H, we always treat them as numbers. We assume T has seconds as unit and ∆h
has Hz as unit . For example, in time domain, the Property I becomes that given T > 0,
H(t) ∈ [1 − δ, 1] if |t −
1
1
T
| ≤ ( − )s3 · T
2
2 s1
In frequency domain, the Property IV becomes
b )) ⊆ [−
supp(H(f
∆h ∆h
s1 `
,
], where ∆h =
.
2 2
s3 T
(35)
Lemma C.8. Let H(t) denote the function defined in Definition C.6. For any t ∈ [− 12 , 12 ], there
e
exists an algorithm that takes O(s1 + ` log(s1 ) + log(1/)) time to output a value H(t)
such that
e
(1 − )H(t) ≤ H(t)
≤ (1 + )H(t).
75
0.06
0.05
0.04
0.03
0.02
0.01
0.00
0.0140
0.15
0.10
0.05
0.00
0.05
0.10 4
Given signal (Frequency domain)
c(f)
H
x ∗ (f)
H
·x ∗ (f)
d
f = ± s1 `/(2s3 )
c
30
20
10
0
10
20
Fourier transform (Time domain)
30
40
H(t)
x ∗ (t)
t = ± 0.5
3
2
1
0
1
2
4
3
Fourier transform (Time domain)
0.15
0.10
0.05
0.00
0.05
0.10 4
0.15
0.10
0.05
0.00
0.05
0.10 4
H(t)
x ∗ (t)
H ·x ∗ (t)
t = ± 0.5
3
2
1
0
1
2
Fourier transform (Time domain)
4
3
H(t)
x ∗ (t)
H ·x ∗ (t)
t = ± 0.5
3
2
1
0
1
2
3
4
Figure 7: Property VI of filter function H(t), the light green area represents RHS(without scalar)
of Property VI of filter H, the light red area represents LHS of Property VI of filter H, the light
yellow area represents the difference. Property VI says the light yellow area is only a small constant
fraction of the light green area.
76
Proof. We will show that using a low degree polynomial with sufficiently large degree is able to
approximate the sinc function. By definition of filter function,
Z +∞
sinc(s1 τ )·` rects2 (t − τ )dτ
H(t) = s0 ·
= s0 ·
s0
=
πs1
s0
=
πs1
=
s0
πs1
−∞
s
t+ 22
Z
s
t− 22
Z
(t+
(
sin(πs1 τ ) `
) dτ
πs1 τ
s2
)πs1
2
s
(t− 22 )πs1
Z
(t+
s2
)πs1
2
s2
)πs1
2
s
(t+ 22 )πs1
(t−
Z
s
(t− 22 )πs1
(
sin(τ ) `
) dτ
τ
∞
X
i=0
τ 2i
(−1)i
(2i + 1)!
!`
dτ
by Taylor expansion
(A + B)` dτ
P
P
τ 2i
i τ 2i
, and B = ∞
where the last step follows by setting A = di=0 (−1)i (2i+1)!
i=d+1 (−1) (2i+1)! .
Denote I + = (t+ s22 )πs1 and I − = (t− s22 )πs1 . Because of t ∈ [−1/2, 1/2], then max(|I + |, |I − |) =
O(s1 ). The goal is to show that
(1 − )
Z
I+
`
(A + B) dτ ≤
I−
Z
I+
I−
`
A dτ ≤ (1 + )
Z
I+
(A + B)` dτ
I−
Let’s prove an upper first,
Z
I+
I−
=
Z
(A + B − B)` dτ
I+
`
(A + B) dτ +
I−
≤
≤
≤
≤
Z
j=1
I+
`
(A + B) dτ +
I−
Z
I+
`
(A + B) dτ +
I+
I−
Z
I+
I−
(A + B)` dτ + `2` ·
I+
I−
X̀ Z
j=1
I+
I−
X̀ Z
j=1
I−
Z
X̀ Z
I+
I−
`
(A + B)`−j (−B)j dτ
j
`
|A + B|`−j |B|j dτ
j
`
|A + B|`−j dτ · max |B|j
j
τ ∈[I − ,I + ]
by |H(t)| ≤ 1 and |B|j ≤ |B|
max |B|
τ ∈[I − ,I + ]
(A + B)` dτ + · (s1 )−Θ(`)
≤(1 + )
Z
by Claim C.9
I+
(A + B)` dτ
by Claim C.10
I−
where all the steps by setting d & s1 +` log(s1 )+log(1/). Similarly, we can prove a lower bound.
P
τ 2i
O(`) .
Claim C.9. Let B(τ ) = +∞
i=d+1 (−1) (2i+1)! , if d & τ +` log(s1 )+log(1/) then |B(τ )| ≤ (1/s1 )
77
Proof. We first show, for any i ≥ d + 1,
τ 2i
(2i + 1)!
τ 2i
≤
e((2i + 1)/e)2i+1
≤ 2−2i
by e(n/e)n ≤ n!
by i & τ
O(`)
≤ (1/s1 )
by i & ` log(s1 ) + log(1/)
Second, we can show that
+∞
X
(−1)
i=d+1
τ 2i
τ 2(d+1)
.
≤ (1/s1 )O(`)
(2i + 1)!
(2(d + 1) + 1)!
Thus, we complete the proof.
Claim C.10. mint∈[−1/2,1/2] |H(t)| ≥ (s1 )−Ω(`) .
Proof. By the property of H(t),
min
1
s <|t|≤ 12
2 3
H(t) = min H(t)
|t|≤ 12
Thus, it suffices to prove a lower bound on H(t) for any t such that 21 s3 < |t| ≤ 12 . Because of
symmetric property, we only need to prove a lower bound for one side. Let’s consider t ∈ [ 12 s3 , 1/2],
s0
H(t) ≥
πs1
≥
s0
πs1
Z
(t+
s2
)πs1
2
s
(t− 22 )πs1
s
(t+ 22 )πs1
Z
(t+
s2
)πs1
4
(
sin(τ ) `
) dτ
τ
(
sin(τ ) `
) dτ
τ
s0
1
s2
s2
· Θ((t + )s1 ) · · π · Θ((t + )πs1 )−`
πs1
2
2
2
−Ω(`)
≥ (s1 )
≥
C.2
b ))
Analysis of filter function (G(t), G(f
b )) in a similar way of (H1 (t), H
c1 (f )) by switching the time domain and the
We construct (G(t), G(f
c1 (f )) and modify the parameters for the permutation hashing Pσ,a,b .
frequency domain of (H1 (t), H
b ) by doing the following
Definition C.11. Given B > 1, δ > 0, α > 0, we construct G(t), G(f
operations,
• s2 =
π
2B ,
• s1 =
B
απ ,
• ` = l = Θ(log(k/δ)).
78
poly log(1/δ)
log2 (1/δ)
log1.5 (1/δ)
∆h
`
log1 (1/δ)
log0.5 (1/δ)
s3
1
1 − 1/k 4
s1
1 − 1/k 5
k
k4
k5
k6
Figure 8: Parameters for s1 , s3 and `.
b ) becomes
Then G(t), G(f
G(t) = b0 · (rects1 (t))∗l · sinc(ts2 )
π
),
2B
b ) = b0 · (sinc(s1 f ))·l ∗ rects (f )
G(f
2
B
·l
π (f ).
= b0 · (sinc( f )) ∗ rect 2B
απ
√
√
b
where the scalar b0 = Θ(s1 l) = Θ(B l/α) satisfying G(0)
= 1.
= b0 · (rect
B
(απ)
(t))∗l · sinc(t
b ))[B, δ, α, l]
Lemma 6.7. Given B > 1, δ > 0, α > 0, we set l = Ω(log(δ/k)). The filter function (G(t), G(f
satisfies the following properties,
b ) ∈ [1 − δ/k, 1], if |f | ≤ (1 − α) 2π .
G(f
2B
2π
2π
b ) ∈ [0, 1], if (1 − α)
Property II : G(f
≤ |f | ≤
.
2B
2B
b ) ∈ [−δ/k, δ/k], if |f | > 2π .
Property III : G(f
2B
l −B l B
Property IV : supp(G(t)) ⊂ [ ·
, ·
].
2 πα 2 πα
Property V : max|G(t)| . poly(B, l).
Property I :
t
Proof. The first five Properties follows from Lemma 6.6 directly.
C.3
Parameters setting for filters
b )).
One-cluster Recovery. In one-cluster, we donot need filter function (G(t), G(f
b )), we
In section C.1, by Equation (34) in the proof of Property VI of filter function (H(t), H(f
1
e 4 ).
need min( 1−s3 , s1 ) ≥ O(k
79
b )), we
In section C.1, by Equation (32) in the proof of Property V of filter function (H(t), H(f
set ` & k log(k/δ).
b )) in Equation (35): ∆h h s1 ` in section
∆h is determined by the parameters of filter (H(t), H(f
s3 T
e 5 log(1/δ))/T .
C.1. Combining the setting of s1 , s3 `, we should set ∆h ≥ O(k
b )).
k-cluster Recovery. Note that in the k-cluster recovery, we need to use filter function (G(t), G(f
We choose l = log(k/δ), α h 1, B h k , and D = l/α.
By proof of Property II of z in Lemma 7.20 from section 7.6, we need T (1 − s3 ) > σBl. By the
same reason in one-cluster recovery, 1 − s3 ≤ e 1 4 . Combining T (1 − s3 ) > σBl and 1 − s3 ≤ e 1 4 ,
O(k )
O(k )
we obtain
T
> σBl
(36)
e
O(k 4 )
1
1
Because in our algorithm, we will sample σ from [ B∆
, 2 ]. Thus, plugging σ = Θ( B∆
) in
h B∆h
h
Equation (36) we have
T
l
>
4
e )
∆h
O(k
which implies another lower bound for ∆h ,
e 4 )l/T
∆h ≥ O(k
Combining the above bound with previous lower bound in one-cluster recovery, we get
e 4 log(1/δ))/T + O(k
e 5 log(1/δ))/T = O(k
e 5 log(1/δ))/T
∆h ≥ O(k
e 4 ) and ` h O(k log(k/δ)).
For s1 and `, we still choose the same setting as before, s1 h O(k
C.4
Analysis of HashToBins
In this section, we explain the correctness of Procedure HashToBins in Algorithm 6. Before giving
the proof of that algorithm, we show how to connect CFT, DTFT and DFT.
Lemma C.12. For any signal W : R → C, let A : Z → C and B : [n] → C be defined as follows:
X
A[i] = W (i), ∀i ∈ Z and B[i] =
A[i + jn], ∀i ∈ [n].
j∈Z
Then we consider the Fourier transform on W, A, and B:
CFT
DTFT
DFT
We have:
b )=
∀f ∈ [0, 1), A(f
X
j∈Z
c : R → C,
W
b : [0, 1] → C,
A
b : [n] → C.
B
c (f + j);
W
80
b =
∀i ∈ [n], B[i]
X
j∈Z
c (i/n + j).
W
Proof. Recall that Combs (t) =
Pc
W (f + j):
j∈Z
b )=
A(f
=
X
P
j∈Z δjs (t).
2πijf A[j] equals to
b ) = P
First, we show A(f
j∈Z e
e2πijf W [j]
j∈Z
Z +∞
−∞
by A[j] = W (j)
e2πif j W (j) · Comb1 (j)dj
= W\
· Comb1 (f )
c ∗ Comb
\ 1 )(f )
= (W
X
c (f + j).
W
=
(37)
j∈Z
b = A(i/n),
b
Next, we prove that ∀i ∈ [n], B[i]
b =
B[i]
=
=
n
X
B[j]e
2πi
ij
n
by DFT
j=1
n X
X
2πi
(
A[j + kn])e n ij
j=1 k∈Z
n X
X
A[j + kn]e
by B[j] =
X
A[j + kn]
k∈Z
2πi
i(j+kn)
n
by e
2πi
·ikn
n
=1
j=1 k∈Z
=
X
j∈Z
i
b
A[j]e2πij n = A(i/n)
by DTFT.
(38)
P
c (j/n + i) for
b = A(j/n)
b
Combining Equation (38) and Equation (37), we obtain that B[j]
= i∈Z W
all j ∈ [n].
P
V [j + (i − 1)B]. Then
Claim C.13. Let u ∈ CB and V ∈ CBD such that for any j ∈ B, u[j] =
i∈[D]
u
b[j] = Vb [jD], ∀j ∈ [B].
Proof. We prove it through the definition of the Fourier transform:
Vb [jD] =
=
BD
X
2πi
V [i] · e BD ·i·(jD)
i=1
B X
D
X
V [i + kB]e
by definition of DFT
2πi
·(i+kB)·j
B
by replacing i by i + kB
i=1 k=1
=
=
B
X
i=1
B
X
i=1
e
2πi
·j·i
B
D
X
k=1
e
2πi
·j·i
B
V [i + (k − 1)B]
u[i] = u
b[j]
by e2πijk = 1
by definition of DFT on u
81
b
A−2,0
(0)
G
(f )
σ,b
1/(Bσ)
b
A0,0
A1,0
A2,0
B−1,0
B0,0
B1,0
B2,0
1/σ
B−2,0
(1)
G
(f )
σ,b
A−1,0
A−2,1
1/(Bσ)
A−1,1
A0,1
A1,1
A2,1
B−1,1
B0,1
B1,1
B2,1
1/σ
B−2,1
b (j) (f ) where the top one is j = 0 and the bottom one is j = 1, Ai,j = [ 1 (2π(i +
Figure 9: G
σ,b
σ
2π
1
2B ), σ (2π(i
+
j
B)
+
2π
2B )],
Bi,j = [ σ1 (2π(i +
j
B)
−
2π(1−α) 1
), σ (2π(i
2B
+
j
B)
+
2π(1−α)
)]
2B
j
B)
−
We use Definition 6.1 and Lemma 6.2 to generalize Lemma C.12,
P
Corollary C.14. If for all j ∈ [n], B[j] =
W (j + in)σ − σa , then ∀j ∈ [n],
i∈Z
X j
j
1
b
c
B[j] =
W ( + i)/σ · e−2πi( n +i)a .
n
σ
i∈Z
If for all j ∈ [n], B[j] =
P
i∈Z
W (j + in)σ − σa e−2πiσb(j+in) , then ∀j ∈ [n],
b =
B[j]
X
i∈Z
j
1
j
c
W ( + i)/σ + b · e−2πi( n +i)a−2πiσab .
n
σ
Remark C.15 (Samples of HashToBins). Procedure HashToBins in Algorithm 6 takes BD
samples in x(t):
x(σ(1 − a)), x(σ(2 − a)), · · · , x(σ(BD − a)).
b )) and Combs (t) = P δsj (t) to define
To analyze our algorithm, we use filter function (G(t), G(f
j∈Z
the discretization of G.
82
b ),
Definition C.16. Define the discretization of G(t) and G(f
Gdis (t) = G(t) · Combs (t)
b ∗ Comb1/s )(f )
b dis (f ) = 1 (G
G
s
b ∗ Comb1 )(f )
= (G
=
π
−B
where | supp(G(t))| =
lB
πα ,
D=
l
πα ,
− (1−α)π
B
(1−α)π
B
π
B
∗
−2
−1
0
1
2
s = | supp(G(t))|/(BD) = l/(παD) = 1.
Definition 6.8. ∀σ > 0, b and j ∈ [B]. Define,
1
G(t/σ)e2πit(j/B−σb)/σ
σ
X
b (j) (f ) = G
b dis ( j − σf − σb) =
b + j − σf − σb)
G
G(i
σ,b
B
B
(j)
Gσ,b (t) =
i∈Z
Lemma 6.9. Let u ∈ CB be the result of HashToBins under permutation Pσ,a,b , and let j ∈ [B].
Define
b (j) ,
zb = x[
·H ·G
σ,b
so
(j)
z = (x · H) ∗ Gσ,b .
Let vector u
b ∈ CB denote the B-dimensional DFT of u, then ∀j ∈ [B],
u
b[j] = zσa .
Proof. Recall B is the number of hash bins. B · D is the number of samples in time signal. Let
W (t) = x · H(t), define vector y ∈ CBD , then ∀j ∈ [BD], define
y[j] = W (σ(j − a))e2πiσbj
Recall G(t) denote the rect∗l
B/α (t) · sinc(t/B), then | supp(G(t))| =
the discretization of G(t), where G0 [i] = G(i). Then, ∀j ∈ [B],
X
u[j] =
V [j + iB]
lB
α.
Let vector G0 ∈ CBD is
i∈[D]
where V [j] = y[j] · G0 [j] and G0 [j] is the value at the jth nonzero point of Gdis (t). Applying Claim
C.13 with the definition of u[j] and V [j + iB], gives u
b[j] = Vb [jD], ∀j ∈ [B].
Because of u is the result of HashToBins(x · H, Pσ,a,b , G) and | supp(G(t))| = BD(choosing
D = l/α), then
X
u[j] =
W (σ(j + iB − a))e−2πiσb(j+iB) G(j + iB)
i∈Z
Then we define G00 (t) = G(t/σ + a)e−2πibσ(t/σ+a) and Y (t) = W (t) · G00 (t), then immediately,
we have
b 00 (f ) = σ G(σ(f
b
c (f ) ∗ G
b 00 (f )
G
− b))e2πiaσf and Yb (f ) = W
83
Thus, we can rewrite u[j] in the following sense,
u[j]
X
=
i∈Z
X
=
i∈Z
X
=
i∈Z
W (σ(j + iB − a))e−2πiσb(j+iB) G(j + iB)
W (σ(j + iB − a))G00 (σ(j + iB − a))
by G00 (t) = G(t/σ + a)e−2πibσ(t/σ+a)
by Y (t) = W (t) · G00 (t)
Y (σ(j + iB − a))
Then
=
=
=
u
b[j]
X
j
j
1
Yb (( + i)/σ) · · e−2πi( B +i)a
B
σ
i∈Z
X Z +∞
c (s) · G
b 00 ( j/B + i − s) · 1 · e−2πi·(j/B+i)a ds
W
σ
σ
i∈Z −∞
Z
X +∞
c (s) · G(j/B
b
W
+ i − σs − σb) · e−2πi·(−σs)a ds
i∈Z −∞
+∞
=
Z
−∞
=
Z
+∞
−∞
c (s) ·
W
X
i∈Z
b
G((j/B
+ i) − σs − σb)e2πiaσs ds
by Corollary C.14
b 00 (f ) = σ G(σ(f
b
by G
− b))e2πiaσf
c (s) · G
b dis ( j − σs − σb)e−2πiaσs ds
W
B
By definition C.16,
b (j) = G
b dis ( j − σs − σb) =
G
σ,b
B
By definition of zb, we have
X
i∈Z
c (f ) ∗ G
b 00 (f )
by Yb (f ) = W
b dis (f ) =
by G
X
i∈Z
b + i)
G(f
b + j − σs − σb)
G(i
B
b (j) (s) = W
c (s) · G
b (j) (s)
zb(s) = x[
· H(s) · G
Then u
b[j] is the (aσ)th inverse Fourier coefficients of zb, basically,
u
b[j] = zaσ = z(aσ)
Thus, we can conclude first computing vector u ∈ CB . Getting vector u
b ∈ CB by using the Discrete
Fourier transform u
b = DFT(u). This procedure allows us to sample from time domain to implicitly
access the time signal’s Fourier transform zb. If z is one-cluster in frequency domain, then apply
one-cluster recovery algorithm.
D
Acknowledgments
The authors would to like thank Aaron Sidford and David Woodruff for useful discussions.
E
Algorithm
This section lists the pseudocode of our algorithms.
84
Algorithm 3
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
procedure GetEmpirical1Energy(z, T, ∆) — Claim 7.11
Rest ← (T ∆)2
for i = 1 → Rest do
Choose αi ∈ [0, T ] uniformly at random
zemp ← zemp + |z(αi )|2
end forp
zemp ← zemp /Rest
return zemp
end procedure
procedure GetLegal1Sample(z, ∆, T, β, zemp ) — Lemma 7.2
Rrepeat ← (T ∆)3 , Sheavy ← ∅
for i = 1 → Rrepeat do
Choose αi ∈ [0, T ] uniformly at random
if |z(αi )| ≥ 0.5 · zemp then
Sheavy ← Sheavy ∪ i
end if
end for
for i ∈ Sheavy do
w(i) ← |z(αi )|2 + |z(αi + β)|2
end for
P
α ← αi with probability w(i)/ j∈Sheavy w(j) for i ∈ Sheavy
return α
end procedure
85
Algorithm 4
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
procedure Locate1Signal(z, T, F, ∆, zemp ) — Lemma 7.15
Set t h log(F T ), t0 = t/4, Dmax h logt0 (F T ), Rloc h log1/c (tc), L(1) = 2F
for i ∈ [Dmax ] do
ts
l h 2F/(t0 )i−1 ∆, s h c, βb = 2∆l
if βb & T /(T ∆)3/2 then
break
else
b zemp , L(i−1) )
L(i) ← Locate1Inner(z, ∆, T, β,
end if
end for
return L(i)
end procedure
b zemp , L)
e
procedure Locate1Inner(z, ∆, T, β,
Let vq ← 0 for q ∈ [t]
while r = 1 → Rloc do
b β]
b uniformly at random
Choose β ∈ [ 12 β,
γ ← GetLegal1Sample(z, ∆, T, β, zemp )
for i ∈ [m] do
e − ∆l/2), β(L
e + ∆l/2)] ∩ Z+ , θi = 1 (φ(x(γ)/x(γ + β)) + 2πsi )
si ∈ [β(L
2πσβ
Let θi belong to region(q)
Then add a vote to region(q) and its two neighbors, i.e., region(q −1) and region(q +1)
end for
end while
qj∗ ← {q|vq > R2loc }
return L ← center of region(qj∗ )
end procedure
procedure FrequencyRecovery1Cluster(z, T, F, ∆) — Theorem 7.5
zemp ← GetEmpirical1Energy(z, T, ∆)
for r = 1 → O(k) do
Lr ← Locate1Signal(z, T, F, ∆, zemp )
end for
return L∗ ← median Lr
r∈[O(k)]
33:
end procedure
86
Algorithm 5 Main algorithm for one-cluster recovery
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
procedure CFT1Culster(x, H, T, F ) — Theorem 8.1
fe0 ← FrequencyRecovery1Cluster(x, H, T, F )
x
e ← SignalRecovery1Cluster(fe0 , poly(k)∆h )
return x
e
end procedure
procedure GenerateIntervals(d)
n ← y0 ← i ← 0, m ← Θ(d)
while yi ≤ 1 − m92 do
√ 2
1−y
yi+1 ← yi + m i , In+1 ← [yi , yi+1 ], In+2 ← [−yi+1 , −yi ]
i ← i + 1, n ← n + 2
end while
In+1 ← [yi , 1], In+2 ← [−yi , −1], n ← n + 2
return n, I
end procedure
procedure RobustPolynomialLearning(x, d, T ) — Theorem 1.2
(n, I) ← GenerateIntervals(d)
for j = 1 → n do
wj ← |Ij |/2
Choose tj from Ij uniformly at random
t +1
zj ← x(T · j 2 )
end for
ej,i ← ti , for each (j, i) ∈ [n] × {0, 1, · · · , d}
A
j
e eb = z, w)
α ← LinearRegressionW(
A,
Pd
Q(t) ← i=0 αi ti
e = Q(T · t+1 )
return Q(t)
2
end procedure
procedure RobustPolynomialLearning+ (x, d, T ) — Theorem 4.5 — a.k.a. SignalRecovery1Cluster
R ← Θ(d)
(n, I) ← GenerateIntervals(d)
wj ← |Ij |/2, for each j ∈ [n]
for i = 1 → R do
Qi ← RobustPolynomialLearning(x, d, T )
end for
Choose tj from Ij uniformly at random, for each j ∈ [n]
for i = 1 → R do
Qi (t1 ), Qi (t2 ), · · · , Qi (tn ) ← MultipointEvaluation(Qi , {t1 , t2 , · · · , tn })
end for
e j ← median Qi (tj ), for each j ∈ [n]
Q
i∈[R]
ej,i ← ti , for each (j, i) ∈ [n] × {0, 1, · · · , d}
A
j
e eb = Q,
e w)
40:
α ← LinearRegressionW(
A,
Pd
i
41:
return Q(t) ← i=0 αi t
42: end procedure
39:
87
Algorithm 6
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
procedure LocateKSignal(x, H, G, T, ∆, σ, b, zemp ) — Clain 7.27
Set t h log(F T ), t0 = t/4, Dmax h logt0 (F T ), Rloc h log1/c (tc), L(1) = 2F
for i ∈ [Dmax ] do
ts
∆l h 2F/(t0 )i−1 , s h c, βb = 2σ∆l
if σ βb & T /(T ∆)3/2 then
break
else
b U, L(i−1) )
L(i) ← LocateKInner(x, H, G, T, ∆, σ, b, zemp β,
end if
end for
return L(i)
end procedure
b U, L)
e
procedure LocateKInner(x, H, G, T, ∆, σ, b, zemp β,
Let vj,q ← 0 for (j, q) ∈ [B] × [t]
for r = 1 → Rloc do
b β]
b uniformly at random
Choose β ∈ [ 12 β,
0
u
b, u
b ← GetLegalKSample(x, H, G, T, ∆, σ, β, zemp )
for j ∈ [B] do
for i ∈ [m] do
1
e j − ∆l/2), σβ(L
e j + ∆l/2)] ∩ Z+
θj,i = 2πσβ
(φ(b
u[j]/ub0 [j]) + 2πsi ), si ∈ [σβ(L
fj,i = θj,i + b (mod F )
suppose fj,i belongs to region(j, q),
add a vote to both region(j, q) and two neighbors nearby that region, e.g.
region(j, q − 1) and region(j, q + 1)
end for
end for
end for
for j ∈ [B] do
qj∗ ← {q|vj,q > R2loc }
Lj ← center of region(j, qj∗ )
end for
return L
end procedure
procedure HashToBins(x,
H, G, Pσ,a,b ) — Lemma 6.9
P
Compute u[j] = i∈D v[j + iB]
u
b ← FFT(u)
return u
b
end procedure
88
Algorithm 7
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
procedure GetEmpiricalKEnergy(x, H, G, T, ∆, σ, b) — Claim 7.26
Rest ← (T ∆)2
for i = 1 → Rest do
Choose α ∈ [0, T ] uniformly at random
u
b ← HashToBins(x, H, G, Pσ,α,b )
for j = 1 → B do
j
j
zemp
← zemp
+ |b
uj |2
end for
end for
for j = 1 →qB do
j
j
zemp
← zemp
/Rest
end for
return zemp .
end procedure
procedure GetLegalKSample(x, H, G, T, ∆, β, zemp ) — Lemma 7.25
Rrepeat ← (T ∆)3 .
j
← ∅, ∀j ∈ [B]
Sheavy
for i = 1 → Rrepeat do
Choose α ∈ [0, T ] uniformly at random
u
bi ← HashToBins(x, H, G, Pσ,α,b )
0
u
b i ← HashToBins(x, H, G, Pσ,α+β,b )
for j = 1 → B do
j
if |b
uij | ≥ 0.5 · zemp
then
j
Sheavy,j ← Sheavy ∪ i
end if
end for
end for
for j = 1 → B do
j
for i ∈ Sheavy
do
0
i
2
w(i) ← |b
uj | + |b
uji |2
end for
P
0
j
(b
vj , vbj0 ) ← (b
uij , u
bji ) with probability w(i)/ i0 ∈S j
w(i0 ) for i ∈ Sheavy
heavy
33:
34:
35:
36:
37:
38:
39:
end for
return vb, vb0 ∈ CB
end procedure
procedure OneStage(x, H, G, σ, b) — Lemma 7.22
zemp ← GetEmpiricalKEnergy(x, H, G, T, ∆, σ, b)
L ← LocateKSignal(x, H, G, T, ∆, σ, b, zemp )
end procedure
89
Algorithm 8 Main algorithm for k-cluster recovery
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
procedure CFTKCluster(x, H, G, T, F )
{fe1 , · · · , fel } ← FrequencyRecoveryKCluster(x, H, G, T, F )
x
e ← SignalRecoveryKCluster+ (fe1 , · · · , fel , ∆ = poly(k, log(1/δ))/T, T )
return x
e as our hypothesis
end procedure
procedure FrequencyRecoveryKCluster(x, H, G) — Theorem 2.6
for r ∈ [R] do
1
Choose σ ∈ [ B∆
, 2 ] uniformly at random
h B∆h
h c]
Choose b ∈ [0, 2πbF/∆
] uniformly at random
(σB)
Lr ← OneStage(x, H, G, σ, b)
end for
L∗ ← MergedStages(L1 , L2 , · · · , LR )
end procedure
procedure SignalRecoveryKCluster(fe1 , · · · , fel , ∆, T )
d ← 5π((∆T )1.5 + k 3 log k + k log 1/δ)
m ← O((kd)C3 · logC3 d) for a constant C3 = 5
for j = 1 → m do
Sample tj from [0, T ] uniformly at random
ej,i ·l+i ← ti1 · e2πifei2 tj for each (i1 , i2 ) ∈ {0, · · · , d} × [l]
A
1
2
j
ebj ← x(tj )
end for
e eb)
α ← LinearRegression(A,
d P
l
P
e
return x
e(t) ←
αi1 ·l+i2 ti1 · e2πifi2 t
i1 =0i2 =1
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
end procedure
procedure SignalRecoveryKCluster+ (fe1 , · · · , fel , ∆, T ) — Theorem 9.1
R ← Θ(k)
d ← 5π((∆T )1.5 + k 3 log k + k log 1/δ)
m ← O((kd)C3 · logC3 d) for a constant C3 = 5
for i = 1 → R do
x
ei (t) ← SignalRecoveryKCluster(fe1 , · · · , fel , ∆, T )
end for
for j = 1 → m do
Sample tj from [0, T ] uniformly at random
ej,i ·l+i ← ti1 · e2πifei2 tj for each (i1 , i2 ) ∈ {0, · · · , d} × [l]
A
1
2
j
ebj ← median x
ei (tj )
i∈[R]
36:
37:
38:
end for
e eb)
α ← LinearRegression(A,
d
l
P P
e
return x
e(t) ←
αi1 ·l+i2 ti1 · e2πifi2 t
i1 =0i2 =1
39:
end procedure
90
| 8cs.DS
|
Learning Aided Optimization for Energy Harvesting
Devices with Outdated State Information
arXiv:1801.03572v1 [math.OC] 10 Jan 2018
Hao Yu and Michael J. Neely
University of Southern California
Abstract—This paper considers utility optimal power control
for energy harvesting wireless devices with a finite capacity
battery. The distribution information of the underlying wireless
environment and harvestable energy is unknown and only outdated system state information is known at the device controller.
This scenario shares similarity with Lyapunov opportunistic
optimization and online learning but is different from both.
By a novel combination of Zinkevich’s online gradient learning
technique and the drift-plus-penalty technique from Lyapunov
opportunistic optimization, this paper proposes a learning-aided
algorithm that achieves utility within O() of the optimal, for
any desired > 0, by using a battery with an O(1/) capacity.
The proposed algorithm has low complexity and makes power
investment decisions based on system history, without requiring
knowledge of the system state or its probability distribution.
I. I NTRODUCTION
Energy harvesting can enable self-sustainable and perpetual
wireless devices. By harvesting energy from the environment
and storing it in a battery for future use, we can significantly
improve energy efficiency and device lifetime. Harvested
energy can come from solar, wind, vibrational, thermal, or
even radio sources [1], [2], [3]. Energy harvesting has been
identified as a key technology for wireless sensor networks [4],
internet of things (IoT) [5], and 5G communication networks
[6]. However, the development of harvesting algorithms is
complex because the harvested energy is highly dynamic and
the device environment and energy needs are also dynamic.
Efficient algorithms should learn when to take energy from
the battery to power device tasks that bring high utility, and
when to save energy for future use.
There have been large amounts of work developing efficient
power control policies to maximize the utility of energy
harvesting devices. In the highly ideal case where the future system state (both the wireless channel sate and energy
harvesting state) can be perfectly predicted, optimal power
control strategies that maximize the throughput of wireless
systems are considered in [7], [8]. In a more realistic case with
only the statistics and causal knowledge of the system state,
power control policies based on Markov Decision Processes
(MDP) are considered in [9], [10]. In the case when the
statistical knowledge is unavailable but the current system state
is observable, work [11] develops suboptimal power control
policies based on approximation algorithms.
Hao Yu and Michael J. Neely are with the Department of Electrical
Engineering, University of Southern California, Los Angeles, USA. This work
is supported in part by grant NSF CCF-1718477.
However, there is little work on the challenging scenario
where neither the distribution information nor the system state
information are known. In practice, the amount of harvested
energy on each slot is known to us only after it arrives and
is stored into the battery. Further, the wireless environment
is often unknown before the power action is chosen. For
example, the wireless channel state in a communication link
is measured at the receiver side and then reported back to
the transmitter with a time delay. If the fading channel varies
very fast, the channel state feedback received at the transmitter
can be outdated. Another example is power control for sensor
nodes that detect unknown targets where the state of targets
is known only after the sensing action is performed.
In this paper, we consider utility-optimal power control
in an energy harvesting wireless device with outdated state
information and unknown state distribution information. This
problem setup is closely related to but different from the
Lyapunov opportunistic power control considered in works
[12], [13], [14] with instantaneous wireless channel state
information. The policies developed in [12], [13], [14] are
allowed to adapt their power actions to the instantaneous
system states on each slot, which are unavailable in our
problem setup. The problem setup in this paper is also closely
related to online convex optimization where control actions
are performed without knowing instantaneous system states
[15], [16], [17]. However, existing methods for online convex
learning require the control actions to be chosen from a fixed
set. This does not hold in our problem since the power to be
used can only be drained from the battery whose backlog is
time-varying and dependent on previous actions.
By combining the drift-plus-penalty (DPP) technique for
Lyapunov opportunistic optimization [18] and the online gradient learning technique for online convex optimization [15],
we develop a novel learning aided dynamic power control
algorithm that can achieve an O() optimal utility by using a
battery with an O(1/) capacity for energy harvesting wireless
devices with outdated state information.
II. P ROBLEM F ORMULATION
Consider an energy harvesting wireless device that operates in normalized time slots t ∈ {1, 2, . . .}. Let ω[t] =
(e[t], s[t]) ∈ Ω represent the system state on each slot t, where
• e[t] is the amount of harvested energy for slot t (for
example, through solar, wind, radio signal, and so on).
• s[t] is the wireless device state on slot t (such as the
vector of channel conditions over multiple subbands).
•
Ω is the state space for all ω[t] = (e[t], s[t]) states.
B. Basic assumption
{ω[t]}∞
t=1
Assume
evolves in an independent and identically
distributed (i.i.d.) manner according to an unknown distribution. Further, the state ω[t] is unknown to the device until the
end of slot t. The device is powered by a finite-size battery.
At the beginning of each slot t ∈ {1, 2, . . .}, the device draws
energy from the battery and allocates it as an n-dimensional
power decision vector p[t] = [p1 [t], . . . , pn [t]]T ∈ P where P
is a compact convex set given by
P = {p ∈ Rn :
n
X
pi ≤ pmax , pi ≥ 0, ∀i ∈ {1, 2, . . . , n}}.
i=1
max
Note that p
is a given positive constant (restricted by
hardware) and represents the maximum total power that can be
used on each slot. The device receives a corresponding utility
U (p[t]; ω[t]). Since p[t] is chosen without knowledge of ω[t],
the achieved utility is unknown until the end of slot t. For
each ω ∈ Ω, the utility function U (p; ω) is assumed to be
continuous and concave over p ∈ P. An example is:
U (p; ω) =
n
X
log(1 + pi [t]si [t])
(1)
i=1
where s[t] = (s1 [t], . . . , sn [t]) is the vector of (unknown)
channel conditions over n orthogonal subbands available to
the wireless device. In this example, pi [t] represents the
amount of power invested over subband i in a rateless coding
transmission scenario, and U (p[t]; ω[t]) is the total throughput
achieved on slot t. We focus on fast time-varying wireless
channels, e.g., communication scenarios with high mobility
transceivers, where s[t] known at the transmitter is outdated
since s[t] must be measured at the receiver side and then
reported back to the transmitter with a time delay.
A. Further examples
The above formulation admits a variety of other useful
application scenarios. For example, it can be used to treat
power control in cognitive radio systems. Suppose an energy
limited secondary user harvests energy and operates over
licensed spectrum occupied by primary users. In this case,
s[t] = (s1 [t], . . . , sn [t]) represents the channel activity of
primary users over each subband. Since primary users are not
controlled by the secondary user, s[t] is only known to the
secondary user at the end of slot t.
Another application is a wireless sensor system. Consider
an energy harvesting sensor node that collects information by
detecting an unpredictable target. In this case, s[t] can be the
state or action of the target on slot t. By using p[t] power for
signaling and sensing, we receive utility U (p[t]; ω[t]), which
depends on state ω[t]. For example, in a monitoring system,
if the monitored target performs an action s[t] that we are
not interested in, then the reward U (p[t]; ω[t]) by using p[t]
is small. Note that s[t] is typically unknown to us at the
beginning of slot t and is only disclosed to us at the end
of slot t.
Assumption 1.
•
•
There exist a constant emax > 0 such that 0 ≤ e[t] ≤
emax , ∀t ∈ {1, 2, . . .}.
Let ∇p U (p; ω) denote a subgradient (or gradient if
U (p; ω) is differentiable) vector of U (p; ω) with respect
∂
U (p; ω), ∀i ∈ {1, 2, . . . , n} denote each
to p and let ∂p
i
component of vector ∇p U (p; ω). There exist positive
∂
constants D1 , . . . , Dn such that | ∂p
U (p; ω)| ≤ Di , ∀i ∈
i
{1, 2, . . . , n} for all ω ∈ Ω and all p p
∈P
P. This further
n
2
implies there exists D > 0, e.g., D =
i=1 Di , such
that k∇p U (p;p
ω)k
Pn≤ D2 for all ω ∈ Ω and all p ∈ P,
where kxk =
i=1 xi is the standard l2 norm.
Such constants D1 , . . . , Dn exist in most cases of interest,
such as for utility functions (1) with bounded si [t] values. 1
The following fact follows directly from Assumption 1.
Fact 1. For each ω ∈ Ω, U (p, ω) is D-Lipschitz over p ∈ P,
i.e.,
|U (p1 , ω) − U (p2 , ω)| ≤ Dkp1 − p2 k, ∀p1 , p2 ∈ P
Proof. By the basic subgradient inequality for concave functions:
U (p2 , ω) + (∇p U (p2 ; ω))T (p1 − p2 ) ≥ U (p1 , ω)
U (p1 , ω) + (∇p U (p1 ; ω))T (p2 − p1 ) ≥ U (p2 , ω)
Rearranging terms and applying the Cauchy-Schwarz inequality yields
U (p1 , ω) − U (p2 , ω) ≤ k∇p U (p2 ; ω)kkp1 − p2 k
U (p2 , ω) − U (p1 , ω) ≤ k∇p U (p1 ; ω)kkp1 − p2 k
Combining the above inequalities and recalling that all subgradients are bounded by D gives |U (p1 , ω) − U (p2 , ω)| ≤
Dkp1 − p2 k.
C. Power control and energy queue model
The finite size battery can be considered as backlog in an
energy queue. Let E[0] be the initial energy backlog in the
battery and E[t] be the energy stored in the battery at the end
of slot t. The power vector p[t] must satisfy the following
energy availability constraint:
Pn
(2)
i=1 pi [t] ≤ E[t − 1], ∀t ∈ {1, 2, . . .}.
which requires the consumed power to be no more than what
is available in the battery.
Let E max be the maximum capacity of the battery. If the
energy availability constraint (2) is satisfied on each slot, the
energy queue backlog E[t] evolves as follows:
Pn
E[t] = min{E[t − 1] − i=1 pi [t] + e[t], E max }, ∀t. (3)
1 This
is always true when si [t] are wireless signal strength attenuations.
D. An upper bound problem
Let ω[t] = (e[t], s[t]) be the random state vector on slot t.
Let E [e] = E [e[t]] denote the expected amount of new energy
that arrives in one slot. Define a function h : P → R by
h(p) = E [U (p; ω[t])] .
Since U (p; ω) is concave in p for all ω by Assumption 1 and
is D-Lipschitz over p ∈ P for all ω by Fact 1, we know h(p)
is concave and continuous.
The function h is typically unknown because the distribution
of ω is unknown. However, to establish a fundamental bound,
suppose both h and E[e] are known and consider choosing a
fixed vector p to solve the following deterministic problem:
max h(p)
(4)
p
s.t.
n
X
pi − E[e] ≤ 0
(5)
i=1
p∈P
(6)
where constraint (5) requires that the consumed energy is no
more than E[e].
Let p∗ be an optimal solution of problem (4)-(6) and U ∗ be
its corresponding utility value of (4). Define a causal policy
as one that, on each slot t, selects p[t] ∈ P based only on
information up to the start of slot t (in particular, without
knowledge of ω[t]). Since ω[t] is i.i.d. over slots, any causal
policy must have p[t] and ω[t] independent for all t. The
next lemma shows that no causal policy p[t], t ∈ {1, 2, . . .}
satisfying (2)-(3) can attain a better utility than U ∗ .
Lemma 1. Let p[t] ∈ P, t ∈ {1, 2, . . .} be yielded by any
causal policy that consumes lessPenergy P
than it harvests in
T
n
the long term, so lim supT →∞ T1 t=1 E [ i=1 pi [t]] ≤ E [e].
Then,
T
1X
E[U (p[t]; ω[t])] ≤ U ∗ .
lim sup
T
T →∞
t=1
Further, since p[t] ∈ P for all slots t, it holds that p̄[T ] ∈ P
for all T > 0. Also,
T
T
1X
(a) 1 X
E[U (p[t]; ω[t])] =
E [h(p[t])]
T t=1
T t=1
i
(b) h P
T
≤ h E T1 t=1 p[t]
= h(p̄[T ])
where (a) holds by (7); (b) holds by Jensen’s inequality for
the concave function h. It follows that:
lim sup
T →∞
T
1X
E[U (p[t]; ω[t])] ≤ lim sup h(p̄[T ]).
T t=1
T →∞
Define θ = lim supT →∞ h(p̄[T ]). It suffices to show that θ ≤
U ∗ . Since p̄[T ] is in the compact set P for all T > 0, the
Bolzano-Wierstrass theorem ensures there is a subsequence of
times Tk such that p̄[Tk ] converges to a fixed vector p0 ∈ P
and h(p̄[Tk ]) converges to θ as k → ∞:
lim p̄[Tk ] = p0 ∈ P
k→∞
lim h(p̄[Tk ]) = θ
k→∞
Continuity of h implies that h(p0 )P= θ. By (8) the vector
n
p0 = [p0,1 , . . . , p0,n ]T must satisfy i=1 p0,i ≤ E [e]. Hence,
p0 is a vector that satisfies constraints (5)-(6) and achieves
utility h(p0 ) = θ. Since U ∗ is defined as the optimal utility
value to problem (4)-(6), it holds that θ ≤ U ∗ .
Note that the U ∗ utility upper bound of Lemma 1 holds for
any policy that consumes no more energy than it harvests in the
long term. Policies that satisfy the physical battery constraints
(2)-(3) certainly consume no more energy than harvested in
the long term. However, Lemma 1 even holds for policies that
violate these physical battery constraints. For example, U ∗ is
still a valid bound for a policy that is allowed to “borrow”
energy from an external power source when its battery is
empty and “return” energy when its battery is full.
Proof. Fix a slot t ∈ {1, 2, . . .}. Then
(a)
III. N EW A LGORITHM
(b)
E [U (p[t]; ω[t])] = E [E [U (p[t]; ω[t])|p[t]]] = E [h(p[t])]
(7)
where (a) holds by iterated expectations; (b) holds because
p[t] and ω[t] are independent (by causality).
For each T > 0 define p̄[T ] = [p̄1 [T ], . . . , p̄n [T ]]T with
This subsection proposes a new learning aided dynamic
power control algorithm that chooses power control actions
based on system history, without requiring the current system
state or its probability distribution.
A. New Algorithm
T
1X
p̄i [T ] =
E [pi [t]] , ∀i ∈ {1, 2, . . . , n}.
T t=1
We know by assumption that:
lim sup
n
X
T →∞ i=1
p̄i [T ] ≤ E [e]
(8)
The new dynamic power control algorithm is described in
Algorithm 1. At the end of slot t, Algorithm 1 chooses p[t+1]
based on ω[t] without requiring ω[t + 1]. To enable these
decisions, the algorithm introduces a (nonpositive) virtual
battery queue process Q[t] ≤ 0, which shall later be shown to
be related to a shifted version of the physical battery queue
E[t].
Algorithm 1 New Algorithm
Let V > 0 be a constant algorithm parameter. Initialize virtual
battery queue variable Q[0] = 0. Choose p[1] = [0, 0, . . . , 0]T
as the power action at slot 1. At the end of each slot t ∈
{1, 2, . . .}, observe ω[t] = (e[t], s[t]) and do the following:
• Update virtual battery queue Q[t]: Update Q[t] via:
Q[t] = min{Q[t − 1] + e[t] −
n
X
pi [t], 0}.
(9)
i=1
•
Power control: Choose
o
n
1
1
p[t + 1] = ProjP p[t] + ∇p U (p[t]; ω[t]) + 2 Q[t]1
V
V
(10)
as the power action for the next slot t + 1 where ProjP {·}
represents the projection onto set P, 1 denotes a column
vector of all ones and ∇p U (p[t]; ω[t]) represents a subgradient (or gradient if U (p; ω[t]) is differentiable) vector of
function U (p; ω[t]) at point p = p[t]. Note that p[t], Q[t]
and ∇p U (p[t]; ω[t]) are given constants in (10).
Note that Algorithm 1 does not explicitly enforce the energy
availability constraint (2). Let p[t + 1] be given by (10), one
may expect to use
Pn
min{ i=1 pi [t + 1], E[t]}
Pn
p̂[t + 1] =
p[t + 1]
(11)
i=1 pi [t + 1]
that scales down p[t + 1] to enforce the energy availability
constraint (2). However, our analysis in Section IV shows that
if the battery capacity is at least as large as an O(V ) constant,
then directly using p[t + 1] from (10) is ensured to always
satisfy the energy availability constraint (2). Thus, there is no
need to take the additional step (11).
B. Algorithm Inuitions
Lemma 2. The power control action p[t + 1] chosen in (10)
is to solve the following quadratic convex program
max V (∇p U (p[t]; ω[t]))T (p − p[t]) + Q[t]1T p
p
V2
−
kp − p[t]k2
2
s.t. p ∈ P
(12)
(13)
Proof. By the definition of projection, equation (10) is to
solve minp∈P kp − p[t] + V1 ∇p U (p[t]; ω[t]) + V12 Q[t]1 k2 .
By expanding the square, eliminating constant terms and
converting the minimization to a maximization of its negative
object, it is easy to show this problem is equivalent to problem
(12)-(13).
The convex projection (10), or equivalently, the quadratic
convex program (12)-(13) can be easily solved. See e.g.,
Lemma 3 in [19] for an algorithm that solves an n-dimensional
quadratic program over set P with complexity O(n log n).
Thus, the overall complexity of Algorithm 1 is low.
1) Connections with the drift-plus-penalty (DPP) technique
for Lyapunov opportunistic optimization: The Lyapunov
opportunistic optimization solves stochastic optimization
without distribution information by developing dynamic
policies that adapt control actions to the current system
state [20], [21], [22], [23], [24], [18]. The dynamic policy
from Lyapunov opportunistic optimization can be interpreted as choosing control actions to maximize a DPP
expression on each slot. Unfortunately, the problem considered in this paper is different from the conventional Lyapunov opportunistic optimization problem since the power
decision cannot be adapted to the unknown current system
state. Nevertheless, if we treat V (∇p U (p[t]; ω[t]))T (p −
2
p[t]) − V2 kp − p[t]k2 as a penalty term and Q[t]1T p as a
drift term, then Lemma 2 suggests that the power control
in Algorithm 1 can still be interpreted as maximizing a
(different) DPP expression. However, this DPP expression
is significantly different from those conventional ones used
in Lyapunov opportunistic optimization [18]. Also, the
penalty term V U (p[t + 1]; ω[t + 1]) used in conventional
Lyapunov opportunistic optimization of [18] is unavailable
in our problem since it depends on the unknown ω[t + 1].
2) Connections with online convex learning: Online convex
learning is a multi-round process where a decision maker
selects its action from a fixed set at each round before
observing the corresponding utility function [15], [16],
[17]. If we assume the wireless device is equipped with
an external free power source with infinite energy, i.e.,
the energy availability constraint (2) is dropped, then the
problem setup in this paper is similar to an online learning
problem where the decision maker selects p[t + 1] ∈ P on
each slot t + 1 to maximize an unknown reward function
U (p[t + 1]; ω[t + 1]) based on the information of previous
reward functions U (p[τ ]; ω[τ ]), τ ∈ {1, . . . , t}. In this
case, Zinkevich’s online gradient method [15], given by
p[t + 1] = ProjP {p[t] + γ∇p U (p[t]; ω[t])}
(14)
where γ is a learning rate parameter, can solve this
idealized problem. In fact, if we ignore V12 Q[t]1 involved
in (10), then (10) is identical to Zinkevich’s learning algorithm with γ = 1/V . However, Zinkevich’s algorithm and
its variations [15], [25], [17] require actions to be chosen
from a fixed set. Our problem requires p[t] chosen on
each slot t to satisfy the energy availability constraint (2),
which is time-varying since E[t] evolves over time based
on random energy arrivals and previous power allocation
decisions.
Now, it is clear why Algorithm 1 is called a learning aided
dynamic power control algorithm: Algorithm 1 can be viewed
as an enhancement of the DPP technique originally developed for Lyapunov opportunistic optimization by replacing its
penalty term with an expression used in Zinkevich’s online
gradient learning.
C. Main Results
While the above subsection provides intuitive connections
to prior work, note that existing techniques cannot be applied
to our problem. The next section develops a novel performance
analysis (summarized in Theorems 1 and 3) to show that
if E[0] = E max = O(V ), then the power control actions
from Algorithm 1 are ensured to satisfy the energy availability
constraint (2) and achieve
t
1
1X
V
E[U (p[τ ]; ω[τ ])] ≥ U ∗ − O( ) − O( ).
t τ =1
t
V
That is, for any desired > 0, by choosing V = 1/ in
Algorithm 1, we can attain an O() optimal utility for all
t ≥ Ω( 12 ) by using a battery with capacity O(1/).
IV. P ERFORMANCE A NALYSIS OF A LGORITHM 1
This section shows Algorithm 1 can attain an O() closeto-optimal utility by using a battery with capacity O(1/).
A. Drift Analysis
Define L[t] = 21 (Q[t])2 and call it a Lyapunov function.
Define the Lyapunov drift as
∆[t] = L[t + 1] − L[t]
Lemma 3. Under Algorithm 1, for all t ≥ 0, the Lyapunov
drift satisfies
∆[t] ≤ Q[t](e[t + 1] −
n
X
1
pi [t + 1]) + B
2
i=1
(15)
with constant B = (max{emax , pmax })2 , where emax is the
constant defined in Assumption 1.
Proof. Fix t ≥ 0. Recall that for any x ∈ R if y = min{x, 0}
then y 2 ≤ x2 . It follows from (9) that
2
(Q[t + 1]) ≤ (Q[t] + e[t + 1] −
n
X
pi [t + 1])2 .
i=1
Expanding the square on the right side, dividing both sides
by
terms yields
+ 1] −
Pn2 and rearranging
Pn ∆[t] ≤ Q[t](e[t
1
2
p
[t
+
1])
+
(e[t
+
1]
−
p
[t
+
1])
.
i=1 i
i=1 i
2
Pn
This lemma follows by noting that
|e[t + 1] − i=1 pi [t +
P
n
1]| ≤ max{emax , pmax } since 0 ≤ i=1 pi [t + 1] ≤ pmax and
max
0 ≤ e[t + 1] ≤ e
.
Recall that a function f : Z 7→ R is said to be strongly
concave with modulus α if there exists a constant α > 0 such
that f (z) + 12 αkzk2 is concave on Z. It is easy to show that if
f (z) is concave and α > 0, then f (z)− α2 kz−z0 k2 is strongly
concave with modulus α for any constant z0 . The maximizer
of a strongly concave function satisfies the following lemma:
Lemma 4 (Corollary 1 in [26]). Let Z ⊆ Rn be a convex
set. Let function f be strongly concave on Z with modulus α
and zopt be a global maximum of h on Z. Then, f (zopt ) ≥
f (z) + α2 kzopt − zk2 for all z ∈ Z.
Lemma 5. Let U ∗ be the utility upper bound defined in Lemma
1 and p∗ be an optimal solution to problem (4)-(6) that attains
U ∗ . At each iteration t ∈ {1, 2, . . .}, Algorithm 1 guarantees
V E[U (p[t]; ω[t])] − ∆[t] ≥ V U ∗ +
D2 + B
V2
E[Φ[t]] −
2
2
where Φ[t] = kp∗ − p[t + 1]k2 − kp∗ − p[t]k2 , D is the
constant defined in Assumption 1 and B is the constant defined
in Lemma 3.
Pn
Proof. Note that i=1 p∗i ≤ E[e]. Fix t ∈
2, . . .}. Note
P{1,
n
that V (∇p U (p[t]; ω[t]))T (p − p[t]) + Q[t] i=1 pi is a linear
function with respect to p. It follows that
n
X
T
V2
V ∇p U (p[t]; ω[t]) (p − p[t]) + Q[t]
pi −
kp − p[t]k2
2
i=1
(16)
is strongly concave with respect to p ∈ P with modulus V 2 .
Since p[t + 1] is chosen to maximize (16) over all p ∈ P, and
since p∗ ∈ P, by Lemma 4 we have
n
X
T
V ∇p U (p[t]; ω[t]) (p[t + 1] − p[t]) + Q[t]
pi [t + 1]
i=1
V2
kp[t + 1] − p[t]k2
−
2
n
X
T
≥V ∇p U (p[t]; ω[t]) (p∗ − p[t]) + Q[t]
p∗i
i=1
V2 ∗
V2 ∗
−
kp − p[t]k2 +
kp − p[t + 1]k2
2
2
n
X
T
V2
Φ[t].
=V ∇p U (p[t]; ω[t]) (p∗ − p[t]) + Q[t]
p∗i +
2
i=1
Subtracting Q[t]e[t + 1] from both sides and rearranging terms
yields
T
V ∇p U (p[t]; ω[t]) (p[t + 1] − p[t])
n
X
+ Q[t](
pi [t + 1] − e[t + 1])
i=1
n
X
T
≥V ∇p U (p[t]; ω[t]) (p∗ − p[t]) + Q[t](
p∗i − e[t + 1])
i=1
V2
V2
+
Φ[t] +
kp[t + 1] − p[t]k2 .
2
2
Adding V U (p[t]; ω[t]) to both sides and noting that
U (p[t]; ω[t]) + (∇p U (p[t]; ω[t]))T (p∗ − p[t]) ≥ U (p∗ ; ω[t])
by the concavity of U (p; ω[t]) yields
T
V U (p[t]; ω[t]) + V ∇p U (p[t]; ω[t]) (p[t + 1] − p[t])
n
X
+ Q[t](
pi [t + 1] − e[t + 1])
i=1
n
X
V2
Φ[t]
≥V U (p∗ ; ω[t]) + Q[t](
p∗i − e[t + 1]) +
2
i=1
+
V2
kp[t + 1] − p[t]k2 .
2
Rearranging terms yields
B. Utility Optimality Analysis
n
X
V U (p[t]; ω[t]) + Q[t](
The next theorem summarizes that the average expected
utility attained by Algorithm 1 is within an O(1/V ) distance
to U ∗ defined in Lemma 1.
pi [t + 1] − e[t + 1])
i=1
n
X
V2
∗
Φ[t]
≥V U (p ; ω[t]) + Q[t](
p∗i − e[t + 1]) +
2
i=1
V2
kp[t + 1] − p[t]k2
2
T
− V ∇p U (p[t]; ω[t]) (p[t + 1] − p[t])
Theorem 1. Let U ∗ be the utility bound defined in Lemma 1.
For all t ∈ {1, 2, . . .}, Algorithm 1 guarantees
+
t
(17)
Note that
T
V ∇p U (p[t]; ω[t]) (p[t + 1] − p[t])
V2
kp[t + 1] − p[t]k2
≤ k∇p U (p[t]; ω[t])k2 +
2
2
(b) 1
V2
≤ D2 +
kp[t + 1] − p[t]k2
(18)
2
2
(a) 1
where (a) follows by using basic inequality xT y ≤ 21 kxk2 +
1
2
n
with x = ∇p U (p[t]; ω[t]) and
2 kyk for all x, y ∈ R
y = V (p[t + 1] − p[t]); and (b) follows from Assumption
1. Substituting (18) into (17) yields
n
X
V U (p[t]; ω[t]) + Q[t](
pi [t + 1] − e[t + 1])
i=1
n
X
1
V2
Φ[t] − D2
≥V U (p∗ ; ω[t]) + Q[t](
p∗i − e[t + 1]) +
2
2
i=1
(19)
By Lemma 3, we have
n
X
B
pi [t + 1] − e[t + 1]) −
−∆[t] ≥ Q[t](
2
i=1
(20)
1X
V (pmax )2
B
D2 + B
E[U (p[τ ]; ω[τ ])] ≥ U ∗ −
−
−
t τ =1
2t
2V t
2V
(23)
where D is the constant defined in Assumption 1 and B is the
constant defined in Lemma 3. This implies,
t
lim sup
t→∞
t
1X
1
E[U (p[τ ]; ω[τ ])] ≥ U ∗ − O(), ∀t ≥ Ω( 2 ).
t τ =1
Proof. Fix t ∈ {1, 2, . . .}. For each τ ∈ {1, 2, . . . , t}, by
Lemma 5, we have
E[V U (p[τ ]; ω[τ ])]−E[∆[τ ]] ≥ V U ∗ +
1X
E[U (p[τ ]; ω[τ ])]
t τ =1
V
E[kp∗ − p[t + 1]k2 − kp∗ − p[1]k2 ]
2t
1
D2 + B
+
E[(Q[t + 1])2 − (Q[1])2 ] −
2V t
2V
V
1
D2 + B
∗
∗
2
≥U − E[kp − p[1]k ] −
E[(Q[1])2 ] −
2t
2V t
2V
max 2
2
(b)
V
(p
)
B
D
+
B
−
−
≥U ∗ −
2t
2V t
2V
=U∗ +
n
X
E[Q[t](
p∗i − e[t + 1])]
− e[t + 1]]
i=1
(a)
(22)
Pn
t
t
1 X
D2 + B
V X
E[Φ[τ ]] +
E[∆[τ ]] −
2t τ =1
V t τ =1
2V
(a)
D2 + B
(21)
2
Note that each Q[t] (depending only on e[τ ], p[τ ] with τ ∈
{1, 2, . . . , t}) is independent of e[t + 1]. Thus,
≥0
D2 + B
V2
E[Φ[τ ]]−
.
2
2
Summing over τ ∈ {1, 2, . . . , t}, dividing both sides by V t
and rearranging terms yields
≥U ∗ +
−
=E[Q[t]]E[
(25)
t
V U (p[t]; ω[t]) − ∆[t]
n
X
V2
Φ[t]
≥V U (p∗ ; ω[t]) + Q[t](
p∗i − e[t + 1]) +
2
i=1
p∗i
(24)
In particular, if we take V = 1/ in Algorithm 1, then
Summing (19) and (20); and cancelling common terms on both
sides yields
i=1
n
X
1X
D2 + B
E[U (p[τ ]; ω[τ ])] ≥ U ∗ −
.
t τ =1
2V
∗
where (a) follows because Q[t] ≤ 0 and
i=1 pi ≤ E[e]
(recall that e[t + 1] is an i.i.d. sample of e).
Taking expectations on both sides of (21) and using (22)
and E[U (p∗ ; ω[t])] = U ∗ yields the desired result.
where (a) follows by recalling that Φ[τ ] = kp∗ − p[τ + 1]k2 −
kp∗ − p[τ ]k2 and ∆[τ ] = 21 (Q[τ + 1])2 − 12p
(Q[τ ])2 ; and (b)
Pn
∗
∗
follows
because kp − p[1]k = kp k =
(p∗i )2 ≤
Pn
Pi=1
n
∗
max
and |Q[1]| = |Q[0] + e[1] −
i=1 piP≤ p
√ i=1 pi [1]| =
n
|e[1] − i=1 pi [1]| ≤ max{emax , pmax } = B where B is
defined in Lemma 3. So far we have proven (23).
Equation (24) follows directly by taking lim sup on both
sides of (23). Equation (25) follows by substituting V = 1
and t = 12 into (23).
C. Lower Bound for Virtual Battery Queue Q[t]
Note that Q[t] ≤ 0 by (9). This subsection further shows that
Q[t] is bounded from below. The projection ProjP {·} satisfies
the following lemma:
and Q[t] ≤ −V (Dmax + pmax ), we know bi ≤ − V1 pmax , ∀i ∈
{1, 2, . . . , n}. By Lemma 6, we have
pi [t + 1] = max{pi [t] + bi , 0}
1
≤ max{pi [t] − pmax , 0}, ∀i ∈ {1, 2, . . . , n}.
V
Lemma 6. For any p[t] ∈ P and vector b ≤ 0, where ≤
between two vectors means component-wisely less than or
equal to, p̃ = ProjP {p[t] + b} is given by
p̃i = max{pi [t] + bi , 0}, ∀i ∈ {1, 2, . . . , n}.
(26)
Proof. Recall that projection ProjP {p[t] + b} by definition is
to solve
min
p
s.t.
n
X
(pi − (pi [t] + bi ))2
i=1
n
X
pi ≤ pmax
(27)
(28)
i=1
pi ≥ 0, ∀i ∈ {1, 2, . . . , n}
(29)
Let I ⊆ {1, 2, . . . , n} be the coordinate index set given by
I
{i ∈ {1, 2, . . . , n} : pi [t] + bj < 0}. For any p such that
P=
n
max
and pi ≥ 0, ∀i ∈ {1, 2, . . . , n}, we have
i=1 pi ≤ p
n
X
=
i=1
X
(pi − (pi [t] + bi ))2
2
(pi − (pi [t] + bi )) +
i∈I
≥
X
X
≥
2
i∈I
where (a) follows because pi [t]
P+ bi < 0 for i 2∈ I and pi ≥
0, ∀i ∈ {1, 2, . . . , n}. Thus,
i∈I (pi [t] + bi ) is an object
value lower bound of problem (27)-(29).
Note that p̃ given by (26) is feasible
Pn to problem
Pn (27)-(29)
since p̃i ≥ 0, ∀i ∈ {1, 2, . . . , n} and i=1 p̃i ≤ i=1 pi [t] ≤
pmax because p̃i ≤ pi [t] for all i and p[t] ∈ P. We further
note that
where emax is the constant defined in Assumption 1 and Dmax
is the constant defined in Corollary 1. Algorithm 1 guarantees
Proof. By virtual queue update equation (9), we know Q[t]
can increase by at most emax and can decrease by at most
pmax on each slot. Since Q[0] = 0, we know Q[t] ≥ −Ql for
all t ≤ V . We need to show Q[t] ≥ −Ql for all t > V . This
can be proven by contradiction as follows:
Assume Q[t] < −Ql for some t > V . Let τ > V be the
first (smallest) slot index when this happens. By the definition
of τ , we have Q[τ ] < −Ql and
Q[τ ] < Q[τ − 1].
•
i∈I
That is, p̃ given by (26) attains the object value lower bound of
problem (27)-(29) and hence is the optimal solution to problem
(27)-(29). Thus, p̃ = ProjP {p[t] + b}.
Corollary 1. If Q[t] ≤ −V (Dmax + pmax ) with Dmax =
max{D1 , . . . , Dn }, then Algorithm 1 guarantees
1 max
p
, 0}, ∀i ∈ {1, 2, . . . , n}.
V
where D1 , . . . , Dn are constants defined in Assumption 1.
pi [t + 1] ≤ max{pi [t] −
Proof. Let b = V1 ∇p U (p[t]; ω[t]) + V12 Q[t]1. Since
∂
∂pi U (p[t]; ω[t]) ≤ Di , ∀i ∈ {1, 2, . . . , n} by Assumption 1
(31)
Now consider the value of Q[τ − V ] in two cases (note that
τ − V > 0).
X
(p˜i − (pi [t] + bi ))2 =
(pi [t] + bi )2 .
i=1
(30)
(pi − (pi [t] + bi ))
(pi [t] + bi )2
n
X
Ql =V (Dmax + 2pmax + emax )
Q[t] ≥ −Ql , ∀t ∈ {0, 1, 2, . . .}.
i∈I
(a) X
Theorem 2. Let V in Algorithm 1 be a positive integer. Define
positive constant Ql , where superscript l denotes “lower”
bound, as
2
i∈{1,2,...,n}\I
(pi − (pi [t] + bi ))
By Corollary 1, if Q[t] ≤ −V (Dmax + pmax ), then each
component of p[t+1] decreases by V1 pmax until it hits 0. That
is, if Q[t] ≤ −V (Dmax + pmax ) for sufficiently many slots,
Algorithm 1 eventually chooses 0 as the power decision. By
virtual
Pn queue update equation (9), Q[t] decreases only when
i=1 pi [t] > 0. These two observations suggest that Q[t]
yielded by Algorithm 1 should be eventually bounded from
below. This is formally summarized in the next theorem.
•
Case Q[τ −V ] ≥ −V (Dmax +pmax +emax ): Since Q[t] can
decrease by at most pmax on each slot, we know Q[τ ] ≥
−V (Dmax + 2pmax + emax ) = −Ql . This contradicts the
definition of τ .
Case Q[τ − V ] < −V (Dmax + pmax + emax ): Since Q[t]
can increase by at most emax on each slot, we know Q[t] <
−V (Dmax + pmax ) for all τ − V ≤ t ≤ τ − 1. By Corollary
1, for all τ − V ≤ t ≤ τ − 1, we have
pi [t + 1] ≤ max{pi [t] −
1 max
p
, 0}, ∀i ∈ {1, 2, . . . , n}.
V
Since the above inequality holds for all t ∈ {τ − V, τ −
V + 1, . . . , τ − 1}, and since at the start of this interval we
trivially have pi [τ − V ] ≤ pmax , ∀i ∈ {1, 2, . . . , n}, at each
step of this interval each component of the power vector
either hits zero or decreases by V1 pmax , and so after the V
steps of this interval we have pi [τ ] = 0, ∀i ∈ {1, 2, . . . , n}.
By (9), we have
Q[τ ] = min{Q[τ − 1] + e[τ ] −
n
X
pi [τ ], 0}
where (a) follows from the induction hypothesis E[t0 ] =
Q[t0 ] + E max and (b) follows from the energy queue dynamic
(3). Thus, (32) holds for t = t0 + 1.
Now observe
i=1
E[t0 + 1] = Q[t0 + 1] + E max
= min{Q[τ − 1] + e[τ ], 0}
(a)
≥ E max − Ql
≥ min{Q[τ − 1], 0}
=Q[τ − 1]
where the final equality holds because the queue is never
positive (see (9)). This contradicts (31).
Both cases lead to contradictions. Thus, Q[t] ≥ −Ql for all
t>V.
D. Energy Availability Guarantee
To implement the power decisions of Algorithm 1 for the
physical battery system E[t] from equations (2)-(3), we must
ensure the energy availability constraint (2) holds on each
slot. The next theorem shows that Algorithm 1 ensures the
constraint (2) always holds as long as the battery capacity
satisfies E max ≥ Ql + pmax and the initial energy satisfies
E[0] = E max . It also explains that Q[t] used in Algorithm 1
is a shifted version of the physical battery backlog E[t].
Theorem 3. If E[0] = E max ≥ Ql + pmax , where Ql is the
constant defined in Theorem 2, then Algorithm 1 ensures the
energy availability constraint (2) on each slot t ∈ {1, 2, . . .}.
Moreover
E[t] = Q[t] + E
max
, ∀t ∈ {0, 1, 2, . . .}.
(32)
Proof.
Pn Note that to show the energy availability constraint
i=1 pi [t] ≤ E[t − 1], ∀t ∈ {1, 2, . . .} is equivalent to show
n
X
pi [t + 1] ≤ E[t], ∀t ∈ {0, 1, 2, . . .}.
(33)
i=1
This lemma can be proven by inductions.
Note that E[0] = E max and Q[0] = 0. It is immediate
that P
(32) holds for t = 0. Since E[0] = E max ≥ pmax
n
and i=1 pi [1] ≤ pmax , equation (33) also holds for t = 0.
Assume (33) and (32) hold for t = t0 and consider t = t0 + 1.
By virtual queue dynamic (9), we have
Q[t0 + 1] = min{Q[t0 ] + e[t0 ] −
n
X
pi [t0 ], 0}
i=1
Adding E max on both sides yields
Q[t0 + 1] + E max
= min{Q[t0 ] + e[t0 + 1] −
(a)
= min{E[t0 ] + e[t0 + 1] −
n
X
i=1
n
X
i=1
(b)
=E[t0 + 1]
pi [t0 + 1] + E max , E max }
pi [t0 + 1], E max }
≥ pmax
n
(b) X
≥
pi [t0 + 2]
i=1
where (a) follows from the fact that Q[t] ≥ −Ql , ∀t ∈
{0, 1, 2, . . .} by Theorem 2; (b) holds since sum power is never
more than pmax . Thus, (33) holds for t = t0 + 1.
Thus, this theorem follows by induction.
E. Utility Optimality and Battery Capacity Tradeoff
By Theorem 1, Algorithm 1 is guaranteed to attain a utility
within an O(1/V ) distance to the optimal utility U ∗ . To obtain
an O()-optimal utility, we can choose V = d1/e, where dxe
represents the smallest integer no less than x. In this case,
Ql defined in (3) is order O(V ). By Theorem 3,we need the
battery capacity E max ≥ Ql + pmax = O(V ) = O(1/)
to satisfy the energy availability constraint. Thus, there is a
[O(), O(1/)] tradeoff between the utility optimality and the
required battery capacity.
F. Extensions
Thus far, we have assumed that ω[t] is known with one
slot delay, i.e., at the end of slot t, or equivalently, at the
beginning of slot t + 1. In fact, if ω(t) is observed with t0 slot
delay (at the end of slot t + t0 − 1), we can modify Algorithm
1 by initializing p[τ ] = 0, τ ∈ {1, 2, . . . , t0 } and
Pnupdating
Q[t − t0 + 1] = min{Q[t − t0 ] + e[t − t0 + 1] − i=1 pi [t −
t0 + 1], 0}, p[t + 1] = ProjP {p[t − t0 + 1] + V1 ∇p U (p[t −
t0 + 1]; ω[t − t0 + 1]) + V12 Q[t − t0 + 1]1} at the end of
each slot t ∈ {t0 , t0 + 1, . . .}. By extending the analysis in
this section (from a t0 = 1 version to a general t0 version), a
similar [O(), O(1/)] tradeoff can be established.
V. N UMERICAL E XPERIMENT
In this section, we consider an energy harvesting wireless
device transmitting over 2 subbands whose channel strength
is represented by s1 [t] and s2 [t], respectively. Our goal is to
decide the power action p[t] to maximize the utility/throughput
given by (1). Let P = {p : p1 + p2 ≤ 5, p1 ≥ 0, p2 ≥ 0}.
Let harvested energy e[t] satisfy the uniform distribution over
interval [0, 3]. Assume both subbands are Rayleigh fading
channels where s1 [t] follows the Rayleigh distribution with
parameter σ = 0.5 truncated in the range [0, 4] and s2 [t]
follows the Rayleigh distribution with parameter σ = 1
truncated in the range [0, 4].
By assuming the perfect knowledge of distributions, we
solve the deterministic problem (4)-(6) and obtain U ∗ =
1.0391. To verify the performance proven in Theorems 1
E max = 50
1
0.8
E max = 20
0.6
E max = 10
Pt
Peoformance of Algorithm 1 using a battery of capacity Ql + pmax
Performance of Algorithm 1 (V = 40) using a small capacity battery
U$
1
t
1.2
1.2
==1 E[U (p[= ]; ![= ])]
and 3, we run Algorithm 1 with V ∈ {5, 10, 20, 40} and
E[0] = E max = Ql + pmax over 1000 independent simulation
runs. In all the simulation runs, the power actions yielded by
Algorithm 1 always satisfy the energy availability constraints.
We also plot the averaged utility performance in Figure 1,
where the y-axis is the running average of expected utility.
Figure 1 shows that the utility performance can approach U ∗
by using larger V parameter.
0.4
U$
1
0.2
0
10 0
V =5
1
t
10 1
10 2
10 3
10 4
10 5
Time slot t
0.6
V = 40
Fig. 2. Utility performance (averaged over 1000 independent simulation runs)
of Algorithm 1 with V = 40 for different E max .
Pt
==1 E[U (p[= ]; ![= ])]
V = 10
0.8
0.4
V = 20
R EFERENCES
0.2
0
10 0
10 1
10 2
10 3
10 4
10 5
Time slot t
Fig. 1. Utility performance (averaged over 1000 independent simulation runs)
of Algorithm 1 with E[0] = E max = Ql + pmax for different V .
In practice, it is possible that for a given V , the battery
capacity E max = Ql + pmax required in Theorem 3 is too
large. If we
Pnrun Algorithm 1 with small capacity batteries
such that
i=1 pi [t + 1] ≥ E[t] for certain slot t, a reasonable choice is to scale down p[t + 1] by (11) and use
p̂[t + 1] as the power action. Now, we run simulations by
fixing V = 40 in Algorithm 1 and test its performance with
small capacity batteries. By Theorem 3, the required battery
capacity to ensure energy availability is E max = 685. In
our simulations, we choose small E max ∈ {10, 20, 50} and
E[0] = 0, i.e., the battery is initially empty. If p[t + 1]
from Algorithm 1 violates energy availability constraint (2),
we use p̂[t + 1] from (11) as the true power action that is
enforced to satisfy (2)Pand update the energy backlog by
n
E[t+1] = min{E[t]− i=1 p̂i [t+1]+e[t+1], E max }. Figure
2 plots the utility performance of Algorithm 1 in this practical
scenario and shows that even with small capacity batteries,
Algorithm 1 still achieves a utility close to U ∗ . This further
demonstrates the superior performance of our algorithm.
VI. C ONCLUSION
This paper develops a new learning aided power control
algorithm for energy harvesting devices, without requiring the
current system state or the distribution information. This new
algorithm can achieve an O() optimal utility by using a
battery with capacity O(1/).
[1] J. A. Paradiso and T. Starner, “Energy scavenging for mobile and
wireless electronics,” IEEE Pervasive Computing, vol. 4, no. 1, pp. 18–
27, 2005.
[2] S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes:
Survey and implications,” IEEE Communications Surveys & Tutorials,
vol. 13, no. 3, pp. 443–461, 2011.
[3] S. Ulukus, A. Yener, E. Erkip, O. Simeone, M. Zorzi, P. Grover, and
K. Huang, “Energy harvesting wireless communications: A review of
recent advances,” IEEE Journal on Selected Areas in Communications,
vol. 33, no. 3, pp. 360–381, 2015.
[4] A. Kansal, J. Hsu, S. Zahedi, and M. B. Srivastava, “Power management
in energy harvesting sensor networks,” ACM Transactions on Embedded
Computing Systems, vol. 6, no. 4, 2007.
[5] P. Kamalinejad, C. Mahapatra, Z. Sheng, S. Mirabbasi, V. C. Leung,
and Y. L. Guan, “Wireless energy harvesting for the internet of things,”
IEEE Communications Magazine, vol. 53, no. 6, pp. 102–108, 2015.
[6] E. Hossain and M. Hasan, “5G cellular: key enabling technologies and
research challenges,” IEEE Instrumentation & Measurement Magazine,
vol. 18, no. 3, pp. 11–21, 2015.
[7] J. Yang and S. Ulukus, “Optimal packet scheduling in an energy harvesting communication system,” IEEE Transactions on Communications,
vol. 60, no. 1, pp. 220–230, 2012.
[8] K. Tutuncuoglu and A. Yener, “Optimum transmission policies for
battery limited energy harvesting nodes,” IEEE Transactions on Wireless
Communications, vol. 11, no. 3, pp. 1180–1189, 2012.
[9] P. Blasco, D. Gunduz, and M. Dohler, “A learning theoretic approach
to energy harvesting communication system optimization,” IEEE Transactions on Wireless Communications, vol. 12, no. 4, pp. 1872–1882,
2013.
[10] N. Michelusi, K. Stamatiou, and M. Zorzi, “Transmission policies for
energy harvesting sensors with time-correlated energy supply,” IEEE
Transactions on Communications, vol. 61, no. 7, pp. 2988–3001, 2013.
[11] W. Wu, J. Wang, X. Wang, F. Shan, and J. Luo, “Online throughput
maximization for energy harvesting communication systems with battery
overflow,” IEEE Transactions on Mobile Computing, vol. 16, no. 1, pp.
185–197, 2017.
[12] M. Gatzianas, L. Georgiadis, and L. Tassiulas, “Control of wireless
networks with rechargeable batteries,” IEEE Transactions on Wireless
Communications, vol. 9, no. 2, pp. 581–593, 2010.
[13] L. Huang and M. J. Neely, “Utility optimal scheduling in energyharvesting networks,” IEEE/ACM Transactions on Networking, vol. 21,
no. 4, pp. 1117–1130, 2013.
[14] R. Urgaonkar, B. Urgaonkar, M. J. Neely, and A. Sivasubramaniam,
“Optimal power cost management using stored energy in data centers,”
Proceedings of ACM SIGMETRICS, 2011.
[15] M. Zinkevich, “Online convex programming and generalized infinitesimal gradient ascent,” in Proceedings of International Conference on
Machine Learning (ICML), 2003.
[16] N. Cesa-Bianchi and G. Lugosi, Prediction, Learning, and Games.
Cambridge University Press, 2006.
[17] S. Shalev-Shwartz, “Online learning and online convex optimization,”
Foundations and Trends in Machine Learning, vol. 4, no. 2, pp. 107–
194, 2011.
[18] M. J. Neely, Stochastic Network Optimization with Application to
Communication and Queueing Systems. Morgan & Claypool Publishers,
2010.
[19] H. Yu and M. J. Neely, “A new backpressure algorithm for joint rate
control and routing with vanishing utility optimality gaps and finite
queue lengths,” in Proceedings of IEEE International Conference on
Computer Communications (INFOCOM), 2017.
[20] L. Tassiulas and A. Ephremides, “Stability properties of constrained
queueing systems and scheduling policies for maximum throughput in
multihop radio networks,” IEEE Transactions on Automatic Control,
vol. 37, no. 12, pp. 1936–1948, 1992.
[21] M. J. Neely, E. Modiano, and C. E. Rohrs, “Dynamic power allocation
and routing for time-varying wireless networks,” IEEE Journal on
Selected Areas in Communications, vol. 23, no. 1, pp. 89–103, 2005.
[22] A. Eryilmaz and R. Srikant, “Joint congestion control, routing, and mac
for stability and fairness in wireless networks,” IEEE Journal on Selected
Areas in Communications, vol. 24, no. 8, pp. 1514–1524, 2006.
[23] A. L. Stolyar, “Maximizing queueing network utility subject to stability:
Greedy primal-dual algorithm,” Queueing Systems, vol. 50, no. 4, pp.
401–457, 2005.
[24] M. J. Neely, E. Modiano, and C.-P. Li, “Fairness and optimal stochastic
control for heterogeneous networks,” IEEE/ACM Transactions on Networking, vol. 16, no. 2, pp. 396–409, 2008.
[25] E. Hazan, A. Agarwal, and S. Kale, “Logarithmic regret algorithms for
online convex optimization,” Machine Learning, vol. 69, pp. 169–192,
2007.
[26] H. Yu and M. J. Neely, “A simple parallel algorithm with an O(1/t)
convergence rate for general convex programs,” SIAM Journal on
Optimization, vol. 27, no. 2, pp. 759–783, 2017.
| 3cs.SY
|
1
Level-Triggered Harvest-then-Consume Protocol
with Two Bits or Less Energy State Information
arXiv:1709.06928v1 [cs.NI] 23 Jul 2017
Sudarshan Guruacharya, Vandana Mittal, and Ekram Hossain
Abstract—We propose a variation of harvest-then-consume
protocol with low complexity where the harvest and consume
phases change when the battery energy level reaches certain
thresholds. The proposed protocol allows us to control the
possible energy outage during consumption phase. Assuming that
the battery is perfect and that the energy arrival is a renewal
process, we analyze the duty cycle and the operating cycle speed
of the protocol. The proposed protocol also allows for limited
battery energy state information. The cases when the system has
two-bits, one-bit, and zero-bit of battery energy state information
are studied in detail. Numerical simulations verify the obtained
formulas.
Index Terms—Energy-harvesting wireless communication,
harvest-then-consume, limited energy state information (ESI)
I. I NTRODUCTION
Harvest-then-consume protocol is an instance of timeswitching architecture of energy harvesting communication,
where harvest and consume phases alternate between each
other [1]. An obvious drawback of harvest-then-consume
protocol is the delay induced by the harvest phase. Despite the
drawback, the harvest-then-consume protocol has been studied
in the context of relay communication [2], [3], cognitive networks [4], sensor networks [5]–[7], and wireless information
and power transfer (WIPT) [8]–[12] where the alternation
between the harvest and consume phase lends itself to the halfduplex nature of these applications. In these work, the size of
the time frame of a harvest-consume cycle is assumed to be
fixed, and the duty cycle is assumed as the control variable,
with the goal of maximizing the system throughput. However,
this can result in energy outage when the harvested energy is
insufficient to power the consumer’s application.
In this letter, we propose a variation of harvest-thenconsume protocol with low complexity where the phase
change happens when the battery energy level crosses certain
threshold; hence the qualifier level-triggered. The protocol we
describe has a built-in guarantee over energy outage. A similar
protocol was used for empirical work in [13] for a time slotted
process. The current work is based on our prior work [14]
where we investigated the recharge time distribution of an
energy harvesting system. Being level-triggered also means
that only finite bits are needed to monitor the changes in the
battery level. This greatly simplifies the system design, since
the circuit needed to monitor the battery’s energy state can be
The authors are with the Department of Electrical and Computer Engineering at the University of Manitoba, Canada (Emails: {Sudarshan.Gurucharya,
Ekram.Hossain}@umanitoba.ca, [email protected]). The work was
supported by a CRD grant (CRDPJ 461412-13) from the Natural Sciences
and Engineering Research Council of Canada (NSERC).
reduced or done away with completely. Also, since the goal of
the proposed protocol is to ensure the energy sufficiency for
a given time frame, complicated optimization problem over
finite or infinite time horizon is avoided.
We will analyze the operating speed and the duty cycle of
the proposed protocol, and establish upper bounds on performance. While these metrics are independent of the purpose of
the energy consumption, nevertheless, we will assume that the
energy is used to transmit messages in a wireless communications system. We will analyze the cases when two bits, one
bit, and zero bit of energy state information (ESI) is available.
The case of zero bit ESI is interesting in itself, since for
this case the harvest and consume durations are deterministic.
Note that such a simple harvest-then-consume protocol will
be particularly suitable for low-complexity energy-harvesting
wireless sensor nodes. To the best of our knowledge, we are
not aware of any prior work that deals with limited ESI.
II. S YSTEM M ODEL
Let U (t) be the energy level of the battery at any given time
t. At any given time, the battery can be in one of the three
possible states: (i) Battery is empty, U (t) = 0, (ii) Battery
is not empty but the energy level is below some required
threshold u, such that 0 < U (t) < u, (iii) Battery energy
level is above the required threshold, U (t) > u. These three
possible energy states can be represented by just two bits of
information. Thus, we have a system with limited ESI.
In this letter, we will consider the following simple version
of the harvest-then-consume protocol:
•
•
The system switches off and goes into harvest phase when
the battery is empty, i.e. after U (t) = 0.
The system switches on and goes into consumption phase
after the battery has acquired u amount of energy, i.e.
after U (t) > u.
Here the harvest and consume phases are triggered when
the battery attains certain fixed levels, hence the name level
triggered.
For simplicity, during the consumption phase we will assume that the rate of energy consumption (or the consumed
power) p is constant. Let τc = inf{t : U (t) > u, U (0) = 0}
be the recharge duration and τd = inf{t : U (τc + t) = 0} be
the discharge duration. Then, the total harvest-consume cycle
duration is T = τc +τd . The performance metric of the harvestthen-consume protocol is taken to be
ω=
1
E[T ]
and ρ =
E[τd ]
.
E[T ]
(1)
2
Here, ω represents the cycle speed at which the protocol
can operate and has the units of Hertz. The other metric,
ρ, is the duty cycle, such that ρ ∈ (0, 1). It represents the
fraction of time that the system does some useful work and is
a dimensionless number. In this paper, we will derive formulas
for these two metrics of interest.
During the harvest phase, we model that the recharge
process for perfect battery (i.e. no self-discharge) as
NA (t)
U (t) =
X
Xi ,
(2)
i=1
where NA (t) = min{k : A0 + A1 + · · · + Ak ≤ t} counts
the number of energy arrivals, Ai≥1 is the inter-arrival time
of energy packets, A0 is the residual time, and Xi is the
energy packet size. The energy arrival is assumed to be a
delayed renewal process. The {Ai≥1 } and {Xi } are assumed
to be independent and identically distributed with finite mean
and variance. Also, we assume that {Ai } and {Xi } are
independent of each other. Lastly, we assume that the joint
distribution of the random vectors {(Ai , Xi )} are identically
distributed as (A, X). For notational convenience, we will
denote λ = 1/E[A] and X̄ = E[X].
With less than two bits of ESI, the system may not be able
tell if the battery has the desired energy level. As such, we
need to impose a statistical guarantee on the energy outage.
Let the energy outage constraint at the switching time tc be
P (U (tc ) ≤ u) = θ1 ,
(3)
where θ1 ∈ (0, 1). Here tc is a fixed duration of recharge, after
which the system is turned on for the consume phase.
Given the recharge process in (2), the mean and variance
of τc for large u are, respectively [14, Eqns. (8), (9)],
E[τc ] ∼ C1 +
u
,
λX̄
and V[τc ] ∼ C2 +
2
λ 2 σA
)/2λ
γ2u
.
X̄ 3
we can solve for p =
(3)
µA
3µA
where Φ(·) is the standard normal distribution.
If the harvested energy is used to transmit information in a
narrow band channel, then in the transmit phase, assuming a
point-to-point wireless communications system with transmit
power p, flat fading channel gain g, and additive white Gaussian noise with power N , we have the signal-to-noise ratio
(SNR) given by Z = gp/N . We assume that the transmitter
always has data to transmit. Let the SNR outage constraint be
(6)
ζN
.
−1
FG
(θ2 )
III. W ITH T WO B ITS OF E NERGY S TATE I NFORMATION
With two bits of ESI, the system can know when the battery
is empty and when it has sufficient energy. The duration that
the recharge process takes to cross the desired energy level u
is τc , where τc is a random variable. Once the required energy
level has been crossed, the system is turned on. The time it
takes to fully discharge the battery is τd = U (τc )/p; and the
c)
total charging and discharging time is T = τc + U (τ
p . Here
again T is a random variable. Also, at the level crossing time
τc , U (τc ) = u + V , where V ≥ 0 is the value by which U (τc )
overshoots the required energy level u. Since U (t) is renewal
process, the overshoot is given by the stationary residual density of X, assuming u is large, as fV (v) = X̄ −1 [1 − FX (v)].
Thus, we have T = τc + (u + V )/p. Here, τc and V are
independent of each other, thus the distribution of their sum
can be obtained by the convolution of their distributions. We
can find the mean value of U as E[U (τc )] = u + C3 , where
2
C3 = (σX
+ X̄ 2 )/2X̄ is the mean of V which does not depend
on u. Also, using(4) and the
mean
of U (τc ), we have the mean
of T as E[T ] = λ1X̄ + p1 u + C1 + Cp3 .
In general, the duty cycle ρ = E[τd ]/E[T ] =
E[U (τc )]/pE[T ] and the system speed ω = 1/E[T ] are
u + C3
,
1+
u + (pC1 + C3 )
−1
1
1
C3
ω=
+
.
u + C1 +
p λX̄
p
ρ=
(4)
Here the constants C1 = (1 +
and C2 =
−
2 2 2
(3)
µA +σA
2
, where µA is the third moment of A; σA and
2µA
2
σX are the variances of A and X, respectively; and γ 2 =
2
2
λ−2 σX
+σA
X̄ 2 . For large u, we can neglect the constant term
and simply write E[τc ] ∼ u/λX̄ and V[τc ] ∼ γ 2 u/X̄ 3 . Thus,
invoking the central limit theorem, for large u the distribution
of τc is given by [14, Eqn. (11)],
!
tc − E[τc ]
p
P (τc (u) ≤ tc ) ≈ Φ
,
(5)
V[τc ]
P (Z ≤ ζ) = θ2 ,
where θ2 ∈ (0, 1) while ζ is the threshold SNR required
for correct decoding of the message signal.
Substituting
the
=
θ
.
Since
expression for SNR in (6), we have P g ≤ ζN
2
p
ζN
=
F
,
where
F
is
the
distribution
of g,
P g ≤ ζN
G
G
p
p
p
λX̄
(7)
(8)
As a special case, as u → 0, we have the duty cycle as
ρ = C3 /(pC1 + C3 ), and the system’s cycle speed as ω =
p/(pC1 + C3 ). This is also the fastest speed that the system
can attain; thus ω ≤ p/(pC1 + C3 ).
1
Likewise, as u → ∞, we can ignore the
constant terms.
1
1
Thus, E[U (τc )] ∼ u and E[T ] ∼ p + λX̄ u. In other words,
larger the required energy, more we need to wait. Also, the
λX̄
duty cycle is ρ ∼ λX̄+p
, and the system’s cycle speed is
pρ
ω∼ u.
Interestingly, this limiting value of ρ is not equal to its value
at u = 0. Setting u = 0 represents an opportunistic scheme
where the harvested energy is immediately consumed. When
u = 0, we have U (τc ) = X and τd = U (τc )/p = X/p. Hence,
E[τd ] = X̄/p. Similarly, E[T ] = E[τc ] + E[τd ] = C1 + X̄/p.
Therefore, ρ = (1 + pC1 /X̄)−1 and ω = (C1 + X̄/p)−1 .
For small values of u, these formulas for ρ and ω will not be
accurate, since we assume stationary residual distribution for
A0 and V , which is valid only for large u. Note that when
u = 0, only one bit is required to check the battery status;
thus this analysis is valid for Section IV as well.
1 Here,
f (x) ∼ g(x) if and only if limx→∞
f (x)
g(x)
= 1.
3
IV. W ITH O NE B IT E NERGY S TATE I NFORMATION
Here, we assume that the system can discern whether or not
the battery is empty. As such, during the charging process, only
statistical guarantee (3) can be given for the energy outage.
Using (4) and (5) in (3), the switching time tc is given by
r
γ2u
u
−1
.
(9)
tc = C1 + Φ (1 − θ1 ) C2 + 3 +
λX̄
X̄
√
The minima at u = 0 is tc,min = C1 + Φ−1 (1 − θ1 ) C2 ,
which gives the minimum waiting time for an energy packet
to arrive.
Once the system is switched on at tc , the duration it takes
for the battery to completely discharge is τd = U (tc )/p. We
can find the distribution for the discharge duration as [15]
ptd − λX̄tc
√
. (10)
P (τd ≤ td ) = P (U (tc ) ≤ ptd ) = Φ
γλ3/2 tc
Thus, the mean discharge time is E[τd ] = λX̄tc /p.
Since the system can detect when the battery is empty, we
can start the recharging process when the battery is completely
discharged. Thus, the cycle duration is T = tc + τd . The
distribution of T is
pt − (p + λX̄)tc
√
P (T ≤ t) = P (τd ≤ t − tc ) = Φ
.
γλ3/2 tc
d = C1 /T . Here, all the constants a, b, c, d ≥ 0. Now, solving
for ρ, we obtain
p
2(1 + a)(1 − d) + b ± b2 + 4(1 + a)((1 + a)c + b(1 − d))
.
ρ=
2(1 + a)2
(13)
When T → ∞, b → 0, c → 0, and d → 0, thus ρ →
1/(1 + a), regardless of the value of θ1 . Thus,
ρ≈
λX̄
,
p + λX̄
for large T.
(14)
A. Feasibility Conditions
For the solution ρ to be feasible, ρ should be within the
interval (0, 1). Thus we need to check the conditions when
ρ > 0 and ρ < 1.
For ρ > 0, from (13), after some simplification, we obtain
the condition (a + 1)2 (c − (d − 1)2 ) > 0. Since a + 1 is always
positive,
c > (d − 1)2 .
(15)
Substituting the definitions of c and d, we find that this
condition reduces to T > tc,min , where tc,min is as given
in Section IV. If the terms c and d were neglected, then the
condition would have reduced to a + 1 > 0, which is always
true.
For ρ < 1, from (13) after some simplification, we obtain
Thus, the mean of T is E[T ] = (p+λX̄)tc /p. Hence, the duty
cycle and the cycle speed are
(a + d)2 > b + c.
p
λX̄
, ω=
.
ρ=
p + λX̄
(p + λX̄)tc
Substituting the expressions for b, c, and d, results in the
condition f (T ) > 0, where f (T ) is a quadratic equation in
terms of T , given as
(11)
Interestingly, since the maximum value of ω is obtained
when tc is minimum at u = 0, we have the upper bound
p
√
ω≤
.
(12)
(p + λX̄)(C1 + Φ−1 (1 − θ1 ) C2 )
If we neglect the constant term and the term with square
pρ
pλX̄
1
root for tc , then we have ω = E[T
] ∼ u(p+λX̄) = u .
V. N O E NERGY S TATE I NFORMATION
In this case, since we do not have any information on the
battery state, we need to rely on the statistical constraints (3).
Unlike other cases, here T is a control parameter. Let the
harvest duration be tc , as given by (9), and the consumption
duration be T − tc > 0. Here, we do not concern ourself
with complete discharge of the battery. Rather, we focus on
the consumption of fixed u amount of energy within the
consumption phase. Once this amount of energy is consumed,
the system reverts to the harvest phase. Thus, some excess
energy may remain in the battery after the consume phase. For
simplicity, we will neglect the excess energy in the analysis.
This is equivalent to assuming that any excess energy after
a complete harvest-consume cycle is wasted or dissipated
unproductively.
Since the consumed power is maintained at fixed p, the
consumed energy is u = ptd =
√ pρT . Substituting this value of
u in (9), we have 1−ρ = d+ c + bρ+aρ, where a = p/λX̄,
b = p(γΦ−1 (1 − θ1 ))2 /X̄ 3 T , c = C2 (Φ−1 (1 − θ1 )/T )2 , and
(16)
f (T ) = KT 2 + LT + M,
where K = a2 , L = 2aC1 + pΦ−1 (1 − θ1 )2 /X̄ 3 , and M =
C12 − C2 Φ−1 (1 − θ1 )2 . The condition f (T ) > 0 is satisfied
for any T if the discriminant of f (T ) is negative. That is, if
L2 − 4KM < 0. When this is not the case, we have T > T+ ,
where T√
+ is the largest root of f (T ) = 0 given by T+ =
(−L + L2 − 4KM )/2K. Had we neglected c and d, the
condition (16) would have simplify to a2 > b; and substituting
the expression for a and b, and solving for T would have given
2 2
γ
us T > λpX̄
[Φ−1 (1 − θ1 )]2 .
Hence, we have the lower bound on T as T >
max(tc,min , T+ ) and the upper bound on ω as
ω<
1
max(tc,min , T+ )
.
(17)
B. Possible Variation
If proper discharge is to be ensured for fixed cycle period
T , then allowed discharge time is td = T − tc . Thus, we have
from (10)
pT − (p + λX̄)tc
√
P (τd ≤ T − tc ) = Φ
.
γλ3/2 tc
Let the probability that battery is fully discharged by time td
be constrained at P (τd ≤ td ) = θ3 . Then, we have
γλ3/2 √ −1
λX̄
T = 1+
tc +
tc Φ (θ3 ).
(18)
p
p
4
0.44
1.2
Duty cycle (ρ)
0.42
0.4
Operating cycle speed (ω)
One bit theoretical
One bit monte carlo
Two bit theoretical
Two bit monte carlo
0.38
0.36
0.34
0.32
One bit theoretical
One bit monte carlo
Two bit theoretical
Two bit monte carlo
1
0.8
0.6
0.4
0.2
0
0
10
20
30
40
Battery energy level (u)
(a)
0
10
20
30
40
Battery energy level (u)
(b)
Fig. 1: (a) Duty cycle versus battery energy level, (b) Operating cycle speed versus battery energy level, when energy packet size X and
inter-arrival time A are uniformly distributed.
If we ignore the square
for large u, we have
root terms,
the approximation T ∼ λ1X̄ + p1 u; and similarly, the duty
λX̄
. Likewise, the cycle speed
cycle ρ = 1 − tc /T is ρ ∼ p+λ
X̄
λX̄p
1
1
of the system is T ∼ u λX̄+p = ρp
u .
VI. N UMERICAL V ERIFICATION
In this section, we verify the obtained formulas with Monte
Carlo simulations for the case of two-bit and one-bit ESI. In
Fig. 1a and Fig. 1b we plot the duty cycle ρ and operating
cycle speed ω with respect to the threshold energy level u. For
the two bit case, since the time τc and τd are known, it is easy
to calculate the charging and discharging time and hence the
duty cycle and operating frequency. However, for one bit case,
charging time tc is calculated using (9) and discharge time τd
is calculated using U (tc )/p. Both the energy packet size and
energy arrival are assumed to follow a uniform distribution
U(0, 2), with unit mean and variance 1/3. We assume that the
power consumption p = 2 and θ1 = 0.1. For a given u, 10,000
simulations are run to obtain a single value of ρ and ω.
The theoretical expressions for ρ and ω for two bit ESI are
given by equations (6) and (7), respectively, and for one bit
ESI, ρ and ω are given by equations (10) and (11), respectively.
From Fig. 1a and Fig. 1b, we see that both ρ and ω do not vary
much for higher values of u. The results from the simulations
match closely with the theoretical predictions.
VII. C ONCLUSION
A level-triggered harvest-then-consume protocol has been
proposed. The duty cycle and operating cycle speed of the
system have been derived for cases when the system has twobits, one-bit, and zero-bit of battery energy state information.
Upper bounds on the system’s speed have been obtained.
Monte Carlo simulations have been performed to verify the
obtained formulas.
R EFERENCES
[1] M.-L. Ku, et al., “Advances in energy harvesting communications: Past,
present, and future challenges,” IEEE Commun. Surveys & Tutorials, no.
2, vol. 18, pp. 1384–1412, 2016.
[2] I. Krikidis, S. Timotheou, and S. Sasaki, “RF energy transfer for cooperative networks: Data relaying or energy harvesting?,” IEEE Commun.
Lett., vol. 16, no. 11, pp. 1772–1775, Nov. 2012.
[3] He Chen et al., “Harvest-then-cooperate: Wireless-powered cooperative
communications,” IEEE Trans. Signal Process., vol. 63, no. 7, pp. 1700–
1711, Apr., 2015.
[4] C. Wu et al., “Energy utilization efficient frame structure for energy
harvesting cognitive radio networks,” IEEE Wireless Commun. Lett., vol.
5, no. 5, pp. 488–491, Oct. 2016.
[5] S. Park, et al., “Optimal mode selection for cognitive radio sensor
networks with RF energy harvesting,” Proc. IEEE 23rd Int. Symp. Pers.
Indoor Mobile Radio Commun. (PIMRC), pp. 2155–2159, Sep. 2012.
[6] N. Jain and V.A. Bohara, “Energy harvesting and spectrum sharing
protocol for wireless sensor networks,” IEEE Wireless Commun. Lett.,
vol. 4, no. 6, pp. 697-700, Jun. 2015.
[7] W. Liu, et al., “Energy harvesting wireless sensor networks: Delay
analysis considering energy costs of sensing and transmission,” IEEE
Trans. Wireless Commun., vol. 15, no. 7, pp. 4635–4650, Jul. 2016.
[8] S. Luo, R. Zhang, and T.J. Lim, “Optimal save-then-transmit protocol for
energy harvesting wireless transmitters,” IEEE Trans. Wireless Commun.,
vol. 12, no. 3, pp. 1196–1207, Feb. 2013.
[9] H. Ju and R. Zhang, “Throughput maximization in wireless powered
communication networks,” IEEE Trans. Wireless Commun., vol. 13, no.
1, pp. 418–428, Jan. 2014.
[10] T.A. Zewde and M.C. Gursoy, “Wireless-powered communication under
statistical quality of service constraints,” IEEE Int. Conf. Commun. (ICC),
22-27 May 2016.
[11] Z.H. Velkov, et al., “Wireless networks with energy harvesting and
power transfer: Joint power and time allocation,” IEEE Signal Process.
Lett., vol. 23, no. 1, pp. 50–54, Jan. 2016.
[12] F. Zhao, L. Wei, and H. Chen, “Optimal time allocation for wireless
information and power transfer in wireless powered communication
systems,” IEEE Trans. Veh. Tech., vol. 65, no. 3, pp. 1830–1835, Mar.
2016.
[13] P. Lee, et al., “Empirical modeling of a solar-powered energy harvesting
wireless sensor node for time-slotted operation,” IEEE Wireless Commun.
Network. Conf. (WCNC), 28-31 March 2011.
[14] S. Guruacharya, V. Mittal, and E. Hossain, “On the battery recharge time
in a stochastic energy harvesting system,” IEEE Wireless Commun. Lett.,
submitted, 2017. Available [Online]: https://arxiv.org/abs/1706.03183
[15] F.E. Beichelt and L.P. Fatti, Stochastic Processes and Their Applications.
CRC Press, 2002.
| 7cs.IT
|
Learning to Design Games: Strategic Environments
in Reinforcement Learning
Haifeng Zhang]∗ , Jun Wang‡ , Zhiming Zhou\ , Weinan Zhang\ , Ying Wen‡ , Yong Yu\ , Wenxin Li†
Peking University, ‡ University College London,\ Shanghai Jiao Tong University
[email protected], [email protected], [email protected]
arXiv:1707.01310v3 [cs.AI] 12 Oct 2017
]
Abstract
In typical reinforcement learning (RL), the environment is assumed given and the goal of the learning is to identify an
optimal policy for the agent taking actions through its interactions with the environment. In this paper, we extend this
setting by considering the environment is not given, but controllable and learnable through its interaction with the agent
at the same time. Theoretically, we find a dual Markov decision process (MDP) w.r.t. the environment to that w.r.t. the
agent, and solving the dual MDP-policy pair yields a policy gradient solution to optimizing the parametrized environment. Furthermore, environments with discontinuous parameters are addressed by a proposed general generative framework. While the idea is illustrated by an extended two-agent
rock-paper-scissors game, our experiments on a Maze game
design task show the effectiveness of the proposed algorithm
in generating diverse and challenging Mazes against different
agents with various settings.
Introduction
Reinforcement learning (RL) is typically concerned with
a scenario where an agent (or multiple agents) taking actions and receiving rewards from an environment (Kaelbling, Littman, and Moore, 1996), and the goal of the learning is to find an optimal policy for the agent that maximizes the cumulative reward when interacting with the environment. Successful applications include playing games
(Mnih et al., 2013; Silver et al., 2016), scheduling traffic signal (Abdulhai, Pringle, and Karakoulas, 2003), regulating ad
bidding (Cai et al., 2017), to name just a few.
In most RL approaches, such as SARSA and Q-learning
(Sutton and Barto, 1998), the model of the environment
is, however, not necessarily known a priori before learning
the optimal policy for the agent. Alternatively, model-based
approaches, such as DYNA (Sutton, 1990) and prioritized
sweeping (Moore and Atkeson, 1993), require establishing
the environment model while learning the optimal policy.
But nonetheless, in either case, the environment is assumed
given and mostly either fixed or non-stationary without a
purposive control (Kaelbling, Littman, and Moore, 1996).
In this paper, we extend the standard RL setting by considering the environment is strategic and controllable. We aim
at learning to design an environment via interacting with an
∗
The work is done during Haifeng Zhang’s visit at UCL. Correspondence to Jun Wang, [email protected]
also learnable agent or multiple agents. This has many potential applications, ranging from designing a game (environment) with a desired level of difficulties in order to fit the
current player’s learning stage (Togelius and Schmidhuber,
2008) and designing shopping space to impulse customers
purchase and long stay (Penn, 2005) to controlling traffic
signals (Ceylan and Bell, 2004).
Our formulation extends the scope of RL by focusing on
the environment modelling and control. We illustrate the
idea by a simple two-agent probabilistic rock-paper-scissors
game. Particularly, in an adversarial case, on the one hand,
the agent aims to minimize its accumulated reward; on the
other hand, the environment tends to minimize the reward
for a given optimal policy from the agent. This effectively
creates a minimax game between the agent and the environment. Given the agent’s playing environment MDP, we, theoretically, find a dual MDP w.r.t. the environment, i.e., how
the environment could decide or sample the successor state
given the agent’s current state and an action taken. Solving
the dual MDP yields a policy gradient solution (Williams,
1992) to optimize the parametric environment achieving its
objective. When the environment’s parameters are not continuous, we propose a generative modeling framework using
reinforcement learning for optimizing the parametric environment, which overcomes the constraints on the environment space. Our experiments on a Maze game generation
task show the effectiveness of generating diverse and challenging Mazes against various types of agents in different
settings. We show that our algorithms would be able to successfully find the weaknesses of the agents and play against
them to generate purposeful environments.
Related Work
Reinforcement learning (RL) (Sutton and Barto, 1998) studies how an intelligent agent learns to take actions through the
interaction with an environment (e.g., the state of a game)
over time. In a typical RL setting, the environment is unknown yet fixed, and the focus is on optimizing the agent
policies. Deep reinforcement learning (DRL) is a marriage
of deep neural networks (LeCun, Bengio, and Hinton, 2015)
and RL; it makes use of deep neural networks as a function
approximator in the decision-making framework of RL to
achieve human-level control and general intelligence (Mnih
et al., 2015). In this paper, instead, we consider a family of
problems that is an extension of RL by considering that the
environment is controllable and strategic. Unlike typical RL,
our subject is the strategic environment not the agent, and the
aim is to learn to design an optimal (game) environment via
the interaction with the intelligent agent.
Technically, our work is related to safe reinforcement
learning, which maximizes the expectation of the return under some safety constraints such as uncertainty (Garcıa and
Fernández, 2015). The uncertainty may be introduced by
two sources: 1) the inherent source from the stochastic state
transition matrix and the stochastic agent policy (Heger,
1994; Gaskett, 2003); and 2) the estimation of the uncertainty from the parameters of MDP (Nilim and El Ghaoui,
2005; Tamar, Xu, and Mannor, 2013; Arjovsky, Chintala,
and Bottou, 2017). Our formulation shares some similarities
with the latter one, due to the used parametric MDPs. However, our problem setting is entirely different from safe RL
as their focus is still on single agent learning in an unknown
environment, whereas our our work is concerned with the
learning of the environment to achieve its own objective.
Our formulation is a general one, applicable in the setting where there are multiple agents (Busoniu and De Schutter) (see our toy example). It is worth mentioning that although multi-agent reinforcement learning (MARL) studies
the strategic interplays among different entities, the game
(either collaborative or competitive) is strictly among multiple agents (Littman, 1994; Hu and Wellman, 2003). By contrast, the strategic interplays in our formulation are between
an agent (or multiple agents) and the environment. The recent work, interactive POMDPs (Gmytrasiewicz and Doshi,
2005), aims to separate beliefs over physical states of the
environment and over models of other agents, but the environment in question is still non-strategic. Our problem, thus,
cannot be formulated directly using MARL as the decision
making of the environment is in an episode-level, while policies of agents typically operate in each time step.
The minimax game formulation can also be found in the
recently emerged generative adversarial nets (GANs), where
a generator and a discriminator play a minimax adversarial
game (Goodfellow et al., 2014). Compared to GANs, our
work addresses a different problem, where the true samples of desired environments are missing in our scenario;
the training of our environment generator is guided by the
behaviours of the agent (corresponding the GAN discriminator) who aims to maximize its cumulative reward in a
given environment. However, the environment generator itself may have a different objective.
In parallel, the design of computer and video games
has been formulated by mechanic, dynamic and aesthetic
(MDA) models (Hunicke, LeBlanc, and Zubek, 2004). For
generating games that conform to design requirements, answer set programming (ASP) is successfully applied (Zook
and Riedl, 2014). Taking optimization goals (such as balancing) into consideration, genetic algorithm (GA) is proposed
as a searcher for optimal game designs (Hom and Marks,
2007). Our work instead formulates the game design problem by extending Markov decision process (MDP) and provides sound solutions for designing video games using reinforcement learning, which shall be explored for future work.
Table 1: Parametrized winning rates of the extended rockpaper-scissors game. The first column and the first row are
the actions of player 1 and 2 respectively. The numbers or
parameters in cells represent winning rates of player 1.
rock paper scissors
rock
0.5
wrp
wrs
paper
wpr
0.5
wps
scissors wsr
wsp
0.5
RL with Controllable Environment
Problem Formulation
Let us first consider the standard reinforcement learning
framework. In this framework there are a learning agent and
a Markov decision process (MDP) M = hS, A, P, R, γi,
where S denotes state space, A action space, P state transition probability function, R reward function and γ discounted factor. The agent interacts with the MDP by taking action a in state s and observe reward r in each timestep, resulting in a trajectory of states, actions and rewards:
H1...∞ = hS1 , A1 , R1 , S2 , A2 , R2 . . .i, St ∈ S, At ∈
A, Rt ∈ R, where P[St+1 = s0 |St = s, At = a] =
P(s, a, s0 ) and E[Rt |St = s, At = a] = R(s, a) hold.1
The agent selects actions according to a policy πφ , where
πφ (a|s) defines the probability that the agent selects action
a in state s. The agent learns
P∞ πφ to maximize the return (cumulative reward) G = t=1 γ t−1 Rt .
In the standard setting, the MDP is given fixed while the
agent is flexible with its policy to achieve its objective. We
extends this setting by also giving flexibility and purpose to
M. Specifically, we parametrize P as Pθ and set the objective of the MDP as O(H), which can be arbitrary based on
the agent’s historical behaviours. We intend to design (generate) a MDP achieves the objective along with the agent
achieving its own objective:
θ∗ = arg max E[O(H)|Mθ = hS, A, Pθ , R, γi;
θ
πφ∗ = arg max E[G|πφ ; Mθ ]].
(1)
πφ
If it is a multi-agent RL setting, the objective function is:
θ∗ = arg max E[O(H)|Mθ = hS, A, Pθ , R, γi;
θ
πφi ∗ = arg max E[G|πφ ; Mθ , πφ−i∗ ], i = 1..n].
(2)
πφ
where πφ−i∗ denotes the joint optimal policy of all the agents
except agent i.
A Toy Example Let us consider a simple extension of the
rock-paper-scissors game to illustrate what the solution of
Eq. (2) looks like. In the Rock-paper-scissors game, two
players select their hand shapes from rock(r), paper(p) and
scissors(s) simultaneously. If a player wins the game, it get
a reward of 1, otherwise 0. Instead of making the winning
rule deterministic: rock > scissors > paper > rock, in
Table 1, we parametrize the game environment by considering certain probabilities for winning. The policy of player
1
In this paper, we use St , At , Rt when they are in a trajectory
while using s, a, r otherwise.
i = 1, 2 is πi (rock) = ari , πi (paper) = api , πi (scissors) =
asi = 1 − ari − api . The objective of the environment is to
make the game as fair as possible, i.e., equalize the expected
rewards of the players:
a1 ,a1
−
p∗
2
max E2 [G|ar2 , ap2 ; ar∗
1 , a1 , w]) ,
ar2 ,ap
2
(3)
p∗
E1 [G|ar1 ,ap1 ; ar∗
2 , a2 , w] =
p∗
s∗
ar1 (ar∗
2 · 0.5 + a2 wrp + a2 wrs )
p∗
s∗
+ap1 (ar∗
2 wpr + a2 · 0.5 + a2 wps )
p∗
E2 [G|ar2 ,ap2 ; ar∗
1 , a1 , w] =
p
r
s
ar∗
1 (a2 · 0.5 + a2 (1 − wrp ) + a2 (1 − wrs ))
p
r
s
+ap∗
1 (a2 (1 − wpr ) + a2 · 0.5 + a2 (1 − wps ))
p
r
s
+as∗
1 (a2 (1 − wsr ) + a2 (1 − wsp ) + a2 · 0.5).
(4)
(5)
Given w, we derive the Nash equilibrium between the two
players by taking partial derivatives of E1 and E2 with regard to corresponding parameters and set them to zero:
∂E1
∂E2
∂E2
∂E1
= 0, p = 0,
= 0, p = 0.
∂ar1
∂a1
∂ar2
∂a2
(6)
The solution is
p∗
p
p∗
p
r
r∗
r
ar∗
1 = f1 (w), a1 = f1 (w), a2 = f2 (w), a2 = f2 (w).
(7)
p∗ r∗
ar∗
1 , a1 , a2
That is to say,
and
can be expressed by
functions of w. For page limit, we omit the detailed form of
the functions. Substituting Eq. (7) into Eq. (4, 5), we get
p∗
r∗ p∗ r∗ p∗
max E1 [G|ar1 , ap1 ; ar∗
2 , a2 , w] = E1 [G|a1 , a1 ; a2 , a2 , w]
ar1 ,ap
1
p∗
max E2 [G|ar2 , ap2 ; ar∗
1 , a1 , w]
ar2 ,ap
2
=
(8)
p∗ r∗ p∗
E2 [G|ar∗
2 , a2 ; a1 , a1 , w]
= f2G (w).
(9)
Thus, the objective function in Eq. (3) can be also expressed
by a function of w. We take the gradient of the objective
function and set it to zero:
2 · ∇(f1G (w) − f2G (w)) = 0.
(10)
Solving this equation, we will get the solution (or solution
set) w∗ which fulfils Eq. (3). This toy example explicitly
illustrates our problem formulation. Specifically, there are
two levels of equilibria. The first one is between the agents,
given a fixed environment, which is solved by Eq. (7). The
ϕ
θ2
Start
G=
ϕ2
G=
G=
ϕ
(agent path evolves)
-28
-40
-42
Train Agent : max G
ϕ1
θ1
G=
G=
ϕn
-22
G=
-
-32
14
G=
-
-40
G=
-38
30
38
m
End
G=
ax
ax
m
min
G
ma
38
=-
30
=xG
14
ma
=xG
start
end
blank
wall
agent path
Figure 1: An example of adversarial maze design. The detailed definition of the maze environment is provided in the
Appendix. In short, an agent tries to find the shortest path
from the start to the end in a given maze map, while the
maze environment tries to design a map to make the path
taken by the agent longer. In the direction of φ, the parameter of an agent policy evolves, whereas in the direction of
θ, the parameter of the maze environment evolves. The cumulative reward G is defined as the opposite number of the
length of path.
second one is between the environment and the agents which
is solved by Eq. (10).
This toy example is simple enough that we can write down
the objective functions of the agents and the environment,
then solve it analytically by taking gradients. However, most
RL environments are more complex that cannot be solved
analytically. We shall discuss how to solve them next.
Adversarial Environment In this paper, we consider a
particular objective of MDP that the MDP acts as an adversarial environment minimizing
return of the
P∞ the expected
t−1
single agent, i.e., O(H) =
Rt = −G. This
t=1 −γ
adversarial objective can be applied to design environments
to analyse the weakness of the agent and its policy learning
algorithms. The objective function is formulated as:
θ∗ = arg min max E[G|πφ ; Mθ = hS, A, Pθ , R, γi].
θ
ap∗
2
= f1G (w),
(maze map evolves)
G=
where w denotes wrp , wrs , wps , wpr , wsr , wsp . Intuitively,
wrs = 1 − wsr , wrp = 1 − wpr , wsp = 1 − wps should be a
sufficient condition of w∗ since it makes the rewards of the
two players symmetrical. Here we briefly derive the solution
analytically to strengthen the understanding of the proposed
problem. We write down the expressions of the expectations
in Eq. (3):
p∗
s∗
+as1 (ar∗
2 wsr + a2 wsp + a2 · 0.5),
θ
θm
p∗
w∗ = arg min ( max
E1 [G|ar1 , ap1 ; ar∗
2 , a2 , w]
r p
w
Train Environment : min max G
φ
(11)
The optimization objective is related to the Nash equilibrium of a zero-sum game between the environment and the
agent, in which the environment decides its parameter θ and
the agent decides its parameter φ simultaneously.
Gradient of Transition Probability
In many cases, Eq. (11) can not be solved analytically. We
propose an approach by which the environment (MDP) and
the agent update their parameters iteratively. In each iteration, the environment updates its parameter by taking a step
in its gradient direction then the agent updates its policy parameter by taking enough steps to be optimal w.r.t. the updated environment, as illustrated by Fig. 1 for learning the
environment of a maze. Since the agent’s policy can be updated using well-studied RL methods, we focus on the update methods for the environment. In each iteration, given
the agent’s policy parameter, the objective of the environment becomes:
θ∗ = arg min E[G|πφ∗ ; Mθ = hS, A, Pθ , R, γi]. (12)
θ
To update environment, we try to find the gradient of the
transition function w.r.t. θ. We derive the gradient by transferring this transition optimization problem to a well-studied
policy optimization problem through a proposed concept of
a duel MDP-policy pair.
Definition 1 (Duel MDP-policy pair). For any MDP-policy
pair hMA , π A i, where MA = hS A , AA , P A , RA , γ A i with
A
start state distribution pA
1 and terminal state set ST , there
E
E
exists a dual MDP-policy pair hM , π i, where ME =
hS E , AE , P E , RE , γ E i with start state distribution pE
1 and
terminal action set AE
satisfying:
T
• S E = S A × AA = {hsA , aA i|sA ∈ S A , aA ∈ AA }, state
in ME corresponds to combination of successive state
and action in MA ;
• AE = S A = {sA |sA ∈ S A }, action in ME corresponds
to state in MA ;
A
A
A
A
• P E (sE
, aE , sE0 ) = P E (hsA
=
j , ak i, s , hsj 0 , ak0 i)
AiA A i A
A
π (ak0 |s ) s = sj 0
, transition in ME depends on
0
sA 6= sA
j0
policy in MA ;
E
E
A A
A
A A A
• RE (sE
i , a ) = R (hsj , ak i, s ) = R (sj , ak ), reE
A
ward in M is the same as in M ;
• γ E = γ A , the discounted factor is the same;
E
E
A A
A A A A A
• pE
1 (s ) = p1 (hs , a i) = p1 (s )π (a |s ), start
E
state distribution in M depends on start state distribution and the first action distribution in MA ;
A A
∈ STA }, terminal action in ME corre• AE
T = {s |s
sponds to terminal state in MA ;
A A
A A A A
• π E (aE |sE ) = π E (sA
i0 |hsi , a i) = P (si , a , si0 ), polE
icy in M corresponds to transition in MA .
We can see that the dual MDP-policy pair in fact describes the same mechanism as the original MDP-policy pair
from another perspective. From the environment’s perspective, the state is the agent’s state-action pair hs, ai and the
action is the next state s0 of the agent; and the policy turns
out to be the next state to transit given the current state and
the agent’s action in the current state.
We give three theorems to derive the gradient of the transition function. The proofs are given in the Appendix.
Theorem 1. For an MDP-policy pair hMA , π A i and its duality hME , π E i, the distribution of trajectory generated by
hMA , π A i is the same as the distribution of a bijective trajectory generated by hME , π E i.
Theorem 2. For an MDP-policy pair hMA , π A i and its
duality hME , π E i, the expected return of two bijective
state-action trajectories, H A from hMA , π A i and H E from
hME , π E i, are equal.
Theorem 3. For an MDP-policy pair hMA , π A i and its duality hME , π E i, the expected return of hMA , π A i is equal
to the expected return of hME , π E i, i.e., E[GA |π A , MA ] =
E[GE |π E , ME ].
Theorem 2 can be understood by the equivalence between
H A and H E and the same generating probability of them as
Environments M θ
A
1. Generate θ
Environment
Generator
µw
4. Agent returns
G1...G 6 guide the
generator update
M θA1
M θA2
M θA3
M θA4
M θA5
M θA6
2. Agent is trained in
each environment
Agent
πϕ
3. Operate in the
environments with
updated policy
Figure 2: Framework dealing with discontinuous transitions.
Generator generates environment parameter θ. For each
MA
θ , the agent policy is trained. Then the policy is tested
in the generated environments and the returns are observed,
which finally guide the generator to update.
given in Theorem 1. Theorem 3 naturally extends Theorem 2
from the single trajectory to the distribution of trajectory according to the same p.m.f. given by Theorem 1.
A
E
E
Now we consider hMA
θ , π i and its duality hM , πθ i,
A
E
where Pθ and πθ are of the same form about θ. Given θ, PθA
and πθE are exactly the same, resulting in E[GA |π A , MA
θ]=
E[GE |πθE , ME ] according to Theorem 3. So optimizing PθA
as Eq. (12) is equivalent to optimizing πθE :
E
E
θ∗ = arg min E[G|πφA∗ ; MA
θ ] = arg min E[G|πθ ; Mφ∗ ].
θ
θ
(13)
We then apply the policy gradient theorem (Sutton et al.,
1999) on πθE and derive the gradient for PθA :
∇θ J(θ) = E[∇θ log πθE (aE |sE )QE (sE , aE )|πθE , ME
φ]
A A
A A
A
A
= E[∇θ log PθA (sA
i , a , si0 )V (si0 )|πφ , Mθ ],
(14)
where J(θ) is cost function, QE (sE , aE ) and V A (sA
)
are
0
i
action-value function and value function of hME , πθE i and
A
hMA
θ , π i respectively; and can be proved equal due to the
equivalence of the two MDPs.
We name the gradient in Eq. (14) as transition gradient.
Transition gradient can be used to update the transition function in an iterative way. In theory, it performs as well as policy gradient since it is equivalent to the policy gradient in the
circumstance of the dual MDP-policy pair.
RL Generator for Discontinuous Transition
Although we have proved the equivalence between the transition optimization and the policy optimization problem in
the above section, two significant characteristics separate the
transition optimization problem from the equivalent policy
optimization problem:
1) In practice, it is hard or impossible to design continuous (differentiable) Pθ subject to the natural environment
constraints, i.e., ensuring Pθ ∈ P for continuous θ where P
may not be continuous. For example, if we want to design a
Maze game, the transition probability between two adjacent
cells is either 0 or 1, depending on whether the target cell
is wall or blank. Also the eligibility of a Maze (e.g. guarantee a valid path) will further make P discretized or pruned.
These constraints result in difficulties in designing continuous Pθ ∈ P. This problem is non-significant in policy optimization problems since the space of a stochastic policy is
naturally continuous.
2) Transition is naturally defined stochastic, while policy is not necessary stochastic. Thus the model of transition
should output a distribution over states while the model of
policy can only decide a choice of action.
These two main differences make it difficult to apply conventional policy-based and value-based methods to the transition optimization problem. Policy-based methods (including policy gradient(Sutton et al., 1999), CMA-ES(Hansen,
Müller, and Koumoutsakos, 2003), etc.) usually assume policy space continuous, so it is assumed transition function is
continuous in our problem. Therefore, the derived transition
gradient method or other policy search methods are not naturally applicable to discontinuous transition. Note that deterministic policy gradient (Silver et al., 2014) may handle the
deterministic transitions, but it is also inapplicable to model
other discontinuous cases. Moreover, value-based methods
cannot naturally apply to most environments because they
do not model probability distribution directly.
To deal with such situation, we propose a generative
framework to find the optimal θ alternative to the gradient
method. In general, we iteratively generate environments
with different θ and test them using the correspondingly
trained agent (illustrated in Fig. 2). Specifically, we generate
environment parameter θ using a generator µw , then optimize w to obtain the (local) optimal w∗ and a corresponding
optimal distribution of θ. Formally, our optimization objective is formulated as
w∗ = arg min Eθ∼µw [E[G|πφ∗ ;
w
A
A
A
A
A
MA
(15)
θ = hS , A , Pθ , R , γ i]].
We model the generation process using an auxiliary MDP
Mµ , i.e., the generator µw generates θ and updates w in a
reinforcement learning way. The reason we adopt reinforcement learning other than supervised learning is that in this
generative task, 1) there is no training data to describe the
distribution of the desired environments so we cannot compute likelihood of generated environments and 2) we can
only evaluate a generated environment through sampling,
i.e., performing agents in the generated environment and get
a score from the trajectory, which can be naturally modeled
by reinforcement learning by viewing the score as a reward
of the actions of the generator.
In detail, the generator µw consists of three elements
µ
hMµ , πw
, f µ i. For generating θ, an auxiliary agent with polµ
icy πw acts in Mµ to generate a trajectory H µ , after that θ is
determined by the transforming function θ = f µ (H µ ), i.e.,
the distribution of θ is given by the distribution of trajectoµ
ries, which are induced by playing πw
in Mµ and transµ
formed through f . For adversarial environments, the reward of the generator is designed to be opposite to the return of the agent got in H µ , which reflects the minimization
objective in Eq. (15). Thus, w can be updated by applying
µ
policy gradient methods on πw
.
There are various ways to designing Mµ . Here we
provide a general design that follows the environment
constraints. We reshape the elements of θ as a vector
θ = hx1 , x2 , . . . , xNθ i, xk ∈ Xk and design Mµ =
hS µ , Aµ , P µ , Rµ , γ µ i to generate θ:
• S µ = {vk = hx1 , x2 , . . . , xk i|k = 0 . . . Nθ , ∃vNθ =
hx1 , x2 , . . . , xk , x0k+1 . . . x0Nθ i = θ, s.t. PθA ∈ PA };
• Aµ =
S
k=1...Nθ
Xk ;
• P µ is defined that for the current state vk =
hx1 , x2 , . . . , xk i and an action xk+1 , if xk+1 ∈ Xk+1 and
vk+1 = hx1 , x2 , . . . , xk+1 i ∈ S µ the next state is vk+1 ,
otherwise vk ;
• Rµ is defined that for terminal state vNθ
=
hx1 , x2 , . . . , xNθ i = θ the reward is the opposite number
of the averaged return got by πφA∗ acting in MA
θ , otherwise the reward is 0;
• γ µ = 1 since there is no intermediate reward unless the
terminal states.
In addition, the start state is v0 = hi and the terminal
states are vNθ = hx1 , x2 , . . . , xNθ i. Corresponding to this
µ
Mµ , πw
(xk+1 |vk ; w) is designed to take an action xk+1 ∈
Xk+1 depending on the previous generated sequence vk ,
and the transforming function f µ is designed as an identical function f µ (vNθ ) = f µ (θ) = θ. Note that due to the
definition of S µ , any partial parameter vt without potential
to be completed as a valid parameter θ s.t. PθA ∈ PA is
avoided to be generated. This ensures any constraint on environment can be followed. On the other hand, any valid θ
µ
is exploratory and of
is probable to be generated once πw
enough expression capacity.
The pseudo code of RL environment generator is provided
in the Appendix.
Experiments with Maze Design
Experiment Setting
In our experiment, we consider a use case of designing
Maze game to test our solutions over the transition gradient
method and the RL generator respectively. As shown in both
Figures 4 and 5, the Maze is a grid world containing a map
of n × n cells. In every time-step, the agent is in a cell and
has four directional actions {N, S, W, E} to select from, and
transitions are made deterministically to an adjacent cell, unless there is a wall (e.g., the black cells as illustrated in Figures 4 and 5), in which case no movement occurs. The minimax game is defined as: the agent should go from the northwest cell to the south-east cell using steps as little as possible, while the goal of the Maze environment is to arrange
the walls in order to maximize the number of steps taken by
the agent. To prevent generating walls that completely block
the agent, the environment generator grows the walls gradually and discards the solutions when there is blocking. The
precise formulation of the Maze game as MDP and the enµ
vironment generation model µw = hMµ , πw
, f µ i are given
in the Appendix.
Note that the above hard wall Maze results in an environment that is discontinuous. In order to also test the case
of continuous environments, we consider a simpler soft wall
Maze as illustrated in Figure 3. Specifically, instead of a hard
wall that blocks the agent completely, each cell except the
end cell has a blockage probability (soft wall) which determines how likely the agent will be blocked in this cell when
it takes transition action from an adjacent cell. In Figure 3,
the blockage probability is indicated by the intensity of the
color in the cell. It is also ensured that the sum of blockage probabilities of all cells is 1 and the maximum blockage
S
S
S
S
S
S
OPT
S
S
S
S
DQN
E
Round 0
E
E
E
E
Round 100
Round 200
Round 300
Round 400
E
Round 0
E
E
E
E
Round 100
Round 200
Round 300
Round 400
Figure 3: Heatmaps of the blockage probability (soft wall) distribution throughout 5 × 5 soft wall Maze learning against the
OPT and DQN agents.
start
end
OPT
blank
wall
agent path
DFS
5×5
Return=16
6×6
Return=22
7×7
Return=30
8×8
Return=38
RHS
5×5
Return=28
6×6
Return=44
7×7
Return=64
8×8
Return=86
5×5
Return=10
6×6
Return=16
7×7
Return=30
8×8
Return=87
DQN
5×5
Return=28
6×6
Return=40
7×7
Return=56
8×8
Return=74
Figure 4: Best mazes against OPT, DFS, RHS and DQN agents with size ranging from 5 × 5 to 8 × 8.
probability for each cell is 0.5. Thus, the task for the adversarial environment in this case is to allocate the soft wall to
each cell to block the agent the most.
The experiment is conducted on PCs with common CPUs.
We implement our experiment environment using Keras-RL
(Plappert, 2016) backed by Keras and Tensorflow. 2
Round 100, ahead of the OPT agent case over 200 rounds.
But then it slows down and it takes around 300 rounds to
achieve the optimal. This is explained by the unnecessary
exploration introduced by the DQN agent in the later stage.
By contrast, the OPT agent only takes 100 rounds from the
nearly optimal (Round 300) to the optimal.
Results for the Transition Gradient Method
Results for Reinforcement Learning Generator
We test the transition gradient method considering the 5 ×
5 soft wall Maze case. We model the transition probability
function by a deep convolutional neural network, which is
updated by the transition gradient following Eq. (14). We
consider the two types of agents: Optimal (OPT) agent and
Deep Q-network learning (DQN) agent. The OPT agent is
not learnable, but always finds the optimal policy against
any generated environment. The DQN agent (Mnih et al.,
2013) is a learnable one, in which the agent’s action-value
function is modeled by a deep neural network, which takes
the whole map and its current position as input, processed
by 3 convolutional layers and 1 dense layer, then outputs the
Q-values over the four directions.
Fig. 3 shows the convergence that our transition gradient method has achieved. The change of the learned environment parameters, in the form of blockage probabilities,
over time are indicated by the color intensity. Intuitively, the
most effective adversarial environment to block the agent is
to place two 0.5 soft walls in the two cells next to the end or
the beginning cell, as this would have the highest blockage
probabilities. We can see that in both cases, using the OPT
agent and the DQN agent, our learning method can obtain
one of the two most optimal Maze environments. Besides,
during the training, we find that if we train the DQN agent to
achieve a reasonable good level then the environment would
converge faster than that of the OPT agent case. This is because the OPT agent always choose the optimal policy, but
the DQN agent may learn it during the play, which would
add the exploration to the environment learning, and bring
more chance to find the optimal learning directions. As evidenced in Fig. 3, the DQN agent case begins to converge
to the two almost optimal adversarial environments from the
We now test our reinforcement learning generator by the
hard wall Maze environment. We follow the proposed genµ
µ
eral framework to design µw = hMµ , πw
, f µ i, where πw
is modeled by a deep neural network. We test our generator against four types of agents each on four sizes of maps
(from 5 × 5 to 8 × 8). Although the objective for every agent
is to minimize the number of steps, not every agent has the
ability to find the optimal policy because of model restrictions of πφ or limitations in the training phase. Therefore,
besides testing our generator against the optimal agent (the
OPT agent) and the DQN agent, we also adopt other two
imperfect agents for our generator to design specific Mazes
in order to understand more about our solution’s behaviors.
They are:
Depth-first search (DFS) agent. The DFS agent searches
the end in a depth-first way. In each time-step, without loss
of generality, the DFS agent is set to select an action according to the priority of East, South, North, West. The DFS
agent takes the highest priority action that leads to a blank
and unvisited cell. If there are none, The DFS agent goes
back to the cell from which it comes.
Right-hand search (RHS) agent. The RHS agent is aware
of the heading direction and follows a strategy that always
ensures that its right-hand cell is a wall or the border. In each
time-step, 1) the RHS agent checks its right-hand cell, if it
is blank, the RHS agent will turn right and step into the cell;
2) if not, then if the front cell is blank, the RHS agent will
step forward; 3) if the front cell is not blank, the RHS agent
will continue turning left until it faces a blank cell, then steps
into that cell.
We also limit the network capacity and training time of
the DQN agent to make it converge differently from the OPT
agent. The learned optimal Mazes are given in Fig. 4 for different agents with different Maze sizes. The strongest Mazes
designed by our generator are found when playing against
2
Our experiment is repeatable and the source code is provided
at https://github.com/pkuzhf/maze-gym.
start
end
blank
wall
OPT
Round 1
Return=14
Round 50
Return=14
Round 100
Return=30
Round 200
Return=32
Round 300
Return=32
Round 400
Return=36
Round 500
Return=36
Round 600
Return=38
Round 1
Return=16
Round 50
Return=58
Round 100
Return=60
Round 150
Return=64
Round 200
Return=66
Round 300
Return=70
Round 400
Return=74
Round 500
Return=86
Round 50
Return=18
Round 100
Return=26
Round 150
Return=36
Round 200
Return=38
Round 300
Return=48
Round 400
Return=58
Round 500
Return=70
Round 600
Return=74
Round 10
Return=19
Round 50
Return=99
Round 75
Return=59
Round 100
Return=16
Round 150
Return=30
Round 200
Return=69
Round 300
Return=87
Round 350
Return=87
DFS
RHS
DQN
Figure 5: Learning to design Mazes against OPT, DFS, RHS and DQN agents in 8 × 8 map.
the OPT agent, shown in Fig. 4 (OPT). We see that in all
cases, from 5 × 5 to 8 × 8, our generator tends to design
long narrow paths without any fork, which makes the optimal paths the longest. By contrast, the generator designs
many forks to trap the DQN agent, shown in Fig. 4 (DQN),
as the DQN agent runs a stochastic policy.
In fact our generator would be able to make use of the
weakness from the agents to design the maps against them.
Fig. 4 (DFS) shows the results that our generator designs
extremely broad areas with only one entrance for the DFS
agent to search exhaustively (visit every cell in the closed
area twice). Fig. 4 (RHS) shows the Mazes generated to trouble the RHS agent the most by creating a highly symmetric
Maze.
Next, Fig. 5 shows the snapshots of the results in different learning rounds.3 They all evolve differently, depending
on the types of the agents. For the OPT agent, we find that
our generator gradually links isolated walls to form a narrow
but long path. For the DFS, our generator gradually encloses
an area then broadens and sweeps it in order to best play
against the policy that has the priority order of their travel
directions. Fig. 5 (RHS) shows that our generator learns to
adjust the wall into zigzag shapes to trouble the RHS agent.
For the DQN agent, with limited network capacity or limited
training time, it is usually the case that it cannot perfectly tell
which road to go during the learning. As such, the generator
tends to generate many forks to confuse the DQN agent.
Furthermore, Fig. 6 shows the process of training our generator against the four agents in 8 × 8 map. We find that for
OPT, DFS and RHS agents, the generator learns rapidly at
first and gradually converges. But for the DQN agent, the
learning curve is tortuous, which illustrates the more adversarial process. This is because the ability of the DQN
agent is gradually improved so it does not accurately and efficiently guide the learning of the generator. Also when the
ability of the DQN agent improves more than the progress
3
A series of video demos that show how the environments
evolve can be found at https://goo.gl/iXbDct.
(1) OPT
(2) DFS
(3) RHS
(4) DQN
Figure 6: Training curves for OPT, DFS, RHS and DQN
agents in 8 × 8 map.
of the generator, the learning curve for the generator may
change its direction for several rounds.
Conclusions
In this paper, we presented an extension of standard reinforcement learning by considering that the environment is
strategic and can be learned based on the interaction with
the agent(s) and the feedback about its performance. We derived a gradient update by introducing a dual MDP-policy
pair for the cases where the transition probability is continuous. To deal with non-continuous transitions, we proposed
a novel generative framework using reinforcement learning.
We evaluated the effectiveness of our solution by considering designing a Maze game. Our experiments showed that
our method can make use of the weaknesses of agents to
learn the environment effectively.
In the future, we plan to apply the proposed method to
practical environment design tasks, such as video game de-
sign (Hom and Marks, 2007), shopping space design (Penn,
2005) and bots routine planning.
References
Abdulhai, B.; Pringle, R.; and Karakoulas, G. J. 2003. Reinforcement learning for true adaptive traffic signal control.
Journal of Transportation Engineering 129(3):278–285.
Arjovsky, M.; Chintala, S.; and Bottou, L. 2017. Wasserstein
gan. arXiv preprint arXiv:1701.07875.
Busoniu, L., and De Schutter, B. A comprehensive survey
of multiagent reinforcement learning.
Cai, H.; Ren, K.; Zhag, W.; Malialis, K.; and Wang, J. 2017.
Real-time bidding by reinforcement learning in display
advertising. In The Tenth ACM International Conference
on Web Search and Data Mining (WSDM). ACM.
Ceylan, H., and Bell, M. G. 2004. Traffic signal timing optimisation based on genetic algorithm approach, including
drivers routing. Transportation Research Part B: Methodological 38(4):329–342.
Garcıa, J., and Fernández, F. 2015. A comprehensive survey on safe reinforcement learning. Journal of Machine
Learning Research 16(1):1437–1480.
Gaskett, C. 2003. Reinforcement learning under circumstances beyond its control.
Gmytrasiewicz, P. J., and Doshi, P. 2005. A framework for
sequential planning in multi-agent settings. J. Artif. Intell.
Res.(JAIR) 24:49–79.
Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.;
Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y.
2014. Generative adversarial nets. In Advances in Neural
Information Processing Systems, 2672–2680.
Hansen, N.; Müller, S. D.; and Koumoutsakos, P. 2003. Reducing the time complexity of the derandomized evolution strategy with covariance matrix adaptation (cma-es).
Evolutionary computation 11(1):1–18.
Heger, M. 1994. Consideration of risk in reinforcement
learning. In Proceedings of the Eleventh International
Conference on Machine Learning, 105–111.
Hom, V., and Marks, J. 2007. Automatic design of balanced
board games. In Proceedings of the AAAI Conference on
Artificial Intelligence and Interactive Digital Entertainment (AIIDE), 25–30.
Hu, J., and Wellman, M. P. 2003. Nash q-learning for
general-sum stochastic games. Journal of Machine learning research 4(Nov):1039–1069.
Hunicke, R.; LeBlanc, M.; and Zubek, R. 2004. Mda: A formal approach to game design and game research. In Proceedings of the AAAI Workshop on Challenges in Game
AI, volume 4, 1722.
Kaelbling, L. P.; Littman, M. L.; and Moore, A. W. 1996.
Reinforcement learning: A survey. Journal of artificial
intelligence research 4:237–285.
LeCun, Y.; Bengio, Y.; and Hinton, G. 2015. Deep learning.
Nature 521(7553):436–444.
Littman, M. L. 1994. Markov games as a framework for
multi-agent reinforcement learning. In Proceedings of the
eleventh international conference on machine learning,
volume 157, 157–163.
Mnih, V.; Kavukcuoglu, K.; Silver, D.; Graves, A.;
Antonoglou, I.; Wierstra, D.; and Riedmiller, M. 2013.
Playing atari with deep reinforcement learning. arXiv
preprint arXiv:1312.5602.
Mnih, V.; Kavukcuoglu, K.; Silver, D.; Rusu, A. A.; Veness, J.; Bellemare, M. G.; Graves, A.; Riedmiller, M.;
Fidjeland, A. K.; Ostrovski, G.; et al. 2015. Humanlevel control through deep reinforcement learning. Nature
518(7540):529–533.
Moore, A. W., and Atkeson, C. G. 1993. Prioritized sweeping: Reinforcement learning with less data and less time.
Machine learning 13(1):103–130.
Nilim, A., and El Ghaoui, L. 2005. Robust control of markov
decision processes with uncertain transition matrices. Operations Research 53(5):780–798.
Penn, A. 2005. The complexity of the elementary interface:
shopping space. In Proceedings to the 5th International
Space Syntax Symposium, volume 1, 25–42. Akkelies van
Nes.
Plappert, M. 2016. keras-rl. https://github.com/
matthiasplappert/keras-rl.
Silver, D.; Lever, G.; Heess, N.; Degris, T.; Wierstra, D.;
and Riedmiller, M. 2014. Deterministic policy gradient
algorithms. In Proceedings of the 31st International Conference on Machine Learning (ICML-14), 387–395.
Silver, D.; Huang, A.; Maddison, C. J.; Guez, A.; Sifre, L.;
Van Den Driessche, G.; Schrittwieser, J.; Antonoglou, I.;
Panneershelvam, V.; Lanctot, M.; et al. 2016. Mastering
the game of go with deep neural networks and tree search.
Nature 529(7587):484–489.
Sutton, R. S., and Barto, A. G. 1998. Reinforcement learning: An introduction, volume 1. MIT press Cambridge.
Sutton, R. S.; McAllester, D. A.; Singh, S. P.; Mansour,
Y.; et al. 1999. Policy gradient methods for reinforcement learning with function approximation. In NIPS, volume 99, 1057–1063.
Sutton, R. S. 1990. Integrated architectures for learning,
planning, and reacting based on approximating dynamic
programming. In Proceedings of the seventh international
conference on machine learning, 216–224.
Tamar, A.; Xu, H.; and Mannor, S. 2013. Scaling up
robust mdps by reinforcement learning. arXiv preprint
arXiv:1306.6189.
Togelius, J., and Schmidhuber, J. 2008. An experiment in
automatic game design. In Computational Intelligence
and Games, 2008. CIG’08. IEEE Symposium On, 111–
118. IEEE.
Williams, R. J. 1992. Simple statistical gradient-following
algorithms for connectionist reinforcement learning. Machine learning 8(3-4):229–256.
Zook, A., and Riedl, M. O. 2014. Automatic game design
via mechanic generation. In AAAI, 530–537.
Appendix
The Definition of the Proposed Maze Game
Formally, a maze has a map defined as mi,j ∈
{BLANK, WALL}, i = 1 . . . n, j = 1 . . . n, where m1,1 and
mn,n are always BLANK as being the start cell and the end
cell respectively. Thus, the agent acts in this maze should
find a policy that walk from the start cell to the end cell using
steps as few as possible. On the other hand, our objective is
to design a maze that maximizes the steps against the agent.
MDP MM = hS M , AM , P M , RM , γ M i describes the rule
of the maze with map m:
• state set S M = {hpx , py i|px , py ∈ Z, 1 ≤ px , py ≤ n},
where hpx , py i denotes the current position in maze;
• action set AM = {h1, 0i, h−1, 0i, h0, −1i, h0, 1i}, which
represents the four directions North, South, West and
East;
• P M is defined that for state s = hpx , py i and action
a = hdx , dy i, the next state s0 = hpx + dx , py + dy i if
hpx + dx , py + dy i is inside the map and mpx +dx ,py +dy =
BLANK, otherwise s0 = s;
• RM is defined that for every step, the agent gets a reward
−1;
• γ M = 1.
The start state h1, 1i and the terminal state is hn, ni. The
map is guaranteed to include a path from the the start cell to
the end cell.
The Definition of the Proposed Generative Model
for Maze Games
As introduced in Sec. 3.3, our framework for solving undifµ
, f µ i.
ferentiable transition includes three elements hMµ , πw
With respect to the maze problem, we firstly define Mµ as
follows:
• S µ consists of all valid maps m, where mi,j ∈
{BLANK, WALL}, i, j ∈ Z, 1 ≤ i, j ≤ n, m1,1 =
BLANK, mn,n = BLANK and there exists a path from
m1,1 to mn,n ;
• Aµ consists of n × n + 1 actions, including n × n twodimensional coordinate hx, yi, x, y ∈ Z, 1 ≤ x, y ≤ n
and an action END;
• P µ is defined that for current state m and an action a, the
next state m0 satisfies:
– m0 = m and the MDP terminates, for a = END, or
a = hx, yi and mx,y = WALL, or setting mx,y =
WALL will isolate m1,1 from mn,n in m;
mi,j
hi, ji =
6 hx, yi
0
– otherwise mi,j =
WALL hi, ji = hx, yi
µ
• R is defined that if the MDP terminates the reward is the
opposite number to the averaged return got by the target
agent acting in MM which is determined by the terminal
state m, otherwise the reward is 0;
• γ µ = 1;
• the start state is m0 that m0i,j = BLANK, i, j ∈ Z, 1 ≤
i, j ≤ n.
µ
Corresponding to Mµ , πw
takes the half-generated map
m as its input and outputs the possibilities of taking the
µ
n × n + 1 actions. πw
is modelled by a deep neural network
consists of three convolutional layers each with 32 filters and
one dense layer.4 The transformation function f µ transforms
m to P M in the same way as described in the definition of
maze above.
Proofs for deriving Transition Gradient
Theorem 1. For an MDP-policy pair hMA , π A i and its duality hME , π E i, the distribution of trajectory generated by
hMA , π A i is the same as the distribution of a bijective trajectory generated by hME , π E i.
Proof. For any state-action trajectory H A
=
hS1 , A1 , S2 , A2 , . . . , ST −1 , AT −1 , ST i
generated
by hMA , π A i, there exists a bijective state-action
trajectory H E generated by hME , π E i: H E
=
hhS1 , A1 i, S2 , hS2 , A2 i, S3 , . . . , hST −1 , AT −1 i, ST i,
whose generating probability is equal to H A :
P[H E |hME , π E i]
(16)
=pE
1 (hS1 , A1 i)
·
−2
TY
π E (St+1 |hSt , At i)P E (hSt , At i, St+1 , hSt+1 , At+1 i)
t=1
· π E (ST |hST −1 , AT −1 i)
−2
TY
A
P A (St , At , St+1 )π A (At+1 |St+1 )
=pA
(S
)π
(A
|S
)
1
1 1
1
t=1
A
· P (ST −1 , AT −1 , ST )
=pA
1 (S1 )
TY
−1
π A (At |St )P A (St , At , St+1 )
t=1
=P[H A |hMA , π A i].
Theorem 2. For an MDP-policy pair hMA , π A i and its
duality hME , π E i, the expected return of two bijective
state-action trajectories, H A from hMA , π A i and H E from
hME , π E i, are equal.
Proof.
E[GE |H E ] =
T
−1
X
(γ E )t−1 RE (hSt , At i, St+1 )
t=1
=
T
−1
X
(γ A )t−1 RA (St , At ) = E[GA |H A ].
t=1
Theorem 3. For an MDP-policy pair hMA , π A i and its duality hME , π E i, the expected return of hMA , π A i is equal
to the expected return of hME , π E i, i.e. E[GA |π A , MA ] =
E[GE |π E , ME ].
4
More details are provided in
https://github.com/pkuzhf/maze-gym.
the
source
code
at
Algorithm 1 Training Adversarial Environment
1: function T RAIN(w, φ) . Where w is parameters of µw ,
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
φ is parameters of πφ
Initialize w, φ
while not converge do
for Env-steps do
µ
play πw
in Mµ to generate a trajectories H µ
get an instance of environment parameter using the transforming function θ = f µ (H µ )
play πφ in MA
θ for several times and get the
averaged return (cumulative reward) G
µ
give −G to πw
as its reward of the generating process and update w using RL methods
end for
for Agent-steps do
µ
use µw = hMµ , πw
, f µ i to generate an inA
stance of environment Mθ as Env-steps do
play πφ in MA
θ and update φ using RL methods
end for
end while
end function
Proof.
E[GE |π E , ME ] =
X
=
X
P[H E |hME , π E i]E[GE |H E ]
HE
P[H A |hMA , π A i]E[GA |H A ]
HA
= E[GA |π A , MA ]
Pseudo code of Environment Generator by RL
Here we provide the pseudo code for the training process
of the proposed RL environment generator. We mainly consider the case that the agent is also learnable (such as DQN
agent). In general, we alternatively train the environment
µ
, f µ i and the agent policy πφ .
generator µw = hMµ , πw
When training environment, the cumulative reward is obtained by playing the agent in the generated environments.
When training agent, it is a regular RL process. The details
are shown in Algorithm 1.
For the cases the agent is fixed (such as OPT, DFS and
RHS agents), we simply remove the Agent-steps in Algorithm 1.
| 2cs.AI
|
Submitted to the Annals of Statistics
arXiv:1709.05603v1 [math.ST] 17 Sep 2017
A SHARP LOWER BOUND FOR MIXED-MEMBERSHIP
ESTIMATION
By Jiashun Jin∗ and Zheng Tracy Ke†
Carnegie Mellon University∗ and University of Chicago†
Consider an undirected network with n nodes and K perceivable
communities, where some nodes may have mixed memberships. We
assume that for each node 1 ≤ i ≤ n, there is a probability mass
function πi defined over {1, 2, . . . , K} such that
πi (k) = the weight of node i on community k,
1 ≤ k ≤ K.
The goal is to estimate {πi , 1 ≤ i ≤ n} (i.e., membership estimation).
We model the network with the degree-corrected mixed membership (DCMM) model [8]. Since for many natural networks, the degrees
have an approximate power-law tail, we allow severe degree heterogeneity in our model.
For any membership estimation {π̂i , 1 ≤ i ≤ n}, since each πi is a
probability mass function, it is natural to measure the errors by the
average ℓ1 -norm
n
1X
kπ̂i − πi k1 .
n i=1
We also consider a variant of the ℓ1 -loss, where each kπ̂i − πi k1 is reweighted by the degree parameter θi in DCMM (to be introduced).
We present a sharp lower bound. We also show that such a lower
bound is achievable under a broad situation. More discussion in this
vein is continued in our forthcoming manuscript [7].
The results are very different from those on community detection.
For community detection, the focus is on the special case where all
πi are degenerate; the goal is clustering, so Hamming distance is the
natural choice of loss function, and the rate can be exponentially fast.
The setting here is broader and more difficult: it is more natural to
use the ℓ1 -loss, and the rate is only polynomially fast.
1. Introduction. Consider an undirected network N = (V, E), where
V = {1, 2, . . . , n} is the set of nodes and E is the set of (undirected) edges.
Let A ∈ Rn,n be the adjacency matrix where
1,
if nodes i and j have an edge,
A(i, j) =
1 ≤ i, j ≤ n,
0,
otherwise.
The diagonals of A are zero since we do not allow for self-edges. Suppose
the network has K perceivable communities (i.e., clusters)
C1 , C2 , . . . , CK ,
1
2
and a node may belong to more than one cluster (i.e., mixed memberships).
For each node 1 ≤ i ≤ n, suppose there exists a Probability Mass Function
(PMF) πi = (πi (1), πi (2), . . . , πi (K))′ ∈ RK such that
πi (k) is the “weight” of node i on Ck ,
1 ≤ k ≤ K.
We call node i “pure” if πi is degenerate (i.e., one entry is 1 and the other
entries are 0) and “mixed” otherwise. The primary interest is to estimate
πi , 1 ≤ i ≤ n.
Estimating mixed memberships is a problem of great interest in social
network analysis [1, 2, 8, 16]. Take the Polbook network [13] for example.
Each node is a book on US politics for sale in Amazon.com, and there is
an edge between two nodes if they are frequently co-purchased. Jin et al.
(2017) [8] modeled this network with a two-community (“Conservative” and
“Liberal”) mixed membership model, where the estimated mixed membership of a node describes how much weight this book puts on “Conservative”
and “Liberal”.
We are interested in the optimal rate of convergence associated with membership estimation. Below, we introduce a model and present a sharp lower
bound. We show that the lower bound is achievable in a broad class of
situations where we allow severe degree heterogeneity.
1.1. Model. Consider the degree-corrected mixed membership (DCMM)
model [8]. Recall that A is the adjacency matrix. DCMM assumes that
(1.1)
{A(i, j) : 1 ≤ i < j ≤ n} are independent Bernoulli variables,
where the Bernoulli parameters are different. For a symmetric non-negative
matrix P ∈ RK,K and a vector θ = (θ1 , θ2 , . . . , θn )′ , where θi > 0 is the
degree heterogeneity parameter of node i, DCMM models
(1.2)
P A(i, j) = 1 = θi θj · πi′ P πj ,
1 ≤ i < j ≤ n.
To ensure model identifiability, we assume
(1.3)
P is non-singular and have unit diagonals.
We now calibrate DCMM with a matrix form. Introduce the two matrices
Θ = diag(θ1 , θ2 , . . . , θn ) ∈ Rn,n and Π = [π1 , π2 , . . . , πn ]′ ∈ Rn,K . Then,
W ,
A = [Ω − diag(Ω)] + |{z}
{z
}
|
“signal”
“noise”
Ω = ΘΠP Π′ Θ,
W = A − E[A].
3
Here Ω is a low-rank matrix (rank(Ω) = K) containing Bernoulli parameters
and W is a generalized Wigner matrix.
DCMM can be viewed as extending the mixed membership stochastic block
model (MMSB) [1] to accommodate degree heterogeneity, and can also be
viewed as extending the degree-corrected block model (DCBM) [11] to accommodate mixed memberships. DCMM is similar to the overlapping continuous
community assignment model (OCCAM) [16], where the difference is that
DCMM regards each membership vector πi as a PMF with a unit ℓ1 -norm
while OCCAM models that each πi has a unit ℓ2 -norm (which seems hard
to interpret).
Remark. The identifiability condition of DCMM is different from that of
DCBM. In DCBM, even when P is singular, the model can still be identifiable. However, in DCMM, since there are many more free parameters, the
full rank assumption of P is required for identifiability.
An example. Let’s look at an example with K = 2 and
a b
P =
.
b c
If nodes i and j are both pure nodes, then there are three cases:
a, i, j ∈ C1 ,
P A(i, j) = 1 = θi θj c, i, j ∈ C2 ,
b, i ∈ C1 , j ∈ C2 or i ∈ C2 , j ∈ C1 .
As a result, in the special case with all nodes being pure, the “signal” matrix
Ω has the form
1 0
θ1
θ1
0 1
a b 1 0 ··· 1
..
..
Ω=
,
.. ..
.
.
. . b c 0 1 · · · 0
θn
θn
1 0
{z
}
|
ΠP Π′
where the matrix ΠP Π′ can be shuffled to a block-wise
some unknown permutation:
a a a
a b a b a
b c b c b
a a a
permute
ΠP Π′ =
a b a b a −→ a a a
b c b c b
b b b
a b a b a
b b b
constant matrix by
b
b
b
c
c
b
b
b
c
c
.
4
In general cases where the nodes have mixed memberships, Ω has a similar
form, except that Π is no longer a matrix of 0’s and 1’s and ΠP Π′ can no
longer be shuffled to a block-wise constant matrix.
0.8 0.2
θ1
θ1
0
1 a b 0.8 0 · · · 0.7
..
..
Ω=
.
..
.. b c 0.2 1 · · · 0.3
.
.
.
.
θn
θn
0.7 0.3
{z
}
|
ΠP Π′
1.2. Loss functions. Given estimators Π̂ = [π̂1 , π̂2 , . . . , π̂n ]′ , since each πi
is a PMF, it is natural to measure the (unweighted) ℓ1 -estimation error:
(1.4)
H(Π̂, Π) = n−1
n
X
i=1
kπ̂i − πi k1 .
We also consider a variant of the ℓ1 -errorP
where kπ̂i − πi k1 is reweighed by
n
−1
the degree parameter θi . Write θ̄ = n
i=1 θi , θmax = max1≤i≤n θi , and
θmin = min1≤i≤n θi . Define the degree-weighted ℓ1 -estimation error as
(1.5)
L(Π̂, Π) = n−1
n
X
i=1
(θi /θ̄)1/2 kπ̂i − πi k1 .
When θmax /θmin is bounded, the above loss functions are equivalent in
the sense that for a constant C > 1,
C −1 H(Π̂, Π) ≤ L(Π̂, Π) ≤ CH(Π̂, Π).
However, when there is severe degree heterogeneity (i.e., θmax /θmin ≫ 1), the
weighted ℓ1 -loss is more convenient to use: The minimax rate for H(Π̂, Π)
depends on all θi in a complicated form, but the minimax rate of L(Π̂, Π) is
a simple function of θ̄.
Remark. The weights in (1.5) are motivated by the study of an oracle
case where all true parameters of DCMM are known except for πi of one
node i. In this case, there exists an oracle estimator π̂i0 such that
kπ̂i0 − πi k1 = (θi /θ̄)−1/2 · O((nθ̄ 2 )−1/2 )
with high probability. It motivates us to re-weight kπ̂i − πi k by (θi /θ̄)1/2 .
5
1.3. Lower bound. In the asymptotic analysis, we fix K and P ∈ RK,K
and let (Θ, Π) change with n. Our results take the form: For each θ in a
broad class Q∗n (K, c) (to be introduced), we provide a minimax lower bound
associated with a class of Π.
Given θ ∈ Rn+ , let θ(1) ≤ θ(2) ≤ . . . ≤ θ(n) be the sorted values of θi ’s. For
a constant c ∈ (0, 1/K), introduce
(1.6)
Q∗n (K, c) = θ ∈ Rn+ : θ̄ ≥ n−1/2 log(n), θ(cKn) ≥ n−1/2 log(n)
Denote by Gn (K) the set of all matrices Π ∈ Rn,K such that each row πi
is a PMF. Given Π ∈ Gn (K), let
Nk = {1 ≤ i ≤ n : πi = ek },
M = {1, 2, . . . , n} \ (N1 ∪ . . . ∪ NK ),
where e1 , e2 , . . . , eK are the standard bases of RK . It is seen that Nk is the
set of pure nodes of community k and M is the set of mixed nodes. Fix
(K, c) and an integer L0 ≥ 1. Introduce
n
Gen (K, c, L0 ; θ) = Π ∈ Gn (K) :
(1.7)
|N
for 1 ≤ k ≤ K;
P k | ≥ cn,
2 ≥ ckθk2 , for 1 ≤ k ≤ K;
θ
i∈Nk i
there is L ≤ L0 , a partition M = ∪L
ℓ=1 Mℓ , PMF’s γ1 , ..., γL ,
where minj6=ℓ kγj − γℓ k ≥ c, min1≤ℓ≤L,1≤k≤K kγℓ − ek k ≥ c, o
such that |Mℓ | ≥ c|M| ≥
log3 (n)
, maxi∈Mℓ
θ̄ 2
kπi − γℓ k ≤
1
log(n)
.
Theorem 1.1 (Lower bound of the weighted ℓ1 -error). Fix K ≥ 2, c ∈
(0, 1/K), and a nonnegative symmetric matrix P ∈ RK,K that satisfies (1.3).
Suppose the DCMM model (1.1)-(1.2) holds. As n → ∞, there are constants
C0 > 0 and δ0 ∈ (0, 1) such that, for any θ ∈ Q∗n (K, c),
C0
√
≥ δ0 .
sup
P L(Π̂, Π) ≥
inf
Π̂ Π∈Gen (K,c,L0 ;θ)
nθ̄ 2
When θmax ≤ Cθmin , the unweighted and weighted ℓ1 -errors are equivalent, and we also have a lower bound for the unweighted ℓ1 -error:
Corollary 1.1 (Lower bound of the unweighted ℓ1 -error). Suppose the
conditions of Theorem 1.1 hold. As n → ∞, there are constants C1 > 0 and
δ0 ∈ (0, 1) such that, for any θ ∈ Q∗n (K, c) satisfying θmax ≤ Cθmin,
C1
sup
P H(Π̂, Π) ≥ √
inf
≥ δ0 .
Π̂ Π∈Gen (K,c,L0 ;θ)
nθ̄ 2
6
Remark. Our results allow for severe degree heterogeneity: for θ ∈ Q∗n (K, c),
it is possible that θmax /θmin ≫ 1. In addition, we allow for sparse networks
because θ ∈ Q∗n (K, c) only requires that the average node degree grows with
n in a logarithmic rate.
Remark. The lower bounds here are different from those on community
detection [15, 3]. For community detection, the focus is on the special case
where all πi are degenerate; the goal is clustering, so Hamming distance is
the natural choice of loss function, and the rate can be exponentially fast.
The setting here is broader and more difficult: it is more natural to use the
ℓ1 -loss, and the rate is only polynomially fast.
1.4. Achievability. Jin et al. [8] proposed a method Mixed-SCORE for
estimating πi ’s. The Mixed-SCORE is a fast and easy-to-use spectral approach, and can be viewed as an extension of Jin’s SCORE [5, 4, 9]. However,
SCORE is originally designed for community detection, and to extend it to
membership estimation, we need several innovations; see [5, 8] for details. It
turns out that Mixed-SCORE is also rate-optimal.
The following theorem follows directly form Theorem 1.2 of [8]:
Theorem 1.2 (Upper bound). Fix K ≥ 2, c ∈ (0, 1/K), and a nonnegative symmetric irreducible matrix P ∈ RK,K that satisfies (1.3). Suppose the
DCMM model (1.1)-(1.2) holds. Let Π̂ be the Mixed-SCORE estimator. As
n → ∞, for any θ ∈ Q∗n (K, c) with θmax ≤ Cθmin and any Π ∈ Gen (K, c, L0 ),
with probability 1 − o(n−3 ),
L(Π̂, Π) ≤ CH(Π̂, Π) ≤
C log(n)
√
.
nθ̄ 2
In the case that θmax ≤ Cθmin , the upper bound and lower bound have
matched, and the minimax rate of convergence for both weighted and unweighted ℓ1 -errors is
(nθ̄ 2 )−1/2 ,
up to a multiple-log(n) factor.
For more general settings where θmax /θmin is unbounded, in a forthcoming
manuscript Jin and Ke [7], we demonstrate that
• The minimax rate of convergence for the weighted ℓ1 -loss L(Π̂, Π) is
still (nθ̄ 2 )−1/2 , up to a multiple-log(n) factor.
• The minimax rate of convergence for the unweighted ℓ1 -loss H(Π̂, Π)
depends on individual θi ’s in a more complicated form.
• Mixed-SCORE achieves the minimax rate for a broad range of settings.
7
At the heart of the upper bound arguments is some new node-wise large
deviation bounds we derived; see our forthcoming manuscript [7]. On a high
level, the technique is connected to the post-PCA entry-wise bounds in Jin
et al. [6] and Ke and Wang [12], but is for very different settings. The main
interest of [6] is on gene microarray analysis, where we discuss three interconnected problems: subject clustering, signal recovery, and global testing;
see also Jin and Wang [10] on IF-PCA. The main interest of [12] is on topic
estimation in text mining.
As far as we know, Jin et al. [6] is the first paper that has carefully
studied post-PCA entry-wise bounds. The bounds are crucial for obtaining
sharp bounds on the clustering errors by PCA approaches.
2. Proof of Theorem 1.1. We introduce a subset of Gen (K, c, L0 ; θ):
Gn∗ (K, c; θ) = Π ∈ Gn (K) : P
|Nk | ≥ cn, for 1 ≤ k ≤ K;
2
2
i∈Nk θi ≥ ckθk , for 1 ≤ k ≤ K;
kπi − (1/K)1K k ≤ 1/ log(n), for i ∈ M .
Since Gn∗ (K, c; θ) ⊂ Gen (K, c, L0 ; θ), for any estimator Π̂,
C0
C0
sup
P L(Π̂, Π) ≥ √
≥
sup
P L(Π̂, Π) ≥ √
.
∗ (K,c;θ)
nθ̄ 2
nθ̄ 2
Π∈Gn
Π∈Gen (K,c,L0 ;θ)
Hence, it suffices to prove the lower bound for Π ∈ Gn∗ (K, c; θ).
We need the following lemma, which is adapted from Theorem 2.5 of [14].
We recall that Gn (K) is the set of all matrices Π ∈ Rn,K each row of which
is a PMF in RK .
Gn∗
Lemma 2.1.
such that:
For any subset Gn∗ ⊂ Gn (K), if there exist Π(0) , Π(1) , . . . , Π(J) ∈
(i) L(Π(j)
, Π(k) ) ≥ 2C0 sn for all 0 ≤ j 6= k ≤ J,
1 PJ
(ii) J+1 j=0 KL(Pj , P0 ) ≤ β log(J),
where C0 > 0, β ∈ (0, 1/8), Pj denotes the probability measure associated
with Π(j) , and KL(·, ·) is the Kullback-Leibler divergence, then
q
√
2β
inf sup P L(Π̂, Π) ≥ C0 sn ≥ 1+√J J 1 − 2β − log(J)
.
∗
Π̂ Π∈Gn
As long as J → ∞ as n → ∞, the right hand side is lower bounded by a
constant.
By Lemma 2.1, it suffices to find Π(0) , Π(1) . . . , Π(J) ∈ Gn∗ (K, c) that satisfy
the requirement of Lemma 2.1. Below, we first consider the case K = 2 and
then generalize the proofs to K ≥ 3.
8
2.1. The case K = 2. We re-parametrize the model by defining a ∈ (0, 1]
and γ = (γ1 , . . . , γn ) ∈ [−1, 1]n through
1 + γ 1 − γ ′
1
1−a
i
i
,
1 ≤ i ≤ n.
,
(2.8) P =
,
πi =
1−a
1
2
2
Since there is a one-to-one mapping between Π and γ, we instead construct
γ (0) , γ (1) , . . . , γ (n) . Without loss of generality, we assume θ1 ≥ θ2 ≥ . . . ≥ θn .
Let n1 = ⌊cn⌋ and n0 = n − 2n1 . Introduce
′
γ ∗ = 0, 0, · · · , 0, 1, 1, · · · , 1, −1, −1, · · · , −1 .
| {z } | {z } |
{z
}
n0
n1
n1
Note that γi∗ ∈ {±1} implies that node i is a pure node and γi∗ = 0 indicates that πi∗ = (1/2, 1/2). From the Varshamov-Gilbert bound for packing
(0)
(1)
(J )
numbers [14, Lemma 2.9], there exist J0 ≥ 2n0 /8 and ω∗ , ω∗ , . . . , ω∗ 0 ∈
(ℓ)
(j)
(0)
{0, 1}n0 such that ω∗ = (0, 0, . . . , 0)′ and kω∗ − ω∗ k1 ≥ n0 /8, for all
(ℓ)
(0)
0 ≤ j 6= ℓ ≤ J0 . Let J = 2J0 , ω (0) = ω∗ , and ω (2ℓ±1) = ±ω∗ for 1 ≤ ℓ ≤ J0 .
(0)
(1)
(J)
Then, the resulting ω , ω , . . . , ω
satisfy that:
(a) min0≤j6=ℓ≤J kω (j) − ω (ℓ) k1 ≥ n0 /8.
P 0
P
(ℓ)
(b) For any real sequence {hi }ni=1 , Jℓ=0 ni=1
hi ωi = 0.
For a properly small constant c0 > 0 to be determined, letting δn = c0 (nθ̄)−1/2 ,
we construct γ (0) , γ (1) , . . . , γ (J) by
(2.9)
1
1
1
.
γ (ℓ) = γ ∗ + δn v ◦ ω (ℓ) , 0, 0, . . . , 0 , with v = √ , √ , . . . , p
| {z }
θ1
θ2
θ n0
2n1
We then use the one-to-one mapping (2.8) to obtain Π(0) , Π(1) , . . . , Π(J) . To
(ℓ)
check that each Π(ℓ) belongs to Gn∗ (K, c; θ), we notice that kπi −(1/2, 1/2)k =
−1/2
−1/2
O(θi
δn ) = O(θn0 δn ) for 1 ≤ i ≤ n0 ; θn0 is the (2cn)-smallest value of
θ1 , . . . , θn and it satisfies that θn0 ≥ n1/2 log(n); additionally, θ̄ ≥ n1/2 log(n);
(ℓ)
it follows that kπi − (1/2, 1/2)k = O(c0 / log(n)); hence, Π(ℓ) ∈ Gn∗ (K, c; θ)
as long as c0 is appropriately small.
What remains is to show that the requirements (i)-(ii) in Lemma 2.1 are
satisfied for sn = (nθ̄ 2 )−1/2 . Consider (i). Note that for any 0 ≤ j 6= ℓ ≤ J,
L(Π(j) , Π(ℓ) ) = min
±
n p
n 1 X
o
(j)
(ℓ)
√
θi |γi ± γi | .
n θ̄ i=1
9
P 0 −1/2
−1/2
For “−”, the term in the brackets is at most δn ni=1
θi
≤ n0 δn θn0 =
o(n); for “+”, this term is at least 4n1 ≥ 4cn. Therefore, the minimum is
achieved at “−”. Furthermore, we have
(j)
L(Π
n
δn
n0 δn
1 Xp
(j)
(ℓ)
θi |γi − γi | = √ kω (j) − ω (ℓ) k1 ≥ √ ,
,Π ) = √
n θ̄ i=1
n θ̄
8n θ̄
(ℓ)
where the last inequality is due to Property (a) of ω (0) , . . . , ω (J) . Since δn =
c0 (nθ̄)−1/2 and n0 ≥ (1 − cK)n, the right hand side is lower bounded by
(c0 ǫ0 /8) · (nθ̄ 2 )−1/2 . This proves (i).
P
(ℓ)
(ℓ)
(0)
We now prove (ii). Note that KL(Pℓ , P0 ) = 1≤i<j≤n Ωij log(Ωij /Ωij ).
Additionally, the parametrization (2.8) yields that
(2.10)
Ωij = θi θj (1 − a/2) + (a/2)γi γj ,
1 ≤ i 6= j ≤ n.
(0)
(ℓ)
(ℓ)
(0)
Since γi = γi for all i > n0 , if both i, j > n0 , then Ωij = Ωij and the
pair (i, j) has no contribution to KL(Pℓ , P). Therefore, we can write
J+1
1 X
KL(Pℓ , P0 )
J +1
=
(2.11)
1
J +1
ℓ=0
J+1
X
X
ℓ=0 1≤i<j≤n0
≡(I) + (II).
X
+
(ℓ)
(ℓ)
(0)
Ωij log(Ωij /Ωij )
1≤i≤n0 ,n0 <j≤n
First, consider (I). From (2.9) and (2.10), for all 1 ≤ i < j ≤ n0 , we have
(0)
Ωij = θi θj (1 − a/2) and
(2.12)
(ℓ)
(0)
(ℓ)
Ωij = Ωij (1 + ∆ij ),
(ℓ)
where ∆ij =
(ℓ)
δ2
a
(ℓ) (ℓ)
p n · ωi ωj .
2 − a θi θj
Write ∆max = max1≤i<j≤n0 ,1≤ℓ≤J |∆ij |. Since θ1 ≥ . . . ≥ θn0 ≫ n−1/2 ,
we have ∆max = O(n1/2 δn2 ) = O((nθ̄ 2 )−1/2 ) = o(1). By Taylor expansion,
(1 + t) ln(1 + t) = t + O(t2 ) ≤ 2|t| for t sufficiently small. Combining the
above gives
(ℓ)
(ℓ)
(0)
(0)
(ℓ)
(ℓ)
Ωij log(Ωij /Ωij ) = Ωij (1 + ∆ij ) ln(1 + ∆ij )
(0)
(ℓ)
≤ 2Ωij |∆ij |
p
(ℓ) (ℓ)
≤ aδn2 · θi θj · |ωi ωj |
10
≤ aδn2
p
θi θj ,
(0)
where the third line is due to (2.12) and the expression Ωij . It follows that
(2.13)
(I) ≤ aδn2
X
1≤i<j≤n0
p
θi θj ≤ aδn2 ·
X p 2
θi .
1≤i≤n0
(0)
Next, consider (II). For i ≤ n0 and j > n0 , Ωij = θi θj (1 − a/2) and
(2.14)
δn
(ℓ)
(0)
(ℓ)
(ℓ)
˜ (ℓ) ), with ∆
˜ (ℓ) = γ (ℓ) · a √
Ωij = Ωij (1 + ∆
· ω and γj ∈ {±1}.
ij
ij
j
2 − a θi i
˜ max = max1≤i≤n ,n <j≤n,1≤ℓ≤J |∆
˜ (ℓ) |. Similar to the bound for ∆max ,
Write ∆
0 0
ij
˜ max = O(n1/4 δn ) = O((nθ̄)−1/4 ) = o(1). Also, by Taylor expanwe have ∆
sion, (1 + t) ln(1 + t) = t + t2 /2 + O(|t|3 ) ≤ t + t2 for t sufficiently small.
Combining the above gives
(ℓ)
(ℓ)
(0)
(0)
(ℓ)
(ℓ)
Ωij log(Ωij /Ωij ) = Ωij (1 + ∆ij ) ln(1 + ∆ij )
(0) ˜ (ℓ)
(0) ˜ (ℓ) 2
≤ Ωij ∆
ij + Ωij (∆ij ) .
(2.15)
Motivated by (2.15), we first bound
(II1 ) ≡
=
n
J n0
X
1 XX
(0) ˜ (ℓ)
Ωij ∆
ij
J +1
ℓ=0 i=1 j=n0 +1
J n0
n
X
p (ℓ)
1 XX
aδn
(ℓ)
· θ j γj · θ i ω i
J +1
2
ℓ=0 i=1 j=n0 +1
=
J X
n p
n
X
aδn X
(ℓ)
(ℓ)
θi ωi
·
θj γj
2(J + 1)
j=n0 +1
ℓ=0 i=1
= 0,
where we have used Property (b) of ω (0) , ω (1) , · · · , ω (J) . We then bound
(II2 ) ≡
=
J n0
n
X
1 XX
(0) ˜ (ℓ) 2
Ωij (∆
ij )
J +1
ℓ=0 i=1 j=n0 +1
J n0
n
X
a2 δn2
1 XX
(ℓ)
· θj · |ωi |
J +1
4 − 2a
ℓ=0 i=1 j=n0 +1
11
n
a2 δn2 X
θj · max kω (ℓ) k1
≤
0≤ℓ≤J
4 − 2a
j=n0 +1
a2 δn2
≤
4 − 2a
n
X
j=n0 +1
θi · n0 .
Combining the above gives
(2.16)
n
X
a2 δn2
(II) ≤
θi .
· n0
4 − 2a
j=n0 +1
(2.13)
PLast, we√combine
Pn0 and (2.16). Using the Cauchy-Schwartz inequality,
2
( 1≤i≤n0 θi ) ≤ n0 i=1 θi . Hence, the right hand side of (2.13) is upper
P 0
θi ). Furthermore,P
since a ∈ (0, 1], the right hand
bounded by aδn2 · n0 ( ni=1
2
side of (2.16) is upper bounded by aδn · n0 ( ni=n0 +1 θi ). As a result,
J
n
X
1 X
KL(Pj , P0 ) ≤ aδn2 · n0
θi = ac0 n0 ,
J +1
j=0
i=1
where we have plugged in δn = c0 (nθ̄)−1/2 . At the same time, log(J) ≥
[log(2)/8]n0 . Hence, the requirement (ii) is satisfied as long as c0 is chosen
appropriately small. The proof for K = 2 is now complete.
2.2. The case of K ≥ 3. The key step is to generalize the construction
of Π(0) , Π(1) , . . . , Π(J) for K = 2 to a general K. Write P̌ = 1K 1′K − P . Let
η ∈ RK be a nonzero vector such that
(2.17)
η ′ 1K = 0,
η ′ P̌ 1K = 0.
Such an η always exists. We assume θ1 ≥ θ2 ≥ . . . ≥ θn without loss of
generality. Let n1 = ⌊cn⌋ and n0 = n − Kn1 . Denote by e1 , . . . , eK the
standard basis vectors of RK . Introduce
′
1
1K , · · · , K1 1K , e1 , · · · , e1 , · · · , eK , · · · , eK .
Π∗ = K
{z
}
|
|
{z
} | {z }
n0
n1
n1
Let ω (0) , ω (1) , . . . , ω (J) ∈ {0, 1}n0 be the same as above. Let δn = c0 (nθ̄)−1/2
for a constant c0 to be determined. Write Π∗ = [π1∗ , . . . , πn∗ ]′ . For each 0 ≤
(ℓ)
(ℓ)
ℓ ≤ J, we construct Π(ℓ) = [π1 , . . . , πn ]′ by
(
√
(ℓ)
ωi · (δn / θi ) · η, 1 ≤ i ≤ n0
(ℓ)
∗
πi = πi +
0K ,
n0 + 1 ≤ i ≤ n.
12
Same as before, we show the requirements (i)-(ii) in Lemma 2.1 are satisfied.
Note that
L(Π(j) , Π(ℓ) ) =
n δ kηk
δn kηk1 (j)
√ kω − ω (ℓ) k1 ≥ 0 n√ 1 = (c0 ǫ0 kηk1 /8) · (nθ̄ 2 )−1/2 .
n θ̄
8n θ̄
Hence, (i) holds for sn = (nθ̄ 2 )−1/2 .
It remains is to prove (ii). Let (I) and (II) be defined in the same way
as in (2.11). We aim to find expressions similar to those in (2.12) and (2.14)
and then generalize the bounds of (I) and (II) for K = 2 to a general K. For
preparation, we first derive an expression for Ωij . Introduce π̌i = πi − K1 1K ∈
RK . Note that πi′ 1K = 1. By direct calculations,
Ωij = θi θj πi′ P πj = θi θj (1 − πi′ P̌ πj )
= θi θj − θi θj (π̌i +
= θi θj (1 −
(2.18)
1 ′
1 P̌ 1K )
K2 K
(0)
Consider (I). Since π̌i
(0)
′
1
K 1K ) P̌ (π̌j
Ωij = θi θj (1 − ǎ),
−
1
K 1K )
1
(π̌i′ P̌ 1K
θi θj K
+
+ π̌j′ P̌ 1K ) − θi θj π̌i′ P̌ π̌j .
is a zero vector for all 1 ≤ i ≤ n0 , we have
with ǎ ≡
1 ′
1 P̌ 1K ,
K2 K
if 1 ≤ i 6= j ≤ n0 .
(ℓ)
Furthermore, for 1 ≤ i ≤ n0 , π̌i ∝ η, where it follows from (2.17) that
η ′ P̌ 1K = 0. Hence, the middle two terms in (2.18) are zero. As a result,
δ2
(ℓ) (ℓ)
(ℓ)
Ωij = θi θj (1 − ǎ) − p n · η ′ P̌ η · ωi ωj ,
θi θj
1 ≤ i 6= j ≤ n0 .
Combining the above, we find that for all 1 ≤ i 6= j ≤ n0 ,
(ℓ)
(0)
(ℓ)
(2.19) Ωij = Ωij (1+∆ij ),
(ℓ)
where ∆ij =
−(η ′ P̌ η)ǎ δn2
(ℓ) (ℓ)
·ω ω .
·p
1 − ǎ
θi θj i j
This provides a counterpart of (2.12) for a general K. Same as before, we
(ℓ)
have the bound: ∆max ≡ max1≤i6=j≤n0 max0≤ℓ≤J |∆ij | = O(n1/2 δn2 ) = o(1).
Following the proof of (2.13), we find that
(2.20)
n0
n0 p
X
X
2
where C1 = |η ′ P̌ η| · |ǎ|.
θi ,
(I) ≤ C1 δn2
θi ≤ C1 n0 δn2
i=1
i=1
Consider (II). In this case, we need to calculate Ωij , 1 ≤ i ≤ n0 , n0 < j ≤ n.
(0)
Recall that π̌i is a zero vector. Write {n0 + 1, n0 + 2, . . . , n} = ∪K
k=1 Nk ,
13
(0)
(0)
where Nk = {1 ≤ j ≤ n : πj = ek }. For j ∈ Nk , it holds that π̌j + K1 1K =
ek . Combining the above with (2.18), we find that for 1 ≤ k ≤ K,
(0)
Ωij = θi θj (1 − b̌k ),
b̌k ≡
with
1 ′
K ek P̌ 1K ,
if 1 ≤ i ≤ n0 , j ∈ Nk .
(ℓ)
(ℓ)
Additionally, we have (π̌i )′ P̌ 1K ∝ η ′ P̌ 1K = 0 and π̌j +
follows from (2.18) that
(ℓ)
(ℓ)
1
K 1K
= ek . It
(ℓ)
Ωij = θi θj (1 − b̌k ) − θi θj (π̌i )′ P̌ π̌j
(ℓ)
ωi δn ′
· η P̌ (ek −
= θi θj (1 − b̌k ) − θi θj · √
θi
1
K 1K )
(ℓ)
ωi δn
· (η ′ P̌ ek ),
= θi θj (1 − b̌k ) − θi θj · √
θi
where the last equality is because of η ′ P̌ 1K = 0. As a result, for 1 ≤ i ≤ n0
and j ∈ Nk ,
(ℓ)
(0)
(ℓ)
˜ (ℓ) = −(η ′ P̌ ek )·
where ∆
ij
˜ ),
(2.21) Ωij = Ωij (1+ ∆
ij
δ
1
√n ·ωi(ℓ) .
1 − b̌k θi
This provides a counterpart of (2.14) for a general K. Same as before, let
˜ max ≡ max1≤i6=j≤n max0≤ℓ≤J |∆
˜ (ℓ) |, and it is seen that ∆
˜ max = O(n1/4 δn ) =
∆
0
ij
o(1). Again, by Taylor expansion, we have (2.15). It follows that
(II) ≤ (II1 ) + (II2 ),
where (II1 ) and (II2 ) are defined the same as before. Using (2.21), we have
J
n
0
1 XX
(II1 ) =
J +1
X
(0) ˜ (ℓ)
Ωij ∆
ij
ℓ=0 i=1 j∈N1 ∪...∪NK
J X
n p
K
X
δn X X
(ℓ)
′
−(η P̌ ek )θj ·
=
θi ωi
J +1
ℓ=0 i=1
k=1 j∈Nk
= 0,
where the last equality is due to Property (b) for ω (0) , ω (1) , . . . , ω (J) . Using
(2.21) again, we have
J
(II2 ) =
n
0
1 XX
J +1
X
ℓ=0 i=1 j∈N1 ∪...∪NK
(0)
(ℓ)
˜ )2
Ωij (∆
ij
14
≤
X
X
δn2
θj · max kω (ℓ) k1
·
(η ′ P̌ ek )2
0≤ℓ≤J
1 − b̌ 1≤k≤K
j∈N
k
≤
δn2
1 − b̌
· max (η ′ P̌ ek )2 ·
1≤k≤K
n
X
j=n0 +1
θi · n0 .
Combining the above gives
(2.22) (II) ≤ C2 n0 δn2
n
X
i=n0 +1
θi ,
where C2 =
1
max (η ′ P̌ ek )2 .
1 − b̌ 1≤k≤K
We note that (2.20) and (2.22) server as the counterpart of (2.14) and (2.16),
respectively. Similarly as in the case of K = 2, we obtain (ii) immediately.
The proof for a general K is now complete.
REFERENCES
[1] Airoldi, E., Blei, D., Fienberg, S. and Xing, E. (2008). Mixed membership
stochastic blockmodels. J. Mach. Learn. Res. 9 1981–2014.
[2] Anandkumar, A., Ge, R., Hsu, D. and Kakade, S. (2014). A tensor approach to
learning mixed membership community models. J. Mach. Learn. Res. 15 2239–2312.
[3] Gao, C., Ma, Z., Zhang, A. Y. and Zhou, H. H. (2016). Community detection in
degree-corrected block models. arXiv:1607.06993.
[4] Ji, P. and Jin, J. (2016). Coauthorship and citation networks for statisticians (with
discussion). Ann. Appl. Statist. 10 1779–1812.
[5] Jin, J. (2015). Fast community detection by SCORE. Ann. Statist. 43 57–89.
[6] Jin, J., Ke, Z. and Wang, W. (2017). Phase transitions for high dimensional clustering and related problems. Ann. Statist., to appear.
[7] Jin, J. and Ke, Z. T. (2017). Optimal estimation of network mixed memberships.
Manuscript.
[8] Jin, J., Ke, Z. T. and Luo, S. (2017). Estimating network memberships by simplex
vertex hunting. arXiv:1708.07852.
[9] Jin, J., Ke, Z. T. and Luo, S. (2017). SCORE+: a refinement of SCORE.
Manuscript.
[10] Jin, J. and Wang, W. (2016). Influential Features PCA for high dimensional clustering (with discussion). Ann. Statist. 44 2323-2359.
[11] Karrer, B. and Newman, M. (2011). Stochastic blockmodels and community structure in networks. Phys. Rev. E 83 016107.
[12] Ke, Z. T. and Wang, M. (2017). A new SVD approach to optimal topic estimation.
arXiv:1704.07016.
[13] Krebs, V. (2004). Unpublished.
[14] Tsybakov, A. B. (2009). Introduction to nonparametric estimation. Revised and
extended from the 2004 French original. Translated by Vladimir Zaiats.
[15] Zhang, A. Y. and Zhou, H. H. (2016). Minimax rates of community detection in
stochastic block models. Ann. Statist. 44 2252–2280.
15
[16] Zhang, Y., Levina, E. and Zhu, J. (2014). Detecting overlapping communities in
networks with spectral methods. arXiv:1412.3432.
J. Jin
Department of Statistics
Carnegie Mellon University
Pittsburgh, Pennsylvania, 15213
USA
E-mail: [email protected]
Z. Ke
Department of Statistics
University of Chicago
Chicago, Illinois, 60637
USA
E-mail: [email protected]
| 10math.ST
|
arXiv:1612.04410v1 [math.GR] 13 Dec 2016
THE DIVISIBILITY GRAPH OF FINITE GROUPS OF LIE
TYPE
ADELEH ABDOLGHAFOURIAN, MOHAMMAD A. IRANMANESH,
AND ALICE C. NIEMEYER
Abstract. The Divisibility Graph of a finite group G has vertex set
the set of conjugacy class lengths of non-central elements in G and two
vertices are connected by an edge if one divides the other. We determine
the connected components of the Divisibility Graph of the finite groups
of Lie type in odd characteristic.
Keywords: Divisibility Graph, Finite Group of Lie Type
MSC2010: primary: 20G40, secondary:05C25
1. Introduction
Given a set X of positive integers, several graphs corresponding to X can
be defined. For example the prime vertex graph Γ(X), the common divisor
graph ∆(X) and the bipartite divisor graph B(X) (see [24, 27] and references
therein for more details). In 2011 Camina and Camina [8] introduced the
divisibility graph D(X) of X as the directed graph with vertex set X\{1}
with an edge from vertex a to vertex b whenever a divides b. For a group G
let cs(G) denote the set of conjugacy class lengths of non-central elements
in G. Camina and Camina asked [8, Question 7] how many components
the divisibility graph of cs(G) has. Clearly it is sufficient to consider the
underlying undirected graph and for the remainder of this paper the Divisibility Graph D(G) of a group G refers to the undirected Divisibility Graph
D(cs(G)).
Note that the set of vertices cs(G) may be replaced by the set C(G) of
orders of the centralisers of non-central elements of G.
In [7, 27] has been shown that the graphs Γ(cs(G)), ∆(cs(G)) and B(cs(G))
have at most two connected components when G is a finite group. Indeed,
when G is a nonabelian finite simple group, then Γ(cs(G)) is complete (see
[5, 17]). It is clear that D(G) is a subgraph of Γ(cs(G))) and we hope that
the structure of D(G) reveals more about the group G.
The first and second authors have shown that for every comparability
graph, there is a finite set X such that this graph is isomorphic to D(X)
in [3]. They found some relationships between the combinatorial properties
of D(X), Γ(X) and ∆(X) such as the number of connected components,
diameter and girth (see [3, Lemma 1]) and found a relationship between
1
2
A.ABDOLGHAFOURIAN, M.A. IRANMANESH, AND A.C. NIEMEYER
D(X × Y ) and product of D(X) and D(Y ) in [3]. They examined the Divisibility Graph D(G) of a finite group G in [1] and showed that when G
is the symmetric or alternating group, then D(G) has at most two or three
connected components, respectively. In both cases, at most one connected
component is not a single vertex.
Here we are interested in the Divisibility Graphs of finite groups of Lie
type. The graph of certain finite simple groups of Lie type is known, namely
the first and second authors described in [2] the structures of the Divisibility
Graphs for PSL(2, q) and Sz(q). Let Ki denote the complete graph on i
vertices. They prove in [2, Theorem 6] that for G = PSL(2, q) the graph
D(G) is either 3K1 or K2 + 2K1 . For the other finite groups of Lie type in
odd characteristic we prove the following theorem:
Theorem 1. Let G be a finite group of Lie type over a finite field of order q
or q 2 in characteristic p where p is an odd prime. Suppose further that the
Lie rank ℓ of G is either as in Table 1 or at least 2. Then the Divisibility
Graph D(G) has at most one connected component which is not a single
vertex.
In particular, we prove that the non-trivial connected component of D(G)
contains the lengths of the conjugacy classes of all non-central involutions
as well as the lengths of all conjugacy classes of all unipotent elements. The
number of connected components consisting of a single vertex corresponds to
the number of conjugacy classes T G of tori T in G for which |T Z(G)/Z(G)|
is odd, coprime to |Z(G)| and for which the centraliser in G of every noncentral element in T is T Z(G).
We now compare the results of the theorem to known results about another type of graph, namely the Prime Graph first introduced by Gruenberg
and Kegel in 1975 in an unpublished manuscript. The vertex set of the
Prime Graph of a finite group G is the set of primes dividing the order of
the group and two vertices r and s are adjacent if and only if G contains an
element of order rs. Williams [37, Lemma 6] investigated Prime Graphs of
finite simple groups in odd characteristic and Kondrat’ev [26] and Lucido
[29] investigated these graphs for even characteristic and for almost simple
groups, respectively. A subgroup T of a group G is called a CC-group if
CG (t) ≤ T for all t ∈ T \Z(G). Williams proved [37, Theorem 1] that the
connected components of the Prime Graph of a finite simple group G of Lie
type in odd characteristic p consist at most of the set {p}, the connected
component containing the prime 2, and a collection of sets consisting of the
primes dividing the order of some torus T for which T Z(G)/Z(G) has odd
order coprime to |Z(G)| and is a CC-group (see Section 3.3). Moreover, for
such groups he showed that {p} is an isolated vertex of the Prime Graph if
and only if G ∼
= PSL(2, q) with q odd. Hence for the groups we consider (see
Section 2.1 where we exclude small dimensions and certain ‘bad’ primes) we
may assume that {p} is not an isolated vertex in the Prime Graph. In this
case, Williams shows that the number of connected components of a finite
THE DIVISIBILITY GRAPH OF FINITE GROUPS OF LIE TYPE
3
simple group of Lie type is at most two, unless G = 2 Dp (3) and p = 2n + 1
a prime for n ≥ 2, for which the Prime Graph has three components and
unless G = E8 (q), for which the Prime Graph can have either four or five
components, depending on whether q ≡ 2, 3 (mod 5) or q ≡ 0, 1, 4 (mod 5),
respectively. We verified the main theorem separately for the case of small
dimensions or the ‘bad’ primes. Hence we obtain the following corollary.
Corollary 2. Let G be a finite group of Lie type over a finite field of order
q or q 2 in characteristic p where p is an odd prime. Suppose further that
the Lie rank ℓ of G is either as in Table 1 or at least 2 and that q 6= 3 if
G is of type E7 . Then the Divisibility Graph D(G) has as many connected
components as the Prime Graph of G.
We note that if G is of adjoint or simply connected type E7 (3), the Divisibility graph of G is connected, whereas the Prime Graph has three connected
components (see [37, Table Id].)
The Prime Graph of G, and thus also the Divisibility Graph of G, is linked
to yet another graph defined for G, the Commuting Graph introduced in
1955 by Brauer and Fowler [6]. The vertices of the Commuting Graph of a
group G are the non-central elements of G and two elements are connected
by an edge if and only if they commute. Due to the work of Morgan and
Parker [31, Theorem 3.7] and Iranmanesh and Jafarzadeh [23, Lemma 4.1]
we also obtain the following corollary.
Corollary 3. Let G be a finite group of Lie type over a finite field of order
q or q 2 in characteristic p where p is an odd prime. Suppose further that
G has trivial centre, that the Lie rank ℓ of G is either as in Table 1 or at
least 2 and that p ≥ 3 if G is of type E7 . Then the G-classes of connected
components of the commuting graph of G are in one to one correspondence
with the connected components of the Divisibility Graph D(G).
1.1. Example. As an example we determine the Divisibility Graph of the
groups PSL(3, q) and PSU(3, q) for q odd. Their conjugacy classes and character tables can be found in [35]. In particular, the centraliser orders of nontrivial elements for both groups are C(G) = {q 3 r ′ , q 2 , qr ′ rs, qr ′ , r 2 , r ′ r, r ′ s, t′ }
where ǫ = 1 for G = PSL(3, q) and ǫ = −1 for G = PSU(3, q) and r = q − ǫ,
s = q + ǫ, t = q 2 + ǫq + 1, a = gcd(3, r), r ′ = r/a, and t′ = t/a. The graph
D(G), shown in Figure 1, depends on a, since when a = 1 the centraliser
orders r 2 and r ′ r agree. The vertex t′ corresponds to a Coxeter torus in G
of odd order which is a CC-group.
2. Background
For a finite group G let cs(G) = {|xG | | x ∈ G}\{1} denote the set
of conjugacy class lengths of non-central elements in G. Let D(G) denote
the Divisibility Graph of G, the graph with vertex set cs(G) and edge set
E(G) = {(|xG |, |y G |); either |xG | divides |y G | or |y G | divides |xG |}.
4
A.ABDOLGHAFOURIAN, M.A. IRANMANESH, AND A.C. NIEMEYER
q 3 r′
q2
qrr′ s
t
qr′
q3r
q2
qr2 s
r′ s
′
qr
rs
t
r2
rr′
r2
Figure 1. The divisibility graph for PSL(3, q) and PSU(3, q)
(left: gcd(3, r) 6= 1 right: gcd(3, r) = 1).
Note that for x, y ∈ G there is an edge in D(G) between |xG | and |y G |
if and only if |CG (y)| divides |CG (x)| or |CG (x)| divides |CG (y)|. Therefore,
the set of vertices cs(G) may be replaced by the set C(G) = {|CG (x)| | x ∈
G, x 6∈ Z(G)}. If (|xG |, |y G |) ∈ E(G) we write x ∼ y. We say non-central
elements x, y ∈ G are equivalent if |xG | =
6 |y G | and |xG | and |y G | are in the
same connected component of D(G). Note that being equivalent induces an
equivalence relation on C(G).
Lemma 4. Let G be a finite group and x, y ∈ G\Z(G).
(1) If gcd(|x|, |y|) = 1 and xy = yx then CG (xy) = CG (x) ∩ CG (y) and
in particular x ∼ xy ∼ y.
(2) x ∼ xm for any m ∈ N\{1} such that xm 6∈ Z(G).
(3) If gcd(|xZ(G)|, |yZ(G)|) = 1 and xyZ(G) = yxZ(G) then x is equivalent to y.
Proof. Note that (1) is [15, Lemma 3]. (2) is obvious. Now consider (3).
By replacing x and y by xmx and y my for some integers mx , my if necessary,
we may assume that gcd(|x|, |y|) = 1. Now suppose z ∈ Z(G) such that
xy = yxz. Then |x| = |xy | = |xz| = lcm(|x|, |z|) and |y| = |y x | = |yz −1 | =
lcm(|y|, |z|). Since |z| divides gcd(|x|, |y|) = 1 we have z = 1. Hence xy = yx
and the result follows from (1).
2.1. The groups we consider. Let q = pf be a power of an odd prime
p and f a non-negative integer. Let G be a connected semisimple (which
implies reductive) algebraic group of rank ℓ with ℓ ≥ 2 defined over the
algebraic closure Fq or Fq2 of the finite field Fq or Fq2 . Let F : G → G be
a Frobenius morphism and let G = GF = {g ∈ G | F (g) = g} be a finite
group of Lie type. Moreover, if G is a classical group of Lie type, we assume
that Ω ≤ G ≤ ∆, where Ω and ∆ and ℓ are as in Table 1. For a field K and a
positive integer n, for convenience we denote both GL(n, K) and GU(n, K)
by GLǫ (n, K), where ǫ = 1 in the former and ǫ = −1 in the latter case.
THE DIVISIBILITY GRAPH OF FINITE GROUPS OF LIE TYPE
Type
Aℓ
2A
ℓ
Bℓ
Cℓ
Dℓ
2D
ℓ
n
ℓ+1
ℓ+1
2ℓ + 1
2ℓ
2ℓ
2ℓ
Ω
∆
dim(∆)
Rank
SL(n, q)
SU(n, q)
Ω(n, q)
Sp(n, q)
Ω+ (n, q)
Ω− (n, q)
GL(n, q)
GU(n, q)
GO(n, q)
GSp(n, q)
GO+ (n, q)
GO− (n, q)
n2
n2
n(n − 1)/2
n(n + 1)/2
n(n − 1)/2
n(n − 1)/2
ℓ≥3
ℓ≥3
ℓ≥2
ℓ≥3
ℓ≥2
ℓ≥3
5
Table 1. Finite classical groups considered in Theorem 1
Note that SL2 (q) ∼
= Sp2 (q) ∼
= SU2 (q) so that we take ℓ ≥ 3 for types
and Cℓ . Moreover, Ω3 (q) ∼
= PSL2 (q) so we also take ℓ ≥ 2 in case
2 ), we take ℓ ≥ 3 in case 2 D . Moreover, Ω+ (q) ∼
∼
Bℓ . As Ω−
(q)
PSL
(q
=
=
2
ℓ
4
4
SL2 (q) ◦ SL2 (q), where ◦ denotes the central product. We determine the
Divisibility Graph of this group directly. In fact, it is easy to see that in this
case the Divisibility Graph is always connected. Thus for the remainder of
the proof we may assume ℓ ≥ 3 in case Dℓ .
For convenience we also record the dimension of the algebraic group corresponding to ∆ in the same table.
For the exceptional groups when p is a bad and odd prime, we obtained the
Divisibility graphs directly, either using the Tables in [12, 16, 18, 19, 34] or
using an explicit list of the generic centraliser orders in E7 (q) as polynomials
in q computed by Frank Lübeck. Using Lübeck’s description as polynomials
in q in GAP [21], we could first determine the lengths of the conjugacy
classes of E7 (q) as polynomials in q together with their multiplicities. For
those conjugacy classes with non-zero multiplicities, we could determine the
divisibility graph. We verified that Theorem 1 holds for all such groups. In
the case of the adjoint group E7 (q) for q a power of 3, the divisibility graph
is connected even when q = 3.
Therefore, from now on we assume that p is odd and a good prime for G,
that is ([9, p. 28])
2A
ℓ
• p 6= 2 when G has type Aℓ , 2 Aℓ , Bℓ , Cℓ , Dℓ , 2 Dℓ ,
• p 6∈ {2, 3} when G has type G2 , F4 , E6 , 2 E6 , E7 ,
• p 6∈ {2, 3, 5} when G has type E8 .
The force of p being a good prime is that for any x ∈ G all unipotent
elements in CG (x) lie in CG (x)◦ .
Let F : G → G be a Frobenius morphism and let G = GF = {g ∈ G |
F (g) = g} be a finite group of Lie type, T0 a fixed F -stable maximal torus
and W = NG (T0 )/T0 the Weyl group of G. Throughout the paper, let G
denote G/Z(G) and g = gZ(G) for g ∈ G.
6
A.ABDOLGHAFOURIAN, M.A. IRANMANESH, AND A.C. NIEMEYER
An element g ∈ G is called regular, if CG (g) has dimension ℓ, where ℓ
denotes the rank of G, i.e. the dimension of a maximal torus of G. More
background on the finite groups of Lie type can be found in [9] and [33].
3. Proof of the main theorem
Our aim is to determine the connected components of the Divisibility
Graph D(G). It is well known that each element g ∈ G has a Jordan
decomposition g = us = su with u unipotent and s semisimple. If g is noncentral then either s is non-central and hence g ∼ s or u is non-trivial and
g ∼ u. Thus we consider non-central semisimple and unipotent elements.
In Section 3.1 we show that all unipotent elements in G are equivalent.
In Section 3.2 we show that every non-central and non-regular semisimple
element is equivalent to a unipotent element in G. In Section 3.3 we recall
the notion of a CC-group introduced by Williams and show that certain
maximal tori in G which are also CC-groups yield isolated vertices of D(G).
Finally we show in Section 3.4 that all regular semisimple elements which
do not lie in such a torus are also equivalent to a unipotent element. Thus
the main theorem follows.
3.1. Unipotent elements. Let G be a finite classical group of Lie type
and let u be a unipotent element in G. We determine the q-part of the order
of the centralizer of u in ∆ when G is one of the groups in Table 1. As q
does not divide |Z(G)| nor [∆ : Ω], the q-part of |CH (u)| is the same for all
groups H with Ω ≤ H ≤ ∆ and for groups H/Z(H).
Lemma 5. Let G = GLn (Fq ), Spn (Fq ) or GOn (Fq ) and let G = GF . Let u
be a unipotent element in G which corresponds in G to a Jordan decomposition ⊕Jiri . In particular, in cases Sp or O the integer ri is even for i odd
or i even, respectively. Then |CG (u)|q = q au , where au is determined by the
following formulas:
X
X
ri (ri + 1)
2
ǫ
(1) G = GL (n, q) then au =
+2
iri rj .
iri −
2
i
i<j
X
X 1
1
1X
(i − )ri2 +
iri rj +
.
(2) G = Sp(n, q) then au =
2
2
4
i even
i
i<j
ǫ
ri odd
(3) G = GO (n, q) for ǫ ∈ {◦, +, −}, then
X 1
X
1X
1
1X
au =
(i − )ri2 +
ri +
iri rj −
.
2
2
2
4
i
i<j
i
i,ri odd
Proof. We first note that it is well known that CG (u) admits a Levi-decomposition,
see for example [22, Prop. 3.2] or [28, Theorem 3.1]. That is, CG (u) = UR,
where U = Ru (CG (u)) is the unipotent radical of CG (u), the group R is
reductive and U ∩ R = {1}. Moreover, U and R are F -stable, CG (u) =
U R, and |U | = q dim U , see [22, Prop. 3.2] or [28, Theorem 7.1]. Thus
we need to determine dim U. Consequently, dim U = dim CG (u) − dim R.
THE DIVISIBILITY GRAPH OF FINITE GROUPS OF LIE TYPE
7
We consider the different cases of classical groups in turn and note that
in each case [28, Theorem 3.1(iii)] determines dim CG (u) and [28, Theorem 3.1(iv),Theorem 7.1(ii)] describe R and thus dim R, and R. It follows
that |CG (u)|q = q dim U |R|q .
P 2
ir +
Case GL: Let G = GLǫ (n, q). Then dim CGLǫ (n,Fq ) (u) =
P 2
Q i ǫi
P
Q
2 i<j iri rj and R = GL(ri , Fq ), thus dim R = i ri , and R = GL (ri , Fq ),
P
P
2
showing that dim U =
i (i − 1)ri + 2
i<j iri rj . Thus |CGL(n,q) (u)|q =
Q
dim(U)
a
u
q =q
i |GL(ri , q)|q . Thus
X
X
X ri (ri − 1)
(i − 1)ri2 + 2
iri rj +
2
i
i<j
i
X
X
ri (ri + 1)
=
+2
iri rj .
iri2 −
2
au =
i
i<j
1 X
1X 2 X
iri +
iri rj +
ri and, more2
2
i
i odd
i<j
Y
Y
P
over, R =
Sp(ri , Fq ) ×
GO(ri , Fq ). Thus, dim R = 12 i odd ri (ri +
i even
P i odd
1) + 21 i even ri (ri − 1), showing that
Case Sp: Here dim CG (u) =
dim U = dim CG (u) − dim R
1 X
1 X
1 X
1X 2 X
iri +
iri rj +
ri −
ri (ri + 1) −
ri (ri − 1)
=
2
2
2
2
i
=
=
1X
2
i odd
i<j
iri2 +
i
X
i<j
i odd
i even
1 X 2 1 X
iri rj −
ri −
ri (ri − 1)
2
2
i odd
i even
X
1X
1 X
(i − 1)ri2 +
iri rj +
ri .
2
2
i
i<j
It follows that q au = q dim(U)
au =
Q
i even
i odd |Sp(ri , q)|q
Q
i even |GO
ǫi (r , q)|
i
q
and
X
X r2
X 1
1X
i
(i − 1)ri2 +
iri rj +
+
.
2
4
4
i even
i
i<j
i
ri odd
1X 2 X
1 X
iri +
iri rj −
ri and further2
2
i
i<j
i odd
Y
Y
P
Sp(ri , Fq ). Thus dim R = i odd ri (r2i −1) +
more R =
GO(ri , Fq ) ×
Case O: Here dim CG (u) =
i odd
i even
8
P
A.ABDOLGHAFOURIAN, M.A. IRANMANESH, AND A.C. NIEMEYER
i even
ri (ri +1)
,
2
showing that
dim U = dim CG (u) − dim R
1X 2 X
1 X
1 X
1 X
=
iri +
iri rj −
ri −
ri (ri − 1) −
ri (ri + 1)
2
2
2
2
i
i<j
i odd
i odd
i even
=
1X 2 X
1 X 2 1 X
iri +
iri rj −
ri −
ri (ri + 1)
2
2
2
=
X
1X
1 X
(i − 1)ri2 +
iri rj −
ri .
2
2
i<j
i
i odd
i
Therefore, q au = q
au =
i<j
Q
dim(U)
i odd |GO
i even
i even
ǫi (r , q)|
i
q
Q
i even |Sp(ri , q)|q
and
X 1
X
1
1X
1X
(i − )ri2 +
ri +
iri rj −
.
2
2
2
4
i
i
i<j
i,ri odd
We now show that all unipotent elements in G are equivalent.
Lemma 6. Let G be a finite group of Lie type as given in Section 2.1. Then
all unipotent elements in G are equivalent.
Proof. Let u be a regular unipotent element in G, which exists by [9, Prop.5.1.7].
Then u lies in a maximal connected F -stable unipotent subgroup U of G
and by [36, III.1.14] CG (u) = Z(G).CU (u) and, since p is good, CU (u) is
connected. Moreover, |CG (u)|q = q ℓ .
Note that it is enough to show this for one of the groups with given Dynkin
Diagram as their q-parts of the centralisers are the same. For the classical
groups we restrict our attention to the groups considered in Lemma 5.
Let v be unipotent with Jordan decomposition ⊕i Jiri . By Lemma 5
|CG (v)|q = q av . If v is not regular, we show that q ℓ divides |CG (v)|q . From
this it follows that v is equivalent to u, as all semisimple elements in CG (u)
lie in Z(G) andPZ(G) ≤ CG (v). Thus we need to show that av − ℓ ≥ 0. We
note that n = i iri .
Suppose first that G = GLǫ (n, q). By Lemma 5 |CGL(n,q) (v)|q = q av with
P
P
av = i iri2 − ri (ri + 1)/2 + 2 i<j iri rj . As iri rj ≥ ri rj , we note that
X
X
(ri + 1)
+
rj .
av − ℓ ≥
ri iri − i −
2
j6=i
i
We show that for any i the i-th summand is non-negative. Clearly this is
the case when ri = 0. So suppose now ri > 0. Then it suffices to show
P
i+ 1
(i − 12 )ri − (i + 21 ) + j6=i rj ≥ 0 if ri ≥ i− 21 = 1 + i−1 1 . For i = 1 this is the
2
2
case for ri ≥ 3 and for i > 1 this is the case for ri ≥ 2. Hence for any such
pair (i, ri ) each summand is non-negative.
THE DIVISIBILITY GRAPH OF FINITE GROUPS OF LIE TYPE
9
Now consider the case i = 1 and ri = 1, 2 or i > 1 and ri = 1. As v is not
regular, i < n. And since n ≥ 3 in either case, there
P is at least one j with
j 6= i such that rj 6= 0. Then (i − 21 )ri − (i + 12 ) + j6=i rj ≥ (i − 21 )ri − i + 12 .
For ri = 1 this expression is 0 and for ri = 2 we have i = 1 and thus
(i − 12 )ri − i + 12 = 1/2 ≥ 0.
Suppose
now that G = Sp(n, q) with n even. As iri rj ≥ ri rj , and ℓ =
1P
1
5(b) |CSp(n,q) (v)|q = q av with av −ℓ ≥
2 n = 2 i iri , we note that by Lemma
P
P
1
1
i ri (i − 2 )ri +
j6=i rj − i .
2
It suffices to show that for any i the i-th summand is non-negative. This
is certainly the case when ri = 0. So suppose ri > 0. This is the case when
ri ≥ 2. Now suppose ri = 1. Then i has to be even and hence
P i ≥ 2. As v
is not regular and v 6= 1, there is a j with rj 6= 0, whence j6=i rj ≥ 1, and
now (i − 12 )ri + 1 − i ≥ 0.
Suppose now that G = GOε (n, q) with ǫ = ◦, if n odd and ǫ = ± if n
even. The Lie rank ℓ of G is 21 (n − 1) when n is odd and 21 n when n is even.
P
In either case, 12 i iri ≥ ℓ.
As iri rj ≥ ri rj , we note that for |CGOε (n,q) (v)|q = q av we have by
Lemma 5(3)
X
X
1
1 X 1
1
ri (i − )ri − 1 − i +
rj +
av − ℓ ≥
2
2
2
2
i
j6=i
i,ri odd
It suffices to show that for any i the i-th summand is non-negative. This
is
P certainly the case when ri = 0. So suppose ri > 0. If ri ≥
P 2, then as
krk ≥ 4, either i ≥ 2 or there is a j with rj 6= 0, whence j6=i rj ≥ 1.
Thus the i-th summand is non-negative as either (i− 12 )2−1−i = i−2 ≥ 0 or
(i− 12 )2−1−i+1 = i−1 ≥ 0. Now suppose ri = 1. Then i has toP
be odd and,
as v is not regular and v 6= 1, there is a j with rj 6= 0, whence j6=i rj ≥ 1.
Thus the i-th summand of av − ℓ is at least 21 (i − 21 − 1 − i + 1) + 14 = 0.
We now consider the remaining cases. Note that it suffices to consider
the simple group.
G2 (q):
Since p is good, gcd(6, p) = 1 and G contains a unipotent element u whose centraliser has order q 2 and q 2 divides the order of the
centraliser of every other non-trivial unipotent element in G by [28, Table 22.2.6].
F4 (q):
Since p is good, gcd(12, p2 ) = 1 and G contains a unipotent
element u whose centraliser has order q 4 and q 4 divides the order of the
centraliser of every other non-trivial unipotent element in G by [28, Table 22.2.4].
E6 (q), 2 E6 (q):
Since p is good, gcd(6, p) = 1 and G contains a unipotent element u whose centraliser has order q 6 and q 6 divides the order of
the centraliser of every other non-trivial unipotent element in G by [28,
Table 22.2.3].
10
A.ABDOLGHAFOURIAN, M.A. IRANMANESH, AND A.C. NIEMEYER
E7 (q):
Since p is good, gcd(12, p2 ) = 1 and G contains a unipotent
element u whose centraliser has order q 7 and q 7 divides the order of the
centraliser of every other non-trivial unipotent element in G by [28, Table 22.2.2].
E8 (q):
Since p is good, gcd(60, p2 ) = 1 and G contains a unipotent
element u whose centraliser has order q 8 and q 8 divides the order of the
centraliser of every other non-trivial unipotent element in G by [28, Table 22.2.1].
3.2. Non-regular semisimple elements.
Lemma 7. Let G be as in Section 2.1. Then every non-central, non-regular
semisimple element is equivalent to a unipotent element of G.
Proof. Let s be a non-central and non-regular semisimple element in G.
Then s lies in an F -stable torus T of G and CG (s)◦ = hT, Xα | α(s) = 1i,
where the Xα denote the root subgroups with respect to the torus T by
[9, Theorem 3.5.3(i)]. Now CG (s)◦ is a connected reductive and F -stable
[9, p. 28] and CG (s)F = CG (s) (see [10, p. 1]). Moreover, since s is not
regular, there exists a root α with respect to T for which α(s) = 1 by [13,
Proposition 14.6]. Therefore the equivalence class A of Xα determines a
non-trivial unipotent subgroup (XA )F of G by [33, Prop. 23.7, Prop. 23.9].
In particular CG (s) contains a unipotent element.
Our next aim is to show that if xZ(G) is an involution in G/Z(G) then
x is equivalent to a unipotent element in G. For the exceptional groups the
proof relies on the knowledge of their classes of involutions.
Lemma 8. Let G be a finite group of Lie type of rank ℓ as given in Section 2.1. If x is an involution in G then x is equivalent to a unipotent
element in G.
Proof. Suppose first that G is a group of type Aℓ , 2 Aℓ , Bℓ , Cℓ . As q odd, an
element of order a power of two is semisimple. Note also that a non-central
semisimple element x ∈ G is regular if and only if it lies in a unique maximal
torus of G. This is the case if and only if xZ(G) is regular in G/Z(G).
A semisimple element in a classical group H in natural characteristic p is
regular in types Aℓ , 2 Aℓ , Bℓ , Cℓ if and only if its minimal polynomial and its
characteristic polynomial are equal. In case Dℓ a stronger condition holds,
see [20, Theorem 3.2.1]. For all values of ℓ we consider (see Section 2.1)
there are no regular involutions in classical groups. Moreover, there are also
no regular elements of order four whose square is a central involution. Thus
in H or H/Z(H) every involution is equivalent to a unipotent element by
Lemma 7.
Now let G be one of the exceptional or twisted simple groups, namely
G2 (q), F4 (q), E6 (q), E7 (q), E8 (q), 2 E6 (q), 3 D4 (q) where q is a power of an odd
prime. Character tables and conjugacy classes of these finite simple groups of
THE DIVISIBILITY GRAPH OF FINITE GROUPS OF LIE TYPE
11
Lie type have been studied extensively, see for example [11, 12, 14, 30, 32, 34]
and it is known that the order of a centraliser of an involution in these
groups is divisible by q. So by Lemma 4, each involution is equivalent to
some unipotent element.
Lemma 9. Let G be as in Section 2.1 and let s be a semisimple element of
G. If |CG (s)| is even, s is equivalent to a unipotent element.
Proof. As |CG (s)| is even, there is an involution x ∈ CG (s).
If |s| is odd, then some power of sb has odd prime order and thus s ∼
sb ∼ x as xs = sx and gcd(|s|, |x|) = 1 by Lemma 4(1). By Lemma 4(3) sb
and thus also s is equivalent to x. If |s| is even then we may choose x to be
the power sb for which sb is an involution in T 0 . Clearly s is equivalent to
x = sb . By Lemma 8 x is equivalent in G to a unipotent element and thus
so is s.
3.3. Isolated vertices. Williams [37, Lemma 5] calls a subgroup T of a
finite group G a CC-group if CG (x) ≤ T for all x ∈ T \Z(G). If G is a group
we denote by G the group G/Z(G), for H ≤ G we let H = HZ(G)/Z(G) and
for g ∈ G we denote gZ(G) by g. Now suppose G is a finite group of Lie type
T is a maximal torus of G and a CC-group such that gcd(|T |, |Z(G)|) = 1.
Williams proved that the set π of prime divisors of |T | forms a connected
component of the Prime Graph of G. Some general properties of CC-groups
are given in [4, Proposition 1.14], where they are called sharp subgroups. In
particular, a torus which is a CC-group is a Hall π-subgroup of G.
Here we prove that if G is a finite group of Lie type and T is a maximal
torus such that T has odd order and is a CC-group in G, then |sG | for an
s ∈ T \Z(G) forms an isolated vertex of D(G). Let T G = ∪h∈G T h .
If the Prime Graph of G has more than one component, let π1 denote the
component containing the prime 2. It follows from [37, p. 487] the number of
components of the Prime Graph of G is at most that of the Prime Graph of
the simple factor corresponding to G. In the following two lemmas we make
use of a result of Williams [37, Lemma 5] in which he showed in particular
that for a (not necessarily simple) group G as in Section 2.1 and a maximal
torus T in G the primes dividing |T | form a complete component of the
Prime Graph of G not containing the prime 2 if and only if |T | is odd,
coprime to |Z(G)| and T is a CC-group.
Lemma 10. Let G be as in Section 2.1. Let T be a maximal torus in G
for which T has odd order coprime to |Z(G)| and is a CC-group. Then the
non-central elements g of T G form the isolated vertex |g G | = |G|/|T Z(G)|
in D(G).
Proof. Let a = gh ∈ T G \Z(G). Then |CG (a)| = |CG (g h )| = |CG (g)| =
|T Z(G)| for every a ∈ T G \Z(G) and thus the non-central elements in T G
form a single vertex in D(G). It remains to see that this vertex is isolated.
12
A.ABDOLGHAFOURIAN, M.A. IRANMANESH, AND A.C. NIEMEYER
Suppose to the contrary that g ∈ T G is non-central and its conjugacy
class length is not isolated in D(G). Then there is a non-central element
x ∈ G\T G with x ∼ g but |xG | =
6 |g G |. Let π denote the set of prime
divisors of |T |. Note that by [37, Lemma 5] π is a complete component of
the Prime Graph of G containing only odd primes. As x 6∈ T G , the elements
g and x have coprime order.
Let b be a prime with b | |x| and b 6∈ π. Thus b divides |CG (x)| but not
|CG (g)|, hence |CG (g)| | |CG (x)| since x ∼ g. Let r ∈ π and let y ∈ CG (x)
be an element of order r. Then x ∈ CG (y). As gcd(|T |, |Z(G)|) = 1 it
follows that y 6∈ Z(G) and hence x ∈ CG (y). Therefore xy is an element of
order br in CG (y) and r is connected to b in the Prime Graph of G. This
is a contradiction to π being a complete connected component of the Prime
Graph.
Lemma 11. Let G be as in Section 2.1. Let T be a maximal torus in G for
which T is a CC-group and gcd(|T |, |Z(G)|) 6= 1. Then |T | is even.
Proof. Let π denote the set of divisors of |T |. If |T | is even, the result holds.
Seeking a contradiction, we now assume |T | is odd. As gcd(|T |, |Z(G)|) 6= 1,
there is an odd prime r ∈ π dividing |Z(G)|. In particular, we are then in
case Aℓ , 2 Aℓ , E6 (q) or 2 E6 (q). In all of these cases, r divides the centraliser
of an involution in G, see the proof of [37, Lemma 5(d)]. In particular,
there is an element g ∈ G for which g has order 2r. Let x = gr and y = g 2 .
Since T is a CC-group it is a Hall π-subgroup there is an element h ∈ G
such that t = yh ∈ T . In particular, t has order a and commutes with xh
which has order 2. Thus 2 divides |CG (t)| = |T |, a contradiction to our
assumption.
3.4. Regular semi-simple elements. We now consider regular semisimple elements. For such an element s we know that CG (s)◦ = T (see [33,
Cor. 14.10]) and hence (CG (s)◦ )F = TF = T.
We first identify those regular semisimple elements in a finite group G
of Lie type whose centralizers are equal to the maximal torus T . Let k
denote the order of the fundamental group of G. Table 2 yields k (see [9,
pp. 25-26]).
The following lemma is [36, Lemma 4.4], [33, Proposition 14.20] and [9,
p. 25]. For a semisimple element we use the notation CG (s) = CG (s)F and
F
CG (s)◦ = CG (s)◦ .
Lemma 12. Let G be as in Section 2.1 and let s be a semisimple element
in G. Then [CG (s) : CG (s)◦ ] divides the order k of the Fundamental group
as in Table 2. If gcd(|s|, k) = 1 then CG (s) = CG (s)◦ . In particular, if
gcd(|s|, k) = 1 and s is regular semisimple, then the unique maximal torus
T containing s is CG (s).
THE DIVISIBILITY GRAPH OF FINITE GROUPS OF LIE TYPE
13
Dynkin Diagram k
Aℓ
gcd(ℓ + 1, q − 1)
2A
gcd(ℓ + 1, q + 1)
ℓ
Bℓ , Cℓ
2
Dℓ
4
G2 , F4 , E8
1
E6
3
E7
2
Table 2. Orders of the fundamental group
Lemma 13. Let G be as in Section 2.1 and let s be a regular semisimple
element in G. If CG (s) > CG (s)◦ = T then s is equivalent to a unipotent
element in G.
Proof. Suppose CG (s) > CG (s)◦ = T . The result follows from Lemma 9
when |CG (s)| is even. Thus we may assume |CG (s)| is odd. In particular,
[CG (s) : CG (s)◦ ] is odd and also divides the order k of the Fundamental
group as in Table 2. This implies we are either in case Aℓ , 2 Aℓ , E6 or 2 E6 .
If |s| is not prime, then we can choose m ∈ N such that |sm | is prime. If sm
is not regular, then sm , and thus also s, is equivalent to a unipotent element
by Lemma 7. Thus we may assume sm is regular. Then T is the unique torus
containing sm . Moreover, sm ∈ T = CG (s)◦ < CG (s) ≤ CG (sm ). Thus, by
Lemma 12, |sm | is prime divisor of k. Hence, replacing s by sm if necessary,
we may now assume that |s| is an odd prime dividing k.
Suppose first we are in case Aℓ or 2 Aℓ . In this case, a semisimple element
in GLǫ (n, q) is regular if and only if its characteristic polynomial is equal to
its minimal polynomial. As k divides q − ǫ, a regular semisimple element
of order dividing k is similar to a diagonal matrix with pairwise distinct
entries in underlying field. Thus the unique torus inside GL(n, q) or GU(n, q)
containing this element is isomorphic to Znq−ǫ . Inside G it follows that |T | is
even for n ≥ 3, contrary to our assumption that |CG (s)| is odd. Hence this
case cannot arise.
Now consider the case that G = E6 (q) or G = 2 E6 (q). By [33, Example 26.11] CG (s) is one of the centralizers in [33, Table 26.1] and divisible
by q. In particular, s is equivalent to a unipotent element.
Lemma 14. Let G be as in Section 2.1 and let s be a regular semisimple
element in G such that CG (s)◦ a maximal torus which is not a CC-group.
Moreover, suppose CG (s) = CG (s)◦ = T . Then s is equivalent to a unipotent
element.
Proof. For any t ∈ T we have Z(G) ≤ CG (t) and, since T is abelian, T ≤
CG (t). Moreover, CG (t) ≤ CG (t). For t = s this implies T ≤ CG (s) ≤
CG (s) = T and hence CG (s) = T.Z(G) and CG (s) ≤ CG (t) for all t ∈ T ,
implying s ∼ t.
14
A.ABDOLGHAFOURIAN, M.A. IRANMANESH, AND A.C. NIEMEYER
As T is not a CC-group, there is one element t ∈ T such that T < CG (t).
If t is non-regular semisimple, then by Lemma 7 there is a unipotent element
u equivalent to t while for t regular semisimple such a u exists by Lemma 13.
Therefore, s and u are also equivalent.
Lemma 15. Let G be as in Section 2.1. Then every non-central semisimple
element s ∈ G is either equivalent to a unipotent element of G or |sG | is an
isolated vertex of D(G).
Proof. As s ∈ G is semisimple, there exists a maximal torus T of G such
that s ∈ T . We consider the following cases:
Case 1: |CG (s)| is even. Then s is equivalent to a unipotent element
by Lemma 9. From now on we assume that |CG (s)| is odd.
Case 2: T is a CC-subgroup. Then gcd(|Z(G)|, |T |) = 1 by Lemma 11
as in particular |T | is odd. Therefore s is related to an isolated vertex by
Lemma 10. From now on we assume that T is not a CC-group.
Case 3: s is regular semisimple, i.e. CG (s)◦ = T Z(G). If CG (s) >
CG (s)◦ = T Z(G) then s is equivalent to a unipotent element in G by
Lemma 13. If CG (s) = CG (s)◦ = T Z(G) then s is equivalent to a unipotent
element in G by Lemma 14.
Case 4: s is a non-regular semisimple element. By Lemma 7 s is equivalent to a unipotent element.
acknowledgements
We thank Frank Lübeck for providing lists of the generic conjugacy class
sizes of the exceptional groups E7 (q) computed in GAP and for helpful discussions. The third author acknowledges the support of the Australian
Research Council Discovery Project DP140100416.
References
[1] A. Abdolghafourian and M. A. Iranmanesh, Divisibility graph for symmetric and
alternating groups. Comm. Algebra 43 (7): 2852-2862, 2015.
[2] A. Abdolghafourian and M. A. Iranmanesh, On the number of connected components
of divisibility graph for certain simple groups. Transactions on Combinatorics, 5
(2),33-40, 2016.
[3] A. Abdolghafourian and M. A. Iranmanesh, On divisibility graphs for finite sets of
positive integers. Rocky Mountain J. Math., (to appear).
[4] L. Babai, P. Pálfy and J. Saxl, On the number of p-regular elements in finite simple
groups, LMS J. Comput. Math., 12, (2009), 82–119.
[5] Bertram, Edward A. and Herzog, Marcel and Mann, Avinoam, On a graph related
to conjugacy classes of groups, Bull. London Math. Soc., 22(6), (1990), 569–575.
[6] R. Brauer and K. A. Fowler, On groups of even order, Annals of Mathematics 62 (3),
(1955), 565–583.
[7] D. Bubboloni, S. Dolfi , M. A. Iranmanesh and C. E. Praeger, On bipartite divisor
graphs for group conjugacy class sizes. J. Pure Appl. Algebra, 213(9) (2009)1722–
1734.
[8] A. R. Camina and R. D. Camina, The influence of conjugacy class sizes on the
structure of finite groups: a survey, Asian-Eur. J. Math. 4(4), (2011), 559–588.
THE DIVISIBILITY GRAPH OF FINITE GROUPS OF LIE TYPE
15
[9] R. W. Carter, Finite groups of Lie type. Wiley Classics Library. John Wiley & Sons
Ltd., Chichester, (1993). Conjugacy classes and complex characters, Reprint of the
1985 original, A Wiley-Interscience Publication.
[10] R. W. Carter, Centralizers of semisimple elements in the finite classical groups. Proc.
London Math. Soc. 3 42 (1981) 1–41.
[11] B. Chang, The conjugate classes of Chevalley groups of type (G2 ). J. Algebra 9 (2)
(1968) 190–211.
[12] D. I. Deriziotis, The Centralizers of Semisimple Elements of the Chevalley Groups
E7 and E8 . Tokyo J. Math. 6 (1) (1983) 191–216.
[13] F. Digne and J. Michel, Representations of finite groups of Lie type, London Mathematical Society Student Texts, 21, Cambridge University Press, Cambridge, (1991).
[14] D. I. Deriziotis and G. O. Michler, Character table and blocks of finite simple triality
groups 3 D4 (q). Trans. Amer. Math. Soc. 303 (1) (1987) 39–70.
[15] S. Dolfi, Arithmatical conditions on the length of the conjugacy classes of a finite
group. J. Algebra 174 (1995) 753–771.
[16] H. Enomoto, The characters of the finite Chevalley group G2 (q), q = 3f . Jpn. J.
Math. 2 (2) (1976) 191–248.
[17] E. Fisman and Z. Arad, A proof of Szep’s conjecture on nonsimplicity of certain
finite groups. J. Algebra, 108(2) (1987) 340–354.
[18] P. Fleischmann and I. Janiszczak. The semisimple conjugacy classes of finite groups
of Lie type E6 and E7 . Comm. Algebra 21.1 (1993), 93-161.
[19] P. Fleischmann and I. Janiszczak, The semisimple conjugacy classes and the generic
class number of the finite simple groups of lie type E8 . Comm. Algebra 22 (6) (1994),
2221–2303.
[20] J. Fulman, P. M. Neumann and C. E. Praeger. A generating function approach to
the enumeration of matrices in classical groups over finite fields. Mem. Amer. Math.
Soc., 176 (830), (2005).
[21] The GAP Group, GAP–Groups, Algorithms, and Programming, 4.7.8, 2015,
http://www.gap-system.org .
[22] S. M. Goodwin and G. Roehrle. On conjugacy of unipotent elements in finite groups
of Lie type, J. of Group Theory 12 (2) (2009), 235–245.
[23] A. Iranmanesh and A. Jafarzadeh, On the commuting graph associated with the
symmetric and alternating groups, J. Algebra Appl., 7(1) (2008) 129–146.
[24] M. A. Iranmanesh and C. E. Praeger, Bipartite divisor graphs for integer subsets.
Graphs Combin., 26(1) (2010) 95–105.
[25] L. S. Kazarin, On groups with isolated conjugacy classes. Izv. Vyssh. Uchebn. Zaved.
Mat., 25(7) (1981) 40–45.
[26] A. S. Kondrat’ev, Prime graph components of finite simple groups, Math. USSR
Sbornik, 67(1) (1990) 235–247.
[27] M. L. Lewis, An overview of graphs associated with character degrees and conjugacy
class sizes in finite groups. Rocky Mountain J. Math., 38(1) (2008) 175–211.
[28] M. W. Liebeck and G. M. Seitz, Unipotent and nilpotent classes in simple algebraic groups and Lie algebras, Mathematical Surveys and Monographs, 180. American
Mathematical Society, Providence, RI, (2012).
[29] M. S. Lucido, Prime graph components of finite almost simple groups. Rend. Sem.
Mat. Univ. Padova, 102, (1999), 1–22.
[30] C. Jansen, K. Lux, R. Parker and R. Wilson. The Atlas of Brauer Characters, Oxford
University Press, Oxford (1995).
[31] G.L. Morgan and C.W. Parker, The diameter of the commuting graph of a finite
group with trivial centre,J. Algebra, 393, (2013) 41–59.
[32] K. Mizuno, The conjugate classes of Chevalley groups of type E6 , J. Fac. Sci. Univ.
Tokyo 24 (1977) 525–563.
16
A.ABDOLGHAFOURIAN, M.A. IRANMANESH, AND A.C. NIEMEYER
[33] G. Malle and D. Testerman, Linear algebraic groups and finite groups of Lie type,
Cambridge Studies in Advanced Mathematics, 133, Cambridge University Press,
Cambridge (2011).
[34] T. Shoji and N. Iwahori, The conjugacy classes of Chevalley groups of type (F4 ) over
finite fields of characteristic p 6= 2. J. Fac. Sci. Univ. Tokyo 21 (1) (1974) 1–17.
[35] W. A. Simpson and J. Sutherland Frame, The character tables for SU (3, q 2 ), PSL
(3, q), PSU (3, q 2 ). Canad. J. Math 25 (3) (1973) 486–494.
[36] T. A. Springer and R. Steinberg, Conjugacy classes, Seminar on Algebraic Groups and
Related Finite Groups (The Institute for Advanced Study, Princeton, N.J., 1968/69),
Lecture Notes in Mathematics, Vol. 131, Springer, Berlin, (1970).
[37] J. S. Williams, Prime graph components of finite groups. J. Algebra 69 (2) (1981)
487–513.
Adeleh Abdolghafourian
Department of Mathematics,
Yazd University, Yazd, 89195-741,
Iran
[email protected]
Mohammad A. Iranmanesh
Department of Mathematics,
Yazd University, Yazd, 89195-741,
Iran
[email protected]
Alice C. Niemeyer
Lehrstuhl B für Mathematik,
Pontdriesch 10-16, RWTH Aachen University, 52062 Aachen,
Germany
[email protected]
| 4math.GR
|
arXiv:1511.07769v2 [math.GR] 5 Oct 2016
A family of irretractable square-free solutions of
the Yang-Baxter equation
D. Bachiller
F. Cedó
E. Jespers
J. Okniński
Abstract
A new family of non-degenerate involutive set-theoretic solutions of
the Yang-Baxter equation is constructed. All these solutions are strong
twisted unions of multipermutation solutions of multipermutation level
at most two. A large subfamily consists of irretractable and square-free
solutions. This subfamily includes a recent example of Vendramin [38, Example 3.9], who first gave a counterexample to Gateva-Ivanova’s Strong
Conjecture [19, Strong Conjecture 2.28(I)]. All the solutions in this subfamily are new counterexamples to Gateva-Ivanova’s Strong Conjecture
and also they answer a question of Cameron and Gateva-Ivanova [21, Open
Questions 6.13 (II)(4)]. It is proved that the natural left brace structure
on the permutation group of the solutions in this family has trivial socle.
Properties of the permutation group and of the structure group associated
to these solutions are also investigated. In particular, it is proved that the
structure groups of finite solutions in this subfamily are not poly-(infinite
cyclic) groups.
2010 MSC: 16T25, 20E22, 20F16.
Keywords: Yang-Baxter equation, set-theoretic solution, brace.
1
Introduction
Since its appearance in a paper of Yang [39], the Yang-Baxter equation has
become an important equation in mathematical physics and also in quantum
group theory. It has stimulated a lot of activity and led to a diversity of new
methods in several related areas of algebra. Recall that a set-theoretic solution of
the Yang-Baxter equation on a non-empty set X is a bijective map r : X ×X −→
X × X such that
r12 r23 r12 = r23 r12 r23 ,
where rij denotes the map X × X × X −→ X × X × X acting as r on the
(i, j) components and as the identity on the remaining component. Drinfeld, in
[14] suggested that is of interest to study set-theoretic solutions of the quantum
Yang-Baxter equation
R12 R13 R23 = R23 R13 R12 .
1
It is known that if τ : X × X −→ X × X is the twist map τ (x, y) = (y, x), then a
map r : X ×X −→ X ×X is a set-theoretic solution of the Yang-Baxter equation
if and only if R = τ ◦ r is a set-theoretic solution of the quantum Yang-Baxter
equation.
In recent years, a special class of solutions of this type, the non-degenerate
involutive solutions, has received a lot of attention [10, 11, 12, 16, 19, 21, 22, 23,
25, 26, 28, 30]. Also, this class of solutions has connections with many topics in
mathematics, such as semigroups of I-type and Bieberbach groups [23], bijective
1-cocycles [16], radical rings [30], triply factorized groups [36], Hopf algebras [15],
regular subgroups of the holomorf and Hopf-Galois extensions [9, 18], groups of
central type [6, 7].
To study involutive non-degenerate set-theoretic solutions of the Yang-Baxter
equation, Rump introduced in [30] a new algebraic structure, called a brace.
Recall that a left brace is a set B with two binary operations, a sum + and a
product ·, such that (B, +) is an abelian group (the additive group of B), (B, ·)
is a group (the multiplicative group of B) and
a · (b + c) + a = a · b + a · c,
for all a, b, c ∈ B. Rump has begun to develop the theory of braces in a series
of papers [29, 31, 32, 33, 34, 35]. The usefulness of this algebraic structure
to solve problems about this type of solutions of the Yang-Baxter equation is
confirmed by the results proven in [11]. Even more, in [4], the classification of
involutive non-degenerate set-theoretic solutions of the Yang-Baxter equation is
reduced to the classification of left braces. Rump [31] and Bachiller [1] classified
some special classes of left braces. These results indicate that the classification
of arbitrary left braces (even in the finite case) seems to be a very difficult
problem. If B is a finite left brace, then it is known that the multiplicative
group of B is solvable [16]. Using some preliminary ideas of Rump, stated in
[35], and developing new ideas on left braces, it has recently been proven in [2]
that there exist finite p-groups which are not multiplicative groups of finite left
braces. This answers in the negative a question which appears implicitly in [16]
and explicitly in [12]. It is an open problem to characterize the finite solvable
groups which are multiplicative groups of left braces.
A possible strategy to classify finite left braces is the following. First, construct and classify the finite simple left braces. Second, develop the theory of
extensions of finite left braces.
Recall that an ideal of a left brace B is a normal subgroup I of the multiplicative group of B such that
ba − b ∈ I,
for all b ∈ B and all a ∈ I. The socle of a left brace B is
Soc(B) = {a ∈ B | ab = a + b for all b ∈ B}.
It is an ideal of B (see [11, page 107]). One says that the left brace B is simple
if B 6= {1} and {1} and B are the only ideals of B. Rump, in [30], has shown
2
that the only simple finite nilpotent left braces (that is, the multiplicative group
of B is nilpotent) are the cyclic groups Z/(p), with p a prime, and it turns out
that the multiplication of the brace is equal to the sum. Recently Bachiller [3]
has developed a method to construct finite non-nilpotent simple left braces and
has given some families of such braces. To apply this method of constructing
new finite simple left braces it is important to discover new families of finite left
braces with trivial socle. Note that, obviously, any finite non-nilpotent simple
left brace should have trivial socle. Some families of finite left braces with trivial
socle have been given in [9, 24]. The natural structure of a left brace on the
permutation group of a finite irretractable solution yields a class of finite braces
with trivial socle (see Lemma 2.1 below). Therefore, to find new families of
finite irretractable solutions, or, more general, new families of finite left braces
with trivial socle is of interest from the point of view of the classification of
finite left braces.
Another key ingredient of the classification would be the theory of extensions
of left braces. However, very little is known about this (see [3, 8]).
Vendramin in [38, Example 3.9] gives a counterexample to a conjecture of
Gateva-Ivanova [19, Strong Conjecture 2.28(I)], see Section 3, by constructing
an irretractable square-free involutive non-degenerate set-theoretic solution of
the Yang-Baxter equation (X, r) with |X| = 8. It is remarkable that among the
2471 square-free non-degenerate involutive set-theoretic solutions on a set X
with |X| ≤ 8 this is the only counterexample to the Gateva-Ivanova conjecture
(see Remark 3.11 in [38]). Furthermore, studying this example of Vendramin
one can check that it is a strong twisted union of two multipermutation solutions
of multipermutation level two. Thus this yields a negative answer to a question posed by Cameron and Gateva-Ivanova [21, Open Questions 6.13 (II)(4)],
although Vendramin did not notice this fact in [38].
In this paper we construct a large family of irretractable square-free involutive non-degenerate solutions of the Yang-Baxter equation that includes the
example of Vendramin. Thus these solutions are new counterexamples to [19,
Strong Conjecture 2.28(I)]. These solutions are strong twisted unions of multipermutation solutions of multipermutation level 2, corresponding to their orbits
under the action of its permutation group. Hence, these solutions also yield a
negative answer to a question posed by Cameron and Gateva-Ivanova in [21,
Open Questions 6.13 (II)(4)]. The natural structure of left brace on the permutation group of these solutions provides a new family of left braces with trivial
socle. We also study another structure associated to a solution of the YangBaxter equation, the so called structure group (which is a solvable Bieberbach
group if the solution is finite) introduced by Etingof, Schedler and Soloviev [16].
In particular, we prove that these groups are not poly-(infinite cyclic). This is
in contrast with the case of multipermutation solutions, whose structure groups
are always poly-(infinite cyclic).
3
2
Preliminary results
We begin by recalling the necessary terminology and notation. Let X be a nonempty set and r : X × X −→ X × X a map, and write r(x, y) = (σx (y), γy (x)).
Recall that (X, r) is said to be a non-degenerate involutive set-theoretic solution
of the Yang-Baxter equation if and only if the following properties hold.
(1) r2 = idX 2 (r is involutive).
(2) σx , γx ∈ Sym(X), for all x ∈ X (r is non-degenerate).
(3) r12 r23 r12 = r23 r12 r23 .
It is easy to check that (1) and (2) imply γy (x) = σσ−1
(x), for all x, y ∈ X.
x (y)
Convention. By a solution of the YBE we mean a non-degenerate involutive
set-theoretic solution of the Yang-Baxter equation.
A solution (X, r) of the YBE is called square-free if r(x, x) = (x, x) for all
x ∈ X. If r(x, y) = (y, x), i.e. if all σx = idX , then r is called the trivial
solution.
The structure group of a solution (X, r) of the YBE is the group G(X, r) =
hX | xy = zt whenever r(x, y) = (z, t)i. The permutation group of (X, r), denoted G(X, r), is the subgroup of the symmetric group Sym(X) on X generated
by {σx | x ∈ X}.
Etingof, Schedler and Soloviev in [16] proposed the following interesting
operator for studying the structure group G(X, r) and to classify solutions of
the YBE. We recall its definition. Given a solution (X, r) of the YBE, with
r(x, y) = (σx (y), γy (x)), define the equivalence relation ∼ on X by
x ∼ y if and only if σx = σy .
We denote by x̄ the ∼-class of x ∈ X. The retraction Ret(X, r) of (X, r) is
the solution (X̄, r̄), where X̄ = X/ ∼ and r̄(x̄, ȳ) = (σx (y), γy (x)). A solution
(X, r) of the YBE is said to be a multipermutation solution if there exists a
positive integer n such that Retn (X, r) is a solution on a set of cardinality 1.
The multipermutation level of a multipermutation solution (X, r) of the YBE
is the smallest positive integer n such that Retn (X, r) is a solution on a set of
cardinality 1. One says that (X, r) is irretractable if Ret(X, r) = (X, r).
Rump in [30] introduced a new algebraic structure, called a left brace. This
allows another possible strategy to attack the problem of constructing and classifying the solutions of the YBE. Recall that a left brace is a set B with two
operations, an addition + and a multiplication ·, such that (B, +) is an abelian
group, (B, ·) is a group and
a · (b + c) + a = a · b + a · c,
4
for all a, b, c ∈ B. It follows that a · (b − c) = a · b − a · c + a, for all a, b, c ∈ B.
For a ∈ B, we denote by λa the map B −→ B defined by λa (b) = a · b − a,
for all b ∈ B. In fact λa ∈ Aut(B, +), and λ : (B, ·) −→ Aut(B, +), defined by
λ(a) = λa , is a group homomorphism (see [11]). The socle, Soc(B), of a left
brace B is defined as
Soc(B) = {a ∈ B | λa = idB }.
It is an ideal of B, i.e. a normal subgroup of (B, ·) that is invariant under all
maps λa . (For the definitions of a homomorphism of left braces, of a right brace
and of related notions we refer to [11]). Note that the maps λa give a useful
link between the two operations in a left brace B, that is
a · b = a + λa (b)
and
a + b = a · λ−1
a (b),
for all a, b ∈ B. By a subgroup of a left brace B we mean a subgroup of the
multiplicative group of B.
Given a solution (X, r) of the YBE, the groups G(X, r) and G(X, r) each have
a natural left brace structure. The additive group of G(X, r) is the free abelian
group with basis X and λx (y) = σx (y), for all x, y ∈ X ⊆ G(X, r). Furthermore,
the map x 7→ σx extends to an onto (multiplicative) group homomorphism
φ : G(X, r) −→ G(X, r)
and Ker(φ) = Soc(G(X, r)) is an ideal of the left brace G(X, r). Hence
G(X, r) ∼
= G(X, r)/ Soc(G(X, r))
has a natural induced left brace structure (see also [37, Section 3] and [20,
Sections 3 and 5]). It follows that φ : G(X, r) −→ G(X, r) is a homomorphism
of left braces and, for every g ∈ G(X, r), the map φ(g) is the restriction of λg
to X. In particular, φ(a + b) = φ(a) + φ(b), where the latter is the sum taken
in the brace G(X, r).
Lemma 2.1 Let (X, r) be a solution of the YBE such that Ret(X, r) = (X, r).
Then Soc(G(X, r)) = {1}.
(x)), for some σx ∈ Sym(X). Let
Proof. We have that r(x, y) = (σx (y), σσ−1
x (y)
g ∈ Soc(G(X, r)). Then there exist x1 , . . . , xn ∈ X and ε1 , . . . , εn ∈ {−1, 1}
such that g = σxε11 · · · σxεnn and gh − g = h for all h ∈ G(X, r). In particular, for
all z ∈ X we have
σz
=
gσz − g
=
σxε11 · · · σxεnn σz − σxε11 · · · σxεnn
=
=
φ(xε11 · · · xεnn z − xε11 · · · xεnn )
φ(λxε11 ···xεnn (z))
=
σλxε1 ···xεn (z) .
1
n
5
Since, by assumption, Ret(X, r) = (X, r) we get that λxε11 ···xεnn (z) = z, for all
z ∈ X. Thus
g = σxε11 · · · σxεnn = φ(xε11 · · · xεnn ) = idX ,
and therefore Soc(G(X, r)) = {1}.
Clearly, the converse of this result is not true. For example, let X = {1, 2}
and let r : X 2 −→ X 2 be defined by r(x, y) = (y, x). Then (X, r) is a solution
of the YBE, Ret(X, r) 6= (X, r) and Soc(G(X, r)) = {1}. What is true is the
following:
Remark 2.2 If B is a left brace with Soc(B) = {1}, then there exists a solution
(X, r) such that G(X, r) ∼
= B as left braces, and Ret(X, r) = (X, r). Indeed,
consider the associated solution of B: X = B and the map r is given by
r : B×B
(a, b)
−→
7→
B×B
(λa (b), λ−1
λa (b) (a)).
Note that,
G(X, r) = hλa | a ∈ Bi = {λa | a ∈ B} ∼
= B/ Soc(B) = B.
Moreover, Ret(X, r) = (X, r), because if λa1 = λa2 , then λa−1 a1 = id, and since
2
the socle is trivial, a−1
2 a1 = 1.
If B is a finite non-trivial two-sided brace, then Soc(B) 6= {1} [11, Proposition 3]. By Lemma 2.1, any such solution (X, r) with G(X, r) ∼
= B satisfies
Ret(X, r) 6= (X, r). In fact, (X, r) is a multipermutation solution, (see [20,
Corollary 5.17] or the proof of [11, Theorem 3] and the comments after this
proof). Hence to study finite non-multipermutation solutions, one should consider only finite left braces which are not two-sided.
Lemma 2.3 Let (X, r) be a solution of the YBE and let {Xi }i∈I be the family
of all orbits of X under the action of G(X, r). Suppose ≤ is a well-order on I.
For each i ∈ I denote by Gi the subgroup of G(X, r) generated by Xi . Then
(i) Gi is a subbrace of G(X, r), invariant by the action of G(X, r).
(ii) Gi Gj = Gj Gi for all i, j ∈ I.
(iii) Every g ∈ G(X, r) \ {1} has a unique presentation as a product g =
g1 · · · gm , where gj ∈ Gij \ {1}, for j = 1, . . . , m, and i1 < · · · < im
are elements of I. Moreover, g can be presented uniquely as a sum g =
h1 + · · · + hm , where hj ∈ Gij \ {1}, for j = 1, . . . , m, and i1 < · · · < im
are elements of I.
6
Proof. Let G+
i be the additive subgroup of G(X, r) generated by Xi . We shall
prove that G+
i = Gi . Let g ∈ Gi . Then there exist x1 , . . . , xk ∈ Xi and integers
ε1 , . . . , εk ∈ {−1, 1} such that g = xε11 · · · xεkk . We have
= xε11 · · · xεkk
εk−1
+ λxε1 ···xεk−1 (xεkk )
= xε11 · · · xk−1
g
1
k−1
= xε11 + λxε11 (xε22 ) + λxε11 xε22 (xε33 ) + · · · + λxε1 ···xεk−1 (xεkk ).
1
k−1
+
Since x−1
= −λx−1 (xi ), it is clear that Gi ⊆ G+
i
i . Let h ∈ Gi . Then there exist
i
y1 , . . . yt ∈ Xi and integers ν1 , . . . , νt ∈ {−1, 1} such that h = ν1 y1 + · · · + νt yt .
We have
h =
=
=
ν1 y 1 + · · · + νt y t
(ν1 y1 + · · · + νt−1 yt−1 )λ−1
ν1 y1 +···+νt−1 yt−1 (νt yt )
−1
−1
(ν1 y1 )λ−1
ν1 y1 (ν2 y2 )λν1 y1 +ν2 y2 (ν3 y3 ) · · · λν1 y1 +···+νt−1 yt−1 (νt yt ).
+
−1
, it is clear that G+
Since −yi = (λ−1
−yi (yi ))
i ⊆ Gi . Hence Gi = Gi . Therefore
Gi is a subbrace of G(X, r) and clearly it is invariant under the action of G(X, r).
This proves (i).
By (i), we know that λg (Gi ) = Gi , for all g ∈ G(X, r) and i ∈ I. Let g ∈ Gi
and h ∈ Gj . By [11, Lemma 2(i)], gh = λg (h)λ−1
λg (h) (g) ∈ Gj Gi . Therefore
Gi Gj ⊆ Gj Gi . Thus (ii) follows by symmetry.
Therefore, for every g ∈ G(X, r) \ {1}, there exist a positive integer m,
i1 , . . . , im ∈ I and gj ∈ Gij , for j = 1, . . . , m, such that i1 < · · · < im and
′
g = g1 · · · gm . Suppose that g1 · · · gm = g1′ · · · gm
, for gj , gj′ ∈ Gij . Then, by the
above,
g1−1 g1′
′ −1
= g2 · · · gm (gm
) · · · (g2′ )−1
′ −1
) )
= g2 + λg2 (g3 ) + · · · + λg2 ···gm−1 (gm ) + λg2 ···gm ((gm
+
+
′ −1
+ · · · + λg2 ···gm (gm
′ )−1 ···(g ′ )−1 ((g )
) ∈ G1 ∩ (G2 + · · · + G+
2
m ).
3
Since the additive group of G(X, r) is free abelian with basis X, we have that
+
+
′
G+
1 ∩ (G2 + · · · + Gm ) = {0}. Hence g1 = g1 . By induction on m, it follows that
′
gj = gj for all j = 1, . . . , m. Let h1 = g1 and hi = λg1 ···gi−1 (gi ), for 1 < i ≤ m.
We have g = h1 + · · · + hm and hj ∈ Gij \ {1}, for j = 1, . . . , m. Therefore, (iii)
follows.
Let (X, r) be a solution of the YBE. We know that G(X, r) is a group
presented with the set of generators X and with relations xy = zt whenever
r(x, y) = (z, t). Since the relations are homogeneous, the group G(X, r) has
a degree function deg : G(X, r) −→ Z such that deg(x) = 1, for all x ∈ X.
Therefore,
for x1 , x2 , . . . , xm ∈ X and n1 , n2 , . . . , nm ∈ Z, deg(xn1 1 · · · xnmm ) =
Pm
l=1 nl .
7
Remark 2.4 Since (G(X, r), +) ∼
define an additive degree
= Z(X) , we can also
Pm
function deg+ as follows deg+ (n1 x1 +· · ·+nm xm ) = l=1 nl , for x1 , . . . , xm ∈ X
and n1 , . . . , nm ∈ Z. In fact, the two functions coincide.
Indeed, take an arbitrary element g = xε11 · · · xεkk of G(X, r), where x1 , . . . xk ∈
P
X and ε1 , . . . , εk ∈ {−1, 1}. Note that deg(g) = ki=1 εi . As seen before, in a
left brace, we can pass from the multiplicative form to the additive form through
the lambda maps, and this in the following way:
g
= xε11 · · · xεkk
= xε11 + λxε11 (xε22 ) + λxε11 xε22 (xε33 ) + · · · + λxε1 ···xεk−1 (xεkk ).
1
k−1
Since x−1
= −λx−1 (xi ) and λh (x) ∈ X for any h ∈ G(X, r) and any x ∈ X, it
i
i
is clear that there exist y1 , . . . , yk ∈ X such that
g = ε1 y1 + · · · + εk yk .
Pk
So we get deg+ (g) = i=1 εi = deg(g), as claimed.
Now that we know that deg+ = deg, we can prove some properties of this
function: for any g, h ∈ G(X, r),
(a) deg(λg (h)) = deg(h). We use the additive definition deg+ of the function.
Assume h = ε1 x1 + · · · + εk xk , where x1 , . . . , xk ∈ X and ε1 , . . . , εk ∈
{−1, 1}. Then λg (h) = ε1 λg (x1 ) + · · · + εk λg (xk ), and since λg (x) ∈ X for
Pk
any g ∈ G and x ∈ X, deg(λg (h)) = deg+ (λg (h)) = i=1 εi = deg(h).
(b) deg(g +h) = deg(g)+deg(h). This is a direct consequence of the application
of the additive function deg+ .
(c) deg(g · h) = deg(g) + deg(h). This follows from the definition of deg.
We will use all these facts in the proof of the next lemma.
Lemma 2.5 Let (X, r) be a solution of the YBE. Let {Xi }i∈I be the family of
all orbits of X under the action of G(X, r). Let Gi be the subgroup of G(X, r)
generated by Xi . Let ≤ be a well-order on I. Let H = {g1 · · · gn | gl ∈ Gil ,
for i1 < i2 < · · · < in in I and deg(gl ) = 0} = {g1 + · · · + gn | gl ∈ Gil , for
i1 < i2 < · · · < in in I and deg(gl ) = 0}. Then H is an ideal of the left brace
G(X, r).
Proof. The set {g1 · · · gn | gl ∈ Gil , for i1 < i2 < · · · < in in I and deg(gl ) = 0}
is equal to {g1 + · · · + gn | gl ∈ Gil , for i1 < i2 < · · · < in in I and deg(gl ) = 0}
because of Lemma 2.3 and the fact that the multiplicative and the additive
degree coincide. Hence H is well-defined.
By the definition of H in multiplicative terms and Lemma 2.3, it is easy to
see that H is a normal subgroup of G(X, r). Let h ∈ H and g ∈ G(X, r). There
8
exist elements i1 < i2 < · · · < in in I and gl ∈ Gil such that h = g1 + · · · + gn
and deg(gl ) = 0 for l = 1, 2, . . . , n. We have that
λg (h) =
=
λg (g1 + · · · + gn )
λg (g1 ) + λg (g2 ) + · · · + λg (gn ).
Now λg (gl ) ∈ Gl and deg(λg (gl )) = deg(gl ) = 0, so H is λg -invariant and the
assertion follows.
3
The main construction
In this section we construct a new family of irretractable square-free solutions
of the YBE. These will be strong twisted unions of multipermutation solutions
of multipermutation level 2.
Strong twisted unions of solutions of the YBE were introduced in [22, Definition 5.1]. In fact, the original definition only covered the union of two quadratic
sets. The general definition appeared later in [21, Definition 3.5]. Recall that
a solution (X, r) of the YBE is a strong twisted union of a set of solutions
{(Xj , rj )S
| j ∈ J}, with 1 < |J|, if the sets Xj are G(X, r)-invariant subsets of
X, X = j∈J Xj , Xj ∩ Xk = ∅ for j 6= k, rj is the restriction of r to Xj2 and,
for all j, k ∈ J such that j 6= k,
σγx (z) (y) = σz (y)
and γσz (x) (t) = γx (t),
(1)
for all x, y ∈ Xj , z, t ∈ Xk , where r(a, b) = (σa (b), γb (a)), for all a, b ∈ X.
Let A and B be a nontrivial (additive) abelian groups. Let I be a set with
|I| > 1 and let X(A, B, I) = A × B × I. Let ϕ1 : A −→ B be a map of sets such
that ϕ1 (−a) = ϕ1 (a) for all a ∈ A. Let ϕ2 : B −→ A be a homomorphism of
groups. For a ∈ A, b ∈ B and i ∈ I, let σ(a,b,i) : X(A, B, I) −→ X(A, B, I) be
the map defined by
(c, d + ϕ1 (a − c), j) if i = j,
σ(a,b,i) (c, d, j) =
(c + ϕ2 (b), d, j)
if i 6= j,
for all c ∈ A, d ∈ B and j ∈ I. Note that σ(a,b,i) is bijective and
(c, d − ϕ1 (a − c), j) if i = j,
−1
(c, d, j) =
σ(a,b,i)
(c − ϕ2 (b), d, j)
if i 6= j,
for all c ∈ A, d ∈ B and j ∈ I.
Let r : X(A, B, I)2 −→ X(A, B, I)2 be defined by
r((a, b, i), (c, d, j)) = (σ(a,b,i) (c, d, j), σσ−1
(a, b, i)).
(a,b,i) (c,d,j)
For i ∈ I, put Xi = A × B × {i}. Clearly we have that r2 = idX(A,B,I)2 , i.e. r
is involutive.
9
Lemma 3.1 For c ∈ A, d ∈ B and j ∈ I, let γ(c,d,j) : X(A, B, I) −→ X(A, B, I)
be the map defined by
γ(c,d,j) (a, b, i) = σσ−1
(a, b, i),
(a,b,i) (c,d,j)
for all a ∈ A, b ∈ B and i ∈ I. Then γ(c,d,j) is bijective and
−1
γ(c,d,j)
= σ(c,d,j) .
Proof. Note that
σσ−1
(a, b, i) =
(a,b,i) (c,d,j)
(a, b − ϕ1 (c − a), i)
(a − ϕ2 (d), b, i)
if i = j,
if i =
6 j.
Therefore the result follows.
Remark 3.2 Lemma 3.1 means that (X(A, B, I), r) satisfies condition lri (see
[20, Definition 2.6]). By [20, Fact 2.7] every square-free solution of the YangBaxter equation satisfies condition lri. But the converse is not true. By
[20, Fact 2.8], since (X(A, B, I), r) is involutive and satisfies lri, we have that
(X(A, B, I), r) also is cyclic (see [20, Definition 2.6]). In [20] Gateva-Ivanova
continued her systematic study of solutions of the Yang-Baxter equation with
these important conditions.
Theorem 3.3 (X(A, B, I), r) is a solution of the YBE and the following conditions hold.
(i) (X(A, B, I), r) is square-free if and only if ϕ1 (0) = 0.
(ii) If ϕ−1
1 (0) = {0} and ϕ2 is injective, then (X(A, B, I), r) is irretractable.
(iii) Every Xi is invariant under the action of G(X(A, B, I), r), and if ri is
the restriction of r to Xi2 , then (Xi , ri ) is a multipermutation solution of
multipermutation level at most two and (X(A, B, I), r) is a strong twisted
union of the solutions (Xi , ri ).
(iv) If ϕ1 (A) generates B as a group and ϕ2 is surjective, then the orbits for
the action of G(X(A, B, I), r) on X(A, B, I) are Xi , for i ∈ I.
Proof. We know that r is involutive. By Lemma 3.1, r is non-degenerate.
Furthermore
−1
r((a, b, i), (c, d, j)) = (σ(a,b,i) (c, d, j), σ(c,d,j)
(a, b, i)).
(2)
By [11, Proposition 2], to prove that (X(A, B, I), r) is a solution of the YBE it
is enough to check that
σ(a,b,i) σσ−1
(a,b,i)
(c,d,j)
= σ(c,d,j) σσ−1
(c,d,j)
10
(a,b,i) ,
for all a, c ∈ A, b, d ∈ B and i, j ∈ I. Since ϕ1 (−a) = ϕ1 (a) and ϕ2 is a
homomorphism of groups, these equalities follow at once from the following two
formulas, where e ∈ A, f ∈ B and k ∈ I.
and
σ(a,b,i) σσ−1 (c,d,j) (e, f, k)
(a,b,i)
(e, f + ϕ1 (c − e) + ϕ1 (a − e), k)
(e + ϕ2 (d − ϕ1 (a − c)) + ϕ2 (b), f, k)
(e + ϕ2 (b), f + ϕ1 (c − ϕ2 (b) − e), k)
=
(e + ϕ2 (d), f + ϕ1 (a − e − ϕ2 (d)), k)
(e + ϕ2 (d) + ϕ2 (b), f, k)
if i = j = k,
if i = j 6= k,
if i 6= j = k,
if j 6= i = k,
if j 6= i, i 6= k
and j 6= k,
σ(c,d,j) σσ−1 (a,b,i) (e, f, k)
(c,d,j)
(e, f + ϕ1 (a − e) + ϕ1 (c − e), k)
(e + ϕ2 (b − ϕ1 (c − a)) + ϕ2 (d), f, k)
(e + ϕ2 (b), f + ϕ1 (c − e − ϕ2 (b)), k)
=
(e + ϕ2 (d), f + ϕ1 (a − ϕ2 (d) − e), k)
(e + ϕ2 (b) + ϕ2 (d), f, k)
if i = j = k,
if i = j 6= k,
if i 6= j = k,
if j 6= i = k,
if j 6= i, i 6= k
and j 6= k.
Therefore, indeed (X(A, B, I), r) is a solution of the YBE.
(i) Note that σ(a,b,i) (a, b, i) = (a, b + ϕ1 (0), i). Hence (X(A, B, I), r) is
square-free if and only if ϕ1 (0) = 0.
(ii) Suppose that ϕ−1
1 (0) = {0} and that ϕ2 is injective. Let (a, b, i), (c, d, j) ∈
X(A, B, I) be two distinct elements. If i 6= j, then σ(a,b,i) (e, f, i) = (e, f +ϕ1 (a−
e), i) and σ(c,d,j) (e, f, i) = (e + ϕ2 (d), f, i). Since A 6= {0}, we can take e ∈ A
such that ϕ1 (a−e) 6= 0. Therefore, if i 6= j, then σ(a,b,i) 6= σ(c,d,j) . Suppose that
i = j. Then (a, b) 6= (c, d). If a 6= c, then σ(a,b,i) (c, d, i) = (c, d+ ϕ1 (a− c), i) and
σ(c,d,j) (c, d, i) = (c, d, i). Thus, in this case, since ϕ1 (a−c) 6= 0, σ(a,b,i) 6= σ(c,d,j) .
Suppose i = j and a = c. Then b 6= d, and for k ∈ I \ {i} there are equalities σ(a,b,i) (0, 0, k) = (ϕ2 (b), 0, k) and σ(c,d,j) (0, 0, k) = (ϕ2 (d), 0, k), which by
the injectivity of ϕ2 , imply again σ(a,b,i) 6= σ(c,d,j) . Thus we have shown that
σ(a,b,i) = σ(c,d,j) if and only if (a, b, i) = (c, d, j). Therefore Ret(X(A, B, I), r) =
(X(A, B, I), r).
(iii) It follows from the definition of r that each Xi is invariant under the
action of G(X(A, B, I), r). Let σ(a,b,i) |Xi be the restriction of σ(a,b,i) to the set
Xi = A × B × {i}. It is easy to check that
σ(a,b,i) |Xi = σ(c,d,i) |Xi
if a = c. Since
r((a, b, i), (c, d, i)) = ((c, d + ϕ1 (a − c), i), (a, b − ϕ1 (c − a), i)),
one easily gets that Ret(Xi , ri ) is a trivial solution and thus Ret2 (Xi , ri ) is
a solution on a set of cardinality 1. Therefore (Xi , ri ) is a multipermutation
11
solution of multipermutation level at most two. Let i, j be distinct elements
in I. To show that (X(A, B, I), r) is a strong twisted union of the solutions
(Xi , ri ), in view of (1) and (2), we should check that
σσ−1
(a,b,i)
and
(c,d,j) (e, f, i)
= σ(c,d,j) (e, f, i)
−1
σσ−1
(a, b, j) = σ(c,d,i)
(a, b, j),
(e,f,j) (c,d,i)
for all a, c, e ∈ A and b, d, f ∈ B. We have
σσ−1
(a,b,i)
(c,d,j) (e, f, i)
= σ(c−ϕ2 (b),d,j) (e, f, i)
= (e + ϕ2 (d), f, i)
= σ(c,d,j) (e, f, i),
σσ−1
(a, b, j)
(e,f,j) (c,d,i)
=
and
−1
σ(c+ϕ
(a, b, j)
2 (f ),d,i)
= (a − ϕ2 (d), b, j)
−1
= σ(c,d,i)
(a, b, j).
Hence, indeed, (X(A, B, I), r) is a strong twisted union of the solutions (Xi , ri ).
(iv) Suppose that ϕ1 (A) generates B and that ϕ2 is surjective. Let a ∈
A and b ∈ B. There exist a1 , . . . , as ∈ A, d ∈ B and z1 , . . . , zn ∈ Z such
that b = z1 ϕ1 (a1 ) + · · · + zs ϕ1 (as ) and ϕ2 (d) = a. Note that if i 6= k, then
z1
zs
σ(0,d,i) (0, 0, k) = (a, 0, k) and σ(a
· · · σ(a
(a, 0, k) = (a, b, k). Hence
1 +a,0,k)
s +a,0,k)
the orbit of (0, 0, k) is A × B × {k}, and this finishes the proof.
Remark 3.4 For A = B = Z/(2) and I = {1, 2} the solution (X(A, B, I), r)
of the above theorem, with ϕ1 = ϕ2 = idA , is isomorphic to the solution of
[38, Example 3.9]. Recall that two solutions (X, r) and (X ′ , r′ ) of the YBE are
isomorphic if there exists a bijective map η : X −→ X ′ such that
r′ (η(x), η(y)) = (η(σx (y)), η(γy (x))),
where r(x, y) = (σx (y), γy (x)), for x, y ∈ X.
Recall that Gateva-Ivanova conjectured that every finite square-free solution
of the YBE is a multipermutation solution [19, Strong Conjecture 2.28(I)]. The
construction of Vendramin, given in Remark 3.4, was the first counterexample to
this conjecture. In fact, he constructed a family of counterexamples consisting
of extensions of the one given above, i.e. square-free solutions (Y, s) such that
there exists a surjective homomorphism Y −→ X of solutions, where X is the
solution (X(A, B, I), r) given in Remark 3.4. Notice that [11, Lemma 4] implies
that these solutions are not multipermutation solutions. However, we do not
know whether these solutions are irretractable.
Perhaps Gateva-Ivanova expected that her conjecture might be too strong,
because the following related question was formulated in [21].
12
Question 1. [21, Open Questions 6.13 (II)(4)] Let (X, r) be a square-free solution of the YBE. Suppose that (X, r) is a strong twisted union of two solutions
of the YBE. Is it true that (X, r) is a multipermutation solution?
Notice also that there are examples of square-free multipermutation solutions
of the YBE which are not a strong twisted union of two solutions of the YBE,
see [10, Theorem 3.1].
Remark 3.5 The solutions (X(A, B, I), r), with ϕ−1
1 (0) = {0} and ϕ2 injective, are new counterexamples to [19, Strong Conjecture 2.28(I)], and if moreover
|I| = 2, then they answer in the negative Question 1. Note that if ϕ−1
1 (0) = {0},
then σ(0,0,i) |Xi 6= σ(a,0,i) |Xi for all a 6= 0, thus, in this case, the multipermutation level of (Xi , ri ) is 2. The solution constructed by Vendramin [38, Example
3.6] also gives a negative answer to Question 1, but this fact is not noticed in
[38].
4
The permutation group and the structure group
of (X(A, B, I), r)
First we will study the structure of the multiplicative group of the left brace
G(X(A, B, I), r).
Let G and H be two abelian groups, and let W be the set of all functions
f : H −→ G. Then W is an abelian group with the sum f1 + f2 defined by
(f1 + f2 )(h) = f1 (h) + f2 (h), for f1 , f2 ∈ W and h ∈ H. Recall that the
complete wreath product G¯≀H can be defined as the semidirect product W ⋊ H
with respect to the action of H on W defined by (hf )(x) = f (x − h), for
f ∈ W and h ∈ H. The wreath product G ≀ H is defined similarly, but replacing
W by the abelian group W ′ of all functions f : H −→ G such that the set
{h ∈ H : f (h) 6= 0} is finite. Obviously, when H is finite, the complete wreath
product and the wreath product coincide.
Proposition 4.1 The permutation group G(X(A, B, I), r) is isomorphic to a
subgroup of the Cartesian product of |I| copies of the complete wreath product
hϕ1 (A)i¯≀A. In particular, if moreover A and B are finite abelian p-groups, and
I is finite, then G(X(A, B, I), r) is a finite p-group.
Proof. Let B1 = hϕ1 (A)i. Let W be the set of all functions f : A −→ B1 .
Consider the set S = {σ(a,b,i) | a ∈ A, b ∈ B, i ∈ I}. Denote the Cartesian
product of |I| copies of W ⋊ A by (W ⋊ A)I and its elements by ((fj , aj ))j∈I ,
with fj ∈ W and aj ∈ A. For each a ∈ A, let fa ∈ W denote the map defined
by fa (x) = ϕ1 (a − x), for all x ∈ A. We define a map ν : S −→ (W ⋊ A)I by
ν(σ(a,b,i) ) = ((fj , aj ))j∈I , where
fa if j = i,
0
if j = i,
fj =
and aj =
0
if j ∈ I \ {i},
ϕ2 (b) if j ∈ I \ {i}.
13
We claim that ν can be extended to an injective homomorphism
ν : G(X(A, B, I), r) −→ (W ⋊ A)I .
Let a1 , . . . , ar ∈ A, b1 , . . . , br ∈ B, i1 , . . . , ir ∈ I and ε1 , . . . , εr ∈ {−1, 1}. To
prove the claim it is enough to prove that
εr
ε1
= idX(A,B,I)
· · · σ(a
σ(a
r ,br ,ir )
1 ,b1 ,i1 )
if and only if
((fr,j , ar,j )−εr )j∈I · · · ((f1,j , a1,j )−ε1 )j∈I = ((0, 0))j∈I ,
where
fk,j =
f ak
0
if j = ik ,
if j ∈ I \ {ik },
and ak,j =
0
ϕ2 (bk )
if j = ik ,
if j ∈ I \ {ik }.
We know that
((fr,j , ar,j )−εr )j∈I · · · ((f1,j , a1,j )−ε1 )j∈I
−εr − 1
=
ar,j (−εr fr,j ), −εr ar,j
2
j∈I
−ε1 − 1
···
a1,j (−ε1 f1,j ), −ε1 a1,j
2
j∈I
r
r
X
X
X
−ε
−
1
l
=
−εk ak,j
εk ak,j
al,j (−εl fl,j ) , −
2
l=1
l<k≤r
k=1
.
j∈I
On the other hand
εr
ε1
(x, y, j)
· · · σ(a
σ(a
r ,br ,ir )
1 ,b1 ,i1 )
ε
ε1
(x + (1 − δir ,j )εr ϕ2 (br ),
· · · σ(ar−1
= σ(a
r−1 ,br−1 ,ir−1 )
1 ,b1 ,i1 )
ε
ε1
· · · σ(ar−2
= σ(a
r−2 ,br−2 ,ir−2 )
1 ,b1 ,i1 )
y + δir ,j εr ϕ1 (ar − x), j)
r
X
x+
(1 − δik ,j )εk ϕ2 (bk ) ,
k=r−1
y+
r
X
l=r−1
=
..
.
=
x+
r
X
δil ,j εl ϕ1 al − x −
X
l<k≤r
(1 − δik ,j )εk ϕ2 (bk ),
y+
l=1
δil ,j εl ϕ1 al − x −
X
l<k≤r
14
(1 − δik ,j )εk ϕ2 (bk ) , j
k=1
r
X
(1 − δik ,j )εk ϕ2 (bk ) , j ,
where δi,j is the Kronecker delta, that is
1 if i = j,
δi,j =
0 if i 6= j.
Hence
εr
ε1
= idX(A,B,I)
· · · σ(a
σ(a
r ,br ,ir )
1 ,b1 ,i1 )
if and only if
r
X
(1 − δik ,j )εk ϕ2 (bk ) = 0
k=1
and
r
X
l=1
δil ,j εl ϕ1 al − x −
X
l<k≤r
for all j ∈ I and all x ∈ A. Note that
r
X
εk ak,j =
k=1
r
X
(1 − δik ,j )εk ϕ2 (bk ) = 0,
(1 − δik ,j )εk ϕ2 (bk )
k=1
and
−ε
−
1
l
al,j (−εl fl,j ) (x)
−εk ak,j
2
l=1
l<k≤r
r
X
X
−ε
−
1
l
εl fl,j x −
= −
al,j −
−εk ak,j
2
l=1
l<k≤r
r
X
X
ε
+
1
l
= −
δil ,j εl ϕ1 al − x −
al,j −
εk ak,j
2
r
X
X
l=1
=
l<k≤r
r
X
εl + 1
ϕ2 (bl )
−
δil ,j εl ϕ1 al − x − (1 − δil ,j )
2
l=1
X
−
(1 − δik ,j )εk ϕ2 (bk )
l<k≤r
=
−
r
X
l=1
δil ,j εl ϕ1 al − x −
X
l<k≤r
(1 − δik ,j )εk ϕ2 (bk ) ,
where, in the last equality, the term (1 − δil ,j ) εl2+1 ϕ2 (bl ) disappears because,
when (1 − δil ,j ) = 1, then δil ,j = 0, whence the term
X
ε
+
1
l
ϕ2 (bl ) −
(1 − δik ,j )εk ϕ2 (bk )
δil ,j εl ϕ1 al − x − (1 − δil ,j )
2
l<k≤r
15
becomes zero, and does not appear in the sum. Therefore the claim is proved.
With some additional hypothesis, we can determine precisely the permutation group of (X(A, B, I), r). Note that the next result gives examples of
solutions of the YBE with permutation group of arbitrarily large nilpotency
class.
Corollary 4.2 Assume that A = B = Z/(k), k > 1, I is a finite set such that
gcd(|I|−1, k) = 1, ϕ2 is surjective, ϕ1 (0) = 0 and ϕ1 (x) = 1 for any x ∈ A\{0}.
Then, G(X(A, B, I), r) ∼
= (A ≀ A)|I| .
In this case, the derived length of G(X(A, B, I), r) is 2. The permutation
group G(X(A, B, I), r) is nilpotent if and only if k = pα for some prime p, and,
in this case, its nilpotency class is equal to (α(p − 1) + 1)pα−1 . In particular, if
A is of prime order p then the nilpotency class is p.
Proof. First, observe that hϕ1 (A)i = A. By Proposition 4.1 (and its proof), we
know that ν : G(X(A, B, I), r) −→ (A ≀ A)|I| is injective. So, in order to prove
the first claim it is enough to show that ν is surjective. As before, denote by fa ,
a ∈ A, the map given by fa (c) = ϕ1 (a − c), for c ∈ A. First we prove that, by
definition of ϕ1 , the set {fa : a ∈ A} generates the abelian group W consisting
of the maps from A to A. Indeed, if f P
is any map from A to itself, we have
to find
z
∈
Z,
a
∈
A,
such
that
f
(c)
=
a
a∈A za fa (c) for any c ∈ A. Observe
P
P
that a∈A za fa (c) = a∈A\{c} za by the definition of ϕ1 . This is a system of
linear equations with coefficients in Z/(k) with k equations in k variables. The
associated matrix of the system is
0 1 ··· 1
.
..
1 0
. ..
∈ Mk (Z/(k)).
Nk = .
.. . . . . . . 1
1 ···
1 0
One can prove that det(Nk ) = (−1)k−1 (k − 1), which is invertible in Z/(k), so
the system has a solution.
Second, we prove that S = {sj (f, b) : j ∈ I, b ∈ A, f ∈ W }, where sj (f, b) =
((fi , bi ))i∈I is an element of (A ≀ A)|I| defined by
(f, 0) if i = j,
(fi , bi ) =
(0, b) if i ∈ I \ {j},
is a set of generators of (A ≀ A)|I|P
. Given an arbitrary element ((fi′ , ci ))i∈I of
|I|
¯
(A≀A) , consider the equations i∈I\{j} xi = cj , one for each j ∈ I, in the
variables {xi }i∈I . Then this is a system of |I| linear equations in |I| variables,
and its associated matrix is N|I| . We know that det(N|I| ) = (−1)|I|−1 (|I| − 1),
so the system has a solution since, by the hypothesis, gcd(|I| − 1, k) = 1. Thus
16
P
there exist elements {bi }i∈I in A such that (P i∈I bi ) − bj = cj for any j ∈ I.
Define functions fj , j ∈ I, by fj (c) = fj′ (c + 1≤k<j bk ) for any c ∈ A. Then,
s1 (f1 , b1 ) · s2 (f2 , b2 ) · · · s|I| (f|I| , b|I| )
=
((f1 , 0), (0, b1 ), . . . , (0, b1 )) · ((0, b2 ), (f2 , 0), (0, b2 ), . . . , (0, b2 )) · · ·
·((0, b|I| ), . . . , (0, b|I| ), (f|I| , 0))
=
(f1 ,
X
bi − b1 ), (b1 f2 ,
i∈I
X
bi − b2 ), . . . ,
i∈I
((b1 + · · · + b|I|−1 )f|I| ,
X
i∈I
=
((fi′ , ci ))i∈I ,
!
bi − b|I| )
showing that S is a set of generators of (A ≀ A)|I| , as claimed.
Thus, to prove that ν is surjective, it is enough to show that, for any si (f, b)
in S, there exists a τ ∈ G(X(A, B, I), r) such that ν(τ ) = si (f, b). By what we
checked
P at the beginning of this proof, there exist integers za , a ∈ A, such that
f = a∈A za fa . Assume A = {a1 , . . . , ak } with a1 = 0, and denote zi = zai .
On the other hand, choose a ∈ A such that ϕ2 (a) = b (using here that ϕ2 is
surjective). Then, define the element
zk
z1 −1 z2
σ(0,a,i) ∈ G(X(A, B, I), r).
τ = σ(0,0,i)
σ(a2 ,0,i) · · · σ(a
k ,0,i)
Note that, by definition of ν, ν(σ(a,b,i) ) = si (fa , ϕ2 (b)). Moreover, si (f, a) ·
si (g, b) = si (f + g, a + b). These two facts explain the following computation:
ν(τ )
=
ν(σ(0,0,i) )z1 −1 · ν(σ(a2 ,0,i) )z2 · · · ν(σ(ak ,0,i) )zk · ν(σ(0,a,i) )
=
si (f0 , 0)z1 −1 · si (fa2 , 0)z2 · · · si (fak , 0)zk · si (f0 , ϕ2 (a))
!
k
X
zi fai + f0 , ϕ2 (a)
si (z1 − 1)f0 +
=
i=2
=
si (f, b).
Hence, this finally shows that ν is surjective.
At this point, we have proved that G(X(A, B, I), r) ∼
= (A ≀ A)|I| . Observe
that A ≀ A is a semidirect product of two abelian groups, so its derived length is
2. It follows that (A ≀ A)|I| also has derived length equal to 2. Concerning the
nilpotency of G(X(A, B, I), r), recall that a direct product of groups G × H is
nilpotent if and only if G and H are nilpotent. Besides, the results of [5] imply
that a wreath product of two finite groups G and H is nilpotent if and only if G
and H are p-groups for the same prime p. So, in our case, (A ≀ A)|I| is nilpotent
if and only if A = Z/(pα ) for some prime p.
Moreover, it is possible to compute the nilpotency class of this wreath product. We use the following known result (see [27, Theorem 5.1]): the nilpotency
17
class of G ≀ H, where G and H are finite abelian p-groups such that G has
exponent pn and H ∼
= Z/(pβ1 ) × · · · × Z/(pβm ) with β1 ≥ · · · ≥ βm , is equal to
(n − 1)(p − 1)pβ1 −1 + 1 +
m
X
(pβi − 1).
i=1
α
Applying this result to G = H = A = Z/(p ), we get that the nilpotency class
of A ≀ A is equal to
(αp − α + 1)pα−1 .
The nilpotency class of (A ≀ A)|I| is also equal to (αp − α + 1)pα−1 because the
class of a direct product G × H is equal to the maximum of the class of G and
the class of H.
Observe that, in particular, for A = Z/(p), we obtain nilpotency class equal
to p.
By Lemma 2.1, we know that Soc(G(X(A, B, I), r)) = {1} if (X(A, B, I), r)
is irretractable. By Theorem 3.3, this happens if ϕ−1
1 (0) = {0} and ϕ2 is
injective. In view of the strategy explained in the introduction, it would be
interesting to know whether under some conditions on A, B, I, ϕ1 and ϕ2 , the
left brace G(X(A, B, I), r) can be simple.
We do not know any new simple left brace of the form G(X(A, B, I), r) and
it seems difficult to study the ideal structure of G(X(A, B, I), r) in general. The
next result shows that there are some non-trivial ideals in some cases where the
socle is trivial. Although these left braces are not simple, maybe they can be
used to construct new families of simple left braces with the techniques used in
[3], because they have trivial socle.
Proposition 4.3 Assume that A = B = Z/(k), k > 1, I is a finite set such
that gcd(|I| − 1, k) = 1, ϕ2 is surjective, ϕ1 (0) = 0 and ϕ1 (x) = 1 for any
x ∈ A \ {0}. Then G(X(A, B, I), r) is not a simple left brace.
Proof. Let I = {i1 , . . . , in } with |I| = n. Let Gi be the subgroup of the
structure group G(X(A, B, I), r) generated by {(a, b, i) | a, b ∈ A}. By Theorem 3(iv), the sets {(a, b, i) | a, b ∈ A} are the orbits of X(A, B, I) under the action of G(X(A, B, I), r). Thus, by Lemma 2.3, every element g of
G(X(A, B, I), r) can be written uniquely as g = g1 · · · gn , for some gl ∈ Gil . Let
H = {g1 · · · gn | gl ∈ Gil and deg(gl ) = 0}. By Lemma 2.5, H is an ideal of the
left brace G(X(A, B, I), r).
Let φ : G(X(A, B, I), r) −→ G(X(A, B, I), r) be the natural map. We shall
prove that φ(H) is a non-trivial proper ideal of the left brace G(X(A, B, I), r).
Note that
k
(x, y, j) = (x + (1 − δi,j )kϕ2 (b), y + δi,j kϕ1 (a − x), j) = (x, y, j),
σ(a,b,i)
k
for all a, b, x, y ∈ A and all i, j ∈ I. Hence φ((a, b, j)k ) = σ(a,b,i)
= id. We claim
that
Ker(φ)H = gr((a, b, i)k | a, b ∈ A, i ∈ I)H.
18
(3)
It is clear that gr((a, b, i)k | a, b ∈ A, i ∈ I)H ⊆ Ker(φ)H. Let g ∈ Ker(φ).
By Lemma 2.3, there exist unique g1 , . . . , gl ∈ G(X(A, B, I), r) such that g =
g1 · · · gn and gl ∈ Gil , for l = 1, . . . , n. Note that
σ(a,b,i) σ(c,d,i) (x, y, j)
= (x + (1 − δi,j )ϕ2 (d + b), y + δi,j (ϕ1 (c − x) + ϕ1 (a − x)), j)
= σ(c,d,i) σ(a,b,i) (x, y, j).
Therefore (a, b, i)−1 (c, d, i)−1 (a, b, i)(c, d, i) ∈ Ker(φ) ∩ H, for all a, b, c, d ∈ A
and all i ∈ I. Hence, to prove that g ∈ gr((a, b, i)k | a, b ∈ A, i ∈ I)H, we may
assume that every gl is of the form
gl
= (0, 0, il )z0,0,l (0, 1, il )z0,1,l · · · (0, k − 1, il )z0,k−1,l
(1, 0, il )z1,0,l (1, 1, il )z1,1,l · · · (1, k − 1, il )z1,k−1,l
· · · (k − 1, 0, il)zk−1,0,l (k − 1, 1, il )zk−1,1,l · · · (k − 1, k − 1, il )zk−1,k−1,l .
Hence
id = φ(g) =
k−1
Y
zp,q,1
σ(p,q,i
1)
p,q=0
!
k−1
Y
·
zp,q,2
σ(p,q,i
2)
p,q=0
!
···
p,q=0
Therefore, for every x, y ∈ A and j ∈ I, we have
!
n
k−1
X
X
(x, y, j) =
x+
(1 − δl,j )ϕ2
zp,q,l q ,
k−1
X
p,q=0
Thus
zp,q,n
σ(p,q,i
n)
!
.
p,q=0
l=1
y+
k−1
Y
n
X
zp,q,j ϕ1 p − x −
(1 − δl,j )ϕ2
k−1
X
p,q=0
l=1
n
X
l=j+1
zp,q,l q
ϕ2
!
k−1
X
p′ ,q′ =0
zp′ ,q′ ,l q ′ , j .
= 0,
for all j ∈ I. Since A is finite and ϕ2 is a surjective endomorphism of A, we
have that ϕ2 is an automorphism of A, and
n
X
l=1
(1 − δl,j )
k−1
X
zp,q,l q = 0,
p,q=0
for all j ∈ I. Note that the system of linear equations
n
X
(1 − δl,j )xl = 0,
l=1
19
for j ∈ I
over A has only the trivial solution xl = 0 for all l, because gcd(|I| − 1, k) = 1.
Therefore
k−1
X
zp,q,l q = 0,
p,q=0
for all l. On the other hand, we also have that
k−1
n
k−1
X
X
X
zp,q,j ϕ1 p − x −
ϕ2
zp′ ,q′ ,l q ′ = 0,
p,q=0
l=j+1
p′ ,q′ =0
for all x ∈ A and all j ∈ I. Therefore
k−1
X
zp,q,j ϕ1 (p − x) = 0,
p,q=0
for all x ∈ A and all j ∈ I. Since ϕ1 (p − x) = 1 − δp,x , and the system of linear
equations
k−1
X
(1 − δp,x )tp = 0 for x ∈ A
p=0
over A has only the trivial solution tp = 0 for all p ∈ A, we get that
k−1
X
zp,q,j = 0 ∈ A,
q=0
Pk−1
for all p ∈ A and all j ∈ I. Hence deg(gl ) = p,q=0 zp,q,l = kzl for some integer
zl . Thus (0, 0, il )−kzl gl ∈ H, for all l. Hence
g = g1 · · · gn ∈ gr((a, b, i)k | a, b ∈ A, i ∈ I)H
and this proves the claim (3). Now it is clear that σ(0,0,i1 ) ∈
/ φ(H). Hence φ(H)
is a proper ideal of G(X(A, B, I), r). Since
−1
−1
(0, 0, i1 ) = (0, 1, i1 ) 6= (0, 0, i1 )
σ(0,0,i1 ) (0, 0, i1 ) = σ(1,0,i
σ(1,0,i
1)
1)
−1
and id 6= σ(1,0,i
σ(0,0,i1 ) ∈ φ(H), the result follows.
1)
Consider now the structure group G(X(A, B, I), r) of the solution (X(A, B, I), r)
of Theorem 3.3. Suppose that A, B and I are finite. Hence X(A, B, I) is finite and by [26, Corollary 8.2.7], G(X(A, B, I), r) is solvable and a Bieberbach
group, i.e. a finitely generated torsion-free abelian-by-finite group, see [13]. It
would be interesting to characterize when the structure group of a solution is
poly-(infinite cyclic). Recall that a multipermutation solution of the YBE on a
finite set X has a structure group that is poly-(infinite cyclic), see [26, Proposition 8.2.12]. It remains an open question whether the converse holds. Farkas in
20
[17, Theorem 23] showed that a Bieberbach group is poly-(infinite cyclic) if and
only if every non-trivial subgroup has a non-trivial center. This is one of the
key ingredients of the proof of the following result. Another key ingredient is
based on the application of the natural structure of a left brace on the structure
group of a solution of the YBE, explained in Section 2.
Theorem 4.4 If A, B and I are finite, ϕ1 (0) = 0, ϕ1 (A) generates B and ϕ2
is an isomorphism, then G(X(A, B, I), r) is not a poly-(infinite cyclic) group.
Proof. Let Gi be the subgroup of G(X(A, B, I), r) generated by {(a, b, i) | a ∈
A, b ∈ B}. By Theorem 3(iv), the sets {(a, b, i) | a ∈ A, b ∈ B} are the
orbits of X(A, B, I) under the action of G(X(A, B, I), r). Let I = {i1 , . . . , in }
with |I| = n. Thus, by Lemma 2.3, every element g of G(X(A, B, I), r) can be
written uniquely as g = g1 + · · · + gn , for some gl ∈ Gil . Let H = {g1 + · · · + gn |
gl ∈ Gil and deg(gl ) = 0}. By Lemma 2.5, H is an ideal of G(X(A, B, I), r)
and it is easy to see that G(X(A, B, I), r)/H ∼
= Zn .
To prove that G(X(A, B, I), r) is not poly-(infinite cyclic), it is sufficient to
show that Z(H) = {1} (see [17, Theorem 23]).
We will show now that indeed Z(H) = {1}. Suppose h ∈ Z(H). Then λh
has finite order, since G(X(A, B, I), r) is a finite group. Let s be the order of
λh . So hs ∈ Z(H) and λhs = id. The group G(X(A, B, I), r) is torsion-free,
therefore, replacing h by hs , we may assume that λh = id. Let g ∈ H. Then
λg (h) = gh − g = hg − g = λh (g) + h − g = g + h − g = h.
Let hl ∈ Gil be such that h = h1 + · · · + hn . Then λg (h) = λg (h1 ) + · · · +
λg (hn ). By Lemma 2.3, Gi is invariant under the action of G(X(A, B, I), r).
Hence λg (hl ) ∈ Gil for all l. Since h = λg (h), comparing their decompositions as
sums of element of the subgroups Gil , we have, by Lemma 2.3, that λg (h1 ) = h1 .
We know that the additive group of G(X(A, B, I), r) is free abelian with basis
X(A, B, I). Thus, we may assume that
h1 = n1 (a1 , b1 , i1 ) + · · · + nm (am , bm , i1 ),
(4)
where (a1 , b1 , i1 ), . . . , (am , bm , i1 ) are m distinct elements of X(A, B, I) and
P
m
l=1 nl = 0. By the hypothesis, ϕ1 (A) generates B, so for every l there exist
c1 , . . . , cs ∈ A and z1 , . . . zs ∈ Z such that bl − b1 = z1 ϕ1 (c1 ) + · · · + zs ϕ1 (cs ).
Let
f
=
(al , bl , i1 )−z1 −...−zs (al + c1 , 0, i1 )z1 · · · (al + cs , 0, i1 )zs
·(0, 0, i2 )−1 (0, ϕ−1
2 (al − a1 ), i2 ).
We have that f ∈ H. Hence, by the above, λf (h1 ) = h1 and
λf ((a1 , b1 , i1 ))
=
zs
−z1 −...−zs z1
σ −1 (al −a1 ),i2 ) ((a1 , b1 , i1 ))
σ −1
σ(al +c1 ,0,i1 ) · · · σ(a
σ(a
l +cs ,0,i1 ) (0,0,i2 ) (0,ϕ
l ,bl ,i1 )
=
zs
−z1 −...−zs z1
((al , b1 , i1 ))
σ(al +c1 ,0,i1 ) · · · σ(a
σ(a
l +cs ,0,i1 )
l ,bl ,i1 )
=
−z1 −...−zs
((al , bl , i1 ))
σ(a
l ,bl ,i1 )
=
(al , bl , i1 ).
2
21
Pm
Hence, from (4) we get that n1 = nl , for all l = 1, . . . , m. Since l=1 nl = 0,
it follows that nl = 0 for all l, and thus h1 = 1. Similarly one can prove that
h2 = · · · = hn = 1, and therefore h = 1, as desired.
The assertion of Theorem 4.4 also seems to be of interest from the point of
view of the Kaplansky conjecture on non-existence of nontrivial units in group
algebras of torsion-free groups. The structure groups G(X(A, B, I), r) provide
some natural nontrivial examples for testing this conjecture.
Acknowledgments
The two first-named authors were partially supported by the grants DGI MICIIN
MTM2011-28992-C02-01, and MINECO MTM2014-53644-P. The third author
is supported in part by Onderzoeksraad of Vrije Universiteit Brussel and Fonds
voor Wetenschappelijk Onderzoek (Belgium). The fourth author is supported
by the National Science Centre grant 2013/09/B/ST1/04408 (Poland).
References
[1] D. Bachiller, Classification of braces of order p3 , J. Pure Appl. Algebra 219
(2015), 3568–3603.
[2] D. Bachiller, Counterexample to a conjecture about braces, J. Algebra 453
(2016), 160–176.
[3] D. Bachiller, Examples of simple left braces, arXiv:1511.08477v1[math.GR].
[4] D. Bachiller, F. Cedó and E. Jespers, Solutions of the Yang-Baxter equation
associated with a left brace, J. Algebra 460 (2016), 80–102.
[5] G. Baumslag, Wreath products and p-groups, Proc. Cambridge Philos. Soc.
55 (1959), 224–231.
[6] N. Ben David, On groups of central type and involutive Yang-Baxter
groups: a cohomological approach, Ph.D. thesis, The Technion-Israel Institute of Technology, Haifa, 2012.
[7] N. Ben David and Y. Ginosar, On groups of central type, non-degenerate
and bijective cohomology classes, Israel J. Math. 172 (2009), 317–335.
[8] N. Ben David and Y. Ginosar, On groups of I-type and involutive YangBaxter groups, J. Algebra 458 (2016), 197–206.
[9] F. Catino and R. Rizzo, Regular subgroups of the affine group and radical
circle algebras, Bull. Aust. Math. Soc. 79 (2009), no. 1, 103–107.
[10] F. Cedó, E. Jespers and J. Okniński, Retractability of the set theoretic
solutions of the Yang-Baxter equation, Adv. Math. 224 (2010), 2472–2484.
22
[11] F. Cedó, E. Jespers and J. Okniński, Braces and the Yang-Baxter equation,
Commun. Math. Phys. 327 (2014), 101–116.
[12] F. Cedó, E. Jespers and Á. del Rı́o, Involutive Yang-Baxter groups, Trans.
Amer. Math. Soc. 362 (2010), 2541–2558.
[13] L.S. Charlap, Bieberbach Groups and Flat Manifolds, Springer-Verlag, New
York, 1986.
[14] V. G. Drinfeld, On unsolved problems in quantum group theory. Quantum
Groups, Lecture Notes Math. 1510, Springer-Verlag, Berlin, 1992, 1–8.
[15] P. Etingof and S. Gelaki, A method of construction of finite-dimensional triangular semisimple Hopf algebras, Mathematical Research Letters 5 (1998),
551–561.
[16] P. Etingof, T. Schedler, A. Soloviev, Set-theoretical solutions to the quantum Yang-Baxter equation, Duke Math. J. 100 (1999), 169–209.
[17] D. R. Farkas, Crystallographic groups and their mathematics, Rocky Mountain J. Math. 11 (1981), 511–551.
[18] S.C. Featherstonhaugh, A. Caranti and L.N. Childs, Abelian Hopf Galois
structures on prime-power Galois field extensions, Trans. Amer. Math. Soc.
364 (2012), 3675–3684.
[19] T. Gateva-Ivanova, A combinatorial approach to the set-theoretic solutions
of the Yang-Baxter equation, J. Math. Phys. 45 (2004), 3828–3858.
[20] T. Gateva-Ivanova, Set-theoretic solutions of the Yang-Baxter equation,
braces, and symmetric groups, arXiv:1507.02602v2[math.QA].
[21] T. Gateva-Ivanova and P. Cameron, Multipermutation solutions of the
Yang-Baxter equation, Comm. Math. Phys. 309 (2012), 583–621.
[22] T. Gateva-Ivanova and S. Majid, Matched pairs approach to set theoretic
solutions of the Yang-Baxter equation, J. Algebra 319 (2008), no. 4, 1462–
1529.
[23] T. Gateva-Ivanova and M. Van den Bergh, Semigroups of I-type, J. Algebra
206 (1998), 97–112.
[24] P. Hegedüs, Regular subgroups of the affine group, J. Algebra 225 (2000),
740–742.
[25] E. Jespers and J. Okniński, Monoids and groups of I-type, Algebr. Represent. Theory 8 (2005), 709–729.
[26] E. Jespers and J. Okniński, Noetherian Semigroup Algebras, Springer, Dordrecht 2007.
23
[27] H. Liebeck, Concerning nilpotent wreath products, Proc. Cambridge Philos. Soc. 58 (1962), 443–451.
[28] W. Rump, A decomposition theorem for square-free unitary solutions of
the quantum Yang-Baxter equation, Adv. Math. 193 (2005), 40–55.
[29] W. Rump, Modules over braces, Algebra Discrete Math. (2006) no. 2, 127–
137.
[30] W. Rump, Braces, radical rings, and the quantum Yang-Baxter equation,
J. Algebra 307 (2007), 153–170.
[31] W. Rump, Classification of cyclic braces, J. Pure Appl. Algebra 209 (2007),
671–685.
[32] W. Rump, Generalized radical rings, unknotted biquandles, and quantum
groups, Colloq. Math. 109 (2007), 85–100.
[33] W. Rump, Semidirect products in algebraic logic and solutions of the quantum Yang-Baxter equation, J. Algebra Appl. 7 (2008), 471–490.
[34] W. Rump, Addendum to “Generalized radical rings, unknotted biquandles,
and quantum groups” (Colloq. Math. 109 (2007), 85–100), Colloq. Math.
117 (2009), 295–298.
[35] W. Rump, The brace of a classical group, Note Math. 34 (2014), 115–144.
[36] Y. P. Sysak, Product of group and the quantum Yang-Baxter equation,
notes of a talk in Advances in Group Theory and Applications, 2011, Porto
Cesareo.
[37] M. Takeuchi, Survey on matched pairs of groups. An elementary approach
to the ESS-LYZ theory, Banach Center Publ. 61 (2003), 305–331.
[38] L. Vendramin, Extensions of set-theoretic solutions of the Yang-Baxter
equation and a conjecture of Gateva-Ivanova, J. Pure Appl. Algebra 220
(2016), no. 5, 2064–2076.
[39] C.N. Yang, Some exact results for the many-body problem in one dimension
with repulsive delta-function interaction, Phys. Rev. Lett. 19 (1967), 1312–
1315.
24
D. Bachiller
Departament de Matemàtiques
Universitat Autònoma de Barcelona
08193 Bellaterra (Barcelona), Spain
[email protected]
F. Cedó
Departament de Matemàtiques
Universitat Autònoma de Barcelona
08193 Bellaterra (Barcelona), Spain
[email protected]
E. Jespers
Department of Mathematics
Vrije Universiteit Brussel
Pleinlaan 2, 1050 Brussel, Belgium
[email protected]
J. Okniński
Institute of Mathematics
Warsaw University
Banacha 2, 02-097 Warsaw, Poland
[email protected]
25
| 4math.GR
|
Adversarial Vulnerability of Neural Networks Increases with Input Dimension
Carl-Johann Simon-Gabriel 1 2 Yann Ollivier 1 Bernhard Schölkopf 2 Léon Bottou 1 David Lopez-Paz 1
arXiv:1802.01421v1 [stat.ML] 5 Feb 2018
Abstract
Over the past four years, neural networks have
proven vulnerable to adversarial images: targeted
but imperceptible image perturbations lead to
drastically different predictions. We show that adversarial vulnerability increases with the gradients
of the training objective when seen as a function
of the inputs. For most current network architectures, we prove that the `1 -norm of these gradients grows as the square root of the input-size.
These nets therefore become increasingly vulnerable with growing image size. Over the course of
our analysis we rediscover and generalize doublebackpropagation, a technique that penalizes large
gradients in the loss surface to reduce adversarial
vulnerability and increase generalization performance. We show that this regularization-scheme
is equivalent at first order to training with adversarial noise. Finally, we demonstrate that replacing strided by average-pooling layers decreases
adversarial vulnerability. Our proofs rely on the
network’s weight-distribution at initialization, but
extensive experiments confirm their conclusions
after training.
1. Introduction
Since the trendsetting paper of Goodfellow et al. (2014),
Convolutional Neural Networks (CNNs) have been found
utterly vulnerable to adversarial examples: an adversary
can drive the performance of state-of-the art CNNs down
to chance-level with imperceptible changes of the inputs.
A number of studies have tried to address this issue, but
only few have stressed that, because adversarial examples
are essentially small input changes that create large output
variations, they are inherently caused by large gradients of
the neural network with respect to its inputs. Two notable
exceptions are Hein and Andriushchenko (2017) and Cisse
et al. (2017), who offer adversarial robustness guarantees
1
Facebook AI Research, Paris/New York 2 Empirical Inference Department, Max Planck Institute for Intelligent Systems,
Tübingen, Germany. Correspondence to: Carl-Johann SimonGabriel <[email protected]>.
that do depend on these gradients.
Contributions. In the same spirit, we present both theoretical arguments and an empirical one-to-one relationship
between the gradient norm of the training objective and adversarial vulnerability. By evaluating those norms based on
the weight-statistics at initialization, we show that, by design, CNNs exhibit increasingly large gradients with input
dimension d, which leaves them more and more vulnerable to adversarial noise. This vulnerability is dampened by
average-pooling layers, but not by strided and max-pooling
ones. Based on the link between large loss-gradients and
adversarial vulnerability, we propose a regularizer that explicitely penalizes the norm of these gradients – a technique
already proposed by Drucker and LeCun (1991) under the
name of double-backpropagation. Just as Tikhonov regularization is equivalent to training with random noise (Bishop,
1995), we show that our proposed regularization-scheme is
equivalent at frist order to training with adversarial noise,
as proposed by Goodfellow et al. (2014). We confirm our
theoretical results by extensive experiments. Overall, our
findings call for a new line of research in the design of neural network architectures with inherently smaller gradients,
as pioneered by Cisse et al. (2017). Our results may thereby
serve as guidelines to practitioners and network-designers.
2. From Adversarial Examples to Large
Gradients
Suppose that a given classifier ϕ classifies an image x as
being in category ϕ(x). By definition, an adversarial image
is a small modification of x, barely noticeable to the human
eye, that suffices to fool the classifier into predicting a class
different from ϕ(x). It is a small perturbation of the inputs,
that creates a large variation of outputs. Adversarial examples are thus seem inherently related to large gradients of
the network. A connection, that we will now clarify.
Adversarial Vulnerability and its Variations. In practice, an adversarial image is constructed by adding a perturbation δ to the original image x such that kδk ≤ for some
(small) number and a given norm k·k over the input space.
We call the perturbed input x + δ an -sized k·k-attack and
say that the attack was successfull when ϕ(x + δ) 6= ϕ(x).
This motivates
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
Definition 1. Given a distribution P over the input-space,
the (total) adversarial vulnerability of a classifier ϕ to an
-sized k·k-attack is the probability that there exists a perturbation δ of x such that
kδk ≤ and
ϕ(x) 6= ϕ(x + δ) .
(1)
When the classifier ϕ classifies all test images x drawn from
P with 100% accuracy, the adversarial vulnerability of ϕ
is exactly the average (for x ∼ P ) increase ∆L0/1 of the
induced zero-one-loss L0/1 (ϕ(x), c) after adversarial perturbation of the test inputs. Here, c designates the true label
of x; and as we shall only consider induced losses, let us
agree to simply call them ‘the loss’ and write for instance
L0/1 (x, c). This equality between the average loss increase
and adversarial vulnerability is only an approximate one if
the classifier has only 90% accuracy on P , because switches
between two wrong classes will be ignored and switchees
from wrong to true classes may compensate switches in the
other direction. In practice however, a practitionner will
be more interested in the accuracy-drop after adversarial
attacks on a trained and accurate classifier, than in the exact
count of class-switches. So we may actually call this average of ∆L0/1 the relevant adversarial vulnerability and
remember that, with rising accuracy, it is an increasingly
good approximation of the total adversarial vulnerability.
In practice, we usually do not train our classifiers with the
zero-one loss. We replace it by a smoother loss L, which
we consider a good enough approximation of L0/1 – for
example the cross-entropy (induced)-loss. But if we agree
that L is a good enough approximation of L0/1 , then we
also ought to agree that the (non-infinitesimal) variations of
L accurately reflect those of L0/1 ; and hence, that the average (for x ∼ P ) of ∆L(x, c) faithfully approximates the
relevant adversarial vulnerability. Consequently, a trained
classifier ϕ can only be robust to adversarial examples if,
on average over x, a small adversarial perturbation δ of x
creates only a small variation δL of the loss. But if kδk ≤ ,
then using a first order Taylor expansion in shows that
δL = max |L(x + δ, c) − L(x, c)|
δ : kδk≤
≈ max |∂x L · δ| = |||∂x L|||,
(2)
δ : kδk≤
where ∂x L denotes the gradient of L with respect to x, and
where the last equality is almost the definition of the dual
norm |||·||| of k·k. Now two remarks. First: the dual norm
only kicks in because we let the input noise δ optimally adjust to the coordinates of ∂x L within its -constraint. This
is the brandmark of adversarial noise: the different coordinates add up, instead of statiscally canceling each other
out as they would with random noise. For example, if we
impose that kδk2 ≤ , then δ will stricly align with ∂x L.
If instead kδk∞ ≤ , then δ will align with the sign of the
coordinates of ∂x L. Second remark: the Taylor expansion
in (2) may actually be dominated by higher-order terms.
Nevertheless, the first-order term shows that if our loss surface has large gradients, then our net will be vulnerable to
adversarial attacks. This explains the importance of:
Lemma 1. At first order approximation in , an -sized adversarial attack generated with norm k·k increases the loss
L at point x by |||∂x L|||, where |||·||| is the dual norm of k·k.
Consequently, we can evaluate the adversarial vulnerability
of a trained classifier to -sized k·k-attacks by Ex |||∂x L|||.
In particular, an -sized `p -attack increases the loss
by k∂x Lkq where 1 ≤ p ≤ ∞ and 1/p + 1/q = 1.
The second paragraph of Lemma 1 follows from the wellknown property that the dual norm of an `p -norm is the
corresponding `q -norm.
A new old regularizer. Since Lemma 1 shows that the
loss of the network after an /2-sized k·k-attack is
L,|||·||| (x, c) := L(x, c) + |||∂x L||| ,
(3)
2
it is natural to take this loss-after-attack as a new training objective. Here we introduced a factor 2 for reasons
that will become clear in a moment. Incidentally, for
k·k = k·k2 , this new loss reduces to an old regularizationscheme proposed by Drucker and LeCun (1991) called
double-backpropagation. At the time, the authors argued
that slightly decreasing a function’s or a classifier’s sensitivity to input perturbations should improve generalization.
In a sense, this is exactly our motivation when defending
against adversarial examples. It is thus not surprising to
end up with the same regularization term. Note that our
reasoning only shows that training with one specific norm
|||·||| in Equation 3 helps to protect against adversarial examples generated from k·k. A priori, we do not know what
will happen for attacks generated with another norm; but
our experiments suggest that training with one norm also
protects against other attacks (see Section 4.2).
Link to adversarially-augmented training. In Equation 1, designates an attack-size threshold, while in (3), it
is a regularization-strength. Rather than a notational conflict,
this reflects an intrinsic duality between two complementary
interpretations of , which we now investigate further. Suppose that, instead of using the loss-after-attack, we augment
our training set with -sized k·k-attacks x + δ, where for
each training point x, the perturbation δ is generated on
the fly to locally maximize the loss-increase. Then we are
effectively training with
1
(L(x, c) + L(x + δ, c)) ,
(4)
2
where by construction δ satisfies Equation 2. We will refer
to this technique as adversarially augmented training. It
L̃,k·k (x, c) :=
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
was first introduced by Goodfellow et al. (2014) with k·k =
k·k∞ under the name of FGSM1 -augmented training. Using
the first order Taylor expansion in of Equation 2, this ’oldplus-post-attack’ loss of Equation 4 simply reduces to our
loss-after-attack.
Proposition 2. Up to first-order approximations in ,
L̃,k·k = L,|||·||| .
For small enough , adversarially-augmented training with
-sized k·k-attacks amounts to penalizing the dual norm |||·|||
of ∂x L with weight /2.
In particular, double-backpropagation corresponds to training with `2 -attacks, while FGSM-augmented training corresponds to an `1 -penalty on ∂x L.
This correspondance between training with preturbations
and using a regularizer can be compared to Tikhonov regularization: Tikhonov regularization amounts to training with
random noise (Bishop, 1995), while training with adversarial noise amounts to penalizing ∂x L.
Calibrating the threshold to the attack-norm k·k. Going back to Lemma 1, we see that adversarial vulnerability
depends on three main factors:
(i) k·k , the norm chosen for the attack
(ii) , the size of the attack, and
(iii) Ex |||∂x L||| , the expected dual norm of ∂x L.
(5)
where ∞ denotes a dimension-independent constant. In
Appendix C we show that this scaling also preserves the average signal-to-noise ratio kxk2 / kδk2 , both accross norms
and dimensions, so that p could correspond to a constant
human perception-threshold. With this in mind, we now turn
to our main contribution, Point (iii): estimating Ex k∂x Lkq
for standard neural nets.
1
FGSM = Fast Gradient Sign Method
In this section, we evaluate the size of k∂x Lkq for standard
neural network architectures. We analyse fully-connected
networks, followed by a more general theorem that in particular encompasses CNNs with or without strided convolutions. We finish by showing that, contrary to strided layers,
average-poolings efficiently decrease the norm of ∂x L.
We start our analysis by showing how changing q affects the
size of k∂x Lkq . Suppose for a moment that the coordinates
of ∂x L have typical magnitude |∂x L|. Then k∂x Lkq scales
like d1/q |∂x L|. Consequently
p k∂x Lkq ∝ p d1/q |∂x L| ∝ d |∂x L| .
(6)
This equation carries two important messages. First, we see
how k∂x Lkq depends on d and q. The dependance seems
highest for q = 1. But once we account for the varying
perceptibility threshold p ∝ d1/p , we see that adversarial
vulnerability scales like d · |∂x L|, whatever `p -norm we use.
Second, Equation 6 shows that to be robust against any type
of `p -attack at any input-dimension d, the average absolute
value of the coefficients of ∂x L must grow slower than 1/d.
Now, here is the catch, which brings us to our core insight.
3.1. Core Idea: One Neuron with Many Inputs
We could see Point (i) as a measure of our sensibility to
image perturbations, (ii) as our sensibility threshold, and
(iii) as the classifier’s expected marginal sensibility to a
unit perturbation. Ex |||∂x L||| hence intuitively captures the
discrepancy between our perception (as modeled by k·k) and
the classifier’s perception for an input-perturbation of small
size . Of course, this viewpoint supposes that we actually
found a norm k·k (or more generally a metric) that faithfully
reflects human perception – a project in its own right, far
beyond the scope of this paper. However, it is clear that
the threshold that we choose should depend on the norm
k·k and hence on the input-dimension d. In particular, for
a given pixel-wise order of magnitude of the perturbations
δ, the `p -norm of the perturbation will scale like d1/p . This
suggests to write the threshold p used with `p -attacks as:
p = ∞ d1/p ,
3. Evaluating Adversarial Vulnerability by
Estimating k∂x Lkq
In order to preserve the activation variance of the neurons
from layer to layer, the neural weights are usually initialized
with a variance that is inversely proportional to the number
of inputs per neuron. Imagine for a moment that the network
consisted only of one output neuron o linearly connected
to all input pixels. For the purpose of this example, we
assimilate o and L. Because we initialize the weights with a
variance of 1/d,
√ their average absolute value |∂x o| ≡ |∂x L|
grows like 1/ d, rather than the required 1/d. By Equation 6, the adversarial vulnerability k∂x okq ≡ k∂x Lkq
√
√
therefore increases like d/ d = d.
This toy example shows that the standard initialization
scheme, which preserves the variance from layer to layer,
√
causes the average coordinate-size |∂x L| to grow like 1/ d
instead of 1/d. When an `∞ -attack tweaks its -sized inputperturbations to align with the coordinate-signs of ∂x L, all
coordinates of ∂x L add up in absolute√value, resulting in an
output-perturbation that scales like d and leaves the network increasingly vulnerable with growing input-dimension.
3.2. Generalization to Deep Networks
Our next theorems generalize the previous toy example to
a very wide class of feedforward nets with ReLU activation functions. For illustrational purposes, we start with
fully connected nets and only then proceed to the broader
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
class, which includes any succession of convolutional and/or
strided layers. In essence, the proofs iterate our insight on
one layer over a sequence of layers. They all rely on the
following set (H) of hypotheses:
H1 There is 1 ReLU after each non-input neuron, which
kills half of its inputs, independently of the weights.
H2 Neurons are partitioned into layers, meaning groups
that each path traverses at most once.
H3 All weights have 0 expectation and variance
2/(in-degree), as in the ‘He-initialization’.
H4 The weights from different layers are independent.
H5 Two weights in a same layer are uncorrelated.
Here, two weights are considered in the same layer if their
end-nodes belong to the same layer. If we follow common
practice and initialize our nets as proposed by He et al.
(2015), then H3-H5 are satisfied at initialization by design,
while H1 is usually a very good approximation (Balduzzi
et al., 2016). After training however, we cannot expect
these hypotheses to hold. While if necessary we may try to
enforce H1 and H3 by adding bathchnorms or penalties on
the layerwise weight averages or variances, it is absurd to
believe that we can keep all weights uncorrelated as implied
by H4-H5. That is why, all our statements in this section are
to be understood as orders of magnitudes that are very well
satisfied at initialization in theory and in practice , and that
we will confirm experimentally after training in Section 4.
Said differently, while our theorems rely on the statistics of
neural nets at initialization, our experiments confirm their
conclusions after training.
Theorem 3 (Vulnerability of Fully Connected Nets).
Consider a succession of fully connected layers with ReLU
activations which takes inputs x of dimension d, satisfies
assumptions (H), and outputs logits fk (x) that get fed to
a final cross-entropy-loss
layer L. Then the coordinates of
√
∂x fk grow like 1/ d, and
1
1
k∂x Lkq ∝ d q − 2
and p k∂x Lkq ∝
√
d.
(7)
Consequently, these networks are increasingly vulnerable
to any kind of `p -attacks with growing input-dimension.
Proof. Let x designate a generic coordinate of x. To evaluate the size of k∂x Lkq , we will evaluate the size of the
coordinates ∂x L of ∂x L by decompose them into
K
K
X
X
∂L ∂fk
∂x L =
=:
∂k L ∂x fk ,
∂fk ∂x
k=1
k=1
where fk (x) denotes the logit-probability of x belonging to
class k. We now investigate the statistical properties of the
logit gradients ∂x fk , and then see how they shape ∂x L.
Step 1: Statistical properties of ∂x fk . Let P(x, k) be
the set of paths p from input neuron x to output-logit k.
Let p − 1 and p be two successive neurons on path p, and
p̃ be the same path p but without its input neuron. Let
wp designate the weight
Q from p − 1 to p and ωp be the
path-product ωp := p∈p̃ wp and W the set of all weights.
Finally, let σp (resp. σp ) be equal to 1 if the ReLU of node
p (resp. if path p) is active for input x, and 0 otherwise.
As previously noticed by Balduzzi et al. (2016) using the
chain rule, we see that ∂xP
fk is the sum of all ωp whose path
is active, i.e. ∂x fk (x) = p∈P(x,k) ωp σp . Consequently:
EW,σ ∂x fk (x)2 =
= |P(x, k)|
Y
p∈p̃
X
Y
p∈P(x,k) p∈p̃
EW wp2 Eσ σp2
Y
Y 1
1
2 1
=
dp ·
= . (8)
dp−1 2
dp−1
d
p∈p̃
p∈p̃
The first equality uses H1 to decouple the expectations
over weights and ReLUs, and then applies Lemma 9 of
Appendix A.1, which uses H3-H5 to kill all cross-terms
and take the expectation over weights inside the product.
The second equality uses H3 and the fact that the resulting
product is the same for all active paths. The third equality
counts the number of paths from x to k and we conclude by
noting that all terms cancel out, except dp−1 from the input
layer which is d.
Step 2: Statistical properties of ∂k L and ∂x L. Defining
fk (x)
qk (x) := PKe efh (x) (the probability of image x belonging
h=1
to class k according to the network), we have, by definition
of the cross-entropy loss, L(x, c) := − log qc (x), where c
is the label of the target class. Thus:
−qk (x)
if k 6= c
∂k L(x) =
and
1 − qc (x)
otherwise,
X
∂x L(x) = (1 − qc ) ∂x fc (x) +
qk (−∂x fk (x)). (9)
k6=c
Using again Lemma 9, we see that the ∂x fk (x) are K centered and uncorrelated variables. So ∂x L(x) is approximately the sum of K uncorrelated variables with
Pzero-mean,
and its total variance is given by (1 − qc )2 + k6=c qk2 /d.
√
Hence the magnitude of ∂x L(x) is 1/ d for all x, so the
`q -norm of the full input gradient is d1/q−1/2 . Equation 6
concludes.
Remark 1. Equation 9 can be rewritten as
∂x L(x) =
K
X
k=1
qk (x) ∂x fc (x) − ∂x fk (x) .
(10)
As the term k = c disappears, the norm of the gradients
∂x L(x) appears to be controlled by the total error probability. This suggests that, even without regularization, trying
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
to decrease the ordinary classification error is still a valid
strategy against adversarial examples. It reflects the fact that
when increasing the classification margin, larger gradients
of the classifier’s logits are needed to push images from
one side of the classication boundary to the other. This is
confirmed by Theorem 2.1 of Hein and Andriushchenko
(2017). See also Equation 14 in Appendix B.
We now turn to a more general theorem, that covers a much
wider class of nets and implies Theorem 3. To do so, we
will use the following symmetry assumption on the neural
connections. For a given path p, let the path-degree dp be
the multiset of encountered in-degrees along path p. For
a fully connected network, this is the unordered sequence
of layer-sizes preceding the last path-node, including the
input-layer. Now consider the multiset {dp }p∈P(x,o) of all
path-degrees when p varies among all paths from input x to
output o. The symmetry assumption (relatively to o) is
(S) All input nodes x have the same multiset {dp }p∈P(x,o)
of path-degrees from x to o.
Intuitively, this means that the statistics of degrees encountered along paths to the output are the same for all input
nodes. This symmetry assumption is exactly satified by fully
connected nets, almost satisfied by CNNs (up to boundary effects, which can be alleviated via periodic or mirror padding)
and exactly satisfied by strided layers, if the layersize is a
multiple of the stride.
Theorem 4 (Vulnerability of Feedforward Nets). Consider any non-recurrent neural network with ReLU activation functions that satisfies assumptions (H) and outputs
logits fk (x) that get fed to the cross-entropy-loss L. Then
k∂x fk k2 is independent of the input dimension d. Moreover,
if the net satifies symmetry assumption
(S), then the coordi√
nates of ∂x fk scale like 1/ d and Equation 7 still holds:
√
1
1
k∂x Lkq ∝ d q − 2 and p k∂x Lkq ∝ d.
Proof in Appendix A.1.
Corollary 5 (Vulnerability of CNNs). Any succession of
convolution and dense layers, strided or not, with ReLU
activations, that satisfies assumptions (H), and outputs
logits that get fed to the cross-entropy-loss
L, has logit√
coordinates that scale like 1/ d and satisfies Equation 7.
In particular, it is increasingly vulnerable with growing
input-resolution to attacks generated with any `p -norm.
3.3. Effects of Strided and Average-Pooling Layers on
Adversarial Vulnerability
It is common practice in CNNs to use average-pooling layers
or strided convolutions to progressively decrease the number
of pixels per channel. Corollary 5 shows that using strided
convolutions does not protect against adversarial examples.
However, how about if we replace strided convolutions by
convolutions with stride 1 plus an average-pooling layer.
An average-pooling introduces deterministic weights of size
1/(pooling-window-size), so Theorem 4, which assumes
random weights, does not apply anymore. Thus we may
wonder: does this replacement help protecting against adversarial examples? The following theorem says it does.
Theorem 6 (Effect of Average-Poolings). Consider any
succession of possibly strided convolution and dense layers
with ReLU activations satisfying assumptions (H). Replacing a strided convolution that selects 1 neuron out of n by a
convolution with stride 1 followed by an average-pooling
over n neurons divides all output and loss gradients by n.
Proof in Appendix A.2. The two previous statements suggest to try and replace any strided convolution by its nonstrided counterpart, followed by an average-pooling layer.
Furthermore, Theorem 6 shows that if we systematically
reduce the number of pixels per channel down to 1 by combining convolutional with average-pooling layers, then the
adversarial vulnerability to `∞ becomes independent of the
input-resolution – provided that assumptions (H) stay valid
after training.
4. Empirical Results
In this section, we first empirically verify that both the average `√
1 -norm of ∂x L and the adversarial vulnerability grow
like d with d the input dimension (Section 4.1) as predicted by Corollary 5. We then compare the loss-gradient
regularization methods with adversarially-augmented training (Section 4.2) and finish by verifying the increased effectiveness of average-pooling over strided layers layers to
decrease adversarial vulnerability (Section 4.3).
For all experiments, we only consider adversarial vulnerability to `∞ -attacks, which we approximate using the FGSMimplementation of the python Foolbox-package (Rauber
et al., 2017). For all test-attacks, we use = .006. Our image datasets being globally normalized, for most images this
perturbation is imperceptible. This -threshold should not
be confused with the regularization-strengths appearing
in (3) and (4), which will be varied in some experiments.
4.1. Vulnerability Grows with Input Resolution
Theorems 3-4 and Corollary 5 predict a linear growth of the
average `1 -norm of ∂x L with the square root of the input
dimension d, and therefore also of adversarial vulnerability
(Lemma 1). To test these predictions, we created a 12class dataset of approximately 80, 000 256 × 256 × 3-sized
RGB-images by merging similar ImageNet-classes, resizing
the smallest image-edge to 256 pixels and center-cropping
the result. We then downsized the images to 32, 64, 128
and 256 pixels per edge, and trained 10 CNNs on each of
these downsized datasets. We then computed adversarial
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
vulnerability and average k∂x Lk1 for each network on a
same held-out test-dataset. Figure 1 summarizes the results.
The dashed-line follows the median of each group of 10
networks; the errorbars show the 10th and 90th quantiles. As
predicted by our theorems, both k∂x Lk
√1 and adversarial
vulnerability grow almost linearly with d.
The growth of |||∂x L||| with d means that for large d we
are outside the linear regime of Lemma 1. And indeed, this
occurs at the largest resolution in our experiments: the crossentropy loss L ranges from 0 to log 12 ≈ 2.5. Meanwhile,
for d = 3·256·256 and = .006, we see that Ex k∂x Lk1 ≈
100, which leads to k∂x Lk1 ≈ 2.7. This is even bigger
than the 2.5-range of L. We are therefore way beyond the
linear approximation validity.
75
80
50
70
25
60
50
32 64
128
image-width /
p
256
d
0
Illustration of Proposition 2. The upper row of Figure 2
plots Ex k∂x L1 k, adversarial vulnerability and accuracy as
a function of d1/p . The excellent match between the adversarial augmentation curve with p = ∞ (p = 2) and its
gradient-regularization dual counterpart with q = 1 (resp.
q = 2) illustrates the duality between as a threshold for
adversarially-augmented training and as a regularization
constant in the regularized loss (Proposition 2).
Confirmation of Equation 5. Still on the upper row, the
curves for p = ∞, q = 1 have no reason to match those
for p = 2, q = 2 when plotted against , because is a
threshold that is relative to a specific attack-norm. However,
Equation 5 suggested that the rescaled thresholds d1/p may
approximately correspond to a same ‘threshold-unit’ accross
`p -norms and accross dimension. This is well confirmed by
the upper row plots: by rescaling the x-axis, the p = q = 2
and q = 1 − 1/p = ∞ curves get almost super-imposed.
Ex k∂x Lk measures adversarial vulnerability. Figure 2d
shows that adversarial vulnerability is almost a function of
Ex k∂x Lk1 which is independent of the training and regularization method used. It is a strikingly clear confirmation that
adversarial examples are primarily caused by large gradients
of the classifier as captured via the induced loss.
100
90
Ex k@xLk1
vulnerability
All networks had exactly the same amount of parameters and
very similar structure accross the various input-resolutions.
The CNNs were a succession of 8 ‘convolution → batchnorm → ReLU’ layers with 64 output channels, followed
by a final full-connection to the 12 logit-outputs. We used
2 × 2-max-poolings after layers 2,4 and 6, and a final maxpooling after layer 8 that fed only 1 neuron per channel to
the fully-connected layer. To ensure that the convolutionkernels cover similar ranges of the images accross each of
the 32, 64, 128 and 256 input-resolutions, we respectively
dilated all convolutions (‘à trous’) by a factor 1, 2, 4 and 8.
tion 15 in Appendix B). Each network has 5 ‘convolution →
batchnorm → ReLU’ layers with 64 output-channels each,
followed by an average-pooling that feeds only 1 neuron per
channel to the final fully-connected linear layer. The results
are summarized in Figure 2. Each point represents one net
trained with a specific adversarial training method and a
specific . Each curve groups all points of a same training
method.
32 64
128
image-width /
p
256
d
Figure 1. Both adversarial vulnerability (left) and Ex k∂x Lk1
(right) increase linearly with the square-root of the imageresolution d, as predicted by Corollary 5. Here, adversarial vulnerability gets dampened at higher dimension, because the first-order
approximation made in Equation 2 becomes less and less valid.
4.2. Gradient Penalty and Adversarial Augmentation
In this section, we check the correspondance between
gradient-regularization and adversarially-augmented training (Proposition 2), and compare these methods to another
gradient-regularizer proposed by Hein and Andriushchenko
(2017): the cross-Lipschitz regularizer. We train several
CNNs with same architecture to classify CIFAR-10 images (Krizhevsky, 2009). For each net, we use a specific
training method with a specific regularization value . The
training methods used were `1 - and `2 -penalization of ∂x L
(Equation 3), adversarial augmentation with `∞ - and `2 - attacks (Equation 4) and the cross-Lipschitz regularizer (Equa-
Adversarial regularization improves generalization.
While adversarial vulnerability steadily decreases with growing (Figure 2b ), Figure 2c shows that, whatever adversarial method used, the accuracy before attack on the
held-out test set first increases (and eventually tumbles
down). Decreasing the adversarial vulnerability thus improves generalization, which we recall was the original motivation for Drucker and LeCun (1991) to introduce doublebackpropagation.
Accuracy-vs-Vulnerability Trade-Off. To concentrate
on the accuracy versus vulnerability trade-offs, Figure 2f
merges Figures 2b and 2c by taking out the irrelevant parameter . Following the curves from right to left now
corresponds to implicitely increasing . The long horizontal plateaus confirm that adversarial vulnerability can be
massively driven down without losing in generalization performance. While all methods perform equally well for small
enought -values, on the long run, the best accuracy-tovulnerability ratios are obtained with the traditional adversarially augmented training using FGSM (p = ∞). By
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
100
50
1
(a)
0
1
log10 (✏d1/p)
80
30
20
2
40
1
(b)
0
1
log10 (✏ d1/p)
4
85
3
80
2
20
1
10
0
50
(d)
100
Ex k@xLk1
150
60
2
accuray
30
70
10
Ex k@xLk2
adversarial vulnerability
0
40
accuracy
Ex k@xLk1
adversarial vulnerability
150
0
1
(c)
0
50
(e)
100
Ex k@xLk1
75
70
150
2
Grad Regu L✏, q = 1
Grad Regu L✏, q = 2
Adv Train Le✏, p = 1
Adv Train Le✏, p = 2
Cross-Lipschitz Regu
65
0
1
log10 (✏ d1/p)
(f)
10
20
30
40
adversarial vulnerability
Figure 2. Average norm Ex k∂x Lk of the loss-gradients, adversarial vulnerability and accuracy (before attack) of various networks trained
with different adversarial regularization methods and regularization strengths . Each point represents a trained network, and each curve a
training-method. Upper row: while accuracy first improves with rising , both the average gradient norms and adversarial vulnerability
steadily decrease. A priori, the regularization parameter has different meanings for each method. The relatively good superposition of
all curves in each upper-row plot illustrates (i) the dual-correspondance between adversarially-augmented training and gradient-loss
penalization and (ii) confirms the rescaling of proposed in Equation 5. (d): adversarial vulnerability is almost a function of the average
loss-gradient norm and does not depend on the training method used. It confirms that large gradients of the loss are the main cause of
adversarial vulnerability. (f ): This figure concentrates on the accuracy-vs-vulnerability trade-offs by taking out from Figures 2b and 2c .
While all methods first perform similarly for small , the best ratios are eventually obtained by the adversarial-augmentation methods (L̃).
(e): The almost perfect linear relation between the `1 - and `2 -norms of ∂x L explain the striking similarity between the q = 1 and q = 2
curves on Figure 2f . It suggests that protecting against a given attack-norm also protects against others.
keeping inside the function L in Equation 4, L̃,p can also
account for higher order Taylor-variation of the original loss
L, which might explain why these methods perform better
for higher than their first-order counter-parts (Equation 3).
The penalty-norm does not matter. We were surprised
to see that on Figures 2d and 2f , the L,q curves are almost identical for q = 1 and 2. This indicates that both
norms can be used interchangeably in (3) (modulo proper
rescaling of via Equation 5), and suggests that protecting
agains a specific attack-norm also protects against others.
Equation 6 may provide an explanation: if the coordinates
of ∂x L behave like centered, uncorrelated variables with
equal variance –which follows from assumptions (H) –,
then the `1 - and `2 -norms of ∂x L are simply proportional.
Plotting Ex k∂x L(x)k2 against Ex k∂x L(x)k1 in Figure 2e
confirms this explanation. The slope is independent of the
training method. Therefore, penalizing the k∂x L(x)k1 during training will not only decrease Ex k∂x Lk1 (as shown in
Figure 2a ), but also drive down Ex k∂x Lk2 and vice-versa.
4.3. Averaging, Subsampling and Max-Pooling
Our theorems show that, contrary to strided layers, averagepoolings should decrease adversarial vulnerability. We
tested this hypothesis on CNNs trained on CIFAR-10, with
6 blocks of ‘convolution → BatchNorm →ReLU’ with 64
output-channels, followed by a final average pooling feeding one neuron per channel to the last fully-connected linear layer. Additionally, after every second convolution, we
placed a pooling layer with stride (2, 2) (thus acting on 2×2
neurons at a time). We tested average-pooling, strided and
max-pooling layers and trained 20 networks per architecture.
Results are shown in Figure 3. As predicted, the networks
with average pooling layers are much more robust to adversarial images than the others. Although all accuracies are
very close, they are slightly better with average pooling than
with striding, but slightly worse than with max-pooling.
5. Related Literature
Goodfellow et al. (2014) already stressed that adversarial
vulnerability increases with growing dimension d. Their argument relies on a ‘one-output-to-many-inputs’-model with
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
vulnerability
accuracy
88.5
88.0
87.5
87.0
30
25
20
average strided
max
average strided
max
Figure 3. Compared effects of average-, strided- and max-pooling
layers on the accuracy before attack (left) and adversarial vulnerability (right). As can be seen, average-pooling layers make
networks more robust to adversarial examples, contrary to strided
and max-pooling ones. This confirms Theorem 6.
dimension-independent weights. They therefore conclude
on a linear growth of adversarial vulnerability with d and
accuse our networks of being “too
√ linear-like”. Although
this linear dependance becomes d when adjusting for a
dimension-dependent weight-initialization, our theory and
experiments nevertheless confirm this point of view, in the
sense that a first-order Talyor expansion is indeed sufficient
to explain the adversarial vulnerability of neural networks.
As suggested by the one-output-to-many-inputs model, the
culprit is that growing dimensionality gives the adversary
more and more room to ‘wriggle around’ with the noise
and adjust to the gradient of the output neuron. One of the
contributions here however is to show that this wriggeling
is still possible when the output is connected to all inputs
only indirectly, even when no neuron is directly connected
to all inputs, like in CNNs.
Incidently, Goodfellow et al. (2014) also already relate adversarial vulnerability to large gradients of the loss L, an insight at the very heart of their FGSM-algorithm. They however do not propose any explicit penalizer on the gradient
of L other than indirectly through adversarially-augmented
training; possibly because the main deep learning libraries
did not support automatized backpropation through gradients at that time. While our penalty stems from an approximation of the relevant adversarial vulnerability, Hein and
Andriushchenko (2017) derived yet another gradient-based
penalty –the cross-Lipschitz-penalty– by considering (and
proving) formal guarantees on the total adversarial vulnerability. While both penalties are similar in spirit, focusing
on the relevant adversarial vulnerability has two main advantages. First, it archieves better accuracy-to-vulnerability
ratios, both in theory and practice, because it ignores classswitches between misclassified examples and penalizes only
those that reduce the accuracy. Second, it allows to deal
with one number only, ∆L0/1 or ∆L, whereas Hein and
Andriushchenko’s cross-Lipschitz regularizer and theoretical guarantees explicitely involve all K logit-functions
(and their gradients). See Appendix B. Penalizing networkgradients is also at the heart of contractive auto-encoders
as proposed by Rifai et al. (2011), where it is used to regularize the encoder-features. Seeing adversarial training as a
generalization method, let us also mention Hochreiter and
Schmidhuber (1995), who propose to enhance generalization by searching for parameters in a “flat minimum region”
of the loss. This leads to a penalty involving the gradient of
the loss, but taken with respect to the weights, rather than
the inputs. In the same vein, a gradient-regularization of
the loss of generative models also appears in Proposition 6
of Ollivier (2014), where it stems from a codelength bound
on the data (minimum description length). Finally, Cisse
et al. (2017) propose new network-architectures that have
small gradients by design, rather than by special training:
an approach that makes all the more sense, considering the
conclusion of Theorems 3 and 4. For further details and
references on adversarial attacks and defenses, we refer
to Yuan et al. (2017).
6. Conclusion
We first showed that adversarial vulnerability increases with
the gradients ∂x L of the loss, which is confirmed by the
near-perfect functional relationship between gradient norms
and vulnerability (Figure 2d ). We then evaluated the size
of k∂x Lkq and showed that usual convolutional or fully
connected architectures are more and more vulnerable to
`p -attacks with growing input dimension d (the image-size).
While using strided convolutions does not help, using sufficiently many average-poolings may significantly robustify the network. Our results rely on the statistical weightdistribution at initialization, but our experiments confirm
their conclusions even after training. We also proposed
to regularize the training by penalizing the (`1 -)norm of
∂x L. Like Tikhonov-regularization being approximately
equivalent to training with random noise, we showed that
our penalization is equivalent to training with adversarial noise, both theoretically at first order, and empirically
(Figure 2). We thereby linked double-backpropagation to
FGSM-augmented training. However, even by combining
all these tricks, the networks remain surprisingly vulnerable
to adversarial attacks, which suggests that on the long run,
we may need to design new network architectures that are
inherently more resolution-invariant.
Acknowledgements
We thank Martı́n Arjovsky, Ilya Tolstikhin and Diego Fioravanti for helpful discussions.
References
D. Balduzzi, B. McWilliams, and T. Butler-Yeoman. Neural
Taylor Approximations: Convergence and Exploration
in Rectifier Networks. arXiv:1611.02345, 2016. arXiv:
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
1611.02345.
C. M. Bishop. Training with noise is equivalent to Tikhonov
regularization. Neural computation, 7(1):108–116, 1995.
M. Cisse, P. Bojanowski, E. Grave, Y. Dauphin, and
N. Usunier. Parseval networks: Improving robustness
to adversarial examples. In ICML, 2017.
H. Drucker and Y. LeCun. Double backpropagation increasing generalization performance. In International Joint
Conference on Neural Networks, 1991.
I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining
and Harnessing Adversarial Examples. arXiv:1412.6572,
2014.
K. He, X. Zhang, S. Ren, and J. Sun. Delving Deep into
Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. arXiv:1502.01852, 2015.
M. Hein and M. Andriushchenko. Formal Guarantees on
the Robustness of a Classifier against Adversarial Manipulation. arXiv:1705.08475, 2017.
S. Hochreiter and J. Schmidhuber. Simplifying neural nets
by discovering flat minima. In NIPS, 1995.
J. Huang. Statistics of Natural Images and Models. PhD
thesis, Brown University, Providence, RI, 2000.
A. Krizhevsky. Learning Multiple Layers of Features from
Tiny Images. Technical report, 2009.
Y. Ollivier. Auto-encoders: reconstruction versus compression. arXiv:1403.7752, 2014.
J. Rauber, W. Brendel, and M. Bethge. Foolbox v0.8.0: A
Python toolbox to benchmark the robustness of machine
learning models. arXiv:1707.04131, 2017.
S. Rifai, P. Vincent, X. Muller, X. Glorot, and Y. Bengio.
Contractive Auto-encoders: Explicit Invariance During
Feature Extraction. In ICML, 2011.
X. Yuan, P. He, Q. Zhu, R. R. Bhat, and X. Li. Adversarial Examples: Attacks and Defenses for Deep Learning.
arXiv:1712.07107, 2017.
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
A. Proofs
A.1. Proof of Theorem 4
Now, let us relate these considerations on graphs to gradients
and use assumptions
(H). We remind that path-product ωp
Q
is the product p∈p̃ wp .
The proof of Theorem 4 is very similar to the one of Theorem 3, but we will need to first generalize the equalities
appearing in Equation 8. To do so, we identify the computational graph of a neural network to an abstract Directed
Acyclic Graph (DAG) which we use to prove the needed
algebraic equalities. We then concentrate on the statistical
weight-interactions implied by assumption (H), and finally
throw these results together to prove the theorem. In all the
proof, o will designate one of the output-logits fk (x).
Lemma 9. Under assumptions (H), each path-product has
expectation 0, two distinct path-products (ωp , ωp0 ) decorrelate, and the variance of a path-product is the product of its
variances:
Lemma 7. Let x be the vector of inputs to a given DAG, o
be any leaf-node of the DAG, x a generic coordinate of x.
Let p be a path from the set of paths P(x, o) from x to o, p̃
the same path without node x, p a generic node in p̃, and
dp be its input-degree. Then:
Proof. Applying H4 then H3 shows that EW [ωp ] = 0.
Now, take two different paths p and p0 that end at a same
node o. Going back from o, consider the first node p at
which p and p0 part. This gives two different incoming
weights wp on p and wp0 on p0 which are in a same layer.
Then:
EW [ωp ωp0 ] = EW ωp\p ωp0 \p EW wp wp0 = 0 ,
X
X
Y 1
=1
dp
(11)
x∈x p̃∈P(x,o) p∈p̃
Proof. We will reason on a random walk starting at o and
going up the DAG by choosing any incoming node with
equal probability. The DAG being finite, this walk will end
up at an input-node xQ
with probability 1. Each path p is
taken with probability p∈p̃ d1p . And the probability to end
up
of all these probabilities, i.e.
P at anPinput-nodeQis the sum
−1
d
,
which
concludes.
x∈x
p∈P(x,o)
p∈p p
The sum over all inputs x in Equation 11 being 1, on average it is 1/d for each x, where d is the total number of
inputs (i.e. the length of x). It becomes an equality under
assumption (S):
Lemma 8. Under the symmetry assumption (S), and with
the previous notations, for any input x ∈ x:
X
(12)
Proof. Let us denote D(x, o) := {dp }x∈P(x,o) . Each path
p in P(x, o) corresponds to exactly one element dp in
D(x, o) and vice-versa. And the
Q elements dp of dp completely determine the product p∈p̃ d−1
p . By using Equation 11 and the fact that, by (S), the multiset D(x, o) is
independent of x, we hence conclude
X
Y 1
X
=
dp
x∈x
x∈x p∈P(x,o) p∈p̃
=d
X
Y
dp ∈D(x,o) dp ∈dp
X
p∈p̃
where the first equality uses H4 and the second uses H5 and
the fact that wp and wp0 have 0 expectation by H3. Moreover
Y
Y
2
EW ωp = EW
wp2 =
EW wp2 ,
p∈p̃
p∈p̃
where the second equality uses H4.
We now have all elements to prove Theorem 4.
Proof. (of Theorem 4) For a given neuron p in p̃, let p − 1
designate the previous node in p of p. Let σp (resp. σp ) be a
variable equal to 0 if neuron p gets killed by its ReLU (resp.
path p is inactive), and 1 otherwise. Then:
X Y
X
∂x o =
∂p−1 p =
ωp σp
p∈P(x,o) p∈p̃
Y 1
1
= .
dp
d
p∈P(x,o) p∈p̃
X
EW [ωp ] = EW [ωp ωp0 ] = 0
Y
EW ωp2 =
EW wp2 .
Y
dp ∈D(x,o) dp ∈dp
1
dp
1
= 1.
dp
Consequently:
EW,σ (∂x o)2 =
=
X
p∈P(x,o)
EW [ωp ωp0 ] Eσ [σp σp0 ]
p,p0 ∈P(x,o)
X
Y
p∈P(x,o) p∈p̃
=
X
EW ωp2 Eσ σp2
(13)
Y 2 1
1
= ,
dp 2
d
p∈P(x,o) p∈p̃
where the firs line uses the independence between the ReLU
killings and the weights (H1), the second uses Lemma 9
and the last uses Lemma 8. The gradient ∂x o thus has coordinates whose squared expectations
scale like 1/d. Thus
√
each coordinate scales like 1/ d and k∂x okq like d1/2−1/q .
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
Conclude on k∂x Lkq and p k∂x Lkq by using Step 2 of the
proof of Theorem 3.
Finally, note that, even without the symmetry assumption
(S), we have:
h
i X
2
EW k∂x ok2 =
X
Y 2 1
=1
dp 2
x∈x p∈P(x,o) p∈p̃
Thus, even without (S), k∂x ok2 is independent of the inputdimension d.
A.2. Proof of Theorem 6
Proof. Pick any strided convolution that selects only 1 neuron out of n, replace it by a convolution with stride 1 plus
average-pooling and consider Equation 13. Going from the
first to the second line uses Lemma 9, which does not apply because the average-pooling weights obviously do not
satisfy H3-H5. Instead, we apply Lemma 10 –a variation
of Lemma 9 proven below–, and proceed with the second
line. Now there are n times more neurons after the convolution, and therefore n times more paths in each P(x, o),
which multiplies the overall variance by n. But the average
pooling divides each path-product ωp by n, which reduces
each path’s variance by n2 . Applying the two last lines
of Equation 13, we see that the overall variance thus gets
2
multiplied by n/n
√ = 1/n, and the average coordinate size
therefore by 1/ n.
Lemma 10 (Variation of Lemma 9). Consider any succession of possibly strided convolution and dense layers
with ReLU activations satisfying assumptions (H). Replace
an arbitrary number of strided convolutions by the corresponding non-strided convolution and average-pooling
layer. Then, for any two different paths p, p0 ∈ P(x, o)
from an input-coordinate x to an output o, the path-products
ωp , ωp0 satisfy:
EW [ωp ] = EW [ωp ωp0 ] = 0
Y
EW ωp2 =
EW wp2 .
p∈p̃
Proof. All path-products are a product of variables satisfying assumptions (H), divided by a product of constants, which correspond to the different average-pooling
weights.
like
Therefore,
in Lemma 9: EW [ωp ] = 0 and
Q
EW ωp2 = p∈p̃ EW wp2 . The difference with Lemma 9
is that, by introducing average-pooling layers, we introduced
paths that may have exactly the same path-products: the reasoning we did in Lemma 9 –“going back from o, consider
the first node p at which p and p0 part...”– does not work, because this node p could be an average-pooling, in which case
we do not get two different weights wp , wp0 . But starting
from o now works: consider the two first nodes p ∈ p and
p0 ∈ p0 that do not both belong to p and p0 . Their weights
wp and wp0 are in a same layer, because we assume that the
architecture is a simple succession of layers; and they cannot be average-pooling weights, because average-poolings
cannot separate one incoming path into two or more paths.
Thus wp and wp0 satisfy H3-H5, so we get, as in the proof
of Lemma 9:
EW [ωp ωp0 ] = EW ωp\p ωp0 \p0 EW [wp wp0 ] = 0 .
B. Comparison to the Cross-Lipschitz
Regularizer
In their Theorem 2.1, Hein and Andriushchenko show that
the minimal = kδkp perturbation to fool the classifier
must be bigger than:
min
k6=c
fc (x) − fk (x)
.
maxy∈B(x,) k∂x fc (y) − ∂x fk (y)kq
(14)
They argue that the training procedure typically already tries
to maximize fc (x) − fk (x), thus one only needs to additionnally ensure that k∂x fc (x) − ∂x fk (x)kq is small. They
then introduce what they call a Cross-Lipschitz Regularization, which corresponds to the case p = 2 and involves the
gradient differences between all classes:
RxLip :=
K
1 X
2
k∂x fh (x) − ∂x fk (x)k2
K2
(15)
k,h=1
In contrast, using (10), (the square of) our proposed regularizer k∂x Lkq from Equation 3 can be rewritten, for
p = q = 2 as:
Rk·k2 (f ) =
K
X
k,h=1
qk (x)qh (x) ∂x fc (x) − ∂x fk (x) ·
· ∂x fc (x) − ∂x fh (x)
(16)
Although both (15) and (16) consist in K 2 terms, corresponding to the K 2 cross-interaction between the K classes,
the big difference is that while in (15) all classes play exactly the same role, in (16) the summands all refer to the
target class c in at least two different ways. First, all gradient
differences are always taken with respect to ∂x fc . Second,
each summand is weighted by the probabilities qk (x) and
qh (x) of the two involved classes, meaning that only the
classes with a non-negligeable probability get their gradient
regularized. This reflects the idea that only points near the
margin need a gradient regularization, which incidentally
will make the margin sharper.
C. Perception Threshold
To keep the average pixel-wise variation constant accross
dimensions d, we saw in Equation 5 that the threshold p of
Adversarial Vulnerability of Neural Networks Increases With Input Dimension
an `p -attack should scale like d1/p . We will now see another
justification for this scaling. Suppose that given an `p -attack
norm, we want to choose p such that the signal-to-noise
ratio (SNR) kxk2 / kδk2 of a perturbation δ with `p -norm
<= p is never greater than a given SNR threshold 1/. For
p = 2 this imposes 2 = kxk2 . More generally, studying
the inclusion of `p -balls in `2 -balls yields
p = kxk2 d1/p−1/2 .
(17)
Note that this gives again p = ∞ d1/p . This explains how
to adjust the threshold with varying `p -attack norm.
Now, let us see how to adjust the threshold of a given `p norm when the dimension d varies. Suppose that x is a
natural image and that decreasing its dimension means either decreasing its resolution or cropping it. Because the
statistics of natural images are approximately resolution
and scale invariant (Huang, 2000), in either case the average squared value of the image pixels
√remains unchanged,
which implies that kxk2 scales like d. Pasting this back
into Equation 17, we again get:
p = ∞ d1/p .
In particular, ∞ ∝ is a dimension-free number, exactly
like in Equation 5 of the main part.
Now, why did we choose the SNR as our invariant reference
quantity and not anything else? One reason is that it corresponds to a physical power ratio between the image and the
perturbation, which we think the human eye is sensible to.
Of course, the eye’s sensitivity also depends on the spectral
frequency of the signals involved, but we are only interested
in orders of magnitude here.
D. A Variant of Adversarially-Augmented
Training
In usual adversarially-augmented training, the adversarial
image x + δ is generated on the fly, but is nevertheless
treated as a fixed input of the neural net, which means that
the gradient does not get backpropagated through δ. This
need not be. As δ is itself a function of x, the gradients
could actually also be backpropagated through δ. As it was
only a one-line change of our code, we used this opportunity
to test this variant of adversarial training (FGSM-variant
in Figure 2) and thank Martı́n Arjovsky for suggesting it.
But except for an increased computation time, we found no
significant difference compared to usual augmented training.
| 1cs.CV
|
NORMALIZED INFORMATION DISTANCE AND THE
OSCILLATION HIERARCHY
arXiv:1708.03583v2 [math.LO] 28 Aug 2017
KLAUS AMBOS-SPIES, WOLFGANG MERKLE, AND SEBASTIAAN A. TERWIJN
Abstract. We study the complexity of approximations to the normalized information distance. We introduce a hierarchy of computable approximations
by considering the number of oscillations. This is a function version of the difference hierarchy for sets. We show that the normalized information distance is
not in any level of this hierarchy, strengthening previous nonapproximability results. As an ingredient to the proof, we also prove a conditional undecidability
result about independence.
1. Introduction
The normalized information distance NID is a metric for binary strings based
on Kolmogorov complexity. It was introduced by Li et al. in [3], and subsequently
studied in a series of papers, cf. [9], [4, Section 8.4]. The NID is interesting for
both theoretical and practical reasons. Despite the fact that it is noncomputable,
it has a number of surprising practical applications.
In this paper we show that for any computable approximation of NID, the
number of oscillations is not bounded by a constant. This improves the nonapproximability result of [8], and also confirms a conjecture made in that paper.
We recall the following definitions from computability theory. A function F :
ω → ω is a ∆02 -function (also called computably approximable, or limit computable)
if there is a computable function f such that
lim f (x, s) = F (x)
s→∞
for every x. By Shoenfield’s Limit Lemma (cf. Odifreddi [5, IV.1.17]), F is limit
computable if and only if F is computable with the Halting Problem ∅′ . The following definition gives a fine hierarchy for ∆02 -functions, by considering the number
of oscillations of the approximations.
Definition 1.1. Suppose that F is a function from ω to ω. (The definition below
applies also to functions with domain or range ω × ω, {0, 1}∗ , or Q.) Suppose that
f is an approximation of F satisfying lims→∞ f (x, s) = F (x) for every x. The
function F is called approximable from below if f can be chosen to be computable
and nondecreasing. Similarly, F is approximable from above if f can be chosen to
be computable and nonincreasing.
−1
Inductively define the function classes Σ−1
n and Πn as follows.
• F ∈ Σ−1
1 if F is approximable from below.
• F ∈ Π−1
1 if F is approximable from above.
• F ∈ Σ−1
n+1 if F has a computable approximation f such that for every x
there exists s such that f (x, t) for t > s is a Π−1
n -approximation, and f (x, t)
for t < s is nondecreasing.
Date: August 29, 2017.
2010 Mathematics Subject Classification. 03D15, 03D32, 03D55, 68Q30,
Key words and phrases. Kolmogorov complexity, information distance, independence.
1
2
K. AMBOS-SPIES, W. MERKLE, AND S. A. TERWIJN
• F ∈ Π−1
n+1 if F has a computable approximation f such that for every x
there exists s such that f (x, t) for t > s is a Σ−1
n -approximation, and f (x, t)
for t < s is nonincreasing.
−1
So Σ−1
n ∪ Πn is the class of functions with a computable approximation that
oscillates no more than n times on every
x, and beginning by going up for Σ−1
n
S −1
−1
the
oscillation
hierarchy.
and by going down for Πn . We call n Σn ∪ Π−1
n
The difference hierarchy over the computably enumerable (c.e.) sets was introduced by Ershov, cf. Odifreddi [5, IV.1.18] and Selivanov [6]. It is a fine hierarchy
for the ∆02 sets, sometimes also referred to as the Boolean hierarchy. It can be seen
as an effective version of a classical hierarchy (introduced by Hausdorff) studied in
descriptive set theory. An analogous hierarchy (defined over NP) is also studied in
complexity theory. Note that for {0, 1}-valued functions (i.e. sets), the oscillation
hierarchy above coincides with the difference hierarchy: Σ−1
1 coincides with the c.e.
sets and Π−1
coincides
with
the
co-c.e.
sets.
Furthermore,
for sets, Σ−1
1
2 coincides
with the class of d.c.e. sets (differences of c.e. sets), and in general Σ−1
n with the
n-c.e. sets. This also explains the notation that we use for these classes, as the
same notation has been used for the difference hierarchy (see e.g. Selivanov [7]).
Just as in the case of sets, an elementary diagonalization argument shows that
the oscillation hierarchy does not exhaust the ∆02 -functions. The main result of
this paper (Theorem 11.1) shows that NID is a natural example of such a function.
Standard techniques also show that the oscillation hierarchy is proper, that is, that
every level of it is strictly included in the next ones. In fact, this follows from the
analogous results for sets.
−1
We note that the definition of the classes Σ−1
n and Πn is easily extended to
−1
real-valued functions. That a function F is Σ1 is then equivalent to saying that
the values F (x) are left c.e. reals in a uniform way.
The information distance D(x, y) between strings x and y is defined as
D(x, y) = min l(p) : U (p, x) = y ∧ U (p, y) = x .
Here U is a fixed universal prefix machine. Like the Kolmogorov complexity function K, the distance function D is a Π−1
1 -function, i.e. computably approximable
from above.
Define
E(x, y) = max K(x|y), K(y|x) .
Here K denotes the prefix-free Kolmogorov complexity. Up to a logarithmic factor,
D and E are equal:
Proposition 1.2. [9, Corollary 3.1] D(x, y) = E(x, y) + O(log E(x, y)).
It was proven in Bennett et al. [1] that E is a metric (more precisely, that it
satisfies the properties of a metric up to a fixed additive constant). Theorem 3.7
of [9] states that E is minimal among all similar distance functions.
The normalized information distance NID(x, y) is defined as
NID(x, y) =
E(x, y)
.
max K(x), K(y)
Note that NID, being the ratio of two Π−1
1 -functions, is computable in the limit.
It was proved in [8] that NID ∈
/ Σ−1
and
NID ∈
/ Π−1
1
1 , that is, NID is neither
computably approximable from below nor from above. (In particular NID is not
computable.) The goal of this paper is to improve this to higher levels of the
oscillation hierarchy.
The outline of the paper is as follows. First, we reprove the nonapproximability
result of [8], using an immunity argument. Next, we discuss the cases Σ−1
2 and
−1
Π2 , that form the basis for later proofs.
NORMALIZED INFORMATION DISTANCE AND THE OSCILLATION HIERARCHY
3
The Σ−1
3 case requires a new ingredient, namely an undecidability result about
independence that is of independent interest (section 7). We show a conditional
undecidability result, namely that, given two random strings, it is undecidable
whether they are independent. In fact, we need a stronger form of this result,
which takes the form of a conditional immunity statement: It is impossible to
effectively generate infinitely many pairs, including infinitely many random pairs,
such that all the random pairs in the enumeration are independent.
In section 10.1 we discuss the Π−1
4 case, which builds on the previous cases. The
proof is set up in such a way that the pattern needed for the induction proof of
the general case is shown. The general case is then treated in section 11.
Our notation is mostly standard. For background about Computability Theory
we refer to Odifreddi [5], and for background about Kolmogorov complexity to Li
and Vitányi [4] and Downey and Hirschfeldt [2]. The natural numbers are denoted
by ω, and the set of binary strings by {0, 1}∗ . We use |x| to denote the length of a
string x. We use 6+ and =+ to denote (in)equality up to fixed additive constants.
For example, f (x) 6+ g(x) means that there is a constant c such that for all x,
f (x) 6 g(x) + c.
2. NID is not Σ−1
1
In this section and below, we consider computable approximations NIDs of NID.
Note that these now take values in Q.
That NID ∈
/ Σ−1
1 was proven in [8], but we give a new proof here.
Lemma 2.1. Consider pairs of strings x,y such that |x| = |y| = n.
(i) lim supn NID(x, y) = 1,
(ii) lim inf n NID(x, y) = 0.
Proof. To prove (i), note that there is a sequence of pairs (x, y) for which NID(x, y)
converges to 1. For example, one can take x random and y = 0n (that is, a sequence
of n zeros). Then NID(x, y) = K(x|n)
K(x) , which for random x converges to 1 when
|x| → ∞.
O(1)
To prove (ii), note that for x = y we have NID(x, x) = K(x)
, which converges
to 0, as K(x) grows unbounded for |x| → ∞.
Recall that a set is immune if it is infinite, but it does not contain an infinite
c.e. subset. Immune sets were introduced by Post, and they play an important
role in computability theory, cf. [5].
Theorem 2.2. (Barzdins) The set of random strings
x | K(x) > 21 |x|
is immune.
Proof. Cf. Theorem 2.7.1 in [4]. The result is stated there for the plain complexity
C, but it equally holds for the prefix-free complexity K.
Proposition 2.3. Let 0 < r < 1 be real. Then the set
X = (x, y) | NID(x, y) > r
is immune.
Proof. First note that X is infinite by Lemma 2.1 (i). Now suppose for a contradiction that A is an infinite c.e. subset of X. Without loss of generality A
is computable. It follows that we can generate an infinite computable sequence
(xn , yn ) such that NID(xn , yn ) > r. Since the sequence is computable we have
4
K. AMBOS-SPIES, W. MERKLE, AND S. A. TERWIJN
that the numerator E(xn , yn ) of NID is bounded by a fixed constant (not depending on n). However, since the denominator max{K(xn ), K(yn )} grows unbounded
for n → ∞ (as K grows unbounded), it is impossible that NID(xn , yn ) > r for
all n.
Theorem 2.4. ([8]) NID is not Σ−1
1 .
Proof. Suppose that NID were
approximable from below. Then we could com
putably enumerate the set (x, y) | NID(x, y) > 12 . Note that this set is infinite
by Lemma 2.1 (i). This contradicts Proposition 2.3.
3. NID is not Π−1
1
Proposition 3.1. ([8]) NID is not Π−1
1 .
Proof. Suppose that NID ∈ Π−1
1 , i.e. that NID has a monotonic nonincreasing
computable approximation. Then also the function NID(x, x) = E(x, x)/K(x) is
Π−1
1 . But E(x, x) is bounded by a constant, and it follows from this that we can
compute K(x) on an infinite computable subset of ω. But this is impossible by
the usual compression argument.
More precisely; suppose E(x, x) 6 c for all x. Then
1
c
, . . . , K(x)
NID(x, x) ∈ K(x)
for every x. Let Ks and NIDs denote the Π−1
1 -approximations of K and NID. As
b
NIDs is rational-valued, if NID(x, x) = K(x) we may never actually see s such that
NIDs (x, x) =
b
Ks (x) .
However, we may fix a precision ε (depending on x) such
b
is
that this holds to within ε. The precision needed to distinguish the values K(x)
1 1
ε < 2 K(x) , but since K(x) is not computable we take a computable upper bound
1
ε = 13 · 2|x|
. Now let 1 6 b 6 c be minimal such that for infinitely many x there is
s such that
NIDs (x, x) ∈ Ksb(x) − ε, Ksb(x) + ε .
(1)
Since for almost all of the x for which (1) holds, Ks (x) is equal to the final
value K(x) (because by monotonicity of the approximations, the final value cannot
change), we see that we can compute K on an infinite computable subset, which
is known to be impossible. Namely, this would entail that we could effectively find
strings of large complexity, which is impossible by Theorem 2.2.
4. NID is not Σ−1
2
Lemma 4.1. There is no computable sequence of pairs (xn , yn ) with |xn | = |yn |,
such that NID(xn , yn ) < n1 for every n > 1.
Proof. Suppose that NID(x, y) < n1 . Then we have
1
E(x, y)
1
> NID(x, y) =
>
,
n
max{K(x), K(y)}
max{K(x), K(y)}
and hence max{K(x), K(y)} > n. Hence the existence of a sequence (xn , yn ) as
in the statement of the lemma would contradict that it is impossible to effectively
generate high-complexity strings: If we could generate pairs of strings (x, y) such
that at least one of x, y was of complexity at least n, then by concatenating x
and y we could generate strings of complexity at least 12 n, again contradicting
Theorem 2.2.
Theorem 4.2. NID is not Σ−1
2 .
NORMALIZED INFORMATION DISTANCE AND THE OSCILLATION HIERARCHY
5
Proof. Suppose for a contradiction that NID is Σ−1
2 , that is, it has a computable
approximation that first goes up and then goes down. As there are infinitely many
pairs (x, y) for which NID(x, y) > 34 , we can effectively find infinitely many pairs
(x, y), with x and y of the same length, such that the approximation initially
indicates (in the going up phase) that NID(x, y) > 34 . If for almost all of these
it would actually hold that NID(x, y) > 43 this would contradict Proposition 2.3.
Hence for infinitely many of the pairs (x, y) the Σ−1
2 -approximation has to come
down subsequently. Indeed, for any given ε > 0 it has to come down to below ε
for some pair (x, y), for otherwise we would again contradict Proposition 2.3. This
means we can thin out the sequence of pairs (x, y) to a computable subsequence
(xn , yn ) such that NID(xn , yn ) < n1 for every n. But this contradicts Lemma 4.1.
5. NID is not Π−1
2
Below we will use the following facts about prefix-free complexity. For a string
x, we let x∗ denote a minimal program for x.
We will use the following theorem (Symmetry of Information), due to Levin and
Gács and also Chaitin:
Theorem 5.1. ([2, p134]) For all strings x, y,
K(x, y) =+ K(x) + K(y|x∗ )
=+ K(y) + K(x|y ∗ ).
Since for strings x and y of the same length we have K(xy) =+ K(x, y), in that
case we have
K(xy) =+ K(x) + K(y|x∗ ).
(2)
Below we also use the elementary fact that for strings x of length n we have
K(x) 6+ n + 2 log n,
K(x|n) 6+ |x|.
(3)
(4)
We will also make use of the following result, which is a consequence of Symmetry
of Information.
Lemma 5.2. Let |x| = |y| = n. Suppose x is random, i.e. K(x) > n, and
y is random relative to x, i.e. K(y|x∗ ) > n. Then x is random relative to y:
K(x|y ∗ ) > n − c log n, where c is a constant not depending on x and y.
Proof. By Theorem 5.1,
K(y) − K(y|x∗ ) =+ K(x) − K(x|y ∗ ).
By (3) and the assumption K(y|x∗ ) > n, the LHS is bounded by 2 log n. Considering the RHS, it then follows from K(x) > n that K(x|y ∗ ) > n − c log n.
Before we prove the theorem, we do some preliminary calculations. Consider
pairs of strings x, y with |x| = |y| = n, with n even, such that x = ax0 and y = ay0
for some string a of length 21 n. That is, x and y both have their first half equal
to a.
For any string a, suppose that x0 and y0 are random relative to each other and
also to a. More precisely, we assume that
K(x0 |a∗ , y0 ) > 21 n − d and K(y0 |a∗ , x0 ) > 12 n − d
(5)
where d is a constant, not depending on a or n. By elementary counting of programs,1 for any ε > 0 there is a constant d such that for all a, the fraction of all
1This holds for the plain Kolmogorov complexity C, so a fortiori it also holds for the prefix-free
complexity K.
6
K. AMBOS-SPIES, W. MERKLE, AND S. A. TERWIJN
pairs x0 ,y0 satisfying (5) exceeds 1 − ε. For definiteness, we choose d such that (5)
holds for a majority of 34 of all pairs x0 and y0 .
We now show that for a random NID(x, y) is large, and that for a compressible
NID(x, y) is small.
First suppose that a is random, i.e. K(a) > |a| = 21 n. By (2) and (5) above we
have
K(ax0 ) =+ K(a) + K(x0 |a∗ )
> 21 n + 12 n − d = n − d.
Similarly we have K(ay0 ) >+ n − d. Using this and (4) gives
max{K(ax0 |ay0 ), K(ay0 |ax0 )}
max{K(ax0 ), K(ay0 )}
max{K(x0 |n), K(y0 |n)} + O(1)
6
max{K(ax0 ), K(ay0 )}
NID(x, y) =
1
2n +
O(1)
≈ 21 .
n − d ± O(1)
Here ≈ denotes asymptotical equality. For the discussion below, we fix a small
number α > 0 such that for almost all x, y as above we have
6
K(a) > 12 n =⇒ NID(x, y) <
1
2
(6)
+ α.
If on the other hand a is compressible (say K(a) 6 21 |a| = 14 n), then using (2)
we have
K(ax0 ) =+ K(a) + K(x0 |a∗ )
6 41 n + K(x0 |a)
6 14 n + K(x0 |n)
6 14 n + 21 n = 34 n,
and similarly K(ay0 ) 6+ 34 n. This gives
max{K(ax0 |ay0 ), K(ay0 |ax0 )}
max{K(x), K(y)}
max{K(x0 |ay0 ), K(y0 |ax0 )}
=+
max{K(x), K(y)}
NID(x, y) =
>
>
1
2n
−d
max{K(x), K(y)}
(by assumption (5))
1
2n −
3
4n
d
≈ 23 .
± O(1)
For the discussion below, we fix a small number β > 0 such that for almost all x, y
as above we have
K(a) 6 41 n =⇒ NID(x, y) > 32 − β.
(7)
Theorem 5.3. NID is not Π−1
2 .
−1
Proof. Suppose for a contradiction that NID ∈ Π−1
2 , and let NIDs denote a Π2 approximation of NID. Note that NIDs consists of an initial going down phase,
followed by a final going up phase. Let α and β be as in (6) and (7). We have two
cases.
Case 1. Suppose that for almost all lengths n, for all pairs x, y of the form
above we have that
if NID(x, y) >
2
3
− β then for all s, NIDs (x, y) >
2
3
− β.
(8)
NORMALIZED INFORMATION DISTANCE AND THE OSCILLATION HIERARCHY
7
We show that this assumption implies that there is an effective procedure to enumerate random strings a as follows. Consider all pairs x = ax0 and y = ay0 , with
|x0 | = |y0 | = |a|. Enumerate a if for most strings x0 and y0 we have found s
such that NIDs (x, y) < 21 + α, where by “most” we mean a majority of 43 of all
pairs x0 , y0 (cf. the discussion following (5) above). By (6), if a is random then
NID(x, y) < 12 + α for most pairs x0 , y0 , hence a is enumerated. Conversely, if a
is enumerated, then by the assumption (8) it cannot be that NID(x, y) > 32 − β,
for then there could be no s such that NIDs (x, y) < 12 for any x0 , y0 . Thus by (7)
we cannot have K(a) 6 21 |a|. Hence we have an enumeration of random strings a
that enumerates all strings with K(a) > |a| and no strings with K(a) 6 21 |a|. By
the immunity of the set of random strings (Theorem 2.2) this is impossible.
Case 2. In the opposite of Case 1, we have that for infinitely many pairs x, y of
the form above, NID(x, y) > 32 − β and there exists s such that NIDs (x, y) 6 23 − β.
This means that the final value of NIDs can only be reached (or approached) in
the final going up phase of the Π−1
2 -approximation. But this implies that we can
effectively enumerate infinitely many pairs x, y such that NID(x, y) > 23 − β,
contradicting the immunity of Proposition 2.3.
Since in both cases we have reached a contradiction, we conclude that NID ∈
/
−1
Π2 .
6. NID is not Π−1
3
This follows with the same argument as for Σ−1
2 , using Lemma 4.1. We need
again that the Π−1
-phase
is
really
used
infinitely
often,
but this was proven already
2
−1
for the Π2 case.
7. Immunity and independence
As an ingredient for the proofs to come, we need a result about the undecidability of independence of random strings. More precisely, we need that there is no
algorithm that, given two random strings of the same length, can decide whether
they are independent or not, where it is agreed that the algorithm may fail to
converge or to give the right answer if one or both of the strings are not random.
In fact, we need something stronger, namely the following immunity property of
independence.
Fix a small real number r > 0. We define the following sets of pairs:
R = (x, y) |x| = |y| = n ∧ K(x) > n ∧ K(y) > n ,
I = (x, y) ∈ R K(x|y) > rn ∨ K(y|x) > rn .
So R is the set of random pairs, and I the set of independent random pairs.2
Definition 7.1. (i) We say that a set Y is decidable conditional to X if there is a
partial computable function ϕ such that for all x, x ∈ X =⇒ ϕ(x) is defined and
Y (x) = ϕ(x).
(ii) We say that a set Y is immune conditional to X if there is no c.e. set A such
that A ∩ X is infinite and A ∩ X ⊆ Y .
Note that for X = ω this is just classical immunity from computability theory.
In the classical case, one can always take A to be computable, since every infinite
c.e. set contains a computable subset. However, in the conditional case this is no
longer true, since A ∩ X need not be c.e.
2By symmetry of information (Theorem 5.1), it actually does not matter much whether we
use “and” or “or” in the definition of I. The amount of independence is given here by r.
8
K. AMBOS-SPIES, W. MERKLE, AND S. A. TERWIJN
For Y ⊆ X, we have that Y is not immune conditional to X if there is an
algorithm that enumerates elements, including infinitely many from X, such that
all the elements from X that are enumerated are in Y . The algorithm may also
enumerate numbers outside X, and these may fail to be included in Y .
Note that if Y ∩ X is infinite, and Y is decidable conditional to X, then Y is
not immune conditional to X. Hence, conditional immunity is a strong form of
conditional undecidability.
Theorem 7.2. The set of independent pairs I is immune conditional to R.
Proof. Suppose for a contradiction that there exists a c.e. set A such that A ∩ R
is infinite and A ∩ R ⊆ I. Consider n such that A contains a random pair (x, y) of
strings of length n. (By assumption, there exist infinitely many such n.) Also, by
assumption, (x, y) ∈ I, and w.l.o.g. K(y|x) > rn. Since K(y|x∗ ) =+ K(y|x, K(x)),
it follows from (2) that
K(xy) >+ K(x) + K(y|x∗ )
>+ K(x) + K(y|x, K(x))
>+ n + K(y|x) − 2 log K(x)
> n + rn − 2 log(2n)
> n + rn − O(log n).
Here we have used K(z) 6+ 2 log z twice. The third inequality holds because
K(y|x) 6+ K(y|x, K(x)) + K(K(x)). We obtain a contradiction by describing xy
with fewer bits as follows.
Consider the “last” string σ of length n found to be nonrandom, that is, for
the approximation Kt of K we have Kt (σ) < n for some t, and for any other
nonrandom string τ of length n we also have Kt (τ ) < n. Hence knowing σ allows
us to decide for every string of length n whether it is random or not. Thus we can
compute from σ the first random pair (x, y) of strings of length n enumerated into
A. Since |σ| = n, we have K(xy) 6+ n. But for large enough n this contradicts
K(xy) >+ n + rn − O(log n).
Note that the undecidability of I, given R, follows from the conditional immunity in the theorem. We can think of A as an algorithm that generates pairs,
including infinitely many random pairs, such that every random pair it includes is
independent. By the theorem, such an algorithm does not exist.
8. NID is not Σ−1
3
The idea of the proof is as follows. As for the Π−1
2 -case, we consider pairs of
strings x, y with |x| = |y| = n, but instead of the first halves being equal to a
common string a we now consider x = a0 x0 and y = b0 y0 , with |a0 | = |b0 | = 12 n.
Assuming that NID is Σ−1
3 , we will have a case distinction as in the proof of
Theorem 5.3, but now instead of enumerating random strings a (as in Case 1 of that
theorem), we will have a procedure to enumerate pairs a0 , b0 that are dependent
but random, which will produce the desired contradiction.
Before we prove the theorem, we perform some auxiliary calculations that will
be used in the proof. Suppose x = a0 x0 , y = b0 y0 are strings of length n such that
|a0 | = |b0 | = |x0 | = |y0 | = 12 n. Assume that x0 and y0 are random relative to each
other, and also to a0 and b0 . More precisely, we assume that
K(x0 |a∗0 , b∗0 , y0 ) > 21 n − d and K(y0 |b∗0 , a∗0 , x0 ) > 12 n − d
(9)
where d is a constant. By elementary counting of programs (in the same way as
for (5)), for any ε > 0 there is a constant d (not depending on a0 , b0 , or n) such
NORMALIZED INFORMATION DISTANCE AND THE OSCILLATION HIERARCHY
9
that the fraction of all pairs x0 , y0 satisfying (9) exceeds 1 − ε. For definiteness,
we choose d such that (9) holds for a majority of 87 of all pairs x0 and y0 .
First suppose that a0 and b0 are random and dependent, more precisely:
• K(a0 |b0 ) 6 rn and K(b0 |a0 ) 6 rn,
• K(a0 ) > |a0 |, K(b0 ) > |b0 |.
1
for definiteness. By (2) and (9)
Here r > 0 is a small rational number, say r = 10
we have
K(a0 x0 ) =+ K(a0 ) + K(x0 |a∗0 )
> 21 n + 12 n − d = n − d.
Similarly we have K(b0 y0 ) >+ n − d. We also have
K(a0 x0 |b0 y0 ) 6+ K(a0 |b0 y0 ) + K(x0 |b0 y0 )
6 K(a0 |b0 ) + K(x0 |n)
6 rn + 12 n.
Similarly we have K(b0 y0 |a0 x0 ) 6+ rn + 12 n. This gives
max{K(a0 x0 |b0 y0 ), K(b0 y0 |a0 x0 )}
max{K(a0 x0 ), K(b0 y0 )}
rn + 12 n + O(1)
.
6
n − d ± O(1)
NID(x, y) =
The last expression is asymptotically smaller than
α such that for almost all a0 , b0 as above we have
NID(x, y) <
1
2
1
2
+ r. We fix a small number
(10)
+ α.
Second suppose that a0 and b0 are independent, more precisely:
• K(a0 |b0 ) > 12 n and K(b0 |a0 ) > 21 n
(Note that this in particular implies that a0 and b0 are random.) By (3) we have
K(a0 x0 ) 6 n + 2 log n + O(1)
and also K(b0 y0 ) 6 n + 2 log n + O(1). Furthermore,
K(a0 x0 |b0 y0 ) =+ K(a0 |b0 y0 ) + K(x0 |a∗0 , b0 y0 )
> K(a0 |b0 y0 ) + 12 n − d
>
1
2n
− O(log n) +
1
2n −
(by assumption (9))
d
(by Lemma 5.2)
= n − O(log n).
and similarly K(b0 y0 |a0 x0 ) >+ n − O(log n). This gives
n − O(log n)
.
n + 2 log n ± O(1)
Since this expression converges to 1, we can let β > 0 be a small number such that
for almost all x, y as above we have
NID(x, y) >
NID(x, y) > 1 − β.
(11)
Theorem 8.1. NID is not Σ−1
3 .
Proof. Assume for a contradiction that NIDs is a Σ−1
3 -approximation of NID. In
−1
the following, we refer to the three phases of the Σ3 -approximation NIDs by phase
1 (going up), 2 (going down), and 3 (going up), respectively. We let α and β be
as in (10) and (11). Consider the set
A = (x, y) | ∃s NIDs (x, y) > 21 + α in phase 1 .
10
K. AMBOS-SPIES, W. MERKLE, AND S. A. TERWIJN
Note that A is a c.e. set, and that A is infinite. In fact, for any r > 0 and for almost
all pairs x, y such that NID(x, y) > r, there exists s such that NIDs (x, y) > r in
phase 1. Namely, if for some r this were not the case, we would see infinitely
many pairs x, y with NIDs (x, y) > r in phase 3, contradicting the immunity of
Proposition 2.3. Now we argue that also infinitely many pairs for which the NID
is small have to enter A. As before, we consider pairs x = a0 x0 , y = b0 y0 of length
n with |a0 | = |b0 | = |x0 | = |y0 |.
Claim: There exist infinitely many pairs a0 , b0 such that for a fraction of at
least 34 of the pairs x0 , y0 ,
• K(a0 |b0 ) 6 rn and K(b0 |a0 ) 6 rn,
• K(a0 ) > |a0 |, K(b0 ) > |b0 |,
• NID(x, y) < 12 + α,
• there exists s such that NIDs (x, y) > 12 + α in phase 1.
To prove the claim, first note that by (10), the third item follows from the first
two, provided x0 and y0 satisfy (9). Suppose the claim is false, so that for almost
every a0 , b0 , for a fraction of > 41 of the x0 , y0 , if the first three items are satisfied
then NIDs (x, y) < 21 + α for every s.
Consider the following enumeration. Enumerate the pair a0 , b0 if for a majority
of at least 78 of the pairs x0 , y0 we have found s such that NIDs (x, y) > 12 + α.
• This enumeration includes all a0 , b0 that are random and independent,
since by (11) for these pairs we have NID(x, y) > 1 − β, and by the assumption following (9) there are at least a fraction of 87 of such x0 , y0 .
• By assumption, if a0 , b0 are random and dependent, for > 41 of the x0 , y0 ,
if NID(x, y) < 21 + α then NIDs (x, y) < 21 + α for every s. Since for x0 , y0
with (9) we have NID(x, y) < 12 + α by (10), we have for a fraction > 81 of
x0 , y0 that NIDs (x, y) < 12 + α for every s. Hence if a0 , b0 are enumerated
and random, at least one of K(a0 |b0 ) > rn or K(b0 |a0 ) > rn must hold.
But the existence of such an enumeration contradicts Theorem 7.2. This concludes
the proof of the claim.
Next we consider the following subset of A:
B = (x, y) ∈ A | ∃s NIDs (x, y) < 21 + α in phase 2 .
By the above Claim, the set B is infinite. Namely, if NID(x, y) < 12 + α and
NIDs (x, y) > 21 + α in phase 1, then we have to see NIDs (x, y) < 21 + α in phase 2,
since this is the going down phase.
Case 1. Suppose that for almost all pairs x, y we have that
if NID(x, y) >
2
3
− γ then for all s, NIDs (x, y) >
2
3
− γ in phase 2.
(12)
Here γ is a small number to be determined later. We show that this implies
that there is an effective procedure to enumerate infinitely many pairs a0 , b0 ,
at least one of which is random. Consider all pairs x = a0 x0 , y = b0 y0 with
|a0 | = |b0 | = |x0 | = |y0 |. Enumerate the pair a0 , b0 if for a majority of at least 34
of the pairs x0 , y0 we have found s such that NIDs (x, y) < 12 + α in phase 2.
First, by the Claim above, there are infinitely many pairs a0 , b0 that are enumerated.
Second, if the pair a0 , b0 is enumerated, then a0 and b0 cannot both be compressible, by the assumption (12). Namely, suppose that both K(a0 ) 6 21 |a0 |
and K(b0 ) 6 21 |b0 |. A calculation as on page 6 shows that K(a0 x0 ) 6+ 34 n and
K(b0 y0 ) 6+ 34 n. Since by (9) we also have that K(a0 x0 |b0 y0 ) > 21 n − d and
K(b0 y0 |a0 x0 ) > 21 n − d, again similarly to the calculation on page 6 we have for
most x0 , y0 that asymptotically NID(x, y) >+ 23 . We fix γ > 0 to be a small
NORMALIZED INFORMATION DISTANCE AND THE OSCILLATION HIERARCHY
11
number such that NID(x, y) > 23 − γ holds for all but finitely many of these a0 , b0 .
Then if (12) holds, for all but finitely many of the enumerated pairs a0 , b0 , at least
one of them has high complexity.
In conclusion, we have an enumeration of infinitely many pairs a0 , b0 , at least
one of which has high complexity. But such an enumeration cannot exist, by the
immunity of the random strings. This shows that Case 1 cannot hold.
Case 2. In the opposite case, we have that (12) does not hold. Hence for
infinitely many pairs x, y we have that NID(x, y) > 23 − γ, but there exists s such
that NIDs (x, y) 6 23 − γ in phase 2. This means that this has to be corrected in
phase 3, so that infinitely often we see NIDs (x, y) > 32 − γ in phase 3. But this
contradicts the immunity from Proposition 2.3.
Since we have reached a contradiction in both cases, we conclude that NID ∈
/
Σ−1
.
3
9. NID is not Σ−1
4
−1
This follows from the Σ−1
3 -case in the same way as the Π3 -case follows from
the Π−1
2 -case, using Lemma 4.1.
10. The case Π−1
4
In this section we prove that Π−1
4 is needed for examples of a specific form. The
proof of this exhibits all the ingredients needed for the general case, which we prove
in section 11.
We start by giving an informal overview of the proof. As before, we let NIDs denote a computable approximation of NID. Assuming NIDs is a Π−1
4 -approximation,
for given x,y we label the four phases of the approximation NIDs (x, y) as follows:
?? ❄
❄❄
??
⑧⑧ ❄❄❄
❄❄
⑧⑧
phase
2
⑧
⑧
❄❄
❄❄
⑧⑧
⑧⑧
❄❄
❄❄
⑧⑧
⑧⑧ phase 4
❄
❄
⑧
⑧
phase 1 ❄❄ ⑧⑧
phase 3 ❄❄❄ ⑧⑧⑧
❄ ⑧⑧
⑧
We consider pairs of strings x = a0 a1 x1 and y = b0 b1 y1 such that |x| = |y| = n,
|ai | = |bi |, and |x1 | = |y1 |. We write x0 = a1 x1 and y0 = b1 y1 .
a0
✤
✤
b0
a1
✤
✤
b1
x1
✤
✤
y1
✤
✤
x
y
The typical situation is as follows: First x and y look random and independent,
forcing NIDs to guess a value close to 1. Next, a dependency between a0 and b0 is
found, which forces NIDs down. Then, if a0 and b0 are found to be compressible,
NIDs is forced up again. This reasoning shows that Π2 is needed for x = a0 x0
and y = b0 y0 when x0 and y0 are random and sufficiently independent. Next, we
repeat this argument, but now with a1 and b1 instead of a0 and b0 . We work in
the subset of strings where phase 2 of the approximation ends high, after a0 and b0
have been found to be compressible. For this to work, we need to assume that a0
and b0 are negligible in length compared to a1 and b1 , so we choose the latter of a
much larger length (in fact, so large that it actually does not matter whether a0 ,b0
are compressible or not). Now we repeat the argument, with x1 and y1 random
and sufficiently independent. First, if a dependency between a1 and b1 is found,
this forces NIDs down (this is phase 3). Finally, if a1 and b1 are found to be
compressible, NIDs is forced up again (phase 4). This shows that four phases are
12
K. AMBOS-SPIES, W. MERKLE, AND S. A. TERWIJN
needed for pairs x,y of this specific form. To implement this informal proof scheme,
we need to specify the extent to which the strings are random, independent, or
compressible, and also what “high” and “low” means.
Keeping the notation from above, we further use the notation
• q0 =
• q1 =
|a0 |
n ,
|a1 |
n .
so that in particular |x1 | = (1 − q0 − q1 )n.
For all of the discussion below, we assume that x1 ,y1 are random and independent:
K(x1 |(a0 a1 )∗ , y) > |x1 | − c and K(y1 |(b0 b1 )∗ , x) > |y1 | − c,
(13)
for some fixed constant c not depending on a0 ,b0 ,a1 ,b1 , or n. This holds for most
x1 ,y1 in the same way as we had before in (9) and (5).
There are three cases.
Case 1. First suppose that x0 ,y0 are independent and random with respect to
a0 ,b0 . More precisely, we assume in this case that the following analogue of (9)
holds:
K(x0 |a∗0 , b∗0 , y0 ) > |x0 | − d and K(y0 |b∗0 , a∗0 , x0 ) > |y0 | − d.
(14)
By (3) we have
K(a0 x0 ) 6+ n + 2 log n
K(b0 y0 ) 6+ n + 2 log n.
Furthermore, by (14) we have
K(a0 x0 |b0 y0 ) >+ K(x0 |b0 y0 ) > |x0 | − d
K(b0 y0 |a0 x0 ) >+ K(y0 |a0 x0 ) > |y0 | − d.
Noting that |x0 | = |y0 | = (1 − q0 )n, this gives
(1 − q0 )n − d
n + 2 log n
→ 1 − q0 .
NID(x, y) >
(15)
Case 2. Second, suppose that a1 ,b1 are random and dependent. Precisely:
• K(a1 |b1 ) 6 r|a1 | and K(b1 |a1 ) 6 r|b1 |,
• K(a1 ) > |a1 |, K(b1 ) > |b1 |.
Here we let r > 0 be a fixed small rational number, say r =
For the denominator of NID(x, y) we have
1
10
for definiteness.
K(a0 a1 x1 ) =+ K(a0 a1 ) + K(x1 |(a0 a1 )∗ )
(by (2))
> K(a1 ) + |x1 | − c
(by (13))
> |a1 | + |x1 | − c
(by assumption)
= n − |a0 | − c
= n − q0 n − c.
Note that (2) still holds if we know q0 and q1 . Likewise, K(b0 b1 y1 ) >+ n − q0 n − c.
NORMALIZED INFORMATION DISTANCE AND THE OSCILLATION HIERARCHY
13
Next, for the numerator of NID(x, y) we have
K(a0 a1 x1 |y) 6+ K(a0 a1 |y) + K(x1 |y)
(by subadditivity)
6 K(a0 a1 |b1 ) + K(x1 |n)
6 K(a0 |b1 ) + K(a1 |b1 ) + |x1 |
(by (4))
6 K(a0 |n) + r|a1 | + |x1 |
6 |a0 | + r|a1 | + |x1 |
6 q0 n + rq1 n + (1 − q0 − q1 )n
= n(1 − (1 − r)q1 ).
Similarly, we have K(b0 b1 y1 |x) 6+ n(1 − (1 − r)q1 ). This gives
n(1 − (1 − r)q1 )
n − q0 n − c
1 − (1 − r)q1
→
.
1 − q0
NID(x, y) 6
(16)
Case 3. Third, suppose that both a1 and b1 are compressible: K(a1 ) 6 21 |a1 |
and K(b1 ) 6 21 |b1 |. We then have
K(a0 a1 x1 ) 6+ K(a0 a1 ) + K(x1 |(a0 a1 )∗ )
(by (2))
6 K(a0 ) + K(a1 ) + K(x1 |n)
(by subadditivity)
6 |a0 | + 2 log |a0 | +
= q0 n + 2 log q0 n +
= 2 log q0 n + n −
1
2 |a1 | + |x1 |
1
2 q1 n + (1 − q0
(by (3))
− q1 )n
1
2 q1 n.
Likewise, K(b0 b1 y1 ) 6+ 2 log q0 n + n − 12 q1 n. For the numerator we have
K(x|y) > K(x1 |y)
> |x1 | − c
(by (13))
= (1 − q0 − q1 )n − c,
and likewise K(y|x) > (1 − q0 − q1 )n − c. Hence
(1 − q0 − q1 )n − c
2 log q0 n + n − 12 q1 n
1 − q0 − q1
→
.
1 − 12 q1
NID(x, y) >
(17)
We now choose the parameters q0 , q1 , and r that suit our needs. For the proof
below we need that there is a gap between the expressions (15), (16), and (17),
namely
1 − q0 − q1
1 − (1 − r)q1
1 − (1 − r)q1
and
.
1 − q0 >
>
1
1 − q0
1 − q0
1 − 2 q1
A straightforward computation shows that both of these inequalities are satisfied
whenever
q0 < rq1 and r 6 13
(18)
so we assume this from now on. (Note that we had previously already fixed r to
1
be 10
.) As the inequalities in (15), (16), and (17) hold asymptotically, for n → ∞,
and for our choice of the parameters q0 , q1 , and r there is a gap between the “low”
expression (16) and the “high” expressions (15) and (17), we can fix two rational
constants H (for high) and L (for low) with H > L so that for almost all lengths n,
NID(x, y) > H in Cases 1 and 3, and NID(x, y) < L in Case 2.
Now we are ready to prove the following
14
K. AMBOS-SPIES, W. MERKLE, AND S. A. TERWIJN
Theorem 10.1. Π−1
4 is needed on examples x = a0 a1 x1 , y = b0 b1 y1 of the above
form. That is, in general any computable approximation of NID(x, y) needs at least
four oscillations on examples of this form.
Proof. We use the notation and the choice of parameters from the discussion above.
In particular we have fixed values for q0 , q1 , r, H, and L. As before, NIDs denotes a
computable approximation of NID. The first four oscillations of NIDs are indicated
by phase 1 (going down), phase 2 (up), phase 3 (down), and phase 4 (up). Let A0
be the set of pairs (x, y) where phase 2 ends high after a0 and b0 have been found
to be compressible:
A0 = (x, y) | a0 , b0 compressible ∧ ∃s NIDs (x, y) > H in phase 2 .
We note that A0 is a c.e. set. The analysis in the proof of Theorem 5.3 shows
that A0 is infinite, namely, A0 includes infinitely many pairs x = a0 x0 , y = b0 y0
for which a0 ,b0 are compressible and x0 ,y0 are random and independent. In Theorem 5.3 this was shown for a0 = b0 and q0 = 12 , but the proof works equally well for
arbitrary q0 , using the computation preceding (15). (The reason that we do not
assume a0 = b0 here is to have a closer analogy with the reasoning at subsequent
stages of the proof.)
Now consider the parts x0 = a1 x1 and y0 = b1 y1 in the set of pairs (x, y) ∈ A0 .
The pairs a1 ,b1 cannot all be random, for then we could enumerate an infinite set
of random strings, as A0 is c.e., so some of these pairs a1 ,b1 must be compressible.
Also, among these pairs a1 ,b1 are infinitely many random pairs (i.e. where both
parts are random), because, as note above, A0 includes infinitely many pairs for
which x0 ,y0 are random. Not all of these random pairs can have a1 ,b1 independent,
because this would violate the conditional immunity of Theorem 7.2.
Hence among these a1 ,b1 are infinitely many random and dependent a1 ,b1 (as in
Case 2 above), which means that NIDs has to go low on these (assuming x1 ,y1 are
random and independent) in phase 3: There exists s such that NIDs (x, y) < L.
If phase 3 would stay high on almost all of the (x, y) ∈ A0 for which a1 ,b1 are
compressible, then if phase 3 goes low on a1 ,b1 for most x1 ,y1 , then at least one
of a1 ,b1 must be random. Since by the above, phase 3 does go low infinitely often
for such a1 ,b1 , this implies that we can enumerate infinitely many random strings,
contradicting Theorem 2.2.
We conclude that phase 3 goes low for infinitely many compressible a1 ,b1 (and
x1 ,y1 random and independent): There are infinitely many (x, y) ∈ A0 such that
a1 and b1 are compressible (as in Case 3) and such that there exists s with
NIDs (x, y) < L in phase 3. As for these x,y actually NID(x, y) > H, this approximation has then to be corrected in phase 4. Consider the set
A1 = (x, y) ∈ A0 | a1 , b1 compressible ∧ ∃s NIDs (x, y) > H in phase 4 .
Again, A1 is c.e., and we conclude from the above that A1 is infinite. As NIDs
was an arbitrary computable approximation, this proves the theorem.
11. NID is not Σ−1
n for any n
The following theorem confirms a conjecture made in [8].
Theorem 11.1. NID is not in Σ−1
n for any n > 1.
Theorem 11.1 immediately follows from the following generalization of Theorem 10.1. We now consider examples x = a0 a1 . . . ak xk , y = b0 b1 . . . bk yk , with
|ai | = |bi | for all i and |xk | = |yk |.
Theorem 11.2. Π−1
2k is needed on examples of the form x = a0 a1 . . . ak−1 xk−1 , y =
b0 b1 . . . bk−1 yk−1 . That is, in general any computable approximation of NID(x, y)
needs at least 2k oscillations on examples of this form.
NORMALIZED INFORMATION DISTANCE AND THE OSCILLATION HIERARCHY
15
Proof. We prove this by induction on k. The case k = 1 was treated in Theorem 5.3, and the case k = 2 in Theorem 10.1. The proof of the latter was set up
in sufficient generality that it will in fact give us the general case without much
extra effort.
We prove a statement that is slightly stronger. The induction hypothesis is
that Π−1
2k is needed on examples a0 a1 . . . ak−1 xk−1 , b0 b1 . . . bk−1 yk−1 , and, more
specifically, that the c.e. set
Ak−1 = (x, y) ∈ Ak−2 | ak−1 , bk−1 compressible ∧
∃s NIDs (x, y) > H in phase 2k .
is infinite. Now the proof of the induction step proceeds exactly as the proof of
Theorem 10.1, with the following translation:
Pk−1
|a0 | is replaced by i=0
|ai |
A0 is replaced by Ak−1
qi = |ani | for every i 6 k.
qi+1 = rqi for every i 6 k − 1.
Pk−1
The latter ensures that the length of ak is sufficiently large to make i=0
|ai |
negligible.
The calculations of Cases 1, 2, and 3, can also be reused, with suitable replacements:
Pk−1
q0 is replaced by i=0
|qi |
q1 is replaced by qk .
The existence of the constants H and L then follows in the same way. The
conclusion is that the c.e. set
Ak = (x, y) ∈ Ak−1 | ak , bk compressible ∧
∃s NIDs (x, y) > H in phase 2(k + 1) .
is infinite. In particular, the theorem holds for k + 1.
Π−1
n
Of course, Theorem 11.1 implies that also NID ∈
/
for any n. This means
that for any computable approximation of NID, the number of oscillations cannot
be bounded by a fixed constant n.
−1
We note that, although the Σ−1
3 -case (Theorem 8.1) is subsumed by the Π4 case (Theorem 10.1), the proof of the former gives more information, as it uses
only the examples a0 x0 , b0 y0 .
From the proof of Theorem 11.1 it is obvious that the examples of pairs of strings
x,y forcing the changes in the approximation NIDs are of rather long length. It
would be interesting to have a more careful analysis of these lengths.
Question 11.3. Relate the number of oscillations of approximations of NID(x, y)
to the length of x and y.
References
[1] C. H. Bennett, P. Gács, M. Li, P. M. B. Vitányi, and W. Zurek, Information distance, IEEE
Trans. Information Theory 44 (1998) 1407–1423.
[2] R. G. Downey and D. R. Hirschfeldt, Algorithmic Randomness and Complexity, SpringerVerlag, 2010.
[3] M. Li, X. Chen, X. Li, B. Ma, P. Vitányi, The similarity metric, IEEE Transactions on
Information Theory, 50(12) (2004) 3250–3264.
[4] M. Li and P. M. B. Vitányi, An introduction to Kolmogorov complexity and its applications,
third edition, Springer-Verlag, 2008.
[5] P. Odifreddi, Classical recursion theory, Vol. 1, Studies in logic and the foundations of
mathematics Vol. 125, North-Holland, 1989.
[6] V. L. Selivanov, Fine Hierarchies and Boolean Terms, Journal of Symbolic Logic 60 (1995)
289–317.
16
K. AMBOS-SPIES, W. MERKLE, AND S. A. TERWIJN
[7] V. L. Selivanov, Difference Hierarchy in φ-Spaces, Algebra and Logic 43(4) (2004) 238–248.
[8] S. A. Terwijn, L. Torenvliet, and P. M. B. Vitányi, Nonapproximability of the normalized
information distance, Journal of Computer and System Sciences 77 (2011) 738–742.
[9] P. M. B. Vitányi, F. J. Balbach, R. Cilibrasi, M. Li, Normalized information distance, pp.
45-82 in: Information Theory and Statistical Learning, F. Emmert-Streib and M. Dehmer
(eds.), Springer-Verlag, 2008.
(Klaus Ambos-Spies) Universität Heidelberg, Institut für Informatik, Im Neuenheimer Feld 205, 69120 Heidelberg
E-mail address: [email protected]
(Wolfgang Merkle) Universität Heidelberg, Institut für Informatik, Im Neuenheimer Feld 205, 69120 Heidelberg
E-mail address: [email protected]
(Sebastiaan A. Terwijn) Radboud University Nijmegen, Department of Mathematics,
P.O. Box 9010, 6500 GL Nijmegen, the Netherlands.
E-mail address: [email protected]
| 7cs.IT
|
arXiv:1803.06914v1 [math.CO] 11 Mar 2018
Discrete Mathematics and Theoretical Computer Science
DMTCS vol. VOL:ISS, , #NUM
Mixing Time of Markov chain of the Knapsack
Problem
Koko K. Kayibi1
1
2
3
S. Pirzada2∗
Carrie Rutherford3
Department of Mathematics and Physics, University of Hull, UK
Department of Mathematics, University of Kashmir, Srinagar, India
Department of Mathematics, University of South Bank, London, UK
received , revised , accepted .
To find the number of assignments of zeros and ones satisfying a specific Knapsack Problem is #P hard, so only
approximations are envisageable. A Markov chain allowing uniform sampling of all possible solutions is given by
Luby, Randall and Sinclair. In 2005, Morris and Sinclair, by using a flow argument, have shown that the mixing time
of this Markov chain is O(n9/2+ǫ ), for any ǫ > 0. By using a canonical path argument on the distributive lattice
structure of the set of solutions, we obtain an improved bound, the mixing time is given as
τx (ǫ) ≤ n3 ln(16ǫ−1 ).
Keywords: Knapsack problem, Markov chain, mixing time
1 Introduction
n
Let a = (a1 , a2 , . . . , an ) be a vector in R and let b be any real number. The Knapsack Problem, denoted
by K(n, b), consists of finding a vector x = (x1 , x2 , . . . , xn ), with xi ∈ {0, 1}, such that
< a, x > = a1 x1 + a2 x2 + . . . + an xn ≤ b.
(1)
Such a vector x is called a feasible solution. In practical terms, the problem can be phrased as follows.
Assume that a traveler disposes of a knapsack which can accommodate at most b kilograms. Let there
be n items such that item i weighs ai kg. How can we pack up so that the total weight does not exceed
the weight accomodable in the knapsack. This example indicates the relevance of the problem in various
walks of life. For instance, in small to medium-sized enterprises, the total weight allowed may be the total
investment available for a list of possible projects. For everyday travelers, the problem consists as how
we can pack up our luggage so not to exceed the total weight allowed by airlines regulations. There are
many applications in business, two of these are explored in [8] (marketing) and [3] (e-commerce).
In mathematical terms, the problem consists of finding the number of solutions of the Inequality 1. The
n
vector (0, 0, · · · , 0) is a trivial solution. Now, it is possible to check exhaustively all the 2 possibilities
∗ Supported
by SERB-DST, New Delhi under the research project number EMR/2015/001047/MS
ISSN subm. to DMTCS
c by the author(s)
Distributed under a Creative Commons Attribution 4.0 International License
2
Koko K. Kayibi, S. Pirzada, Carrie Rutherford
and see which of these satisfy Inequality 1. This scheme may work for small n. In case n is large, say
n > 100, to check exhaustively all the possible solutions will take a long time, and by the time we are
finished, the problem is obsolete (the flight is gone or the investment opportunities wasted). Hence, it is
advisable to sample a solution uniformly at random.
Let Ω(n, b) be the set of all solutions of the problem K(n, b). One way to sample a solution at random
is to construct a random walk which starts at a given element of Ω(n, b), then moves to another element
according to some simple rule which changes one element to another. One such simple change consists
of changing the entry xi to xi + 1 (mod 2). We call such a change a flip on the position i. A flip at a
point i is positive if xi goes from 0 to 1 and a flip at a point i is negative if xi goes from 1 to 0. With this
terminology, we let a neutral flip to be the fact of leaving an entry unchanged. After running the random
walk for t steps, we output the element reached by the random walk and consider this element, the tth
state reached, as a random sample. Such a random walk is given in [6, 11] as follows.
Take the solution x = (0, 0, . . . , 0) as the starting point. Stay on x with probability 12 , and with
probability 21 choose a position i in the vector x and perform a flip on the position i. Call the new vector
obtained by a single flip as y. Now, if y satisfies Equation 1, move from x to y, else stay at x. It is routine
to check that this random walk defines a Markov chain that can reach all the solutions and is aperiodic.
Thus, it is an ergodic Markov chain that tends asymptotically towards a stationary state π. So it may be
used to sample uniformly a solution of K(n, b) problem. The key issue now is to know the time taken to
reach the steady state.
Since the successive states of a Markov chain are not independent, an unbiased sample can only be
obtained if the chain reaches stationarity, the stage when the probability of sampling a particular configuration is fixed in time. The mixing time of a Markov chain is the number of steps necessary to reach that
stationary state. To know the mixing time of a particular Markov chain is crucial to avoid either getting a
biased sample (if stationarity is not reached), or to avoid the computational cost of running the chain more
than necessary. In [7], it is shown that the mixing time of Markov chain defined above is polynomial on
the length of the solutions of the problem. In this paper, by using a different approach which makes use
of the lattice structure of the solution set Ω(n, b), we obtain a better bound.
2 Preliminaries
Let M be a Markov chain on a set of states Ω. Let P be the matrix of transitions from one state to another.
We can visualize the Markov chain M as a weighted directed graph G, where the states are vertices of
G and there is an edge of weight P (x, y) if the transition from state x to state y has probability P (x, y).
A Markov chain is irreducible if, for all pairs of states x and y, there is an integer t, depending on the
pair (x, y), such that P t (x, y) > 0. In terms of the graph G, the Markov chain is irreducible if there is
a path between every pair of vertices of G. A Markov chain is aperiodic if for all states x, there is an
′
integer t such that for all t′ ≥ t, P t (x, x) > 0. That is, after sufficient number of iterations, the chain
has a positive probability of staying on x at every subsequent step. This ensures that the return to state x
is not periodic. In terms of the graph G, this can be achieved by having a loop at every vertex. A Markov
chain is ergodic if it is irreducible and aperiodic. If P is the matrix of transitions of an ergodic Markov
chain, there is an integer t such that for all the pairs of states (x, y), P t (x, y) > 0. (Notice that t does
not depend on the pair). It can also be proved that for every ergodic Markov chain with transition matrix
P , the largest eigenvalue of P is equal to 1. Using this, it can be proved that there is a unique probability
vector π, such that πP = π. The vector π is the stationary distribution of M. A chain is reversible if
Mixing Time of Markov chain of the Knapsack Problem
3
P (x, y)π(x) = P (y, x)π(y) for every pair of states x and y. If a chain is reversible and the matrix P is
symmetric, then π(x) = π(y) for every pair x and y. The stationary state is then said to be uniform. In
the obvious way, we say a matrix P is irreducible, ergodic, aperiodic, etc. if the Markov chain ruled by
P is irreducible, ergodic, aperiodic, etc.
Let λ1 = 1 and λ2 be the largest and the second largest eigenvalue of the transition matrix P . The real
number λ1 − λ2 , that is, 1 − λ2 is the spectral gap of the matrix P . It can be shown that the bigger the
gap, the faster the mixing time of the chain. The analysis of the mixing time of a Markov chain is based
on the intuition that a random walk on the graph G mixes fast ( i.e., reaches all the states quickly) if G has
no bottleneck. That is, there is no cut between any set of vertices S and its complement which blocks the
flow of the Markov chain and thus prevents the Markov chain from reaching easily to some states. See
[2, 4, 6, 11] for a better exposition on the topic, and for definitions in graph theory we refer to [10].
Denoting the probability of x at stationarity by π(x) and the probability of moving from x to y by
P (x, y), the capacity of the arc e = (x, y), denoted by c(e), is given as c(e) = π(x)P (x, y). Let Px,y
denote the set of all simple paths p from x to y (paths that containPevery vertex at most once). A flow in G is
a function φ, from the set of simple paths to the reals, such that p∈P φ(p) = π(x)π(y) for all vertices
x,y
P
x, y of G with x 6= y. An arc-flow along an arc e, denoted by φ′ (e), is then defined as φ′ (e) = p∋e φ(p).
For a flow φ, a measure of existence of an overload along an arc is given by the quantity ρ(e), where
′
(e)
, and the cost of the flow φ, denoted by ρ(φ), is given by ρ(φ) = maxe ρ(e).
ρ(e) = φc(e)
If a network G representing a Markov chain can support a flow of low cost, then it cannot have any bottlenecks, and hence its mixing time should be small. This intuition is confirmed by the following theorem
[11].
Theorem 2.1 [11] Let M be an ergodic reversible Markov chain with holding probabilities P (x, x) ≥
at all states x. The mixing time of M satisfies
1
1
τx (ǫ) ≤ ρ(φ)|p| ln
+ ln ,
π(x)
ǫ
1
2
where |p| is the length of a longest path carrying non-zero flow in φ.
Thus one way to prove that a Markov chain mixes fast is to produce a flow along some paths where the
paths and the maximal overload on edges are polynomials on the size of the problem.
3 Canonical path and Mixing time
Let G denote the graph whose vertices are all the solutions of the K(n, b), and there is an edge connecting
a vertex x to a vertex y if x can be turned into y by changing a single entry of the solution vector x. So G
is just an Hasse diagram, denoted as L(n, b), which is a top-truncated lattice. For illustration, let a1 = 5,
a2 = 3, a3 = 2, a4 = 1, and b = 9. The set of solutions of the inequality 5x1 + 3x2 + 2x3 + x4 ≤ 9 is
given in Figure 1.
Now, we define formally a random walk on G as follows. Start at any vertex z. If the walk is at x, move
1
if there is an edge connecting x to y in G and < y, a > ≤ b, and
from x to y with probability Px,y = 2n
stay at x otherwise. Obviously, the probability of staying at x is at least 12 . Now, since G is just the lattice
of flats of the free matroid Un,n minus the upper part which contains the points y such that < y, a > 6≤ b,
4
Koko K. Kayibi, S. Pirzada, Carrie Rutherford
(1,1,0,1)
(1,1,0,0)
(1,0,1,0)
(1,0,0,0)
(0,1,1,0)
(0,1,0,0)
(1,0,1,1)
(1,0,0,1)
(0,0,1,0)
(0,1,1,1)
(0,1,0,1)
(0,0,1,1)
(0,0,0,1)
(0,0,0,0)
Fig. 1: The solutions of the inequality 5x1 + 3x2 + 2x3 + x4 ≤ 9 ordered by containence.
so G is connected. Therefore the matrix P is irreducible. Moreover, since P (x, x) ≥ 21 , there is a loop
on every point x, the matrix P is aperiodic. So the random walk is an ergodic Markov chain. Finally, P
is a symmetric matrix. Thus, the Markov chain has a stationary uniform distribution π. We denote this
Markov chain as Mknap .
The following theorem is our main result.
Theorem 3.1 Let L(n, b) be the Hasse diagram of the solutions of the inequality < a, x > ≤ b, for any
real vector a of length n. The Markov chain Mknap on L(n, b) mixes fast and the mixing time τ , is given
by
τx (ǫ) ≤ n3 ln(16ǫ−1 ).
The proof is similar to that of papers [4, 5], and requires the following background. A path from v
to w is a sequence of binary vectors (x(1) , x(2) , . . . , x(r) ), such that x(1) = v and x(r) = w, and x(i+1)
is obtained from x(i) by a single flip. We prove this result by showing that, between any two vertices
of L(n, b), there is a path, the canonical path (defined below), that is not ‘too long’ and where no edge
is overloaded by the ‘probability fluid’. To put this in a more formal setting, we need the following
definitions and lemmas.
Let xi denote the ith entry of x. Let x and y be two solutions of the inequality < a, z > ≤ b. We say
that x and y are matched at the ith entry if xi = yi . If x and y are not matched at the entry k, to match
the point k means to do the operation yk = xk + 1(mod 2). That is, if xk = 0, change it into 1, and
vice-versa.
The canonical path from v to w is the path that consists of matching the entries in increasing order of
the indices, starting from i = 1 up to i = n. But, if flipping the position i positively (changing its entry
from 0 to 1) leads to a vector that is not in L(n, b), then we first flip negatively (change 1 to 0) the nearest
positions k1 , k2 , . . . , kr , such that k1 < k2 < . . . < kr , kj > i, for 1 ≤ j ≤ r, vk = 1 and wk = 0, and
j
j
P
ak ≥ ai . That is, the canonical path first flips as many nearest positions greater than i as possible, so
j
j
that flipping afterwards the position i does not lead outside L(n, b). In simpler words, the Canonical path
is given by Algorithm 1.
Algorithm 1. To get from one feasible solution v = (v1 , · · · , vn ) to another w = (w1 , · · · , wn ), one
scans from left to right, flipping variables from 0 to 1 and vice-versa, as required. Thus an intermediate
Mixing Time of Markov chain of the Knapsack Problem
5
state is a feasible solution (w1 , · · · , wi−1 , vi , · · · , vn ). However, changing vi = 0 to wi = 1 might violate
the linear inequality. In this case, one jumps vi , and continues scanning right, changing 1s to 0s, as
necessary, until there is enough slack to allow vi to be flipped.
(Informally, one may visualize the canonical path as follows. Suppose that the items one wants to pack
up in a bag are ranked from 1 to n, according to their importance. Let there be two bags, v and w, both
satisfying the requirement of not being overloaded. Assume that one would like to make v carry the same
items as w. Then starting from the most valuable item, say item 1, one checks whether or not the item i
is contained in the bag w. If not, and v also does not contain it, then nothing to do. In case v contains it,
then remove this from v. Suppose that w contains item i. If v also contains it, then leave it alone. If not,
check whether inserting item i into v would overload the bag. If there is no overload, insert item i into v.
If there is an overloading, then first remove from v the items less valuable than the item i and that are not
in w. One has to remove as many items as it is necessary to avoid the item i to overload the bag.)
With this definition, suppose that Algorithm 1 is moving from v to w, and at some instant, the highest
entry flipped is vl . Then the sequence of indices (1, 2, . . . , n) can be partitioned in three parts. Part one,
called the Matched Zone, is the ordered sequence of indices (1, 2, . . . , k) such that vi = wi , for all i from
1 to k. Part two, called the Heap Zone, is the sequence of indices (k + 1, k + 2, . . . , l), where l is the
index of the last entry flipped along the canonical path, and where some entries are matched (flipped from
1 to 0 to avoid running into a binary vector that does not satisfy the inequality), and where some other
entries are not flipped (they are jumped by Algorithm 1, either because vj = 0 or vj = wj = 1). Finally,
part three, the Untouched Zone, consists of the sequence (l + 1, l + 2, . . . , n), the sequence of indices of
entries not yet touched by Algorithm 1.
Lemma 3.2 For all pairs of binary vectors v and w belonging to L(n, b), there is a canonical path
(x(0) , x(1) , . . . , x(r) ) such that x(0) = v and x(r) = w, and for all 0 ≤ j ≤ r, x(j) ∈ L(n, b).
Proof. Let v and w belong to L(n, b) but there is no such a canonical path from v to w. That is, Algorithm
1 gets stalled at some point. This is possible only if there is an entry vi whose flipping positively will
P lead
to a vector v ′ such that v ′ 6∈ L(n, b), but for all j such that j > i , vj = 1, wj = 0, we have ai > j aj .
That is, even after Algorithm
P1 had flipped negatively all such entries vj , flipping vi would still lead to a
binary vector v ′ where b < i vi′ .ai (That is, v ′ is outside L(n, b)).
Now, suppose that the canonical path is at its tth iteration when it has to flip the entry vi . Thus, there are
n − t positions left to match before reaching w. Let there be s positions j such that j > i , vj = 1, wj = 0
(points that may be flipped to avoid exiting L(n, b)). ( Obviously, n − t ≥ s, since there may still be positions k such that vk = 1, wk = 1 or vk = 0, wk = 0, or vk = 0, wk = 1). Since flipping all the
P s positions
j > i such that vj = 1, wj = 0 still leads to a binary vector outside L(n, b), we have ai > j aj . Thus,
w 6∈ L(n, b). This is clearly a contradiction.
✷
One of the most important properties of Algorithm 1 is that once matched, an entry can not be unmatched anymore. That is, the matched zone only increases along the canonical path. This entails the
following lemma.
Lemma 3.3 For all binary vectors v and w that belong to L(n, b), the length of the canonical path from
v to w is at most n.
6
Koko K. Kayibi, S. Pirzada, Carrie Rutherford
Proof. Since every entry is matched at most once, the length of the path is at most the number of entries
in v.
✷
In order for the canonical paths argument to work, we need the number of paths through any edge of
the upper-truncated lattice L(n, b) to be bounded above by a small polynomial of n times N , the number
of valid knapsack solutions. And indeed, in what follows, for any arc e = (z, y) of L(n, b), the number
of different canonical paths passing through e is at most 2N . But, before proving this crucial result in
Corollary 3.7, we need the following definitions and lemmas. Let v and z be elements of L(n, b). We say
an entry vi is determined by another entry zi if it is possible to know the value of vi from the value of zi .
That is, vi and zi are not independent, since either vi = zi , or vi = zi + 1(mod 2). We say that the index
i is free in a vector v if vi = 0 or vi = 1 independently of any entry of z. We recall that the canonical path
(v, . . . , w) passes through e = (z, y) means that to change v into w, Algorithm 1 must make a sequence
of flips that change v into z, then z into y, then y into w. See Figure 2 for an illustration.
vs
wk
vs−1
e
v4
z
y
v3
w3
v2
w2
v1
w1
Fig. 2: Different canonical paths that pass through the arc e.
Lemma 3.4 Let G be the upper-truncated lattice whose vertices are all the elements of L(n, b). Let z
and y be two such vertices, such that zi = yi for all the indices except r, and let e = (z, y) be the edge
that represents the flip matching the rth entry of z (That is, the flip transforming z into y). For any two
vertices v and w, the canonical path (v, . . . , w) passes through e only if the set of indices that are free in
v is disjoint from the set of indices that are free in w.
Proof. Suppose that the canonical path from v to w passes through the arc e at the rth entry of z. Consider
the Matched Zone, the Untouched Zone and the Heap Zone at the instant when the flip is performed at r.
We aim at showing the following.
(1) The Matched Zone (from index 1 to k) is free in v and determined by y in w.
(2) The Untouched Zone (from index l + 1 to n) is free in w and determined by z in v.
(3) In the Heap Zone, entries that are flipped carry 0 in z, and must carry 1 in v.
(4) In the Heap Zone, entries that are not flipped and carry 0 in z must carry 0 in v. In the Heap Zone,
entries that are not flipped and carry 1 in z must carry 1 in v.
Mixing Time of Markov chain of the Knapsack Problem
7
Indeed, the canonical path from v to w passes through e = (z, y) only if the sequence of flips from the
first one to the lth flip change v to z. Thus, (1), all the entries of v in the Untouched Zone must be equal to
the corresponding entries in z, since there are untouched, but they are are free in w. Now, (2), after changing z to y, the remaining flips must change y to w. Since the canonical path does not un-match positions
that are already matched, entries of w in the Matched Zone must be equal to the corresponding entries in
y, while they are free in v. In the Heap zone, entries are either flipped or left untouched (jumped). If, (3),
they were flipped, thus they are 0 in z and they must be 1 in v. Those are the entries that are flipped from
1 to 0 to get some slack. If, (4), they are jumped, they may be 0 or 1 in z. If they are 0 in z, they must be
0 in v, and if they are 1 in z, they are also 1 in v, since the Canonical Path from v to w reaches z at that
instant means that the untouched entries are equal. Hence in the Heap zone, entries of v are determined
by entries of z.
✷
Lemma 3.5 Let Ω(n, b) be the set of solutions of the inequality a1 x1 + a2 x2 + · · · + an xn ≤ b and let
|Ω| = N . Then
N = N ′ + N ′′ ,
(2)
where N ′ is the number of solutions of the inequality a2 x2 + . . . + an xn ≤ b and N ′′ is the number of
solutions of the inequality a2 x2 + . . . + an xn ≤ b − a1 .
Proof. (First, note that there is nothing special with choosing to remove x1 . Any other variable would do
as well.) The set of solutions of the inequality a1 x1 + a2 x2 + . . . + an xn ≤ b can be partitioned into the
sets A and B, where A is the set of solutions with x1 = 0 and B is the set of solutions with x1 = 1. The set
A is in one-one correspondence with the set of solutions of the the inequality a2 x2 + . . . + an xn ≤ b, and
the set B is in one-one correspondence with the set of solutions of inequality a2 x2 +. . .+an xn ≤ b−a1 . ✷
Corollary 3.6 Let z be a given feasible solution, let α + β = n, let |Ω| = N , and let V be the set of all
α
the solutions where the first β entries are determined by z. Then |V | is less or equal to (2N ) n .
Proof. (Obviously, |V | is just the number of solutions aβ+1 xβ+1 + . . . + an xn ≤ b − a1 − a2 − · · · − aβ ).
The proof uses induction on α, the number of indices that are free. If α = 0 (i.e., all the variables xi
0
are determined), the number of solutions is less or equal to N = 1. (We may also check the formula by
n
taking α = n. If all the variables are free, there are N n = N possible feasible solutions). For induction,
assume that the result holds for α − 1. Let M denote the number of feasible solutions where α variables
are free, and suppose that xn is free. Then, by Lemma 3.5, M = M ′ + M ′′ , where M ′ is the number of
solutions of the inequality a1 x1 + . . . + an xn−1 ≤ b and M ′′ is the number of solutions of the inequality
a1 x1 + . . . + an xn−1 ≤ b − an . We have
M
=
M ′ + M ′′
≤
N′
≤
≤
α−1
n
(2N ′ )
+ N ′′
α−1
n
α−1
n
by inductive hypothesis
since N ′ ≥ N ′′
α
n
(2N ) .
✷
8
Koko K. Kayibi, S. Pirzada, Carrie Rutherford
Corollary 3.7 If e = (z, y) is an arc of L(n, b), the number of different canonical paths passing through
e is at most 2N , where N is the number of vertices of L(n, b).
Proof. Let the canonical path from v to w passes through the arc e. By Lemma 3.4, the set of indices that
are free in v is disjoint from the set of indices that are free in w. Let α and β, with α + β ≤ n, be the
number of indices that are free in v and w, respectively. By Corollary 3.6, for a fixed feasible solution w′ ,
α
there are at most N n vectors v such that the canonical path (v, . . . , w′ ) passes through e = (z, y), and for
β
a fixed vertex v ′ , there are at most N n vectors w such that the canonical path (v ′ , . . . , w) passes through
e = (z, y). Therefore, for a fixed vector z and fixed index k such that e is the edge consisting of flipping
α
β
the kth entry of z, the total number of canonical paths passing through e is at most (2N ) n (2N ) n ≤ 2N .
✷
Proof of Theorem 3.1. In order to prove Theorem 3.1 by using Theorem 2.1, we show that there is a
flow φ such that ρ(φ) is a polynomial in n, the size of a vector solution of K(n, b). Indeed, if x and y are
two vertices of G, let p̂xy denote the canonical path from x to y and let φ be a flow consisting of injecting
π(x)π(y) units of flow along p̂xy . Then, for all arcs e, we have
X
φ′ (e) =
π(x)π(y),
where the sum is over all the pairs {x, y} such that e ∈ p̂xy . Since by Corollary 3.7, there are at most 2N
canonical paths through e, and π(x) = π(y) = N1 , as the distribution π is uniform, we have
φ′ (e) ≤ 2N π(x)π(y) ≤
Moreover, since Px,y =
1
2n
2
.
N
if there is an edge from x to y, we have
1
.
2nN
(3)
maxe φ (e)
≤ 4n.
ρ(φ) ≤
mine c(e)
(4)
c(e) = π(x)Px,y ≥
Thus,
′
Now, using Theorem 2.1, Lemma 3.3 and Equation 4, we have
τx (ǫ) ≤ (n)(4n)(ln
Finally, using the fact that
1
π(x)
1
1
+ ln ).
π(x)
ǫ
n
= N ≤ 2 , we have
1
τx (ǫ) ≤ 4n2 (n ln 2 + ln ) ≤ n3 ln(16ǫ−1 ).
ǫ
✷
Acknowledgements. The research of second author is supported by SERB-DST, New Delhi under the
research project number EMR/2015/001047/MS.
Mixing Time of Markov chain of the Knapsack Problem
9
References
[1] L. K. Grover and P. Kaur: An improved estimator of the finite population mean in simple random
sampling, Model Assisted Statistics and Applications 6, 1 (2011) 47-55.
[2] M. Jerrum and A. Sinclair: Approximating the permanent, SIAM J. Comput. 18, 6 (1989) 11491178.
[3] S. Kameshwaran, Algorithms for Piecewise Linear Knapsack Problems with Applications in Electronic Commerce, PhD Thesis, IISc 2004.
[4] K. K. Kayibi and S. Pirzada: T-Tetrominoes tiling’s Markov chain mixes fast, Theor. Comp. Sci. 714
(2018) 1-14.
[5] K. K. Kayibi and S. Pirzada: Sampling Contingency Tables, AKCE Internat. J. Graphs Comb. to
appear.
[6] M. Luby, D. Randall and A. Sinclair: Markov chain algorithms for planar lattice structures, SIAM
J. Comput. 31, 1 (2001) 167-192.
[7] B. Morris and A. Sinclair: Random walks on truncated cubes and sampling 0-1 knapsack solutions,
Siam J. Comp. 34 (2005) 195–226.
[8] E. A. Owoloko and E. T. Sagoe: Optimal advert placement slot using the knapsack problem model,
Am. J. Sci. Ind. Res. 1, 1 (2010) 51-55
[9] J. G. Oxley: Matroid Theory, Oxford University Press, New York (1992).
[10] S. Pirzada: An Introduction to Graph Theory, Universities Press, Orient Blackswan, Hyderabad,
2012.
[11] A. Sinclair: Convergence rates for Monte Carlo Experiments, Numerical Methods for Polymeric
Systems, S.G. Whittington, ed., IMA Volumes in Mathematics and its Applications, (1997) 1-18.
| 8cs.DS
|
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
arXiv:1711.10365v1 [math.AC] 27 Nov 2017
Ilaria Del Corso1 and Roberto Dvornicich2
Abstract. In [Fuc60, Problem 72] Fuchs posed the problem of characterizing the groups
which are the groups of units of commutative rings. In the following years, some partial
answers have been given to this question in particular cases. In a previous paper [DCD17]
we dealt with finite characteristic rings. In this paper we consider Fuchs’ question for
finite groups and we address this problem in two cases. Firstly, we study the case of
torson-free rings and we obtain a complete classification of the finite groups of units which
arise in this case. Secondly, we examine the case of characteristic zero rings obtaining, a
pretty good description of the possible groups of units equipped with families of examples
of both realizable and non-realizable groups. The main tools to deal with this general
case are the Pearson and Schneider splitting of a ring [PS70], our previous results on finite
characteristic rings [DCD17] and our classification of the groups of units of torsion-free
rings.
As a consequence of our results we completely answer Ditor’s question [Dit71] on the
possible cardinalities of the group of units of a ring.
MSC: 16U60,13A99, 20K15.
1. Introduction
In this paper we consider the famous problem posed by Fuchs in [Fuc60, Problem 72]:
Characterize the groups which are the groups of all units in a commutative and
associative ring with identity.
A partial approach to this problem was suggested by Ditor [Dit71] in 1971, with the
following less general question:
Which whole numbers can be the number of units of a ring?
In the following years, these questions have been considered by many authors. A
first result is due to Gilmer [Gil63], who considered the case of finite commutative rings,
classifying the possible cyclic groups that arise in this case. A contribution towards the
solution of the problem can be derived from the results by Hallett and Hirsch [HH65],
and subsequently by Hirsch and Zassenhaus [HZ66], combined with [Cor63]. In fact,
Hallett and Hirsch dealt with the different problem of describing the group Aut(G) when
G is a torsion free group, and found necessary conditions for a finite group to be the
automorphism group of a torsion free group. Now, if G is an abelian group, then Aut(G)
is the group of units of the ring End(G) and a deep result of Corner [Cor63] shows that
any countable, torsion-free and reduced ring is in fact isomorphic to End(G) for some
countable and torsion-free abelian group G. It follows that the results in [HH65] produce
a proof (although rather indirect) that if a finite group is the group of units of a reduced
1
(I. Del Corso): Dipartimento di Matematica Università di Pisa, e-mail: [email protected]
(R. Dvornicich): Dipartimento di Matematica Università di Pisa, e-mail: [email protected]
2
1
2
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
and torsion free ring, then it must satisfy some necessary conditions, namely, it must be
a subgroup of a direct product of groups of a given family.
Later on, Pearson and Schneider [PS70] combined the result of Gilmer and the result
of Hallett and Hirsch to describe explicitly all possible finite cyclic groups that can occur
as A∗ for a commutative ring A.
Other instances on this subject were considered also by other authors, like Eldridge
and Fischer [EF67], who considered rings with the descending chain condition, or Dolžan
[Dol02], who classified the finite rings A for which A∗ is a finite group whose order is
either a prime power or not divisible by 4. Recently, Chebolu and Lockridge [CL15]
[CL17] have studied Fuchs’ problem in a different setting, and have been able to classify
the indecomposable abelian groups and the finite dihedral groups that are realizable as the
group of units of a ring.
For short, we shall call realizable a finite abelian group which is the group of units of
some commutative ring.
In a previous paper [DCD17] we considered Fuchs’ and Ditor’s questions in the case
when A is a finite characteristic ring, obtaining necessary conditions ([DCD17, Thm 3.1])
for a group to be realizable in this case, and therefore to produce infinite families of nonrealizable groups. On the other hand, we also gave positive answers; in fact, we exhibited
large classes of groups that are realizable. Moreover, our results allowed us to completely
answer Ditor’s question for finite characteristic rings.
In this paper we consider Fuchs’ question for finite groups and rings of any characteristic.
In Section 3 we classify the possible groups of units in the case of integral domains, proving
the following (see Theorem 3.2)
Theorem A The finite abelian groups that occur as group of units of an integral
domain A are:
i) the multiplicative groups of the finite fields if char(A) > 0;
ii) the cyclic groups of order 2,4, or 6 if char(A) = 0.
In Section 4 we study the case of torson-free rings and we obtain a complete classification
of the finite groups of units which arise in this case. The result is the following (see
Theorem 4.1)
Theorem B The finite abelian groups which are the group of units of a torsion-free
ring A, are all those of the form
(Z/2Z)a × (Z/4Z)b × (Z/3Z)c
where a, b, c ∈ N, a + b ≥ 1 and a ≥ 1 if c ≥ 1.
In particular, the possible values of |A∗ | are the integers 2d 3c with d ≥ 1.
The more relevant part of the proof consists in producing explicit examples of rings
realizing these groups of units. In fact, for c = 0 one can simply choose A = Za ×
Z[i]b , whereas, for c ≥ 1, one needs a more sophisticated construction, as described in
Proposition 4.7.
In Section 5 we consider Fuchs’ question in the general case when A is a characteristic
zero ring and A∗ is finite. Defining ε(A) as the minimum exponent of 2 in the decomposition of the 2-Sylow of A∗ as direct sum of cyclic groups, we prove (see Theorem
5.1)
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
3
Theorem C The finite abelian groups which are the group of units of a ring A of
characteristic 0 have the form
Z/2ε Z × H
(1)
where ε = ε(A) = 1, 2 and H is an abelian group.
It is known that all groups of the form Z/2Z × H, where H is any abelian group, are
realizable; on the other hand, we show that, in the case when ε(A) = 2, this is no longer
true and a more subtle analysis is necessary. A fundamental tool for the study of this
general case is Proposition 5.9 (due to Pearson and Schneider) which allows to reduce
the study to finite rings and to rings whose torsion elements are nilpotent (which we call
type 2 rings). Finite rings were studied in [DCD17], whereas the study of type 2 rings is
done by using the results of Section 4 on torsion-free rings. The outcome is a pretty good
description of the situation, which allows us to find families of non-realizable groups (see
Proposition 5.13) and to produce large families of realizable groups (see Proposition 5.8).
In particular, we can easily reobtain the classification of the cyclic realizable groups.
Finally, we include an Appendix where we present some results concerning the densities
of the cardinalities of realizable groups A∗ in the set of the natural numbers. In Proposition
A.2 we prove that the density of the cardinalities of all realizable groups is equal to 21 ,
whereas in Proposition A.3 we show that restricting to reduced rings the density is zero.
2. Notation and preliminary results
Let A be a ring with 1 and, as usual, denote by A∗ its multiplicative group, which we
shall assume to be finite and abelian throughout the paper. The following remark shows
that for the study of all possible groups A∗ we may restrict to a particular class of rings.
Remark 2.1. Let A0 be the fundamental subring of A (namely Z or Z/nZ depending on
whether the characteristic of A is 0 or n) and let B = A0 [A∗ ]. Then trivially B ∗ = A∗ .
Hence, without loss of generality, we shall assume that A = A0 [A∗ ].
It follows that for our purposes we can assume that A is commutative, finitely generated
and integral over A0 . We will make this assumption throughout the paper
For a commutative ring A, we denote by N = N(A) its nilradical, namely, the ideal of
the nilpotent elements. A ring is called reduced if its nilradical is trivial.
We note that replacing A with A0 [A∗ ] does not change the property of being reduced
because they have the same nilradical. In fact, the nilradical of A0 [A∗ ] is obviously
contained in the nilradical of A; on the other hand, if a ∈ A is such that an = 0 for some
n ∈ N, then 1 + a ∈ A∗ , so a ∈ A0 [A∗ ].
An element a ∈ A is a torsion element if its additive order is finite. A is torsion-free if
0 is its only torsion element.
For each n ≥ 2 we will denote by ζn a complex primitive n-th root of unity.
Lemma 2.2. Let A be a commutative ring with 1 and assume that A∗ is finite. Let
x ∈ N, then x is a torsion element. In particular, if A is torsion-free, then it is reduced.
Proof. We know that 1 + N ⊆ A∗ ; this yields that 1 + N, and hence N, is finite. Let
x ∈ N; the elements of the Z-module {nx}n∈Z are all nilpotents, hence they can not be
all distinct. This means that x is a torsion element.
4
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
In the following we shall make repeated use of the following proposition (see [DCD17,
Prop.2.2]):
Proposition 2.3. Let A be a commutative ring and let N be its nilradical. Then the
sequence
φ
1 → 1 + N ֒→ A∗ → (A/N)∗ → 1,
(2)
where φ(x) = x + N, is exact.
Although for finite characteristic rings the exact sequence (2) always splits (see [DCD17,
Thm 3.1]), this is no longer true in general, as shown in Section 5 (see Examples 2).
3. Integral Domains
In this section we characterize the finite abelian groups A∗ when A is a domain. The
fundamental result to deal with this case is the famous Dirichlet’s units theorem. For the
sake of completeness, we quote here the generalization of the classical Dirichlet’s units
theorem to orders of a number field (see [Neu99, Ch.1, §12] for a proof).
Theorem 3.1 (Dirichlet). Let K be a finite extension of Q such that [K : Q] = n and
denote by R an order of the algebraic integers of K. Assume that among the n embeddings
of K in Q̄, r are real (namely map K into R) and 2s are non-real (n = r + 2s). Then
R∗ ∼
= T × Zr+s−1
where T is the group of the roots of unity contained in R.
Theorem 3.2. The finite abelian groups that occur as group of units of an integral
domain A are:
i) the multiplicative groups of the finite fields if char(A) > 0;
ii) the cyclic groups of order 2,4, or 6 if char(A) = 0.
Proof. If A is a domain, then clearly also A0 [A∗ ] is a domain, so we may assume A =
A0 [A∗ ].
If the characteristic of A is a prime number p, then A is a finite domain, hence A is a
field. It follows that A∗ is the multiplicative group of a finite field. Trivially, any finite
field Fq produces its multiplicative group F∗q .
If char(A) = 0, then A is a finitely generated Z-module, so its quotient field K is a
number field and, by Remark 2.1, A is contained in the ring of integers OK of K. This
guarantees that A is an order of K and the structure of its group of units is described by
Dirichlet’s units theorem. In particular, if A∗ is finite, then r + s − 1 = 0, so the quotient
field K of A = Z[A∗ ] is Q or a quadratic imaginary field. It follows that, if A is a domain
of characteristic 0 and A∗ is finite, then A∗ is a subgroup of the cyclic group of roots of
unity contained in an imaginary quadratic field. Now, since A∗ has even order (Z ⊆ A),
the possibilities for A∗ are {±1}, hζ4 i and hζ6 i. On the other hand, it is clear that each
of those groups does occur as A∗ for some domain A; for example, we can take A = Z,
Z[ζ4 ] and Z[ζ6 ], respectively.
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
5
4. Torsion-free rings
The case of torsion-free rings has already been considered in the literature. Actually,
if A∗ is finite and A is torsion-free then, by Lemma 2.2, it is also reduced, and one can
deduce from [HH65] (by using the deep result of Corner [Cor63]), that, if A∗ is finite
and abelian, then it is a subgroup of a group of the form (Z/2Z)a × (Z/4Z)b × (Z/3Z)c .
However, the question of which groups actually occur remained open: in the next theorem
we find a condition on the exponents a, b, c which is necessary and sufficient for the group
to be realizable, so we completely classify the groups of units in this case.
We present an independent proof of the necessity of the condition. As to the sufficiency, we produce examples of all realizable groups by means of the rather sophisticated
construction given in Proposition 4.7.
Theorem 4.1. The finite abelian groups which are the group of units of a torsion-free
ring A, are all those of the form
(Z/2Z)a × (Z/4Z)b × (Z/3Z)c
where a, b, c ∈ N, a + b ≥ 1 and a ≥ 1 if c ≥ 1.
In particular, the possible values of |A∗ | are the integers 2d 3c with d ≥ 1.
In the following lemmas we assume A to be a torsion-free ring such that A∗ is finite.
Lemma 4.2. Let α ∈ A∗ and let ϕα : Z[x] → A be the homomorphism defined by
p(x) 7→ p(α). Then ker(ϕα ) is a principal ideal generated by a non-constant polynomial.
Proof. Let I = ker(ϕα ) and let I = (f1 (x), . . . , fn (x)). Denote by IQ[x] the extension of
I to Q[x]. Then, IQ[x] = (d(x)), where
d(x) =
n
X
ai (x)fi (x)
(3)
i=1
for some ai (x) ∈ Q[x]. Up to multiplication by a non-zero constant (this does not change
the ideal generated by d(x) in Q[x]), in equation (3) we can assume that ai (x) ∈ Z[x] for
i = 1, . . . , n, and therefore d(x) ∈ I. Let d(x) = cd1 (x), where c ∈ Z \ {0} and d1 ∈ Z[x]
is primitive. Since d(α) = cd1 (α) = 0 and A is torsion-free, we have that d1 (α) = 0, so
d1 (x) ∈ I. On the other hand, if f ∈ I, then d1 (x) divides f (x) in Q[x] and, by Gauß
Lemma, d1 (x) divides f (x) in Z[x] too. In conclusion I = (d1 (x)).
To conclude the proof, we note that d1 (x) has positive degree, since A is torsion-free
and therefore I ∩ Z = {0} .
Remark 4.3. The previous lemma is not true if the ring A is not torsion-free. In fact,
let A = Z[x]/I where I = (x2 , 2x) and let α = x̄ ∈ A; then I is the kernel of the map
p(x) 7→ p(α) but it is not principal.
Lemma 4.4. Let p be a prime, let k > 0 and assume that pk 6= 2, 3, 4. Then in A∗ there
are no elements of order pk .
Proof. Assume, on the contrary, that there exists a ring A fulfilling the hypotheses and
an element α ∈ A∗ of order pk , with pk 6= 2, 3, 4. Note that in this case φ(pk ) > 2. By
Lemma 4.2, we know that Z[x]/ ker(ϕα ) is a subring of A and ker(ϕα ) is a principal ideal
6
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
generated by a non constant polynomial d(x) ∈ Z[x], say. We claim that Z[x]/(d(x))
contains a unit of infinite order.
k
The element α has order pk , so ker(ϕα ) = (d(x)) ⊇ (xp − 1), namely,
k
d(x) | (xp − 1) = (xp
k−1
− 1)Φpk (x),
where Φpk (x) denotes the pk -th cyclotomic polynomial. Since Φpk (x) is irreducible in Z[x]
we have two cases: either (d(x), Φpk (x)) = 1 or (d(x), Φpk (x)) = Φpk (x).
k−1
k−1
= 1, contrary to our
However, if (d(x), Φpk (x)) = 1, then d(x) | xp − 1, hence αp
assumption.
It follows that necessarily (d(x), Φpk (x)) = Φpk (x). Set d(x) = h(x)Φpk (x) for some
polynomial h(x) ∈ Z[x]. The Chinese Remainder Theorem gives an injection
ψ : Z[x]/(d(x)) → Z[x]/(h(x)) × Z[x]/(Φpk (x)).
(4)
Clearly the composition of the map ψ with the projection to each factor is surjective,
whereas Im(ψ) = {(ā, b̄) | a − b ∈ (h(x), Φpk (x))}, where ā and b̄ denote the projections
of elements of Z[x] onto their respective quotients.
We note that (h(x), Φpk (x)) = (p, h(x)) : in fact, one sees immediately that Φpk (x) ≡ p
k−1
k−1
(mod xp − 1), so p ∈ (xp − 1, Φpk (x)) ⊆ (h(x), Φpk (x)). On the other hand, Φpk (x) ≡
k−1
k−1
(xp − 1)p−1 (mod p), hence Φpk (x) ∈ (p, xp − 1) ⊆ (p, h(x)). Summing up we have
Im(ψ) = {(ā, b̄) | a − b ∈ (p, h(x))}.
We shall obtain a contradiction by showing that Im(ψ), and hence Z[x]/(d(x)), contains
a unit of infinite order. Since φ(pk ) > 2, by Dirichlet’s unit Theorem (Thm.3.1) Z[ζpk ] ∼
=
Z[x]/(Φpk (x)) contains a unit ε of infinite order. Let g(x) ∈ Z[x] be a representative of
ε, i.e., ε = g(ζpk ); clearly, since g(ζpk ) is a unit, it does not belong to the prime ideal
(p, ζpk − 1), hence g(ζpk ) ≡ g(1) 6≡ 0 (mod (p, ζpk − 1)) and therefore g(1) 6≡ 0 (mod p).
k
k−1
k−1
It follows that g(x)φ(p ) ≡ g(xp )p−1 ≡ g(1)p−1 ≡ 1 (mod (p, xp − 1)), and, since
k−1
φ(pk )
h(x) divides xp −1, the same congruence holds also modulo (p, h(x)). Hence (1̄, g(x)
Im(ψ) is a unit of infinite order, a contradiction.
)∈
Lemma 4.5. A does not contain a unit α of multiplicative order 12 such that α6 = −1.
Proof. Assume, on the contrary, that there exists an element α ∈ A∗ such that α6 = −1
and consider the substitution homomorphism ϕα : Z[x] → A of Lemma 4.2. Let ker(ϕα ) =
(d(x)); we have that d(x)|x6 + 1, hence d(x) = x2 + 1, x4 − x2 + 1, or x6 + 1. The first
two cases must be excluded; in fact, the first one would imply α4 = 1 and the second case
would imply that Z[x]/ ker(ϕα ) ∼
= Z[ζ12 ] which contains units of infinite order (see Thm.
3.1). So we are left to examine the case d(x) = x6 + 1. We consider again the injection
given by the Chinese Remainder Theorem
ψ : Z[x]/(x6 + 1) → Z[x]/(x2 + 1) × Z[x]/(x4 − x2 + 1).
Arguing as in the proof of Lemma 4.4, we have
Im(ψ) = {(a(x), b(x)) | a(x) − b(x) ∈ (x2 + 1, x4 − x2 + 1) = (3, x2 + 1)},
we show that Im(ψ) contains a unit of infinite order.
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
7
Clearly, Z[x]/(x4 − x2 + 1) ∼
= Z[ζ12 ] so it contains a unit g(x) of infinite order, which,
in particular, does not belong to any proper ideal of Z[x]/(x4 − x2 + 1). This gives that
any representative g(x) of the class g(x) does not belong to the proper ideal (3, x2 + 1)
of Z[x]. Since Z[x]/(3, x2 + 1) ∼
= F9 , we have that 1 − g(x)8 ∈ (3, x2 + 1), therefore,
8
(1̄, g(x) ) ∈ Im(ψ) and is a unit of infinite order, contradicting the finiteness of A∗ .
Remark 4.6. The condition α6 = −1 of the lemma can not be relaxed to α12 = 1; for
example the ring Z[ζ3 ]×Z[i] is a torsion free ring, has a finite group of units and α = (ζ3 , i)
has order 12.
Proof of Theorem 4.1. Lemma 4.4 ensures that the cyclic factors of A∗ can have only
order 2, 3 or 4. It follows that
A∗ ∼
= (Z/2Z)a × (Z/4Z)b × (Z/3Z)c
for some a, b, c ≥ 0. Moreover, Z∗ ⊆ A∗ so 2 divides |A∗ | and a + b ≥ 1. To show that, if
c ≥ 1, then a ≥ 1, we use Lemma 4.5. Assume, by contradiction, that a = 0; in this case
−1 is the square of an element of α ∈ A∗ , and since c ≥ 1 there exist γ ∈ A∗ of order 3.
It follows that (αγ)6 = −1: this gives a contradiction by Lemma 4.5.
To conclude the proof we have to show that all abelian groups of the form
(Z/2Z)a × (Z/4Z)b × (Z/3Z)c ,
with a + b ≥ 1 and a ≥ 1 if c ≥ 1, actually occur as groups of units of suitable rings of
characteristic 0, without nilpotents and torsion free. In fact, for c = 0 one can choose
A = Za × Z[i]b ; for c ≥ 1 the following proposition provides a family of rings {An }n such
that A = Za−1 × Z[i]b × Ac−1 has group of units of the required form.
Proposition 4.7. For n ≥ 0 let
An = Z[x, y1 , . . . , yn ]/(x2 + x + 1, {yi2 + yi + 1}i=1,...,n ).
Then An is a ring without nilpotents and torsion free and
n+1
.
A∗ ∼
= Z/2Z × (Z/3Z)
n
Proof. Let I = (x2 + x + 1, {yi2 + yi + 1}i=1,...,n ). For each S ⊆ {1, . . . , n} let
Y
yS =
yi ;
i∈S
then {yS + I, xyS + I}S⊆{1,...,n} is a Z-basis for An . It follows that the set P of polynomials of Z[x, y1 , . . . , yn ] of degree less than 2 in each of the variables is a complete and
irredundant set of representatives of the elements of An .
For 1 ≤ i ≤ n define fi = yi − x and gi = yi − x2 . Moreover, for each S ⊆ {1, . . . , n},
put
IS = I + ({fi }i6∈S , {gi }i∈S })
and
BS = Z[x, y1 , . . . , yn ]/IS .
8
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
Define also
B=
Y
BS ,
S
where the product is taken over all subsets S of {1, . . . , n}.
Lemma 4.8. For each S ⊆ {1, . . . , n},
BS ∼
= Z[ζ3 ]
n
n
n
whence, B ∼
= (Z[ζ3 ])2 and B ∗ ∼
= (Z/2Z)2 × (Z/3Z)2 .
Proof. In BS every yi is congruent to either x or to x2 according to i 6∈ S or i ∈ S. In any
case, BS ∼
= Z[ζ3 ] for all S.
= Z[x]/(x2 + x + 1) ∼
The statement about B is clear since B is the product of 2n factors all isomorphic to
Z[ζ3 ]. Finally, by Theorem 3.1, Z[ζ3 ]∗ ∼
= Z/6Z and the result on B ∗ follows.
Let πS : Z[x, y1 , . . . , yn ] → BS be the canonical projection and let
π : Z[x, y1 , . . . , yn ] → B.
defined by π(ξ) = (πS (ξ))S for all ξ ∈ Z[x, y1 , . . . , yn ].
Lemma 4.9. The map π induces an injective ring homomorphism
ψ : An → B.
Proof. The map π induces an injective homomorphism An → B if and only if ker(π) = I.
By the Chinese Remainder Theorem, this is the case if and only if I = ∩S IS .
Clearly, I ⊆ ∩S IS and we are left to verify the reverse inclusion.
Let f ∈ ∩S IS ; and let r ∈ P be its canonical representative in A, namely f + I = r + I.
We claim that r = 0, so f ∈ I.
Denote by Q̄ an algebraic closure of Q and let I˜ and I˜S be the extensions of I and IS to
˜ = ∪S V (I˜S )
the ring Q̄[x, y1 , . . . , yn ]. Considering the sets of zeroes in Q̄n+1 we have V (I)
and, since the ideals I˜ and I˜S are radical, this yields I˜ = ∩S I˜S .
˜
Now, ∩S IS ⊆ ∩S I˜S ∩ Z[x, y1 , . . . , yn ] = I˜ ∩ Z[x, y1 , . . . , yn ], so f ∈ I˜ and hence r ∈ I.
˜ xyS + I}
˜ S⊆{1,...,n} is a Q̄-basis for Q̄[x, y1 , . . . , yn ]/I,
˜ and
Considering that {yS + I,
X
˜
I˜ = r + I˜ =
(hS yS + gS xyS + I),
S
we get r = 0.
It follows that ∩S IS ⊆ I and the equality holds.
n
From Lemmas 4.8 and 4.9 it follows that An is isomorphic to a subring of (Z[ζ3 ])2 (in
particular this shows that An is torsion-free) and that A∗n is isomorphic to a subgroup of
n
n
(Z/2Z)2 × (Z/3Z)2 . To prove the Proposition we are left to show that the cardinality
of the 2-Sylow Gn,2 and of the 3-Sylow Gn,3 of A∗n are 2 and 3n+1, respectively.
Arguing as in the proof of Lemma 4.4 we can easily prove the following
(i) if (aS + IS )S ∈ Im(ψ), then aS − aT ∈ IS + IT for all S 6= T ;
(ii) IS + IT = (3, x − 1, y1 − 1, . . . , yn − 1) for all S 6= T .
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
9
We show that the group Gn,2 has 2 elements. In fact, ψ(Gn,2 ) is a subgroup of {(±1 +
IS )S }; on the other hand, if (aS + IS )S ∈ ψ(Gn,2 ), by (i), aS − aT ∈ IS + IT for all S, T :
since aS = ±1 and 2 6∈ IS + IT for all S 6= T , then an element of ψ(Gn,2 ) must have all
coordinates represented by 1 or all represented by −1, hence Gn,2 = {±1 + I}.
We shall now prove that |Gn,3| = 3n+1 . Firstly, Gn,3 contains the set U = {xε0 y1ε1 · · · ynεn +
I | εi ∈ {0, 1, 2}} and so |Gn,3 | ≥ 3n+1 , since the elements of U are all distinct. To prove the
last assertion it is enough to note that the canonical representatives in P of the elements
of U are all distinct, since they have different factorization in Z[x, y1 , . . . , yn ].
We have to show the converse inequality. Clearly, ψ(Gn,3 ) is a subgroup of the 3-torsion
subgroup of B ∗ , namely, {(xεS + IS )S | εS = 0, 1, 2}.
Consider the projection
Y
BS /3BS ,
pr : B → B/3B =
S
let (b̄S )S = pr((bS + IS )S ) and ψ̄ = pr ◦ ψ. When restricting pr to the 3-torsion elements of
B ∗ we obtain an injective map (for each S the elements 1, x, x2 are pairwise not congruent
modulo 3BS ), whence |Gn,3| = |ψ̄(Gn,3 )|.
Q
n
Since B/3B ∼
= (Z[ζ3 ]/3Z[ζ3 ])2 and dimF3 Z[ζ3 ]/3Z[ζ3 ] = 2, then B/3B = S BS /3BS
is a vector space over F3 of dimension 2n+1 . For each S consider the base of BS /3BS
given by the classes of 1 and of x − 1. We shall use the notation b̄ = (b̄S )S ∈ B/3B and
b̄S = b0,S + b1,S (x − 1) where bi,S ∈ F3 . Note that the projection of the 3-torsion elements
of B ∗ is a subset of B/3B contained in the affine subspace
W = {(b̄S )S | b0,S = 1̄}.
Put V = ψ̄(An ); then V is a vector subspace of B/3B, so ψ̄(Gn,3 ) ⊆ V ∩ W . We shall
derive an upper bound for |ψ̄(Gn,3)| by bounding the dimension of the affine subspace
V ∩ W.
We now determine a set of linearly independent relations fulfilled by the elements of V .
By (i) and (ii), if v = (v0,S + v1,S (x − 1))S ∈ V then
v0,S = v0,T
∀S 6= T.
(5)
Considering S = ∅ and T 6= ∅ , we get 2n − 1 independent linear relations among
the elements of V , whereas the other relations depend on these. This gives dim V ≤
2n+1 − 2n + 1 = 2n + 1.
Further, we show that the elements of V verify the following set of relations:
X
(−1)|T | vT = 0
∀v = (vT )T ∈ V
(6)
T ⊆U
for all U ⊆ {1, . . . , n} with |U| ≥ 2.
Note that {ψ̄(yS ), ψ̄(xyS )}S is a set of generators of V , hence it is enough to show that
relations (6) are verified by these elements. Clearly ψ̄T (yS ) = x̄|S|+|S∩T |. If U 6⊆ S, then,
for each r ≥ 0, among the subsets T of U with |S ∩ T | = r one half has even cardinality
and one half has odd cardinality, so the sum clearly vanishes. If U ⊆ S, and |U| = m
10
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
then the relation becomes
m
X
X
i m
|T | |S|+|T |
x̄|S|+i = x̄|S| (1̄ − x̄)m = 0,
(−1)
(−1) x̄
=
i
i=0
T ⊆U
since (1 − x)m ∈ 3B for m ≥ 2. An analogous computation proves that (6) holds also for
v = ψ̄(xyS ).
Each of the equations (6) splits into the two equations
X
(−1)|T | vi,T = 0
∀v = (vT )T ∈ V, i = 0, 1.
(7)
T ⊆U
The relations with i = 0 follow directly from equations (5).
The relations with i = 1 are clearly independent from the previous ones and also among
themselves (for example because one can order the subsets U so that the matrix of the
relations becomes triangular). There is one relation for each subset U of {1, . . . , n} of
cardinality at least 2, and hence we get 2n − n − 1 relations. This gives
dim V ≤ 2n + 1 − 2n + n + 1 = n + 2.
Since V 6⊆ W , the dimension of V ∩ W is at most n + 1 and therefore |ψ̄(Gn,3 )| ≤ 3n+1 ,
as desired.
5. General characteristic zero rings
In this section we consider the general case of a characteristic zero ring A. In Theorem
5.1 we prove that all groups of the form Z/2Z × H, where H is any abelian group, are
realizable as groups of units. Moreover, when A∗ does not have a cyclic factor of order 2
in its decomposition, then it must have a cyclic direct factor of order 4, and a more subtle
analysis is required. For this case we give some necessary conditions and some sufficient
conditions for the group to be realizable (see Proposition 5.8, Proposition 5.13 and the
summary in Section 5.3).
Definition 1. For a ring A we define ε(A) as the minimum exponent of 2 in the decomposition of the 2-Sylow of A∗ as direct sum of cyclic groups.
The following theorem describes almost precisely the group of units in the case of
characteristic zero rings. The second statement of the theorem can be found also in
[Fuc70, Thm 129.1]; we include the proof using our notation, since we will need a similar
construction in the proof of Proposition 5.8. However, the main result of next theorem is
the fact that, for every ring A of characteristic 0, necessarily ε(A) ≤ 2.
Theorem 5.1. The finite abelian groups which are the group of units of a ring A of
characteristic 0 have the form
Z/2ε Z × H
(8)
where ε = ε(A) = 1, 2 and H is an abelian group.
Conversely for each finite abelian group H there exists a ring A such that
A∗ ∼
= Z/2Z × H.
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
11
Lemma 5.2. Let α ∈ A∗ and consider the homomorphism ϕα : Z[x] → A defined by
p(x) 7→ p(α). If ker(ϕα ) contains an irreducible polynomial µ(x), then ker(ϕα ) = (µ(x))
and Z[α] is a domain.
Proof. If µ(x) ∈ ker(ϕα ), then ker(ϕα ) = (µ(x), f1 (x), . . . , fk (x)) for some fi ∈ Z[x].
Now, by Gauß Lemma, µ(x) is irreducible in Q[x]; moreover, 1 6∈ ker(ϕα )Q[x] (in fact,
ker(ϕα ) ∩ Z = {0} since char(A) = 0), therefore ker(ϕα )Q[x] = (µ(x)) and µ(x) divides
fi (x) in Q[x], for each i. Finally, since µ(x) is primitive, we can conclude that µ(x) divides
fi (x) also in Z[x], hence ker(ϕα ) = (µ(x)) and Z[α] ∼
= Z[x]/(µ(x)).
Remark 5.3. We note that the homomorphism ϕα induces an injection of Z[α] into A
preserving the identity. This property will be crucial in the following since it implies that
(Z[α])∗ is a subgroup of A∗ .
Proof of Theorem 5.1. Since char(A) = 0, then Z ⊆ A and 2 = |Z∗ | divides |A∗ |, so
ε(A) ≥ 1; we have to show that ε ≤ 2. Assume the contrary; then −1 = α4 for some
α ∈ A∗ . Let again ϕα : Z[x] → A be defined by p(x) 7→ p(α). Then the irreducible
polynomial µ(x) = x4 + 1 belongs to ker(ϕα ). By Lemma 5.2, we get ker(ϕα ) = (µ(x)). It
follows that A contains a ring isomorphic to Z[x]/(x4 + 1) ∼
= Z[ζ8 ] and, by Remark 5.3, A∗
∗
contains Z[ζ8 ] as a subgroup: this is not possible, since A∗ is finite and Z[ζ8 ]∗ is infinite
by Dirichlet Theorem.
As to the converse, let H ∼
= Z/a1 Z × · · · Z/an Z where a1 , . . . , an are positive integers.
Define
Z[x1 , . . . , xn ]
A=
.
(ai xi , xi xj )1≤i,j≤n
Let N be the nilradical of A; then N = (x̄1 , . . . , x̄n ) where, as usual, x̄j denotes the class
of xj in A. Clearly, A = Z[N]. The exact sequence given in (2) in this case specifies to
1 → 1 + N → A∗ → {±1} → 1,
so the sequence splits and A∗ ∼
= Z/2Z×(1 + N). Moreover, in this case the map n → 1 + n
is an isomorphism from the additive group N to the multiplicative group (1 + N), hence
A∗ ∼
= Z/2Z × Z/a1 Z × · · · Z/an Z.
= Z/2Z × N ∼
From the previous proposition we immediately get the following
Corollary 5.4. The possible values of |A∗ |, when A is a characteristic 0 ring with finite
group of units, are all the even positive integers.
The last corollary and [DCD17, Cor. 3.4] allows to completely answer Ditor’s question
for rings of any characteristic.
Corollary 5.5. The possible values of |A∗ |, when A is a ring with finite group of units,
are all the even positive integers and the finite products of integers of the form 2λ − 1
with λ ≥ 1.
As a particular case we reobtain the following result of Ditor [Dit71]
Corollary 5.6. If |A∗ | = p is prime, then p = 2 or p is a Mersenne prime.
12
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
5.1. The case ε(A) = 2. In Theorem 5.1 we classify the groups of units of the rings A
with ε(A) = 1, showing that any abelian group H can appear in equation (8) in this
case. This is no longer true for rings with ε = 2: for example, in this case we can not
have H ∼
= Z/11Z, since the cyclic group Z/44Z is not realizable (see [PS70]). Many other
examples can be derived from Proposition 5.13 and Corollary 5.14.
In the following remark we point out an important necessary condition on A∗ for rings
with ε = 2 and in the next proposition we exhibit a large class of groups H which occur
in this case.
Remark 5.7. Let ε(A) = 2, then the ring A must contain Z[i] as a subring. In fact, in this
case −1 = α2 for some α ∈ A and so, by Lemma 5.2, we have Z[i] ∼
= Z[x]/(x2 + 1) ֒→ A.
We will see that this property is crucial for showing that not all abelian groups H can
occur in (8) in this case (see Proposition 5.13).
For any finite group H and for any prime number p we denote by Hp the p-Sylow
subgroup of H.
Proposition 5.8. Let H be a finite abelian group with the following properties:
a) ∀p ≡ 3 (mod 4) the p-Sylow Hp is the square of a group;
b) H2 is isomorphic to P or to Z/4Z × P where
P ∼
= Z/2e1 Z × · · · × Z/2e2r Z
with 2 ≤e1 ≤ · · · ≤ e2r and e2j − e2j−1 ≤ 1 ∀j = 1, . . . , r.
Then Z/4Z × H is the group of units of a characteristic 0 ring.
Proof. To prove this proposition we need to refine the proof of Theorem 5.1.
Let a1 , . . . , an be any finite sequence of non-zero elements of the ring Z[i]. Similarly to
above, define
Z[i][x1 , . . . , xn ]
A=
.
(ai xi , xi xj )1≤i,j≤n
By the same argument we get
A∗ ∼
= Z/4Z × Z[i]/(a1 ) × · · · Z[i]/(an ).
= Z/4Z × N ∼
= Z[i]∗ × (1 + N) ∼
To understand which groups can be obtained when the aj ’s vary, it is enough to describe
the quotients Z[i]/(π)h when π is a prime of Z[i]. Let (p) = (π) ∩ Z and let N : Z[i] → Z
denote the usual norm. It is well known that N(π) = p2 or p according to p ≡ 3 (mod 4)
or not; from this, and from (2) = (1 + i)2, we easily get the following group isomorphisms:
Z/ph Z
if p ≡ 1 (mod 4)
h
h
Z[i] ∼ Z/p Z × Z/p Z
if p ≡ 3 (mod 4)
(9)
=
k
k
h
(π)
Z/2 Z × Z/2 Z
if p = 2 and h = 2k
Z/2k+1 Z × Z/2k Z if p = 2 and h = 2k + 1.
Now, it is clear that, by suitable choices of the aj ’s, one can obtain any p-group if p ≡
1 (mod 4), and any square of a p-group if p ≡ 3 (mod 4). Moreover, the groups P
described in (b) are precisely the groups that can be obtained as a finite product of
groups Z[i]/(1 + i)h with h ≥ 4. Finally, if A is a ring such that (A∗ )2 is isomorphic to
P , then, putting B = A × Z[i], we have (B ∗ )2 ∼
= P × Z/4Z.
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
13
To conclude, we note that in our construction H has no direct summand isomorphic to
Z/2Z, hence the rings we constructed have ε = 2.
To further investigate the structure of A∗ in the case when ε(A) = 2, it is convenient
to use the splitting of a ring proved by Pearson and Schneider, which we recall below.
We recall that, by Remark 2.1, we can assume that our ring fulfils the hypothesis of the
proposition.
Proposition 5.9. ([PS70, Prop. 1]) Let A be a commutative ring which is finitely generated and integral over its fundamental subring. Then A = A1 ⊕ A2 , where A1 is a finite
ring and the torsion ideal of A2 is contained in its nilradical.
In the paper [DCD17] we studied the groups of units of finite rings, so in the following
we shall concentrate on the study of the second factor of the decomposition.
Definition 2. A commutative ring is called of type 2 if its torsion ideal is contained in
the nilradical.
Remark 5.10. (i) A type 2 ring has characteristic 0, since otherwise 1 would be a torsion
element and hence nilpotent.
(ii) Let A be a type 2 ring such that A∗ is finite. Then, by Lemma 2.2, the torsion
elements of A are precisely its nilpotent elements. In particular, A is reduced if and only
if it is torsion free.
5.2. Type 2 rings with ε(A) = 2. For any ring of type 2 consider the exact sequence
(2) of finite groups
1 → 1 + N → A∗ → (A/N)∗ → 1
and for each prime p the exact sequence induced on the p-Sylow
1 → (1 + N)p → (A∗ )p → (A/N)∗p → 1
Since A is of type 2, the ring B = A/N is torsion-free and B ∗ is described by Theorem
4.1, so its p-Sylow (B ∗ )p is trivial for p > 3. The following proposition shows that also
(B ∗ )3 is trivial in this case.
Proposition 5.11. Let A be a type 2 ring with ε(A) = 2, and let B = A/N. Then the
3-Sylow of B ∗ is trivial.
Proof. Assume the contrary and let β ∈ B ∗ an element of order 3. By Lemma 5.2, the
ring A contains the domain Z[α] as a subring and they have the same identity. Let
ψ : Z[α] → B be the homomorphism obtained by composing the inclusion Z[α] ֒→ A with
the projection A → A/N, namely, ψ(p(α)) = p(α) + N. Clearly, ker(ψ) = N ∩ Z[α]
is trivial since Z[α] is a domain and ψ(1) = 1 + N = 1B . Let iB = ψ(i); then iB has
order 4, i2B = −1B and it is immediate to check that iB β is a unit of order 12 such that
(iB β)6 = −1B , contradicting Lemma 4.5.
Now, note that (1 + N)p = 1 + Np ; in fact, these two groups have the same cardinality,
so it is enough to prove that 1 + Np ⊆ (1 + N)p : this can be checked by noting that,
if x ∈ Np and k, l ≥ 0 are such that pk x = 0 and xl = 0, then for h ≥ k + l it results
14
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
Pph ph j
h
(1 + x)p =
j=0 j x = 1, and hence 1 + x ∈ (1 + N)p . It follows that the exact
sequence on the p-Sylow reads as
and, in particular, we get
1 → 1 + Np → (A∗ )p → (B ∗ )p → 1.
(10)
(A∗ )p = 1 + Np for p ≥ 3.
(11)
Proposition 5.12.
We note that the rings constructed in the proof of Proposition 5.8, are in fact type 2
rings, so the result proved there holds also if we restrict to type 2 rings. In particular, A∗p
can be any abelian p-group when p ≡ 1 (mod 4), and we are left to analyze the p-Sylow
of A∗ for p = 2 and p ≡ 3 (mod 4). In this last case, the following proposition gives a
constraint on the structure and on the cardinality of 1 + Np .
Proposition 5.13. Let A be a ring with ε(A) = 2 and let p ≡ 3 (mod 4). For each
j ≥ 1, the quotient (1 + Njp )/(1 + Nj+1
p ) has a filtration such that all its quotients are
Fp2 -vector spaces.
In particular |1 + Np | is a square.
Proof. For all j ≥ 1 the map x 7→ 1 + x induces an isomorphism between Xj = Njp /Nj+1
p
and (1 + Njp )/(1 + Nj+1
).
p
Now, Xj is a Z[i]-module and an abelian finite p-group, so for a suitable integer rj we
have the filtration
Xj ⊃ pXj ⊃ · · · ⊃ prj Xj = {0},
It follows that
whose quotients pl Xj /pl+1 Xj are Z[i]/(p) ∼
= Fp2 -vector spaces for all j, l. P
there exist integers kj such that |Xj | = p2kj , so |1 + Np | = p2k where k = kj .
Corollary 5.14. Let ε(A) = 2 and p ≡ 3 (mod 4), then 1 + Np is not a cyclic group.
Proof. All quotients of a cyclic group are cyclic, hence they can not admit a filtration
whose quotients are Fp2 -vector spaces.
The previous proposition can not be generalized to a result on the structure of the
group; in fact, the following example shows that for p ≡ 3 (mod 4) the group 1 + Np is
not necessarily the square of a group.
Example 1. Let p ≡ 3 (mod 4) and A = Z[i][x]/(p2 x, px2 , xp + px). Clearly, N = Np =
(x̄), where x̄ denotes the class of x in A, and as a Z[i]-module
p−2
p−1
Np ∼
= Z[i]/(p2 ) × (Z[i]/(p)) .
= ⊕ hx̄j i ∼
j=1
2p
Therefore, |1 + Np | = |Np | = p .
Pp−1
j
Let n ∈ Np , then n can be written as n =
j=1 λj x̄ ∈ Np , where λj ∈ Z[i] and
are uniquely determined modulo p2 if j = 1 and modulo p if 2 ≤ j ≤ p − 1. Then
(1+n)p = 1+px(λ1 −λp1 ) = 1 if and only if the reduction of λ1 modulo p belongs to Fp ; this
means that there are p3 possibilities for λ1 and p2 possibilities for λj for j ≥ 2. It follows
that 1+Np contains exactly p2p−1 elements of exponent p, so 1+Np ∼
= Z/p2 Z×(Z/pZ)2p−2
and it is not a square of a group.
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
15
As to the case p = 2, the following example and Remark 5.15 show that the exact
sequence (10) and Theorem 4.1 are not sufficient to describe the 2-component of A∗ .
Example 2. Let A = Z[i][x, y]/(x2 − y − 1, (1 + i)y, y 3). Clearly A is a type 2 ring,
N = N2 = (ȳ) and as a Z[i]-module
N∼
= (Z[i]/(1 + i))2 .
= hȳi ⊕ hȳ 2 i ∼
Therefore, |1 + N| = |N| = 22 . By direct computation, one sees that 1 + ȳ has order 4,
hence the group 1 + N is isomorphic to Z/4Z.
Consider now the quotient ring B = A/N; we have B = Z[i, x]/(x2 − 1) and
B ∗ = B ∗ = {±1, ±i, ±x̄, ±ix̄} ∼
= Z/4Z × Z/2Z,
2
in particular ε(B) = 1. The exact sequence (2), that in this case coincides with (10),
becomes
1 → Z/4Z → A∗ → Z/4Z × Z/2Z → 1
and this gives |A∗ | = 25 . Finally, since x̄ is a unit of order 8, one gets A∗ ∼
= hx̄i × hii ∼
=
Z/8Z × Z/4Z.
.
Remark 5.15. Example 2 shows that
• the list of 2-groups given in Proposition 5.8 is not exhaustive, since it does not
contain the group A∗ ∼
= Z/8Z × Z/4Z;
• it may happen that ε(A) = 2 while ε(A/N) = 1;
• the exact sequence (10) does not split in general for p = 2.
5.3. Concluding remarks. To conclude, while there is a complete classification of the
possible finite abelian groups A∗ when ε(A) = 1, in the case when ε(A) = 2 we have found
some necessary and some sufficient conditions, which are indeed very strict.
We summarize the results for the case when A is a ring of type 2 with ε(A) = 2 as
follows:
• If p ≡ 1 (mod 4) then (A∗ )p = 1 + Np and can be any abelian p-group.
• If p ≡ 3 (mod 4) then (A∗ )p = 1 + Np but it can not be any p-group in in view
of the condition on the filtration given in Proposition 5.13. In particular, the
cardinality of 1 + Np must be a square and all squares of a p-group are realizable
(see Proposition 5.8).
• If p = 2 then we have an exact sequence
1 → 1 + N2 → (A∗ )2 → (Z/2Z)a × (Z/4Z)b → 1
where a+b ≥ 1. Moreover, Proposition 5.8 gives a list (not exhaustive) of realizable
2-Sylow subgroups of A∗ .
Since our description is not complete, we have tried to guess what could be a complete
classification. For the case p = 2, what we have just said shows that the situation is far
from clear, so we do not know what to expect. The case p ≡ 3 (mod 4) appears simpler;
since we have not found any example of a p-group with a filtration as in Proposition 5.13
which is not realizable as (A∗ )p , one might conjecture that the condition on the filtration
could be sufficient for realizability.
16
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
By virtue of the decomposition A = A1 ⊕ A2 given in Proposition 5.9, the group of
units of a general ring A is the product of the groups of units of a finite ring and of a
type 2 ring. Therefore, the results of this section together with those of [DCD17] allow to
give a fairly precise description of A∗ for a general ring A: in fact, we are able to produce
families of new realizable abelian groups which can not be written as a product of cyclic
realizable factors (already known as a consequence of [PS70]); on the other hand, our
constraints allow to show that many families of abelian groups can not be realized. We
give below one instance of each type.
Example 3. The group Z/4Z × (Z/(11Z))2 is realizable (see Proposition 5.8), although
Z/4Z × Z/11Z and Z/11Z are not.
Example 4. The group G = Z/4Z × Z/4Z × Z/11Z is not realizable. In fact, assume by
contradiction that G is the group of units of a ring A = A1 ⊕ A2 as in the decomposition
by Pearson and Schneider. Then its cyclic factor Z/11Z must come either from A1 or
from A2 . It cannot come from A1 : in fact, if 11 | |A∗1 | then either 10 | |A∗1 | or A∗1 is cyclic
of order pλ − 1 for suitable p and λ (see [DCD17, Theorem 3.1]) and both these cases are
impossible since A∗1 is a subgroup of G. It cannot come from A2 either, because in this
case we would have ε(A2 ) = 2, and, by Proposition 5.13 the 11-Sylow subgroup of A∗2
must have a cardinality which is a square. So G is not realizable.
Proposition 5.9, [DCD17, Cor. 3.2] and Theorem 4.1 immediately give the following
corollary which classifies the groups of units of reduced rings (see also Remark 5.10 (ii)).
Corollary 5.16. The finite abelian groups which are groups of units of a reduced ring
are the finite products of multiplicative groups of finite fields and possibly a group of the
form (Z/2Z)a × (Z/4Z)b × (Z/3Z)c where a, b, c ∈ N, a + b ≥ 1 and a ≥ 1 if c ≥ 1.
Finally, we note that combining the previous results with [DCD17, Cor. 4.2] we easily
reobtain, as a particular case of our results, the complete classification of the finite cyclic
groups which occur as group of units of a ring (firstly given in [PS70]).
Corollary 5.17. A finite cyclic group is the group of units of ring if and only if its order
is the product of a set of pairwise coprime integers of the following list:
a) pλ − 1 where p is a prime and λ ≥ 1;
b) (p − 1)pk where p > 2 is a prime and k ≥ 1;
c) 2d where d > 0 is odd;
d) 4d where d is an odd integer and each of its prime factors is congruent to 1 mod
4.
Proof. By [DCD17, Cor. 4.2], the cyclic groups of units which are realizable with finite
characteristic ring are those whose order is the product of pairwise coprime integers as in
(a), (b) and, possibly, 2 or 4. Consider now rings of characteristic 0: by Theorem 5.1 all
cardinalites listed in (c) are possible; by Proposition 5.8 and Corollary 5.14 the cardinality
4d is possible if and only if d is an odd integer and each of its prime factors is congruent
to 1 mod 4.
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
17
Appendix A. Densities
In this section we study the distribution of the possible values of the cardinality of A∗
where A∗ is a finite abelian group. The first simple remark regards the prime numbers p
that occur as the cardinality of A∗ : by Corollary 5.6, we know that the only possibilities
are p = 2 or p a Mersenne prime.
The next observation is that odd numbers can be very seldom the cardinality of A∗ for
some ring A. For a set X ⊂ N and for n ∈ N we denote, as usual, X(n) := X ∩ {1, . . . , n}
and the density of X as the limit
|X(n)|
lim
n→∞
n
whenever this limit exists. We have:
Proposition A.1. The set of all possible odd values of |A∗ | has density zero.
Proof. Let
{ |A∗ | : |A∗ | is odd }. By [DCD17, Corollary 3.4] and Theorem 5.1,
Qm X k=
X = { i=1 (2 i − 1) | m ∈ N, ki ≥ 2}. For any positive number M we estimate the
cardinality of the set X(2M ) = {n ∈ X | n ≤ 2M }. Clearly,
2
Pm
i=1 (ki −1)
=
m
Y
2
ki −1
i=1
Qm
≤
Pm
m
Y
i=1
(2ki − 1),
and if n = i=1 (2ki − 1) ∈ X(2M ), then i=1 (ki − 1) ≤ M. It follows that |X(2M )| is at
P
most M
h=1 p(h), where p(h) denotes the number of partitions of the positive integer h.
By [And84, Thm 6.3]
12 !
1
1
2
p(h) ∼ √ exp π
h2 ,
3
4h 3
so
!
21
M
1
1
1 X
2
M2
p(h) ≪ M log M exp π
2M h=1
2
3
and the last expression goes to zero as M → ∞.
Theorem 5.1 says that any even number can be the cardinality of A∗ for some ring A.
Together with Proposition A.1 this gives
Proposition A.2. The density of the possible cardinalities of A∗ is equal to 21 .
As we have seen before, the possible groups A∗ are much less if we restrict to the case
where the ring A has no non-zero nilpotents. In fact, we have the following
Proposition A.3. The set of possible cardinalities if A∗ , where A ia a reduced ring, has
density zero.
Proof. We need the following
Lemma A.4. For each h ≥ 0, the set Xh of possible cardinalities of A∗ such that
2h ||card(A∗ ), where A is a reduced ring, has density zero.
18
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
Proof. For h = 0, the statement has already been proved in Proposition A.1. For h > 0,
observe that, by Corollary 5.16, the elements of m ∈ Xh are obtained as products of type
m=2
a+2b
s
t
Y
Y
λ
ki
(2 − 1) (pj j − 1) ,
i=1
(12)
j=1
Q
λ
where a+2b ≤ h, p1 , . . . , ps are odd primes and 2h−a−2b || tj=1 (pj j −1) (the possible factor
Q
3c given by Theorem 4.1 can be absorbed in the product si=1 (2ki − 1), since 22 − 1 = 3).
Q
We write a number m ∈ Xh (n) as m = 2a+2b m1 m2 , where m1 = si=1 (2ki − 1) and
Q
λ
m2 = tj=1 (pj j − 1) and we partition Xh (n) according to the value of m1 ≤ n. We have
P
|Xh (n)| ≤ m1 ≤n |{2a+2b m2 ≤ n/m1 }|. To estimate the cardinality of each summand we
first note that m2 ≤ mn1 and, since all numbers pλj − 1 are even, we have t ≤ h. Moreover,
Q
we have trivially pλ ≤ 23 (pλ − 1), so tj=1 pλj ≤ ( 23 )t mn1 ≤ ( 32 )h mn1 .
It follows that, for any m1 ≤ n and any given pair (a, b) with a + 2b ≤ h, theQ
number of
possible m2 for which m ∈ Xh (n) does not exceed the cardinality of the set { tj=1 pλj ≤
( 32 )h mn1 }. For any positive real number x and any positive integer t, let
ωt (x) = {pλ1 1 . . . pλt t ≤ x | λj ≥ 0} .
Clearly ωt (x) ≤ ωh (x) for t ≤ h. By [Ram17, Lemma B], we have
x(log log x)h−1
.
ωh (x) = O
log x
Putting x = ( 23 )h mn1 and summing over the finitely many possibilities for the pairs (a, b),
we obtain that
!
h
n
3
|{2a+2b m2 ≤ n/m1 }| ≤ c(h) ωh
2
m1
for some positive constant c(h) depending only on h.
Next, we split the numbers in Xh (n) into two subsets, according to the value of m1 .
We have
3 h
3 h
X
X
(2) n
(2) n
ωh
ωh
|Xh (n)| ≤ c(h)
+ c(h)
.
(13)
m1
m1
√
√
m1 ≤
( 32 )h n
m1 >
( 32 )h n
( 32 )h n
m1
For a generic term of the first sum in (13) we have log
≫ log n, so, for
n sufficiently large, the contribution of each of these terms is less than or equal to
log n)h−1
for some positive constant c1 (h). Now,
c1 (h) m11 n (loglog
n
X
X 1
1
≤
:= S
k
1
m1 k ,...,k (2 − 1) · · · (2ks − 1)
m ≤n
1
1
s
and the series S converges, since
∞
∞
Y
1
2k − 1 Y
1+ k
=
S≤
k −2
2
2 −2
k=1
k=1
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
and the series of logarithms
P∞
k=1
log(1 +
1
)
2k −2
∼
P∞
1
k=1 2k −2
19
converges. It follows that
h−1
log n)
there is a positive constant c2 such that the first sum is less than or equal to c2 n(loglog
n
and therefore is o(n).
1
As to the second sum, each term in it is clearly ≪ n 2 , so we are reduced to estimate
the number of terms is this sum.
We now write the number m1 as (22 − 1)µ1 (23 − 1)µ2 · · · (2r+1 − 1)µr ; since m1 ≤
n, then a fortiori, 2µ1 22µ2 · · · 2rµr ≤ n, so the non-negative integers µi are such that
n
µ1 + 2µ2 + · · · + rµr ≤ log
. It is well-known that the number of integer points (µ1 , . . . , µr )
log 2
satisfying the preceding inequality is approximately equal to the volume of the region
log n
{(u1 , . . . , ur ) ∈ Rr | ui ≥ 0, u1 + 2u2 + · · · + rur ≤
}.
log 2
More precisely, by [Lan94, Chapter 6, Thm 2], we get that the number of r-tuples
(µ1 , . . . , µr ) such that 2µ1 22µ2 · · · 2rµr ≤ n is less than
(log n)r−1
1 (log n)r
+O r
) .
(r!)2 (log 2)r
(r − 1)!2
By the trivial inequality (2r)! ≥ (r!)2 we get
(log n)r
(log n)r
≤
(r!)2
(2r)!
1
=
((log n) 2 )2r
(2r)!
1
≤ e(log n) 2
1
≪ (log n)e(log n) 2 = o(nε )
for any ε > 0. Taking into account that r ≪ log n, the error term is estimated similarly,
so the second sum is also o(n).
Putting together the last two estimates we get the lemma.
Denote by Xh′ the set of all possible cardinalities of A∗ divisible by 2h , where A is
a reduced ring. For any h ≥ 0, we can obviously partition the set X of all possible
cardinalities of A∗ as
X = X0 ∪ X1 ∪ · · · ∪ Xh−1 ∪ Xh′ .
By the lemma, the densities δ(Xi ) of Xi are zero for i = 0, . . . , h = 1, so the density of
1
X equals the density of Xh′ . But trivially δ(Xh′ ) ≤ 2h+1
, so, taking the limit for h → ∞,
we get δ(X) = 0.
References
[And84] G. E. Andrews, The theory of partitions, Cambridge University Press, 1984.
[CL15] Sunil K. Chebolu and Keir Lockridge, Fuchs’ problem for indecomposable abelian groups, J.
Algebra 438 (2015), 325–336.
[CL17]
, Fuchs’ problem for dihedral groups, JPAA 221 (2017), 971–982.
[Cor63] A. L. S. Corner, Every countable reduced torsion-free ring is an endomorphism ring, Proc. Lond.
Math. Soc. 13 (1963), no. 3, 687–710.
20
ON FUCHS’ PROBLEM ABOUT THE GROUP OF UNITS OF A RING
[DCD17] I. Del Corso and R. Dvornicich, Finite groups of units of finite characteristic rings, to appear
in Annali di Matematica https://doi.org/10.1007/s10231-017-0697-5 (2017).
[Dit71] S. Z. Ditor, On the group of units of a ring, The American Mathematical Monthly 78 (1971),
no. 5, 522–523.
[Dol02] D. Dolžan, Groups of units in a finite ring, J. Pure Appl. Algebra 170 (2002), no. 2-3, 174–183.
[EF67] K. E. Eldridge and I. Fischer, Dcc rings with a cyclic group of units, Duke Math. J. 34 (1967),
243–248.
[Fuc60] L. Fuchs, Abelian groups, 3rd ed., Pergamon, Oxford, 1960.
[Fuc70]
, Infinite abelian groups, Pure and applied mathematics, no. v. 2, Academic Press, 1970.
[Gil63] R. W. Gilmer, Finite rings with a cyclic group of units, Amer. J. Math. 85 (1963), 447–452.
[HH65] J. T. Hallett and K. A. Hirsch, Torsion-free groups having finite automorphism groups, J. of
Algebra 2 (1965), 287–298.
[HZ66] K. A. Hirsch and H. Zassenhaus, Finite automorphism groups of torsion free groups, J. London
Math. Soc. 41 (1966), 545–549.
[Lan94] Serge Lang, Algebraic number theory, 2nd ed., Graduate Texts in Mathematics, vol. 110,
Springer-Verlag, New York, 1994.
[Neu99] Jürgen Neukirch, Algebraic number theory, Grundlehren der mathematischen Wissenschaften,
vol. 322, Springer-Verlag, 1999.
[PS70] K. R. Pearson and J. E. Schneider, Rings with a cyclic group of units, J. of Algebra 16 (1970),
243–251.
[Ram17] S. Ramanujan, The normal number of prime factors of a number n, Quartely Journal in Mathematics XLVIII (1917), 76–92.
| 0math.AC
|
arXiv:1711.05734v2 [cs.DC] 20 Feb 2018
C HIPMUNK: A Systolically Scalable 0.9 mm2,
3.08 Gop/s/mW @ 1.2 mW Accelerator for
Near-Sensor Recurrent Neural Network Inference
Francesco Conti∗†
[email protected]
Lukas Cavigelli∗
[email protected]
Igor Susmelj∗
[email protected]
Gianna Paulin∗
[email protected]
Luca Benini∗†
[email protected]
Abstract
Recurrent neural networks (RNNs) are state-of-the-art in voice awareness/understanding and speech recognition. On-device computation of RNNs
on low-power mobile and wearable devices would be key to applications such
as zero-latency voice-based human-machine interfaces. Here we present C HIP MUNK , a small (<1 mm2 ) hardware accelerator for Long-Short Term Memory
RNNs in UMC 65 nm technology capable to operate at a measured peak efficiency
up to 3.08 Gop/s/mW at 1.24 mW peak power. To implement big RNN models
without incurring in huge memory transfer overhead, multiple C HIPMUNK engines can cooperate to form a single systolic array. In this way, the C HIPMUNK
architecture in a 75 tiles configuration can achieve real-time phoneme extraction
on a demanding RNN topology proposed in [1], consuming less than 13 mW of
average power.
1
Introduction
In the last few years, we have witnessed an “artificial intelligence” revolution that has been fueled
by the concurrent availability of huge amounts of training data, computing power to learn upon it,
and evolution of “smart” algorithms, in particular those based on deep learning. Within this field,
recurrent neural networks (RNNs), particularly Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), are receiving increasing attention: They have shown state-of-the-art accuracy
in tasks such as speech recognition [1, 2] and language translation [3], making them the forefront of
the “intelligent” user interfaces of products such as Amazon Alexa, Google Assistant, Apple Siri,
Microsoft Cortana and others.
One of the key limitations of the current generation of commercial products based on RNNs is that
these embedded, edge devices depend on remote servers taking care of the computational workload
necessary for the deployment of these algorithms. Moreover, when RNNs are used as a component
of human-machine interfaces, the intrinsic latency of network communication can also be problematic, as people expect the “smart” devices to reply not only accurately, but also timely.
For these reasons, it is very attractive to integrate RNN capabilities locally in embedded mobile
and wearable platforms, making them capable of state-of-the-art voice and speech recognition autonomously and independent from external servers. Nonetheless, while much attention has recently
been dedicated to the deployment of embedded low-power inference accelerators for forward-only
deep networks deployment [4–8], making RNNs energy-efficient is a fundamentally harder problem:
the necessity to keep and update an internal state and the widespread usage of densely connected
layers translate to very large memory footprint and high bandwidth requirements.
1
2
Integrated Systems Laboratory, ETH Zurich.
Energy-Efficient Embedded Systems Laboratory, University of Bologna.
In this work, we present a twofold contribution towards the deployment of RNN-based algorithms in
devices such as smartphones, smartwatches and wearables. First, we designed C HIPMUNK, a small
and low-energy hardware accelerator engine targeted at real-time speech recognition and capable to
operate autonomously on moderate size LSTM networks. We present silicon results from a prototype
chip containing a C HIPMUNK engine, which has been fabricated in UMC 65 nm technology; the
chip can achieve up to 3.8 Gop/s at maximum efficiency operating point (@0.75 V), consuming
only 1.24 mW.
Second, we conceived a scalable computing architecture, apt to operate on bigger LSTM models
as well. As the main limitation to the deployment of big RNNs in embedded scenarios stems from
their memory boundedness, we designed the C HIPMUNK engines so that they can be replicated
in a systolic array, cooperating on a single bigger LSTM network. This methodology allows the
acceleration of large-scale RNNs, which can be made fast enough to operate in real-time under
realistically tight time, memory and battery constraints without requiring complex, power hungry
and expensive high-bandwidth main memory interfaces.
2
Related Work
A recent thorough survey of efforts on hardware acceleration and design of efficient shows that few
efforts have been focused on RNN inference [9]. We thus focus on this application, surveying stateof-the-art implementations from data-center to ultra-low power accelerators in the remainder of this
section.
Data center workloads for RNNs are often offloaded to GPUs or specialized semi-independent coprocessors such as Google’s Tensor Processing Unit (TPU) [10] consuming in the order of 50-300 W.
The TPU is a unified architecture to target DNNs with convolutional and densely connected layers
as well as LSTMs. However, TPUs suffer from low utilization when running RNNs. Yet 29% of the
workload running on Google’s TPUs is devoted to RNN inference [10], showing their relevance in
commercial applications.
In a lower power range (tens of Watts), several FPGA implementations can be found. The Efficient Speech-recognition Engine (ESE) [11] targets the deployment of RNNs on a Xilinx UltraScale
FPGA. To maximize efficiency and address the memory boundedness of RNNs, it heavily focuses
on network quantization and pruning of the recurrent topologies and thus this accelerator engine
is mainly targeted at sparse matrix-vector operations. Rybalkin et al. [12] also target bidirectional
LSTMs in their FPGA accelerator. Bidirectional LSTMs have been shown to obtain better accuracy
in some cases [1] but are less attractive for an online, real-time scenario as they inherently increase
the network latency. Finally, DeepStream [13] is a small hardware accelerator deployed on a Xilinx
Zynq 7020 targeted at text recognition with RNNs. It requires to continuously stream in weights,
which makes it impractical for big RNN topologies with millions of weights.
The only published ultra-low power (few mW) implementation, the DNPU [14], uses two separate
special-purpose engines for convolutional layers (called CP), on one side, and fully-connected and
recurrent ones on the other (called FRP). The FRP does not include any particular facilities to address
the stateful nature of RNNs, and it includes only a small amount of memory (10 kB) making external
memory accesses necessary for even small RNNs, thus limiting peak performance by introducing a
serious bandwidth bottleneck.
3
3.1
Architecture
Operating principle
Long Short-Term Memory (LSTM) network layers [15] are often described with the following set of
canonical equations:
it
ft
ct
ot
ht
= σ(Wxi xt + Whi ht−1 + wci ct−1 + bi )
= σ(Wxf xt + Whf ht−1 + wcf ct−1 + bf )
= ft ct−1 + it tanh(Wxc xt + Whc ht−1 + bc )
= σ(Wxo xt + Who ht−1 + wco ct + bo )
= ot tanh(ct )
(1)
(2)
(3)
(4)
(5)
where x is the input state vector; i, f , o are called input, forget and output gates respectively; c
and h are the cell and hidden states. The subscript indicates either the current state t or the previous
2
xt
ht−1
•
Wxf
•
Whf
ct−1
•
σ
Wxi
•
Whi
x(0)
ft
•
σ
Wxc
•
Whc
bc
•
•
chip (0,1) y(0)
W(0,1)
chip (1,0)
W(1,0)
chip (1,1) y(1)
W(1,1)
ct
wco
Who
bo
computed across all chips
chip (0,0)
W(0,0)
it
tanh
Wxo
x(1)
wci
bi
•
Systolic Matrix Multiplication
wcf
bf
•
•
tanh
σ
ot
•
ht
•
Why
σ
yt
comp. only in right-most chips
Figure 1: Data dependency graph of a LSTM. The majority of computations are the vector-matrix
mult. (green) and can be distributed across multiple chips. Top-right: distribution of a vector matrix
mult. to a systolic array of chips.
t − 1, and denotes element-wise multiplication1 . The characteristic dimensions of all vectors and
matrices depend on the size of the input state (Nx ) and on that of the hidden state (Nh ). Multiple
LSTM layers can be connected by using the hidden state of one layer as input of the next. Finally,
LSTM networks often include a final densely connected layer without recurrence: yt = σ(Why ht ).
In C HIPMUNK, we exploit two distinct observations regarding LSTMs. First, all compute steps
are based on the same set of basic operations: i) matrix-vector products, ii) element-wise vector
products, and iii) element-wise non-linear activations. The internal datapath of C HIPMUNK can be
configured to execute these three basic operations (Section 3.2) and the LSTM state parameters are
stored on-chip. Second, the vast amount of data required to compute one time step of a RNN are
the weights. Storing them on-chip is thus essential to achieve high energy efficiency. To this end,
we a large share of the overall chip area is dedicated to SRAM to keep the weights local. For larger
LSTMs not fitting on a single chip, we allow operation in a systolic mode where the weights are
split across multiple connected chips and only the much smaller intermediate results are exchanged
as further discussed in Section 3.3.
3.2
Tile architecture
A product between a matrix of size A × B and a vector of size B is composed of two nested loops,
i.e. in pseudo-code:
for a in range(0, A):
# row loop
for b in range(0, B): # column loop
z += W[a,b] * x[b]
In C HIPMUNK, the row loop is executed on multiple parallel units, while the inner loop is executed
sequentially.
Fig. 2a shows a high-level diagram of the C HIPMUNK LSTM datapath that implements this functionality. Nlstm parallel LSTM units are used to execute all the iterations of the row loop at the same
time. Each LSTM unit is composed of an embedded memory bank to store weights (W ), registers
for storing the ot , ft , it and ct values locally, a multiply-accumulate unit and two lookup tables to
implement the non-linear activation functions. xt and ht are kept outside of the LSTM units, in a
bank of Nlstm registers. At each cycle of a column loop, one element of the input state and one
of the hidden state are selected depending on the iteration index and broadcast to all LSTM units.
Fig. 2b shows the basic operation loops composing a LSTM network deployed on C HIPMUNK.
1
In most literature Eqs. (1), (2) and (4) use matrix notation for Wci , Wcf and Wco ; however as these
matrices are diagonal by construction, we use the element-wise product notation here for consistence with what
is actually implemented in the C HIPMUNK hardware.
3
(a) LSTM datapath.
(b) Sequence of datapath basic operation loops.
Figure 2: LSTM datapath used in C HIPMUNK and typical sequence of operations. The datapath can
be used to implement the operations in Eqs. (1) to (5) by appropriately controlling the muxes and
clearing the register states.
All state variables use 8 bit fixed point precision, while 16 bits are used within the multiplyaccumulate block to minimize overflows. I/O is performed via an input stream port and an output
stream port, each consisting of 8 bits of data and 2 bits to enable a simple ready/valid handshake.
Weights are loaded at the beginning of the computation of a LSTM layer, and inputs are streamed in
sequentially. The internal state of the LSTM cell in terms of cell state and hidden state is retained
between consecutive LSTM input “frames” to implement the recurrent nature of the network. A
C HIPMUNK engine can be used to implement a full LSTM network with Nx , Nh ≤ Nlstm storing
the weights on chip. Larger networks require to stream them in from an external source.
3.3
Systolic scaling
As the main target of the C HIPMUNK accelerator is to enable ultra-low latency applications such
as on-device real-time speech recognition, the computing power of a single engine might not be
sufficient. A single engine cannot be arbitrarily scaled up: LSTM units are all coupled to the same
set of registers via simple multiplexers, making it impractical to increase Nlstm above a few hundred
units. Instead, to provide a more scalable and elegant solution, we designed C HIPMUNK so that
multiple engines can be connected as tiles and share the burden of the RNN computation in a spatial
fashion.
Fig. 3 shows how the computation is split between multiple tiles in the case of a 3×3 array. The input
state is split into vectors of size Nlstm and each vector is broadcast vertically along a column. The
4
Figure 3: C HIPMUNK tile I/O and operation of a 3×3 systolic array during the load of the input state
xt , computation of the new it , ot , ft , ct , ht state values, and redistribution of the updated hidden
state ht .
CHIPMUNK (1 tile)
WEIGHT
SRAM #0
WEIGHT
SRAM #3
WEIGHT
SRAM #6
WEIGHT
SRAM #4
WEIGHT
SRAM #2
WEIGHT
SRAM #11
WEIGHT
SRAM #10
WEIGHT
SRAM #7
WEIGHT
SRAM #5
core pitch: 0.96 mm
LSTM CELLS
WEIGHT
SRAM #1
WEIGHT
SRAM #8
WEIGHT
SRAM #9
TECHNOLOGY
UMC 65nm
HVT std cells
AREA (core)
0.926 mm2
AREA (die)
1.568 mm2
I/O
32 logic pins
8 power pins
81.7 kB
MEMORY
PERFORMANCE 32 Gop/s @1.24 V
POWER (core)
OP. RANGE
3.8 Gop/s @0.75V
29 mW @1.24 V
1.2 mW @0.75V
0.75V - 1.24V
die pitch: 1.25 mm
Figure 4: Microphotograph of a C HIPMUNK die.
new value for the internal gates/states is computed by accumulating the results computed by each
row. Finally, the last column can compute the output hidden state, which is broadcasted vertically
to the columns for the next iteration (cf. Fig. 3c). For a given network size/systolic configuration,
these connections can be hard-wired such that no external multiplexing is required.
4
Results & Discussion
4.1
Silicon prototype & Comparison with State-of-the-Art
We designed and built a silicon prototype based on a single C HIPMUNK tile as described in Section
3.2. The prototype chip was fabricated in UMC 65 nm technology, using high voltage threshold
cells to minimize leakage. It features Nlstm = 96 LSTM units, which hold their weight and bias
parameters in 12 separate SRAM banks (81.7 kB in total). The full chip, shown in Fig. 4, occupies
1.57 mm2 including the pads. The chip exposes the interface described in Section 3.3 for tile-to-tile
communication, so that it would be possible to prototype a systolic array using many discrete chips.
Fig. 5 shows the experimental results obtained by testing the C HIPMUNK prototype at room temperature (25 ◦C). The prototype is fully functional in an operating range between 0.75 V (limited
by SRAM failure) and 1.24 V, corresponding to a range of 20 to 168 MHz of maximum clock
frequency and from 1.24 to 29 mW of power consumption. The peak performance in terms of operations per second2 of one C HIPMUNK chip is 32.2 Gop/s (at 1.24 V) and the peak energy efficiency
(3.08 Gop/s/mW) is reached at 0.75 V.
Table 1 compares architectural parameters and synthetic results between C HIPMUNK and the existing VLSI and FPGA-based implementations for which performance and energy numbers have been
published. Our work reaches comparable performance with the DNPU proposed by Shin et al. [14].
Performance is obviously below that claimed by Google TPU [10], but this is mostly due to the
different size. In fact, despite the TPU uses 28 nm integration, C HIPMUNK has 2.8× better area
2
As customary for neural network accelerators, we count 1 multiply-accumulate as 2 operations.
5
6
‡
†
*
UMC 65 nm CMOS
core: 0.93 mm2
die: 1.57 mm2
82 kB
8-16 bit
96
1.24 V / 0.75 V
168 / 20 MHz
29.03 / 1.24 mW
32.3 / 3.8 Gop/s
1.11 / 3.08 Gop/s/mW
34.4 Gop/s/mm2
65 nm CMOS
core*: ∼2.0 mm2
die: 16.0 mm2
10 kB
4-7 bit
64
1.1 V / 0.77 V
200 / 50 MHz
21 / 2.6 mW
25 / 6.25 Gop/s
1.10 / 2.22 Gop/s/mW
12.5 Gop/s/mm2
DNPU*[14]
28 nm CMOS
–
die: ∼300 mm2
28 MB
8-16 bit
66k
–
700 MHz
40-28 W
3.7-2.8 Top/s
<0.13 Gop/s/mW
12.3-9.3 Gop/s/mm2
Google TPU†[10]
Xilinx XCKU060
294k LUT
453k FF
4.2 MB
12 bit, pruned
–
–
200 MHz
41 W
2.5 Top/s (equiv.)‡
0.061 Gop/s/mW‡
–
Han et al.‡[11]
Xilinx Z7045
33k LUT
15k FF, 33 DSP
332 kB
5-16 bit
–
–
166 MHz
∼10 W
152 Gop/s
0.0152 Gop/s/mW
–
Rybalkin et al. [12]
Xilinx Z7020
7.6k LUT
13k FF, 50 DSP
–
16 bit
4
–
142 MHz
2.3 W
0.389 Gop/s
0.000146 Gop/s/mW
–
Chang et al. [13]
The DNPU is a mixed CNN-RNN processor. We report here only the figures related to the RNN subunit.
We present here the values from [10] based on the two LSTMs for which they measured the performance. For both, the TPU is severely memory bandwidth
limited.
They assume a well-structured sparsity of 11.2% in the weight matrices. Reported numbers are dense-equivalent throughput. Underlying compute throughput:
282 Gop/s.
On-chip memory
Arithmetic
Number of MACs
Core voltage
Frequency
Power
Peak performance
Energy efficiency
Area efficiency
Technology
Area
THIS WORK
Table 1: Comparison to Existing VLSI and FPGA Implementations
3.5
1.2
highest performance
32 Gop/s @ 1.24 V , 168 MHz
1.0
0.9
2.5
core power [mW]
core voltage [V]
1.1
2.0
1.5
0.8
0.7
1.0
highest efficiency
3.08 Gop/s/mW @ 0.75 V , 20 MHz
20
40
60
80
100
120
frequency [MHz]
140
160
180
0.7
0.8
0.9
1.0
1.1
1.2
energy efficiency [Gop/s/mW]
3.0
0.5
1.3
core voltage [V]
Figure 5: Frequency, power and performance of the C HIPMUNK prototype versus operating voltage
at room temperature (25 ◦C). The left shmoo plot shows core voltage versus operating frequency;
the color shade corresponds to the core power consumption (darker=less power). The right plot
shows energy efficiency versus core voltage; the color shade of the scattered dots corresponds to the
core power, while their size is proportional to the maximum frequency.
Table 2: CTC-3L-421H-UNI Speech Recognition LSTM Executed on C HIPMUNK With a 10 ms
Constraint
Configuration
PERF @1.24 V
EFF @0.75 V
Execution time
systolic 3×5×5
systolic 5×5
single
0.09 ms
1.59 ms
38.23 ms
0.76 ms
13.31 ms
321.14 ms
Peak power
systolic 3×5×5
systolic 5×5
single
1833.75 mW
611.25 mW
24.45 mW
165.75 mW
55.25 mW
2.21 mW
systolic 3×5×5
systolic 5×5
16.53 mW
96.89 mW
12.55 mW
-
Average power
efficiency - and a performance-wise “TPU-equivalent” array with ∼115 C HIPMUNK engines would
consume only 3.33 W, an order of magnitude less than the TPU. C HIPMUNK advances the state-ofthe-art energy efficiency with respect to the DNPU, showing a 39% improvement. Moreover, the
DNPU does not include any provision to address the fundamental memory boundedness of RNNs,
which C HIPMUNK addresses via systolic scaling. All FPGA implementations [11–13] are at least
two orders of magnitude less energy-efficient.
In terms of arithmetic precision we have chosen to use 8 bit fixed-point representations for storage
and perform the MAC operations with 16 bit precision. This is in line with Google’s TPU and higher
than the 4-7 bit of the DNPU.
4.2
Real-world speech recognition
To evaluate C HIPMUNK on a real-world problem, we targeted CTC-3L-421H-UNI, a 3-layer, 421hidden units per layer LSTM topology introduced by Graves et al. [1], which takes as input a stream
of 123 Mel-Frequency Cepstral Coefficients (MFCCs) extracted from an audio stream and identifies
phonemes with an error rate of 19.6%, evaluated on the TIMIT database. The MFCC input “frames”
are produced with a 10 ms rate, which means that any embedded low-latency real-time RNN implementation should be able to elaborate the full network in less than this time. We evaluate three
different C HIPMUNK configurations: a systolic array of 75 units, divided in 3 sub-arrays of 5 × 5
engines; a single array of 5 × 5 engines; and a single C HIPMUNK engine. The largest configuration
can host the full topology in a spatial fashion; each of the sub-arrays hosts one layer of the RNN.
After the initial programming phase, it does not need any reprogramming. The smaller arrays need
7
to be reconfigured at each new layer (in the 5 × 5 array case) or multiple times per layer (in the
single unit case).
Table 2 reports execution time and power for these three configurations. Execution times include
both computation and reconfiguration, excluding only the initial configuration which doesn’t need
to be repeated for each new frame/layer. Bold time/power values indicate configurations that can
meet the 10 ms deadline. As the CTC-3L-421H-UNI topology has ∼ 3.8 × 106 weights, a 3 × 5 × 5
systolic configuration is best used (all weights stored locally). Smaller configurations imply a >
80% overhead for reloading weights.
Average power, also shown in Table 2, is computed under the assumption that the array is perfectly
duty cycled when not in use over the 10 ms window. Even in the assumption that the C HIPMUNK
array is always-on, the 12.55 mW required to process this network would only add ∼4% to idle
power on a typical smartphone (300 to 400 mW [16]). Adding a filter to drop clearly uninteresting
input (e.g. silence) would likely decrease this overhead by an order of magnitude.
5
Conclusion
We have presented an architecture and silicon measurement results for a small (0.9 mm2 ) RNN hardware accelerator providing 3.8 Gop/s at 1.2 mW in 65 nm digital CMOS technology, resulting in new
state-of-the-art energy and area efficiencies of 3.08 Gop/s/mW and 34.4 Gop/s/mm2 . The systolic
design is scalable to accommodate also large RNNs efficiently by connecting multiple identical
chips on the circuit board.
Acknowledgements
This work was supported in part by the EU project ExaNoDe under grant H2020-671578, and in
part by the Swiss National Science Foundation project Micropower Deep Learning.
References
[1] A. Graves, A.-R. Mohamed, and G. Hinton, “Speech Recognition With Deep Recurrent Neural
Networks,” in Proc. IEEE ICASSP, 2013.
[2] W. Xiong, J. Droppo, X. Huang, F. Seide, M. Seltzer, A. Stolcke, D. Yu, and G. Zweig, “The
Microsoft 2016 Conversational Speech Recognition System,” in Proc. IEEE ICASSP, 2017,
pp. 5255–5259.
[3] K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio, “Learning Phrase Representations using RNN EncoderDecoder for Statistical Machine
Translation,” in Proc. ACL EMNLP, 2014, pp. 1724–1734.
[4] L. Cavigelli and L. Benini, “A 803 GOp/s/W Convolutional Network Accelerator,” IEEE
TCSVT, 2016.
[5] F. Conti, R. Schilling, P. D. Schiavone, A. Pullini, D. Rossi, F. K. Gurkaynak, M. Muehlberghuber, M. Gautschi, I. Loi, G. Haugou, S. Mangard, and L. Benini, “An IoT Endpoint Systemon-Chip for Secure and Energy-Efficient Near-Sensor Analytics,” IEEE TCAS, vol. 64, no. 9,
pp. 2481–2494, 9 2017.
[6] R. Andri, L. Cavigelli, D. Rossi, and L. Benini, “YodaNN: An Architecture for Ultra-Low
Power Binary-Weight CNN Acceleration,” IEEE TCAD, 2017.
[7] Y.-H. Chen, T. Krishna, J. Emer, and V. Sze, “Eyeriss: An Energy-Efficient Reconfigurable
Accelerator for Deep Convolutional Neural Networks,” in Proc. IEEE ISSCC, 2016, pp. 262–
263.
[8] Z. Du, R. Fasthuber, T. Chen, P. Ienne, L. Li, T. Luo, X. Feng, Y. Chen, and O. Temam,
“ShiDianNao: Shifting Vision Processing Closer to the Sensor,” in Proc. ACM/IEEE ISCA,
2015, pp. 92–104.
[9] V. Sze, Y.-H. Chen, T.-J. Yang, and J. Emer, “Efficient Processing of Deep Neural Networks:
A Tutorial and Survey,” arXiv:703.09039, 2017.
[10] N. P. Jouppi, A. Borchers, R. Boyle, P.-l. Cantin, C. Chao, C. Clark, J. Coriell, M. Daley,
M. Dau, J. Dean, B. Gelb, C. Young, T. V. Ghaemmaghami, R. Gottipati, W. Gulland, R. Hagmann, C. R. Ho, D. Hogberg, J. Hu, R. Hundt, D. Hurt, J. Ibarz, N. Patil, A. Jaffey, A. Jaworski,
8
[11]
[12]
[13]
[14]
[15]
[16]
A. Kaplan, H. Khaitan, D. Killebrew, A. Koch, N. Kumar, S. Lacy, J. Laudon, J. Law, D. Patterson, D. Le, C. Leary, Z. Liu, K. Lucke, A. Lundin, G. MacKean, A. Maggiore, M. Mahony,
K. Miller, R. Nagarajan, G. Agrawal, R. Narayanaswami, R. Ni, K. Nix, T. Norrie, M. Omernick, N. Penukonda, A. Phelps, J. Ross, M. Ross, A. Salek, R. Bajwa, E. Samadiani, C. Severn, G. Sizikov, M. Snelham, J. Souter, D. Steinberg, A. Swing, M. Tan, G. Thorson, B. Tian,
S. Bates, H. Toma, E. Tuttle, V. Vasudevan, R. Walter, W. Wang, E. Wilcox, D. H. Yoon,
S. Bhatia, and N. Boden, “In-Datacenter Performance Analysis of a Tensor Processing Unit,”
in Proc. ACM ISCA, 2017.
S. Han, J. Kang, H. Mao, Y. Hu, X. Li, Y. Li, D. Xie, H. Luo, S. Yao, Y. Wang, H. Yang,
and W. J. Dally, “ESE: Efficient Speech Recognition Engine with Sparse LSTM on FPGA,” in
Proc. ACM/SIGDA FPGA, 2016.
V. Rybalkin, N. Wehn, M. R. Yousefi, and D. Stricker, “Hardware Architecture of Bidirectional
Long Short-Term Memory Neural Network for Optical Character Recognition,” in Proc. IEEE
DATE, 2017, pp. 1390–1395.
A. X. M. Chang and E. Culurciello, “Hardware Accelerators for Recurrent Neural Networks
on FPGA,” in Proc. IEEE ISCAS, 2017.
D. Shin, J. Lee, J. Lee, and H. J. Yoo, “DNPU: An 8.1TOPS/W Reconfigurable CNN-RNN
Processor for General-Purpose Deep Neural Networks,” in Proc. IEEE ISSCC, vol. 60, 2017,
pp. 240–241.
S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, vol. 9,
no. 8, pp. 1735–1780, 1997.
A. Carroll and G. Heiser, “An Analysis of Power Consumption in a Smartphone,” in USENIX
ATC, 2010.
9
| 9cs.NE
|
1
An Approximate Pareto Set for Minimizing the
Maximum Lateness and Makespan on Parallel
Machines
arXiv:1802.10488v1 [cs.DS] 28 Feb 2018
Gais Alhadi1 , Imed Kacem2 , Pierre Laroche3 , and Izzeldin M. Osman4
Abstract—We consider the two-parallel machines scheduling
problem, with the aim of minimizing the maximum lateness and
the makespan. Formally, the problem is defined as follows. We
have to schedule a set J of n jobs on two identical machines.
Each job i ∈ J has a processing time pi and a delivery time
qi . Each machine can only perform one job at a given time.
The machines are available at time t = 0 and each of them
can process at most one job at a given time. The problem is
to find a sequence of jobs, with the objective of minimizing the
maximum lateness Lmax and the makespan Cmax . With no loss
of generality, we consider that all data are integers and that
jobs are indexed in non-increasing order of their delivery times:
q1 ≥ q2 ≥ . . . ≥ qn . This paper proposes an exact algorithm
(based on a dynamic programming) to generate the complete
Pareto Frontier in a pseudo-polynomial time. Then, we present
an FPTAS (Fully Polynomial Time Approximation Scheme) to
generate an approximate Pareto Frontier, based on the conversion
of the dynamic programming. The proposed FPTAS is strongly
polynomial. Some numerical experiments are provided in order
to compare the two proposed approaches.
I. I NTRODUCTION
We consider the two-parallel machines scheduling problem, with the aim of minimizing the maximum lateness and
makespan. Formally, the problem is defined as follows. We
have to schedule a set J of n jobs on two identical machines.
Each job i ∈ J has a processing time pi and a delivery time
qi . The machines are available at time t=0 and each of them
can process at most one job at a time. The problem is to
find a sequence of jobs, with the objective of minimizing the
maximum lateness Lmax and the makespan Cmax . With no
loss of generality, we consider that all data are integers and
that jobs are indexed in non-increasing order of their delivery
times q1 ≥ q2 ≥ . . . ≥ qn .
For self-consistency, we recall some necessary definitions
related to the approximation area. An algorithm A is called
a ρ−approximation algorithm for a given problem, if for any
instance I of that problem the algorithm A yields, within a
polynomial time, a feasible solution with an objective value
A(I) such that: |A(I) − OP T (I)| ≤ .OP T (I), where
1 Gais Alhadi is a member
Computer Sciences, University
of Faculty of Mathematical and
of Gezira, Wad-Madani, Sudan
[email protected]
Imed
Kacem2
and
Pierre
Laroche3
are
members
of
the
LCOMS
Laboratory,
University
of
Lorraine,
F-57045
Metz,
France
[email protected],
[email protected]
Izzeldin M. Osman4 is from Sudan University of Science and Technology,
Khartoum, Sudan [email protected]
OP T (I) is the optimal value of I and ρ is the performance
guarantee or the worst-case ratio of the approximation algorithm A. It can be a real number greater or equal to 1
for the minimization problems ρ = 1 + (that it leads
to inequality A(I) ≤ (1 + )OP T (I)), or it can be real
number from the interval [0, 1] for the maximization problems
ρ = 1 − (that it leads to inequality A(I) ≥ (1 − )OP T (I)).
The Pareto-optimal solutions are the solutions that are not
dominated by other solutions. Thus, we can consider that
the solution is Pareto-optimal if there does not exist another
solution that is simultaneously the best for all the objectives.
Noteworthy, Pareto-optimal solutions represent a range of
reasonable optimal solutions for all possible functions based
on the different objectives. A schedule is called Pareto-optimal
if it is not possible to decrease the value of one objective
without increasing the value of the other.
It is noteworthy that during the last decade the multi-objective
scheduling problems have attracted numerous researchers from
all the world and have been widely studied in the literature. For the scheduling problems on serial-batch machine,
Geng et al.[17] studied scheduling problems with or without
precedence relations, where the objective is to minimize
makespan and maximum cost. They have provided highly
efficient polynomial-time algorithms to generate all Pareto
optimal points. An approximate Pareto set of minimal size
that approximates within an accuracy for multi-objective
optimization problems have been studied by Bazgan et al.[2].
They proposed a 3-approximation algorithm for two objectives and also proposed a study of the greedy algorithm
performance for a three-objective case when the points are
given explicitly in the input. They showed that the threeobjective case is NP-hard. Chen and Zou [5] proposed a
runtime analysis of a (µ + 1) multi-objective evolutionary
algorithm for three multi-objective optimization problems with
unknown attributes. They showed that when the size of the
population is less than the total number of Pareto-vector, the
(µ + 1) multi-objective evolutionary algorithm cannot obtain
the expected polynomial runtime for the exact discrete multiobjective optimization problems. Thus, we must determine the
size of the population equal to the total number of leading
ones, trailing zeros. Furthermore, the expected polynomial
runtime for the exponential discrete multi-objective optimization problem can be obtained by the ratio of n/2 to µ − 1
over an appropriate period of time. They also showed that the
(µ + 1) multi-objective evolutionary algorithm can be solved
efficiently in polynomial runtime by obtaining an − adaptive
2
Pareto front. Florios and Mavrotas [6] used AUGMECON2,
a multi-objective mathematical programming method (which
is suitable for general multi-objective integer programming
problems), to produce all the Pareto-optimal solutions for
multi-objective traveling salesman and set covering problems.
They showed that the performance of the algorithm is slightly
better than it already exists. Moreover, they showed that their
results can be helpful for other multi-objective mathematical
programming methods or even multi-objective meta-heuristics.
In [10], Sabouni and Jolai proposed an optimal method for
the problem of scheduling jobs on a single batch processing
machine to minimize the makespan and the maximum lateness.
They showed that the proposed method is optimal when the
set with maximum lateness objective has the same processing times. They also proposed an optimal method for the
group that has the maximum lateness objective and the same
processing times. Geng et al.[3] considered the scheduling
problem on an unbounded p-batch machine with family jobs
to find all Pareto-optimal points for minimizing makespan and
maximum lateness. They presented a dynamic programming
algorithm to solve the studied problem. He et al.[4] showed
that the Pareto optimization scheduling problem on a single
bounded serial-batching machine to minimize makespan and
maximum lateness is solvable in O(n6 ). They also presented
an O(n3 )- time algorithm to find all Pareto optimal solutions
where the processing times and deadlines are agreeable. For
the bi-criteria scheduling problem, He et al.[8] showed that the
problem of minimizing maximum cost and makespan is solvable in O(n5 ) time. The authors presented a polynomial-time
algorithm in order to find all Pareto optimal solutions. Also,
He et al.[9] showed that the bi-criteria batching problem of
minimizing maximum cost and makespan is solvable in O(n3 )
time. The bi-criteria scheduling problem on a parallel-batching
machine to minimize maximum lateness and makespan have
been considered in [11]. The authors presented a polynomialtime algorithm in order to find all Pareto optimal solutions.
Allahverdi and Aldowaisan [13] studied the no-wait flow-shop
scheduling problem with bi-criteria of makespan or maximum
lateness. They also proposed a dominance relation and a
branch-and-bound algorithm and showed that these algorithms
are quite efficient.
The remainder of this paper is organized as follows. In
Section 2, we describe the proposed dynamic programming
(DP) algorithm. Section 3, provides the description and the
analysis of the FPTAS. In Section 4, we present a practical
example for DP and FPTAS. Finally, Section 5 concludes the
paper.
II. DYNAMIC P ROGRAMMING A LGORITHM
The following dynamic programming algorithm A, can
be applied to solve exactly this problem. This algorithm A
generates iteratively some sets of states. At every iteration i,
a set χi composed of states is generated (0 ≤ i ≤ n). Each
state [k, Lmax , Cmax ] in χi can be associated to a feasible
partial schedule for the first i jobs. Let variable k ∈ {0, 1}
denote the most loaded machine, Lmax denote the maximum
lateness and Cmax denote the maximum completion time
of the corresponding schedule. The dynamic programming
algorithm can be described as follows.
A LGORITHM A
1) Set χ1 = {[1, p1 + q1 , p1 ]}.
2) For i ∈ {2, 3, ..., n},
a) χi = φ.
b) For every state [k, Lmax , Cmax ] in χi−1 :
• (schedule job i on machine k)
add [k,max{Lmax ,Cmax + pi + qi },Cmax + pi ]
to χi
• (schedule job i on machine 1 − k)
Pi
if (Cmax ≥ j=1 pj − Cmax )
Pi
– add [k,max{Lmax , j=1 pj − Cmax +
qi },Cmax ] to χi
else
Pi
– add [1 − k,max{Lmax , j=1 pj − Cmax +
Pi
qi }, j=1 pj − Cmax ] to χi
c) For every k, for every Cmax : keep only one state
with the smallest possible Lmax .
d) Remove χi−1 .
3) Return the Pareto front of χn , by only keeping nondominated states.
Remark: To destroy the symmetry, we start by χ1 = {[1, p1 +
q1 , p1 ]} (i.e., we perform job 1 on the first machine).
III. A PPROXIMATE PARETO F RONTIER
The main idea of the Approximate Pareto Frontier is to
remove a special part of the states generated by the dynamic
programming algorithm A. Therefore, the modified algorithm
A0 described in Lemma III.1 produces an approximation
solution instead of the optimal solution.
Given an arbitrary > 0, we define the following parameters:
δ1 =
P/2
,
n
and
(P + qmax )/3
.
n
where qmax is the maximum delivery time and P is the total
sum of processing times.
δ2 =
∗
Let L∗max and Cmax
be the optimal solutions for our
two objectives. Let LM AX and CM AX be the upper
bounds for the two considered criteria (scheduling all the
jobs on the same machine), such that,
0 ≤ LM AX = P + qmax ≤ 3L∗max
∗
0 ≤ CM AX = P ≤ 2Cmax
We divide the intervals [0, CM AX] and [0, LM AX] into
equal sub-intervals respectively of lengths δ1 and δ2 . Then,
an FPTAS is defined by following the same procedure as in
the dynamic programming, except the fact that it will keep
only one representative state for every couple of the defined
subintervals produced from [0, CM AX] and [0, LM AX].
Thus, our FPTAS will generate approximate sets χ#
i of states
3
instead of χi . The following lemma shows the closeness of
the result generated by the FPTAS compared to the dynamic
programming.
Lemma III.1. For every state [k, Lmax , Cmax ] ∈ χi there
#
#
exists at least one approximate state [m, L#
max , Cmax ] ∈ χi
such that:
L#
max ≤ Lmax + i. max{δ1 , δ2 },
Cmax − i.δ1 ≤
≤ Cmax + i.δ1 .
Proof. By induction on i.
First, for i = 0 we have χ#
i = χ1 . Therefore, the statement is
trivial. Now, assume that the lemma holds true up to level i−1.
Consider an arbitrary state [k, Lmax , Cmax ] ∈ χi . Algorithm
A introduces this state into χi when job i is added to some
0
0
0
feasible state for the first i − 1 jobs. Let [k , Lmax , Cmax ] be
the above feasible state. Three cases can be distinguished:
0
0
0
1) [k, Lmax , Cmax ] = [k , max{Lmax ,Cmax + pi + qi },
0
Cmax + pi ]
Pi
0
0
0
2) [k, Lmax , Cmax ] = [k , max{Lmax , j=1 pj − Cmax +
0
qi }, Cmax ]
Pi
0
0
3) [k, Lmax , Cmax ] = [1 − k , max{Lmax , j=1 pj −
Pi
0
0
Cmax + qi }, j=1 pj − Cmax ]
We will prove the statement for level i in the three cases.
0
0
0
st
• 1
Case: [k, Lmax , Cmax ] = [k , max{Lmax ,Cmax +
0
pi + qi }, Cmax + pi ]
0
0
0
∈
χi−1 , there exists
Since [k ,Lmax ,Cmax ]
0
0
0
#
#
] ∈ χ#
, Cmax
,
such
that:
[k # , Lmax
i−1
0
0
#
≤
Lmax + (i − 1) max{δ1 , δ2 } and
Lmax
0
0
0
#
≤ Cmax + (i − 1)δ1 .
Cmax − (i − 1)δ1 ≤ Cmax
0
0
0
#
#
#
+ pi +
, Cmax
Consequently, the state [k , max{Lmax
0
0
#
qi }, Cmax + pi ] is created by algorithm A at iteration i.
However, it may be removed when reducing the state subset.
Let [α, λ, µ] be the state in χ#
same box as the
i that is in the
0
0
0
0
#
#
#
+ pi ]. Hence,
+ pi + qi }, Cmax
, Cmax
sate [k # , max{Lmax
we have:
0
0
0
0
≤ Lmax + i. max{δ1 , δ2 }
(1)
In addition,
Pi
0
0
0
0
#
#
#
Here, the state [k # , max{Lmax
, j=1 pj −Cmax
+qi },Cmax
]
0
is created by algorithm A at iteration i. However, it may be removed when reducing the state subset. Let [α, λ, µ] P
be the state
0
0
i
#
#
in χ#
that
is
in
the
same
box
as
[k
,
max{L
,
max
i
j=1 pj −
0
0
#
#
Cmax + qi }, Cmax ]. Hence, we have:
λ
0
#
,
≤ max{Lmax
i
X
0
#
+ qi } + δ2
pj − Cmax
j=1
0
≤ max{Lmax + (i − 1) max{δ1 , δ2 },
i
X
pj
j=1
0
−(Cmax − (i − 1)δ1 ) + qi } + δ2
i
X
0
0
pj − Cmax + qi }
≤ max{Lmax ,
j=1
+(i − 1) max{δ1 , δ2 } + δ2
≤ Lmax + (i − 1) max{δ1 , δ2 } + δ2
<
Lmax + i. max{δ1 , δ2 }
(4)
Moreover,
0
0
0
0
#
µ ≥ Cmax
− δ1 ≥ Cmax − (i − 1)δ1 − δ1 = Cmax − iδ1 . (6)
0
#
Cmax
+ pi + δ1
(2)
Consequently, [α, λ, µ] is an approximate state verifying the
two conditions.
Pi
0
0
#
#
• Sub-case 2.2:
j=1 pj − Cmax > Cmax
(3)
Pi
0
0
0
#
#
Here, the state [1 − k # , max{Lmax
, j=1 pj − Cmax
+
Pi
0
0
#
qi }, j=1 pj − Cmax ] is created by algorithm A at iteration i.
However, it may be removed when reducing the state subset.
#
Let [α, λ, µ] be the state
is in P
the same box as
Pi in χi 0that
0
0
0
i
#
#
#
#
[1−k , max{Lmax , j=1 pj −Cmax
+qi }, j=1 pj −Cmax
].
Hence, we have:
0
Cmax + (i − 1)δ1 + pi + δ1 = Cmax + iδ1 .
and,
≥
0
And,
+(i − 1). max{δ1 , δ2 } + δ2
µ ≥
0
Since [k ,Lmax ,Cmax ]
∈
χi−1 , there exists
0
0
0
#
#
[k # , Lmax
, Cmax
] ∈ χ#
,
such
that:
i−1
0
0
#
Lmax
≤
Lmax + (i − 1)max{δ1 , δ2 } and
0
0
0
#
Cmax − (i − 1)δ1 ≤ Cmax
≤ Cmax + (i − 1)δ1 .
#
µ ≤ Cmax
+ δ1 ≤ Cmax + (i − 1)δ1 + δ1 = Cmax + iδ1 . (5)
≤ max{Lmax , Cmax + pi + qi }
≤
0
=
#
#
≤ max{Lmax
, Cmax
+ pi + qi } + δ2
λ
µ ≤
2nd
Case:P
[k, Lmax , Cmax ]
0
0
0
0
i
[k ,max{Lmax , j=1 pj − Cmax + qi },Cmax ]
Consequently, two sub-cases can occur:
Pi
0
0
#
#
• Sub-case 2.1:
j=1 pj − Cmax ≤ Cmax
and
#
Cmax
•
0
0
#
Cmax
+ pi − δ1 ≥ Cmax − (i − 1)δ1 + pi − δ1
Cmax − iδ1 .
Consequently, [α, λ, µ] is an approximate state verifying the
two conditions.
4
0
0
#
[1−k # ,max{Lmax
,
Hence, we have:
0
#
≤ max{Lmax
,
λ
i
X
0
#
pj − Cmax
+ qi } + δ2
λ
≤
j=1
#
pj −Cmax
+qi },
j=1
0
#
max{Lmax
,
i
X
Pi
0
j=1
#
pj −Cmax
].
0
#
pj − Cmax
+ qi } + δ2
j=1
0
≤ max{Lmax + (i − 1) max{δ1 , δ2 },
i
X
pj
≤
j=1
0
max{Lmax + (i − 1)max{δ1 , δ2 },
i
X
pj
j=1
0
−(Cmax + (i − 1)δ1 ) + qi } + δ2
i
X
0
0
≤ max{Lmax ,
pj − Cmax + qi }
0
≤
j=1
−(Cmax − (i − 1)δ1 ) + qi } + δ2
i
X
0
max{Lmax ,
pj
j=1
+(i − 1) max{δ1 , δ2 } + δ2
0
−Cmax + qi } + (i − 1) max{δ1 , δ2 } + δ2
≤ Lmax + (i − 1) max{δ1 , δ2 } + δ2
<
0
Pi
Lmax + i. max{δ1 , δ2 }
≤ Lmax + (i − 1) max{δ1 , δ2 } + δ2
(7)
≤ Lmax + i max{δ1 , δ2 }
(11)
and
Moreover,
µ ≤
i
X
i
X
µ ≤
0
#
+ δ1
pj − Cmax
0
#
+ δ1
pj − Cmax
j=1
j=1
≤
(8)
i
X
0
pj − (Cmax − (i − 1)δ1 ) + δ1
j=1
≤ Cmax + iδ1 .
0
0
#
≥ Cmax − (i − 1)δ1 , then the following relation
Since Cmax
holds
i
X
µ ≤
(12)
In the other hand, we have
µ ≥
0
pj − Cmax + (i − 1)δ1 + δ1 ≤ Cmax + iδ1
i
X
0
#
pj − Cmax
− δ1
j=1
j=1
(9)
≥
i
X
0
pj − (Cmax + (i − 1)δ1 ) − δ1
j=1
(since
And,
0
0
Pi
j=1
≥ Cmax − iδ1 .
pj − Cmax ≤ Cmax ).
i
X
µ ≥
0
0
#
#
− δ1
− δ1 ≥ Cmax
pj − Cmax
j=1
≥
Cmax − (i − 1)δ1 − δ1 = Cmax − iδ1 .
(10)
Thus, [α, λ, µ] verifies the necessary conditions.
(13)
Thus, [α, λ, µ] fulfills the conditions.
Pi
0
0
#
#
• Sub-case 3.2:
j=1 pj − Cmax < Cmax
P
0
0
0
0
i
#
#
#
]
, j=1 pj −Cmax
+qi },Cmax
Here, the state [k # , max{Lmax
0
is created by algorithm A at iteration i. However, it may be removed when reducing the state subset. Let [α, λ, µ] P
be the state
0
0
i
#
#
in χ#
that
is
in
the
same
box
as
[k
,
max{L
,
max
i
j=1 pj −
0
0
#
#
Cmax + qi },Cmax ]. Hence, we have:
0
•
3rd Case:P
[k, Lmax , Cmax ] P =
[1 − k ,
0
0
0
i
i
max{Lmax , j=1 pj − Cmax + qi }, j=1 pj − Cmax ]
0
0
Pi
0
j=1
0
#
max{Lmax
,
i
X
0
#
pj − Cmax
+ qi }
j=1
+δ2
≤
0
max{Lmax + (i − 1) max{δ1 , δ2 },
i
X
pj
j=1
0
0
#
#
pj − Cmax
≥ Cmax
Pi
0
0
0
#
#
Here, the state [1 − k # ,max{Lmax
, j=1 pj − Cmax
+
Pi
0
0
#
qi }, j=1 pj − Cmax ] is created by algorithm A at iteration i.
However, it may be removed when reducing the state subset.
Let [α, λ, µ] be the state in χ#
i that is in the same box as
Sub-case 3.1:
≤
0
Since [k ,Lmax ,Cmax ]
∈
χi−1 , there exists
0
0
0
#
#
#
#
[k , Lmax , Cmax ] ∈ χi−1 , such that:
0
0
#
Lmax
≤
Lmax + (i − 1) max{δ1 , δ2 } and
0
0
0
#
Cmax − (i − 1)δ1 ≤ Cmax
≤ Cmax + (i − 1)δ1 .
Consequently, two sub-cases can occur:
•
λ
≤
−(Cmax − (i − 1)δ1 ) + qi } + δ2
i
X
0
0
max{Lmax ,
pj − Cmax + qi }
j=1
+(i − 1) max{δ1 , δ2 } + δ2
≤
Lmax + (i − 1) max{δ1 , δ2 } + δ2
≤ Lmax + i max{δ1 , δ2 }
(14)
5
0
and
µ ≤
≤
0
0
#
Cmax
+ δ1 ≤ Cmax + (i − 1)δ1 + δ1
Cmax + iδ1 .
(15)
In the other hand, we have
0
#
µ ≥ Cmax
− δ1 ≥
i
X
0
#
pj − Cmax
− δ1
j=1
≥
i
X
0
pj − Cmax − iδ1 = Cmax − iδ1 .
(16)
j=1
Therefore, [α, λ, µ] fulfills the conditions.
In conclusion, the statement holds also for level i in
the third case, and this completes our inductive proof.
Based on the lemma, we deduce easily that for every nondominated state [k, Lmax , Cmax ] ∈ χn , it must remain a close
#
#
state [m, L#
max , Cmax ] ∈ χn such that:
L#
max ≤ Lmax + n. max{δ1 , δ2 } ≤ (1 + ).Lmax ,
and
#
Cmax
≤ Cmax + n.δ1 ≤ (1 + ).Cmax .
Moreover, it is clear that the FPTAS runs polynomially in n
and 1/. The overall complexity of our FPTAS is O(n3 /2 ).
IV. R ESULTS
The following results have been obtained after testing the
performance of the proposed algorithms. The code has been
done in Java and the experiments were performed on an
Intel(R) Core(TM)-i7 with 8GB RAM. We randomly generate
five sets of instances, with different numbers of jobs and
various processing and delivery times:
• number of jobs: from 5 to 25, 26 to 50, 51 to 75, 76 to
100 and 100 to 200
• processing times : from 1 to 20, 1 to 100 and 1 to 1000
• delivery times : from 1 to 20, 1 to 100 and 1 to 1000
That gave us 135 instances in each set of instances. Finally,
the FPTAS has been tested with two values of : 0.3 and 0.9.
To ensure the consistency of running times, each test has been
run three times.
Figure 1 presents a comparison of FPTAS and Dynamic
Programming. The left part of this figure shows the average
size of the Pareto Front (i.e. the number of solutions) found
by the Dynamic Programming algorithm and our FPTAS with
the two values we used. The sizes are given for our five sets
of instances, from small instances (5-25 instances) to bigger
ones (100-200 jobs). We can see that the number of solutions
decreases as the number of jobs increases. With a lot of jobs,
it is more likely to obtain very similar solutions, a lot of them
being dominated by others. At the opposite, the number of
jobs has no real influence on the Pareto front sizes found by
our FPTAS algorithm, whatever the value of .
On the right part of the same figure are given the average
0
∗
quality of the two objectives of our study: LA
max /Lmax and
A
∗
Cmax
/Cmax
. We can see that the FPTAS algorithms are
finding solutions closer to the optimal ones when the number
of jobs is increasing. Cmax values are closer to the optimal
than Lmax values, which is not a surprise, as Lmax depends
on Cmax . Worth to mention, our FPTAS with = 0.3 gives
better results than with = 0.9, which is consistent with the
theory.
We have also studied the influence of processing and
delivery times, in Table I and II. These tables present average
results for our benchmark, considering the 5 sets of instances.
Results are presented for = 0.3; the analysis is the same
with = 0.9. Table I shows that instances composed of jobs
with various processing times lead to optimal solutions with
a bigger Pareto front, as seen in column 2. At the opposite,
columns 3 to 5 show that our FPTAS algorithm is not really
influenced by this parameter. Table II shows that delivery
times ranges have more influence on the results : the size of
the Pareto front of non-dominated solutions grows faster. Our
FPTAS algorithm has the same behavior, and we can see that
the results are more close to the optimal with smaller values
of qi .
Table I: Quality of FPTAS as a function of processing time
ranges (for = 0.3)
pi range
1-20
1-100
1-500
DP size of
Pareto front
2.26
4.07
5.65
FPTAS size of
Pareto front
1.23
1.84
1.73
0
0
A
∗
Cmax
/Cmax
∗
LA
max /Lmax
1.0006
1.0008
1.0005
1.003
1.007
1.004
Table II: Quality of FPTAS as a function of delivery time
ranges (for = 0.3)
qi range
1-20
1-100
1-500
DP size of
Pareto front
2.02
3.03
6.93
FPTAS size of
Pareto front
1.25
1.57
1.99
0
0
A
∗
Cmax
/Cmax
∗
LA
max /Lmax
1.0007
1.0007
1.0005
1.001
1.002
1.008
Computing times are given in Tables III, IV and V. They
compare our Dynamic Programming algorithm and our FPTAS, considering two values : 0.3 and 0.9. All values
are in milliseconds. Table III shows that all algorithms are
slower when the number of states is growing. Table IV
shows an interesting result: while the Dynamic Programming
algorithm becomes slower when the processing times ranges
are growing, the FPTAS has an opposite behavior. The FPTAS
with = 0.3 is even slower than the exact algorithm for the
smallest range of processing times. Table V shows that the
delivery times ranges have a smaller influence on the Dynamic
Programming algorithm: computing times are growing, but
slower than the delivery times ranges. The FPTAS computing
times are also growing in function of the delivery times ranges.
Note that tables IV and V are based on mean values from all
the set of instances.
6
Quality of Cmax and Lmax
Size of Pareto Front
Figure 1: Quality of our FPTAS algorithm with = 0.3 and = 0.9
Table III: Average computing times vs size of instances (ms)
#jobs
5-25
26-50
51-75
76-100
100-200
DP
67
1278
7917
24937
164332
FPTAS
= 0.3 = 0.9,
0.9
0.3
5.4
0.9
22
3.4
58
7.8
281
32
Table IV: Average computing times (ms) vs processing time
ranges
pi ranges
1-20
1-100
1-500
DP
91
2668
116362
FPTAS
= 0.3
= 0.9,
144
15
51
7
25
5
V. C ONCLUSIONS AND P ERSPECTIVES
The two-parallel machines scheduling problem has been
considered to minimize the maximum lateness and the
makespan. We have proposed an exact algorithm (based on
dynamic algorithm) to generate the complete Pareto Frontier
in a pseudo-polynomial time. Then, we present an FPTAS
(Fully Polynomial Time Approximation Scheme) to generate
an approximate Pareto Frontier, based on the conversion of
the exact dynamic programming. For the proposed algorithms,
we randomly generated several instances with different ranges,
and, for each job Ji , its processing time pi and delivery time qi
are sets to be integer numbers.The results of the experiments
showed that the proposed algorithms for the considered problem are very efficient. It is clear that optimizing the maximum
lateness (Lmax ) implies to minimize implicitly the makespan
(Cmax ). Moreover, the values of and processing and delivery
times play an important role in the results (i.e., big processing
Table V: Average computing times (ms) vs delivery time
ranges
qi ranges
1-20
1-100
1-500
DP
31447
40482
47190
FPTAS
= 0.3 = 0.9,
26
5
46
7
149
15
times, small delivery times and big make the FPTAS faster
and vice versa).
In our future works, the study of the multiple-machine scheduling problems seems to be a challenging perspective in the
extension of our work.
R EFERENCES
[1] G. Sapienza, G. Brestovac, R. Grgurina, and T. Seceleanu. “On applying
multiple criteria decision analysis in embedded systems design.” Design
automation for embedded systems,20(3), Vol 20(3), pp 211–238 (2016).
[2] C. Bazgan, F. Jamain, and D. Vanderpooten. “Approximate Pareto sets
of minimal size for multi-objective optimization problems.” .Operations
Research Letters, Vol 43(1), pp 1–6 (2015).
[3] Z. Geng and J. Yuan. “Pareto optimization scheduling of family jobs
on a p-batch machine to minimize makespan and maximum lateness.”
Theoretical Computer Science, Vol 570, pp 22–29 (2015).
[4] C. He, H. Lin, and Y. Lin. “Bounded serial-batching scheduling for
minimizing maximum lateness and makespan.” Discrete Optimization,
Vol 16, pp 70–75 (2015).
[5] Y. Chen and X. Zou. “Runtime analysis of a multi-objective evolutionary
algorithm for obtaining finite approximations of Pareto fronts.” Information Sciences, Vol 262, pp 62–77 (2014).
[6] K. Florios and G. Mavrotas. “Generation of the exact Pareto set in
Multi-Objective Traveling Salesman and Set Covering Problems.” Applied
Mathematics and Computation, Vol 237, pp 1–19 (2014).
[7] Q. Feng, J. Yuan, H. Liu, and C. He. “A note on two-agent scheduling on
an unbounded parallel-batching machine with makespan and maximum
lateness objectives.” Applied Mathematical Modelling, Vol 37(10–11), pp
7071–7076 (2013).
[8] C. He, H. Lin, Y. Lin, and J. Tian. “Bicriteria scheduling on a seriesbatching machine to minimize maximum cost and makespan.” Central
European Journal of Operations Research, pp 1–10 (2013).
[9] C. He, X. M. Wang, Y. X. Lin, and Y. D. Mu. “An Improved Algorithm for
a Bicriteria Batching Scheduling Problem.” RAIRO-Operations Research,
Vol 47(1), pp1–8 (2013).
[10] M. T. Y. Sabouni and F. Jolai. “Optimal methods for batch processing
problem with makespan and maximum lateness objectives.” Applied
Mathematical Modelling, Vol 34(2), pp 314–324 (2010).
[11] C. He, Y. Lin, and J. Yuan. “Bicriteria scheduling on a batching machine
to minimize maximum lateness and makespan.” Theoretical Computer
Science, Vol 381(1-3), pp 234–240 (2007).
[12] D. Sarkar and J. M. Modak. “Pareto-optimal solutions for multi-objective
optimization of fed-batch bioreactors using nondominated sorting genetic
algorithm.” Chemical Engineering Science, Vol 60(2), pp 481–492 (2005).
[13] A. Allahverdi and T. Aldowaisan. “No-wait flowshops with bicriteria
of makespan and maximum lateness.” European Journal of Operational
Research,Vol 152(1), pp 132–147 (2004).
[14] C. M. Sil and E. C. Biscaia. “Genetic algorithm development for multiobjective optimization of batch free-radical polymerization reactors.”
Computers & chemical engineering, Vol 27(8), pp 1329–1344 (2003).
[15] P. Kumar. “A Framework for Multi-objective Optimization and Multicriteria Decision Making for Design of Electrical Drives.” Universiteit
Delft. (2008).
[16] S. Chakhar and J. Martel. “Multi-Criteria Evaluation Functions Inside
Geographical Information Systems Towards a Spatial Decision Support
System.”(2006).
7
[17] Z. Geng and J. Yuan. “Scheduling with or without precedence relations
on a serial-batch machine to minimize makespan and maximum cost.”
submitted, 2017
| 8cs.DS
|
E cient textualrepresentation of structure
Brenton Chapin
arXiv:1706.00862v1 [cs.PL] 2 Jun 2017
ABSTRACT
This paper attempts a more formal approach to the legibility of text
based programming languages, presenting, with proof, minimum
possible ways of representing structure in text interleaved with
information. This presumes that a minimalist approach is best for
purposes of human readability, data storage and transmission, and
machine evaluation.
Several proposals are given for improving the expression of interleaved hierarchical structure. For instance, a single colon can
replace a pair of brackets, and bracket types do not need to be repeated in both opening and closing symbols or words. Historic and
customary uses of punctuation symbols guided the chosen form
and nature of the improvements.
KEYWORDS
programming language design, structured programming, human
readability, syntax, notation, history, data compression, minification
ACM Reference format:
Brenton Chapin. 2016. Efficient textual representation of structure. In
Proceedings of ACM Conference, Washington, DC, USA, July 2017 (Conference’17), 11 pages.
DOI: 10.1145/nnnnnnn.nnnnnnn
1
INTRODUCTION
Information is almost always more useful when organized, and
structure is key to that. Therefore efficient and clear representation of structure is of paramount importance. Structured programming languages are only one use of structure to organize one kind
of information, source code, but they make such unprecedentedly
elaborate use of structure that they have exposed deficiencies in
our methods of expression.
The languages for programming and math developed in an evolutionary manner, with much borrowing from earlier work, and
the reasons for various decisions became buried in custom and
history. Studies on choices of characters and legibility have been
patchy, with some questions overlooked because they were thought
relatively unimportant. Pioneers of programming languages hurriedly made expedient adaptations of existing notations for similar
problems. Most heavily borrowed was mathematical notation.
With many important questions needing settling with the creation of the first programming languages, issues of symbology were
mostly passed over as unimportant and arbitrary. As Bob Bemer
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
Conference’17, Washington, DC, USA
© 2016 ACM. 978-x-xxxx-xxxx-x/YY/MM. . . $15.00
DOI: 10.1145/nnnnnnn.nnnnnnn
noted, “much documentation is lost, and it was characteristic of
the times that nobody seemed to think character sets a very important feature of computers [12].” There are many studies on the
readability and other properties of various fonts and color combinations, but when discussed in relation to source code, the term
“readability” refers more to comprehensibility [21].
Punctuation is the class of symbol most closely associated with
showy, interleaved structure. Positioning is the other major method
used to indicate structure, and among the other intended purposes,
“control characters” attempted to provide ways to position text.
Currently, Python is the most popular programming language that
relies on text positioning rather than punctuation to indicate structure. Visual programming goes further yet, replacing textual indicators of structure and flow with graphical ones.
The programming language wars are still hot today, with new
languages emerging and gaining followers. One cause of the passionate debates is the tendency of language designers to resort to
an evangelical approach to justify their choices of design elements
for which they have little compelling technical reason. Sometimes
the designers make overlarge and unsubstantiated claims [19]. For
many programming languages, one of the defining features is the
choice and usage of symbols. These choices are not modifiable by
the programmers, so that if such changes are desired, a whole new
programming language may need to be created, another factor in
the very proliferation of programming languages that the ALGOL
designers were hoping to avoid.
The ideas in this paper aim at the foundation, the symbolic representation of the structure. Structure is chosen as the crucial concept that must be addressed to improve representation. Minimalism is the chosen guide.
Too much minimalism is certainly possible, for instance by expecting people to work with information that has been minimized
by data compression techniques which transform the data into a
very compact but human unreadable form. The minification techniques of removing unnecessary whitespace and shortening variable names is another example. The target of these minimization
efforts is the representation of the structure of the source code, not
the source code itself. Further, this is not about rewriting and rearranging to place information in more efficient structures, this is
about making the representation more efficient regardless of the
structure chosen. Punctuation has always been minimal, using
smaller and less obtrusive symbols than those used to represent
the letters of a language, and the syntax of structural elements follows that pattern.
Some more items to note are splits between textual representations used for source code, versus those used for markup, as in
HTML, and data organization, as in XML and YAML. Within programming languages, there is the dot (or arrow) notation of Object
Oriented Programming and the completely different notations for
Structured Programming, such as the curly braces. Yet those splits
Conference’17, July 2017, Washington, DC, USA
seem artificial, as hierarchical structure is used in all. Many programming languages are needlessly poor at expressing data. Several of the improvements in C++11 and C++14 touch on this issue, allowing more flexible constructions of constants and initializations of arrays and objects. One intent of JSON is to bridge this
divide.
Also, these are all interleaved formats, meaning the symbols
that denote the structure are interleaved with the symbols of the
data. Goals in data storage are minimal size, and fast access. An
obvious and common method to achieve both is to exclude all complex structure from the data, using an external representation. The
disadvantage is that they require some connection, often expressed
in fixed sizes and padding, which can end up using more space
than an interleaved method. Over the years, the attention paid to
brevity has varied with the cost and availability of storage.
By 1963, the American Standard Code for Information Interchange (ASCII) was set to a fixed size of 7 bits, though at least
two variable codes, Morse code (1844) and Huffman coding (1952)
existed at the time.
One of the goals of XML was human readability. Minimalism
was thought orthogonal or possibly even antithetical to the goal
of human readability, and the resulting language ironically suffers
from excessive verbosity that obscures essentials, rendering the result less human readable. XML and COBOL show that a negative
attitude towards minimalism (”10. Terseness in XML markup is
of minimal importance. [14]”), that regarding minimalism as unrelated or even an impediment to comprehension, is not correct.
Minimalism is also central to Information Theory, in which it was
demonstrated that the crude redundancy of repeating information
over and over, is very wasteful and poor at preserving the fidelity
of data against errors in the transmission. If repetition is a poor
means of ensuring the fidelity of data, perhaps it is also a poor
means of representing structure in ways easy for humans to read.
Another demonstration of the limited usefulness of repetition is
the FAT file system, which despite allocating room for a copy of
the directories and file names, is actually one of the most fragile
and easily corrupted file systems currently in use.
Of particular note is the C programming language. So many
programming languages adopted the C syntax that they have been
tagged with a moniker of their own, the “curly-brace” languages.
Perhaps one of the reasons curly brace syntax eclipsed Pascal and
ALGOL is the use of a single character each, rather than the words
BEGIN and END, to delimit blocks. The designers of C did not restrict themselves to curly braces only, they also used square brackets, parentheses, and even angle brackets, for array indexing, lists
of function parameters, and macros respectively. Why that choice
of symbol assignment? Why not use parentheses for all, and rely
on context or some other means to distinguish between a parameter list and a block of code? If there is any doubt that it is possible
to use only parentheses, the LISP programming language is proof.
Or, why not copy FORTRAN in the use of parentheses for array
indices? One kind of answer is that in C these different kinds of
brackets serve as sigils, to distinguish between identifiers for functions, arrays, and variables. But that only begs the question of why
have sigils? And it still does not answer why any particular symbol
was chosen for a particular use.
Brenton Chapin
2 HISTORY
For answers, one must dig into the history of computation and
mathematics. In the case of C, the chain of preceding languages is
roughly B, BCPL (Basic CPL), CPL (Combined Programming Language), and finally ALGOL (Algorithmic Language). The paper on
ALGOL 58 [6] says of the choice to use square brackets to delimit
array indices, only that “subscripted variables” (the term used in
ALGOL for what today we call an array variable, or simply an array), “designate quantities which are components of multidimensional arrays” and that “The complete list of subscripts is enclosed
in the subscript brackets [].” But why did they pick square brackets? FORTRAN, the oldest programming language to achieve wide
acceptance, uses parentheses, not square brackets.
For that matter, why use any bracket at all? No one says. It
seems likely that they would rather have used actual subscripted
text, just like in mathematical notation, but early computers could
not do it. Square brackets was a notational device to indicate subscription without actually presenting the text so. Apart from computer limitations, a big problem with subscripting is that the notation doesn’t nest well, at 3 or more levels becoming too small for
the human eye to read. One can surmise from the use of the term
“subscript” that this was another borrowing, from linear algebra in
which a matrix is denoted with square brackets. And indeed the
original name of ALGOL 58, is International Algebraic Language.
The only deviation in the use of square brackets for array indexes
from ALGOL to C was BCPL, which among the many simplifications of CPL it introduced, attempted to repurpose square brackets
for code blocks, using only pointer arithmetic to access array elements [8].
ASCII codified the glyphs used for nearly all programming languages. A notable exception is APL, which makes use of mathematical symbols, mainly from Set Theory and Vector calculus, that
were not put in ASCII [7]. Unlike EBCDIC, ASCII at least organized the alphabet into a contiguous block. But the exact set of
punctuation symbols is unclear, ranging from all symbols that are
not letters, numbers, and control characters, to only those used to
clarify the structure and meaning of sentences and text. There are
no formal, ordered, centuries old lists of punctuation symbols.
The ASCII ordering and choice of punctuation is derived from
the QWERTY keyboard layout, which dates to the late 19th century. The notion that QWERTY was deliberately arranged to slow
typists down is a popular but wrong myth [23]. Morse Code and
many other factors were considered, and over the years small changes
have been made to accommodate new uses. For instance, “shift-2”
is the double quote mark on many older keyboards, but today is
‘@’ on most keyboards.
We could go further back, and ask why mathematical notation
uses parentheses for functions, and square brackets for matrices.
Why is y = fx the customary, canonical expression for a function,
and why in particular the use of parentheses to bracket the independent variable x? In A History of Mathematical Notations [3],
Cajori credits Euler (1707-1783) with the first use of parentheses
to bracket the variable of a function, in a 1734 paper. That paper is
E44 [1] in the numbering scheme created to refer to Euler’s works.
However, examining E44 and several others of Euler’s papers, one
finds no such use of parentheses, and the exact phrase and formula
Efficient textual representation of structure
Cajori quoted is not present. Euler uses parentheses to group parts
of equations, but not to separate function names and variables. Euler’s notation is y = fx, and it is up to the reader to understand
that x and y are variables, and f is a function.
Note also the choice of the letter f because it is the first letter of
the word “function”, a custom followed in many places, such as the
decision in FORTRAN to use the first letter of a variable name to
indicate integer (name begins with “I” for integer, through “Z”) or
floating point (name begins with “A” through “H”). This desire to
match functionality to the first letter of an appropriate term was
taken to extremes, so that more than one early game employed
a lucky placement of keys on the QWERTY keyboard,’W’, ’E’, ’S’,
plus ’3’, to refer to west, east, south, and north respectively.
By 1837, in a major work on Number Theory which is regarded
as also an important paper on the modern definition of a function, Dirichlet (1805-1859) used parentheses around the independent variable [2]. But why did mathematicians pick those symbols,
that format? They too engaged in expedience, adopting the idea of
parentheses from still earlier scholars. Mathematical notation has
a long evolutionary history, and while fascinating, the main point
here is that many choices of symbols and syntax were made long
before any possible use in programming languages was conceived.
While 1837 is also the year that Babbage proposed the Analytical
Engine, arguably the first computer, functioning computation machinery would not be built until many years later. Therefore symbols and syntax certainly could not have been chosen based on
experiences in computer programming.
That was about as far as the early pioneers went in exploring
questions of how best to symbolize code and data. None of the
terms and areas of study, not semiotics, symbology, linguistics,
grammar, lexicology, punctuation, readability, typography, legibility, notation, expressiveness, or rubrication, quite address these
questions. Studies of notation and syntax get the closest, but even
there syntax is confined to issues of context.
Most programming languages use a hierarchical structure to organize code. Possibly the earliest and simplest formally specified
language for expressing hierarchy is Dyck Language. Object Oriented Programming and Functional Programming did not abandon
this fundamental organization, they only added to it. Declarative
programming, as represented in Prolog and SQL, at first glance
seems not to need much structure. A point of confusion is order vs
structure vs hierarchy. Declarative programming needs structure,
but not order and not necessarily hierarchy. Hierarchic structure,
of programs and data, can be more efficiently represented with several changes.
The advent of markup languages revived interest in hierarchical data storage, which was introduced in the 1960s, before the
relational database model. No longer were interleaving structural
symbols just for programs, they were harnessed to organize data.
Traditionally, data has been organized into fixed size elements so
that no symbols need be reserved for explicit denotation of structure, and, even more importantly, so that random access is quick,
taking O 1 time to retrieve any one element. This is also true of the
pre-computer era, which used tables extensively, carefully lining
up columns to aid the human eye. Where one-size-fits-all is inadequate, the expedient method used is to have a small fixed size field
Conference’17, July 2017, Washington, DC, USA
to hold a value for the size of a variable length field. Packet networking is an example of this organization of data. The roughly
analogous method in writing is the technique of employing any
of a variety of superscripted symbols such as an asterisk, *, or a
dagger, y, to indicate there is a footnote.
XML and HTML are the most well known of these markup languages, and like programming languages, their history is also evolutionary. Both trace back to Standard Generalized Markup Language (SGML) which was standardized in 1986, predating the World
Wide Web. Like so many other decisions in languages, the creators
of the Web seized upon SGML out of expediency. SGML in turn
descends from GML, an IBM effort to manage technical documentation and data, based upon ideas first articulated circa 1966 [16].
But as many have complained over the years, these markup
languages have undesirable features, and among the biggest is extreme verbosity. The rules they force upon users, to come closer to
the goal of “human readability”, often have the opposite effect. On
the scales of minimalism, XML and relatives are extremely poor because their representations are highly redundant. Not only must
brackets be balanced in “proper” HTML and XML, but the matching tags must repeat the tag name. Why did the designers do it?
Ironically, those rules have done much to add clutter and thereby
reduce the human readability that was their intended goal. YAML
(YAML Ain’t Markup Language) was motivated in part by recognition that XML is burdened with design constraints that have little
purpose in data serialization [17]. Lightweight markup languages
such as Markdown are an acknowledgment that the human readability of HTML could be better.
Most popular programming languages are poor at expressing
data. Here are some examples to illustrate this. A list of the first 10
chemical elements can be encoded in a JavaScript array like this:
c o n s t CE = [ ” ? ” , ”H” , ” He ” , ” L i ” , ” Be ” , ” B ” ,
”C ” , ” N ” , ” O ” , ” F ” , ” Ne ” ] ;
A simple trick yields a much cleaner representation:
c o n s t CE = ” ? H He L i Be B C N O F Ne ”
. split (” ”);
But this is the very sort of trick that makes programming needlessly difficult for professional programmers unfamiliar with the
arcana of a particular language.
One problem is that the default, unquoted meaning of an alphanumeric sequence is to treat it as the name of a variable. The
double quote mark changes the mode, but that mode has no support for structural elements, so only a simple string can be encoded.
The programmer is forced to change modes over and over, entering
string mode to give a short string, leaving string mode to impart a
tiny amount of structure, then entering string mode again to give
the next string. Or the programmer can use a clever trick such as
the split function, or create a function to parse a string into a
complicated object, or even employ a library such as YAML.
Another example, of a family tree, in Python:
c l a s s t n : # t n means ” t r e e node ”
def
init
( s e l f , name , c h i l d =None ) :
i f c h i l d ==None : s e l f . c = [ ]
else : self . c = child
Conference’17, July 2017, Washington, DC, USA
s e l f . n = name
familytree =
[ tn ( ” grandmother ” ,
[ tn ( ” older uncle ” ,
[ tn ( ” o l d e st 1 st cousin ” ) ,
t n ( ” 2 nd o l d e s t 1 s t c o u s i n ” ) ] ) ,
tn ( ” fat her ” ,
[ tn ( ” older s i s t e r ” ,
[ tn ( ” niece ” ) ,
t n ( ” nephew ” ) ] ) ,
t n ( ” you ” ,
[ t n ( ” son ” ,
[ tn ( ” granddaughter ” ) ] ) ,
tn ( ” daughter ” ,
[ tn ( ” grandson ” ) ] ) ] ) ,
t n ( ” younger b r o t h e r ” ) ] ) ,
...
This terrible encoding is littered with alternating brackets of 2
kinds, as well as double quote marks and commas. This shows
that Python can be even worse than LISP, for those who thought
Python’s use of indentation lead to clean code in all cases, and that
LISP had too many parentheses. To get clean looking code, the expert programmer resorts to using functions to read a simple string
(which may be a data file) into a complicated object. Employing a
data serialization library such as YAML, is a common method of
handling this issue. Should it be the preferred method? Shouldn’t
programming languages be able to do better with their native syntax? After all, native handling of regular expressions is what made
Perl popular. Improvements in the representation of structure are
applicable both to coding and to data representation.
3
ELIMINATING RUNS OF BRACKETS
The first change addresses a problem most languages have, but
which is perhaps most obvious in LISP, and for which it has been
criticized in the “backronym” of Lots of Idiotic Spurious Parentheses. Often, brackets cluster, as several structures all start or end
simultaneously. They can add to the visual clutter without adding
to the ease of comprehension.
There are many solutions to this problem, among them operator
precedence, and postfix notation, also known as Reverse Polish notation, first conceived in 1924[4]. A limitation of these Polish notations is that to make brackets unnecessary, the number of operands
must be fixed, an inflexibility that is insufficiently general for the
structures used in programming.
A popular short cut is use of context and knowledge about the
permitted or sensible content of subtrees. For instance, in HTML
the paragraph indicator, <p>, cannot be nested. This is often used
to omit the matching closing bracket, </p>, when the next structure is another paragraph, or something else that cannot be inside
a paragraph, such as a header. Such omissions are not officially
sanctioned in HTML, but are so popular that web browsers had to
support them anyway. Obvious problems with this approach are
that knowledge of every exception to the rules for indicating the
nesting may be very large, and may change.
Brenton Chapin
The approach taken in Perl 6 is to allow all kinds of shortcuts
that do not greatly complicate the parser. Compared to Perl 5,
some brackets are no longer required. In particular, the parentheses of the if and for statements are optional [18]. Effectively, this
change is a recognition that if and for are enough by themselves
to indicate structure, that they are in fact now part of the set of
symbols used to denote structure.
One could employ 2 sets of brackets, perhaps () and [], in a
scheme in which a closing bracket closes its matching opening
bracket, and all the open brackets of the other kind in between.
For example, [a [b]] becomes [a (b], [d [e [f]]] becomes
(d [e [f). This idea can work in the other direction. [[g] h]
becomes [g) h]. It even works in both directions at once, with
((j)(k)) becoming [j)(k]. However, the best this idea can do
for ((m)) is [(m].
An issue is that 2 more symbols are needed. We can employ
only one more symbol, eliminating only one of the excess opening or closing brackets, and still clean up most of clutter. Call a 3
symbol system that eliminates excess closing brackets a “closing 3”,
and a 3 symbol system that eliminates excess opening brackets an
“opening 3”. Using colon, :, for this 3rd symbol in a closing 3 system, because that approximately matches the traditional use of the
colon in written natural languages, changes (a (b)) into (a : b).
((m)) becomes (:m), (((n))) becomes (::n), and ((j)(k)) becomes ((j):k). Additionally, the brackets are still balanced, with
equal numbers of opening and closing brackets in all the systems.
For a slightly larger example, consider this Ackermann function,
from the classic textbook Structure and Interpretation of Computer
Programs, exercise 1.8 [10]:
( d e f i n e (A x y )
( cond ( ( = y 0 )
((= x 0)
((= y 1)
( e l s e (A
0)
(∗ 2 y ))
2)
(
x 1)
(A x (
y 1))))))
Employing a closing 3 system as suggested above, gives this:
( d e f i n e (A x y )
: cond ( ( = y 0 )
((= x 0)
((= y 1)
: else :A
0)
:∗ 2 y)
2)
(
x 1)
:A x :
y 1)
6 colons have replaced 6 opening brackets. The 6 matching closing brackets have been removed. Indeed, there is never a need for
multiple adjacent closing brackets, as proven next.
Theorem 3.1. Given a sequence S of arbitrary symbols over an
alphabet A in which 2 symbols, an “opening” and a “closing” symbol,
are reserved to denote hierarchy in a format that interleaves data
and structure, and S is properly balanced, the hierarchy can always
be represented in a system with 3 reserved symbols in which there are
no runs (sequences of length 2 or greater) of the closing symbol.
Proof. WLOG, let ‘(’ and ‘)’, the parentheses, represent the
opening and closing symbols in both systems, and let ‘:’, the colon,
Efficient textual representation of structure
represent the 3rd symbol in the 3 symbol system. To allow elimination of all runs of 2 or more closing symbols, assign ‘:’ the same
meaning as ‘(’, the opening of a subtree, except that the matching
closing symbol for ‘:’ is an already necessary ‘)’ that matches an
existing ‘’ which precedes the ’:’.
Then, instances of the sequence “( s1 ( s2 ))” in which s1
and s2 are arbitrary sequences which may include balanced occurrences of ‘(’ and ‘)’ and ‘:’, may be replaced with “( s1 : s2
)”.
The replacement symbols are sufficient to represent all the relationships. The symbols still indicate that s1 is the parent of s2 ,
preserve all relationships s1 and s2 have with all other sequences
before and after because none of them need change and no additional context is needed, and preserve all relationships contained
within and between s1 and s2 also because none of them change,
nor add any contextual dependencies.
This replacement can be applied repeatedly, to reduce any number of adjacent closing brackets to 1 closing bracket. Each replacement preserves the property of balance for all remaining parentheses, as exactly one pair of matched parentheses is replaced with a
single colon.
The corollary that no runs of the opening bracket are needed in
an opening 3 system, is obvious.
A pushdown automaton can easily transform from an opening
3 to a 2, or from a 2 to a closing 3, if the data is processed in reverse order. Of course, a pushdown automaton can easily reverse
a string. In practice, the C++ style of comment delimited by a 2
slashes takes more work to detect from back to front. Nor can the
start of a C style comment be simply determined working from
back to front, because “/*” can be within a comment.
A natural question is why not use a 4 symbol system, as originally outlined above with the 2 sets of bracket symbols, and eliminate all runs of opening and closing brackets? Simply put, the
additional savings is not significant, as can be seen in that it is no
help at all on the examples of ((m)) and (((n))).
As to why, it is not possible to employ any finite set of symbols to represent infinitely many numbers with just 1 symbol each,
no matter what form the representation takes. If the representation takes the form of n opening brackets followed by n closing
brackets, all of one kind of bracket can be collapsed, because n is
preserved in the other. If both are collapsed, then n must be represented some other way. That is why the idea of using 2 sets of
brackets does not work to reduce all runs of opening and of closing
brackets to 1.
Thus we see that the idea of replacing each run of closing brackets with a single closing bracket is really the removal of a redundancy, the redundancy of specifying the depth twice, first with
opening brackets, then with an equal number of closing brackets.
That redundancy is no longer available to remove once one kind
of bracket has been reduced.
The 3 symbol system need not be exclusive, can mix with 2 symbol usage as in (a (b : c)). In practice, in coding it will likely be
preferable to use the 3rd symbol only for subtrees that are known
in advance to be the terminal child. For other uses, such as minification of JavaScript or JSON, may want to use the 3rd symbol
everywhere possible.
Conference’17, July 2017, Washington, DC, USA
Removing the redundancies of the 2 symbol system can be of
some value in data compression. Since the amount of information
encoded is the same, an ideal data compression algorithm should
produce the same size compressed output whether a 2 or a 3 symbol system is used. In practice, the output sizes vary, sometimes
better for the 3 symbol system, and sometimes worse. To better test
whether the more efficient representation helps with data compression, can try a much larger example. Biologists have organized millions of species into a Tree of Life [28], using Newick format [11],
an interleaved hierarchical format. Tests upon grafted solution.tre
from version 9.1, the file with the highest percentage of interleaved
structural symbols relative to data, containing 100,734 parentheses in 721,324 characters total, show an “opening 3” system does
reduce size even after compression.
compression
none
gzip
bzip2
xz
system
original 2 symbol
721,324
250,142
218,717
211,812
opening 3
690,077
241,169
213,341
203,724
A final note about whether to prefer an opening 3 or a closing
3 system. The closing 3 is the better fit with our customs. For instance, in curly brace languages, the name of an array is given before the index of the desired element. It is arr[39] not [39]arr. It
is the same with function names and parameters– the name comes
first.
4 UNIVERSAL BRACKET
A sequence such as “[x(y]z)” in which 2 different sets of brackets are interwoven, is almost always an error, not valid in any
mainstream language. An analogous sequence in HTML could be
“<b>x<i>y</b>z</i>”, which is not valid, even though its meaning can in this case make sense. The HTML specification calls this
“misnesting”. This invalidity is used in an ad hoc fashion to reduce some of HTML’s redundancy. A common case is the closing of an outer structure that implies an inner structure must also
close, as in this example: “<tr>x<td>y</tr>”. Some omissions require knowledge that some structure is not allowed. For instance,
“<p><p></p></p>” is not valid because the ‘p’ element (p for paragraph) can’t be the direct child of another ‘p’ element. Therefore
“<p>x<p>” always implies a closing tag preceding the 2nd opening
tag: “<p>x</p><p>”. This usage is acknowledged in HTML5, but
still recommended against: “..the closing tag is considered optional.
Never rely on this. It might produce unexpected results and/or errors if you forget the end tag.” [25]
A combination opening and closing tag in one, called a “selfclosing” tag, is meant for an empty element, and has been in XML
from the start [14]. As of version 5, HTML has adopted a variation
of this idea. The XML self-closing tag requires a slash character immediately before the closing angle bracket. In HTML5, 15 element
types were singled out as making sense only as empty (void), and
HTML does not use or permit the penultimate slash character in
those tags.
Another solution to some of HTML’s verbosity is to omit the
name from the end tags, using only “</>”, which works fine since
misnesting is not allowed or often sensible anyway. SGML has
Conference’17, July 2017, Washington, DC, USA
this feature in its SHORTTAG constructs, calling it the empty end
tag. But HTML does not allow it. This idea of a universal closing
bracket or, alternatively, a universal opening bracket, can be employed in any language containing 2 or more sets of bracket symbols and in which interweaving is invalid. It eliminates misnesting,
as interweaving is no longer possible. And it reduces the alphabet
size.
If we choose the square bracket for the universal closing symbol, then a sequence such as “(x[y]z)” could become “(x[y]z]”,
and the closing parenthesis symbol would be unused, and could
be repurposed. (Note that this change does not reduce the number
of closing brackets, there are still 2 in the example. It reduces the
required size of the alphabet.)
There can still be a need for other closing characters, such as
an “unindent” invisible control character. The universal closing
bracket could still be used for that, but would want it to be invisible
in that case.
Converting back and forth between a representation that uses
closing brackets that match the opening brackets, and a representation that uses only a universal closing bracket is easily done with
a pushdown automaton. The type of the node is preserved in the
choice of opening bracket, and having the type repeated with a
matching closing bracket is merely redundant.
Having established that a universal bracket symbol is workable,
several more questions naturally arise. Does it make code easier
to understand, more human readable? Many have expressed the
sentiment that requiring a closing tag to repeat the type given in
the opening tag helps prevent human mistakes, and is therefore
good. The issue is confused by the practice of entirely omitting
tags in specific situation. With a means of representing structure
that is not so tiresomely redundant, these ugly short cuts can be
made unnecessary.
5
TYPES FOR NODES
Often the representational capability of a hierarchical structure is
enhanced by adding some means of representing different kinds of
children. An example is the “red–black tree” in which tree nodes
have been assigned an additional property, a color. This can be
and is often done independently of the structure, by means of an
additional data item. Another very popular method is sigils in the
form of different kinds of brackets. ASCII has 3 sets of symbols
meant solely for brackets: the parenthesis, square bracket, and
curly braces. One more set, the angle brackets, doubles as the mathematical symbols “greater than” and “less than”, and for that reason
was used gingerly. Further sets can be contrived, for instance ‘n’
and ‘/’, and for that matter of course any two arbitrary characters
could be chosen to serve as brackets. The obvious complaint is that
4 sets is far too few. Even if a dozen plus from Unicode are added,
it still isn’t enough.
In any case, programming language designers used all the ASCII
symbols meant for brackets early on. The curly brace languages
employ curly braces to denote blocks of code, square brackets to
denote array indices, and parentheses for parameter lists. The dual
purpose symbols for angle brackets did not go unused as brackets,
being employed in the C preprocessor, and later in the markup
languages SGML, XML, and HTML.
Brenton Chapin
These SGML markup languages expanded the number of bracket
types infinitely, by allowing multiple character brackets. Although
that solves the problems caused by finite quantities of different
bracket symbols, the means and requirements chosen add greatly
to the verbosity, a common criticism often expressed in abuse of
the rules rather than in words. It is possible that the desire for a
visual match between the opening and closing bracket led to the
SGML requirement that both the opening and closing brackets contain copies of the string employed to give them a type, despite the
obvious redundancy.
An efficient way is to designate one of the bracket sets as ”typed”,
the start of a multicharacter bracket. That allows the other brackets
to remain bare to be used same as traditionally used in programming languages, and still allows infinite bracket types. Which symbol is best employed, and where should it be positioned? Between
bracket and name, or after the name? Or, should it be a combined
symbol, a bracket that indicates a name is to follow, since there
is more than one kind of bracket available? Possibly the most efficient use of existing symbols is to keep parentheses as is, bare,
and designate the square bracket or curly brace as the indicator
for a child structure with a type, with the name to follow, similar
to HTML.
Another method is to reserve a symbol to indicate a type name
only, no structure. ‘$’ is often used similarly.
Whichever method is chosen to indicate the start of a type name,
how is the name to be ended? The name could be just like variable
names in programming languages, with only letters and numbers
(and the underscore character) allowed in the name so that any
other symbol, such as the space character, automatically ends the
name. The method of using a special symbol, as done with the
closing angle bracket of HTML, is also workable.
But the designers of HTML did not let tag names be only names.
They crammed additional structured information into “attributes”.
An example is “<ul class="vs1" id="item1"> content </ul>”.
This information could have been in the same system, for instance
something like “<ul> <attr> class = "vs1" id = "item1"
</attr> content </ul>”, or even “<ul> <attr> <nm>class </nm>
<val> vs1 </val> <nm> id </nm> <val> item1 </val> </attr>
content </ul>”. The only purposes this alternate subsystem really
serves is visual distinction and less verbosity, though it’s claimed
to maintain the distinction between data and metadata. HTML has
evolved towards lighter use of attributes, moving much formatting
information from the tags to CSS, where it is also less redundant.
6 REPRESENTING SIBLINGS AND COUSINS
The list is well known and has a long history. Each item in a list
can be considered a sibling of each other item. Traditionally, each
item is on its own line, or is separated by a comma. LISP means
“LISt Processor”, and is built around the idea of making lists the
fundamental building block with which to organize both data and
code. Comma Separated Values (CSV) notation [20] is a simple
data format based on one list with items separated by, of course,
commas. One of the most notorious departures from the use of
commas is multidimensional arrays in C, in which the syntax to
access an element at index x;y is not x;y, it is xy.
Efficient textual representation of structure
The idea of separating items in a list with a single symbol (or
word) seems simple, but turns out to have several surprisingly
tricky issues.
Consider how to represent a list in a language that does not have
any symbol analogous to the comma, Dyck Language interleaved
with data. How is the sibling relationship expressed? (First, note
the convention is to place the parent before the child, as in p(c),
although the opposite, (c)p is just as expressive.) One way is to
wrap brackets around each individual data item. Then the number
of brackets needed to represent a relationship must be increased
by 1 for all depths, so that (a)(b) means a and b are siblings, and
((c))((d)) means c and d are 1st cousins. A 2x2 array would be
((p)(q))((r)(s)). Although it works, it is far more verbose. Additionally it spoils the abbreviation of allowing siblings to be separated by a child, as in a(e)b, which must instead be (a(e))(b).
So, a better way is to always separate siblings with a child, using
a null child if the older sibling has no children, as in a()b. Then a
2x2 array can be represented with (p()q)(r()s).
Expanding to cousins is still a problem. With the addition of the
comma as a sibling separator, (p()q)(r()s) becomes (p,q)(r,s).
The sequence still has a “)(”, which the comma does not help reduce. An obvious extension is to introduce another symbol, say
semicolon, to separate 1st cousins. Then the sequence can become
(p,q;r,s).
What to do for 2nd cousins? Just add brackets to the semicolon, as in );(? Or employ yet another symbol to replace ))((?
How many symbols should be so employed? The ASCII committee
settled on 4, ASCII characters 28 through 31, though 8 were proposed [12]. They were called Information Separators [9]. We can
do better than that.
There are several issues with having 2 or more Information Separators that merit careful consideration.
First, consider the sequence p(q,r;s). q and r are siblings to
each other, and descendants of p, and s is 1st cousin to q and r.
There are several different more precise meanings this could have.
The semicolon can be given higher precedence than the brackets,
that is, all three of q, r, and s are children of p. In that case, this
particular sequence is invalid, because r and s cannot be children
of p and 1st cousins to each other. All children of the same parent
must be siblings.
Another interpretation is to allow a single opening bracket to
separate a variable number of generations instead of always one
generation. Then, since grandchildren of p can be 1st cousins to
one another, all 3 of q, r, and s must be grandchildren of p. But
this idea has the big disadvantage of adding context dependency
to the grammar. Whether q is a child or a grandchild of p cannot
be known until all the characters between the opening and closing
brackets are scanned. If a semicolon is found on the same level as
q, then q is a grandchild of p. If there are even deeper separators,
q is a great grandchild or even more distant descendant of p. If
none are found, then q is a child of p.
Best is to consider the semicolon as a combined open and close
bracket, )(, having the same precedence as any other bracket. In
that case, sis not a descendant of p, sis a nephew of p. That meaning does not add context. This does have more invalid strings, for
instance the simple sequence r;s is invalid because the brackets
are not balanced.
Conference’17, July 2017, Washington, DC, USA
Second, consider how to combine separators with colons. The
colon is a bracket, and should have the same precedence. Then a
sequence such as (p:q;r) means that p is parent to q, and not
parent or sibling to r. p is uncle to r, q is 1st cousin to r, and
r’s parent is null. Perhaps the easiest way to see this is to reverse
the colon transform to get (p(q;r)), then reverse the separator
transform to get (p(q)(r)). If p and rare supposed to be siblings,
and q a child of p, the correct way to represent that is not to use
semicolon or colon, it is p(q)r.
The 2 transforms, colon and separator, are mostly complementary, but in some cases can compete to reduce the same redundancies. The following table shows the results of transforming each of
the 14 Dyck words of length 8 (replacing ][ with a comma rather
than a semicolon, for greater visual clarity.)
Dyck word colon
separator both
1 [[[[]]]]
[:::]
[[[[]]]]
[:::]
2 [][[[]]]
[][::]
[,[[]]]
[,::]
3 [[][[]]]
[[]::]
[[,[]]]
[:,:]
4 [[]][[]]
[:][:]
[[],[]]
[[],:]
5 [[[][]]]
[:[]:]
[[[,]]]
[::,]
6 [[[]][]]
[[:]:]
[[[],]]
[:[],]
7 [[[]]][]
[::][]
[[[]],]
[[:],]
8 [][[][]]
[][[]:]
[,[,]]
[,:,]
9 [][[]][]
[][:][]
[,[],]
[,[],]
10 [[][][]]
[[][]:]
[[,,]]
[:,,]
11 [[][]][]
[[]:][]
[[,],]
[[,],]
12 [[]][][]
[:][][]
[[],,]
[[],,]
13 [][][[]]
[][][:]
[,,[]]
[,,:]
14 [][][][]
[][][][] [,,,]
[,,,]
The last column shows the result of applying the semicolon transform, followed by the colon transform. Applied second, the transform to colon can be blind to the presence of any separators, and
be correct and achieve maximum reduction. A separator acts as a
bridge, so that a colon can start a list, a natural looking use, rather
than opening the last item of a list.
If the separator transform is second, then to achieve maximum
reduction, as well as a correct transformation, it has to be done
with awareness of colons. A colon may be opening the last item
in a list, and it can be moved to the head. [[]:] can become [:,]
by replacing ]:, which is an open close pair of brackets, with a
separator, and then, replacing the opening bracket of the previous
item in the list with a colon. This can be repeated until the colon
has migrated to the front of the list. If the separator transform is
done blindly on a sequence with colons, it can be incorrect. [[]][]
is [:][] but then replacing the ][ with a separator gives [:,],
which is not correct. Correct is [[],]. Undoing [:,] shows that
sequence is actually [[][]], a list of 2 items.
Applying both transforms to the Ackermann function given earlier replaces a total of 11 bracket pairs with either a single separator
(the comma was used in this example) or a single colon:
( define :A x
cond : ( = y
(= x
(= y
else
y,
0)
0,
1)
:A
0,
∗ 2 y) ,
2,
:
x 1,
A x :
y 1)
Conference’17, July 2017, Washington, DC, USA
Third, what of types? Should the separated items be the same
types? The traditional meaning of a comma is as a separator only,
of untyped data.
Or, should text adjacent to a comma be interpreted as a type
name? A way to resolve that question is to provide another means
to add a type if desired, and let separators remain separators only.
For example, as mentioned in the section on types, the ‘$’ character could be used to indicate an alphanumeric sequence is a type
name. Deeper separators would need a list of types, or restrictions
on elements for which they can specify a new type, and while notations for that can of course be invented, there is little point when
opening brackets can accomplish that with reasonable efficiency
and without any additional rules.
Fourth, there are different potential meanings for runs of a separator symbol. 2 adjacent semicolons could mean that there is an
empty element in the middle, like for(;;i++) in C. Or, it could
mean that the separation between the data elements on either side
is deeper, that is, they are 2nd cousins instead of 1st cousins. What
should 2 semicolons mean, )()( or ))((? The former is the more
widely used meaning. The latter is accomplished by the limited
method of having more Information Separator symbols, which cannot neatly handle great depths. It seems useful to have clear and
concise ways to express either meaning. One way to do this is to
have 2 Information Separators, one for siblings and one for cousins.
The wrinkle is that repetition of these symbols would have the 2
different meanings. n of the sibling separator can mean there are
n 1 siblings who have been omitted from the list, while n of the
cousin separator can mean the cousins are nth cousins, being 1st
cousins only when n = 1. This approach, combined with an efficient way to express quantities, discussed next, can express both
meanings.
However, another way is not to use the system for expressing
quantities, and then assign different meanings to those quantities,
but to use typing. A semicolon could be followed by an integer
to indicate the depth of the divide, eg. “;3” means the adjacent
elements are 3rd cousins. Then a run of n semicolons can mean
that there are n 1 1st cousins in the middle, same as a run of n
commas means n 1 middle siblings. This makes it slightly harder
to support typed separators, but of course it can still be done, That
point is moot if sticking with the traditional meaning of separators
being typeless.
A minor matter is that separators have an inherent off-by-one
issue. A comma separated list usually contains one fewer commas
than data items. Often, specifications allow a meaningless trailing
comma to be present, for the convenience of programmers.
A big reason to support efficient representation of a cousin relationship and even reserve symbols especially for it rather than rely
on brackets is that it is a natural way to map multidimensional arrays to a hierarchical structure. Another reason is that people are
familiar with and like separators.
7
EFFICIENT REPRESENTATION OF
ARBITRARY QUANTITIES
Infinitely many numbers cannot be represented with single symbols from a finite set of symbols.
Brenton Chapin
Though we can’t collapse arbitrary quantities to single symbols,
we can however do better than using n symbols to represent n,
by employing the same principle used in the Arabic numbering
system that replaced unary numbering systems such as the Roman
one and hash marks. All this is well known, as is that a binary
numbering system has the minimum number of symbols needed
to represent quantities of n with log n symbols.
Can we do even better than log n, represent any arbitrary quantity of size n with even fewer symbols? No. For this question, the
Pigeonhole principle applies. As in data compression, to be able
to represent some quantities of amount n with fewer than log n
symbols (from a finite set of symbols), other quantities must be
represented with more than log n symbols. When the amounts are
averaged over all quantities n, the size is log n, or greater.
Numbering systems can be employed to represent structure. Rather
than come up with more and more symbols to represent greater
and greater quantities, as the ASCII committee did with their 4
separator symbols, can employ 2 symbols in a binary code.
Obviously any one symbol which may be repeated can be made
one member of a set of 2 symbols to be used in a binary encoding. But if there are many symbols which may be repeated, finding
enough symbols becomes a problem.
Since quantities are so useful, and unused symbols so precious, a
better idea is to reserve 2 symbols for a binary code for quantities
only, for any other symbol that may be repeated. For example,
instead of using 2 kinds of open bracket symbol in a binary code
as in something like [(([ to represent 9 open brackets, have 1001(
mean 9 open brackets, 1101* mean 13 asterisks, and so on.
Still better is to use an escape character and a decimal representation. The backslash can be used for this, as the only backslash
escape sequence that uses a number is n0, to mean the NULL character, ASCII 0. Then 9 open brackets can be represented with n9(.
One desirable additional symbol to allow is the minus sign, for negative quantities. If only integers are allowed, then there is no need
to overload the meaning of an escaped period for a decimal point
character.
This sort of representation is the well known idea of run-length
encoding (RLE) [5]. RLE is simple and easy, even relatively easy
for a person to understand without computer aid.
Of course there is the minor problem that the numeric symbols
themselves cannot be the object of a RLE escape sequence. There
are several easy ways to resolve that issue. Easiest is to simply not
support repetition of the digit characters, forcing the use of the
traditional method if repetition of a digit is wanted. Perhaps next
easiest is to employ a terminal symbol. To keep the representation
one character shorter, the terminal symbol can be optional, used
only if needed to remove ambiguity.
But RLE is very limited in what it can express. That keeps it
dirt simple and easy to read, but perhaps more expressiveness is
desirable, for such uses as repeating patterns, not only single characters. One simple example of such a sequence is the CR/LF pair.
With a trivial amount of additional syntax, it is possible to efficiently encode repeating patterns. A further use is as a repetition
of an escape. Suppose one has a string consisting of many, perhaps over half, of characters that must be escaped. One traditional
method is to inflate by up to double the quantity of characters by
preceding each special character with an escape character. That
Efficient textual representation of structure
can get difficult for a programmer to read, as seen in Perl’s regular expressions. A quantity that can be applied to indicate how
many characters are to be escaped can supersede the traditional escape character method. This notion is fairly obvious and has been
proposed on a number of occasions, for instance by Rivest in his
draft for S-expressions [15], for what he called “Verbatim representation”, and with Hollerith constants in FORTRAN 66 [13]. Perl’s
regular expressions has a similar mechanism.
While it is trivial to extend run length encoding to handle repeating patterns, there are still many other highly redundant strings
that this extended RLE cannot encode, yet are simple to describe.
The question is how far to go, how much complexity is really useful and can still be easily encoded? And, would it still be human
readable?
Perhaps an efficient way to represent “))(())((” and larger
combinations is also desirable? To encode such things, more is
required. A simple idea is to support the encoding of lists of quantities, a vector, rather than a single quantity. The escape character
can be employed as a separator. Then what is needed is agreement
on the meanings to assign to the multiple quantities. For example,
to encode 5 repetitions of a string of length 4, “abcd”, should it be
“n4n5abcd” or “n5n4abcd” or something else?
But if a vector of quantities is such a good idea, why not a tree
of quantities? Takes only 2 symbols to represent the structure of
a tree. However, the additional complexity is almost certainly too
much to remain human readable, and there’s the question of what
uses could we make of a tree of quantities?
One use for a vector of quantities is for the sizes of the dimensions of a multidimensional array. Such a usage creeps into the domain of external representation of structure. The interleaving can
be reduced to a single character used as a separator, or removed entirely. For instance, a 2x3 array with variable sized elements could
be notated as n2n3n? 1a,1b,1c,2a,2b,2c, using the same separator symbol every time, with the division between 1c and 2a known
to be deeper than the rest only because that info was given in the
vector of quantities. Or that 2x3 array with fixed sized elements
could be notated as n2n3n2 1a1b1c2a2b2c.
If means to represent something analogous to a Hollerith constant are provided, some probably will use it for very complicated
objects. Just serialize the data, and use the total length of the resulting string as the size. Supporting runs of the same symbol, and
blocks of data analogous to Hollerith constants, provides enough
for further elaboration if desired, while keeping the notation simpler.
We get away with unary representations, because we stick to
small and simple structures. If we seldom count higher than 3 or 5,
and almost never higher than 10, tally marks work fine. A check of
the Firefox source code reveals that only a few programs reached
a nesting depth of 15, with most never exceeding 10, so RLE for
opening brackets and colons, and quantities to indicate the depths
of separators are not going to remove much clutter. But perhaps
flatter structuring has been chosen to avoid the clutter that would
result from deeper nestings. And block escapes are still a viable
use of quantities.
Conference’17, July 2017, Washington, DC, USA
8 REPRESENTING STRUCTURE WITH
POSITIONING
The only ASCII control characters still really useful are the 2 for
indicating a new line of text. Next most used is tab, which has
ambiguous meaning and is easily and often replaced with spaces,
except in a few special cases such as Makefiles. The rest of the
ASCII control characters are very seldom seen, and when present,
modern applications may simply ignore their meanings [27]. Of
the 132,231 text files in the Firefox 50 source code, just 121 have
ASCII control characters other than the 3 most common: LF, tab,
and CR. A mere 5 files use ANSI escape sequences, which start with
the Escape character (ctrl-[, ASCII 27), and that only to set text
colors.
ASCII’s minimal means of positioning text is sufficient but not
efficient or neat. One of the worst inefficiencies is the very repetitive use of spaces to indent lines of text. Some ANSI escape sequences address this issue, but not well. The VT100 series text
terminals became popular in large part because they adopted and
extended the ANSI escape sequences. Yet they have not been much
used outside of terminal control. They did not grow beyond that
niche to become common within text files. Colored text is the ANSI
escape sequence most used, yet it is rare. One of the most common uses of colored text, highlighting of source code, does not use
ANSI at all, even though editing may still be done in a terminal
that supports ANSI. Rather, text editors parse the source code being edited to compute which colors to assign, updating in real time
as the user makes changes. HTML and CSS can specify colors directly, and are not limited to a tiny 16 color palette. That and word
processor options have become the way to set text and other colors in documents. ASCII and ANSI must use a fixed width font to
position text accurately, and consequently, source code is almost
always viewed in such fonts.
What sort of positioning information would be most useful?
Means of clear, easy, and minimal description of position that best
supports useful structures should be leading contenders. Indentation is the most popular way to express hierarchy through position
alone. It is so common that even though curly brace languages do
not use indentation, coders are exhorted to use “proper indentation” anyway, so that their source code is more readable. Perhaps
the most prominent and distinctive feature of the Python programming language is the use of pure positioning to indicate code structure. Another major use is the alignment of columns, usually for
tables. The ASCII tab character does not do either of these well.
Superscripting and subscripting can be considered a kind of positioning. It has a major limitation in that it does not scale. Each
successively deeper nesting requires progressively smaller text, which
soon becomes too small to read.
A proposal is to reassign 4 ASCII control characters for indentation. 3 of them can be increase indent (push), revert indent (pop),
and boost indent, analogous to the 2 brackets and colon in a closing 3 system. These characters can be invisible and have a width of
zero, not directly affecting the position of text. They only change
the level of indentation. The 4th character can mean forward to
the next indentation, replacing the leading spaces on all indented
lines of text. It could also mean advance to the next line, but that
Conference’17, July 2017, Washington, DC, USA
would be less flexible, wouldn’t support situations in which text
such as a line number is wanted before the indentation.
These characters do not specify the size of an indentation, only
the number of levels. This would allow individual users to set the
indentation size themselves without affecting others. It could also
make variable width fonts usable, as the problem of what unit to
use to specify indentation sizes is entirely avoided. It does add one
item to the state a text editor or viewer must maintain: a stack of
indentation levels.
Indentation characters could ease editing of source code. There
would be no more need to shift blocks of code several spaces right
or left by changing the number of leading spaces on each line,
whether by manually adding or deleting each space, or by using
some smart editor function such as that assigned to the tab key in
EMACS.
For the columns of tables, need better tab functionality. It would
be desirable not to rely on the use of a monospace font to recreate
the intended horizontal alignments. A limitation to dump is any
sort of tiny maximum distance of 8 spaces. Further, it should be
independent of any indentation setting. The C1 set of control characters contains several intended for tabular positioning, but they
do not do enough. One problem is that the state they set up is
global. Another is that they still implicitly depend upon a fixed
font, using the cursor position for fixing the location of tab stops.
It is basically a copy of the ideas and means used in the most advanced typewriters, with all their limitations.
The means HTML provides for laying out tables is fairly comprehensive and flexible, and proven over many years of use. If
better handling of tables is desired in plain text, copying HTML’s
handling and a subset of capabilities concerning tables into control character functionality seems a good approach. Lightweight
markup languages such as Markdown [26] and Bulletin Board Code
(bbcode) [22], arose to satisfy the desire to be able to create lists
and tables in text based forums more easily than with HTML. This
shows that many users like text editing to have such capabilities.
9
CONCLUSION
This paper proposed several changes in standard textual notations
to eliminate redundancy that may be hampering the human readability of structured documents such as source code. Proving that
human readability is improved was not attempted. Instead, the paper surmises that some kinds of redundancy merely add clutter,
showed where and how redundancy lurks, and proposed ways to
eliminate it. Good answers to Lots of Idiotic Spurious Parentheses
have been desired for a long time, and perhaps until now have not
been satisfactory.
Notation that scales and adds expressiveness, and allows much
more brevity without sacrificing clarity, is especially preferred. Punctuation with long histories in natural languages, especially English,
was tapped as a guide, in part because those uses are familiar to
people literate in those languages.
The first proposed change was to add a 3rd kind of bracket symbol roughly equivalent to the meaning of the colon in English, so
that a parent–child relationship represented as “(p(c))” is instead
represented as “(p:c)”. Proof was given that this 3 symbol system
Brenton Chapin
can collapse all runs of 2 or more closing brackets to a single closing bracket.
The idea of a universal closing bracket was presented. ”fa (b
[c] d) eg” can be represented as “fa (b [c] d] e]”, reducing the number of different symbols required, as ‘)’ and ‘g’ are no
longer needed.
More use of separators was proposed to replace sequences of
closing brackets followed by opening brackets. “((a()b)(c()d))”
can be represented with “((a,b;c,d))”.
Ways of adding types to the structure were discussed.
Positioning was recognized as an important way of denoting
structure. It is observed that means of expressing position have
been neglected. Markup languages limit themselves to data, and
are not much used for writing of other structured information such
as source code. Moreover, by using visible text to express position,
and requiring translation with special tools such as a web browser,
they fail at the goal of using position alone to express structure.
The means provided in ASCII work only with monospace fonts,
and require much wasteful redundancy. Repurposing some of the
unused control characters to better support indentation and tabular structure was proposed.
Together, these changes reduced the number and quantity of
symbols needed. They improved the amount of data compression
obtained by general purpose data compression programs. They reduced the size of the source code. Whether the goal of greater
human readability was also achieved was not studied, but it was
surmised that removing redundancies in the notation does help
with readability.
REFERENCES
[1] Leonhard Euler. “De infinitis curvis eiusdem generis seu methodus inveniendi aequationes pro infinitis curvis eiusdem generis,” Commentarii
academiae scientiarum Petropolitanae 7, 1740, pp. 174-189. Original available
at http://eulerarchive.maa.org/docs/originals/E044.pdf, English translation at
http://www.17centurymaths.com/contents/euler/e044tr.pdf .
[2] Dirichlet, P. G. L., “Beweis des Satzes, dass jede unbegrenzte arithmetische Progression, deren erstes Glied und Differenz ganze Zahlen
ohne gemeinschaftlichen Factor sind, unendlich viele Primzahlen
enthlt”, Abhand. Ak. Wiss. Berlin, 1837. English translation available at
http://arxiv.org/pdf/0808.1408v2.pdf.
[3] Florian Cajori. “A History of Mathematical Notations.” 1928-1929. Paragraph
643, Vol II, pg. 268.
[4] Jan Lukasiewicz. “Uwagi o aksjomacie Nicoda i ’dedukcji uoglniajfbcej’ ”
Ksiga pamifbtkowa Polskiego Towarzystwa Filozoficznego, Lww 1931.
[5] Oliver, B. M. “Efficient Coding.” Bell System Technical Journal, 31: 724-750.
doi:10.1002/j.1538-7305.1952.tb01403.x
[6] A. J. Perlis and K. Samelson. “Preliminary Report – International Algebraic
Language,” Communications of the ACM, July 1958.
[7] Kenneth E. Iverson. A Programming Language. John Wiley & Sons, Inc., New
York, NY, USA. 1962.
[8] Richards, Martin. “BCPL: A tool for compiler writing and system programming”
[9] Winett, J., ”EBCDIC Codes and Their Mapping to ASCII”, RFC 183, DOI
10.17487/RFC0183, July 1971, http://www.rfc-editor.org/info/rfc183
[10] Abelson, Harold and Sussman, Gerald Jay, with Sussman, Julie. “Structure
and Interpretation of Computer Programs.” 1985.
[11] Felsenstein,
Joe
et
al.
The
Newick
tree
format.
http://evolution.genetics.washington.edu/phylip/newicktree.html
[12] Bemer,
Robert.
“The
Great
Curly
Brace
Trace
Chase”
http://www.bobbemer.com/BRACES.HTM
[13] FORTRAN 77 4.0 Reference Manual. SunSoft. Part No.: 802-2998-10 Revision
A, November 1995. p. 38. https://archive.org/details/Fortran77Manual
[14] Tim Bray,
C. M.
Sperberg-McQueen,
et.
al.
Extensible
Markup Language (XML),
W3C
Working
Draft 14-Nov-96.
https://www.w3.org/TR/WD-xml-961114
[15] Rivest, R. “S-Expressions” Network Working Group, Internet Draft, May 4,
1997. Section 4.1. http://people.csail.mit.edu/rivest/Sexp.txt
Efficient textual representation of structure
[16] Goldfarb, Charles F. “SGML: The Reason Why and the First Published Hint.”
Journal of the American Society for Information Science, volume 48, number
7 (July 1997). Annotated reprint of Goldfarb, Charles F., Mosher, Edward J.,
and Peterson, Theodore I. “An Online System for Integrated Text Processing”
Journal of the American Society for Information Science, volume 7 (Oct 1970).
http://www.sgmlsource.com/history/jasis.htm.
[17] Oren Ben-Kiki, Clark Evans, Brian Ingerson. YAML Ain’t Markup Language
(YAML) 1.0. Final Draft 2004-JAN-29. http://yaml.org/spec/1.0/
[18] Wall,
Larry.
“Synopsis
4:
Blocks
and
Statements”
http://design.perl6.org/S04.html
[19] Markstrum, Shane. “Staking Claims: A History of Programming Language
Design Claims and Evidence.”
[20] Shafranovich, Y., ”Common Format and MIME Type for Comma-Separated
Values (CSV) Files”, RFC 4180, DOI 10.17487/RFC4180, October 2005,
http://www.rfc-editor.org/info/rfc4180
[21] Buse, Raymond P.L. and Weimer, Westley R. “A Metric for Software Readability.” Proceedings of the 2008 International Symposium on Software Testing
and Analysis, pp. 121-130.
[22] anonymous authors. Bulletin Board Code. http://www.bbcode.org
[23] Yasuoka, Kiochi and Yasuoka, Motoko. “On the Prehistory of QWERTY.” ZINBUN (2011), 42: 161-174. https://doi.org/10.14989/139379
[24] ECMA
404
The
JSON
Data
Interchange
Format.
https://www.ecma-international.org/publications/files/ECMA-ST/ECMA-404.pdf
[25] HTML Elements. http://www.w3schools.com/html/html elements.asp
[26] MacFarlane, John. CommonMark Spec. http://spec.commonmark.org/
[27] Leonard, S., ”Guidance on Markdown: Design Philosophies, Stability Strategies, and Select Registrations”, RFC 7764, DOI 10.17487/RFC7764, March 2016.
http://www.rfc-editor.org/info/rfc7764
[28] Open Tree of Life, v9.1. http://files.opentreeoflife.org/synthesis/opentree9.1/opentree9.1 tree.tgz
Conference’17, July 2017, Washington, DC, USA
| 7cs.IT
|
Computation of the first Chow group of
a Hilbert scheme of space curves
arXiv:1103.0122v2 [math.AG] 20 Jun 2014
Gerd Gotzmann
Abstract
An earlier wrong formula for the dimension of A1 (Hd,g ) ⊗ Q is corrected.
Introduction
The results stated in ([T4], pp. 1) have to be corrected as follows: Let H = Hd,g =
HilbP (P3C ) be the Hilbert scheme , which parametrizes the curves in P3C of degree d and
genus g (i.e., the closed subschemes of P3C with Hilbert polynomial P (T ) = dT − g + 1).
It is always assumed that d ≥ 3 and g is not maximal, i.e. that g < (d − 1)(d − 2)/2.
Theorem 0.1. Let be g(d) := (d − 2)2 /4. Then dimQ A1 (Hd,g ) ⊗ Q = 3 (resp. = 4),
if g ≤ g(d) (resp. if g > g(d)).
Corollary 0.1. NS(H) ≃ Zρ and Pic(H) ≃ Zρ ⊕ Cr , where r := dimC H 1 (H, OH )
and ρ = 3, if g ≤ g(d). If g > g(d), then ρ = 3 or ρ = 4.
Theorem 0.2. Let C ֒→ H×P3 be the universal curve over H. Then dimQ A1 (C) ⊗ Q =
Z
dimQ A1 (H) ⊗ Q + 1.
Z
Corollary 0.2. NS(C) = Zρ+1 and Pic(C) ≃ Zρ+1 ⊕Cs , where s := dimC H 1 (C, OC )
and ρ is defined as in Corollary 1.
That means, the formula (d − 2)(d − 3)/2 for the bound g(d) in ([T4], p.1) is wrong
and has to be replaced by the above formula.
March 2, 2011
Contents
Chapter 1. Summary of earlier results
1
Chapter 2. Subschemes of points in P2 and their Hilbert functions
5
Chapter 3. A rough description of ideals invariant under Γ · T (ρ)
27
Chapter 4. The α-grade.
43
Chapter 5. Estimates of the α-grade in the case ρ1 < 0, ρ2 > 0.
55
Chapter 6. Estimates of the α-grade in the case ρ1 > 0, ρ2 > 0 and r ≥ 1.
65
Chapter 7. Estimates of the α-grade in the case ρ1 > 0, ρ2 > 0 and r = 0.
79
Chapter 8. Borel-invariant surfaces and standard cycles
87
Chapter 9. Relations between B-invariant 1-cycles
95
Chapter 10. Proof of Theorem 1.1
101
Chapter 11. Surfaces in H invariant under Ga · T (4; k)
103
Chapter 12. Surfaces in H invariant under B(4; k)
107
Chapter 13. Relations in B(4; k)-invariant surfaces
115
Chapter 14. Necessary and sufficient conditions
119
Chapter 15. The case d ≥ 5
123
Chapter 16. The cases d = 3 and d = 4
125
Chapter 17. Correction of the results of [T4] and summary
129
Appendix A. Notations
131
Appendix B. Hilbert functions without Uniform Position Property
135
Appendix C. Ideals with many monomials
137
Appendix D. Unipotent groups acting on polynomial rings
139
Appendix E. Standard bases
143
Appendix. Bibliography
145
i
CHAPTER 1
Summary of earlier results
1.1. Description of the starting situation
The result are the same as in [T1]-[T4] and are summed up in Appendix A. The
ground field is C, and H = Hd,g is the Hilbert scheme which parametrizes the curves
C ⊂ P3C of degree d and genus g (i.e. the closed subschemes of P3C with Hilbert polynomial
P (T ) = dT − g + 1). According to F.S.Macaulay, Hd,g is not empty if and only if the
“complementary” Hilbert polynomial Q(T ) = T +r
− P (T ) either has the form Q(T ) =
T −a+2
Tr −a+2 T −b+1
T −1+3
T −1+3
+
or the form Q(T ) =
+
+ 1 , where a is an integer ≥
3
2
3
2
1, respectively a and b are integers (Macaulay coefficients), such that 2 ≤ a ≤ b. Between
the degree and genus on the one hand and the Macaulay coefficients on the other hand, one
has the following relations d = a, g = (d − 1)(d − 2)/2, if Q(T ) = T −1+3
+ T −a+2
, and
3
2
T −a+2 T −b+1
T −1+3
2
d = a−1, g = (a −3a+4)/2−b, if Q(T ) =
+
+
, respectively. One
3
2
1
sees that the first case occurs if and only if one is dealing with plane curves, in which case
the groups A1 (H) and NS(H) both have the rank 2 (cf. [T1], Satz 2a, p. 91). Therefore
in the following we always suppose that d ≥ 3 and g < (d − 1)(d − 2)/2, that means, the
complementary Hilbert polynomial has the form Q(T ) = T −1+3
+ T −a+2
+ T −b+1
,
3
2
1
where 4 ≤ a ≤ b.
We also write HQ instead of Hd,g in order to express that this Hilbert scheme likewise
parametrizes the ideals I ⊂ OP3 with Hilbert polynomial Q(T ), or equivalently, the
saturated graded ideals in C[x, y, z, t] with Hilbert polynomial Q(T ).
In [T1]-[T4] it was tried to describe the first Chow group A1 (H), where we always
take rational coefficients, and we write A1 (H) instead of A1 (H) ⊗ Q. The starting point
Z
is the following consideration: If the Borel group B = B(4; k) operates on H = HQ
in the obvious way, then one can deform each 1-cycle on H in a 1-cycle, whose prime
components are B-invariant, irreducible, reduced and closed curves on H. It follows that
A1 (H) is generated by such B-invariant 1-prime cycles on H. This is a partial statement
of a theorem of Hirschowitz. (Later on we will have to use the general statement, whereas
the partial statement can be proved in a simple way, see [T1], Lemma 1, p. 6.) Now such
a B-invariant 1-prime cycle (i.e. closed, irreducible and reduced curve) C on H can be
formally described as follows: Either each point of C is invariant under ∆ := U(4; k), or
one has C = Gia · η, where η is a closed point of H, which is invariant under T = T (4; k)
and the group Gi , i ∈ {1, 2, 3}. Here Gia is the group Ga , acting by
ψα1 : x 7−→ x, y 7−→ y, z 7−→ z, t 7−→ αz + t
ψα2 : x 7−→ x, y −
7 → y, z 7−→ αy + z, t 7−→ t
ψα3 : x −
7 → x, y −
7 → αx + y, z 7−→ z, t 7−→ t,
1
respectively, on P = k[x, y, z, t], and Gi is
Gia , that means, one defines
1
∗
∗
∗
0
1
∗
∗
,
G
:=
G1 :=
2
0
0
1
0
0 0 0 1
the subgroup of ∆, which is complementary to
∗
1
0
0
1
0
0
0
∗
0
1
0
∗
∗
,
∗
1
G3 :=
1
0
0
0
0
1
0
0
∗
∗
1
0
∗
∗
.
∗
1
If C has this form, then C is called a curve or a 1-cycle of type i, where i ∈ {1, 2, 3}.
A(H) := Im(A1 (H∆ ) → A1 (H)) is called the “algebraic part” and A1 (H) := A1 (H)/A(H)
is called the “combinatorial part” of the first Chow group of H. Here H∆ denotes the
fixed point scheme which, just as all other fixed point schemes that will occur later on,
is supposed to have the induced reduced scheme structure. (This convention is valid also
for the Hilbert scheme H d := Hilbd (P2C ), see below.)
In order to formulate the results obtained so far, one has to introduce the following
”tautological” 1-cycles on H:
−
C1 = (x, y a , y a−1 z b−a (αz + t)) α ∈ k
C2 = {(x, y a−1 (αy + z), y a−2 z b−a+1 (αy + z))|α ∈ k}−
C3 = {(xa , αx + y, xa−1 z b−a+1 )|α ∈ k}−
D = {(x2 , xy, y a−1, z b−2a+4 (y a−2 + αxz a−3 ))|α ∈ k}−
E = {(x2 , xy, xz, y a, y a−1 z b−a+1 , xtb−2 + αy a−1z b−a )|α ∈ k}−
For the sake of simplicity, we now suppose d ≥ 5 (i.e. a ≥ 6). (The cases d = 3 and d = 4
will be treated separately in Chapter 16.) Then one has the following results:
1. If b < 2a − 4, i.e. if g > γ(d) := (d − 2)(d − 3)/2, then A(H) is generated by E, and
A1 (H) is generated by E, C1 , C2 , C3 .
2. If b ≥ 2a − 4, i.e if g ≤ γ(d), then A(H) is generated by E and D and A1 (H) is
generated by E, D, C2 and C3 ( see [T1], Satz 2, p. 91; [T3], Proposition 4, p. 22; [T4],
Satz 1 and Proposition 2, p. 26).
From reasons of degree it follows that [C2 ] can not lie in the vector space spanned by
[E], [D], [C3 ], so the problem is to decide, if [C3 ] ∈ A(H).
In ([T4], Proposition 3, p. 32) it was erroneously claimed that [C3 ] ∈ A(H), if b ≥
2a − 4. (The error is the wrong computation of the degree in ([T4], p. 28, line 21 to line
30.) Therefore the bound for the genus in ([T4], p. 1) is wrong.
Actually, in ([T2], 3.3.2) it had been proved, that [C3 ] ∈ A(H), if a ≥ 6 is even and
b ≥ a2 /4, i.e. if d ≥ 5 is odd and g ≤ (d − 1)(d − 3)/4. In the case d ≥ 6 even , in
Conclusion 14.3 it will follow that [C3 ] ∈ A(H), if g ≤ (d − 2)2 /4. (This means the bound
of [T2], 3.3.3 is valid if d ≥ 6, already . ). One sees that the condition for g in both cases
can be summed up to g ≤ (d − 2)2 /4.
The major part of the following text serves for the proof that this sufficient condition
is a necessary condition, too (cf. Conclusion 14.1).
2
1.2. Technical tools
The formulas in ([T2], p. 134) and of ([T3], Anhang 2, p. 50) show that it is not
possible to decide by means of the computation of degrees, whether [C3 ] lies in A(H).
Therefore we try to get a grasp of the relations among the B-invariant 1-cycles on H with
the help of the theorem of Hirschowitz ([Hi], Thm. 1, p. 87). We sketch the procedure.
1.2.1. The Theorem of Hirschowitz. There is a closed and reduced subscheme
Z = Z(H) of H, such that Z(k) = {x ∈ H(k)| dim ∆ · x ≤ 1} (cf. [Ho], p. 412 and [T3],
Lemma 1, p. 35). Then one can show, with the help of the theorem of Hirschowitz, that
∼
A1 (Z) → A1 (H) (cf. [T2], Lemma 24, p. 121). As was explained in (1.1), A1 (H) has a
generating system consisting of B-invariant 1-cycles which lie in Z, automatically. As ∆
is normalized by B, B operates on Z and therefore one can form the so called equivariant
Chow group AB
1 (Z), which is isomorphic to A1 (Z) ([Hi], loc. cit.). And the relations
among B-invariant 1-cycles on Z are generated by relations among such cycles, which lie
on B-invariant surfaces V ⊂ Z ( see [Hi], Mode d’emploi, p. 89).
1.2.2. The Restriction morphism. Let Ut ⊂ H be the open subset consisting of
the ideals I ⊂ OP3 with Hilbert polynomial Q, such that t is not a zero divisor of OP3 /I.
Then there is a so called restriction-morphism h
: Ut →
H d := Hilbd (P2C ), defined by
I 7→ I ′ := I + tOP3 (−1)/tOP3 (−1). E.g., if G :=
1
0
0
0
0
1
0
0
0
0
1
0
∗
∗
∗
1
< ∆, then the fixed point
scheme F := HG is contained in Ut , and the restriction of h to F is denoted by h, again.
In ([G4], Abschnitt 6, p. 672f) the following description of Im(h) is given:
(i) There is a finite set F of Hilbert functions of ideals of colength d on P2 such that
S
Im(h) = {H≥ϕ |ϕ ∈ F }.
(ii) If k = k and if I ⊂ OP2k is an ideal of colength d and Hilbert function ϕ, then
d
P
I ∈ Im(h) ⇐⇒ g ∗ (ϕ) :=
ϕ(n) − d+3
+ d2 + 1 ≥ g.
3
0
(iii) If ϕ ∈ F and if ψ is the Hilbert function of an ideal on P2 of colength d such that
ϕ(n) ≤ ψ(n) for all n ∈ N, then ψ ∈ F .
(iv) Let be ϕ ∈ F and I ⊂ OP2k an ideal with Hilbert function ϕ. Let I ∗ be the ideal
n
in OP3k defined by H 0 (I ∗ (n)) = ⊕ tn−i H 0 (I(i)), then V+ (I ∗ ) ⊂ P3k is a curve of degree d
and genus g ∗ (ϕ).
i=0
Here Hϕ ⊂ H d is the locally closed subscheme (with the reduced induced scheme
structure), which parametrizes the ideals I ⊂ OP2 of colength d with Hilbert function
S
ϕ, and H≥ϕ := {Hψ |ψ ≥ ϕ} is a closed subscheme ( with the induced reduced scheme
structure). The image of C3 under h is the 1-cycle c3 := {(xd , αx + y)|α ∈ k}− . One has
S
Theorem 1.1. Let be d ≥ 5, H := {Hϕ ⊂ H d |g ∗(ϕ) > g(d)} and
A(H) := Im(A1 (HU (3;k) ) → A1 (H)). Then [c3 ] ∈
/ A(H).
3
The proof extends over the chapters 2 to 10 and essentially rests on the apparently
strong condition for an ideal I to have a Hilbert function ϕ such that g ∗(ϕ) > g(d).
1.2.3. Standard cycles on H d . It has been shown, respectively it will be shown
that [C3 ] ∈ A(Hd,g ), if g ≤ g(d) (cf. 1.1). Therefore, we can suppose that g > g(d). If
J ∈ Ut and the restriction ideal I := J ′ has the Hilbert function ϕ, then from (ii) in
(1.2.2) it follows that g ∗ (ϕ) > g(d). It will be shown in Chapter 2 that this implies there
is a linear form ℓ ∈ S1 − (0), an ideal K ⊂ OP2 of colength c and a form f ∈ H 0 (K(m))
such that I = ℓK(−1) + f OP2 (−m), c + m = d and m ≥ c + 2.
Let be C = Ga · η ⊂ H a 1-cycle of type 3 and let be J ↔ η the corresponding
′
′
ideal in OP3 with Hilbert polynomial
Q. Thenthe
ideal I := J ↔ η := h(η) is
1 0 ∗
invariant under T (3; k) and Γ := 0 1 ∗ < U(3, k). It follows that either
0 0 1
I = xK(−1) + y m OP2 (−m) or I = yK(−1) + xm OP2 (−m), where K is a monomial
ideal. We say, I has x-standard form or I has y-standard form, respectively, and we call
C ′ := Ga · η ′ a x-standard cycle or y-standard cycle on H d , respectively. With the help of
the theorem of Hirschowitz one can again try to describe the relations between B(3; k)invariant y-standard cycles on H, and one obtains that such relations cannot make the
y-standard cycle c3 disappear modulo A(H) (cf. Proposition 9.1) from which Theorem
0.1 will follow.
1.2.4. 1-cycles of proper type 3. Let be C = Ga · η, η ↔ J , be a 1-cycle of type
3 on H = Hd,g , such that d ≥ 5 and g > g(d). C is called a 1-cycle of proper type 3, if
C ′ = Ga · η ′ is a y-standard cycle on H. Corresponding to Hirschowitz’s theorem one has
to consider B(4; k)-invariant surfaces V ⊂ Z(H), which contain a 1-cycle of proper type
3. It turns out that then V is pointwise invariant under G3 and therefore V is contained
in Ut . Then one can map relations between B-invariant 1-cycles on V by h∗ into relations
between B(3; k)-invariant 1-cycles on h(V ), and one obtains with the aid of Proposition
9.1 the main result of the second part of the paper ( Theorem 14.1), which corresponds
to Theorem 0.1. In Chapter 15 there is complete description of A1 (Hd,g ) if d ≥ 5, and in
Chapter 16 this is done in the cases d = 3 and d = 4 (Theorem 15.1 and Theorem 16.1,
respectively).
4
CHAPTER 2
Subschemes of points in P2 and their Hilbert functions
2.1. General properties
The ground field is C and k denotes an extension field. A closed subscheme Z ⊂ P2k of
− d. If
length d > 0 is defined by an ideal I ⊂ OP2k with Hilbert polynomial Q(n) = n+2
2
0
0
2
the Hilbert function h (I(n)) = dimk H (Pk ; I(n)), n ∈ N, of I is denoted by ϕ(n), then
ϕ′ (n) := ϕ(n) − ϕ(n − 1), n ∈ N, denotes the difference function. If ϕ : N −→ N is any
function, such that ϕ(n) = n+2
− d for n ≫ 0, then the ideals I ⊂ OP2k with Hilbert
2
function ϕ form a locally closed subset Hϕ of the Hilbert scheme H d = Hilbd (P2C ), and we
take Hϕ as a subscheme of H d with the induced reduced scheme structure.
Iarrobino has shown ([I], Lemma 1.3, p.8) that Hϕ 6= ∅ if and only if the difference
function fulfils the following two conditions:
(a) ϕ′ (n) ≤ n + 1, for all n ∈ N and
(b) ϕ′ (n) ≤ max(ϕ′ (n + 1) − 1, 0), for all n ∈ N.
If α = α(ϕ) := min{n ∈ N | ϕ(n) > 0}, then (b) is equivalent to:
(b’) ϕ′ (n) + 1 ≤ ϕ′ (n + 1), for all n ≥ α.
The (Mumford-)regularity e of an ideal I with Hilbert function ϕ as before is characterized by e = reg(ϕ) = min{n ∈ N | ϕ′ (n + 1) = n + 1} (cf. Appendix B, Lemma
2). In principle, the graph of ϕ′ has the shape of Fig. 2.1. If ∅ =
6 Hϕ ⊂ H d , then d is
n+2
determined by the condition ϕ(n) = 2 − d, n ≫ 0, and we call d the colength of ϕ.
It is known that reg(ϕ) ≤ d ([G1], Lemma 2.9, p. 65), and reg(ϕ) = d is equivalent with
I being generated by a linear form ℓ ∈ S1 and a form f ∈ Sd , not divisible by ℓ. Here
S = k[x, y, z] is the graded polynomial ring. Another characterization of reg(ϕ) = d is
that the graph of ϕ′ has the shape of Fig.2.2. One notes that the colength of ϕ is equal to
the number of ”monomials” between the graph of ϕ′ and the line y = x + 1. (For this and
other properties , see [T1]-[T4].) In the following we write P2 instead of P2k and denote by
I an ideal in OP2 , whose finite colength (resp. whose Hilbert function) usually is denoted
by d (resp. by ϕ).
2.2. Numerical and algebraic properties
Lemma 2.1. Let be k = k, I ⊂ OP2 an ideal with colength d, Hilbert function ϕ and
regularity m. We assume that there is a number ε ∈ N, 0 ≤ ε < m−2, such that ϕ′ (n) = n
for all n ∈ N, such that ε + 1 ≤ n ≤ m − 1. Then there is a linear form ℓ = S1 , an ideal
K ⊂ OP2 of colength c and a form f ∈ H 0 (K(m)), such that I = ℓK(−1) + f OP2 (−m). If
5
ℓ1 , ℓ2 are any linear forms in S1 such that ℓ, ℓ1 , ℓ2 is a basis of the k-vector space S1 and
if R := k[ℓ1 , ℓ2 ] is the subring of S, isomorphic to k[x, y], then d = c + m and
(
ℓH 0 (K(n − 1))
if n < m,
H 0 (I(n)) =
0
ℓH (K(n − 1)) ⊕ f Rn−m if n ≥ m.
Proof. By assumption, the graph of ϕ′ has the shape as in Fig.2.3. Then there is a
ℓ ∈ S1 − (0) and an ideal K ⊂ OP2 of regularity ≤ ε such that H 0 (I(n)) = ℓH 0 (K(n − 1))
for all n ≤ m − 1 (cf. [G4] and Appendix B, Lemma 1). If ψ is the Hilbert function of
K, then ϕ′ (n) = ψ ′ (n − 1) for 1 ≤ n ≤ m − 1, and because of the shape of the graphs
of ϕ′ and ψ ′ it follows that ϕ(n) = ψ(n − 1) + (n − m + 1) for all n ≥ m. Therefore
H 0 (I(m)) = ℓH 0 (K(m − 1)) ⊕ f · k, where f ∈ H 0 (I(m)) is a suitable section. Because
of the m-regularity of I it follows that H 0 (I(n)) = ℓH 0 (K(n − 1)) + f Sn−m , n ≥ m. If
n = m + 1, then from ϕ(m + 1) = ψ(m) + 2 it follows that S1 f ∩ ℓH 0 (K(m)) has the
dimension 1. Thus there is a h ∈ S1 − (0), such that hf ∈ ℓH 0 (K(m)). If ℓ would be a
divisor of f , then it would follow that I ⊂ ℓOP2 (−m) and thus I would not have a finite
colength in OP2 . Therefore we may suppose that h = ℓ, and it follows that f ∈ H 0 (K(m)).
We choose ℓ1 , ℓ2 ∈ S1 such that ℓ, ℓ1 , ℓ2 are linear independent and we put R := k[ℓ1 , ℓ2 ].
If there would be a r ∈ Rν − (0) such that rf ∈ ℓH 0 (K(m + ν − 1)), then it would follow
that ℓ is a divisor of f , contradiction. Between the graph of ψ ′ (n − 1) and the line y = x
there are exactly c := colength(ψ) monomials, and therefore d = c + m (cf. Fig.2.3).
Corollary 2.1. The assumptions and notations are as in Lemma 2.1. Then one
has:
(i) κ := reg(K) ≤ ε, especially κ ≤ m − 3.
(ii) K (respectively the linear form ℓ) is uniquely determined (respectively uniquely up to
a factor out of k different from zero).
(iii) f is uniquely determined up to a factor out of k different from zero, modulo ℓH 0 (K(m−
1)).
(iv) κ and ε are uniquely determined by ϕ.
Proof. (i) ε + 1 = ϕ′ (ε + 1) = ψ ′ (ε). From (Appendix B, Lemma 2) it follows that
κ ≤ ε.
(ii) The regularity only depends on the Hilbert function, and therefore κ = reg(K1 ) =
reg(K2 ) < m − 1. Thus from ℓ1 H 0 (K1 (m − 2)) = ℓ2 H 0 (K2 (m − 2)) it follows that
ℓ1 K1 (−1) = ℓ2 K2 (−1).
If ℓ1 would not be a divisor of ℓ2 then one would have K2 ⊂ ℓ1 OP2 (−1) contradiction.
From this assertion (ii) does follow, and (iii) and (iv) are clear.
Remark 2.1. If ϕ and ψ are two Hilbert functions of colength d, then from ϕ < ψ
(that means ϕ(n) ≤ ψ(n) for all n ∈ N and ϕ(n) < ψ(n) for at least one n ∈ N) it follows
that g ∗ (ϕ) < g ∗ (ψ). This follows immediately from the definition of g ∗(ϕ) in (1.2.2).
6
e−2
P
Remark 2.2. If e := reg(ϕ), d := colength(ϕ) and s(ϕ) :=
ϕ(i), then g ∗(ϕ) =
i=0
n+2
s(ϕ) − e+1
+
d(e
−
2)
+
1.
Because
of
ϕ(n)
=
− d for all n ≥ e − 1 this follows
3
2
from a simple computation with binomial coefficients.
2.2.1. Hilbert functions of colength ≤ 4. We use the formula of Remark 2.2 and
orientate ourselves by the figures 2.4–2.7.
d = 1 There is only one Hilbert function (cf. Fig. 2.4).
2
∗
+ 1 · (1 − 2) + 1 = 0.
e = 1, s(ϕ) = 0, g (ϕ) = 0 −
3
d = 2 There is again only one Hilbert function (cf. Fig. 2.5).
3
∗
+ 2 · 0 + 1 = 0.
e = 2, s(ϕ) = 0, g (ϕ) = 0 −
3
d = 3 There are two Hilbert functions (Fig. 2.6 a and Fig. 2.6 b).
3
∗
+ 3 · 0 + 1 = 0,
e1 = 2, s(ϕ1 ) = 0, g (ϕ1 ) = 0 −
3
4
∗
+ 3 · 1 + 1 = 1.
e2 = 3, s(ϕ2 ) = 1, g (ϕ2 ) = 1 −
3
d = 4 There are two Hilbert functions (Fig. 2.7 a and Fig. 2.7 b).
4
∗
+ 4 · 1 + 1 = 1,
e1 = 3, s(ϕ1 ) = 0, g (ϕ1 ) = 0 −
3
5
∗
+ 4 · 2 + 1 = 3.
e2 = 4, s(ϕ2 ) = 4, g (ϕ2 ) = 4 −
3
2.2.2. Two special ideals. First case: If d ≥ 6 is even, then let be e := d/2 + 1 and
I := (x2 , xy e−2, y e ). The Hilbert function χ can be read from Fig. 2.8. One notes that
n−1
P
colength(I) and reg(I) really are equal to d and e, respectively, and χ(n) =
i = n2 ,
1
if 1 ≤ n ≤ e − 2. Therefore s(χ) =
−
e+1
3
1
i
2
=
e−1
3
and it follows that
e
e
e+1
+ 2(e − 1)(e − 2) + 1 = e−1
−
+
−
+ 2e2 − 6e + 5
3
3
3
3
1
1
= − (e − 1)(e − 2) − e(e − 1) + 2e2 − 6e + 5 = e2 − 4e + 4
2
2
1
= (e − 2)2 = (d − 2)2 .
4
g ∗(χ) =
e−1
3
e−2
P
Second case: If d ≥ 5 is odd, then let be e := (d + 1)/2 and I := (x2 , xy e−1 , y e).
The Hilbert function χ can be read from Fig. 2.9. One notes that colength (I) and reg(I)
are equal to d and e, respectively, and χ(n) = n2 , if 1 ≤ n < e.
7
Therefore s(χ) =
e−2
P
2
g ∗ (χ) =
e−1
3
−
i
2
=
e+1
3
e−1
3
and it follows that
+ (2e − 1)(e − 2) + 1 = −
e−1
2
−
e
2
+ (2e − 1)(e − 2) + 1
3
1
= −(e − 1)2 + 2e2 − 5e + 3 = e2 − 3e + 2 = (d + 1)2 − (d + 1) + 2
4
2
1 2
= (d − 4d + 3).
4
Definition 1. If d ≥ 5, then we set
(
1
(d − 2)2
if d ≥ 6 is even,
g(d) := 41
(d − 1)(d − 3) if d ≥ 5 is odd.
4
g(d) is called the deformation bound for ideals in OP2 of colength d.
The rest of the article is to justify this notation.
2.2.3.
Lemma 2.2. Let be k = k, I ⊂ OP2 an ideal of colength d ≥ 5 and regularity m. Let
be ϕ the Hilbert function of I. If g ∗(ϕ) > g(d), then the assumptions of Lemma 2.1 are
fulfilled by I.
Proof. Let be χ the Hilbert function defined by Fig. 2.8 and Fig. 2.9, respectively.
Let be m = reg(ϕ). If ϕ′ (m) − ϕ′ (m − 1) > 2, then ϕ′ (i) ≤ χ′ (i) and therefore ϕ(i) < χ(i)
for all i, and it would follow that g ∗ (ϕ) ≤ g(χ) (Remark 2.1). If ϕ′ (m)−ϕ′ (m−1) = 1, then
ϕ′ (m−1) = ϕ′ (m)−1 = (m+1)−1 = (m−1)+1, therefore reg(I) ≤ m−1 (cf. Appendix
B, Lemma 2). It follows that ϕ′ (m) − ϕ′ (m − 1) = 2, therefore ϕ′ (m − 1) = m − 1. If
ϕ′ (m−2) = m−2, as well, then the assumptions of Lemma 2.1 are fulfilled with ε := m−3,
for instance. Thus without restriction of generality one can assume ϕ′ (m − 2) ≤ m − 3.
Case 1: ϕ′ (m − 2) < m − 3. Figure 2.10 represents the Hilbert function ϕ as well as the
B(3; k)-invariant ideal M with Hilbert function ϕ. Then one makes the deformation
E(H 0 (M(m))) 7→ E(H 0 (M(m))) − u ∪ v =: E(H 0 (N (m))),
where N is a B(3; k)-invariant ideal with Hilbert function ψ > ϕ. But then it follows
g ∗ (ϕ) < g ∗(ψ) ≤ g ∗ (χ) = g(d), contradiction.
Case 2: ϕ′ (m − 2) = m − 3. If the graph of ϕ′ would have a shape different from that
in Fig. 2.11 a, i.e., if the graph of ϕ′ would have a “jumping place” n < m − 2 (cf. the
terminology in [T1], p. 72), then as in the first case one could make a deformation
E(H 0 (M(m))) 7→ E(H 0 (M(m))) − u ∪ v =: E(N (m)))
(cf. Fig. 2.11b) and would get a contradiction, again. It only remains the possibility
represented in Fig. 2.11a. But then ϕ = χ, which contradicts the assumption g ∗(ϕ) >
g(d).
8
2.3. Numerical conclusions from g ∗ (ϕ) > g(d)
2.3.1. At first we describe the starting situation: In this section we suppose that
g(d) is defined, i.e. d ≥ 5. Moreover let be ϕ a Hilbert function such that Hϕ 6= ∅,
colength(ϕ) = d, reg(ϕ) = m and g ∗ (ϕ) > g(d). Then the assumptions of Lemma 2.2 are
fulfilled for an ideal I, which can be supposed to be monomial. Therefore the assumption
k = k is superfluous. As m and d are uniquely determined by ϕ, c := d − m is uniquely
determined, too.
The aim in this section is to prove the inequality m ≥ c + 2. By Lemma 2.1 and
Lemma 2.2, respectively, one can write I = ℓK(−1) + f OP2 (−m), and c is equal to the
colength of the Hilbert function ψ of K. (As I is monomial, K is monomial, too, and
without restriction one has ℓ ∈ {x, y, z}.)
Lemma 2.3. Let be ψ the Hilbert function of an ideal K of colength c ≥ 5, κ :=
reg(ψ), and m ≥ κ + 2 an integer. If one defines ϕ by ϕ′ (n) := ψ ′ (n − 1), 0 ≤ n ≤
m − 1, ϕ′ (n) := n + 1, n ≥ m, then Hϕ 6= ∅, colength(ϕ) = c + m, reg(ϕ) = m and
g ∗ (ϕ) = g ∗(ψ) + 12 m(m − 3) + c.
Proof. We orientate ourselves by Figure 2.3, but the weaker assumption m ≥ κ + 2
takes the place of the assumption m ≥ κ + 3.
Without restriction one can assume that K is B(3; k)-invariant. Then y m ∈ H 0 (K(m))
and I := xK(−1) + y mOP2 (−m) has the Hilbert function ϕ, the regularity m and the
colength c + m. This follows by considering the figure mentioned above (and has been
shown in a more general situation in [G4], Lemma 4, p. 660). We compute g ∗(ϕ) (cf.
Remark 2.2):
s(ϕ) =
m−2
X
κ−1
X
ϕ(i) =
i=0
i=0
=
κ−2
X
ψ(i) +
i=0
= s(ψ) +
ψ(i − 1) +
m−3
X
m−2
X
i=κ
ψ(i) = s(ψ) +
i=0
i+2
2
m−3
X
i=κ−1
i=κ−1
m−3
X
ψ(i − 1)
−
κ−2
X
i=0
i+2
2
i+2
2
−c
− (m − κ − 1)c.
By Remark 2.2 it follows that:
m+1
g ∗ (ϕ) = s(ψ) + m3 − κ+1
−
(m
−
κ
−
1)c
−
+ (c + m)(m − 2) + 1
3
3
κ+1
m
m+1
= s(ψ) − 3 + c(κ − 2) + 1 + 3 − 3 − mc + c + (c + m)m − 2m
1
= g ∗ (ψ) − m2 + m2 − 2m + c = g ∗ (ψ) + m(m − 3) + c.
2
2.3.2. The cases c ≤ 4. By Lemma 2.2 (resp. by Corollary 2.1 of Lemma 2.1) one
has ϕ′ (n) = ψ ′ (n − 1), 0 ≤ n ≤ m − 1, ϕ′ (n) = n + 1, n ≥ m, and κ = reg(K) ≤ m − 3.
Then the assumptions of Lemma 2.3 are fulfilled.
9
We use the formula given there and orientate ourselves by the Figures 2.4 - 2.7. The
regularity of the Hilbert function considered each time will now be denoted by κ.
If
If
If
If
c ∈ {0, 1}, then because of d = m + c it follows that m ≥ 4.
c = 2, then κ = 2 and m ≥ κ + 3 = 5.
c = 3 and κ = 2 or κ = 3, then m ≥ κ + 3 = 5.
c = 4, then κ = 3 or κ = 4 and m ≥ κ + 3 ≥ 6.
Thus in the cases 0 ≤ c ≤ 4 one has m ≥ c + 2.
2.3.3. The case g ∗ (ψ) ≤ g(c). This notation implies that c ≥ 5. If κ is the regularity
and c is the colength of any Hilbert function, then because of 1 + 2 + · · · + κ ≥ c, one
√
2c − 1. By Lemma 2.2 the assumptions of
always has κ+1
≥
c,
and
therefore
κ
≥
2
Lemma 2.1 are fulfilled, therefore by Corollary 2.1 it follows that m ≥ κ + 3 > 5.16.
1st case: c and m are even.
By the formulas for g(d) and g ∗ (ϕ) it follows that:
1 2
(c
4
− 4c + 4) + 12 m(m − 3) + c > 41 [(c + m)2 − 4(c + m) + 4]
⇐⇒
1
m(m
2
− 3) + c > 14 [2cm + m2 − 4m]
⇐⇒ m2 − 2(c + 1)m + 4c > 0.
The solutions of the corresponding quadratic equation are 0 and 2c.
Therefore m ≥ 2c + 1 > c + 2.
2nd case: c is even, m is odd.
One obtains the inequality:
1 2
(c
4
− 4c + 4) + 12 m(m − 3) + c > 41 [(c + m)2 − 4(c + m) + 3]
m2 − 2(c + 1)m + 4c + 1 > 0.
√
The solutions of the corresponding quadratic equation are m = c + 1 ± c2 − 2c ≥ 0.
√
√
Because of c + 1 − c2 − 2c < 3, if c ≥ 5, it follows that m ≥ c + 1 + c2 − 2c. Because of
√
c + 1 + c2 − 2c > 2c − 1, if c ≥ 5, it follows that m ≥ 2c, therefore m ≥ 2c + 1 > c + 2.
⇐⇒
3rd case: c is odd, m is even.
One obtains the inequality :
1 2
(c
4
− 4c + 3) + 21 m(m − 3) + c > 14 [(c + m)2 − 4(c + m) + 4]
⇐⇒
m2 − 2(c + 1)m + 4c > 0.
It follows that m ≥ 2c + 1 > c + 2.
4th case: c and m are odd.
One obtains the inequality:
1 2
(c
4
⇐⇒
− 4c + 3) + 21 m(m − 3) + c > 14 [(c + m)2 − 4(c + m) + 4]
m2 − 2(c + 1)m + 4c − 1 > 0.
10
p
The solutions of thepcorresponding quadratic equation are m = c +p1 ± (c − 1)2 + 1.
Because of c + 1 − (c − 1)2 + 1 < 2 it follows that m ≥ c + 1 + (c − 1)2 + 1 > 2c,
therefore m ≥ 2c + 1 ≥ c + 2.
2.3.4. The case g ∗ (ψ) ≥ g(c). As in the proof of Lemma 2.3 one can write I =
xK(−1)+y m OP2 (−m), K a B(3; k)-invariant ideal of colength c, κ = reg(K), d = colength(I) =
c + m, and again m ≥ κ + 3 (Corollary 2.1). We represent the Hilbert function ψ by the
ideal K = xJ (−1) + y κ OP2 (−κ), J a B(3; k)-invariant ideal of colength b and of regularity ε, where κ ≥ ε + 3 (cf. Corollary 2.1). If the Hilbert function of J is denoted by
ϑ, then in principle one has the situation represented by Fig. 2.12. If one assumes that
(m−1) −κ ≤ colength(J ) = b, then one could bring the monomials denoted by 1, 2, 3, . . .
in the positions denoted by 1, 2, 3, . . . (cf. Fig. 2.12). In the course of this the Hilbert
function increases and therefore g ∗ (ϕ) < g ∗ (ϕ1 ) < · · · < g(d), contradiction. Thus one
has (m − 1) − κ > b, i.e., m ≥ κ + b + 2 = c + 2.
2.3.5. Summary.
Lemma 2.4. (Notations as in Lemma 2.1 and Lemma 2.2) If g(d) < g ∗(ϕ), then
m ≥ c + 2.
From the proof of Lemma 2.4 we can conclude one more statement:
Corollary 2.2. Let be g ∗(ψ) ≤ g(c) (which notation implies c ≥ 5). Then m ≥
2c + 1.
2.4. Additional group operations
2.4.1. General auxiliary lemmas. The group Gl(3; k) operates on S = k[x, y, z],
and therefore on H d = Hilbd (P2k ). If ρ = (ρ0 , ρ1 , ρ2 ) ∈ Z3 is a vector such that ρ0 +ρ1 +ρ2 =
0, then
:= { (λ0 , λ1 , λ2 ) ∈ (k ∗ )3 | λρ00 λρ11 λρ22 = 1 } is a subgroup of T = T (3; k), and
nT (ρ) o
1 0 ∗
0 1 ∗
is a subgroup of U(3; k). We let the additive group Ga operate on S by
Γ :=
0 0 1
ψα : x 7→ x, y 7→ αx + y, z 7→ z, α ∈ k,
and σ : Gm → T nearly always denotes the operation σ(λ) : x 7→ x, y 7→ y, z 7→ λz, λ ∈ k ∗ .
Auxiliary Lemma 1. If V ⊂ Sd is a vector space, invariant under G := Γ · T (ρ)
where ρ = (0, −1, 1) or ρ = (−1, 0, 1), then V is monomial, i.e. invariant under T .
Proof. We first consider the case ρ = (0, −1, 1). We take a standard basis of V
consisting of T (ρ)-semi-invariants (see Appendix E). Assuming that the assertion above
is wrong , we conclude that there is a T (ρ)-semi-invariant f ∈ V , such that the monomials
occurring in f are not in V . Then there is such a form with smallest z-degree. From the
invariance of V under T (ρ) it follows that V is invariant under the Gm -action τ (λ) : x 7→
P
λx, y → y, z 7→ z, λ ∈ k ∗ . We write f = Mp, where M = xℓ y m, p = ni=0 ai y n−i z i , ℓ + m +
P
n = d, n ≥ 1 and an 6= 0. It follows that y∂f /∂z = yM n1 iai y n−i z i−1 ∈ V . Now y∂f /∂z
is also a T (ρ)-semi-invariant with smaller z-degree than f . According to the choice of f
11
it follows that g := Myz n−1 ∈ V , therefore y∂g/∂z = (n − 1)My 2 z n−2 ∈ V , etc. One gets
My i z n−i ∈ V, 1 ≤ i ≤ n, therefore My n ∈ V , contradiction.
P
In the case ρ = (−1, 0, 1) we write f = xℓ y m ni=0 ai xn−i z i . Because of x∂f /∂z =
P
xM ni=1 iai xn−i z i−1 we can argue as before.
Auxiliary Lemma 2. Let be I ⊂ OP2 an ideal of colength d, which is invariant under
G := Γ · T (ρ). If ρ0 + ρ1 + ρ2 = 0, ρ0 < 0 and ρ1 < 0, then I is invariant under T .
Proof. Let be n the smallest natural number, such that H 0 (I(n)) is not T -invariant.
Then we have without restriction that n ≥ 1. As H 0 (I(n)) has a standard basis , there
is a proper semi-invariant in H 0 (I(n)), i.e. a form f ∈ H 0 (I(n)) of the shape f =
M(1+a1 X ρ +a2 X 2ρ +· · ·+ar X rρ ), M a monomial, ar 6= 0, r ≥ 1, and no monomial MX iρ is
in H 0 (I(n)), if ai 6= 0. If M would be divisible by z, then g := z −1 f ∈ H 0 (I(n−1)) would
be a proper semi-invariant, too, because from z −1 MX iρ ∈ H 0(I(n − 1)) it follows that
MX iρ ∈ H 0 (I(n)). Therefore, M is not divisible by z. From the proper semi-invariants of
H 0 (I(n)) we chose one, say f , such that the z-degree is minimal. Now from f ∈ H 0 (I(n)),
because of the Γ-invariance, it follows that x∂f /∂z = xM(a1 ρ2 z −1 X ρ +· · ·+rar ρ2 z −1 X rρ)
and y∂f /∂z = yM(a1 ρ2 z −1 X ρ + · · · + rar ρ2 z −1 X rρ ) is in H 0 (I(n)), i.e., g := xMX ρ z −1 p
and h := yMX ρ z −1 p are in H 0 (I(n)), where p(X ρ ) := a1 ρ2 +2a2 ρ2 X ρ +· · ·+rar ρ2 X (r−1)ρ .
As the z-degree of g and of h is smaller than the z-degree of f, g and h are no longer
proper semi-invariants, i.e. the monomials which occur in g or in h, all are in H 0 (I(n)). It
follows that u := z −1 xMX rρ and v := z −1 yMX rρ are in H 0 (I(n)). From the Γ-invariance
|ρ1 |
|ρ |
it follows by applying the operators x∂/∂z and y∂/∂z repeatedly, that xz |ρ00| · yz |ρ1 | · MX rρ ∈
H 0 (I(n)).
Now X ρ = x−|ρ0 | y −|ρ1 | z ρ2 and ρ0 + ρ1 + ρ2 = 0, therefore MX (r−1)ρ ∈ H 0 (I(n)). Applying
the operators mentioned before again gives MX (r−2)ρ ∈ H 0 (I(n)), etc. It follows that
MX iρ ∈ H 0 (I(n)), 0 ≤ i ≤ r − 1, and therefore MX rρ ∈ H 0 (I(n)), contradiction.
2.4.2. Successive construction of Γ-invariant ideals. At first we consider a general situation: Let be K ⊂ OP2 an ideal of colength c and of regularity e; z is supposed
to be a non-zero divisor of OP2 /K; let be R := k[x, y] and ℓ ∈ R1 a linear form. Let be
m > e an integer and f ∈ H 0 (K(m)) a section, whose leading term is not divisible by ℓ,
i.e., if one writes f = f 0 + zf 1 + · · · + z m f m , where f i ∈ Rm−i , then f 0 is not divisible by
ℓ.
Lemma 2.5. The ideal I := ℓK(−1) + f OP2 (−m) has the following properties:
(i) z is not a zero-divisor of OP2 /I.
(ii) H 0 (I(n)) = ℓH 0 (K(n − 1)), if n < m, and
H 0 (I(n)) = ℓH 0 (K(n − 1)) ⊕ f k[x, z]n−m , if n ≥ m and ℓ = αx + y
(respectively H 0 (I(n)) = ℓH 0 (K(n − 1)) ⊕ f k[y, z]n−m , if n ≥ m and ℓ = x).
(iii) colength(I) = c + m, reg(I) = m.
Proof. If ℓ = x, these are the statements of ([G4], Lemma 4, p. 660). If ℓ =
αx + y, α ∈ k ∗ , then let be u the automorphism x 7→ ℓ, y 7→ y, z 7→ z of S. By applying
12
(loc.cit) to
one gets
Now applying u gives
K := u−1 (K), I := u−1(I), f := u−1 (f )
H 0 (I(n)) = xH 0 (K(n − 1))) ⊕ f k[y, z]n−m .
H 0(I(n)) = ℓH 0 (K(n − 1)) ⊕ f k[y, z]n−m .
As k[y, z] = k[ℓ − αx, z] and ℓf ∈ ℓH 0 (K(m)), the statement (ii) follows, if α 6= 0.
If α = 0, we take the automorphism x 7→ y, y 7→ x, z 7→ z and argue as before.
We would like to put the section f in a certain normal form. We first consider the
case ℓ = αx + y. Then we can write f 0 = xm + ℓu, u ∈ Rm−1 , without restriction. As
m − 1 ≥ e, there is v = v 0 + zv 1 + z 2 v 2 + · · · ∈ H 0 (K(m − 1)) such that v 0 = u. As f is
determined only modulo ℓH 0 (K(m − 1)), we can replace f by f˜ := f − ℓv, therefore we
can assume without restriction, that f = f 0 + zf 1 + · · · + z m f m , where f 0 = xm .
We now suppose that K is invariant under Γ, and will formulate conditions that I
is Γ-invariant, too. This is equivalent to the condition that f is Γ-invariant modulo
ℓH 0 (K(m − 1)). By ([T2], Hilfssatz 1, p. 142) this is equivalent to the condition that
hx, yi∂f /∂z ⊂ ℓH 0 (K(m − 1)). It follows that ℓ is a divisor of f i , 1 ≤ i ≤ n, i.e., one has
f = xm + ℓzg, g ∈ Sm−2 .
Write g = g 0 + zg 1 + z 2 g 2 + · · · , where g i ∈ Rm−2−i and choose u = u0 + zu1 + · · · ∈
H 0 (K(m − 2)) such that u0 = g 0 . This is possible, if m − 2 ≥ e. As f is determined only
modulo ℓH 0 (K(m − 1)), one can replace f by f˜ = f − ℓzu. It follows that one can assume
without restriction f = xm + ℓz 2 g, where g ∈ Sm−3 .
Choose u ∈ H 0 (K(m − 3)), where u0 = g 0 ; this is possible, if m − 3 ≥ e. If this is the
case, replace f by f˜ = f − ℓz 2 u. It follows that one can assume without restriction f =
xm +ℓz 3 g, where g ∈ Sm−4 , etc. Finally one obtains f = xm +z m−e ℓg, ℓ = αx+y, g ∈ Se−1 ,
and the Γ-invariance of f modulo ℓH 0 (K(m−1)) is equivalent to hx, yi[(m−e)z m−e−1 ℓg +
z m−e ℓ∂g/∂z] ⊂ ℓH 0 (K(m − 1)), i.e. equivalent to:
(2.1)
hx, yi[(m − e)g + z∂g/∂z] ⊂ H 0 (K(e))
In the case ℓ = x, because of Rm = xRm−1 ⊕ y m · k, one can write f 0 = y m + xu, and
the same argumentation shows that one can write f = y m + z m−e xg, g ∈ Se−1 , and the
Γ-invariance can again be expressed by the inclusion (2.1).
2.4.3. Standard forms. Let I ⊂ OP2 have the colength d and Hilbert function ϕ,
and let I be invariant under G := Γ · T (ρ), where ρ2 > 0. Moreover, we assume that
g ∗ (ϕ) > g(d). By Lemma 2.2 it follows that I = ℓK(−1)+f OP2 (−m), if k = k is supposed.
As H 0 (I(m − 1)) = ℓH 0 (K(m − 2)) is then invariant under G and m − 1 > e = reg(K),
it follows that hℓi and K are G-invariant. Assume that hax + by + zi is Γ-invariant. But
then hax + by + zi = h(a + α)x + (b + β)y + zi, ∀ α, β ∈ k, which is not possible. Thus we
have ℓ = ax + by. From hλ0 ax + λ1 byi = hax + byi, ∀ (λ0 , λ1 , λ2 ) ∈ T (ρ) it follows that
λ0 /λ1 = 1 ∀ (λ0 , λ1 , λ2 ) ∈ T (ρ), if a and b both were different from 0. But then it would
13
follow T (ρ) ⊂ T (1, −1, 0), and therefore ρ2 = 0, contradiction. Therefore we have ℓ = x
or ℓ = y, without restriction.
We consider the case ℓ = x, for example. As it was shown in (2.4.2) we can write
f = y m + z m−e xg, e = reg(K), g ∈ Se−1 .
From Appendix E it follows that xH 0 (K(m − 1)) has a standard basis of T (ρ)-semiinvariants fi = mi pi (X ρ ), i.e., mi is a monomial, pi is a polynomial in one variable with
constant term 1, and such that mi does not occur in fj any longer, if i 6= j. Now each fi is
divisible by x, therefore y m does not occur in fi . If the initial monomial mi of fi appears
in f , then mi has to appear in z m−e xg. By choosing α ∈ k in a suitable way, one can
achieve that mi does not occur in f˜ := f − αfi . As ρ2 > 0 and fi is divisible by x, f˜ still
has the shape y m + z m−e xg̃, g̃ ∈ Se−1 . By repeating this procedure one can achieve that
none of the mi does occur in f = y m + z m−e xg (and f is still invariant under Γ modulo
xH 0 (K(m − 1)). The same argumentation as in the proof of the lemma in Appendix E
then shows that f is automatically a T (ρ)-semi-invariant with initial monomial y m , and
f together with the fi forms a standard basis of H 0 (I(m)). We summarize:
Lemma 2.6. Let be I ⊂ OP2k an ideal of colength d, with Hilbert function ϕ, which
is invariant under G = Γ · T (ρ), where ρ2 > 0. Assume that g(d) < g ∗(ϕ). (It is not
assumed that k = k.) Then I = xK(−1) + f OP2 (−m) or I = yK(−1) + f OP2 (−m),
where K is a G-invariant ideal with colength(K) = c, reg(K) = e, and c + m = d.
Moreover f ∈ H 0 (K(m)) can be written in the form f = y m + z m−e xg respectively in the
form f = xm + z m−e yg, where g ∈ Se−1 . We have hx, yi∂f /∂z ⊂ xH 0 (K(m − 1)) or
hx, yi∂f /∂z ⊂ yH 0(K(m − 1)), respectively, and each of these inclusions are equivalent
to the inclusion (2.1) in section (2.4.2). One has
(
xH 0 (K(n − 1))
if n < m,
0
H (I(n)) =
0
xH (K(n − 1)) ⊕ f k[y, z]n−m if n ≥ m,
respectively
(
yH 0(K(n − 1))
H 0 (I(n)) =
yH 0(K(n − 1)) ⊕ f k[x, z]n−m
if n ≤ m,
if n ≥ m.
If one chooses a standard basis {fi } of xH 0 (K(m−1)) or of yH 0(K(m−1)), respectively,
then one can choose f in such a way that f has the form and the properties mentioned
above and together with the fi ’s forms a standard basis of H 0 (I(m)).
Proof. If k = k this follows from the foregoing discussion. One has, e.g. I ⊗ k =
k
0
yK(−1)+f OP2 (−m), where K ⊂ OP2 ⊗k and f ∈ H (K(m)) have the properties mentioned
above. One sees that K = I ⊗ k : yOP2 ⊗k (−1) and therefore one has the exact sequence:
0 → (OP2 ⊗k /K)(−1) → OP2 ⊗k /I ⊗ k → OP2 ⊗k /I ⊗ k + yOP2⊗k (−1) → 0.
If K := I : yOP2 (−1), then the sequence
0 → (OP2 /K)(−1) → OP2 /I → OP2 /I + yOP2 (−1) → 0
14
is exact, too. Tensoring this sequence with k one obtains a commutative diagram
0
/
(OP2 ⊗k /K ⊗ k)(−1)
·y
/
OP2 ⊗k /I ⊗ k
/
OP2 ⊗k /I ⊗ k + yOP2⊗k (−1)
/
0
OP2 ⊗k /I ⊗ k
/
OP2 ⊗k /I ⊗ k + yOP2⊗k (−1)
/
0
0
/
OP2 ⊗k /K(−1)
/
with exact rows, where the first vertical arrow is obtained from the canonical injection
∼
K ⊗ k ֒→ K by tensoring with OP2 ⊗k (−1). It follows from this, that K ⊗ k → K.
∼
Because of H 0 (I(m)) ⊗ k → H 0 (I(m) ⊗ k) one obtains a standard basis of T (ρ)-semiinvariants of H 0 (I(m) ⊗ k) by tensoring a standard basis of H 0 (I(m)) with ⊗ 1k . As the
k
elements of a standard basis are uniquely determined up to constant factors, it follows
that f = f ⊗k 1k , where f ∈ H 0 (K(m)). Therefore f has the form xm + z m−e yg, g ∈
Se−1 , if f has the form xm + z m−e yg, g ∈ Se−1 ⊗ k. For reasons of dimension it follows
H 0 (I(n)) = yH 0(K(n − 1)), n < m, and H 0 (I(n)) = yH 0(K(n − 1)) + f k[x, z]n−m , n ≥ m.
As the G-invariance of K follows from the G-invariance of I, the remaining statements of
Lemma 2.6 follow by the same argumentation as in the case k = k.
Definition 2. The (uniquely determined) decomposition I = xK(−1)+f OP2 (−m) or
I = yK(−1)+f OP2 (−m) of Lemma 2.6 is called x-standard form or y-standard form of I,
respectively.( N.B. For the Hilbert function ϕ of I this definition implies that g(d) < g ∗ (ϕ)
is fulfilled.)
Corollary 2.3. Let be R = k[x, y]. If I has x-standard form (resp. y-standard
form), then xRm−2 (resp. yRm−2 ) is contained in H 0(I(m − 1)) and thus xRm−1 (resp.
yRm−1 ) is contained in H 0 (I(m)).
Proof. One has m − 2 ≥ c by Lemma 2.4 and thus xRm−2 ⊂ xH 0 (K(m − 2)) (resp.
yRm−2 ⊂ yH 0(K(m − 2))) by Appendix C, Remark 2.
Remark 2.3. If I has x-standard from or y-standard form, respectively, and if Gm acts
on S by σ(λ) : x 7→ x, y 7→ y, z 7→ λz, λ ∈ k ∗ , then I0 := lim σ(λ)I again has x-standard
λ→0
form or y-standard form, respectively. This follows from h0 (I0 (n)) = h0 (I(n)), n ∈ N (cf.
[G2], Lemma 4, p. 542), because then H 0 (I0 (n)) is generated by the initial monomials of
a standard basis of H 0 (I(n)), n ∈ N. The Figures 2.13a and 2.13b show the typical shape
of the pyramid formed by the initial monomials.
Remark 2.4. An ideal cannot have x-standard form and y-standard form at the same
time. For m = reg(I) and c = colength(I) are determined by the Hilbert function ϕ of
I. If I would have x-standard form as well as y-standard form, then E(H 0 (I0 (d))) has
the form shown in Figure 2.14a. As I and I0 have the same Hilbert function, ϕ has the
form shown in Figure 2.14b, and therefore g ∗ (ϕ) ≤ g ∗ (χ) = g(d), where χ is the Hilbert
function of (2.2.2).
Remark 2.5. (Notations and assumption as in Remark 2.3) I∞ := lim σ(λ)I has
λ→∞
again x-standard form or y-standard form, respectively. The reasoning is as follows:
15
The Hilbert function ϑ of I∞ (respectively the regularity µ of I∞ ) is ≥ ϕ (respectively
≥ m). This follows from the semicontinuity theorem. By Lemma 2.4 it follows that
m ≥ c + 2, therefore Rm−2 ⊂ H 0 (K(m − 2)) (cf. Appendix C, Remark 2). If I has xstandard form (respectively y-standard form), then xRm−2 ⊂ H 0 (I(m − 1)) (respectively
yRm−2 ⊂ H 0 (I(m − 1))) follows. Therefore xRm−2 ⊂ H 0 (I∞ (m − 1)) (respectively
yRm−2 ⊂ H 0 (I∞ (m − 1))). It follows that xRµ−2 ⊂ H 0 (I∞ (µ − 1)) (respectively yRµ−2 ⊂
H 0 (I∞ (µ − 1))). Now I∞ has x-standard form or y-standard form, in any case, and the
above inclusions show that I and I∞ have the same kind of standard form.
2.5. The type of a G-invariant ideal
Let I ⊂ OP2k be an ideal of colength d, with Hilbert function ϕ and invariant under
G = Γ · T (ρ), where ρ2 > 0 and k is an extension field of C.
2.5.1.
Definition 3. 1◦ I has the type (−1), if one of the following cases occurs:
1st case: g ∗ (ϕ) ≤ g(d), where d ≥ 5 by convention.
2nd case: 0 ≤ d ≤ 4
2◦ We now assume g ∗ (ϕ) > g(d), which notation implies d ≥ 5. Then one has
I = ℓ0 I1 (−1) + f0 OP2 (−m0 ), where (ℓ0 , I1 , f0 , m0 ) is as in Lemma 2.6. If I1 has type
(−1), then we say I has type 0. If I1 is not of type (−1), then I1 = ℓ1 I2 (−1)+f1 OP2 (−m1 ),
where (ℓ1 , I2 , f1 , m1 ) is as in Lemma 2.6. If I2 has the type (−1), then we say I has the
type 1, etc. As d = colength(I) = colength(I1 ) + m0 , etc., the colengths of the ideals in
question decrease and the procedure will terminate. We have
Lemma 2.7. If I has not the type (−1), then one has a sequence of decompositions
I =: I0 = ℓ0 I1 (−1) + f0 OP2 (−m0 ),
(2.2)
I1 = ℓ1 I2 (−1) + f1 OP2 (−m1 ),
·································
Ir−1 = ℓr−1 Ir (−1) + fr−1 OP2 (−mr−1 ),
Ir = ℓr K(−1) + fr OP2 (−mr ).
For a given ideal Ii , (ℓi , Ii+1 , fi , mi ) is
Ir+1 := K. If di and ϕi is the colength
the inequality g(di ) < g ∗ (ϕi ) is fulfilled.
Hilbert function, the regularity) of K is
defined as in Lemma 2.6, where 0 ≤ i ≤ r and
and the Hilbert function, respectively, of Ii , then
The ideal K has the type (−1). The colength (the
denoted by c (by ψ and κ, respectively).
3◦ We have already noted in Corollary 2.1 that the decompositions of I0 , I1 , · · · , are
uniquely determined, in essence. Therefore the number r is determined uniquely. It is
called the type of I. The types of the ideals occurring in (2.2), their Hilbert functions and
the numbers m0 , · · · , mr are uniquely determined by the Hilbert function ϕ.
16
2.5.2. Properties of an ideal of type r ≥ 0. The assumptions and notations are
as before, and we assume that I has the type r ≥ 0.
Lemma 2.8. In the decompositions (2.2) of I one has:
(a)
(b)
(c)
(d)
colength(Ii ) = colength(Ii+1 ) + mi , i.e., colength(Ii ) = c + mr + · · · + mi , 0 ≤ i ≤ r.
If r = 0, then m0 ≥ c + 2, where c = colength(K).
If r ≥ 1, then m0 ≥ c + 2 + mr + · · · + m1 = colength(I1 ) + 2.
If r ≥ 0, then m0 ≥ 2r (c + 2).
Proof. (a) follows from the decompositions (2.2) and from Lemma 2.6.
(b) follows from Lemma 2.4.
(c) As the statement only depends on the Hilbert functions of I0 , I1 , · · · , Ir+1 = K , one
can assume without restriction ℓi = x, 0 ≤ i ≤ r, and Ii is B(3; k)-invariant, 0 ≤ i ≤ r +1.
Then one has
I = xI1 (−1) + y m0 OP2 (−m0 ) and I1 = xI2 (−1) + y m1 OP2 (−m1 ),
where I2 = K, if r = 1.
We argue as in (2.3.4) and we orientate ourselves by Figure 2.15. If one would have
m0 − 1 − m1 ≤ colength(I2 ), then one could make the deformations
1 7→ 1, . . . , m0 − 1 − m1 7→ m0 − 1 − m1 .
Then we would get g ∗ (ϕ) < · · · ≤ g(d), contradiction, because from type r ≥ 0 it follows
that g ∗ (ϕ) > g(d). Therefore one has m0 − 1 − m1 > colength(I2 ). Now colength(I2 ) = c,
if r = 1 (respectively colength(I2 ) = c + mr + · · · + m2 , if r ≥ 2) as was shown in (a).
(d) If r = 0, this is statement (b). If r = 1, then by (b) and (c) it follows that
m0 ≥ (c + 2) + m1 ≥ 2(c + 2). Now we assume r ≥ 2. We argue by induction and assume
that mr ≥ c + 2, mr−1 ≥ 2(c + 1), . . . , m1 ≥ 2r−1 (c + 2). By (c) it follows that
m0 ≥ (c + 2) + (c + 2) + 2(c + 2) + · · · + 2r−1 (c + 2) = 2r (c + 2).
In the case r = 1, the statement (c) is valid even if K = OP2 , i.e., if c = 0. Because
then one can assume without restriction again
I = xI1 (−1) + y m0 OP2 (−m0 ) and I1 = xOP2 (−1) + y m1 OP2 (−m1 ),
and then, because of colength(I1 ) = m1 , by Lemma 2.4 it follows that m0 ≥ m1 + 2, i.e.,
one gets the statement (c).
Corollary 2.4 (of the proof of Lemma 2.8). If I has type r ≥ 1, then mj + j <
mi + i − 1 for all 0 ≤ i < j ≤ r.
Proof. We refer to Lemma 2.8 and use the same notations. If the sequence of
decompositions (2.2) in (2.5.1) begins with Ii instead of I0 , then from Lemma 2.8c it
follows that mi−1 ≥ c + 2 + mr + · · · + mi . It follows that mi−1 − mi ≥ 2, therefore
mi−1 + (i − 1) > mi + i. One gets
mr + r < mr−1 + (r − 1) < · · · < m1 + 1 < m0 .
17
If mj + j = mi + (i − 1), then it would follow j = i + 1 and therefore mi+1 = mi − 2. We
will show, that this is not possible.
The ideal Ii has the type r −i ≥ 1, the colength c + mr + · · ·+ mi = di , and the Hilbert
function ϕi of Ii fulfils the inequality g(di ) < g ∗(ϕi ) (cf. Lemma 2.7). As to the Hilbert
function ϕi , one obtains the situation of Figure 2.16, which corresponds to the Hilbert
function ϕ. But then it follows that g ∗ (ϕi ) ≤ g ∗ (χ′ ) = g(di ), where the graph of χ′ is
denoted by a dotted line (cf. the first case in (2.2.2), Figure 2.8, and the argumentation
in the proof of Lemma2.2).
18
Fig. 2.1
0
1
2
3
Fig. 2.2
α ... ... ... ... e
Fig. 2.3
0
1
2
3
4 ... ... ... ... d
Fig. 2.4
0
1
2
Fig. 2.5
0
1
2
3
4
ε ε+1 . . . . . . . . . m
Fig. 2.6a
0
1
2 ...
Fig. 2.6b
0
1
2
3
19
0
1
2
Fig. 2.7b
Fig. 2.7a
0
1
2
3 ...
Fig. 2.8
0
1
2
3
4 ...
Fig. 2.9
ye
ye
xy e−2
0
x2
1 2 3 ... ... ... ... ... ... e
20
0
x2
1 2 3
4 ... ... ... ... ... e
Fig. 2.10
u
v
0
1 . . . . . . . . . . . . . . . . . . . . . . . . . . . m−1 m
21
Fig. 2.11a
Fig. 2.11b
u
v
0
2 · · · · · · · · · · · ·m−1 m
1
0
1 ... ... n ... ... ... m
Fig. 2.12
6
5
4
3
ψ ′ (n − 1)
2
1
1
2
3
ϑ′ (n − 2)
4
5
6
0
1
2
3
4
5
6
7
8 ε+2 10 11 12 13 14 15 16 17 18 m
22
Fig. 2.13a
Fig. 2.13b
ym
xm
0
1
2
3
4
5 κ+1 7
8
0
1
2
3
Fig. 2.14a
0
1
2
3
4
5
4
5
6 κ+1 8
Fig. 2.14b
6
7 m
0
23
1
2
3
4
5
6
7 m
Fig. 2.15
m0 −1−m1
2
graph of ϕ′1 (n − 1)
1
1
graph of ϕ′2 (n − 2)
0
1
2
3
4
5
6
7
8
m2 +2 or κ+2
2
10 11 m1 +1 13 14 15 16 17 m0 19
#{monomials between graph of ϕ′2 (n − 1) and line y = x − 1} = colength(I2 )
24
mi
mi+1 +1
Fig. 2.16
25
CHAPTER 3
A rough description of ideals invariant under Γ · T (ρ)
It seems impossible to characterize ideals of colength d in OP2 , which are invariant
under G := Γ · T (ρ). If I is an ideal of type r in the sense of the definition in (2.5.1),
however, then one can give a rough description of the forms f0 , · · · , fr , which occur in the
decomposition (2.2) in (2.5.1). This description will be used later on for estimating the
so called α-grade of I which will be defined in the next chapter.
3.1. Assumptions
Let I ⊂ OP2 be an ideal of colength d, invariant under Γ · T (ρ), where ρ2 > 0, as
usual. We assume that I has the type r ≥ 0. Then according to Lemma 2.7 one has a
decomposition:
I =: I0 = ℓ0 I1 (−1) + f0 OP2 (−m0 ),
I1 = ℓ1 I2 (−1) + f1 OP2 (−m1 ),
(Z)
······························
Ii = ℓi Ii+1 (−1) + fi OP2 (−mi ),
······························
Ir = ℓr Ir+1 (−1) + fr OP2 (−mr ).
Using the notations of (loc.cit.) the ideal Ir+1 := K has the colength c and regularity κ.
If ℓi = x, then
H 0 (Ii (n)) = xH 0 (Ii+1 (n − 1)) + fi k[y, z]n−mi
and if ℓi = y, then
H 0 (Ii (n)) = yH 0(Ii+1 (n − 1)) + fi k[x, z]n−mi
(cf. Lemma 2.6). From (Z) follows
H 0 (I1 (mi + i − 1)) = ℓ1 H 0 (I2 (mi + i − 2)),
H 0 (I2 (mi + i − 2)) = ℓ2 H 0 (I3 (mi + i − 3)),
·································
H 0(Ii−1 (mi + 1)) = ℓi−1 H 0 (Ii (mi )),
H 0 (Ii (mi )) = ℓi H 0 (Ii+1 (mi − 1)) + hfi i,
for by Corollary 2.4 mi + 2 < mi−1 , i = 1, . . . , r. If one starts with H 0 (I1 (mi + i − 2)),
then one obtains a similar system of equations, whose last line is
H 0 (Ii (mi − 1)) = ℓi H 0 (Ii+1 (mi − 2)).
27
Conclusion 3.1. If 2 ≤ i ≤ r (if 1 ≤ i ≤ r, respectively), then H 0 (I1 (mi + i − 1)) =
ℓ1 · · · ℓi−1 H 0 (Ii (mi )) (H 0 (I1 (mi + i − 2)) = ℓ1 · · · ℓi H 0 (Ii+1 (mi − 2)), respectively).
3.2. Notations
We orientate ourselves by Figure 3.1, which shows the initial monomials of H 0 (I(m0 )).
The set of all monomials in Sm0 with z-degree ≥ m0 − (c + r) (with z-degree ≤ m0 − (c +
r + 1), respectively) is called the left domain and is denoted by LB (is called right domain
and is denoted by RB, respectively). The monomials in Sc−1 − in (H 0 (K(c − 1))) form a
basis of Sc−1 /H 0 (K(c − 1)). If we put ℓ = ℓ0 · · · ℓr , then ℓ[Sc−1 /H 0 (K(c − 1))] has a basis
consisting of the c monomials in ℓSc−1 − ℓ· in (H 0 (K(c − 1))).
The initial monomial of ℓ0 · · · ℓi−1 fi · z m0 −mi −i is Miup := xi−ι(i) y mi +ι(i) z m0 −mi −i or
Midown := xmi +i−ι(i) y ι(i) z m0 −mi −i , if ℓi = x or if ℓi = y, respectively. Here we have put, if
1 ≤ i ≤ r + 1, ι(i) := number of indices 0 ≤ j < i such that ℓj = y. We also put ι(0) = 0,
so the formulas give the right result for i = 0, too.
For instance, in Figure 3.1 we have
r = 5, ℓ0 = x, ℓ1 = y, ℓ2 = ℓ3 = x, ℓ4 = ℓ5 = y,
and therefore
ι(1) = 0, ι(2) = ι(3) = ι(4) = 1, ι(5) = 2, ι(6) = 3.
(N.B. ℓ0 · · · ℓi−1 = xi−ι(i) y ι(i) , if 0 < i ≤ r + 1.)
As always we assume ρ2 > 0. If Midown occurs, then Ii = yIi+1(−1) + fi OP2 (−mi ).
If Miup occurs, then Ii = xIi+1 (−1) + fi OP2 (−mi ). By Lemma 2.8 mi ≥ (c + 2) + mr +
· · · + mi+1 , if r ≥ 1 (m0 ≥ c + 2, if r = 0).The colength of Ii+1 equals c + mr + · · · + mi+1 ,
and therefore Rn ⊂ H 0 (Ii+1 (n)) if n ≥ mi − 2 ≥ c + mr + · · · + mi+1 in the case
r ≥ 1(Rn ⊂ H 0(K(n)) if n ≥ c in the case r = 0). This follows from Remark 2 in
Appendix C. The next generating element fi has the initial monomial Miup or Midown .
Conclusion 3.2. Suppose ρ2 > 0 and ρ1 is arbitrary.Then the columns of total degree
mi + i − 2 and mi + i − 1 (in the variables x and y ),which occur in in(H 0(I(m0 ))) , also
occur in H 0 (I(m0 )). Thus xMiup or yMidown , when it exists, is contained in I.
Proof. This follows from , Conclusion 3.1, the decompositions (Z) and the foregoing
discussion.
As always we assume ρ2 > 0. We have to distinguish between two cases (cf. 2.4.1
Auxiliary Lemma 1 and Auxiliary Lemma 2):
Main Case I: ρ0 > 0 and ρ1 < 0
Main Case II: ρ0 < 0 and ρ1 > 0
28
3.3. Description of f0 in Case I
If the initial monomial of f0 equals xm0 , then f0 = xm0 and ℓ0 = y. Therefore
we may assume that f0 is a proper semi-invariant with initial monomial y m0 . Then
I = I0 = xI1 (−1) + f0 OP2 (−m0 ), and we write f0 = y m0 + z µ x·G, µ ∈ N maximal, G a
T (ρ)-semi-invariant (cf. 2.4.3). If G0 is the initial monomial of G, then N : = z µ · x · G0
is called the vice-monomial of f0 , which by assumption is not equal to zero .
The inclusion (2.1) of (2.4.2) reads
(3.1)
hx, yi[µG + z ∂G/∂z] ⊂ H 0 (I1 (m0 − µ))
As for the initial monomial G0 of G, one has
(3.2)
hx, yiG0 ⊂ in(H 0 (I1 (m0 − µ))) = H 0 (in(I1 (m0 − µ)))
where in denotes the subspace or the initial ideal, respectively, generated by the initial
monomials. (Equality follows, for example, from in(I) = lim σ(λ)I and [G2], Lemma 3
λ→0
and 4.) Without restriction we may assume that G0 ∈
/ in(H 0 (I1 (m0 − µ − 1))), because
we can reduce f0 modulo xH 0 (I1 (m0 − 1)).
(N.B.: The same statements are analogously valid in the Case II, that means, if ρ1 > 0,
I = yI1 (−1) + f0 OP2 (−m0 ), f0 = xm0 + z µ yG, etc.)
It follows that one of the following cases occurs, where 1 ≤ i ≤ r:
1◦ N = Nidown := (z/x)Midown = xmi +i−ι(i)−1 y ι(i) z m0 −mi −i+1
2◦ N = Niup := (z/y)Miup = xi−ι(i) y mi +ι(i)−1 z m0 −mi +i+1
3◦ hx, yiN ⊂ ℓin(H 0 (K(m0 − r))), where ℓ := ℓ0 · · · ℓr and N is a monomial in
L := ℓSm0 −(r+1) − ℓin(H 0 (K(m0 − r − 1))).
Notabene: One has L = [ℓSc−1 − ℓ · in(H 0(K(c − 1)))]·z m0 −r−c , because of
m0M
−r−1
ν=c
z m0 −r−1−ν Rν ⊂ H 0 (K(m0 − r − 1)).
As µ is maximal and ρ2 > 0, G0 is not divisible by z. In the cases 1◦ and 2◦ we therefore
have µ = m0 − mi − i + 1. The case 3◦ will be treated in (3.3.6), and we assume the case
1◦ or 2◦ . Then (3.1) can be written as
hx, yi[µG + z ∂G/∂z] ⊂ H 0 (I1 (mi + i − 1)).
If we put h := ℓ1 · · · ℓi−1 , then h = xi−1−ι(i) y ι(i) , because of ℓ0 = x one has ι(i) = number
of indices j such that 0 < j < i and ℓj = y. If i = 1, then h := 1. By Conclusion 1
it follows from (3.1), that G is divisible by h, that means, one has G = hg, g ∈ Smi −1 .
Therefore we can write
f0 = y m0 + z µ xhg
and (3.1) is equivalent to:
(3.3)
hx, yi[µg + z ∂g/∂z] ⊂ H 0 (Ii (mi )).
29
3.3.1. We assume the case 1◦ . As Nidown occurs , Ii = yIi+1 (−1) + xmi OP2 (−mi ). If
g 0 is the initial monomial of the form g (cf. the inclusion (3.3)), then z µ xG0 = hg 0xz µ =
xi−ι(i) y ι(i) z µ g 0 = Nidown . Because of µ = m0 − mi − i + 1 it follows that g 0 = xmi −1 .
Representing the initial monomials of H 0 (Ii (mi )) in the same way as the initial monomials
of H 0 (I0 (m0 )), one sees that southwest of Nidown there is no further monomial which occurs
in f0 .
Conclusion 3.3. Let be ρ1 < 0. If the vice-monomial of f0 equals Nidown , then
f0 = y m0 + αNidown , where α ∈ k. As fi = Midown is a monomial, by Conclusion 3.2
it follows that all monomials which have the same z-degree as Mi and are elements of
in(H 0 (I(m0 ))) also are elements of H 0 (I(m0 )). Therefore we have (x, y)y m0 ⊂ I.
3.3.2. We assume the case 2◦ . Then y m0 X νρ = Niup , where ν > 0 and 1 ≤ i ≤ r.
This statement is equivalent to the equations:
(3.4)
νρ0 = i − ι(i), νρ1 = mi + ι(i) − 1 − m0 , νρ2 = m0 − mi − i + 1.
Lemma 3.1. If one has f0 = y m0 + αNiup + · · · , where 1 ≤ i ≤ r and α ∈ k − (0), then
ρ2 > mi+1 = reg(Ii+1 ), where we put Ir+1 := K and mr+1 := κ = reg(K). Especially Ij is
a monomial ideal for all j > i.
Proof. As we have assumed ρ1 < 0, by (2.4.1 Auxiliary Lemma 2) one has ρ0 > 0.
Therefore ν ≤ i. On the other hand ρ2 = (m0 − mi − i + 1)/ν ≥ (m0 − mi − i + 1)/i; thus
it suffices to show (m0 − mi − i + 1)/i > mi+1 , i.e.
(3.5)
m0 > mi + i · mi+1 + (i − 1).
We start with i = r, in which case one has to show m0 > mr + rκ + (r − 1). This is valid
if r = 0, so we may assume r ≥ 1. By Lemma 2.8 one has m0 ≥ c + 2 + mr + · · · + m1 ,
thus it suffices to show c + 2 + mr + · · · + m1 > mr + rκ + (r − 1). If r = 1, then this
inequality is equivalent to c + 2 > κ, and this is true. Therefore one can assume without
restriction that r > 1, and has to show
c + 2 + mr−1 + · · · + m1 > rκ + (r − 1).
Because of κ ≤ c it suffices to show 2 + mr−1 + · · · + m1 > (r − 1)(κ + 1) which is true
because of κ ≤ c < mr−1 < · · · < m1 (cf. Lemma 2.4 and Corollary 2.4).
Now we assume 1 ≤ i < r, in which case it suffices to show:
c + 2 + mr + · · · + m1 > mi + i · mi+1 + (i − 1)
⇐⇒
(c + 2) + (mr + · · · + mi+2 ) + (mi−1 + · · · + m1 ) > (i − 1)mi+1 + (i − 1).
On the left side of this inequality the second summand (the third summand, respectively)
does not occur, if i + 1 = r (if i = 1, respectively). If i = 1 and r = 2, the inequality
reads c + 2 > 0. If i = 1 and r ≥ 3, the inequality reduces to (c + 2) + mr + · · · + m3 > 0.
Thus the case i ≥ 2 remains, and it suffices to show mi−1 + · · · + m1 > (i − 1)(mi+1 + 1),
which is true because of mi+1 < mi < · · · < m1 (loc.cit.).
30
3.3.3. Description of the ideal Ii . The assumptions are the same as in Lemma
3.1, but we slightly change the notations and write I = Ii , K = Ii+1 , m = mi , e = reg(K).
(Thus e = mi+1 or e = κ.) We have the following situation: I = xK(−1) + f OP2 (−m) is
Γ · T (ρ)-invariant, K is monomial, ρ1 < 0, ρ2 > e = reg(K). From the results of (2.4.2)
and Lemma 2.6 it follows that f = y m + z m−e xg, where g ∈ Se−1 and
(3.6)
hx, yi[(m − e)g + z ∂g/∂z] ⊂ H 0 (K(e))
where f ∈ H 0 (K(m)) is determined only modulo xH 0 (K(m − 1) and we can assume that
f is a T (ρ)-semi-invariant. Then g is a T (ρ)-semi-invariant, too. Thus we can write
g = N(1 + a1 X ρ + · · · ), where N ∈ Se−1 is a monomial. If we assume a1 6= 0 for example,
then NX ρ ∈ Se−1 would have a z-degree ≥ ρ2 . As ρ2 > e (cf. Lemma 3.1 ), this is
impossible.
Conclusion 3.4. Under the assumptions mentioned above, the form g in (3.6) equals
a monomial N ∈ Se−1 − H 0 (K(e − 1)) such that hx, yiN ⊂ H 0 (K(e)). It follows that
(x, y)y m ⊂ I.
3.3.4. We go back to the notations of Lemma 3.1, i.e. one has N = Niup . As in the
case 1◦ one has µ = m0 − mi − i + 1. As y m0 is the initial monomial of f0 , ℓ0 equals x and
h := ℓ1 · · · ℓi−1 equals xi−1−ι(i) y ι(i) as before. If one puts g = µg + z∂g/∂z, then (3.3) can
be written as
(3.7)
hx, yig ⊂ xH 0 (Ii+1 (mi − 1)) + hfi i
where fi has the initial monomial y mi ; thus
(3.8)
g ∈ H 0 (Ii+1 (mi − 1)).
The initial monomial of g equals the initial monomial of g; by construction this is equal
to y mi −1 . (Proof: The initial monomial of xhgz µ = xGz µ is equal to Niup by assumption.
Thus the initial monomial g 0 of g fulfils the equation
xhg 0 z µ = xi−ι(i) y mi +ι(i)−1 z m0 −mi −i+1 .
As xh = xi−ι(i) y ι(i) and µ = m0 − mi − i + 1, it follows that g 0 = y mi −1 .) From (3.7) it
follows that
(3.9)
yg = xF + αfi , F ∈ H 0 (Ii+1 (mi − 1)), α ∈ k ∗ .
Now we can write fi = y mi + z mi −mi+1 xu (cf. the last sentence in (2.4.2)), where u is either
a monomial in Smi+1 −1 − H 0 (Ii+1 (mi+1 − 1)) (cf. Conclusion 2.4), or u = 0.
We consider the first case. Then it follows that z mi −mi+1 xu = βy mi X νρ , where β ∈ k ∗
and ν ≥ 1. As fi can be reduced modulo xH 0 (Ii+1 (mi − 1)), we have z mi −mi+1 xu ∈
/
mi −1
0
,
xH (Ii+1 (mi − 1)) without restriction. From (3.9) it follows that in g, except for y
mi −mi+1
the monomial u := z
xu/y occurs.
Suppose, there is another monomial v in g. Then one would have yv ∈
xH (Ii+1 (mi − 1)), and from (3.8) it would follow that v is an element of the monomial
subspace H 0 (Ii+1 (mi − 1)). Figure 3.2 shows H 0 (Ii (mi )) = xH 0 (Ii+1 (mi − 1)) + hfi i
marked with — in black and H 0 (Ii+1 (mi )) marked with — in blue. Suppose that
0
31
v ∈ H 0 (Ii (mi − 1)) = xH 0 (Ii+1 (mi − 2)). By construction v occurs in g, therefore
xhv = ℓ0 · · · ℓi v occurs in xG and z µ xhv occurs in z µ xG. On the other hand, z µ xhv ∈
z µ xℓ1 · · · ℓi−1 xH 0 (Ii+1 (mi − 2)) = z µ xH 0 (I1 (mi + i − 2)) ⊂ xH 0 (I1 (m0 − 1)), because ℓi = x and so Conclusion 3.1 can be applied. As f0 can be reduced modulo
xH 0 (I1 (m0 − 1)), one can assume without restriction, that v does not occur in g and
therefore does not occur in the inclusions (3.7) or (3.8), which have to be fulfilled. Thus
we can assume without restriction that v ∈ H 0 (Ii+1 (mi − 1)) − H 0 (Ii (mi − 1)). From
yv ∈ xH 0 (Ii+1 (mi − 1)) it follows that v is equal to one of the monomials in Figure 3.2,
which are denoted by ?. Therefore the z-degree of v is ≥ mi − mi+1 .
By construction, the z-degree of u is ≥ mi − mi+1 , too. As u and v occur in the semiinvariant g, both monomials differ by a factor of the form X νρ . As ρ2 > mi+1 (Lemma
3.1), it follows that ν = 0, i.e., u and v differ by a constant factor. Therefore one has
g = βy mi−1 + γz mi −mi+1 xu/y ; β, γ ∈ k ∗ .
We have to describe the position of xu more exactly. From xu ∈
/ xH 0 (Ii+1 (mi+1 − 1)) but
hx, yixu ∈ xH 0 (Ii+1 (mi+1 )) and from the Γ-invariance of Ii+1 it follows that z m0 −mi+1 xu
equals Njdown or Njup , where j is an index r ≥ j > i, or equals a monomial L ∈ L (cf. 3.3).
Suppose z m0 −mi+1 ·xu = Njdown . Then u = z mi −mi+1 xu/y is southwest of Njdown /z m0 −mi
and does not occur in the monomial subspace H 0 (Ii+1 (mi − 1)), which contradicts (3.8).
Finally we note that the monomials of g agree with the monomials of g.
Conclusion 3.5. Assume that f0 has the vice-monomial Niup and that fi is not a
monomial. Then (up to a constant factor) fi = Miup + αNjup , where 1 ≤ i < j ≤ r and
α ∈ k ∗ , or fi = Miup + αL, where L ∈ L is a monomial such (x, y)L ⊂ ℓK(−r − 1), and
α ∈ k∗ .
Then it follows that f0 = y m0 + βNjup + γNjup · (z/y) or f0 = y m0 + βNiup + γL · (z/y),
respectively. Here β and γ are elements of k ∗ . Finally, all monomials, which occur in x f0
or in y 2 f0 also occur in I.
Proof. The shape of f0 and of fi results from the forgoing argumentation. Conclusion
3.4 gives (x, y)L ⊂ ℓK(−r − 1). The statements concerning the monomials in xf0 follow
from Conclusion 3.2. By Lemma 3.1, Ij is monomial, thus yNjup = z Mjup ∈ I and
therefore y Miup ∈ I. Furthermore y 2Njup · (z/y) = y Njup z ∈ I and as well y 2L · (z/y) =
yLz ∈ I. From this the assertion concerning the monomials of y 2 f0 follows.
Now we come to the case that fi = y mi . The above reasoning had shown that y mi −1
occurs in g. Suppose that another monomial v occurs in g. The same argumentation as
before shows that v equals one of the monomials in Figure 3.2 denoted by ?. Thus v is
equal to Njup or Njdown for an index i < j ≤ r, or is equal to a monomial L ∈ L, such that
(x, y)L ⊂ ℓK(−r − 1). Furthermore, there can be only one such monomial v.
Conclusion 3.6. If f0 has the vice-monomial Niup and if fi = y mi , then f0 = y m0 +
αNiup +βNjup or f0 = y m0 +αNiup +βNjdown , where 1 ≤ i < j ≤ r, or f0 = y m0 +αNiup +βL,
32
where L ∈ L is a monomial such that (x, y)L ⊂ ℓK(−r − 1). All monomials occurring in
xf0 or yf0 also occur in I.
Proof. The statements concerning the shape of f0 follow from the foregoing discussion. As Ii+1 is monomial by Lemma 3.1 and as fi is a monomial, Ii is monomial and
(x, y)L, (x, y)Niup and (x, y)Njdown are contained in I.
3.3.5. Suppose in the forms f0 and fj , where 1 ≤ j ≤ r , as in Conclusion 3.5
or Conclusion 3.6 there are three monomials with coefficients different from zero. We
call f0 (and fj , respectively) a trinomial and we then have f0 = y m0 + αNiup + βE0
and fj = Mjup + γNkup + δEj , where the “final monomials” E0 and Ej have the shape
described in Conclusion 3.5 and Conclusion 3.6, and where α, β, γ, δ ∈ k ∗ . If i > k,
then mi ≤ mk+1 < ρ2 (Lemma 3.1). As we are in the case ρ0 > 0, ρ2 > 0, it follows that
|ρ1 | > mi , and as Niup and E0 both occur in the semi-invariant f0 , one has E0 = Niup · X νρ ,
where ν ≥ 1. Looking at Figure 3.2, one sees that then E0 cannot be an element of Sm0 ,
contradiction. In the same way it follows that i < k is not possible, and thus we have
i = k. As f0 and fj are T (ρ)-semi-invariants, it follows that y m0 · X νρ = Mjup , where
ν ≥ 1. We conclude from this that
νρ0 = j − ι(j), νρ1 = mj + ι(j) − m0 , νρ2 = m0 − mj − j,
where 1 ≤ j ≤ r. We want to show that this implies ρ2 > mj+1 . Similarly as in the proof
of Lemma 3.1 it suffices to show m0 > mj + jmj+1 + j. We start with the case j = r.
Then again mr+1 := κ and one has to show m0 > mr + rκ + r. The case r = 0 cannot
occur. If r = 1, the inequality reads m0 > m1 + κ + 1, and because of m0 ≥ c + 2 + m1
(Lemma 2.8) and κ ≤ c this is right. Therefore we can assume r ≥ 2.
Because of (loc.cit.) it suffices to show c + 2 + mr + · · · + m1 > mr + rκ + r. As κ ≤ c it
suffices to show 2+mr−1 +· · ·+m1 > (r−1)κ+r ⇐⇒ 1+mr−1 +· · ·+m1 > (r−1)(κ+1).
Because of κ ≤ c < mr < · · · < m1 this is true (cf.Lemma 2.4 and Corollary 2.4).
We now assume 1 ≤ j < r, in which case it suffices to show:
c + 2 + mr + · · · + m1 > mj + jmj+1 + j
If j = 1 this inequality reads c + 2 + mr + · · · + m1 > m1 + m2 + 1, and this is true,
because r ≥ 2. Thus we can assume 2 ≤ j < r and the inequality which has to be shown
is equivalent to
(c + 1) + (mr + · · · + mj+2 ) + (mj−1 + · · · + m1 ) > (j − 1)(mj+1 + 1)
Because of mj+1 < mj < · · · < m1 (loc.cit.), this is true.
We thus have proved that from y m0 X νρ = Mjup
(3.10)
ρ2 > mj+1
follows. As k > j by definition, we have mk ≤ mj+1 < ρ2 and the same argumentation as
before carried out with Niup and E0 shows that Nkup and Ej cannot simultaneously occur
in fj .
33
Conclusion 3.7. Among the forms fi , 0 ≤ i ≤ r, which have the shape described in
Conclusion 3.5 and Conclusion 3.6, there can only occur at most one trinomial.
3.3.6. We now consider the case 3◦ . The following argumentation is first of all
independent of the sign of ρ1 . We write f0 = f 0 + z u g, where f 0 = y m0 if ℓ0 = x and
f 0 = xm0 , if ℓ0 = y. If we choose µ maximal then we have µ ≥ m0 − (κ + r). For as
Rκ ⊂ in(H 0 (K(κ))), the initial monomial of g has a z-degree > m0 − (κ + r + 1), i.e., the
initial monomial occurs in a column of total degree in x and y smaller or equal κ + r.
From the Γ-invariance of f0 modulo ℓ0 H 0 (I1 (m0 − 1)) it follows that hx, yi∂f /∂z =
hx, yi [µz µ−1 g + z µ ∂g/∂z] ⊂ ℓ0 H 0 (I1 (m0 − 1)). From the decompositions in (Z) (cf. 3.1)
we conclude that H 0 (Ii (n)) = ℓi H 0 (Ii+1 (n − 1)) if n < mi . Now
m0 − µ < κ + r + 1 < mr + r < · · · < m1 + 1 < m0
(cf. Corollary 2.4) and thus
H 0 (I1 (m0 − µ)) = ℓ1 H 0 (I2 (m0 − µ − 1)),
H 0 (I2 (m0 − µ − 1)) = ℓ2 H 0 (I3 (m0 − µ − 2)),
0
················································
H (Ir−1 (m0 − µ − r + 2)) = ℓr−1 H 0 (Ir (m0 − µ − r + 1)),
H 0 (Ir (m0 − µ − r + 1)) = ℓr H 0 (K(m0 − µ − r)).
It follows that H 0 (I1 (m0 − µ)) = ℓ1 · · · ℓr H 0 (K(m0 − µ − r)) and therefore
hx, yi [µg + z ∂g/∂z] ⊂ ℓH 0 (K(m0 − µ − r)) where ℓ := ℓ0 · · · ℓr = xa y b
and a (respectively b) is the number of ℓi = x (respectively the number of ℓi = y). This
implies that g is divisible by ℓ, and changing the notation we can write f0 = f 0 + ℓz µ g,
where ℓ = ℓ0 · · · ℓr , µ ≥ m0 − (κ + r) is maximal, g ∈ Sm0 −µ−r−1 and
hx, yi [µg + z ∂g/∂z] ⊂ H 0 (K(m0 − µ − r)).
Now f0 ∈ H 0 (K(m0 )) or f0 ∈ H 0 (I1 (m0 )) and m0 ≥ colength(K) + 2, if r = 0 or
m0 ≥ colength(I1 ) + 2, if r ≥ 1, respectively (cf. Lemma 2.6 and Lemma 2.8). Therefore
Rm0 ⊂ H 0 (K(m0 )) or Rm0 ⊂ H 0 (I1 (m0 )), respectively (cf. Appendix C, Remark 2). It
follows ℓz µ g ∈ H 0 (K(m0 )) (or ℓz µ g ∈ H 0 (I1 (m0 )) and thus ℓg ∈ H 0 (K(m0 − µ)) (or
ℓg ∈ H 0 (I1 (m0 − µ)) = ℓ1 · · · ℓr H 0 (K(m0 − µ − r)), respectively).
From this we conclude that ℓ0 g ∈ H 0 (K(m0 − µ − r)), in any case. We also note that
g∈
/ H 0 (K(m0 − µ − r − 1)) without restriction, because otherwise
ℓg ∈ ℓH 0 (K(m0 − µ − r − 1)) = ℓ0 H 0 (I1 (m0 − µ − 1)),
and thus z µ ℓg ∈ ℓ0 H 0 (I1 (m0 − 1)) would follow. But then ℓg could be deleted.
34
Conclusion 3.8. If the vice-monomial of f0 is different from all monomials Niup and
Njdown , then we can write
f0 = f 0 + z µ ℓg, where µ ≥ m0 − (κ + r)is maximal, ℓ := ℓ0 · · · ℓr ,
f 0 = y m0 , if ℓ0 = x (or f = xm0 , if ℓ0 = y, respectively),
g ∈ Sm0 −µ−r−1 − H 0 (K(m0 − µ − r − 1)), ℓ0 g ∈ H 0 (K(m0 − µ − r))
and hx, yi [µg + z ∂g/∂z] ⊂ H 0 (K(m0 − µ − r)).
3.4. Order of f0 in the case 3◦
We keep the notations of Conclusion 3.8, but now we write h := ℓ0 · · · ℓr . In order to
simplify the notations , the cases ρ1 > 0 and ρ1 < 0 will be treated separately. As always
we assume ρ2 > 0.
3.4.1. Let be ρ1 > 0 (and thus ρ0 < 0). Conclusion 3.8 is still valid, if x and y are
interchanged. Thus one can write f0 = xm0 + z µ hg, ℓ0 = y and
H 0(I(d)) = yH 0(I1 (d − 1)) ⊕ h{f0 xn z d−m0 −n |0 ≤ n ≤ d − m0 }i,
where Ii := K if r = 0 (cf. Lemma 2.6).
Suppose there is a 0 ≤ ν ≤ d−m0 such that xν+1 f0 ≡ xm0 +ν+1 modulo hH 0 (K(m0 +ν−r)).
Because of
hH 0 (K(m0 + ν − r)) = yℓ1 · · · ℓr H 0 (K(m0 + ν − r)) ⊂ yH 0(I1 (m0 + ν))
then follows that
H 0 (I(d)) = yH 0(I1 (d − 1))
⊕ h{f0 xn z d−m0 −n |0 ≤ n ≤ ν}i ⊕ h{xm0 +n z d−m0 −n |ν + 1 ≤ n ≤ d − m0 }i.
˙ α (H 0 (I(d)) (cf. Chapter 4), then
If one wants to determine the so called α-grade of ∧ψ
it is easier to estimate the contribution coming from the monomials in H 0 (I(d)). In the
following definition we assume ρ2 > 0 and the sign of ρ1 is arbitrary.
Definition 4. The order of f0 is the smallest natural number ν such that xν+1 f0 ≡
xm0 +ν+1 modulo hH 0 (K(m0 + ν − r)) (such that y ν+1 f0 ≡ y m0 +ν+1 modulo hH 0 (K(m0 +
ν − r)), respectively) if ρ1 > 0 (if ρ1 < 0, respectively).
Here we have put h := ℓ0 · · · ℓr (cf. (Z) in 3.1).
Remark 3.1. If ρ1 > 0, then ρ0 < 0, and from Lemma 2.6 it follows that f0 has
the initial monomial M0down = xm0 . Then Conclusion 3.2 gives yxm0 ∈ I. On the other
hand, if ρ1 < 0 the same argumentation shows that f0 has the initial monomial y m0 and
xy m0 ∈ I.
We now take up again the assumption ρ1 > 0 from the beginning of (3.4.1).
Subcase 1: ℓ0 = x. As ρ0 < 0 one has f0 = y m0 .
Subcase 2: ℓ0 = y. Putting h := xa y b one can write
f0 = xm0 + z µ hg = xm0 (1 + xa−m0 y b z µ g).
35
As f0 is a T (ρ)-semi-invariant, one has
xa−m0 y bz µ g = X γρ p(X ρ ), where γ ∈ N−(0), p(X ρ ) = a0 +a1 X ρ +. . .+as X sρ and ai ∈ k.
We now assume f0 is not a monomial. If both µ and γ are chosen maximal, then a0 6= 0,
and we can write g = Mp(X ρ ), where M := xm0 −a+γρ0 y γρ1 −b and µ = γρ2 . From the
Γ-invariance of f0 modulo ℓ0 H 0 (I1 (m0 − 1)) it follows that
hx, yihz µ−1 [µg + z∂g/∂z] ⊂ ℓ0 H 0 (I1 (m0 − 1)),
thus
hx, yiℓ1 · · · ℓr [µg + z∂g/∂z] ⊂ H 0 (I1 (m0 − µ)).
We had already obtained H 0 (I1 (m0 − µ)) = ℓ1 · · · ℓr H 0 (K(m0 − µ − r)) in (3.3.6).
We conclude that
hx, yi[µg + z∂g/∂z] ⊂ H 0 (K(m0 − µ − r))
and therefore:
xg ≡ − µ1 xM[ρ2 a1 X ρ + 2ρ2 a2 X 2ρ + · · · + sρ2 as X sρ ] modulo H 0 (K(m0 − µ − r))
Assume that s = degree (p) > 0. Then as 6= 0 and j := min {i > 0|ai 6= 0} is an integer
between 1 and s inclusive. Then one can write :
xg ≡ − µ1 xMX jρ [jρ2 aj + (j + 1)ρ2 aj+1 X ρ + · · · + sρ2 as X (s−j)ρ ] modulo H 0 (K(m0 − µ − r))
From this it follows that
xf0 = xm0 +1 + h z µ xg ≡ f˜ modulo hH 0 (K(m0 − r)),
where f˜ := xm0 +1 + h z µ̃ M̃ p̃(X ρ ) and µ̃ := µ + jρ2 > µ, M̃ := xMxjρ0 y jρ1 ,
p̃(X ρ ) := ã0 + ã1 X ρ + · · · + ãs−j X (s−j)ρ
and finally ã0 := − µ1 jρ2 aj 6= 0, · · · , ãs−j := − µ1 sρ2 as 6= 0.
Remark 3.2. f˜ is again a T (ρ)-semi-invariant, because
x−(m0 +1) hz µ̃ M̃ = x−m0 −1 xa y b z µ+jρ2 x · xm0 −a+γρ0 · y γρ1 −b · xjρ0 · y jρ1
= x(γ+j)ρ0 y (γ+j)ρ1 z (γ+j)ρ2 .
Remark 3.3. f˜ is only determined modulo ℓ0 H 0 (I1 (m0 )), as hH 0 (K(m0 − r)) ⊂
ℓ0 H 0 (I1 (m0 )).
We continue the above discussion and get at once xf0 ≡ xm0 +1 modulo hH 0 (K(m0 −r)),
if s = 0. In any case we have degree (p̃) < degree (p), and continuing in this way, we get
finally p̃·(s+1) = 0. This means, there is an integer 0 ≤ ν ≤ s such that
(3.11)
xν+1 f0 ≡ xm0 +ν+1 modulo hH 0 (K(m0 − r + ν)).
36
3.4.2. We now assume ρ1 < 0. If f0 is a monomial, then the order of f0 is zero
by definition. Thus we may assume without restriction that ρ0 > 0 (c.f. 2.4.1 Auxiliary
Lemma 2).
Subcase 1: ℓ0 = y. Then f0 = xm0 (Lemma 2.6).
Subcase 2: ℓ0 = x. With notations analogous to that of (3.4.1) one has: f0 = y m0 (1 +
xa y b−m0 z µ g) is a semi-invariant, therefore xa y b−m0 z µ g = X γρ p(X ρ ), γ ∈ N − (0). If f0 is
not a monomial and µ and γ are chosen maximal, then one can write g = Mp(X ρ ), where
µ = γρ2 , M = xγρ0 −a y m0 −b+γρ1 and p(X ρ ) = a0 +a1 X ρ +· · ·+as X sρ . Furthermore one gets
yf0 ≡ f˜ modulo hH 0 (K(m0 − r)), where f˜ := y m0 +1 + hz µ̃ M̃ p̃(X ρ ), µ̃ := µ + jρ2 , M̃ :=
yMxjρ0 y jρ1 and p̃ is defined as in (3.4.1). It is clear that one gets the corresponding
statements as in (3.4.1), if x and y as well as a and b are interchanged. This means, there
is an integer 0 ≤ ν ≤ s such that
(3.12)
y ν+1f0 ≡ y m0 +ν+1 modulo hH 0 (K(m0 − r + ν)).
3.4.3. Estimate of the order of f0 . The order has been defined as the smallest
natural number ν such that xν+1 f0 ≡ xm0 +ν+1 (respectively y ν+1f0 ≡ y m0 +ν+1 ) modulo
hH 0 (K(m0 − r + ν)). We keep the notations of (3.4.1) and (3.4.2). From γρ2 = µ ≥
m0 − (κ + r) (cf. 3.3.6) it follows that γ ≥ [m0 − (κ + r)]/ρ2 . On the other hand,
xm0 −a+γρ0 · xsρ0 has to be a polynomial in case that ρ1 > 0 ( y m0 −b+γρ1 · y sρ1 has to be
a monomial in case that ρ1 < 0, respectively). That means m0 − a + (γ + s)ρ0 ≥ 0
(m0 − b + (γ + s)ρ1 ≥ 0, respectively). Now one has ρ0 < 0, if ρ1 > 0, and ρ0 > 0, if
ρ1 < 0 (cf. 2.4.1 Auxiliary Lemma 2). It follows that
γ + s ≤ (m0 − a)/|ρ0 | if ρ1 > 0,
γ + s ≤ (m0 − b)/|ρ1 | if ρ1 < 0.
Here a (respectively b) is the number of ℓi , 0 ≤ i ≤ r, such that ℓi = x (respectively
ℓi = y).
In case that ρ1 > 0 we obtain , because of |ρ0 | = ρ1 + ρ2 :
s≤
m0 − a m0 − (κ + r)
−
ρ1 + ρ2
ρ2
And in case that ρ1 < 0, because of |ρ1 | = ρ0 + ρ2 , we obtain
s≤
m0 − b m0 − (κ + r)
−
.
ρ0 + ρ2
ρ2
From the congruences (3.11) and (3.12) we get
Conclusion 3.9. The order s of f0 fulfils the inequality:
a
κ − m0 1 − 1
+ ρr2 − ρ1 +ρ
if ρ1 > 0,
ρ2
ρ2
ρ1 +ρ2
2
s≤
b
κ − m0 1 − 1
+ ρr2 − ρ0 +ρ
if ρ1 < 0.
ρ2
ρ2
ρ0 +ρ2
2
37
3.5. Summary in the Case I
We recall that this means ρ0 > 0, ρ2 > 0 and ρ1 < 0.
Proposition 3.1. Let be ρ2 > 0 and ρ1 < 0.
(a) The following cases can a priori occur:
1st case: One of the fi , say fi0 has as vice-monomial an “upper empty corner” Njup
. Then
0
ρ2 > κ, thus K is monomial, and the fi have one of the following forms:
(1)
(2)
(3)
(4)
(5)
fi = Midown
fi = Miup + αNjdown
fi = Miup + αNjup + βNkup
fi = Miup + αNjup + βNkdown
fi = Miup + αNjup + βL, L ∈ L a monomial
such that
(x, y) · L ⊂ ℓK(−r − 1), ℓ := ℓ0 · · · ℓr .
(6) fi = Miup + αNjup + βNkup · (z/y), α, β ∈ k ∗
(7) fi = Miup + αNjup + βL · (z/y), L ∈ L a monomial
such that
(x, y) · L ⊂ ℓK(−r − 1), α, β ∈ k ∗ .
Here α and β are (a priori) arbitrary elements of k and 0 ≤ i < j < k ≤ r.
2nd case: Not any of the fi has as vice-monomial an “upper empty corner” Njup . Then
each of the fi has one of the following forms:
(1) fi = Miup + αNjdown
(2) fi = Midown
P
(3) fi = Miup + F, F = αj Lj , Lj ∈ L monomial, αj ∈ k ∗ , such that xF ∈ ℓK(−r −
P
1), and for suitable βj ∈ k ∗ and G := βj Lj one has (x, y) · G ⊂ ℓK(−r − 1).
(b) In the 1st case there is only one trinomial at most, i.e. there is only one fi at most,
which has the shape as in one of the cases 1.3 - 1.7, with α and β different from zero.
S
(c) Let be M :=
{monomials in H 0 (I(n))} and let be hMi the subspace of S generated
n≥0
by these monomials. In the 1st case one has:
• xfi ∈ hMi, for all 0 ≤ i ≤ r;
• yfi ∈ hMi, if fi has the shape as in one of the cases 1.1–1.5
• y 2 fi ∈ hMi, if fi has the shape as in one of the cases 1.6 and 1.7.
(d) In the 2nd case one has hx, yifi ⊂ hMi, if fi has the shape as in case 2.1.
If fi has the shape as in case 2.3, then xMiup ∈ M , and the order s of fi fulfils the
inequality
1
1
r
κ
− mi
−
+ ,
s≤
ρ2
ρ2 ρ0 + ρ2
ρ2
where 0 ≤ i ≤ r.
38
Proof. (1.1) and (1.2) result from Conclusion 3.2 and Conclusion 3.3; (1.3)–(1.5)
result from Conclusion 3.6; and (1.6), and (1.7) results from Conclusion 3.5. (2.1) results
again from Conclusion 3.2 and Conclusion 3.3; (2.2) is clear, and (2.3) results from Conclusion 3.8, for if Miup occurs, then ℓi = x, and this corresponds to the linear form ℓ0 in
Conclusion 3.8.
(b) results from Conclusion 3.7.
(c) results from the Conclusions 3.3, 3.5 and 3.6.
(d) results in the case (2.1) from the fact that fj = Mjdown is a monomial and therefore all
monomials with the same z-degree as Mjdown , which occur in in(H 0(I(m0 ))), also occur in
H 0 (I(m0 )) (Conclusion 3.2). In the case (2.3) the first part of the statement results from
Conclusion 3.2, and the statement concerning the order results from Conclusion 3.9.
Remark 3.4. If r = 0, then only the cases (2.2) and (2.3) can occur, i.e., one has
either f0 = xm0 or f0 = y m0 + F , where F has the properties mentioned above.
3.6. Summary in the Case II
We recall that this means ρ0 < 0, ρ1 > 0, ρ2 > 0.
We translate formally the results of (3.5): As always ρ2 > 0 is assumed, one has ρ0 < 0.
By definition ι(i) = #{ℓj = y|0 ≤ j < i} and therefore i − ι(i) = #{ℓj = x|0 ≤ j < i}.
We conclude that if one interchanges x and y in the monomial Miup (respectively Midown )
and if one simultaneously replaces ι(i) by i − ι(i) and vice versa i − ι(i) by ι(i), then one
obtains Midown (respectively Miup ). Therefore the statements of Proposition 3.1 can be
translated to the Case II by interchanging x and y and “up” and “down”:
Proposition 3.2. Let be ρ2 > 0 and ρ1 > 0.
(a) The following cases can a priori occur:
1st case: One of the fi , say fi0 has as vice-monomial a “lower empty corner” Njdown .
Then ρ2 > κ, thus K is monomial and the fi have one of the following forms:
fi = Miup
fi = Midown + αNjup
fi = Midown + αNjdown + βNkdown
fi = Midown + αNjdown + βNkup
fi = Midown + αNjdown + βL, L ∈ L a monomial
such that (x, y) · L ⊂ ℓK(−r − 1), ℓ := ℓ0 · · · ℓr
(6) fi = Midown + αNjdown + βNkdown · (z/x), α, β ∈ k ∗
(7) fi = Midown + αNjdown + βL · (z/x), L ∈ L a monomial
such that (x, y) · L ⊂ ℓK(−r − 1), α, β ∈ k ∗ .
(1)
(2)
(3)
(4)
(5)
2nd case: Not any of the fi has as vice-monomial a “ lower empty corner” Njdown . Then
each of the fi has one of the following forms:
(1) fi = Midown + αNjup
(2) fi = Miup
39
P
(3) fi = Miup + F, F =
αj Lj , Lj ∈ L monomial,αj ∈ k ∗ , such that yF ∈ ℓK(−r −
P
1), and for suitable βj ∈ k ∗ and G := βj Lj one has (x, y) · G ⊂ ℓK(−r − 1).
(b) The same statement as in Proposition 3.1.
(c) The same statement as in Proposition 3.1, if x and y are interchanged.
(d) The same statement as in Proposition 3.1, if one replaces xMiup by yMidown .
And the order s of fi fulfils the inequality
r
1
1
κ
+ ,
− mi
−
s≤
ρ2
ρ2 ρ1 + ρ2
ρ2
where 0 ≤ i ≤ r.
Remark 3.5. If r = 0, then only the cases (2.2) and (2.3) can occur.
40
Fig. 3.1
M0
M2
N2
M3
N3
N5 M 5
N4 M 4
N1 M 1
7
8
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
Explanation: reg(K) ≤ c ⇒ Rn ⊂ H 0 (K(n)), n ≥ c
Sc−1 /H 0 (K(c − 1)) has a basis of c monomials, namely the monomials
of Sc−1 − inH 0 (K(c − 1)) ⇒ ℓ[Sc−1 /H 0(K(c − 1))] has a basis
of c monomials, namely the monomials of ℓSc−1 − ℓ · inH 0 (K(c − 1))
They generate a subspace L. ℓ := ℓ0 · · · ℓr = xr+1−ι(r+1) y ι(r+1)
41
m0
6
m1 +1
5
m2 +2
4
m3 +3
3
m4 +4
2
m5 +5
1
κ+6
0
Fig. 3.2
y mi
Niup
Ii+1
?
?
Ii = xIi+1 (−1) + fi OP2 (−mi )
?
?
?
?
42
mi
mi+1 +1
mi+1
?
CHAPTER 4
The α-grade.
4.1. Notations.
We let Gm (respectively Ga ) operate on S = k[x, y, z] by σ(λ) : x → x, y 7→ y, z 7→ λz
(respectively by ψα : x 7→ x, y 7→ αx + y, z 7→ z).
m
m
V
V
Let be G = Grassm (Sn ). If V ∈ G(k), then V has the dimension 1 and V 7→ V
m
m
V
V
p
defines a closed immersion G → P( Sn ) = PN , N := dim Sn − 1, the so called Plücker
embedding.
If one numbers the monomials in Sn , then one gets a basis {e1 , e2 , · · · } of Sn , and
m
V
therefore e(i) = ei1 ∧ · · · ∧ eim is a basis of Sn , where (i) = (i1 , · · · , im ) runs through all
. If one puts
sequences of natural numbers, such that 1 ≤ i1 < · · · < im ≤ n+2
2
ψα (e(i) ) := ψα (ei1 ) ∧ · · · ∧ ψα (eim ), then Ga operates in an equivariant manner on G
and PN , and the same is valid for the operation σ of Gm . If ξ ∈ G(k) corresponds to
m
V
the vector space V ⊂ Sn , then Cξ := { ψα (V )|α ∈ k}− is a point or a curve in PN ,
and there are polynomials f0 , · · · , fN in one variable with coefficients in k, such that
Cξ = {(f0 (α) : · · · : fN (α))|α ∈ k}− . At least one of the fi is equal to 1, and if ξ is
not invariant under Ga , then Cξ is a Ga -invariant closed curve in PN of degree equal to
max{deg(fi )|0 ≤ i ≤ N} (cf. [T1], Bemerkungen 2 und 3, p. 11). This number is denoted
by α-grade (V ).
Let now be ei , 1 ≤ i ≤ ℓ := n+2
, the monomials in Sn , ordered in the inverse
2
ℓ
P
lexicographic manner. If fi =
aji ej , 1 ≤ i ≤ m, is a basis of V , then f1 ∧ · · · ∧ fm =
j=1
a
·
·
·
a
i
1
i
m
1
1
P
P(i) e(i) , where P(i) = det · · · · · · · · · · · · is the Plücker coordinate for the index
(i)
aim 1 · · · aim m
m
V
P
(i) = (i1 , · · · , im ). It follows that ψα ( V ) = h P(i) ψα (e(i) )i and we conclude from this:
(i)
α − grade (V ) ≤ max{α − grade ψα (e(i) ) | P(i) 6= 0}
(4.1)
(i)
where we define the α-grade of ψα (e(i) ) to be the α-grade of the monomial subspace
ℓ
P
pjν (α)ej , 1 ≤
hei1 , · · · , eim i of Sn . This can be computed as follows: Write ψα (eiν ) =
j=1
ν ≤ m, where the pνj are polynomials in one variable with coefficients in Z. Then
P
ψα (e(i) ) = P(j) (α)e(j) , where the P(j) (α) are the Plücker coordinates of the vector space
(j)
43
hψα (ei1 ), · · · , ψα (eim )i. The P(j) are polynomials in one variable with coefficients in Z. As
P(i) (α) = 1, the α-grade of hei1 , · · · , eim i is equal to max{deg(P(j) )}.
(j)
Whereas it seems practically impossible to find a formula for the α-grade of an arbitrary vector space V ⊂ Sn , the α-grade of a monomial subspace V ⊂ Sn can be computed
n
L
as follows: Write V =
z n−i Vi where Vi ⊂ Ri is a monomial subspace and R = k[x, y].
i=0
If we put m(i) := dim Vi then Vi has a basis of the form {xi−aij y aij |1 ≤ j ≤ m(i)}, where
n
P
0 ≤ ai1 < · · · < aim(i) ≤ i is a sequence of integers. As α-grade (V ) =
α-grade (Vi ), we
i=0
can consider V as a graded vector space in R, which has a basis of monomials of different
degrees. In ([T1], 1.3, p. 12f.) it was shown:
If 0 ≤ c1 < · · · < cr ≤ i are integers, and
W := hxi−c1 y c1 · · · , xi−cr y cr i ⊂ Ri ,
then α-grade (W ) = (c1 + · · · + cr ) − (1 + · · · + r − 1).
(4.2)
Later on we will need an estimate of the α-grade of an ideal I ⊂ OP2 of colength d,
which is invariant under G = Γ · T (ρ). This will be done by estimating the α-grade of
the vector space V = H 0 (I(n)), if n is sufficiently large. By means of (4.1) the estimate
will be reduced to the computation of the α-grade of monomial subspaces, and because
of (4.2) this can be regarded as a combinatorial problem, the formulation of which needs
some more notations.
4.2. The weight of a pyramid.
Definition 5. A pyramid with frame c and colength d, 1 ≤ d ≤ c, is a set P of
monomials in R = k[x, y] with the following properties: The i-th “column” Si of P
consists of monomials xi−aij y aij , 0 ≤ ai1 < . . . < aim(i) ≤ i, for all 0 ≤ i ≤ c − 1, such that
the following conditions are fulfilled:
(1)
c−1
S
Si = P
i=0
(2) #({xi−j y j |0 ≤ j ≤ i ≤ c − 1} \ P ) = d
Then w(Si ) := (ai1 + · · · + aim(i) ) − (1 + · · · + m(i) − 1) is called the weight of the i-th
c−1
P
column Si , and w(P ) :=
w(Si ) is called the weight of the pyramid P .
i=0
Example. Let be I ⊂ OP2 an ideal of colength d which is invariant under T (3; k).
Then reg(I) ≤ d, therefore h0 (I(d − 1)) = d+1
− d, and we can write H 0 (I(d − 1)) =
2
d−1
L d−1−i
z
Vi , where Vi ⊂ Ri are monomial subspaces. If Si is the set of monomials in
i=0
Vi , then P :=
d−1
S
Si is a pyramid with frame and colength equal to d. From (4.2) one
i=0
concludes that w(P ) = α-grade (H 0 (I(d − 1)). (N.B. One has Rn ⊂ H 0 (I(n)) if n ≥ d.)
44
Remark 4.1. Let 0 ≤ c1 < · · · < cr ≤ i be a sequence of integers. w(c1, · · · , cr ) :=
(c1 + · · · + cr ) − (1 + · · · + r − 1) is maximal if and only if cν = i − r + ν, 1 ≤ ν ≤ r, i.e.
if (c1 , · · · , cr ) = (i − r + 1, · · · , i).
The aim is to determine those pyramid P of type (c, d), i.e., with frame c and colength
d, for which w(P ) is maximal. Because of Remark 4.1 we will consider without restriction
only pyramids with Si = {xi−a(i) y a(i) , · · · , xy i−1 , y i }, where a(i) := i+1−m(i) is a number
between 0 and i inclusive. We call xi−a(i) y a(i) the initial monomial and a(i) initial degree of
Si . For simplicity we write Si = (a(i), a(i) + 1, · · · , i) and P = hxi−a(i) y a(i) |0 ≤ i ≤ c − 1i.
Remark 4.2. w(Si ) = ia(i) + a(i) − a(i)2 .
Proof. w(Si ) = (a(i) + · · · + i) − (1 + · · · + i − a(i)) = (1 + · · · + i) − (1 + · · · + a(i) −
a(i)
i−a(i)+1
1) − (1 + · · · + i − a(i)) = i+1
−
−
, and a direct computation gives the
2
2
2
assertion.
Taking away xi−a(i) y a(i) from Si and adding xj−a(j)+1 y a(j)−1 to Sj , if Si 6= ∅, a(j) > 0
and j 6= i, then gives a pyramid
P ′ = P − {xi−a(i) y a(i) } ∪ {xj−a(j)+1 y a(j)−1 }.
We express this as Si (P ) = (a(i), · · · , i), Sj (P ) = (a(j), · · · , j), Si (P ′ ) = (a(i) + 1, · · · , i),
Sj (P ′ ) = (a(j) − 1, · · · , j) and we get w(Si (P )) = ia(i) + a(i) − a(i)2 ; w(Si (P ′ )) =
i(a(i) + 1) + (a(i) + 1) − (a(i) + 1)2 ; w(Sj (P )) = ja(j) + a(j) − a(j)2 , w(Sj (P ′ )) = j(a(j) −
1) + a(j) − 1 − (a(j) − 1)2 . It follows that w(P ′) − w(P ) = [w(Si (P ′)) − w(Si (P ))] +
[w(Sj (P ′ )) − w(Sj (P ))] = [i + 1 − 2a(i) − 1] + [−j − 1 + 2a(j) − 1]. We get the following
formula:
(4.3)
w(P ′) − w(P ) = 2(a(j) − a(i)) − (j − i) − 2,
where we have made the assumption that i 6= j, Si (P ) 6= ∅, and a(j) > 0.
Now let P be a pyramid of type (c, d), such that w(P ) is maximal. Then for all
deformations P 7→ P ′ as above, the right side of (4.3) is ≤ 0, i.e. a(j) − a(i) ≤ 12 (j − i) + 1.
If i < j (if i > j, respectively) this is equivalent to
1
1
a(j) − a(i)
1
1
a(j) − a(i)
≤ +
and
≥ +
, respectively.
j−i
2 j−i
j−i
2 j−i
Putting in j = i + 1 (respectively j = i − 1) gives a(i + 1) − a(i) ≤ 1.5 (a(i) − a(i − 1) ≥
−0.5, respectively). As the left side of these inequalities are integers, it follows that
a(i + 1) − a(i) ≤ 1 (a(i) − a(i − 1) ≥ 0, respectively) for all 0 ≤ i < c − 1 (for all
0 < i ≤ c − 1, respectively). Note that we can apply (4.3) only under the assumption
a(i + 1) > 0 (respectively a(i − 1) > 0). But if a(i + 1) = 0 (respectively a(i − 1) = 0),
then the two last inequalities are true, too. So we get
Remark 4.3. If the pyramid P of type (c, d) has maximal weight, then one has
a(i) ≤ a(i + 1) ≤ a(i) + 1, for all 0 ≤ i ≤ c − 2.
Remark 4.4. If P = {xi−aij y aij |i, j} is a pyramid of type (c, d), then P ′ = yP :=
{xi−aij y aij +1 |i, j} is called shifted pyramid, and w(P ′) = w(P ) + #P .
45
Proof. The column Si = (ai1 , · · · , aim(i) ) goes over to Si′ = (ai1 + 1, · · · , aim(i) + 1),
and w(Si′ ) = (ai1 + 1) + · · · + (aim(i) + 1) − (1 + · · · + m(i) − 1) = w(Si ) + m(i). Thus
c−1
P
w(P ′) = (w(Si ) + m(i)) = w(P ) + #P .
i=0
Remark 4.5. A pyramid with maximal weight does not contain a step of breadth
≥ 4, i.e. it is not possible that a(j) = · · · = a(i) > 0, if i − j ≥ 3.
Proof. If one would have the situation shown in Figure 4.1, then one could make the
deformation xi−a(i) y a(i) 7→ xj−a(j)+1 y a(j)−1 and then from (4.3) one would get: w(P ′) −
w(P ) = 2 · 0 − (j − i) − 2 = i − j − 2 ≥ 1, contradiction.
Remark 4.6. In any pyramid two consecutive “normal” steps can be replaced by a
step of breadth 3, without a change of weight. In the course of this, the sum over the
x-degrees of all monomials in the pyramid increases by 2, however.
Proof. (cf. Fig. 4.2). By assumption one has a(i) + 1 = a(i + 1) and a(i + 1) + 1 =
a(i + 2). If one makes the deformation xi−a(i) y a(i) 7→ xj−a(j)+1 y a(j)−1 , where j = i + 2,
then one gets w(P ′) − w(P ) = 2 · 2 − 2 − 2 = 0.
N.B. The possibility a(i) = 0 is not excluded.
Remark 4.7. There is a pyramid of maximal weight, which does not contain two
consecutive normal steps, that means, there is no index i, such that a(i + 1) = a(i) +
1, a(i + 2) = a(i + 1) + 1. This is called a “prepared” pyramid. (N.B. a(i) may equal zero.)
Proof. Apply Remark 4.6 for several times.
Remark 4.8. A prepared pyramid of maximal weight does not contain two steps of
breadth 3.
Proof. We consider two cases:
1st case: The steps of breadth 3 are situated side by side. Then one makes the deformation
described in Fig. 4.3 and one gets: w(P ′) − w(P ) = 2[a(i − 5) − a(i)] + 5 − 2 = 1,
contradiction.
2nd case: Between the steps of breadth 3 there are ν ≥ 1 double steps. One then makes
the deformation described in Fig. 4.4. Putting j = i − 2(ν + 1) one gets: w(P ′) − w(P ) =
2(a(j) − a(i)) − (j − i) − 2 = 2(−ν) + 2(ν + 1) − 2 = 0. Then P ′ would have maximal
weight, too. But P ′ contains a step of breadth 4, contradicting Remark 4.5.
Remark 4.9. Each positive natural number d can uniquely represented either in the
form d = n(n + 1) − r, 0 ≤ r < n (1st case) or in the form d = n2 − r, 0 ≤ r < n (2nd
case). Both cases exclude each other.
Proof. Choosing n sufficiently large, then the sequence n(n+1), n(n+1)−1, · · · , n(n+
1) − (n − 1), n(n + 1) − n = n2 , n2 − 1, · · · , n2 − (n − 1), n2 − n = (n − 1)n, · · · contains
any given set {1, · · · , m} ⊂ N.
46
We now assume that P is a prepared pyramid of type (c, d), with d ≥ 3 and maximal
weight. If P does not contain a step of breadth 3, then P has the form described either
in Fig. 4.5a or in Fig. 4.5b, according to if either d = n(n + 1) or d = n2 . (One has
n
n
P
P
2 · ν = n(n + 1) and (2ν − 1) = n2 .) If there is 1 step of breadth 3, then from the
1
1
Remarks 3,5,6,7 and 8 it follows that P has the form described either in Fig. 4.6a or in
Fig. 4.6b. Here the step of breadth 3 may lie quite left or right in Fig. 4.6a or Fig 4.6b,
respectively. One sees that Fig. 4.6a and Fig. 4.6b result from Fig. 4.5a and Fig. 4.5b,
respectively, by removing the monomials marked by −.
We first compute the weights of the pyramids shown in Fig. 4.5a and Fig. 4.5b, and
then the weights of the “reduced” pyramids P in Fig. 4.6a and Fig. 4.6b.
1st case: (Fig. 4.5a)
i
a(i)
c−1
n
c−2
n
.......... .....
c − 2ν − 1 n − ν
c − 2ν − 2 n − ν
.......... .....
c − 2n + 1
1
c − 2n
1
w(Si ) = ia(i) − a(i)(a(i) − 1)
(c − 1)n − n(n − 1)
(c − 2)n − n(n − 1)
........................................
(c − 2ν − 1)(n − ν) − (n − ν)(n − ν − 1)
(c − 2ν − 2)(n − ν) − (n − ν)(n − ν − 1)
........................................
(c − 2n + 1) · 1 − 1 · 0
(c − 2n) · 1 − 1 · 0
2nd case: (Fig 4.5b)
i
c−1
c−2
c−3
..........
c − 2ν
c − 2ν − 1
..........
c − 2n + 2
c − 2n + 1
a(i)
n
n−1
n−1
.....
n−ν
n−ν
.....
1
1
w(Si ) = ia(i) − a(i)(a(i) − 1)
(c − 1)n − n(n − 1)
(c − 2)(n − 1) − (n − 1)(n − 2)
(c − 3)(n − 1) − (n − 1)(n − 2)
........................................
(c − 2ν)(n − ν) − (n − ν)(n − ν − 1)
(c − 2ν − 1)(n − ν) − (n − ν)(n − ν − 1)
........................................
(c − 2n + 2) · 1 − 1 · 0
(c − 2n + 1) · 1 − 1 · 0
1st case: We sum up w(Si ) and w(Si−1 ) if i = c − 2ν − 1 and we get:
n−1
X
ν=0
(n − ν)(2c − 2n − 2ν − 1) =
n
X
ν=1
ν(2c − 4n + 2ν − 1)
= (2c − 4n − 1) · 12 n(n + 1) + 13 n(n + 1)(2n + 1)
= n(n + 1)(c − 2n − 0.5 + 32 n + 13 )
= n(n + 1)(c − 34 n − 16 ).
In the reduced pyramid the initial terms a(i) of the column S i , if i = c−2, c−4, · · · , c−
2r, then are equal to n − 1, n − 2, · · · , n − r. This means, a(i) = n − ν, if i = c − 2ν,
and w(S i ) = (c − 2ν)(n − ν) − (n − ν)(n − ν − 1), 1 ≤ ν ≤ r. If i = c − 2ν we get:
47
w(S i )−w(Si ) = (c−2ν)(n−ν)−(n−ν)(n−ν −1)−[(c−2ν)(n−ν +1)−(n−ν +1)(n−ν)] =
−(c − 2ν) + (n − ν) · 2 = 2n − c.
It follows that w(P ) − w(P ) = r(2n − c).
2nd case: We sum up w(Si ) and w(Si−1 ) if i = c − 2ν and we get:
n−1
X
ν=1
(n − ν)(2c − 2n − 2ν + 1) =
n−1
X
ν=1
ν(2c − 4n + 2ν + 1)
1
1
= (2c − 4n + 1) · (n − 1)n + (n − 1)n(2n − 1)
2
3
2
1
= (n − 1)n(c − 2n + 0.5 + n − )
3
3
4
1
= n(n − 1)(c − n + ).
3
6
Beside this, we have to add the weight (c − 1)n − n(n − 1) = n(c − n), if i = c − 1,
and we get w(P ) = n[(c + 0.5)n − 43 n2 − 61 ].
In the reduced pyramid the initial terms a(i) of the column S i , if i = c−1, c−3, · · · , c−
2r + 1, then are equal to n − 1, n − 2, · · · , n − r. This means a(i) = n − ν, if i = c − 2ν + 1,
and w(S i ) = (c − 2ν + 1)(n − ν) − (n − ν)(n − ν − 1), 1 ≤ ν ≤ r. If i = c − 2ν + 1 we get:
w(S i ) − w(Si ) = (c − 2ν + 1)(n − ν) − (n − ν)(n − ν − 1) − [(c − 2ν + 1)(n − ν + 1) − (n −
ν + 1)(n − ν)] = 2n − c − 1, 1 ≤ ν ≤ r. We get w(P ) − w(P ) = r(2n − c − 1).
Remark 4.10. The maximal weight of a pyramid P of type (c, d) is equal to
n[(c − 1, 5)n + (c + 2r − 1/6) − 4/3 · n2 ] − rc
(4.4)
if d = n(n + 1) − r and 0 ≤ r < n and it is equal to
n[(c + 0.5)n + (2r − 1/6) − 4/3 · n2 ] − r(c + 1)
(4.5)
if d = n2 − r and 0 ≤ r < n.
Proof. If d ≥ 3, this follows from the foregoing computation. If d = 2 or if d = 1,
then n = 1 and r = 0. The formula (4.4) and the formula (4.5) give the weights 2c − 3
and c − 1, which is confirmed by Fig. 4.7a and Fig. 4.7b, respectively.
Remark 4.11. The formulas (4.4) and (4.5) agree in the ends of the intervals. This
is shown by putting in r = n in (4.4) and r = 0 in (4.5) and by putting in n − 1 instead
of n and r = 0 in (4.4) and r = n in (4.5), respectively, and then checking equality.
We denote the maximal weight of a pyramid of type (c, d) by w(Pc,d).
Remark 4.12.
(
− 43 n3 − 1, 5n2 + (2r − 16 )n + dc,
w(Pc,d) =
− 34 n3 + 0.5n2 + (2r − 16 )n − r + dc,
if d = n(n + 1) − r,
if d = n2 − r
Thus w(Pc,d) is a strictly increasing function of c ≥ d, if d is fixed.
48
Remark 4.13. Fixing the integer c ≥ 5, then w(Pc,d) is a strictly increasing function
of 1 ≤ d ≤ c.
Proof. w(Pc,d) = − 43 n3 − 1, 5n2 − 16 n + cn(n + 1) + r(2n − c), if d = n(n + 1) − r, 0 ≤
r ≤ n, and w(Pc,d) = − 34 n3 + 0.5n2 − 16 n + cn2 + r(2n − c − 1), if d = n2 − r, 0 ≤ r ≤ n.
√
From n(n + 1) − r = d ≤ c and from n2 − r = d ≤ c it follows that n ≤ c. From c ≥ 5
we conclude 2n − c < 0. From d ≥ 1 it follows that n ≥ 1. As a function of r both terms
for w(Pc,d) are strictly decreasing in the intervals 0 ≤ r ≤ n, and the assertion follows
from Remark 4.11 and Remark 4.12.
Remark 4.14. Fixing the integer 1 ≤ c ≤ 4, w(Pc,d) is an increasing function of
1 ≤ d ≤ c.
Proof. By drawing the possible patterns one finds:
d
1 2
w(P2,d ) 1 1
d
1 2 3
w(P3,d) 2 3 3
d
1 2 3 4
w(P4,d ) 3 5 6 7
We define Pc := Pc,c and w(∅) = 0.
Proposition 4.1. w(Pc ) ≤ (c − 1)2 for all c ∈ N.
Proof. 1st case: c = d = n(n + 1) − r, 0 ≤ r < n. Because of Remark 4.12 one has
to show:
− 34 n3 − 1, 5n2 + (2r − 16 )n + c2 ≤ (c − 1)2 ⇔
4 3
n + 1, 5n2 − (2r − 61 )n − 2[n(n + 1) − r] + 1 ≥ 0 ⇔
3
4 3
n − 0.5n2 − (2r + 11
)n + 2r + 1 ≥ 0 ⇐
3
6
4 3
11
2
n − 0.5n − (2n + 6 )n + 2n ≥ 0 ⇐
3
4 3
n − 2, 5n2 + 61 n ≥ 0. This is true if n ≥ 2. If n = 1, then r = 0, and by substituting
3
one can convince oneself that the inequality is true in this case, too.
2nd case: c = d = n2 − r, 0 ≤ r < n. One has to show:
− 43 n3 + 0.5n2 + (2r − 16 )n − r + c2 ≤ (c − 1)2
(4.6)
⇔ 43 n3 − 0.5n2 − (2r − 61 )n + r − 2[n2 − r] + 1 ≥ 0
⇔ 43 n3 − 2, 5n2 − (2r − 61 )n + 3r + 1 ≥ 0.
Assuming n ≥ 2, this inequality follows from
4 3
n
3
− 2, 5n2 − (2n − 61 )n + 3n + 1 ≥ 0
⇐
⇔
4 3
n −
3
4 2
n
3
4, 5n2 + 3 16 n ≥ 0
− 4, 5n + 3 61 ≥ 0.
This is true if n ≥ 3. Putting n = 1 and n = 2 in (4.6) gives the inequalities r ≥ 0 and
2 − r > 0, respectively, which are true by assumption. As P1,1 = ∅, the assertion is true
if c = 0 or c = 1, too.
49
4.3. Preview
Let be V ⊂ Sn a m-dimensional subspace and V ↔ ξ ∈ G(k) the corresponding point
in G = Grassm (Sn ). We let Gm and Ga operate on S as described in (4.1). We assume
V not to be invariant under Gm or Ga . Then C := {ψα (ξ)|α ∈ k}− is a closed irreducible
curve, which is to have the induced reduced scheme structure and which we imagine as a
curve in PN by means of the Plücker embedding p. Let be h the Hilbert polynomial of
p(C) ⊂ PN , X = Hilbh (G) ֒→ Hilbh (PN ) and σ : Gm → X the morphism λ 7→ σ(λ)C. It
has an extension σ : P1 −→ X , which induces a family of curves
C ֒→ G × P1
ց
f
↓ p2
P1
such that f is flat and Cλ := f −1 (λ) = σ(λ)C for all λ ∈ P1 − {0, ∞}. As f is dominant,
we get (cf. [Fu], p.15): If C0/∞ := p1 (C0/∞ ), then [C0 ] = [C∞ ] in A1 (G).
Let be ξ0/∞ = lim σ(λ)ξ and C0/∞ := {ψα (ξ0/∞ )|α ∈ k}− . The central theme of
λ→0/∞
[T1]–[T4] is the question, what is the connection of [C0 ] and [C∞ ]. The essential tool,
which was already used in [T1] is the α-grade of V , which is nothing else than the degree
of C, imbedded in PN by means of p (cf. [T1], 1.3).
We paraphrase the recipe for estimating the α-grade of V : Let M1 < · · · < Mℓ , ℓ =
ℓ
P
n+2
,
be
the
inverse-lexicographically
ordered
monomials
of
S
,
and
let
be
f
=
aij Mj ,
n
i
2
j=1
1 ≤ i ≤ m, a basis of V . If Mj1 < · · · < Mjm is a sequence of monomials in Sn , then the
Plücker coordinate PV (Mj1 , · · · , Mjm ) is defined to be the determinant of (aijν )1≤i,ν≤m .
In the following , V is a T (ρ)-invariant subspace of Sn , and fi = Mi (1 + a1i X ρ + a2i X 2ρ +
· · · + aiν(i) X ν(i)ρ ), 1 ≤ i ≤ m, is a basis of T (ρ)-semi-invariants. From formula (4.1) in
(4.1) it follows that α − grade(V ) ≤ max{α − grade hM1 X j(1)ρ , · · · , Mm X j(m)ρ i}
(j)
where (j) = (j(1), · · · , j(m)) ∈ [1, ℓ]m ∩ Nm runs through all sequence such that
PV (M1 X j(1)ρ , · · · , Mm X j(m)ρ ) 6= 0.
It is possible to choose the semi-invariants fi so that the initial monomials Mi (the
final monomials Mi X ν(i)ρ =: Ni , respectively) are linearly independent, i.e. different from
each other (cf. Appendix E or the proof of Hilfssatz 6 in [T2], Anhang 1, p. 140). As
the Plücker coordinates of a subvector space, up to a factor different from zero, do not
depend on the basis which one has chosen, one has
(4.7)
PV (M1 , · · · , Mm ) 6= 0 and PV (N1 , · · · , NM ) 6= 0.
Define V0 = hM1 , · · · , Mm i ↔ ξ0 and V∞ := hN1 , · · · , Nm i ↔ ξ∞ . As the function deg
is constant on flat families of curves, we get
α-grade(V ) = deg(C) = deg(C0 ) = deg(C∞ )
As C0/∞ ⊂ C0/∞ we conclude that
(4.8)
α-grade(V ) ≥ max(α-grade(V0 ), α-grade(V∞ )).
50
Now it is scarcely possible to see if PV (M1 X j(1)ρ , · · · , Mm X j(m)ρ ) is different from zero.
Therefore we introduce the number
j(1)
j(m)ρ
max −α-grade(V ) := max{α − grade ha1 M1 X j(1)ρ , . . . , aj(m)
i}
m Mm X
(j)
where (j) runs through all sequences (j(1), · · · , j(m)) ∈ Nm , such that 0 ≤ j(i) ≤ ν(i)
for all 1 ≤ i ≤ m and a0i := 1.
Remark 4.15. (a) Clearly α-grade (V ) ≤ max −α-grade (V ).
(b) In the definition, the monomials need not be ordered.
j(i)
(c) If one coefficient ai is equal to zero or if two of the monomials Mi X j(i)ρ are equal
for different indices i, then the m-times exterior product of the monomial space and its
α-grade are zero (cf. 4.1).
To say it differently, take from each semi-invariant fi a monomial Mi X j(i)ρ , whose
m
j(i)
coefficient ai is different from zero, form ∧ ψα (Mi X j(i)ρ ), and determine the highest
i=1
power of α occurring in such an exterior product. Finally, define max −α-grade (V ) to
be the maximum of these degrees, if (j) runs through all such sequences.
Accordingly, one defines
j(1)
j(m)ρ
min −α − grade(V ) := min{α-gradeha1 M1 X j(1)ρ , · · · , aj(m)
i}
m Mm X
(j)
where (j) runs through all sequences of the kind described above and which give an
α-grade different from zero.
As α-grade (V0/∞ ) = deg(C0/∞ ), from (4.7) we conclude that
(4.9)
min −α-grade(V ) ≤ min(deg C0 , deg C∞ ).
Later on, the vector space V will always be equal to H 0 (I(n))), where I ⊂ OP2 is a
G = Γ · T (ρ)-invariant ideal of y-standard form (cf. 2.4.3 Definition 2). We will see that
max −α-grade (I) := max −α-grade (H 0 (I(n))) and min −α-grade (I) := min −α-grade
(H 0 (I(n))) not only are independent of n ≥ colength (I), but also can be computed with
the help of smaller numbers n. The real aim of the following estimates is to prove the
following inequality: If I ⊂ OP2 is an ideal of y-standard form and if reg(I) = m, then
(!)
Q(m − 1) + min −α-grade(I) > max −α-grade(I).
From this it will follow that C0 and C∞ do not contain any y-standard cycle besides
C0 and C∞ , respectively (cf. Lemma 9.2).
51
Fig. 4.1
Fig. 4.2
Fig. 4.3
a(i+2)
a(j)
Fig. 4.4
a(i)
a(i)
a(i)
a(i−5)
a(i)
a(j)
Fig. 4.5a
n
n−1 n−1
n−ν n−ν
52
−
c−1
c−2
n
c−2ν−1
c−2ν−2
c−2n+1
c−2n
−
−
−
−
−
n
e−2n
c−2n+1
e−2n+1
c−2n+2
−
n−ν n−ν
Fig. 4.5b
−
Fig. 4.6a
53
c−2ν−1
c−2ν
−
−
c−2
c−1
−
n
e−1
−
n−1 n−1
e−2
n
n
Fig. 4.6b
Fig. 4.7a
e−1
e−2
e−2n+2
e−2n+1
n
Fig. 4.7b
c−1
c−1
54
CHAPTER 5
Estimates of the α-grade in the case ρ1 < 0, ρ2 > 0.
5.1. Preliminary remarks.
We refer to Proposition 3.1 in section 3.5 and we treat case (I.1) at first. If in fi the
vice-monomial Njup occurs, then Ik is monomial for all k ≥ j + 1. Especially, fj+1 , · · · fr
are monomials, which do not cause a deformation of the pyramid.
We show that the z-degree of such a monomial Njup cannot be equal to the z-degree
of an initial monomial Mk . For then it would follow mj + j − 1 = mk + k for another
index k, which is not possible by Corollary 2.4. For the same reason it is not possible
that the z-degree of Njup · (z/y) is equal to the z-degree of Nkdown or of Mk . The corresponding statements are true, if “up” and “down” are exchanged, as it follows from the
corresponding definition in (3.2) and (3.3).
Finally, if there occurs a deformation of the form (1.6) of Proposition 3.1, then it can
be that the final monomial Nkup (z/y) of fi has the same z-degree as the initial monomial
Mℓ of fℓ . But then ℓ > k > j > i, and therefore Iℓ is a monomial ideal by Lemma 3.1
. But then fℓ does not define a deformation, at all. It follows from this remarks that
one can separately consider the deformations defined by the different fi , if one wants to
determine the changes of α-grade caused by these deformations. (N.B. This statement is
analogously valid in the situation described by Proposition 3.2, too).
At first we determine the change of the α-grade, if in one fi the initial monomial Miup
is replaces by another monomial occurring in fi :
1◦ Miup 7−→ Njdown , if 0 ≤ i < j ≤ r;
2◦ Miup 7−→ Njup , if 0 ≤ i < j ≤ r;
3◦ Miup 7−→ L, L ∈ L monomial such that (x, y)L ⊂ ℓK(−r − 1), 0 ≤ i ≤ r;
4◦ Miup 7−→ Nkup · (z/y) = Mkup · (z/y)2 , 0 ≤ i < k ≤ r;
5◦ Miup 7−→ L · (z/y), L ∈ L monomial such that (x, y)L ⊂ ℓK(−r − 1), 0 ≤ i ≤ r.
The deformation 4◦ (resp. 5◦ ) comes from the case 1.6 (resp. 1.7) of Proposition 3.1,
and therefore there is at most one such a deformation, whereas in the deformations 1◦ and
2◦ (resp. 3◦ ) the index i may a priori run through all integers 0 ≤ i < r (resp. 0 ≤ i ≤ r).
Then for the index j in the cases 1◦ and 2◦ (resp. for the monomial L in the case 3◦ )
there are several possibilities. But if one has chosen i, then one has to decide for an index
j (resp. for a monomial L), and we will give an uniform estimate of the corresponding
changes of α-grades.
We identify the set L, which was introduced in Section (3.3) with the vector space
generated by the monomials in this set.
55
We denote by LB (left domain) the vector space generated by all monomials in Sm0
with z-degree ≥ m0 − (c + r), i.e., generated by all monomials xa y bz m0 −(a+b) , where
a + b ≤ c + r.
As to the deformation 4◦ (resp. 5◦ ), there is still the possibility yMiup 7→ yMkup (z/y)2
(resp. yMiup 7→ yL · (z/y)). This is because in the case 1.6 (resp. 1.7) of Proposition 3.1,
fi has the order 1, whereas in the remaining cases fi has the order 0.
Remember that (cf. Figure 3.1 and 5.1):
Miup = xi−ι(i) y mi +ι(i) z m0 −mi −i
Niup = Miup · (z/y) = xi−ι(i) y mi +ι(i)−1 z m0 −mi −i+1
Midown = xmi +i−ι(i) y ι(i) z m0 −mi −i
Nidown = Midown · (z/x) = xmi +i−ι(i)−1 y ι(i) z m0 −mi −i+1
Ekup := Mkup · (z/y)2 = xk−ι(k) y mk +ι(k)−2 z m0 −mk −k+2
Ekdown := Mkdown · (z/x)2 = xmk +k−ι(k)−2 y ι(k) z m0 −mk −k+2 .
5.2. Estimates in the case I.
We determine one after the other the changes of the α-grade in the deformations:
1 Miup → Njdown .
First we note ϕ′ (mi + i) = mi + 1 and ϕ′ (mj + j − 1) = mj − 1 (see Fig.5.2 ). The α-grade
of the column, in which Miup occurs changes by −(mi + ι(i)) + ϕ′ (mi + i) − 1 = −ι(i)
(cf. the formula (4.2) in 4.1). The α-grade of the column, to which Njdown is added,
changes by ι(j) − ϕ′ (mj + j − 1) = ι(j) − mj + 1 ( loc . cit.). Therefore the α-grade
changes by −mi − ι(i) + ϕ′ (mi + i) − 1 + ι(j) − ϕ′ (mj + j − 1) = ι(j) − ι(i) − mj + 1.
As 0 ≤ ι(i) ≤ ι(j) ≤ j , the absolute value of this difference is ≤ max(j, mj − 1), where
0 ≤ i < j ≤ r.
2◦ Miup → Njup .
The α-grade of the column, to which Miup belongs changes by −(mi +ι(i))+ϕ′ (mi +i)−1 =
−ι(i); the α-grade of the column, to which Njup is added, changes by mj +ι(j)−1−ϕ′ (mj +
j − 1) = ι(j). Therefore the change of α-grade is equal to 0 ≤ ι(j) − ι(i) ≤ j, where
0 ≤ i < j ≤ r.
3◦ Miup 7−→ L ∈ L.
The α-grade of the column, to which Miup belongs, changes by −ι(i). From the domain L
the monomial L is removed, such that by Proposition 4.1 one gets the following estimate
of the α-grade: The α-grade after the deformation 3◦ of that part of the pyramid, which
belongs to the left domain is ≤ (c − 1)2 + ι(r + 1)[ c+1
− (c − 1)]. For one has #L = c, and
2
c+1
0
there are 2 − c initial monomials of ℓH (K(c − 1)) in the left domain LB. Therefore
the expression in the bracket gives the number of monomials in the corresponding part
of the pyramid after the deformation. Besides this one has to take into account that the
pyramid is pushed upwards by ι(r + 1) units (cf. Remark 4.4). We recall that LB (resp.
RB) is the vector space generated by the monomials of total degree ≤ c + r (resp. of
total degree ≥ c + r + 1) in x and y (cf. Fig. 5.3). The change of α-grade caused by
the deformation 3◦ can be expressed as follows: The change in the domain RB is −ι(i);
◦
56
the α-grade of the left domain of the pyramid after the deformation is estimated as given
above.
4◦ Miup 7−→ Ekup .
At first we consider the case that Ekup belongs to the right domain, i.e. mk +k−2 ≥ c+r+1,
and we orientate ourselves by Figure 5.1. The deformation 4◦ changes the α-grade of the
column of Miup by −ι(i) (cf. 3◦ ), and the α-grade of the column to which Ekup is added
changes by mk +ι(k)−2−ϕ′ (mk +k −2). As we have remarked above, ϕ′ (mk +k) = mk +1
and therefore ϕ′ (mk +k−2) = mk −2. Therefore the α-grade of the column of Ekup changes
by ι(k). Altogether the deformation 4◦ causes a change of α-grade by 0 ≤ ι(k) − ι(i) ≤ k,
where 0 ≤ i < k ≤ r.
This deformation occurs only once, yet one has to take into account the deformation
4◦ bis (y/z)Miup 7→ Nkup ( Proposition 3.1c). In the column of yMiup this gives a change of
the α-grade by −(mi + ι(i) + 1) + ϕ′ (mi + i + 1) − 1 = −mi − ι(i) − 1 + mi + 2 − 1 = −ι(i).
In the column of Nkup the α-grade changes by mk + ι(k) − 1 − ϕ′ (mk + k − 1) = mk +
ι(k) − 1 − (mk − 1) = ι(k). Altogether the deformation 4◦ bis gives a change of α-grade
by 0 ≤ ι(k) − ι(i) ≤ k.
Now to the case mk + k − 2 ≤ c + r. Due to the deformation 4◦ (resp. 4◦ bis) the
α−grade in the right domain of the pyramid changes by −ι(i). In any case the deformation
4◦ (resp. 4◦ bis) gives a change of α-grade in the right domain of absolute value ≤ r.
5◦ Miup 7−→ L · (z/y).
Removing Miup (resp. (y/z)Miup ) gives a change of α-grade by −ι(i) in the corresponding
column (cf. case 4◦ ). The changes in the left domain will be estimated later on.
The deformations 1◦ − 5◦ exclude each other, i.e., there are at most r + 1 such deformations plus two deformations 4◦ bis and 5◦ bis. The changes in the right domain can be
estimated in the cases 1◦ and 2◦ by max(j, mj − 1) ≤ r + mi+1 , where i runs through the
numbers 0, · · · , r − 1. The absolute value of the change in the case 3◦ can be estimated
by r, and the same is true for the deformations 4◦ , 4◦ bis, 5◦ and 5◦ bis.
We now consider possible trinomials.
6 We assume there is a trinomial of the form 1.3. We want to determine the change of
α-grade,if Njup is replaced by Nkup , where we start from a pyramid containing Njup instead
of Miup . The changes of α-grade in the following diagram follow from the computation in
2◦ .
Miup
ι(j) − ι(i) ւ
ց ι(k) − ι(i)
δ
up
Nj
−→ Nkup
◦
The change of α-degree is therefore 0 ≤ δ := ι(k) − ι(j) ≤ r.
7◦ The trinomial has the form 1.4.
Miup
ι(j) − ι(i) ւ
ց ι(k) − ι(i) − mk + 1
δ
up
Nj
−→ Nkdown
(cf. 1◦ and 2◦ ) Therefore δ = ι(k) − ι(j) − mk + 1, and as in 1◦ we obtain the estimate
0 ≤ |δ| ≤ max(k, mk − 1) ≤ mk + k < mi+1 + r.
57
8◦ The trinomial has the form (1.5).
Miup
ι(j) − ι(i) ւ
ց −ι(i)
δ
up
Nj
−→ L
(cf. 3◦ ) Therefore δ = −ι(j) and 0 ≤ |δ| ≤ r.
9◦ The trinomial has the form (1.6).
Miup
ι(j) − ι(i) ւ
ց ι(k) − ι(i)(resp. − ι(i))
δ
up
Nj
−→ Nkup · (z/y)
(cf. 4◦ ) It follows δ = ι(k) − ι(j) (resp. δ = −ι(j)) and therefore |δ| ≤ r.
10◦ The trinomial has the form (1.7).
Miup
ι(j) − ι(i) ւ
ց −ι(i)
δ
up
Nj
−→ L · (z/y)
(cf. 5◦ ) It follows that δ = −ι(j) and |δ| ≤ r.
N.B. Because of Njup · (y/z) = Mjup the cases 9◦ bis and 10◦ bis do not occur.
Summarizing the cases 1◦ − 10◦ one sees that the total change of α-grade in the right
r
P
domain has an absolute value ≤ (r + 1)r + 2r +
mi . In order to formulate this result
i=1
in a suitable manner, we have to introduce some notations.
We take up the decomposition (Z) of Section (3.1) and choose a standard basis of
H (K(c)). Then we multiply the elements in this basis as well as the forms fi by monomials
in the variables x, y, z to obtain a basis of T (ρ)- semi-invariants of H 0 (I(n)) with different
initial monomials. By linearily combining one can achive that the initial monomial of
each semi-invariant does not appear in any other of these semi-invariant, i.e., one gets a
standard basis. (Fig. 3.1 is to show these initial monomials.) The set of all monomials
which occur in this basis form a pyramid, which is denoted by P. Here n ≫ 0, e.g. n ≥ d.
0
From each element of the basis we take a monomial the coefficient of which is different
from zero and such that the monomials are different from each other .Then we compute
the α-grade of the vector space generated by these monomials. The maximum and the
minimum of the α-grades which one obtains in this way had been denoted by
max −α − grade(V ) and by
min −α − grade(V ),
respectively. One chooses from such a sequence of monomials, which gives the maximal αgrade (which gives the minimal α- grade , respectively) those monomials the total degree
in x and y of which is ≥ c + (r + 1). Then one forms the α-grade of the subspaces
generated by these monomials and denotes it by max −α-grade (P ∩ RB) ( by min −αgrade (P ∩ RB), respectively ). If one chooses from such sequences of monomials those
58
monomials the total degree in x and y of which is ≤ c + r, then max −α-grade (P ∩ LB)
and min −α-grade (P ∩ LB) are defined analogously.
Of course this is valid in the case ρ1 > 0, too, but the assumption ρ2 > 0 is essential .
We make the
Definition 6. A := max −α-grade (P ∩ RB) − min −α-grade (P ∩ RB).
Then we can formulate the result obtained above as
Conclusion 5.1. In the case I.1 one has
r
X
mi .
A ≤ r(r + 3) +
i=1
N.B.. If r = 0 one has actually A = 0.
Now to the case I.2 (cf. Proposition 3.1, 2nd case).
1 Miup 7−→ Njdown gives a change of α-grade of absolute value ≤ max(r, mi+1 ), where
0 ≤ i ≤ r − 1 (cf. the case I.1).
2◦ Miup 7−→ L ∈ L gives a change of α-grade in the right domain by −ι(i) (see above).
Further possible deformations are yMiup 7→∈ L, y 2Miup 7→∈ L, · · · , y ν Miup 7→∈ L, so long
as mi + i + ν < mi−1 + (i − 1) − 1 (cf. Conclusion 3.2 ). This gives in the column of yMiup
(of y 2 Miup , · · · , y ν Miup , respectively) a change of α-grade by
◦
−(mi + ι(i) + 1) + [ϕ′ (mi + i + 1) − 1] = −(mi + ι(i) + 1) + [mi + 1] = −ι(i)
(by − (mi + ι(i) + 2) + [ϕ′ (mi + i + 2) − 1] = −(mi + ι(i) + 2) + [mi + 2] = −ι(i),
··· ,
−(mi + ι(i) + ν) + [ϕ′ (mi + i + ν) − 1] = −(mi + ι(i) + ν) − [mi + ν] = −ι(i), respectively)
as long as mi + i + ν < mi−1 + (i − 2), see above.) This procedure can be repeated at
most c times, until L is full. As ι(r) ≤ r, the total change of α-degree in the right domain
caused by deformations 2◦ has an absolute value ≤ cr. If A is defined as before one gets
Conclusion 5.2. In the case I.2 one has
r
X
A ≤ r(r + c) +
mi .
i=1
N.B. If r = 0, then one really has A = 0. For removing M0up = y m0 does not change
the α-grade of the column of z-degree 0 (this α-grade is equal to 0), or one has f0 = xm0 .
5.3. The case r ≥ 1.
We start with an ideal I of type r ≥ 0 such that ℓ0 = y, and we refer to Proposition
3.1 again. The aim is to prove the inequality (!) in (4.3), where now V = H 0 (I(n)). In
the course of the following computations it will turn out that α-grade (V ) is independent
of n, if n is sufficient large, e.g. if n ≥ d.
59
If one simply writes α-grade (I) instead of α-grade (H 0 (I(n))), where n is sufficient
large, then one has to show:
Q(m0 − 1) + min −α − grade (I) > max −α − grade (I)
(!)
We orientate ourselves by Fig. 5.4. From the Remarks 4.4 , 4.13 , 4.14 and Proposition
4.1 it follows that
c+1
− c]
min −α-grade(LB ∩ P) ≥ ι(r + 1)[
2
c+1
2
− (c − s − 1)]
max −α-grade(LB ∩ P) ≤ (c − 1) + ι(r + 1)[
2
where s + 1 = total number of all deformations Miup 7→∈ L, · · · , y ν Miup 7→∈ L, even
with different indices i (s = −1 means, there are no such deformations). For proving (!)
it is sufficient to show
Q(m0 − 1) > (c − 1)2 + (s + 1) · ι(r + 1) + A
(*)
where A = max −α-grad (P ∩ RB) − min −α-grade (P ∩ RB) by definition (cf. Section
5.2 for the notations).
N.B. If c = 0 there are no deformations into the left domain of the pyramid. Therefore
the inequality (*) reduces to
Q(m0 − 1) > A.
(*bis)
These statements are independent of ρ1 < 0 (Case I) or ρ1 > 0 (Case II).
Unfortunately one has to distinguish these two cases in the following estimates and we
start with Case I.1 (cf. Proposition 3.1).
Because of 0 ≤ s + 1 ≤ c, 1 ≤ ι(r + 1) ≤ r + 1 and the estimates of
as well as Lemma 2.8) it is sufficient to show:
r
r
X
X
m0 +1
2
− (c +
mi ) > (c − 1) + c(r + 1) + r(r + 3) +
mi
2
i=0
i=1
1
m0 (m0 − 1) > c2 + cr + r(r + 3) + 1
2
The case r = 0 has to be treated separately (see below Section 5.4), such that we can
assume r ≥ 1. If c = 0, then m0 ≥ 2r+1 ( Lemma 2.8), and (5.1) follows from 2r (2r+1 −1) >
r(r + 3) + 1, which inequality is true if r ≥ 1. Therefore we can assume c > 0. Because of
m0 ≥ 2r (c+2) the inequality follows from 2r−1 (c+2)·2r (c+1) > c2 +cr +r(r +3)+1 ⇐⇒
22r−1 (c2 + 3c + 2) > c2 + cr + r(r + 3) + 1, which is true if r ≥ 1 and c > 0.
(5.1)
⇐⇒
Conclusion 5.3. In the case I.1 the inequality (*) is valid, if r ≥ 1.
Now to the case I.2. Because of Conclusion 5.2 one has to replace in the inequality
above r(r + 3) by r(r + c), i.e. one has to show that 22r−1 (c2 + 3c + 2) > c2 + 2rc + r 2 + 1,
if r ≥ 1. This is true, if c ≥ 0.
Conclusion 5.4. In the case I.2 the inequality (*) is valid, if r ≥ 1.
60
5.4. The case r = 0.
As always we suppose g ∗(ϕ) > g(d). As ℓ0 = y and ρ1 < 0 by assumption, one has
f0 = xm0 (cf. Fig. 5.5). Therefore there are no deformations into L, i.e. A = 0 and (*)
reads Q(m0 − 1) > (c − 1)2 . If for simplicity one writes m instead of m0 , then one has to
show
(5.2)
m(m − 1) > 2c2 − 2c + 2
Now c = colength (K), K an ideal of type (−1) and the following cases can occur:
1. If ψ is the Hilbert function of K, then g ∗ (ψ) ≤ g(c). Especially one has c ≥ 5 and by
Corollary 2.2 m ≥ 2c + 1, from which (5.2) follows.
2. One has c ≤ 4. If c = 0, then I is monomial and there are no deformations at all. It
follows that (∗bis) is fulfilled.
A little consideration shows that because of m ≥ c + 2 the inequality (5.2) is fulfilled
in the cases 1 ≤ c ≤ 4, too. Thus (*) and (∗bis) are proved in the case r = 0, also. Using
Conclusion 5.4 gives
Proposition 5.1. Assume that I has the type r ≥ 0 and has y-standard form; assume
ρ1 < 0 and ρ2 > 0. Then (*) and (∗bis) are fulfilled, respectively.
N.B. Hence the inequality (!) follows (see the corresponding argumentation in 5.3).
61
Fig. 5.1
Mkup
Nkup
mk +k
mk +k−2
Ekup
Fig. 5.2
ϕ′ (mi + i) = mi + 1
ϕ′
0
1
2
3
4
5
6
m2 +2
8
m1 +1
10 m0
62
Fig. 5.3
c+r+1
c+r
column with c + 1 monomials
is complete
mr +r
right domain
#{ monomials in the left domain } = h0 (K(c − 1)) = c+1
−c
2
n+2
as K has the Hilbert polynomial 2 − c.
m0
left domain
#{ monomials in the right domain } = Q(m0 − 1) − #{ monomials in the left domain }
= m02+1 − d c+1
−c
2
63
Fig. 5.4a
Fig. 5.4b
· · · · · · · · · · · · · · · · · · m1 +1 · · · · · · m0
· · · · · · · · · · · · · · · · · · m1 +1 · · · · · · m0
Fig. 5.5 (r = 0)
column filled
with monomials
#L=c
c c+1
LB
RB
64
CHAPTER 6
Estimates of the α-grade in the case ρ1 > 0, ρ2 > 0 and r ≥ 1.
We refer to Proposition 3.2 in (3.6) and take over the notations from there. As has
been remarked in (5.1) one can compute the changes of α-grade by the single deformations
fi separately.
6.1. Estimates in the case II
At first we compute the changes of the α-grade , if we replace the initial monomial
Mi of fi by another monomial occurring in fi (cf. 5.1):
1◦ Midown 7−→ Njup , if 0 ≤ i < j ≤ r.
In the column of Midown there is a change of the α-grade by −ι(i)+ϕ′ (mi +i)−1 = mi −ι(i).
In the column of Njup there is a change of the α-grade by mj + ι(j) − 1 − ϕ′ (mj + j − 1) =
mj +ι(j)−1−(mj −1) = ι(j). Therefore the deformation 1◦ gives a change of the α-grade
by mi + ι(j) − ι(i), 0 ≤ i < j ≤ r.
2◦ Midown 7−→ Njdown , 0 ≤ i < j ≤ r.
In the column of Midown there is a change of the α-grade by mi − ι(i); in the column of
Njdown is a change of α-grade by ι(j) − ϕ′ (mj + j − 1) = ι(j) − mj + 1. Therefore the
deformation 2◦ gives a change of α-grade, whose absolute value is |ι(j)−ι(i)+mi −mj +1| ≤
max(|mi + ι(j)|, |mj + ι(i) − 1|) ≤ mi + j, where 0 ≤ i < j ≤ r.
3◦ Midown 7−→ L ∈ L.
In the column of Midown is a change of α-grade by 0 < mi − ι(i) ≤ mi + i, 0 ≤ i ≤ r. The
change of α-grade in the left domain can be estimated as in Case I.
4◦ Midown 7−→ Ekdown = Mkdown · (z/x)2 , 0 ≤ i < k ≤ r.
At first we consider the case that Ekdown belongs to the right domain, i.e. mk + k − 2 ≥
c+r +1. The change of α-grade in the column of Midown is mi −ι(i); the change of α-grade
in the column of Ekdown is ι(k) − ϕ′ (mk + k − 2) = ι(k) − (mk − 2). Therefore the absolute
value of the change of α-grade by the deformation 4◦ is |mi − ι(i) + ι(k) − mk + 2| ≤
max(|mi + ι(k)|, |mk + ι(i) − 2|) = mi + ι(k) ≤ mi + k, if 0 ≤ i < k ≤ r. This deformation
can occur only once, yet one has to take into account the deformation
4◦ bis (x/z)Midown 7→ Nkdown (cf. Proposition 3.2c).
In the column of xMidown this gives a change of the α-grade by −ι(i) + ϕ′ (mi + i + 1) − 1 =
mi − ι(i) + 1. In the column of Nkdown the α-grade changes by ι(k) − ϕ′ (mk + k − 1) =
ι(k) −mk + 1. Thus the absolute value of the change of α-grade in the right domain due to
4◦ bis is |mi −ι(i)+1+ι(k)−mk +1) ≤ max(|mi +ι(k)|, |mk +ι(i)−2|) = mi +ι(k) ≤ mi +k,
where 0 ≤ i < k < r. This deformation occurs only once.
Now to the case that mk + k − 2 ≤ c + r. Removing Midown ( resp. (x/z)Midown ) gives a
65
change of α-grade by mi − ι(i) ( by mi − ι(i) + 1, respectively), whose absolute value is
bounded by mi + r.
5◦ Midown 7−→ L · (z/x).
Removing Midown (resp. (x/z)Midown ) causes a change of α-grade of the column of Midown
(resp. (x/z)Midown ) by mi − ι(i) (resp. by mi − ι(i) + 1), which are estimated by mi + i
(resp. mi + i + 1), where 0 ≤ i ≤ r. The deformation 5◦ (resp. 5◦ bis) can occur only
once. The changes in the left domain will be estimated later on.
The deformation 1◦ −5◦ exclude each other, i.e. there are at most r+1 such deformation
plus two deformations 4◦ bis and 5◦ bis. The changes of α-grade in the right domain in the
cases 1◦ − 3◦ have an absolute value ≤ mi + r, 0 ≤ i ≤ r. The same estimate is valid for
the deformations 4◦ and 4◦ bis, even if Ek belongs to the left domain, as we have assumed
r ≥ 1. As for the deformations 5◦ (resp. 5◦ bis) we estimate the change of the α-grade by
mi + r (resp. mi + r + 1).
We now consider possible trinomials.
6 We assume there is a trinomial of the form 1.3. Similarly as in the Case I in Chapter
5, we have a diagram
Midown
γ(j) ւ
ց γ(k)
δ
down
Nj
−→ Nkdown
where γ(j) := mi − mj − ι(i) + ι(j) + 1 and γ(k) := mi − mk − ι(i) + ι(k) + 1 (cf. 2◦ ).
Therefore δ = mj − mk − ι(j). It follows that |δ| ≤ max(|mj + ι(k)|, |mk + ι(j)|) ≤ mi + r.
◦
7◦ The trinomial has the form 1.4.
Midown
γ(j) ւ
ցβ
δ
down
−→ Nkup
Nj
where β := mi − ι(i) + ι(k) (cf. 1◦ and 2◦ ). It follows that δ = mj − ι(j) + ι(k) − 1 and
|δ| ≤ mi + r.
8◦ The trinomial has the form 1.5.
Midown
γ(j) ւ
ցβ
δ
down
Nj
−→ L
where β := mi − ι(i) (cf. 3◦ ). It follows that δ = mj − ι(j) − 1 and |δ| ≤ mi + r.
9◦ The trinomial has the form 1.6.
Midown
γ(j) ւ
ցβ
δ
down
Nj
−→ Nk · (z/x)
where β := mi − mk − ι(i) + ι(k) + 2, respectively β := mi − ι(i) (cf. 4◦ ). It follows that
δ = mj − mk − ι(j) + ι(k) + 1 (resp. δ = mj − ι(j) − 1) and |δ| ≤ max(|mj − ι(k)|, |mk +
66
ι(j) − 1|) ≤ mj + r.
10◦ The trinomial has the form 1.7.
Midown
γ(j) ւ
ց mi − ι(i)
δ
down
Nj
−→ L · (z/x)
(cf. 5◦ ). It follows that δ = mj − ι(j) − 1 and |δ| < mi + r.
Notabene. Because of Njdown · (x/z) = Mjdown the case 9◦ bis or 10◦ bis does not occur.
Summarizing the cases 1◦ − 10◦ one sees that the total change of α-grade in the
r
P
mi . If one estimates the changes
right domain has an absolute value ≤ (r + 1)r +
i=0
in the cases 4◦ bis and 5◦ bis by mi + r and mi + r + 1, respectively, one obtains A ≤
r
P
r(r+3)+1+3m0 + mi . As we have assumed that r ≥ 1, we have m0 ≥ c+2+mr +· · ·+m1
i=1
( Lemma 2.8) and we obtain
Conclusion 6.1. If r ≥ 1 is assumed,in the case II.1 one has A ≤ 4m0 + r(r + 3) −
c − 1.
Now we come to the case II.2 (cf. Proposition 3.2, 2nd case).
1◦ Midown 7−→ Njup again gives a change of the α-grade in the right domain, whose
absolute value is ≤ mi + r, 0 ≤ i ≤ r − 1 (see above).
2◦ Midown 7−→ L ∈ L gives a change of α-grade in the right domain by mi − ι(i), 0 ≤
i ≤ r (see above). Further possible deformations are xMidown 7→∈ L, x2 Midown 7→∈
L, · · · , xν Midown 7→∈ L, as long as mi + i + ν < mi−1 + (i − 2) (cf. Conclusion 3.2).
This gives in the column of xMidown (of x2 Midown , · · · , xν Midown , respectively) a change
of α-grade by −ι(i) + ϕ′ (mi + i + 1) − 1 = −ι(i) + (mi + 2) − 1 = mi − ι(i) + 1 (by
−ι(i) + ϕ′ (mi + i + 2) − 1 = mi − ι(i) + 2, · · · , −ι(i) + ϕ′ (mi + i + ν) − 1 = mi − ι(i) + ν,
respectively), as long as mi + i + ν < mi−1 + (i − 2) and ν ≤ c − 1.
Remark 6.1. One has |mi − ι(i)| ≤ m0 for all 0 ≤ i ≤ r.
Proof. The inequality −ι(i) + mi ≤ mi ≤ m0 is true, and ι(i) − mi ≤ i − mi ≤
r − mi ≤ r ≤ m0 is true if r = 0. If r ≥ 1 one has m0 ≥ c + 2 + mr + · · · + m1 > r (Lemma
2.8).
From this we conclude: Replacing Midown , xMidown , · · · , xν(i) Midown by monomials in L,
even with different indices i as long as ν(i) ≤ c − 1, gives a change of α-grade in the right
ν(i)
P
(m0 + j), because |mi − ι(i) + j| ≤ m0 + j by the
domain whose absolute value is ≤
j=0
remark above . One gets A ≤
r−1
P
(mi + r) +
i=0
r ν(i)
P
P
(m0 + j). As ν(0) + 1 + · · ·+ ν(r) + 1 ≤ c,
i=0 j=0
67
it follows that A ≤
we obtain
r−1
P
i=1
r−1
P
(mi +r)+
i=0
c−1
P
(m0 +j). As m0 ≥ c+2+mr +· · ·+m1 and mr ≥ c+2,
j=0
mi ≤ m0 − 2(c + 2). This estimate is valid if r ≥ 1. In the case r = 0 one
only has the deformations M0down 7→∈ L, · · · , xs M0down 7→∈ L, and s can be estimated as
in ( Proposition 3.2).
If M0down occurs, the α-grade in the column of M0down , · · · , xs M0down increases by
m0 , · · · , m0 + s, respectively.
Conclusion 6.2. In the case II.2 one has A ≤ (c + 1)m0 + r 2 − 2(c + 2) + 2c , if r ≥ 1.
If r = 0, if I has y-standard form and κ := reg(K), then A ≤ (s + 1)m0 + s+1
, where
2
s ≤ κ/ρ2 − m0 (1/ρ2 − 1/(ρ1 + ρ2 )).
6.2. The case r ≥ 2.
We recall that we started from an ideal I of type r ≥ 0 with y-standard form, and
the aim was to show the inequalities
Q(m0 − 1) > (c − 1)2 + (s + 1)ι(r + 1) + A
(*)
and
respectively (cf. Section 5.3).
Q(m0 − 1) > A (*bis)
At first we treat the case II.1, where A ≤ 4m0 + r(r + 3) − c − 1, if r ≥ 1.
Auxiliary Lemma 1. If I has the type r = 2 (the type r = 1, respectively), then
m0 ≥ 14 ( m0 ≥ 7, respectively).
Proof. We use the results of Lemma 2.8 and write in the case r = 2 : I = I0 =
yI1 (−1) + f0 OP2 (−m0 ), I1 = ℓ1 I2 (−1) + f1 OP2 (−m1 ), I2 = ℓ2 K(−1) + f2 OP2 (−m2 ). I2
has the type 0, therefore colength (I2 ) = c + m2 ≥ 5 and it follows that m2 ≥ 5 − c. As
I1 has the type 1, we get m1 ≥ m2 + c + 2 ≥ 7. Because of m0 ≥ c + 2 + m2 + m1 it
follows that m0 ≥ c + 2 + 5 − c + 7.
To begin with, let c = 0. Then (∗bis) reads 21 m0 (m0 + 1) − (mr + · · · + m0 ) >
r
P
4m0 + r(r + 3) − 1. Because of m0 − (c + 2) ≥
mi it is sufficient to prove
1
(6.1)
m0 (m0 − 11) > 2r(r + 3) − 6
If r = 2 (respectively r = 3) the right side of (6.1) is equal to 14 (equal to 30, respectively).
As m0 ≥ 14 by the Auxiliary Lemma 1 (m0 ≥ 23 (c + 2) = 16 by Lemma 2.8, respectively),
(6.1) is true in these cases.
Let be r ≥ 4. Then m0 ≥ 2r+1 and (6.1) follows from 2r (2r+1 − 11) > r(r + 3). Now
2r > r if r ≥ 2, and 2r+1 − 11 > r + 3 is equivalent to 2r+1 > r + 14, which is true if r ≥ 4.
68
Thus we can assume c ≥ 1 in the following. Because of 1 ≤ ι(r + 1) ≤ r + 1, 0 ≤
r
P
s + 1 ≤ c, mi ≤ m0 − (c + 2) the inequality (∗) will follow from:
i=1
(6.2)
1
m0 (m0 + 1) − (2m0 − 2) > (c − 1)2 + c(r + 1) + 4m0 + r(r + 3) − c − 1
2
⇐⇒ m0 (m0 − 11) > 2c2 + 2(r − 2)c + 2r(r + 3) − 4
Because of m0 ≥ 2r (c + 2) it suffices to show:
(6.3)
2r (c + 2)[2r c + 2r+1 − 11] > 2c2 + 2(r − 2)c + 2r(r + 3) − 4
If r = 2 this inequality reads 4(c + 2)(4c − 3) > 2c2 + 16 ⇔ 7c2 + 10c − 20 > 0 and
this is true if c ≥ 2.
In the case r = 2, c = 1, the inequality (6.2) reads m0 (m0 − 11) > 18, which is
true because m0 ≥ 14 (cf. Auxiliary Lemma 1). Therefore we can now suppose without
restriction r ≥ 3, c ≥ 1. But 2r+1 > 11 and thus (6.3) follows from 2r (c + 2) · 2r c >
2c2 + 2cr + 2r(r + 3) which is equivalent to :
(22r − 2)c2 + (22r+1 − 2r)c > 2r(r + 3)
(6.4)
The left side of (6.4) is a monotone function of c, and if c = 1, then (6.4) reads
2 − 2 + 22r+1 − 2r > 2r(r + 3) ⇔ 22r + 22r+1 > 2r 2 + 8r + 2 ⇔ 22r−1 + 2r > r 2 + 4r + 1.
This is true if r ≥ 3. Summarizing all subcases we obtain
2r
Conclusion 6.3. In the case II.1 the inequality (∗) is fulfilled for all r ≥ 2.
We now consider the case II.2 and assume r ≥ 2. With the help of Conclusion 6.2 and
r
P
the estimates
mi ≤ m0 − (c + 2), s + 1 ≤ c, ι(r + 1) ≤ r + 1 one sees that (∗) follows
i=1
from
1
c
2
2
.
m0 (m0 + 1) − (2m0 − 2) ≥ (c − 1) + c(r + 1) + (c + 1)m0 + r − 2(c + 2) +
2
2
A simple computation shows that this is equivalent to
(6.5)
m0 (m0 − 2c − 5) > 3c2 + 2cr − 7c − 10 + 2r 2 .
Now we have m0 ≥ 2r (c + 2) ≥ 4(c + 2). If c = 0, then (6.5) follows from 2r+1 >
−10 + r 2 ,which is true for all r ≥ 2. Therefore we can assume c > 0. Then (6.5) follows
from 2r (c + 2)(2c + 3) > 3c2 + 2cr + 2r 2 ⇔ 2r (2c2 + 7c + 6) > 3c2 + 2cr + 2r 2 which is
true for all c ≥ 1 and r ≥ 2. We get
Conclusion 6.4. In the case II.2 the inequality (*) is fulfilled for all r ≥ 2.
6.3. The case r = 1.
Then I = yI1 (−1) + f0 OP2 (−m0 ), I1 = ℓ1 K(−1) + f1 OP2 (−m1 ), where I1 has the type
0 and K has the type −1.
69
6.3.1.
We start with the case II.1 of Proposition 3.2 .
Subcase 1: ℓ1 = y: Then one has the situation shown in Figure 6.1 and there are the
following possibilities (case II 1.5 and II 1.7, respectively):
1◦ f0 = xm0 + αN1down + βL, L ∈ L monomial such that (x, y)L ⊂ ℓK(−2).
2◦ f0 = xm0 + αN1down + βL · (z/x), L ∈ L monomial such that (x, y)L ⊂ ℓK(−2).
We treat the case 1◦ . At first, one has the possibility xm0 7→ N1down . The α-grade
of the column of xm0 changes by m0 ; the α-grade of the column of N1down changes by
ι(1) − ϕ′ (m1 ) = 1 − (m1 − 1) = 2 − m1 . Therefore the change of α-grade in the right
domain is m0 − m1 + 2. The deformation xm0 7→ L gives a change of α-grade by m0 in
the right domain. As the order of f0 is equal to 0 in the case II 1.5 (cf. Proposition 3.2c),
there are no other changes of α-grade caused by f0 .
By Proposition 3.2 again, it follows that f1 has the form of case II.1.5, where α = 0,
and the order of f1 is equal to 0. The deformation M1down 7→∈ L gives a change of αgrade by −ι(1) + ϕ′ (m1 + 1) − 1 = −ι(1) + m1 = m1 − 1. Thus in the case 1◦ one has
A ≤ max(m0 − m1 + 2, m0 ) + m1 − 1 = m0 + m1 − 1, because m1 ≥ c + 2 (Lemma 2.4).
2◦ At first, f0 defines a deformation as in the case 1◦ and gives a change of α-grade
≤ max(m0 , m0 − m1 + 2) = m0 in the right domain. But as f0 has the order ≤ 1 by
Proposition 3.2c, there is still the possibility xm0 +1 7→∈ L, which gives a change of αgrade by m0 + 1 in the right domain. As f1 again has the same form as in the case 1◦ , it
follows that A ≤ 2m0 + m1 .
Because of s + 1 ≤ c the inequality (*) follows from
1
m0 (m0 + 1) − (c + m0 + m1 ) > (c − 1)2 + 2c + 2m0 + m1 .
2
As m1 ≤ m0 − (c + 2), this inequality follows from m0 (m0 − 9) > 2c2 − 2c − 6. Because of
m0 ≥ 2c + 4 it suffices to show (2c + 4)(2c − 5) > 2c2 − 2c − 6 ⇔ 2c2 > 14. This inequality
is fulfilled, if c ≥ 3. The cases c = 0, 1, 2 will be treated later on (see below).
Subcase 2: ℓ1 = x: Then from Figure 6.2 we conclude that only the second case of
Proposition 3.2 can occur, and we have
Conclusion 6.5. In the case II.1 the inequality (*) is fulfilled except if ℓ0 = y, ℓ1 = y
and c = {0, 1, 2}.
6.3.2.
We now treat the case II.2 of Proposition 3.2).
Subcase 1: ℓ1 = y: Figure 6.1 shows that only the case II.2.3 is possible. Then
there are s + 1 deformations M0 7→∈ L, · · · , xs M0 7→∈ L ( and t + 1 deformations
M1 7→∈ L, · · · , xt M1 7→∈ L, respectively). The changes of α-grade in the columns of
M0 , · · · , xs M0 (of M1 , · · · , xt M1 , respectively) is m0 , · · · , m0 +s ( and m1 −1, m1 , · · · , m1 +
t − 1, respectively). Here s and t fulfil the inequalities of ( Proposition 3.2d). Thus the
total change of α-grade in the right domain fulfils the inequality: A ≤ (s + 1)m0 + s+1
+
2
70
(t + 1)m1 +
t
2
− 1, where
κ
1
1
1
− m0 ( −
)+
and
ρ2
ρ2 ρ1 + ρ2
ρ2
κ
1
1
t≤
− m1 ( −
).
ρ2
ρ2 ρ1 + ρ2
Estimate of s: Because of κ ≤ c and m0 ≥ 2(c + 2) one obtains :
1
1
1
c
− 2(c + 2)( −
)+
s≤
ρ2
ρ2 ρ1 + ρ2
ρ2
c+1
1
1
1
1
=
− 2(c + 1)( −
) − 2( −
)
ρ2
ρ2 ρ1 + ρ2
ρ2 ρ1 + ρ2
1
1
1
2
− ) − 2( −
)
= (c + 1)(
ρ1 + ρ2 ρ2
ρ2 ρ1 + ρ2
s≤
We first consider the possibility s ≥ 0. This implies
2
ρ1 +ρ2
−
1
ρ2
> 0, i.e. ρ2 > ρ1 .
2
− x1 ; i.e. x corresponds to ρ2 and a corresponds to ρ1 , therefore 1 ≤
Let be fa (x) = x+a
√
√
a < x. fa′ (x) = −2/(x+a)2 +1/x2 < 0 ⇔ 2x2 > (x+a)2 ⇔ 2x > x+a ⇔ x > a(1+ 2).
√
√
It follows that fa (x) has the maximum for x = a(1 + 2), and fa (a(1 + 2)) = 0.171···
.
a
Therefore s ≤ 0.172(c + 1).
Estimate of t: Because of m1 ≥ c + 2 (cf. Lemma 2.4) one has
c
1
1
t≤
− (c + 2)
−
ρ2
ρ2 ρ1 + ρ2
c
1
1
1
c
−c
−
−2
−
=
ρ2
ρ2 ρ1 + ρ2
ρ2 ρ1 + ρ2
c
1
1
=
−2
−
ρ1 + ρ2
ρ2 ρ1 + ρ2
Therefore t ≤ c/2, if ρ2 ≤ ρ1 and t ≤ c/3, if ρ2 > ρ1 .
First possibility: ρ2 ≤ ρ1 : Then s < 0, i.e. there are no deformations defined by f0 ,
and A ≤ (c/2 + 1)m1 + c/2
− 1.
2
Second possibility: ρ1 < ρ2 . Then s ≤ 0.172(c + 1), t ≤ c/3 and A ≤ (0.172c +
c/3
1, 172)m0 + 0.172c+1,172
+
(c/3
+
1)m
+
− 1.
1
2
2
As m1 ≤ m0 − (c + 2) (Lemma 2.8), one obtains the following estimates:
First possibility: A ≤ (c/2+1)[m0 −(c+2)]+c2/8−1 ⇒ A ≤ (0.5c+1)m0 −3/8c2 −2c−3
Second possibility: A ≤ (0.172c + 1, 172)m0 + 0.5(0.172c + 1, 172)2 + (c/3 + 1)[m0 −
(c + 2)] + c2 /18 − 1 ⇒ A ≤ (0.51c + 2, 172)m0 − 0.25c2 − 1, 46c − 2, 3
As we have assumed that ℓ0 = y, ℓ1 = y it follows that ι(r + 1) = ι(2) = 2. As
mentioned above, only the case II.2.3 of Proposition 3.2 can occur. If c = 0, there are no
deformations at all, so that one can assume c ≥ 1. Replacing s + 1 by c, one sees that it
suffices to show:
(6.6)
Q(m0 − 1) > c2 + 1 + A
71
First possibility: A ≤ (0.5c + 1)m1 + 0.5c
− 1 (see above). One sees that it suffices
2
to show:
1
m (m0 + 1) − (c + m0 + m1 ) > c2 + 1 + (0.5c + 1)m1 + 81 c2 − 1 ⇐⇒ 12 m0 (m0 + 1) −
2 0
m0 − (0.5c + 2)m1 > 89 c2 + c.
Because of m1 ≤ m0 −(c+2) this follows from 21 m0 (m0 +1)−(0.5c+3)m0 +(0.5c+2)(c+2) >
9 2
c + c ⇐⇒ m0 (m0 − c − 5) > 1, 25c2 − 4c − 8.
8
As m0 ≥ 2c+4 this follows from (2c+4)(c−1) > 1, 25c2 −4c−8 ⇐⇒ 0.75c2 +6c+4 > 0.
This inequality is fulfilled for all c.
Second possibility: A ≤ (0.51c + 2, 172)m0 − 0.25c2 − 1, 46c − 2, 3. Putting this into
(6.6), one has to show m0 (m0 +1)−2(c+m0 +m1 ) > 1, 5c2 −2, 92c−2, 6+(1.02c+4, 344)m0.
Because of m1 ≤ m0 −(c+2) this follows from m0 (m0 +1)−2(2m0 −2)−(1, 02c+4, 344)m0 >
1, 5c2 − 2, 92c − 2, 6 ⇐⇒ m0 (m0 − 1, 02c − 7, 344) > 1, 5c2 − 2, 92c − 6, 6.
As m0 ≥ 2c + 4 this follows from (2c + 4)(0.98c − 3, 344) > 1, 5c2 − 2, 92c − 6, 6 ⇐⇒
0.46c2 + 0.152c − 6, 776 > 0.
This inequality is fulfilled if c > 3, 676 and the cases c = 0, 1, 2, 3 will be treated later on
in (6.3.3).
Subcase 2: ℓ1 = x. Figure 6.2 shows that in Proposition 3.2 f1 = M1up has to be a
monomial and that for f0 one of the two cases 2.1 or 2.3 can occur. We first treat the case
2.1, i.e., one has the deformation xm0 7→ N1up . The α-grade of the column of xm0 changes
by m0 and the α-grade of the column of N1up changes by 1. Therefore A ≤ m0 + 1. There
are no further deformations in the right domain. Now in the inequality (*) one has s = 0
and ι(2) = 1 and therefore has to show:
1
m (m0 + 1) − (c + m0 + m1 ) > (c − 1)2 + 1 + m0 + 1. Because of m1 ≤ m0 − (c + 2)
2 0
this follows from m0 (m0 − 5) > 2(c − 1)2 . As m0 ≥ 2c + 4 this follows from (2c +
4)(2c − 1) > 2(c − 1)2 ⇐⇒ 2c2 + 10c − 6 > 0. This inequality is fulfilled, if c ≥ 1.
In the case c = 0 one has (∗bis), i.e., one has to prove Q(m0 − 1) > A, i.e. to prove
1
m (m0 + 1) − (c + m1 + m0 ) > m0 + 1.
2 0
One sees that this follows from m0 (m0 − 5) > −2. As one has m0 ≥ 7 by (6.2 Auxiliary
Lemma 1), this inequality is fulfilled.
Now we treat the case 2.3, that means f0 = M0 + F as in Proposition 3.2. The only
possible deformations are M0 7→∈ L, xM0 7→∈ L, · · · , xs M0 7→∈ L. As c = 0 implies that
I is monomial, we can assume c > 0.
We again distinguish two cases:
First possibility: ρ2 ≤ ρ1 . Then s + 1 = 0, i.e. there is no deformation at all, and
therefore one has A = 0. Then (*) reads 21 m0 (m0 + 1) − (c + m0 + m1 ) > (c − 1)2. Because
of m1 ≤ m0 − (c + 2) this follows from m0 (m0 − 3) > 2c2 − 4c − 2. Because of m0 ≥ 2c + 4
it suffices to show (2c + 4)(2c + 1) > 2c2 − 4c − 2 ⇐⇒ 2c2 + 14c + 6 > 0 which is true
for all c.
Second possibility: ρ1 < ρ2 . Then s ≤ 0, 172(c + 1), as was shown above. We have
already remarked at the beginning of (6.3.2) that the s + 1 deformations noted above give
a change of α-grade A ≤ (s + 1)m0 + s+1
. It follows that A ≤ (0.172c + 1, 172)m0 +
2
72
0.5(0.172c + 1, 172)2.
In order to prove (*) it suffices to show Q(m0 − 1) > (c − 1)2 + (0.172c + 1, 172)[1 + m0 +
0.5·(0.172c+1, 172)]. Because of m1 ≤ m0 −(c+2) this follows from m0 (m0 +1)−2(2m0 −
2) > 2(c − 1)2 + (0.344c + 2, 344)(m0 + 0.086c + 1, 586) ⇐⇒ m0 (m0 − 0.344c − 5, 344) >
2, 029584c2 − 3, 252832c + 1, 717584.
Now c ≥ 1 and m0 ≥ 2c + 4 so that it suffices to show (2c + 4)(1, 656c − 1, 344) >
2, 029584c2 − 3, 252832c + 1, 717584 ⇐⇒ 1, 282416c2 + 7, 188832c − 7, 093584 > 0.
This is true if c ≥ 1. Therefore the inequality (*)is fulfilled in Subcase 2 .
Conclusion 6.6. In the case II, if r = 1, then (*) is fulfilled except in the case II.1
if ℓ0 = y, ℓ1 = y and c ∈ {0, 1, 2} or in the case II.2 if ℓ0 = y, ℓ1 = y and c ∈ {0, 1, 2, 3}.
6.3.3. The cases 0 ≤ c ≤ 4. c = 0 From Figure 6.3 it follows that M0 7→ N1
is the only possible deformation, which is the case II.1. The change of α-grade is
A = m0 + m1 − 2(m1 − 1) = m0 − m1 + 2 and the inequality (∗bis) reads Q(m0 − 1) >
m0 −m1 +2 ⇐⇒ m0 (m0 −3) > 4, and this is true, as m0 ≥ 7 by (6.2 Auxiliary Lemma 1).
c = 1 At first, we note that K is monomial. From Figure 6.4 it follows that there are
three deformations, which can occur, with the “total” change of α-grade B:
1. M0 7→ L, B = m0 + 2
2. M0 7→ N1 and M1 7→ L, B = (m0 − m1 + 2) + (2m1 − m1 − 1) + 2 = m0 + 3
3. M1 7→ L, B = m1 + 1.
Then (∗bis) follows from Q(m0 −1) > m0 +3 ⇐⇒ 12 m0 (m0 +1)−(1+m0 +m1 ) > m0 +3.
As m1 ≤ m0 − 3 (Lemma 2.8), one sees that it suffices to show m0 (m0 − 5) > 2. This is
true, because of m0 ≥ 2(c + 2) = 6 (loc.cit.).
c = 2 Then K is monomial, and the possible deformations are shown in Figure 6.5
a, b. If the case II.1 occurs, then f0 = m0 + αN1 + βF and f1 = M1 + γL. One sees that
the change of α-grade in the right domain becomes maximal, if the deformations M0 7→ F
and M1 7→ L occur. Therefore A ≤ m0 + 2m1 − m1 − 1 = m0 + m1 − 1. The inequality (*)
follows from Q(m0 − 1) > 12 + 2 · 2 + (m0 + m1 − 1) ⇐⇒ 21 m0 (m0 + 1) − (2 + m0 + m1 ) >
4 + m0 + m1 ⇐⇒ 21 m0 (m0 + 1) − 2m0 − 2m1 > 6. Because of m1 ≤ m0 − (c + 2) = m0 − 4
it suffices to show 12 m0 (m0 + 1) − 4m0 > −2 ⇐⇒ m0 (m0 − 7) > −4. This is true as
m0 ≥ 2c + 4 = 8.
If the Figure 6.5a occurs, then only the case 1◦ of (6.3.1) is possible and the change
of α-grade in the right domain is A ≤ m0 + m1 − 1. One gets the same estimate of A if
Figure 6.5a occurs in the case II.2, because xF and xL are elements of K and thus the
order of f0 and of f1 is equal to 0. Then (*) reads Q(m0 − 1) > 12 + 2 · 2 + (m0 + m1 − 1),
and this is true as was shown just before.
If Figure 6.5b occurs, then the order of f0 can be equal to 1, but it is not possible
that f1 = M1 + αF , where α 6= 0, because K is monomial and F ∈
/ K, whereas f1 ∈ K by
Lemma 2.6.
73
We want to sharpen the estimate of (6.3). M0 7→ F and xM0 7→ L yield A =
2m0 + 1; M0 7→ F, M1 7→ L yield A = m0 + m1 − 1 and M0 7→ N1 , M1 7→ L yield
A = (m0 − m1 + 2) + m1 − 1. Therefore A ≤ 2m0 + 1, and (*) reads 12 m0 (m0 + 1) − (2 +
m0 +m1 ) > 12 +2·2+2m0 +1. Because of m1 ≤ m0 −4 it suffices to show m0 (m0 −7) > 8.
As m0 ≥ 2c + 4 = 8, the case c = 2, m1 = 4, m0 = 8 remains (cf. Fig. 6.5c). One sees
that the deformations M0 7→ F, xM0 7→ L give the maximal change of “total” α-grade
B = (8 + 9) + 1 + 2 = 20. As Q(7) = 92 − 14 = 22 the inequality (!) in (5.3) is fulfilled.
c = 3 Because of Conclusion 6.6 one has only to consider deformations of the form
f0 = M0 + F, f1 = M1 + G, where F, G ∈ L (cf. Proposition 3.2). From the computations
in (6.3.2) it follows that the order of f0 is ≤ 0.172(3 + 1) < 1 and the order of f1 is ≤ 1.
Therefore the change of α-grade in the right domain is A ≤ m0 + (m1 − 1) + m1 . If one
replaces m1 by m0 − 5 ≥ m1 in the inequality (*) of (5.3), then one has to show:
1
m (m0 + 1) − (2m0 − 2) > 22 + 3 · 2 + m0 + 2(m0 − 5) − 1 ⇐⇒ m0 (m0 − 9) > −6. This
2 0
is true because of m0 ≥ 2 · (3 + 2) = 10 (cf. Lemma 2.8).
Conclusion 6.7. Even in the cases c ∈ {0, 1, 2, 3} the inequalities (*) and (!) respectively, are fulfilled.
Summarizing the Conclusions 6.1–6.7, we obtain:
Proposition 6.1. If ρ1 > 0, ρ2 > 0, I has the type r ≥ 1 and has y-standard form,
then the inequality (!) of Section (4.3) is fulfilled.
74
Fig. 6.1
monomials
N1 M1
0
1
2
3
4
5
6
7
m1 m1 +1 10
11
12
13
M0
m0
c c+1
LB
RB
Fig. 6.2
M1
N1
monomials
0
1 . . . . . . . . . c+1c+2 . . .
LB
m1 m1 +1 . . .
RB
75
. . . . . . . . . m0
Fig. 6.3
N1 M1
0
1
2
3
M0
4 m1 m1 +1 . . . m0
Fig. 6.4
L
N1 M1
0
1
2 . . . . . . . . .m1 +1. . . m0
Fig. 6.5a
L
F
N1 M1
0
1
2
3
4
5 m1 +1. . . m0
76
Fig. 6.5b
F
Fig. 6.5c
L
F
L
N1 M1
0
1
2
3
4
5
6
N1 M1
M0
7 8
0
77
1
2
3
4
5
6
M0
7 8
CHAPTER 7
Estimates of the α-grade in the case ρ1 > 0, ρ2 > 0 and r = 0.
Let I be an ideal of type r = 0 with y-standard form. Then one has by definition:
I = yK(−1) + f OP2 (−m), colength (K) = c, reg(K) = κ, colength (I) = d = c + m, I
and K invariant under G = Γ · T (ρ), and if the Hilbert functions of K and I are ψ and
ϕ, respectively, then g ∗ (ϕ) > g(d). By definition, K has the type (−1), that means, the
following cases can occur:
1st case: g ∗ (ψ) ≤ g(c), where c ≥ 5 by convention.
2nd case: 0 ≤ c ≤ 4.
As usual the aim is to prove (*) in (5.3), and we write m and f instead of m0 and f0 .
7.1. The case g ∗ (ψ) ≤ g(c).
7.1.1. We first assume that there is no deformation into L, at all. That means f =
x (cf. Fig. 7.1), A = 0 and one has to prove the inequality (*), i.e. Q(m − 1) > (c − 1)2 .
Because of Q(m−1) = 21 m(m+1)−(c+m), this is equivalent to m(m−1) > 2c2 −2c+2. As
m ≥ 2c+1 by Corollary 2.2, it suffices to show (2c+1)·2c > 2c2 −2c+2 ⇐⇒ c2 +2c−1 > 0,
and this is true for c ≥ 5.
m
m
s
7.1.2. We now assume that there are
M0 = x 7→∈ L, · · · , x M0 7→∈
deformations
1
L. Then by Proposition 3.2d s ≤ ρκ2 − m ρ12 − ρ1 +ρ
.
2
Auxiliary lemma 1. If g ∗ (ψ) ≤ g(c) and if there is the deformation M0 7→∈ L, then
ρ1 < ρ2 .
Proof. Because M0 X νρ = L is a monomial in L, the slope of the line connecting M0
and L is ≤ the slope of the line connecting M0 and L0 (see Figure 7.1, take into account
the inclusion (3.1) in 3.3 and the following remark concerning the vice-monomial of f .)
It follows that ρ1 /ρ2 ≤ κ/(m − κ). Now κ/(m − κ) < 1 is equivalent to 2κ < m, and
because of κ ≤ c and Corollary 2.2 this is true.
Auxiliary lemma 2. If g ∗ (ψ) ≤ g(c), then κ = reg(K) ≤ c − 2.
Proof. One has κ ≤ c and from κ = c it follows that g ∗ (ψ) = (c − 1)(c − 2)/2 > g(c)
(cf. [T1], p. 92 and Anhang 2e, p. 96). By considering the figures 7.2a and 7.2b one can
convince oneself that κ = c − 1 implies g ∗ (ψ) > g(c).
79
It is clear that the change of α-grade in the right domain caused by the deformations
mentioned above is equal to
s+1
.
(7.1)
A = m + (m + 1) + · · · + (m + s) = (s + 1)m +
2
Because of κ ≤ c − 2 and m ≥ 2c + 1 (Corollary 2.2) one has
1
1
c−2
.
− (2c + 1)
−
s≤
ρ2
ρ2 ρ1 + ρ2
It follows that
c−2
s≤
− 2(c − 2)
ρ2
and therefore
(7.2)
1
1
−
ρ2 ρ2 + ρ2
−5
2
1
s < (c − 2)
−
ρ1 + ρ2 ρ2
1
1
−
ρ2 ρ2 + ρ2
Auxiliary lemma 3. If g ∗ (ψ) ≤ g(c) and if there is a deformation M0 7→∈ L, then
s < 61 (c − 2).
Proof. As 1 ≤ ρ1 < ρ2 (cf. Auxiliary lemma 1), one has to find an upper bound for
2
the function f : [N − {0}] × [N − {0}] → R, f (x, y) := x+y
− y1 , on the domain 1 ≤ x < y.
One can convince oneself that f (1, 2) is the maximal value.
In the case r = 0 the inequality (*) reads 12 m(m + 1) − (c + m) > (c − 1)2 + (s + 1) + A.
Putting in the expression (7.1), one gets 12 m(m + 1) − m > c2 − c + s + 2 + (s + 1)m +
s+1
⇐⇒ m(m + 1) − 2m > 2c2 − 2c + 2(s + 2) + 2(s + 1)m + s(s + 1) ⇐⇒ m(m − 1) >
2
2c2 − 2c + 2(s + 2) + 2(s + 1)m + s(s + 1). Because of the Auxiliary lemma 3 it suffices to
73 2 29
c − 18 c + 28
. As m ≥ 2c + 1 (Corollary 2.2) this follows from:
show: m m − 13 c − 37 > 36
9
5
4
73 2
29
28
(2c + 1) 3 c − 3 > 36 c − 18 c + 9 ⇐⇒ 47
c2 + 11
c − 40
> 0. As this is true if c ≥ 2, we
36
18
9
have proven (*) in the 1st case.
7.2. The cases 0 ≤ c ≤ 4.
If c = 0, then I = (y, xm ) is monomial and (*) is true, obviously. Unfortunately,
one has to consider the cases c = 1, · · · , 4 separately. The ideal K can have the Hilbert
functions noted in (2.2.2).
7.2.1. c = 1 . From Figure 7.3 we see that deg C0 = 2 + · · · + m − 1 = m2 − 1
and deg C∞ = 1 + · · · + m = m+1
. The deformation of C0 into C∞ is defined by
2
m
m−1
f0 = x + yz
, and this is a simple deformation in the sense of ([T1], 1.3).
One has α-grade (I) = deg(C) = max(deg C0 , deg C∞ ) (loc.cit. Hilfssatz 2, p. 12).
In order to prove (*) one has to show: Q(m − 1) > deg C∞ − deg C0 = m + 1 ⇐⇒
1
m(m + 1) − (1 + m) > m + 1 ⇐⇒ m2 − 3m − 4 > 0. This is true if m ≥ 5.
2
Conclusion 7.1. If c = 1, then (*) is fulfilled except in the case c = 1, m = 4, which
will be treated in (9.3).
80
7.2.2. c = 2 . There are two subcases, which are shown in Fig. 7.4a and Fig. 7.4b.
Here only simple deformations occur, again.
1st subcase (Fig. 7.4a). One gets deg C0 = 1 + 3 + 4 + · · · + m − 1 = m2 − 2; deg C∞ =
2 + 3 + · · · + m = m+1
− 1. The same argumentation as in (7.2.1) shows that one has
2
to show Q(m − 1) > m + 1, i.e. m2 − 3m − 6 > 0. This is fulfilled if m ≥ 5.
In the case m = 4, by the formula of Remark 2.2, it follows that g ∗ (ϕ) = 4. As g(6) = 4
and g ∗(ϕ) > g(d) by assumption, the case m = 4 cannot occur.
2nd subcase (Fig. 7.4b). deg C0 = 2 + · · · + m − 1, deg C∞ = 2 + · · · + m; Q(m − 1) >
deg C∞ − deg C0 = m ⇐⇒ m2 − 3m − 4 > 0. This is true if m ≥ 5, and the case m = 4
cannot occur.
Conclusion 7.2. If c = 2, then (*) is fulfilled.
7.2.3. c = 3 . Here 5 deformations are possible, which are all simple (Fig. 7.5a-7.5e).
1st subcase (Fig. 7.5a). deg C0 = 3 + · · · + m − 1 = m2 − 3; deg C∞ = 2 + 4 + 4 + · · · +
m − 1 = m2 . Q(m − 1) > 3 ⇐⇒ m2 − m > 12. This is true because of m ≥ c + 2 = 5.
2nd subcase (Fig. 7.5b). deg C0 = 3 + · · · + m − 1; deg C∞ = 1 + 3 + 4 + · · · + m.
Q(m − 1) > deg C∞ − deg C0 = m + 1 ⇐⇒ m2 − 3m − 8 > 0. This is true because of
m ≥ 5.
3rd subcase (Fig. 7.5c). deg C0 = 3 + · · · + m − 1; deg C∞ = 2 + · · · + m. Q(m − 1) >
deg C∞ − deg C0 = m + 2 ⇐⇒ m2 − 3m − 10 > 0. This is fulfilled if m ≥ 6. As
m ≥ c + 2 = 5, the case m = 5 remains. But if m = 5, then from Remark 2.2 it follows
that g ∗ (ϕ) = 8 < g(8) = 9, which contradicts the assumption.
4th subcase (Fig. 7.5d). deg C0 = 1 + 2 + 4 + · · · + m − 1; deg C∞ = 1 + 3 + · · · +
m; Q(m − 1) > deg C∞ − deg C0 = m + 1 ⇐⇒ m2 − 3m − 8 > 0. This is fulfilled if
m ≥ 5.
5th subcase (Fig. 7.5e). deg C0 = 2+4+4+· · ·+m−1; deg C∞ = 2+· · ·+m; Q(m−1) >
deg C∞ − deg C0 = m − 1 ⇐⇒ m2 − 3m > 4. This is fulfilled if m ≥ 5.
Conclusion 7.3. If c = 3, then (*) is fulfilled.
7.2.4. c = 4 . There are 8 subcases which are shown in Fig. 7.6a1 –7.6e. At first we
make the additional assumption that m ≥ 7. The slope ρ1 /ρ2 of the line connecting the
monomials denoted by + and - is equal to 1/2 in Fig. 7.6b1 , is ≤ 1/4 in Fig. 7.6b2 and
is ≤ 2/5 in Fig. 7.6b3 . It follows that the deformations of Fig. 7.6b1 and of Fig. 7.6b2 (
respectively the deformations of Fig. 7.6b1 and of Fig. 7.6b3 ) cannot occur simultaneously,
that means, we have only simple deformations.
1st subcase (Fig. 7.6a1 ). deg C0 = 2 + 4 + · · · + m − 1; deg C∞ = 1 + 2 + 4 + · · · +
m; Q(m − 1) > deg C∞ − deg C0 = m + 1 ⇐⇒ m2 − 3m − 10 > 0. This is fulfilled if
m ≥ c + 2 = 6.
2nd subcase (Fig. 7.6a2 ). deg C0 = 2 + 4 + · · · + m − 1; deg C∞ = 3 + 4 + · · · +
m; deg C∞ − deg C0 = m + 1, etc, as in the first subcase.
81
3rd subcase (Fig 7.6b1 ). deg C0 = 4 + 4 + 5 + · · · + m − 1 = m2 − 2; deg C∞ =
2 + 4 + 6 + 5 + · · · + m − 1; Q(m − 1) > deg C∞ − deg C0 = 4 ⇐⇒ m(m − 1) > 16. This
is fulfilled if m ≥ 5.
4th subcase (Fig. 7.6b2 ). deg C0 = 4 + 4 + 5 + · · · + (m − 1); deg C∞ = 3 + · · · +
m; Q(m − 1) > deg C∞ − deg C0 = m − 1 ⇐⇒ m2 − 3m − 6 > 0. This is fulfilled if
m ≥ 5.
5th subcase (Fig. 7.6b3 ). deg C0 = 4 + 4 + 5 + · · · + m − 1; deg C∞ = 2 + 4 + 4 + · · · +
m; Q(m − 1) > m + 2 ⇐⇒ m2 − 3m − 12 > 0. This is fulfilled if m ≥ 6.
6th subcase (Fig 7.6c). deg C0 = 3 + · · · + m − 1; deg C∞ = 3 + · · · + m; Q(m − 1) >
m ⇐⇒ m2 − 3m − 8 > 0. This is fulfilled if m ≥ 5.
7th subcase (Fig. 7.6d). deg C0 = 1 + 2 + 3 + 5 + . . . + m − 1; deg C∞ = 1 + 2 + 4 +
· · · + m; Q(m − 1) > m + 1 ⇐⇒ m2 − 3m − 10 > 0. This is fulfilled if m ≥ 6.
8th subcase (Fig. 7.6e). deg C0 = 2 + 4 + 6 + 5 · · · + m − 1 = m2 + 2; deg C∞ =
; Q(m − 1) > m − 2 ⇐⇒ m2 − 3m − 4 > 0. This is fulfilled,
2 + 4 + 4 + · · · + m = m+1
2
if m ≥ 5.
As m ≥ c + 2 = 6, the case m = 6 remains. All inequalities are fulfilled if m ≥ 6;
therefore the possibility remains that in Fig. 7.6b1 and Fig. 7.6b3 , if m = 6 and ρ =
(−3, 1, 2), the deformations f = (x3 y + ay 2 z 2 ) · z 2 and g = x6 + by 2 z 4 simultaneously
occur. Now f ∧ g = x3 yz 2 ∧ x6 + bx3 yz 2 ∧ y 2 z 4 + ay 2 z 4 ∧ x6 and therefore
max −α-grade (I) ≤ max(deg C0 in Fig. 7.6b1 , deg C∞ in Fig. 7.6b3 , deg C∞ in Fig. 7.6b1 )
m
m
m
m+1
= max m2 − 2, m+1
,
+
2
and
min
−
α
−
grade(I)
=min
−
2,
,
+
2
.
2
2
2
2
2
m+1
m
m+1
Now 2 > 2 + 2, if m ≥ 3, and therefore max −α − grade(I) = 2 , and
m
min − α − grade(I) = m2 − 2. Thus (*) follows from Q(m − 1) > m+1
−
+ 2 ⇐⇒
2
2
2
m − 3m − 12 > 0, and this is true if m ≥ 6.
Conclusion 7.4. If c = 4, then (*) is fulfilled.
7.2.5.
Summarizing the Conclusions 7.1–7.4 we obtain
Proposition 7.1. If ρ1 > 0, ρ2 > 0 and r = 0, then (*) is fulfilled except in the case
c = 1, m = 4, I = (y(x, y), x4 + yz 3 ).
Notabene. Hence the inequality (!)in Section 4.3 is fulfilled with one exception .
82
Fig. 7.1
first column filled
with monomials
of H 0 (K(c))
L0
monomial domain
0
1
2
3
4
5
6
7
8
κ κ+1
M0
m0
c c+1
LB
RB
Fig. 7.2b
Fig. 7.2a
0
1 2 3 4 5 6 7 8 9
colength(K) = 10, reg(K) = 9
0
83
1 2 3 4 5 6 7 8 9 10
colength(K) = 11, reg(K) = 10
Fig. 7.3
Fig. 7.4a
Fig. 7.4b
−
0
1
2
M
3 ... ... m
0
+
2 ... ... m
1
−
0
1
2
+
3 ... ... m
Fig. 7.5c
Fig. 7.5a
−
0
1
2
Fig. 7.5b
−
+
3 ... ... m
0
1
2
−
+
3 ... ... m
0
1
Fig. 7.5d
2
+
3 ... ... m
Fig. 7.5e
−
0
1
2
+
3 ... ... m
84
−
0
1
2
+
3 ... ... m
Fig. 7.6a1
Fig. 7.6a2
−
−
0
1
2
3
+
4 ... ... m
0
1
2
Fig. 7.6b1
−
0
1
2
Fig. 7.6b2
−
+
3
3
+
4 ... ... m
4 ... ... m
0
85
1
2
3
+
4 ... ... m
Fig. 7.6b3
Fig. 7.6c
−
0
1
2
−
3
+
4 ... ... m
0
1
2
Fig. 7.6d
3
+
4 ... ... m
Fig. 7.6e
−
0
1
2
3
−
+
4 ... ... m
0
86
1
2
3
+
4 ... ... m
CHAPTER 8
Borel-invariant surfaces and standard cycles
8.1. Preliminary remarks.
d
The group GL(3; k) operates on S by
matrices, thus
GL(3; k) operates on H =
1 0 ∗
d
2
Hilb (Pk ). We recall the subgroups Γ = 0 1 ∗ < U(3; k) and T (ρ) < T :=
0 0 1
T (3; k), already introduced in (2.4.1). As each subspace U ⊂ Sn is invariant under
D := {(λ, λ, λ)|λ ∈ k ∗ }, each ideal I ⊂ OP2 and each point of H d is invariant under D.
Instead of the operation of T on H d one can consider the operation of the group T /D,
which is isomorphic to T (2; k).
According to the theory of Hirschowitz, the 1-cycles in H d which are invariant under
B = B(3; k), form a generating system of the first Chow group of H d , and relations among
them are realized in B-invariant surfaces V ⊂ H d ([Hi], Mode d’emploi, p. 89).
We distinguish between the cases, whether V is pointwise invariant under the Gm operation σ(λ) : x 7→ x, y 7→ y, z 7→ λz, or not. This we call the homogeneous case and
the inhomogeneous case, respectively.
8.2. The inhomogeneous case.
We let T = T (3; k) operate by diagonal matrices and let Ga operate by ψα : x 7→
x, y 7→ αx + y, z 7→ z on S. Then the group G = Ga · T operates on H d . Let V ⊂ H d be
a G-invariant, closed, two-dimensional variety, which is not pointwise invariant under Ga
and is not pointwise invariant under the Gm -operation σ(λ) : x 7→ x, y 7→ y, z 7→ λz. (For
abbreviation we speak of an inhomogeneous surface.)
We suppose k = k. If ξ ∈ H d (k) is a closed point, then Tξ and Gξ denote the inertia
group of ξ in T and G, respectively.
8.2.1. Auxiliary lemma 1. There is a point ξ ∈ V (k) such that V = T · ξ.
Proof. If dim T · ξ < 2 for all ξ ∈ V (k), then ξ is T -invariant or T · ξ is a T -invariant,
irreducible curve, ∀ ξ ∈ V (k). The first case can occur for only finitely many points; in the
second case one has Tξ = T (ρ), ρ ∈ Z3 − (0), such that ρ0 + ρ1 + ρ2 = 0 ([T1], Bemerkung
1, p.2).
There are only finitely many ρ, such that there exists an ideal I ⊂ OP2 of fixed
colength d, which is invariant under T (ρ), but not invariant under T (see [T2], Hilfssatz
6, p. 140). In other words, there are only finitely many ρ, such that (H d )T (ρ) 6⊇ (H d )T .
87
S
From the assumption follows V = {V T (ρi ) |i ∈ I}, where I is a finite set of indices. As
V is irreducible, it follows that V = V T (ρ) for a certain ρ. Now one has
Gg(ξ) = gGξ g −1 ∀ ξ ∈ V (k), ∀g ∈ G(k),
and therefore
Gg(ξ) ⊃ T (ρ) ∪ gT (ρ)g −1, ∀ ξ ∈ V (k),
∀ g ∈ G(k).
We show there are λ0 6= λ1 in k ∗ such that τ = (λ0 , λ1 , 1) ∈ T (ρ). We assume that
(λ0 , λ1 , λ2 ) ∈ T (ρ) implies λ0 /λ2 = λ1 /λ2 , i.e. λ0 = λ1 . Then each element in T (ρ) has
the form (λ, λ, λ2), thus T (ρ) ⊂ Gm · D, where D = {(λ, λ, λ)|λ ∈ k ∗ }. But then ρ2 = 0
and V is pointwise Gm -invariant, contradiction.
We thus take τ = (λ0 , λ1 , 1) ∈ T (ρ) such that λ0 =
6 λ1 and take
1 α 0
0 1 0 ∈ Ga < G, where α 6= 0.
g=
0 0 1
Then
gτ g −1
is an element of Gg(ξ) and thus
τ −1 gτ g −1
λ0 (λ1 − λ0 )α 0
λ1
0
= 0
0
0
1
1 β 0
= 0 1 0
0 0 1
β := λ−1
0 (λ1 − λ0 )α 6= 0
is an element of Gg(ξ) , too. It follows that g(ξ) is invariant under ψβ , thus invariant under
Ga and therefore ξ is invariant under Ga . This is valid for all ξ ∈ V (k), contradicting the
assumption.
8.2.2. If one lets T (2; k) operate on S by x 7→ λ0 x, y 7→ λ1 y, z 7→ z then one sees
that G = T · Ga operates on H d in the same way as the group T (2; k) · Ga which for
simplicity is again denoted by G (and is equal to B(2; k)). If ξ is as in the Auxiliary
lemma 1, then from V = G · ξ it follows that Gξ < G is 1-dimensional. There is the
·
S
decomposition Gξ = gi H, H := G0ξ , in finitely many cosets. As H is a 1-dimensional
connected group, two cases can occur:
1st case: H ≃ Ga . Now the unipotent elements in B(2; k) are just the matrices
ψα = ( 10 α1 ). It follows that there is α 6= 0 such that ψα ∈ Gξ . But then ξ is invariant
under Ga . As Ga is normalized by T , T ·ξ is pointwise Ga -invariant. Because of V = T · ξ,
V is pointwise Ga -invariant, contradiction.
2nd case: H ≃ Gm . As each element of Gm is semi-simple, so is each element of
the isomorphic image H. Thus the commutative group H < GL(2; k) consists of semisimple elements. Then there is an g ∈ GL(2; k), such that g −1Hg consists of diago∼
nal matrices ([K], Lemma 1, p. 150). Because of Gm ≃ H → g −1 Hg < T (2; k) one
has a 1-psg f : Gm → T (2; k). Let p : T (2; k) → Gm be the projection onto the
88
n
first component. Then
p ◦ af : Gm → Gm has the form λ → λ , n a suitable inteλ 0
λ ∈ k ∗ =: T (a, b), a and b suitable integers. It
ger. Thus g −1 Hg =
0 λb
a11 a12
−1
follows H = gT (a, b)g ⊂ G = Ga · T (2; k) = B(2; k). Writing g =
gives
a21 a22
a22 −a12
−1
−1
g =D
, D := det(g). We compute:
−a21
a11
a
a11 a12
λa 0
λ a11 λb a12
and
=
a21 a22
0 λb
λa a21 λb a22
a
λ a11 λb a12
a22 −a12
λa 0
−1
−1
=
g =D
g
−a21
a11
0 λb
λa a21 λb a22
a
λ a11 a22 − λb a12 a21 −λa a11 a12 + λb a11 a12
=
D −1
λa a a − λb a21 a22 −λa a12 a21 + λb a11 a22
a 21 22
λ a11 a22 − λb a12 a21 (λb − λa )a11 a12
.
D −1
(λa − λb )a21 a22
λb a11 a22 − λa a12 a21
This matrix is an upper triangular matrix if and only if (λa − λb )a21 a22 = 0, ∀λ ∈ k ∗ .
Subcase 1. a = b ⇒ H = T (a, a).
Subcase 2. a 6= b and a21 = 0. Write
a11 a12
a11 0
1 c
g=
= uτ, where τ =
,u =
and c := a12 /a22 .
0 a22
0 a22
0 1
Then H = gT (a, b)g −1 = uT (a, b)u−1 .
λa 0
0 λb
Subcase 3. a 6= b and a21 6= 0. Then a22 = 0 and g
g −1 =
−λb a12 a21 (λb − λa )a11 a12
−1
D
. As D = −a12 a21 , this matrix equals
0
−λa a12 a21
b
λ (λa − λb )c
, where c = a11 a12 /a12 a21 = a11 /a21 .
0
λa
b
λ (λa − λb )c
−1
∗
Thus H = gT (a, b)g =
λ ∈ k , where c := a11 /a21 . If u :=
0
λa
b
1 −c
λ (λa − λb )c
λb 0
−1
, then u
u =
and thus H = u−1 T (b, a)u.
0 1
0
λa
0 λa
We have proved
Auxiliary lemma 2. There is an element u ∈ Ga such that H = uT (a, b)u−1, where a
and b are suitable integers.
8.2.3. Let ξ and u be as in Auxiliary lemma 1 and 2, respectively. Set ζ := u−1 (ξ).
Then Gζ = u−1 Gξ u ⊃ u−1 Hu = T (a, b) < T (2; k) and thus dim T (2; k) · ζ ≤ 1. If this
dimension would be equal to 0, then Gζ and Gξ would have the dimension 2. Because
of V = G · ξ the dimension of V would be 1, contradiction. Thus dim T · ζ = 1 and
(Appendix E, Corollary) gives
89
Auxiliary lemma 3. The inertia group Tζ of ζ in T (3; k) has the form T (ρ), where
ρ = (ρ0 , ρ1 , ρ2 ) and ρ2 6= 0.
8.2.4. Summary.
Proposition 8.1. Let V ⊂ H d be a closed 2-dimensional subvariety, invariant under
G = Ga · T (3; k), where Ga operates by ψα : x 7→ x, y 7→ αx + y, z 7→ z and T (3; k)
operates by diagonal matrices on S. We suppose that V is not pointwise invariant under
this Ga -operation and is not pointwise invariant under the Gm -operation σ(λ) : x 7→
x, y 7→ y, z 7→ λz. Then there is a point ξ ∈ V (k) and u ∈ Ga such that :
(i) V = T (3; k) · ξ (ii) The inertia group Tζ of ζ := u(ξ) in T (3; k) has the form T (ρ),
where ρ = (ρ0 , ρ1 , ρ2 ) and ρ2 6= 0. (iii) V = Gm · Ga · ζ.
Proof. The statements (i) and (ii) follow from (8.2.1) – (8.2.3). Put G := Gm · Ga =
Gm × Ga . If the statement (iii) were wrong, then the inertia group Gζ would have
a dimension ≥ 1 and thus would contain a subgroup H isomorphic to Gm or Ga . If
H ≃ Gm , then ζ would be invariant under p1 (H) = {(λn , 1)|λ ∈ k ∗ }, n ∈ Z − (0). Then
ζ and ξ would be invariant under Gm and thus V would be pointwise Gm - invariant,
contradiction.
If H ≃ Ga then ζ and ξ would be invariant under Ga . As Ga is normalized by T (3; k),
T (3; k) · ξ would be pointwise Ga -invariant and the same would follow for V .
8.3. The homogeneous case.
We now assume V ⊂ H d is a 2-dimensional subvariety, invariant under G :=
Ga · T (3; k), not pointwise invariant under Ga , but now pointwise invariant under the
Gm -operation σ as in (8.1). As there are only finitely many T (3; k)-fixed points in H d , it
follows that V is not pointwise fixed by the Gm -operation τ (λ) : x 7→ λx, y 7→ y, z 7→ z.
Let ξ ∈ V (k) be not Ga -invariant and not Gm -invariant. We define a morphism f by
f : Ga × Gm → V , (α, λ) 7→ ψα τ (λ)ξ.
Assume that f has an infinite fibre. Then there is an element (β, µ) ∈ Ga × Gm such
that ψα τ (λ)ξ = ψβ τ (µ)ξ, i.e.
(8.1)
ψ(α−β)λ−1 (ξ) = τ (λ−1 µ)ξ
for infinitely many (α, λ) in Gm × Ga .
By assumption, C := {ψα (ξ)|α ∈ k}− and D := {τ (λ)(ξ)|λ ∈ k ∗ }− are curves in V . If
one assumes that only finitely many different λ can occur in (8.1), then on the left side
of (8.1) also only finitely many α can occur. For ξ is not a fixed point of Ga , so that from
ψα (ξ) = ψβ (ξ) it follows that α = β. This contradicts the last assumption. Thus C and
D have in common infinitely many points and hence are equal (as subschemes with the
induced reduced scheme structure).
The fixed-point set of C under Gm consists of the two points ξ0/∞ := lim τ (λ)ξ. As
λ→0/∞
C has an unique Ga -fixed-point, and as Ga is normalized by Gm , one of the two points, say
90
ξ∞ , is fixed by Ga . Thus C = {ψα (ξ0)|α ∈ k}− and ξ0 corresponds to a monomial ideal.
n
S
There are only finitely many T (3; k)-fixed points ξi , 1 ≤ i ≤ n, in H d . Set M := Ga · ξi .
1
Then C \ M is open and non empty, and choosing ξ ∈ C \ M it follows that f as defined
above has no infinite fiber.
Proposition 8.2. Let V ⊂ H d be a closed 2-dimensional subvariety, invariant under
G = Ga · T (3; k), not pointwise Ga -invariant, but pointwise invariant under the Gm operation σ(λ) : x 7→ x, y 7→ y, z 7→ λz. Then V is not pointwise invariant under the
Gm -operation τ (λ) : x 7→ λx, y 7→ y, z 7→ z. And for all closed points ζ in an open non
empty subset of V one has V = Ga · Gm · ζ.
8.4. Standard cycles.
In the following we suppose d ≥ 5 and we recall to memory the closed subscheme
[
H = {Hϕ ⊂ H d | g ∗ (ϕ) > g(d)}
of H d (cf. 1.2.2). As usual, we let Ga operate by ψα : x 7→ x, y 7→ αx + y, z 7→ z and we
let Gm operate by σ(λ) : x 7→ x, y 7→ y, z 7→ λz (by σ(λ) : x 7→ λx, y 7→ y, z 7→ z) on S in
the inhomogeneous case (in the homogeneous case, respectively).
8.4.1.
Definition 7. Let C ⊂ H be a B = B(3; k) -invariant curve, which is not pointwise
−
Ga -invariant. Then C={ψα (ξ)|α ∈k}
, where ξ ↔ I is an ideal, invariant under
1 0 ∗
T = T (3; k) and Γ = 0 1 ∗ < ∆ := U(3; k), which is uniquely determined by
0 0 1
C (cf. [T1], Proposition 0, p. 3). We call C a x-standard cycle respectively a y-standard
cycle, if I has x-standard form respectively y-standard form (see 2.4.3 Definition 2).
8.4.2. Let V ⊂ Z = Z(H) be a 2-dimensional, B-invariant subvariety, where Z is
defined as in (1.2.1). We suppose, that V contains a y-standard cycle. Then V is not
pointwise Ga -invariant, so that we can write V = Gm · Ga · ζ, where ζ ∈ V (k) is as in
Proposition 8.1 or Proposition 8.2 , respectively. By the definition of Z, the orbit ∆ · ζ
has a dimension
ζ is not ∆-invariant,the inertia group ∆ζ of ζ in ∆ has the form
≤ 1. As
1 α ∗
G(a : b) = 0 1 β ∈ ∆|aα + bβ = 0 , where a, b ∈ k and not both elements are
0 0 1
zero (cf. Appendix D, Lemma 1). Let I be the ideal corresponding to ζ. If ϕ is the Hilbert
function of I, then g ∗ (ϕ) > g(d) by the definition of H and thus I = ℓK(−1)+f OP2 (−m),
where K ⊂ OP2 has the colength c, f ∈ Sm , c+m = d and m = reg(I) ≥ c+2 ( see Lemma
2.1 – Lemma 2.4). If ν := min{n|H 0(I(n)) 6= (0)}, then ν < m . This follows from Lemma
2.2 and Corollary 2.1 . As G(a : b) is unipotent, there is an eigenvector f ∈ H 0(I(ν)).
From x ∂f /∂z ∈ hf i and bx ∂f /∂y − ay∂f /∂z ∈ hf i we conclude that f = xν , if we
assume b 6= 0, which we do now. Let be η ∈ V (k) any point. If L is the corresponding
91
ideal, then xν ∈ H 0 (L(ν)). ( Proof : This is first of all true for all η ∈ W := Gm · Ga · ζ.
By means of the mapping J 7→ H 0 (J (d)) we can regard H d as a closed subscheme of
G = GrassQ(d) (Sd ). If J ∈ H d is any ideal, the condition xν ∈ H 0 (J (ν)) is equivalent to
the condition xν Sd−ν ⊂ H 0 (J (d)). An element of G(Spec A) is a subbundle L ⊂ Sd ⊗ A
of rank Q(d), and the condition xν Sd−ν ⊂ L defines a closed subscheme of G. Thus the
condition xν ∈ H 0 (L(ν)) defines a closed subset of V . As V = W , this condition is fulfilled
for all points of V .) Assume that L has y-standard form. Then L = y·M(−1)+gOP2 (−n),
where e := colength (M), n := reg(L) ≥ m, and e + n = d. As ν < m ≤ n we get
xν ∈ H 0 (L(ν)) = yH 0(M(ν − 1)), contradiction.
Lemma 8.1. If V ⊂ Z(H) is a B-invariant surface which contains a y-standard cycle,
then V is point-wise invariant under Γ.
Proof. From the foregoing reasoning it follows that b = 0, i.e., ζ is invariant under
Γ = G(1 : 0). As Γ is normalized by Ga and T , it follows that Gm · Ga · ζ is point-wise
invariant under Γ, and the same is true for V .
8.4.3. We suppose that V ⊂ Z(H) is a B-invariant surface, which contains a ystandard cycle. We represent V in the form V = Gm · Ga · ζ, according to Proposition
8.1 or Proposition 8.2, respectively.
Lemma 8.2. (a) One can assume without restriction that J ↔ ζ has y-standard form.
(b) I0/∞ are monomial ideals and have y-standard form.
Proof. Let be ζ ↔ I as in Proposition 8.1 (resp. in Proposition 8.2 ).
(a) By Lemma 2.6), I has x- or y-standard form. First we consider the inhomogeneous
case . From I = xK(−1) + f OP (−m) and the Γ-invariance of I (Lemma 8.1), the Γinvariance of K follows. By (Appendix C, Remark 2) we have Rc ⊂ H 0 (K(c)), and
because of m ≥ c + 2 we get Rm−2 ⊂ H 0 (K(m − 2)), thus xRm−2 ⊂ xH 0 (K(m − 2)) =
H 0 (I(m − 1)). We conclude that xRm−2 ⊂ H 0 (J (m − 1)), hence xRm−2 · Sd−m+1 ⊂
H 0 (J (d)), if J ↔ η is any point of Gm · Ga · ζ. The same reasoning as in (8.4.2) shows
that xRm−2 · Sd−m+1 ⊂ H 0 (J (d)) and hence xRm−2 ⊂ H 0 (J (m − 1)) is true for all
J ↔ η ∈ V . As J = ℓM(−1) + gOP2 (−n), e = colength (M), n = reg(J ) ≥ m, e + n = d
and ℓ ∈ R1 − (0), it follows that H 0 (J (m − 1)) = ℓH 0 (M(m − 2)) ⊂ xRm−2 , hence ℓ = x.
It follows, that no ideal in V has y-standard form, contradiction. Thus the statement (a)
is proved in the inhomogeneous case. If the homogeneous occurs and if we assume that
ζ ↔ I would have x-standard form, the same argumentation as in the inhomogeneous case
gives a contradiction. By Lemma 2.2 we have a representation I = ℓK(−1) + f OP2 (−m),
and if the homogeneous case occurs, because of the Γ-invariance of I, we conclude that
ℓ = −βx + y, where β ∈ k. Replacing I by ψβ (I) = yψβ (K)(−1) + ψβ (f )OP2 (−m), we
can assume without restriction that I has y-standard form.
(b) If I = yK(−1) + f OP2 (−m), colength (K) = c, reg(I) = m, c + m = d, m ≥ c + 2,
then Rm−2 ⊂ H 0 (K(m − 2)) (Appendix C, Remark 2), thus yRm−2 ⊂ H 0 (I(m − 1)),
hence yRm−2 ⊂ H 0 (σ(λ)I(m − 1)), λ ∈ k ∗ . As yRm−2 ⊂ H 0 (J (m − 1)) is equivalent
to yRm−2 · Sd−m+1 ⊂ H 0 (J (d)), the condition yRm−2 ⊂ H 0 (J (m − 1)) defines a closed
92
subscheme of V , which is invariant under the Gm -operation σ, as one sees by the same
reasoning as in the proof of Lemma 8.1 . It follows that yRm−2 ⊂ H 0(I0/∞ (m − 1)). By
Lemma 2.2 we can write I0 = ℓM(−1) + gOP2 (−n), where n = reg(I0 ) ≥ m. But then
yRm−2 ⊂ H 0 (I0 (m − 1)) = ℓH 0 (M(m − 2)), showing that ℓ = y. The same argument
shows that I∞ has y-standard form, too.
In the inhomogeneous case ζ is fixed by T (ρ), where ρ2 6= 0 ( Proposition 8.1) , hence
ζ0/∞ are fixed by T (ρ) · Gm = T (3; k)
In the homogeneous case ζ0/∞ are invariant under the two Gm -operations σ and τ , hence
are invariant under T (3; k).
93
CHAPTER 9
Relations between B-invariant 1-cycles
9.1. Generalities on relations between 1-cycles
We describe the method of Hirschowitz in our special situation. B = B(3; k) operates
on H d and we take a closed subscheme X ⊂ H d , which is B-invariant. Z1B (X) is the
free group generated by B-invariant curves in X. It is easy to see that the canonical
homomorphism Z1B (X) → A1 (X) is surjective (cf. [T1], Lemma 1, p. 6). By ([Hi],
Theorem 1, p.87) the kernel RatB
1 (X) can be described as follows. One has to consider all
1
operations of B on P and all two dimensional subvarieties B ⊂ X × P1 with the following
k
properties:
(i) p2 : B → P1 is dominant, hence surjective and flat.
(ii) The operation of B on P1 fixes 0 and ∞.
(iii) B is invariant under the induced operation (ξ, µ) 7→ (gξ, gµ), ξ ∈ X, µ ∈ P1 , g ∈ B,
of B on X ×k P1 .
N.B. According to a general theorem of Fogarty, the fixed-point scheme (P1 )U (3;k) is
connected; hence from (iii) it follows that U(3; k) operates trivially on P1 .
If Bµ := p−1
2 (µ) is the fibre and Bµ := p1 (Bµ ) is the isomorphic image in X, then one
has:
g(Bµ ) = { (gξ, gµ) | ξ ∈ X, such that (ξ, µ) ∈ B }
= { (ξ, gµ) | ξ ∈ X, such that (ξ, gµ) ∈ B }
= Bgµ , for all g ∈ B, µ ∈ P1
We conclude that
(9.1)
gBµ = Bgµ , for all g ∈ B, µ ∈ P1 .
From property (ii) it follows that B0 and B∞ are B-invariant 1-cycles in X. Then
B
RatB
1 (X) is defined to be the subgroup of Z1 (X) generated by all elements of the form B0 −
B
B
B∞ , and the theorem of Hirschowitz says that AB
1 (X) := Z1 (X)/ Rat1 (X) is canonically
isomorphic to A1 (X) (loc.cit.). We consider V := p1 (B) as a closed subscheme of X with
the induced reduced scheme structure. As B ⊂ V ×k P1 , dim V ≥ 1. If dim V = 1, then
B = V × P1 and B0 − B∞ is equal to zero in Z1B (X). Thus we can assume without
restriction that V is a B-invariant, 2-dimensional subvariety of X.
95
9.2. The standard construction
Let X ⊂ H d be a B(3; k)-invariant subvariety. Start with a subvariety B ⊂ X ×k P1 as
in (9.1), such that V := p1 (B) ⊂ H d is a 2-dimensional subvariety, which is automatically
B(3; k)-invariant. Assume V is not pointwise invariant under the Ga -operation as in 8.2.
Write
(9.2)
V = Ga · Gm · ξ
where ξ ∈ V (k) is a suitable point and Gm operates by σ(λ) : x 7→ x, y 7→ y, z 7→ λz or by
σ(λ) : x 7→ λx, y 7→ y, z 7→ z, respectively (cf. Proposition 8.1 and Proposition 8.2). Then
Gm is a subgroup of T and fixes 0 = (0 : 1) and ∞ = (1 : 0) by assumption. Assume
that σ operates trivially on Gm . From (9.1) it follows that σ(λ)Bµ = Bµ , ∀λ ∈ k ∗ .
If ξ is the point of (9.2), then one chooses µ in such a way that (ξ, µ) ∈ Bµ . Then
ψα (ξ, µ) = (ψα (ξ), µ) ∈ Bµ , ∀α ∈ k, hence ψα (ξ) ∈ Bµ , ∀α ∈ k. Because of (9.1) and (9.2)
we would get V ⊂ Bµ . Then the closed subscheme Bµ ⊂ B has the dimension two and
it follows Bµ = B, contradiction, as p2 is dominant by assumption. This argumentation
also shows ξ ∈
/ B0 ∪ B∞ , i.e. there is µ ∈ P1 − {0, ∞} such that (ξ, µ) ∈ Bµ . Thus one
can find λ ∈ k ∗ such that σ(λ)µ = (1 : 1). Replacing ξ and µ by ξ ′ = σ(λ)ξ and σ(λ)µ,
respectively, then (9.2) is fulfilled with ξ ′ instead of ξ. Thus one can assume without
restriction that µ = (1 : 1). As A1 = P1 − {∞} is fixed by σ, Gm operates by σ on A1 and
fixes (0 : 1). Then there is m ∈ Z such that σ(λ)(a : 1) = (λm a : 1) for all a ∈ k, λ ∈ k ∗ .
As the action of Gm on P1 is not trivial, m 6= 0.
C := {ψα (ξ)|α ∈ k}− ⊂ B(1:1) is a curve in V ; let h be its Hilbert polynomial
with respect to the usual embedding of H d in a projective space (cf. [T2], 4.1.1). The
association λ 7→ σ(λ)C defines a morphism Gm → X := Hilbh (V ). It has an extension to
a morphism P1 → X , which defines a flat family of curves
C ֒→ V ×k P1
↓ p2
P1
∗
If Cλ := p−1
2 (λ), then Cλ := p1 (Cλ ) = σ(λ)C, ∀λ ∈ k and
(9.3)
[p1 (C0 )] = [C] = [p1 (C∞ )] ∈ A1 (V ).
The finite morphism A1 − (0) → A1 − (0) defined by λ 7→ λm has an extension
f : P1 → P1 such that (λ : 1) 7→ (λm : 1), ∀λ ∈ k, and (1 : 0) 7→ (1 : 0). For simplicity, the
morphism 1V × f : V × P1 → V × P1 is denoted by f , again. By construction C ⊂ B(1:1)
and hence σ(λ)C ⊂ σ(λ)B(1:1) = Bσ(λ)(1:1) . The fibre Cλ = σ(λ)C ×{(λ : 1)} is mapped by
f into σ(λ)C × {(λm : 1)} = σ(λ)(C × {(1 : 1)}) ⊂ σ(λ)(B(1:1) × {(1 : 1)}) = σ(λ)B(1:1) =
Bσ(λ)(1:1) , ∀λ ∈ k ∗ .
This construction of the family C is called standard construction and the proof in
([T1], Lemma 1, p.6) shows that C ⊂ V ×k P1 is a subvariety. As C0 and C∞ are closed in
o
C, C := C − (C0 ∪ C∞ ) is open in C. Because C and B are irreducible and f is projective,
96
o
from f (C) ⊂ B we conclude that f (C) = B. As V × {0} and V × {∞} are mapped by f
into itselve, C0 ⊂ V × {0} and C∞ ⊂ V × {∞} are mapped by f into B ∩ V × {0} = B0
and B ∩ V × {∞} = B∞ , respectively. As f (C) = B, we get f (C0 ) = B0 and f (C∞ ) = B∞ ,
i.e. C0 = B0 and C∞ = B∞ . The curves p1 (C0 ) and p1 (C∞ ) are called the limit curves (of
the standard construction) and are denoted by C0 = lim σ(λ)C and C∞ = lim σ(λ)C,
λ→0
respectively, and one has
B0 − B∞ = C0 − C∞
(9.4)
λ→∞
in Z1B (X).
We note that the relation (9.4) was derived under the assumption that V is not pointwise invariant under Ga . If V is pointwise invariant under Ga , the standard construction
cannot be carried out. We get
P
Lemma 9.1. Let X be as before. Elements of RatB
qi Ci ,
1 (X) are either of the form
Ga
where qi ∈ Q and Ci ⊂ X is a B(3; k)-invariant 1-prime cycle in X or they occur in
the following way: One considers all B(3; k)-invariant 2-dimensional subvarieties V ⊂ X,
which are not pointwise invariant under Ga . Then there are points ξ ∈ V (k) such that
V = Gm · Ga · ξ, where Gm operates by σ(λ) : x 7→ x, y 7→ y, z 7→ λz on S (inhomogeneous
case) or by σ(λ) : x 7→ λx, y 7→ y, z 7→ z (homogeneous case). C := {ψα (ξ)|α ∈ k}− is
a curve in V (with the reduced induced scheme structure). By means of the standard
construction it defines a family of curves C ⊂ V ×k P1 , which is flat over P1 and with the
Ga
fibres Cλ = σ(λ)C for all λ ∈ P1 − {0, ∞}. RatB
1 (X) is generated by the relations in X
noted above as well as by the relations C0 − C∞ , where C0/∞ = lim σ(λ)C are the limit
λ→0/∞
curves of C in V and these are Ga -invariant 1-cycles.
Proof. As σ(λ)ψα = ψαλ σ(λ) in the homogeneous case, σ(λ)C = {ψα σ(λ)ξ|α ∈ k}−
is Ga -invariant, ∀λ ∈ k ∗ . And the same is true in the inhomogeneous case .
9.3. Relations containing a y-standard cycle
Suppose d ≥ 5. The closed subscheme H ⊂ H d (cf. Section 8.4) is clearly B-invariant.
As U(3; k) is normalized by T (3; k), the closed subscheme X = Z(H) is B-invariant,
too (cf.Section 1.2). From Lemma 8.1 it follows that relations in Z1B (X) containing a
y-standard cycle are defined by 2-dimensional B-invariant surfaces V ⊂ X, which are
pointwise Γ-invariant but not pointwise Ga -invariant.
We can write V = Gm · Ga · ζ, where ζ ∈ V (k) is a point as in Proposition 8.1
or Proposition 8.2. By Lemma 8.2 we can assume that I ↔ ζ has y-standard form.
Let be ζ0/∞ = lim σ(λ)ζ , where σ(λ) denotes one of the two Gm - operations menλ→0/∞
tioned in Lemma 9.1 .Then carry out the standard construction by using the curve
C := { ψα (ζ) | α ∈ k }− . By Lemma 8.2 C0/∞ are y-standard cycles. We conclude from
Proposition 8.1 and Proposition 8.2 that W := Gm · Ga · ζ is B-invariant and therefore
V − W is a union of B-invariant curves and points. As C0/∞ is invariant under Gm and
Ga , from W ∩ C0/∞ 6= ∅ it would follow that V ⊂ C0/∞ . Hence C0/∞ ⊂ V − W , and
C0/∞ is a union of B-invariant curves, too. As ζ0/∞ ∈ C0/∞ , one has C0/∞ ⊂ C0/∞ . Now
97
from the Propositions 5.1, 6.1 and 7.1 it follows that the inequality (!) in ( 4.3) is fulfilled
with one exception (cf. Proposition 7.1 ). Let be m = reg(I). Because of max −α-grade
(I) ≥ α-grade (I) = deg C = deg C0 = deg C∞ and min(deg C0 , deg C∞ ) ≥ min −αgrade (I) (cf. the definitions and the inequality (4.9) in Section 4.3), from (!) it follows
that
Q(m − 1) + min(deg C0 , deg C∞ ) > deg C0 = deg C∞ .
(**)
As V = W , each ideal J in V has a regularity n ≥ m. If J is monomial and has ystandard form , then from Remark 4.4 it follows that the y-standard cycle generated by
J has a degree ≥ Q(n − 1) ≥ Q(m − 1).
Lemma 9.2. C0 and C∞ are the only y-standard cycles contained in the B-invariant
cycles C0 and C∞ , respectively, and they both occur with the multiplicity 1.
Proof. Except when c = 1, m = 4 the statements follow from (**) and the foregoing
discussion. In the remaining exceptional case (cf. Proposition 7.1 ) we carry out the
standard construction with I = (y(x, y), x4 + yz 3 ) ↔ ζ. Then ζ0 ↔ (y(x, y), x4), ζ∞ ↔
(y, x5 ) and C0 and C∞ are y-standard cycles of degrees 5 and 10, respectively, as one can
see from Figure 9.1.
If η := lim ψα (ζ), the corresponding ideal L is invariant under Ga , hence invariant
λ→∞
under U(3; k). One sees that xz 3 ∈ L, and hence x ∈ L follows. One concludes from this
that L = (x, y 5 ).
As σ and ψα commute, it follows that η is the only point in σ(λ)C for all λ, which is
S
S
U(3; k)-invariant. From V = p1 (C) = {p1 (Cλ )|λ ∈ P1 } = {σ(λ)C|λ ∈ k ∗ } ∪ C0 ∪ C∞
it follows that
V = W ∪ {η} ∪ C0 ∪ C∞
where W := Gm · Ga · ζ. Now η0 := lim ψα (ζ0 ) ↔ (x(x, y), y 4) and η∞ := lim ψα (ζ∞ ) ↔
α→∞
α→∞
(x, y 5 ) are the only B(3; k)-invariant points in C0 and C∞ , respectively. By the theorem
of Fogarty V U (3;k) is connected, hence is a union of pointwise U(3; k)-invariant, closed
and irreducible curves. If E is such a curve, then E ⊂ C0 ∪ C∞ follows. Now deg C0 =
deg C∞ = deg C = 10. Hence C∞ = C∞ and E ⊂ C0 follows. As each y-standard cycle
in V has a degree ≥ Q(3) = 5, C0 is the only y-standard cycle contained in C0 , and C0
occurs with multiplicity 1.
Proposition 9.1. Let be X = Z(H). Relations in RatB
1 (X) , which are defined by a
B-invariant surface in X and which contain a y-standard cycle, are generated by relations
of the form C1 − C2 + Z where C1 and C2 are y-standard cycles and Z is a 1-cycle in X,
whose prime components are B-invariant but are not y-standard cycles.
Proof. This follows from Lemma 9.1 and Lemma 9.2 .
98
Fig. 9.1
Fig. 9.2
C∞
C0
−
0
1
2
3
+
4
0
1
2
3
99
4
5
CHAPTER 10
Proof of Theorem 1.1
This had been formulated in Chapter 1 and we repeat the notations and assumptions
S
introduced there: d ≥ 5, g(d) = (d − 2)2 /4, H = {H≥ϕ ⊂ H d |g ∗(ϕ) > g(d)}, A(H) =
Im(A1 (HU (3;k) ) → A1 (H)), c3 = {(αx + y, xd )|α ∈ k}− .
P
We make the assumption [c3 ] ∈ A(H). Then [c3 ] =
qi [Ui ] in A1 (H), where qi ∈ Q,
and Ui ⊂ H are T -invariant curves, which are pointwise U(3; k)-invariant. (A1 (HU (3;k) )
is generated by such curves, as follows from the theorem of Hirschowitz or more directly
∼
from an application of Lemma 1 in [T1], p. 6.) If X := Z(H), then A1 (X) → A1 (H)
∼
([T2], Lemma 24, p.121) and AB
1 (X) → A1 (X) ([Hi], Theorem 1, p.87).
P
Hence we can regard α := c3 − qi Ui as a 1-cycle in Z1B (X) ,whose canonical imB
B
B
age in AB
1 (X) := Z1 (X)/ Rat1 (X) vanishes. (We recall that Z1 (X) is the free group
generated by all closed, reduced and irreducible curves in X.) Define B(X) to be the subgroup of Z1B (X), which is generated by all B(3; k)-invariant curves in X (closed, reduced,
irreducible), which are not y-standard cycles. Define:
F1 (X) = Z1B (X)/B(X), R1 (X) = RatB
1 (X) + B(X)/B(X)
F1 (X) is the free group generated by the y-standard cycles, and by Proposition 9.1 R1 (X)
is generated by elements of F1 (X) of the form C1 − C2 , where C1 and C2 are (different)
y-standard cycles. It follows that the canonical image of α in F1 (X)/R1 (X) on the one
hand vanishes and on the other hand is equal to the canonical image of c3 in this Qvector space. Hence c3 ∈ R1 (X). We will show that this is not possible. To each of the
finitely many y-standard cycles we associate a canonical basis element ei ∈ Qn , 1 ≤ i ≤ n.
Especially, we associate c3 to the element en = (0, . . . , 1). From c3 ∈ R1 (X) it follows
that en ∈ h{ei − ej |1 ≤ i < j ≤ n}i, which is not possible. All in all we have
Theorem 10.1. [c3 ] ∈
/ A(H).
Corollary 10.1. dimQ A1 (H) ≥ 3.
Proof. E := {(x2 , xy, y d−1 + αxz d−2 )|α ∈ k}− and F := {(x, y d−1 (αy + z))|α ∈ k}−
are 1-cycles in H, which are pointwise invariant under U(3; k) and G(0, 1), respectively
(see the notation in Appendix D). If [c3 ] = q1 [E] + q2 [F ], q1 , q2 ∈ Q, we compute the
intersection numbers with the tautological line bundles and get
d
= q1 + q2 (n − d + 1)
2
for all n ≥ d − 1. Hence q2 = 0. If q1 6= 0, then [c3 ] ∈ A(H) would follow.
101
CHAPTER 11
Surfaces in H invariant under Ga · T (4; k)
We let Ga operate by ψα : x 7→ x, y 7→ αx + y, z 7→ z, t 7→ z and let T (4; k) operate by
diagonal matrices on P = k[x, y, z, t], hence on H = HQ . Let V ⊂ H be a 2-dimensional
subvariety, which is invariant under Ga · T (4; k) but not pointwise invariant under Ga .
11.1. The inhomogeneous case
We suppose that V is not pointwise invariant under the Gm -operation σ(λ) : x 7→
x, y 7→ y, z 7→ z, t 7→ λt.
11.1.1. Auxiliary Lemma 1. There is a ξ ∈ V (k) such that V = T (4; k) · ξ.
Proof. If dim T (4; k) · ξ ≤ 1, then the inertia group Tξ ⊂ T (4; k) has the dimension
≥ 3. If the dimension is 4, then ξ corresponds to a monomial ideal with Hilbert polynomial
Q, and there are only finitely many such ideals. If dim Tξ = 3, then Tξ = T (ρ) ([T2],
Hilfssatz 7, p.141). If ξ corresponds to the ideal I with Hilbert polynomial Q, then
H 0 (I(b)) has a basis of the form fi = mi pi (X ρ ), where m is a monomial and pi is a
polynomial in 1 variable. At least for one index i the polynomial pi (X ρ ) contains a
positive power of X ρ . Then mi X ρ ∈ Pb . As mi ∈ Pb and ρ = (ρ0 , . . . , ρ3 ) ∈ Z4 − (0) is a
vector such that ρ0 + · · ·+ ρ3 = 0, this implies |ρi | ≤ b. Hence there are only finitely many
such vectors ρ such that the fixed-point scheme HT (ρ) does not consist of only finitely
many T (4; k)-fixed points.
n
S
Suppose that dim T (4; k) · ξ ≤ 1 for all ξ ∈ V (k). Then V = V T (ρi ) , ρi ∈ Z4 − (0),
1
hence V = V T (ρ) , ρ ∈ Z4 − (0) suitable vector. We show there is τ = (λ0 , . . . , λ3 ) ∈ T (ρ)
such that λ0 6= λ1 . If not, it follows that T (ρ) ⊂ T (1, −1, 0, 0), hence νρ = (1, −1, 0, 0), ν ∈
Z. But then ρ3 = 0 and V would be pointwise invariant under Gm , contradiction.
1 α(λ−1
−1
−1
0 λ1 − 1)
= ψβ ,
Take this τ and an arbitrary α 6= 0. Then τ ψα τ ψα =
0
1
where β := α(λ−1
0 λ1 −1) 6= 0. The same argumentation as in the proof of ( 8.2.1 Auxiliary
Lemma 1) gives a contradiction.
11.1.2. If 0 ≤ i < j ≤ 3 define Tij = {(λ0 , λ1 , λ2 , λ3 ) ∈ T (4; k)|λℓ = 1 if ℓ 6= i and
ℓ 6= j}. We let G := Ga · T23 operate von V . If ξ ∈ V (k) is as in Auxiliary Lemma 1,
let Gξ be the inertia group of ξ in G. From 1 ≤ dim Gξ ≤ 2 it follows that there is a
1-dimensional connected subgroup H < G0ξ . Then H ≃ Ga or H ≃ Gm . In the first case it
103
follows that H = Ga and hence T (4; k) · ξ is pointwise Ga -invariant, contradiction. Hence
H ≃ Gm . In the diagram
i
Gm ≃ H ֒→ G = Ga × T23
p1 ւ
ց p2
Ga
T23
p1 ◦ i is the trivial map, hence H = {(1, 1, λr , λs )|λ ∈ k ∗ } =: T23 (r, s), where r, s ∈ Z are
not both equal to zero.
λ α
∗
11.1.3. Let be G :=
λ, µ ∈ k , α ∈ k and let ξ be as in Auxiliary
0 µ
Lemma 1. Then there is again a subgroup H < G0ξ isomorphic to Gm . It has the form
a
λ (λb − λa )c
∗
H=
λ ∈ k , where a, b ∈ Z are not both equal to zero and c ∈ k
0
λb
1 −c 0 0
0 1 0 0
is a fixed element (cf. 8.2.2). Let be u =
0 0 1 0.
0 0 0 1
a
λ
b
λ
∗
−1
=: T01 (a, b). If ζ := u(ξ) one gets
λ
∈
k
Then uHu =
1
1
Tζ ⊃ T01 (a, b). From (11.1.2) it follows Tζ ⊃ T23 (r, s), too. As Tζ contains the diagonal
group, dim Tζ ≥ 3 follows. If ζ were fixed by T (4; k), then ξ = u−1 (ζ) would be fixed by
T23 , contradiction. It follows that Tζ = T (ρ). Here ρ3 6= 0, because ξ is not invariant
under σ, for otherwise V would be pointwise Gm -invariant. If c = 0, then Tξ = T (ρ),
which contradicts the choice of ξ.
Conclusion 11.1. The element u is different from 1, and putting ζ := u(ξ) one has
V = Ga · Gm · ζ.
Proof. Put G := Ga · Gm = Ga × Gm . If dim Gζ ≥ 1, then G0ζ would contain a
subgroup H isomorphic to Ga or Gm . But then ζ would be invariant under Gm or Ga ,
and then the same would be true for ξ, which gives a contradiction as above.
Conclusion 11.2. V \ Ga · Gm · ζ is a union of points and curves which are invariant
under Ga · T (4; k).
Proof. If τ ∈ T (ρ), then τ · Ga · Gm · ζ = Ga · τ · Gm · ζ = Ga · Gm · τ · ζ = Ga · Gm · ζ,
and if τ ∈ Gm , the same is true.
11.2. The homogeneous case
11.2.1. We first suppose that V is pointwise invariant under the Gm -operation σ(λ) :
x 7→ x, y 7→ y, z 7→ z, t 7→ λt, but not pointwise invariant under the Gm -operation τ (λ) :
x 7→ x, y 7→ y, z 7→ λz, t 7→ t. Then V is invariant under T (3; k) ≃ {(λ0 , λ1 , λ2 , 1)|λi ∈ k ∗ }
and we have the same situation as in Chapter 8. We use the same notations introduced
104
there
ξ (8.2.1 Auxiliary Lemma 1). We let G := Ga · T (2; k) =
and get V = T (2; k) ·
λ α
λ, µ ∈ k ∗ , α ∈ k operate on V . Then there is a subgroup H of Gξ , which
0 µ
is isomorphic to Ga or Gm . In the first case ξ would
bea Ga -invariant,
hence Vwould
λ (λb − λa )c
be pointwise Ga -invariant. In the second case H =
λ ∈ k ∗ . The
0
λb
same argumentation as in (11.1.3) shows that the inertia group of ζ = u(ξ) in T (3; k) has
a dimension ≥ 2. If Tζ = T (3; k), then ζ would be invariant under the Gm -operation τ ,
hence ξ invariant under τ , too, and V would be pointwise invariant under τ . It follows
that Tζ = T (ρ), where ρ = (ρ0 , ρ1 , ρ2 , 0) and ρ2 6= 0. We note that u 6= 1, for otherwise
the inertia group of ξ in T (4; k) would have a dimension ≥ 3.
Conclusion 11.3. The element u is different from 1 and putting ζ = u(ξ) one has
V = Ga · Gm · ζ.
The same argumentation as in (11.1.3) gives
Conclusion 11.4. V \ Ga · Gm · ζ is a union of points and curves which are invariant
under Ga · T (4; k).
11.2.2. We now suppose V is pointwise invariant under T23 = { (1, 1, λ2, λ3 ) | λi ∈ k ∗ }.
Then V is not pointwise invariant under the Gm -operation σ(λ) : x 7→ λx, y 7→ y, z 7→
Gm
z, t 7→ t. Let ξ ∈ V \ (V
∪ V Ga ) be a closed point, and put G := Ga ⋉ Gm =
λ α
λ ∈ k∗ , α ∈ k .
0 1
Assume that dim Gξ ≥ 1. Then H := G0ξ is 1-dimensional and connected. As ξ is not
a
λ (1 − λa )c
∗
, a ∈ Z − (0). Putting
Ga -invariant, H ≃ Gm hence H =
λ∈k
0
1
1 −c 0 0
0 1 0 0
−1
a
∗
u =
0 0 1 0 and ζ = u(ξ) we get uHu = { (λ , 1, 1, 1) | λ ∈ k }. It follows
0 0 0 1
that ζ is Gm -invariant, hence T (4; k)-invariant. Thus ξ corresponds to an ideal I such
that H 0 (I(b)) has a generating system of the form Mi (y − cx)ni , where Mi is a monomial
without y and ni ∈ N. There are only finitely many points ζi ∈ V (k) which are T (4; k)
S
invariant and it follows that ξ ∈ Ga · ζi .
S
Conclusion 11.5. If ξ ∈ V \ [ Ga · ζi ∪ V Ga ∪ V Gm ] is a closed point, then V =
Ga · Gm · ξ.
11.3. Summary
Let V ⊂ HQ be a 2-dimensional subvariety, invariant under G := Ga · T (4; k) but not
pointwise invariant under Ga . We distinguish between three cases:
1. V is not pointwise invariant under the Gm -operation σ(λ) : x 7→ x, y 7→ y, z 7→ z, t 7→
λt.
105
2. V is pointwise invariant under the Gm -operation σ as in the first case, but not pointwise
invariant under the Gm -operation τ (λ) : x 7→ x, y 7→ y, z 7→ λz, t 7→ t.
3. V is pointwise invariant under the Gm -operations σ and τ as in the 1th and 2nd case.
Then V is not pointwise invariant under the Gm -operation ω(λ) : x 7→ λx, y 7→ y, z 7→
z, t 7→ t.
Lemma 11.1. (a) In the 1st case (respectively in the 2nd case) there is ζ ∈ V (k) with
the following properties:
(i) The inertia group Tζ of ζ in T (4; k) has the form T (ρ) where ρ3 > 0 (respectively
ρ2 > 0 and ρ3 = 0).
(ii) V = Ga · Gm · ζ and V \ Ga · Gm · ζ is a union of points and curves invariant under
Ga · T (4; k).
(iii) There is u ∈ Ga , different from 1, such that V = T (4; k) · ξ, where ξ := u(ζ).
(b) In the 3rd case, if one chooses ζ ∈ V (k) such that ζ is neither Ga – nor Gm -invariant
and does not lie in the set ξ ∈ V (k) ∃u ∈ Ga and ∃µ ∈ V T (4;k) (k) such that ξ = u(µ) ,
then V = Ga · Gm · ζ.
106
CHAPTER 12
Surfaces in H invariant under B(4; k)
12.1. The operation of the unipotent group on H
12.1.1.
Let be p = (a : b : c) ∈ P2 (k)
1 α ∗
0 1 β
G(p) :=
0
0 1
0 0 0
and
∗
∗
aα + bβ + cγ = 0 .
γ
1
This is a 5-dimensional subgroup of ∆ := U(4 : k) and each 5-dimensional subgroup of ∆
has this form, where p ∈ P2 (k) is uniquely determined (Appendix D, Lemma 1).
Especially one has the groups
G
: 0
i = G(pi ), where
p1 = (0
1
1
∗
∗
∗
0
0 1 ∗ ∗
, G2 =
0), p3 = (1 : 0 : 0) and G1 =
0
0
0
1
0
0
0 0 0 1
1
0
∗
∗
0
1
∗
∗
.
0 0 1 ∗
0 0 0 1
1 α 0 0
0 1 0 0
−1
Remark 12.1. If ψα =
0 0 1 0, then ψα G(p)ψα = G(p).
0 0 0 1
: 1), p2
=(0 : 1 :
∗ ∗ ∗
1 0 ∗
, G3 =
0 1 ∗
0 0 1
Remark 12.2. If τ = (λ0 , λ1 , λ2 , λ3 ) ∈ T (4; k), then τ G(p)τ −1 = G(τ p), where τ p :=
−1
−1
(aλ−1
0 λ1 : bλ1 λ2 : cλ2 λ3 ).
12.1.2. In ([T2], 3.2.1) and ([T3], 10.2) we had introduced a closed, reduced subscheme Z = Z(HQ ) of HQ such that Z(k) = {x ∈ HQ (k)| dim ∆ · x ≤ 1}. From the
∼
theorem of Hirschowitz it follows that A1 (Z) −→ A1 (HQ ) ([T2], Lemma 24, p. 121). In
the following we consider a surface V ⊂ Z (i.e. a closed 2-dimensional subvariety) which
is B(4; k)-invariant, but not pointwise invariant under ∆ = U(4; k).
12.1.3. Auxiliary Lemma 1. Let be ξ ∈ Z(k) but ξ not invariant under ∆ ,hence
∆ξ = G(p) . If τ ∈ Tξ , then τ p = p .
Proof. If τ ∈ Tξ , then τ G(p)τ −1 τ ξ = τ ξ and thus G(τ p)ξ = ξ. If τ p 6= p , then
G(τ p) 6= G(p) and ξ would be fixed by the subgroup of ∆ , which is generated by G(p)
and G(τ p) , i.e. fixed by ∆ .
107
12.1.4. Auxiliary Lemma 2. Let ξ be as in Auxiliary Lemma 1 . If p = (a : b : 0)
and a, b 6= 0 , then Tξ ⊂ T (1, −2, 1, 0) .
Proof. This follows from Remark 12.2 and Auxiliary Lemma 1 .
12.2. The case p = (a : b : c) where a, b, c 6= 0.
Let be V ⊂ Z a B-invariant surface and ξ ∈ V (k) a point such that ∆ξ = G(a : b : c),
−2 3
2
∗
where a, b, c 6= 0. Then Tξ ⊂ Λ := {(λ0 , λ1 , λ−1
0 λ1 , λ0 λ1 )|λ0 , λ1 ∈ k }. As dim T (4; k)·ξ ≤
2, one has dim Tξ ≥ 2 and thus Tξ = Λ.
λ
α
0
0
0
0
λ
0
0
1
∗
operate on V .
α ∈ k, λ0 , λ1 ∈ k
We let G := Ga · T01 =
0 0 1 0
0 0 0 1
Remark 12.3. T (4; k)ξ ∩ G = (1).
−2 3
−1 2
−2 3
2
Proof. If (λ0 , λ1 , λ−1
0 λ1 , λ0 λ1 ) ∈ G, then λ0 λ1 = λ0 λ1 = 1 which implies λ0 /λ1 =
1, and then λ0 = λ1 = 1 follows.
Remark 12.4. V = T01 · ξ.
Proof. Because of Remark (12.3) dim T01 · ξ < 2 is not possible.
From Remark (12.4) it follows that Gξ < G is 1-dimensional. If H := G0ξ would be
isomorphic to Ga
, then
be invariant
under
Ga , hence invariant under ∆, contradic mξ would
λ
(λn − λm )c
tion. Thus H =
λ ∈ k ∗ , m, n ∈ Z not both equal to zero, c ∈ k
0
λn
m
1 −c
λ
0
−1
∗
(see 8.2.2 Auxiliary Lemma 2). If u :=
, then uHu =
λ ∈ k =:
0 1
0 λn
T01 (m, n). Putting ζ := u(ξ) one gets Tζ ⊃ T01 (m, n). Because of Remark (12.1) from
∆ξ = G(p) it follows that ∆ζ = G(p), too. Hence Tζ = Λ ⊃ T01 (m, n), which is not
possible. We have proved
Lemma 12.1. If V ⊂ Z is a B(4; k)-invariant surface, which is not pointwise∆ invariant , then there is no point ξ ∈ V (k), whose inertia group ∆ξ in ∆ has the form
G(a : b : c), where a, b, c 6= 0.
12.3. 1-cycles of proper type 3.
12.3.1. Recalling the restriction morphism h. The ideals I ⊂ OP3 with Hilbert
polynomial Q such that t is not a zero divisor of OP3 /I form an open non empty subset
108
Ut ⊂ HQ , and I 7→ I ′ := I + tOP3 (−1)/tO
P3 (−1)
1 0
0 1
phism h : Ut → H d = Hilbd (P2 ). If Γ :=
0 0
0 0
in Ut .
defines
the so called restriction mor0 ∗
0 ∗
< ∆, then HΓ is contained
Q
1 ∗
0 1
12.3.2. 1-cycles of proper type 3. We recall, respectively introduce, the notations:
An ideal J ⊂ OP3k with Hilbert polynomial Q corresponds to a point ξ ∈ HQ (k). J has
the type 3, if J is invariant under T (4; k) and G3 = G(1 : 0 : 0), but not invariant under
∆. The curve C = Ga · ξ = {ψα (ξ)|α ∈ k}− in HQ is called the 1-cycle of type 3 defined
by ξ. We say C is a 1-cycle of proper type 3, if I := J ′ = h(J ) has y-standard form (cf.
2.4.3 Definition 2). If ϕ is the Hilbert function of I ↔ ξ ′ ∈ H d (k), then g ∗(ϕ) > g(d)
m
by definition, and one has I = yK(−1) + x
O
K ⊂ OP2 has the colength
P2 (−m),where
1 0 ∗
′
c and is invariant under T (3; k) and G3 := 0 1 ∗ < U(3; k). Moreover one has
0 0 1
d = m + c and m ≥ c + 2.
T −a+2
T −b+1
If Q(T ) = T −1+3
+
+
is the Hilbert polynomial of J , then the
3
2
1
3
closed subscheme V+ (J ) ⊂ P defined by J has the “complementary” Hilbert polynomial
P (n) = dn − g + 1, where d = a − 1, g = g(J ) = (a2 − 3a + 4)/2 − b (cf. [T1], p. 92).
Lemma 12.2. Let J be of proper type 3 and put ν := min{n|H 0 (J (n)) 6= (0)}. If
xν ∈ H 0 (J (ν)), then g(J ) < 0.
Proof. We start with an ideal J fulfilling these conditions. There are subspaces
Ui ⊂ Si , invariant under T (3; k)·G′3 such that S1 Ui ⊂ Ui+1 , i = 0, 1, 2, . . . and H 0 (J (n)) =
n
L
tn−i Ui , for all n ∈ N. Besides this, Un = H 0 (I(n)), at least if n ≥ b, where I = J ′
i=0
is the restriction ideal of J (cf. [G78], Lemma 2.9, p. 65). We replace Un by H 0 (I(n))
for all ν ≤ n ≤ b − 1 and we get an ideal J ∗ ⊃ J such that H 0(J ∗ (n)) = (0), if
n
L
n < ν, and H 0 (J ∗ (n)) =
tn−i H 0 (I(i)), if n ≥ ν. (N.B.: xν ∈ H 0 (J (ν)) ⇒ xν ∈
0
−→
i=ν
0
Im(H (J (ν)) res H (I(ν)))). Then (J ∗ )′ is equal to I, J ∗ is of proper type 3, the
n−a+2
n−b∗ +1
Hilbert polynomial of J ∗ is equal to Q∗ (n) = n−1+3
+
+
, the length of
3
2
1
∗
∗
∗
∗
2
J /J is equal to Q (n) − Q(n) = b − b ≥ 0, and because of g(J ) = (a − 3a + 4)/2 − b∗
one has g(J ∗ ) ≥ g(J ). Thus it suffices to show g(J ∗ ) < 0 We thus can assume without
n
L
restriction J = J ∗ , i.e. H 0 (J (n)) =
tn−i H ( I(i)) for all n ≥ ν, and xν ∈ H 0(I(ν)).
i=ν
Using the terminology of [T1]–[T4], one can say the pyramid E(J ) is complete up to the
level ν over the platform H 0 (I(n)), where n ≥ b, for instance (cf. Fig. 12.1). Further
note that
(12.1)
H 0 (I(n)) = yH 0(K(n − 1)) ⊕ xm k[x, z]n−m
for all n ∈ N (cf. Lemma 2.6). We associate to I the ideal Ĩ represented in Figure
12.2; that means Ĩ arises from I by shifting yH 0(K(n − 1)) into yH 0(K̃(n − 1)), where
109
the Hilbert functions of K and K̃ agree and hence I and Ĩ have the same Hilbert function
n
L
ϕ. Then J˜ is defined by H 0 (J˜(n)) =
tn−i H 0 (Ĩ(i)) and H 0 (J˜(n)) = (0), if n < ν.
i=ν
Then Ĩ and J˜ fulfil the same assumptions as I and J do, and g(J ) = g(J˜). Thus we
can assume without restriction that I has the shape as represented by Fig. 12.2. Then
•
•
one makes the graded deformations • 7→ 1 (or • → 2 , etc.) in E(J ). Each of the
orbits which are to be exchanged, have the same length. One gets an ideal J˜ with the
same Hilbert polynomial, which is again of proper type 3. If ϕ̃ is the Hilbert function of
Ĩ := (J˜)′ , then ϕ̃ > ϕ, hence g ∗(ϕ̃) > g ∗ (ϕ) > g(d) (cf. Remark 2.1). As J and J˜ have
the same Hilbert polynomial, one has g(J˜) = g(J ). The colength of Ĩ in OP2 is the same
as the colenght d of I in OP2 , hence the coefficient a, remains unchanged. Now one can
again pass to (J˜)∗ and has again xν ∈ H 0((J˜)∗ (ν)), ν = min{n|H 0 ((J˜)∗ (n)) 6= (0)}, (J˜)∗
of proper type 3. Thus it suffices to show g((J˜)∗ ) < 0. Continuing in this way one sees
that one can assume without restriction: J is of proper type 3, I = J ′ has the shape of
n
L
Figure 12.4, H 0 (J (n)) =
tn−i H 0 (I(i)), for all n ≥ ν, H 0 (J (n)) = (0), if n < ν and
i=ν
xν ∈ H 0 (J (ν)), i.e., xν ∈ H 0 (I(ν)). From (12.1) it follows that ν ≥ m. As m ≥ c + 2, we
have h0 (K(n)) = n+2
− c, n ≥ m − 1, hence h0 (I(n)) = h0 (K(n − 1)) + (n − m + 1) =
2
n−1+2
n−1+2
n−1+1
n+2
−
c
+
(n
−
m
+
1)
=
+
+
1
−
(c
+
m)
=
− d, n ≥ m. But then
2
2
1
2
n
n
X
X
i+2
−d
h0 (J (n)) =
h0 (I(i)) =
2
i=ν
i=ν
X
ν−1
n
X
i+2
i+2
− (n − ν + 1)d
−
=
2
2
i=0
i=0
ν+2
n+3
− (n − ν + 1)d.
−
=
3
3
From this we get P (n) = (n − ν + 1)d + ν+2
, thus:
3
(12.2)
g(J ) = (ν − 1)d − ν+2
+1.
3
We regard g(J ) as a function of ν ≥ m, and we
1
1
g(x) := (x − 1)d − x3 − x2 −
6
2
have to determine the maximum of
1
x + 1, x ≥ m.
3
q
1 2
1
′
We have g (x) = − 2 x − x + (d − 3 ) = 0 ⇔ x = −1 ± 2d + 31 , and we show
q
m ≥ −1 + 2d + 31 . This last inequality is equivalent to (m + 1)2 > 2d + 31 ⇔ m2 + 23 ≥ 2c
(because of d = c + m). As m ≥ c + 2, this is true.
It follows that g ′(x) ≤ 0, if x ≥ m, hence g(x) is monotone decreasing if x ≥ m. Now
1
1
1
g(m) = (m − 1)d − m3 − m2 − m + 1
6
2
3
1 3 1 2 1
= (m − 1)(c + m) − m − m − m + 1
6
2
3
1 3 1 2 1
≤ (m − 1)(2m − 2) − m − m − m + 1.
6
2
3
110
The right side of this inequality is smaller than 0 if and only if 18 < m3 − 9m2 + 26m,
which is true as m ≥ c + 2 ≥ 2.
12.4. B(4; k)-invariant surfaces containing a 1-cycle of proper type 3.
Let be V ⊂ Z = Z(HQ ) a B(4; k)-invariant surface containing a 1-cycle D of proper
type 3. Then V is not pointwise invariant under the Ga -operation ψα : x 7→ x, y 7→
αx+y, z 7→ z, t 7→ t. According to Lemma 11.1 one can write V = Ga · Gm · ζ. The inertia
group ∆ζ has the form G(p) and by Lemma 12.1 it follows that p = (a : b : c), a, b, c 6= 0,
is not possible.
12.4.1. The case p = (a : 0 : c), a, c 6= 0. Then V is not pointwise invariant under
the Gm -operation σ (notations as in Lemma 11.1), as the following argumentation will
show: Let J ↔ ζ be the corresponding ideal, let P be an associated prime ideal, which is
G(p)-invariant. If t ∈ P, then from (Appendix D, Lemma 2) it follows that P = (x, y, z, t),
contradiction. Thus t is a non zero divisor of OP3 /J . If J would be invariant under σ,
then J would be generated by elements of the form f · tn , where f ∈ Sm , m, n ∈ N. It
follows that f ∈ J and thus J is invariant under Γ (cf. 12.3.1). But by assumption
∆ζ = G(a : 0 : c) 6⊃ Γ.
The inertia group Tζ ⊂ T (4; k) has the form T (ρ), where ρ3 > 0 ( Lemma 11.1 a). By
Appendix E H 0 (J (b)) has a standard basis fi = Mi pi (X ρ ).
Let be W := Ga · Gm · ζ The morphism his defined on V ∩ Ut ⊃ W . As W = V , it
follows that h(W ) = {ψα (ζ ′)|α ∈ k} is dense in h(V ∩ Ut ), where ζ ′ = h(ζ) ↔ J ′ . As
ζ ∈ Ut , it follows that ζ0 = lim σ(λ)ζ ∈ Ut too (see [G88], Lemma 4, p. 542, and [G89],
λ→0
1.6, p.14). As ρ3 > 0 it follows that ζ0′ = ζ ′. Now write D = {ψα (η)|α ∈ k}− , where
η ∈ V (k) is invariant under T (4; k) and G3 , hence C ⊂ V ∩Ut and h(η) ∈ {ψα (ζ ′)|α ∈ k}− .
′
As η is not Ga -invariant , h(η) is not Ga -invariant, hence
h(η)
∈ k}. Then
∈ {ψα (ζ )|α
1 0 ∗ ∗
0 1 ∗ ∗
′
′
< G(p) is
(Appendix D, Lemma 3) shows h(η) = ζ = ζ0 . H :=
0 0 1 0
0 0 0 1
normalized by σ, hence ζ0 is H-invariant. As ζ0 ∈ Ut is Gm -invariant , it follows that ζ0
is Γ-invariant. But then ζ0 is invariant under G3 , and as ζ0′ = η ′ corresponds to an ideal
in y-standard form, J0 ↔ ζ0 is of proper type 3.
As G(p) is unipotent there is an eigenvector f 6= 0 in H 0 (J (ν)), ν := min{n|H 0(J (n)) 6=
(0)}. From x∂f /∂t ∈ hf i it follows that ∂f /∂t = 0. From y∂f /∂z ∈ hf i it follows
∂f /∂z = 0 and from cx∂f /∂y − az∂f /∂t ∈ hf i we deduce f = xν (cf. Appendix D,
Lemma 2). But then xν ∈ J0 , too. Now h0 (J0 (n)) = h0 (J (n)), n ∈ Z (cf. [G88], Lemma
4, p.542 and [G89], 1.6, p.14). But then from Lemma 12.2 it follows that g < 0.
Conclusion 12.1. In the case p = (a : 0 : c), a, c 6= 0, V does not contain a 1-cycle
of proper type 3, if g > g(d) is supposed.
111
12.4.2. The case p = (a : b : 0), a, b 6= 0. As Tζ ⊂ T (1, −2, 1, 0) (cf. Auxiliary
Lemma 2 ), it follows from Lemma 11.1 that V is pointwise invariant under σ(λ) : x 7→
x, y 7→ y, z 7→ z, t 7→ λt. As V is pointwise invariant under Γ, one has V ⊂ GΦ , where Φ
is the Hilbert function of J ↔ ζ and GΦ is the corresponding “graded Hilbert scheme”.
This had been defined in ([G4], Abschnitt 2) as follows: Gm operates on HQ by σ. Then
it is shown in (loc. cit.) that G := (HQ )Gm ∩ Ut is a closed subscheme of HQ , and G is a
disjunct union of closed subschemes GΦ , where GΦ parametrizes the corresponding ideals
with Hilbert function Φ.
Now suppose D = {ψα (η)|α ∈ k}− ⊂ V is a 1-cycle of proper type 3, where η
corresponds to an ideal L of proper type 3. Then L and J ↔ ζ have the same Hilbert
function.
Let ν and f ∈ H 0 (J (ν)) be defined as in (12.4.1). From x∂f /∂z ∈ hf i and y∂f /∂t ∈
hf i it follows that ∂f /∂z = ∂f /∂t = 0. But then from bx∂f /∂y − ay∂f /∂z ∈ hf i it
follows that ∂f /∂y = 0, hence f = xν . (cf. Appendix D, Lemma 2). If I corresponds to
any point ξ ∈ Ga · Gm · ζ, then xν ∈ H 0 (I(ν)), and the same argumentation as in (8.4.2)
shows this is true for all points in V . But from xν ∈ H 0 (L(ν)) it follows that g < 0.
Conclusion 12.2. In the case p = (a : b : 0), a, b 6= 0, V does not contain a 1-cycle
of proper type 3, if g > g(d) is supposed.
12.4.3. If V contains a 1-cycle of proper type 3, then V cannot be pointwise Ga invariant, so the case p = (0 : b : c) cannot occur, at all.
Lemma 12.3. Suppose g > g(d). If V ⊂ Z(HQ ) is a B(4; k)-invariant surface containing a 1-cycle of proper type 3, then V is pointwise invariant under G3 = G(1 : 0 : 0).
Fig.: 12.1
c monomials are missing
xm z n−m
112
Fig.: 12.2
xm z n−m
Fig.: 12.3
2
1
xm z n−m
113
Fig.: 12.4
c monomials are missing
xm z n−m
114
CHAPTER 13
Relations in B(4; k)-invariant surfaces
We suppose g > g(d) and let V ⊂ X = Z(HQ ) be a B(4; k)-invariant surface, which
contains a 1-cycle of proper type 3. From it follows that V is pointwise G3 -invariant. Then
from ([T1], Proposition 0, p.3) we conclude that any B-invariant 1-prime cycle D ⊂ V is
either pointwise ∆-invariant or a 1-cycle of type 3. The aim is to describe the relations in
Z1B (X) defined by the standard construction of Section 9.2 carried out with regard to V .
13.1. First case : V is not pointwise invariant under the Gm -operation σ
We use the notation of (11.1). According to Lemma 11.1 V = Gm · Ga · ζ, where
Tζ = T (ρ) and ρ3 > 0, for otherwise ζ would be Gm -invariant and hence V would be
pointwise Gm -invariant. Let J ↔ ζ and C := {ψα (ζ)|α ∈ k}− . If one chooses a standard
basis of H 0 (J (b)) consisting of T (ρ)-semi-invariants, then one sees that h(V ) = C ′ :=
{ψα (ζ ′)|α ∈ k}− , where ζ ′ = h(ζ). As V contains a 1-cycle of proper type 3, C ′ is a
y-standard cycle, generated by ζ ′ ↔ J ′ .
Now if D ⊂ V is a 1-cycle of proper type 3, then h(D) = C ′ . If D is of type 3, but
not of proper type 3, then h(D) is not a y-standard cycle. Write D = Ga · η, η ∈ V (k)
invariant under T (4; k). It follows that η ′ = h(η) is one of the two T (3; k)-fixed points of
C ′ (cf. Appendix D, Lemma 3). If η ′ = ζ ′, then D would be of proper type 3. Hence η ′
is the unique point of C ′ , invariant under B(3; k) and h(D) equals the point η ′ .
If D is pointwise ∆-invariant, then h(D) is pointwise invariant under U(3; k) and hence
equals the point η ′ , too .
Let C0/∞ = lim σ(λ)C be the B-invariant limit curves of C coming from the stanλ→0/∞
dard construction. We write in Z1B (V ):
X
C0 =
mi Ai + Z0
and C∞ =
X
nj Bj + Z∞
where Ai and Bj are the 1-cycles of proper type 3, occurring in C0 and C∞ , respectively,
and mi , nj ∈ N − (0). All the other prime cycles occurring in C0 and C∞ are summed up
in Z0 and Z∞ , respectively.
Let Mn be a tautological line bundle on HQ . Then (Mn · Ai ) = δn + ai , (Ln · Bj ) =
δn + bj . Here δ ∈ N − (0) is independent of i and j, as h(Ai ) = h(Bj ) = C ′ for all i, j,
whereas the constants ai and bj still depend on Ai and Bj , respectively. From [C0 ] = [C∞ ]
P
P
it follows that
mi (δn + ai ) =
nj (δn + bj ) + c for all n ≫ 0, where c depends only
P
P
on Z0 and Z∞ (see Appendix D, Lemma 4) . It follows that
mi = nj . Therefore we
115
can write
(13.1)
C0 − C∞ =
X
k
(Ek − Fk ) +
X
Gk
k
where (E1 , E2 , . . . ) := (A1 , . . . , A1 , A2 , . . . , A2 , . . . ) and (F1 , F2 , . . . ) := (B1 , . . . , B1 , B2 ,
. . . , B2 , . . . ). Here A1 (respectively B1 ) are to occur m1 -times (respectively n1 -times) etc.
P
P
By the way, the arbitrary association Ek 7→ Fk is possible because of
mi =
nj . If
Ek = Fk , the summand Ek − Fk is left out. Gk is composed of either pointwise U(4; k)invariant curves or 1-cycles of type 3, which are not proper.
13.2. Second case: V is pointwise invariant under the Gm -operation σ, but
not pointwise invariant under the Gm -operation τ
We use the same notations as in (11.3).
1st subcase: h(V ) is not 2-dimensional. By assumption V contains a 1-cycle D of proper
type 3, hence h(V ) = h(D) =: D ′ is a y-standard cycle and all the other 1-cycles in V ,
which are of proper type 3, are mapped by h onto D ′ . Then one carries out the standard
construction by means of the operation τ and one gets formally the same relations as
(13.1).
′
2nd subcase: h(V
H d is a B(3; k)-invariant surface, which is pointwise invariant
) = V ⊂
1 0 ∗
under G′3 = 0 1 ∗ < U(3; k). As V ′ contains a y-standard cycle, V ′ is not
0 0 1
pointwise Ga -invariant.
At first, the standard construction is carried out in V = Gm · Ga · ζ (cf. Lemma
11.1): C := {ψα (ζ)|α ∈ k}− is a closed curve in V with Hilbert polynomial p and
λ 7→ τ (λ)C defines a morphism Gm → Hilbp (V )Ga , whose extension to P1 defines a
∗
flat family C ֒→ V × P1 over P1 such that Cλ := p−1
2 (λ) = τ (λ)C × {λ}, if λ ∈ k .
C0/∞ =: p1 (C0/∞ ) are called the limit curves of C and [C0 ] − [C∞ ] is the “standard
relation” defined by V .
Put U := {τ (λ)ψα (ζ)|α ∈ k, λ ∈ k ∗ }.
Put U ′ := {τ (λ)ψα (ζ ′ )|α ∈ k, λ ∈ k ∗ }.
Then U = V and U ′ = V ′ . Carrying out the standard construction by means of C ′ :=
p2
{ψα (ζ ′)|α ∈ k}− one gets a flat family C ′ ֒→ V ′ × P1 −→ P1 . One has a morphism
h×id
C ֒→ V × P1 −→ V ′ × P1 , which is denoted by f .
Put U := {(τ (λ)ψα (ζ), λ)|α ∈ k, λ ∈ k ∗ } ⊂ C.
Put U′ := {(τ (λ)ψα (ζ ′), λ)|α ∈ k, λ ∈ k ∗ } ⊂ C ′ .
C and C ′ are reduced and irreducible (see [T1], proof of Lemma 1, p.6). Hence U = C and
U′ = C ′ . As f (U) = U′ and f is projective, f (C) = C ′ follows. As the diagram
C
f
p2 ց
−→
P1
116
ւ p2
C′
′
is commutative, f (C0 ) = C0′ and f (C∞ ) = C∞
follows. As the diagram
h×id
V × P1 −→ V ′ × P1
p1 y
V
h
−→
yp1
V′
′
is commutative, it follows that C′0/∞ := p1 (C0/∞
) = p1 f (C0/∞ ) = h(C0/∞ ). Let be
′
′
ζ0/∞ := lim τ (λ)ζ and ζ0/∞ := lim τ (λ)ζ .
λ→0/∞
λ→0/∞
′
′
Now from Lemma 9.2, it follows that C0/∞
:= {ψα (ζ0/∞
)|α ∈ k}− are the only ystandard cycles in C′0/∞ and they both occur with multiplicity 1. We want to show that
C0/∞ := {ψα (ζ0/∞ )|α ∈ k}− are the only 1-cycles of proper type 3 in C0/∞ , and they both
occur with multiplicity 1. In order to show this, we consider the extension of τ : λ 7→ τ (λ)ζ
to a morphism τ : A1 → V . Then ζ0 = τ (0). As there is a commutative diagram
Gm
τ
−→
τ′ ց
V
yh
V′
where τ ′ (λ) := τ (λ)ζ ′ and as the extensions are determined uniquely, it follows that
τ ′ = h ◦ τ , hence h(ζ0 ) = ζ0′ . By constructing the corresponding extensions to P1 , it
′
′
follows in the same way that h(ζ∞ ) = ζ∞
. Hence we get h(C0/∞ ) = C0/∞
. Therefore C0
and C∞ are 1-cycles of proper type 3, and from ζ0/∞ ∈ C0/∞ we conclude C0/∞ ⊂ C0/∞ .
Assume there is another 1-cycle D of proper type 3 contained in C0 , or assume that
C0 occurs with multiplicity ≥ 2 in C0 . Then there is an equation [C0 ] = [C0 ] + [D] + · · ·
in Z1B (V ), where D = C0 , if C0 occurs with multiplicity ≥ 2. From Lemma 9.2 it follows
that h(D) = C0′ . It follows that
h∗ ([C0 ]) = h∗ ([C0 ]) + h∗ ([D]) + · · ·
= deg(h|C0 )[h(C0 )] + deg(h|D)[h(D)] + · · ·
= h∗ ([C]) = deg(h|C)[C ′ ].
As C = {ψα (ζ)|α ∈ k}− and α 7→ ψα (ζ) is injective and the same is true for C ′ and
α 7→ ψα (ζ ′ ), h|C is an isomorphism. The same argumentation shows that h|C0 and h|C∞
are isomorphisms. Hence we get [C0′ ] + [C0′ ] + · · · = [C ′ ] = [C′0 ], which means that C0′
occurs with multiplicity ≥ 2 in C′0 , contradiction.
The same argumentation shows that C∞ is the only 1-cycle of proper type 3 in C∞ , and
it occurs with multiplicity 1. This gives a relation of the form
(13.2)
C0 − C∞ = C 0 − C ∞ − Z
where C0 and C∞ are 1-cycles of proper type 3 and Z is a 1-cycle whose components are
either pointwise U(4; k)-invariant curves or 1-cycles of type 3, which are not proper.
117
13.3. Third case:V is pointwise invariant under σ and τ
Then we write V = Gm · Ga · ζ as in the 3rd case of Lemma 11.1.
1st subcase : h(V ) is not 2-dimensional. The same argumentation as in the first subcase
of (13.2), using the Gm -operation ω as in the third case of Lemma 11.1 instead of the
Gm -operation τ , gives relations of the form (13.1).
2nd subcase: If h(V ) is 2-dimensional, the same argumentation as in the second subcase
of 13.2 , with ω instead of τ , gives relations of the form (13.2).
Proposition 13.1. Assume g > g(d) and let V ⊂ X := Z(HQ ) be a B(4; k)-invariant
surface containing a 1-cycle of proper type 3. Then each relation in Z1B (X) defined by V ,
which contains a 1-cycle of proper type 3 is a sum of relations of the form C1 − C2 + Z.
Here C1 and C2 are 1-cycles of proper type 3, and Z is a 1-cycle whose prime components
either are pointwise ∆-invariant or 1-cycles of type 3 , which are not proper .
Proof. This follows from the foregoing discussion.
118
CHAPTER 14
Necessary and sufficient conditions
We take up the notations introduced in Chapter 1. Let be d ≥ 5, g(d) = (d −
2) /4, H = Hd,g , A(H) = Im(A1 (HU (4;k) ) −→ A1 (H)). Obviously, the cycle C3 =
{(xa , αx + y, xa−1 z b−a+1 )|α ∈ k}− , where d = a − 1, g = (a2 − 3a + 4)/2 − b, is a 1cycle of proper type 3.
2
14.1. The necessary condition.
In the proof of Theorem 1.1 in Chapter 10 we replace H, U(3; k), B(3; k) and “ystandard cycle” by H, U(4; k), B(4; k) and “1-cycle of proper type 3”, respectively. Then
using Proposition 13.1 instead of Proposition 9.1 , the same reasoning as in the proof of
Theorem 1.1 gives the
Conclusion 14.1. If d ≥ 5 and g > g(d), then [C3 ] ∈
/ A(H).
14.2. The sufficient condition.
14.2.1. The case d ≥ 5 and d odd. In ([T2], 3.3.2, Folgerung, p.129) it had been
shown [C3 ] ∈ A(H), if a ≥ 6 and b ≥ a2 /4. We express this condition by means of the
formulas in (1.1):
1
1
1
b ≥ a2 /4 ⇔ g ≤ [(d + 1)2 − 3(d + 1) + 4] − (d + 1)2 = (d2 − 4d + 3).
2
4
4
1
2
As d is odd, this is equivalent to g ≤ g(d) = 4 (d − 2) .
Conclusion 14.2. If d ≥ 5, d is odd and g ≤ g(d), then [C3 ] ∈ A(Hd,g ).
14.2.2. The case d ≥ 6 and d even. We will show that the bound given in
([T2], 3.3.3 Folgerung, p.129) is already valid, if a ≥ 7.
1◦ We consider the Hilbert function ϕ of the monomial ideal (x2 , xy e−2 , y e), e := d/2 + 1 ≥
4, which is represented in Fig. 14.1. In Section (2.2.3) this Hilbert function had been
denoted by χ and it had been shown that g ∗ (ϕ) = 41 (d−2)2 =: g(d). The figures 14.1 - 14.6
show all possible monomial ideals with Hilbert function ϕ. Besides this, we consider the
ideal I := (xy, xe , y e, xe−1 + y e−1) which corresponds to a closed point ξ in the Iarrobino
variety Iϕ . This variety parametrizes all sequences (U0 , U1 , . . . ) of subspaces Ui ⊂ Ri with
dimension ϕ′ (i) = ϕ(i) − ϕ(i − 1), such that R1 Ui ⊂ Ui+1 , i ∈ N . Here R is the polyno mial ring in two variables .
2◦ We show that V = Ga · Gm · ξ is 2-dimensional, where Gm operates by τ (λ) : x 7→
f
λx, y 7→ y, z 7→ z. If this would not be the case, then Ga × Gm −→ Iϕ defined by
119
(α, λ) 7→ ψα τ (λ)ξ would have a fibre with infinitely many points. The argumentation in
(8.3) then showed that one of the points ξ0/∞ = lim τ (λ)ξ is invariant under Ga . As
λ→0/∞
ξ0 ↔ I0 = (xy, xe , y e−1) and ξ∞ ↔ I∞ = (xy, xe−1, y e ), this is wrong as e ≥ 4. Thus we
can carry out the standard construction with C = {ψα (ξ)|α ∈ k}− , if e ≥ 4. As we have
already noted in the proof of Lemma 9.1 , the curves τ (λ)C are Ga -invariant, hence the
construction of the family C takes place in (Hilbp (Iϕ ))Ga . Therefore the limit curves C0/∞
are invariant under B(3; k).
As usual, we put C0/∞ = {ψα (ξ0/∞ )|α ∈ k}− .
3◦ Let I ↔ (U0 , U1 , · · · ), where Ui = xyRi−2 , if 0 ≤ i ≤ e − 2, Ue−1 = xyRe−3 + hxe−1 +
y e−1i, Ui = Ri , i ≥ e. We get
ψα (Ue−1 ) = x(αx + y)Re−3 + hxe−1 + (αx + y)(αx + y)e−2i
⇒
= x(αx + y)Re−3 + hxe−1 + (αx + y)y e−2i
˙ α (Ue−1 ) = αxe−1 ∧ αxe−2 y ∧ · · · ∧ αx2 y e−3 ∧ αxy e−2
∧ψ
+ terms with smaller powers of α.
Hence α-grade (Ue−1 ) = e − 1. On the other hand, by formula (4.2) in 4.1 we get α-grade
(hxe−2 y, . . . , y e−1i) = e − 1, too. From this it follows that C0 = C0 .
4◦ Besides C∞ , the limit cycle C∞ contains other prime components D1 , D2 , . . . which
are B(3; k)-invariant curves in Iϕ .
As char(k) = 0 is supposed, if m ≤ n, G := Grassm (Rn ) has only one Ga -fixed point,
namely the subspace hxn , · · · , xn−m+1 y m−1 i. Then ([T1], Proposition 0, p.3) shows that
each B(2; k)-invariant curve in G has the form Ga · η, where η ∈ G(k) corresponds to a
monomial subspace of Rn . It follows that each B(3; k)-invariant curve in Iϕ also has the
form Ga · η, where η ∈ Iϕ (k) corresponds to a monomial ideal.
5◦ Figure 14.1 - Figure 14.6 show all possible monomial ideals with Hilbert function ϕ,
and using formula (4.2) we compute the degrees of the corresponding 1-cycles:
e−2
e−2
e−2
deg c2 = e−2
,
deg
c
=
+
(e
−
1),
deg
c
=
2
+
(e
−
1),
deg
c
=
2
+ (e −
3
4
5
2
2
2
2
2), deg c6 = 1.
We see that c2 (respectively c3 ) is equal to C∞ (respectively C0 ). If C0 (respectively
Di ) occurs in C∞ with the multiplicity n (respectively ni ), then from deg C∞ = deg C0 =
e−2
deg C0 it follows that e−2
+
(e
−
1)
=
n
+ n1 deg D1 + · · · where n > 0, ni ≥ 0.
2
2
e−2
1st case: e ≥ 6. Then 2 > e − 1, hence n = 1. From D1 ∈ {c4 , c5 , c6 } and e − 1 ≥
n1 deg D1 we conclude that D1 = c6 and
[C0 ] = [C∞ ] + (e − 1)[c6 ]
2nd case: e = 5. Then deg C0 = 7, deg C∞ = 3, deg c4 = 10, deg c5 = 9, and we get
[C0 ] = 2[C∞ ] + [c6 ] or [C0 ] = [C∞ ] + 4[c6 ].
3rd case: e = 4. Then deg C0 = 4, deg C∞ = 1, deg c4 = 5, deg c5 = 4. Therefore one gets
[C0 ] = n[C∞ ] + m[c6 ]
where n, m ∈ N, n ≥ 1, n + m = 4.
6◦ One has a closed immersion Iϕ → Hd,g(d) defined by I 7→ I ∗ , where I ∗ is the ideal
120
generated by I in OP3 . That means H 0 (P3 ; I ∗ (n)) =
case, one gets an equation
∗
[C0∗ ] = n[C∞
] + m[c∗6 ],
n
L
i=0
tn−i H 0 (P2 ; I(i)), n ∈ N. In any
m, n ∈ N, n ≥ 1.
∗
Now from ([T1], Lemma 5, p.45) it follows that one can deform C0∗ (respectively C∞
) by
a sequence of graded deformations of type 3 (cf. [T1], p.44) modulo A(H) in the cycle C3
(respectively in the zero cycle). The cycle c∗6 is equal to the cycle Dα , α = (d + 2)/2 = e,
which had been introduced in ([T4], p.20f). By ([T4], Lemma 5, p.25) one gets De ≡ D2
modulo A(H), and in ([T4], Abschnitt 3.3, p.25) it had been noted that [D2 ] = [D], where
D is the cycle introduced in (1.1). From ([T4], Satz 1, p.26) it follows [C3 ] ∈ A(Hd,g(d) ).
Applying the shifting morphism (cf. [T3], Folgerung, p.55 and Proposition, p.56) we get:
Conclusion 14.3. If d ≥ 6 is even and g ≤ g(d), then [C3 ] ∈ A(Hd,g ).
14.3. Summary.
Theorem 14.1. Let be d ≥ 5. Then [C3 ] ∈ A(Hd,g ) if and only if g ≤ g(d).
Proof. This follows from Conclusion 14.1- 14.3.
Fig.: 14.2
Fig.: 14.1
c2
0
1
2
3
4 ... ... e
0
121
1
2
3
4 ... ... e
Fig.: 14.3
Fig.: 14.4
c3
0
1
2
3
c4
4 ... ... e
0
1
2
3
4 ... ... e
Fig.: 14.6
Fig.: 14.5
c5
c6
0
1
2
3
4 ... ... e
0
122
1
2
3
4 ... ... e
CHAPTER 15
The case d ≥ 5
We suppose d ≥ 5, i.e. a ≥ 6, and g <
d−1
2
, i.e. g not maximal.
From ([T1], Satz 2, p.91) and ([T4], Proposition 2, p.26) it follows that A1 (Hd,g )/A(Hd,g )
is generated by the so-called combinatorial cycles C1 , C2 , C3 . Using the formulas in (1.1),
one shows that g ≤ γ(d) := d−2
is equivalent to b ≥ 2a − 4.
2
1st case: g > γ(d). This is equivalent to b < 2a − 4. By ([T4], Satz 1, p.26) one has
A(Hd,g ) = hEi. Assume there is a relation q0 [E] + q1 [C1 ] + q2 [C2 ] + q3 [C3 ] = 0, qi ∈ Q.
Computing the intersection numbers with the tautological line bundles Ln by means of
the formulas (1)-(5) in ([T2], p.134f) and the formula in ([T3], Hilfssatz 1, p.50), one
sees that q2 = 0. Put r := 2a − 4 − b and denote the r-times applied shifting morphism by f (see [T3], p.52 ). The images of E, C1 , C3 under f in Hd,γ(d) are denoted
by e, c1 , c3 . By ([T2], 3.2.2 Folgerung, p.124) and ([T3], Anhang 2, p.54f) it follows that
[f (E)] ≡ [e], hf (C1)i ≡ hc1 i and hf (C3 )i ≡ hc3 i modulo A(Hd,γ(d) ). By ([T3], Proposition
4, p. 22) [c1 ] ∈ A(Hd,γ(d) . As γ(d) > g(d) if d ≥ 5, from Theorem 14.1 it follows that
[c3 ] ∈
/ A(Hd,γ(d) ). Hence f (C3 ) is not a single point, hence deg(f |C3 ) 6= 0. Applying f∗
to the relation q0 [E] + q1 [C1 ] + q3 [C3 ] = 0 then gives q3 deg(f |C3 ) · [c3 ] ∈ A(Hd,γ(d) ). If
q3 6= 0 it would follow that [c3 ] ∈ A(Hd,γ(d) ), contradiction. But from q0 [E] + q1 [C1 ] = 0
it follows q0 = q1 = 0 (cf. [T2], 4.1.3).
2nd case: g(d) < g ≤ γ(d). Then b ≥ 2a − 4 and in (1.1) it was explained that in this case
A1 (Hd,g ) is generated by E, D, C2 and C3 . If q0 [E]+q1 [D]+q2 [C2 ]+q3 [C3 ] = 0, then q2 = 0
by ([T2], loc.cit.). If one assumes that q3 6= 0, it follows that [C3 ] ∈ hE, Di ⊂ A(Hd,g ),
contradicting Theorem 14.1. As in the first case we get q0 = q1 = 0.
3rd case: g ≤ g(d). Then A(Hd,g ) = hE, Di, [C1] ∈ A(Hd,g ) and [C3 ] ∈ A(Hd,g )
(see.[T3], Proposition 4, p.22 ; [T4], Satz 1, p.26 ; and finally Theorem 14.1 ). Thus
A1 (Hd,g ) = hE, D, C2 i.
All in all we get: Suppose that d ≥ 5 and g < d−1
, i.e. g is not maximal. Put
2
d−2
2
g(d) := (d − 2) /4 and γ(d) := 2 .
Theorem 15.1. (i) If g > γ(d), then A1 (Hd,g ) is freely generated by E, C1 , C2 , C3.
(ii) If g(d) < g ≤ γ(d), then A1 (Hd,g ) is freely generated by E, D, C2 , C3 .
(iii) If g ≤ g(d), then A1 (Hd,g ) is freely generated by E, D, C2.
N.B.: If g = d−1
, then A1 (Hd,g ) is freely generated by C1 := {(x, y d (αy + z))|α ∈ k}−
2
and C2 := {(αx + y, xd+1 )|α ∈ k}− ([T1], Satz 2, p.91).
123
CHAPTER 16
The cases d = 3 and d = 4
16.1. The case d = 3.
T −4+2
T −b+1
If g is not maximal, i.e., if g ≤ 0, then Q(T ) = T −1+3
+
+
, where
3
2
1
b ≥ 4. If b = 4, then in ([T2], p.137) it had been shown that [C3 ] ∈ A(H). Applying the
shifting morphism , (see [T3] ,p .54 ) then shows that this statement is true for all g < 0
(cf. [T3], Anhang 2, Folgerung, p.55 and Proposition, p.56). By ([T1], Satz, p.91, [T3],
Prop. 4, p.22 and Satz 1, p.26) it follows that A1 (H3,g ) is generated by [E], [D], [C2], if
g ≤ 0. The three cycles are linear independent, as already was noted in Chapter 15.
16.2. The case d = 4
If g is not maximal, i.e., if g < 3, then Q(T ) = T −1+3
+ T −5+2
+ T −b+1
and b ≥ 5.
3
2
1
4
If b = 5, then g = 2, and then A1 (H) = Q , as had been shown in ([T3], Anhang 1).
We now treat the case b = 6, i.e., g = 1. As the condition b ≥ 2a − 4 is fulfilled,
A(H) = hE, Di by ([T4], Satz 1, p.26). The quotient A1 (H)/A(H) is generated by C1 , C2
and further cycles C3 , C4, C5 of type 3 (cf. [T1], Satz 2c, p.92). By ([T3], Proposition 4,
p.22) [C1 ] ∈ A(H), and we are going to simplify the reductions of C3 , C4 , C5 described in
([T3], Abschnitt 8).
16.2.1. The cycle C3 . By definition a cycle of type 3 has the form C = Ga · ξ, where
ξ corresponds to a monomial ideal J with Hilbert polynomial Q, which is invariant under
the subgroup G3 < U(4; k) (cf. Chapter 1). Here Ga operates as usual by ψα : x 7→
x, y 7→ αx + y, z 7→ z, t 7→ t. Let I := J ′ ∈ H 4 (k) be the image of J under the restriction
morphism. If I is Ga -invariant, then I is B(3; k)-invariant, hence deg(Ln |C) = constant
and [C] ∈ A(H) (cf. [T3], Anhang 2, Hilfssatz 2, p.50). Therefore we can assume without
restriction that I is not Ga -invariant. As the colength of I in OP2 is equal to 4, the Hilbert
function of I is equal to one of the functions represented in Figure 2.7a and Figure 2.7b.
In (2.2.1) we had obtained g ∗ (ϕ1 ) = 1 and g ∗ (ϕ2 ) = 3. The Hilbert function ϕ1 leads to
two possible 1-cycles of proper type 3, namely to
F1 := {ψα (ξ1 )|α ∈ k}− ,
F2 := {ψα (ξ2 )|α ∈ k}− ,
ξ1 ↔ (xy, y 2, x3 ),
and
ξ2 ↔ (x2 , y 2).
The Hilbert function ϕ2 leads to different 1-cycles of proper type 3, which however by
means of admissible G3 -invariant deformations all can be deformed modulo A(H) in
C3 = {ψα (ξ3 )|α ∈ k}− , ξ3 ↔ (x5 , y, x4z 2 ).
125
With such admissible G3 -invariant deformations we can deform C3 in the cycle generated
by (x4 , xy, y 2, yz 2 ), and afterwards this cycle can be deformed by the graded deformation
(yz 2 , yz 3 , . . . ) 7→ (x3 , x3 z, . . . ) in the cycle F1 .
We deform F1 into F2 in the following way: Put K := (x3 , x2 y, xy 2, y 3). If L ⊂
R2 = k[x, y]2 is a 2-dimensional subspace, then hx, yiL ⊂ H 0 (K(3)) = R3 and z n L ∩
H 0 (K(n + 2)) = (0) for all n ≥ 0. It follows that the ideal (L, K) ⊂ OP3 has the Hilbert
polynomial Q, and L 7→ (L, K) defines a closed immersion P2 ≃ Grass2 (R2 ) → HQ . Let
hxy, y 2i ↔ η1 ∈ P1 , hx2 , y 2i ↔ η2 ∈ P2 , ℓi := {ψα (ηi )|α ∈ k}− , i = 1, 2. Because of
reasons of degree one has [ℓ1 ] = 2[ℓ2 ] in A1 (P2 ), hence [F1 ] = 2[F2 ] in A1 (HQ ). Now
F2 = {(x2 , xy 2, y 3 , (αx + y)2 |α ∈ k}− = {(x2 , xy 2 , y 3, λxy + µy 2 )|(λ : µ) ∈ P1 } is equal to
the cycle D3 which had been introduced in ([T4], Abschnitt 3.2, p.20). By Lemma 5 in
(loc.cit.) D3 ≡ D2 modulo A(H), where D2 := {(x2 , xy, y 4, λxz 2 + µy 3)|(λ : µ) ∈ P1 } is
the cycle , which had been denoted by D in Chapter 1. It follows that [C3 ] ∈ A(H), and
applying the shifting morphism gives
Conclusion 16.1. [C3 ] ∈ A(H4,g ) for all g ≤ 1.
Remark 16.1. If d = 3 and g > g(3) = 1/4, C3 does not occur at all. If d = 4 and
g > g(4) = 1, then A(H4,2 ) = hEi ([T4], Satz 1, p.26) and hence [C3 ] ∈
/ A(H4,2 ).
16.2.2.
[C4 ] ∈ A(H) is true for arbitrary d and g ([T4], Proposition 2, p.26).
16.2.3. The cycle C5 . In (loc.cit.) [C5 ] ∈ A(H) was shown, too, but the proof
required tedious computations, which can be avoided by the following argumentation.
One has only to treat the cases g = 0 and g = −2, that means, the cases b = 7 and b = 9
([T1], Satz 2c(iii), p.92).
We start with b = 7. Then C5 = Ga · ξ0 , where ξ0 ↔ J0 := (y 2, K), K := (x3 , x2 y, xy 2,
y 3, x2 z, y 2 z). Put J := (x2 + y 2, K) ↔ ξ. As J0 and J have the same Hilbert function, ξ0
and ξ both lie in the same graded Hilbert scheme Hϕ (cf. Appendix A). If Gm operates
by τ (λ) : x 7→ λx, y 7→ y, z 7→ z, t 7→ t, then one sees that indeed ξ0 = lim τ (λ)ξ and
λ→0
ξ∞ := lim τ (λ)ξ ↔ J∞ := (x2 , K). Put C := {ψα (ξ)|α ∈ k}− , C0/∞ := {ψα (ξ0/∞ )|α ∈
λ→∞
k}− . One sees that α-grade H 0 (J (n)) = α-grade H 0 (J0 (n)) = α-grade H 0 (J∞ (n)) + 2,
for all n ≥ 2, hence deg C = deg C0 = deg C∞ + 2, relative to an appropriate imbedding
of Hϕ into PN .
Put V := Gm · Ga · ξ. As ξ is invariant under G := G3 · T23 and as G3 and T23 =
{(1, 1, λ2, λ3 )|λ2 , λ3 ∈ k ∗ } are normalized by Gm · Ga , V is pointwise G-invariant. Let p
be the Hilbert polynomial of C ⊂ PN . Then the standard construction by means of V
takes place in X := Hilbp (V )Ga , as C is Ga -invariant. Thus the limit curves C0/∞ are
pointwise G-invariant and invariant under Ga and Gm , hence they are B(4; k)-invariant.
Now C0/∞ ⊂ C0/∞ and the above computation of the degrees shows that C0 = C0 ,
whereas C∞ except from C∞ has to contain a further 1-cycle Z of degree 2, i.e., C∞
has to contain an irreducible curve F of degree ≥ 1, which is B(4; k)-invariant. As V is
pointwise G-invariant, F is either pointwise U(4; k)-invariant or an 1-cycle of type 3. As
126
deg(Ln |F ) is constant it follows that [F ] ∈ A(H) (see [T3], Anhang 2, Hilfssatz 2, p.50).
It follows that Z ∈ A(H).
From [C5 ] = [C0 ] = [C∞ ] + Z and C∞ = C4 (cf. [T1], Satz 2, p. 91), because of
(16.2.2) it follows that [C5 ] ∈ A(H).
We now treat the case b = 9. Here C5 = Ga · η, where η ↔ (x3 , x2 y, y 2, x2 z 3 ). By
the G3 -admissible deformation y 2 7→ x2 z 2 C5 is deformed in the cycle C5′ := Ga · ξ0 ,
where ξ0 ↔ J0 := (x3 , x2 y, xy 2, y 3 , y 2z, x2 z 2 ). By ([T1], 1.4.4) [C5 ] = q[C5′ ] modulo
A(H). Hence we can argue with C5′ and J0 in the same way as in the case b = 7: Let
K := (x3 , x2 y, xy 2, y 3, x2 z 2 , y 2z 2 ) and J := (x2 z + y 2 z, K) ↔ ξ. If ξ0/∞ := lim τ (λ)ξ ,
λ→0/∞
then ξ0 is as above and ξ∞ ↔ J∞ = (x3 , x2 y, xy 2, y 3, x2 z, y 2 z 2 ).
Obviously, one has the same computation of degree as in the case b = 7 and the
same argumentation shows that [C5′ ] = [C∞ ] modulo A(H). C∞ is deformed by the G3 admissible deformation y 2 z 2 7→ x2 in the cycle Ga · ζ, where ζ ↔ (x2 , xy 2 , y 3, y 2z 3 ) As
this is the cycle C4 , by 16.2.2 we get [C5 ] ∈ A(H).
Conclusion 16.2. [C5 ] ∈ A(H4,g ) if g = 0 or g = −2.
16.2.4. Summary. The results of 16.1 and 16.2 give
Theorem 16.1. (i) If g ≤ 0, then A1 (H3,g ) is freely generated by [E], [D], [C2 ].
(ii) A1 (H4,2 ) ≃ Q4 and if g ≤ 1, then A1 (H4,g ) is freely generated by [E], [D], [C2 ].
127
CHAPTER 17
Correction of the results of [T4] and summary
In [T4] the computation of the degree on page 28 is wrong, and hence Conclusions 1-3
and Proposition 3 on the pages 29-32 are wrong. As well ([T4], Lemma 7, p.33) is wrong.
The statement of ([T4], Proposition 4, p.33) is right, but with regard to Theorem 15.1(i) it
is irrelevant. The statement in ([T4], Satz 2, p.35) is wrong and has to be replaced by the
statements of Theorem 15.1 and Theorem 16.1 , respectively. The results of the sections 8
and 9 in [T4] remain valid, if one replaces the old bound by the bound g(d) = (d − 2)2 /4.
Furthermore, ([T4], Satz 3, p.35) has to be replaced by:
Theorem 17.1. Let be d ≥ 3 and g not maximal, let C be the universal curve with
degree d and genus g over Hd,g .
(i) If g > γ(d) := d−2
, then A1 (C) is freely generated by E ∗ , C1∗ , C2∗ , C3∗ and L∗ .
2
(ii) If g(d) < g ≤ γ(d), then A1 (C) is freely generated by E ∗ , D ∗ , C2∗ , C3∗ and L∗ .
(iii) If g ≤ g(d), then A1 (C) is freely generated by E ∗ , D ∗ , C2∗ and L∗ .
The statements of ([T4], Satz 4 and Satz 5, p. 36) are correct, if the bound mention
there is replaced by g(d) = (d − 2)2 /4. The reason is that the arguments used in (loc.cit.)
formally do not depend on the bound.
All in all, one gets the results which had been stated in the introduction.
Concluding Remark: Having arrived at this point, it is not so difficult any more to
explicitly determine the cone of curves and the ample cone of Hd,g (and of C).
129
APPENDIX A
Notations
The ground field is C; all schemes are of finite type over C; k denotes an extension
field of C. P = k[x, y, z, t], S = k[x, y, z], R = k[x, y] are the graded polynomial rings.
T = T (4; k) group of diagonal matrices
∆ = U(4; k) unitriangular group
B = B(4; k) Borel group
T (ρ)
subgroup
of T
(3;k) or of T (4; k) (cf. 2.4.1 and [T1], p.2).
1
0
0
∗
0
1
0
∗
< U(4; k)
Γ=
0 0 1 ∗
0 0 0 1
G1 , G2 , G3 subgroups of U(4; k) (cf. 1.1)
H = Hd,g Hilbert scheme of curves in P3 with degree d ≥ 1 and genus g, i.e. H =
Hilbp (P3k ), where P (T ) = dT − g + 1.
Q(T ) = T +3
− P (T ) complementary Hilbert polynomial
3
HQ = Hilbert scheme of ideals I ⊂ OP3 with Hilbert polynomial Q(T ), i.e. H = Hd,g =
HQ .
HQ 6= ∅ if and only if Q(T ) = T −1+3
+ T −a+2
or Q(T ) = T −1+3
+ T −a+2
+ T −b+1
,
3
2
3
2
1
where a and b are natural numbers 1 ≤ a ≤ b. The first case is equivalent with d = a
and g = (d − 1)(d − 2/2, i.e., equivalent with the case of plane curves. We consider
only the case g < (d − 1)(d − 2)/2. In this case we have the relations d = a − 1 and
g = (a2 − 3a + 4)/2 − b.
It was not possible to reserve the letter d for denoting the degree of a curve. If
necessary d denotes a number large enough, e.g. d ≥ b = bound of regularity of all ideals
in OP3 with Hilbert polynomial Q (cf. [G1], Lemma 2.9, p.65).
G = Grassm (Pd ) Grassmann scheme of m-dimensional subspaces of Pd .
Let ϕ : N → N be a function with the following properties: There is an ideal I ⊂ OP2
of finite colength with Hilbert function h(n) = h0 (I(n)), such that 0 ≤ ϕ(n) ≤ h(n) for
all n ∈ N and ϕ(n) = h(n) for n ≥ d, where n is large enough, e.g. n ≥ d := colength(I).
On the category of k-schemes a functor is defined by
Hϕ (Spec A) = {(U0 , · · · , Ud )|Un ⊂ Sn ⊗ A subbundle of rank ϕ(n) such that S1 Un−1 ⊂
Un , 1 ≤ n ≤ d}
Hϕ is a closed subscheme of a suitable product of Grassmann schemes; it is called
graded Hilbert scheme.
131
To each ideal J ⊂ OP3k with Hilbert polynomial Q corresponds a point ξ ∈ H(k),
which we denote by ξ ↔ J .
h(J ) denotes the Hilbert function of J , that means h(J )(n) = dimk H 0 (J (n)), n ∈ N.
If ϕ is the Hilbert function of an ideal in OP2k of colength d, then
Hϕ := {I ⊂ OP2k |h0 (I(n)) = ϕ(n), n ∈ N}
is a locally closed subset of Hilbd (P2 ), which we regard to have the induced reduced scheme
structure.
If G is a subgroup of GL(4; k), then HG denotes the fixed-point scheme, which is to
have the induced reduced scheme structure. The same convention is to be valid for all
fixed-point subschemes of H d = Hilbd (P2 ) .
If C ֒→ H is a curve, then by means of the Grothendieck-Plücker embedding H −→ PN
we can regard C as a curve in a projective space, whose Hilbert polynomial has the form
deg(C) · T + c. Here deg(C) is defined as follows : If I is the universal sheaf of ideals
on X = H × P3k , then F := OX /I is the structure sheaf of the universal curve C over
H, and the direct image π∗ (F (n)) is locally free on H of rank P (n) for all n ≥ b. The
˙ ∗ (F (n)) are called the tautological line bundles on H,which are
line bundles Mn := ∧π
very ample and thus define the Grothendieck - Plücker embeddings in suitable projective
spaces. Here ∧˙ is to denote the highest exterior power. Then deg(C) is the intersection
number deg(Mn |C) := (Mn · C). (If C is a so called tautological or basis cycle one can
compute this intersection number directly, see [T2], Section 4.1.)
After these more or less conventional notations we introduce some notations concerning
monomial ideals. If J ⊂ OP3 is T -invariant, then H 0 (P3k ; J (d)) ⊂ Pd is generated by
monomials . To each monomial xd−(a+b+c) y a z b tc in H 0 (J (d)) we associate the cube [a, a +
1] × [b, b + 1] × [c, c + 1] in an y − z − t - coordinate system, and the union of these
cubes gives a so called pyramid, which is denoted by E(J (d)). Usually we assume that
d
L
J is invariant under ∆ or Γ. Then we can write H 0 (J (d)) =
td−n Un , where Un ⊂ Sn
n=0
are subspaces such that S1 · Un ⊂ Un+1 , 0 ≤ n ≤ d − 1, which we call the layers of the
pyramid. (In [T1]–[T4] we made extensive use of this concept, but here it occurs only
once in 12.3.3.)
A1 (−) denotes the group of rational equivalence classes with coefficients in Q.
132
Fig.: A.1
1
0
0
1
2 ... ... ... ... ... ...
133
APPENDIX B
Hilbert functions without Uniform Position Property
Lemma B.1. Let be k be an algebraically closed field, I ⊂ OP2k an ideal of finite
colength with Hilbert function ϕ and difference function ϕ′ (n) = ϕ(n) − ϕ(n − 1). Let m
be a natural number such that ϕ′ (m + 1) = ϕ′ (m) + 1. The ideal J ⊂ OP2k generated by
H 0 (I(m)) has the following properties:
(i) J is m-regular;
(ii) H 0 (J (n)) = H 0 (I(n)) for all n ≦ m + 1;
(iii) If δ := m + 1 − ϕ′ (m) > 0, then there is a form f ∈ Sδ and an ideal L ∈ OP2 of finite
colength such that J = f · L(−δ).
Proof. Let be In := H 0 (I(n)), I :=
∞
L
In . The ideal I is called Borel-normed, if
n=0
in(I) is invariant under B(3; k), where in(I) is the ideal generated by the initial monomials
of all forms in I. According to a theorem of Galligo, there is a g ∈ GL(3; k) such
that g(I) is Borel-normed. (In [G4], Anhang IV, in the special case of three variables,
there is an ad-hoc-proof.) Therefore we can assume without restriction that I is Borelnormed . Then Fig.A.1 shows not only the graph of ϕ′ , but also the monomials in
H 0 (I0 (n)), where I0 := [in(I)]∼ (cf. [G4], Anhang V, Hilfssatz 1, p.116). One sees that
S1 in(Im ) = in(Im+1 ) and this implies the statements (i) and (ii) (cf. loc.cit., Lemma, p.
116 or [Gre], Proposition 2.28, p. 41).
Let ψ be the Hilbert function of J . Then ψ is also the Hilbert function of J0 = in(J )
(cf. [G4], Hilfssatz 1, p.114), and the further development of the graph of ψ ′ is marked
by · · · in Fig. A.1. The line l : y = x − δ + 1 is marked by - - - - . If c is the number
of monomials between the graphs of ψ ′ and l, then ψ(n) = n−δ+2
− c, n ≥ m. Then the
2
n+2
Hilbert polynomial of OP2 /J is equal to p(n) = 2 − ψ(n) = δn + 1.5δ − 0.5δ 2 + c.
Hence V+ (J ) ⊂ P2k is 1- codimensional and there is an irreducible component which is
equal to a hypersurface V (f ), f ∈ Sν an irreducible form (Hauptidealsatz of Krull). From
J ⊂ Rad(J ) ⊂ (f ) it follows J = f · K(−ν), where K ⊂ OP2 is an ideal with Hilbert
function χ(n) := ψ(n + ν) = n−(δ−ν)+2
− c. If δ − ν > 0, one argues with K in the same
2
way as with J , and one finally gets the statement (iii) .
Lemma B.2. Let the assumptions and the notations be as before. Then reg(I) =
min { n ∈ Z | ϕ′ (n) = n + 1 }.
Proof. As in the proof of Lemma 1, we can assume that I is Borel-normed. We let
2
Gm operate on S by σ(λ) : x 7→ x, y 7→ λg y, z 7→ λg z, where g is a high enough natural
number. Then lim σ(λ)I is equal to the ideal I0 as in the proof of Lemma 1, I and I0
λ→0
have the same Hilbert function, and reg(I) = reg(I0 ) ( cf. [Gre], Theorem 2.27). Hence
135
one can assume without restriction that I is monomial. But then the statement follows
from ([T1], Anhang 2), for instance.
136
APPENDIX C
Ideals with many monomials
If k is a field, let be S = k[x1 , . . . , xr , t] and R = k[x1 , . . . , xr ]. Gm operates on S by
σ(λ) : xi 7→ xi , 1 ≤ i ≤ r, and t 7→ λt, λ ∈ k ∗ . Let H be the Hilbert scheme of ideals
− Q(n) is
I ⊂ OPr with Hilbert polynomial Q, i.e., H = HilbP (Prk ), where P (n) = n+r
r
r
the complementary Hilbert polynomial of the subscheme V+ (I) ⊂ P . We suppose that
H is not empty. Then the ideals I ⊂ OPr with Hilbert polynomial Q, for which t is a
non-zero divisor of OP r /I, form an open, non-empty subset Ut ⊂ H.
If K/k is an extension field and if I ∈ H(K), then the limit ideals I0/∞ := lim σ(λ)I
λ→0/∞
are in H(K) again, and if I ∈ Ut , then I0 ∈ Ut , too (cf. [G2], Lemma 4). We say that I
fulfils the limit condition, if I∞ ∈ Ut .
Remark C.1. If I is fixed by the subgroup Γ : xi 7→ xi , t 7→ α1 x1 + · · · + αr xr + t of
U(r + 1; k), then I does fulfil the limit condition (cf. [G2], proof of Lemma 3, p. 541).
If I ∈ Ut , then I ′ := I + tOPr (−1)/tOPr (−1) can be regarded as an ideal in OPr−1
with Hilbert polynomial Q′ (T ) = Q(T ) − Q(T − 1).
C.0.5. Lemma. Let I ∈ H(k) ∩ Ut be an ideal which fulfils the limit condition.
(i) If d ≥ max(reg(I0 ), reg(I∞ )), then H 0 (Prk , I(d)) ∩ Rd has the dimension Q′ (d).
(ii) If d ≥ reg(I ′ ) and H 0 (I(d))∩Rd has a dimension ≥ Q′ (d), then d ≥ max(reg(I0 ), reg(I∞ )).
Proof. (i) There is a basis of M := H 0 (I(d)) of the form gi = tei gi0 + tei −1 gi1 + · · · ,
such that 0 ≤ e1 ≤ e2 ≤ · · · ≤ em , m := Q(d), gij ∈ R and gi0 ∈ Rd−ei , 1 ≤ i ≤ m, linear
independent. Then M∞ := lim σ(λ)M = h{tei gi0 |1 ≤ i ≤ m}i (limit in Grassm (Sd )) has
λ→∞
the dimension m. As d ≥ reg(I∞ ) by assumption, it follows that Q(d) = h0 (I∞ (d)), and
L 0
hence M∞ = H 0 (I∞ (d)). Now t is a non-zero divisor of S/
H (I∞ (n) by assumption,
n≥0
Thus it follows that H 0 (I∞ (n)) = h{tei −(d−n) gi0 |ei ≥ d − n}i for all 0 ≤ n ≤ d. If n = d−1
one gets H 0 (I∞ (d −1)) = h{tei −1 gi0 |ei ≥ 1}i, hence Q(d −1) = |{i|ei ≥ 1}|. It follows that
Q′ (d) = |{i|ei = 0}|. Thus M ∩ Rd ⊃ h{gi0|ei = 0}i has a dimension ≥ Q′ (d). Because
of reg(I ′ ) ≤ reg(I) one has h0 (I ′ (d)) = Q′ (d) and the canonical restriction mapping
ρd : M = H 0 (I(d)) −→ H 0 (I ′ (d)) is injective on M ∩ Rd . It follows that the dimension
of M ∩ Rd can not be greater than Q′ (d).
(ii) From the exact sequence
(1)
·t
0 −→ I(−1) −→ I −→ I ′ −→ 0
137
it follows that H i (I(n − i)) = (0) if i ≥ 2 and n ≥ e := reg(I ′ ) (see [M], p.102). The
sequence
ρ
d
0 −→ H 0 (I(d − 1)) −→ H 0 (I(d)) −→
H 0 (I ′ (d)) −→ H 1 (I(d − 1)) −→ H 1 (I(d)) −→ 0
is exact as d ≥ e, where ρ is induced by the canonical restriction mapping S −→
S/tS(−1) = R. As ρd is injective on H 0 (I(d)) ∩ Rd and h0 (I ′ (d)) = Q′ (d), it follows
from the assumption that ρd is surjective. From the e - regularity of I ′ it follows that
R1 H 0 (I ′ (n)) = H 0 (I ′ (n + 1)), for all n ≥ e. Hence ρn is surjective for all n ≥ d. Hence
0 −→ H 1 (I(n − 1)) −→ H 1 (I(n)) −→ 0 is exact for all n ≥ d, thus H 1 (I(n − 1)) = (0)
for all n ≥ d. It follows that reg(I) ≤ d. One again has the exact sequences:
(2)
·t
′
0 −→ I0/∞ (−1) −→ I0/∞ −→ I0/∞
−→ 0
As (I ′ )0/∞ = (I0/∞ )′ ⊃ I ′ and all these ideals have the Hilbert polynomial Q′ , it follows
that (I ′ )0/∞ = I ′ . As H 0 (I(d)) ∩ Rd is fixed by σ(λ), it follows that H 0 (I(d)) ∩ Rd ⊂
H 0 (I0/∞ (d)). Then one argues as before, using (2) instead of (1).
Remark C.2. Let I ⊂ OP2 be an ideal of colength d, let be S = k[x, y, z], R = k[x, y],
and let Gm operate by σ(λ) : x 7→ x, y 7→ y, z 7→ λz . We assume I to be invariant under
Γ (see above). As d ≥ reg(I) for all ideals I ⊂ OP2 of colength d, the assumption of part
(i) of the lemma is fulfilled, hence H 0 (I(n)) ∩ Rn has the dimension Q′ (n) = n+1
for all
1
n ≥ d and therefore:
H 0 (I(n)) ⊃ Rn
(3)
for all n ≥ d.
This inclusion has been used in the text for several times, e.g. in Section (2.2).
Remark C.3. Let I ⊂ OP2 be an ideal of finite colength, with Hilbert function ϕ,
which is invariant under Γ · T (ρ). Let Gm operate on S by σ(λ) : x 7→ x, y 7→ y, z 7→ λz. If
in(I) is the initial ideal with regard to the inverse lexicographical order, then in(I) is equal
to the limit ideal I0 = lim σ(λ)I. As h0 (I0 (n)) = ϕ(n), it follows that in(H 0 (I(n))) =
λ→0
H 0 (I0 (n)), for all n ∈ N (cf. Appendix E and [G2], Lemma 3 and Lemma 4, pp. 541).
Thus the number of the initial monomials and of the monomials in H 0 (I0 (n)), which are
represented in our figures, can be determined by means of the Hilbert function, alone.
138
APPENDIX D
Unipotent groups acting on polynomial rings
Lemma D.1. The 5-dimensional subgroups of ∆ = U(4; k) have the form
1 α ∗ ∗
0
1
β
∗
G(p) :=
aα + bβ + cγ = 0
0 0 1 γ
0 0 0 1
where p = (a : b : c) ∈ P2 (k) is uniquely determined.
∗
∗
⊂ ∆ is a normal subgroup. Let G be a 5-dimensional
0
1
−→ ∆/N
is
an injective homomorphism
and ∆/N ≃ G3a .
1 0 x y
0
1
0
z
x, y, z ∈ k, ax + by + cz = 0
First case: dim G∩N = 2. Then G∩N =
0 0 1 0
0 0 0 1
1 α x y
0 1 β z
2
, where
where (a : b : c) ∈ P (k) is a suitable point. It follows that G =
0 0 1 γ
0 0 0 1
α,
β, γ are any element
of k, and x, y, z ∈ k have to fulfil the conditions noted above. If
′
′
′
1
α
x
y
′
′
0
1
β
z
0 0 1 γ ′ is any other element of G, then
0 0 0 1
1
0
Proof. N :=
0
0
0 ∗
1 0
0 1
0 0
subgroup of ∆. Then G/G ∩ N
a(x′ + αβ ′ + x) + b(y ′ + αz ′ + xγ ′ + y) + c(z ′ + βγ ′ + z) = 0.
As α, β, γ, α′, β ′ , γ ′ are any elements of k, we conclude that a = b = c = 0, contradiction.
Second case: N ⊂ G. Then G/N ֒→ G3a is 2-dimensional, and one concludes from this
that G has the form noted above and that p ∈ P2 (k) is uniquely determined. Furthermore
it is easy to see that G(p) is a subgroup.
Lemma D.2. Let be P = k[x, y, z, t], let V ⊂ P be a subspace which is invariant under
G(p). If f ∈ P is invariant under G(p) modulo V , then the polynomials
x∂f /∂z, y∂f /∂t, x∂f /∂t, bx∂f /∂y − ay∂f /∂z, cx∂f /∂y − az∂f /∂t, cy∂f /∂z − bz∂f /∂t all
lie in V .
139
1 α 0 0
0 1 β 0
r s m n
Proof. If g =
0 0 1 γ ∈ G(p) and M = x y z t , then g(M) − M =
0 0 0 1
r
s
m
x (αx+y) (βy +z) (γz +t)n −M = αsxr+1 y s−1z m tn +βmxr y s+1z m−1 tn +γnxr y s z m+1 tn−1
+ terms containing αi , β i , γ i , where i ≥ 2.
If g(f )−f ∈ V , then it follows that αx∂f /∂y +βy∂f /∂z +γz∂f /∂t+ terms containing
α , β i , γ i , where i ≥ 2, lies in V .
First case: c 6= 0. Then γ = −(aα/c + bβ/c), hence αx∂f /∂y + βy∂f /∂z − (aα/c +
bβ/c)z∂f /∂t + terms containing α2 , αβ, β 2 · · · ∈ V . It follows that α(cx∂f /∂y−az∂f /∂t)+
β(cy∂f /∂z − bz∂f /∂t)+ terms containing α2 , αβ, β 2, · · · ∈ V . Put α = 0 and β = 0,
respectively. It follows that cy∂f /∂z − bz∂f /∂t ∈ V and cx∂f /∂y − az∂f /∂t ∈ V . Multiplication by a and b, respectively, and then subtracting the two polynomials from each
other gives c(bx∂f /∂y − ay∂f /∂z) ∈ V . As c 6= 0, the last three statements of the assertion follow.
Second case : c = 0, b 6= 0. Then we can choose γ ∈ k arbitrarily and −aα/b = β.
It follows that αx∂f /∂y − (aα/b)y∂f /∂z + γz∂f /∂t + terms containing α2 , γ 2 , · · · ∈ V .
Putting α = 0 gives z∂f /∂t ∈ V . Putting γ = 0 gives bx∂f /∂y − ay∂f /∂z ∈ V .
Third case: b = 0, c = 0. Then a = 1 and β and
γare any elements
of k, whereas α = 0.
1 0 ∗ ∗
0 1 0 ∗
⊂ G(p), the same reaThis gives y∂f /∂z and z∂f /∂t ∈ V . As N =
0 0 1 0
0 0 0 1
i
soning as in the proof of ([T2], Hilfssatz 1, p. 142) shows that x∂f /∂z, x∂f /∂t, y∂f /∂t ∈
V.
Lemma D.3. Let I ⊂ OP2 be a monomial ideal of colength d > 0 and let ξ be the
corresponding point in H d (k). If ξ is not invariant under the Ga -operation ψα : x 7→
x, y 7→ αx + y, z 7→ z, then C := Ga · ξ contains exactly two fixed points under T (3; k),
namely the point ξ and the Ga - fixed point ψ∞ (ξ) := lim ψα (ξ).
α−→∞
Proof. Embedding H d in Grassq (Sd ), where q := d+2
−d, one sees that it is sufficient
2
to prove the corresponding statement for a T (3; k)-invariant q - dimensional subspace
d
L
U ⊂ Sd and the corresponding point in Grassq (Sd ). As one can write U =
z d−i Ui ,
i=0
where Ui ⊂ Ri , is a subspace, it suffices to prove the corresponding statement for an
r-dimensional subspace U ⊂ Rn , which is invariant under T (2; k), but not invariant under
Ga . As lim ψα (U) is a Ga - invariant subspace and as char(k) = 0, it follows that this
α−→∞
subspace is equal to hxn , xn−1 y, . . . , xn−r+1 y r−1i. It follows that C has two fixed-points
under T (2; k). In order to prove there are no more fixed-points, it suffices to show the
following: If there is an element α 6= 0 in k such that ψα (U) is T (2; k)-invariant, then
ψα (U) is T (2; k)-invariant for all α 6= 0. If one takes a monomial M = xn−r y r ∈ U, then
ψα (M) = xn−r (αx + y)r ∈ ψα (U). As this is a monomial subspace by assumption, it
140
follows that xn−r xi y r−i ∈ ψα (U), 0 ≤ i ≤ r. Thus one has M ∈ ψα (U) and it follows
that U = ψα (U), But this implies ψnα (U) = U for all n ∈ N and thus ψα (U) = U for all
α ∈ k.
Lemma D.4. Suppose C ⊂ HQ is a B-invariant, irreducible, reduced curve, which is
pointwise invariant under Γ such that the image of C under the restriction morphism h
is a single point. Then (Mn · C) is constant for n ≫ 0 .
Proof. From ([T1], Proposition 0, p.3) it follows that C is either pointwise ∆invariant or a 1-cycle of type 3. We consider the first case. Then C = {σ(λ)ξ|λ ∈ k ∗ }− ,
where σ is a suitable Gm -operation and ξ ∈ H(k) corresponds to a ∆-invariant ideal J .
Let be I = h(J ) and I ∗ the ideal on P3 ,which is generated by I (cf. 1.2.2). Then
n
H 0 (J (n)) ⊂ ⊕ tn−i H 0 (I(i)) for all n (cf. [G5], Hilfssatz 3.2, p.295). Now the assertion
i=0
follows, as I is fixed by σ.
In the second case on has C = {ψα (ξ)|α ∈ k}− , where ψα is the usual Ga -operation
and ξ ∈ H(k) corresponds to a monomial ideal J . Then one argues as in the first case,
with ψα instead of σ.
141
APPENDIX E
Standard bases
Let k be an extension field of C. If ρ = (ρ0 , . . . , ρr ) ∈ Zr+1 − (0), then T (ρ) :=
{ λ = (λ0 , · · · , λr ) | λi ∈ k ∗ and λρ := λρ00 · · · λρr r = 1 } is a subgroup of dimension r of
Gr+1
m .
Auxiliary lemma: If σ, τ ∈ Zr+1 − (0) such that T (σ) ⊂ T (τ ), then there is an integer
n such that τ = n · σ.
Proof. Write σ = (a0 , · · · , ar ), τ = (b0 , · · · , br ) . As the dimension of T (σ) is equal
to r, there is an index i such that ai 6= 0 and bi 6= 0.Without restriction one can assume
that a0 6= 0 and b0 6= 0 . Choose p, q ∈ Z such that pa0 = qb0 and (p, q) = 1. Then
1 −qb1
r −qbr
λpa
· · · λpa
= 1 for all λ ∈ T (σ) follows. Because of dim T (σ) = r one gets
1
r
pai − qbi = 0, 0 ≤ i ≤ r, and thus σ = qρ, τ = pρ, where ρ ∈ Zr+1 − (0) is a suitable
vector. If ǫ is any q-th root of unity in C, one can choose λ ∈ (C∗ )r+1 such that λρ = ǫ.
From λqρ = λσ = 1 it follows that ǫp = λpρ = λτ = 1, too, and q = 1 follows.
We let Gr+1
operate on S = k[X0 , · · · , Xr ] by Xi 7→ λi Xi . If ρ = (ρ0 , . . . , ρr ), then
m
ρ0
ρ
ρr
X := X0 · · · Xr .
Lemma E.1. Let V ⊂ Sd be a T (ρ)-invariant subspace. Then V has a basis of the
form fi = mi · pi (X ρ ), where the mi are different monomials, pi is a polynomial in one
variable with constant term 1 and mi does not appear in mj · pj (X ρ ) if i 6= j.
Proof. By linearly combining any basis of V one obtains a basis fi = mi + gi , where
the mi are different monomials, each gi is a sum of monomials, each of which is greater
than mi in the inverse lexicographic order, and mi does not appear in gj . If g ∈ T (ρ),
then g(fi ) contains the same monomials as fi and from g(fi ) ∈ h{fi }i we conclude that
each fi is a semi-invariant, i.e, hg(fi )i = hfi i for all g ∈ T (ρ) .
P
P
ai λαi X αi =
Now let be f =
ai X αi any T (ρ)-semi-invariant. Let be λ ∈ T (ρ). Then
P
α (0)
α (r)
c(λ) · ai X αi , where λαi := λ0 i . . . λr i . It follows that λαi = c(λ), if ai 6= 0, and
therefore λαi −αj = 1 for all i, j, such that ai 6= 0 and aj 6= 0. Thus T (ρ) ⊂ T (αi − αj ), and
the Auxiliary lemma gives αi − αj = nij ρ, nij ∈ Z, if ai 6= 0 and aj 6= 0. One sees that
P
there is an exponent α0 ∈ {αi } and natural numbers ni , such that f = X α0 · ai X ni ρ .
Corollary E.1. Let V ⊂ Sd be a m-dimensional subspace, let x ∈ Grassm (Sd ) be
the closed point of Grassm (Sd ) defined by V . If the orbit T · x has the dimension 1, then
the inertia group Tx of x has the form T (ρ), where ρ ∈ Zr+1 − (0).
143
Proof. This follows by similar argumentations as before (see[T2], Hilfssatz 7,p.141).
144
Bibliography
[F] Fogarty, J.: Algebraic families on an algebraic surface. Amer. J. Math. 90, 511–521 (1968).
[Fu] Fulton, W.: Intersection theory. Springer-Verlag 1984.
[G1] Gotzmann, G.: Eine Bedingung fr die Flachheit und das Hilbertpolynom eines graduierten Ringes.
Math. Z. 158, 61–70 (1978).
: A stratification of the Hilbert scheme of points in the projective plane. Math. Z. 199,
[G2]
539–547 (1988),
: Einfacher Zusammenhang der Hilbertschemata von Kurven im komplex-projectiven Raum.
[G3]
Invent.math. 99, 655–675 (1990).
[G4]
: Topologische Eigenschaften von Hilbertfunktion–Strata. Habilitationsschrift, Universität
Münster, 1993.
: Einige einfach-zusammenhängende Hilbertschemata. Math. Z. 180, 291–305 (1982)
[G5]
[Gre] Green, M.: Generic initial ideals. In: Six lectures on commutative algebra. Progress in Mathematics
166, 119–186, Birkhäuser, Basel (1998).
[HM] Harris, J., Morrison, I.: Moduli of curves, Springer–Verlag 1998.
[Ha] Hartshorne, R.: Algebraic geometry, Springer–Verlag 1977.
[Hi] Hirschowitz, A.: Le group de Chow équivariant. C.R. Acad, Sc. Paris, t.298, Serie I, no. 5, 87–89
(1984).
[Ho] Horrocks, G.: Properties of schemes inherited from orbits. J. Alg. 40, 819–823 (1976).
[I]
Iarrobino, A.: Punctual Hilbert schemes. Mem. Amer. Math. Soc. Vol.10, N. 188 (1977).
[K] Kraft, H. : Geometrische Methoden in der Invariantentheorie, Verlag Vieweg 1985.
[L] Lella, P.: A network of rational curves on the Hilbert scheme. Preprint, available at
http://arxiv.org./abs/1006.5020.
[LR] Lella, P., Roggero, M.: Rational components of Hilbert schemes. Preprint, available at
http://arxiv.org/abs/0903.1029.
[Mu] Mumford, D.: Lectures on curves on an algebraic surface, Pinceton, 1966.
[T1] Gotzmann, G.: Der kombinatorische Teil der ersten Chowgruppe eines Hilbertschemas von
Raumkurven, Schriftenreihe des Mathematischen Instituts der Universität Münster, 3. Serie, Heft
13, September 1994.
: Der algebraische Teil der ersten Chowgruppe eines Hilbertschemas von Raumkurven, ibid.,
[T2]
Heft 19, Februar 1997.
: Die Néron–Severi–Gruppe eines Hilbertschemas von Raumkurven und der universellen
[T3]
Kurve, ibid., Heft 23, Januar 1999.
: Die erste Chowgruppe eines Hilbertschemas von Raumkurven, ibid., Heft 25, März 2000.
[T4]
145
| 0math.AC
|
F R O N T MA T T E R
Time-Ordered Product Expansions
for Computational Stochastic Systems Biology
Eric Mjolsness
UC Irvine Department of Computer Science
[email protected]
September 2012
BODY
Abstract
The time-ordered product framework of quantum field theory can also be used to understand salient
phenomena in stochastic biochemical networks. It is used here to derive Gillespie’s Stochastic
Simulation Algorithm (SSA) for chemical reaction networks; consequently, the SSA can be interpreted in terms of Feynman diagrams. It is also used here to derive other, more general simulation and
parameter-learning algorithms including simulation algorithms for networks of stochastic reaction-like
processes operating on parameterized objects, and also hybrid stochastic reaction/differential equation
models in which systems of ordinary differential equations evolve the parameters of objects that can
also undergo stochastic reactions. Thus, the time-ordered product expansion (TOPE) can be used
systematically to derive simulation and parameter-fitting algorithms for stochastic systems.
1
Introduction
The master equation for a continuous-time stochastic dynamical system may be expressed as
d p ê d t = W ÿ p where W is the time-evolution operator, often an infinite-dimensional matrix. Particular
choices for W lead to the special case of the “chemical master equation” for stochastic chemical kinetics,
often useful in bioligical applications; we will see this and other applications below. The general master
equation has the formal solution pHtL = expHt WL ÿ pH0L . If W can be decomposed as a sum W0 + W1 , then
there is a perturbation theory for expHt WL in terms of expHt W0 L and its perturbations by W1 . The timeordered product expansion (which we refer to by the acronym TOPE) gives a formula for the solution of a
master equation [1-3] which can be expressed as follows [4]:
exp Ht WL ÿ p0 = exp Ht HW0 + W1 LL ÿ p0
ÄÅ t
ÉÑ
tk
t2
Å
Ñ
= ‚ ÅÅÅŇ d tk ‡ d tk-1 ‡ d t1 expHHt - tk L W0 L ÿ W1 ÿ expHHtk - tk-1 L W0 L ÿ W1 ÿ expHt1 W0 L ÿ p0 ÑÑÑÑ
ÅÇ 0
ÑÖ
0
0
k=0
¶
(1)
This expression can be derived (as in [4]) by expanding in powers of W1 , each expanded to all orders in
W0 , and using the normalization formula for the Dirichlet distribution to subdivide the time interval @0, tDL
into k subintervals.
Since W0 and W1 do not generally commute, this expression involves alternation from right to left
of W0 and W1 related operations. Using the “time-ordered exponential” of operators [5], this result can be
compactly reexpressed as:
i i
yy
expHt HW0 + W1 LL = expHt W0 L jjexpjj‡ W1 Ht L d tzzzz
k k 0
{{+
t
(2)
Q-BioPaperMjolsness2012V23TR.nb
where
W1 Ht L ª expH-t W0 L W1 expHt W0 L .
(3)
Here IexpIŸ0 GHtL d tMM is obtained term by term from the Taylor series for the operator exponential, by
+
reordering all monomials containing terms evaluated at different times so that they are indexed by ordered
sequences of times Htk , ..., t1 L that increase right to left (details reviewed in Section 3.5.1 below). In field
theory it is standard to prefer Equation 2 over Equation 1 for theoretical calculations, but for algorithmic
concreteness this paper will favor the more explicit expression Equation 1 where possible. Indeed, each
summand of Equation 1 already looks like a Markov chain in which the matrix or operator product operation
“ÿ ”. which sum over states, is supplemented by integration over an extra time variable. This observation will
be made precise in Section 3. In general there is a risk that the infinite sum over terms could diverge. However, the master equation must conserve total probability and this constrains W to have zero column-sums and
also constrains the spectrum of W to have nonpositive real parts. In this setting some decompositions
W = W0 + W1 converge well enough, as we will see by example below.
t
One particular specialization of TOPE lets us derive Gillespie’s Stochastic Simulation Algorithm
`
(SSA): take W0 = -D = the diagonal part of W , and W1 = W = the off-diagonal part of W . Then for chemical
reaction networks TOPE generates Feynman-like diagrams. An example is illustrated below for the simple
reaction network with just two reactions, the forwards and backwards parts of the generic trivalent reaction
A + B F C , to which others can be reduced.
Figure 1. A time history of the reaction A + B F C . Time flows left to right. Open circles represent reaction
events, with probability factor µ W1 . In between reaction events are unimolecular particle propagators
expHHtk - tk-1 L W0 L, labelled by arrows and particle names (repeated for clarity). This is a non-spatial version
of the Lee model in quantum field theory (cf. for example [6]).
The TOPE (Equation 1 or Equation 2) can be applied recursively, since it reduces one operator
exponential expHt WL to another one expHt W0 L. This fact will be exploited in Section 3 below. But eventually
one must get to an operator exponential that is tractable by other means. One way to do this is to let W0 = D =
the diagonal part of W , as in the SSA algorithm derivation below.
2
Q-BioPaperMjolsness2012V23TR.nb
2
2.1
Methods
Creation/annihilation operator notation
We will use operator notation for molecule (or other
changes [1-4]. Here we just review the notation as used in [4].
(respectively) to destroy and create identical particles of a given
elements have the entirely off-diagonal expressions
a = jd
and a` = d
for all i, j in
ij
i j-1
ij
reactant) creation and annihilation state
The elementary operators a and a` act
type. In the particle-number basis their
i j+1
80, 1, 2, ...< .
(4)
Here di j is the Kronecker delta function. The creation and annihilation operators satisfy the Heisenberg
algebra @a, a` D = I but are different from those of quantum mechanics because they are not conjugates or
transposes of one another. (This is the reason we do not denote the creation operator a† , as it is in quantum
mechanics, or a* .) Instead of being conjugate to `a , the annihilator a encodes the chemical law of mass action
since its nonzero entries are equal to the number of particles available to react or decay. The diagonal
“number operator” is N ª a` a .
The creation and annihilation operators may be represented in terms of their action on probability
n
generating functions gHzL = ڦ
n=0 pn z , where pn is the probability that there exist n particles of a given type.
In this case:
a = and a` = z µ
(5)
z
In the presence of different types of particles (eg. molecules or other objects) the
creation/annihilation operator notation is generalized, eg to aa and a` b for molecule types Aa , in which all
operators for unequal types commute:
@a , a` D ª a a` - a` a = I d
Operating on an empty “vacuum” state †0\ with no objects, the monomials in the creation operators a` b
span a Fock space. Molecule or object types indexed by a may even be taken to include arbitrary discretevalued molecular attributes (or attributes of other objects) such as phosphorylation state or integer-valued
parameters. Continuous-valued parameters such as position (in quantum field theory it would more naturally
be the conserved momentum, unlike the typical viscous-medium dynamics in biology) may be encoded into a
real-valued vector argument xwhich requires a Dirac delta function instead of a Kronecker delta function, so
for example:
@aa HxL, a` b HyLD = I da b dHx - yL
(6)
a
b
a
b
b
a
ab
A non-molecular example of such parameterized objects would be: cells of a given real-valued volume or
lengthscale as in Section 3.5.5 below.
However for some attributes such as real-valued object positions one may wish to limit the state
space to between zero or nmax, a molecules (or other objects) at each unique real value. The resulting commutator is still diagonal as described in [4]. The particular case nmax, a = 1 is not a stochastic version of fermions
because particles with different types or values of the attributes still commute rather than anticommuting.
The basic rule for translating chemical reactions into creation/annihilation operator notiation is:
first, annihilate all objects on the incoming or left hand side of a reaction; then create all the objects on the
outgoing or right hand side of a reaction. Thus, the off-diagonal part of the operator for a reaction
8AaHpL HpL » 1 b p b pmax <* ö 8A bHqL HqL » 1 b q b qmax <* with reaction rate rr
3
Q-BioPaperMjolsness2012V23TR.nb
that converts an incoming multiset 8<* of numerically parameterized rectants 8AaHpL HpL » p œ lhsHrL<*
(reactants can appear multiple times in a multiset) into an outgoing multiset 8A bHqL HqL » q œ rhsHrL<* , is:
ÄÅ
ÉÑ ÄÅ
ÉÑ
ÅÅ
ÑÑ ÅÅ
ÑÑ
`
ÅÅ
ÑÑ ÅÅ
Ñ
`
Or = rr ÅÅÅ ‰ a bHqL Hyq LÑÑÑ ÅÅÅ ‰ aaHpL Hxq LÑÑÑÑ .
(7)
ÅÅqœrhsHrL
ÑÑ ÅÅ pœlhsHrL
ÑÑ
ÅÇ
ÑÖ ÅÇ
ÑÖ
There is one such operator for every possible set of values for the numerical parameters. Since timeevolution operators for different processes just add, a generic operator for all parameter values must sum
and/or integrate the operator of Equation 7 over all the parameters, in the Cartesian product of measure spaces
in which they take values. Equation 7 adds probability to the new state of the system, but does not take it
away from the old state of the system before a reaction. That job requires a negative diagonal matrix as shown
in Equation 9 below. In the case of Equation 7, the corresponding diagonal operator is
Dr = rr A¤aœlhsHrL NaHpL Hx p LE . Examples are provided in [4] and below.
2.2
Solvable example: An exact solution for SSA behavior
For a few very simple examples, we can not only solve analytically for the behavior of the biochemical system, but we can even add in the behavior of the SSA simulation algorithm and solve for that exactly as
well. For example consider the minimal bidirectional reaction Aõ Ø. This case is analytically solvable,
including the complete statistics of its SSA algorithm simulation. It has forward synthesis and backwards
decay reactions. The operator expression is therefore:
W = k Ha a` - IL + k Ha a - NL
(8)
s
d
Here a = 1 is the generating function variable for the total number of reactions, corresponding to offdiagonal matrix elements of W . Power series in a will decompose total probability according to this number.
Translating the master equation for Equation 8 into a PDE in the two variables t and z using
representation Equation 5, and solving analytically, this model has the exact solution
ks
m
gm Hz, t » aL = Ia + Hz - aL e-kd t M expC- ÅÅÅÅÅÅÅÅ IH1 - aL kd t + Iz a - a2 M Ie-kd t - 1MMG
kd
ks
m
= Ia + Hz - aL e-kd t M expC ÅÅÅÅÅÅÅÅ Hz a - 1L I1 - e-kd t MOG
kd
ks
äexpC ÅÅÅÅÅÅÅÅ Ia2 - 1M Ikd t + e-kd t - 1MOG
kd
= Binomial initial condition with decay * Poisson on forward reactions
* Poisson on forward ê backward reaction pairs.
As usual z is the generating function variable whose exponent is the total number nA of A molecules or
particles, m is the initial number of molecules, and t is continuous time. The * operation is a convolution of
probability distributions. A product of generating functions with the same variable is a convolution of distributions [7]. Note the interpretation in terms of Binomials and Poissons with time-varying parameters. The third
factor represents a linearly increasing number of canceling forward/backward reaction pairs as a function of
time - a kind of random walk.
The full derivation below will generalize this solvable example, again separating the diagonal from
the off-diagonal terms in W .
4
Q-BioPaperMjolsness2012V23TR.nb
2.3
Notation for SSA rederivation from TOPE
One specialization of TOPE lets us derive SSA for biochemical reaction networks, as follows. First
decompose W into nonegative off-diagonal and nonpositive diagonal parts, as must be possible by the conservation and nonnegativity of probability. For example conservation of probability implies
" p 0 = dH1 ÿ pL ê d t = H1 ÿ WL ÿ p fl 1 ÿ W = 0 . Then
`
`
W = W - D, where D = diagI1 ÿ W M, i.e.
(9)
`
`
W I J = H1 - dI J L WI J and D I J = dI J ‚ W K J
K
where I and J index the possible states of the system. To prevent negative probabilities from evolving
`
under the master equation, all entries of W and therefore D must be nonnegative. In this circumstance the
TOPE becomes:
exp It IẀ - DMM =
ÄÅ
ÉÑ
k
ÅÅ t
t
ÑÑ
ij
yz
`
`
ÅÅ
Ñ
j
z
k
„ ÅÅŇ ‡ IPq=0 d tq M djjjt - ‚ tq zzz expH-tk DL W expH-t1 DL W expH-t0 DL ÑÑÑÑ
ÅÅÅ 0
ÑÑÑ
0
Ç
k q=0 {
Ö
k=0
¶
= „ ‡ ‡
¶
t
0
t
IPkq=0
0
ÅÄÅ 0
ÑÉÑ
k
ij
yz
ÅÅ
ÑÑ
`
j
z
Å
j
z
Å
d tq M djt - ‚ tq zz expH-tk DLÅÅ ‰ W expH-tq DLÑÑÑÑ
j
Å
ÑÑ
ÅÅÇ q=k-1
ÑÖ
k q=0 {
We define the conditional probability distribution (where @tq Dk0 ª @t0 , ... tk D denotes an ordered contiguous sequence of time intervals)
ÄÅ 0
ÉÑ
k
l
ÑÑ i
y|
ÅÅÅ
o
o
`
o
k
ÅÅ ‰ W expH-t DL ÑÑÑ djjjjt - ‚ t zzzzo
PrII, @tq D0 , k » J, tM = m
expH-t
DL
}
k
q
q
Å
Ñ
(10)
j
z
o
o
ÑÑ
o
o
ÅÅÅ
Ñ
q=0
q=k-1
Å
Ñ
n
Ç
Ö k
{~I,J
k=0
For DI I 0,
PrII,
@tq Dk0 ,
ÅÄÅ 0
ÑÉÑ i
k
l
y|
o
ÅÅ
ÑÑ jj
o
o
-1
Å
ÑÑ djjt - ‚ t zzzzo
Å
k » J, tM = m
expH-t
DL
IẀ
D
M
HD
expH-t
DLL
‰
k
Å
q
Ñ
q
z}
o
o
ÅÅ
Ñ j
o
Ñ
ÅÅÇ q=k-1
ÑÑÖ k q=0 {o
n
~I,J
Either way,
PrHI » J, tL = ‚ ‡ ‡ IPkq=0 d tq M PrII, @tq Dk0 , k » J, tM = AexpIt IẀ - DMME .
I,J
¶
k=0
2.4
t
0
t
0
Semigroup property
Suppose t = t1 + t2 , all nonnegative. Then for any time-evolution equation we must have the semigroup
property:
PrHI » J, tL = ‚ PrHI » K, t2 L PrHK » J, t1 L
K
i.e.
AexpIt IẀ - DMMEI,J = ‚ AexpIt2 IẀ - DMMEI,K AexpIt1 IẀ - DMME
K
5
K,J
.
Q-BioPaperMjolsness2012V23TR.nb
Is there a k -event version of this rule, for k = k1 + k2 ? In other words, can we add (nonnegative) numbers
of reaction events rather than time intervals ?
We observe (where again @tq Dk0 ª @t0 , ... tk D )
PrHI, k » J, tL = ‡ ‡ IPkq=0 d tq M PrII, @tq Dk0 , k » J, tM
t
0
and
t
0
k
‚ ‡ d t PrII, Atk£ 1 , tk1 +1 , ... tk E, k2 » K, tM PrIK, @tq D01 , k1 » J, t - tM
t
ÄÅ k +1
ÉÑ
l
ÅÅ 1
ÑÑ
o
`
o
ÅÅ
ÑÑ
Å
ÑÑAẀ expI-t£k1 DM E
= · dtm
expH-t
DL
W
expH-t
DL
‰
k
q
o
ÅÅÅ
ÑÑ
o
ÅÅÇ q=k-1
ÑÑÖ
n
0
Ä
ÉÑ
ÅÅ 0
k1 -1
ÑÑ ij
ij
ij k
yzyz
yz|
Å
o
`
Å
Ñ
o
j
j
z
z
ä djjjt - jjj ‚ tq + t£k1 zzzzzz expH-tk1 DLÅÅÅÅ ‰ W expH-tq DL ÑÑÑÑ djjjjt - t - ‚ tq zzzzo
}
o
Å
Ñ
ÅÅÇ q=k1 -1
ÑÑÖ k
q=0
k
kq=k1 +1
{{
{~ I,J
ÄÅ k +1
ÉÑ
l
ÅÅ 1
ÑÑ
o
`
o
ÅÅ
ÑÑ
Å
ÑÑAẀ expI-It£k1 + tk1 M DM E
=m
expH-t
DL
W
expH-t
DL
‰
k
ÅÅ
q
o
ÑÑ
o
ÅÅ q=k-1
ÅÇ
ÑÑÖ
n
ÄÅ 0
ÉÑ
k1 -1
ÅÅ
ÑÑ ij ij k
yzyz|
o
`
Å
Ñ
o
ä ÅÅÅÅ ‰ W expH-tq DL ÑÑÑÑ djjjjt - jjjj ‚ tq + t£k1 + ‚ tq zzzzzzzz}
o
o
ÅÅ q=k -1
ÑÑ
q=0
ÅÇ 1
ÑÖ k kq=k1 +1
{{~I,J
K
0
t
= PrII, A@tq Dk01 -1 , t£k1 + tk1 , @tq Dkk
E, k » J, tM .
Thus, if k = k1 + k2 and for any t£k1 œ @0, tk1 D , there is a semigroup law:
1 +1
PrII, @tq Dk0 , k » J, tM =
‚ ‡ d t PrII, Atk£ 1 , tk1 +1 , ... tk E, k2 » K, tM PrIK, At0 , ... tk1 -1 , tk1 - t£k1 E, k1 » J, t - tM
t
(11)
0
In this derivation there is an arbitrary choice of t£k1 from the interval @0, tk1 D. Consistent uniform possibilities include t£k1 = 0 and t£k1 = tk1 . We will choose the latter option, so that the last t in a series @tq D is zero.
The result used in the next section will be a semigroup law for “just-fired” probabilities, right after a reaction
or (more generally) a rule firing.
K
3
Results and discussion
Given the foregoing semigroup expression from TOPE, we complete the derivation of a Markov
chain representing the SSA algorithm. We then consider extensions of this result, including parameterized
reactants, but focussing mainly on hybrid stochastic event/ordinary differential equation dynamical systems.
3.1
3.1.1
Derivation of a Markov chain
Just-fired probabilities and Bayes’ rule
Define the just-fired probabilities
6
Q-BioPaperMjolsness2012V23TR.nb
è
PrII, @tq Dk-1
, k » J, tM = PrHI, @t0 , ... tk-1 , tk = 0D, k » J, tL
0
ÄÅ 0
ÉÑ
k
ÅÅ
ÑÑ
ij
yz
`
ÅÅ
Ñ
= ÅÅÅ ‰ W expH-tq DL ÑÑÑÑ djjjjt - ‚ tq zzzz
ÅÅ q=k-1
ÑÑ
ÇÅ
ÖÑI,J k q=0 {
and calculate their conditional distributions given the number of events k , using Bayes’ rule :
è
PrII, @tq Dk-1
… J, t, kM = PrII, A@tq Dk-1
, tk = 0E … J, t, kM
0
0
PrII, A@tq Dk-1
, 0E, k » J, tM
0
= ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
t
t
k-1
£
£
£
‚ £ Ÿ0 Ÿ0 IPk-1
q=0 d tq M PrJI , BAtq E0 , 0F, k » J, tN
I
0
`
B‰
W expH-tq DL F dIt - ⁄kq=0 tq M
q=k-1
I,J
= ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ .
t ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
0
`
k
k-1
£
£
W expI-tq DM F dIt - ⁄q=0 t£q M
„ ‡ ‡ IPq=0 d tq MB ‰
I£
q=k-1
0
I £ ,J
With tk£ 1 = tk1 and tk = 0 the semigroup law becomes:
PrII, A@tq Dk-1
, 0E, k » J, tM
0
= ‚ ‡ d t PrHI, @tk1 , ... tk-1 , 0D, k2 » K, tL PrHK, @t0 , ... tk1 -1 , 0D, k1 » J, t - tL
t
K
i.e.
0
t
è
è
è
k-1
k1 -1
PrII, @tq Dk-1
,
k
»
J,
tM
=
‚
‡ d t PrII, @tq Dk , k2 » K, tM PrIK, @tq D0 , k1 » J, t - tM
0
K
1
0
(12)
For k2 = 1 and k > 1 ,
t
è
è
è
PrII, @tq Dk-1
,
k
»
J,
tM
=
d t PrHI, tk-1 , 1 » K, tL PrIK, @tq Dk-2
, k - 1 » J, t - tM
‚
‡
0
0
K
(13)
0
Suppose now that we wish to sample from this dynamical system for a total time t that is drawn from an
exponential distribution with mean T :
è
PrHtL = expH-t ê TL ê T
We will take T to be much longer than other characteristic times in the system given by nonzero matrix
elements of W . Then
ÅÄÅ 0
ÑÉÑ
k
ij
yz
Å
ÑÑ
è
è
è
`
k-1
k-1
-1 Å
Å
PrII, @tq D0 , t, k » JM = PrII, @tq D0 , k » J, tM PrHtL = T ÅÅÅ ‰ W expH-tq HD + ! ê TLL ÑÑÑÑ djjjjt - ‚ tq zzzz
ÅÅ q=k-1
ÑÑ
ÅÇ
ÑÖI,J k q=0 {
where ! is the identity matrix.
Then
è
PrII, @tq Dk-1
, k » JM = ‡
0
0
¶
ÅÄÅ 0
ÑÉÑ
Å
ÑÑ
`
è
k-1
-1 Å
Å
d t PrII, @tq D0 , t, k » JM = T ÅÅÅ ‰ W expH-tq HD + ! ê TLL ÑÑÑÑ
ÅÅ q=k-1
ÑÑ
ÅÇ
ÑÖI,J
ÄÅ 0
ÉÑ
ÄÅ 0
ÉÑ
ÅÅ
ÑÑ
ÅÅ
ÑÑ
¶
è
`
`
ÑÑ
ÅÅ
Ñ
-1
-1 Å
-1
Å
PrHI, k » JL = T ÅÅÅ ‰ W ‡ d tq expH-tq HD + ! ê TLL ÑÑÑ = T ÅÅÅ ‰ W HD + ! ê TL ÑÑÑÑ
ÅÅ q=k-1
Ñ
Å
ÑÑ
0
ÑÑÖ
ÅÅÇ q=k-1
ÅÇ
ÑÖI,J
I,J
Assuming T p k ê DI I for every element D I I of D (assuming the reaction network has been modified as in
[2] to formally eliminate terminal states),
7
Q-BioPaperMjolsness2012V23TR.nb
HD + ! ê TL-1 = D-1 I! + D-1 ë TM
and
-1
k
è
PrHI, k » JL = T -1 B IẀ D-1 M F
º D-1 I! - D-1 ë T + ...M
I,J
I1 + OIk D-1 ë TMM
è
è
PrHk » JL = ‚ PrHI, k » JL º T -1 I1 + OIk D-1 ë TMM
Small t ê T :
I
è
è
PrHt » JL = PrHtL = expH-t ê TL ê T = T -1 H1 + OHt ê TLL
Bayes’ rule:
è
PrHk » JL
è
è
k-1
PrII, @tq Dk-1
,
k
»
J,
tM
=
Pr
II,
@t
D
,
t
»
J,
kM
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
q
è ÅÅÅÅÅÅÅÅ
0
0
PrHtL
Substituting this expression into Equation 13 everywhere,
è
PrHk » JL
è
PrII, @tq Dk-1
,
t
»
J,
kM
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
è ÅÅÅÅÅÅÅÅ
0
PrHtL
è
è
PrH1 » KL PrHk - 1 » JL è
è
k-2
= „ ‡ d t ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
è ÅÅÅÅÅÅÅÅÅÅ ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
è ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ PrHI, tk-1 , t » K, 1L PrIK, @tq D0 , t - t » J, k - 1M
PrHtL
PrHt - tL
0
t
K
è
è
è
t
ij PrH1 » KL PrHk - 1 » JL yz ij
yz
PrHtL
è
j
z
j
PrII, @tq Dk-1
,
t
»
J,
kM
=
d
t
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅ
Å
ÅÅÅ
Å
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
„
j
z
j
‡
è
è ÅÅÅÅÅÅÅÅÅ zzz
j
zj è
0
PrHk » JL
0
k
{ k PrHtL PrHt - tL {
K
è
è
ä PrHI, tk-1 , t » K, 1L PrIK, @tq Dk-2
, t - t » J, k - 1M
0
But we can evaluate the ratios
è
PrHtL
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
è ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
è ÅÅÅÅÅÅÅÅÅ = T
PrHtL PrHt - tL
è
è
PrH1 » KL PrHk - 1 » JL
-1
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
è ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ = T H1 + OHt ê TLL
PrHk » JL
Thus to leading (zero’th) order in t ê T ,
t
è
è
è
k-2
PrII, @tq Dk-1
,
t
»
J,
kM
º
‚
‡ d t PrHI, tk-1 , t » K, 1L PrIK, @tq D0 , t - t » J, k - 1M
0
(14)
0
K
and therefore (by integration over tq )
t
è
è
è
PrHI, t » J, kL º ‚ ‡ d t PrHI, t » K, 1L PrHK, t - t » J, k - 1L
K
(15)
0
This can be interpreted as an infinite-dimensional (first-order) Markov chain which increments k each
iteration, and which updates I and t according to the conditional probability operator
8
Q-BioPaperMjolsness2012V23TR.nb
è
"HI, t £ » J, tL = PrHI, t £ - t » J, 1L = ‡
¶
è
d t£ PrHI, t£ , t £ - t » J, 1L
è
PrHt£ - tL
è
d t£ PrHI, t£ , 1 » J, t £ - tL ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
è ÅÅÅÅÅÅÅÅÅÅÅ
PrH1 » JL
0
¶
expH-Ht £ - tL ê TL ê T
è
º ‡ d t£ PrHI, t£ , 1 » J, t£ - tL ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ .
1êT
0
=‡
Using
0
¶
`
è
PrHI, t£ , 1 » J, t£ - tL = A W expH-t£ DL EI,J dHt£ - t - t£ L
we calculate
"HI, t£ » J, tL º ‡
¶
`
d t£ A W expH-t£ DL EI,J dHt £ - t - t£ L expH-Ht £ - tL ê TL
`
= A W expH-Ht £ - tL DL EI,J expH-Ht £ - tL ê TL QHt £ r tL
`
= B W expH-Ht£ - tL HD + ! ê TL DI,J QHt £ r tL
`
º A W expH-Ht £ - tL DL EI,J QHt £ r tL .
0
Thus,
`
"HI, t £ » J, tL º W I,J expH-Ht £ - tL DJ J L QHt £ r tL .
This expression is properly normalized:
‚‡
I
¶
-¶
d t£ "HI, t £ » J, tL = DJ J ‡
= DJ J ‡
¶
-¶
¶
(16)
d t£ expH-Ht£ - tL DJ J L QHt£ r tL
d t expH-t DJ J L = 1
0
as claimed. For DJ J 0 , we may factor " into a time update followed by a state update, each normalized:
M1 ª IẀ I,J DJ J
"HI, t £ » J, tL º M1 M0 where
M and M0 ª @DJ J expH-Ht £ - tL DJ J L QHt£ r tLD .
-1
(17)
Here matrix M0 is diagonal, so the elementwise product is also a matrix product. Thus
t
è
è
PrHI, t » J, kL º ‚ ‡ d t "HI, t » K, t - tL PrHK, t - t » J, k - 1L
= ‚‡
K
K
¶
0
è
d told "HI, t » K, told L PrHK, told » J, k - 1L, where
0
è
è
PrHt » J, kL ª " Î PrHt - t » J, k - 1L
and finally the algorithmic Markov chain expression
è
è
PrHt » J, kL = " k ÎPrH0 » J, 0L .
3.1.2
(18)
(19)
SSA Theory Summary
We use the Time-Ordered Product Expansion with the diagonal decomposition:
`
`
W = W - D and D = diagI1 ÿ W M .
Then an event-oriented simulation requires the use of Bayes’ rule to convert from
PHa, @tq » 0 b q b k - 1D, k » c, tL to PHa, @tq » 0 b q b k -91D, t » c, kL , and this in turn requires a distribution on
total simulation time t . To this end, suppose there is a small constant probability per unit time, e = 1 ê Tlong ,
that a simulation of the stochastic process defined by W comes to an end. This termination process will
provide an exponential prior distribution on the time variable t . Taking the limit in which Tlong is much longer
than the time t of interest to us, so that e t = t ê Tlong Ø 0 , the parameter Tlong drops out of the calculation and
the Bayes-transformed simulation statistics are unaffected by the probability of pending termination.
Q-BioPaperMjolsness2012V23TR.nb
Then an event-oriented simulation requires the use of Bayes’ rule to convert from
PHa, @tq » 0 b q b k - 1D, k » c, tL to PHa, @tq » 0 b q b k - 1D, t » c, kL , and this in turn requires a distribution on
total simulation time t . To this end, suppose there is a small constant probability per unit time, e = 1 ê Tlong ,
that a simulation of the stochastic process defined by W comes to an end. This termination process will
provide an exponential prior distribution on the time variable t . Taking the limit in which Tlong is much longer
than the time t of interest to us, so that e t = t ê Tlong Ø 0 , the parameter Tlong drops out of the calculation and
the Bayes-transformed simulation statistics are unaffected by the probability of pending termination.
è
Define a prior on the total simulation time with T Ø ¶ : PrHtL = expH-t ê TL ê T . Then from the
TOPE, the just-fired probabilities form a Markov chain:
ÄÅ 0
ÉÑ
ÅÅ
ÑÑ
¶
è
`
Ñ
-1 Å
Å
PrHI, k » JL = T ÅÅÅ ‰ W ‡ d tq expH-tq HD + ! ê TLL ÑÑÑÑ
ÅÅ q=k-1
ÑÑ
0
ÅÇ
ÑÖI,J
ÅÄÅ 0
ÑÉÑ
Å
ÑÑ
`
-1
-1 Å
Å
= T ÅÅÅ ‰ W HD + ! ê TL ÑÑÑÑ
ÅÅ q=k-1
ÑÑ
ÅÇ
ÑÖ
k
è
PrHI, k » JL = T -1 B IẀ D-1 M F
I,J
I,J
I1 + OIk D-1 ë TMM
The joint probabilities of system states and elapsed time also form a Markov chain:
"HI, t £ » J, tL º IẀ I,J DJ J -1 M@DJ J expH-Ht £ - tL DJ J L QHt £ r tLD and
è
è
PrHt » J, kL ª " Î Pr Ht - t » J, k - 1L .
(20)
where the “Î ” inner product operation combines both a sum over all states and an integral over all nonnegative times. The off-diagonal operator is a sum of reaction propensities kHrL
J multiplied by reaction-specific
matrices SHrL
specifying
their
effects
on
particle
numbers,
so
we
may
write
IJ
`
` HrL
HrL
W I,J = ‚ W I J = ‚ kHrL
J SI J , where
SHrL
IJ
HrL
œ 80, 1< and SHrL
I J = 1 fl kJ > 0
r
r
HrL
-1
Now using SHrL
M in Equation 20:
I I = 0 and ‚ SI J b 1 , we have a Markov interpretation of IẀ I,J DJ J
`
W I,J DJ J -1 = ‚ ProbHr » JL SHrL
I J = ProbHI » JL , where
I
r
£
kHrL
J
ProbHr » JL = ÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅ and kJHtotalL = ‚ kHrJ L .
HtotalL
kJ
r£
Of course the factor @DJ J expH-Ht £ - tL DJ J L QHt£ r tLD in Equation 20 is just the SSA exponential distribution of non-negative waiting times between reaction events. Finally iterating the recursion relation of Equation
è
è
20, we find the algorithmic Markov chain expression for SSA (Equation 19) : PrHt » J, kL = "k Î PrH0 » J, 0L ,
which expresses the iteration of a Markov chain.
This Markov chain expression has also been used as the starting point for the derivation of exact
accelerated stochastic simulation algorithms [10,11] that execute many reactions per step (i.e. they “leap”
forward) and thus go much faster than SSA, while also sampling from the exact probability distribution given
by the just-fired probabilities above. These derivations proceed by algebraic rearrangement of terms to
express computationally efficient versions of rejection sampling. The algorithm of [11] has been parallelized,
which is often difficult for discrete-event simulations.
The foregoing derivation was outlined in far less detail in [8]. A similar equation for SSA was
reached by very different methods in [9], Theorem 10.1. To our knowledge this is the first complete derivation of SSA from field theory methods like TOPE.
10
Q-BioPaperMjolsness2012V23TR.nb
3.2
Algorithm: SSA
The SSA algorithm represented by the Markov chain in Equations 19 and 20 above may be written out in
pseudocode as follows:
repeat {
compute propensities kHrL
compute kHtotalL = ⁄r kHrL
draw waiting time D t from kHtotalL expH-D t kHtotalL L
draw reaction r from distribution kHrL ê kHtotalL and execute reaction r
t := t + D t ;
// advance the clock by D t
} until t r tmax
3.3
Extension: Parameterized rule and graph grammar SSA-like algorithm
For biological modeling, including spatial and mechanical modeling of biological systems, it is
very important to generalize from pure particles to particles with both discrete and continuous attributes. The
complication is that reaction or process rates can then depend on the attributes both of the incoming and
outgoing objects. A non-molecular example of such parameterized objects would be cells of a given size,
whose propensity to divide may actually depend on their real-valued size parameter (as in Section 3.5.5).
More generally, this capability enables agent-based modeling and simulation since it allows interacting
objects to have dynamic internal state and even (as explained in Section 3.3.2 below) dynamic relationships.
Generalizing from Equation 7, as in [4], we include all possible instantiations of parameters x and
y, the integrated off-diagonal process representation operator
ÄÅ
ÉÑ ÄÅ
ÉÑ
ÅÅ
ÑÑ ÅÅ
ÑÑ
`
ÅÅ
ÑÑ ÅÅ
Ñ
`
Or = ‡ ‡ d 8x< d 8y< rr H@x p D, @yq DL ÅÅÅ ‰ a bHqL Hyq LÑÑÑ ÅÅÅ ‰ aaHpL Hxq LÑÑÑÑ
(21)
ÅÅqœrhsHrL
ÑÑ ÅÅ pœlhsHrL
ÑÑ
ÅÇ
ÑÖ ÅÇ
ÑÖ
The generalization is conceptually trivial because we have simply used a function rr H@xa D, @yb DL to express
the possibly infinite number of different reaction rates that pertain to objects that differ only in their attributes.
Again (as in Equation 7 ), summation over all discrete-valued parameters and integration over all continuousvalued parameters generalizes the operator to handle all possible sets of parameter values.
3.3.1
Algorithm: SSA with parametrized reactant objects
The resulting variant of the SSA algorithm for parameterized reactions can be expressed in
pseudocode as follows (outlined briefly in [8]):
forall reactions r factor rHrL Hxin , xout L = kHrL Hxin L pHrL Hxout » xin L;
compute SSA propensities as kHrL Hxin L;
repeat {
compute kHtotalL = ⁄r kHrL Hxin L ;
draw waiting time D t from kHtotalL expH-D t kHtotalL L ;
draw reaction r from distribution kHrL Hxin L ê kHtotalL ;
draw xout from pHrL Hxout » xin L and execute reaction r;
t := t + D t ;
// advance the clock by D t
11
Q-BioPaperMjolsness2012V23TR.nb
} until t r tmax
3.3.2
Structural matching
The functions rHx, yL appearing in Equation 21 may impose constraints including equality of
variables; equivalently we may allow some variables to appear multiple times in object parameter lists. Either
way there follows a mechanism to encode structural relations - graphs and labelled graphs - in the input and
output variable lists. Object attributes can include Object ID codes which other objects can also include in
their parameter lists. (Of course, the numeric values of Object IDs can be globally permuted without changing the structural relationships among extant objects.) In this way, the integrated version of the parameterized
reaction operator above encodes structural pattern matching, including variable-binding in logical formulate,
among the preconditions that can be enforced before such a generalized reaction or “rule” can fire.
From this point of view, syntactic variable-binding has the semantics of multiple integration [4]. In
this way we can entrain pattern-matching systems such as the computer algebra system Mathematica, or logicbased programming languages, to the job of simulating complex process rules. As in rule-based expert
systems, when multiple rules might fire the Rete algorithm [10] can be used to speed up the computations
required to maintain knowledge of their relative rates.
The resulting systems have the power to model and simulate dynamic labelled graphs including
growing multicellular tissues with dynamical cell-neighbor relationships [4] and molecular complexes with
dynamical binding structure [12,13,14]. Thus, the TOPE operator algebra approach also explains why and
how structural (graph-) matching computations arise naturally in biochemical and multicellular biological
simulation.
3.4
Hybrid SSA/ODE setup
As will be shown in Section 3.4 below, the operator formulation for a system of ordinary differential
equations is [4]:
ÄÅ
ÉÑ
ÅÅ
ij
yzÑÑÑÑ
ÅÅ
`
`
Odrift = - · · d 8x< d 8y< aH8y<L aH8x<LÅÅÅÅ„ “ yi jjjvi H8y<L ‰ dHyk - xk L zzzÑÑÑÑ
(22)
ÅÅ
k
k
{ÑÑÑÖ
ÇÅ i
Here and in the calculations that follow, the Dirac delta function can be considered as a Gaussian with very
small variance, which participates in a limiting process by which, at the end of each calculation, the limit of
zero variance is taken.
In [4] this operator expression is generalized from ordinary differential equations to stochastic
differential equations, for example those pertaining to the diffusion of particles, as equivalently represented by
the Fokker-Planck equation.
3.4.1
Computation of matrix elements
From the commutator
H
L
@aHyL, a` HxLD = dHy - xL I + QHNHxL » nmax L NHxL ,
`
we may calculate matrix elements of Odrift in Equation 22 such as:
12
`
Yw » Odrift
Q-BioPaperMjolsness2012V23TR.nb
ƒƒ
ƒƒ
i
y`
ƒƒ
ƒƒ
j
z
`
ƒ
j
z
» z] = -[8w< ƒƒƒ ‡ d 8x< ‡ d 8y< jj‚ “ yi vi H8y<L dH8y< - 8x<L zz aH8y<L aH8x<L aH8z<L ƒƒƒƒ 0_
ƒƒ
ƒƒ
k i
{
ƒ
ƒ
ƒƒ
ƒƒ
ij
yz
ƒƒ
ƒƒ
= -[8w< ƒƒƒƒ ‡ d 8x< ‡ d 8y< jjj‚ “ yi vi H8y<L dH8y< - 8x<L zzz a` H8y<L dH8x< - 8z<L@I + QHNHxLL NH8x<LD ƒƒƒƒ 0_
ƒƒ
ƒƒ
k i
{
ƒ
ƒ
ij
yz
= - ‡ d 8x< ‡ d 8y< jjj‚ “ yi vi H8y<L dH8y< - 8x<L zzz dH8x< - 8z<L X8w< » 8y<\
k i
{
ij
yz
= - ‡ d 8y< jjj‚ “ yi vi H8y<L dH8y< - 8z<L zzz dH8w< - 8y<L
k i
{
ij
yz
ij
yz
= + ‡ d 8y< dH8y< - 8z<L jjj‚ vi H8y<L “ yi dH8w< - 8y<Lzzz - ‡ d 8y< jjj‚ vi H8y<L dH8y< - 8z<L dH8y< - 8w<Lzzz
k i
{
k i
{
= ‚ vi H8z<L “ zi dH8w< - 8z<L + boundary term H Ø 0 hereL
i
The easiest treatment for the boundary terms is to add the assumptions that boundaries are at infinity in the
space of parameters x, y and z, and that initial conditions place zero probability there, and that finite velocities
vHxL ensure the probability remains zero at infinity at finite times. In that case boundary terms can be
`
`
neglected. Alternatively, we can define Odrift =Odrift - diag(1·Odrift L which in this case subtracts off the
boundary term. Then
ÄÅ
ÉÑ
ÅÅ
ij
yzÑÑÑÑ
ÅÅ
`
Odrift = - · · d 8x< d 8y< HàH8y<L aH8x<L - aH8x<L aH8x<LLÅÅÅÅ„ “ yi jjjvi H8y<L ‰ dHyk - xk L zzzÑÑÑÑ
(23)
ÅÅ
k
k
{ÑÑÖÑ
ÇÅ i
If we define xHtL as a time-varying version of z, satisfying
xi
ÅÅÅÅÅÅÅÅÅÅÅÅÅ = vi H8xk <L,
t
then
xi
d
‚ vi H8z<L “ xi dH8w< - 8xHtL<L = ‚ K ÅÅÅÅÅÅÅÅÅÅÅÅ O K ÅÅÅÅÅÅÅÅÅÅÅÅ O dH8w< - 8xHtL<L = K ÅÅÅÅÅÅÅÅÅ O dH8w< - 8xHtL<L
t
x
d
t
i
i
i
Next we calculate Xw » expHt Odrift L » z\ . To this end, Taylor’s theorem may be written
tn d n
Shiftt Î f HtL = f Ht + tL > ‚ ÅÅÅÅÅÅÅÅÅ K ÅÅÅÅÅÅÅÅÅ O f HtL > e Ht Dt L f HtL
n! d t
n=0
¶
if t is a constant. For small t we have
Xw » expHt Odrift L » x\ = Xw » x\ + t Xw » Odrift » x\ + OIt2 M
d
= K1 + tK ÅÅÅÅÅÅÅÅÅ OO dH8w< - 8xHtL<L + OIt2 M
dt
= Shiftt dH8w< - 8xHtL<L ª dH8w< - 8xHt + tL<L + OIt2 M
For larger t we have
ƒƒ n
ƒƒ
ƒƒ
ƒƒ
t
Xw » expHt Odrift L » x\ = limnض [w ƒƒƒƒ ‰ expJ ÅÅÅÅÅ Odrift N ƒƒƒƒ x_
ƒƒ i=1
n
ƒƒ
ƒ
ƒ
= ‡ ‡ d xn-1 d x1 @Shifttên dH8w< - 8xn-1 <LD @Shifttên dH8x1 < - 8x<LD
13
Q-BioPaperMjolsness2012V23TR.nb
ij n
yz
= limnض jjj ‰ Shifttên zzz dH8w< - 8xHtL<L
k i=1
{
= Shiftt dH8w< - 8xHtL<L ª dH8w< - 8xHt + tL<L
i
i
yy
= djj8w< - jjzHt = 0L + ‡ vi HzHtLL d t zzzz
k
k
{{
0
t
Thus (where “IC” means initial condition)
ij
yz
Xw » expHt O8DE< L » z\ = expjjjt ‚ vi H8z<L “ zi zzz dH8w< - 8z<L
k i
{
i
i
yy
= djj8w< - jj8zH0L = z< + ‡ vi HzHt£ LL d tzzzz
k
k
{{
0
xi
= dK8w< - KSolution of ÅÅÅÅÅÅÅÅÅÅÅÅÅ = vi H8xk <L with IC zH0L = zOO
t
t
(24)
QED.
As far as we know this derivation has not appeared previously. As a corrollary, using Equation 24
we may multiply by f HwL and integrate over w to calculate
i
i
yy
exp Ht v H8z<L ÿ “ z L d Hw - zL = djjw - jjzH0L + ‡ vHzHt£ LL d t £ zzzz
k
k
{{
0
t
i
y
exp Ht v H8z<L ÿ “ z L f HzL = f jjzH0L + ‡ vHzHt £ LL d t £ zz .
k
{
0
t
fl
3.5
(25)
Hybrid SSA/ODE: Operator algebra derivation
We now derive a new SSA-like simulation algorithm for hybrid combinations of discrete events
and ODE dynamics, using operator algebra. The main idea is to replace the exponential distribution factor
t
expH-t D L with expI- Ÿ0 DHt £ L d t £ M , and add an extra ODE to the system of ODEs in order to keep track of the
integral. We will use the more compact formulation of TOPE in Equation 2 to derive this method.
3.5.1
Heisenberg picture
Let the operators, rather than the states, evolve in time according to W0 according to Equation 3.
This is traditionally called the “Heisenberg picture” in distinction to the “Schroedinger picture” in quantum
mechanics. Recall Equation 2 and Equation 3, where HL+ is the time-ordering super-operator :
l OHti L O Ht j L if ti r t j
HOHti L O Ht j L L+ = m
n OHt j L O Hti L if ti b t j
(and likewise for higher order products). Note that if OHti L and OHt j L commute for all pairs of times ti and
t j , then HOHti L O Ht j L L+ = OHti L O Ht j L , and the time-ordering operator HL+ can be dropped.
3.5.2
Application to ODE + decay clock
The hybrid system consisting of chemical reactions (possibly parameterized) together with ordinary
differential equations has the combined operator W = IÒreact - Dreact M + ODE , which we can regroup as
`
W = HODE - Dreact L + Oreact
14
Q-BioPaperMjolsness2012V23TR.nb
and then apply TOPE to ODE - Dreact first with W00 = ODE and W01 Htk L = -Dreact Htk L, and then again to
`
`
HODE - Dreact L + Oreact with W0 = W00 + W01 = ODE - Dreact and W1 = Oreact .
In the first application of TOPE to ODE - Dreact with W00 = ODE , the opererators
W01 Htk L = -Dreact Htk L defined at different times are all diagonal in the same (particle number basis and therefore commute:
@Dreact Hti L, Dreact Ht j LD = 0 .
In this circumstance, we can simply drop the time-ordering super-operator HL+ in Equation 2 and write
i
y
expHt HODE - Dreact LL = expHt ODE L expjj- ‡ d t £ Dreact Ht£ Lzz
k 0
{
t
(26)
where, as in Equation 3, Dreact Ht £ L = expH-t £ ODE L Dreact expHt£ ODE L . In our case, (Equation 26) specializes
to :
t
ij
yz
i
y
Xw » expHt HO8DE< - Dreact LL » z\ = expjjjt ‚ vi H8z<L “ zi zzz expjj- ‡ d t £ Dreact Ht£ Lzz dH8w< - 8z<L
k 0
{
k i
{
t
t
t
i
i
yy i
i
yy
= expjjj- ‡ d t £ Dreact jjjzH0L + ‡ vH8z<L d t zzzzzz djjw - jjzH0L + ‡ vHzHt £ LL d t£ zzzz
k
k
{{
0
0
k 0
k
{{
£
(27)
This result looks very similar to Equation 25 applied to f HzL = expI- Ÿ0 d t £ Dreact Ht£ LM dH8w< - 8z<L, and we
now aim to understand and exploit this similarity.
t
3.5.3
Equivalent ODE
Consider the dynamics expressed in Equation 27. Can we obtain the first factor from ODE’s alone?
Yes, if we introduce a new state variable t involved in every ODE-related rule. Set tH0L = 0 as the new
variable’s initial condition, and augment the ODE operators as follows
è
O8DE<
Z = Hz, tL
VHzL = HvH8z<L, -DHzLL
“ Z = H“ z , t L
= VHZL “ Z = vH8z<L ÿ “ z +DHzL t
(28)
In other words, we have added a differential equation for t to the ODE system
xi
dt
ÅÅÅÅÅÅÅÅÅÅÅÅÅ = vi H8xk <L and ÅÅÅÅÅÅÅÅÅÅ = DHzL .
t
dt
This equation is solvable in terms of a “warped time” coordinate
tHtL = ‡ Dreact HzHt £ LL d t£ .
t
0
There are degenerate cases Dreact = 0 only if there are terminal states in the reaction network.
To see that this is the correct procedure, calculate from Equation 24:
15
(29)
[K
Q-BioPaperMjolsness2012V23TR.nb
w
tmax
è
O Ã expIt O8DE< M expH-tmax L Ã K
zH0L
O_
tH0L = 0
ij ij
yzyz
= expjjjt jjj ‚ vi H8z<L “ zi +D HzL t zzzzzz dH8w< - 8z<L dHtmax - tL expH-tmax L
kk i
{{
i
i
yy i
y
= djj8w< - jjzH0L + ‡ vi HzHt £ LL d t zzzz djjtmax - ‡ Dreact HzHt £ LL d t £ zz expH-tmax L
k
k
{{ k
{
0
0
t
(30)
t
= dI8w< - ISolution of Equation 29 , with IC zH0L, tH0L = 0MM µ expH-tmax L
This expression agrees with Equation 27, as required. But how do we insure the IC on t ? That can be done
as follows:
[K
z
è
O Ã expIt O8DE< M expH-tmax L Ã J N_ =
tmax
0
w
z
z£
z
è
[K
O Ã expIt O8DE< M expH-tmax L Ã J N_ µ K1 = ‡ d t£ d z£ [K £ O Ã J N_O
tmax
0
t
t
£
£
w
z
z
z
è
= [K
O Ã expIt O8DE< M expH-tmax L K‡ d t£ d z£ Ã K O_ [K £ O ÃO Ã J N_
tmax
0
t
t
w
z
è
= [K
O Ã expIt O8DE< M expH-tmax L Pt:=0 Ã J N_
tmax
t
where
w
Pt:=0 ª ‡ d t£ d z£ Ã K
(31)
z£
z£
O_ [K £ O Ã
0
t
is a projection operator (i.e. one that satisfies P ÿ P = P) that resets the variable t to zero after each use. In
summary,
[K
z
è
O Ã expIt O8DE< M expH-tmax L Pt:=0 Ã J N_ =
tmax
t
w
dI8w< - ISolution of Equation 29 , with IC zH0L, tH0L = 0MM µ expH-tmax L
(32)
Clearly this result is equivalent to Equation 27 and is in the correct form for a Markov chain that can
represent a computation. Of course, the matrix element calculated is only relevant if tmax as drawn from the
w
exponential is constrained to be equal to the final value of t in the final state YI tmax
M … as solved by the ODE
è
system O8DE< . We can implement this constraint with a factor of dHt - tmax L in the Markov chain over states
and times. Thus a step in the Markov chain in between reactions can be written as:
è
"between reactions = dHt - tmax L expIt O8DE< M Pt:=0 expH-tmax L QHtmax r 0L
`
" = Oreact ÿ "between reactions
`
As in the SSA derivation, the reaction step is given by factors of Oreact which need to be normalized by
Dreact . Using dHt - tmax Htmax LL d t = dHt - tmax L d t and d t ê d t = Dreact HtL, we find
M01 = expH-tmax L QHtmax r 0L
è
M00 = dHt - tmax Htmax LL expIt O8DE< M ÿ Pt:=0
`
M1 = Oreact ê Dreact
" = M1 ÿ M00 ÿ M01
where " represents the Markov chain corresponding to the simulation algorithm.
In implementations so far [14] we have used instead the equivalent differential operator
è
O8DE< = VHZL “ Z = vH8z<L ÿ “ z -DHzL p p
with p = expH-tL, initialized at p0 = 1 , and a uniform distribution on pfinal œ @0, 1D .
16
(33)
Q-BioPaperMjolsness2012V23TR.nb
3.5.4
Algorithm: Hybrid SSA/ODE solver
By Equation 33 above, a Markov chain algorithm for simulating the hybrid system can be represented in the
following SSA-like pseudocode:
factor rHrL Hxin , xout L = kHrL Hxin L pHrL Hxout » xin L ;
initialize SSA propensities as kHrL Hxin L;
repeat {
initialize kHtotalL := ⁄r kHrL Hxin L;
initialize t := 0 ;
draw effective waiting time tmax from expH-tmax L
dt
ÅÅÅÅ
ÅÅÅÅ = kHtotalL HtL
dt
solve ODE system, including an extra ODE updating t:
draw reaction r from distribution kHrL Hxin L ê kHtotalL ;
until t = tmax
draw xout from pHrL Hxout » xin L and execute reaction r;
} until t r tmax
3.5.5
Application: Cell division
As a simplified model of stochastic cell division, we may consider constant growth of a linear dimension l
of each cell, d l ê d t = v, coupled with a stochastic cell division rule whose propensity depends on the ratio of
l to a threshold length l0 for likely division:
cellHlL Ø cellHl ê 2L, cellHl ê 2L
with
rdivision sHb Hl ê l0 - 1LL
with a sigmoidal function such as sHxL = 1 ê H1 + expH-xLL. In this model the parameter b varies the sharpness of the threshold, and rdivision is the maximal propensity for division. Experimental evidence for stochastic dependence of division events on cell size in plant cells is reviewed in [15].
The differential equation for length can also be put in the form of a reaction rule that includes an
ODE:
cellHlL Ø cellHl + d l L solving d l ê d t = vHlL
as described in [14]. Clearly this model could be augmented with other parameters such as growth signals
with their own dynamics. This was done in models of biological stem cell niches in mouse olfactory epithelium and plant root growth, using the foregoing cell division rules. These systems were studied and simulated
using the hybrid SSA/ODE algorithm above, in [14,16].
3.5.6
Application: Time-varying propensity for complete polymerization
Consider the n-step polymerization reaction
8 A Ø X1 with k1 , X1 Ø X2 with k2 , ... , Xn-1 Ø B with kn <
There is an nHmaxL .
ti = t ê n and
ki = n k.
`
W = l a` n+1
17
Q-BioPaperMjolsness2012V23TR.nb
`
W = lHHa` n+1 + cn+1 L - In+1 L = lIbn+1 - In+1 M
`
where cn+1 is all zeros except for a “1” entry in the lower right corner. Since b and I are matrices that
`
commute, expHt WL = expIt l bn+1 M expH-t l In+1 L and we easily compute
ln t n-1 e-l t
PHt » t, nL = @expHt WLD1, n+1 = ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
Hn - 1L !
This is the distribution on polymer completion times. It is an Erlang distribution (a Gamma distribution
with integral values of n). If t is held fixed and n tends towards infinity, this distribution approaches a delta
function dHt - tL, which can lead to differential-delay equation models for reaction networks involving
polymerization processes such as transcription [17]. This probability distribution for termination times also
corresponds to the time-varying propensity function
t
ÅÄÅ
ÑÉÑ ln t n-1 e-l t
rHt » t, nL = PHt » t, nL ì ÅÅÅÅ1 - ‡ PHt » t, nL d t ÑÑÑÑ = ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ ,
(34)
ÅÇ
ÑÖ
GHn, t lL
0
which increases monotonically in time.
As in Equation 18, the resulting time-varying propensity still fits within the framework of a Markov
chain "HI, t£ » J, tL that advances the time variable by an increment that is a random variable. The method of
the previous section can be used to implement an SSA-like algorithm, with differential equations that govern
propensities replaced by algebraic equations (Equation 34) or, if differential equations are also present, by
differential-algebraic equations.
Figure 2. Erlang-derived time-dependent propensities for completion of a multistage process
t = 1, n œ 81, ..., 10< . Horizontal axis: time, t . Vertical axis: propensity, rHt » t, nL . Plots for varying n are
superimposed. For larger n there is a “maturation” phenomenon whereby completion at small times is very
unlikely, and when a process is “overdue” for completion then its propensity becomes very high. By comparison, propensities for very small n increase rapidly at first and are then relatively flat.
18
Q-BioPaperMjolsness2012V23TR.nb
4
Conclusion and outlook
We have shown that the time-ordered product expansion (TOPE) can be used systematically to
derive computational simulation and parameter-fitting algorithms for stochastic systems, connecting two
seemingly distant areas of research. In doing so we have developed the means to translate formally between
field theory language and the language of computable Markov chains in which randomized algorithms can be
expressed and derived. By this means we hope to open the door to the use of TOPE and related methods from
quantum and statistical field theory in the computational simulation of stochastic biochemical kinetics, with
broad applicability in physically based biology. The particular hybrid stochastic process/ordinary differential
equation simulation algorithm derived here is very different from interleaving and operator splitting algorithms which are intrinsically approximate; instead, this algorithm is exact in the same sense that SSA is (that
is, it draws from the same distribution of just-fired reactions), except for errors introduced by the ODE solver
and in the solver’s detection of the ODE stopping criterion, which is that some variable reaches a threshold
value. A future prospect for the field theory approach is for application to reaction-diffusion systems in which
the propagator for particles between reactions is the heat kernal Green’s function for the diffusion equation.
The result may be an alternative avenue for derivation of novel particle-based, off-grid stochastic numerical
solvers for reaction-diffusion problems as treated in [2], which, like the algorithms shown here, are also
amenable to generalizations to exact “leaping” acceleration and to hybrid stochastic/differential equation
solution algorithms.
5
Acknowledgements
Research was supported by NIH grants R01 GM086883 and P50 GM76516 to UC Irvine. I also
wish to acknowledge the hospitality, travel support, and research environments provided by the Center for
Nonlinear Studies (CNLS) at the Los Alamos National Laboratory, the Sainsbury Laboratory Cambridge
University, and the Pauli Center for Theoretical Studies at ETH Zürich and the University of Zürich.
BACKMATTER
A
References
[1] Doi, J. (1976) Second quantization representation for classical many-particle system. 1976 J. Phys. A:
Math. Gen. 9 1465 .
[2] Doi, J. (1976) Stochastic theory of diffusion-controlled reaction 1976 J. Phys. A: Math. Gen. 9 1479 .
[3] Mattis D.C. and Glasser M.L. (1998) The uses of quantum field theory in diffusion-limited reactions.
Rev.Mod. Phys. 70, 979–1001
[4] Mjolsness E. and Yosiphon G (2006) Stochastic Process Semantics for Dynamical Grammars. Annals
of Mathematics and Artificial Intelligence, 47(3-4)
[5] H. M. Fried (2002), Green’s Functions and Ordered Exponentials, Cambridge University Press
[6] C. M. Bender, S. F. Brandt, J.-H. Chen, and Q. Wang (2005), "Ghost Busting: PT-Symmetric Interpretation of the Lee Model". Physical Review D 71, 025014
[7] Xueying Zhang, Katrien De Cock, Mónica F. Bugallo, and Petar M. Djuri# (2005), A general method
for the computation of probabilities in systems of first order chemical reactions. J. Chem. Phys. 122, 104101
(2005).
19
Q-BioPaperMjolsness2012V23TR.nb
[8] Yosiphon G. and Mjolsness E. (2010) Towards the Inference of Stochastic Biochemical Network and
Parameterized Grammar Models. In Learning and Inference in Computational Systems Biology, (Lawrence
N., Girolami M., Rattray M., and Sanguinetti G., Eds.) MIT Press.
[9] Darren J. Wilkinson (2006). Stochastic Modelling for Systems Biology. Chapman & Hall/CRC Press,
Boca Raton, Florida.
[10] E. Mjolsness, D. Orendorff, P. Chatelain, P. Koumoutsakos (2009), “An Exact Accelerated Stochastic
Simulation Algorithm”, Journal of Chemical Physics 130, 144110.
[11] Orendorff, David (2012). “Exact and Hierarchical Reaction Leaping: Asymptotic Improvements to the
Stochastic Simulation Algorithm”. PhD thesis, UC Irvine Computer Science Department, June 2012. Thesis
available at: http://computableplant.ics.uci.edu/~dorendorff/thesis .
[12] Hlavacek WS, Faeder JR, Blinov ML, Posner RG, Hucka M, Fontana W (2006) Rules for modeling
signal-transduction systems. Science’s STKE 2006:re6.
[13] Danos V, Feret J, Fontana W, Harmer R, Krivine J (2007) Rule-based modelling of cellular signaling.
Lect Notes Comput Sci 4703:17-41.
[14] Yosiphon, G. (2009), “Stochastic Parameterized Grammars: Formalization, Inference, and Modeling
Applications”, PhD Thesis, UC Irvine Computer Science Department, June 2009. Thesis and software :
http://computableplant.ics.uci.edu/~guy/Plenum.html .
[15] Roeder, A. H. K. (2012), When and where plant cells divide: a perspective from computational
modeling. Current Opinion in Plant Biology 2012, 15:1–7 .
[16] V.V. Mironova, Nadya A Omelyanchuk, Guy Yosiphon, Stanislav I Fadeev, Nikolai A Kolchanov,
Eric Mjolsness and Vitaly A Likhoshvai (2010). “A plausible mechanism for auxin patterning along the
developing root”. BMC Systems Biology 4:98.
[17] Likhoshvai, V. A.; Demidenko, G. V.; Fadeev, S. I. (2006), Modeling of Gene Expression by the
Delay Equation. Bioinformatics of Genome Regulation and Structure II (2006): Part 3, 421-431, DOI:
10.1007/0-387-29455-4_40 .
[18] Wang Y, Christley S., Mjolsness E., and Xie X. (2010) Parameter inference for discretely observed
stochastic kinetic models using stochastic gradient descent. BMC Systems Biology 4:99.
6
Appendix: Maximum likelihood parameter inference
Application of the TOPE to maximum-likelihood parameter learning in stochastic reaction networks has previously been presented [16]. Here, for completeness of presentation for a different audience, we
just show the essential gradient calculation step.
Suppose we have observations of the state of a chemical reaction network at times 8ts <, and wish to
improve the probability PHData » ModelL of a reaction network model for the flow of probability at intermediate times. We will use the TOPE for each time interval in between observation times ts :
Be
è
Hts+1 -ts L W
F HxHts+1 L, xHts LL = „ ‡
¶
k=0
è
tn D
äBe
`
W ... e
è
t1 D
ts+1
ts
`
We
‡
ts
è
t0 D
ts+1
n
ij
yz
d @tDn0 djjjjHts+1 - ts L - ‚ t p zzzz
p=0
k
{
F Hx Hts+1 L, x Hts LL
We will need to compute the derivatives of this probability with respect to reaction rates:
20
(35)
Q-BioPaperMjolsness2012V23TR.nb
ts+1
ts+1
rr ÅÅÅÅÅÅÅÅÅÅÅÅÅÅ AeHts+1 -ts L W E HxHts+1 L, xHts LL = „ ‡
‡
d @tDn0
rr
ts
ts
¶
k=0
n
ij
yz
djjjHts+1 - ts L - ‚ t p zzzz
j
p=0
k
{
`
`
`
ä ‚ Ae-tn D W ... e-t p+1 D Irr W r M e-t p D ... e-t1 D W e-t0 D E HxHts+1 L, xHts LL
n
p=0
-„ ‡
¶
k=0
ts
ts+1
‡
ts+1
d @tDn0
ts
n
ij
yz
j
j
djjHts+1 - ts L - ‚ t p zzzz
p=0
k
{
`
`
`
`
ä ‚ Ae-tn D W ... W Hrr t p Dr L e-t p D W ... e-t1 D W e-t0 D E HxHts+1 L, xHts LL
n
p=0
ij r @Ẁ D
yz
j
r
r IJ
z
rr @Ẁ r D I J = jjjj ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅ zzzz I@ẀDI J M = br I J @ẀDI J
j ‚ rr @Ẁ r D z
IJ {
k
r
where we defined the “branching ratio”
jij rr @Ẁ r DI J zyz
br I J ª jjjj ÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅÅ
ÅÅÅÅÅÅÅÅÅÅ zzz = Xdr,RHI,JL \ pHI»JL
j ‚ rr @Ẁ r D zz
IJ {
k
r
for reaction r in state J , assuming each reaction r results in just one output state I per input state J . Here
RHI, JL is the random variable denoting the actual reaction chosen in transitioning from state J to state I . Then
n
ts+1
ts+1
ij
yz
è
rr ÅÅÅÅÅÅÅÅÅÅÅÅÅÅ BeHts+1 -ts L W F HxHts+1 L, xHts LL = „ ‡
‡
d @tDn0 djjjjHts+1 - ts L - ‚ t p zzzz
rr
ts
ts
p=0
k
{
k=0
¶
`
`
`
`
ä ‚ Aetn D W ... et p+1 D Ibr W - W rr t p Dr M et p D ... et1 D W et0 D E HxHts+1 L, xHts LL
n
p=0
or
è
rr ÅÅÅÅÅÅÅÅÅÅÅÅÅÅ BeHts+1 -ts L W F HxHts+1 L, xHts LL =
rr
‚ ‚ Xbr Hreaction event p out of nL\W` , xHts+1 L,xHts L - rr ‚ ‚ Xt p Dr \W` , xHt
¶
¶
n
k=0 p=0
n
k=0 p=0
s+1 L,xHts L
(36)
This finally is a quantity that is easy to compute as a running average during a simulation of the
network with incorrect values of the parameters, thereby contributing to the calculation of an improved set of
parameter values in a stochastic gradient descent algorithm. This is the key update equation in a learning
algorithm for reaction rates in stochastic biochemical networks (extensible to other process networks).
Algorithmic details can be found in [18], noting particularly Equation 2.4 therein. A related stochastic learning algorithm is proposed in [8].
21
| 5cs.CE
|
arXiv:1710.04321v1 [math.GR] 11 Oct 2017
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE
DISCRETELY VALUED FIELDS
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
A BSTRACT. Let K be a complete discretely valued field with residue field k with char(k) 6=
2. Assuming that the norm principle holds for extended Clifford groups Ω(q) for every
even dimensional non-degenerate quadratic form q defined over any finite extension of k,
we show that it holds for extended Clifford groups Ω(Q) for every even dimensional nondegenerate quadratic form Q defined over K.
1. I NTRODUCTION
Let K be a field and T , a commutative linear algebraic group defined over K. Given
L/K, a finite separable field extension,
L/K :
Q one can define the norm homomorphism Nsep
T (L) → T (K) which sends t ; γ γ(t) where γ runs over cosets of Gal (K /L)
in Gal (K sep /K). The definition of the norm homomorphism can be extended to K-étale
algebras in a similar manner. Note that if T = Gm , then NL/K : T (L) → T (K) is precisely
the usual norm NL/K : L∗ → K ∗ .
Now let G be a linear algebraic group defined over K and let f : G → T be an algebraic
group homomorphism defined over K. Consider the following diagram:
G(L)
f (L)
T (L)
NL/K
G(K)
f (K)
T (K)
We say that the norm principle holds for f : G → T over a finite separable field extension
(or étale algebra) L/K if NL/K (Im f (L)) ⊆ Im f (K). We say that the norm principle
holds for f : G → T if for every finite separable field extension (equivalently for every
étale algebra) L/K , NL/K (Im f (L)) ⊆ Im f (K).
Suppose further that the commutator subgroup G0 of G is defined over K. Then every
homomorphism f : G → T factors through the natural homomorphism f˜ : G → G/G0
The second author was partially supported by the Canada Research Chairs Program and an NSERC research grant. The work of the third author has been supported by the NSF grant DMS #1160206.
1
2
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
and it is an easy check that the norm principle for f˜ (over L/K) implies the norm principle
for f (over L/K). We say that the norm principle holds for G (over L/K) if it holds for f˜
(over L/K).
Let Q be a quadratic form over K. The classical norm principle of Scharlau which asserts
that norms of similarity factors of QL are themselves similarity factors of Q can be restated
in this context to say that the norm principle holds for the multiplier map M : GO(Q) →
Gm . Similarly Knebusch’s norm principle which states that norms of spinor norms of QL
are spinor norms of Q can be reformulated as the norm principle holding for the spinor
norm map µ : Γ+ (Q) → Gm .
Norm principles have been previously studied in ([Gi93], [Me96]), especially in conjunction with the rationality or the R-triviality of the algebraic group in question. In ([BM00]),
it was established that the norm principle holds in general for all reductive groups of classical type without Dn components. The Dn case was investigated in ([Bh16]) and a scalar
obstruction defined up to spinor norms, whose vanishing would imply the norm principle,
was given. However, the triviality of this scalar obstruction is far from clear and the question whether the norm principle holds for reductive groups with type Dn components still
remains open.
If K is a number field, the norm principle was proved in full generality for all reductive
groups by P. Gille ([Gi97]), so the first widely open and very interesting case is when K
is the function field k(C) of a curve C defined over a number field k and the group G in
question is of classical type with the semisimple part G0 = Spin(Q). As we show in the
last section of the paper, the validity of the norm principle Q
over K is closely related to
the triviality of the kernel of the natural map H1 (K, G0 ) → v H1 (Kv , G0 ) where v runs
through a set of discrete valuations of K. Therefore the (traditional) local-global approach
leads us first to look in detail over completions Kv .
With this motivation in mind, in this paper, we investigate the Dn case over an arbitrary
complete discretely valued field K with residue field k and char(k) 6= 2, restricting ourselves to type Dn groups arising from quadratic forms. In the main result of the paper, we
show that if the norm principle holds for such groups defined over all finite extensions of the
residue field k, then it holds for such groups defined over K (c.f. Theorem 5.1). This yields
examples of complete discretely valued fields with residue fields of virtual cohomological
dimension ≤ 2 over which the norm principle holds for the groups under consideration
(c.f. Corollary 6.3). As a further application, we also relate the possible failure of the norm
principle to the nontriviality of certain Tate-Shaferevich sets (Section 7).
Notations. Let K be a field of characteristic not 2 and (V, Q), a non-degenerate quadratic
space of dimension 2n over K where n ∈ Z>0 . Let Z/K denote the discriminant extension
of Q and let ψ denote the non-trivial K-automorphism of Z.
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS
3
Let µ denote the center of Spin(Q).
Recall that µ = RZ/K (µ2 ) when n is even and µ =
Norm
µ4[Z] := Ker RZ/K µ4 −−−→ µ4 when n is odd. Let Ω(Q) be the extended Clifford group1
of Q. If dim Q = 2n ≥ 4, this reductive group has center RZ/K Gm and is an envelope
of Spin(Q) ([BM00], Ex 4.4). Further it is an extension of the projective group of proper
similitudes PGO+ (Q) by RZ/K Gm :
χ0
1 → RZ/K Gm → Ω(Q) −
→ PGO+ (Q) → 1
2. R EDUCTIONS
2.1. Another formulation of the norm principle. Let 1 → J → G1 → G2 → 1 be
a central K-isogeny of reductive groups G1 , G2 /K. We recall another formulation of the
norm principle for the connecting map δ : G2 (−) → H1 (−, J) and its relation to the norm
principle of an associated map f : G → T . This discussion is taken from ([Me96], 3.10).
Since J is commutative, one can define norm maps (correstriction) NL/K : H1 (L, J) →
H1 (K, J) for finite separable extensions L/K. Consider the following diagram:
G2 (L)
δ(L)
H1 (L, J)
NL/K
G2 (K)
δ(K)
H1 (K, J)
We say that the norm principle holds for δ : G2 (−) → H1 (−, J) over a finite separable
field extension (or étale algebra) L/K if NL/K (Im δ(L)) ⊆ Im δ(K). We say that the
norm principle holds for δ if the norm principle holds for δ over every finite separable field
extension (equivalently over every étale algebra) L/K.
Let T 0 be a quasi-trivial torus2 containing J. The associated map f : G → T (where G/K
is reductive and T is commutative) is determined by the following commutative diagram
(Figure 1) with exact rows and columns:
1
If dim Q = 2, we set Ω(Q) to be the commutative group RZ/K Gm , for which the norm principle holds.
Product of Weil restrictions of split tori.
2
4
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
1
1
1
J
G1
G2
1
id
T0
1
G
f0
G2
1
f
id
T
h
1
T
1
F IGURE 1. Restating the norm principle
Lemma 2.1. Let L/K be a finite separable extension and let G1 , G2 , J and f : G → T be
as above. Then the norm principle holds for f : G → T over L/K if and only if it holds
for δ : G2 (−) → H1 (−, J) over L/K.
h
Proof. Let 1 → T 0 → G →
− G2 → 1 be the exact row as in the diagram above. Since T 0
is a quasi-trivial torus, this exact sequence induces surjective maps h(L) : G(L) → G2 (L)
for all L/K. Let δ1 : T (−) → H1 (−, J) be the connecting map of the first exact column in
the diagram above. By ([Me96], Lemma 3.11), the following square is anticommutative.
g ∈ G(L)
h(L)
δ(L)
f (L)
t ∈ T (L)
g2 ∈ G2 (L)
δ1 (L)
j ∈ H1 (L, J)
Assume that the norm principle holds for f over L/K. Let g2 ∈ G2 (L) and δ(L)(g2 ) = j.
Since h(L) is surjective, pick g ∈ G(L) such that h(L)(g) = g2 and let t = f (L)(g) ∈
T (L). Thus δ1 (L)(t) = j −1 and δ1 (K)(NL/K (t)) = NL/K (j −1 ). Since the norm principle
holds for f , NL/K (t) = f (K)(g̃) for some g̃ ∈ G(K). Then NL/K (j) = δ(K)[h(K)(g̃)].
Thus NL/K (j) ∈ Im δ(K) and hence the norm principle holds for δ over L/K.
Conversely, let the norm principle hold for δ over L/K. Let g ∈ G(L) and f (L)(g) = t. Set
h(L)(g) = g2 ∈ G2 (L) and δ(L)(g2 ) = j. Thus δ1 (L)(t) = j −1 and δ1 (K)(NL/K (t)) =
NL/K (j −1 ) as before. Since the norm principle holds for δ, NL/K (j) = δ(K)(g˜2 ) for some
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS
5
g˜2 ∈ G2 (K). Since h(K) is surjective, pick g̃ ∈ G(K) such that h(K)(g̃) = g˜2 . Then
NL/K (t) = f (K)(g̃)f 0 (K)(t0 ) for some t0 ∈ T 0 (K). Thus NL/K (t) ∈ Im f (K) and hence
the norm principle holds for f over L/K.
Let G be a reductive group defined over K whose simple components are of classical type
and let T be a commutative group defined K. Then (Thm 1.1, [BM00]) establishes that if
the Dynkin diagram of G does not contain connected components Dn for n ≥ 4, the norm
principle holds for any group homomorphism G → T . We would like to investigate the
norm principle for G → T in the remaining case when G has simple components of type
Dn , under the further simplifying assumption that these simple components have simply
connected covers Spin(Qi ) arising from quadratic forms Qi over K.
By following the reductions in ([BM00]), it is easy to see that we need only to check
whether the norm principle holds for the group Ω(Q), i.e. for the canonical map Ω(Q) →
Ω(Q)
for any non-degenerate even dimensional quadratic form Q/K. If dim Q = 2,
[Ω(Q),Ω(Q)]
as noted before, the norm principle holds for the commutative group Ω(Q).
2.2. Maps S and α. We now restate the norm principle for Ω(Q) when dim Q = 2n ≥ 4
in two other equivalent forms, which will be used in the rest of the paper.
Ω(Q)
Let T denote [Ω(Q),Ω(Q)]
and let T 0 denote the quasi-trivial torus RZ/K Gm containing µ.
Using the fact that the semisimple part of Ω(Q) is Spin(Q), Figure 1 above yields the
following commutative diagram:
1
1
1
µ
Spin(Q)
PGO+ (Q)
1
id
1
RZ/K Gm
Ω(Q)
f0
T
1
f
id
T
1
h
PGO+ (Q)
1
6
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
Let S : PGO+ (Q)(−) → H1 (−, µ) denote the connecting map of the first exact row. By
Lemma 2.1, the norm principle for Ω(Q) → T over L/K holds if and only if it holds for S
over L/K.
For any L/K, the short exact sequence 1 → µ → Spin(Q) → PGO+ (Q) → 1 gives rise
to the long exact sequence
S(L)
α(L)
. . . → PGO+ (Q)(L) −−→ H1 (L, µ) −−→ H1 (L, Spin(Q)) → . . . .
Hence we can deduce that the norm principle holds for S over L/K if and only if the
following, which we call the norm principle for α : H1 (−, µ) → H1 (−, Spin(Q)) over
L/K, holds:
α(L)
1
1
For every u ∈ Ker H (L, µ) −−→ H (L, Spin(Q)) , the element NL/K (u) belongs to
α(K)
1
1
Ker H (K, µ) −−−→ H (K, Spin(Q)) .
Thus the above discussion gives the following:
Lemma 2.2. Let L/K be a finite separable field extension. Then the following are equivalent
(1) The norm principle holds for Ω(Q) over L/K
(2) The norm principle holds for S : PGO+ (Q)(−) → H1 (−, µ) over L/K
(3) The norm principle holds for α : H1 (−, µ) → H1 (−, Spin(Q)) over L/K
2.3. Auxiliary maps i and j. We recall the (explicit) definitions of useful auxiliary maps
i : H1 (−, µ2 ) → H1 (−, µ) and j : H1 (−, µ) → H1 (−, µ2 ).
Recall the following commutative diagram with the two complete rows and columns exact:
1
µ2
1
µ
1
id
µ2
Spin(Q)
PGO+ (Q)
1
id
1
µ2
O+ (Q)
1
1
PGO+ (Q)
1
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS
7
For every L/K, the long exact sequence of cohomology gives rise to maps i(L) : H1 (L, µ2 ) →
H1 (L, µ) and j(L) : H1 (L, µ) → H1 (L, µ2 ) which fit into the diagram
H1 (L, µ2 )
id
H1 (L, µ2 )
i0 (L)
i(L)
PGO+ (Q)(L)
S(L)
H1 (L, µ)
α(L)
H1 (L, Spin(Q))
j(L)
id
PGO+ (Q)(L)
M(L)
H1 (L, µ2 )
H1 (L, O+ (Q))
F IGURE 2. Auxiliary maps i and j
Here M(L) : PGO+ (Q)(L) → H1 (L, µ2 ) is the multiplier map. The natural map i0 (L) :
H1 (L, µ2 ) → H1 (L, Spin(Q)) in the above diagram will also be used later.
Let Sn(L) : O+ (Q)(L) → H1 (L, µ2 ) denote the spinor norm map. The maps i and j fit
into the following commutative diagram with exact columns ([KMRT98], Prop 13.33 &
13.36).
O+ (Q)(L)
Sn(L)
H1 (L, µ2 )
i(L)
π(L)
PGO+ (Q)(L)
S(L)
H1 (L, µ)
M(L)
j(L)
H1 (L, µ2 )
H1 (L, µ2 )
=
F IGURE 3. Relation to spinor norms
The explicit descriptions of i and j depend on the parity of n = dim(Q)/2.
∗
∗
∗
- If n is even, H1 (K, µ) = ZZ∗2 . Then i(K) : KK∗2 → ZZ∗2 is the inclusion map and
∗
∗
j(K) : ZZ∗2 → KK∗2 sends [z] ; [NZ/K (z)] for z ∈ Z ∗ .
- If n is odd, H1 (K, µ) = UU0(K)
where U ⊂ Gm × RZ/K Gm is the subgroup defined
(K)
∗
∗ 4
by U (K) = {(f, z) ∈ K × Z |f = NZ/K (z)} and U0 ⊂ U ⊂ Gm × RZ/K Gm is
the algebraic subgroup defined by U0 (K) = {(NZ/K (z), z 4 ) ∈ K ∗ × Z ∗ |z ∈ Z ∗ }.
8
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
K∗
is the map sending f K ∗2 ; [f, f 2 ] for f ∈
→ UU0(K)
K ∗2
(K)
∗
U (K)
→ KK∗2 sends [f, z] ; NZ/K (z0 )K ∗2 where z0 ∈ Z ∗
U0 (K)
−1
−2
Then i(K) :
K ∗ . Finally
j(K) :
z0 ψ(z0 )
is such that
=f
z.
We end this section with one more useful reduction, which is a direct consequence of
Knebusch’s norm principle.
Lemma 2.3. Let L/K be a finite separable extension and u ∈ Ker(α(L)). If j(L)(u) =
1 ∈ H1 (L, µ2 ), then NL/K (u) ∈ Ker(α(K)).
Proof. Since u ∈ Ker (α(L)), there exists [g] ∈ PGO+ (Q)(L) such that S(L)([g]) =
u. Since j(L)(u) = 1, by Figure 3 above, we see that u is the image of a spinor norm
of QL , i.e u = i(L)(Sn(h)) for some h ∈ O+ (Q)(L). By Knebusch’s norm
principle,
+
NL/K (Sn(h)) = Sn(h̃) for some h̃ ∈ O (Q)(K). Hence NL/K (u) = i(K) Sn(h̃) =
S(K)(π(h̃)). Therefore α(K)(NL/K (u)) = 1.
2.4. A square diagram. Using the auxiliary maps i and j just defined, we construct a
square diagram (SQ) by piecing together the two classical norm principles of Scharlau
and Knebusch. It is shown that the norm principle holds for Ω(Q) precisely when the
square is commutative for all finite separable extensions. This reformulation yields further
reductions for the norm principle question.
Using Figure 2 in Section 2.3, for each finite separable extension L/K, define the subgroup
H(L) ⊆ H1 (L, µ) as follows:
H(L) := {x ∈ H1 (L, µ) | j(L)(x) ∈ Im (M(L))}
= {x ∈ H1 (L, µ) | α(L)(x) ∈ Im (i0 (L))}.
Recall that i0 (L) : H1 (L, µ2 ) → H1 (L, Spin(Q)) induces the group homomorphism i00 (L) :
L∗ /L∗2 → L∗ / Sn(QL ) ⊆ H1 (L, Spin(Q)) sending f L∗2 ; [f ] for each f ∈ L∗ . Thus
α(L) induces the following map
α̃(L) : H(L) → L∗ / Sn(QL ).
Lemma 2.4. α̃(L) : H(L) → L∗ / Sn(QL ) is a group homomorphism.
Proof. For i = 1, 2, let zi ∈ H(L) with α̃(L)(zi ) = xi ∈ L∗ / Sn(QL ). By the definition of
H(L) and a diagram chase of Figure 2 of Section 2.3, there exist [gi ] ∈ PGO+ (L) such that
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS
9
S([gi−1 ])zi = i(L)(yi ) for yi ∈ H1 (L, µ2 ). Thus α̃(L)(S([gi−1 ])zi ) = i0 (L)(yi ) = i00 (L)(yi ).
Since i00 (L) is a group homomorphism and H(L) is abelian, we see that
α̃(L)(S([g1−1 ])z1 )α̃(L)(S([g2−1 ])z2 ) = i00 (L)(y1 )i00 (L)(y2 )
= i00 (L)(y1 y2 )
= α̃(L)(S([g1−1 ])z1 S([g2−1 ])z2 )
= α̃(L)(S([g2−1 g1−1 ])z1 z2 ).
Since µ = Z(Spin(Q)), we have α̃(L)(S([gi−1 ]zi ) = α̃(L)(zi ) and α̃(L)(S([g2−1 g1−1 ]z1 z2 ) =
α̃(L)(z1 z2 ) ([KMRT98], Corollary 28.4, Pg 386), we conclude α̃(L) is a group homomorphism.
From the definition of H(L), it follows that Im (i(L)) ⊆ H(L). Further by a chase of the
following square which is part of Figure 2, we see that α̃(L)(i(L)[f L∗2 ]) = [f ] for each
f ∈ L∗ .
id
f L∗2 ∈ H1 (L, µ2 )
f L∗2 ∈ H1 (L, µ2 )
i0 (L)
i(L)
H1 (L, µ)
α(L)
[f ] ∈ H1 (L, Spin(Q))
Similarly it is immediate to see that Im (S(L)) = Ker (α(L)) ⊆ H(L) and further is
exactly Ker (α̃(L)).
Scharlau’s norm principle implies that the norm map NL/K : H1 (L, µ) → H1 (K, µ) induces
the map NL/K : H(L) → H(K). Similarly Knebusch’s norm principle implies that the
norm map NL/K : L∗ → K ∗ induces the map NL/K : L∗ / Sn(QL ) → K ∗ / Sn(Q). Thus,
the following square diagram (labelled SQ for L/K) is defined:
H(L)
α̃(L)
L∗ / Sn(QL )
NL/K (Scharlau)
H(K)
α̃(K)
NL/K (Knebusch)
K ∗ / Sn(Q)
Theorem 2.5. Let L/K be a finite separable field extension. Then the following are equivalent:
(1) The norm principle holds for Ω(Q) over L/K.
(2) The square diagram (SQ for L/K) commutes.
10
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
Proof. (1) =⇒ (2): By Lemma 2.2, the norm principle holds for α over L/K. Let
z ∈ H(L) and α̃(L)(z) = [x] ∈ L∗ / Sn(QL ) for x ∈ L∗ . As observed above, x ∈
H(L) and α̃(L)(x) = [x] ∈ L∗ / Sn(QL ). Since α̃(L) is a group homomorphism, zx−1 ∈
Ker(α̃(L)) = Ker(α(L)). Since we assume that the norm principle holds for α over L/K,
α(K)(NL/K (zx−1 )) = 1. As NL/K (z), NL/K (x), NL/K (zx−1 ) ∈ H(K) and α(K) induces
α̃(K), α̃(K)(NL/K (z)) = α̃(K)(NL/K (x)) = [NL/K (x)] ∈ K ∗ / Sn(Q). Hence (SQ for
L/K) commutes.
z ∈ H(L)
α̃(L)
NL/K
NL/K (z) ∈ H(K)
[x] ∈ L∗ / Sn(QL )
NL/K
α̃(K)
[NL/K (x)] ∈ K ∗ / Sn(Q)
(2) =⇒ (1): Let u ∈ Ker(α(L)) = Ker(α̃(L)). By the commutativity of the square (SQ
for L/K), we have α̃(K) NL/K (u) = 1. Hence the norm principle holds for α over L/K
and by Lemma 2.2, it holds for Ω(Q) over L/K.
If Q/K is isotropic, then Sn(Q) = K ∗ and hence the square (SQ for L/K) commutes for
all L/K yielding the following:
Corollary 2.6. If Q/K is isotropic, then the norm principle holds for Ω(Q).
2.4.1. Reduction to quadratic extensions. We now show that to prove the norm principle
for Ω(Q), it suffices to consider separable quadratic extensions. More precisely, we prove
the following:
Theorem 2.7. Let Q/K be a quadratic form of even dimension. Suppose that for every
finite separable field extension M/K, the norm principle holds for the M -group Ω(Q)M
over every separable quadratic field extension M 0 /M . Then the norm principle holds for
Ω(Q).
Proof. Let L/K be any finite separable field extension. By Theorem 2.5, it suffices to show
that the square (SQ for L/K) commutes.
There exists a separable field extension
M/K with [M : K] = 2m + 1 for3 some m ≥ 0
Qr
such that LM := L ⊗K M ' i=1 Mi where each [Mi : M ] = 2mi with Mi /M separable
field extensions filtered by quadratic extensions.
By assumption and Lemma 2.2, the norm principle holds for the map α over each Mi /M
and hence over the étale algebra LM/M . Hence, by Theorem 2.5, the square (SQ for
LM/M ) commutes.
3
For instance, take M to be the fixed field of a 2-sylow subgroup of Gal(N/L) where N is the Galois
closure of L over K.
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS 11
Note that the natural map K ∗ / Sn(Q) → M ∗ / Sn(QM ) is injective. This is because if f ∈
K ∗ becomes a spinor norm from QM , then by Knesbusch’s norm principle, NM/K (f ) =
f 2m+1 is a spinor norm from Q and hence [f ] = 1 ∈ K ∗ / Sn(Q) to begin with.
Look at the following cuboid which has commutative front vertical face (SQ for LM/M )
as well as commutative side, top and bottom faces. A diagram chase and the injectivity
of the map K ∗ / Sn(Q) → M ∗ / Sn(QM ) shows that the back vertical face (SQ for L/K)
commutes.
H(L)
α̃(L)
N
H(K)
L∗ / Sn(QL )
N
α̃(K)
H(LM )
K ∗ / Sn(Q)
α̃(LM )
(LM )∗ / Sn(QLM )
N
H(M )
N
α̃(M )
M ∗ / Sn(QM )
3. A N INDUCTIVE APPROACH
In this section, we outline a possible inductive approach to the norm principle question.
Let Q/K be a quadratic form of even dimension as before and let L/K be a finite separable
field extension. Let u ∈ Ker(α(L)). We would like to show NL/K (u) ∈ Ker(α(K)). Set
j(L)(u) = [λ] ∈ H1 (L, µ2 ) for some λ ∈ L∗ . It follows from Figure 2 in Section 2.3 that
QL ' λQL .
Suppose that there exist even dimensional quadratic forms fu , gu defined over K such that
Q ' fu ⊥ gu and (fu )L ' λ (fu )L . Note that this immediately implies (gu )L ' λ (gu )L .
Let Ru := O+ (fu ) × O+ (gu ) ⊂ O+ (Q). Define an intermediate new group R̃u to be the
preimage of Ru under the canonical homomorphism Spin(Q) → O+ (Q).
12
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
We have the following diagram with exact rows:
1
µ2
µ
µ2
1
R̃u
Ru
1
Spin(Q)
O+ (Q)
1
id
1
µ2
id
1
µ2
Let γu : H1 (−, µ) → H1 (−, R̃u ) be the induced map from the inclusion µ ,→ R̃u .
Lemma 3.1. Let L/K be a finite separable field extension and u ∈ Ker(α(L)). Assume that there exist even dimensional quadratic forms fu , gu /K such that Q ' fu ⊥
gu and j(L)(u) = [λ] for λ ∈ L∗ with (fu )L ' λ (fu )L and (gu )L ' λ (gu )L . If
NL/K (Ker (γu (L))) ⊂ Ker (γu (K)), then NL/K (u) ∈ Ker (α(K)).
Proof. Let γu (L)(u) = v ∈ H1 (L, R̃u ). Since (fu )L ' λ (fu )L and (gu )L ' λ (gu )L , [λ]
goes to 1 in H1 (L, Ru ). Hence v goes to 1 in H1 (L, Ru ) and therefore there exists a ∈ L∗
such that [a] ∈ H1 (L, µ2 ) goes to v.
i(L)
H1 (L, µ2 )
id
u ∈ H1 (L, µ)
j(L)
[λ] ∈ H1 (L, µ2 )
γu (L)
1
1
v ∈ H (L, R̃u )
H (L, µ2 )
α(L)
1 ∈ H1 (L, Ru )
id
H1 (L, µ2 )
1 ∈ H1 (L, Spin(Q))
H1 (L, O+ (Q))
As µ ⊆ Z(R̃u ), we have i([a])−1 u is in the image of R̃u /µ (L) → H1 (L, µ) ([KMRT98],
Corollary 28.4, Pg 386). Hence γu (L) (i([a])−1 u) = 1 and therefore α(L)(i([a])−1 u) =
1 ∈ H1 (L, Spin(Q)). Since α(L)(u) = 1, we see that α(L)(i[a]) = 1, i.e. a is a spinor
norm of QL . By Knebusch’s
norm principle, NL/K (a) is a spinor norm of Q and hence
α(K) i[NL/K (a)] = 1.
By assumption on γu , we have γu (K) NL/K (i([a])−1 u) = 1 ∈ H1 (K,R̃u ) and hence this
element dies in H1 (L, Spin(Q)) also. That is, α(K) NL/K (i([a])−1 u) = 1. This implies
α(K)(NL/K (u)) = 1.
This leads to the following:
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS 13
Theorem 3.2. Let L/K be a finite separable field extension and u ∈ Ker(α(L)). Assume
that there exist even dimensional quadratic forms fu , gu /K such that Q ' fu ⊥ gu and
j(L)(u) = [λ] for λ ∈ L∗ with (fu )L ' λ (fu )L and (gu )L ' λ (gu )L . If the norm principle
holds for Ω(fu ) and Ω(gu ) over L/K, then NL/K (u) ∈ Ker (α(K)).
Proof. Recall the exact sequence of algebraic K-groups 1 → µ → R̃u → R̃u /µ → 1. By
Lemma 3.1, it suffices to check that NL/K (Ker (γu (L))) ⊂ Ker (γu (K)), which is equivalent to verifying that the norm principle holds for R̃u /µ(−) → H1 (−, µ) over L/K (c.f.
Lemma 2.2). As in Section 2.1, construct f : G → T given by the following commutative
diagram with exact rows and columns.
1
1
1
µ
R̃u
R̃u /µ
1
id
1
RZ/K Gm
G
1
R̃u /µ
1
f
f0
T
h
id
T
1
By Lemma 2.1, it suffices to check the norm principle for f : G → T over L/K or more
generally, the norm principle for G over L/K. Then ([BM00], Proposition 5.2) along with
the hypothesis that the norm principle holds for Ω(fu ) and Ω(gu ) concludes the proof.
4. R- EQUIVALENCE CLASSES OF PGO+ (Q)(K)
Let G be a linear algebraic group defined over K and L/K, a field extension. Then x, y ∈
G(L) are said to be R-equivalent if there exists an L-rational map f : A1L 99K G defined at
0 and 1 sending 0 ; x and 1 ; y. This defines an equivalence relation on G(L) ([Gi97],
Section II.1). Let RG(L) denote the normal subgroup of elements in G(L) which are Requivalent to the identity eG . We denote the quotient group G(L)/RG(L) by G(L)/R.
This is the group of R-equivalence classes introduced by Manin for the L-points of the
variety underlying group G.
14
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
The norm principles in ([Gi93], [Me96]) are stated in general for R-trivial elements (i.e,
elements R-equivalent to eG ).
Theorem 4.1 (Gille, [Gi93]). Let 1 → µ → G̃ → G → 1 be an isogeny of semisimple
algebraic groups over K and NL/K : H1 (L, µ) → H1 (K, µ) be the induced norm map
for a field extension L/K. Let RG(L) (resp. RG(K)) denote the elements of G(L) (resp
G(K)) which are R-equivalent to the identity.
RG(L)
δ(L)
H1 (L, µ)
NL/K
RG(K)
δ(K)
H1 (K, µ)
Then NL/K Im δ(L) : RG(L) → H1 (L, µ) ⊆ Im δ(K) : RG(K) → H1 (K, µ) .
A similar statement holds for the norm principles of morphisms f : G → T where G is a
reductive linear algebraic group ([Me96], Thm 3.9).
In [Me96(2)], the group of R-equivalence classes was computed for adjoint semisimple
classical groups. These computations applied to the adjoint group PGO+ (Q) translates to
a natural isomorphism PGO+ (Q)(K)/R ' G(Q)/ (K ∗2 Hyp(Q)) where,
- Q/K is a non-degenerate form of dim 2n,
- G(Q) is the group of similarities {λ ∈ K ∗ | λQ ' Q},
- Hyp(Q) is the subgroup generated by hNE/K (E ∗ ) | QE ' Hn i where E runs over
finite extensions of K.
The following lemma identifies some quadratic forms which give rise to adjoint groups of
type Dn with trivial group of R-equivalence classes over the base field.
Lemma 4.2. Let K be a complete discretely valued field with ring of integers OK and
∗
residue field k with char(k) 6= 2. Let τ = ha1 , a2 , . . . , ar i where each ai ∈ OK
. Let
+
∗
π ∈ K be a parameter of K and set ξ = τ ⊗ h1, πi. Then PGO (ξ)(K)/R is trivial.
∗
Proof. Let θ ∈ G(ξ). Thus up to squares in K ∗ , θ = xπ for some x ∈ OK
and ∈
+
∗2
{0, 1}. Since PGO (ξ)(K)/R ' G(ξ)/ (K Hyp(ξ)), we would like to show xπ ∈
Hyp(ξ)K ∗2 .
√
√
The totally ramified extension L0 = K( −π) splits ξ and π = NL0 /K ( −π). Thus π ∈
∗
Hyp(ξ). Thus we can assume θ = x ∈ G(ξ) for x ∈ OK
.
Let x denote the image of x in k and τ , the image of τ in W (k). Recall the second residue
homomorphism δ2,π : W(K) → W(k) with respect to the parameter π. Then we have
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS 15
[τ ] = δ2,π (ξ) = δ2,π (xπ ξ) = δ2,π (xξ) = [xτ ] ∈ W (k). This shows that [τ ⊗h1, −xik ] = 0
in W (k).
√
Look at L00 = K( −xπ) which is a complete discretely valued field with residue field
k. Recall the first residue homomorphism δ1,y : W(L00 ) → W(k) with respect to some
parameter y of L00 . Note that ξL00 ' τL00 ⊗ h1, −xiL00 and δ1,y (ξL00 ) = [τ ⊗ h1, −xik ] = 0 ∈
W (k). Thus by Hensel’s Lemma, [ξL00 ] = 0 ∈ W (L00 ).
√
Finally as xπ = NL00 /K ( −xπ), xπ ∈ Hyp(ξ). Since Hyp(ξ) is the subgroup generated by
norms from finite extensions which split S, x ∈ Hyp(ξ), which concludes the proof.
5. OVER COMPLETE DISCRETELY VALUED FIELDS
In this section, we work over a complete discretely valued field K. We fix the convention of
letting ? denote the image of ? in the residue field for ? defined over the ring of integers of
a complete discretely valued field. We also let OX denote the ring of integers of a complete
discretely valued field X. We now state the main result of this paper.
Theorem 5.1. Let K be a complete discretely valued field with residue field k with char(k) 6=
2. Assume that the norm principle holds for Ω(q) for every non-degenerate quadratic form
q of even dimension defined over any finite extension of k. Then the norm principle holds
for Ω(Q) for every non-degenerate quadratic form Q of even dimension over K.
By Theorem 2.7, it suffices to show that the norm principle holds for Ω(Q) over separable quadratic extensions L/K. Fix a uniformizing parameter t of K. Write the nondegenerate quadratic form Q/K in the form q ⊥ tp with q ' ha1 , a2 , . . . , adim q i and
∗
. Since dim Q = 2n, dim q and dim p have the same
p ' hb1 , b2 , . . . , bdim p i for ai , bj ∈ OK
parity.
Note that L/K is a complete discretely valued field. Let `/k denote its residue field. Then
one of the following holds:
- L/K is an unramified extension and θ := t is a parameter of L.
√
∗
and
- L/K √
is totally ramified and ` = k. Further L ' K( ct) for some c ∈ OK
θ := ct is a parameter of L.
By Lemma 2.2, it suffices to verify the norm principle for α over L/K. Let u ∈ Ker (α(L)) ⊆
H(L). Set j(L)(u) = [λ] ∈ H1 (L, µ2 ) ' L∗ /L∗2 for some representative λ ∈ L∗ . We
would like to show that α(K) NL/K (u) = 1 ∈ H1 (K, Spin(Q)). By Corollary 2.6, we
can assume q and p are anisotropic over K.
5.1. Lemmata. We begin by reducing to the case when λ is a unit in L.
Lemma 5.2. Up to squares, λ can be assumed to be in OL∗ .
16
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
Proof. Without loss of generality, we can assume λ ∈ OL∗ or λ = λ̃θ for λ̃ ∈ OL∗ where θ
is a parameter of L.
If L/K is unramified and λ = λ̃θ where θ = t, using λQL ' QL and the second residue
map δ2,t : W (L) → W (`) for the parameter t of L, we see that
[p` ] = δ2,t (QL ) = δ2,t (λQL ) = [λ̃q ` ] ∈ W (`).
Thus QL ' qL ⊗h1, λ̃tiL . By Lemma 4.2, PGO+ (Q)(L)/R = {1} and hence RPGO+ (Q)(L) =
PGO+ (Q)(L). Then Theorem 4.1 implies that the norm principle holds for our required
map S : PGO+ (Q) → H1 (−, µ) over the extension L/K and hence for Ω(Q) over L/K
(Lemma 2.2).
√
√
∗
and λ = λ̃θ where θ := ct and QL ' qL ⊥ cpL , using
If L ' K( ct) for some c ∈ OK
λQL ' QL and the second residue map δ2,θ : W (L) → W (k) for the parameter θ of L, we
see that
0 = δ2,θ (QL ) = δ2,θ (λQL ) = [λ̃q ⊥ λ̃cp] ∈ W (k).
Thus [q ⊥ cp] = 0 ∈ W (k) and hence [QL ] = 0 ∈ W (L). This implies PGO+ (Q)(L)/R =
{1} and hence RPGO+ (Q)(L) = PGO+ (Q)(L). Then Theorem 4.1 and Lemma 2.2 imply
that the norm principle holds for Ω(Q) over L/K.
Next, we describe the shape of the element u ∈ Ker (α(L)) under consideration when QL
is unramified. We construct a related element u0 ∈ H1 (OL , µ) which in fact also lives in
Ker (α(L)). To do so, we make use of the explicit description of the maps i(L) and j(L)
given in Section 2.3.
Lemma 5.3. Let L be a complete discretely valued field with rings of integers OL , parameter θ and residue field `. Let Q0 = hx1 , x2 , . . . , x2r i be a quadratic form over L where each
xi ∈ OL∗ . Let u be an element in the kernel of α(L) : H1 (L, µ0 ) → H1 (L, Spin(Q0 )) where
µ0 is the center of Spin(Q0 ). Assume j(L)(u) = [λ] ∈ H1 (L, µ2 )for some λ ∈ OL∗ . Then
0
there exist u0 ∈ H1 (OL , µ0 ) and 0 ∈ Z such that u = u0 i(L)[θ ] .
Proof. Let Z 0 denote the discriminant of Q0 . Since Q0 is unramified over L, Z 0 is an unramified (possibly split) quadratic extension of L. Thus θ is still a parameter of Z 0 .
U (L)
Suppose first that dim Q0 ∼
= 2 mod 4 and Z 0 is a field. Since H1 (L, µ0 ) = U0 (L) , there
exist ∈ Z, f ∈ OL∗ and z ∈ OZ∗ 0 with NZ 0 /L (z) = f 4 such that u = [f θ , zθ2 ]. Define
u0 := [f, z] in H1 (OL , µ0 ). Since i(L)(θ) = [θ, θ2 ], it is clear that u = u0 (i(L)[θ ]).
The case when dim Q0 ∼
= 2 mod 4 but Z 0 ' L × L is a little more delicate. Again, clearly
there exist , 1 ∈ Z, f, z1 , z2 ∈ OL∗ with z1 z2 = f 4 such that u = [f θ , z1 θ1 , z2 θ4−1 ].
Define u0 := [f, z1 , z2 ] in H1 (OL , µ0 ).
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS 17
Since j(L)(u) = [λ] and up to squares λ ∈ OL∗ , we see that 1 introduced above is even4.
Since [ab, a4 , b4 ] = [1] ∈ H1 (L, µ0 ) and i(L)(θ) = [θ, θ2 , θ2 ], we see that
(
u0
if 1
u=
0
u (i(L)[θ]) if 1
∼
= 0 mod 4,
∼
= 2 mod 4.
Suppose now that dim Q0 ∼
= 0 mod 4. Since H1 (L, µ0 ) = Z 0∗ /Z 0∗2 , there exist , 1 ∈ Z
and z ∈ OZ∗ 0 (resp. z1 , z2 ∈ OL∗ ) such that u = [zθ ] (resp. u = [z1 θ , z2 θ1 ]) if Z 0 is a field
(resp. a split extension). Define u0 := [z] (resp. [z1 , z2 ]) in H1 (OL , µ0 ). As λ ∈ OL∗ up to
squares, we can assume5 that = 1 . Thus u = u0 (i(L)[θ ]) ∈ H1 (L, µ0 ).
The following lemma shows that u0 defined above still lives in the kernel of α(L).
Lemma 5.4. Let u, u0 be as in Lemma 5.3. Then u0 ∈ Ker (α(L)).
0
Proof. We have u = u0 i(L)[θ ] ∈ H1 (L, µ0 ).
Since Q0 is unramified over L, we can view Q0 as the generic fiber of a diagonal quadratic
form Q0OL . Further Spin(Q0L ) is the generic fiber of a smooth reductive group scheme
G = Spin(Q0OL ) over OL .
√
√
Set E := L( −θ) if dim Q0 ∼
= 2 mod 4 and E := L( θ) if dim Q0 ∼
= 0 mod 4. In
either case, E is a totally ramified extension of L with residue field `. Since i(E)(θ) = [1],
it is clear that u0 is the image of u in H1 (E, µ0 ). Since α(L)(u) = 1, we have α(E)(u0 ) = 1.
u ∈ H1 (L, µ0 )
α(L)
1 ∈ H1 (L, Spin(Q0 ))
u0 ∈ H1 (E, µ0 )
α(E)
1 ∈ H1 (E, Spin(Q0 ))
Since u0 is defined over OL and hence OE and the kernel of the natural map H1 (OE , G) →
H1 (E, Spin(Q0 )) is trivial ([Ni84]), we see that u0 ∈ Ker α(OE ) : H1 (OE , µ0 ) → H1 (OE , G) .
u0 ∈ H1 (OE , µ0 )
u0 ∈ H1 (E, µ0 )
4
α(OE )
α(E)
H1 (OE , G)
1 ∈ H1 (E, Spin(Q0 ))
[f −2 z1 θ1 −2 , f −2 z2 θ2−1 ] = [f −2 z1 θ1 −2 , 1]ψ[f −2 z1 θ1 −2 , 1]−1 and [λ] = [f −2 z1 θ1 −2 ].
5
[λ] = [z1 z2 θ+1 ] when Z 0 ' L × L.
18
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
Specializing to the residue field of E, we see that u0 ∈ Ker α(`) : H1 (`, µ0 ) → H1(`, G) .
By Hensel’s lemma, this implies u0 ∈ Ker α(OL ) : H1 (OL , µ0 ) → H1 (OL , G) which
shows that α(L)(u0 ) = 1.
5.2. Proof of the Theorem 5.1. We proceed by induction on dim Q. For the base case
when dim Q = 2, the norm principle holds for Ω(Q) because it is a commutative group. Assume that the norm principle holds for Ω(Q̃) over L/K for all even dimensional quadratic
forms Q̃/K with dim Q̃ < dim Q. We break up the proof into separate cases depending on
the ramification of L/K.
Case I : L/K is unramified. Recall that we have reduced to the case where λ ∈ OL∗ up to
squares. Using the second residue map again, we see that [p` ] = δ2,t (QL ) = δ2,t (λQL ) =
[λp` ] ∈ W (`). Similarly q ` ' λq ` . Thus λ ∈ G(qL ) ∩ G(pL ).
Note that if dim q is odd, since [h1, −λiL ⊗ qL ] = 0 ∈ W (L), then by ([Sch85], Theorem
10.13), we have λ ∈ L∗2 , i.e [λ] = 1 ∈ H1 (L, µ2 ). Thus by Lemma 2.3, the norm principle
holds.
So we assume that dim q is even. Therefore dim p is even too. Set fu = q, gu = tp,
Ru := O+ (fu ) × O+ (gu ) ⊂ O+ (Q) and R̃u , the preimage of Ru under the canonical
homomorphism Spin(Q) → O+ (Q). By Theorem 3.2, it suffices to show that the norm
principle holds for Ω(fu ) and Ω(gu ) over L/K, which holds by induction if dim q, dim p 6=
0.
Without loss of generality6, suppose that dim p = 0 and QK ' qK . Since Q is unramified
over K and hence QL is unramified over
L, using Lemmata 5.3 and 5.4, we find u0 ∈
0
H1 (OL , µ) such that u = u0 i(L)[t ] for some 0 ∈ Z and α(L)(u0 ) = 1. The proof
of Lemma 5.4 in fact shows that we can specialize to the residue field and get u0 in the
kernel of α(`) : H1 (`, µ) → H1 (`, Spin(QL )). Since the norm principle holds for α for
the quadratic form Q over `/k by assumption and Lemma 2.2, N`/k (u0 ) is in the kernel
of α(k) : H1 (k, µ) → H1 (k, Spin(Q). Thus by Hensel’s Lemma, α(K) NL/K (u0 ) = 1.
0
Lemma 2.3 implies that NL/K (i(L)[t ]) ∈ Ker (α(K)). Thus NL/K (u) ∈ Ker (α(K)).
Case II : L/K is ramified. Recall once again that we have reduced to the case where
∗
λ ∈ OL∗ up to squares. Using Hensel’s lemma, we can assume in fact that λ ∈ OK
. Further
we can also assume λ 6∈ L∗2 as otherwise, we would be done by Lemma 2.3.
Subcase IIa: We first look at the situation when dim p 6= dim q. Then the following
Lemma in conjunction with Theorem 3.2 and our induction hypothesis finishes the proof
in this case.
6
If dim q = 0, Q is similar to the unramified form p over K and Ω(Q) ' Ω(p). The same proof works in
this case.
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS 19
Lemma 5.5. Suppose that dim p 6= dim q. Then q ⊗ h1, −λi or p ⊗ h1, −λi is isotropic
over K. Further QK ' fK ⊥ gK for f, g even dimensional quadratic forms over K with
λfL ' fL and λgL ' gL .
Proof. Since λ is a multiplier for the form q ⊥ cp, we have [h1, −λiL ⊗ q ⊥ h1, −λiL ⊗
cpL ] = 0 ∈ W (L). Without loss of generality, assume dim q > dim p. Hence q ⊗ h1, −λi
is isotropic over L.
∗
∗
Recall that q ' ha1 , a2 , . . . , adim q i for ai ∈ OK
and λ ∈ OK
. Thus q ⊗ h1, −λi '
0
0
0
0
ha1 , a2 , . . . , a2 dim q i for some ai ∈ OK ∗. Hence there exist wi ∈ OL∗ ∪ {0} not all 0 and
P
ti ∈ Z such that a0i wi2 θ2ti = 0 where θ is a parameter of L. By cancelling factors of θ2
if necessary, we can assume that each tj ≥ 0 and at least one ti = 0 (with corresponding
wi 6= 0). That is q ⊗h1, −λi is isotropic over k. By Hensel’s lemma, q ⊗h1, −λi is isotropic
over K.
Since we have assumed q is anisotropic over K, this implies q(u) = λq(v) for u, v ∈
K dim q and q(u), q(v) 6= 0. Note that if u, v are linearly dependent over K, then λ ∈ K ∗2
contradicting our assumption that λ is not a square. Thus u, v span a two dimensional
K-vector space W . Let f be the K-quadratic form q restricted to W .
Let q(v) = a, q(u) = λa and bq (u, v) = x. Then for orthogonal basis {v, u − xa v},
2
x
x2
fK ' ha, λa − xa iK and for orthogonal basis {u, v − λa
u}, fK ' hλa, a − λa
iK . Thus
λfL ' fL and hence the Lemma follows.
Subcase IIb: Assume now that dim p = dim q. Since QL is unramified
over L, 0use Lem1
0
0
0
mata 5.3 and 5.4 to find u ∈ H (OL , µ) such that u = u i(L)[θ ] for some ∈ Z and
α(L)(u0 ) = 1.
Lemma 5.6. NL/K (u0 ) = 1 ∈ H1 (K, µ).
Proof. We only give the proof in the case ZL is a field. The proof when ZL ' L × L is
similar.
If dim q = dim p is odd, then the discriminant Z of Q = q ⊥ tp is a totally ramified
quadratic extension of K and ZL := Z ⊗K L is an unramified (possibly split) quadratic
extension of L. Thus θ is still a parameter of ZL . Further, if the residue field of ZL = `0 ,
then the norm maps NZL /L : ZL∗ → L∗ and NZL /Z : ZL∗ → Z ∗ induce the same map
N`0 /k : `0 ∗ → k ∗ at the residue field level.
L
ram
K
unram
ZL
unram
ram
Z
20
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
Recall that H1 (L, µ) =
U (L)
U0 (L)
and u0 = [f, z] for some f ∈ OL∗ and z ∈ OZ∗ L with
4
NZL /L (z) = f 4 . Thus we have NZL /Z (z) = N`0 /k (z) = f . By Hensel’s Lemma, there
∗
exist f˜ ∈ OK
and x ∼
= 1 mod (mOL ) in OL∗ such that f = xf˜. Similarly, there exist
y, y 0 ∼
= 1 mod (mOZ ) in OZ∗ such that NZL /Z (z) = f˜4 y 0 = f˜4 y 4 .
Now NL/K (u0 ) = [NL/K (f ), NZL /Z (z)] = [f˜2 NL/K (x), f˜4 y 4 ] = [f˜2 , f˜4 ][NL/K (x), y 4 ] ∈
H1 (K, µ) where NZ/K (y 4 ) = NL/K (x)4 . Setting a = NZ/K (y) NL/K (x)−1 , we see that
a4 = 1 and a ∼
= 1 mod mOK . By Hensel’s Lemma yet again, a = 1 and hence NZ/K (y) =
NL/K (x). Since [N(b), b4 ] = 1 ∈ H1 (K, µ) for every b ∈ K ∗ , we see that NL/K (u0 ) = 1.
If dim q = dim p is even, then the discriminant Z of Q = q ⊥ tp is an unramified (possibly
split) quadratic extension of K as also ZL /L := Z ⊗K L/L. Thus θ is still a parameter of
ZL . Note that ZL ' L × L if and only if Z ' K × K.
L
ram
K
unram
ZL
ram
unram
Z
Recall that H1 (L, µ) = ZL∗ /ZL∗2 and u0 = [z] in H1 (L, µ) for some z ∈ OZ∗ L if ZL is a field
(resp. a split extension). Since norms of units of totally ramified quadratic extensions are
squares and ZL /Z is ramified, NL/K (u0 ) = 1 ∈ H1 (K, µ).
0
Lemma 2.3 implies that NL/K (i(L)[θ ]) ∈ Ker (α(K)). Since NL/K (u) = NL/K (u0 ) NL/K
we have NL/K (u) ∈ Ker (α(K)) which concludes the proof of Theorem 5.1.
6. E XAMPLES
Let G be a semisimple simply connected linear algebraic group defined over a field k. Then
H1 (k, G) is trivial if k is a p-adic field ([K65]) or a global field of positive characteristic
([H75]). More generally, suppose that char(k) 6= 2 and cd(k) ≤ 2. Then Bayer-Parimala’s
proof of Serre’s conjecture II shows that H1 (k, G) is trivial if G is further assumed to be of
classical type ([BP95]). Set G := Spin(q) where q is any even dimensional nondegenerate
quadratic form over k and let µ be its center. It is immediate therefore that the norm
principle holds for the map α : H1 (−, µ) → H1 (−, Spin(q)) and hence that it holds for the
group Ω(q) defined over k.
Now let G be a semisimple simple adjoint linear algebraic group of classical type defined
over a number field k. Then, G(k)/R, the group of R-equivalence classes of the k-points
of G, is trivial ([Gi97] Corollaire III.4.2 and [KP08], pg 1). Set G := PGO+ (q) where
0
i(L)[θ ]
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS 21
q is any even dimensional nondegenerate quadratic form over k and let µ be the center
of Spin(q). It follows from Theorem 4.1 that the norm principle holds for the map S :
PGO+ (q)(−) → H1 (−, µ) and hence that it holds for the group Ω(q) defined over k.
Recall that a field k is said to √
have virtual cohomological dimension (vcd) ≤ n if the
cohomological dimension of k −1 is ≤ n. Examples of vcd ≤ 2 fields include cd ≤ 2
fields and number fields. We begin by showing the following:
Lemma 6.1. Let F be a real closed field and q 0 , a non-degenerate even dimensional quadratic form over F . Then PGO+ (q 0 )(F )/R is trivial.
Proof. Without loss of generality we can assume that q 0 is anisotropic over F . Since for
every a ∈ F , either a or −a is a square in F , we can further assume that q 0 ' h1, . . . , 1i.
Then the variety of PGO+ (q 0 ) is stably rational over F ([Ch94]), whence the claim.
Proposition 6.2. Let k be a field with vcd(k) ≤ 2 and q, a non-degenerate even dimensional quadratic form over k. Then the norm principle holds for Ω(q).
Proof. If cd(k) ≤ 2, the discussion above already gives the proof. Hence we can assume
that cd(k) 6= vcd(k) and hence that char(k) = 0 and k can be ordered (Theorem 1.1,
[BP98]).
Let `/k be a finite separable extension, µ, the center of Spin(q), and let α(−) : H1 (−, µ) →
H1 (−, Spin(q)) be the natural map between the Galois cohomology sets. Let ξ ∈ Ker (α(`))
and set η := N`/k (ξ). We would like to show that α(k) (η) is trivial in H1 (k, Spin(q)).
Let Ω denote the set of all orderings v of k and kv , the real closure of k at v. Then the
Hasse principle result of Bayer-Parimala over perfect fields of vcd ≤ 2 for semisimple
simply connected groups
of classical type ([BP98]) gives in particular that the natural map
Q
1
H (k, Spin(q)) → v∈Ω H1 (kv , Spin(q)) has trivial kernel.
Thus, it suffices to show that the image of η in H 1 (kv , Spin(q)) is trivial for each v ∈ Ω.
Note that Reskv (η) = N`⊗k kv /kv (ξ). By Lemma 6.1, PGO+ (q)(` ⊗k kv )/R is trivial and
hence as the norm principle holds for R-trivial elements (Theorem 4.1), we can conclude
that the image of Reskv (η) in H 1 (kv , Spin(q)) is trivial.
Since virtual cohomological dimension behaves well with respect to finite extensions,
Proposition 6.2 in conjunction with Theorem 5.1 immediately yields the following:
Corollary 6.3. Let K be a complete discretely valued field with residue field k such that
char(k) 6= 2 and vcd(k) ≤ 2. Then the norm principle holds for Ω(Q) for every even
dimensional non-degenerate quadratic form Q/K.
22
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
7. O N THE TRIVIALITY OF THE TATE -S HAFEREVICH SET
Let G be a semisimple algebraic group over a number field K and V K , the set of all places
of K. One of the main finiteness results in the arithmetic theory of linear algebraic groups
states that the natural global-to-local map
Y
ρG : H1 (K, G) →
H1 (Kv , G)
v∈V K
is proper, i.e. the preimage of any finite set is finite; in particular, the corresponding TateShafarevich set X(G) := Ker ρG is finite. Moreover, if in addition G is simply connected,
then X(G) = 1, i.e. ρG is injective.
A natural question to ask is if, and to what extent, the above finiteness property can be
extended to fields other than number fields. More precisely, let K be a finitely generated
field. Can one equip K with a “natural” set V of discrete valuations such that for a given
absolutely almost simple K-group G, the natural global-to-local map relative to V
Y
ρG,V : H1 (K, G) →
H1 (Kv , G)
v∈V
is proper ? If the answer is affirmative, is it true that for a simply connected group G the
kernel XV (G) of ρG,V is trivial ?
A natural candidate for such a V appears to be the set of discrete valuations associated to the
prime divisors of a model of K, i.e. a smooth affine arithmetic scheme with function field
K (we call such sets divisorial). It is known that divisorial sets V indeed work for adjoint
inner forms of type A` , i.e. for G = PGL`+1 , provided that char K does not divide ` + 1
(cf. [CRR13]); this relies on the finiteness of the unramified Brauer group (`+1) Br(K)V
([CRR16]). Until recently no other types have been considered.
Relating the norm principle and the Tate-Shaferevich set of spinor groups. The first interesting and widely open case is the one when K is the function field of a curve C defined
over a number field and G = Spin(f ) is the spinor group of a quadratic form f over K. In
a recent paper ([CRR16(2)]) it was proved that for the special orthogonal group O+ (f ) and
for a divisorial set V , the global-to-local map ρO+ (f ),V is proper. This result in conjunction
with the exact sequence
O+ (f )(K) −→ H1 (K, µ2 ) −→ H1 (K, Spin(f )) −→ H1 (K, O+ (f )),
and twisting shows that the Tate-Shafarevich set XV (Spin(f )) is finite if and only if the
group LGC(g) = {[a] ∈ K × / Sn(g) | a is a spinor norm of g over Kv for all v ∈ V } is
finite for all quadratic forms g over K. Note that LGC(g) is a subset of XV (Spin(g)).
Note that the residue field κ(v) of any valuation v ∈ V is either a number field or the
function field of a curve over a finite field, i.e., vcd(κ(v)) ≤ 2. Theorefore by Corollary
THE NORM PRINCIPLE FOR TYPE Dn GROUPS OVER COMPLETE DISCRETELY VALUED FIELDS 23
6.3, the norm principle for Ω(f ) holds over the completion Kv . Thus by Theorem 2.5,
the square diagram (SQ for X/Kv ) is commutative for every finite separable extension
X/Kv . Let L/K be a finite separable extension. It follows then that the obstruction to the
commutativity of the square diagram (SQ for L/K) is a subgroup in LGC(f ) having the
property: it is trivial if and only if the norm principle holds for the extension L/K. Thus,
the failure of the norm principle for Spin(f ) would imply that XV (Spin(f )) is nontrivial.
R EFERENCES
[BM00] P. Barquero and A. Merkurjev, Norm Principle for Reductive Algebraic Groups, J. Proceedings of
the International Colloquium on Algebra, Arithmetic and Geometry TIFR, Mumbai (2000).
[BP95] E. Bayer-Fluckiger and R. Parimala, Galois cohomology of the classical groups over fields of cohomological dimension ≤ 2, Invent. Math. 2 (1995) : pgs 195-229.
[BP98] E. Bayer-Fluckiger and R. Parimala, Classical groups and the Hasse principle, Annals of mathematics
147(3) (1998) : pgs 651-693.
[Bh16] N. Bhaskhar, On Serre’s injectivity question and norm principle, Commentarii Mathematici Helvetici
91(1) (2016) : pgs 145-161.
[Ch94] V. Chernousov, The group of multipliers of the canonical quadratic form and stable rationality of the
variety PSO, Matematicheskie zametki 55 (1994) : pgs 114-119.
[CRR13] V.I. Chernousov, A.S. Rapinchuk and I.A. Rapinchuk, The genus of a division algebra and the
unramified Brauer group, Bull. Math. Sci. 3 (2013): pgs 211-240.
[CRR16] V.I. Chernousov, A.S. Rapinchuk and I.A. Rapinchuk, On the size of the genus of a division algebra, Proc. Steklov Inst. of Math. 292(1) (2016) : pgs 63-93.
[CRR16(2)] V.I. Chernousov, A.S. Rapinchuk and I.A. Rapinchuk, On some finiteness properties of algebraic groups over finitely generated fields, C. R. Acad. Sci. Paris, Ser. I, 354 (2016) : pgs 869-873.
[Gi93] P. Gille, R-équivalence et principe de norme en cohomologie galoisienne, Comptes rendus de
l’Académie des sciences. Série 1, Mathmatique 316(4) (1993) : pgs 315-320.
[Gi97] P. Gille, La R-équivalence sur les groupes algébriques réductifs, Publications mathématiques de
lIHÉS 86 (1997) : pgs 199-235.
[H75] G. Harder, Über die Galoiskohomologie halbeinfacher algebraischer Gruppen. III. Journal für die
reine und angewandte Mathematik 274 (1975) : pgs 125-138.
[K65] M. Kneser, Galois-Kohomologie halbeinfacher algebraischer Gruppen über p-adischen Körpern. I,
Math. Z. 88 (1965) : pgs 40-47, II ibid 89 (1965) : pgs 250-272.
[KMRT98] M.-A. Knus, A. Merkurjev, M. Rost, and J.-P. Tignol, The book of involutions, American Mathematical Society Colloquium Publications 44 (1998), AMS.
[KP08] A. Kulshrestha and R. Parimala, R-equivalence in adjoint classical groups over fields of virtual cohomological dimension 2, Transactions of the American Mathematical Society 360(3) (2008) : pgs
1193-1221.
[Me96] A. Merkurjev, A norm principle for algebraic groups, St Petersburg Math 7 (2) (1996) : pgs 243-264.
[Me96(2)] A. Merkurjev, R-equivalence and rationality problem for semisimple adjoint classical groups,
Publications Mathématiques de l’IHÉS 84 (1996) : pgs 189-213.
[Ni84] Y. Nisnevich, Rationally Trivial Principal Homogeneous Spaces and Arithmetic of Reductive Group
Schemes Over Dedekind Rings, C. R. Acad. Sci. Paris, Série I, 299 , no. 1 (1984) : pgs 58. Astérisque
227 (783(4)) (1995) : pgs 229-257.
[Sch85] W. Scharlau, Quadratic and Hermitian forms, Springer-Verlag (1985).
24
NIVEDITA BHASKHAR, VLADIMIR CHERNOUSOV, AND ALEXANDER MERKURJEV
D EPARTMENT OF M ATHEMATICS U NIVERSITY OF C ALIFORNIA AT L OS A NGELES , L OS A NGELES , CA
90095-1555
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICAL S CIENCES , U NIVERSITY OF A LBERTA , E DMONTON , A LBERTA , C ANADA
T6G 2G1
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS U NIVERSITY OF C ALIFORNIA AT L OS A NGELES , L OS A NGELES , CA
90095-1555
E-mail address: [email protected]
| 4math.GR
|
DeepLung: Deep 3D Dual Path Nets for
Automated Pulmonary Nodule Detection and Classification
1
Wentao Zhu1
Chaochun Liu2
Wei Fan3
2
University of California, Irvine
Baidu Research
arXiv:1801.09555v1 [cs.CV] 25 Jan 2018
{wentaoz1, xhx}@ics.uci.edu
[email protected]
Abstract
In this work, we present a fully automated lung computed
tomography (CT) cancer diagnosis system, DeepLung.
DeepLung consists of two components, nodule detection
(identifying the locations of candidate nodules) and classification (classifying candidate nodules into benign or malignant). Considering the 3D nature of lung CT data and
the compactness of dual path networks (DPN), two deep
3D DPN are designed for nodule detection and classification respectively. Specifically, a 3D Faster Regions with
Convolutional Neural Net (R-CNN) is designed for nodule detection with 3D dual path blocks and a U-net-like
encoder-decoder structure to effectively learn nodule features. For nodule classification, gradient boosting machine
(GBM) with 3D dual path network features is proposed.
The nodule classification subnetwork was validated on a
public dataset from LIDC-IDRI, on which it achieved better performance than state-of-the-art approaches and surpassed the performance of experienced doctors based on
image modality. Within the DeepLung system, candidate
nodules are detected first by the nodule detection subnetwork, and nodule diagnosis is conducted by the classification subnetwork. Extensive experimental results demonstrate that DeepLung has performance comparable to experienced doctors both for the nodule-level and patient-level
diagnosis on the LIDC-IDRI dataset.1
1. Introduction
Lung cancer is the most common cause of cancer-related
death in men. Low-dose lung CT screening provides an effective way for early diagnosis, which can sharply reduce
the lung cancer mortality rate. Advanced computer-aided
diagnosis systems (CADs) are expected to have high sensitivities while at the same time maintaining low false positive
rates. Recent advances in deep learning enable us to rethink
the ways of clinician lung cancer diagnosis.
1 https://github.com/uci-cbcl/DeepLung.git
3
Xiaohui Xie1
Tencent Medical AI Lab
[email protected]
Current lung CT analysis research mainly includes nodule detection [6, 5], and nodule classification [26, 25, 14,
33]. There is few work on building a complete lung CT
cancer diagnosis system for fully automated lung CT cancer diagnosis using deep learning, integrating both nodule
detection and nodule classification. It is worth exploring a
whole lung CT cancer diagnosis system and understanding
how far the performance of current deep learning technology differs from that of experienced doctors. To our best
knowledge, this is the first work for a fully automated and
complete lung CT cancer diagnosis system using deep nets.
The emergence of large-scale dataset, LUNA16 [24],
accelerated the nodule detection related research. Typically, nodule detection consists of two stages, region proposal generation and false positive reduction. Traditional
approaches generally require manually designed features
such as morphological features, voxel clustering and pixel
thresholding [20, 15]. Recently, deep ConvNets, such as
Faster R-CNN [21, 17] and fully ConvNets [18, 37, 31, 30,
29], are employed to generate candidate bounding boxes
[5, 6]. In the second stage, more advanced methods or complex features, such as carefully designed texture features,
are used to remove false positive nodules. Because of the
3D nature of CT data and the effectiveness of Faster R-CNN
for object detection in 2D natural images [13], we design a
3D Faster R-CNN for nodule detection with 3D convolutional kernels and a U-net-like encoder-decoder structure to
effectively learn latent features [22]. The U-Net structure is
basically a convolutional autoencoder, augmented with skip
connections between encoder and decoder layers [22]. Although it has been widely used in the context of semantic
segmentation, being able to capture both contextual and local information should be very helpful for nodule detections
as well. Because 3D ConvNet has too large a number of parameters and is difficult to train on public lung CT datasets
of relatively small sizes, 3D dual path network is employed
as the building block since deep dual path network is more
compact and provides better performance than deep residual network at the same time [3].
Before the era of deep learning, manual feature engi-
Figure 1. The framework of DeepLung. DeepLung first employs 3D Faster R-CNN to generate candidate nodules. Then it uses deep 3D
DPN to extract deep features from the detected and cropped nodules. Lastly, GBM with deep features, detected nodule size, and raw pixels
is employed for classification. Patient-level diagnosis can be achieved by fusing the classification results of detected nodules in the CT.
neering followed by classifiers was the general pipeline for
nodule classification [10]. After the large-scale LIDC-IDRI
[2] dataset became publicly available, deep learning-based
methods have become the dominant framework for nodule classification research [25, 35]. Multi-scale deep ConvNet with shared weights on different scales has been proposed for the nodule classification [26]. The weight sharing
scheme reduces the number of parameters and forces the
multi-scale deep ConvNet to learn scale-invariant features.
Inspired by the recent success of dual path network (DPN)
on ImageNet [3, 4], we propose a novel framework for CT
nodule classification. First, we design a deep 3D dual path
network to extract features. As gradient boosting machines
(GBM) are known to have superb performance given effective features, we use GBM with deep 3D dual path features,
nodule size, and cropped raw nodule CT pixels for the nodule classification [8].
Finally, we built a fully automated lung CT cancer diagnosis system, henceforth called DeepLung, by combining the nodule detection network and nodule classification
network together, as illustrated in Fig. 1. For a CT image, we first use the detection subnetwork to detect candidate nodules. Next, we employ the classification subnetwork to classify the detected nodules into either malignant
or benign. Finally, the patient-level diagnosis result can be
achieved for the whole CT by fusing the diagnosis result of
each nodule.
Our main contributions are as follows: 1) To fully exploit the 3D CT images, two deep 3D ConvNets are designed for nodule detection and classification respectively.
Because 3D ConvNet contains too many parameters and is
difficult to train on relatively small public lung CT datasets,
we employ 3D dual path networks as the neural network
architecture since DPN uses less parameters and obtains
better performance than residual network [3]. Specifically,
inspired by the effectiveness of Faster R-CNN for object
detection [13], we propose 3D Faster R-CNN for nodule
detection based on 3D dual path network and U-net-like
encoder-decoder structure, and deep 3D dual path network
for nodule classification. 2) Our classification framework
achieves better performance compared with state-of-the-art
approaches, and surpasses the performance of experienced
doctors on the public dataset, LIDC-IDRI. 3) Our fully automated DeepLung system, nodule classification based on
detection, is comparable to the performance of experienced
doctors both on nodule-level and patient-level diagnosis.
2. Related Work
Traditional nodule detection involves hand-designed features or descriptors [19] requiring domain expertise. Recently, several works have been proposed to use deep ConvNets for nodule detection to automatically learn features,
which is proven to be much more effective than handdesigned features. Setio et al. proposes multi-view ConvNet for false positive nodule reduction [23]. Due to the
3D nature of CT scans, some work proposed 3D ConvNets
to handle the challenge. The 3D fully ConvNet (FCN) is
proposed to generate region candidates, and deep ConvNet
with weighted sampling is used for false positive reduction
[6]. Ding et al. and Liao et al. use the Faster R-CNN to
generate candidate nodules followed by 3D ConvNets to remove false positive nodules [5, 17]. Due to the effective
performance of Faster R-CNN [13, 21], we design a novel
network, 3D Faster R-CNN with 3D dual path blocks, for
the nodule detection. Further, a U-net-like encoder-decoder
scheme is employed for 3D Faster R-CNN to effectively
learn the features [22].
Nodule classification has traditionally been based on
segmentation [7] and manual feature design [1]. Several
works designed 3D contour feature, shape feature and texture feature for CT nodule diagnosis [32, 7, 10]. Recently,
deep networks have been shown to be effective for medical
images. Artificial neural network was implemented for CT
nodule diagnosis [28]. More computationally effective network, multi-scale ConvNet with shared weights for different scales to learn scale-invariant features, is proposed for
Figure 2. Illustration of dual path connection [3], which benefits
both from the advantage of residual learning [11] and that of dense
connection [12] from network structure design intrinsically.
nodule classification [26]. Deep transfer learning and multiinstance learning is used for patient-level lung CT diagnosis [25, 36]. A comparative study on 2D and 3D ConvNets
is conducted and 3D ConvNet is shown to be better than
2D ConvNet for 3D CT data [33]. Furthermore, a multitask learning and transfer learning framework is proposed
for nodule diagnosis [14]. Different from their approaches,
we propose a novel classification framework for CT nodule diagnosis. Inspired by the recent success of deep dual
path network (DPN) on ImageNet [3], we design a novel
3D DPN to extract features from raw CT nodules. In part to
the superior performance of GBM with complete features,
we employ GBM with different levels of granularity ranging from raw pixels, DPN features, to global features such
as nodule size for the nodule diagnosis. Patient-level diagnosis can be achieved by fusing the nodule-level diagnosis.
3. DeepLung Framework
Our fully automated lung CT cancer diagnosis system
consists of two parts: nodule detection and classification.
We design a 3D Faster R-CNN for nodule detection, and
propose GBM with deep 3D DPN features, raw nodule CT
pixels and nodule size for nodule classification.
3.1. 3D Faster R-CNN with Deep 3D Dual Path Net
for Nodule Detection
Inspired by the success of dual path network on the ImageNet [3, 4], we design a deep 3D DPN framework for lung
CT nodule detection and classification in Fig. 3 and Fig.
4. Dual path connection benefits both from the advantage
of residual learning and that of dense connection [11, 12].
The shortcut connection in residual learning is an effective
way to eliminate vanishing gradient phenomenon in very
deep networks. From a learned feature sharing perspective,
residual learning enables feature reuse, while dense connection has an advantage of exploiting new features [3]. Additionally, densely connected network has fewer parameters
than residual learning because there is no need to relearn
redundant feature maps. The assumption of dual path con-
nection is that there might exist some redundancy in the exploited features. And dual path connection uses part of feature maps for dense connection and part of them for residual learning. In implementation, the dual path connection
splits its feature maps into two parts. One part, F(x)[d :],
is used for residual learning, the other part, F(x)[: d], is
used for dense connection as shown in Fig. 2. Here d is a
hyper-parameter for deciding how many new features to be
exploited. The dual path connection can be formulated as
y = G([x[: d], F(x)[: d], F(x)[d :] + x[d :]),
(1)
where y is the feature map for dual path connection, G is
used as ReLU activation function, F is convolutional layer
functions, and x is the input of dual path connection block.
Dual path connection integrates the advantages of the two
advanced frameworks, residual learning for feature reuse
and dense connection for the ability to exploit new features,
into a unified structure which obtained success on the ImageNet dataset[4]. We design deep 3D neural nets based on
3D DPN because of its compactness and effectiveness.
The 3D Faster R-CNN with a U-net-like encoderdecoder structure and 3D dual path blocks is illustrated in
Fig. 3. Due to the GPU memory limitation, the input of 3D
Faster R-CNN is cropped from 3D reconstructed CT images with pixel size 96 × 96 × 96. The encoder network
is derived from 2D DPN [3]. Before the first max-pooling,
two convolutional layers are used to generate features. After that, eight dual path blocks are employed in the encoder
subnetwork. We integrate the U-net-like encoder-decoder
design concept in the detection to learn the deep nets efficiently [22]. In fact, for the region proposal generation, the
3D Faster R-CNN conducts pixel-wise multi-scale learning
and the U-net is validated as an effective way for pixel-wise
labeling. This integration makes candidate nodule generation more effective. In the decoder network, the feature
maps are processed by deconvolution layers and dual path
blocks, and are subsequently concatenated with the corresponding layers in the encoder network [34]. Then a convolutional layer with dropout (dropout probability 0.5) is used
in the second to the last layer. In the last layer, we design 3
anchors, 5, 10, 20, for scale references which are designed
based on the distribution of nodule sizes. For each anchor,
there are 5 parts in the loss function, classification loss Lcls
for whether the current box is a nodule or not, regression
loss Lreg for nodule coordinates x, y, z and nodule size d.
If an anchor overlaps a ground truth bounding box with
the intersection over union (IoU) higher than 0.5, we consider it as a positive anchor (p? = 1). On the other hand,
if an anchor has IoU with all ground truth boxes less than
0.02, we consider it as a negative anchor (p? = 0). The
multi-task loss function for the anchor i is defined as
L(pi , ti ) = λLcls (pi , p?i ) + p?i Lreg (ti , ti ? ),
(2)
Figure 3. The 3D Faster R-CNN framework contains 3D dual path
blocks and a U-net-like encoder-decoder structure. We design 26
layers 3D dual path network for the encoder subnetwork. The
model employs 3 anchors and multi-task learning loss, including
coordinates (x, y, z) and diameter d regression, and candidate box
classification. The numbers in boxes are feature map sizes in the
format (#slices*#rows*#cols*#maps). The numbers above the
connections are in the format (#filters #slices*#rows*#cols).
Figure 4. The deep 3D dual path network framework in the nodule
classification subnetwork, which contains 30 3D dual path connection blocks. After the training, the deep 3D dual path network
feature is extracted for gradient boosting machine to do nodule
diagnosis. The numbers are of the same formats as Fig. 3.
where pi is the predicted probability for current anchor i
being a nodule, ti is the predicted relative coordinates for
nodule position, which is defined as
x − xa y − ya z − za
d
ti = (
,
,
, log( )),
da
da
da
da
(3)
where (x, y, z, d) are the predicted nodule coordinates and
diameter in the original space, (xa , ya , za , da ) are the coordinates and scale for the anchor i. For ground truth nodule
position, it is defined as
t?i = (
x? − xa y ? − ya z ? − za
d?
,
,
, log( )),
da
da
da
da
(4)
where (x? , y ? , z ? , d? ) are nodule ground truth coordinates
and diameter. The λ is set as 0.5. For Lcls , we used binary
cross entropy loss function. For Lreg , we used smooth l1
regression loss function [9].
3.2. Gradient Boosting Machine with 3D Dual Path
Net Feature for Nodule Classification
For CT data, advanced method should be effective to extract 3D volume feature [33]. We design a 3D deep dual
path network for the 3D CT lung nodule classification in
Fig. 4. The main reason we employ dual modules for detection and classification is that classifying nodules into benign and malignant requires the system to learn finer-level
features, which can be achieved by focusing only on nodules. In addition, it allows to introduce extra features in the
final classification. We first crop CT data centered at predicted nodule locations with size 32 × 32 × 32. After that,
a convolutional layer is used to extract features. Then 30
3D dual path blocks are employed to learn higher level features. Lastly, the 3D average pooling and binary logistic
regression layer are used for benign or malignant diagnosis.
The deep 3D dual path network can be used as a classifier for nodule diagnosis directly and it can also be employed to learn effective features. We construct feature by
concatenating the learned deep 3D DPN features (the second from the last layer (2,560 dimension)), nodule size, and
raw 3D cropped nodule pixels. Given complete and effective features, GBM is a superb method to build an advanced
classifier [8]. We validate the feature combining nodule size
with raw 3D cropped nodule pixels in combination with the
GBM classifier and obtained 86.12% average test accuracy.
Lastly, we employ GBM with the constructed feature and
achieve the best diagnosis performance.
3.3. DeepLung System: Fully Automated Lung CT
Cancer Diagnosis
The DeepLung system includes the nodule detection using the 3D Faster R-CNN and nodule classification using
GBM with constructed feature (deep 3D dual path features,
nodule size and raw nodule CT pixels) as shown in Fig. 1.
Due to the GPU memory limitation, we first split the
whole CT into several 96 × 96 × 96 patches, process them
through the detector, and combine the detected results together. We only keep the detected boxes of detection probabilities larger than 0.12 (threshold as -2 before sigmoid
function). After that, non-maximum suppression (NMS) is
adopted based on detection probability with the intersection
over union (IoU) threshold as 0.1. Here we expect to not
miss too many ground truth nodules.
After we get the detected nodules, we crop the nodule
with the center as the detected center and size of 32 × 32 ×
32. The detected nodule size is kept as a feature input for
later downstream classification. The deep 3D DPN is employed to extract features. We use the GBM and construct
features to conduct diagnosis for the detected nodules. For
pixel feature, we use the cropped size of 16 × 16 × 16
and center as the detected nodule center in the experiments.
For patient-level diagnosis, if one of the detected nodules is
positive (cancer), the patient is classified as having cancer.
Conversely, if all detected nodules are negative, the patient
is considered non-cancer.
4. Experiments
We conduct extensive experiments to validate the
DeepLung system. We perform 10-fold cross validation using the detector on LUNA16 dataset. For nodule classification, we use the LIDC-IDRI annotation, and employ the
LUNA16’s patient-level dataset split. Finally, we also validate the whole system based on the detected nodules both
on patient-level diagnosis and nodule-level diagnosis.
In the training, for each model, we use 150 epochs in
total with stochastic gradient descent optimization and momentum as 0.9. The batch size parameter is limited by GPU
memory. We use weight decay as 1 × 10−4 . The initial
learning rate is 0.01, 0.001 after half the total number of
epoch, and 0.0001 after epoch 120.
4.1. Datasets
LUNA16 dataset is a subset of the largest publicly available dataset for pulmonary nodules, LIDC-IDRI [2, 24].
LUNA16 dataset only has the detection annotations, while
LIDC-IDRI contains almost all the related information for
low-dose lung CTs including several doctors’ annotations
on nodule sizes, locations, diagnosis results, nodule texture,
nodule margin and other informations. LUNA16 dataset
removes CTs with slice thickness greater than 3mm, slice
spacing inconsistent or missing slices from LIDC-IDRI
dataset, and explicitly gives the patient-level 10-fold cross
validation split of the dataset. LUNA16 dataset contains
888 low-dose lung CTs, and LIDC-IDRI contains 1,018
low-dose lung CTs. Note that LUNA16 dataset removes
the annotated nodules of size smaller than 3mm.
For nodule classification, we extract nodule annotations
from LIDC-IDRI dataset, find the mapping of different doctors’ nodule annotations with the LUNA16’s nodule annotations, and obtained the ground truth of nodule diagnosis by
averaging different doctors’ diagnosis (discarding 0 score
for diagnosis which corresponds to N/A.). If the final average score is equal to 3 (uncertain about malignant or benign), we remove the nodule. For the nodules with score
greater than 3, we label them as positive. Otherwise, we
label them as negative. Because CT slides were annotated
by anonymous doctors, the identities of doctors (referred to
as Drs 1-4 as the 1st-4th annotations) are not strictly consistent. As such, we refer them as “simulated” doctors. To
make our results reproducible, we only keep the CTs within
LUNA16 dataset, and use the same cross validation split as
LUNA16 for classification.
4.2. Preprocessing
Three automated preprocessing steps are employed for
the input CT images. First, we clip the raw data into
[−1200, 600]. Second, we transform the range linearly
into [0, 1]. Finally, we use LUNA16’s given segmentation
ground truth and remove the background.
Figure 5. The 3D Faster R-CNN network with 3D residual blocks.
It contains several 3D residual blocks. We employ a deep 3D residual network of 18 layers as the encoder subnetwork, which is an
extension from 2D Res18 net [11].
4.3. DeepLung for Nodule Detection
We train and evaluate the detector on LUNA16 dataset
following 10-fold cross validation with given patient-level
split. In training, we augment the dataset by randomly flipping the image and use cropping scale betweeb 0.75 to 1.25.
The evaluation metric, FROC, is the average recall rate at
the average number of false positives at 0.125, 0.25, 0.5,
1, 2, 4, 8 per scan, which is the official evaluation metric
for LUNA16 dataset [24]. In the test phase, we use detection probability threshold as -2 (before sigmoid function),
followed by NMS with IoU threshold as 0.1.
To validate the performance of proposed deep 3D dual
path network for detection, we employ a deep 3D residual
network as a comparison in Fig. 5. The encoder part of
this baseline network is a deep 3D residual network of 18
layers, which is an extension from 2D Res18 net [11]. Note
that the 3D Res18 Faster R-CNN contains 5.4M trainable
parameters, while the 3D DPN26 Faster R-CNN employs
1.4M trainable parameters, which is only 14 of 3D Res18
Faster R-CNN.
The FROC performance on LUNA16 is visualized in
Fig. 6. The solid line is interpolated FROC based on true
prediction. The 3D DPN26 Faster R-CNN achieves a FROC
score of 84.2% without any false positive nodule reduction stage, which is better than the previous 83.9% using
two-stage training [6]. The 3D DPN26 Faster R-CNN using only 14 of the parameters performs better than the 3D
Res18 Faster R-CNN, which demonstrates the superior suitability of the 3D DPN for detection. Ding et al. obtains
89.1% FROC using 2D Faster R-CNN followed by extra
false positive reduction classifier [5], while we only employ
enhanced Faster R-CNN with deep 3D dual path for detection. We have recently applied the 3D model to Alibaba
Tianchi Medical AI on nodule detection challenge and were
able to achieve top accuracy on a hold-out dataset.
4.4. DeepLung for Nodule Classification
We validate the nodule classification performance of
the DeepLung system on the LIDC-IDRI dataset with the
LUNA16’s split principle, 10-fold patient-level cross vali-
Table 2. Nodule-level diagnosis accuracy (%) between nodule
classification subnetwork in DeepLung and experienced doctors
on doctor’s individually confident nodules.
Doctors
DeepLung
Dr 1
93.44
93.55
Dr 2
93.69
93.30
Dr 3
91.82
93.19
Dr 4
86.03
90.89
Average
91.25
92.74
Table 3. Statistical property of predicted malignant probability for
borderline nodules (%)
Prediction
Frequency
Figure 6. Sensitivity (Recall) rate with respect to false positives per
scan. The FROC (average recall rate at the false positives as 0.125,
0.25, 0.5, 1, 2, 4, 8) of 3D Res18 Faster R-CNN is 83.4%, while
the FROC of 3D DPN26 Faster R-CNN is 84.2% with only 14 of
the parameters as 3D Res18 Faster R-CNN. The 3D Res18 Faster
R-CNN has a total recall rate 94.6% for all the detected nodules,
while 3D DPN26 Faster R-CNN has a recall rate 95.8%.
Models
Multi-scale CNN [26]
Slice-level 2D CNN [33]
Nodule-level 2D CNN [33]
Vanilla 3D CNN [33]
Multi-crop CNN [27]
Deep 3D DPN
Nodule Size+Pixel+GBM
All feat.+GBM
Accuracy (%)
86.84
86.70
87.30
87.40
87.14
88.74
86.12
90.44
Year
2015
2016
2016
2016
2017
2017
2017
2017
dation. There are 1,004 nodules of which 450 are positive.
In the training, we first pad the nodules of size 32 × 32 × 32
into 36 × 36 × 36, randomly crop 32 × 32 × 32 from the
padded data, horizontal flip, vertical flip, z-axis flip the data
for augmentation, randomly set 4 × 4 × 4 patch to zero, and
normalize the data with the mean and standard deviation
obtained from training data. The total number of epochs is
1,050. The initial learning rate is 0.01, and reduce to 0.001
after epoch 525, and finally to 0.0001 after epoch 840. Due
to time and resource limitation for training, we use the fold
1, 2, 3, 4, 5 for test, and the final performance is the average
performance on the five test folds. The nodule classification
performance is concluded in Table 1.
From the table 1, our deep 3D DPN achieves better performance than those of Multi-scale CNN [26], Vanilla 3D
CNN [33] and Multi-crop CNN [27], because of the strong
power of 3D structure and deep dual path network. GBM
with nodule size and raw nodule pixels with crop size as
16 × 16 × 16 achieves comparable performance as multiscale CNN [26] because of the superior classification per-
< 0.2 or
> 0.8
80.14
< 0.3 or
> 0.7
89.75
< 0.4 or
> 0.6
94.80
formance of GBM. Finally, we construct feature with deep
3D dual path network features, 3D Faster R-CNN detected
nodule size and raw nodule pixels, and obtain 90.44% accuracy, which shows the effectiveness of deep 3D dual path
network features.
4.4.1
Table 1. Nodule classification comparisons on LIDC-IDRI dataset.
< 0.1 or
> 0.9
64.98
Compared with Experienced Doctors on Their
Individual Confident Nodules
We compare our predictions with those of four “simulated”
experienced doctors on their individually confident nodules
(with individual score not 3). Note that about 1/3 annotations are 3. Comparison results are concluded in Table 2.
From Table 2, these doctors’ confident nodules are easy
to be diagnosed nodules from the performance comparison
between our model’s performances in Table 1 and Table
2. To our surprise, the average performance of our model
is 1.5% better than that of experienced doctors even on
their individually confident diagnosed nodules. In fact, our
model’s performance is better than 3 out of 4 doctors (doctor 1, 3, 4) on the confident nodule diagnosis task. The result validates deep network surpasses human-level performance for image classification [11], and the DeepLung is
better suited for nodule diagnosis than experienced doctors.
We also employ Kappa coefficient, which is a common
approach to evaluate the agreement between two raters, to
test the agreement between DeepLung and the ground truth
[16]. The kappa coefficient of DeepLung is 85.07%, which
is significantly better than the average kappa coefficient of
doctors (81.58%). To evaluate the performance for all nodules including borderline nodules (labeled as 3, uncertain
between malignant and benign), we compute the log likelihood (LL) scores of DeepLung and doctors’ diagnosis. We
randomly sample 100 times from the experienced doctors’
annotations as 100 “simulated” doctors. The mean LL of
doctors is -2.563 with a standard deviation of 0.23. By
contrast, the LL of DeepLung is -1.515, showing that the
performance of DeepLung is 4.48 standard deviation better
than the average performance of doctors, which is highly
statistically significant. It is important to analysis the sta-
Table 4. Comparison between DeepLung’s nodule classification
on all detected nodules and doctors on all nodules.
Method
Acc. (%)
TP Set
81.42
FP Set
97.02
Doctors
74.05-82.67
tistical property of predictions for borderline nodules that
cannot be conclusively classified by doctors. Interestingly,
64.98% of the borderline nodules are classified to be either
malignant (with probability > 0.9) or benign (with probability < 0.1) in Table 3. DeepLung classified most of the
borderline nodules of malignant probabilities closer to zero
or closer to one, showing its potential as a tool for assisted
diagnosis.
4.5. DeepLung for Fully Automated Lung CT Cancer Diagnosis
We also validate the DeepLung for fully automated lung
CT cancer diagnosis on the LIDC-IDRI dataset with the
same protocol as LUNA16’s patient-level split. Firstly, we
employ our 3D Faster R-CNN to detect suspicious nodules.
Then we retrain the model from nodule classification model
on the detected nodules dataset. If the center of detected
nodule is within the ground truth positive nodule, it is a positive nodule. Otherwise, it is a negative nodule. Through
this mapping from the detected nodule and ground truth
nodule, we can evaluate the performance and compare it
with the performance of experienced doctors. We adopt the
test fold 1, 2, 3, 4, 5 to validate the performance the same
as that for nodule classification.
Different from pure nodule classification, the fully automated lung CT nodule diagnosis relies on nodule detection.
We evaluate the performance of DeepLung on the detection true positive (TP) set and detection false positive (FP)
set individually in Table 4. If the detected nodule of center
within one of ground truth nodule regions, it is in the TP
set. If the detected nodule of center out of any ground truth
nodule regions, it is in FP set. From Table 4, the DeepLung
system using detected nodule region obtains 81.42% accuracy for all the detected TP nodules. Note that the experienced doctors obtain 78.36% accuracy for all the nodule
diagnosis on average. The DeepLung system with fully automated lung CT nodule diagnosis still achieves above average performance of experienced doctors. On the FP set,
our nodule classification subnetwork in the DeepLung can
reduce 97.02% FP detected nodules, which guarantees that
our fully automated system is effective for the lung CT cancer diagnosis.
4.5.1
Compared with Experienced Doctors on Their
Individually Confident CTs
We employ the DeepLung for patient-level diagnosis further. If the current CT has one nodule that is classified as
Table 5. Patient-level diagnosis accuracy(%) between DeepLung
and experienced doctors on doctor’s individually confident CTs.
Doctors
DeepLung
Dr 1
83.03
81.82
Dr 2
85.65
80.69
Dr 3
82.75
78.86
Dr 4
77.80
84.28
Average
82.31
81.41
positive, the diagnosis of the CT is positive. If all the nodules are classified as negative for the CT, the diagnosis of
the CT is negative. We evaluate the DeepLung on the doctors’ individually confident CTs for benchmark comparison
in Table 5.
From Table 5, DeepLung achieves 81.41% patient-level
diagnosis accuracy. This is 99% of the average performance of four experienced doctors and better than Dr 4 altogether. This performance gives confidence that DeepLung
can be a useful tool to assist doctors’ in their diagonsis. We further validate our method against the four doctors’ individual confidential CTs. The Kappa coefficient
of DeepLung is 63.02%, while the average Kappa coefficient of the doctors is 64.46%. It implies the predictions
of DeepLung are of good agreement with ground truths for
patient-level diagnosis, and are comparable with those of
experienced doctors.
5. Discussion
In this section, we will argue the utility of DeepLung by
visualizing the nodule detection and classification results.
5.1. Nodule Detection
We randomly pick nodules from test fold 1 and visualize them in red circles in the first row of Fig. 7. Detected
nodules are visualized in blue circles of the second row. Because CT is 3D voxel data, we can only plot the central slice
for visualization. The third row shows the detection probabilities for the detected nodules. The central slice number
is shown below each slice. The diameter of the circle is
relative to the nodule size.
From the central slice visualizations in Fig. 7, we observe the detected nodule positions including central slice
numbers are consistent with those of ground truth nodules.
The circle sizes are similar between the nodules in the first
row and the second row. The detection probability is also
very high for these nodules in the third row. It shows 3D
Faster R-CNN works well to detect the nodules from test
fold 1.
5.2. Nodule Classification
We also visualize the nodule classification results from
test fold 1 in Fig. 8. We choose nodules that is predicted
right, but annotated incorrectly by some doctors. The first
seven nodules are benign nodules, and the remaining nodules are malignant nodules. The numbers below the fig-
Figure 7. Visualization of central slices for nodule ground truths and detection results. We randomly choose nodules (red circle boxes in
the first row) from test fold 1. Detection results are shown in the blue circles of second row. The center slice numbers are shown below the
images. The last row shows detection probability. The DeepLung performs well for nodule detection.
ures are the DeepLung predicted malignant probabilities
followed by which annotation of doctors is wrong. For the
DeepLung, if the probability is larger than 0.5, it predicts
malignant. Otherwise, it predicts benign. For an experienced doctor, if a nodule is large and has irregular shape, it
has a high probability to be a malignant nodule.
itives. In addition, doctors’ own internal bias may play a
role in how confident he/she predicts these scans while being limited to observing only one slice at a time. Machine
learning-based methods can overcome these limitations and
are able to learn complicated rules and high dimensional
features while utilizing all input slices at once without much
problem. From this perspective, DeepLung can potentially
be of great use to doctors in their effort to make consistent
and accurage diagonsis.
6. Conclusion
Figure 8. Visualization of central slices for nodule classification
results on test fold 1. We choose nodules that are predicted right
by the DeepLung, but annotated incorrectly by some doctors. The
numbers below the nodules are model predicted malignant probabilities followed by which annotation of doctors is wrong. The first
seven nodules are benign nodules. The rest nodules are malignant
nodules. The DeepLung performs well for nodule classification.
From Fig. 8, we can observe that doctors mis-diagnose
some nodules. The reason may be be that humans are not fit
to process 3D CT data which are of low signal to noise ratio. Perhaps some doctors cannot find some weak irregular
boundaries or erroraneously consider some normal tissues
as nodule boundaries leading to false negatives or false pos-
In this work, we propose a fully automated lung CT cancer diagnosis system based on deep learning. DeepLung
consists of two parts, nodule detection and classification.
To fully exploit 3D CT images, we propose two deep 3D
convolutional networks based on 3D dual path networks,
which is more compact and can yield better performance
than residual networks. For nodule detection, we design a
3D Faster R-CNN with 3D dual path blocks and a U-netlike encoder-decoder structure to detect candidate nodules.
The detected nodules are subsequently fed to nodule classification network. We use a deep 3D dual path network
to extract classification features. Finally, gradient boosting machine with combined features are trained to classify
candidate nodules into benign or malignant. Extensive experimental results on public available large-scale datasets,
LUNA16 and LIDC-IDRI datasets, demonstrate the superior performance of the DeepLung system.
References
[1] H. J. Aerts et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nature communications, 5, 2014.
[2] S. G. Armato et al. The lung image database consortium
(lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical physics, 38(2):915–931, 2011.
[3] Y. Chen, J. Li, H. Xiao, X. Jin, S. Yan, and J. Feng. Dual
path networks. In NIPS, 2017.
[4] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. FeiFei. Imagenet: A large-scale hierarchical image database. In
CVPR. IEEE, 2009.
[5] J. Ding, A. Li, Z. Hu, and L. Wang. Accurate pulmonary
nodule detection in computed tomography images using
deep convolutional neural networks. In MICCAI, 2017.
[6] Q. Dou, H. Chen, Y. Jin, H. Lin, J. Qin, and P.-A. Heng.
Automated pulmonary nodule detection via 3d convnets with
online sample filtering and hybrid-loss residual learning. In
MICCAI, 2017.
[7] A. El-Baz et al. 3d shape analysis for early diagnosis of
malignant lung nodules. In IPMI, 2011.
[8] J. H. Friedman. Greedy function approximation: a gradient
boosting machine. Annals of statistics, 2001.
[9] R. Girshick. Fast r-cnn. In ICCV, pages 1440–1448, 2015.
[10] F. Han, G. Zhang, H. Wang, B. Song, H. Lu, D. Zhao,
H. Zhao, and Z. Liang. A texture feature analysis for diagnosis of pulmonary nodules using lidc-idri database. In
Medical Imaging Physics and Engineering. IEEE, 2013.
[11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In CVPR, 2016.
[12] G. Huang, Z. Liu, K. Q. Weinberger, and L. van der Maaten.
Densely connected convolutional networks. In CVPR, 2017.
[13] J. Huang et al. Speed/accuracy trade-offs for modern convolutional object detectors. In CVPR, 2017.
[14] S. Hussein, K. Cao, Q. Song, and U. Bagci. Risk stratification of lung nodules using 3d cnn-based multi-task learning.
In IPMI, 2017.
[15] C. Jacobs et al. Automatic detection of subsolid pulmonary
nodules in thoracic computed tomography images. Medical
image analysis, 2014.
[16] J. R. Landis and G. G. Koch. The measurement of observer
agreement for categorical data. biometrics, 1977.
[17] F. Liao, M. Liang, Z. Li, X. Hu, and S. Song. Evaluate the
malignancy of pulmonary nodules using the 3d deep leaky
noisy-or network. arXiv preprint arXiv:1711.08324, 2017.
[18] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. In CVPR, 2015.
[19] E. Lopez Torres et al. Large scale validation of the m5l lung
cad on heterogeneous ct datasets. Medical physics, 2015.
[20] K. Murphy et al. A large-scale evaluation of automatic pulmonary nodule detection in chest ct using local image features and k-nearest-neighbour classification. Medical image
analysis, 2009.
[21] S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards
real-time object detection with region proposal networks. In
NIPS, 2015.
[22] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015.
[23] A. A. A. Setio et al. Pulmonary nodule detection in ct images: false positive reduction using multi-view convolutional
networks. IEEE TMI, 2016.
[24] A. A. A. Setio et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge.
Medical Image Analysis, 2017.
[25] W. Shen, M. Zhou, F. Yang, D. Dong, C. Yang, Y. Zang, and
J. Tian. Learning from experts: Developing transferable deep
features for patient-level lung cancer prediction. In MICCAI,
2016.
[26] W. Shen, M. Zhou, F. Yang, C. Yang, and J. Tian. Multi-scale
convolutional neural networks for lung nodule classification.
In IPMI, 2015.
[27] W. Shen, M. Zhou, F. Yang, D. Yu, D. Dong, C. Yang,
Y. Zang, and J. Tian. Multi-crop convolutional neural networks for lung nodule malignancy suspiciousness classification. Pattern Recognition, 2017.
[28] K. Suzuki, F. Li, S. Sone, and K. Doi. Computer-aided diagnostic scheme for distinction between benign and malignant
nodules in thoracic low-dose ct by use of massive training
artificial neural network. IEEE TMI, 2005.
[29] Z. Wang et al. Exploring fisher vector and deep networks for
action spotting. In CVPRW, 2015.
[30] Z. Wang et al. Weakly supervised patchnets: Describing and
aggregating local patches for scene recognition. IEEE TIP,
2017.
[31] Z. Wang et al. Structed triplets learning with pos-tag guided
attention for visual question answering. In WACV, 2018.
[32] T. W. Way, L. M. Hadjiiski, B. Sahiner, H.-P. Chan,
P. N. Cascade, E. A. Kazerooni, N. Bogot, and C. Zhou.
Computer-aided diagnosis of pulmonary nodules on ct scans:
Segmentation and classification using 3d active contours.
Medical Physics, 2006.
[33] X. Yan, J. Pang, H. Qi, Y. Zhu, C. Bai, X. Geng, M. Liu,
D. Terzopoulos, and X. Ding. Classification of lung nodule
malignancy risk on computed tomography images using convolutional neural network: A comparison between 2d and 3d
strategies. In ACCV, 2016.
[34] M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus. Deconvolutional networks. In CVPR. IEEE, 2010.
[35] W. Zhu et al. Co-occurrence feature learning for skeleton
based action recognition using regularized deep lstm networks. In AAAI, 2016.
[36] W. Zhu, Q. Lou, Y. S. Vang, and X. Xie. Deep multi-instance
networks with sparse label assignment for whole mammogram classification. In MICCAI, 2017.
[37] W. Zhu, X. Xiang, T. D. T. Tran, G. D. H. Hager, and X. Xie.
Adversarial deep structured nets for mass segmentation from
mammograms. IEEE International Symposium on Biomedical Imaging, 2018.
| 9cs.NE
|
arXiv:quant-ph/0511145v1 15 Nov 2005
Semantics and Simulation
of Communication in
Quantum Programming
Diploma Thesis
Wolfgang Mauerer1
Quantum Information Theory Group
Institute for Theoretical Physics I and
Max Planck Research Group for Optics, Information and Photonics
University Erlangen-Nuremberg, May 2005
Abstract
We present the quantum programming language cQPL which is an extended version of
QPL [Sel04b]. It is capable of quantum communication and it can be used to formulate
all possible quantum algorithms. Additionally, it possesses a denotational semantics based
on a partial order of superoperators and uses fixed points on a generalised Hilbert space
to formalise (in addition to all standard features expected from a quantum programming
language) the exchange of classical and quantum data between an arbitrary number of participants. Additionally, we present the implementation of a cQPL compiler which generates
code for a quantum simulator.
PACS numbers: 03.67.-a, 03.67.Hk, 03.67.Lx, 89.20.Ff
Revision 1.1 (15. September 2005)
1
eMail: [email protected]
Contents
1 Introduction
1
2 Quantum programming with QPL and cQPL
2.1 Programming and quantum physics . . . . . . . . . . .
2.1.1 Computability . . . . . . . . . . . . . . . . . .
2.1.2 Characteristics of quantum computers . . . . .
2.1.3 Static typing, functionality and runtime errors
2.2 Introduction to QPL and cQPL . . . . . . . . . . . . .
2.2.1 Model of computation . . . . . . . . . . . . . .
2.2.2 Language elements . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
3
4
5
5
5
6
Identifiers and variables • Arithmetic and logical expressions • Procedures •
Gates • Control flow • Other features.
2.2.3
Modelling quantum communication with cQPL . . . . . . . . . . . . . 10
Quantum channels • Modules and communication primitives.
2.3
Example programs . . . . . . . . .
2.3.1 Random coin tossing . . . .
2.3.2 Distribution of an EPR pair
2.3.3 Quantum teleportation . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11
12
12
13
3 A compiler for cQPL
3.1 Structure and implementation . . . . . . .
3.1.1 Compiler technique . . . . . . . . .
3.1.2 The implementation . . . . . . . .
3.1.3 Implementation of communication
3.2 Using the compiler . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
15
17
18
18
4 Mathematical structures
4.1 Algebraic structures . . . . . . . . . .
4.1.1 Fundamentals . . . . . . . . . .
4.2 Linear operators . . . . . . . . . . . .
4.2.1 General . . . . . . . . . . . . .
4.2.2 Hilbert-Schmidt operators . . .
4.3 Connection with quantum mechanics .
4.3.1 States and effects . . . . . . . .
4.3.2 Observables . . . . . . . . . . .
4.3.3 Classical components . . . . . .
4.3.4 Composite and hybrid systems
4.4 Domain theory . . . . . . . . . . . . .
4.4.1 Basic definitions . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
21
21
21
21
22
23
23
24
25
25
26
26
i
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4.5
4.4.2 A fixed point theorem . . . . . . . .
4.4.3 Constructions on domains . . . . . .
cp-Maps and their representation . . . . . .
4.5.1 Operator-sum representation . . . .
4.5.2 Equivalence of Kraus operators . . .
4.5.3 A partial order for Kraus operators .
4.5.4 Kraus aggregations . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
27
28
28
29
30
30
31
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
39
39
41
41
41
42
43
45
46
A partial order for Kraus aggregations • Equivalence of Kraus aggregations.
5 Formal denotational semantics
5.1 Fundamentals of denotational semantics . . .
5.2 Survey of QPL . . . . . . . . . . . . . . . . .
5.2.1 Notational conventions . . . . . . . . .
5.2.2 Language elements . . . . . . . . . . .
5.2.3 Semantics . . . . . . . . . . . . . . . .
5.2.4 Limitation: Quantum communication
5.3 Denotational semantics of cQPL . . . . . . .
5.3.1 Formal definitions . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Typing context • Probabilistic environment • Kraus aggregations.
5.3.2
Some examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.3.3
Extension to multipartite systems . . . . . . . . . . . . . . . . . . . . 55
Semantics of sequential programs • Communication with EPR pairs.
Kraus Aggregation • Typing Context • Probabilistic Environment • Quantum
channels.
5.3.4
5.3.5
5.3.6
Existence of fixed points . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Types of interpretational equivalence . . . . . . . . . . . . . . . . . . . 59
Semantics of the language components . . . . . . . . . . . . . . . . . . 60
Some notational remarks • State transformations and fixed points • Classical
operations • Quantum operations.
5.3.7
5.3.8
Explicit transformations of density matrices . . . . . . . . . . . . . . . 78
The type system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Quantum variable tuples • Application of operators • If conditionals and while
loops • Measurements • Communication.
5.4
Avoidance of runtime errors . . . . . . . . . . . . . . .
5.4.1 Unique ownership of qbits . . . . . . . . . . . .
5.4.2 Prevention of cloning and unphysical situations
5.4.3 Unavoidable non-termination conditions . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
81
81
82
82
6 Prospects
85
6.1 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2 Latest developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
A List of symbols
87
B Glossary
89
C Formal syntax
91
Bibliography
93
ii
Dich sah ich, und die milde Freude
Floß von dem süßen Blick auf mich;
Ganz war mein Herz an deiner Seite
Und jeder Atemzug für dich.
Johann Wolfgang von Goethe, Willkommen und Abschied
1
Introduction
Although there is no formal proof that quantum computers offer greater computational
power than classical ones, there are a few quantum computer algorithms which provide
efficient solutions for problems which are up to now believed to be classically NP-hard, i.e.,
they cannot be solved in polynomial time. One of these problems – computing discrete
logarithms – is the cornerstone of basically every modern classical cryptographic algorithm,
so the increased interest in quantum computing is obvious, both from a fundamental and a
practical point of view.
How can quantum algorithms be described in an efficient, readable and precise manner?
Usually, this is done with quantum circuits which are combined into a quantum network, but
especially for larger programs, this is a very cumbersome and error-prone method. Much
research has been performed on the construction, implementation and conception of classical programming languages; therefore, a great amount of alternatives are available. The
situation is totally different in quantum computing: Only a handful of languages have been
proposed so far [AG04, BSC01, Öme98, SP00, Sel04b, vT04], and only two working implementations based on quantum computer simulators are available [Öme98, BSC01] (recently,
a third, but still rough implementation was presented in [AG05]). Both are based on imperative/object oriented languages (C, Pascal, C++, . . . ), with quantum features as extensions,
whereas QPL [Sel04b] models a basically functional language.1
In this work, we present an extension of QPL with abilities to cover quantum communication, i.e., the transmission of quantum mechanical states and exploitation of their highly
non-classical properties like superpositions and entanglement. To distinguish our approach
from QPL, we call it cQPL for communication capable QPL. In contrast to quantum computers of which only some very elementary parts have been experimentally realised until now,
implementations of quantum communication are not only available in several laboratories
around the world, but can even be obtained commercially.
To provide some orientation where our approach can be located in contrast to work done
by other contributors to the field, Table 1.1 presents a comparison of quantum programming
languages and their features.
In a nutshell, the contribution of our work to the field of quantum programming languages
is twofold:
❏ We provide a compiler which can serve as a testbed for new ideas in quantum programming, to teach concepts of quantum algorithms or act as an aid to the intuition
1 We need to note, though, that the term functional should not be overestimated in this context. Important features like higher order functions, which are considered to be key elements of classical functional
languages, are missing in QPL. The most interesting part of the language from a physicist’s point of view is
the ability to guarantee freedom against runtime errors already at compile time, no matter how this goal is
achieved.
1
Chapter 1. Introduction
Reference
New language
Respects physicsa
Implemented
Formal semantics
Communication
Universal
QCL
[Öme98]
✗
✗
✓
✗
✗
✓
Q Language
[BSC01]
✗
✗
✓
✗
✗
✓
qGCL
[SP00]
✓
✗
✗
✓
✓
✓
QML
[AG04]
✓
✓
✓b
✓
✗
✓
QPL
[Sel04b]
✓
✓
✓c
✓
✗
✓
cQPL
✓
✓
✓
✓
✓
✓
a Respecting physics is meant in the sense that it is not possible to syntactically specify programs which
would create unphysical situations; of course, such a state will force the simulation to abort with an error
and needs thus to be avoided if possible.
b A partially finished implementation was available at the time of writing.
c If we consider our cQPL compiler to be a QPL compiler as well.
Table 1.1: Comparison of quantum programming languages defined in other approaches to the problem, their features and their shortcomings.
of users who want to experiment (in the sense of goal oriented playing, not laboratory)
with quantum protocols.
❏ We present an alternate approach to the compositional semantics of QPL and provide
the possibility to include and formalise quantum communication as part of the language. This is a necessary step towards the formalisation of open-world programs, but
may also prove itself useful in fields like quantum process calculi, automated protocol
analysis or similar – cf. Chapter 6 for further prospects.
The layout of this thesis is as follows: In Chapters 2 and 3, we provide an overview
about quantum programming, present the language cQPL and describe the compiler and
its implementation. Chapter 4 presents the mathematical tools and requisites necessary
for a denotational semantics of cQPL, and Chapter 5 develops the semantical description.
Chapter 6 finally provides some short remarks on possible further directions that may be
pursued based on the results of this work. All chapters are interlaced with short introductions
to topics, tools and techniques which are uncommon in physics, but necessary for our work;
Appendices A and B provide a list of symbols and a glossary, respectively. The formal
syntax is presented in Appendix C. To aid the reader in staying on track, we have provided
short summaries in grey boxes at some points along the way.
Since the topic dealt with does not only contain problems of physical nature, but also
touches the fields of computer science and mathematics, we have tried to make the text as
self-contained as possible for readers with any of these backgrounds.2 Obviously it was not
possible to present everything in as much detail as required without repeating the introductory textbooks, but numerous references to the literature are provided which hopefully
alleviates any arising problems.
2 In this context, it is interesting to note that 48% of all entries in the bibliography are of physical nature
and 36% can be counted to computer science; the remaining 16% belong to mathematics or are of general
interest.
2
When someone says, “I want a programming
language in which I need only say what I
wish done,” give him a lollipop.
Alan Perlis, Epigrams on Programming
2
Quantum programming with
QPL and cQPL
This chapter presents an overview about classical and quantum computers, the underlying
models of computation and some principal remarks on quantum programming languages.
Afterwards, a short introduction to QPL and cQPL is given. Note that this chapter is
intentionally kept as terse as possible to allow a more detailed coverage of other topics dealt
with in this thesis. Therefore, no attempt is made to to provide special rigour in this chapter.
2.1
2.1.1
Programming and quantum physics
Computability
Classical computers can be described by several models which are apt for different purposes
(for a more detailed description, cf., e.g., [AB02, Sch01]):
❏ Turing machines
❏ General recursive functions
❏ Register machines
❏ Lambda calculus
❏ Logical gates
❏ Universal programming languages
They can all be brought to a common denominator by showing that they are able to
compute respectively solve the same class of problems. They are computationally equivalent
because one model can be used to efficiently simulate any other model; a Turing machine is
normally taken to be the normative instance among them. The fact that a turing machine
can compute everything which is computable in principle is captured in the Church-Turing
thesis which is one of the fundamental axioms of computer science:1
Hypothesis 2.1.1 (Church-Turing). Every function which would naturally be regarded
as computable can be computed by a Turing machine.
This definition places computability in a purely abstract setting without regard to the
laws of physics. Deutsch [Deu85] realised that if the laws of (quantum) physics are used
as basis for computation, an improved version of the Turing machine might lead to greater
computational power. This requires an extended version of the Church-Turing thesis as well:
1 To be precise, one would have to note that Church’s thesis states that “a function of positive integers
is effectively calculable only if recursive”. This is equivalent to Turing’s thesis, though, and the name
Church-Turing thesis is conventionally used in the literature.
3
Chapter 2. Quantum programming with QPL and cQPL
Hypothesis 2.1.2 (Church-Turing-Deutsch). Every physical process can be simulated
by a universal computing device.
Greater computational power is meant in the sense that there are some problems which
can be efficiently solved by a universal computing device according to Church-TuringDeutsch but which cannot be efficiently solved by a Turing machine, e.g., the simluation cost
for a universal computing device on a normal turing machine would be at least exponential
for the most general case.2
Until now, it has not been proven if quantum Turing machines have greater computational power than Turing machines or not; cf. Refs. [Cle99, Aha98] for further information
on this and related topics.
2.1.2
Characteristics of quantum computers
The basic building block for quantum computers are quantum bits (qbits), i.e., quantum
mechanical objects which can be represented by two different basis states, usually denoted
by |0i and |1i. This can, e.g., be realised by two-spin systems such as electrons, with the
two different polarisations of a photon etc. In contrast to classical machines and models of
computation, quantum computing can draw from two additional resources:
❏ Superpositions: The state of a quantum bit can be in a superposition α |0i + β |1i
with α, β ∈ C and |α|2 + |β|2 = 1. If every qbit of a quantum register consisting of n
elements is brought into a symmetric superposition with α = β = 2−1/2 , the register
contains all numbers from 0 to 2n − 1 at the same time. Manipulations of the register
thus manipulate all these numbers in one step, whereas classically, the number of
required manipulations would grow exponentially with the register size. This feature
is conventionally referred to as quantum parallelism.
❏ Entanglement: Two parts of a quantum system can be in an entangled state such that
manipulations on one part of the system influence the other system although they may
be spatially separated.
While superpositions are useful for computational problems, entanglement is suitable for
tasks like secret key growing, but is for example also necessary to connect input with output
registers for quantum function application.
In general, there are several equivalent models for quantum computers which are abstracted from their physical realisation:
❏ Quantum Turing machines
❏ Quantum gates
❏ Universal quantum programming languages
All these models are equivalent, cf., e.g., Refs.[NC00, Aha98, Pre99] for more detailed explanations.
2 Note that there is a very prominent problem which illuminates this field: Factorising integers is possible
at polynomial cost on a quantum computer, but the currently best known algorithms require exponential cost
on a classical computer. This does not prove, though, that quantum computers are per se more powerful than
classical ones because there is, for example, no proof that there is no classical polynomial-time factorisation
algorithm.
4
2.2. Introduction to QPL and cQPL
2.1.3
Static typing, functionality and runtime errors
Selinger proves in [Sel04b] that QPL has the ability to avoid runtime errors by detecting
them at compile time (and can thus reject the program) which is not possible in other
language proposals presented at the time of writing. He argues that this must be attributed
to the static type system of the language. Although the proof for this is solely based on
properties of the static type system, the functional style of QPL does have its merits in this
respect in our opinion, too: It restricts the language to elements which allow to express
universal programs, but do not allow constructions that can produce runtime errors that
cannot be detected at compile time.3
When work on this thesis was started, one of the points we wanted to investigate was if
and how principles of functional languages could be advantageous for quantum languages.
This did not reach fruition, though, because we could not find a simple way to transfer any
advanced functional method like higher-order functions, recursively defined data types etc.
to the quantum case.4
From a physical perspective, the absence of runtime-errors is much more interesting than
any computer science related question like the type system or functionality. In this thesis,
we thus concentrated on preserving the possibilities of static checking provided by QPL as
far as possible while enhancing usability of the language and providing means to handle
commuication.
2.2
Introduction to QPL and cQPL
This section presents a very short introduction to cQPL; some of the things stated in the
following also hold for a (regularised version) of block QPL5 as presented in [Sel04b] which
is used as basis for cQPL. Note that cQPL is explicitely meant to be an experimental
compiler. Only very modest effort was made to make the compiler easy to use (the error
messages provided can, for example, often only be understood if the user is familiar with the
inner working of the compiler), and only very few classical operations which are considered
standard components of programming languages (e.g., numerical operations like sin, cos etc.)
were implemented to save time since this is purely routine work. Nevertheless, care has been
taken to design the compiler for easy extensibility.6
2.2.1
Model of computation
Although quantum gates are the most widespread model in the literature on quantum algorithms, the QRAM model suggested by Knill [Kni96] is more apt as basis for quantum
programming languages. The model consists of two components: A classical computer for
conventional tasks like program flow control, classical calculations etc., and a quantum memory controlled by the classical computer that can not only store quantum states, but also
apply any unitary operator and perform measurements. Figure 2.1 visualises the approach.
3A
static type system is in general not enough to ensure this property: Just think of the many possible
ways to generate runtime-errors in C, which is statically typed as well. This is partially caused by the fact
that typing in C is weak; nevertheless, even strongly typed languages as, e.g., Java and C# still cannot
prevail runtime errors. Thus, static typing alone is not sufficient to avoid all possible runtime errors.
4 To our knowledge, no fully satisfying mechanism for any of these problems has been found until now
although the topic is addressed in several papers.
5 The variant of QPL with a textual structure that includes blocks; alternatively, there is a textual variant
without blocks and a flow diagram representation.
6 For example, integration of the n-qbit Fourier gate could be accomplished by adding only 8 lines of code
to the compiler and 10 to the runtime library.
5
Chapter 2. Quantum programming with QPL and cQPL
Figure 2.1: Hybrid architecture for a quantum computer which consists of a classical computer and a quantum memory with the ability to apply unitary
operators and perform measurements at the disposal of the classical
system (image taken from [Öme03]).
cQPL is centred on the assumption that the control flow of a program can be described
by classical means. Although this is considered to be a loss of generality by some authors
(most notably [Öme03]), it does not represent a real restriction because all known quantum
algorithms can be expressed in this framework. Data is, of course, quantum mechanical,
and can be modified by unitary operators which are equivalently called gates in analogy to
the gate model of quantum computation.
2.2.2
Language elements
The following remarks provide an introduction to cQPL for ordinary users; a formal version
of the syntax will be presented in Appendix C. The compiler distribution provides some
example programs which demonstrate all features.
2.2.2.1
Identifiers and variables
Identifiers for variables can be denoted by strings consisting of the characters A–Z, a–z,
and 0–9, where no digit must be at the beginning. Allocation of new variables is done with
the operator new:
new int loop := 10;
new qbit b1 := 1;
Note that it is mandatory to provide an initial value both for classical and quantum data
types. By default, the data types bit, int, float, qbit and qint are available.
Assigning values to classical variables is possible using the operator :=:
loop := loop - 1;
6
2.2. Introduction to QPL and cQPL
2.2.2.2
Arithmetic and logical expressions
Arithmetic expressions can be given as in most programming languages (e.g., 7+3*x-5).
The operators +, -, * and / are available, they are overloaded to work with any classical
numerical data type.
Operators resulting in a logical value (true or false) are <, >, <=, >=, = and !=; logical
negation is available using !, and predicates can be combined using & (and) and | (or).
Operator priority is defined as usual in arithmetic and logic; parentheses can be used to
explicitely modify this.
2.2.2.3
Procedures
Procedures can take an arbitrary number of input parameters (including none) which are
specified as name:type tuples. Declarations take the following form:
proc test: a:int, b:bit, c:float, d:qbit {
...
}
Procedure calls are achieved with the keyword call:
(eins, zwei, drei) := call test(a0, a1, a2, b1);
Note that the parameters are passed by value so that modifications of the classical
variables given to the procedure (in this case a0, a1 and a2) are not visible in the caller’s
scope. This means that no matter what test does, these variables have the same values
before and after the procedure is called. This is not the case with b1 because it is a quantum
variable and cannot be cloned to implement call-by-value; any modifications performed by
test on the state ofb1 are visible for the caller after test returns.
The result of a procedure is a tuple containing the values the classical parameters had at
the end of procedure execution; the example stores these in the local variables eins, zwei
and drei. Note that the result of a procedure call may be ignored as well by the caller, so
the following variant is also possible:
call test(a0, a1, a2, b1);
The example program proc test.qpl which accompanies the compiler further illustrates
the described behaviour, so we refer the reader to it.
2.2.2.4
Gates
Gate application is performed using the operator *= as in the following example:
q *= Not;
A small number of elementary gates is built into the language core (remember that additional
ones can be added with really little effort as we already mentioned before):
❏ H Hadamard transformation on a qbit.
❏ FT(n) Fourier transformation on n qbits, i.e., the n-fold tensor product of Hadamard
transforms.
7
Chapter 2. Quantum programming with QPL and cQPL
❏ NOT Logical negation on one qbit.
❏ CNot Controlled Not on two qbits.
❏ Phase Phase shift gate on one qbit; the desired shift is given as parameter.
The dimension of the gate and the destination must match. Variable tuples where
identifiers are combined with commas (,) can be used to combine several quantum variables
as in the following example:
new qbit test1 := 0;
new qbit test2 := 1;
test1, test2 *= CNot;
test1 *= Phase 0.5;
User-defined gates can be defined by enclosing a list of (complex) numbers in [[ and ]]
as in the following example:
test1,test2 *= [[0.5, 0.5,
0.5, 0.5,
0.5, 0.5i, -0.5, -0.5i,
0.5, -0.5,
0.5, -0.5,
0.5, -0.5i, -0.5, 0.5i]];
2.2.2.5
Control flow
If-then-else and While are available for directing the control flow.
new int loop := 10;
while (loop > 5) do {
print loop;
loop := loop - 1;
};
if (loop = 3) then {
print "3";
}
else {
print "Nicht 3";
}
The meaning of these operations is the same as in classical languages.
2.2.2.6
Other features
Several more features which do not fit into any of the above categories are available:
❏ dump takes one or more quantum variable identifiers as argument and provides a dump
of the current probability spectrum in the canonical basis |0i , |1i:
dump eins, zwei;
To demonstrate the effect of this command, consider the following example:
8
2.2. Introduction to QPL and cQPL
new qbit a := 0;
new qbit b := 0;
print "State before FT:"; dump a, b;
a, b *= FT(2);
print "State after FT:"; dump a, b;
The program fragment produces the following output when run:
State before FT: 1 |00>
State after FT: 0.25 |00>, 0.25 |01>, 0.25 |10>, 0.25 |11>
Note that there is a fundamental difference between dumping the state of quantum
variables and measuring the state and printing the result, as the following example
shows:
measure a then { print "a is |0>"; } else { print "a is |1>"; };
print "State of b:"; dump b;
print "State of (a,b):"; dump a,b;
If the fragment above is supplemented by these lines, running the program either yields
this
a is |0>
State of b: 0.5 |0>, 0.5 |1>
State of (a,b): 0.5 |00>, 0.5 |01>
or that output (the output of the statements before is omitted):
a is |1>
State of b: 0.5 |0>, 0.5 |1>
State of (a,b): 0.5 |10>, 0.5 |11>
The result of the command dump b never changes, no matter how often the program is
run. The result of the measure command, however, will change so that every output
appears with probability 0.5 for a large number of executions.
❏ skip does nothing, but can be used to fulfil syntactic requirements if, for example,
one branch of an if-then-else-statement is supposed to do nothing:
if (condition) then {
skip;
}
else {
...
}
❏ print prints the value of a variable or an arithmetic expression (e.g., print 5+7;
print a;), but can also be used to output text strings enclosed in quotation marks
to the console (e.g., print "Hello, world!";).
9
Chapter 2. Quantum programming with QPL and cQPL
❏ measure measures a quantum variable and returns a classical result. The result is
governed by a probability distribution according to the state the quantum variable
is in; thus, successive program runs will in general return different results when the
function is called. Note that the measurement is always performed in the standard
basis |0i , |1i for each contributing qbit.
2.2.3
Modelling quantum communication with cQPL
Although quantum communication has already been implemented with several physical
schemes at the time of writing, cQPL does not consider any of these solutions. Instead,
the model presented in Section 2.2.1 is extended in such a way that the peculiarities of
communication can be replaced by reasonable simulation alternatives. In a real physical
setting, communication between two parties can be implemented by using whatever kind
of quantum channel which allows to transfer quantum states. Since this is not quite easy
to implement experimentally (quantum states are very fragile objects), diverse effects like
channel loss, decoherence etc. need to be taken into account in a real-world setting. This is
accounted for by a channel model which describes these effects mathematically.
2.2.3.1
Quantum channels
Obviously, no spatial transmission of quantum states is performed in the simulation.7 A
quantum channel is thus replaced with a label on each qbit present in the simulation which
denotes the respective owner. Sending a qbit is thus equivalent to changing the label of
it. Note that this definition differs from the definition of a channel which is used in some
contributions to the literature, e.g. [Key02]. Here, a channel is seen as something that modifies the state, for example by decoherence, loss or the influence of eavesdroppers, whereas
our definition captures the notion of a channel as a means of unambiguous quantum state
transfer. Inclusion of eavesdroppers (or any modification of the quantum state) is possible
if a quantum channel is replaced by two quantum channels which are connected by a third
module (which we E for Eve following the usual convention) which receives quantum states
from Alice, performs appropriate modifications and resends the states to Bob. Obviously,
Eve can collect multiple qbits that pass the channel and manipulate them collectively to pursue different attack strategies, so this is in no way a restriction to intercept-resend attacks.
Figure 2.2 provides a visualisation of the communication model.
2.2.3.2
Modules and communication primitives
Communicating systems are specified in terms of modules which contain the code the participants execute. Every module is identified by a unique label which identifies it among the
participants. Note that module definitions must only occur at the top-level; modules must
not contain other module definitions. Two parties which are named Alice and Bob can be
implemented by this structure:
module Alice {
...
};
module Bob {
7 It would be possible, for example, to consider networked computers between which simulated quantum
states can be transferred. This would add no new physical insights to the problem, but only add technical
difficulties, so we did not implement such a scheme.
10
2.3. Example programs
Figure 2.2: Model of communication used in cQPL. Transmission of quantum data
is replaced by a quantum heap shared by the communicating parties;
all qbits have a unique label that identifies to whom they belong at
the moment. Sending and receiving can be modelled by changing the
label, the channel itself is modelled by a third party.
...
};
Channels for sending and receiving are implicitly opened between all participants. The
send command is provided for sending variables. Two parameters need to be supplied: A
list of variables (which may be classical or quantum) and the identifier of the receiver. For
example, the following code can be placed in module Alice:
new qbit q1 := 0;
new qbit q2 := 1;
...
send q1,q2 to Bob;
Receiving works similar; consider for example the following code which might be located
in module Bob:
...
receive var1:qbit, var2:qbit from Alice;
...
// Do something with var1, var2
Note the a receive commands implicitly introduces new variables into the present frame,
in this case var1 and var2. The data type of the received quantities must be specified after
the variable name where a colon is used as separator.
2.3
Example programs
We present three small, but complete examples together with their output generated by
executing them to allow a look at the cQPL syntax without resorting to program fragments.
11
Chapter 2. Quantum programming with QPL and cQPL
2.3.1
Random coin tossing
This simple algorithm puts a quantum bit into a symmetric superposition and measures it
to simulate the effect of tossing a perfect coin:
new qbit q := 0;
q *= H;
measure q then { print "Tossed head"; } else { print "Tossed tail"; };
The output is with equal probability either Tossed head or Tossed tail, obviously.
2.3.2
Distribution of an EPR pair
Distributing EPR pairs is one possible method to establish a secret key between Alice and
Bob that can, e.g., be used for a Vernam cipher. The following cQPL program shows how
to do this for a single EPR pair:
module Alice {
proc createEPR: a:qbit, b:qbit {
b *= Not;
b *= H;
a,b *= CNot;
} in {
new qbit first := 0;
new qbit second := 0;
call createEPR(first, second);
send second to Bob;
measure first then { print "Alice’s qbit is |1>"; }
else { print "Alice’s qbit is |0>"; };
};
};
module Bob {
receive q:qbit from Alice;
measure q then { print "Bob’s qbit is |1>"; }
else { print "Bob’s qbit is |0>"; };
};
Running the program creates one of the following two outputs with equal probability:
Alice’s qbit is |0>
Bob’s qbit is |0>
Alice’s qbit is |1>
Bob’s qbit is |1>
Note that the order of Alice and Bob’s output may be inverted as well because after
and before the send/receive synchronisation, the execution order of the threads representing
Alice and Bob is indeterministic.8
8 To be precise: The execution order depends on how the computer used for the simulation implements
threads and their parallel execution, so it is effectively indeterminate. On machines with at least two CPUs
and assuming that each thread runs on one of them, the indeterminism is real.
12
2.3. Example programs
2.3.3
Quantum teleportation
Quantum teleportation is an algorithm that enables to transfer an unknown quantum state
between two parties if both of them share one part of an EPR state (in this case, |β00 i) and
can communicate classically. An easy calculation as is, for example, given in [NC00, p. 27]
or any other quantum information text shows how this works, so we will not repeat it here.
The implementation in cQPL is as follows:
module Alice {
proc createEPR: a:qbit, b:qbit {
a *= H;
b,a *= CNot; /* b: Control, a: Target */
} in {
new qbit teleport := 0; /* Apply unitary operations to set the qbit
to any other desired state */
new qbit epr1 := 0;
new qbit epr2 := 0;
call createEPR(epr1, epr2);
send epr2 to Bob;
teleport, epr1 *= CNot;
new bit m1 :=
new bit m2 :=
m1 := measure
m2 := measure
/* teleport: Control, epr1: Target */
0;
0;
teleport;
epr1;
/* Transmit the classical measurement results to Bob */
send m1, m2 to Bob;
};
};
module Bob {
receive q:qbit from Alice;
receive m1:bit, m2:bit from Bob;
if (m1 = 1) then {
q *= [[ 0,1,1,0 ]];
};
if (m2 = 1) then {
q *= [[ 1,0,0,-1 ]];
};
/* Apply sigma_x */
/* Apply sigma_z */
/* The state is now teleported */
print "Teleported state:";
dump q;
};
13
14
Denn was wir tun müssen, nachdem wir es
gelernt haben, das lernen wir, indem wir es
tun.
Aristoteles, Nikomachische Ethik
3
A compiler for cQPL
A compiler for cQPL was implemented as part of this thesis; since no quantum computers
are available yet, it is obviously targeted at simulators for such. This chapter presents a very
short overview about the compiler’s implementation and the limitations which arise from the
experimental nature of it. Note that this chapter is intentionally kept as terse as possible;
it is not supposed to be a detailed description of the techniques used in implementing the
compiler nor is its intention to go down to the source code level.
3.1
Structure and implementation
Since the initial intention when work on the thesis started was to examine the aptness
of functional methods in programming languages for quantum computers, we decided to
implement the compiler for cQPL in a functional language as well (observe the remarks in
Section 2.1.3 why functionality was realised not to be the important factor). The choice after
testing several alternatives fell on Objective Caml (OCaml) which is, for example, described
on http://caml.inria.fr. One of the reasons for this choice was that very good automated
generators for lexers and parsers are available as part of the compiler distribution which
considerably reduce routine work. As simulation backend, we used the routines supplied
with Ömer’s QCL compiler [Öme98] because this library was the most advanced one at that
time. In the meanwhile, a number of other libraries appeared and existing ones matured,
so most likely, we would have chosen a different simulation backend now because the QCL
library comes effectively without any documentation for the programmer (only the QCL
compiler built on top of the library is documented) which resulted in quite a few technical
obstacles.
A simple compiler for a simple classical language was implemented as part of this thesis
to provide an example for explaining compiler technique and to get used to the OCaml
language and the associated tool-chain.
3.1.1
Compiler technique
As usual in compiler technique, the work is separated into several passes which are shown
in Figure 3.1.
Lexical analysis is performed by a lexer generated with OCamllex. The lexer transforms
the input stream of characters into a stream of tokens where tokens are, for example, keywords like int, measure and so on. This simplifies the parsing process because it does not
need to deal with a program representation at the level of single characters, but can already
work with less elementary units. A more interesting part is the syntactical analysis where
the parser determines if the program is valid according to the rules given in Section C. The
15
Chapter 3. A compiler for cQPL
Figure 3.1: Structure of the cQPL compiler
syntax for cQPL is specified as a LALR(1) grammar which is a special kind of an LR(1)
grammar which, in turn, is a variant of a context free grammar as introduced in Definition 5.1.1. We do not want to go into more technical details here, but refer to Appendix B
and the usual literature on the topic, e.g.. Refs. [ASU86, App04, GBJL02, WM95] for the
exact definitions of such grammars.
After the syntactical correctness of a program has been ensured, the compiler can analyse
the generated parse tree to infer the meaning of the program and perform compile time
checks. These try to find as many errors in the program as possible to ensure that they
will not appear at runtime and lead to abortion of the program with an error. Such checks
include, for example:
❏ Making sure that all variables were declared before use.
❏ Checking that procedures are called with proper arguments.
❏ Matching the dimension of quantum gates with the dimension of variables they are
applied to.
❏ Ensuring that tuples of quantum variables are disjoint.
Many more checks of this kind can be found in the source code. Some (e.g., the first
mentioned two) appear in classical programming languages as well, whereas others (e.g., the
last mentioned two) are specific to quantum languages or necessary only for our approach.
The goal of the analysis phase is to provide a representation of the program that is augmented
with everything necessary so that the code-generation backend can produce its result directly
from this representation. An additional advantage of this strategy is that the code generation
can be replaced by a variant targeting a different simulator.1 The structure of the annotated
parse tree which is passed from the parser to the code generation backend is documented in
the source code.
1 For a very early version of the compiler, a code generation backend for the Fraunhofer simulator
[RAMK+ 04] was provided as well, but we dropped support for this to not spend too much time on implementation issues.
16
3.1. Structure and implementation
3.1.2
The implementation
We do not want to go into any implementation details here, but just provide an overview
about the source files which constitute the compiler. Every detail of the implementation can
obviously be found there.
❏ lexer.mll is an input file for OCamllex which generates the lexical analyser for QPL
from the rules given in the file.
❏ parser.mly is an input file for OCamlyacc which is used to transfer the rules (and
code) given in the file into a parser for cQPL syntax.
❏ parser defs.ml defines the data structures used by the parser to build the abstract
syntax tree. For every element of the grammar, a separate data structure is defined.
Most of them can be augmented with additional information which is inferred during
analysis, e.g., lists of type conversions.
❏ cqpl.ml plays a twofold role: On the one hand, it implements the user interaction
part which handles command line processing and directs the different passes of the
compiler. On the other hand, it implements the semantic analysis where for every
grammar production, a corresponding function is available to perform the appropriate
checks.
❏ gen qcl.ml implements the code generation backend for the QCL library. The generated code depends on the C++ code in qpl_runtime.{cc,h} which provides the
runtime environment for programs. This handles, for example, quantum memory management and provides implementations of standard operators. qpl runtime template.h contain the implementation of runtime routines which use templates and thus
cannot be compiled to a static object; the routines to implement the communication operations (i.e., thread handling, data exchange and locking) can be found in
qpl\_runtime\_comm.{cc,h}.
❏ type.ml contains the type checker and routines to perform lossy and non-lossy conversion between different data types. Especially interesting here is the conversion
between classical and quantum variables. The routines provided by this file are much
more general than required by the compiler and would in principle be able to handle
mixed data types as well.
❏ stacked env.ml provides a stackable environment for the semantic analysis.
Since the compiler is only one part of this thesis, we deliberately accepted some limitations which would have to be removed for a production quality compiler. None of them
would present a problem in principle, but would require some tedious effort that does not
justify the gain:
❏ Error messages are not detailed, and no effort is made to continue translation as far
as possible or perform error recovery after a mistake was spotted.
❏ No specific error productions are provided. Syntactical errors are therefore in general
reported by the lexer and not properly by the parser.
❏ There is absolutely no optimisation for execution speed.
17
Chapter 3. A compiler for cQPL
❏ Communication was not too intensively tested because we concentrated more on the
formal aspects of this. Additionally, it uses big locks instead of finer-grained solutions
which reduces performance.
❏ There is no set of standard routines or a standard library of any kind which would be
required for real-world applications.
3.1.3
Implementation of communication
Since we do not use a real network of quantum and classical computers to simulate communicating parties, but only a single system, it is necessary to find an approach that emulates
the characteristic properties of real-world systems in the simulation environment. It has
already been shown that the QRAM model can – together with an appropriate labelling
mechanism for the qbits – provide a suitable replacement for a quantum channel. The remaining issue that needs to be addressed is how the model the independent execution of
the parties together with the synchronisation when send/receive operations take place is to
be implemented. This is, obviously, a problem of the backend, so the solution presented is
specific to the QCL backend.
For every module defined in a program, a separate thread2 is created which has access to
the global memory management routines and communication channels. Quantum memory
management is performed by the main process; appropriate locking techniques ensure that
no race conditions can occur when quantum bits are allocated by the modules. Sending
quantum data can in this scheme be performed by exchanging pointers to the respective
positions on the quantum heap; the static analysis of the compiler guarantees that no more
than one process is in possession of a given qbit at a time. Again, appropriate mechanisms in
the form of mutexes are used to provide the required synchronisation between the modules
which is required to implement send and receive operations. For every pair of modules, a
separate bi-directional queue is provided to handle the exchange of quantum and classical
data.
3.2
Using the compiler
As all well behaved programs, the compiler can provide a list of its options:
wolfgang@meitner> ./cqpl --help
Quantum Programming Language v1.0
Usage: qpl [<input>] [<output>] [--debug] [--nonative] [--norun]
[--backend qcl] [--qheap size]
<input>: Input filename
<output>: Output filename
--debug Print debug messages
--backend Simulation backend (Only qcl is supported at the moment)
--nonative Generate only backend code, don’t create a native executable
--norun Do not execute the generated native code
--qheap Size of quantum heap (default: 200 qbits)
2 A thread is – depending on the implementation by the underlying library and the operating system
kernel – a lightweight process that shares most ressources he has with other threads that execute in parallel,
but has its own control flow.
18
3.2. Using the compiler
-help Display this list of options
--help Display this list of options
❏ <input> and <output> provide the names of the input and output file. If no input file
is specified, the compiler reads from the stdin channel. If no output if specified, the
filename of the input file together with some appropriate extension according to the
output mode is used.
❏ --debug reports lots of debug messages. Console output is really noisy with this
option enabled, but provides some insight into what the compiler does. Obviously
mainly useful for debugging the compiler itself.
❏ --nonative specifies that no native code is generated, i.e., the program is not transformed into machine code by the backend. For the QCL backend, this means that
only a C++ version of the quantum program is generated, but the C++ compiler is
not called to generate a native executable.
❏ --norun does not execute the generated program, but only leaves the executable file
which can be run later.
❏ --qheap size sets the size of the quantum heap to size qbits.
The compiler source code will be made available in the near future on http://kerr.
physik.uni-erlangen.de/qit/qpl.html once the remaining changes to make the source
code ready for public distribution have been performed.
19
20
A mathematician is a device for turning
coffee into theorems.
Pál Erdős
4
Mathematical structures
Because our work involves the theoretical parts of physics and computer science, many different notations and conventions enter the game. To eschew obfuscation, it behoves to define
a consistent notation to be used in this work, which we shall do in the following sections. Additionally, this chapter introduces some mathematical structures and their properties which
are required for the description of the denotational semantics of cQPL.
4.1
4.1.1
Algebraic structures
Fundamentals
The concept of an algebra is convenient to cover the properties of quantum mechanics in an
abstract setting, so we remind the reader of two elementary definitions for terms that are
often used sloppily in physics.
Definition 4.1.1 (Algebra). Let K be a field. An associative K-algebra A over K is a
nonempty set A together with three operations called addition +, multiplication × and scalar
multiplication · (the last two operations are usually denoted by juxtaposition of symbols) for
which the following properties hold:
❏ A is a linear space under addition and scalar multiplication.
❏ A is a ring under addition and multiplication.
❏ If r ∈ K and a, b ∈ A, then r · (a × b) = (r · a) × b = a × (r · b).
Definition 4.1.2 (Subalgebra). S ⊆ A is subalgebra of an algebra A if it has the properties
of an algebra and is closed under operations of A.
4.2
4.2.1
Linear operators
General
Much of our work is based on the grounds of linear operators, so we present some fundamental
definitions and cite a theorem from the literature which will be useful for us later on.
Remark 4.2.1. Note that we use only linear operators in this work, so operator is used as
a synonym for linear operator without mentioning this explicitely in the following chapters.
21
Chapter 4. Mathematical structures
Definition 4.2.1 (Linear operator). A linear operator T from a normed space X to
another normed space Y is a linear map from D(T ) ⊆ X (the domain of T ) to Y with the
following property for x, y ∈ D(T ), α, β ∈ K:
T (αx + βy) = αT (x) + βT (y).
(4.1)
Definition 4.2.2 (Bounded operator). An operator is called bounded if ∃C ≥ 0, C ∈ R
so that
||T x|| ≤ C · ||x||
(4.2)
for all x ∈ D(T ) with ||x|| ≤ 1.
Theorem 4.2.1. Let X, Y be normed spaces. For a linear operator T : X → Y , the
following properties are equivalent:
❏ T is continuous in every point of D(T ).
❏ T is continuous at 0.
❏ T is bounded.
Proof. Cf. Ref. [Wei00, Theorem 2.1].
4.2.2
Hilbert-Schmidt operators
Let X, Y be Hilbert spaces. An operator K ∈ B(X, Y ) is called Hilbert-Schmidt-operator
if there exists an orthonormal basis {eα : α ∈ A} (where A is some index set) with
P
2
†
α∈A ||Keα || < ∞. In a more physical notation, this means that tr K K < ∞. This
is obviously fulfilled if K ∈ B(H) and dim(H) < ∞.
Theorem 4.2.2 (Hilbert space of Hilbert-Schmidt operators). For Hilbert-Schmidt
operators K, L of a Hilbert space X to a Hilbert space Y , ||·||HS is a norm on this space
induced by the scalar product
X
hK, LiHS :=
hKeα , Leα i .
(4.3)
α
Physicists generally write this as:
hK, LiHS = tr K † L.
(4.4)
Proof. If K is a Hilbert-Schmidt operator, aK is a Hilbert-Schmidt operator as well for
every a ∈ K. If K, L are HS operators, then for every orthonormal basis {eα }, the following
equation holds:
X
X
2
2
2
(4.5)
||(K + L)eα || ≤ 2 ·
||Keα || + ||Leα || < ∞,
α
α
i.e., K + L is a Hilbert-Schmidt operator as well. By h·, ·i, we denote the scalar product
1/2
in the space of Hilbert-Schmidt operators, and ||K||HS = hK, KiHS (as usual, the scalar
product induces a metric).
A comparison of both spaces can be found in Table 4.1.1
1 Note that there would be many other different choices how the norm for Hilbert spaces could be defined
which all fulfil the required properties of a norm that can, e.g., be found in [AG81].
22
4.3. Connection with quantum mechanics
State
Operator
Norm
Operator
norm
Hilbert space
|f i ∈ H
H → H:pD̂ |xi = |x′ i
||f i| = hf |f i
||D̂||sup = sup |D̂ |f i |
|f i∈H
|f |≤1
Hilbert space of Hilbert-Schmidt operators
D̂ : H → H
Λ : D̂ → D̂√≡ (H → H) → (H → H)
||D̂||HS = tr D† D
||Λ|| = sup Λ(D̂) = sup tr Λ(D̂)† Λ(D̂)
D̂∈H
||D̂||≤1
D̂∈H
||D̂||≤1
Table 4.1: A comparison between general Hilbert spaces and Hilbert spaces with
Hilbert-Schmidt operators as basis.
4.3
Connection with quantum mechanics
Quantum mechanics is one of the most advanced theories in physics; most research performed
today is centred around it. There are many ways to describe the theory mathematically;
for our purposes, an abstract algebraic formulation based on the principles and structures
introduced in the previous part seems to be most apt. In the following, we present a concise
overview about the central elements of the theory, following the structure of the review given
in Ref. [Key02]. More elaborate introductions can be found in the usual textbooks on the
topic, e.g., [Sak94]. For those especially interested on the impact of quantum mechanics on
information theory, Refs. [NC00, Pre99, Key02] can be especially recommended.
Probabilistic processes lie at the very heart of quantum mechanics: Predictions made by
the theory hold only in the sense that probabilities for outcomes of measurements can be
provided. A large number of repeated experiments with the same preparation parameters
results in a distribution that states how many times a certain outcome will appear. Exactly
this distribution can be predicted by theory. The outcome of a single measurement can in
general never be forecast with certainty.
4.3.1
States and effects
Ideally, an experiment resulting in a probability distribution can be carried out by repeating
the following two processes until a sufficient amount of statistics has been gathered:
❏ Preparation of a quantum mechanical state according to some fixed procedure that
can be repeated a sufficient number of times.
❏ Measurement of some observable quantity, e.g., spin, energy, etc. Effects are a special
class of measurements which can result in either the answer “yes” or “no” according
to some probability distribution.
Note that we do not only deal with purely quantum mechanical states, but may also
encounter a mixture between classical and quantum mechanical properties (which are usually
termed hybrid systems) that our formalism must be able to account for. Measurement results
fall into the classical category since gauges in the macroscopic world are used to infer them
from the quantum system.2
Every quantum system can be completely characterised by its observable quantities which
in turn are characterised by self-adjoint operators. These operators form an algebra A as
2 The problem of how measurements of a quantum system are to be interpreted (or even how the whole
process can be described consistently) has been and still is one of the fundamental philosophical problems
of quantum mechanics. We take a pragmatic point of view here and do not consider the problem in greater
detail, but refer to the literature, e.g., [Aul00].
23
Chapter 4. Mathematical structures
introduced in Section 4.1.1; since we do only deal with finite-dimensional Hilbert spaces
here, we can restrict ourselves to subalgebras of B(H), i.e., A ⊂ B(H). A is called the
observable algebra of the system and is often identified with the system itself because it is
possible to deduce all properties of the system from its observable algebra. The dual algebra
of A is denoted by A∗ and is the algebra defined on the dual space.
To capture the notions of state and effect mathematically, we introduce two sets according to the following definition:
S(A) = { ̺ ∈ A∗ ̺ ≥ 0 ∧ ̺(1) = 1 }
E(A) = { A ∈ A A ≥ 0 ∧ A ≤ 1 }
(4.6)
(4.7)
S represents the set of states, while E contains all effects. For every tuple (̺, A) ∈ S×E, there
exists a map (̺, A) → ̺(A) ∈ [0, 1] which gives the probability p = ̺(A) that measuring
an effect A on a (system prepared in the) state ̺ results in the answer “yes”. Accordingly,
the probability for the answer “no” is given by 1 − p. ̺(A) is called the expectation value
of a state A; states are thus defined as expectation value functionals from an abstract point
of view. These expectation value functionals can by uniquely connected with a normalised
trace-class3 operator ̺ such that ̺(A) = tr(̺A). In principle, it would be necessary to
introduce two different symbols for the expectation value functional and the operator, but
for simplicity, we omit this complication.
We have to distinguish between two different kinds of states: Pure and mixed ones. This
is a consequence of the fact that both S and E are convex spaces: For two states ̺1 , ̺2 ∈ S(A)
and λ ∈ R, 0 ≤ λ ≤ 1, the convex combination λ̺1 + (1 − λ)̺2 is also an element of S(A).
The same statement holds for the elements of E(A). This decomposition provides a very nice
insight into the structure of both spaces. Extremal points in this space cannot be written
as a proper convex decomposition, i.e., x = λy + (1 − λ)z → λ = 1 ∨ λ = 0 ∨ x = y = z.
These can be interpreted as follows:
❏ For S(A), they are pure states with no associated classical uncertainty.
❏ For E(A), they describe measurements which do not allow any fuzziness as is, e.g.,
introduced by a detector which detects some property not with certainty, but only up
to some finite error (alas, all real-world detectors).
4.3.2
Observables
Until now, we have only been talking about effects, i.e., “yes”/“no” measurements, but not
about measurements with a more complicated result range which are necessary to describe
general observables. Although we would have to consider an infinite (even uncountable)
number of possible outcomes for a general description of quantum mechanics, it is sufficient
to consider only observables with a finite range for our purposes.4 Such observables are
represented by maps which connect elements x of a finite set R to some effect Ex ∈ E(A);
this in turn gives rise to a probability distribution px = ̺(Ex ). More formally, we can put
it as in the following:
A family E = {Ex },P
x ∈ R of effects Ex ∈ A if called a positive operator valued measurement (POVM) on R if x∈R Ex = 1.
3 For
trace-class operators, the trace is independent of the basis chosen to evaluate the trace.
is justified because quantum computers process states of the type (|0i , |1i)⊗n . Although quantum
computers can possess an arbitrary number of qbits, it is still a fixed and (which is most important) finite
number; additionally, we do not care for any continuous quantum properties of these objects.
4 This
24
4.3. Connection with quantum mechanics
Note that the Ex need not necessarily be projectors, i.e., Ex2 = Ex . Should this nevertheless be the case ∀x, the measurement is called a projective measurement.
Observables of this kind can be described by self-adjoint operators of the underlying
Hilbert space H which can (without any claim of formal correctness or even a proof) be
seen as follows: Every self-adjoint operator A on a Hilbert space H of finite dimension can
(because
P of the spectral theorem, cf., e.g., [AG81, Wei00]) be decomposed into the form
A = λ∈σ(A) λPλ where σ(A) denotes the spectrum of A and Pλ the projectors onto the
P
corresponding eigenspace. The expectation value λ λ̺(Pλ ) of P for a given state ̺ can
equivalently be calculated by ̺(A) = tr(̺A). Since this is the standard way of formulating
the expectation value of an operator, both points of view coincide.
4.3.3
Classical components
Systems consisting solely of quantum components are generally not to be found: At the
latest after a measurement has been performed, classical probabilities need to be accounted
for. Therefore, we need to pay attention to hybrid systems composed from quantum and
classical parts as well. Obviously, we have to orient ourselves along the lines of Section 4.3.1
to provide proper grounding for both possibilities. Consider a finite set X of elementary
events, i.e., all possible outcomes of an experiment. Again, S(A) and E(A) define the set of
states and effects, respectively, but this time, the observable algebra is given by all complex
valued functions from the set X to C as defined by
A = C(X) = { f : X → C } .
By identifying the function f with the operator fˆ given by
X
fˆ =
fx |xi hx|
(4.8)
(4.9)
x∈X
where |xi denotes a fixed orthonormal basis, the probability distribution can be interpreted
as an operator algebra similar to the quantum mechanical case because fˆ is obviously an
element of B(H). Thus, C(X) can be used as an observable algebra A along any other
quantum mechanical or classical constituent of a multipartite composite system.
4.3.4
Composite and hybrid systems
Since quantum mechanical and classical systems can be described with very similar structures, the presented formalism is obviously well suited for the presentation of composite
systems. Let A ⊂ B(H) and B ⊂ B(K) be systems given in terms of their observable
algebras; the composite system is then given by
A ⊗ B ≡ span { A ⊗ B A ∈ A, B ∈ B } .
(4.10)
Three cases for the choice of H, K can be distinguished:
❏ If both systems are quantum, then A ⊗ B = B(H ⊗ K).
❏ If both systems are classical, then A ⊗ B = C(X × Y ) with C as defined by Eqn. 4.8
❏ If A is classical and B is quantum mechanical, we have a hybrid system; the composite
observable algebra is then given by C(X)⊗B(H) which cannot be simplified any further.
Observables are operator-valued functions in this case, as expected.
25
Chapter 4. Mathematical structures
4.4
Domain theory
The definition of a proper and sound mathematical semantics for a programming language
necessitates apt structures which can be used as a solid ground underlying the work. The
method we use – denotational semantics – is conventionally based on semantic domains,
which in turn rely on partial orders and recursion theory. The purpose of this section is to
introduce the elements required for the semantic description of cQPL which will be given in
Chapter 5; more details are available, e.g., in [GS90, Win93].
4.4.1
Basic definitions
Definition 4.4.1 (Partial order). A partial order (P, ⊑)is a set P on which there is a
binary relation ⊑ for which the following properties hold ∀p, q, r ∈ P :
❏ p ⊑ p (reflexive)
❏ p ⊑ q and q ⊑ r ⇒ p ⊑ r (transitive)
❏ p ⊑ q and q ⊑ p ⇒ p = q (antisymmetric)
Definition 4.4.2 (Upper bound). For a partial order (P, ⊑) and a subset X ⊆ P , p ∈ P
is an upper bound of X if and only if ∀q ∈ X : q ⊑ p.
The element p is a least upper bound if:
❏ p is an upper bound of X
❏ For all upper bounds q of X, p ⊑ q
Remark 4.4.1. Note that it follows from the definition that the least upper bound is unique.
Definition 4.4.3 (ω-chain). Let (D, ⊑D ) be a partial order. An ω-chain of the partial
order is an increasing chain d0 ⊑D d1 ⊑D · · · ⊑D dn ⊑ · · · of elements of the partial order.
Note that ω represents the increasing chain of natural numbers N0 .
Definition 4.4.4 (Complete partial order). The partial order (D, ⊑D ) is a complete
partial order (cpo) if it has least upper bounds of all ω-chains, i.e., any increasing chain
{dn |n ∈ ω} of elements in D has a least upper bound ⊔{dn |n ∈ ω}, written as ⊔n∈ω dn .
(D, ⊑D ) is a cpo with bottom if it is a cpo which has a bottom element (often also called
least element) ⊥D for which ⊥D ⊑ d ∀d ∈ D holds.
Definition 4.4.5 (Directed-complete partial order). A partial order (D, ⊑) in which
every directed subset has a supremum is called directed-complete partial order (dcpo).
Definition 4.4.6 (Monotone function). A function f : D → E between cpos D and E
is monotonic if and only if ∀d, d′ ∈ D:
d ⊑ d′ ⇒ f (d) ⊑ f (d′ ).
(4.11)
Definition 4.4.7 (Continuous function). A function f : D → E between cpos D and E
is continuous if and only if it is monotonic and for all chains do ⊑ d1 · · · ⊑ dn ⊑ · · · in D
there holds
G
f (dn ) = f (⊔n∈ω dn ) .
(4.12)
n∈ω
26
4.4. Domain theory
4.4.2
A fixed point theorem
Definition 4.4.8 (Fixed point). Let f : D → D be a continuous function on a cpo D
with bottom ⊥D . A fixed point of f is an element d ∈ D such that f (d) = d. A prefixed
point of f is an element d ∈ D such that f (d) ⊑ d.
Theorem 4.4.1 (Fixed-point theorem). Let f : D → D be a continuous function on a
cpo with a bottom D. Define
G
fix (f ) =
f n (⊥).
(4.13)
n∈ω
Then fix (f ) is a fixed point of f and the least prefixed point of f , i.e.:
❏ f (fix (f )) = fix (f )
❏ If f (d) ⊑ d then fix (f ) ⊑ d
Consequently, fix (f ) is the least fixed point of f .
Proof. It follows from continuity of f that
f (fix (f )) = f (⊔n∈ω f n (⊥)) =
G
f n+1 (⊥)
(4.14)
n∈ω
=
G
⊥ ∪ f n+1 (⊥)|n ∈ ω
n∈ω
= fix (f ) .
=
G
f n (⊥)
(4.15)
n∈ω
(4.16)
Thus fix (f ) is a fixed point because f (fix (f )) = fix (f ) is exactly the required property
of a fixed point (adding ⊥ in step 4.15 is justified because the least upper bound is not
influenced by this). Suppose d is a prefixed point. Certainly, ⊥⊑ d. By monotonicity,
f (⊥) ⊑ f (d). But d is a prefixed point, i.e., f (d) ⊑ d, so f (⊥) ⊑ d, and by induction
f n (⊥) ⊑ d ∀n ∈ ω. Thus, fix (f ) = ⊔n∈ω f n (⊥) ⊑ d
Remark 4.4.2. Note that it is customary to define
G
YD f ≡
f n (⊥)
(4.17)
n∈ω
such that Y D is obviously a function (D → D) → D which maps functions to their fixed
points; it is thus termed the fixed point combinator.
Definition 4.4.9 (Scott topology5 ). 6 Let D be a dcpo. A subset A is called Scott closed
if it is a lower set7 and is closed under suprema of directed subsets.8 Complements of closed
sets are called Scott open; they are the elements of σD , the Scott topology on D.
Theorem 4.4.2. A function f : Dn → Dm is continuous in the sense of definition 4.4.7
(i.e., Scott continuous) if it is topologically continuous with respect to the Scott topology.
Proof. Cf. Ref. [AJ94, Theorem 2.3.4]
The last definition and theorem are quite technical, but we need them for the proof of
Theorem 4.5.4 later on.
5 This
definition is taken from [AJ94].
topological space is a set X together with a collection T of subsets where the empty set and X are in
T , the union of any collection of sets in T is in T and the intersection of any pair of sets in T is also in T .
7 A lower set is a finite, non-empty downward-closed subset of a partial order, i.e., {x|x ⊑ y}.
8 A subset A of a poset is directed if it is nonempty and each pair of elements has an upper bound in A.
6A
27
Chapter 4. Mathematical structures
4.4.3
Constructions on domains
We will use the following constructions on dcpos to create new dcpos:
❏ D1 × D2 × · · · × Dn denotes n-tuples respectively cartesian domains. The weaker-than
relation is defined such that
(x1 , x2 , . . . , xn ) ⊑D1 ×D2 ×···×Dn (y1 , y2 , . . . , yn ) ⇔ xi ⊑Di yi
(4.18)
for i = 1, . . . , n and xi ∈ Di , yi ∈ Di .
❏ D1 ⊗D2 ⊗· · ·⊗Dn represents the smash product which identifies all tuples that contain
one or more ⊥-elements. Example: (d1 , ⊥2 ), (⊥1 , d2 ) and (⊥1 , ⊥2 ) are all identified
with a new bottom element ⊥D1 ⊗D2 for di ∈ Di Formally, the new domain D ⊗ E is
the set
{(x, y) ∈ D × E|x 6= ⊥ and y 6= ⊥} ∪ {⊥D⊗E }.
(4.19)
❏ D1 + D2 + · · · + Dn is the separated sum domain which consists of all elements in Di
together with a new bottom symbol ⊥D1 +D2 +···+Dn (usually abbreviated to ⊥D ).
❏ The coalesced sum D1 ⊕ D2 ⊕ · · · ⊕ Dn is similar to the separated sum, but the new
bottom element ⊥D is gained by identifying all elements d1 + d2 + · · · + dn (di ∈ Di )
which contain one or more of ⊥Di .
Summary
❏ Lifting is the operation that adds bottom element to a domain D; the result is denoted
by D⊥ ; continuity is not influenced by this.
We have provided the basic framework required to build the denotational semantics of cQPL. This framework is composed of two parts: On the one hand, we
need an abstract representation of quantum mechanics to account for the physical properties of the programming language. On the other hand, the concept
of partial orders builds the basis for defining semantic domains, i.e., the space
which will be used to place the equations describing the semantics of cQPL in.
The choice of partial orders for that is especially justified by the fact that fixed
points can be constructively obtained in them. These in turn are required to solve
recursive domain equations that will be needed to give a denotation for several
language constructs of cQPL, most important the communication features.
4.5
cp-Maps and their representation
In quantum mechanics, time evolution is described by transformations of density matrices
with an operator Λ that is called a superoperator [Pre99, NC00, Key02].
Definition 4.5.1 (Superoperator). A superoperator Λ : B(H) → B(H) has the following
properties for all density operators ̺ ∈ D with ̺′ = Λ(̺):
❏ Λ is linear.
❏ ̺† = ̺ ⇒ ̺′† = ̺′ (hermeticity is preserved).
❏ tr ̺′ = 1 if tr ̺ = 1 (trace preserving).
28
4.5. cp-Maps and their representation
❏ Λ⊗1 is semidefinite positive (∀n ∈ N : Λ⊗1n ≥ 0), i.e., Λ is a completely positive map.
In other words, this means that Λ is not only semidefinite positive (̺′ is nonnegative
if ̺ is nonnegative) on HA , but also on any possible extension HA ⊗ HB .
Note that if dissipative processes (e.g., postselection of observed events) are considered,
the second condition is loosened to tr(̺′ ) ≤ 1.
4.5.1
Operator-sum representation
Kraus [Kra83] proved a result about the decomposability of completely positive maps which
is ubiquitous in quantum information theory:
Theorem 4.5.1 (Kraus representation theorem). A superoperator Λ as defined in
PN
Def. 4.5.1 can be written as a partition of 1 = k=1 A†k Ak where Ak are linear operators
acting on the Hilbert space of the system such that
̺′ = Λ(̺) =
N
X
k=1
Ak ̺A†k ∀̺ ∈ D
(4.20)
for any density matrix ̺ that represents a mixed or a pure state.
Proof. Cf. Ref. [NC00, Pre99, Kra83]
To illustrate this representation, consider the situation that the system under consideration is in contact with a much larger environment, a common situation for physical problems.
Together, both systems form a closed quantum system. State transformations in this combined system can be described by a unitary transformation U ∈ U (dim(H) · dim(Henv ))
where H denotes the Hilbert space of the system under consideration and Henv the Hilbert
space of the environment. Assume that the environment is in a pure state |e0 i he0 |.9 The
density operator of the system under consideration after the unitary operation was applied
to the total system can be recovered by tracing out the environment:
̺′ = Λ(̺) = tr(U ̺ ⊗ |e0 i he0 | U † )
X
=
hek | U (̺ ⊗ |e0 i he0 |)U † |e0 i
(4.21)
(4.22)
k
=
X
k
=
X
hek | U |e0 i ̺ he0 | U † |ek i
(4.23)
Ak ̺A†k .
(4.24)
k
In the last step, Ak is defined by Ak ≡ hek | U |e0 i.
Remark 4.5.1. We say that a set of Kraus operators {Ak } implements a cp-map Λ if
P
∀̺ ∈ D : k Ak ̺A†k = Λ(̺). This simplifies the further description.
Theorem 4.5.2. The operation elements of a given superoperator Λ are not unique: If {Ej }
is a set of Kraus operators, then a different set of Kraus operators {Fk } describes the same
9 This assumption holds without loss of generality because it can be shown that a system can be purified
by introducing extra dimensions which do not have any physical consequences.
29
Chapter 4. Mathematical structures
operation if and only if there exists a unitary matrix U ∈ U (n) with n = card({Ek }) (where
card(X) is the cardinality of the set X) such that
X
Ukj Ej .
(4.25)
Fk =
j
Note that the shorter set may be padded with zero elements until the cardinality of both
matches.
Proof. Cf., e.g., Ref. [NC00, Theorem 8.2] or Ref. [Pre99].
Remark 4.5.2. Let {Ak } be a set of Kraus operators that represents the cp-map Λ. Note
that if any number of elements Ai is taken from {Ak }, the set still remains a completely
positive map, but is not trace preserving any more.
Remark 4.5.3. Note that superoperators are elements of B(H) which makes it possible to
apply many theorems of linear operator algebra to superoperators. In fact, superoperators can
be used as elements of a Hilbert space as defined in Section 4.2.2. The distinction between
operators and superoperators in physics is therefore in general superfluous.
Remark 4.5.4. It can be shown that the number of Kraus elements needed to express any
arbitrary completely positive map T : B(H1 ) → B(H2 ) is bounded by dim(H1 ) · dim(H2 ),
confer, e.g., [Pre99, p. 102]).
4.5.2
Equivalence of Kraus operators
The unitary connection between two sets of Kraus operators defined in Equation 4.25 gives
rise to an equivalence relation between such sets. Two sets {Aj } and {Bk } are members of
the same equivalence class if there is a unitary matrix which connects both representations:
n
X
A∼
= B ⇐⇒ ∃U ∈ U (n) : Ai =
Uij Bj with i = 1, . . . , n.
(4.26)
j=1
The set of all sets of Kraus operators inducing the same map Λ is defined in the obvious
way:
n
o
X
K(Λ) ≡ {Ak }
Ak ̺A†k = Λ(̺) ∀̺ ∈ D .
(4.27)
k
If we talk about a set of Kraus operators or simply Kraus operators in the following, we
always mean an arbitrary set which is an element of the equivalence class inducing the same
cp-map (i.e., an element of K(Λ)), but will not mention this explicitly every time.
4.5.3
A partial order for Kraus operators
The Löwner partial order [Löw34] for two density operators A, B : B(H1 ) → B(H2 ) is given
by
A ⊑ B ⇐⇒ (B − A) > 0.
(4.28)
This partial order can be extended to sets of Kraus operators by defining
{Ai } ⊑ {Bi } ⇐⇒∀̺ ∈ D ∀n ∈ N :
X
i
(Bi ⊗ 1n )̺(Bi ⊗ 1n ) −
†
30
X
k
(Ak ⊗ 1n )̺(Ak ⊗ 1n )
†
!
> 0.
(4.29)
4.5. cp-Maps and their representation
Partial orders are often interpreted as approximations: If an element A is weaker than
B (A ⊑ B), then A is said to approximate B. This point of view will come handy when we
consider solutions of fixed point equations in the denotational description.
It is necessary for our work to see that Kraus operators form a complete partial order.
For this, observe first the following theorem:
Theorem 4.5.3. The partial order on all density operators ̺ ∈ D given by the Löwner
partial order ⊑ is complete.
Proof. Cf. Ref. [Sel04b, Proposition 3.6]
From this, we can deduce the required statement:
Theorem 4.5.4. The partial order for cp-maps defined by the extended Löwner partial order
given by Eqn. 4.29 is complete, i.e., it forms a cpo.
Proof. Let {A1 } ⊑ {A2 }, . . . be an increasing chain of topologically continuous (and therefore
monotone because of 4.4.2, 4.4.7 and 4.4.6) Kraus operators. Because of Definition 4.29, the
relation ̺1 ⊑ ̺2 is preserved by applying {A1 }, {A2 } with {A1 } ⊑ {A2 } to ̺1 , ̺2 . An ωchain of density operators is conserved if an increasing chain of Kraus operators is applied to
it. Because of Theorem 4.5.3, the fact that the previous consideration applies to all density
operators in D and the uniqueness of the least upper bound, the extended Löwner partial
order is complete as well.
4.5.4
Kraus aggregations
We mentioned that superoperators applied to density matrices describe quantum mechanical processes. Operations performed one after another can therefore be described by the
consecutive application of the corresponding superoperators:
If the sets
given by
{A1k }
̺′ = Λ1 (̺), ̺′′ = Λ2 (̺′ ) ⇒ ̺′′ = Λ2 (Λ1 (̺))
and
{A2k }
(4.30)
implement Λ1 and Λ2 , then the same state transformation is
̺′′ =
XX
k
†
†
A2k A1l ̺A1l A2k .
(4.31)
l
We call a collection of sets of Kraus operators that are to be applied subsequently an
aggregation of (sets of ) Kraus operators or simply Kraus aggregation; the Kraus sets involved
are written as a list of the form
Γ = {A1k }, {A2k }, . . . , {Ank }
The list Γ gives rise to the following quantum mechanical operation:
XX X
†
†
Γ(̺) = ̺′ =
···
Ankn · · · A2k2 A1k1 ̺A1k1 A2k2 · · · Ankn †
k1
k2
(4.32)
(4.33)
kn
List concatenation is formally described by the operator ◦:
Γ1 = {A1k }, {A2k }, . . . , {Ank },
(4.34)
Γ2 = {Bk1 }, {Bk2 }, . . . , {Bkm }
⇒ Γ1 ◦ Γ2 ≡ {A1k }, {A2k }, . . . , {Ank }, {Bk1 }, {Bk2 }, . . . , {Bkm }
31
(4.35)
(4.36)
Chapter 4. Mathematical structures
i.e., the effect of Γ1 ◦ Γ2 on a state ̺ is the same as if first Γ1 and then Γ2 would have been
applied. Note (since this is a potential source of confusion) that the list is “processed” from
left to right, not from right to left!
A Kraus aggregation can also consist of multiple sub-aggregations which are prefixed by
some scalar. Formally, we use the operator + to denote this:
Γ′ = p 1 · Γ1 + · · · + p n · Γn .
(4.37)
1
1
· {NOT} + · {1}.
2
2
(4.38)
10
P If the pi ∈′ R are to be interpreted as probabilities, the normalisation condition is
n pn ≤ 1. Γ can thus be seen as a formal combination of lists. The interpretation of such
an aggregation is straightforward: With probability pk , the Kraus aggregation Λk is selected
whenever Λ′ acts on a density operator. Obviously, lists of this form are apt to introduce
mixed states into the Kraus list formalism. Consider, for example, the aggregation
∆=
The effect of it is to apply the unconditional not-operation (which maps |0i → |1i and
|1i → |0i and may, for example, be implemented with σ̂x ) with probability 0.5 and to leave
the state unchanged with the same probability. If this aggregation is applied to, e.g., the
following (pure) density operator
̺ = |0i h0| ,
(4.39)
the resulting state is the impure density operator given by
1
1
{NOT}(̺) + {1}(̺)
2
2
1
1
1
1
= |1i h1| + |0i h0| = {|1i} + {|0i}
2
2
2
2
̺′ = ∆(̺) =
(4.40)
(4.41)
which describes an impure mixture between {|0i} and {|1i}.
Remark 4.5.5. Note that we will use Kraus lists prefixed with probabilities to describe
different measurement outcomes when we provide the semantics of cQPL in Chapter 5. The
physical way to think about such operations is to take a density operator ̺ and apply the
Kraus elements for the projective measurements on it; this results in the state
X
X
̺′ =
Mk ̺Mk† =
Em (̺)
(4.42)
k
k
where Mk are the projection operators and Em (̺) ≡ Mk ̺Mk† . The probability to obtain the
measurement outcome k is given by
p(m) = tr(Em (̺)).
(4.43)
The probability factors in Kraus aggregations can be calculated in exactly this way; both
points of view therefore provide the same information.
Note that we allow the pre-factors of the sub-aggregations to depend on parameters which
make the complete aggregation dependent on the disjoint union of the set of parameters used
for the sub-aggregations. This is necessary to describe Kraus aggregations dependent on
probability distributions which are unknown before a initial state is given or the outcomes
10 The sum can be smaller than 1 to account for the possibility of non-termination which will happen with
P
probability 1 − pi . It also allows to describe non trace-preserving effects.
32
4.5. cp-Maps and their representation
of some measurements are known. The following example shows a Kraus aggregation where
the first sublist depends on the parameters a1 and a2 and the second on a1 and a3 ; the
complete aggregation obviously depends on a1 , a2 and a3 :
Γ(a1 , a2 , a3 ) = p(a1 , a2 ) · Γ1 + p(a1 , a3 ) · Γ2 .
(4.44)
For a Kraus aggregation of the most general form (where {ai } denotes the set of parameters
for the ith sub-aggregation) given by
Γ(∪i {ai }) =
X
i
pi ({ai })Γi ,
(4.45)
the normalisation condition is obviously still given by
X
i
pi ({ai }) ≤ 1
(4.46)
which necessitates that 0 ≤ pi ≤ 1 ∀i (this is supposed to hold for all p used in the following).
It is possible to contract Kraus (sub-)aggregations which consist of more than one element
to a shorter form because two Kraus sets {Ai } and {Bi } can be contracted to a new set
{Ck } which describes the subsequent application of both initial sets, as the following simple
calculation shows:
({Ak }, {Bi })(̺) = {Bi }({Ak }(̺)) =
N X
N
X
B̂k Âi ̺†i B̂k†
(4.47)
k=1 i=1
2
=
N
X
Ĉn ̺Ĉn†
(4.48)
n=1
with
Ĉn ≡ B̂⌈n/N ⌉ Ân mod N .
(4.49)
Recall that different set cardinalities can be compensated by adding an appropriate number
of zero operators to the smaller set. Since the calculation is valid ∀̺ ∈ D, the new single
element aggregation {Ci } is a unique replacement for the aggregation {Ai }, {Bi }.
Based on this contraction, it is possible to define a standard representation for Kraus
aggregations which is easier to handle formally when aggregations must, for example, be
compared.
With
n
o
X
P ≡ {pi ({ai })} aik ∈ R∀i, k, 0 ≤ pi ({ai }) ≤ 1,
pi ({ai }) ≤ 1
(4.50)
i
being the set of all possible parametrised probability distributions and
K ≡ { Λ Λ is a cp-map }
(4.51)
being the set of all unparametrised Kraus aggregations contracted to the normal form given
by Eqn. 4.48, we can finally define the set of all possible Kraus aggregations formally by
nX
o
pi Λi {pi } ∈ P ∧ Λk ∈ K .
A≡
(4.52)
i
33
Chapter 4. Mathematical structures
4.5.4.1
A partial order for Kraus aggregations
For a Kraus aggregation of the contracted normal form Λ = {Ck }, the definition for a
partial order can be directly transferred from Equation 4.28. If the aggregation contains
sub-aggregations, ⊑ is formally a function dependent on the parameters of the aggregation:
For Γ1 = Γ1 (A1 , . . . , An ) and Γ2 = Γ2 (B1 , . . . , Bn ), the partial comparison Γ1 ⊑ Γ2 becomes a function (A1 , . . . , An , B1 , . . . , Bn ) → {true, false}, i.e., the comparison depends on
the parameters of both sets of parameters involved. Note that this does not concern Kraus
aggregations where all coefficients have defined scalar values. Basically, the parametrised
comparision is nothing else than a comparision of all elements of an unfolded Kraus aggregation as defined in Section 5.3.6.2 followed by folding everything back afterwards.
4.5.4.2
Equivalence of Kraus aggregations
One possible task of denotational semantics is to decide wether two programs which look
different perform the same actions, i.e., if their semantics coincide. This question is in
general complicated to answer constructively. Nevertheless, it is possible for some cases. We
will consider this problem in more detail in Chapter 5. At this point, we are interested in
the question when two Kraus aggregations are semantically equivalent, i.e., induce the same
physical operations. The method used for this is almost identical to the method used for
Kraus sets. Consider two aggregations Γ1 and Γ2 given in the contracted normal form, i.e.,
Γ1 =
N
X
p1k Λ1k
(4.53)
p2k Λ2k .
(4.54)
k=1
Γ2 =
N
X
k=1
Let Sym(M ) by the symmetric group over the finite set M . Both lists are equivalent if
(but not only if) the following condition holds:
Γ1 ∼
= Γ2 ⇔∃ϕ ∈ Sym([1, . . . , N ])∀k ∈ [1, N ]∀̺ ∈ D :
p1k (A1k )
=
p2P(k) (A2P(k) )
∧
Λ1k (̺)
=
Λ2P(k) (̺).
(4.55)
(4.56)
Note that this equivalence requires that the same Kraus operators are used in both lists; it is
nevertheless possible that a different set of Kraus operators prefixed by another probability
distribution induces the same action. The criterion given here is thus sufficient, but not
necessary.
The set E of all aggregations that are equivalent in this sense can be defined analogous
to Eqn. 4.27:
(4.57)
E(Λ) ≡ { Λi ∈ A Λi ∼
= Λ }.
This definition is not very satisfying from a constructive point of view: There is no
simple way to systematically decide if the effects of two aggregations coincide. This can be
improved by giving an explicit criterion for the equivalence between two Kraus aggregations.
We consider the special case of two lists which are composed of the same operators, but are
ordered differently. This happens, for example, when statements in a program are reordered.
With the method given below, we can thus get a criterion to decide if such reorderings
preserve the semantics of programs which is a very important case.
Unparametrised Kraus aggregations can always be written in the standard form given
by Eqn. 4.48 and are thus equivalent to a Kraus set; this again is equivalent to some cpmap Λ. Because we have seen in Section 4.2.2 that such cp-maps form a Hilbert space, it
34
4.5. cp-Maps and their representation
is reasonable to define a commutator (analogous to the case of regular operators) for two
Hilbert-Schmidt operators Λ1 , Λ2 by setting:
[Λ1 , Λ2 ] ≡ Λ1 Λ2 − Λ2 Λ1 .
(4.58)
The following theorem provides a condition for the identity between a list of operators and a
permutation of it which is based on elementary commutators of the elements. Unfortunately,
this is not a general solution since the effect of the theorem might just be to rephrase the
problem in different terms if the structure of the commutators is not apt.
Theorem 4.5.5. Let A1 , A2 , . . . , An be operators and let ϕ ∈ Sym(n) be a permutation of
the index set. Then the difference between the commuted product Aϕ(1) · Aϕ(2) · · · Aϕ(n) and
A1 · A2 · · · An can be written as11
Aϕ(1) · Aϕ(2) · · · Aϕ(n) = A1 · A2 · · · An +
X
(s,t)
Xs,t · [As , At ] · Ys,t · Zs
(4.59)
where (s,t) runs over all inversions of ϕ, i.e., 1 ≤ t < s ≤ n and ϕ−1 (s) < ϕ−1 (t) and where
Xs,t =
Y
Ys,t =
Aϕ(i) ,
Y
Zs =
Aϕ(i) ,
Y
Ak .
(4.60)
s<k≤n
1≤i≤n
1≤i≤n
ϕ(i)<s, i<ϕ−1 (t)
ϕ(i)<s, i>ϕ−1 (t)
Proof. We prove this statement by induction on the list length. The cases n = 0 and n = 1
are trivial. The induction step n → n + 1 can be seen as follows. Let j ∈ [0, . . . , n + 1] such
that ϕ(j) = n + 1. Then,
Aϕ(1) · Aϕ(2) · · · Aϕ(j) · · · Aϕ(n+1) =
Aϕ(1) · · · Aϕ(j−1) · Aϕ(j+1) · Aϕ(j) · Aϕ(j+2) · · · Aϕ(n+1) +
Aϕ(1) · · · Aϕ(j−1) · [Aϕ(j) , Aϕ(j+1) ] · Aϕ(j+2) · · · Aϕ(n+1) = ... =
X
Xn+1,t · [An+1 , At ] · Yn+1,t
Aϕ(1) · · · Aϕ(j−1) · Aϕ(j+1) · · · Aϕ(n+1) Aϕ(j) +
{z
} | {z } n+1,t
|
I.H.
(4.61)
A(n+1)
where 1 ≤ t < n + 1, ϕ(n + 1) < ϕ−1 (t),
Xn+1,t =
Y
Yn+1,t =
Aϕ(i) ,
Y
Aϕ(i)
1≤i≤n+1
1≤i≤n+1
ϕ(i)<n+1, i<ϕ−1 (t)
ϕ(i)<n+1, i>ϕ−1 (t)
and n + 1 is obviously fixed. The final resulting equation thus resembles exactly the form
given by Eqn. 4.59, but we have not used the induction hypothesis yet. Now, by using the
induction hypothesis, it follows that
Aϕ(1) · · · Aϕ(j−1) · Aϕ(j+1) · · · Aϕ(n+1) = A1 · · · An +
X
(s′ ,t′ )
Xs′ ,t′ · [As′ , At′ ] · Ys′ ,t′ · Zs′ (4.62)
11 This representation (which is much more elegant than the one derived by the author) was provided by
Volker Strehl.
35
Chapter 4. Mathematical structures
where the primed identifiers are defined by 1 ≤ t′ < s′ ≤ n, and ϕ−1 (s′ ) < ϕ−1 (t′ ). By
placing this into the part of Eqn. 4.61 marked by I.H., we see that
X
Xs′ ,t′ · [As′ , At′ ] · Ys′ ,t′ Zs′ · An+1 +
Aϕ(1) · · · Aϕ(n+1) = A1 · · · An +
(s′ ,t′ )
X
(n+1,t)
Xn+1,t · [An+1 , At ] · Yn+1,t
= A1 · · · An An+1 +
= A1 · · · An+1 +
X
(s′ ,t′ )
X
Xs′ ,t′ · [As′ , At′ ] · Ys′ ,t′ Zs′ An+1 +
| {z }
Zs
X
Xn+1,t · [An+1 , At ] · Yn+1,t
(n+1,t)
Xs,t [As , At ]Ys,t Zs
(4.63)
(s,t)
where the unprimed variables are now given by 1 ≤ t < s ≤ 1 and ϕ−1 (s) < ϕ−1 (t); the
condition for k in Zs is now obviously s < k ≤ n + 1. The resulting Equation 4.63 has thus
the form for n + 1 as required by the statement.
To illustrate this theorem (note, additionally, that a little program to calculate all elements of the commutator sum is available), consider the permutation given by
1 2 3 4 5
.
(4.64)
5 2 3 1 4
The inversions (s, t) are all pairs of elements in the permuted list where a bigger element
is on the left side of a smaller element, in this case: (5, 2), (5, 3), (5, 1), (5, 4), (2, 1), (3, 1).
Note that the inversions characterise the list completely, cf., e.g., [Knu98, Section 5.1.1].
The method defined above is a variant of insertion sort which is a standard sorting method,
covered, e.g., in [SF96]. This can be seen by inspecting the conditions imposed by the
products defining X, Y and Z:
❏ For Xs,t , ϕ(i) < s, i < ϕ−1 (t) selects all i such that the corresponding elements in the
permuted list are smaller than the element s of the inversion and are placed on the
left hand side of the element t in the permuted list. For (5, 1), the condition would
select i = 2, 3.
❏ The conditions for Y make sure that again only elements which are smaller than s are
selected. This time, they additionally have to be on the right hand side of t in the
permuted list.
❏ Z specifies all elements which are on the right hand side of s in the unpermuted list.
By applying these rules, we can calculate the following sets for each inversion:
(5, 1) → X : i = 2, 3; Y : i = 5
(5, 2) → Y : i = 3, 4, 5
(5, 3) → X : i = 2; Y : i = 4, 5
(5, 4) → X : i = 2, 3, 4
(3, 1) → Z : k = 4, 5
(2, 1) → Z : k = 3, 4, 5
36
4.5. cp-Maps and their representation
This leads to the following identity that is provided by Eqn. 4.59 (note that we use i instead
of Ai to simplify the notation):
52314 = 12345 + 231[5, 4] + 2[5, 3]14 + [5, 2]314 + 23[5, 1]4 + 2[3, 1]45 + [2, 1]345
= 12345 + 23154 − 23145 + 25314 − 23514 + 52314 − 25314 + 23145−
23145 + 23145 − 21345 + 21345 − 12345
= 12345 − 12345 + 23154 − 23154 + 25314 − 25314 + 23145 − 23145+
21345 − 21345 + 23514 − 23514 + 52314
= 52314
It it also instructive to observe the following two identies because they illuminate the
induction step:
14532 = 12345 + 14[5, 3]2 + 143[5, 2] + 1[4, 3]25 + 13[4, 2]5 + 1[3, 2]45
1432 = 1234 + 1[4, 3]2 + 13[4, 2] + 1[3, 2]4
Remark 4.5.6. Because the proof has only made use of general properties of permutations
and of the definition of the commutator, it is not only applicable to Hilbert-Schmidt-operators
as we need, but also for any other objects fulfilling the mentioned properties.
37
Summary
We have explained how to represent quantum operations by cp-maps and these in
turn by a sum of Kraus operators. The Löwner partial order defined for density
matrices was generalised to Kraus operators; this order is complete and is therefore
a cpo as introduced in the beginning of this chapter. Since the denotational
semantics of cQPL will require lists of Kraus operators, we have introduced Kraus
aggregations to handle this formally. Since it is one of the problems of denotational
semantics to decide whether two given programs are equal or not, we have also
derived general and specific criteria for the equivalence of Kraus aggregations.
38
A map is not the territory.
Alfred Korzybski, Science and Sanity – An
Introduction to Non-Aristotelian Systems
and General Semantics
5
Formal denotational semantics
In this chapter, we are going to define a denotational semantics for cQPL, the communication
capable version of QPL [Sel04b]. Before we get into the details, we will give an overview
about the ideas of denotational semantics in general, present a survey of the denotational
semantics of QPL (because we reuse some ideas for the semantics of cQPL) and show why
the approach of annotation-based QPL must fail for communicating programs.
5.1
Fundamentals of denotational semantics
Denotational semantics is a well-understood standard method of theoretical computer science which is used to assign precise and mathematically sound and rigorous semantics to syntactically specified programs; introductions are, e.g., given in Refs. [Mos90, Win93, Rey98].
In this section, we will try to present an elementary introduction to the field. We align our
description along the lines of [Mos90, Section 1–3].
Computer programs are (usually) specified in the form of a textual description; this
description must follow certain rules defined by a grammar. Usually, context-free grammars
are used for this purpose because they are the most apt choice for that kind of problem.
They are defined as follows regarding to [AB02, Sch01]:
Definition 5.1.1 (Context-free grammar). A context-free grammar G is a four-tuple
(N, T, P, s0 ) where N is a finite set of nonterminal symbols, T is a finite set of terminal
symbols with T ∩ N = ∅, P ⊆ N × (N ∪ T )∗ is a finite set of productions and s0 ∈ N is the
start symbol.
As a very simple example, consider a grammar for binary strings of the form 0, 01,
100110, . . . which is recursively given by1
B ::= ’0’ | ’1’ | B’0’ | B’1’.
(5.1)
The terminal symbols2 are 0 and 1, the non-terminal3 symbol is B, and the start symbol is
1 In general, one has to distinguish between abstract and concrete syntax respectively grammars defining
these. The latter is used to specify a representation of programs that can be processed with parsers;
for that, some syntactical elements for disambiguation of certain constructions needs to be introduced.
Additionally, the capabilities and, especially, limitations of different parsing techniques need to be considered
when specifying a concrete grammar. Abstract syntax, on the other hand, is a representation of a program
stripped down to the bare minimum that is able to include all available information; additionally, the
structure of the syntax can be chosen such that it is not most suited for parsing, but for further analysis and
processing of the program. Usually, data structures in the form of trees are used to realise abstract syntax.
2 A constant symbol which cannot be resolved any further, cf. Appendix B.
3 A symbol whose definition consists of a chain of terminal and (at least one) non-terminal symbols and
can thus be resolved further.
39
Chapter 5. Formal denotational semantics
obvious because there is only one non-terminal. The productions are defined by Equation 5.1;
explicitely, they are given by{B × 0, B × 1, B × B0, B × B1}.
This grammar defines the syntactical representation of binary numerals. The really
interesting thing, however, is not how numerals look like, but instead what they mean – in
other words, the semantics of numerals. Obviously, the meaning of a binary numeral is some
natural number, so finding semantics for a binary string is equivalent with constructing a
method which assigns the appropriate natural number to a given syntactical representation
of a binary numeral. The constitutional parts of which the grammar is made up of are called
phrases. In our case, these are given by the strings 0 and 1 and the productions B0 and B1.
Denotational semantics assigns a meaning to sentences constructed according to a given
grammar by assigning a meaning to every elementary phrase of a grammar. The meaning of
phrases which are constructed from multiple sub-phrases (e.g., B0 in the example grammar)
is given by the meaning of these sub-phrases. The meaning of a complete program is thus
determined by the meaning of its constituents. The denotational approach is – in a nutshell
– characterised by the following points:
❏ Denotational semantics assigns some appropriate semantic object to every phrase of
the grammar; the object is called the denotation of the phrase.
❏ Valuation functions are used to connect syntactical objects with their semantical counterparts. For example, BIN is a valuation function that maps text strings consisting
of a series of ’0’ and ’1’ to a natural number.
❏ The denotation of compound phrases must only depend on the denotations of the
sub-phrases, i.e., JA1 , . . . , An K = f (JA1 K, . . . , JAn K). This is also known as the compositionality principle.
The valuation functions for binary numerals can be represented by the following equations:
BIN J0K
BIN JB0K
=
=
0
2 · (BIN JBK)
BIN J1K
BIN JB1K
=
=
1
(2 · (BIN JBK)) + 1
The double brackets JK are used to distinguish between the realms of syntax and semantics, while the valuation function BIN is used to map the phrases in these brackets to natural
numbers, their denotations. Thus, the domain of this function is the semantic domain N.
In Chapter 4, the required material for the specification of domains suitable to support the
denotational semantics of cQPL has been presented; it will be put to use in this chapter.
Especially note that the denotations of the composite phrases B0 and B1 are defined only in
terms of the denotations of their sub-phrases as required by the compositionality principle.
To clarify the effect of the denotational definitions, consider how the meaning of the
numeral 101 is denoted; the abstract syntax generates the tree shown in Figure 5.1 as
representation. This leads to the following denotation (observe that the Bs used in the
equations are not identical):
BIN JBK = BIN JB1K = 2 · BIN JBK + 1
= 2 · BIN JB0K + 1 = 2 · 2 · BIN JBK + 1
= 2 · 2 · BIN J1K + 1 = 2 · 2 · 1 + 1 = 5
Since B = 101, the final denotation is given by BIN J101K = 5. This is precisely the
expected result.
40
5.2. Survey of QPL
Figure 5.1: Derivation tree for the binary numeral 101 generated by the abstract
grammar given in Equation 5.1.
5.2
Survey of QPL
The semantics of QPL is based on the idea of annotating a flow graph that represents a
quantum program with density matrices for the quantum mechanical parts and tuples of
probabilities covering the classical components. Additionally, a typing context is used to
keep track of all variables together with their types that are in use at a certain stage of a
program.
Since our work is based on QPL, it seems appropriate to summarise its central concepts.
The original definition of QPL [Sel04b] provides a more detailed description than given
here; an alternative review can be found in [Sch04]. We align our summary on both sources.
Note that it is nevertheless useful to have some familiarity with the paper introducing QPL
because we can obviously not repeat everything here.
5.2.1
Notational conventions
QPL operates on finite-dimensional quantum states represented by vectors over C. The
basis states for qbits are defined as |0i = (1, 0)t and |1i = (0, 1)t . Combination of multiple
qbits are as usual represented by tensor products of these states. Density matrices are used
as basis for any manipulations performed by the language. If a state is defined by some
n
vector u ∈ C2 , the corresponding density matrix is given by ̺ = uu† and may also be
denoted by {u}. Mixed states are represented by linear combinations of pure states, e.g.,
λ1 u1 u†1 + · · · + λn un u†n . Given four matrices A1 , A2 , A3 , A4 of identical dimension, they can
be concatenated horizontally and vertically by
A1 A2
(5.2)
A3 A4
which is used to specify composite density matrices. This notation is used to specify the
semantics of actions possible in QPL which will be introduced in the following sections.
5.2.2
Language elements
QPL programs are given in terms of quantum flow charts 4 where each edge is supplemented
with all the information necessary to unambiguously specify the meaning of a program.
Every edge is augmented with
❏ a typing context, i.e., a mapping from identifiers of variables to the types of these. It
is written as a list of identifiers followed by their type, e.g., a, b, c : bit, d : int. Typing
4 There is also a textual representation for programs, but this is only considered as an aside in the
definition of QPL.
41
Chapter 5. Formal denotational semantics
contexts encapsulating variables which are not related to the present considerations
are denoted by Γ.
❏ an annotation, i.e., a tuple of density matrices which specifies the state of the system.
The annotation of a classical bit is given by (A, B) where A + B = 1 and A represents
the probability that the value of the bit is 0, whereas B is the probability that the value is
1. The annotation for a quantum bit is of the form given by Eqn. 5.2.
All classical operations possible with QPL and their flow graph representations are shown
in Figure 5.2. Figure 5.3 depicts the quantum mechanical parts.
Figure 5.2: Summary of all classical operations of QPL, taken from [Sel04b]. Note
that the symbol “=” is used to separate typing context and the annotation, which can be confusing at times because it is not associated
with Γ alone.
5.2.3
Semantics
The semantics of a QPL program can be directly inferred from the flow graph representation.
The explicit transformation of a density matrix given in the annotation of the edges serves
as a unique representation of the meaning of a program. [Sel04b] proofs that this approach
is indeed well-defined and also works for recursion and loops, which can be included into the
language. Categorical structures that allow to accommodate superoperators and morphisms
42
5.2. Survey of QPL
Figure 5.3: Summary of all quantum mechanical operations of QPL, taken
from [Sel04b].
to manipulate these according to the possibilities of QPL are used as a formal basis for the
definition of the semantics. This categorical superstructure5 is not too interesting for our
purposes. It suffices to know that the valuation functions for the diverse language elements
are defined as shown in Figure 5.4 and that they indeed fulfil everything which is necessary
for a sound and well-defined interpretation. Note that the way how the semantics is specified
in Figures 5.2 and 5.3 is not equivalent to the method of Figure 5.4: While the first one
relies on explicit transformations of density matrices, the second one uses a more abstract
representation in form of superoperators and is almost completely identical to the basis of
our approach (the functions computed by both approaches of Ref. [Sel04b] are nevertheless
identical except for loops and recursion). Especially, the second variant is compositional,
which is a necessary condition to describe multipartite systems in such a way that the
description of one part is independent of other parts.
5.2.4
Limitation: Quantum communication
Before we lay out the denotational semantics of cQPL, it is advisable to sketch in which sense
the different approaches used in QPL do not work for programs dealing with communication.
5 Note that we are only referring to category theory here, not to the compositional semantics presented
by Selinger.
43
Chapter 5. Formal denotational semantics
Figure 5.4: Valuation functions which define the denotational semantics of QPL,
taken from [Sel04b].
First of all, let ̺AB denote the density matrix of the state shared by Alice and Bob.
The information available for each party can be inferred by calculating the partial trace:
̺A = trB (̺AB ) and ̺B = trA (̺AB ). The bipartite density matrix can never be recovered
from these partial density matrices because there are many bipartite density matrices which
give rise to the same partial density matrices.
One of the goals of denotational semantics is to assign sufficient information to every edge
of a quantum flow graph such that the complete semantics of a program can be reconstructed
by combining only the information given by the edges constituting the program. The denotation of a statement composed of several sub-statements must be completely determined
only by a function of the denotations of the sub-statements.
This is impossible in the annotation-based semantics of QPL because transformations
between explicit density matrices are considered. Since a combination of the partial density
matrices ̺A , ̺B which were manipulated by Alice and Bob does not restore the total bipartite state ̺AB , the QPL annotation would obviously not comply with the physical state
afterwards.
A possible solution is the annotation of the complete flow graph, i.e., of both paths
representing the control flow for Alice and Bob. In this case, the operations performed
by Alice and Bob would be written as tensor products of the type  ⊗ 1B and 1A ⊗ B̂
which act on the complete density matrix ̺AB . This way, we could assign semantics to the
program as a whole, but would loose the ability to construct the denotation of a phrase from
the denotations of its subphrases. This means that the semantics of the complete program
could not be constructed from the denotation of Alice’s and Bob’s program alone which is
in contrast to the key idea of denotational semantics.
Therefore, we need to seek a solution that does not characterise quantum operations by
showing transformations of explicit density matrices, but uses something that captures the
44
5.3. Denotational semantics of cQPL
notion of a transformation in a more abstract sense. Completely positive maps represented
by a set of Kraus operators fulfil this need as we will show in Section 5.3.1; this is basically
the same approach as used for the compositional semantics of QPL. Nevertheless, QPL does
not provide any means of parallel composition, communication and other details which are
necessary to describe quantum communication as we will do in the remaining parts of this
thesis.
5.3
Denotational semantics of cQPL
cQPL is an extended variant of QPL with the ability to express and formalise quantum communication, i.e., the ability to describe multiple parallel flow graphs that exchange quantum
and classical information at well-defined points, but do otherwise know nothing about each
other. To achieve this, we have to base the semantic description on three components in
contrast to the two components (typing context and tuples of density matrices) of QPL:
❏ A Kraus aggregation K as defined in Section 4.5.4 which is used to keep track of the
quantum operations performed on the qbits of the system.
❏ A typing context T used to specify which quantum and classical variables are allocated
at a given moment and which data type they have. This is also important to describe
communication because it allows to uniquely determine to which party a variable
belongs at a given stage of a program.
❏ A probabilistic environment mapping identifiers to values. Since the interaction between quantum and classical parts of the system introduces probability, the values of
classical variables are subject to such a distribution. In general, only the range of
possible values together with the fact that it is governed by a probability distribution
is known in the semantical description.
We refer to these three elements as the three-tuple (K, T, E).
A Kraus aggregation K specifies a quantum mechanical operation which has the same
effect for all density matrices ̺ (in the sense that the application of a Hadamard gate will
yield different effects according to the state it was applied to. Nevertheless, it is still a
Hadamard gate in every case, and this is the really important thing). Thus, the operation
is completely characterised without the need to specify any density matrix at all. This is
exactly what is required when spatially separated operations performed by several parties
on multipartite states are to be described, as we will see later in greater detail.
To realise the benefits of this approach, consider how the generation of a new qbit in
state |0i subsequently followed by the application of a Hadamard gate is described in QPL
(we do not show the complete flow graph, but only the relevant parts of the annotation):
create new qbit q
Γ = A −−−−−−−−−−−→ q : qbit, Γ =
A 0
0 0
45
apply H on q
−−−−−−−−→ q : qbit, Γ =
1
2
A A
A A
Summary
We have presented a quick summary of QPL and the associated denotational
semantics which is based on partial orders of density operators. Additionally, we
have shown why this approach is not suitable to describe quantum communication
respectively the interaction of spatially separated systems where the combined
density matrix is not available as whole in the framework of the annotation-based
semantics.
Chapter 5. Formal denotational semantics
Although only the newly created qbit is concerned, the state of the remaining system is
still implicitly present in A. This is more than needed: It suffices to consider the application
of two operations given by the following Kraus sets:
{Ci }#q ; {Hi }#q
(5.3)
where {Ci }#q stands for “create a new qbit with label q” and {Hi }#q for “apply a Hadamard
gate on q”. With these, we can describe the same operation without resorting to a density
matrix or any other part of the system unconcerned by the operation at all.
The typing context T is basically adopted from QPL. An extension to the framework
used by QPL is the probabilistic environment. For every allocated classical variable in the
current frame, it is used to specify a probability distribution that maps the variable name to
the range of possible values. This distribution is parametrised by density operators because
it depends on the initial conditions of the program fragment and on the path taken in the
flow graph (an example explaining this will follow in the next section). The probabilistic
environment could in principle be replaced by the tuples for classical states as used in QPL,
but this works only well for data types with a very low number of bits. Because of this
reason, QPL tries to hide these tuples most of the time, so we eliminate them completely
and replace them by the probabilistic environment.
The probabilistic environment also deals with quantum variables: For every such variable,
the position of the allocated qbits in the global quantum heap is given by the probabilistic
environment. This is necessary because quantum variables cannot be characterised by a
value as it is possible for classical variables because they do not have a state as such. The
state is replaced by the series of operations which have been performed on the variables;
since these operations need some location to act on, every quantum variable needs to have
a unique position on the quantum heap, i.e., where the qbits are stored.
Note that this approach is somewhat contrary to the spirit of functional programming
because it introduces stateful global variables, but a closer examination reveals that QPL
implicitly uses the same model and that compile-time checking (and thus the protection
against runtime errors) is not affected by this.
In addition to allocated variables, branches in programs are also present in the probabilistic environment. They are identified by a unique ID which is assigned to every branching
node.6 This is necessary because the branching conditions – being based on comparisons of
probability distributed quantities – are in general not represented by some fixed values, but
represented by a probability distribution as well.
5.3.1
Formal definitions
In this section, we will present some methods to characterise and describe the semantic
components of cQPL. Note that in the following, we use Dn to denote the set of all density
operators of dimension n, dropping the subscript if the exact dimension is not important or
can be deduced from the context.
5.3.1.1
Typing context
Let σ be a list of numbers ni ∈ N+ , i = 1, . . . , k as given by σ = nτ1 , nτ2 , . . . , nτk . σ is also
called the signature of a data type. An associated Hilbert space Hσ is given by
6 To
Hσ ≡ H1 ⊗ · · · ⊗ Hk
(5.4)
be precise: Which is assigned whenever the branching node is transversed because we need to account,
e.g., for branches in loops where the same branch might be traversed multiple times.
46
5.3. Denotational semantics of cQPL
where Hi is either B(H) for quantum or C(X) for classical data (cf. Section 4.3.3) where
both are distinguished by the index τ : τ = q for quantum variables and τ = c for classical
q
variables. The dimension of the i-th space is given by 2ni for quantum mechanical and nci
for classical variables. Since we restrict ourselfs to finite-dimensional Hilbert spaces, this
means that we can use Hi = Cni for quantum mechanical and X = [0, 1, . . . , ni − 1] for
classical data. To distinguish between both cases, we define the function q : nτi → [0, 1]
given by
(
1 if τ = q
τ
q(ni ) =
(5.5)
0 otherwise
Note that although our formalism allows to define data types which consist of both quantum
mechanical and classical components, we do not exploit this possibility because we could
not find any reasonable application for this in our work. To keep the formalism as general
as possible, we will nevertheless still retain the possibility as long as no noteworthy effort is
necessary to do so.
We can define a function tq : σ → N to compute the total number of qbits necessary for
a given signature (card(σ) denotes the cardinality of σ):
card(σ)
tq (σ) =
X
i=1
q(nki ) · nki
(5.6)
The analogous function tc for the classical components is given by
card(σ)
tc (σ) =
X
i=1
(1 − q(nki )) · nki
(5.7)
Finally, two functions q(σ) and c(σ) to check if a given data type is purely classical or purely
quantum are necessary:7
q(σ) =
(
1 tc (σ) = 0 ∧ tq (σ) > 0
0 otherwise
(5.8)
c(σ) =
(
1 tq (σ) = 0 ∧ tc (σ) ≥ 0
0 otherwise
(5.9)
For simplicity, we label data types required for practical use with special mnemonics; some
examples can be found in Table 5.1. Note that the type void can, for example, be used to
formally describe statements which return no value and thus have no type.
Since two finite-dimensional Hilbert spaces H1 and H2 are isomorphic if the sum over
the dimensions of their subsystems is equal (this is, e.g., proved in [Wei00, Theorem 2.62]),
i.e.,
O
O
X
X
H1k = H1 ∼
H2l ⇐⇒
dim(H1k ) =
dim(H2l ),
(5.10)
= H2 =
k
l
k
l
the description of types is not unique. For example, the types given by (2q , 2q , 2q , 2q ) and
qshort = 8q are identical and provide only different aspects of the same thing. This
7 Note
that we define a data type consisting of 0 quantum and 0 classical bits, i.e., void, as classical.
47
Chapter 5. Formal denotational semantics
Mnemonic
bit
qbit
short
qshort
int
qint
void
Signature
2c
2q
8c
8q
16c
16q
0(c)
Table 5.1: Signatures and the mnemonics commonly used in programming languages for data types supported by cQPL.
equivalence can also be extended to mixed data types:
X
X
nτ1 , . . . , nτK ∼
q(nτk ) · nτk =
q(mτl ) · mτl
= mτ1 , . . . , mτL ⇐⇒
k
∧
X
k
l
(1 −
q(nτk ))
·
nτk
=
X
l
(1 − q(mτl )) · mτl
(5.11)
This creates an equivalence class for data types which will be useful in Section 5.3.8. The
class of all data types equivalent to σ is given by
tc (σ)
q (σ)
tM
M
inϕ(i+tq ) (Ω#i)c
(5.12)
inϕ(i) (Ξ#i)q ⊕
T (σ) ≡
i=1
i=1
such that Ξ ∈ S(tq (σ)), Ω ∈ S(tc (σ)), ϕ ∈ Sym([1, . . . , tq (σ) + tc (σ)])
where M #i denotes the ith element of the ordered set M and S(k) is the decomposition of
the scalar value k into all possible sums given by
n
S(k) ≡ Z is a set with members ∈ N
card(Z)
X
i=1
o
Z#i = k .
(5.13)
If σ2 ∈ T (σ1 ), we write σ1 ∼
= σ2 .
To illustrate the effect of Eqn. 5.12, consider a data type which consists of 3 quantum
and 3 classical bits. Structurally, it does not make any difference how these components
are ordered, e.g., (1q , 1c , 2q , 2c ) is identical with (3q , 3c ) in this sense.8 Eqn. 5.12 is a
generalisation of this idea: The scalar 3 can be decomposed as 1 + 1 + 1, 2 + 1 and 3
as given by S(3), so there is no difference between any of these groupings. Additionally,
it is not interesting how the components are ordered, e.g., (2 + 1) is equivalent to (1 + 2).
Finally, the quantum and classical components can be arbitrarily interchanged, so we have
to consider this as well. The effect of Eqn. 5.12 is to construct all equivalent representations
of a data types following these considerations.
Note that QPL uses a Cartesian product of complex vector spaces given by Cn1 ×n1 ×
· · · × Cnk ×nk for both classical and quantum mechanical signatures (the set of complex d × d
8 Note that there is a difference between these orderings from the compiler’s point of view because the
different components are located at different locations in memory if different orderings are used. The
semantics is nevertheless unconcerned by this.
48
5.3. Denotational semantics of cQPL
matrices is used to represent the complex Hilbert space of dimension d). This does not
reflect the relationship between corresponding quantum and classical objects directly. For
example, the data type for bits is given by bit = (1, 1), whereas for qbits, the definition
is qbit = 2). This leads to appropriate spaces for these objects, but does not present the
relation between them directly.
We thus used a different approach that makes correspondences more clear which is important when, for example, quantum variables are measured and the result is stored in
a classical variable. Besides, it fits better into the more abstract description of quantum
mechanics as introduced in Section 4.3.
Let Σ be the set of finite strings over the alphabet α. With this and the notion of types,
we can define the typing context used in the semantic description of cQPL.
Definition 5.3.1 (Typing context). A typing context τ is a three-tuple τ = (ι, θ, χ) where
ι is a set of identifiers in Σ, θ is a set of types and χ : ι → θ is a surjective mapping which
assigns a type to every identifier. Identifiers starting with # must not be used by programs.
Because cQPL is strongly and statically typed (i.e., the type of an expression is completely determined by the types of its components and the type of an elementary component
cannot be changed after it has been declared, cf. Appendix B), information contained in
the typing context cannot be modified any more once it has been introduced. Note that this
does not hinder the possibility of overshading entries. This happens when, e.g., a variable
declared in an inner block has the same name as a variable declared in an outer block.
Although both have identical names, their types do not need to match because they are
otherwise completely unconnected.
Typing contexts are modified when new variables are declared (and thus added to the
context) or when variables are removed from the scope (and thus have to be removed from
the typing context). Since it is obvious how this influences a given context τ , we only note
that it is easy to define appropriate morphisms τ → τ ′ which perform the desired job.
Formally, we use the notation
τ → τ ′ = τ ⊕ (ξ → qbit)
(5.14)
to introduce some new identifier ξ with type qbit into the context τ . Equivalently, the
notation τ ⊖ ξ is used to remove ξ which is needed to describe sending quantum variables.
5.3.1.2
Probabilistic environment
The probabilistic environment can be defined formally as follows:
Definition 5.3.2 (Probabilistic environment). Let π be a probability
distribution on a
P
finite set X with probabilities pi for every element of X such that i pi = 1. Let ι be a set of
identifiers, P be a set of probability distributions and M be a surjective map M : ι → P ∪ ⊥.
Then E = (ι, P, M ) is a probabilistic environment.
As usual in denotational semantics, the symbol ⊥ is used to denote an undefined identifier, i.e., a variable which is not present in the environment. To specify components of a
probabilistic environment, we use the notation
π
x
range(x)
x −→
(5.15)
to denote an element of the probabilistic environment where x is the identifier, πx the
associated probability distribution and range(x) the set of possible values which obviously
49
Chapter 5. Formal denotational semantics
depends on the data type of x. Adding a new binding to a given environment E is once more
done with the operator ⊕ which is formally a morphism E = (ι, P, M ) → (ι′ , P ′ , M ′ ) = E ′ :
π
x
range(x)
E ′ = E ⊕ x −→
(5.16)
Note that multiple inclusion of variables overrides the previous definition. Thus, the
meaning of
π′
π
x
x
range(x)) ⊕ x −→
range(x)
E ′ = (E ⊕ x −→
(5.17)
is to create a probabilistic environment which contains πx′ as probability distribution for
the variable x. The previous distribution πx can then not be recovered any more in E ′ .
In direct analogy to Kraus aggregations, probabilistic environments can be combined
with +; the summands are prefixed by some constraint that states which one has to be
chosen with which probability:
p1 ({c1 }) · E1 + p2 ({c2 }) · E2 + · · ·
(5.18)
P
where {ci } are the conditions which determine the values of p and i pi = 1. This construction is necessary for the description of, e.g., if-conditions when it is not a priori determined
which path will be selected. If both paths of an if-condition perform a modification on the
same variable that already existed before the branch, then the variable will have different
values after the merge point. The entries of the probabilistic environment which record this
assignment are then prefixed by the branching probability. This is also one of the reasons
why the branching probability needs to be kept even after the branched paths are merged
again.
Definition 5.3.3 (Distributivity of ⊕ over +). The operation ⊕ is defined to be distribuπx
πx
πx
range(x).
range(x) + E2 ⊕ x −→
range(x) = E1 ⊕ x −→
tive over +, i.e., (E1 + E2 ) ⊕ x −→
This ensures that adding a new binding to a sum of environments results in adding the
binding to all contributing environments automatically.
Remark 5.3.1. This formalism is not equivalent with the functionality introduced by stores
(cf. e.g., [Mos90, Rey98]). It does still not make use of stateful variables per se, but rather
updates the binding of a variable, i.e., the value it is associated with.
Observe that the probabilistic environment is only necessary for classical, but not for
quantum variables: The state (or, rather: the history of all operations performed until the
present moment) of the quantum mechanical constituents of the computation can be reconstructed with the aid of the Kraus aggregation. Nevertheless, the probabilistic environment
is necessary to keep track of quantum variables in a different way which will be introduced
in a moment.
Note that the view on quantum variables differs slightly from that on classical ones: It
is not only necessary to keep track of the structure of a variable (as is done by the typing
context), but also of the position within the quantum heap – this is necessitated by the
underlying model of computation as introduced in Section 2.2.3.9 The environment can be
used to provide this kind of information by supplying a map
9 The
# : ι ∋ v → (i1 , i2 )
(5.19)
value of a quantum variable can obviously not be directly stored in an environment because the
state might be in a superposition. The operations performed on the quantum bit are recorded in the Kraus
aggregation and unambiguously specify the state.
50
5.3. Denotational semantics of cQPL
where ik are integer numbers with 0 ≤ ik < Q and Q is the size of the quantum heap. The
tuple i1 , i2 denotes the interval [i1 , i2 ] which contains i2 − i1 + 1 quantum bits. Obviously,
the number of qbits allocated in the quantum heap must agree with the number of qbits
necessary for the type of the variable as given by the typing context.
In theory, it is possible to assume that the quantum heap can always be partitioned into
consecutive intervals; we do not need to take care of issues like fragmentation which does
obviously appear in implementations and simulations. We assume that the hardware of the
quantum memory take care of this issue by acting like a memory management unit.10 Note
that if the user is allowed to directly address the components of the quantum heap, it is
possible to cause run-time errors as in QCL. Therefore, we do not allow this.
Consider a subset M = [0, n] of N. The set of all interval partitions is given by11
I(M ) ≡ {m ⊆ P(M )|∀n ∈ [1, . . . , card(m) − 1] :
((m#n)#0 − (m#(n − 1))#(card(m#(n − 1)) − 1)) = 1}
(5.20)
where we suppose that the contents of all sets m#i is sorted in ascending order. This
can be used to formally define how the probabilistic environment can be adapted to the
requirements for quantum variables:
Definition 5.3.4 (Quantum part of the probabilistic environment). Let (ι, P, M ) be
a probabilistic environment. It can be extended to fulfil the requirements for the description
of quantum variables by the following construction:
❏ P is extended to P ⊕ (I([0, Q − 1]) ∪ Σ∗ ) where Q is the total number of quantum
bits present in a system. The set of intervals is used to represent quantum variables
which reside on the local quantum heap, i.e., which were allocated in the module the
probabilistic environment belongs to. Σ∗ is used to denote the originating module for
variables which were received from some other party.12
❏ Let Q = { v ∈ ι q(χ(v)) = 1 } be the set of all identifiers for variables
T with quantum
data type. Then, M ′ is an injective morphism Q → I for which range(M ′ ) = ∅
(this ensures that quantum variables do not overlap on the quantum heap) must hold.
Then M is replaced by M ⊕ M ′ in the previous definition.
Note that this definition reflects a fundamental difference between classical and quantum
variables: While a classical variable is nothing else than a mapping between an identifier
and a value that can be governed by a probability distribution, such a mapping is in general impossible for quantum variables because they do not have a value per se, but only a
certain quantum state. To describe this quantum state precisely (disregarding the principal
impossibility of implementing a measurement that delivers this information by inspecting
10 This component of a processor creates a view of the available memory such that every application –
roughly – thinks that it would have an own linear address space which is as big as the the one available for
the whole system.
11 An example might illustrate this definition: Consider the set {[0, 1], [2], [3, 4]}. This is a proper partition
since no elements overlap and the boundaries are adjacent. These conditions can be ensured by considering
the last element of the ith set given by (m#i)#(card(m#n) − 1) and the first element of the (i + 1)th set
given by (m#(i + 1))#0. If the difference between these is +1, then both the adjacency and no overlap
conditions are fulfilled. If this holds for all subsets, we have a proper partition.
12 We have to make sure that every quantum bit in the system belongs to exactly one place in a quantum
heap. This is simple for single-party programs, but gets more complicated when communicating programs
are considered because the case of sending the same quantum bit back and forth between participants must
be taken into account.
51
Chapter 5. Formal denotational semantics
a single copy of a quantum system), one would need an infinite amount of classical information because even a simple system such as a qbit takes values in a continuous space as
a consequence of quantum superpositions. This makes the classical approach of mapping
the identifier to a probability distribution of values impossible. Nevertheless, the quantum
variable is completely characterised if its location on the quantum heap together with the
operations performed on its initial state are known.
Remark 5.3.2. Although we retain the name probabilistic environment also for the version
of the environment extended to quantum variables, there are no probabilities involved in the
connection between variable names and the allocated positions on the quantum heap. The
convention just simplifies the notation.
Remark 5.3.3. Support for mixed quantum/classical types would require a little more effort
compared to the case of full separation because with the introduction of such types, the direct
decomposability of the probabilistic environment would not be feasible any more. Nevertheless, no fundamental difficulties would be associated with this.
Adding a new quantum variable with name ν which occupies the quantum heap positions
given by (q1 , . . . , qn ) is written as
E ⊕q ν : (q1 , . . . , qn )
(5.21)
Removing a quantum variable is denoted by ⊖q ; this is required when qbits are transmitted and thus cannot be accessed any more (we drop the index if there is no danger of
confusion). Note that there is no need for a corresponding operation for classical variables
because overlays provide the required functionality, as will be shown later.
5.3.1.3
Kraus aggregations
Summary
We can directly adopt the definition of Kraus operators as given in Section 4.5.4. Nothing
needs to be modified for our purposes (note that the composition of two Kraus aggregations
was denoted by ◦ instead of ⊕ as used for the other elements of the three-tuple (K, T, E) to
avoid confusion with the symbol + used to combine sub-aggregations).
It is possible to show that this semantic framework can be used to formalise standard
QPL, but we omit the details here.
We have introduced the structures which are necessary to specify the denotational semantics of cQPL; they fulfil the required properties as described in Chapter 4. The semantic framework consists of three components: A Kraus aggregation
which is used to store all quantum mechanical operations performed by commands
of cQPL, a probabilistic environment which maps identifiers to values (possibly
governed by a probability distribution) and provides mechanisms to ensure that
quantum variables do not interfere with each other, and a typing context whose
information is the basis for compile-time correctness checks of programs. These
are grouped in a (K, T, E) tuple which will be omnipresent in the following. Additionally, we have derived some criteria for the identity of data types.
5.3.2
Some examples
Before we commence to extend the formal definitions for multipartite systems, we want to
demonstrate the introduced concepts with some examples which should aid the reader to
see the rationale behind their definition.
52
5.3. Denotational semantics of cQPL
5.3.2.1
Semantics of sequential programs
Consider the following program fragment which applies a Hadamard matrix to the quantum
variable p or q depending on the result of a branch based on a comparison of two classical
variables x and y:
if (x > y) {
q *= H;
}
else {
p *= H;
};
The flow graph representation of the fragment is given in Figure 5.5. To shorten the annotation of the edges and to avoid repeating the same information over and over, we introduce
the injection functions ini . ini are 0-based injections into the ith element of an n-tuple. If we
want to add the contribution {U } to the element K of the tuple (K, T, E) = ξ, we can write
this as ξ ⊕ in0 ({U }). The initial (K, T, E) tuple of the example flow graphs is abbreviated
by ξ; modifications derived from this are denoted by ξ ′ , ξ ′′ , . . ..
{{Γ}} ; {τ ; x, y : int; p, q : qbit} ;
o
n
πy
πx
range(x); y −−→ range(y) ≡ ξ
E ⊕ x −−→
x>y
Unique node ID #42
ξ ′′ ≡ ξ ⊕ in2 (#42 → false)
⊕ in1 (#42 : bit)
ξ ′ ≡ ξ ⊕ in2 (#42 → true)
⊕ in1 (#42 : bit)
p∗ =H
q∗ =H
´
`
ξ ′ ⊕ in0 {H}#p
´
`
ξ ′′ ⊕ in0 {H}#q
Implicit merge for branch #42
n
“
”
“
”o
p (val(#42) = true) · {Γ} ; {H}#p + p (val(#42) = false) · {Γ} ; {H}#q
;
{τ
o
n ; x, y : πint; #42 : bit; p, qπy: qbit} ;
π#42
x
range(x); y −−→ range(y); #42 −
−−−→ range(#42)
E ⊕ x −−→
Figure 5.5: Flow graph of a simple branching operation to demonstrate the elements of the semantic framework: A 3-tuple (K, T, E) is used to annotate every edge of the graph; K is a list (or aggregation) of Kraus
operators, T is the typing context and E the probabilistic environment.
Note that the annotation of the graph uses several abbreviations as
defined in the text.
The initial configuration of the 3-tuple (K, T, E) is given by Γ (initial list of Kraus
operators), E (probabilistic environment) and τ (typing context). We don’t care about
their contents in detail, they can represent the semantics of any valid program fragment that
might be placed before our example. Things which are of interest for our code fragment are:
53
Chapter 5. Formal denotational semantics
❏ The two classical variables x and y, both of type int.
❏ The probability distributions πx , πy which map the variables x and y to a value contained in [0, 2bits per int − 1].
The probability distribution for classical variables arises because we work with Kraus operators describing quantum operations instead of density matrix transformations. Consider
the measurement of a single qbit whose result is stored in a classical bit variable: We know
that the range of the measurement outcome is {0, 1}, but we don’t know with which probability the “0” and the “1” will appear because we do not have an explicit density matrix to
describe the qbit. This piece of information can only be gained when the final semantics of
the program (in form of a total set of Kraus operators) is “applied” to a well-defined initial
configuration; only then quantitative statements about the distribution are feasible. All we
know is that the measurement result x will be governed by a probability distribution πx
with a certain well-defined range, so we preserve that information.
Since the values of x and y are given by a probability distribution, the result of the
comparison x > y (with outcome range {true, false}) can only be specified by another
probability distribution which can be deduced from πx and πy . Since we need to reference
the outcome of the comparison at a later point in the flow graph (when the two edges of the
branch are merged), a unique identifier for the node is created (#42 in this case) and the
probabilistic environment is extended accordingly.
Depending on the outcome of the comparison, a Hadamard gate is applied on either p or
q. This does not change the probabilistic environment or the typing context, but is recorded
by placing an appropriate Kraus operator in the Kraus aggregation (#p and #q denote the
position of the quantum bits in the quantum heap).
After every if-then-else construction, an implicit merge operation which unites the two
branches takes place. The probability distribution of the branching condition is preserved
in the probabilistic environment under the label assigned to the branch statement; the
Kraus aggregation is transformed into a sum that formally resembles a mixed state: With
probability val(#42) == true (which is the probability that x > y evaluated to true),
the operation {Γ̂}; {H}#p was performed, while with probability val(#42) == false, the
operation was {Γ̂}; {H}#q .
5.3.2.2
Communication with EPR pairs
Consider the following (pseudo-)code which describes how Alice creates an EPR pair and
transmits half of it to Bob:
module Alice {
new qbit p := 0;
new qbit q := 0;
createEPR(p,q);
send q to Bob;
new bit b := measure(p);
if (b) { ... } else { ... };
};
module Bob {
receive m from Alice;
new bit b := measure(m);
if (b) { ... } else { ... };
};
54
5.3. Denotational semantics of cQPL
The corresponding flow diagram is given in Figure 5.6.
new qbits p, q := 0
createEPR(p, q)
n
{Γ} ; {C}p ; {C}q {EP R}#p,#q
{τ ; p, q : qbit} ; {E}
send q
nn
Γ′
o
;
receive m
oo
nn
;
{τ ; p : qbit} ; {E}
measure p
ID: #23
o
; {R}#m ;
n
o n
o
′
τ ; m : qbit ; E ′
Γ′′
o
ID: #42
measure m
Figure 5.6: Flow diagram which describes the creation of an EPR pair by Alice;
she keeps the first half, while the second half is sent to Bob. Afterwards, both of them measure their qbit. {Γ′ } is a shorthand for
{Γ}; {C}p ; {C}q ; {EPR}#p,#q , {Γ′′ } is the initial Kraus aggregation
of Bob, τ and E respectively τ ′ and E ′ are the initial typing contexts
and probabilistic environments of Alice and Bob.
Note that the labelling formalism is reduced to the basic necessities in this example
in order to emphasise the central elements. Because Kraus operators are used to describe
the quantum mechanical operations, it is possible to perform spatially disjoint actions by
parties who do not know the total state. When the edges are merged together, the operations
performed by Alice and Bob can (as was mentioned before) be factorised as  ⊗ 1B , 1A ⊗ B̂
for the combined semantics of both branches; the total density matrix is not involved in this,
contrary to annotation-based QPL. The framework to generate the semantics of the total
system from the semantics of the components will be introduced in the following section.
5.3.3
Extension to multipartite systems
Since we want to consider the formal semantics of programs which deal with communication,
we need to extend the previous definitions to the multi-party case. For this, observe that
the number of participants can naturally be assumed to be finite which makes it possible to
label each party with a unique identifier of finite length. For formal simplicity, we assume
that the set of labellings for participants Lc is disjoint with the standard labels used for
variables, i.e., Lc ∩ Σ∗ = ∅. Let Lc (i) denote the unique label given by the ith entry of Lc .
Assume that we have n communicating parties which are labelled with l1 , . . . , ln ∈ Lc .
According to the principle of compositionality (we will explain this for the context of communication in more detail in Section 5.3.6.4 on Page 70), there is a three-tuple (K, T, E)
for every participant, i.e., (K1 , T1 , E1 ), . . . , (Kn , Tn , En ). The combined three-tuple for the
complete system is then given by
(⊗ni=1 Ki , ⊗ni=1 Ti , ⊗ni=1 Ei )
55
(5.22)
Chapter 5. Formal denotational semantics
We will consider how the tensor product needs to be defined for each component to
provide a sound basis for our further needs.13
5.3.3.1
Kraus Aggregation
Consider, for simplicity, two Kraus aggregations
Λ1 = {A1k }, {A2k }, . . . , {Ank }
Λ2 =
{Bk1 }, {Bk2 }, . . . , {Bkm }
(5.23)
(5.24)
(in case of n 6= m, the shorter list can be padded with zero elements so that m = n can
be assumed, but note that it is only for notational convenience). If both lists operate on a
disjoint set of qbits, i.e., no send/receive (pseudo-)operations are contained in the lists, then
∀i, j : [{Aik } ⊗ 1B , 1A ⊗ {Bkj }] = 0 holds. The Kraus aggregation for the composite system
can be written as:
Λ1 ⊗ Λ2 = {A1k } ⊗ {Bk1 }, . . . , {Ank } ⊗ {Bkn }
(5.25)
Note that members of type {Ak } ⊗ {Bk } can be written as {A} ⊗ 1B + 1A ⊗ {B}. If this
is done for all list elements, we see that all combinations of A ⊗ 1 with 1 ⊗ B commute;
this induces many equivalent orderings of the lists – in addition to the normal commutative
equivalences as given by Eqn. 4.57 for the sublists – which needs to be considered when
multi-party aggregations are tested for semantic equality in Section 5.3.5.
Formally, the equivalence class of compatible Kraus aggregations for two independent
parallel systems is given by
)
( n
n
M
M
i
i
inϕ(n+i) 1A ⊗ {Bk }
inϕ(i) {Ak } ⊗ 1B ⊕
(5.26)
i=1
i=1
with ϕ ∈ Sym(2n) : ϕ(i) < ϕ(i + 1)∀i ∈ [1, . . . , n] ∧ ∀i ∈ [n + 1, . . . , 2n]
where the additional constraints on the permutation make sure that the order of {Aik }
and {Bki } is preserved. We can generalise this approach to n Kraus aggregations given
in the contracted normal form (padding is applied as usual to compensate for different
cardinalities):
PN
Definition 5.3.5 (Tensor product for Kraus aggregations). Let Λ1 = k1 =1 p1k1 Λ1k1 ,
PN
. . . , Λn = kn =1 pnkn Λnkn be Kraus aggregations in the contracted normal form. The tensor
product of these is given by
n
O
i=1
Λi ≡
N
X
k1 =1
···
N
X
kn =1
p1k1 · · · pnkn · Λ1k1 ⊗ · · · ⊗ Λnkn
(5.27)
Note that the naming of qbits changes when multipartite systems are merged. For Kraus
sets which symbolically refer to qbits, the labels must be updated accordingly (the exact renaming scheme is given in Definition 5.3.6). Two aggregations are equivalent if they are
member of the same equivalence class as given by Eqn. 5.26 or member of the equivalence
class given by the same formula, but induced by a compatible ordering of one or more of Λi
as defined by Eqn. 5.25.
Note that we will show in Section 5.3.6.4 how send and receive operations can be integrated into this formalism.
13 Note that although formally, different tensor products are used for each element of the (K, T, E) tuple,
we use the same symbol for all of them to simplify notation.
56
5.3. Denotational semantics of cQPL
5.3.3.2
Typing Context
Definition 5.3.6 (Tensor product for typing contexts). The tensor product of n typing
contexts (ι1 , θ1 , χ1 ), . . . , (ιn , θn , χn ) is given by
!
n
n
n
n
[
[
[
O
(5.28)
Lc (i)χi
θi ,
Lc (i)ιi ,
(ιi , θi , χi ) =
i=1
i=1
i=1
i=1
where Lc (i)M denotes a set which contains all elements of M prefixed by the unique
identifier Lc (i); Lc (i)χ denotes the morphism χ adapted to the new naming scheme.
5.3.3.3
Probabilistic Environment
The probabilistic environment can be adapted to multipartite systems analogous to the
typing context, i.e., by prefixing the variable names with appropriate labels and adapting
the morphism used to connect the set of identifiers with the set of data types. Obviously,
the prefix for each subsystem must be the same as used for the typing context.
The quantum part needs to be modified as well: The local quantum heaps must be united
to a single one; necessarily, the positions of all variables on the new heap still need to be
disjoint. This is simple to achieve: If E0 uses the range [0, n1 ] and E1 the range [0, n2 ], the
new range is given by [0, n1 + n2 + 1] and the morphism M ′ needs to be adapted such that
the mapping remains unchanged for variables originating from E0 and the constant offset
n1 + 1 is added to its codomain for variables originating from E1 .14 Likewise, the function
# which associates variable names with quantum heap positions needs to be updated such
that the modified variables names are mapped to the modified positions.
We will not consider the renamings any more in the following parts, but just take them
as given; this simplifies the notation considerably.
5.3.3.4
Quantum channels
Quantum channels are used to exchange information between processes.15 Although not
only quantum data, but also classical variables can be sent, we restrict our considerations to
the first case because the second one is not too interesting from a physical point of view and
would only obstruct the view on the central elements. In particular, classical communication
can be seen as a special case of quantum communication (cf., e.g., [Key02, Section 6.2.2]),
hence the generality of the approach does not suffer from this restriction). Besides, the topic
of classical communication has been investigated in classical programming language theory
for a long time, so we can refer the reader to the wealth of existing literature about this
topic, e.g., Ref. [Rey98].
We have already introduced quantum channels informally in Section 2.2.3.1; here, we
consider the concept formally.
Definition 5.3.7 (Quantum channel). A quantum channel is a five-tuple (O, D, S, R, F )
where O is the origin and D the destination for a quantum variable (these can, e.g., be
14 Note that both local quantum heaps could have already used the full number of available quantum
bits (Q); we do not consider this problem any further because it is alway possible to limit the number of
quantum bits for n communicating systems to n·Q because both n and Q are finite. We are not too concerned
about this problem because we always assume that there are enough qbits available. The reason behind
the restriction to a finite, but fixed number of qbits is to ensure the boundedness of all Kraus operation as
explained in Chapter 4.
15 A user in a real-world implementation is nothing else than a process in the simulation. We thus use
both terms interchangeably.
57
Chapter 5. Formal denotational semantics
represented by processes), F is a FIFO16 containing objects of type (̺, σ), S is a morphism
to place two-tuples (̺, σ) in F and R is a morphism to retrieve two-tuples (̺, σ) from F .
As usual, ̺ represents the density matrix of a quantum variable and σ the associated type.
Thus, the quantum channel can be used to make sure that not only typing is guaranteed
to be preserved along communication (for this, confer further Section 5.4), but also that
quantum data does not appear multiple times in a composite system at the same time
which is necessary to avoid unphysical situations in the simulation.
Remark 5.3.4. Note that quantum channels are only necessary when the parallel composition of two or more processes is considered. For single modules, the functions used to deposit
respectively request information from the channel together with an abstract representation of
the channel (e.g., an identifier) are sufficient.
Remark 5.3.5. Also note that most descriptions of quantum communication protocols do
not consider typing of the exchanged data explicitely, it is nevertheless implicitly implied by
the physical realisation of the protocol, e.g., in the measurement process, by the hardware
used to realise the communication channel or in the way the quantum part is implemented
in general.
5.3.4
Existence of fixed points
Because fixed points are important for the denotational description, we need to prove the
following theorem which states a condition for the existence of such. In the following, the
condition can shown to be fulfilled for every element of the semantics.
Theorem 5.3.1. Let T be a linear operator D → D acting on a complete partial order D
with bottom ⊥D . If T is bounded, then a fixed point of T exists.
Proof. Since T is bounded, we can see from Theorem 4.2.1 that it is continuous as well.
Topological continuity implies Scott continuity as was shown in Theorem 4.4.2. The existence of a fixed point is now given by Theorem 4.4.1, as required.
Remark 5.3.6. Note that the same proof could have been deduced at a slightly more abstract
level using the notion of a pointed dcpo, i.e., a dcpo with a least element. For structures
fulfilling this condition, Theorem 2.1.19 in [AJ94] ensures that the desired least fixed points
exist. Ref. [Sel04b] uses a somewhat similar reasoning in the description of recursive procedures where the existence of least fixed point solutions for these is explained by the fact that
for Scott-continuous endofunctions on pointed complete partial orders, these always exist.
Remark 5.3.7. For those who want to be extra sure, the classical fixed point theorem by
Schauder which states that any continuous map with a countably compact image on a compact, convex subset of a Banach space has a fixed point could also be used to derive the
required property of cp-maps.
We will need fixed points to solve recursive equations which occur in denotations specified by recursive equations. These are required for loops and the combined semantics of
communicating systems.
16 First in, first out queue. Informally, this is a queue where objects can be put in on one side and taken
out on the other side. The object which is put in first comes out first, the second one comes out second etc.
58
5.3. Denotational semantics of cQPL
5.3.5
Types of interpretational equivalence
The term “equivalence” is not unambiguous without further specification. Under which
conditions can two programs or, respectively, the denotations of two programs be considered
as equivalent? QPL has to distinguish between two different types of equivalences as noted
in [Sel04b, Section 6.6]; likewise, we can define several types of equivalence:17
❏ Two programs are textually equivalent if there is a bijective mapping between the set
of all variables the programs use and for every constituent of program A, there is a
constituent of program B such that Ai = Bi , i.e., the programs are identical already
at the level of the syntax. The ordering of these constituents must be identical.
For example, the two program fragments new int a; a:=1; and new int b; b :=
1; are equivalent because the sequence of commands is identical if the replacement
a ↔ b is applied to the variables.
❏ Two programs are denotationally equivalent if their denotations are identical.
The second definition only shifts the problem because it leaves the question of how
to identify equivalent denotations. This is problematic for cQPL because we do consider
multiple representations of superoperators which have identical effects; some care needs to be
taken to exactly specify the meaning of “identical ” in this setting. Denotational equivalence
can be refined to the following cases for cQPL:
❏ Direct denotational equivalence: We can distinguish three different scenarios which
exhibit direct denotational equivalence for (K1 , T1 , E1 ) and (K2 , T2 , E2 ) given as denotations of A1 , A2 :
1. card(K1 ) = card(K2 ), ∀i ∈ [0, . . . , card(K)[: K1 i = K2 i and E1 = E2 , T1 = T2 .
This means that both programs have have the same number of Kraus sets as
denotation, the probabilistic environment and the typing context contain the
same information and the denotations of the statements are pairwise identical.
2. E1 = E2 , T1 = T2 , card(K1 ) = card(K2 ), ∃ϕ ∈ Sym(n) : ∀i ∈ [0, . . . , card(K)[:
K1 i = K2 ϕ(i) such that the sum of commutator products calculated according to
the method given in Section 4.5.4.2 vanish identically, i.e., only permutations with
vanishing commutator have been used. This equality holds if only commuting
statements have been exchanged to match the lists, the total denotation is thus
identical.
3. ∀̺ ∈ D : (K1 , T1 , E1 )(̺) = (K2 , T2 , E2 )(̺) where (K, T, E)(̺) means the application of the Kraus set in K on ̺ where the information contained in E is utilised
to construct the proper superoperators because the exact representation of K in
general depends on information given in T and E. Note that the initial state
resolves any symbolic parametrisations which may be present in K.
The first condition obviously implies the second and third condition; the second implies
the third, but the other direction is not true in general, so equivalences of decreasing
strength are defined by this enumeration.
❏ Heap-permutative respectively variable-permutative denotational equivalence is given
if there exists a permutation of the quantum heap positions (respectively the variable
names) such that direct denotational equivalence holds. These permutation schemes
can be used to align different probabilistic environments to each other.
17 Note
that our definitions of equivalence do not coincide with the types of equivalence given by Selinger.
59
Chapter 5. Formal denotational semantics
Summary
Note that textual equivalence implies denotational equivalence, but the converse statement is not valid in general.
A last form of equivalence that needs to be considered concerns the parallel execution of programs. If {Ai } represents a set of n communicating modules, the order in
which the subsystems are given does not make any difference, i.e., JA1 ||A2 || · · · ||An K ∼
=
JAϕ(1) ||Aϕ(2) || · · · ||Aϕ(n) K for any ϕ ∈ Sym(n) and an according update of the references to
the other subsystems Ak , k 6= m in Am for all m. Likewise, relabelling of communicating
modules does not change the meaning of parallel execution if the reference names in all participating modules are updated correspondingly. This type of equivalence can be referred
to as communicative equivalence.
The problem of how to detect equivalence between different representations will emerge
several times in the following remarks and is not easy to solve constructively.
We have augmented the definitions of the semantic basis (K, T, E) with the elements required to represent communicating systems, i.e., cQPL programs which
are generally independent of each other, but can exchange quantum mechanical
and classical data in a well-defined way. Additionally, we have shown that fixed
points exist in this framework; they are necessary to assign semantics to numerous
components of the language as explained in Chapter 4. Criteria for the equivalence of cQPL programs were specified as well; this allows to check if programs
which are specified by different sequences of commands have the same effect.
5.3.6
Semantics of the language components
Chapter 2 gave an informal18 introduction the the language elements of cQPL. In this chapter, we will use the mathematical and semantical formalism introduced in the preceding
sections to give a rigorous mathematical meaning to these statements. By the compositionality of denotational semantics, this means that all cQPL programs (which are, necessarily,
composed of cQPL statements) have a defined semantics. There are two possible representations for cQPL programs in form of textual descriptions and graphical flow charts; we
resort to the particular representation that is more convenient for the desired purpose in
the following remarks. Establishing a formal correspondence between both possible representations of cQPL is obvious and follows exactly the argumentation used in [Sel04b]; we
will thus not bore the reader with details on how to relate both representations uniquely.
Note that we try to keep the purely classical formalism as terse as possible because
most problems related with this are not too interesting from a physical point of view. More
elaborate descriptions or gentle introductions can be found, e.g., in Refs. [Mos90, Rey98,
Win93].
5.3.6.1
Some notational remarks
Some conventions and notations widespread in semantics are uncommon in physics, thus we
want to make two short remarks before proceeding further.
Typed lambda calculus Computer science literature habitually uses the typed lambda
calculus (cf., e.g., [RP02]) to formulate the equations for valuation functions; this is useful to
18 It should be noted that although informal may sound a little fuzzy, such a description is normally the
maximal level of accuracy with which users of programming languages (and in most cases, implementors as
well) are confronted.
60
5.3. Denotational semantics of cQPL
not only clarify which parameters are used, but also to determine their type. We, in contrast,
use a different notation. Observe, for example, the denotation of the dyadic operator + which
adds two natural numbers:
DOJ+K(n1 , n1 ) = sum(n1 , n2 )
(5.29)
It is intuitively clear that we are talking about a function which takes two natural numbers
as arguments and computes another natural number as result. Nevertheless, we did not
formally specify the data types of the arguments nor of the result of the function.
This can be solved by using the typed lambda calculus in which the function would be
written as:
(5.30)
DOJ+K = λn1 ∈ N.λn2 ∈ N.sum(n1 , n2 )
This very simple example already demonstrates that the representation requires a considerable notational effort. Since nearly all valuation functions for the semantics of cQPL defined
in the following require (K, T, E) tuples in addition to the effective parameters, this would
unduly inflate the length of equations which does not really increase lucidity. Thus, we
stick to a simplified notation which follows the algebraic convention for functions as presented above. The domains where parameters originate from are normally clear from the
context; we will mention it explicitely should this not be the case since typing is obviously
not explicitely part of the simplified description.
Currying/Schönfinkeln Another point we want to mention is the insight that functions of
more than one parameter may always be composed by a number of subsequent functions
which take exactly one parameter. Thus, a function f (x1 , x2 , x3 ) = y with xi , y ∈ N which
is an element of (N × N × N) → N can also be seen as a mapping (N → N → N) → N, the
parentheses may also be omitted. The technique of transforming a function with multiple
arguments into a function which takes only one argument, but returns another function
which requires the remaining arguments and returns the result is conventionally termed
currying, although it was first introduced by Schönfinkel [Sch24]. The process can obviously
be repeated so that there are only functions left which take exactly one argument. In the
following, we will use the form which is more apt for the respective purpose.
5.3.6.2
State transformations and fixed points
Valuation functions for top-level elements of cQPL (i.e., those for expressions) work on a
three-tuple (K, T, E) and produce a new three-tuple (K ′ , T ′ , E ′ ) as we will see in the course
of the following remarks. Thus, these tuples form the domain which is the basis of semantics.
Since we will need fixed points as solution of several recursive domain equations, we need to
show how to calculate them for (K, T, E) tuples. We have already shown that fixed points
exist for Kraus aggregations which fulfil certain conditions. Now, we need to transfer this
to (K, T, E) tuples.
For this, note that the typing context is not involved in the calculation of fixed points:
Its purpose is to ensure well-typedness of programs (which will be explained in more detail
in Section 5.3.8) and (as a consequence) to make sure that ownership of qbits is unique.
Otherwise, it has no semantical meaning.
To consider the contribution of the probabilistic environment, observe that we could do
without it in principle by using a different notation as is, for example, the case in QPL
(we did not adopt this approach because it quickly leads to very long annotations which are
cumbersome to handle; additionally, our approach has a greater similarity with the standard
notation commonly used in denotational semantics). For every possible value of a variable
61
Chapter 5. Formal denotational semantics
in the probabilistic environment, a specific Kraus aggregation can be inferred. Consider
a measurement of two qbits whose result is stored in a classical variable of two bits. A
πx :PROJ (q,T )
probability distribution x −−−−−−−−−−→ [0, 1, 2, 3] is stored in the probabilistic environment
where PROJ (q, T ) represents the information that the quantum variable q contained in
the typing context T was measured (this will be covered in more detail in Section 5.3.6.4
on page 69). Since the exact form of the probability distribution depends on the state of
the measured variable about which nothing is known in the worst case (if, e.g., the variable
was received from a remote party which did not characterise it any further), we have to
account for all possible cases, i.e., for all values the variable can have in principle. Let
{Γ}; {M }#q be the contents of the Kraus list K from the (K, T, E) tuple immediately after
the measurement where {M }#q denotes the Kraus set for a projective measurement. This
list can with the help of the probabilistic environment be rewritten into a four-tuple
({Γ}; {P0 }, {Γ}; {P1}, {Γ}; {P2 }, {Γ}; {P3})
(5.31)
where {Pi } = |ii hi| is one of the projection operators which constitute the POVM elements
of the measurement. Note that the Kraus aggregations contained in tuples of this kind are
always unparametrised.19
If there is now an operation performed which is independent of the measured variable, the
contribution to the Kraus set is appended to all list components in this picture. Consider, for
example, the application of some operator U to another quantum variable v. The resulting
four-tuple of Kraus aggregations then looks like:
({Γ}; {P0 }; {U }#v , {Γ}; {P1 }; {U }#v , {Γ}; {P2 }; {U }#v , {Γ}; {P3 }; {U }#v ).
(5.32)
The first entry belongs to the case that x = 0, the second to x = 1, and so on. Obviously,
it is much simpler to write this in our notation as
({Γ}; {M }#q ; {U }#v )
(5.33)
from which the tuple representation can be reconstructed. This also works if operations
are considered that depend on the state of a classical variable governed by a probability
distribution. Consider, for example, the case that an operator U is applied to some quantum
variable v if the value of x is 2 (the program code for such an operation would be if (x = 2)
then v *= U;). In our notation, the branching condition is preserved in the probabilistic
environment as shown in the example given by Figure 5.5. From this information, the
corresponding tuple representation
({Γ}; {P0 }, {Γ}; {P1 }, {Γ}; {P2}; {U }#v , {Γ}; {P3 }; ).
(5.34)
can be constructed. Note that U is only applied in the case x = 2.
The transfer from our representation to tuples of Kraus aggregations is easier when
classical variables with defined values and no associated uncertainty are considered, so we
will not show an explicit example for this.
As the forgoing considerations have demonstrated, the probabilistic environment and
the Kraus aggregation contained in the (K, T, E) tuple can be used to construct a tuple of
Kraus aggregations where the tuple contains one entry for every combination of values the
classical variables can be in. Fixed points of (K, T, E) triples must therefore be calculated
separately for all possible Kraus aggregations that can be constructed from the triple because
19 This kind of split is one of the reasons why it is more convenient to work with Kraus representations of
cp-maps instead of the cp-maps alone which would also be possible in principle.
62
5.3. Denotational semantics of cQPL
each of them represents another possible meaning of the program. This is possible with the
methods introduced before. After the fixed points have been derived, the usual (K, T, E)
representation can be used again. Thus, calculation of fixed points effectively only requires
the properties of Kraus sets as introduced before. In the following, we need thus only make
sure that the conditions for the existence of fixed points of Kraus aggregations as given in
Theorem 5.3.1 are fulfilled to ensure the existence of fixed points for (K, T, E) tuples.
5.3.6.3
Classical operations
The classical subsystem of cQPL consists of the following parts:
❏ Allocation and (implicit) deallocation of classical variables.
❏ If-then-else expressions.
❏ While-loops (this and the previous point require the evaluation of boolean conditions
which must also be accounted for by the semantics).
❏ Sequential composition.
❏ Do-nothing-operation (skip).
❏ Sending and receiving of classical states.
❏ Assignment to classical variables.
❏ Calling procedures which manipulate classical data.
❏ Blocks.
Note that we do not cover sendig and receiving of classical data because this has extensively been covered in the literature. Additionally, it is in principle always possible to
achieve the same effects with the transmission of quantum mechanical information. In the
following, we cover only the valuation functions which are either absolutely indispensable or
are influenced by the quantum properties of our language.
Sequential composition Modifications made to the typing context, the probabilistic environment and the Kraus aggregation made by the first statement must be taken into account
when the semantics of the second statement is calculated:
EX PJS1 ; S2 K(K, T, E) = EX PJS2 K(EX PJS1 K(K, T, E))
|
{z
}
(5.35)
=(K ′ ,T ′ ,E ′ )
This is utilised many times in the denotational equations for quantum communication.
Blocks Blocks are used to introduce multiple levels of scope into programs. This can
happen both implicitly (e.g., in loops) and explicitely (by syntactical specification of blocks),
but there is no need to distinguish between these cases from a denotational point of view.
Superficially, a block looks just like a collection of multiple statements which are executed
one after another; but some additional points need to be taken into account:
❏ New variables (both quantum and classical) may be declared inside blocks, but they
cease to exist once the block’s scope is left.
63
Chapter 5. Formal denotational semantics
❏ New variables do overshade old ones if they share the identifier.
❏ Changed bindings of already existing variables are also visible after the control flow
has left the block’s scope.
Thus, the probabilistic environment is partially affected by a block. The typing context
before and after the block is identical and thus unaffected by the block’s effect,20 and the
Kraus aggregation records everything that has been done inside the block.
Since the mentioned problems appear in every programming language featuring blocks,
standard solutions are available in every textbook (as usual, cf. Refs. [Mos90, Rey98,
Win93]), so we will not explictely present them here to save some formal overhead.
Conditionals and Operators Dyadic operators combine two subexpressions into one result,
as, e.g., all arithmetic operations do. Conditionals are operators which result in a boolean
variable, i.e., they evaluate to one of the values true or false which are represented by 1
and 0. In contrast to most classical languages, the result of both types need not be fixed
with certainty, but can be governed by a probability distribution. Note that conditionals
may not be used as stand-alone expressions, but can only be part of conditional statements.
This is why their semantic valuation does not result in the usual (K, T, E) tuple, but in
a probability distribution for the possible results which is, e.g., apt for inclusion into the
probabilistic environment. The basic valuation functions are given by:
OPJa DO bK(T, E) = DOJDOK(OP (a)(T, E) ⊗ OP(b)(T, E))
OPJMO aK(T, E) = MOJM OK(OP (a)(T, E))
(5.36)
(5.37)
where DO ∈ { +, −, ∧, ∨, . . . } and MO ∈ { ¬, − }. The meaning of the operations is defined
as usual, but the probability distribution nature of the arguments needs to be taken into
account:
M
M
πa (v1 )πb (v2 )DO(v1 , v2 ).
(5.38)
DOJDOK(a, b) =
v1 ∈range(a) v2 ∈range(b)
This expression results in a new probability distribution that can be used by the elements
further up in the evaluation hierarchy.
Note that the denotation of a single variable is given by the corresponding probability
distribution which can be found in the probabilistic environment:
OPJvK(T, E) = E(v)
(5.39)
This definition ensures that chains of expressions using dyadic and monadic operators
(e.g., 42 + 23 + 4) are covered by the semantics because a and b in Eqns. 5.36, 5.37 can either
be values or other operator expressions.
Also note that the eventual action of the respective operators (+, ∧, . . . ) can be seen
intuitively, so we abstain from further formalisation and rely on the reader to use his common
mathematical sense.
Remark 5.3.8. Note that we do neither consider any problems of numerical accuracy nor
of limited ranges of representable numbers for the specific data types. Consequently, we also
do not care for the problem of division by zero. We are aware that such pitfalls exist, but
are not interested in their solution in this context since their nature is purely classical.
20 But
note that the typing context within the block may well be different than the one outside.
64
5.3. Denotational semantics of cQPL
If-Statements The semantic description of the if-statement is simplified by introducing
the following helper function:
f ((K0 ,T0 , E0 ), (K1 , T1 , E1 ), π, ν, (K, T, E)) =
p(ν = 0) · K ◦ (K0 − K) + p(ν = 1) · K ◦ (K1 − K),
T ⊕ ν : bit, p(ν = 0) · E ⊕ ν : π ⊕ E0 + p(ν = 1) · E ⊕ ν : π ⊕ E1
(5.40)
which eases selection of components of (K, T, E) tuples gained by other evaluations and
additionally circumvents repeated semantic evaluations of some components. The tuple
(K0 , T0 , E0 ) is the result of the evaluation of the if-branch, while (K1 , T1 , E1 ) is for the
then branch. π is the probability distribution governing the branch, and ν is the identifier
which is used to represent this distribution in the probabilistic environment. (Ki − K)
represents the Kraus aggregation that contains only the elements that were appended to Ki
in comparision to K; this ensures that only new contributions introduced in the branches
are added to the Kraus aggregation finally. The denotational description for the if-statement
then reads as
Jif c then C0 else C1 K(K, T, E) =
uid , (K, T, E))
f (EX PJC0 K(K, T, E), EX PJC1 K(K, T, E), OPJcK(T, E), |{z}
{z
} |
{z
} |
{z
}
|
(K0 ,T0 ,E0 )
(K1 ,T1 ,E1 )
π
(5.41)
ν
where uid is a unique identifier for the branch which can be chosen at will, but must not be
identical with other identifiers already in use. Such a choice is simple for non-communicating
programs. The extension to communicating systems is possible if every identifier is given a
unique prefix for each partner as described before.
Note that although dyadic operators might syntactically be used to describe arithmetic
operations and not necessary conditionals, this source of mistake is ruled out by the type
system which only allows boolean typed expressions for c.
While-Statements While statements can be solved using the fixed-point theorem given in
Eqn. 4.4.1. For this, note that the denotation of the while function can be rewritten using
the previously considered if-function together with an explicit block (c denotes a boolean
condition and S a statement):
JwK(K, T, E) ≡ Jwhile c do SK(K, T, E)
JwK(K, T, E) = Jif c then {c; w} else skipK(K, T, E)
(5.42)
(5.43)
Eqn. 5.43 is a recursive equation (c appears both on the left hand and the right hand side)
whose solution is a fixed point. The agreement is that the least fixed point is taken to be
the denotational solution, and this is exactly what the fixed point theorem delivers. To
write this formally is now a standard exercise of denotational semantics [Rey98, Mos90],
but we show it nevertheless because it is an instructive example for the technique of solving
recursive equations which will be necessary for the denotation of quantum communication.
Consider the function F given by
F f (K, T, E) = if JcK(K,T,E) then f (JSK(K, T, E)) else (K, T, E).
(5.44)
Then the fixed point solution can be formally written as
Jwhile c do SK = YΣ→Σ⊥ F,
65
(5.45)
Chapter 5. Formal denotational semantics
where Σ is any (K, T, E) tuple as usual; assume that f is defined like
(
⊥
if x = ⊥
f (x) =
f (x) otherwise
(5.46)
since we have to account for the case that the argument of f is ⊥ because of the recursion
(note that we would have to use two different symbols for f to write this überproperly).
Assignments Assignments in cQPL can be seen as a convenience mechanism which extends
the simple binding of identifiers to values; it is not necessary to introduce the concept of
stateful variables to the language to be able to formalise assignments. The description can
be simplified if some syntactical transformations are applied. For this, first consider the
following program fragment:
new int a := 10;
// Do something using a (part 1)
a := a + 4;
// Do something using a (part 2)
The assignment is equivalent to introducing a new identifier a’:
new int a := 10;
// Do something using a (part 1)
new int a’ := a + 4;
// Do something using a’ (and replace all a by a’) (part 2)
This strategy also works when blocks are taken into consideration:
new int a := 10;
if (...) {
new int a := 5;
a := a + 1;
...
}
else {
new int a := 1;
a := a + 1;
...
}
Note that not possible to employ a static renaming scheme in this case because the
identifiers a in both subsequent blocks would then end up with identical names which leads
to problems. We thus have to postulate that new identifiers are always chosen such that
they do not overlap with previous identifiers and must, of course, also not overlap with
identifiers which can be assigned by the user. This is possible if the set of identifier strings
is denoted by Σ∗1 , we introduce a second set of strings Σ∗2 with Σ1 ∩ Σ2 = ∅; each time
an identifier is overshaded, it is replaced in its complete scope with a new one in Σ∗2 that
has not been used before. While this policy is hard to implement in practice,21 it does not
21 This is exactly the reason why mechanisms like call-by-name disappeared as a curiosity some thirty
years ago.
66
5.3. Denotational semantics of cQPL
present any problem in theory (even cases which require an infinite number of replacements
are no problem because there are infinitely many unique identifiers).
Most important, the approach is also valid when loops are introduced because these can
be rewritten using a (possibly infinite) chain of appropriate if-then constructions as is done
in the denotation of them.
In conclusion, we do not need to take care for the obstacles introduced by overshading, but can simply ignore the problem in the denotation of assignments (ν denotes some
identifier):
Jν := ǫK(K, T, E) = (K, T, E ⊕ ν : EQN JǫK(K, T, E))
(5.47)
where ǫ is some arbitrary arithmetic expression (which includes single identifiers); the denotation of this is obvious and will not be considered further. Note that this valuation function
does not cover the case that the result of a quantum variable measurement is stored in a
classical variable; this will be covered later when we describe the denotation of the measure
funtion on Page 69.
Allocating and destroying variables As we have noted in the previous remarks, we do not
need to take the problem of overshading into account when dealing with assignments; this
obviously also applies to allocations. We refrain from a more detailled description of this
topic because everything necessary for the solution can be readily found in the literature, e.g.,
[Win93, Rey98, Mos90]. Note that allocating new variables does not present any problems
for the boundedness of the Kraus aggregation because the number of qbits is limited by Q,
an arbitrary, but finite quantity.
Procedure handling Procedures in cQPL follow the standard scheme of classical languages
for the non-quantum part. The denotation of such can therefore be directly taken from the
usual textbooks [Rey98, Mos90, Win93], so we will not repeat this here. The interesting
problem is given by recursive procedures, especially when they act on quantum parameters.
For simplicity, we consider directly recursive procedures with quantum parameters; the case
of indirect recursion is in princple identical, but necessitates more formal effort, so we skip
it here. The solution is an adaption of the method presented in [Sel04b, Sections 5.5 and
6.5] for our purposes.
Consider a procedure Y which depends on itself, e.g.,
Y = X(Y )
(5.48)
If Y is given as a flow chart, this can be interpreted as shown in Figure 5.7: A “hole” in the
representation of X is replaced by Y with another hole, this is again replaced by the same,
...
proc rec: test:qbit {
A
if (cond) call rec(test’);
else { ... }
B
}
Formally, we can thus define an approximation relation given by
Yi+1 = X(Yi ).
67
(5.49)
Chapter 5. Formal denotational semantics
Figure 5.7: Unwinding a recursive flow chart is done by placing the flow chart with
a hole (given by the white rectangle) into the hole and repeating the
process again and again. The limit of this sequence is given by a fixed
point as explained in the text.
where Yi are approximations of Y which get better with increasing i. Y0 = ⊥ is the crudest
approximation which simply does not terminate. The solution of Equation 5.49 is another
case for the fixed-point theorem. Once a solution for Y has been found, it can be applied to
any (K, T, E) tuple to calculate the required state transformation. Note that although the
procedure depends only on quantum variables, there nevertheless needs to be a termination
condition. In the example, this condition is given by the if-then-else statement in which
the recursive call is wrapped; since the procedure only takes a quantum parameter, the
condition is either trivial or depends on a measurement of quantum states. In the first case,
the condition depends only on classical variables which were allocated within the procedure,
the condition could thus be determined at compile-time and the recursion unrolled because
the recursion depth is already known. In the second case, we really need to unwind the
procedure.
5.3.6.4
Quantum operations
The quantum mechanically relevant operations of cQPL are :
68
5.3. Denotational semantics of cQPL
❏ Application of (unitary) operators.
❏ Calling procedures which manipulate quantum data.
❏ Sending and receiving quantum states.
❏ Measurements.
❏ Creating new and destroying existing quantum states, where the last operation is not
explicitely, but only implicitly possible when quantum variables drop out of the present
frame. Sending quantum variables makes them disappear from the scope so that they
can not be accessed any more, but does not destroy respectively deallocate them.
We will give formal denotations for these constructs in the remaining part of this section.
Obviously, the description of communication is our foremost concern, so we elaborate this
in most detail.
Unitary operators Unitary operators induce isometries, so the corresponding Kraus set is
obviously bounded and filfills the requirement for fixed points. Every unitary transformation can be expressed by a Kraus set with only one element. The semantics of a unitary
transformation acting on the qbits q1 , . . . , qk is given by
EX PJ(q1 ,...,qk ) *= UK(K, T, E) = (K ◦ {U }(#q1 ,...,#qk ) , T, E).
(5.50)
Note that the type system ensures that the variable/operator dimension on both sides of
the expression matches as required, i.e., the classical variable has the proper size to hold all
potential measurement outcomes.
Measurements Measurements are described by the following semantic equation:
πa :PROJ (q,T )
EX PJa := measure qK(K, T, E) = K ◦ {M }#q , T, E ⊕ a −−−−−−−−−−→ range(a) .
(5.51)
Hereby, PROJ (q, T ) denotes the function to generate the set of projectors for the type of
q which can be resolved from the typing context T . The required basis for the projection
t (χ(q))
(which is actually nothing else than the basis for F2q
in Dirac notation, also named the
standard basis) is given by
(χ(q))
tq O
B=
|in i ∀in ∈ {0, 1} .
(5.52)
n=0
The corresponding Kraus elements are obvious. Accordingly, πa : PROJ (q, T ) denotes the
probability distribution which connects the possible values of a with a probability distribution induced by the projectors. Since the measurement operators work on discrete states in
a finite-dimensional Hilbert space, they are obviously bounded.
Remark 5.3.9. Note that although we restrict measurements to projections onto the standard basis, projective measurements in arbitrary bases can be realised by applying appropriate
unitary transformations prior to the measurement.
69
Chapter 5. Formal denotational semantics
Sending and receiving qbits To consider sending quantum variables, we first split commands which send lists of quantum variables into a list of commands that send one quantum
variable each:
EX PJsend q1 , . . . , qn to moduleK(K, T, E) =
EX PJsend q1 to module; send q2 to module; . . . ; send qn to moduleK(K, T, E).
(5.53)
The denotation of a single send command is given by
EX PJsend q to moduleK(K, T, E) = K ◦ {S}#q , T ⊖ q, E ⊖ q .
(5.54)
{S}#q is not a real physical map as other cp-maps are, but only a “placeholder” to note that
qbits have been sent. This will become important when the semantics of parallel execution
is considered further below. The interesting thing here is that the sent qbit is removed from
the typing context and from the probabilistic environment. Thus, it is not visible any more
in the remaining statements. Further access to it can be detected as erroneous at compile
time; the typing context gives the formal basis for this.
Receiving qbits is described in a similar manner. First, a receive operation with multiple
quantum variables is split into a sequence of single-variable receive operations:
EX PJreceive q1 : qtype1 , . . . , qn : qtypen from moduleK(K, T, E) =
(5.55)
EX PJreceive q1 : qtype1 from module; . . . ; receive qn : qtypen from moduleK(K, T, E).
The denotation of a single-variable receive command is given by:
EX PJreceive q : qtype from moduleK(K, T, E) =
K ◦ {R}#q , T ⊕ q : qtype, E ⊕ q : module .
(5.56)
Again, {R}#q is a placeholder required to denote the semantics of parallel composition.
Parallel composition The compositionality principle of denotational semantics implies that
the formal denotation of sending and receiving qbits must be independent of the conjugate
action, i.e., sending must be independent from reception and reception must be independent from sending. This is fulfilled in the formalism of cQPL as we have shown on page
70. Nevertheless, the denotation of the communication as a whole requires (at last) two
communicating partners. It needs thus make use of both of them to assign semantics to
communication.
Before we start with the formal details, we want to motivate why parallel composition
is necessary at all. For this, consider the case of two processes where A sends a quantum
bit (which we call a1 ) to B and, later on, receives a quantum bit from B (which we call a2 ).
We must distinguish two different cases (note that for more difficult cases with an arbitrary
number of send and receive statements, the conditions become more complicated because
we need to account for more general schemes of mutual influence. These conditions will be
developed stepwise in the following):
❏ The returned quantum bit was not identical with the sent one, a1 6= a2 . The inequality
refers to the positions occupied by the qbits on the combined quantum heap, it is not
related with the names of the qbits.
❏ B returned the quantum bit it got from A, a1 = a2 .
70
5.3. Denotational semantics of cQPL
Consider the consequences for the semantics of parallel composition when the interaction
of both processes is considered (both cases can be treated identically if only the separated
semantics is taken into account): While in the first case (a1 6= a2 ), operations on a2 and a1
end up on different physical locations, the same operations must in the second case (a1 = a2 )
be applied to the same physical location. Therefore, we must be able to construct enough
information from the composed systems such that it is possible to distinguish between both
cases.
Additionally, the distinctness of qbits influences denotational equivalence; many different
orderings of the actions performed by A and B lead to the same semantics (the exact
conditions for this will be formulated later on), but the class of possible reorderings is
usually bigger if there was no interaction on the same physical location in communicating
modules. Figure 5.8 presents a visualisation of this fact using a pseudo flow diagram.
Figure 5.8: Flow graph for two communicating processes. A sends a qbit to B
and B sends one to A; if there are no operations on the same physical
locations, the shaded regions commute. This is obviously not the case
if one or more identical quantum heap positions are modified in both
paths: Since the shaded regions operate on the same physical qbit
then, their actions need not necessarily be interchangeable because
non-commuting operations may have been performed on the same
physical location.
There is no explicit statement in the syntax of cQPL which describes the combination of
two processes. As explained in Chapter 2, communication is realised implicitly using modules
which interact among each other. These modules must thus be given some semantical
meaning; this will be developed in the following. Note that we restrict the description to
two communicating processes which we call A and B (this might stand as an abbreviation
for the omnipresent Alice and Bob) at first. Consider the following cQPL fragment:
71
Chapter 5. Formal denotational semantics
module A {
...
};
module B {
...
};
Any valid cQPL code (except the definition of new modules) may be contained in the
bodies of module A and module B; both entities thus constitute regular cQPL programs
whose meaning is made up by the meanings of all statements they consist of. Thus, we
can write PROGJAK and PROGJBK for the denotation of the code as if the modules were
regular, uncommunicating cQPL programs.
To consider the communicative interactions between both, we introduce the operation
PROGJA||BK ≡ COMM(PROGJAK, PROGJBK)(K∅ , T∅ , E∅ )
(5.57)
where (K∅ , T∅ , E∅ ) denotes the inital (K, T, E) tuple without content. The valuation function COMM is used to compute the combined denotation of A and B which we also call
parallel execution. It must obviously only depend on the denotations JAK and JBK, i.e.,
(K, T, E) tuples. Since we do not want the semantics of communication to depend on the
order in which the communicating partners are specified, the operator || must necessarily
!
be commutative: PROGJA||BK = PROGJB||AK. This equivalence can easily be achieved:
It suffices to define the operator in such a way that the modules are ordered lexicographically, then the order in which they are specified does not influence the denotation; the
commutation relation is thus automatically fulfilled.
PROGJAK is used to evaluate a semantic equation with empty initial context; this is
obviously the case when the top-level of a program is considered (as is the case for modules)
where no definitions can have been made yet. Thus, PROGJAK ≡ EX PJAK(K∅ , T∅ , E∅ ).
Consider the evaluation of JAK and JBK which both do as usual depend on (K, T, E)
tuples. COMM results in the creation of a valuation function which depends on the tensor
product of these tuples, i.e.,
VALJAK(K0 , T0 , E0 )
VALJBK(K1 , T1 , E1 )
⇒ COMM(VALJAK, VALJBK) (K, T, E) .
| {z }
(5.58)
(K0 ⊗ K1 , T0 ⊗ T1 , E0 ⊗ E1 )
Both representations convey the same information. This is immediately obvious if the expressions are written as direct λ-abstractions as defined in Section 5.3.6.1 together with
an appropriate “untensoring” function which separates the tensor product of the combined
(K, T, E) tuple into two components. We will not show this explicitely to avoid introducing
even more symbols.
Note that termination of A and B alone does not imply termination of PROGJA||BK.22
This can be seen by considering the following simple program:
22 Here, the question arises how termination should be defined for communicating processes if only one
part (A) of a total system is considered. We could, for instance, define a demonic partner which always
supplies the required number of qbits the process needs and absorbs any number of qbits sent, but this will
not necessary lead to termination of A. The exact solution of the problem depends on the context in which
it is considered; we will not deal with it any further here, but define a process as terminating if there exists
a demonic partner that behaves in such a way that the process terminates.
72
5.3. Denotational semantics of cQPL
module A {
new qbit q1;
send q1 to B;
};
module B {
receive a1:qbit from A;
receive a2:qbit from A;
};
While both processes represent valid cQPL programs, they will obviously result in some
non-terminating program (which should thus denote ⊥) because B will wait forever for the
second qbit (a2 ) to be sent. Thus, COMM(x, y) = ⊥ is possible although both x 6= ⊥ and
y 6= ⊥.
Extensions of the semantical components (K, T, E) to the multi-party case were defined
in Section 5.3.3; they provide the basis for the denotation of parallel execution which will be
developed stepwise by considering bipartite systems without communication, bipartite systems with single send/receive pairs, bipartite systems with arbitrary send/receive statements
and finally, arbitrary multipartite systems.
Thus, let A and B be two programs which do not use any communication. The semantics
of their parallel execution can be readily denoted:
PROGJAK = (K0 , T0 , E0 ), PROGJBK = (K1 , T1 , E1 )
⇒
(5.59)
PROGJA||BK = (K0 ⊗ K1 , T0 ⊗ T1 , E0 ⊗ E1 ).
Note that the following equivalences hold if Ai , Bi are parts of a program which do not
contain any send/receive operations (we omit the required EX Ps in the second line to
simplify the notation):
EX PJA1 ||B1 ; A2 ||B2 K = EX PJA1 ; A2 ||B1 ; B2 K
{z
} |
{z
}
|
(COMM (JA1 K, JB1 K) ; COMM (JA2 K, JB2 K)) = COMM(JA1 ; A2 K, JB1 ; B2 K).
(5.60)
(5.61)
This ensures that it does not make any difference if we consider the sequential combination
of two parallel executions or the parallel execution of two sequential combinations as shown
in the formula; we will make use of this later on.
The situation gets more complicated if a single send/receive pair is allowed, i.e., if a single
quantum variable can be transferred from A to B (the case B to A is nearly identical, so
we restrict ourselfs to the first case). Let {S} and {R} denote the Kraus pseudo-operations
for sending and receiving. To see how these operations influence the possible equivalent
compositions of a program, observe the following two schematic Kraus aggregations:
ΛA = {A1 }, . . . , {An }, {S}, {An+1}, . . . , {Am }
ΛB = {B1 }, . . . , {Bn′ }, {R}, {Bn′+1 }, . . . , {Bm′ }.
(5.62)
(5.63)
To consider how these aggregations can be rearranged, we define the following function:
(
1 if A and B are operations on disjoint qbits
D(A, B) =
.
(5.64)
0 otherwise
73
Chapter 5. Formal denotational semantics
Remark 5.3.10. The same effect could have been achieved with the standard commutator
in principle, but we want the notation to emphasise additionally that not the simultaneous diagonalisability, but the fact that the operations work on disjoint bases is responsible
for the exchangeability. Additionally, using the standard commutator to show equivalence
between permutations of communicating systems would have interferred with the standard
permutation rules that would then have become more complicated.
Since the two processes are necessarily totally uncorrelated before the transmission takes
place, the parts given by the sub-aggregations
Λ1A = {A1 }, . . . , {An }
Λ1B
(5.65)
= {B1 }, . . . , {Bn′ }
(5.66)
Λ2A = {An+1 }, . . . , {Am }
(5.67)
can be composed as in Eqn. 5.59 because D(Λ1A , Λ1B ) = 1. Since the parts after sending
respectively receiving the quantum variable are uncorrelated as well (they both work on
disjoint sets of qbits; this property remains valid in the combined semantics because the
combination of the probabilistic environments creates an appropriate combined quantum
heap as described in Section 5.3.3), the sub-aggregations
Λ2B
= {Bn′ +1 }, . . . , {Bm′ }
(5.68)
can likewise be parallel composed as in Eqn. 5.59. The only thing which needs to be taken
into account is that references to the position of the received qbit must be replaced by the
position of the qbit on the combined quantum heap in the combined probabilistic environment. This can formally be achieved by the following function (here, S denotes a send and
R a receive statement):
′
⊗
⊗
SR(COMM(JSK, JRK))(K ⊗ , T ⊗ , E ⊗ ) = (K ⊗ , T ⊗ , E ⊗ ).
(5.69)
⊗
Here, K , T , E denotes the usual (K, T, E) parameter tuple with the additional requirement that it must have the structure which is gained by combining two (K, T, E) tuples
with a tensor product as given in Section 5.3.3. SR itself is responsible for two things: On
the one hand, it applies the effect of COMM(JSK, JRK) to the parameter tuple, and on the
other hand, it replaces the portion of E which contains the information about the received
quantum variable so that it now points to the position of the sent quantum variable on the
combined quantum heap; the resulting probabilistic environment is denoted by E ′ . This
is obviously possible since SR does have access to the information provided both by send
and receive. To illustrate the effect of SR, consider the following example: Let the sent
quantum variable be denoted by q and the received one by r. The names of these variables
will have thus been changed to Lc (1)q and Lc (2)r. The receiving module does not have
any information about the received quantum variable except its type and its local name;
the position on the combined quantum heap is unknown. This can now be changed by SR;
for that, it inserts the position of Lc (1)q on the combined quantum heap into the combined probabilistic environment such that Lc (2)r points to it. Nothing more is necessary to
identify the received quantum variable with the sent one.
Note that RS defines the analogous function for the receive/send case. This is necessary
when bidirectional communication between processes is considered. Except the inverted
direction of data flow, the function is completely equivalent to SR.
Thus, the semantics of parallel execution for processes with a single send/receive operation can be reformulated as
PROGJA||BK = PROGJA1 ||B1 ; S||R; A2 ||B2 K
74
(5.70)
5.3. Denotational semantics of cQPL
where JAi K and JBi K denote the parts of the program which induce the operations described
by ΛiA,B as given by Eqns. 5.65–5.68. It is already known how to compute the semantics of
JA1 ||B1 K. To compute the denotation of the complete statement, we first consider how to
include the send/receive pair into the description:
PROGJA1 ||B1 ; S||RK = SR(COMM(JSK, JRK))
⊗
⊗
(COMM(PROGJA1 K, PROGJB1 K)(K∅
, T∅⊗ , E∅
)).
(5.71)
Note that although PROG with argument JA1 K, JB1 K appears both on the left and right hand
side of this equation, it is not truly recursive, but can here be seen as just a breakdown into
simpler cases. A general recursive formula will be derived in the following.
By using Eqn. 5.70, we can now give the denotation of the complete parallel composition
as defined in Eqn. 5.61; for this, we define the righthand side of Eqn. 5.71 to be denoted by
ξ to increase clarity:
PROGJA1 ||B1 ; S||R; A2 ||B2 K = COMM(PROGJA2 K, PROGJB2 K)(ξ).
(5.72)
Note that establishing the semantics of a given description is one thing we need to do;
finding equivalent descriptions for a given communication is another task. For this, we need
to consider all rearrangements that preserve semantics. This allows us to decide if two
communicating programs are identical because we can check if they can be brought to the
same form.
For the case of a single send/receive pair, the operations can be shifted in the Kraus
aggregation if certain conditions hold:
❏ The send operation can be postponed to the end of the aggregation (at least in the case
where no more send/receive operations take place) or brought forward by k positions
in the Kraus aggregation if D({An−l }, S) = 1 ∀l = 0, . . . , k − 1.
❏ The receive operation can be brought forward to the first position of the Kraus aggregation (again, this relies on the fact that only a single send/receive operations takes
place) or postponed by k positions if D({Bn+1+l }, R) = 1 ∀l = 0, . . . , k − 1.
Performing shifts characterised by these operations together with reorderings as given
by Eqn. 5.26 generates the equivalence class of all programs with a single send/receive pair
that posses the same semantics.
The next step is to include arbitrary send/receive operations into the communication
between A and B. As in the case of a single send/receive combination, the send/receive
statements act as synchronisation points; the denotation needs thus be aligned along them.
The problem to solve is now given by
PROGJA1 ||B1 ; S1 ||R1 ; A2 ||B2 ; S3 ||R3 ; · · · ; Sn−1 ||Rn−1 ; An ||Bn K.
(5.73)
First, we will establish the semantics for the given ordering; semantics-preserving permutations will be considered afterwards.
Note that we only consider the denotation of data flow in one direction for simplicity;
the semantics of the case R||S which may appear mixed with the other form is gained by
replacing SR with RS at the appropriate places). To find a solution for this equation, define
A′ ||B ′ ≡ A2 ||B2 ; Sn ||Rn ; · · · ; Sn−1 ||Rn−1 ; An ||Bn .
(5.74)
Eqn. 5.73 then has the form
PROGJA1 ||B1 ; S1 ||R1 ; A′ ||B ′ K.
75
(5.75)
Chapter 5. Formal denotational semantics
According to Eqn. 5.72, the solution of Eqn. 5.75 is given by
EX PJA′ ||B ′ K(SR(COMM(JS1 K, JR1 K))
⊗
⊗
(COMM(PROGJA1 K, PROGJB1 K))(K∅
, T∅⊗ , E∅
)).
(5.76)
By introducing ξ1 as abbreviation for the part following EX PJA′ ||B ′ K and expanding A′ ||B ′
to A2 ||B2 ; S2 ||R2 ; A′′ ||B ′′ , the formula reads as
EX PJA2 ||B2 ; S2 ||R2 ; A′′ ||B ′′ K(ξ1 ).
(5.77)
This type of equation is already well-known; it can be further resolved to
EX PJA′′ ||B ′′ K(SR(COMM(JS2 K, JR2 K))
(COMM(PROGJA2 K, PROGJB2 K))(ξ1 )).
(5.78)
By recursively defining
⊗
⊗
ξ0 = (K⊥
, T⊥⊗ , E⊥
)
(5.79)
ξi = (SR(COMM(JSi K, JRi K))(COMM(PROGJAi K, PROGJBi K))(ξi−1 ))
(5.80)
we see that the final solution of Eqn. 5.73 is given by
EX PJAn ||Bn K(ξn )
(5.81)
where ξn needs to be expanded as defined above.
The whole process thus leads to a recursive valuation function given by
EX PJA||B; S||R; C||DK(K, T, E) =
EX PJC||DK(SR(COMM(JSK, JRK))(COMM(EX PJAK, EX PJBK))(K, T, E)
(5.82)
whose solution can be found by resolving the recursion in the usual way.
Now, consider which alternative orderings of the Kraus aggregations preserve semantics
in the multi send/receive and receive/send case. For this, observe the following symbolic
representation of two Kraus aggregations (to save some notational effort and to increase
lucidity, we represent the Kraus sets which are not concerned with communication by boxes.
Although the boxes have identical widths, they do not need to contain the same number of
Kraus sets):
A1
S1A
B1
R1B
A2
R1A
A3
B2
S1B
B3
(5.83)
.
(5.84)
As in the case of a single send/receive pair, we know that the blocks A1 and B1 (considering again the compatible displacements of S1A and R1B ) can be arbitrarily combined
because they work on disjoint subsets of the quantum heap; the same holds for A2 , B2 and
A3 , B3 . In addition to the previously given rules, the following restrictions hold for shifting
send and receive operations in multi send/receive scenarios:
❏ Two consecutive send statements can only be interchanged if the corresponding receive
statements are interchanged, and vice versa.
76
5.3. Denotational semantics of cQPL
❏ If D(Ai , Bi+1 ) = D(Bi , Ai+1 ) = 1 the blocks Ai , Ai+1 and Bi , Bi+1 can be taken like
single blocks when possible orderings according to Eqn. 5.26 are considered. Note that
the validity of this condition can only be detected if the combined quantum heap is
considered. If it holds, then the sender does nothing to the sent and the receive does
nothing to the received quantum bit; thus, the statements contained in the blocks can
be executed in arbitrary order.
Two systems are identical if their denotations can be unified by performing rearrangements
according to these rules.
Another possibility that needs to be considered is the case where the number of send/receive pairs in the parallel processes does not match. As we have described before, this leads
to a non-terminating process because either the sending process wants to ship a quantum
variable, but cannot deliver it and thus blocks or the receiver wants to get a variable, but
blocks indefinitely because there is no sender for one. The semantics should thus be given
by ⊥ for both cases.
For the case of two parties (where we do not need to consider the case that a balanced
number of send/receive statements is present, but the distribution among the parties is
unmatched), this is covered by the following definition:
EX PJA||BK = ⊥ if R(JAK) 6= S(JAK) ∨ S(JAK) 6= R(JBK)
(5.85)
where S denotes the number of send and R the number of receive statements. The denotations JAK and JBK are supposed to be those derived for the parallel composition. The
extension to higher-dimensional systems is obvious, so we omit it here.
Finally, we can describe the semantics for multi-party communication with arbitrary
send/receive statements which is the most general case and therefore includes everything
considered before. The problem which needs to be evaluated is given by
PROGJA1 ||A2 || · · · ||An K,
(5.86)
where (note the different notation compared to before!) Ai represents all statements given
in module i; this may contain any number of send/receive statements that are now denoted
by Ski and Rki where i is the receiver for S and destination for R and k the sequence number
within the other send/receive statements of the communication channel the statement works
in (if we consider for example three parties A1 ,A2 and A3 , then there are the channels A1 –A2 ,
A1 –A3 , A2 –A3 ).
The semantic context the evaluation is based on is given by the tensor product of the
semantic contexts of the subsystems, i.e., (K ⊗ , T ⊗ , E ⊗ ) = (K1 ⊗ · · · ⊗ Kn , T1 ⊗ · · · ⊗
Tn , E1 ⊗ · · · ⊗ En ) and equivalent for the initial context. The valuation function COMM
given by Eqn. 5.57 can be extended from the two-party case to the n-party case without any
problems, we denote this by COMM⊗n . Since communication still takes place between two
partners (although there are now many choices for such two-partner subsystems), there is
always a pair of corresponding send/receive respectively receive/send statements. To take
the n-dimensional semantical context into account, the definition of SR (the version for
two parties is given in Eqn. 5.69) needs to be adjusted as follows when sending a quantum
variable from system m to system m′ is to be covered:
′
′
′
SRm,m (COMM⊗n (JS m K, JRm K))(K ⊗ , T ⊗, E ⊗ ) = (K ⊗ , T ⊗ , E ⊗ ).
(5.87)
SR is again responsible to apply the effect of COMM⊗n (JSK, JRK) to the parameter
tuple; note that in this case, do-nothing-operations 1 are used for all systems except m and
77
Chapter 5. Formal denotational semantics
m′ because these are not concerned with the communication. This is to ensure that the
dimensionality matches.
Additionally, SR replaces the portion of E which contains the information about the
received quantum variable so that it now points to the position of the sent quantum variable
on the combined quantum heap. This is identical to the effect in the simplified case for two
systems. If in this case the sent quantum variable is denoted by q in system m and the
received one by r in system m′ , then the names of these variables will have been changed to
Lc (m)q and Lc (m′ )r. SR simply inserts the position of Lc (m)q on the combined quantum
heap into the combined probabilistic environment such that Lc (m′ )r points to it.
With this, we can generalise the recursive definition of Eqn. 5.82 to the case with an
arbitrary number of participants:
′
EX PJA1 || · · · ||An ; S m ||Rm ; B1 || · · · ||Bn K(K, T, E) =
′
′
EX PJB1 || · · · ||Bn K(SRm,m (COMM⊗n (JS m K, JRm K))
(COMM(EX PJA1 K, . . . , EX PJAn K))(K, T, E).
(5.88)
Finally, this is the solution to the most general case of communication which can be
expressed in cQPL.
5.3.7
Explicit transformations of density matrices
Although the abstract view on quantum operations which we have presented in this work
is quite suitable for reasoning about general formal properties of quantum systems, the demands of practical work are usually of a different nature: Here, one is interested in the
calculation of explicit states and probabilities which determine a system and allow predictions about its past, present and future behaviour. This goal is usually achieved by specifying
the initial state of the system, subjecting this to diverse transformations and measuring the
required properties which give rise to the desired explicit probability distributions.
The semantics of a cQPL program can be used to generate exactly this information:
The abstract transformation given by the semantical denotation of a program is a function
which maps the density matrix of the input state to the density matrix of the output state.
Obviously, the election of a certain density matrix as initial state implies loss of generality,
but in turn allows to infer real-world information, not just abstract properties of generalised
systems.
5.3.8
The type system
Typing judgements make statements about the connection between expressions and their
types; cf., e.g.,[Car97] for an introduction. For our purposes, the following two building
blocks are necessary to describe the properties of cQPL:
E ⊢ e : T ⇔ Expression e has type T in E
E ⊢ F ⇔ F is well-typed in E.
(5.89)
(5.90)
Proper typing is necessary to eliminate certain runtime errors by applying appropriate compile time checks (cf. Section 5.4). Additionally, it is the key to showing that our formalism
ensures that quantum bits can – especially in communicating systems – be only manipulated
by one party at a time (a similar line of reasoning, albeit for a quite different formalism,
was used in [GN05]). The typing context provided by T in the (K, T, E) tuple is the basis
for this.
78
5.3. Denotational semantics of cQPL
Properties of the type system are customary expressed with judgements of the following
general form:
P1 · · · Pn
(5.91)
C
where the Pi are called the premises and C the conclusion. If all premises are true, the
conclusion is fulfilled. Such judgements can be used to deduce the type of a given composite
expression in an automated, formal manner. The following elementary typing judgements
hold for cQPL:23
Scalars
new t n := v
C1 ; C2
Conditionals
Arithmetic
i∈N
(analogous for bits, floats etc.)
E ⊢ i : int
E⊢v:t
E⊢n:t
E ⊢ C1
E ⊢ C2
(Composition preserves well-typedness)
E ⊢ (C1 ; C2 ) : void
E ⊢ v1 : t1 ∧ E ⊢ v2 : t2 c(t1 ) = c(t2 ) = 1
(op ∈ { <, >, =, . . . })
E ⊢ op(v1 , v2 ) : bit
E ⊢ v1 : t1 ∧ E ⊢ v2 : t2 c(t1 ) = c(t2 ) = 1
(op ∈ { +, −, ·, :, . . . })
E ⊢ op(v1 , v2 ) : max(t1 , t2 )
(5.92)
(5.93)
(5.94)
(5.95)
(5.96)
Again, we do not consider division by zero or overflows; up- and downcasting of data
types and procedure handling is also skipped. An equivalence relation between two types
was given by Eqn. 5.11 in Section 5.3.1.1; this can be immediately carried forward to typing
judgements:
E ⊢ x : σ1
.
(5.97)
σ1 ∼
= σ2 ⇒
E ⊢ x : σ2
Note that we do not consider these equivalences explicitely in the following to simplify the
notation, all statements are automatically supposed to hold for all equivalent types as well
without further noting this.
Also note that subtyping (i.e., considering one type as a subtype of another and allowing
appropriate conversions) is not explicitely taken into account because this is also a problem
which is specific to the classical data types of cQPL and thus not of too much interest here.
5.3.8.1
Quantum variable tuples
Tupling of variables must make sure that no component appears more than once in the list
because this would allow to write programs which violate the no-cloning principle and thus
lead to runtime errors. Formally, the requirement is given by24
∀k = 1, . . . , n : q(σk ) = 1 ∧
#xk ∩ (#x1 ∪ · · · ∪ #xk−1 ∪ #xk+1 ∪ · · · ∪ #xn ) = ∅
P
.
E ⊢ (x1 , . . . , xn ) : i σi
(5.98)
The meaning of this is as follows: E ⊢ xi : σi formulates the requirement that all variables
are well-defined. The condition ∀k = 1, . . . , n : q(σk ) = 1 requires that all components are
E ⊢ x1 : σ1 · · · E ⊢ xn : σn
23 Remember:
c(x) = 1 ensures that the data type of x is purely classical.
q(k) = 1 ensures that the data type does not contain any classical components, #q denotes
the positions in the quantum heap occupied by a quantum variable.
24 Remember:
79
Chapter 5. Formal denotational semantics
quantum data types; the tuple thus has no classical components which is justified by our
abdication of mixed types. The condition #xk ∩(#x1 ∪· · ·∪#xk−1 ∪#xk+1 ∪· · ·∪#xn ) = ∅
is a formal version of the requirement that no variable may appear more than once in the
list of variables. The conclusion which can be drawn from these premises is that (x1 , . . . , xn )
is a proper quantum variable tuple, i.e., a well-typed expression in the current environment.
5.3.8.2
Application of operators
The application of unitary operators requires that the dimension of the operator matches
the dimension of the variable or variables it is applied to. Formally, this is written as
E ⊢ (x1 , . . . , xn ) : q tq (q) = dim(U ) ∧ U ∈ U (n)
.
∀k : E ⊢ xk : qk ∧ ((x1 , . . . , xn )*=U ) : void
(5.99)
The statement additionally ensures that the typing of the qbits involved is not influenced
by the operator application. Note that we do not explicitely specify a formal condition for the
unitarity of an operator U given in terms of a function of its components. The membership
in U (n) is sufficient for our purposes. The distinctness of the destination variables for the
transformation is already ensured by the tupling requirements given above, so it does not
need to be checked explicitely.
5.3.8.3
If conditionals and while loops
The condition for this construction must have type bit, whereas both possible paths must
be well-typed:
E ⊢ c : bit
E ⊢P ∧E ⊢Q
.
(5.100)
E ⊢ (if c then P else Q) : void
A similar condition holds for the while loop:
E ⊢ c : bit
E⊢P
.
E ⊢ (while(c) do P ) : void
5.3.8.4
(5.101)
Measurements
The classical data type used to store the result of a measurement must have the same number
of bits as there are qbits in the quantum variable. This is represented by the condition
E ⊢ a : σ1 , c(σ1 ) = 1 ∧ E ⊢ b : σ2 , q(σ2 ) = 1 tq (σ2 ) = tc (σ1 )
E ⊢ (a := measure b) : void
5.3.8.5
(5.102)
Communication
Sending qbits When qbits are sent, the type system has to make sure that no qbit is
sent twice because this would result in the same effects as if operators could be applied to
multiple copies of the same qbit; using a tuple to combine the sent qbits automatically solves
this problem:
E ⊢ (x1 , . . . , xn ) : σ, q(σ) = 1
.
(5.103)
E ⊢ (send q1 , . . . , qn ):void
Note that the type system is not concerned with the actual receiver of the qbits; this
information is only required for the denotation of the expression, but not to ensure welltypedness.
80
5.4. Avoidance of runtime errors
Receiving qbits When qbits are received, the type system must make sure that the destination variables are not yet defined in the receiver’s context, i.e., they must not be well-typed
expressions. Afterwards, the variables used in the receive statement are well-defined in the
typing context and have the data type required by the statement. This can be formally
written as
∀i : ¬(E ⊢ xi ), q(σi ) = 1
.
(5.104)
E ⊢ (receive x1 : σ1 , . . . , xn : σn ):void ∧ ∀i : E ⊢ xi : σi
As in the case of sending, it is not interesting for the type system from which communication
partner the qbits originate.
5.4
Avoidance of runtime errors
QPL is a functional language with a static type system which guarantees the absence of
runtime errors (note that functionality is subject to a precise definition of the term; it is
certain that classical languages need to have additional properties – most important higherorder functions – to be called fully functional. But this is not really relevant from a physicist’s
point of view, as we have discussed before). It is very desirable that runtime errors can be
avoided as far as in principle possible, from a physicists point of view, it does not matter
how this is achieved. This desire was brought forward into cQPL and manifests itself in two
points: Cloning is (as in QPL) prevented already at the syntactical level, and communication
does not allow different processes to access qbits concurrently.
5.4.1
Unique ownership of qbits
Observation 5.4.1. No part of the quantum heap is accessible to two or more parties at
the same time during parallel execution of arbitrary cQPL programs.
Rationale: When a module is considered stand-alone, it is obvious that all qbits present
in the system are owned by one party. Uniqueness of the ownership is guaranteed by the
quantum part of the probabilistic environment which ensures (as described in Section 5.3.1)
that it is impossible for two or more names to refer to overlapping sets of qbits.
Parallel composition of systems is performed by always considering pairs of send/receive
statements; while sending removes the sent qbit from the typing context of the originating
system, receiving adds it to the typing context of the destination. Since both commands are
always considered in pairs and denoted atomically,25 a quantum variable may not be in two
typing contexts at the same time. Access to quantum variables is only possible for a user
when the variable is present in his typing context, this ensures (together with the fact that
quantum heaps of communicating systems cannot overlap by virtue of Definition 5.3.3) that
it is impossible for two or more modules to access identical qbits at a time.
It is possible to prove this statement formally based on the observations in Section 5.3.8.
Since this is on the one hand a general problem of semantics theory and on the other
hand burdened with many technical difficulties, we omit a precise proof here, but refer to
25 This
means that nothing can happen in between sending and receiving the quantum bit.
81
Summary
We have given the denotational semantics of all language components of cQPL excluding some standard cases that are readily available in the literature. Together
with the definition of the type system (by intentional omission of all technical
details), this completes the effort of assigning a precise meaning to quantum programs written in cQPL.
Chapter 5. Formal denotational semantics
[WF94] where the exact details can be found. [GN05] is one source where the proofs of
the aforementioned reference have been adapted to a quantum system which fulfils exactly
the same properties as ours (basically, not too much except notational details needs to be
changed).
X
5.4.2
Prevention of cloning and unphysical situations
One of the fundamental consequences of quantum mechanics is that it is impossible to
define a unitary operator that can duplicate arbitrary quantum states with perfect fidelity; a
straightforward calculation shown in nearly every text on quantum mechanics (e.g., [NC00,
Pre99]) proves this. Obviously, quantum programming languages must make sure that
cloning is forbidden because otherwise, processes contrary to the laws of physics could be
simulated. Since most other approaches to quantum programming (e.g., [Öme98, BSC01,
Kni96, SP00]) require the possibility to address quantum bits via references or pointers, they
cannot ensure at compile-time that two distinct variables do not share the same quantum bit;
they must provide appropriate checks at runtime which ensure this condition. Aside from
efficiency considerations, this is unsatisfying because especially for long-running programs,
termination with an error which was caused by a programming mistake is undesirable.
The static typing of QPL allows together with some syntactical checks to ensure that
once a program was approved to be correct by the static syntactical and semantical analysis
of the compiler, no runtime errors caused by unphysical cloning of quantum states can
happen. A similar statement can be observed for cQPL:
Observation 5.4.2. cQPL programs which do not use communication primitives can be
guaranteed to execute without runtime errors if the syntactic and semantic analysis deems
them correct.
Rationale: No quantum bit can be referred to by multiple identifiers in cQPL, as was
shown in the previous section. Thus, the distinctness of the quantum components of a list
of identifiers (which is used to specify the list of qbits an operator works on or given as
parameter to a procedure) can be guaranteed by ensuring that the same identifier does not
appear multiple times in the list. Therefore, the same line of reasoning for the impossibility
of cloning or generating unphysical situations as in [Sel04b, Section 4.8] applies.
X
Remark 5.4.1. Note that non-termination is something different than a runtime error.
Remark 5.4.2. Note that static typing can also prevent the possibility for some runtime
errors which originate from the classical parts of the language; this is well-known in programming language theory (cf., e.g., Refs. [RP02, WM95, App04] for details) so that we will
not dwell into this any further here.
5.4.3
Unavoidable non-termination conditions
Albeit cQPL tries to prevent runtime errors as good as possible, the introduction of communication opens the possibility of writing programs that cannot be checked at compile time
if they will terminate at runtime although nothing would hinder the separate modules to
terminate. Nevertheless, by restricting the code to a certain subset of cQPL,26 it is still
possible to produce programs which will execute guaranteed without termination problems
and without runtime errors. Note that non-termination is not considered as a runtime error.
26 Which can, in principle, solve all problems that might arise in quantum programming, but is not a very
practical.
82
5.4. Avoidance of runtime errors
If a program of the form while (1) do skip is provided, then executing the skip command forever (and thus doing nothing forever) is exactly the intention of the program and
therefore the correct behaviour which should be reflected by the denotation.
In Section 5.3.6.4, we have already considered an example of a non-terminating program.
The culprit here was the different number of sent versus received qbits, but since the number
of sent and received qbits is fixed at compile time on both sides, this error can obviously
be detected by the semantic analysis; the program can be rejected. Unfortunately, this
possibility is not always the case because the exact number of how many qbits will be
sent and how many will be received can not be decided in general. Consider the following
example:
module A {
new qword nq;
nq *= H(8);
new word n := measure nq;
while (n >= 0) {
new qbit q;
send q to B;
n := n-1;
}
};
module B {
receive q1:qbit, q2;qbit, q3:qbit;
};
Since n in module A may contain (with equal probability) any value in [0, 28 − 1], the
number of sent qbits cannot be determined with certainty, but is governed by the probability
distribution of n. It may be the case that the program terminates (namely, if exactly three
qbits are sent by A), but it may also be the case that less or more than three qbits are sent.
This results in either a blocking process A which cannot find a receiver for the qbits it wants
to transmit, or in a blocking process B which is not satisfied with a proper number of qbits
and blocks to wait for the missing ones.
Fortunately, there are only three commands in cQPL that allow to execute a sequence
of communication commands for which it is not possible to determine at runtime how many
there will be, so we can make the following observation:
Observation 5.4.3. cQPL programs using communication can be guaranteed to execute
without runtime errors if the syntactic and semantic analysis deems them correct and the
following possibilities of the language are not used:
❏ While-loops with a termination condition that contains a probabilistic variable.
❏ If-conditionals that are based on a probabilistic variable.
❏ Recursive procedures whose recursion depth cannot be determined at compile time.
Rationale: All send and receive statements which are given as a sequence of commands
(which may include the use of blocks) can be counted at compile time; their order is obviously
also known. If the if-statement is used with a condition that can be computed at compile
time, one path can be eliminated. Thus, the statement is nothing else than a regular
contribution to the list of statements. While-loops with a compile-time computable number
83
Chapter 5. Formal denotational semantics
of iterations can be replaced by inlining the loop body the appropriate number of times, so
they also become only a regular contribution to a sequence of commands. If the recursion
depth of a procedure can be calculated, it be converted to an iteration where the number
of steps and thus the number of communication commands are known. Therefore, it is also
just a regular contribution to a sequence of commands.
X
Remark 5.4.3. Note that the checks required to determine the number of loop iterations
etc. at compile-time are based on well-understood analysis techniques in computer science;
nevertheless, we did not actually implement these checks in the cQPL compiler because it is
nothing else than a routine task with little benefit and no gain of any valuable insight, but
just a technical problem.
84
I ain’t no physicist, but I know what matters.
Popeye the sailor
6
6.1
Prospects
Outlook
The field of quantum programming languages is – as everything connected with quantum
information – still a young one, and many things that are standard in classical programming
languages still need to be adapted for these. Some ideas which were tried to be realised
during the work on this thesis, but did not reach fruition are:
❏ Integration of higher-order functions. Everything we tried ended up in requiring closures for an implementation, but this is (to our knowledge) impossible to achieve
because of the no-cloning theorem. Having them would be quite desirable for many
applications.
❏ The ability to describe the complete loss of quantum bits caused by imperfect channels
or eavesdroppers. This is obviously hard to integrate into a programming language,1
but should be possible by heading for a protocol specification variant of cQPL.
❏ Consideration of more general eavesdropping models where the strategy needs not be
fixed, but can be one of multiple independent alternatives. The work provided in
[dH02] would possibly provide a suitable basis with demonic choices.
❏ Most texts about quantum programming languages extensively use categories to describe the underlying structures. We found that this does not really add any substantial
points, but merely more notation and nomenclature, so we did not follow this style
although some effort was made in the beginning to become familiar with the field.
Nevertheless, the material provided here could serve as starting point for the following
possible extensions:
❏ Quantum instead of classical control, i.e., allowing conditions to be based on quantum
and not classical logic. It would be possible to simulate this with QCL, but the benefit
is questionable because no known quantum algorithm makes use of such a feature.
❏ The method presented here could provide a basis to formulate quantum process algebras as, e.g., presented in [GN05, AM05].
❏ An extension from discrete to continuous systems would allow the simulation of general quantum systems and could thus be useful for a much wider range of quantum
information applications.
1 Just think about the situation that would arise if variables in classical programming languages could
randomly disappear. . .
85
Chapter 6. Prospects
❏ Faulty hardware models could be integrated at the simulation layer, but this would
presumably be very challenging at the semantic level.
Initially, it was planned to also investigate the possibility of integrating the semantic
framework into a theorem prover which could possibly facilitate automated analysis techniques for quantum protocols. Some preliminary experiments were performed by describing
the BB84 protocol in a classical protocol simulator, but this has only shown that the gap
between the requirements for such an automatisation and what is currently available is still
very wide for all approaches to quantum programming.
6.2
Latest developments
After this thesis was finished, another QPL compiler written by D. Williams was presented in
a joint work by Nagarajan, Papanikolaou and Williams [NPW05]. Since both efforts work on
closely related fields, it seems apt to sketch similarities and differences of them (to distinguish
it from our implementation, we call William’s compiler sqrQPL2 in the following):
❏ Both compilers use the quantum computer model defined by Knill [Kni96] as basic
architecture.
❏ sqrQPL provides an own quantum simulator which is called sequential quantum random
access machine. Code generated for this architecture resembles machine language quite
closely.
❏ sqrQPL provides support for a smaller subset of QPL than cQPL.
❏ sqrQPL has the ability to automatically decompose arbitrary unitary matrices into a
set of standard gates. As a result (and in addition to theoretical elegance) of this, the
simulator needs only provide support for very few different elementary gates.
❏ The semantics of sqrQPL is fully covered by the one given for QPL, while cQPL needs
to provide additional semantics for the added features.
❏ cQPL already includes support for communication and concurrency, whereas work to
bring these abilities to sqrQPL will be started in the future according to [NPW05].
It would be interesting (and should be possible without too much effort) to provide an
SQRAM-backend for cQPL; since Ref. [NPW05] states that support for communication and
concurrency is (at least in preliminary form) already present in their simulator, no major
obstacle does seem to exist to hinder such an endeavour. In summary (and, obviously, seen
from the author’s subjective point of view) sqrQPL is a straight implementation of QPL
where the ability to decompose complicated gates into a set of simpler ones is the essential feature. The focus of cQPL is mainly on the semantics of (quantum) communication;
although the cQPL compiler seems (at the time of writing) to provide a bigger and more
versatile language core than sqrQPL, it is more or less a by-product of the actual work.
2 Because their compiler targets a virtual machine which is termed sequential quantum random access
machine.
86
. . . und Lasse sagte, die Sprache der Jungen
sei sowieso die einzig wahre.
Astrid Lindgren, Wir Kinder aus Bullerbü
A
List of symbols
The following presents a list of symbols used in this work. Note that the meanings given
here are not necessarily the only ones with which they were used.
A . . . . . . . . . . . Observable algebra
#number . . . Unique node id
A . . . . . . . . . . . Finite ordered set
#string . . . . . Position(s) occupied by
quantum variable variable
on the quantum heap
JK . . . . . . . . . . . Separate syntax and
semantics
|| . . . . . . . . . . . . Parallel composition
$ . . . . . . . . . . . . Superoperator on B(H)
F
. . . . . . . . . . . Least upper bound
A#b . . . . . . . . The bth element of the
ordered set A
B(H) . . . . . . . . Set of all bounded
operators on Hilbert
space H
COMM . . . . Valuation function for
parallel execution
C(X) . . . . . . . . Complex-valued functions
X →C
c(σ) . . . . . . . . . Check if a given data
type is purely classical
card(X) . . . . . Cardinality of X
⊥ . . . . . . . . . . . Least element of a partial
order
∼
= . . . . . . . . . . . Equivalence, reflexive and
transitive
⊑ . . . . . . . . . . . Binary partial order
(K∅ , T∅ , E∅ ) Initial (K, T, E) tuple
Γ . . . . . . . . . . . List of Kraus sets
Λ . . . . . . . . . . . Completely positive map
Dn . . . . . . . . . . Set of all density
operators of dimension n
S(k) . . . . . . . . Set of all decompositions
of k ∈ N
D(A, B) . . . . . Determine if A and B
operate on disjoint qbits
DO . . . . . . . . . Valuation function for
dyadic operators
E . . . . . . . . . . . Environment
χ(v) . . . . . . . . Type associated with a
variable v
ω . . . . . . . . . . . Increasing chain of
natural numbers
ϕ . . . . . . . . . . . A permutation
π : B . . . . . . . . Probability distribution
obtained by applying a
projective measurement
defined by the basis B
̺ . . . . . . . . . . . . A density operator
E ⊢ x : T . . . . x has type T is valid in
environment E
E ⊢ F . . . . . . . F is well-typed in
environment E
E(A) . . . . . . . . Effects of A
Σ . . . . . . . . . . . State in form of a
(K, T, E) tuple
σ . . . . . . . . . . . Signature for types
EQN . . . . . . . Valuation function for
arithmetic expression
EX P . . . . . . . . Valuation function for
expressions
A . . . . . . . . . . . Set of all possible Kraus
aggregations
87
Appendix A. List of symbols
F . . . . . . . . . . . Set of all fixed points of a
permutation
F2 . . . . . . . . . . Binary group/ring/field
fix . . . . . . . . . . Fixed point
in . . . . . . . . . . . Injection
I(M ) . . . . . . . Set of all intervals in M
K . . . . . . . . . . . Kraus aggregation
q(σ) . . . . . . . . . Check if a given data
type is purely quantum
qtype . . . . . . Arbitrary quantum data
type
R . . . . . . . . . . . Number of receive
statements in a Kraus
aggregation
K . . . . . . . . . . . Set of all unparametrised
Kraus agregations
Lc . . . . . . . . . . Set of labels for
communication partners
M . . . . . . . . . . Finite set
RS . . . . . . . . . Valuation function for a
receive/send pair
MO . . . . . . . . Valuation function for
monadic operators
OP . . . . . . . . . Valuation function for
operators
P(M ) . . . . . . . Powerset of M
S(A) . . . . . . . . States of A
S . . . . . . . . . . . Number of send
statements in a Kraus
aggregation
smash . . . . . . . Smash product (with a
single bottom element)
SR . . . . . . . . . Valuation function for a
send/receive pair
nc . . . . . . . . . . . Classical data type with
n bits
nq . . . . . . . . . . Quantum data type with
n qbits
pos(x, L) . . . . Position of x in the list L
Sym(M ) . . . . Symmetric group over M
T (σ) . . . . . . . . Set of data types
equivalent to σ
T . . . . . . . . . . . Typing context
PROG . . . . . . Valuation function for
programs
PROJ . . . . . Generate Kraus set with
projection operators for a
quantum type
Q . . . . . . . . . . . Size of the quantum heap
(global constant!)
Q . . . . . . . . . . . Set of all quantum
variables in a typing
context
q(nτ ) . . . . . . . Distinguish between
classical and quantum
components of a data
type
tq (σ) . . . . . . . . Number of quantum bits
contained in a data type
σ
tc (σ) . . . . . . . . Number of bits contained
in a data type σ
U (n) . . . . . . . . Unitary group of degree n
VAL . . . . . . . . Arbitrary valuation
function
X . . . . . . . . . . . Finite set
x : t . . . . . . . . . Variable x with type t
Y D . . . . . . . . . Fixed point combinator
on cpo D
88
Die Bedeutung eines Wortes ist das, was die
Erklärung der Bedeutung erklärt.
Ludwig Wittgenstein, Philosophische
Grammatik
B
Glossary
Some terms used in this work are not too commonplace in physics, so we collected the most
important definitions to remind the reader of their meaning if it cannot be immediately
recollected. Some of the definitions were inspired by [Wik05].
Abstract syntax Grammar used to specify the possible shapes of the parse
tree.
EBNF Extended Backus-Naur Form
Environment Structure which provides a
mapping between identifiers of variables and the values associated with
them.
Backus-Naur Form Metasyntax with a
standardised set of symbols and notations which is used to express contextfree grammars.
FIFO Queue with first-in, first-out behaviour, i.e., the output of the queue
is in the same order as the input.
Compiler-Compiler A program used to
generate a parser which can perform
syntax analysis on programs that follow a given grammar.
Functional languages do not work on explicit states, but use transformations
that map input to output parameters
(→ referential transparency). Assignment to variables is not possible since
they represent immutable bindings for
values. In our notation, functionality
is exploited as far as it is necessary for
the ability to guarantee freedom from
runtime errors. Classical examples of
functional languages include Lisp, ML,
and Haskell.
Compile time refers to all actions which
are performed by the compiler before
the program is executed, e.g., syntactical and semantical analysis, scoping
rule enforcement, type analysis, optimisation, code generation etc.
Concrete syntax Syntax in which textual
representations of programs must be
specified.
Identifier Name of a variable in a program.
Context free grammar Formal grammar
in which every production rule needs
to be of the form V → w where V is a
non-terminal and w a list of terminal
and non-terminal symbols.
Imperative languages work on a global
state that is modified during runtime.
The most widespread languages (C,
C++, Pascal etc.) follow this approach. It is normally impossible to
decide if a program written in an imperative language will terminate or
produce errors without executing it.
Data type A data type is a name or label
for a set of values and some operations
which can be performed on that set of
values.
89
Appendix B. Glossary
Lexer A lexer is the part of a compiler
which takes the source code of a program (in textual form) and disseminates it into a stream of tokens which
is fed to the parser.
Scope Rules used to determine what, if
any, entity a given occurrence of an
identifier in a program refers to.
Semantic analysis is the part of a compiler that adds semantic information
to the parse tree (for example, the required space in memory for variables)
and performs sanity checks which may
detect errors in the code before it is
executed.
Lexicographic order Two strings x, y ∈
Σ∗ can be ordered such that x > y
if x#i − y#i > 0 for the first i ∈ N for
which x#i 6= y#i.
LALR(1) Certain class of context-free
grammars that needs to be specified
subject to some constraints on its
form, but can be handled by Yacc-style
parsers.
Static typing means that once a type has
been assigned to an object, it cannot
be changed any more.
Strong typing means that not only values,
but also identifiers are typed.
Mutex Mutual exclusion. A technique realised with the aid of special variables
which ensures that only one component of a parallel program can be in
the region protected by the mutex at
a time.
Terminal symbol A symbol of a grammar
that represents a constant.
Token Tokens are the smallest elementary
parts of a program form the parser’s
point of view. While int is a threeletter word respectively a string of
characters in the source code, the
parser regards it as a single entity
which describes the data type of integers.
Non-terminal symbol A symbol that is
composed of terminal symbols and
possibly other non-terminal symbols.
Parser The parser is the part of a compiler
which analyses the grammatical structure of a program (which is fed to him
in the form of tokens produced by the
lexer ). This process is also known as
syntactical analysis.
Type Also called data type. It is a label for
a set of values together with some operations that can be performed on the
set.
Type system Set of rules that determines
which type a given object has, how
types can be combined etc.
Parse tree Representation of a program
which is generated by the parser. Since
tree-based data structures are used to
represent the information, the abstract
syntax of the language (which is easier
to analyse) can be utilised.
Type checking is the pass of a compiler
which ensures that all operations of
a program are applied to variables of
proper type; it can, e.g., ensure that
string concatenation is not tried to be
performed on integers.
Runtime refers to the time when a program is executed and the compiler has
no more influence on what happens.
Alternatively, it may denote a library
with helper functions supplied by the
compiler which are necessary for the
generated code to work (this may be
also referred to as runtime library).
Yacc Yet another compiler compiler. One
of the early approaches to automated
parser generation.
Most modern
parser generators follow the concept of
this program.
90
C
Der Satz ist der sprachliche Ausdruck dafür,
dass sich die Verbindung mehrerer Vorstellungen in der Seele des Sprechenden vollzogen
hat, und das Mittel dazu, die nämliche
Verbindung der nämlichen Vorstellungen in
der Seele des Hörenden zu erzeugen.
H. Paul, Prinzipien der Sprachgeschichte
Formal syntax
The formal syntax for cQPL is defined by the following rules which are used to generate the
parser. Words in typewriter face denote tokens recognised by the lexer, whereas slanted
text is used for productions. identifiers are given by a letter followed by an arbitrary number
of letters, digits and underscores. The empty production is denoted by ǫ.
program: stmt list EOF
| module list EOF
stmt list: statement;
| stmt list statement;
module list: module def ;
| module list module def ;
module def : module identifier { stmt list }
proc decl: proc identifier:context -> context block in statement
| proc identifier:context block in statement
context: identifier:var type more context | ǫ
nonempty context: identifier:var type more context
more context: , identifier:var type more context | ǫ
block: { stmt list }
var type: bit | qbit | qint | int | float
send stmt: send args to identifier
receive stmt: receive nonempty context from identifier
allocate stmt: new var type identifier := arith expr
arith expr: int value
| float value
| true
| false
| identifier
| (arith expr)
| arith expr + arith expr
| arith expr - arith expr
| arith expr * arith expr
| arith expr / arith expr
| arith expr < arith expr
| arith expr > arith expr
| arith expr <= arith expr
| arith expr >= arith expr
| arith expr == arith expr
| arith expr != arith expr
91
Appendix C. Formal syntax
| arith expr & arith expr
| arith expr | arith expr
| - arith expr
| ! arith expr
proc call: call identifier (args)
| (var list) := call identifier (args)
args: identifier more args | ǫ
more args: , identifier more args | ǫ
if stmt: if arith expr then statement
| if arith expr then statement else statement
measure stmt: measure identifier then statement else statement
assign stmt: identifier := arith expr
assign measure stmt: identifier := measure identifier
while stmt: while arith expr do statement
gate stmt: var list *= gate
gate: H | CNot | Not | Phase float value
| FT (int value)
| [[ number list ]]
number list: sign float value
| sign int value
| sign float value PLUS sign imaginary value
| sign int value PLUS sign imaginary value
| sign imaginary value
| number list, sign float value
| number list, sign int value
| number list, sign imaginary value
| number list, sign float value + sign imaginary value
| number list, sign int value + sign imaginary value
sign: - | + | ǫ
var list: identifier | var list, identifier
skip stmt: skip
print stmt: print "string"
| print arith expr
| dump var list
statement: proc call
| proc decl
| while stmt
| allocate stmt
| if stmt
| print stmt
| assign stmt
| assign measure stmt
| measure stmt
| skip stmt
| block
| gate stmt
| send stmt
| receive stmt
92
Bibliography
[AB02]
Alexander Asteroth and Christel Baier. Theoretische Informatik. Pearson
Studium, 2002.
[AC04a]
Samson Abramsky and Bob Coecke. A categorial semantics of quantum protocols. arXiv:quant-ph/0402130, pages 1–20, 2004.
[AC04b]
Samson Abramsky and Bob Coecke. A categorical semantics of quantum protocols. In LICS ’04: Proceedings of the 19th Annual IEEE Symposium on Logic
in Computer Science (LICS’04), pages 415–425, Washington, DC, USA, 2004.
IEEE Computer Society.
[AG81]
N.I. Achieser and I.M. Glasmann. Theorie der linearen Operatoren im HilbertRaum. Verlag Harri Deutsch, 1981.
[AG04]
T. Altenkirch and J. Grattage. A functional quantum programming language.
arXiv:quant-ph/0409065, 2004.
[AG05]
Thorsten Altenkirch and Jonathan Grattage. QML: Quantum data and control. Submitted for publication, Febuary 2005.
[Aha98]
Dorit Aharonov. Quantum computation. arXiv:quant-ph/9812037, 1998.
[AJ94]
Samson Abramsky and Achim Jung. Handbook for Logic in Computer Science,
volume 3, chapter Domain Theory. Clarendon Press, Oxford, 1994.
[AM05]
P. Adão and P. Mateus. A process algebra for reasoning about quantum security. In Electronic Notes in Theoretical Computer Science. Springer, 2005.
Preliminary version to be presented at 3rd International Workshop on Quantum Programming Languages.
[App04]
Andrew W. Appel. Modern Compiler Implementation in ML. Cambridge
University Press, New York, NY, USA, 2004.
[ASU86]
Alfred V. Aho, Ravi Sethi, and Jeffrey D. Ullman. Compilers: principles,
techniques, and tools. Addison-Wesley Longman Publishing Co., Inc., Boston,
MA, USA, 1986.
[Aul00]
Gennaro Auletta. Foundations and Interpretation of Quantum Mechanics.
World Scientific, 2000.
[BSC01]
S. Betelli, L. Serafini, and T. Calarcoet. Toward an architecture for quantum
programming. arXiv:cs.pl/0103009, 2001.
93
Bibliography
[BW]
Björn Butscher and Hendrik Weimer. Simulation eines Quantencomputers.
Universität Stuttgart.
[Car97]
Luca Cardelli. Type Systems, chapter 103. Handbook of computer science and
engineering. CRC Press, 1997.
[Cle99]
Richard Cleve. An introduction to quantum complexity theory. arXiv:quantph/9906111, 1999.
[Deu85]
David Deutsch. Quantum theory, the Church-Turing principle and the universal quantum computer. Proceedings of the Royal Society of London Ser. A,
A400:97–117, 1985.
[dH02]
J. den Hartog. Probabilistic extension of semantical models. PhD thesis, Vrije
Universiteit Amsterdam, 2002.
[DJ92]
David Deutsch and Richard Jozsa. Rapid solutions of problems by quantum
computation. Proceedings of the Royal Society of London, pages 553– 558,
1992.
[DS63]
Nelson Dunford and Jacob T. Schwartz. Linear Operators. Interscience Publishers, 1963.
[GA05]
Jonathan Grattage and Thorsten Altenkirch. A compiler for a functional quantum programming language. submitted for publication, January 2005.
[GBJL02]
Dick Grune, Henri E. Bal, Ceriel J. H. Jacobs, and Koen Langendoen. Modern
Compiler Design. John Wiley, 2002.
[GN04]
S.J. Gay and R. Nagarajan. Communicating quantum processes. Proceedings
of the conference for quantum programming languages, pages 91–107, 2004.
[GN05]
Simon J. Gay and Rajagopal Nagarajan. Communicating quantum processes.
In POPL ’05: Proceedings of the 32nd ACM SIGPLAN-SIGACT sysposium
on Principles of programming languages, pages 145–157, New York, NY, USA,
2005. ACM Press.
[Gro96]
Lov K. Grover. A fast quantum mechanical algorithm for database search.
Proceedings of the Twenty-Eighth Annual ACM Symposium on Theory of Computing, pages 212–219, 1996.
[Gru99]
Jozef Gruska. Quantum Computing. McGraw–Hill International, 1999.
[GS90]
C.A. Gunther and D.S. Scott. Semantic Domains, chapter 12, pages 635–674.
Elsevier Science Publishers, 1990.
[Hol82]
Alexander S. Holevo. Probabilistic and Statistical Aspects of Quantum Theory,
volume 1 of North-Holland series in statistics and probability. North-Holland,
Amsterdam, 1982. First publ. in Russian in 1980.
[JL04]
Philippe Jorrand and Marie Lalire. Toward a quantum process algebra. In
CF’04: Proceedings of the first conference on computing frontiers, pages 111–
119, New York, NY, USA, 2004. ACM Press.
[Key02]
Michael Keyl. Fundamentals of quantum information theory. arXiv:quantph/0202122, 369(5):431–548, 2002.
94
Bibliography
[KN00]
E.H. Knill and M.A. Nielsen. Encyclopedia of Mathematics, Supplement III,
chapter Theory of quantum computation. Kluwer Academic Publishers, 2000.
[Kni96]
E. Knill. Conventions for quantum pseudocode. Technical Report LAUR-962724, 1996.
[Knu98]
Donald E. Knuth. Art of Computer Programming, Volume 3: Sorting and
Searching (2nd Edition). Addison-Wesley Professional, April 1998.
[Kra83]
Karl Kraus. States, Effects and Operations. Fundamental Notions of Quantum
Theory. Academic Press, Berlin, 1983.
[Lou03]
Kenneth C. Louden. Programming Languages: Principles and Practice. Thomson, Pacific Grove, second edition, 2003.
[Löw34]
K. Löwner. Über monotone Matrixfunktionen. Mathematische Zeitschrift,
38:177–216, 1934.
[MB01]
S-C. Mu and R. S. Bird. Quantum functional programming. 2nd Asian Workshop on Programming Languages and Systems, 2001.
[Mer98]
Eugen Merzbacher. Quantum Mechanics. Wiley, John & Sons, 3 edition, 1998.
[Mos90]
Peter D. Mosses. Denotational semantics, chapter 11, pages 577–629. Elsevier
scientific publishers, 1990.
[NC00]
Michael L. Nielsen and Isaac L. Chuang. Quantum computation and quantum
information. Cambridge University Press, New York, NY, USA, 2000.
[NPW05]
Rajagopal Nagarajan, Nikolaos Papanikolaou, and David Williams. Simulating
and compiling code for the sequential quantum random access machine. In
Selinger [Sel05].
[Öme98]
Bernhard Ömer. A procedural formalism for quantum computing. Master’s
thesis, TU Vienna, 1998.
[Öme00]
Bernhard Ömer. Quantum Programming in QCL. Master’s thesis, TU Vienna,
2000.
[Öme03]
Bernhard Ömer. Structured quantum programming. PhD thesis, TU Vienna,
2003.
[Pre99]
John Preskill. Lecture notes for the course quantum computation (physics
229). www.theory.caltech.edu/people/preskill/ph229, 1999.
[RAMK+ 04] Helge Rosé, Torsten Asselmeyer-Maluga, Matthias Kolbe, Falk Niehoerster, and Andreas Schramm. The fraunhofer quantum computing portal.
arXiv:quant-ph/0406089, 2004.
[Rey98]
John C. Reynolds. Theories of programming languages. Cambridge University
Press, 1998.
[RP02]
Peter Rechenberg and Gustav Pomberger. Informatik-Handbuch. Hanser Fachbuch, 2002.
[Sak94]
Jun John Sakurai. Modern Quantum Mechanics. Addison-Wesley, 1994.
95
Bibliography
[Sch24]
Moses Schönfinkel. Über die Bausteine mathematischer Logik. Math. Ann. 92,
pages 305–316, 1924.
[Sch01]
Uwe Schöning. Theoretische Informatik - kurzgefasst. Spektrum, Akad. Verl.,
4 edition, 2001.
[Sch04]
Andreas Schroeder. Quantenflussdiagramme und die Quantenprogrammiersprache QPL. Seminar der Lehr- und Forschungseinheit für theoretische Informatik, LMU München, 2004.
[Sel04a]
Peter Selinger. A brief survey of quantum programming languages. In Lecture
Notes in Computer Science 2998. Springer, 2004.
[Sel04b]
Peter Selinger. Towards a quantum programming language. Mathematical.
Structures in Comp. Sci., 14(4):527–586, 2004.
[Sel05]
Peter Selinger. Proceedings of the 3rd international workshop on quantum
programming languages. In Peter Selinger, editor, Proceedings of the 3rd International Workshop on Quantum Programming Languages, Electronic Notes
in Theoretical Computer Science. Elsevier Science, 2005.
[SF96]
Robert Sedgewick and Philippe Flajolet. An Introduction to the Analysis of
Algorithms. Addison-Wesley, 1996.
[Sho94]
Peter W. Shor. Algorithms for quantum computation: Discrete logarithms
and factoring. IEEE Symposium on Foundations of Computer Science, pages
124–134, 1994.
[SP00]
J.W. Sanders and P.Zuliani. Quantum programming. Lecture notes in computer science, 1837, 2000.
[Sto87]
Joseph E. Stoy. Denotational semantics. MIT Press, 4 edition, 1987.
[Str00]
Christopher Strachey. Fundamental concepts in programming languages.
Higher Order Symbol. Comput., 13(1-2):11–49, 2000.
[vT04]
Andre van Tonder. A lambda calculus for quantum computation. SIAM Journal on Computing, 33:1109–1135, 2004.
[Wei00]
Joachim Weidmann. Lineare Operatoren in Hilberträumen, volume 1. B.G.
Teubner, 2000.
[WF94]
Andrew K. Wright and Matthias Felleisen. A syntactic approach to type soundness. Information and Computation, 1994.
[Wik05]
Wikipedia community.
2005.
[Win93]
Glynn Winskel. The formal semantics of programming languages: an introduction. MIT Press, Cambridge, MA, USA, 1993.
[WM95]
Reinhard Wilhelm and Dieter Maurer. Compiler Design. Addison Wesley
Longman Publishing Co., Inc., Redwood City, CA, USA, 1995.
Wikipedia online dictionary, www.wikipedia.net,
96
Thanks
❏ To PD Dr. Norbert Lütkenhaus for taking the peril of a journey into the strange and
unaccustomed, his support by discussions and suggestions and for giving me complete
freedom in deciding what to work on.
❏ To Prof. Dr. Dr. Volker Strehl for many pointers into the right direction and for
sacrificing time for a student from another faculty.
❏ To Tobias Moroder, Hans Loehr, Martin Trini, Johannes Rigas, Volkher Scholz and
Markus Diefenthaler for proofreading and many valuable corrections and suggestions.
❏ To the guys in my office (Tobi Moroder, Johannes Rigas and Dr. Matthias Jakob)
for providing a pleasant environment to work in, many inspiring level eights and our
shared pleasure of working under illumination provided by the moon.
❏ To an anonymous reviewer for encouraging comments.
❏ To dict.leo.org for countless suggestions on english vocabulary; quick answers to
many questions were given by www.wikipedia.net.
❏ To Dr. Peter Selinger for detailled explanations regarding his work.
❏ To the red and the green forrest fairy because they are way too mythical to not be
thanked; without any doubt, the same holds for HM Queen Elizabeth II.
❏ To all members of the QIT group (Tobi, Johannes, Philippe, Matthias 1, Matthias
2, Joe, Geir Ove, Marcos, Ivan) for their help, valuable discussions and the pleasant
working environment, not to forget the shared fun among some of us in chasing little
bouncing objects on diverse courts.
❏ To my parents and my family for their overall and ubiquitous love and support in any
aspect of life.
97
| 2cs.AI
|
Explaining How a Deep Neural Network Trained with
End-to-End Learning Steers a Car
arXiv:1704.07911v1 [cs.CV] 25 Apr 2017
Mariusz Bojarski
NVIDIA Corporation
Holmdel, NJ 07733
Krzysztof Choromanski
Google Research
New York, NY 10011
Philip Yeres
NVIDIA Corporation
Holmdel, NJ 07733
Bernhard Firner
NVIDIA Corporation
Holmdel, NJ 07733
Anna Choromanaska
New York University
New York, NY 10012
Lawrence Jackel
NVIDIA Corporation
Holmdel, NJ 07733
Urs Muller
NVIDIA Corporation
Holmdel, NJ 07733
Abstract
As part of a complete software stack for autonomous driving, NVIDIA has created
a neural-network-based system, known as PilotNet, which outputs steering angles
given images of the road ahead. PilotNet is trained using road images paired with
the steering angles generated by a human driving a data-collection car. It derives
the necessary domain knowledge by observing human drivers. This eliminates the
need for human engineers to anticipate what is important in an image and foresee
all the necessary rules for safe driving. Road tests demonstrated that PilotNet
can successfully perform lane keeping in a wide variety of driving conditions,
regardless of whether lane markings are present or not.
The goal of the work described here is to explain what PilotNet learns and how
it makes its decisions. To this end we developed a method for determining which
elements in the road image most influence PilotNet’s steering decision. Results
show that PilotNet indeed learns to recognize relevant objects on the road.
In addition to learning the obvious features such as lane markings, edges of roads,
and other cars, PilotNet learns more subtle features that would be hard to anticipate and program by engineers, for example, bushes lining the edge of the road
and atypical vehicle classes.
1
Introduction
A previous report [1] described an end-to-end learning system for self-driving cars in which a convolutional neural network (CNN) [2] was trained to output steering angles given input images of the
road ahead. This system is now called PilotNet. The training data were images from a front-facing
camera in a data collection car coupled with the time-synchronized steering angle recorded from
a human driver. The motivation for PilotNet was to eliminate the need for hand-coding rules and
instead create a system that learns by observing. Initial results were encouraging, although major
improvements are required before such a system can drive without the need for human intervention.
To gain insight into how the learned system decides what to do, and thus both enable further system
improvements and create trust that the system is paying attention to the essential cues for safe steering, we developed a simple method for highlighting those parts of an image that are most salient
1
Output: vehicle control
Fully-connected layer
Fully-connected layer
Fully-connected layer
10 neurons
50 neurons
100 neurons
1164 neurons
Flatten
Convolutional
feature map
64@1x18
3x3 kernel
Convolutional
feature map
64@3x20
3x3 kernel
5x5 kernel
Convolutional
feature map
48@5x22
Convolutional
feature map
36@14x47
5x5 kernel
Convolutional
feature map
24@31x98
5x5 kernel
Normalized
input planes
3@66x200
Normalization
Input planes
3@66x200
Figure 1: PilotNet architecture.
in determining steering angles. We call these salient image sections the salient objects. A detailed
report describing our saliency detecting method can be found in [3]
Several methods for finding saliency have been described by other authors. Among them are sensitivity based approaches [4, 5, 6], deconvolution based ones [7, 8], or more complex ones like
layer-wise relevance propagation (LRP) [9]. We believe the simplicity of our method, its fast execution on our test car’s NVIDIA DRIVETM PX 2 AI car computer, along with its nearly pixel level
resolution, makes it especially advantageous for our task.
1.1
Training the PilotNet Self-Driving System
PilotNet training data contains single images sampled from video from a front-facing camera in
the car, paired with the corresponding steering command (1/r), where r is the turning radius of the
vehicle. The training data is augmented with additional image/steering-command pairs that simulate
the vehicle in different off-center and off-orientationpoistions. For the augmented images, the target
steering command is appropriately adjusted to one that will steer the vehicle back to the center of
the lane.
Once the network is trained, it can be used to provide the steering command given a new image.
2
PilotNet Network Architecture
The PilotNet architecture is shown in Figure 1. The network consists of 9 layers, including a normalization layer, 5 convolutional layers and 3 fully connected layers. The input image is split into YUV
2
Pointwise
multiplication
Averaging
al
Sc
eu
p
Final visualization mask
1@66x200
Figure 2: Block diagram of the visualization method that identifies the salient objects.
planes and passed to the network. The first layer of the network performs image normalization. The
normalizer is hard-coded and is not adjusted in the learning process.
The convolutional layers were designed to perform feature extraction and were chosen empirically
through a series of experiments that varied layer configurations. Strided convolutions were used in
the first three convolutional layers with a 2×2 stride and a 5×5 kernel and a non-strided convolution
with a 3×3 kernel size in the last two convolutional layers.
The five convolutional layers are followed with three fully connected layers leading to an output
control value that is the inverse turning radius. The fully connected layers are designed to function
as a controller for steering, but note that by training the system end-to-end, there is no hard boundary
between which parts of the network function primarily as feature extractors and which serve as the
controller.
3
Finding the Salient Objects
The central idea in discerning the salient objects is finding parts of the image that correspond to
locations where the feature maps, described above, have the greatest activations.
The activations of the higher-level maps become masks for the activations of lower levels using the
following algorithm:
1. In each layer, the activations of the feature maps are averaged.
2. The top most averaged map is scaled up to the size of the map of the layer below. The
up-scaling is done using deconvolution. The parameters (filter size and stride) used for the
3
deconvolution are the same as in the convolutional layer used to generate the map. The
weights for deconvolution are set to 1.0 and biases are set to 0.0.
3. The up-scaled averaged map from an upper level is then multiplied with the averaged map
from the layer below (both are now the same size). The result is an intermediate mask.
4. The intermediate mask is scaled up to the size of the maps of layer below in the same way
as described Step 2.
5. The up-scaled intermediate map is again multiplied with the averaged map from the layer
below (both are now the same size). Thus a new intermediate mask is obtained.
6. Steps 4 and 5 above are repeated until the input is reached. The last mask which is of the
size of the input image is normalized to the range from 0.0 to 1.0 and becomes the final
visualization mask.
This visualization mask shows which regions of the input image contribute most to the output of
the network. These regions identify the salient objects. The algorithm block diagram is shown in
Figure 2.
The process of creating the visualization mask is illustrated in Figure 3. The visualization mask
is overlaid on the input image to highlight the pixels in the original camera image to illustrate the
salient objects.
Results for various input images are shown in Figure 4. Notice in the top image the base of cars as
well as lines (dashed and solid) indicating lanes are highlighted, while a nearly horizontal line from
a crosswalk is ignored. In the middle image there are no lanes painted on the road, but the parked
cars, which indicate the edge of the road, are highlighted. In the lower image the grass at the edge
of the road is highlighted. Without any coding, these detections show how PilotNet mirrors the way
human drivers would use these visual cues.
Figure 5 show a view inside our test car. At the top of the image we see the actual view through the
windshield. A PilotNet monitor is at the bottom center displaying diagnostics.
Figure 6 is a blowup of the PilotNet monitor. The top image is captured by the front-facing camera.
The green rectangle outlines the section of the camera image that is fed to the neural network.
The bottom image displays the salient regions. Note that PilotNet identifies the partially occluded
construction vehicle on the right side of the road as a salient object. To the best of our knowledge,
such a vehicle, particularly in the pose we see here, was never part of the PilotNet training data.
4
Analysis
While the salient objects found by our method clearly appear to be ones that should influence steering, we conducted a series of experiments to validate that these objects actually do control the
steering. To perform these tests, we segmented the input image that is presented to PilotNet into two
classes.
Class 1 is meant to include all the regions that have a significant effect on the steering angle output
by PilotNet. These regions include all the pixels that correspond to locations where the visualization
mask is above a threshold. These regions are then dilated by 30 pixels to counteract the increasing
span of the higher-level feature map layers with respect to the input image. The exact amount of
dilation was determined empirically. The second class includes all pixels in the original image
minus the pixels in Class 1. If the objects found by our method indeed dominate control of the
output steering angle, we would expect the following: if we create an image in which we uniformly
translate only the pixels in Class 1 while maintaining the position of the pixels in Class 2 and use this
new image as input to PilotNet, we would expect a significant change in the steering angle output.
However, if we instead translate the pixels in Class 2 while keeping those in Class 1 fixed and feed
this image into PilotNet, then we would expect minimal change in PilotNet’s output.
Figure 7 illustrates the process described above. The top image shows a scene captured by our
data collection car. The next image shows highlighted salient regions that were identified using the
method of Section 3. The next image shows the salient regions dilated. The bottom image shows a
test image in which the dilated salient objects are shifted.
4
Figure 3: Left: Averaged feature maps for each level of the network. Right: Intermediate visualization mask for each level of the network.
Figure 4: Examples of salient objects for various image inputs.
5
Figure 5: View inside our test car
The above predictions are indeed born out by our experiments. Figure 8 shows plots of PilotNet
steering output as a function of pixel shift in the input image. The blue line shows the results when
we shift the pixels that include the salient objects (Class 1). The red line shows the results when we
shift the pixels not included in the salient objects. The yellow line shows the result when we shift
all the pixels in the input image.
Shifting the salient objects results in a linear change in steering angle that is nearly as large as
that which occurs when we shift the entire image. Shifting just the background pixels has a much
smaller effect on the steering angle. We are thus confident that our method does indeed find the most
important regions in the image for determining steering.
5
Conclusions
We describe a method for finding the regions in input images by which PilotNet makes its steering
decisions, i. e., the salient objects. We further provide evidence that the salient objects identified by
this method are correct. The results substantially contribute to our understanding of what PilotNet
learns.
Examination of the salient objects shows that PilotNet learns features that “make sense” to a human,
while ignoring structures in the camera images that are not relevant to driving. This capability is
derived from data without the need of hand-crafted rules. In fact, PilotNet learns to recognize subtle
6
Figure 6: The PilotNet monitor from Figure 5 above
Figure 7: Images used in experiments to show the effect of image-shifts on steer angle.
7
Figure 8: Plots of PilotNet steering output as a function of pixel shift in the input image.
features which would be hard to anticipate and program by human engineers, such as bushes lining
the edge of the road and atypical vehicle classes.
References
[1] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal,
Lawrence D. Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao, and Karol Zieba.
End to end learning for self-driving cars, April 25 2016. URL: http://arxiv.org/abs/1604.
07316, arXiv:arXiv:1604.07316.
[2] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten zip code recognition. Neural Computation, 1(4):541–551, Winter 1989.
URL: http://yann.lecun.org/exdb/publis/pdf/lecun-89e.pdf.
[3] Mariusz Bojarski, Anna Choromanska, Krzysztof Choromanski, Bernhard Firner, Larry Jackel, Urs Muller,
and Karol Zieba. VisualBackProp: visualizing CNNs for autonomous driving, November 16 2016. URL:
https://arxiv.org/abs/1611.05418, arXiv:arXiv:1611.05418.
[4] D. Baehrens, T. Schroeter, S. Harmeling, M.i Kawanabe, K. Hansen, and K.-R. Müller. How to explain
individual classification decisions. J. Mach. Learn. Res., 11:1803–1831, 2010.
[5] K. Simonyan, A. Vedaldi, and A. Zisserman. Deep inside convolutional networks: Visualising image
classification models and saliency maps. In Workshop Proc. ICLR, 2014.
[6] P. M. Rasmussen, T. Schmah, K. H. Madsen, T. E. Lund, S. C. Strother, and L. K. Hansen. Visualization of
nonlinear classification models in neuroimaging - signed sensitivity maps. BIOSIGNALS, pages 254–263,
2012.
[7] M. D. Zeiler, G. W. Taylor, and R. Fergus. Adaptive deconvolutional networks for mid and high level
feature learning. In ICCV, 2011.
[8] M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014.
[9] S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. Müller, and W Samek. On pixel-wise explanations
for non-linear classifier decisions by layer-wise relevance propagation. PLOS ONE, 10(7):e0130140, 2015.
URL: http://dx.doi.org/10.1371/journal.pone.0130140.
8
| 9cs.NE
|
COUNTING SUBGRAPHS IN FFTP GRAPHS WITH SYMMETRY
arXiv:1802.05213v1 [math.GR] 14 Feb 2018
YAGO ANTOLÍN
Abstract. Following ideas that go back to Cannon, we show the rationality of various generating functions of growth sequences counting embeddings of convex subgraphs in locallyfinite, vertex-transitive graphs with the (relative) falsification by fellow traveler property
(fftp). In particular, we recover results of Cannon, of Epstein, Iano-Fletcher and Zwick, and
of Calegari and Fujiwara. One of our applications concerns Schreier coset graphs of hyperbolic groups relative to quasi-convex subgroups, we show that these graphs have rational
growth, the falsification by fellow traveler property, and the existence of a lower bound for
the growth rate independent the quasi-convex subgroup (provided it that has infinite index)
and the generating set.
1. Introduction
In the celebrated paper [8], Cannon showed that the growth of groups acting properly
and cocompactly on Hn is rational, i.e, the generating function of the sequence counting the
number of elements in the ball of radius n is a rational function. His ideas were successively
used by Gromov [17] showing that (word) hyperbolic groups have rational growth, by Epstein,
Iano-Fletcher and Zwick [13] who improved results indicated by Saito [29] showing that
the generating function counting the number of embeddings of finite subgraphs in geodesic
automatic Cayley graphs is rational, and recently by Calegari and Fujiwara [3] obtaining the
previous result for vertex-transitive hyperbolic graphs, not necessarily Cayley graphs.
Neumann and Shapiro [23] observed that one can use Cannon’s arguments under an hypothesis weaker than hyperbolicity, called the falsification by fellow traveler property (fftp).
A graph has this property if there is a constant M , such that every non-geodesic path M fellow travels with a shorter one. For Cayley graphs, this property has been widely studied
and there are several examples beyond hyperbolicity. The following families of groups have
Cayley graphs with fftp for at least one generating set: virtually abelian groups and geometrically finite hyperbolic groups [23], Coxeter groups and groups acting simply transitively on
the chambers of locally finite buildings [25], groups acting cellularly on locally finite CAT(0)
cube complexes where the action is simply transitive on the vertices [24], Garside groups [18]
and Artin groups of large type [19]. The property of having a generating set with fftp is
preserved under relative hyperbolicity [1] and arguing similarly as in [22] it follows that it is
also preserved under graph products.
Date: February 15, 2018.
2010 Mathematics Subject Classification. 20F65, 20F10, 20F67,05A15, 05C25.
Key words and phrases. hyperbolic groups, Cayley graphs, Schreier graphs, growth series, languages of
geodesics, falsification by fellow traveler property, automatic groups, uniform exponential growth.
The author is supported by the Juan de la Cierva grant IJCI-2014-22425, and acknowledges partial support from the Spanish Government through grant number MTM2014-54896 and through the Severo Ochoa
Programme for Centres of Excellence in R&D (SEV-20150554).
1
2
YAGO ANTOLÍN
This paper aims to push the ideas of Cannon to the limit; we will present a common
generalization of the previous results for counting “convex” subgraphs (not necessarily finite)
on locally finite, vertex-transitive graphs1 (not necessarily Cayley graphs) with the (relative)
falsification by fellow traveler property.
Let Γ be a locally finite, connected, vertex-transitive graph. Let dΓ denote the combinatorial graph metric on V Γ, the vertices of Γ. Let Z be some graph and let eZ (n) be the
number of different embeddings of Z as a complete subgraph in Bv0 (n), the ball of radius n
of Γ with center at v0 , i.e.
eZ (n) =
]{f : Z → Γ | f injective graph morphism , f (Z) ⊆ Bv0 (n)}
.
|Aut(Z)|
For example, if Z = • is a vertex, e• (n) counts the number of vertices in the ball of radius
n, and Cannon’s result asserts
that when Γ is the Cayley graph of a group acting properly
P
and cocompactly in Hn , n≥0 e• (n)tn ∈ Z[[t]] is a rational function. i.e. an element of Q(t).
In the case Z is infinite, eZ (n) is equal to zero for all n, and to deal with this, we will
count embeddings of Z with non-trivial intersection with the ball of radius n. However, we
need to restrict to some family of embeddings.
Definition 1.1. Let Γ be a graph and G 6 Aut(Γ) acting vertex transitively. A subgraph
Z of Γ has G-proper embeddings if for all v ∈ V Γ, the set {gZ | g ∈ G, v ∈ gZ} is finite.
Given G 6 Aut(Γ) and a subgraph Z of Γ with G-proper embeddings, we will consider
the function i(G,Z) (n) counting the number of gZ ⊆ Γ, g ∈ G, such that gZ intersects nontrivially the ball of radius n centered at v0 . In a similar way, we denote by e(G,Z) (n) the
number of elements of the G-orbit of Z embedded in the ball of radius n. In order to count
infinite subgraphs, we will need some mild convexity properties, which we define below.
Definition 1.2. Let Γ be a graph and Z a subgraph. For v ∈ V Γ, the closest point
projections of v onto Z is denoted by πZ (v) := {z ∈ V Z | d(v, z) = d(v, V Z)}.
The set Z has M -fellow projections if for every u, v ∈ V Γ with d(u, v) = 1 and every
zv ∈ πZ (v) there is zu ∈ πZ (u) such that d(zv , zu ) ≤ M .
The set Z has M -bounded projections if for every u, v ∈ V Γ with d(u, v) = 1 the diameter
of πZ (u) ∪ πZ (v) is bounded above by M .
Note M -bounded projections implies M -fellow projections, and that any finite subgraph Z
have both properties. We will see other examples with these properties in Section 7, including
quasi-convex subgraphs of hyperbolic graphs, parabolic subgroups of relatively hyperbolic
groups, parabolic subgroups of raags and standard parabolic subgroups of Coxeter groups.
Fix a vertex v0 . Let g ∈ G. A gZ-geodesic path is a path from v0 to gZ that realizes the
distance d(v0 , gZ). We will say that a path p is a (G, Z)-geodesic path if it is a gZ-geodesic
path for some g ∈ G. Let g(G,Z) (n) be the number of (G, Z)-geodesics of length ≤ n.
We will introduce the relative falsification by fellow traveler property in Section 3. The
following, is a particular case of our main result that will be stated in Theorem 3.5.
1Let Γ be a graph. We will say that G 6 Aut(Γ) is vertex-transitive (or that G acts vertex transitively) if
there is a single orbit of vertices of Γ by the action of G. If Aut(Γ) is vertex-transitive, we will say that Γ is
vertex-transitive.
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
3
Theorem A. Let Γ be a locally finite graph with the falsification by fellow traveler property.
Let G 6 Aut(Γ) be a vertex-transitive subgroup and Z a graph with G-proper embeddings. If
Z has fellow projections, then
P
(1) the (G, Z)-geodesic growth function, n≥0 g(G,Z) (n)tn ∈ Z[[t]], and
P
(2) the (G, Z)-embeddings growth function, n≥0 e(G,Z) (n)tn ∈ Z[[t]],
are rational functions, and moreover, if Z has bounded projections, then
P
(3) the (G, Z)-intersections growth function, n≥0 i(G,Z) (n)tn ∈ Z[[t]],
is rational.
Particular cases previously known of Theorem A are: (1) for Cayley graphs with the
falsification by fellow traveler property and Z = • in [23]; and (2) for Γ hyperbolic and finite
graph Z in [3]; using a stronger version of this theorem (Theorem 3.5) one recovers (2) for
geodesically automatic Cayley graphs and Z finite, which was proved in [13].
Observe that Theorem A can be used to understand the geometry of Schreier coset graphs.
Indeed, suppose that Γ is a Cayley graph of a group G with respect to some generating set
X. If Z is (the subgraph spanned by) a subgroup, then it has G-proper embeddings and
i(G,Z) (n) counts how many left cosets of Z intersect non-trivially a ball of radius n in the
Cayley graph. If the generating set X is symmetric, then the previous number is also equal
to the number of right cosets meeting the ball of radius n, which in turn, is the number of
elements in the ball of radius n of the Schreier coset graph Γ(G, Z, X). We will prove
Theorem B. Let G be a group and X a finite symmetric generating set of G. Suppose that
the Cayley graph Γ(G, X) has the falsification by fellow traveler property. Let H 6 G be a
subgroup of G such that H (as a subgraph) has fellow projections in Γ(G, X). Then:
(1) the Schreier coset graph Γ(G, H, X) has the falsification by fellow traveler property
relative to the family of all paths starting at the coset H;
(2) the set of words
Geo(H\G, X) = {w ∈ X ∗ | `(w) ≤ `(u) ∀u ∈ X ∗ , u ∈G Hw}
is a regular language;
(3) if moreover, H
Phas bounded projections, then for the Schreier coset graph Γ(G, H, X),
the function n≥0 e• (n)tn is rational.
The Schreier graphs for hyperbolic groups relative to quasi-convex subgroups were studied
in [20] by I. Kapovich. There it is proved that if H is quasi-convex of infinite index in a nonelementary hyperbolic group G, then Γ(G, H, X) is non-amenable. A number of consequences
are derived from this fact, and in particular bounds on the co-growth rate are obtained.
Here, we explore the growth rate of these Schreier graphs. Using ideas of [13] we obtain
Theorem C. Let G be hyperbolic, H a quasi-convex subgroup and X a finite symmetric
generating set of G. Let Γ = Γ(G, H, X) be the Schreier coset graph and let e• (n) = |V BΓ (n)|
be the number of vertices in the ball of radius n centered at H in Γ.
There exists a polynomial QX (t) depending on (G, X) and a constant λ > 1 depending
only on G such that the following hold:
P
P
(1) n≥0 e• (n)tn is a rational function with denominator QX (t) (i.e. QX (t) n≥0 e• (n)tn
is a polynomial).
4
YAGO ANTOLÍN
(2) If G is non-elementary and H is of infinite index, then
p
lim sup n e• (n) ≥ λ.
2. Notations, conventions and definitions
In this paper, a digraph Γ is a 4-tuple (V Γ, EΓ, (·)− , (·)+ ), where V Γ is a non-empty set
whose elements are called vertices, EΓ is a set, whose elements are called oriented edges,
(·)− , (·)+ : EΓ → V Γ are functions that are called incidence functions. A graph is a digraph
−1
with an involution (·)−1 : EΓ → EΓ satisfying that for all e ∈ EΓ, e−1
− = e+ and e+ = e− .
A combinatorial path p in a digraph Γ is a sequence v0 , e1 , v1 , e2 , . . . , en , vn where vi ∈ V Γ,
ei ∈ EΓ and (ei )− = vi−1 and (ei )+ = vi . The length of the path p is denoted by `(p) and
is the number of edges in the sequence. We extend the adjacent functions to paths, setting
p− = v0 and p+ = vn . We write p(i) to denote vi and if p is a path of length n, we use the
convention that p(k) = p+ for all k ≥ n. If Γ is a graph, we also extend (·)−1 to paths, being
−1
p−1 = vn , e−1
n , . . . , e1 , v0 .
The combinatorial distance between u, v in V Γ, denoted dΓ (u, v), is the infimum of the
length of the combinatorial paths p with p− = u and p+ = v. A path realizing the combinatorial distance between two vertices is a geodesic.
In a graph, one should think e and e−1 as a single un-oriented edge and view Γ as metric
space that topologically a 1-dimensional CW-complex in which 0-cells are the elements of
V Γ and 1-cells are elements of EΓ/(·)−1 , where [e] = {e, e−1 } is attached to e− and e+ . The
metric arises by making each 1-cell isometric to the interval [0, 1] of the real line. With this
setting the combinatorial distance agrees with the induced metric on vertices.
Let λ ≥ 1 and c ≥ 0. A path p is a (λ, c)-quasi-geodesic if for any subpath q of p we have
`(q) ≤ λd(q− , q+ ) + c.
Let p, q be paths in Γ and M ≥ 0. We say that p, q asynchronously M -fellow travel if there
exist non-decreasing functions φ : N → N and ψ : N → N such that d(p(t), q(φ(t))) ≤ M
and d(p(ψ(t)), q(t)) ≤ M for all t ∈ N. We say that p, q synchronously M -fellow travel if
d(p(t), q(t)) ≤ M for all t ∈ N.
Let G be a group, H a subgroup and X a symmetric generating set for G. The Schreier
coset graph for G relative to H with respect to X is a graph Γ(G, H, X) that has vertex set
H\G and edges (H\G) × X where (Hg, x) is an oriented edge from Hg to Hgx with label
x. Since X is symmetric, there is an edge (Hgx, x−1 ) that we define to be (Hg, x)−1 . When
H = {1}, Γ(G, {1}, X) is just the Cayley graph of G with respect to X and we write Γ(G, X).
We use the metric on the Cayley graph to define the length of g ∈ G as |g|X := d(1, g).
2.1. Regular languages. A finite state automaton is a 5-tuple (S, A, s0 , X, τ ), where S is
a set whose elements are called states, A is a subset of S of whose states are called accepting
states, a distinguished element s0 ∈ S called initial state, a finite set X called the input
alphabet and a function τ : S × X → S called the transition function.
Let X be a set. We denote by X ∗ the free monoid generated by X. Given a non-negative
integer M , we denote by X ≤M ⊂ X ∗ the set of words in X of length at most M .
We extend τ to a function τ : S × X ∗ → S recursively, by setting τ (s, wx) = τ (τ (s, w), x)
where w ∈ X ∗ , x ∈ X and s ∈ S.
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
5
A language L over X is a subset of X ∗ . A language is regular if there is a finite state
automaton (S, A, s0 , X, τ ) such that
L = {w ∈ X ∗ | τ (s0 , w) ∈ A}.
To a finite state automaton, we can associate a rooted X-labeled digraph ∆, whose vertices
are the set of states, the root is the initial state, and edges are of the form e = (s, x, τ (s, x))
where e− = s and e+ = τ (s, x) and the label is x. In particular, a word w ∈ X ∗ codifies a
path in ∆ starting at the root.
Suppose that ρ : E∆ → R is a function. We can extend ρ to a function on paths in ∆
starting at s0 (and hence to words in X) by multiplying the values of each edge Q
of the path,
i.e if p = v0 , e1 , v1 , e2 , . . . , vn is path in ∆ starting at the root, we define ρ(p) = ρ(ei ).
3. Labelling paths in vertex-transitive graphs
Let Γ be a locally finite, connected, vertex-transitive graph and let v0 be a vertex of Γ.
Definition 3.1. A presentation for the paths starting at v0 is a pair (X, θ) where X is a set
in bijection with (v0 )−1
− = {e ∈ EΓ | e− = v0 }, via x 7→ e(x) and e 7→ xe , together with a
map θ : X → Aut(Γ) satisfying that for each xe ∈ X, θ(xe )(v0 ) = e+ .
Let (X, θ) be a presentation for paths in Γ starting at v0 . Let F = F (X) be the free
group freely generated by X. Then θ extend to an homomorphism θ : F → hθ(X)i. Let T
be the Cayley graph of F with respect to X. The maps f ∈ F = V T 7→ θ(f )(v0 ) ∈ Γ, and
(f, x, f x) ∈ ET 7→ θ(f )(e(x)) define a surjective graph homomorphism T → Γ that is a local
isomorphism and hence a covering map. We call this covering map θ again. Thus, any path
starting at v0 lifts to a unique path in T starting at 1, and conversely, a paths in T starting
to 1 map to paths in Γ starting at v0 .
Since T is a Cayley graph of a free group generated by X, paths starting at 1 in T
correspond to words over X. Conversely, given a path in Γ starting at v0 , the lift in T gives
a word in X corresponding to the lifted path.
Remark 3.2. We can not use the labels of edges of T and the map θ to define a labeling
on edges of Γ since in general this labeling would not be hθ(X)i-invariant. If it were, then
Γ would be the Cayley graph of hθ(X)i with respect to θ(X), however there are vertextransitive locally-finite graphs (for example Diestel Leader graphs [14]) that are not even
quasi-isometric to Cayley graphs.
In summary, given a presentation for paths at v0 , (X, θ), there is a bijection between words over X and paths starting at v0 given on words as follows: for a word
w ≡ x1 . . . xn in X we assign the combinatorial path pw consisting on the sequence
v0 , e(x1 ), θ(x1 )(v0 ), θ(x1 )(e(x2 )), θ(x1 x2 )(v0 ), . . . , θ(x1 . . . xn )(v0 ).
For sake of notation, we will usually drop the θ and the function. Thus a word w in X can
be seen as an element of F or an element of Aut(Γ) under θ and we will write wv0 instead
of θ(w)(v0 ). The meaning will be clear from the context.
Let us denote the words giving geodesic paths as
Geo(Γ, X, θ) = {w ∈ X ∗ | pw is a geodesic path}.
6
YAGO ANTOLÍN
In general, if P if a collection of paths in Γ starting at v0 , we denote by L(P) = L(P, X, θ)
the language defined by the paths in P, i.e.
L(P) = {w ∈ X ∗ | pw ∈ P}.
3.1. Relative falsification by fellow traveler property.
Definition 3.3. Let Γ be a graph, v0 a vertex, P a family of paths in Γ and M ≥ 0.
The family P is v0 -spanning if for all v ∈ V Γ there is a geodesic path p ∈ P from v0 to v.
We denote by P + to the union of P and the one-edge continuations of paths in P, i.e.
= P ∪ {p, e, e+ | p ∈ P, e ∈ EΓ, e− = p+ }
P+
We say that Γ has the falsification by M -fellow traveler property relative to P (Γ is M -fftp
relative to P) if every path p ∈ P + asynchronously M -fellow travel with a path q ∈ P with
the same end points, and moreover if p is not geodesic, then `(q) < `(p).
If Γ is M -fftp relative to the collection of all paths, we just say that Γ is M -fftp.
The original definition of the falsification by fellow traveler property requires that the
paths asynchronously M -fellow travel, however, it was observed by Elder [9] that the original
definition is equivalent to the synchronous definition (up to increasing constants).
The main example of graphs with the falsification by fellow traveler property with respect
to a spanning family of paths are Cayley graphs of groups with a generating set admitting a
geodesically automatic structure.
Example 3.4. Let G be a group and X a finite generating set for X. Recall that L ⊆ X ∗ is a
geodesic automatic structure if L is a regular language that surjects onto G with the natural
evaluation map, each word w ∈ L labels a geodesic path in the Cayley graph Γ(G, X) and
there is a constant M , such that for every w ∈ L and every x ∈ X, there exists u ∈ L such
that wx =G u and the paths labeled by wx and u synchronously M -fellow travel (see [12] for
details on geodesic automatic structures). Clearly P = {pw | w ∈ L} is a 1G -spanning family
of paths and Γ(G, X) has the falsification by fellow traveler property with respect to P.
We now can state the full version of Theorem A.
Theorem 3.5. Let Γ be a locally finite graph and G 6 Aut(Γ) be a vertex-transitive subgroup.
Let (X, θ) be a presentation for paths in Γ starting at a vertex v0 . Let P be a v0 -spanning
collection of paths such that Γ has the falsification by fellow traveler property relative to P
and L(P) is a regular language. Let and Z a subgraph of Γ with proper G-embeddings and
fellow projections. Then
(1) {w ∈ X ∗ | pw is a (G, Z)-geodesic} ∩ L(P)
P is regular,
(2) the (G, Z)-embeddings growth function n≥0 e(G,Z) (n)tn is a rational function,
P
(3) if Z has bounded projections, then the (G, Z)-intersections function n≥0 i(G,Z) (n)tn
is rational.
4. The fftp Automaton and the n-type directed graph
Throughout this section Γ is a locally-finite vertex-transitive graph. We start by fixing a
vertex v0 of V Γ and (X, θ) a presentation for paths in Γ starting at v0 .
Let M ≥ 0. For each x ∈ X, a ∈ X ≤M with av0 ∈ Bxv0 (M ), and b ∈ X ≤M we set
dx (a, b) := min{`(p) | p− = av0 , p+ = xbv0 , p ⊆ Bxv0 (M )} ⊆ [0, M ].
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
7
Definition 4.1. Let M ≥ 0. The fftp-automaton for (Γ, X, θ) with parameter M ≥ 0 and
accepting states A, is defined as follows:
(1) the input alphabet is X,
(2) the set of states S consists on a fail state {%} and states of the form φ : X ≤M →
[−M, M ],
(3) a distinguished initial state φ0 given by φ0 (w) = d(v0 , wv0 ) for w ∈ X ≤M ,
(4) a transition map τ : S × X → S defined as τ (%, x) = %, τ (φ, x) = % if φ(x) 6= 1.
Otherwise τ (φ, x) = ψ where for b ∈ X ≤M
ψ(b) = min{φ(a) + dx (a, b) − 1 | a ∈ X ≤M , av0 ∈ Bxv0 (M )},
(5) a subset A of S of accepting states.
By the fftp digraph we will refer to the digraph associated to the fftp automaton.
For the sake of simplifying the notation, given w ∈ X ∗ , we will use φw to denote the state
τ (φ0 , w) ∈ S.
We now clarify the meaning of the states of the fftp-automaton.
Proposition 4.2. Let (X, τ, φ0 , S, A) be the fftp automaton for (Γ, X, θ) with parameter M .
Then φw 6= % if and only if pw does not asynchronously M -fellow travel with a shorter path,
moreover if φw 6= %, then for u ∈ X ≤M
(1)
φw (u) = min{`(q) − `(pw ) | q− = v0 , q+ = wuv0 and pw , q asyn M − fellow travel}.
Note that in particular, if uv0 = u0 v0 then φw (u) = φw (u0 ).
Proof. We prove the proposition by induction on `(w), being the base of induction the case
`(w) = 0, where pw is the path of length zero starting and ending at v0 and φw is equal to
φ0 . Observe that the proposition holds in this case.
So assume that the proposition holds for w, and we have to show it for wx, where x ∈ X.
Consider first the case φwx = %. We have two subcases. If φw = %, then by induction
hypothesis, pw asynchronously M -fellow travels with a shorter path from v0 to wv0 and so
does pwx since pw is a subpath of pwx . If φw 6= %, then φw (x) 6= 1 and by (1) there is a path
q from v0 to wxv0 such that pw and q asynchronously M -fellow travel and `(q) − `(pw ) < 1.
It follows that `(q) < `(pw ) + 1 = `(pwx ) and pwx and q asynchronously M -fellow travel.
Consider now the case that φwx 6= %. Then φw (x) = 1. Suppose that there is a path q from
v0 to wxv0 that M -fellow travels with pwx and is shorter than pwx . Then `(q) − `(pw ) ≤ 0
and, q also asynchronously M -fellow travels with pw . Since 1 = φw (x) ≤ `(q) − `(pw ) we get
a contradiction.
We now show that (1) holds for φwx . Let b ∈ X ≤M . By definition φwx (b) = min{φw (a) +
− 1 | a ∈ X ≤M , av0 ∈ Bxv0 (M )}.
dx (a, b)
Let a ∈ X ≤M realizing the minimum. Then, by induction hypothesis, there exists a
path q1 from v0 to wav0 such that `(qa ) − `(pw ) = φw (a), (qa )+ ∈ Bwxv0 (M ) and q1 and
pwx asynchronously M -fellow travel. By definition of dx , there is a path q2 from wav0 to
wxbv0 of length dx (a, b) that asynchronously M -fellow travel with wxv0 . Thus q = q1 q2
asynchronously M -fellow travels with pwx and `(q) − `(pwx ) = `(q1 ) − `(pw ) + `(q2 ) − 1. We
have shown that φwx (b) is greater or equal than the right-hand side of (1).
8
YAGO ANTOLÍN
Let q be a path of minimal length from v0 to wxbv0 that asynchronously M -fellow travel
with pwx . Write q as q1 q2 where q1 be the longest initial subpath of q that asynchronously
M -fellow travel with pw , and q2 might be an empty subpath. Note that q2 ⊆ Bwxv0 (M )
and that (q1 )+ ∈ Bwxv0 . By the minimality of q one has that `(q1 ) − `(pw ) = φw (a) where
a ∈ X ≤M is such that wav0 = (q1 )+ and `(q2 ) = dx (a, b). Thus we get that `(q) − `(pwx ) =
`(q1 ) + `(q2 ) − `(pw ) − 1 = φw (a) + dx (a, b) − 1 ≥ φwx (b). Thus φwx (b) is less or equal than
the right-hand side of (1). This completes the proof.
Remark 4.3. The transition function τ in Definition 3.3 does not completely agree with
the one of Neumann and Shapiro of [23, Proposition 4.1]. In their transition function the
minimum is taken over the a ∈ X ≤M such that d(av0 , xbv0 ) ≤ 1 (which is equivalent to
dx (a, b) ≤ 1), and later it is claimed without proof that Proposition 4.2 holds for that
automaton. Our definition of the transition function for the fftp automaton resembles more
to the one of [3, Lemma 3.7] where essentially it describes the transition between different
tournaments. We remark that it is not exactly equal because without some extra structure
(such as hyperbolicity) we do not know that dx (a, b) agrees with d(av0 , xbv0 ).
4.1. Fftp graphs and the fftp-automaton. We will now show the power of this automaton
when Γ is a fftp graph.
Hypothesis 4.4. Let Γ be a vertex transitive graph, v0 ∈ V Γ, and (X, θ) a presentation of
paths starting at v0 . Let P be a v0 -spanning family of paths and assume that Γ is M -fftp
relative to P. Let (X, τ, φ0 , S, A) be the fftp automaton for (Γ, X, θ) with parameter M 2 .
The importance of M 2 will be evident soon. First, we need the following fact, whose proof
consist on repeatedly using the definition.
Lemma 4.5. With Hypothesis 4.4. Let B ≥ 1. Let p be a concatenation of a geodesic path
p0 in P starting at v0 and a path of length at most B starting at p0− . Then there exists a
geodesic path q in P with the same endpoints as p and such that p and q asynchronously
B · M -fellow travel.
Proposition 4.6. With Hypothesis 4.4. Let w ∈ L(P). Then φw 6= % if and only if pw is
geodesic. Moreover, if φw is not the fail state, then for all u ∈ X ≤M ,
(2)
φw (u) = dΓ (v0 , wuv0 ) − dΓ (v0 , wv0 ).
Proof. Let w ∈ L(P). If pw is geodesic, then it can not asynchronously fellow travel with a
shorter path, and by Proposition 4.2, φw 6= %. If pw is not geodesic, by the definition of fftp
relative to P, there exists a shorter path that asynchronously M -fellow travels with pw and
thus φw = %.
Now assume that φw 6= % and let u ∈ X ≤M . The path pwu is a concatenation of a geodesic
pw and a path of length at most M , and hence by Lemma 4.5 it asynchronously M 2 -fellow
travels with a geodesic q, and (2) follows from (1).
It follows from Proposition 4.6 that the fftp-automaton accepts those w ∈ L(P) that are
geodesic. Hence the language Geo(Γ, X, θ) ∩ L(P) is regular (this will be a particular case of
Corollary 5.3). Therefore, the generating function of the number of geodesic paths in P is a
rational function.
We aim to use the fftp-automaton not just to count geodesic paths of length ≤ n but also
to count the number of vertices in Bv0 (n). It will be convenient to understand which states
accept geodesics ending in the same vertex.
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
9
Definition 4.7 (Cannon’s N -type). Let G 6 Aut(Γ). Two vertices v and u of Γ have the
same N -type mod G, if there is an automorphism α ∈ G of Γ, such that α(v) = u, and
d(v0 , v 0 ) − d(v0 , v) = d(v0 , α(v 0 )) − d(v0 , u)
for every v 0 ∈ V Γ with d(v, v 0 ) ≤ N .
Definition 4.8. Let G 6 Aut(Γ). Two functions φ, φ0 : Bv0 (m) → [−m, m] are M -equivalent
mod G if there is an automorphism α of Γ such that α ∈ G, and φ0 |Bv0 (M ) = (φ ◦ α)|Bv0 (M ) ,
i.e. they agree in the ball of radius M up to automorphism α ∈ G of Γ.
We will denote by ∼k both equivalent relations, where k is the appropriate constant.
From now on, we fix G = hθ(X)i. It follows from Proposition 4.2 that the states of the
2
fftp-automaton φw : X ≤M → [−M 2 , M 2 ] induce a function Bv0 (M 2 ) → [−M 2 , M 2 ], uv0 7→
φw (u). For the sake of simplifying the notation, we will also denote this function by φw .
Lemma 4.9. With Hypothesis 4.4. If pw and pw0 are geodesic paths from v0 to v, then φw
and φw0 are M -equivalent mod G and wv0 and w0 v0 have the same M -type mod G.
Proof. Seeing w and w0 as automorphism of Γ, we let α = w−1 w0 ∈ G = hθ(X)i 6 Aut(Γ)
and we have that α(v0 ) = v0 . By Proposition 4.6, for every u ∈ X ≤M , φw (α(u)) =
d(v0 , wα(u)v0 ) − d(v0 , wv0 ) = d(v0 , w0 uv0 ) − d(v0 , w0 v0 ) = φw0 (u).
4.2. Random geodesic combings.
Definition 4.10. A random geodesic combing of a graph Γ is a set probability measures
{µv | v ∈ V Γ} where each µv has support on the set of geodesic paths starting at v0 and
ending at v. A geodesic combing is a random geodesic combing where each probability
measure has support on a single geodesic path.
Since paths starting at v0 are codified by words in X, we can think that µv is a probability
measure defined on X ∗ , whose support is contained in the set {w ∈ Geo(Γ, X, θ) | wv0 = v}.
We will say that a random geodesic combing {µv | v ∈ V Γ} is Markov if there is a finite
rooted X-labeled digraph ∆ and a function ρ : E∆ → [0, 1], such that for every w ∈ X ∗ ,
ρ(w) = µwv0 (pw ), where ρ is extended to (labels of) paths in ∆ as in subsection 2.1.
Theorem 4.11. With Hypothesis 4.4. Let ∼M denote the equivalence relation on the vertices
[v]
of Γ consisting on having the same M -type mod G = hθ(X)i. Let e(G,•) (n) := ]{u ∼M v |
[v]
d(v0 , u) ≤ n} and e(G,•) (n) := ]{u ∼M v | d(v0 , u) = n}.
(i) Let ∆ be the fftp-digraph. There is a function ρ : E∆ → [0, 1] defining a Markov
random geodesic combing for the graph Γ.
(ii) There is a square non-negative matrix A, and for every v ∈ V Γ there are non-negative
[v]
vectors u and v such that e(G,•) = vT An v.
(iii) For every v ∈ V Γ, the series
X [v]
X [v]
e(G,•) (n)tn and
e(G,•) (n)tn
n≥0
n≥0
are rational functions.
Proof. (i). Let v0 , v ∈ V Γ. A v0 -parent of v is an edge e adjacent to v in a geodesic from v0
to v. If w and w0 are words accepted in equivalent states of the fftp-automaton, then wv0 and
w0 v0 have the same number of parents, which is the number of x ∈ X for which φw (x) = −1.
10
YAGO ANTOLÍN
We assign weights on the edges of the fftp digraph ∆ as follows: for an edge going from a
1
state φw to a state φwx 6= % we assign the weight ]parents
φwx ; for an edge going to the state
% we assign weight 0. We now extend this weights to paths in ∆ by multiplying the weights
of along a path. Denote ρ this weight function on paths. For convenience, the path of length
zero has weight 1.
We get that each path in ∆ starting at φ0 and not ending in % codifies a word labeling a
geodesic path in Γ starting at v0 , and it has an associated weight. So we can think that ρ is
a function on paths on Γ starting at v0 that assigns a weight with values in [0, 1] and that a
necessary and sufficient condition for ρ(p) = 0 is that p is a non-geodesic path in Γ.
P
We need to show that for each v ∈ V Γ, p∈Geo(v0 ,v) ρ(p) = 1. The proof is by induction
on d(v0 , v). The base of induction is d(v0 , v) = 1. Let x1 , . . . , xn ∈ X such that xi v0 = v.
−1
Then [φxi ] = [φxj ] for 1 ≤ i, j ≤ n and φxi (y) = −1 if and only if y ∈ {x−1
1 , . . . , xn }. Thus
the number of parents of v is n and that is the number of paths in ∆ that codify a geodesic
path from v0 to v and hence, each of this paths have weight n1 and the claim follow in this
case.
Now assume that the result has been proved for d(v0 , v 0 ) ≤ m with m ≥ 1 and assume
that d(v0 , v) = m + 1. Let v1 , . . . , vk be the vertices in Γ with d(vi , v) = 1 and vi in some
geodesic from v0 to v. Let ei,j , i = 1, . . . , k and j = 1, . . . , n(j) be edges from vi to v. Notice
that these are all the parents of v, and therefore they correspond to an edge êi,j in the fftp
1
automaton of weight ] parents
of v .
Then
X
ρ(p) =
n(j)
k X
X
i=1 j=1
p∈Geo(v0 ,v)
=
n(j)
k X
X
X
ρ(êi,j )
ρ(p)
p∈Geo(v0 ,vi )
ρ(êi,j ) = (] parents of v)
i=1 j=1
1
= 1.
] parents of v
(ii). We continue with the notation of (i).
Let M = (mi,j ) be the transition of matrix the fftp digraph i.e. mij is the number of
edges starting in the state i and ending in the state j. We let the index 0 correspond to the
state φ0 . Form a new matrix A where for j 6= %, aij is equal to mi,j divided by the number
of parents of the state corresponding to the index j. Using the theory of Markov Chains, it
follows that if a0j (n) is the (0, j)-term of An , we have that a0j (n) is sum of the weights of
the paths in ∆ of length n starting at φ0 and ending in the state j, which correspond to the
number of geodesic paths in Γ starting a v0 and accepted at the state j.
Let v ∈ V Γ and let T a set of those φw 6= ρ with wv0 ∼M v i.e. the states corresponding
to paths ending in vertices with the same M -type as v. Let u be the column vector with
all entries equal to zero except from the entry corresponding to φ0 which is equal to 1.
Similarly, let v beP
the column vector with vi = 1 if i ∈ T and vi = 0 otherwise. Then for
n ≥ 1, vT An u = j∈T a0j (n) represents the sum of the weights of the paths in Γ starting
at v0 and finishing at a vertex at distance n with the same M -type as v.
(iii). We continue with the notation of (ii). First notice that
X [v]
X [v]
e(G,•) (n)tn .
(3)
(1 − t)
e(G,•) (n)tn =
n≥0
n≥0
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
11
So to prove the rationality of the series, it will be enough to show the rationality of the
right-hand side one
Thus
X
n≥0
[v]
e(G,•) (n)tn =
X
vT An utn = vT (Id − At)−1 u =
n≥0
vT Adj(Id − At)u
det(Id − At)
which is a quotient of two polynomials in Q[t].
4.3. Cannon’s proof: Cone-types and N -types. The fftp digraph used in Theorem
4.11 might be unnecessarily big. In fact, Cannon [8] shows that under hyperbolicity, for N
sufficiently big, the N -types determine the cone types and hence there are finitely cone types.
The rationality of the growth series follows from a recursion involving the cone types, which
is similar to the argument in the proof of Theorem 4.11. Neumann and Shapiro [23] realized
that one just needs fftp to show that for N sufficiently big, the N -type determine the cone
type. In this subsection, for the sake of completeness, we will see that there are covering
digraph maps from fftp digraph to the N -type digraph, and from the N -type digraph to the
cone-type digraph.
Lemma 4.12. With Hypothesis 4.4. Let (X, τ, φ0 , S, A) be the fftp automaton for (Γ, X, θ).
Let ∆ be the fftp digraph. The projection map S → S/ ∼M extends to digraph map π : ∆ → ∆
such that for any two paths in the fftp digraph that represent geodesic paths to the same vertex
of Γ, their projection to ∆ end in the same state.
We will say that ∆ is a M -type digraph.
Proof. We note that the canonical projection map on vertices given by φ 7→ [φ], does not
extend to a canonical map on edges in general, and the construction will depend on the
choices of automorphism realizing the M -equivalence between vertices of ∆.
We fix a representative ψ0 for each M -equivalent class of vertices [ψ], and for each ψ ∈ [ψ0 ]
we fix an automorphism αψ ∈ G such that ψ |Bv0 (M ) = (ψ0 ◦ αψ )|Bv0 (M ) . Recall the notation
of Definition 3.1. We have a bijection from X to (v0 )−1
− given by x 7→ e(x) and with inverse
map e 7→ xe . Let x ∈ X and let x0 ∈ X such that ex = α(ex0 ). We map the edges f of ∆,
with f− = ψ ∈ [ψ0 ] and f+ = τ (ψ, x0 ) to the same edge e ∈ ∆ from [ψ0 ] to [τ (ψ0 , x)]. By
construction, this map is surjective and a local isomorphism, and thus a covering map.
The fftp digraph is an X-labeled graph. If two words w, w0 ∈ X ∗ represent geodesics
paths starting at v0 and ending at v, then by Lemma 4.9 the vertices φw and φw0 of ∆ are
M -equivalent mod G.
Remark 4.13. In the case of Cayley graphs, the group of automorphism considered preserves
the labeling of the edges and hence two functions are in the same M -type if and only if they
are equal and the covering map ∆ → ∆ is canonical.
Definition 4.14. Let Γ be a vertex transitive graph and v0 a vertex.
The cone of v ∈ V Γ, denoted cone(v), is the set paths p starting at v that continue a
geodesic from v0 to v. Two vertices v and v 0 have the same cone type mod G if there is an
automorphism of α ∈ G such that α(v) = v 0 and α(cone(v)) = α(cone(v 0 )).
The following is an standard argument and goes back to Cannon [8].
12
YAGO ANTOLÍN
Lemma 4.15. With the hypothesis 4.4 and assume further that P is the collection of all
paths. If φw and φu are M -equivalent mod G then (pw )+ , and (pu )+ have the same cone type
mod G.
Proof. Let v = (pu )+ and v 0 = (pw )+ . Suppose that there is an automorphism β ∈ G
satisfying that β(v) = v 0 and φw (β(a)) = φu (a) for all a ∈ X ≤M . We will prove β(cone(v)) =
cone(v 0 ). Note that by symmetry, it is enough to show that β(cone(v)) ⊆ cone(v 0 ).
Suppose that there is a path q in cone(v) such that β(q) does not belong to cone(v 0 ).
Without loss of generality, we can assume that `(q) is minimal with this property. Let t be
path consisting on pw followed by β(q). Then t is a minimal non-geodesic i.e. a path all
whose proper subpaths are geodesic, and hence it M -fellow travels with a geodesic path r.
Let r be broken into two subpaths r1 from v0 to some point a at distance at most M from v 0
and r2 from v to β(q)+ . Let s1 be a geodesic path in from v0 to β −1 (a) and s2 be β −1 (r2 ).
Let s be the concatenation of s1 and s2 . See Figure 1.
q ∈ cone(v)
v
s2 = β −1 (r2 )
β −1 (a)
β(q)
r2
v 0 = β(v)
a
s1
pu
pw
r1
v0
Figure 1. Paths in the proof of Lemma 4.15.
By (2),
`(s1 ) − `(pu ) = φu (β −1 (a)) = φw (a) = `(r1 ) − `(pw ).
Then `(s) − `(pu ) − `(q) = `(s1 ) − `(pu ) + `(s2 ) − `(q) = `(r1 ) − `(pw ) + `(r2 ) − `(β(q)) =
`(r) − `(t) = −1. Thus pu followed by q is not geodesic, contradicting that q ∈ cone(v).
Assume the hypothesis of the previous lemma. Then Γ has finitely many cone types, and
can we construct the cone-type digraph ∆ as follows. The vertices are the different cone types
(i.e. V ∆ = {cone(v) | v ∈ V Γ}) and and there is an edge from cone(v) to cone(u) if there is
an edge in cone(v) starting at v and ending at u0 with cone(u0 ) equivalent to cone(u) mod G.
Thus, given a covering π : ∆ → ∆ from the fftp digraph to an M -type digraph, we have a
natural map from V ∆ to V ∆ and it is easy to extend it to a digraph covering map ∆ to ∆
by using π to determine to which type of edge in Γ correspond traversing a given edge in ∆.
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
13
Remark 4.16. It is worth recalling the example of Elder [10] of a virtually abelian group with
a generating set that does not have the falsification by fellow traveler property, but it has
finitely many cone-types. Thus fftp is strictly stronger than having finitely many cone-types.
5. Choosing the accepting states and proof of Theorem 3.5
Throughout this section we assume Hypothesis 4.4.
Lemma 5.1. Let Z be a subset of Γ with M -fellow projections.
For every path p in P from v0 to Z such that `(p) > d(v0 , Z), there is a path q ∈ P from
v0 to Z such that `(q) < `(p), d(p+ , q+ ) ≤ M , and p and q asynchronously M 2 -fellow travel.
Proof. Let p ∈ P from v0 to Z such that `(p) > d(v0 , Z).
If p is not geodesic, then by fftp there is a shorter path q ∈ P with the same endpoints
that asynchronously M 2 -fellow travel with it. So the claims of the lemma are satisfied.
So assume that p is geodesic. Recall that p(i), i = 0, 1, . . . , `(p) denotes the ith vertex
in the combinatorial path defined by p. Since `(p) > d(v0 , Z), p+ ∈
/ πZ (v0 ) = πZ (p− ). Let
j = max{i | p+ ∈
/ πZ (p(i))}, that is p(j) is the closest vertex of p whose projection to Z
does not contain the vertex p+ . See Figure 5. Note that j 6= `(p) and by definition of j we
have that p+ ∈ πZ (p(j + 1)). Since d(p(j), p(j + 1)) = 1 by the fellow projections, there is
zj ∈ πZ (p(j)) such that d(zj , p+ ) ≤ M . Let r be a path from p+ to zj . Since p ∈ P is geodesic
and `(r) ≤ M , by Lemma 4.5, the concatenation of p and t asynchronously M 2 -fellow travel
with a geodesic path q ∈ P from v0 to zj .
p+
p(j)
zj ∈ πZ (p(j))
v0
∈ πZ (v0 )
Z
Figure 2. Paths in the proof of Lemma 5.1.
Let G 6 Aut(Γ) be vertex-transitive. Without loss of generalization we will assume that
θ(X) ⊆ G, where (X, θ) is the presentation for paths starting at v0 . Let H = hθ(X)i 6 G.
Since G acts by automorphism on the graph Γ, it preserves the distance d, and we see that
πgZ (gv) = πZ (v) for every g ∈ G, v ∈ V Γ. It follows that for every g ∈ G, if Z has M -fellow
projections/M -bounded projections then so does gZ.
Throughout the rest of the section we assume that Z is a subgraph of Γ with M -fellow
projections and G-proper embeddings in Γ. Since for all v ∈ V Γ, {gZ | v ∈ P
gZ} is finite,
there are finitely many g1 , . . . , gs ∈ G such that tHgi Z = GZ. Thus e(G,Z) = si=1 e(H,gi Z)
14
YAGO ANTOLÍN
P
and i(G,Z) = si=1 i(H,gi Z) and therefore the rationality of e(G,Z) and i(G,Z) will follow from
those of e(H,gi Z) and i(H,gi Z) . Thus from now on, we will assume that G = H = hθ(X)i.
5.1. Regularity of language of (G, Z)-geodesics. We are going to select a set of states
of the fftp-automaton that exactly accept the (G, Z)-geodesics.
Suppose that v0 ∈ Z and that w ∈ X ∗ is such that pw is a geodesic path. It might happen
that `(pw ) = d(v0 , wv0 ) < d(v0 , wZ) but pw is a (G, Z)-geodesic since there is g ∈ G such
that wv0 ∈ gZ and `(pw ) = d(v0 , wv0 ) = d(v0 , gZ). To deal with this, we will consider the
set {gZ | v0 ∈ gZ}, which by the G-proper embedding assumptions, is a finite set Z1 , . . . , Zn .
Then, wZ1 , . . . , wZn is the set of subgraphs {gZ | wv0 ∈ gZ} and thus pw is a (G, Z)-geodesic
if and only if there is i ∈ {1, . . . , n} such that `(pw ) = d(v0 , wZi ).
Recall that all the states of the fftp automaton are functions φw with w ∈ X ∗ that records
the distances from v0 to the vertices in Bwv0 (M ). We let A be the set of those φw that detect
that wv0 minimizes the distance from the vertices of wZi to v0 . That is
(4)
A = {φw | ∃i ∈ {1, . . . , n} s. t. φw (u) ≥ 0 ∀u ∈ X ≤M with wuv0 ∈ wZi }.
Lemma 5.2. Let w ∈ L(P ). Then pw is a (G, Z)-geodesic if and only if φw ∈ A.
Proof. If φw ∈ A, then φw 6= %, and hence pw is a geodesic.
The path pw is a (G, Z)-geodesic if and only if there is g ∈ G such that (pw )+ ∈ gZ and
d(v0 , gZ) = `(p), and hence, if and only if there is i such that (pw )+ ∈ wZi and d(v0 , gZ) =
`(p). By the Lemma 5.1, the latter is equivalent to the non-existence of a path q, with
`(q) < `(pw ), q+ ∈ wZi , d((pw )+ , q+ ) ≤ M and such that q and pw asynchronously M 2 -fellow
travel. Finally, by (1), this is equivalent to φw (u) ≥ 0 for all u ∈ {s ∈ X ≤k | wsv0 ∈ wZi }
which is equivalent to φw ∈ A.
Corollary 5.3. If L(P ) is regular then the language
{w ∈ X ∗ | pw is a (G, Z)-geodesic} ∩ L(P)
is regular.
Proof. Let L be the language accepted by the fftp-automaton with parameter M 2 and accepting states A defined in (4). By definition, L is a regular language. Since the intersection
of regular languages is regular, L ∩ L(P ) is regular (see for example [12, Lemma 1.4.1]). By
Lemma 5.2, L ∩ L(P ) = {w ∈ X ∗ | pw is a (G, Z)-geodesic} ∩ L(P).
5.2. Rationality of (G, Z)-embeddings. Suppose that Z is a finite subgraph of Γ. Without of loss of generality, we can assume that diam(Z) ≤ M and v0 ∈ Z. We will deduce the
rationality of e(G,Z) from the one of e(G,•) . The idea is that the number of gZ ⊆ Bv0 (n) with
v ∈ gZ is always the same number, say C, except when v is close to the border of the Bv0 (n)
where copies of Z might be not embeddable. We will approximate e(G,Z) (n) by Ce(G,•) and
then correct the overcounting coming from the vertices near the border.
Let g, h ∈ G and suppose that gZ = hZ with gv0 6= hv0 . Then there is automorphism of
Z that sends v0 to g −1 hv0 . Let O = GZ v0 be the orbit of v0 under GZ = {g ∈ G | gZ = Z},
the G-stabilizer of Z. Then, we have that
X
1
e(G,Z) (n) =
]{gZ | gv0 = v, gZ ⊆ Bv0 (n)}.
|O|
v∈Bv0 (n)
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
15
Suppose that d(v0 , v) ≥ n − M . Then the ’shape’ of Bv0 (n) ∩ Bv (M ) depend on the M -type
of v. For example, in the Cayley graph of Z × Z with respect to a basis {(1, 0), (0, 1)} one has
that B(0,0) (3) ∩ B(3,0) (1) contains two vertices, while B(0,0) (3) ∩ B(2,1) (1) contains 3 vertices.
However, since the M -type records all the relative distances from v0 to the vertices of Bv (n),
the ’shape’ of Bv0 (n) ∩ Bv (M ) is completely determined by the M -type of v and d(v0 , v).
So we split the previous number as follows
X
]{gZ | gv0 = v, gZ ⊆ Bv0 (n)}
|O|e(G,Z) (n) =
v∈Bv0 (n−M )
+
M
X
X
i=1 σ∈V /∼M
X
]{gZ | gv0 = v, gZ ⊆ Bv0 (n)}.
v of type σ
d(v0 ,v)=n−M +i
Note that the first term of the sum above is equal to e(G,•) (n−M ) multiplied by some constant
E, and E is equal to the number of gZ with gv0 fixed. Similarly, each summand on the second
term is equal to eσ(G,•) (n−M +i) multiplied by some other constant Eiσ , and Eiσ is equal to the
number of gZ with gv0 = v of type σ and max{d(v0 , u) | u ∈ gZ} ≤ n = d(v0 , gv0 ) + M − i.
We have written the last equality to emphasize that since φw contains all the information
of the M -type, one can read from there the number of different copies of gZ of Z with
gv0 = w0 in Bwv0 (M ) ∩ Bv0 (n) for every n, that is, the values of the numbers Eiσ .
We have expressed e(G,Z) (n) as finitely many sums of the eσ(G,•) (n) and eσ(G,•) (n). It follows
P
by Theorem 4.11 that n≥0 e(G,Z) (n)tn is a sum of finitely many rational functions.
5.3. Rationality of (G, Z)-intersections. Now we assume also that Z has M -bounded
projections. For each v ∈ V Γ there is finite number of gZ, say Cv , satisfying that v = gv0
and d(v0 , v) = d(v0 , gZ). The number Cv depend only on the M -type of v. Indeed, by the
the G-properness, there exists Z1 , . . . , Zn such that if v = hv0 then any gZ with gv0 = v
is equal to one of the hZ1 , . . . , hZn . By Lemma 5.1 we can decide for which i’s one has
d(v0 , hZi ) = d(v0 , v), and moreover this number of i’s be read from the M -type of v.
P
[v]
One can estimate i(G,Z) as v∈V /∼M C[v] e(G,•) however we might produce some overcounting. If gv0 = v and d(v0 , gZ) = d(v0 , v), then v ∈ πgZ (v0 ) and by the M -bounded projections
assumption any other u ∈ πgZ (v0 ) is at distance ≤ M and again, one can read the number
of such u from the M -type of v, say that the number is Dv .
We have
X
i(G,Z) (n)tn =
n≥0
X
X
n≥0 σ∈V Γ/∼M
Cσ σ
e
(n)tn .
Dσ (G,•)
By Theorem 4.11, the right-hand side of the above expression (and hence the left-hand side)
is a rational function.
5.4. Proof of Theorem 3.5.
Proof of Theorem 3.5. Item (1) follows from Corollary 5.3 and the fact that the growth series
of a regular language is rational. Item (2) follows from the discussion of Subsection 5.2 and
Item (3) follows from the discussion of Subsection 5.3.
16
YAGO ANTOLÍN
6. Schreier coset graphs
Suppose that X is symmetric, i.e. X = X −1 , then the involution on X, extends to an
−1
−1
involution on X ∗ , also denoted by −1 , defined by x1 x2 . . . xn 7→ x−1
n xn−1 . . . x1 . Observe
that Geo(G/H, X) = Geo(H\G, X)−1 , where Geo(G/H, X) = {w ∈ X ∗ | `(w) ≤ `(u) ∀u ∈
X ∗ , u ∈ wH} and Geo(H\G, X) was defined similarly in Theorem B. In particular, since the
reverse of a regular language is regular [12, Theorem 1.2.8], we have that Geo(G/H, X) is
regular if and only if Geo(H\G, X) is regular.
Let |gH|X = min{`(w) | w ∈ X ∗ , w ∈ gH} and analogously |Hg|X = min{`(w) | w ∈
w ∈G Hg}. Note that |Hg −1 |X = |gH|X and hence, for each n ∈ N |B(H\G,X) (n)| =
|B(G/H,X) (n)| where B(G/H,X) (n) = {gH ∈ G/H | |gH|X ≤ n}.
X ∗,
One advantage of working with left cosets as subsets of Γ = Γ(G, X) is that the left action
on G on G/H is isometric, i.e. dΓ (aH, bH) = dΓ (caH, cbH) for all a, b, c ∈ G.
As noticed before, H has D-fellow projections if and only if gH has D-fellow projections.
Remark 6.1. Let Γ = Γ(G, X) be a Cayley graph with X symmetric. Then G 6 Aut(Γ)
is vertex-transitive, and there is a canonical presentation for paths which is induced by the
canonical group homomorphism from F (X), the free group on X, to G. Here the vertex v0
will be 1G . If Z is a subgraph of Γ, then we will denote the set of (G, Z)-geodesics simply
by Geo(G/Z, X). We note that this agrees with the notation above.
Corollary 6.2. If (G, X) has fftp and H 6 G has fellow projections in (G, X), then
Γ(G, H, X) has fftp relative to the collection of paths starting at H.
Proof. Let M be the fftp constant for (G, X) and the fellow projections constant for H in
(G, X). We will see that Γ(G, H, X) has M 2 -fftp.
Let p be a path in Γ(G, H, X) that is not geodesic. Let w ∈ X ∗ be the label of p. Then the
path pw−1 ∈ Γ(G, X) from 1 to w−1 H is not an H-geodesic. By Lemma 5.1, since Z = w−1 H
has M -fellow projections, there is u ∈ X ∗ such that `(u) < `(w), uH = w−1 H, and pu and
pw−1 asynchronously M 2 -fellow travel and dX ((pu )+ , (pw−1 )+ ) = |u−1 w−1 |X ≤ M . Take
h ∈ H such that uh = w−1 . Then the path pw in Γ(G, X) starting at 1 and labeled by w,
M 2 -fellow travels with the path h−1 pu−1 starting at h−1 and labeled by u−1 . Note that the
there is a graph map ρ : Γ(G, X) → Γ(G, H, X) that preserves labels and does not increase
distances. Thus, ρ(h−1 pu−1 ) and ρ(pw ) have the same endpoints, and asynchronously M 2 fellow travel.
Proof of Theorem B. Item (1) is Corollary 6.2. By Theorem 3.5 (1), {w ∈ X ∗ |
pw is a (G, Z)-geodesic} ∩ L(P)) is regular (since Γ(G, X) has fftp relative to P the set
of all the paths P starting at H, L(P) = X ∗ ) and it is equal to Geo(G\H, X), and by the
discussion above Geo(H/G, X) is also regular. Finally, (3) follows from Theorem 3.5 (3).
6.1. Shortlex coset transversals. Fix an order on X and denote by ≤SL the shortlex
order on X ∗ , i.e. w ≤SL v if and only if `(w) < `(v) or `(w) = `(v) and w precedes v
lexicographically (with the fixed order on X). Let ShortLex(G/H, X) = {w ∈ X ∗ | w ≤SL
v, ∀v ∈ X ∗ with vH = wH} and similarly ShortLex(H\G, X) = {w ∈ X ∗ | w ≤SL v, ∀v ∈
X ∗ with Hv = Hw}.
Proposition 6.3. Suppose that (G, X) is shortlex automatic and H 6 G is a subgroup with
bounded projections. Then ShortLex(G/H, X) is a regular language.
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
17
Proof. Without loss of generalization, we can assume that M is the fftp constant and the
bounded projections constant. Recall that if (G, X) is shortlex automatic it means that
Shortlex(G, X) is a geodesic automatic structure. Let P be the paths in Γ(G, X) with label
in Shortlex(G, X). By the Example 3.4 Γ(G, X) has fftp relative to P and P is 1G -spanning.
By Corollary 5.3, we have that L = Geo(G/H, X) ∩ Shortlex(G, X)({ε} ∪ X) is a rational
language. It is clear that ShortLex(G/H, X) ⊆ L. Now suppose that there are w, w0 ∈ L
with wH = w0 H. By the bounded projection property dX (w, w0 ) ≤ M and since w, w0 ∈
ShortLex(G, X), w and w0 M 2 -fellow travel.
In particular
Shortlex(G/H, X) = {w ∈ L | w ≤SL w0 , ∀w0 ∈ L, w0 H = wH}
and bearing in mind that asynchronous fellow travel of geodesics imply synchronous fellow
travel (with potentially a larger constant), there is an standard argument showing that righthand set is a regular language (for example [12, Proof of Theorem 2.5.1]).
Example 6.4 (Redfern). Suppose that G is hyperbolic, X is an ordered generating set and
H is a quasi-convex subgroup (and as discussed in the next section, it has bounded projections). Hyperbolic groups are fftp and shortlex automatic, and hence it follows that
ShortLex(G/H, X) is regular. Redfern in his PhD thesis [28] proved that (G, H, X) is shortlex coset automatic, which implies that ShortLex(H\G, X) is regular.
7. Examples
The falsification by fellow traveler property is known to depend on the generating set [23]
(see also [11]). For some families of fftp graphs, we will provide examples of subgraphs with
bounded projections. This subgraphs are typically quasi-convex, i.e. recall that a subgraph
Z of Γ is σ-quasi-convex, if for every x, y ∈ Z every geodesic path p from x to y lies in the
σ-neighborhood of Z.
7.1. Quasi-convex subsets of hyperbolic spaces. It follows immediately from the thintriangle definition, that if a graph is δ-hyperbolic, then it has δ-fftp. The following follows
from [6, Corollary 2.3].
Lemma 7.1. Let Γ be a δ-hyperbolic graph, and Z a σ-quasi-convex subset. There exists
a constant k, depending only of δ and σ such that for any x, y ∈ Z and any zx ∈ πZ (x),
zy ∈ πZ (y) one has that d(zx , zy ) ≤ k + d(x, y).
Corollary 7.2. Quasi-convex subgroups of hyperbolic groups have bounded projections.
As we have seen in the proof of Theorem 3.5, the fftp constant and the bounded projection
constant determine the size of the digraph codifying the random Markov geodesic combing.
We will use now the idea of [13, Lemma 8.2], to see that in the case of δ-hyperbolic graphs the
construction of the Markov geodesic combing depends mainly on δ and no other parameter.
The idea is that if Z is a subset of Γ with D-bounded projections and v ∈ V Γ is a vertex far
away from Z, then Z-geodesics from v δ-fellow travel except from a constant time depending
on D. Indeed, let p, q be geodesics from v to Z realizing the minimum distances between v
and Z. Let t be a geodesic from p+ to q+ . Then we have a geodesic triangle. As we are in
hyperbolic space, the triangle has a δ 0 -center (where δ 0 only depends on δ), that is, there is
a point x that is at distance δ 0 of the other three sides. Say that xp ∈ p, xq ∈ q and xt ∈ t
are three vertices at distance at most 2δ 0 of each other. It follows that xp (resp. xq ) is at
18
YAGO ANTOLÍN
distance at most D + 2δ 0 of p+ (resp. xq ). In particular the subpath of p from v to xp and
the subpath of q from v to xq synchronously fellow travel with a constant depending only on
δ (one can take 2δ 0 · δ). We have shown:
Lemma 7.3. Let Γ be a δ-hyperbolic graph, and Z a set with D-bounded projections. There
exists constants M = M (δ) only depending on δ and R = R(D, δ) only depending on D and δ
such that for any v ∈ V Γ and any two Z-geodesic paths p, q from v to Z, the initial subpaths
p0 and q 0 of p and q of length d(v, Z) − R synchronously M -fellow travel.
Proof of Theorem C. We now assume that G is an hyperbolic group, X is a finite symmetric
generating set, Γ = Γ(G, X) the Cayley graph of G and Z is a quasi-convex subgroup of G.
Let M and R the constants of the previous lemma.
We will show
P that there is na polynomial QX (t) ∈ Z[t], only depending on G and X, such
that QX (t) · ( n≥0 i(G,Z) (n)t ) is a polynomial. First observe that arguing as in (3), it is
P
enough to show that QX (t)( n≥0 i(G,Z) (n)tn ) is a polynomial where i(G,Z) (n) = i(G,Z) (n) −
i(G,Z) (n − 1) for n ≥ 0 and setting i(G,Z) (−1) = 0.
Let n > R. Then i(G,Z) (n) is the cardinality of S = {gZ | gZ ∩ B1G (n) 6= ∅ = gZ ∩
B1G (n − 1)}. Let v ∈ Γ, with d(1G , v) = n − R and let Sv the subset of S of those gZ
for which there is a geodesic path from 1G to gZ going through v. By the Lemma 7.3, the
cardinality of Sv only depends on the M -type of v. Similarly, if there are is also u ∈ Γ with
d(1G , v) = d(1G , u) = n − R and there is gZ ∈ S with geodesic paths p, q from 1G to gZ
going through u and v respectively, by Lemma 7.3 then d(u, v) ≤ M and the cardinality of
such u is determined by the M -type of v. Thus, there are constants Cσ , Dσ such that
X
X
i(G,Z) (n)tn =
n≥R
X
n≥M σ∈V Γ/∼M
Cσ σ
e
(n − R)tn .
Dσ (G,•)
By Theorem 4.11, there are polynomials Pσ and Qσ over Z[t] such that Pσ (t)/Qσ (t) =
Cσ Pσ (t)tR P
σ
n
= n≥R eσ(G,•) (i − R)tn . Therefore
n≥0 e(G,•) (n)t and thus
Dσ Qσ (t)
P
X
i(G,Z) tn =
Q
σ
X
i(G,Z) tn +
i=0
n≥0
and we can take QX (t) =
R−1
X
σ∈V Γ/∼M
Cσ Pσ (t)tR
Dσ Qσ (t)
Qσ (t). This completes the proof of (1).
To show (2), it follows
q from the previous computation, Theorem 4.11 and [15, Proposition
3.5.] that lim supn→∞ n i(G,Z) (n) = max{ρA , 1} where A is the matrix provided by Theorem
4.11 and ρA is the Perron-Frobenius eigenvalue. If G is non-elementary and H is of infinite
index, then Γ(G, H, X) grows exponentially (Kapovich [20, Theorem 1.5] shows there is a
g
quasiconvex free subgroup F ≤ G of rank two such
qthat H ∩ F = {1} for every g ∈ G) and
hence ρA > 1 and thus we have that lim supn→∞
n
i(G,H) (n) = ρA > 1.
Recall that Koubi [21] showed that if G is non-elementary, then
there is λ > 1 such that
p
n
for any finite generating Y set of G it holds that lim supn→∞ |BY (n)| > λ. Thus, we get
that ρA > λ for all generating sets X.
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
19
7.2. Parabolic subgroups of relatively hyperbolic groups. Let G be a group hyperbolic relative to a collection of subgroups {Hω }ω∈Ω . The subgroups Hω , ω ∈ Ω are called
parabolic subgroups (see [26] for the details). Let H = ∪Hω .
In this subsection we will make intense use of results of [1] where it is shown that fftp
is preserved under relative hyperbolicity in the following way. Suppose that Y is a finite
generating set for G. There is a finite subset H0 ⊆ H such that for any finite generating X
satisfying that Y ∪ H ⊆ X ⊆ Y ∪ H and that Γ(Hω , X ∩ Hω ) is fftp, it follows that Γ(G, X)
is fftp. We remark that the condition Γ(Hω , X ∩ Hω ) is fftp, it is relatively easy to achieve,
since if Hω has fftp for some generating set, then any finite generating set of Hω can be
enlarged to have fftp (see [1, Proposition 3.2]).
Let p be a path in the Cayley graph Γ(G, X ∪ H). An Hλ -component of p, is a subpath s
of p with the property that the label of the path s is an element of the free monoid Hλ∗ and
it is not properly contained in any other subpath of p with this property. Two components s
and r (not necessarily in the same path) are connected if both are Hω -components for some
ω ∈ Ω and (s− )Hω = (r− )Hω . A component in a closed path that is not connected to other
component is called isolated.
We will use the following result, which is a version of [27, Proposition 3.2].
Lemma 7.4. Let G be hyperbolic relative to {Hω }ω∈Ω and X a finite generating set of G.
There exists D = D(G, X, λ, c) > 0 such that the following hold. Let P = p1 p2 · · · pn be an
n-gon in Γ(G, X ∪ H) and I a distinguished subset of sides of P such that if pi ∈ I, pi is an
isolated component in P, and if pi ∈
/ I, pi is a (λ, c)-quasi-geodesic. Then
X
dX ((pi )− , (pi )+ ) ≤ Dn.
i∈I
Lemma 7.5. Let G be hyperbolic relative to {Hω }ω∈Ω and Y a finite symmetric generating
set. Then there is a finite subset H0 ⊆ H such that for every generating set X of G satisfying
that Y ∪ H0 ⊆ X ⊆ Y ∪ H, any Hω has bounded projections (and hence fellow projections)
in Γ(G, X).
Proof. From the Generating Set Lemma [1, Lemma 5.3], there is a finite set H0 of H and
constants λ ≥ 1, c ≥ 0 with the property that for any finite symmetric generating set X
such that Y ∪ H0 ⊆ X ⊆ Y ∪ H one has that for any geodesic word w ∈ X ∗ there is a
word w
b in (X ∪ H)∗ with w =G w,
b w
b labels a (λ, c)-quasi-geodesic in Γ(G, X ∪ H) and such
that any prefix of w
b represents a prefix of w in the following way: if w ≡ x1 . . . xn and
w
b ≡ z1 . . . zm , m ≤ n, there is an increasing function f : {1, . . . , m} → {1, . . . , n} such that
z1 . . . zi =G x1 . . . xf (i) . In particular, the set of group elements that appears as vertices in
the path in Γ(G, X ∪ H) starting at g and labeled by w
b is a subset of the group elements
that appears as vertices in the path in Γ(G, X) starting at g and labeled by w.
Fix a symmetric generating set X as above. Fix ω ∈ Ω. We will use π X to denote
projections in the space Γ(G, X) and π X∪H to denote projections in Γ(G, X ∪ H). Let
X (a) and z ∈ π X (b) and let zb ∈ π X∪H and
a, b ∈ G, with dX (a, b) = 1. Let za ∈ πH
a
b
Hω
Hω
ω
X∪H (b).
zbb ∈ πH
ω
We will first show that dX (zba , zbb ) is uniformly bounded, and then that dX (za , zba ) and
dX (zb , zbb ) are uniformly bounded.
Let D be the constant of Lemma 7.4. An important fact is that there is constant m > 0
such that if an Hω -component s is connected to an Hµ -component, and dX (s− , s+ ) > m,
20
YAGO ANTOLÍN
e
a
qa
ra
za
zba
b
qb
Z
zbb
Figure 3. Paths involved in the proof of 7.5
then µ = ω (see for example, [1, Lemma 4.2]). Without loss of generality, we can assume
m < D.
Let qa and qb be geodesics in Γ(G, X ∪ H) from a to zba and from b to zbb respectively.
Let e be the edge with e− = a, e+ = b (this edge has its label in X). Let f be the edge
with f− = zba and f+ = zbb (and label in Hω ) and consider the geodesic 4-gon with sides
e, f, qa , qb . We have to bound dX (f− , f+ ). Suppose that dX (f− , f+ ) > m. If f is not isolated
in the 4-gon, say it is connected to a component of qa , then this component must be an
Hω -component and hence dX∪H (a, zba ) > dX∪H (a, Hω ) a contradiction. We obtain a similar
contradiction if f is connected to a component of qb . If e and f are connected, then e is an
Hω -component and {a, b} ∈ Hω and then a = zba and b = zbb and dX (za , zb ) = 1. Thus, we
are left with the case where f is isolated, and hence by Lemma 7.4, dX (zba , zbb ) ≤ 3D.
Now let w ∈ X ∗ a label of a geodesic path from a to za . Let w
b be as above, and ra the
paths from a to za in Γ(G, X ∪ H) with label w.
b Let t be the edge from za to zba (with
label in Hω ). Consider the (λ, c)-quasi-geodesic triangle with sides ra , qa , t. Assume that
dX (t− , t+ ) > m. By the above argument, t can not be connected to a component of qa . If t
is connected to a component of ra , then there is vertex of ra in Hω , which means that there
is a prefix of w0 of w such that aw0 ∈ Hω , which means that dX (a, pa ) > `(w0 ) ≥ dX (a, Hω )
giving a contradiction. Then t is isolated in the triangle and dX (pa , zba ) ≤ 2D.
The same arguments shows that dX (zb , zbb ) ≤ 2D and hence dX (za , zb ) ≤ 7D.
Combining the lemma with the results of [1] discussed in the introduction we have:
Corollary 7.6. Let G finitely generated and hyperbolic relative to {Hω }ω∈Ω such that each
Hω has the falsification by fellow traveler property respect to some finite generating set. Then
there is a finite generating set X of G such that Γ(G, X) is fftp and each Hω has bounded
projections in Γ(G, X).
7.3. Quasi-convex subsets of CAT(0) cube complexes. Let C be a locally finite,
CAT(0) cube complex. Let Γ be the 1-skeleton of C. Then Γ is an fftp graph. This fact
appears in the proof [24, Theorem 1.1.], where there is an extra hypothesis to make Γ a
Cayley graph, however, the fact that it is a Cayley graph is not used through the proof (for
showing fftp), the point being that a path in Γ is geodesic if and only if it does not cross
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
21
twice the same wall. Using the CAT(0)-cubical geometry, Noskov shows that one can take 2
as (synchronous) fftp constant.
Lemma 7.7. Let Γ be the 1-skeleton of a locally finite CAT(0) cube complex. Let Z be a
σ-quasi-convex subset of Γ. Then Z has bounded projections.
Proof. For x, y ∈ V Γ, let I[x, y] denote the set of vertices that appear in combinatorial
geodesics from x to y. Note that Γ is a median graph [5, Theorem 6.1.], that is given any
3 vertices x, y, z ∈ V Γ, the intersection I(x, y) ∩ I(x, z) ∩ I(z, y) consist on a single vertex
denoted µ(x, y, x) and called the median of x, y, z.
Let x, y ∈ V Γ, dΓ (x, y) = 1 and zx ∈ πZ (x), zy ∈ πZ (y). Let mx be the median of x, zx , zy
and my the median of y, zy , zx . Since mx is in a geodesic from zx to zy , dΓ (mx , Z) ≤ σ.
Since zx ∈ Z and mx is in a geodesic path that realizes the minimum distance from x to Z,
we have that dΓ (mx , zx ) ≤ σ. Similarly, dΓ (my , zy ) ≤ σ.
Now, the median map µ is 1-Lipschitz [4, Corollary 2.15], and thus since d(x, y) = 1, we
have that d(µ(x, zx , zy ), µ(y, zx , zy )) ≤ 1 and thus d(zx , zy ) ≤ 1 + 2σ.
7.4. Parabolic subgroups of right angled Artin groups. Recall that a right angled
Artin group G has a standard presentation hX k Ri in which the only relations of R consist
on commuting relations among the elements of X. A parabolic subgroup P 6 G is a subgroup
that is G-conjugate to a subgroup hY i where Y is some subset of X. The Cayley graph
of Γ(G, X) is the 1-skeleton of the Salvetti Complex, that is a locally finite CAT(0) cube
complex. Therefore Γ(G, X) has fftp. By [16, Lemma 3.6], P is quasi-convex in Γ(G, X).
Thus we obtain the following corollary of Lemma 7.7, which gives another interesting
family of examples for Theorem B.
Corollary 7.8. Let G be a finitely generated right angled Artin group, X the standard
generating set, and P a parabolic subgroup of G. Then Γ(G, X) has the falsification by
fellow traveler property and P has bounded projections in Γ(G, X).
7.5. Standard parabolic subgroups of Coxeter groups. Let S be a finite set. A Coxeter
matrix over S is a square matrix M = (ms,t )s,t∈S with coefficients in N ∪ {∞} such that
ms,s = 1 for all s ∈ S and ms,t = mt,s > 2 for all s, t ∈ S, s 6= t. The Coxeter group
associated to M is the group given by the following presentation:
W = hS | s2 = 1 for all s ∈ S , (st)ms,t = 1 for all s, t ∈ S such that s 6= t and ms,t 6= ∞i .
Let X be a subset S. We denote by WX the subgroup of W generated by X. By Bourbaki
[2], WX is a Coxeter group (associated to the submatrix of M corresponding to X). It is
called a standard parabolic subgroup of W . We recall now some standard facts about Coxeter
groups.
We say that w ∈ W is X-reduced if it is of minimal length in its coset WX w.
Lemma 7.9. (Bourbaki [2])
(1) There exists a unique X-reduced element in each coset c ∈ WX \W .
(2) Let u ∈ W . Then u is X-reduced if and only if |su|S = |u|S + 1 for all s ∈ X.
(3) Let w ∈ W . Let u be the (unique) X-reduced element lying in WX w and let v ∈ WX
such that w = vu. Then |w|S = |v|S + |u|S .
22
YAGO ANTOLÍN
We will also make use of the “Exchange condition” that characterizes the Coxeter groups
(see [7, Chapter 4] for example).
Lemma 7.10. Let w ∈ W and let s, t ∈ S such that |sw|S = |wt|S = |w|S + 1 and |swt|S <
|w|S + 1. Then sw = wt.
The following proof was provided by Luis Paris.
Proposition 7.11. Let W be a Coxeter group, S a set of Coxeter generators and X ⊆ S.
Then WX has 1-bounded projections in Γ(W, S).
Proof. Let w ∈ W . Write w = vu where u is X-reduced and v ∈ WX . By Lemma 7.9, for all
v 0 ∈ WX , we have
dS (v 0 , w) = |v 0
−1
vu|S = |v 0
−1
v|S + |u|S > |u|S = dS (v, w) ,
and we have equality if and only if |v 0 −1 v|S = 0, that is, if and only if v 0 = v. So, πWX (w) =
{v}.
Let w0 ∈ W such that dS (w0 , w) = 1. Let t ∈ S such that w0 = wt. Upon exchanging
w0 and w we may assume that |w0 |S = |wt|S = |w|S + 1. If ut is X-reduced, then, by the
above, πX (w0 ) = {v} (since wt = vut) and the diameter of πWX (w) ∪ πWX (w0 ) = {v} is 0.
So, we may assume that ut is not X-reduced. By Lemma 7.9 there exists s ∈ X such that
|sut|S < |ut|S = |u|S +1. We also have |ut|S = |u|S +1 by hypothesis and |su|S = |u|S +1 since
u is X-reduced. By Lemma 7.10 this implies that su = ut, hence w0 = vsu. Since u is Xreduced, it follows that πWX (w0 ) = {vs}, hence the diameter of πWX (w) ∪ πWX (w0 ) = {v, vs}
is 1.
Acknowledgments This work was originated by a question of Ian Leary, who asked the
author about growth Schreier graphs of relatively hyperbolic groups with respect to parabolic
subgroups, in the ferry returning from the conference in Ventotene 2015. The relative falsification by fellow traveler property appeared in the paper after a question of Ilya Gekhtman
about the case of automatic groups. Finally, the author thanks Luis Paris for his interest
and for providing the proof of Proposition 7.11.
References
[1] Y. Antolı́n and L. Ciobanu, Finite generating sets of relatively hyperbolic groups and applications to geodesic
languages, Trans. Amer. Math. Soc. 368 (2016), no. 11, 7965–8010.
[2] N. Bourbaki, Éléments de mathématique. Fasc. XXXIV. Groupes et algèbres de Lie. Chapitre IV: Groupes
de Coxeter et systèmes de Tits. Chapitre V: Groupes engendrés par des réflexions. Chapitre VI: systèmes
de racines, Actualités Scientifiques et Industrielles, No. 1337, Hermann, Paris, 1968.
[3] D. Calegari and K. Fujiwara, Counting subgraphs in hyperbolic graphs with symmetry. J. Math. Soc. Japan
67 (2015), no. 3, 1213–1226.
[4] I. Chatterji, C. Drutu, F. Haglund, Kazhdan and Haagerup properties from the median viewpoint. Adv.
Math. 225 (2010), no. 2, 882–921.
[5] V. Chepoi, Graphs of some CAT(0) complexes. Adv. in Appl. Math. 24 (2000), no. 2, 125 –179.
[6] M. Coornaert, T. Delzant, and A. Papadopoulos, Geometrie et theorie des groupes, Lecture Notes in
Math., vol.1441, Springer Verlag (1990).
[7] M W. Davis, The geometry and topology of Coxeter groups, London Mathematical Society Monographs
Series, 32. Princeton University Press, Princeton, NJ, 2008.
[8] J. W. Cannon, The combinatorial structure of cocompact discrete hyperbolic groups, Geom. Dedicata, 16
(1984), 123–148.
[9] M. Elder, Finiteness and the falsification by fellow traveler property. Geometriae Dedicata 95 (2002)
103–113.
GROWTH IN VERTEX-TRANSITIVE, FFTP GRAPHS
23
[10] M. Elder, Regular languages and the falsification by fellow traveler property. Algebraic and Geometric
Topology Vol 5 (2005), paper no. 8, pages 129–134.
[11] A. Elvey-Price, A Cayley graph for F2 × F2 which is not minimally almost convex. arXiv:1611.00101.
[12] D. B. A. Epstein, J. Cannon, D. Holt, S. Levy, M. Paterson, M. and W. Thurston, W., Word Processing
in Groups. Jones and Bartlett, Boston, 1992.
[13] D. B. A. Epstein, A. R. Iano-Fletcher, and U. Zwick, Growth functions and automatic groups. Experiment.
Math. 5 (1996), no. 4, 297–315.
[14] A. Eskin, D. Fisher and K. Whyte, Coarse differentiation of quasi-isometries I: Spaces not quasi-isometric
to Cayley graphs. Ann. of Math. (2) 176 (2012), no. 1, 221–260.
[15] F. Dahmani, D. Futer, D. T. Wise, Growth of quasiconvex subgroups, arXiv:1602.08085
[16] T. Hsu and D. T. Wise, Separating quasiconvex subgroups of right-angled Artin groups. Math. Z. 240
(2002), no. 3, 521–548.
[17] M. Gromov, Hyperbolic groups. Essays in group theory, 75–163, Math. Sci. Res. Inst. Publ., 8, Springer,
New York, 1987.
[18] D. F. Holt, Garside groups have the falsification by fellow-traveler property. Groups Geom. Dyn. 4 (2010),
777–784.
[19] D. F. Holt and S. Rees, Artin groups of large type are shortlex automatic with regular geodesics. Proc.
Lond. Math. Soc. (3) 104 (2012), no. 3, 486–512.
[20] I. Kapovich The nonamenability of Schreier graphs for infinite index quasiconvex subgroups of hyperbolic
groups. Enseign. Math. (2) 48 (2002), no. 3-4, 359–375.
[21] M. Koubi, Croissance uniforme dans les groupes hyperboliques. Ann. Inst. Fourier (Grenoble) 48 (1998),
no. 5, 1441–1453.
[22] J. Loeffler, J. Meier and J. Worthington, Graph products and Cannon pairs, Internat. J. Algebra Comput.
12 (2002), 6, 747–754.
[23] W. D. Neumann and M. Shapiro, Automatic structures, rational growth, and geometrically finite hyperbolic groups. Invent. Math. 120 (1995), no. 2, 259–287.
[24] G. A. Noskov, Growth of certain non-positively curved cube groups, Europ.J. Combinatorics 21 (2000),
659–666.
[25] G. A. Noskov, Bounded shortening in Coxeter complexes and buildings. pp. 1014 in: A. K. Guts (Ed.),
Mathematical structures and modeling, No. 8, Omsk. Gos. Univ., Omsk 2001. Available electronically at
http://cmm.univer.omsk.su/sbornik/sborn8.html.
[26] D. Osin, Relatively hyperbolic groups: intrinsic geometry, algebraic properties, and algorithmic problems.
Mem. Amer. Math. Soc. 179 (2006), no. 843, vi+100 pp.
[27] D. Osin, Peripheral fillings of relatively hyperbolic groups. Invent. Math. 167 (2007), no. 2, 295–326.
[28] I. D. Redfern Automatic Coset Systems, PhD Thesis, University of Warwick, 1993.
[29] K, Saito, The limit element in the configuration algebra for a discrete group: a prcis. Proceedings of the
International Congress of Mathematicians, Vol. I, II (Kyoto, 1990), 931–942, Math. Soc. Japan, Tokyo,
1991.
Departamento de Matematicas, Universidad Autonoma de Madrid and Instituto de Ciencias
Matematicas, CSIC-UAM-UC3M-UCM .
E-mail address, Yago Antolı́n: [email protected]
| 4math.GR
|
arXiv:1602.06613v2 [math.AC] 3 Mar 2016
A DUALITY IN BUCHSBAUM RINGS AND
TRIANGULATED MANIFOLDS
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
Abstract. Let ∆ be a triangulated homology ball whose boundary complex is ∂∆.
A result of Hochster asserts that the canonical module of the Stanley–Reisner ring of
∆, F[∆], is isomorphic to the Stanley–Reisner module of the pair (∆, ∂∆), F[∆, ∂∆].
This result implies that an Artinian reduction of F[∆, ∂∆] is (up to a shift in grading)
isomorphic to the Matlis dual of the corresponding Artinian reduction of F[∆]. We
establish a generalization of this duality to all triangulations of connected orientable
homology manifolds with boundary. We also provide an explicit algebraic interpretation
of the h′′ -numbers of Buchsbaum complexes and use it to prove the monotonicity of h′′ numbers for pairs of Buchsbaum complexes as well as the unimodality of h′′ -vectors of
barycentric subdivisions of Buchsbaum polyhedral complexes. We close with applications
to the algebraic manifold g-conjecture.
1. Introduction
In this paper, we study an algebraic duality of Stanley–Reisner rings of triangulated
homology manifolds with non-empty boundary. Our starting point is the following (unpublished) result of Hochster — see [25, Ch. II, §7]. (We defer most of definitions until
later sections.) Let ∆ be a triangulated (d − 1)-dimensional homology ball whose boundary complex is ∂∆, let F[∆] be the Stanley–Reisner ring of ∆, and let F[∆, ∂∆] be the
Stanley–Reisner module of (∆, ∂∆). (Throughout the paper F denotes an infinite field.)
Hochster’s result asserts that the canonical module ωF[∆] of F[∆] is isomorphic to F[∆, ∂∆].
In the last decade or so, this result had a lot of impact on the study of face numbers of
simplicial complexes, especially in connection with the g-conjecture for spheres, see, for
instance, the proof of Theorem 3.1 in a recent survey paper by Swartz [28].
One numerical consequence of Hochster’s result is the following symmetry of h-numbers
of homology balls: hi (∆, ∂∆) = hd−i (∆). The h-numbers are certain linear combinations
of the face numbers; they are usually arranged in a vector called the h-vector. In fact,
Hochster’s result implies a stronger statement: it implies that there is an isomorphism
∨
F[∆, ∂∆]/ΘF[∆, ∂∆] ∼
(1)
= F[∆]/ΘF[∆] (−d),
where Θ is a linear system of parameters for F[∆] and N ∨ is the (graded) Matlis dual of N
(see e.g. [16, Lemma 3.6]). As the F-dimensions of the ith graded components of modules
Murai’s research is partially supported by Grant-in-Aid for Scientific Research (C) 25400043.
Novik’s research is partially supported by NSF grant DMS-1361423.
Yoshida’s research is partially supported by Grant-in-Aid for Scientific Research (C) 25400050.
1
2
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
in (1) are equal to hi (∆, ∂∆) and hd−i (∆), respectively, the above-mentioned symmetry,
hi (∆, ∂∆) = hd−i (∆), follows.
Hochster’s result was generalized to homology manifolds with boundary by Gräbe [8].
To state Gräbe’s result, we recall the definition of homology manifolds. We denote by Sd
and Bd the d-dimensional sphere and ball, respectively. A pure d-dimensional simplicial
complex ∆ is an F-homology d-manifold without boundary if the link of each nonempty
face τ of ∆ has the homology of Sd−|τ | (over F) . An F-homology d-manifold with boundary
is a pure d-dimensional simplicial complex ∆ such that (i) the link of each nonempty face
τ of ∆ has the homology of either Sd−|τ | or Bd−|τ | , and (ii) the set of all boundary faces,
that is,
∂∆ := τ ∈ ∆ : the link of τ has the same homology as Bd−|τ | ∪ {∅}
is a (d−1)-dimensional F-homology manifold without boundary. A connected F-homology
e d (∆, ∂∆) is
d-manifold with boundary is said to be orientable if the top homology H
isomorphic to F.
Gräbe [8] proved that if ∆ is an orientable homology manifold with boundary, then
F[∆, ∂∆] is the canonical module of F[∆]. Gräbe also established a symmetry of hnumbers for such a ∆ (see [9]). While Gräbe’s original statement of symmetry is somewhat complicated, it was recently observed by the first two authors [15] that it takes the
following simple form when expressed in the language of h′′ -numbers.
Theorem 1.1. Let ∆ be a connected orientable F-homology (d − 1)-manifold with nonempty boundary ∂∆. Then h′′i (∆, ∂∆) = h′′d−i (∆) for all i = 0, 1, . . . , d.
The h′′ -numbers are certain modifications of h-numbers (see Section 3 for their definition). Similarly to the h-numbers, the h′′ -numbers are usually arranged in a vector, called
the h′′ -vector. For homology manifolds, this vector appears to be a “correct” analog of
the h-vector. Indeed, many properties of h-vectors of homology balls and spheres are now
known to hold for the h′′ -vectors of homology manifolds (with and without boundary),
see recent survey articles [11, 28]. In light of Gräbe’s result from [8] and Theorem 1.1, it
is natural to ask if Theorem 1.1 can be explained by Matlis duality. The first goal of this
paper is to provide such an explanation.
To this end, the key object is the submodule Σ(Θ; M) defined by Goto [7]. Several
definitions are in order. Let S = F[x1 , . . . , xn ] be a graded polynomial ring over a field
F with deg xi = 1 for i = 1, 2, . . . , n. Let M be a finitely generated graded S-module of
Krull dimension d and let Θ = θ1 , . . . , θd be a homogeneous system of parameters for M.
The module Σ(Θ; M) is defined as follows:
!
d
X
Σ(Θ; M) = ΘM +
(θ1 , . . . , θ̂i , . . . , θd )M :M θi ⊆ M.
i=1
This module was introduced by Goto in [7] and has been used in the study of Buchsbaum
local rings. Note that if M is a Cohen–Macaulay module, then Σ(Θ; M) = ΘM. We first
show that this submodule is closely related to the h′′ -vectors. Specifically, we establish
the following explicit algebraic interpretation of h′′ -numbers.
A DUALITY IN BUCHSBAUM RINGS AND TRIANGULATED MANIFOLDS
3
Theorem 1.2. Let (∆, Γ) be a Buchsbaum relative simplicial complex of dimension d − 1
and let Θ be a linear system of parameters for F[∆, Γ]. Then
dimF F[∆, Γ]/Σ Θ; F[∆, Γ]) j = h′′j (∆, Γ) for all j = 0, 1, . . . , d.
Theorems 1.1 and 1.2 suggest that when ∆ is a homology manifold with boundary
there might be a duality between the quotients of F[∆] and F[∆, ∂∆] by Σ(Θ; F[∆]) and
Σ(Θ; F[∆, ∂∆]), respectively. We prove that this is indeed the case. In fact, we prove a
more general algebraic result on canonical modules of Buchsbaum graded algebras. Let
m = (x1 , . . . , xn ) be the graded maximal ideal of S. If M is a finitely generated graded
S-module of Krull dimension d, then the canonical module of M, ωM , is the module
∨
ωM := Hmd (M) ,
where Hmi (M) denotes the ith local cohomology module of M. We prove that the following
isomorphism holds for all Buchsbaum graded algebras.
Theorem 1.3. Let R = S/I be a Buchsbaum graded F-algebra of Krull dimension d ≥
2,
Pdlet Θ = θ1 , . . . , θd ∈ S be a homogeneous system of parameters for R, and let δ =
i=1 deg θi . If depth R ≥ 2, then
∨
ωR /Σ(Θ; ωR ) ∼
= R/Σ(Θ; R) (−δ).
As we mentioned above, if ∆ is a connected orientable homology manifold, then (by
Gräbe’s result) the module F[∆, ∂∆] is the canonical module of F[∆]; furthermore it is not
hard to see that F[∆] satisfies the assumptions of Theorem 1.3. (Indeed, the connectivity
of ∆ implies that depth F[∆] ≥ 2, and the fact that F[∆] is Buchsbaum follows from
Schenzel’s theorem — see Theorem 3.1 below.) Hence we obtain the following corollary
that generalizes (1).
Corollary 1.4. Let ∆ be a connected orientable F-homology (d − 1)-manifold with nonempty boundary ∂∆ and let Θ be a linear system of parameters for F[∆]. Then
∨
F[∆, ∂∆]/Σ(Θ; F[∆, ∂∆]) ∼
= F[∆]/Σ(Θ; F[∆]) (−d).
We also consider combinatorial and algebraic applications of Theorems 1.2 and 1.3.
Specifically, we prove the monotonicity of h′′ -vectors for pairs of Buchsbaum simplicial
complexes, establish the unimodality of h′′ -vectors of barycentric subdivisions of Buchsbaum polyhedral complexes, provide a combinatorial formula for the a-invariant of Buchsbaum Stanley–Reisner rings, and extend the result of Swartz [28, Theorem 3.1] as well as
the result of Böhm and Papadakis [5, Corollary 4.5] related to the sphere g-conjecture to
the generality of the manifold g-conjecture. More precisely, Swartz’s result asserts that
most of bistellar flips when applied to a homology sphere preserve the weak Lefschetz property while Böhm–Papadakis’ result asserts that stellar subdivisions at large-dimensional
faces of homology spheres preserve the weak Lefschetz property; we extend both of these
results to bistellar flips and stellar subdivisions performed on connected orientable homology manifolds.
The structure of the paper is as follows. In Section 2 we prove Theorem 1.3 (although we
defer part of a proof to the Appendix). In Section 3, we study Stanley–Reisner rings and
4
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
modules of Buchsbaum simplicial complexes. There, after reviewing basics of simplicial
complexes and Stanley–Reisner rings and modules, we verify Theorem 1.2 and derive
several combinatorial consequences. Section 4 is devoted to applications of our results to
the manifold g-conjecture. Finally, in the Appendix, we prove a graded version of Goto’s
result [7, Proposition 3.6] — a result on which our proof of Theorem 1.3 is based.
2. Duality in Buchsbaum rings
In this section, we prove Theorem 1.3. We start by recalling some definitions and results
pertaining to Buchsbaum rings and modules.
Let S = F[x1 , . . . , xn ] be a graded polynomial ring with deg xi = 1 for i = 1, 2, . . . , n
and let m = (x1 , . . . , xn ) be the graded maximal ideal of S. Given a graded S-module N,
we denote by N(a) the module N with grading shifted by a ∈ Z, that is, N(a)j = Na+j .
If M is a finitely generated graded S-module of Krull dimension d, then a homogeneous
system of parameters (or h.s.o.p.) for M is a sequence Θ = θ1 , . . . , θd ∈ m of homogeneous elements such that dimF M/ΘM < ∞. A sequence θ1 , . . . , θr ∈ m of homogeneous
elements is said to be a weak M-sequence if
(θ1 , . . . , θi−1 )M :M θi = (θ1 , . . . , θi−1 )M :M m
for all i = 1, 2, . . . , r. We say that M is Buchsbaum if every h.s.o.p. for M is a weak
M-sequence.
Let M be a finitely generated graded S-module and let Θ be its h.s.o.p. The function
PΘ,M : Z → Z defined by
PΘ,M (n) := dimF M/(Θ)n+1 M
is called the Hilbert–Samuel function of M w.r.t. the ideal (Θ). It is known that there is a
polynomial in n of degree d, denoted by pΘ,M (n), such that PΘ,M (n) = pΘ,M (n) for n ≫ 0
(see [6, Proposition 4.6.2]). The leading coefficient of the polynomial pΘ,M (n) multiplied
by d! is called the multiplicity of M w.r.t. the ideal (Θ), and is denoted by eΘ (M). We
will make use of the following known characterization of Buchsbaum property, see [26,
Theorem I.1.12 and Proposition I.2.6].
Lemma 2.1. A finitely generated graded S-module M of Krull dimension d is Buchsbaum
if and only if, for every h.s.o.p. Θ of M,
d−1
X
d−1
dimF Hmi (M).
dimF M/ΘM − eΘ (M) =
i
i=0
We also recall some known results on canonical modules of Buchsbaum modules. For a
finitely generated graded S-module M, the depth of M is defined by depth(M) := min{i :
Hmi (M) 6= 0}.
Lemma 2.2. Let M be a finitely generated graded S-module of Krull dimension d. If M
is Buchsbaum, then the following properties hold:
(i) ωM is Buchsbaum.
(ii) Hmi (ωM ) ∼
= (Hmd−i+1 (M))∨ (as graded modules) for all i = 2, 3, . . . , d − 1.
A DUALITY IN BUCHSBAUM RINGS AND TRIANGULATED MANIFOLDS
5
(iii) If depth(M) ≥ 2, then (Hmd (ωM ))∨ ∼
= M (as graded modules).
(iv) If Θ is an h.s.o.p. for M, then Θ is also an h.s.o.p. for ωM and eΘ (ωM ) = eΘ (M).
See [26, Theorem II.4.9] for (i), [23, Korollar 3.13] or [4, (1.16)] for (ii) and (iii), and [27,
Lemma 2.2] for (iv).
The following theorem is a graded version of Goto’s result [7, Proposition 3.6]. The
original result by Goto is a statement about Buchsbaum local rings. It may be possible
to prove Theorem 2.3 in the same way as in [7] by replacing rings with modules and
by carefully keeping track of grading. However, since we could not find any literature
allowing us to easily check this statement, we will provide its proof in the Appendix.
Theorem 2.3. Let M be a finitely generated graded S-module of Krull dimension d > 0.
Assume further that MPis Buchsbaum and that Θ = θ1 , . . . , θd is an h.s.o.p. for M with
deg θi = δi . Let δC := i∈C δi for C ⊆ [d] = {1, 2, . . . , d}. Then
L
|C|
(i) Σ(Θ; M)/ΘM ∼
Hm (M)(−δC ), and
=
C([d]
(ii) there is an injection M/Σ(Θ; M) → Hmd (M)(−δ[d] ).
We are now in a position to prove Theorem 1.3.
Proof of Theorem 1.3. We first prove that R/Σ(Θ; R) and ωR /Σ(Θ; ωR ) have the same
F-dimension. Indeed,
dimF R/Σ(Θ; R) = dimF (R/ΘR) − dimF Σ(Θ; R)/ΘR
d−1
d−1
X
X
d
d−1
i
dimF Hmi (R)
dimF Hm (R) −
= eΘ (R) +
i
i
i=0
i=0
d−1
X d−1
dimF Hmi (R),
= eΘ (R) −
i
−
1
i=1
where we use Lemma 2.1 and Theorem 2.3(i) for the second equality. Similarly, since ωR
is Buchsbaum, the same computation yields
d−1
X
d−1
dimF Hmi (ωR ).
dimF ωR /Σ(Θ; ωR ) = eΘ (ωR ) −
i
−
1
i=1
Now, since depth(R) ≥ 2 by the assumptions of the theorem and since depth(ωR ) ≥ 2
always holds, see [3, Lemma 1], parts (ii) and (iv) of Lemma 2.2 guarantee that
dimF R/Σ(Θ; R) = dimF ωR /Σ(Θ; ωR ) .
Thus, to complete the proof of the statement, it suffices to show that there is a surjection
from R/Σ(Θ; R)(+δ) to (ωR /Σ(Θ; ωR ))∨ . By Theorem 2.3(ii), there is an injection
ωR /Σ(Θ; ωR ) → Hmd (ωR )(−δ).
Dualizing and using Lemma 2.2(iii), we obtain a surjection
∨
∨
(2)
R(+δ) ∼
= Hmd (ωR )(−δ) → ωR /Σ(Θ; ωR ) .
6
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
Since ωR is an R-module, it follows from the definition of Σ(Θ; ωR ) that Σ(Θ; R) · ωR ⊆
Σ(Θ; ωR ). Hence
Σ(Θ; R) · ωR /Σ(Θ; ωR ) = 0,
which in turn implies
∨
(3)
Σ(Θ; R) · ωR /Σ(Θ; ωR ) = 0.
Now (2) and (3) put together guarantee the existence of a surjection
∨
R/Σ(Θ; R) (+δ) → ωR /Σ(Θ; ωR ) ,
as desired.
Remark 2.4. The above proof also works in the local setting: it shows that if R is a
Buchsbaum Noetherian local ring with depth R ≥ 2 (and if the canonical module of R
exists), then ωR /Σ(Θ; ωR ) is isomorphic to the Matlis dual of R/Σ(Θ; R).
3. Duality in Stanley–Reisner rings of manifolds
In this section, we study Buchsbaum Stanley–Reisner rings and modules. Some objects
in this section such as homology groups, Betti numbers, h′ - and h′′ -numbers depend on
the characteristic of F; however, we fix a field F throughout this section, and omit F from
our notation.
We start by reviewing basics of simplicial complexes and Stanley–Reisner rings. A
simplicial complex ∆ on [n] is a collection of subsets of [n] that is closed under inclusion.
A relative simplicial complex Ψ on [n] is a collection of subsets of [n] with the property
that there are simplicial complexes ∆ ⊇ Γ such that Ψ = ∆ \ Γ. We identify such a pair of
simplicial complexes (∆, Γ) with the relative simplicial complex ∆ \ Γ. Also, a simplicial
complex ∆ will be identified with (∆, ∅). A face of (∆, Γ) is an element of ∆ \ Γ. The
dimension of a face τ is its cardinality minus one, and the dimension of (∆, Γ) is the
maximal dimension of its faces. A relative simplicial complex is said to be pure if all its
maximal faces have the same dimension.
e i (∆, Γ) the ith reduced homology group of the pair (∆, Γ) computed
We denote by H
e ∗ (∆, Γ) is the usual relative homology of a pair and
with coefficients in F: when Γ 6= ∅, H
e ∗ (∆, Γ) = H
e ∗ (∆) is the reduced homology of ∆. The Betti numbers of
when Γ = ∅, H
e i (∆, Γ). If ∆ is a simplicial complex on [n] and
(∆, Γ) are defined by β̃i (∆, Γ) := dimF H
τ ∈ ∆ is a face of ∆, then the link of τ in ∆ is
lk∆ (τ ) := {σ ∈ ∆ : τ ∪ σ ∈ ∆, τ ∩ σ = ∅}.
For convenience, we also define lk∆ (τ ) = ∅ if τ 6∈ ∆.
Let ∆ be a simplicial complex on [n]. The Stanley–Reisner ideal of ∆ (in S) is the ideal
I∆ = (xτ : τ ⊆ [n], τ 6∈ ∆) ⊆ S,
Q
where xτ = i∈τ xi . If (∆, Γ) is a relative simplicial complex, then the Stanley–Reisner
module of (∆, Γ) is the S-module
F[∆, Γ] = IΓ /I∆ .
A DUALITY IN BUCHSBAUM RINGS AND TRIANGULATED MANIFOLDS
7
When Γ = ∅, the ring F[∆] = F[∆, ∅] = S/I∆ is called the Stanley–Reisner ring of ∆.
A relative simplicial complex (∆, Γ) is said to be Buchsbaum if F[∆, Γ] is a Buchsbaum
module. The following characterization of Buchsbaum property was given by Schenzel
[22, Theorem 3.2]; a proof for relative simplicial complexes appears in [1, Theorem 1.11].
Theorem 3.1 (Schenzel). A pure relative simplicial complex (∆, Γ) of dimension d is
e i lk∆ (τ ), lkΓ (τ ) = 0 for every non-empty face τ ∈ ∆ \ Γ and
Buchsbaum if and only if, H
all i 6= d − |τ |.
In particular, if ∆ is a homology manifold with boundary, then ∆ and (∆, ∂∆) are Buchsbaum (relative) simplicial complexes. (However, most of Buchsbaum complexes are not
homology manifolds.)
Next, we discuss face numbers of Buchsbaum simplicial complexes. For a relative
simplicial complex (∆, Γ) of dimension d − 1, let fi (∆, Γ) be the number of i-dimensional
faces of (∆, Γ) and let
j
X
j−i d − i
fi−1 (∆, Γ) for j = 0, 1, . . . , d.
(−1)
hj (∆, Γ) =
d−j
i=0
For convenience, we also define hj (∆, Γ) = 0 for j > dim(∆, Γ) + 1. The h-numbers
play a central role in the study of face numbers of Cohen–Macaulay simplicial complexes.
On the other hand, for Buchsbaum simplicial complexes, the following modifications of
h-numbers, called h′ -numbers and h′′ -numbers, behave better than the usual h-numbers.
Recall that a linear system of parameters (or l.s.o.p.) is an h.s.o.p. Θ = θ1 , . . . , θd
consisting of linear forms. Note that when F is infinite, any finitely generated graded
S-module has an l.s.o.p. The following result, established by Schenzel [22, §4] (a proof for
Stanley–Reisner modules appears in [1, Theorem 2.5]), is known as Schenzel’s formula.
Theorem 3.2 (Schenzel). Let (∆, Γ) be a Buchsbaum relative simplicial complex of dimension d − 1 and let Θ be an l.s.o.p. for F[∆, Γ]. Then, for j = 0, 1, . . . , d,
X
j−1
d
dimF F[∆, Γ]/ΘF[∆, Γ] j = hj (∆, Γ) −
(−1)j−i β̃i−1 (∆, Γ).
j i=1
In view of Schenzel’s formula, we define the h′ -numbers of a (d −1)-dimensional relative
simplicial complex (∆, Γ) by
X
j−1
d
′
(−1)j−i β̃i−1 (∆, Γ).
hj (∆, Γ) = hj (∆, Γ) −
j i=1
Furthermore, we define the h′′ -numbers of (∆, Γ) by
(
h′j (∆, Γ) − dj β̃j−1 (∆, Γ), if 0 ≤ j < d,
′′
hj (∆, Γ) =
h′d (∆, Γ),
if j = d.
We are now ready to prove Theorem 1.2.
8
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
Proof of Theorem 1.2. Let M = F[∆, Γ]. Observe that
dimF M/Σ(Θ; M) j = dimF (M/ΘM)j − dimF Σ(Θ; M)/ΘM j .
e i−1 (∆, Γ) for i < d (see [1, Theorem 1.8]), Theorem 2.3(i)
Since Hmi (M) = (Hmi (M))0 ∼
=H
implies
d
β̃j−1(∆, Γ), if 0 ≤ j ≤ d − 1,
j
dimF Σ(Θ; M)/ΘM j =
0,
if j ≥ d.
The desired statement then follows from Theorem 3.2 asserting that dimF (M/ΘM)j =
h′j (∆, Γ) for all j.
It was proved in [19, Theorem 3.4] that the jth graded component of the socle of
F[∆]/ΘF[∆] has dimension at least dj β̃j−1 (∆). This implies that the h′′ -numbers of a
Buchsbaum simplicial complex form the Hilbert function of some quotient of its Stanley–
Reisner ring. A new contribution and the significance of Theorem 1.2 is that it provides
an explicit algebraic interpretation of h′′ -numbers via a submodule Σ(Θ; −).
In the rest of this section we discuss a few algebraic and combinatorial applications of
our results. As was proved by Kalai and, independently, Stanley (see [25, Ch. III, §9] and
[24]), if ∆ ⊇ Γ are Cohen–Macaulay simplicial complexes of the same dimension, then
hi (∆) ≥ hi (Γ) for all i. The interpretation of the h′′ -numbers given in Theorem 1.2 allows
us to prove the following generalization of this fact.
Theorem 3.3. Let ∆ ⊇ Γ be Buchsbaum simplicial complexes of the same dimension.
Then h′′i (∆) ≥ h′′i (Γ) for all i.
Proof. We may assume that ∆ and Γ are simplicial complexes on [n] and that F is infinite.
Let d = dim ∆ + 1. Then there is a common linear system of parameters Θ = θ1 , . . . , θd
for F[∆] and F[Γ]. By the definition of Σ(Θ; −),
Pd
F[∆]/Σ(Θ; F[∆]) = S/ (Θ) +
k=1 (θ1 , . . . , θ̂k , . . . , θd ) + I∆ :S θk
and an analogous formula holds for F[Γ]/Σ(Θ; F[Γ]). Since IΓ ⊇ I∆ , the above formula
implies that F[Γ]/Σ(Θ; F[Γ]) is a quotient ring of F[∆]/Σ(Θ; F[∆]). The desired statement
then follows from Theorem 1.2.
Corollary 3.4. Let ∆ be a Buchsbaum simplicial complex and let τ ∈ ∆ be any non-empty
face. Then h′′i (∆) ≥ hi (lk∆ (τ )) for all i.
Proof. Consider the star of τ , st∆ (τ ) = {σ ∈ ∆ : σ ∪ τ ∈ ∆}. Then st∆ (τ ) is Cohen–
Macaulay (by Theorem 3.1 and Reisner’s criterion) and has the same dimension as ∆,
and so by Theorem 3.3, h′′i (∆) ≥ h′′i (st∆ (τ )) for all i. The result follows since lk∆ (τ ) and
st∆ (τ ) have the same h-numbers (this is because st∆ (τ ) is just a cone over lk∆ (τ ), cf. [25,
Corollary III.9.2]) and since for Cohen–Macaulay simplicial complexes the h′′ -numbers
coincide with the h-numbers.
Let R = S/I be a graded F-algebra. The a-invariant of R is the number
a(R) = − min{k : (ωR )k 6= 0}.
A DUALITY IN BUCHSBAUM RINGS AND TRIANGULATED MANIFOLDS
9
This number is an important invariant in commutative algebra. When R is Cohen–
Macaulay, it is well-known that a(R) = max{i : hi (R) 6= 0} − dim R, where hi (R) is
the ith h-number of R. (See [6, §4.1] for the definition of h-numbers for modules.) The
following result provides a generalization of this fact.
Theorem 3.5. Let R = S/I be a Buchsbaum graded F-algebra of Krull dimension d with
depth(R) ≥ 2 and let Θ = θ1 , . . . , θd be an l.s.o.p. for R. Then
a(R) = max{k : (R/Σ(Θ; R))k 6= 0} − d.
In particular, for any connected Buchsbaum simplicial complex ∆ of dimension d − 1,
a(F[∆]) = max{k : h′′k (∆) 6= 0} − d.
Proof. Let m = max{k : (R/Σ(Θ; R))k 6= 0}. Then by Theorem 1.3,
min{k : (ωR /Σ(Θ; ωR ))k 6= 0} = d − m.
As min{k : (ωR )k 6= 0} = min{k : (ωR /ΘωR )k 6= 0}, the theorem would follow if we prove
that min{k : (ωR /ΘωR )k 6= 0} = d − m. To this end, it is enough to show that
Σ(Θ; ωR )/ΘωR k = 0 for k ≤ d − m − 1.
(4)
Since mHmi (R) = 0 for i < d (see [26, Proposition I.2.1]), we conclude from Theorem 2.3(i) that m(Σ(Θ; R)/ΘR) = 0. Furthermore, since (R/ΘR)k = (Σ(Θ; R)/ΘR)k for
k ≥ m + 1, it follows that for k ≥ m + 2,
Σ(Θ; R)/ΘR k = (R/ΘR)k = m(R/ΘR) k = m Σ(Θ; R)/ΘR k = 0.
The isomorphism
Σ(Θ; R)/ΘR ∼
=
M
Hm|C|(R)(−|C|)
C([d]
established in Theorem 2.3(i) then implies that
Hmi (R)j = 0 for all i ≤ d − 1 and j ≥ m + 2 − i.
Therefore,
Hmd−i+1 (R)∨ (−i)
and so
(5)
k
= 0 for all 2 ≤ i ≤ d − 1 and −(k − i) ≥ m + 2 − (d − i + 1),
Hmd−i+1 (R)∨ (−i)
k
= 0 for all 2 ≤ i ≤ d − 1 and k ≤ d − m − 1.
Finally, since depth(ωR ) ≥ 2, we infer from Theorem 2.3(i) and Lemma 2.2(ii) that
M
M
Hmd−|C|+1(R)∨ (−|C|).
Hm|C| (ωR )(−|C|) ∼
Σ(Θ; ωR )/ΘωR ∼
=
=
C([d], |C|≥2
C([d], |C|≥2
The above isomorphisms and (5) then yield the desired property (4).
10
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
Observe that by Hochster’s formula on local cohomology [6, Theorem 5.3.8], if ∆ is a
(d − 1)-dimensional Buchsbaum simplicial complex, then −a(F[∆]) equals the minimum
cardinality of a face whose link has a non-vanishing top homology. Thus the “in particular”
part of Theorem 3.5 can be equivalently restated as follows: if for every face τ ∈ ∆ of
dimension < d−k, the link of τ has vanishing top homology, then h′′k (∆) = 0. For the case
of homology manifolds with boundary, faces with vanishing top homology are precisely
the boundary faces; in this case, the above statement reduces to [14, Theorem 3.1].
A sequence h0 , h1 , . . . , hm of numbers is said to be unimodal if there is an index p such
that h0 ≤ h1 ≤ · · · ≤ hp ≥ · · · ≥ hm . It was proved in [13] that the h′′ -numbers of the
barycentric subdivision of any connected Buchsbaum simplicial complex form a unimodal
sequence. Here we use the Matlis duality established in Theorem 1.3 to generalize this
result to Buchsbaum polyhedral complexes. (We refer our readers to [13] for the definition
of barycentric subdivisions.)
Theorem 3.6. Let ∆ be the barycentric subdivision of a connected polyhedral complex Γ of
dimension d − 1. Suppose that the characteristic of F is zero and ∆ is Buchsbaum. Then,
for a generic choice of linear forms Θ = θ1 , . . . , θd , and θd+1 ∈ F[∆], the multiplication
×θd+1 : F[∆]/Σ(Θ; F[∆]) i → F[∆]/Σ(Θ; F[∆]) i+1
is injective for i ≤ d2 − 1 and is surjective for i ≥
h′′0 (∆), h′′1 (∆), . . . , h′′d (∆) is unimodal.
d
.
2
In particular, the sequence
Proof. By genericity of linear forms, Θ = θ1 , . . . , θd is a common l.s.o.p. for F[∆] and
ωF[∆] . Let P be the face poset of Γ. Then the Stanley–Reisner ring F[∆] is a squarefree
P -module (a notion introduced in [16, Definition 2.1]); furthermore, by [16, Theorem 3.1]
ωF[∆] is also a squarefree P -module. It then follows from [16, Theorem 6.2 (ii)] that the
multiplication maps
(6)
× θd+1 : F[∆]/ΘF[∆] i → F[∆]/ΘF[∆] i+1
and
×θd+1 : ωF[∆] /(ΘωF[∆])
i
→ ωF[∆] /(ΘωΘF[∆])
i+1
.
are surjective for i ≥ d2 . (While Cohen–Macaulayness was assumed in [16, Theorem 6.2],
one can see from the proof given in [16] that this assumption was used only in the proof of
part (i) and is unnecessary to derive surjectivity.) Since ΘF[∆] is contained in Σ(Θ; F[∆]),
the map in (6) remains surjective if we replace ΘF[∆] with Σ(Θ; F[∆]), and a similar
statement holds for ωF[∆]. These surjectivities and the Matlis duality F[∆]/Σ(Θ; F[∆]) ∼
=
∨
(ωF[∆] /Σ(Θ; ωF[∆])) (d) of Theorem 1.3 yield the desired statement.
In fact, in view of results from [10], it is tempting to conjecture that if ∆ is a barycentric
subdivision of a Buchsbaum
of dimension
d − 1, then even the
regular CW-complex
sequence h′′0 (∆)/ d0 , h′′1 (∆)/ d1 , h′′2 (∆)/ d2 , . . . , h′′d (∆)/ dd is unimodal.
We close this section with a couple of remarks.
A DUALITY IN BUCHSBAUM RINGS AND TRIANGULATED MANIFOLDS
11
Remark 3.7. Our results on h′′ -vectors can be generalized to the following setting. Consider a finitely generated graded S-module M of Krull dimension d such that
(7)
Hmi (M) = Hmi (M) 0 for all 0 ≤ i < d.
Define the h′ - and h′′ -numbers of M in the same way as for the Stanley–Reisner modules
but with dimF (Hmj (M)) used as a replacement for β̃j−1(∆, Γ). Then suitably modified
statements of Theorems 1.2 and 3.2 continue to hold for such an M. Furthermore, if
M satisfies (7), then M must be Buchsbaum (see [26, Proposition I.3.10]), in which case
ωM also satisfies (7) by Lemma 2.2(ii). In particular, if ∆ is an arbitrary connected
Buchsbaum simplicial complex of dimension d − 1, then
h′′i (ωF[∆] ) = h′′d−i (F[∆]) for all i = 0, 1, . . . , d.
Remark 3.8. A statement analogous to Corollary 1.4 also holds for homology manifolds
without boundary. Indeed, if ∆ is an orientable homology manifold without boundary,
then by Gräbe’s result [8], F[∆] is isomorphic to its own canonical module, and hence, by
Theorem 1.3, F[∆]/Σ(Θ; F[∆]) is an Artinian Gorenstein algebra. This fact was essentially
proved in [18, Theorem 1.4].
Let us also point out that if ∆ is a connected orientable homology manifold with or
without boundary, then for M = F[∆], the statement of part (ii) of Lemma 2.2 is a simple
consequence of the Poincaré–Alexander–Lefschetz duality along with Gräbe’s result [8]
that ωM ∼
= F[∆, ∂∆].
4. Applications to the manifold g-conjecture
In this section we discuss connected orientable homology manifolds without boundary.
One of the most important open problems in algebraic combinatorics is the algebraic gconjecture; it asserts that every homology sphere has the weak Lefschetz property (the
WLP, for short). Kalai proposed a far-reaching generalization of this conjecture [17,
Conjecture 7.5] to homology manifolds. Using the Σ(Θ, −) module allows to restate
Kalai’s conjecture as follows.
Conjecture 4.1. Let ∆ be a connected orientable F-homology (d − 1)-manifold without boundary. Then, for a generic choice of linear forms Θ = θ1 , . . . , θd , the ring
F[∆]/Σ(Θ; F[∆]) has the WLP, that is, for a generic linear form ω, the multiplication
×ω : F[∆]/Σ(Θ; F[∆]) ⌊d/2⌋ → F[∆]/Σ(Θ; F[∆]) ⌊d/2⌋+1
is surjective.
It is worth mentioning that while Kalai’s original statement of the conjecture did not
involve Σ(Θ, −), the two statements are equivalent (see [18] and Remark 3.8 above).
Somewhat informally, we say that ∆ (or F[∆]) has the WLP if ∆ satisfies the conclusions
of the above conjecture. Enumerative consequences of this conjecture are discussed in [18,
§1].
Given a finite set A, we denote by A the simplex on A, i.e., the simplicial complex
whose set of faces consists of all subsets of A. When A = {a} consists of a single vertex,
12
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
we write a to denote the vertex a, viewed as a 0-dimensional simplex. If ∆ and Γ are
two simplicial complexes on disjoint vertex sets, then the join of ∆ and Γ, ∆ ∗ Γ, is the
simplicial complex defined by
∆ ∗ Γ := {σ ∪ τ : σ ∈ ∆, τ ∈ Γ}.
Finally, if ∆ is a simplicial complex and W is a set, then ∆W = {σ ∈ ∆ : σ ⊆ W } is the
subcomplex of ∆ induced by W .
Let ∆ be a (d − 1)-dimensional homology manifold, and let A and B be disjoint subsets
such that |A| + |B| = d + 1. If ∆A∪B = A ∗ ∂B, then the operation of removing A ∗ ∂B
from ∆ and replacing it with ∂A ∗ B is called a (|B| − 1)-bistellar flip. (For instance,
a 0-flip is simply a stellar subdivision at a facet.) The resulting complex is a homology
manifold homeomorphic to the original complex.
The following surprising property was proved by Pachner: if ∆1 and ∆2 are two PL
homeomorphic combinatorial manifolds without boundary, then they can be connected
by a sequence of bistellar flips (see [20, 21]). Since the boundary complex of a simplex
has the WLP, Pachner’s result suggests the following inductive approach to the algebraic
g-conjecture for PL spheres: prove that bistellar flips applied to PL spheres preserve
the WLP. In [28, §3], Swartz showed that most of bistellar flips applied to homology
spheres preserve the WLP. Here we extend Swartz’s results to the generality of orientable
homology manifolds without boundary:
Theorem 4.2. Let ∆ be a (d − 1)-dimensional, connected, orientable homology manifold
without boundary. Suppose that ∆′ is obtained from ∆ via a (p − 1)-bistellar flip with
p 6= (d + 1)/2 if d is odd and with p ∈
/ {d/2, (d + 2)/2} if d is even, and let Θ be a
common l.s.o.p. for F[∆] and F[∆′ ]. Then F[∆′ ]/Σ(Θ; F[∆′ ]) has the WLP if and only if
F[∆]/Σ(Θ; F[∆]) has the WLP.
If ∆ is a simplicial complex and σ is a face of ∆, then the stellar subdivision of ∆ at σ
consists of (i) removing σ and all faces containing it from ∆, (ii) introducing a new vertex
a, and (iii) adding new faces in a ∗ ∂σ ∗ lk∆ (σ) to ∆:
sdσ (∆) := ∆ \ st∆ (σ) ∪ a ∗ ∂σ ∗ lk∆ (σ) .
A classical result due to Alexander [2] asserts that two simplicial complexes are PL homeomorphic if and only if they are stellar equivalent, that is, one of them can be obtained
from another by a sequence of stellar subdivisions and their inverses. Thus, a different approach to the algebraic g-conjecture (at least for PL manifolds) is to show that the WLP
is preserved by stellar subdivisions and their inverses. Böhm and Papadakis [5, Corollary
4.5] proved that this is the case if one applies stellar subdivisions at faces of sufficiently
large dimension (or their inverses) to homology spheres. We extend their result to the
generality of orientable homology manifolds without boundary:
Theorem 4.3. Let ∆ be a (d−1)-dimensional, connected, orientable F-homology manifold
without boundary and let σ be a face of ∆ with dim σ > d/2. Then F[∆] has the WLP if
and only if F[sdσ (∆)] has the WLP.
A DUALITY IN BUCHSBAUM RINGS AND TRIANGULATED MANIFOLDS
13
The structure of the proofs of both theorems is similar to the proof of [28, Theorem
3.1]. The new key ingredient is given by the following lemma. Recall that ∆ is an Fhomology (d − 1)-sphere if ∆ is an F-homology manifold whose homology over F coincides
with that of Sd−1 . Similarly, ∆ is an F-homology ball if (i) ∆ is an F-homology manifold
with boundary, (ii) the homology of ∆ (over F) vanishes, and (iii) the boundary of ∆ is
an F-homology sphere.
Lemma 4.4. Let ∆ be a (d − 1)-dimensional F-homology manifold without boundary, Γ
a full-dimensional subcomplex of ∆, and Θ an l.s.o.p. for F[∆]. Assume further that Γ
is an F-homology ball, and let D be the simplicial complex obtained from ∆ by removing
the interior faces of Γ. Then D is an F-homology manifold with boundary and the natural
surjection F[∆] → F[Γ] induces the following short exact sequence
0 → F[D, ∂D]/Σ(Θ; F[D, ∂D]) → F[∆]/Σ(Θ; F[∆]) → F[Γ]/ ΘF[Γ] → 0.
Proof. The fact that D is a homology manifold with boundary follows by a standard
Mayer-Vietoris argument (note that ∂D = ∂Γ is a homology (d−2)-sphere). Furthermore,
by excision and since Γ has vanishing homology,
(8)
β̃i (D, ∂D) = β̃i (∆, Γ) = β̃i (∆) for all i.
Now, using that the homology ball Γ is a full-dimensional subcomplex of ∆, we conclude
as in the proof of Theorem 3.3 that there is a natural surjection
(9)
φ : F[∆]/Σ Θ; F[∆] → F[Γ]/ ΘF[Γ] = F[Γ]/Σ Θ; F[Γ] .
Thus, to finish the proof
of the lemma,
F[D, ∂D]/Σ Θ; F[D, ∂D] . This, in turn,
facts:
(i) Ker(φ) is a quotient of F[D, ∂D]/Σ
(ii) dimF (Ker(φ))j = dimF F[D, ∂D]/Σ
it suffices to show that the kernel of φ is
would follow if we verify the following two
Θ; F[D, ∂D]
, and
Θ; F[D, ∂D] j for all j.
For brevity, let R = F[∆] and let I = F[D, ∂D]. Since (D, ∂D) = (∆, Γ) as relative
simplicial complexes, it follows that I is an ideal of R, and R/I is F[Γ]. Thus our surjection
φ is the projection φ : R/Σ(Θ; R) → R/(I + ΘR). Hence
Ker(φ) = (I + ΘR)/Σ(Θ; R) = (I + Σ(Θ; R))/Σ(Θ; R) = I/(I ∩ Σ(Θ; R)),
which together with an observation that I ∩ Σ(Θ; R) contains Σ(Θ; I) yields assertion (i).
To see that I ∩ Σ(Θ; R) ⊇ Σ(Θ; I), note that since I is a subset of R, ΘI is contained
in ΘR, and (θ1 , ..., θ̂i , . . . , θd )I :I θi is contained in (θ1 , ..., θ̂i , . . . , θd )R :R θi for all i ∈ [d];
therefore Σ(Θ; R) ⊇ Σ(Θ; I).
As for assertion (ii), the dimension of the kernel of φ can be computed as follows: by
(9) and by Theorem 1.2,
(10)
dimF (Ker(φ))j = h′′j (∆) − hj (Γ)
Now, since each face of ∆ is either a face of (D, ∂D) or a face of Γ (but not of both),
fi (∆) = fi (D, ∂D) + fi (Γ) for all i ≥ −1, and so
(11)
hj (D, ∂D) = hj (∆) − hj (Γ) for all j ≥ 0.
14
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
Equations (8), (10) and (11) yield
dimF (Ker(φ))j = h′′j (D, ∂D) = dimF F[D, ∂D]/Σ Θ; F[D, ∂D]
j
for all j ≥ 0,
where the last equality is another application of Theorem 1.2. The assertion follows.
Another ingredient needed for both proofs is the following immediate consequence of
the snake lemma.
Lemma 4.5. Let 0 → L → N → M → 0 be an exact sequence of graded S-modules,
let ω ∈ S be a linear form, and let k be a fixed integer. Assume also that the map
×ω : Mk → Mk+1 is bijective. Then the map ×ω : Nk → Nk+1 is surjective if and only if
the map ×ω : Lk → Lk+1 is surjective.
We are now ready to prove both of the theorems.
Proof of Theorem 4.2. We are given that ∆′ = ∆ \ (A ∗ ∂B) ∪ ∂A ∗ B , where |B| = p.
Let D be the simplicial complex obtained from ∆ by removing the interior faces of the
ball Γ1 = A ∗ ∂B; equivalently, D is obtained from ∆′ by removing the interior faces of
Γ2 = ∂A ∗ B. Since Γ1 and Γ2 are full-dimensional subcomplexes of ∆ and ∆′ , and Θ is
an l.s.o.p. for both F[∆] and F[∆′ ], it is also an l.s.o.p. for both F[Γ1 ] and F[Γ2 ]. Hence
1, if j ≤ p − 1,
dimF (F[Γ1 ]/ΘF[Γ1 ])j = hj (A ∗ ∂B) = hj (∂B) =
0, if j > p − 1,
which implies that F[Γ1 ]/ΘF[Γ1 ] ∼
= F[x]/(xd−p+1 ).
= F[x]/(xp ). Similarly, F[Γ2 ]/ΘF[Γ2 ] ∼
d d+1 d+2
These isomorphisms together with the assumption p 6∈ { 2 , 2 , 2 } yield that, for a
generic choice of a linear form w, the multiplication map
×w : (F[Γi ]/ΘF[Γi ])⌊ d ⌋ → (F[Γi ]/ΘF[Γi ])⌊ d ⌋+1
2
2
is bijective for i ∈ {1, 2}.
Applying Lemma 4.4 to ∆ and Γ1 , we conclude from Lemma 4.5 that F[∆]/Σ(Θ; F[∆])
has the WLP if and only if, for a generic linear form w, the multiplication
×w : F[D, ∂D]/Σ(Θ; F[D, ∂D]) ⌊ d ⌋ → F[D, ∂D]/Σ(Θ; F[D, ∂D]) ⌊ d ⌋+1
2
2
′
is surjective. On the other hand, by Lemma 4.4 applied to ∆ and Γ2 , the latter condition
is equivalent to the WLP of F[∆′ ]/Σ(Θ; F[∆′ ]).
Proof of Theorem 4.3. Let D be the homology manifold obtained from ∆ by removing the
interior faces of the homology ball Γ1 = st∆ (σ); equivalently, D is obtained from sdσ (∆)
by removing the interior faces of the homology ball Γ2 = a ∗ ∂σ ∗ lk∆ (σ). As the proof of
Theorem 4.2 shows, to verify Theorem 4.3, it suffices to check that, for a generic choice
of linear forms Θ = θ1 , . . . , θd and another generic linear form w, the multiplication
(12)
×w : F[Γi ]/ΘF[Γi ] ⌊ d ⌋ → F[Γi ]/ΘF[Γi ] ⌊ d ⌋+1
2
2
is bijective for each i ∈ {1, 2}.
In the case of i = 1, the desired bijection is an immediate consequence of the fact that
dimF (F[Γ1 ]/ΘF[Γ1 ])j = hj (st∆ (σ)) = hj (lk∆ (σ)) = 0 for j ≥ d − |σ| + 1
A DUALITY IN BUCHSBAUM RINGS AND TRIANGULATED MANIFOLDS
15
and the assumption |σ| > d2 + 1. We now prove that the map in (12) is also bijective for
i = 2. To this end, observe that the homology (d − 2)-sphere ∂σ ∗ lk∆ (σ) is the boundary
of the homology (d − 1)-ball σ ∗ lk∆ (σ), and hence ∂σ ∗ lk∆ (σ) is (d − |σ|)-stacked. (A
homology (d − 2)-sphere is called (d − k)-stacked if it is the boundary of a homology ball
that has no interior faces of size ≤ k − 1.) It then follows from [28, Corollary 6.3] that
∂σ ∗ lk∆ (σ) has the WLP. This, in turn, implies that the map in (12) is surjective as F[Γ2 ]
is a polynomial ring over F[∂σ∗lk∆ (σ)]. Also, since |σ| > d2 +1, we infer from [12, Theorem
2] and the (d−|σ|)-stackedness of ∂σ ∗lk∆ (σ) that h⌊ d ⌋ (∂σ ∗lk∆ (σ)) = h⌊ d ⌋+1 (∂σ ∗lk∆ (σ)).
2
2
As Γ2 and ∂σ ∗ lk∆ (σ) have the same h-vector, we conclude that h⌊ d ⌋ (Γ2 ) = h⌊ d ⌋+1 (Γ2 ),
2
2
and therefore that the map in (12) is bijective.
Let ∆ be a triangulation of a closed surface and assume that one of the vertices of
∆ is connected to all other vertices of ∆. Then by [18, Theorem 1.6], ∆ has the WLP.
However, the problem of whether every triangulation of a closed surface other than the
sphere possesses the WLP is at present wide open.
Remark 4.6. Note that if ∆ is an arbitrary odd-dimensional Buchsbaum complex (e.g., a
homology manifold), then there exists a simplicial complex Γ that is PL homeomorphic to
∆ and has the WLP in characteristic 0. Simply take Γ to be the barycentric subdivision
of ∆: the WLP of Γ is guaranteed by Theorem 3.6.
Appendix. Proof of Theorem 2.3
The goal of this Appendix is to verify Theorem 2.3. Our proof is based on the proof
of [19, Theorem 2.2], and so some details are omitted. Let M be a finitely generated
Buchsbaum graded S-module of Krull dimension
d, and let Θ = θ1 , . . . , θd be an h.s.o.p.
P
for M with deg θi = δi . We write δC = i∈C δi for C ⊆ [d]. We need the following result
(see [26, Lemma II.4.14′ ]).
Lemma A.1. There is an isomorphism
M
Hm|C|+i (M)(−δC )
Hmi (M/((θ1 , . . . , θj )M)) ∼
=
for all i, j with i + j < d.
C⊆[j]
We use the following notation: if C ⊆ [d], then MhCi is defined by
MhCi = M/ (θi : i 6∈ C)M .
By [26, Proposition I.2.1], if C ( [d] and s ∈ [d]\C, then Hm0 (MhC ∪{s}i) = 0 :M hC∪{s}i θs .
This leads to the following short exact sequence
(13)
×θ
π
s
s
0 −→ MhC ∪ {s}i/Hm0 (MhC ∪ {s}i)(−δs ) −→
MhC ∪ {s}i −→
MhCi −→ 0,
where πs is a natural projection. This short exact sequence gives rise to exact sequences
(14)
π∗
ϕ∗
s
s
0 −→ Hmk (MhC ∪ {s}i) −→
Hmk (MhCi) −→
Hmk+1(MhC ∪ {s}i)(−δs ) −→ 0
16
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
for k < |C|, and
ϕ∗
π∗
s
s
Hm|C|+1(MhC ∪ {s}i)(−δs ),
Hm|C| (MhCi) −→
0 −→ Hm|C| (MhC ∪ {s}i) −→
(15)
where we denote by ϕ∗s the connecting homomorphism.
Similarly, if C ( [d], |C| ≤ d−2, and s, t ∈ [d]\C, we obtain the following commutative
diagram
0 →
M hC ∪ {s}i/Hm0 (M hC ∪ {s}i)(−δs )
↑ πt
0 → M hC ∪ {s, t}i/Hm0 (M hC ∪ {s, t}i)(−δs )
×θs
→
×θs
→
M hC ∪ {s}i
↑ πt
π
→s
M hCi
↑ πt
π
M hC ∪ {s, t}i →s M hC ∪ {t}i → 0.
This diagram, in turn, induces the commutative diagram
(16)
ϕ∗
Hmi (MhCi)
↑ πt∗
Hmi+1 (MhC ∪ {s}i)(−δs )
↑ πt∗
s
−→
ϕ∗
s
Hmi (MhC ∪ {t}i) −→
Hmi+1 (MhC ∪ {s, t}i)(−δs ).
Now, for k = 2, 3, . . . , d, define the maps φk and ψk as compositions
ϕ∗k−1
ϕ∗
ϕ∗
φk : Hm0 (Mh∅i) →1 Hm1 M [1] (−δ[1] ) →2 · · · → Hmk−1 M [k − 1] (−δ[k−1] )
and
ϕ∗k−1
ϕ∗
ϕ∗
ψk : Hm0 (M h{k}i) →1 Hm1 M [1] ∪ {k} (−δ[1] ) →2 · · · → Hmk−1 M [k] (−δ[k−1] ).
Then the commutativity of (16) implies the commutativity of
φk
Hm0 (Mh∅i) −→
Hmk−1 M [k − 1] (−δ[k−1] )
↑ πk∗
(17)
ψ
k
Hm0 (Mh{k}i) −→
Hmk−1
↑ πk∗
M [k] (−δ[k−1] ).
To prove part (ii) of Theorem 2.3, we consider the following diagram
Hm0 (Mh{1}i)
ψ2
Hm0 Mh{2}i −→
Hm1 M [2] (−δ[1] )
→0
π∗
1
−→
Hm0 (Mh∅i)
↓ ϕ∗1
π2∗
−→
Hm1 M [1] (−δ[1] )
↓ ϕ∗2
..
.
↓ ϕ∗d−1
ψd
πd∗
Hmd−1 M [d − 1] (−δ[d−1] )
Hm0 Mh{d}i −→
Hmd−1 M [d] (−δ[d−1] ) −→
↓ ϕ∗d
Hmd M [d] (−δ[d] ).
A DUALITY IN BUCHSBAUM RINGS AND TRIANGULATED MANIFOLDS
17
The surjectivity of ϕ∗s in (14) implies that ψk is surjective. Hence in each horizontal line
of the diagram,
Im(πk∗ ◦ ψk ) = Im(πk∗ ).
(18)
Also, since the sequence in (15) is exact, it follows that in the diagram,
Ker(ϕ∗k ) = Im(πk∗ ).
(19)
Define φd+1 = ϕ∗d ◦ · · · ◦ ϕ∗1 to be the composition of the vertical maps in the diagram.
By (15), each πk∗ (in the diagram) is injective, and we conclude that
Ker(φd+1 ) ∼
=
(20)
d
M
Im(πk∗ )
∼
=
d
M
Hmk−1 M [k]
k=1
k=1
(−δ[k−1] )
as F-vector spaces. In addition, using (18) and the commutativity of (17), we obtain that
Ker(φd+1 ) is the sum of the images of
πk∗ : Hm0 (Mh{k}i) → Hm0 (Mh∅i) = M/ΘM.
Finally, since
Hm0 (Mh{k}i) = (θ1 , . . . , θ̂k , . . . , θd )M :M θk /(θ1 , . . . , θ̂k , . . . , θd )M
and since πk∗ is a natural projection, it follows that
Ker(φd+1 ) = Σ(Θ; M)/ΘM.
This proves part (ii) of the statement since φd+1 is a map from M/ΘM to Hmd (M)(−δ[d] ).
It remains to verify part (i). Recall that Ker(φd+1 ) is the sum of images of Hm0 (Mh{k}i)
and that by [26, Proposition I.2.1], m · Hm0 (Mh{k}i) = 0. Hence the modules in (20) are
in fact isomorphic as S-modules (since they are direct sums of copies of F). Thus we infer
from (20) and Lemma A.1 that
Σ(Θ; M)/ΘM = Ker(φd+1 ) ∼
=
d
M
k=1
∼
=
d
M
k=1
∼
=
M
Hmk−1 M [k]
M
C⊆[d]\[k]
(−δ[k−1] )
Hm|C|+k−1(M)(−δC∪[k−1] )
Hm|C| (M)(−δC ),
C([d]
as desired.
Remark. As the proof in this Appendix is based on the proof of [19, Theorem 2.2], it
is worth pointing out that there is a minor mistake in [19]. Indeed, the short “exact”
sequence that appears three lines after the statement of Theorem 2.4 in [19] is not necessarily exact. However, this mistake can be easily corrected by replacing this sequence
with the short exact sequence of equation (13).
18
SATOSHI MURAI, ISABELLA NOVIK, AND KEN-ICHI YOSHIDA
Acknowledgments: We thank Ed Swartz for raising a question of whether there is an
algebraic version of Theorem 1.1 and for bringing the result of Böhm and Papadakis to
our attention. We are also grateful to Shiro Goto for helpful conversations.
References
[1] Karim A. Adiprasito and Raman Sanyal. Relative Stanley-Rreisner theory and upper bound theorems
for Minkowski sums. arXiv1405.7368, 2014.
[2] James W. Alexander. The combinatorial theory of complexes. Ann. Math., 31:292–320, 1930.
[3] Yoichi Aoyama. On the depth and the projective dimension of the canonical module. Japan. J. Math.
(N.S.), 6(1):61–66, 1980.
[4] Yoichi Aoyama and Shiro Goto. Some special cases of a conjecture of Sharp. J. Math. Kyoto Univ.,
26(4):613–634, 1986.
[5] Janko Böhm and Stavros Argyrios Papadakis. Weak Lefschetz property and stellar subdivisions of
Gorenstein complexes. arXiv.1501.01513, 2015.
[6] Winfried Bruns and Jürgen Herzog. Cohen-Macaulay rings, volume 39 of Cambridge Studies in
Advanced Mathematics. Cambridge University Press, Cambridge, 1993.
[7] Shiro Goto. On the associated graded rings of parameter ideals in Buchsbaum rings. J. Algebra,
85(2):490–534, 1983.
[8] Hans-Gert Gräbe. Über den Stanley-Reisner-Ring von Quasimannigfaltigkeiten. Math. Nachr.,
117:161–174, 1984.
[9] Hans-Gert Gräbe. Generalized Dehn-Sommerville equations and an upper bound theorem. Beiträge
Algebra Geom., (25):47–60, 1987.
[10] Martina Juhnke-Kubitzke and Satoshi Murai. Balanced generalized lower bound inequality for simplicial polytopes. arXiv1503.06430, 2015.
[11] Steven Klee and Isabella Novik. Face enumeration on simplicial complexes. arXiv:1505.06380, 2015.
[12] Peter McMullen and David W. Walkup. A generalized lower-bound conjecture for simplicial polytopes. Mathematika, 18:264–273, 1971.
[13] Satoshi Murai. On face vectors of barycentric subdivisions of manifolds. SIAM J. Discrete Math.,
24(3):1019–1037, 2010.
[14] Satoshi Murai and Eran Nevo. On r-stacked triangulated manifolds. J. Algebraic Combin., 39(2):373–
388, 2014.
[15] Satoshi Murai and Isabella Novik. Face numbers of manifolds with boundary. arXiv:1509.05115,
2015.
[16] Satoshi Murai and Kohji Yanagawa. Squarefree P -modules and the cd-index. Adv. Math., 265:241–
279, 2014.
[17] Isabella Novik. Upper bound theorems for homology manifolds. Israel J. Math., 108:45–82, 1998.
[18] Isabella Novik and Ed Swartz. Gorenstein rings through face rings of manifolds. Compos. Math.,
145(4):993–1000, 2009.
[19] Isabella Novik and Ed Swartz. Socles of Buchsbaum modules, complexes and posets. Adv. Math.,
222(6):2059–2084, 2009.
[20] Udo Pachner. Konstruktionsmethoden und das kombinatorische Homöomorphieproblem für Triangulationen kompakter semilinearer Mannigfaltigkeiten. Abh. Math. Sem. Univ. Hamburg, 57:69–86,
1987.
[21] Udo Pachner. P.L. homeomorphic manifolds are equivalent by elementary shellings. European J.
Combin., 12(2):129–145, 1991.
[22] Peter Schenzel. On the number of faces of simplicial complexes and the purity of Frobenius. Math.
Z., 178(1):125–142, 1981.
[23] Peter Schenzel. Dualisierende Komplexe in der lokalen Algebra und Buchsbaum-Ringe, volume 907
of Lecture Notes in Mathematics. Springer-Verlag, Berlin-New York, 1982.
A DUALITY IN BUCHSBAUM RINGS AND TRIANGULATED MANIFOLDS
19
[24] Richard P. Stanley. A monotonicity property of h-vectors and h∗ -vectors. European J. Combin.,
14(3):251–258, 1993.
[25] Richard P. Stanley. Combinatorics and commutative algebra, volume 41 of Progress in Mathematics.
Birkhäuser Boston, Inc., Boston, MA, second edition, 1996.
[26] Jürgen Stückrad and Wolfgang Vogel. Buchsbaum rings and applications. Springer-Verlag, Berlin,
1986.
[27] Naoyoshi Suzuki. The Koszul complex of Buchsbaum modules. RIMS Kokyuroku, 446:15–25, 1981.
[28] Ed Swartz. Thirty-five years and counting. arXiv.1411.0987, 2014.
Satoshi Murai, Department of Pure and Applied Mathematics, Graduate School of
Information Science and Technology, Osaka University, Suita, Osaka, 565-0871, Japan
E-mail address: [email protected]
Isabella Novik, Department of Mathematics, University of Washington, Seattle, WA
98195-4350, USA.
E-mail address: [email protected]
Ken-ichi Yoshida, Department of Mathematics, College of Humanities and Sciences,
Nihon University, Setagaya-ku, Tokyo, 156-8550, Japan
E-mail address: [email protected]
| 0math.AC
|
Published as a conference paper at ICLR 2016
S PARK N ET: T RAINING D EEP N ETWORKS IN S PARK
arXiv:1511.06051v4 [stat.ML] 28 Feb 2016
Philipp Moritz∗, Robert Nishihara∗, Ion Stoica, Michael I. Jordan
Electrical Engineering and Computer Science
University of California
Berkeley, CA 94720, USA
{pcmoritz,rkn,istoica,jordan}@eecs.berkeley.edu
A BSTRACT
Training deep networks is a time-consuming process, with networks for object recognition often requiring multiple days to train. For this reason, leveraging the resources of a cluster to speed up training is an important area of
work. However, widely-popular batch-processing computational frameworks
like MapReduce and Spark were not designed to support the asynchronous and
communication-intensive workloads of existing distributed deep learning systems.
We introduce SparkNet, a framework for training deep networks in Spark. Our implementation includes a convenient interface for reading data from Spark RDDs,
a Scala interface to the Caffe deep learning framework, and a lightweight multidimensional tensor library. Using a simple parallelization scheme for stochastic
gradient descent, SparkNet scales well with the cluster size and tolerates very
high-latency communication. Furthermore, it is easy to deploy and use with no
parameter tuning, and it is compatible with existing Caffe models. We quantify
the dependence of the speedup obtained by SparkNet on the number of machines,
the communication frequency, and the cluster’s communication overhead, and we
benchmark our system’s performance on the ImageNet dataset.
1
I NTRODUCTION
Deep learning has advanced the state of the art in a number of application domains. Many of the
recent advances involve fitting large models (often several hundreds megabytes) to larger datasets
(often hundreds of gigabytes). Given the scale of these optimization problems, training can be timeconsuming, often requiring multiple days on a single GPU using stochastic gradient descent (SGD).
For this reason, much effort has been devoted to leveraging the computational resources of a cluster
to speed up the training of deep networks (and more generally to perform distributed optimization).
Many attempts to speed up the training of deep networks rely on asynchronous, lock-free optimization (Dean et al., 2012; Chilimbi et al., 2014). This paradigm uses the parameter server model (Li
et al., 2014; Ho et al., 2013), in which one or more master nodes hold the latest model parameters
in memory and serve them to worker nodes upon request. The nodes then compute gradients with
respect to these parameters on a minibatch drawn from the local data shard. These gradients are
shipped back to the server, which updates the model parameters.
At the same time, batch-processing frameworks enjoy widespread usage and have been gaining in
popularity. Beginning with MapReduce (Dean & Ghemawat, 2008), a number of frameworks for
distributed computing have emerged to make it easier to write distributed programs that leverage the
resources of a cluster (Zaharia et al., 2010; Isard et al., 2007; Murray et al., 2013). These frameworks
have greatly simplified many large-scale data analytics tasks. However, state-of-the-art deep learning
systems rely on custom implementations to facilitate their asynchronous, communication-intensive
workloads. One reason is that popular batch-processing frameworks (Dean & Ghemawat, 2008;
Zaharia et al., 2010) are not designed to support the workloads of existing deep learning systems.
SparkNet implements a scalable, distributed algorithm for training deep networks that lends itself to
batch computational frameworks such as MapReduce and Spark and works well out-of-the-box in
bandwidth-limited environments.
∗
Both authors contributed equally.
1
Published as a conference paper at ICLR 2016
Figure 1: This figure depicts the SparkNet architecture.
The benefits of integrating model training with existing batch frameworks are numerous. Much of
the difficulty of applying machine learning has to do with obtaining, cleaning, and processing data as
well as deploying models and serving predictions. For this reason, it is convenient to integrate model
training with the existing data-processing pipelines that have been engineered in today’s distributed
computational environments. Furthermore, this approach allows data to be kept in memory from
start to finish, whereas a segmented approach requires writing to disk between operations. If a
user wishes to train a deep network on the output of a SQL query or on the output of a graph
computation and to feed the resulting predictions into a distributed visualization tool, this can be
done conveniently within a single computational framework.
We emphasize that the hardware requirements of our approach are minimal. Whereas many approaches to the distributed training of deep networks involve heavy communication (often communicating multiple gradient vectors for every minibatch), our approach gracefully handles the
bandwidth-limited setting while also taking advantage of clusters with low-latency communication.
For this reason, we can easily deploy our algorithm on clusters that are not optimized for communication. Our implementation works well out-of-the box on a five-node EC2 cluster in which
broadcasting and collecting model parameters (several hundred megabytes per worker) takes on the
order of 20 seconds, and performing a single minibatch gradient computation requires about 2 seconds (for AlexNet). We achieve this by providing a simple algorithm for parallelizing SGD that
involves minimal communication and lends itself to straightforward implementation in batch computational frameworks. Our goal is not to outperform custom computational frameworks but rather
to propose a system that can be easily implemented in popular batch frameworks and that performs
nearly as well as what can be accomplished with specialized frameworks.
2
I MPLEMENTATION
Here we describe our implementation of SparkNet. SparkNet builds on Apache Spark (Zaharia et al.,
2010) and the Caffe deep learning library (Jia et al., 2014). In addition, we use Java Native Access
class Net {
def Net(netParams: NetParams): Net
def setTrainingData(data: Iterator[(NDArray,Int)])
def setValidationData(data: Iterator[(NDArray,Int)])
def train(numSteps: Int)
def test(numSteps: Int): Float
def setWeights(weights: WeightCollection)
def getWeights(): WeightCollection
}
Listing 1: SparkNet API
2
Published as a conference paper at ICLR 2016
val netParams = NetParams(
RDDLayer("data", shape=List(batchsize, 1, 28, 28)),
RDDLayer("label", shape=List(batchsize, 1)),
ConvLayer("conv1", List("data"), kernel=(5,5), numFilters=20),
PoolLayer("pool1", List("conv1"), pool=Max, kernel=(2,2), stride=(2,2)),
ConvLayer("conv2", List("pool1"), kernel=(5,5), numFilters=50),
PoolLayer("pool2", List("conv2"), pool=Max, kernel=(2,2), stride=(2,2)),
LinearLayer("ip1", List("pool2"), numOutputs=500),
ActivationLayer("relu1", List("ip1"), activation=ReLU),
LinearLayer("ip2", List("relu1"), numOutputs=10),
SoftmaxWithLoss("loss", List("ip2", "label"))
)
Listing 2: Example network specification in SparkNet
var trainData = loadData(...)
var trainData = preprocess(trainData).cache()
var nets = trainData.foreachPartition(data => {
var net = Net(netParams)
net.setTrainingData(data)
net)
var weights = initialWeights(...)
for (i <- 1 to 1000) {
var broadcastWeights = broadcast(weights)
nets.map(net => net.setWeights(broadcastWeights.value))
weights = nets.map(net => {
net.train(50)
net.getWeights()}).mean() // an average of WeightCollection objects
}
Listing 3: Distributed training example
for accessing Caffe data and weights natively from Scala, and we use the Java implementation of
Google Protocol Buffers to allow the dynamic construction of Caffe networks at runtime.
The Net class wraps Caffe and exposes a simple API containing the methods shown in Listing 1.
The NetParams type specifies a network architecture, and the WeightCollection type is
a map from layer names to lists of weights. It allows the manipulation of network components
and the storage of weights and outputs for individual layers. To facilitate manipulation of data
and weights without copying memory from Caffe, we implement the NDArray class, which is a
lightweight multi-dimensional tensor library. One benefit of building on Caffe is that any existing
Caffe model definition or solver file is automatically compatible with SparkNet. There is a large
community developing Caffe models and extensions, and these can easily be used in SparkNet. By
building on top of Spark, we inherit the advantages of modern batch computational frameworks.
These include the high-throughput loading and preprocessing of data and the ability to keep data in
memory between operations. In Listing 2, we give an example of how network architectures can
be specified in SparkNet. In addition, model specifications or weights can be loaded directly from
Caffe files. An example sketch of code that uses our API to perform distributed training is given in
Listing 3.
2.1
PARALLELIZING SGD
To perform well in bandwidth-limited environments, we recommend a parallelization scheme for
SGD that requires minimal communication. This approach is not specific to SGD. Indeed, SparkNet
works out of the box with any Caffe solver.
3
Published as a conference paper at ICLR 2016
The parallelization scheme is described in Listing 3. Spark consists of a single master node and a
number of worker nodes. The data is split among the Spark workers. In every iteration, the Spark
master broadcasts the model parameters to each worker. Each worker then runs SGD on the model
with its subset of data for a fixed number of iterations τ (we use τ = 50 in Listing 3) or for a fixed
length of time, after which the resulting model parameters on each worker are sent to the master and
averaged to form the new model parameters. We recommend initializing the network by running
SGD for a small number of iterations on the master. A similar and more sophisticated approach to
parallelizing SGD with minimal communication overhead is discussed in Zhang et al. (2015).
The standard approach to parallelizing each gradient computation requires broadcasting and collecting model parameters (hundreds of megabytes per worker and gigabytes in total) after every SGD
update, which occurs tens of thousands of times during training. On our EC2 cluster, each broadcast
and collection takes about twenty seconds, putting a bound on the speedup that can be expected
using this approach without better hardware or without partitioning models across machines. Our
approach broadcasts and collects the parameters a factor of τ times less for the same number of
iterations. In our experiments, we set τ = 50, but other values seem to work about as well.
We note that Caffe supports parallelism across multiple GPUs within a single node. This is not a
competing form of parallelism but rather a complementary one. In some of our experiments, we use
Caffe to handle parallelism within a single node, and we use the parallelization scheme described in
Listing 3 to handle parallelism across nodes.
3
E XPERIMENTS
In Section 3.2, we will benchmark the performance of SparkNet and measure the speedup that our
system obtains relative to training on a single node. However, the outcomes of those experiments
depend on a number of different factors. In addition to τ (the number of iterations between synchronizations) and K (the number of machines in our cluster), they depend on the communication
overhead in our cluster S. In Section 3.1, we find it instructive to measure the speedup in the idealized case of zero communication overhead (S = 0). This idealized model gives us an upper bound
on the maximum speedup that we could hope to obtain in a real-world cluster, and it allows us to
build a model for the speedup as a function of S (the overhead is easily measured in practice).
3.1
T HEORETICAL C ONSIDERATIONS
Before benchmarking our system, we determine the maximum possible speedup that could be obtained in principle in a cluster with no communication overhead. We determine the dependence of
this speedup on the parameters τ (the number of iterations between synchronizations) and K (the
number of machines in our cluster).
3.1.1
L IMITATIONS OF NAIVE PARALLELIZATION
To begin with, we consider the theoretical limitations of a naive parallelism scheme which parallelizes SGD by distributing each minibatch computation over multiple machines (see Figure 2b).
Let Na (b) be the number of serial iterations of SGD required to obtain an accuracy of a when training with a batch size of b (when we say accuracy, we are referring to test accuracy). Suppose that
computing the gradient over a batch of size b requires C(b) units of time. Then the running time
required to achieve an accuracy of a with serial training is
Na (b)C(b).
(1)
A naive parallelization scheme attempts to distribute the computation at each iteration by dividing
each minibatch between the K machines, computing the gradients separately, and aggregating the
results on one node. Under this scheme, the cost of the computation done on a single node in a single
iteration is C(b/K) and satisfies C(b/K) ≥ C(b)/K (the cost is sublinear in the batch size). In a
system with no communication overhead and no overhead for summing the gradients, this approach
could in principle achieve an accuracy of a in time Na (b)C(b)/K. This represents a linear speedup
in the number of machines (for values of K up to the batch size b).
In practice, there are several important considerations. First, for the approximation C(b/K) ≈
C(b)/K to hold, K must be much smaller than b, limiting the number of machines we can use to
4
Published as a conference paper at ICLR 2016
effectively parallelize the minibatch computation. One might imagine circumventing this limitation
by using a larger batch size b. Unfortunately, the benefit of using larger batches is relatively modest.
As the batch size b increases, Na (b) does not decrease enough to justify the use of a very large value
of b.
Furthermore, the benefits of this approach depend greatly on the degree of communication overhead.
If aggregating the gradients and broadcasting the model parameters requires S units of time, then
the time required by this approach is at least C(b)/K + S per iteration and Na (b)(C(b)/K + S) to
achieve an accuracy of a. Therefore, the maximum achievable speedup is C(b)/(C(b)/K + S) ≤
C(b)/S. We may expect S to increase modestly as K increases, but we suppress this effect here.
3.1.2
L IMITATIONS OF S PARK N ET PARALLELIZATION
The performance of the naive parallelization scheme is easily understood because its behavior is
equivalent to that of the serial algorithm. In contrast, SparkNet uses a parallelization scheme that is
not equivalent to serial SGD (described in Section 2.1), and so its analysis is more complex.
SparkNet’s parallelization scheme proceeds in rounds (see Figure 2c). In each round, each machine
runs SGD for τ iterations with batch size b. Between rounds, the models on the workers are gathered
together on the master, averaged, and broadcast to the workers.
We use Ma (b, K, τ ) to denote the number of rounds required to achieve an accuracy of a. The
number of parallel iterations of SGD under SparkNet’s parallelization scheme required to achieve
an accuracy of a is then τ Ma (b, K, τ ), and the wallclock time is
(τ C(b) + S)Ma (b, K, τ ),
(2)
where S is the time required to gather and broadcast model parameters.
To measure the sensitivity of SparkNet’s parallelization scheme to the parameters τ and K, we
consider a grid of values of K and τ . For each pair of parameters, we run SparkNet using a modified
version of AlexNet on a subset of ImageNet (the first 100 classes each with approximately 1000 data
points) for a total of 20000 parallel iterations. For each of these training runs, we compute the ratio
τ Ma (b, K, τ )/Na (b). This is the speedup achieved relative to training on a single machine when
S = 0. In Figure 3, we plot a heatmap of the speedup given by the SparkNet parallelization scheme
under different values of τ and K.
Figure 3 exhibits several trends. The top row of the heatmap corresponds to the case K = 1, where
we use only one worker. Since we do not have multiple workers to synchronize when K = 1, the
number of iterations τ between synchronizations does not matter, so all of the squares in the top
row of the grid should behave similarly and should exhibit a speedup factor of 1 (up to randomness
in the optimization). The rightmost column of each heatmap corresponds to the case τ = 1, where
we synchronize after every iteration of SGD. This is equivalent to running serial SGD with a batch
size of Kb, where b is the batchsize on each worker (in these experiments we use b = 100). In this
column, the speedup should increase sublinearly with K. We note that it is slightly surprising that
the speedup does not increase monotonically from left to right as τ decreases. Intuitively, we might
expect more synchronization to be strictly better (recall we are disregarding the overhead due to
synchronization). However, our experiments suggest that modest delays between synchronizations
can be beneficial.
This experiment capture the speedup that we can expect from the SparkNet parallelization scheme
in the case of zero communication overhead (the numbers are dataset specific, but the trends are of
interest). Having measured these numbers, it is straightforward to compute the speedup that we can
expect as a function of the communication overhead.
In Figure 4, we plot the speedup expected both from naive parallelization and from SparkNet on
a five-node cluster as a function of S (normalized so that C(b) = 1). As expected, naive parallelization gives a maximum speedup of 5 (on a five-node cluster) when there is zero communication
overhead (note that our plot does not go all the way to S = 0), and it gives no speedup when the
communication overhead is comparable to or greater than the cost of a minibatch computation. In
contrast, SparkNet gives a relatively consistent speedup even when the communication overhead is
100 times the cost of a minibatch computation.
5
Published as a conference paper at ICLR 2016
(a) This figure depicts a serial run of SGD. Each block corresponds to a single SGD update with
batch size b. The quantity Na (b) is the number of iterations required to achieve an accuracy of a.
(b) This figure depicts a parallel run of SGD on K = 4 machines under a naive parallelization
scheme. At each iteration, each batch of size b is divided among the K machines, the gradients
over the subsets are computed separately on each machine, the updates are aggregated, and the new
model is broadcast to the workers. Algorithmically, this approach is exactly equivalent to the serial
run of SGD in Figure 2a and so the number of iterations required to achieve an accuracy of a is the
same value Na (b).
(c) This figure depicts a parallel run of SGD on K = 4 machines under SparkNet’s parallelization
scheme. At each step, each machine runs SGD with batch size b for τ iterations, after which the
models are aggregated, averaged, and broadcast to the workers. The quantity Ma (b, K, τ ) is the
number of rounds (of τ iterations) required to obtain an accuracy of a. The total number of parallel
iterations of SGD under SparkNet’s parallelization scheme required to obtain an accuracy of a is
then τ Ma (b, K, τ ).
Figure 2: Computational models for different parallelization schemes.
The speedup given by the naive parallelization scheme can be computed exactly and is given by
C(b)/(C(b)/K +S). This formula is essentially Amdahl’s law. Note that when S ≥ C(b), the naive
parallelization scheme is slower than the computation on a single machine. The speedup obtained
by SparkNet is Na (b)C(b)/[(τ C(b) + S)Ma (b, K, τ )] for a specific value of τ . The numerator is
the time required by serial SGD to achieve an accuracy of a from Equation 1, and the denominator is
the time required by SparkNet to achieve the same accuracy from Equation 2. Choosing the optimal
value of τ gives us a speedup of maxτ Na (b)C(b)/[(τ C(b) + S)Ma (b, K, τ )]. In practice, choosing
τ is not a difficult problem. The ratio Na (b)/(τ Ma (b, K, τ )) (the speedup when S = 0) degrades
6
Published as a conference paper at ICLR 2016
speedup to accuracy 20%
1.1
1.3
1.2
1.3
1.2
1.1
0.8
1.3
2.75
2
1.4
1.6
1.6
1.8
1.6
2.0
2.0
1.8
1.6
2.50
3
1.3
1.4
1.9
2.1
2.1
2.1
2.2
2.6
1.7
4
1.2
1.7
1.8
1.9
2.2
2.4
2.6
2.1
1.9
5
1.1
1.6
1.9
1.9
2.4
3.0
3.0
2.4
1.9
6
1.1
1.7
1.7
1.8
2.6
2.8
3.0
2.4
2.0
1000
500
100
25
10
5
2
1
K
1.2
2500
3.00
1
2.25
2.00
1.75
1.50
1.25
1.00
τ
Figure 3: This figure shows the speedup τ Ma (b, τ, K)/Na (b) given by SparkNet’s parallelization
scheme relative to training on a single machine to obtain an accuracy of a = 20%. Each grid square
corresponds to a different choice of K and τ . We show the speedup in the zero communication
overhead setting. This experiment uses a modified version of AlexNet on a subset of ImageNet
(100 classes each with approximately 1000 images). Note that these numbers are dataset specific.
Nevertheless, the trends they capture are of interest.
5
Naive
SparkNet
No Speedup
speedup
4
3
2
1
0
10−2
10−1
100
101
communication overhead S
102
103
Figure 4: This figure shows the speedups obtained by the naive parallelization scheme and by
SparkNet as a function of the cluster’s communication overhead (normalized so that C(b) = 1).
We consider K = 5. The data for this plot applies to training a modified version of AlexNet on
a subset of ImageNet (approximately 1000 images for each of the first 100 classes). The speedup
obtained by the naive parallelization scheme is C(b)/(C(b)/K + S). The speedup obtained by
SparkNet is Na (b)C(b)/[(τ C(b) + S)Ma (b, K, τ )] for a specific value of τ . The numerator is the
time required by serial SGD to achieve an accuracy of a, and the denominator is the time required by
SparkNet to achieve the same accuracy (see Equation 1 and Equation 2). For the optimal value of τ ,
the speedup is maxτ Na (b)C(b)/[(τ C(b) + S)Ma (b, K, τ )]. To plot the SparkNet speedup curve,
we maximize over the set of values τ ∈ {1, 2, 5, 10, 25, 100, 500, 1000, 2500} and use the values
Ma (b, K, τ ) and Na (b) from the experiments in the fifth row of Figure 3. In our experiments, we
have S ≈ 20s and C(b) ≈ 2s.
slowly as τ increases, so it suffices to choose τ to be a small multiple of S (say 5S) so that the
algorithm spends only a fraction of its time in communication.
When plotting the SparkNet speedup in Figure 4, we do not maximize over all positive integer
values of τ but rather over the set τ ∈ {1, 2, 5, 10, 25, 100, 500, 1000, 2500}, and we use the values
7
45
40
35
30
25
20
15
10
5
0
accuracy
accuracy
Published as a conference paper at ICLR 2016
Caffe
SparkNet 3 node
SparkNet 5 node
SparkNet 10 node
0
5
10
hours
15
20
Figure 5: This figure shows the performance
of SparkNet on a 3-node, 5-node, and 10-node
cluster, where each node has 1 GPU. In these
experiments, we use τ = 50. The baseline
was obtained by running Caffe on a single GPU
with no communication. The experiments are
performed on ImageNet using AlexNet.
60
50
40
30
20
10
0
Caffe 4 GPU
SparkNet 3 node 4 GPU
SparkNet 6 node 4 GPU
0
20
40
60
hours
80
100
120
Figure 6: This figure shows the performance of
SparkNet on a 3-node cluster and on a 6-node
cluster, where each node has 4 GPUs. In these
experiments, we use τ = 50. The baseline uses
Caffe on a single node with 4 GPUs and no
communication overhead. The experiments are
performed on ImageNet using GoogLeNet.
of Na (b) and Ma (b, K, τ ) corresponding to the fifth row of Figure 3. Including more values of τ
would only increase the SparkNet speedup. The distributed training of deep networks is typically
thought of as a communication-intensive procedure. However, Figure 4 demonstrates the value of
SparkNet’s parallelization scheme even in the most bandwidth-limited settings.
The naive parallelization scheme may appear to be a straw man. However, it is a frequently-used approach to parallelizing SGD (Noel et al., 2015; Iandola et al., 2015), especially when asynchronous
updates are not an option (as in computational frameworks like MapReduce and Spark).
3.2
T RAINING B ENCHMARKS
To explore the scaling behavior of our algorithm and implementation, we perform experiments on
EC2 using clusters of g2.8xlarge nodes. Each node has four NVIDIA GRID GPUs and 60GB
memory. We train the default Caffe model of AlexNet (Krizhevsky et al., 2012) on the ImageNet
dataset (Russakovsky et al., 2015). We run SparkNet with K = 3, 5, and 10 and plot the results
in Figure 5. For comparison, we also run Caffe on the same cluster with a single GPU and no
communication overhead to obtain the K = 1 plot. These experiments use only a single GPU on
each node. To measure the speedup, we compare the wall-clock time required to obtain an accuracy
of 45%. With 1 GPU and no communication overhead, this takes 55.6 hours. With 3, 5, and 10
GPUs, SparkNet takes 22.9, 14.5, and 12.8 hours, giving speedups of 2.4, 3.8, and 4.4.
We also train the default Caffe model of GoogLeNet (Szegedy et al., 2015) on ImageNet. We run
SparkNet with K = 3 and K = 6 and plot the results in Figure 6. In these experiments, we
use Caffe’s multi-GPU support to take advantage of all four GPUs within each node, and we use
SparkNet’s parallelization scheme to handle parallelism across nodes. For comparison, we train
Caffe on a single node with four GPUs and no communication overhead. To measure the speedup,
we compare the wall-clock time required to obtain an accuracy of 40%. Relative to the baseline of
Caffe with four GPUs, SparkNet on 3 and 6 nodes gives speedups of 2.7 and 3.2. Note that this is
on top of the speedup of roughly 3.5 that Caffe with four GPUs gets over Caffe with one GPU, so
the speedups that SparkNet obtains over Caffe on a single GPU are roughly 9.4 and 11.2.
Furthermore, we explore the dependence of the parallelization scheme described in Section 2.1 on
the parameter τ which determines the number of iterations of SGD that each worker does before
synchronizing with the other workers. These results are shown in Figure 7. Note that in the presence
of stragglers, it suffices to replace the fixed number of iterations τ with a fixed length of time, but in
our experimental setup, the timing was sufficiently consistent and stragglers did not arise. The single
GPU experiment in Figure 5 was trained on a single GPU node with no communication overhead.
8
accuracy
Published as a conference paper at ICLR 2016
45
40
35
30
25
20
15
10
5
0
20 iterations
50 iterations
100 iterations
150 iterations
0
2
4
hours
6
8
10
Figure 7: This figure shows the dependence of the parallelization scheme described in Section 2.1
on τ . Each experiment was run with K = 5 workers. This figure shows that good performance can
be achieved without collecting and broadcasting the model after every SGD update.
4
R ELATED W ORK
Much work has been done to build distributed frameworks for training deep networks. Coates et al.
(2013) build a model-parallel system for training deep networks on a GPU cluster using MPI over
Infiniband. Dean et al. (2012) build DistBelief, a distributed system capable of training deep networks on thousands of machines using stochastic and batch optimization procedures. In particular,
they highlight asynchronous SGD and batch L-BFGS. Distbelief exploits both data parallelism and
model parallelism. Chilimbi et al. (2014) build Project Adam, a system for training deep networks
on hundreds of machines using asynchronous SGD. Li et al. (2014); Ho et al. (2013) build parameter
servers to exploit model and data parallelism, and though their systems are better suited to sparse
gradient updates, they could very well be applied to the distributed training of deep networks. More
recently, Abadi et al. (2015) build TensorFlow, a sophisticated system for training deep networks
and more generally for specifying computation graphs and performing automatic differentiation.
Iandola et al. (2015) build FireCaffe, a data-parallel system that achieves impressive scaling using
naive parallelization in the high-performance computing setting. They minimize communication
overhead by using a tree reduce for aggregating gradients in a supercomputer with Cray Gemini
interconnects.
These custom systems have numerous advantages including high performance, fine-grained control
over scheduling and task placement, and the ability to take advantage of low-latency communication
between machines. On the other hand, due to their demanding communication requirements, they
are unlikely to exhibit the same scaling on an EC2 cluster. Furthermore, due to their nature as custom
systems, they lack the benefits of tight integration with general-purpose computational frameworks
such as Spark. For some of these systems, preprocessing must be done separately by a MapReduce
style framework, and data is written to disk between segments of the pipeline. With SparkNet,
preprocessing and training are both done in Spark.
Training a machine learning model such as a deep network is often one step of many in real-world
data analytics pipelines (Sparks et al., 2015). Obtaining, cleaning, and preprocessing the data are
often expensive operations, as is transferring data between systems. Training data for a machine
learning model may be derived from a streaming source, from a SQL query, or from a graph computation. A user wishing to train a deep network in a custom system on the output of a SQL query
would need a separate SQL engine. In SparkNet, training a deep network on the output of a SQL
query, or a graph computation, or a streaming data source is straightforward due to its general purpose nature and its support for SQL, graph computations, and data streams (Armbrust et al., 2015;
Gonzalez et al., 2014; Zaharia et al., 2013).
Some attempts have been made to train deep networks in general-purpose computational frameworks, however, existing work typically hinges on extremely low-latency intra-cluster communica9
Published as a conference paper at ICLR 2016
tion. Noel et al. (2015) train deep networks in Spark on top of YARN using SGD and leverage cluster
resources to parallelize the computation of the gradient over each minibatch. To achieve competitive
performance, they use remote direct memory accesses over Infiniband to exchange model parameters
quickly between GPUs. In contrast, SparkNet tolerates low-bandwidth intra-cluster communication
and works out of the box on Amazon EC2.
A separate line of work addresses speeding up the training of deep networks using single-machine
parallelism. For example, Caffe con Troll (Abuzaid et al., 2015) modifies Caffe to leverage both
CPU and GPU resources within a single node. These approaches are compatible with SparkNet and
the two can be used in conjunction.
Many popular computational frameworks provide support for training machine learning models
(Meng et al., 2015) such as linear models and matrix factorization models. However, due to the
demanding communication requirements and the larger scale of many deep learning problems, these
libraries have not been extended to include deep networks.
Various authors have studied the theory of averaging separate runs of SGD. In the bandwidth-limited
setting, Zinkevich et al. (2010) analyze a simple algorithm for convex optimization that is easily
implemented in the MapReduce framework and can tolerate high-latency communication between
machines. Zhang et al. (2015) define a parallelization scheme that penalizes divergences between
parallel workers, and they provide an analysis in the convex case. Zhang & Jordan (2015) propose a general abstraction for parallelizing stochastic optimization algorithms along with a Spark
implementation.
5
D ISCUSSION
We have described an approach to distributing the training of deep networks in communicationlimited environments that lends itself to an implementation in batch computational frameworks like
MapReduce and Spark. We provide SparkNet, an easy-to-use deep learning implementation for
Spark that is based on Caffe and enables the easy parallelization of existing Caffe models with
minimal modification. As machine learning increasingly depends on larger and larger datasets,
integration with a fast and general engine for big data processing such as Spark allows researchers
and practitioners to draw from a rich ecosystem of tools to develop and deploy their models. They
can build models that incorporate features from a variety of data sources like images on a distributed
file system, results from a SQL query or graph database query, or streaming data sources.
Using a smaller version of the ImageNet benchmark we quantify the speedup achieved by SparkNet
as a function of the size of the cluster, the communication frequency, and the cluster’s communication overhead. We demonstrate that our approach is effective even in highly bandwidth-limited
settings. On the full ImageNet benchmark we showed that our system achieves a sizable speedup
over a single node experiment even with few GPUs.
The code for SparkNet is available at https://github.com/amplab/SparkNet. We invite
contributions and hope that the project will help bring a diverse set of deep learning applications to
the Spark community.
ACKNOWLEDGMENTS
We would like to thank Cyprien Noel, Andy Feng, Tomer Kaftan, Evan Sparks, and Shivaram
Venkataraman for valuable advice. This research is supported in part by NSF grant number DGE1106400. This research is supported in part by NSF CISE Expeditions Award CCF-1139158, DOE
Award SN10040 DE-SC0012463, and DARPA XData Award FA8750-12-2-0331, and gifts from
Amazon Web Services, Google, IBM, SAP, The Thomas and Stacey Siebel Foundation, Adatao,
Adobe, Apple, Blue Goji, Bosch, Cisco, Cray, Cloudera, EMC2, Ericsson, Facebook, Fujitsu,
Guavus, HP, Huawei, Informatica, Intel, Microsoft, NetApp, Pivotal, Samsung, Schlumberger,
Splunk, Virdata and VMware.
10
Published as a conference paper at ICLR 2016
R EFERENCES
Abadi, Martı́n, Agarwal, Ashish, Barham, Paul, et al. TensorFlow: Large-scale machine learning on
heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from
tensorflow.org.
Abuzaid, Firas, Hadjis, Stefan, Zhang, Ce, and Ré, Christopher. Caffe con Troll: Shallow ideas to
speed up deep learning. arXiv preprint arXiv:1504.04343, 2015.
Armbrust, Michael, Xin, Reynold S, Lian, Cheng, Huai, Yin, Liu, Davies, Bradley, Joseph K, Meng,
Xiangrui, Kaftan, Tomer, Franklin, Michael J, Ghodsi, Ali, et al. Spark SQL: Relational data
processing in Spark. In Proceedings of the 2015 ACM SIGMOD International Conference on
Management of Data, pp. 1383–1394. ACM, 2015.
Chilimbi, Trishul, Suzue, Yutaka, Apacible, Johnson, and Kalyanaraman, Karthik. Project Adam:
Building an efficient and scalable deep learning training system. In 11th USENIX Symposium on
Operating Systems Design and Implementation, pp. 571–582, 2014.
Coates, Adam, Huval, Brody, Wang, Tao, Wu, David, Catanzaro, Bryan, and Andrew, Ng. Deep
learning with cots hpc systems. In Proceedings of the 30th International Conference on Machine
Learning, pp. 1337–1345, 2013.
Dean, Jeffrey and Ghemawat, Sanjay. MapReduce: simplified data processing on large clusters.
Communications of the ACM, 51(1):107–113, 2008.
Dean, Jeffrey, Corrado, Greg, Monga, Rajat, Chen, Kai, Devin, Matthieu, Mao, Mark, Ranzato,
Marc’Aurelio, Senior, Andrew, Tucker, Paul, Yang, Ke, Le, Quoc V., and Ng, Andrew Y. Large
scale distributed deep networks. In Advances in Neural Information Processing Systems, pp.
1223–1231, 2012.
Gonzalez, Joseph E, Xin, Reynold S, Dave, Ankur, Crankshaw, Daniel, Franklin, Michael J, and
Stoica, Ion. Graphx: Graph processing in a distributed dataflow framework. In Proceedings of
OSDI, pp. 599–613, 2014.
Ho, Qirong, Cipar, James, Cui, Henggang, Lee, Seunghak, Kim, Jin Kyu, Gibbons, Phillip B, Gibson, Garth A, Ganger, Greg, and Xing, Eric P. More effective distributed ML via a stale synchronous parallel parameter server. In Advances in Neural Information Processing Systems, pp.
1223–1231, 2013.
Iandola, Forrest N, Ashraf, Khalid, Moskewicz, Mattthew W, and Keutzer, Kurt. FireCaffe:
near-linear acceleration of deep neural network training on compute clusters. arXiv preprint
arXiv:1511.00175, 2015.
Isard, Michael, Budiu, Mihai, Yu, Yuan, Birrell, Andrew, and Fetterly, Dennis. Dryad: Distributed
data-parallel programs from sequential building blocks. In Proceedings of the 2nd ACM SIGOPS/EuroSys European Conference on Computer Systems, pp. 59–72, 2007.
Jia, Yangqing, Shelhamer, Evan, Donahue, Jeff, Karayev, Sergey, Long, Jonathan, Girshick, Ross,
Guadarrama, Sergio, and Darrell, Trevor. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the ACM International Conference on Multimedia, pp. 675–678.
ACM, 2014.
Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, pp. 1097–1105,
2012.
Li, Mu, Andersen, David G, Park, Jun Woo, Smola, Alexander J, Ahmed, Amr, Josifovski, Vanja,
Long, James, Shekita, Eugene J, and Su, Bor-Yiing. Scaling distributed machine learning with the
parameter server. In 11th USENIX Symposium on Operating Systems Design and Implementation,
pp. 583–598, 2014.
Meng, Xiangrui, Bradley, Joseph, Yavuz, Burak, Sparks, Evan, Venkataraman, Shivaram, Liu,
Davies, Freeman, Jeremy, Tsai, DB, Amde, Manish, Owen, Sean, et al. MLlib: Machine learning
in Apache Spark. arXiv preprint arXiv:1505.06807, 2015.
11
Published as a conference paper at ICLR 2016
Murray, Derek G, McSherry, Frank, Isaacs, Rebecca, Isard, Michael, Barham, Paul, and Abadi,
Martı́n. Naiad: a timely dataflow system. In Proceedings of the Twenty-Fourth ACM Symposium
on Operating Systems Principles, pp. 439–455. ACM, 2013.
Noel, Cyprien, Shi, Jun, and Feng, Andy. Large scale distributed deep learning on Hadoop
clusters, 2015. URL http://yahoohadoop.tumblr.com/post/129872361846/
large-scale-distributed-deep-learning-on-hadoop.
Russakovsky, Olga, Deng, Jia, Su, Hao, Krause, Jonathan, Satheesh, Sanjeev, Ma, Sean, Huang,
Zhiheng, Karpathy, Andrej, Khosla, Aditya, Bernstein, Michael, Berg, Alexander C., and Fei-Fei,
Li. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer
Vision, pp. 1–42, 2015.
Sparks, Evan R., Venkataraman, Shivaram, Kaftan, Tomer, Franklin, Michael, and Recht, Benjamin.
KeystoneML: End-to-end machine learning pipelines at scale. 2015.
Szegedy, Christian, Liu, Wei, Jia, Yangqing, Sermanet, Pierre, Reed, Scott, Anguelov, Dragomir,
Erhan, Dumitru, Vanhoucke, Vincent, and Rabinovich, Andrew. Going deeper with convolutions.
In Computer Vision and Pattern Recognition, 2015.
Zaharia, Matei, Chowdhury, Mosharaf, Franklin, Michael J, Shenker, Scott, and Stoica, Ion. Spark:
cluster computing with working sets. In Proceedings of the 2nd USENIX conference on Hot topics
in cloud computing, volume 10, pp. 10, 2010.
Zaharia, Matei, Das, Tathagata, Li, Haoyuan, Hunter, Timothy, Shenker, Scott, and Stoica, Ion.
Discretized streams: Fault-tolerant streaming computation at scale. In Proceedings of the TwentyFourth ACM Symposium on Operating Systems Principles, pp. 423–438. ACM, 2013.
Zhang, Sixin, Choromanska, Anna E, and LeCun, Yann. Deep learning with elastic averaging SGD.
In Advances in Neural Information Processing Systems, pp. 685–693, 2015.
Zhang, Yuchen and Jordan, Michael I. Splash: User-friendly programming interface for parallelizing
stochastic algorithms. arXiv preprint arXiv:1506.07552, 2015.
Zinkevich, Martin, Weimer, Markus, Li, Lihong, and Smola, Alex J. Parallelized stochastic gradient
descent. In Advances in Neural Information Processing Systems, pp. 2595–2603, 2010.
12
| 9cs.NE
|
On the sub-Gaussianity of the Beta and Dirichlet distributions
Olivier Marchal1 and Julyan Arbel2
1
Université de Lyon, CNRS UMR 5208, Université Jean Monnet,
Institut Camille Jordan, France
2
arXiv:1705.00048v2 [math.ST] 25 Sep 2017
Inria Grenoble Rhône-Alpes,
Laboratoire Jean Kuntzmann, Université Grenoble Alpes, France
September 26, 2017
Abstract
We obtain the optimal proxy variance for the sub-Gaussianity of Beta distribution, thus
proving upper bounds recently conjectured by Elder (2016). We provide different proof techniques
for the symmetrical (around its mean) case and the non-symmetrical case. The technique in the
latter case relies on studying the ordinary differential equation satisfied by the Beta momentgenerating function known as the confluent hypergeometric function. As a consequence, we derive
the optimal proxy variance for the Dirichlet distribution, which is apparently a novel result. We
also provide a new proof of the optimal proxy variance for the Bernoulli distribution, and discuss
in this context the proxy variance relation to log-Sobolev inequalities and transport inequalities.
1
Introduction
The sub-Gaussian property (Buldygin and Kozachenko, 1980, 2000; Pisier, 2016) and related concentration inequalities (Boucheron et al., 2013; Raginsky and Sason, 2013) have attracted a lot of
attention in the last couple of decades due to their applications in various areas such as pure mathematics, physics, information theory and computer sciences. Recent interest focused on deriving
the optimal proxy variance for discrete random variables like the Bernoulli distribution (Buldygin
and Moskvichova, 2013; Kearns and Saul, 1998; Berend and Kontorovich, 2013) and the missing
mass (McAllester and Schapire, 2000; McAllester and Ortiz, 2003; Berend and Kontorovich, 2013;
Ben-Hamou et al., 2017). Our focus is instead on two continuous random variables, the Beta and
Dirichlet distributions, for which the optimal proxy variance was not known to the best of our knowledge. Some upper bounds were recently conjectured by Elder (2016) that we prove in the present
article by providing the optimal proxy variance for both Beta and Dirichlet distributions. Similar
concentration properties of the Beta distribution have been recently used in many contexts including
Bayesian adaptive data analysis (Elder, 2016), Bayesian nonparametrics (Castillo, 2016) and spectral
properties of random matrices (Perry et al., 2016).
We start by reminding the definition of sub-Gaussian property for random variables:
Definition 1 (Sub-Gaussian variables). A random variable X with finite mean µ = E[X] is subGaussian if there is a positive number σ such that:
2 2
λ σ
E[exp(λ(X − µ))] ≤ exp
for all λ ∈ R.
(1)
2
Such a constant σ 2 is called a proxy variance (or sub-Gaussian norm), and we say that X is σ 2 -subGaussian. If X is sub-Gaussian, one is usually interested in the optimal proxy variance:
2
σopt
(X) = min{σ 2 ≥ 0 such that X is σ 2 -sub-Gaussian}.
2
Note that the variance always gives a lower bound on the optimal proxy variance: Var[X] ≤ σopt
(X).
2
In particular, when σopt (X) = Var[X], X is said to be strictly sub-Gaussian.
1
Every compactly supported distribution, as is the Beta(α, β) distribution, is sub-Gaussian. This
can be seen by Hoeffding’s classic inequality: any random variable X supported on [0, 1] with mean
µ satisfies
i
h
∀λ ∈ R,
E eλ(X−µ) ≤ e
λ2
8
,
thus exhibiting 14 as an upper bound to the proxy variance. This bound can be improved by taking
into account the location of the mean µ within the interval [0, 1]. An early step in this direction is
the second inequality in Hoeffding (1963) paper, indexed (2.2). It states that if µ < 1/2, then for any
2
positive , P(X − µ > ) ≤ e− g(µ) , where
g(µ) =
1−µ
1
ln
1 − 2µ
µ
(2)
1
thus indicating that X has a right tail lighter than a Gaussian tail of variance 2g(µ)
. Hoeffding’s
result was strengthened by Kearns and Saul (1998) to comply with Definition 1 of sub-Gaussianity1
as follows
2
λ
E[exp(λ(X − µ))] ≤ exp
for all λ ∈ R,
(3)
4g(µ)
1
is a distribution-sensitive proxy variance for any [0, 1]-supported random
thus indicating that 2g(µ)
variable with mean µ (see also Berend and Kontorovich, 2013, for a detailed proof of this result). If
this is the optimal proxy variance for the Bernoulli distribution (see Theorem 2.1 and Theorem 3.1
of Buldygin and Moskvichova, 2013), it is clear from our result that it does not hold true for the
α
Beta distribution. However, fixing α+β
= µ and letting α → 0, β → 0, the Beta(α, β) distribution
concentrates to the Bern(µ) distribution, and we show that we recover the optimal proxy variance
for the Bernoulli distribution (Theorem 2).
An interesting common feature between optimal proxy variances for the Bernoulli distribution:
1
2g(µ) , and that of the Beta distribution derived later on, is that they deteriorate in a similar fashion
as the mean µ goes to 0 or 1, see for instance the left panel of Figure 1. We briefly present here
classical proof techniques for sub-Gaussianity hinging on certain tools from functional analysis. We
show how they apply in the Bernoulli setting, and let as an interesting open problem how our proof
in the Beta distribution setting could be supplemented by these same functional analysis tools.
Essentially two (related) functional inequalities allow one to derive a sub-Gaussian property: logSobolev inequalities, which date back to Gross (1975), and transport inequalities. The relation with
the former inequalities is called Herbst’s argument. It states that if a probability measure satisfies a
log-Sobolev inequality with some constant, then it is sub-Gaussian with the same constant as a proxy
variance2 (see for instance Ledoux, 1999, Section 2.3 and Proposition 2.3). The optimal constant
in the log-Sobolev inequality satisfied by the Bernoulli distribution also produces its optimal proxy
variance (Ledoux, 1999, Corollary 5.9).
The relation with transport inequalities is usually referred to as Marton’s argument (see for
instance Raginsky and Sason, 2013, Section 3.4). Define the Wasserstein distance between two probability measures P and Q on a space X by
Z
W (P, Q) = inf
d(x, y)π(dx, dy),
π∈Π(P,Q)
X ×X
where Π(P, Q) is the set of probability measures on X × X with fixed marginal distributions respectively P and Q. The Wasserstein distance depends on some choice of a distance d on X . A probability
measure P is said to satisfy a transport inequality with constant c, if for any probability measure Q
dominated by P ,
p
W (P, Q) ≤ 2cD(Q||P ),
(4)
where D(Q||P ) is the entropy, or Kullback–Leibler divergence, between P and Q. The transport
inequality (4) is denoted by T(c).
Bobkov and Götze (1999) proved that T(c) implies c-sub-Gaussianity. See also Proposition 3.6
and Theorem 3.4.4 of Raginsky and Sason (2013) for general results. Further developments in the
2
−
1 Note
indeed that Equation (1), together with Markov inequality, imply P(X − µ > ) ≤ e 2σ2 .
2 The implied predicate is actually stronger than sub-Gaussianity, but it is not useful for our purposes.
2
discrete X setting are interesting for our purposes. Equip a discrete space X with the Hamming
metric, d(x, y) = 1{x6=y} . The induced Wasserstein distance then reduces to the total variation
distance, W (P, Q) = kP − QkTV . In that setting, Ordentlich and Weinberger (2005) proved the
distribution-sensitive transport inequality:
s
1
D(Q||P ),
(5)
kP − QkTV ≤
g(µP )
where the function g is defined in Equation (2) and the coefficient µP is called the balance coefficient of
P , and is defined by µP = max min{P (A), 1 − P (A)}. In particular, the Bernoulli balance coefficient
A⊂X
is easily
shown to
coincide with its mean. Hence, applying the result of Bobkov and Götze (1999)
1
1
to the T 2g(µP ) transport inequality (5) yields a distribution-sensitive proxy variance of 2g(µ)
for
the Bernoulli with mean µ. It is optimal, see for instance Theorem 3.4.6 of Raginsky and Sason
(2013). This viewpoint highlights the key role played by the balance coefficient in the non-uniformity
of the optimal proxy variance for discrete distributions such as the Bernoulli. However, it is not
clear how this argument would carry over to non discrete distributions such as the Beta distribution
for explaining similar sensitivity to the mean. However, to quote Raginsky and Sason (2013), the
general approach may not produce optimal concentration estimates, that often require case-by-case
treatments. This is the route followed in this note for the Beta distribution.
The outline of the note is as follows. We introduce the Beta distribution and state the main result
(Theorem 1) in Section 2.1. We then prove our result depending on whether α = β (Section 2.2) or
α 6= β (Section 2.3). In the first case, the proof is elementary and based on comparing the coefficients
of the entire series representations of the functions of both sides of inequality (1). However, it does
not directly carry over to the second case, whose proof requires some finer analysis tool: the study
of the ordinary differential equation (ODE) satisfied by the confluent hypergeometric function 1 F1 .
Although the second proof also covers the case α = β upon slight modifications, the independent
proof for the symmetric case is kept owing to its simplicity. As a by-product, we derive the optimal
proxy variance for the Bernoulli and the Dirichlet distributions in Section 3. The R code for the plots
presented in this note and for a function deriving the optimal proxy variance in terms of α and β is
available at http://www.julyanarbel.com/software.
2
2.1
Optimal proxy variance for the Beta distribution
Notations and main result
The Beta(α, β) distribution, with α, β > 0, is characterized by a density on the segment [0, 1] given
by:
f (x) =
1
xα−1 (1 − x)β−1 ,
B(α, β)
R∞
where B(α, β) = 0 xα−1 (1 − x)β−1 dx = Γ(α)Γ(β)
Γ(α+β) is the Beta function. The moment-generating
function of a Beta(α, β) distribution is given by a confluent hypergeometric function (also known as
Kummer’s function):
E[exp(λX)] = 1 F1 (α; α + β; λ) =
∞
X
j=0
Γ(α + j)Γ(α + β) j
λ .
(j!)Γ(α)Γ(α + β + j)
(6)
This is equivalent to say that the j th raw moment of a Beta(α, β) random variable X is given by:
E[X j ] =
(α)j
,
(α + β)j
(7)
where (x)j = x(x + 1) · · · (x + j − 1) = Γ(x+j)
Γ(x) is the Pochhammer symbol, also known in the literature
as a rising factorial. In particular, the mean and variance are given by:
E[X] =
α
,
α+β
Var[X] =
3
αβ
(α +
β)2 (α
+ β + 1)
.
The Beta distribution is ubiquitous in statistics. It plays a central role in the binomial model in
Bayesian statistics where it is a conjugate prior distribution (the associated posterior distribution is
also Beta): if X ∼ Binomial(θ, N ) and θ ∼ Beta(α, β), then θ|X ∼ Beta(α + X, β + N − X). It is
also key to Bayesian nonparametrics where it embodies, among others, the distribution of the breaks
in the stick-breaking representation of the Dirichlet process and the Pitman–Yor process; marginal
distributions of Polya trees (Castillo, 2016); the posterior distribution of discovery probabilities under
a Bayesian nonparametrics model (Arbel et al., 2017). Our main result opens new research avenues
for instance about asymptotic (frequentist) assessments of these procedures.
Our main result regarding the Beta distribution is the following:
Theorem 1 (Optimal proxy variance for the Beta distribution). For any α, β > 0, the Beta distri2
2
bution Beta(α, β) is σopt
(α, β)-sub-Gaussian with optimal proxy variance σopt
(α, β) given by:
α
1 F1 (α+1;α+β+1;x0 )
2
−
1
(α, β) = (α+β)x
σopt
F
(α;α+β;x
)
0
1 1
0
where x0 is the unique solution
of the equation
(8)
1 F1 (α+1;α+β+1;x0 )
ln(1 F1 (α; α + β; x0 )) = αx0
1+
.
2(α+β)
1 F1 (α;α+β;x0 )
2
A simple and explicit upper bound to σopt
(α, β) is given by σ02 (α, β) =
1
2
- for α 6= β we have Var[Beta(α, β)] < σopt
(α, β) < 4(α+β+1)
1
2
(α, α) = 4(2α+1)
- for α = β we have Var[Beta(α, α)] = σopt
.
1
4(α+β+1) :
Equation (8) defining x0 is a transcendental equation, the solution of which is not available in
closed form. However, it is simple to evaluate numerically. The values of the variance, optimal proxy
variance and its simple upper bound are illustrated on Figure 1. Note that for a fixed value of the
sum of the parameters, α + β = S, the optimal proxy variance deteriorates when α, or equivalently
β, gets close to 0 or to S. This is reminiscent of the Bernoulli optimal proxy variance behavior which
deteriorates when the success probability moves away from 21 (Buldygin and Moskvichova, 2013).
The intuition
of the proofcan be seen from Figure 2 (Section 2.3.3) where we represent the difference
λ 7→ exp E[X]λ +
σ2 2
2 λ
− E[exp(λX)] for various values of σ 2 . The main argument is that the
optimal proxy variance is obtained for the curve (in magenta) whose positive local minimum equals
zero, thus leading to the system of equations of Theorem 1.
Corollary 1. The Beta distribution Beta(α, β) is strictly sub-Gaussian if and only if α = β.
σ2opt, α+β=1
σ2opt
1/4
1
2
3
4
1/4
0.15
0.1
0.05
0
0
4
0
0.5
1
0
0.5
3
1
2
1
µ
µ
1
2
Figure 1: Left: curves of Var[Beta(α, β)] (green), σopt
(α, β) (purple) and 4(α+β+1)
(dotted black) for
2
the Beta(α, β) distribution with α + β set to 1, σopt (µ) for the Bern(µ) distribution (blue); varying
2
2
mean µ on the x-axis. Center : curves of σopt
(µ) for the Bern(µ) distribution (blue), and of σopt
(α, β)
for the Beta(α, β) distribution with α + β varying on a log scale from 0.1 (purple) to 10 (red); varying
2
mean µ on the x-axis. Right: surfaces of Var[Beta(α, β)] (green) and σopt
(α, β) (purple), for values
of α and β varying in [0.2, 4].
4
As a direct consequence, we obtain the strict sub-Gaussianity of the uniform, the arc-sine and the
Wigner semicircle distributions, as special cases up to a trivial rescaling of the Beta(α, α) distribution
respectively with α equal to 1, 12 and 32 .
2.2
The Beta(α, α) distribution is strictly sub-Gaussian
Let σ02 (α) = Var[Beta(α, α)] =
1
4(2α+1) .
Since a random variable X ∼ Beta(α, α) is symmetric around
λ2 σ02 (α)
1
2 , only its even centered moments are non-zero. The reason why E[exp(λ(X −E[X]))] ≤ exp
2
is because the coefficients of the series expansions at λ = 0 of each side:
∞
X
λ2j
E[exp(λ(X − E[X]))] =
,
E (X − 1/2)2j
(2j)!
j=0
2 2
X
∞
∞
X
λ σ0 (α)
σ 2j λ2j
λ2j
exp
=
=
,
j
2j
j
2
2 j!
2 2 (2α + 1)j (j!)
j=0
j=0
(9)
(10)
satisfy the inequalities:
"
E
1
X−
2
2j #
1
σ 2j (α) 1
≤ 0 j
.
(2j)!
2
j!
(11)
Indeed, algebra yields:
"
E
1
X−
2
2j #
=
(2j)! (α)j
(2j)! Γ(2α)Γ(α + j)
= 2j
.
22j j! Γ(α)Γ(2(α + j))
2 j! (2α)2j
(12)
Combining the expression of the raw moments (7) with the following inequality:
(α)j
=
(2α)2j
1
2j
j
Q
≤
(2α + 2l − 1)
1
,
(2(2α + 1))j
(13)
l=1
in (9) concludes the proof.
Remark 1. The non-symmetrical distribution with α 6= β has even centered moments whose expressions are not as simple as (12). Moreover, it has obviously non-zero odd centered moments. For this
last reason, the present proof does not carry over to the case α 6= β.
2.3
2.3.1
Optimal proxy variance for the Beta(α, β) distribution
Connection with ordinary differential equations
In this section, we assume that X ∼ Beta(α, β) with β 6= α. We denote σ02 =
dependence on α and β for compactness) and define for all t ∈ R:
σt2 =
1
4(α+β+1)
(we omit the
(1 − t)
αβ
1
(β − α)2
+t
=
+
t.
4(1 + α + β)
(α + β)2 (α + β + 1)
4(1 + α + β) 4(α + β)2 (1 + α + β)
In other words, the decreasing function t 7→ σt2 maps the interval [0, 1] to the interval [σ12 , σ02 ] with
σ12 = Var[X]. Then, we introduce the function ut defined by:
α
σt2 2
def
ut (x) = exp
x+
x − E[exp(xX)] , ∀ x ∈ R,
α+β
2
where σt2 -sub-Gaussianity amounts to non negativity of ut on R. Since the confluent hypergeometric
function y : x 7→ y(x) = 1 F1 (α, α+β; x) satisfies the linear second order ordinary differential equation
xy 00 (x) + (α + β − x)y 0 (x) − αy(x) = 0, we obtain together with equation (6) that ut is the unique
solution of the Cauchy problem:
00
0
xut (x) + (α + β − x)ut (x) − αut (x)
σ2
α
= 16(α+β)4x(1+α+β)2 exp α+β
x + 2t x2 P2 (x; t),
(14)
ut (0) = 0 and u0t (0) = 0,
5
where P2 is a polynomial of degree 2 in x:
P2 (x; t) = 4(1 − t)(α2 − β 2 )2 (1 + α + β)2
2
− 4(β 2 − α2 )(1 + α + β) (α + β)2 − t(β − α)2 x + (α + β)2 − t(β − α)2 x2 .
For normalization purposes, we also define:
α
σt2 2
vt (x) = 16(α + β) (1 + α + β) ut (x) exp −
x−
x ,
α+β
2
4
2
= 16(α + β) (1 + α + β) 1 − E[exp(xX)] exp −
4
2
α
σ2
x − t x2
α+β
2
. (15)
The function vt is the unique solution of the Cauchy problem:
xvt00 (x) + Q1 (x; t)vt0 (x) + Q0 (x; t)vt (x) = xP2 (x; t),
vt (0) = 0 and vt0 (0) = 0,
(16)
with:
β−α
(α + β)2 − t(β − α)2 2
x+
x ,
α+β
2(α + β)2 (1 + α + β)
1
xP2 (x; t).
Q0 (x; t) =
16(α + β)4 (1 + α + β)2
Q1 (x; t) = α + β −
Note that ut and vt have the same sign hence proving that ut is positive (resp. negative) is equivalent
to proving that vt is positive (resp. negative). From standard theory on ODEs (Birkhoff and Rota,
1989; Robinson, 2004), we get that the functions ut and vt are C ∞ (R). Indeed, the only possible
singularity is at x = 0 but the initial conditions imply that the function is regular at this point. In
particular, a Taylor expansion at x = 0 shows that:
vt00 (0) =
P2 (0; t)
= 4(β 2 − α2 )2 (1 + α + β)(1 − t).
1 + Q1 (0; t)
(17)
We also observe that the discriminant of the polynomial x 7→ P2 (x; t) is given by:
∆t = 16t(1 + α + β)2 (β 2 − α2 )2 (α + β)2 − t(β − α)2
2
.
Hence we conclude that for t > 0, P2 admits two distinct real zeros that are positive, while for t < 0
it remains strictly positive on R. For t = 0, P2 admits a double zero and thus remains positive on R
appart from its zero.
By definition (15), we want to study the sign of vt on R. Indeed, showing that vt is positive on R
then X is equivalent to showing σt2 -sub-Gaussianity. We first observe that we may restrict the sign
study on R+ . Indeed, if we prove that:
λα
λ2 σt2
∀ λ ≥ 0 , E[exp(λX)] = 1 F1 (α; α + β; λ) ≤ exp
+
.
(18)
α+β
2
then, the case λ < 0 is automatically obtained by noting that 1 − X ∼ Beta(β, α), whose mean is
β
α
α+β = 1 − α+β . Therefore, applying (18) to 1 − X gives that for all λ < 0:
E[exp(λX)] = exp(λ)E[exp(−λ(1 − X))]
≤ exp(λ) exp −λ(1 −
α
σ 2 λ2
)+ t
α+β
2
α
σ 2 λ2
= exp λ
+ t
α+β
2
.
Eventually, in agreement with the general theory, we observe that for t > 1 (i.e. σt2 < Var[X]), X is
not σt2 -sub-Gaussian. Indeed, the series expansion at x = 0 (17), shows that for t > 1, vt is strictly
negative in a neighborhood of 0. On the contrary, for t < 1, the function vt is strictly positive in a
neighborhood of 0 so that we may not directly conclude. Note also that for any value of t, we always
have lim vt (x) = +∞.
x→∞
6
2.3.2
Proof that the Beta(α, β) distribution is σ02 -sub-Gaussian
In this section, we take t = 0. As explained above, this corresponds to a case where P2 is positive on R
(apart from its double zero). We prove that u0 (x) > 0 for x > 0 by proceeding by contradiction. Let
us assume that there exists x1 > 0 such that u0 (x1 ) = 0. Since the non-empty set {x > 0 / u0 (x) = 0}
is compact (because u0 ∈ C ∞ (R)) and excludes a neighborhood of 0, we may define x0 = min{x >
0 such that u0 (x) = 0} > 0. Let us now define the set:
M = {0 < x < x0 such that u00 (x) = 0 and u00 changes sign at x}.
Since u0 (0) = u0 (x0 ) = 0 and the facts that u0 is strictly positive in a neighborhood of 0 and u00 is
continuous on R, Rolle’s theorem shows that M is not empty and that:
m = min{x ∈ M } exists and 0 < m < x0 .
Evaluating the ODE (14) at x = m and using the fact that the polynomial P2 is positive on R (appart
from its double zero) leads to:
mu000 (m) + 0 − αu0 (m) ≥ 0 ⇒ u000 (m) ≥
α
u0 (m) > 0.
m
(19)
However, combined with u00 (m) = 0, this contradicts the fact that u00 changes sign at x = m.
Thus, we conclude that there cannot exist x1 > 0 such that u0 (x1 ) = 0. Since u0 is strictly positive
in a neighborhood of 0 and continuous on R, we conclude that it must remain strictly positive on R∗+ .
Remark 2. In this proof, the case β = α requires an adaptation since u000 (0) = 0. Thus, we must
(4)
(3)
determine u0 (0) > 0 (u0 (0) = 0 by symmetry) to ensure that the function u0 is locally convex and
remains strictly positive in a neighborhood of 0. Apart from this minor verification, the rest of the
proof applies also to this case.
2.3.3
Proof of the optimal proxy variance for the Beta(α, β) distribution
In this section we assume that β 6= α. From general theorems regarding ODEs, we have that the
application:
[0, +∞) × [0, +∞) → R
g:
(20)
(x, t)
7→ g(x, t) = vt (x),
is smooth (g ∈ C ∞ ([0, +∞) × [0, +∞))). Indeed, the t-dependence of the coefficients of the ODE (16)
is polynomial and thus smooth. The x-dependence of the coefficients of the ODE (16) is polynomial
and as explained above, the only possible singularity in x is at x = 0 but initial conditions ensure
that the solutions x 7→ vt (x) are always regular there. Since for all t ≥ 0 we have lim vt (x) = +∞,
x→∞
we also have that the function:
[0, +∞) → R
h:
t
7→ min{vt (x), x ∈ R∗+ },
is continuous.
We now observe that for any 0 ≤ t < 1, the functions vt are strictly positive in a neighborhood
of 0. More precisely, if we choose a segment [0, t0 ] with t0 < 1, then for all 0 ≤ t ≤ t0 we have that
vt00 (0) ≥ vt000 (0) = 4(β 2 − α2 )2 (1 + α + β)(1 − t0 ) > 0. Hence, we may choose η > 0 such that for all
0 ≤ t ≤ t0 , we have vt is strictly positive on ]0, η]. Moreover, since lim v0 (x) = +∞, v0 is bounded
x→∞
from below on [η, +∞[ by a constant A > 0. Thus, since g is continuous, there exists a neighborhood
of t = 0 in which all solutions vt remain greater than A2 on [η, +∞[ and thus strictly positive on R∗+ .
This shows that for t > 0, σ02 is not optimal.
Let us now introduce the set:
T+ = {t ≥ 0 such that vt is positive on R∗+ }.
Then, from the results presented above, we know that T+ is non-empty, that it contains a neighborhood of 0 and that it is bounded from above by 1. Moreover, by connection with the initial problem
(15), T+ is an interval and thus is of the form [0, topt ] with 0 < topt < 1. Indeed t ∈ T+ implies
7
by construction that for all s ≤ t, s ∈ T+ . Note also that t < 1 since v1 is strictly negative in a
neighborhood of 0 and thus min{v1 (x), x ∈ R∗+ } < 0. Hence, the continuity of h shows that there
exists a neighborhood of t = 1 in which the solutions vt are non-positive on R∗+ . For t = topt the
function vtopt must have a zero on R∗+ otherwise by continuity of h we may find a neighborhood
of topt for which min{vt (x), x ∈ R∗+ } remains strictly positive thus contradicting the maximality of
topt . Since vtopt must remain positive, the zero is at least a double zero and therefore we find that
there exists x0 > 0 such that vtopt (x0 ) = 0, vt0 opt (x0 ) = 0 and vt00opt (x0 ) ≥ 0. From (6) and (15), the
conditions vtopt (x0 ) = 0, vt0 opt (x0 ) = 0 are equivalent to the following system of equations (we use
here the contiguous relations for the confluent hypergeometric function: 1 F10 (a; b; x) = ab 1 F1 (a; b; x)):
σt2opt
α
2
1 F1 (α; α + β; x0 ) = exp
α+β x0 +
2 x0 ,
(21)
α
2
α 1 F1 (α + 1; α + β + 1; x0 ) =
+
σ
x
F
(α;
α
+
β;
x
).
1 1
0
topt 0
α+β
α+β
This is equivalent to say that x0 ≡ x0 (α, β) is the solution of the transcendental equation:
αx0
1 F1 (α + 1; α + β + 1; x0 )
1+
,
1 F1 (α; α + β; x0 ) = exp
2(α + β)
1 F1 (α; α + β; x0 )
and that σt2opt is given by:
σt2opt =
α
(α + β)x0
+ 1; α + β + 1; x0 )
−1 .
1 F1 (α; α + β; x0 )
1 F1 (α
Note that by symmetry, we have x0 (β, α) = −x0 (α, β) hence, σt2opt (β, α) = σt2opt (α, β). Moreover, if
β > α then x0 (α, β) > 0 while α > β implies x0 (α, β) < 0. We may illustrate the situation with
Figure 2 which displays the difference function x 7→ ut (x).
Remark 3. The system of equations (21) admits only one solution on R∗+ . Indeed, let us transpose
the problem from vtopt to utopt using (15) and assume that there exist two points 0 < x0 < x1 such that
utopt (x0 ) = 0, u0topt (x0 ) = 0 and utopt (x1 ) = 0, u0topt (x1 ) = 0 with utopt strictly positive on (x0 , x1 ) (hence
u00topt (x0 ) ≥ 0 and u00topt (x1 ) ≥ 0). Using (14), this implies that P2 (x0 ; topt ) ≥ 0 and P2 (x1 ; topt ) ≥ 0.
If we denote x− < x+ the potential distinct positive zeros of x 7→ P2 (x; topt ) we may exclude that
x0 ≤ x− . Indeed, if x0 ≤ x− then we may apply the same argument to utopt on the interval [0, x0 ]
as the one developed for u0 in Section 2.3.2 and obtain a contradiction. Thus, the only remaining
case is to assume x1 > x0 > x+ . In that case, since utopt (x0 ) = 0, u0topt (x0 ) = 0, utopt (x1 ) = 0 and
x 7→ P2 (x; topt ) is positive on [x0 , x1 ], we may apply the same argument to utopt on the interval [x0 , x1 ]
as the one developed for u0 in Section 2.3.2 and obtain a contradiction.
3
Relations to other distributions
3.1
Optimal proxy variance for the Bernoulli distribution
We show that our proof technique can be used to recover the optimal proxy variance for the Bernoulli
distribution, known since Kearns and Saul (1998). This is illustrated by the center panel of Figure 1.
Theorem 2 (Optimal proxy variance for the Bernoulli distribution). For any µ ∈ (0, 1), the Bernoulli
2
distribution with mean µ is sub-Gaussian with optimal proxy variance σopt
(µ) given by:
2
σopt
(µ) =
Proof. In the limit α → 0 with
u00t,µ (x)
α
α+β
(1 − 2µ)
.
2 ln 1−µ
µ
(22)
fixed equal to µ, the differential equation (14) simplifies into:
x2
x2 (2µ − 1)2 t
− ut,µ (x) = exp µx +
−
8
8
2
(2µ − 1) (1 − t) (2µ − 1)(1 − t + 4tµ(1 − µ))
(1 − t + 4tµ(1 − µ))2 2
+
x−
x
4
4
16
8
2
x10−3
0
1
u0
utopt
utnon opt
u1
●
−1
x0
0
1
2
3
Figure 2: Difference function x 7→ ut (x). For t = 0 (simple upper bound σ02 ), the curve [dotted
2
black] remains strictly positive. For t = topt (optimal proxy variance σopt
), the curve [magenta] has
zero minimum (at x0 ). For t = 1 (leading to the variance), the curve [dashed green] has negative
second derivative at x = 0, hence is directly negative around 0. The intermediate case with tnon opt
in the interval (topt , 1) produces a curve [orange, dash and dots] which is first positive, then negative,
and positive again.
with the Cauchy initial conditions ut,µ (0) = 0 and u0t,µ (0) = 0. The solution of this Cauchy problem
is explicit and given by:
x2 (2µ − 1)2 t
x2
−
ut,µ (x) = exp µx +
− µex + µ − 1
(23)
8
8
2
(µ) = 14 − 14 x(2µ − 1)2 t0 where t0 is determined
Therefore the optimal proxy variance is given by σopt
0
by the system of equations: ut0 ,µ (x0 ) = 0 and ut0 ,µ (x0 ), thus defining implicitly t0 and x0 as functions
of µ. Inorder to solve
explicitly the last system of equations, we perform the change of variables:
s
2t̃
(µ, t) = s+1 , s+1 + 1 so that the solution (23) is now given by:
ut̃,s (x) = exp
s
t̃
x−
x2
s+1
4(s + 1)
−
s x
1
e −
s+1
s+1
Consequently, we have to solve the system:
(s + 1) exp s x0 − t̃0 x20 = sex0 + 1
ut̃0 ,s (x0 ) = 0
4(s+1)
s+1
⇔
u0t̃0 ,x (x0 ) = 0
s − x0 t̃0 exp s x − t̃0 x2 = sex0
s+1 0
2
4(s+1) 0
Introducing another change of variable (x0 , y0 ) = (x0 , x0 t̃0 ), the last system is equivalent to:
(
(
x0
y0 +2
s − y0 = s(1 +
y
)e
0
s
−
y
=
s
exp
ln(1
+
y
)
0
0
y0 −2s
⇔
x0
s − y20
1 + y0 = exp − s+1
x0 = ys+1
0 −s ln(1 + y0 )
2
We now observe that y0 = s − 1 and x0 = −2 ln s is a solution of the former system. Performing back
4(1−s)
the various changes of variables, this is equivalent to say that t̃0 = xy00 = 2(1−s)
ln s so that t0 = (s+1) ln s +1
1
2
or equivalently t0 = (2µ−1)
. Consequently, the optimal proxy variance is given by:
2 +
(2µ−1) ln 1−µ
µ
2
σopt
(µ) =
1 1
(1 − 2µ)
− (2µ − 1)2 t0 =
4 4
2 ln 1−µ
µ
which is precisely the optimal proxy variance of a Bernoulli random variable with mean µ.
9
3.2
Optimal proxy variance for the Dirichlet distribution
We start by reminding the definition of sub-Gaussian property for random vectors:
Definition 2 (Sub-Gaussian vectors). A random d-dimensional vector X with finite mean µ = E[X]
is σ 2 -sub-Gaussian if the random variable u> X is σ 2 -sub-Gaussian for any unit vector u in the
d
P
simplex S d−1 = {u ∈ [0, 1]d / ui = 1}. This is equivalent to say that:
i=1
>
E[exp(λ (X − µ))] ≤ exp
d
P
where kλk2 =
i=1
kλk2 σ 2
2
for all λ ∈ Rd .
λ2i . Eventually, a random vector X is said to be strictly sub-Gaussian, if the random
variables u> X are strictly sub-Gaussian for any unit vector u ∈ S d−1 .
Let d ≥ 2. The Dirichlet distribution Dir(α), with positive parameters α = (α1 , . . . , αd )> , is
characterized by a density on the simplex S d−1 given by:
d
1 Y αi −1
f (x1 , . . . , xd ; α1 , . . . , αd ) =
x
,
B(α) i=1 i
where B(α) =
1
Γ(ᾱ)
d
Q
Γ(αi ) and ᾱ =
i=1
d
P
αi . It generalizes the Beta distribution in the sense that the
i=1
components are Beta distributed. More precisely, for any non-empty and strict subset I of {1, . . . , d}:
X
X X
X = (X1 , . . . , Xd )> ∼ Dir(α) =⇒
Xi ∼ Beta αi ,
αj .
i∈I
i∈I
j ∈I
/
However, we remind the reader that the components (Xi )1≤i≤d are not independent and the variance/covariance matrix is given by:
∀ i 6= j : Cov[Xi , Xj ] = −
αi (ᾱ − αi )
αi αj
and ∀ 1 ≤ i ≤ d , Var[Xi ] = 2
.
ᾱ2 (1 + ᾱ)
ᾱ (1 + ᾱ)
Eventually, if we define n = (n1 , . . . , nd )> ∈ Nd , then the moments of the Dirichlet distribution are
given by:
" d
#
Y
B(α + n)
ni
=
E
Xi
.
B(α)
i=1
This is equivalent to say that the moment-generating function of the Dirichlet distribution is:
E[exp(λ> X)] =
∞
X
X
m=0 n1 +···+nd
=
∞
X
X
m=0 n1 +···+nd
=
∞
X
X
m=0 n1 +···+nd
B(α + n) λn1 1 . . . λnd d
,
B(α) (n1 )! . . . (nd )!
=m
d
λn1 1 . . . λnd d
Γ(ᾱ) Y Γ(αi + ni )
,
(n1 )! . . . (nd )! Γ(ᾱ + n̄) i=1 Γ(αi )
=m
d
λn1 1 . . . λnd d
1 Y
(αi )ni ,
(n1 )! . . . (nd )! (ᾱ)n̄ i=1
=m
where we have defined λ = (λ1 , . . . , λd )> ∈ Rd and n̄ =
d
P
ni .
i=1
Let us define ei the ith canonical vector of Rd and X = (X1 , . . . , Xd )> ∼ Dir(α). From Definition 1
and the results regarding the Beta(α, β) distribution obtained in Section 2.3, we immediately get that
def
2
2
2
e>
i X = Xi is σi -sub-Gaussian with σi = σopt (αi , ᾱ − αi ) defined from Theorem 1. Moreover, in
2
direction ei , σi is the optimal proxy variance. Therefore, the remaining issue is to generalize these
results for arbitrary unit vectors on S d−1 . We obtain the following result:
10
Theorem 3 (Optimal proxy variance for the Dirichlet distribution). For any parameter α, the Dirich2
let distribution Dir(α) is sub-Gaussian with optimal proxy variance σopt
(α) given from Theorem 1
and:
2
2
σopt
(α) = σopt
(αmax , ᾱ − αmax ) where αmax = max {αi }.
1≤i≤d
2
σopt
(αi , βi
Proof. We first observe that the computations of
= ᾱ − αi ) correspond to cases where the
2
sum αi + βi is fixed to ᾱ and thus independent of i. Therefore, σopt
(αi , ᾱ − αi ) is maximal when
1
|(ᾱ − αi ) − αi | is minimal, i.e. when the distance from αi to 2 ᾱ is minimal. It is easy to see that
this corresponds to choosing αi = αmax = max{αi , 1 ≤ i ≤ d} (by looking at the two possible cases
αmax ≤ 21 ᾱ and αmax > 12 ᾱ).
2
We then observe that σmax
cannot be improved. Indeed, let us denote i0 one of the components
for which the maximum is obtained. Then, if we take u = ei0 , the discussion presented above shows
2
is the optimal proxy variance in this direction. Hence the optimal proxy variance
that σi20 = σmax
2
cannot be lower than σmax
.
2
Let us now prove that X is σmax
-sub-Gaussian. Let u = (u1 , . . . , ud )> be a unit vector on S d−1 and
λ ∈ R. We define for clarity λ = λu. We have:
"
!#
d
h
i
X
>
>
E exp λu X = E exp λ X = E exp
λi Xi
,
i=1
=
∞
X
X
m=0 n1 +···+nd
d
λn1 1 . . . λnd d
1 Y
(αi )ni . (24)
(n1 )! . . . (nd )! (ᾱ)n̄ i=1
=m
Note that we also have:
∞
X
(α
)
i j
E [exp(λi Xi )] =
λji
(j!)(
ᾱ)
j
i=1
i=1
j=0
d
Y
d
Y
=
∞
X
X
m=0 n1 +···+nd
d
λn1 1 . . . λnd d Y (αi )ni
.
(n1 )! . . . (nd )! i=1 (ᾱ)ni
=m
(25)
Moreover we have the inequality:
d
Y
(ᾱ)ni ≤ (ᾱ)n̄ ,
i=1
because both sides have the same number of terms in the product (i.e. n̄) but those of the right hand
side are always greater or equal to those of the left hand side. Hence, from (24) and (25), we find:
d
h
i Y
E exp λ> (X − µ) ≤
E [exp(λi (Xi − µi ))] .
i=1
Using the optimal proxy variance of the Beta distribution proven in Theorem 1, we find:
2 2
d
d
h
i Y
Y
λ i σi
>
E exp λ (X − µ) ≤
E [exp(λi (Xi − µi ))] ≤
exp
2
i=1
i=1
2 2
2
d
Y
λi σmax
σmax kλk2
≤
exp
= exp
,
2
2
i=1
2
thus showing that X is σmax
-sub-Gaussian and concluding the proof.
Note that using Theorem 3, we obtain the following corollary:
Corollary 2. For any integer d ≥ 2, the Dirichlet distribution Dir(α1 , . . . , αd ) is strictly sub-Gaussian
if and only if d = 2 and α1 = α2 .
def
Indeed, we first need to require α1 = · · · = αd = α so that all directions have the same optimal
proxy variance. Then, each component satisfies Xi = e>
i X ∼ Beta(α, (d − 1)α) and Theorem 1 shows
that σi2 is the optimal proxy variance for Xi if and only if α = (d − 1)α, i.e. if and only if d = 2.
11
Acknowledgments
We wish to thank Stéphane Boucheron for an enlightening discussion of our results and an anonymous
referee for insightful suggestions. O.M. would like to thank Université de Lyon, Université Jean Monnet and Institut Camille Jordan for financial support. This work was supported by the LABEX MILYON (ANR-10-LABX-0070) of Université de Lyon, within the program “Investissements d’Avenir”
(ANR-11-IDEX-0007) operated by the French National Research Agency (ANR). J.A. would like to
thank Inria Grenoble Rhône-Alpes and Laboratoire Jean Kuntzmann, Université Grenoble Alpes for
financial support. J.A. is also member of Laboratoire de Statistique, CREST, Paris. This work was
partially conducted during a scholar visit of J.A. at the Department of Statistics & Data Science of
the University of Texas at Austin.
References
Arbel, J., Favaro, S., Nipoti, B., and Teh, Y. W. (2017). Bayesian nonparametric inference for
discovery probabilities: credible intervals and large sample asymptotics. Statistica Sinica, 27:839–
858.
Ben-Hamou, A., Boucheron, S., and Ohannessian, M. I. (2017). Concentration inequalities in the
infinite urn scheme for occupancy counts and the missing mass, with applications. Bernoulli,
23(1):249–287.
Berend, D. and Kontorovich, A. (2013). On the concentration of the missing mass. Electronic
Communications in Probability, 18(3):1–7.
Birkhoff, G. and Rota, G.-C. (1989). Ordinary Differential Equations. John Wiley and Sons Editions.
Bobkov, S. G. and Götze, F. (1999). Exponential integrability and transportation cost related to
logarithmic sobolev inequalities. Journal of Functional Analysis, 163(1):1–28.
Boucheron, S., Lugosi, G., and Massart, P. (2013). Concentration inequalities: A nonasymptotic
theory of independence. Oxford University Press.
Buldygin, V. V. and Kozachenko, Y. V. (1980). Sub-Gaussian random variables. Ukrainian Mathematical Journal, 32(6):483–489.
Buldygin, V. V. and Kozachenko, Y. V. (2000). Metric characterization of random variables and
random processes, volume 188. American Mathematical Society, Providence, Rhode Island.
Buldygin, V. V. and Moskvichova, K. (2013). The sub-Gaussian norm of a binary random variable.
Theory of probability and mathematical statistics, 86:33–49.
Castillo, I. (2016). Pólya tree posterior distributions on densities. Annales de l’Institut Henri
Poincaré, to appear.
Elder, S. (2016). Bayesian adaptive data analysis guarantees from subgaussianity. arXiv preprint:
arXiv:1611.00065.
Gross, L. (1975). Logarithmic sobolev inequalities. American Journal of Mathematics, 97(4):1061–
1083.
Hoeffding, W. (1963). Probability inequalities for sums of bounded random variables. Journal of the
American statistical association, 58(301):13–30.
Kearns, M. and Saul, L. (1998). Large deviation methods for approximate probabilistic inference. In
Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence, pages 311–319.
Ledoux, M. (1999). Concentration of measure and logarithmic sobolev inequalities. Lecture notes in
mathematics - Springer Verlag-, pages 120–216.
McAllester, D. A. and Ortiz, L. (2003). Concentration inequalities for the missing mass and for
histogram rule error. Journal of Machine Learning Research, 4:895–911.
12
McAllester, D. A. and Schapire, R. E. (2000). On the convergence rate of Good-Turing estimators.
In COLT, pages 1–6.
Ordentlich, E. and Weinberger, M. J. (2005). A distribution dependent refinement of pinsker’s inequality. IEEE Transactions on Information Theory, 51(5):1836–1840.
Perry, A., Wein, A. S., and Bandeira, A. S. (2016). Statistical limits of spiked tensor models. arXiv
preprint: arXiv:1612.07728.
Pisier, G. (2016). Subgaussian sequences in probability and Fourier analysis.
arXiv:1607.01053.
arXiv preprint:
Raginsky, M. and Sason, I. (2013). Concentration of measure inequalities in information theory,
communications, and coding. Foundations and Trends in Communications and Information Theory,
10(1-2):1–246.
Robinson, J. C. (2004). An introduction to Ordinary Differential Equations. Cambridge University
Press.
13
| 10math.ST
|
A theory of passive linear systems with no assumptions
Timothy H. Hughes a,?,??
arXiv:1611.06140v3 [cs.SY] 22 Jan 2018
a
Department of Mathematics, University of Exeter, Penryn Campus, Penryn, Cornwall, TR10 9EZ, UK
Abstract
We present two linked theorems on passivity: the passive behavior theorem, parts 1 and 2. Part 1 provides necessary and
sufficient conditions for a general linear system, described by a set of high order differential equations, to be passive. Part 2
extends the positive-real lemma to include uncontrollable and unobservable state-space systems.
Key words: Passive system; Positive-real lemma; Linear system; Controllability; Observability; Behavior.
1
Introduction
A system is called passive if there is an upper bound on
the net energy that can be extracted from the system
from the present time onwards. This is a fundamental
property of many physical systems. In systems and control theory, the concept of passivity has its origins in the
study of electric networks comprising resistors, inductors, capacitors, transformers, and gyrators (RLCTG
networks). In contemporary systems theory, passive systems are more familiar through their role in the positivereal lemma. This lemma proves the equivalence of: (i) an
integral condition related to the energy exchanged with
the system; (ii) a condition on the transfer function for
the system (the positive-real condition); and (iii) a linear
matrix inequality involving the matrices in a state-space
realization for the system. As well as being relevant to
passive systems, the lemma also gives necessary and sufficient conditions for the existence of non-negative definite solutions to an important linear matrix inequality
and algebraic Riccati equation, and has links with spectral factorisation. However, these results are all subject
to one caveat: the system is assumed to be controllable.
? A simpler version of Theorem 13 in this paper, for singleinput single-output systems, was presented at the European
Control Conference, Aalborg, 2016 (see Hughes, 2016b).
?? c
2017. This manuscript version is made
available under the CC-BY-NC-ND 4.0 license
http://creativecommons.org/licenses/by-nc-nd/4.0/. This
is the accepted version of the manuscript: Hughes, T.H.:
A theory of passive linear systems with no assumptions,
Automatica, 86, 87-97 (2017).
Email address: [email protected] (Timothy H.
Hughes ).
Preprint submitted to Automatica
As emphasised by Çamlibel et al. (2003); Willems
(2007); Hughes and Smith (2017), there is no explicit
connection between the concepts of passivity and controllability. Moreover, the a-priori assumption of controllability in the positive-real lemma leaves open several
questions of physical significance. In particular, it is not
known what uncontrollable behaviors can be realized as
the driving-point behavior of an electric (RLCTG) network. Similarly, necessary and sufficient conditions for
the existence of a non-negative definite solution to the
linear matrix inequality (and algebraic Riccati equation) considered in the positive-real lemma are unknown
when the state-space realization under consideration
is uncontrollable. There have been many papers in the
literature that have aimed to relax the assumption of
controllability in the positive-real lemma, e.g., Pandolfi
(2001); Collado et al. (2001); Kunimatsu et al. (2008)
(and many papers have studied uncontrollable cyclodissipative systems, e.g., Ferrante and Pandolfi (2002);
Çamlibel et al. (2003); Ferrante (2005); Pal and Belur
(2008)), but all of these papers contain other a-priori
assumptions. The objective of this paper is to provide
a complete theory of passive linear systems with no superfluous assumptions. Our main contributions are: 1. a
new trajectory-based definition of passivity (Definition
5); and 2. two linked theorems that we call the passive
behavior theorem, parts 1 and 2. Part 1 (Theorem 9) provides necessary and sufficient conditions for the passivity of a general linear system (described by a differential
d
d
)i = Q( dt
)v for some square
equation of the form P ( dt
polynomial matrices P and Q). This generalizes classical results that are restricted to controllable behaviors
(where P and Q are left coprime). Part 2 (Theorem
13) extends the positive-real lemma by removing the apriori controllability and observability assumptions. As
24 January 2018
all t ∈ R. We also consider the function space
a corollary of these results, we find that any passive (not
necessarily controllable) behavior can be realized as the
driving-point behavior of an electric (RLCTG) network
EC−
N nX
i −1
X
R, R :={w | w(t)=<
w̃ij tj eλi t for all t∈R
k
i=1 j=0
with w̃ij ∈ Ck , λi ∈ C− , and N, ni integers},
The structure of the paper is as follows. In Section 2, we
discuss the positive-real lemma and its limitations. Section 3 discusses our new definition of passivity. Then, in
Section 4, we introduce the new concept of a positive-real
pair, and we state our two passive behavior theorems. It
is shown that our new concept of a positive-real pair provides the appropriate extension of the positive-real concept to uncontrollable systems. Specifically, for any pair
of square polynomial matrices P and Q, we show that the
system corresponding to the solutions to the differential
d
d
equation P ( dt
)i = Q( dt
)v is passive if and only if (P, Q)
is a positive-real pair. The proofs of the passive behavior
theorems are in Section 6, and some preliminary results
appear in Section 5. Finally, the paper is strongly influenced by the behavioral approach to dynamical systems
(see Polderman and Willems, 1998). Therefore, to make
the paper accessible to the reader unfamiliar with behavioral theory, we provide four short appendices containing relevant background on linear systems, behaviors, and polynomial matrices. These contain numbered
notes (A1, A2, and so forth) that will be referred to in
the text. The reader who wishes to follow the proofs in
Sections 5 and 6 is advised to first read these appendices.
and note that EC− R, Rk ⊂ C∞ R, Rk ⊂ Lloc
R, Rk .
1
We consider behaviors (systems) defined as the set of
weak solutions to a linear differential equation:
d
)w=0}, R ∈ Rl×k [ξ]. (1.1)
B={w ∈ Lloc
R, Rk | R( dt
1
Here, if R(ξ) = R0 + R1 ξ + . . . + RL ξ L and w ∈
dL w
d
)w = R0 w +R1 dw
C∞ R, Rk , then R( dt
dt +. . .+RL dtL
(see Polderman and Willems, 1998, Definition 2.3.7 for
d
the meaning of a weak solution to R( dt
)w = 0 when w
is not necessarily differentiable). Particular attention is
paid to the special class of state-space systems:
n
loc
n
loc
Bs ={(u, y, x) ∈ Lloc
R, Rd
1 (R, R ) ×L1 (R, R ) ×L1
such that
with A ∈ R
d×d
dx
dt
= Ax + Bu and y = Cx + Du},
, B ∈ Rd×n , C ∈ Rn×d , D ∈ Rn×n . (1.2)
Several properties of state-space systems are listed in
Appendix D. In particular, from note D1, if (u, y, x) ∈
Bs , then x satisfies the variation of the constants formula
almost everywhere, which determines the value x(t1 ) of x
at an instant t1 ∈ R. Finally, we also consider behaviors
obtained by permuting and/or eliminating variables in a
behavior B as in (1.1). For example, associated with the
state-space system Bs in (1.2) is the corresponding exter(u,y)
= {(u, y) | ∃x with (u, y, x) ∈ Bs }.
nal behavior Bs
More generally, for any given T1 ∈ Rl1 ×k , . . . , Tn ∈
Rln ×k such that col(T1 · · · Tn ) ∈ Rk×k is a permutation matrix, and integer 1 ≤ m ≤ n, we denote the projection of B onto T1 w, . . . , Tm w by
The notation is as follows. R (C) denotes the real (complex) numbers; C+ (C+ ) denotes the open (closed)
right-half plane; C− (C− ) denotes the open (closed)
left-half plane. R[ξ] (R(ξ)) denotes the polynomials (rational functions) in the indeterminate ξ with real coefficients. Rm×n (resp., Cm×n , Rm×n [ξ], Rm×n (ξ)) denotes
the matrices with m rows and n columns with entries
from R (resp., C, R[ξ], R(ξ)), and the number n is omitted whenever n = 1. If H ∈ Cm×n , then <(H) (=(H))
denotes its real (imaginary) part, and H̄ its complex
conjugate. If H ∈ Rm×n , Cm×n , Rm×n [ξ] or Rm×n (ξ),
then H T denotes its transpose; and if H is nonsingular (i.e., det(H) 6≡ 0), then H −1 denotes its inverse.
We let col(H1 · · · Hn ) (diag(H1 · · · Hn )) denote the
block column (block diagonal) matrix with entries
H1 , . . . , Hn . If M ∈ Cm×m , then M > 0 (M ≥ 0) indicates that M is Hermitian positive (non-negative) definite, and spec(M ) := {λ ∈ C | det(λI−M ) = 0}. If G ∈
Rm×n (ξ), then normalrank(G) := maxλ∈C (rank(G(λ))),
G? (ξ) := G(−ξ)T , G is called para-Hermitian if G = G? ,
and proper if limξ→∞ (G(ξ)) exists. Lloc
R, Rk and
1
C∞ R, Rk denote the (k-vector-valued) locally integrable and infinitely-often differentiable functions (Polderman and Willems, 1998, Definitions 2.3.3, 2.3.4). We
equate any two locally integrable functions that differ
only on a set of measure zero. If w ∈ Lloc
R, Rk , then
1
T
T
w denotes the function satisfying w (t) = w(t)T for
B (T1 w,...,Tm w) = {(T1 w, . . . , Tm w) | ∃(Tm+1 w, . . . , Tn w)
such that w ∈ B}.
2
The positive-real lemma
The central role of passivity in systems and control is
exemplified by the positive-real lemma (see Lemma 1).
The name positive-real (PR) describes a function G ∈
Rn×n (ξ) with the properties: (i) G is analytic in C+ ;
and (ii) G(λ̄)T + G(λ) ≥ 0 for all λ ∈ C+ (see Anderson and Vongpanitlerd, 1973, Theorem 2.7.2 for a well
known equivalent condition). The positive-real lemma
then considers a state-space system as in (1.2) and provides necessary and sufficient conditions for the transfer
function G(ξ) = D + C(ξI−A)−1 B to be PR. Notably,
2
for all ω ∈ R ∪ ∞; and (ii) the existence of a real symmetric X ≥ 0 such that Π(X) = 0 and spec(A + B(D +
DT )−1 (B T X − C)) ∈ C− (Zhou et al., 1996, Corollary 13.27). Third, if spec(A) ∈ C− , then condition 3 in
Lemma 1 is equivalent to condition 4 together with the
additional condition (Pandolfi, 2001, equation (4)) (this
condition will be discussed in Remark 22).
it is assumed that (A, B) is controllable and (C, A) is
observable (see notes D2 and D4).
Lemma 1 (Positive-real lemma) Let Bs be as in
(1.2) and let (A, B) be controllable and (C, A) observable. Then the following are equivalent:
1. Given any x0 ∈ Rd , there exists
Z tS1 a (x0 ) ∈ R with
T
Sa (x0 ) :=
sup
−
u (t)y(t)dt .
t1 ≥t0 ∈R, (u,y,x)∈Bs
with x(t0 )=x0
2.
sup
t1 ≥t0 ∈R, (u,y,x)∈Bs
with x(t0 )=0
Z
−
t1
Nevertheless, the results in these references, and other
similar results in the literature (e.g., Collado et al., 2001;
Kunimatsu et al., 2008), do not cover several important systems. In particular, they do not consider systems
whose transfer functions possess imaginary axis poles.
We consider one such system in Example 4. Other important examples include conservative systems, whose
transfer functions are lossless PR (see Anderson and
Vongpanitlerd, 1973, Chapter 2).
t0
uT (t)y(t)dt = 0 .
t0
3. There exist real matrices X, LX , WX such that X >
T
0, −AT X − XA = LTX LX , C − B T X = WX
LX ,
T
T
and D + D = WX WX .
4. G(ξ) := D + C(ξI−A)−1 B is PR.
Example 4 Let Bs be as in (1.2) with
If, in addition, D + DT > 0, then the above conditions
are equivalent to:
A=
5. There exists a real X > 0 such that Π(X) :=
−AT X−XA−(C T −XB)(D+DT )−1 (C−B T X) =
0 and spec(A + B(D + DT )−1 (B T X − C)) ∈ C− .
h0
0 1
0 0 1
0 −1 0
i
, B=
h1i
0
0
, C = [1 1 0], and D = 1.
Here, (A, B) is not controllable. We now show that conditions 2 and 4 of Lemma 1 hold for this example, yet
condition 1 does not. First, direct calculation verifies
that G(ξ) = 1 + 1/ξ, and so condition 4 is satisfied.
Second, from the variation of the constants formula (see
note D1), y(t) = x1 (t0 ) + (2 cos(t − t0 ) − 1)x2 (t0 ) +
2 sin(t − t0 )x3 (t0 ) + u(t) + ∫tt0 u(τ )dτ for all t ≥ t0 .
Hence, if x(t0 ) = 0 and t1 ≥ t0 , then ∫tt01 u(t)y(t)dt =
∫tt01 u2 (t)dt + 21 (∫tt01 u(τ )dτ )2 ≥ 0, and so condition 2 is
satisfied. Third, with x1 (0) = x2 (0) = 0, x3 (0) = −1,
and u(t) = sin(t) for all t ≥ 0, then y(t) = − sin(t) −
cos(t) for all t ≥ 0. Thus, for any given positive integer n,
− ∫0nπ y(t)u(t)dt = ∫0nπ sin2 (t)dt + ∫0nπ sin(t) cos(t)dt =
1
2 nπ. It follows that condition 1 does not hold. Furthermore, it will follow from Theorem 13 of this paper that
condition 3 of Lemma 1 does not hold for this system.
For a proof of the positive-real lemma, we refer to
Willems (1972b); Anderson and Vongpanitlerd (1973).
These references also describe links with spectral factorization, which is the concern of the following well
known result (Youla, 1961, Theorem 2):
Lemma 2 (Youla’s spectral factorisation result)
Let H ∈ Rn×n (ξ) be para-Hermitian; let H(jω) ≥ 0 for
all ω ∈ R, ω not a pole of H; and let normalrank(H) = r.
There exists a Z ∈ Rr×n (ξ) such that (i) H = Z ? Z;
(ii) Z is analytic in C+ ; and (iii) Z(λ) has full row
rank for all λ ∈ C+ . Moreover, if H ∈ Rn×n [ξ], then
Z ∈ Rr×n [ξ]; if H(jω) is analytic for all ω ∈ R, then Z
is analytic in C+ ; and if Z1 ∈ Rr×n (ξ) also satisfies (i)–
(iii), then there exists a T ∈ Rr×r such that Z1 = T Z
and T T T = I. We call any Z ∈ Rr×n (ξ) that satisfies
(i)–(iii) a spectral factor of H.
One of the main contributions of this paper is a generalization of the positive-real lemma to include state-space
systems that are not necessarily controllable or observable (Theorem 13). In contrast to other papers on this
subject, we do not introduce any superfluous assumptions. However, as we will argue in the next section, a
state-space system is not a natural starting point for the
study of passive systems. Thus, a second major contribution of this paper is a necessary and sufficient condition
for the passivity of a general linear system, described by
a set of high order differential equations (Theorem 9).
Remark 3 When G is as in Lemma 1 with D +DT > 0,
T
there exists WX ∈ Rn×n with D + DT = WX
WX . Then,
with X as in condition 5 of Lemma 1, it can be shown
T −1
that ZX (ξ) := WX + (WX
) (C − B T X)(ξI−A)−1 B is
a spectral factor of G + G? (see Willems, 1972b).
The assumptions in Lemma 1 can be relaxed in three particularly notable ways. First, from (Willems, 1971, Theorems 1, 3, 8), conditions 1–4 of Lemma 1 are equivalent
even if (C, A) is not observable, but X may then be singular in condition 3. Second, the following are equivalent
irrespective of whether (A, B) is controllable or (C, A) is
observable: (i) spec(A) ∈ C− and G(−jω)T +G(jω) > 0
3
Passivity
The concept of passivity is relevant to systems
whose variables can be partitioned into two sets
n
loc
n
i ∈ Lloc
1 (R, R ) and v ∈ L1 (R, R ) with the property
3
that − ∫tt01 iT (t)v(t)dt is the net energy extracted from
the system in the interval from t0 to t1 . Passivity has
its origins in the study of electric RLCTG networks,
for which i represents the driving-point currents and
v the corresponding driving-point voltages. As shown
in Hughes (2017a), for any given RLCTG network, the
driving-point currents and voltages are related by a
linear differential equation of the form:
that this is consistent with Definition 5 when considering
systems with a state-space realization as in (1.2), where
i = u and v = y. However, as mentioned earlier, there
are systems that are passive in the sense of Definition
5 that cannot be represented in this form. Specifically,
as will be shown in Lemma 12, condition 1 of Lemma 1
only applies to systems of the form:
n
loc
n
B̃ = {(u, y) ∈ Lloc
1 (R, R ) × L1 (R, R ) |
n
loc
n
B = {(i, v) ∈ Lloc
1 (R, R ) × L1 (R, R ) |
d
d
P̃ ( dt
)u = Q̃( dt
)y, where P̃ , Q̃ ∈ Rn×n [ξ],
d
d
P ( dt
)i = Q( dt
)v, for some P, Q ∈ Rn×n [ξ]}. (3.1)
Q̃ is nonsingular, and Q̃−1 P̃ is proper}. (3.2)
Note that (i, v) need not be an input-output partition in
the sense of Polderman and Willems (1998). For example: (i) Q is singular for a transformer; 1 and (ii) Q−1 P
is not proper for an inductor. 2 Yet it is common for passivity to be defined for systems described using a statespace or input-output representation. This implies assumptions that (i) Q is nonsingular; and (ii) Q−1 P is
proper. Accordingly, we provide a new definition of passivity for the general system in (3.1) that does not depend on such assumptions. Note that this definition extends naturally to non-linear and time-varying systems.
Thus, this condition does not cover systems of the form of
(5) for which either Q is singular or Q−1 P is not proper.
Definition 5 is similar to a definition for dissipativity proposed in (Willems, 2007, Section 8) and used by Hughes
and Smith (2017) (note that it is straightforward to generalize Definition 5 to the framework of dissipative systems). In Hughes and Smith (2017), the system B in (3.1)
was called passive if, given any (i, v) ∈ B and any t0 ∈ R,
there exists a K ∈ R (dependent on (i, v) and t0 ) such
that − ∫tt01 iT (t)v(t)dt < K for all t1 ≥ t0 . Evidently, if
B in (3.1) is passive in the sense of Definition 5, then B
is also passive in the sense of Willems (2007); Hughes
and Smith (2017). It can also be shown that the converse is true. 3 However, Definition 5 is a more accurate
statement of the physical property of passivity (when extended to time-varying and non-linear systems), as the
following example demonstrates.
Definition 5 (Passive system) The system B in (3.1)
is called passive if, for any given (i, v) ∈ B and t0 ∈ R,
there exists a K ∈ R (dependent on (i, v) and t0 ) such
that if (î, v̂) ∈ B satisfies (î(t), v̂(t)) = (i(t), v(t)) for all
t < t0 , then − ∫tt01 îT (t)v̂(t)dt < K for all t1 ≥ t0 .
In words, a system is passive if there is an upper bound
to the net energy that can be extracted from the system
from t0 onwards. The upper bound depends on the past
of the trajectory, but, given this past, the same upper
bound applies to all possible future trajectories.
Example 6 Consider the behavior B = {(u, y) ∈
loc
loc
Lloc
1 (R, R) × L1 (R, R) | ∃x ∈ L1 (R, R) with (i)
x(t) = 0 and y(t) = 0 for all t < 0; (ii) dx
dt (t) = u(t)
and y(t) = 0 for all 0 ≤ t < 1; (iii) dx
(t)
= u(t) and
dt
dx
y(t) = 2x(t) for all 1 ≤ t < 2; and (iv) dt (t) = 0 and
y(t) = 0 for all t ≥ 2}. Thus, if either t0 ≥ 2 or t1 ≤ 1,
then − ∫tt01 u(t)y(t)dt = 0; and if instead t1 > t0 , t0 < 2,
min(2,t )
and t1 > 1, then − ∫tt01 u(t)y(t)dt = −[x2 (t)]max(1,t10 ) =
A detailed discussion of the issues with existing definitions of passivity (and dissipativity) was provided in
(Willems, 2007, Section 8). However, for reasons detailed
at the end of this section, our definition differs from a
similar definition proposed by Willems (2007). First, we
compare Definition 5 to the conditions of the positivereal lemma. Note that it is not essential to follow the
discussion in the remainder of this section to understand
the main results in the paper.
min(2,t )
max(1,t )
0
u(τ )dτ )2 . It fol−[(∫0t u(τ )dτ )2 ]max(1,t10 ) ≤ (∫0
lows that, given any t0 ∈ R, there exists a K ∈ R depending on t0 and (u, y) such that − ∫tt01 u(t)y(t)dt < K
for all t1 ≥ t0 , and so B is passive in the sense of
Hughes and Smith (2017). On the other hand, for
any given (u, y) ∈ B, t0 < 1, and K > 0, there
exists (û, ŷ) ∈ B with (û(t), ŷ(t)) = (u(t), y(t)) for
all t < t0 such that − ∫tt01 û(t)ŷ(t)dt ≥ K (e.g., let
√
t0
û(t) = ( K −
√ ∫0 u(τ )dτ )/(1−t0 ) for all t0 ≤ t < 1,
and û(t) = − K for all t ≥ 1). Thus, if t0 < 1, then
an arbitrarily large amount of energy can be extracted
Condition 2 of Lemma 1 is sometimes stated as the definition of passivity for the system in (1.2) (e.g., Anderson and Vongpanitlerd, 1973, Section 2.3). However, the
system in Example 4 satisfies this condition but is not
passive in the sense of Definition 5. In other papers, condition 1 of Lemma 1 is stated as the definition for passivity (e.g., Willems, 1972b). It is shown in Hughes (2017b)
1
The behavior of a transformer with turns-ratio matrix
T ∈ Rn1 ×n2 is determined by the equations v1 = T T v2 , and
i2 = −T i1 , with v = col(v1 v2 ) and i = col(i1 i2 )).
2
For an inductor with inductance L, then Q−1 P (ξ) = Lξ.
3
Minor adjustments can be made to the proof given in this
paper to show that if B is passive in the sense of Hughes and
Smith (2017), then condition 2 of Theorem 13 holds.
4
from this system from t0 onwards, and this system is
not passive in the sense of Definition 5.
3. There exist compatible partitions i = (i1 , i2 ) and v =
(v1 , v2 ) such that B̃ := B (col(i1 v2 ),col(v1 i2 )) takes the
form of (3.2), and B̃ is passive.
Motivated by electric (RLCTG) networks, we have introduced a definition for passivity for the system in (3.1).
The classical theory of electric networks provides necessary and sufficient conditions on P and Q for the system
in (3.1) to be realized by an RLCTG network providing
P and Q are left coprime. Yet, as emphasised in Çamlibel
et al. (2003), such conditions are unknown in cases when
P and Q are not left coprime. More fundamentally, in
these cases, necessary and sufficient conditions on P and
Q for the system in (3.1) to be passive are also unknown.
Such conditions are provided in Theorem 9 of this paper.
4
Remark 10 It is also the case that the conditions in
Theorem 9 hold if and only if B is the driving-point behavior of an electric RLCTG network (Hughes, 2017a).
Remark 11 In the terminology of behavioral theory,
condition 3 of Theorem 9 implies that if B in (3.1) is
passive then there exists an input-output partition with
the property that iT v = uT y (in the context of electric networks, the input col(i1 v2 ) contains exactly one
variable, either current or voltage, for each port of the
network). It is well known that, if B is as in (3.1) and
normalrank([P −Q]) = n, then there exists an inputn
output partitioning of col(i v) into u ∈ Lloc
1 (R, R ) and
loc
n
(u,y)
y ∈ L1 (R, R ), for which B̃ := B
takes the form of
(3.2) (Polderman and Willems, 1998, Section 3.3). However, this does not suffice to show condition 3 in Theorem 9. For example, for the system
The passive behavior theorem
In this section, we present our new passive behavior theorem in two parts. The theorems use our new concept of
a positive-real pair, which we define as follows:
Definition 7 Let P, Q ∈ Rn×n [ξ]. We call (P, Q) a
positive-real pair if the following conditions hold:
"
0
0
1. P (λ)Q(λ̄)T + Q(λ)P (λ̄)T ≥ 0 for all λ ∈ C+ .
2. rank([P −Q](λ)) = n for all λ ∈ C+ .
3. If p ∈ Rn [ξ] and λ ∈ C satisfy pT (P Q? + QP ? ) = 0
and p(λ)T [P −Q](λ) = 0, then p(λ) = 0.
d
dt
+1
0
#" #
i1
i2
"
=
0
0
0
d
dt
+2
#" #
v1
,
v2
it can be shown that there is no input-output partition
with the property that i1 v1 + i2 v2 = uT y.
Theorem 9 allows us to apply the following results from
Willems (1986); Rapisarda and Willems (1997); Hughes
(2016a) on state-space realizations of behaviors.
Remark 8 A key result in behavioral theory is that
any behavior B as in (1.1) has a controllable part (Bc
in Lemma 17) and an autonomous part (Ba in Lemma
17). As will be shown in Section 5, the conditions in
Definition 7 can be understood in terms of Bc and Ba .
Roughly speaking, the passivity of Bc implies condition
1; the stability of Ba implies condition 2, as does the
stabilizability of B (see note B3); and condition 3 is a
coupling condition between the trajectories in Ba and the
so-called lossless trajectories in Bc . In particular, if the
transfer function from i to v is lossless PR (see Anderson
and Vongpanitlerd, 1973, Chapter 2), then P Q? +QP ? =
0, and condition 3 implies that P and Q are left coprime,
so B is controllable (see note B3).
Lemma 12 Let Bs be as in (1.2). Then there exist polynomial matrices M̃ , Ñ , P̃ and Q̃ such that
1. M̃ ∈ Rn×n [ξ] and Ñ ∈ Rn×d [ξ] are left coprime;
2. M̃ (ξ)C = Ñ (ξ)(ξI − A);
3. P̃ := Ñ B + M̃ D and Q̃ := M̃ .
Furthermore, if M̃ , Ñ , P̃ and Q̃ satisfy conditions 1–3,
(u,y)
then B̃ := Bs
takes the form of (3.2).
Now, let B̃ take the form of (3.2). Then there exists Bs
(u,y)
as in (1.2) such that B̃ = Bs
. Also, for any such Bs ,
there exist M̃ and Ñ such that conditions 1–3 hold.
We note that condition 1 of Definition 7 is a natural
generalization of a positive-real transfer function Q−1 P
to the case with Q singular. Yet, as discussed in Section 3,
this condition is not sufficient for the behavior B in (3.1)
to be passive. As the following theorem demonstrates,
conditions 2 and 3 are also required to obtain a necessary
and sufficient condition for passivity.
In the next theorem, we consider the state-space system Bs in (1.2), and we provide necessary and sufficient
(u,y)
conditions for Bs
to be passive. This generalizes the
positive-real lemma (Lemma 1) to state-space systems
that need not be controllable or observable.
Theorem 9 (Passive behavior theorem, Part 1)
Let B be as in (3.1). Then the following are equivalent:
Theorem 13 (Passive behavior theorem, Part 2)
Let Bs be as in (1.2); let P̃ , Q̃ be as in Lemma 12; and
let G(ξ) := D + C(ξI − A)−1 B. Then the following are
equivalent:
1. B is passive.
2. (P, Q) is a positive-real pair.
5
(u,y)
1. B̃ := Bs
is passive.
2. (P̃ , Q̃) is a positive-real pair.
3. There exist real matrices X, LX , WX such that X ≥
T
0, −AT X − XA = LTX LX , C − B T X = WX
LX ,
T
T
and D + D = WX WX .
4. There exist real matrices X, LX , WX as in condition 3 that have the additional property that WX +
LX (ξI−A)−1 B is a spectral factor of G + G? .
then BC ∞ := B ∩ C ∞ (R, Rn ) × C ∞ (R, Rn ) is called
cyclo-dissipative with respect to the supply rate iT v
(or cyclo-passive) if there exists a quadratic differend
tial form Qψ such that iT v ≥ dt
Qψ (col(i v)) for all
(i, v) ∈ BC ∞ (Pal and Belur, 2008, Definition 3.1). Also,
BC ∞ is called strictly cyclo-dissipative with respect to
the supply rate iT v (or strictly cyclo-passive) if there exists a quadratic differential form Qψ and an > 0 such
d
Qψ (col(i v)) + (iT i + vT v) for all (i, v) ∈
that iT v ≥ dt
BC ∞ (Pal and Belur, 2008, Definition 3.2). In these definitions, Qψ is called a storage function (Trentelman
and Willems, 1997, Definition 4.2), which is called nonnegative if Qψ (col(i v))(t) ≥ 0 for all (i, v) ∈ BC ∞ and
all t ∈ R. Çamlibel et al. (2003) considered cyclo-passive
single-input single-output systems, while Pal and Belur
(2008) considered a class of strictly cyclo-dissipative systems that includes the strictly cyclo-passive systems. 4
If, in addition, D + DT > 0, then the above conditions
are equivalent to:
5. There exists a real X ≥ 0 such that Π(X) :=
−AT X−XA−(C T −XB)(D+DT )−1 (C−B T X)=0.
Now, suppose conditions 1–4 hold. Then:
(i) If (C, A) is observable and X is as in condition 3,
then (a) X > 0; and (b) spec(A) ∈ C− .
(ii) If D + DT > 0 and X is as in condition 4, then (a)
Π(X) = 0; and (b) spec(A + B(D + DT )−1 (B T X −
C)) ∈ C− if and only if spec(A) ∈ C− .
It can be shown that there are cyclo-passive systems that
are not passive, and there are passive systems that are
not strictly cyclo-passive. Thus the problems considered
in Çamlibel et al. (2003); Pal and Belur (2008) are not
equivalent to the problem considered in this paper. It can
also be shown from Theorems 9 and 13 and Remark 15
that BC ∞ is passive in accordance with Definition 5 if and
only if BC ∞ is cyclo-passive with a non-negative storage
function. However, there are two notable reasons why we
have not defined a passive system as a cyclo-passive system with a non-negative storage function. First, as discussed in Willems (2007), it is preferable to define passivity without invoking an a-priori assumption of the existence of a quadratic storage function. This is one of the
main benefits of Definition 5. Second, we note that there
is no consensus on the appropriate definition of a cyclodissipative system. This concerns the issue of whether
to allow for unobservable storage functions, as arise in
electric networks (see Willems, 2004). As shown in that
paper, there are systems that are not cyclo-dissipative
(with respect to a given supply rate), but do possess an
unobservable storage function with respect to that supply rate (Willems, 2004, Section VI). This issue does not
arise with the definition of passivity given in this paper.
Remark 14 Note that, if the conditions in Theorem
(u,y)
,
13 hold for one state-space realization Bs of B̃ := Bs
then they hold for all state-space realizations of B̃. Note
also that P̃ and Q̃ are not uniquely defined in that theorem, but it is straightforward to show that condition 2
is invariant of the specific choice of matrices.
Remark 15 Let X, LX , WX be as in condition 3 of Theorem 13, let (u, y, x) ∈ Bs , and let t0 ≤ t1 ∈ R. Since x
is absolutely continuous, then integration by parts gives
Z
t1
t1
uT (t)y(t) + yT (t)u(t)dt − xT (t)Xx(t) t
0
t0
Z t1
=
(LX x + WX u)T (t)(LX x + WX u)(t)dt ≥ 0.
t0
With the notation S(x) := 21 xT Xx for all x ∈ Rd , it
is straightforward to verify that S is a storage function with respect to the supply rate uT y in the sense
of (Willems, 1972a, Definition 2). It follows from The(u,y)
orem 13 that, if B̃ := Bs
is passive (in accordance
with the trajectory-based Definition 5), then Bs has a
(non-negative) quadratic state storage function.
We also note that Çamlibel et al. (2003); Pal and Belur
(2008) invoke assumptions that are not present in this
paper. In Çamlibel et al. (2003), only single-input singleoutput systems are considered (i.e., n = 1 for B in (3.1)),
for which condition 3 in Definition 7 takes the much
simpler form: if P Q? + QP ? = 0, then [P −Q](λ) has
full row rank for all λ ∈ C. Also, Çamlibel et al. (2003)
assume that there are no uncontrollable imaginary axis
modes (i.e., rank([P −Q](jω)) is constant for all ω ∈
R). In contrast, we prove that this condition must hold if
Remark 16 It is instructive to compare Theorems 9
and 13 with papers by Çamlibel et al. (2003); Pal and
Belur (2008), which consider cyclo-dissipativity in the
behavioral framework. The reader who is unfamiliar with
these papers may prefer to skip straight to Section 5.
4
Note that these papers use the word dissipative for what
we call cyclo-dissipative systems. We reserve the word dissipative for systems that have a non-negative storage function,
as in Willems (1972a)).
In Çamlibel et al. (2003); Pal and Belur (2008), cyclodissipativity is defined using the formalism of quadratic
differential forms (see Appendix C). With B as in (3.1),
6
d
d
d
d
n
Lloc
1 (R, R ) | P ( dt )i = Q( dt )v and U ( dt )i = −V ( dt )v};
d
loc
n
loc
n
(ii) Bc := {(i, v) ∈ L1 (R, R )×L1 (R, R ) | P̃ ( dt )i =
d
Q̃( dt
)v}; and (iii) B̂ := {(i, v, i1 , v1 , i2 , v2 ) | (i1 , v1 ) ∈
Ba , (i2 , v2 ) ∈ Bc , i = i1 + i2 and v = v1 + v2 }. Then
B is passive (note, however, that there may exist ω ∈ R
such that det(P (jω)) = 0 and/or det(Q(jω)) = 0).
In Pal and Belur (2008), only strictly cyclo-dissipative
systems are considered. If B in (3.1) is strictly cyclopassive, then it can be shown that 1. Q(λ) and P (λ) are
nonsingular for all λ ∈ C+ ; and 2. P (jω)Q(−jω)T +
Q(jω)P (−jω)T is nonsingular for all ω ∈ R. The first
condition implies that condition 2 of Definition 7 holds
(but the converse implication does not hold). Similarly,
the second condition implies that condition 3 of Definition 7 holds (again, the converse implication does not
hold). Also, the proof of the main results in Pal and Belur
(2008) used algebraic Riccati equations and Hamiltonian
matrices. This approach cannot be used in this paper as
it is possible that D + DT is singular in Theorem 13.
5
Bc ∩C ∞ (R, Rn ) ×C ∞ (R, Rn ) ={(i, v) | ∃w∈C ∞ (R, Rn )
d
d
such that i = M ( dt
)w and v = N ( dt
)w},
Ba = {(i, v) | ∃z ∈ C
such that i =
∞
n
(R, R ) with
d
X( dt
)z
and v =
d
)z =
F ( dt
d
Y ( dt )z},
and B = B̂ (i,v) .
(5.3)
0,
(5.4)
(5.5)
PROOF. The decomposition in the first part of the
lemma statement is not unique, but one such decomposition is obtained by computing a lower echelon form
for [P −Q] (see note A4). This gives a unimodular
W ∈ R2n×2n [ξ] such that [F 0] = [P −Q]W . Then
W −1 =: Ŵ ∈ R2n×2n [ξ], and by suitably partitioning
Ŵ (resp., W ) we obtain the polynomial matrices in the
first (resp., second) block matrix in (5.2).
Passive behaviors and positive-real pairs
In Section 2, we showed that the system in Example 4
has a positive-real transfer function, yet is not passive.
(u,y)
=: B̃ =
For that system, it can be shown that Bs
d2
d
loc
loc
{(u, y) ∈ L1 (R, R) × L1 (R, R) | ( dt2 + 1)( dt
+ 1)u =
d2
d
( dt
2 + 1) dt y}. In particular, if B̃ is passive, then B̃c =
dy
d
loc
{(u, y) ∈ Lloc
1 (R, R) × L1 (R, R) | ( dt + 1)u = dt }
must be passive, and it follows that the transfer function
G(ξ) = 1 + 1/ξ must be PR. But this condition is not
sufficient for B̃ to be passive since there are trajectories
dy
dy
d2
d
d
in B̃ with ( dt
2 +1)(( dt +1)u− dt ) = 0 but ( dt +1)u 6≡ dt .
To show the second part of the lemma, we note initially
that (5.3)–(5.4) are easily shown from (5.1)–(5.2) and
(Polderman and Willems, 1998, Theorem 3.2.15). Now,
consider the compatibly partitioned matrices
#
"
" #
F P̃ −F Q̃
U
V
0
0
I
0
0
I
Z:=
As the preceding example indicates, the transfer function does not always determine the behavior of the system. In contrast, the behavior is always determined by
the polynomial matrices corresponding to the differential equations governing the system (i.e., by P and Q
in (3.1)). Thus, passivity will impose requirements on
these polynomial matrices. The purpose of this section
is to determine these requirements, resulting in Lemma
21. We will first prove some alternative requirements in
Lemma 18, which we then show to be equivalent to the
conditions in Lemma 21. These alternative requirements
relate to the following decomposition of the behavior B
in (3.1) into controllable and autonomous parts:
Lemma 17 Let B in (3.1) satisfy normalrank([P −Q]) =
n. Then there exist F, P̃ , Q̃, M, N, U, V, X, Y ∈Rn×n [ξ]
such that
P = F P̃ , Q = F Q̃, and
(5.1)
"
#"
# "
# "
#"
#
P̃ −Q̃ X M
In 0
X M P̃ −Q̃
=
=
. (5.2)
U V
Y N
0 In
Y N
U V
0 0
0 0
P̃ −Q̃ , R:=
I 0
0 I
0
0
0
I
0
0
0
0
0
I
, W :=
I
0
0
0
0
0
I
0
0
0
F −F P̃ F Q̃
0
0
0
I
0
0
0
I
0
0
0
I
,
and note that B̂ is the set of locally integrable solud
d
tions to R( dt
)col(i v) = Z( dt
)col(i1 v1 i2 v2 ). Next,
2n×2n
let Z2 ∈ R
[ξ] be formed from the last four block
rows of Z. It is straightforward to verify from (5.2) that
Z2 is unimodular. As W is unimodular, then by premultiplying R and Z by W we conclude that B̂ is the
d
d
set of locally integrable solutions to P ( dt
)i = Q( dt
)v
d
and col(0 0 i v) = Z2 ( dt )col(i1 i2 v1 v2 ) (see note
B1). In particular, (i, v) ∈ B, and it remains to
show that, for any given (i, v) ∈ B, there exist locally integrable (i1 , v1 , i2 , v2 ) such that col(0 0 i v) =
d
)col(i1 i2 v1 v2 ). Accordingly, for any given
Z2 ( dt
H ∈ Rm×n [ξ] with normalrank(H) = m, we let ∆(H)
denote the maximum degree of all determinants composed of m columns of H. Then, from (Polderman,
1997, Theorem 2.8), it suffices to show that there exists a determinant of degree ∆([R Z]) formed from the
columns in Z together with some of the columns in R.
It can be shown that ∆([R Z]) = ∆([P̃ −Q̃]) +
deg(det(F )) (this follows since any non-zero determinant formed from columns of [R Z] must contain: (i)
the 2n non-zero columns from the first two block rows of
[R Z], which form a nonsingular matrix whose determinant is det(F ); and (ii) at least n non-zero columns from
Now, let F, P̃ , Q̃, M, N, U, V, X, Y ∈ Rn×n [ξ] satisfy
n
(5.1)–(5.2), and let (i) Ba := {(i, v) ∈ Lloc
1 (R, R ) ×
7
the third block row). Next, let ∆([P̃ −Q̃]) be the degree
of the determinant formed from columns i1 , . . . , in of
[P̃ −Q̃]. It can be shown that the degree of the determinant formed from columns i1 , . . . , in and 2n + 1, . . . , 6n
of [R Z] equals that of the determinant formed from
columns i1 , . . . , in and 2n + 1, . . . , 6n of W [R Z], which
equals ∆([P̃ −Q̃]) + deg(det(F )) = ∆([R Z]). 2
that (P (λ) + Q(λ))z = 0. Then, with the notation v(t) = zeλt + z̄eλ̄t and i(t) = −v(t) for all
t ∈ R, we find that (i, v) ∈ B. Also, for any given
t1 ≥ t0 ∈ R, then − ∫tt01 iT (t)v(t)dt=2<(zT z ∫tt01 e2λt dt)+
2(z̄T z) ∫tt01 e2<(λ)t dt. By considering separately the cases
=(λ) = 0 and =(λ) 6= 0, it can be shown that for
any given K ∈ R there exists t1 ≥ t0 ∈ R such that
− ∫tt01 iT (t)v(t)dt ≥ K, whence B is not passive. Thus, if
B is passive, then rank([P −Q](λ)) = n for all λ ∈ C+ .
Equations (5.3)–(5.4) represent the infinitely-often differentiable part of the behavior B in terms of the five matrices M, N, X, Y and F ∈ Rn×n [ξ]. In the next lemma,
we provide three conditions on these matrices for B to
be passive. These correspond to the conditions:
Proof of condition 1. Consider a fixed but arbitrary
λ ∈ C+ and c ∈ Cn ; let z(t) = ceλt + c̄eλ̄t for all t ∈
d
d
R; let i := M ( dt
)z and v := N ( dt
)z; let Ψ(η, ξ) :=
T
T
T
M (η) N (ξ) + N
(η) M (ξ); and let α := c Ψ(λ, λ)c and
T
β := c̄ Ψ λ̄, λ c. Then (i, v) ∈ B by Lemma 17, and
1. Bc is passive.
2. Ba is stable. (i.e., (ia , va ) ∈ Ba ⇒ ia (t) → 0 and
va (t) → 0 as t → ∞).
3. If t0 ≤ t1 ∈ R, (ia , va ) ∈ Ba , and (il , vl ) ∈
Bc ∩C ∞ (R, Rn )×C ∞ (R, Rn ) with il (t) = vl (t) = 0
for all t < t0 and ∫tt01 iTl (t)vl (t)dt = 0, then
∫tt01 (iTa (t)vl (t) + vaT (t)il (t))dt = 0.
Z
t1
Z
iT (t)v(t)dt = < α
t1
Z
e2λt dt +β
e2<(λ)t dt. (5.6)
t0
t0
t0
t1
We will show that if there exists a λ ∈ C+ and c ∈ Cn
such that β = c̄T Ψ(λ̄, λ)c < 0, then for any given K ∈ R
there exists a t1 ≥ t0 with − ∫tt01 iT (t)v(t)dt ≥ K. This
will prove condition 1.
Condition 1 is to be expected since Bc ⊆ B. Condition
2 is equivalent to B being stabilizable. 5 Condition 3
is a coupling condition between the lossless trajectory
(il , vl ) and the autonomous trajectory (ia , va ). In fact,
this condition also holds when Ba is replaced by B ∩
EC− (R, Rn ) × EC− (R, Rn ) (an observation which is used
in the proof of Theorem 13), and provides the intuition
behind the third condition of the following lemma:
Let λ = σ + jω for some σ, ω ∈ R with σ ≥ 0,
and consider a fixed but arbitrary K ∈ R. We consider the cases (i) ω = 0; and (ii) ω 6= 0. In case
(i), let λ = λ̄ and c = c̄, so α = β. Then, from
(5.6), ∫tt01 iT (t)v(t)dt = 2β ∫tt01 e2λt dt, and ∫tt01 e2λt dt =
(1/2λ)(e2λt1 −2λt0 ) if <(λ) 6= 0, and t1 − t0 otherwise.
In case (ii), for any given integer n, we let T (n) ∈ R
satisfy 2ωT (n) = 2π(n + 1/4) − arg(α/(σ + jω)) (note,
if T (n) ≥ t0 , then n ≥ ωt0 /π + 1/4 when ω > 0, and
n ≤ ωt0 /π − 3/4 when ω < 0). Then arg(αe2λT (n) /λ) =
T (n)
π/2, so from (5.6) we find that ∫t0 iT (t)v(t)dt =
(β(e2<(λ)T (n) −e2<(λ)t0 )/2<(λ)) − <(αe2λt0 /2λ) if
<(λ) 6= 0, and β(T (n) − t0 ) − <(αe2λt0 /2λ) otherwise.
In both cases (i) and (ii), if β < 0, then by taking t1
sufficiently large (and letting t1 = T (n) in case (ii)) we
obtain − ∫tt01 iT (t)v(t)dt ≥ K.
Lemma 18 Let B be as in (3.1) and let B be passive.
Then normalrank([P −Q]) = n. Furthermore, with
M, N and F as in Lemma 17, then
1. M (λ̄)T N (λ) + N (λ̄)T M (λ) ≥ 0 for all λ ∈ C+ .
2. F (λ) is nonsingular for all λ ∈ C+ .
3. If (is , vs ) ∈ B ∩ EC− (R, Rn ) × EC− (R, Rn ) and
b ∈ Rn [ξ] satisfies b? (M ? N + N ? M ) = 0, then
d
d
d
)(M ? ( dt
)vs + N ? ( dt
)is ) = 0.
b? ( dt
Proof of condition 3.
Let b ∈ Rn [ξ] satisfy
b? (M ? N + N ? M ) = 0, let t0 ≤ t1 ∈ R, and consider a fixed but arbitrary (is , vs ) ∈ B ∩ EC− (R, Rn ) ×
EC− (R, Rn ) and z ∈ C ∞ (R, R). Then, with the notation
d
d
d
d
)b( dt
)z + is and v := N ( dt
)b( dt
)z + vs , it
i := M ( dt
follows that (i, v) ∈ B by Lemma 17. Also, with
PROOF. We first show that n = rank([P −Q](λ)) =
rank(F (λ)[P̃ −Q̃](λ)) for all λ ∈ C+ . This implies that
normalrank([P −Q]) = n and condition 2 holds. We
then show condition 1, and finally condition 3.
Proof that rank([P −Q](λ)) = n for all λ ∈ C+ .
Suppose instead that there exists λ ∈ C+ such
that rank([P −Q](λ)) < n. Then rank(P (λ) +
Q(λ)) < n, and so there exists 0 6= z ∈ Cn such
Z
t1
J1 :=
t
Z 0t1
d
d
d
d
((M ( dt
)b( dt
)z)T (N ( dt
)b( dt
)z))(t)dt, and
d
d
d
d
((M ( dt
)b( dt
)z)T vs +(N ( dt
)b( dt
)z)T is )(t)dt,
t
Z 0 t1
Z t1
then
(iT v)(t)dt = J1 +J2 +
(iTs vs )(t)dt.
(5.7)
J2 :=
5
In fact, it was established in Hughes and Smith (2017) that
any passive behavior is stabilizable. However, as discussed
in Section 2, the definition of passivity in Hughes and Smith
(2017) differs from the definition in this paper.
t0
8
t0
Since b? (M ? N + N ? M ) = 0 then, from note C3,
t1
d
d
)z)(t)+LΦN b (z, (M b)( dt
)z)(t) t ,
J1 = 21 LΦM b (z, (N b)( dt
0
Z t1
? d
? d
? d
(z(b ( dt )(M ( dt )vs + N ( dt )is )))(t)dt
and J2 =
t0
F (λ) is non-singular for all λ ∈ C+ , and so Ba ⊆
B∩EC− (R, Rn )×EC− (R, Rn ) by Lemma 17 and (Polderman and Willems, 1998, Section 3.2.2). Then 2 ⇒ 1 since,
by Lemma 17, if (is , vs ) ∈ B∩EC− (R, Rn )×EC− (R, Rn ),
then there exists (ia , va ) ∈ Ba and z ∈ C∞ (R, Rn ) such
d
d
that is = M ( dt
)z + ia and vs = N ( dt
)z + va .
t
+ [LΦM b (z, vs )(t)+LΦN b (z, is )(t)]t10 .
2 ⇐⇒ 3.
To see that 2 ⇒ 3, note initially from
Lemma 17 that condition 2 implies that if b ∈ Rn [ξ]
n
satisfies b? (M ? N +N ? M ), and z ∈ Lloc
1 (R, R ) satisfies
d
d
? d
?
?
F ( dt )z = 0, then b ( dt )(M Y + N X)( dt )z = 0. From
note B2, this implies that there exists p ∈ Rn [ξ] such
that b? (M ? Y + N ? X) = pT F . Similarly, from Lemma
17, it is straightforward to show that 3 ⇒ 2.
d
d
d
Now, let g := b ( dt
)(M ? ( dt
)vs + N ? ( dt
)is ); let ψ ∈
∞
C (R, R) and t0 ≤ t1 ∈ R satisfy ψ(t) = 0 for all
k
t ≤ t0 , and ddtkψ (t1 ) = 0 (k = 0, 1, 2, . . .); let z := gψ;
∞
:= 2
?
and let f
g . Then, z, f ∈ C (R, R); f (t) ≥ 0 for
all t ∈ R; J1 = 0; and J2 = ∫tt01 (f ψ)(t)dt. Moreover,
since (is , vs ) ∈ B ∩ EC− (R, Rn ) × EC− (R, Rn ), then it
is straightforward to show that iTs vs ∈ EC− (R, R), and
that there exists an M ∈ R such that ∫tt0 (iTs vs )(t)dt < M
for all t ≥ t0 . Thus, from (5.7), there exists an M ∈ R
such that − ∫tt01 (iT v)(t)dt > − ∫tt01 (f ψ)(t)dt−M . Finally,
we will show that, for any given K ∈ R, there exist ψ
and t1 with the properties outlined above that satisfy
− ∫tt01 (f ψ)(t)dt > K+M . This proves condition 3.
3 ⇐⇒ 4.
h
Note initially from (5.2) that
"
#"
#"
#
i −Q̃? V ? Y ? X ? X M
h
i
= In 0 .
P̃ −Q̃
P̃ ? U ? N ? M ? Y N
(5.8)
To prove that 3 ⇒ 4, note that if c ∈ Rn [ξ] satisfies
cT (P̃ Q̃? + Q̃P̃ ? ) = 0, then b? := cT (P̃ V ? − Q̃U ? ) satisfies b? (M ? N +N ? M ) = 0 by (5.8). Thus, from condition
3, there exists p ∈ Rn [ξ] such that b? (M ? Y + N ? X) =
pT F . But b? (M ? Y + N ? X) = cT (P̃ V ? − Q̃U ? )(M ? Y +
N ? X), and cT = cT (P̃ V ? − Q̃U ? )(M ? Y +N ? X) = pT F
by (5.8). The proof of 4 ⇒ 3 is similar.
2
Let φ(t) = e1/(t −1) for −1 < t < 1 with φ(t) = 0
otherwise. Also, for any given integer k, let gk (t) :=
f (t)φ(t − 1 − t0 − 2k) for all t ∈ R. Note that gk ∈
t +2(k+1)
C ∞ (R, R) and ∫t00+2k
gk (t)dt > 0 (k = 0, 1, . . .).
Now, let N be a positive integer with N > K+M , and let
PN −1
t +2(k+1)
ψ(t) = − k=0 φ(t−1−t0 −2k)/(∫t00+2k
gk (t)dt) for
all t ∈ R. It can be verified that ψ ∈ C ∞ (R, R); ψ(t) = 0
l
for all t ≤ t0 ; ddtψl (t0 + 2k) = 0 for k, l = 0, 1, 2, . . .; and
− ∫tt00 +2N (f ψ)(t)dt = N > K + M . 2
4 ⇐⇒ 5.
To see that 4 ⇒ 5, we let r :=
?
normalrank(P̃ Q̃ + Q̃P̃ ? ), and we let the rows of V1 ∈
R(n−r)×n [ξ] be a basis for the left syzygy of P̃ Q̃? + Q̃P̃ ?
(see note A3). Then condition 4 implies that there exists V̂1 ∈ R(n−r)×n [ξ] such that V1 = V̂1 F . Since V1 (λ)
has full row rank for all λ ∈ C, then so too must V̂1 (λ).
Since, in addition, P Q? + QP ? = F (P̃ Q̃? + Q̃P̃ ? )F ? ,
then V̂1 (P Q? + QP ? ) = V1 (P̃ Q̃? + Q̃P̃ ? )F ? = 0 and
normalrank(P Q? + QP ? ) = normalrank(P̃ Q̃? + Q̃P̃ ? ),
and we conclude that the rows of V̂1 are a basis
for the left syzygy of P Q? +QP ? . It follows that if
pT (P Q? +QP ? ) = 0, then there exists g∈R(n−r) [ξ] such
that pT = gT V̂1 . If, in addition, p(λ)T [P −Q](λ) = 0,
then p(λ)T F (λ) = 0 since P̃ and Q̃ are left coprime,
whence g(λ)T V̂1 (λ)F (λ) = g(λ)T V1 (λ) = 0. But V1 (λ)
has full row rank for all λ ∈ C, and we conclude that
g(λ) = 0 and so p(λ) = 0. Finally, to show that 5 ⇒ 4,
we let the rows of V̂1 ∈ R(n−r)×n [ξ] be a basis for the
left syzygy of P Q? + QP ? . Then, from condition 5, we
conclude that c(λ)T V̂1 (λ)F (λ) = 0 ⇒ c(λ)T = 0, and it
follows that V̂1 (λ)F (λ) has full row rank for all λ ∈ C. In
a similar manner to before, it can then be shown that the
rows of V̂1 F are a basis for the left syzygy of P̃ Q̃? + Q̃P̃ ? .
Hence, if c ∈ Rn [ξ] satisfies cT (P̃ Q̃? + Q̃P̃ ? ) = 0, then
there exists g ∈ R(n−r) [ξ] such that cT = gT V̂1 F , and
by letting pT := gT V̂1 we obtain condition 4. 2
In the next lemma, we present several equivalent conditions to the third condition in Lemma 18. This leads to
two algebraic tests for this condition (see Remark 20),
and the main result in this section (see Lemma 21).
Lemma 19 Let B be as in (3.1); let rank([P −Q](λ)) =
n for all λ ∈ C+ ; and let F, P̃ , Q̃, M, N, U, V, X, Y , and
Ba be as in Lemma 17. Then the following are equivalent:
1. Condition 3 of Lemma 18 holds.
2. If (ia , va ) ∈ Ba and b ∈ Rn [ξ] satisfies b? (M ? N +
d
d
d
)(M ? ( dt
)va + N ? ( dt
)ia ) = 0.
N ? M ) = 0, then b? ( dt
n
?
?
?
3. If b ∈ R [ξ] satisfies b (M N +N M ) = 0, then there
exists p ∈ Rn [ξ] such that b? (M ? Y + N ? X) = pT F .
4. If c ∈ Rn [ξ] satisfies cT (P̃ Q̃? + Q̃P̃ ? ) = 0, then there
exists p ∈ Rn [ξ] such that cT = pT F .
5. If p ∈ Rn [ξ] and λ ∈ C satisfy pT (P Q? + QP ? ) = 0
and p(λ)T [P −Q](λ) = 0, then p(λ) = 0.
PROOF. 1 ⇐⇒ 2.
That 1 ⇒ 2 follows since
rank([P −Q](λ)) = n for all λ ∈ C+ implies that
9
2 ⇒ 4. To prove this implication, we will show conditions (i) and (ii) below. The notation in those conditions
is as follows. We let T = col(T1 T2 ) be such that C̃ =
[C̃1 0] = CT −1 and à = T AT −1 have the observer staircase form indicated in note D2, and we let T B =: B̃ and
T1 B =: B̃1 . Then, with Ã11 ∈ Rd1 ×d1 as in note D2, we
let T̃ ∈ Rd1 ×d1 be such that T̃ Ã11 T̃ −1 = diag(As Au ),
where spec(As ) ∈ C− and spec(Au ) ∈ C+ (Gantmacher,
1980, Chapter VII). 6 We partition T̃ B̃1 and C̃1 T̃ −1
compatibly with T̃ Ã11 T̃ −1 = diag(As Au ) as T̃ B̃1 =
col(Bs Bu ) and C̃1 T̃ −1 = [Cs Cu ]. We then let Gs (ξ) =
D + Cs (ξI−As )−1 Bs and Gu (ξ) = Cu (ξI−Au )−1 Bu ,
and direct calculation shows that G(ξ) = D + C̃1 (ξI −
Ã11 )−1 B̃1 = Gs (ξ) + Gu (ξ). We will show the following.
Remark 20 Let B, F, P̃ , Q̃, M, N, U, V, X, Y , and Ba be
as in Lemma 19 (with rank([P −Q](λ)) = n for all
λ ∈ C+ ). Lemma 19 leads to two tests that can be
implemented by a standard symbolic algebra program
(using exact arithmetic if the polynomial matrix coefficients are rational). As in the proof of Lemma 19,
let r := normalrank(P Q? +QP ? ), and note that it is
easily shown from the proof of that lemma that r =
normalrank(M ? N +N ? M ). The two tests are as follows.
1. Using the matrices M, N, X, Y and F :
(a) Compute a V ∈ R(n−r)×n whose rows are a basis
for the left syzygy of M ? N +N ? M .
(b) Condition 3 of Lemma 19 holds if and only if
V (M ? N +N ? M ) is divisible on the right by F .
2. Using the matrices P and Q:
(a) Compute a V ∈ R(n−r)×n whose rows are a basis
for the left syzygy of P Q? +QP ? .
(b) Condition 5 of Lemma 19 holds if and only if
V (λ)[P −Q](λ) has full row rank for all λ ∈ C.
(i) There exists a real Xu > 0 such that −ATu Xu −
Xu Au = 0 and CuT − Xu Bu = 0.
(ii) There exist real matrices Xs , L, W such that Xs >
0, −ATs Xs − Xs As = LT L, CsT − Xs Bs = LT W ,
and D + DT = W T W , where W + L(ξI−As )−1 Bs
is a spectral factor of G + G? .
Lemma 21 Let B be as in (3.1). If B is passive, then
(P, Q) is a positive-real pair.
We note that T̂ := col(T̃ T1 T2 ) = diag(T̃ I)T is nonsingular. Then, with  := T̂ AT̂ −1 , B̂ := T̂ B, Ĉ := C T̂ −1 ,
X̂ := diag(Xs Xu 0), LX̂ := [L 0 0], and WX̂ := W ,
it can be verified that X̂ ≥ 0; −ÂT X̂ − X̂ Â =
diag((−ATs Xs −Xs As ) (−ATu Xu −Xu Au ) 0) = LTX̂ LX̂ ,
PROOF. That (P, Q) satisfy condition 2 in Definition
7 follows from condition 2 of Lemma 18, by noting from
(5.1)–(5.2) that [P −Q]col(X Y ) = F . Then condition 3 in Definition 7 follows from Lemmas 18–19. It remains to show that condition 1 of Definition 7 holds. To
see this, first note from condition 1 of Lemma 18 that
((M +N )(λ̄)T (M +N )(λ)−(N −M )(λ̄)T (N −M )(λ)) ≥
0 for all λ ∈ C+ . Next, let λ ∈ C+ and z ∈ Cn satisfy (M + N )(λ)z = 0. Then −z̄T ((N − M )(λ̄)T (N −
M )(λ))z ≥ 0, which implies that (N − M )(λ)z = 0, and
so M (λ)z = 0 and N (λ)z = 0. Then from (5.2) we obtain z = U (λ)M (λ)z + V (λ)N (λ)z = 0. We conclude
that (M + N )(λ) is nonsingular for all λ ∈ C+ . Accordingly, with the notation H := (N − M )(M + N )−1 , then
I ≥ H(λ̄)T H(λ) for all λ ∈ C+ . This implies that I ≥
H(λ)H(λ̄)T for all λ ∈ C+ (to see this, let X = H(λ),
and note that I −X X̄ T = (I −X X̄ T )(I −X X̄ T )+X(I −
X̄ T X)X̄ T ). Then, noting that P M = QN implies that
(P +Q)H = (P −Q)(M +N )(M +N )−1 = P −Q, we find
that (P +Q)(λ)(P +Q)(λ̄)T ≥ (P +Q)(λ)H(λ)H(λ̄)(P +
Q)(λ̄)T = (P − Q)(λ)(P − Q)(λ̄)T for all λ ∈ C+ . We
conclude that condition 1 of Definition 7 holds. 2
6
Ĉ T − X̂ B̂ = col((CsT −Xs Bs ) (CuT −Xu Bu ) 0) =
T
LTX̂ WX̂ and D + DT = WX̂
WX̂ ; and ZX̂ (ξ) =
WX̂ +LX̂ (ξI−Â)−1 B̂ is a spectral factor of G+G? . Finally, with X:=T̂ T X̂ T̂ , LX :=LX̂ T̂ , and WX :=WX̂ , it
can be verified that X, LX , and WX satisfy condition 4.
We first prove (i). Direct calculation verifies that G =
Q̃−1 P̃ . Since (P̃ , Q̃) is a positive-real pair and Q̃ is nonsingular, then G is PR. To see this, note that if G is analytic in C+ then G(λ) + G(λ̄)T = Q̃−1 (λ)(P̃ (λ)Q̃(λ̄)T +
Q̃(λ)P̃ (λ̄)T )(Q̃−1 )(λ̄)T ≥ 0 for all λ ∈ C+ , so G is PR.
But suppose instead that G has a pole at some λ ∈ C+ .
By considering the Laurent series for G about λ, it can
be shown that, for any > 0, there exists z ∈ Cn and an
η ∈ C with |η| ≤ such that z̄T (G(λ+η)+G(λ̄+ η̄)T )z <
0: a contradiction.
Since G is analytic in C+ and G = Gu +Gs with Gs (ξ) =
D + Cs (ξI−As )−1 Bs (whose poles are all in C− ) and
Gu (ξ) = Cu (ξI−Au )−1 Bu (whose poles are all in C+ ),
then the poles of Gu must all be on the imaginary axis.
Since, in addition, G is PR, then Gu and Gs are both PR
and Gu + G?u = 0 (Anderson and Vongpanitlerd, 1973,
Passive behavior theorem
In this final section, we prove Theorems 9 and 13.
PROOF OF THEOREM 13 (see p. 5). We first prove
that 1 ⇒ 2 ⇒ 4 ⇒ 3 ⇒ 1.
6
This can alternatively be shown using the real Jordan
form. Here, letting ds denote the number of columns (and
rows) of As , then the first ds rows (resp., last d1 − ds rows)
of T̃ span the stable (resp., unstable) left eigenspace of Ã11 .
1 ⇒ 2. By Lemma 12, B̃ takes the form of (3.2). Hence,
(P̃ , Q̃) is a positive-real pair by Lemma 21.
10
(u,y)
Section 5.1). Next, note that B̃ := Bs
is stabilizable
(by condition 2 of Definition 7), and has the observable
realization in note D3, whence [λI−Ã11 B̃1 ] has full row
rank for all λ ∈ C+ (this follows from note D4). It is
then easily shown that [λI−Au Bu ] has full row rank for
all λ ∈ C, so (Au , Bu ) is controllable. Similarly, it can
be shown that (Cu , Au ) is observable since (C̃1 , Ã11 ) is.
Thus, Gu (ξ) = Cu (ξI−Au )−1 Bu is PR with Gu +G?u = 0
and with (Au , Bu ) controllable and (Cu , Au ) observable,
and so (i) holds by (Willems, 1972b, Theorem 5).
and Willems (1997); Hughes (2016a), it follows that
(y)
d
n
B̂s = {y ∈ Lloc
1 (R, R ) | U ( dt )y = 0} (c.f., Lemma
d
n
12). Thus, if vs ∈ Lloc
1 (R, R ) satisfies U ( dt )vs = 0,
n
n
then (0, vs ) ∈ B̃ ∩ EC− (R, R ) × EC− (R, R ). Next, note
from Lemmas 19 and 21 that condition 3 of Lemma
18 holds. Also, since H2 K ? = 0, then H2 K ? K =
n
H2 (M ? N + N ? M ) = 0. Thus, if vs ∈ Lloc
1 (R, R ) satisd
d
d
fies U ( dt )vs = 0, then H2 ( dt )M ? ( dt )vs = 0. It follows
from note B2 that there exists S ∈ R(n−r)×n [ξ] such
that H2 M ? = SU , whence H2 M ? Cs = SU Cs = SV As .
With J2 := SV , we obtain condition (a)(ii), which
completes the proof of condition (a).
Next, let As (ξ) := ξI −As ; let M and N be as in Lemma
17 (so, in particular, M is invertible, and N M −1 =
Q−1 P = G, which is PR); let r = normalrank(M ? N +
N ? M ); and let K ∈ Rr×n [ξ] be a spectral factor for
M ? N +N ? M (i.e., K(λ) has full row rank for all λ ∈ C+ ,
and K ? K = M ? N + N ? M ). To prove condition (ii), we
will show the following four conditions.
Condition (b) follows as spec(As )∈C− implies that Xs =
T
∫0∞ eAs t LT LeAs t dt ≥ 0 satisfies −ATs Xs −Xs As =LT L.
To show (c), we first let λ ∈ C+ and z ∈ Cn satisfy
M (λ)z = 0. We recall that N M −1 = G is PR, so G is
analytic in C+ , whence N (λ)z = G(λ)M (λ)z = 0. Then,
from (5.2), it follows that (U (λ)M (λ) + V (λ)N (λ))z =
z = 0. We conclude that M (λ) is nonsingular for all
λ ∈ C+ . Since, in addition, K is a spectral factor of
M N ? + N M ? , then it is straightforward to show that Z
is a spectral factor of G + G? = M −1 N + N ? (M −1 )? .
(a) There exist J ∈ Rn×ds [ξ] and L ∈ Rr×ds such that
K ? L + JAs = M ? Cs .
(b) With L as in (a), there exists Xs ∈ Rds ×ds such that
−ATs Xs − Xs As = LT L.
(c) Z := KM −1 is a spectral factor of G + G? and, with
W := limξ→∞ Z(ξ), then Z = W + LA−1
s Bs . In
particular, D + DT = W T W .
(d) With Z, W, L as in (a)–(c), then CsT −Xs Bs =LT W .
We next let W := limξ→∞ Z(ξ), and we will show that:
(c)(i) W + LA−1
s Bs − Z has no poles in C− ; and (c)(ii)
W + LA−1
B
− Z has no poles in C+ . Since, in ads
s
dition, W = limξ→∞ (Z(ξ)), then W + LA−1
s Bs = Z.
It then follows that W T W = limξ→∞ (Z ? (ξ)Z(ξ)) =
limξ→∞ (G(ξ) + G? (ξ)) = D + DT .
To show (a), recall that K ? ∈ Rn×r [ξ] satisfies
normalrank(K ? ) = r, and let col(H1 H2 ) = H ∈
Rn×n [ξ] be a unimodular matrix such that the rows
of H2 ∈ R(n−r)×n [ξ] are a basis for the left syzygy
of K ? (e.g., consider the upper echelon form for K ? ,
see note A4). Then H2 K ? = 0, H1 K ? ∈ Rr×r [ξ],
and normalrank(H1 K ? ) = r. We will show that:
(a)(i) there exists L ∈ Rr×ds and J1 ∈ Rr×ds [ξ]
such that H1 K ? L + J1 As = H1 M ? Cs ; and (a)(ii)
there exists J2 ∈ R(n−r)×ds [ξ] such that J2 As =
H2 M ? Cs . Since H2 K ? = 0 and H is unimodular, then
J:=H −1 col(J1 J2 ) and L as above satisfy condition (a).
To show (c)(i), we note that K ? (W + LA−1
s Bs − Z) =
? ?
?
?
−1
K ? W + K ? LA−1
s Bs − M Z Z = K W + K LAs Bs −
? −1 T
T
).
Clearly,
)
C
(A
B
+
B
M ? (D + DT + Cs A−1
s
s
s
s
s
(A?s )−1 has no poles in C− , and from (a) it follows that
?
−1
K ? LA−1
s − M Cs As = −J, which has no poles in C− .
?
−1
Thus, K (W + LAs Bs − Z) has no poles in C− . Since
K ? (λ) has full column rank for all λ ∈ C− , then we
conclude that W + LA−1
s Bs − Z has no poles in C− .
To see (a)(i), note from the definitions of H, K and
As that (H1 K ? )(λ) is nonsingular for all λ ∈ C− and
As (λ) is nonsingular for all λ ∈ C+ . Furthermore, from
(Gantmacher, 1980, pp. 77–79), there exist E ∈ Rr×ds [ξ]
and F ∈ Rr×ds such that H1 M ? Cs = EAs + F . Then,
from (Feinstein and Bar-Ness, 1980, Theorem II), there
exist L ∈ Rr×ds and R ∈ Rr×ds [ξ] such that H1 K ? L +
RAs = F . With J1 := E +R, we obtain condition (a)(i).
To see (c)(ii), we note that, since G + G? = D + DT +
T
? −1 T
? −1
Cs A−1
Cs , and A−1
) has
s Bs + Bs (As )
s (resp., (As )
?
no poles in C+ (resp., C− ), then (G+G )(jω) is analytic
for all ω ∈ R, whence Z is analytic in C+ . It follows that
W + LA−1
s Bs − Z has no poles in C+ . This completes
the proof of condition (c).
Finally, to show condition (d), note initially from (b)
that Xs A−1
+ (A?s )−1 Xs = (A?s )−1 LT LA−1
s
s . Next,
?
note that M ? (W T L + BsT Xs − Cs )A−1
=
M
(W T +
s
T
? −1 T
−1
?
−1
? T
? −1
Bs (As ) L )LAs − M Cs As − M Bs (As ) Xs .
Also, M ? (W T + BsT (A?s )−1 LT ) = M ? Z ? = K ? by
(c). Thus, from (a), we find that M ? (W T L + BsT Xs −
? T
? −1
Cs )A−1
= (K ? L − M ? Cs )A−1
Xs =
s
s − M Bs (As )
? T
? −1
−J − M Bs (As ) Xs . It follows that M ? (W T L +
n
To see (a)(ii), let B̂s := {(y, xs ) ∈ Lloc
1 (R, R ) ×
dxs
loc
ds
L1 (R, R ) | dt = As xs and y = Cs xs }. Note
that, if (vs , xs ) ∈ B̂s , then vs ∈ EC− (R, Rn ) and
(0, vs , T̂ −1 col(xs 0 0)) ∈ Bs , and so (0, vs ) ∈ B̃ ∩
EC− (R, Rn ) × EC− (R, Rn ). Then, let U ∈ Rn×n [ξ]
and V ∈ Rn×ds [ξ] be left coprime matrices satisfying U Cs = V As , so, from Willems (1986); Rapisarda
11
−1
Then yT (λI−(A − BWX
LX ))(λI−A)−1 B = yT B +
−1
−1
T
−1
y BWX LX (λI−A) B = yT B + yT BWX
(ZX (λ) −
−1
−1
T
WX ) = y BWX ZX (λ) = 0. But WX ZX (λ) has
full row rank, so yT B = 0. Thus, yT (λI−A) =
−1
−1
yT (λI−(A − BWX
LX )) − yT BWX
LX = 0, which
implies that y = 0. 2
BsT Xs − Cs )A−1
has no poles in C− . But M ? (λ) is
s
nonsingular for all λ ∈ C− , and we conclude that
(W T L + BsT Xs − Cs )A−1
s has no poles in C− . It is then
straightforward to show that W T L + BsT Xs − Cs = 0.
4 ⇒ 3.
Immediate.
3 ⇒ 1.
Consider a fixed but arbitrary (u, y, x) ∈ Bs
(u,y)
and t0 ∈ R, and let (û, ŷ) ∈ B̃ = Bs
satisfy û(t) =
u(t) and ŷ(t) = y(t) for all t < t0 . Then, from note
D3, there exists (û, ŷ, x̂) ∈ Bs with x̂(t0 ) = x(t0 ). From
Remark 15, since x̂T (t1 )X x̂(t1 ) ≥ 0 and x̂(t0 ) = x(t0 ),
then − ∫tt01 ûT (t)ŷ(t)dt ≤ 12 xT (t0 )Xx(t0 ). This inequality holds for all (û, ŷ) ∈ B̃ that satisfy (û(t), ŷ(t)) =
(u(t), y(t)) for all t < t0 , so B̃ is passive.
Remark 22 We note that the matrix L in the above
theorem can be obtained by considering the Jordan
chains of As . Specifically, let As have eigenvalues
λ1 , . . . , λn with Jordan chains (v1,1 , . . . , v1,N (λ1 ) ), . . . ,
(vn,1 , . . . , vn,N (λn ) ). Also, for any given H ∈ Rm×n (ξ)
and λ ∈ C such that λ is not a pole of H, let Hλ,j denote
the j+1th term in the Taylor expansion for H about λ,
dj
i.e., Hλ,j = j!1 ( dξ
j H)(λ) for j = 0, 1, . . .. Then L can be
obtained by solving the equations
We next assume that D + DT > 0, and we prove that 4
⇒ 5 ⇒ 3. First, let X, LX , WX and ZX be as in condition
4. Since n ≥ normalrank(G + G? ) ≥ rank(D + DT ) = n,
then normalrank(G + G? ) = n, so ZX ∈ Rn×n (ξ), and
T
WX = limξ→∞ (Z(ξ)) ∈ Rn×n . As WX
WX = D + D T ,
which is nonsingular, then WX is nonsingular. We then
find that −AT X − XA − (C T − XB)(D + DT )−1 (C −
−1
T
T −1
LX = 0.
) WX
(WX
B T X) = LTX LX − LTX WX WX
Next, suppose X ≥ 0 is real and satisfies Π(X) = 0;
let WX be a real nonsingular matrix with D + DT =
T
T −1
WX ; and let LX := (WX
) (C − B T X). Then
WX
X, LX and WX satisfy condition 3.
k−1
X
(Kλ?i ,j Lvi,k−j − Mλ?i ,j Cs vi,k−j ) = 0,
(6.1)
j=0
for i = 1, . . . , n and k = 1, . . . , N (λi ). It can then be
shown that, if spec(A) ∈ C− and condition 3 of Theorem
13 holds, then (Pandolfi, 2001, equation (4)) must hold.
To show (6.1), we let As (ξ) := ξI − As , and we consider
the Jordan chain for As corresponding to an eigenvalue
λ: As (λ)v1 = 0, As (λ)vj + vj−1 = 0 (j = 2, . . . , N (λ)).
If J ∈ Rn×ds [ξ] and L ∈ Rr×ds satisfy K ? L + JAs =
dj
?
?
M ? Cs , then Kλ,j
L − Mλ,j
Cs = − j!1 dξ
j (JAs )(λ) =
−Jλ,j As (λ)−Jλ,j−1 (where Jλ,−1 := 0). Thus, for
Pk−1 ?
?
k = 1, . . . , N (λ),
j=0 (Kλ,j Lvk−j − Mλ,j Cs vk−j )
Pk−1
Pk−1
= − j=0 (Jλ,j As (λ)vk−j ) − i=1 (Jλ,i−1 vk−i ) =
Pk−2
− j=0 (Jλ,j (As (λ)vk−j +vk−j−1 ))−Jλ,k−1 As (λ)v1 =0.
We now prove condition (i). Accordingly, suppose condition 3 holds, and let X, LX and WX be as in that condition. To show condition (i)(a), suppose that (C, A) is
observable and there exists z ∈ Rd with Xz = 0. Since
X is symmetric, then zT X = 0. Thus, zT (−AT X −
XA)z = zT LTX LX z = 0, whence LX z = 0. It follows
T
that (C − B T X)z = WX
LX z = 0, so Cz = 0. Also,
T
T
(−A X −XA)z = LX LX z = 0, so XAz = 0. By replacing z with Az in the preceding argument, we find that
LX Az = 0, CAz = 0, and XA2 z = 0. Proceeding inductively gives CAk z = 0 (k = 0, 1, 2, . . .). Since (C, A)
is observable, then z = 0, and we conclude that X > 0.
To show condition (i)(b), let λ ∈ C+ and z ∈ Cd satisfy
(λI−A)z = 0. Then z̄T LTX LX z = z̄T (−AT X − XA)z =
−(λ̄ + λ)z̄T Xz ≤ 0, whence LX z = 0 and z̄T Xz = 0.
Since X > 0 by condition (i)(a), then z = 0, and we
conclude that spec(A) ∈ C− .
PROOF OF THEOREM 9 (see p. 5). That 1 ⇒ 2
was shown in Lemma 21. Here, prove that 2 ⇒ 3 ⇒ 1.
2 ⇒ 3. Let P̂ := P −Q and Q̂ := P +Q. Since (P, Q) is
a positive-real pair, then Q̂(λ)Q̂(λ̄)T − P̂ (λ)P̂ (λ̄)T ≥ 0
and rank([P̂ −Q̂](λ)) = n for all λ ∈ C+ . We will show
that: (i) Q̂(λ) is nonsingular for all λ ∈ C+ ; and (ii)
Q̂−1 P̂ is proper. To see (i), suppose z ∈ Cn and λ ∈ C+
satisfy zT Q̂(λ) = 0. Then −zT P̂ (λ)P̂ (λ̄)T z̄ ≥ 0, which
implies that zT P̂ (λ) = 0. Since rank([P̂ −Q̂](λ)) = n,
then this implies that z = 0. To see (ii), note that,
since Q̂(λ) is nonsingular for all λ ∈ C+ , then I −
(Q̂−1 P̂ )(λ)(Q̂−1 P̂ )(λ̄)T ≥ 0 for all λ ∈ C+ , and it is
then easily shown that Q̂−1 P̂ is proper.
It remains to prove condition (ii). Condition (ii)(a) was
shown in the proof of 4 ⇒ 5. To see condition (ii)(b), note
−1
that A+B(D +DT )−1 (B T X −C) = A−BWX
LX , and
consider a fixed but arbitrary λ ∈ C+ . From the proof
of condition (i)(b), if z ∈ Cd satisfies (λI−A)z = 0,
−1
then LX z = 0, whence (λI−(A − BWX
LX ))z =
(λI−A)z = 0. It remains to show that if λI−A is
−1
nonsingular, then λI−(A − BWX
LX ) is nonsingular. Accordingly, suppose that λI−A is nonsingular
−1
and y ∈ Cd satisfies yT (λI−(A − BWX
LX )) = 0.
Let R ∈ Rm×n [ξ] with normalrank(R) = m, and
recall the notation ∆(R) from the proof of Lemma
17. If R is partitioned as R = [R1 R2 ] where
R2 ∈ Rm×m [ξ] is nonsingular, then R2−1 R1 is proper
12
7
if and only if deg(det(R2 )) = ∆(R) (Polderman and
Willems, 1998, Theorem 3.3.22). Thus, deg(det(Q̂)) =
∆([P̂ −Q̂]). But det(Q̂) = det(P + Q), which is the
sum of all the determininants composed of columns
of P together with the complementary columns
of Q (i.e., det([p1 p2 · · · ]) + det([q1 p2 · · · ]) +
det([p1 q2 · · · ]) + . . ., where pk (resp., qk ) denotes
the kth column of P (resp., Q)). From among the
determinants in this sum, we pick one of greatest degree, we let T = col(T1 T2 ) be a permutation matrix
such that T1T (resp., T2T ) selects the columns from Q
(resp., P ) appearing in this determinant, and we define
Q̃ := [QT1T −P T2T ] and P̃ := [P T1T −QT2T ]. Then
deg(det(Q̃)) ≥ deg(det(P + Q)) = ∆([P̂ −Q̂]). Furthermore, T1T T1 + T2T T2 = I as T is a permutation
matrix, and with the notation
"
S1 :=
T1T 0
0 T2T
0 T2T T1T 0
#
"
, and S2 :=
1
2
I I
−I I
Conclusions
The positive-real lemma links the concepts of passivity,
positive-real transfer functions, spectral factorisation,
linear matrix inequalities, and algebraic Riccati equations. However, the lemma only considers systems described by a controllable state-space realization, which
leaves important questions unanswered. For example, it
does not specify which uncontrollable systems are passive. In this paper, we sought to answer this question
and others by proving two new theorems: the passive
behavior theorem, parts 1 and 2.
Acknowledgements
This research was conducted in part during a Fellowship supported by the Cambridge Philosophical Society,
http://www.cambridgephilosophicalsociety.org.
#
,
A
Polynomial and rational matrices
Several of the results in this paper depend on the properties of polynomial matrices that we describe here.
we find that S1 S1T = 2S2 S2T = I, and [P̃ −Q̃] =
[P −Q]S1 = [P̂ −Q̂]S2 S1 . Then, from the BinetCauchy formula, we obtain ∆([P̃ −Q̃]) = ∆([P̂ −Q̂])
(Hughes, 2016a, proof of Theorem 7.2). Since, in addition, ∆([P̃ −Q̃]) ≥ deg(det(Q̃)) ≥ ∆([P̂ −Q̂]), then
deg(det(Q̃)) = ∆([P̃ −Q̃]), so Q̃−1 P̃ is proper.
A1 U ∈ Rl×l [ξ] is called unimodular if there exists
V ∈ Rl×l [ξ] such that U V = I (whence V U = I). U is
unimodular if and only if det(U ) is a non-zero constant.
A2 Let R1 ∈ Rl×n1 [ξ] and R2 ∈ Rl×n2 [ξ]. We say that
R1 and R2 are left coprime if [R1 R2 ](λ) has full row
rank for all λ ∈ C.
A3 Let R ∈ Rl×n [ξ]. The left syzygy of R is the set of
c ∈ Rl [ξ] that satisfy cT R = 0. If normalrank(R) = m,
then there exists V ∈ R(l−m)×l [ξ] such that (i) V (λ)
has full row rank for all λ ∈ C; and (ii) V R = 0. If
V ∈ R(l−m)×l [ξ] satisfies (i) and (ii), then c ∈ Rl [ξ]
is in the left syzygy of R if and only if there exists a
p ∈ Rl−m [ξ] such that pT V = cT ; and we say that the
rows of V are a basis for the left syzygy of R.
A4 Given any R ∈ Rl×n [ξ] with normalrank(R) =
m, there exists a unimodular U ∈ Rl×l [ξ] (resp., V ∈
Rn×n [ξ]) such that U R = col(R̃ 0(l−m)×n ) (resp., RV =
[R̂ 0l×(n−m) ]), where R̃ ∈ Rm×n [ξ] is in either (i) upper echelon form, or (ii) row reduced form (resp., R̂ ∈
Rl×m [ξ] is in either (ib) lower echelon form, or (iib) column reduced form) (see, e.g., Gantmacher (1980) Chapter VI and Wolovich (1974)). The last l − m rows of
U are a basis for the left syzygy of R. Evidently, if
R is para-Hermitian, then U RU ? = diag(Φ 0) where
Φ ∈ Rm×m [ξ] is para-Hermitian and nonsingular.
d
Since S1 S1T = I, then [P̃ −Q̃]( dt
)S1T col(i v) =
d
[P −Q]( dt )col(i v). Thus, with i1 = T1 i, i2 = T2 i,
v1 = T1 v, and v2 = T2 v, it follows that i and v have
the compatible partitions i := (i1 , i2 ) and v := (v1 , v2 ),
and B̃ := B (col(i1 v2 ),col(v1 i2 )) takes the form of (3.2).
Thus, from Lemma 12, there exists a state-space sys(u,y)
. Moreover, it
tem Bs as in (1.2) such that B̃ = Bs
is easily verified that (P̃ , Q̃) is a positive-real pair since
(P, Q) is, so B̃ is passive by Theorem 13.
3 ⇒ 1. Note from the preceding discussion that B takes
the form of (3.1) where [P −Q] = [P̃ −Q̃]S1T . Now,
let (i, v) ∈ B and t0 ∈ R; let (î, v̂) ∈ B be a fixed but
arbitrary trajectory satisfying (î(t), v̂(t)) = (i(t), v(t))
for all t < t0 ; and let u := col(T1 i T2 v), y :=
col(T1 v T2 i), û := col(T1 î T2 v̂), and ŷ := col(T1 v̂ T2 î).
Then (u, y) ∈ B̃, (û, ŷ) ∈ B̃, and (û(t), ŷ(t)) =
(u(t), y(t)) for all t < t0 . Since B̃ is passive, there exists a K ∈ R such that − ∫tt01 ûT (t)ŷ(t)dt < K for all
t1 ≥ t0 . Since, in addition T1T T1 + T2T T2 = I, then
− ∫tt01 îT (t)v̂(t)dt = − ∫tt01 ûT (t)ŷ(t)dt < K, and we
conclude that B is passive. 2
B
Linear systems and behaviors
Here, we provide relevant results from behavioral theory
(see Polderman and Willems, 1998).
13
d
dt LΦR (w, x),
d
k
B1 Let B1 = {w ∈ Lloc
)w = 0}
| R1 ( dt
1 R, R
d
k
loc
and B2 = {w ∈ L1 R, R | R2 ( dt )w = 0} for some
R1 , R2 ∈ Rl×k [ξ]. Then B1 = B2 if and only if there
exists a unimodular U ∈ Rl×l [ξ] such that R1 = U R2
(Polderman and Willems, 1998, Theorem 3.6.2). The requirement that R1 and R2 have the same number of rows
is of little consequence since the addition or deletion of
rows of zeros doesn’t alter the behavior.
B2 Let F ∈ Rm1 ×n [ξ], G ∈ Rm2 ×n [ξ], and B:={z ∈
d
d
loc
L1 (R, Rn ) | F ( dt
)z = 0}. If z ∈ B implies G( dt
)z =
m2 ×m1
0, then there exists H ∈ R
[ξ] such that G =
n
HF . To see this, note that B = {z ∈ Lloc
1 (R, R ) |
d
col(F G)( dt )z = 0}. It follows from note B1 that there
exists a unimodular matrix U with U col(F 0m2 ×n ) =
col(F G). We form H from the last m2 rows and first
m1 columns of U to obtain G = HF .
B3 Consider a system B as in (1.1). B is called controllable if, for any two trajectories w1 , w2 ∈ B and t0 ∈ R,
there exists w ∈ B and t1 ≥ t0 such that w(t) = w1 (t)
for all t ≤ t0 and w(t) = w2 (t) for all t ≥ t1 (Polderman and Willems, 1998, Definition 5.2.2); and stabilizable if for any w1 ∈ B there exists w ∈ B such that
w(t) = w1 (t) for all t ≤ t0 and limt→∞ w(t) = 0 (Polderman and Willems, 1998, Definition 5.2.29). B is controllable (resp., stabilizable) if and only if the rank of
R(λ) is the same for all λ ∈ C (resp., λ ∈ C+ ) (Polderman and Willems, 1998, Theorems 5.2.10, 5.2.30).
C
Z
so, for any given t1 ≥ t0 ∈ R,
t1
T
d
R dt
w (t)x(t)dt
t0
Z t1
t
d
=
wT (t) RT − dt
x (t)dt + [LΦR (w, x)(t)]t10 .
t0
d
d
Note that if R( dt
) = dt
, then ΦR = 1, and this becomes
the formula for integration by parts.
D
States and state-space systems
In this final appendix, we provide several useful definitions and results concerning state-space systems.
D1 Let Bs be as in (1.2). Then, for any given u ∈
n
d
Lloc
1 (R, R ), x0 ∈ R , and t0 ∈ R, there exists a unique
(u, y, x) ∈ Bs with x(t0 ) = x0 , which is given by the
variation of the constants formula: x(t) = eA(t−t0 ) x0 +
∫tt0 eA(t−τ ) Bu(τ )dτ for all t ≥ t0 ; x(t) = eA(t−t0 ) x0 −
∫tt0 eA(t−τ ) Bu(τ )dτ for all t < t0 ; and y = Cx + Du.
D2 Let Bs be as in (1.2). We call the pair (C, A) observable if (u, y, x), (u, y, x̂) ∈ Bs imply x = x̂ (Polderman
and Willems, 1998, Definition 5.3.2). With the notation
Vo := col(C CA · · · CAd−1 ), then (C, A) is observable
if and only if rank(Vo ) = d (Polderman and Willems,
1998, Theorem 5.3.9). Now, let rank(Vo ) = d1 < d;
let the columns of S2 be a basis for the set {z ∈ Rd |
Vo z = 0}; let S = [S1 S2 ] be nonsingular; and partition
T := S −1 compatibly with S as T = col(T1 T2 ). Then
Bilinear and quadratic differential forms
"
Bilinear and quadratic differential forms were introduced in Willems and Trentelman (1998), and are useful
for studying dissipativity. Some relevant definitions and
results are presented here.
T1
T2
#
h
"
i
A S1 S2 =:
Ã11 0
Ã21 Ã22
#
h
i
h
i
, C S1 S2 =: C̃1 0 ,
where (C̃1 , Ã11 ) is observable: the observer staircase
form (Polderman and Willems, 1998, Corollary 5.3.14).
D3 Let Bs be as in (1.2); let Vo , S and T be as in
(u,y)
note D2; let B := Bs
; and let t0 ∈ R. If (u, y, x) ∈
Bs , (û, ŷ) ∈ B, and (û(t), ŷ(t)) = (u(t), y(t)) for all
t < t0 , then there exists (û, ŷ, x̂) ∈ Bs with x̂(t0 ) =
x(t0 ). This follows from the following two observations,
which are easily shown from the variation of the constants formula: (i) if (u, y) ∈ B and z ∈ Rd−d1 , then
there exists (u, y, x) ∈ Bs with T2 x(t0 ) = z; and (ii) if
(u, y, x), (û, ŷ, x̂) ∈ Bs , and (u(t), y(t)) = (û(t), ŷ(t))
for all t < t0 , then T1 x(t0 ) = T1 x̂(t0 ).
Also, with Ã11 and C̃1 as in note D2, and B̃1 := T1 B,
then it follows from the variation of the constants for(u,y)
mula that B = B̃s
, with
C1 A bilinear differential form is a mapping from
C∞ (R, Rm ) × C∞ (R, Rn ) to C∞ (R, R) of the form
PM PN di−1 w T
dj−1 x
LΦ (w, x) :=
j=1 ( dti−1 ) Φij ( dtj−1 ), for
i=1
some positive integers M, N and Φij ∈ Rm×n . It is
naturally associated with the two variable polynomial matrix Φ ∈ Rm×n [ξ, η] defined as Φ(ξ, η) :=
PM PN
i−1 j−1
η . If Φ ∈ Rm×m [ξ, η], then
j=1 Φij ξ
i=1
Qφ (w):=Lφ (w, w) is called a quadratic differential form.
C2 Let Φ ∈ Rm×n [ξ, η] and let Ψ(ξ, η) := (ξ +
η)Φ(ξ, η). Then the product rule of differentiation gives
d
dt LΦ (w, x) = LΨ (w, x).
C3 Associated with a given R ∈ Rm×n [ξ] is the bilinear differential form LΦR , where ΦR (ξ, η) := (R(ξ) −
R(−η))/(ξ +η). Since ξ = −η implies R(ξ)−R(−η) = 0,
then ΦR ∈ Rm×n [ξ, η] fromthe factor theorem. Further
d
d
w)T x − wT (RT − dt
x) =
more, from note C2, (R dt
n
loc
n
loc
B̃s ={(u, y, x̃) ∈ Lloc
R, Rd1
1 (R, R ) ×L1 (R, R ) ×L1
such that
14
dx̃
dt
= Ã11 x̃+B̃1 u and y = C̃1 x̃+Du}.
n
loc
D4 Let B̂ = {(u, x) ∈ Lloc
R, Rd |
1 (R, R ) × L1
dx
:= [B AB · · · Ad−1 B]. If
dt = Ax + Bu} and let Vc
B is controllable, then we also call the pair (A, B) controllable. From (Polderman and Willems, 1998, Section
5.2.1), the following are equivalent: (i) (A, B) is controllable; (ii) [λI−A B] has full row rank for all λ ∈ C;
and (iii) rank(Vc ) = d. Now, let Bs be as in (1.2); let
(u,y)
B:=Bs
; and let (C, A) be observable. Then B is controllable (resp., stabilizable) if and only if [λI−A B] has
full row rank for all λ ∈ C (resp., λ ∈ C+ ). The proof is
similar to (Hughes, 2016a, proof of Theorem 5.2).
Pal, D., Belur, M. N., 2008. Dissipativity of uncontrollable systems, storage functions, and Lyapunov functions. SIAM Journal on Control Optim. 47 (6), 2930–
2966.
Pandolfi, L., Mar 2001. An observation on the positive
real lemma. Journal of Mathematical Analysis and
Applications 255 (2), 480–490.
Polderman, J. W., 1997. Proper elimination of latent
variables. Systems and Control Letters 32 (5), 261–
269.
Polderman, J. W., Willems, J. C., 1998. Introduction
to Mathematical Systems Theory: A Behavioral Approach. New York : Springer-Verlag.
Rapisarda, P., Willems, J. C., 1997. State maps for linear systems. SIAM Journal on Control Optim. 35 (3),
1053–1091.
Trentelman, H. L., Willems, J. C., December 1997. Every storage function is a state function. Systems and
Control Letters 32 (5), 249–259.
Willems, J. C., December 1971. Least squares stationary
optimal control and the algebraic Riccati equation.
IEEE Trans. on Automatic Control 16 (6), 621–634.
Willems, J. C., 1972a. Dissipative dynamical systems,
Part I: General theory. Arch. Ration. Mech. Anal. 45,
321–351.
Willems, J. C., 1972b. Dissipative dynamical systems,
Part II: Linear systems with quadratic supply rates.
Arch. Ration. Mech. Anal. 45, 352–393.
Willems, J. C., Sept. 1986. From time series to linear
system—Part I. Finite dimensional linear time invariant systems. Automatica 22 (5), 561–580.
Willems, J. C., 2004. Hidden variables in dissipative systems. Proceedings of the 43rd IEEE Conference on
Decision and Control, 358–363.
Willems, J. C., 2007. Dissipative dynamical systems. European Journal on Control 13, 134–151.
Willems, J. C., Trentelman, H. L., September 1998. On
quadratic differential forms. SIAM Journal on Control
Optim. 36 (5), 1703–1749.
Wolovich, W. A., 1974. Linear Multivariable Systems.
New York : Springer-Verlag.
Youla, D. C., July 1961. On the factorization of rational
matrices. IRE Trans. Information Theory 7 (3), 172–
189.
Zhou, K., Doyle, J. C., Glover, K., 1996. Robust and
Optimal Control. New Jersey : Prentice Hall.
References
Anderson, B. D. O., Vongpanitlerd, S., 1973. Network
Analysis and Synthesis. Upper Saddle River, NJ:
Prentice-Hall.
Çamlibel, M. K., Willems, J. C., Belur, M. N., Dec.
2003. On the dissipativity of uncontrollable systems.
In: Proceedings of the 42nd IEEE Conference on Decision and Control, Hawaii. pp. 1645–1650.
Collado, J., Lozano, R., Johansson, R., July 2001.
On Kalman-Yakubovich-Popov lemma for stabilizable systems. IEEE Trans. on Automatic Control
46 (7), 1089–1093.
Feinstein, J., Bar-Ness, Y., 1980. On the uniqueness of
the minimal solution to the matrix polynomial equation A(λ)X(λ) − Y (λ)B(λ) = C(λ). Journal of the
Franklin Institute 310 (2), 131–134.
Ferrante, A., May 2005. Positive real lemma: Necessary
and sufficient conditions for the existence of solutions
under virtually no assumptions. IEEE Trans. on Automatic Control 50 (5), 720–724.
Ferrante, A., Pandolfi, L., October 2002. On the solvability of the positive real lemma equations. Systems
and Control Letters 47 (3), 211–219.
Gantmacher, F. R., 1980. The Theory of Matrices. Vol. I.
New York : Chelsea.
Hughes, T., 2017a. Passivity and electric circuits: a
behavioral approach. Proceedings of the 20th IFAC
World Congress, Toulouse.
Hughes, T. H., 2016a. Behavioral realizations using companion matrices and the Smith form. SIAM Journal
on Control Optim. 54(2), 845–865.
Hughes, T. H., 2016b. Controllability of passive singleinput single-output systems. In: 2016 European Control Conference (ECC). pp. 1087–1092.
Hughes, T. H., 2017b. On the optimal control of
passive or non-expansive systems. IEEE Trans.
on Automatic Control, accepted. Pre-print:
https://arxiv.org/abs/1703.08153.
Hughes, T. H., Smith, M. C., March 2017. Controllability of linear passive network behaviors. Systems and
Control Letters 101, 58–66.
Kunimatsu, S., Sang-Hoon, K., Fujii, T., Ishitobi, M.,
2008. On positive real lemma for non-minimal realization systems. Proceedings of the 17th IFAC World
Congress, Seoul, 5868–5873.
Timothy H. Hughes
received
the M.Eng. degree in mechanical
engineering, and the Ph.D. degree
in control engineering, from the
University of Cambridge, U.K., in
2007 and 2014, respectively.
From 2007 to 2010 he was employed
as a Mechanical Engineer at The
Technology Partnership, Hertfordshire, U.K; and from 2013 to 2017
he held a Research Fellowship at the University of Cam-
15
bridge. He is now a Lecturer at the Department of Mathematics at the University of Exeter.
16
| 3cs.SY
|
Faster Parallel Solver for Positive Linear Programs via
Dynamically-Bucketed Selective Coordinate Descent
arXiv:1511.06468v1 [cs.DS] 20 Nov 2015
Di Wang ∗
Michael W. Mahoney †
Nishanth Mohan ‡
Satish Rao §
Abstract
We provide improved parallel approximation algorithms for the important class of packing and covering
linear programs. In particular, we present new parallel ǫ-approximate packing and covering solvers which run
in Õ(1/ǫ2 ) expected time, i.e., in expectation they take Õ(1/ǫ2 ) iterations and they do Õ(N/ǫ2 ) total work,
where N is the size of the constraint matrix and ǫ is the error parameter, and where the Õ hides logarithmic
factors. To achieve our improvement, we introduce an algorithmic technique of broader interest: dynamicallybucketed selective coordinate descent (DB-SCD). At each step of the iterative optimization algorithm, the
DB-SCD method dynamically buckets the coordinates of the gradient into those of roughly equal magnitude,
and it updates all the coordinates in one of the buckets. This dynamically-bucketed updating permits us to take
steps along several coordinates with similar-sized gradients, thereby permitting more appropriate step sizes at
each step of the algorithm. In particular, this technique allows us to use in a straightforward manner the recent
analysis from the breakthrough results of Allen-Zhu and Orecchia [2] to achieve our still-further improved
bounds. More generally, this method addresses “interference” among coordinates, by which we mean the
impact of the update of one coordinate on the gradients of other coordinates. Such interference is a core issue
in parallelizing optimization routines that rely on smoothness properties. Since our DB-SCD method reduces
interference via updating a selective subset of variables at each iteration, we expect it may also have more
general applicability in optimization.
1 Introduction
Packing and covering problems are important classes of linear programs with many applications, and they have
long drawn interest in computer science in general and theoretical computer science in particular. In their generic
form, fractional packing problems can be written as the linear program (LP):
max{cT x : Ax ≤ b},
x≥0
m×n
where c ∈ Rn≥0 , b ∈ Rm
are all non-negative. Without loss of generality, one can scale the
≥0 , and A ∈ R≥0
coefficients, in which case one can write this LP in the standard form:
max{~1T x : Ax ≤ ~1},
x≥0
(1)
where A ∈ Rm×n
≥0 . The dual of this LP, the fractional covering problem, can be written in the standard form as:
min{~1T y : AT y ≥ ~1}.
y≥0
(2)
We denote by OPT the optimal value of the primal problem (1) (which is also the optimal value of the dual
problem (2)). In this case, we say that a vector x is a (1 − ǫ)-approximation for the packing LP if Ax ≤ ~1
∗ Department of Electrical Engineering and Computer Sciences, University of California at Berkeley, Berkeley, CA 94720.
[email protected]
† International Computer Science Institute and Department of Statistics, University of California at Berkeley, Berkeley, CA 94720.
[email protected]
‡ Department of Electrical Engineering and Computer Sciences, University of California at Berkeley, Berkeley, CA 94720.
[email protected]
§ Department of Electrical Engineering and Computer Sciences, University of California at Berkeley, Berkeley, CA 94720.
[email protected]
1
Email:
Email:
Email:
Email:
Luby and Nisan [13]
Allen-Zhu and Orecchia [2]
Our main results
Number of distributed iterations
2
O logǫ4 N
2
O logǫ3 N
2
log N log 1ǫ
O
ǫ2
Total work
2
O logǫ4 N × (N log n)
2
O logǫ3 N × N
2
log N log 1ǫ
×N
O
ǫ2
Table 1: Running time for several parallel solvers for packing and covering LP problems. N is the total number of non-zero elements in the constraint matrix. Our algorithm is randomized, so the number of distributed
iterations and total work are in expectation.
and ~1T x ≥ (1 − ǫ) OPT, and we say that y is a (1 + ǫ)-approximation for the covering LP if Ay ≥ ~1 and
~1T y ≤ (1 + ǫ) OPT.
In this paper, we describe improved parallel algorithms for packing LPs and covering LPs, improving the dependence on the error parameter ǫ from Õ(1/ǫ3 ) to Õ(1/ǫ2 ) for both the total work and the distributed iteration
count for both problems (1) and (2). Our approach follows the general approach of transforming non-smooth
LPs to smooth convex optimization problems and then applying an efficient first-order optimization algorithm.
Unfortunately, the smoothed objective that arises does not have particularly Lipschitz continuity properties, and
thus we are unable to use traditional optimization methods to improve the (parallel) convergence rate beyond
Õ(1/ǫ3 ). Thus, to achieve our improvement to Õ(1/ǫ2 ), we develop the dynamically-bucketed selective coordinate descent (DB-SCD) method. This descent method involves partitioning coordinates into buckets based on
the magnitudes of their gradients and updating the coordinates in one of the buckets. This permits us to make
relatively large gradient moves along a subset of the coordinates for which we can control the smoothness of
gradients within the range of our step. Given that controlling the smoothness properties of functions is central
to controlling the convergence rate of continuous optimization algorithms, we expect that this method will be
useful more generally.
1.1 Overview of Prior Methods
Although one can use general LP solvers such as interior point methods to solve packing and covering problems
with a convergence rate of O(log(1/ǫ)), such algorithms usually have very high per-iteration cost, as methods
such as the computation of the Hessian and matrix inversion are involved. In the setting of large-scale problems,
low precision iterative solvers are often more popular choices. Such solvers usually run in time with a much
better dependence on the problem size, but they have the much worse poly(1/ǫ) dependence on the approximation parameter. Most such work falls into one of two categories. The first category follows the approach of
transforming LPs to convex optimization problems, then applying efficient first-order optimization algorithms.
Examples of work in this category include [1, 2, 4, 14, 15, 18], and all except [1, 2] apply to more general classes
of LPs. The second category is based on the Lagrangian relaxation framework, and some examples of work
in this category include [8, 10, 13, 17, 25, 26]. For a more detailed comparison of this prior work, see Table 1
in [1]. Based on whether the running time depends on the width ρ, a parameter which typically depends on
the dimension or the largest entry of A, these algorithms can also be divided into width-dependent solvers and
width-independent solvers. Width-dependent solvers are usually pseudo-polynomial, as the running time depends on ρ OPT, which itself can be large, while width-independent solvers are more efficient, in the sense that
they provide truly polynomial-time approximation solvers.
The line of research associated with width-independent
solvers was initiated by Lubyand Nisan [13], where
the authors gave a parallel algorithm runs in Õ 1/ǫ4 distributed iterations and Õ N/ǫ4 total work. Note that,
since we are most interested here in the dependence on the error parameter ǫ, to simplify the discussion and
notation, we will follow the standard practice of using Õ to hide poly-log factors. For readers interested in the
more precise results, see Table 1.1 in this section as well as our analysis below. For sequential algorithms, on
the total work front, a recent breakthrough gives an Õ(N/ǫ) sequential algorithm for packing and a different
Õ(N/ǫ1.5 ) sequential algorithm for covering [1]. Our recent work improved this by developing a diameter
reduction method that leads to a unified framework that achieves an Õ(N/ǫ) sequential algorithm for both
packing and covering [23]. In terms of parallel algorithms, improvement over the Õ(1/ǫ4 ) iteration count and
Õ(N/ǫ4 ) total work in the original paper of Luby and Nisan [13] was only achieved recently. In particular,
Allen-Zhu and Orecchia [2] gave a deterministic algorithm with Õ(1/ǫ3 ) iterations and Õ(N/ǫ3 ) total work.
2
1.2 Our Main Results
In this paper, we describe improved parallel algorithms for packing LPs and for covering LPs, improving the
dependence on the error parameter ǫ for total both work and distributed iteration count for both problems (1)
and (2) from Õ(1/ǫ3 ) to Õ(1/ǫ2 ). In particular, we present a stochastic parallel solver that provides a (1 − ǫ)approximation for primal packing LPs of the form (1). The solver is a width-independent first-order method.
It is a stochastic method, and it converges not only in expectation, but also with at least constant probability.
Furthermore, our solver has the additional guarantee that the objective function value is non-increasing across
iterations. In general, stochastic first order methods, such as coordinate descent and stochastic gradient descent,
show the expectation of the objective function converges to optimum, without the monotonicity guarantee on
the actual objective (e.g., [7, 12, 16, 20]). In practice, when the constraints in the problem are ill-conditioned or
highly non-separable, the objective function value may fluctuate heavily during execution of stochastic methods,
and this has motivated the development of more robust stochastic algorithms (e.g., [9]).
More precisely, here is our main theorem for the fractional packing problem.
Theorem 1.1. There is a randomized algorithm that, with probability at least 9/10, computes a (1 − O(ǫ))approximation to the fractional packing problem, has Õ(N/ǫ2 ) total work, and is implementable in Õ(1/ǫ2 )
distributed iterations.
In addition to this result for the primal packing problem, as given by (1), we can use the primal packing solver
to get a (1 + ǫ)-approximation to the dual covering problem, as given by (2). Here is our main theorem for the
fractional covering problem.
Theorem 1.2. There is a randomized algorithm that, with probability at least 9/10, computes a (1 + O(ǫ))approximation to the fractional covering problem, has Õ(N/ǫ2 ) total work, and is implementable in Õ(1/ǫ2 )
distributed iterations.
That is, our packing solver and our covering solver have Õ(1/ǫ2 ) expected iterations and Õ(N/ǫ2 ) expected
total work. Among other things, this gives an expected improvement of Õ(1/ǫ) over the current fastest parallel
algorithm of Allen-Zhu and Orecchia [2], in terms of both iteration count and total work. See Table 1.1 for a
more detailed comparison with their results as well as the results of Luby and Nisan [13]. See also Section 2,
which contains a more detailed statement of these results as well as the algorithms that achieve these results.
1.3 Our Main Techniques
The general approach of transforming non-smooth LPs into smooth convex optimization problems and then
applying efficient first-order optimization algorithm has been done by Nesterov [15], Allen-Zhu and Orecchia [1,
2], and many others. In particular, following [1, 2], to find a (1 − ǫ)-approximation of a packing LP, we will
approximately minimize, over the region x ≥ ~0, the following convex function:
fµ (x) = −~1T x + µ
m
X
1
exp( ((Ax)j ) − 1).
µ
j=1
(3)
In Section 3.1, we will discuss the motivation of using fµ (x) to solve our packing LP as well as properties of the
parameter µ and the function fµ (x). Here, we will focus on the main techniques we had to introduce in order to
obtain our improved solver.
To do so, recall that first-order methods in optimization exploit the idea of using the negative of the gradient
as the descent direction for each step of an iterative algorithm. In order to lower bound the improvement of
successive steps of the algorithm, smoothness is at the core of the analysis. The reason is basically since we
want to move in the descent direction as far as possible, without changing the gradient by too much. This is
most commonly captured by proving the Lipschitz continuity property with parameter L on the function f , i.e.,
we want to find an L ∈ R+ such that
k∇f (x) − ∇f (y)k∗ ≤ Lkx − yk
∀x, y ∈ domain(f ).
(4)
While popular, the Lipschitz continuity property is often much stronger than necessary. For example, instead of
controlling the properties of the gradient for all x, y pairs, in most gradient based methods, we actually only need
to show the smoothness of gradients within the range of our step—which, by design, we can control (indeed,
3
step sizes are often chosen based on this). This motivates the use of weaker continuity properties by focusing on
more constrained cases of x, y pairs that are still sufficient to lower bound the improvement of successive steps
of the algorithm.
Our basic improvement with the DB-SCD method comes from updating only a carefully-chosen subset of
variables in each iteration. In particular, we bucket the coordinates of the gradient into buckets of roughly
equal gradient magnitude, according to (9) below, and then we update the coordinates in one of the buckets,
according to Step 7 of Algorithm 1 below. While our particular approach is novel, the basic idea is hardly new
to optimization. Indeed, for most non-trivial functions, variables “interfere” with each other, in the sense that
variable i’s gradient will be affected by the update of variable j (e.g., [6]). Thus, if we aim to move the variables
while maintaining smoothness of the gradients, we have to take interference into consideration. In general, this
limits the possible step size. The global Lipschitz continuity parameter usually suffers from interference, since
the Lipschitz property (4) needs to hold for all x, y pairs, which allows arbitrary interference.
One way to alleviate the problem of interference is to update fewer variables in each iteration. This potentially
permits better control over the changes of gradients for the updated variables, since for the variables not updated,
the changes of their gradients don’t affect the objective improvement for that iteration. One extreme of this idea
is the coordinate descent method [24], where in each iteration only one variable is updated. In this case, the
step length of the update on the single variable is often larger than the step length when all variables are moved
simultaneously. On the other hand, in most cases the computation of n successive partial derivatives can be
more expensive than the computation of all the n partial derivatives for a fixed x, limiting the applicability of
the coordinate descent method. When the tradeoff is good between the gain in the step length versus the loss in
the computation, coordinate descent can be better than gradient descent in terms of total work ( [1, 11]). More
generally, in the context of solving linear systems, we have the example of Jacobi iterations versus Gauss-Seidel
iterations, and similar tradeoffs between interference and running time arise ( [3, 5], Chapter 4 of [21]). Still
more generally, various efforts to parallelize coordinate descent can be seen as explorations of tradeoffs among
the smoothness parameter, the amount of computation, as well as the distributed iteration count ( [6, 7, 19, 20]).
To the best of our knowledge, all such works mentioned exhibit an inverse relationship between the number
of variables updated each iteration, and the number of total iterations required, i.e., when fewer variables are
updated, then more iterative steps are needed. This is what one would naturally expect. Moreover, the prior
works mentioned mostly choose the subset of variables to update either by some fix order, e.g., uniformly at
random, or according to some static partition constructed from the sparsity structure of the given instance, e.g.,
the objective function is separable or the matrix in the problem is block diagonal ( [22]). Rarely, if at all, is
a subset of variables chosen dynamically using knowledge of the actual values of the gradients. Again, this is
what one would naturally expect. For example, as Nesterov wrote in his seminal accelerated coordinate descent
work [16], if one already computed the whole gradient, then full-gradient methods seem to be better options.
With respect to both of these considerations, our method of selective coordinate descent is novel and quite
different. First, at least for the case of packing and covering LPs, we can achieve better parallel running time and
better total work by updating fewer (carefully-selected) variables each iteration. Second, our work shows that the
extra computation of the whole gradient can help us select a better subset dynamically (even if we don’t update
all coordinates). Our results in this paper show that both of these directions are worth additional exploration.
Finally, we emphasize that a less obvious benefit of our approach is that the gradients in most cases contain
useful information about the dual problem, and if we have the whole gradients from all iterations, then we can
exploit the primal-dual structure of the packing problem to obtain a solution to the covering problem.
2 Faster Parallel Solver for Packing LPs and Covering LPs
In this section, we will present our main results, including a statement of our main algorithm and theorem for
a parallel solver for packing LPs (in Section 2.1), a statement of our main algorithm and theorem for a parallel
solver for covering LPs (in Section 2.2), and a description of the main technical ideas underlying our proofs (in
Section 2.3).
2.1 Algorithm and Main Theorems for Parallel Packing LP Solver
Our main algorithm to find a (1 − ǫ)-approximation of a packing LPs is specified in Algorithm 1. This algorithm
takes as input the matrix A as in (1), the smoothed objective function fµ , and the approximation parameter ǫ. It
1
returns as output xT , such that with constant probability 1+ǫ
xT is a (1 − O(ǫ))-approximation to the packing
4
problem. In this algorithm, we use x[i] to denote the i-th coordinate of vector x, except with matrix-vector
multiplications, where we use (Ax)i to denote the value of the i-th component of Ax.
Algorithm 1 Stochastic and Parallelizable Packing LP Solver
n
Input: A ∈ Rm×n
≥0 , fµ , ǫ ∈ (0, 1/2] Output: x ∈ R≥0
µ
ǫ
1: µ ← 4 log(nm/ǫ) , α ← 20
⌈10 log(1/ǫ) log(2n)⌉
2: T ←
= Õ( ǫ12 ), w ← log( 1ǫ )
αǫ
1−ǫ/2
3: x0 ← nkA k
:i ∞
4: for k = 0 to T − 1 do
5:
Select t ∈ {0, . . . , w − 1} uniformly at random
6:
for i = 1 to n do
(t)
7:
Gradient truncation and coordinate selection: Compute ∇i fµ (xk ), and get ξk [i], as defined in (9)
def
(t)
(t)
8:
Update step: xk+1 [i] ← xk+1 [i] = xk [i] exp(−αξk [i])
9:
end for
10: end for
11: return xT .
To understand the steps of Algorithm 1, recall that the function fµ (x) referred to in the algorithm is given
by Eqn. (3) and is a smoothed version of the packing objective of (1). Then, following [2], in each iteration
k ∈ [0, T − 1], for each variable xk [i], we can break the gradients ∇i fµ (xk ) into small, medium, and large
components. That is, we can let
∇i fµ (xk ) ∇i fµ (xk ) ∈ [−ǫ, ǫ]
ζk [i] =
(5)
0
otherwise
∇i fµ (xk ) ∈ [−ǫ, ǫ]
0
ξk [i] =
∇i fµ (xk ) ∇i fµ (xk ) ∈ [−1, 1]\[−ǫ, ǫ]
(6)
1
∇i fµ (xk ) > 1
∇i fµ (xk ) − 1 ∇i fµ (xk ) > 1
ηk [i] =
(7)
0
otherwise
denote, respectively, the small, medium, and large components of the gradient. In particular, from this decomposition, we have
∇fµ (xk ) = ζk + ξk + ηk .
(8)
(It was by adopting this partitioning that previous work achieved their Õ(1/ǫ3 ) running time [2].)
In Lemma 3.4 below, we will establish that if the gradients are all within a factor of (say) 2 from each
other, then we can take a multiplicative step with step size α = Θ(µ) = Õ(ǫ). To exploit this algorithmically,
and to lead to our improved algorithmic results, we will further partition the variables into groups such that,
for variables in the same group, the absolute values of their truncated gradients will be within a factor 2 of
each other. In particular, we will further partition the medium component into groups or buckets in which the
truncated gradients are of roughly equal magnnitude. To do this, for t ∈ {0, . . . , ⌈log( 1ǫ ) − 1⌉}, we let
ξk [i] ξk [i] ∈ (ǫ2t , ǫ2t+1 ]
ηk [i] t = ⌈log( 1ǫ )⌉ − 1
(t)
(t)
t+1
t
∪[−ǫ2 , −ǫ2 )
.
(9)
ξk [i] =
and ηk [i] =
0
otherwise
0
otherwise
Then, in each iteration of Algorithm 1, we will pick a bucket t uniformly at random, and we will update all
(t)
variables using ξk .
Our main result for Algorithm 1 is summarized in the following theorem, the proof of which we will present
in Section 3.
Theorem 2.1. Algorithm 1 outputs xT satisfying E[fµ (xT )] ≤ −(1 − 5ǫ) OPT, and the algorithm can be
log2 N
log2 N
) iterations with total work O(N × log(1/ǫ)
), where N is the total
implemented with O( log(1/ǫ)
ǫ2
ǫ2
number of non-zeros in A.
5
Remark. Our main Theorem 1.1 follows almost immediately from Theorem 2.1, when combined with a standard
application of Markov bound and part (5) of Lemma 3.3 below. In particular, by Lemma 3.3(2), for every
x ≥ 0, fµ (x) ≥ −(1 + ǫ) OPT. From Theorem 2.1, we have that fµ (xT ) + (1 + ǫ) OPT is a non-negative
random variable with expectation at most 4ǫ. Using Markov’s inequality, with at least probability 9/10, we have
fµ (xT ) ≤ −(1 − 41ǫ) OPT, giving a (1 − O(ǫ) approximation by Lemma 3.3(5). The total work and iteration
count follow directly from Theorem 2.1, thus proving our main Theorem 1.1.
2.2 Algorithm and Main Theorems for Parallel Covering LP Solver
A benefit of computing all the gradients in each iteration of Algorithm 1 is that—even if we don’t use them
for our packing solver—we can exploit the same primal-dual structure as in [2] to get a covering LP solver. In
particular, given a covering LP instance in the form of (2), we can construct its dual, which is a packing LP. If
we then run Algorithm 1 on the packing instance for T iterations, then the average of the exponential penalties
used in the computation of gradients, i.e.,
ȳ =
T −1
1 X −−−→
p(xk ) ≥ 0,
T
(10)
k=0
will, with constant probability, be a (1+ǫ)-approximation of the covering problem, after some simple fixing step.
The fixing step is to post-processing ȳ to enforce feasibility of the covering solution, as ȳ will only be feasible
−−−→
with high probability. Here p(xk ) is the vector of all the exponential penalties of the packing constraints, as
defined in Lemma 3.2, which we compute in each iteration of Algorithm 1 to get the gradients. To obtain the
dual solution, the primal-dual property we will exploit is that the slackness of the i-th covering constraint with ȳ
is the average gradient of the i-th variable in the primal packing LP:
(AT ȳ)i − 1 =
T −1
T −1
1 X
1 X T −−−→
(A p(xk ))i − 1 =
∇i fµ (xk ).
T
T
k=0
k=0
Following a similar approch as in [2, 27], we present our fixing procedure in Algorithm 2. Note, in particular,
that this explicitly makes all the dual constraints satisfied.
Algorithm 2 Post-processing ȳ to enforce feasibility of the dual solution for the Covering LP solver
m
T
~
Input: A ∈ Rm×n , ǫ ∈ (0, 1/10], ȳ ∈ R≥0
Output: y ∈ Rm
≥0 such that A y ≥ 1.
≥0
1:
2:
3:
4:
5:
6:
ȳ ′ ← ȳ
def
for all i such that λi = (AT ȳ)i − 1 + ǫ ≤ −2ǫ do
′
Let j = argmaxj Ai,j ′ , i.e. Ai,j = kA:i k∞ .
−λi
ȳj′ ← ȳj′ + A
.
i,j
end for
ȳ ′
return y = 1−3ǫ
.
Overall, given a covering LP instance in the form of (2), the entire covering solver consists of the following.
• First, construct its dual, which is a packing LP in the form of (1).
• Then, run Algorithm 1 for T iterations, with T ≥ max{ 6w
αǫ ,
2w 2 log
ǫ2
n
ǫ
}.
• Finally, fix the ȳ as in (10) with Algorithm 2, which takes O(log(N )) time and O(N ) work.
If y is the output of Algorithm 2, then y is feasible by construction, i.e., y ≥ 0, AT y ≥ ~1, and moreover we can
establish the following result.
Theorem 2.2. E[~1T y] ≤ (1 + 10ǫ) OPT, and y ≥ 0, AT y ≥ ~1.
Remark. Our main Theorem 1.2 follows almost immediately from Theorem 2.2, when combined with a standard
application of Markov bound and several lemmas below. From Theorem 2.2, we have E[~1T y] ≤ (1 + 10ǫ) OPT,
and ~1T y > OPT, since y is always feasible. Then y − OPT is a non-negative random variable with expectation
6
at most 10ǫ. Using Markov’s inequality, with at least probability 9/10, we have ~1T y ≤ (1 + 100ǫ) OPT, giving
a (1 + O(ǫ) approximation. The expected total work and iteration count is dominated by Algorithm 1. We
show in Lemma 4.1 and Lemma 4.3 that runing Algorithm 1 for T = max{ 6w
αǫ ,
iterations is sufficient. This proves our main Theorem 1.2.
2w 2 log
ǫ2
n
ǫ
} = O(
log 2 ( 1ǫ ) log N
)
ǫ2
2.3 Discussion of Main Technical Ideas Underlying Our Proofs
Before proceeding with our proofs of Theorems 2.1 and 2.2 in Sections 3 and 4, repsectively, we provide here a
discussion of the main technical ideas underlying our methods.
At a high level, we (as well as Allen-Zhu and Orecchia [1, 2]) use the same two-step approach of Nesterov [15]. The first step involves smoothing, which transforms the constrained problem into a convex and
smooth objective function with trivial or no constraints. (By smooth, we mean that the gradient of the objective
function has some property in the flavor of Lipschitz continuity.) In our case, we optimize the function fµ (x)
with the trivial constraints x ≥ ~0, and the smoothness property of fµ (x) is specified in Lemma 3.4. Once
smoothing is accomplished, the second step uses one of several first order methods for convex optimization in
order to obtain an approximate solution. Examples of standard application of this approach to packing and covering LPs includes the width-dependent solvers of [14, 15] as well as multiplicative weights update solvers [4].
In these examples, the dependence of OPT is introduced, when the entropy function is used in the smoothing
step. The dependence on ρ is from using the full gradient, which can be large if there are large entries in A.
We will see in Section 3.1 that fµ (x) is the objective function derived from a different smoothing step,
which avoids the dependence of OPT. The function fµ (x) is used in [2], which is the first width-independent
result following the optimization approach, as well as in later works [1, 23]. The different smoothing alone is
not enough for width-independence, however, since in the optimization step, the convergence of the first order
method will depend on the largest absolute value of the gradient or feedback vector. We use the same idea as
in [1, 2] to use truncated gradient in our algorithm, thus effectively reducing the width to 1. More specifically,
the regret term in standard mirror descent analysis depends on the width, while in Lemma 3.8, our regret term
α2 OPT doesn’t.
A major issue that arises with this approach is that the convex objective function fµ (x) is not smooth in
the standard Lipschitz continuity sense. The authors in [2] work around it by showing that the gradient is
multiplicatively smooth within some local neighborhood, and they constrain their update steps to stay inside that
region. The bottleneck of their convergence, as we see it, is that the step size of the update is too small due to
interference (recall that interference is the dependence of variable i’s gradient on the value of other variables)
between different coordinates. In a typical iteration, they aim to move all variables simultaneously proportional
to the respective gradients, without changing the gradient of any variable by too much. When the gradients of
the variables are not on the same scale, a natural obstacle arises: when variable i has a large gradient, we would
like to move i by a large step in order to harness the gradient improvement, but we are prevented from doing so,
due to the interference of i with some other variable j, which has a tiny gradient.
We tackle this bottleneck by designing in a dynamic manner selective coordinate descent steps designed to
reduce interference. In each iteration, we group all the variables according to the magnitudes of their gradients
such that variables in the same group all have approximately the same magnitudes up to some constant factor, as
specified in (9) in Section 2.1. If we then only update variables of one randomly chosen group, as in Step 7 of
Algorithm 1, then we can take larger steps for the subset of variables we update. In addition, we show that we
only need log(1/ǫ) groups, so each iteration we update a large fraction of the variables on expectation.
To make our analysis work out, we follow [2], and we interpret the update step as both a gradient descent step
and a mirror descent step. (We note, though, that this is different from [1,28], where the gradient descent step and
the mirror descent step are separate steps, and the algorithm takes a linear coupling of the two steps to achieve
acceleration as in Nesterov’s accelerated gradient descent. In [2] and in our algorithm, it is a single update step
each iteration, interpreted as both a gradient step and a mirror descent step.) This two-way interpretation is
required when we use convexity to upper bound the gap between fµ (xk ) and fµ (u) for some xk , u. Recall we
break the gradients into small, medium and large components as in (8), so we have:
fµ (xk ) − fµ (u) ≤ h∇fµ (xk ), xk − ui = hζk , xk − ui + hξk , xk − ui + hηk , xk − ui.
We will see that in expectation, the loss incurred by the medium component will be bounded using the mirror
descent interpretation in Lemma 3.8, and the loss incurred by the large component will be bounded using the
gradient descent interpretation in Lemma 3.5.
7
A benefit of our DB-SCD method is that it is modular and thus it can be coupled cleanly with existing
methods. To illustrate this, since our interest in this particular problem was inspired by the original improvement
of [2], we will to the extent possible adopt the techniques from their work, pointing out similarities in the
following sections whenever possible.
3 Analysis of Algorithm for Packing LPs: Proof of Theorem 2.1
In this section, we will provide a proof of Theorem 2.1. Recall that we denote by OPT the optimal value of (1)
and that Algorithm 1 will compute a (1 − ǫ)-approximation x where Ax ≤ ~1 and ~1T x ≥ (1 − ǫ) OPT. In
Section 3.1, we’ll present some preliminaries and describe how we perform smoothing on the original packing
objective function. We’ll analyze the update step as a gradient descent step in Section 3.2, and we’ll analyze the
same update step as a mirror descent step in Section 3.3. Finally, in Section 3.4, we’ll show how to combine
the two analyses to complete the proof of Theorem 2.1. Some of the following results are technically-tedious
but conceptually-straightforward extensions of analogous results from [2], and some of the results are restated
from [2]. For completeness, we provide the proof of all of these results, with the latter relegated to Appendix A.
3.1 Preliminaries and Smoothing the Objective
To start, let’s assume that
min kA:i k∞ = 1.
i∈[n]
This assumption is without loss of generality: since we are interested in multiplicative (1 − ǫ)-approximation,
we can simply scale A for this to hold without sacrificing approximation quality. With this assumption, the
following lemma holds. (This lemma is the same as Proposition 2.2.(a) in [2], and its proof is included for
completeness in Appendix A.)
Lemma 3.1. OPT ∈ [1, n]
With OPT being at least 1, the error we introduce later in the smoothing step will be small enough that the
smoothing function approximates the packing LP well enough with respect to ǫ around the optimum.
We will turn the packing LP objective into a smoothed objective function fµ (x), as used in [1, 2], and we
are going to find a (1 − ǫ)-approximation of the packing LP by approximately minimizing fµ (x) over the region
x ≥ 0. The function fµ (x) is
fµ (x) = −~1T x + max{y T (Ax − ~1) + µH(y)},
def
y≥0
and it is a smoothed objective in the sense that it turns the packing constraints
P into soft penalties, with H(y)
being a regularization term. Here, we use the generalized entropy H(y) = − j yj log yj + yj , where µ is the
smoothing parameter balancing the penalty and the regularization. It is straightforward to compute the optimal
y, and write fµ (x) explicitly, as stated in the following lemma.
Pm
def
Lemma 3.2. fµ (x) = −~1T x + µ j=1 pj (x), where pj (x) = exp( µ1 ((Ax)j ) − 1).
Optimizing fµ (x) gives a good approximation to OPT, in the following sense. If we let x∗ be an optimal
def
solution, and u∗ = (1 − ǫ/2)x∗ , then we have the properties in the following lemma. (This lemma is the same
as Proposition 2.2 in [2], and its proof is included for completeness in Appendix A.)
Lemma 3.3. Setting the smoothing parameter µ =
ǫ
4 log(nm/ǫ) ,
we have
1. fµ (u∗ ) ≤ −(1 − ǫ) OPT.
2. fµ (x) ≥ −(1 + ǫ) OPT for every x ≥ 0.
3. Letting x0 ≥ 0 be such that x0 [i] =
1−ǫ/2
nkA:i k∞
for each i ∈ [n], we have fµ (x0 ) ≤ − 1−ǫ
n .
4. For any x ≥ 0 satisfying fµ (x) ≤ 0, we must have Ax ≤ (1 + ǫ)~1, and thus ~1T x ≤ (1 + ǫ) OPT.
5. If x ≥ 0 satisfies fµ (x) ≤ −(1 − O(ǫ)) OPT, then
8
1
1+ǫ x
is a (1 − O(ǫ))-approximation to the packing LP.
6. The gradient of fµ (x) is
−−→
∇fµ (x) = −~1 + AT p(x) where
and ∇i fµ (x) = −1 +
P
j
1
def
pj (x) = exp( ((Ax)j − 1),
µ
Aji pj (x) ∈ [−1, ∞].
Although fµ (x) gives a good approximation to the packing LP without introducing dependence of OPT, we
cannot simply apply the standard (accelerated) gradient descent algorithm to optimize it, as fµ (x) doesn’t have
the necessary Lipschitz-smoothness property. (Indeed, our DB-SCD method was designed to address this issue.)
def
(t)
(t)
As before [23], we interpret our update step xk+1 [i] ← xk+1 [i] = xk [i] exp(−αξk [i]) in Algorithm 1 as
both a gradient descent step as well as a mirror descent step. That is, in order to prove the theorem, our analysis
will view it from both perspectives. We proceed now with the respective analysis for the two interpretations.
3.2 Gradient Descent Step
We will first analyze the update step in Algorithm 1 as a gradient descent step. As in most gradient descent
analysis, we need to bound our step’s impact on the gradients. To do so, we will show fµ (x) is locally multiplicative Lipschitz continuous, in a sense quantified by the following lemma. Note the result is a stronger version
of Proposition 3.6 in [2], in the sense that the step size α is 1/ǫ larger. This improvement is achieved due to the
reduced interference from our DB-SCD updating method.
(t)
(t)
Lemma 3.4. Let xk+1 [i] = xk [i] exp(−αξk [i]), for any t = 0, . . . , w − 1, as in Algorithm 1. Let Bt =
(t)
(t)
{i|ξk [i] > 0} be the set of variables we update. If Axk ≤ (1 + ǫ)~1, then for any x = τ xk + (1 − τ )xk+1 where
τ ∈ [0, 1], we have ∀i ∈ Bt , ∇i fµ (x) is between 12 ∇i fµ (xk ) and 23 ∇i fµ (xk ).
(t)
Proof. Because for all i ∈ Bt , ξk [i] ∈ (ǫ2t , ǫ2t+1 ] ∪ [−ǫ2t+1 , −ǫ2t ), each variable changes multiplicatively
by at most exp(±αǫ2t+1 ), and since αǫ2t+1 ≤ 1/4, we must have for all i,
8
8
x[i] ∈ xk [i] · [1 − αǫ2t , 1 + αǫ2t ].
3
3
Now we look at the impact of the step on the exponential penalties
(11)
1
pj (x) = exp( ((Ax)j − 1)).
µ
Due to (11), and (Axk )j ≤ (1 + ǫ) for any j, we have
|(Ax)j − (Axk )j | ≤
8
10
αǫ2t (Axk )j ≤
αǫ2t .
3
3
Then by our choice of α, we have
pj (x) ≥ pj (xk ) exp(−
t
Since ǫ2t ≤ 1, we have exp(− ǫ26 ) ≥ (1 −
ǫ2t
4 ).
10αǫ2t
ǫ2t
) = pj (xk ) exp(− ).
3µ
6
By a similar calculation for the upper bound, we have
pj (x) ∈ pj (xk ) · [1 −
ǫ2t
ǫ2t
,1 +
].
4
4
(12)
(t)
For any i ∈ Bt , if ξk [i] ∈ (ǫ2t , ǫ2t+1 ], we have
∇i fµ (x) = (AT p(x))i − 1
ǫ2t
)−1
4
ǫ2t
= (∇i fµ (xk ) + 1)(1 −
)−1
4
∇i fµ (xk )
,
≥
2
> (AT p(xk ))(1 −
(t)
where the last step is due to ǫ2t ≤ 1 and ∇i fµ (xk ) ≥ ξk [i] > ǫ2t . By similar calculation, we get ∇i fµ (x) ≤
(t)
3
t+1
, −ǫ2t ).
2 ∇i fµ (xk ). The same holds for the case ξk [i] ∈ [−ǫ2
9
We will see in Claim 3.6 that the condition of Axk ≤ (1 + ǫ)~1 holds for all k = 0, . . . , T . Once we establish
smoothness of the gradients within the range of our update step, we can lower bound the improvement we make.
(t)
In particular, the term hαηk , xk − ui is the loss incurred from the truncation, as our update step doesn’t act on
the truncated part, but it shows up when we use convexity to bound the gap to optimality.
Lemma 3.5. For all t = 0, . . . , w − 1, any u ≥ 0
(t)
(t)
hαηk , xk − ui ≤ 4(fµ (xk ) − fµ (xk+1 )).
Proof. First observe that
(t)
fµ (xk ) − fµ (xk+1 ) =
=
Z
0
1
(t)
(t)
(t)
h∇fµ (xk+1 + τ (xk − xk+1 )), xk − xk+1 idτ
1
XZ
(t)
i∈Bt
(t)
(t)
∇i fµ (xk+1 + τ (xk − xk+1 ))dτ × (xk [i] − xk+1 [i]),
0
(t)
(t)
where the last equality is because xk [i] − xk+1 [i] = 0 for i 6∈ Bt . By Lemma 3.4, we have that ∇i fµ (xk+1 +
(t)
τ (xk − xk+1 ))
(t)
has the same sign as ∇i fµ (xk ) for all τ ∈ [0, 1]. Furthermore, by our update rule, xk [i] − xk+1 [i]
(t)
also has the same sign as ∇i fµ (xk ), and so we have fµ (xk ) − fµ (xk+1 ) ≥ 0 for all t. If t < w − 1, then we
(t)
know η = ~0, and thus
k
(t)
(t)
hαηk , xk − ui = 0 ≤ 4(fµ (xk ) − fµ (xk+1 )).
(t)
When t = w − 1, let B = {i|∇i fµ (xk ) > 1} ⊇ Bt be the set of variables with nonzero ηk [i], we know for
(t)
(t)
i ∈ B, ξk [i] = 1, so xk+1 [i] = xk [i] exp(−α), and
(t)
fµ (xk ) − fµ (xk+1 ) =
≥
XZ
1
i∈Bt
0
XZ
1
i∈B
0
(t)
(t)
(t)
∇i fµ (xk+1 + τ (xk − xk+1 ))dτ × (xk [i] − xk+1 [i])
(t)
(t)
(t)
∇i fµ (xk+1 + τ (xk − xk+1 ))dτ × (xk [i] − xk+1 [i])
X1
∇i fµ (xk ) × xk [i](1 − exp(−α))
2
Xα
≥
∇i fµ (xk )xk [i].
4
≥
i∈B
i∈B
The first inequality is due to Bt ⊆ B, and every i has non-negative contribution to the sum. The second
inequality is from Lemma 3.4, and the last inequality is because (1 − exp(−α)) > α/2 when α < 1/10. Then
we have
X
(t)
(t)
hαηk , xk − ui ≤
α∇i fµ (xk )xk [i] ≤ 4(fµ (xk ) − fµ (xk+1 )),
i∈B
(t)
where the first inequality is because ∇i fµ (xk ) > ηk ≥ 0 in this case, and u ≥ 0.
(t)
We see fµ (xk ) − fµ (xk+1 ) ≥ 0 for any t = 0, . . . , w − 1, we have the following.
Claim 3.6. fµ (xk ) is non-increasing with k. By part (3), (4) of Lemma 3.3, Axk ≤ (1 + ǫ)~1, and ~1T xk ≤
(1 + ǫ) OPT for all k.
3.3 Mirror Descent Step
We now interpret the update step as a mirror descent step. We use the same proximal setup as in [2]. The distance
generating function will be the generalized entropy function, where
X
def
x[i] log(x[i]) − x[i],
w(x) =
i∈[n]
10
and the corresponding Bregman divergence function will be
X
Vx (y) =
(y[i] log
i∈[n]
y[i]
+ x[i] − y[i]).
x[i]
This is the standard proximal setup when one works with L1 -norm with the simplex as the feasible region. In our
case, since the feasible region is x ≥ 0, we don’t have the standard strong convexity of the Bregman divergence,
but one can verify
Vx (y) =
X
X (x[i] − y[i])2
y[i]
+ x[i] − y[i]) ≥
.
x[i]
2 max{x[i], y[i]}
(y[i] log
i∈[n]
(13)
i∈[n]
To interpret the update step as a mirror descent step, the following claim is used. It is the same as Claim 3.7
in [2] applied to different vectors. It is fairly straightforward to verity, and we include the proof in Appendix A.
Claim 3.7. For all t = 0, . . . , w − 1, we have
(t)
(t)
xk+1 = argmin{Vxk (z) + hz − xk , αξk i}.
z≥0
Once we see the update step is indeed a mirror descent step, we can derive the following result from the
textbook mirror descent analysis (or, e.g., Lemma 3.3 in [2]).
Lemma 3.8. For all t = 0, . . . , w − 1, we have for any u ≥ 0
(t)
hαξk , xk − ui ≤ α2 OPT +Vxk (u) − Vx(t) (u).
k+1
Proof. The lemma follows from the following chain of equalities and inequalities.
(t)
(t)
(t)
(t)
(t)
(t)
(t)
(t)
(t)
hαξk , xk − ui = hαξk , xk − xk+1 i + hαξk , xk+1 − ui
(t)
(t)
≤ hαξk , xk − xk+1 i + h−∇Vxk (xk+1 ), xk+1 − ui
(t)
≤ hαξk , xk − xk+1 i + Vxk (u) − Vx(t) (u) − Vxk (xk+1 )
k+1
(t)
≤
X
(t)
αξk [i](xk [i]
−
(t)
xk+1 [i])
−
i∈[n]
≤
(xk [i] − xk+1 [i])2
(t)
2 max{xk [i], xk+1 [i]}
!
+ Vxk (u) − Vx(t) (u)
k+1
X (αξk(t) [i])2 max{xk [i], x(t)
k+1 [i]}
+ Vxk (u) − Vx(t) (u)
k+1
2
i∈[n]
2 2~ T
α 1 xk + Vxk (u) − Vx(t) (u)
k+1
3
≤ α2 OPT +Vxk (u) − Vx(t) (u).
≤
k+1
(t)
The first equality follows by adding and subtracting xk+1 . The first inequality is due to the the minimality of
(t)
xk+1 , which gives
(t)
(t)
(t)
h∇Vxk (xk+1 ) + αξk , u − xk+1 i ≥ 0
∀u ≥ 0,
the second inequality is due to the standard three point property of Bregman divergence, that is ∀x, y ≥ 0
h−∇Vx (y), y − ui = Vx (u) − Vy (u) − Vx (y),
the third inequality is from (13), the fourth inequality follows from 2ab − a2 ≤ b2 , the next inequality is due to
(t)
(t)
xk+1 [i] ≤ xk [i](1 + ǫ), and ξk [i] ≤ 1. The last inequality is by Claim 3.6, ~1T xk ≤ (1 + ǫ) OPT.
11
3.4 Coupling of Gradient and Mirror Descent
In this section we show convergence using the results we derived by analyzing the update step as both a gradient
descent step and a mirror descent step.
Recall we break the gradients into small, medium and large components. The proof follows a similar approach as Lemma 3.4 of [2], where we bound the three components respectively,
and telescope along all iterations. Furthermore, we divide the medium and large components into w = log( 1ǫ ) groups, as follows:
∇fµ (xk ) = ζk + ξk + ηk
and
(t)
(t)
ξk = wEt [ξk ] and ηk = wEt [ηk ].
We bound the gap to optimality at iteration k, as follows:
α(fµ (xk ) − fµ (u∗ )) ≤hα∇fµ (xk ), xk − u∗ i
=αhζk , xk − u∗ i + αhξk , xk − u∗ i + αhηk , xk − u∗ i
(t)
(t)
=αhζk , xk − u∗ i + wEt [hαξk , xk − u∗ i] + wEt [hαηk , xk − u∗ i].
The first line is due to convexity. The next two lines just break and regroup the terms. Now we upperbound each
of the three terms
Lemma 3.9. We have the following.
1. hζk , xk − u∗ i ≤ 3ǫ OPT.
(t)
2. ∀t, hαξk , xk − u∗ i ≤ α2 OPT +Vxk (u∗ ) − Vx(t) (u∗ ).
k+1
(t)
hαηk , xk
3. ∀t,
− u∗ i ≤ 4(fµ (xk ) −
(t)
fµ (xk+1 )).
Proof. We establish each result in turn.
1. We know |ζk [i]| ≤ ǫ for all i, ~1T xk ≤ (1 + ǫ) OPT from Claim 3.6, and ~1T u∗ ≤ OPT, thus
hζk , xk − u∗ i ≤ ǫ(~1T xk + ~1T u∗ ) ≤ 3ǫ OPT
2. This is just Lemma 3.8 applied to u = u∗ .
3. This is just Lemma 3.5 applied to u = u∗ .
Given this, we have
(t)
(t)
α(fµ (xk ) − fµ (u∗ )) ≤αhζk , xk − u∗ i + wEt [hαξk , xk − u∗ i] + wEt [hαηk , xk − u∗ i]
(t)
≤3αǫ OPT +wEt [α2 OPT +Vxk (u∗ ) − Vx(t) (u∗ )] + 4wfµ (xk ) − 4wEt [fµ (xk+1 )].
k+1
The above inequality holds for xk following any sequence of random choices (i.e., t’s in the first k −1 iterations).
Let Ik denote the random choices over the first k iterations, and we take expectation of the above inequality to get
α(EIk [fµ (xk )] − fµ (u∗ )) ≤3αǫ OPT +α2 w OPT +wEIk [Vxk (u∗ )] − wEIk+1 [Vx(t) (u∗ )]
k+1
+ 4wEIk [fµ (xk )] −
(t)
4wEIk+1 [fµ (xk+1 )].
(14)
Telescoping (14) for k = 0, . . . , T − 1, we get
α
T
−1
X
(t)
(EIk [fµ (xk )] − fµ (u∗ )) ≤3T αǫ OPT +wT α2 OPT +wVx0 (u∗ ) + 4wfµ (x0 ) − 4wEIT [fµ (xT )]
k=0
(t)
≤3T αǫ OPT +wT α2 OPT +2w log 2n OPT −4wEIT [fµ (xT )],
where the last inequality is due to fµ (x0 ) < 0, and
12
(15)
Claim 3.10. Vx0 (u∗ ) ≤ 2 log 2n OPT
Proof.
Vx0 (u∗ ) =
X
u∗ [i] log
i
X
u∗ [i]
u∗ [i]
u∗ [i] log
+ x0 [i] − u∗ [i] ≤
+ x0 [i]
x0 [i]
x0 [i]
i
X
1/kA:i k∞
1 − ǫ/2
u∗ [i] log
≤
+
(1
−
ǫ/2)/(nkA
k
)
nkA
:i ∞
:i k∞
i
≤ ~1T u∗ log(2n) + 1 ≤ 2 log(2n) OPT
where we have used u∗i ≤
OPT ≥ 1.
1
kA:i k∞
in the second line, since Au∗ ≤ ~1. The third line is due to ~1T u∗ ≤ OPT, and
(t)
(t)
We prove that EIT [fµ (xT )] ≤ −(1 − 5ǫ) OPT (i.e. Theorem 2.1) by contradiction. If EIT [fµ (xT )] >
(t)
−(1 − 5ǫ) OPT, we have −4wEIT [fµ (xT )] ≤ 4w OPT. If we divide both sides of (15) by αT , then we have
T −1
1 X
α(EIk [fµ (xk )] − fµ (u∗ ))
Tα
k=0
1
(t)
(3T αǫ OPT +wT α2 OPT +2w log 2n OPT −4wEIT [fµ (xT )])
Tα
4w
2w log 2n
OPT +
OPT .
≤3ǫ OPT +wα OPT +
Tα
Tα
≤
Recall α = µ/20 =
and
4w
Tα
ǫ
20 log
mn
ǫ
, we have wα ≤ ǫ/20. By our choice of T =
10w log 2n
,
αǫ
we have
2w log 2n
Tα
< ǫ,
< ǫ. Thus
T −1
1 X
(EIk [fµ (xk )] − fµ (u∗ )) ≤ 4ǫ OPT
T
k=0
From part (1) of Lemma 3.3, we know fµ (u∗ ) ≤ −(1 − ǫ) OPT, which suggests there exists a xk , such that
(t)
EIT [fµ (xk )] ≤ −1(1 − 5ǫ) OPT. This gives a contradiction of EIT [fµ (xT )] > −(1 − 5ǫ) OPT by Claim 3.6,
as fµ (xk ) is non-increasing, so is EIk [fµ (xk )].
The running time guarantee in Theorem 2.1 comes directly from our choice of T , and that in each iteration
of Algorithm 1, the gradients can be computed in O(log N ) distributed iterations and O(N ) total work.
4 Analysis of Algorithm for Covering LPs: Proof of Theorem 2.2
In this section, we will provide a proof of Theorem 2.2. To do so, first recall some properties from the analysis
of the packing algorithm:
1. ∀u ≥ 0,
1
T E[
P
k
fµ (xk )] − fµ (u) ≤
≤
PT −1
1
k=0 h∇fµ (xk ), xk − ui]
T E[
w
4w
(f
(x
µ
0 ) − E[fµ (xT )]) + αT
αT
Vx0 (u) + 2ǫ OPT +ǫ~1T u.
(16)
This is simply telescoping (14) for a general u ≥ 0 instead of u∗ . Notice Lemma 3.9(1) for general u ≥ 0
gives hζk , xk − ui ≤ ǫ~1T u + ǫ~1T xk ≤ 2ǫ OPT +~1T u.
2. Axk ≤ (1 + ǫ)~1, ~1T xk ≤ (1 + ǫ) OPT and fµ (xk ) ≥ −(1 + ǫ) OPT hold for all k and any outcome of
1−ǫ/2
random choices. This follows from Claim 3.6, and Lemma 3.3(2). Also x0 [i] = nkA
for each i ∈ [n],
:i k∞
and fµ (x0 ) ≤ 0.
We start with the following lemma which states that ~1T ȳ is close to OPT on expectation.
Lemma 4.1. For any T ≥
6w
αǫ ,
we have E[~1T ȳ] ≤ (1 + 5ǫ) OPT
13
The proof of this lemma follows directly from Lemma D.1 of [2], only with the additional w = log( 1ǫ ) due
to our dynamic grouping, and the expectation. The expectation holds since all the inequalities used in the proof
hold universally (i.e., in any outcome of the random choices). We omit the detailed proof here, and encourage
interested readers to look at [2].
Now we look at the i-th constraint of the covering LP, which corresponds to the variable x[i] in the dual
(i)
packing instance. Let Zk be the indicator random variable of whether x[i] is in the group being updated in
iteration k of Algorithm 1, and let
Si = w
T
−1
X
(i)
Zk (min{∇i fµ (xk ), 1} + ǫ).
k=0
We can obtain a lower bound on the random variable Si as follows:
Lemma 4.2.
Si ≥ −
w log 2n
α
∀i.
(17)
PT −1 (t)
Proof. Using the notations of Algorithm 1, we have, for all i, Swi ≥ k=0 ξk [i]. We know the cumulative
1+ǫ
update on variable x[i] must be bounded, due to Claim 3.6, xk [i] ≤ kA:i k∞ for all k, and in particular
T −1
X (t)
Si
Si
1+ǫ
1 − ǫ/2
≥ xT [i] = x0 [i] exp(−α
ξk [i]) ≥ x0 [i] exp(−α ) =
exp(−α ).
kA:i k∞
w
nkA:i k∞
w
k=0
The bound in (17) follows.
Notice that the slackness of the i-th covering constraint with the solution ȳ is
(AT ȳ)i − 1 + ǫ =
T −1
1 X T −−−→
(A p(xk ))i − 1 + ǫ
T
k=0
=
≥
1
T
T
−1
X
∇i fµ (xk ) + ǫ
k=0
T −1
1 X
min{∇i fµ (xk ), 1} + ǫ
T
k=0
1
= Ui ,
T
PT −1
with the definition of the random variable Ui = k=0 min{∇i fµ (xk ), 1} + ǫ. If the i-th variable is updated in
(i)
2n
all iterations, i.e., Zk = 1 for all k, then we have Ui = Si in that case, and when T ≥ 6w log
, we have that
αǫ
def
(AT ȳ)i − 1 + ǫ ≥
w log 2n
1
Si ≥ −
≥ −ǫ.
T
αT
Thus, we know that (AT ȳ)i ≥ 1 − 2ǫ, for all i, which means all covering constraints are approximately feasible.
However, we don’t always update variable i in all iterations of Algorithm 1, and so we need to bound the
difference Si − Ui . We do so with the following lemma.
Lemma 4.3. For any T ≥
2w 2 log
ǫ2
n
ǫ
, we have Pr[Si − Ui ≥ ǫT ] ≤
ǫ
n.
Proof. The randomness of Ui and Si comes from the random choice of which group to update in each iteration
of Algorithm 1. Let
X (i)
X
(i)
Dk = w
Zk′ (min{∇i fµ (xk ), 1} + ǫ) −
min{∇i fµ (xk ), 1} + ǫ.
k′ ≤k
k′ ≤k
14
(i)
Let Gk be the random choice (i.e., the group to update) made at k-th iteration of Algorithm 1. Since Zk is an
(i)
indicator random variables with probability of w1 being 1, and it is independent from G0 , . . . , Gk−1 , Dk is a
martingale with respect to Gk , as
(i)
(i)
E[Dk |G0 , . . . , Gk−1 ] = Dk−1 +
and so we have
1
(i)
w(min{∇i fµ (xk ), 1} + ǫ) − (min{∇i fµ (xk ), 1} + ǫ) = Dk−1 ,
w
(i)
D0 = E[Si − Ui ] = 0,
(i)
(i)
D T = Si − U i .
(i)
Furthermore, |Dk − Dk+1 | ≤ w for all k, so we can apply Azuma’s inequality, and get
Pr[Si − Ui ≥ ǫT ] ≤ exp(
ǫ
ǫ2 T 2
)≤ ,
2T w2
n
from which the lemma follows.
The above lemma, Lemma 4.3, shows that with high probability, ȳ satisfies the i-th covering constraint up
to −3ǫ. In the rare case it doesn’t, we use Algorithm 2 to fix it, and get ȳ ′ . We show on expectation this step
doesn’t add too much to the total cost.
Lemma 4.4. E[~1T y¯′ ] ≤ (1 + 6ǫ) OPT.
Proof. When Si − Ui ≤ ǫT , we have
(AT ȳ)i − 1 + ǫ ≥
1
w log 2n
(Si − ǫT ) ≥ −
− ǫ ≥ −2ǫ,
T
αT
and so we don’t need to fix the i-th constraint. When that is not the case, since (AT ȳ)i ≥ 0, and kA:i k∞ ≥ 1, we
need to add at most 1 to some variable ȳj′ to fix the i-th covering constraint. For all the n covering constraints,
we add on expectation at most n nǫ ≤ ǫ ≤ ǫ OPT to ȳ to get ȳ ′ . Together with Lemma 4.1, we have E[~1T y¯′ ] ≤
(1 + 6ǫ) OPT.
We complete the proof of Theorem 2.2 by noticing (AT ȳ ′ )i ≥ 1 − 3ǫ, for all i. Thus, the output of Algoȳ ′
rithm 2, 1−3ǫ
, satisfies the properties stated in Theorem 2.2.
Acknowledgments. DW was supported by ARO Grant W911NF-12-1-0541 and NSF Grant CCF-1528174,
SR was funded by NSF Grant CCF-1528174, and MWM acknowledges the support of the NSF, AFOSR, and
DARPA.
Appendix A
Missing Proofs
The following proofs can be found in [2], and we include them here for completeness.
Lemma 3.1. OPT ∈ [1, n]
Proof. By the assumption mini∈[n] kA:i k∞ = 1, we know at least one variable has all coefficients at most 1, so
we can just set that variable to 1, which gives OPT ≥ 1. On the other hand, since each variable has a coefficient
of 1 in some constraint, no variable can be larger than 1, thus OPT ≤ n.
Lemma 3.3. Setting the smoothing parameter µ =
ǫ
4 log(nm/ǫ) ,
we have
1. fµ (u∗ ) ≤ −(1 − ǫ) OPT.
2. fµ (x) ≥ −(1 + ǫ) OPT for every x ≥ 0.
3. Letting x0 ≥ 0 be such that x0 [i] =
1−ǫ/2
nkA:i k∞
for each i ∈ [n], we have fµ (x0 ) ≤ − 1−ǫ
n .
4. For any x ≥ 0 satisfying fµ (x) ≤ 0, we must have Ax ≤ (1 + ǫ)~1, and thus ~1T x ≤ (1 + ǫ) OPT.
15
5. If x ≥ 0 satisfies fµ (x) ≤ −(1 − O(ǫ)) OPT, then
1
1+ǫ x
is a (1 − O(ǫ))-approximation to the packing LP.
6. The gradient of fµ (x) is
−−→
∇fµ (x) = −~1 + AT p(x) where
and ∇i fµ (x) = −1 +
P
j
1
def
pj (x) = exp( ((Ax)j − 1),
µ
Aji pj (x) ∈ [−1, ∞].
Proof. We establish each result in turn.
1. Since Ax∗ ≤ ~1, and u∗ = (1 − ǫ/2)x∗ , we have (Au∗ )j − 1 ≤ −ǫ/2 for all j. Then pj (u∗ ) ≤
Pm
ǫ 2
ǫ 2
) , and fµ (u∗ ) = −~1T u∗ + µ j=1 pj (u∗ ) ≤ −(1 − ǫ/2) OPT +µm( mn
) ≤
exp(− µ1 2ǫ ) = ( mn
−(1 − ǫ) OPT.
2. By contradiction, suppose fµ (x) < −(1 + ǫ) OPT, since fµ (x) > −~1T x, we must have ~1T x > (1 +
ǫ) OPT. Suppose ~1T x = (1 + v) OPT for some v > ǫ. There must exits a j, such that (Ax)j > 1 + v.
4 v/ǫ
Then we have pj (x) > exp(v/µ) = (( mn
, which implies
ǫ ) )
fµ (x) ≥ −(1 + v) OPT +µpj (x) ≥
mn 4 v/ǫ
ǫ
((
) ) − (1 + v) OPT > 0,
4 log(mn/ǫ)
ǫ
since OPT ≤ n, and v > ǫ. This gives a contradiction.
3. The x0 we use satisfies Ax0 − ~1 ≤ −ǫ/2 − ~1, thus
fµ (x0 ) = µ
X
pj (x0 ) − ~1T x0 ≤
j
1 − ǫ/2
1−ǫ
µm
−
≤−
.
(nm)2
n
n
4. By contradiction, suppose there is some j such that (Ax)j − 1 ≥ ǫ. Let v > ǫ be the smallest v such that
Ax ≤ (1+v) OPT, and denote j the constraint that has (Ax)j −1 = v. We must have ~1T x ≤ (1+v) OPT
by definition of OPT. Then
fµ (x) ≥ µpj (x) − (1 + v) OPT ≥
mn 4 v/ǫ
ǫ
((
) ) − (1 + v) OPT > 0,
4 log(mn/ǫ)
ǫ
which gives a contradiction.
x
5. By the above part, fµ (x) ≤ −(1 − O(ǫ)) OPT ≤ 0 suggests 1+ǫ
is feasible. Furthermore, −~1T x <
x
T
T
fµ (x) ≤ −(1 − O(ǫ)) OPT gives ~1 x ≥ (1 − O(ǫ)) OPT, thus ~1 1+ǫ
≥ (1 − O(ǫ)) OPT is approximately optimal.
6. This is by straightforward computation.
Claim 3.7. For all t = 0, . . . , w − 1, we have
(t)
(t)
xk+1 = argmin{Vxk (z) + hz − xk , αξk i}.
z≥0
Proof. Since the function Vx(k) (z), the dot product and the constraint z ≥ 0 are all coordinate-wise separable,
we look at each coordinate independently. Thus we only need to check
(t)
xk+1 [i] = argmin{(z[i] log
z[i]≥0
z[i]
(t)
+ xk [i] − z[i]) + αξk [i](z[i] − xk [i])}.
xk [i]
This univariate function being optimized is convex and has a unique minimizer. We find it by taking the derivative
to get
z[i]
(t)
+ αξk [i] = 0,
log
xk [i]
(t)
def
(t)
which gives xk+1 [i] = z[i] = xk [i] exp(−αξk [i]).
16
References
[1] Zeyuan Allen-Zhu and Lorenzo Orecchia. Nearly-linear time positive LP solver with faster convergence
rate. In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC ’15,
pages 229–236, 2015. Newer version available at http://arxiv.org/abs/1411.1124.
[2] Zeyuan Allen-Zhu and Lorenzo Orecchia. Using optimization to break the epsilon barrier: A faster and
simpler width-independent algorithm for solving positive linear programs in parallel. In Proceedings of the
Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’15, pages 1439–1456, 2015.
[3] Pierluigi Amodio and Francesca Mazzia. A parallel Gauss-Seidel method for block tridiagonal linear
systems. SIAM J. Sci. Comput., 16(6):1451–1461, November 1995.
[4] Sanjeev Arora, Elad Hazan, and Satyen Kale. The multiplicative weights update method: a meta-algorithm
and applications. Theory of Computing, 8(6):121–164, 2012.
[5] Dimitri P. Bertsekas and John N. Tsitsiklis. Some aspects of parallel and distributed iterative algorithms—
A survey. Automatica, 27(1):3–21, 1991.
[6] Joseph K. Bradley, Aapo Kyrola, Danny Bickson, and Carlos Guestrin. Parallel coordinate descent for l1regularized loss minimization. In Proceedings of the 28th International Conference on Machine Learning,
ICML 2011, Bellevue, Washington, USA, June 28 - July 2, 2011, pages 321–328, 2011.
[7] Olivier Fercoq and Peter Richtárik.
abs/1312.5799, 2013.
Accelerated, parallel and proximal coordinate descent.
CoRR,
[8] Lisa Fleischer. A fast approximation scheme for fractional covering problems with variable upper bounds.
In Proceedings of the Fifteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2004, New
Orleans, Louisiana, USA, January 11-14, 2004, pages 1001–1010, 2004.
[9] Kimon Fountoulakis and Rachael Tappenden. Robust block coordinate descent. CoRR, abs/1407.7573,
2014.
[10] Christos Koufogiannakis and Neal E. Young. A nearly linear-time PTAS for explicit fractional packing and
covering linear programs. Algorithmica, 70(4):648–674, 2014.
[11] Yin Tat Lee and Aaron Sidford. Efficient accelerated coordinate descent methods and faster algorithms
for solving linear systems. In 54th Annual IEEE Symposium on Foundations of Computer Science, FOCS
2013, 26-29 October, 2013, Berkeley, CA, USA, pages 147–156, 2013.
[12] Ji Liu, Steve J. Wright, Christopher Ré, Victor Bittorf, and Srikrishna Sridhar. An asynchronous parallel
stochastic coordinate descent algorithm. In Proceedings of the 31th International Conference on Machine
Learning, ICML 2014, Beijing, China, 21-26 June 2014, pages 469–477, 2014.
[13] Michael Luby and Noam Nisan. A parallel approximation algorithm for positive linear programming. In
Proceedings of the Twenty-Fifth Annual ACM Symposium on Theory of Computing, May 16-18, 1993, San
Diego, CA, USA, pages 448–457, 1993.
[14] Arkadi Nemirovski. Prox-method with rate of convergence O(1/t) for variational inequalities with Lipschitz continuous monotone operators and smooth convex-concave saddle point problems. SIAM Journal
on Optimization, 15(1):229–251, 2004.
[15] Yurii Nesterov. Smooth minimization of non-smooth functions. Math. Program., 103(1):127–152, 2005.
[16] Yurii Nesterov. Efficiency of coordinate descent methods on huge-scale optimization problems. SIAM
Journal on Optimization, 22(2):341–362, 2012.
[17] Serge A. Plotkin, David B. Shmoys, and Éva Tardos. Fast approximation algorithms for fractional packing
and covering problems. In 32nd Annual Symposium on Foundations of Computer Science, San Juan, Puerto
Rico, 1-4 October 1991, pages 495–504, 1991.
17
[18] James Renegar. Efficient first-order methods for linear programming and semidefinite programming.
CoRR, abs/1409.5832, 2014.
[19] Peter Richtárik and Martin Takác. Parallel coordinate descent methods for big data optimization. CoRR,
abs/1212.0873, 2012.
[20] Peter Richtárik and Martin Takác. Iteration complexity of randomized block-coordinate descent methods
for minimizing a composite function. Math. Program., 144(1-2):1–38, 2014.
[21] Yousef Saad. Iterative methods for sparse linear systems. SIAM, 2003.
[22] Paul Tseng and Sangwoon Yun. A coordinate gradient descent method for nonsmooth separable minimization. Math. Program., 117(1-2):387–423, 2009.
[23] Di Wang, Satish Rao, and Michael W. Mahoney. Unified acceleration method for packing and covering
problems via diameter reduction. CoRR, abs/1508.02439, 2015.
[24] Stephen J. Wright. Coordinate descent algorithms. Math. Program., 151(1):3–34, 2015.
[25] Neal E. Young. Sequential and parallel algorithms for mixed packing and covering. In 42nd Annual
Symposium on Foundations of Computer Science, FOCS 2001, 14-17 October 2001, Las Vegas, Nevada,
USA, pages 538–546, 2001.
[26] Neal E. Young. Nearly linear-time approximation schemes for mixed packing/covering and facility-location
linear programs. CoRR, abs/1407.3015, 2014.
[27] Zeyuan Allen Zhu, Yin Tat Lee, and Lorenzo Orecchia. Using optimization to obtain a width-independent,
parallel, simpler, and faster positive SDP solver. CoRR, abs/1507.02259, 2015.
[28] Zeyuan Allen Zhu and Lorenzo Orecchia. Linear coupling: An ultimate unification of gradient and mirror
descent. CoRR, abs/1407.1537, 2014.
18
| 8cs.DS
|
arXiv:1711.01032v2 [cs.DM] 8 Nov 2017
A Simply Exponential Upper Bound
on the Maximum Number of Stable Matchings
Anna R. Karlin
University of Washington
[email protected]
Shayan Oveis Gharan
University of Washington
[email protected]
Robbie Weber
University of Washington
[email protected]
November 9, 2017
Abstract
Stable matching is a classical combinatorial problem that has been the subject of intense
theoretical and empirical study since its introduction in 1962 in a seminal paper by Gale and
Shapley [GS62]. In this paper, we provide a new upper bound on f (n), the maximum number of
stable matchings that a stable matching instance with n men and n women can have. It has been
a long-standing open problem to understand the asymptotic behavior of f (n) as n → ∞, first
posed by Donald Knuth in the 1970s [Knu76]. Until now the best lower bound was approximately
2.28n , and the best upper bound was 2n log n−O(n) . In this paper, we show that for all n,
f (n) ≤ cn for some universal constant c. This matches the lower bound up to the base of the
exponent. Our proof is based on a reduction to counting the number of downsets of a family of
posets that we call “mixing”. The latter might be of independent interest.
1
1
Introduction
Stable matching is a classical combinatorial problem that has been the subject of intense theoretical
and empirical study since its introduction in a seminal paper by Gale and Shapley in 1962 [GS62].
Variants of the algorithm introduced in [GS62] are widely used in practice, e.g. to match medical residents to hospitals. Stable matching is even the focus of the 2012 Nobel Prize in Economics [Ram12].
A stable matching instance with n men and n women is defined by a set of preference lists, one
per person. Person i’s preference list gives a ranking over members of the opposite sex. The Stable
Matching Problem is to find a matching (i.e., a bijection) between the men and the women that is
stable, that is, has no blocking pairs. A man m and a woman w form a blocking pair in a matching if
they are not matched to each other, but both prefer the other to their partner in the matching. Gale
and Shapley [GS62] showed that a stable matching always exists and gave an efficient algorithm to
find one.1 Since at least one stable matching always exists, a natural question is to determine the
maximum number of stable matchings an instance of a given size can have. This problem was posed
in the 1970s in a monograph by Knuth [Knu76], and was the first of Gusfield and Irving’s twelve
open problems in their 1989 textbook [GI89]. We denote the maximum number of stable matchings
an instance with n men and n women can have by f (n).
Progress on determining the asymptotics of f (n) has been somewhat slow. The best lower
bound is approximately 2.28n , and the best upper bound prior to this paper was 2n log n−O(n) . See
the related work section for a detailed history.
In this paper, we present an improved upper bound.
Theorem 1.1. There is a universal constant c such that f (n), the number of stable matchings in
an instance with n men and n women, is at most cn .
To prove this theorem, we use a result due to Irving and Leather [IL86] that shows that there is a
bijection between the stable matchings of an instance I and the downsets2 of a particular partiallyordered set (poset) associated with I known as the rotation poset. We show that the rotation poset
associated with a stable matching instance has a particular property that we call n-mixing, and that
any poset with this property has at most cn downsets. All the steps in our proof are elementary.
The bound extends trivially to stable roommates instances. In the stable roommates problem,
a set of n agents rank the other n − 1 agents in the set. The agents are paired off into roommate
pairs, which are stable if no two agents would like to leave their partners and be matched to each
other. A construction of Dean and Munshi [DM10], demonstrates that a stable roommates instance
with n agents can be converted into a stable matching instance with n men and n women, such
that the stable roommate assignments correspond to a subset of the stable matchings in the new
instance. Using this construction, we can apply our upper bound to Stable Roommates.
Theorem 1.2. There is a universal constant c, such that the number of stable assignments in a
stable roommate instance with n agents is at most cn .
1.1
Related Work
Lower bounds: It is trivial to provide instances with 2n/2 stable matchings by combining disjoint
instances of size 2. Irving and Leather constructed a family of instances [IL86] which has since been
1
2
Stable matching algorithms were actually developed and used as early as 1951 to match interns to hospitals [Sta53].
See section 2 for definitions of all the relevant terminology.
2
shown by Knuth3 to contain at least Ω(2.28n ) matchings. Irving and Leather’s family only has
instances for n which is a power of 2. Benjamin, Converse, and Krieger also provided a lower bound
√
on f (n) by creating a family of instances with Ω(2n n) matchings [BCK95]. While this is fewer
matchings than the instances in [IL86], Benjamin et al.’s family has instances for every even n, not
just powers of 2. In 2002, Thurber extended Irving and Leather’s lower bound to all values of n.
For n powers of 2, Thurber’s construction exactly coincides with Irving and Leather’s. For all other
n, the construction produces a lower bound of 2.28n /clog n for some constant c [Thu02]. To date,
this lower bound of Ω(2.28n ) is the best known. We refer the reader to Manlove’s textbook for a
more thorough description of the history of these lower bounds [Man13].
Upper bounds: Trivially, there are at most n! stable matchings (as there are at most n! bijections
between the men and women). The first progress on upper bounds that we are aware of was made
by Stathoupolos in his 2011 Master’s thesis [Sta11], where he proves that the number of stable
matchings is at most O(n!/cn ) for some constant c. A more recent paper of Drgas-Burchardt and
Świtalski shows a weaker upper bound of approximately 34 n! [DBŚ13]. All previous upper bounds
have the form 2n log n−O(n) .
Restricted preferences: The number of possible stable matchings has also been studied under
various models restricting or randomizing the allowable preference lists. If all preference lists are
equally likely and selected independently for each agent, Pittel shows that the expected number
of stable matchings is O(n log n) [Pit89]. Applying Markov’s Inequality shows that the number of
stable matchings is polynomial in n with probability 1 − o(1). Therefore, the lower bound instances
described above are a vanishingly small fraction of all instances. Work of Hoffman, Levy, and
Mossel (described in Levy’s PhD thesis [Lev17]) shows that under a Mallows model [Mal57], where
preference lists are selected with probability proportional to the number of inversions in the list,
the number of stable matchings is C n with high probability (where the constant C depends on the
exact parameters of the model).
The number of attainable partners4 a person can have has also been the subject of much research.
Knuth, Motwani, and Pittel show that the number of attainable partners is O(log n) with high
probability if the lists are uniformly random [KMP90]. Immorlica and Mahdian show that if agents
on one side of the instance have random preference lists of length k (and consider all other agents
unacceptable) the expected number of agents with more than one attainable partner depends only on
k (and not on n) [IM05]. Ashlagi, Kanoria, and Leshno show that if the number of men and women
is unbalanced, with uniformly random lists, the fraction of agents with more than one attainable
partner is 1 − o(1) with high probability [AKL17].
Counting: A natural computational problem is to count the number of stable matchings in a
given instance as efficiently as possible. Irving and Leather show that finding the exact number
of matchings is #P -complete [IL86], so finding an approximate count is a more realistic goal.
Bhatnagar, Greenberg, and Randall consider instances where preference lists come from restricted
models [BGR08]; for example, those in which the preference lists reflect linear combinations of k
“attributes” of the other set, or where every agent appears in a “range” of k consecutive positions in
3
Personal communication, as described in [GI89].
Woman w is an attainable partner of man m if there is a stable matching in which they are matched to each
other.
4
3
every preference list. In both of these cases, they show that a natural Markov Chain Monte Carlo
approach does not produce a good approximation (as the chain does not mix efficiently). As part of
their proof, they show the number of stable matchings can still be as large as cn for some constant
c < 2.28, even in these restricted cases.
A formal hardness result was later shown by Dyer, Goldberg, Greenhill, and Jerrum [DGGJ04].
They show approximately counting the number of stable matchings is equivalent to approximately
counting for a class of problems, canonically represented by #BIS.5 This hardness result was
strengthened by Chebolu, Goldberg, and Martin [CGM12] to hold even if the instances come from
some of the restricted classes of [BGR08].
The heart of all of these results is the rotation poset (originally developed in [IL86]), which we
use and describe in section 4.
Stable matching in general: See the books by Roth and Sotomayor [RS92], Gusfield and Irving [GI89], Manlove [Man13], and Knuth [Knu97] for more about the topic of stable matching. For
many examples of stable matching in the real world, see [Rot15].
2
Preliminaries and main technical theorem
In this section, we review standard terminology regarding partially ordered sets, describe the key
property of a poset we will use, and state our main technical theorem (Theorem 2.5).
Definition 2.1 (Poset). A partially ordered set (or poset) (V, ≺), is defined by a set V and a binary
relation, ≺, on V satisfying:
• Antisymmetry: For all distinct u, v ∈ V if u ≺ v then v 6≺ u, and
• Transitivity: for all u, v, w ∈ V if u ≺ v and v ≺ w then u ≺ w.
Two elements u, v ∈ V are comparable if u ≺ v or v ≺ u. They are incomparable otherwise.
If u ≺ v, we say u is dominated by v and v dominates u.
Definition 2.2 (Chain). A set S of elements is called a chain if each pair of elements in S is
comparable. In other words, for ` > 0, a chain of length ` is a sequence of elements v1 ≺ v2 ≺ v3 ≺
· · · ≺ v` .
A set of elements is called an antichain if they are pairwise incomparable.
Definition 2.3 (Downset). A downset of a partial order is an antichain and all elements dominated
by some element of that antichain.
Observe that a downset is closed under ≺. That is, for any downset S, if v ∈ S and u ≺ v then
u ∈ S.
The following is the key property of the posets associated with stable matching instances that
we will use in the proof.
5
More specifically, they show an FPRAS for the number of stable matchings exists if and only if one exists for
#BIS, approximately counting the number of independent sets in a bipartite graph. Goldberg and Jerrum conjecture
that no such FPRAS exists [GJ12]. See [CGM12] for formal definitions.
4
n
n
Figure 1: A 2n-mixing poset with respect to chains defined by the red and blue paths. A path
in the graph from v to u indicates that u ≺ v. That is, the poset
√ is the transitive closure of the
arrows shown. For any set U of k elements, there are at least 2 k chains that contain one of these
elements.
Definition 2.4 (n-mixing). A poset (V, ≺) is n-mixing if there exist n chains C1 , . . . , Cn (not
necessarily disjoint) such that
i) Every element of V belongs to at least one chain, i.e., ∪ni=1 Ci = V ,
p
ii) For any U ⊆ V , there are at least 2 |U | chains each containing an element of U .
Observe that if a poset is formed by n disjoint chains, each of length `, it has about `n downsets,
and ` could be arbitrarily bigger than n. But such a poset is not mixing, since taking U to be the
set of elements on one of the chains violates property ii) of mixing. For an example of a mixing
poset, see Figure 1. We can now state our main technical theorem.
Theorem 2.5. There is a universal constant c, such that if a poset is n-mixing, then it has at most
cn downsets.
Note that the n-mixing property immediately implies that the poset has at most n2 elements;
just let U = V in the above definition. A poset with n2 elements covered by n chains can have at
most (n + 1)n downsets (this is achieved for n equal length chains). So, the main contribution of
the above theorem is to improve this trivial upper bound to cn for some constant c.
Theorem 2.5 is the main technical contribution of this work. The proof is contained in section 3.
To complete the proof of Theorem 1.1 we use the following theorem relating mixing posets to stable
matchings.
5
Theorem 2.6. For every stable matching instance I with a total of n men and women, there exists
an n-mixing poset (R, ≺), called the rotation poset, such that the number of downsets of the poset
is equal to the number of stable matchings of I.
Note that Theorem 2.5 and Theorem 2.6 immediately imply Theorem 1.1. We prove Theorem 2.6
in section 4, by combining existing observations about the rotation poset.
3
Proof of main technical theorem
In this section we prove Theorem 2.5. The proof proceeds by finding an element v of the poset
which dominates and is dominated by many elements. We then count downsets by considering
those downsets that contain v and those that do not. Since v dominates and is dominated by many
elements, the size of each remaining instance is significantly smaller, yielding the bound.
Formally, we say an element is α-critical if it dominates α elements and is dominated by α
elements. The key lemma is that there is always an Ω((|V |/n)3/2 )-critical element.
Lemma 3.1. Let (V, ≺) be an n-mixing poset with respect to chains C1 , . . . , Cn , and define d = |Vn | .
For some universal constants d0 > 1 and c0 > 0, there is an element v ∈ V such that v is (c0 d3/2 )critical as long as d ≥ d0 .
We prove Lemma 3.1 in subsection 3.1 via a counting argument.
In the rest of this section we prove Theorem 2.5 using Lemma 3.1. We bound the number of
downsets by induction on d.
Our base case is when d = d0 . In this case, the number of downsets is maximized when the
chains are all the same length, so we have an upper bound of (d0 + 1)n .
For larger d, first we identify a c0 d3/2 -critical element v. The number of downsets containing
v is the number of downsets in the poset remaining after we delete v and everything it dominates.
Similarly, the number of downsets not containing v is the number of downsets in the poset remaining
after we delete v and everything dominating v. In both cases, the resulting poset is still n-mixing,
so we can induct. We call such a step (choosing a critical element) an iteration.
It remains to bound the number of downsets that this process enumerates. By the n-mixing
property, there are at most n2 elements in the initial poset. We partition the iterations into phases,
2
2
where in phase i we reduce the size of the poset from n2i elements to 2ni+1 elements. By definition,
n
in phase i, d ≥ 2i+1
. So, we can bound the number of iterations required in phase i (call it ki ) by:
c0
n 3/2
n2
k
>
.
i
2i+1
2i+1
√
Rearranging, we see that it suffices to choose ki = 2(i+1)/2 n/c0 . We continue until d = d0 .
Summing across all phases, the number of choices to make is at most
√ log
n
n X (i+1)/2
n
2
<5 .
c0
c0
i=0
So, the algorithm enumerates at most (d0 +1)n downsets in the base case and it makes at most 5n/c0
choices during the inductive process. Thus, the number of downsets is at most (d0 + 1)n 25n/c0 = cn
as required.
6
3.1
Proof of main technical lemma
Finally, we prove Lemma 3.1, i.e. we show that any n-mixing poset (V, ≺) contains a c0 d3/2 -critical
element as long as d ≥ d0 . (Recall that d = |V |/n.)
We make use of the standard graph representation of the partial order: In this graph there is
a node for each element of the partial order, and a directed edge from v to u if u ≺ v. Of course,
this directed graph is acyclic. Henceforth, we refer only to this DAG rather than to the poset and
partition the nodes of the DAG into levels as follows: Level 1 nodes are those with no outgoing
edges, i.e., sinks of the DAG, and level i nodes are those whose longest path to a sink (i.e., a level
1 node) has exactly i nodes. Note that each level is an antichain.
Next, we create n disjoint subchains S1 , . . . , Sn . The subchain Si will be a subset of the nodes
in Ci . We perform the assignment of nodes to subchains by processing up the DAG level by level.
Initialize every subchain Si to be empty. For each node u, consider the set of indices I(u) = {j :
u ∈ Cj }. Assign u to the subchain for an index in I(u) which currently has the fewest nodes among
those chains, i.e. arg minj∈I(u) |Sj |, breaking ties arbitrarily. If u is the k th node assigned to a
subchain Si , then we say the height of u is k. By construction, the following properties hold:
(a) The Si ’s are chains, since Si ⊆ Ci .
(b) The Si ’s are disjoint.
(c) If u has height h, then u dominates at least h − 1 nodes in each of the subchains Sj such
that j ∈ I(u) (i.e., {j : u ∈ Cj }).
Claim 3.2. Let D be the set of nodes of height at least bd/2c. Each node u ∈ D dominates at least
c0 · d3/2 nodes, for d ≥ d0 , where c0 and d0 are universal constants.
Proof. Suppose that u ∈ Si ∩ D is at height ` ≥ bd/2c and let D(u)
p be the nodes in Si of height
d`/2e through `. By the mixing property, these nodes lie on at least 2 b`/2c chains C. Moreover, by
construction, on each subchain Sj ∈ C, at least b`/2c−1 nodes are dominatedpby some node in D(u)
and hence are dominated by u. Therefore, u dominates at least (b`/2c − 1) · 2 b`/2c + (b`/2c − 1) =
Ω(d3/2 ) nodes.
Since the number of nodes in the DAG is dn, we conclude:
Corollary 3.3. There is a set D of strictly more than |V |/2 nodes, that each dominate Ω(d3/2 )
nodes.
A symmetric argument in which subchains are built starting from the sources of the DAG shows
that there is a set D of strictly more than |V |/2 nodes that are dominated by Ω(d3/2 ) nodes.
Therefore, there is some node v in the intersection of D and D. This is the c0 d3/2 -critical node we
seek.
Remark 3.4. A crude analysis shows that c0 ≥ 1/8 when d0 > 25.
4
Rotations and the rotation poset
In this section we present a key theorem of Irving and Leather [IL86] and show how it can be used
to prove Theorem 2.6.
7
Theorem 2.6. For every stable matching instance I with a total of n men and women, there exists
an n-mixing poset (R, ≺), called the rotation poset, such that the number of downsets of the poset
is equal to the number of stable matchings of I.
We begin with the definitions needed to prove Theorem 2.6.
Definition 4.1 (Rotation). Let k ≥ 2. A rotation ρ is an ordered list of pairs
ρ = ((m0 , w0 ), (m1 , w1 ), . . . , (mk−1 , wk−1 ))
that are matched in some stable matching M with the property that for every i such that 0 ≤ i ≤ k−1,
woman wi+1 (where the subscript is taken mod k) is the highest ranked woman on mi ’s preference
list satisfying:
i) man mi prefers wi to wi+1 , and
ii) woman wi+1 prefers mi to mi+1 .
In this case, we say ρ is exposed in M .
We will sometimes abuse notation and think of a rotation as the set containing those pairs. Also,
we will need the following facts about rotations later.
Lemma 4.2 ([IL86, Lemma 4.7]). A pair (m, w) can appear in at most one rotation.
Lemma 4.3 ([GI89, Lemma 2.5.1]). If ρ is a rotation with consecutive pairs (mi , wi ), and (mi+1 , wi+1 ),
and w is a woman between wi and wi+1 in mi ’s preference list, then there is no stable matching
containing the pair (mi , w).
Definition 4.4 (Elimination of a Rotation). Let ρ = ((m0 , w0 ), . . . , (mk−1 , wk−1 )) be a rotation
exposed in stable matching M . The rotation ρ is eliminated from M by matching mi to w(i+1) mod k ,
for all 0 ≤ i ≤ k − 1, leaving all other pairs in M unchanged, i.e., matching M is replaced with
matching M 0 , where
M 0 := M \ρ ∪ {(m0 , w1 ), (m1 , w2 ), . . . , (mk−1 , w0 )}.
Note that when we eliminate a rotation from M , the resulting matching M 0 is stable.6
Irving and Leather studied the following process: Fix a stable matching instance I. Starting
at the man-optimal matching7 , choose a rotation in the current matching, and eliminate it. They
show that for any stable matching M , there is a set of rotations, R(M ), one can eliminate (starting
from the man-optimal matching) that will yield M .
Switching from M to M 0 makes all the women in ρ happier and all the men in ρ less happy. It is easy to check
that this switch cannot create a blocking pair inside the set ρ. The only other possibility for a blocking pair is a man
in ρ with a woman outside ρ. For (mi ∈ ρ, w 6∈ ρ) to become a blocking pair, mi would have to prefer w to wi+1 ,
but by the definition of rotation, wi+1 was the first woman on m’s list who would prefer to be matched to him, so he
cannot prefer w to wi+1 . See also [GI89, Lemma 2.5.2].
7
A fascinating fact about stable matching is that there is a matching, known as the man-optimal matching, in
which each man is matched with his favorite attainable partner. Recall that woman w is attainable for man m if
there is some stable matching in which they are matched.
6
8
However, there is a partial order on the set of rotations – some must be eliminated before
others.8 If ρ must be eliminated before ρ0 , we write ρ ≺ ρ0 . Let R be the set of rotations for a stable
matching instance, and let ≺ be that partial order on the rotations defined by elimination order.
We call this poset the rotation poset. See Figure 2 for an example of a stable matching instance
and the corresponding rotation poset.
For our purposes, the important result relating the rotation poset to stable matchings is the
following.
Theorem 4.5 ([IL86, Theorem 4.1]). For any stable matching instance, there is a one-to-one
correspondence between downsets of the rotation poset and the set of stable matchings. In other
words, the number of stable matchings is exactly equal to the number of downsets of (R, ≺).
Indeed, the downset corresponding to M is exactly the set R(M ) discussed above.
Thus, to prove Theorem 2.6, it remains to show that the rotation poset associated to any
stable matching instance with a total of n men and women is n-mixing. First we construct the
chains C1 , . . . , Cn . We are going to have one chain for each agent (man or woman), where the
corresponding chain contains all the rotations that include that agent. Call these sets C1 , . . . , Cn .
To prove that C1 , . . . , Cn is n-mixing, we first need to show that each Ci is indeed a chain, i.e.,
every pair of rotations where a specific agent appears are comparable and second we need to show
that C1 , . . . , Cn satisfy property (ii) of Definition 2.4.
Claim 4.6. If two rotations share an agent, then they are comparable.9
Proof. First, suppose that the shared agent is a man. If it is a woman, we can just switch the
designations of “men” and “women” and use the same proof on the “reversed” version of the rotation
graph. Let ρ1 , ρ2 be rotations sharing an agent m, and let m be matched to w1 in ρ1 and w2 in ρ2 ,
where m prefers w1 to w2 .
For the sake of contradiction assume that ρ1 and ρ2 are incomparable. We show that there
exists a rotation ρ which causes m to skip over w1 . This would contradict Lemma 4.3 as it implies
that (m, w1 ) belongs to no stable matching.
Suppose we start from the man optimal stable matching and eliminate all rotations dominated by
ρ2 and let M be the resulting stable matching. By the correspondence in Theorem 4.5, (m, w2 ) ∈ M .
Since m prefers w1 to w2 , there must be a rotation ρ that we eliminated which caused m to be
matched to someone worse than w1 for the first time. Since ρ2 and ρ1 are incomparable, ρ1 6= ρ.
Therefore, by Lemma 4.2 (m, w1 ) ∈
/ ρ; so, ρ caused m to skip over w1 . This is a contradiction.
√
Claim 4.7. Every set of k rotations contains at least 2 k agents.
√
Proof. We argue by contrapositive. Suppose we have a set of rotations involving fewer than 2 k
agents. Every rotation contains a (man,
woman) pair, who (by Lemma 4.2) have not appeared
√
together before. With fewer than 2 k agents, there are strictly less than k (man, woman) pairs
which can appear, and thus fewer than k rotations in the set.
8
For example, a rotation containing a pair (m, w) is not exposed (and thus cannot be eliminated) until that pair
is matched, so a rotation with consecutive pairs (m, w0 ), (m0 , w) must be eliminated first. The details of exactly
when one rotation must be eliminated before another are not of direct use to us (we only require the rather coarse
description in Claim 4.6), so we do not describe them here. See [IL86] or [GI89] for a full description of the poset.
9
This observation is not novel; for example, it is implicit in the discussion of [GI89], but we have not seen the
statement explicitly written down, so we prove it here.
9
5
Conclusion
We have shown there is some constant c such that f (n) ≤ cn . We have not made a significant effort
to optimize the constants in our argument, favoring ease of exposition over the exact result. By
making a few minor changes to the argument, we obtain f (n) ≤ 217n for sufficiently large n. A
more careful argument could probably improve this constant somewhat, but this approach will not
get a constant c close to the (approximately) 2.28 we would need to match the best known lower
bound. Determining the precise asymptotic behavior of f (n) remains an interesting open problem.
Acknowledgements
The first author gratefully acknowledges the support of NSF grant CCF 1420381. The second author
gratefully acknowledges the support of NSF grant CCF-1552097 and ONR-YIP grant N00014-17-12429.
References
[AKL17]
Itai Ashlagi, Yash Kanoria, and Jacob D Leshno. Unbalanced random matching markets:
The stark effect of competition. Journal of Political Economy, 125(1):69–98, 2017. 3
[BCK95]
Arthur T Benjamin, Cherlyn Converse, and Henry A Krieger. How do I marry thee?
Let me count the ways. Discrete Applied Mathematics, 59(3):285–292, 1995. 3
[BGR08]
Nayantara Bhatnagar, Sam Greenberg, and Dana Randall. Sampling stable marriages:
why spouse-swapping won’t work. In Proceedings of the nineteenth annual ACM-SIAM
symposium on Discrete algorithms, pages 1223–1232. Society for Industrial and Applied
Mathematics, 2008. 3, 4
[CGM12] Prasad Chebolu, Leslie Ann Goldberg, and Russell Martin. The complexity of approximately counting stable matchings. Theoretical Computer Science, 437:35–68, 2012. 4
[DBŚ13]
Ewa Drgas-Burchardt and Zbigniew Świtalski. A number of stable matchings in models
of the gale–shapley type. Discrete Applied Mathematics, 161(18):2932–2936, 2013. 3
[DGGJ04] Martin Dyer, Leslie Ann Goldberg, Catherine Greenhill, and Mark Jerrum. The relative
complexity of approximate counting problems. Algorithmica, 38(3):471–500, 2004. 4
[DM10]
Brian C Dean and Siddharth Munshi. Faster algorithms for stable allocation problems.
Algorithmica, 58(1):59–81, 2010. 2
[GI89]
Dan Gusfield and Robert W Irving. The stable marriage problem: structure and algorithms. MIT press, 1989. 2, 3, 4, 8, 9
[GJ12]
Leslie Ann Goldberg and Mark Jerrum. Approximating the partition function of the
ferromagnetic potts model. Journal of the ACM (JACM), 59(5):25, 2012. 4
[GS62]
David Gale and Lloyd S Shapley. College admissions and the stability of marriage. The
American Mathematical Monthly, 69(1):9–15, 1962. 1, 2
10
Men’s
m1
m2
m3
m4
m5
m6
m7
m8
preferences:
w1 w2 w3
w2 w1 w4
w3 w4 w1
w4 w3 w2
w5 w6 w7
w6 w5 w8
w7 w8 w5
w8 w7 w6
Women’s preferences:
w1 m8 m7 m6
w2 m7 m8 m5
w3 m6 m5 m8
w4 m5 m6 m7
w5 m4 m3 m2
w6 m3 m4 m1
w7 m2 m1 m4
w8 m1 m2 m3
w4
w3
w2
w1
w8
w7
w6
w5
m5
m6
m7
m8
m1
m2
m3
m4
w5
w6
w7
w8
w1
w2
w3
w4
m4
m3
m2
m1
m8
m7
m6
m5
w6
w5
w8
w7
w2
w1
w4
w3
m3
m4
m1
m2
m7
m8
m5
m6
w7
w8
w5
w6
w3
w4
w1
w2
m2
m1
m4
m3
m6
m5
m8
m7
w8
w7
w6
w5
w4
w3
w2
w1
m1
m2
m3
m4
m5
m6
m7
m8
Figure 2: A size-8 instance of stable matching with its rotation poset. This is part of the family of
instances described by Irving and Leather, which produces the Ω(2.28n ) lower bound.
11
[IL86]
Robert W Irving and Paul Leather. The complexity of counting stable marriages. SIAM
Journal on Computing, 15(3):655–667, 1986. 2, 3, 4, 7, 8, 9
[IM05]
Nicole Immorlica and Mohammad Mahdian. Marriage, honesty, and stability. In Proceedings of the sixteenth annual ACM-SIAM symposium on Discrete algorithms, pages
53–62. Society for Industrial and Applied Mathematics, 2005. 3
[KMP90] Donald E Knuth, Rajeev Motwani, and Boris Pittel. Stable husbands. Random Structures & Algorithms, 1(1):1–14, 1990. 3
[Knu76]
Donald Ervin Knuth. Mariages stables et leurs relations avec d’autres problèmes combinatoires: introduction à l’analyse mathématique des algorithmes. Montréal: Presses de
l’Université de Montréal, 1976. 1, 2
[Knu97]
Donald Ervin Knuth. Stable marriage and its relation to other combinatorial problems:
An introduction to the mathematical analysis of algorithms, volume 10. American Mathematical Soc., 1997. English translation of Mariages stables. 4
[Lev17]
Avi Levi. Novel uses of the Mallows model in coloring and matching. PhD thesis,
University of Washington, 2017. 3
[Mal57]
Colin L Mallows. Non-null ranking models. i. Biometrika, 44(1/2):114–130, 1957. 3
[Man13]
David F Manlove. Algorithmics of matching under preferences, volume 2. World Scientific, 2013. 3, 4
[Pit89]
Boris Pittel. The average number of stable matchings. SIAM Journal on Discrete
Mathematics, 2(4):530–549, 1989. 3
[Ram12]
Catherine Rampell. 2 From U.S. win Nobel in economics. The New York Times, October
15, 2012. 2
[Rot15]
Alvin E Roth. Who Gets What–and Why: The New Economics of Matchmaking and
Market Design. Houghton Mifflin Harcourt, 2015. 4
[RS92]
Alvin E Roth and Marilda A Oliveira Sotomayor. Two-sided matching: A study in
game-theoretic modeling and analysis. Cambridge University Press, 1992. 4
[Sta53]
John M Stalnaker. The matching program for intern placement. Academic Medicine,
28(11):13–19, 1953. 2
[Sta11]
Georgios K. Stathopoulos. Variants of stable marriage algorithms, complexity and structural properties. Master’s thesis, University of Athens, Department of Mathematics,
2011. 3
[Thu02]
Edward G Thurber. Concerning the maximum number of stable matchings in the stable
marriage problem. Discrete Mathematics, 248(1-3):195–219, 2002. 3
12
| 8cs.DS
|
arXiv:1412.0442v3 [math.ST] 30 Sep 2016
Bernoulli 23(1), 2017, 479–502
DOI: 10.3150/15-BEJ750
Exact confidence intervals and hypothesis
tests for parameters of discrete distributions
MÅNS THULIN1 and SILVELYN ZWANZIG2
1
Department of Statistics, Uppsala University, 751 05 Uppsala, Sweden.
E-mail: [email protected]
2
Department of Mathematics, Uppsala University, 751 05 Uppsala, Sweden.
E-mail: [email protected]
We study exact confidence intervals and two-sided hypothesis tests for univariate parameters of
stochastically increasing discrete distributions, such as the binomial and Poisson distributions.
It is shown that several popular methods for constructing short intervals lack strict nestedness,
meaning that accepting a lower confidence level not always will lead to a shorter confidence
interval. These intervals correspond to a class of tests that are shown to assign differing pvalues to indistinguishable models. Finally, we show that among strictly nested intervals, fiducial
intervals, including the Clopper–Pearson interval for a binomial proportion and the Garwood
interval for a Poisson mean, are optimal.
Keywords: binomial distribution; confidence interval; expected length; fiducial interval;
hypothesis test; Poisson distribution
1. Introduction
Hypothesis testing and interval estimation of parameters in discrete distributions are
two of the classic statistical problems, particularly for the binomial and Poisson distributions, which remain two of the most important statistical models. The fact that these
distributions are discrete makes it impossible to construct non-randomized confidence
intervals that have coverage equal to 1 − α for all values of the unknown parameter θ,
and, equivalently, impossible to construct two-sided tests with size equal to α for all pairs
(α, θ0 ), where θ0 denotes the value of θ under the null hypothesis. It is however possible
to construct confidence intervals that have coverage at least equal to 1 − α for all values
of the unknown parameter, and tests that have size at most equal to α. Such intervals
and tests are called exact, and are the topic of this paper.
Given an observation x, the classic method of constructing exact confidence intervals
for parameters of some common discrete distributions is to use the fiducial interval of
This is an electronic reprint of the original article published by the ISI/BS in Bernoulli,
2017, Vol. 23, No. 1, 479–502. This reprint differs from the original in pagination and
typographic detail.
1350-7265
c
2017 ISI/BS
2
M. Thulin and S. Zwanzig
Fisher [17, 36]: (θL , θU ) where θL and θU are such that
X
X
PθL (X = k) = α/2 and
PθU (X = k) = α/2.
k≤x
(1)
k≥x
For the binomial parameter, the fiducial interval is known as the Clopper–Pearson interval
[11] and for the mean of a Poisson distribution it is known as the Garwood interval [18].
The hypothesis H0 : θ = θ0 can be tested against the alternative H1 : θ 6= θ0 by checking
whether θ0 is contained in the fiducial interval. The p-value λf (θ0 , x) of this test is two
times the smaller p-value of two one-sided tests:
X
X
λf (θ0 , x) = min 2 ·
Pθ0 (X = k), 2 ·
(2)
Pθ0 (X = k), 1 .
k≤x
k≥x
In their seminal paper on binomial confidence intervals, Brown et al. [7] write: “The
Clopper–Pearson interval is wastefully conservative and is not a good choice for practical
use, unless strict adherence to the prescription C(p, n) > 1 − α is demanded,” where
C(p, n) denotes the coverage probability. Instead they recommend using approximate
intervals, which obtain the nominal confidence level 1 − α in some average sense, but
have lower coverage for some values of θ. Such intervals are typically shorter than exact
intervals, and their corresponding tests typically have higher power. These advantages
comes at the cost that the actual confidence levels may be much lower than stated and
that the size of tests may be inflated. For popular approximate intervals, the deviations
in coverage from 1 − α may be non-negligible even for large sample sizes [31]. For this
reason, some statistician prefer to use exact methods like those discussed in this paper,
in order to guarantee that confidence levels are not exaggerated and type I error rates
are not understated.
When other criteria than merely coverage levels and expected lengths are considered,
exact confidence intervals can moreover compare favourably to approximate intervals
[24, 32]. Finally, even if one prefers to use average coverage as a criterion for comparing
confidence intervals, it is of interest to study exact intervals due to the facts that these
intervals can be adjusted to have coverage 1 − α on average, and that such adjusted
intervals tend to have shorter expected length than other approximate intervals [25, 30].
For comparisons of exact and approximate intervals in the binomial setting, and further
arguments for using exact methods for discrete distributions, see [31].
Regarding the fiducial Clopper–Pearson interval, Brown et al. [7] also write “better
exact methods are available; see, for instance, [5] and [9].” Fiducial intervals are equaltailed, meaning that the lower bound is a 1 − α/2 lower confidence bound and that
the upper bound is a 1 − α/2 upper confidence bound. Several authors, including those
mentioned by Brown et al. [7] in the above quote, have proposed shorter exact intervals
that improve upon fiducial intervals by letting the tail-coverages vary for different x,
so that their bounds no longer are 1 − α/2 confidence bounds [4, 5, 9, 10, 13, 14, 19,
21, 26, 28, 35]. Such intervals, known as strictly two-sided intervals, tend to have less
conservative coverage and are typically shorter than fiducial intervals. Their use has been
advocated by [1, 2, 15, 16, 20, 25, 27] and [22], among others.
Exact intervals and tests for discrete distributions
3
Figure 1. p-values and interval bounds for the mean of a Poisson distribution, when x = 9 has
been observed. The strictly two-sided Sterne [28] method is shown in black, and the fiducial
Garwood [18] method is shown in grey.
Unlike the equal-tailed fiducial intervals, the p-values of tests corresponding to strictly
two-sided confidence intervals can not be written as two times the smaller p-value of
two one-sided tests. Instead, for some test statistic T (θ0 , X) satisfying mild regularity
conditions detailed in Section 2, the p-value of a strictly two-sided test is defined as
λ(θ0 , x) = Pθ0 (T (θ0 , X) ≥ T (θ0 , x)).
If the null distribution of T (θ0 , X) is asymmetric, the level α rejection region of such a
test is not the intersection of the rejection regions of two one-sided level α/2 tests.
The main goal of this paper is to show that strictly two-sided confidence intervals and
hypothesis tests suffer from several problems. These are illustrated in Figure 1, in which
the p-values and interval bounds for the mean of a Poisson distribution are shown for
two tests and their corresponding confidence intervals. The first of these is the strictly
two-sided Sterne [28] interval, the other being the fiducial Garwood [18] interval.
In the spirit of Birnbaum [3], the p-values are plotted as a function of the value θ0 of
the parameter under the null hypothesis. In the Poisson model, it is reasonable to expect
that a small change in the null value of θ should lead to a small change in the p-value,
since Pθ (X = x) is continuous in θ, so that there is no concernable difference between
the Poisson(θ) and Poisson(θ + ε) models when ε is infinitesimal. This is not the case for
the strictly two-sided test: its p-value is discontinuous when viewed as a function of θ0 .
The evidence against two models, which for all practical purposes are indistinguishable,
can therefore differ greatly. Several examples of this are seen in Figure 1; the p-value for
θ0 = 4.954163, for instance, is 0.0722, so that the null hypothesis is rejected at the 10%
level, while the highly similar hypothesis θ0 = 4.954164 cannot be rejected as its p-value
is 0.1071.
Moreover, we would expect that the p-value increases as θ0 goes from 0 to the observed
x, and that it thereafter decreases, since this would mean that the p-value becomes smaller
when the null hypothesis agrees less with the data. This is not the case for the strictly
two-sided test. Instead, the p-value sometimes increases when the null θ is changed to
4
M. Thulin and S. Zwanzig
agree less with the observed x. As an example, consider the p-values shown in Figure 1.
When x = 9 has been observed from a Poisson distribution, the p-value when θ0 = 15.6
is 0.0993, so that the null hypothesis is rejected at the 10% level. However, even though
x = 9 disagrees even more with the null hypothesis θ0 = 15.95, the p-value for this θ0 is
0.1011, and the hypothesis can not be rejected. The test corresponding to the fiducial
interval does not suffer from either of these problems.
The strictly two-sided confidence interval is no better than its corresponding test.
When the interval bounds are plotted as functions of the confidence level 1 − α, we
see two phenomenons. The first is that the interval bounds are discontinuous in 1 − α,
meaning that a small change in α can cause one of the interval bounds to leap. The second
is that the bounds sometimes are constant, meaning that a change in α not necessarily
will lead to a change in the bounds. For some α, both bounds remain unchanged in an
interval (α − ε, α + ε). There is therefore no guarantee that accepting a larger α will lead
to a shorter interval; we say that the interval is not strictly nested. The fiducial interval
does not suffer from either of these problems.
These properties can also cause strictly two-sided test and intervals to behave strangely
as more data is collected. As an example, consider the Blaker [4] test for the negative
binomial proportion θ. When k = 19 successes are observed after x = 38 trials, the maximum likelihood estimator is θ̂ = 0.5 and Blaker p-value for the test of the hypothesis
θ = 0.625 is 0.0.0929, causing us to reject the null hypothesis at the 10% level. If we then
decide to collect more data by requiring that k = 20 successes should be observed, and
observe one failure and one success so that x = 40, θ̂ is still 0.5. We would now expect
the p-value to decrease as this outcome appears to be even less in line with θ = 0.625.
Instead, the Blaker p-value for k = 20 and x = 40 is 0.106, and we can no longer reject the
null hypothesis at the 10% level. Analogous problems arise for confidence intervals. The
90% Blaker confidence interval for θ given k = 19 and x = 38 is (0.35992, 0.62279), while
for k = 20 and x = 40 it is (0.36202, 0.62689). The latter interval is not, as we normally
would expect, a subset of the former. Moreover, the interval based on more data is wider
than the interval based on less data: the interval widths are 0.263 and 0.265, respectively.
As we will see, intervals lacking strict nestedness is equivalent to their corresponding
p-values being discontinuous in θ. Consequently, intervals which are not strictly nested
correspond to tests that attach widely differing evidence to indistinguishable hypotheses.
We believe that this is an unacceptable property of a hypothesis test, and argue that
such intervals and tests should be avoided.
In this paper, we show that these problems are universal for strictly two-sided intervals
and tests, when the data is generated by a class of discrete distributions that includes
the binomial, Poisson and negative binomial distributions. They also carry over to exact analysis of contingency tables and discrete models with nuisance parameters, when
such analyses are based on conditioning that reduces the problem to a one-parameter
framework.
In Section 2, we give a formal description of the setting for our results. We then show
that the p-values of strictly two-sided tests are discontinuous, and that their corresponding intervals have bounds that are not strictly monotone. Finally, we show that strictly
two-sided intervals never are strictly nested, meaning that both interval bounds simultaneously may remain unchanged when α is changed. Section 3 is devoted to showing that
Exact intervals and tests for discrete distributions
5
strictly two-sided intervals typically have bounds that moreover are discontinuous in α,
and that the corresponding p-values lack desirable monotonicity properties. In Section 4,
it is then demonstrated that fiducial intervals not only are strictly nested but also are
the shortest equal-tailed intervals. The paper concludes with a discussion in Section 5.
Most proofs and some technical details are contained in two appendices.
2. The lack of strict nestedness and its implications
2.1. Setting
This section is concerned with nestedness. We start by defining this concept.
Definition 1. A confidence interval is nested if the 1 − α interval is a subset of the
1 − α0 interval when 1 > α > α0 > 0, and strictly nested if the 1 − α interval always is a
proper subset of the 1 − α0 interval.
If an interval is not strictly nested, accepting a lower confidence level does not always
yield a shorter interval, so that sometimes nothing is gained by increasing α. Despite the
importance of nestedness, this property has not been discussed much in the literature,
likely because it is taken for granted. Notable exceptions are Blaker [4], who proved that
the binomial Blyth–Still–Casella interval is not strictly nested and Vos and Hudson [33],
who showed by example that the Blaker interval for a binomial proportion lacks strict
nestedness.
Next, we give some definitions and state the assumptions under which strictly twosided intervals are not strictly nested. We will limit our study to parameters of discrete
distributions Pθ belonging to a class P(Θ, X ).
Definition 2. Let θ ∈ Θ denote an unknown parameter, with Θ being a connected open
subset of R, and let X ⊆ Z be a sample space consisting of consecutive integers. A family
of distributions Pθ on X parameterized by θ ∈ Θ belongs to P(Θ, X ) if
A1. ∀(θ, x) ∈ Θ × X , Pθ (X = x) > 0,
A2. Pθ is stochastically increasing, i.e. Pθ (X ≤ x) is strictly decreasing in θ for any
fixed x ∈ X \ sup X ,
A3. For any fixed x ∈ X , Pθ (X = x) is differentiable in θ.
Conditions A1–A3 are satisfied by for instance the binomial, Poisson and negative
binomial distributions as long as Θ is the natural parameter space, that is, as long as
it has not been restricted. This follows directly from the proposition below, the proof
of which is given in Appendix B. The conditions are typically also satisfied for other
common parameterizations.
Proposition 1. If Pθ constitutes a regular discrete one-parameter exponential family
with an increasing likelihood ratio, where θ is the natural parameter, then Pθ ∈ P(Θ, X ).
6
M. Thulin and S. Zwanzig
To fully understand the implications of the lack of nestedness, we will study the hypothesis tests to which non-nested intervals correspond, so-called strictly two-sided tests:
Definition 3. Consider a two-sided test of H0 : θ = θ0 versus H1 : θ 6= θ0 , with a
test statistic T (θ0 , x). The test is called strictly two-sided if the p-value of the test
is λ(θ0 , x) = Pθ0 (T (θ0 , X) ≥ T (θ0 , x)) and it satisfies conditions B1–B2 below. Moreover, in case λ(θ, x), viewed as a function of θ, has a jump at θ0 we define λ(θ0 , x) =
lim inf θ→θ0 λ(θ, x).
B1. For any x ∈ X , there exists a θx ∈ Θ such that T (θx , x) < T (θx , y) for all y ∈
X \ {x}.
B2. There exists a θ0 ∈ Θ such that there does not exist a µ ∈ Θ for which
Pθ0 (T (θ0 , X) = µ − k) = Pθ0 (T (θ0 , X) = µ + k) for all k : µ ± k ∈ X .
Condition B1 is included to ensure that the test does not yield the same result for
all x and θ. The name strictly two-sided comes from condition B2, which ensures that
the p-value must be computed by comparing the test statistic to both tails of the null
distribution simultaneously.
The p-value of a strictly two-sided test can be written as
λ(θ, x) =
X
k∈Aθ,x
Pθ (X = k)
where Aθ,x = {k ∈ X : T (θ, k) ≥ T (θ, x)}.
(3)
For simplicity, we will assume that the test statistic is such that
B3. For any θ ∈ Θ, there exists xθ ∈ X such that T (θ, x) is decreasing in x when x < xθ
and increasing in x when x > xθ .
Under B3, the set Aθ,x has a particularly simple form.
Proposition 2. Under B3, the functions k1 (θ, x) := min{k ≥ xθ : T (θ, k) ≥ T (θ, x)} and
k2 (θ, x) := max{k ≤ xθ : T (θ, k) ≥ T (θ, x)} are such that
Aθ,x = {k ∈ X : k ≥ k1 (θ, x)} ∪ {k ∈ X : k ≤ k2 (θ, x)}.
(4)
For any x, at least one of k1 (θ, x) and k2 (θ, x) is non-constant in θ.
The proof of the proposition is given in Appendix B.
When x is fixed and θ is varying we will refer to λ(θ, x) as the p-value function. We
define the corresponding confidence interval using the convex hull of {θ : λ(θ, x) > α} to
ensure that it in fact is an interval; as we will see in Section 3, {θ : λ(θ, x) > α} itself
is not always connected. The interval in the following definition is guaranteed to be
nested: if α > α0 the convex hull of {θ : λ(θ, x) > α} is a subset of the convex hull of
{θ : λ(θ, x) > α0 }.
7
Exact intervals and tests for discrete distributions
Definition 4. The 1 − α confidence interval Iα (x) = (Lα (x), Uα (x)) corresponding to a
test is
Iα (x) = (inf{θ : λ(θ, x) > α}, sup{θ : λ(θ, x) > α}).
(5)
A confidence interval is said to be strictly two-sided if it is based on the inversion of a
strictly two-sided test.
2.2. Examples of strictly two-sided tests
We will focus on four commonly used strictly two-sided tests, which satisfy conditions B1,
B2 and B3 for some common discrete distributions, including the binomial, Poisson and
negative binomial distributions. These tests are briefly described below. Further details,
as well as conditions for B1–B3 to hold, are given in Appendix A.
The likelihood ratio test, for which T (θ, x) is the likelihood ratio statistic [20, 27].
The score test, for which T (θ, x) is the score statistic [20, 27].
The Sterne test, for which T (θ, x) = 1/Pθ (X = x) [28].
The Blaker test, which in fact is a class of tests. Given a statistic S(x), the Blaker statistic is T (θ, x) = 1/ min{Pθ (S(X) ≤ S(x)), Pθ (S(X) ≥ S(x))}, was introduced in Blaker [4].
See also [37] for a interpretation based on confidence curves. In the binomial, negative
binomial and Poisson settings, we will use the sufficient statistic S(x) = x, as is common.
In Section 2.5, we will discuss confidence intervals that have varying tail-coverage but
are based on minimization algorithms rather than test inversion. Because these intervals
do not fall under Definition 4 we will refer to them as being of strictly two-sided-type
rather than as being strictly two-sided.
2.3. Lack of strict nestedness and its interpretation
We will now show that strictly two-sided intervals lack strict nestedness, and that this is
caused by jumps in the p-value function λ(θ, x), viewed as a function of θ.
Proposition 3. Assume that Pθ ∈ P(Θ, X ). Let λ(θ, x) be the p-value function of a
strictly two-sided test and let Iα (x) denote its corresponding strictly two-sided confidence
interval. Then for any x ∈ X :
(a) λ(θ, x) is not continuous in θ,
(b) the bounds of Iα (x) are not strictly monotone in α,
(c) Iα (x) is not strictly nested.
First, we show that λ(θ, x) has jumps. For any fixed x ∈ X , by Proposition 2 we have,
under B3,
X
X
X
Pθ (X = k),
(6)
Pθ (X = k) +
λ(θ, x) =
Pθ (X = k) =
k∈Aθ,x
k≥k1 (θ,x)
k≤k2 (θ,x)
8
M. Thulin and S. Zwanzig
where at least one of the ki (θ, x) is non-constant in θ. ki (θ, x) are integer-valued stepfunctions. Thus, for ε > 0 whenever ki (θ, x) < ki (θ + ε, x), ki must have a jump between
θ and θ + ε. This induces a jump in the p-value function as well. To see this, assume
without loss of generality that k1 (θ + ε, x) = k1 (θ, x) and k2 (θ + ε, x) = k2 (θ, x) + 1. Then
λ(θ + ε, x) =
X
k≥k1 (θ,x)
Pθ+ε (X = k) +
X
Pθ+ε (X = k) + Pθ+ε (X = k2 (θ, x) + 1),
k≤k2 (θ,x)
but by A1 and A3,
lim λ(θ + ε, x) = λ(θ, x) + Pθ (X = k2 (θ, x) + 1) > λ(θ, x).
εց0
Thus λ(θ + ε, x) 6ց λ(θ, x) as ε ց 0 and the function is hence not continuous in θ. In
particular, we have shown that λ(θ, x) has the following property:
Lemma 1. Under the assumptions of Theorem 3, λ(θ, x) as a function of θ has a jump
whenever a point is added to or removed from Aθ,x .
Values of α for which Iα (x) is not strictly nested correspond to the jumps in λ(θ, x).
To see this, note that if the interval (α0 , α1 ) ⊆ (0, 1) is such that
{θ : λ(θ, x) ∈ (α0 , α1 )} = ∅
(7)
then for α ∈ (α0 , α1 ), we have λ(θ, x) > α if and only if λ(θ, x) > α1 , which means that
the lower interval bound
Lα (x) = inf{θ : λ(θ, x) > α} = inf{θ : λ(θ, x) > α1 } = Lα1 (x)
so that Lα (x) is not strictly monotone in α. By definition, the interval is not strictly
nested if there exists an α such that both Lα (x) and the upper interval bound Uα (x)
simultaneously are constant in a neighbourhood of α. The proof that there always exists
such an α is somewhat technical, and is deferred to Appendix B.
In particular, Proposition 3 holds when the test and its corresponding confidence interval are exact. The proposition is illustrated for exact tests and intervals in Figures 2–3.
In Figure 2, p-values for the strictly two-sided [4, 28], likelihood ratio and score tests
[20, 27] are compared to the p-values of the non-strictly two-sided test that corresponds
to the fiducial interval in the Poisson and binomial settings. It is readily verified that the
strictly two-sided tests satisfy B1–B3; see Appendix A. In Figure 3, the interval bounds
of some strictly two-sided intervals are compared to the bounds of the fiducial interval.
In the Poisson case, the Sterne, Blaker, likelihood ratio, score, Crow–Gardner [10, 14]
and Kabaila–Byrne [21] (the latter two being of strictly two-sided-type) intervals are
compared to the Garwood interval. In the binomial case, the Sterne, Blaker, likelihood
ratio, score, Crow [5, 9, 13] (which is of strictly two-sided-type) and Göb and Lurz [19]
intervals are compared to the Clopper–Pearson interval.
Exact intervals and tests for discrete distributions
9
Figure 2. Unlike the p-values for the fiducial test (shown in grey in all plots), the strictly
two-sided Sterne, Blaker, likelihood ratio (LR) and score p-values are discontinuous and not
bimonotone. In (a), the p-values are shown when x = 2 is an observation from a Poisson distribution with null mean θ. In (b), the p-values are shown when x = 2 is an observation from a
null Bin(20, θ)-distribution.
10
M. Thulin and S. Zwanzig
Figure 3. Interval bounds of several strictly two-sided and strictly two-sided-type confidence
intervals. The intervals are compared to the fiducial interval, the bounds of which are plotted in grey. In (a), the intervals are shown when x = 2 is an observation from a Poisson distribution with mean θ. In (b), the intervals are shown when x = 2 is an observation from a
Bin(20, θ)-distribution.
Exact intervals and tests for discrete distributions
11
2.4. The largest α for which an interval is strictly nested
Proposition 3 tells us that strictly two-sided confidence intervals lack strict nestedness
and that their bounds are not strictly monotone in α. This may however not be a great
problem if the lack of strict nestedness and monotonicity occurs only for α close to 1.
Under some stronger assumptions on T (θ, x), X and Pθ we can derive expressions for
the largest α for which Iα (x) is strictly nested and the largest α for which each interval
bound is strictly monotone. As we will see, these bounds for α are usually close to 0,
meaning that the lack of strict nestedness and monotonicity occurs also for α that are
used in practice.
We restrict our attention to samples spaces of the form X = {0, 1, 2, . . .} or X =
{0, 1, 2, . . ., n}, for some known n < ∞. Moreover, we will require some additional conditions, which essentially make up stronger versions of A2 and B3:
A2+ . Pθ (X ≤ x) is strictly decreasing in θ for any x ∈ X \ sup X .
B3+ . (i) For any θ ∈ Θ, there exists xθ ∈ X such that T (θ, x) is strictly decreasing
in x when x < xθ and strictly increasing in x when x > xθ .
(ii) For any x ∈ X , there exists a θx ∈ Θ such that λ(θx , x) = 1 and T (θ, x) is
strictly decreasing in θ when θ < θx and strictly increasing in θ when θ > θx .
(iii) xθ is an increasing function of θ.
Proposition 4. Assume that X = {0, 1, 2, . . .} or X = {0, 1, 2, . . ., n}. Under A2+ , B3+
and the assumptions of Proposition 3 it holds that
(a) There exists an αnest > 0 such that Iα (x) is strictly nested for all x ∈ X and
α ≤ αnest .
(b) Let αL = inf x∈X inf θ∈{θ:T (θ,0)>T (θ,x)} λ(θ, x). Then (i) αL > 0, (ii) for all x > 0,
Lα (x) is continuous and strictly increasing in α when α ≤ αL , and (iii) there exists an
x > 0 and an ε > 0 such that Lα (x) is constant in (αL , αL + ε).
(c) For X = {0, 1, 2, . . ., n}, let αU = inf x∈X supθ∈{θ:T (θ,n)>T (θ,x)} λ(θ, x). Then (i)
αU > 0, (ii) for all x < n, Uα (x) is continuous and strictly decreasing in α when α ≤ αU ,
and (iii) there exists an x > 0 and an ε > 0 such that Uα (x) is constant in (αU , αU + ε).
Proposition 4 deals with α guaranteeing strict monotonicity and nestedness for all x.
We can also study monotonicity and nestedness for fixed x. For any x ∈ X , let αL (x)
denote the largest α for which Lα (x) is strictly monotone, and αU (x) denote the largest α
for which Uα (x) is strictly monotone. Finally, let αnest (x) be the largest α for which Iα (x)
is strictly nested. In Figure 4(a), these quantities are shown for the Blaker interval for a
binomial proportion, with n = 20 and x ∈ {1, 2, . . . , 19}. In this example, αnest (x) < 0.1
for most x. As is seen, αnest (x) is often equal to or very close to max(αL (x), αU (x)).
Figures for other intervals, other n and other distributions are similar.
Figure 4(b) shows αnest for the binomial Blaker interval as a function of the sample size
n. It is seen that when 7 ≤ n ≤ 100 we have αnest < 0.01 for the Blaker interval, meaning
that the interval lacks strict nestedness for virtually all values of α that actually are used
in practice for these sample sizes.
12
M. Thulin and S. Zwanzig
Figure 4. (a) The largest α for which the lower and upper bounds of the Blaker interval
for a binomial proportion are strictly monotone (αL (x) and αU (x)), and the largest α for
which the interval is nested conditioned on x (αnest (x)), when n = 20. The common choices
α ∈ {0.01, 0.05, 0.1} are shown as dashed lines. (b) αnest , the largest α for which the Blaker
interval for a binomial proportion is strictly nested, as a function of n.
2.5. Confidence intervals not based on test-inversion
An interesting class of confidence intervals are based on minimization algorithms. This
class includes [5, 9, 10, 13, 14, 21] and [26] intervals. For such intervals, the shortest
interval is determined for each α. What typically occurs for these intervals is that they
correspond to inversion of different tests for different α. Often this will result in intervals
that lack nestedness (and not only strict nestedness), as it leads to some values of θ
having multiple p-values attached to them. This can be seen in Figure 3: neither the
Crow interval for the binomial parameter nor the Crow–Gardner and Kabaila–Byrne
intervals for the Poisson parameter are nested.
If a two-sided 1 − α interval is (θℓ , θu ), then the p-values for the corresponding twosided tests of the hypotheses θ0 = θℓ and θ0 = θu are α. Using this relationship, we can
plot the p-value functions of tests corresponding to intervals that are not defined in terms
of test inversion, such as minimization-based intervals. The lack of nestedness means that
the p-value function λ(θ, x) of the corresponding test is not a proper function for x ∈ X
fixed, since some values of θ are mapped to more than one p-value. For some intervals,
this problem becomes extreme. Two examples of this are the Kabaila–Byrne and Crow–
Gardner intervals for a Poisson mean, shown in Figure 5. For other intervals, the lack of
nestedness results in less extreme p-value functions. An example of this is the Schilling–
Doi interval for a binomial proportion; in Figure 5 the jumps in its p-value function are
shown as vertical lines, in order to make the consequences of the non-nestedness easier
to spot.
Exact intervals and tests for discrete distributions
13
Figure 5. Comparison between p-values corresponding to fiducial intervals (grey) and p-values
corresponding to some minimization-based intervals (black). The p-values of the tests corresponding to the Kabaila–Byrne and Crow–Gardner intervals are shown for x = 2 being an
observation from a Poisson distribution with null mean θ, and the p-values of the test corresponding to the Schilling–Doi interval are shown for x = 8 being an observation from a null
Bin(20, θ)-distribution.
3. Continuity and bimonotonicity
For Θ ⊆ R, we say that a function f : Θ → R is strictly bimonotone on Θ if there exist
θ0 , θ1 ∈ Θ such that f is strictly increasing on (inf Θ, θ0 ), constant on (θ0 , θ1 ) and strictly
decreasing on (θ1 , sup Θ).
As have been argued for example, by Hirji [20] and Vos and Hudson [33], this type of
bimonotonicity is a highly desirable property of p-values when viewed as a function of
θ. Ideally λ(θ, x) should increase monotonically from 0 to 1 and then decreases monotonically to 0, just like the p-values of the tests corresponding to fiducial intervals do in
Figure 2. One reason that this property is desirable is the following result.
Proposition 5. The bounds of a confidence interval are discontinuous in α if their
corresponding p-value function is not strictly bimonotone in θ.
Proof. Assume without loss of generality that there exist θ0 < θ1 < inf{θ : λ(θ, x) = 1}
such that λ(θ, x) is increasing in θ in the interval (inf Θ, θ0 ) and decreasing or constant
in the interval (θ0 , θ1 ). Let α0 = λ(θ0 , x). Then θ1 = inf{θ > θ0 : λ(θ, x) > α0 }. Thus
Lα0 (x) = θ0 but for all ε > 0, Lα+ε (x) ≥ θ1 , meaning that Lα (x) has a jump of length
θ1 − θ0 > 0 at α = α0 . An analogous argument holds for the upper bound.
Hirji [20] mentions that p-value functions of strictly two-sided tests need not be bimonotone, whereas Vos and Hudson [33] showed by example that the Blaker test for a
binomial proportion lacks bimonotonicity. Upon closer inspection of Figures 2 and 3, it
can be seen that all the strictly two-sided tests considered here suffer from this problem.
Next, we give a condition under which the p-value function of a strictly two-sided
test is strictly bimonotone for fixed x, the derivation of which is given in Appendix B.
The bimonotonicity condition requires the following additional assumptions, which are
satisfied by the binomial, negative binomial and Poisson distributions.
14
M. Thulin and S. Zwanzig
A4. For x ∈ X \ sup X , limθ→inf Θ Pθ (X ≤ x) = 1 and limθ→sup Θ Pθ (X ≤ x) = 0.
P
P
A5. For k1 , k2 ∈ X such that k1 ≥ k2 + 2, k≥k1 Pθ (X = k) + k≤k2 Pθ (X = k) has a
unique minimum in the interior of Θ.
Proposition 6. Under the assumptions and notation of Proposition 3, assume that Pθ
satisfies conditions A4 and A5. Let θr (θ0 , x) be the solution to
k1 (θ0 ,x)−1
X
k=k2 (θ0 ,x)+1
d
Pθ (X = k) = 0
dθ
(8)
in the interior of Θ. Then
(a) λ(θ, x) is strictly bimonotone in θ for any fixed x ∈ X \ sup X .
(b) The bounds of Iα (x) are continuous in α,
if and only if there does not exist (θ0 , x) such that either
θ0 < inf{θ : λ(θ, x) = 1}
and
θ0 < θr (θ0 , x),
θ0 > sup{θ : λ(θ, x) = 1}
and
θ0 > θr (θ0 , x).
or
(9)
For any given Pθ , we can evaluate numerically whether the bimonotonicity condition
(9) is violated for a pair (θ0 , x). We have not been able to find a strictly two-sided test
that passes (9) for any x. Proposition 6 is illustrated in the Poisson and binomial settings
in Figures 2–3. When x = 2 from a Poisson random variable has been observed, the pvalue functions of the Sterne and Blaker tests are non-bimonotone for the first time when
θ = 3. For the likelihood ratio√test, the first occurrence is at θ = 1 and for the score test
the first occurrence is at θ = 12.
A consequence of λ(θ, x) lacking bimonotonicity is that the confidence “interval” {θ :
λ(θ, x) > α} may contain holes, and therefore not be an interval at all. The common
remedy for this is to redefine the intervals as the convex hull of {θ : λ(θ, x) > α}, as we
did in Definition 4. This does not change the infimum or supremum of the set, and does
therefore not affect nestedness or continuity of the bounds. Similarly, Fay [15] proposed
handling the problem of non-bimonotone p-value functions by redefining the p-values
using the convex hull of {θ : λ(θ, x) > α}. The redefined p-values are constant where
they previously were non-monotone. By Proposition 5, the bounds of the corresponding
intervals are however still discontinuous in α.
For the binomial and negative binomial distributions, the left-hand side of (8) is a
polynomial of order k1 (θ0 , x) − k2 (θ0 , x) − 1. For the Poisson distribution, it is straightforward to find a general solution to (8), which yields the following proposition, the proof
of which is omitted.
Proposition 7. For X ∼ Poisson(θ), the p-value function λ(θ, x) belonging to a strictly
two-sided test is bimonotone in θ if and only if there does not exist (θ, x) such that either
15
Exact intervals and tests for discrete distributions
)1/(k1 (θ,x)−k2 (θ,x)−1) , or
• θ < inf{θ : λ(θ, x) = 1} and θ < ( (k1k(θ,x)−1)!
2 (θ,x)!
)1/(k1 (θ,x)−k2 (θ,x)−1) .
• θ > sup{θ : λ(θ, x) = 1} and θ > ( (k1k(θ,x)−1)!
2 (θ,x)!
Note that if we let n = k1 (θ, x) − k2 (θ, x) − 1 then
(k1 (θ, x) − 1)!
k2 (θ, x)!
1/(k1 (θ,x)−k2 (θ,x)−1)
k1 (θ,x)−1
=
Y
k=k2 (θ,x)+1
k
!1/n
,
the geometric mean of Acθ,x .
4. Some results for fiducial intervals
4.1. Fiducial intervals are strictly nested and have continuous
bounds
The test corresponding to the fiducial intervals is not strictly two-sided. Its p-values are
defined by (2). The following proposition, the proof of which can be found in Appendix B,
states that fiducial intervals do not suffer from the problems associated with strictly twosided intervals.
Proposition 8. Under A1–A3 and A4, fiducial intervals are strictly nested. Moreover,
for any x ∈ X the bounds of the interval are continuous in α and λf (θ, x) is continuous
in θ.
4.2. Optimality results
For a binomial proportion, Wang [34] presented results claiming that under certain conditions on α and n the fiducial Clopper–Pearson interval is the shortest interval in the
class of exact confidence intervals with monotone bounds. A counterexample to the optimality result of [34] is the strictly two-sided Blaker interval [4], which always is contained
in the Clopper–Pearson interval. Among equal-tailed intervals however, fiducial intervals
posses length optimality properties. We expect that this is known, but have not been able
to find such results in the literature, for which reason we briefly cover length optimality
below.
Our main tool for showing length optimality is a theorem due to [6]. Under assumptions A1, A2 and A3, consider the class ML,α of one-sided 1 − α confidence bounds
(Lα (x), ∞) ∩ Θ for θ ∈ Θ based on an observation x of X ∼ Pθ satisfying the following
three criteria:
C1. Lα (x) ≤ Lα (x + 1),
C2. inf θ∈Θ Pθ (Lα (x) ≤ θ) ≥ 1 − α,
16
M. Thulin and S. Zwanzig
C3. Lα (x) only depends on x, α and Pθ .
Criterion C3 rules out randomized bounds, which can be shorter while maintaining
exact coverage, but rely on conditioning on information not contained in the sufficient
statistic; see, for example, [29]. C3 is implicit in Bolshev’s paper; we have added it here
for clarity. ML,α is the class of monotone exact lower confidence bounds. We call an
interval (or a bound) Iα (x) in a class of intervals K the smallest interval in K if, for
any other interval Iα∗ (x) ∈ K, Iα (x) \ Iα∗ (x) = ∅. For the ML,α class, Bolshev [6] proved
that the one-sided lower fiducial bound is the smallest bound in ML,α . Under analogous
conditions, the upper fiducial bound is similarly the smallest bound in the set MU,α of
exact monotone upper confidence bounds.
The extension of Bolshev’s theorem to two-sided confidence intervals is straightforward
and does not require the additional conditions that Wang [34] used in the binomial setting. Consider the class Mα of exact equal-tailed confidence intervals (Lα/2 (x), Uα/2 (x))
for θ based on an observation x of X ∼ Pθ satisfying
D1. Lα/2 (x) ≤ Lα/2 (x + 1) and Uα/2 (x) ≤ Uα/2 (x + 1),
D2. inf θ∈Θ Pθ (Lα/2 (x) ≤ θ) ≥ 1 − α/2 and inf θ∈Θ Pθ (Uα/2 (x) ≥ θ) ≥ 1 − α/2,
D3. (Lα/2 (x), Uα/2 (x)) only depends on x, α and Pθ .
Note that if an interval belongs to Mα then it is the intersection of a bound in ML,α/2
and a bound in MU,α/2 .
Proposition 9. The fiducial interval is the smallest interval in Mα .
Proof. Let Iα (x) = (Lα/2 (x), Uα/2 (x)) denote the fiducial interval and assume that there
∗
is an interval Iα∗ (x) = (L∗α/2 (x), Uα/2
(x)) in Mα such that Iα (x) \ Iα∗ (x) 6= ∅. Then
∗
L∗α/2 (x) > Lα/2 (x) or Uα/2
(x) < Uα/2 (x). Consequently, at least one of the one-sided
∗
∗
bounds (Lα/2 (x), ∞) ∩ Θ or (−∞, Uα/2
(x)) ∩ Θ is smaller than the corresponding fiducial bound. By Bolshev’s theorem, this means that Iα∗ (x) is not in Mα , which is a
contradiction.
Similar results can be obtained for intervals with fixed but unequal tails, in a completely
analogue manner.
Finally, the fact that the fiducial interval is the smallest interval in Mα leads to the
following proposition, in which the smallness is expressed in the more familiar terms of
the interval length Uα/2 (x) − Lα/2 (x).
Proposition 10. Among the intervals in Mα , the fiducial interval minimizes the expected length for all θ ∈ Θ as well as the length for all x ∈ X .
∗
Proof. For an interval (L∗α/2 (x), Uα/2
(x)) ∈ Mα to have shorter length than the fiducial
∗
interval (Lα/2 (x), Uα/2 (x)) it must hold that L∗α/2 (x) > Lα/2 (x) or Uα/2
(x) < Uα/2 (x).
By Proposition 9 neither condition can be fulfilled. Since the fiducial interval thereP
fore minimizes the length for each x, it also minimizes the expected length k Pθ (X =
k)(Uα/2 (k) − Lα/2 (k)).
Exact intervals and tests for discrete distributions
17
A consequence of Proposition 10 is that, in the class of equal-tailed two-sided tests
of θ = θ0 , the test that corresponds to the fiducial interval is admissible in the sense of
Cohen and Strawderman [12].
5. Conclusion
There exist a large number of methods for obtaining exact confidence intervals that are
shorter than the equal-tailed fiducial intervals. The use of such an interval comes at the
cost of losing control over the balance between the coverage levels of the corresponding
lower and upper confidence bounds. In many situations it is preferable to use an equaltailed interval, in order to guard equally against overestimation and underestimation
and not to bias the inference in some direction. The case for equal-tailed intervals is
further strengthened by the fact that strictly two-sided confidence intervals lack strict
nestedness. This causes difficulties with the interpretation of the intervals: what does it
mean that, for a particular x, the 92% interval equals the 95% interval? Which confidence
level should be reported for such an interval? More seriously, we have also seen that such
intervals may yield highly disparate conclusions for two indistinguishable models Pθ and
Pθ+ε . From a hypothesis testing perspective, this occurs when the null hypothesis θ0 is
changed slightly. From a confidence interval perspective, it can occur for small changes
in α, since the bounds of strictly two-sided intervals typically are discontinuous in α.
These problems have been pointed out for specific intervals in the past [4, 33]. We have
shown that they in fact are inherent to strictly two-sided confidence intervals.
The problems discussed in this paper arise also for strictly two-sided methods for
discrete distributions not covered by Definition 2. Examples include the hypergeometric
distribution and the joint distribution of two binomial proportions. We have restricted
our attention to the class of distributions given by Definition 2 in order to keep the proofs
reasonably short.
Strictly two-sided and equal-tailed confidence intervals are the most commonly used
types of two-sided confidence intervals. We have seen that strictly-two sided intervals lack
strict nestedness and that an extension of Bolshev’s theorem shows that the standard
fiducial intervals are the shortest equal-tailed exact intervals. While fiducial intervals
have been criticized for being overly conservative and too wide [1, 7, 8], the conclusion
of this paper is that they for practical purposes in fact are the optimal strictly nested
intervals.
Appendix A: Strictly two-sided tests
A.1. The likelihood ratio and Sterne tests
Let L(θ, x) = Pθ (X = x) be the likelihood function of Pθ . The likelihood ratio statistic is
TLR (θ0 , x) =
supθ∈Θ L(θ, x)
L(θ0 , x)
18
M. Thulin and S. Zwanzig
and the Sterne statistic is
TSt (θ0 , x) = 1/L(θ0 , x).
Both these statistics are minimized when θ0 is the maximum likelihood estimator of θ
given x. Thus for B1 to be satisfied it suffices that the maximum likelihood estimator of
θ is well-defined and strictly monotone in x. B2 is satisfied when there exists a θ such
that L(θ, x) is an asymmetric function of x. By definition, B3 is satisfied if there exists
an x0 such that L(θ, x) is increasing when x < x0 and decreasing when x > x0 . This is
guaranteed if Pθ has a monotone likelihood ratio.
The binomial, negative binomial and Poisson distributions all have well-defined and
strictly monotone maximum likelihood estimators and monotone likelihood ratios. Moreover, their probability functions are in general asymmetric in x. The likelihood ratio and
Sterne tests therefore satisfy conditions B1–B3 for these models.
A.2. The score test
Let U (θ, x) =
statistic is
∂
∂θ
ln L(θ, x) and let I(θ) be the Fisher information of Pθ . The score test
TSc (θ0 , x) =
(U (θ0 , x))2
.
I(θ0 )
If the maximum likelihood estimator of θ exists and is unique, then B1 is satisfied, with
θx being the maximum likelihood estimator of θ given x. B2 is satisfied if the distribution
of U (θ, x)2 is asymmetric for some θ. B3 is satisfied if there exists an x0 such that U (θ, x)2
is decreasing when x < x0 and increasing when x > x0 .
If Pθ is a regular exponential family with natural parameter θ, then U (θ, x) = x−Eθ (X)
and I(θ) = Varθ (X). B2 is satisfied if the distribution of X 2 is asymmetric for some θ and
B3 is satisfied since (x − Eθ (X))2 is convex in x. B1–B3 are therefore satisfied for the binomial, negative binomial and Poisson distributions, using the natural parametrizations.
These conditions are also satisfied for the most commonly used alternative parametrizations.
A.3. The Blaker test
The Blaker statistic is
TB (θ0 , x) ∝ (λT (θ0 , x))
−1
,
where λT (θ0 , x) is the p-value of a test with a rejection region that is the union of the
rejection regions of two one-sided level α/2 tests. The properties of T (θ0 , x) therefore
depend on the choice of λT (θ0 , x). A typical choice is the P
fiducial p-value (2).
Under A3 and A4, for any x ∈ X there exist θx such that k≤x−1 Pθx (X = k) < 1/2 and
P
P
P
k≥x Pθx (X =
k≤x Pθx (X = k) ≥ 1/2 and
k≥x+1 Pθx (X = k) < 1/2. Then we have
k) ≥ 1/2, so that λf (θ, x) = 1. Let Θ1 (x) denote the set of such θx .
19
Exact intervals and tests for discrete distributions
P
P
Pθx (X = k) < 1/2 but
Now, let y = x − 1. Then
k≥y+1 Pθx (X = k) =
k≤y−1
P
(X
=
k)
≥
1/2,
so
if
θ
∈
Θ
(x)
then
θ
∈
/
Θ
(y).
Similarly,
if we let y = x + 1,
P
x
1
x
1
Pk≥x θx
P
/ Θ1 (y). Thus,
k≤y−1 Pθx (X = k) =
k≤x Pθx (X = k) ≥ 1/2, so if θx ∈ Θ1 (x) then θx ∈
A3 and A4 are sufficient for B1 to hold for the Blaker statistic based on the fiducial
p-value.
B2 holds if the distribution of λT (θ, x) is asymmetric in x for some θ. For λf (θ, x) this
holds if Pθ (X = x) is asymmetric as a function of x for some θ.
Finally, B3 is satisfied since the monotonicity of Pθ (X ≤ x) in x implies that λT (θ, x)
is a bimonotone function of x. B1–B3 are therefore satisfied for the binomial, negative
binomial and Poisson distributions.
Appendix B: Proofs
B.1. Proof of Proposition 1
If Pθ is a discrete one-parameter exponential family in natural form, for (θ, x) ∈ Θ × X
its probability function can be written as
pθ (x) = Pθ (X = x) = exp{θT (x) − K(θ)}h(x),
(B.1)
where T : X 7→ R is a function that does not depend on θ and K : Θ 7→ R is infinitely
often differentiable in Θ since Pθ is regular [23], Theorem 1.17. θ, T (x) and K(θ) are all
finite for (θ, x) ∈ Θ × X , and thus (B.1) is strictly positive when (θ, x) ∈ Θ × X , yielding
A1. Moreover, A3 follows from the fact that when x is fixed (B.1) is differentiable in θ
since θT (x) and K(θ) are infinitely differentiable.
To see that an increasing likelihood ratio implies A2, let ℓ(x) = pθ2 (x)/pθ1 (x) for θ1 < θ2
in Θ. The likelihood ratio ℓ(x) is increasing in x. Let
X
Fθ (x) = Pθ (X ≤ x) =
Pθ (X = k)
k≤x
and
Gθ (x) = Pθ (X > x) =
X
Pθ (X = k).
k>x
We consider the cases when ℓ(x) ≤ 1 and ℓ(x) ≥ 1 separately.
If ℓ(t) ≤ 1 then for s < t, Fθ2 (s) ≤ Fθ1 (s) since pθ2 (x) ≤ pθ1 (x) for all x < t.
If ℓ(t) ≥ 1 then pθ2 (x) ≥ pθ1 (x) when x > t and for s > t, Gθ2 (s) ≥ Gθ1 (s). Since Fθ (x) =
1 − Gθ (x), it follows that Fθ2 (s) ≤ Fθ1 (s).
B.2. Proof of Proposition 2
First, assume that k ≥ xθ . Then by B3 T (θ, ·) is increasing at k. There are two possible
scenarios:
20
M. Thulin and S. Zwanzig
(i) k ≥ k1 (θ, x): By definition, T (θ, k1 (θ, x)) ≥ T (θ, x). Since T (θ, ·) is increasing for
x ≥ xθ it follows that T (θ, k) ≥ T (θ, k1 (θ, x)) ≥ T (θ, x), meaning that k ∈ Aθ,x .
(ii) k < k1 (θ, x): it follows from the definition of k1 (θ, x) that T (θ, k) < T (θ, x), so
k∈
/ Aθ,x .
In summary, if k ≥ xθ then k ∈ Aθ,x if and only if k ≥ k1 (θ, x). An analogous argument
shows that if k ≤ xθ then k ∈ Aθ,x if and only if k ≤ k2 (θ, x), and the first part of the
proposition follows.
To see that at least one of k1 (θ, x) and k2 (θ, x) is non-constant in θ, note that by B1,
for any pair (x, y) ∈ X 2 there exist (θx , θy ) ∈ Θ2 such that x ∈
/ Aθx ,y but x ∈ Aθy ,y . The
set Aθ,x is therefore not constant in θ, and thus at least one of k1 (θ, x) and k2 (θ, x) must
be non-constant in θ.
B.3. Proof of Proposition 3
(a) and (b) were proved in Section 2.3. We will now prove (c). Let Lα (x) and Uα (x)
denote the lower and upper bounds of the interval. We will show that for any x ∈ X
there exists an α0 ∈ (0, 1) such that Lα (x) and Uα (x) simultaneously are constant in a
neighbourhood of α0 , so that the confidence interval is not strictly nested.
We introduce the mutually disjoint sets
Θ1 (x) := {θ : λ(θ, x) = 1},
Θℓ (x) := {θ : θ ≤ inf Θ1 (x)}
and
(B.2)
Θu (x) := {θ : θ ≥ sup Θ1 (x)},
which are such that Θℓ (x) ∪ Θ1 (x) ∪ Θu (x) = Θ. We also define
Θα (x) := {θ : λ(θ, x) > α}.
By condition B1, given x ∈ X there exists θx ∈ Θ such that
λ(θx , x) = Pθx (T (θx , X) ≥ T (θx , x)) = 1 − Pθx (T (θx , X) < T (θx , x)) = 1,
so Θ1 (x) is non-empty.
First, we investigate the behaviour of the bounds when either Θℓ (x) or Θu (x) is empty.
Let Θ̄α (x) be the closure of Θα (x). Since Θ1 (x) ⊆ Θ̄α (x) for all α ∈ (0, 1),
sup Θ1 (x) ∈ Θ̄α (x)
and
inf Θ1 (x) ∈ Θ̄α (x).
If inf Θ1 (x) = inf Θ, then Θℓ (x) = {θ : θ < inf Θ} = ∅. Then
Lα (x) = inf Θα (x) ≤ inf Θ1 (x) = inf Θ ≤ inf Θα (x),
so Lα (x) = inf Θ for all α ∈ (0, 1). Similarly, if sup Θ1 (x) = sup Θ then Θu (x) = {θ : θ >
sup Θ} = ∅, and
Uα (x) = sup Θα (x) ≥ sup Θ1 (x) = sup Θ ≥ sup Θα (x),
21
Exact intervals and tests for discrete distributions
so Uα (x) = sup Θ for all α ∈ (0, 1). Thus, when Θℓ (x) is empty Lα (x) is constant and
when Θu (x) is empty Uα (x) is constant. In this case, whether or not the interval is strictly
nested therefore depends on whether there exists an α ∈ (0, 1) such that the other bound
is constant in a neighbourhood of α. We will therefore without loss of generality assume
that neither Θℓ (x) nor Θu (x) are empty.
Let
αℓ =
lim inf
θ→inf Θ1 (x)
λ(θ, x)
and αu =
lim inf
θ→sup Θ1 (x)
λ(θ, x).
Since Aθ,x 6= X for θ < inf Θ1 (x), by A1 αℓ < 1, and similarly αu < 1. Thus, a point is
added to or removed from Aθ,x at θ = inf Θ1 (x) and θ = sup Θ1 (x), and by Lemma 1 the
p-value function λ(θ, x) must have jumps at θ = inf Θ1 (x) and at θ = sup Θ1 (x). Then
for α0 = max(αℓ , αu ), there is an α1 ∈ (α0 , 1) for which there exists δ > 0 such that
{θ ∈ Θ : λ(θ, x) ∈ (α1 − δ, α1 + δ)} = ∅.
Thus both the upper and the lower bound of Iα (x) are constant in a neighbourhood of
α = α1 , and the interval is not strictly nested.
B.4. An auxiliary lemma
The following auxiliary lemma will be used in the proof of Proposition 4.
Lemma 2. With X as in Proposition 4, for any y, x ∈ X such that 0 ≤ y < x, let
θy,x = inf{θ : T (θ, x) ≤ T (θ, y)}
(B.3)
and define θx,x := inf Θ1 (x). Under A2+ and B3+ ,
θ0,x ≤ θ1,x ≤ · · · ≤ θx−1,x ≤ θx,x .
(B.4)
Moreover,
Aθ,x = {k ∈ X : k ≥ x}
if and only if
θ ∈ (inf Θ, θ0,x),
(B.5)
and
Aθ,x = {k ∈ X : k ≤ y} ∪ {k ∈ X : k ≥ x}
if and only if
θ ∈ [θy,x , θy+1,x ).
(B.6)
Proof. First, we establish some facts about θy,x and the behaviour of T (θ, ·) for such
θ. If θ ∈ Θ1 (x), then it follows from (3) that T (θ, x) ≤ T (θ, y). Thus, by (B.3) we have
θy,x ≤ inf Θ1 (x), so by (B.2), θy,x ∈ Θℓ (x).
With xθ as defined in B3+ (i), T (θ, ·) is increasing at x if x > xθ . If θ ∈ Θ1 (x) then
xθ = x. By B3+ (iii), xθ is an increasing function of θ. Thus, if θ ≤ inf Θ1 (x), i.e. θ ∈ Θℓ (x),
we have x > xθ . Since θy,x ∈ Θℓ (x), T (θy,x , ·) is increasing at x.
22
M. Thulin and S. Zwanzig
It now follows that for any y < x, T (θ, x) < T (θ, y) can happen only if T (θ, ·) is
decreasing at y. Whenever T (θ, x) < T (θ, y) and T (θ, ·) is decreasing at y, we have
T (θ, x) < T (θ, y) < T (θ, y − 1), and (B.4) follows since {θ : T (θ, x) ≤ T (θ, y)} ⊂ {θ :
T (θ, x) ≤ T (θ, y − 1)}.
Let x ≥ 1 be fixed. If θ ≤ θ0,x then T (θ, x) < T (θ, 0) and (B.4) ensures that T (θ, x) <
T (θ, y) for all other y < x as well. (B.5) now follows from (3).
Next, for some y < x, let θ ∈ Θℓ (x) be such that θ > θy,x . Under B3+ (ii) we have
θ < θx ∈ Θ1 (x), so T (·, x) is decreasing in Θℓ (x). However, if T (θ, x) < T (θ, y) then y <
xθ , so θ ≥ sup Θ1 (y). Thus θx,y > θy , implying that T (·, y) is increasing at θy,x . Thus
T (θ, x) < T (θ, y) for all θ > θy,x . Equations (B.5) and (B.6) now follow from (3).
B.5. Proof of Proposition 4
We start by showing (b) and finish by proving (a). The proof of (c) is analogous to the
proof of (b), and is therefore omitted.
(b) We wish to find the largest αL such that, for all x ∈ X , Lα (x) is strictly monotone
in α when α < αL . For a given x, let αL (x) be the largest α such that Lα (x) is strictly
monotone in α when α < αL (x). Then αL ≤ αL (x) for all x, with equality for some x. We
therefore show the statement by showing that αL (x) = inf θ∈{θ:T (θ,0)>T (θ,x)} λ(θ, x) > 0.
As in the proof of Lemma 2, it suffices to study θ ∈ Θℓ (x), where Θℓ (x) is defined as
in (B.2).
(i) Let x ≥ 1 be fixed. If θ ≤ θ0,x , defined as in (B.3), then by (B.5), Aθ,x = {k : k ≥ x}.
Thus, by (3), λ(θ, x) = Pθ (X ≥ x). The p-value function λ(·, x) is therefore non-negative
(by A1), yielding (i).
(ii) λ(·, x) is strictly increasing (by A2+ ) and continuous (by A3). We extend the
p-value function by defining λ(inf Θ, x) := limθցinf Θ λ(θ, x). Then λ(θ, x) is a continuous strictly monotone bijection from the compact set [inf Θ, θ0,x ] to the compact set
[λ(inf Θ, x), αL (x)]. It is therefore a homeomorphism, and it follows that its inverse Lα (x)
is continuous and strictly monotone in α, which yields (ii).
(iii) From Lemmas 2 and 1, it follows that the first discontinuity in λ(θ, x) occurs at
θ0,x . From Definition 4, it follows that there exists an ε > 0 such that Lα (x) is constant
in (αL (x), αL (x) + ε) if and only if λ(θ, x) > αL (x) for all θ ∈ [θ0,x , inf Θ1 (x)].
By (B.6) for any θ ∈ [θ0,x , inf Θ1 (x)], there exists a y < x such that θ ∈ [θy,x , θy+1,x ),
so that
λ(θ, x) = Pθ (X ≤ y) + Pθ (X ≥ x) > ε + Pθ (X ≥ x)
(B.7)
> ε + Pθ0,x (X ≥ x) = ε + αL (x) > αL (x),
where the first inequality follows from A1 and the second inequality follows from A2+ .
(iii) now follows.
23
Exact intervals and tests for discrete distributions
(a) For any x ∈ X , let AL (x) denote the set of α for which Lα (x) is locally constant in
α and AU (x) denote the set of α for which Uα (x) is locally constant in α. By definition,
the largest α for which Iα (X) is strictly nested is
[
αnest = inf α : ∃ε > 0 for which (α, α + ε) ⊆
(AL (x) ∩ AU (x)) .
x∈X
Using Proposition 3(c), AL (x) ∩ AU (x) has a connected uncountable subset for all x, so
αnest always exists. By part (b) of Proposition 4, αnest ≥ αL > 0.
B.6. Proof of Proposition 6
By Proposition 3(a) and A3, λ(θ, x) is a piecewise continuous function. By Lemma 1, it is
not continuous at the boundaries of the set Θ1 (x). Hence λ(θ, x) can only be bimonotone
if it is monotone whenever it is continuous. Each of its continuous parts can be represented
by equation (6) with fixed k1 and k2 . Such a part can be written as
X
X
1−
Pθ (X = k).
(B.8)
Pθ (X = k) +
k≤k1 −1
k≤k2
P
By A2, 1 − k≤k1 −1 Pθ (X = k) is strictly increasing and k≤k2 Pθ (X = k) is strictly
decreasing. By condition A4 (B.8) equals 1 at the boundaries of Θ. If it is not constant,
it must therefore by condition A5 have a unique minimum in the interior of Θ. Rewriting
the expression again, we have
X
Pθ (X = k),
(B.8) = 1 −
P
k2 +1≤k≤k1 −1
so that the minimum is given by the root θr of the equation
d
dθ
kX
1 −1
k=k2 +1
Pθ (X = k) =
kX
1 −1
k=k2 +1
d
Pθ (X = k) = 0
dθ
(B.9)
that is in the interior of Θ. Next, we let k1 and k2 vary as functions of (θ, x) and use
θr (θ, x) to denote the solution of (B.9) with k1 = k1 (θ, x) and k2 = k2 (θ, x).
By Proposition 3(a), λ(θ, x) has jumps corresponding to changes in k1 (θ, x) or k2 (θ, x).
λ(θ, x) fails to be bimonotone if
(k1 (θ, x), k2 (θ, x)) = (k1 (θr (θ, x) + ε, x), k2 (θr (θ, x) + ε, x))
for some ε > 0,
i.e. if it does not jump before the root θr (θ, x) that corresponds to (k1 (θ, x), k2 (θ, x)),
d
λ(θ, x),
since for λ′ (θ, x) = dθ
sign(λ′ (θr (θ, x) − ε)) 6= sign(λ′ (θr (θ, x) + ε)).
24
M. Thulin and S. Zwanzig
Assume that θ ∈ Θu (x). Then λ(θ, x) should be decreasing in θ. If θ > θr (θ, x) then
(B.8) with k1 = k1 (θ, x) and k2 = k2 (θ, x) is increasing, so that λ(θ, x) is not bimonotone.
If instead we assume that θ ∈ Θℓ (x) so that λ(θ, x) is in its increasing part, we similarly
get that λ(θ, x) is not bimonotone if θ < θr (θ, x). This establishes (a). Part (b) then
follows from Proposition 5.
B.7. Proof of Proposition 8
Let R be the extended real line and Θ ⊆ R be the closure of Θ. Let F (θ, x) = Pθ (X ≤ x).
By A1–A4, F (·, x) is a continuous monotone bijection from Θ to [0, 1] for all x ∈ X \sup X .
Since Θ and [0, 1] both are compact, it follows that F (·, x) is a homeomorphism, which
ensures that the bounds given by (1) are continuous in α. The monotonicity of Fθ (·, x)
ensures that both Fθ−1 (·, x) and the bounds are monotone, so that the interval is strictly
nested.
Finally, by condition A3, the p-value function (2) is continuous in θ when x ∈ X is
fixed.
Acknowledgments
The authors wish to thank the Editor and the reviewers for comments that helped improve the paper.
References
[1] Agresti, A. (2003). Dealing with discreteness: Making “exact” confidence intervals for
proportions, differences of proportions, and odds ratios more exact. Stat. Methods Med.
Res. 12 3–21. MR1977232
[2] Agresti, A. and Min, Y. (2001). On small-sample confidence intervals for parameters in
discrete distributions. Biometrics 57 963–971. MR1863460
[3] Birnbaum, A. (1961). Confidence curves: An omnibus technique for estimation and testing
statistical hypotheses. J. Amer. Statist. Assoc. 56 246–249. MR0121904
[4] Blaker, H. (2000). Confidence curves and improved exact confidence intervals for discrete
distributions. Canad. J. Statist. 28 783–798. MR1821434
[5] Blyth, C.R. and Still, H.A. (1983). Binomial confidence intervals. J. Amer. Statist.
Assoc. 78 108–116. MR0696854
[6] Bol’šev, L.N. (1965). On the construction of confidence limits. Teor. Verojatnost. i Primenen. 10 187–192. MR0210218
[7] Brown, L.D., Cai, T.T. and DasGupta, A. (2001). Interval estimation for a binomial
proportion. Statist. Sci. 16 101–133. MR1861069
[8] Byrne, J. and Kabaila, P. (2005). Comparison of Poisson confidence intervals. Comm.
Statist. Theory Methods 34 545–556. MR2181200
[9] Casella, G. (1986). Refining binomial confidence intervals. Canad. J. Statist. 14 113–129.
MR0849867
Exact intervals and tests for discrete distributions
25
[10] Casella, G. and Robert, C. (1989). Refining Poisson confidence intervals. Canad. J.
Statist. 17 45–57. MR1014090
[11] Clopper, C.J. and Pearson, E.S. (1934). The use of confidence or fiducial limits illustrated in the case of the binomial. Biometrika 26 404–413.
[12] Cohen, A. and Strawderman, W.E. (1973). Admissibility implications for different criteria in confidence estimation. Ann. Statist. 1 363–366. MR0348875
[13] Crow, E.L. (1956). Confidence intervals for a proportion. Biometrika 43 423–435.
MR0093077
[14] Crow, E.L. and Gardner, R.S. (1959). Confidence intervals for the expectation of a
Poisson variable. Biometrika 46 441–453. MR0125681
[15] Fay, M.P. (2010). Confidence intervals that match Fisher’s exact or Blaker’s exact tests.
Biostatistics 11 373–374.
[16] Fay, M.P. (2010). Two-sided exact tests and matching confidence intervals for discrete
data. R Journal 2 53–58.
[17] Fisher, R.A. (1930). Inverse probability. Proc. Camb. Philos. Soc. 26 528–535.
[18] Garwood, F. (1936). Fiducial limits for the Poisson distribution. Biometrika 28 437–442.
[19] Göb, R. and Lurz, K. (2014). Design and analysis of shortest two-sided confidence intervals
for a probability under prior information. Metrika 77 389–413. MR3175131
[20] Hirji, K.F. (2006). Exact Analysis of Discrete Data. Boca Raton, FL: Chapman &
Hall/CRC. MR2193238
[21] Kabaila, P. and Byrne, J. (2001). Exact short Poisson confidence intervals. Canad. J.
Statist. 29 99–106. MR1834489
[22] Lecoutre, B. and Poitevineau, J. (2014). New results for computing Blaker’s exact
confidence interval for one parameter discrete distributions. Comm. Statist. Simulation
Comput. DOI:10.1080/03610918.2014.911900.
[23] Liese, F. and Miescke, K.-J. (2008). Statistical Decision Theory: Estimation, Testing,
and Selection. Springer Series in Statistics. New York: Springer. MR2421720
[24] Newcombe, R.G. (2011). Measures of location for confidence intervals for proportions.
Comm. Statist. Theory Methods 40 1743–1767. MR2781501
[25] Reiczigel, J. (2003). Confidence intervals for the binomial parameter: Some new considerations. Stat. Med. 22 611–621.
[26] Schilling, M.F. and Doi, J.A. (2014). A coverage probability approach to finding an
optimal binomial confidence procedure. Amer. Statist. 68 133–145. MR3246552
[27] Sommerville, M.C. and Brown, R.S. (2013). Exact likelihood ratio and score confidence
intervals for the binomial proportion. Pharmaceutical Statistics 12 120–128.
[28] Sterne, T.E. (1954). Some remarks on confidence or fiducial limits. Biometrika 41 275–
278. MR0062387
[29] Thulin, M. (2014). On split sample and randomized confidence intervals for binomial
proportions. Statist. Probab. Lett. 92 65–71. MR3230474
[30] Thulin, M. (2014). Coverage-adjusted confidence intervals for a binomial proportion.
Scand. J. Stat. 41 291–300. MR3207171
[31] Thulin, M. (2014). The cost of using exact confidence intervals for a binomial proportion.
Electron. J. Stat. 8 817–840. MR3217790
[32] Vos, P.W. and Hudson, S. (2005). Evaluation criteria for discrete confidence intervals:
Beyond coverage and length. Amer. Statist. 59 137–142. MR2133560
[33] Vos, P.W. and Hudson, S. (2008). Problems with binomial two-sided tests and the associated confidence intervals. Aust. N. Z. J. Stat. 50 81–89. MR2414657
26
M. Thulin and S. Zwanzig
[34] Wang, W. (2006). Smallest confidence intervals for one binomial proportion. J. Statist.
Plann. Inference 136 4293–4306. MR2323417
[35] Wang, W. (2014). Exact optimal confidence intervals for hypergeometric parameters. J.
Amer. Statist. Assoc. To appear. DOI:10.1080/01621459.2014.966191.
[36] Wang, Y.H. (2000). Fiducial intervals: What are they? Amer. Statist. 54 105–111.
MR1803120
[37] Xie, M.-g. and Singh, K. (2013). Confidence distribution, the frequentist distribution
estimator of a parameter: A review. Int. Stat. Rev. 81 3–39. MR3047496
Received December 2014 and revised March 2015
| 10math.ST
|
arXiv:1207.1735v1 [math.NT] 6 Jul 2012
A few conjectures about the multiple zeta values
German Combariza
July 10, 2012
Abstract
The multiple zeta values (MZV) are a set of real numbers with a
beautiful structure as an algebra over the rational numbers. They are
related to maybe the most important conjecture on mathematics today,
the Riemann hypothesis. In this paper we will show partial solutions to
five well known conjectures about the MZV which we partially proved
using just linear systems and shuffles products. We saw that to partially
solve four of this conjectures we only need linear algebra.
1
Preliminary
Let s = (s1 , s2 , · · · , sk ) be a tuple of natural numbers with si > 0 and s1 > 1.
Consider the real numbers also known as the multiple zeta values (MZV)
X
1
ζ(s) =
.
s1 s2
n n · · · nskk
n >n >···>n >0 1 2
1
2
k
Our main interest is the problem concerning the polynomial relations over Q of
the MZV. For MZV of length 1 and even degree there is a beautiful formula due
to Euler
(2πi)s Bs
for s = 2, 4, 6, · · ·
ζ(s) = −
2s!
which express the values of the zeta functions at even points in terms of the
Bernoulli numbers Bs ∈ Q defined by the generating function
∞
t X ts
t
=
1
−
+
Bs
Bs = 0 for odd s ≥ 3.
et − 1
2 s=2
s!
Therefore the transcendence degree of the ring Q[ζ(2), ζ(4), · · · ] over Q is 1.
Much less is known on the arithmetic nature of values of the zeta function at
odd integers ζ(3), ζ(5), · · · . We believe that each one of these numbers is transcendent, we only know by Apery’s theorem that ζ(3) is an irrational number.
On the other hand for the multi-index MZV there are more relations explained
by a rich algebraic structure over an associate algebra which surjects the algebra
of the MZV over Q. For example for the MZV ζ(2, 1) Euler proved the identity
ζ(2, 1) = ζ(3).
1
The big problem is: Find all the relations between the MZV. Apparently all the
relations are explained by the shuffle and stuffle products and maybe the main
conjecture about it is the Zagier’s conjecture which determine the dimension
of the Q-algebra spam by the MZV of a fixed degree over Q. We will start
this paper by defining the shuffle and stuffle products in section 2. We will
also state the Zagier’s conjecture. The goal of section 3 is explain the KanekoNoro-Tsurumaki conjecture [2]. In section 4 we will explain the Petitot-Minh
approach to the Zagier’s conjecture using the xtaylor algorithm. In section 5
we will talk about the transcendence degree over Q and the 2-3 conjecture.
Finally in section 6 the Broadhurst-Kreimer conjecture and decomposability of
the MZV. We started this project inspired by the Petitot-Minh paper [4] and
we will follow its notation and definitions.
2
A source of relations
Consider the alphabet X = x0 , x1 and encode the multi-index s = (s1 , · · · , sk )
by the rule
s = (s1 , · · · , sk ) 7→ xs01 −1 x1 xs02 −1 x1 · · · x0sk −1 x1 .
(1)
Then the degree of the MZV ζ(s) is by definition the degree of the monomial
on the variables x0 , x1 , i.e. s1 + s2 + · · · sk . Let H1 and H2 , be the sub-algebras
Q ⊕ QhXix1 and Q ⊕ x0 QhXix1 , respectively. Let us define the the shuffle
product
and the stuffle or second shuffle product ∗ over QhXi by the rules
1
w =w1= w
1 ∗ w = w ∗ 1 = w.
and
xi u xj v
yi u ∗ yj v
= xi (u xj v) + xj (xi u v)
= yi (u ∗ yj v) + yj (yi u ∗ v) + yi+j (u ∗ v)
(2)
(3)
Inductive arguments enable us to prove that each of the above products is
commutative and associative. The motivation for this products are explained
in the following examples.
Example 1 (The Stuffle product). Using Σ notation, the MZV ζ(2)2 can be
expressed as
!
!
!
X 1
X 1
X
X
X
1
2
ζ(2) =
=
+
+
2
2
2 n2
m
n
m
m=n
m>0
n>0
m>n>0
n>m>0
This is
ζ(2)2 = 2ζ(2, 2) + ζ(4).
Using stuffle notation, should be clear that
y2 ∗ y2 = y2 (1 ∗ y2 ) + y2 (y2 ∗ 1) + y4 = 2y2 y2 + y4 .
2
The MZV can also be expressed as multiple integrals.
Z
Z
ζ(s1 , s2 , · · · , sk ) = · · ·
w1 (t1 )w2 (t2 ) · · · wp (tp ).
1>t1 >···>tp >0
with p = s1 + s2 + · · · sk and wi (t) = dt/(1 − t) if i ∈ {s1 , s1 + s2 , · · · , s1 + s2 +
· · · + sk } and wi (t) = dt/t otherwise.
Example 2 (The Shuffle product). Using the integral notation above for the
MZV ζ(2)2 we have
Z
Z
dt1 dt2
dr1 dr2
ζ(2)2 =
0>t1 >t2 >1 t1 (1 − t2 )
0>r1 >r2 >1 r1 (1 − r2 )
simplifying
Z
4
0>t1 >t2 >t3 >t4 >1
Z
dt1 dt2 dt3 dt4
dt1 dt2 dt3 dt4
+2
t1 t2 (1 − t3 )(1 − t4 )
0>t1 >t2 >t3 >t4 >1 t1 t3 (1 − t2 )(1 − t4 )
i.e.
ζ(2)2 = 4ζ(3, 1) + 2ζ(2, 2).
Using the shuffle product the same identity is
x0 x1
x x
0 1
2
ζ(2)
x x ) + x (x x x ) = 2x (x x x )
2x [x (1 x x ) + x (x x )] = 2x x x x + 4x x
= x0 (x1
0 1
0
=
0 1
0 1
= 2ζ(2, 2) + 4ζ(3, 1).
0 1
0
1
1
1
0
1
0 1
0 1 0 1
2 2
0 1
Consider the function ζ as the Q-linear map ζ : H2 → R induce by:
X
ζ : xs1 −1 yxs2 −1 y · · · xsk −1 y 7→ ζ(s) =
n1 >n2 >···>nk
1
.
s1 s2
n
n
· · · nskk
>0 1 2
Theorem 3. Under the shuffle product, the ζ map H2 → R is an homomorphism.
ζ(w v) = ζ(w)ζ(v)
The same is true for the stuffle product. For the proof of this theorems we
refer the reader to [4]
Theorem 4. Under the stuffle product, the ζ map H2 → R is an homomorphism.
ζ(w ∗ v) = ζ(w)ζ(v)
Corollary 5. For any two words w, w′ in H2
ζ(w
w − w ∗ w ) = 0.
′
′
3
This beautiful corollary is the main source of relations for the multiple zeta
values and our object of study is the ideal generated by this difference. Denote
ζp the Q-algebra form by all the MZV of weight p, i.e. ζp = Q[ζ(s)] for all s
with s1 + · · · + sk = n. Denote ζp+ the Q-algebra spam by all the the MZV of
weight less or equal p, i.e ζp+ = ζp ⊕ ζp−1 ⊕ · · · ⊕ ζ2 .
Definition 6. The weight of the MZV ζ(s1 , s2 , · · · , sk ) is the natural number
s1 + s2 + · · · sk .
Conjecture 7 (Zagier). Let dp be the Q-dimension of ζp . Then these are given
by the recurrence d1 = 0, d2 = d3 = 1 and
dn = dn−2 + dn−3 for n ≥ 4.
Here are some examples of this conjecture.
• p = 2, d2 = 1, ζ2 = Q[ζ(2)].
• p = 3, d3 = 1, ζ3 = Q[ζ(3)].
• p = 4, d4 = 1, ζ4 = Q[ζ(2)2 ].
• p = 5, d5 = 2, ζ5 = Q[ζ(2)ζ(3), ζ(5)].
• p = 6, d6 = 2, ζ6 = Q[ζ(2)3 , ζ(3)2 ].
• p = 7, d7 = 3, ζ7 = Q[ζ(7), ζ(2)ζ(5), ζ(2)2 ζ(3)].
• p = 8, d8 = 4, ζ8 = Q[ζ(2)4 , ζ(3)ζ(5), ζ(2)ζ(3)2 , ζ(6, 2)].
• p = 9, d9 = 5, ζ9 = Q[ζ(9), ζ(2)ζ(7), ζ(2)2 ζ(5), ζ(2)3 ζ(3), ζ(3)3 ].
4
3
Kaneko-Noro-Tsurumaki conjecture and the
regularization process
So far we have a nice way to generate relations between the MZV with the
“shuffle - stuttle” process. The question now is the domain for this relations.
One can try
w w′ − w ∗ w′
′
2
for every w, w ∈ H but this is going to produce more variables than relations.
To solve this problem Kaneko-Noro-Tsurumaki introduce the notion of regularization. Below is a brief introduction to the subject. To view this in detail see
[2].
It is known that (H1 , ) is a commutative algebra and isomorphic to the
polynomial algebra over H2 with one variable, i.e.
H1 ≃ H2 [x1 ].
Let reg : H1 → H2 be the map “constant term” from the above isomorphism
of shuffle algebras.
reg :
n
X
wi
x
i
1
7→ w0 .
(4)
i=0
Example 8.
H1 ≃
x1 x0 x1 =
H2 [x1 ]
−2x0 x21 + x0 x1
x
1
7 H2
→
7→ −2x0 x21
This process will reduce the number of variables to elements only in H2 after
the next theorem.
Theorem 9 (From [2]).
ζ(reg(w1
w
0
− w1 ∗ w0 )) = 0
For any w0 ∈ H2 and w1 ∈ H1 .
Given an integer n let w, w′ ∈ H1 such that degree(w) + degree(w′ ) = n.
Then the regularization process is giving a linear system of relations of the MZV
with variables the monomials in H2 of degree n. The number of variables is 2n−1 .
This works, but we have limitations of space then there is a new problem: Too
many equations. This is where the Kaneko-Noro-Tsurumaki conjectures plays
its role.
Conjecture 10 (Kaneko, Noro, Tsurumaki). In order to have all MZV relations
of degree n it is enough to take w1 ∈ H2 and w0 ∈ {x1 , x0 x1 , x20 x1 , x0 x21 } with
degree(w0 ) + degree(w1 ) = n.
For n > 7 the Kaneko-Noro-Tsurumaki linear system has 2n−1 relations and
2
equations which null space is generated by the MZV relations of degree n.
We verify this dimension with the Zagier conjecture. The files and the process
on this conjecture is at the end of this paper.
n−1
5
4
Petitot-Minh approach and Zagier conjecture
In this section we will try to explain how Petitot and Minh approached to the
Zagier’s conjecture using the xtaylor algorithms. We will also discuss their
conjecture about the algebraic structure of the algebra of the MZV over the
rationals. More details about this can be found at [4]. This papers was the first
motivation for this work and we follow its ideas and notation.
Definition 11. A word l ∈ Qhx0 , x1 i on the letters x0 , x1 is a Lyndon word if
it is less that any of its proper right factors in the lexicographic order.
l = uv ⇒ l < v.
The Lyndon words plays an important role in the associative algebra (Qhx0 , x1 i,
and determine a basis for the calculations of the MZV relations. A few examples
of Lyndon are x0 , x1 , x0 x1 , x20 x1 , x0 x31 , x0 x1 x0 x21 and examples of non Lyndon
words are x0 x1 x0 x1 , x0 x1 x30 x1 . Before state the Minh-Petitot conjecture we will
see that the set of Lyndon words form a basis for Qhx0 , x1 i and we will express
any words in terms of this basis.
Theorem 12. The set L of all Lyndon words over x0 , x1 is a polynomial basis
for the commutative shuffle algebra Qhx0 , x1 i.
Qhx, yi ≃ Q[L] .
For the proof of this theorem and more detail we reefer the reader to [4].
This the idea they followed. If one is able to express any word in H1 in term of
Lyndon words, the variables in the systems of MZV will be reduce to the Lyndon
words. To do that we need a couple of definitions that make sense the idea of a
differential and the will use the same idea behind the Taylor algorithms.
Definition 13 (The bracket). For any Lyndon word l, the bracketed form [l],
is defined as follows:
[l] = [[u], [v]]
with [x0 ] = x0 ,[x1 ] = x1 and [x0 , x1 ] = x0 x1 − x1 x0 and for l = uv, where u, v
are Lyndon words and v is the longest Lyndon word such that l = uv.
Definition 14 (The right residual). The right residual (p⊲q) of p by q is defined
by: (p ⊲ q|w) = (p|qw) for every word w. Where (u|v) = δvu .
We are now ready to explain the xtaylor algorithms. This algorithms appeared first at Petitot’s doctoral thesis. This are the steps that we will follow.
First we need a derivation then find a set of possible divisor. At this stage
we improve the xtaylor algorithms as appears in [4] by looking to a small set
of factors for ζ(s1 , s2 , · · · , sk ) to all sub-sequence of s1 , s2 , · · · , sk that are also
Lyndon words i.e. {ζ(s1 ), ζ(s1 , s2 , · · · , ζ(s1 , · · · , sk−1 ))∩ Lyndon. Once a divisor in found one proceed to do the same process for the residuals we will see
below. This process follow the idea of the Taylor algorithms the only missing
concept here was the differential. To those elements whose differentiation by
them is not zero we are calling them here Lyndon divisors.
6
)
Lemma 15 (2.1 [4]). The operation − ✄ [−] is a differential for the shuffle
product.
We illustrate the algorithms with this example. ζ(2, 4) is not a Lyndon
word. The set of Lyndon divisors is just {ζ(2)}. The Taylor process follow
using ζ(2, 4) − ζ(2, 4) ✄ [ζ(2)] instead. The final stage is
ζ(2, 4) = 4ζ(5, 1) + 2ζ(4, 2) − ζ(3)
ζ(3) + ζ(2) ζ(4).
We are now ready to see the Minh-Petitot conjecture and their explanation.
Lemma 16 (4.1 [4]). Let R = k[X1 , X2 , · · · , Xn ] be a polynomial ring; I be an
ideal of R and G ⊂ R be a Gröbner. basis of I for any admissible order. If
the leading terms of polynomials of G are single indeterminates then the algebra
R/I is free and the ideal I is prime.
Conjecture 17. The Q-algebra of the MZV is a polynomial algebra.
Example 18. ζ≤6 = Q[ζ(2), ζ(3), ζ(5)].
5
Transcendence degree of ζp over Q and the 2-3
conjecture
The algebra ζ6 of MZV of degree 6 or lower over Q is generated for the MZV
ζ(2), ζ(3), ζ(5). Therefore it’s transcendence degree is 3. The natural question
is to generalize this degree for higher algebras of MZV. Here another application
of the Lyndon words will help to state a conjecture about it.
Conjecture 19. The set of MZV ζ(s1 , s2 , · · · , sk ) with k ≥ 1 and sj ∈ {2, 3}
for 1 ≤ j ≤ k, such that s1 , s2 , · · · , sk is a Lyndon word on the alphabet {2, 3},
give a transcendence basis of the Q-algebra of the MZV.
If this conjecture is true we can tell what is going to be the transcendence
degree if each step of this tower of Q-algebras.
Corollary 20. The number N (p) of elements of weight p in a transcendence
basis of the M ZV is given by
N (p) =
1X
µ(p/l)Pl
p
l|p
where Pl = Pl−2 + Pl−3 and P1 = 0, P2 = 2, P3 = 3.
This is the conjecture exemplified.
7
Degree
p = 2;
p = 3;
p = 4;
p = 5;
p = 6;
p = 7;
p = 8;
p = 9;
Lyndon
x;
y;
MZV
ζ(2)
ζ(3)
xy;
ζ(2, 3)
x2 y;
xy 2 ;
x3 y;
ζ(2, 2, 3)
ζ(2, 3, 3)
ζ(2, 2, 2, 3)
In this context the algebra ζ6 is just the polynomial algebra Q[ζ(2), ζ(3), ζ(2, 3)].
6
Deep vs weight and the Broadhurst-Kreimer
conjecture
This last conjecture is about the “decomposability” of the MZV. For example
ζ(2, 3) can be write in term of the MZV ζ(2), ζ(3) and ζ(5) this are MZV of
deep 1 but the MZV of depth 2 ζ(6, 2) can not be decompose in MZV of depth
1.
Definition 21. The depth of the MZV s1 , s2 , · · · , sk is k.
The problem here is what and where. What MZV are not decomposable in
term of of lower depth and when can we find them.
Conjecture 22. The number Dn,k of MZV of weight n and depth k that are
not reducible to MZV of lesser depth is generated by
1−
YY
x12 y 2 (1 − y 2 )
x3 y
+
=
(1 − xn y k )Dn,k .
2
4
6
1−x
(1 − x )(1 − x )
n≥3 k≥1
This is the table about the depth indecomposability.
w/d
2
3
4
5
6
7
8
9
1
1
1
2
1
Irr. MZV
ζ(2)
ζ(3)
ζ(5)
1
1
1
8
ζ(7)
ζ(6, 2)
ζ(9)
7
What have we done?
7.1
Kaneko-Noro-Tsurumaki Conjecture
. To solve the Kaneko-Noro-Tsurumaki conjecture stated in section 3 I used
Gap.I generate the matrices for the MZV relations from degree 7 until degree
22. This matrices are nothing else but the difference of the shuffle and stuffle
product
w1 w2 − w1 ∗ w2
2
for w1 ∈ H and w2 ∈ {x1 , x0 x1 , x20 x1 , x0 x21 } with degree(w1 )+degree(w2 ) = n.
The rank of this matrices equals the the dimensions of the Zagier’s conjecture.
Showing that this is number of relations is enough and solving the conjecture
until degree 22.
• Program: GAP.
• files in: .http://www.csd.uwo.ca/∼gcombar/ResearchMZV/GAP/Regularization.
• function: RelationWithRegularizationTruncatedUpToDegree( degree, filename).
7.2
Zagier’s Conjecture and Minh-Petitot Conjecture
Let us fix a degree n. In order to prove this conjecture for degree n this are the
steps I did
1. Find all Lyndon words of degree n.
2. Find all non-Lyndon words in H2 and factorize them as shuffle products
of Lyndon words. For example
ζ(2, 4) = 4ζ(5, 1) + 2ζ(4, 2) − ζ(3)
ζ(3) + ζ(2) ζ(4).
3. I called ζ(2), ζ(3) and ζ(4) the divisors of ζ(2, 4). Here I realize that the
sub-words intersect the divisors. This was a improvement in my original
xtaylor algorithm.
4. Find all the relations w1 w2 − w1 ∗ w2 for w1 , w2 Lyndon words with
degree(w1 ) + degree(w2 ) = n.
5. Replace all non-Lyndon words in w1 w2 − w1 ∗ w2 for its correspondence
factorization in Lyndon words, for example
6. Solve the system will find a Gröbner. basis for the ideal generated by all
the relations of degree n.
7. This Gröbner. basis should have a linear basis of the same dimension as
the Zagier’s conjecture numbers.
9
8. A big improvement here is use the Gröbner. basis for smaller degrees ≤ n
in the Lyndon factorization, for example
2
ζ(2, 4) = 4ζ(5, 1) + 2ζ(4, 2) − ζ(3) ζ(3) + ζ(2)
ζ(2) ζ(2) .
5
After this process find the Gröbner. basis is reduced to solve a linear system
where the variables are products and powers of Lyndon words. Now it is easy
to check the Minh-Petitot condition.
• Program: Magma.
• Function: Replacing( , , ).
• file: http://www.csd.uwo.ca/∼gcombar/ResearchMZV/Magma/All/LinearSystem4.mag
7.3
The 2-3 Conjecture
This conjecture is easily seen from the Gröbner. basis and the linear generators
in each degree for the algebra ζn . For degrees n = 2, 3 the 2-3 Lyndon words
and Linear generators for ζn are the same: ζ(2) and ζ(3) respectively.
For n = 5 there is one Lyndon word ζ(2, 3). The linear generators for ζ5
are ζ(5) and ζ(2)ζ(3). The connection is made by the corresponding relation
in the Gröbner. basis of degree 5. i.e. ζ(2, 3) = 92 ζ(5) − 2ζ(2)ζ(3). This shows
that the transcendental degree of ζ5+ is three with generators the three Lyndon
words ζ(2), ζ(3), ζ(5).
ζ5+
=
ζ5 ⊕ ζ4 ⊕ ζ3 ⊕ ζ2 = ζ5 ⊕ ζ3 ⊕ ζ2
=
=
Q[ζ(5), ζ(3)ζ(2)] ⊕ Q[ζ(3)] ⊕ Q[ζ(2)]
Q[ζ(5), ζ(3), ζ(2)]
=
Q[ζ(2, 3), ζ(3), ζ(2)].
Just to be clear: The dimension of ζn+ is the transcendental degree over Q.
But the dimension of ζn is the linear dimension over Q. For degree 6 there are
nothing to say since there are not 2-3 Lyndon words of weight 6.
For degree 7 there is a 2-3 Lyndon word ζ(2, 2, 3) = − 291
16 ζ(7) + 12ζ(5)ζ(2) −
The ζ7 space is generated by ζ(7), ζ(3)ζ(2)2 , ζ(5)ζ(2).
3
2
5 ζ(3)ζ(2) .
The generators for ζ8 are ζ(2)4 , ζ(5)ζ(3), ζ(3)2 ζ(2), ζ(6, 2) and the only 2-3
Lyndon word is
ζ(2, 3, 3) =
45
1111
27
ζ(6, 2) − ζ(5)ζ(3) + 2ζ(3)2 ζ(2) +
ζ(2)4 .
4
2
350
(5)
In each case the new 2-3 Lyndon word is providing with the new element in
each algebra. For ζ9+ the generators are ζ(7)ζ(2), ζ(3)3 , ζ(5)ζ(2)2 , ζ(3)ζ(2)3 , ζ(9)
10
and the new transcendental element
ζ(2, 2, 2, 3) =
641
18
3
ζ(9) − 30ζ(7)ζ(2) + ζ(5)ζ(2)2 − ζ(3)ζ(2)3
16
5
35
And for degree 10 they are ζ(8, 2), ζ(3)2 ζ(2)2 , ζ(7)ζ(3), ζ(5)2 , ζ(5)ζ(3)ζ(2), ζ(2)5 , ζ(6, 2)ζ(2)
and
ζ(2, 2, 3, 3) = −
7.4
2037
1737
3
56643
873
ζ(8, 2) +
ζ(7)ζ(3) +
ζ(5)2 − 24ζ(5)ζ(3)ζ(2) + 3/5ζ(3)2 ζ(2)2 −
ζ(2)5 (6)
64
32
32
5
7700
The Broadhurst-Kreimer conjecture
I used the 2-3 conjecture to find a transcendental basis for the algebra of the
MZV as follow.
Degree
2;
3;
5;
7;
8;
9;
10;
11;
12;
7.5
2-3 Lyndon word
ζ(2)
ζ(3)
ζ(2, 3)
ζ(2, 2, 3)
ζ(2, 3, 3)
ζ(2, 2, 2, 3)
ζ(2, 2, 3, 3)
ζ(2, 3, 3, 3), ζ(2, 2, 2, 2, 3)
ζ(2, 2, 2, 3, 3), ζ(2, 2, 3, 2, 3)
Transcendent basic element
ζ(2)
ζ(3)
ζ(5)
ζ(7)
ζ(6, 2)
ζ(9)
ζ(8, 2)
ζ(8, 2, 1), ζ(9, 2)
ζ(8, 2, 1, 1), ζ(10, 2)
Conclusion
I used the shuffle minus stuffle product to generate all the known relations of the
MZV. Using the xtaylor algorithm I expressed a non-Lyndon word as a product
and sum of Lyndon words. Finally I saw that finding the Gröbner. basis of
degree n for the ideal of the relations is equivalent to solve the linear system
given by the relations of degree n without non-Lyndon words (i.e replacing each
non-Lyndon words by its factorization) and the Gröbner. basis of degree less or
equal n. All the programs and this paper are available online at:
http://www.csd.uwo.ca/∼gcombar/ResearchMZV/
You can check the Gröbner with the relations found by Petitot and Minh in the
maple file http://www.csd.uwo.ca/∼gcombar/ResearchMZV/mzv16.m. I would
like to continue with this research in the future with a more power machine.
Germán Combariza.
Computer science department.
Western University. London ON Canada.
[email protected]
11
References
[1] M. E. Hoffman, The algebra of multiple harmonic series, J. Algebra 194:2
(1997), 477495.
[2] M. Kaneko, M. Noro, AND K. Tsurumaki, On a conjecture for the dimension of the space of the multiple zeta values, Software for Algebraic
Geometry, IMA 148 (2008), 4758.
[3] H.M. Minh, G. Jacob, M. Petitot, and N. E. Oussous, Aspects combinatoires
des polylogarithmes et des sommes dEulerZagier, S m. Lothar. Combin. 43
(1999)
[4] H.N. Minh, M. Petitot Lyndon words, polylogarithms and the Riemann ζ
function, Discrete Mathematics, Volume 217, Number 1, 28 April 2000 ,
pp. 273-292(20).
[5] H.N. Minh , M. Petitot, J. Van Der Hoeven, Shuffle algebra and polylogarithms, Proc. of FPSAC98, 10th International Conference on Formal Power
Series and Algebraic Combinatorics,Toronto, June 1998.
[6] R. Ree, Lie elements and an algebra associated with shuffles, Ann. Math.
68 (1958) 210220.
[7] M. Waldschmidt, Valeurs zeta multiples: une introduction, J. Th or. Nombres Bordeaux 12:2 (2000), 581595.
[8] W. Zudilin, Algebraic relations for multiple zeta values, Russian Mathematical Surveys Volume 58, Number 1.
12
| 0math.AC
|
arXiv:1304.5248v1 [math.AG] 18 Apr 2013
Gorenstein in codimension 4 –
the general structure theory
Miles Reid
Abstract
I describe the projective resolution of a codimension 4 Gorenstein
ideal, aiming to extend Buchsbaum and Eisenbud’s famous result in
codimension 3. The main result is a structure theorem stating that the
ideal is determined by its (k+1)×2k matrix of first syzygies, viewed as
a morphism from the ambient regular space to the Spin-Hom variety
SpHk ⊂ Mat(k + 1, 2k). This is a general result encapsulating some
theoretical aspects of the problem, but, as it stands, is still some way
from tractable applications.
This paper introduces the Spin-Hom varieties SpHk ⊂ Mat(k + 1, 2k)
for k ≥ 3, that I define as almost homogeneous spaces under the group
GL(k + 1) × O(2k) (see 2.4). These serve as key varieties for the (k + 1) × 2k
first syzygy matrixes of codimension 4 Gorenstein ideals I in a polynomial
ring S plus appropriate presentation data; the correspondence takes I to its
matrix of first syzygies. Such ideals I are parametrised by an open subscheme of SpHk (S) = Mor(Spec S, SpHk ). The open condition comes from
the Buchsbaum–Eisenbud exactness criterion “What makes a complex exact?” [BE1]: the classifying map α : Spec S → SpHk must hit the degeneracy
locus of SpHk in codimension ≥ 4.
The map α has Cramer-spinor coordinates Li and σJ in standard reprek−1
sentations kk+1 and k2
of GL(k + 1) and Pin(2k) (see 3.3), and the k × k
minors of M1 (I) are in the product ideal I · Sym2 ({σJ }). The spinors themselves should also be in I, so that the k×k minors of M1 (I) are in I 3 ; this goes
some way towards explaining the mechanism that makes the syzygy matrix
M1 (I) “drop rank by 3 at one go” – it has rank k outside V (I) = Spec(S/I)
and ≤ k − 3 on V (I).
1
The results here are not yet applicable in any satisfactory way, and raise
almost as many questions as they answer. While Gorenstein codimension 4
ideals are subject to a structure theorem, that I believe to be the correct
codimension 4 generalisation of the famous Buchsbaum–Eisenbud theorem
in codimension 3 [BE2], I do not say that this makes them tractable.
Thanks I am grateful to Chen Jungkai for inviting me to AGEA, and to
him and his colleagues at University of Taiwan for generous hospitality. My
visit was funded by Korean Government WCU Grant R33-2008-000-10101-0,
which also partly funded my work over the last 4 years, and I am very grateful
to Lee Yongnam for setting up the grant and administering every aspect of it.
I wish to thank Fabrizio Catanese, Eduardo Dias, Sasha Kuznetsov and Liu
Wenfei for contributing corrections, questions and stimulating discussion. I
owe a particular debt of gratitude to Alessio Corti for detailed suggestions
that have helped me improve the layout and contents of the paper.
Website See www.warwick.ac.uk/staff/Miles.Reid/codim4 for material accompanying this paper.
1
Introduction
Gorenstein rings are important, appearing throughout algebra, algebraic
geometry and singularity theory. A common source is Zariski’s standard
constructionL
of graded ring over a polarised variety X, L: the graded ring
0
R(X, L) =
n≥0 H (X, nL) is a Gorenstein ring under natural and fairly
mild conditions (cohomology vanishing plus KX = kX L for some kX ∈ Z,
see for example [GW]). Knowing how to construct R(X, L) by generators
and relations gives precise answer to questions on embedding X ֒→ Pn and
determining the equations of the image.
1.1
Background and the Buchsbaum–Eisenbud result
I work over a field k containing 12 (such as k = C, but see 4.5 for the more
general case). Let S = k[x1 , . . . , xn ] be a positively graded polynomial ring
with wt xi = ai , and R = S/IR a quotient of S that is a Gorenstein ring.
Equivalently, Spec R ⊂ Spec S = Ank is a Gorenstein graded scheme. By
the Auslander–Buchsbaum form of the Hilbert syzygies theorem, R has a
2
minimal free graded resolution P• of the form
0 ← P0 ← P1 ← · · · ← Pc ← 0
↓
R
(1.1)
where P0 = S → R = S/IR is the quotient map, and P1 → S gives a
minimum set of generators of the ideal IR . Here the length c of the resolution
equals n − depth R, and each Pi is a graded free module
of rank bi . I write
Lbi
⊕bi
Pi = bi S (as an abbreviation for S ), or Pi =
j=1 S(−dij ) if I need to
keep track of the gradings. The condition depth R = dim R that the depth
is maximal characterises the Cohen–Macaulay case, and then c = codim R =
codim(Spec R ⊂ Spec S). If in addition Pc is a free module of rank 1, so that
Pc ∼
number, then R is a Gorenstein ring of
= S(−α) with α the adjunction
P
canonical weight κR = α − ai ; for my purposes, one can take this to be
the definition of Gorenstein.
Duality makes the resolution (1.1) symmetric: the dual complex (P• )∨ =
HomS (P• , Pc ) resolves the dualising module ωR = ExtcS (R, ωS ), which
P is isomorphic to R (or, as a graded module, to R(κR ) with κR = α − ai ), so
that P• ∼
= (P• )∨ . In particular the Betti numbers bi satisfy the symmetry
bc−i = bi , or
Pc−i = HomS (Pi , Pc ) ∼
=
bi
M
S(−α + dij ),
j=1
where Pi =
bi
M
S(−dij ).
j=1
The Buchsbaum–Eisenbud symmetriser trick [BE2] adds precision to this
(this is where the assumption 12 ∈ S comes into play):
There is a symmetric perfect pairing S 2 (P• ) → Pc inducing the
duality P• ∼
= (P• )∨ .
The idea is to pass from P• as a resolution of R to the complex P• ⊗ P•
(the total complex of the double complex) as a resolution of R ⊗S R (left
derived tensor product), then to replace P• ⊗ P• by its symmetrised version
S 2 (P• ). In the double complex P• ⊗ P• , one decorates the arrows by signs
±1 to make each rectangle anticommute (to get d2 = 0). The symmetrised
complex S 2 (P• ) then involves replacing the arrows by half the sum or differences of symmetrically placed arrows. (This provides lots of opportunities
for confusion about signs!)
3
For details, see [BE2]. The conclusion is that P• has a ±-symmetric
bilinear form that induces perfect pairings Pi ⊗ Pc−i → Pc = S for each i,
compatible with the differentials.
The Buchsbaum–Eisenbud structure theorem in codimension 3 is a simple
consequence of this symmetry, and a model for what I try to do in this paper.
Namely, in codimension 3 we have
0 ← P0 ← P1 ← P2 ← P3 ← 0,
(1.2)
with P0 = S, P3 ∼
= P1∨, and the matrix M defining
= S, P2 = Hom(P1 , P3 ) ∼
the map P1 ← P2 is skew (that is, antisymmetric). If I set P1 = nS then the
respective ranks of the differentials in (1.2) are 1, n − 1 and 1; since M is
skew, his rank must be even, so that n = 2ν + 1. Moreover, the kernel and
cokernel are given by the Pfaffians of M, by the skew version of Cramer’s
rule.
Generalising the Buchsbaum–Eisenbud Theorem to codimension 4 has
been a notoriously elusive problem since the 1970s.
1.2
Main aim
This paper starts by describing the shape of the resolution of a codimension 4
Gorenstein ring by analogy with (1.2). The first syzygy matrix M1 : P1 ← P2
is a (k + 1) ×2k matrix whose k + 1 rows generically span a maximal isotropic
space of the symmetric quadratic form on P2 . The ideal IR is generated by
the entries of the map L : P0 ← P1 , which is determined by the linear algebra
of quadratic forms as the linear relation that must hold between the k + 1
rows of M1 .
This is all uncomplicated stuff, deduced directly from the symmetry trick
of [BE2]. It leads to the definition of the Spin-Hom varieties SpHk in the
space of (k + 1) × 2k matrixes (see Section 2.4). The first syzygy matrix M1
is then an S-valued point of SpHk , or a morphism α : Spec S → SpHk .
The converse is more subtle, and is the main point of the paper. By
construction, SpHk supports a short complex P1 ← P2 ← P3 of free modules
with a certain universal property. If we were allowed to restrict to a smooth
open subscheme S 0 of SpHk meeting the degeneracy locus SpHdgn
in codik
mension 4, the reflexive hull of the cokernel of M1 and the kernel of M2 would
provide a complex P• that resolves a sheaf of Gorenstein codimension 4 ideals
in S 0 . (This follows by the main proof below).
4
Unfortunately, this is only an adequate description of codimension 4
Gorenstein ideals in the uninteresting case of complete intersection ideals.
Any other case necessarily involves smaller strata of SpHk , where SpHk is
singular. Thus to cover every codimension 4 Gorenstein ring, I am forced
into the logically subtle situation of a universal construction whose universal
space does not itself support the type of object I am trying to classify, namely
a Gorenstein codimension 4 ideal. See 4.3 for further discussion of this point.
Main Theorem 2.5 gives the universal construction. To paraphrase: for a
polynomial ring S graded in positive degrees, there is a 1-to-1 correspondence
between:
(1) Gorenstein codimension 4 graded ideals I ⊂ S and
(2) graded morphisms α : Spec S → SpHk for which α−1 (SpHdgn
k ) has codimension ≥ 4 in Spec S.
I should say at once that this is intended as a theoretical structure result. It
has the glaring weakness that it does not so far make any tractable predictions
even in model cases (see 4.7 for a discussion). But it is possibly better than
no structure result at all.
1.3
Contents of the paper
Section 2.1 describes the shape of the free resolution and its symmetry, following the above introductory discussion. Section 2.4 defines the Spin-Hom
variety SpHk ⊂ Mat(k + 1, 2k), to serve as my universal space. The definition takes the form of a quasihomogeneous space for the complex Lie group
G = GL(k + 1) × O(2k) or its spin double cover GL(k + 1) × Pin(2k). More
explicitly, define SpHk as the closure of the G-orbit SpH0k = G · M0 of the
typical matrix M0 = ( I0k 00 ) under the given action of G = GL(k + 1) × O(2k)
on Mat(k + 1, 2k).
The degeneracy locus SpHdgn
is the complement SpHk \ SpH0k . Once these
k
definitions are in place, Section 2.5 states the main theorem, and proves it
based on the exactness criterion of [BE1].
The Spin-Hom varieties SpHk have a rich structure arising from representation theory. A matrix M1 ∈ SpH0k can be viewed as an isomorphism
between a k-dimensional space in kk+1 and a maximal isotropic space for ϕ
in k2k . This displays SpH0k as a principal GL(k) bundle over Pk ×OGr(k, 2k).
Section 3 discusses the properties of the SpHk in more detail, notably their
5
symmetry under the maximal torus and Weyl group. The spinor and nonspinor sets correspond to the two different spinor components OGr(k, 2k)
and OGr′ (k, 2k) of the maximal isotropic Grassmannian.
I introduce the Cramer-spinor coordinates σJ in 3.3; the main point is
that, for a spinor subset J ∪ J c , the (k + 1) × k submatrix of M1 ∈ SpHk
formed by those columns has top wedge factoring as (L1 , . . . , Lk+1 ) · σJ2 where
L : P0 ← P1 is the vector of equations (see Lemma 3.3.2). Ensuring that the
appropriate square root σJ is defined as an element σJ ∈ S involves the
point that, whereas the spinor bundle defines a 2-torsion
Weil divisor on the
V
affine orthogonal Grassmannian a OGr(k, 2k) ⊂ k k2k (the affine cone over
OGr(k, 2k) in Plücker space) and on SpHk , its birational transform under
the classifying maps α : Spec S → SpHk of Theorem 2.5 is the trivial bundle
on Spec S.
The spinor coordinates vanish on the degeneracy locus SpHdgn
and define
k
0
k+1
2k−1
an equivariant morphism SpHk → k
⊗k
. At the same time, they
vanish on the nonspin variety SpH′k , corresponding to the other component
OGr′ (k, 2k) of the Grassmannian of maximal isotropic subspaces; this has
nonspinor coordinates, that vanish on SpHk . Between them, these give set
theoretic equations for SpHk and its degeneracy locus.
The final Section 4 discusses a number of issues with my construction and
some open problems and challenges for the future.
2
The main result
For a codimension 4 Gorenstein ideal I with k + 1 generators, the module P2
of first syzygies is a 2k dimensional orthogonal space with a nondegenerate
(symmetric) quadratic form ϕ. The k + 1 rows of the first syzygy matrix
M1 (R) span an isotropic subspace in P2 with respect to ϕ. Since the maximal isotropic subspaces are k-dimensional, this implies a linear dependence
relation (L1 , . . . , Lk+1 ) that bases coker M1 and thus provides the generators
of I. A first draft of this idea was sketched in [Ki], 10.2.
2.1
The free resolution
Let S = k[x1 , . . . , xN ] be the polynomial ring over an algebraically closed
field k of characteristic 6= 2, graded in positive degrees. Let IR be a homogeneous ideal with quotient R = S/IR that is Gorenstein of codimension 4;
6
equivalently, IR defines a codimension 4 Gorenstein graded subscheme
V (IR ) = Spec R ⊂ AN
k = Spec S.
Suppose that IR has k + 1 generators L1 , . . . , Lk+1 . It follows from the
Auslander–Buchsbaum form of the Hilbert syzygies theorem and the symmetriser trick of Buchsbaum–Eisenbud [BE2] that the free resolution of R
is
0 ← P0 ← P1 ← P2 ← P3 ← P4 ← 0,
(2.1)
where P0 = S, P4 ∼
= P1∨; and moreover, P2 has a
= S, P3 = Hom(P1 , P4 ) ∼
nondegenerate symmetric bilinear form ϕ : S 2 P2 → P4 compatible with the
complex P• , so that P2 → P1 is dual to P3 → P2 under ϕ. The simple cases
of 2.3, Examples 2.1–2.3 give a sanity check (just in case you are sceptical
about the symmetry of ϕ).
A choice of basis of P2 gives ϕ the standard block form1 ( 0I I0 ). Then
the first syzygy matrix in (2.1) is M1 (R) = (A B), where the two blocks are
(k + 1) × k matrixes satisfying
(A B) ( 0I I0 ) t (A B) = 0,
(2.2)
that is, A tB + B tA = 0, or A tB is skew. I call this a (k + 1) × 2k resolution
(meaning that the defining ideal IR has k + 1 generators yoked by 2k first
syzygies).
The number of equations in (2.2) is k+2
. For example, in the typical case
2
k = 8, the variety defined by (2.2) involves k+2
= 45 quadratic equations
2
in 2k(k + 1) = 144 variables. The scheme Vk defined by (2.2) appears in
the literature as the variety of complexes. However it is not really the right
object – it breaks into 2 irreducible components for spinor reasons, and it is
better to study just one, which is my SpHk .
2.2
The general fibre
Let ξ ∈ Spec S = AN be a point outside V (IR ) = Spec R with residue field
K = k(ξ) (for example, a k-valued point, with K = k, or the generic point,
1
In the graded case this is trivial because ϕ is homogeneous of degree 0, so is basically
a nondegenerate quadratic form on a vector space V2 with P2 = V2 ⊗ S. See the discussion
in 4.5 for the more general case.
7
with K = Frac S). Evaluating (2.1) at ξ gives the exact sequence of vector
spaces
0 ← V0 ← V1 ← V2 ← V3 ← V4 ← 0
(2.3)
over K, where V0 = K, V4 ∼
= V1∨ ,
= K, V1 = (k + 1)K, V3 = Hom(V1 , V4 ) ∼
and V2 = 2kK with the nondegenerate quadratic form ϕ = ( 0I I0 ). Over K,
the maps in (2.3) can be written as the matrixes
0!
..
Ik 0
0 0
. .
0 ... 0 1
(2.4)
0 0
Ik 0
0
1
This data determines a fibre bundle over AN \ V (IR ) with the exact complex
(2.3) as fibre, and structure group the orthogonal group of the complex, which
I take to be GL(k + 1) × O(2k) or its double cover GL(k + 1) × Pin(2k).
2.3
Simple examples
Example 2.1 A codimension 4 complete intersection has L = (x1 , x2 , x3 , x4 )
and Koszul syzygy matrix
−x4
.
.
.
x3 −x2
.
−x4
.
−x3
.
x1
.
(2.5)
(A B) =
.
.
−x4
x2 −x1
.
x1
x2
x3
.
.
.
V
In this choice, A = M1,2,3 has rank 3 and 3 A = x24 · (x1 , . . . , x4 ). See 3.3
for spinors. A spinor subset J ∪ J c has an odd number i of columns from A
and the complementary 3V− i columns from B. For example, columns 1, 5, 6
give a 4 × 3 matrix with 3 M1,5,6 = x21 · (x1 , x2 , x3 , x4 ).
Example 2.2 Another easy case is that of a hypersurface section h = 0 in
a codimension 3 ideal given by the Pfaffians Pf i of a (2l + 1) × (2l + 1) skew
matrix M. The syzygy matrix is
!
−hI2l+1
M
(A B) =
.
(2.6)
Pf 1 . . . Pf 2l+1
0...0
One sees that a spinor σJ corresponding to 2l + 1 − 2i columns from A and
a complementary 2i from B is of the form hl−i times a diagonal 2i × 2i
Pfaffian of M. Thus the top wedge of the left-hand block A of (2.6) equals
σ 2 · (h, Pf 1 , . . . , Pf 2l+1 ) where σ = hl .
8
Example 2.3 The extrasymmetric matrix
a b
d
e
f
c
e
g
h
f
h
i
M =
−λa −λb
−λc
(2.7)
with a single multiplier λ is the simplest case of a Tom unprojection (see [TJ],
Section 9 for details). Let I be the ideal generated by the 4 × 4 Pfaffians
of M. The diagonal entries d, g, i of the 3 × 3 symmetric top right block
are all unprojection variables; thus i appears linearly in 4 equations of the
form i · (a, d, e, g) = · · · , and eliminating it projects to the codimension 3
Gorenstein ring defined by the Pfaffians of the top left 5 × 5 block.
If λ ∈ S is a perfect square, I is the ideal of√Segre(P2√× P2 ) ⊂ P8 up to
a coordinate change, but the Galois symmetry λ 7→ − λ swaps the two
factors. See [TJ], Section 9 for more details, and for several more families
of examples; in any of these cases, writing out the resolution matrixes (A B)
with the stated isotropy property makes a demanding but rewarding exercise
for the dedicated student.
By extrasymmetry, out of the 15 entries of M, 9 are independent and 6
repeats. His 4 × 4 Pfaffians follow a similar pattern. I write the 9 generators
of the ideal I of Pfaffians as the vector L =
λac + eh − f g, −λab − dh + ef, λa2 + dg − e2 ,
ah − bg + ce, −af + be − cd, λb2 + di − f 2 ,
λbc + ei − f h, λc2 + gi − h2 , ai − bh + cf
9
Its matrix of first syzygies M1 is the transpose of
.
a
b
−a .
c
−b −c .
−d −e −f
−e −g −h
−h .
f −h
.
f
.
.
.
d
e
f
.
λa
e
g
h
−λa
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
λc
.
.
g −e
−λb λc −g .
d
.
−λb e −d .
i
.
.
.
.
.
i
h
.
.
.
.
i
.
.
.
.
.
i
.
.
.
−λc
.
i
.
.
c
−b
−c
.
.
.
.
.
.
−h
f
h
.
.
h
.
.
c
.
.
.
.
.
.
.
.
(2.8)
−h f −λc
−f .
λb
e −d −λa
−c b
−h
−b .
f
−a .
.
. −a .
.
. −a
d
e
g
M1 is of block form (A B) with two 9×8 blocks, and one checks that LM1 = 0,
and M1 is isotropic for the standard quadratic form J = ( 0I I0 ), so its kernel is
t
M2 = tB
. The focus in (2.8) is on i as an unprojection variable, multiplying
A
d, e, g, a. One recognises its Tom3 matrix as the top 5 × 5 block, and the
Koszul syzygy matrix of d, e, g, a as Submatrix([6, 7, 8, 14, 15, 16], [6, 7, 8, 9]);
compare [KM].
For some of the spinors (see Section 3), consider the 8 × 9 submatrixes
formed by 4 out of the first 5 rows of (2.8), and the complementary 4 rows
from the last 8. One calculates their maximal minors with a mild effort:
V8
M1,2,3,4,13,14,15,16 = a2 (af − be + cd)2 · L,
V8
M1,2,3,5,12,14,15,16 = a2 (ah − bg + ce)2 · L,
V8
(2.9)
M1,2,4,5,11,14,15,16 = a2 (−λa2 − dg + e2 )2 · L,
V8
M1,3,4,5,10,14,15,16 = a2 (−λab − dh + ef )2 · L,
V8
M2,3,4,5,9,14,15,16 = a2 (−λac − eh + f g)2 · L.
The factor a comes from the 3 × 3 diagonal block at the bottom right, and
the varying factors are the 4 × 4 Pfaffians of the first 5 × 5 block. Compare
4.4 for a sample Koszul syzygy.
10
Exercise 2.4 Apply column and isotropic row operations to put the variable f down a main diagonal of B; check that this puts the complementary
A in the form of a skew 8 × 8 matrix and a row of zeros. Hint: order
the rows as 15, 16, 12, 11, 6, 2, 1, 5, 7, 8, 4, 3, 14, 10, 9, 13 and the columns as
1, 2, −3, 4, 5, −7, 8, 9, 6. (See the website for the easy code.) Do the same for
either variable e, h, and the same for any of a, b, c (involving the multiplier
λ).
Thus the isotropy condition tMJM can be thought of as many skew symmetries.
These examples provide useful sanity checks, with everything given by
transparent calculations; it is reassuring to be able to verify the symmetry of
the bilinear form on P2 asserted in Proposition 1, the shape of A tB in (2.2),
which parity of J gives nonzero spinors σJ , and other minor issues of this
nature.
I have written out the matrixes, spinors, Koszul syzygies etc. in a small
number of more complicated explicit examples (see the website). It should
be possible to treat fairly general Tom and Jerry constructions in the same
style, although so far I do not know how to use this to predict anything
useful. The motivation for this paper came in large part from continuing
attempts to understand Horikawa surfaces and Duncan Dicks’ 1988 thesis
[Di], [R1].
2.4
Definition of the Spin-Hom variety SpHk
Define the Spin-Hom variety SpHk ⊂ Mat(k + 1, 2k) as the closure under
G = GL(k + 1) × O(2k) of the orbit of M 0 = I0k 00 , the second matrix
in (2.4). It consists of isotropic homomorphisms V1 ← V2 , in other words
matrixes M1 whose k + 1 rows are isotropic and mutually orthogonal vectors
in V2 w.r.t. the quadratic form ϕ, and span a subspace that is in the given
component of maximal isotropic subspaces if it is k-dimensional.
In more detail, write SpH0k = G · M 0 ⊂ Mat(k + 1, 2k) for the orbit, SpHk
for its closure, and SpHdgn
= SpHk \ SpH0k for the degeneracy locus, consisting
k
of matrixes of rank < k. Section 3 discusses several further properties of SpHk
and its degeneracy locus SpHdgn
k .
11
2.5
The Main Theorem
Assume that S is a polynomial ring graded in positive degrees. Let I be
a homogeneous ideal defining a codimension 4 Gorenstein subscheme X =
V (I) ⊂ Spec S. Then a choice of minimal generators of I (made up of k + 1
elements, say) and of the first syzygies between these defines a morphism
α : Spec S → SpHk such that α−1 (SpHdgn ) has the same support as X, and
hence codimension 4 in Spec S.
Conversely, let α : Spec S → SpHk ⊂ Mat(k + 1, 2k) be a morphism
for which α−1 (SpHdgn ) has codimension ≥ 4 in Spec S. Assume that α is
graded, that is, equivariant for a positively graded action of Gm on SpHk ⊂
Mat(k + 1, 2k). Let M1 = (A B) be the matrix image of α (the matrix entries
of M1 or the coordinates of α are elements of S). Then by construction M1
and J tM1 define the two middle morphisms of a complex. I assert that this
extends to a complex
L
J tM
M
tL
1
1
0 ← P0 ←
− P1 ←−−
P2 ←−−−
P3 ←
− P4 ← 0.
(2.10)
in which P0 , P4 ∼
= S, the complex is exact except at P0 , and the image of L =
(L1 , . . . , Lk+1 ) generates the ideal of a Gorenstein codimension 4 subscheme
X ⊂ Spec S.
2.6
Proof
The first part follows from what I have already said. The converse follows
by a straightforward application of the exactness criterion of [BE1].
The complex P• of (2.10) comes directly from M1 . Namely, define P0 as
the reflexive hull of coker{M1 : P1 ← P2 } (that is, double dual); it has rank 1
because M1 has generic rank k. A graded reflexive module of rank 1 over a
graded regular ring is free (this is the same as saying that a Weil divisor on
a nonsingular variety is Cartier), so P0 ∼
= S. Given P3 ∼
= P1∨ , the generically
surjective map S ∼
= P0 ← P1 is dual to an inclusion S ֒→ P3 that maps to
the kernel of P2 ← P3 .
The key point is to prove exactness of the complex
ϕ1
ϕ2
ϕ3
ϕ4
P0 ←− P1 ←− P2 ←− P3 ←− P4 ← 0,
where I write ϕ1 = (L1 , . . . , Lk+1 ), ϕ2 = M1 , etc. to agree with [BE1]. The
modules and homomorphisms P0 , ϕ1 , P1 , ϕ2 , P2 , ϕ3 , P3 , ϕ4 , P4 of this complex
12
have respective ranks 1, 1, k + 1, k, 2k, k, k + 1, 1, 1, which accords with an
exact sequence of vector spaces, as in (2.3–2.4); this is Part (1) of the criterion
of [BE1], Theorem 1.
The second condition Part (2) requires the matrixes of ϕi to have maximal
nonzero minors generating an ideal I(ϕi ) that contains a regular sequence
of length i. However, P• is exact outside the degeneracy locus, that is, at
any point ξ ∈ Spec S for which α(ξ) ∈
/ SpHdgn
k , and by assumption, the locus
of such points has codimension ≥ 4. Thus the maximal minors of each ϕi
generate an ideal defining a subscheme of codimension ≥ 4. In a Cohen–
Macaulay ring, an ideal defining a subscheme of codimension ≥ i has height
≥ i. Q.E.D.
3
Properties of SpHk and its spinors
This section introduces the spinors as sections of the spinor line bundle S on
SpHk . The nonspinors vanish on SpHk and cut it out in Vk set theoretically.
The spinors vanish on the other component SpH′k and cut out set theoretically
the degeneracy locus SpHdgn
in SpHk .
k
The easy bit is to say that a spinor is the square root of a determinant on
Vk ⊂ Mat(k + 1, 2k) that vanishes to even order on a divisor of SpHk because
it is locally the square of a Pfaffian. The ratio of two spinors is a rational
function on SpHk .
The tricky point is that the spinors are sections of the spinor bundle S
on SpHk that is defined as a Pin(2k) equivariant bundle, so not described
by any particularly straightforward linear or multilinear algebra. As everyone knows, the spinor bundle S on OGr(k, 2k) is the ample generator of
Pic(OGr(k, 2k)), with the property that S ⊗2 is the restriction of the Plücker
bundle O(1) on Gr(k,V2k). On the affine orthogonal Grassmannian in Plücker
space a Gr(k, 2k) ⊂ k k2k , it corresponds to a 2-torsion Weil divisor class.
I write out a transparent treatment of the first example in 3.2.
I need to argue that the spinors pulled back to my regular ambient Spec S
by the appropriate birational transform are elements of S (that is, polynomials), rather than just sections of a spinor line bundle. The reason that
I expect to be able to do this is because I have done many calculations like
the Tom unprojection of 2.3, Example 2.3, and it always works. In the final
analysis, I win for the banal reason that the ambient space Spec S has no
2-torsion Weil divisors in its class group (because S is factorial), so that the
13
birational transform of the spinor bundle S to Spec S = AN is trivial.
The Cramer-spinor coordinates of the syzygy matrix M1 = (A B) have
the potential to clarify many points about Gorenstein codimension 4: the
generic rank of M1 is k, but it drops to k − 3 on Spec R; its k × k minors are
in IR3 . There also seems to be a possible explanation of the difference seen
in examples between k even and odd in terms of the well known differences
between the Weyl groups Dk (compare 3.1.3).
3.1
Symmetry
View GL(k+1) as acting on the first syzygy matrix M1 (R) by row operations,
and O(2k) as column operations preserving the orthogonal structure ϕ, or
the matrix ( 0I I0 ). The maximal torus Gk+1
m and Weyl group Ak = Sk+1 of the
first factor GL(k + 1) act in the obvious way by scaling and permuting the
rows of M1 .
I need some standard notions for the symmetry of O(2k) and its spinors.
For further details, see Fulton and Harris [FH], esp. Chapter 20 and [CR],
Section 4. Write V2 = k2k for the 2k dimensional vector space with basis
e1 , . . . , ek and dual basis f1 , . . . , fk , making the quadratic form ϕ = ( I0 I0 ).
Write U = U k = he1 , . . . , ek i, so that V2 = U ⊕ U ∨ . The orthogonal Grassmannian OGr(k, 2k) is defined as the variety of k-dimensional isotropic subspaces that intersect U in even codimension, that is, in a subspace of dimension ≡ k modulo 2.
3.1.1
The Dk symmetry of OGr(k, 2k) and SpHk
I describe the Dk Weyl group symmetry of the columns in this notation
(compare [CR], Section 4). The maximal torus Gkm of O(2k) multiplies ei
by λi and fi by λ−1
i , and acts likewise on the columns of M1 = (A B). The
Weyl group Dk acts on the ei , fi and on the columns of M1 = (A B) by
permutations, as follows: the subgroup Sk permutes the ei simultaneously
with the fi ; and the rest of Dk swaps evenly many of the ei with their
corresponding fi , thus taking U = he1 , . . . , ek i to another coordinate k-plane
in OGr(k, 2k). Exercise: The younger reader may enjoy checking that the
k − 1 permutations si = (i, i + 1) = (ei ei+1 )(fi fi+1 ) together with sk =
(ek fk+1 )(ek+1fk ) are involutions satisfying the standard Coxeter relations of
type Dk , especially (sk−1 sk )2 = 1 and (sk−2 sk )3 = 1.
14
3.1.2
Spinor and nonspinor subsets
The spinor sets J ∪ J c index the spinors σJ (introduced in 3.3). Let {ei , fi }
be the standard basis of k2k with form ϕ = ( I0 I0 ). There are 2k choices of
maximal isotropic subspaces of k2k based by a subset of this basis; each is
based by a subset J of {e1 , . . . , ek } together with the complementary subset
J c of {f1 , . . . , fk }. The spinor subsets are those for which #J has the same
parity as k, or in other words, the complement #J c is even; the nonspinor
subsets are those for which #J has the parity of k −1. The spinor set indexes
a basis σJ of the spinor space of OGr(k, 2k), and similarly, the nonspinor set
indexes the nonspinors σJ′ ′ of his dark twin OGr′ (k, 2k).
The standard affine piece of OGr(k, 2k) consists of k-dimensional spaces
based by k vectors that one writes as a matrix (I A) with A a skew k × k
matrix. The spinor coordinates of (I A) are the 2i × 2i diagonal Pfaffians
of A for 0 ≤ i ≤ [ k2 ]. They correspond in an obvious way to the spinor sets
just defined and they are the spinors apart from the quibble about taking an
overall square root and what bundle they belong to.
3.1.3
Even versus odd
The distinction between k even or odd is crucial for anything to do with
O(2k), Dk , spinors, Clifford algebras, etc. The spinor and nonspinor sets
correspond to taking a subset J of {e1 , . . . , ek } and the complementary set
J c of {f1 , . . . , fk }. The 2k choices correspond to the vertices of a k-cube.
When k is even this is a bipartite graph; the spinors and nonspinors form
the two parts. By contrast, for odd k, both spinors and nonspinors are
indexed by the vertices of the k-cube divided by the antipodal involution
([CR], Section 4 writes out the case k = 5 in detail).
For simplicity, I assume that k is even in most of what follows; the common case in applications that I really care about is k = 8. Then J = ∅ and
J c = {1, . . . , k} is a spinor set, and the affine pieces represented by (I X) and
(Y I) (with skew X or Y ) are in the same component of OGr(k, 2k). The
odd case involves related tricks, but with some notable differences of detail
(compare [CR], Section 4).
3.1.4
The other component OGr′ and SpH′k
I write OGr′ (k, 2k) for the other component of the maximal isotropic Grassmannian, consisting of subspaces meeting U in odd codimension. Swapping
15
oddly many of the ei and fi interchanges OGr and OGr′ . Likewise, SpH′k is
the closure of the G-orbit of the matrix M0′ obtained by interchanging one
corresponding pair of columns of M0 .
Claim 3.1 Write Vk for the scheme defined by (2.2) (that is, the “variety of
complexes”). It has two irreducible components Vk = SpHk ∪ SpH′k containing
matrixes of maximal rank k. The two components are generically reduced
and intersect in the degenerate locus SpHdgn
k . (But one expects Vk to have
embedded primes at its smaller strata, as in the discussion around (3.5).)
This follows from the properties of spinor minors ∆J discussed in Exercise 3.2.1: the ∆J are k × k minors defined as polynomials on Vk , and vanish
on SpH′k but are nonzero on a dense open subset of SpHk .
3.2
A first introduction to OGr(k, 2k) and its spinors
The lines on the quadric surface provide the simplest calculation, and already
have lots to teach us about OGr(2, 4) and OGr(k, 2k): the conditions for the
2 × 4 matrix
a b x y
(3.1)
N=
c d z t
to be isotropic for ( 0I I0 ) are
ax + by = 0,
az + bt + cx + dy = 0,
cz + dt = 0.
(3.2)
Three equations (3.2) generate an ideal IW defining a codimension 3 complete
intersection W ⊂ A8 that breaks up into two components Σ⊔Σ′ , corresponding to the two pencils of lines on the quadric surface: the two affine pieces of
OGr(2, 4) that consist of matrixes row equivalent to (I A) or (A I), with A
a skew matrix, have one of the spinor minors ∆1 = ad − bc or ∆2 = xt − yz
nonzero, and
dx − bz = at − cy = 0 and dy − bt = −(az − cx)
(3.3)
on them. This follows because all the products of ∆1 , ∆2 with the nonspinors
minors dx − bz, at − cy are in IW , as one checks readily. Thus if ∆1 6= 0 (say),
I can multiply by the adjoint of the first block to get
∆1 0 dx − bz dy − bt
a b x y
d −b
(3.4)
=
0 ∆1 az − cx at − cy
c d z t
−c a
16
where the second block is skew. Note that
∆1 · ∆1 ∆2 − (az − cx)2 ∈ IW .
(3.5)
If ∆1 6= 0, the relations (3.2) imply that we are in Σ. The ideal of Σ is
obtained from (3.2) allowing cancellation of ∆1 ; in other words IΣ = [IW : ∆1 ]
is the colon ideal with either of the spinor minors ∆1 or ∆2 .
The second block in (3.4) is only skew mod IW after cancelling one of
a, b, . . . , t; similarly ∆1 ∆2 − (az − cx)2 ∈
/ IW , so that (3.5) involves cancelling
∆1 . Thus a geometric description of Σ, Σ′ ⊂ Mat(k, 2k) should usually lead
to ideals with embedded primes at their intersection or its smaller strata.
Now by relation (3.5), the Plücker embedding takes OGr(2, 4) to the conic
XZ = Y 2 , with X = ∆1 = ad − bc, Y = az − cx, Z = ∆2 = xt − yz. This
is (P1 , O(2)) parametrised by u2 , uv, v 2 where u, v base H 0 (P1 , O(1)). Thus
X = u2 , Y = uv and Z = v 2 on OGr(2, 4); the spinors are u and v. The ratio
u : v equals X : Y = Y √
: Z. Each √
of ∆1 and ∆2 vanishes on a double divisor,
but the quantities u = ∆1 , v = ∆2 are not themselves polynomial.
The conclusion is that the minors ∆1 and ∆2 are spinor squares, that is,
squares of sections u, v of a line bundle S, the spinor bundle on OGr(2, 4). If
we view OGr(2, 4) as a subvariety of Gr(2, 4), only S ⊗2 extends to V
the Plücker
line bundle O(1). Embedding OGr(2, 4) in the Plücker space P( 2 C4 ) and
taking the affine cone gives the affine spinor variety a OGr(2, 4) as the cone
over the conic, and S with its sections u, v as the ruling.
In fact a OGr(2, 4) and his dark twin a OGr′ are two ordinary
V quadric
cones in linearly disjoint vector subspaces of the Plücker space 2 C4 , and
the spinor bundle on the union has a divisor class that is a 2-torsion Weil
divisor on each component. This picture is of course the orbifold quotient of
±1 acting on two planes A2 meeting transversally in A4 .
3.2.1
Exercise
Generalise the above baby calculation to the subvariety Wk ⊂ Mat(k, 2k)
of matrixes (A X) whose k rows span an isotropic space for ( 0I I0 ), or in
equations, the k × k product A tX is skew. Assume k is even.
(1) Wk ⊂
Mat(k, 2k) is a complete intersection subvariety of codimension
k+1
. [Hint: Just a dimension count.]
2
(2) Wk breaks up into two irreducible components Σ∪Σ′ , where Σ contains
the space spanned by (I X) with X skew, or more generally, by the span
17
of the columns J ∪ J c for J a spinor set; its nondegenerate points form
a principal GL(k) bundle over the two components OGr ⊔ OGr′ of the
maximal isotropic Grassmannian.
(3) For J a spinor set, the k ×k spinor minor ∆J of (A X) (the determinant
of the submatrix formed by the columns J ∪ J c ) is a polynomial on
Mat(k × 2k) that vanishes on Σ′ , and vanishes along a double divisor
of Σ, that is, twice a prime Weil divisor DJ .
(4) The Weil divisors DJ1 and DJ2 corresponding to two spinor sets J1 and
J2 are linearly equivalent. [Hint: First suppose that J1 is obtained
from J by exactly two transpositions, say (e1 f2 )(e2 f1 ), and argue as in
(3.5) to prove that σJ σJ1 restricted to Σ is the square of either minor
obtained by just one of the transpositions.]
3.2.2
Spinors on OGr(k, 2k)
The orthogonal Grassmann variety OGr(k, 2k) has a spinor embedding into
k−1
P(k2 ), of which the usual Plücker embedding
OGr(k, 2k) ⊂ Gr(k, 2k) ֒→ P
k
^
k−1
k2k
is the Veronese square. The space of spinors k2
is a representation of the
spin double cover Pin(2k) → O(2k).
A point W ∈ OGr(k, 2k) is a k-dimensional subspace W k ⊂ k2k isotropic
for ( I0 I0 ) and intersecting U = he1 , . . . , en i in even codimension. I can write a
basis as the rows of a k × 2k matrix NW . If I view W as a point of Gr(k, 2k),
its Plücker coordinates are all the k × k minors of NW . There are 2k
of
k
these (that is, 12870 if k = 8), a fraction of which vanish OGr(k, 2k), as the
determinant of a skew matrix of odd size.
The finer embedding of OGr(k, 2k) is by spinors. The spinors σJ are
sections of the spinor line bundle S, 2k−1 of them (which is 128 if k = 8,
about 1/100 of the number of Plücker minors). Each comes by taking a k × k
submatrix formed by a spinor subset of columns of NW (in other words,
restricting to an isotropic coordinate subspace of k2k in the
k specified component OGr(k, 2k)), taking its 2κ × 2κ minor (where κ = 2 ) and factoring
it as the perfect square of a section of S. The only general reason for a
2κ × 2κ minor to be a perfect square is that the submatrix is skew in some
18
basis; in fact, as in (3.4), after taking one fixed square root of a determinant,
and making a change of basis, the maximal isotropic space can be written as
(I X) with X skew, and the spinors are all the Pfaffians of X.
3.3
3.3.1
Cramer-spinor coordinates on SpHk
Geometric interpretation
A point of the open orbit SpH0k ⊂ SpHk is a matrix M of rank k; it defines
an isomorphism from a k-dimensional subspace of V1 (the column span of
M) to its row span, a maximal isotropic subspace of V2 in the specified
component OGr(k, 2k). Therefore the nondegenerate orbit SpH0k ⊂ SpHk
has a morphism to P(V1∨ ) × OGr(k, 2k) that makes it a principal GL(k)
bundle. The product P(V1∨ ) × OGr(k, 2k) is a projective homogeneous space
under G = GL(k + 1) × Pin(2k)
k−1
It embeds naturally in the projectivisation of kk+1 ⊗k2 , with the second
factor the space of spinors. This is the representation of G with highest weight
vector v = (0, . . . , 0, 1) ⊗ (1, 0, . . . , 0). The composite
k−1
SpH0k → P(V1∨ ) × OGr(k, 2k) ֒→ P(kk+1 ⊗ k2
)
(3.6)
takes the typical matrix M0 (or equivalently, the complex (2.4)) to v.
The Cramer-spinor coordinates of α ∈ SpHk (S) are the bihomogeneous
coordinates under the composite map (3.6).
3.3.2
Spinors as polynomials
The spinors σJ occur naturally as sections of the spinor line bundle S on
OGr(k, 2k), and so have well defined pullbacks to SpH0k or to any scheme T
with a morphism α : T → SpH0k . For σJ to be well defined in H 0 (OT ), the
pullback of the spinor line bundle to T must be trivial.
Lemma 3.2 Let α ∈ Mor(Spec S, SpHk ) = SpHk (S) be a classifying map as
in Theorem 2.5 and write M1 ∈ Mat(S, k + 1, 2k) for its matrix (with entries
in S). Then for a spinor set J ∪ J c (as in 3.1.2), the (k + 1) × k submatrix
NJ of M1 with columns J ∪ J c has
^k
NJ = L · σJ2 ,
(3.7)
where L = (L1 , . . . , Lk+1 ) generates the cokernel of M1 , and σJ ∈ S.
19
3.4
Proof
A classifying map α ∈ SpHk (S) as in Theorem 2.5 restricts to a morphism α
from the nondegenerate locus Spec S \ V (IR ) to SpH0k ; on the complement of
V (IR ), the matrix M1 has rank k, and its kth wedge defines the composite
morphism to the product Pk × Gr(k, 2k) in its Segre embedding:
Spec S \ V (IR ) → SpH0k → Pk × OGr(k, 2k)
^k
֒→ Pk × Gr(k, 2k) ⊂ P kk+1 ⊗
V 2k . (3.8)
V
The entries of k NJ are k + 1 coordinates of this morphism, and are of the
form Li · σJ2 already on the level of Pk × OGr(k, 2k).
Note that Spec S \ V (IR ) is the complement in Spec S = AN of a subset
of codimension ≥ 4 so has trivial Pic. Each maximal minor of NJ splits as Li
times a polynomial that vanishes on a divisor that is a double (because it is
the pullback of the square of a spinor); therefore the polynomial is a perfect
square in S. QED
The following statement is the remaining basic issue that I am currently
unable to settle in general.
Conjecture 3.3 Under the assumptions of Lemma 3.3.2, σJ ∈ IR .
This is clear when R is reduced, that is, IR is a radical ideal. Indeed if
σJ is a unit at some generic point ξ ∈ V (IR ) = Spec R, then (3.7) implies
that IR is generated at ξ by the k × k minors of the (k + 1) × k matrix
NJ ; these equations define a codimension 2 subscheme of Spec S, which is
a contradiction. This case is sufficient for applications to construction of
ordinary varieties, but not of course to Artinian subschemes of A4 .
The conjecture also holds under the assumption that IR is generically a
codimension 4 complete intersection. Indeed, the resolution of IR near any
generic point ξ ∈ V (IR ) is then the 4 × 6 Koszul resolution of the complete
intersection direct sum some nonminimal stuff that just add invertible square
matrix blocks. Then both the Li and the σJ are locally given by Example 2.1.
At present, the thing that seems to make the conjecture hard is that the
definition of the σJ and the methods currently available for getting formulas
for them consists of working on the nondegenerate locus of SpHk : choose a
block diagonal form and take the Pfaffian of a skew complement, . . . This is
just not applicable at points σ ∈ V (IR ).
20
The conjecture could possibly be treated by a more direct understanding
of the spin morphism Spec S → k2k defined by spinors and nonspinors, not
passing via the square
Vk root of the Plücker morphism as I do implicitly in
Lemma 1 by taking
.
4
4.1
Final remarks, open problems
Birational structure and dimension of SpHk
A general M = (A B) ∈ SpHk has k + 1 rows that span a maximal isotropic
space U ∈ OGr(k, 2k) and 2k columns that span a k-dimensional vector
subspace of kk+1 , that I can view as a point of Pk ; thus SpH0k is a principal
GL(k) bundle over Pk ×OGr(k, 2k). In particular, dim SpHk = k 2 +k + k2 =
3k 2 +k
.
2
Ik 0
is calculated
The tangent space to SpHk at the general point
M
=
0
0
0
A′
B′
k
k
by writing an infinitely near matrix as M0 + ak+1
bk+1 ; here the blocks
A′k and Bk′ are k × k matrixes, and ak+1 and bk+1 are 1 × k rows. Then
the tangent space to Vk defined by A tB = 0 is the affine subspace obtained
by setting
Bk′ to be skew and bk+1 = 0. Therefore SpHk has codimension
2
k+1
+ k and dimension 2k(k + 1) − k+1
− k = 3k 2+k .
2
2
′
It is interesting to observe that equations (2.2) express SpH
k ∪ SpHk as
k+1
an almost complete intersection. Namely, (2.2) is a set of 2 equations in
2
A2k(k+1) vanishing on a variety of dimension 3k 2+k , that is, of codimension
k+1
− 1.
2
4.2
Intermediate rank
The Spin-Hom variety SpHk certainly contains degenerate matrixes M1 of
rank k − 1 or k − 2, but any morphism Spec S → SpHk that hits one of these
must hit the degeneracy locus in codimension ≤ 3, so does not correspond to
anything I need here. The following claim must be true, but I am not sure
where it fits in the logical development.
Claim 4.1 Every point P ∈ SpHk corresponds to a matrix M1 = (A B) of
rank ≤ k. If a morphism α : Spec S → SpHk takes ξ to a matrix M1 of
rank k + 1 − i for i = 1, 2, 3, 4 then α−1 (SpHdgn
k ) has codimension ≤ i in
a neighbourhood of ξ. In other words, a morphism α that is regular in the
21
sense of my requirement never hits matrixes M1 of rank intermediate between
k and k − 3; and if α is regular then α−1 (SpHdgn
k ) has codimension exactly 4.
4.3
The degeneracy locus as universal subscheme
The proof in 2.6 doesn’t work for SpHk itself in a neighbourhood of a point
of SpHdgn
k , because taking the reflexive hull, and asserting that P0 is locally
free works only over a regular scheme. Moreover, it is not just the proof that
goes wrong. I don’t know what happens over the strata of SpHdgn
where M1
k
drops rank by only 1 or 2.
We discuss the speculative hope that SpHdgn
⊂ SpHk has a description
k
as a kind of universal codimension 4 subscheme, with the inclusions enjoying
some kind of Gorenstein adjunction properties. But if this is to be possible at
all, we must first discard uninteresting components of SpHdgn
corresponding
k
to matrixes of intermediate rank k − 1 or k − 2.
It is possible that there is some universal blowup of some big open in SpHk
that supports a Gorenstein codimension 4 subscheme and would be a universal space in a more conventional sense. Or, as the referee suggests, there
might be a more basic sense in which appropriate codimension 4 components
Γ of the degeneracy locus are universal Gorenstein embeddings, meaning that
the adjunction calculation ωΓ = Ext4OSpH (OΓ , ωSpH ) for the dualising sheaf is
locally free and commutes with regular pullbacks.
4.4
Koszul syzygies
Expressing the generators of I as a function
of the entries of the syzygy
V2
matrix is essentially given by the map
P1 → P2 that writes the Koszul
syzygies as linear combinations of the minimal syzygies.
The Li are certainly linear combinations of the entries of M1 . More
precisely, since the 2k columns of M1 provide a minimal basis for the syzygies,
they cover in particular the Koszul syzygies Li · Lj − Lj · Li ≡ 0. This means
that for every i 6= j there is column vector vij with entries in S such that
M1 vij = (. . . , Lj , . . . , Li , . . . ) is the column vector with Lj in the ith place
and Li in the jth and 0 elsewhere. For example, referring to Example 2.3,
you might enjoy the little exercise in linear algebra of finding the vector
v = (−λc, λb, 0, 0, 0, d, e, g, 0, 0, 0, 0, 0, 0, 0, 0) for which
v tM1 = (−λab − dh + ef, −λac − eh + f g, 0, 0, 0, 0, 0, 0, 0),
22
where tM1 is the matrix of (2.8), and similarly for 35 other values of i, j.
4.5
More general ambient ring S
I restrict to the case of ideals in a graded polynomial ring over a field of
characteristic 6= 2 in the belief that progress in this case will surely be followed
by the more general case of a regular local ring. Then P2 is still a free module,
with a perfect symmetric bilinear form S 2 (P2 ) → P4 , with respect to which
P1 ← P2 is the dual of P2 ← P3 . This can be put in the form ( 0I I0 ) over
the residue field k0 = S/mS of S if we assume that k(S) is algebraically
closed and contains 12 ; we can do the same over S itself if we assume that S
is complete (to use Hensel’s Lemma). At some point if we feel the need for
general regular rings, we can probably live with a perfect quadratic form ϕ
and the dualities it provides, without the need for the normal form ( I0 I0 ).
4.6
More general rings and modules
Beyond the narrow question of Gorenstein codimension 4, one could ask for
the structure of any free resolution of an S-module M or S-algebra R. As in
2.2, one can say exactly what the general fibre is, and think of the complex
P• as a fibre bundle over S \ Supp M with some product of linear groups
as structure group. If we are doing R-algebras, the complex P• also has a
symmetric bilinear structure, that reduces the structure group. My point
is that if we eventually succeed in making some progress with Gorenstein
codimension 4 rings, we might hope to also get some ideas about Cohen–
Macaulay codimension 3 and Gorenstein codimension 5.
For example, in vague terms, there is a fairly clear strategy how to find a
key variety for the resolution complexes of Gorenstein codimension 5 ideals,
by analogy with my Main Theorem 2.5. In this case, the resolution has the
shape
0 ← P0 ← P1 ← P2 ← P3 ← P4 ← P5 ← 0,
(4.1)
with P0 = S, P1 = (a + 1)S, P2 = (a + b)S and P3 , . . . , P5 their duals.
The complex is determined by two syzygy matrixes M1 ∈ Mat(a + 1, a + b) of
generic rank a defining P1 ← P2 and a symmetric (a+b)×(a+b) matrix M2 of
generic rank b defining P2 ← P3 = P2∨ , constrained by the complex condition
M1 M2 = 0. The “general fibre” is given by the pair M1 = ( I0a 00 ), M2 = 00 I0b ,
the appropriate key variety is its closed orbit under GL(a + 1) × GL(a + b).
23
The maximal nonzero minors of M1 and M2 define a map to a highest weight
orbit in
a
a
b
^
^
^
2
Hom
P2 , P1 × Sym
P2 .
4.7
Difficulties with applications
I expand what the introduction said about the theory currently not being applicable. We now possess hundreds of constructions of codimension 4 Gorenstein varieties, for example, the Fano 3-folds of [TJ], but their treatment
(for example, as Kustin–Miller unprojections) has almost nothing to do with
the structure theory developed here. My Main Theorem 2.5 does not as
it stands construct anything, because it does not say how to produce morphisms α : Spec S → SpHk , or predict their properties. The point that must
be understood is not the key variety SpHk itself, but rather the space of
morphisms Mor(Spec S, SpHk ), which may be intractable or infinitely complicated (in the sense of Vakil’s Murphy’s law [Va]); there are a number of
basic questions here that I do not yet understand.
Even given α, we do not really know how to write out the equations
(L1 , . . . , Lk+1 ), other than by the implicit procedure of taking hcfs of k × k
minors. One hopes for a simple formula for the defining relations Li as a
function of the first syzygy matrix M1 = (A B). Instead, one gets
Vk the vector
(L1 , . . . , Lk+1 ) by taking out the highest common factor from
MI for any
spinor subset I, asserting that it is a perfect square σJ2 . The disadvantage is
that as it stands this is only implicitly a formula for the Li .
4.8
Obstructed constructions
One reason that Mor(S, SpHk ) is complicated is that the target is big and
singular and needs many equations. However, there are also contexts in which
S-valued points of much simpler varieties already give families of Gorenstein
codimension 4 ideals that are obstructed in interesting ways.
Given a 2V
× 4 matrix A = ( ab11 ab22 ab33 ab44 ) with entries in a regular ring S, the
6 equations 2 A = 0 define a Cohen–Macaulay codimension 3 subvariety
V ⊂ Spec S. An elephant X ∈ |−KV | is then a Gorenstein subvariety of
codimension 4 with a 9 × 16 resolution. If we are in the “generic” case with 8
independent indeterminate entries, V is the affine cone over Segre(P1 × P3 ),
and X is a cone over a divisor of bidegree (k, k + 2) in Segre(P1 × P3 ).
24
Although X ⊂ V is a divisor, if we are obliged to treat it by equations
in the ambient space Spec S, it needs 3 equations in “rolling factors format”.
The general case of this is contained in Dicks’ thesis [Di], [R1]: choose two
vectors m1 , m2 , m3 , m4 and n1 , n2 , n3 , n4 , and assume that the identity
X
X
ai ni ≡
bi mi
(4.2)
holds as an equality in the ambient ring S. Then the 3 equations
X
X
X
X
ai mi =
bi mi ≡
ai ni =
bi ni = 0
(4.3)
define a hypersurface X ⊂ V that is an elephant X ∈ |−KV | and thus a
Gorenstein subvariety with 9 × 16 resolution.
The problem in setting up the data defining X is then to find solutions
in S of (4.2). In other words, these are S-valued points of the affine quadric
cone Q16 , or morphisms Spec S → Q16 . How to map a regular ambient space
to the quadratic cone Q16 is a small foretaste of the more general problem of
the classifying map Spec S → SpHk . This case is discussed further in [Ki],
Example 10.8, which in particular writes out explicitly the relation between
(4.3) and the classifying map Spec S → SpHk of Theorem 2.5.
There are many quite different families of solutions to this problem, depending on what assumptions we make about the graded ring S, and how
general we take the matrix A to be; different solutions have a number of
important applications to construction and moduli of algebraic varieties, including my treatment of the Horikawa quintic n-folds.
Another illustration of the phenomenon arises in a recent preprint of
Catanese, Liu and Pignatelli [CLP]. Take the 5 × 5 skew matrix
v u z2 D
z1 y m25
(4.4)
M =
l m35
m45
with entries in a regular ring S0 , and suppose that v, u, z2, D forms a regular
sequence in S. Assume that the identity
z1 m45 − ym35 + lm25 ≡ av + bu + cz2 + dD
(4.5)
holds as an equality in S0 . The identity (4.5) puts the Pfaffian Pf 23.45 in the
ideal (v, u, z2 , D); the other 4 Pfaffians are in the same ideal for the trivial
reason that every term involves one entry from the top row of M.
25
This is a new way of setting up the data for a Kustin–Miller unprojection:
write Y ⊂ Spec S0 for the codimension 3 Gorenstein subscheme defined by
the Pfaffians of M. It contains the codimension 4 complete intersection
V (v, u, z2 , D) as a codimension 1 subscheme, and unprojecting V in Y adjoins
an unprojection variable x2 having 4 linear equations x2 · (v, u, z2, D) = · · · ,
giving a codimension 4 Gorenstein ring with 9 × 16 resolution.
The problem of how to fix (4.5) as an identity in S0 is again a question
of the S0 -valued points of a quadric cone, this time a quadric Q14 of rank 14.
[CLP], Proposition 5.13 find two different families of solutions, and exploit
this to give a local description of the moduli of their surfaces.
At first sight this looks a bit like a Jerry15 unprojection. In fact one of
the families of [CLP] (the one with c0 = Bx = 0) can easily be massaged to
a conventional Jerry15 having a double Jerry structure (compare [TJ], 9.2),
but this does not seem possible for the more interesting family in [CLP] with
Dx = (l/c0 )Bx .
Question Do these theoretical calculations contain the results of [Di], [CLP]
and the like?
Answer Absolutely not. They may provide a framework that can produce
examples, or simplify and organise the construction of examples. To get
complete moduli spaces, it is almost always essential to use other methods,
notably infinitesimal deformation calculations or geometric constructions.
Question The fact that S can have various gradings seems to add to the
complexity of the space Mor(S, SpHk ), doesn’t it?
Answer That may not be the right interpretation – we could perhaps think
that Mor(S, SpHk ) (or even the same just for Mor(S, Q2k ) into a quadric of
rank 2k ≥ 4) is infinite dimensional and infinitely complicated, so subject
to Murphy’s law [Va], but that when we cut it down to graded in given
degrees, it becomes finitely determined, breaking up into a number of finite
dimensional families that may be a bit singular, but can be studied with
success in favourable cases.
4.9
4.9.1
Problem session
Computing project
It is a little project in computer algebra to write an algorithm to put the
projective resolution (2.1) in symmetric form. This might just be a straight26
forward implementation of the Buchsbaum–Eisenbud symmetrised complex
S 2 P• outlined in Section 1. Any old computer algebra package can do syzygies, but as far as I know, none knows about the symmetry in the Gorenstein
case.
We now have very many substantial working constructions of codimension 4 Gorenstein varieties. We know in principle that the matrix of first
syzygies can be written out in the (A B) form of (2.8), but as things stand,
it takes a few hours or days of pleasurable puzzling to do any particular case.
4.9.2
Linear subvarieties
What are the linear subvarieties of SpHk ? The linear question may be
tractable, and may provide a partial answer to the quest for an explicit
structure result.
The Spin-Hom variety SpHk is defined near a general point by quadratic
equations, so its linear subspaces can be studied by the tangent-cone construction by analogy with the linear subspaces of quadrics, Segre products
or Grassmannians: the tangent plane TP at P ∈ V intersects V in a cone, so
that linear subspaces of V through P correspond to linear subspaces in the
base of the cone. Now choose a point of the projected variety and continue.
Presumably at each stage there are a finite number of strata of the variety
in which to choose our point P , giving a finite number of types of Π up to
symmetry. I believe that the two famous cases of the Segre models of P2 × P2
and P1 × P1 × P1 are maximal linear space of SpH8 .
It is possible that this method can be used to understand more general
morphisms Spec S → SpHk from the regular space Spec S. In this context,
it is very suggestive that Tom and Jerry [TJ] are given in terms of linear
subspaces of Gr(2, 5). In this case, the intersection with a tangent space is
a cone over P1 × P2 , so it is clear how to construct all linear subspaces of
Gr(2, 5), and equally clear that there are two different families, and how they
differ.
4.9.3
Breaking the Ak and Dk symmetry
Experience shows that the bulk constructions of Gorenstein codimension 4
ideals do not have the symmetry of the Buchsbaum–Eisenbud Pfaffians in
codimension 3. The equations and syzygies invariably divide up into subsets
that one is supposed to treat inhomogeneously. For example, in the 9 × 16
27
unprojection cases, the defining equations split into two sets, the 5 Pfaffian
equations of the variety in codimension 3 not involving the unprojection
variable s, and the 4 unprojection equations that are linear in s.
The columns of the syzygy matrix (A B) are governed by the algebraic
group Spin(2k) of type Dk , whereas its rows are governed by GL(k + 1)
of type Ak . The common bulk constructions of Gorenstein codimension 4
ideals seem to to accommodate the Ak symmetry of the rows of M1 and
the Dk symmetry of its columns by somehow breaking both to make them
compatible. This arises if you try to write the 128 spinor coordinates σJ as
linear combinations of the 9 relations (L1 , . . . , Lk+1), so relating something
to do with the columns of M1 to its rows. This symmetry breaking and its
effect is fairly transparent in 2.3, Example 2.2, (2.6).
Example 2.3 is more typical. (This case comes with three different Tom
projections, so may be more amenable.) Of the 128 spinors σJ , it turns out
that 14 are zero, 62 are of the form a monomial times one of the relations
Li (as in (2.9)), and the remainder are more complicated (probably always
a sum of two such products). Mapping this out creates a correspondence
from spinor sets to relations, so from the rows of M1 to its columns; there
is obviously a systematic structure going on here, and nailing it down is
an intriguing puzzle. How this plays out more generally for Kustin–Miller
unprojection [KM], [PR] and its special cases Tom and Jerry [TJ] is an
interesting challenge.
4.9.4
Open problems
To be useful, a structure theory should make some predictions. I hope that
the methods of this paper will eventually be applicable to start dealing with
issues such as the following:
• k = 3. A 4 × 6 resolution is a Koszul complex.
• k = 4. There are no almost complete intersection Gorenstein ideals.
Equivalently, a 5 × 8 resolution is nonminimal: if X is Gorenstein
codimension 4 and (L1 , . . . , L5 ) generate IX then the first syzygy matrix
M1 has a unit entry, making one of the Li redundant. This is a well
known theorem of Kunz [K], but I want to deduce it by my methods.
• k = 5. Is it true that a 6 × 10 resolution is a hypersurface in a 5 × 5
Pfaffian as in 2.3, Example 2.2?
28
The same question for more general odd k: are hypersurfaces in a
codimension 3 Gorenstein varieties the only cases? Is this even true for
all the known examples in the literature? This might relate to my even
versus odd remark in 3.1.3.
• k = 6. I would like to know whether every case of 7 × 12 resolution is
the known Kustin–Miller unprojection from a codimension 4 complete
intersection divisor in a codimension 3 complete intersection.
• k = 8. As everyone knows, the main case is 9×16. How do we apply the
theory to add anything useful to the huge number of known examples?
There are hints that something along these lines may eventually be possible, but it is not in place yet.
References
[TJ]
Gavin Brown, Michael Kerber and Miles Reid, Fano 3-folds in codimension 4, Tom and Jerry. Part I, Compositio Math. 148 (2012)
1171–1194
[BE1] David Buchsbaum and David Eisenbud, What makes a complex exact? J. Algebra 25 (1973) 259–268
[BE2] David Buchsbaum and David Eisenbud, Algebra structures for finite
free resolutions, and some structure theorems for ideals of codimension 3, Amer. J. Math. 99 (1977) 447–485
[CLP] Fabrizio Catanese, LIU Wenfei, Roberto Pignatelli, The moduli space
of even surfaces of general type with K 2 = 8, pg = 4 and q = 0,
arXiv:1209.0034, 29 pp.
[CR]
A. Corti and M. Reid, Weighted Grassmannians, in Algebraic geometry, de Gruyter, Berlin, 2002, pp. 141–163
[Di]
Duncan Dicks, Surfaces with pg = 3, K 2 = 4 and extensiondeformation theory, 1988, Warwick PhD thesis, 125 + vi pp. (available from my website)
[FH]
W. Fulton and J. Harris, Representation theory. A first course, Grad.
Texts in Math. 129, Springer-Verlag, New York 1991
29
[GW] GOTO Shiro and WATANABE Keiichi, On graded rings. I, J. Math.
Soc. Japan 30 (1978) 179–213
[K]
Ernst Kunz, Almost complete intersections are not Gorenstein rings,
J. Algebra 28 (1974) 111–115
[KM]
A. Kustin and M. Miller, Constructing big Gorenstein ideals from
small ones, J. Algebra 85 (1983) 303–322
[PR]
Stavros Argyrios Papadakis and Miles Reid, Kustin–Miller unprojection without complexes, J. Algebraic Geom. 13 (2004) 563–577,
Preprint math.AG/0011094
[Ki]
M. Reid, Graded rings and birational geometry, in Proc. of algebraic
geometry symposium (Kinosaki, Oct 2000), K. Ohno (Ed.), 1–72, get
from www.warwick.ac.uk/staff/Miles.Reid/3folds
[R1]
M. Reid, Surfaces with pg = 3, K 2 = 4 according to E. Horikawa
and D. Dicks, in Proceedings of Algebraic geometry mini-symposium
(Tokyo, Dec 1989), 1–22
[Va]
Ravi Vakil, Murphy’s law in algebraic geometry: badly-behaved deformation spaces, Invent. Math. 164 (2006) 569–590
Miles Reid,
Mathematics Institute, University of Warwick,
Coventry CV4 7AL, England
e-mail: [email protected]
30
| 0math.AC
|
arXiv:1310.0570v5 [math.AC] 30 Oct 2017
CANONICAL SYSTEMS OF BASIC INVARIANTS FOR UNITARY
REFLECTION GROUPS
NORIHIRO NAKASHIMA, HIROAKI TERAO, AND SHUHEI TSUJIE
Abstract. It has been known that there exists a canonical system for every
finite real reflection group. The first and the third authors obtained an explicit
formula for a canonical system in the previous paper. In this article, we first define
canonical systems for the finite unitary reflection groups, and then prove their
existence. Our proof does not depend on the classification of unitary reflection
groups. Furthermore, we give an explicit formula for a canonical system for every
unitary reflection group.
1. Introduction
Let V be an n-dimensional unitary space, and W ⊆ U(V ) a finite unitary reflection group. Each reflection fixes a hyperplane in V pointwise. Let S denote the
symmetric algebra S(V ∗ ) of the dual space V ∗ , and Sk the vector space consisting
of homogeneous polynomials of degree k together with the zero polynomial. Then
W acts contravariantly on S by (w · f )(v) = f (w −1 · v) for v ∈ V , w ∈ W and
f ∈ S. The action of W on S preserves the degree of homogeneous polynomials,
and W acts also on Sk . The subalgebra R = S W of W -invariant polynomials of S
is generated by n algebraically independent homogeneous polynomials by Chevalley
[2]. A system of such generators is called a system of basic invariants of R.
Let x1 , . . . , xn be an orthonormal basis for V ∗ , and ∂1 , . . . , ∂n the basis for V ∗∗
dual to x1 , . . . , xn . The symmetric algebra of V ∗∗ acts naturally on S as differential
operators
P (e.g., Kane [8, §25-2]). Let c denote the complex conjugate of c ∈ C. For
f = a ca xa ∈ S, a differential operator f ∗ is defined by
(1.1)
f ∗ := f (∂) :=
X
ca ∂ a ,
a
1991 Mathematics Subject Classification. Primary 13A50; Secondary 20F55.
Key words and phrases. basic invariants; invariant theory; finite unitary reflection groups.
1
2
NORIHIRO NAKASHIMA, HIROAKI TERAO, AND SHUHEI TSUJIE
where a = (a1 , . . . , an ) ∈ Zn≥0 , ca ∈ C, xa = xa11 · · · xann , and ∂ a = ∂1a1 · · · ∂nan . Note
that (cf )∗ = cf ∗ , w · (f ∗ ) = (w · f )∗ and w · (f ∗ g) = (w · f )∗(w · g) for c ∈ C, w ∈ W ,
f, g ∈ S.
Flatto and Wiener introduced canonical systems to solve a mean value problem
related to vertices for polytopes in [3, 4, 5]. They proved that there exists a canonical system for every finite real reflection group. Later in [7], Iwasaki gave a new
definition of the canonical systems as well as explicit formulas for canonical systems
for some types of reflection groups. The first and the third authors, in their previous work [9], obtained an explicit formula for a canonical system for every reflection
group. In this article, we extend the definition of canonical systems to the finite
unitary reflection groups as follows.
Definition 1.1. A system {f1 , . . . , fn } of basic invariants is said to be canonical
if it satisfies the following system of partial differential equations:
(1.2)
fi∗ fj = δij
for i, j = 1, . . . , n, where δij is the Kronecker delta.
Our main result is the following existence theorem.
Theorem 1.2. There exists a canonical system for every finite unitary reflection
group.
Our proof of Theorem 1.2 is classification free. Furthermore, we give an explicit
formula (Theorem 4.3) for a canonical system which is also classification free. This
formula is the same as one obtained in [9] for the real case, and we improve the
proof.
The organization of this article is as follows. In Section 2, we introduce Lemma
2.2 which will play an important role in Section 3 when we prove the existence
theorem 1.2. In Section 4, we give an explicit formula for a canonical system.
2. Basic invariants
Let R+ be the ideal of R generated by homogeneous elements of positive degrees,
and I = SR+ the ideal of S generated by R+ . We define a unitary inner product
h·, ·i : S × S → C by
(2.1)
hf, gi = f ∗ g|x=0 = f (∂)g|x=0
(f, g ∈ S),
where x = (x1 , . . . , xn ) and ∂ = (∂1 , . . . , ∂n ). Let eH be the order of the cyclic group
generated by the reflection sH ∈ W corresponding to a reflecting hyperplane H. Fix
CANONICAL SYSTEMS OF BASIC INVARIANTS
3
LH ∈ V ∗ satisfying ker LH = H. Let ∆ denote the product of LeHH −1 as H runs over
the set of all reflecting hyperplanes. Then ∆ is skew-invariant, i.e., w · ∆ = (det w)∆
for any w ∈ W . Set
H := {f ∗ ∆ | f ∈ S}.
(2.2)
The following lemma is obtained by Steinberg [12].
Lemma 2.1. Let f ∈ S be a homogeneous polynomial. Then we have the following:
(1) f ∈ I if and only if f ∗ ∆ = 0,
(2) g ∗f = 0 for all g ∈ I if and only if f ∈ H.
It follows from Lemma 2.1 (2) that I is the orthogonal complement of H with
respect to the inner product (2.1) degreewise.
In the rest of this paper, we assume that W acts on V irreducibly. We fix a
W -stable graded subspace U of S such that S = I ⊕ U. It is known that the
U is isomorphic to the regular representation of W (see Bourbaki [1, Chap. 5
§5 Theorem 2].) Hence the multiplicity of V in U is equal to dimC V = n. Let
π : S → U be the second projection with respect to the decomposition S = I ⊕ U.
Then π is a W -homomorphism. Let {h1 , . . . , hn } be a system of basic invariants
with deg h1 ≤ · · · ≤ deg hn . The multiset of degrees mi := deg hi (i = 1, . . . , n) does
not depend on a choice of basic invariants.
There exists a unique linear map d : S → S ⊗C V ∗ satisfying d(f g) = f d(g) +
gd(f ) for f, g ∈ S and dL := 1 ⊗ L ∈ C ⊗C V ∗ for L ∈ V ∗ . The map d is called the
differential map. The differential 1-form dh is expressed as
dh =
n
X
∂j h ⊗ xj =
n
X
(∂j h)dxj
j=1
j=1
for h ∈ S. Note that dh is invariant if h is invariant. Define a W -homomorphism
ε : (S ⊗C V ∗ )W → R+
by
ε
n
X
j=1
hj dxj
!
=
n
X
xj hj .
j=1
Then ε ◦ d(h) = (deg h)h for any homogeneous polynomial h. The projection π :
S →P
U is extended to P
a W -homomorphism π̃ : (S ⊗ V ∗ )W → (U ⊗ V ∗ )W defined
n
by π̃( j=1 gj ⊗ xj ) := nj=1 π(gj ) ⊗ xj .
4
NORIHIRO NAKASHIMA, HIROAKI TERAO, AND SHUHEI TSUJIE
Lemma 2.2. Let {h1 , . . . , hn } be a system of basic invariants. Put fi := (ε◦ π̃)(dhi ),
and {f1 , . . . , fn } is a system of basic invariants.
Proof. Since h1 , . . . , hn are invariants, so are the 1-forms dh1 , . . . , dhn . Thus each fi
is invariant because both ε and π̃ are W -homomorphisms.
Next we prove that
Pn{f1 , . . . , fn } is a system of basic invariants. Define fij :=
π(∂j hi ). Then fi = j=1 xj fij . For j = 1, . . . , n, we express ∂j hi = fij + rij for
P
P
some rij ∈ I. Then rij = ℓk=1 hk gijk for some gijk ∈ S. Put ri := nj=1 xj rij for
i = 1, . . . , n. Then we have
mi hi =
n
X
xj ∂j hi = fi + ri .
j=1
Since fi is invariant, the polynomial ri = mi hi − fi is also invariant. This implies
ri = ri♯ =
n
ℓ
X
X
k=1
xj gi,j,k
j=1
!♯
hk ∈ I 2 ∩ R,
where ♯ denotes the averaging operator, i.e.,
f♯ =
1 X
w·f
|W | w∈W
for f ∈ S. Thus we have ∂j ri ∈ I for i, j ∈ {1, . . . , n}. Let J(g1 , . . . , gn ) denote the
Jacobian for g1 , . . . , gn ∈ S. Then
J(m1 h1 , . . . , mn hn ) = det[∂j fi + ∂j ri ]i,j
≡ det[∂j fi ]i,j = J(f1 , . . . , fn )
(mod I).
It immediately follows that ∆ ∈
/ I by Lemma 2.1 (1). Since J(h1 , . . . , hn ) is a nonzero
constant multiple of ∆, we obtain J(f1 , . . . , fn ) ∈
/ I, and thus J(f1 , . . . , fn ) 6= 0.
By the Jacobian criterion (e.g., [6, Proposition 3.10]), {f1 , . . . , fn } is algebraically
independent. Therefore {f1 , . . . , fn } is a system of basic invariants because deg fi =
deg hi for i = 1, . . . , n.
Remark. There exists a W -stable subspace U ′ of S such that S = I ⊕ U ′ and
df1 , . . . , dfn ∈ (U ′ ⊗C V ∗ )W . However, we do not know whether U ′ coincides with U
or not. In section 3, we see that U and U ′ coincide when U = H.
CANONICAL SYSTEMS OF BASIC INVARIANTS
5
3. Existence of a canonical system
In this section, we prove Theorem 1.2 which is the existence theorem of a canonical system. The following lemma is widely known.
Lemma 3.1. Let g ∈ S be a homogeneous polynomial, and put gj := ∂j g for j =
1, . . . , n. Then, for any h ∈ S, we have
g ∗ (xj h) = xj g ∗h + gj∗h.
(3.1)
Proof. We only need to verify the assertion when g is a monomial. We verify it by
induction on deg g. Let a = (a1 , . . . , an ) be a multi-index with |a| = deg g. Then
∂ a (xj h) = ∂ a−ej ∂j (xj h) = ∂ a−ej h + ∂ a−ej (xj ∂j h)
= ∂ a−ej h + xj ∂ a−ej ∂j h + (aj − 1)∂ a−2ej ∂j h
= xj ∂ a h + aj ∂ a−ej h.
By Lemma 2.1, I is the orthogonal complement of H with respect to the inner
product (2.1), and the W -stable graded space S is decomposed into the direct sum
of the W -stable graded subspaces I and H, i.e., S = I ⊕ H. Let π : S → H be the
second projection with respect to the decomposition S = I ⊕ H. Let h1 , . . . , hn be
an arbitrary system of basic invariants. Put fijP:= π(∂j hi ) ∈ H = {f ∗ ∆ | f ∈ S}
n
for i, j = 1, . . . , n, and fi := (ε ◦ π̃)(dhi ) =
j=1 xj fij for i = 1, . . . , n. Then
{f1 , . . . , fn } is a system of basic invariants by Lemma 2.2. We are now ready to give
a proof of Theorem 1.2.
Let g ∈ R+ be a homogeneous invariant polynomial with deg g < mi . By using
Lemma 2.1 and Lemma 3.1 we obtain
n
n
n
X
X
X
∗
∗
∗
∗
g fi =
g (xj fij ) =
xj g fij + gj fij =
gj∗ fij ∈ H.
j=1
j=1
j=1
∗
Meanwhile, g fi is an invariant polynomial of positive degree since g and fi are
invariant and deg g < mi . Therefore we have
g ∗ fi ∈ H ∩ I = {0}.
In particular, when g = fj with deg fj < mi , we have fj∗ fi = 0. It immediately
follows that fj∗ fi = 0 if deg fj > deg fi . Applying the Gram-Schmidt orthogonalization with respect to the inner product (2.1), we obtain a canonical system of basic
invariants. This completes our proof of Theorem 1.2.
The subspace spanned by a canonical system can be characterized as follows:
6
NORIHIRO NAKASHIMA, HIROAKI TERAO, AND SHUHEI TSUJIE
Proposition 3.2. Let {f1 , . . . , fn } be a canonical system and F := hf1 , . . . , fn iC .
Then
(1) F =
L∞
k=1 {f
∈ Rk | g ∗ f = 0 for g ∈ Rℓ with 0 < ℓ < k}, where Rk := R∩Sk ,
(2) hdf1, . . . , dfn iC = (H ⊗C V ∗ )W ,
(3) F = ε((H ⊗C V ∗ )W ) .
Proof. Define
∞
M
{f ∈ Rk | g ∗ f = 0 for g ∈ Rℓ with 0 < ℓ < k}.
G :=
k=1
Let f ∈ G be a homogeneous polynomial. For any g ∈ I, it is not hard to see
that g ∗ (∂j f ) = ∂j (g ∗f ) = 0. Thus we have ∂j f ∈ H for j = 1, . . . , n by Lemma
2.1. This implies d(G) ⊆ (H ⊗C V ∗ )W . The inclusion F ⊆ G follows immediately
because {f1 , . . . , fn } is a canonical system. Hence we have the following commutative
diagram:
d
⊆
(3.2)
(H ⊗C V ∗ )W
⊆
G →
d
F → hdf1 , . . . , dfn iC .
Since ker(d) = C and G does not contain any nonzero constant, the horizontal
maps d are both injective. Recall that V is an irreducible representation and that H
affords the regular representation of W . Hence we have dim(H ⊗C V ∗ )W = dim V =
n because (H ⊗C V ∗ )W ≃ HomW (V, H); this isomorphism is referred from [10] or
the proof of [11, Lemma 6.45]. By comparing the dimensions, we have F = G, which
is (1). Sending both sides of this equality by d, we obtain
(3.3)
hdf1 , . . . , dfn iC = d(F ) = d(G) = (H ⊗C V ∗ )W .
So the equality (2) is proved. Moreover, we have
(3.4)
hf1 , . . . , fn iC = F = G = ε((H ⊗C V ∗ )W )
by applying ε to (3.3). This verifies (3).
CANONICAL SYSTEMS OF BASIC INVARIANTS
7
4. An explicit construction of a canonical system
The following is a key to our explicit formula for a canonical system.
Definition 4.1 (c.f. [9]). Define a linear map
φ : S→H
by φ(f ) := (f ∗ ∆)∗ ∆ for f ∈ S. The map φ induces a W -homomorphism
φ̃ : (S ⊗ V ∗ )W → (H ⊗ V ∗ )W
P
P
defined by φ̃( f ⊗ x) := φ(f ) ⊗ x.
One has
w · φ(f ) = ((w · f )∗ (w · ∆))∗ (w · ∆) = ((w · f )∗ (det(w)∆))∗ (det(w)∆)
=det(w) det(w)((w · f )∗ ∆)∗ (∆) = φ(w · f )
for w ∈ W and f ∈ S. Therefore φ is a W -homomorphism, and so is φ̃.
Let {h1 , . . . , hn } be an arbitrary system of basic invariants, and assume deg hi =
mi for i = 1, . . . , n. Let {f1 , . . . , fn } be a canonical system with deg fi = mi . We
have already shown in Proposition 3.2 (1) that (H ⊗ V ∗ )W = hdf1 , . . . , dfn iC .
Lemma 4.2. . The restriction
φ̃|hdh1 ,...,dhn iC : hdh1 , . . . , dhn iC → (H ⊗ V ∗ )W = hdf1 , . . . , dfn iC
is isomorphic.
Proof. It is enough to prove the injectivity. Fix h ∈ hh1 , . . . , hn iC with φ̃(dh) = 0. It
follows from Lemma 2.1 that ker φ̃ = (I ⊗C V ∗ )W . Then we have dh ∈ (I ⊗C V ∗ )W .
At the same time, since {f1 , . . . , fn } is a system of basic invariants, we may write
h=
n
X
λk fk + P,
k=1
where λk ∈ C and P ∈ I ∩ R. Then the 1-form dP lies in (I ⊗C V ∗ )W , and
dfk ∈ (H ⊗C V ∗ )W for k = 1, . . . , n by Proposition 3.2. Hence we have
2
n
X
λk dfk = dh − dP ∈ (H ⊗C V ∗ )W ∩ (I ⊗C V ∗ )W = {0}.
k=1
This implies λk = 0 for all k = 1, . . . , n since {df1 , . . . , dfn } is linearly independent
over C. Thus we have h = P ∈ I 2 ∩ R. The algebraic independence of h1 , . . . , hn
implies hh1 , . . . , hn i ∩ I 2 = {0}. Therefore h = 0.
8
NORIHIRO NAKASHIMA, HIROAKI TERAO, AND SHUHEI TSUJIE
The image of φ̃|hdh1 ,...,dhn iC coincides with hdf1 , . . . , dfn iC by Lemma 4.2. Therefore
we have a chain of the linear maps:
(4.1)
φ̃
d
ε
hh1 , . . . , hn iC → hdh1 , . . . , dhn iC → hdf1 , . . . , dfn iC → hf1 , . . . , fn iC .
The image of {h1 , . . . , hn } by the composition of all the maps in (4.1) forms a basis
for hf1 , . . . , fn iC . Thus we have the following explicit formula for a canonical system
of basic invariants.
Theorem 4.3 (c.f. [9]). Let h1 , . . . , hn be an arbitrary system of basic invariants.
Applying the Gram-Schmidt orthogonalization to
)
(
n
X
xj φ(∂j hi ) i = 1, . . . , n
ε ◦ φ̃(dhi ) =
j=1
with respect to the inner product (2.1), we obtain a canonical system of basic invariants.
Remark. Theorem 4.3 asserts the same formula as [9, Theorem 3.4] for the real
case. In [9], to prove the theorem, we showed the symmetricity of φ with respect to
the inner product (2.1) and considered eigenvectors of φ̃. In contrast, the proof of
this paper does not require these arguments.
References
[1] N. Bourbaki, Groupes et Algèbres de Lie. Chapitres 4, 5 et 6, Hermann, Paris, 1968.
[2] C. Chevalley, Invariants of finite groups generated by reflections. Amer. J. Math, 77(1955),
no. 4, 778–782.
[3] L. Flatto, Basic sets of invariants for finite reflection groups. Bull. Amer. Math. Soc. 74(1968),
no. 4, 730–734.
[4] L. Flatto, Invariants of finite reflection groups and mean value problems. II. Amer. J. Math.
92(1970), 552–561.
[5] L. Flatto and M. M. Wiener, Invariants of finite reflection groups and mean value problems.
Amer. J. Math. 91(1969), no. 3, 591–598.
[6] J. E. Humphreys, Reflection Groups and Coxeter Groups. Cambridge Univ. Press, Cambridge,
New York, 1990.
[7] K. Iwasaki, Basic invariants of finite reflection groups. J. Algebra. 195(1997), no. 2, 538–547.
[8] R. Kane, Reflection groups and invariant theory. CMS Books in Mathematics/Ouvrages de
Mathématiques de la SMC, 5, Springer-Verlag, New York, 1997.
CANONICAL SYSTEMS OF BASIC INVARIANTS
9
[9] N. Nakashima and S. Tsujie, A canonical system of basic invariants of a finite reflection group.
J. Algebra, 406(2014), 143–153.
[10] P. Orlik and L. Solomon, Unitary reflection groups and cohomology. Invent. Math. 59(1980),
no. 1, 77–94.
[11] P. Orlik and H. Terao, Arrangements of Hyperplanes. Grundlehren dermatematischen Wissenschaften 300, Springer-Verlag, 1992.
[12] R. Steinberg, Differential equations invariant under finite reflection groups. Trans. Amer.
Math. Soc. 112(1964), no. 3, 392–400.
School of Information Environment, Tokyo Denki University, Inzai, 270-1382,
Japan
E-mail address: [email protected]
Department of Mathematics, Hokkaido University, Sapporo, 060-0810, Japan
E-mail address: [email protected]
Department of Mathematics, Hokkaido University, Sapporo, 060-0810, Japan
E-mail address: [email protected]
| 0math.AC
|
When Hypermutations and Ageing Enable Artificial Immune
Systems to Outperform Evolutionary Algorithms1
Dogan Corus, Pietro S. Oliveto, Donya Yazdani
Department of Computer Science, University of Sheffield
Sheffield, UK
arXiv:1804.01314v1 [cs.NE] 4 Apr 2018
S1 4DP
Abstract
We present a time complexity analysis of the Opt-IA artificial immune system (AIS). We first
highlight the power and limitations of its distinguishing operators (i.e., hypermutations with
mutation potential and ageing) by analysing them in isolation. Recent work has shown that
ageing combined with local mutations can help escape local optima on a dynamic optimisation benchmark function. We generalise this result by rigorously proving that, compared to
evolutionary algorithms (EAs), ageing leads to impressive speed-ups on the standard Cliff
benchmark function both when using local and global mutations. Unless the stop at first
constructive mutation (FCM) mechanism is applied, we show that hypermutations require
exponential expected runtime to optimise any function with a polynomial number of optima.
If instead FCM is used, the expected runtime is at most a linear factor larger than the upper bound achieved for any random local search algorithm using the artificial fitness levels
method. Nevertheless, we prove that algorithms using hypermutations can be considerably
faster than EAs at escaping local optima. An analysis of the complete Opt-IA reveals that
it is efficient on the previously considered functions and highlights problems where the use
of the full algorithm is crucial. We complete the picture by presenting a class of functions
for which Opt-IA fails with overwhelming probability while standard EAs are efficient.
Keywords: Artificial Immune Systems, Opt-IA, Runtime Analysis, Evolutionary
Algorithms, Hypermutation, Ageing
1. Introduction
Artificial Immune Systems (AIS) are a class of bio-inspired computing techniques that
take inspiration from the immune system of vertebrates [2]. Burnet’s clonal selection theory
1
An extended abstract of this paper has been published at the 2017 Genetic and Evolutionary Computation Conference [1].
Email addresses: [email protected] (Dogan Corus), [email protected] (Pietro S.
Oliveto), [email protected] (Donya Yazdani)
Preprint submitted to Journal of LATEX Templates
April 5, 2018
[3] has inspired various AIS for function optimisation. The most popular ones are Clonalg
[4], the B-Cell Algorithm [5] and Opt-IA [6].
After numerous successful applications of AIS were reported, a growing body of theoretical work has gradually been built to shed light on the working principles of AIS. While
initial work derived conditions that allowed to prove whether an AIS converges or not [7],
nowadays rigorous time complexity analyses of AIS are available. Initial runtime analyses
focused on studying the performance of typical AIS operators in isolation to explain when
and why they are effective. Such studies have been extensively performed for the contiguous somatic hypermutation operator employed by the B-Cell algorithm [8, 9], the inversely
proportional hypermutation operator of Clonalg [10, 11] and the ageing operator used by
Opt-IA [12, 13, 14]. These studies formed a foundational basis which allowed the subsequent
analysis of the complete B-Cell algorithm as used in practice for standard NP-hard problems
[15, 16].
Compared to the relatively well understood B-Cell algorithm, the theoretical understanding of other AIS for optimisation is particularly limited. In this paper we consider
the complete Opt-IA algorithm [17, 6]. This algorithm has been shown to be successful at
optimising problems such as protein structure prediction [6], graph colouring [18] and hitting
set [19]. The main distinguishing features of Opt-IA compared to other AIS is their use of
an ageing operator and of hypermutations with mutation potentials 2 . In this work we will
first analyse the characteristics of these operators respectively in isolation and afterwards
consider a simple, but complete, Opt-IA algorithm. The aim is to highlight function characteristics for which Opt-IA and its main components are particularly effective, hence when
it may be preferable to standard Evolutionary Algorithms (EAs).
The idea behind the ageing operator is that old individuals should have a lower probability of surviving compared to younger ones. Ageing was originally introduced as a mechanism
to maintain diversity. Theoretical analyses have strived to justify this initial motivation because the new random individuals (introduced to replace old individuals) typically have very
low fitness and die out quickly. On the other hand, it is well understood that ageing can
be used as a substitute for a standard restart strategy if the whole population dies at the
same generation [12] and to escape local optima if most of the population dies except for one
survivor that, at the same generation, moves out of the local optimum [14]. This effect was
shown for a Random Local Search (RLS) algorithm equipped with ageing on the Balance
dynamic optimisation benchmark function. An evolutionary algorithm (EA) using standard
bit mutation (SBM) [20, 21] and ageing would not be able to escape the local optima due
to their very large basin of attraction. Herein, we carefully analyse the ability of ageing to
escape local optima on the more general Cliff benchmark function and show that using the
operator with both RLS and EAs can make a difference between polynomial and exponential
runtimes.
Hypermutation operators are inspired by the high mutation rates occurring in the immune system. In Opt-IA the mutation potential is linear in the problem size and in different
2
The cloning operator is a common feature in AIS and the hypermacromutation operator is essentially
the same as the well studied contiguous somatic mutation operator of the B-Cell algorithm.
2
algorithmic variants may either be static or increase by a factor that is proportional to
the fitness of the solution (i.e., the b-cell) undergoing the mutation or decrease by a factor
that is inversely proportional to the fitness. The theoretical understanding of hypermutations with mutation potential is very limited. To the best of our knowledge the only runtime
analysis available is [22], where inversely proportional hypermutations were considered, with
and without the stop at first constructive mutation (FCM) strategy3 . The analysis revealed
that, without FCM, the operator requires exponential runtime to optimise the standard
OneMax function, while by using FCM the algorithm is efficient. We consider a different
hypermutation variant using static mutation potentials and argue that it is just as effective
if not superior to other variants. We first show that the use of FCM is essential by rigorously
proving that a (1+1) EA equipped with hypermutations and no FCM requires exponential
expected runtime to optimise any function with a polynomial number of optima. We then
consider the operator with FCM for any objective function that can be analysed using the
artificial fitness level (AFL) method [20, 21] and show an upper bound on its runtime that
is at most by a linear factor larger than the upper bound obtained for any RLS algorithm
using AFL. To achieve this, we present a theorem that allows to derive an upper bound
on the runtime of sophisticated hypermutation operators by analysing much simpler RLS
algorithms with an arbitrary neighbourhood. As a result, all existing results achieved via
AFL for RLS may be translated into upper bounds on the runtime of static hypermutations.
Finally, we use the standard Cliff and Jump benchmark functions to show that hypermutations can achieve considerable speed-ups for escaping local optima compared to well
studied EAs.
We then concentrate on the analysis of the complete Opt-IA algorithm. The standard
Opt-IA uses both hypermutations and hypermacromutation (both with FCM) mainly because preliminary experimental studies for trap functions indicated that this setting led to
the best results [17, 6]. Our analysis reveals that it is unnecessary to use both operators for
Opt-IA to be efficient on trap functions. To this end, we will consider the simple version
using only static hypermutations as in [17]. We will first consider the algorithm with the
simplification that we allow genotypic duplicates in the population, to simplify the analysis
and enhance the probabilities of ageing to create copies and escape from local optima. Afterwards we extend the analysis to the standard version using a genotype diversity mechanism.
Apart from proving that the algorithm is efficient for the previously considered functions,
we present a class of functions called HiddenPath, where it is necessary to use both ageing
and hypermutations in conjunction, hence the use of Opt-IA in its totality is crucial. Having
shown several general settings where Opt-IA is advantageous compared to standard EAs, we
conclude the paper by pointing out limitations of the algorithm. In particular, we present
a class of functions called HyperTrap that is deceptive for Opt-IA while standard EAs
optimise it efficiently w.o.p.
Compared to its conference version [1], this paper has been improved in several ways.
Firstly, we have extended our analyses of the Ageing operator and Opt-IA to include the
3
An analysis of high mutation rates in the context of population-based evolutionary algorithms was
performed in [23].
3
Algorithm 1 Opt-IA [6]
1: Initialisation: t = 0
Create P (t) = {x1 , ..., xµ }, a population of µ b-cells uniformly at random.
xage
= 0 for i = {1, ...µ}
i
2: while Termination condition is not reached do
3:
Cloning: \\clones each b-cell dup times.
P (clo) = Cloning (P (t) , dup)
4:
Variation: \\hypermutates each clone.
P (hyp) = Hypermutation (P (clo) , c, n)
P (macro) = Hypermacromutation (P (clo) , n)
5:
Ageing: \\increases the age of b-cells and removes the old ones.
Ageing (P (t) , P (hyp) , P (macro) , τ )
6:
Selection: \\(µ + λ) – selection with genotype diversity
P (t+1) = Selection (P (t) , P (hyp) , P (macro) ) :
a) Remove genotypic duplicates in P (t+1) ,
b) While |P (t+1) | > µ, remove the b-cell with the lowest fitness breaking ties
uniformly at random,
c) While |P (t+1) | < µ, create a new b-cell uniformly at random.
7:
t := t + 1
8: end while
genotype diversity mechanism as in the algorithm proposed in the literature [17, 6]. Another
addition is the introduction of a class of functions where Opt-IA fails to find the optimum
efficiently allowing us to complete the picture by highlighting problem characteristics where
Opt-IA succeeds and where it does not. Finally, this paper includes some proofs which were
omitted from the conference version due to page limitations.
The rest of the paper is structured as follows. In Section 2, we introduce and define OptIA and and its operators. In Section 3, we present the results of our analyses of the Static
Hypermutation operator in a simple framework to shed light on its power and limitations
in isolation. In Section 4, we present our analyses of the Ageing operator in isolation and
highlight its ability to escape from local optima. In Section 5, we present the results of
our analyses of the complete algorithm. In Section 6, we extend the analyses to include
the genotype diversity mechanism as applied in the standard Opt-IA [17, 6]. Finally, we
conclude the paper with a discussion of the results and directions for future work.
2. Preliminaries
In this section we present the standard Opt-IA as applied in [6] for the maximisation of
f : {0, 1}n → R. Its pseudo-code is given in Algorithm 1. Opt-IA is initialised with a population of µ b-cells generated uniformly at random, each with age = 0 . In each generation,
the algorithm creates a new parent population consisting of dup copies of each b-cell (i.e.,
Cloning) which will be the subject of variation. The pseudo-code of the Cloning operator is
given in Algorithm 2.
4
Algorithm 2 Cloning (P (t) , dup)
1: for all xi ∈ P (t) do
2:
Clone xi dup times
3:
Add the clones to P (clo)
4: end for
The variation stage in Opt-IA uses a hypermutation operator with mutation potential
sometimes followed by hypermacromutation [6], sometimes not [17]. If both are applied,
they act on the clone population (i.e., not in sequence) such that they generate µ mutants
each. The number of bits M that are flipped by the hypermutation operator is determined by
a function called mutation potential. Three different potentials have been considered in the
literature: static where the number of bits that are flipped is linear in the problem size and
does not depend on the fitness function4 , proportional (i.e., a linear number of bits are always
flipped but increasing proportionally with the fitness of the mutated b-cell) and inversely
proportional (i.e., to the fitness of the mutated b-cell). The latter potential was previously
theoretically analysed in [22]. What is unclear from the literature is whether the M bits to
be flipped should be distinct or not and, when using FCM, whether a constructive mutation
is a strictly improving move or whether a solution of equal fitness suffices. In this paper we
will consider the static hypermutation operator with pseudo-code given in Algorithm 3. In
particular, the M flipped bits will always be distinct and both kinds of constructive mutation
will be considered. At the end of the variation stage all created individuals have age = 0
if their fitness is higher than that of their parent cell, otherwise they inherit their parent’s
age. Then the whole population (i.e., parents and offspring) undergoes the ageing process in
which the age of each b-cell is increased by one. Additionally, the ageing operator removes
old individuals. Three methods have been proposed in the literature: static ageing, which
deterministically removes all individuals who exceed age τ ; stochastic ageing, which removes
each individual at each generation with probability pdie ; and the recently introduced hybrid
ageing [14], where individuals have a probability pdie of dying only once they reach an age
of τ . In [14] it was shown that the hybrid version allows to escape local optima, hence we
employ this version in this paper and give its pseudo-code in Algorithm 4.
The generation ends with a selection phase. If the total number of b-cells that have
survived the ageing operator is larger than µ, then a standard (µ + λ) selection scheme is
used with the exception that genotype duplicates are not allowed. If the population size
is less than µ, then a birth phase fills the population up to size µ by introducing random
b-cells of age = 0. In this paper we give evidence that disallowing genotypic duplicates may
be detrimental because, as we will show, genotypic copies may help the ageing operator to
escape local optima more efficiently.
4
In [17] the mutation potential is declared to be a constant 0 < c ≤ 1. This is obviously a typo: the
authors intended the mutation potential to be cn, where 0 < c ≤ 1.
5
Algorithm 3 Static hypermutation (P (clo) , c, n)
1: M = cn (Mutation potential)
2: for all xi ∈ P (clo) do
3:
if FCM is not used then
4:
create yi by flipping M distinct bits selected uniformly at random
5:
else
6:
create yi by flipping at most M distinct bits selected uniformly at random one after
another until a constructive mutation happens
7:
end if
8:
If f (yi ) > f (xi ) then yiage = 0, else yiage = xage
i
9:
Add yi to P (hyp)
10: end for
Algorithm 4 Hybrid ageing (P (t) , P (hyp) , µ, τ )
1: for all xi ∈ (P (t) , P (hyp) ) do
2:
xage
= xage
+1
i
i
age
3:
if xi > τ then
4:
remove xi with probability pdie = 1 − 1/µ
5:
end if
6: end for
3. Static Hypermutation
The aim of this section is to highlight the power and limitations of the static hypermutation operator in isolation. For this purpose we embed the operator into a minimal AIS
framework that uses a population of only one b-cell and creates exactly one clone per generation. The resulting (1+1) IAhyp , depicted in Algorithm 6, is essentially a (1+1) EA that
applies the static hypermutation operator instead of using SBM. We will first show that,
without the use of FCM, hypermutations are an inefficient variation operator for virtually
any optimisation function of interest. From there on we will only consider the operator
equipped with FCM. Then we will prove that the (1+1) IAhyp has a runtime that is at most
a linear factor larger than that obtained for any RLS algorithm using the artificial fitness
levels method. Intuitively, this happens because, if the operator flips some bits that were
already set correctly at the beginning of its iteration, then it will perform at most cn useless
fitness function evaluations before one hypermutation process is concluded. We formalise
this result in Theorem 2 when FCM only accepts strict improvements, and formally call the
hyp
algorithm (1+1) IAhyp
> in this case. For the (1+1) IA≥ , where FCM also considers points
of equal fitness as constructive solutions, we prove in Theorem 3 that the algorithm cannot
be too slow compared to the standard RLS1 (i.e., flipping one bit per iteration). We show
that the presented results are tight for some standard benchmark functions by proving that
the (1+1) IAhyp has expected runtimes of Θ(n2 log n) for OneMax and Θ(n3 ) for LeadingOnes, respectively, versus the expected Θ(n log n) and Θ(n2 ) fitness function evaluations
6
Algorithm 5 Selection (P (t) , P (hyp) , µ)
1: P (t+1) = P (t) ∪ P (hyp)
2: if genotype diversity is used then
3:
Do not accept any offspring with the same genotype as individuals in P (t)
4: end if
5: if |P (t+1) | > µ then
6:
Remove the |P (t+1) | − µ individuals with the lowest fitness breaking ties uniformly
at random.
7: end if
8: if |P (t+1) | < µ then
9:
Add µ − |P (t+1) | individuals initialised uniformly at random.
10: end if
Algorithm 6 (1 + 1)IAhyp (c, n)
1: Initialisation (µ = 1)
2: while an optimum is not found do
3:
P (clo) = Cloning (P (t) , dup = 1)
4:
P (hyp) = Static hypermutation (P (clo) , c, n)
5:
If f (P (hyp) ) ≥ f (P (t) ), then P (t+1) := P (hyp) . Else, P (t+1) := P (t)
6:
t := t + 1
7: end while
required by RLS1 [20]. Nevertheless, we conclude the section by showing for the standard
benchmark functions Jump and Cliff that the (1+1) IAhyp can be particularly efficient on
functions with local optima that are generally difficult to escape from.
We start by highlighting the limitations of static hypermutation when FCM is not used.
Since M = cn distinct bits have to be flipped at once, the outcome of the hypermutation
operator is characterised by a uniform distribution over the set of all solutions which have
Hamming distance M to the parent. Since M is linear in n, the size of this set of points is
exponentially large and thus the probability of a particular outcome is exponentially small.
In the following theorem, we formalise this limitation.
Theorem 1. For any function with a polynomial number of optima, the (1+1) IAhyp without
FCM needs expected exponential time to find any of the optima.
Proof. We analyse the expected time of the last step before an optimal solution is found. To
do this we optimistically assume that all the optima are at Hamming distance M = cn from
the current b-cell. Otherwise, if no optima were at Hamming distance M, the probability
of reaching an optimum would be zero. Then,
given that the number of different points at
n
Hamming distance cn from any point is cn
and they all are reached with equal probability,
≤ poly(n)
=
the probability of finding this optimum in the last step, for any c 6= 1, is p ≤ poly(n)
n
eΩ(n)
(cn
)
e−Ω(n) . By a simple waiting time argument, the expected time to reach any optimum is at
7
least eΩ(n) . If c = 1, all the bits will be flipped by the operator. As a result the algorithm
will only ever see the initial point and its exact reverse. Hence, an optimum is only found if
the b-cell is randomly initialised on one of the optima or on the reverse of one of the optima.
This happens with probability p = (2 · poly(n)) /(2−n ) = 2−Ω(n) which means the expected
time in this case is also at least in the order of 2Ω(n) .
The theorem explains why poor results were achieved in previous experimental work both
on benchmark functions and real world applications such as the hard protein folding problem
[17, 6]. The authors indeed state that “With this policy, however, and for the problems which
are faced in this paper, the implemented IA did not provide good results” [6]. Theorem 1
shows that this is the case for any optimisation function of practical interest. In [22] it had
already been shown that inversely proportional hypermutations cannot optimise OneMax
in polynomial time. Although static hypermutations are the focus of this paper, we point out
that Theorem 1 can easily be extended to both the inversely proportional hypermutations
considered in [22] and to the proportional hypermutations from the literature [17]. From
now on we will only consider hypermutations coupled with FCM. We start by showing that
hypermutations cannot be too slow compared to local search operators. We first state and
prove the following helper lemma.
Lemma 1. The probability that the static hypermutation applied to x ∈ {0, 1}n either evaluates a specific y ∈ {0, 1}n with Hamming distance k ≤ cn to x (i.e., event Ey ), or that it stops
−1
earlier on a constructive mutation (i.e., event Ec ) is lower bounded by P r{Ey ∨ Ec } ≥ nk .
Moreover, if there
are no constructive mutations with Hamming distance smaller than k, then
n −1
P r{Ey } = k .
Proof. Since the bits to be flipped are picked without replacement, each successive bit-flip
increases the Hamming distance between the current solution and the original solution by
one. Without loss of generality, let π = (π1 , π2 , . . . , πcn ) be the sequence of bit-flips that
will be executed by the static hypermutation, which are sampled uniformly from the space
of all permutations of n bit positions. The lower bound is based on the fact that the first
k bit positions in π have nk different and equally probable outcomes. Since the only event
that can prevent the static hypermutation to evaluate the k-th solution in the sequence is
the discovery of a constructive mutation in one of the first k − 1 evaluations, the probability
−1
P r{Ey ∨ Ec } is larger than nk . If no such constructive mutation exists (i.e. P r{Ec } = 0)
−1
then, the probability P r{Ey } is exactly equal to nk .
We are ready to show that static hypermutations cannot be too slow compared to any
upper bound obtained by applying the artificial fitness levels (AFL) method on the expected
runtime of the (1+1) RLSk which flips exactly k bits to produce a new solution and applies
non-strict elitist
S selection. AFL require a partition of the search space X into mutually
exclusive sets i∈{1,...,m} Ai = X such that ∀i < j, x ∈ Ai ∧ y ∈ Aj =⇒ f (x) < f (y).
The expected runtime of a (1 + 1) algorithm A with variation operator Var(x): X → X to
8
solve any function defined on X can be upper bounded via AFL by E(T ) ≤
!
m
S
pi = min P r{Var(x) ∈
Aj } [20, 21].
x∈Ai
Pm
1
j=1 pi ,
where
j=i+1
AF L
Theorem 2. Let E T(1+1)RLS
be any upper bound on the expected runtime of the (1 +
k
1)RLSk established via the artificial fitness levels method. Then, E T(1+1)IAhyp ≤ cn ·
>
AF L
E T(1+1)RLSk .
Proof. Let the function cj (x) for solution x ∈ Ai return the number of solutions which are
at Hamming distance j away from x and belong to set Ak for some k > i. The upper bound
on the expected runtime
of the(1+1) RLSk to solve any function obtained by applying the
Pm 1
n
AF L
AFL method is: E T(1+1)RLS
≤
,
where
p
=
min
c
(x)/
. Since FCM wastes
i
k
j=1 pi
k
k)
x∈Ai
at most cn bit mutations when it fails to improve, to prove our claim it is sufficient to show
that for any current
solution x ∈ Ai , the probability that (1+1) IAhyp
> finds an improvement
n
is at least ck (x)/ k . The theorem follows from Lemma 1 and the definition of a constructive
mutation for the (1+1) IAhyp
> , since for each one of the ck (x) search points, the probability
−1
of either finding it or finding a constructive mutation is lower bounded by nk .
A similar result with the additional restriction that k = 1 can be obtained for the
(1+1) IAhyp
≥ .
AF L
Theorem 3. E T(1+1)IAhyp ≤ cn · E T(1+1)RLS1 ,
≥
AF L
where E T(1+1)RLS1 is any upper bound on the expected runtime of the (1 + 1)RLS1 established via the artificial fitness levels method.
Proof. The probability that static hypermutation produces a particular Hamming neighbour
of the input solution in the first mutation step is 1/n, which is equal to the probability that
RLS1 produces the same solution. The theorem follows from the fact that in every failure
to obtain an improvement in the first step, static hypermutation wastes at most cn fitness
evaluations.
In the following we show that the upper bounds of the previous theorems are tight for
well-known benchmark functions.
hyp
Theorem 4. The
of the (1 + 1)IAhyp
> and of the (1 + 1)IA≥ to optimise
Pnexpected runtime
OneMax(x):= i=1 xi is Θ(n2 log n).
Proof. The upper bounds for the > and ≥ FCM selection versions follow respectively from
theorems 2 and 3 since it is easy to derive an upper bound of O(n log n) for RLS using
AFL [20]. For the lower bound of the > FCM selection version, we follow the analysis in
Theorem 3 of [22] for inversely proportional hypermutation (IPH) with FCM to optimise
9
ZeroMin. The proof there relies on IPH wasting cn function evaluations every time it fails
to find an improvement. This is obviously also true for static hypermutations (albeit for a
different constant c), hence the proof also applies to our algorithm. The lower bound also
holds for the ≥ FCM selection algorithm as this algorithm cannot be faster on OneMax
by accepting solutions of equal fitness (i.e., the probability of finding an improved solution
does not increase).
We now turn to the LeadingOnes function, which simply returns the number of consecutive 1-bits before the first 0-bit.
Pn Qi
Theorem 5. The expected runtime of the (1 + 1)IAhyp
≥ on LeadingOnes:=
i=1
j=1 xi
3
is Θ(n ).
Proof. The upper bound is implied by Theorem 3 because AFL gives an O(n2 ) runtime of
RLS for LeadingOnes [20]. Let E[fi ] be the expected number of fitness function evaluations
until an improvement is made, considering that the initial solution has i LeadingOnes.
The initial solutions consist of i leading 1-bits, followed by a 0-bit and n − i − 1 bits which
each can be either one or zero with equal probability. Let events E1 , E2 and E3 be that the
first mutation step flips one of the leading ones (with probability i/n), the first 0-bit (with
probability 1/n) or any of the remaining bits (with probability (n−i−1)/n), respectively. If
E1 occurs, then the following mutation steps cannot reach any solution with fitness value i or
higher and all cn mutation steps are executed. Since no improvements have been achieved,
the remaining expected number of evaluations will be the same as the initial expectation
(i.e., E[fi |E1 ] = cn + E[fi ]). If E2 occurs, then a new solution with higher fitness value is
acquired and the mutation process stops (i.e., E[fi |E2 ] = 1). However if E3 occurs, since the
number of leading 1-bits in the new solution is i, the hypermutation operator stops without
any improvement (i.e., E[fi |E3 ] = 1 + E[fi ]). According to the law of total expectation:
(1 + E[fi ]). When, this equation is solved for E[fi ], we
E[fi ] = ni (cn + E[fi ]) + n1 · 1 + n−i−1
n
obtain, E[fi ] = icn + n − i. Since the expected number of consecutive 1-bits that follow the
leftmost 0-bit is less than two, the probability of not skipping a level i is Ω(1). We obtain
n
n
P
P
a lower bound,
pi · E[fi ] = Ω(1) (icn + n − i) = Ω(n3 ) by summing over all fitness
i=1
i=1
levels.
We now focus on establishing that hypermutations may produce considerable speed-ups
if local optima need to be overcome. The Jump(k,n) function, introduced in [24], consists of
a OneMax slope with a gap of length k bits that needs to be overcome for the optimum to
be found. The function is formally defined as:
Pn
Pn
k + i=1 xi if Pi=1 xi ≤ n − k
Jump(k,n) (x) :=
or ni=1 xi = n
P
n − ni=1 xi otherwise
for n > 1 and k ∈ {1...n}. Mutation-based EAs require Θ(nk ) function evaluations to
optimise the function and recently a faster upper bound by a linear factor has been proved
10
for standard crossover-based steady-state GAs [25]. Hence, EAs require increasing runtimes
as the length of the gap increases, from superpolynomial to exponential as soon as k = ω(1).
The following theorem shows that hypermutations allow speed-ups by an exponential factor
of (e/k)k , when the jump is hard to perform. A similar result has been shown for the recently
introduced fast-GA [26].
Theorem 6. Let cn > k. Then the expected runtime of the (1 + 1)IAhyp to optimise Jump
k+1 k
is at most O( n kk·e ).
Proof. The (1+1) IAhyp reaches the fitness level n−k (i.e., the local optimum) in O(n2 log n)
steps according to Theorem 4. All local optima have Hamming distance k to the optimum and the probability
that static hypermutation finds the optimum is lower bounded
n −1
in Lemma 1 by k . Hence,
total expected time to find the optimum is, E(T ) ≤
k+1 the
n
n
·ek
2
.
O(n log n) + cn · k = O
kk
Obviously, hypermutations can jump over large fitness valleys also on functions with
other characteristics. For instance the Cliffd function was originally introduced to show
when non-elitist EAs may outperform elitist ones [27].
(
OneMax(x)
if OneMax(x) ≤ n − d
Cliffd (x) =
OneMax(x) − d + 1/2 otherwise
Similarly to Jump, the function has a OneMax slope with a gap of length d bits that
needs to be overcome for the optimum to be found. Differently to Jump though, the local
optimum is followed by another OneMax slope leading to the optimum. Hence, algorithms
that accept a move jumping to the bottom of the cliff, can then optimise the following slope
and reach the optimum. While elitist mutation-based EAs obviously have a runtime of
Θ(nd ) (i.e., they do not accept the jump to the bottom of the cliff), the following corollary
shows how hypermutations lead to speed-ups that increase exponentially with the distance
d between the cliff and the optimum.
hyp
Corollary
to optimise Cliff
Let cn > d. Then the expected runtime of the (1 + 1)IA
d+1 d1.
is O n dd·e .
The analysis can also be extended to show an equivalent speed-up compared to the
(1+1) EA for crossing the fitness valleys of arbitrary length and depth recently introduced
in [28]. In the next section we will prove that the ageing operator can lead to surprising
speed-ups for Cliff and functions with similar characteristics.
4. Ageing
It is well-understood that the ageing operator can allow algorithms to escape from local
optima. This effect was shown on the Balance function from dynamic optimisation for
an RLS algorithm embedding a hybrid ageing operator [14]. However, for that specific
11
Algorithm 7 (µ+1) RLSageing
p
1: Initialisation
2: while the optimum is not found do
3:
Select x ∈ P (t) uniformly at random
4:
With probability 1/2 < 1 − p ≤ 1 create y by flipping one bit. Otherwise, y is a copy
of x.
5:
Add y to P (t)
6:
Hybrid ageing (P (t) , µ, τ )
7:
Selection (P (t) , µ) without genotype diversity
8: end while
function, an SBM operator would fail to escape, due to the large basin of attraction of the
local optima. In this section we highlight the capabilities of ageing in a more general setting
(i.e., the standard Cliff benchmark function) and show that ageing may also be efficient
when coupled with SBM.
Ageing allows to escape from a local optimum if one not locally optimal b-cell is created
and survives while all the other b-cells die. For this to happen it is necessary that all the
b-cells are old and have similar age. This is achieved on a local optimum by creating copies
of the locally optimal b-cell (i.e., the b-cells will inherit the age of their parent). Hence,
the ability of a mutation operator to create copies enhances this particular capability of the
ageing operator. To this end we first consider a modified RLS algorithm that with some
constant probability p = Ω(1) does not flip any bit and implements the ageing operator
and present its pseudopresented in Algorithm 4. We call this algorithm (µ+1) RLSageing
p
code in Algorithm 7. Apart from making ageing more effective, this slight modification to
the standard RLS considerably simplifies the proofs of the statements we wish to make. In
Section 6 we will generalise the result to an RLS algorithm that does not allow genotype
duplicates as in the standard Opt-IA.
The Cliff benchmark function is generally used to highlight circumstances when nonelitist EAs outperform elitist ones. Algorithms that accept inferior solutions can be efficient
for the function by jumping to the bottom of the cliff and then optimising the OneMax
slope. This effect was originally shown for the (1, λ) EA that can optimise the function
in approximately O(n25 ) fitness function evaluations if the population size λ is not too
large nor too small [27]. This makes the difference between polynomial and exponential
expected runtimes compared to elitist EAs (i.e., Θ(nd )) if the cliff is located far away from the
optimum. A smaller, but still exponential, speed-up was recently shown for the population
genetics inspired SSWM algorithm with runtime at most nd /eΩ(d) [29]. In the previous
section we have already shown that hypermutations are faster than SSWM. The following
theorem proves a surprising result for the considered (µ+1) RLSageing
for Cliff. Not only
p
is the algorithm very fast, but our upper bound becomes lowest (i.e., O(n log n)) when
the function is most difficult (i.e., when the cliff is located at distance d = Θ(n) from
the optimum). In this setting the algorithm is asymptotically as fast as any evolutionary
algorithm using standard bit mutation can be on any function with unique optimum [30].
12
Theorem 7. For µ = O(log n), p < 1 a constant and τ = Θ(n log n), the (µ+1) RLSageing
p
2 3
µ n log n
optimises Cliff in expected time O
if d < n/4−n for any constant 0 < < 1/4.
d2
Proof. We follow the proof of Theorem 10 in [14] of the (µ + 1) RLS for the Balance
function and adapt the arguments therein to the OneMax landscape and to the RLS
operator we use. Given that there are i < µ individuals with j 1-bits in the population,
the probability of creating a new individual with j 1-bits is at least (i/µ)p because the
RLS operator creates a copy with a constant probability p. Hence we follow the proof in
[14] to show that in O(µn + n log n) expected steps the population climbs up the OneMax
slope (i.e., samples a solution with n − d 1-bits) and subsequently the whole population
will be taken over by the local optimum in O(µ log µ) expected generations. Now we can
apply Lemma 5 in [14] to show that in expected O(µ3 ) steps the whole population will have
the same age. As a result after another, at most, τ = Θ(n log n) generations the whole
population will reach age τ simultaneously because no improvements may occur unless the
optimum is found. Overall, the total expected time until the population consists only of
local optima with age τ is at most O(µn + n log n + µ log µ + µ3 + τ ) = O(µn + n log n).
Now we calculate the probability that in the next step one individual jumps to the bottom
of the cliff and the rest
die in the same generation. The first event happens with probability
(1 − p)(d/n) = Ω nd (i.e., an offspring solution with n − d + 1 1-bits is created by flipping
one of the d 0-bits in the parent solution). The probability that the rest of the population
dies is 1/µ · (1 − 1/µ)µ . We now require that the survivor creates an offspring with higher
fitness (i.e. with age = 0) by flipping one of its 0-bits (i.e. it climbs one step up
second
the
d
slope). This event happens with probability at least (1 − p)(d − 1)/(µn) = Ω µn and in
the same generation with probability (1 − 1/µ) = Ω(1) the parent of age τ + 1 dies due to
ageing. Finally, the new solution (i.e. the “safe” individual) takes over the population in
O(µ log µ) expected generations by standard arguments. Using Markov’s inequality we show
that the probability of the take-over happening before any of the new random individuals are
improved n times is at least 1 − O((µ
the overall probability of this series
log)µ/n).
µ
Hence,
µ
d
d2
of consecutive events is Ω nd · µ1 1 − µ1 · Ω µn
· (1 − µ1 ) · 1 − O µ log
=
Ω
2
2
n
n µ
and the expected number of trials (i.e. climbs
up
and
restarts)
until
we
get
a
survivor
2 2
which is safe at the bottom of the cliff is O ndµ2 . Every time the set of events fails
to happen, we wait for another O(µn + n log n) fitness evaluations until the population
reaches a configuration where all individuals are locally optimal and have age τ . Once a safe
individual has taken over the population, the expected time to find the global optimum will
be at most O(µn + n log n). Overall, the total expected time to optimise Cliff conditional
on the best individual
never dying when climbing up the slopes is E(Ttotal ) ≤ O(µn +
2 3
n2 µ2
n log n) · O d2 + O (µn + n log n) = O µ nd2log n . Finally, we consider the probability
that the best individual in the population never dies when climbing up the slopes due
to ageing. After any higher fitness level is discovered it takes O(µ log µ) generations in
expectation and at most n1/2 generations with overwhelming probability until the whole
population takes over the level. For the first n − log n levels, the probability of improving a
13
solution is at least Ω(log n/n) and the probability that this improvement does not happen in
2
τ − n1/2 = Ω(n log n) generations is at most (1 − Ω(log n/n))Ω(n log n) = e−Ω(log n) = n−Ω(log n) .
For the remaining fitness levels, the probability of reaching age τ before improving is similarly
(1 − Ω(1/n))Ω(n log n) = e−Ω(log n) = n−Ω(1) . By the union bound over all levels the probability
that the best solution is never lost due to ageing is at least 1 − o(1). We pessimistically
assume that the whole optimisation process has to restart if the best individual reaches age
τ . However, since at most 1/(1 − o(1)) = O(1) restarts are necessary in expectation, our
bound on the expected runtime holds.
We conclude the section by considering the (µ+1) EAageing which differs from the (µ+1) RLSageing
p
by using standard bit mutation with mutation rate 1/n instead of flipping exactly one bit.
SBM allows copying individuals but, since it is a global operator, it can jump back to the
local optima from anywhere in the search space with non-zero probability. Nevertheless,
the following theorem shows that, for not too large populations, the algorithm is still very
efficient when the cliff is at linear distance from the optimum. Its proof follows similar arguments to those of Theorem 7. The main difference in the analysis is that it has to be shown
that once the solutions have jumped to the bottom of the cliff, they have a good probability
of reaching the optimum before jumping back to the top of the cliff.
Theorem 8. The (µ+1) EAageing optimises Cliff in expected O(n1+ ) time if d = (1/4)(1−
c)n for some constant 0 < c < 1, τ = Θ(n log n) and µ = Θ(1), where is an arbitrarily
small positive constant.
Proof. The probability that the SBM operator increases the number of 1-bits in a parent
solution with Ω(n) 0-bits is at least Ω(n)/(ne) = Ω(1). Following the same arguments
as in the proof of Theorem 7 while considering n/4 > d = Θ(n) and µ = O(1) we can
show that with constant probability and in expected O(n log n) time the algorithm reaches
a configuration where there is a single individual at the bottom of the cliff with n − d + 2
1-bits and age zero while the rest of the population consists of solutions which have been
randomly sampled in the previous iteration.
We will now show that the newly generated individuals have worse fitness than the
solution at the bottom of the cliff when they are initialised. Since d = (1/4)(1 − c)n, the
fitness value of the solutions with more than n − d 1-bits is at least n − (1/4)(1 − c)n −
(1/4)(1 − c)n = (n/2)(1 + c). Due to Chernoff bounds, the newly created individuals have
less than (n/2)(1 + (c/2)) < (n/2)(1 + c) 1-bits w.o.p.
Next we prove that for any constant , there exist some positive constant c∗ , such that the
best solution at the bottom of the cliff (the leading solution) will be improved consecutively
for c∗ log n iterations with probability at least n− . The leading solution with i 0-bits is
selected for mutation and the
a single 0-bit with probability pi := (1/µ) · (i/n)(1 −
QSBM flips
∗ log n
n−1 )n−1 . With probability n−d+1+c
pi , c∗ log n consecutive improvements occur. We
i=n−d+2
can bound this probability from below by the final improvement probability pf raised to
the power of c∗ log n since the improvement probability is inversely proportional to the
number of 0-bits. Considering that pf ≥ (n − d + 1 + c∗ log n)/(nµe) = Ω(1), we can
∗
set c∗ := − logpf 2 = Ω(1), which yields pcf log n = n− . Here, we note that immediately
14
Algorithm 8 Opt-IA(c, n, µ, τ, dup)
1: Initialisation:
t=0
Create P (t) = {x1 , ..., xµ }, a population of µ b-cells uniformly at random
xage
= 0 for i = {1, ...µ}
i
2: while the optimum is not found do
3:
P (clo) = Cloning (P (t) , dup)
4:
P (hyp) = Static hypermutation (P clo , c, n)
5:
Hybrid ageing (P (t) , P (hyp) , µ, τ )
6:
Selection (P (t) , P (hyp) , µ) without genotype diversity
7:
t := t + 1
8: end while
after c∗ log n consecutive improvements, all the individuals in the population have at least
n − d + 1 + c∗ log n − µ 1-bits. More precisely there will be one and only one individual in
the population with j bits for all j in [n − d + 1 + c∗ log n − µ, n − d + 2 + c∗ log n]. Since µ is
constant and SBM flips at least k bits with probability n−k nk , the probability that a single
operation of SBM decreases the number of 1-bits in any individual below n − d + 1 is in the
order of O(1/(log n)!). Since this probability is not polynomially bounded, with probability
at least 1 − O(1/n) it will not happen in any polynomial number of iterations. After the
consecutive improvements occur it takes O(n log n) until the second slope is climbed and
the optimum is found.
The asymptotic bound on the runtime is obtained by considering that algorithm will
reach the local optimum n times in expectation before the consecutive improvements are
observed. Since the time to restart after reaching the local optimum is in the order of
O(n log n), the expected time is O(n1+ log n). Since is an arbitrarily small constant, the
order O(n1+ log n) is equivalent to the order O(n1+ ).
5. Opt-IA
After having analysed the operators separately, in this section we consider the complete
Opt-IA. According to the obtained results and reasonings in the previous sections, the
considered Opt-IA, shown in Algorithm 8, uses static hypermutation coupled with FCM
as variation operator, hybrid ageing and standard (µ + λ) selection. Also, a mutation is
considered constructive if it results in creating an equally fit solution or a better one.
In this section we first show that Opt-IA is efficient for all the functions considered previously in the paper. Then, in Subsection 5.1 we present a problem where the use of the whole
Opt-IA is crucial. In Subsection 5.2 we show limitations of Opt-IA by presenting a class
of functions where standard EAs are efficient while Opt-IA is not w.o.p. We conclude the
section with Subsection 5.3 where we present an analysis for trap functions which disproves
previous claims in the literature about Opt-IA’s behaviour on this class of problems.
15
The following theorem proves that Opt-IA can optimise any benchmark function considered previously in this paper. The theorem uses that the ageing parameter τ is set large
enough such that no individuals die with high probability before the optima are found.
Theorem 9. Let τ be large enough. Then the following upper bounds on the expected runtime
of Opt-IA hold:
E[TOneMax ] = O (µ · dup · n2 log n) for τ = Ω(n2 ),
E[TLeadingOnes ] =
· n3) for τ = Ω(n2 ),
O (µ · dup
k+1
k
for τ = Ω(nk+1 ),
E[TJumpk ] = O µ · dup · n kk·e
nd+1 ·ed
and E[TCliffd ] = O µ · dup · dd
for τ = Ω(nd+1) .
Proof. The claims use that if τ is large enough (i.e., Ω(n2 ) for OneMax and LeadingOnes,
Ω(nk+1 ) for Jumpk and Ω(nd+1 ) for Cliffd ), then with probability 1 − o(1) the current
best solution will never reach age τ and die due to ageing before an improvement is found.
For the Opt-IA to lose the current best solution due to ageing, it is necessary that the best
solution is not improved for τ generations consecutively. If the improvement probability
for any non-optimal solution is at least pmin and if the age τ is set to be p−1
min n, then the
probability that a solution will reach age τ before creating an offspring with higher fitness
is at most (1 − pmin )τ ≤ e−n . By the union bound it is also exponentially unlikely that this
occurs in a polynomial number of fitness levels that need to be traversed before reaching the
optimum. Since the suggested τ for each function is larger than the corresponding p−1
min n,
the upper bounds of the (1+1) IAhyp (which does not implement ageing) for OneMax,
LeadingOnes, Jump and Cliff are valid for Opt-IA when multiplied by µ · dup to take
into account the population and clones.
The following theorem shows that the presented upper bound for OneMax is tight for
sub-logarithmic population and clone sizes (i.e., µ = o(log n) and dup = o(log n)).
Theorem 10. Opt-IA needs at least cn2 /2 · log n − c0 µ · dup · n2 expected fitness function
evaluations for any mutation potential c to optimise OneMax.
Before proving the theorem, we state the following helper result under the name of Ballot
problem [31]. “ Suppose that, in a ballot, candidate P scores p votes and candidate Q scores
q votes, where p > q. The probability that throughout the counting there are always more
votes for P than for Q equals (p − q)/(p + q)” [31].
Proof. By Chernoff bounds, the initial individuals have at least i = n/3 0-bits with overwhelming probability. To calculate the probability of an improvement (i.e., flipping equal
or more 0-bits than 1-bits during one mutation operation), we use the Ballot theorem in
a similar way to [22]. Considering the number of 0-bits as i = q and the number of 1bits as n − i = p, the probability of an improvement is at most 1 − (p − q)/(p + q) =
1 − (n − 2i)/n = 2i/n according to the Ballot theorem 5 . Hence, the probability that at
5
Like in [22] we consider that using cn < n mutations, the probability can only be lower than the result
stated in the Ballot theorem.
16
least one out of dup · µ individuals succeeds is P ≤ dup · µ2i/n by the Union bound. We
optimistically assume that the rest of the individuals also improve their fitness after such
event happens. Recall that the mutation operator wastes cn fitness function evaluations
every time it does not improve
thefitness. Therefore, the expected time to see an improve
n
ment is E(Timprove ) ≥ dup·2µi − 1 · µcn · dup. Since the mutation operator stops at the
first constructive mutation (i.e., when the number of 1-bits is increased by one), it is necessary to improve at least n/3 times. So the total expected time to optimise OneMax is
P
2
0
2
E(Ttotal ) ≥ n/3
i=1 E(Timprove ) = cn /2 · log n − c µ · dup · n .
If individuals were to be removed via ageing, then the runtime may only increase.
5.1. Opt-IA Can Be More Efficient
In this section, we present the function HiddenPath to illustrate a problem where
the use of static hypermutation and ageing together is crucial. When either of these two
characteristic operators of Opt-IA is not used, we will prove that the expected runtime is
at least superpolynomial. HiddenPath : {0, 1}n → R can be described by a series of
modifications to the well-know ZeroMax function. The distinguishing solutions are those
with five 0-bits and those with n − 1 0-bits, along with log n − 3 solutions of the form 1n−k 0k
for 5 ≤ k ≤ log n + 1. The solutions with exactly n − 1 0-bits constitute the local optima
of HiddenPath, and the solutions with exactly five 0-bits form a gradient with fitness
increasing with more 0-bits in the rightmost five bit positions. For any constant < 1, the
HiddenPath function is formally defined as follows:
Pn
(1−xi )
n − + i=n−4n
if ZeroMax(x) = 5 and x 6= 1n−5 05
if ZeroMax(x) < 5 or ZeroMax(x) = n
0
HiddenPath(x) = n − + k/ log n
if 5 ≤ k ≤ log n + 1 and x = 1n−k 0k
if ZeroMax(x) = n − 1
n
ZeroMax(x)
otherwise
Since the all 0-bits string returns fitness value zero, there is a drift towards solutions with
n − 1 0-bits while the global optimum is the 1n−log n−1 0log n+1 bit string. The solutions with
exactly five 0-bits work as a net that stops any static hypermutation that has an input
solution with less than five 0-bits. The namesake path to the global optimum consists of
log n−3 Hamming neighbours and the first solution on the path has five 0-bits. This function
is illustrated in Figure 1.
Theorem 11. For c = 1, dup = 1, µ = O(log n) and τ = Ω(n2 log n), Opt-IA needs expected
O(τ µn + µn7/2 ) fitness function evaluations to optimise HiddenPath.
Proof. For convenience we will call any solution of the form 1n−k 0k for 5 ≤ k ≤ log n
an Sp solution and other solutions with i 0-bits Si solutions. After O(n log n) generations
in expectation an Sn−1 solution is found by optimising ZeroMax. Assuming the global
optimum is not found first, consider the generation when an Sn−1 solution is found for the
17
Figure 1: HiddenPath
first time. Another Sn−1 solution is created and accepted by Opt-IA with probability at
least 1/n since it is sufficient to flip the single 1-bit in the first mutation step and any 0-bit
in the second step will be flipped next with probability 1. Thus, Sn−1 solutions take over the
population in expected O(nµ) generations. Since, apart from the optimum, no other solution
has higher fitness than Sn−1 solutions, the population consists only of Sn−1 solutions after
the takeover occurs (no solutions die with overwhelming probability by standard Markov’s
inequality arguments). We now bound the expected time until all the population has the
same age. Considering that the probability of creating another Sn−1 solution is Θ(1/n),
the
µ
log2 n
1
probability of creating two copies in one single generation is 2 O n2 = O n2 . With
constant probability this event does not happen in o(n2 / log2 n) generations. Conditional
on that at most one additional Sn−1 solution is created in every generation, we can follow
a similar argument as in the proof of Theorem 7. Hence, we can show that in expected
O(µ3 n) iterations after the takeover, the whole population reaches the same age. When
the population of Sn−1 solutions with the same age reaches age τ , with probability 1/µ ·
(1 − (1/µ))2µ−1 = Θ(1/µ) a single new clone survives while the rest of the population dies.
With probability 1−O(1/n) the survived clone has hypermutated all n bits (i.e., the survived
clone is an S1 solution). In the following generation, the population consists of an S1 solution
and µ − 1 randomly sampled solutions. With probability 1 the S1 solution produces an S5
solution via hypermutation. On the other hand, with√
overwhelming probability the randomly
sampled solutions still have fitness value n/2 ± O( n), hence the S1 solution is removed
from the population while the S5 b-cell is kept. Overall, after Θ(µ)+o(1) expected restarts a
solution in S5 will be found in a total expected runtime of O[(n log n + µn + µ3 n + τ + 1)µ] =
O(µτ ) generations.
We momentarily ignore the event that the S5 solution reaches an Sn−1 point via hypermutation which happens with probability O(1/n4 ). The
√ population consists of a single S5
solution and µ − 1 solutions with at most n/2 + O( n) 0-bits. We will now bound the
expected time until S5 solutions take over the population conditional on no Sn−1 solutions
18
being created. Clones of any S5 solutions are also S5 solutions with O(1/n) probability
after hypermutation. Moreover, if the outcome of hypermutation is neither an S5 , an Sp or
an Sn−1 solution, then it is an Sn−5 solution since all n bit-flips will have been executed.
Since Sn−5 solutions have higher fitness value than the randomly sampled solutions they
stay in the population. In the subsequent generation, if hypermutation does not improve an
Sn−5 solution (which happens with probability O(1/n)), it executes n bit-flips to create yet
another S5 solution unless the path is found.
This feedback causes the number of Sn−5 and S5 solutions to double in each generation
with constant probability until they collectively take over the population in O(log µ) generations in expectation. Then, with constant probability all the Sn−5 solutions produce an S5
solution via hypermutation
and consequently the population consists only of S5 solutions.
√
Hence, in O( n log µ) generations w.o.p by applying standard Markov’s inequality iteratively, all the population is on S5 conditional on Sn−4 solutions not being created before.
Since the probability that the static hypermutation creates an Sn−4 solution
from an Sn−5
√
solution is less than (4/n)√· µ, and the probability it happens in O( n log µ) generations
is less than (4/n) · µ · O( n log µ) = o(1), the takeover occurs with high probability, i.e.,
1 − o(1).
After the whole population consists of only S5 solutions, except for the global optimum
and the local optima, only other points on the gradient have higher fitness. The probability
of improving on the gradient is at least 1/n2 which is the probability of choosing two specific
bits to be flipped (a 0-bit and a 1-bit) leading towards the b-cell with the best fitness on
the gradient. Considering that the total number of improvements on the gradient is at most
5, in O(n2 · 5) = O(n2 ) generations in expectation the first point of SP will be found. Now
we consider the probability of jumping back to the local optima from the gradient, which
n
is Θ(1)/ n−4
= O(n−4 ), before finding the best point of the SP. Using Markov’s inequality
√
iteratively, the probability
that the time to find the first point of SP is more than 2 n · cn2
√
is less than 1/2 n . The probability of jumping to a local optimum in these number of steps
−4
√
√
4 ncn2 n
√
n
is 1 − (1 − n )
= 1 − o(1). Therefore, SP is found before any local optima in at
5/2
most O(n ) generations with probability 1 − o(1). Hence, the previously excluded event
has now been taken into account.
After 1n−5 05 is added to the population, the best solution on the path is improved with
probability Ω(1/n) by hypermutation and in expected O(n log n) generations the global
optimum is found. Since all Sp and S5 solutions have a Hamming distance smaller than
n − 4 and larger than n/2 to any Sn−1 solution, the probability that a local optimum is
−1
= o(1) by the Union bound.
found before the global optimum is at most O(n log n) · µn n4
Hence, with probability 1 − o(1) the time to find the optimum is O(n log n) generations.
We pessimistically assume that we start over when this event does not occur, which implies
that the whole process until that point should be repeated 1/o(1) times in expectation.
Overall the dominating term in the runtime is O(τ + n5/2 ) generations. By multiplying with
the maximum possible wasted fitness evaluations per generation (µcn), the upper bound is
proven.
19
In the following two theorems we show that hypermutations and ageing used in conjunction are essential.
Theorem 12. Opt-IA without ageing (i.e., τ = ∞) with µ = O(log n) and dup = O(1)
cannot optimise HiddenPath in less than nΩ(log n) expected fitness function evaluations.
Proof. The only points with higher fitness than solutions with more than n/2 0-bits are path
points and S5 points. If all solutions in the population have less than n − c1 log n 0-bits and
less than n−c1 log n 1-bits for some c1 > 1, then the Hamming distance of any solution in the
population to any path solution or an S5 solution is at least Ω(log n). Thus, the probability
that a path or an S5 solution will be discovered in a single hypermutation operation is
exponentially small according to Lemma 1. As a result we condition on not seeing these
points and the algorithm will find a solution with n − c1 log n 0-bits in O(µ · dup · n2 log n)
fitness function evaluations by optimising ZeroMax (i.e., Theorem 9). In the rest of the
proof we will calculate the probability of reaching an Sn−1 solution before finding either a
path point or S5 point, or a complementary solution of a path point. The latter solutions, if
accepted, with high probability would hypermutate into path points. We call an event bad
where any of the mentioned points are discovered.
Any solution in the search space can have at most two Hamming neighbours which are on
the path or at most two Hamming neighbours with complementary bit-strings that are on the
path. According to Lemma 1, the probability of reaching a solution with Hamming distance
at least d is O(1/nd ) when d < n/2. Therefore, the probability that any of the log n path
points (or their complementary bit-strings) are discovered is O(1/n) + O(1/n2 ) · O(log n) =
O(1/n). For any initial solution with less than n − 10 0-bits, the probability of finding a
−1 n
solution with five 0-bits is at most n6
= O(1/n). Since the probability of reaching an
5
S5 point from solutions with more than n − 10 0-bits may be much larger, we first calculate
the probability of finding n − 11 0-bits before a bad event occurs.
The probability of finding a solution which improves the ZeroMax value is at least
Θ(1/n) (even when we exclude the potential improvements whose complementary bit-strings
are on the path). Since at every static hypermutation, the probabilities of finding an S5
solution, a path solution or its complementary bit-string are all in the order of O(1/n) (up
to n − 10 0-bits), the conditional probability that a solution that improves the current best
ZeroMax value is found before any other improvement is in the order of Θ(1/n)/(Θ(1/n)+
O(dup · µ/n)) = Ω(1/(dup · µ)). The probability that it happens c1 log n times immediately
after the first solution with more than n−c1 log n 0-bits is added to the population is at least
Ω(dup · µ)−c1 (log n) = (dup · µ)−O(log n) . This sequence of c1 log n improvements implies that a
solution with n − 11 0-bits is added to the population. We now consider the probability that
one individual finds the local optimum (i.e., an Sn−1 solution) by improving its ZeroMax
value in 10 consecutive generations and that none of the other individuals improve in these
generations. The probability that the current best individual improves in the next step
is at least 1/n and the probability that the other individuals do not improve is at least
(1 − c/n)µ ≥ 1/e. Hence, the probability that this happens consecutively in the next 10
steps is at least (1/ne)10 = Ω(n−10 ).
20
Once a local optimal solution is found, with high probability it takes over the population
before the optimum is found and the expected runtime conditional on the current population
n)log n
consisting only of solutions with n − 1 0-bits is at least logn n ≥ (n−log
. By the law
(log n)log n
of total expectation, the unconditional expected runtime is lower bounded by: E[T ] ≥
n)log n
. This expression is in the order of nΩ(log n) for any
(1 − o (1)) n−10 (dup · µ)−O(log n) (n−log
(log n)log n
µ = O(log n) and dup = O(1).
Theorem 13. With Probability at least 1 − e−Ω(n) , Opt-IA using SBM and ageing cannot
optimise HiddenPath in less than nΩ(n) fitness function evaluations.
Proof. By Chernoff bounds, the probability that an initial solution has less than n/4 1-bits
is bounded above by e−Ω(n) . Hence, w.o.p the population is at least a linear distance away
from any solution on the path and any solution with five 0-bits (Sp ∪ S5 ) by the Union
bound. Therefore, the probability
of finding either any of the log n points on Sp or one of
0
0
the n5 points on S5 is log n + n5 · 1/nc n (1 − 1/n)n−c n ≤ n−Ω(n) . Since accepting any
improvements, except for the Sp ∪S5 solutions, increases the distance to the Sp ∪S5 solutions,
the probability of jumping to an Sp ∪ S5 solution further decreases throughout the search
process.
5.2. When Opt-IA is detrimental
In this section we present a function class for which Opt-IA is inefficient. The class of
functions, which we call HyperTrap, is illustrated in Figure 2. It is inspired by the SpTarget function introduced in [32] and used in [22] as an example where hypermutations
outperform SBM. Compared to Sp-Target, HyperTrap has local and global optima
inverted with the purpose of trapping hypermutations. Also there are some other necessary
modifications to prevent the algorithm from finding the global optimum via large mutations.
In HyperTrap, the points with |x|1 < n/2 are evaluated by OneMax, while the points
with |x|1 ≥ n/2 are evaluated as follows:
• Points of the form 1i 0n−i with n/2 ≤ i ≤ n, shape a path called Sp (short path). The
last point of Sp, i.e., 1n is the global optimum (Opt).
• Points with |x|1 ≥ 3n/4 and a Hamming distance of at least γn to all points in Sp,
form the local optima (Lo). In the formal definition of the function, H(x, SP ) is the
minimum Hamming distance of the individual x to all Sp points.
• Points with |x|1 = n/2 are ranked among each other such that bit strings with more 1bits in the beginning have higher fitness. This ranking forms a gradient from 0n/2 1n/2 ,
which has the lowest fitness, to 1n/2−1 0n/2+1 , which has the highest fitness.
• Points which do not belong to any of the above sub-spaces are evaluated by ZeroMax.
21
Figure 2: HyperTrapγ
The function is formally defined as follows:
OneMax(x)
if
Pn
i=1 (n−i)·(xi )
n/2 +
if
n
n2 · |x|
if
1
HyperTrapγ (x) =
n3
if
4
n
if
ZeroMax(x)
if
|x|1 < n/2
|x|1 = n/2
x ∈ Sp := {x | x = 1i 0n−i & n/2 ≤ i < n}
x ∈ Lo := {x | |x|1 ≥ 3n/4 & H(x, Sp) ≥ γn}
x = 1n = Opt
|x|1 > n/2 & x ∈
/ (Sp ∪ Lo ∪ Opt)
We will show that there exists a HyperTrapγ such that Opt-IA with mutation potential
cn gets trapped in the local optima.
√
Theorem 14. With probability 1 − 2−Ω( n) , Opt-IA with mutation potential cn cannot optimise HyperTrapc/8 in less than exponential time.
Proof. We will show that the population will first follow the ZeroMaxP(or OneMax)
gradient until it samples a solution with n/2 0-bits and then the gradient of ni=1 (n−i)·(xi )
until it samples an Sp point with approximately n/2 0-bits. Afterwards, we will prove
that large jumps on the path are unlikely. Finally, we will show that with overwhelming
probability a locally optimal solution is sampled before a linear number of path points are
traversed. We optimistically assume that τ is large enough such that individuals do not die
before finding the optimum.
With overwhelmingly high probability, all randomly initialised individuals have (1/2±)n
0-bits for any arbitrarily small = θ(1). Starting with such points, the probability of jumping
to the global optimum is exponentially small according to Lemma 1. Being optimistic, we
assume that the algorithm does not sample a locally optimal point for now. Hence, until
an Sp point is found, the current best fitness can only be improved if a point with either a
22
P
higher ZeroMax (or OneMax) or a higher ni=1 (n − i) · (xi ) value P
than the current best
2
individual is sampled. Since, there are less than n different values of ni=1 (n − i) · (xi ), and
less than n different values of ZeroMax (or OneMax), the current best individual cannot
be improved more than n2 + n times after initialisation without sampling an Sp point.
Let Splower denote the Sp points with less than (1/2 + 2)n 1-bits, and Spupper denote
Sp \ Splower . We will now show that with overwhelmingly high probability an Splower point
will be sampled before an Spupper point. The search points between (1/2 ± )n 1-bits have at
least a distance of n to the Spupper points. Using Lemma 1, we can bound the probability
n −1
of sampling an Spupper by n
from any input solution with (1/2 ± )n 1-bits. Excluding
Sp points, a solution with (1/2 ± )n 1-bits has an improvement probability of at least 1/n2
(i.e., min{1/n2 , Θ(1)} = 1/n2 with the second term being the probability of improving the
ZeroMax or OneMax value and the first term the probability of improving when it has
exactly n/2 0-bits). Thus, the conditional probability that the best solution is improved
before any search point in the population is mutated into an Spupper point is (n−2 )/(n−2 +
n −1
dup · µ · n
) = 1 − 2−Ω(n) . This implies that an Splower point is sampled before an Spupper
point with overwhelmingly high probability. Indeed, by the union bound, this event happens
n2 + n times consecutively after initialisation with probability at least 1 − (n2 + n) · 2−Ω(n) =
1 − 2−Ω(n) , thus a point in Splower is sampled before an Spupper point with overwhelmingly
high probability.
Each point of Sp has at least n/2 1-bits at the beginning of its bit string. In order
to improve by any value in one mutation operation, 1-bits should never be touched before
the mutation operator stops. Hence, the probability of improving
by X on Sp is p(X) <
√
X
(1/2) . This√yields that the probability of improving by n in one generation is at most
dup · µ · (1/2) n , as there are µ individuals in the population. We have now shown that it is
exponentially unlikely that even an arbitrarily small fraction of the path points are avoided
by jumping directly to path points with more 1-bits.
Let t be the first iteration when the current Sp solution
√ has at least 99n/100 1-bits for the
n new 1-bits are added given that
first time. The conditional probability that more than
√
an improving path point is sampled is at most (1/2) n−1 since the improvement probability
is at least
1/n. Therefore, in generation t, the current best individual
cannot have more
√
√
n−1
than n + 99n/100 1-bits with √
probability 1 − dup · µ · (1/2)
. By the Union bound,
0
the probability of √
improving by n − 200
bits
in
any
of
the
next
c
n
generations is at most
√
0
n−200
−Ω( n)
0
c n · µ · dup · (1/2) √ = dup · µ · 2
for any c = poly(n). Therefore, the probability
√
−Ω( n)
that no jump of size n−200 happens in c0 n generations is at least 1−dup·µ·2
. So we
√
can conclude that the number of generations to
add another n/200 − n 1-bits to the prefix
√
√
n/200− n
of the current best solution is at least √n−200 = n/200 with overwhelming probability.
Now, we show that during this number of generations, the algorithm gets trapped in the
local optima with overwhelming probability with the optimistic assumption that only the
best b-cell in the population can be mutated into a locally optimal solution. Considering
the best current individual, in each generation a 1-bit is flipped with probability √
at least
99/100 in the first mutation step.
Thus,
the
probability
of
not
touching
a
1-bit
in
n/200
√
n/200
generations is less than (1/100)
. Overall, the probability of flipping a 1-bit in this
23
number of generations and of not locating the optimum in the same amount of time is
√
√
n
−Ω( n)
P ≥ 1−
· dup · µ · 2
·
200
√
= (1 − dup · µ · 2−Ω(
n)
√
1
1−
100
√
)(1 − 2−Ω(
n)
n/200
!
) ≥ 1 − 2−Ω(
√
n)
After flipping a 1-bit, the mutation operator mutates at most cn bits until it finds an
improvement. All sampled solutions in the first n/4 − n/100 mutation steps will have more
than 3n/4 1-bits and thus satisfy the first condition of the local optima. If the Hamming
distance between one of the first n/4 − n/100 sampled solutions and all Sp points is at
least γn, then the algorithm is in the trap. We only consider the Sp points with a prefix
containing more than 3n/4 − γn 1-bits since Sp points with less than 3n/4 − γn 1-bits have
already Hamming distance more than γn to the local optima.
Similarly to the analysis done in [22], we consider the kth mutation step where, k :=
6cn
≤ cn where the last expression is due to γ = c/8. After
4γn(1 + 1/5)/(3 − 4γ) = 30−5c
k steps, the expected number of bits flipped in the prefix of length n(3/4 − γ) is at least
6cn
≤ n/4 − n/100.
k(3/4 − γ) = (1 + 1/5)γn. For any mutation potential 0 < c ≤ 1, 30−5c
Using Chernoff bounds, we can show that the probability of having less than γn 0-bits in
−γn·6
the prefix is P (X ≤ (1 − 0.1)E[X]) ≤ e 1000 = e−Ω(n)
√
√
Altogether, with probability (1 − e−Ω(γn) ) · (1 − 2−Ω( n) ) = 1 − e−Ω( n) a point in the
local optima is sampled. In O(log µ) = O(log n) generations this individual takesover the
n
population. Once in the local optima, the algorithm needs at least dup · µ · 1/ γn
time to
find the global optimum.
Theorem 15. The (1+1) EA optimises HyperTrap in O(n3+ ) steps w.o.p., 1 − e−Ω(n )
for > 0.
√
Proof. With overwhelming probability, the algorithm is initialised within n/2±c n 1-bits by
Chernoff bounds. Since the distance to any trap point is linear, the probability of mutating
into a trap point is exponentially small. As the algorithm approaches the bit string with
n/2 1-bits, the distance to the trap remains
√ linear. Conditional on not entering the trap, by
standard AFL it takes the (1+1) EA O( n) steps in expectation to find a point with n/2 1bits, i.e., on the gradient. After finding such a point, the (1+1) EA improves on the gradient
with probability at least 1/n2 · (1 − 1/n)n−2 ≥ 1/en2 which is the probability of flipping
the rightmost 1-bit and the leftmost 0-bit while leaving the rest of the bits untouched. To
reach the first point of Sp, there are a linear number of 1-bits that need to be shifted to
the beginning of the bit string. It therefore takes O(n3 ) steps in expectation to find the Sp.
The (1+1) EA improves on Sp with probability at least 1/n · (1 − 1/n)n−1 ≥ 1/en, which
is the probability of flipping the leftmost 0-bit and not touching the other bits. Hence, the
global optimum is found in O(n2 ) steps in expectation giving a total expected runtime of
O(n3 ) conditional on not falling into the trap. By applying Markov’s inequality iteratively,
24
Table 1: “The best results obtained by IA with static, proportional, inversely proportional hypermutation
and hypermacromutation with M-mutations (M-mut) strategy” [17].
Trap
S(1)
Static
Proportional
Inversely
S(4)
0
0
0
100, 4334.94
τ =1
5, 5626.8
τ = 50
5, 174458.6
τ = 50
0
S(5)
0
0
0
0
S(2)
S(3)
100, 576.81
100, 604.33
100, 504.76
τ = 25, c = 0.9 τ = 100, c = 0.2 τ = 5, c = 0.3
34, 82645.85
100, 50266.61
97, 58092.70
τ = 20, c = 1.0 τ = 25, c = 0.8 τ = 20, c = 0.2
0
0
0
Hypermacro M-mut
the probability that the optimum is not found within ecn3+ steps is less than e−n with an
arbitrarily small constant. Since the probability of finding a local optima from the gradient
points or Sp points is n−Ω(n) in each step, the probability of not falling into the trap in n3+
0
steps is less than (n3+ · n−c n ) by the Union bound. Overall, the total probability of finding
0
the optimum within O(n3+ ) steps is bigger than (1 − e−n ) · (1 − n3+ · n−c n ) = 1 − e−Ω(n ) .
5.3. On Trap Functions
In [17], where Opt-IA was originally introduced, the effectiveness of the algorithm was
tested for optimising the following simple trap function:
(
a
(z − OneMax(x))
if OneMax(x) ≤ z
Simple Trap(x) := z b
(OneMax(x) − z) otherwise
n−z
where z ≈ n/4, b = n − z − 1 and 3b/2 ≤ a ≤ 2b and the optimal solution is the 0n bit
string.
Table 1 shows the experimental results obtained by Opt-IA in [17] for optimising the
Simple Trap with the following five parameter settings; S(1): {n = 10, z = 3, a = 12, b =
6}, S(2): {n = 20, z = 5, a = 20, b = 14}, S(3): {n = 50, z = 10, a = 80, b = 39}, S(4):
{n = 75, z = 20, a = 80, b = 54} and S(5): {n = 100, z = 25, a = 100, b = 74}. There are
4 values at each entry of Table 1. The first value shows the success rate (SR), the second
value shows the average number of fitness evaluations (AES), the third value (τ ) shows the
best found value for the ageing parameter among {1, 5, 10, 15, 20, 25, 50, 100, 150, 200, ∞}
and the forth value (c) is the best found value for the hypermutation parameter c among
{0.1, 0.2, ..., 1}. The reported results are averaged over 100 independent runs each with a
termination criterion of reaching 5 × 105 fitness function evaluations. For all of the results,
the population size is 10 and dup = 1.
25
According to these experimental results, Opt-IA using either hypermutations or hypermacromutation cannot optimise Simple Trap at all for problem sizes n ≥ 50. Additionally,
static hypermutations and hypermacromutation show low success rates already for n = 20.
The following theorem shows that Opt-IA indeed optimises Simple Trap efficiently.
Theorem 16. Opt-IA needs O(µn2 log n) expected fitness function evaluations to optimise
Simple Trap with τ = Ω(µn1+ ) for arbitrary = Ω(1/n), c = 1 and dup = 1.
√
Proof. With overwhelming probability, the individuals are initialised within n/2± n 1-bits
by the Chernoff bounds. Let i be the number of 0-bits of the current best individual. The
probability of an improvement is at least i/n regardless of which branch the best solution is
on. Whenever another individual on the opposite branch with better or equal fitness value is
introduced, then the branch where the current best individual belongs to is switched. Hence,
the probability of improving in the next step will not change (i.e., min{(n−i)/n, i/n} = i/n).
This implies that w.o.p no individual will die in each generation.
By following the proof of Theorem 4, at least one individual will reach 1n or 0n in
O(µn2 log n) fitness function evaluations in expectation as long as no individual has died
due to ageing. Pessimistically assuming that 1n is found first, the algorithm finds the global
optimum in one step with probability 1 by flipping all bits and performing n fitness function
evaluations.
Altogether, w.o.p the optimum is found within µn1+ generations which is µn2+ fitness
function evaluations, that is before the ageing operator is ever triggered (i.e., τ = Ω(µn1+ )).
Given that Opt-IA was tested in [17] also with the parameters suggested by Theorem
16 (i.e., c = 1, dup = 1, τ = ∞), we speculate that either FCM was mistakingly not used
or the stopping criterion (i.e., the total number of allowed fitness evaluations, i.e., 5 × 105 )
was too small. We point out that, for large enough τ also using hypermacromutation as
mutation operator would lead to an expected runtime of O(µn2 log n) for Trap functions
[8], with or without FCM. In any case, it is not necessary to apply both hypermutations and
hypermacromutation together to efficiently optimise Trap functions. On the other hand,
the inversely proportional hypermutation operator considered in [22] would fail in optimising
this function because it cannot flip n bits when on the local optimum.
6. Not Allowing Genotype Duplicates
None of the algorithms considered in the previous sections use the genotype diversity
mechanism. In this section, we do not allow genotype redundancies in the population as
proposed in the original algorithm [17, 6]. This change potentially affects the behaviour of
the algorithm. In particular, the probability of creating copies of individuals may change
considerably. In the following, we will first analyse the ageing operator in isolation and
afterwards the complete Opt-IA both with genotype diversity.
26
Algorithm 9 (µ+1) RLSageing
1: Initialisation
2: while the optimum is not found do
3:
Select x ∈ P (t) uniformly at random
4:
Create y by flipping one bit
5:
Add y to P (t)
6:
Hybrid ageing (P (t) , µ, τ )
7:
Selection (P (t) , µ) with genotype diversity
8: end while
6.1. (µ+1) RLSageing with genotype diversity
In this subsection, we analyse (µ+1) RLSageing (Algorithm 9) with genotype diversity
without genotype diversity was previously
for optimising Cliff for which (µ+1) RLSageing
p
analysed in Section 4. The main difference compared to the analysis there is that taking
over the population on the local optima is now harder since identical individuals are not
allowed. The proof of the following theorem shows that the algorithm can still take over and
use ageing to escape from the local optima as long as the population size is not too large.
Theorem 17. For constant µ = Θ(1) and τ = Θ(n), the (µ+1) RLSageing with genotype
diversity optimises Cliff in expected O(n log n) fitness function evaluations for any linear
d ≤ n/4 − and any µ.
Proof. With overwhelming probability, the initial individuals are sampled with n(1/2 ± )
1-bits for any arbitrarily small 0 < = Θ(1). Since the population size is constant and
there is a constant probability of improving in the first mutation step, a local optimum is
found in at most O(n) steps by picking the best individual and improving it O(n) times. If
there is a single locally optimal solution in the population, then with probability 1/µ this
individual is selected as parent and produces an offspring with fitness n − d − 1 (i.e., one
bit away) with probability (n − d − i)/n ≤ (n − d − µ)/n = Θ(1) where i is the number of
individuals already with fitness n − d − 1. In the next generation, with probability (i + 1)/µ
one of the individuals on the local optimum or one step away from it is selected as parent
and produces an offspring on either the local optimum or one bit away with probability at
least (n − d − µ)/n = Θ(1). Hence, in expected time µ · Θ(1) = Θ(µ) all µ individuals are
on the local optimum.
When the last inferior solution is replaced by an individual on the local optimum, the
other individuals have ages in the order of O(µ). Thus, the probability that the rest
of the population does not die until the youngest individual reaches age τ , is at least
(1/µ)(µ−1)·O(µ) = Θ(1), the probability that µ − 1 individuals above age τ survive O(µ)
consecutive generations.
In the first generation when the last individual reaches age τ , with probability d/n an
offspring is created at the bottom of the cliff (i.e., with a fitness value of n − d + 1) and with
probability 1/µ · (1 − 1/µ)µ = Θ(1) all the parents die together at that step and the offspring
survives. The rest of the proof follows the same arguments as the proof of Theorem 7.
27
Overall, the total expected time to optimise Cliff is dominated by the time to climb
the second OneMax slope which takes O(n log n) steps in expectation.
6.2. Opt-IA with genotype diversity
In this subsection, we analyse Opt-IA (Algorithm 8) with genotype diversity to optimise
all the functions for which Opt-IA without genotype diversity was analysed in Section 5. The
following are straightforward corollaries of Theorem 9 and Theorem 16. Since the ageing
mechanism never triggers here and the proofs of those theorems do not depend on creating
genotype copies, the arguments are still valid for Opt-IA with genotype diversity.
Corollary 2. The upper bounds on the expected runtime of Opt-IA with genotype diversity
and ageing parameter τ large enough for OneMax, LeadingOnes, Jumpk and Cliffd
are as follows:
E[TOneMax ] = O (µ · dup · n2 log n) ,
E[TLeadingOnes ]= O (µ · dup · n3) ,
k+1 k
E[TJumpk ] = O µ · dup · n kk·e ,
nd+1 ·ed
and E[TCliffd ] = O µ · dup · dd
.
Corollary 3. Opt-IA with genotype diversity needs O (µn2 log n) expected fitness function
evaluations to optimise Simple Trap with τ = Ω(µn1+ ), c = 1 and dup = 1.
The following theorem shows the same expected runtime for Opt-IA with genotype diversity for HiddenPath as that of the Opt-IA without genotype diversity proven in Theorem
11. However, we reduce the population size to be constant. The proof follows the main
arguments of the proof of Theorem 11. Here we only discuss about the probability of events
which should be handled differently considering that genotype duplicates are not allowed.
Theorem 18. For c = 1, dup = 1, µ = Θ(1) and τ = Ω(n2 log n), Opt-IA with genotype
diversity needs O(τ µn + µn9/2 ) expected fitness function evaluations to optimise HiddenPath.
Proof. We follow the analysis of the proof of Theorem 11 for Opt-IA with genotype diversity.
Although the analysis did not benefit from genotype duplicates, not allowing them potentially affects the runtime of the events where the population takes over. The potentially
affected events are:
• For Sn−1 solutions to take over the population, the probabilities are different here since
a new Sn−1 solution will not be accepted if it is identical to any current Sn−1 solutions.
Here, after finding the first Sn−1 solution, the rest are created and accepted by Opt-IA
with probability at least Ω(1/n · (n − µ)/(n − 1)) = Ω(1/n). Therefore, the argument
made in the proof of Theorem 11 does not change the runtime asymptotically.
28
• The arguments about the expected time needed for Sn−1 solutions to reach the same
age after the takeover are the same as in the proof of Theorem 11; without genotype duplicates, the probability of creating another Sn−1 is still Ω(1/n). Hence, the
probability of creating two copies in the same generation is still unlikely and we can
adapt the arguments made in the proof of Theorem 11. Hence, in expected O(µ3 n)
generations after the takeover of Sn−1 , the population reaches the same age.
• The expected time for S5 solutions to take over the population of recently initialised
individuals is O(log µ) generations in the proof of Theorem 11. However, this time is
different here. With genotype diversity, the probability of creating another S5 solution
) ≥ 4/n = Ω(1/n) because there can be at most µ = Θ(1) different
is at least ( n5 · n−5−µ
n−1
genotypes in the population. So in expected O(µn) generations the S5 solutions take
over the population conditional on Sn−1 solutions not having been found before. If
an S5 solution is not improved, then with probability 1 the hypermutation operator
flips all bits and produces an Sn−5 solution. This solution will be accepted because
it has higher fitness than the randomly initialised individuals. For our purposes, we
pessimistically assume that µ − 1 individuals are Sn−5 solutions and we prove that
with constant probability the population is taken over by S5 solutions. Now we need
to show that the takeover of S5 solutions happens before an Sn−1 solution is found.
Let A be the event of creating a new S5 solution and B the event of improving an
Sn−5 solution by increasing the number of 0-bits. Then P (A) ≥ 4/n = Ω(1/n) and
P (B) ≤ 2i/n ≤ 10/n by the Ballot theorem when i is the number of 1-bits. Therefore,
in each generation the probability that event A happens before either B or A+B is at
least P (A|A ∪ B ∪ A + B) ≥ (4/n)/ ((10µ/n) + (4/n) + O(1/n2 )) ≥ (1/3µ) = Θ(1).
Hence, the probability that µ S5 solutions are found before any Sn−1 solution is at least
(1/3µ)µ . Recalling that hypermutation stops at the first improvement, Sn−3 cannot
be found before Sn−4 . So with constant probability the population is taken over on S5
before any point in Sn−1 is found.
The rest of the proof of Theorem 11 is not affected by genotype diversity.
7. Conclusion
We have presented an analysis of the standard Opt-IA artificial immune system. We first
highlighted how both the ageing and hypermutation operators may allow to efficiently escape
local optima that are particularly hard for standard evolutionary algorithms. Concerning
hypermutations, we proved that FCM is essential to the operator and suggested considering
a mutation constructive if the produced fitness is at least as good as the previous one. The
reason is that far away points of equal fitness should be attractive for the sake of increasing
exploration capabilities. Concerning ageing, we showed for the first time that the operator
can be very efficient when coupled with standard bit mutations and hypermutations. To
the best of our knowledge, the operator allows the best known runtime (i.e., O(n log n))
for hard Cliff functions. Afterwards, we presented a class of functions where both the
29
characteristics of ageing and hypermutation are crucial, hence Opt-IA is efficient while standard evolutionary algorithms are inefficient even if coupled with one extra AIS operator
(either cloning, ageing, hypermutation or contiguous somatic mutation). Finally, we prove
that all the positive results presented for the Opt-IA without genotype diversity also hold
for the Opt-IA with genotype diversity as used in the original algorithm. However, small
population sizes may be required if ageing has to be triggered to escape from local optima
with genotype diversity. To complete the picture we present a class of problems where the
use of hypermutations and ageing is detrimental while standard evolutionary algorithms are
efficient.
Our analysis shows that for easy problems for which local search strategies are efficient,
using hypermutations and ageing may be detrimental. We have shown this effect for the
simple OneMax and LeadingOnes functions for which we have proven a linear slow-down
in the expected runtime compared to local search strategies and simple evolutionary algorithms. While such a slow-down may be considered an acceptable expense in exchange for
being efficient on more complicated optimisation problems, we have shown for HyperTrap
that the consequences may be more drastic, by making the difference between polynomial
and exponential runtimes. Indeed, the function is easy for local search algorithms with
neighbourhoods of size greater than 1 and for simple EAs. However, hypermutations make
Opt-IA fail with overwhelming probability and ageing does not help the algorithm to escape
from the trap permanently.
On the other hand, for more complicated multimodal functions, with closer characteristics to optimisation problems that occur in practice, we have shown several advantages
of hypermutations and ageing to escape from local optima (i.e., Jump and Cliff). Furthermore, the combination of hypermutation and ageing may be advantageous to locate
new promising basins of attraction that are hard to find via more traditional optimisation
techniques such as local search or EAs. We have illustrated such effect in the analysis of
HiddenPath, where the combination of hypermutations and ageing allow Opt-IA to locate
a new basin of attraction that initially has lower fitness than the easy-to-find local optima.
However, in the long run, having identified this new basin of attraction allows the algorithm
to find the global optimum.
Overall, we believe this work is a significant contribution towards the understanding
of which kind of problems it is advantageous to use artificial immune systems on rather
than evolutionary algorithms and for which it is detrimental. Future work should focus
on providing such advantages and disadvantages for classical combinatorial optimisation
problems.
Acknowledgments: The research leading to these results has received funding from
the EPSRC under grant agreement no EP/M004252/1.
References
References
[1] D. Corus, P. S. Oliveto, D. Yazdani, On the runtime analysis of the opt-ia artificial immune system,
in: Proc. of GECCO 2017, 2017, pp. 83–90.
30
[2] L. N. de Castro, J. Timmis, Artificial Immune Systems: A New Computational Intelligence Paradigm,
Springer-Verlag, Secaucus, NJ, USA, 2002.
[3] F. M. Burnet, The Clonal Selection Theory of Acquired Immunity, Cambridge University Press, 1959.
[4] L. N. de Castro, F. J. V. Zuben, Learning and optimization using the clonal selection principle, IEEE
Transaction on Evolutionary Computation 6 (3) (2002) 239–251.
[5] J. Kelsey, J. Timmis, Immune inspired somatic contiguous hypermutation for function optimisation,
in: Proc. of GECCO 2003, 2003, pp. 207–218.
[6] V. Cutello, M. Pavone, J. Timmis, An immune algorithm for protein structure prediction on lattice
models, IEEE Transactions on Evolutionary Computation 10 (2006) 844–861.
[7] V. Cutello, G. Nicosia, M. Romeo, P. S. Oliveto, On the convergence of immune algorithms, in: Proc.
of FOCI 2007, 2007, pp. 409–415.
[8] T. Jansen, C. Zarges, Analyzing different variants of immune inspired somatic contiguous hypermutations, Theoretical Computer Science 412 (6) (2011) 517–533.
[9] D. Corus, J. He, T. Jansen, P. S. Oliveto, D. Sudholt, C. Zarges, On easiest functions for mutation
operators in bio-inspired optimisation, Algorithmicadoi:10.1007/s00453-016-0201-4.
[10] C. Zarges, Rigorous runtime analysis of inversely fitness proportional mutation rates, in: Proc. of
PPSN X, 2008, pp. 112–122.
[11] C. Zarges, On the utility of the population size for inversely fitness proportional mutation rates, in:
Proc. of FOGA 2009, 2009, pp. 39–46.
[12] T. Jansen, C. Zarges, On the role of age diversity for effective aging operators, Evolutionary Intelligence
4 (2) (2011) 99–125.
[13] C. Horoba, T. Jansen, C. Zarges, Maximal age in randomized search heuristics with aging, in: Proc. of
GECCO 2009, 2009, pp. 803–810.
[14] P. S. Oliveto, D. Sudholt, On the runtime analysis of stochastic ageing mechanisms, in: Proc. of
GECCO 2014, 2014, pp. 113–120.
[15] T. Jansen, P. S. Oliveto, C. Zarges, On the analysis of the immune-inspired B-Cell algorithm for the
vertex cover problem, in: Proc. of ICARIS 2011, 2011, pp. 117–131.
[16] T. Jansen, C. Zarges, Computing longest common subsequences with the B-Cell Algorithm, in: Proc.
of ICARIS 2012, 2012, pp. 111–124.
[17] V. Cutello, G. Nicosia, M. Pavone, Exploring the capability of immune algorithms: A characterization
of hypermutation operators, in: Proc. of ICARIS 2004, 2004, pp. 263–276.
[18] V. Cutello, G. Nicosia, M. Pavone, A hybrid immune algorithm with information gain for the graph
coloring problem, in: Proc. of GECCO 2003, 2003, pp. 171–182.
[19] V. Cutello, G. Nicosia, A clonal selection algorithm for coloring, hitting set and satisfiability problems,
Neural Nets 3931 (2006) 324–337.
[20] P. S. Oliveto, X. Yao, Runtime analysis of evolutionary algorithms for discrete optimisation, in: Theory
of Randomized Search Heuristics: Foundations and Recent Developments, World Scientific, 2011, Ch. 2,
pp. 21–52.
[21] T. Jansen, Analyzing Evolutionary Algorithms: The Computer Science Perspective, Springer, 2013.
[22] T. Jansen, C. Zarges, Variation in artificial immune systems: Hypermutations with mutation potential,
in: Proc. of ICARIS 2011, 2011, pp. 132–145.
[23] P. S. Oliveto, P. K. Lehre, F. Neumann, Theoretical analysis of rank-based mutation-combining exploration and exploitation, in: Proc. of CEC 2009, 2009, pp. 1455–1462.
[24] S. Droste, T. Jansen, I. Wegener, On the analysis of the (1+ 1) evolutionary algorithm, Theoretical
Computer Science 276 (1-2) (2002) 51–81.
[25] D.-C. Dang, T. Friedrich, T. Kötzing, M. S. Krejca, P. K. Lehre, P. S. Oliveto, D. Sudholt, A. M.
Sutton, Emergence of diversity and its benefits for crossover in genetic algorithms, to appear in IEEE
Transactions on Evolutionary Computationdoi:10.1109/TEVC.2017.2724201.
[26] B. Doerr, H. P. Le, R. Makhmara, T. D. Nguyen, Fast genetic algorithms, in: Proc. of GECCO 2017,
2017, pp. 777–784.
[27] J. Jägersküpper, T. Storch, When the plus strategy outperforms the comma strategy and when not,
31
in: Proc. of FOCI 2007, 2007, pp. 25–32.
[28] P. S. Oliveto, T. Paixão, J. Pérez Heredia, D. Sudholt, B. Trubenová, How to escape local optima in black box optimisation: When non-elitism outperforms elitism, Algorithmicadoi:10.1007/
s00453-017-0369-2.
[29] T. Paixão, J. Pérez Heredia, D. Sudholt, B. Trubenová, Towards a runtime comparison of natural and
artificial evolution, Algorithmica 78 (2) (2017) 681–713.
[30] D. Sudholt, A new method for lower bounds on the running time of evolutionary algorithms, IEEE
Transactions on Evolutionary Computation 17 (2012) 418–435.
[31] W. Feller, An Introduction to Probability Theory and Its Applications, John Wiley & Sons, 1968.
[32] F. Neumann, D. Sudholt, C. Witt, Rigorous analyses for the combination of ant colony optimization and
local search, in: Proc. of International Conference on Ant Colony Optimization and Swarm Intelligence
2008, Vol. 5217, 2008, pp. 132–143.
32
| 9cs.NE
|
Resilient Monotone Submodular Maximization
arXiv:1703.07280v2 [math.OC] 31 Oct 2017
Vasileios Tzoumas,1 Konstantinos Gatsis,1 Ali Jadbabaie,2 George J. Pappas1
Abstract— In this paper, we focus on applications in machine
learning, optimization, and control that call for the resilient
selection of a few elements, e.g. features, sensors, or leaders,
against a number of adversarial denial-of-service attacks or
failures. In general, such resilient optimization problems are
hard and cannot be solved exactly in polynomial time, even
though they may involve objective functions that are monotone
and submodular. In this paper, we provide for the solution
of such optimization problems the first scalable approximation
algorithm that is valid for any number of attacks or failures
and which, for functions with low curvature, guarantees superior approximation performance. Notably, the curvature has
been known to tighten approximations for several non-resilient
optimization problems, yet its effect on resilient optimization
had hitherto been unknown. We complement our theoretical
analyses with empirical evaluations.
During the last decade, researchers in machine learning,
optimization, and control have focused on questions such as:
• (Sensor selection) How many sensors do we need to
deploy in a large water distribution network to detect a
contamination outbreak as fast as possible? [1]
• (Feature selection) Which few features do we need to
select from the data flood to optimally train a spam e-mail
classifier? [2]
• (Leader selection) Which UAVs in a multi-UAV system do
we need to choose as leaders so the system can complete
a surveillance task despite communication noise? [3]
The effort to answer such questions has culminated in a
plethora of papers on topics such as actuator placement for
controllability [4]–[9]; sensor scheduling for target tracking [10]–[15]; and visual cue selection for robotic navigation [16], [17]. Notably, in all aforementioned papers the
underlying optimization problem is of the form
max
f (A),
max
min
A⊆V,|A|≤α B⊆A,|B|≤β
I. I NTRODUCTION
A⊆V,|A|≤α
But sensors and actuators fail [19]; features can become
obsolete [20]; and leaders can be attacked [21]. For example,
sensors may fail due to malfunctions or adversarial attacks.
In such scenarios, questions such as the following arise:
• Where to place a few actuators in a system, when some
of them may fail? [22]
• Which sensors to activate to track an adversarial target that
can jam a fraction of the activated sensors? [23], [24]
• Or, which features to select to train a robust machine
learning model to changing features? [25]
In such scenarios, the optimization problem we need to
address takes the form
(1)
where the set function f exhibits monotonicity and submodularity, a diminishing returns property [1]–[17]. In words,
Problem (1) aims to find a set A of α elements from the
finite ground set V, such that A maximizes f . This problem
is NP-hard, yet several good approximation algorithms have
been proposed for its solution, such as the greedy [18].
1 The authors are with the Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104-6228 USA
(email: {vtzoumas, kgatsis, pappasg}@seas.upenn.edu).
2 The author is with the Institute for Data, Systems and Society, Massachusetts Institute of Technology, Cambridge, MA 02139 USA (email:
[email protected]).
This work was supported in part by TerraSwarm, one of six centers of
STARnet, a Semiconductor Research Corporation program sponsored by
MARCO and DARPA, and in part by AFOSR Complex Networks Program.
f (A \ B),
(2)
where β ≤ α, which is a resilient formulation of Problem (1).
In words, Problem (2) aims to find a set A of α elements
such that A is resilient against the worst possible removal
of β of its elements. Importantly, this formulation is suitable
when we have no prior on the failure or attack mechanism.
The most relevant papers on the resilient monotone
submodular optimization Problem (2) are [19] and [26].
In particular, Problem (2) was introduced in [19], where
the authors proposed an approximation algorithm for the
more general problem maxA⊆V,|A|≤α mini∈{1,2,...,m} fi (A),
where each fi is monotone submodular. In more detail, the
algorithm in [19] guarantees a high value for Problem (2) by
allowing sets A up to size α[1 + 2 log(αβ log(|V|))], instead
β
of α. Nonetheless, it runs with O(|V|2 ( α
β ) ) evaluations of
f , which is exponential in the number of possible removals
β, and quadratic in the number of available elements for
resiliency |V|. This limits its applicability in large-scale
settings where both β and |V| can be in the order of several
thousands [27]. In [26], and its arxiv version [28], the authors
prove that Problem (2) is NP-hard. In addition, they provide
an algorithm for Problem (2) that runs with O(|V|(α − β))
evaluations of f , and which, for α, β −→ +∞, guarantees
an approximate value at least ≃ 29% the optimal, the first
approximation performance bound for Problem (2). However,
in [26] the proposed algorithm is valid only √
when the number
β of possible failures and attacks is up to 2α, whereas in
practice the number β of failures and attacks can be of the
same order as the number α of selected sensors, actuators,
etc. For example, it may be the case that β is up to α/2 [29].
In this paper, we show how the notion of the curvature —
deviation from modularity (additivity)— of a function can
be used to provide a scalable algorithm for the resilient
maximization Problem (2) that has maximum resiliency,
and for submodular functions with low curvature, superior
approximation performance. Notably, low curvature submodular functions are involved in a series of applications [12],
[30], such as sensor placement for mutual information maximization [10]; feature selection for Gaussian process regression [12]; and active learning for speech recognition [17].
In more detail, exploiting the curvature of the monotone
submodular function f , denoted henceforth by κf , in the
resilient maximization Problem (2), we provide an algorithm
(Algorithm 1) with the properties:
• Algorithm 1 is valid for any number of selections for
resiliency α and any number of failures or attacks β;
• Algorithm 1 runs with O(|V|(α − β)) evaluations of f ;
• Algorithm 1 guarantees the approximation performance
bound
fAlgorithm 1
1
1
≥ max 1 − κf ,
1 − e−κf ,
⋆
f
β + 1 κf
where f ⋆ is the (optimal) value of Problem (2), and
fAlgorithm 1 is the (approximate) value achieved by Algorithm 1 for Problem 2. Notably, the notion of curvature
we use for the monotone submodular function f is such
that the curvature κf takes only the values 0 ≤ κf ≤ 1.
Overall, Algorithm 1 improves upon the state-of-the-art
algorithms for Problem (2) as follows:
• High resiliency: Algorithm 1 is the first scalable algorithm
for Problem (2) with bounded approximation performance
for any number of failures or attacks β ≤ α.
• High approximation performance: For low curvature values (κf ≤ 0.71), Algorithm 1 is the first scalable algorithm
to exhibit approximation performance for Problem (2) at
least 29% the optimal.
For example, for the central problem in machine learning
of Gaussian process regression with RBF kernels [10],
[31], Algorithm 1 guarantees almost exact approximation
performance (approximate value ≃ 100% the optimal).
II. R ESILIENT S UBMODULAR M AXIMIZATION
We state the resilient maximization problem considered
in this paper. To this end, we start with the definitions of
monotone and submodular set functions.
Notation: For any set function f : 2V 7→ R on a ground
set V, and any element x ∈ V, f (x) denotes f ({x}).
Definition 1 (Monotonicity): Consider any finite ground
set V. The set function f : 2V 7→ R is non-decreasing if
and only if for any A ⊆ A′ ⊆ V, f (A) ≤ f (A′ ).
N
In words, a set function f : 2V 7→ R is non-decreasing if
and only if adding more elements in any set A ⊆ V cannot
decrease the value of f (A).
Definition 2 (Submodularity [32, Proposition 2.1]):
Consider any finite ground set V. The set function
f : 2V 7→ R is submodular if and only if
′
′
• for any sets A ⊆ V and A ⊆ V, it is f (A) + f (A ) ≥
′
′
f (A ∪ A ) + f (A ∩ A );
′
• equivalently, for any sets A ⊆ A ⊆ V and any element
x ∈ V, it is f (A ∪ {x}) − f (A) ≥ f (A′ ∪ {x}) − f (A′ ).
N
In words, a set function f : 2V 7→ R is submodular if and
only if it satisfies the following intuitive diminishing returns
property: for any element x ∈ V, the marginal gain f (A ∪
{x})− f (A) diminishes as the set A grows; equivalently, for
any set A ⊆ V and any element x ∈ V, the marginal gain
f (A ∪ {x}) − f (A) is non-increasing.
In this paper, we consider the problem of resilient monotone submodular maximization, defined as follows.
Problem 1: Consider
a finite ground set V;
a submodular and monotone set function f : 2V 7→ R
such that (without loss of generality) f is non-negative
and f (∅) = 0;
• and two integers α and β such that 0 ≤ β ≤ α ≤ |V|.
The problem of resilient monotone submodular maximization
is to select a set A of α elements from the ground set V, such
that f (A) is resilient against the worst possible removal B
of β of A’s elements. In mathematical notation,
•
•
max
min
A⊆V,|A|≤α B⊆A,|B|≤β
f (A \ B).
N
The resilient maximization Problem 1 may be interpreted
as a two-stage perfect information sequential game [33,
Chapter 4], where the player that plays first chooses a set A,
and the player that plays second, knowing A, chooses a setremoval B from A.
III. A LGORITHM
FOR RESILIENT SUBMODULAR
MAXIMIZATION
We present the first scalable algorithm for the resilient
maximization Problem 1, and show that this algorithm is
valid for any number of failures and attacks, and that for
functions with low curvature it guarantees superior approximation performance. We begin by presenting the definition
of curvature of monotone submodular functions.
A. Curvature of monotone submodular functions
We define the curvature of monotone submodular functions, which we use to quantify the approximation performance of the proposed algorithm in this paper. To this end,
we start with the definition of modular set functions.
Definition 3 (Modularity): Consider any finite ground
set V. The set function f : 2V 7→ RPis modular if and only
N
if for any set A ⊆ V, it is f (A) = v∈A f (v).
In words, a set function f : 2V 7→ R is modular if through
f all elements in the ground set V cannot substitute each
other, since Definition 3 implies that for any set A ⊆ V and
any element x ∈ V \ A, it is f ({x} ∪ A) − f (A) = f (x);
that is, in the presence of A, x retains its contribution
to the value of f ({x} ∪ A). In contrast, for a submodular set function g : 2V 7→ R, the elements in V can
substitute each other, since Definition 2 implies g({x} ∪
A) − g(A) ≤ g(x); that is, in the presence of A, x may
lose part of its contribution to the value of g({x} ∪ A).
Definition 4 (Curvature): Consider a finite ground set V
and a monotone submodular set function f : 2V 7→ R such
that (without loss of generality) for each element v ∈ V, it
is f (v) 6= 0. The curvature of f is defined as
κf = 1 − min
v∈V
f (V) − f (V \ {v})
.
f (v)
N
In words, the curvature of a monotone submodular function f : 2V 7→ R measures how far f is from modularity:
in particular, per Definition 2 of submodularity, it follows
that curvature takes values 0 ≤ κf ≤ 1, and
• κf = 0 if and only if for all elements v ∈ V, it is f (V) −
f (V \ {v}) = f (v), that is, f is modular.
• κf = 1 if and only if there exist an element v ∈ V such
that f (V) = f (V \ {v}), that is, in the presence of V \ {v},
v loses all its contribution to the overall value of f (V).
An example of a monotone submodular function with zero
total curvature is the trace of the controllability matrix, which
captures the control effort to drive the system in the state
space [5]. A function with curvature 1 is the matroid rank
function [34]. At the same time, many practically interesting
functions have curvature strictly smaller than 1, such as the
concave over modular functions [34, Section 2.1], and the
log det of positive-definite matrices [12], [35], which are
used in applications such as speech processing [36], computer vision [37], feature selection for Gaussian process regression [10], and sensor scheduling for target tracking [14].
The notion of total curvature has served to tighten bounds
for several submodular maximization problems, e.g., for the
non-resilient optimization problem maxA⊆V,|A|≤α f (A) the
approximation bound of the greedy is tightened from the
value 1 − 1/e to κ1f (1 − e−κf ) [34], [38], [39]. Nonetheless,
for resilient submodular maximization problems, such as
Problem 1, the role of the curvature has not been addressed
yet. We provide the first results to this end in the next section.
B. Algorithm for resilient submodular maximization
We exploit the curvature Definition 4 to provide for the
resilient maximization Problem 1 Algorithm 1. Algorithm 1
1
g(κf ) =
1−κf
κf
(1 − e−κf )
0.8
g(κf )
Algorithm 1 Algorithm for Problem 1.
Input: Per Problem 1, three are the inputs to Algorithm 1:
• finite ground set V = {v1 , v2 , . . . , vm }, where m =
|V|;
V
• submodular and monotone set function f : 2 7→ R
such that f is non-negative and f (∅) = 0;
• and two integers α and β such that 0 ≤ β ≤ α ≤ |V|.
Output: Set ARES , with properties per Theorem 1.
1: A1 ← ∅, A2 ← ∅
′
2: Sort elements in V such that V = {v1′ , v2′ , . . . , vm
} and
′
′
′
f (v1 ) ≥ f (v2 ) ≥ . . . ≥ f (vm )
′
3: A1 ← {v1′ , v2′ , . . . , vβ
}
4: while |A2 | < α − β do
5:
x ∈ arg maxy∈V\(A1 ∪A2 ) f (y|A2 )
6:
A2 ← {x} ∪ A2
7: end while
8: AR ES ← A1 ∪ A2
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
κf
Fig. 1. Plot of g(κf ) versus curvature κf of a monotone submodular
function f . By definition, the curvature κf of a monotone submodular
function f takes values between 0 and 1. g(κf ) increases from 0 to 0.5 as
κf decreases from 1 to 0.
returns a solution for Problem 1, denoted by ARES , in two
steps: first, in lines 1-3 Algorithm 1 selects a set A1 , which is
composed of β elements from the ground set V. Specifically,
per line 2, each element v ∈ A1 is such that for all elements
v ′ ∈ V \ A1 , it is f (v) ≥ f (v ′ ). Second, in lines 4-8,
Algorithm 1 selects greedily from the set V \ A1 a set A2 ,
which is composed of α − β elements, and then, in line 8,
Algorithm 1 returns set ARES as the union of A1 and A2 .
Algorithm 1’s performance is quantified in Theorem 1,
whose proof can be found in the appendix. The intuition
behind Algorithm 1 is discussed in Section III-C.
Theorem 1: Per Problem 1, let
•
•
the real number f ⋆ equal to the (optimal) value of Problem 1, i.e., f ⋆ = maxA⊆V,|A|≤α minB⊆A,|B|≤β f (A \ B);
and for any set A ⊆ V, the set B ⋆ (A) equal to the
optimal set-removal of β elements from A, i.e., B ⋆ (A) ∈
minB⊆A,|B|≤β f (A \ B).
The following two hold on Algorithm 1’s performance:
1) Algorithm 1 returns a set ARES ⊆ V such that |ARES | ≤ α,
and
f (ARES \ B ⋆ (ARES )) ≥
1
1
max 1 − κf ,
1 − e−κf f ⋆ ,
β + 1 κf
and in particular, for κf = 0, f (ARES \ B ⋆ (ARES )) = f ⋆ .
2) Algorithm 1 runs in O(|V|(α − β)) evaluations of f . N
Remark 1: Given any finite ground set V and monotone
submodular function f : 2V 7→ R in the resilient maximization Problem 1, the approximation bound of Algorithm 1
depends on the curvature value which is computed with
O(|V|) evaluations of f according to its Definition 4.
N
Remark 2: Given any finite ground set V, finite number
of failures or attacks β and monotone submodular function
f : 2V 7→ R in the resilient maximization Problem 1, the
approximation bound
1 is non-zero, since for
of Algorithm
1
1
≥ β+1
> 0, and for all
finite β it is max 1 − κf , β+1
1
−κf
) ≥ 1 − 1/e ≃ 0.63 > 0. N
0 ≤ κf ≤ 1, it is κf (1 − e
Remark 3: Per Theorem 1, Algorithm 1’s approximation performance bound is in the worst case equal to
1−κf
−κf
), which is plotted in Fig. 1, and is approxκf (1 − e
imately equal to −κf + 1.
N
Remark 4: For zero curvature, Algorithm 1 solves Problem 1 exactly. For non-zero curvature, in which case Problem 1 is NP-hard [28], Algorithm 1’s approximation bound
1−κ
is in the worst-case equal to κf f (1 − e−κf ), which tends
to 1 as κf −→ 0. Overall, Algorithm 1’s curvature-dependent
approximation bound makes a first step towards separating the class of monotone submodular functions into the
functions for which the resilient maximization Problem 1
can be approximated well (low curvature functions), and
the functions for which it cannot (high curvature functions).
The reason is that Algorithm 1’s approximation bound increases as the curvature decreases, and for zero curvature
it becomes 1 —i.e., for zero curvature, Algorithm 1 solves
Problem 1 exactly. This role of the curvature in Problem 1 is
similar to the role that the curvature has played in the nonresilient variant of Problem 1, i.e., the optimization problem
maxA⊆V,|A|≤α f (A), where it has been used to separate the
class of submodular functions into the functions for which
maxA⊆V,|A|≤α f (A) can be approximated well (low curvature functions), and the functions for which it cannot (high
curvature functions) [34], [38]. Hence, Theorem 1 also supports the intuition that the resilient maximization Problem 1
is easier when the non-resilient variant maxA⊆V,|A|≤α f (A)
is easy, and it is harder when maxA⊆V,|A|≤α f (A) is hard.
N
Theorem 1 implies for Algorithm 1’s performance:
•
•
Algorithm 1 is the first scalable algorithm for Problem 1
that is valid for any number of failures or attacks β, and
any number of selections for resiliency α. In particular,
the previously proposed algorithms for Problem 1 in [19]
and [26] are such that: the algorithm in [19] runs in
exponential time in β, and quadratic in the cardinality of
the ground set V and, as a result, has limited applicability
in large-scale
√ settings. The algorithm in [26] is valid only
for β ≤ 2α and, as a result, has limited applicability in
decision and control settings where the number of failures
and attacks β can be up to the number of placed sensors,
√
actuators, etc. α. For example, the inequality β ≤ 2α is
violated when among 100 placed sensors, 20 may fail.
Algorithm 1 is the first scalable algorithm for Problem 1 that for non-zero curvature values κf ≤ 0.71,
and any number of failures and attacks β, guarantees
approximation performance more than at least 29% the
optimal. In particular, the previously proposed algorithms
for Problem 1 in [19] and [26] are such that: the algorithm
in [19] runs in exponential time in β, and quadratic in the
cardinality of the ground set V and, as a result, has limited
applicability in large-scale settings. The algorithm in [26],
√
under the constraint β ≤ 2α, and for α, β −→ +∞,
guarantees an approximate value up to at least 29% the
optimal. In particular, for α, β < +∞, its approximation
performance is less than at least 29% the optimal.
An example of a central problem in machine learning,
with applications, e.g., in sensor placement, where Algorithm 1 guarantees almost exact approximation performance
(≃ 100% the optimal) is that of Gaussian process regression
for Gaussian processes with RBF kernels [10], [31]. The
reason is that in this class of problems the objective function
is the entropy of the selected measurements, which for
Gaussian processes with RBF kernels was shown recently
to have curvature values close to zero [12, Theorem 5].
C. Intuition behind curvature-dependent algorithm for resilient maximization
We explain the intuition behind Algorithm 1 for Problem 1, and give also an illustrative example. To this end,
we focus only on the NP-hard case where the curvature of
Problem 1’s objective function is non-zero [28, Lemma 3].
In addition, we use the notation introduced in the statements
of Algorithm 1 and of Theorem 1.
We explain how Algorithm 1 works first for the case where
the optimal set-removal B ⋆ (ARES ) is equal to the set A1 , and
then for the case where B ⋆ (ARES ) 6= A1 . Notably, the case
B ⋆ (ARES ) = A1 is possible since in lines 1-3 Algorithm 1
selects a set A1 such that |A1 | = β, and per Problem 1,
the number β is the number of possible removals.
a) Intuition behind Algorithm 1 for B ⋆ (ARES ) = A1 :
The two parts of Algorithm 1 operate in tandem to guarantee
1
1 − e−κf f ⋆ .
(3)
f (ARES \ B ⋆ (ARES )) ≥
κf
This happens as follows: B ⋆ (ARES ) = A1 implies that
ARES \ B ⋆ (ARES ) = A2 , since ARES = A1 ∪ A2 , per line 8
of Algorithm 1. Therefore, f (ARES \ B ⋆ (ARES )) = f (A2 ).
But in lines 4-7 of Algorithm 1, A2 is selected greedily and,
as a result, using [38, Theorem 5.4] we have
1
1 − e−κf
max
f (A).
f (A2 ) ≥
κf
A⊆V\A1 ,|A|≤α−β
Finally, it is maxA⊆V\A1 ,|A|≤α−β f (A) ≥ f ⋆ , since the lefthand-side of this inequality is the maximum value one can
achieve for f by choosing α − β elements from the ground
set V knowing that A1 has been removed from V, whereas
the right-hand-side is the maximum value one can achieve
for f by choosing α elements from V not knowing which β
of them will be optimally removed; a mathematical version
of the latter proof can be found in [28, Lemma 2].
b) Intuition behind Algorithm 1 for B ⋆ (ARES ) 6= A1 :
The two parts of Algorithm 1 operate in tandem to guarantee,
f (ARES \ B ⋆ (ARES )) ≥
1
1
1 − e−κf f ⋆ ,
max 1 − κf ,
β + 1 κf
(4)
This happens as follows: B ⋆ (ARES ) =
6 A1 implies, along
with |B ⋆ (ARES )| = |A1 | = β, that if µ elements in A1
Overall, in the worst-case Algorithm 1 guarantees for the
resilient maximization Problem 1 the approximation performance bound in ineq. (4), as stated in Theorem 1, since for all
values of the curvature κf , the bound in ineq. (4) is smaller
than the bound in ineq. (3), since max [1 − κf , 1/(β + 1)] ≤
1, and we do not know a priori which of the two preceding
bounds holds at a problem instance.
Example 1: We use an instance of Problem 1 to illustrate
how Algorithm 1 finds an approximate solution to Problem 1,
as well as, how it performs. We consider the following
instance: let α = 2, β = 1, V = {v1 , v2 , v3 }, and f such
that f (∅) = 0, f (v3 ) > 0, f (v1 ) = f (v3 ) + 1, f (v2 ) =
f (v3 ) + 0.5, f ({v1 } ∪ {v2 }) = f (v3 ) + 1, f ({v1 } ∪ {v3 }) =
f (V) = 2f (v3 ) + 1, and f ({v2 } ∪ {v3 }) = 2f (v3 ) + 0.5.
For the aforementioned instance, the curvature is κf =
1 and, as a result, Algorithm 1 is guaranteed to return a
set ARES such that the approximation ratio is either at least
1 − 1/e, per bound in ineq. (3), or at least (1 − 1/e)/2, per
bound in ineq. (4).
Algorithm 1 selects ARES = {v1 , v2 }, which in this example is the exact solution to Problem 1. In particular, in lines 23, Algorithm 1 selects A1 = {v1 }, and in lines 4-7, it selects
A2 = {v2 }. The optimal set-removal is B ⋆ (ARES ) = {v1 }.
The reason that Algorithm 1 performs optimally in this
example, which is not expected by Theorem 1 since it has
the worst curvature value κf = 1, is as follows: In lines 1-3
Algorithm 1 selects A1 = {v1 }, which for ARES = {v1 , v2 }
is the element that will be included in the optimal removal
B ⋆ (ARES ). That is, in this example B ⋆ (ARES ) = A1 , which
implies that the approximation performance of Algorithm 1 is
bounded by 1−1/e, as in ineq. (3). This is the first important
observation towards explaining the optimal performance of
Algorithm 1 in this example; the second and final necessary
observation is as follows: In lines 4-7, Algorithm 1 selects
the best element in V assuming that the element in A1 will
be included in B ⋆ (ARES ), i.e., removed from ARES , since
it selects greedily from V \ A1 . In contrast, if in lines 4-7
Algorithm 1 would select greedily from V without taking
into account that the element in A1 is going to be included
in B ⋆ (ARES ), then it would select ARES = {v1 , v3 } which is
suboptimal for Problem 1.
N
100
Empirical approximation ratio (%)
are not included in B ⋆ (ARES ), exactly µ elements in A2 are
included in B ⋆ (ARES ). Therefore, f (ARES \ B ⋆ (ARES )) can
take a bounded value similar to the one in ineq. (3) only
if the µ elements in A1 that are not included in B ⋆ (ARES )
can compensate for the µ elements in A2 that are included
in B ⋆ (ARES ). For this reason, in line 2, Algorithm 1 chooses
the elements in A1 so that they a have higher value than those
in V \A1 : In particular, using the fact that the elements in A1
have a higher value than those in V \ A1 , and the definition
of the curvature κf , in the proof of Theorem 1 we bound
how much value the elements in A1 that are not included in
B ⋆ (ARES ) can compensate for the value of the elements in
A2 that are included in B ⋆ (ARES ), and conclude ineq. (4).
99
98
β
β
β
β
β
β
97
96
95
8
9
10
11
12
13
=1
=2
=3
=4
=5
=6
14
15
|V|
Fig. 2. Empirical approximation performance of Algorithm 1. For details,
see Section IV.
IV. S IMULATIONS
We empirically test Algorithm 1’s approximation performance for Problem 1 against an exact, brute force algorithm.
As a test function, with consider one of the following form:
Given a finite ground set V, and |V| positive semi-definite
matrices D1 , D2 , . . . , D|V| ,P
for any set A ⊆ V, let the set
function f (A) = log det( i∈A Di + I), where I is the
identity matrix. Functions of this form appear in applications
such as sensor selection for Gaussian process regression [10],
and actuator placement for bounded control effort [7], [40].
To run our simulations, and be able to compute the exact
value to Problem 1, we select small sizes of |V| from 8 to 15.
In addition, we fix the number of selections for resiliency α
to 7, vary the number of attacks/failures β from 1 to 6, and
for each of the aforementioned cases, generate 10 random
instances of the matrices D1 , D2 , . . . , D|V| of size 20 × 20.
Our simulations are summarized in Fig. 2, where Algorithm 1 is seen to perform close to 98% the optimal, and
its approximation performance to degrade as β increases up
to α, which is equal to 7. Notably, for all generated instances
of f , f ’s curvature takes values larger than 0.9, for which,
Algorithm 1’s worst-case theoretical approximation bound
in Theorem 1 is at least 14%, since for the maximum value
of β, which is equal to 6 in this simulation example, Algorithm 1’s approximation bound per Theorem 1 is max[1/(β+
1), 1 − κf ] = 1/(β + 1) = 0.14. Overall, Algorithm 1’s
approximation performance in Fig. 2 is in accordance with
the empirical observation that greedy-type algorithms for
submodular maximization often outperform in practice their
worst-case theoretical approximation guarantees [10], [12].
V. C ONCLUDING
REMARKS
& F UTURE
WORK
We focused on the resilient submodular maximization
Problem 1, which is central in machine learning, optimization, and control, in applications such as feature selection for
classifier training, and sensor scheduling for target tracking.
In particular, exploiting the notion of curvature, we provided
the first scalable algorithm for Problem 1, Algorithm 1,
which is valid for any number of attacks or failures, and
which, for functions with low curvature, guarantees superior
approximation performance. In addition, for functions with
zero curvature, Algorithm 1 solves Problem 1 exactly. Overall, Algorithm 1’s approximation bound makes a first step
to characterize the curvature’s effect on approximations for
resilient submodular maximization problems, complementing
that way the current knowledge on the curvature’s effect
on non-resilient submodular maximization. Future work will
focus on Algorithm 1’s extension to matroid constraints.
A PPENDIX : P ROOF
OF
T HEOREM 1
Notation: Given a set function f : 2V 7→ R, for any
sets X ⊆ V and X ′ ⊆ V, the f (X |X ′ ) denotes f (X ∪ X ′ ) −
f (X ′ ). The set A⋆ denotes a solution to Problem 1, i.e.,
A⋆ ∈ arg maxA⊆V,|A|≤α minB⊆A,|B|≤β f (A \ B).
A. Proof of Algorithm 1’s approximation performance:
We complete the proof first for curvature value κf equal
to 0, and then for κf 6= 0.
1) Proof of Algorithm 1’s approximation performance for
κf = 0: In this case, Algorithm 1 solves Problem 1 exactly.
This follows from the two observations below:
•
•
For κf = 0, Algorithm 1 returns a set ARES such that for all
elements v ∈ ARES and v ′ ∈ V \ ARES , it is f (v) ≥ f (v ′ ).
The reasons are two: first,the set A1 is such that for all
elements v ∈ A1 and v ′ ∈ V \ A1 , f (v) ≥ f (v ′ ); and
second the set A2 is such that for all elements v ∈ A2
and v ′ ∈ V \ (A1 ∪ A2 ), it is f (v) ≥ f (v ′ ), since in line 5
of Algorithm 1 it is f (y|A1 ∪A2 ) = f (y), which is implied
from the fact that κf = 0 implies that f is modular.
For κf = 0, the set A⋆ is also such that for all elements
v ∈ A⋆ and v ′ ∈ V \ A⋆ , it is f (v) ≥ f (v ′ ). To explain
why, we first make two observations: first, κf = 0 implies
that f is modular, which in turn implies
P that for any sets
A ⊆ V and B ⊆ A, it is f (A \ B) = v∈A\B f (v); and
second, Problem 1 considers (without loss of generality)
that f is non-negative, which implies that for all v ∈ V,
it is f (v) ≥ 0. Therefore, for any A, the optimal setremoval B ⋆ (A) contains β elements in A such that for all
elements x ∈ B ⋆ (A) and y ∈ A, it is f (x) ≥ f (y). Thus,
A⋆ maximizes f (A\B ⋆ (A)) if and only if for all elements
v ∈ A and v ′ ∈ V \ A, it is f (v) ≥ f (v ′ ).
2) Proof of Algorithm 1’s approximation performance for
κf =
6 0: We use the symbols B1⋆ and B2⋆ , defined in Fig. 3.
We complete the proof of Algorithm 1’s approximation
bound for κf 6= 0 in two steps: First, we consider that
B ⋆ (ARES ) = A1 , i.e., B2⋆ = ∅. Second, we consider that
B ⋆ (ARES ) 6= A1 , i.e., B2⋆ 6= ∅.
For the case B2⋆ = ∅, we prove
1
1 − e−κf f (A⋆ \B ⋆ (A⋆ )). (5)
f (ARES \B ⋆ (ARES )) ≥
κf
V
A1
B1⋆
A2
B2⋆
Fig. 3.
Venn diagram, where A1 , A2 , B1⋆ , B2⋆ are as follows: Per
Algorithm 1, A1 and A2 are such that ARES = A1 ∪A2 , and A1 ∩A2 = ∅.
Also, B1⋆ and B2⋆ are such that B1⋆ = B⋆ (ARES ) ∩ A1 , and B2⋆ =
B⋆ (ARES ) ∩ A2 ; by definition, B1⋆ ∩ B2⋆ = ∅ and B⋆ (ARES ) = B1⋆ ∪ B2⋆ .
In particular, we prove ineq. (5) by making the following
three consecutive observations: first, f (ARES \ B ⋆ (ARES )) =
f (A2 ), since B ⋆ (ARES ) = A1 , because B2⋆ = ∅. Second,
1
1 − e−κf
max
f (A), (6)
f (A2 ) ≥
κf
A⊆V\A1 ,|A|≤α−β
since in line 5 Algorithm 1 constructs greedily the set A2
using elements from V \ A1 [38, Theorem 5.4]; and third,
max
A⊆V\A1 ,|A|≤α−β
f (A) ≥ f (A⋆ \ B ⋆ (A⋆ )),
(7)
because of [28, Lemma 2], where we recall that the set A⋆
denotes an (optimal) solution to Problem 1. The above three
consecutive observations complete the proof of ineq. (5), and
end our focus on the case where B2⋆ = ∅.
For the case B2⋆ 6= ∅, we prove first that
f (ARES \ B ⋆ (ARES )) ≥
1 − κf
1 − e−κf
κf
f (A⋆ \ B ⋆ (A⋆ )), (8)
and then we complete the proof of Algorithm 1’s approximation performance bound by also proving that
f (ARES \ B ⋆ (ARES )) ≥
1 1
1 − e−κf f (A⋆ \ B ⋆ (A⋆ )).
β + 1 κf
(9)
To prove ineq. (8), we use the following lemma.
Lemma 1: Consider a finite ground set V and a monotone
set function f : 2V 7→ R such that f is non-negative and
f (∅) = 0. For any set A ⊆ V,
X
f (A) ≥ (1 − κf )
f (a).
N
a∈A
Proof of Lemma 1: Let A = {a1 , a2 , . . . , a|A| }. We prove
Lemma 1 by proving the following two inequalities:
f (A) ≥
|A|
X
i=1
|A|
X
i=1
f (ai |V \ {ai }),
f (ai |V \ {ai }) ≥ (1 − κf )
|A|
X
i=1
f (ai ).
(10)
(11)
We begin with the proof of ineq. (10):
f (A) = f (A|∅)
≥ f (A|V \ A)
(12)
(13)
X
f (ai |V \ {ai , ai+1 , . . . , a|A| })
(14)
X
f (ai |V \ {ai }),
(15)
η=
|A|
=
i=1
|A|
≥
i=1
where ineqs. (13) to (15) hold for the following reasons:
ineq. (13) is implied by eq. (12) because f is submodular
and ∅ ⊆ V \ A; eq. (14) holds since for any sets X ⊆ V
and Y ⊆ V it is f (X |Y) = f (X ∪ Y) − f (Y), and it also
{a1 , a2 , . . . , a|A| } denotes the set A; and ineq. (15) holds
since f is submodular and V \ {ai , ai+1 , . . . , aµ } ⊆ V \ {ai }.
These observations complete the proof of ineq. (10).
We now prove ineq. (11) using the Definition 4 of κf , as
follows: since κf = 1 − minv∈V f (v|V\{v})
, it is implied that
f (v)
for all elements v ∈ V it is f (v|V \ {v}) ≥ (1 − κf )f (v).
Therefore, adding the latter inequality across all elements
a ∈ A completes the proof of ineq. (11).
In addition to the above lemma, to prove ineq. (8) we use
the following two notations: first, let the set A+
1 denote
the elements in A1 not included in the optimal set-removal
⋆
B ⋆ (ARES ); notably, A+
1 is non-empty, since B (AR ES ) inter⋆
sects with the set A2 (because B2 6= ∅) and |B ⋆ (ARES )| =
|A1 | and, as a result, at least one element in A1 is not
included in B ⋆ (ARES ). In addition, let the set A+
2 denote
the elements in A2 not included in B ⋆ (ARES ).
We prove ineq. (8) first using Lemma 1 and then taking
into account ineqs. (6) and (7), as follows:
f (ARES \ B ⋆ (ARES ))
+
= f (A+
1 ∪ A2 )
X
≥ (1 − κf )
(16)
f (v)
(17)
≥ (1 − κf )
≥ (1 −
X
f (v) +
v∈A2 \A+
2
κf )f [(A2 \ A+
2)
= (1 − κf )f (A2 ),
X
v∈A+
2
∪ A+
2]
f (v)
f (B2⋆ |ARES \ B ⋆ (ARES ))
f (A2 )
(21)
Later in this proof, we prove that 0 ≤ η ≤ 1. We prove
ineq. (9) by first observing that
f (ARES \ B ⋆ (ARES )) ≥ max{f (ARES \ B ⋆ (ARES )), f (A+
1 )},
(22)
and then proving the following three inequalities:
f (ARES \ B ⋆ (ARES )) ≥ (1 − η)f (A2 ),
1
f (A+
1 ) ≥ η f (A2 ),
β
1
1
max{(1 − η), η } ≥
.
β
β+1
(23)
(24)
(25)
Specifically, if we substitute ineqs. (23), (24) and (25)
to (22), and take into account that f (A2 ) ≥ 0, then
f (ARES \ B ⋆ (ARES )) ≥
1
f (A2 ),
β+1
which implies ineq. (9) after taking into account ineqs. (6) and (7).
In the remaining paragraphs, we complete the proof of
ineq. (9) by proving 0 ≤ η ≤ 1, and ineqs. (23), (24)
and (25), respectively.
a) Proof of ineq. 0 ≤ η ≤ 1: We first prove that η ≥ 0,
and then that η ≤ 1: i) η ≥ 0, since by definition η =
f (B2⋆ |ARES \ B ⋆ (ARES ))/f (A2 ), and f is non-negative; and
ii) η ≤ 1, since f (A2 ) ≥ f (B2⋆ ), due to monotonicity of f
and that B2⋆ ⊆ A2 , and f (B2⋆ ) ≥ f (B2⋆ |ARES \ B ⋆ (ARES )),
due to submodularity of f and that ∅ ⊆ ARES \ B ⋆ (ARES ).
b) Proof of ineq. (23): We complete the proof of
ineq. (23) in two steps. First, it can be verified that
f (ARES \ B ⋆ (ARES )) = f (A2 )−
+
v∈A+
1 ∪A2
To complete the proof of Algorithm 1’s approximation
performance bound it remains to prove ineq. (9). To this
end, we denote
(18)
(19)
(20)
where eq. (16) to (20) hold for the following reasons:
+
eq. (16) follows from the definitions of the sets A+
1 and A2 ;
ineq. (17) follows from ineq. (16) due to Lemma 1; ineq. (18)
follows from ineq. (17) because for all elements v ∈ A+
1
′
and all elements v ′ ∈ A2 \ A+
2 it is f (v) ≥ f (v ) (note that
+
+
due to the definitions of the sets A+
1 and A2 it is |A1 | =
+
|A2 \ A2 |, that is, the number of non-removed elements in
A1 is equal to the number of removed elements in A2 );
finally, ineq. (19) follows from ineq. (18) because the set
function f is submodular and, as a result, the submodularity
Definition 2 implies that for any sets A ⊆ V and A′ ⊆ V, it
is f (A) + f (A′ ) ≥ f (A ∪ A′ ). Now, ineq. (8) follows from
ineq. (20) by taking into account ineqs. (6) and (7).
f (B2⋆ |ARES \ B ⋆ (ARES )) + f (A1 |A2 ) − f (B1⋆ |ARES \ B1⋆ ),
(26)
since for any X ⊆ V and Y ⊆ V, f (X |Y) = f (X ∪
Y) − f (Y). Second, (26) implies (23), since f (B2⋆ |ARES \
B ⋆ (ARES )) = ηf (A2 ), and f (A1 |A2 )−f (B1⋆ |ARES \B1⋆ ) ≥ 0.
The latter is true due to the following two observations:
i) f (A1 |A2 ) ≥ f (B1⋆ |A2 ), since f is monotone and B1⋆ ⊆
A1 ; and ii) f (B1⋆ |A2 ) ≥ f (B1⋆ |ARES \ B1⋆ ), since f is
submodular and A2 ⊆ ARES \ B1⋆ (see also Fig. 3).
c) Proof of ineq. (24): We use the following lemma.
Lemma 2: Consider any finite ground set V, a monotone
submodular function f : 2V 7→ R and non-empty sets Y, P ⊆
V such that for all elements y ∈ Y and all elements p ∈ P
it is f (y) ≥ f (p). Then,
f (P|Y) ≤ |P|f (Y).
N
Proof of Lemma 2: Consider any element y ∈ Y (such
an element exists since Lemma 2 considers that Y is non-
empty); then,
B. Proof of Algorithm 1’s running time
f (P|Y) = f (P ∪ Y) − f (Y)
≤ f (P) + f (Y) − f (Y)
= f (P)
X
≤
f (p)
(27)
(28)
(29)
p∈P
≤ |P| max f (p)
p∈P
≤ |P|f (y)
≤ |P|f (Y),
(30)
(31)
where eq. (27) to ineq. (31) hold for the following reasons:
eq. (27) holds since for any sets X ⊆ V and Y ⊆ V, it is
f (X |Y) = f (X ∪ Y) − f (Y); ineq. (28) holds since f is
submodular and, as a result, the submodularity Definition 2
implies that for any set A ⊆ V and A′ ⊆ V, it is f (A ∪
A′ ) ≤ f (A) + f (A′ ); ineq. (29) holds for the same reason
as ineq. (28); ineq. (30) holds since or all elements y ∈ Y
and all elements p ∈ P it is f (y) ≥ f (p); finally, ineq. (31)
holds because f is monotone and y ∈ Y.
To prove ineq. (24), since it is B2⋆ 6= ∅ (and, as a result,
+
also A+
1 6= ∅) and for all elements a ∈ A1 and all elements
⋆
b ∈ B2 it is f (a) ≥ f (b), from Lemma 2 we have
+
⋆
f (B2⋆ |A+
1 ) ≤ |B2 |f (A1 )
≤ βf (A+
1 ),
since
|B2⋆ |
(32)
≤ β. Overall,
1
f (B2⋆ |A+
1)
β
1
+
≥ f (B2⋆ |A+
1 ∪ A2 )
β
1
= f (B2⋆ |ARES \ B ⋆ (ARES ))
β
1
= η f (A2 ),
β
f (A+
1)≥
(33)
(34)
(35)
(36)
where ineq. (33) to eq. (36) hold for the following reasons:
ineq. (33) follows from ineq. (32); ineq. (34) holds since f
+
+
is submodular and A+
1 ⊆ A1 ∪ A2 ; eq. (35) holds due to
+
+
the definitions of the sets A1 , A2 and B ⋆ (ARES ); finally,
eq. (36) holds due to the definition of η. Overall, the latter
derivation concludes the proof of ineq. (24).
d) Proof of ineq. (25): Let b = 1/β. We complete the
proof first for the case where (1 − η) ≥ ηb, and then for the
case (1−η) < ηb: i) When (1−η) ≥ ηb, max{(1−η), ηb} =
1−η and η ≤ 1/(1+b). Due to the latter, 1−η ≥ b/(1+b) =
1/(β + 1) and, as a result, (25) holds. ii) When (1 − η) < ηb,
max{(1 − η), ηb} = ηb and η > 1/(1 + b). Due to the latter,
ηb > b/(1 + b) and, as a result, (25) holds.
We completed the proof of 0 ≤ η ≤ 1, and ineqs. (23), (24) and (25). Thus, we also completed the proof
of ineq. (9), that is, the last part of the proof of Algorithm 1’s
approximation bound for κf 6= 0.
We complete the proof in two steps, where the time for
each evaluation of f is denoted as τf : we first compute the
time that line 2 of Algorithm 1 needs to be executed, and
then the time that lines 4-7 need to be executed. Line 2 needs
m log(m) + mτf + m + O(log(m)) time, since it asks for m
evaluations of f , and their sorting, which takes m log(m) +
m + O(log(m)) time, using, e.g., the merge sort algorithm.
Lines 4-7 need m(α − β)τf + (α − β)m time, since they ask
for at most m(α − β) evaluations of f , since the while loop
is repeated α − β times, and during each loop at most m
evaluations of f are needed by line 5, and for a maximal
element in line 5, which needs m time to be found. Overall,
Algorithm 1 runs in m(α − β)τf + (α − β)m + m log(m) +
mτf + m + O(log(m)) = O(m(α − β)τf ) time.
R EFERENCES
[1] A. Krause, J. Leskovec, C. Guestrin, J. VanBriesen, and C. Faloutsos,
“Efficient sensor placement optimization for securing large water
distribution networks,” Journal of Water Resources Planning and
Management, vol. 134, no. 6, pp. 516–526, November 2008.
[2] A. Das and D. Kempe, “Submodular meets spectral: Greedy algorithms
for subset selection, sparse approximation and dictionary selection,”
in Proceedings of the 28th International Conference on Machine
Learning, 2011, pp. 1057–1064.
[3] A. Clark, B. Alomair, L. Bushnell, and R. Poovendran, Submodularity
in dynamics and control of networked systems. Springer, 2016.
[4] A. Olshevsky, “Minimal controllability problems,” IEEE Transactions
on Control of Network Systems, vol. 1, no. 3, pp. 249–258, 2014.
[5] F. Pasqualetti, S. Zampieri, and F. Bullo, “Controllability metrics,
limitations and algorithms for complex networks,” IEEE Transactions
on Control of Network Systems, vol. 1, no. 1, pp. 40–52, March 2014.
[6] S. Pequito, S. Kar, and A. Aguiar, “A framework for structural input/output and control configuration selection in large-scale systems,”
IEEE Trans. on Automatic Control, vol. 61, no. 2, pp. 303–318, 2016.
[7] T. H. Summers, F. L. Cortesi, and J. Lygeros, “On submodularity and
controllability in complex dynamical networks,” IEEE Transactions
on Control of Network Systems, vol. 3, no. 1, pp. 91–101, 2016.
[8] V. Tzoumas, M. A. Rahimian, G. J. Pappas, and A. Jadbabaie,
“Minimal actuator placement with bounds on control effort,” IEEE
Trans. on Control of Network Systems, vol. 3, no. 1, pp. 67–78, 2016.
[9] E. Nozari, F. Pasqualetti, and J. Cortes, “Time-varying actuator
scheduling in complex networks,” arXiv preprint:1611.06485, 2016.
[10] A. Krause, A. Singh, and C. Guestrin, “Near-optimal sensor placements in gaussian processes: Theory, efficient algorithms and empirical
studies,” J.l of Machine Learning Research, vol. 9, pp. 235–284, 2008.
[11] S. T. Jawaid and S. L. Smith, “Submodularity and greedy algorithms in
sensor scheduling for linear dynamical systems,” Automatica, vol. 61,
pp. 282–288, 2015.
[12] D. Sharma, A. Kapoor, and A. Deshpande, “On greedy maximization
of entropy,” in Proceedings of the 32nd International Conference on
Machine Learning, 2015, pp. 1330–1338.
[13] V. Tzoumas, A. Jadbabaie, and G. J. Pappas, “Near-optimal sensor
scheduling for batch state estimation,” in IEEE 55th Conference on
Decision and Control, 2016, pp. 2695–2702.
[14] V. Tzoumas, N. A. Atanasov, A. Jadbabaie, and G. J. Pappas,
“Scheduling nonlinear sensors for stochastic process estimation,” in
IEEE American Control Conference, 2017, to appear.
[15] H. Zhang, R. Ayoub, and S. Sundaram, “Sensor selection for kalman
filtering of linear dynamical systems: Complexity, limitations and
greedy algorithms,” Automatica, vol. 78, pp. 202 – 210, 2017.
[16] L. Carlone and S. Karaman, “Attention and anticipation in fast visualinertial navigation,” arXiv preprint:1610.03344, 2016.
[17] R. Iyer, S. Jegelka, and J. Bilmes, “Fast semidifferential-based submodular function optimization,” in Proceedings of the 30th International Conference on International Conference on Machine Learning,
2013, pp. 855–863.
[18] G. L. Nemhauser and L. A. Wolsey, “Best algorithms for approximating the maximum of a submodular set function,” Mathematics of
operations research, vol. 3, no. 3, pp. 177–188, 1978.
[19] A. Krause, H. B. McMahan, C. Guestrin, and A. Gupta, “Robust
submodular observation selection,” Journal of Machine Learning Research, vol. 9, pp. 2761–2801, 2008.
[20] A. Globerson and S. Roweis, “Nightmare at test time: Robust learning
by feature deletion,” in Proceedings of the 23rd international conference on Machine learning, 2006, pp. 353–360.
[21] A. Clark, L. Bushnell, and R. Poovendran, “Leader selection games
under link noise injection attacks,” in Proceedings of the 1st International Conf. on High Confidence Networked Systems, 2012, pp. 31–40.
[22] S. Pequito, G. Ramos, S. Kar, A. P. Aguiar, and J. Ramos, “On
the Exact Solution of the Minimal Controllability Problem,” arXiv
preprint:1401.4209, 2014.
[23] A. Das and D. Kempe, “Sensor selection for minimizing worst-case
prediction error,” in Proceedings of the 7th International conference
on Information processing in sensor networks, 2008, pp. 97–108.
[24] A. Laszka, Y. Vorobeychik, and X. Koutsoukos, “Resilient observation
selection in adversarial settings,” in 2015 54th IEEE Conference on
Decision and Control (CDC), Dec 2015, pp. 7416–7421.
[25] W. Xu, K. Ma, W. Trappe, and Y. Zhang, “Jamming sensor networks,”
IEEE Network, vol. 20, no. 3, pp. 41–47, 2006.
[26] J. B. Orlin, A. S. Schulz, and R. Udwani, “Robust monotone submodular function maximization,” in 18th International Conf. on Integer
Programming and Combinatorial Optimization, 2016, pp. 312–324.
[27] Y. Shoukry, M. Chong, M. Wakaiki, P. Nuzzo, A. L. SangiovanniVincentelli, S. A. Seshia, J. P. Hespanha, and P. Tabuada, “SMT-based
observer design for cyber-physical systems under sensor attacks,” in
Proc. of the 7th International Conf. on Cyber-Physical Systems, 2016.
[28] J. B. Orlin, A. S. Schulz, and R. Udwani, “Robust monotone submodular function maximization,” arXiv preprint:1507.06616, 2015.
[29] M. S. Chong, M. Wakaiki, and J. P. Hespanha, “Observability of linear
systems under adversarial attacks,” in American Control Conference,
2015, pp. 2439–2444.
[30] S. Jegelka, “Combinatorial problems with submodular coupling in
machine learning and computer vision,” Ph.D. dissertation, ETH
Zurich, 2012.
[31] C. M. Bishop, Pattern recognition and machine learning. Springer,
2006.
[32] G. Nemhauser, L. Wolsey, and M. Fisher, “An analysis of approximations for maximizing submodular set functions – I,” Mathematical
Programming, vol. 14, no. 1, pp. 265–294, 1978.
[33] R. B. Myerson, Game theory: Analysis of conflict. Harvard University
Press, 2013.
[34] R. K. Iyer, S. Jegelka, and J. A. Bilmes, “Curvature and optimal
algorithms for learning and minimizing submodular functions,” in
Advances in Neural Inform. Processing Systems, 2013, pp. 2742–2750.
[35] M. Sviridenko, J. Vondrák, and J. Ward, “Optimal approximation for
submodular and supermodular optimization with bounded curvature,”
in Proceedings of the 26th Annual ACM-SIAM Symposium on Discrete
Algorithms, 2015, pp. 1134–1148.
[36] H. Lin and J. Bilmes, “A class of submodular functions for document
summarization,” in Proceedings of the 49th Annual Meeting of the
Association for Computational Linguistics: Human Language Technologies – Volume 1, 2011, pp. 510–520.
[37] S. Jegelka and J. Bilmes, “Submodularity beyond submodular energies: Coupling edges in graph cuts,” in IEEE Conference on Computer
Vision and Pattern Recognition, 2011, pp. 1897–1904.
[38] M. Conforti and G. Cornuéjols, “Submodular set functions, matroids
and the greedy algorithm: Tight worst-case bounds and some generalizations of the rado-edmonds theorem,” Discrete Applied Mathematics,
vol. 7, no. 3, pp. 251 – 274, 1984.
[39] J. Vondrák, “Submodularity and curvature: The optimal algorithm,”
RIMS Kokyuroku Bessatsu, vol. 23, pp. 253–266, 2010.
[40] V. Tzoumas, M. Rahimian, G. Pappas, and A. Jadbabaie, “Minimal actuator placement with bounds on control effort,” arXiv
preprint:1409.3289, 2014.
| 3cs.SY
|
1
Layered Space-Time Index Coding
Yu-Chih Huang, Yi Hong, Emanuele Viterbo, and Lakshmi Natarajan
arXiv:1709.04379v1 [cs.IT] 13 Sep 2017
Abstract
Multicasting K independent messages via multiple-input multiple-output (MIMO) channels to multiple users
where each user already has a subset of messages as side information is studied. A general framework of constructing
layered space-time index coding (LSTIC) from a large class of space-time block codes (STBC), including perfect
STBC, is proposed. We analyze the proposed LSTIC and show that it provides minimum determinant gains that
are exponential with the amount of information contained in the side information for any possible side information.
When constructed over a perfect STBC, the proposed LSTIC is itself a perfect STBC and hence many desired
properties are preserved. To illustrate, we construct LSTIC over the following well-known STBCs: Golden code;
3 × 3, 4 × 4, 6 × 6 perfect STBCs; and Alamouti code. Simulation results show that the obtained side information
gain can be well predicted by our analysis.
Index Terms
Index coding, broadcast channels, side information, space-time block codes, MIMO channel.
I. I NTRODUCTION
The index coding problem [1], [2] studies the problem of optimally broadcasting independent messages
via noiseless links to multiple receivers where each receiver demands a subset of messages and already
has another subset of messages as side information. The side information at a receiver is described by
an index set and could be obtained from various means depending on the application. For example, in
retransmissions in broadcast channel [1], the side information is decoded from the previous received
signals; in the coded caching technique [3], [4], the side information is prefetched into users’ local cache
memories during off-peak hours; and in wireless relay networks [5]–[7], the side information is the users’
own data and/or is decoded/overheard from the previous sessions.
At the physical layer, the index coding problem can in fact be modeled as the noisy broadcast channel
with receiver side information. This problem has recently been investigated from two different perspectives
and most of the prior works can be categorized accordingly into two groups. The first one including [5], [6],
[8]–[11] focuses on characterizing the capacity region of the AWGN broadcast channel with message side
information. The capacity region of the two-user broadcast channel with receiver message side information
has been completely characterized [5], [8]. However, since the number of possible index sets increases
exponentially with the number of users in the network, the problem quickly becomes intractable as the
number of users increases. As a result, the capacity region for the three-user case has not been fully
characterized for some index sets [9]–[11] and our knowledge about more than three users is limited to
some special cases [12], [13].
The second category including [14]–[18] considers designing codes/constellations that possess some
desired properties in the finite dimension regime. The main objective is to design codes such that the
probability of error will decrease by an amount that is proportional to the amount of information contained
in the side information. In [14], Mahesh and Rajan consider the AWGN broadcast channel and assume
that the transmitter knows all the index sets, i.e., the side information configuration is available at the
transmitter. The scheme proposed in [14] consists of a linear index coding followed by an algorithm that
Y.-C. Huang is with the Department of Communication Engineering, National Taipei University, 237 Sanxia District, New Taipei City,
Taiwan (email: [email protected]).
Y. Hong and E. Viterbo are with the Department of Electrical and Computer System Engineering, Monash University, VIC 3800, Australia
(e-mail: {yi.hong, emanuele.viterbo}@monash.edu).
L. Natarajan is with the Department of Electrical Engineering, Indian Institute of Technology Hyderabad, Sangareddy 502285, India
(e-mail: [email protected]).
2
maps coded bits onto a phase shift keying (PSK) modulation. It is shown in [14] that this scheme indeed
can provide a reduction in probability of error proportional to the amount of side information.
Another line of research within this category ([15]–[18]), which seamlessly scales to any number of
users, considers the scenario where the transmitter is oblivious of the index sets. This enables to handle
large numbers of users, when the index sets to feedback to the transmitter require excessive resources
and/or the complexity of designing the specific index code becomes excessive. The objective then becomes
designing coding schemes that are fair to every possible index set. As a starting point, only the multicasting
case is considered in [15]–[18] where all the receivers demand all the messages.
In [15] and [16], Natarajan et al. study code design for the AWGN broadcast channel where minimum
distance is one of the most crucial parameters to be maximized. They first propose a coding scheme
in [15] by partitioning multi-dimensional pulse amplitude modulation (PAM) into subsets via computer
search for up to five messages with the message size up to 26 . Exploiting the algebraic structure induced
by the Chinese remainder theorem (CRT), a novel coding scheme, lattice index coding, is then proposed
in [16] to accommodate any number of messages with message sizes relatively prime to each other. Both
the schemes in [15] and [16] are shown to provide gains in minimum distance exponential with the rate
of the side information, for any index set.
In [17], Huang considers the same multicasting problem with message side information, where each
link experiences a Rayleigh fading channel on top of the AWGN noise. It is well-known that in contrast
to the AWGN channel, maximizing minimum distance alone is far from enough for the Rayleigh fading
channel and the minimum product distance dominates the performance [19]–[21]. The lattice index coding
scheme proposed in [17] generalizes the idea of [16] from some famous principal ideal domains to any
ring of algebraic integers. It is shown that codes thus constructed over rings of algebraic integers of
totally real number fields provide gains in minimum product distance that is exponential with the rate of
the side information for any index set. The multicasting problem with message side information is then
considered in [18] under the 2×2 MIMO setting where the transmitter and the receivers are equipped with
two antennas. For such a MIMO setting, the minimum determinant of the code serves as one of the most
important parameters to be maximized [21], [22] and algebraic space-time block codes (STBC) constructed
from cyclic division algebras [23]–[26] are a class of codes that possess many desired properties. Since
CRT does not hold for non-commutative rings such as cyclic division algebras, the trick used in [16]
and [17] does not work here in general. In [18], the problem is circumvented by using the bijective
mapping between the Golden algebra and a commutative ring found in [27] together with some special
ideals whose group structure is preserved by the mapping. As a result, we successfully construct Goldencoded index coding from Golden code, a subclass of perfect codes for the case with two transmitter and
receiver antennas, and show that the minimum determinant increases exponentially with the rate of the
side information for any index set.
A. Contributions
In this work, we consider the problem of multicasting over MIMO channel with message side information. We propose layered space-time index coding (LSTIC), a general framework of constructing
lattice space-time index codes from algebraic STBC. We exploit the algebraic structure of these codes to
encode the different messages into subcodes, which preserve all the good properties of the STBC, such
as non-vanishing determinant and power efficiency.
Any receiver that has some of the messages as side information will be decoding a subcode that has an
improved performance in terms of error probability. We provide a lower bound on the side information
gain for any side information configuration. The side information gain essentially measures the SNR
reduction (normalized by the rate of the side information) to achieve the same error probability, given
the side information. This lower bound implies an exponential increase of minimum determinant and is
universal in the sense that it holds for any possible index set.
We apply the proposed framework with the Golden code, 3×3, 4×4, 6×6 perfect STBCs, and Alamouti
code, and show that our analysis well predicts the actual side information gains obtained from simulations.
3
For each of the above codes, we also provide a table of the corresponding prime ideal factorizations for
p < 100, over which the LSTIC can be constructed according to the desired message sizes.
We note that the technique used in [18] requires the code to be constructed over prime ideals whose
group structure are preserved by the bijective mapping of [27] and thereby limits the possible message
sizes. In contrast to the Golden-coded index coding in [18] working only for Golden code with the message
sizes confined to some particular prime powers, the proposed LSTIC is quite general in the following
senses: 1) it works for a large class of STBC constructed from cyclic division algebras; and 2) it has less
restriction on the message size. We would like to emphasize here that when specialized to the Golden
code, the proposed LSTIC is not a special case of the Golden-coded index coding in [18] and vice versa,
due to the different message sizes.
B. Notations
Throughout the paper,
√ the following notations are used. Matrices are written in capital boldface, for
example X. Let i , −1 and ω , ei2π/3 be the primitive cube root of unity. We denote by Z, Z[i] ,
{a + bi|a, b ∈ Z}, and Z[ω] , {a + bω|a, b ∈ Z} the ring of integers, the ring of Gaussian integers, and the
ring of Eisenstein integers, respectively. Also, we denote by Q, R, and C the field of rational numbers,
the field of real numbers, and the field of complex numbers, respectively.
C. Organization
The rest of the paper is organized as follows. In Section II, we state the problem of physical-layer index
coding over MIMO channel and formally define the side information gain, the performance measure that
we will use throughout the paper. Background knowledge on algebra, algebraic number theory, and cyclic
division algebra is given in Section III. The LSTIC is then proposed and analyzed in Section IV. In
Sections V-IX, we construct LSTIC over Golden code, 3 × 3 perfect STBC, 4 × 4 perfect STBC, 6 × 6
perfect STBC, and Alamouti code. We then conclude the paper in Section X.
II. P ROBLEM S TATEMENT
Consider the network shown in Fig. 1 where there is a base station broadcasting messages to L users.
The base station is equipped with nt antennas and each user is equipped with nr antennas. There are K
independent messages {w1 , . . . , wK } collocated at the base station and each wk is uniformly distributed
over {1, . . . , Mk }. Each user demands all the K messages and already has a subset of the messages as
side information. For user ℓ, we denote by Sℓ ⊆ {1, . . . , K} the index set and the side information at the
user is wSℓ , {ws |s ∈ Sℓ }. The base station encodes the messages across space (nt antennas) and time
(T symbol durations) into an nt × T codeword matrix X where each entry xjt ∈ C and the codeword
is subject to the power constraint E[kXk2 ] = nt T . In a space-time code, each codeword X is used to
transmit r information-bearing real symbols. We denote by Rk = log2 (Mk )/r the rate of the message wk
measured in bits per real symbol. The signal model between the base station and the user ℓ is given by
Yℓ = Hℓ X + Zℓ ,
where Yℓ is of size nr × T , Hℓ is a random nr × nt matrix with each element i.i.d. ∼ CN (0, 1), and Zℓ is
a random nr × T matrix with each element i.i.d. ∼ CN (0, σl2 ). Each user is assumed to know the channel
matrix Hℓ associated with its received signal, i.e., channel state information at the receiver is assumed.
The signal-to-noise power ratio (SNR) is defined as SNRl , σn2t .
l
Let φ be a bijective encoding function that maps the messages (w1 , . . . , wK ) to the transmitted signal
X. The codebook C is the collection of codewords given by
C = {X = φ(w1 , . . . , wK )|wk ∈ {1, . . . , Mk }, ∀k} .
4
+
H1
X
Hℓ
Y1
Z1
+
Yℓ
(1)
User ℓ
(ℓ)
(ℓ)
(L)
(L)
{ŵ1 , . . . , ŵK }
wSℓ
..
.
HL
{w1 , . . . , wK }
(1)
{ŵ1 , . . . , ŵK }
wS1
..
.
Zℓ
+
User 1
YL
ZL
User L
{ŵ1 , . . . , ŵK }
wSL
Fig. 1. Multicasting {w1 , . . . , wK } over MIMO channel to L receivers where each receiver ℓ ∈ {1, . . . , L} has a subset of messages wSℓ
as side information.
(ℓ)
(ℓ)
Based on the received signal Yℓ and side information wSℓ , the user ℓ forms {ŵ1 , . . . , ŵK } (or equivalently
X̂(ℓ) ) an estimate of {w1 , . . . , wK } (or equivalently X). The probability of error is defined as
(ℓ)
(ℓ)
pe(ℓ) , Pr{{w1, . . . , wK } =
6 {ŵ1 , . . . , ŵK }}
= Pr{X 6= X̂(ℓ) },
where the second expression is often called the codeword error rate (CER). We emphasize here that the
index set Sℓ can be any subset of {1, . . . , K} and is oblivious to the base station. This makes the problem
of every ℓ identical for the base station. We therefore focus on a generic user and drop the subscript
(superscript in some cases) ℓ. The dummy variable ℓ is then released for later use.
Following [22], we define A , (X − X′)(X − X′ )† for any pair of codeword matrices X, X′ ∈ C. Let r
be the rank of A. For the generic user with S = ∅, in the high SNR regime, the probability of mistaking
X′ for X can be bounded as
−rnr
SNR∆1/r
′
,
Pr(X → X ) ≤
4nt
Q
where ∆ = rm=1 λm with λ1 , . . . , λm being the non-zero eigenvalues of A. Moreover, for full rank codes,
i.e., r = nt and
nt
Y
λm = det(A) 6= 0,
∆=
m=1
we define the minimum determinant of C as follows,
δ(C) , min′ det(A).
X6=X ∈C
If C is carved from a lattice Λ [28], we have
δ(C) = min det(X)2 .
X6=0∈Λ
(1)
To estimate the probability of error more accurately, let us define NX the number of codewords X′ ∈ C
resulting in det(A) = δ(C) and define
1 X
NC ,
NX ,
(2)
|C| X∈C
5
the average of NX over X ∈ C. For a STBC carved from a lattice, we can now approximate the probability
of error as
!
[
1 X
X → X′
pe =
P
|C| X∈C
X′ 6=X
−nt nr
(a) 1 X
SNRδ(C)1/nt
≈
NX
|C| X∈C
4nt
−nt nr
SNRδ(C)1/nt
= NC
,
(3)
4nt
where the approximation in (a) will become quite accurate in the high SNR regime.
Having had the approximation in (3), we can now follow [18] to derive the side information gain as
follows. We first note that with the knowledge of side information ws = vs , ∀s ∈ S, the generic user
can throw away all the codewords that do not correspond to this side information. The codebook then
becomes
dk = vk ,
k ∈ S;
,
CS , X = φ(d1 , . . . , dK )
dk ∈ {1, . . . , Mk }, otherwise.
a subcode of C. Since CS ⊆ C, the minimum determinant of CS , δ(CS ), will be no less than δ(C). Let
us now see how gains in minimum determinant can be translated into SNR gains. Following [18], we let
SNR and SNRS be the SNR required for the codebooks C and CS , respectively, to achieve a same error
probability pe . Then (3) says that
−nt nr
−nt nr
SNRS δ(CS )1/nt
SNRδ(C)1/nt
≈ NCS
NC
4nt
4nt
(⇔) 10 log10 (SNR) − 10 log10 (SNRS ) ≈
1
NC
δ(CS )
1
,
(4)
+ 10 log10
10 log10
nt nr
NCS
nt
δ(C)
which represents the SNR gain in dB provided by the side information wS . As mentioned in [18] and
many other work in the space-time code literature, it is in general not an easy task to keep tracking both
NCS and δ(CS ) for lattice codes; we thereby focus solely on δ(CS ) as our design guideline and define the
SNR gain as 10 log10 (δ(CS )/δ(C))nt dB. To get a fair comparison for every possible side information, we
then normalize this side information gain by the rate of the side information and define the normalized
side information gain as
δ(CS )
10 log10 δ(C)
Γ(C, S) ,
,
(5)
nt RS
P
where the rate of the side information is defined as RS , s∈S Rs and is measured in bits per real symbol,
which makes the normalized side information gain having the unit “dB/bits per real symbol”. The side
information gain essentially serves as an approximation of the SNR gain provided by side information
wS , normalized by the rate of wS . We note that involving the first term of (4) into the definition of side
information gain results in a better approximation. Hence, although we use (5) as the design guideline
throughout the paper, (4) is also used to confirm the simulation results.
III. BACKGROUND
In this section, we first review basic knowledge including algebra and algebraic number theory. We
then focus on cyclic division algebra and its connection to lattice STBC. To make the paper concise, we
only review the minimum required background for understanding the discussion that follows. For details,
please refer, for example, to [20], [22], [29]–[31].
6
A. Algebra
Let R be a commutative ring equipped with two operations addition + and multiplication ·. An ideal I
of R is an additive subgroup of R with respect to + that absorbs the multiplication of R, i.e., it satisfies
a · r ∈ I for a ∈ I and r ∈ R. An ideal I is a principal ideal if it can be generated by a singleton, i.e.,
I = aR for some a ∈ R. A proper ideal I is an ideal that is at the same time, a proper subset of R, i.e.,
∅=
6 I ⊂ R.
For an ideal I and any two elements a, b ∈ R, a is congruent to b modulo I if and only if a − b ∈ I,
which defines an equivalence relation. The quotient ring R/I of R by I is the collection of equivalence
classes with addition and multiplication defined as the original ones followed by modulo I operation as
follows,
(a + I) + (b + I) = (a + b) + I, and
(a + I) · (b + I) = (a · b) + I,
respectively. A prime ideal p of R is a proper ideal satisfying that whenever ab ∈ p for a, b ∈ R, then
either a ∈ p or b ∈ p. We now define the sum and product of ideals. Let I1 and I2 be two ideals of R,
the sum of two ideals is itself an ideal and is defined as
I1 + I2 , {a + b : a ∈ I1 , b ∈ I2 } .
The product of I1 and I2 is again an ideal and is defined as
( n
)
X
I1 I2 ,
ai bi : ai ∈ I1 , bi ∈ I2 , n ∈ N .
i=1
In general, I1 I2 ⊆ I1 ∩ I2 . Two ideals are said to be relatively prime if R = I1 + I2 . When I1 and I2 are
relatively prime, we further have I1 I2 = I1 ∩ I2 . We say I1 divides I2 , denoted as I1 |I2 , if I2 = I1 I3
for some ideal I3 and consequently I2 ⊆ I1 .
Consider two commutative rings R1 and R2 with two operations (+, ·) and (⊕, ⊙), respectively. A
ring homomorphism between R1 and R2 is a function σ : R1 → R2 such that
σ(a + b) = σ(a) ⊕ σ(b), ∀a, b ∈ R1 ,
σ(a · b) = σ(a) ⊙ σ(b), ∀a, b ∈ R1 .
In other words, a ring homomorphism preserves the ring structure. A homomorphism is a monomorphism
if it is injective and is an isomorphism if it is bijective. Moreover, an isomorphism σ : R1 → R1 is called
automorphism.
We now review two classical results in ring theory whose proofs can be found in a standard textbook.
Lemma 1 (Second isomorphism theorem [29, Theorem 2.12]). Let R be a commutative ring, I1 and I2
be two ideals. We have the following isomorphism,
I1 /(I1 ∩ I2 ) ∼
= (I1 + I2 )/I2 .
In fact, the second isomorphism theorem holds for the more general case where I1 is only a subring
and not necessarily an ideal.
Lemma 2 (Chinese remainder theorem [29, Corollary 2.27]). Let I1 , . . . , In be ideals of a commutative
ring R. Moreover, I1 , . . . , In are relatively prime. We have
R/Πni=1 Ii ∼
= R/I1 × . . . × R/In .
where × stands for Cartesian product and the operations of the right hand side are defined componentwise.
We provide a quick example for what have been reviewed above.
7
Example 3. Consider Z the set of all integers with ordinary addition + and multiplication ·. Clearly, it
forms a commutative ring. 2Z is the principal ideal of Z consisting of all the even integers. Moreover,
it is a prime ideal. The quotient Z/2Z = Z2 forms a ring with addition + mod 2Z and multiplication ·
mod 2Z. Also, for 3Z another principal prime ideal of Z, we have the quotient ring Z/3Z = Z3 . Since
2 · (−1) + 3 · 1 = 1, 2Z + 3Z = Z and thus 2Z and 3Z are relatively prime. One can easily verify
that 2Z ∩ 3Z is precisely 6Z. Now, the CRT guarantees the existence of a ring isomorphism between
Z6 = Z/6Z and Z2 × Z3 . One can verify that M(v1 , v2 ) = 3v1 − 2v2 mod 6Z where v1 ∈ Z2 and
v2 ∈ Z3 is a ring isomorphism.
B. Algebraic Numbers and Algebraic Integers
An algebraic number is a complex number that is a root of some polynomial with coefficients in Z.
Let L be a field and K ⊂ L be a subfield; L is said to be a field extension of K, which is usually denoted
as L/K. L can be viewed as a vector space over K. The degree of L over K, denoted by [L : K], is
defined as the dimension of the vector space L over K. A number field is a field extension of Q with
finite degree, i.e., a finite extension K/Q. Every number field K can be generated from Q by adjoining
an algebraic number θ, i.e., K = Q(θ). An algebraic integer is a complex number that is a root of some
polynomial with the leading coefficient 1 and other coefficients in Z. For a number field K, we denote
by OK the ring of integers of K which comprises all the algebraic integers in K.
Let L/K be a field extension of K with degree [L : K] = n. Throughout the paper, we will further
assume that L/K is a Galois extension. There are exactly n distinct K-automorphisms σi : L → L for
i ∈ {1, . . . , n}, i.e., automorphisms that fix K. Such automorphisms are called (relative) embeddings. It
can be shown that Gal(L/K) , {σ1 , . . . , σn } form a group under function composition, which is called
the Galois group. For α ∈ L, we define the norm of α as
NL/K (α) =
n
Y
σi (α),
i=1
where σ2 (α), . . . , σn (α) are called the conjugates of σ1 (α) = α. Let {α1 , . . . , αn } be an integral basis for
OL , such that any element in OL can be uniquely written as a linear combination of the basis element
with coefficients Z. The discriminant of a number field L is defined as
2
σ1 (α1 ) σ1 (α2 ) . . . σ1 (αn )
σ2 (α1 ) σ2 (α2 ) . . . σ2 (αn )
.
dL , det
..
..
..
..
.
.
.
.
σn (α1 ) σn (α2 ) . . . σn (αn )
Let I be an ideal in OL , then I can be generated by at most two elements, i.e., I = αOL + βOL for
some α, β ∈ OL . The norm of I is defined as
N(I) , |OL /I|.
Moreover, if I = αOL is principal, N(I) = |NL/Q (α)|.
Let p be a prime ideal in OL , the ring of integers of L with [L : Q] = n. We say that p lies above a
prime p if p ∩ Z = pZ. For a prime p, the principal ideal pOL can be factorized into 1 ≤ g ≤ n prime
ideals as
pOL = pe11 · . . . · pegg ,
where ei , i ∈ {1, . . . , g}, is the ramification index of pi . Also, for each pi , we havePN(pi ) = pfi and
g
OL /pi ∼
= Fpfi where 1 ≤ fi ≤ n is the inertial degree. Overall, it can be shown that i=1 ei fi = n. For
a Galois extension, we have e1 = e2 = . . . = eg = e and f1 = f2 = . . . = fg = f , which implies that
ef g = n. A prime p is ramified in OL if not all ei = 1 in the factorization of pOL . Ramified primes in
OL are precisely those p that divides the discriminant dL .
8
Example 4. Consider Q(i) the field extension obtained from Q by adjoining i. Every element in Q(i) has
the form a + bi where a, b ∈ Q; thus, it is a number field with degree 2. The two Q-automorphisms are
σ1 (a + bi) → a + bi and σ2 (a + bi) → a − bi. The Galois group is cyclic and can be generated by σ2 .
Since σ1 is the identity mapping and σ2 sends an element to its complex conjugate, the norm defined in
this number field coincides with the Euclidean norm. The ring of integers is Z[i], the Gaussian integers,
having integral basis {1, i}. The discriminant is computed as follows,
2
1 i
dQ(i) = det
= −4.
1 −i
Since 2|dQ(i), 2Z[i] = p2 ramifies where p = (1 + i)Z[i]. This is the only ramified prime in Q(i). Also,
5Z[i] = p1 p2 splits into two prime ideals p1 = (1 + 2i)Z[i] and p2 = (1 − 2i)Z[i] with e = 1 and f = 1.
Another example is that 3Z[i] is itself a prime ideal with e = 1 and f = 2. In each case, we have ef g = 2.
C. Cyclic division algebra and lattice space-time codes
An algebra A over a field L is a set satisfying: i) it is a vector space over L; ii) it is a ring with respect
to addition and multiplication by elements of A; and iii) (αa)b = a(αb) = α(ab) for any α ∈ L and
a, b ∈ A. Let L/K be a field extension of K of degree n whose Galois group is a cyclic group generated
by σ. One can construct a cyclic algebra A = (L/K, σ, γ) as
A = (L/K, σ, γ) = x0 + x1 e + . . . + xn−1 en−1 |x0 , . . . , xn−1 ∈ L ,
where en = γ ∈ K and λe = eσ(λ) for λ ∈ L. A is said to be a division algebra if every non-zero
element of A is invertible. A cyclic division algebra is a cyclic algebra that is at the same time a division
algebra. In the space-time coding literature (see [22] and reference therein), a cyclic division algebra is
usually constructed from a cyclic algebra A = (L/K, σ, γ) with carefully chosen γ such that none of
γ, γ 2 , . . . , γ n−1 are norms of some element of L.
Consider nt = nr = T = n, an n × n STBC carved from A corresponds to a finite subset of
ĀI = x0 + x1 e + . . . + xn−1 en−1 |x0 , x1 , . . . , xn−1 ∈ I ,
(6)
where I is an ideal in OL . More specifically, an n × n STBC thus constructed can be obtained by putting
ĀI into the matrix form given by
x0
x1
...
xn−1
γσ(xn−1 )
σ(x0 )
σ(xn−2 )
(7)
CI =
..
..
..
x0 , . . . , xn−1 ∈ I .
.
.
.
γσ n−1 (x1 ) γσ n−1 (x2 ) . . . σ n−1 (x0 )
A layer ℓ ∈ {0, . . . , n − 1} of the codeword in CI is the collection of the entries in positions (m, (ℓ + m)
mod (n)) for m ∈ {1, . . . , n}. We note that each layer ℓ ∈ {0, . . . , n − 1} corresponds to the same xℓ ∈ I.
Here, we use the subscript I in ĀI and CI to emphasize that the elements xℓ for all ℓ are restricted to the
ideal I. For transmission with finite input power constraint, one carves a subset from (a possibly shifted
and scaled version of) CI to form the codebook. From this point onward, we restrict the discussion to
K = Q(i) or Q(ω), which corresponds to the case where each xℓ is a linear combination of n QAM or
HEX constellation symbols. One observes that each codeword X ∈ CI conveys n symbols of L, where
each symbol xℓ is a linear combination of n QAM or HEX symbols. Therefore, the STBC thus constructed
is full-rate. i.e., it uses an n × n matrix to transmit n2 symbols. Another consequence of having each xℓ
being a linear combination of n QAM or HEX symbols is that the code may not be energy-efficient as
compared to sending QAM or HEX symbols directly. This drawback can often be overcome by choosing
a suitable ideal I such that CI becomes a scaled and rotated version of Z[i]n or Z[ω]n .
The determinant of the codeword X ∈ CI corresponding to x ∈ A is called the reduced norm of x.
What is important about having the structure of cyclic division algebra is that when γ ∈ OK not the norm
9
of an element in L, it guarantees that the code is fully diverse and has non-vanishing determinant (NVD).
This is evident from [26, Corollary1 and Corollary 2], which states that the reduced norm of x ∈ ĀOL
belongs to OK and thus δ(COL ) = 1. Now, since I ⊆ OL , one has that δ(CI ) ≥ 1. In fact, one can obtain
better bounds on δ(CI ) as follows.
Lemma 5 ([26, Corollary 3 and Corollary 4]). Let CI be a STBC built over the cyclic division algebra
A = (L/K, σ, γ) as in (7), where γ ∈ OK not the norm of an element in L. Then,
N(I) ≤ δ(CI ) ≤ min NL/Q (x).
x∈I
We end this section by providing the definition of a perfect STBC as follows.
Definition 6. A n × n STBC is called a perfect STBC if i) it is full-rate; ii) it is fully diverse and has
NVD property; iii) the energy used to send the coded symbol on each layer is equal to that for sending
the uncoded symbol themselves; and iv) all the coded symbols have the same average energy.
IV. P ROPOSED L AYERED S PACE -T IME I NDEX C ODING
In this section, we propose the LSTIC and show that for any index set, it can provide SNR gain that
is proportional to the information contained in the side information. In the proposed scheme, instead of
directly tackling ĀOL as done in [18], we recognize the layered structure of STBC reviewed in Section III-C
and perform partition layer by layer. More specifically, we split each message wk , k ∈ {1, . . . , K}, into n
sub-messages, namely wk,ℓ for ℓ ∈ {0, . . . , n − 1}, and encode w1,ℓ , . . . , wK,ℓ into xℓ the layer ℓ. The main
advantage of this approach is that now each layer’s signal is in OL and thereby one can apply CRT for
partitioning. In what follows, we focus solely on cyclic division algebras with γ ∈ OK , such that none of
γ, γ 2 , . . . , γ n−1 are norms of element in L. We split the discussion into two parts depending on whether
I is principal or not. The first case includes constructions from 2 × 2, 3 × 3, and 4 × 4 perfect STBC
while the second case encompasses constructions from the 6 × 6 perfect STBC. The similar approach can
also be applied to Alamouti code for constructing Layered Alamouti-coded index coding, which will be
discussed in Section IX.
Remark 7. We emphasize that the approach that we propose in the following in fact applies to any cyclic
division algebra with the non-norm element γ with K = Q(i) or Q(ω). For instance, the STBC design
with non-norm element γ ∈ K in [32] can also be used as the base STBC of our LSTIC. The main reason
that we particularly focus on γ ∈ OK is so that we can rely on Lemma 5 to prove a lower bound on the
side information gain. Apart from this, the proposed method does not require γ ∈ OK .
A. LSTIC with principal I
Without loss of generality, we assume that I is generated by some α ∈ OL , i.e., I = αOL . Then, (6)
becomes
ĀI = x0 + x1 e + . . . + xn−1 en−1 |x0 , x1 , . . . , xn−1 ∈ αOL ,
= αx0 + αx1 e + . . . + αxn−1 en−1 |x0 , x1 , . . . , xn−1 ∈ OL ,
and (7) can be rewritten as
x0
x1
...
xn−1
γσ(xn−1 )
σ(x0 )
σ(xn−2 )
D(α) ·
..
..
..
x0 , . . . , xn−1 ∈ OL ,
.
.
.
γσ n−1 (x1 ) γσ n−1 (x2 ) . . . σ n−1 (x0 )
(8)
10
where
α
0
...
0
0 σ(α) . . .
0
D(α) ,
.
.
..
.
..
..
..
.
n−1
0
0
. . . σ (α)
We emphasize here that, as mentioned in Section III-C, the codebook that we actually use should be
a scaled version of the above codebook to satisfy the power constraint. However, in our analysis, what
we really care is the ratio between the minimum determinants of the codebooks with and without side
information, where the scaling does not make any difference. Therefore, throughout the paper, when
analyzing the proposed scheme, we ignore the scaling factor for the sake of brevity. On the other hand, in
our simulations, we do take the scaling into account and normalize the codebook to make the parameters
reflect the actual SNR.
We can now use the technique in [17] to partition OL . Let q1 , . . . , qK be K ideals in OL that are
relatively prime and have N(qk ) = qk , k ∈ {1, . . . , K}. Note that qk s are not necessarily prime ideals
and qk s are not necessarily prime. We have q1 ∩ . . . ∩ qK = q1 · . . . · qK , q. From CRT, we have
OL /q ∼
= Bq1 × . . . × BqK ,
= OL /q1 × . . . × OL /qK ∼
where Bqk = OL /qk is a commutative ring1 with size qk . Let M be an isomorphism that maps Bq1 ×
. . . × BqK to a complete set of coset leaders of OL /q having minimum energy.
Now, for k ∈ {1, . . . , K}, let wk ∈ Bnqk which can be represented as wk = (wk,0, . . . , wk,n−1) where
each wk,ℓ ∈ Bqk . The encoder collects w1,ℓ , . . . , wK,ℓ to form the signal of the layer ℓ ∈ {0, . . . , n − 1} as
xℓ = M(w1,ℓ , . . . , wK,ℓ ) ∈ OL /q,
ℓ ∈ {0, . . . , n − 1}.
The overall codebook corresponds to
Ā = αx0 + αx1 e + . . . + αxn−1 en−1 |x0 , . . . , xn−1 ∈ OL /q ,
a subset of ĀI and has the matrix form as that in (8) with x0 , . . . , xn−1 ∈ OL /q.
For the proposed LSTIC within this class, we can show the following theorem.
Theorem 8. For any S ⊂ {1, . . . , K}, the proposed LSTIC with principal I provides a side information
gain at least 6 dB/bits per real symbol, i.e., Γ(C, S) ≥ 6 dB/bits per real symbol. Moreover, if all qk ,
k ∈ {1, . . . , K}, are principal, then Γ(C, S) = 6 dB/bits per real symbol.
Proof. We first note that in the proposed scheme, each message is spread onto n layers of signals, which
are then mapped to a n × n complex codeword matrix. i.e., 2n2 real symbols. Therefore, the rate of the
message wk is given by
1
1
log2 (qkn ) =
log2 (qk ), bits per real symbol.
(9)
2
2n
2n
Consider a generic receiver with index set S, let the messages be ws = vs for s ∈ S. This means that
ws,ℓ = vs,ℓ for all ℓ ∈ {0, . . . , n − 1} are known at the receiver. Let us first take S = {s} for example.
The ℓth layer’s signal can then be rewritten as
Rk =
{s}
xℓ
= M(w1,ℓ , . . . , ws−1,ℓ , vs,ℓ , ws+1,ℓ , . . . , wK,ℓ )
(a)
{s}
= M(0, . . . , 0, vs , 0, . . . , 0) + M(w1,ℓ , . . . , ws−1,ℓ , 0, ws+1,ℓ , . . . , wK,ℓ ) + ζℓ
{s}
= ξℓ
{s}
+ x̃ℓ ,
1
Depending on the ideal qk , this ring could be a finite field, a product of finite fields, a product of finite rings and finite fields, or others.
But it is always commutative since a quotient ring of a commutative ring is always commutative. Throughout the paper, we do not use the
ring property of the messages and therefore, we do not emphasize which type of ring it is.
11
{s}
{s}
{s}
{s}
where ζℓ ∈ q, x̃ℓ , M(w1,ℓ , . . . , ws−1,ℓ , 0, ws+1,ℓ, . . . , wK,ℓ)+ζℓ , and ξℓ , M(0, . . . , 0, vs , 0, . . . , 0)
is known at the receiver. The equality (a) above holds because M is an isomorphism. From CRT, we
have
{s}
(xℓ
{s}
which implies that xℓ
{s}
− ξℓ )
mod qs = 0,
belongs to a shifted version of qs . For the general S, we can similarly show that
xSℓ = M(d1,ℓ , . . . , dK,ℓ) + M(u1,ℓ , . . . , uK,ℓ) + ζℓS
= ξℓS + x̃Sℓ ,
where ζℓS ∈ q, x̃Sℓ = M(u1,,ℓ, . . . , uK,ℓ) + ζℓS , and ξℓS , M(d1,ℓ , . . . , dK,ℓ) with
vk,ℓ , k ∈ S;
dk,ℓ =
0,
k ∈ Sc,
and
uk,ℓ =
0,
k ∈ S;
wk,ℓ , k ∈ S c .
Note that ξℓS is known at the receiver. We now have
xSℓ − ξℓS
mod qs = 0,
(10)
(11)
(12)
for all s ∈ S,
which shows that xSℓ belongs to a shifted version of ∩s∈S qs = Πs∈S qs . Therefore, after revealing wS , the
code CS corresponds to
S
S
α(ξ0 + . . . + ξn−1
en−1 ) + α(x̃S0 + . . . + x̃Sn−1 en−1 )|x̃S0 , . . . , x̃Sn−1 ∈ Πs∈S qs ,
Hence, thanks to that σ is a homomorphism, each codeword X ∈ CS has the matrix form given by
X = VS + X̃S ,
where
ξ0S
S
γσ(ξn−1
)
S
V = D(α) ·
..
.
γσ
and
n−1
(ξ1S )
x̃S0
γσ(x̃Sn−1 )
X̃S = D(α) ·
..
.
γσ
n−1
(x̃S1 )
ξ1S
σ(ξ0S )
γσ
n−1
(ξ2S )
x̃S1
σ(x̃S0 )
γσ
n−1
(x̃S2 )
...
..
S
ξn−1
S
)
σ(ξn−2
,
..
.
.
. . . σ n−1 (ξ0S )
...
..
x̃Sn−1
σ(x̃Sn−2 )
.
..
.
.
. . . σ n−1 (x̃S0 )
Note that the second part of X̃S is a codeword of the code
x0
x1
...
xn−1
γσ(xn−1 )
σ(x0 )
σ(xn−2 )
CΠs∈S qs =
..
..
..
x0 , . . . , xn−1 ∈ Πs∈S qs ,
.
.
.
γσ n−1 (x1 ) γσ n−1 (x2 ) . . . σ n−1 (x0 )
whose minimum determinant can be bounded by Lemma 5 as follows,
δ(CΠs∈S qs ) ≥ N(Πs∈S qs ).
(13)
12
The receiver can now subtract the known VS and compute the minimum determinant as
δ(CS ) = | det(D(α))|2δ(CΠs∈S qs )
= |NL/K (α)|2δ(CΠs∈S qs )
(a)
= N(α)δ(CΠs∈S qs ),
where (a) follows from the fact that K = Q(i) or Q(ω) is a quadratic extension. Plugging (13) into the
above equation results in
δ(CS ) ≥ N(α)N(Πs∈S qs )
= N(α)Πs∈S N(qs ) = N(α)Πs∈S qs ,
(14)
where the last equality follows from the fact that the ideal norm is multiplicative. Moreover, without
revealing any side information, the overall codebook would have
δ(C) = N(α)N(1) = N(α).
(15)
Combining (9), (14), and (15) results in
10 log10 (Πs∈S qs )
P
1
n 2n
s∈S log2 (qs )
P
s∈S 20 log10 (qs )
= P
= 6 dB/bits per real symbol.
s∈S log2 (qs )
Γ(C, S) ≥
To prove the second statement, we note that if the ideal Πs∈S qs is principal, then we can indeed find
an element in the ideal such that the inequality in (14) holds with equality. Hence, if q1 , . . . , qK are all
principal, Γ(C, S) = 6 dB for every S.
B. LSTIC with non-principal I
We now construct LSTIC from a STBC based on a cyclic division algebra A = (L/K, σ, γ) and a
non-principal ideal I in OL as described in (6). Let q1 , . . . , qK be K ideals in OL that are relatively
prime and have norm N(qk ) = qk , k ∈ {1, . . . , K}. We again let q1 · . . . · qK = q. We further assume that
each qk and I are relatively prime, which also implies that q and I are relatively prime. From the second
isomorphism theorem [29] and CRT, we have
(b)
(a)
I/Iq = I/I ∩ q ∼
= (I + q)/q
(c)
(d)
= OL /q ∼
= OL /q1 × . . . × OL /qK
∼
= Bq1 × . . . × BqK ,
where both (a) and (c) are due to the fact that q and I are relatively prime, (b) follows from the second
isomorphism theorem, and (d) follows from CRT. We use Bqk to denote the quotient ring that is isomorphic
to OL /qk which has size qk . Let M be an isomorphism that maps elements in Bq1 ×. . .×BqK to a complete
set of coset leaders of I/Iq.
For k ∈ {1, . . . , K}, we again enforce wk = (wk,0 , . . . , wk,n−1) ∈ Bnqk where each ℓ ∈ {0, . . . , n − 1}.
The sub-messages w1,ℓ , . . . , wK,ℓ are collected and encoded into xℓ the signal of the ℓ ∈ {0, . . . , n − 1}
layer as
xℓ = M(w1,ℓ , . . . , wK,ℓ ) ∈ I/Iq, ℓ ∈ {0, . . . , n − 1}.
The overall codebook now corresponds to {x0 + x1 e + . . . + xn−1 en−1 |x0 , . . . , xn−1 ∈ I/Iq} a subset of
ĀI and has the matrix form as that in (7) with x0 , . . . , xn−1 ∈ I/Iq.
For the proposed LSTIC within this class, we can show the following theorem.
13
Theorem 9. For any S ⊂ {1, . . . , K}, the side information gain achieved by the proposed LSTIC with
non-principal ideal I is lower bounded as
Γ(C, S) ≥ 6 + γI dB/bits per real symbol,
where
γI = 20 log10
N(I)
minx∈I NL/Q (x)
is negative and is only a function of I and is independent of S.
,
(16)
Proof. We again note that the rate of the message wk is given by
1
1
log2 (qk ), bits per real symbol.
(17)
Rk = 2 log2 (qkn ) =
2n
2n
We consider a generic receiver having index set S. Suppose the messages ws = vs for s ∈ S are known,
which means that ws,ℓ = vs,ℓ for all ℓ ∈ {0, . . . , n − 1} are known at the receiver. Similar to (10), we have
xSℓ = M(d1,ℓ , . . . , dK,ℓ) + M(u1,ℓ , . . . , uK,ℓ) + ζℓS
= ξℓS + x̃Sℓ ,
where dk,ℓ and uk,ℓ are defined in (11) and (12), respectively, and ζℓS ∈ Iq. Therefore, we have
xSℓ − ξℓS
mod Iqs = 0, for all s ∈ S,
which means that xSℓ belongs to a shifted version of
(a)
∩s∈S Iqs = ∩s∈S (I ∩ qs )
(b)
= I ∩ (∩s∈S qs ) = IΠs∈S qs ,
where (a) follows from that I and qs are relatively prime for each s and (b) is due to the fact that
q1 , . . . , qK are relatively prime.
After revealing wS , the code CS would correspond to
S
S
(ξ0 + . . . + ξn−1
en−1 ) + (x̃S0 + . . . + x̃Sn−1 en−1 )|x̃S0 , . . . , x̃Sn−1 ∈ IΠs∈S qs ,
Therefore, each codeword X ∈ CS has the matrix form given by
X = VS + X̃S ,
where
and
S
ξ0S
ξ1S
...
ξn−1
S
S
γσ(ξn−1
)
σ(ξ0S )
σ(ξn−2
)
S
,
V =
..
..
..
.
.
.
n−1 S
n−1 S
n−1 S
γσ (ξ1 ) γσ (ξ2 ) . . . σ (ξ0 )
x̃S0
x̃S1
...
x̃Sn−1
γσ(x̃Sn−1 )
σ(x̃S0 )
σ(x̃Sn−2 )
S
.
X̃ =
..
..
..
.
.
.
n−1 S
n−1 S
n−1 S
γσ (x̃1 ) γσ (x̃2 ) . . . σ (x̃0 )
We can again note that X̃S belongs to
x0
x1
...
xn−1
γσ(xn−1 )
σ(x0 )
σ(xn−2 )
CIΠs∈S qs =
..
..
..
x0 , . . . , xn−1 ∈ IΠs∈S qs ,
.
.
.
γσ n−1 (x1 ) γσ n−1 (x2 ) . . . σ n−1 (x0 )
14
whose minimum determinant can be bounded via Lemma 5 by
δ(CIΠs∈S qs ) ≥ N(IΠs∈S qs ).
One can now remove the contribution of VS from the received signal and bound the minimum determinant
as
δ(CS ) ≥ N(I)N(Πs∈S qs )
= N(I)Πs∈S N(qs ) = N(I)Πs∈S qs .
(18)
When no side information is available, we can again use Lemma 5 to bound the minimum determinant
as
N(I) ≤ δ(C) ≤ min NL/Q (x).
(19)
x∈I
Combining (17), (18), and (19) results in
N (I)
10 log10 N(Πs∈S qs ) minx∈I
NL/Q (x)
P
Γ(C, S) ≥
1
n 2n s∈S log2 (ps )
N (I)
P
20 log10 minx∈I NL/Q (x)
s∈S 20 log10 (qs )
P
= P
+
s∈S log2 (qs )
s∈S log2 (ps )
= 6 + γI,S dB/bits per real symbol.
Noting that γI,S ≤ 0 from (19) and γI,S ≥ γI completes the proof.
V. L AYERED G OLDEN -C ODED I NDEX C ODING
In this section, we propose layered Golden-coded index coding, a family of LSTIC constructed from
Golden code. To provide a concrete illustration of how the proposed scheme works, we will walk through
this example in detail. Before proceeding, we note that the layered Golden-coded index coding proposed
here is different, in essence, from the Golden-coded index coding in [18]. Here, we partition the code
layer by layer while in [18] we directly tackle the Golden algebra. We would like to emphasize that
neither of these two schemes subsumes the other as a special case; however, the approach taken in [18]
only works for some
√ particular primes.
Let
5) a quadratic extension of K = Q(i) and consider the non-trivial Q(i)-automorphism
√ L = Q(i,
√
σ : 5 → − 5. Also, let γ = i. The Golden code is built from the Golden algebra given by
n
√
√ o
G = (Q(i, 5)/Q(i), σ, i) = x0 + x1 e|x0 , x1 ∈ Q(i, 5) ,
√
where e2 = i and ze = eσ(z). The ring of integers of L is OL = Z[i][θ] where θ = 1+2 5 . Let I = αOL
be the principal ideal generated by α = 1 + iθ̄ where θ̄ , σ(θ). The Golden code [25] corresponds to
GI = {x0 + x1 e|x0 , x1 ∈ αOL } ,
which can be put into the matrix form
1
αx0
αx1
x0 , x1 ∈ Z[i][θ]
CI = √
iσ(αx1 ) σ(αx0 )
5
1
α(a + bθ)
α(c + dθ)
a, b, c, d ∈ Z[i] .
= √
5 iσ(α)(c + dθ̄) σ(α)(a + bθ̄)
15
The proposed layered Golden-coded index coding can be categorized into the class in Section IV-A.
Let q1 , q2 , . . . , qK be prime ideals in OL that are relatively prime. Let q1 . . . qK , q. Also, let |OK /qk | =
N(qk ) , qk for k ∈ {1, . . . , K} where qk s are not necessarily primes. From CRT, we have
OK /q ∼
= Bq1 × . . . × BqK ,
= OK /q1 × . . . × OK /qK ∼
where Bqk = OK /qk is a commutative ring with size qk . This guarantees the existence of M : Bq1 × . . . ×
BqK → OL /q an isomorphism that maps the messages to a complete set of coset leaders of OL /q with
minimum energy. In the proposed layered Golden-coded index coding scheme, we let wk ∈ B2qk and split
it into wk,0, wk,1 ∈ Bqk .
The sub-messages w1,ℓ , . . . , wK,ℓ, for ℓ ∈ {0, 1}, are encoded onto OL /q via M to form
xℓ = M(w1,ℓ , . . . , wK,ℓ) ∈ OL /q,
ℓ ∈ {0, 1}.
The overall codebook becomes a Golden code
1
αx0
αx1
C= √
x0 , x1 ∈ OL /q .
5 iσ(αx1 ) σ(αx0 )
(20)
(21)
From Theorem 8, we obtain the following corollary. Note that the proof of this corollary is almost
identical to that of Theorem 8. However, as mentioned earlier, in order to provide a complete illustration,
we still present the proof.
Corollary 10. For any S ⊂ {1, . . . , K}, the proposed layered Golden-coded index coding provides
Γ(C, S) = 6 dB/bits per real symbol.
Proof. The rate of the message wk is given by
Rk =
1
log2 (N(qk )2 ) bits per real symbol.
8
(22)
Suppose some messages wS , {wk = vk |k ∈ S} are known; this means that both wS,ℓ , {wk,ℓ =
vk,ℓ |k ∈ S} for ℓ = 0 and ℓ = 1 are known. Therefore, from Section IV-A, xℓ , ℓ ∈ {0, 1}, belongs to a
shifted version of Πk∈S qk . Thus, after revealing wS , the code CS becomes a shifted version of
1
αx0
αx1
√
x0 , x1 ∈ Πk∈S qk .
5 iσ(αx1 ) σ(αx0 )
For every codeword X̃S ∈ CS corresponding to x0 , x1 ∈ Πk∈S qk , the determinant is given by
1
αx0
αx1
S
det(X̃ ) = det
iσ(αx1 ) σ(αx0 )
5
(a) 1
αx0
αx1
= det
iσ(α)σ(x1 ) σ(α)σ(x0 ))
5
1
α
0
x0
x1
= det
det
0 σ(α)
iσ(x1 ) σ(x0 )
5
1
x0
x1
,
= Nrd (α) det
iσ(x1 ) σ(x0 )
5
where (a) is due to that σ is a homomorphism. Now, plugging |Nrd (α)|2 = 5 results in
2
1
x0
x1
δ(CS ) = det
iσ(x1 ) σ(x0 )
5
(a) 1
(b) 1
= N(Πk∈S qk ) = Πk∈S N(qk ),
5
5
(23)
16
TABLE I
P RIME FACTORIZATION OF p < 100 IN Z[i][θ]
p
2
3
5
7
11
13
17
19
23
29
31
37
41
43
47
53
59
61
67
71
73
79
83
89
97
WHERE
θ=
√
1+ 5
.
2
p
(1 + i)
(θ̄ − iθ), (θ̄ + iθ)
(1 + iθ̄), (1 − iθ̄)
((1 + θ) + i(1 + θ̄)), ((1 + θ) − i(1 + θ̄))
(3iθ − i), (3iθ̄ − i)
(2 + 3i), (2 − 3i)
(4 + i), (4 − i)
(4iθ − i), (4iθ̄ − i)
((3θ̄ − 1) + i(3θ − 1)), ((3θ̄ − 1) − i(3θ − 1))
(2i + θ), (2i + θ̄), (θ̄ − 2i), (θ − 2i)
(2 − 5θ̄), (2 − 5θ)
(6 + i), (6 − i)
(θ̄ + i(2θ − 1)), (θ̄ − i(2θ − 1)), (θ − i(2θ − 1)), (θ + i(2θ − 1))
((4 + θ) + i(4 + θ̄)), ((4 + θ̄) + i(4 + θ))
¯ − i(2 + 3θ))
((2 + 3θ) − i(2 + 3θ̄)), ((2 + 3θ)
(7 + 2i), (7 − 2i)
(7θ − 2), (7θ̄ − 2)
((2θ̄ − 1) + i(θ + 1)), ((2θ − 1) + i(θ̄ + 1)), ((2θ − 1) + i(θ + 1)), ((2θ̄ − 1) + i(θ̄ + 1))
((5θ̄ − 1) + i(5θ − 1)), ((5θ − 1) + i(5θ̄ − 1))
(8 + θ), (8 + θ̄)
(3 + 8i), (3 − 8i)
(8θ − 3), (8θ̄ − 3)
((4 + 3θ) + i(4 + 3θ̄)), ((4 + 3θ̄) + i(4 + 3θ))
(2θ̄ − i(θ + 1)), (2θ − i(θ̄ + 1)), (2θ̄ + i(θ + 1)), (2θ̄ + i(θ̄ + 1))
(9 + 4i), (9 − 4i)
f
2
2
1
2
2
2
2
2
2
1
2
2
1
2
2
2
2
1
2
2
2
2
2
1
2
where (a) follows from [26, Corollary 3] and the fact that OL = Z[i][θ] is a principal ideal domain and
(b) follows from the fact that algebraic norm is multiplicative. Now, combining what we have obtained
in (22) and (23) and the fact that δ(C) = 1/5 result in
Γ(C, S) =
10 log10 (Πk∈S N(qk ))
P
= 6 dB/bits per real symbol.
2 14 k∈S log2 N(qk )
A. Examples and Simulation Results
In Table I, we factorize each prime p < 100 into prime ideals in OL via Magma [33]. Any pair of
ideals in this table is relatively prime and thus qk can be chosen as product of some prime ideals that
have not been selected for some qk′ , k ′ 6= k. In Table I, we show ideals and their inertial degrees f . The
ramification index of each prime ideal lying above p 6= 2, 5 is 1 and is 2 for prime ideals lying above
2, 5. This can be seen by observing that
d L = 52 · 42 ,
which has prime factors 2 and 5. Moreover, since OL is a principal ideal domain, so every pOL can be
factorized into principal prime ideals.
Simulation results for the proposed layered Golden-coded index coding are provided in Fig. 2. In this
figure, three sets of simulations are performed. In the first one, we constructed the layered Golden-coded
index coding with two principal ideals generated by β1 = (θ̄ − iθ) and β2 = (θ̄ + iθ), respectively. From
Table I, we see that each of these ideals corresponds to p = 3 and has inertial degree 2; thus, it has norm
equal to 32 = 9. Thus, each message wk ∈ B29 , which is then split into sub-messages wk,1 , wk,2 ∈ B9 . The
sub-messages w1,ℓ and w2,ℓ are then encoded into xℓ via (20), which is then put into the matrix form in
(21). Moreover, from Table I, we know that 3OL = β1 β2 OL . Therefore, the overall codebook corresponds
17
0
10
Layered GCIC, p=3, S = φ
Layered GCIC, p=3, S = {1}
Layered GCIC, p=3, S = {2}
Layered GCIC, p=5, S = φ
Layered GCIC, p=5, S = {1}
Layered GCIC, p=5, S = {2}
Layered GCIC, p=7, S = φ
Layered GCIC, p=7, S = {1}
Layered GCIC, p=7, S = {2}
−1
10
−2
CER
10
−3
10
−4
10
7.3 dB
10 dB
−5
12.1 dB
10
−6
10
10
15
20
25
30
35
40
45
Es/No
Fig. 2. CER performance for the proposed layered Golden-coded index coding.
to (21) with x0 , x1 ∈ OL /3OL . Simulation results in Fig. 2 show that revealing either message to the
receiver provides roughly 7.3 dB of SNR gain. This conforms with the analysis that when reveal either
message, we expect to achieve SNR gain
118
1
1
+ 10 log10 (9) ≈ 7.45 dB,
10 log10
4
10
2
where 118 and 10 inside the first logarithm are NC and NCS , respectively and the 9 inside the second
logarithm is the ratio of δ(CS ) and δ(C).
In the second set of simulations, the two principal ideals are replaced by those generated by β1 = (1+iθ̄)2
and β2 = (1 − iθ̄)2 , respectively. From Table I, we see that (1 + iθ̄) and (1 − iθ̄) are both corresponding
to p = 5 with inertial degree 1; thus, β1 OL and β2 OL both have norm equal to 52 = 25. Moreover,
5OL = β1 β2 OL ; thereby, the overall codebook corresponds to (21) with x0 , x1 ∈ OL /5OL . Simulation
results in Fig. 2 show that revealing either message to the receiver provides roughly 10 dB of SNR gain.
This again coincides with the analysis which says that by revealing one side information , we can expect
an SNR gain of
1
1
656
+ 10 log10 (25) ≈ 10.27 dB,
10 log10
4
32
2
where 656 and 32 inside the first logarithm are NC and NCS , respectively and the 25 inside the second
logarithm is the ratio of δ(CS ) and δ(C). In the last set of simulations, the two prime ideals corresponding
to p = 7 is considered. Simulation results show that a roughly 12.1 dB SNR gain can be obtained by
revealing either of the message. This again can be well predicted by the analysis which indicates that we
can expect an SNR gain of
1
2042
1
+ 10 log10 (49) ≈ 12.69 dB,
10 log10
4
41
2
where 2042 and 41 inside the first logarithm are NC and NCS , respectively and the 49 inside the second
logarithm is the ratio of δ(CS ) and δ(C).
18
Remark 11. We end this section by showing that the proposed layered Golden-coded index coding is not
a special case of the Golden-coded index coding in [18] and vice versa. The Golden-coded index coding in
[18] is constructed over Z[e][θ] with ideals of the form (α + βe)Z[e][θ] where α, β ∈ Z[i]. Consider p = 17
for which [18, Example 6] indicates that 17Z[e][θ] can be partitioned into 4 ideals, each with norm 172 .
So the Golden-coded index coding can take messages of size 172 . To do the same for our layered scheme,
it requires an ideal in Z[i][θ] to have norm 17, which is impossible from the result in Table I. Now, let us
consider p = 29 where Table I shows that 29Z[i][θ] can be partitioned into four ideals, each with norm 29.
Hence, the proposed layered Golden-coded index coding can take messages of size 292 . This will require
29Z[e][θ] to be partitioned into ideals of the form α + βe with norm 292 . However, using Magma, we
obtain that 29Z[e][θ] = I1 I2 I3 I4 with I1 = (θ̄ + 2i)Z[e][θ], I2 = (θ̄ − 2i)Z[e][θ], I3 = (θ + 2i)Z[e][θ], and
I4 = (θ − 2i)Z[e][θ], where none of these satisfies the form required by the Golden-coded index coding.
VI. LSTIC BASED ON 3 × 3 PERFECT STBC
Let ζ7 be the 7th root of unity and let θ , ζ7 + ζ7−1 = 2 cos 2π
. Also, let K = Q(ω) and let
7
L = Q(ω, θ) the field extension of K with [L : K] = 3. Consider the cyclic division algebra
A = (L/K, σ, γ) = {x0 + x1 e + x2 e2 |x0 , . . . , x2 ∈ L},
where σ : ζ7 + ζ7−1 → ζ72 + ζ7−2 and e3 = γ , j. A 3 × 3 perfect STBC is constructed from
ĀI = {αx0 + αx1 e + αx2 e2 |x0 , . . . , x2 ∈ OL },
where α = 1 + ω + θ. The code will have the matrix form shown in (8).
One can now follow Section IV-A to construct LSTIC based on 3 × 3 perfect STBC. As a result, we
have the following corollary whose proof is identical to that of Theorem 8 together with the fact that
OL = Z[ω][θ] is a principal ideal domain.
Corollary 12. For any S ⊂ {1, . . . , K}, the proposed LSTIC based on 3 × 3 perfect STBC provides
Γ(C, S) = 6 dB/bits per real symbol.
A. Examples and Simulation Results
Here, we again factorize each prime p < 100 into prime ideals via Magma. We show ideals and their
inertial degrees f . The ramification index of each prime ideal lying above p is given by
2, p = 3;
3, p = 7;
e=
1, otherwise.
This can be justified by observing that
d L = 33 74 ,
which has prime factors 3 and 7. Again, since OL is a principal ideal domain, every pOL can be factorized
into principal prime ideals as shown in Table II.
Simulation results for the 3 × 3 case are presented in Fig. 3 where we construct LSTIC from the
3 × 3 perfect STBC with two principal ideals generated by β1 = ((ω − 1)θ2 + (ω − 1)θ − ω + 2) and
β2 = ((−ω + 1)θ2 − (ω − 1)θ + 2ω − 1). From Table II, we learn that both β1 and β2 correspond to p = 7
and we have β1 β2 OL = 7OL . Hence, the overall codebook corresponds to (8) with x0 , x1 , x2 ∈ OL /3OL .
Fig. 3 indicates that by revealing either of the message to the receiver, one obtains a roughly 10.5 dB
SNR reduction. On the other hand, our analysis shows that the SNR reduction one can expect is roughly
5.9 × 1010
1
1
+ 10 log10 (343) ≈ 13.95 dB,
10 log10
9
652428
3
where the parameters inside the first and second logarithms are corresponding to gains in NC and δ(C),
respectively. The difference between the simulation results and our analysis is largely due to the fact that
the SNR gain is measured at 10−4 CER, which is far from the asymptotic regime for a 3 × 3 STBC. This
is evident by observing that the CER curves have not even exhibited the promised diversity order of 9.
19
TABLE II
P RIME FACTORIZATION OF p < 100 IN Z[ω][θ] WHERE θ = ζ7 + ζ7−1 .
p
2
3
5
7
11
13
17
19
23
29
31
37
41
43
47
53
59
61
67
71
73
79
83
89
97
p
(2)
(1 + ω)
(5)
((ω − 1)θ2 + (ω − 1)θ − ω + 2), ((−ω + 1)θ2 − (ω − 1)θ + 2ω − 1)
(11)
(ωθ2 + (ω − 1)θ − ω − 1), ((ω − 1)θ2 − θ − ω + 1), (−θ2 − ωθ + 2)
(ωθ2 + θ − 2ω + 1), (−ωθ2 − θ + ω), (ωθ2 + θ − 2ω)
(17)
(3 − 5ω), (3ω − 5)
(23)
((2ω − 2)θ2 − (ω − 1)θ − 4ω + 4), (3ωθ2 + 2ωθ − 4ω), (3ωθ2 + ωθ − 4ω)
(ω + 5), (5ω + 1)
(7ω − 4), (3ω + 4)
((ω − 1)θ2 − (2ω − 2)θ − 4ω + 4), (3θ2 + θ − 3), ((2 − 2ω)θ2 − (3ω − 3)θ + 4ω − 4)
((ω − 1)θ2 + θ − 2ω + 2), (θ2 + (−ω + 2)θ − 1), ((ω + 1)θ2 + θ − 2ω − 1)
(θ2 + (ω − 1)θ − 2), (−θ2 + (−ω − 1)θ + 1), ((−ω + 2)θ2 + θ + 2ω − 3)
(47)
(53)
(59)
(5ω + 4), (4ω + 5)
(7 − 9ω), (7ω − 9)
(θ2 + θ + 3), (4θ2 + 3θ − 5), ((ω − 1)θ2 − 6ω + 6)
(8ω − 9), (9 − ω)
(7ω + 3), (3ω + 7))
(2θ2 − 2θ − 5), (4θ2 + 2θ − 5), (2ωθ2 + 4jθ − 3ω)
(89)
(−θ2 − θ − 2ω + 3), (θ2 − 2ω), (θ − 2ω + 2)
(−θ2 − θ + 2ω + 1), (θ2 + 2ω − 2), (θ + 2ω)
0
10
3 × 3 LSTIC, p=7, S = φ
3 × 3 LSTIC, p=7, S = {1}
3 × 3 LSTIC, p=7, S = {2}
−1
10
−2
CER
10
−3
10
−4
10
10.5 dB
−5
10
0
5
10
15
20
Es/No
Fig. 3. CER performance for the proposed LSTIC constructed from 3 × 3 STBC.
25
30
35
f
6
3
6
1
6
1
6
3
6
2
3
3
2
1
6
6
6
3
3
2
3
3
2
6
1
20
0
10
4 × 4 LSTIC, p=3, S = φ
4 × 4 LSTIC, p=3, S = {1}
4 × 4 LSTIC, p=3, S = {2}
4 × 4 LSTIC, p=5, S = φ
4 × 4 LSTIC, p=5, S = {1}
4 × 4 LSTIC, p=5, S = {2}
−1
10
−2
CER
10
−3
10
−4
10
8 dB
5.5 dB
−5
10
−6
10
5
10
15
20
25
Es/No
Fig. 4. CER performance for the proposed LSTIC constructed from 4 × 4 STBC.
VII. LSTIC BASED ON 4 × 4 PERFECT STBC
−1
. Also, let K = Q(i) and let
Let ζ15 be the 15th root of unity and let θ , ζ15 + ζ15
= 2 cos 2π
15
L = Q(i, θ) the field extension of K with [L : K] = 4. Consider the cyclic division algebra
A = (L/K, σ, γ) = {x0 + x1 e + x2 e2 + x3 e3 |x0 , . . . , x3 ∈ L},
−1
−2
2
where σ : ζ15 + ζ15
→ ζ15
+ ζ15
and e4 = γ , i. A 4 × 4 perfect STBC is constructed from
ĀI = {αx0 + αx1 e + αx2 e2 + αx3 e3 |x0 , . . . , x3 ∈ OL },
where α = (1 − 3i) + iθ2 . The code will have the matrix form shown in (8).
One can now follow Section IV-A to construct LSTIC based on 4 × 4 perfect STBC. As a result, we
have the following corollary.
Corollary 13. For any S ⊂ {1, . . . , K}, the proposed LSTIC based on 4 × 4 perfect STBC provides
Γ(C, S) ≥ 6 dB/bits per real symbol. Moreover, if all qk , k ∈ {1, . . . , K}, are principal, then Γ(C, S) = 6
dB/bits per real symbol.
A. Examples and Simulation Results
Here, we factorize each prime p < 100 into prime ideals via Magma. In Table III, we show ideals and
their inertial degrees f . The ramification index of each prime ideal lying above p is given by
2, p = 2, 3;
4, p = 5;
e=
1, otherwise.
This can be justified by observing that
d L = 28 34 56 ,
which has prime factors 2, 3, and 5. Also, note that in this case, p = 3, 5, 29, 89 are factorized into
non-principal prime ideals.
21
TABLE III
−1
P RIME FACTORIZATION OF p < 100 IN Z[i][θ] WHERE θ = ζ15 + ζ15
. F OR p = 29, 89, WE ONLY LIST ONE OF THE EIGHT IDEALS DUE
THE SPACE LIMITATION ; THE OTHER SEVEN IDEALS CAN BE OBTAINED AS THE CONJUGATES .
p
2
3
5
7
11
13
17
19
23
29
31
37
41
43
47
53
59
61
67
71
73
79
83
89
97
p
(1 + i)
(3, (5i + 2)θ3 + 7iθ2 + (4i + 4)θ + 7i + 7), (3, 2θ3 + 2iθ2 + (7i + 5)θ + 8i + 6)
(5, (14i + 21)θ3 + (10i + 1)θ2 + (12i + 21)θ + 4i + 22)
(5, (20i + 7)θ3 + (6i + 9)θ2 + (21i + 8)θ + 8i + 5)
((−i + 1)θ3 + (3i − 3)θ − 2i − 1), ((i + 1)θ3 − (3i + 3)θ + 2i − 1)
(2iθ3 + iθ2 − 6iθ − i + 1), (iθ3 − 2iθ + 1)
((i + 1)θ3 + iθ2 − (3i + 4)θ − i), (θ3 − θ2 − 3θ − i + 3)
(2 + 3i), (2 − 3i)
(4 + i), (4 − i)
(iθ3 − (4i − 1)θ + i), (iθ3 − (1 − i)θ2 − 3iθ − 2i + 2)
(iθ3 + (i + 1)θ2 − 3iθ − 2i − 2), (iθ3 − (4i + 1)θ + i)
((3i + 3)θ3 − (9i + 9)θ + 2i + 1), ((3i − 3)θ3 − (9i − 9)θ + 2i − 1)
(29, (813i + 779)θ3 + (812i + 793)θ2 + (755i + 41)θ + 814i + 5)
(−2iθ2 + 5i), (2iθ3 + 2iθ2 − 6iθ − 3i), (2θ3 − 8θ + 1), (2θ − 1)
(6 + i), (6 − i)
((i + 2)θ3 − (3i + 6)θ + i + 1), ((2i − 1)θ3 + (3 − 6i)θ + i)
((2i + 1)θ3 − (6i + 3)θ + i + 1), ((2i + 1)θ3 − (6i + 3)θ + i)
((i − 1)θ3 − (3i − 3)θ + 5i + 4), ((i + 1)θ3 − (3i + 3)θ + 5i − 4)
((3i + 3)θ3 − (9i + 9)θ + 5i − 2), ((3i − 3)θ3 − (9i − 9)θ + 5i + 2)
(2 + 7i), (2 − 7i)
(−2θ3 − θ2 + 7θ − 1), (−iθ3 − iθ2 + 2iθ + 4i), (−iθ2 + iθ + 4i), (−θ3 + θ2 + 4θ − 1)
(θ + i), (θ3 − 4θ + i + 1), (−θ3 − θ2 + 3θ + i + 2), (θ2 + i − 2)
(θ − i), (θ3 − 4θ − i + 1), (−θ3 − θ2 + 3θ − i + 2), (θ2 − i − 2)
((5i + 5)θ3 − (15i + 15)θ + 4i + 1), ((5i − 5)θ3 − (15i − 15)θ + 4i − 1)
((3i + 1)θ3 + (i + 1)θ2 − (10i + 3)θ − i), ((i − 3)θ3 − 2θ2 + (9 − 4i)θ + i + 2)
((i + 3)θ3 + 2θ2 − (4i + 9)θ + i − 2), ((3i − 1)θ3 + (i − 1)θ2 − (10i − 3)θ − i)
(3 + 8i), (3 − 8i)
((2i + 1)θ3 − (7i + 2)θ + i), ((i + 2)θ3 − (i − 1)θ2 − (3i + 6)θ + 3i − 1)
((i − 2)θ3 − (i + 1)θ2 − (3i − 6)θ + 3i + 1), ((1 − 2i)θ3 + (7i − 2)θ − i)
((3i − 3)θ3 − (9i − 9)θ + 7i + 4), ((3i + 3)θ3 − (9i + 9)θ + 7i − 4)
(89, (27i + 82)θ3 + (31i + 117)θ2 + (77i + 7669)θ + 7896i + 7771)
(9 + 4i), (9 − 4i)
TO
f
4
2
1
4
2
4
4
2
4
1
2
4
2
4
4
4
2
1
4
2
4
2
4
1
4
In Fig. 4, two sets of simulation results are presented. Let us consider ideals I1 = (3, (5i + 2)θ3 + 7iθ2 +
(4i+4)θ+7i+7) and I2 = (3, 2θ3 +2iθ2 +(7i+5)θ+8i+6). From Table III and the ramification index of 3,
we learn that 3OL = I21 I22 where N(I21 ) = N(I22 ) = 81. Moreover, with some computation, we have that I21
and I22 are principal ideals with generators β1 = (i + 1)θ3 − 3(i + 1)θ + 1 and β2 = (i − 1)θ3 − 3(i − 1)θ − 1,
respectively. In the first set, we construct LSTIC from 4 × 4 perfect STBC with two principal ideals
corresponding to p = 3 generated by β1 and β2 , respectively. Each message consists of four sub-messages
from Z81 and the overall codebook corresponds to the one in (8) with x0 , x1 , x2 , x3 ∈ OL /3OL . Fig. 4
indicates a roughly 5.5 dB SNR gain by revealing either message to the receiver. We note that the analysis
predicts a roughly
1
1
4.89 × 109
+ 10 log10 (81) ≈ 8.35 dB,
10 log10
16
9099
4
where again the parameters inside the first and second logarithms are corresponding to gains in NC and
δ(C), respectively.
In the second set of simulations, we consider ideals I1 = (5, (14i + 21)θ3 + (10i + 1)θ2 + (12i + 21)θ +
4i + 22) and I2 = (5, (20i + 7)θ3 + (6i + 9)θ2 + (21i + 8)θ + 8i + 5) that correspond to p = 5. Again
from III and the ramification index of 5, we learn that 5OL = I41 I42 where N(I21 ) = N(I22 ) = 625. We
have that I41 and I42 are principal ideals with generators β1 = 2i −1 and β2 = 2i + 1, respectively. We again
construct LSTIC from 4 × 4 perfect STBC with two principal ideals generated by β1 and β2 , respectively.
22
Simulation result in Fig. 4 shows a roughly 8 dB SNR gain obtained by revealing either message to the
receiver. We again note that the analysis predicts a SNR gain of roughly
1
1
4.65 × 1014
+
10 log10
10 log10 (625) ≈ 12.19 dB.
16
2.18 × 106
4
In both the cases, one observes that there is a difference between the simulation results and the analysis.
This again can be explained by that the CER where we measure the side information gain is far from the
asymptotic regime for a 4 × 4 STBC, which is again evident by observing that the CER curves have not
exhibited the promised diversity order of 16.
VIII. LSTIC BASED ON 6 × 6 PERFECT STBC
−1
π
Let ζ28 be the 28th root of unity and let θ , ζ28 + ζ28
= 2 cos 14
. Also, let K = Q(ω) and let
L = Q(ω, θ) the field extension of K with [L : K] = 6. Consider the cyclic division algebra
A = (L/K, σ, γ) = {x0 + x1 e + . . . + x5 e5 |x0 , . . . , x5 ∈ L},
−1
−5
5
where σ : ζ28 + ζ28
→ ζ28
+ ζ28
and e6 = γ , −ω. A 6 × 6 perfect STBC is constructed from
ĀI = {x0 + x1 e + . . . + x5 e5 |x0 , . . . , x5 ∈ I},
where I is such that 7OL = I6 Ī6 .
One can now follow Section IV-B to construct LSTIC based on 6 × 6 perfect STBC. As a result, we
have the following corollary.
Corollary 14. For any S ⊂ {1, . . . , K}, the side information gain achieved by the proposed LSTIC based
on 6 × 6 perfect STBC with non-principal ideal I is lower bounded as
Γ(C, S) ≥ 6 + γI dB/bits per real symbol,
where γI is as shown in (16).
In Table IV, we again factorize each prime p < 100 into prime ideals via Magma. We show ideals and
their inertial degrees f . The ramification index of each prime ideal lying above p is given by
2, p = 2, 3;
6, p = 7;
e=
1, otherwise.
This can be justified by observing that
dL = 212 36 710 ,
which has prime factors 2, 3, and 7. In this case, for p < 100, we note that p = 3, 7, 19, 31 are factorized
into non-principal prime ideals.
IX. L AYERED A LAMOUTI -C ODED I NDEX C ODING
In this section, we construct space-time index codes for 2 × 1 MISO channel from Alamouti code [23].
Alamouti code can be regarded as codes constructed over Hamilton quaternions [34], the R-algebra of
dimension 4 given by
H = {a + bi + cj + dk|a, b, c, d ∈ R},
where i2 = −1, j2 = −1, k2 = −1, and k = ij = −ji. We note that H is a cyclic division algebra
H = (Q(i)/Q, σ, −1) = {x0 + jx1 |x0 , x1 ∈ Q(i)},
23
TABLE IV
−1
.
P RIME FACTORIZATION OF p < 100 IN Z[ω][θ] WHERE θ = ζ28 + ζ28
p
2
3
5
7
11
13
17
19
23
29
31
37
41
43
47
53
59
61
67
71
73
79
83
89
97
p
(θ4 − 5θ2 + θ + 5)
(3, (7ω + 6)θ5 + (3ω + 6)θ4 + (3ω + 2)θ3 + (3ω + 8)θ2 + (8ω + 4)θ + 3ω + 5)
(3, (7ω + 8)θ5 + (ω + 1)θ4 + (ω + 4)θ3 + (7ω + 7)θ2 + (3ω + 7)θ + 5ω + 6)
((ω − 1)θ5 + (ω − 2)θ4 + (4 − 5ω)θ3 + (9 − 5ω)θ2 + (5ω − 2)θ + 6ω − 8)
((1 − ω)θ5 + (1 − 2ω)θ4 + (4ω − 5)θ3 + (9ω − 5)θ2 + (5 − 2ω)θ − 8ω + 6)
(7, (32ω + 15)θ5 + (22ω + 21)θ4 + (23ω + 14)θ3 + (44ω + 10)θ2 + (18ω + 21)θ + 3ω + 20)
(7, (36ω + 12)θ5 + (22ω + 27)θ4 + (26ω + 36)θ3 + (23ω + 42)θ2 + (14ω + 43)θ + 9ω + 29)
((ω + 1)θ5 − (2ω + 1)θ4 − (4ω + 5)θ3 + (9ω + 5)θ2 + (2ω + 5)θ − 8ω − 6)
((ω + 1)θ5 + (2ω + 1)θ4 − (4ω + 5)θ3 − (9ω + 5)θ2 + (2ω + 5)θ + 8 + 6)
4
(ωθ − (5ω − 1)θ2 + 5ω − 3), ((1 − ω)θ4 + (4ω − 5)θ2 − 2ω + 4), ((ω − 1)θ4 + (5 − 4ω)θ2 + 3ω − 5)
((ω − 1)θ4 + (4 − 5ω)θ2 + 5ω − 2), (ωθ4 − (4ω + 1)θ2 + 2ω + 2), ((1 − ω)θ4 + (5ω − 4)θ2 − 5ω + 3)
((ω − 1)θ4 − (2ω + 1)θ3 + (5 − 2ω)θ2 + (3ω + 3)θ + 2ω − 5)
((ω − 1)θ5 + 3θ4 + (3 − 6ω)θ3 + (ω − 13)θ2 + (8ω + 1)θ − 4ω + 10)
(19, (61ω + 187)θ5 + (107ω + 256)θ4 + (123ω + 152)θ3 + (87ω + 168)θ2 + (100ω + 76)θ + 172ω + 278)
(19, (89ω + 144)θ5 + 176ωθ4 + (198ω + 167)θ3 + (42ω + 90)θ2 + (42ω + 214)θ + 134ω + 293)
(19, (103ω + 89)θ5 + (27ω + 254)θ4 + (229ω + 360)θ3 + (296ω + 260)θ2 + (100ω + 197)θ + 61ω + 239)
(19, (158ω + 258)θ5 + (98ω + 84)θ4 + (119ω + 3)θ3 + (159ω + 249)θ2 + (234ω + 184)θ + 28ω + 19)
((3 − 2ω)θ4 − (ω + 1)θ3 + (9ω − 16)θ2 + (ω + 6)θ − 9ω + 16)
((3 − ω)θ4 + (2ω − 1)θ3 + (7ω − 16)θ2 + (6 − 7ω)θ − 7ω + 16)
5
(θ − θ4 − 5θ3 + 4θ2 + 5θ − 3), (θ4 − θ3 − 5θ2 + 3θ + 4), (ωθ2 − ωθ − 3ω)
(ωθ2 + ωθ − 3ω), (θ4 + θ3 − 5θ2 − 3θ + 4), (θ5 + θ4 − 5θ3 − 4θ2 + 5θ + 3)
(31, (724ω + 833)θ5 + (545ω + 827)θ4 + (656ω + 170)θ3 + (771ω + 171)θ2 + (715ω + 907)θ + 680ω + 916)
(31, (45ω + 21)θ5 + (266ω + 398)θ4 + (942ω + 59)θ3 + (506ω + 472)θ2 + (43ω + 69)θ + 210ω + 417)
(31, (927ω + 236)θ5 + (56ω + 700)θ4 + (151ω + 808)θ3 + (525ω + 9)θ2 + (749ω + 157)θ + 951ω + 828)
(31, (194ω + 143)θ5 + (848ω + 7)θ4 + (305ω + 255)θ3 + (521ω + 168)θ2 + (378ω + 357)θ + 861ω + 890)
((1 − ω)(θ5 − θ4 − 5θ3 ) + (5 − 6ω)(θ2 + θ) + 5ω − 3), ((1 − ω)θ4 + (ω − 2)θ3 + (4ω − 3)θ2 + (5 − 2ω)θ − 4ω + 3)
((ω − 1)(θ5 − θ4 − 5θ3 ) + (5ω − 6)(θ2 + θ) − 3ω + 5), (ωθ4 − (ω + 1)θ3 − (4ω − 1)θ2 + (2ω + 3)θ + 4ω − 1)
(θ4 − θ3 − (4 − ω)θ2 − (ω − 3)θ − ω + 2), (θ5 + ωθ4 − (ω + 5)θ3 − (5ω + 1)θ2 + (3ω + 5)θ + 6ω + 2)
(ωθ4 + (ω − 1)θ3 − 4ωθ2 + (3 − 3ω)θ + 4ω − 2), (ωθ4 − θ3 − (4ω + 1)θ2 + (4 − ω)θ + 2ω + 3)
(θ3 − (ω + 1)θ2 − 2θ + 3ω + 1), (ωθ5 + ωθ4 − 5ωθ3 + (1 − 5ω)θ2 + (5ω + 1)θ + 5ω − 2)
((2 − ω)θ4 + (5ω − 9)θ2 − 5ω + 7), ((1 − ω)θ4 + (6ω − 5)θ2 − 7ω + 5), (θ4 + (ω − 4)θ2 − 2ω + 2)
((2ω − 1)θ4 + (5 − 9ω)θ2 + 7ω − 5), ((1 − ω)θ4 + (5ω − 6)θ2 − 5ω + 7), ((ω − 1)θ4 + (4 − 3ω)θ2 − 2)
((1 − ω)(2θ5 − 2θ4 − 7θ3 + 5θ2 + θ + 8), ((ω − 1)(2θ5 + 2θ4 − 7θ3 − 5θ2 + θ − 8))
(3θ4 + 5θ3 − 17θ2 − 13θ + 17), ((ω − 1)(3θ5 − 2θ4 − 20θ3 + 5θ2 + 30θ + 8))
(3ωθ5 − 18ωθ3 + 21ωθ + 2ω), (3ωθ5 − 18ωθ3 + 21ωθ − 2ω)
(5 − 9ω), (4 − 9ω)
(7ω + 2), (2ω + 7)
((2 − ω)θ4 + (ω − 1)θ3 + (5ω − 9)θ2 + (4 − 3ω)θ − 5ω + 7), (−θ5 + θ4 + (6 − ω)θ3 − 6θ2 + (3ω − 9)θ − ω + 8)
(θ5 − θ4 − 5θ3 + (6 − ω)θ2 + (6 − ω)θ + 2ω − 7), (θ5 + θ4 − 5θ3 + (ω − 6)θ2 + (6 − ω)θ − 2ω + 7)
((ω − 1)(θ5 + θ4 − 6θ2 ) + (5 − 6ω)θ3 + (9ω − 6)θ + 8ω − 7), (ωθ4 − θ3 − 4ωθ2 + 3θ + 2ω − 2)
(ω + 8), (8ω + 1)
(7 − 10ω), (3 − 10ω)
(θ5 − 5θ3 − θ2 + 6θ + 4), ((2 − 2ω)θ4 + (ω − 1)θ3 + (9ω − 9)θ2 + (2 − 2ω)θ − 8ω + 8)
((ω − 1)(θ5 + θ4 − 4θ3 − 4θ2 + 2θ + 4), (θ5 − 5θ3 + θ2 + 6θ − 4)
(2θ4 + θ3 − 9θ2 − 2θ + 8), (θ5 − θ4 − 4θ3 + 4θ2 + 2θ − 4)
(5ωθ4 − (2ω + 1)θ3 + (1 − 18ω)θ2 + (2 − ω)θ + 18ω − 1))
((3 − 2ω)θ5 + (3 − 2ω)θ4 + (17ω − 23)θ3 + (15ω − 20)θ2 + (39 − 31ω)θ − 32ω + 38)
(θ2 + 2ω − 4), (ωθ4 − 4ωθ2 + 2ω + 2), ((1 − ω)θ4 + (5ω − 5)θ2 − 3ω + 5)
(θ2 − 2ω − 2), (θ4 − 4θ2 + 2ω + 2), (ωθ4 − 5ωθ2 + 3ω + 2)
f
6
3
6
6
1
6
2
6
3
6
2
3
3
2
2
6
6
6
6
6
2
6
6
2
6
6
2
24
where σ : i → −i and λj = jσ(λ). This induces a layered structure of the Alamouti code. Now, consider
H̄ = {x0 + jx1 |x0 , x1 ∈ Z[i]}, an Alamouti code corresponds to a finite subset of
x0 −x∗1
CZ[i] ,
x0 , x1 ∈ Z[i] .
x1 x∗0
Thus, Alamouti code does not belong to the family of codes considered in Section IV (which have base
fields K = Q(i) or Q(j)). Fortunately, one can follow the same approach and obtain Alamouti-coded index
coding as follows.
H = {x0 + jx1 |x0 , x1 ∈ C}.
which In what follows, we propose and analyze the layered Alamouti-coded index code using an approach
similar to that in Section IV.
Note that Z[i] is a principal ideal domain; so every ideal can be generated by a singleton. Let φ1 , . . . , φK
be K elements in Z[i] that are relatively prime. Also, define q = ΠK
k=1 φk and define N(φk ) = qk for
k ∈ {1, . . . , K} where qk s are not necessarily primes. From CRT, we have
Z[i]/qZ[i] ∼
= Z[i]/φ1 Z[i] × . . . × Z[i]/φK Z[i] ∼
= Bq × . . . × Bq ,
1
K
where Bqk = Z[i]/φk Z[i] is a commutative ring with size qk . Let M be an isomorphism that maps the
messages onto a complete set of coset leaders of Z[i]/qZ[i] with minimum energy. For k ∈ {1, . . . , K},
we enforce wk ∈ B2qk which can be represented as wk = (wk,0, wk,1) where each wk,ℓ ∈ Bqk . The encoder
maps w1,ℓ , . . . , wK,ℓ into the signal of the layer ℓ ∈ {0, 1} as
xℓ = M(w1,ℓ , . . . , wK,ℓ ) ∈ Z[i]/qZ[i],
ℓ ∈ {0, 1}.
The overall codebook becomes a subset of CZ[i] given by
x0 −x∗1
C,
x0 , x1 ∈ Z[i]/qZ[i] .
x1 x∗0
(24)
For the proposed layered Alamouti-coded index coding, we provide the following result without proof.
The proof is essentially identical to the proof of Theorem 8.
Theorem 15. For any S ⊂ {1, . . . , K}, the proposed Alamouti-coded index coding provides Γ(C, S) = 6
dB/bits per real symbol.
A. Examples and Simulation Results
Here, we list choices of φk lying above a prime p < 100. In Table V, we show principal ideals and
their inertial degrees f . From dQ(i) = 4, we know that the ramification index of each prime ideal lying
above p 6= 2 is 1 and is 2 for prime ideals lying above 2.
Simulation results for using the proposed layered Alamouti-coded index coding over the 2 × 1 MISO
channel are provided in Fig. 5. In this figure, we construct the proposed layered Alamouti-index coding
with two ideals generated by β1 = 1 + 2i and β2 = 1 − 2i, respectively. From Table V, we know that
5Z[i] = β1 β2 Z[i] and each ideal has norm equal to p = 5. Each message consists of two sub-messages in
Z5 and we encode the sub-messages of the same layer into the signal of that layer. The overall codebook
becomes (24) with x0 , x1 ∈ Z[i]/5Z[i]. The results in Fig. 5 indicates a roughly 8.1 dB SNR gain when
either message is revealed to the receiver. This can be accurately predicted by our analysis that revealing
either message leads to an SNR gain given by
4
1
1
+ 10 log10 (25) ≈ 8.49 dB,
10 log10
2
2
2
where 4 and 2 in the first logarithms are NC and NCS , respectively, and 25 inside the second logarithm
corresponds to the gain in determinant.
25
TABLE V
P RIME FACTORIZATION OF p < 100 IN Z[i].
p
2
3
5
7
11
13
17
19
23
29
31
37
41
43
47
53
59
61
67
71
73
79
83
89
97
(φ)
(1 + i)
(3)
(1 + 2i), (1 − 2i)
(7)
(11)
(2 + 3i), (2 − 3i)
(1 + 4i), (1 − 4i)
(19)
(23)
(2 + 5i), (2 − 5i)
(31)
(1 + 6i), (1 − 6i)
(5 + 4i), (5 − 4i)
(43)
(47)
(2 + 7i), (2 − 7i)
(59)
(5 + 6i), (5 − 6i)
(67)
(71)
(3 + 8i), (3 − 8i)
(79)
(83)
(5 + 8i), (5 − 8i)
(4 + 9i), (4 − 9i)
f
1
2
1
2
2
1
1
2
2
1
2
1
1
2
2
1
2
1
2
2
1
2
2
1
1
0
10
Layered ACIC, p=5, S = φ
Layered ACIC, p=5, S = {1}
Layered ACIC, p=5, S = {2}
−1
10
−2
CER
10
−3
10
−4
10
−5
10
0
5
10
15
20
25
8.1 dB
30
35
Es/No
Fig. 5. CER performance for the proposed layered Alamouti-coded index coding (ACIC).
X. C ONCLUSIONS
In this paper, we have studied the problem of multicasting K independent messages via MIMO links
to multiple receivers where each of them already has a subset of messages as side information. A novel
scheme, LSTIC, constructed over STBC has been proposed for exploiting side information without prior
knowledge of the side information configuration. It has been shown that the proposed LSTIC possesses
26
the nice property that for any possible side information the minimum determinant increases exponentially
as the rate of the side information increases. Moreover, when constructed over perfect STBC, the perfect
STBC properties are preserved by our construction and therefore the LSTIC is itself a perfect STBC.
Examples including constructions of LSTIC over Golden code, 3 × 3 perfect STBC, 4 × 4 perfect STBC,
6 × 6 perfect STBC, and Alamouti code have been provided and simulations have been conducted to
corroborate our analysis.
R EFERENCES
[1] Y. Birk and T. Kol, “Informed-source coding-on-demand (ISCOD) over broadcast channel,” in Proceedings of the IEEE INFOCOM,
Mar. 1998, pp. 1257–1264.
[2] ——, “Coding on demand by an informed source (ISCOD) for efficient broadcast of different supplemental data to caching clients,”
IEEE Transactions on Information Theory, vol. 52, no. 6, pp. 2825–2830, Jun. 2006.
[3] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” IEEE Transactions on Information Theory, vol. 60, no. 5, pp.
2856–2867, May 2014.
[4] ——, “Coding for caching: Fundamental limits and practical challenges,” IEEE Commununications Magazine, vol. 54, no. 8, pp. 23–29,
Aug. 2016.
[5] T. Oechtering, C. Schnurr, I. Bjelakovic, and H. Boche, “Broadcast capacity region of two-phase bidirectional relaying,” IEEE
Transactions on Information Theory, vol. 54, no. 1, pp. 454–458, Jan. 2008.
[6] G. Kramer and S. Shamai, “Capacity for classes of broadcast channels with receiver side information,” in Proceedings of the IEEE
Information Theory Workshop, Sep. 2007, pp. 313–318.
[7] W.-C. Kuo and C.-C. Wang, “Two-flow capacity region of the cope principle for wireless butterfly networks with broadcast erasure
channels,” IEEE Transactions on Information Theory, vol. 59, no. 11, pp. 7553–7575, Nov. 2013.
[8] Y. Wu, “Broadcasting when receivers know some messages a priori,” in Proceedings of the IEEE International Symposium on Information
Theory, Jun. 2007, pp. 1141–1145.
[9] J. W. Yoo, T. Liu, and F. Xue, “Gaussian broadcast channels with receiver message side information,” in Proceedings of the IEEE
International Symposium on Information Theory, Jun. 2009, pp. 2472–2476.
[10] J. Sima and W. Chen, “Joint network and Gelfand-Pinsker coding for 3-receiver Gaussian broadcast channels with receiver message
side information,” in Proceedings of the IEEE International Symposium on Information Theory, Jun. 2014, pp. 81–85.
[11] B. Asadi, L. Ong, and S. J. Johnson, “Optimal coding schemes for the three-receiver AWGN broadcast channel with receiver message
side information.” IEEE Transactions on Information Theory, vol. 61, no. 10, pp. 5490–5503, Oct. 2015.
[12] E. Tuncel, “Slepian-Wolf coding over broadcast channels,” IEEE Transactions on Information Theory, vol. 52, no. 4, pp. 1469–1482,
Apr. 2006.
[13] L. Natarajan, Y. Hong, and E. Viterbo, “Capacity optimality of lattice codes in common message gaussian broadcast channels with
coded side information,” in Proceedings of the IEEE International Symposium on Information Theory, Jun. 2017, pp. 1833–1837.
[14] A. A. Mahesh and B. S. Rajan, “Index coded PSK modulation,” in Proceedings of the IEEE Wireless Communications and Networking
Conference, Apr. 2016, pp. 1–7.
[15] L. Natarajan, Y. Hong, and E. Viterbo, “Index codes for the Gaussian broadcast channel using quadrature amplitude modulation,” IEEE
Communications Letters, vol. 19, no. 8, pp. 1291–1294, Aug. 2015.
[16] ——, “Lattice index coding,” IEEE Transactions on Information Theory, vol. 61, no. 12, pp. 6505–6525, Dec. 2015.
[17] Y.-C. Huang, “Lattice index codes from algebraic number fields,” IEEE Transactions on Information Theory, vol. 63, no. 4, pp.
2098–2112, Apr. 2017.
[18] Y.-C. Huang, Y. Hong, and E. Viterbo, “Golden-coded index coding,” in Proceedings of the IEEE International Symposium on
Information Theory, Jun. 2017, pp. 2548–2552.
[19] J. Boutros, E. Viterbo, C. Rastello, and J.-C. Belfıore, “Good lattice constellations for both Rayleigh and Gaussian channels,” IEEE
Transactions on Information Theory, vol. 42, no. 2, pp. 502–518, Mar. 1996.
[20] F. Oggier and E. Viterbo, “Algebraic number theory and code design for Rayleigh fading channels,” Foundations and Trends in
Communications and Information Theory, vol. 1, no. 3, pp. 333–415, 2004.
[21] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge University Press, 2005.
[22] F. Oggier, J.-C. Belfiore, and E. Viterbo, “Cyclic division algebras: A tool for space-time coding,” Foundations and Trends in
Communications and Information Theory, vol. 4, no. 1, pp. 1–95, 2007.
[23] S. M. Alamouti, “A simple transmit diversity technique for wireless communication,” IEEE Journal on Selected Areas in Communications, vol. 16, no. 8, pp. 1451–1458, Oct. 1998.
[24] B. A. Sethuraman, B. S. Rajan, and V. Shashidhar, “Full-diversity, high-rate space-time block codes from division algebra,” IEEE
Transactions on Information Theory, vol. 49, no. 10, pp. 2596–2616, Oct. 2003.
[25] J.-C. Belfiore, G. Rekaya, and E. Viterbo, “The golden code: A 2 × 2 full-rate space-time code with nonvanishing determinants,” IEEE
Transactions on Information Theory, vol. 51, no. 4, pp. 1432–1436, Apr. 2005.
[26] F. Oggier, G. Rekaya, J.-C. Belfiore, and E. Viterbo, “Perfect spacetime block codes,” IEEE Transactions on Information Theory,
vol. 52, no. 9, pp. 3885–3902, Sep. 2006.
[27] D. Champion, J.-C. Belfiore, G. Rekaya, and E. Viterbo, “Partitioning the Golden code: A framework to the design of space-time coded
modulation,” in Proc. Canadian Workshop on Inf. Theory, Jun. 2005.
[28] J. H. Conway and N. J. A. Sloane, Sphere Packings, Lattices, and Groups. Springer Verlag, 1999.
[29] T. W. Hungerford, Algebra (Graduate Texts in Mathematics). Springer, 1974.
[30] I. Steward and D. Tall, Algebraic Number Theory and Fermat’s Last Theorem. A K Peters/CRC Press, 2001.
27
[31] S. Lang, Algebraic Number Theory (Graduate Texts in Mathematics). Springer, 1994.
[32] P. Elia, B. A. Sethuraman, and P. V. Kumar, “Perfect space-time codes for any number of antennas,” IEEE Transactions on Information
Theory, vol. 53, no. 11, pp. 3853–3868, Nov. 2007.
[33] W. Bosma, J. Cannon, and C. Playoust, “The Magma algebra system. I. The user language,” J. Symbolic Comput., vol. 24, no. 3-4,
pp. 235–265, 1997. [Online]. Available: http://dx.doi.org/10.1006/jsco.1996.0125
[34] J. H. Conway and D. A. Smith, On Quaternions and Octonions. CRC Press, 2003.
| 7cs.IT
|
arXiv:1402.2642v1 [math-ph] 10 Feb 2014
A comprehensive analysis of the geometry of
TDOA maps in localisation problems ‡
Marco Compagnoniχ , Roberto Notariχ §
Fabio Antonacciς , Augusto Sartiς
χ
Dipartimento di Matematica,
Dipartimento di Elettronica, Informazione e Bioingegneria,
Politecnico di Milano, Piazza L. Da Vinci 32, I-20133 Milano, Italia
ς
E-mail: [email protected], [email protected],
[email protected], [email protected]
Abstract. In this manuscript we consider the well-established problem of
TDOA-based source localization and propose a comprehensive analysis of its
solutions for arbitrary sensor measurements and placements. More specifically,
we define the TDOA map from the physical space of source locations to the space
of range measurements (TDOAs), in the specific case of three receivers in 2D
space. We then study the identifiability of the model, giving a complete analytical
characterization of the image of this map and its invertibility. This analysis has
been conducted in a completely mathematical fashion, using many different tools
which make it valid for every sensor configuration. These results are the first step
towards the solution of more general problems involving, for example, a larger
number of sensors, uncertainty in their placement, or lack of synchronization.
1. Introduction
The localization of radiant sources based on a spatial distribution of sensors has been
an important research topic for the past two decades, particularly in the area of spacetime audio processing. Among the many solutions that are available in the literature,
those based on Time Differences Of Arrival (TDOAs) between distinct sensors of a
signal emitted by the source are the most widespread and popular. Such solutions,
in fact, are characterized by a certain flexibility, a reasonably modest computational
cost with respect to other solutions and a certain robustness against noise. Popular
TDOA-based solutions are [2, 7, 10, 24, 26, 27, 31–33, 41, 44, 46, 48, 49, 52, 54, 55].
Let us consider the problem of planar source localization in a homogeneous
medium with negligible reverberation. From elementary geometry, the locus of
putative source locations that are compatible with a TDOA measurement between
two sensors in positions mi and mj is one branch of a hyperbola of foci mi and mj ,
whose aperture depends on the range difference (TDOA × speed of sound). A single
TDOA measurement is, therefore, not sufficient for localizing a source, but it narrows
‡ This is an author-created, un-copyedited version of an article published in Inverse Problems.
IOP Publishing Ltd is not responsible for any errors or omissions in this version of the
manuscript or any version derived from it.
The Version of Record is available online at
doi:10.1088/0266-5611/30/3/035004.
§ M.Compagnoni and R.Notari should be equally considered as first coauthors of the present work.
Geometry of TDOA maps
2
down the set of locations that are compatible with that measurement by reducing its
dimensionality.
Multiple measurements do enable localization but measurement errors cause the
corresponding hyperbola branches not to meet at a single point, thus ruling out simple
geometric intersection as a solution to the localization problem [14]. This is why
research has focused on techniques that are aimed at overcoming this problem while
achieving robustness. Examples are Maximum Likelihood (ML) [16, 26, 53]; Least
Squares (LS) [2]; and Constrained Least Squares (CLS) [34, 48], which offer accurate
results for the most common configurations of sensors.
There are many situations, however, in which it is necessary to minimize the
number of sensors in use, due to specific sensor placement constraints, or cost
limitations. In these cases it becomes important to assess how the solutions to the
localization problem “behave” (and how many there are) as the measurements or the
sensor geometry vary. This problem has been partially addressed in the case of the
localization of a radio-beacon receiver in LORAN navigation systems [49] and in the
context of the Global Positioning System (GPS), where measurements are of Time Of
Arrivals (TOAs) instead of TDOAs (see [1, 8, 9, 15, 18, 19, 28, 30, 37, 38]). In particular,
these studies provide the solution for the case of planar (2D) source localization with
three receivers (i.e. with two TDOAs) and they recognize the possibility of dual
solutions in some instances, as two different source positions could correspond to the
same pair of TDOA measurements.
Recently, in [51] the author focused on the assessment of the ill-posedness of the
localization problem in the case of 2D minimal sensor configurations, i.e. on quantify
how changes in the measurements propagate onto changes in the estimated source
location. In particular, in the same quoted manuscript it has been introduced the
space of TDOA measurements and it has been shown that in this space there exist
small regions associated with dual solutions corresponding to large regions in physical
space. This assessment, however, is performed in a simulative fashion and for one
specific sensor geometry, and it would be important to extend its generality further.
What we propose in this manuscript is a generalization of the discussion contained
in [51] based on a fully analytical and mathematically rigorous approach. We encode
the TDOA localization problem into a map, called the TDOA map, from the space
of source locations to the space of TDOA measurements and we offer a complete
characterization of such a map. Not only it is our goal to analytically derive results
shown in [51] (irrespective of the geometry of the acquisition system), but also to
complete the characterization of the TDOA map by analyzing the properties of its
image and pre-image, finding closed-form expressions for the boundaries of the regions
of interest. We observe that this approach to the problem fits into the research field
of structural identifiability of complex systems (see for example [11, 40]), where one
is interested in studying if the parameters of a model (in our case, the coordinates of
the source) can be fully retrieved from the experimental data. A similar analysis of
the source localization problem has been proposed and investigated very recently also
in [5, 17], the latter in the context of the TOA–based target tracking.
We believe that characterizing the TDOA map to its fullest extent, even in
the simplest case of three calibrated and synchronous sensors, is a necessary step
for developing new mathematical tools for a wide range of more general problems.
One immediate consequence of this gained knowledge is the possibility to study
how to optimize sensor placement in terms of robustness against noise or measuring
errors. More importantly, this study paves the way to new venues of research.
Geometry of TDOA maps
3
For example, it enables the statistical analysis of error propagation in TDOA-based
localization problems; and it allows us to approach more complex scenarios where
the uncertainty lies with sensor synchronization or spatial sensor placement. This
prospective investigation, in fact, is in line with the recently revamped interest of the
research community in self-calibrating and self-synchronizing spatial distributions of
sensors [16, 45, 47].
Our analysis starts from [20], where a different perspective on the localization
problem is offered through the adoption of the Space–Range Differences (SRD)
reference frame, where the wavefront propagation is described by a (propagation) cone
whose vertex lies on the source location. As range difference measurements (TDOA ×
propagation speed) are bound to lie on the surface of the propagation cone, localizing
a source in the SRD space corresponds to finding the vertex of the cone that best fits
the measured data. The SRD reference frame is also used in [12] to offer geometric
interpretations to the underlying principles behind the most common TDOA-based
localization solutions. Although not explicitly claimed, the localization problem is
described in [12, 20] in terms of null surfaces and planes in the 3D Minkowski space.
This suggests us that exterior algebra can give us powerful tools for approaching our
problem as well. We therefore begin our analysis by showing how the SRD reference
frame can be better represented within the framework of exterior algebra, and we show
how the newly gained tools allow us to derive a global analytical characterization of the
TDOA map. Working with exterior algebra in the Minkowski space is not unheard
of in the literature of space-time signal processing. In [18, 19], for example, this
representation is used for approaching source localization in the GPS context.
The manuscript is organized as shown in Fig. 1. Section 2 introduces the concept
of TDOA map. Two are the TDOA maps defined: τ2 , where the TDOAs are referred to
a common reference microphone; and τ2∗ , which considers the TDOAs between all the
pairs of microphones. The two maps are, in fact, equivalent in absence of measurement
errors. This is why most of the techniques in the literature work with τ2 . However,
in the presence of measurement noise, adopting τ2∗ helps gain robustness. For this
reason we decided to consider both τ2 and τ2∗ . In order to introduce our mathematical
formalisms with some progression, in the first part of the manuscript our analysis will
concern τ2 . Section 3 focuses on the local analysis of the TDOA map τ2 . In practice,
we show what can be accomplished using “conventional” analysis tools (analysis of
the Jacobian matrix). This analysis represents the first step towards the study of the
invertibility of τ2 . In Section 4 we move forward with our representation by defining
the TDOA mapping in the Space - Range Difference (SRD) reference frame. This
is where we show that the Minkowski space is the most natural representation for
a mapping that “lives” in the SRD reference frame. Section 5 describes the early
properties of τ2 , with particular emphasis on the fact that its image is contained in a
compact polygonal region. Section 6 offers a complete description of the mapping τ2
for the case of non-aligned microphones. In particular, Subsection 6.1 shows that the
preimage (inverse image) of τ2 can be described in terms of the non-negative roots of
a degree-2 equation, while 6.3 describes Im(τ ) and the cardinality of the pre-image.
Finally, Subsection 6.4 shows the pre-image regions in τ2 and the bifurcation curve Ẽ
that divides the region of cardinality 1 from the regions of cardinality 0 or 2. Similar
results are derived for the case of aligned microphones in Section 7. In Section 8 we
use the previous results on τ2 to describes the image and the preimages of the map τ2∗ .
Section 9 discusses the impact of this work and offers an example aimed at showing
that the global analysis on τ2 (or τ2∗ ) gives new insight on the localization problem,
Geometry of TDOA maps
4
which could not be derived with a local approach. Finally, Section 10 draws some
conclusions and describes possible future research directions that can take advantage
of the analysis presented in this manuscript.
In order to keep the manuscript as self-contained as possible, in Appendix A
we give an overview on exterior algebra on a vector space. For similar reasons, we
also included an introduction to plane algebraic geometry in Appendix B. These two
Sections, of course, can be skipped by the readers who are already familiar with these
topics. Finally, in Appendix C we included the code for computing the cartesian
equation of the bifurcation curve Ẽ.
Figure 1. Organization of the manuscript.
2. From the physical model to its mathematical description
As mentioned above, we focus on the case of coplanar source and receivers, with
synchronized receivers in known locations and with anechoic and homogenous
propagation. The physical world can therefore be identified with the Euclidean plane,
here referred to as the x–plane. This choice [12,20] allows us to approach the problem
with more progression and visualization effectiveness.
After choosing an orthogonal Cartesian co-ordinate system, the Euclidean x–plane
can be identified with R2 . On this plane, mi = (xi , yi ), i = 0, 1, 2 are the positions
of the microphones and x = (x, y) is the position of the source S. The corresponding
displacement vectors are
di (x) = x − mi ,
dji = mj − mi ,
i, j = 0, 1, 2,
(1)
whose moduli are di (x) and dji , respectively. Generally speaking, given a vector v,
we denote its norm ||v|| with v and with ṽ = vv the corresponding unit vector.
Without loss of generality, we assume the speed of propagation in the medium
to be equal to 1. For each pair of different microphones, the measured TDOA τ̂ji (x)
turns out to be equal to the pseudorange (i.e. the range difference)
τji (x) = dj (x) − di (x), i, j = 0, 1, 2,
(2)
Geometry of TDOA maps
5
plus a measurement error ji :
τ̂ji (x) = τji (x) + ji ,
i, j = 0, 1, 2.
(3)
A wavefront originating from a source in x will produce a set of measurements
(τ̂10 (x), τ̂20 (x), τ̂21 (x)). As the measurement noise is a random variable, we are
concerning with a stochastic model.
Definition 2.1 The complete TDOA model is
τ̂2∗ (x) = (τ̂10 (x), τ̂20 (x), τ̂21 (x)).
The deterministic part of this model is obtained by setting ji = 0 in
gives us the complete TDOA map:
τ2∗ : R2
x
→
→
R3
(τ10 (x), τ20 (x), τ21 (x)).
(4)
τ̂2∗ (x),
which
(5)
The target set is referred to as the τ ∗ –space.
In this manuscript we approach the deterministic problem, therefore we only consider
the complete TDOA map. Using the above definition, localization problems can be
readily formulated in terms of τ2∗ . For example, given a set of measurements, we are
interested to know if there exists a source that has produced them, if such a source is
unique, and where it is. In a mathematical setting, these questions are equivalent to:
• given τ2∗ ∈ R3 , does there exist a source in the x–plane such that τ2∗ (x) = τ ∗ ,
i.e. τ ∗ ∈ Im(τ2∗ )?
• If x exists, is it unique, i.e. |τ2∗ −1 (τ )| = 1?
• If so, is it possible to find the coordinates of x? i.e. given τ ∗ , can we find the
only x that solves the equation τ2∗ (x) = τ ∗ ?
With these problems in mind, we focus on the study of the image of the TDOA map
τ2∗ and of its global properties. In particular, we are interested in finding the locus of
points where the map becomes 1–to–1. Moreover, as solving the localization problem
consists of finding the inverse image of τ ∗ ∈ Im(τ2∗ ), we aim at giving an explicit
description of the preimages, also called the fibers, of τ2∗ .
The complete model τ̂2∗ (x) takes into account each one of the three TDOA that
can be defined between the sensors. This, in fact, becomes necessary when working
in a realistic (noisy) situation [50]. We should keep in mind, however, that there is
a linear relationship between the pseudoranges (3), which allows us to simplify the
deterministic problem.
Definition 2.2 Let (τ10 , τ20 , τ21 ) be the coordinates of the τ ∗ –space. Then, H is the
plane of equation τ10 − τ20 + τ21 = 0.
Lemma 2.3 The image Im(τ ∗2 ) is contained in H.
Proof. For each x ∈ R2 we have
τ10 (x) − τ20 (x) + τ21 (x) = 0
(6)
from the definition (2) of pseudoranges.
In the literature, Lemma 2.3 is usually presented by saying that there are only
two linearly independent pseudoranges and (τ10 (x), τ20 (x)), for example, are sufficient
for completely encoding the deterministic TDOA model. This suggests us to define a
reduced version of the above definition:
Geometry of TDOA maps
6
Definition 2.4 The map from the position of the source in the x-plane to the linearly
independent pseudoranges
τ2 : R2
x
−→
−→
R2
(τ10 (x), τ20 (x))
(7)
is called the TDOA map. The target set is referred to as the τ –plane.
In τ2 we consider only the pseudoranges involving receiver m0 , which we call reference
microphone. If pi : H → R2 is the projection that takes care of forgetting the i–th
coordinate, we have that τ2∗ is related to τ2 by τ2 = p3 ◦ τ2∗ . As pi is clearly 1–to–1,
it follows that all the previous questions about the deterministic localization problem
can be equivalently formulated in terms of τ2 and its image Im(τ2 ) (see Figure 12
in Section 8 for an example of Im(τ2∗ ) and its projection Im(τ2 ) via p3 ). Analogous
considerations can be done if we consider p1 ◦τ2∗ or p2 ◦τ2∗ , that is equivalent to choose
m2 or m1 as reference point, respectively.
In Sections from 3 to 7, we will focus on the study of τ2 and we will complete the
analysis of τ2∗ in Section 8. For reasons of notational simplicity, when we study the
map τ2 we will drop the second subscript and simply write τh (x) = τh0 (x), h = 1, 2.
Moreover, as we focus on the deterministic model, in the rest of the manuscript we
will interchangeably use the terms pseudorange and TDOA.
3. Local analysis of τ2
In this Section, we present a local analysis of the TDOA map τ2 . From a mathematical
standpoint, this is the first natural step towards studying of the invertibility of τ2 .
In fact, as stated by the Inverse Function Theorem, if the Jacobian matrix J(x) of
τ2 is invertible in x, then τ2 is invertible in a neighborhood of x. Studying the
invertibility of a map through linearization (i.e. studying its Jacobian matrix) is a
classical choice when investigating the properties of a complex (non–linear) model.
In the case of acoustic source localization, for example, [20, 45] adopt this method
to study the accuracy of various statistical estimators for the TDOA model. As a
byproduct of our study, at the end of the section we will discuss how the accuracy in
a noisy scenario is strictly related to the existence of the so-called degeneracy locus,
which is the locus where the rank of J(x) drops.
The component functions τi (x) of τ2 are differentiable in R2 \ {m0 , m1 , m2 },
therefore so is τ2 . The i–th row of J(x) is the gradient ∇τi (x), i.e.
x − x0 y − yi
y − y0
x − xi
−
,
−
= d̃i (x) − d̃0 (x).
(8)
∇τi (x) =
di (x)
d0 (x)
di (x)
d0 (x)
Definition 3.1 Let us assume that m0 , m1 , m2 are not collinear. Let r0 , r1 , r2 be
the lines that pass through two of such three points, in compliance with the notation
mi ∈
/ ri , i = 0, 1, 2. Let us split each line in three parts as r0 = r0− ∪ r00 ∪ r0+ , where r00
is the segment with endpoints m1 and m2 , r0− is the half–line originating from m2 and
not containing m1 , and r0+ is the half–line originating from m1 and not containing
m2 . Similar splittings are done for r1 , r2 , with r1+ , r2+ having m0 as endpoint.
Let us now assume that m0 , m1 , m2 belong to the line r. Then, r0 is the smallest
segment containing all three points and rc is its complement in r.
Theorem 3.2 Let J(x) be the Jacobian matrix of τ2 at x 6= m0 , m1 , m2 . Then,
Geometry of TDOA maps
7
Figure 2. A general and a collinear configuration of the microphones mi , i =
0, 1, 2. Left-hand side: line ri+ , ri− and ri0 refer to the portions of the line joining
the microphones j and k, j, k 6= i.
(i) if m0 , m1 , m2 are not collinear, then
1 if x ∈ ∪2i=0 (ri− ∪ ri+ ) ,
rank(J(x)) =
2 otherwise;
(ii) if m0 , m1 , m2 are collinear, then
0
1
rank(J(x)) =
2
if x ∈ rc ,
if x ∈ r0 ,
otherwise.
Proof. Assume x 6= mi , for i = 0, 1, 2. As explained in Section 2, the x–plane is
equipped with the Euclidean inner product, therefore we can use the machinery of
Appendix A. As claimed in Proposition A.3, ∗(det(J(x))) = ∇τ1 (x) ∧ ∇τ2 (x). Hence,
we work in the exterior algebra of the 2–forms. From eq. (8) and the general properties
of 2–forms, we obtain
∇τ1 (x) ∧ ∇τ2 (x) = (d̃1 (x) − d̃0 (x)) ∧ (d̃2 (x) − d̃0 (x)) =
= d̃1 (x) ∧ d̃2 (x) − d̃0 (x) ∧ d̃2 (x) − d̃1 (x) ∧ d̃0 (x).
(9)
Let us first assume that d̃1 (x), d̃2 (x) are linearly independent or, equivalently,
that x ∈
/ r0 . In this case there exist a1 , a2 ∈ R such that
d̃0 (x) = a1 d̃1 (x) + a2 d̃2 (x).
After simplifying equation (9), we get
det(J(x)) = ∇τ1 (x) ∧ ∇τ2 (x) = (−a1 − a2 + 1) d̃1 (x) ∧ d̃2 (x), (10)
therefore det(J(x)) = 0 if, and only if, a1 + a2 = 1, because the linear independence
of d̃1 (x), d̃2 (x) implies d̃1 (x) ∧ d̃2 (x) 6= 0. Furthermore, from d̃0 (x) = a1 d̃1 (x) +
a2 d̃2 (x), we obtain
1 = kd̃0 (x)k2 = a21 + a22 + 2a1 a2 d̃1 (x) · d̃2 (x).
After simple calculations, the previous equality becomes
2a1 a2 (d̃1 (x) · d̃2 (x) − 1) = 0 ,
Geometry of TDOA maps
8
therefore either a1 = 0 or a2 = 0, because the third factor is different from zero. If
a1 = 0, then a2 = 1 and d̃0 (x) = d̃2 (x), i.e. x ∈ r1+ ∪ r1− . Otherwise, if a2 = 0, then
a2 = 1 and d̃0 (x) = d̃1 (x), i.e. x ∈ r2+ ∪ r2− .
On the other hand, if x ∈ r0 , then d̃1 (x) = d̃2 (x) if x ∈ r0+ ∪ r0− , and
d̃1 (x) = −d̃2 (x) if x ∈ r00 . Therefore, the equality (9) becomes
0
if x ∈ r0+ ∪ r0− ,
∇τ1 (x) ∧ ∇τ2 (x) =
−2d̃0 (x) ∧ d̃2 (x) if x ∈ r00 .
In conclusion, if m0 , m1 , m2 are not collinear, then det(J(x)) = 0 for each x ∈
∪2i=0 (ri+ ∪ ri− ), proving the first claim. If, on the other hand, m0 , m1 , m2 lie on the
line r, then det(J(x)) = 0 for all x ∈ r. Furthermore, d̃0 (x) = d̃1 (x) = d̃2 (x) if and
only if x ∈ rc , therefore ∇τ1 (x) = ∇τ2 (x) = (0, 0), i.e. J(x) is the null matrix.
Theorem 3.2 has an interesting geometric interpretation.
Definition 3.3 Let τ ∈ R. The set
Ai (τ ) = {x ∈ R2 | τi (x) = τ }
(11)
is the level set of τi (x) in the x–plane.
Lemma 3.4 If |τ | > di0 , then Ai (τ ) = ∅. Moreover, if 0 < |τ | < di0 , then Ai (τ ) is
the branch of hyperbola with foci m0 , mi and parameter τ, while
+
r
if τ = di0 ,
j
−
Ai (τ ) = rj
if τ = −di0 ,
aj
if τ = 0,
where j 6= i, {i, j} = {1, 2}, and aj is the line that bisects the line segment rj0 .
Proof. By definition, we have τi (x) = di (x) − d0 (x), therefore the first claim follows
from the classical inequalities between the sides of the triangle of vertices x, mi , m0 .
The second claim follows from a classical result: given any hyperbola with foci mi , m0
and parameter c ∈ R+ , the two branches are defined by either one of the two equations
di (x) − d0 (x) = c
and
di (x) − d0 (x) = −c
.
The last claim is a straightforward computation.
Fig. 3(a) shows the hyperbola branches with foci m0 , mi . By definition of level set,
each point in the domain of τi lies on exactly one branch Ai (τ ) for some τ ∈ [−di0 , di0 ]
(by abuse of notation, we consider Ai (0), Ai (±di0 ) as branches of hyperbolas as well).
This means that, given τ = (τ1 , τ2 ), the source is identified as the intersection points
A1 (τ1 )∩A2 (τ2 ). As a direct consequence, the quality of the localization depends on the
type of intersection: in a noisy scenario, an error on the measurements τ changes the
shape of the related hyperbolas, therefore the localization accuracy is strictly related
to the incidence angle between the hyperbolas branches (see [12] for a similar analysis
of the localization problem).
Notation: We denote the tangent line to a curve C at a smooth point x ∈ C as Tx,C .
Geometry of TDOA maps
9
Remark 3.5 (1) ∇τi (x) = 0 if, and only if, x ∈ rj+ ∪ rj− , with j 6= 0, i. In fact,
∇τi (x) = 0 is equivalent to d̃i (x) = d˜0 (x), i.e. x ∈ rj+ ∪ rj− . Hence Ai (±di0 ) is
nowhere smooth.
(2) Assume that x ∈
/ rj+ ∪ rj− . Then, it is well-known that ∇τi (x) is orthogonal to the
line Tx,Ai (τi ) and that it bisects the angle m\
0 xmi , where m0 , mi are the foci of the
hyperbola. Consequently, the tangent line is parallel to the vector d̃i (x) + d̃0 (x) and,
quite clearly, ∇τi (x) = d̃i (x) − d̃0 (x) is orthogonal to the previous vector (as we can
see in Fig. 3(b), if we draw the unit vectors d̃i (x) and d˜0 (x), their sum lies on the
tangent line Tx,Ai (τi ) while their difference is the gradient ∇τi (x)).
(a)
(b)
Figure 3. (a) Level sets Ai (τ ); (b) Gradient and tangent line Tx,Ai (τ ) .
Proposition 3.6 Let x ∈ A1 (τ1 ) ∩ A2 (τ2 ). Then,
(i) if m0 , m1 , m2 are not collinear, then Tx,A1 (τ1 ) 6= Tx,A2 (τ2 ) , or equivalently, A1 (τ1 )
and A2 (τ2 ) meet transversally at x if, and only if, x ∈ R2 \ {∪2i=0 (ri+ ∪ ri− )};
(ii) if m0 , m1 , m2 lie on r, then A1 (τ1 ) ∩ A2 (τ2 ) is finite if, and only if, x ∈ R2 \ rc .
Furthermore A1 (τ1 ) and A2 (τ2 ) meet transversally at x if, and only if, x ∈ R2 \ r.
Proof. The loci A1 (τ1 ) and A2 (τ2 ) meet transversally at x, i.e. Tx,A1 (τ1 ) 6= Tx,A2 (τ2 )
if, and only if, ∇τ1 (x) and ∇τ2 (x) are linearly independent. That last condition
is equivalent to det(J(x)) 6= 0. The claim concerning transversal intersection is
therefore equivalent to Theorem 3.2. Finally, if x ∈ rc , then either A1 (τ1 ) ⊂ A2 (τ2 )
or A2 (τ2 ) ⊂ A1 (τ1 ).
In Fig. 4 we showed a case of tangential intersection of A1 (τ1 ) and A2 (τ2 ). From
Proposition 3.6, we gather new insight on source localization in realistic scenarios.
The above discussion, in fact, allows us to predict the existence of unavoidable poor
localization regions centered on each half–line forming the degeneracy locus. We will
return on this topic in Section 9.
4. The 3–dimensional Minkowski space
As discussed in Section 3, TDOA–based localization is mathematically equivalent to
computing the intersection points of some hyperbola branches. This can be treated as
an algebraic problem in the x–plane by simply considering the full hyperbolas. In this
Geometry of TDOA maps
10
Figure 4. The hyperbola branches intersect tangentially on the degeneracy
locus. This configuration can lead to poor localization accuracy when TDOAs
are affected by measurement errors.
case, however, it is not easy to manipulate the system of two quadratic equations and
remain in full control of all the intersection points. In particular, there could appear
extra (both real and complex) intersection points with no meaning for the problem,
and there is no systematic way to select the ones that are actually related to the
localization.
In order to overcome such difficulties, we manipulate the equations that define
the level sets Ai (τi ) (see Def. 3.3), to obtain an equivalent, partially linear, problem
in a 3D space (see [12] for an introduction on the topic). In order to find the points
in A1 (τ1 ) ∩ A2 (τ2 ), we need to solve the system
τ1 = d1 (x) − d0 (x),
τ2 = d2 (x) − d0 (x).
We introduce a third auxiliary variable τ , and rewrite it as
τ1 − τ = d1 (x),
τ2 − τ = d2 (x),
τ = −d0 (x).
Again, this is not an algebraic problem, because of the presence of Euclidean distances.
However, by squaring both sides of the equations, we obtain the polynomial system
(τ1 − τ )2 = d1 (x)2 ,
(τ2 − τ )2 = d2 (x)2 ,
2
τ = d0 (x)2 .
In geometric terms, this corresponds to studying the intersection of three cones in
the 3D space described by the triplets (x, y, τ ). As described in [12, 20] this problem
representation is given in the space–range reference frame. For the given TDOA
measurements (τ1 , τ2 ), a solution (x̄, ȳ, τ̄ ) of the system gives an admissible position
(x̄, ȳ) of the source in the x–plane and the corresponding time of emission τ̄ of the
signal, with respect to the time of arrival at the reference microphone m0 . We are
actually only interested in the solutions with τ̄ ≤ min(τ1 , τ2 , 0), i.e. in the points that
Geometry of TDOA maps
11
lie on the three negative half–cones. Then, we can use the third equation to simplify
the others, to obtain
d2 −τ 2
d10 · d0 (x) − τ1 τ = 102 1 ,
d2 −τ 2
d20 · d0 (x) − τ2 τ = 202 2 ,
(12)
d0 (x)2 − τ 2 = 0,
τ ≤ min(τ1 , τ2 , 0).
We conclude that, from a mathematical standpoint, that of TDOA-based localization
is a semi–algebraic and partially linear problem, given by the intersection of two planes
(a line) and a half–cone. This is shown in Fig. 5. Notice that the equations in system
(12) involve expressions that are very similar to the standard 3D scalar products and
norms, up to a minus sign in each monomial involving the variable τ or (τ1 , τ2 ). This
suggests that, in order to describe and handle all the previous geometrical objects, an
appropriate mathematical framework is the 3D Minkowski space. In the rest of the
manuscript, we will explore this approach and, in particular, we will carry out our
analysis using the exterior algebra formalism (see also [18, 19] for a similar analysis).
We refer to Appendix A for a concise illustration of the mathematical tools we are
going to use.
Figure 5. The intersection of the two negative half–cones C0 (τ )− and C1 (τ )−
is a curve contained in the plane Π1 (τ ). The curve projects onto the hyperbola
branch A1 (τ ) in the x–plane.
Let e1 , e2 and e3 be the unit vectors of the axes x, y and τ , respectively. Given
the pair τ = (τ1 , τ2 ) on the τ –plane, we define the points Mi (τ ) = (xi , yi , τi ), i = 1, 2,
and M0 = (x0 , y0 , 0). Given a generic point X = (x, y, τ ) in 3D space, the
displacement vectors are defined as Di (X, τ ) = X − Mi (τ ). Furthermore, we set
Dji (τ ) = Mj (τ ) − Mi (τ ), for 0 ≤ i < j ≤ 2. Notice that, in order to render the
notation more uniform, we left all points and vectors as functions of τ , although
many of them actually depend on a single TDOA.
Definition 4.1 For i = 0, 1, 2, we set
(i) Ci (τ ) = {X ∈ R2,1 | k Di (X, τ ) k2 = 0};
(ii) Ci (τ )− = {X ∈ Ci (τ ) | hDi (X, τ ), e3 i ≥ 0}.
Geometry of TDOA maps
12
Moreover, for i = 1, 2, we set
Πi (τ ) = {X ∈ R2,1 | hDi0 (τ ), D0 (X, τ )i =
1
k Di0 (τ ) k2 }
2
and L21 (τ ) = Π1 (τ ) ∩ Π2 (τ ).
Ci (τ ) is a right circular cone with Mi (τ ) as vertex, and Ci (τ )− is a half–cone, while
Πi (τ ) is a plane through (M0 (τ ) + Mi (τ ))/2. Using the exterior algebra formalism
(see eq. (A.3) and the preceding discussion in Appendix A), Πi is given by
1
(13)
iD0 (X) (Di0 (τ )[ ) = k Di0 (τ ) k2 .
2
Finally, if D10 (τ ) and D20 (τ ) are linearly independent, then L21 (τ ) is the line of
equation
1
1
iD0 (X) (D10 (τ )[ ∧ D20 (τ )[ ) = k D10 (τ ) k2 D20 (τ )[ − k D20 (τ ) k2 D10 (τ )[ . (14)
2
2
We are now ready to discuss the link that exists between the geometry of the
Minkowski space and the TDOA–based localization. As above, we set Ai (τ ) = Ai (τi ).
Theorem 4.2 Let π : R2,1 → R2 be the projection onto the x–plane. Then
(i) π(C0− ∩ Ci (τ )− ) = Ai (τ ) if 0 ≤ |τi | ≤ di0 , for i = 1, 2;
Ai (τ )
if − di0 < τi ≤ di0
(ii) π(C0− ∩ Πi (τ )) =
with i 6= j.
Ai (τ ) ∪ rj0 if τi = −di0
Proof. Let x = π(X). We therefore have X = (x, τ ). According to Definition 4.1,
we obtain X ∈ C0− if, and only if, k D0 (X, τ ) k2 = 0 and hD0 (X, τ ), e3 i > 0, which
means that d0 (x)2 − τ 2 = 0, −τ > 0, therefore we finally obtain d0 (x) = −τ, τ < 0.
Similarly, X ∈ Ci (τ )− is equivalent to di (x) = −(τ − τi ), τ < τi . As a consequence,
X ∈ C0− ∩ Ci (τ )− if, and only if, di (x) − d0 (x) = τi , τ < min(0, τi ), i.e. x ∈ Ai (τ ),
therefore the first claim follows.
Then, we remark that Di (X, τ ) = D0 (X, τ ) + Di0 (τ ), that implies
k Di (X, τ ) k2 =k D0 (X, τ ) k2 +2hD0 (X, τ ), Di0 (τ )i+ k Di0 (τ ) k2 .
Hence, X ∈ C0 ∩Π(τ ) if, and only if, X ∈ C0 ∩Ci (τ ), and, using the first claim, we get
π(C0− ∩ Πi (τ )) ⊇ Ai (τ ). C0− ∩ Πi (τ ) is degenerate, precisely a half–line, if, and only if,
M0 ∈ Πi (τ ), i.e. 0 = hDi0 (τ ), 0i = 21 k Di0 (τ ) k2 . The last condition is equivalent to
Mi (τ ) ∈ C0 , or τi2 = di0 . Hence, if τi2 6= d2i0 , π(C0− ∩ Πi (τ )) is a hyperbola branch and
the first equality follows. Otherwise, if τi2 = d2i0 , then π(C0 ∩ Π(τ )) = rj . It is easy to
check that (mi , −di0 ) ∈ C0− and that (mi , di0 ) ∈ C0 \C0− . So, π(C0 ∩Π(τ )) = Ai (τ )∪rj0
if τi = −di0 .
5. First properties of the image of τ2
We now study the set of admissible pseudoranges, i.e. the image Im(τ2 ) of the TDOA
map, in the τ –plane. In particular, in this Section we begin with focusing on the
dimension of the image and then we prove that Im(τ2 ) is contained within a bounded
convex set in the τ –plane. These preliminary results are quite similar for both cases
of generic and collinear microphone configurations, which is the reason why we collect
them together in this Section. For the definition and properties of convex polytopes,
see [39] among the many available references.
Geometry of TDOA maps
13
Theorem 5.1 Im(τ2 ) is locally the τ –plane.
Proof. Let us assume that x̄ is a point where τ2 is regular, i.e. where the Jacobian
matrix J(x̄) has rank 2 (see Theorem 3.2). The map τ2 can be written as
d1 (x) − d0 (x) = τ1
d2 (x) − d0 (x) = τ2
and τ̄ = τ2 (x̄) is a solution of the system. The Implicit Function Theorem guarantees
that there exist functions x = x(τ ) and y = y(τ ), which are defined in a neighborhood
of τ̄ and take on values in a neighborhood of x̄ so that the given system will be
equivalent to
x = x(τ )
,
y = y(τ )
therefore the claim follows.
In Lemma 3.4 we showed that the TDOAs are constrained by the triangular
inequalities. In the rest of this Section we will show that, as a consequence of these
inequalities, τ2 maps the x–plane onto a specific bounded region in the τ –plane.
Definition 5.2 Let
P2 = {τ ∈ R2 | kDji (τ )k2 ≥ 0, 0 ≤ i < j ≤ 2},
and k ∈ {0, 1, 2} be different from i, j, for 0 ≤ i < j ≤ 2. We define
Fk+ = {τ ∈ P2 | kDji (τ )k2 = 0, hDji (τ ), e3 i < 0},
Fk− = {τ ∈ P2 | kDji (τ )k2 = 0, hDji (τ ), e3 i > 0},
R0 = F1+ ∩ F2+ ,
R1 = F0+ ∩ F2− ,
R2 = F0− ∩ F1− .
Before we proceed with studying the relation between P2 and Im(τ2 ), let us describe
the geometric properties of this set. In Fig. 6, we show some examples of P2 (in gray),
for different positions of the points m0 , m1 and m2 .
Theorem 5.3 P2 is a polygon (a 2–dimensional convex polytope). Moreover, if the
points m0 , m1 and m2 are not collinear, then P2 has exactly 6 facets Fk± , which drop
to 4 if the points are collinear.
Proof. As a first step we notice that
kDji (τ )k2 = d2ji − (τj − τi )2 ,
therefore
kDji (τ )k2 ≥ 0
⇔
dji ≥ |τj − τi |.
(15)
The set P2 is a 2–dimensional convex polytope because, according to (15), it is the
intersection of half–planes and it contains an open neighborhood of 0 = (0, 0) ∈ R2 .
In fact, the coordinates of 0 satisfy all the finitely many strict inequalities defining
P2 , which implies that also a sufficiently small open disc centered at 0 belongs to P2 .
In order to prove the rest of the statement, we need to show that the inequalities
defining P2 are redundant if, and only if, m0 , m1 , m2 are collinear. Let us consider
−d10 ≤ τ1 ≤ d10
−d20 ≤ τ2 ≤ d20
.
−d21 ≤ τ2 − τ1 ≤ d21
Geometry of TDOA maps
14
The first two inequalities define a rectangle whose sides are parallel to the τi axes.
The lines τ2 − τ1 = d21 and τ2 = d20 meet at (d20 − d21 , d20 ), which lies between
(−d10 , d20 ) and (d10 , d20 ) if, and only if, |d20 − d21 | < d10 , i.e. when the points
m0 , m1 , m2 are not collinear. Through a similar reasoning we can show that, if
the three points mi are not collinear, the line τ2 − τ1 = d21 meets τ1 = −d10 at
(−d10 , d21 − d10 ), while τ2 − τ1 = −d21 meets τ1 = d10 at (d10 , −d21 + d10 ) and
τ2 = −d20 at (d21 − d20 , −d20 ). An easy check proves that P2 is a hexagon of vertices
(d10 , d20 ), (d20 −d21 , d20 ), (−d10 , d21 −d10 ), (−d10 , −d20 ), (d21 −d20 , −d20 ), (d10 , −d21 +
d10 ).
On the other hand, if m0 , m1 , m2 are collinear, P2 ends up having 4 sides. There
are three possible configurations: (i) m0 lies between m1 and m2 ; (ii) m1 lies between
m0 and m2 ; (iii) m2 lies between m0 and m1 . In case (i) we have that d21 = d10 +d20 ,
therefore −d21 ≤ τ2 −τ1 ≤ d21 are redundant. In case (ii) we have that d20 = d10 +d21
and −d20 ≤ τ2 ≤ d20 give no restrictions to the others. In case (iii), −d10 ≤ τ1 ≤ d10
are redundant as it follows from d10 = d20 + d21 .
Figure 6. Left-hand side: polygon P2 (in shaded gray) under the assumption
that the points m0 , m1 and m2 are not collinear. Center: polygon P2 (in shaded
gray) in the case of three collinear points with m0 between m1 and m2 . Righthand side: polygon P2 (in shaded gray) when the sensors lie on a line, but with
m1 between m0 and m2 . The case with m2 between m0 and m1 can be obtained
from the image on the right by swapping the role of τ1 and τ2 .
For further reference, we name the vertices of the rectangle −d10 ≤ τ1 ≤ d10 , −d20 ≤
τ2 ≤ d20 , recalling that R0 = (d10 , d20 ) (see Definition 5.2).
Definition 5.4 Let R∗ = (−d10 , d20 ), R10 = (−d10 , −d20 ), and R1∗ = (d10 , −d20 ).
We are now ready to present the main result of this section.
Proposition 5.5 Im(τ2 ) ( P2 . Moreover, τ2 −1 (Fk± ) = rk± , k = 0, 1, 2, and, if
m0 , m1 , m2 are not collinear, then τ2 −1 (Rk ) = mk , k = 0, 1, 2.
Proof. The first statement is a direct consequence of Definition 5.2, relation (15) and
Lemma 3.4. Let us now consider x such that ±dji = τj (x) − τi (x) = dj (x) − di (x).
Using Lemma 3.4 we get x ∈ rk± , as claimed. As the preimage of the intersection of
two sets is equal to the intersection of the respective preimages, the last statement
follows from Definition 3.1. Finally, the vertices of P2 that are different from R0 , R1
and R2 are not in Im(τ2 ), because the corresponding half–lines do not meet, as it
Geometry of TDOA maps
15
is easy to verify in all the possible cases. For example, if m0 , m1 and m2 are not
collinear, then r1+ and r0+ do not meet, which implies (d20 − d21 , d20 ) ∈ P2 \ Im(τ2 ).
6. The localization problem in the general case
In this Section we offer further insight on the TDOA map under the assumption
that m0 , m1 , m2 are not collinear. Subsections 6.1 and 6.2 contain some preliminary
mathematical results. In Subsection 6.1 we show how the preimages of the τ2 map are
strictly related to the non–positive real roots of a degree-2 equation, whose coefficients
are polynomials in τ (see eq. (18) and the proof of Theorem 6.16). In order to use
the Descartes’ rule of signs for the characterization of the roots, in Subsection 6.2 we
give the necessary background on the zero sets of such coefficients and on the sign
that the polynomials take on in the τ –plane. The main results of this Section are
offered in Subsections 6.3 and 6.4. In the former we completely describe Im(τ2 ) and
the cardinality of each fiber, while in the latter we derive a visual representation of
the different preimage regions of τ2 in the x–plane, and find the locus where τ2 is
1–to–1. The two Subsections 6.3 and 6.4 also offer an interpretation of such results
from the perspective of the localization problem.
This Section is, in fact, quite central for the manuscript, and the results included
here are mainly proven using techniques coming from algebraic geometry. A brief
presentation of the tools of algebraic geometry that are needed for this purpose is
included in Appendix B. In order to improve the readability of this Section, we
collected some of the proofs in Subsection 6.5.
6.1. The quadratic equation
As discussed in the previous Sections, τ ∈ Im(τ2 ) if, and only if, A1 (τ ) ∩ A2 (τ ) 6= ∅.
According to Theorem 4.2, we have A1 (τ ) ∩ A2 (τ ) ⊆ π(C0− ∩ L21 (τ )), therefore the
analysis of the intersection C0− ∩L21 (τ ) plays a crucial role in characterizing the TDOA
map. We begin with studying the line L21 (τ ) of defining eq. (14).
Assuming that the microphones are not aligned, we have
D1 (X, τ ) ∧ D2 (X, τ ) = (d1 (x) + τ1 e3 ) ∧ (d2 (x) + τ2 e3 ) =
(16)
= d1 (x) ∧ d2 (x) + (τ2 d1 (x) − τ1 d2 (x)) ∧ e3 6= 0
because d1 (x) and d2 (x) are linearly independent. Consequently D1 (X, τ ) and
D2 (X, τ ) are linearly independent as well for every τ ∈ R2 . Let
Ω = D10 (τ ) ∧ D20 (τ ) ∧ e3 = d10 ∧ d20 ∧ e3 6= 0 ,
(17)
which is a 3–form (see Section A.2 in Appendix A). With no loss of generality,
we can assume that Ω is positively oriented, i.e. Ω = kω with k > 0, therefore
hΩ, ωi = −k < 0.
Lemma 6.1 For any τ ∈ R2 , L21 (τ ) = Π1 (τ ) ∩ Π2 (τ ) is a line. A parametric
representation of L21 (τ ) is X(λ; τ ) = L0 (τ ) + λv(τ ), where
v(τ ) = ∗(D10 (τ ) ∧ D20 (τ )) = ∗((d10 ∧ d20 ) + (τ2 d10 − τ1 d20 ) ∧ e3 )
and the displacement vector of L0 (τ ) is
1
∗ kD10 (τ )k2 D20 (τ ) − kD20 (τ )k2 D10 (τ ) ∧ e3 =
D0 (L0 (τ )) =
2∗Ω
∗ kD10 (τ )k2 d20 − kD20 (τ )k2 d10 ∧ e3
= −
.
2kd10 ∧ d20 k
Geometry of TDOA maps
16
Proof. See Subsection 6.5.
Remark 6.2 The point L0 (τ ) is the intersection between L21 (τ ) and the x–plane.
In fact, from the properties of the Hodge ∗ operator, we know that the component of
D0 (L0 (τ )) along e3 is zero.
We can turn our attention to the study of C0− ∩ L21 (τ ). From the definition of
C0 and Lemma 6.1 follows that a point X(λ; τ ) of the line L21 (τ ) lies on C0 if the
vector D0 (L0 (τ )) + λv(τ ) is isotropic with respect to the bilinear form b. This means
that kD0 (L0 (τ )) + λv(τ )k2 = 0 or, more explicitly,
kv(τ )k2 λ2 + 2λhD0 (L0 (τ )), v(τ )i + kD0 (L0 (τ ))k2 = 0.
(18)
This equation in λ ∈ R has a degree that does not exceed 2, and coefficients that
depend on τ .
Definition 6.3 Let
(i) a(τ ) = kv(τ )k2 = kτ2 d10 − τ1 d20 k2 − kd10 ∧ d20 k2 ;
hτ2 d10 − τ1 d20 , kD20 (τ )k2 d10 − kD10 (τ )k2 d20 i
;
(ii) b(τ ) = hD0 (L0 (τ )), v(τ )i =
2kd10 ∧ d20 k
(iii) c(τ ) = kD0 (L0 (τ ))k2 =
kD10 (τ )k2 d20 − kD20 (τ )k2 d10
4kd10 ∧ d20 k2
2
.
Eq. (18) can be rewritten as
a(τ )λ2 + 2b(τ )λ + c(τ ) = 0.
(19)
The explicit solution of eq. (18) will be derived in Subsection 6.4.
6.2. The study of the coefficients
In order to study the solutions of the quadratic equation (18), we need to use Descartes’
rule of signs. To apply it, we first describe the zero set of the coefficients a(τ ), b(τ ),
and c(τ ); then we study the sign of these coefficients wherever they do not vanish. As
stated above, the main mathematical tools that are used in this Subsection come from
algebraic geometry because a(τ ), b(τ ) and c(τ ) are polynomials with real coefficients
(see Appendix B for a short introduction or [13, 22, 29]).
Let us first describe the vanishing locus of c(τ ), over both R, where it is
particularly simple, and C.
Proposition 6.4 c(τ ) ≥ 0 for every τ ∈ R2 . Moreover, c(τ ) = 0 if and only if
τ ∈ {R0 , R∗ , R1∗ , R10 }. On the complex field C, c(τ ) factors as the product of two
degree-2 polynomials.
Proof. See Subsection 6.5.
In order to analyze the sign of a(τ ), we need to introduce some notation.
Definition 6.5 We define three subsets of the τ –plane, according to the sign of a(τ ):
• E = {τ ∈ R2 | a(τ ) = 0};
• E + = {τ ∈ R2 | a(τ ) > 0};
• E − = {τ ∈ R2 | a(τ ) < 0}.
Proposition 6.6 E ⊂ P2 is an ellipse centered in 0 = (0, 0), and it represents the
only conic that is tangent to all sides of the hexagon P2 .
Geometry of TDOA maps
17
Proof. See Subsection 6.5.
The ellipse E and some specific points on the polytope P2 are shown in Fig. 7. As the
tangency points will eventually show up in the study of the vanishing locus of b(τ ),
we define them here for further reference.
Definition 6.7 Let 0 ≤ i, j, k ≤ 2 with k < j and k 6= j. Then
Ti+ = hd10 , d˜jk i, hd20 , d˜jk i
and Ti− = −hd10 , d˜jk i, −hd20 , d˜jk i ,
where, according to our current notation, d˜jk =
djk
djk .
Figure 7. The ellipse E (in blue) is tangent to each side of the hexagon P2 (in
gray). We have 11 distinguished points: the center 0 of E, the six tangency points
Ti± , i = 0, 1, 2, and the vertices of the rectangle R10 , R0 , R∗ and R1∗ .
Remark 6.8 For every non-collinear choice of m0 , m1 , m2 , E is smooth. In fact,
∇a(τ ) = (0, 0) is the homogeneous linear system
2
d20 τ1 − hd10 , d20 iτ2 = 0
−hd10 , d20 iτ1 + d210 τ2 = 0
whose only solution is 0 ∈
/ E, because the matrix of the coefficients of the variables
has determinant kd10 ∧ d20 k2 6= 0.
We conclude the analysis of the sign of a(τ ) by noticing that the set E − contains
the origin 0, therefore it is the bounded connected component of R2 \ E. Similarly,
R0 ∈ E + therefore E + is the unbounded connected component of R2 \ E.
The analysis of the sign of the last coefficients b(τ ) is a bit more involved. Let us
define the notations as done for a(τ ).
Definition 6.9 We define three subsets of the τ –plane, according to the sign of b(τ ):
• C = {τ ∈ R2 | b(τ ) = 0};
• C + = {τ ∈ R2 | b(τ ) > 0};
• C − = {τ ∈ R2 | b(τ ) < 0}.
Geometry of TDOA maps
18
As our aim is to study the relative position of P2 and the sets C, C + , and C − , we need
more of an in-depth understanding of the curve C (see Figure 8 for some examples
of this curve). We will first analyze the role of the 11 distinguished points marked
in Fig. 7 for the study of c(τ ) = 0 and a(τ ) = 0 in connection with b(τ ) = 0. We
will then look for special displacement positions of m0 , m1 and m2 , which force C to
be non–irreducible. In fact, the irreducibility of C has an impact on the topological
properties of C + and C − , particularly on their connectedness by arcs. We will finally
study the connected components of C.
Proposition 6.10 C is a cubic curve with 2–fold rotational symmetry with respect
to 0, which contains T0± , T1± , T2± , R0 , R10 , R∗ , R1∗ and 0. The tangent lines to C at
R0 , R10 , R∗ , R1∗ are orthogonal to F0+ , therefore C is smooth at the above four points.
Finally, C transversally intersects both E and the lines that support the sides of P2 .
Proof. See Subsection 6.5.
Proposition 6.11 C is a smooth curve, unless d10 = d20 . In this case C is the
union of the line L : τ1 + τ2 = 0 and the conic E 0 : τ12 − (hd̃10 , d̃20 i + 1)τ1 τ2 + τ22 +
d210 (hd̃10 , d̃20 i − 1) = 0.
Proof. See Subsection 6.5.
For the sake of completeness, we now investigate the uniqueness of this cubic
curve by showing that C is completely determined by the positions of the points m0 ,
m1 , m2 .
Proposition 6.12 C is the unique cubic curve that contains the points T0± , T1± , T2± ,
0, R0 , R10 , R∗ , R1∗ .
Proof. See Subsection 6.5.
Remark 6.13 Due to the 2–fold rotational symmetry around 0, and the fact that C
is smooth at 0, we can conclude that 0 is an inflectional point for C.
The cubic curve C, where smooth, has genus 1. Therefore, in the τ –plane, it
can have either 1 or 2 ovals, in compliance with Harnack’s Theorem B.31 (see Fig.
8). Depending on the position of m0 , m1 , m2 , both cases are possible. Following
standard notation, the two ovals are called Co , the odd oval, and Ce , the even one (this
could be missing), and, at least in the projective plane P2R , they are the connected
components of C. The importance of studying the connected components of C rests
on the fact that C divides every neighborhood of a point P ∈ C in two sets, one in
C + , the other in C − . Therefore, we need to locate Co and Ce with respect to P2 .
Proposition 6.14 The points T0± , T1± , T2± , 0, R0 , R∗ , R10 , R1∗ belong to the same
connected component Co of C, which is the only one that intersects P2 .
Proof. See Subsection 6.5.
Now we can complete the study of the sign of b(τ ) within P2 . Let us first assume
that C is smooth. Due to the rotational symmetry of C, the component Co is connected
in the affine plane R2 as well, and it divides the τ –plane in two disjoint sets, which
we name Co+ and Co− . Due to Proposition 6.14, b(τ ) does not change sign on P2 ∩ Co+
and P2 ∩ Co− , therefore we have P2 ∩ C + = P2 ∩ Co+ and P2 ∩ C − = P2 ∩ Co− (possibly
with Co+ , Co− in swapped order). In particular, evaluating b(τ ) at the vertices of P2 ,
we have that Co+ is the connected component of R2 \ Co containing R1 , R2 .
Geometry of TDOA maps
19
Figure 8. Examples of cubics C: on the left-hand side it is singular. In the centre
it is an oval and on the right-hand side are two ovals. The curve E is shown in
blue, C in red, and the hexagon P2 in shaded gray. The 11 distinguished points
are marked in the first two pictures, but not in the last one because 4 of them are
very close to each other on the upper–right vertex of the rectangle, and similarly
for the 4 ones, which are close to the opposite vertex. In all the three cases P2 is
an hexagon, but it exhibits two very short sides in the right-hand picture.
Finally, if C is singular we have C = L ∪ E 0 (see Fig. 8). There are four disjoint
regions in the τ –plane having different signs. Again by evaluating b(τ ) at the vertices
of P2 , we obtain that C + is the union of the region outside E 0 in the half plane
containing R1 , R2 plus the region inside E 0 in the complementary half plane.
6.3. The image of τ2
In this Subsection we achieve one of the main goals of the manuscript, as we derive
the complete and explicit description of Im(τ2 ), i.e. the set of admissible TDOAs.
These results are summarized in Fig. 9. In the following, we will denote the closure
of a set U as Ū and its interior as Ů .
Definition 6.15 The set P˚2 ∩ E + ∩ C + is the union of three disjoint connected
components that we name U0 , U1 , U2 , where Ri ∈ Ūi for i = 0, 1, 2.
Theorem 6.16 Im(τ2 ) = E − ∪ Ū0 ∪ Ū1 ∪ Ū2 \ {T0± , T1± , T2± }. Moreover,
(
2
if τ ∈ U0 ∪ U1 ∪ U2 ,
|τ2 −1 (τ )| =
1
if τ ∈ Im(τ2 ) \ U0 ∪ U1 ∪ U2 .
Proof. Consider the equation (19)
a(τ )λ2 + 2b(τ )λ + c(τ ) = 0,
with τ ∈ P2 . The reduced discriminant ∆(τ )/4 = b(τ )2 − a(τ )c(τ ) is a degree6 polynomial that vanishes if L21 (τ ) is tangent to the cone C0 . According to
Theorem 4.2, this condition is equivalent to A1 (τ ) and A2 (τ ) intersecting tangentially.
According to Proposition 3.6, this happens exactly if x ∈ r0± ∪ r1± ∪ r2± . Hence,
τ ∈ τ2 (r0± ∪ r1± ∪ r2± ) = F0± ∪ F1± ∪ F2± , which implies ∆(τ ) = 0 if, and only if,
τ ∈ ∂P2 . On the other hand, ∆(0) = −a(0)c(0) > 0 because 0 ∈ E − , therefore
Geometry of TDOA maps
20
∆(τ ) > 0 for τ ∈ P̊2 . As a consequence, equation (19) has real solutions for any
τ ∈ P2 .
According to Theorem 4.2, we are looking for τ that satisfies C0− ∩ L21 (τ ) 6= ∅:
0 ≤ hD0 (L0 (τ )) + λv(τ ), e3 i = hD0 (L0 (τ )), e3 i + λhv(τ ), e3 i =
= λh∗(d10 ∧ d20 ), e3 i = λh∗(d10 ∧ d20 ), ∗(e1 ∧ e2 )i =
= −λhd10 ∧ d20 , e1 ∧ e2 i = λhd10 ∧ d20 ∧ e3 , ωi = λhΩ, ωi,
which narrows down to λ ≤ 0, as hΩ, ωi < 0.
Let us first consider the case λ = 0, which is equivalent to c(τ ) = 0. From
Proposition 6.4, we know that c(τ ) ≥ 0 for any τ ∈ P2 , and c(τ ) = 0 if, and only
if, τ ∈ {R0 , R∗ , R10 , R1∗ }. At the four considered points, also b(τ ) = 0 and ∆(τ ) = 0.
Hence, λ = 0 is the only solution with multiplicity 2, if τ = R0 or τ = R10 , the other
two points not being in P2 . However, the half–lines r1+ and r2+ meet at x = m0 , while
r1− ∩ r2− = ∅. Consequently, τ2 −1 (R0 ) = m0 , while R10 ∈
/ Im(τ2 ).
Let us now assume λ 6= 0, i.e. c(τ ) > 0, and consider all the possible cases, one at a
time. The main (and essentially unique) tool is Descartes’ rule of signs for determining
the number of positive roots of a polynomial equation, with real coefficients and real
roots.
Case (i): a(τ ) = b(τ ) = 0.
Eq. (19) has no solution, therefore E ∩ C = {T0± , T1± , T2± } is not in Im(τ2 ).
Case (ii): a(τ ) = 0, b(τ ) 6= 0.
Eq. (19) has the only solution λ = −c(τ )/2b(τ ) for each τ ∈ E \ {T0± , T1± , T2± }.
Moreover, λ < 0 if, and only if, b(τ ) > 0 i.e. τ ∈ E ∩ C + = (∂U0 ∪ ∂U1 ∪ ∂U2 ) ∩
E \ {T0± , T1± , T2± }.
Case (iii): a(τ ) < 0.
Equation (19) has one negative root and one positive root, thus E − ⊂ Im(τ2 )
and |τ2 −1 (τ )| = 1 for each τ ∈ E − .
Case (iv): a(τ ) > 0, b(τ ) < 0.
Eq. (19) has two positive roots, thus E + ∩ C − ∩ Im(τ2 ) = ∅.
Case (v): a(τ ) > 0, b(τ ) > 0, ∆(τ ) = 0.
Eq. (19) has one negative root with multiplicity 2, thus |τ2 −1 (τ )| = 1 for each
τ ∈ E + ∩ C + ∩ ∂P2 . In particular, τ2 −1 (Rj ) = mj for j = 1, 2.
Case (vi): a(τ ) > 0, b(τ ) > 0, ∆(τ ) > 0.
Eq. (19) has two distinct negative roots, therefore |τ2 −1 (τ )| = 2 for any
τ ∈ U0 ∪ U1 ∪ U2 .
Remark 6.17 Theorem 6.16 agrees with Theorem 4.2. The exact relationship
between A1 (τ ) ∩ A2 (τ ) and π(C0− ∩ L21 (τ )) is the following:
(1) If τ ∈
/ F2− ∪ F1− , then π(C0− ∩ L21 (τ )) = A1 (τ ) ∩ A2 (τ ).
(2) If τ ∈ F2− \ {R10 }, then τ1 = −d10 and −d20 < τ2 ≤ d21 − d10 . Thus
π(C0− ∩ L21 (τ )) = (A1 (τ ) ∩ A2 (τ )) ∪ (r20 ∩ A2 (τ )),
where A1 (τ ) = r2− . If x ∈ r20 ∩ A2 (τ ), then d0 (x) = d10 − d1 (x) and
τ2 (x) = d2 (x) − d0 (x) = d2 (x) + d1 (x) − d10 ≥ d21 − d10 ,
where we used the triangular inequality. It follows that τ2 (x) = d21 − d10 and x =
m1 , so r20 ∩ A2 (τ ) = A1 (τ ) ∩ A2 (τ ) and, again, π(C0− ∩ L21 (τ )) = A1 (τ ) ∩ A2 (τ ).
The case τ ∈ F1− \ {R10 } is similar.
Geometry of TDOA maps
21
Figure 9. The image of τ2 is the gray subset of P2 . In the light gray region
marked as E − the map τ2 is 1–to–1, while in the medium gray regions U0 ∪U1 ∪U2
the map τ2 is 1–to–2. The continuous part of ∂P2 and E, and the vertices Ri ,
are in the image, where τ2 is 1–to–1. The dashed part of ∂P2 and E, and the
tangency points Ti± , do not belong to Im(τ2 ). We remark that the triangles
U0 , U1 , U2 stay on the same connected component of R2 \ Co (see Fig. 8)
(3) If τ = R10 ∈
/ Im(τ2 ), then A1 (τ )∩A2 (τ ) = ∅ while π(C0− ∩L21 (τ )) = r10 ∪r20 = m0 .
Theorem 6.16 can be nicely interpreted in terms of the two-dimensional and the
three-dimensional intersection problems. Here we use some standard Minkowski and
relativistic conventions used, for example, in [3, 43].
(i) τ ∈ E if, and only if, v(τ ) is isotropic, or light–like. In this case, the line L21 (τ )
is parallel to a generatrix of the cone, therefore it meets C0 at an ideal point.
On the x–plane this means that the level sets A1 (τ ) and A2 (τ ) have one parallel
asymptote. With respect to the localization problem, τ ∈ E means that there
could exist a source whose distance from the microphones is large compared to
d10 and d20 . Along E, the two TDOAs are not independent and we are able to
recover information only about the direction of arrival of the signal, and not on
the source location. Things complicate further if τ ∈ E ∩ Im(τ2 ), as the level
sets A1 (τ ) and A2 (τ ) also meet at a point at finite distance, corresponding to
another admissible source location.
(ii) τ ∈ E − if, and only if v(τ ) is time–like, pointing to the interior of the cone C0 .
In this case, the line L21 (τ ) intersects both half–cones and, on the x–plane, the
level sets A1 (τ ) and A2 (τ ) meet at a single point. This is the most desirable case
for localization purposes: a τ corresponds to a unique source position x.
(iii) τ ∈ E + if, and only if, v(τ ) is space–like, pointing to the exterior of the cone
C0 . In this case, the line L21 (τ ) intersects only one half–cone, depending on the
position of the point L0 (τ ) and the direction of v(τ ). On the x–plane, the level
sets A1 (τ ), A2 (τ ) either do not intersect or intersect at two distinct points. In the
last case, for a given τ there are two admissible source positions. Following the
discussion at point (i), a source runs away to infinity as τ gets close to E, while
the other remains at a finite position, which suggests a possible way to distinguish
between them if one has some a-priori knowledge on the source location. Finally,
we observe that the two solutions overlap if τ ∈ ∂P2 , which corresponds to x in
the degeneracy locus.
Geometry of TDOA maps
22
If τ ∈ E − , the localization is still possible even in a noisy scenario, but we experience
a loss in precision and stability as τ approach E (see also the discussion in Section 9).
6.4. The inverse image
We are now ready to reverse the analysis. In fact, the description of Im(τ2 ) allows us
to analyze the dual situation in the physical x–plane. For any given τ ∈ Im(τ2 ) and
a negative solution λ of eq. (18), we have the corresponding preimage in the x–plane
x(τ ) = L0 (τ ) + λ ∗ ((τ2 d10 − τ1 d20 ) ∧ e3 ),
(20)
where ∗((τ2 d10 − τ1 d20 ) ∧ e3 ) is the projection of v(τ ) on the subspace spanned by
e1 , e2 . Roughly speaking, we can identify two distinct regions: the preimage of the
interior of the ellipse, where the TDOA map is 1–to–1 and the source localization
is possible, and the preimage of the three triangles Ui , i = 0, 1, 2, where the map is
2–to–1 and there is no way to locate the source. The region of transition is also known
in the literature as the bifurcation region [19]. In this subsection we offer a complete
geometric description of the above sets.
Notice that that formula (20) gives the exact solutions x to the localization
problem for any given measurements τ , and it can be used as the starting point and
building block for a local error propagation analysis in the case of noisy measurements
or even with sensor calibration uncertainty.
Definition 6.18 Let E be the ellipse in the τ –plane defined by a(τ ) = 0. We call Ẽ
its inverse image contained in the x–plane, and we refer to it as the bifurcation curve.
As we said in the discussion at the end of Subsection 6.3, for τ ∈ E we have an
admissible source position at an ideal point of the x–plane and, possibly, one more at
a finite distance from the sensors. In the affine plane, the curve Ẽ is exactly the set
of these last points. According to Definition 6.18, Ẽ is the preimage of E, therefore
it can be studied using formula (20). We recall that for τ ∈ E we have a(τ ) = 0,
therefore eq. (18) has a unique solution in λ(τ ) = −c(τ )/2b(τ ), which corresponds to
the unique preimage
c(τ )
∗ ((τ2 d10 − τ1 d20 ) ∧ e3 ).
(21)
x(τ ) = L0 (τ ) −
2b(τ )
In the next Theorem, we show that the function (21) restricted on E is a rational
parametrization of degree 5 of the bifurcation curve Ẽ. This means that Ẽ admits a
characterization as an algebraic curve.
Theorem 6.19 Ẽ is a rational degree–5 curve, whose ideal points are the ones of the
lines r0 , r1 , r2 and the cyclic points of P2C .
Proof. See Subsection 6.5.
In Fig. 10 we show two examples of the quintic Ẽ. The real part of Ẽ consists of
three disjoint arcs, one for each arc of E contained in Im(τ2 ). The points m0 , m1 , m2
do not belong to Ẽ, as their images via τ2 are not on E. Notice that no arc is
bounded, as Ẽ has genus 0. In particular, when τ approaches a point Ti± in E ∩ C the
denominator of x(τ ) approaches to zero and Ẽ goes to infinity. As for the smoothness
of Ẽ, the curve has no self–intersection because each point of the x–plane has one
image in the τ –plane. Furthermore, it is quite easy to show that cusps are not allowed
on Ẽ either. In fact, E is regularly parameterized and the Jacobian matrix of τ2 is
invertible on Ẽ, which implies the regularity of x(τ ). Quite clearly, on the complex
Geometry of TDOA maps
23
plane Ẽ is bound to have singular points, as Ẽ is an algebraic rational quintic curve.
In Appendix C we include the source code in Singular [23] language for computing
the Cartesian equation (further analysis of the properties of the bifurcation curve is
contained in [21]).
From Fig. 9 and Fig. 10, we immediately recognize what was assessed through
simulations and for a specific sensor configuration in [51]. These results, however, have
been here derived in closed form and for arbitrary sensor geometries, which allows us
to characterise the pre–image in an exhaustive fashion.
The curve Ẽ separates the regions of the x–plane where the map τ2 is 1–to–
1 or 2–to–1. We complete the analysis in terms of TDOA–based localization after
introducing and analyzing the preimage of the open subsets E − , U0 , U1 , U2 of Im(τ2 ).
Definition 6.20 Let Ũi be the inverse image of Ui via τ2 , for i = 0, 1, 2, and Ẽ − be
the inverse image of E − .
The continuity of τ2 implies that Ẽ − , Ũ0 , Ũ1 , Ũ2 are open subsets of the x–plane,
which are separated by the three arcs of Ẽ. Let F (x, y) = 0 be a Cartesian equation
of Ẽ: a point x ∈ Ẽ − if F (x)F (m0 ) < 0, while x ∈ Ũ0 ∪ Ũ1 ∪ Ũ2 if F (x)F (m0 ) > 0.
Now, let us focus on the open sets Ũi . In this case, without loss of generality, we
consider i = 0, the other two ones having the same properties.
Proposition 6.21 Ũ0 has two connected components separated by r1+ ∪ r2+ , and τ2 is
1–to–1 on each of them.
Proof. See Subsection 6.5.
Figure 10. Two examples of the different localization regions and the quintic Ẽ
in the x–plane. The microphones m0 , m1 and m2 are the in the points marked
with black dots. Locations of the microphones are m0 = (0, 0), m1 = (2, 0),
m2 = (2, 2) and m2 = (−2, 2) on the left and the right, respectively. Each
quintic Ẽ separates the light gray region Ẽ − , where the map τ2 is 1–1 and it is
possible to localize the source, and the dark gray region Ũ0 ∪ Ũ1 ∪ Ũ2 , where τ2 is
2–1 and the localization is not unique. The dashed lines represent the degeneracy
locus of τ2 .
Remark 6.22 The previous Proposition can be restated by saying that τ2 : Ũi → Ui
is a double cover, for every i = 0, 1, 2. The ramification locus is the union of the
two half–lines through mi , while the branching locus is union of the two facets of P2
through Ri .
Geometry of TDOA maps
24
The source localization is possible if τ ∈ E − and, consequently, x ∈ Ẽ − .
Otherwise, assume τ ∈ U0 . According to Proposition 6.21, there are two admissible
sources in the two disjoint components of Ũ0 . As τ comes close to E, one of its inverse
images approaches a point on Ẽ, while the other one goes to infinity. Conversely, if τ
approaches ∂P2 , the inverse images of τ come closer to each other and converge to a
point on the degeneracy locus r1+ ∪ r2+ . As we said above, in a realistic noisy scenario,
we end up with poor localization in the proximity of Ẽ.
6.5. Proofs of the results
Proof of Lemma 6.1. As remarked before eq. (14), L21 (τ ) is a line because D10 (τ )
and D20 (τ ) are linearly independent. Thus, the equation of L21 (τ ) is (14):
iD0 (X) (D10 (τ )[ ∧ D20 (τ )[ ) =
1
1
k D10 (τ ) k2 D20 (τ )[ − k D20 (τ ) k2 D10 (τ )[ .
2
2
A vector v(τ ) is parallel to L21 (τ ) if it is a solution of
iv(τ ) (D10 (τ )[ ∧ D20 (τ )[ ) = 0
From Corollary A.6, this is equivalent to
v(τ ) = t ∗ (D10 (τ )[ ∧ D20 (τ )[ )] = t ∗ (D10 (τ ) ∧ D20 (τ ))
for t ∈ R. We prove the first claim of the Lemma by setting t = 1.
Then, let L0 (τ ) be the intersection point between L21 (τ ) and the x–plane. This
implies that
iD0 (L0 (τ )) Ω[ =
1
k D10 (τ ) k2 D20 (τ )[ − k D20 (τ ) k2 D10 (τ )[ ∧ e3 [ .
2
Therefore, the second claim follows from Lemma A.7.
Proof of Proposition 6.4. As a real function, c(τ ) ≥ 0 because kD10 (τ )k2 d20 −
kD20 (τ )k2 d10 is in the subspace spanned by e1 , e2 , where b is positive-defined, and
d10 ∧ d20 is parallel to e1 ∧ e2 , whose module is equal to 1. Furthermore, d10 , d20
are linearly independent, thus c(τ ) = 0 if, and only if, kD10 (τ )k2 = kD20 (τ )k2 = 0,
i.e. τ12 = d210 , τ22 = d220 , and the claim follows.
The gradient of c(τ ) is
∇c(τ ) =
1
h−τ1 d20 , kD10 (τ )k2 d20 − kD20 (τ )k2 d10 i,
kd10 ∧ d20 k2
hτ2 d10 , kD10 (τ )k2 d20 − kD20 (τ )k2 d10 i ,
therefore it vanishes if kD10 (τ )k2 = kD20 (τ )k2 = 0. Hence, in A2C , c(τ ) = 0 is a
quartic algebraic curve with four singular points, thus it cannot be irreducible (see
Theorem B.22). After some simple computations, we obtain
c(τ ) = d20 e−iθ (τ12 − d210 ) − d10 eiθ (τ22 − d220 ) d20 eiθ (τ12 − d210 ) − d10 e−iθ (τ22 − d220 )
where 2θ ∈ (0, π) is the angle between d10 and d20 .
Geometry of TDOA maps
25
Proof of Proposition 6.6. The equation that defines E, i.e.
a(τ ) = −kd10 ∧ d20 k2 + kτ2 d10 − τ1 d20 k2 = 0 ,
has degree 2, therefore E is a conic in the τ –space. Considering the assumption
of non-collinearity, kτ2 d10 − τ1 d20 k2 is a positively-defined quadratic form and
kd10 ∧ d20 k2 > 0. E is therefore a non-degenerate ellipse containing real points,
whose center is at 0. Moreover, it is a simple matter of computation to verify that the
intersection between E and Fi+ is the point Ti+ with multiplicity 2, for i = 0, 1, 2, and
analogously for Fi− and Ti− . E is therefore tangent to each side of P2 . This implies
also that E ⊂ P2 . In order to prove the uniqueness of E, we embed the τ –plane
R2 into a projective plane P2R , and take the dual projective plane P̌2R (see Definition
B.24). In P̌2R there exists one conic through the 6 points corresponding to the sides of
P2 and it is the dual conic Ě (see Definition B.25 and Proposition B.26). Moreover,
Ě is unique by Corollary B.29. We conclude that the uniqueness of Ě is equivalent to
the uniqueness of E.
Proof of Proposition 6.10. C is defined by the degree–3 polynomial equation
b̄(τ ) = hτ2 d10 − τ1 d20 , kD20 (τ )k2 d10 − kD10 (τ )k2 d20 i = 0,
therefore it is a cubic curve. It is easy to verify that
• the equation does not change if we replace τi with −τi , i = 1, 2, therefore C has
a 2–fold rotational symmetry with respect to 0;
• C contains all the 11 points of the statement.
The partial derivatives of b̄ are
∂ b̄
= −hd20 , kD20 (τ )k2 d10 − kD10 (τ )k2 d20 i + hτ2 d10 − τ1 d20 , 2τ1 d20 i,
∂τ1
∂ b̄
= hd10 , kD20 (τ )k2 d10 − kD10 (τ )k2 d20 i − hτ2 d10 − τ1 d20 , 2τ2 d10 i.
∂τ2
After simple calculations, we obtain ∇b̄(R0 ) = 2d10 d20 (d10 d20 − hd10 , d20 i) (1, 1) and
∇b̄(R∗ ) = −2d10 d20 (d10 d20 + hd10 , d20 i) (1, 1). The gradient of b̄ is therefore non-zero
at both R0 and R∗ , i.e. R0 and R∗ are smooth on C. Moreover, the tangent lines to
C at R0 and R∗ are orthogonal to (1, 1) therefore they are orthogonal to F0+ . For
symmetry, the same holds at R10 and R1∗ .
In compliance with Bézout’s Theorem B.16, C and E meet at 6 points after
embedding the τ –plane into P2C , but C ∩ E = {T0± , T1± , T2± }, thus C and E intersect
transversally. Moreover, we use Bézout’s Theorem also to prove that C is not tangent
to any line among F2± , and F1± , because the points where the curve C meets each line
are known.
Finally, the line containing F0+ meets C at T0+ plus two other points whose
coordinates solve τ2 = τ1 + d21 , (d21 τ1 − hd21 , d10 i)2 + kd10 ∧ d20 k2 = 0, therefore
they cannot be real. As a consequence, according to Bézout’s Theorem, we obtain
that C and F0+ are not tangent. By symmetry, F0− is not tangent to C either.
Proof of Proposition 6.11. The gradient at 0 is ∇b̄(0) = (−d220 hd21 , d10 i, d210 hd21 , d20 i) 6=
(0, 0). Hence, if C is not smooth, there are at least two singular points, because of the 2–fold rotational symmetry, and so C is reducible. As C contains
Geometry of TDOA maps
26
T0± , T1± , T2± , R0 , R∗ , R10 , R1∗ , 0, the only possible splitting of C is E 0 ∪ L, with L the
line through R∗ , R1∗ , 0, T0± , and E 0 the conic through T2± , T1± , R0 , R10 . The point T0+
is collinear with 0 and R∗ if, and only if, there exists t ∈ R such that
hd10 , d̃21 i, hd20 , d̃21 i = t(−d10 , d20 ),
therefore t = hd̃20 , d̃21 i and (d20 −d10 )(hd10 , d20 i+d10 d20 ) = 0. Since m0 , m1 , m2 are
not collinear, the second factor is non–zero and C is singular if, and only if, d10 = d20 .
The equations of L and E 0 are then straightforward.
Proof of Proposition 6.12. If C is not smooth, any cubic curve containing the given
points contains the line L, according to Bézout’s Theorem. The remaining points lie
on a unique conic E 0 , therefore C = E 0 ∪ L is unique.
Let us now assume that C is smooth. We embed the τ –plane into P2C , and
let X = {T0± , T1± , T2± } and Y = {0, R0 , R10 , R∗ , R1∗ }. The defining ideal IX of
X is generated by ah (τ ), b̄h (τ ) obtained by homogenizing a(τ ) and b̄(τ ), because
X = E ∩ C, as proven earlier (see Theorem B.30). The ideal IY of Y is generated by
a degree 2 and two degree 3 homogeneous polynomials (see Theorem B.30). Let L1
be the line through R0 , R10 , 0, L2 be the line through R∗ , R1∗ , and let Q = L1 ∪ L2 . Q
is a reducible conic that is singular at 0. With abuse of notation, we also call Q the
defining polynomial of Q and so Q ∈ IY by Definition B.27. Moreover, Y ⊂ C, and so
b̄h ∈ IY , as well. Finally, let C 0 be a further cubic curve, whose defining polynomial
is equal to C 0 , with abuse of notation, so that IY = hQ, b̄h , C 0 i.
Claim: C 0 is not a combination of Q, ah , and b̄h .
If we assume the contrary, we have C 0 = q1 ah +q2 Q+q3 b̄h with deg(q1 ) = deg(q2 ) = 1,
and deg(q3 ) = 0. Consequently, we have q1 (p) = 0 for every p ∈ Y, because ah (p) 6= 0
for each p ∈ Y . So, q1 ∈ IY therefore q1 = 0 because IY does not contain linear
forms. Then, C 0 is a combination of Q and b̄h , but this is not possible because C 0 is
a minimal generator of IY , therefore the claim holds true.
Hence, IX + IY is minimally generated by ah , Q, b̄h , C 0 . Moreover, since two conics
meet at four points, (C[τ0 , τ1 , τ2 ]/(ah , Q))3 has dimension 4, and we obtain
C[τ0 , τ1 , τ2 ]
= 10 − 8 = 2.
dimC
IX + IY
3
From the exactness of the short sequence of vector spaces (B.2)
C[τ0 , τ1 , τ2 ]
C[τ0 , τ1 , τ2 ]
C[τ0 , τ1 , τ2 ]
C[τ0 , τ1 , τ2 ]
→
⊕
→
→0
0→
IX ∩ IY
IX
IY
IX + IY
3
3
3
3
we conclude that the dimension of the first item is 6 + 5 − 2 = 9 and, finally,
2]
dimC (IX ∩ IY )3 = dimC (C[τ0 , τ1 , τ2 ])3 − dimC C[τIX0 ,τ∩I1 ,τ
= 10 − 9 = 1.
Y
3
P2R .
Proof of Proposition 6.14. We embed the τ –plane into
The oval Co meets all the
lines of the projective plane either in 1 or in 3 points, up to count the points with their
intersection multiplicity, as discussed after Harnack’s Theorem B.31 in Appendix B.
This implies that Co contains the inflectional point 0 of C. Moreover, from the proof
of Proposition 6.10, the lines supporting F0± meet C at 1 point each, thus T0± ∈ Co .
The possible second oval Ce meets every line at an even number of points (0 is
allowed) and it cannot meet Co . By contradiction, let us assume that R∗ ∈ Ce . Hence,
Geometry of TDOA maps
27
the line F2− meets Ce either at T2− or R10 . This implies that Ce meets the line F0+ ,
which is a contradiction. We conclude that R∗ and, symmetrically, R1∗ lie on Co .
Again by contradiction, we assume that R0 ∈ Ce . By looking at the intersection
points of Ce with the sides of P2 , we obtain T1+ , T2+ ∈ Ce and, symmetrically,
R10 , T1− , T2− ∈ Ce . We also observe that Ce meets the tangent line to C at R0
exclusively at the point R0 itself, and the same holds true at R10 . As Ce does not
meet F0± , Ce is constrained into the quadrangle formed by F0± and the tangent lines
to C at R0 , R10 . This quadrangle contains 0, therefore either Ce is the union of two
disjoint ovals, or Ce meets Co . Both cases are not allowed, thus R0 ∈ Co . This implies
that all the remaining points lie on Co and the first claim is proven.
We finish the proof by noting that Ce does not meet any side of P2 and, on the
other hand, Ce cannot be contained in P2 .
Proof of Theorem 6.19. If τ ∈ E, its preimage is given by (21):
x(τ ) = L0 (τ ) −
c(τ )
∗ ((τ2 d10 − τ1 d20 ) ∧ e3 ).
2b(τ )
Hence, x(τ ) gives a point in the x–plane both if τ ∈ E ∩ Im(τ2 ) and if τ ∈
E \ (Im(τ2 ) ∪ C). Moreover, because of the symmetry properties of the polynomials
and vectors involved, we have x(−τ ) = x(τ ), which means that x(τ ) is a 2–to–1 map
from E to Ẽ.
In order to obtain a parametrization of Ẽ, we consider a parametrization of E
via the pencil of lines through 0. Let τ1 = µ1 t, τ2 = µ2 t be a line through 0 in the
τ –plane, with µ = (µ1 , µ2 ) ∈ R2 \ {(0, 0)}. The line intersects the ellipse E at the two
points t = ±kd10 ∧ d20 k/kµ2 d10 − µ1 d20 k, which are symmetrical with respect to
0. Let P0 (µ) = kµ2 d10 − µ1 d20 k2 . This is a degree–2 homogeneous polynomial that
vanishes at the ideal points of E, therefore it is irreducible over R. By substituting
τ = kd√10 ∧d20 k µ, all the functions depend on µ, therefore we obtain
P0 (µ)
kD10 (µ)k2 =
1
hµ1 d20 − µ2 d10 , d10 i2 ,
P0 (µ)
kD20 (µ)k2 =
1
hµ1 d20 − µ2 d10 , d20 i2
P0 (µ)
which are both ratios of degree–2 homogeneous polynomials. For our convenience, we
set P1 (µ) = hµ1 d20 − µ2 d10 , d10 i2 and P2 (µ) = hµ1 d20 − µ2 d10 , d20 i2 . As τ depends
on µ, the polynomials b(τ ), c(τ ) can be computed as depending on µ, obtaining
c(µ) =
1
kP1 (µ)d20 − P2 (µ)d10 k2 ,
4kd10 ∧ d20 k2 P0 (µ)2
1
b(µ) = p
hµ2 d10 − µ1 d20 , P2 (µ)d10 − P1 (µ)d20 i =
2 P0 (µ)3
1
= p
hµ2 d10 − µ1 d20 , d10 ihµ2 d10 − µ1 d20 , d20 ihµ2 d10 − µ1 d20 , d21 i.
2 P0 (µ)3
Moreover,
∗((τ2 d10 − τ1 d20 ) ∧ e3 ) =
kd10 ∧ d20 k
p
∗ ((µ2 d10 − µ1 d20 ) ∧ e3 )
P0 (µ)
Geometry of TDOA maps
28
and
D0 (L0 (µ)) = −
1
∗ ((P1 (µ)d20 − P2 (µ)d10 ) ∧ e3 ) .
2kd10 ∧ d20 kP0 (µ)
It follows that D0 (x(µ)) is a ratio of two degree–5 homogeneous polynomials.
The denominator is, up to a non zero scalar, P0 (µ)hµ2 d10 − µ1 d20 , d10 ihµ2 d10 −
µ1 d20 , d20 ihµ2 d10 − µ1 d20 , d21 i. It is easy to check that, if µ is such that hµ2 d10 −
µ1 d20 , dij i = 0, for 0 ≤ j < i ≤ 2, then c(µ) 6= 0 because c is non–zero on E, and
∗ ((µ2 d10 − µ1 d20 ) ∧ e3 ) does not vanish. Hence, the numerator does not vanish at
the given µ. We remark that they give exactly the ideal points of the lines r0 , r1 , r2 .
The ideal points of E are the roots of P0 (µ), i.e. µ1 = (d10 e−iθ , d20 eiθ ) and µ2 =
(d10 eiθ , d20 e−iθ ) (same notation of Theorem 6.4). Here we analyze D0 (x(µ1 )), being
the other case analogous. After tedious, though fairly straightforward computations,
the numerators of the coefficients of ∗(d10 ∧ e3 ) and ∗(d20 ∧ e3 ) turn out to be
−d610 d720 eiθ (eiθ d10 − e−iθ d20 )2 sin4 (2θ),
d710 d620 e−iθ (eiθ d10 − e−iθ d20 )2 sin4 (2θ).
Without loss of generality, we choose a reference system where
d10 = d10 (cos θ e1 + sin θ e2 )
∗(d10 ∧ e3 ) = d10 (− sin θ e1 + cos θ e2 )
⇒
.
d20 = d20 (cos θ e1 − sin θ e2 )
∗(d20 ∧ e3 ) = d20 (sin θ e1 + cos θ e2 )
Therefore, x(µ1 ) is the ideal point
d710 d720 (eiθ d10 − e−iθ d20 )2 sin5 (2θ)(1 : i : 0).
It is simple to prove that the coefficient cannot vanish for a value of θ ∈ (0, π/2).
We conclude that x(µ1 ) (and similarly x(µ2 )) is a cyclic point of P2C . Furthermore,
the parametric representation of Ẽ is given by ratios of degree–5 polynomials without
common factors, and the claim follows.
Proof of Proposition 6.21. The closure Ũ0 of Ũ0 contains m0 , r1+ ∪ r2+ , and the arc
of Ẽ inverse image of the arc of E ∩ Im(τ2 ) with endpoints T1+ , T2+ . Furthermore,
Ũ0 ∪ r1+ ∪ r2+ ∪ m0 is connected because equal to an oval of Ẽ intersected with the
Euclidean x–plane, but Ẽ ∩ (r1+ ∪ r2+ ∪ m0 ) is the empty set, because their images in
the τ –plane do not meet. Hence, Ũ0 has two connected components, and τ2 : Ũ0 → U0
is a cover.
Let us now assume that the two inverse images of τ0 ∈ U0 belong to the same
connected component of Ũ0 . As U0 is path–connected, from the Path Lifting Theorem
(see [36]), it follows that the inverse images of any other point τ ∈ U0 belong to the
same connected component of Ũ0 as well. Let x0 be a point in the other connected
component of Ũ0 , with τ 0 = τ2 (x0 ). Hence, τ 0 has three inverse images, contradicting
Theorem 6.16. Thus, τ2 is 1–to–1 on each connected component of Ũ0 , as claimed.
7. The localization problem for special configurations
In this Section we study the behaviour of the TDOA map τ2 , particularly of its
image, under the hypothesis that m0 , m1 , and m2 lie on a line r. This is equivalent
to assuming that d20 = kd10 for some k ∈ R, k 6= 0, 1. If k < 0, then m0 lies between
Geometry of TDOA maps
29
m1 and m2 , if 0 < k < 1, then m2 lies between m0 and m1 , and finally, if k > 1,
then m1 lies between m0 and m2 . As discussed in Section 5, in this configuration,
the polygon P2 has only four sides.
Let us first consider the case in which D10 (τ ) and D20 (τ ) are linearly dependent.
Lemma 7.1 The vectors D10 (τ ) and D20 (τ ) are linearly dependent if, and only if,
d10 τ2 − sgn(k)d20 τ1 = τ2 − kτ1 = 0.
Proof. By definition we have Di0 (τ ) = di0 + τi e3 . Under the assumption d20 = kd10
of this Section, D10 (τ ) and D20 (τ ) are linearly dependent if, and only if, τ2 = kτ1
or, equivalently, d10 τ2 − sgn(k)d20 τ1 = 0, as claimed.
The line d10 τ2 − sgn(k)d20 τ1 = τ2 − kτ1 = 0 contains the origin 0 of the τ –plane, and
two opposite vertices of P2 : if k > 0, then it contains (d10 , d20 ), while, if k < 0, it
contains (−d10 , d20 ).
Proposition 7.2 Assume D10 (τ ) and D20 (τ ) are linearly dependent. Then, either
d10 6= ±τ1 , and the intersection of the planes Π1 (τ ) and Π2 (τ ) is empty, or d10 = ±τ1 ,
and Π1 (τ ) = Π2 (τ ) 3 M0 .
Proof. By assumption, we have τ2 = kτ1 , with d20 = kd10 , k 6= 0, 1. As a
consequence D20 (τ ) = kD10 (τ ), therefore both D20 (τ )[ = kD10 (τ )[ and kD20 (τ )k2
= k 2 kD10 (τ )k2 . Let X ∈ Π1 (τ ) ∩ Π2 (τ ). From equation (13) it follows that
1
1 2
k kD10 (τ )k2 = kD20 (τ )k2 = iD0 (X) (D20 (τ )[ ) = hD0 (X), D20 (τ )i =
2
2
1
k kD10 (τ )k2
2
Hence, either k 2 = k, which is not allowed because k 6= 0, 1, or kD10 (τ )k2 = 0. The
second condition implies d10 = ±τ1 and Π1 (τ ) 3 M0 , which completes the proof.
Proposition 7.2 implies that the points τ on the line τ2 − kτ1 = 0, with τ1 6= ±d10 ,
are not in Im(τ2 ). Furthermore, with notation of Definition 3.1, we have
= k hD0 (X), D10 (τ )i = k iD0 (X) (D10 (τ )[ ) =
Proposition 7.3 τ2 (x) = (d10 , sgn(k)d20 ) if, and only if, x ∈ rc , and hd0 (x), d10 i <
0, while τ2 (x) = (−d10 , −sgn(k)d20 ) if, and only if, x ∈ rc , and hd0 (x), d10 i > 0.
Proof. τ2 (x) = ±(d10 , sgn(k)d20 ) if, and only if, x ∈ rc . Moreover, given x ∈ rc ,
τ1 (x) = d10 is equivalent to m0 lying between m1 ad x.
Now, we assume that τ does not belong to the line τ2 − kτ1 = 0.
Lemma 7.4 Assume that D10 (τ ) and D20 (τ ) are linearly independent. Then, the
parametric equation of the line L21 (τ ) = Π1 (τ ) ∩ Π2 (τ ) is L0 (τ ) + λv(τ ), where
v(τ ) = ∗(d10 ∧ e3 ),
D0 (L0 (τ )) = −
1
∗ v(τ ) ∧ kD20 (τ )k2 D10 (τ ) − kD10 (τ )k2 D20 (τ ) .
2d210 (kτ1 − τ2 )
Proof. We use the same reasoning as in Lemma 6.1.
The line L21 (τ ) is parallel to the x–plane, because hv(τ ), e3 i = 0, thus it is not
possible for it to intersect both half–cones C0+ , C0− . As for the general case, the line
L21 (τ ) intersects the cone C0 if and only if
kv(τ )k2 λ2 + 2hv(τ ), D0 (L0 (τ ))iλ + kD0 (L0 (τ ))k2 = 0.
(22)
Geometry of TDOA maps
30
In this case, kv(τ )k2 = −hd10 ∧ e3 , d10 ∧ e3 i = d210 > 0, hv(τ ), D0 (L0 (τ ))i = 0, and
kD0 (L0 (τ ))k2 = −
=−
(d210 − τ12 )(d220 − τ22 )(d221 − (τ1 − τ2 )2 )
=
4d210 (kτ1 − τ2 )2
kD10 (τ )k2 kD20 (τ )k2 kD21 (τ )k2
.
4d210 (kτ1 − τ2 )2
As a consequence, the line L21 (τ ) intersects the cone C0 if, and only if, c(τ ) ≤ 0.
Moreover, the two intersections belong to C0− if, and only if, hD0 (L0 (τ )), e3 i > 0,
which means that
kτ12 − τ22 + d210 (k 2 − k)
> 0.
2(τ2 − kτ1 )
Now, we are able to describe the image of τ2 . The results of the next theorem are
illustrated in Fig. 11, in the subcase with k < 0, i.e. m0 between m1 and m2 (the
other two subcases are similar).
Theorem 7.5 Let us assume that d20 = kd10 , k 6= 0, 1 and let Ri be the image
of the point mi in the interior of r0 . Then, the image of τ2 is the triangle T
with vertices (d10 , sgn(k)d20 ), (−d10 , −sgn(k)d20 ), Ri minus the open segment with
endpoints (d10 , sgn(k)d20 ), (−d10 , −sgn(k)d20 ). Moreover, given τ ∈ Im(τ2 ), we have
∞ if τ = ±(d10 , sgn(k)d20 ),
|τ2 −1 (τ )| =
2
if τ ∈ T̊ ,
1
otherwise.
Proof. The case τ = ±(d10 , sgn(k)d20 ) has already been studied, as well as the
case τ on the line through them. Let us assume that τ does not lie on the line
d10 τ2 − sgn(k)d20 τ1 = 0. Eq. (22) has two real distinct roots if, and only if, c(τ ) < 0.
Quite clearly this happens if, and only if, τ ∈ P˚2 . Furthermore, the same equation has
a multiplicity–two root if, and only if, c(τ ) = 0, i.e. τ ∈ ∂P2 . Finally, the intersection
kτ 2 −τ22 +d210 (k2 −k)
points of L21 (τ ) and C0 are in C0− if, and only if, 1 2(τ
> 0.
2 −kτ1 )
2
2
2
2
0
The equation e(τ ) = kτ1 −τ2 +d10 (k −k) = 0 defines a conic C through the four
points (±d10 , ±d20 ). If k < 0, C 0 is an ellipse with real points, and so P2 is inscribed
into C 0 . Moreover, e(0) > 0, and so e(τ ) > 0 for each τ ∈ P2 except the four points
(±d10 , ±d20 ). If k > 0, C 0 is a hyperbola. The tangent line to C 0 at R0 = (d10 , d20 )
is F0+ if k > 1, (F0− if 0 < k < 1, respectively), while the tangent line to C 0 at R2
is F0− if k > 1 (F0+ if 0 < k < 1, respectively). Finally, if 0 < k < 1 then R0 and
(d10 , −d20 ) belong to the same branch of C 0 ((−d10 , d20 ) if k > 1, respectively). As a
consequence, e(τ ) does not change sign in P2 . More precisely, e(τ ) has the same sign
as k 2 − k for each τ ∈ P2 , except for τ = ±(d10 , d20 ), where it vanishes.
On the other hand, after a rather strightforward computation we find that the
linear polynomial τ2 − kτ1 has the same sign as k 2 − k at the vertex Ri , therefore the
kτ 2 −τ22 +d210 (k2 −k)
ratio 1 2(τ
is positive at any point in the interior of the triangle T . This
2 −kτ1 )
proves that each point τ in T̊ has two distinct preimages.
Finally, for τ on the two remaining sides of T , eq. (22) has only one root of
multiplicity 2, which implies |τ2 −1 (τ )| = 1.
The preimages of τ ∈ Im(τ2 ) in the x–plane are
x(τ ) = π(L0 (τ )) + λ v(τ ),
Geometry of TDOA maps
31
Figure 11. The image of τ2 under the assumption that m0 lies on the segment
between m1 and m2 . In the gray region T the map τ2 is 2–to–1. Along the
horizontal and vertical sides of T the map is 1–to–1, with the exception of the
vertices R1 , R2 , where the fibers of τ2 are not finite. Finally, the dashed side of
T is not in Im(τ2 ).
where λ = ±kD10 (τ )k kD20 (τ )k kD21 (τ )k/2d210 |kτ1 − τ2 | and π is the projection
onto the x–plane. Moreover, we have
D0 (π(L0 (τ ))) =
kD20 (τ )k2 τ1 − kD10 (τ )k2 kτ2
d10 .
2d210 (kτ1 − τ2 )
In order to interpret the results, we notice that in the aligned configuration, the
foci of A1 (τ ), A2 (τ ) belong to the line r, therefore the two level sets A1 (τ ), A2 (τ ) are
both symmetrical with respect to r. We are in the 1–to–1 situation if, and only if,
the source x belongs to r0 , corresponding to A1 (τ ), A2 (τ ) tangentially intersecting at
x. In the general case, when τ ∈ T̊ , the level sets meet at two distinct symmetrical
points. This agrees with the classical statement that it is not possible to distinguish
between symmetric configuration of the source, with respect to r, using a linear array
of receivers.
The degenerate situation occurs for x ∈ rc , dual to τ equal both to
(d10 , sgn(k)d20 ) and (−d10 , −sgn(k)d20 ). In this case the localization of the source
is totally unavailable, because the preimages contain infinitely many points of the
x–plane. Finally, the points on the interior of the side τ2 − kτ1 = 0 correspond to
A1 (τ ), A2 (τ ) with parallel asymptotic lines and empty intersection.
8. The image of the complete TDOA map
In Section 2 we explained that the relation between τ2 and τ2∗ is given by the projection
p3 from the plane H ⊂ R3 to R2 via the equality τ2 = p3 ◦ τ2∗ . As p3 is invertible, it
holds that τ2∗ = p−1
3 ◦ τ2 and consequently we have the following result:
∗
∗ −1
(τ ∗ ) =
Theorem 8.1 Im(τ2∗ )=p−1
3 (Im(τ2 )). More precisely, let τ ∈ H, then τ2
−1
∗
τ2 (τ ), where τ = p3 (τ ).
Theorem 8.1 allows us to give the explicit description of Im(τ2∗ ), thus reaching one of
the main objectives we set ourselves in Section 2. We start by defining the relevant
subsets of H.
Geometry of TDOA maps
32
Definition 8.2 Assuming 0 ≤ i, j, k ≤ 2 distinct, in the τ ∗ –space we set:
• P2 = {τ ∗ ∈ H | kDji (τ ∗ )k2 ≥ 0 for every i, j};
• Fk+ = {τ ∗ ∈ P2 | kDji (τ ∗ )k2 = 0, hDji (τ ∗ ), e3 i < 0};
• Fk− = {τ ∗ ∈ P2 | kDji (τ ∗ )k2 = 0, hDji (τ ∗ ), e3 i > 0};
• Ek = {τ ∗ ∈ P2 | kDik (τ ∗ ) ∧ Djk (τ ∗ )k2 = 0}.
As the above definitions are stated using the exterior algebra formalism, for
completeness we observe that H can also be described in similar terms:
H = {τ ∗ ∈ R3 | D10 (τ ∗ ) − D20 (τ ∗ ) + D21 (τ ∗ ) = 0}
In Fig. 12 we show an example of Im(τ2∗ ) along with its projection Im(τ2 ).
Figure 12. The image of τ2∗ is the subset of P2 in green, while the image of
τ2 is the subset of P2 in red. There is a 1–to–1 correspondence between Im(τ2∗ )
and Im(τ2 ) via the projection map p3 . In the lightly shaded regions, the TDOA
maps are 1–to–1, while in the more darkly shaded regions the maps are 2–to–1.
As explained in Section 10, three noisy TDOAs define a point τ ∗ outside P2 . The
Maximum Likelihood Estimator computes the projection τ̄ = pH (τ ∗ ) on P2 , then
the estimated source position is computed as x̄ = τ2∗ −1 (τ̄ ) = τ2 −1 (p3 (τ̄ )).
As a consequence of Theorem 8.1, the structure of Im(τ2∗ ) turns out to be similar
to that of Im(τ2 ), thus we can omit the proofs and go over the main facts about τ2∗ .
• τ2∗ is a local diffeomorphism between the x–plane and H, with the exception of
the degeneracy locus ∪2i=0 (ri− ∪ ri+ ), as described in Theorem 3.2.
±
−1
±
• P2 is the convex polygon given by p−1
3 (P2 ), whose facets are Fk = p3 (Fk ).
∗
The image of τ2 is a proper subset of P2 and, in particular, the image of the
degeneracy locus is a subset of the facets.
Geometry of TDOA maps
33
• Ek does not depend on k. If the points m0 , m1 and m2 are not aligned, then
we have Ek = p−1
3 (E). Therefore, Ek is the unique ellipse that is tangent to each
facet of the hexagon P2 . The cardinality of each fiber of τ2∗ is equal to that of
the corresponding fiber of τ2 , as described in Theorem 6.16 and in Proposition
6.21.
• If the points m0 , m1 , m2 are aligned, then Ek is one of the diagonals of the
quadrangle P2 . The cardinality of each fiber of τ2∗ is equal to that of the
corresponding fiber of τ2 , as described in Theorem 7.5.
Remark 8.3 In the definition of τ2∗ we notice a natural symmetry among the points
m0 , m1 and m2 , which is lost in τ2 as we elected m0 to be the reference microphone.
As noticed in Section 2, by taking p1 ◦ τ ∗2 or p2 ◦ τ ∗2 we define different TDOA maps,
with different reference microphones. Quite obviously, their properties are similar to
those of τ2 studied in Sections 6 and 7, in fact p1 ◦ p−1
and p2 ◦ p−1
are invertible
3
3
maps between the images of the TDOA maps, factorizing on Im(τ2 ∗ ). Although such
maps are, in fact, equivalent, some of their properties could be more or less difficult
to check depending on the chosen reference point. For example, the lines F0± become
parallel to the reference axes when applying pi ◦ p−1
3 for i = 1 or 2.
Remark 8.4 The previous remark implies that pi ◦ p−1
3 sends the ellipse E onto the
ellipse associated to the TDOA map pi ◦ τ ∗2 , but this does not happen for the cubic
curve C. In fact, both the cubics associated to pi ◦ τ ∗2 , i = 1, 2 do not contain
(the transformations of) 4 of the 11 points characterizing C in Proposition 6.12:
R0 , R∗ , R10 , R1∗ . This, however, is not an issue for localization purposes, as the image
of any TDOA map only depends on C ∩ E = E ∩ P2 .
9. Impact assessment
As anticipated in the Introduction, a complete characterization of the TDOA map
constitutes an important building block for tackling a wide range of more general and
challenging problems. For example, we could optimize sensor placement in terms of
robustness against noise or measuring errors. More generally, we could embark into
a statistical analysis of error propagation or consider more complex scenarios where
the uncertainty lies with sensor synchronization or spatial sensor placement. While
these general scenarios will be the topic of future contributions, in this Section we can
already show an example of how to jointly use local and global analysis to shed light
on the uncertainty in localisation problems.
A possible approach to the study of the accuracy of localization is based on the
linearization of the TDOA model (see [20, 45]). Usually this analysis is pursued in
a statistical context, but it essentially involves the analysis of the Jacobian matrix
J(x) of τ2 and its determinant det(J(x)). In the differential geometry interpretation,
the absolute value of Jacobian determinant is the ratio between the areas of two
corresponding infinitesimal regions in the τ –plane and in the x–plane, under the action
of the map τ2 . As a consequence, the quality of the localization is best in the regions
of maximum of |det(J(x))|, where the TDOAs are highly sensitive to differences of
source position. This local analysis is equivalent, up to a costant factor, to that of
the map τ2∗ . Starting from expression (10), in Fig. 13 we display the level sets of
|det(J(x))| along with the geometric configuration of sensors and with the curves that
we displayed earlier in Fig. 10. Fig. 13 shows that the local error analysis does not
Geometry of TDOA maps
34
Figure 13. Level sets of the absolute value of the Jacobian determinant
| det(J(x))| of τ2 (values are marked next to them). The picture also shows
the bifurcation curve Ẽ and the degeneracy locus det(J(x)) = 0, on the x–plane.
Given the microphones in m0 = (0, 0), m1 = (2, 0), and m2 = (2, 2), the best
accuracy of the localization is obtained in the region that lies the closest to the
center of the triangle. Notice that, although det(J(x)) is not affected by Ẽ, in the
proximity to this curve the localization fails due to the global properties of τ2 .
take count of the global aspects of the localization. In particular, | det(J(x))| becomes
quite large in the proximity of the sensors. In these areas, however, localisation is not
accurate because of their proximity to the bifurcation curve Ẽ and the overlapping to
the sets Ũi . Having access to a complete global characterisation of the TDOA map
allows us to predict this behaviour.
10. Conclusions and perspectives
In this manuscript we offered an exhaustive mathematical characterization of the
TDOA map in the planar case of three receivers. We began with defining the non–
algebraic complete TDOA map τ2∗ . We then derived a complete characterization of
both Im(τ2∗ ) ⊂ R3 and the various preimage regions in the x–plane. We found that
Im(τ2∗ ) is a bounded subset of the plane H and, in particular, we showed that the
image is contained in the convex polygon P2 . We also described the subsets of the
image in relation to the cardinality of the fibers, i.e. the loci where the TDOA map
is 1–to–1 or 2–to–1, which provided a complete analysis of the a-priori identifiability
problem. On the x–plane, we defined the degeneracy locus, where τ2∗ is not a local
diffeomorphism, and we described the sets where τ2∗ is globally invertible and those
where it is not.
We conducted our analysis using various mathematical tools, including multilinear
algebra, the Minkowski space, algebraic and differential geometry. Indeed, these
tools may seem too sophisticated for a problem as “simple” as that of TDOA-based
localization. After all, this is a problem that, in the engineering literature, is commonly
Geometry of TDOA maps
35
treated as consolidated or even taken for granted. As explained in the Introduction,
however, the purpose of this work was twofold:
(i) to derive analytically and in the most general sense what was preliminarily shown
in a fully simulative fashion in [51], and to make the analysis valid for arbitrary
sensor geometries;
(ii) to offer a complete characterisation of the TDOA map, to be applied to the
solution of more general problems.
The first purpose was amply proven throughout the manuscript (Sections 2 to
8). What remains to be shown is how this analysis can pave the way to a deeper
understanding of the localization problem in more general settings, such as in the
presence of noisy measurements (propagation of uncertainty) or even in the presence
of uncertainty in the sensor calibration and/or in their synchronization. An early
discussion in this direction was offered in Section 9 where we described how errors
propagate in a three-sensor setup based on local analysis and showed that, without
a global perspective on the behaviour of the TDOA map we could be easily led to
drawing wrong conclusions.
The authors are currently working on the extension to arbitrary distributions of
sensors both in the plane and in the 3D Euclidean space, using similar techniques and
notations. In particular, the model can be encoded as well in a TDOA map τn∗ , whose
image is a real surface and a real threefold, respectively. We expect the bifurcation
locus and the 2–to–1 regions to become thinner as the number of receivers in general
position increase, although they do not immediately disappear (for example, in the
planar case and n = 3, there are still curves in the x–plane where the localization
is ambiguous). The precise description of τn∗ is needed also for the study of the
localization with partially synchronized microphones. In fact, in this scenario not all
TDOAs are available and, in our description, this is equivalent to considering some
kind of projection of Im(τn∗ ), just like the relationship between τ2∗ and τ2 , explained
in Sections 2 and 8.
In a near future we also want to pursue the study of the nonlinear statistical
model, following a similar gradual approach to the one of this manuscript. Even in
this respect, the knowledge of the noiseless measurements set Im(τn∗ ) constitutes the
starting point for any further advances on the study of the stochastic model. Roughly
speaking, a vector of measurements τ ∗ affected by errors corresponds to a point that
lies close to the set Im(τn∗ ) and the localization is a two-step procedure: we can first
estimate τ̄ ∈ Im(τn∗ ) from τ ∗ , then we evaluate the inverse map τn∗ −1 (τ̄ ).
We can give a real example of this process in the case of the complete TDOA model
defined through the map τ2∗ . In a noisy scenario (e.g. with Gaussian errors), a set of
three TDOAs gives a point τ ∗ in the three dimensional τ ∗ –space, that with probability
1 is not on the plane H. A standard approach to obtain an estimation x̄ of the
source position is through Maximum Likelihood Estimator (MLE). With respect to the
discussion of the previous paragraph, it is well known that the estimated τ̄ ∈ Im(τ2∗ )
given by MLE is the orthogonal projection of τ ∗ on the noiseless measurements set,
i.e. the projection pH (τ ∗ ) on H (see [6, 42] and Fig. 12) therefore x̄ = τ2∗ −1 (pH (τ ∗ )).
A similar reasoning applies to the more complex case of τn∗ . In particular,
any estimator has a geometrical interpretation and the relative accuracy depends
on the (non trivial) shape of Im(τn∗ ). Possible techniques to be applied come
from Information Geometry [6, 35], an approach that proved successful in similar
situations and that is based on the careful description of Im(τn∗ ). With this in mind,
Geometry of TDOA maps
36
notice that our characterization of the TDOA model in algebraic geometry terms
becomes instrumental for understanding and computing the MLE. Very recently, novel
techniques have been developed and applied to to similar situations, in cases where
scientific and engineering models are expressed as sets of real solutions to systems
of polynomial equations (see, for example, [4, 25, 35]). The somewhat surprising fact
that, although the TDOA map is not polynomial, all the loci involved in the analysis
of τ2∗ are algebraic or semi–algebraic, is a promising indicator of the effectiveness of
this approach.
Acknowledgments
The authors would like to thank E. Schlesinger, F. Belgiorno, S. Cacciatori, M. Conti
and V. Pata for the useful discussions and suggestions during the preparation of this
work and the anonymous referees for the valuable remarks on the preliminary version
of the paper.
Appendices
A. The exterior algebra formalism
In this appendix, we recall the main definitions and some useful results about the
exterior algebra of a real vector space. At the end, we analyze in full detail the two
main examples that we use in the paper, namely the 2–dimensional Euclidean case
and the 3–dimensional Minkowski one. The literature on the subject is wide and we
mention [3] among the many possible references.
Let V be a n–dimensional R–vector space. Adopting a standard notation, ∧V is
the exterior algebra of V (hence, ∧k V = 0 for each k ≥ n + 1 while ∧k V has dimension
n
k for k = 0, . . . , n). Roughly speaking, the symbol ∧ is skew–commutative, and
linear with respect to each factor. Hence, given the basis (e1 , . . . , en ) of V, the
reader can simply think at ∧k V as the R–vector space spanned by ei1 ∧ . . . ∧ eik
for 1 ≤ i1 < . . . < ik ≤ n.
We choose a non–degenerate, symmetric bilinear form b : V × V → R, and an
orthonormal basis B = (e1 , . . . , en ) with respect to b, which means
if i = j = 1, . . . , r,
1
−1 if i = j = r + 1, . . . , n,
b(ei , ej ) =
0
if i 6= j.
The couple (r, n − r) is the signature of b. By setting hu, vi = b(u, v) and k u k2 =
b(u, u), we can easily compute their expression in coordinates with respect to B, and
we have
Pn
Pn
h i=1 ui ei , j=1 vi ei i = u1 v1 + . . . + ur vr − ur+1 vr+1 − . . . − un vn
(A.1)
Pn
k i=1 ui ei k2 = u21 + . . . + u2r − u2r+1 − . . . − u2n
The inner product in ∧k V is defined by
hu1 , v1 i . . .
..
hu1 ∧ . . . ∧ uk , v1 ∧ . . . ∧ vk i = det
.
huk , v1 i . . .
hu1 , vk i
..
.
huk , vk i
(A.2)
Geometry of TDOA maps
37
and extended by linearity. For example, (e1 ∧ e2 , e1 ∧ e3 , . . . , en−1 ∧ en ) is an
orthonormal basis of ∧2 V, while (ω = e1 ∧ . . . ∧ en ) is an orthonormal basis of ∧n V
with k ω k2 = (−1)n−r .
Finally, from the choice of ω as positive basis of ∧n V, and from the fact that the
natural concatenation of a k–form and a (n − k)–form gives a n–form, one recovers the
classical Hodge ∗ operator R → ∧n V , V → ∧n−1 V, . . . , ∧n−1 V → V , and ∧n V → R,
that are all isomorphisms.
Definition A.1 Given ω ∈ ∧n V, ω 6= 0, there exists a unique linear map ∗ : ∧k V →
∧n−k V that verifies the condition
x ∧ ∗y = hx, yiω
for every x, y ∈ ∧k V.
Theorem A.2 The map ∗ : ∧k V
→ ∧n−k V satisfies both ∗ ◦ ∗ =
n−r+k(n−k)
n−r
(−1)
id∧k V and h∗x, ∗yi = (−1)
hx, yi for every x, y in ∧k V, and for
any k = 0, . . . , n.
We now consider the dual space V ∗ of V , i.e. the R-vector space of the R–
linear maps from V to R. Given the basis (e1 , . . . , en ) of V, the dual space can be
identified with the n × 1 row matrices whose entries are the values that the map takes
at ei , i = 1, . . . , n.
We use the form b to construct an isomorphism between V and V ∗ . Given u ∈ V ,
we define u[ ∈ V ∗ by setting u[ (v) = hu, vi. It is easy to prove that [ : V → V ∗ is
an isomorphism, and so B [ = (e1 [ , . . . , en [ ) is a basis of V ∗ . We want V and V ∗ to
be isometric. Therefore we choose the non–degenerate, symmetric, bilinear form b[ on
V ∗ as
b[ (u[ , v[ ) = b(u, v)
for every u[ , v[ ∈ V ∗ .
]
∗
In such a way, B [ is orthonormal with the
Pnsame signature
Pnas B. We define : V [→ V
[
[ ]
to be the inverse isomorphism of , i.e. ( i=1 ui ei ) = i=1 ui ei . We can estend and
]
to the associated exterior algebra ∧V ∗ , obtaining the isomorphisms ∧k V → ∧k V ∗
and ∧k V ∗ → ∧k V :
(u1 ∧ . . . ∧ uk )[ = u1 [ ∧ . . . ∧ uk [
and
(α1 ∧ . . . ∧ αk )] = α1 ] ∧ . . . ∧ αk ] .
As for ∧k V , we follow a similar procedure, and extend b[ to ∧k V ∗ . Finally, after
choosing ω [ as positive basis of ∧n V ∗ , we define the Hodge ∗ operator on ∧k V ∗ , as
∗(x) = ∗(x] )
[
for each x ∈ ∧k V ∗ .
As last general topic, we consider the evaluation of a k–form in ∧k V ∗ on u ∈ V.
Such operation gives rise to the linear map iu : ∧k V ∗ → ∧k−1 V ∗ defined as
iu (α1 ∧ . . . ∧ αk ) =
k
X
ci ∧ . . . ∧ αk
(−1)i−1 αi (u) α1 ∧ . . . ∧ α
i=1
ci means that the item is missing.
where α
(A.3)
Geometry of TDOA maps
38
A.1. The Euclidean vector space of dimension 2
Let V be a 2–dimensional vector space on R, let b a non–degenerate bilinear form with
signature (2, 0), and let B = (e1 , e2 ) be an orthonormal basis with respect to b. Then,
∧k V = 0 for k ≥ 3, and ∧2 V has dimension 1 with (ω = e1 ∧ e2 ) as orthonormal
basis. On the natural bases, the Hodge operator is defined as:
∗(1) = ω, ∗(e1 ) = e2 , ∗(e2 ) = −e1 , ∗(ω) = 1.
Analogously, V ∗ has dimension 2, with basis (e1 [ , e2 [ ), where
ei [ (u1 e1 + u2 e2 ) = b(ei , u1 e1 + u2 e2 ) = ui
for i = 1, 2,
and
∗(1) = ω [ , ∗(e1 [ ) = e2 [ , ∗(e2 [ ) = −e1 [ , ∗(ω [ ) = 1.
Proposition A.3 Let u
u2 e2 [ , v[ = v1 e1 [ + v2 e2 [
u1
∗(u ∧ v) = det
u2
= u1 e1 + u2 e2 , v = v1 e1 + v2 e2 ∈ V and u[ = u1 e1 [ +
∈ V ∗ . Then,
v1
u1 u2
[
[
and
∗ (u ∧ v ) = det
.
v2
v 1 v2
We adopt the usual convention that the components of a vector in V are written
as columns, while the components of a vector in V ∗ are written as rows. Of course,
the images of the two 2–form are equal because of the properties of the determinant
of a matrix. The proof is an easy computation and we do not write the details.
A.2. The Minkowski vector space of dimension 3
Let V be a 3–dimensional vector space on R, let b a non–degenerate bilinear form with
signature (2, 1), and let B = (e1 , e2 , e3 ) be an orthonormal basis with respect to b.
Then, ∧k V = 0 for k ≥ 4, and ∧2 V has dimension 3 with (e1 ∧ e2 , e1 ∧ e3 , e2 ∧ e3 )
as orthonormal basis with signature (1, 2), while ∧3 V has dimension 1 and (ω =
e1 ∧ e2 ∧ e3 ) is an orthonormal basis of this last space. On the natural bases the ∗
operator acts as follows:
∗(1) = ω, ∗(ω) = −1,
∗(e1 ) = e2 ∧ e3 , ∗(e2 ) = −e1 ∧ e3 , ∗(e3 ) = −e1 ∧ e2 ,
∗(e1 ∧ e2 ) = e3 , ∗(e1 ∧ e3 ) = e2 , ∗(e2 ∧ e3 ) = −e1 .
As before, we compute the images of the elements of the bases of ∧k V ∗ via the Hodge
∗ operator, and we get:
∗(1) = ω [ , ∗(ω [ ) = −1,
∗(e1 [ ) = e2 [ ∧ e3 [ , ∗(e2 [ ) = −e1 [ ∧ e3 [ , ∗(e3 [ ) = −e1 [ ∧ e2 [ ,
∗(e1 [ ∧ e2 [ ) = e3 [ , ∗(e1 [ ∧ e3 [ ) = e2 [ , ∗(e2 [ ∧ e3 [ ) = −e1 [ .
Now we state some results that we use in the body of the paper.
Geometry of TDOA maps
39
Lemma A.4 Let u, v, w ∈ V be linearly independent, so Ω = u ∧ v ∧ w 6= 0. Then,
∗(u) = −
1
hu, wiu ∧ v + kuk2 v ∧ w + hu, viw ∧ u
∗(Ω)
and
∗ (u ∧ v) = −
1
hu ∧ v, v ∧ wiu + hu ∧ v, w ∧ uiv + ku ∧ vk2 w .
∗(Ω)
Proof. From the linear independence of u, v, and w, it follows that u ∧ v, v ∧ w, w ∧ u
is a basis of ∧2 V . Hence, there exist elements in R such that ∗(u) = au ∧ v + bv ∧
w + cw ∧ u. From the definition of ∗ follows that u ∧ ∗(u) = kuk2 ω. By substituting
the expression of ∗(u), and using the properties of ∧ we obtain b Ω = kuk2 ω or,
equivalently, b ∗ (Ω) = −kuk2 , which gives b = −kuk2 / ∗ (Ω). Through a similar
computation we can derive a and c, therefore the first claim follows. The second claim
can be proven through the same arguments, therefore we can skip the details.
Lemma A.5 Let u, v ∈ V, and γ ∈ V ∗ . Then, γ(∗(u ∧ v)) = ∗(u ∧ v ∧ γ ] ).
Proof. We can verify the equality by using components with respect to B, B [ .
∗
Corollary A.6
Let α, β ∈ V be linearly independent. Then iu (α ∧ β) = 0 if, and
]
only if, u ∈ L (∗(α ∧ β)) , where L(. . .) is the subspace generated by the vectors in
parenthesis.
]
Proof. Assume u = t (∗(α ∧ β)) , for some t ∈ R. Then
]
α(u) = tα (∗(α ∧ β)) = ∗(α] ∧ β ] ∧ α] ) = 0.
A similar argument proves that β(u) = 0. By definition, iu (α ∧ β) = α(u)β − β(u)α
therefore the claim follows.
Conversely, assume that α(u)β − β(u)α = 0. Then α(u) = β(u) = 0 because
α, β are linearly independent. Hence, hα] , ui = hβ ] , ui = 0. This implies that
]
u = t (∗(α ∧ β)) for some t ∈ R, which completes the proof.
Lemma A.7 Let Θ ∈ ∧3 V ∗ be a non–zero 3–form. Then, iu (Θ) = α ∧ β if, and only
]
1
(∗(α ∧ β)) .
if, u = ∗(Θ)
Proof. There exists t ∈ R, t 6= 0, such that Θ = tω [ , therefore,
iu (Θ) = t e1 [ (u)e2 [ ∧ e3 [ − e2 [ (u)e1 [ ∧ e3 [ + e3 [ (u)e1 [ ∧ e2 [ = α ∧ β.
This implies
∗(α ∧ β) = t −he1 , uie1 [ − he2 , uie2 [ + he3 , uie3 [ = −tu[
and u = − 1t ∗ (α ∧ β)] . Moreover, ∗(Θ) = t ∗(ω [ ) = −t, therefore one side of the claim
follows. The converse is proven through a straightforward computation.
Geometry of TDOA maps
40
B. A brief introduction to plane algebraic geometry
In this Appendix, we recall the main definitions and results concerning curves in the
affine or projective plane.
B.1. Affine spaces and algebraic subsets
Definition B.1 Let K be a field and let V be a K–vector space. Let Σ be a non–empty
set. A map φ : Σ × Σ → V that verifies
i. φ(A, B) + φ(B, C) = φ(A, C) for every A, B, C ∈ Σ
ii. φA : Σ → V defined as φA (X) = φ(A, X) is 1–to–1 for every A ∈ Σ
is an affine structure on Σ, and the couple (Σ, φ) is called affine space and named
A(V ).
The main example we use is the following: let Σ = V, and define φ(u, v) = v − u.
φ is an affine structure on V and so we get the affine space A(V ). If dim(V ) = n, we
say that A(V ) has dimension n, as well. The advantage to have an affine structure on
a set of points is that we can easily define the coordinates of the points.
Definition B.2 Let A(V ) be an affine space of dimension n. A reference frame is a
couple R = (O, B), where O is a point, and B = (v1 , . . . , vn ) is a basis of V. Given
P ∈ A(V ), its coordinates in the frame R are the components of φ(O, P ) with respect
to B.
Thanks to the properties of the affine structure, once R is given, there is a 1–to–
1 correspondence between points in A(V ) and elements in Kn . So, usually, the two
spaces are identified. When this happens, one denotes Kn as AnK . We remark that
the identification imply the choice of the reference frame, and so some care has to be
taken if one works with more than one reference frame.
If K = R, and V is an Euclidean vector space, then A(V ) is referred to as
Euclidean space, and indicated with E(V ), or En emphasizing just the dimension of V.
In this setting, the set–theoretical equality between En and AnR is evident. However,
if one switches from En to AnR , then one is not allowed to use distances and angles.
Another standard construction is the following. Given the real vector space V,
one can consider the complex vector space V̄ = C ⊗R V. Roughly speaking, we allow
complex numbers to multiply the vectors of V. As example of the previous construction,
we remark that Cn = C ⊗R Rn . It holds dimC V̄ = dimR V, and dimR V̄ = 2 dimR V.
Of corse, we have a set–theoretical inclusion V ⊂ V̄ , that is not a linear map. The
inclusion of vector spaces provides an inclusion of the corresponding affine spaces, that
can be written as AnR ⊂ AnC , up to the choice of a reference frame with the same origin
O ∈ AnR , and the same basis B ⊂ V both for V and for V̄ .
In the paper, we mainly use the affine space with n = 2, namely the affine plane.
The geometrical objects in A2K that are studied in the algebraic geometry framework
are (algebraic) curves and their intersections. We recall the definition of algebraic
curve.
Definition B.3 Let f1 , . . . , fr ∈ K[x, y] be polynomials.
V (f1 , . . . , fr ) of the given polynomials is
The vanishing locus
V (f1 , . . . , fr ) = {P ∈ A2K | fi (P ) = 0 for every i = 1, . . . , r}.
Geometry of TDOA maps
41
The evaluation of a polynomial f at P, f (P ), simply consists in substituting the
coordinates of P in the expression of f.
Definition B.4 A non–empty subset C ⊂ A2K is an algebraic curve if there exists a
polynomial f ∈ K[x, y] of degree ≥ 1 such that C = V (f ).
We get a line when the degree of f is 1, a conic when the degree of f is 2. From degree
3 on, a curve is named according to the degree of f, e.g. there are cubic curves, quartic
ones, and so on.
The advantage of considering curves in A2C is that some unpleasant phenomenon
do not happen: the vanishing locus of x2 + y 2 is a single point in A2R and a couple of
lines in A2C , the vanishing locus of x2 + y 2 + 1 is empty in A2R , and a conic in A2C .
In greater generality, we can consider algebraic subsets.
Definition B.5 A subset X ⊂ A2K is algebraic if there exist f1 , . . . , fr ∈ K[x, y] such
that X = V (f1 , . . . , fr ).
For example, the intersection of the curves Ci = V (fi ), i = 1, . . . , r, is the
algebraic set X = V (f1 , . . . , fr ). It is possible to prove that algebraic sets are the
closed sets of a topology, the Zariski topology, on A2K .
B.2. Projective spaces and algebraic subsets
Roughly speaking, the points of a projective space are the 1–dimensional subspaces
of a vector space. Hence, we have to identify all the vectors belonging to the same
subspace. The mathematical machinery is the following one.
Definition B.6 Let V be a n + 1–dimensional vector space over the ground field K.
We define the relation ∼ in V \ {0} as
u∼v
if there exists t ∈ K, t 6= 0, such that u = tv.
It is easy to check the following.
Proposition B.7 ∼ is an equivalence relation.
Definition B.8 The projective space of dimension n over V is the set of equivalence
classes of V \ {0} modulo ∼, that is to say,
P(V ) = (V \ {0})/ ∼ .
As for the affine space, we define a reference frame in the projective space.
Definition B.9 Let P(V ) be a projective space of dimension n. A reference frame is
R = (B), where B = (v0 , . . . , vn ) is a basis of V. Given P ∈ P(V ), its homogeneous
coordinates with respect to R are the components of v ∈ P with respect to B, and we
set P = (x0 : . . . : xn ).
The homogeneous coordinates of a point P ∈ P(V ) are not unique. In fact, if
v ∈ P, then P contains also tv for every t ∈ K, t 6= 0. The components of tv are the
ones of v times t, and so the homogeneous coordinates of a point are unique up to a
scalar factor, i.e. if (x0 : . . . : xn ) are the homogeneous coordinates of P, then also
(tx0 : . . . : txn ) are so, for every t 6= 0.
A first property is the following one.
Geometry of TDOA maps
42
Proposition B.10 Let V be a real vector space of dimension n + 1. Then
P2 (V ) ⊂ P2 (V̄ = C ⊗R V ).
Proof. From the definition of V̄ , it follows that vectors that are proportional in V are
proportional also in V̄ , and so there is a natural way to identify a point P ∈ P(V )
with a point in P(V̄ ). This identification gives an inclusion. We remark that the two
spaces are not equal for n ≥ 1.
Hence, we restrict to P(V ) where V is a vector space over the complex field
and we stress the properties that behave differently in a projective space over a real
vector space. Moreover, once a reference frame is given, we identify the points with
their homogeneous coordinates. In this case, we simply write PnC or PnR to stress the
dimension and the ground field. Motivated again from the case considered in the
paper, we focus on the projective space of dimension 2, i.e. on the projective plane
P2C .
In the projective setting, the polynomial to be considered are the homogeneous
ones. In fact, due to the construction of the homogeneous coordinates, the evaluation
of a polynomial at a points is in general a meaningless concept. However, it is
meaningful to check if a polynomial vanishes at a projective point P.
Proposition B.11 Let f ∈ K[x0 , x1 , x2 ] be equal to f = f0 + f1 + . . . + fd where fi
is homogeneous of degree i. Let P ∈ P2K be the point with homogeneous coordinates
(x0 : x1 : x2 ). Then, f (P ) = 0 if, and only if, fi (P ) = 0 for each i.
Proof. We have f (P ) = f (x0 : x1 : x2 ) = f0 (x0 : x1 : x2 )+f1 (x0 : x1 : x2 )+. . .+fd (x0 :
x1 : x2 ). The point P, however, is represented also from the coordinates (tx0 : tx1 : tx2 )
for every t 6= 0, and so we have f (tx0 : tx1 : tx2 ) = f0 (tx0 : tx1 : tx2 ) + f1 (tx0 : tx1 :
tx2 ) + . . . + fd (tx0 : tx1 : tx2 ) = f0 (x0 : x1 : x2 ) + tf1 (x0 : x1 : x2 ) + . . . + td fd (x0 :
x1 : x2 ). So, if K contains infinitely many elements, then fi (x0 : x1 : x2 ) = 0 for every
i = 0, . . . , d.
This proposition justifies the fact that we restrict to homogeneous polynomials.
Definition B.12 Let f1 , . . . , fr ∈ K[x0 , x1 , x2 ] be homogeneous polynomials. Then,
their vanishing locus is
V (f1 , . . . , fr ) = {P ∈ P2K | fi (P ) = 0 for every i = 1, . . . , r}.
We are now ready to define the projective algebraic sets, and projective curves in
particular.
Definition B.13 A non–empty subset C ⊂ P2K is a projective plane curve if there
exists a homogeneous polynomial f such that C = V (f ).
A subset X ⊆ P2K is a projective algebraic set if there exist homogeneous polynomials
f1 , . . . , fr such that X = V (f1 , . . . , fr ).
As in the case of the affine plane, it is possible to prove that the projective
algebraic sets are the closed sets of a topology, the Zariski topology, on the projective
plane. A line in P2K is the vanishing locus of a degree 1 homogeneous polynomial, a
conic is the vanishing locus of a degree 2 homogeneous polynomial, and so on.
As further step, we show that it is possible to include the affine plane in the
projective one, as an open subset.
Geometry of TDOA maps
43
Proposition B.14 Let H0 = V (x0 ) ⊂ P2K be a line, and let U0 be the open
complement of H0 . Then, U0 can be identified with the affine plane, and the
identification preserves the algebraic sets.
Proof. Let P = (x0 : x1 : x2 ) ∈ U0 . From the definition of U0 , it follows that x0 6= 0,
and so we can take (1 : x1 /x0 : x2 /x0 ) as the homogeneous coordinates of P. If we
set x = x1 /x0 , y = x2 /x0 , then we can identify P with the point Q ∈ A2K whose
coordinates are (x, y). Conversely, each point Q(x, y) ∈ A2K can be identified with the
point P ∈ U0 whose homogeneous coordinates are (1 : x : y). Hence, there is a 1–to–1
correspondence between U0 and A2K . To prove the remaining part of the statement,
we start considering curves. So, let C ⊂ P2K be a curve different from H0 , and let
f (x0 , x1 , x2 ) ∈ K[x0 , x1 , x2 ] be a homogeneous polynomial such that C = V (f ). Let
g(x, y) = f (1, x, y) ∈ K[x, y], and let D ⊂ A2K be the affine curve it defines. Then, the
previous correspondence maps the points in C ∩U0 to points in D, and conversely. So a
curve in the projective plane is transformed in a curve in the affine plane. Conversely,
let D = V (g) be a curve in A2K , with g ∈ K[x, y]. Let d be the degree of g, and
let f (x0 , x1 , x2 ) = xd0 g(x1 /x0 , x2 /x0 ). It is easy to check that f ∈ K[x0 , x1 , x2 ] is
a degree d homogeneous polynomial. Let C = V (f ) be the corresponding curve in
the projective plane. Then, it is straightforward to prove that the points of D are
mapped to points in C ∩ U0 . As the algebraic sets are union of finitely many curves
or intersection of curves, the proof is complete, because it holds on curves.
Notice that a consequence of the previous proof is that the complement of
whatever line in P2K is an affine plane. Conversely, the projective plane is the union of
an affine plane and a projective line. The points of the projective line are called ideal
points of the affine plane and are thought to as directions of the lines in the affine
plane.
From now on, we deal with the geometry of algebraic sets in the projective plane
P2C because we have the chains of inclusions
E2 = A2R ⊂ A2C ⊂ P2C
and
E2 = A2R ⊂ P2R ⊂ P2C ,
and we underline the problems when restricting to smaller ambient spaces.
B.3. Intersection of curves and singular points
A line L in P2C is the vanishing locus of a linear homogeneous equation a0 x0 + a1 x1 +
a2 x2 = 0, and so it has parametric equation xi = bi s + ci t, s, t ∈ C, i = 0, 1, 2.
Given a curve C = V (f ), the intersection L ∩ C is given by solving the equation
f (b0 s + c0 t, . . . , b2 s + c2 t) = F (s, t) homogeneous of the same degree of f. Hence, we
have proved
Proposition B.15 A line L intersects a degree d curve C at exactly d points, up to
count each point with its multiplicity.
Proof. The Fundamental Theorem of Algebra states that a degree d equation in one
variable has exactly d roots over C, up to count each root with its multiplicity. We
can apply it to F (1, t), after computing the largest power of s that divides F.
2
Of course, when restricting to PR , the complex roots do not give contribution,
and so a line meets a degree d curve in at most d points, up to count each one of them
with multiplicity. When considering the intersection in A2C , the roots that correspond
to ideal points give no contribution, and so again a line meets a degree d curve in at
Geometry of TDOA maps
44
most d points. When restricting to A2R , one has to take care of both problems. As an
example, we consider the conic C = V (x21 − x0 x2 ) ⊂ P2C , and the line L1 = V (x0 + x2 ).
They meet at A1 (i : 1 : −i) and B1 (−i : 1 : i) where i2 = −1. So, C ∩ L1 is empty,
when the intersection is considered in P2R . If we consider the line L2 = V (x0 − x1 ),
the intersection of C and L2 consists of the points A2 (0 : 0 : 1) and B2 (1 : 1 : 1). Let
A2C be identified with U1 complement of the line H1 = V (x1 ). Then the conic C (the
line L2 , respectively) has equation xy − 1 = 0 (x = 1, respectively). The intersection
contains the point (1, 1), only. In fact, A2 is an ideal point for the identification of the
affine plane with U1 .
The previous result can be generalized to the intersection of two curves of
arbitrary degree.
Theorem B.16 (Bézout’s Theorem) Let C, C 0 be projective plane curves of degree
d and d0 , respectively. If C and C 0 have a finite number of common points, then there
are exactly dd0 intersection points, up to count them with their multiplicity.
The proof of Bézout’s Theorem can be found in [29, Ch.I, Corollary 7.8], and goes
beyond the scopes of this introduction.
Now, we can define a smooth point on a curve.
Definition B.17 Let C ⊂ P2C be a curve and let P ∈ C be a point. P is a smooth
point of C if there exists a line L containing P that intersects C at P with multiplicity
1. A curve C whose points are all smooth is said to be smooth, singular otherwise.
The property is local, so we can reduce to an affine plane, by taking the
complement of a line that do not contain P. Moreover, we choose a reference frame
in such affine plane so that P is the origin. Hence, C = V (f ) for a suitable
f ∈ K[x, y], with f (0, 0) = 0. The lines through the origin have parametric equation
x = lt, y = mt, and the intersection between C and one of such line is described by
f (lt, mt) = 0. By McLaurin expansion, we have 0 = (lfx (0, 0) + mfy (0, 0))t+ higher
degree terms. Hence, P is smooth if there exists l, m such that lfx (0, 0)+mfy (0, 0) 6= 0,
or equivalently, the gradient ∇f (0, 0) is not zero. We have then proved
Proposition B.18 A point P ∈ C = V (f ) is smooth for C if ∇f (P ) 6= 0, where f
is a homogeneous polynomial that defines C.
Proposition B.18 allows us to prove that the singular locus Sing(C) of C = V (f )
is algebraic and it holds Sing(C) = V (f, fx0 , fx1 , fx2 ). Thanks to Euler formula for
homogeneous functions
dF (x0 , x1 , x2 ) = x0 Fx0 (x0 , x1 , x2 ) + x1 Fx1 (x0 , x1 , x2 ) + x2 Fx2 (x0 , x1 , x2 ) (B.1)
where d is the degree of F, we have that Sing(C) = V (fx0 , fx1 , fx2 ). In the affine plane,
if C = V (f ), we have Sing(C) = V (f, fx , fy ).
Also if intuition suggests that the singular points on a curve are finitely many
special points, there are examples of curves with a subcurve of singular points.
For example, consider the curve C = V (x0 x21 ) in the projective plane P2R . Then,
Sing(C) = V (2x0 x1 , x21 ) = V (x1 ). Hence, C has the line V (x1 ) as its singular locus.
The curve C is the union of the line V (x0 ) and of the conic V (x21 ). The conic, however,
is a double line (twice the line V (x1 )), and so the singular locus of C is equal to the
double line. This phenomenon can be easily generalized. Before giving the definitions
on curves to handle it, we recall some properties of polynomials. We state them for
polynomials in two variables but they can be extended without effort to homogeneous
polynomials in 3 variables.
Geometry of TDOA maps
45
Definition B.19 A polynomial f ∈ K[x, y] is irreducible if it cannot be written as
product of two non–constant polynomials.
Theorem B.20 Every polynomial f ∈ K[x, y] can be written as product of powers of
irreducible polynomials, in a unique way, up to some non–zero constant.
Now, we translate the previous results in geometrical terms.
Definition B.21 A curve C is
(i)
(ii)
(iii)
(iv)
reduced and irreducible if C = V (f ) with f irreducible;
irreducible and non–reduced if C = V (f m ) with f irreducible and m ≥ 2;
reduced, non–irreducible if C = V (f1 · · · fr ) with fi irreducible for every i;
non–reduced and non–irreducible if C = V (f1m1 · · · frmr ) with fi irreducible and
m1 + . . . + mr ≥ r + 1.
Going back to the study of the singular locus of a curve, we have the following result.
Theorem B.22 The singular points of a curve C are finitely many, or C is smooth
if and only if C is reduced. Moreover,
C isreduced and irreducible (reduced,
if
d−1
d
respectively), there are at most
(
, respectively) singular points on
2
2
C, where d is the degree of C.
For example, a reduced and irreducible conic is smooth, while a reduced non–
irreducible conic has exactly one singular point (the point where the two lines, whose
union is the conic, meet). Furthermore, a reduced and irreducible cubic can have at
most one singular point, a reduced and irreducible quartic curve can have at most 3
singular points, and so on.
A notion, apparently non related to the singular locus of a curve, is the rationality
of a curve.
Definition B.23 A curve C = V (f ) ⊂ P2C is rational if there exist
g0 (s, t), g1 (s, t), g2 (s, t) ∈ C[s, t], homogeneous of degree equal to the one of f, and
without common factors of positive degree, such that f (g0 , g1 , g2 ) is identically zero.
A rational curve is then a curve whose points have coordinates that can be
expressed via the parameter functions gi (s, t), i = 0, 1, 2. It is possible to prove that
a reduced irreducible curve is rational if it has as much singular points as its degree
allows. E.g., smooth conic, cubic with one singular point, quartic with three singular
points, quintic with 6 singular points, are all examples of rational curves. From one
hand, more than the number of singular points, the rationality depends on the kind
of singularities of the curve itself, on the other hand, a deeper study of the singular
points of a curve goes further the scope of this introduction, and so we do not go on
along these lines.
B.4. Dual projective plane
As explained earlier, a line in the projective plane is the vanishing locus of a non–zero
degree 1 homogeneous polynomial a0 x0 +a1 x1 +a2 x2 = 0. The coefficients a0 , a1 , a2 are
defined up to a scalar. In fact, for each k 6= 0, (ka0 )x0 + (ka1 )x1 + (ka2 )x2 = 0 defines
the same line as the previous equation. Hence, the coefficients can be interpreted as
points of a projective plane.
Geometry of TDOA maps
46
Definition B.24 Given the projective plane P(V ), the dual plane is defined as P(V ∗ ),
where V ∗ is the dual vector space of V. Once a reference frame R = (B) is given in
P(V ), the dual basis B ∗ defines the dual reference frame R∗ in P(V ∗ ). With this in
mind, we set P̌2K = P(V ∗ ), and the points of P̌2K are the coefficients of the lines of P2K ,
or the lines of P2K , for short.
As (V ∗ )∗ ∼
= V, the dual of the dual projective plane in the initial one. We can
now define the dual curve.
Definition B.25 Let C ⊂ P2K be a reduced and irreducible curve. The dual curve
Č ⊂ P̌2K is the unique algebraic curve that contains the tangent lines at the points of
C.
Assume that C is smooth of degree d. Then, Č has degree d(d − 1). We can assume
K = C because the degree does not depend on the ground field. A line in P̌2C is a
point in P2C . Hence, we have to compute how many tangent lines to C pass through
A
a
2
the same point A(xA
0 : x1 : x2 ) in PC . A tangent line contains A if, and only if,
A
A
A
x0 fx0 + x1 fx1 + x2 fx2 = 0. This last curve has degree d − 1 and is called the polar
curve to C with respect to the pole A. The intersection points of C and the polar
curve are d(d − 1) by Bézout Theorem, and so the degree of Č is d(d − 1), as claimed.
Proposition B.26 Let C be a smooth curve. Then, C and Č have the same degree
if and only if C is a conic.
Proof. The solutions of the equation d(d − 1) = d are d = 0 and d = 2, and d = 0
cannot be accepted.
B.5. Hilbert function and the geometry of a set of points
Definition B.27 Let X ⊂ P2K be an algebraic set. We set
IX = {f ∈ K[x0 , x1 , x2 ] | f is homogeneous and f (P ) = 0 for every P ∈ X}.
Moreover, we call
S(X) =
K[x0 , x1 , x2 ]
IX
the homogeneous coordinate ring of X. The function
K[x0 , x1 , x2 ]
HX : t ∈ Z → dimK
∈Z
IX
t
is called the Hilbert function of X.
The homogeneous elements of IX can be interpreted as the curves that vanish at all
the points of X. Hence, the homogeneous elements of the quotient ring S(X) are the
curves that do not vanish at all the points of X. Finally, if we fix the degree t, S(X)t
is a K–vector space, whose dimension is equal to dimK (K[x0 , x1 , x2 ])t − dimK (IX )t .
To illustrate the importance of the Hilbert function of an algebraic subset, we
connect it to some known results. At first, we recall without proof a general result on
the Hilbert function of a finite set of points.
Theorem B.28 Let X be a finite subset of points, eventually with multiplicities, and
assume that the sum of the multiplicities of all the points of X is d. Then,
Geometry of TDOA maps
47
(i) HX (t − 1) ≤ HX (t) for each t ≥ 1;
(ii) if HX (t) = HX (t + 1) for a suitable t, then HX (t) = HX (t + j) for every j ≥ 0;
(iii) HX (t) = d for t ≥ d − 1.
Now, we can state the announced results.
Corollary B.29 Given five distinct points, there exists at least a conic that contains
them. Moreover, it is unique if, and only if, no four points of the given five ones are
collinear.
Proof. Let X be a set of five distinct points. From Theorem B.28, it follows that
HX (t) ≤ 5 for every t, and so, in particular, HX (2) ≤ 5. Then, dimC (IX )2 =
dimC (C[x0 , x1 , x2 ])2 −HX (2) ≥ 1, and the first claim is proved. Assume now that there
are two different conics C1 , C2 through X. Then, by Bézout Theorem, the intersection
C1 ∩ C2 contains infinitely many points, and so we have C1 = L ∪ L1 , C2 = L ∪ L2 ,
with L, L1 , L2 lines, and eventually L = L1 , or L = L2 . So, at most one point in X is
L1 ∩ L2 , and then at least four ones belong to L. Conversely, if four points in X are
contained in a line L, and L1 ∩ L2 is the fifth point, then L ∪ L1 and L ∪ L2 are two
distinct conics containing X.
Similarly, it is possible to prove also results on the generators of an ideal.
Theorem B.30 Let X be a set of 6 distinct points lying on exactly one conic
C = V (f ). Then, IX = hf, gi where V (g) is a cubic curve, and f, g without common
factors.
Let X be a set of 5 distinct points, lying on exactly one conic C = V (f ). Then,
IX = hf, g1 , g2 i where g1 , g2 are homogeneous polynomials of degree 3, such that g1 , g2
are linearly independent in S(C).
It is possible to prove that IX is an ideal in K[x0 , x1 , x2 ], and it is easy to prove
that, if X and Y are algebraic sets, then IX∪Y = IX ∩ IY . Moreover, IX + IY ⊆ IX∩Y ,
and it is possible to prove that F ∈ IX + IY if, and only if, F ∈ IX∩Y under the
assumption that the degree of F is large enough. The coordinate rings of X, Y, X ∪ Y
and X ∩ Y are related each other from the short exact sequence of vector spaces
K[x0 , x1 , x2 ]
β
α
→ 0 (B.2)
0 → S(X ∪ Y )t −→ S(X)t ⊕ S(Y )t −→
IX + IY
t
where the first linear map α is defined as α(F ) = (F, F ), and the second linear map β
is defined as β(F, G) = F − G. A direct consequence of the exactness of (B.2) is that
K[x0 , x1 , x2 ]
dimK
= HX (t) + HY (t) − HX∪Y (t).
IX + IY
t
B.6. Topology of real algebraic curves
The study of the topological properties of algebraic curves in the real projective plane
in full generality is outside the scope of this appendix, so we consider only the cases
of smooth conic and cubic curves. In particular we focus on the problem of connected
components of a curve with respect to the Euclidean topology. The main result is by
Harnack, that found a bound on the number of such connected components.
Theorem B.31 (Harnack’s Theorem) Let C ⊂ P2R be a smooth algebraic curve.
If C is a conic, then either C is empty, or C is a closed connected curve. If C is a
cubic curve, then either C is connected, or C is the disjoint union of two connected
components.
Geometry of TDOA maps
48
Let H be a line in P2R , and let A2R be the complement of H. Moreover, let C ⊂ P2R
be a smooth conic with a real point, so that C is a connected curve. If H meets C at
two non–real points (H is tangent to C or H ∩ C consists of two real distinct points,
respectively), then C is an ellipse (a parabola or an hyperbola, respectively).
For cubic curves, the picture is as follows. We call oval each connected component
of C. Then, C has either one oval or two ones. In both cases, one of the two ovals, the
only one if C is connected, meets all the lines of P2R at 1 or 3 points, counted with their
multiplicities, and so it is called the odd oval Co . The second oval, if it exists, meets all
the lines at an even number of points, eventually the intersection with a line is empty,
and so it is called the even oval Ce . The oval Co , when we restrict to A2R , is either
connected and unbounded (if the ideal line meets Co at 1 point), or the union of three
unbounded arcs (if the ideal line meets Co at 3 distinct points). The even oval Ce
does not contain real inflectional points, i.e. smooth points P ∈ C with the property
that the tangent line at P to C meets C at P with multiplicity 3. So, it behaves like
conics: if the ideal line meets Ce at 2 non–real points, then Ce is topologically like an
ellipse, if the ideal line meets Ce at two real distinct points, then Ce is topologically
like a hyperbola, and finally, if the ideal line is tangent to Ce , then Ce is topologically
like a parabola (we remind that two curves behave topologically the same if the first
one can be deformed with continuity to the second one).
C. An algorithm for the bifurcation curve
This Appendix lists the source code in Singular language [23] for computing the Cartesian equation of the bifurcation curve Ẽ (see Definition 6.18 and Theorem 6.19). It
requires specifying the location of the sensors and assigning the components of the
displacement vectors d10 and d20 .
ring r=0,(x,y,z,m1,m2),dp;
LIB"linalg.lib";
matrix d1[2][1];
matrix d2[2][1];
poly p0=(m2*m2*transpose(d1)*d1-2*m1*m2*transpose(d1)*d2+m1*m1*transpose(d2)*d2)[1,1];
poly q1=(m1*transpose(d2)*d1-m2*transpose(d1)*d1)[1,1];
poly p1=q1*q1;
poly q2=(m1*transpose(d2)*d2-m2*transpose(d1)*d2)[1,1];
poly p2=q2*q2;
poly pv=det(concat(d1,d2));
poly p3=(p1*p1*transpose(d2)*d2-2*p1*p2*transpose(d1)*d2+p2*p2*transpose(d1)*d1)[1,1];
poly den=4*pv*p0*q1*q2*(q1-q2);
poly numx=2*q1*q2*(q1-q2)*(p1*d2[2,1]-p2*d1[2,1])+p3*(m2*d1[2,1]-m1*d2[2,1]);
poly numy=2*q1*q2*(q1-q2)*(-p1*d2[1,1]+p2*d1[1,1])-p3*(m2*d1[1,1]-m1*d2[1,1]);
ideal ii=x-numx, y-numy, z-den;
ideal jj=elim(ii,m1*m2);
jj=reduce(jj,std(z-1));
References
[1] J. Abel and J. Chauffe. Existence and uniqueness of GPS solutions. IEEE Transactions on
Aerospace and Electronic Systems, 27:952–956, November 1991.
Geometry of TDOA maps
49
[2] J. Abel and J. Smith. The spherical interpolation method for closed-form passive source
localization using range difference measurements. In IEEE International Conference on
Acoustics, Speech, and Signal Processing, ICASSP ’87., volume 12, pages 471 – 474, apr
1987.
[3] R. Abraham, J. Marsden, and T. Ratiu. Manifolds, tensor analysis, and applications, Second
Edition. Springer Verlag, 1988.
[4] C. Aholt, B. Sturmfels, and R. Thomas. A Hilbert Scheme in Computer Vision. Canad. J.
Math., 65(5):961–988, 2013.
[5] X. Alameda-Pineda and R. Horaud. A Geometric Approach to Sound Source Localization from
Time-Delay Estimates. 1311.1047.
[6] S. Amari and H. Nagaoka. Methods of Information Geometry. American Mathematical Society,
2000.
[7] F. Antonacci, D. Riva, D. Saiu, A. Sarti, M. Tagliasacchi, and S. Tubaro. Tracking multiple
acoustic sources using particle filtering. In In Proc. of EUROPEAN SIGNAL PROCESSING
CONFERENCE (EUSIPCO), 2006.
[8] J.L. Awange and J. Shan. Algebraic Solution of GPS Pseudo-Ranging Equations. GPS
Solutions, 5(4):20–32, 2002.
[9] S. Bancroft. An Algebraic Solution of the GPS Equations. IEEE Transactions on Aerospace
Electronic Systems, 21:56–59, January 1985.
[10] A. Beck, P. Stoica, and Jian Li. Exact and approximate solutions of source localization problems.
IEEE Transactions on Signal Processing, 56(5):1770 –1778, May 2008.
[11] R. Bellman and K. Astrom. On structural identifiability. Mathematical Biosciences, (7):329–
339, 1970.
[12] P. Bestagini, M. Compagnoni, F. Antonacci, A. Sarti, and S. Tubaro. Tdoa-based acoustic
source localization in the space–range reference frame. Multidimensional Systems and Signal
Processing, Mar. 2013.
[13] R. Bix. Conics and cubics. A concrete introduction to algebraic curves. Undergraduate Texts
in Mathematics. Springer Verlag, 1998.
[14] A. Canclini, F. Antonacci, A. Sarti, and S. Tubaro. Acoustic source localization with distributed
asynchronous microphone networks. IEEE Transactions on Audio, Speech, and Language
Processing, 21(2):439–443, 2013.
[15] J. Chauffe and J. Abel. On the exact solution of the pseudorange equations. IEEE Transactions
on Aerospace and Electronic Systems, 30:1021–1030, October 1994.
[16] J.C. Chen, R.E. Hudson, and Kung Yao. Maximum-likelihood source localization and unknown
sensor location estimation for wideband signals in the near-field. IEEE Transactions on
Signal Processing, 50(8):1843 –1854, August 2002.
[17] Y. Cheng, X. Wangb, M. Morelande, and B. Moran. Information geometry of target tracking
sensor networks. Information Fusion, 14:311–326, 2013.
[18] B. Coll, J. Ferrando, and J. Morales-Lladosa. Positioning systems in minkowski space-time: from
emission to inertial coordinates. Classical Quantum Gravity, 27:065013, 2010, 0910.2568.
[19] B. Coll, J. Ferrando, and J. Morales-Lladosa. Positioning systems in minkowski space-time:
Bifurcation problem and observational data. Phys. Rev. D, 86:084036, Oct 2012.
[20] M. Compagnoni, P. Bestagini, F. Antonacci, A. Sarti, and S. Tubaro. Localization of acoustic
sources through the fitting of propagation cones using multiple independent arrays. IEEE
Transactions on Audio, Speech, and Language Processing, 20(7):1964 –1975, sept. 2012.
[21] M. Compagnoni and R. Notari. TDOA-based localization in two dimension: the bifurcation
curve. 2014, 1402.1530.
[22] D.A. Cox, J. Little, and D. OShea. Ideals, Varieties, and Algorithms: An Introduction to
Computational Algebraic Geometry and Commutative Algebra. Springer Verlag, New York,
2007.
[23] W. Decker, G.-M. Greuel, G. Pfister, and H. Schönemann. Singular 3-1-6 — A computer
algebra system for polynomial computations. 2012. http://www.singular.uni-kl.de.
[24] Ju-Yong Do, M. Rabinowitz, and P. Enge. Robustness of TOA and TDOA positioning under
suboptimal weighting conditions. IEEE Transactions on Aerospace and Electronic Systems,
43(3):1177 –1180, Jul. 2007.
[25] J. Draisma, E. Horobet, G. Ottaviani, B. Sturmfels, and R.R. Thomas. The Euclidean distance
degree of an algebraic variety. 2013, 1309.0049.
[26] W.H. Foy. Position-location solutions by taylor-series estimation. IEEE Transactions on
Aerospace and Electronic Systems, AES-12(2):187 –194, Mar. 1976.
[27] M. Gillette and H. Silverman. A linear closed-form algorithm for source localization from timedifferences of arrival. IEEE Signal Processing Letters, 15:1–4, 2008.
Geometry of TDOA maps
50
[28] E.W. Grafarend and J. Shan. GPS Solutions: Closed Forms, Critical and Special Configurations
of P4P. GPS Solutions, 5(3):29–41, 2002.
[29] R. Hartshorne. Algebraic Geometry. Springer Verlag, New York, 1977.
[30] J. Hoshen. The GPS Equations and the Problem of Apollonius. IEEE Transactions on
Aerospace and Electronic Systems, 32(3):1116–1124, July 1996.
[31] Y. Huang and J. Benesty.
Audio Signal Processing for Next Generation Multimedia
Communication Systems. Kluwer Academic Publishers, 2004.
[32] Y. Huang, J. Benesty, and G.W. Elko. Passive acoustic source localization for video
camera steering. In 2000 IEEE International Conference on Acoustics, Speech, and Signal
Processing, 2000. ICASSP ’00. Proceedings., volume 2, pages II909 –II912 vol.2, 2000.
[33] Y. Huang, J. Benesty, G.W. Elko, and R.M. Mersereati. Real-time passive source localization: a
practical linear-correction least-squares approach. IEEE Transactions on Speech and Audio
Processing, 9(8):943 –956, November 2001.
[34] Y. Huang, J. Benesty, and G.Elko. Source Localization, chapter 9, pages 229–253. Kluwer
Academic Publishers, 2004.
[35] K. Kobayashi and H.P. Wynn. Computational algebraic methods in efficient estimation.
1310.6515.
[36] C. Kosniowski. A first course in Algebraic Topology. Cambridge University Press, Cambridge,
1980.
[37] L.O. Kraus. A direct solution to GPS-type navigation equations. IEEE Transactions on
Aerospace and Electronic Systems, AES-23(2):223–232, March 1987.
[38] J. Leva. An alternative closed form solution to the GPS pseudorange equation. In Proceedings
of the Institute of Navigation National Technical Meeting, pages 269–271, Anaheim, CA,
January 1995.
[39] J. Matousek. Lectures on discrete geometry. Springer Verlag, New York, 2012.
[40] H. Miao, X. Xia, A.S. Perelson, and H. Wu. On identifiability of nonlinear ode models and
applications in viral dynamics. SIAM Rev., (53):3–39, 2011.
[41] C. Militello and S. R. Buenafuente. An exact noniterative linear method for locating sources
based on measuring receiver arrival times. The Journal of the Acoustical Society of America,
121(6):3595–3601, 2007.
[42] A. Pázman. Nonlinear statistical models, volume 254 of Mathematics and its Applications.
Kluwer Academic Publishers Group, Dordrecht, 1993.
[43] R. Penrose. The Road to Reality: A Complete Guide to the Laws of the Universe. Vintage
Books, 2007.
[44] A. Pourmohammad and S.M. Ahadi. Tde-based 2d real time high accuracy sound source location
calculation using a special microphones arrangement. In in proc. of 2nd Internationan
Conference on Signal Processing Systems (ICSPS), volume 1, pages V1–378 –V1–382, Jul.
2010.
[45] V. Raykar, I. Kozintsev, and R. Lienhart. Position calibration of microphones and loudspeakers
in distributed computing platforms. IEEE Transactions on Speech and Audio Processing,
13(1):70–83, Jan. 2005.
[46] S.S Reddi. An exact solution to range computation with time delay information for arbitrary
array geometries. IEEE Transactions on Signal Processing, 41:485–486, 1993.
[47] A. Redondi, M. Tagliasacchi, F. Antonacci, and A. Sarti. Geometric calibration of distributed
microphone arrays. In IEEE International Workshop on Multimedia Signal Processing
(MMSP), 2009.
[48] H. Schau and A. Robinson. Passive source localization employing intersecting spherical
surfaces from time-of-arrival differences. IEEE Transactions on Acoustics, Speech and Signal
Processing, 35(8):1223 – 1225, August 1987.
[49] R.O. Schmidt. A new approach to geometry of range difference location. IEEE Transactions
on Aerospace and Electronic Systems, AES-8(6):821 –835, Nov. 1972.
[50] Hing Cheung So, Yiu Tong Chan, and F.K.W.Chan. Closed–form Formulae for Time–
Difference–of–Arrival Estimation. IEEE Transactions on Signal Processing, 56(6):2614–2620,
2007.
[51] S.J. Spencer. The two-dimensional source location problem for time differences of arrival at
minimal element monitoring arrays. J. Acoust. Soc. Am., 121(6):3579–3594, June 2007.
[52] P. Teng, A. Lombard, and W. Kellermann. Disambiguation in multidimensional tracking of
multiple acoustic sources using a gaussian likelihood criterion. In 2010 IEEE International
Conference on Acoustics Speech and Signal Processing (ICASSP), pages 145 –148, Mar. 2010.
[53] D.J. Torrieri. Statistical theory of passive location systems. IEEE Transactions on Aerospace
and Electronic Systems, AES-20(2):183 –198, Mar. 1984.
Geometry of TDOA maps
51
[54] Ying Yu and H.F. Silverman. An improved TDOA-based location estimation algorithm for large
aperture microphone arrays. In IEEE International Conference on Acoustics, Speech, and
Signal Processing, 2004. Proceedings. (ICASSP ’04)., volume 4, pages iv–77 – iv–80 vol.4,
May 2004.
[55] C.M. Zannini, A. Cirillo, R. Parisi, and A. Uncini. Improved TDOA disambiguation techniques
for sound source localization in reverberant environments. In Proceedings of 2010 IEEE
International Symposium on Circuits and Systems (ISCAS), pages 2666 –2669, 302010-june2
2010.
| 0math.AC
|
The Fourier Transform of Poisson Multinomial Distributions
and its Algorithmic Applications
arXiv:1511.03592v2 [cs.DS] 22 Jun 2016
Ilias Diakonikolas∗
University of Southern California
[email protected]
Daniel M. Kane†
University of California, San Diego
[email protected]
Alistair Stewart∗
University of Southern California
[email protected]
June 23, 2016
Abstract
We study Poisson Multinomial Distributions – a fundamental family of discrete distributions
that generalize the binomial and multinomial distributions, and are commonly encountered in
computer science. Formally,
P an (n, k)-Poisson Multinomial Distribution (PMD) is a random
variable of the form X = ni=1 Xi , where the Xi ’s are independent random vectors supported
on the set {e1 , e2 , . . . , ek } of standard basis vectors in Rk . In this paper, we obtain a refined
structural understanding of PMDs by analyzing their Fourier transform. As our core structural
result, we prove that the Fourier transform of PMDs is approximately sparse, i.e., roughly
speaking, its L1 -norm is small outside a small set. By building on this result, we obtain the
following applications:
Learning Theory. We design the first computationally efficient learning algorithm for
PMDs with respect to the total variation distance. Our algorithm learns an arbitrary (n, k)ek (1/ǫ2 ), and runs in time
PMD within variation distance ǫ using a near-optimal sample size of O
ek (1/ǫ2 ) · log n. Previously, no algorithm with a poly(1/ǫ) runtime was known, even for k = 3.
O
Game Theory. We give the first efficient polynomial-time approximation scheme (EPTAS)
for computing Nash equilibria in anonymous games. For normalized anonymous games with n
players and k strategies, our algorithm computes a well-supported ǫ-Nash equilibrium in time
3
3
k−1
nO(k ) · (k/ǫ)O(k log(k/ǫ)/ log log(k/ǫ)) . The best previous algorithm for this problem [DP08,
k
2
DP14] had running time n(f (k)/ǫ) , where f (k) = Ω(k k ), for any k > 2.
Statistics. We prove a multivariate central limit theorem (CLT) that relates an arbitrary
PMD to a discretized multivariate Gaussian with the same mean and covariance, in total variation distance. Our new CLT strengthens the CLT of Valiant and Valiant [VV10, VV11] by
completely removing the dependence on n in the error bound.
Along the way we prove several new structural results of independent interest about PMDs.
These include: (i) a robust moment-matching lemma, roughly stating that two PMDs that
approximately agree on their low-degree parameter moments are close in variation distance;
(ii) near-optimal size proper ǫ-covers for PMDs in total variation distance (constructive upper
bound and nearly-matching lower bound). In addition to Fourier analysis, we employ a number
of analytic tools, including the saddlepoint method from complex analysis, that may find other
applications.
∗
Part of this work was performed while the author was at the University of Edinburgh. Supported in part by a
Marie Curie Career Integration Grant and EPSRC grant EP/L021749/1.
†
Supported by NSF Award CCF-1553288. This work was initiated while visiting the University of Edinburgh.
1
Introduction
1.1 Background and Motivation The Poisson Multinomial Distribution (PMD) is the discrete
probability distribution of a sum of mutually independent categorical random variables over the
same sample space. A categorical random variable (k-CRV) describes the result of a random event
that takes on oneP
of k ≥ 2 possible outcomes. Formally, an (n, k)-PMD is any random variable
of the form X = ni=1 Xi , where the Xi ’s are independent random vectors supported on the set
{e1 , e2 , . . . , ek } of standard basis vectors in Rk .
PMDs comprise a broad class of discrete distributions of fundamental importance in computer
science, probability, and statistics. A large body of work in the probability and statistics literature
has been devoted to the study of the behavior of PMDs under various structural conditions [Bar88,
Loh92, BHJ92, Ben03, Roo99, Roo10]. PMDs generalize the familiar binomial and multinomial
distributions, and describe many distributions commonly encountered in computer science (see,
e.g., [DP07, DP08, Val08, VV11]). The k = 2 case corresponds to the Poisson binomial distribution
(PBD), introduced by Poisson [Poi37] as a non-trivial generalization of the binomial distribution.
Recent years have witnessed a flurry of research activity on PMDs and related distributions,
from several perspectives of theoretical computer science, including learning [DDS12, DDO+ 13,
DKS15a, DKT15, DKS15b], property testing [Val08, VV10, VV11], computational game theory [DP07, DP08, BCI+ 08, DP09, DP14, GT14], and derandomization [GMRZ11, BDS12, De15,
GKM15]. More specifically, the following questions have been of interest to the TCS community:
1. Is there a statistically and computationally efficient algorithm for learning PMDs from independent samples in total variation distance?
2. How fast can we compute approximate Nash equilibria in anonymous games with many players
and a small number of strategies per player?
3. How well can a PMD be approximated, in total variation distance, by a discretized Gaussian
with the same mean and covariance matrix?
The first question is a fundamental problem in unsupervised learning that has received considerable recent attention in TCS [DDS12, DDO+ 13, DKS15a, DKT15, DKS15b]. The aforementioned
works have studied the learnability of PMDs, and related distribution families, in particular PBDs
(i.e., (n, 2)-PMDs) and sums of independent integer random variables. Prior to this work, no
computationally efficient learning algorithm for PMDs was known, even for the case of k = 3.
The second question concerns an important class of succinct games previously studied in the
economics literature [Mil96, Blo99, Blo05], whose (exact) Nash equilibrium computation was recently shown to be intractable [CDO15]. The formal connection between computing Nash equilibria
in these games and PMDs was established in a sequence of papers by Daskalakis and Papadimitriou [DP07, DP08, DP09, DP14], who leveraged it to gave the first PTAS for the problem. Prior
to this work, no efficient PTAS was known, even for anonymous games with 3 strategies per player.
The third question refers to the design of Central Limit Theorems (CLTs) for PMDs with
respect to the total variation distance. Despite substantial amount of work in probability theory,
the first strong CLT of this form appears to have been shown by Valiant and Valiant [VV10, VV11],
motivated by applications in distribution property testing. [VV10, VV11] leveraged their CLT to
obtain tight lower bounds for several fundamental problems in property testing. We remark that
the error bound of the [VV10] CLT has a logarithmic dependence on the size n of the PMD (number
of summands), and it was conjectured in [VV10] that this dependence is unnecessary.
1
1.2 Our Results The main technical contribution of this work is the use of Fourier analytic
techniques to obtain a refined understanding of the structure of PMDs. As our core structural
result, we prove that the Fourier transform of PMDs is approximately sparse, i.e., roughly speaking,
its L1 -norm is small outside a small set. By building on this property, we are able to obtain various
new structural results about PMDs, and make progress on the three questions stated in the previous
subsection. In this subsection, we describe our algorithmic and structural contributions in detail.
We start by stating our algorithmic results in learning and computational game theory, followed
by an informal description of our structural results and the connections between them.
Distribution Learning. As our main learning result, we obtain the first statistically and computationally efficient learning algorithm for PMDs with respect to the total variation distance. In
particular, we show:
Theorem 1.1 (Efficiently Learning PMDs). For all n, k ∈ Z+ and ǫ > 0, there is an algorithm for
learning (n, k)-PMDs with the following performance
guarantee: Let P be an unknown (n, k)-PMD.
2k
4k
2
The algorithm uses m = O k log (k/ǫ)/ǫ samples from P, runs in time1 O k6k log3k (k/ǫ)/ǫ2 ·
log n, and with probability at least 9/10 outputs an ǫ-sampler for P.
We remark that our learning algorithm outputs a succinct description of its hypothesis H, via
b which is supported on a small size set. We show that
its Discrete Fourier Transform (DFT), H,
the DFT gives both an efficient ǫ-sampler and an efficient ǫ-evaluation oracle for P.
Our algorithm learns an unknown (n, k)-PMD within variation distance ǫ, with sample complexek (1/ǫ2 ), and computational complexity O
ek (1/ǫ2 )·log n. The sample complexity of our algorithm
ity O
is near-optimal for any fixed k, as Ω(k/ǫ2 ) samples are necessary, even for n = 1. We note that
recent work by Daskalakis et al. [DKT15] established a similar sample upper bound, however their
k+1
5k
algorithm is not computationally efficient. More specifically, it runs in time (1/ǫ)Ω(k log (1/ǫ)) ,
which is quasi-polynomial in 1/ǫ, even for k = 2. For the k = 2 case, in recent work [DKS15a] the
2 ). Prior to this
e
authors of this paper gave an algorithm with sample complexity and runtime O(1/ǫ
work, no algorithm with a poly(1/ǫ) sample size and runtime was known, even for k = 3.
Our learning algorithm and its analysis are described in Section 3.
Computational Game Theory. As our second algorithmic contribution, we give the first efficient polynomial-time approximation scheme (EPTAS) for computing Nash equilibria in anonymous
games with many players and a small number of strategies. In anonymous games, all players have
the same set of strategies, and the payoff of a player depends on the strategy played by the player
and the number of other players who play each of the strategies. In particular, we show:
Theorem 1.2 (EPTAS for Nash in Anonymous Games). There is an EPTAS for the mixed Nash
equilibrium problem for normalized anonymous games with a constant number of strategies. More
precisely, there exists an algorithm with the following performance guarantee: for all ǫ > 0, and
any normalized anonymous game G of n players and k strategies, the algorithm runs in time
3
3
k−1
(kn)O(k ) (1/ǫ)O(k log(k/ǫ)/ log log(k/ǫ)) , and outputs a (well-supported) ǫ-Nash equilibrium of G.
k2
6k
The previous PTAS for this problem [DP08, DP14] has running time nO(2 (f (k)/ǫ) ) , where
2
f (k) ≤ 23k−1 kk +1 k!. Our algorithm decouples the dependence on n and 1/ǫ, and, importantly,
its running time dependence on 1/ǫ is quasi-polynomial. For k = 2, an algorithm with runtime
1
We work in the standard “word RAM” model in which basic arithmetic operations on O(log n)-bit integers are
assumed to take constant time.
2
2
poly(n)(1/ǫ)O(log (1/ǫ)) was given in [DP09], which was sharpened to poly(n)(1/ǫ)O(log(1/ǫ)) in the
recent work of the authors [DKS15a]. Hence, we obtain, for any value of k, the same qualitative
runtime dependence on 1/ǫ as in the case k = 2.
Similarly to [DP08, DP14], our algorithm proceeds by constructing a proper ǫ-cover, in total
variation distance, for the space of PMDs. A proper ǫ-cover for Mn,k , the set of all (n, k)-PMDs,
is a subset C of Mn,k such that any distribution in Mn,k is within total variation distance ǫ from
some distribution in C. Our main technical contribution is the efficient construction of a proper
ǫ-cover of near-minimum size (see Theorem 1.4). We note that, as follows from Theorem 1.5, the
quasi-polynomial dependence on 1/ǫ and the doubly exponential dependence on k in the runtime
are unavoidable for any cover-based algorithm. Our cover upper and lower bounds and our Nash
approximation algorithm are given in Section 4.
Statistics. Using our Fourier-based machinery, we prove a strong “size-free” CLT relating the
total variation distance between a PMD and an appropriately discretized Gaussian with the same
mean and covariance matrix. In particular, we show:
Theorem 1.3. Let X be an (n, k)-PMD with covariance matrix Σ. Suppose that Σ has no eigenvectors other than 1 = (1, 1, . . . , 1) with eigenvalue less than σ. Then, there exists a discrete Gaussian
G so that
dTV (X, G) ≤ poly(k)/poly(σ).
As mentioned above, Valiant and Valiant [VV10, VV11] proved a CLT of this form and used
it as their main technical tool to obtain tight information-theoretic lower bounds for fundamental
statistical estimation tasks. This and related CLTs have since been used in proving lower bounds
for other problems (see, e.g., [CST14]). The error bound in the CLT of [VV10, VV11] is of the form
poly(k)/poly(σ) · (1 + log n)2/3 , i.e., it has a dependence on the size n of the underlying PMD. Our
Theorem 1.3 provides a qualitative improvement over the aforementioned bound, by establishing
that no dependence on n is necessary. We note that [VV10] conjectured that such a qualitative
improvement may be possible.
We remark that our techniques for proving Theorem 1.3 are orthogonal to those of [VV10,
VV11]. While Valiant and Valiant use Stein’s method, we prove our strengthened CLT using the
Fourier techniques that underly this paper. We view Fourier analysis as the right technical tool
to analyze sums of independent random variables. An additional ingredient that we require is the
saddlepoint method from complex analysis. We hope that our new CLT will be of broader use as
an analytic tool to the TCS community. Our CLT is proved in Section 5.
Structure of PMDs. We now provide a brief intuitive overview of our new structural results
for PMDs, the relation between them, and their connection to our algorithmic results mentioned
above. The unifying theme of our work is a refined analysis of the structure of PMDs, based on
their Fourier transform. The Fourier transform is one of the most natural technical tools to consider
for analyzing sums of independent random variables, and indeed one of the classical proofs of the
(asymptotic) central limit theorem is based on Fourier methods. The basis of our results, both
algorithmic and structural, is the following statement:
Informal Lemma (Sparsity of the Fourier Transform of PMDs.) For any (n, k)-PMD P, and any
b
ǫ > 0 there exists a “small” set T = T (P, ǫ), such that the L1 -norm of its Fourier transform, P,
outside the set T is at most ǫ.
We will need two different versions of the above statement for our applications, and therefore we
do not provide a formal statement at this stage. The precise meaning of the term “small” depends
3
on the setting: For the continuous Fourier transform, we essentially prove that the product of the
volume of the effective support of the Fourier transform times the number of points in the effective
support of our distribution is small. In particular, the set T is a scaled version of the dual ellipsoid
b has an effective
to the ellipsoid defined by the covariance matrix of P. Hence, roughly speaking, P
support that is the dual of the effective support of P. (See Lemma 4.2 in Section 4 for the precise
statement.)
In the case of the Discrete Fourier Transform (DFT), we show that there exists a discrete set
with small cardinality, such that L1 -norm of the DFT outside this set is small. At a high-level,
to prove this statement, we need the appropriate definition of the (multidimensional) DFT, which
turns out to be non-trivial, and is crucial for the computational efficiency of our learning algorithm.
More specifically, we chose the period of the DFT to reflect the shape of the effective support of
our PMD. (See Proposition 3.8 in Section 3 for the statement.)
With Fourier sparsity as our starting point, we obtain new structural results of independent
interest for PMDs. The first is a “robust” moment-matching lemma, which we now informally
state:
Informal Lemma (Parameter Moment Closeness Implies Closeness in Distribution.) For any pair
of (n, k)-PMDs P, Q, if the “low-degree” parameter moment profiles of P and Q are close, then
P, Q are close in total variation distance.
See Definition 2.2 for the definition of parameter moments of a PMD. The formal statement
of the aforementioned lemma appears as Lemma 4.6 in Section 4.1. Our robust moment-matching
lemma is the basis for our proper cover algorithm and our EPTAS for Nash equilibria in anonymous
games. Our constructive cover upper bound is the following:
Theorem 1.4 (Optimal Covers for PMDs). For all n, k ∈ Z+ , k > 2, and ǫ > 0, there exists an
ǫ-cover Mn,k,ǫ ⊆ Mn,k , under the total variation distance, of the set Mn,k of (n, k)-PMDs of size
2
k−1
|Mn,k,ǫ | ≤ nO(k ) · (1/ǫ)O(k log(k/ǫ)/ log log(k/ǫ)) . In addition, there exists an algorithm to construct
3
3
k−1
the set Mn,k,ǫ that runs in time nO(k ) · (1/ǫ)O(k log(k/ǫ)/ log log(k/ǫ)) .
A sparse proper cover quantifies the “size” of the space of PMDs and provides useful structural
information that can be exploited in a variety of applications. In addition to Nash equilibria
in anonymous games, our efficient proper cover construction provides a smaller search space for
approximately solving essentially any optimization problem over PMDs. As another corollary of our
cover construction, we obtain the first EPTAS for computing threat points in anonymous games.
Perhaps surprisingly, we also prove that our above upper bound is essentially tight:
Theorem 1.5 (Cover Lower Bound for PMDs). For any k > 2, ǫ > 0 sufficiently small as a
function of k, and n = Ωk (log(1/ǫ)/ log log(1/ǫ))k−1 , any ǫ-cover for Mn,k has size at least nΩ(k) ·
k−1
(1/ǫ)Ωk (log(1/ǫ)/ log log(1/ǫ)) .
We remark that, in previous work [DKS15a], the authors proved a tight cover size bound of
n · (1/ǫ)Θ(k log(1/ǫ)) for (n, k)-SIIRVs, i.e., sums of n independent scalar random variables each
supported on [k]. While a cover size lower bound for (n, k)-SIIRVs directly implies the same lower
bound for (n, k)-PMDs, the opposite is not true. Indeed, Theorems 1.4 and 1.5 show that covers
for (n, k)-PMDs are inherently larger, requiring a doubly exponential dependence on k.
1.3 Our Approach and Techniques At a high-level, the Fourier techniques of this paper can
be viewed as a highly non-trivial generalization of the techniques in our recent paper [DKS15a]
on sums of independent scalar random variables. We would like to emphasize that a number of
4
new conceptual and technical ideas are required to overcome the various obstacles arising in the
multi-dimensional setting.
We start with an intuitive explanation of two key ideas that form the basis of our approach.
Sparsity of the Fourier Transform of PMDs. Since the Fourier Transform (FT) of a PMD is
the product of the FTs of its component CRVs, its magnitude is the product of terms each bounded
from above by 1. Note that each term in the product is strictly less than 1 except in a small region,
unless the component CRV is trivial (i.e., essentially deterministic). Roughly speaking, to establish
the sparsity of the FT of PMDs, we proceed as follows: We bound from above the magnitude of
the FT by the FT of a Gaussian with the same covariance matrix as our PMD. (See, for example,
Lemma 3.10.) This gives us tail bounds for the FT of the PMD in terms of the FT of this Gaussian,
and when combined with the concentration of the PMD itself, yields the desired property.
Approximation of the logarithm of the Fourier Transform. A key ingredient in our proofs
is the approximation of the logarithm of the Fourier Transform (log FT) of PMDs by low-degree
polynomials. Observe that the log FT is a sum of terms, which is convenient for the analysis. We
focus on approximating the log FT by a low-degree Taylor polynomial within the effective support
of the FT. (Note that outside the effective support the log FT can be infinity.) Morally speaking,
the log FT is smooth, i.e., it is approximated by the first several terms of its Taylor series. Formally
however, this statement is in general not true and requires various technical conditions, depending
on the setting. One important point to note is that the sparsity of the FT controls the domain
in which this approximation will need to hold, and thus help us bound the Taylor error. We will
need to ensure that the sizes of the Taylor coefficients are not too large given the location of the
effective support, which turns out to be a non-trivial technical hurdle. To ensure this, we need to
be very careful about how we perform this Taylor expansion. In particular, the correct choice of
the point that we Taylor expand around will be critical for our applications. We elaborate on these
difficulties in the relevant technical sections. Finally, we remark that the degree of polynomial
approximation we will require depends on the setting: In our cover upper bounds, we will require
(nearly) logarithmic degree, while for our CLT degree-2 approximation suffices.
We are now ready to give an overview of the ideas in the proofs of each of our results.
Efficient Learning Algorithm. The high-level structure of our learning algorithm relies on the
sparsity of the Fourier transform, and is similar to the algorithm in our previous work [DKS15a] for
learning sums of independent integer random variables. More specifically, our learning algorithm
estimates the effective support of the DFT, and then computes the empirical DFT in this effective
support. This high-level description would perhaps suffice, if we were only interested in bounding
the sample complexity. In order to obtain a computationally efficient algorithm, it is crucial to use
the appropriate definition of the DFT and its inverse.
In more detail, our algorithm works as follows: It starts by drawing poly(k) samples to estimate
the mean vector and covariance matrix of our PMD to good accuracy. Using these estimates, we
can bound the effective support of our distribution in an appropriate ellipsoid. In particular, we
show that our PMD lies (whp) in a fundamental domain of an appropriate integer lattice L = M Zk ,
where M ∈ Zk×k is an integer matrix whose columns are appropriate functions of the eigenvalues
and eigenvectors of the (sample) covariance matrix. This property allows us to learn our unknown
PMD X by learning the random variable X (mod L). To do this, we learn its Discrete Fourier
transform. Let L∗ be the dual lattice to L (i.e., the set of points ξ so that ξ · x ∈ Z for all
5
b of our PMD X ∼ P on the dual lattice L∗ , that is,
x ∈ L). Importantly, we define the DFT, P,
b : L∗ /Zk → C with P(ξ)
b
P
= E[e(ξ · X)]. A useful property of this definition is the following: the
probability that X (mod L) attains a given value
by the inverse DFT, defined on the
P x is given
1
b
lattice L, namely Pr [X (mod L) = x] = | det(M )| ξ∈L∗ /Zk P(ξ)e(−ξ · x).
The main structural property needed for the analysis of our algorithm is that there exists an
explicit set T with integer coordinates and cardinality (k log(1/ǫ))O(k) that contains all but O(ǫ)
b Given this property, our algorithm draws an additional set of samples of size
of the L1 mass of P.
O(k)
2
(k log(1/ǫ))
/ǫ from the PMD, and computes the empirical DFT (modulo L) on its effective
support T. Using these ingredients, we are able to show that the inverse of the empirical DFT
defines a pseudo-distribution that is ǫ-close to our unknown PMD in total variation distance.
Observe that the support of the inverse DFT can be large, namely Ω(nk−1 ). Our algorithm does
not explicitly evaluate the inverse DFT at all these points, but outputs a succinct description of
b We emphasize that this succinct description suffices to efficiently
its hypothesis H, via its DFT H.
obtain both an approximate evaluation oracle and an approximate sampler for our target PMD
P. Indeed, it is clear that computing the inverse DFT at a single point can be done in time
O(|T |) = (k log(1/ǫ))O(k) , and gives an approximate oracle for the probability mass function of P.
b as a
By using additional algorithmic ingredients, we show how to use an oracle for the DFT, H,
black-box to obtain a computationally efficient approximate sampler for P.
Our learning algorithm and its analysis are given in Section 3.
Constructive Proper Cover and Anonymous Games. The correctness of our learning algorithm easily implies (see Section 3.3) an algorithm to construct a non-proper ǫ-cover for PMDs of
2
O(k)
. While this upper bound is close to being best possible (see Section 4.5),
size nO(k ) ·(1/ǫ)log(1/ǫ))
it does not suffice for our algorithmic applications in anonymous games. For these applications,
it is crucial to obtain an efficient algorithm that constructs a proper ǫ-cover, and in fact one that
works in a certain stylized way.
To construct a proper cover, we rely on the sparsity of the continuous Fourier Transform of
PMDs. Namely, we show that for any PMD P, with effective support S ⊆ [n]k , there exists an
b is at
appropriately defined set T ⊆ [0, 1]k such that the contribution of T to the L1 -norm of |P|
most ǫ/|S|. By using this property, we show that any two PMDs, with approximately the same
variance in each direction, that have continuous Fourier transforms close to each other in the set T,
are close in total variation distance. We build on this lemma to prove our robust moment-matching
result. Roughly speaking, we show that two PMDs, with approximately the same variance in each
direction, that are “close” to each other in their low-degree parameter moments are also close in
total variation distance. We emphasize that the meaning of the term “close” here is quite subtle: we
need to appropriately partition the component CRVs into groups, and approximate the parameter
moments of the PMDs formed by each group within a different degree and different accuracy for
each degree. (See Lemma 4.6 in Section 4.1.)
Our algorithm to construct a proper cover, and our EPTAS for Nash equilibria in anonymous
games proceed by a careful dynamic programming approach, that is based on our aforementioned
robust moment-matching result.
Finally, we note that combining our moment-matching lemma with a recent result in algebraic
geometry gives us the following structural result of independent interest: Every PMD is ǫ-close to
another PMD that is a sum of at most O(k + log(1/ǫ))k distinct k-CRVs.
The aforementioned algorithmic and structural results are given in Section 4.
6
Cover Size Lower Bound. As mentioned above, a crucial ingredient of our cover upper bound
is a robust moment-matching lemma, which translates closeness between the low-degree parameter
moments of two PMDs to closeness between their Fourier Transforms, and in turn to closeness in
total variation distance. To prove our cover lower bound, we follow the opposite direction. We
construct an explicit set of PMDs with the property that any pair of distinct PMDs in our set
have a non-trivial difference in (at least) one of their low-degree parameter moments. We then
show that difference in one of the parameter moments implies that there exists a point where the
probability generating functions have a non-trivial difference. Notably, our proof for this step is
non-constructive making essential use of Cauchy’s integral formula. Finally, we can easily translate
a pointwise difference between the probability generating functions to a non-trivial total variation
distance error. We present our cover lower bound construction in Section 4.5.
Central Limit Theorem for PMDs. The basic idea of the proof of our CLT will be to compare
the Fourier transform of our PMD X to that of the discrete Gaussian G with the same mean
and covariance. By taking the inverse Fourier transform, we will be able to conclude that these
distributions are pointwise close. A careful analysis using a Taylor approximation and the fact that
b and G
b have small effective support, gives us a total variation distance error independent of
both X
the size n. Alas, this approach results in an error dependence that is exponential in k. To obtain
an error bound that scales polynomially with k, we require stronger bounds between X and G at
points away from the mean. Intuitively, we need to take advantage of cancellation in the inverse
Fourier transform integrals. To achieve this, we will use the saddlepoint method from complex
analysis. The full proof of our CLT is given in Section 5.
1.4 Related and Prior Work There is extensive literature on distribution learning and computation of approximate Nash equilibria in various classes of games. We have already mentioned
the most relevant references in the introduction.
Daskalakis et al. [DKT15] studied the structure and learnability of PMDs. They obtained a
5k
k+2
2
non-proper ǫ-cover of size nk · 2O(k log(1/ǫ) ) , and an information-theoretic upper bound on the
learning sample complexity of O(k5k log(1/ǫ)k+2 /ǫ2 ). The dependence on 1/ǫ in their cover size is
also quasi-polynomial, but is suboptimal as follows from our upper and lower bounds. Importantly,
the [DKT15] construction yields a non-proper cover. As previously mentioned, a proper cover
construction is necessary for our algorithmic applications. We note that the learning algorithm
of [DKT15] relies on enumeration over a cover, hence runs in time quasi-polynomial in 1/ǫ, even for
k = 2. The techniques of [DKT15] are orthogonal to ours. Their cover upper bound is obtained by a
clever black-box application of the CLT of [VV10], combined with a non-robust moment-matching
lemma that they deduce from a result of Roos [Roo02]. We remind the reader that our Fourier
techniques strengthen both these technical tools: Theorem 1.3 strengthens the CLT of [VV10], and
we prove a robust and quantitatively essentially optimal moment-matching lemma.
In recent work [DKS15a], the authors used Fourier analytic techniques to study the structure
and learnability of sums of independent integer random variables (SIIRVs). The techniques of this
paper can be viewed as a (highly nontrivial) generalization of those in [DKS15a]. We also note
that the upper bounds we obtain in this paper for learning and covering PMDs do not subsume the
ones in [DKS15a]. In fact, our cover upper and lower bounds in this work show that optimal covers
for PMDs are inherently larger than optimal covers for SIIRVs. Moreover, the sample complexity
of our SIIRV learning algorithm [DKS15a] is significantly better than that of our PMD learning
algorithm in this paper.
7
1.5 Concurrent and Independent Work Concurrently and independently to our work, [DDKT16]
obtained qualitatively similar results using different techniques. We now provide a statement of
the [DDKT16] results in tandem with a comparison to our work.
[DDKT16] give a learning algorithm for PMDs with sample complexity (k log(1/ǫ)O(k) /ǫ2 ) and
2
runtime (k/ǫ)O(k ) . The [DDKT16] algorithm uses the continuous Fourier transform, exploiting
its sparsity property, plus additional structural and algorithmic ingredients. The aforementioned
runtime is not polynomial in the sample size, unless k is fixed. In contrast, our learning algorithm
runs in sample–polynomial time, and, for fixed k, in nearly-linear time. The [DDKT16] learning
algorithm outputs an explicit hypothesis, which can be easily sampled. On the other hand, our
algorithm outputs a succinct description of its hypothesis (via its DFT), and we show how to
efficiently sample from it.
[DDKT16] also prove a size-free CLT, analogous to our Theorem 1.3, with error polynomial in
k and 1/σ. Their CLT is obtained by bootstrapping the CLT of [VV10, VV11] using techniques
from [DKT15]. As previously mentioned, our proof is technically orthogonal to [VV10, VV11,
DDKT16], making use of the sparsity of the Fourier transform combined with tools from complex
analysis. It is worth noting that our CLT also achieves a near-optimal dependence in the error as
a function of 1/σ (up to log factors).
Finally, [DDKT16] prove analogues of Theorems 1.2, 1.4, and 1.5 with qualitatively similar
bounds to ours. We note that [DDKT16] improve the dependence on n in the cover size to an
optimal nO(k), while the dependence on ǫ in their cover upper bound is the same as in [DKT15]. The
cover size lower bound of [DDKT16] is qualitatively of the right form, though slightly suboptimal
as a function of ǫ. The algorithms to construct proper covers and the corresponding EPTAS for
anonymous games in both works have running time roughly comparable to the PMD cover size.
1.6 Organization In Section 3, we describe and analyze our learning algorithm for PMDs.
Section 4 contains our proper cover upper bound construction, our cover size lower bound, and
the related approximation algorithm for Nash equilibria in anonymous games. Finally, Section 5
contains the proof of our CLT.
2
Preliminaries
In this section, we record the necessary definitions and terminology that will be used throughout
the technical sections of this paper.
def
For n ∈ Z+ , we will denote [n] = {1, . . . , n}. For a vector v ∈ Rn , and p ≥ 1, we will
def Pn
denote kvkp = ( i=1 |vi |p )1/p . We will use the boldface notation 0 to denote the zero vector or
matrix in the appropriate dimension.
Notation.
Poisson Multinomial Distributions.
We start by defining our basic object of study:
Definition 2.1 ((n, k)-PMD). For k ∈ Z+ , let ej , j ∈ [k], be the standard unit vector along dimension j in Rk . A k-Categorical Random Variable (k-CRV) is a vector random variable supported
on the set {e1 , e2 , . . . , ek }. A k-Poisson Multinomial
Distribution of order n, or (n, k)-PMD, is any
P
vector random variable of the form X = ni=1 Xi where the Xi ’s are independent k-CRVs. We will
denote by Mn,k the set of all (n, k)-PMDs.
We will require the following notion of a parameter moment for a PMD:
8
P
Definition 2.2 (mth -parameter moment of a PMD). Let X = ni=1 Xi be an (n, k)-PMD such that
for 1 ≤ i ≤ n and 1 ≤ j ≤ k we denote pi,j = Pr[Xi = ej ]. For m = (m1 , . . . , mk ) ∈ Zk+ , we define
Q
P
def P
m
the mth -parameter moment of X to be Mm (X) = ni=1 kj=1 pi,jj . We will refer to |m|1 = kj=1 mj
as the degree of the parameter moment Mm (X).
(Pseudo-)Distributions and Total Variation Distance. P
A function P : A → R, over a finite
set A, is called a distribution ifPP(a) ≥ 0 for all a ∈ A, and a∈A P(a) = 1. The function P is
called
a pseudo-distribution if a∈A P(a) = 1. For S ⊆ A, we sometimes write P(S) to denote
P
P(a).
A distribution P supported on a finite domain A can be viewed as the probability
a∈S
mass function of a random variable X, i.e., P(a) = PrX∼P [X = a].
The total variation distance between two pseudo-distributions P and Q supported on a finite
P
def
domain A is dTV (P, Q) = maxS⊆A |P(S) − Q(S)| = (1/2)·kP−Qk1 = (1/2)· a∈A |P(a)−Q(a)|.
If X and Y are two random variables ranging over a finite set, their total variation distance
dTV (X, Y ) is defined as the total variation distance between their distributions. For convenience,
we will often blur the distinction between a random variable and its distribution.
Covers. Let (X , d) be a metric space. Given ǫ > 0, a subset Y ⊆ X is said to be a proper ǫ-cover
of X with respect to the metric d : X 2 → R+ , if for every x ∈ X there exists some y ∈ Y such that
d(x, y) ≤ ǫ. (If Y is not necessarily a subset of X , then we obtain a non-proper ǫ-cover.) There may
exist many ǫ-covers of X , but one is typically interested in one with the minimum cardinality. The
ǫ-covering number of (X , d) is the minimum cardinality of any ǫ-cover of X . Intuitively, the covering
number of a metric space captures the “size” of the space. In this work, we will be interested on
efficiently constructing sparse covers for PMDs under the total variation distance metric.
Distribution Learning. We now define the notion of distribution learning we use in this paper. Note that an explicit description of a discrete distribution via its probability mass function
scales linearly with the support size. Since we are interested in the computational complexity of
distribution learning, our algorithms will need to use a succinct description of their output hypothesis. A simple succinct representation of a discrete distribution is via an evaluation oracle for the
probability mass function:
Definition 2.3 (Evaluation Oracle). Let P be a distribution over [n]k . An evaluation oracle for P
is a polynomial size circuit C with m = O(k log n) input bits z ∈ [n]k such that for each z ∈ [n]k ,
the output of the circuit C(z) equals the binary representation of the probability P(z). For ǫ > 0,
an ǫ-evaluation oracle for P is an evaluation oracle for some pseudo-distribution P′ which has
dTV (P′ , P) ≤ ǫ.
One of the most general ways to succinctly specify a distribution is to give the code of an efficient
algorithm that takes “pure” randomness and transforms it into a sample from the distribution. This
is the standard notion of a sampler:
Definition 2.4 (Sampler). Let P be a distribution over [n]k . An ǫ-sampler for P is a circuit C
with m = O(k log n + log(1/ǫ)) input bits z and m′ = O(k log n) output bits y which is such that
when z ∼ Um then y ∼ P′ , for some distribution P′ which has dTV (P′ , P) ≤ ǫ.
We can now give a formal definition of distribution learning:
9
Definition 2.5 (Distribution Learning). Let D be a family of distributions. A randomized algorithm AD is a distribution learning algorithm for class D, if for any ǫ > 0, and any P ∈ D, on
input ǫ and sample access to P, with probability 9/10, algorithm AD outputs an ǫ-sampler (or an
ǫ-evaluation oracle) for P.
Remark 2.6. We emphasize that our learning algorithm in Section 3 outputs both an ǫ-sampler
and an ǫ-evaluation oracle for the target distribution.
Anonymous Games and Nash Equilibria. An anonymous game is a triple (n, k, {uiℓ }i∈[n],ℓ∈[k])
where [n], n ≥ 2, is the set of players, [k], k ≥ 2, a common set of strategies available to all players,
and uiℓ the payoff function of player i when she plays strategy ℓ. This function maps the set of
P
partitions Πkn−1 = {(x1 , . . . , xk ) | xℓ ∈ Z+ for all ℓ ∈ [k] ∧ kℓ=1 xℓ = n − 1} to the interval [0, 1].
That is, it is assumed that the payoff of each player depends on her own strategy and only the
number of other players choosing each of the k strategies.
We denote by ∆kn−1 the convex hull of the set Πkn−1 , i.e., ∆kn−1 = {(x1 , . . . , xk ) | xℓ ≥
P
def
0 for all ℓ ∈ [k] ∧ kℓ=1 xℓ = n − 1}. A mixed strategy is an element of ∆k = ∆k1 . A mixed
strategy profile is a mapping δ from [n] to ∆k . We denote by δi the mixed strategy of player i
in the profile δ and δ−i the collection of all mixed strategies but i’s in δ. For ǫ ≥ 0, a mixed
strategy profile δ is a (well-supported) ǫ-Nash equilibrium iff for all i ∈ [n] and ℓ, ℓ′ ∈ [k] we have:
Ex∼δ−i [uiℓ (x)] > Ex∼δ−i [uiℓ′ (x)] + ǫ =⇒ δi (ℓ′ ) = 0. Note that given a mixed strategy profile δ, we
can compute a player’s expected payoff in time poly(nk ) by straightforward dynamic programming.
Note that the mixed strategy δi of player i ∈ [n] defines the k-CRV Xi , i.e., a random vector
supported in the set {e1 , . . . , ek }, such that Pr[Xi = eℓ ] = δi (ℓ), for all ℓ. Hence, if (X1 , . . . , Xn )
is ha mixed
profile, the expected payoff of player i ∈ [n] for using pure strategy ℓ ∈ [k] is
P strategy i
i
.
E uℓ
j6=i,j∈[n] Xj
Multidimensional Fourier Transform. Throughout this paper, we will make essential use of
the (continuous and the discrete) multidimensional Fourier transform. For x ∈ R, we will denote
def
e(x) = exp(−2πix). The (continuous) Fourier
Transform (FT) of a function F : Zk → C is the
P
function Fb : [0, 1]k → C defined as Fb(ξ) = x∈Zk e(ξ · x)F (x). For the case that F is a probability
mass function, we can equivalently write Fb(ξ) = Ex∼F [e(ξ · x)] .
For computational purposes, we will also need the Discrete Fourier Transform (DFT) and its
inverse, whose definition is somewhat more subtle. Let M ∈ Zk×k be an integer k × k matrix. We
def
consider the integer lattice L = L(M ) = M Zk = {p ∈ Zk | p = M q, q ∈ Zk }, and its dual lattice
def
L∗ = L∗ (M ) = {ξ ∈ Rk | ξ · x ∈ Z for all x ∈ L}. Note that L∗ = (M T )−1 Zk , and that L∗ is not
necessarily integral. The quotient Zk /L is the set of equivalence classes of points in Zk such that
two points x, y ∈ Zk are in the same equivalence class iff x − y ∈ L. Similarly, the quotient L∗ /Zk
is the set of equivalence classes of points in L∗ such that any two points x, y ∈ L∗ are in the same
equivalence class iff x − y ∈ Zk .
The Discrete Fourier Transform (DFT) modulo M ,PM ∈ Zk×k , of a function F : Zk → C
is the function FbM : L∗ /Zk → C defined as FbM (ξ) = x∈Zk e(ξ · x)F (x). (We will remove the
subscript M when it is clear from the context.) Similarly, for the case that F is a probability
mass function, we can equivalently write Fb(ξ) = Ex∼F [e(ξ · x)] . The inverse DFT of a function
b : L∗ /Zk → C is the function G : A → C defined on a fundamental domain A of L(M ) as
G
P
1
b
follows: G(x) = | det(M
ξ∈L∗ /Zk G(x)e(−ξ · x). Note that these operations are inverse of each
)|
other, namely for any function F : A → C, the inverse DFT of Fb is identified with F.
10
Pn
Let X =
an (n, k)-PMD such that for 1 ≤ i ≤ n and 1 ≤ j ≤ k we denote
i=1 Xi be P
pi,j = Pr[Xi = ej ], where kj=1 pi,j = 1. To avoid clutter in the notation, we will sometimes use
the symbol X to denote
mass function. With this convention, we
Q the corresponding
Q Pprobability
k
b
ci (ξ) = n
can write that X(ξ)
= ni=1 X
e(ξ
)p
.
j
i,j
i=1
j=1
Basics from Linear Algebra. We remind the reader a few basic definitions from linear algebra that
we will repeatedly use throughout this paper. The Frobenius norm of A ∈ Rm×n is
qP
def
m×n is defined as kAk def
2
kAkF =
2 =
i,j Ai,j . The spectral norm (or induced L2 -norm) of A ∈ R
p
maxx:kxk2 =1 kAxk2 = λmax (AT A). We note that for any A ∈ Rm×n , it holds kAk2 ≤ kAkF . A
symmetric matrix A ∈ Rn×n is called positive semidefinite (PSD), denoted by A 0, if xT Ax ≥ 0
for all x ∈ Rn , or equivalently all the eigenvalues of A are nonnegative. Similarly, a symmetric
matrix A ∈ Rn×n is called positive definite (PD), denoted by A ≻ 0, if xT Ax > 0 for all x ∈ Rn ,
x 6= 0, or equivalently all the eigenvalues of A are strictly positive. For two symmetric matrices
A, B ∈ Rn×n we write A B to denote that the difference A − B is PSD, i.e., A − B 0. Similarly,
we write A ≻ B to denote that the difference A − B is PD, i.e., A − B ≻ 0.
3
Efficiently Learning PMDs
In this section, we describe and analyze our sample near-optimal and computationally efficient
learning algorithm for PMDs. This section is organized as follows: In Section 3.1, we give our
main algorithm which, given samples from a PMD P, efficiently computes a succinct description
of a hypothesis pseudo-distribution H such that dTV (H, P) ≤ ǫ/3. As previously explained, the
b which is supported on a discrete set T of cardinality
succinct description of H is via its DFT H,
O(k)
b
|T | = (k log(1/ǫ))
. Note that H provides an ǫ-evaluation oracle for P with running time O(|T |).
b in a black-box manner, to efficiently obtain an ǫ-sampler
In Section 3.2, we show how to use H,
for P, i.e., sample from a distribution Q such that dTV (Q, P) ≤ ǫ. Finally, in Section 3.3 we show
how a nearly–tight cover upper bound can easily be deduced from our learning algorithm.
3.1 Main Learning Algorithm In this subsection, we give an algorithm Efficient-Learn-PMD
establishing the following theorem:
Theorem 3.1. For all n, k ∈ Z+ and ǫ > 0, the algorithm Efficient-Learn-PMD has the following
performance guarantee: Let P be an unknown (n, k)-PMD. The algorithm uses O k4k log2k (k/ǫ)/ǫ2
b of a
samples from P, runs in time O k6k log3k (k/ǫ)/ǫ2 + k4 log log n , and outputs the DFT H
pseudo-distribution H that, with probability at least 9/10, satisfies dTV (H, P) ≤ ǫ/3.
Our learning algorithm is described in the following pseudo-code:
11
Algorithm Efficient-Learn-PMD
Input: sample access to an (n, k)-PMD X ∼ P and ǫ > 0.
b : T → C of a
Output: A set T ⊆ (R/Z)k of cardinality |T | ≤ O(k2 log(k/ǫ))k , and the DFT H
pseudo-distribution H such that dTV (H, P) ≤ ǫ/3.
Let C > 0 be a sufficiently large universal constant.
b the sample
1. Draw m0 = O(k4 ) samples from X, and let µ
b be the sample mean and Σ
covariance matrix.
b i.e., an orthonormal eigenbasis vi
2. Compute an approximate spectral decomposition of Σ,
with corresponding eigenvalues λi , i ∈ [k].
3. Let M ∈ Zk×k be the matrix
whose ith column is the closest integer point to the vector
q
k ln(k/ǫ)λi + k2 ln2 (k/ǫ) vi .
C
4. Define T ⊆ (R/Z)k to be the set of points ξ = (ξ1 , . . . , ξk ) of the form ξ = (M T )−1 · v+Zk ,
for some v ∈ Zk with kvk2 ≤ C 2 k2 ln(k/ǫ).
5. Draw m = O k4k log2k (k/ǫ)/ǫ2 samples si , i ∈ [m], from X, and output the empirical
1 Pm
b : T → C, i.e., H(ξ)
b
DFT H
=m
i=1 e(ξ · si ).
b is a succinct description of the pseudo-distribution H, the inverse DFT of
/* The DFT H
P
1
b defined by: H(x) =
b
H,
H(ξ)e(−ξ
· x), for x ∈ Zk ∩ (b
µ + M · (−1/2, 1/2]k ),
| det(M )|
ξ∈T
and H(x) = 0 otherwise. Our algorithm does not output H explicitly, but implicitly via
its DFT.*/
Let X be the unknown target (n, k)-PMD. We will denote by P the probability mass function
of X, i.e., X ∼ P. Throughout this analysis, we will denote by µ and Σ the mean vector and
covariance matrix of X.
First, note that the algorithm Efficient-Learn-PMD is easily seen to have the desired sample
and time complexity. Indeed, the algorithm draws m0 samples in Step 1 and m samples in Step 5,
for a total sample complexity of O(k4k log2k (k/ǫ)/ǫ2 ). The runtime of the algorithm is dominated
by computing the DFT in Step 5 which takes time O(m|T |) = O(k6k log3k (k/ǫ)/ǫ2 ). Computing
an approximate eigendecomposition can be done in time O(k 4 log log n)(see, e.g., [PC99]). The
remaining part of this section is devoted to proving the correctness of our algorithm.
Remark 3.2. We remark that in Step 4 of our algorithm, the notation ξ = (M T )−1 · v+Zk refers
to an equivalence class of points. In particular, any pair of distinct vectors v, v ′ ∈ Zk satisfying
kvk2 , kv ′ k2 ≤ C 2 k2 ln(k/ǫ), and (M T )−1 ·(v −v ′ ) ∈ Zk correspond to the same point ξ, and therefore
are not counted twice.
Overview of Analysis. We begin with a brief overview of the analysis. First, we show (Lemma 3.3)
that at least 1 − O(ǫ) of the probability mass of X lies in the ellipsoid with center µ and coe = O(k log(k/ǫ))Σ + O(k log(k/ǫ))2 I. Moreover, with high probability over the
variance matrix Σ
b and µ
samples drawn in Step 1 of the algorithm, the estimates Σ
b will be good approximations of
Σ and µ (Lemma 3.4). By combining these two lemmas, we obtain (Corollary 3.5) that at least
1 − O(ǫ) of the probability mass of X lies in the ellipsoid with center µ
b and covariance matrix
′
2
b
Σ = O(k log(k/ǫ))Σ + O(k log(k/ǫ)) I.
12
By the above, and by our choice of the matrix M ∈ Zk×k , we use linear-algebraic arguments to
prove (Lemma 3.6) that almost all of the probability mass of X lies in the set µ
b + M (−1/2, 1/2]k ,
k
a fundamental domain of the lattice L = M Z . This lemma is crucial because it implies that, to
learn our PMD X, it suffices to learn the random variable X (mod L). We do this by learning the
Discrete Fourier transform of this distribution. This step can be implemented efficiently due to the
sparsity property of the DFT (Proposition 3.8): except for points in T , the magnitude of the DFT
will be very small. Establishing the desired sparsity property for the DFT is the main technical
contribution of this section.
Given the above, it it fairly easy to complete the analysis of correctness. For every point in T we
√
can learn the DFT up to absolute error O(1/ m). Since the cardinality of T is appropriately small,
this implies that the total error over T is small. The sparsity property of the DFT (Lemma 3.14)
completes the proof.
Detailed Analysis. We now proceed with the detailed analysis of our algorithm. We start
by showing that PMDs are concentrated with high probability. More specifically, the following
lemma shows that an unknown PMD X, with mean vector µ and covariance matrix Σ, is effectively
supported in an ellipsoid centered at µ, whose principal axes are determined by the eigenvectors
and eigenvalues of Σ and the desired concentration probability:
Lemma 3.3. Let X be an (n, k)-PMD with mean vector µ and covariance matrix Σ. For any
e = k ln(k/ǫ)Σ+k2 ln2 (k/ǫ)I. Then, with probability
0 < ǫ < 1, consider the positive-definite matrix Σ
e −1 · (X − µ) = O(1).
at least 1 − ǫ/10 over X, we have that (X − µ)T · Σ
P
Proof. Let X = ni=1 Xi , where the Xi ’s are independent k-CRVs. We can write µ = E[X] =
P
n
k
i=1 µi , where µi = E[Xi ]. Note that for any unit vector u ∈ R , kuk2 = 1, we have that the scalar
random variable u · (X − µ) is
Pa sum of independent, mean 0, bounded random variables. Indeed,
we have that u · (X − µ) = ni=1 u · (Xi − µi ), and that E[u · (Xi − µi )] = u · (E[Xi ] − µi ) = 0.
Moreover, we can write
√
√
|u · (Xi − µi )| ≤ kuk2 · kXi − µi k2 ≤ kXi k2 + kµi k2 ≤ 1 + kkµi k1 ≤ 2 k ,
where we used the Cauchy-Schwartz inequality twice, the triangle inequality, and the fact that a
k-CRV Xi with mean µi by definition satisfy kXi k2 = 1, and kµi k1 = 1.
variance of u · (X − µ). By Bernstein’s inequality, we obtain that for t =
p Let ν be the √
2ν ln(10k/ǫ) + 2 k ln(10k/ǫ) it holds
t2 /2
ǫ
√
Pr [|u · (X − µ)| > t] ≤ exp −
≤
.
(1)
10k
ν + 2 kt/3
Let Σ, the covariance matrix of X, have an orthonormal eigenbasis uj ∈ Rk , kuj k2 = 1, with
corresponding eigenvalues νj , j ∈ [k]. Since Σ is positive-semidefinite, we have that νj ≥ 0, for all
j ∈ [k]. We consider the random variable uj · (X − µ). In addition to being a sum of independent,
mean 0, bounded random
variables, we
claim that Var[uj · (X − µ)] = νj . First, it is clear that
Var[uj · (X − µ)] = E (uj · (X − µ))2 . Moreover, note that for any vector v ∈ Rk , we have that
E[(v · (X − µ))2 ] = v T · Σ · v. For v = uj , we thus get E (uj · (X − µ))2 = uTj · Σ · uj = νj .
√
p
Applying (1) for uj · (X − µ), with tj = 2νj ln(10k/ǫ) + 2 k ln(10k/ǫ), yields that for all
j ∈ [k] we have
!
t2j /2
ǫ
√
.
≤
Pr [|uj · (X − µ)| > tj ] ≤ exp −
10k
νj + 2 ktj /3
13
By a union bound, it follows that
Pr [∀j ∈ [k] : |uj · (X − µ)| ≤ tj ] ≥ 1 − ǫ/10 .
We condition on this event.
Since uj and νj are the eigenvectors and eigenvalues of Σ, we have that Σ = U T · diag (νj ) · U,
where U has j th column uj . We can thus write
e = U T · diag kνj ln(k/ǫ) + k2 ln2 (k/ǫ) · U.
Σ
Therefore, we have:
T
e −1
(X − µ) · Σ
· (X − µ) = diag
q
≤ k diag
j∈[k]
kνj ln(k/ǫ) + k2 ln (k/ǫ)
q
= max q
2
−1
2
kνj ln(k/ǫ) + k2 ln (k/ǫ)
|uj · (X − µ)|
2
νj ln(k/ǫ) + k ln (k/ǫ)
= O(1) ,
2
2
T
· U · (X − µ)
−1
2
2
T
· U · (X − µ)
∞
where the √
last inequality follows
√ √from our conditioning, the definition of tj , and the elementary
√
inequality a + b ≥ ( a + b)/ 2, a, b ∈ R+ . This completes the proof of Lemma 3.3.
Lemma 3.3 shows that an arbitrary (n, k)-PMD X puts at least 1 − ǫ/10 of its probability mass
e −1 · (x − µ) ≤ c}, where c > 0 is an appropriate
in the ellipsoid E = {x ∈ Rk : (x − µ)T · (Σ)
universal constant. This is the ellipsoid centered at µ, whose principal semiaxes are parallel to
e −1 . The length of the principal semiaxis
the uj ’s, i.e., the eigenvectors of Σ, or equivalently of (Σ)
e −1 , and is equal to c−1/2 ·
corresponding to uj is determined by the corresponding eigenvalue of (Σ)
q
kνj ln(k/ǫ) + k2 ln2 (k/ǫ).
Note that this ellipsoid depends on the mean vector µ and covariance matrix Σ, that are
unknown to the algorithm. To obtain a bounding ellipsoid that is known to the algorithm, we will
b are good
use the following lemma (see Appendix A for the simple proof) showing that µ
b and Σ
approximations to µ and Σ respectively.
Lemma 3.4. With probability at least 19/20 over the samples drawn in Step 1 of the algorithm,
b + I (Σ + I)/2.
we have that (b
µ − µ)T · (Σ + I)−1 · (b
µ − µ) = O(1), and 2(Σ + I) Σ
b Concretely,
We also need to deal with the error introduced in the eigendecomposition of Σ.
T
b
we factorize Σ as V ΛV, for an orthogonal matrix V and diagonal matrix Λ. This factorization
b by a constant factor, we
is necessarily inexact. By increasing the precision to which we learn Σ
T
b in terms of our computed
can still have 2(Σ + I) V ΛV + I (Σ + I)/2. We could redefine Σ
T
b
orthonormal eigenbasis, i.e., Σ := V ΛV . Thus, we may henceforth assume that the decomposition
b = V T ΛV is exact.
Σ
For the rest of this section, we will condition on the event that the statements of Lemma 3.4 are
satisfied. By combining Lemmas 3.3 and 3.4 , we show that we can get a known ellipsoid containing
the effective support of X, by replacing µ and Σ in the definition of E by their sample versions.
More specifically, we have the following corollary:
14
b + k2 ln2 (k/ǫ)I. Then, with probability at least 1 − ǫ/10 over
Corollary 3.5. Let Σ′ = k ln(k/ǫ)Σ
X, we have that (X − µ
b)T · (Σ′ )−1 · (X − µ
b) = O(1).
b + I (Σ + I)/2. Hence, we have that
Proof. By Lemma 3.4, it holds that 2(Σ + I) Σ
b + k2 ln2 (k/ǫ)I 1 k ln(k/ǫ)Σ + k2 ln2 (k/ǫ)I .
2 k ln(k/ǫ)Σ + k 2 ln2 (k/ǫ)I k ln(k/ǫ)Σ
2
e By standard results, taking inverses reverses the
e this is 2Σ
e Σ′ 1 Σ.
In terms of Σ′ and Σ,
2
positive semi-definite ordering (see e.g., Corollary 7.7.4 (a) in [HJ85]). Hence,
e −1 Σ′−1 1 Σ
e −1 .
2Σ
2
Combining the above with Lemma 3.3, with probability at least 1 − ǫ/10 over X we have that
e −1 · (X − µ) = O(1) .
(X − µ)T · Σ′−1 · (X − µ) ≤ 2(X − µ)T · Σ
(2)
(b
µ − µ)T · Σ′−1 · (b
µ − µ) ≤ (b
µ − µ)T · (Σ + I)−1 · (b
µ − µ) = O(1) .
(3)
e Σ + I, and therefore (Σ′ )−1 (Σ + I)−1 , Lemma 3.4 gives that
Since Σ′ 21 Σ
We then obtain:
(X − µ
b)T · (Σ′ )−1 · (X − µ
b) = ((X − µ) − (b
µ − µ))T · (Σ′ )−1 · ((X − µ) − (b
µ − µ))
= (X − µ)T · Σ′−1 · (X − µ) + (b
µ − µ)T · Σ′−1 · (b
µ − µ)
− (X − µ)T · Σ′−1 · (b
µ − µ) − (b
µ − µ)T · Σ′−1 · (X − µ) .
Equations (2) and (3) yield that the first two terms are O(1). Since Σ′−1 is positive-definite,
xT Σ′−1 y as a function of vectors x, y ∈ Rk is an inner product. So, we may apply the CauchySchwartz inequality to bound each of the last two terms from above by
q
((X − µ)T · Σ′−1 · (X − µ)) · ((b
µ − µ)T · Σ′−1 · (b
µ − µ)) = O(1) ,
where the last equality follows from (2) and (3). This completes the proof of Corollary 3.5.
Corollary 3.5 shows that our unknown PMD X puts at least 1 − ǫ/10 of its probability mass
in the ellipsoid E ′ = {x ∈ Rk : (x − µ
b)T · (Σ′ )−1 · (x − µ
b) ≤ c}, for an appropriate universal
constant c > 0. This is the ellipsoid centered at µ
b, whose principal semiaxes are parallel to the
b and the length of the principal semiaxis parallel to vj is equal to
vj ’s, the
eigenvectors
of
Σ,
q
c−1/2 ·
kλj ln(k/ǫ) + k2 ln2 (k/ǫ).
Our next step is to to relate the ellipsoid E ′ to the integermatrix M ∈ Zk×k used inour
q
algorithm. Let M ′ ∈ Rk×k be the matrix with j th column C
k ln(k/ǫ)λj + k2 ln2 (k/ǫ) vj ,
where C > 0 is the constant in the algorithm statement. The matrix M ∈ Zk×k is obtained
by rounding each entry of M ′ to the closest integer point. We note that the ellipsoid E ′ can be
equivalently expressed as E ′ = {x ∈ Rk : k(M ′ )−1 · (x − µ
b)k2 ≤ 1/4}. Using the relation between M
′
′
and M , we show that E is enclosed in the ellipsoid {x ∈ Rk : k(M )−1 · (x − µ
b)k2 < 1/2}, which is in
turn enclosed in the parallelepiped with integer corner points {x ∈ Rk : k(M )−1 · (x − µ
b)k∞ < 1/2}.
k
This parallelepiped is a fundamental domain of the lattice L = M Z . Formally, we have:
15
Lemma 3.6. With probability at least 1 − ǫ/10 over X, we have that X ∈ µ
b + M (−1/2, 1/2]k .
q
2
′
k×k
2
Proof. Let M ∈ R
be the matrix with columns C
k ln(k/ǫ)λi + k ln (k/ǫ) vi , where C > 0
is the constant in the algorithm statement. Note that,
b + C 2 k2 ln2 (k/ǫ)I = C 2 Σ′ .
M ′ (M ′ )T = C 2 k ln(k/ǫ)Σ
(4)
k(M ′ )−1 · (X − µ
b)k22 = C −2 (X − µ
b)T · (Σ′ )−1 · (X − µ
b) ≤ 1/16 .
(5)
For a large enough constant C, Corollary 3.5 implies that with probability at least 1 − ǫ/10,
Note that the above is an equivalent description of the ellipsoid E ′ . Our lemma will follow from the
following claim:
Claim 3.7. For any x ∈ Rk , it holds
kM −1 xk2 < 2k(M ′ )−1 xk2 and kM T xk2 < 2k(M ′ )T xk2 .
(6)
Proof. By construction, M and M ′ differ by at most 1 in each entry, and therefore M − M ′ has
Frobenius norm (and, thus, induced L2 -norm) at most k. For any x ∈ Rk , we thus have that
q
q
k(M ′ )T xk2 = xT · M ′ · (M ′ )T · x ≥ xT · (C 2 k2 I) · x > 2kkxk2 ,
and therefore
1
kM T xk2 ≥ k(M ′ )T xk2 − k(M − M ′ )T k2 kxk2 > k(M ′ )T xk2 .
2
T
′
T
′
T
Similarly, we get kM xk2 ≤ k(M ) xk2 + k(M − M ) k2 kxk2 < 2k(M ′ )T xk2 . In terms of the PSD
ordering, we have:
1 ′
M (M ′ )T ≺ M M T ≺ 4M ′ (M ′ )T .
(7)
4
Since M ′ (M ′ )T I, both M ′ (M ′ )T and M M T are positive-definite, and so M and M ′ are invertible.
Taking inverses in Equation (7) reverses the ordering, that is:
1
(M ′−1 )T M ′−1 ≺ (M −1 )T M −1 ≺ 4(M ′−1 )T M ′−1 .
4
The claim now follows.
Hence, Claim 3.7 implies that with probability at least 1 − ǫ/10, we have:
kM −1 (X − µ
b)k∞ ≤ kM −1 (X − µ
b)k2 < 2k(M ′ )−1 (X − µ
b)k2 < 1/2 ,
where the last inequality follows from (5). In other words, with probability at least 1 − ǫ/10, X
lies in µ
b + M (−1/2, 1/2]k , which was to be proved.
Recall that L denotes the lattice M Zk . The above lemma implies that it is sufficient to learn
the random variable X (mod L). To do this, we will learn its Discrete Fourier transform. Let
L∗ be the dual lattice to L. Recall that the DFT of the PMD P, with X ∼ P, is the function
b : L∗ /Zk → C defined by P(ξ)
b
P
= E[e(ξ · X)]. Moreover, the probability that X (mod L) attains
a given value x is given by the inverse DFT, namely
X
1
b
P(ξ)e(−ξ
· x) .
Pr [X (mod L) = x] =
| det(M )|
∗
k
ξ∈L /Z
The main component of the analysis is the following proposition, establishing that the total contribution to the above sum coming from points ξ 6∈ T is small. In particular, we prove the following:
16
Proposition 3.8. We have that
P
ξ∈(L∗ /Zk )\T
b
|P(ξ)|
< ǫ/10.
To prove this proposition, we will need a number of intermediate claims and lemmas. We start
with the following claim, showing that for every point ξ ∈ Rk , there exists an integer shift whose
coordinates lie in an interval of length strictly less than 1:
Claim 3.9. For each ξ ∈ Rk , there exists a ∈ Z+ with 0 ≤ a ≤ k, and b ∈ Zk such that
ik
h
a
, a+k
ξ − b ∈ k+1
k+1 .
Proof. Consider the k fractional parts
′ of the
coordinates of ξ, i.e., ξi − ⌊ξi ⌋, for 1 ≤ i ≤ k. Now
′
a
−1
a
consider the k + 1 intervals Ia′ = k+1 , k+1 , for 1 ≤ a′ ≤ k + 1. By the pigeonhole principle,
there is an a′ such that ξi − ⌊ξi ⌋ ∈
/ Ia′ , for all i, 1 ≤ i ≤ k. We define a = a′ when a′ < k + 1, and
a = 0 when a′ = k + 1.
i
i h
h
a
a−1
,1
∪ k+1
For any i, with 1 ≤ i ≤ k, since ξi − ⌊ξi ⌋ ∈
/ Ia′ , we have that ξi − ⌊ξi ⌋ ∈ 0, k+1
(taking the first hinterval toi be empty if a = 0). Hence, by setting one of bi = ⌊ξi ⌋, or bi = ⌊ξi ⌋ − 1,
a
we get ξi − bi ∈ k+1
, a+k
k+1 . This completes the proof.
The following lemma gives a “Gaussian decay” upper bound on the magnitude of the DFT, at
points ξ whose coordinates lie in an interval of length less than 1. Roughly speaking, the proof of
Proposition 3.8 proceeds by applying this lemma for all ξ ∈
/ T.
Lemma 3.10. Fix δ ∈ (0, 1). Suppose that ξ ∈ Rk has coordinates lying in an interval I of length
b
1 − δ. Then, |P(ξ)|
= exp(−Ω(δ2 ξ T · Σ · ξ)).
P
Proof. Since P is a PMD, we have X = ni=1 Xi , where Xi ∼ Pi for independent k-CRV’s Pi , we
Q
b
ci (ξ)|. Note also that ξ T · Σ · ξ = Var[ξ · X] = Pn Var[ξ · Xi ]. It therefore
have that |P(ξ)|
= ni=1 |P
i=1
suffices to show that for each i ∈ [n] it holds
ci (ξ)| = exp(−Ω(δ2 Var[ξ · Xi ])) .
|P
Let Xi′ ∼ Pi be an independent copy of Xi , and Yi = Xi − Xi′ . We note that Var[ξ · Xi ] =
ci (ξ)|2 = E[e(ξ · Yi )]. Since Yi is a symmetric random variable, we have
(1/2)E[(ξ · Yi )2 ], and that |P
that
ci (ξ)|2 = E[e(ξ · Yi )] = E[cos(2πξ · Yi )] .
|P
We will need the following technical claim:
Claim 3.11. Fix 0 < δ < 1. For all x ∈ R with |x| ≤ 1 − δ, it holds 1 − cos(2πx) ≥ δ2 x2 .
Proof. When 0 ≤ |x| ≤ 1/4, sin(2πx) is concave, since its second derivative is −4π 2 sin(2πx) ≤ 0.
So, we have sin(2πx) ≥ (1 − 4x) sin(0) + 4x sin(π/2) = 4x. Integrating the latter inequality, we
obtain that (1 − cos(2πx))/2π ≥ 2x2 , i.e., (1 − cos(2πx)) ≥ 4πx2 . Thus, for 0 ≤ |x| ≤ 1/4, we have
1 − cos(2πx)) ≥ 4πx2 ≥ δ2 x2 .
When 1/4 ≤ |x| ≤ 3/4, we have (1 − cos(2πx)) ≥ 1 ≥ δ2 x2 . Finally, when 3/4 ≤ |x| ≤ 1 − δ,
we have 0 ≤ 1 − |x| ≤ δ ≤ 1/4, and therefore 1 − cos(2πx) = 1 − cos(2π(1 − |x|)) ≥ 1 − cos(2πδ) ≥
4πδ2 ≥ δ2 x2 . This establishes the proof of the claim.
ci (ξ)|2 is
Since ξ · Yi is by assumption supported on the interval [−1 + δ, 1 − δ], we have that |P
E[cos(2πξ · Yi )] = E[1 − Ω(δ2 (ξ · Yi )2 )] ≤ exp(−Ω(δ2 E[(ξ · Yi )2 ])) = exp(−Ω(δ2 Var[ξ · Xi ])) .
This completes the proof of Lemma 3.10.
17
We are now ready to prove the following crucial lemma, which shows that the DFT of P is
effectively supported on the set T.
Lemma 3.12. For integers 0 ≤ a ≤ k, we have that
X
b
|P(ξ)|
<
k
a
ξ∈L∗ ∩[ k+1
\(T +Zk )
, a+k
k+1 ]
ǫ
.
10(k + 1)
We start by providing a brief overview of the proof. First note that Claim 3.9 and Lemma 3.10
b
together imply that for ξ ∈ [a/(k + 1), (a + k)/(k + 1)]k , if ξ T · Σ · ξ ≥ k2 log(1/ǫ′ ), then |P(ξ)|
≤ ǫ′ ,
for any ǫ′ > 0. Observe that the set {ξ ∈ Rk : ξ T · Σ · ξ ≤ k2 log(1/ǫ′ )} is not an ellipsoid,
because Σ is singular. However, by using the fact that M and M ′ are close to each other, – more
specifically, using ingredients from the proof of Lemma 3.6 – we are able to bound its intersection
with [a/(k + 1), (a + k)/(k + 1)]k by an appropriate ellipsoid of the form {ξ T · (M · M T ) · ξ ≤ r}.
The gain here is that M T ξ ∈ Zk . This allows us to provide an upper bound on the cardinality
of the set of lattice points in L∗ in one of these ellipsoids. Note that, in terms of v = M T ξ, these
are the integer points that lie in a sphere of some radius r > r0 = C 2 k2 ln(k/ǫ). Now, if we consider
the set 2t r0 ≤ ξ T · (M M T ) · ξ < 2t+1 r0 , we have both an upper bound on the magnitude of the
DFT and on the number of integer points in the set. By summing over all values of t ≥ 0, we get
an upper bound on the error coming from points outside of T.
Proof of Lemma 3.12.
/ T + Zk , we have that kM T ξk2 > C 2 k2 ln(k/ǫ). Since the coorh Since iξ ∈
a
, a+k
dinates of ξ lie in k+1
k+1 , an interval of length 1 − 1/(k + 1), we may apply Lemma 3.10 to
obtain:
b
|P(ξ)|
= exp −Ω k−2 ξ T · Σ · ξ
b − I) · ξ
= exp −Ω k−2 ξ T · (Σ
(by Lemma 3.4)
T ′
ξ M (M ′ )T ξ
= exp −Ω k−2
− k log(k/ǫ)kξk22
(by Equation (4))
C 2 k log(k/ǫ)
k(M ′ )T ξk22
2
2
−2
− k log(k/ǫ)kξk∞
= exp −Ω k
C 2 k log(k/ǫ)
kM T ξk22
2
2
−2
− k log(k/ǫ)kξk∞
(by Equation (6))
= exp −Ω k
C 2 k log(k/ǫ)
= exp −Ω C −2 k−3 log−1 (k/ǫ)kM T ξk22 − log(k/ǫ)
(since kξk∞ ≤ 2)
= exp −Ω C −2 k−3 log−1 (k/ǫ)kM T ξk22 .
(since kM T ξk2 > C 2 k2 ln(k/ǫ))
Next note that for ξ ∈ L∗ we have that M T ξ ∈ Zk . Thus, letting v = M T ξ, it suffices to show that
X
ǫ
exp −Ω C −2 k−3 log−1 (k/ǫ)kvk22 <
.
(8)
10(k
+ 1)
k
2 2
v∈Z ,kvk2 ≥C k ln(k/ǫ)
Although the integer points in the above sum are not in the sphere kvk2 ≤ C 2 k2 ln(k/ǫ), they lie
in some sphere kvk2 ≤ 2t+1 C 2 k2 ln(k/ǫ), for some integer t > 0. The number of integral points in
one of these spheres is less than that of the appropriate enclosing cube. Namely, we have that
n
o
# v ∈ Zk , kvk2 ≤ 2t+1 C 2 k2 ln(k/ǫ)
n
o
≤ # v ∈ Zk , kvk∞ ≤ 2t+1 C 2 k2 ln(k/ǫ)
k
= 1 + 2⌊2t+1 C 2 k2 ln(k/ǫ)⌋ .
(9)
18
Inequality (8) is obtained by bounding the LHS from above as follows:
∞
X
t=0
=
∞
X
t=0
≤
=
∞
X
t=0
∞
X
t=0
n
o
exp −Ω(C −2 k−3 log−1 (k/ǫ)(2t C 2 k2 ln(k/ǫ))2 ) · # v ∈ Zk , kvk2 ≤ 2t+1 C 2 k2 ln(k/ǫ)
n
o
exp −Ω(Ck log(k/ǫ)4t ) · # v ∈ Zk , kvk2 ≤ 2t+1 C 2 k2 ln(k/ǫ)
k
exp −Ω(Ck log(k/ǫ)4t ) · 2t+2 C 2 k2 log(k/ǫ)
(by Equation (9))
exp −Ω(Ck log(k/ǫ)4t ) exp (O(k(t + log k + log log(k/ǫ))))
≤ exp (− ln((k + 1)/10ǫ)) ·
ǫ2
∞
X
√
∞
X
t=0
√
exp(−k C(4t − t)))
exp(− C(4t − t))
·
10(k + 1) t=0
ǫ
.
<
10(k + 1)
≤
This completes the proof of Lemma 3.12.
We are now prepared to prove Proposition 3.8.
Proof of Proposition 3.8. Let Ta be the set of points ξ ∈ Rk which have a lift with all coordinates in
ik
h
S
a
, for some integer 0 ≤ a ≤ k. By Claim 3.9, we have that a Ta = (L∗ /Zk ).
, a+k
the interval k+1
k+1
By Lemma 3.12, for all 0 ≤ a ≤ k,
X
ǫ
b
|P(ξ)|
<
,
10(k + 1)
ξ∈Ta \T
and so we have:
X
ξ∈(L∗ /Zk )\T
b
|P(ξ)|
≤
k
X
X
a=0 ξ∈Ta \T
b
|P(ξ)|
< ǫ/10 .
Our next simple lemma states that the empirical DFT is a good approximation to the true
DFT on the set T.
Lemma 3.13. Letting m = P
(C 5 k4 ln2 (k/ǫ))k /ǫ2 , with 19/20 probability over the choice of m samb
b
ples in Step 5, we have that ξ∈T |H(ξ)
− P(ξ)|
< ǫ/10.
b
Proof. For any given ξ ∈ T, we note that H(ξ)
is the average of m samples from e(ξ · X), a random
b
variable whose distribution has mean P(ξ)
and variance at most O(1). Therefore, we have that
√
b
b
E[|H(ξ)
− P(ξ)|]
≤ O(1)/ m.
Summing over ξ ∈ T, and noting that |T | ≤ O(C 2 k2 log(k/ǫ))k , we get that the expectation of the
quantity in question is less than ǫ/400. Markov’s inequality completes the argument.
19
Finally, we bound from above the total variation distance between P and H.
Lemma 3.14. Assuming that the conclusion of the previous lemma holds, then for any x ∈ Zk /L
we have that
Pr[X ≡ x (mod L)] −
Proof. We note that
=
≤
X
1
ǫ
b
H(ξ)e(−ξ
· x) ≤
.
| det(M )|
5| det(M )|
ξ∈T
Pr[X ≡ x
(mod L)] −
1
| det(M )|
X
1
| det(M )|
ξ∈L∗ /Zk
X
ξ∈T
b
P(ξ)e(−ξ
· x) −
ξ∈L∗ /Zk ,ξ6∈T
ǫ
,
≤
5| det(M )|
X
1
b
H(ξ)e(−ξ
· x)
| det(M )|
b
|P(ξ)|
−
X
1
b
H(ξ)e(−ξ
· x)
det(M )
1
| det(M )|
ξ∈T
X
ξ∈T
b
b
|P(ξ)
− H(ξ)|
where the last line follows from Proposition 3.8 and Lemma 3.13.
It follows that, for each x ∈ µ
b + M (−1/2, 1/2]k , our hypothesis pseudo-distribution H(x) equals
ǫ
words, the pseudothe probability that X ≡ x (mod L) plus an error of at most 5| det(M
)| .In other
ǫ
k
distribution defined by H (mod L) differs from X (mod L) by at most 5| det(M
)| |Z /L| = ǫ/5. On
the other hand, letting X ′ ∼ P′ be obtained by moving a sample from X to its unique representative
modulo L lying in µ
b + M (−1/2, 1/2]k , we have that X = X ′ with probability at least 1 − ǫ/10.
Therefore, dTV (P, P′ ) ≤ ǫ/10. Note that X (mod L) = X ′ (mod L), and so dTV (H (mod L), P′
(mod L)) < ǫ/5. Moreover, H and P′ are both supported on the same fundamental domain of L,
and hence dTV (H, P′ ) = dTV (H (mod L), P′ (mod L)) < ǫ/5. Therefore, assuming that the above
high probability events hold, we have that dTV (H, P) ≤ dTV (H, P′ ) + dTV (P, P′ ) ≤ 3ǫ/10.
This completes the analysis and the proof of Theorem 3.1.
3.2 An Efficient Sampler for our Hypothesis The learning algorithm of Section 3.1 outputs
a succinct description of the hypothesis pseudo-distribution H, via its DFT. This immediately
provides us with an efficient evaluation oracle for H, i.e., an ǫ-evaluation oracle for our target PMD
P. The running time of this oracle is linear in the size of T, the effective support of the DFT.
Note that we can explicitly output the hypothesis H by computing the inverse DFT at all the
b the support of H can
points of the support of H. However, in contrast to the effective support of H,
be large, and this explicit description would not lead to a computationally efficient algorithm. In this
subsection, we show how to efficiently obtain an ǫ-sampler for our unknown PMD P, using the DFT
representation of H as a black-box. In particular, starting with the DFT of an accurate hypothesis
H, represented via its DFT, we show how to efficiently obtain an ǫ-sampler for the unknown target
distribution. We remark that the efficient procedure of this subsection is not restricted to PMDs,
but is more general, applying to all discrete distributions with an approximately sparse DFT (over
any dimension) for which an efficient oracle for the DFT is available.
In particular, we prove the following theorem:
20
Theorem 3.15. Let M ∈ Zk×k , m ∈ Rk , and S = m + M (−1/2, 1/2] k ∩ Zk . Let H : S → R be
b which is supported on
a pseudo-distribution succinctly represented
via its DFT (modulo M ), H,
P
b
b
a set T, i.e., H(x) = (1/| det(M )|) · ξ∈T e(−ξ · x)H(ξ),
for x ∈ S, with 0 ∈ T and H(0)
=
1. Suppose that there exists a distribution P with dTV (H, P) ≤ ǫ/3. Then, there exists an ǫsampler for P, i.e., a sampler for a distribution Q such that dTV (P, Q) ≤ ǫ, running in time
O(log(| det(M )|) log(| det(M )|/ǫ) · |T | · poly(k)).
We remark that the ǫ-sampler in the above theorem statement can be described as a randomized
b
algorithm that takes as input M , T , H(ξ),
for ξ ∈ T, and the Smith normal form decomposition of
M (see Lemma 3.21).
We start by observing that our main learning result, Theorem 1.1, follows by combining Theorem 3.1 with Theorem 3.15. Indeed, note thatqthe matrix M in the definition of our PMD
e ≤ (nk log(1/ǫ))O(k) . Also recall that
algorithm in Section 3.1 satisfies | det(M )| ≤ O( det(Σ))
|T | = O(k2k logk (k/ǫ)). Since M has largest entry n, by Lemma 3.21, we can compute its Smith
normal form decomposition in time poly(k) log n. Hence, for the case of PMDs, we obtain the
following corollary, establishing Theorem 1.1:
Corollary 3.16. For all n, k ∈ Z+ and ǫ > 0, there is an algorithm with the following performance
guarantee: Let P be an unknown (n, k)-PMD. The algorithm uses O k4k log2k (k/ǫ)/ǫ2 samples
from P, runs in time O k6k log3k (k/ǫ)/ǫ2 · log n , and with probability at least 9/10 outputs an
ǫ-sampler for P. This ǫ-sampler runs (i.e., produces a sample) in time poly(k)O(k 2k logk+1 (k/ǫ)) ·
log2 n.
This section is devoted to the proof of Theorem 3.15. We first handle the case of one-dimensional
distributions, and then appropriately reduce the high-dimensional case to the one-dimensional.
b
Remark 3.17.
P We remark that the assumption that H(0) = 1 in our theorem statement, ensures that x∈S H(x) = 1, and so, for any distribution P over S the total variational distance
P
def 1 P
|H(x)−P(x)|
is
well
behaved
in
the
sense
that
d
(H,
P)
=
dTV (H, P)
=
TV
x∈S
x:P(x)>H(x) (P(x)−
2
P
H(x)) = x:P(x)<H(x) (H(x) − P(x)). This fact will be useful in the correctness of our sampler.
We start by providing some high-level intuition. Roughly speaking, we obtain the desired
sampler by considering an appropriate definition of a Cumulative Distribution Function (CDF)
corresponding to H. For the 1-dimensional case (i.e., the case k = 1 in our theorem statement), the
definition of the CDF is clear, and our sampler proceeds as follows: We use the DFT to obtain a
closed form expression for the CDF of H, and then we query the CDF using an appropriate binary
search procedure to sample from the distribution. One subtle point is that H(x) is a pseudodistribution, i.e. it is not necessarily non-negative at all points. Our analysis shows that this does
not pose any problems with correctness, by using the aforementioned remark.
For the case of two or more dimensions (k ≥ 2), we essentially provide a computationally
efficient reduction to the 1-dimensional case. In particular, we exploit the fact that the underlying
domain is discrete, to define an efficiently computable bijection from the domain to the integers,
and consider the corresponding 1-dimensional CDF. To achieve this, we efficiently decompose the
integer matrix M ∈ Zk×k using a version of the Smith Normal Form, effectively reducing to the case
that M is diagonal. For the diagonal case, we can intuitively treat the dimensions independently,
using the lexicographic ordering.
Our first lemma handles the 1-dimensional case, assuming the existence of an efficient oracle
for the CDF:
21
Lemma 3.18. Given a pseudo-distribution H supported on [a, b] ∩ Z, a, b ∈ Z, with CDF cH (x) =
P
i:a≤i≤x H(i) (which satisfies cH (b) = 1), and oracle access to a function c(x) so that |c(x) −
cH (x)| < ǫ/(10(b − a + 1)) for all x, we have the following: If there is a distribution P with
dTV (H, P) ≤ ǫ/3, there is a sampler for a distribution Q with dTV (P, Q) ≤ ǫ, using O(log(b + 1 −
a)+log(1/ǫ)) uniform random bits as input, and running in time O((D+1)(log(b+1−a))+log(1/ǫ)),
where D is the running time of evaluating the CDF c(x).
Proof. We begin our analysis by producing an algorithm that works when we are able to exactly
sample cH (x).
We can compute an inverse to the CDF dH : [0, 1] → [a, b] ∩ Z, at y ∈ [0, 1], using binary search,
as follows:
1. We have an interval [a′ , b′ ], initially [a − 1, b], with cH (a′ ) ≤ y ≤ cH (b′ ) and cH (a′ ) < cH (b′ ).
2. If b′ − a′ = 1, output dH (y) = b′ .
3. Otherwise, find the midpoint c′ = ⌊(a′ + b′ )/2⌋.
4. If cH (a′ ) < cH (c′ ) and y ≤ cH (c′ ), repeat with [a′ , c′ ]; else repeat with [c′ , b].
The function dH can be thought of as some kind of inverse to the CDF cH : [a − 1, b] ∩ Z → [0, 1]
in the following sense:
Claim 3.19. The function dH satisfies: For any y ∈ [0, 1], it holds cH (dH (y) − 1) ≤ y ≤ cH (dH (y))
and cH (dH (y) − 1) < cH (dH (y)).
Proof. Note that if we don’t have cH (a′ ) < cH (c′ ) and y ≤ cH (c′ ), then cH (c′ ) < y ≤ cH (b′ ). So,
Step 4 gives an interval [a′ , b′ ] which satisfies cH (a′ ) ≤ y ≤ cH (b′ ) and cH (a′ ) < cH (b′ ). The initial
interval [a − 1, b] satisfies these conditions since cH (a − 1) = 0 and cH (b) = 1. By induction, all
[a′ , b′ ] in the execution of the above algorithm have cH (a′ ) ≤ y ≤ cH (b′ ) and cH (a′ ) < cH (b′ ). Since
this is impossible if a′ = b′ , and Step 4 always recurses on a shorter interval, we eventually have
b′ − a′ = 1. Then, the conditions cH (a′ ) ≤ y ≤ cH (b′ ) and cH (a′ ) < cH (b′ ) give the claim.
Computing dH (y) requires O(log(b−a+1)) evaluations of cH , and O(log(b−a+1)) comparisons
of y. For the rest of this proof, we will use n = b − a + 1 to denote the support size.
Consider the random variable dH (Y ), for Y uniformly distributed in [0, 1], whose distribution
we will call Q′ . When dH (Y ) = x, we have cH (x − 1) ≤ Y ≤ cH (x), and so when Q′ (x) > 0, we
have Q′ (x) ≤ Pr [cH (x − 1) ≤ Y ≤ cH (x)] = cH (x) − cH (x − 1) = H(x). So, when H(x) > 0, we
have H(x) ≥ Q′ (x). But when H(x) ≤ 0, we have Q′ (x) =P0, since then cH (x) < cH (x − 1) and no
y has cH (x − 1) ≤ y ≤ cH (x). So, we have dTV (Q′ , H) = x:H(x)<0 −H(x) ≤ dTV (H, P) ≤ ǫ/3.
We now show how to effectively sample from Q′ . The issue is how to simulate a sample from
the uniform distribution on [0, 1] with uniform random bits. We do this by flipping coins for the
bits of Y lazily. We note that we will only need to know more than m bits of Y if Y is within 2−m
of one of the values of cH (x) for some x. By a union bound, this happens with probability at most
n2−m over the choice of Y. Therefore, for m > log2 (10n/ǫ), the probability that this will happen is
at most ǫ/10 and can be ignored.
Therefore, the random variable dH (Y ′ ), for Y ′ uniformly distributed on the multiples of 2−r in
[0, 1) for r = O(log n + log(1/ǫ)), has distribution Q′ that satisfies dTV (Q, Q′ ) ≤ ǫ/10. This means
that dTV (P, Q′ ) ≤ dTV (P, H) + dTV (H, Q) + dTV (Q, Q′ ) ≤ 9ǫ/10. That is, we obtain an ǫ-sampler
that uses O(log n + log(1/ǫ)) coin flips, O(log n) calls to cH (x), and has the desired running time.
22
We now need to show how this can be simulated without access to cH , and instead only having
access to its approximation c(x). The modification required is rather straightforward. Essentially,
we can run the same algorithm using c(x) in place of cH (x). We note that all comparisons with Y
will produce the same result, unless the chosen Y is between c(x) and cH (x) for some value of x.
Observe that because of our bounds on their difference, the probability of this occurring for any
given value of x is at most ǫ/(10n). By a union bound, the probability of it occurring for any x is
at most ǫ/10. Thus, with probability at least 1 − ǫ/10 our algorithm returns the same result that
it would have had it had access to cH (x) instead of c(x). This implies that the variable sampled
by this algorithm has variation distance at most ǫ/10 from what would have been sampled by our
other algorithm. Therefore, this algorithm samples a Q with dTV (P, Q) ≤ ǫ.
We next show that we can efficiently compute an appropriate CDF using the DFT. For the
1-dimensional case, this follows easily via a closed form expression. For the high-dimensional case,
we first obtain a closed form expression for the case that the matrix M is diagonal. We then reduce
the general case to the diagonal case, by using a Smith normal form decomposition.
Proposition 3.20. For H as in Theorem 3.15, we have the following:
(i) If
P k = 1, there is an algorithm to compute the CDF cH : [a, b] ∩ Z → [0, 1] with cH (x) =
i:a≤i≤x H(i) to any precision δ > 0, where a = m − ⌈M/2⌉ + 1 and b = m + ⌊M/2⌋,
M ∈ Z+ . The algorithm runs in time O(|T | log(1/δ)).
(ii) If M ∈ Zk×k is diagonal, there is an algorithm which computes
the CDF to any precision
P
δ > 0 under the lexicographic ordering ≤lex , i.e., cH (x) = y∈T :y≤lex x H(y). The algorithm
runs in time O(k2 |T | log(1/δ)).
k×k , there is an explicit ordering ≤ for which we can compute the CDF
(iii) For any M
g
P∈ Z
cH (x) =
H(y)
to
any
precision
δ
>
0.
This
computation can be done in time
y∈T :y≤g x
2
O(k |T | log(1/δ) + poly(k)).
In cases (ii) and (iii), we can also compute the embedding of the corresponding ordering onto the
integers [| det M |] = {1, 2, ..., | det M |}, i.e., we can give a monotone bijection f : S → [| det M |] for
which we can efficiently compute f and f −1 (i.e., with the same running time bound we give for
computing cH (x)).
Proof. Recall that the PMF of H at x ∈ S is given by the inverse DFT:
X
1
b
e(−ξ · x)H(ξ)
.
H(x) =
| det M |
(10)
ξ∈T
Proof of (i): For (i), the CDF is given by:
cH (x) =
1
M
X X
i:a≤i≤x ξ∈T
X
1 Xb
=
H(ξ)
e(−ξx)
M
i:a≤i≤x
ξ∈T
When ξ 6= 0, the term
have:
P
i:a≤i≤x e(−ξ
X
i:a≤i≤x
b
e(−ξx)H(ξ)
· x) is a geometric series. By standard results on its sum, we
e(−ξx) =
e(−ξa) − e(−ξ(x + 1))
.
1 − e(−ξ)
23
P
b
When ξ = 0, e(−ξ) = 1, and we get a≤i≤x e(−ξx) = i + 1 − a. In this case, we also have H(ξ)
= 1.
Putting this together we have:
X
1
e(−ξa)
−
e(−ξ(x
+
1))
b
.
cH (x) =
H(ξ)
i+1−a+
(11)
M
1 − e(−ξ)
ξ∈T \{0}
Hence, we obtain a closed form expression for the CDF that can be approximated to desired
precision in time O(|T | log(1/δ)).
Q
Proof of (ii): For (ii), we can write M = diag(Mi ), 1 ≤ i ≤ k, and S = ki=1 ([ai , bi ] ∩ Z), where
ai = mi − ⌈|Mi |/2⌉ + 1 and bi = mi + ⌊|Mi |/2⌋. With our lexicographic ordering, we have:
X
cH (x) =
H(y)
y∈S:y≤lex x
xX
1 −1
=
b2
X
y1 =a1 y2 =a2
+
xX
2 −1
b3
X
y2 =a2 y3 =a3
...
xX
k −1
+
···
···
bk
X
H(y)
yk =ak
bk
X
H(x1 , y2 , . . . , yk )
yk =ak
H(x1 , . . . , xk−1 , yk )
yk =ak
+ H(x) .
To avoid clutter in the notation, we define cH,i (x) to be one of these sums, i.e.,
cH,i (x)
def
=
xX
i −1
bi+1
X
yi =ai yi+1 =ai+1
=
···
xX
i −1
1
·
| det(M )| y =a
i
=
=
H(x1 , . . . , xi−1 , yi , . . . , yk )
yk =ak
bi+1
X
yi+1 =ai+1
···
bk X
X
yk =ak ξ∈T
b
−
H(ξ)e
i−1
X
j=1
ξ j xj −
!
k
X
j=i
ξj y j
bi
xX
i−1
k
i −1
X
X
X
Y
1
b
·
ξ j xj
e(−ξj yj )
H(ξ)e −
e(−ξi yi )
| det(M )|
yi =ai
j=1
j=i+1 yi =ai
ξ∈T
i−1
k
X
Y
X
1
b
ξj xj si (ai , xi − 1)
sj (aj , bj ) ,
H(ξ)e −
·
| det(M )|
ξ∈T
where si (a′i , b′i ) :=
i
bk
X
Pb′i
yi =a′i
j=1
j=i+1
e(−ξi yi ). As before, this is a geometric series, so either ξi = 0, when we
e(−ξ a′ )−e(−ξ(b′ +1))
i i
i
.
have si = b′i + 1 − a′i , or si =
1−e(−ξ)
We can thus evaluate cH,i in O(|T |k) arithmetic operations and so compute
cH (x) to desired
Q
accuracyPin O(|T |k2Q
log(1/δ)) time. We also note that f : S → {0, 1, ..., ( i |Mi |) − 1}, defined by
f (x) := i (xi − ai ) ij=1 Mj is a strictly monotone bijection, and that f and f −1 can be computed
Q
in time O(k). Now, cH (f −1 (y) is the CDF of the distribution on y ∈ {0, 1, ..., ( i |Mi |) − 1} whose
PMF is given by H(f −1 (y)).
24
Proof of (iii): We will reduce (iii) to (ii).
factorization of integer matrices:
To do this, we use Smith normal form, a canonical
Lemma 3.21 (See [Sto00], [Sto96])). Given any integer matrix M ∈ Zk×k , we can factorize M
as M = U · D · V , where U, D, V ∈ Zk×k with D diagonal and U ,V unimodular, i.e., | det(U )| =
| det(V )| = 1, and therefore U −1 , V −1 ∈ Zk×k . This factorization can be computed in time
poly(k) log max |Mi,j |.
i,j
Note that the Smith normal form satisfies additional conditions on D than those in Lemma
3.21, but we are only interested in finding such a decomposition where D is a diagonal integer
matrix.
We note that the integer lattices M Zk and U DZk are identical, since if x = M b for b ∈ Zk ,
then x = U Dc for c = V b ∈ Zk . For any ξ ∈ (M T )−1 Zk , ξ = (U T )−1 (D T )−1 (V T )−1 b, for b ∈ Z.
Then, if we take ν = U T ξ, we have ν ∈ (DV )−1 Zk = D −1 Zk .
Hence, we can re-write (10) as follows:
H(x) =
X
1
T −1
b
e −νU −1 x H((U
) ν) .
| det M |
T
ν∈U T
Since U T T ⊆ D −1 Zk , substituting y = U −1 x almost gives us the conditions which would allow us to
apply (ii). The issue is that for x ∈ m + M (−1/2, 1/2]k , we have U −1 X ∈ U −1 m + DV (−1/2, 1/2]k ,
but we do not necessarily have U −1 X ∈ U −1 m+D(−1/2, 1/2]k . The following claim gives the details
of how to change the fundamental domain:
Claim 3.22. Given a non-singular M ∈ Zk×k and m, x ∈ Rk , then x′ = x + M R M −1 (m − x)
is the unique x′ ∈ m + M (−1/2, 1/2]k with x − x′ ∈ M Zk , where R(x) is x with each coordinate
rounded to the nearest integer, rounding half integers up, i.e., (R(x))i := 21 + ⌈xi − 21 ⌉.
def
So we take y = g(x) = U −1 x + DR (U D)−1 m − (U D)−1 x , which using Claim 3.22 has
g(x) ∈ U −1 m + D(−1/2, 1/2]k . We need the inverse function of g : m + M (−1/2, 1/2]k → U −1 m +
D(−1/2, 1/2]k . Note that g−1 (y) = U y + D −1 R(U −1 m − y), which again by Claim 3.22 is in
m + M (−1/2, 1/2]k .
So, if y = g(x), since | det(M )| = | det(U )| · | det(D)| · | det(V )| = | det(D)|, we have:
H(g −1 (y)) =
X
1
T −1
b
e(ν · y)H((U
) ν) .
| det(D)|
T
(12)
ν∈U T
Now, we can take H(g−1 (y) to be a function of y supported on U −1 m+D(−1/2, 1/2]k with a sparse
DFT modulo D supported on U T T ⊆ D −1 Zk . At this point, we can apply the algorithm of (ii),
which gives a way to compute the CDF of H(g −1 (y)) with respect to the lexicographic ordering on
y. Note that g and g−1 can be computed in time poly(k), or more precisely the running time of
matrix multiplication and inversion.
For the
P ordering on x ∈ S, which has x1 ≤g x2 when g(x1 ) ≤lex g(x2 ), we can compute
cH (x) = y∈S:y≤g x H(y) by applying the algorithm in (ii) to the function given in (12) applied at
g(x). So, we can compute cH (x) in time O(k2 |T | + poly(k)). Again, the function given by f (g(x)),
where f is as in (ii) is a monotone bijection from S to {0, 1, . . . , det(M ) − 1}, and we can calculate
this function and its inverse in time poly(k).
Now we can prove the main theorem of this subsection.
25
Proof of Theorem 3.15. By Proposition 3.20 (iii), there is a bijection f which takes the support S
of H to the integers {0, 1, . . . , |S| − 1}, and we can efficiently calculate the CDF of the distribution
considered on this set of integers. So, we can apply Lemma 3.18 to this CDF on this distribution.
This gives us an ǫ-sampler for this distribution, which we can then apply f −1 to each sample to
get an ǫ-sampler for H. To find the time it takes to compute each sample, we need to substitute
D = O(poly(k) + k2 |T | log(| det(M )|/ǫ)) from the running time of the CDF in Proposition 3.20 (iii)
into the bound in Lemma 3.18, yielding
O(log(| det(M )|) log(| det(M )|/ǫ) · |T | · poly(k))
time. This completes the proof.
3.3 Using our Learning Algorithm to Obtain a Cover As an application of our learning
algorithm in Section 3.1, we provide a simple proof that the space Mn,k of all (n, k)-PMDs has an ǫO(k)
2
cover under the total variation distance of size nO(k ) ·2O(k log(1/ǫ))
. Our argument is constructive,
yielding an efficient algorithm to construct a non-proper ǫ-cover of this size.
Two remarks are in order: (i) the non-proper cover construction in this subsection does not
suffice for our algorithmic applications of Section 4. These applications require the efficient construction of a proper ǫ-cover plus additional algorithmic ingredients. (ii) The upper bound on the
cover size obtained here is nearly optimal, as follows from our lower bound in Section 4.5.
The idea behind using our algorithm to obtain a cover is quite simple. In order to determine
its hypothesis, H, our algorithm Efficient-Learn-PMD requires the following quantities:
b ∈ Rk×k satisfying the conclusions of Lemma 3.4.
• A vector µ
b ∈ Rk and a PSD matrix Σ
P
b
b
b
• Values of the DFT H(ξ),
for all ξ ∈ T, so that
ξ∈T |H(ξ) − P(ξ)| < ǫ/10. Recall that
def
T = ξ = (M T )−1 v | v ∈ Zk ∧ kvk2 ≤ C 2 k2 ln(k/ǫ) .
Given this information, the analysis in Section 3.1 carries over immediately. The algorithm
Efficient-Learn-PMD works by estimating the mean and covariance using samples, and then
b
taking H(ξ)
to be the sample Fourier transform. If we instead guess the values of these quantities
using an appropriate discretization, we obtain an ǫ-cover for Mn,k . More specifically, we discretize
the above quantities as follows:
def
• To discretize theP
mean vector, we consider a 1-cover Y1 of the set Y = {y = (y1 , . . . , yk ) ∈
Rk : (yi ≥ 0) ∧ ( ki=1 yi = k)} with respect to the L2 norm. It is easy to construct such a
cover with size |Y1 | ≤ O(n)k .
def
• To discretize the covariance matrix, we consider a 1/2-cover S1/2 of the set of matrices S =
{A ∈ Rk×k : (A 0) ∧ kAk2 ≤ n}, with respect to the spectral norm k · k2 . Note that we can
construct such a 1/2-cover with size |S1/2 | ≤ (4n + 1)k(k+1)/2 . This is because any maximal
1/4-packing of this space (i.e., a maximal set of matrices in S with pairwise distance under the
spectral norm at least 1/4) is such a 1/2-cover. Observe that for any such maximal packing,
the balls of radius 1/4 centered at these points are disjoint and contained in the ball (under
the spectral norm) about the origin of radius (n + 1/4). Since the ball of radius (n + 1/4) has
volume (4n + 1)k(k+1)/2 times as much as the ball of radius 1/4, a simple volume argument
completes the proof.
26
• Finally, to discretize the Fourier transform, we consider a δ-cover Cδ of the unit disc on the
complex plane C, with respect to the standard distance on C, where
def
δ = ǫ(2C 2 k2 log(k/ǫ))−k /10 = ǫ/(10t) ≤ ǫ/(10|T |).
def
We note that t = (2C 2 k2 log(k/ǫ))k is an upper bound on |T |.
We claim that there is an ǫ-cover of the space of (n, k)-PMDs indexed by Y1 × S1/2 × Cδt . Such
b be
a cover is clearly of the desired size. The cover is constructed as follows: We let µ
b and Σ
the selected elements from Y1 and S1/2 , respectively. We use these elements to define the matrix
b
M ∈ Zk×k as in the algorithm description. We then use our elements of Cδ as the values of H(ξ)
for ξ ∈ T (noting that |T | ≤ t).
We claim that for any (n, k)-PMD P there exists a choice of parameters, so that the returned
distribution H is within total variation distance ǫ of P. We show this as follows: Let µ and Σ be
the true mean and covariance matrix of P. We have that µ ∈ Y and that Σ ∈ S. Therefore, there
b ∈ S1/2 so that |µ − µ
b −I/2. It is easy to see that these
exist µ
b ∈ Y1 and Σ
b|2 ≤ 1 and I/2 Σ − Σ
conditions imply the conclusions of Lemma 3.4. Additionally, we can P
pick elements of Cδ in order
b
b
b
b
to make |H(ξ)
− P(ξ)|
< ǫ/(10|T |) for each ξ ∈ T. This will give that ξ∈T |H(ξ)
− P(ξ)|
< ǫ/10.
In particular, the hypothesis H indexed by this collection of parameters will be within variation
distance ǫ of P. Hence, the set we have constructed is an ǫ-cover, and our proof is complete.
4
Efficient Proper Covers and Nash Equilibria in Anonymous Games
In this section, we give our efficient proper cover construction for PMDs, and our EPTAS for
computing Nash equilibria in anonymous games. These algorithmic results are based on new
structural results for PMDs that we establish. The structure of this section is as follows: In
Section 4.1, we show the desired sparsity property of the continuous Fourier transform of PMDs,
and use it to prove our robust moment-matching lemma. Our dynamic-programming algorithm
for efficiently constructing a proper cover relies on this lemma, and is given in Section 4.2. By
building on the proper cover construction, in Section 4.3 we give our EPTAS for Nash equilibria
in anonymous games. In Section 4.4, we combine our moment-matching lemma with recent results
in algebraic geometry, to show that any PMD is close to another PMD with few distinct CRV
components. Finally, in Section 4.5 we prove out cover size lower bound.
4.1 Low-Degree Parameter Moment Closeness Implies Closeness in Variation Distance In this subsection, we establish the sparsity of the continuous Fourier transform of PMDs,
and use it to prove our robust moment-matching lemma, translating closeness in the low-degree
parameter moments to closeness in total variation distance.
At a high-level, our robust moment-matching lemma (Lemma 4.6) is proved by combining
the sparsity of the continuous Fourier transform of PMDs (Lemma 4.2) with very careful Taylor
approximations of the logarithm of the Fourier transform (log FT) of our PMDs. For technical
reasons related to the convergence of the log FT, we will need one additional property from our
PMDs. In particular, we require that each component k-CRV has the same most likely outcome.
This assumption is essentially without loss of generality. There exist at most k such outcomes,
and we can express an arbitrary PMD as a sum of k independent component PMDs whose k-CRV
components satisfy this property. Formally, we have the following definition:
Definition 4.1. We say that a k-CRV W is j-maximal, for some
P j ∈ [k], if for all ℓ ∈ [k] we
have Pr[W = ej ] ≥ Pr[W = eℓ ]. We say that an (n, k)-PMD X = ni=1 Xi is j-maximal, for some
j ∈ [k], if for all 1 ≤ i ≤ n Xi is a j-maximal k-CRV.
27
Pk
i
i
Any
(n,
k)-PMD
X
can
be
written
as
X
=
i=1 X , where X is an i-maximal (ni , k)-PMD,
P
with i ni = n. For the rest of this intuitive explanation, we focus on two (n, k)-PMDs X, Y that
are promised to be i-maximal, for some i ∈ [k].
b Yb have roughly the same effective support, we also assume that they
To guarantee that X,
have roughly the same variance in each direction. We will show that if the low-degree parameter
moments of X and Y are close to each other, then X and Y are close in total variation distance.
We proceed by partitioning the k-CRV components of our PMDs into groups, based on their
maximum probability element ej , with j 6= i. The maximum probability of a k-CRV quantifies
its maximum contribution to the variance of the PMD in some direction. Roughly speaking, the
smaller this contribution is, the fewer terms in the Taylor approximation are needed to achieve
a given error. More specifically, we consider three different groups, partitioning the component
k-CRVs into ones with small, medium, and large contribution to the variance in some direction.
For the PMD (defined by the CRVs) of the first group, we only need to approximate the first 2
parameter moments. For the PMD of the second group, we approximate the low-degree parameter
moments up to degree Ok (log(1/ǫ)/ log log(1/ǫ)). Finally, the third group is guaranteed to have
very few component k-CRVS, hence we can afford to approximate the individual parameters.
To quantify the above, we need some more notation and definitions. To avoid clutter in the
notation, we focus without loss of generality
on the case i = k, i.e., our PMDs are k-maximal. For
P
a k-maximal (n, k)-PMD, X, let X = ni=1 Xi , where the Xi is a k-CRV with pi,j = Pr[Xi = ej ]
P
for 1 ≤ i ≤ n and 1 ≤ j ≤ k. Observe that kj=1 pi,j = 1, for 1 ≤ i ≤ n, hence the definition of
k-maximality implies that pi,k ≥ 1/k for all i. Note that thePj th component of the random vector
X is a PBD with parameters pi,j , 1 ≤ i ≤ n. Let sj (X) = ni=1 pi,j be the expected value of the
j th component of X. We can assume that sj (X) ≥ ǫ/k, for all 1 ≤ j ≤ k − 1; otherwise, we can
remove the corresponding coordinates and introduce an error of at most ǫ in variation distance.
th
Note that, for j 6= k, the variance
Pnof the j coordinate of X is in [sj (X)/2, sj (X)]. Indeed,
the aforementioned variance equals i=1 pi,j (1 − pi,j ), which is clearly at most sj (X). The other
direction follows by observing that, for all j 6= k, we have 1 ≥ pi,k + pi,j ≥ 2pi,j , or pi,j ≤ 1/2, where
we again used the k-maximality of X. Therefore, by Bernstein’s inequality and a union bound,
there is a set S ⊆ [n]k of size
|S| ≤
k−1
Y
1/2
1 + 12sj (X)
j=1
k−1
Y
(k−1)
(1 + 12sj (X)1/2 ) ,
·
ln(2k/ǫ) ≤ O log(k/ǫ)
j=1
so that X lies in S with probability at least 1 − ǫ.
We start by showing that the continuous Fourier transform of a PMD is approximately sparse,
namely it is effectively supported on a small set T. More precisely, we prove that there exists a
set T in the Fourier domain such that the integral of the absolute value of the Fourier transform
outside T multiplied by the size of the effective support |S| of our PMD is small.
Lemma 4.2 (Sparsity of the FT for PMDs). Let X be k-maximal (k, n)-PMD with effective support
S. Let
n
o
def
T = ξ ∈ [0, 1]k : [ξj − ξk ] < Ck(1 + 12sj (X))−1/2 log1/2 (1/ǫ) ,
where [x] is the distance between x and the nearest integer, and C > 0 is a sufficiently large universal
constant. Then, we have that
Z
b ≪ ǫ/|S|.
|X|
T
Proof. To prove the lemma, we will need the following technical claim:
28
Claim 4.3. For all ξ = (ξ1 , . . . , ξk ) ∈ [0, 1]k , for all 1 ≤ i ≤ n and 1 ≤ j ≤ k − 1, it holds:
bi (ξ)| = exp(−Ω(pi,j [ξj − ξk ]2 /k)) .
|X
Proof. The claim follows from the following sequence of (in-)equalities:
k
k
X
X
X
ci (ξ)|2 =
|X
pi,j e(ξj )
pi,j ′ e(−ξj ′ ) =
pi,j pi,j ′ e(ξj − ξj ′ )
j ′ =1
j=1
=
X
j,j ′
=1 −
j,j ′
pi,j pi,j ′ cos(2π(ξj − ξj ′ )) = 1 −
X
1 − Ω([ξj − ξj ′ ])2
pi,j pi,j ′
j6=j ′
= exp −Ω
X
j6=j ′
X
j6=j ′
pi,j pi,j ′ 1 − cos(2π(ξj − ξj ′ ))
(by Claim 3.11 with δ = 12 )
pi,j ′ pi,j [ξj − ξj ′ ]2
X
= exp −Ω
pi,j pi,k [ξj − ξk ]2
j<k
= exp −Ω pi,j [ξj − ξk ]2 /k
,
where the last lines uses the fact pi,k ≥ 1/k, which follows from k-maximality.
As a consequence of Claim 4.3, we have that
b
|X(ξ)|
=
n
Y
i=1
ci (ξ)| = exp(−Ω(sj (X)[ξj − ξk ]2 /k)) .
|X
(13)
R
b we proceed as follows: For ℓ ∈ Z+ ,
Let T = [0, 1]k \ T be the complement of T . To bound T |X|,
we define the sets
n
o
T ℓ = ξ : max([ξj − ξk ]/(Ck(1 + 12sj (X))−1/2 log1/2 (1/ǫ))) ∈ [2ℓ , 2ℓ+1 ] ,
ℓ
b
and observe that T ⊆ ∪ℓ∈Z+ T ℓ . Now, Equation (13) implies that for ξ ∈ T ℓ it holds |X(ξ)|
≤ ǫ10k2 ,
where we used the assumption that the constant
R that the
Q C is sufficiently large. It is easy to see
b from
volume of T ℓ is at most O(2ℓ k log1/2 (1/ǫ))k j<k (1 + 12sj (X))−1/2 . We can bound T |X|
b within T ℓ times the volume of T ℓ , namely
above by the sum over ℓ of the maximum value of |X|
Z
T
b ≤
|X|
∞ Z
X
ℓ=0
Tℓ
b ≤
|X|
∞
X
ℓ=0
!
b
sup |X(ξ)|
Vol(T ℓ ) ≤ ǫk
ξ∈T ℓ
Y
j<k
(1 + 12sj (X))−1/2 ≪ ǫ/|S|.
This completes the proof of Lemma 4.2.
We now use the sparsity of the Fourier transform to show that if two k-maximal PMDs, with
similar variances in each direction, have Fourier transforms that are pointwise sufficiently close to
each other in this effective support, then they are close to each other in total variation distance.
29
Lemma 4.4. Let X and Y be k-maximaln (k, n)-PMDs, satisfying 1/2 ≤ (1+sj (X))/(1+sj (Y )) ≤o2
def
for all j, 1 ≤ j ≤ k − 1. Let T = ξ ∈ [0, 1]k : [ξj − ξk ] < Ck(1 + 12sj (X))−1/2 log1/2 (1/ǫ) ,
where [x] is the distance between x and the nearest integer, and C > 0 is a sufficiently large
b
universal constant. Suppose that for all ξ ∈ T it holds |X(ξ)
− Yb (ξ)| ≤ ǫ(Ck log(k/ǫ))−2k . Then,
dTV (X, Y ) ≤ ǫ.
Proof. We start with an intuitive explanation of the proof. Since 1 + sj (X), 1 + sj (Y ) are within a
factor of 2 for 1 ≤ j ≤ k − 1, it follows from the above
both effectively supported
Qk−1that X and Y are
on a set S ⊆ [n]k of size |S| ≤ O (log(k/ǫ))k−1 · j=1
1 + 12sj (X)1/2 . Therefore, to prove the
lemma, it is sufficient to establish that kX − Y k∞ ≤ O(ǫ/|S|).
b and
We prove this statement in two steps by analyzing the continuous Fourier transforms X
b
Y . The first step of the proof exploits the fact that the Fourier transforms of X and Y are
each essentially supported on the set T of the lemma statement. Recalling the assumptionR that
b
(1 + sj (X))/(1 + sj (Y )) ∈ [1/2, 2], 1 ≤ j ≤ k − 1, an application of Lemma 4.2 yields that T |X|
R
and T |Yb | are both at most ǫ/|S|. Thus, we have that
Z
Z
Z
b − Yb | ≤
b +
|X
|X|
|Yb | ≪ ǫ/|S| .
T
T
T
b
In the second step of the proof,R we use theR assumption that the absolute difference |X(ξ)−
Yb (ξ)|,
b
b
b
b
ξ ∈ T , is small, and the fact that T |X| and T |Y | are individually small, to show that kX − Y k1 ≤
b − Yb k1 combined with the concentration
O(ǫ/|S|). The straightforward inequality kX − Y k∞ ≤ kX
of X, Y completes the proof.
b − Yb k1 , it suffices to show that the integral
Given the aforementioned, in order to bound kX
b
b
over T is small. By the assumption of the lemma, we have
R that |X(ξ) − Y (ξ)| is point-wise at most
−2k
b − Yb | by multiplying this quantity by
ǫ(Ck log(k/ǫ))
over T . We obtain an upper bound on T |X
Q
the volume of T . Note that the volume of T is at most O(k log1/2 (1/ǫ))k−1 j<k (1+ 12sj (X))−1/2 ).
Hence,
Z
Y
b − Yb | ≤ ǫ · Θ(log(k/ǫ))−k ·
|X
(1 + 12sj (X))−1/2 ≪ ǫ/|S|.
T
j<k
b − Yb k1 = O(ǫ/|S|), which implies that the
Combining the above, we get that kX − Y k∞ ≤ kX
L1 distance between X and Y over S is O(ǫ). The contribution of S to the L1 distance is at
most ǫ, since both X and Y are in S with probability at least 1 − ǫ. This completes the proof of
Lemma 4.4.
We use this lemma as technical tool for our robust moment-matching lemma. As mentioned in
the beginning of the section, we will need to handle separately the component k-CRVs that have
a significant contribution to the variance in some direction. This is formalized in the following
definition:
P
Definition 4.5. Let X be a k-maximal (n, k)-PMD with X = ni=1 Xi and 0 < δ ≤ 1. For a given
ℓ ∈ [n], we say that a particular component k-CRV Xℓ , with pℓ,j =p
Pr[Xℓ = ej ], is δ-exceptional if
there exists a coordinate j, with 1 ≤ j ≤ k − 1, such that pℓ,j ≥ δ · 1 + sj (X). We will denote by
E(δ, X) ⊆ [n] the set of δ-exceptional components of X.
Recall that the variance of the j th coordinate of X is in [sj (X)/2, sj (X)]. Therefore, the above
definition states that the j th coordinate of Xi has probability mass which is at least a δ-fraction of
the standard deviation across the j th coordinate of X.
30
We remark that for any (n, k)-PMD X, at most k/δ2 of its component k-CRVs are δ-exceptional.
To see this, we observe that the number of δ-exceptional components is at most k − 1 times
the number of (δ, j)-exceptional components, i.e., the k-CRVs Xi which are δ-exceptional for the
same value of j. We claim that for any j, 1 ≤ j ≤ k − 1, the number of (δ, j)-exceptional
components
is at most 1/δ2 . Indeed, let
corresponding
P
PEj ⊆ [n] denote the P
Pset. Then,Pwe have
that i∈Ej p2i,j ≥ δ2 |Ej |sj (X) = δ2 |Ej | ni=1 pi,j . Noting that i∈Ej p2i,j ≤ ni=1 p2i,j ≤ ni=1 pi,j ,
we get that δ2 |Ej | ≤ 1, thus yielding the claim.
We now have all the necessary ingredients for our robust moment-matching lemma. Roughly
speaking, we partition the coordinate k-CRVs of our k-maximal PMDs into three groups. For
appropriate values 0 < δ1 < δ2 , we have: (i) k-CRVs that are not δ1 -exceptional, (ii) k-CRVs
that are δ1 -exceptional, but not δ2 -exceptional, and (iii) δ2 -exceptional k-CRVs. For group (i),
we will only need to approximate the first two parameter moments in order to get a good Taylor
approximation, and for group (ii) we need to approximate as many as Ok (log(1/ǫ)/ log log(1/ǫ))
degree parameter moments. Group (iii) has Ok (log3/2 (1/ǫ)) coordinate k-CRVs, hence we simply
approximate the individual (relatively few) parameters each to high precision. Formally, we have:
Lemma 4.6. Let X and Y be k-maximal (n, k)-PMDs, satisfying 1/2 ≤ (1+sj (X))/(1+sj (Y )) ≤ 2
for all j, 1 ≤ j ≤ k − 1. Let C be a sufficiently large constant. Suppose that the component kCRVs of X and Y can be partitioned into three groups, so that X = X (1) + X (2) + X (3) and
Y = Y (1) + Y (2) + Y (3) , where X (t) and Y (t) , 1 ≤ t ≤ 3, are PMDs over the same number of
k-CRVs. Additionally assume the following: (i) for t ≤ 2 the random variables X (t) and Y (t)
def
def
have no δt -exceptional components, where δ1 = δ1 (ǫ) = ǫ(Ck log(k/ǫ))−3k−3 and δ2 = δ2 (ǫ) =
k−1 log−3/4 (1/ǫ), and (ii) there is a bijection between the component k-CRVs of X (3) with those in
Y (3) , so that corresponding k-CRVs have total variation distance at most ǫ/3n3 , where n3 is the
number of such k-CRVs.
Finally, suppose that for t ≤ 2, and all vectors m ∈ Zk+ with mk = 0 and |m|1 ≤ Kt it holds
def
|Mm (X (t) ) − Mm (Y (t) )|(2k)|m|1 ≤ γ = γ(ǫ) = ǫ(Ck log(k/ǫ))−2k−1 ,
where K1 = 2 and K2 = K2 (ǫ) = C(log(1/ǫ)/ log log(1/ǫ) + k). Then dTV (X, Y ) ≤ ǫ.
P
Proof. First, note that dTV (X, Y ) ≤ 3t=1 dTV (X (t) , Y (t) ), so it suffices to show that dTV (X (t) , Y (t) ) =
ǫ/3, for t = 1, 2, 3. This holds trivially for t = 3, by assumption. To prove the statement for t = 1, 2,
d
d
(t) and Y
(t) are point-wise close on the set T , namely
by Lemma 4.4, it is sufficient to show that X
d
d
that for all ξ ∈ T it holds |X (t) (ξ) − Y (t) (ξ)| ≤ ǫ(Ck log(k/ǫ))−2k . To show this, we show separately
d
d
(t) is close to Y
(t) for each t = 1, 2.
that X
P
(t)
Let X = i∈At Xi , where At ⊆ [n] with |At | = nt . We have the following formula for the
31
Fourier transform of X (t) :
d
(t) (ξ) =
X
k
YX
e(ξj )pi,j
i∈At j=1
= e(nt ξk )
Y
i∈At
k−1
X
1 −
(1 − e(ξj − ξk ))pi,j
= e(nt ξk ) exp
j=1
X
i∈At
log 1 −
X
= e(nt ξk ) exp −
∞
X
i∈At ℓ=1
k−1
X
j=1
1
ℓ
(1 − e(ξj − ξk ))pi,j
ℓ
(1 − e(ξj − ξk ))pi,j
k−1
X
j=1
k−1
Y
X
|m|1
1
(1 − e(ξj − ξk ))mj .
Mm (X (t) )
= e(nt ξk ) exp −
m |m|1
k−1
(14)
j=1
m∈Z+
(t) . To prove the lemma, we will show that, for all ξ ∈ T , the two
An analogous formula holds for Yd
corresponding expressions inside the exponential of (14) agree for X (t) and Y (t) , up to a sufficiently
small error.
We first deal with the terms with |m|1 ≤ Kt , t = 1, 2. By the statement of the lemma, for any
two such terms we have that |Mm (X (t) ) − Mm (Y (t) )| ≤ (2k)−|m|1 · ǫ(Ck log(k/ǫ))−2k . Hence, for
any ξ ∈ [0, 1]k , the contribution of these terms to the difference is at most
X
|m|1
−2k−1
ǫ(Ck log(k/ǫ))
(2k)−|m|1 2|m|1 ≤ Kt ǫ(Ck log(k/ǫ))−2k−1 ≤ ǫ(Ck log(k/ǫ))−2k .
m
k−1
m∈Z+
, |m|1 ≤Kt
To deal with the remaining terms, we need the following technical claim:
k−1
with |m|1 ≥ 2, we have that
Claim 4.7. Let X (t) be as above. For t ≤ 2 and m ∈ Z+
|Mm (X (t) )|
k−1
Y
j=1
mj
|m| −2
(1 + sj (X))−1/2 log1/2 (1/ǫ)
≤ log|m|1 /2 (1/ǫ) · δt 1 .
Qk−1 mj
P
pi,j . Thus, the claim is equivalent to
Proof. By definition we have that Mm (X (t) ) = i∈At j=1
showing that
mj
X k−1
Y
|m| −2
pi,j · (1 + sj (X))−1/2 (X)
≤ δt 1 .
i∈At j=1
Since, by definition, X (t) does not contain any δt -exceptional k-CRV components, we have that for
all i ∈ At and all j ∈ [k − 1] it holds pi,j · (1 + sj (X))−1/2 (X) ≤ δt . Now observe that decreasing any
component of m by 1 decreases the left hand side of the above by a factor of at least δt . Therefore,
it suffices to prove the desired inequality for |m|1 = 2, i.e., to show that
X
pi,j1 (1 + sj1 (X))−1/2 pi,j2 (1 + sj2 (X))−1/2 ≤ 1.
i∈At
32
Indeed, the above inequality holds true, as follows from an application of the Cauchy-Schwartz
inequality, and the fact that
X
i∈At
p2i,j
≤
n
X
pi,j = sj (X)≤ sj (X) + 1.
i=1
This completes the proof of Claim 4.7.
Now, for ξ ∈ T , the contribution to the exponent of (14), coming from terms with |m|1 > Kt ,
is at most
k−1
X
X
Y
X
ℓ
1
O([ξj − ξk ])mj ≤
Mm (X (t) )
(15)
kℓ · logℓ/2 (1/ǫ) · δtℓ−2 .
ℓ
m
k−1
ℓ>Kt m∈Z
+
j=1
:|m|1 =ℓ
ℓ>Kt
Equation (15) requires a few facts to be justified. First, we use the multinomial identity
(k −
1)ℓ .
We also require the fact that
P
ℓ
m∈Zk−1
+ :|m|1 =ℓ m
|1 − e(ξj − ξk )| ≤ O([ξj − ξk ]) ,
ξ ∈ [0, 1]T , and recall that [ξj − ξk ] < Ck(1 + 12sj (X))−1/2 log1/2 (1/ǫ), for ξ ∈ T. Combining the
above with Claim 4.7 gives (15).
Finally, we claim that
X
kℓ · logℓ/2 (1/ǫ) · δtℓ−2 ≤ ǫ(Ck log(k/ǫ))−2k ,
ℓ>Kt
where the last inequality holds for both t = 1, 2, as can be readily verified from the definition
of δ1 , δ2 , K1 , K2 . Combining with the bounds for smaller |m|1 , we get that the absolute difference
d
d
(t) and Y
(t) on T is at most ǫ(Ck log(k/ǫ))−2k . Therefore, Lemma 4.2 implies that
between X
dTV (X (t) , Y (t) ) ≤ ǫ. A suitable definition of C in the statement of the lemma to make this ǫ/3
completes the proof of Lemma 4.6.
Remark 4.8. We note that the quantitive statement of Lemma 4.6 is crucial for our algorithm:
(i) The set of non δ1 -exceptional components can contain up to n k-CRVs. Since we only need to
approximate only the first 2 parameter moments for this set, this only involves poly(n) possibilities.
(ii) The set of δ1 -exceptional but not δ2 -exceptional k-CRVs has size O(k/δ12 ), which is independent
of n. In this case, we approximate the first Ok (log(1/ǫ)/ log log(1/ǫ)) parameter moments, and the
total number of possibilities is independent of n and bounded by an appropriate quasipolynomial
function of 1/ǫ. (ii) The set of δ2 -exceptional components is sufficiently small, so that we can afford
to do a brute-force grid over the parameters.
4.2 Efficient Construction of a Proper Cover As a warm-up for our proper cover algorithm,
we use the structural lemma of the previous section to show the following upper bound on the cover
size of PMDs.
Proposition 4.9. For all n, k ∈ Z+ , k > 2, and ǫ > 0, there exists an ǫ-cover of the set of
3
k−1
(n, k)-PMDs of size nO(k ) (1/ǫ)O(k log(1/ǫ)/ log log(1/ǫ)) .
33
=
Remark 4.10. We remark that, for the sake of simplicity, we have not optimized the dependence
of our cover upper bound on the parameter n. With a slightly more careful argument, one can
2
k−1
easily obtain a cover size upper bound nO(k ) (1/ǫ)O(k log(1/ǫ)/ log log(1/ǫ)) . On the other hand, the
asymptotic dependence of our upper bound on the error parameter ǫ is optimal. In Section 4.5, we
k−1
show a lower bound of (1/ǫ)Ωk (log(1/ǫ)/ log log(1/ǫ)) .
P
Proof of Proposition 4.9. Let X be an arbitrary (n, k)-PMD. We can write X as ki=1 X i , where
P
X i is an i-maximal (n(i) , k)-PMD, where ki=1 n(i) = n. By the subadditivity of the total variation
distance for independent random variables, it suffices to show that the set of i-maximal (n, k)-PMDs
2
k−1
has an ǫ/k-cover of size nO(k ) (1/ǫ)O(k log(k/ǫ)/ log log(k/ǫ)) .
To establish the aforementioned upper bound on the cover size of i-maximal PMDs, we focus
without loss of generality on the case i = k. The proof proceeds by an appropriate application of
Lemma 4.6 and a counting argument. The idea is fairly simple: for a k-maximal (n, k)-PMD X,
we start by approximating the means sj (X), 1 ≤ j ≤ k − 1, within a factor of 2, and then impose
an appropriate grid on its low-degree parameter moments.
We associate to such a k-maximal (n, k)-PMD X the following data, and claim that if X and
Y are two k-maximal PMDs with the same data, then their total variational distance is at most
def
ǫ′ = ǫ/k. An ǫ′ -cover for the set of k-maximal (n, k)-PMDs can then be obtained by taking one
def
representative X for each possible setting of the data in question. Let us denote δ1′ = δ1 (ǫ′ ),
def
def
def
def
δ2′ = δ2 (ǫ′ ), γ ′ = γ(ǫ′ ), K1′ = K1 = 2, and K2′ = K2 (ǫ′ ), where the functions δ1 (ǫ), δ2 (ǫ), γ(ǫ),
and K2 (ǫ) are defined in the statement of Lemma 4.6.
In particular, for any X, we partition the coordinates of [n] into the sets A1 = E(δ1′ , X),
A2 = E(δ2′ , X) \ A1 , and A3 = E(δ2′ , X). We use these subsets to define X (1) , X (2) and X (3) on
n1 , n2 , n3 k-CRVs respectively.
Now, to X we associate the following data:
• n1 , n2 , n3 .
• The nearest integer to log2 (sj (X) + 1) for each j, 1 ≤ j ≤ k − 1.
• The nearest integer multiple of γ ′ /(2k)|m|1 to each of the Mm (X (1) ) for |m|1 ≤ 2.
• The nearest integer multiple of γ ′ /(2k)|m|1 to Mm (X (2) ) for |m|1 ≤ K2′ .
• Rounding of each of the pi,j for i ∈ A3 to the nearest integer multiple of ǫ′ /(kn3 ).
First, note that if X and Y are k-maximal (n, k)-PMDs with the same associated data, then they
are partitioned as X = X (1) + X (2) + X (3) , Y = Y (1) + Y (2) + Y (3) , where X (t) , Y (t) , t ≤ 2,
have no δt′ -exceptional variables and have the same number of component k-CRVs. Furthermore,
1 + sj (X) and 1 + sj (Y ) differ by at most a factor of 2 for each j. We also must have that
|Mm (X (t) ) − Mm (Y (t) )|(2k)|m|1 ≤ γ ′ for |m|1 ≤ Kt′ , and there is a bijection between the variables in
X (3) and those in Y (3) so that corresponding variables differ by at most ǫ′ /(kn3 ) in each parameter
(and, thus, differ by at most ǫ′ /n3 in total variation distance). Lemma 4.6 implies that if X and
Y have the same data, then dTV (X, Y ) ≤ ǫ′ . Hence, this does provide an ǫ′ -cover for the set of
k-maximal (n, k)-PMDs.
We are left to prove that this cover is of the appropriate size. To do that, we need to prove a
bound on the number of possible values that can be taken by the above data. We have at most
n choices for each ni , and O(log(n)) choices for each of the k rounded values of log2 (sj (X) +
1) (since each is an integer between 0 and log2 (n) + 1). X (1) has O(k2 ) parameter moments
with |m|1 ≤ 2, and there are at most O(kn/γ ′ ) options for each of them (since each parameter
34
moment is at most n). There are O((k + K2′ )k−1 ) parameter moments of X (2) that need to be
considered. By Claim 4.7, each such parameter moment has magnitude at most O(k/δ1′ 2 ), and, by
′
our aforementioned rounding, needs to be evaluated to additive accuracy at worst γ ′ /(2k)K2 . Finally,
2
note that n3 = |A3 | ≤ k/δ2′ , since the coordinates of A3 are δ2′ -exceptional under X. Each of the
corresponding O(k2 /δ2′ 2 ) parameters pi,j for i ∈ A3 need to be approximated to precision ǫ′ /(kn3 ).
We remark that the number of such parameters is less than O(k log(1/ǫ′ )/ log log(1/ǫ′ ))k−1 , since
k ≥ 3. Putting this together, we obtain that the number of possible values for this data is at most
2
′
′ k−1
nO(k ) (1/ǫ′ )O(k log(1/ǫ )/ log log(1/ǫ )) . This completes the proof of Proposition 4.9.
The proof of Proposition 4.9 can be made algorithmic using Dynamic Programming, yielding
an efficient construction of a proper ǫ-cover for the set of all (n, k)-PMDs.
Theorem 4.11. Let S1 , S2 , . . . , Sn be sets of k-CRVs. Let S be the set of (n, k)-PMDs of the form
P
n
ℓ=1 Xℓ , where Xℓ ∈ Sℓ . There exists an algorithm that runs in time
3
nO(k ) · (k/ǫ)O(k
3
log(k/ǫ)/ log log(k/ǫ))k−1
· max |Sℓ | ,
ℓ∈[n]
and returns an ǫ-cover of S.
Observe that if we choose each Si to be a δ-cover for the set of all k-CRVs, with δ = ǫ/n, by
the subadditivity of the total variation distance for independent random variables, we obtain an
ǫ-cover for Mn,k , the set of all (n, k)-PMDs. It is easy to see that the set of k-CRVs has an explicit
δ-cover of size O(1/δ)k . This gives the following corollary:
Corollary 4.12. There exists an algorithm that, on input n, k ∈ Z+ , k > 2, and ǫ > 0, computes
3
3
k−1
.
a proper ǫ-cover for the set Mn,k and runs in time nO(k ) · (k/ǫ)O(k log(k/ǫ)/ log log(k/ǫ))
Proof of Theorem 4.11. The high-level idea is to split each such PMD into its i-maximal PMD
def
components and approximate each to total variation distance ǫ′ = ǫ/k. We do this by keeping
track of the appropriate data, along the lines of Proposition 4.9, and using dynamic programming.
For the sake of readability, we start by introducing the notation that is used throughout this
proof. We use X to denote a generic (n, k)-PMD, and X i , 1 ≤ i ≤ k, to denote its i-maximal
PMD components. For an (n, k)-PMD and a vector m = (m1 , . . . , mk ) ∈ Zk+ , we denote its mth
P
P
m
parameter moment by Mm (X) = nℓ=1 kj=1 pℓ,jj . Throughout this proof, we will only consider
parameter moments of i-maximal PMDs, in which case the vector m of interest will by construction
satisfy mi = 0, i.e., m = (m1 , . . . , mi−1 , 0, mi+1 , . . . , mk ).
In the first step of our algorithm, we guess approximations to the quantities 1 + sj (X i ) to
within a factor of 2, where X i is intended to be the i-maximal PMD component of our final PMD
X. We represent these guesses in the form of a matrix G = (Gi,j )1≤i6=j≤k . Specifically, we take
Gi,j = (2ai,j + 3)/4 for each integer ai,j ≥ 0, where each ai,j is bounded from above by O(log n). For
each fixed guess
P G, we proceed as follows: For h ∈ [n], we denote by Sh the set of all (h, k)-PMDs
of the form hℓ=1 Xℓ , where Xℓ ∈ Sℓ . For each h ∈ [n], we compute the set of all possible (distinct)
data DG (X), where X ∈ Sh . The data DG (X) consists of the following:
• The number of i-maximal k-CRVs of X, for each i, 1 ≤ i ≤ k.
• Letting X i denote the i-maximal PMD component of X, we partition the k-CRV components
of X i into three sets based on whether or not they are δ1′ -exceptional or δ2′ -exceptional with
respect to our guess matrix G for 1 + sj (X i ). Formally, we have the following definition:
35
P
Definition 4.13. Let X i be an i-maximal (hi , k)-PMD with X i = ℓ∈Ai Xℓ and 0 < δ ≤ 1.
We say that a particular component k-CRV Xℓ , ℓ ∈ Ai , is δ-exceptional with
p respect to
G = (Gi,j ), if there exists a coordinate j 6= i, 1 ≤ j ≤ k, such that pℓ,j ≥ δ · Gi,j . We will
denote by E(δ, G) ⊆ Ai the set of δ-exceptional coordinates of X i .
With this notation, we partition Ai into the following three sets: Ai1 = E(δ1′ , G), Ai2 =
E(δ2′ , G) \ Ai1 , and Ai3 = E(δ2′ , G). For each i, 1 ≤ i ≤ k, we store the following information:
– ni1 = |Ai1 |, ni2 = |Ai2 |, and ni3 = |Ai3 |.
– Approximations sej,i of the quantities sj (X i ), for each j 6= i, 1 ≤ j ≤ k to within an
additive error of (h/4n).
– Approximations of the parameter moments Mm (X i )(1) , for all m = (m1 , . . . , mk ) ∈ Zk+
with mi = 0 and |m|1 ≤ 2, to within an additive γ ′ /(2k)|m|1 · (ni1 /n).
– Approximations of the parameter moments Mm (X i )(2) , for all m = (m1 , . . . , mk ) ∈ Zk+
with mi = 0 and |m|1 ≤ K2′ to within an additive γ ′ /(2k)|m|1 · (ni2 · δ1′ 2 /2k).
– Rounding of each of the parameters pℓ,j , for each k-CRV Xℓ , ℓ ∈ Ai3 , to the nearest
integer multiple of ǫ′ δ2′ 2 /2k2 .
Note that DG (X) can be stored as a vector of counts and moments. In particular, for the data
associated with k-CRVs in Ai3 , 1 ≤ i ≤ k, we can store a vector of counts of the possible roundings
of the parameters using a sparse representation.
We emphasize that our aforementioned approximate description needs to satisfy the following
property: for independent PMDs X and Y , we have that DG (X + Y ) = DG (X) + DG (Y ). This
property is crucial, as it allows us to store only one PMD as a representative for each distinct data
vector. This follows from the fact that, if the property is satisfied, then DG (X + Y ) only depends
on the data associated with X and Y.
Pn
To ensure
this
property
is
satisfied,
for
a
PMD
X
=
ℓ=1 Xℓ , where Xℓ is a k-CRV, we define
Pn
DG (X) = ℓ=1 DG (Xℓ ). We now need to define DG (W ) for a k-CRV W . For DG (W ), we store
the following information:
• The value of i for which W is i-maximal.
• Whether or not W is δ1′ -exceptional and δ2′ -exceptional with respect to G.
• sj (W ) = Pr[W = j] rounded down to a multiple of 1/4n, for each j 6= i, 1 ≤ j ≤ k.
• If W is not δ1′ -exceptional with respect to G, the nearest integer multiple of γ ′ /(n(2k)|m|1 ) to
Mm (W ) for each m ∈ Zk+ , with mi = 0 and |m|1 ≤ 2.
but not δ2′ -exceptional with respect to G, the nearest integer multiple
• If W is δ1′ -exceptional
2
′
|m|
′
1
of γ /(2k)
· (δ1 /2k) to Mm (W ), for each m ∈ Zk+ with mi = 0 and |m|1 ≤ K2′ .
• If W is δ2′ -exceptional with respect to G, we store roundings of each of the probabilities
Pr[W = j] to the nearest integer multiple of ǫ′ δ2′ 2 /2k.
Given the above detailed description, we are ready to describe our dynamic programming based
algorithm. Recall that for each h, 1 ≤ h ≤ n, we compute sets of all possible (distinct) data DG (X),
where X ∈ Sh . We do the computation by a dynamic program that works as follows:
36
• At the beginning of step h, we have a set Dh−1 of all possibilities of DG (X) for PMDs of
P
D (X)
the form X = h−1
ej,iG
≤ 2Gi,j . Moreover, for each
ℓ=1 Xℓ , where Xℓ ∈ Sℓ , that have 1 + s
D ∈ Dh−1 , we have a representative PMD YD , given in terms of its k-CRVs, that satisfies
DG (YD ) = D.
• To compute Dh , we proceed as follows: We start by computing DG (Xh ), for each Xh ∈ Sh .
Then, we compute a list of possible data for Dh as follows: For each D ∈ Dh−1 and Xh ∈ Sh ,
we compute the data D + DG (Xh ), and the k-CRVs of the PMD YD + Xh that has this data,
since DG (YD + Xh ) = D + DG (Xh ). We then remove duplicate data from this list, arbitrarily
keeping one PMD that can produce the data. We then remove data D where 1 + seD
j,i ≥ 2Gi,j .
This gives our set of possible data Dh . Now, we note that Dh contains all possible data of
P
D (X)
PMDs of the form hℓ=1 Xℓ , where each Xℓ ∈ Sℓ , that have 1 + sej,iG
≤ 2Gi,j , and for each
distinct possibility, we have an explicit PMD that has this data.
• After step n, for each D ∈ Dn , we output the data and the associated explicit PMD, if the
following condition is satisfied:
Condition 4.14. For each i, j ∈ Z+ , with 1 ≤ i 6= j ≤ k, it holds (a) Gi,j ≤ 1 +
D (X)
D (X)
D (X)
max{0, sej,iG
− 1/4} and (b) 1 + sej,iG
≤ 2Gi,j , where sej,iG
is the approximation to
i
i
sj (X ) in DG (X), and Gi,j is the guess for 1 + sj (X ) in G.
We claim that the above computation, performed for all values of G, outputs an ǫ-cover of the set
S. This is formally established using the following claim:
Claim 4.15. (i) For any X, Y ∈ S, if DG (X) = DG (Y ) and DG (X) satisfies Condition 4.14,
then dTV (X, Y ) ≤ ǫ.
(ii) For any X ∈ S, there exists a G such that DG (X) satisfies for i, j ∈ Z+ , with 1 ≤ i 6= j ≤ k,
D (X)
D (X)
(a) Gi,j ≤ 1 + max{0, sej,iG
− 3/4} and (b) 1 + sej,iG
≤ 2Gi,j , hence also Condition 4.14.
Remark 4.16. Note that Condition (a) in statement (ii) of the claim above is slightly stronger
than that in Condition 4.14. This slightly stronger condition will be needed for the anonymous
games application in the following section.
Proof. To prove (i), we want to use Lemma 4.6 to show that that for all i ∈ [k], the i-maximal
components of X and Y are close, i.e., that dTV (X i , Y i ) ≤ ǫ/k. To do this, we proceed as follows:
D (X)
We first show that 12 ≤ (1 + sj (X i ))/(1 + sj (Y i )) ≤ 2. By the definition of sej,iG , for each
P
D (X)
1
i-maximal k-CRV Xℓ ∈ Ai , we have sj (Xℓ ) − 4n
≤ sej,iG
≤ sj (Xℓ ). Thus, for X i = ℓ∈Ai Xℓ , we
D (X)
have that sj (X i )−(1/4) ≤ sej,iG
D (X)
sej,iG
≤ sj (X i ). Since sj (X i ) ≥ 0, we have that max{0, sj (X i )−1/4} ≤
≤ sj (X i ). Combining this with Condition 4.14 yields that
Gi,j ≤ 1 + sj (X i ) ≤ 2Gi,j .
(16)
Since an identical inequality holds for Y, we have that 12 ≤ (1 + sj (X i ))/(1 + sj (Y i )) ≤ 2.
We next show that the set of coordinates Ai1 for X does not contain any δ1′ exceptional variables
i , since ℓ is not δ ′ -exceptional with respect to G, using (16), we have that
for X i . Forp
all ℓ ∈ Ap
1
1
pℓ,j ≤ δ1′ · Gi,j ≤ 1 + sj (X i ). Similarly, it follows that Ai2 for X does not contain any δ2′ exceptional variables. The same statements also directly follow for Y.
37
Now, we obtain bounds on the size of the Ait ’s, t = 1, 2, 3. We trivially have |Ai1 | ≤ n. From (16),
i are δ ′ -exceptional with respect to G . If we denote by E ⊆ Ai
we have that all variables in Ap
i,j
j
2
1
2
′
i
the set of ℓ ∈ A2 with pℓ,j ≥ δ1 Gi,j , then using (16), we have that
sj (X i ) =
X
ℓ∈Ai
pℓ,j ≥
X
ℓ∈Ai
p2ℓ,j ≥
X
ℓ∈Ej
p2ℓ,j ≥ δ1′2 |Ej |Gi,j ≥ δ1′2 |Ej |(1 + sj (X i ))/2 ≥ sj (X i ) · δ1′2 |Ej |/2.
S
Thus, |Ej | ≤ 2/δ1′2 . Since Ai2 = kj=1 Ej , we have |Ai2 | ≤ 2k/δ1′2 . Similarly, we have |Ai3 | ≤ 2k/δ2′2 .
For ℓ ∈ Ai1 , DG (Xℓ ) contains an approximation to Mm (Xℓ ) for each m ∈ Zk+ with mi = 0 and
|m|1 ≤ 2, to within accuracy γ′ /(2n(2k)|m|1 ). Since |Ai1 | ≤ n, we have that DG (X i ) contains an
approximation to Mm (X i )(1) to within γ ′ /(2(2k)|m|1 ). Since a similar bound holds for (Y i )(1) ,
and DG (Y i )(1) = DG (X i )(1) , we have that |Mm (X i )(1) − Mm (Y i )(1) | ≤ γ ′ /(2k)|m|1 .
Similarly, for ℓ ∈ Ai2 , DG (Xℓ ) contains an approximation to Mm (Xℓ ) for each m ∈ Zk+ with
mi = 0 and |m|1 ≤ K2′ to within accuracy (1/2) · γ ′ /(2k)|m|1 · (δ1′ 2 /2k). Since |Ai2 | ≤ 2k/δ2′2 ,
DG (X i ) contains an approximation to Mm (X i )(2) to within γ ′ /(2(2k)|m|1 ). Since a similar bound
holds for (Y i )(2) and DG (Y i )(2) = DG (X i )(2) , we have that |Mm (X i )(2) − Mm (Y i )(2) | ≤
γ ′ /(2k)|m|1 .
Finally, for ℓ ∈ Ai3 , DG (Xℓ ) contains an approximation to pℓ,j for all j 6= i to within ǫ′ δ2′ 2 /4k2 .
The counts of variables with these approximations are the same in (X i )(3) and (Y i )(3) . So, there
i (X) to Ai (Y ) such that an ℓ′ = f (ℓ) has D (X ) = D (Y ′ ). Then, we have
is bijection f from AP
G
G ℓ
ℓ
3
3
that dTV (Xℓ , Yℓ′ ) ≤ j6=i ǫ′ δ2′ 2 /2k2 ≤ ǫ′ δ2′ 2 /2k ≤ ǫ′ /|Ai3 |.
We now have all the necessary conditions to apply Lemma 4.6, yielding that dTV (X i , Y i ) ≤ ǫ/k.
By the sub-additivity of total variational distance, we have dTV (X, Y ) ≤ ǫ, proving statement (i)
of the claim.
To prove (ii), it suffices to show that for any i, j there is a Gi,j that satisfies the inequalities
claimed. Recall that Gi,j takes values of the form (2a + 3)/4 for an integer a ≥ 0. For a = 0,
D (X)
D (X)
− 3/4} is satisfied for any value of sej,iG .
Gi,j = 1 and the inequality Gi,j ≤ 1 + max{0, sej,iG
D (X)
When a ≥ 1, Gi,j > 1, so the inequality Gi,j ≤ 1 + max{0, sej,iG
Gi,j ≤ 1 +
D (X)
sej,iG
inequality in (i)
D (X)
− 3/4, i.e., when sej,iG
D (X)
is satisfied when sej,iG
≤
− 3/4} is only satisfied when
≥ Gi,j − 1/4 = (2a + 2)/4 = (2a−1 + 1)/2. The second
2Gi,j − 1 = 2 · (2a + 1)/4 = (2a + 1)/2.
D (X)
D (X)
Summarizing, for a = 0, we need that sej,iG
∈ [0, 1], and for a ≥ 1, we need that sej,iG
∈
[(2a−1 + 1)/2, (2a + 1)/2]. So, there is a Gi,j = (2ai,j + 3)/4 for which the required inequalities are
satisfied. Thus, there is a G for which we get the necessary inequalities for all 1 ≤ i, j ≤ k with
i 6= j. This completes the proof of (ii).
We now bound the running time:
Claim 4.17. For a generic (n, k)-PMD X, the number of possible values taken by DG (X) considered
3
3
k−1
is at most nO(k ) (k/ǫ)O(k log(1/ǫ)/ log log(1/ǫ)) .
Proof. For a fixed G, we consider the number of possibilities for DG (X i ) for each 1 ≤ i ≤ k.
For each j 6= i, we approximate sj (X i ) up to an additive 1/(4n). Since 0 ≤ sj (Xi ) ≤ n, there
are at most 4n2 possibilities. For all such j we have O(n2k ) possibilities.
We approximate the parameter moments of Mm (X i )(1) as an integer multiple of γ ′ /(n(2k)|m|1 )
for all m with m1 ≤ 2. For each such m, we have 0 ≤ Mm (X i )(1) ≤ n, so there are at
most n2 (2k)|m|1 /γ ′ = n2 (k log(1/ǫ))O(k) (1/ǫ) possibilities. There are O(k2 ) such m, so we have
2
3
2
nO(k ) · (k log(1/ǫ)O(k ) (1/ǫ)O(k ) possibilities.
38
We approximate the parameter moments of Mm (X i )(2) as a multiple of γ ′ /(2k)|m|1 ·(δ1′ 2 /2k)
for each m with |m|1 ≤ K2′ . The number of k-CRVs in (X i )(2)
is |Ai2 | ≤ 2k/δ1′2 from the proof
of Claim 4.15. So, for each m, we have 0 ≤ Mm (X i )(2) ≤ |Ai2 |, and there are at most
′
(2k)K2 +2 /(γ ′ δ1′2 δ2′2 ) = kO(k+ln(k/ǫ)/ ln ln(k/ǫ)) ln(1/ǫ)O(k) /ǫ = (k/ǫ)O(k) possibilities. Since there are
at most
K2′k−1 = O((ln(k/ǫ)/ ln ln(k/ǫ) + k)k−1
2 k−1
possibilities.
such moments, there are (k/ǫ)O(k ln(k/ǫ)/ ln ln(k/ǫ)+k )
i
We approximate each Xℓ for ℓ ∈ A3 as a k-CRV whose probabilities are multiples of ǫδ2′ 2 /2k2 .
So, there are (2k2 /(ǫδ2′ )k = (k/ǫ)O(k) possible k-CRVs. Since there may be |Ai3 | ≤ 2k/δ2′2 =
3/2
3
2k2 log3/2 (k/ǫ) such k-CRVs, there are (k/ǫ)O(k log (k/ǫ)) possibilities.
2
2 k−1
Multiplying these together, for every G, there are at most nO(k ) (k/ǫ)O(k ln(k/ǫ)/ ln ln(k/ǫ)+k )
3
3
k−1
possible
possible values of DG (X i ). Hence, there are at most nO(k ) (k/ǫ)O(k ln(k/ǫ)/ ln ln(k/ǫ))
2
k
values of DG (X) for a given G. Finally, there are O(log n) possible values of Gi,j , since Gi,j =
(2ai,j + 3)/4, for integers ai,j , and we do not need to consider Gi,j > n. Therefore, the number of
3
3
k−1
possible values of DG (X) is at most nO(k ) · (k/ǫ)O(k ln(k/ǫ)/ ln ln(k/ǫ)) .
The runtime of the algorithm is dominated by the runtime of the substep of each step h, where
we calculate D + DG (Xh ) for all D ∈ Dh−1 and Xh ∈ Sh . Note that D and DG (Xh ) are vectors
with O(K2′k ) = O(log(k/ǫ)/ log log(k/ǫ) + k)k non-zero coordinates. So, the runtime of step h is at
most
3
3
k−1
|Dh−1 | · |Sh | · O((K2′ )k ) = |Sh | · nO(k ) · (k/ǫ)O(k ln(k/ǫ)/ ln ln(k/ǫ))
,
by Claim 4.17. The overall runtime of the algorithm is thus nO(k
maxh |Sh |. This completes the proof of Theorem 4.11.
3)
· (k/ǫ)O(k
3
ln(k/ǫ)/ ln ln(k/ǫ))k−1
·
4.3 An EPTAS for Nash Equilibria in Anonymous Games In this subsection, we describe
our EPTAS for computing Nash equilibria in anonymous games:
3
3
k−1
Theorem 4.18. There exists an nO(k ) ·(k/ǫ)O(k log(k/ǫ)/ log log(k/ǫ)) -time algorithm for computing
a (well-supported) ǫ-Nash Equilibrium in an n-player, k-strategy anonymous game.
This subsection is devoted to the proof of Theorem 4.18.
We compute a well-supported ǫ-Nash equilibrium, using a procedure similar to [DP14]. We
start by using a dynamic program very similar to that of our Theorem 4.11 in order to construct
an ǫ/10-cover. We iterate over this ǫ/10-cover. For each element of the cover, we compute a set of
possible ǫ/5-best responses. Finally, we again use the dynamic program of Theorem 4.11 to check if
we can construct this element of the cover out of best responses. If we can, then we have found an
ǫ-Nash equilibrium. Since there exists an ǫ/5-Nash equilibrium in our cover, this procedure must
produce an output.
In more detail, to compute the aforementioned best responses, we use a modification of the
algorithm in Theorem 4.11, which produces output at the penultimate step. The reason for this
modification is the following: For the approximate Nash equilibrium computation, we need the
data produced by the dynamic program, not just the cover of PMDs. Using this data, we can
subtract the data corresponding to each candidate best response. This allows us to approximate
the distribution of the sum of the other players strategies, which we need in order to calculate the
players expected utilities.
Recall that a mixed strategy profile for a k-strategy anonymous game can be represented as
a set of k-CRVs, {Xi }i∈[k] , where the k-CRV Xi describes the mixed strategy for player i. Recall
39
that a mixed strategy profile is an ǫ-approximate NashPequilibrium, if for each player i we have
E[uiXi (X−i )] ≥ E[uiℓ (X−i )] − ǫ, for ℓ ∈ [k], where X−i = j∈[n]\{i} Xj is the distribution of the sum
of other players strategies. A strategy profile is an ǫ-well-supported Nash equilibrium if for each
player i, E[uiℓ′ (X−i )] ≥ E[uiℓ (X−i )] − ǫ for each ℓ ∈ [k] and eℓ′ in the support of Xi . If this holds for
one player i, then we call Xi an ǫ(-well-supported) best response to X−i .
Lemma 4.19. Suppose that Xi is a δ-best response to X−i for player i. Then, if an n − 1 PMD
Y−i has dTV (X−i , Y−i ) ≤ ǫ, Xi is a (δ + 2ǫ)-best response to Y−i . If, additionally, a k-CRV Yi has
Pr[Yi = ej ] = 0 for all j with Pr[Xi = ej ] = 0, then Yi is a (δ + 2ǫ)-best response to Y−i .
Proof. Since uiℓ (x) ∈ [0, 1] for ℓ ∈ [k] and any x in the support of X−i , we have that E[uiℓ (X−i )] −
E[uiℓ (Y−i )] ≤ dTV (X−i , Y−i )). Similarly, we have E[uiℓ (X−i )] − E[uiℓ (Y−i )] ≤ dTV (X−i , Y−i ). Thus,
for all eℓ′ in the support of Xi and all ℓ ∈ [k], we have
E[uiℓ′ (Y−i )] ≥ E[uiℓ′ (X−i )] − ǫ ≥ E[uiℓ (X−i )] − ǫ − δ ≥ E[uiℓ (Y−i )] − 2ǫ − δ.
That is, Xi is a (δ + 2ǫ)-best response to Y−i . Since the support of Yi is a subset if the support of
Xi , Yi is also a (δ + 2ǫ)-best response to Y−i .
We note that by rounding the entries of an actual Nash Equilibrium, there exists an ǫ/5-Nash
equilibrium where all the probabilities of all the strategies are integer multiples of ǫ/(10kn):
Claim 4.20. There is an ǫ/5-well-supported Nash equilibrium {Xi }, where the probabilities Pr[Xi =
ej ] are multiples of ǫ/(10kn), for all 1 ≤ i ≤ n and 1 ≤ j ≤ k.
Proof. By Nash’s Theorem, there is a Nash equilibrium {Yi }. We construct {Xi } from {Yi } as
follows: If Yi is ℓ-maximal, then for every j 6=P
ℓ, we set Pr[Xi = ej ] to be Pr[Yi = ej ] rounded down to
a multiple of ǫ/(10kn) and Pr[Xi = eℓ ] = 1− j6=ℓ Pr[Xj = ej ]. Now, we have dTV (Xi , Yi ) ≤ ǫ/(10n)
and the support of Xi is a subset of the support of Yj . By the sub-additivity of total variational
distance, for every i we have dTV (X−i , Y−i ) ≤ ǫ/10. Since {Yi } is a Nash equilibrium, for all players
i, Yi is a 0-best response to Y−i . By Lemma 4.19, we have that Xi is a 2ǫ/10-best response to X−i
for all players i. Hence, {Xi } is an ǫ/5-well supported Nash equilibrium.
Let S be the set of all k-CRVs whose probabilities are multiples of ǫ/(10kn). We will require
def
def
a modification of the algorithm from Theorem 4.11 (applied with Si = S for all i, and ǫ = ǫ/5),
which produces output at both step n and step n−1. Specifically, in addition to outputting a subset
VG,n ⊆ DG,n of the data of possible (n, k)-PMDs that satisfy conditions (a) and (b) of Claim‘4.15
(ii), we output the subset VG,n−1 ⊆ DG,n−1 of the data of possible (n − 1, k)-PMDs that satisfy the
slightly weaker conditions (a) and (b) of Condition 4.14 .
In more detail, we need the following guarantees about the output of our modified algorithm:
P
P
Claim 4.21. For every PMD X = ni=1 Xi and X−j = i∈[n]\j Xi , for some 1 ≤ j ≤ n, and any
Xi ∈ S, for 1 ≤ i ≤ n, we have:
• There is a guess G, such that DG (X) ∈ VG,n .
• For any G such that DG (X) ∈ VG,n , we also have DG (X) − DG (Xj ) = DG (X−j ) ∈ VG,n−1 .
• If DG (X) ∈ VG,n , for any PMD Y with DG (Y ) = DG (X) or DG (Y ) = DG (X−j ), we have
dTV (X, Y ) ≤ ǫ/5 or dTV (X−j , Y ) ≤ ǫ/5 respectively.
40
Proof. By Claim 4.15 (ii), there is a G such that DG (X) satisfies conditions (a) and (b) and so
DG (X) ∈ VG,n .
We note that by the correctness of the dynamic program, since X−j is a sum of n − 1 many
k-CRVs in S, we have DG (X−j ) ∈ DG,n−1 . To show that it is in VG,n−1 , we need to show that all
D (X )
D (X)
seh,iG −j satisfy Condition 4.14, for all 1 ≤ i, h ≤ k and h 6= i. We know that seh,iG
satisfies the
stronger conditions (a) and (b) of Claim 4.15 (ii). All we need to show is that
D (X)
seh,iG
D (X−j )
− 1/2 ≤ seh,iG
D (X)
≤ seh,iG
.
This condition is trivial unless Xj is i-maximal. If it is, we note that Pr[Xj = eh ] ≤ Pr[Xj = ei ],
D (X )
and so sej,iG j ≤ Pr[Xj = eh ] ≤ 1/2. Thus, DG (X−j ) = DG (X) − DG (Xj ) ∈ VG,n−1 .
We now have that both DG (X) and DG (X−j ) satisfy Condition 4.14. Therefore, Claim 4.15 (i)
yields the third claim.
We note that we can calculate the expected utilities efficiently to sufficient precision:
Claim 4.22. Given an anonymous game (n, k, {uiℓ }i∈[n],ℓ∈[k]) with each utility given to within an
additive ǫ/2 using O(log(1/ǫ)) bits, and given
P a PMD X in terms of its constituent k-CRVs Xi , we
i
can approximate the expected utility E[uℓ ( j6=i Xi )] for any player i and pure strategy ℓ to within
ǫ in time O(nk+1 · k log(n) · polylog(1/ǫ)).
P
Proof. We can compute the probability mass function of X−i = j6=i Xj by using the FFT on
Q
bi , calculate the DFT of X−i , X
d
bj , and
[n]k . We calculate the DFT of each Xi , X
X
−i (ξ) =
j6=i
finally compute the inverse DFT. To do this within ǫ/2 total variational error needs time O(nk+1 ·
k log(n)polylog(1/ǫ)), since we need to use the FFT algorithm
n + 1 times. We then use this
P
approximate pmf to compute the expectation E[uiℓ (X−i )] = x uiℓ (x)Xi (x). This takes time O(nk )
and gives error ǫ.
Henceforth, we will assume that we can compute these expectations exactly, but it should be
clear that computing them to within a suitably small O(ǫ) error suffices.
Proof of Theorem 4.18. We use the modified dynamic programming algorithm given above to produce an ǫ/5-cover with explicit sets VG,n , VG,n−1 of data and PMDs which produce each output
data.
Then, for each G and for each D ∈ VG,n , we try to construct an ǫ-Nash equilibrium whose
associated PMD X has DG (X) = D. Firstly, for each player i we compute a set Si ⊆ S of best
responses to X. To do this, we check each Xi ∈ S individually. We first check if DG (X) − DG (Xi ) ∈
VG,n−1 . If it is not, then Claim 4.21
Pnimplies that there is no set of strategies for the other players
Xj ∈ S, for j 6= i, such that DG ( i=1 Xi ) = D. In this case, we do not put this Xi ∈ Si . If we
do have D−i := D − DG (Xi ) ∈ VG,n−1 , then we recall that the algorithm gives us an explicit YD−i
such that D(YD−i ) = D−i . Now, we calculate the expected utilities E[uiℓ (YD−i )] for each 1 ≤ ℓ ≤ k.
If Xi is a 3ǫ/5-best response to YD−i , then we add it to Si .
When we have calculated the set of best responses Si for each player, we use the algorithm
from Theorem 4.11 with these Si ’s and this guess G. If the set of data it outputs contains P
D, then
we output the explicit PMD X := YD that does so in terms of its constituent CRVs X = ni=1 Xi
and terminate.
To prove correctness, we first show that {Xi } is an ǫ-Nash equilibrium, and second that that
the algorithm always produces an output. We need to show that Xi is an ǫ-best response to
41
P
X−i = j∈[n]\i Xj . When we put Xi in Si , we checked that Xi was a 3ǫ/5-best response to YD−i ,
where D−i = D − DG (Xi ). But note that
DG (YD−i ) = D − DG (Xi ) = DG (X) − DG (Xi ) = DG (X−i ).
Since D ∈ VG,n and DG (YD−i ) = DG (X−i ), Claim 4.21 yields that dTV (X−i , YD−i ) ≤ ǫ/5. So, by
Lemma 4.19, we indeed have that Xi is an ǫ-best response to X−i . Since this holds for all Xi , X
is an ǫ-Nash equilibrium.
′
′
By Claim 4.20, there
Pnexists′ an ǫ/5-Nash equilibrium {Xi }, ′ with each Xi ∈ S. By Claim 4.21,
′
we have that for X = i=1 Xi , there is a guess G with DG (X ) ∈ VG,n . So, if the algorithm does
′
not terminate successfully first, it eventually considers G and
P D := D′G (X ). We ′next show that
′
′
that the algorithm puts Xi in Si . For each 1 ≤ i ≤ n, X−i = j∈[n]\i Xj has DG (X−j ) ∈ VG,n−1 by
′ ), and we
Claim 4.21, since DG (X ′ ) ∈ VG,n . So, D−i = D − DG (Xi′ ) = DG (X ′ ) − DG (Xi′ ) = DG (X−i
′
′
have D−i ∈ VG,n−1 . Hence, the algorithm will put Xi in Si if Xi is an 4ǫ/5-best response to YD−i . By
Claim 4.21, since DG (X) ∈ VG,n and DG (X−i ) = DG (YD−i ), we have dTV (YD−i , X−i ) ≤ ǫ/5. Since
′ is an ǫ/5-best response to X ′ . Since d
{Xi′ } is an ǫ/5-Nash equilibrium, X−i
TV (YD−i , X−i ) ≤ ǫ/5,
−i
by Lemma 4.19, this implies that Xi′ is a 3ǫ/5-best response to YD−i . Thus, the algorithm puts
Xi′ in Si . Since each Xi′ satisfies Xi′ ∈ Si , by Theorem 4.11, the algorithm from that theorem
outputs a set of data that includes DG (X ′ ) = D. Therefore, if the algorithm does not terminate
successfully first, when it considers G and D, it will produce an output. This completes the proof
of Theorem 4.18.
Threat points in anonymous games. As an additional application of our proper cover construction, we give an EPTAS for computing threat points in anonymous games [BCI+ 08].
Definition 4.23. The threat point of an anonymous game (n, k, {uiℓ }i∈[n],ℓ∈[k]) is the vector θ with
θi =
min
max E[uij (X−i )].
X−i ∈Mn−1,k 1≤j≤k
Intuitively, If all other players cooperate to try and punish player i, then they can force her
expected utility to be θi but no lower, so long as player i is trying to maximize it. This notion has
applications in finding Nash equilibria in repeated anonymous games.
Corollary 4.24. Given an anonymous game (n, k, {uiℓ }i∈[n],ℓ∈[k]) with k > 2, we can compute a θ̃
3
3
k−1
with kθ − θ̃k∞ ≤ ǫ in time nO(k ) · (k/ǫ)O(k log(k/ǫ)/ log log(k/ǫ)) . Additionally,
P for each player i,
we obtain strategies Xi,j for all other players j 6= i such that max1≤ℓ≤k E[uiℓ ( j6=i Xi,j )] ≤ θi + ǫ.
Proof. Using the dynamic programming algorithm of Theorem 4.11, we can construct an ǫ-cover C
of Mn−1,k . For each player i, we then compute θ̃i = minX−i ∈C max1≤j≤k E[uij (X−i )] by brute force.
Additionally, we return the k-CRVs Xi,j that the algorithm gives us as the explicit parameters
of the PMD X−i which achieves this minimum, i.e., with θ̃i = max1≤j≤k E[uij (X−i )]. The running
time of this algorithm is dominated by the dynamic programming step.
We now show correctness. Let Y−i ∈ Mn−1,k be such that θi = max1≤j≤k E[uij (X−i )]. Then,
′
i
′ ∈ C with d
′
i
there exists a Y−i
TV (Y−i , Y−i ) ≤ ǫ, and so we have |E[uj (Y−i )] − E[uj (Y−i )]| ≤ ǫ.
Therefore,
′
θi = max E[uij (Y−i )] ≥ max E[uij (Y−i
)] − ǫ ≥ θ̃i − ǫ .
1≤j≤k
1≤j≤k
Similarly, there is an X−i ∈ C with θ̃i = max1≤j≤k E[uij (X−i )] ≤ θi . And so we have |θ̃i − θi | ≤ ǫ,
P
P
as required. Additionally, for the j6=i Xi,j = X−i we have max1≤ℓ≤k E[uiℓ ( j6=i Xi,j )] = θ˜i ≤
θi + ǫ.
42
4.4 Every PMD is close to a PMD with few distinct parameters In this section, we
prove our structural result that states that any PMD is close to another PMD which is the sum of
k-CRVs with a small number of distinct parameters.
Theorem 4.25. Let n, k ∈ Z+ , k > 2, and ǫ > 0. For any (n, k)-PMD X, there is an
P (n, k)-PMD
Y such that dTV (X, Y ) ≤ ǫ satisfying the following property: We can write Y = ni=1 Yi where
each k-CRV Yi is distributed as one of
O ((log(k/ǫ)/(log log(k/ǫ)) + k))k
distinct k-CRVs.
The main geometric tool used to prove this is the following result from [GRW15]:
Lemma 4.26 (Theorem 14 from [GRW15]). Let f (x) be a multivariate polynomial with variables
xi,j , for 1 ≤ i ≤ n and 1 ≤ j ≤ k, which is symmetric up to permutations of the i’s, i.e., such that
for any permutation σ ∈ Sn , we have that, for (xσ )i,j := xσ(i),j , for all 1 ≤ i ≤ n and 1 ≤ j ≤ k,
f (xσ ) = f (x).PLet w ∈ Zk>0 . Suppose that f has weighted w degree at most d, i.e., each monomial
Q
ai,j
′
n
i,j xi,j has
i,j wj ai,j ≤ d. Suppose that the minimum of f (x) is attained by some x ∈ R , i.e.,
∗
∗
′
that f (x ) = minx∈Rn×k f (x). Then, there is a point x with f (x ) = minx∈Rn×k
j fk(x), such that the
Q
k
number of distinct y ∈ Rk of the form yj = x∗i,j , for some i, is at most j=1 wdj .
Proof of Theorem 4.25. Firstly we’re going to divide our PMD into i-maximal PMDs. We assume
wlog that X is k-maximal below.
We divide this PMD X into component PMDs X (1) , X (2) , X (3) according to whether these are
δ1 and δ2 , as in the proof of Proposition 4.9. We want to show that there exists a Y (1) ,Y (2) such
that X (1) and Y (1) agree on the first 2 moments, X (2) and Y (2) agree on the first K2 moments, but
each has few distinct CRVs. Then Y = Y (1) + Y (2) + X (3) is close to X by Lemma 4.6 (because
the first moments agree, i.e., we have sj (X) = sj (Y )).
We are going to use Lemma 4.26 to show that P
we can satisfy some polynomial equations
pl (x) = 0 by setting f to be a sum of squares f (x) = l pl (x)2 . Then if the polynomial equations
have a simultaneous solution at x, f attains its minimum of 0 at x. Some of these pl ’s are going
to be symmetric in terms of i. For the rest, we are going to have identical equations that hold for
each individual i, so f overall will be symmetric.
We have X (t) for t = 1, 2, and we want to construct a Y (t) with few distinct k-CRVs. That is, we
want to find pi,j , the probability that Yi = pi,j , for 1 ≤ i ≤ n, 1 ≤ j ≤ k. These pi,j ’s have to satisfy
(t)
certain inequalities to ensure each Yi is a non-δt exceptional k-CRV and pi,1 ≤ pi,2 ≤ . . . ≤ pi,k .
To do this, we will need to introduce variables whose square is the slack in each of these inequalities.
The free variables of these equations will be pi,1 , . . . , pi,k , xi,1 , . . . , xi,3k . The equations we consider are as follows:
(t)
The following two equations mean that Yi is a k-CRV with the necessary properties: For each
1 ≤ i ≤ n and 1 ≤ j ≤ k − 1,
pi,j = x2i,j
(17)
pi,j + x2i,j+k = pi,k
and
pi,j + x2i,j+2k = δt
43
q
1 + sj (X) .
(18)
(19)
For each 1 ≤ i ≤ n,
k
X
pi,j = 1 .
(20)
j=1
We need an equation that the
mth
moment of Y (t) is identical to the mth moment of X (t) , i.e.,
X Y mj
pi,j − Mm (X (t) ) = 0 ,
(21)
i
j
for each moment m with |m|1 ≤ Kt .
If these equations have a solution for real pi,j ’s and xi,j ’s, then the pi,j ’s satisfy all the inequalities
we need. We square all these expressions and sum them giving f. Note that the slack variables
xi,j only appear in monomials of degree 4 in f . We set the weights wj of the pi,j to be 1 and
the weights of the xi,j to be Kt /2. Then, f has w degree 2Kt : (21) has degree Kt in terms of
pi,j , so when we square it to put it in f , it has degree 2Kt . So we have that, for d = 2Kt ,
Qk j d k
k 3k = O(K )k Now f is symmetric in terms of the n different values of i, so we
t
j=1 wj = (2Kt ) 4
can apply Lemma 4.26, which yields that there is a minimum with O(Kt )k distinct (k + 1)-vectors
provided that there is any minimum.
However, note that if we set p′i,j = Pr[Xi = ej ] and define the x′i,j appropriately, we obtain an
x′ such that f (x′ ) = 0. Since f is a sum of squares f (x) ≥ 0. So, there is an x∗ with f (x∗ ) = 0, but
such that x∗ has O(Kt )k distinct 4k-vectors (p∗i,1 , . . . , p∗i,k , x∗i,1 , . . . , x∗i,k ).
Using the p∗i,j ’s in this solution, we have a Y (t) with O(Kt )k distinct CRVs. So, the Y which is
O(ǫ) close to X has (O(K1 )k + O(K2 )k + k(log 1/ǫ)2 ) distinct k-CRVs. Overall, we have that any
PMD is O(kǫ)-close to one with
k · O(K2 )k = O ((log(1/ǫ)/(log log(1/ǫ)) + k))k
distinct constituent k-CRVs. Thus, every PMD is ǫ-close to one with k·O ((log(k/ǫ)/(log log(k/ǫ)) + k))k
distinct constituent k-CRVs. This completes the proof.
4.5 Cover Size Lower Bound for PMDs In this subsection, we prove our lower bound on
the cover size of PMDs, which is restated below:
Theorem 4.27. (Cover Size Lower Bound for (n, k)-PMDs) Let k > 2, k ∈ Z+ , and ǫ be sufficiently
small as a function of k. For n = Ω((1/k) · log(1/ǫ)/ log log(1/ǫ))k−1 any ǫ-cover of Mn,k under
k−1
the total variation distance must be of size nΩ(k) · (1/ǫ)Ω((1/k)·log(1/ǫ)/ log log(1/ǫ)) .
Theorem 4.27 will follow from the following theorem:
Theorem 4.28. Let k > 2, k ∈ Z+ , and ǫ be sufficiently small as a function of k. Let n =
Ω((1/k) · log(1/ǫ)/ log log(1/ǫ))k−1 . There exists a set S of (n, k)-PMDs so that for x, y ∈ S, x 6= y
k−1
implies that dTV (x, y) ≥ ǫ, and |S| ≥ (1/ǫ)Ω((1/k)·log(1/ǫ)/ log log(1/ǫ)) .
The proof of Theorem 4.28 is quite elaborate and is postponed to the following subsection.
def
We now show how Theorem 1.5 follows from it. Let n0 = Θ((1/k) · log(1/ǫ)/ log log(1/ǫ))k−1 . By
Theorem 4.28, there exists a set Sn0 of size (1/ǫ)Ω(n0 ) consisting of (n0 , k)-PMDs that are ǫ-far
from each other.
We construct (n/n0 )Ω(k) appropriate “shifts” of the set Sn0 , by selecting appropriate sets of
n − n0 deterministic component k-CRVs. These sets shift the mean vector of the corresponding
44
PMD, while the remaining n0 components form an embedding of the set Sn0 . We remark that the
PMDs corresponding to different shifts have disjoint supports. Therefore, any ǫ-cover must contain
disjoint ǫ-covers for each shift, which is isomorphic to Sn0 . Therefore, any ǫ-cover must be of size
k−1
(n/n0 )Ω(k) · (1/ǫ)Ω((1/k)·log(1/ǫ)/ log log(1/ǫ))
k−1
= nΩ(k) · (1/ǫ)Ω((1/k)·log(1/ǫ)/ log log(1/ǫ))
,
where the last inequality used the fact that nk0 = o((1/ǫ)n0 ), if the parameter ǫ is sufficiently small
as a function of k. This completes the proof. The following subsection is devoted to the proof of
Theorem 4.28.
4.5.1 Proof of Theorem 4.28. Let k > 2, k ∈ Z+ , and ǫ be sufficiently small as a function of
k. Let n = Θ((1/k) · log(1/ǫ)/ log log(1/ǫ))k−1 .
We express an (n, k)-PMD X as a sum of independent k-CRVs Xs , where s ranges over some
index set. For 1 ≤ j ≤ k − 1, we will denote ps,j = Pr[Xs = ej ]. Note that Pr[Xs = ek ] =
Pk−1
1 − j=1
ps,j .
We construct our lower bound set S explicitly as follows. Let 0 < c < 1 be an appropriately
def
small universal constant. We define the integer parameters a = ⌊c ln(1/ǫ)/2k ln ln(1/ǫ))⌋ and
def
t = ⌊ǫ−c ⌋. We define the set S to have elements indexed by a function f : [a]k−1 → [t], where the
function f corresponds to the PMD
X
def
Xf =
Xsf ,
s∈[a]k−1
and the k-CRV Xsf , s = (s1 , . . . , sk−1 ) ∈ [a]k−1 , has the following parameters:
pfs,j =
sj + δj,1 ǫ3c f (s)
,
lnk (1/ǫ)
(22)
for 1 ≤ j ≤ k − 1. (Note that we use δi,j to denote the standard Kronecker delta function, i.e.,
δi,j = 1 if and only if i = j).
Let F = {f | f : [a]k−1 → [t]} be the set of all functions from [a]k−1 to [t]. Then, we have that
def
S = {X f : f ∈ F}.
That is, each PMD in S is the sum of ak−1 many k-CRVs, and there are t possibilities for each
k-CRV. Therefore,
k−1
k−1
= (1/ǫ)Ω((1/k)·log(1/ǫ)/ log log(1/ǫ)) .
|S| = ta
Observe that all PMDs in S are k-maximal. In particular, for any f ∈ F, s ∈ [a]k−1 , and 1 ≤ j ≤
k − 1, the above definition implies that
pfs,j ≤
1
1
·
.
k lnk (1/ǫ)
(23)
An important observation, that will be used throughout our proof, is that for each k-CRV Xsf , only
the first out of the k −1 parameters pfs,j , 1 ≤ j ≤ k −1, depends on the function f . More specifically,
the effect of the function f on pfs,1 is a very small perturbation of the numerator. Note that the first
summand in the numerator of (22) is a positive integer, while the summand corresponding to f is
at most ǫ2c = o(1). We emphasize that this perturbation term is an absolutely crucial ingredient
of our construction. As will become clear from the proof below, this term allows us to show that
distinct PMDs in S have a parameter moment that is substantially different.
The proof proceeds in two main conceptual steps that we explain in detail below.
45
First Step. In the first step, we show that for any two distinct PMDs in S, there exists a
k−1
, we recall that
parameter moment in which they differ by a non-trivial amount. For m ∈ Z+
P
def
th
the m parameter moment of a k-maximal PMD X =
s∈S Xs is defined to be Mm (X) =
P
Qk−1 mj
f
g
s∈S
j=1 ps,j . In Lemma 4.29 below, we show that for any distinct PMDs X , X ∈ S, there
exists m ∈ [a]k−1 such that their mth parameter moments differ by at least poly(ǫ).
Lemma 4.29. If f, g : [a]k−1 → [t], with f 6= g, then there exists m ∈ [a]k−1 so that
|Mm (X f ) − Mm (X g )| ≥ ǫ4c .
We now give a brief intuitive overview of the proof. It is clear that, for f 6= g, the PMDs
X f and X g have distinct parameters. Indeed, since f 6= g, there exists an s ∈ [a]k−1 such that
f (s) 6= g(s), which implies that the k-CRVs Xsf and Xsg have pfs,1 6= pgs,1 .
We start by pointing out that if two arbitrary PMDs have distinct parameters, there exists a
parameter moment where they differ. This implication uses the fact that PMDs are determined
by their moments, which can be established by showing that the Jacobian matrix of the moment
function is non-singular. Lemma 4.29 is a a robust version of this fact, that applies to PMDs in S,
and is proved by crucially exploiting the structure of the set S.
Our proof of Lemma 4.29 proceeds as follows: We start by approximating the parameter moments Mm (X f ), X f ∈ S, from above and below, using the definition of the parameters of X f . This
approximation step allows us to express the desired difference Mm (X f ) − Mm (X g ) (roughly) as
the product of two terms: the first term is always positive and has magnitude poly(ǫ), while the
second term is L · (f − g), for a certain linear transformation (matrix) L. We show that L is the
tensor product of matrices Li , where each Li is a Vandermonde matrix on distinct integers. Hence,
each Li is invertible, which in turn implies that L is invertible. Therefore, since f 6= g, we deduce
that L · (f − g) 6= 0. Noting that the elements of this vector are integers, yields the desired lower
bound.
Proof of Lemma 4.29. We begin by approximating the mth parameter moment of X f . We have
that
X
f
Mm (X ) =
k−1
Y
s∈[a]k−1 j=1
= ln
−kkmk1
sj + δj,1 ǫ3c f (s)
lnk (1/ǫ)
(1/ǫ)
X
mj
3c
m1
(s1 + ǫ f (s))
k−1
Y
mj
sj
.
j=2
s∈[a]k−1
P 1 m1 m1 −i 3c
Note that in the expression (s1 +ǫ3c f (s))m1 = m
(ǫ f (s))i , the ratio of the (ǫ3c f (s))i+1
i=0 i s1
3c
i
3c
2c
term to the (ǫ f (s)) term is (m1 − i)ǫ f (s)/s1 i ≤ aǫ ≤ 1/2. So, we have
3c
m1
(s1 + ǫ f (s))
=
m1
X
m1
i
i=0
≤
1
sm
1
≤
1
sm
1
1 −i
sm
(ǫ3c f (s))i
1
+
1 −1 3c
m1 s m
ǫ f (s) +
1
+
1 −1 3c
m1 s m
ǫ f (s) +
1
46
(m1 (m1 −
a 4c
a ǫ
.
1)/2)s1m1 −2 (ǫ3c f (s))2
m
1 −2
X
i=0
2−i
We can therefore write
f
Mm (X ) = ln
−kkmk1
(1/ǫ)
X
Y mj
m1 k−1
sj
s1 + ǫ f (s)
3c
s∈[a]k−1
j=2
k−1
X
Y mj
m1 −1 3c
1
≤ ln−kkmk1 (1/ǫ)
sm
ǫ f (s) + aa ǫ4c
sj
1 + m1 s 1
s∈[a]k−1
≤ ln−kkmk1 (1/ǫ)
X
k−1
Y
s∈[a]k−1 j=1
j=2
m
sj j + ǫ3c
X
1 −1
m1 f (s)sm
1
k−1
Y
j=2
s∈[a]k−1
m
sj j + aka ǫ4c .
Note that aka = exp (ak ln a) ≤ exp (ak ln ln 1/ǫ) ≤ exp (c ln ǫ/2) = (1/ǫ)c/2 , and so finally we have
k−1
Y mj
X
Y mj
X k−1
1 −1
sj + ǫ7c/2 ,
sj + ǫ3c
m1 f (s)sm
Mm (X f ) ≤ ln−kkmk1 (1/ǫ)
1
s∈[a]k−1 j=1
and that
Mm (X f ) ≥ ln−kkmk1 (1/ǫ)
X
j=2
s∈[a]k−1
k−1
Y
s∈[a]k−1 j=1
m
sj j + ǫ3c
X
1 −1
m1 f (s)sm
1
s∈[a]k−1
k−1
Y
j=2
m
sj j .
An analogous formula holds for the parameter moments of X g and therefore
k−1
Y mj
X
1 −1
sj (f (s) − g(s)) + O(ǫ7c/2 ) .
m1 s m
Mm (X f ) − Mm (X g ) = ln−kkmk1 (1/ǫ) ǫ3c
1
j=2
s∈[a]k−1
k−1
, the integer
We need to show that, for at least one value of m ∈ Z+
X
k−1
Y
mj −1
sj
s∈[a]k−1 j=1
is non-zero, since log−kkmk1 (1/ǫ) · ǫ3c m1
Qk−1
j=1
(f (s) − g(s))
sj > 0 for all s and m.
We observe that these integers are the coordinates of L(f − g), where L : R[a]
the linear transformation with
def
L(h)m =
X
s∈[a]k−1
k−1
Y
mj −1
sj
k−1
→ R[a]
k−1
is
(h(s)) ,
j=1
for m ∈ [a]k−1 . It should be noted that L is the tensor product of the linear transformations
Li : Rm → Rm , with
X m −1
Li (h)mi =
si i h(s) ,
si ∈[a]
for 1 ≤ i ≤ k −1. Moreover, each Li (h) is given by the Vandermonde matrix on the distinct integers
1, 2, . . . , a, which is non-singular. Since each Li is invertible, the tensor product L is also invertible.
Therefore, L(f − g) is non-zero. That is, there exists an m ∈ [a]k−1 with (L(f − g))m 6= 0, and so
Mm (X f ) − Mm (X g ) = log−kkmk1 (1/ǫ) · ǫ3c · m1
47
k−1
Y
j=1
sj (L(f − g))m 6= 0 .
Since m1
Qk−1
j=1
sj (L(f − g))m is an integer, |m1
Qk−1
j=1
sj (L(f − g))m | ≥ 1. So, we get
|Mm (X f ) − Mm (X g )| ≥ ln−kkmk1 (1/ǫ)ǫ3c .
Finally, we note that ln−kkmk1 (1/ǫ) = exp(−kkmk1 ln ln 1/ǫ) ≥ exp(−ka ln ln 1/ǫ) ≥ ǫc/2 . We
therefore conclude that |Mm (X f ) − Mm (X g )| ≥ ǫ4c , as required.
Second Step. In the second step of the proof, we show that two PMDs in S that have a parameter
moment that differs by a non-trivial amount, must differ significantly in total variation distance.
In particular, we prove:
Lemma 4.30. Let f, g : [a]k−1 → [t], with f 6= g. If |Mm (X f ) − Mm (X g )| ≥ ǫ4c for some
m ∈ [a]k−1 , then dTV (X f , X g ) ≥ ǫ.
We establish this lemma in two sub-steps: We first show that if the mth parameter moments
of two PMDs in S differ by a non-trivial amount, then the corresponding probability generating
functions (PGF) must differ by a non-trivial amount at a point. An intriguing property of our
proof of this claim is that it is non-constructive: we prove that there exists a point where the
PGF’s differ, but we do not explicitly find such a point. Our non-constructive argument makes
essential use of Cauchy’s integral formula. We are then able to directly translate a distance lower
bound between the PGFs to a lower bound in total variation distance.
For a random variable W = (W1 , . . . , Wk ) taking values in Zk and z = (z1 ,h. . . , zk−1 ) i∈ Ck−1 ,
Qk−1 Wi
def
. For a
we recall the definition of the probability generating function: P (W, z) = E
i=1 zi
PMD X f , we have that
P (X f , z) = E
Y
k−1
Y
s∈[a]k−1 i=1
f
Xs,i
zi
=
Y
s∈[a]k−1
1+
k−1
X
i=1
pfs,i (zi
!
− 1)
.
We start by establishing the following crucial claim:
Claim 4.31. Let f, g : [a]k−1 → [t], with f 6= g. If |Mm (X f )−Mm (X g )| ≥ ǫ4c for some m ∈ [a]k−1 ,
then there exists z ∗ ∈ Ck−1 with kz ∗ k∞ ≤ 2 such that |P (X f , z ∗ ) − P (X g , z ∗ )| ≥ ǫ5c .
Before we proceed with the formal proof, we provide an intuitive explanation of
the argument.
The proof of Claim 4.31 proceeds as follows: We start by expressing ln P (X f , z) , the logarithm
of the PGF of a PMD X f ∈ S, as a Taylor series whose coefficients depend on its parameter
moments Mm (X f ). We remark that appropriate bounds on the parameter moments of X f ∈ S
imply that this series is in fact absolutely convergent in an appropriate region R. Note that, using
k−1
,
the aforementioned Taylor expansion,we can express each parameter moment Mm (X f ), m ∈ Z+
as a partial derivative of ln P (X f , z) . Hence, if X f and X g are distinct PMDs in S, the difference
|Mm (X f ) − Mm (X g )| can also be expressed
value of the partial derivative of the
as the absolute
f
g
difference between the PGFs ln P (X , z) − ln (P (X , z)) . We then use Cauchy’s integral formula
to express this partial derivative as anintegral, which we can further be absolutely bounded from
above by the difference ln
P (X f , z ∗ ) − ln (P (X g , z ∗ )) , for some point z ∗ ∈ R. Finally, we use
the fact that ln P (X f , z) is absolutely bounded for all z ∈ R to complete the proof of the claim.
48
Proof of Claim 4.31. For z ∈ Ck−1 , let w ∈ Ck−1 be defined by wi = zi − 1, 1 ≤ i ≤ k − 1. When
|wi | ≤ 5, for all 1 ≤ i ≤ k − 1, we take logarithms and obtain:
!
k−1
X
X
f
f
ps,i wi
ln(P (X , z)) =
ln 1 +
i=1
s∈[a]k−1
=
∞
X X
(−1)ℓ+1
ℓ
k−1
s∈[a]
ℓ=1
k−1
X
pfs,i wi
i=1
!ℓ
k−1
∞
k−1
X X
(−1)ℓ+1 X
ℓ Y f mi Y mi
=
wi
(ps,i )
m
ℓ
k−1
i=1
i=1
ℓ=1
kmk
=ℓ
s∈[a]
1
!
!
k−1
Y f
X
Y m
X (−1)kmk1 +1 kmk1 k−1
mi
i
(ps,i )
wi
=
m
kmk1
k−1
k−1
i=1
a∈Z≥0
a6=0
X (−1)kmk1 +1 kmk1
=
kmk1
m
k−1
a∈Z≥0
a6=0
k−1
Y
s∈[a]
wimi
i=1
!
i=1
Mm (X f ).
We note that
|Mm (X f )| =
X
k−1
Y
s∈[a]k−1 i=1
(pfs,i )mi ≤ ak−1 ln−(k−1)kmk1 (1/ǫ) < (10k)−kmk1 ,
where the first inequality follows from (23). Therefore, for kzk∞ ≤ 4 and so kwk∞ ≤ 5, we obtain
that
∞
∞
∞
X
X
X
kmk1
1 X
ℓ ℓ
−ℓ
f
ℓ
−ℓ
2−ℓ = 1 . (24)
(k − 1) 5 (10k) ≤
| ln(P (X , z))| ≤
kwk∞ (10k) ≤
m
ℓ
k−1
ℓ=1
ℓ=1
ℓ=1
a∈Z+
kmk1 =ℓ
This suffices to show that all the series above are absolutely convergent when kzk∞ ≤ 4, and thus
that their manipulations are valid, and that we get the principal branch of the logarithm. To
summarize, for all z ∈ Ck−1 with kzk∞ ≤ 4, and w ∈ Ck−1 with wi = zi − 1, i ∈ [k − 1], we have:
!
Y m
X (−1)kmk1 +1 kmk1 k−1
f
i
Mm (X f ) .
(25)
wi
ln(P (X , z)) =
m
kmk
1
k−1
i=1
a∈Z≥0
a6=0
Qk−1 mi
In particular, we have that the i=1
wi coefficient of ln(P (X f , z)) is an integer multiple of
Mm (X f )/kmk1 . This expansion is a Taylor series in the wi ’s, so this coefficient is equal to a partial
derivative, which we can extract by Cauchy’s integral formula. Now, suppose that X f , X g are
49
distinct elements of S. We have that:
f
g
|Mm (X ) − Mm (X )|/kmk1
kmk1
≤ |Mm (X ) − Mm (X )|
/kmk1
m
!
k−1
Y
∂ kmk1 ln(P (X f , z)) − ln(P (X g , z))
1/mi !
=
mk−1
∂w1m1 · · · ∂wk−1
i=1
I
I
k−1
Y m
wi i dw1 . . . dwk−1
= (1/2πi)k−1 . . . (ln(P (X f , z)) − ln(P (X g , z)))/
f
g
γ
γ
≤
=
max
|w1 |=1,...,|wk−1 |=1
i=1
| ln(P (X f , z)) − ln(P (X g , z)))/
f
max
|w1 |=1,...,|wk−1 |=1
g
k−1
Y
i=1
wimi |
| ln(P (X , z)) − ln(P (X , z)))| ,
where the third line above follows from Cauchy’s integral formula, and γ is the path round the unit
circle.
Now suppose that there exists an m ∈ [a]k−1 , i.e., with kmk1 ≤ (k − 1)a, such that it holds
∗ )
|Mm (X f ) − Mm (X g )| ≥ ǫ4c . By the above, this implies that there is some w∗ = (w1∗ , . . . , wk−1
∗
∗
with |wi | = 1 for all i so that for the corresponding z ,
| ln(P (X f , z ∗ )) − ln(P (X g , z ∗ ))| ≥ ǫ4c /kmk1 ≥ ǫ4c /(ka) .
(26)
Note that kz ∗ k∞ ≤ kw∗ k∞ +1 = 2. Hence, z ∗ ∈ R. Applying (24), at this z ∗ , we have | ln(P (X f , z ∗ ))| ≤
1 and | ln(P (X g , z ∗ ))| ≤ 1. Therefore, by Equation (26), for this z ∗ with kz ∗ k∞ ≤ 2, we have that
|P (X f , z ∗ ) − P (X g , z ∗ )| = Ω ǫ4c /(ka) ≥ ǫ5c ,
where the last inequality follows from our definition of a. This completes the proof of Claim 4.31.
We are now ready to translate a lower bound on the distance between the PGFs to a lower
bound on total variation distance. Namely, we prove the following:
Claim 4.32. If there exists z ∗ ∈ Ck−1 with kz ∗ k∞ ≤ 2 such that |P (X f , z ∗ ) − P (X g , z ∗ )| ≥ ǫ5c ,
then dTV (X f , X g ) ≥ ǫ.
The main idea of the proof of Claim 4.32 is this: By Equation (24), we know that | ln(P (X f , z))| ≤
1, for all z ∈ R. We use this fact to show that the contribution to the value of the PGF P (X f , z ∗ )
coming from the subset of the probability space {X f > T } is at most O(2−T ). On the other
hand, the contribution to the difference | ln(P (X f , z ∗ )) − ln(P (X g , z ∗ ))| coming from the set
{X f ≤ T, X g ≤ T } can be easily bounded from above by 2T · dTV (X f , X g ). The claim follows by
selecting an appropriate value of T = Θ(log(1/ǫ)), balancing these two terms.
Proof. First note that exponentiating Equation (24) at z = (4, 4, . . . , 4), and using the definition
of the PGF we get:
h Pk−1 f i h Pk−1 g i
E 4 i=1 Xi , E 4 i=1 Xi ≤ e .
Therefore, for any z with kzk∞ ≤ 2 and any T ∈ Z+ we have that
X
k−1
Y
x,|x|1 ≥T i=1
zixi
f
X
T
Pr[X = x] ≤ (1/2)
k−1
Y
x,|x|1 ≥T i=1
50
4|x|1 Pr[X f = x] ≤ e(1/2)T .
A similar bound holds for X g . By assumption, there exists such a z ∗ so that
ǫ5c ≤ P (X f , z ∗ ) − P (X g , z ∗ )
=
Y
X k−1
(z ∗ )xi i Pr[X f = x] − Pr[X g = x]
x
=
i=1
Y
X k−1
(z ∗ )xi i Pr[X f = x] − Pr[X g = x] + 2e(1/2)T
|x|1 <T i=1
≤ 2T
T
X
|x|1 <T
Pr[X f = x] − Pr[X g = x] + 2e(1/2)T
≤ 2 dTV (X f , X g ) + 2e/2T .
Taking T = ⌈5c log 2 (1/ǫ)⌉, we get
dTV (X f , X g ) ≥ Ω(ǫ10c ) ≥ ǫ.
This completes the proof Claim 4.32.
Lemma 4.30 follows by combining Claims 4.31 and 4.32. By putting together Lemmas 4.29
and 4.30, it follows that any two distinct elements of S are ǫ-separated in total variation distance. This completes the proof of Theorem 4.28, establishing the correctness of our lower bound
construction.
5
A Size–Free Central Limit Theorem for PMDs
In this section, we prove our new CLT thereby establishing Theorem 1.3. For the purposes of this
section, we define a discrete Gaussian in k dimensions to be a probability distribution supported
on Zk so that the probability of a point x is proportional to eQ(x) , for some quadratic polynomial
Q. The formal statement of our CLT is the following:
Theorem 5.1. Let X be an (n, k)-PMD with covariance matrix Σ. Suppose that Σ has no eigenvectors other than 1 = (1, 1, . . . , 1) with eigenvalue less than σ. Then, there exists a discrete Gaussian
G so that
q
7/2
dTV (X, G) ≤ O(k
log3 (σ)/σ).
We note that our phrasing of the theorem above is slightly different than the CLT statement
of [VV10]. More specifically, we work with (n, k)-PMDs directly, while [VV10] work with projections
of PMDs onto k − 1 coordinates. Also, our notion of a discrete Gaussian is not the same as the
one discussed in [VV10]. At the end of the section, we show how our statement can be rephrased
to be directly comparable to the [VV10] statement.
Proof of Theorem 5.1. We note that unless σ > k7 that there is nothing to prove, and thus we
will assume this throughout the rest of the proof.
The basic idea of the proof will be to compare the Fourier transform of X to that of the
discrete Gaussian with density proportional to the pdf of N (µ, Σ) (where µ is the expectation of
X). By taking the inverse Fourier transform, we will be able to conclude that these distributions
are pointwise close. A careful analysis of this combined with the claim that both X and G have
small effective support will yield our result.
51
We start by providing a summary of the main steps of the proof. We start by bounding the
effective support of X under our assumptions (Lemma 5.2 and Corollary 5.3). Then, we describe
the effective support of its Fourier transform (Lemma 5.5). We further show that the effective
support of the distribution X and the Fourier transform of the discrete Gaussian G are similar (see
Lemmas 5.9 and 5.10). We then obtain an estimate of the error between the Fourier transforms of
X and a Gaussian with the same mean and covariance (Lemma 5.8). The difference between the
distributions of X and G at a point, as given by the inverse Fourier transform, is approximately
equal to the integral of this error over the effective support of the Fourier transform of X and G.
If we take bounds on the size of this integral naively, we get a weaker result than Theorem 5.1,
concretely that dTV (X, G) ≤ O(log(σ))k σ −1/2 (Proposition 5.11). Finally, we are able to show the
necessary bound on this integral by using the saddlepoint method.
We already have a bound on the effective support of a general PMD (Lemma 3.3). Using this
lemma, we obtain simpler bounds that hold under our assumptions.
Lemma 5.2. Let X be an (n, k)-PMD with mean µ and covariance matrix Σ, where all non-trivial
eigenvalues of Σ are at least σ, then for any ǫ > exp(−σ/k), with probability 1 − ǫ over X we have
that
(X − µ)T (Σ + I)−1 (X − µ) = O(k log(k/ǫ)).
Proof. From Lemma 3.3, we have that (X − µ)T (k ln(k/ǫ)Σ + k2 ln2 (k/ǫ)I)−1 (X − µ) = O(1) with
probability at least 1 − ǫ/10.
By our√assumptions on Σ, we can write Σ = U T diag(λi )U, for an orthogonal matrix U with kth
column 1/ k and λi ≥ σ for 1 ≤ i ≤ k − 1, and λk = 0.
By our assumptions on ǫ, we have that σ ≥ k ln(1/ǫ), and so for 1 ≤ i ≤ k − 1, we have
λi + 1 ≥ 12 (λi + k ln(1/ǫ)).
Since 1T (X − µ) = 0, we have that (U T (X − µ))k = 0, and so we can write
(X − µ) · (Σ + I)−1 (X − µ) = (U T (X − µ))T diag(1/(λi + 1))U T (X − µ)
≤ (U T (X − µ))T diag(2/(λi + k ln(1/ǫ)))U T (X − µ)
−1
= 2k ln(k/ǫ) · (X − µ)T k ln(k/ǫ)Σ + k 2 ln2 (k/ǫ)I
(X − µ)
= O(k ln(k/ǫ)).
Specifically, if we take ǫ = 1/σ, we have the following:
Corollary 5.3. Let X be as above, and let S be the set of points x ∈ Zk where (x − µ)T 1 = 0 and
(x − µ)T (Σ + I)−1 (x − µ) ≤ (Ck log(σ)) ,
for some sufficiently large constant C. Then, X ∈ S with probability at least 1 − 1/σ, and
p
|S| = det(Σ + I) · O(log(σ))k/2 .
Proof. Noting that ln(kσ) = O(log σ) since σ > k, by Lemma 5.2, applied with ǫ = 1/σ, it follows
that x ∈ S with probability 1 − 1/σ.
The remainder of the claim is a standard counting argument where we need to bound the
number of integer lattice points within a continuous region (in this case, an ellipsoid). We deal
with this by way of the standard technique of erecting a unit cube about each of the lattice points
and bounding the volume of the union. Note that for x ∈ Zk , the cubes x + (−1/2, 1/2)k each have
52
volume one and are disjoint. Thus, if we define S ′ to be the set of y such that there exists an x ∈ S
with ky − xk∞ < 21 , then |S| = Vol(S ′ ). For any y ∈ S ′ , there is an x ∈ S with:
(y − µ) · (Σ + I)−1 (y − µ) = (x − µ) · (Σ + I)−1 (x − µ) + (y − x) · (Σ + I)−1 (y − x) +
2(y − x) · (Σ + I)−1 (x − µ)
≤ O((x − µ) · (Σ + I)−1 (x − µ) + (y − x) · (Σ + I)−1 (y − x))
≤ O(Ck log(σ) + (y − x) · I(y − x))
≤ O(Ck log(σ) + kky − xk2∞ )
≤ O(Ck log(σ) + k)
= O(Ck log(σ)) .
That is, S ′ is contained in the ellipsoid (y − µ) · (Σ + I)−1 (y − µ) ≤ O(Ck log(σ)). The corollary
follows by bounding the volume of this ellipsoid. We have the following simple claim:
Claim
of the ellipsoid xT A−1 x ≤ ck for a symmetric k × k matrix A and c > 0
p 5.4. The volume
is det(A) · O(c)k/2 .
Proof. We can factorize A√= U T diag(λi )U T for some orthogonal matrix U. Then, the ellipsoid is
the set of x with kdiag(1/ ckλi )U T xk2 ≤ 1. The volume of the ellipsoid is
p
p
det(diag(1/ ckλi )U T )−1 Vk = det(A) · (ck)k/2 Vk ,
where Vk is the volume of the unit sphere. By standard results, Vk = π k/2 /Γ(1 + k/2) = Ω(k)−k/2 ,
using Stirling’s approximation
p
Γ(1 + k/2) = 2π/(1 + k/2)((1 + k/2)/e)1+k/2 (1 + O(2/(k + 2)).
p
Therefore, the volume is O( det(A) · (c)k/2 ).
volume of the ellipsoid (y − µ) · (Σ + I)−1 (y
p − µ) ≤ O(Ck log(σ)) is
p As a consequence, the
k/2
det(Σ + I) · O(C log σ) . Thus, we conclude that |S| ≤ Vol(S ′ ) ≤ det(Σ + I) · O(log σ)k/2 .
This completes the proof of the corollary.
b has
Next, we proceed to describe the Fourier support of X. In particular, we show that X
a relatively small effective support, T . Our Fourier sparsity lemma in this section is somewhat
different than in previous section, but the ideas are similar. The proof will similarly need Lemma
3.10.
def
Lemma 5.5. Let T = {ξ ∈ Rk | ξ · Σξ ≤ Ck log(σ)}, for C some sufficiently large constant. Then,
we have that:
p
(i) For all ξ ∈ T, the entries of ξ are contained in an interval of length 2 Ck log(σ)/σ.
(ii) Letting T ′ = T ∩ {ξ ∈ Rk | ξ1 ∈ [0, 1]}, it holds Vol(T ′ ) = det(Σ + I)−1/2 · O(C log(σ))k/2 .
R
b
(iii) [0,1]k \(T +Zk ) |X(ξ)|dξ
≤ 1/(σ|S|).
53
Proof. We define ξ̃ to be the projection of ξ onto the plane where the coordinates sum to 0, i.e.,
ξ̃ = ξ + α1 for some α ∈pR and ξ˜ · 1 = 0. Then, we have that ξ · Σξ ≥ σ|ξ̃|22 . Hence, for ξ ∈ T, we
have that |ξ̃|∞ ≤ |ξ̃|2 ≤ Ck log(σ)/σ. This implies that for any i, j it holds
p
|ξi − ξj | = |ξ̃i − ξ˜j | ≤ 2|ξ̃|∞ ≤ 2 Ck log(σ)/σ.
This proves (i).
p
In particular, for√any ξ ∈√T, and all i, we have |ξ1 − ξi | ≤ 2 Ck log(σ)/σ ≤ 2C. And so if
ξ ∈ T ′ , then ξi ∈ [−2 C, 1+2 C]. Thus, for ξ ∈ T ′ , it holds ξ ·ξ ≤ O(C). Thus, for ξ ∈ T ′ , we have
ξ · (Σ + I) · ξ ≤ O(Ck log(σ)). By Claim 5.4, we get that Vol(T ′ ) ≤ det(Σ + I)−1/2 O(C log(σ))k/2 .
This proves (ii).
By Claim 3.9, for every ξ ∈ [0, 1)k , there is an interval Iξ of length 1 − 1/(k + 1) such that
′
ξ = ξ + b, for some b ∈ Zk , has coordinates in Iξ . Let Tm be the set of ξ such that there is such a
ξ ′ with
2m+1 Ck log(σ) ≥ ξ ′ · Σξ ′ ≥ 2m Ck log(σ) ,
and ξi = ξi′ − ⌊ξi′ ⌋ for all 1 ≤ i ≤ k. Then, for every ξ ∈ [0, 1]k , we either have ξ ∈ T + Zk or else
ξ ∈ Tm for some m ≥ 0. Hence,
Z
[0,1]k \(T +Zk )
b
|X(ξ)|dξ
≤
∞
X
m=0
b
Vol(Tm ) sup |X(ξ)|
.
(27)
ξ∈Tm
To bound the RHS above, we need bounds on the volume of each Tm . These can be obtained
using a similar argument to (ii) along with some translation.
Claim 5.6. We have that Vol(Tm ) ≤ det(Σ + I)−1/2 · O(2m+1 C log(σ))k/2 .
Proof. Let Um be the set of ξ such that there is a ξ ′ with 2m+1 Ck log(σ) ≥ ξ ′ · Σξ ′ and ξi = ξi′ − ⌊ξi′ ⌋
′ be the set of ξ ′ with ξ ∈ [0, 1] and 2m+1 Ck log(σ) ≥
for all 1 ≤ i ≤ k. Note that Tm ⊆ Um . Let Um
1
′
′
′′
m+1
ξ · Σξ . Note that for any ξ with 2
Ck log(σ) ≥ ξ ′′ · Σξ ′′ , we have that ξ ′′ + λ1 also satisfies
′ .
2m+1 Ck log(σ) ≥ (ξ ′′ + λ1) · Σ(ξ ′′ + λ1) for any λ ∈ R. In particular ξ ′ = ξ ′′ − (⌊ξ1′′ ⌋)1) ∈ Um
′′
′′
′
′
′
Note that ξi − ⌊ξi ⌋ = ξi for all i. So Um is the set of ξ such that there is a ξ ∈ Um with
Qk
′ ) since U
′
ξi = ξi′ − ⌊ξi′ ⌋. Then, Vol(Um ) ≤ Vol(Um
m = ∪b∈Zn (Um ∩
i=1 [bi , bi + 1)) − b, and so
Qk
P
′
′
Vol(Um ) ≤ b∈Zn Vol(Um ∩ i=1 [bi , bi + 1))) = Vol(Um ).
Note that by Lemma 5.5 (ii) applied with C := 2m+1 C gives the bound
′
Vol(Um
) ≤ det(Σ + I)−1/2 · O(2m+1 C log(σ))k/2 .
′ ) ≤ det(Σ + I)−1/2 · O(2m+1 C log(σ))k/2 . This
Therefore, we have Vol(Tm ) ≤ Vol(Um ) ≤ Vol(Um
completes the proof.
b
by using Lemma 3.10.
Next, we obtain bounds on supξ∈Tm |X(ξ)|
b
Claim 5.7. For ξ ∈ Tm , it holds |X(ξ)|
≤ exp(−Ω(C2m log(σ)/k)). If additionally we have m ≤
m
b
4 log2 k, then |X(ξ)| = exp(−Ω(C2 k log(σ))).
Proof. Note that ξ ′ has coordinates in an interval of length 1 − 1/k, so we may apply Lemma 3.10,
yielding
b
b ′ )| ≤ exp(−Ω(ξ ′T · Σ · ξ ′ /k2 )) = exp(−Ω(C2m log(σ)/k)).
|X(ξ)|
= |X(ξ
54
To get the stronger bound, we need to show that for small m, all the coordinates of ξ ′ are in a
shorter interval. As before, we consider ξ˜′ , the projection of ξ ′ onto the set of x with x · 1 = 0.
Similarly, we have ξ ′ · Σξ ′ ≥ σ|ξ˜′ |22 . So, for any i, j, it holds
p
p
|ξi′ − ξj′ | ≤ |ξ˜′ i − ξ˜′ j | ≤ 2|ξ˜′ |∞ ≤ 2|ξ˜′ |2 ≤ ξ ′ · Σξ ′ /σ ≤ C2m+1 k log(σ)/σ.
For m ≤ log2 (σ/Ck log(σ)) − 3, we have that the coordinates of ξ lie in an interval of length 1/2.
Now, Lemma 3.10 gives that
b
b ′ )| ≤ exp(−Ω(ξ ′T · Σ · ξ ′ )) = exp(−Ω(Ck2m log(σ)/k)).
|X(ξ)|
= |X(ξ
Finally, note that 4 log2 k ≤ log2 (σ/Ck log(σ)) − 3, when σ ≥ Ck 3 . This completes the proof of the
claim.
Using the above, we can write
Z
[0,1]k \(T +Zk )
b
|X(ξ)|dξ
≤
∞
X
m=0
b
Vol(Tm ) sup |X(ξ)|
ξ∈Tm
−1/2
≤ det(Σ + I)
· O(C log(σ))
We divide this sum into two pieces:
4 log2 k
X
mk/2
2
m=0
b
sup |X(ξ)|
≤
≤
∞
X
m=0
b
2mk/2 sup |X(ξ)|
.
ξ∈Tm
log2 (σ/Ck log(σ))−3
ξ∈Tm
≤
k/2
X
2mk/2 exp(−Ω(C2m k log(σ)))
m=0
4 log2 k
X
m=0
4 log2 k
X
exp(−Ω(C(2m − m)k log(σ)))
2−m exp(−Ω(Ck log(σ)))
m=0
≤ exp(−Ω(Ck log(σ))) = σ −Ω(Ck) ,
and
∞
X
m=4 log2 k
b
2mk/2 sup |X(ξ)|
≤
ξ∈Tm
≤
≤
≤
We thus have
R
b
[0,1]k \(T +Zk ) |X(ξ)|dξ
∞
X
2mk/2 exp(−Ω(C2m log(σ)/k))
m=4 log2 k
∞
X
m=4 log2 k
∞
X
m=4 log2 k
∞
X
m=4 log2 k
exp(−Ω(C(2m − m) log(σ)/k))
exp(−Ω(C(k2 + m) log(σ)/k))
2−m exp(−Ω(Ck log(σ))) ≤ σ −Ω(Ck) .
≤ det(Σ + I)−1/2 O(C log(σ))k/2 log(σ)−O(Ck) .
55
The previous lemma establishes that the contribution to the Fourier transform of X coming
from points outside of T is negligibly small. We next claim that, for ξ ∈ T, it is approximated by
a Gaussian.
Lemma 5.8. For ξ ∈ T, we have that
q
2
3/2
7/2
3
b
X(ξ)
= exp 2πiµ · ξ − 2π ξ · Σξ + O(C k
log (σ)/σ) .
This also holds for complex ξ, under the assumption that the coordinate-wise complex and real parts
of ξ are in T, i.e., such that Re(ξ) · ΣRe(ξ), Im(ξ) · ΣIm(ξ) ≤ O(Ck log(σ)).
Q P
b
Proof. Recall that X(ξ)
= ni=1 kj=1 e(ξj )pij . Let mi be the element of [k] so that pimi is as
large as possible for each i. In particular,
Pk pimi ≥ 1/k. We will attempt to approximate the above
product by approximating the log of j=1 e(ξj )pij by its Taylor series expansion around the point
(ξmi , ξmi , . . . , ξmi ). In particular, by Taylor’s Theorem, we find that
k
X
j=1
e(ξj )pij = exp 2πi ξmi +
k
X
j=1
2
k
k
X
X
pij (ξj − ξmi ) + Ei ,
pij (ξj − ξmi )2 + 2π 2
pij (ξj − ξmi ) − 2π 2
j=1
j=1
ci (ξ)) at some
where Ei is the third directional derivative in the ξ − (ξmi , . . . , ξmi ) direction of log(X
˜
point ξ along the line between ξ and (ξmi , . . . , ξmi ). Note that the above is exactly
exp 2πi(ξ · E[Xi ]) − 2π 2 (ξ · Cov(Xi )ξ) + Ei .
Thus, taking a product over i, we find that
b
X(ξ)
= exp 2πiµ · ξ − 2π 2 ξ · Σξ +
n
X
i=1
Ei
!
.
We remark that the coefficients of this Taylor series are (up to powers of −2πi) the cumulants
of X.
P
Since the coordinates of ξ̃ lie in an interval of length at most 1/2, we have that kj=1 e(ξ̃j )pij
is bounded away from 0. Therefore, we get that
|Ei | = O
X
k
j=1
pij |ξ̃j − ξmi |3 +
+
k
X
j1 ,j2 =1
k
X
pij1 pij2 |ξ̃j1 − ξmi |2 |ξ̃j2 − ξmi |
j1 ,j2 ,j3 =1
pij1 pij2 pij3 |ξ̃j1 − ξmi ||ξ̃j2 − ξmi ||ξ̃j3 − ξmi | .
Next note that
Var(Xi · ξ̃) ≥ pij pimi |ξ̃j − ξmi |2 ≥ pij |ξ̃j − ξmi |2 /k.
Additionally, note that
Ck log(σ) ≥ ξ · Σξ = Var(X · ξ) =
56
n
X
i=1
Var(Xi · ξ).
Therefore,
n
X
i=1
2
pij |ξ̃j − ξmi | ≤
n
X
i=1
pij |ξj − ξmi |2 ≤ Ck 2 log(σ).
p
Thus, since |ξ̃j − ξmi | = O( Ck log(σ)/σ) for all i, j we have that
n X
k
X
i=1 j=1
3
pij |ξ̃j − ξmi | ≤ O(C
3/2 7/2
k
q
log3 (σ)/σ).
We have that
k
n
X
X
i=1 j1 ,j2 =1
pij1 pij2 |ξ̃j1 − ξmi ||ξ̃j2 − ξmi | ≤
≤
n
X
i=1
n
X
i=1
k
X
j=1
k
X
j=1
3
2
pi,j |ξ̃j − ξmi |
k
X
pi,j
pi,j |ξ̃j − ξmi |2
j=1
= O(Ck log(σ)).
Therefore,
k
n
X
X
i=1 j1 ,j2 =1
and
k
X
j1 ,j2 ,j3 =1
pij1 pij2 |ξ̃j1 − ξmi |2 |ξ̃j2 − ξmi |
pij1 pij2 pij3 |ξ̃j1 − ξmi ||ξ̃j2 − ξmi ||ξ̃j3 − ξmi |
are both
O(C
3/2 7/2
k
q
log3 (σ)/σ).
We now define G to be the discrete Gaussian supported on the set of points in Zk whose
coordinates sum to n, so that for such a point p we have:
Z
−(k−1)/2
′ −1/2
−1
e(−p · ξ) exp(2πi(ξ · µ) − 2π 2 ξ · Σξ)
G(p) = (2π)
det(Σ )
exp((p − µ) · Σ (p − µ)/2) = P
ξ, ξj =0
Z
e(−p · ξ) exp(2πi(ξ · µ) − 2π 2 ξ · Σξ) ,
=
ξ,ξ1 ∈[0,1]
where Σ′ = Σ + 11T restricted to the space of vectors whose coordinates sum to 0.
b equal
We let G
b
G(ξ)
:= exp(2πi(ξ · µ) − 2π 2 ξ · Σξ).
b and X
b do
Next, we claim that G and X have similar effective supports and subsequently that G
as well. Firstly, the effective support of the distribution of G is similar to that of X, namely S:
Lemma 5.9. The sum of the absolute values of G at points not is S is at most 1/σ.
57
Proof. For this it suffices to prove a tail bound for G analogous to that satisfied by X. In particular,
assuming
that Σ has unit eigenvectors vi with eigenvalues λi , it suffices to prove that |(G − µ) · vi | <
√
λi t except with probability at most exp(−Ω(t2 )). Recall that
G(p) = (2π)−(k−1)/2 det(Σ′ )−1/2 exp((p − µ) · Σ−1 (p − µ)/2).
Let G̃ be the continuous probability density defined by
G̃(x) = (2π)−k/2 det(Σ′ )−1/2 exp((x − µ) · Σ′−1 (x − µ)/2).
Note that for any p with (p − µ) · 1 = 0, and x ∈ [−1/2, 1/2]k , we have that
G(p) = O G̃(p + x) + G̃(p − x) .
Therefore, we have that
G(p) = O
Z
!
G̃(x) .
x∈p+[−1/2,1/2]k
√
√
Applying
this
formula
for
each
p
with
(p−µ)·v
≥
λ
t
and
noting
that
(x−µ)·v
≥
(p−µ)·v
−
k≥
i
i
i
i
√
√
λi t − k yields
p
p
√
Pr(|(G − µ) · vi | > λi t) = O(Pr(|(G̃ − µ) · vi | > λi t − k = exp(−Ω(t2 )).
Taking a union bound over 1 ≤ i ≤ k yields our result.
Secondly, the effective support of the Fourier Transform of G is similar to that of X, namely T :
b
Lemma 5.10. The integral of |G(ξ)|
over ξ with ξ1 ∈ [0, 1] and ξ not in T is at most 1/(|S|σ).
Proof. We consider the integral over ξ ∈ Tm , where
Tm := {ξ : ξ1 ∈ [0, 1] | ξ · Σξ ∈ [2m Ck log(σ), 2m+1 Ck log(σ)]}.
b
We note that it has volume 2mk kO(k) logO(k) (σ)/|S|, and that within Tm it holds |G(ξ)|
= exp(−Ω(Ck log(σ)2m )).
From this it is easy to see that the integral over Tm is at most 2−m−1 /(|S|σ). Summing over m
yields the result.
We now have all that is necessary to prove a weaker version of our main result.
Proposition 5.11. We have the following:
dTV (X, G) ≤ O(log(σ))k σ −1/2 .
Proof. First, we bound the L∞ of the difference. In particular, we note that for any p with integer
coordinates summing to n we have that
Z
b
e(−p · ξ)X(ξ)dξ
,
X(p) =
ξ∈Rk ,ξ1 ∈[0,1],ξi ∈[ξ1 −1/2,ξ1 +1/2]
and
G(p) =
Z
ξ,ξ1 ∈[0,1]
e(−p · ξ) exp(2πi(ξ · µ) − 2π 2 ξ · Σξ).
58
We note that in both cases the integral for ξ not in T is at most 1/(|S|σ). To show this, we need to
note that any ξ with ξ1 ∈ [0, 1], ξi ∈ [ξ1 − 1/2, ξ1 + 1/2] equivalent to a point of T modulo Zk must
lie in T itself. This is because the element of T must have all its coordinates differing by at most
1/4, and thus must differ from ξ by an integer multiple of (1, 1, . . . , 1). Therefore, we have that
|X(p) − G(p)| =
≤
≤
Z
Z
Z
ξ∈T,ξ1 ∈[0,1]
ξ∈T,ξ1 ∈[0,1]
b
b
+ O(1/(|S|σ))
e(−p · ξ)(X(ξ)
− G(ξ))dξ
b
b
|X(ξ)
− G(ξ)|dξ
+ O(1/(|S|σ))
O(C
3/2 7/2
k
ξ∈T,ξ1 ∈[0,1]
≤ det(Σ + I)−1/2 (C 3/2
p
q
log3 (σ)/σ)dξ + O(1/(|S|σ))
log(σ)/σ)O(C log(σ))k/2 .
Therefore, the sum of |X(p) − G(p)| over p ∈ S is at most
p
(C 3/2 log(σ)/σ)O(C 2 log2 (σ))k/2 .
The sum over p 6∈ S is at most O(1/σ). This completes the proof.
The proof of the main theorem is substantially the same as the above. The one obstacle that
we face is that above we are only able to prove L∞ bounds on the difference between X and G, and
these bounds are too weak for our purposes. What we would like to do is to prove stronger bounds
on the difference between X and G at points p far from µ. In order to do this, we will need to take
advantage of cancellation in the inverse Fourier transform integrals. To achieve this, we will use
the saddle point method from complex analysis.
Proof of Theorem 5.1. For p ∈ S we have as above that
|X(p) − G(p)| =
Z
ξ∈T,ξ1 ∈[0,1]
b
b
+ O(1/(|S|σ)).
e(−p · ξ)(X(ξ)
− G(ξ))dξ
Let ξ0 ∈ Rk be such that ξ0 .1 = 0 and so that Σξ0 = (µ − p)/(2π) (i.e., take ξ0 = (Σ + 11T )−1 µ −
p)/(2π)). We think of the integral above as an iterated contour integral. By deforming the contour
associated with the innermost integral, we claim that it is the same as the sum of the integrals over
ξ with Re(ξ) ∈ T, Re(ξ1 ) ∈ [0, 1] and Im(ξ) = ξ0 and the integral over Re(ξ) ∈ δT, Re(ξ1 ) ∈ [0, 1]
and Im(ξ) = tξ0 for some t ∈ [0, 1] (the extra pieces that we would need to add at Re(ξ1 ) = 0 and
Re(ξ1 ) = 1 cancel out).
R
b
b
Claim 5.12. ξ∈δT,ξ1 ∈[0,1] e(−p · ξ)(X(ξ)
− G(ξ))dξ
equals
Z
ξ∈T,ξ1 ∈[0,1]
plus
Z
ξ∈δT,ξ1 ∈[0,1]
Z
1
t=0
b + iξ0 ) − G(ξ
b + iξ0 ))dξ
e(−p · (ξ + iξ0 ))(X(ξ
b + itξ0 ) − G(ξ
b + itξ0 ))d(tξ0 ) · dξ.
e(−p · (ξ + itξ0 ))(X(ξ
59
b
b
Proof. We write f (ξ) = e(−p · ξ)(X(ξ)
− G(ξ)).
Let O be an orthogonal matrix with kth column
ξ0 /kξ0 k2 . Then, we change variables from ξ to ν = O T ξ, yielding
Z
Z
f (OT ν)dν
f (ξ)dξ =
ν∈O T T ′
ξ∈T ′
We can consider this as an iterated integral where νi is integrated from ai (ν1 , . . . , νi−1 ) to bi (ν1 , . . . , νi−1 ).
Z
T
f (O ν)dν =
ν∈O T T ′
Z
b1
a1
Z
b2 (ν1 )
...
a2 (ν1 )
Z
bk (ν1 ,...,νk−1 )
f (O T ν)dνk . . . dν2 dν1 .
ak (ν1 ,...,νk−1 )
We consider the innermost integral. The function f (OT ν) is a linear combination of exponentials
and so is holomorphic on all of Cn . Let C be the contour which consists of three straight lines,
from ak (ν1 , . . . , νk−1 ) via ak (ν1 , . . . , νk−1 ) + ikξ0 k2 and bk (ν1 , . . . , νk−1 ) + ikξ0 k2 to bk (ν1 , . . . , νk−1 ).
Then, by standard facts of complex analysis, we have:
Z
bk (ν1 ,...,νk−1 )
f (OT ν)dνk
ak (ν1 ,...,νk−1 )
=
Z
f (OT ν)dνk
C
=
Z
1
0
Z
f (OT (ν1 , . . . , νk−1 , ak (ν1 , . . . , νk−1 ) + ikξ0 k2 t))ikξ0 k2 dt
bk (ν1 ,...,νk−1 )
f (OT (ν + ikξ0 k2 ek ))dνk
ak (ν1 ,...,νk−1 )
Z 1
f (OT (ν1 , . . . , νk−1 , bk (ν1 , . . . , νk−1 ) +
+
0
+
ikξ0 k2 (1 − t′ )))ikξ0 k2 dt′
The middle part of this path gives the first term in the statement of the claim:
Z
b1
a1
=
Z
Z
b2 (ν1 )
a2 (ν1 )
ν∈O T T ′
=
Z
ξ∈T ′
...
Z
bk (ν1 ,...,νk−1 )
ak (ν1 ,...,νk−1 )
f (OT (ν + ikξ0 k2 ek ))dνk . . . dν2 dν1
f (OT (ν + ikξ0 k2 ek ))
f (ξ + iξ0 ) .
A change of variables allows us to express the sum of the contributions from the first and third
part of the path:
Z 1
f (O T (ν1 , . . . , νk−1 , ak (ν1 , . . . , νk−1 ) + ikξ0 k2 t))ikξ0 k2 dt
0
Z 1
f (O T (ν1 , . . . , νk−1 , bk (ν1 , . . . , νk−1 ) + ikξ0 k2 t′ ))ikξ0 k2 dt′ .
−
0
Changing variables to replace (ν1 , . . . , νk−1 , ak (ν1 , . . . , νk−1 )) or (ν1 , . . . , νk−1 , bk (ν1 , . . . , νk−1 )) with
ξ ∈ δT or ξ ∈ T ∩ {0, 1} we get an appropriate integral of ±if (ξ + itξ0 ). We note that the volume
form for ξ0 assigns to a surface element the volume of the projection of that element in the ξ0
direction. Multiplying by kξ0 k2 and the appropriate sign yields exactly the measure ξ0 · dξ. Thus,
60
we are left with an integral of f (ξ + itξ0 )d(tξ0 ) · dξ. However, it should be noted that the measures
ξ0 · dξ are opposite on ξ1 = 0 and ξ1 = 1 boundaries (as dξ is the outward pointing normal). Since
f (ξ + itξ0 ) = f (ξ + 1+ itξ0), the integrals over these regions cancel, leaving exactly with the claimed
integral.
In order to estimate this difference, we use Lemma 5.8, which still applies. Furthermore, we
note that
ξ0 · Σξ0 = (p − µ) · Σ−1 (p − µ)/(4π 2 ) = O(Ck log(σ)) ,
because p ∈ S.
Therefore, we have that |X(p) − G(p)| is O(1/(|S|σ)) plus
Z
q
7/2
3
O(k
exp 2πi(µ − p) · ξ − 2π 2 ξ · Σξ dξ.
log (σ)/σ)
C
Now, when Im(ξ) = ξ0 , we have that
exp 2πi(µ − p) · ξ − 2π 2 ξ · Σξ = exp −(p − µ)Σ−1 (p − µ)/2 − 2π 2 Re(ξ) · ΣRe(ξ) .
Integrating, we find that the difference over this region is at most times
Z
q
q
−1
2
7/2
7/2
3
log (σ)/σ) exp −(p − µ)Σ (p − µ)/2 − 2π Re(ξ) · ΣRe(ξ) dξ = O(k
log3 (σ)/σ)G(p).
O(k
The contribution from the part of the contour where Re(ξ) is on the boundary of T is also easy to
b
b
bound after noting that both |X(ξ)|
and √
|G(ξ)|
are O(σ −k ). We furthermore claim that the total
volume of the region of integration is O( kVol(T )). Together, these would imply that the total
integral over this region is O(1/(σ|S|)). To do this, we note that the total volume of the region
being integrated over is at most the volume of the projection of T in the direction perpendicular
to ξ0 times the length of ξ0 . In order to analyze this we consider each slice of T given by ξ · 1 = α
separately. Noting that |α| ≤ 2 for all ξ ∈ δT, it suffices to consider only a single slice. In particular,
since for all such α, we have that ξ˜ ∈ T, it suffices to consider the slice α = 0. Along this slice, we
have that T is an ellipsoid. Note that ξ0 · Σξ0 = (p − µ) · (Σ + 11T )−1 (p − µ) ≤ Ck log(σ). Therefore
ξ0 ∈ T.
Next, we claim that if E is any ellipsoid in at most k dimensions, and if v is a vector with
v ∈ E, then the product
of the length of v times the volume of the projection of E perpendicular
√
to v is at most O( kVol(E)). This follows after noting that the claim is invariant under affine
transformations, and thus it suffices to consider E the unit ball for which it is easy to verify.
Therefore, for p ∈ S, we have that
q
7/2
log3 (σ)/σ)G(p).
|X(p) − G(p)| ≤ O(1/|S|σ) + O(k
From this it is easy to see that it is also
|X(p) − G(p)| ≤ O(1/|S|σ) + O(k
3/2
q
Summing over p ∈ S gives a total difference of at most
q
O(k7/2 log3 (σ)/σ).
log3 (σ)/σ)X(p).
Combining this with the fact that the sum of X(p) and G(p) for p not in S is at most 1/σ gives us
that
q
7/2
dTV (X, G) = O(k
log3 (σ)/σ).
This completes the proof of Theorem 5.1.
61
Comparison to the [VV10] CLT. We note that the above statement of Theorem 5.1 is not
immediately comparable to the CLT of [VV10]. More specifically, we work with PMDs directly,
while [VV10] works with projections of PMDs onto k − 1 coordinates. Also, our notion of a discrete
Gaussian is not the same as the one discussed in [VV10]. However, it is not difficult to relate the
two results. First, we need to relate our PMD (supported on integer vectors whose coordinates sum
to n) to theirs (which are projections of PMDs onto k − 1 coordinates). In particular, we need to
show that this projection does not skew minimum eigenvalue in the wrong direction. This is done
in the following simple proposition:
Proposition 5.13. Let X be an (n, k)-PMD, and X ′ be obtained by projecting X onto its first
k − 1 coordinates. Let Σ and Σ′ be the covariance matrices of X and X ′ , respectively, and let σ and
σ ′ be the second smallest and smallest eigenvalues respectively of Σ and Σ′ . Then, we have σ ≥ σ ′ .
Proof. Note that, since 1 is in the kernel of Σ, σ is the minimum of v orthogonal to 1 of
Whereas, σ ′ is the minimum over w ∈ Rk−1 \{0} of
Rk
w T Σw
.
wT w
kth
w T Σ′ w
.
wT w
vT Σv
.
vT v
This is the same as the minimum over
w in
with
coordinate equal to 0 of
Let the minimization problem defining σ be obtained by some particular v orthogonal to 1. In
T
particular, a v so that σ = vvTΣv
. Let w be the unique vector of the form v + a1 so that w has last
v
coordinate 0. Then, we have that
σ′ ≥
wT Σw
v T Σv
v T Σv
=
≥
= σ.
wT w
wT w
vT v
This completes the proof.
Next, we need to relate the two slightly different notions of discrete Gaussian.
Proposition 5.14. Let G be a Gaussian in Rk with covariance matrix Σ, which has no eigenvalue
smaller than σ. Let G′ be the discrete Gaussian obtained by rounding the values of G to the nearest
lattice point. Let G′′ be the discrete distribution obtained by assigning each integer lattice point
mass proportional to the probability density function of G. Then, we have that
p
dTV (G′ , G′′ ) ≤ O(k log(σ)/σ).
Proof. We note that the probability density function of G is proportional to exp(−(x · Σ−1 x)/2)dx.
Suppose that y is another vector with kx − yk∞ < 1. We would like to claim that the probability
density function at y is approximately the same as at x. In particular, we write y = x + z and note
that
y · Σ−1 y = x · Σ−1 x + 2z · Σ−1 x + z · Σ−1 z
= x · Σ−1 x + 2(Σ−1/2 x) · (Σ−1/2 z) + O(|z|22 σ −1 )
p
= x · Σ−1 x + O(|Σ−1/2 x|2 k/σ + kσ −1 )
p
= x · Σ−1 x + O( (k/σ)x · Σ−1 x + kσ −1 ).
Note that, for lattice points x, G′ (x) is the average over y in a unit cube about
p x of the pdf of G−1at
′′
y, while G (x) is just the pdf of G at x. These quantities are within a 1+O( (k/σ)x · Σ−1 x+kσ )
multiple of each other by the above so long as the term in the “O”
p is o(1). Therefore, for all x
with x · Σ−1 x ≪ k log(σ), we have that G′ (x) = G′′ (x)(1 + O(k log(σ)/σ)). We note however
62
that G′ has only a 1/σ probability of x being outside of this range. Furthermore, we claim that
G′′ (x) = O(G′ (x)) for all x. To see this, note that for any v with kvk∞ ≤ 1/2, we have
p
T −1
G(x) = e−v Σ v/2 G(x + v)G(x − v) ≤ e−(k/2σ) max{G(x + v), G(x − v)}.
We assume that σ ≥ k2 or else we have nothing to prove. Then, we have G(x) = O(G(x + v) +
G(x − v)), and by considering the integral that defines G′ , we have G′′ (x) = O(G′ (x)). Thus, G′′
−1
similarly has O(1/σ) mass
p outside of the range x · Σ x ≪ k log(σ). Therefore, the L1 difference
inside the range is O(k log(σ)/σ) and the L1 error from outside is O(1/σ). This completes the
proof.
Armed with these propositions, we have the following corollary of Theorem 5.1:
Corollary 5.15. Let X be an (n, k)-PMD, and X ′ be obtained by projecting X onto its first k − 1
coordinates. Let Σ′ be the covariance matrix of X ′ . Suppose that Σ′ has no eigenvectors with
eigenvalue less than σ ′ . Let G′ be the distribution obtained by sampling from N (E[X ′ ], Σ′ ) and
rounding to the nearest point in Zk . Then, we have that
q
′
′
7/2
log3 (σ ′ )/σ ′ ).
dTV (X , G ) ≤ O(k
Proof. Let Σ be the covariance matrix of X. Since X is a PMD, 1 is an eigenvector of Σ with
′
eigenvalue 0. By Proposition
q 5.13, the other eigenvalues of Σ are at least σ . Theorem 5.1 now yields
that dTV (X, G) ≤ O(k 7/2 log3 (σ ′ )/σ ′ ), where G is a discrete Gaussian in k dimensions (as defined
in the context of the theorem statement). Let G′′ be the discrete Gaussian obtained
by projecting G
q
onto the first k − 1 coordinates. Then, we have that dTV (X ′ , G′′ ) ≤ O(k7/2 log3 (σ ′ )/σ ′ ). From the
proof of Theorem 5.1, G is proportional to the pdf of N (E[X], Σ). Note that G′′ is proportional
to
p
′ , G′′ ) ≤ O(k log(σ ′ )/σ ′ ).
the pdf of N (E[X ′ ], Σ′ ). Then, by Proposition 5.14, it follows that
d
(G
TV
q
So, by the triangle inequality, we have dTV (X ′ , G′ ) ≤ O(k7/2
6
log3 (σ ′ )/σ ′ ), as required.
Conclusions and Open Problems
In this work, we used Fourier analytic techniques to obtain a number of structural results on PMDs.
As a consequence, we gave a number of applications in distribution learning, statistics, and game
theory. We believe that our techniques are of independent interest and may find other applications.
Several interesting open questions remain:
• What is the precise complexity of learning PMDs? Our bound is nearly-optimal when the
dimension k is fixed. The case of high dimension is not well-understood, and seems to require
different ideas.
• Is there an efficient proper learning algorithm, i.e., an algorithm that outputs a PMD as its
hypothesis? This question is still open even for k = 2; see [DKS15b] for some recent progress.
• What is the optimal error dependence in Theorem 1.3 as a function of the dimension k?
• Is there a fully-polynomial time approximation scheme (FPTAS) for computing ǫ-Nash equilibria in anonymous games? We remark that cover-based algorithms cannot lead to such a
result, because of the quasi-polynomial cover size lower bounds in this paper, as well as in
our previous work [DKS15a] for the case k = 2. Progress in this direction requires a deeper
understanding of the relevant fixed points.
63
References
[Bar88]
A. D. Barbour. Stein’s method and poisson process convergence. Journal of Applied
Probability, 25:pp. 175–184, 1988.
[BCI+ 08]
C. Borgs, J. T. Chayes, N. Immorlica, A. T. Kalai, V. S. Mirrokni, and C. H. Papadimitriou. The myth of the folk theorem. In STOC, pages 365–372, 2008.
[BDS12]
A. Bhaskara, D. Desai, and S. Srinivasan. Optimal hitting sets for combinatorial shapes.
In 15th International Workshop, APPROX 2012, and 16th International Workshop,
RANDOM 2012, pages 423–434, 2012.
[Ben03]
V. Bentkus. On the dependence of the Berry-Esseen bound on dimension. Journal of
Statistical Planning and Inference, 113:385–402, 2003.
[BHJ92]
A.D. Barbour, L. Holst, and S. Janson. Poisson Approximation. Oxford University
Press, New York, NY, 1992.
[Blo99]
M. Blonski. Anonymous games with binary actions. Games and Economic Behavior,
28(2):171 – 180, 1999.
[Blo05]
M. Blonski. The women of cairo: Equilibria in large anonymous games. Journal of
Mathematical Economics, 41(3):253 – 264, 2005.
[CDO15]
X. Chen, D. Durfee, and A. Orfanou. On the complexity of nash equilibria in anonymous
games. In STOC, 2015.
[CST14]
X. Chen, R. A. Servedio, and L.Y. Tan. New algorithms and lower bounds for monotonicity testing. In FOCS, pages 286–295, 2014.
[DDKT16] C. Daskalakis, A. De, G. Kamath, and C. Tzamos. A size-free CLT for poisson multinomials and its applications. In Proceedings of STOC’16, 2016.
[DDO+ 13] C. Daskalakis, I. Diakonikolas, R. O’Donnell, R.A. Servedio, and L. Tan. Learning Sums
of Independent Integer Random Variables. In FOCS, pages 217–226, 2013.
[DDS12]
C. Daskalakis, I. Diakonikolas, and R.A. Servedio. Learning Poisson Binomial Distributions. In STOC, pages 709–728, 2012.
[De15]
A. De. Beyond the central limit theorem: asymptotic expansions and pseudorandomness
for combinatorial sums. In FOCS, 2015.
[DKS15a] I. Diakonikolas, D. M. Kane, and A. Stewart. Optimal learning via the fourier transform
for sums of independent integer random variables. CoRR, abs/1505.00662, 2015. To
appear in COLT 2016.
[DKS15b] I. Diakonikolas, D. M. Kane, and A. Stewart. Properly learning poisson binomial distributions in almost polynomial time. CoRR, 2015. To appear in COLT 2016.
[DKT15]
C. Daskalakis, G. Kamath, and C. Tzamos. On the structure, covering, and learning of
poisson multinomial distributions. In FOCS, 2015.
[DP07]
C. Daskalakis and C. H. Papadimitriou. Computing equilibria in anonymous games. In
FOCS, pages 83–93, 2007.
64
[DP08]
C. Daskalakis and C. H. Papadimitriou. Discretized multinomial distributions and nash
equilibria in anonymous games. In FOCS, pages 25–34, 2008.
[DP09]
C. Daskalakis and C. Papadimitriou. On Oblivious PTAS’s for Nash Equilibrium. In
STOC, pages 75–84, 2009.
[DP14]
C. Daskalakis and C. H. Papadimitriou. Approximate nash equilibria in anonymous
games. Journal of Economic Theory, 2014.
[GKM15] P. Gopalan, D. M. Kane, and R. Meka. Pseudorandomness via the discrete fourier
transform. In FOCS, 2015.
[GMRZ11] P. Gopalan, R. Meka, O. Reingold, and D. Zuckerman. Pseudorandom generators for
combinatorial shapes. In STOC, pages 253–262, 2011.
[GRW15]
P. Gorlach, C Riener, and T. Weisser. Deciding positivity of multisymmetric polynomials. Journal of Symbolic Computation, 2015. Also available as arxiv report
http://arxiv.org/abs/1409.2707.
[GT14]
P. W. Goldberg and S. Turchetta. Query complexity of approximate equilibria in anonymous games. CoRR, abs/1412.6455, 2014.
[HJ85]
R. A. Horn and C. R. Johnson. Matrix Analysis. Cambridge University Press, 1985.
[Loh92]
W. Loh. Stein’s method and multinomial approximation. Ann. Appl. Probab., 2(3):536–
554, 08 1992.
[Mil96]
I. Milchtaich. Congestion games with player-specific payoff functions. Games and Economic Behavior, 13(1):111 – 124, 1996.
[PC99]
V. Y. Pan and Z. Q. Chen. The complexity of the matrix eigenproblem. In Proceedings
of the Thirty-first Annual ACM Symposium on Theory of Computing, pages 507–516,
1999.
[Poi37]
S.D. Poisson. Recherches sur la Probabilitè des jugements en matié criminelle et en
matiére civile. Bachelier, Paris, 1837.
[Roo99]
B. Roos. On the rate of multivariate poisson convergence. Journal of Multivariate
Analysis, 69(1):120 – 134, 1999.
[Roo02]
B. Roos. Multinomial and krawtchouk approximations to the generalized multinomial
distribution. Theory of Probability & Its Applications, 46(1):103–117, 2002.
[Roo10]
B. Roos. Closeness of convolutions of probability measures. Bernoulli, 16(1):23–50,
2010.
[Sto96]
A. Storjohann. Near optimal algorithms for computing smith normal forms of integer
matrices. In Proceedings of the 1996 international symposium on Symbolic and algebraic
computation, pages 267–274, 1996.
[Sto00]
A. Storjohann. Algorithms for matrix canonical forms. PhD thesis, Diss., Technische
Wissenschaften ETH Zürich, Nr. 13922, 2001, 2000.
65
[Val08]
P. Valiant. Testing symmetric properties of distributions. In STOC, pages 383–392,
2008.
[VV10]
G. Valiant and P. Valiant. A CLT and tight lower bounds for estimating entropy.
Electronic Colloquium on Computational Complexity (ECCC), 17(179), 2010.
[VV11]
G. Valiant and P. Valiant. Estimating the unseen: an n/ log(n)-sample estimator for
entropy and support size, shown optimal via new CLTs. In STOC, pages 685–694, 2011.
Appendix
A
Proof of Lemma 3.4
Lemma 3.4 follows directly from the following statement:
b be the sample
Lemma A.1. If we take O(k4 /ǫ2 ) samples from an (n, k)-PMD and let µ
b and Σ
k
mean and sample covariance matrix, then with probability 19/20, for any y ∈ R , we have:
q
T
|y (b
µ − µ)| ≤ ǫ y T (Σ + I)y ,
and
b − Σ)y| ≤ ǫy T (Σ + I)y .
|y T (Σ
The above lemma and its proof follow from a minor modification of an analogous lemma in
[DKT15]. We include the proof here for the sake of completeness. We will use the following simple
lemma:
Lemma A.2 (Lemma 21 from [DKT15]). For any vector y ∈ Rk , given sample access to an
(n, k)-PMD P with mean µ and covariance matrix Σ, there exists an algorithm
p which can produce
b−
b such that with probability at least 19/20: |y T (b
y T Σy and |y T (Σ
estimates µ
b and
Σ,
µ
−
µ)|
≤
ǫ
q
Σ)y| ≤ ǫy T Σy
1+
yT y
.
y T Σy
The sample and time complexity are O(1/ǫ2 ).
Proof of Lemma A.1. The proof will follow by applying Lemma A.2 to k2 carefully chosen vectors
simultaneously using the union bound. Using the resulting guarantees, we show that the same
estimates hold for any direction, at a cost of rescaling ǫ by a factor of k. Let S be the set of k2
vectors {vi }, for 1 ≤ i ≤ k, and { √λ1+1 vi + √1 vj }, for each i 6= j, where the vi ’s are an orthonormal
i
λj
eigenbasis for Σ with eigenvalues λi . From Lemma A.2 and a union bound, with probability 9/10,
for all y ∈ S, we have
p
|y T (b
µ − µ)| ≤ (ǫ/k) y T Σy ,
and
s
T
b − Σ)y| ≤ (ǫ/3k)(y T Σy) 1 + y y .
|y T (Σ
T
y Σy
We claim that the latter implies that:
b − Σ)y| ≤ (ǫ/3k)(y T (Σ + I)y) .
|y T (Σ
b = 0, since then y T X is a constant for a PMD random
Note that if y T Σy = 0, we must have y T Σy
variable X. Otherwise,
s
q
q
yT y
= (y T Σy)(y T Σy + y y y) = (y T Σy)(y T (Σ + I)y) ≤ (y T (Σ + I)y) .
(y T Σy) 1 + T
y Σy
66
b now follows from Lemma A.2.
The claim about the accuracy of Σ
We now prove that the mean estimator µ
b is accurate. Consider P
an arbitrary vector y, which
can be decomposed into a linear composition of the eigenvectors y = i αi vi .
Then,
s X
X
X
p
√
T
T
y (b
µ − µ) =
αi vi (b
µ − µ) ≤ (ǫ/k)
|αi | λi + 1 ≤ (ǫ/k) k k
α2i (λi + 1) ,
i
but
P
2
i αi (λi
i
i
+ 1) = y T (Σ + I)y, so we have y T (b
µ − µ) ≤ ǫ
p
y T (Σ + I)y as required.
We are now ready to complete the proof of the desired lemma.
Lemma 3.4. With probability 19/20, we have that (b
µ − µ)T (Σ + I)−1 (b
µ − µ) = O(1), 2(Σ + I) ≥
b
Σ + I ≥ (Σ + I)/2.
b − Σ)y| ≤ ǫy T (Σ + I)y, that is
Proof. We apply Lemma A.1 with ǫ := 1/2. For all y, we have |y T (Σ
1 T
b + I)y ≤ 3 y T (Σ + I)y .
y (Σ + I)y ≤ y T (Σ
2
2
b + I ≤ 3 (Σ + I) as required.
Thus, we have 12 (Σ + I) ≤ Σ
2
Note that since Σ + I is positive definite, it is non-singular. Setting y =
1
(Σ +
(b
µ−µ)T (Σ+I)−1 (b
µ−µ)
I)−1 (b
µ − µ), we have y T µ
b − µ) = 1 and y T (Σ + I)y = 1/(b
µ − µ)T (Σ + I)−1 (b
µ − µ). So, Lemma A.1
gives us:
q
1
|y T (b
µ − µ)| ≤
y T (Σ + I)y.
2
Substituting the above:
q
1
1/(b
µ − µ)T (Σ + I)−1 (b
µ − µ) .
1≤
2
Therefore, we have (b
µ − µ)T (Σ + I)−1 (b
µ − µ) ≤ 1/4, as required.
67
| 8cs.DS
|
CONCURRENCY AND COMPUTATION: PRACTICE AND EXPERIENCE
Concurrency Computat.: Pract. Exper. 2012; 00:1–19
Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/cpe
Accelerating R-based Analytics on the Cloud†
Ishan Patel, Andrew Rau-Chaplin and Blesson Varghese∗
arXiv:1308.2787v1 [cs.DC] 13 Aug 2013
Risk Analytics Laboratory, Faculty of Computer Science
Dalhousie University, Halifax, Nova Scotia, Canada
E-mail: {patel, arc, varghese}@cs.dal.ca
SUMMARY
This paper addresses how the benefits of cloud-based infrastructure can be harnessed for analytical
workloads. Often the software handling analytical workloads is not developed by a professional programmer,
but on an ad hoc basis by Analysts in high-level programming environments such as R or Matlab. The goal
of this research is to allow Analysts to take an analytical job that executes on their personal workstations, and
with minimum effort execute it on cloud infrastructure and manage both the resources and the data required
by the job. If this can be facilitated gracefully, then the Analyst benefits from on-demand resources, low
maintenance cost and scalability of computing resources, all of which are offered by the cloud. In this paper,
a Platform for Parallel R-based Analytics on the Cloud (P2RAC) that is placed between an Analyst and a
cloud infrastructure is proposed and implemented. P2RAC offers a set of command-line tools for managing
the resources, such as instances and clusters, the data and the execution of the software on the Amazon
Elastic Computing Cloud infrastructure. Experimental studies are pursued using two parallel problems and
the results obtained confirm the feasibility of employing P2RAC for solving large-scale analytical problems
on the cloud.
Copyright c 2012 John Wiley & Sons, Ltd.
Received . . .
KEY WORDS: cloud computing; data analytics; R script; catastrophe bonds
1. INTRODUCTION
Cloud based infrastructure has proven to be very effective in providing on-demand computational
resources to both commercial applications and a wide range of large-scale scientific applications.
Applications in climate simulation and analysis [1], biomedical image processing [2], satellite
data processing [3], astronomy [4], and disaster response systems [5] have all been successfully
supported using cloud infrastructure. Such successes in both the commercial and scientific settings
have depended on bringing to bear the talents of computer scientists and expert developers in order
to efficiently exploit cloud-based infrastructure.
In this paper, we explore how a different class of users with a different kind of workload might
be able to take advantage of the cloud. In particular, we study how Analysts, who are domain
experts with quantitative/mathematical skills, but often with software skills limited to high-level
programming environments like R [6], Matlab [7] or Octave [7], might be supported in harnessing
the cloud for ad hoc analytical workloads.
† This
research was fully financially supported by the Natural Sciences and Engineering Research Council of Canada
(NSERC) and Flagstone Re, Halifax, Canada under the Collaborative Research and Development grant CRDPJ 41288911, and was supported in part by an Amazon Web Service Education Grant Program.
∗ Correspondence to: [email protected]
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls [Version: 2010/05/13 v3.00]
2
Analytical workloads abound in application domains ranging from computational finance and
risk analytics to engineering and manufacturing settings. In our experience, these workloads which
often involve simulation [8] and optimisation [9] tasks share common features as follows:
(a) The associated codes are developed by Analysts, not professional developers, in high-level
programming environments such as R or Matlab.
(b) These codes and the related input data are generally created by Analysts for either one time
use or are heavily modified each time they are used to adapt them to the analytical question at
hand.
(c) The codes are often computationally intensive or require a large number of independent runs
with varying input parameters making some form of parallelism attractive.
Cloud computing is a potential solution that can be beneficial not only to meet the computational
requirements of analytical workloads [10] but also to achieve speed-up [11]. A wide range of
analytical workloads are already harnessing the benefits of cloud computing. For example, analytical
workloads related to data processing [12], online games [13], climate [14], medical records [15], risk
[16], social networks [17] and neuroscience [18].
Our goal has been to develop a platform that allows Analysts to take an analytical job (both
the code and associated data) that runs on their personal workstations and with minimum effort
and minimum change to the code have them run on large-scale parallel cloud infrastructure. If this
can be facilitated gracefully, the Analyst can solve larger problems or perform more experiments
in less time. Our approach is somewhat different from other ‘cluster on cloud’ projects such as
[19][20][21][22] in that our focus is to simplify an Analyst’s use of cloud infrastructure, rather
than provide a fully-configurable high-performance computing cluster on clouds for developers. In
particular, we have explored a platform for facilitating R-based risk analytics on the Amazon EC2
cloud. However, we believe that the basic platform and the experienced gained can be generalised
to a wider class of analytics and cloud-based infrastructure.
The Platform for Parallel R-based Analytics on Cloud infrastructure (P2RAC) has been designed
to harness on-demand compute and storage resources available on the cloud for analytical
workloads, and at the same time simplify an Analyst’s use of the cloud infrastructure. It provides
an interface/API for Analysts to (i) set-up both individual machines and clusters of machines in
the cloud, (ii) associate with machines both persistent and short term storage, (iii) transfer both
code and data, (iv) execute both batch and interactive jobs, and (v) manage execution and resource
termination. The goal is to allow Analysts to submit jobs to the cloud with minimum effort, and
manage both the resources and the data required to solve a problem and execute it on the cloud. The
current implementation of P2RAC provides support for the resources available on the Amazon AWS
infrastructure and offers a set of both core and diagnostic tools, and could also be implemented on
other cloud platforms.
In our interaction with Analysts working in industry we have observed a number of commonly
arising work patterns. In each of these Analysts start by prototyping an analytical code on their
personal workstations, and then as they progress in their work may require additional computing
resources. Additional resources are required (i) to perform production runs on a single workstation
with more memory or compute speed, (ii) to perform a parameter sweep in which the same code is
run hundreds or thousands of times with different input parameters, or (iii) to speed-up a single long
running optimisation or simulation task by exploiting co-operative parallelism that may be built
into the optimisation or simulation library being used. P2RAC has been designed to address each of
these cases. In the later two cases it is important to balance easy-of-use with parallel speed-up. The
experimental section of this paper explores this trade-off and shows that reasonable speed-up can
often be obtained on cloud infrastructure without undue complexity.
The remainder of this paper is organised as follows. Section 2 presents the software structure,
the sequence of steps that need to be initiated to execute an analytical task on the cloud, and two
example workflows for using P2RAC when compute resources, such as instances and clusters are
employed. Section 3 presents the implementation of P2RAC and the set of commands offered by the
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
3
Figure 1. Conceptual design of P2RAC
platform. Section 4 describes two analytical problems that are executed on the cloud using P2RAC
and the results obtained from the experiments. Section 5 concludes the paper.
2. P2RAC: PLATFORM AND USAGE
In this section, we discuss the software structure of Platform for Parallel R-based Analytics on
Cloud Infrastructure (P2RAC), and its usage. Figure 1 illustrates how the platform fits in coherently
between an Analyst, who is the typical user, and the cloud infrastructure. The Analyst’s project
comprising R scripts and large data files is passed on to P2RAC as a task for execution on the cloud.
The platform then gathers and initialises a pool of computational and storage resources on the cloud
for executing the task. The platform subsequently transfers the task onto the resources and manages
task execution. After the task is executed, the platform gathers results which may be spread across
the resources in the cloud.
The platform provides support both for an individual computing resource as well as for clusters
(a collection of computing resources that work collaboratively). The platform comprises three
components, (a) the core tools, (b) the diagnostic tools and (c) the configuration files. The core tools
provide functionalities for resource management, data management and execution management.
The diagnostic tools provide functionalities for checking the computational environment. The
configuration files provide support for the functionalities of the core and diagnostic tools.
A sequence of five steps need to be initiated on an Analyst site to execute a job on the cloud. In
the first step, computing and storage resources are initialised on the cloud. In the second step, the
analytical project is sent to the resources. In the third step, the scripts within the project are executed
on the resources. In the fourth step, results which are generated on the resources are gathered onto
the analyst site. In the fifth step, all resources initialised on the cloud are released.
2.1. Example Workflows
The illustration of the above five steps using the command-line tools of P2RAC is shown as two
workflows, the first for executing a task on an instance and the second for executing a task on a
cluster. For running an R-script on an instance, the sequence of commands are shown in Figure 2. An
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
4
Figure 2. Workflow of commands using P2RAC. * besides a command indicates that the command can be
executed multiple times.
Analyst in possession of R scripts utilising SNOW library firstly creates an instance (corresponds to
the first step) on the Amazon Cloud. The scripts and the data required by the script are then provided
to the instance (corresponds to the second step). The script is executed on the instance and produces
the results on the instance (corresponds to the third step). The results can then be fetched by the
Analyst (corresponds to the fourth step). Multiple tasks can be executed and their results generated
and retrieved from the same instance. Finally, the instance is terminated (corresponds to the fifth
step).
Figure 3 shows the sequence of the commands and the order in which they are executed at the
Analyst site for executing a task on the cluster. A cluster with 4 instances (1 master and 3 worker
instances) is created (corresponds to the first step). The Analyst secondly sends the R scripts and
the data required by the scripts to the cluster (corresponds to the second step). The script is then
executed and the result of execution is generated on the master instance (corresponds to the third
step). When the execution of the R scripts are completed then the Analyst gathers the results to his
site (corresponds to the fourth step). Multiple tasks can be executed on the same cluster, and the
results generated can be retrieved from the cluster. The cluster is finally terminated (corresponds to
the fifth step).
3. SYSTEM DESIGN AND IMPLEMENTATION
This section considers firstly, how Amazon Web Service (AWS) is supported on P2RAC, and
secondly, the tools offered by P2RAC. The tools of P2RAC are separated out as (i) the core tools,
(ii) the diagnostic tools and (iii) the configuration files supporting both instances and clusters. This
section considers the implementation of the tools that support the cluster and instance. The core and
diagnostic tools are implemented as commands which are executed from the command-line.
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
5
Figure 3. Workflow of commands using P2RAC. * besides a command indicates that the command can be
executed multiple times.
3.1. AWS Support
The resources provided by a cloud are accessible through Infrastructure as a Service (IaaS) [25],
and Amazon is a leading provider of IaaS. The computational resources on Amazon are referred
to as Elastic Compute Cloud (EC2) and are available as instances. These resources are available
on-demand and are paid for on the basis of their usage.
The Amazon instances are initialized using Amazon Machine Images (AMI) [28]. Two Ubuntu
AMIs are used in this research. The first AMI, supports Cluster Compute instances offered by
Amazon. The Cluster Compute instances are built on Hardware Virtual Machine (HVM) type
virtualisation. The second Ubuntu AMI offers support for non-HVM type virtualisation. Any
additional libraries which an Analyst needs to include in his project can be installed on the instance
by specifying library packages in a configuration file.
The storage resources on Amazon are referred to as the Elastic Block Storage (EBS) [27]. Similar
to EC2, EBS is available on-demand and are paid for on the basis of data transfer and volume of
storage. One note worthy feature of EBS includes its ability to provide persistent data storage. This
feature is exploited when large volumes of data need to be used in an Analyst’s project and thereby
eliminates the need for frequent transfer of data that may not be changed over time. Another feature
of EBS is that it can be attached (mounted) as a local storage onto an EC2 instance in addition to its
storage. This feature eliminates the need for making data locally available to the instance.
The EC2 and EBS services are employed by P2RAC. The P2RAC platform mediates between
an Analyst’s project and the web services offered by Amazon. The platform is built with the
Python programming language and draws heavily on two libraries. Firstly, the BOTO library, which
provides P2RAC a Python interface to the Amazon web services. Secondly, the Fabric library, which
facilitates remote administration of the instances on the cloud for P2RAC.
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
6
3.2. Core Tools
The core tools of P2RAC provide functionalities for resource management, data management on the
resource, and execution management of a task on the resource. An interface is provided to an Analyst
to access the cloud which allows significant computational resources to be brought to bear while
greatly reducing the complexities associated with directly working with the cloud infrastructure.
This is necessary since an Analyst is less likely to have knowledge and experience of working with
computational clouds.
3.2.1. Instance Support is offered using tools for instance management, data management on the
instance and execution management of a task on the instance.
Instance Management tools configure an instance on the cloud, offers the instance for
the execution of a task and finally terminates the instance after the task it was executing is
completed. P2RAC offers two tools for managing instances, namely ec2createinstance and
ec2terminateinstance.
The ec2createinstance command is responsible for configuring an instance on the cloud
and making it available to an analyst.
The syntax of the ec2createinstance command is
ec2createinstance [-h] [-v] [-iname INSTANCE NAME]
[-ebsvol EBS VOLUME | -snap EBS SNAP] [-type INSTANCE TYPE]
[-desc INSTANCE DESCRIPTION]
The optional arguments of the ec2createinstance command are iname, ebsvol, snap,
type and desc. iname specifies the name of the instance that is created. ebsvol specifies the
EBS volume ID which is provided by Amazon when an EBS volume is created. snap specifies
the EBS snapshot ID from which an EBS volume can be created. snap and ebsvol cannot be
specified at the same time. ebsvol can be specified when a EBS volume is available, while if snap
is specified then a new EBS volume is created from the snapshot specified. If both arguments are
not provided, then a default snapshot from a configuration file is used. type defines the Amazon
EC2 instance type which is specified based on the computational requirements of the task. For
example, a High-memory Quadruple Extra Large Instance, offers 68.4 GB of memory, 26 EC2
compute units, 1690 GB of instance storage, with high input/output performance, with instance
type as m2.4xlarge, and was employed for a number of experiments in the work reported in this
paper. The desc argument can be used to provide a description of the instance.
For example, if a command such as
ec2createinstance -iname ’hpc instance’ -ebsvol
’vol-xxxxxxxx’ -type ’m2.4xlarge’ -desc ’For Trial
Simulation Run’
is executed, then a sequence of activities follow. One EC2 instance of m2.4xlarge type is
initialised using the AMI specified in the configuration file, and is tagged. The EBS volume,
vol-xxxxxxxx (volume ID is masked in this paper) is attached on to the instance. A configuration
file at the Analyst site is updated with instance information such as the public DNS names of the
instance, EBS volume ID and description of the instance. Should the optional arguments be not
provided then the default values which are defined in a configuration file is chosen.
The multiple execution of the ec2createinstance command facilitates the creation of
multiple instances. Since an EBS volume can only be attached to one instance, and therefore, the
need for multiple EBS volumes arises when multiple instances are created. Should multiple EBS
volumes require the same data then they need to snapshot from the same source located on Simple
Storage Service (S3) offered by Amazon [30]. Multiple instances cannot have the same name when
the ec2createinstance command is executed more than once.
When a task has completed execution on the instance it is essential to safely release
the instance. The ec2terminateinstance command is provided. The syntax of the
ec2terminateinstance is
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
7
ec2terminateinstance [-h] [-v] [-iname INSTANCE NAME]
[-deletevol]
The optional arguments of the ec2terminateinstance command are iname and
deletevol. iname specifies the name of the instance that needs to be terminated. The
deletevol switch deletes an EBS volume attached to the instance being terminated. To terminate
all the instances on the cloud ec2terminateall command is provided which will be considered
in cluster management.
For example, if a command such as
ec2terminateinstance iname ’hpc instance’
is executed, then a sequence of activities follows. Firstly, the EBS volume attached to the instance
is no more available. Further to this, the instance is terminated. The section containing the instance
information of hpc instance in the configuration file is removed. Should the deletevol
switch be included in the command then the EBS volume, vol-xxxxxxxx is deleted.
Data Management on the instance is required to transfer the script and the data from the Analyst’s
site onto the instance and then receive results. The Secure Copy Protocol (SCP) or the rsync protocol
can be employed for data transfer. However, rsync, is chosen as it transfers data quicker than SCP.
Moreover, rsync in subsequent data transfers only synchronises the data changed at the source. This
feature of rsync makes it suitable for P2RAC since data at the Analyst’s site changes frequently. The
ec2senddatatoinstance command is provided for data transfer and the results are retrieved
using the ec2getresultsfrominstance command.
The syntax of the ec2senddatatoinstance command is
ec2senddatatoinstance [-h] [-v] [-iname INSTANCE NAME]
[-projectdir PROJECT DIRECTORY]
The optional arguments of the ec2senddatatoinstance command are iname and
projectdir. iname specifies the name of the instance to which the project directory will be
synchronised. The source project directory is specified as the argument projectdir. If the
instance name is not provided by the Analyst then the instance from the configuration file is
employed. Should the project directory not be specified then the current working directory at the
Analyst site is used as the source project directory. The destination directory is not specified since
the project directory is synchronised at the home directory of the root user.
The project directory comprises a set of R scripts which need to be executed, a set of data files
required by the scripts and a sub-directory that will contain results after the execution of the script.
Large volumes of data which are less likely to change in a short course of time are stored on the EBS
volume. On the other hand smaller chunks of data that frequently change are synchronised from an
Analyst’s site on to the local storage of the instance.
The syntax of the ec2getresultsfrominstance is
ec2getresultsfrominstance [-h] [-v] [-iname INSTANCE NAME]
[-projectdir PROJECT DIRECTORY] [-runname RUN NAME]
The optional arguments are iname and projectdir. The iname argument specifies the name
of the instance from where the results have to be fetched. The projectdir specifies the location
of the source project directory at the Analyst site. The command utilises the name of the project
from the path of the source project directory to fetch data from the corresponding project directory
on the instance. If no project directory is specified then the path of the current working directory at
the Analyst site is used.
The mandatory argument for the ec2getresultsfrominstance command is runname
which indicates the name of a run that was specified during execution and whose results need to be
gathered. This argument is used if the same R script has been executed a number of times and each
execution had to be differentiated.
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
8
Execution Management command, namely the ec2runoninstance runs the R script on the
instance. This command locks the instance onto the R script and does not permit any additional use
of the instance until the script has completed execution or the instance is manually unlocked using
ec2resourcelock considered later. The syntax of ec2runoninstance is
ec2runoninstance [-h] [-v] [-iname INSTANCE NAME]
[-projectdir PROJECT DIRECTORY] [-rscript R SCRIPT]
[-runname RUN NAME]
The optional arguments of ec2runoninstance are iname, projectdir and rscript.
The iname argument specifies the name of the instance on which the R script needs to be executed.
The projectdir specifies the location of the source project directory at the Analyst site. The
command utilises the name of the project from the path of the source project directory to execute
an R script from the corresponding project directory on the instance. rscript specifies the name
of the R script to be executed from projectdir. If rscript is not provided then the user is
prompted to select from a list of R scripts that may be available in the project directory.
The mandatory argument for ec2runoninstance is runname which indicates the name of
a run.
3.2.2. Cluster Support is offered using tools for cluster management, data management on the
cluster and execution management of a task on the cluster.
Cluster Management in the core tools are a set of functionalities that range from gathering a pool
of instances on the cloud, followed by configuring the instances as a cluster, offering the cluster
for task execution and finally terminating the cluster when the task executing on the cluster has
completed. P2RAC offers two core tools for cluster management, namely ec2createcluster
and ec2terminatecluster.
The ec2createcluster tool is responsible for gathering and configuring the pool of
instances as a cluster on the cloud.
The syntax of the ec2createcluster command is
ec2createcluster [-h] [-v] [-cname CLUSTER NAME] [-csize
CLUSTER SIZE] [-ebsvol EBS VOLUME | -snap EBS SNAP] [-type
INSTANCE TYPE] [-desc CLUSTER DESCRIPTION]
The optional arguments of the ec2createcluster command are cname, csize, ebsvol,
snap, type and desc. cname specifies the name of the cluster that is created. csize specifies
the size of the cluster. ebsvol specifies the EBS volume ID which is provided by Amazon when
an EBS volume is created. snap specifies the EBS snapshot ID from which an EBS volume can
be created. snap and ebsvol cannot be specified at the same time. ebsvol can be specified
when a EBS volume is available, while if snap is specified then a new EBS volume is created
from the snapshot specified. If both arguments are not provided, then a default snapshot from a
configuration file is used. type defines the Amazon EC2 instance type which is specified based on
the computational requirements of the task. The desc argument can be used to provide a description
of the cluster.
For example, if a command such as
ec2createcluster -cname ’hpc cluster’ -csize ’10’ -ebsvol
’vol-xxxxxxxx’ -type ’m2.4xlarge’ -desc ’For Trial
Simulation Run’
is executed, then a sequence of activities follow. Ten EC2 instances of m2.4xlarge type are
initialized using the AMI specified in the configuration file. The cluster of the ten EC2 instances
is referred to as hpc cluster. One instance in the hpc cluster is denoted as the master and
tagged as hpc cluster Master, while the remaining nine instances are denoted as workers
and are tagged as hpc cluster Workers. The EBS volume, vol-xxxxxxxx (volume ID is
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
9
masked in this paper) is attached on to the master instance. Network File System (NFS) is employed
to share the attached EBS volume among the nine worker instances. A configuration file at the
Analyst site is updated with cluster information such as the public DNS names of the master and
worker instances, size of the cluster, EBS volume ID, description of the cluster and whether the
cluster is in use for executing a script. Should the optional arguments be not provided then the
default values which are defined in a configuration file is chosen.
The multiple execution of the ec2createcluster command facilitates the creation of
multiple clusters. Since an EBS volume can only be attached to the master instance of one cluster
the need for multiple EBS volumes arises when multiple clusters are created. Should multiple EBS
volumes require the same data then they need to snapshot from the same source located on Simple
Storage Service (S3) offered by Amazon [30]. Multiple clusters cannot have the same name when
the ec2createcluster command is executed more than once.
When a task has completed execution it is essential to safely release the resources which are
utilised by the cluster. To this end, the ec2terminate cluster command is provided. The syntax
of the ec2terminatecluster command is
ec2terminatecluster [-h] [-v] [-cname CLUSTER NAME]
[-deletevol]
The optional arguments of the ec2terminatecluster command are cname and
deletevol. cname specifies the name of the cluster that needs to be terminated. The
deletevol switch deletes an EBS volume attached to the cluster being terminated.
For example, if a command such as
ec2terminatecluster cname ’hpc cluster’
is executed, then whether a cluster is in use is firstly checked. If the cluster is in use, then the
cluster cannot be terminated. If the cluster is not in use, then a sequence of activities follows.
The EBS volume that has been shared with the worker instances through NFS is no more
shared. Further to this, the worker instances are terminated such that they do not exist. The EBS
volume vol-xxxxxxxx is detached from the master node and the master instance is terminated.
The section containing the cluster information of hpc cluster in the configuration file is
removed. Should the deletevol switch be included in the command then the EBS volume,
vol-xxxxxxxx is deleted.
When all resources, both of the instance and the cluster need to be terminated the
ec2terminateall command is provided. The syntax of ec2terminateall is
ec2terminateall [-h] [-v] [-instances] [-clusters]
[-ebsvolumes] [-snapshots]
The optional switches are instances, clusters, ebsvolumes and snapshots which
terminates all the instances, clusters, EBS volumes and snapshots respectively.
Data management on the cluster is required to transfer the task (both the scripts and data) from
the Analyst site to the cluster on the cloud, and thereafter receive results to the Analyst site. The
transfer of task may be to the entire pool of resources in the cluster or to a specific instance on
the cluster. Two feasible routes are to use the Secure Copy (SCP) protocol or the rsync protocol.
The rsync protocol is employed owing to the quicker synchronisation of data between a source and
a destination site. Therefore, two commands, namely the ec2senddatatoclusternodes and
ec2senddatatomaster based on the rsync protocol are provided. In order to receive the results
to the Analyst site the ec2getresults command is developed.
The ec2senddatatoclusternodes command enables an Analyst’s project to be
synchronised with all instances of a cluster. This stands different to the data that is stored in an
EBS volume mounted on the master instance and shared with the worker instances. In this research,
large volumes of data which are less likely to change in a short course of time are stored on the EBS
volume. On the other hand smaller chunks of data that frequently change are synchronised from an
Analyst’s site on to the local storage of the cluster instances.
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
10
The structure of a project at the Analyst’s site is worthwhile to be noted. A directory
comprising a set of R scripts which need to be executed, a set of data files required by
the scripts and a sub-directory that will contain results after the execution of the script. The
ec2senddatatoclusternodes command synchronises the directory from the Analyst’s site
to all the instances of the cluster. In other words, every instance contains a project directory.
The syntax of the ec2senddatatoclusternodes command is
ec2senddatatoclusternodes [-h] [-v] [-cname CLUSTER NAME]
[-projectdir PROJECT DIRECTORY]
The optional arguments of the ec2senddatatoclusternodes command are cname and
projectdir. The cname argument specifies the name of the cluster whose instances will be
synchronised with the project directory. The source project directory is specified as the argument
projectdir. If the cluster name is not provided by the Analyst then the cluster name from the
configuration file is employed. Should the project directory not be specified then the current working
directory at the Analyst site is used as the source project directory. The destination directory is not
specified since the project directory is synchronised at the home directory of the root user.
Owing to the nature of the task to be executed, it may not be necessary that the project directory be
provided to all the instances of a cluster. For example, consider a task in which the master instance
receives data from the Analyst site and distributes it to the worker instances in the cluster. In such
a case it would be inefficient to synchronise the source project directory with all the instances of
the cluster, but would be sufficient for the master instance alone to have the project directory. To
facilitate this, ec2senddatatomaster command is provided.
The syntax of the ec2senddatatomaster is
ec2senddatatomaster [-h] [-v] [cname CLUSTER NAME]
[-projectdir PROJECT DIRECTORY]
The optional arguments of ec2senddatatomaster are similar to that of
ec2senddatatoclusternodes.
Based on the nature of the R scripts that are executed there are three possible scenarios for
generating results. It is assumed that the R scripts generate results in a sub-directory within the
project directory. In the first scenario, the master instance aggregates the results from the worker
instances and stores them at the master instance. In the second scenario, however, the results are
only generated on the worker instances. In the third scenario, the results are generated on both
the master and worker instances. In both the scenarios, the results need to be obtained at the
Analyst site. Therefore, the ec2getresults command is provided. To address the first scenario,
ec2getresults gathers results from the master instance and provides it at the Analyst site. To
address the second scenario, ec2getresults gathers results from the worker instances. In the
third scenario, ec2getresults gathers results from both the master and all the worker instances.
The aggregated results are stored in a directory at the same hierarchical level of the project directory
at the Analyst site.
The syntax of the ec2getresults is
ec2getresults [-h] [-v] [cname CLUSTER NAME] [-projectdir
PROJECT DIRECTORY] [-runname RUN NAME] -frommaster |
-fromworkers | -fromall
The optional arguments are cname, projectdir and a switch. The cname argument specifies
the name of the cluster from where the results have to be fetched. The projectdir specifies the
location of the source project directory at the Analyst site. The command utilises the name of the
project from the path of the source project directory to fetch data from the corresponding project
directory on the cluster. If no project directory is specified then the path of the current working
directory at the Analyst site is used. The switch specifies the instances from where the results need
to be gathered. If frommaster is specified, then results are gathered as in the first scenario. If
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
11
fromworkers is specified, then results are gathered as in the second scenario. If fromall is
specified, then results are gathered as in the third scenario. If no switch is specified then the results
are gathered as in the first scenario.
The mandatory argument for the ec2getresults command is runname which indicates the
name of a run that was specified during execution and whose results need to be gathered. This
argument is used if the same R script has been executed a number of times and each execution had
to be differentiated.
Execution Management assigns the Analyst task onto a cluster and further runs task on the
cluster. This command locks the cluster for execution of the R script specified by the execution
command and does not permit any additional use of the cluster until the script has completed
execution or the cluster is manually unlocked using ec2resourcelock considered in the next
section. For this, the ec2runoncluster command is provided. The syntax of ec2runscript
is
ec2runoncluster [-h] [-v] [-cname CLUSTER NAME]
[-projectdir PROJECT DIRECTORY] [-rscript R SCRIPT]
[-runname RUN NAME] [-bynode | -byslot]
The optional arguments of ec2runoncluster are cname, projectdir and rscript.
The cname argument specifies the name of the cluster where the R script needs to be executed.
The projectdir specifies the location of the source project directory at the Analyst site. The
command utilises the name of the project from the path of the source project directory to execute an
R script from the corresponding project directory on the cluster. rscript specifies the name of the
R script to be executed from projectdir. If rscript is not provided then the user is prompted
to select from a list of R scripts that may be available in the project directory.
Two optional switches to manage the scheduling of slave processes onto the cores of the cluster
nodes are available. The bynode switch assigns the processes in a round-robin fashion while the
byslot switch assigns all processes on a node until all of its cores are exhausted. In MPI, the
default scheduling is byslot, whereas in P2RAC, bynode is chosen as the default scheduling
mechanism if the switch is not specified. bynode switch is required to meet the memory constraints
of large processes.
The mandatory argument for ec2runoncluster is runname which indicates the name of a
run.
3.3. Diagnostic Tools
The diagnostic tools support the access of instances and clusters and are available as follows:
(a) ec2listinstances, ec2listclusters and ec2listallresources - for listing
the instances and clusters created by the Analyst on the Amazon cloud
(b) ec2logintoinstance and ec2logintomaster - for accessing an instance or the
master instance of a cluster
(c) ec2resoucelock - to lock an instance or a cluster for a specific task or to unlock them
from use
The syntax of ec2listinstance is
ec2listinstance [-h] [-v] [-names]
The optional switch names provides the names of the instance. If the switch is not provided then
the list of the instances, their public DNS names, volume ID of the EBS volume shared with the
instances and the description of the cluster is provided.
The syntax of ec2listclusters is
ec2listclusters [-h] [-v] [-names]
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
12
The optional switch names provides the names of the clusters on the cloud. If the switch is not
provided then the list of the clusters along with the size of the cluster, public DNS name of all
instances, volume ID of the EBS volume shared with the instances of the cluster and the description
of the cluster.
The syntax of ec2listallresources is
ec2listallresources [-h] [-v] [-instances] [-ebsvols]
[-snapshots] [-amis]
The switches instances, ebsvols, snapshots and amis provides the names of the
instances, EBS volumes, EBS snapshots and AMIs on the cloud.
The syntax of ec2logintoinstance is
ec2logintoinstance [-h] [-v] [-iname INSTANCE NAME]
The optional argument iname specifies the name of the instance that needs to be accessed. The
connection to the instance is facilitated through Secure Shell (SSH). If the name of the instance is
not provided then the instance listed in the configuration file is used.
The syntax of ec2logintocluster is
ec2logintocluster [-h] [-v] [cname CLUSTER NAME]
The optional argument cname specifies the name of the cluster whose master instance needs to
be accessed. The connection to the master instance is also facilitated through Secure Shell (SSH).
If the name of the cluster is not provided then the master instance of the default cluster listed in the
configuration file is used.
The syntax of ec2resourcelock is
ec2resourcelock [-h] [-v] [-iname INSTANCE NAME | cname
CLUSTER NAME] [-free | -inuse]
The optional arguments iname specifies the name of the instance and cname the name of the
cluster that needs to be locked or unlocked. The resources are locked using the -inuse switch and
unlocked using the -free switch.
3.4. Configuration Files
There are four files that support the core and diagnostic tools which reside on the Analyst site.
Firstly, a file that contains a list of variables that are required by the command line tools along with
a number of directory paths and references to access keys for Amazon resources. Secondly, a file
that provide support for instances, and includes the name of the instance created, its public DNS
name, Volume ID, the description of the instance and whether the instance is in use. Thirdly, a file
that provides support for clusters, and includes the names of the clusters created, their size, the
public DNS names of all their instances, Volume ID of the EBS volume shared by the master with
the workers, the description of the clusters and whether the cluster is in use. Fourthly, a file that
contains a list of R libraries which are required by an Analyst’s project. These libraries are installed
on the instances of the cluster when it is created. This is required in addition to the pre-installed
libraries of the base AMI.
The P2RAC platform offers support both for instances and clusters. In the case of instance
support, P2RAC enables the management of instances, which includes the creation and termination
of single and multiple instances. In the case of cluster support, P2RAC facilitates creation and
termination of single and multiple clusters. In either case, management of data is facilitated by
making large and small chunks of data available to the executing task by sharing persistent data
volumes and by synchronising data from the Analyst site to the instances or cluster.
The platform is designed for both batch mode execution and interactive mode execution. Batchmode execution in P2RAC supports time consuming production tasks. The core commands are
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
13
Table I. Resources Utilised for Experimental Studies
Resource
Provided by
Desktop A
Dalhousie
University
Desktop B
Flagstone Re
Instance A
Instance B
Cluster A
Cluster B
Cluster C
Cluster D
Amazon
Amazon
Amazon
Amazon
Amazon
Amazon
Processor
/
Amazon Type
Intel (R) Core
(TM) i7-2600 @
3.4 GHz
Intel (R) Xeon
X5660 @ 2.8
GHz
m2.2xlarge
m2.4xlarge
m2.2xlarge X 2
m2.2xlarge X 4
m2.2xlarge X 8
m2.2xlarge X 16
No. of
cores
8
Memory
Storage
System
type
64 bit
16GB
1.8 TB
6
24GB
2 TB
64 bit
4
8
8
16
32
64
34.2 GB
68.4 GB
68.4 GB
136.8 GB
273.6 GB
547.2 GB
850 GB
1690 GB
1.7 TB
3.4 TB
6.8 TB
13.6 TB
64 bit
64 bit
64 bit
64 bit
64 bit
64 bit
listed in a script and the script is executed without the intervention of an Analyst. The interactive
mode execution on the other hand allows an Analyst to experiment with his scripts and supports
execution of ad hoc tasks. The core commands are executed from the command line by the Analyst.
All commands in P2RAC utilise two switches, one of which is -h that provides a description of the
use and arguments of the command, and the other which is -v that provides the version of P2RAC.
4. EXPERIMENTAL STUDIES
This section presents the platform, the two problems employed on P2RAC and the results obtained
from experiments. The experimental studies is performed for a twofold reason, firstly, to evaluate
the feasibility of P2RAC and qualitatively assess its usage and secondly, with respect to parallelism
evaluate the speed up that can be expected for both the easy case of independent parallelism and the
more challenging case of cooperative parallelism.
Table I presents the name of the resource and its location, the processor type, the number of
cores for the processor, memory, storage capacity and operating system time specifications of the
resources. Two desktop computers, two Amazon cloud instances, and four clusters comprising
Amazon cloud instances are the computational resources utilised for the evaluation of P2RAC. The
Amazon cloud instances are high-memory instances, namely the high-memory double extra large
instance (m2.2xlarge) and the high-memory quadruple extra large instance (m2.4xlarge). The four
clusters have 2, 4, 8 and 16 nodes (each cluster has one master node) and 8, 16, 32 and 64 available
cores (each node has 4 cores) respectively for executing a task. Each m2.2xlarge instance is charged
$0.9 per hour and each m2.4xlarge instance is charged $1.8 per hour.
To evaluate the feasibility of P2RAC, two kinds of experimental problems which are analytical in
nature were employed. The first problem is a computationally intensive task employing co-operative
parallelism, referred to as Catastrophe Bond Optimisation (CATopt). The problem belongs to the
domain of reinsurance analytics and is typical of the optimisation problems found in this domain in
that (a) it is a large-scale highly non-linear optimisation problem in several thousands of dimensions,
(b) it is likely to be executed a few times a year by a sponsoring reinsurance company, and (c) it
may require large-scale executions and analysis over the course of several weeks before an actuary
can sign-off the results.
Catastrophe bonds (also known as Cat Bonds) [31] are risk-linked securities that transfer a
specified set of risks, typically associated with catastrophic loss of property, from a sponsoring
insurer or reinsurer to sophisticated investors. If no catastrophe occurs, the sponsor pays a coupon
or interest payments to the investors, who made a profitable return. If a catastrophe occurs that meets
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
14
Figure 4. Speed-up achieved for the CATopt and Parameter Sweep Problems using P2RAC
the conditions described when the bond is issued (referred to as trigger), then the principal would
be forgiven and the sponsoring insurance company would use this money to pay claim-holders.
When a cat bond is issued with a parametric trigger, then the investors are made available
with (a) regions and perils exposed and the sponsor’s share in those region-perils, and (b) the
probability of attachment and expected loss. Consider there are m region-peril combinations, for
example, Alabama Residential or Florida Commercial. Then there are m market shares or weights
corresponding to the region-perils denoted wj , where j = 1, 2, . . . , m. The Industry Losses from an
event i can be denoted as IL(i,j) , where j = 1, 2, . . . , m, and P
the sponsor’s loss from that event is sli .
m
The loss based on which the sponsor will get his payment is j=1 wj IL(i,j) and the recovery value
Pm
is Recoveryi = M in(M ax( j=1 wj IL(i,j) − Att, 0), Limit), where Att is an attachment point
which is a deductible and Limit is the maximum payout defined P
contractually. So the sponsor faces
m
‘basis risk’ since the actual loss cli could be very different from j=1 wj IL(i,j) , and consequently
receive more/less recovery than required.
In the experimental studies, an R-based example of the CATopt problem which seeks to identify
a set of weights that minimizes basis risk is employed. The dimension of the optimisation space is
typically 2000-4000 region-perils combinations and there are a number of non-linear constraints that
must be applied thereby making the CATopt problem a challenging and computationally intensive
problem. The CATopt R script is structured as a distributed genetic algorithm using the rgenoud
[29] R package which combines evolutionary search algorithms with derivative-based (Newton or
quasi-Newton) methods. The input data to the CATopt problem is approximately 300 MB.
The second problem is a parameter sweep task [32, 33] that runs multiple independent jobs
without any data dependency between the runs. An R-based example of a simple Monte Carlo
simulation was employed. The input data to the parameter sweep task is approximately only 3 MB.
Figure 4 is a graph showing the relative speed-up achieved for both the experimental problems
with increasing number of Amazon instances. Figure 5 is a bar graph that shows the timing for the
best results on the two desktops, the two Amazon instances and the four Amazon clusters. The best
performance is achieved on Cluster D.
For the CATopt problem in these experiment, the population size is set to 200 and the maximum
generations is set to 50. In both the problems there is a near 100% efficiency for up to 4 Amazon
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
15
Figure 5. Best-case timing results of the CATopt and Parameter Sweep Problems using P2RAC
instances, after which there is a drop in the parallel efficiency. The reduction in parallel efficiency is
due to increase in communication overheads between virtualised cloud instances. The acceleration
achieved for both the problems is satisfactory considering the low cost of the infrastructure
employed and that the problems can be directly deployed from an Analyst’s site without any
additional tuning.
Figure 6 and Figure 7 are bar graphs for the CATopt and parameter sweep problems showing (a)
the time taken to create the Amazon resource, (b) the time taken to submit the project to an instance
or to the master node of a cluster, (c) the time taken to submit the project to all the nodes of a cluster,
(d) the time taken for fetching results from an instance or from the master node of a cluster, (e) the
time taken for fetching results from all the nodes of a cluster, and (f) the time taken for terminating
the resources.
There is an increase in the time taken for creating Amazon instances and clusters. Though the
creation of resources occur in parallel, nearly 7 minutes are required to initialise a 8 node m2.2xlarge
cluster and approximately 8 minutes are required to initialise a 16 node m2.2xlarge cluster. This
indicates that the time taken for configuring large-scale clusters will also increase. Alternative
techniques will need to be investigated to reduce this time and incorporated within the underlying
interface that manages AWS. The time taken for terminating the instances and clusters remain the
same.
The time taken to synchronise the project directory comprising the R script and the data (for the
first experiment 300 MB and for the second 3 MB) on an instance or on the master node of the
cluster and the time taken for fetching results from an instance or from the master node of a cluster
remains the same on all resources considered in the experiment. However, there is an increase in
both the times for submitting a job and for retrieving results when all the nodes of the cluster are
considered. Though the submission of the job and the retrieval of results are parallel in nature there
is an increase in time, eliciting the need for investigating other parallel methods that can reduce this
time. For large jobs like CATopt that need to run for many hours the additional minutes do not seem
significant, however, it may not be worthwhile to spend a lot of time for creating and moving data
around resources for small jobs.
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
16
Figure 6. Time taken for creating resource, submitting project, fetching results and terminating resources for
the CATopt problem on Amazon instances and clusters
Figure 7. Time taken for creating resource, submitting project, fetching results and terminating resources for
the parameter sweep problem on Amazon instances and clusters
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
17
5. DISCUSSION AND CONCLUSIONS
There are a large number of cluster on cloud frameworks supporting a variety of applications.
Frameworks such as ElasticR [19], CycleCloud [20] and Pegasus [37] provide a large dashboard for
configuring a high-performance cluster on the cloud. These frameworks are useful for developers
who have advanced knowledge of the technicalities of the cloud, and can adapt their analytical
workloads for such frameworks. Ad hoc analytics will not be easy on such tools as they are mostly
performed by Analysts who have limited technical skills, and therefore their prototyping requires
a simpler framework. Frameworks such as CloudBLAST [38] and CloudBurst [39] support easy
workflows without having to specifically adapt analytical workloads for the cloud, but are aimed at
computational biological applications and thereby limit their use for Analysts. Frameworks such
as Cloudfoundry [21] and Starcluster [22] do not provide explicit support for R programming
language and therefore a lot of manual procedures are required to run an Analyst’s workload. Such
frameworks do not provide seamless execution of a task on the cloud.
In this paper, a platform that (i) provides an Analyst with an interface to seamlessly submit an
analytical job and collect its results, and (ii) provides the flexibility to manage the resources and jobs
is ideal for an Analyst who needs to exploit the benefits of the cloud. Analytical jobs can benefit from
harnessing the benefits of cloud computing such as on-demand availability of resources, scalability
of the resources and low costs for maintenance. The Platform for Parallel R-based Analytics on
the Cloud (P2RAC) proposed and implemented in this paper provides an interface for an Analyst
who needs to submit a R-based analytical job on the cloud. P2RAC is therefore placed in between
an Analyst and the cloud infrastructure. P2RAC provides support for instances and clusters on the
Amazon cloud and offers a set of core and diagnostic tools which can be used from the commandline.
The core tools support resource management, data management on the resource and execution
management. Resource management functionalities range from gathering resources required for
an analytical job and releasing them after their use. Data management ranges from providing the
resources with data required for executing an R-script and gathering results from the resources
onto the Analyst’s site once the R-script has completed execution. Execution management provides
functionalities for executing an R-script on the resources acquired.
The diagnostic tools support the testing of resources and provide information of the resources
employed by an Analyst. This is facilitated by providing access to log in to the acquired resources.
The feasibility of P2RAC is validated by considering two analytical problems. The algorithm
for solving the first problem incorporates co-operative parallelism for optimisation, while in the
second problem, multiple independent jobs are employed for parameter sweep. Both the problems
are provided from an Analyst work site to the Amazon cloud using P2RAC. The results obtained
from P2RAC is an evidence that R-based analytical problems can benefit from cloud computing and
P2RAC facilitates the execution of the analyst job on the cloud.
The installation package and the source code for P2RAC is hosted via the Python Package
Index (PyPI) http://pypi.python.org/pypi. The installation is built for the Linux OS, supports Virtual
Python Environment [34] and can be installed on a desktop using Easy Install [35]. After installation,
P2RAC can be configured automatically using ec2configurep2rac.
In the future, it is anticipated that another version of P2RAC will be released that incorporates
the support for spot instances. Fault tolerance through distributed checkpointing for spot instance
used in P2RAC will be explored. Further, the dynamic scaling of clusters, i.e., increase and
decrease the size of a cluster when required by a job that is executing on the cluster, will be
investigated for implementation. While packages such as those reported in [24, 36] support explicit
parallelism, efforts will be made to provide a dedicated interface that can exploit Amazon EC2 for
R programming.
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
18
ACKNOWLEDGEMENT
The authors would like to thank Dr. Georg Hoffman and Dr. Oliver Baltzer of Flagstone Re, Halifax,
Canada for their support and participation in this research.
REFERENCES
1. Evangelinos C, Lermusiax P.F.J., Xu J., Haley Jr P.J., and Hill C.N. 2011. Many Task Computing for Real-Time
Uncertainty Prediction and Data Assimilation in the Ocean. IEEE Transactions on Parallel and Distributed Systems,
Volume 22, No. 6, 2011, pp. 1012-1024.
2. Zhang C., De Sterck H., Aboulnaga A., Djambazian H. and Sladek R. 2010. Case Study of Scientific Data
Processing on a Cloud Using Hadoop. Proceedings of the 23rd International Conference on High Performance
Computing Systems and Applications, pp. 400-415.
3. Li J., Humphrey M., Cheah Y.-W., Ryu Y., Agarwal D., Jackson K. and Van Ingen C. 2010. Fault Tolerance
and Scaling in e-Science Cloud Applications: Observations from the Continuing Development of MODISAzure.
Proceedings of the 24th IEEE International Parallel and Distributed Processing Symposium, Atlanta, USA, pp.
246-253.
4. Hoffa C., Mehta G., Freeman T., Deelman E., Keahey K., Berriman B. and Good J. 2008. On the Use of Cloud
Computing for Scientific Workflows. Proceedings of the 4th IEEE International Conference on eScience, USA, pp.
640-645.
5. Kelly S., Mazyck C., Pfeiffer K. and Shing M.-T. 2011. A Cloud Computing Application for Synchronized Disaster
Response Operations. Proceedings of the IEEE World Congress on Services, pp. 612-616.
6. Venables W.N., Smith D.M. and the R Development Core Team. 2012. An Introduction to R. Notes on R: A
Programming Environment for Data Analysis and Graphics, Version 2.14.2.
7. Quarteroni A. and Saleri F. 2006. Scientific Computing with MATLAB and Octave. Springer, 2nd Edition.
8. Grossi P. and Kunreuther H. 2005. Catastrophe Modeling: A New Approach to Managing Risk. Springer, 1st Edition.
9. Perez M.J. 2007. Multi-Objective Optimization Evolutionary Algorithms in Insurance-Linked Derivatives.
Handbook of Research on Nature Inspired Computing for Economics and Management, Edited by Rennard J.-P.,
IGI Global, pp. 885-908.
10. Iosup A., Ostermann S., Yigitbasi M.N., Prodan R., Fahringer T. and Epema D.H.J. 2011. Performance Analysis
of Cloud Computing Services for Many-Tasks Scientific Computing. IEEE Transactions on Parallel and Distributed
Systems, Vol. 22, No. 6.
11. Habich D., Lehner W., Richly S. and Assmann U. 2010. Using Cloud Technologies to Optimize Data-Intensive
Service Applications. Proceedings of the 3rd IEEE International Conference on Cloud Computing, pp. 19-26.
12. Barga R.S., Ekanayake J. and Lu W. 2012. Project Daytona: Data Analytics as a Cloud Service. Proceedings of the
28th IEEE International Conference on Data Engineering, pp. 1317-1320.
13. Nae V., Iosup A. and Prodan, R. 2011. Dynamic Resource Provisioning in Massively Multiplayer Online Games.
IEEE Transactions on Parallel and Distributed Systems, Vol. 22, No. 3.
14. Lu S., Li R.M., Tjhi W.C., Lee K.K., Wang L., Li X. and Ma D. 2011. A Framework for Cloud-Based LargeScale Data Analytics and Visualization: Case Study on Multiscale Climate Data. Proceedings of the 3rd IEEE
International Conference on Cloud Computing Technology and Science, pp. 618-622.
15. Tancer J. and Varde A.S. 2011. The Deployment of MML for Data Analytics over the Cloud. Proceedings of the
11th IEEE international Conference on Data Mining Workshops, pp. 188-195.
16. Kim H., Chaudhari S., Parashar M. and Marty C. 2009. Online Risk Analytics on the Cloud. Proceedings of the 9th
IEEE/ACM International Symposium on Cluster Computing and Grid, pp. 484-489.
17. Xue W., Shi J. and Yang B. 2010. X-RIME: Cloud-Based Large Scale Social Network Analysis. Proceedings of the
IEEE International Conference on Services Computing, pp. 506-513.
18. Watson P., Hiden H. and Woodman S. 2010. e-Science Central for CARMEN: Science as a Service. Concurrency
and Computation: Practice and Experience, Vol. 22, pp. 2369-2380.
19. Chine K. 2009. Scientific Computing Environments in the Age of Virtualization, Toward a Universal Platform for
the Cloud. Proceedings of the IEEE International Workshop on Opensource Software for Scientific Computation,
pp. 44-48.
20. CycleCloud website: http://cyclecomputing.com/cyclecloud/overview [Last checked: 10th August 2012]
21. Cloud Foundry Website: http://www.cloudfoundry.com/ [Last checked: 10th August 2012]
22. StarCluster Website: http://web.mit.edu/star/cluster/ [Last checked: 10th August 2012]
23. Pacheco P. 2011. An Introduction to Parallel Programming (1st Edition). Morgan Kaufmann.
24. Simple Network of Workstations (SNOW) website: http://www.sfu.ca/∼sblay/R/snow.html [Last checked: 8th
August 2012]
25. Leavitt N. 2009. Is Cloud Computing Really Ready for Prime Time? IEEE Computer, Vol. 42, Issue 1, pp. 15-20.
26. Amazon Elastic Compute Cloud (EC2) website: http://aws.amazon.com/ec2/ [Last checked: 10th August 2012]
27. Amazon Elastic Block Store (EBS) website: http://aws.amazon.com/ebs/ [Last checked: 10th August 2012]
28. Amazon Machine Images (AMI) website: http://aws.amazon.com/amis [Last checked: 10th August 2012]
29. Mebane, Jr. W. R. and Sekhon J. S. 2011. Genetic Optimization Using Derivatives: The rgenoud Package for R.
Journal of Statistical Software, Vol. 42, Issue 1.
30. Amazon Simple Storage Service (S3) website: http://aws.amazon.com/s3/ [Last checked: 10th August 2012]
31. Cummins J.D. 2008. Cat Bonds and Other Risk-Linked Securities: State of the Market and Recent Developments.
Risk Management and Insurance Review, Vol. 11, No. 1, pp. 23-47.
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
19
32. Bliss N.T. and Kepner J. 2007. pMATLAB Parallel MATLAB Library. International Journal of High Performance
Computing Applications. Vol. 21, Issue 3, pp. 336-359.
33. Casanova H., Zagorodnov D., Berman F. and Legrand A. 2000. Heuristics for Scheduling Parameter Sweep
Applications in Grid Environments. Proceedings of the 9th Workshop on Heterogeneous Computing.
34. Virtual Python Environment website: http://pypi.python.org/pypi/virtualenv/ [Last checked: 10th August 2012]
35. Python Easy Install website: http://packages.python.org/distribute/easy install.html [Last checked: 10th August
2012]
36. Knaus J, Porzelius C, Binder H and Schwarzer G. 2009. Easier Parallel Computing in R with snowfall and sfCluster.
The R Journal, Vol. 1, Issue 1, pp. 54-59.
37. Lee K, Paton N.W., Sakellariou R., Deelman E., Fernandes A.A.A., Mehta G. 2009. Adaptive Workflow Processing
and Execution in Pegasus. Concurrency and Computation: Practice and Experience, Volume 21, Issue 16, 2009, pp.
1965-1981.
38. Matsunaga A., Tsugawa M. and Fortes J. 2008. CloudBLAST: Combining MapReduce and Virtualization on
Distributed Resources for Bioinformatics Applications. Proceedings of the 4th IEEE International Conference on
eScience, pp. 222-229.
39. Schatz M.C. 2009. CloudBurst: Highly Sensitive Read Mapping with MapReduce. Bioinformatics, Vol. 25, Issue
11, pp. 1363-1369.
Copyright c 2012 John Wiley & Sons, Ltd.
Prepared using cpeauth.cls
Concurrency Computat.: Pract. Exper. (2012)
DOI: 10.1002/cpe
| 5cs.CE
|
arXiv:0908.3280v2 [cs.IR] 29 Sep 2009
On the Relationship between Trading Network
and WWW Network: A Preferential Attachment
Perspective
Andri Mirzal
Graduate School of Information Science and Technology,
Hokkaido University, Kita 14 Nishi 9, Kita-Ku,
Sapporo 060-0814, Japan
December 28, 2017
Abstract
This paper describes the relationship between trading network and
WWW network from preferential attachment mechanism perspective. This
mechanism is known to be the underlying principle in the network evolution and has been incorporated to formulate two famous web pages
ranking algorithms, PageRank and HITS. We point out the differences
between trading network and WWW network in this mechanism, derive
the formulation of HITS-based ranking algorithm for trading network as
a direct consequence of the differences, and apply the same framework
when deriving the formulation back to the HITS formulation that turns
to become a technique to accelerate its convergences.
1
Introduction
The researches on the analysis of preferential attachment and network structure
can be dated back to 50’s with the work of Solomonoff and Rapoport (1951),
where the authors presented the first systematic study of a class of networks
known as random graphs. Actually the study of graph itself has a long history
in mathematics as Euler introduced the using of vertices and edges to model
the famous Königsberg bridge problem in 1736. However, different from the
classical studies, the modern network studies have some interesting additional
features: (1) focusing on much larger problems that can contain million vertices
so it is natural to consider statistical properties of the networks, (2) dealing
with real networks like Internet topology (Faloutsos et al., 1999), WWW network (Albert et al., 1999; Broder et al., 2000), metabolic networks (Jeong et al.,
2000; Wagner and Fell, 2001), scientific collaboration networks (Price, 1965;
Newman, 2001; Barabási et al., 2002; Jeong et al., 2003), and epidemic spreading networks (Ball et al., 1997; Keeling, M.J., 1999; Kuperman and Abramson,
2001; Pastor-Satorras and Vespignani, 2001) among others, and (3) studying dynamical properties of the networks as many real networks are not static entities,
but grow according to some rules (Bianconi and Barabási, 2001; Barabási et al.,
2002; Jeong et al., 2003).
1
The foundation of modern random graph theory which focuses on structure and statistical properties of very large random graphs was set by Erdős
and Renyı́ (Erdős and Renyı́, 1959, 1960, 1961). The random graph is a very
influential model in modeling the real networks because it can describe many
phenomena including phase transition, short paths between most of vertex pairs,
and the existence of a giant component. Prior to the finding of scale-free network
(Barabási and Albert, 1999), many network designs were based on the random
graph model, including Internet data protocols design (Newman et al., 2006)
and social network experiments setting (Travers and Milgram, 1969; Pool and Kochen,
1978).
The first widely known challenge to the random graph came from the study of
WWW network topology by Albert, Jeong, and Barabási1 (Albert et al., 1999;
Barabási and Albert, 1999). While the first paper’s main focus is in the short
path between any pair of pages that supports the finding of Watts and Strogatz
(1998), the second paper is the first to explicitly challenge the effectiveness of
the random graph in modeling the real networks by showing experimentally the
ubiquitous existence of power-law degree distributions (pk ∝ k −γ , where pk is
the probability of vertices with degree k, and γ is the exponent that usually in
the range 2-3 in the real networks) in a variety of real networks including WWW
network, scientific collaboration networks, and film actors networks that led to
the famous scale-free hypothesis.
The power-law degree distributions found in many real networks are considered to be the most important remark that shows the discrepancy between
random graph prediction of the degree distribution (which should follow Poisson
distribution, pk ∝ mk e−m /k!, where m is the mean degree of vertices) and the
real situation. This discrepancy outclasses other good predictions made by the
random graph including short path, diameter of the graph, phase transition,
and size of the giant component, and generates a very large number of scientific
publications on such networks vary from mathematics, physics, computer science, economics to sociology. And in turn creates a new field of study: complex
networks.
There are several fundamental reasons behind the curiosity in the ubiquity
of the scale-free phenomenon. We enlist some of them here:
1. Different from the Poisson distribution, mean value and standard deviation of the power-law distribution doesn’t imply centrality and data dispersion. As we know, these metrics can be very useful in describing a
distribution without having to plot it. But in the case of the power-law,
these metrices can be misleading because the mean value doesn’t reflect
the centrality and the standard deviation doesn’t tell us about the range
where most of the data lies.
2. There is no peak value in the power-law, pk decreases monotonically as k
increases.
1 We must note here actually Price (Price, 1965) is the first person to show that real
networks, scientific collaboration networks, are following power-law degree distribution, and
his finding was long before the works of Barabási et al. (Barabási et al., 2002; Jeong et al.,
2003). Interestingly, Price didn’t seem to know about the famous random graph model of
Erdős and Renyı́, and also Barabási et al. didn’t know about the previous work of Price when
studying scientific collaboration networks.
2
λ = 2.1
λ = 2.3
λ = 2.5
λ = 2.7
λ = 2.9
m=1
m=3
m=5
m=7
m=9
0
2
4
6
8
10
1
(a) Poisson distribution
10
100
(b) Power-law distribution
Figure 1: Poisson and power-law distribution plots for several m and λ respectively. Note that x-axis is the degree k, y-axis is the probability pk , and (b) is
in log-log scale.
3. The power-law has a large tail that decays much slower than the Poisson
distribution, thus there are some vertices with very high degree. These
vertices are the hubs and have role to keep the integrity and robustness
of the network.
Of the three reasons above, the existence of hubs is considered to be the
most important and surprising finding because:
1. Prior to the works of Barabási and Albert (1999), it was very natural
to think that in general real datasets have Poisson distribution family
(including binomial, normal, and Gaussian distribution).
2. The existence of very large hubs implies that virtually there is no limit for
vertices to create and receive new edges (this is the reason for scale-free
term picked by Barabási and Albert).
3. There should be some fundamental principles that govern the evolution
of the scale-free networks. The principles are described as growth and
preferential attachment (Barabási and Albert, 1999), where the probability of receiving a new edge is proportional to the number of edges a vertex
already has.
In this paper, we study the preferential attachment mechanism in trading
networks. By using the supply and demand principle, we show that the preferential attachment in trading networks is opposite to the corresponding mechanism in WWW network. Because the preferential attachment is the principle
behind the formulation of link structure ranking algorithms like PageRank and
HITS (see section 3 for details), we will use the differences to define a ranking
algorithm for trading networks. The proposed algorithm will be HITS-based
because there are two type of transactions to be captured, sellings and buyings.
In network term, sellings are equivalent to creating new outlinks and buyings
are equivalent to receiving new inlinks (see section 4 for details about resource
flows). So out of these algorithms, only HITS which produces two type of scores,
authority scores that correspond with inlinks (buyings) and hub scores that correspond with outlinks (sellings), can be extended to trading networks. And by
3
using the same framework when deriving the proposed algorithm for trading
networks, we present a new approach to accelerate HITS computation. The
preliminary results can be found in (Mirzal, 2009a,b)
2
Preferential attachment
The preferential attachment is a concept introduced by Yule in 1925 (Yule, 1925)
and then is used to describe a class of mechanisms in which the probability of
receiving a quantity is proportional to the number of that quantity the object
already has. It has appeared in several fields under different names; in information science it is known as the cumulative advantage (Price, 1976), in sociology
as the Matthew effect (Merton, 1968), and in economics as the Gibrat principle
(Simon, 1955). The preferential attachment is long known to be the principle
behind the power-law distribution exhibited by some real datasets, for example
the distributions of wealth accumulation (Reed, 2001), the distribution of the
number of species per genus (Yule, 1925), the distributions of word frequencies
used in books and documents (Zipf, 1935; Konchady, 2006), and the number of
collaborators in scientific collaboration networks (Price, 1965; Newman, 2001;
Barabási et al., 2002; Jeong et al., 2003) among others.
However, the presence of the preferential attachment in the network evolution doesn’t always produce the power-law degree distribution. If there are some
constraints in generating new edges, usually the degree distribution will not be
following the power-law because there are not many vertices with very high
number of degree. But usually it will not be following the Poisson distribution
either. Instead it will have non power-law but still right-skewed degree distribution. For example, power grid and air traffic have exponential distributions,
friendship networks have Gaussian distributions, and movie actors network has
an exponentially truncated power-law distribution (Amaral et al., 2000).
Barabási et al. (2002) provide a robust test to detect the preferential attachment in the network evolution by observing the change of degree ∆k as
a function of k (∆k ∝ k v ) for every vertex over some time intervals (thus
requiring dynamic data which is not always available). If the preferential attachment exists, v will be bigger than 0. For perfectly scale-free network, v will
be equal to 1. And if the mechanism doesn’t exist, v will be equal to 0 (note
that actually Barabási et al. (2002) use integral of the probability of receiving
Rk
new edges, κ(k) = 1 Π(k ′ )dk ′ where Π(k ′ ) ∝ k v to define the test). Different
from usual simple test by only plotting pk , this method can distinguish networks
with preferential attachment from merely random graphs with power-law degree
distributions (Newman et al., 2001). However, on many occasions it is usually
sufficient to utilize pk distribution, and actually this is the most common approach used by the researchers to detect the presence of preferential attachment
in the real networks (Faloutsos et al., 1999; Broder et al., 2000; Jeong et al.,
2000; Wagner and Fell, 2001; Newman, 2001; Pastor-Satorras and Vespignani,
2001; Liljeros et al., 2001; Barabási and Albert, 1999; Price, 1965; Kleinberg,
1999; Albert and Barabási, 2000; Price, 1976; Merton, 1968; Zipf, 1935).
4
3
PageRank and HITS
Around the same time with the finding of the preferential attachment in the
network evolution, two groups of researchers started to realize the role of link
structure of WWW network in determining the values of the pages. Links
in WWW network are the hyperlinks created by the site owners to point to
other relevant pages, favorite pages, popular pages, or pages that contain useful
information (these were especially true in the beginning of WWW era where
most hyperlinks were created by human and link spammers were rare). So, the
hyperlinks reflect the opinion about the values of the pages; the more valuable
the pages, the more inlinks they have. Thus, the hyperlink structure can be
utilized to distinguish important pages from less important ones.
The first link structure ranking algorithm was proposed by Brin and Page
(Brin et al., 1999; Brin and Page, 1998), known as PageRank, a popularity measure based on hypothesis of a random surfer that is infinitely following the hyperlink structure of WWW network. In the long run, the proportion of time
a random surfer spends on a page depends on the number of inlinks the page
has and on the number of inlinks other pages that point to it have. This is
intuitive because the number of inlinks of a page reflects its reachability from
other pages. And because the proportion of time it spends on a page reflects the
value of the page, the PageRank score of a page is proportional to the number
of inlinks the page has and to the number of inlinks other pages that point to
it have. On the other hand, because the hyperlinks are the opinions or the
recommendations created by the site owners to other pages, the values of the
recommendations should be dropped if there are too many of them on a page.
Thus, the PageRank score of a page is inversely proportional to the number of
outlinks other pages that point to it have. Mathematically, PageRank is defined
with the following equation:
X prj
(1)
pri =
outdegj
j∈Bi
where pri denotes PageRank score of page i, outdegi denotes outdegree of i,
and Bi denotes set of pages that point to i.
The above equation is a circular statement: the score of a page depends on
the scores of other pages that point to it, and in turn the scores of those pages
depend on the scores of other pages that point to them. To solve it, usually
iterative procedure is employed with each page is given an initial value (usually
set to 1/N , where N is the number of pages).
(k)
(k+1)
pri
=
X
j∈Bi
prj
,
outdegj
k = 1, . . . , K
(2)
where K denotes the final iteration where the predefined criterion is satisfied.
To get a more compact form, Equation (2) is rewritten in the matrix form.
pr(k+1)T = pr(k)T Do−1 L
(3)
where Do = diag(outdeg1 , . . . , outdegN ), pr(k)T denotes the 1 × N PageRank
vector at iteration k, and L denotes the adjacency matrix induced from WWW
network where [L]ij = 1 if there is a hyperlink from i to j, and 0 otherwise.
5
Equation (3) is the problem of finding the dominant eigenvector of (Do−1 L)T
by using the power method (Barret et al., 1994). By Markov chains theory,
Equation (3) converges to a unique positive PageRank vector pr iff Do−1 L is
stochastic, irreducible, and aperiodic (Langville and Meyer, 2006).
A matrix is stochastic iff there is no zero row and all the rows are normalized.
So, the first adjustment is to modify Do−1 L into a stochastic matrix. Let d be
N × 1 dangling vector where its nth (n = 1, 2, . . . N ) entry is 1 if n is a dangling
page and 0 otherwise, and eT be all-one 1 × N vector. The stochastic version
of Do−1 L is S = Do−1 L + (1/N )deT .
A matrix is irreducible iff its directed graph is strongly connected; for every
pair of vertices, there is at least one path connecting them. And a matrix is
aperiodic iff there is only one principal eigenvalue on the spectral circle. The
irreducibility and aperiodicity properties can be enforced by replacing all zero
entries of S with small positive numbers. Thus, the stochastic, irreducible, and
aperiodic version of Do−1 L is P = α S + (1/N )(1 − α)eeT , where 0 < α < 1
denotes a scalar that controls proportion of time the random surfer follows the
hyperlinks as opposed to teleporting (usually set to 0.85). And Equation (3)
can be rewritten as:
pr(k+1)T = pr(k)T P
= αpr(k)T Do−1 L + (1/N ) αpr(k)T d + 1 − α eT
(4)
The second ranking algorithm, HITS (Hypertext Induced Topic Search) was
introduced by Kleinberg (Kleinberg, 1999). Different from PageRank, HITS
produces two metrics associated with every page, authority and hub. Authority
scores determine pages’ popularity and hub scores are used to find portal pages,
pages that link to popular (thus useful) pages.
HITS is defined with the following statement: authority score of a page is
the sum of hub scores of others that point to it and hub score of a page is the
sum of authority scores of others that are pointed to by it (Kleinberg, 1999).
Like PageRank, this is also a circular statement, the authority scores depend
on the hub scores and vice versa. To solve it, the following equation is used.
X (k+1)
X (k)
(k+1)
(k+1)
(5)
=
aj
=
hj , and hi
ai
j∈Fi
j∈Bi
(k)
(k)
where ai and hi denote the authority and hub score of page i at iteration k,
Bi denotes the set of pages that point to i, and Fi denotes the set of pages that
are pointed to by i. In the matrix from, HITS formulation can be rewritten as:
a(k+1)T = h(k)T L, and h(k+1)T = a(k+1)T LT
(6)
where aT denotes 1×N authority vector and hT denotes 1×N hub vector.
In HITS both authority matrix, LT L (a(k+1)T = a(k)T LT L) and hub matrix, LLT (h(k+1)T = h(k)T LLT ) are nonnegative. Thus by Perron theorem
for nonnegative matrices, aT and hT exist but there is no guarantee of the
uniqueness. To ensure the uniqueness, the authority and hub matrices must
be modified into positive matrices (Farahat et al., 2006; Langville and Meyer,
2006). Let  and Ĥ be the positive version of the authority matrix and hub
matrix respectively. We can define them as  = ζ LT L + (1/N )(1 − ζ) e eT ,
6
Figure 2: Labeled-link network model of the trading activities.
and Ĥ = ζ L LT + (1/N )(1 − ζ) e eT , where 0 < ζ < 1 denotes a constant
that should be set near to 1 to preserve the hyperlink structure information.
Thus, unique and positive authority and hub vectors can be calculated by using
a(k+1)T = a(k)T Â and h(k+1)T = h(k)T Ĥ.
4
Link structure ranking algorithm for trading
networks
The trading activities are the exchanges of different goods and/or services (we
will refer goods/services as resources for the rest of the paper) involving at least
two agents. These activities can be modeled with labeled-link network where
the vertices are the agents and the directed edges are the flows of the resources.
Figure 2 shows the network model of the trading activities. Note that actually
the transactions are mutual; there are two opposite flows for each transaction,
the flow of the resource and the flow of the payment. However because the
price is a better unit of account in the market and generally is used to measure
the quantity of the resources, each transaction can be described by only one
directed edge, the flow of the resource weighted with the price.
There are some differences between trading network and WWW network
that are worth to be noted. First, in trading network every vertex has at least
one type of resource and a new edge is created when two vertices exchange their
resources. Consequently, the amount of resources limits the number and the
weight of edges a vertex can have. On the other hand, in WWW network the
creation of edges is simply the creation of new hyperlinks on the web pages,
so there is no resource needs to be allocated. Second, different from trading
network, the creation of edges in WWW network is not a mutual process; if
page A has a hyperlink to page B, it doesn’t necessary that B also has a
hyperlink to A. Third, every edge in trading network is labeled with resource
description and is weighted with the price. On the other hand, the edges in
7
WWW network are usually unlabeled and weighted with either 1 or 02 . And
fourth, while the purpose of edges creation in trading network is to maximize
the transaction benefits, in WWW network is to get hyperlinks from popular
pages.
The last difference is directly related to the preferential attachment mechanisms. Before we define the preferential attachment in trading network, we will
enlist some assumptions that have to be taken in order to simplify the complex
interactions among agents.
1. All transactions are carried out under ceteris paribus condition. So the
prices depend only on demands and supplies of the corresponding resources, not other substitute or complementary resources.
2. The perfect market condition is met and the prices have already reached
the equilibrium states.
3. The amount of resources owned by an agent is reflected in its buying and
selling volumes of the corresponding resources.
The first assumption allows us to form and analyze one network for each
resource independently. The second assumption guarantees that resources availability is the main motivation in choosing business partners, not the price differences. And the third assumption allows us to estimate the resources availability
for future transactions by using current and past buying and selling volumes of
the corresponding resources, which is reflected in the weights of the inlinks and
outlinks.
Note that both first and second assumptions are very common in the trading
network analysis and the economics in general. So, we will only discuss the
reasons behind the last assumption. The last assumption is the heart of the
proposed algorithm formulation because it allows us to (1) model the trading
activities completely with the labeled-link network which is a standard model
in graph theory, (2) relate the amount of resources owned by an agent to the
weights of the corresponding inlinks and outlinks, and (3) define the preferential
attachment in trading network by using the number of inlinks and outlinks
(more specifically, total weights of those links) so that it can be compared to the
preferential attachment in WWW network induced from the HITS formulation
(see Figure 3), and in turn allowing us to formulate a ranking algorithm for
trading network.
4.1
Proposed algorithm formulation
In trading activities, there are costs associated with every transaction. Thus,
every agent must implement an optimal preferential attachment strategy to
maximize the benefits. In the real situation, every transaction conducted by
an agent influences its financial states, including transactions from different
resources. However, by assumption 1 we can isolate the influences and form
2 There are some works that are devoted to the analysis of WWW network labels (the
hypertexts). For example: Eiron and McCurley (2003), Kolda et al. (2005), and Fujii (2008).
But because the hyperlinks are the recommendations, they are alike, and in some cases can
be ignored safely, including the calculations of PageRank and HITS. Conversely, in trading
network the labels are the inherent information of the transactions that cannot be ignored at
any cost
8
(a) Trading network
(b) WWW network
Figure 3: The preferential attachment mechanisms in trading network and
WWW network (HITS’s version).
one labeled-link network for each different resource. So, if there are x type of
resources traded among the agents, there will be x labeled-link networks that
can be analyzed separately. Then by assumption 2, each agent should buy
(receive inlinks) from others with abundant resources, and should sell (create
outlinks) to others that are lack of the resources. And by assumption 3, agents
with abundant resources are the agents with many inlinks and agents that lack
of the resources are the agents with many outlinks.
Thus, we can define the preferential attachment in trading network with
the following statement: an agent should receive new inlinks from others with
many inlinks and should create new outlinks to others with many outlinks. This
statement is interesting because it resembles HITS’s version of preferential attachment in WWW network. As discussed in section 3, in HITS good authorities (pages with many inlinks) are pointed to by good hubs (pages with many
outlinks) and good hubs point to good authorities. Thus, HITS’s version of preferential attachment is: a page should receive new inlinks from others with many
outlinks and should create new outlinks to others with many inlinks. Figure 3(a)
and 3(b) show the preferential attachment in trading network and WWW network respectively. As we can see, the preferential attachment in trading network
is opposite to the HITS’s version of preferential attachment in WWW network,
so it can be utilized to formulate the proposed algorithm.
The proposed algorithm is defined with the following statement: a vertex
becomes more important if being pointed to by others with many inlinks and
points to others with many outlinks. This statement is derived directly from the
preferential attachment in trading network defined above. And by comparing
the preferential attachments in both networks (see Figure 3) and the HITS
9
formulation (see Equation 5), the proposed algorithm can be written as:
X (k)
X (k)
(k+1)
=β
ri
rj caj + (1 − β)
rj chj , where
j∈Bi
(7)
j∈Fi
indegi
|indegi − outdegi |pi ,
degi
outdegi
chi =
|indegi − outdegi |−pi , and
degi
1 if indegi > outdegi
pi =
−1 if indegi < outdegi
0 otherwise
cai =
(8)
(9)
(10)
(k)
where ri denotes ranking score of vertex i at iteration k; indegi , outdegi , and
degi denote indegree, outdegree, and degree of i; and 0 < β < 1 is a scalar used
to determine which link is more important. If outlink (selling) is more important
than inlink (buying), β < 0.5; if inlink is more important than outlink, β > 0.5;
and β = 0.5 otherwise.
The constants ca and ch are introduced to favour the preferential attachment. As shown in Equation 8 and 9, ca will be bigger for vertices with many
inlinks, and ch will be bigger for vertices with many outlinks. Thus, by Equation 7, vertices that are pointed to by others with many inlinks and point to
others with many outlinks (following the preferential attachment) will have bigger scores than vertices that do the opposite (not following the preferential
attachment).
P
(k)
Note that the first term of the right hand part of Equation 7, j∈Bi rj caj ,
describes the fraction of scores a vertex receives from its inlinks, and the second
P
(k)
term of the right hand part,
j∈Fi rj chj , describes the fraction of scores
a vertex receives from its outlinks. So, the first term can be defined as the
authority part and the second term as the hub part.
The proposed algorithm will be represented in matrix to allow necessary
adjustments be applied in order to ensure the convergence. Let M = βF + (1 −
β)G, where F = KD−1 Di L be the authority part, and G = K−1 D−1 Do LT
be the hub part. Then, Equation 7 can be rewritten as:
r(k+1)T = r(k)T M
= r(k)T βKD−1 Di L + (1 − β)K−1 D−1 Do LT
(11)
where L denotes the induced adjacency matrix, r(k)T denotes 1×N ranking vector at interation k, Di = diag(indeg1 , . . . , indegN ), Do = diag(outdeg1 , . . . , outdegN ),
D = Di + Do, and K is a diagonal matrix where [K]ii = |(Di − Do)ii |pi . Note
that different from WWW network, in trading network entries of L are the
weights of the corresponding links which are usually nonnegative real numbers.
As shown in Equation 11, M has no zero row, but is not a stochastic matrix
because the rows are not normalized. Therefore the stochasticity
adjustment is
P
required. Let N be a diagonal matrix where [N]ii = j∈V Mij (V denotes the
set of all vertices in the network), the stochastic version of M can be written
as M = N−1 M. And the irreducibility and aperiodicity adjustments can be
done by replacing all zero entries of M with small positive numbers: R =
10
ζ M + (1/N )(1 − ζ)eeT , where 0 < ζ < 1 is equivalent to α in PageRank and
should be set near to 1. Thus, the proposed algorithm can be rewritten as:
r(k+1)T = r(k)T R
(12)
As we can see, R is identical to P in Equation 4, and by choosing a positive
initial vector (for example r(k=1)T = (1/N )eT ) the Equation 12 is guaranteed
to converge to a unique positive ranking vector r(K)T (Farahat et al., 2006).
The proposed algorithm only accommodates the flowing resources. If we
have data about the amount of resources owned by the agents which is not
from the transactions, for example natural resources like gas, oil, coal, gold,
etc (we will refer these as reserved resources for the rest of the paper), this
information can also be included in the final scores. Let uT be 1 × N vector
where [u]i corresponds to the amount of the reserved resource of agent i. Then
the final ranking vector can be written as: r̂T = c rT + (1 − c)uT , where u is the
normalized version of u, and 0 < c < 1 is a control parameter that determines
which vector is more important.
We can also introduce a scaling constant similar to the work of Bianconi and Barabási
(2001) associated with every agent to the final score to describe its competitiveness. These constants can be used not only to favour the competitive agents,
but also to handle some issues related to the trading activities like reliability
and trust issues.
Occasionally, agents’ scores as the buyers and/or the sellers are more desirable than the overall scores. By inspecting Figure 3(a) and Equation (7),
ranking vector as the buyers, bT , and as the sellers, sT , can be written as:
b(k+1)T = b(k)T Ca L, and s(k+1)T = s(k)T Ch LT
(13)
where Ca = diag(ca1 , . . . , caN ), and Ch = diag(ch1 , . . . , chN ).
4.2
Experimental results
We will examine the proposed algorithm performance by using international
trading datasets from the United Nations (United Nations, 1996, 1999). There
are several good reasons in choosing these datasets. First, the size of the networks are small compared to other datasets like online auction networks, therefore the errors produced in each iteration can be minimized. Second, the classification of products is clear, so the adjacency matrix for every product can be
easily constructed. And third, the prices of the products in the same category
are almost the same, complying with the second assumption.
As stated earlier R is stochastic, irreducible, and aperiodic. Thus, the power
method applied to Equation (12) is guaranteed to converge to a unique positive
ranking vector r(K)T for any positive starting vector. Therefore, the question
left is “will it converge to something that makes sense in the context of measuring
the degree of importance of agents in trading network”. We will answer this
question by calculating the similarity between vector of our proposed algorithm
r, and standard measure, vector of total export and import t. This vector is
chosen as the standard measure not only because it is the simplest and common
way in measuring the degree of importance, but also because the most active
agents are usually the most connected ones which are conventionally considered
to be the most important vertices in the graph theory. And as the similarity
11
measures, cosine criterion cos θ and Spearman rank order correlation coefficient
ρ will be used.
PN
6 i=1 ([o(r)]i − [o(t)]i )2
r·t
, and ρ = 1 −
(14)
cos θ =
krk2 ktk2
N (N 2 − 1)
where k ∗ k2 denotes 2-norm of vector ∗, and o(∗) denotes the ordering induced
from vector ∗. For example, if ∗ = [0.3397, 0.1819, 0.3328], then o(∗) = [1, 3, 2].
Thus, while the cosine criterion measures the distance between two vectors, the
Spearman correlation measures the similarity between orderings induced from
the vectors.
To get insight about the computational performance, the number of iterations required by the proposed algorithm to achieve the same residual level will
be compared to the results of PageRank and HITS. In the experiments, the
residual level is set to 10−8 and β is set to 0.5. The number of iterations is
chosen instead of computational time because the sizes of trading networks are
very small, so the power method produces negligible computational time. Table
1 gives summary of the results, and Table 2 and 3 show lists of top ten countries in hydrogen peroxide trading (the least similar to the standard measure in
the cosine criterion) and medicinal products (the most similar to the standard
measure in the cosine criterion).
As shown in Table 1, the proposed algorithm takes more iteration steps to
converge. But because trading networks are usually much smaller than WWW
network, this is unlikely to become a problem (the computational times of these
three algorithms are practically zero). And the similarity measures both in the
cosine criterion and the Spearman correlation give promising results with average around 89% and 91% respectively. This high similarities are also confirmed
by the top ten countries shown in Table 2 and 3. Thus, it can be conferred
that the proposed algorithm gives meaningful results in measuring the degree
of importance of vertices in trading networks.
However, an important issue arises concerning the usefulness of the proposed
algorithm. If the total volumes can describe the degree of importance, one
can argue about the meaning of using the proposed algorithm which is clearly
much more expensive to compute. Before answering this question, we should
make clear that in general the problem of assigning the degree of importance to
vertices in a graph doesn’t have correct solution. Rather, the “correct” issue is
Table 1: The performances of the proposed algorithm.
Data
Steel products
Ethylene
Propylene
Sodium
Hydrogen peroxide
Carbon
Radio-active
Plastics
Medicinal products
Average
#Vert.
97
43
38
49
47
51
53
53
53
54
#Edg.
2627
169
144
268
261
535
717
1410
1504
848
HITS
26
7
10
11
51
22
25
20
9
20
12
#Iterations
PR
Prop. Alg.
54
42
44
54
40
143
53
143
61
99
37
65
23
26
37
39
18
14
41
69
Similarity
cos θ
ρ
0.862
0.874
0.849
0.916
0.974
0.905
0.808
0.850
0.752
0.902
0.912
0.929
0.884
0.927
0.985
0.968
0.989
0.965
0.891
0.915
Table 2: Top ten countries in hydrogen peroxide trading.
Ordered by stand. meas.
Country
Score
Netherlands
0.132290
Canada
0.095014
United States
0.088694
Moldova
0.065088
Austria
0.059850
China
0.054194
Japan
0.048676
Italy
0.045744
Colombia
0.037772
Turkey
0.037353
Ordered by prop. alg.
Country
Score
Japan
0.172970
Norway
0.123360
Netherlands
0.114200
Canada
0.082261
Turkey
0.053170
United States
0.047059
Rep. Korea
0.043684
Moldova
0.038344
China
0.036916
Thailand
0.034545
Table 3: Top ten countries in medicinal products trading.
Ordered by stand. meas.
Country
Score
Germany
0.133530
United States
0.114520
United Kingdom
0.096001
France
0.092408
Switzerland
0.083244
Italy
0.067707
Belg-Luxemb.
0.056696
Netherlands
0.051564
Japan
0.049308
Sweden
0.033573
Ordered by prop. alg.
Country
Score
Germany
0.139490
United Kingdom
0.107270
United States
0.098509
Switzerland
0.095938
France
0.085463
Italy
0.064711
Belg-Luxemb.
0.051169
Netherlands
0.047270
Ireland
0.043663
Sweden
0.041134
how to find the useful solution. This issue has been extensively studied in WWW
network where there are numerous methods which can roughly be classified into
query-dependent scores and query-independent scores. For example content
scores are query-dependent and PageRank is query-independent. And if the
user satisfaction is considered to be the usefulness standard, PageRank seems
to be more useful than HITS.
Hence, the main purpose of the proposed algorithm is to present a new
method to compute ranking scores in trading networks which will become crucial if the problem involving finding the most important and relevant users in a
large trading network like online auction network (this is the recommendation
problem which arises as one of the most important problem in the computer science researches (Pan et al., 2006)). And because the proposed algorithm uses
the network structure, an uncaptured information in the total volumes method,
the amount of the resource is not only the factor, the link structure information is also important in determining the final scores. Thus in the proposed
algorithm’s viewpoint, a well connected vertex which can be considered an important vertex in the graph term are more favourable than a less connected
vertex with the same amount of resource.
13
5
Acceleration method for HITS
As shown in Figure 3 the preferential attachment in WWW network induced
from the HITS definition is opposite to the preferential attachment in trading
network. Therefore, the same framework when deriving the ranking algorithm
for trading network can be applied back to HITS. To derive the modified HITS
formulation, we first discuss Equation (13) because it separates the ranking vector r into buying vector b and selling vector s, so it is in the same shape with the
HITS formulation in Equation (6). By comparing the preferential attachment
in trading network in Figure 3(a) with Equation (13), we can get insight about
the relationship between the preferential attachment and the buying and selling
vectors.
5.1
Modified HITS formulation
As shown in the left hand side of Figure 3(a), an agent prefers other with many
inlinks when receiving a new inlink. And in the first part of Equation (13),
ranking score of an agent as a buyer is the sum of ranking scores of others as
buyers weighted with ca of the corresponding agents from which it receives the
resources. Because ca is bigger if an agent has many inlinks than outlinks, the
first part of Equation (13) says an agent should receive new inlinks from others
with many inlinks, which is identical to the preferential attachment shown in
the left hand side of Figure 3(a).
This is also true for the selling part (right hand side of Figure 3(a)); an
agent prefers other with many outlinks when creating a new outlink. And in
the second part of Equation (13), ranking score of an agent as a seller is the
sum of ranking scores of others as sellers weighted with ch of the corresponding
agents to which it delivers the resources. Because ch is bigger if an agent
has many outlinks, second part of Equation (13) says an agent should creates
new outlinks to others with many outlinks, which is identical to the preferential
attachment shown in the right hand side of Figure 3(a). Thus, it is clear that the
preferential attachments in Figure 3 can be utilized directly to the formulation
of the ranking algorithms.
We will use this connection to define the modified HITS. As shown in the left
hand side of Figure 3(b), a page prefers other with many outlinks when receving
a new inlink. Because in WWW network inlink corresponds with authority
concept and outlink corresponds with hub concept, this preferential attachment
implies that authority score of a page is the sum of hub scores of others that
point to it weighted with ch of the corresponding pages. And in the right hand
side of Figure 3(b), a page prefers other with many inlinks when creating a new
outlink. Consequently, hub score of a page is the sum of authority scores of
others that are pointed to by it weighted with ca of the corresponding pages.
Thus, the proposed algorithm can be written as:
X (k)
X (k+1)
(k+1)
(k+1)
=
ai
=
hj chj , and hi
caj
(15)
aj
j∈Bi
j∈Fi
And in the matrix form:
a(k+1)T = h(k)T Ch L, and h(k+1)T = a(k+1)T Ca LT
14
(16)
Scores
Final distribution
Initial distribution
0
10
20
30
40
50
60
70
80
90
100
Pages ordering
Figure 4: The distances between initial and final distributions.
As shown in the Equation (15), the proposed algorithm is HITS with the
introduction of two constants to every page. Because ca is bigger for an authoritative page and ch is bigger for a hubby page, the pages that follow the
preferential attachment will collect their scores faster as the iterations progress
under the proposed algorithm than under HITS. Thus, it can be expected that
the proposed algorithm will converge faster in the datasets that are following the
preferential attachment. As shown in the previous works (Broder et al., 2000;
Albert et al., 1999; Kleinberg et al., 1999), WWW network does have power-law
degree distributions for both inlinks and outlinks, so the preferential attachment
exists.
Usually, a uniform distribution is used as the starting vector (Langville and Meyer,
2006). Thus, the distances between initial and final scores are not uniform. For
some very authoritative and hubby pages, it takes more iterations to reach the
final scores. This is also true for pages that have very low final authority or
hub scores. Figure 4 describes such condition; the distances between initial and
final scores of the pages that ordered in the top and bottom are greater than the
pages in the middle positions. Because the authority and hub scores are proportional to ca and ch3 , the distances between final and initial authority and
hub scores are proportional to ca and ch respectively. Thus, the pages ordered
in the top (bottom) will reach the stationary values faster under the proposed
algorithm due to the bigger (smaller) ca and ch.
5.2
Experimental results
Due to the limited space, we only present the experimental results and analysis
briefly. More detailed discussions can be found in (Mirzal and Furukawa, 2010).
There are six datasets used in the experiments that consist of around 10
thousands to 225 thousands pages with average degree from 4 to 47. Except
wikipedia (Segaran, 2006), all datasets were crawled by using our crawling sys3 As
shown in (Ding et al., 2002, 2004), authority (hub) scores are proportional to the
number of inlinks (outlinks), and by definition ca (ch) values are proportional to the number
of inlinks (outlinks). Thus the authority and hub scores are proportional to the ca and ch
respectively.
15
Table 4: Datasets summary.
Data
Crawled
britannica
09/2008
jobs
12/2008
opera
12/2008
scholarpedia
06/2008
stanford
12/2008
wikipedia
09/2006
Average
#Pages
21104
16056
49749
74243
225441
10431
66170
#Links
994554
187957
437748
1077781
2196441
46152
46152
AD
47.1
11.7
8.8
14.5
9.7
4.4
16
Table 5: Top 10 results with query “programming” for wikipedia dataset.
No.
1
2
3
4
5
6
7
8
9
10
HITS
Programming_language
Categorical_list_of_
programming_languages
C_programming_language
Functional_programming
Object-oriented_programming
Programming_paradigm
Java_programming_language
Generic_programming
Lisp_programming_language
Ada_programming_language
Prop. Alg.
Programming_language
Categorical_list_of_
programming_languages
C_programming_language
Functional_programming
Object-oriented_programming
Java_programming_language
Programming_paradigm
Generic_programming
Lisp_programming_language
Ada_programming_language
tem (Mirzal, 2009b). All datasets, but britannica, have a typical WWW network average degree, around 4 to 15 (Langville and Meyer, 2006; Kamvar et al.,
2003). Table 4 summarizes the datasets where AD denotes the average degree.
The experiments are conducted by using a notebook with 1.86 GHz Intel
Processor and 2 GB RAM. The codes are written in python by extensively using
database to store lists of adjacency matrices, score vectors, and other related
data. Figure 5 shows the convergence rates and Figure 6 shows processing times
to achieve the same corresponding residual levels. Note that the uniform starting
vectors are used for all datasets, and ca and ch computations have already been
included in the processing times.
As shown in Figure 5 and 6, the proposed algorithm in general can give
improvements to both convergence rates and processing times. While in the
processing times there are still some cases where the proposed algorithm cannot
do better than HITS, in the convergence rates the proposed algorithm performs
better than HITS in all cases. Table 5 gives examples of top ten pages returned
by HITS and the proposed algorithm with query “programming” for wikipedia
dataset. Note that for brevity only file names are displayed. To get full URLs,
each name has to be prefixed with “http://en.wikipedia.org/wiki/”.
6
Conclusion
We present a link structure ranking algorithm for trading network which is derived from analyzing the preferential attachment mechanism in the network. We
show that the mechanism in trading network is opposite to the corresponding
mechanism in WWW network induced from the HITS definition. The differ16
0
0
HITS
Prop. alg.
−1
−1
−2
−2
−3
−3
−4
−4
−5
−5
−6
1
2
3
4
5
6
−6
1
7
2
(a) http://www.britannica.com
3
4
5
6
7
8
9
(b) http://www.jobs.ac.uk
0
1
0
−1
−1
−2
−2
−3
−3
−4
−4
−5
−6
2
4
6
8
10
12
14
16
18
−5
1
20
(c) http://www.opera.com
2
3
4
5
6
(d) http://www.scholarpedia.org
1
0
0
−1
−1
−2
−2
−3
−3
−4
−4
−5
−5
−6
1
2
3
4
5
6
7
8
9
10
11
−6
12
(e) http://www.stanford.edu
2
4
6
8
10
12
14
16
(f) http://www.wikipedia.org
Figure 5: Convergence rates comparison. Note that x-axis is the number of
iterations and y-axis is the residual in log scale.
ences come from the fact that in trading network the links are the flows of the
resources driven by the supply and demand principle, a fundamental principle
behind the trading activities, and in WWW network the links are the hyperlinks
that can be created without exchanging any resource. Because the preferential
attachment is the underlying principle behind the HITS formulation, by utilizing the differences we are able to define a link structure ranking algorithm for
trading networks. The distinct feature of our algorithm is the using of network
structure in determining the ranking scores which is a popular method in the
WWW network researches.
There are some possible applications of the proposed algorithm. The most
obvious one is to use it as a metric to determine the degree of importance of
agents involved in the trading activities. Different from the standard method of
using aggregate transaction volumes, the proposed algorithm which makes use
of the network structure will favour agents that are highly connected or link to
17
1600
HITS
Prop. alg.
1400
1200
1000
800
600
400
200
0
Britannica
Jobs
Opera Scholarpedia Stanford
Wikipedia
Figure 6: Processing times (in second) to reach the same corresponding residual
levels.
(are linked by) other highly connected countries. Thus, the network structure
which is an invaluable information in the graph theory but uncaptured in the
standard method will become an essential factor in determining the degree of
importance. The second possible application is to design a recommendation
scheme; for example in online auction network where the number of users is
enormous, the proposed algorithm will be helpful in focusing efforts to only the
most important users that are relevant to the search queries.
In the WWW network part, we show that the modified HITS which favours
the preferential attachment in general has better convergence rates than the
original HITS, thus it can be used to improve the HITS computations. This is
an interesting subject on its own and has been studied in our other work. The
readers can refer to (Mirzal and Furukawa, 2010) for detailed discussions.
References
Albert, R. and Barabási, A.-L. (2000) ‘Topology of Evolving Networks: Local
Events and Universality’, Phys. Rev. Lett., Vol. 85, pp.5234–5237.
Albert, R., Jeong, H. and Barabási, A.-L. (1999) ‘Diameter of the World Wide
Web’, Nature, Vol. 401, pp.130–131.
Amaral, L.A.N., Scala, A., Barthélémy, M. and Stanley, H.E. (2000) ‘Classes of
Small-world Networks’, Proc. Natl. Acad. Sci. USA, Vol. 97, pp.11149–11152.
Ball, F., Mollison, D. and Scalia-Tomba, G. (1997) ‘Epidemics with Two Levels
of Mixing’, Annals of Applied Probability, Vol. 7, pp.46–89.
Barabási, A.-L. and Albert, R. (1999) ‘Emergence of Scaling in Random Networks’, Science, Vol. 286, pp.509–512.
Barabási, A.-L., Jeong, H., Néda, Z., Ravasz, E., Schubert, A. and Vicsek, T.
(2002) ‘Evolution of the Social Network of Scientific Collaborations’, Physica
A, Vol. 311, pp.590–614.
Barret, R., Berry, M., Chan, T.F., Demmel J., Donato J., Dongarra, J., Eijkhout, V., Pozo. R., Romine, C. and Van der Vorst, H. (1994) Templates for
18
the Solution of Linear Systems: Building Blocks for Iterative Methods, 2nd
Edition, SIAM, Philadelphia.
Bianconi, G. and Barabási, A.-L. (2001) ‘Competition and Multiscaling in
Evolving Networks’, Europhys. Lett., Vol. 54, pp.436–442.
Brin, S. and Page, L. (1998) ‘The Anatomy of a Large-scale Hypertextual Web
Search Engine’, Computer Networks and ISDN Systems, Vol. 33, pp.107–117.
Brin, S., Page, L., Motwami R. and Winograd, T. (1999) ‘The PageRank Citation Ranking: Bringing Order to the Web’, Technical Report 1999-0120,
Computer Science Department, Stanford University.
Broder, A., Kumar, R., Maghoul, F., Raghavan, P., Rajagopalan, S., Stata, R.,
Tomkins, A. and Wiener, J. (2000) ‘Graph Structure in the Web’, Computer
Networks, Vol. 33, pp.309–320.
Ding, C., He, X., Zha, H. and Simon H. (2002) ‘PageRank, HITS and a unified
framework for the link analysis’, ACM SIGIR, pp.353–354.
Ding, C., Zha, H., He, X., Husbands, P. and Simon H. (2004) ‘Link analysis:
Hubs and Authorities on the World Wide Web’, SIAM Review, Vol. 46 No.
2, pp.256–268.
Eiron, N. and McCurley, K.S. (2003) ‘Analysis of Anchor Text for Web Search’,
Proc. SIGIR, ACM Press, pp.459–460.
Erdős, P. and Renyı́, A. (1959) ‘On Random Graphs’, Publicationes Mathematicae, Vol. 6, pp.290–297.
Erdős, P. and Renyı́, A. (1960) ‘On the Evolution of the Random Graphs’, Publication of the Mathematical Institute of the Hungarian Academy of Sciences,
Vol. 5, pp.17–61.
Erdős, P. and Renyı́, A. (1961) ‘On the Strength of Connectedness of a Random
Graph’, Acta Mathematica Scientia Hungary, Vol. 12, pp.261–267.
Faloutsos, M., Faloutsos, P. and Faloutsos, C. (1999) ‘On Power-Law Relationships of the Internet Topology’, Computer Communications Review, Vol. 29,
pp.251–262.
Farahat, A., Lofaro, T., Miller, J.C., Rae, G. and Ward, L.A. (2006) ‘Authority
Rankings from HITS, PageRank, and SALSA: Existence, Uniqueness, and
Effect of Initialization’, SIAM Journal of Scientific Computing, Vol. 27, No.
4, pp.1181–1201.
Fujii, A. (2008) ‘Modeling Anchor Text and Classifying Queries to Enhance Web
Document Retrieval’, Proc. 17th WWW Conference, pp.337–346.
Jeong, H., Néda, Z. and Barabási, A.-L. (2003) ‘Measuring Preferential Attachment in Evolving Networks’, Europhys. Lett., Vol. 61, pp.567–572.
Jeong, H., Tombor, B., Albert, R., Oltvai, Z.N. and Barabási, A.-L. (2000) ‘The
Large-Scale Organization of Metabolic Networks’, Nature, Vol. 407, pp.651–
654.
19
Kamvar, S.D., Haveliwala, T.H., Manning, C.D. and Golub, G.H. (2003) ‘Exploiting the Block Structure of the Web for Computing PageRank’, Technical
Report 2003-17, Stanford University.
Keeling, M.J. (1999) ‘The Effects of Local Spatial Structure on Epidemiological
Invasion’, Proc. R. Soc. London B, Vol. 266, pp.859–867.
Kleinberg, J.M. (1999) ‘Authoritative Sources in A Hyperlink Environment’,
Journal of ACM, Vol. 46, pp.604–632.
Kleinberg, J.M., Kumar, R., Raghavan, P., Rajagopalan, S. and Tomkins, A.
(1999) ‘The Web as a Graph: Measurements, Models, and Methods’, COCOON, pp.1–17.
Kolda, T.G., Bader, B.W. and Kenny, J.P. (2005) ‘Higher-Order Web Link Analysis Using Multilinear Algebra’, Proc. 5th IEEE International Conference on
Data Mining, pp.242–249.
Konchady, M. (2006) Text Mining Application Programming, Charles River Media, Boston, USA.
Kuperman, M. and Abramson, G. (2001) ‘Small World Effect in an Epidemiological Model’, Phys. Rev. Lett., Vol. 86, pp.2909–2912.
Langville, A.N. and Meyer, C.D. (2006) Google’s PageRank and Beyond, Princeton University Press.
Liljeros, F., Edling, C.R., Amaral, L.A.N., Stanley, H.E. and Arberg, Y. (2001)
‘The Web of Human Sexual Contacts’, Nature, Vol. 411, pp.907–908.
Merton, R.K. (1968) ‘The Matthew Effect in Science’, Science, Vol. 159, pp.56–
63.
Mirzal, A. (2009a) ‘Link Structure Ranking Algorithm for Trading Networks’,
Proc. International Conference on Complex, Intelligent, and Software Intensive Systems, IEEE Comp. Soc., pp.120–127.
Mirzal, A. (2009b) ‘PyThinSearch: A Simple Web Search Engine’, Proc. International Conference on Complex, Intelligent, and Software Intensive Systems,
IEEE Comp. Soc., pp.1–7.
Mirzal, A. and Furukawa, M. (2010) ‘A Method for Accelerating the HITS
Algorithm’, To appear in Journal of Advanced Computational Intelligence
and Intelligent Informatics, arXiv:0909.0572.
Newman, M.E.J. (2001) ‘The Structure of Scientific Collaboration Networks’,
Proc. Natl. Acad. Sci. USA, Vol. 98, pp.404–409.
Newman, M.E.J., Barabási, A.-L. and Watts, D.J. (2006) The Structure and
Dynamics of Networks, Princeton University Press, USA.
Newman, M.E.J., Strogatz, S.H. and Watts, D.J. (2001) ‘Random Graphs with
Arbitrary Degree Distributions and Their Applications’, Phys. Rev. E, Vol.
64, 026118.
20
Pan, Z., Xu, B., Yang, H. and Zhang, M. (2006) ‘Content-based personalised
recommendation in virtual shopping environment’, International Journal of
Business Intelligence and Data Mining, Vol. 1, No. 4, pp.430–449.
Pastor-Satorras, R. and Vespignani A. (2001) ‘Epidemic Spreading in Scale-Free
Networks’, Phys. Rev. Lett., Vol. 86, pp.3200–3203.
Pool, I. de S. and Kochen, M. (1978) ‘Contacts and Influence’, Social Networks,
Vol. 1, pp.1–48.
Price, D.J. de S. (1965) ‘Networks of Scientific Papers’, Science, Vol. 149,
pp.510–515.
Price, D.J. de S. (1976) ‘A General Theory of Bibliometric and other Cumulative
Advantage Processes’, J. Amer. Soc. Inform. Sci., Vol. 27, pp.292–306.
Reed, W.J. (2001) ‘The Pareto, Zipf and other power laws’, Economics Letters,
Vol. 74, pp.15–19.
Segaran, T. (2006) ‘http://kiwitobes.com/db/searchindex.db’.
Simon, H.A. (1955) ‘On a Class of Skew Distribution Functions’, Biometrika,
Vol. 42, pp.424–440.
Solomonoff, R. and Rapoport, A. (1951) ‘Connectivity of Random Nets’, Bulletin of Mathematical Biophysics, Vol. 13, pp.107–117.
Travers, J. and Milgram, S. (1969) ‘An Experimental Study of the Small World
Problem’, Sociometry, Vol. 32, pp.425–443.
United Nations (1996) Annual Bulletin of Trade in Chemical Products 1996.
United Nations (1999) Annual Bulletin of Statistics of World Trade in Steel
1998. Destination of Export of Steel Products (Total).
Wagner, A. and Fell, D. (2001) ‘The Small World inside Large Metabolic Networks’, Proc. Royal Society London B, Vol. 268, pp.1803–1810.
Watts, D.J. and Strogatz, S.H. (1998) ‘Collective Dynamics of ’Small-world’
Networks’, Nature, Vol. 393, pp.440–442.
Yule, G. U. (1925) ‘A Mathematical Theory of Evolution, based on the Conclusions of Dr. J. C. Willis, F.R.S.’, Philosophical Transactions of the Royal
Society of London, Ser. B 213, pp.21–87.
Zipf, G.K. (1935) The Psycho-biology of Language, Boston, USA.
21
| 5cs.CE
|
Towards “Propagation = Logic + Control”
Sebastian Brand1 and Roland H.C. Yap2
1
National ICT Australia, Victoria Research Lab, Melbourne, Australia
School of Computing, National University of Singapore, Singapore
arXiv:cs/0608015v1 [cs.PL] 3 Aug 2006
2
Abstract. Constraint propagation algorithms implement logical inference. For efficiency, it is essential to control whether and in what order
basic inference steps are taken. We provide a high-level framework that
clearly differentiates between information needed for controlling propagation versus that needed for the logical semantics of complex constraints
composed from primitive ones. We argue for the appropriateness of our
controlled propagation framework by showing that it captures the underlying principles of manually designed propagation algorithms, such
as literal watching for unit clause propagation and the lexicographic ordering constraint. We provide an implementation and benchmark results
that demonstrate the practicality and efficiency of our framework.
1
Introduction
Constraint programming solves combinatorial problems by combining search and
logical inference. The latter, constraint propagation, aims at reducing the search
space. Its applicability and usefulness relies on the availability of efficiently executable propagation algorithms.
It is well understood how primitive constraints, e. g. indexical constraints,
and also their reified versions, are best propagated. We also call such primitive constraints pre-defined, because efficient, special-purpose propagation algorithms exist for them and many constraint solving systems provide implementations. However, when modelling problems, one often wants to make use of
more complex constraints whose semantics can best be described as a combination of pre-defined constraints using logical operators (i. e. conjunction, disjunction, negation). Examples are constraints for breaking symmetries [FHK+ 02]
and channelling constraints [CCLW99].
Complex constraints are beneficial in two aspects. Firstly, from a reasoning
perspective, complex constraints give a more direct and understandable highlevel problem model. Secondly, from a propagation perspective, the more more
global scope of such constraints can allow stronger inference. While elaborate
special-purpose propagation algorithms are known for many specific complex
constraints (the classic example is the alldifferent constraint discussed in [Rég94]),
the diversity of combinatorial problems tackled with constraint programming in
practice implies that more diverse and rich constraint languages are needed.
Complex constraints which are defined by logical combinations of primitive constraints can always be decomposed into their primitive constituents and
Boolean constraints, for which propagation methods exist. However, decomposing in this way may
(A) cause redundant propagation, as well as
(B) limit possible propagation.
This is due to the loss of a global view: information between the constituents of
a decomposition is only exchanged via shared constrained variables.
As an example, consider the implication constraint x = 5 → y 6= 8 during
constructive search. First, once the domain of x does not contain 5 any more,
the conclusion y 6= 8 is irrelevant for the remainder of the search. Second, only
an instantiation of y is relevant as non-instantiating reductions of the domain of
y do not allow any conclusions on x. These properties are lost if the implication
is decomposed into the reified constraints (x = 5) ≡ b1 , (y 6= 8) ≡ b2 and the
Boolean constraints not(b1 , b′1 ), or(b′1 , b2 ).
Our focus is point (A). We show how shared control information allows a
constraint to signal others what sort of information is relevant to its propagation
or that any future propagation on their part has become irrelevant to it. We
address (B) to an extent by considering implied constraints in the decomposition.
Such constraints may be logically redundant but not operationally so. Control
flags connecting them to their respective antecedents allow us to keep track of
the special status of implied constraint, so as to avoid redundant propagation
steps. Our proposed control framework is naturally applicable not only to the
usual tree-structure decomposition but also to those with a more complex DAG
structure, which permits stronger propagation.
Our objective is to capture the essence of manually designed propagation
algorithms, which implicitly merge the separate aspects of logic and control.
We summarise this by Propagation = Logic + Control in the spirit of [Kow79].
The ultimate goal of our approach is a fully automated treatment of arbitrary
complex constraints specified in a logic-based constraint definition language. We
envisage such a language to be analogous to CLP but focused on propagation.
Our framework would allow users lacking the expertise in or the time for the development of specialised propagation to rapidly prototype and refine propagation
algorithms for complex constraints.
Preliminaries
Consider a finite sequence of different variables X = x1 , . . . , xm with respective
domains D(x1 ), . . . , D(xm ). A constraint C on X is a pair hS, Xi. The set
S is an m-ary relation and a subset of the Cartesian product of the domains,
that is, S ⊆ D(x1 ) × . . . × D(xm ). The elements of S are the solutions of the
constraint, and m is its arity. We assume m > 1. We sometimes write C(X) for
the constraint and often identify C with S.
We distinguish pre-defined, primitive constraints, such as x = y, x 6 y,
and complex constraints, constructed from the primitive constraints and the
logical operators ∨, ∧, ¬ etc. For each logical operator there is a corresponding
Boolean constraint. For example, the satisfying assignments of x∨y = z are the
solutions of the constraint or(x, y, z). The reified version of a constraint C(X)
is a constraint on X and an additional Boolean variable b reflecting the truth
of C(X); we write it as C(X) ≡ b. Complex constraints can be decomposed
into a set of reified primitive constraints and Boolean constraints, whereby new
Boolean variables are introduced. For example, the first step in decomposing
C1 ∨ C2 may result in the three constraints C1 ≡ b1 , C2 ≡ b2 , and or(b1 , b2 , 1).
Constraint propagation aims at inferring new constraints from given constraints. In its most common form, a single constraint is considered, and the
domains of its variables are reduced without eliminating any solution of the constraint. If every domain is maximally reduced and none is empty, the constraint
is said to be domain-consistent (DC). For instance, x < y with D(x) = {1, 2},
D(y) = {1, 2, 3} can be made domain-consistent by inferring the constraint
y 6= 1, leading to the smaller domain D(y) = {2, 3}.
Decomposing a complex constraint may hinder propagation. For example,
DC-establishing propagation is guaranteed to result in the same domain reductions on a constraint and its decomposition only if the constraint graph of the
decomposition is a tree [Fre82]. For instance, the constraints of the decomposition of the constraint (x > y) ∧ (x < y) considered in isolation do not indicate
its inconsistency.
2
Logic and Control Information
A complex constraint expressed as a logical combination of primitive constraints
can be decomposed into its primitive parts. However, such a naive decomposition
has the disadvantage that it assigns equal relevance to every constraint. This may
cause redundant reasoning to take place for the individual primitive constraints
and connecting Boolean constraints. We prevent this by maintaining fine-grained
control information on whether the truth or falsity of individual constraints
matters. We say that a truth status of a constraint is relevant if it entails the
truth status of some other constraint.
We focus on the disjunction operator first.
Proposition 1. Suppose C is the disjunctive constraint C1 ∨ C2 . Consider the
truth status of C in terms of the respective truth statuses of the individual constraints C1 , C2 .
– If the falsity of C is asserted then the falsity of C1 and C2 can be asserted.
– If the truth of C is asserted then the falsity of C1 and C2 is relevant, but not
their truth.
– If the truth of C is queried then the truth of C1 and C2 is relevant, but not
their falsity.
– If the falsity of C is queried then the falsity of only one of C1 or C2 is
relevant, but not the their truth.
Proof. Let the reified version of C be (C1 ∨C2 ) ≡ b and its partial decomposition
be C1 ≡ b1 , C2 ≡ b2 , or(b1 , b2 , b). The following cases can occur when asserting
or querying C.
Case b = 0. Then C1 and C2 must both be asserted to be false.
Case b = 1.
– Suppose C1 is found to be true. This means that both the truth and the
falsity of C2 , hence C2 itself, have become irrelevant for the remainder
of the current search. Although this simplifies the representation of C to
C1 , it does not lead to any inference on it. In this sense, the truth of C1
is useless information.
The case of C2 being true is analogous.
– Suppose C1 is found to be false. This is useful information as we now
must assert the truth of C2 , which may cause further inference in C2 .
The case of C2 being false is analogous.
Only falsity of C1 or C2 is information that may cause propagation. Their
truth is irrelevant in this respect.
Case b unknown. We now assume that we know what aspect of the truth
status of C is relevant: its truth or its falsity. If neither is relevant then we
need not consider C, i. e. C1 and C2 , at all. If both the truth and falsity of
C are relevant, the union of the individual cases applies.
Truth of C is queried:
– Suppose C1 or C2 is found to be true. This means that C is true,
and knowing either case is therefore useful information.
– Suppose C1 is found to be false. Then the truth of C depends on the
truth of C2 . The reasoning for C2 being false is analogous.
The truth of both C1 and C2 matters, but not their falsity.
Falsity of C is queried:
– Suppose C1 or C2 is found to be true. While this means that C is
true, this is not relevant since its falsity is queried.
– Suppose C1 is found to be false. Then the falsity of C depends on the
falsity of C2 . Now suppose otherwise that C1 is queried for falsity
but not found to be false. If C1 is not false then C cannot be false.
It is important to realise that this reasoning is independent of C2 .
The reasoning for C2 being false is symmetric.
In summary, to determine the falsity of C, it suffices to query the falsity
of just one of C1 or C2 .
⊓
⊔
Fig. 1 shows the flow of control information through a disjunction. There,
and throughout the rest of this paper, we denote a truth query by chk-true and
a falsity query by chk-false.
Analogous studies on control flow can be conducted for all other Boolean
operators. The case of a negated constraint is straightforward: truth and falsity
swap their roles. Conjunction is entirely symmetric to disjunction due to De
Morgan’s law. For example, a query for falsity of the conjunction propagates to
both conjuncts while a query for truth need only be propagated to one conjunct.
We remark that one can apply such an analysis to other kinds of operators
including non-logical ones. Thus, the cardinality constraint [HD91] can be handled
within this framework.
0
1
chk-false
chk-true
or
or
or
or
0
0
chk-false
chk-false
chk-false
chk-true
chk-true
Fig. 1. Control flow through a disjunction
2.1
Controlled Propagation
Irrelevant inference can be prevented by distinguishing whether the truth or
the falsity of a constraint matters. This control information arises from truth
information and is propagated similarly. By controlled propagation we mean
constraint propagation that (1) conducts inference according to truth and falsity
information and (2) propagates such information.
We now characterise controlled propagation for a complex constraint in decomposition. We are interested in the effective propagation, i. e. newly inferred
constraints (such as smaller domains) on the original variables rather than on
auxiliary Boolean variables. We assume that only individual constraints are propagated1 . This is the usual case in practice.
Theorem 1 (Controlled Propagation). Controlled and uncontrolled propagation of the constraints of the decomposition of a constraint C are equivalent
with respect to the variables of C if only single constraints are propagated.
⊓
⊔
Proof. Proposition 1 and analogous propositions for the other Boolean operators.
In the following, we explain a formal framework for maintaining and reacting
to control information.
Control store. Constraints communicate truth information by shared Boolean
variables. Similarly, we think of control information being communicated between constraints by shared sets of control flags. As control flags we consider the
truth status queries chk-true, chk-false and the additional flag irrelevant signalling permanent irrelevance. In this context, ‘permanently’ refers to subsidiary parts of the search, that is, until the next back-tracking. Note that the
temporary absence of truth and falsity queries on a constraint is not the same
as its irrelevance. We write
C with F S
to mean that the constraint C can read and update the sequence of control flag
sets F S. One difference between logic and control information communication
is that control flows only one way, from a producer to a consumer.
1
E. g., path-consistency enforcing propagation considers two constraints at a time.
Propagating control. A set of control flags F is updated by adding or deleting
flags. We abbreviate the adding operation F := F ∪ {f } as F ∪= f . We denote
by F1
F2 that from now on permanently changes to the control flags in F1
are reflected in corresponding changes to F2 ; e. g. an addition of f to F1 leads
to an addition of f to F2 .
We employ rules to specify how control information is attached to the constituents of a decomposed complex constraint, and how it propagates. The rule
A ⇒ B denotes that the conditions in A, consisting of constraints and associated
control information, entail the constraints and the updates of control information
specified in B. We use delete statements in the conclusion to explicitly remove
a constraint from the constraint store once it is solved or became permanently
irrelevant.
Relevance. At the core of controlled propagation is the principle that reasoning
effort should be made only if it is relevant to do so, that is, if the truth or falsity
of the constraint at hand is asserted or queried. We reflect this condition in the
predicate
is relevant (b, F )
:=
b = 1 or chk-true ∈ F
or
b = 0 or chk-false ∈ F .
(is rel)
It applies to constraints in the form C ≡ b with F . We show later that this
principle can be applied to primitive constraints.
2.2
Boolean Constraints
We again focus on disjunctive constraints. The following rule decomposes the
constraint (C1 ∨ C2 ) ≡ b only if the relevance test is passed. In this case the
shared control sets are initialised.
is relevant (b, F )
⇒ or(b, b1 , b2 ) with hF , F1 , F2 i,
C1 ≡ b1 with F1 , F1 := ∅,
C2 ≡ b2 with F2 , F2 := ∅.
(ordec )
The following rules specify how control information propagates through this
disjunctive constraint in accordance with Proposition 1:
b=1
b1 = 0
⇒
⇒
F1 ∪= chk-false, F2 ∪= chk-false;
F
F2 , delete or(b, b1 , b2 );
b2 = 0
b1 = 1
⇒
⇒
F
F1 , delete or(b, b1 , b2 );
F2 ∪= irrelevant, delete or(b, b1 , b2 );
b2 = 1
chk-false ∈ F
⇒
⇒
F1 ∪= irrelevant, delete or(b, b1 , b2 );
F1 ∪= chk-false;
chk-true ∈ F
irrelevant ∈ F
⇒
⇒
F1 ∪= chk-true, F2 ∪= chk-true;
F1 ∪= irrelevant, F2 ∪= irrelevant, delete or(b, b1 , b2 ).
(orcf )
In rule (orcf ), we arbitrarily select the first disjunct to receive chk-false. For
comparison and completeness, here are the rules propagating truth information:
b1 = 0
⇒
b = b2 ;
b1 = 1
⇒
b = 1;
b2 = 0
b=0
⇒
⇒
b = b1 ;
b1 = 0, b2 = 0.
b2 = 1
⇒
b = 1;
Control propagation for the negation constraint not(b, bN ) with hF , FN i is
straightforward:
b = 1 or b = 0 or bN = 1 or bN = 0
chk-false ∈ F
⇒
⇒
delete not(b, bN );
FN ∪= chk-true;
chk-true ∈ F
irrelevant ∈ F
⇒
⇒
FN ∪= chk-false;
FN ∪= irrelevant.
The rules for other Boolean operators are analogous. Note that a move from
binary to n-ary conjunctions or disjunctions does not affect the control flow in
principle, in the same way that the logic is unaffected.
Both chk-true and chk-false can be in the control set of a constraint at the
same time, as it might be in a both positive and negative context. An example
is the condition of an if-then-else. On the other hand, if for instance a constraint
is not in a negated context, chk-false cannot arise.
2.3
Primitive Constraints
Asserting and querying other primitive constraints can be controlled similarly
to Boolean constraints. In particular, the relevance condition (is rel) must be
satisfied before inspecting a constraint. We furthermore deal with irrelevant ∈ F
as expected, by not asserting the primitive constraint or by deleting it from the
set of currently queried or asserted constraints.
When a query on a primitive constraint is inconclusive, it is re-evaluated
whenever useful. This can be when elements from a variable domain are removed
or when a bound changes. We rely on the constraint solving environment to signal
such changes.
Deciding the truth or the falsity of a constraint in general is an expensive
operation that requires the evaluation of every variable domain. A primitive
C(X) is guaranteed to be true if and only if C(X) ⊆ D(X) and C(X) is nonempty. C is guaranteed to be false if and only if C(X) ∩ D(X) = ∅, where X =
x1 , . . . , xn and D(X) = D(x1 ) × . . . × D(xn ). For some primitive constraints we
can give complete but simpler evaluation criteria, similarly to indexicals [CD96];
see Tab. 1.
Practical constraint solving systems usually maintain domain bounds explicitly. This makes answering the truth query for equality constraints and the
queries for ordering constraints very efficient. Furthermore, the re-evaluation of
a query can be better controlled: only changes of the respective bounds are an
event that makes a re-evaluation worthwhile.
Constraint
true if
false if
x∈S
x=a
x=y
x6y
D(x) ⊆ S
|D(x)| = 1, D(x) = {a}
|D(x)| = |D(y)| = 1, D(x) = D(y)
max(D(x)) 6 min(D(y))
D(x) ∩ S = ∅
a∈
/ D(x)
D(x) ∩ D(y) = ∅
min(D(x)) > max(D(y))
Table 1. Primitive constraint queries (S is a constant set, a is a constant value)
3
Implied Constraints
Appropriate handling of implied constraints fits naturally into the control propagation framework. Suppose the disjunctive constraint C1 ∨ C2 implies C⊲ ; that
is, (C1 ∨ C2 ) → C⊲ is always true. Logically, C⊲ is redundant. In terms of
constraint propagation, it may not be, however.
Consider the disjunction (x = y) ∨ (x < y), which implies x 6 y. Assume the
domains are D(x) = {4, 5}, D(y) = {3, 4, 5}. Since the individual disjuncts are
not false, there is no propagation from the decomposition. In order to conclude
x 6 y and thus D(y) = {4, 5} we associate the constraint with its implied
constraint.
We write a disjunctive constraint annotated with an implied constraint as
C1 ∨ C2 ⊲ C⊲ .
To benefit from the propagation of C⊲ , we could represent this constraint as
(C1 ∨ C2 ) ∧ C⊲ . However, this representation has the shortcoming that it leads
to redundant propagation in some circumstances. Once one disjunct, say, C1 ,
is known to be false, the other disjunct, C2 , can be imposed. The propagation
of C⊲ is then still executed, however, while it is subsumed by that of C2 . It
is desirable to recognise that C⊲ is operationally redundant at this point. We
capture this situation by enhancing the decomposition rule (ordec ) as follows:
(C1 ∨ C2 ⊲ C⊲ ) ≡ b with F
⇒ or⊲ (b, b1 , b2 , b⊲ ) with hF , F1 , F2 , F⊲ i,
C1 ≡ b1 with F1 , F1 := ∅,
C2 ≡ b2 with F2 , F2 := ∅,
C⊲ ≡ b⊲ with F⊲ , F⊲ := ∅.
Additionally to the control rules for regular disjunctive constraints shown earlier,
we now also use the following four rules:
b⊲ = 0 ⇒ b = 0;
b = 1 ⇒ b⊲ = 1;
b1 = 0 ⇒ F⊲ ∪= irrelevant, delete or⊲ (b, b1 , b2 , b⊲ );
b2 = 0 ⇒ F⊲ ∪= irrelevant, delete or⊲ (b, b1 , b2 , b⊲ ).
We envisage the automated discovery of implied constraints, but for now we
assume manual annotation.
4
Subconstraint Sharing: From Trees to DAGs
The straightforward decomposition of complex constraints can contain unnecessary copies of the same subconstraint in different contexts. The dual constraint
graph (whose vertices are the constraints and whose edges are the variables) is
a tree, while often a directed acyclic graph (DAG) gives a logically equivalent
but more compact representation. See, for example, CDDs [CY05].
We can apply controlled propagation to complex constraints represented in
DAG form. We need to account for the multiplicity of a constraint when handling
queries on it: the set of control flags now becomes a multiset, and in effect, we
maintain reference counters for subconstraints. Control flags need to be properly
subtracted from the control set of a constraint. For the sake of a simple example,
consider the constraint (C ∨ C1 ) ∧ (C ∨ C2 ). Fig. 2 shows a decomposition of it.
chk-false
chk-false
and
chk-false
or
chk-false
C1
C
and
chk-false
or
chk-false
C
C2
chk-false
or
chk-false 2
C
chk-false
or
C1
C2
Fig. 2. Left: no sharing. Right: sharing with reference counting
Another example is the condition in an if-then-else constraint. Opportunities
for shared structures arise frequently when constraints are defined in terms of
subconstraints that in turn are constructed by recursive definitions.
5
Case Studies
We examine several constraints studied in the literature and show that their
decomposition benefits from controlled propagation.
Literal Watching. The DPLL procedure for solving the SAT problem uses
a combination of search and inference and can be viewed as a special case of
constraint programming. Many SAT solvers based on DPLL employ unit propagation with 2-literal watching, e. g. Chaff [MMZ+ 01]. At any time, only changes
to two literals per clause are tracked, and consideration of other literals is postponed.
Let us view a propositional clause as a Boolean constraint. We define
clause(x1 , . . . , xn )
:=
x1 = 1 ∨ clause(x2 , . . . , xn )
and show in Fig. 3 the decomposition of clause(x1 , . . . , xn ) as a graph for controlled and uncontrolled propagation (where D(xi ) = {0, 1} for all xi ). Both
propagation variants enforce domain-consistency if the primitive equality constraints do and the variables are pairwise different. This corresponds to unit
propagation.
1
1
or
x1 = 1
or
chk-false
x1 = 1
chk-false
x2 = 1
or
or
x2 = 1
x3 = 1
...
chk-false
or
clause(x3 , . . . , xn )
xn = 1
Fig. 3. Uncontrolled versus controlled decomposition of clause
Uncontrolled decomposition expands fully into n − 1 Boolean or constraints
and n primitive constraints xi = 1. Controlled decomposition only expands into
two or constraints and the first two primitive constraints x1 = 1, x2 = 1. The
leaf node marked clause(x3 , . . . , xn ) is initially not expanded as neither assertion
nor query information is passed to it. The essence is that the first or constraint
results in two chk-false queries to the subordinate or constraint which passes this
query on to just one disjunct. This structure is maintained with respect to new
information such as variable instantiations. No more than two primitive equality
constraints are ever queried at a time. A reduction of inference effort as well as
of space usage results.
Controlled propagation here corresponds precisely to 2-literal watching.
Disequality of Tuples. Finite domain constraint programming generally focuses on variables over the integers. Sometimes, higher-structured variable types,
such as sets of integers, are more appropriate for modelling. Many complex constraints studied in the constraint community are on a sequence of variables and
can thus naturally be viewed as constraining a variable whose type is tuple-ofintegers. The recent study [QW05] examines how some known constraint propagation algorithms for integer variables can be lifted to higher-structured variables. One of the constraints examined is alldifferent on tuples, which requires
a sequence of variables of type tuple-of-integers to be pairwise different. Its
straightforward definition is
^
alldifferent tp(hX1 , . . . , Xn i) :=
different tp(Xi , Xj ),
i,j∈1,...,n, i<j
where
different tp(hx1 , . . . , xm i, hy1 , . . . , ym i) :=
_
xi 6= yi .
i∈1,...,m
Let us examine these constraints with respect to controlled propagation. The
different tp constraint is a large disjunction, and it behaves thus like the clause
constraint studied in the previous section – at most two disjuncts xi 6= yi are
queried for falsity at any time.
Deciding the falsity of a disequality constraint is particularly efficient when
the primitive constraints in Tab. 1 are used, i. e. falsity of disequality when the
domains are singletons. If the domains are not singletons, re-evaluation of the
query is only necessary once that is the case. In contrast, a truth query for
a disequality is (more) expensive as the domains must be intersected, and, if
inconclusive, should be re-evaluated whenever any domain
change occurred.
The alldifferent tp constraint is a conjunction of n2 different tp constraints.
Therefore, controlled propagation queries at most n(n − 1) disequality constraints for falsity at a time. Uncontrolled propagation asserts all n(n − 1)m/2
reified disequality constraints and in essence queries truth and falsity of each.
Using controlled rather than uncontrolled decomposition-based propagation for
alldifferent tp saves substantial effort without loss of effective propagation.
We remark that a specialised, stronger but non-trivial propagation algorithm
for this case has been studied in [QW05]. The controlled propagation framework
is then useful when specialised algorithms are not readily available, for example
due to a lack of expertise or resources in the design and implementation of
propagation algorithms.
Lexicographic Ordering Constraint. It is often desirable to prevent symmetries in constraint problems. One way is to add symmetry-breaking constraints
such as the lexicographic ordering constraint [FHK+ 02]. A straightforward definition is as follows:
lex(hx1 , . . . , xn i, hy1 , . . . , yn i) := x1 < y1
∨
x1 = y1 ∧ lex(hx2 , . . . , xn i, hy2 , . . . , yn i)
∨
n=0
With this definition, propagation of the decomposition does not always enforce
domain-consistency. Consider lex(hx1 , x2 i, hy1 , y2 i) with the domains D(x1 ) =
D(x2 ) = D(y2 ) = {3..5} and D(y1 ) = {0..5}. Controlled decomposition results
in the reified versions of x1 < y1 , x1 = y1 , x2 < y2 connected by Boolean
constraints. None of these primitive constraints is true or false. Yet we should
be able to conclude x1 6 y1 , hence D(y1 ) = {3..5}, from the definition of lex.
The difficulty is that the naive decomposition is weaker than the logical definition because it only reasons on the individual primitive constraints. However,
it is easy to see that x1 6 y1 is an implied constraint in the sense of Section 3,
and we can annotate the definition of lex accordingly:
lex(hx1 , . . . , xn i, hy1 , . . . , yn i) :=
x1 < y1
∨
x1 = y1 ∧ lex(hx2 , . . . , xn i, hy2 , . . . , yn i)
⊲ x1 6 y1
∨
n=0
We state without proof that propagation of the constraints of the decomposition
enforces domain-consistency on lex if the annotated definition is used.
Tab. 2 represents a trace of lex on the example used in [FHK+ 02], showing
the lazy decomposing due to controlled propagation. We collapse several atomic
inference steps and omit the Boolean constraints, and we write vi..j to abbreviate vi , . . . , vj . Observe how the implied constraints xi 6 yi are asserted, made
irrelevant and then deleted. The derivation ends with no constraints other than
x3 < y3 queried or asserted.
Asserted
Set of constraints queried for
falsity
Variable domains
x1
x2
x3
y1
y2
y3
x4
x4
x5
y5
lex(hx1..5 i,
hy1..5 i)
{2}
{1, 3, 4} {1..5} {1..2} {3..5}
{0..2} {1}
{0..4} {0..1} {0..2}
x 1 6 y1
x 2 6 y2
x 3 6 y3
x 3 6 y3
x 3 6 y3
x 1 < y1 , x 1 = y1 , x 2 < y2
{2}
{2}
{1, 3, 4} {1..5} {1..2} {3..5}
{1}
{0..4} {0..1} {0..2}
{2}
{2}
{1}
{1}
{1..5} {1..2} {3..5}
{0..4} {0..1} {0..2}
{2}
{2}
{1}
{1}
{1..4} {1..2} {3..5}
{1..4} {0..1} {0..2}
{2}
{2}
{1}
{1}
{1..4} {1..2} {3..5}
{1..4} {0..1} {0..2}
{2}
{2}
{1}
{1}
{1..4} {1..2} {3..5}
{1..4} {0..1} {0..2}
{2}
{2}
{1}
{1}
{1..3} {1..2} {3..5}
{2..4} {0..1} {0..2}
x 2 < y2 , x 2 = y2 , x 3 < y3
x 3 < y3 , x 3 = y3 , x 4 < y4
x 3 < y3 , x 3 = y3 , x 4 = y4 , x 5 < y5
x 3 < y3 , x 3 = y3 , x 4 = y4 , x 5 = y5
x 3 < y3
Table 2. An example of controlled propagation of the lex constraint
6
Implementation and Benchmarks
We implemented a prototype of the controlled propagation framework in the
CLP system ECLi PSe [WNS97], using its predicate suspension features and attributed variables to handle control information. The implementation provides
controlled propagation for the basic Boolean and primitive constraints, and it
handles implied constraints. Structure-sharing by a DAG-structured decomposition is not supported.
We conducted several simple benchmarks to compare controlled and uncontrolled propagation on constraint decompositions, using the clause, different tp,
alldifferent tp and lex constraints. A benchmark consisted of finding a solution
clause
nb. of variables
runtime (%)
different tp
alldifferent tp
lex
5 10 20 50 5 10 20 50 5 10 20 50
5 10 20 50
100 69 50 38 88 84 67 62 66 38 23 11 138 92 69 54
Table 3. Benchmark results: controlled propagation (uncontrolled prop. =
100%)
to a single constraint. For the uncontrolled propagation benchmark, the constraint was simply decomposed into built-in Boolean and primitive constraints
of ECLi PSe , and implied constraints (in lex) were conjunctively added to their
respective premise.
The number of variables in the respective tuple(s) was varied between five and
50. For the alldifferent tp benchmark, we chose 20 tuples. The variables ranged
over the interval {1..10} (except for clause). Solutions to the constraints were
searched by randomly selecting a variable and a value in its domain. This value
was either assigned or excluded from its domain; this choice was also random. To
obtain meaningful averages, every individual solution search was run a sufficient
number of times (typically a few 10000) so that the total computation time was
roughly 15 s. Each of these runs used a new initialisation of the pseudo-random
number generator resulting in a possibly different solution, while the benchmark
versions (controlled vs. uncontrolled propagation) used the same initial value
to obtain identical search trees. Every experiment was repeated five times. In
Tab. 3, we give the relative solving time with controlled propagation, based on
the corresponding uncontrolled propagation benchmark taken to be 100%.
The benchmarks show that controlling propagation can reduce the propagation time. The reduction is especially substantial for high-arity constraints. For
low-arity constraints, the extra cost of maintaining control information in our
implementation can outweigh the saving due to less propagation. While we have
not measured the space usage of the two propagation approaches, it follows from
the analyses in Section 5 that using controlled propagation for the considered
constraints often also requires less space, since constraints are decomposed only
when required.
We remark that efficiency was a minor concern in our high-level, proof-ofconcept implementation; consequently we expect that it can be improved considerably. For example, for constraints that are in negation normal form (all
constraints in our benchmark), the control flag chk-true is never created. A
simpler subset of the control propagation rules can then be used.
7
Final Remarks
Related Work. In terms of foundations, the controlled propagation framework
can be described as a refined instance of the CLP scheme (see [JM94]), by a
subdivision of the set of active constraints according to their associated truth
and falsity queries. Concurrent constraint programming (CCP) [Sar93], based on
asserting and querying constraints, is closely related; our propagation framework
can be viewed as an extension in which control is explicitly addressed and dealt
with in a fine-grained way. A practical CCP-based language such as CHR [Frü98]
would lend itself well to an implementation. For example, control propagation
rules with delete statements can be implemented as simplification rules.
A number of approaches address the issue of propagation of complex constraints. The proposal of [BW05] is to view a constraint as an expression from
which sets of inconsistent or valid variable assignments (in extension) can be
computed. It focuses more on the complexity issues of achieving certain kinds
of local consistencies. The work [BCP04] studies semi-automatic construction
of propagation mechanisms for constraints defined by extended finite automata.
An automaton is captured by signature (automaton input) constraints and state
transition constraints. Signature constraints represent groups of reified primitive constraints and are considered pre-defined. They communicate with state
transition constraints via constrained variables, which correspond to tuples of
Boolean variables of the reified constraints in the signature constraints. Similarly to propagating the constraint in decomposition, all automata constraints
are propagated independently of each other.
Controlled propagation is similar to techniques used in NoClause, a SAT
solver for propositional non-CNF formulas [TBW04], which in turn lifts techniques such as 2-literal watching from CNF to non-CNF solvers. We describe
here these techniques in a formal, abstract framework and integrate non-Boolean
primitive constraints and implied constraints, thus making them usable for constraint propagation.
Conclusion. We have proposed a new framework for propagating arbitrary
complex constraints. It is characterised by viewing logic and control as separate
concerns. We have shown that the controlled propagation framework explains
and generalises some of the principles on which efficient manually devised propagation algorithms for complex constraints are based. By discussing an implementation and benchmarks, we have demonstrated feasibility and efficiency. The
practical benefits of the controlled propagation framework are that it provides
automatic constraint propagation for arbitrary logical combinations of primitive
constraints. Depending on the constraint, controlling the propagation can result
in substantially reduced usage of time as well as space.
Our focus in this paper has been on reducing unnecessary inference steps. The
complementary task of automatically identifying and enabling useful inference
steps in our framework deserves to be addressed. It would be interesting to
investigate if automatic reasoning methods can be used to strengthen constraint
definitions, for instance by automatically deriving implied constraints.
Acknowledgements
We thank the anonymous reviewers for their comments. This paper was written
while Roland Yap was visiting the Swedish Institute of Computer Science and
their support and hospitality are gratefully acknowledged. The research here is
supported by a NUS ARF grant.
References
BCP04.
N. Beldiceanu, M. Carlsson, and T. Petit. Deriving filtering algorithms from
constraint checkers. In Wallace [Wal04], pages 107–122.
BW05.
F. Bacchus and T. Walsh. Propagating logical combinations of constraints.
In L. P. Kaelbling and A. Saffiotti, editors, Proc. of International Joint
Conference on Artificial Intelligence (IJCAI’05), pages 35–40, 2005.
CCLW99. B. M. W. Cheng, K. M. F. Choi, J. H.-M. Lee, and J. C. K. Wu. Increasing constraint propagation by redundant modeling: An experience report.
Constraints, 4(2):167–192, 1999.
CD96.
P. Codognet and D. Diaz. Compiling constraints in clp(FD). Journal of
Logic Programming, 27(3):185–226, 1996.
CY05.
K. C. K. Cheng and R. H. C. Yap. Constrained decision diagrams. In M. M.
Veloso and S. Kambhampati, editors, Proc. of 20th National Conference on
Artificial Intelligence (AAAI’05), pages 366–371. AAAI Press, 2005.
FHK+ 02. A. M. Frisch, B. Hnich, Z. Kiziltan, I. Miguel, and T. Walsh. Global constraints for lexicographic orderings. In P. Van Hentenryck, editor, Proc. of
8th International Conference on Principles and Practice of Constraint Programming (CP’02), volume 2470 of LNCS, pages 93–108. Springer, 2002.
Fre82.
E. C. Freuder. A sufficient condition for backtrack-free search. Journal of
the ACM, 29(1):24–32, 1982.
Frü98.
T. Frühwirth. Theory and practice of Constraint Handling Rules. Journal
of Logic Programming, 37(1-3):95–138, 1998.
HD91.
P. Van Hentenryck and Y. Deville. The Cardinality operator: A new logical
connective for constraint logic programming. In K. Furukawa, editor, Proc.
of 8th International Conference on Logic Programming (ICLP’91), pages
745–759. MIT Press, 1991.
JM94.
J. Jaffar and M. J. Maher. Constraint logic programming: A survey. Journal
of Logic Programming, 19 & 20:503–582, 1994.
Kow79.
R. A. Kowalski. Algorithm = Logic + Control. Communications of the
ACM, 22(7):424–436, 1979.
MMZ+ 01. M. W. Moskewicz, C. F. Madigan, Y. Zhao, L. Zhang, and S. Malik. Chaff:
Engineering an efficient SAT solver. In Proc. of 38th Design Automation
Conference (DAC’01), 2001.
QW05.
C.-G. Quimper and T. Walsh. Beyond finite domains: The All Different and
Global Cardinality constraints. In P. van Beek, editor, Proc. of 11th International Conference on Principles and Practice of Constraint Programming
(CP’04), volume 3709 of LNCS, pages 812–816. Springer, 2005.
Rég94.
J.-C. Régin. A filtering algorithm for constraints of difference in csps.
In Proc. of 12th National Conference on Artificial Intelligence (AAAI’94),
pages 362–367. AAAI Press, 1994.
Sar93.
V. A. Saraswat. Concurrent Constraint Programming. MIT Press, 1993.
TBW04. Chr. Thiffault, F. Bacchus, and T. Walsh. Solving non-clausal formulas
with DPLL search. In Wallace [Wal04], pages 663–678.
Wal04.
M. Wallace, editor. Proc. of 10th International Conference on Principles
and Practice of Constraint Programming (CP’04), volume 3258 of LNCS.
Springer, 2004.
WNS97.
M. G. Wallace, S. Novello, and J. Schimpf. ECLiPSe: A platform for constraint logic programming. ICL Systems Journal, 12(1):159–200, 1997.
| 6cs.PL
|
Single Cell and Multi-cell Performance Analysis of OFDM Index
Modulation
Shangbin Wu1,* and Maziar Nekovee1
arXiv:1803.02730v1 [cs.IT] 7 Mar 2018
1
*
Samsung R&D Institute UK
[email protected]
Abstract: This paper addresses the achievable rate of single cell and sum rate of multi-cell orthogonal frequency division multiplexing (OFDM) index modulation (IM). The single cell achievable
rate of OFDM-IM with Gaussian input is calculated using a multi-ary symmetric channel. Then,
the cumulative distribution function (CDF) of multi-cell OFDM-IM is investigated by stochastic
geometry. Furthermore, it is proved in this paper that the probability density function (PDF) of
noise plus intercell-interference (ICI) in multi-cell OFDM-IM with quadrature amplitude modulation (QAM) follows a mixture of Gaussians (MoG) distribution. Next, parameters of the MoG
distribution are estimated using a simplified expectation maximization (EM) algorithm. Upper
bound of sum rates of multi-cell OFDM-IM is derived. Furthermore, analytic and simulated results are compared and discussed.
1.
Introduction
The fifth generation (5G) cellular network is emerging to satisfy the unprecedented growth in data
traffic and the number of connected devices. A key performance indicator (KPI) of 5G is the ability to provide smooth quality of user experience at cell edges, which requires above 1 Gbps data
rates. Recently, orthogonal frequency division multiplexing (OFDM) index modulation (IM) [1]
was proposed as one of the 5G enabling technologies, as it has advantages such as increase in
per subcarrier transmit power and reduction in inter-cell interference (ICI). It was reported in [2]
and [3] that the ICI of legacy OFDM networks follows a Gaussian distribution which caused most
throughput degradation. However, the ICI of OFDM-IM is not Gaussian distributed. Therefore,
OFDM-IM is able to provide room to optimize data rate at cell edges. The philosophy behind
OFDM-IM is that only one of a number of OFDM subcarriers is active when transmitting symbols. In addition to the information carried by the transmitted symbol, the subcarrier index can
also be used to convey information. The idea of conveying information via indexes was first proposed in the spatial domain, i.e., spatial modulation (SM) [4]. A variant version of OFDM-IM
was proposed in [5], where bits were divided into blocks of bits before the OFDM-IM modulator.
Authors in [6] studied practical implementation issues of OFDM-IM such as maximum likelihood
(ML) detector, log-likelihood ratio (LLR) detector, and impact of channel estimation errors. A
low complexity ML detector for OFDM with in-phase/quadrature IM was discussed in [7], which
was implemented with a priori knowledge of noise variance. In [8], additional interleaving was
introduced to subcarriers in correlated channels, in order to provide extra diversity gain to the
OFDM-IM. Later, the combination of OFDM-IM and SM was proposed in [9], where the symbol
domain, subcarrier domain, and spatial domain formed a three dimensional signal space. Then,
1
the generalized OFDM-IM was introduced in [10], where multiple subcarriers were active. The
information conveyed by the index relies on different combinations of active subcarrier indexes.
When the transmitted symbols of OFDM-IM are quadrature-amplitude modulation (QAM) symbols, this type of OFDM-IM was named frequency QAM (FQAM) [2]. Studies of OFDM–IM
in [1], [2], [6]–[9] focused on bit error rate (BER) and frame error rate (FER) performance.
Achievable rate of OFDM–IM with different settings were reported in [11]–[13]. OFDM-IM
with finite constellation input was discussed in [11], where a closed-form lower bound of the
achievable rate was derived. The application of OFDM-IM for underwater acoustic communications as well as achievable rate with finite constellation input were investigated in [12]. Although
achievable rate of OFDM-IM with Gaussian input was investigated in [13], closed-form expression
was not provided in [13].
Also, the performance of OFDM-IM in multi-cell scenarios has been less studied. System level
simulations (SLSs) over typical hexagonal multi-cell network are able to provide certain insights of
multi-cell performance of wireless networks with OFDM-IM. However, there are two drawbacks
of SLSs. First, SLSs are time consuming. Second, the hexagonal multi-cell network layout is
usually not fulfilled in realistic situations, where base stations (BSs) are approximately distributed
in a uniform manner. Therefore, it is beneficial to have analytic results on the performance of
multi-cell scenarios. This can be investigated via a mathematics tool called stochastic geometry
[14]–[19], where BSs were assumed to be distributed following a Poisson Point Process (PPP).
In this case, the cumulative distribution function (CDF) of the signal to interference plus noise
(SINR) was expressed in closed form in different scenarios, such as ad hoc networks [16], cellular
networks [18], and cooperative networks [19].
Authors in [2] studied the statistics of ICI in multi-cell OFDM-IM with QAM inputs. The generalized Gaussian distribution was used to approximate the distribution of noise plus ICI. However,
the exact distribution of noise plus ICI of multi-cell OFDM-IM with QAM inputs was missing in
the literature. In this paper, this exact distribution will be found.
The contributions of this paper are listed as below:
1. The subcarrier index detection error probability of single cell OFDM-IM with Gaussian input
is conducted. Then, closed-form single cell achievable rate of OFDM-IM with Gaussian input
is derived.
2. The CDF of SINR of multi-cell OFDM-IM is derived using stochastic geometry.
3. The distribution of ICI of multi-cell OFDM-IM with QAM input is derived, showing that
it follows a mixture of Gaussians (MoG) distribution. In addition, the parameters of the
probability density function (PDF) of ICI are computed using a simplified expectation maximization (EM) algorithm. Then, the upper bound of sum rates of multi-cell OFDM-IM with
QAM input is studied.
The rest of this paper is structured as follows. Section 2 gives a general description of the
system model. Achievable rate of single cell OFDM-IM with Gaussian input will be investigated
in Section 3. Section 4 will study the CDF of SINR of multi-cell OFDM-IM with stochastic
geometry. Also, the distribution and its parameters of ICI are analyzed. Simulation results and
analysis are presented in Section 5. Conclusions are drawn in Section 6.
2
2.
System Model
Let us consider a downlink multi-cell network using OFDM-IM, where a target user equipment
(UE) is located at the origin. BSs are distributed as a homogeneous PPP with density λ. The
set of all BSs is denoted as S and the set of BSs interfering the target UE is denoted as S ′ .
Let α be the pathloss coefficient. Assume that the OFDM-IM system has NF subcarriers, then
log2 NF bits are conveyed by subcarrier indexes. NB is the number of BSs. Let F be the subcarrier index, which is a uniformly distributed random variable defined on {1, 2, · · · , NF } and let
Hξ = {hξ,1 , hξ,2, · · · , hξ,NF } be the set of all channel coefficients from the ξth BS to the target user
on these subcarriers. The channel coefficients hξ,k (1 6 k 6 NF ) are independently and identically distributed (i.i.d.) zero mean unit variance complex Gaussian random variables. In practice,
this i.i.d. channel assumption can be achieved by introducing interleaving between subcarriers
as [8]. The interleaving can be done via a pseudo random sequence, which is shared by the BS
and UE, such that the BS and UE can map or de–map between the original subcarrier indices and
the interleaved subcarrier indices. Throughout the paper, we assume that the target UE has perfect
knowledge of the channel coefficients from the associated BS but no knowledge from other BSs.
Let the ξth BS be the associated BS of the target UE. The distance between the ξth BS and the
target UE is d and the distance between the lth BS and the target UE is dl (∀l 6= ξ). Then, the
received signal Y of the target UE can be expressed as
p
Y = Tξ Hξ Xξ + N + I,
(1)
where N is a zero mean complex Gaussian noise with variance σN2 , Hξ is uniformly distributed
random variable defined on Hξ , Xξ is the transmitted symbol from the ξth BS to the target user,
Tξ is the average received power (including transmit power, path loss, and shadow fading) from
the ξth BS to the target UE, and I is the interference from other BSs. Thus, interference I can be
written as
Xp
I=
Tl Hl Xl ζl ,
(2)
l6=ξ
where ζl = 1 if the lth BS is transmitting on the same subcarrier as the ξth BS and ζl = 0 if the
lth BS is not transmitting on the same subcarrier as the ξth BS. This is due to a basic property of
OFDM-IM, which activates only one subcarrier in each transmission period.
3.
Single Cell OFDM-IM
Achievable rate of OFDM-IM with QAM input and other finite constellation inputs can be found
in [2], [11]–[13]. However, closed-form expression for achievable rate of sing-cell OFDM-IM
with Gaussian input is missing in the literature. Therefore, to fill this gap, single cell achievable
rate of OFDM-IM with Gaussian input is analyzed in this section. Since single cell is considered,
subscript ξ representing the ξth BS is dropped for brevity and the interference term I equals 0. Information of OFDM-IM is conveyed by the symbol X and the frequency index F . The achievable
rate r in this paper is defined by the average maximum achievable mutual information between the
information source (X, F ) and the destination Y , which can be characterized as
r = E [max I(X, F ; Y )]
≈ E [max I(X; Y |F )] + E [max I(F ; Y )] = r1 + r2 .
3
(3)
Fig. 1. Diagram of an NF -ary symmetric channel.
Since the channel coefficient H is determined once the subcarrier index F is determined, the
achievable rate of the single cell OFDM-IM generated by the symbol in terms of signal to noise
ratio (SNR) ρ can be calculated using the achievable rate of the Rayleigh fading channel, i.e.,
r1 (ρ) = E [max I(X; Y |F )] = E [max I(X; Y |H)]
Z∞
1
1
1
−z
= log2 (1 + zρ) e dz = −
exp( ),
Ei −
ln2
ρ
ρ
(4)
0
R ∞ −t
where Ei (·) is the exponential integral [20] defined by Ei (−z) = − z e t dt. The SNR ρ can
easily be controlled by adjusting the BS transit power to compensate path loss and shadow fading
in a single cell scenario.
The calculation of r2 is equivalent to determining how much information is retrieved from Y
when the information is conveyed on a certain subcarrier F . The information retrieving process is
not perfect because of the existence of noise. As a result, the subcarrier index may be incorrectly
detected. Let F̂ be the detected subcarrier index and PF 6=F̂ be the subcarrier index detection error
probability and denote PF =k∩F̂ =l as the probability that the kth subcarrier is used whereas the lth
subcarrier is detected. With the assumption that the channel coefficients on subcarriers are i.i.d.,
P F̂
it can be observed that PF =fk ∩F̂ =fl = NFF6=−1
(∀l 6= k). Hence, the channel between F and Y can
be abstracted by a NF -ary symmetric channel as depicted in Fig. 1. The achievable rate r2 (ρ) of a
NF -ary symmetric channel depends on the subcarrier index detection error probability and can be
presented as [21]
h
i
r2 (ρ) = log2 NF − Hb PF 6=F̂ (ρ) − PF 6=F̂ (ρ) log2 (NF − 1),
(5)
where Hb [µ] is the binary entropy function defined by Hb [µ] = −µ log2 µ − (1 − µ) log2 (1 − µ).
In order to calculate r2 (ρ), we need to compute PF 6=F̂ (ρ) with the following lemma.
4
Lemma 1. The subcarrier index detection error probability can be presented as
h
q
i
NF −k
r NF −1
1
−
Φ
2ρ(NF −k−1)
π X k
NF − k
NF −k−1
p
,
PF 6=F̂ (ρ) = 1 −
C
(−1)
exp
2 k=0 NF −1
2ρ(NF − k − 1)
ρ(NF − k − 1)(NF − k)
(6)
where CNk F −1 is the binomial coefficient defined by CNk F −1 =
Rz
2
function defined by Φ(z) = √2π e−t dt.
(NF −1)!
k!(NF −k−1)
and Φ(z) is the error
0
Proof. Conditioning on X, errors occur in the detection of the subcarrier index when the received
signal power on the intended subcarrier is less then any one of the rest subcarriers. Due to the
symmetry, it is sufficient to calculate the subcarrier index detection error probability assuming that
the first subcarrier is used. Let Yk (1 6 k 6 NF ) be the received signal on the kth subcarrier. Then,
Y1 = HX + N and Y2 , Y3 , ..., YNF have the same distribution as N. With the condition X = x,
|Y1 |2 is an exponential random variable with mean x2 + σN2 and |Yk6=1|2 are exponential random
variables with mean σN2 . Let G|Y2 |2 be the CDF of |Y2 |2 . Then, PF 6=F̂ |X=x (ρ) is calculated as
PF 6=F̂ |X=x (ρ) = Pr |Y1|2 < max |Y2 |2 , · · · , |YNF |2 |X = x
Z∞
F −1
= 1 − p|Y1 |2 (z)GN
|Y2 |2 (z)dz
0
n
o
Z∞ exp − 2 z 2 NX
F −1
− 12 (NF −k−1)z
x +σN
NF − 1
NF −k−1
σ
N
=1−
dz
(−1)
e
k
x2 + σN2
k=0
0
n
o
Z∞ exp − 2 z 2 − z2 (NF − k − 1)
NX
F −1
x +σN
σN
NF − 1
(−1)NF −k−1
=1−
dz
2
2
k
x
+
σ
N
k=0
0
=1−
=1−
NX
F −1
k=0
NX
F −1
k=0
NF − 1
(−1)NF −k−1
k
CNk F −1 (−1)NF −k−1
x2
2 (NF
σN
x2 ρ(NF
1
− k − 1) + NF − k
1
.
− k − 1) + NF − k
(7)
Since x follows a Gaussian distribution with zero mean and variance one, the PDF of x2 is the
z
1
e− 2 . Unconditioning the
Gamma distribution with degree of freedom one, i.e., pX 2 (z) = √2πz
5
frequency index detection error probability, PF 6=F̂ (ρ) can be obtained as
PF 6=F̂ (ρ) =
=
Z∞
−∞
Z∞
PF 6=F̂ |X=x (ρ)pX (x)dx
PF 6=F̂ |X=z (ρ)pX 2 (z)dz
0
=1−
∞
N
F −1 Z
X
k=0 0
CNk F −1 (−1)NF −k−1
1
√
e−z/2 dz.
zρ(NF − k − 1) + NF − k 2πz
(8)
Solving the integral [20], (6) can be obtained.
4.
Multi-Cell OFDM-IM
After deriving the single cell sum rate of OFDM-IM with Gaussian inputs in Section 3, multi-cell
OFDM-IM is investigated in this section. In practical system, finite alphabet inputs are usually
used instead of Gaussian inputs. Moreover, the impact of ICI needs to be studied. Therefore,
OFDM-IM with QAM inputs is assumed in this section and other types of constellations can be
obtained in a similar manner. The CDF of SINR, the PDF of noise plus ICI, and the multi-cell sum
rate will be derived.
4.1.
CDF of SINR
Since only one subcarrier of the target UE and the associated BS is active, the density of interfering
BSs to the target UE is one NF th of the original BS density. Let ρ̃ denote the SINR and PT denote
the transmit power of each BS. The derivation of the CDF of SINR is directly generalized from
[16]. The normalized interference power I from other BSs transmitting on the same subcarrier to
the target user can be computed as
X
|hl |2 |dl |−α .
(9)
I=
l∈S ′
The channel hξ between the target user and his associated BS follows Rayleigh distribution. Hence,
|hξ |2 is an exponentially distributed random variable. The CDF Gρ̃ (ρ̃) of SINR can then be computed as [16]
PT |hξ |2 d−α
Gρ̃ (ρ̃) = 1 − Pr
> ρ̃
2
σN
+ PT I
2
= 1 − exp −PT−1 dα ρ̃σN
E [exp (−dα ρ̃I)] .
(10)
Using the Laplace transform of the exponential function and the probability generating func-
6
tional [16], the expectation part in (10) can be expressed as
E [exp (−dα ρ̃I)] = E [exp (−zI)]|z=dα ρ̃
Z
λ
1
= exp −
dx
NF R2 1 + z −1 |y|α
|z=dα ρ
2
2π
λ
.
= exp − z 2/α
NF
α sin(2π/α) |z=dα ρ̃
(11)
Hence, the CDF Gρ̃ (ρ̃) of multi-cell OFDM-IM can be expressed as
λ 2 2
2π 2
−1 α
2
Gρ̃ (ρ̃) = 1 − exp −PT d ρ̃σN exp − d ρ̃ α
.
NF
α sin(2π/α)
(12)
To avoid 0 at the denominator in (12), α should be larger than 2. Typical values of α are between
2 to 4.
4.2.
PDF of Noise plus Interference
The exact PDF of noise plus ICI of multi-cell OFDM-IM with QAM inputs remained unanswered
in the literature. Generalized Gaussian distributions were used to approximate the PDF in [2] and
[3]. In this section, the exact PDF of noise plus ICI will be derived, showing that noise
P √plus ICI is
NB −1
MoG distributed. Let Q = 2
and denote ψ as the noise plus ICI, i.e., ψ = N +
Tl Hl Xl ζl .
l6=ξ
Theorem 2. The PDF pψ of noise plus ICI in multi-cell OFDM-IM with QAM inputs follows a
MoG distribution, i.e.,
pψ (z) =
Q
X
k=1
ωk N (z; 0, σk2 ).
(13)
Proof. In this proof, 4QAM is used for simplicity. The distribution of one ICI term , i.e. ψl =
hl Xl ζl , is first computed. It can be easily shown that hl Xl ∼ CN (0, 1). Also, the PDF of ζl for
4QAM can be expressed as pζl (ζl ) = N1F δ(ζl − 1) + NNF −1
δ(ζl ) which is directly obtained from the
F
fact that the subcarrier index is chosen uniformly. Next, according to the product distribution,
phl Xl ζl (z) =
Z∞
−∞
1
1
pζl (τ ) exp(−|z/τ |2 ) dτ
π
|τ |
1 1
NF − 1
|z|2
1
2
=
lim
exp(−|z| ) +
exp −
NF π
NF ǫ→0 πǫ
ǫ
NF − 1
1 1
exp(−|z|2 ) +
δ(z).
=
NF π
NF
(14)
Hence, the PDF of one interference term is a weighted sum of a Gaussian function and a Dirac delta
function. Using the properties of convolution, the PDF of the total interference can be expressed
7
as
ph1 X1 ζ1 (z) ∗ · · · ∗ phξ−1 Xξ−1 ζξ−1 (z) ∗ phξ+1 Xξ+1 ζξ+1 (z) ∗ · · · ∗ phNB XNB ζNB (z)
Q−1
=
X
k=1
ω̃k CN (z; 0, σ̃k2 ) + ω̃Q δ(z),
where ∗ denotes the convolution operator and ω̃k > 0 with
(15)
P
ω̃k = 1. By adding the noise term,
k
the PDF pψ of noise plus ICI can be calculated as
"Q−1
#
X
pψ (z) = CN (z; 0, σN2 ) ∗
ω̃k CN (z; 0, σ̃k2 ) + ω̃Q δ(z)
k=1
=
B −1
2N
X
k=1
where ωk > 0 with
P
ωk CN (z; 0, σk2 ),
(16)
ωk = 1.
k
Lemma 3. The PDF pY of the received signal Y in multi-cell OFDM-IM with QAM inputs follows
a MoG distribution, i.e.,
pY (y) =
Q
X
k=1
ωk′ CN (y; 0, σk′2).
(17)
Proof. This can be obtained directly from Theorem 2 and (1), because the desired signal also
follows a Gaussian distribution.
2
Next, parameters Ω = [ω1 ω2 · · · ωQ ] and v = σ12 σ22 · · · σQ
need to be estimated. It is well
known that the EM algorithm has been widely used to estimate parameters of MoG distributions.
Detailed derivations of EM algorithm for MoG parameter estimation are beyond the scope of this
paper. Interested readers can find details in [22]. In particular, for FQAM, the traditional EM
algorithm [22] can be further simplified because the means of the channel, the transmitted symbol,
the ICI, and noise are all zero. The simplified EM algorithm is presented in Fig. 2. Assuming
s
samples of the noise plus ICI are measured as {τa }N
a=1 . First, Ω and v are randomly chosen
as initialization. Second, E step and M step are operated iteratively until a certain convergence
condition is met. The number of Gaussian functions in (16) grows exponentially with the number
of BSs, which is impractical due to high complexity. However, only a small number of BSs are
dominating the total interference. In this case, a small number Q′ (Q′ << Q) of Gaussian functions
will be sufficient to approximatethe distribution of noise plus
of the PDF pY
ICI. The parameters
′
′2
of the received signal, i.e., Ω′ = ω1′ ω2′ · · · ωQ
and v = σ1′2 σ2′2 · · · σQ
, can also be estimated
via the same procedure.
4.3.
Multi-Cell Sum Rate
The multi-cell sum rate r3 of OFDM-IM is defined by the mutual information between the received
signal and the information source as
r3 = E [max I(X, F ; Y )] = E [H(Y )] − E [H(Y |X, F )] = H(Y ) − H(Y |X, F ),
8
(21)
1:
2:
Initialize Ω and v randomly
E step: Compute
ηak =
ωk CN (τa ; 0, σk2 )
Q
P
ωk CN (τa ; 0, σk2)
(18)
k=1
3:
M step: Update σk2,new and ωknew according to
σk2,new
Ns
1 X
ηak |τa |2
= N
Ps
ηak a=1
(19)
a=1
ωknew
4:
=
Ns
P
ηak
a=1
.
Ns
(20)
Repeat E step and M step until convergence condition is met.
Fig. 2. EM algorithm for parameter estimation of noise plus ICI in OFDM-IM systems.
where H() is the entropy function. The maximum operator in (21) is removed because of the QAM
symbol and the expectation operator is removed because the entropies of Y and N + I are already
expectation values. Although closed-form integral of (21) is not feasible because of the summation
inside the logarithm function, upper bounds of the sum rate can be evaluated.
Since the entropy function H() is concave, using Jensen’s inequality, H(Y |X, F ) is lower
bounded by
H(Y |X, F ) = H
Q
X
k=1
ωk CN (z; 0, σk2 )
!
>
Q
X
k=1
ωk H CN (z; 0, σk2 )
Q
1 X
= +
ωk log2 (2πeσk ).
2 k=1
(22)
On the contrary, H(Y ) is upper bounded by [24]
!
Z X
Q
Q
X
′
′2
ωk′ CN (y; 0, σk′2) log2 ωk′ CN (y; 0, σk′2) dy
ωk CN (y; 0, σk ) 6 −
H(Y ) = H
k=1
=
1
+
2
k=1
Q
X
k=1
9
ωk′ log2 (2πeσk′ /ωk′ ).
(23)
Subcarrier index detection error probability
100
10-1
Analytic results (NF = 8)
Simulated results (NF = 8)
Analytic results (NF = 4)
Simulated results (NF = 4)
Analytic results (NF = 2)
Simulated results (NF = 2)
10-2
0
5
10
15
20
25
30
35
40
SNR (dB)
Fig. 3. Subcarrier index detection error probability of single cell OFDM-IM.
To sum up, the multi-cell sum rate of OFDM-IM r3 is upper bounded by
r3 6
Q
X
ωk′ ′
log2 (2πeσk′ ′ /ωk′ ′ )
k ′ =1
−
Q
X
ωk log2 (2πeσk ).
(24)
k=1
This upper bound provides insights in sum rate performance of multi-cell OFDM-IM.
5.
Results and Analysis
Subcarrier index detection error probability and achievable rates of single cell OFDM-IM with different number of subcarriers are depicted in Fig. 3 and Fig. 4, respectively. In both figures, analytic
results match simulated results well. It can be observed in Fig. 4 that when the SNR is relatively
low, the achievable rate contributed by subcarrier indexes is not significant. However, when SNR
increases gradually, the gaps between achievable rates with different numbers of subcarriers first
increase and then become stable.
Fig. 5 shows single cell OFDM-IM achievable rate contributed by subcarrier indexes with
different numbers of subcarriers and different SNR values. From (5), r2 tends to zero when NF
tends to infinity. This suggests that the benefit of increasing NF is diminishing and there is an
optimal number of NF such that r2 reaches its peak. Closed-from expression of the optimal value
of NF is not available. However, this can be calculated by numerical results. It is shown in Fig. 5
that the optimal value of NF increases with the SNR.
In multi-cell simulations, thermal noise with power density −173 dBm/Hz [23] is assumed in
the multi-cell network. Also, the bandwidth of each subcarrier is 15 kHz [23]. In this case, the
noise power per subcarrier can be calculated by σN2 = 7.5 × 10−11 W.
10
18
Simulated results (NF = 4)
Analytic results (NF = 4)
Simulated results (NF = 8)
Analytic results (NF = 8)
Simulated results (NF = 16)
Analytic results (NF = 16)
16
Achievable rate (bps/Hz)
14
12
10
8
6
4
2
0
0
5
10
15
20
25
30
35
40
SNR (dB)
Fig. 4. Achievable rates of single cell OFDM-IM with different number of subcarriers.
0.6
SNR=-5 dB
SNR=0 dB
SNR=3 dB
0.5
Optimal NF = 1024
r2 (bps/Hz)
0.4
Optimal NF = 512
0.3
0.2
Optimal NF = 32
0.1
0
1
2
3
4
5
6
7
8
9
10
11
12
log2 (NF )
Fig. 5. The achievable rates contributed by different numbers of subcarriers and the optimal NF .
11
1
Simulated results, NF = 2
Analytic results, NF = 2
Simulated results, NF = 4
Analytic results, NF = 4
0.9
0.8
0.7
CDF
0.6
0.5
0.4
0.3
0.2
0.1
0
-30
-20
-10
0
10
20
30
SINR (dB)
Fig. 6. Comparison of CDFs of SINR with respect to different numbers of subcarriers of multi-cell
OFDM-IM (λ = 10−4 , α = 3, PT = 40W, d = 50m).
Fig. 6 illustrates the CDFs of SINR in terms of different values of NF in the multi-cell OFDMIM scenario. The SINR is larger when NF is larger, because the target UE has a smaller probability
of being interfered, where ICI is more spreaded in the frequency domain. In addition, the simulated
results reasonably well align with analytic results.
The PDF of the real part of noise plus ICI of multi-cell OFDM-IM with 4QAM inputs is illustrated in Fig. 7. The total number of BSs is 19 and the inter site distance (ISD) is 100m. It
be can observed that the MoG distribution derived in this paper fits simulation excellently. The
generalized Gaussian model proposed in [2] [3] and the Gaussian model are also shown in Fig.
7. However, these two models fail to match the realistic noise plus ICI well. The generalized
Gaussian model is able to capture the peak while it does not model the spread well. On the other
hand, the Gaussian model has better alignment with the PDF spread than the generalized Gaussian
model, but it is not accurate to model the peak.
The upper bound of sum rates of multi-cell OFDM-IM with 4QAM is depicted in Fig. 8. Single
cell achievable rates with Gaussian input and 4QAM are included as reference. It can be observed
that the sum rate of multi-cell OFDM-IM is at least approximately 20% worse than single cell
results because of ICI. When the SNR is smaller, although the multi-cell result outperforms the
other two single cell results, this is caused by the upper limit.
6.
Conclusions
In this paper, single cell achievable rate and multi-cell statistical properties of SINR, ICI, and
sum rate for OFDM-IM have been studied. It has been shown that the increase of the number of
subcarriers does not improve the single cell achievable rate in the low SNR regime. The main
12
50
Simulation
MoG model
Gaussian model
Generalized Gaussian model
Probability density (dB)
45
40
35
30
25
20
15
-1
0
1
Real (N + I)
×10-4
Fig. 7. PDF of the real part of noise plus ICI of multi-cell OFDM-IM with 4QAM inputs (NB = 19,
α = 3, PT = 40W, d = 50m, Q′ = 4).
8
Single cell with Gaussian input
Multi cell with 4QAM (upper bound)
Single cell with 4QAM
7
Sum rate (bps/Hz)
6
5
4
3
2
1
0
0
5
10
15
20
25
SNR (dB)
Fig. 8. Upper bound of sum rates of multi-cell OFDM-IM (NB = 19, NF = 4).
13
30
benefit of more subcarriers appears in multi-cell scenarios, where less BSs will be interfering the
target UE. In addition, the PDF of noise plus ICI has been derived for OFDM-IM with QAM input,
showing that noise plus ICI follows a MoG distribution. The parameters of the MoG distribution
have been estimated by a simplified EM algorithm in this paper. Later, upper bound of sum rate of
multi-cell OFDM-IM has been derived, which can be used as performance guideline for OFDMIM networks. For future work, it will be practical to use the PDF of noise plus ICI to develop
multi-cell OFDM-IM signal detection algorithms. Also, the analysis method in this paper can be
extended to NF -ary asymmetric channel to investigate achievable rates of generalized OFDM-IMs.
7.
Acknowledgement
This work has been performed in the framework of the Horizon 2020 project FANTASTIC-5G
(ICT-671660) receiving funds from the European Union. The authors would like to acknowledge
the contributions of their colleagues in the project, although the views expressed in this contribution
are those of the authors and do not necessarily represent the project.
8.
References
[1] Abu-adhiga, R. and Haas, H.: “Subcarrier-index modulation OFDM,” in Proc. PIMRC’09,
Tokyo, Japan, Sep. 2009, pp. 177–181.
[2] Hong, S., Sagong, M., Lim, C., et al: “Frequency and quadrature-amplitude modulation for
downlink cellular OFDMA networks,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, June 2014,
pp. 1256–1267.
[3] Seol, C. and Cheun, K.: “A statistical inter-cell interference model for downlink cellular
OFDMA networks under log-normal shadowing and multipath Rayleigh fading,” IEEE Trans.
Commun., vol. 57, no. 10, Oct. 2009, pp. 3069–3077.
[4] Mesleh, R. Y., Haas, H., Sinanovic, S., et al: “Spatial modulation,” IEEE Trans. Veh. Technol.,
vol. 57, no. 4, July 2008, pp. 2228–2241.
[5] Tsonev, D., Sinanovic, S., and Haas, H.; ’Enhanced subcarrier index modulation (SIM)
OFDM’. Proc. Globecom’11, Huston, USA, Dec. 2011, pp. 1–5.
[6] Basar, E., Aygolu, U., Panayirci, E., et al: “Orthogonal frequency division multiplexing with
index modulation,” IEEE Trans. Signal Process., vol. 61, no. 22, Nov. 2013, pp. 5536–5549.
[7] Zheng, B., Chen, F., Wen, M., et al: “Low-complexity ML detector and performance analysis
for OFDM with in-phase/quadrature index modulation,” IEEE Commun. Lett., vol. 19, no. 11,
Nov. 2015, pp. 1893–1896.
[8] Xiao, Y., Wang, S., Dan, et al: “OFDM with interleaved subcarrier-index modulation,” IEEE
Commun. Lett., vol. 18, no. 8, Aug. 2014, pp. 1447–1450.
[9] Datta, T., Eshwaraiah, H. S., and Chockalingam, A.: “Generalized space and frequency index
modulation,” IEEE Trans. Veh. Technol., vol. 65, no. 7, pp. 4911–4924, July 2016.
14
[10] Fan, R., Yu, Y. J., and Guan, Y. L.: “Generalization of orthogonal frequency division multiplexing with index modulation,” IEEE Trans. Wireless Commun., vol. 14, no. 10, Oct. 2015,
pp. 5350–5359.
[11] Wen, M., Cheng, X., and Ma, M.: “On the achievable rate of OFDM with index modulation,”
IEEE Trans. Signal Process., vol. 64, no. 8, Dec. 2015, pp. 1919–1932.
[12] Wen, M., Cheng, X., Yang, L., et al: “Index modulated OFDM for underwater acoustic
communications,” IEEE Comm. Mag., vol. 54, no. 5, May. 2016, pp. 132–137.
[13] Ishikawa, N., Sugiura, S., and Hanzo, L.: “Subcarrier-index modulation aided OFDM - will
it work?,” IEEE Access, vol. 4, 2016, pp. 2580–2593.
[14] Baccelli, F. and Blaszczyszyn, B.: Stochastic Geometry and Wireless Networks Volume I:
Theory., Foundations and Trends R in Networking: vol. 3: no. 3–4, 2010, pp. 249–449.
[15] Baccelli, F., and Blaszczyszyn, B.: Stochastic Geometry and Wireless Networks Volume II:
Applications., Foundations and Trends R in Networking: vol. 4: no. 1–2, 2010, pp. 1–312.
[16] Baccelli, F., Blaszczyszyn, B., and Muhlethaler, P.: “An Aloha protocol for multihop mobile
wireless networks,” IEEE Trans. Inf. Theory., vol. 52, no. 2, Feb. 2006, pp. 421–436.
[17] ElSawy, H., Hossain, E., and Haenggi, M.: “Stochastic geometry for modeling, analysis,
and design of multi-tier and cognitive cellular wireless networks: a survey,” IEEE Commun.
Survey & Tutorials, vol. 15, no. 3, Third quarter. 2013, pp. 996–1019.
[18] Andrews, J. G., Baccelli, F., and Ganti, R. K.: “A tractable approach to coverage and rate in
cellular networks,” IEEE Trans. Commun., vol. 59, no. 11, Nov. 2011, pp. 3122–3134.
[19] Baccelli, F. and Giovanidis, A.: “A stochastic geometry framework for analyzing pairwisecooperative cellular networks,” IEEE Trans. Wireless Commun., vol. 14, no. 2, Feb. 2015, pp.
794–808.
[20] Gradshteyn, I. S. and Ryzhik, I. M.: Table of Integrals, Series, and Products., 7th ed, Academic Press, Burlington, 2007.
[21] Weidmann, C. and Lechner, G.: “A fresh look at coding for q-ary symmetric channels,” IEEE
Trans. Inf. Theory., vol. 58, no. 11, Nov. 2012, pp. 6959–6967.
[22] Bishop, C. M.: Pattern Recognition and Machine Learning., Springer, New York, 2007.
[23] 3GPP T. R. 36.814, Further advancements for E-UTRA physical layer aspects., V1.7.0, Feb.
2010.
[24] Huber, M. F., Bailey, T. Durrant-Whyte, H., et al; ’On entropy approximation for Gaussian
mixture random variables’. Proc. MFI’08, Seoul, Korea, Aug. 2008, pp. 181–188.
15
| 7cs.IT
|
A New Model for Evaluating Range-Based Anomaly Detection Algorithms
Nesime Tatbul 1 2 Tae Jun Lee 3 * Stan Zdonik 4 Justin Gottschlich 1
arXiv:1803.03639v1 [cs.LG] 8 Mar 2018
Abstract
Classical anomaly detection (AD) is principally
concerned with point-based anomalies, anomalies
that occur at a single point in time. While pointbased anomalies are useful, many real-world
anomalies are range-based, meaning they occur
over a period of time. Therefore, applying classical point-based accuracy measures to range-based
AD systems can be misleading. In this paper, we
present a new mathematical model that more accurately gauges the classification correctness of
AD systems for range-based anomalies. Unlike
prior work, our mathematical definitions are a superset of the classical AD definitions, enabling
our system to also subsume point-based anomalies. Moreover, our system is broadly generalizable and provides a number of specialization
functions that can control the application’s bias
along a multi-dimensional axis.
1. Introduction
Anomaly detection (AD) is the process of identifying abnormal events. Anomalies are in general domain- or problemspecific. For example, in medical applications, one might
wish to represent each unique disease or illness as its own
anomaly. In the cyber-security space, each individually
known vulnerability might be classified as its own anomaly
as a means to track frequency, or grouped together based
on commonality, rarity, and emergency. In parallel software
correctness, one might wish to represent each type of race
condition as its own anomaly.
Although the problems in each of the prior examples are
different, they all have one thing in common. Each of their
anomalies are events that occur over a period of time rather
than at a fixed point. Disease and illness manifest gradually (Kourou et al., 2015). Cyber-attacks are almost always
multi-step processes (AlEroud & Karabatis, 2012). Soft1
Intel Labs, Santa Clara, CA, USA 2 MIT, Cambridge,
MA, USA 3 Microsoft, Seattle, WA, USA 4 Brown University,
Providence, RI, USA. ∗ The work was done while a Brown student.
Correspondence to: Nesime Tatbul <[email protected]>.
Copyright 2018 by the author(s).
ware correctness and performance bugs generally emerge
from the execution of several serial operations in a precise
sequence (Alam & Muzahid, 2016). It is therefore critical
that the fundamental accuracy measurements for anomalies,
and the systems that detect them, reason about them as they
occur over a period of time. We call such events rangebased anomalies, which are a subset of collective anomalies
(Chandola et al., 2009). Unfortunately, the classical metrics
for anomaly detection, which are widely used today, were
designed to handle only single-point anomalies (Aggarwal,
2013). The core of this paper addresses this problem.
An AD algorithm is essentially a form of binary classification. It recognizes certain patterns in its input and classifies
them as either normal or not. For this general class of algorithms, measurements of Recall and Precision are widely
used for evaluating the accuracy of the results. They are formally defined as follows (where TP , FP , FN are number of
true positives, false positives, false negatives, respectively):
Recall = T P ÷ (T P + F N )
P recision = T P ÷ (T P + F P )
(1)
(2)
Informally, Recall is the rate at which a system can identify
real anomalies without excluding any, while Precision is
the rate a system can identify anomalies without including
any non-anomalous ones. In this sense, Recall and Precision are complementary. This characterization proves useful
when they are combined, such as in F1 score, as such combinations help gauge the quality of both anomalous and
non-anomalous predictions.
While useful for point-based anomalies, classical recall and
precision suffer from the inability to capture and bias the
classification of domain-specific time series anomalies. This
is beginning to have profound negative side-effects on the
advancement of both practical and research AD systems. In
particular, many time series AD systems’ accuracy are being
misrepresented, because point-based recall and precision
are being used to measure their effectiveness for rangebased anomalies. Moreover, the need to accurately identify
time series anomalies is growing in importance due to the
explosion of streaming and real-time systems (Anava et al.,
2015; Malhotra et al., 2015; Twitter, 2015; Guha et al., 2016;
Yu et al., 2016; Ahmad et al., 2017).
To address this, we redefine recall and precision to encompass range-based anomalies. Unlike prior work (Lavin &
A New Model for Evaluating Range-Based Anomaly Detection Algorithms
(a) Precision = 0.6, Recall = 0.5
(b) Precision = ?, Recall = ?
Figure 1: Point-based vs. Range-based Anomalies
Ahmad, 2015; Ahmad et al., 2017), our mathematical definitions are a superset of the classical point-based definitions,
enabling our system to subsume point-based anomalies. Further, our system is broadly generalizable by providing specialization functions that can control a domain’s bias along a
multi-dimensional axis that is necessary to properly accommodate the needs of a specific domain or problem space.
The key contribution of this paper is a new mathematical
model, which can be used to evaluate, rank, and compare
results of AD algorithms. Although outside of the scope
of this paper, our model can also be used as the objective
function for machine learning (ML) training, which may
have a profound impact on AD training strategies, giving
rise to fundamentally new ML techniques in the future.
In what follows, we first set the context by detailing the
problem and giving an overview of prior work. Then in
Sections 4 and 5, we formally present our new model. Finally, Section 6 provides an experimental study of our new
model in comparison to the classical model as well as a
recent scoring model provided by the Numenta Anomaly
Benchmark (Lavin & Ahmad, 2015).
2. Problem Motivation and Design Goals
Classical recall and precision are defined for sets of independent points (Figure 1a), which is sufficient for point-based
anomalies. Time series AD algorithms, on the other hand,
work with sequences of time intervals (Figure 1b). Unfortunately, due to the design of the classical model, important
time series specific characteristics cannot be captured by it.
Many of these distinctions are due to partial overlaps. In
the point-based case, a predicted anomaly point Pi is either
a member of the set of real anomalies (a TP in Figure 1a)
or not (an FP in Figure 1a). In the range-based case, it
may happen that a predicted anomaly range might partially
overlap with a real one. In this case, a single prediction
range is partially a TP and partially an FP at the same time.
The size of this partial overlap needs to be quantified, but
there may be additional criteria. For example, one may
want to consider the position of the overlap. After all, a
range consists of an ordered collection of points and the
order might be meaningful for the application. For instance,
detecting the earlier portion of an anomaly (i.e., its “frontend”) might be more critical for a real-time application to
reduce the time to react to it. Furthermore, overlaps are
no longer 1-to-1 as in the classical model. One or more
predicted anomaly ranges may (partially) overlap with one
or more real ones. Figure 1b illustrates one such situation.
The specific domain might care if each independent anomaly
range is detected as a single unit or not. Thus, we may want
to also capture cardinality when measuring overlap.
If all anomalies were point-based, or if there were no partialoverlap situations, then we could reuse the classical model,
by simply treating each range in a sequence like a point in a
set. For general range-based anomalies, the classical model
falls short. Motivated by these observations, we propose a
new model that meets the following design goals:
• Expressive: captures criteria that are unique to rangebased anomalies such as overlap position and cardinality.
• Flexible: supports adjustable weights across multiple
criteria for domain-specific needs.
• Extensible: supports inclusion of additional domainspecific criteria that cannot be known a priori.
Intuitively, our model should enable ranking the results of
time series AD algorithms in a way that is as close to human
judgment as possible.
3. Related Work
There is a growing body of research emphasizing the importance of time series AD. Bailis et al.’s work focuses on prioritization techniques when analyzing fast, streaming data
including time series outlier detection (Bailis et al., 2017).
Malhotra et al. and Guha et al. have proposed new ML
techniques to handle time series anomalies, ranging from
space shuttles to Amazon web services (Malhotra et al.,
2015; Guha et al., 2016). Lipton et al. have investigated
the viability of existing ML-based AD techniques on medical time series data (Lipton et al., 2016), while Patcha and
Park have drawn out weaknesses for AD systems to properly handle cyber-attacks in a time series setting, amongst
others (Patcha & Park, 2007).
Despite these advances, the techniques designed to judge
the efficacy of time series AD systems are insufficient. They
largely rely on classical recall and precision, which are
severely limited when applied to range-based anomalies.
This is a notable shortcoming in the measurement, and even
training, of future-generation ML models and algorithms.
A New Model for Evaluating Range-Based Anomaly Detection Algorithms
We are not the first to make this observation. Lavin and
Ahmad have also discussed the lack of proper time series
measurement techniques for AD in their Numenta Anomaly
Benchmark (NAB) (Lavin & Ahmad, 2015). Similar to our
findings, they note the limitations of classical recall and
precision in accurately measuring AD systems. Focusing
specifically on real-time streaming applications, they state
that: (i) only early anomaly detection matters (i.e., frontend bias) and (ii) only the first point of the anomaly range
matters (i.e., an emphasis of precision over recall and a
bias toward single-point true positive prediction systems).
They then propose an anomaly scoring system based on
anomaly windows, application profiles, and a sigmoidal
scoring function, which is specifically designed to reward
detections earlier in a window.
While such an approach may be suitable for their restricted
domain, it is not generally applicable to range-based anomalies. For example, in some time series scenarios such as the
identification of medical anomalies (e.g., cancer detection),
it is critical to capture when an illness is recessing due to the
use of life threatening medical treatments used to address
such illnesses (Formenti & Demaria, 2017). Under Lavin
and Ahmad’s single-point reward system, there is no clear
distinction for such state changes, because the single-point
prediction algorithm and the scoring system that they propose only indicate the presence of an anomalous state, not
the beginning or ending of it. Furthermore, as discussed in
detail in a follow-up analysis by Singh and Olinsky (Singh
& Olinsky, 2017), the NAB scoring system has a number of
limitations that makes it challenging to use in real-world applications, even within its restricted domain (e.g., the need
to determine window size in advance, ambiguities in scoring
function, and magic numbers).
Unlike Lavin and Ahmad’s approach, our accuracy model
does not have a fixed disposition to a single domain. Rather,
it is general and flexible enough to support multiple application scenarios, including the classical recall/precision
measures and Numenta’s front-end bias scoring system,
among potentially many others, as we demonstrate with
experiments in Section 6.
4. Range-based Recall
The purpose of the Recall measurement is to reward a prediction system when anomalies are successfully identified (i.e.,
TP) and to penalize it when they are not (i.e., FN). In the
classical point-based model, Recall is computed by counting
the number of real anomaly points that are predicted by an
AD system, which is then divided by the total number of real
anomaly points. As such, this model is not sensitive to real
anomalies that may be organized as a range of contiguous
points but collectively indicate a single anomalous event. In
this section, we propose a new way to compute the recall
score for such range-based anomalies. Table 1 provides a
Notation
R
Ri
P
Pj
N
Nr
Np
α
β
γ()
ω()
δ()
Table 1: Notation
Description
set of real anomaly ranges
the ith real anomaly range
set of predicted anomaly ranges
the j th predicted anomaly range
total number of all points
number of real anomaly ranges
number of predicted anomaly ranges
relative weight of existence reward
relative weight of overlap reward
overlap cardinality function
overlap size function
positional bias function
summary of the notation used in our equations.
Given a set of real anomaly ranges R = {R1 , .., RNr } and
a set of predicted anomaly ranges P = {P1 , .., PNp }, our
RecallT (R, P ) formulation iterates over the set of all real
anomaly ranges (R), computing a recall score for each real
anomaly range (Ri ∈ R) and adding them up into a total
recall score. This total score is then divided by the total
number of real anomalies (Nr ) to obtain an average recall
score for the whole time series.
PNr
RecallT (R, P ) =
i=1
RecallT (Ri , P )
Nr
(3)
When computing the recall score RecallT (Ri , P ) for a single real anomaly range Ri , we consider the following:
• Existence: Catching the existence of an anomaly (even
by predicting only a single point in Ri ) by itself might be
valuable for the application.
• Size: The larger the size of the correctly predicted portion
of Ri , the higher the recall score will likely be.
• Position: In some cases, not only size, but also the relative
position of the correctly predicted portion of Ri might
matter to the application.
• Cardinality: Detecting Ri with a single prediction range
Pj ∈ P may be more valuable than doing so with multiple different ranges in P .
We capture all of these points as a sum of two main reward
terms weighted by α and β, respectively, where 0 ≤ α, β ≤
1 and α + β = 1. α represents the relative importance
of rewarding existence, whereas β represents the relative
importance of rewarding size, position, and cardinality, all
of which stem from the actual overlap between Ri and the
set of all predicted anomaly ranges (Pi ∈ P ).
RecallT (Ri , P ) = α × ExistenceReward(Ri , P )
+ β × OverlapReward(Ri , P )
(4)
If anomaly range Ri is caught (i.e., |Ri ∩ Pj | ≥ 1 across
all Pj ∈ P ), then an existence reward of 1 is earned.
A New Model for Evaluating Range-Based Anomaly Detection Algorithms
function δ(i, AnomalyLength)
return 1
function δ(i, AnomalyLength)
return AnomalyLength - i + 1
function δ(i, AnomalyLength)
return i
function δ(i, AnomalyLength)
if i ≤ AnomalyLength/2 then
return i
else
return AnomalyLength - i + 1
function ω(AnomalyRange, OverlapSet, δ)
MyValue ← 0
MaxValue ← 0
AnomalyLength ← length(AnomalyRange)
for i ← 1, AnomalyLength do
Bias ← δ(i, AnomalyLength)
MaxValue ← MaxValue + Bias
if AnomalyRange[i] in OverlapSet then
MyValue ← MyValue + Bias
return MyValue/MaxValue
(a) Overlap Size
. Flat bias
. Front-end bias
. Back-end bias
. Middle bias
(b) Positional Bias
Figure 2: Example Functions for ω() and δ() in Range-based Recall
(
ExistenceReward(Ri , P ) =
P Np
1, if j=1 |Ri ∩ Pj | ≥ 1
(5)
0, otherwise
Additionally, a cumulative overlap reward that depends
on three application-defined functions 0 ≤ γ() ≤ 1,
0 ≤ ω() ≤ 1, and δ() ≥ 1 is earned. These capture the cardinality (γ), size (ω), and position (δ) aspects of the overlap.
More specifically, the cardinality term serves as a scaling
factor for the rewards earned from overlap size and position.
OverlapReward(Ri , P ) = CardinalityF actor(Ri , P )
×
Np
X
ω(Ri , Ri ∩ Pj , δ)
(6)
cancer detection scenario or real-time anomaly detection),
then the front-end bias function should be used instead. If
later ones are more important (e.g., missile defense alarm
system scenario), then the back-end bias function should
be used. Finally, the middle bias function is for prioritizing
mid-portions of anomalies.
Our recall formulation for range-based anomalies subsumes
the classical one for point-based anomalies, i.e., RecallT ≡
Recall when:
(i) all Ri ∈ R and Pj ∈ P are represented as unit-size
ranges (e.g., range [1, 3] as [1, 1], [2, 2], [3, 3]), and
(ii) α = 0, β = 1, γ() = 1, ω() is as in Figure 2a, and δ()
returns flat positional bias as in Figure 2b.
j=1
The cardinality factor gets the largest value of 1 when Ri
overlaps with at most one predicted anomaly range (i.e., it
is identified by a single prediction range). Otherwise, it
receives a value 0 ≤ γ() ≤ 1 defined by the application.
CardinalityF actor(Ri , P ) =
1
, if Ri overlaps
with at most
(7)
one Pj ∈ P
γ(Ri , P ), otherwise
It is important to note that the RecallT constants (α and β)
and functions (γ(), ω(), and δ()) are tunable according to
the needs of the application. Below, we provide an example
for these functions to illustrate how they can be customized.
Intuitively, the cardinality factor should be inversely proportional to the number of distinct prediction ranges that a
real anomaly range Ri overlaps. Let us denote this number by Card (Ri ). A typical example for γ() could then be
1 /Card (Ri ).
Figure 2 provides an example for the ω() function for size,
which can be used with many different δ() functions, four of
which we provide, for positional bias. If all index positions
are equally important for the application, then the flat bias
function should be used. In this case, the larger the size
of the overlap, the higher the reward will be. If earlier
index positions are more important than later ones (e.g.,
5. Range-based Precision
In the classical point-based accuracy model, the precision
score is computed by counting the number of successful prediction points (i.e., TP) in proportion to the total number of
prediction points (i.e., TP+FP). As such, the key difference
between Precision and Recall is that Precision penalizes
FPs instead of FNs. In this section, we extend the classical
precision formulation to range-based anomalies over timeseries. Our formulation follows a similar structure as the
one for RecallT presented in the previous section.
Given a set of real anomaly ranges R = {R1 , .., RNr } and
a set of predicted anomaly ranges P = {P1 , .., PNp }, our
P recisionT (R, P ) formulation iterates over the set of all
predicted anomaly ranges (P ), computing a precision score
for each predicted anomaly range (Pi ∈ P ) and adding
them up into a total precision score. This total score is then
divided by the total number of predicted anomalies (Np ) to
obtain an average precision score for the whole time-series.
PNp
P recisionT (R, P ) =
i=1
P recisionT (R, Pi )
Np
(8)
When computing the precision score P recisionT (R, Pi )
for a single predicted anomaly range Pi , there is no need
for an existence reward, since precision by definition emphasizes prediction quality, and existence by itself is too
A New Model for Evaluating Range-Based Anomaly Detection Algorithms
low a bar for judging the quality of a prediction. On the
other hand, we still need the overlap reward to capture the
cardinality, size, and position aspects of a prediction.
P recisionT (R, Pi ) = CardinalityF actor(Pi , R)
×
Nr
X
ω(Pi , Pi ∩ Rj , δ)
(9)
j=1
Like in our recall formulation, the γ(), ω(), and δ() functions are tunable according to application semantics. Furthermore, P recisionT ≡ P recision under the same settings provided in the previous section (except the α and β
constants, which are not used in P recisionT ).
It is important to notice that the positional bias function (δ())
used in P recisionT refers to relative index positions in a
predicted anomaly range, not in a real anomaly range (unlike
RecallT ). Furthermore, while δ() provides a potential knob
for positional bias, we believe that in many domains a flat
bias function will suffice for P recisionT , as an F P is
typically considered uniformly bad wherever it appears in a
prediction range.
6. Experimental Study
In this section, we present an experimental study of our
range-based model applied to AD results of a state-of-the-art
ML system over a collection of diverse time series datasets.
We aim to demonstrate two key features of our model: (i) its
expressive power for range-based anomalies compared to
the classical model, (ii) its flexibility in supporting diverse
application semantics in comparison to the Numenta scoring
model (Lavin & Ahmad, 2015). Furthermore, we provide
a cost analysis for computing our new metrics, which is
important for their practical application.
6.1. Experimental Setup
System: All experiments were run on a Windows 10 maTM
chine with an Intel R Core i5-6300HQ processor running
at 2.30 GHz with 4 cores and 8 GB of RAM.
Datasets: We used a mixture of real and synthetic datasets.
The real datasets are taken from the NAB Data Corpus
(Numenta, 2017), whereas the synthetic datasets are generated by the Paranom tool (Gottschlich, 2018). All data
is time-ordered, univariate, numeric time series, for which
anomalous points/ranges (i.e., “the ground truth”) are already known.
NYC-Taxi: A real dataset collected by the NYC Taxi and
Limousine Commission. The data represent the number
of passengers over time recorded as 30-minute aggregates.
There are five anomalies: the NYC Marathon, Thanksgiving,
Christmas, New Year’s day, and a snow storm.
Twitter-AAPL: A real dataset with a time series of the number of tweets that mention Apple’s ticker symbol (AAPL),
recorded as 5-minute aggregates.
Machine-Temp: A real dataset with readings from a temperature sensor attached to a large industrial machine. There
are three anomalies: a planned shutdown, an unidentified
error, and a catastrophic failure.
ECG: A dataset from Paranom based on a real Electrocardiogram (ECG) dataset (Malhotra et al., 2015; Keogh, 2005);
augmented to include additional synthetic anomalies over
the original single pre-ventricular contraction anomaly.
Space-Shuttle: A dataset based on sensor measurements
from valves on a NASA space shuttle. It is also generated
by Paranom based on a real dataset from the literature with
multiple additional synthetic anomalies (Malhotra et al.,
2015; Keogh, 2005).
Sine: A dataset from Paranom that includes a sine wave
oscillating between 0.2 and 0.5 with a complete period over
360 timestamps. It contains many stochastic anomalous
events ranging from 50 − 100 time intervals.
Time-Guided: A Paranom dataset that monotonically increases a univariate value loosely associated with the time
interval. Multiple ranged-based stochastic anomalies appear
with inverted negative values.
For training and testing, we carefully partition each of the
datasets to ensure each anomaly range is kept intact in spite
of the segmentation.
Anomaly Detection Algorithm: We trained a TensorFlowimplemented version of LSTM-AD, a long short-term memory AD model (Malhotra et al., 2015), across all datasets.
We interpret adjacent anomalous points as part of a single
anomaly range. Thus, the LSTM-AD testing phase output is
a sequence of predicted anomaly ranges for a given dataset.
Compared Approaches: We evaluated the prediction accuracy of each output from our LSTM-AD runs using three
models: classical point-based precision/recall, the Numenta
scoring model (Lavin & Ahmad, 2015) (see Section 3),
and our range-based precision/recall model. Unless otherwise stated, we use the following default parameter settings for computing our RecallT and PrecisionT equations:
α = 0, β = 1, γ() = 1, ω() is as in Figure 2a, and δ()
returns flat positional bias as in Figure 2b.
6.2. Comparison to the Classical Point-Based Model
First, we compare our range-based model with the classical
point-based model. The goal of this experiment is twofold:
(i) verify that our model subsumes the classical one, (ii)
show that our model can capture additional variations for
anomaly ranges that the classical one cannot.
We present our results in Figure 3 with three bar
charts, which show Recall, Precision, and F1
Score values for our LSTM-AD testing over the
A New Model for Evaluating Range-Based Anomaly Detection Algorithms
each metric. Note that γ() is set to 1/Card(Ri ).
Recall: In Figure 3a, we provide RecallT measurements
for four positional biases: flat, front, back, and middle (see
Figure 2b). Recall is perfect (= 1.0) in all runs for the
Machine-Temp dataset. When we analyzed the real and
predicted anomaly ranges for this dataset, we see two real
anomaly ranges for which LSTM-AD predicts as a single
anomaly. Both real anomalies are fully predicted (i.e., no
false negatives and Card(Ri ) = 1 for both ranges). Therefore, recall has the largest possible value independent from
what δ() is, and is the expected behavior for this scenario.
(a) Recall
For all remaining datasets except Time-Guided, RecallT
has a smaller value than Recall , which we expected for the
following reasons. Real anomaly ranges are rarely captured
(i) entirely by LSTM-AD (i.e., overlap reward is partially
earned) and (ii) by exactly one predicted anomaly range
(i.e., cardinality factor downscales the overlap reward). Analyzing the real and predicted anomalies for Time-Guided reveals that LSTM-AD predicts 7/12 ranges fully, 2/12 ranges
half-way, and none of the remaining 3/12. As such, differences amongst RecallT biases are mainly due to the two
half-predicts. Both half-predicts lie at back-ends, which
explains why Recall T Back is larger than other biases.
This demonstrates the positional bias sensitivity of RecallT ,
something that cannot be captured by the classical Recall .
(b) Precision
(c) F1 Score
Figure 3: Our Model vs. the Classical Point-based Model
seven datasets, respectively.
Recall values are
computed using Equation 1 for Recall (bar labeled
Recall Classical) and Equation 3 for RecallT (bars
labeled Recall T *). Similarly, Precision values
are computed using Equation 2 for P recision (bar labeled Precision Classical) and Equation 8 for
P recisionT (bars labeled Precision T *). Finally, F1
Score values are computed using the following wellknown F-score equation, which essentially represents the
harmonic mean of Precision and Recall:
F1 = 2 ×
P recision × Recall
P recision + Recall
(10)
The first observation we make is that in all three graphs the
first two bars are equal for all datasets. This demonstrates
that our model, when properly configured as described in the
last paragraph of Section 4, subsumes the classical model.
Secondly, we observe the impact of positional bias (δ()) on
We make similar observations for ECG, Space-Shuttle, and
Sine. That is, different positional bias leads to different RecallT values. These demonstrate the sensitivity of
RecallT to the positions of correctly predicted portions of
anomalies. For NYC-Taxi and Twitter-AAPL, the differences
are not as pronounced. Data indicates that these datasets
contains few real anomaly ranges (2 and 3, respectively),
but LSTM-AD predicts a larger number of small anomaly
ranges. As such, overlaps are small and highly fragmented,
likely making δ() less dominant than ω() and γ().
Precision: In Figure 3b, we present PrecisionT only with
flat positional bias, as we expect this to be the typical use
case (see Section 5). PrecisionT is typically smaller than
Precision in all but one dataset (NYC-Taxi). When reviewing the predictions for NYC-Taxi, we found many narrowrange predictions, with a majority of them being false positives. Because Np < N dominates the rewards earned
for true positives, the value of PrecisionT is larger than
Precision for this dataset. Furthermore, all precision values are similar for Sine and Time-Guided. For these two
datasets, both real and predicted anomaly ranges are narrow,
diminishing the differences between points and ranges.
Precision Classical and Precision T flat
follow a similar pattern, except the latter ranks higher on
Sine than Space-Shuttle. Comparing these two runs, we notice that Space-Shuttle contains a larger number of predicted
A New Model for Evaluating Range-Based Anomaly Detection Algorithms
(a) Numenta Standard
(b) Numenta Reward-Low-FP
(c) Numenta Reward-Low-FN
Figure 4: Our Model vs. the Numenta Model
anomaly ranges where many of them were complete mispredictions. This has a more significant negative impact on
the precision score in range-based Precision T flat
than in point-based Precision Classical. This scenario demonstrates that PrecisionT is more range-aware
than Precision and can, therefore, judge exactness of rangebased predictions more accurately.
F1 Score: In Figure 3c, we present F1 scores. The general
behavior is as in the classical use case, because F1 is the
harmonic mean of respective recall and precision values in
both models. A noteworthy observation is that the F1 scores
display similar sensitivity to positional bias variations as
in recall graph of Figure 3a. This demonstrates that our
range-based model’s expressiveness for recall and precision
carries into combined metrics like F1 scores.
6.3. Comparison to the Numenta Scoring Model
In this section, we report results from our comparative experiment with the Numenta scoring model (Lavin & Ahmad,
2015). To our knowledge, the Numenta model is currently
the only alternative to the classical one for time series AD.
Our goals are: (i) to determine if our model can mimic the
Numenta model, (ii) to explore the potential of our model
in providing further flexibilities beyond Numenta.
Numenta is domain-specific. It focuses solely on rewarding
early predictions, similar to our front-end bias function.
Moreover, it is structurally different from precision/recallbased models as it provides a single, sigmoidal scoring
function, whose weights can be adjusted according to three
pre-defined application profiles (Standard, Reward-Low-FP,
Reward-Low-FN). In each profile, contribution of TP, FP,
and FN points are weighed differently. Although Numenta is
primarily point-based in nature, it has the concept of a fixedsize anomaly window. Each point prediction is mapped to its
closest preceding anomaly window and earns a reward based
on its position within that window. However, in practice
anomaly ranges vary and cannot be fixed to a single value in
advance. Numenta uses additional window padding to help
manage this. Finally, Numenta’s upper bound scoring value
is 1, but it has no fixed lower bound.
Given these irregularities (and several others (Singh & Olinsky, 2017)), it is challenging to make a direct and unbiased
comparison to the Numenta model. Instead, we focus on
comparing general trends of our Fβ score measurements to
Numenta’s scores, rather than their absolute values.
To obtain a behavior that is similar to Numenta’s, we
used the following settings for our model: α = 0, β = 1,
γ() = 1, ω() is as in Figure 2a, δ() is front-bias for Recall
and flat-bias for Precision. Because Numenta makes pointbased predictions, we represent Pi as points. For each run,
we compute Recall T Front and Precision T Flat and
their F score corresponding to the Numenta application profile under consideration (i.e., F1 for Numenta Standard ,
F0.5 for Numenta Reward Low FP , and F2 for
Numenta Reward Low FN ).
Figure 4 contains three bar charts, one for each Numenta
profile. The datasets on the x-axis are presented in descending order of the Numenta accuracy scores. For ease of
exposition, we scaled the negative Numenta scores by a factor of 100 and lower-bounded the y-axis at -1. For all charts,
we added approximate trends. At a high level, although the
accuracy values vary slightly across the charts, the trend
curves follow a similar pattern. Therefore, we only discuss
Figure 4a.
In Figure 4a, both models trend in a decreasing fashion
across the datasets from left to right, with one visible exception from the Space-Shuttle dataset. Analyzing the real
and predicted data for the Space-Shuttle, we noted a small
number of mid-range real anomalies and a much larger number of mid-range predicted anomalies. Contrarily, when
we analyzed similar data for Sine, we found it had many
narrow-range real and predicted anomalies that appear close
together. These findings reveal that Space-Shuttle data had
more FPs with a greater probability of TPs, which is therefore reflected in its smaller PrecisionT score and larger
RecallT score (see Figure 3). Moreover, because Numenta
favors precision over recall, this discrepancy is further magnified. Overall, this experiment shows that, not only can our
model mimic Numenta, but also catch additional intricacies.
To illustrate this further, we investigated how the two models
behave under two contrasting positional bias scenarios: (i)
anomaly predictions overlap with front-end portions of real
anomaly ranges (Front Predicted) and (ii) anomaly predic-
A New Model for Evaluating Range-Based Anomaly Detection Algorithms
tions overlap with back-end portions of real anomaly ranges
(Back Predicted) as shown in Table 2. As these scenarios
did not exist in our real and Paranom-generated datasets,
we artificially generated the desired ranges in a symmetric
fashion to make them directly comparable. We then scored
them using Numenta Standard , F1 T Front Flat, and
F1 T Back Flat. As shown in Table 2, Numenta’s scoring
function is not sufficiently sensitive to distinguish between
the two scenarios. This is not too surprising, as Numenta
was designed to reward early detections (closer to Front Predicted). However, our model clearly distinguishes between
the two scenarios when its positional bias is set appropriately. This experiment demonstrates our model’s flexibility
and generality over the Numenta model in evaluating a
wider range of time series AD algorithms and scenarios.
(a) Naive
Table 2: Sensitivity to Positional Bias
Front Predicted
Back Predicted
Numenta Std
0.67
0.63
F1 T Front
0.42
0.11
F1 T Back
0.11
0.42
6.4. Cost Analysis
In this section, we analyze the computational overhead of
our RecallT and P recisionT equations, both mathematically and experimentally.
Naive: In our RecallT and P recisionT equations, we need
to keep track of two sets, R and P . A naive algorithm, which
compares each Ri ∈ R with all Pj ∈ P for RecallT (and
vice versa for P recisionT ) has a computational complexity
of O(Nr × Np ).
Optimization1: We can eliminate many of these comparisons by taking advantage of sequential relationships among
ranges. Assume that all ranges are represented as timestamp
pairs, ordered by their first timestamps. We can iterate over
R and P simultaneously, only comparing those pairs that
are in close proximity in terms of their timestamps. This
optimized algorithm brings the computational complexity
to linear (O(Nr ) if Nr ≥ Np , otherwise O(Np )).
Optimization2: An additional optimization is possible in
computing the positional bias functions. Instead of computing them for each point, we can use them in a closed form,
leading to a single computation per range.
We note that these optimizations are not exhaustive. In fact,
many others (e.g., using bit vectors and bitwise operations)
can be applied. These are beyond the goals of our analysis.
We implemented five alternative ways of computing recall
and precision, and analyzed how the computation time for
each of them scales with increasing number of real and predicted anomaly ranges. Two of these are naive algorithms
for both classical and our recall/precision equations. The
rest are their optimized variants. Ranges used in this experiment are randomly generated from a time series of 50K
data points. Figure 5 shows our results. We plot naive and
(b) Optimized
Figure 5: Cost Analysis
optimized approaches in separate graphs, as their costs are
orders of magnitude different. In Figure 5a, we see that
computing our range-based metrics naively is significantly
more costly than doing so for the classical ones. Fortunately, we can optimize our model’s computational cost.
The optimized results in Figure 5b provide an improvement
in performance. Both the classical metrics and ours can be
computed in almost three orders of magnitude less time than
their corresponding naive baselines when optimized. There
is still a factor of 2-3 difference, but our model’s values
are notably closer at this scale to the classical model. This
analysis shows that our new range-based metrics can be
computed efficiently, and their overhead compared to the
classical ones is acceptably low.
7. Conclusion
Precision/recall measures are commonly used to evaluate
accuracy of algorithms with imprecise results, like anomaly
prediction. In this paper, we argued that classical precision/recall were invented for point-based data, whereas for
time series data, anomalies are in the form of ranges. In
response, we offer a new formal model that accounts for
range-specific issues such as partial overlaps between real
vs. predicted ranges as well as their relative positions. Moreover, our approach allows users to supply customizable bias
functions that enable weighing multiple criteria differently.
Experiments with diverse datasets compared to two existing models show that our new model is expressive, flexible,
and extensible. Future work includes designing new ML
algorithms that are optimized for our new accuracy model.
A New Model for Evaluating Range-Based Anomaly Detection Algorithms
Acknowledgements
We thank Eric Metcalf for his help with the cost experiments.
This research has been funded in part by Intel.
References
Aggarwal, C. C. Outlier Analysis. Springer, 2013.
Ahmad, S., Lavin, A., Purdy, S., and Agha, Z. Unsupervised Real-time Anomaly Detection for Streaming Data.
Neurocomputing, 262:134–147, 2017.
Alam, M. M. U. and Muzahid, A. Production-Run Software
Failure Diagnosis via Adaptive Communication Tracking. In ACM/IEEE Annual International Symposium on
Computer Architecture (ISCA), pp. 354–366, 2016.
AlEroud, A. and Karabatis, G. A Contextual Anomaly
Detection Approach to Discover Zero-Day Attacks. In
International Conference on CyberSecurity (ICCS), pp.
40–45, 2012.
Anava, O., Hazan, E., and Zeevi, A. Online Time Series Prediction with Missing Data. In International Conference
on Machine Learning (ICML), pp. 2191–2199, 2015.
Bailis, P., Gan, E., Madden, S., Narayanan, D., Rong, K.,
and Suri, S. MacroBase: Prioritizing Attention in Fast
Data. In ACM International Conference on Management
of Data (SIGMOD), pp. 541–556, 2017.
Chandola, V., Banerjee, A., and Kumar, V. Anomaly Detection: A Survey. ACM Computing Surveys, 41(3):15:1–
15:58, 2009.
Formenti, S. and Demaria, S. Systemic Effects of Local
Radiotherapy. The Lancet Oncology, pp. 718–726, 2017.
Gottschlich, J. Paranom: A Parallel Anomaly Dataset Generator. https://arxiv.org/abs/1801.03164/, 2018.
Guha, S., Mishra, N., Roy, G., and Schrijvers, O. Robust Random Cut Forest Based Anomaly Detection on Streams. In
International Conference on Machine Learning (ICML), pp.
2712–2721, 2016.
Keogh, E. UCR Time-Series Datasets. http://www.cs.ucr.
edu/˜eamonn/discords/, 2005.
Kourou, K., Exarchos, T. P., Exarchos, K. P., Karamouzis, M. V.,
and Fotiadis, D. I. Machine Learning Applications in Cancer Prognosis and Prediction. Computational and Structural
Biotechnology Journal, 13:8–17, 2015.
Lavin, A. and Ahmad, S. Evaluating Real-Time Anomaly Detection Algorithms - The Numenta Anomaly Benchmark. In IEEE
International Conference on Machine Learning and Applications (ICMLA), pp. 38–44, 2015.
Lipton, Z. C., Kale, D. C., Elkan, C., and Wetzel, R. C. Learning to Diagnose with LSTM Recurrent Neural Networks. In
International Conference on Learning Representations (ICLR),
2016.
Malhotra, P., Vig, L., Shroff, G., and Agarwal, P. Long Short Term
Memory Networks for Anomaly Detection in Time Series. In
European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), pp. 89–94,
2015.
Numenta. NAB Data Corpus. https://github.com/
numenta/NAB/tree/master/data/, 2017.
Patcha, A. and Park, J. An Overview of Anomaly Detection Techniques: Existing Solutions and Latest Technological Trends.
Computer Networks, 51(12):3448–3470, 2007.
Singh, N. and Olinsky, C. Demystifying Numenta Anomaly Benchmark. In International Joint Conference on Neural Networks
(IJCNN), pp. 1570–1577, 2017.
Twitter. AnomalyDetection R Package. https://github.
com/twitter/AnomalyDetection/, 2015.
Yu, H., Rao, N., and Dhillon, I. S. Temporal Regularized Matrix
Factorization for High-dimensional Time Series Prediction. In
Annual Conference on Neural Information Processing Systems
(NIPS), pp. 847–855, 2016.
| 2cs.AI
|
arXiv:1606.07976v1 [math.AC] 25 Jun 2016
TOTALLY ACYCLIC APPROXIMATIONS
PETTER A. BERGH, DAVID A. JORGENSEN, AND W. FRANK MOORE
Abstract. Let R be a commutative local ring. We study the subcategory
of the homotopy category of R-complexes consisting of the totally acyclic Rcomplexes. In particular, in the context where Q → R is a surjective local ring
homomorphism such that R has finite projective dimension over Q, we define
an adjoint pair of functors between the homotopy category of totally acyclic
R-complexes and that of Q-complexes, which are analogous to the classical
adjoint pair between the module categories of R and Q. We give detailed
proofs of the adjunction in terms of the unit and counit. As a consequence,
one obtains a precise notion of approximations of totally acyclic R-complexes
by totally acyclic Q-complexes.
Introduction
This paper was motivated by the desire to approximate in a meaningful way totally acyclic complexes over a commutative local ring R by simpler totally acyclic
complexes, possibly even periodic ones. The subcategory Ktac (R) of the homotopy
category of R-complexes consisting of totally acyclic R-complexes is a thick subcategory, and therefore is also a triangulated category. To facilitate an approximation,
we assume that there exists a surjective local ring homomorphism ϕ : Q → R such
that Q is Gorenstein and R has finite projective dimension as a Q-module. Our
approximation is then achieved through an adjoint pair of triangle functors
Sϕ
Ktac (Q) o
Tϕ
/ K (R)
tac
which are consonant with the classical adjoint pair of functors between the module
categories of Q and R. Indeed, as is the case in the classical setting, Sϕ is simply
the base change functor. However, as nontrivial totally acyclic R-complexes are
never totally acyclic Q-complexes when R 6= Q, obtaining the right adjoint Tϕ
requires a modification of the forgetful functor. Its construction is given in Section
2, and we prove in detail that it is a triangle functor. We remark also that the
condition of pdQ R < ∞ is key to the existence of Tϕ . In Section 3 we prove that
Sϕ and Tϕ form an adjoint pair via the unit and counit natural transformations. In
Section 4 we recall the precise notion of approximation, in terms of the counit of the
adjunction. The approximations are especially meaningful in Remark 4.7, where
totally acyclic complexes over a quotient of a hypersurface ring are approximated
by period 2 complexes. Our interest in this project was also motivated by recent
work in [Kr], [Ne] on approximations and adjoints in related categories.
Date: February 2, 2018.
2010 Mathematics Subject Classification. 13D02, 13H10, 13C14.
Key words and phrases. Totally acyclic complex, Approximation, Adjoint functors, Support
variety .
1
2
PETTER A. BERGH, DAVID A. JORGENSEN, AND W. FRANK MOORE
To put the functors S = Sϕ and T = Tϕ in further perspective, Let Db (R)
denote the bounded derived category of R and Db (Q) that of Q. The classical
adjoint pair of the base change functor and the forgetful functor for the module
categories extend to an adjoint pair of the bounded derived categories. Since the
projective dimension of R over Q is finite, a perfect complex in Db (R) under the
forgetful functor is a perfect complex in Db (Q). Thus the adjoint pair extends to
the verdier quotients of the bounded derived category modulo the corresponding
thick subcategories of perfect complexes, in other words the singularity categories
Dbsg (Q) and Dbsg (R). This explains the right pair of vertical arrows in the diagram
below. The top horizontal arrow is the equivalence of categories proved in [Bu]; the
functor is defined by hard truncation of a totally acyclic complex to the right of
zero. It is proved in [BeJoOp] that the same functor is fully faithfully in general,
which explains the bottom arrow.
∼
=
Ktac (Q)
O
S
Ktac (R)
τ
σ
T
/ Db (Q)
sg
O
/ Dbsg (R)
The adjoint pair S and T of this paper is the corresponding adjoint pair making
the square commute.
1. Preliminaries
Let R be a ring. By an R-complex C we mean a sequence of left R-module
homomrphisms
C
∂n+1
∂C
C : · · · → Cn+1 −−−→ Cn −−n→ Cn−1 → · · ·
graded homologically, where each Cn is a left R-module, and n is its homological
degree. Without prior stipulation, an arbitrary R-module is assumed also to be a
complex concentrated in homological degree 0.
Recall that a semi-free R-complex F is one whose underlying graded
F R-module
F ♮ has graded basis E which can be written as a disjoint union E = n≥0 E n such
Sn−1
that ∂ F (E n ) ⊆ Rh i=1 E i i for every n ≥ 1.
Definition. A semi-free resolution of an R-complex C is a quasi-isomorphism
F → C of complexes where F is a semi-free R-complex.
We now state some important facts, and simply give a reference when the proofs
are available in the literature.
1.1. [ChFoHo, 5.1.7, 5.1.13] Every R-complex C has a semi-free resolution π : F →
C with Fn = 0 for all n < inf{i | Hi (C) 6= 0}. Moreover, π can be chosen to be
surjective. If R is left Noetherian and C is a bounded below complex of finitely
generated R-modules, then π can even be chosen to be surjective with Fn finitely
generated for all n, and Fn = 0 for all n < inf{i | Ci 6= 0}.
The next result is the so-called ‘Comparison Theorem.’
1.2. [ChFoHo, 5.2.9] Let π ′ : F ′ → C ′ be a quasiisomorphism of R-complexes, and
α : F → C ′ a morphism of R-complexes with F semi-free. Then there exists a
morphism of complexes µ : F → F ′ such that π ′ µ ∼ α. If π ′ is surjective, then µ
TOTALLY ACYCLIC APPROXIMATIONS
3
can be chosen such that π ′ µ = α. Moreover, µ is homotopic to any other morphism
of R-complexes µ′ with π ′ µ′ = α.
From this point on we assume that R is a commutative Noetherian ring.
Definition. Recall from [AvMa] that an R-complex C of finitely generated free
modules is called totally acyclic if
H(C) = 0 = H(HomR (C, R)).
We denote by Ktac (R) the homotopy category of totally acyclic R-complexes; the
objects in Ktac (R) are the totally acyclic R-complexes, and the morphisms are
homotopy equivalence classes of morphisms of R-complexes. For a morphism f :
C → C ′ of R-complexes we write [f ] for its homotopy equivalence class. Thus for
two morphisms f, g : C → C ′ of R-complexes, one has f ∼ g if and only if [f ] = [g].
The following facts we use often in the rest of the paper. Given an R-complex
C, from now on we write C ∗ for the R-complex HomR (C, R).
1.3. Suppose that s is an integer, C, C ′ ∈ Ktac (R) and f : C → C ′ is a morphism
′
C′
of complexes. If there exist maps σn : Cn → Cn+1
satisfying fn = σn−1 ∂nC +∂n+1
σn
for all n > s, then the σn can be extended to a homotopy showing that f ∼ 0.
′
Proof. By induction it suffices to show there exists a map σs : Cs → Cs+1
such
′
C
C
that fs+1 = σs ∂s+1 + ∂s+2 σs+1 .
′
Note that since C is a totally acyclic complex and Cs+1
is free, the complex
′
HomR (C, Cs+1 ) is exact. We have
′
HomR (C,Cs+1
)
C
C′
C′
∂s+1
(fs+1 − ∂s+2
σs+1 ) = fs+1 − ∂s+2
σs+1 ∂s+2
C
C′
C′
= fs+1 ∂s+2
− ∂s+2
fs+2 − ∂s+3
σs+2
C
C′
C′
C′
= fs+1 ∂s+2
− ∂s+2
fs+2 − ∂s+3
∂s+3
σs+2
=0
′
′
HomR (C,Cs+1
)
C
σs+1 ∈ Ker ∂s+1
Thus fs+1 − ∂s+2
′
such that
exists a map σs : Cs → Cs+1
C
C′
other words fs+1 = σs ∂s+1 + ∂s+2
σs+1 .
′
HomR (C,Cs+1
)
= Im ∂s
′
HomR (C,Cs+1
)
∂s
(σs )
. Therefore there
′
C
σs+1 , in
= fs+1 − ∂s+2
1.4. Suppose that s is an integer and C, C ′ ∈ Ktac (R). If there exist maps fn :
′
Cn → Cn′ satisfying fn−1 ∂nC = ∂nC fn for all n > s, then the fn can be extended
to a morphism of complexes f : C → C ′ . Any two such extensions are homotopic,
that is, if g : C → C ′ is a morphism of complexes such that there exist maps
′
′
σn : Cn → Cn+1
with fn − gn = σn−1 ∂nC + ∂nC σn for all n ≥ s, then f ∼ g.
Proof. The second statement follows immediately from 1.3. The proof of the first
statement is essentially that of 1.3, and is left to the reader.
Definition. A complete resolution (see, for example, [AvMa]) of an R-complex C
is a diagram of morphisms of R-complexes
ρ
π
U−
→F −
→C
such that U ∈ Ktac (R), F is a semi-free resolution of C, and ρn is bijective for all
n ≫ 0. We will often abuse terminology and call U a complete resolution of C.
4
PETTER A. BERGH, DAVID A. JORGENSEN, AND W. FRANK MOORE
The following is a slightly generalized version of [AvMa, 5.3]. The proof is
essentially the same as that in ibid.
1.5. Complete resolutions are well-defined up to homotopy equivalence. Specifiρ
π
ρ′
π′
cally, if U −
→F −
→ C and U ′ −→ F ′ −→ C ′ are complete resolutions of R-complexes
′
C and C , and µ : C → C ′ is a morphism of complexes, then by the comparison theorem 1.2 there exists a unique-up-to-homotopy morphism µ making the right-hand
square of the diagram
/F
/C
U
µ
b
µ
µ
/ F′
/ C′
U′
commute up to homotopy, and for each choice of µ there exists a unique-up-tohomotopy morphism µ
b making the left-hand square commute up to homotopy. If
b are homotopy equivalences.
µ = IdC , then µ and µ
1.6. Suppose that Q and R are commutative Noetherian rings, and ϕ : Q → R
is a surjective ring homomorphism. Let C be a complex of finitely generated free
Q-modules. Then
HomQ (C, Q) ⊗Q R and HomR (C ⊗Q R, R)
are isomorphic as R-complexes.
Proof. By Hom-tensor adjunction for complexes, and after making the canonical
identification of HomR (R, R) with R, we immediately get that HomR (C ⊗Q R, R)
and HomQ (C, R) are isomorphic. Therefore it suffices to prove that HomQ (C, Q)⊗Q
R and HomQ (C, R) are isomorphic.
Define maps αn : HomQ (C−n , Q) ⊗Q R → HomQ (C−n , R) by αn (f ⊗ r)(x) =
ϕ(f (x))r, and βn : HomQ (C−n , R) → HomQ (C−n , Q) ⊗Q R by βn (g) = g ′ ⊗Q 1R
where ϕg ′ = g. Thus αn (βn (g))(x) = αn (g ′ ⊗ 1R )(x) = ϕ(g ′ (x)) = g(x), and so
αn βn = IdHomQ (C−n ,R) . Also, βn (αn (f ⊗ r)) = g ′ ⊗ 1R where ϕ(g ′ (x)) = ϕ(f (x))r,
and so βn αn = IdHomQ (C−n ,Q)⊗Q R . Finally one just needs to check that αn and βn
commute with the differentials, and this is left to the reader.
2. The ‘forgetful’ triangle functor
Let Q be a commutative local Gorenstein ring, and ϕ : Q → R a surjective local
ring homomorphism such that pdQ R < ∞.
The main objective of this section is to define the ‘forgetful’ functor
T = Tϕ : Ktac (R) → Ktac (Q)
and prove that it is a triangle functor. The definition of T is as follows.
Definition. Let C ∈ Ktac (R). Then T C ∈ Ktac (Q) is a complete resolution of
Im ∂0C over Q (which is uniquely defined by 1.5.) Given a morphism [f ] : C → C ′ in
′
Ktac (R), we have the map µ : Im ∂0C → Im ∂0C induced by the morphism f : C → C ′
of R-complexes. Then T [f ] : T C → T C ′ is the homotopy equivalence class [b
µ] of the
comparison map µ
b : T C → T C ′ between complete resolutions (which is uniquely
determined by µ, by 1.5.)
Proposition 2.1. T : Ktac (R) → Ktac (Q) is an additive functor.
TOTALLY ACYCLIC APPROXIMATIONS
5
Proof. The only verifications of the axioms of additive functor that are possibly not
obvious follow easily from the fact that T is well-defined on morphisms in Ktac (R),
which we now prove. Let [f ] : C → C ′ be the zero morphism in Ktac (R). Let
′
µ : Im ∂0C → Im ∂0C be the map induced by f . Then there exists a homotopy
′
′
′
σ ∈ HomR (C, C ′ )1 such that ∂0C (σ−1 |Im ∂0C ) = µ, where ∂0C : C0′ → Im ∂0C is the
′
′
map induced by ∂0C . Let F and F ′ be Q-free resolutions of Im ∂0C and Im ∂0C ,
respectively, and K a minimal Q-free resolution of C0′ . Then any chain map µ :
F → F ′ lifting µ is homotopic to a lifting of the composition
···
/ F1
···
/ K1
···
/ F1′
∂1F
/ F0
/ Im ∂0C
∂1K
/ K0
/ C0′
′
∂1F
/ F0′
/ Im ∂0C ′
/0
σ−1 |Im ∂ C
0
∂0C
/0
′
/0
which is eventually zero since K is a finite complex. This shows that µ
b ∼ 0, in
other words, T [f ] = 0.
Theorem 2.2. T : Ktac (R) → Ktac (Q) is a triangle functor.
Proof. We first show that T commutes with shifts, that is, there is a natural isomorphism between T Σ and ΣT . Let M = Im ∂0C . From Corollary 2.4 below we see
that a minimal free resolution of ΩR
1 (M ) agrees with that of M beginning at degree
pdQ R + 1. Specifically if F and F ′ are minimal free resolutions of M and ΩR
1 (M )
′
, for c = pdQ R. This
over Q, respectively, then we have Σ−1 (F≥c+2 ) ∼
= F≥c+1
implies that Σ−1 (T C) ≃ T (Σ−1 C), which is another way of stating the result. It is
clear that this isomorphism respects morphisms in Ktac (R), that is, if [f ] : C → C ′
is a morphism in Ktac (R), then the following diagram commutes.
T ΣC
T Σ[f ]
/ T ΣC ′
≃
≃
ΣT C
ΣT [f ]
/ ΣT C ′
Next we show that T takes distinguished triangles to distinguished triangles.
Any distinguished triangle in Ktac (R) is isomorphic as a triangle to one of the form
[f ]
C −−→ C ′ → cone([f ]) → ΣC. We need to complete the diagram
T [f ]
TC
T [f ]
TC
/ T C′
/ T cone([f ])
/ Σ(T C)
/ T C′
/ cone(T [f ])
/ Σ(T C)
so that the second two squares commute.
ρ
π
ρ′
π′
′
Let T C −
→ F −
→ C≥−1 and T C ′ −→ F ′ −→ C≥0
be complete resolutions over
′
Q, and µ : C≥−1 → C≥0 be the map induced by f . Then by 1.5 there exists
6
PETTER A. BERGH, DAVID A. JORGENSEN, AND W. FRANK MOORE
a morphism of Q-complexes µ : F → F ′ such that all squares in the following
diagram commute (we can assume that π ′ is surjective.)
/ F1
F2
F2′
π2
µ2
z
/ F′
1
π2′
z
µ1
/ F0
π1
/ F′
π−1
π0′
/ C1
t
tt
ztttµ1
/ C′
1
t
tt
ztttµ2
C2′
π0
µ0
0
π1′
C2
z
/ F−1
/ C−1
/ C0
t
tt
ztttµ0
/ C′
0
This diagram gives rise to a commutative diagram of short exact sequences of
morphisms of Q-complexes
/ F′
0
/ cone(µ)
π′
/ cone(µ)
/ C′
≥0
0
π′ 0
0 Σπ
/ ΣF
/0
Σπ
/ Σ(C≥−1 )
/0
The resulting commutative diagram of long exact sequences of homology shows that
the morphism of Q-complexes cone(µ) → cone(µ) is a quasiisomorphism. Thus
cone(f )
cone(µ) is a semi-free resolution of cone(µ), and therefore also of Im ∂0
. It
follows that
cone(f )
cone(b
µ) → cone(µ) → Im ∂0
cone(f )
is a complete resolution of Im ∂0
, where µ
b : T C → T C ′ is a morphism of
Q-complexes as in 1.5. We now have that the diagram of morphisms
TC
µ
b
/ T C′
/ T cone(f )
/ Σ(T C)
TC
µ
b
/ T C′
/ cone(b
µ)
/ Σ(T C)
which commutes up to homotopy. The result follows.
For an arbitrary commutative local ring A, consider a short exact sequence of
π
finitely generated A-modules 0 → X → Y −
→ Z → 0. Let F and G be minimal free
resolutions of Y and Z, respectively. Let cone(φ) denote the mapping cone of the
morphism φ : F → G lifting the surjection π:
−∂1F 0
φ1 ∂2G
( φ0 ∂1G )
cone(φ) : · · · → F1 ⊕ G2 −−−−−−−−→ F0 ⊕ G1 −−−−−−→ G0 .
Lemma 2.3. In the notation of the previous discussion, the truncated mapping
cone
−∂2F 0
φ2 ∂3G
−∂1F 0
φ1 ∂2G
trcone(φ) : · · · → F2 ⊕ G3 −−−−−−−−→ F1 ⊕ G2 −−−−−−−−→ ker φ0
is (after shift) a free resolution of X.
∂1G
TOTALLY ACYCLIC APPROXIMATIONS
7
Proof. We first note that Since F and G are both minimal free resolutions, the map
φ0 is surjective. Therefore so is the map ( φ0 ∂1G ), and so ker ( φ0 ∂1G ) is free.
From the short exact sequence of complexes 0 → G → cone(φ) → ΣF → 0, we
get the long exact sequence of homology
π
· · · → H1 (G) → H1 (cone(φ)) → H0 (F ) −
→ H0 (G) → H0 (cone(φ)) → 0.
Since ( φ0 ∂1G ) is surjective, H0 (cone(φ)) = 0. Thus this long exact sequence of
homology reduces to the short exact sequence
π
0 → H1 (cone(φ)) → Y −
→ Z → 0,
and so X ∼
= H1 (cone(φ)). Since Hi (cone(φ)) = Hi (trcone(φ) for all i ≥ 1, the result
follows, that is Σ−1 trcone(φ) is a free resolution of X.
Corollary 2.4. Let K be a minimal free resolution of R over Q, F a minimal
resolution of M over Q, and µ = rank F0 . Consider the morphism of complexes
φ : K µ → F lifting the surjection F0 → M . Then Σ−1 trcone(φ) is a free resolution
−1
of ΩR
trcone φ)≥c+1 = Σ−1 (F≥c+2 )
1 (M ) over Q. If c = pdQ R < ∞ then (Σ
3. Adjunction
Keeping the same assumptions on Q and R as in the previous section, the goal of
this section is to compare Ktac (R) with Ktac (Q). This will be done by constructing
an adjoint pair of triangle functors between the two categories.
The descension functor is easy:
S = Sϕ : Ktac (Q) → Ktac (R)
is defined by SC = C ⊗Q R and S[f ] = [f ⊗Q R] for C an object and [f ] a morphism
in Ktac (Q). This is a triangle functor due in part to 1.6. Indeed, if C is acyclic
complex of free Q-modules, then C ⊗Q R is an acyclic complex of free R-modules
(since pdQ R < ∞), and HomQ (C, Q) being acyclic implies HomQ (C, Q) ⊗Q R
is acyclic. Hence HomR (C ⊗Q R, R) is acyclic by 1.6. It is easy to see that S
takes homotopic morphisms of complexes to homotopic morphisms of complexes,
commutes with shifts and takes distinguished triangles to distinguished triangles.
The ascension functor
T = Tϕ : Ktac (R) → Ktac (Q)
is the functor defined in Section 2.
Our main result for this section is the following.
Theorem 3.1. The triangle functors S and T form an adjoint pair, that is, they
satisfy the following property: for all C ∈ Ktac (R) and D ∈ Ktac (Q) there exist a
bijection
(1)
HomKtac (Q) (D, T C) → HomKtac (R) (SD, C)
which is natural in each variable.
Before engaging the proof, we observe that for D ∈ Ktac (Q) one has
T SD ≃ D ⊗Q K
where K is a Q-free resolution of R. Indeed, one has that D≥0 ⊗Q K is a Q-free
resolution of Im ∂0SD ∼
= Im ∂0D ⊗Q R. The assertion is now clear.
8
PETTER A. BERGH, DAVID A. JORGENSEN, AND W. FRANK MOORE
Proof. In order to prove the theorem we define natural transformations
η : IdKtac (Q) → T S
and
ǫ : ST → IdKtac (R)
— the unit and counit, respectively, of the adjunction — as follows. For D ∈
Ktac (Q) define ηD : D → T SD to be the
Lnmorphism of complexes embedding
Dn into the first component of T SDn =
i=0 Dn−i ⊗Q Ki for all n. And for
C ∈ Ktac (R) define ǫC : ST C → C to be the morphism induced by the comparison
map F → C≥0 , where F is a free resolution of Im ∂0C over Q. It follows from 1.4
and 1.3 that η and ǫ are natural in their arguments.
We just need to show that
T ǫC ◦ ηT C ∼ IdT C
and
ǫSD ◦ SηD ∼ IdSD
First we discuss the map T ǫC . We have a morphism of complexes
···
/ T C1
···
/ F1
µ
b1
∂1T C
∂1F
/ T C0
µ
b0
/ F0
/ Im ∂0T C
/0
ν
/ Im ∂ C
0
/0
where F is a Q-free resolution of Im ∂0C . Suppose that ker ϕ = (x1 , . . . , xr ). For each
1 ≤ i ≤ r, let σi : T C0 → F1 be the beginning of homotopies expressing the fact that
the morphisms xi µ
b are null homotopic, that is, ∂1F ◦ σi = xi µ
b0 for 1 ≤ i ≤ r. Now
let K be a minimal free resolution of R over Q and define maps u0 : T C0 ⊗Q K0 →
F0 by u0 (a ⊗ b) = µ
b0 (a)b, and u1 (a ⊗ (b1 , . . . , br )) = σ1 (a)b1 + · · · + σc (a)br for
a ⊗ (b1, . . . , br ) ∈ T C0 ⊗Q K1 , u1 (a ⊗ b) = µ
b1 (a)b for a ⊗ b ∈ T C1 ⊗Q K0 . Then one
checks easily that the following diagram commutes.
T C0 ⊗Q K1
❙❙❙
❙❙T❙C0 ⊗∂1K
L
❙❙❙
❙❙❙
❙❙)
TC
∂1 ⊗K0
/ T C0 ⊗Q K0
T C1 ⊗Q K0
u0
u1
F1
/ Im ∂0T C ⊗Q R
∂1F
/ F0
/0
ν⊗R
/ Im ∂ C
0
/0
This gives rise to a morphism of Q-complexes (T C≥0 ⊗Q K) → F such that un (a ⊗
b) = µ
bn (a)b for a ⊗ b ∈ T Cn ⊗Q K0 , the former complex being a Q-free resolution
of Im ∂0T C ⊗Q R, It follows that we may achieve the morphism T ǫC : T ST C → T C
satisfying (T ǫC )n (a ⊗ b) = ab for a ⊗ b ∈ T Cn ⊗Q K0 .
We have the natural embedding ηT C : T C → T ST C with ηT C (a) = a ⊗ 1 ∈
T Cn ⊗Q K0 for all a ∈ T Cn . Thus we have shown that T ǫC ◦ ηT C ∼ IdT C .
The morphism
SηD : SD → ST SD embeds SDn into the first component of
Lc
ST SD ≃ i=0 SDn−i ⊗Q SKi for all L
n. And the morphism ǫSD : ST SD → SD
c
takes the first component of ST SD ≃ i=0 SDn−i ⊗Q SKi to SDn for all n ∈ Z.
Thus we have ǫSD ◦ SηD ∼ IdSD .
TOTALLY ACYCLIC APPROXIMATIONS
9
4. Approximations of totally acyclic complexes
Our main application of Theorem 3.1 is a resulting notion of approximation in
the homotopy category of totally acyclic complexes. We now recall the notion of
approximation we use, due to Auslander and Smalø [AuSm], and independently,
Enochs [En],
Let X be a full subcategory of a category C. Then a right X -approximation of
ǫ
C ∈ C is a morphism X −
→ C, with X ∈ X , such that for all objects Y ∈ X , the
Hom(Y,ǫ)
sequence HomC (Y, X) −−−−−−→ HomC (Y, C) → 0 is exact.
Dually, one has the concept of left X -approximations. Specifically, a morphism
µ
C −
→ X, with X ∈ X , is called a left X -approximation of C ∈ C if for all objects
Hom(µ,Y )
Y ∈ X , the sequence HomC (X, Y ) −−−−−−→ HomC (C, Y ) → 0 is exact.
The full subcategory X is called functorially finite in C if for every object C ∈ C,
there exists a right X -approximation of C and a left X -approximation of C.
We let S Ktac (Q) = {D ⊗Q R | D ∈ Ktac (Q)}. Our main application of Theorem
3.1 is the following.
Theorem 4.1.
S Ktac (Q) is functorially finite in Ktac (R).
Proof. That every C ∈ Ktac (R) has a right S Ktac (Q)-approximation follows immediately from Theorem 3.1: the morphism [ǫC ] : ST C → C is a right approximation in Ktac (R). Indeed, if [f ] : SD → C is any morphism in Ktac (R) with
D ∈ Ktac (Q), then from the natural transformation ǫ : ST → IdKtac (R) we have
equality [ǫC ] ◦ ST [f ] = [f ] ◦ [ǫSD ]. Composing on the right with S[ηD ] we obtain
[ǫC ] ◦ ST [f ] ◦ S[ηD ] = [f ], and thus ST [f ] ◦ S[ηD ] : SD → ST C is the morphism
we seek.
Now we show that every C ∈ Ktac (R) has a left approximation. This can be
done by simply dualizing a right approximation. For this we will use several times
that for any given D ∈ Ktac (Q), one has the natural isomorphism of complexes
HomR (D ⊗Q R, R) ∼
= HomQ (D, Q)⊗Q R from 1.6. We have the right approximation
[ǫC ∗ ] : ST C ∗ → C ∗ of C ∗ . The claim is that [ǫ∗C ∗ ] : C ∼
= C ∗∗ → (ST C ∗ )∗ is
∗
a left approximation of C. Note that the target of ǫC ∗ is in S Ktac (Q) by the
aforementioned isomorphism of 1.6. Now let E ∈ Ktac (Q) and f : C → SE be
a morphism in Ktac (R). Then we have the morphism f ∗ : (SE)∗ → C ∗ , with
(SE)∗ in S Ktac (Q). Therefore we have that f ∗ ∼ ǫC ∗ g for some morphism g :
(SE)∗ → ST C ∗ . Dualizing back we have that f ∼ g ∗ ǫ∗C ∗ , which is what we needed
to show.
Motivated by results along the lines of [Ne, Proposition 1.4], we ask the following:
Question 4.2. Is S Ktac (Q) a thick subcategory of Ktac (R)?
We illustrate Theorem 4.1 with an example.
Example 4.3. Let R = k[x, y]/(x2 , y 2 ), and C be the totally acyclic R-complex
with Im ∂0C = Rxy ∼
= k:
x 0 −y
0 y x
( xy )
(x y)
( xy )
C : · · · → R −−−−−−−→ R2 −−−−→ R −−−→ R −−−→ R2 → · · ·
3
Then a free resolution of Im ∂0C over Q = k[x, y]/(x2 ) is given by
x −y
0 x
x −y
( x xy ) 2 0 x
(x y)
→ Q −−−−−→ Q2 −−−−→ Q → 0
F : · · · → Q −−−−−→ Q −−0−−
2
2
10
PETTER A. BERGH, DAVID A. JORGENSEN, AND W. FRANK MOORE
The right approximation ǫC : ST C → C takes the form
···
/ R2
···
x −y
0 x
10
00
01
/ R3
( x0 xy )
/ R2
/ R2
IdR2
x 0 −y
0 y x
/ R2
x −y
0 x
/ R2
(x y)
/ R2
(y 0)
(1 0)
/R
( x0 xy )
/R
( xy )
/ ···
y 0
0 0
/ R2
( xy )
/ ···
Since C is self-dual in this example, that is C ∼
= Σ−1 (C ∗ ), the left approximation
∗
: C → (ST C) takes the form
[ǫ∗C ]
···
/ R3
···
x 0 −y
0 y x
y 00
0 00
/ R2
x 0
−y x
(x y)
/ R2
/ R2
y 0
0 0
( xy )
/R
( y0 )
x 0
y x
/ R2
( xy )
/R
/ R2
( 10 )
x 0
−y x
/ R2
/ ···
IdR2
x 0
y x
/ R2
/ ···
Approximations may be trivial, in particular, when the projective dimension of
Im ∂0C is finite over Q, as is the case in the next example.
Example 4.4. Let R = k[x, y]/(x2 , y 2 ) and C the totally acyclic R-complex with
Im ∂0C = Ry:
(y)
(y)
(y)
(y)
C : · · · → R −−→ R −−→ R −−→ R −−→ R → · · ·
Then for Q = k[x, y]/(x2 ), pdQ Im ∂0C < ∞ and the approximation is [ǫC ] : 0 → C.
ǫ
Recall (from [AuSm], for example) that a morphism X −
→ C is called right
f
minimal if for every morphism X −
→ X such that ǫf = ǫ, we have that f is an
isomorphism. We now point out that the right approximation [ǫC ] : ST C → C may
or may not be right minimal.
Proposition 4.5. Suppose that D ∈ Ktac (Q). Then [ǫSD ] : ST SD → SD is not a
minimal approximation.
Proof. Let K be a Q-free resolution of R. As described in
proof of 3.1, ǫSD :
Lthe
c
ST SD → SD takes the first component of ST SD ≃
i=0 SDn−i ⊗Q SKi to
SDn for all n ∈ Z. Thus taking as [f ] : ST SD → ST SD the morphism sending
SDn ⊗Q SK0 to itself and everything else to zero, we have [ǫSD ] ◦ [f ] = [ǫSD ] and
[f ] is not an isomorphism in Ktac (R).
Example 4.6. Let R = k[x, y]/(x2 , y 2 ) and C the totally acyclic complex
x
x
x
C : ··· → R −
→R−
→R−
→ R → ···
Then M = Im ∂0C = Rx and a free resolution of M over Q = k[x, y]/(x2 ) is given
by
2
x −y 2
0 x
2
x y2
0 x
2
x −y 2
0 x
( x y2 )
· · · → Q −−−−−−→ Q −−−−−→ Q −−−−−−→ Q2 −−−−→ Q → 0
TOTALLY ACYCLIC APPROXIMATIONS
11
Thus ST C takes the form
( x x0 ) 2 ( x0 x0 ) 2 ( x0 x0 ) 2
· · · → R2 −−0−−
→ R −−−−→ R −−−−→ R → · · ·
and ǫC : ST C → C is given by (ǫC )n = ( 1 0 ) for all n. This is not a minimal
right approximation. Indeed, consider the morphism f : ST C → ST C given by
fn = ( 10 00 ). Then one has ǫC f = ǫC and f is not a homotopy equivalence.
4.7. Approximations by period 2 complexes. Recall that a local ring Q
is a hypersurface ring if Q is the quotient of a regular local ring by a principal
ideal; hypersurface rings are Gorenstein. In this case, Eisenbud [Ei] has shown that
totally acyclic complexes are always periodic of period at most two. Thus in the
setup where Q is a hypersurface (and, as always, that pdQ R < ∞ via ϕ : Q → R),
our approximations compare nonperiodic totally acyclic complexes with those of
period two. This setup occurs when R has an embedded deformation [Av].
We next state a few results for later reference. They have to do with compositions
of approximations, in two different senses.
Proposition 4.8. Consider a sequence of finite local ring homomorphisms
ϕ
ψ
Q−
→ R′ −
→R
such that Q and R′ are Gorenstein, pdQ R′ < ∞, and pdR′ R < ∞. Then Sψϕ and
Tψϕ are naturally isomorphic to Sψ Sϕ and Tϕ Tψ , respectively.
Proof. This follows from the fact that the assertion is clear for the S functors, and
from uniqueness of adjoints.
4.9. Upon computing the right approximation [ǫC ] : ST C → C, one may iterate
this process. Indeed, complete [ǫC ] to a triangle in Ktac (R) and rotate it to obtain
Σ−1 cone([ǫC ]) → ST C → C → .
Now compute a right approximation of Σ−1 ST C, and repeat. One then obtains a
sequence of maps in Ktac (R):
B : · · · → B3 → B2 → B1 → B0 → C
where B0 is a right approximation of C, B1 is a right approximation of Σ−1 cone(B0 →
C), etc. Note that since composing two consecutive maps in an exact triangle is
the zero map, one has that the same holds for the maps in B.
Proposition 4.10. Let ϕ : Q → R be a surjective local homomorphism of Gorenstein rings with pdQ R = 1, and let C ∈ Ktac (R). If [ǫC ] : ST C → C is the right
approximation of C, then cone([ǫC ]) is isomorphic to Σ2 C.
Proof. By the assumptions of the proposition, we have that R = Q/(x) for some
nonzerodivisor x contained in the maximal ideal of Q. Now let t : C → Σ2 C be the
cohomology operator defined by the embedded deformation R = Q/(x) (see [Av]).
Then one can show that we have an exact triangle
t
C−
→ Σ2 C → ΣST C →
where the last arrow is the approximation map Σ[ǫC ]. Therefore, after rotation we
have
[ǫC ]
t
ST C −−→ C −
→ Σ2 C →
which proves the claim.
12
PETTER A. BERGH, DAVID A. JORGENSEN, AND W. FRANK MOORE
The following proposition follows from the previous one.
Proposition 4.11. Let ϕ : Q → R be a surjective local homomorphism of Gorenstein rings with pdQ R = 1, and let C ∈ Ktac (R). Then the triangle resolution, as
in 4.9, of C in Ktac (R) with respect to Ktac (Q) has the form
· · · → Σ−2 ST C → Σ−1 ST C → ST C → C.
References
[AuSm]
M. Auslander and S. O. Smalø, Preprojective modules over Artin algebras J. Algebra
66(1) (1980), 61122.
[Av]
L. Avramov, Homological asymptotics of modules over local rings, Commutative Algebra (Berkeley, 1987), MSRI Publ. 15, Springer, New York 1989; pp. 33–62.
[AvMa] L. Avramov and A. Martsinkovsky, Absolute, relative, and Tate cohomology of modules
of finite Gorenstein dimension Proc. London Math. Soc. (3) 85 (2002), 393–440.
[BeJoOp] P. A. Bergh, D. A. Jorgensen and S. Oppermann, The Gorenstein defect category Quarterly Journal of Mathematics 6 (2015), no. 2, 459–471.
[Bu]
R.-O. Buchweitz, Maximal Cohen-Macaulay modules and Tate-cohomology over Gorenstein rings, 1987, 155 pp; see
https://tspace.library.utoronto.ca/handle/1807/16682.
[ChFoHo] L. W. Christensen, H.-B. Foxby and H. Holm, Derived category methods in commutative
algebra in preparation.
[Ei]
D. Eisenbud, Homological algebra on a complete intersection, with an application to
group representations, Trans. Amer. Math. Soc. 260 (1980), no. 1, 35–64.
[En]
E. E. Enochs, Injective and at covers, envelopes and resolvents, Israel J. Math. 39
(1981), no. 3, 189–209.
[Kr]
H. Krause, Approximations and adjoints in homotopy categories, Math. Ann. 353
(2012), no. 3, 765–781.
[Ne]
A. Neeman, Some adjoints in homotopy categories, Ann.Math (2) 171(3) (2010), 2143–
2155.
Petter Andreas Bergh, Institutt for matematiske fag, NTNU, N-7491 Trondheim,
Norway
E-mail address: [email protected]
David A. Jorgensen, Department of mathematics, University of Texas at Arlington,
Arlington, TX 76019, USA
E-mail address: [email protected]
W. Frank Moore, Department of mathematics, University of Texas at Arlington,
Arlington, TX 76019, USA
E-mail address: [email protected]
| 0math.AC
|
NEURAL IDEALS IN SAGEMATH
arXiv:1609.09602v1 [q-bio.NC] 30 Sep 2016
ETHAN PETERSEN, NORA YOUNGS, RYAN KRUSE, DANE MIYATA, REBECCA GARCIA,
AND LUIS DAVID GARCÍA PUENTE
Abstract. A major area in neuroscience research is the study of how the brain processes spatial
information. Neurons in the brain represent external stimuli via neural codes. These codes often
arise from stereotyped stimulus-response maps, associating to each neuron a convex receptive field.
An important problem consists in determining what stimulus space features can be extracted directly from a neural code. The neural ideal is an algebraic object that encodes the full combinatorial
data of a neural code. This ideal can be expressed in a canonical form that directly translates to
a minimal description of the receptive field structure intrinsic to the code. In here, we describe a
SageMath package that contains several algorithms related to the canonical form of a neural ideal.
1. Introduction
Due to many recent technological advances in neuroscience, researchers’ ability to collect neural
data has increased dramatically. With this comes a need for new methods to process and understand
this data. One major question faced by researchers is to determine how the brain encodes spatial
features of its environment through patterns of neural activity, as with place cell codes [4]. In the
recent paper [2], Curto et al. phrase this question as, “What can be inferred about the underlying
stimulus space from neural activity alone?” To answer this question, Curto et al. introduced the
neural ring and a related neural ideal, algebraic objects that encode the full combinatorial data of
a neural code. They further show that the neural ideal can be expressed in a canonical form that
directly translates to a minimal description of the receptive field structure intrinsic to the code.
In this article we describe a SageMath [5] package that implements several algorithms to compute
the neural ideal and its canonical form, featuring a new iterative algorithm which proves to be
much more efficient than the original canonical form algorithm outlined by Curto et al. in [2]. This
package also provides an algorithm to compute the primary decomposition of a pseudo-monomial
ideal. Accompanying these functions are others that calculate related objects, such as the Gröbner
basis, Gröbner fan, universal Gröbner basis, and the neural ideal itself. In Section 2 we give a short
introduction to the algebraic geometry of neural codes. Section 3 describes a new and improved
algorithm to compute the canonical form of a neural ideal from a neural code. Finally, Section 4
provides a tour of the functionality within our code, centered on the canonical form algorithm.
2. Background
In this section we give a brief introduction to the neural ring and the neural ideal of a neural
code. A more thorough background, including necessary theorems and proofs, can be found in [2].
2.1. Neural Codes and the Neural Ideal. A neural code C ⊂ {0, 1}n is a set of binary strings
that represent neural activity. A ‘1’ represents a firing neuron, while a ‘0’ represents an idle neuron.
For example, the presence of the word 1011 in a 4-neuron code would indicate an instance when
neurons 1, 3, and 4 were firing but neuron 2 was not. Given a neural code C ⊂ {0, 1}n , the
corresponding neural ideal JC in the ring F2 [x1 , . . . , xn ] is defined by the following set of generators.
For any v ∈ {0, 1}n , consider the polynomial
Date: October 3, 2016.
1
2
E. PETERSEN, N. YOUNGS, R. KRUSE, D. MIYATA, R. GARCIA, AND L. D. GARCÍA PUENTE
ρv =
n
Y
(1 − vi − xi ) =
i=1
Y
xi
{i|vi =1}
Y
(1 − xj ).
{j|vj =0}
Notice that ρv (x) acts as a characteristic function for v, since it satisfies ρv (v) = 1 and ρv (x) = 0
for any x 6= v ∈ {0, 1}n . The neural ideal JC ⊆ F2 [x1 , . . . , xn ] associated to the neural code C is the
ideal generated by the polynomials ρv with v ∈
/ C, that is,
JC = hρv | v ∈
/ Ci.
2.2. Realizations and the Canonical Form. Many systems of neurons react to stimuli which
have a natural geographic association. One example is head direction cells in rats, where each
neuron responds to a preferred range of angles; these stimuli come from the 1-dimensional set
of possible angles for the head. Another example is place cells (in rats), where each neuron is
associated to a place field or region of the rat’s 2-dimensional environment. In such a geographic
setup, we would assume that if two neurons are observed to fire together, then the sets of stimuli
for these neurons must overlap. The idea of a realization for a code formalizes these notions.
Suppose U = {U1 , . . . , Un } is a collection of open sets with each Ui ⊂ X. Here, X ⊂ Rn represents
the space of possible stimuli, and Ui is the receptive field of the ith neuron, the set of stimuli which
will cause that neuron to fire. We say that U is a realization for a code C, or that C = C(U ), if
\
[
C = {v ∈ {0, 1}n | (
Ui )\
Uj };
vi =1
vj =0
that is, C represents the set of regions defined by U . It turns out that by considering the ideal
JC for the code C, we can determine complete information about the interaction of the Ui in any
realization U of C. This is facilitated by the canonical form of the neural ideal JC .
A polynomial f ∈ F2 [x1 , . . . , xn ] is a pseudo-monomial if f has the form
Y Y
(1 − xj ),
xi
f=
j∈τ
i∈σ
where σ ∩ τ = ∅. An ideal J ⊂ F2 [x1 , . . . , xn ] is a pseudo-monomial ideal if J can be generated by
a set of finitely many pseudo-monomials. Let J ⊂ F2 [x1 , . . . , xn ] be an ideal, and f ∈ J a pseudomonomial. Then f is a minimal pseudo-monomial of J if there is no other pseudo-monomial g ∈ J
with deg(g) < deg(f ) such that f = gh for some h ∈ F2 [x1 , . . . , xn ]. The canonical form of a
pseudo-monomial ideal J, denoted CF(J), is the set of all minimal pseudo-monomials of J.
The set CF(J) is unique for any given pseudo-monomial ideal J and J = hCF(J)i. It is important
to note that even though CF(J) is made up of minimal pseudo-monomials, it does not necessarily
imply that CF(J) is a minimal set of generators for J. The neural ideal JC for a neural code C is
a pseudo-monomial ideal since JC = hρv | v ∈
/ Ci and each of the ρv ’s is a pseudo-monomial. The
following theorem describes the set of relations on any realization U of C which CF(J) provides.
Theorem 2.1. Let C ⊂ {0, 1}n be a neural code, and let U = {U1 , . . . , Un } be any
T collection of
open sets in a stimulus space X such that C = C(U ). Given σ ⊆ {1, . . . , n}, let Uσ = i∈σ Ui . Then
the canonical form of JC is:
JC = {xσ | σ is minimal w.r.t. Uσ = ∅} ,
)
(
Y
[
[
(1 − xi ) | σ, τ 6= ∅, σ ∩ τ = ∅, Uσ 6= ∅,
Ui 6= X, and σ, τ are minimal w.r.t. Uσ ⊆
Ui ,
xσ
i∈τ
i∈τ
i∈τ
)
(
Y
[
(1 − xi ) | τ is minimal w.r.t. X ⊆
Ui
.
i∈τ
i∈τ
NEURAL IDEALS IN SAGEMATH
3
We call the above three (disjoint) sets of relations comprising CF(JC ) the minimal Type 1, Type
2 and Type 3 relations, respectively. Since the canonical form is unique, by Theorem 2.1, any
receptive field representation of the code C = C(U ) satisfies the following relationships:
Type 1: xσ ∈ CF(JC ) implies Uσ = ∅, but all intersections Uγ where γ ( σ are non-empty.
Q
S
Type 2: xσ S i∈τ (1−xi ) ∈ CF(JC ) implies Uσ ⊆ i∈τ Ui ,S
but no lower-order intersection is contained
in i∈τ Ui and all the Ui s are needed for Uσ ⊆ i∈τ Ui .
Q
S
S
Type 3: i∈τ (1 − xi ) ∈ CF(JC ) implies X ⊆ i∈τ Ui , but X 6⊆ i∈γ Ui for any γ ( τ .
3. The Iterative Algorithm
In [2], the authors provided a first algorithm to obtain the canonical form via the primary
decomposition of the neural ideal. Here we present an alternative algorithm. Rather than using
the primary decomposition, this algorithm begins with the canonical form for a code consisting of
a single codeword, and iterates by adding the remaining codewords one by one and adjusting the
canonical form accordingly. We describe the process for adding in a new codeword in Algorithm 1.
Algorithm 1: Iterative step to update a given canonical form after adding a code word c
Input : CF(JC ) = {f1 , . . . , fk }, where C ⊂ {0, 1}n is a code on n neurons, and a codeword
c ∈ {0, 1}n
Output: CF(JC∪{c} )
begin
L ←− {}, M ←− {}, N ←− {}
for x ←− 1 to k do
if fi (c) = 0 then
L ←− L ∪ {fi }
else
M ←− M ∪ {fi }
end
end
for f ∈ M do
for j ←− 1 to n do
if f (xj − cj ) is not a multiple of an element of L and (xj − cj − 1) ∤ f then
N ←− N ∪ {f (xj − cj )}
end
end
end
return L ∪ N = CF(JC∪{c} )
end
In summary, each pseudo-monomial f from CF(JC ) for which f (c) = 0 is automatically in the new
canonical form CF(JC∪{c} ). For those pseudo-monomials f ∈ CF(JC ) with f (c) = 1, we consider
the product of f with all possible linear terms (xj − cj ) (so this product will be 0 when evaluated
at c) as a possible candidate for CF(JC ); but we remove any such products which are redundant.
Here, a redundant pseudo-monomial is one which is either a multiple of another already known to
be in the canonical form, or which is a multiple of a Boolean polynomial.
Certainly, any polynomial f output by this algorithm will have the property that f (v) = 0 for
all v ∈ C ∪ {c}. A proof that this algorithm outputs exactly CF(JC∪{c} ) is found in the Appendix.
4
E. PETERSEN, N. YOUNGS, R. KRUSE, D. MIYATA, R. GARCIA, AND L. D. GARCÍA PUENTE
We have developed SageMath code that computes the canonical form using this iterative algorithm; see Section 4 for an in-depth tutorial of our package. We find that this iterative algorithm
performs substantially better than the original algorithm from [2]. We have also implemented our
algorithm in Matlab [6].
Table 1 displays some runtime statistics regarding our iterative canonical form algorithm. These
runtime statistics were obtained by running our SageMath implementation on 100 randomly generated sets of codewords for each dimension n = 4, . . . , 10. These computations were performed
on SageMath 7.2 running on a Macbook Pro with a 2.8 GHz Intel Core i7 processor and 16 GB
of memory. We performed a similar test for our implementation of the original canonical form
algorithm in [2] and on the Matlab implementation of our iterative method. However, even in
dimension 5 the original algorithm performs poorly. In our tests, we found several codes for which
the original algorithm took hundreds or even thousands of seconds to compute the canonical form.
For example, the iterative algorithm takes 0.01 seconds to compute the canonical form of the code
below, but the original method takes 1 hour and 8 minutes to perform the same computation.
10000, 10001, 01011, 01010, 10010, 01110, 01101, 01100, 11111,
11010, 11011, 01000, 01001, 00111, 00110, 00001, 00010, 00011, 00101.
We also found several codes on dimension 5 for which the original algorithm halted due to lack
of memory. One such code is
01110, 01111, 11010, 11100, 11101, 01011, 01000, 00110, 00001, 00011, 10011, 00100, 00101, 10111.
The iterative algorithm computes the canonical form of the previous code in 0.0089 seconds. In
our Matlab implementation we also found several examples in dimension 6 for which the canonical
form took thousands of seconds to be computed.
Dimension min
max
4
0.000077 0.0034
5
0.000087 0.014
6
0.00012 0.108
7
0.00012 0.621
8
0.000097 4.011
9
0.698
39.28
10
0.229
350.5
Table 1. Runtime statistics (in seconds) for
mean median std
0.0016 0.0018 0.00076
0.0076 0.0082 0.0034
0.049 0.051
0.024
0.298 0.323
0.135
1.964 2.276
1.036
24.86 27.38
9.976
237.45 271.3
87.1
the iterative CF algorithm in SageMath.
Our code has many features beyond the computation of the canonical form via the new iterative
algorithm. In the following tutorial, we show how to use the code to compute any of the following:
(1) The neural ideal.
(2) The canonical form for a neural ideal via the new iterative algorithm.
(3) The canonical form for a neural ideal via the primary decomposition algorithm.
(4) A tailored method to compute the primary decomposition of pseudo-monomial ideals.
(5) Gröbner bases and the Gröbner fan of a neural ideal.
(6) A method to test whether a code is a simplicial complex.
4. SageMath Tutorial
We decided to develop our code on SageMath due to its open-source nature, extensive functionality, and ease of use [5]. In this tutorial, we’ll begin with installing the NeuralCodes package.
Then, we’ll walk through the most important functions of the package.
NEURAL IDEALS IN SAGEMATH
5
4.1. Installation. We will assume that SageMath is properly installed on the system. In our
tutorial, the files, iterative canonical.spyx, neuralcode.py and examples.py, are downloaded
in the folder NeuralIdeals. Now, we can load the package by:
sage: load("NeuralIdeals/iterative_canonical.spyx")
sage: load("NeuralIdeals/neuralcode.py")
sage: load("NeuralIdeals/examples.py")
The first file contains the iterative algorithm in Cython, so loading will compile the code. Note
that they must be loaded in this order, as the Cython file must be compiled for the tests in
neuralcode.py to run. The second file, neuralcode.py, holds all of the code that we will be
demonstrating. The third, examples.py has some additional examples that can be loaded with
sage: neuralcodes(). We expect to include our code in SageMath, but currently it can be easily
obtained at https://github.com/e6-1/NeuralIdeals. We also want to note that running the
package in the free SageMathCloud (https://cloud.sagemath.com) is even easier. One has to
create a new project, upload all three files above and then simply run load('neuralcode.py').
4.2. Examples. First, we define a neural code:
sage: neuralCode = NeuralCode(['001','010','110'])
Now, we can perform a variety of useful operations. We can compute the neural ideal:
sage: neuralIdeal = neuralCode.neural_ideal()
sage: neuralIdeal
Ideal (x0*x1*x2 + x0*x1 + x0*x2 + x0 + x1*x2 + x1 + x2 + 1, x0*x1*x2 + x1*x2,
x0*x1*x2 + x0*x1 + x0*x2 + x0, x0*x1*x2 + x0*x2, x0*x1*x2) of Multivariate
Polynomial Ring in x0, x1, x2 over Finite Field of size 2
We can compute the primary decomposition using a custom algorithm:
sage: pm_primary_decomposition(neuralIdeal)
[Ideal (x2 + 1, x1, x0) of Multivariate Polynomial Ring in x0, x1, x2
over Finite Field of size 2,
Ideal (x2, x1 + 1) of Multivariate Polynomial Ring in x0, x1, x2
over Finite Field of size 2]
We can compute the canonical form of the neural ideal.
sage: canonicalForm = neuralCode.canonical()
sage: canonicalForm
Ideal (x1*x2, x1*x2 + x1 + x2 + 1, x0*x1 + x0, x0*x2) of
Multivariate Polynomial Ring in x0, x1, x2 over Finite Field of size 2
The method canonical() will use the iterative algorithm by default. If we want to use the
procedure described in [2], one needs to specify it with canonical(algorithm="original"). This
procedure uses by default the tailored pseudo-monomial primary decomposition, but one can make
this explicit with canonical(algorithm="original", decomposition_algorithm="pm").
sage: neuralCode.canonical(algorithm="original", decomposition_algorithm="pm")
Ideal (x1*x2, x0*x1 + x0, x1*x2 + x1 + x2 + 1, x0*x2) of
Multivariate Polynomial Ring in x0, x1, x2 over Finite Field of size 2
Besides the tailored pseudo-monomial primary decomposition method, we can also use the standard primary decomposition methods implemented in SageMath such as Shimoyama-Yokoyama
and Gianni-Trager-Zacharias with the flags "sy" and "gtz", respectively. Table 2 compares the
runtimes (in seconds) for the primary decomposition of neural ideals using these three methods.
Each entry in this table is the mean value of the runtimes for 50 randomly generated codes.
6
E. PETERSEN, N. YOUNGS, R. KRUSE, D. MIYATA, R. GARCIA, AND L. D. GARCÍA PUENTE
Dimension Pseudo-Monomial PD Shimoyama-Yokoyama Gianni-Trager-Zacharias
4
0.16
0.028
0.027
5
0.85
0.14
0.07
6
6.43
2.98
0.27
7
63.1
74.9
1.9
8
578
3040
45
Table 2. Comparisons of runtimes for primary decompositions of neural ideals.
We found that, in general, the custom pseudo-monomial primary decomposition algorithm does
not outperform the Gianni-Trager-Zacharias algorithm. Nevertheless, this procedure is implemented to be used in characteristic 0. It also works in fields of large positive characteristic. But in
small characteristic this procedure may not terminate.
We also want to observe that the canonical() method returns an Ideal object whose generators
are not factored and hence not easy to interpret in our context. In order to obtain the generators
in the canonical form in factored form, we use factored_canonical():
sage: neuralCode.factored_canonical()
[x2 * x1, (x1 + 1) * x0, (x2 + 1) * (x1 + 1), x2 * x0]
From this output we can easily read off the RF structure of the neural code. However, we also
provide a different command that parses the output to explicitly describe the RF structure.
sage: neuralCode.canonical_RF_structure()
Intersection of U_['2', '1'] is empty
X = Union of U_['2', '1']
Intersection of U_['0'] is a subset of Union of U_['1']
Intersection of U_['2', '0'] is empty
We can also compute the Gröbner basis and the Gröbner fan of the neural ideal. Note that we
could compute the neural ideal and use the built in groebner_basis() method, but that approach
will not impose the conditions of the Boolean ring (i.e. x2 + x = 0). Our method uses the built-in
groebner_basis() command but also reduces modulo the Boolean equations.
sage: neuralCode.groebner_basis()
Ideal (x0*x2, x1 + x2 + 1) of Multivariate Polynomial Ring in x0, x1, x2
over Finite Field of size 2
sage: neuralIdeal.groebner_basis()
[x0*x2, x1 + x2 + 1, x2^2 + x2]
sage: neuralCode.groebner_fan()
[Ideal (x1 + x2 + 1, x0*x2) of Multivariate Polynomial Ring in x0, x1, x2
over Finite Field of size 2]
A neural code is called convex if its codewords correspond to regions defined by an arrangement
of convex open sets in Euclidean space. Convex codes have been observed experimentally in many
brain areas. Hence there has been increased interest in understanding what makes a neural code
convex [1]. It has also been observed that if a code is a simplicial complex then it is convex. One
can check if a code is a simplicial complex with the command is_simplicial(Codes):
sage: is_simplicial(['001','010','110'])
False
sage: is_simplicial(['000','001','010','100','110','011','101','111'])
True
NEURAL IDEALS IN SAGEMATH
7
Appendix: Proof of Iterative Algorithm
Here, we show that the process described in Algorithm 1 gives CF(JC∪{c} ) from CF(JC ) and c.
Throughout, we use the following conventions and terminology: C and D are neural codes on the
same number of neurons; so, C, D ⊆ {0, 1}n . A monomial xα is square-free if αi ∈ {1, 0} for all
i = 1, . . . , n. A polynomial is square-free if it can be written as the sum of square-free monomials.
For example: x1 x2 + x4 + x1 x3 x2 is square-free. There is a unique square-free representative of
every equivalence class of F2 [x1 , . . . , xn ]/hxi (1 − xi )i. For h ∈ F2 [x1 , . . . , xn ], let hR denote this
unique square-free representative of the equivalence class of h in F2 [x1 , . . . , xn ]/hxi (1 − xi )i.
Then, for CF(JC ) = {f1 , . . . , fr } and CF(JD ) = {g1 , . . . , gs }, we define the set of reduced products
def
P (C, D) = {(fi gj )R | i ∈ [r], j ∈ [s]}.
Note that as pseudo-monomials are square-free, for each pair i, j we have either (fi gj )R = 0 or
(fi gj )R is a multiple of both fi and gj . We define the minimal reduced products as
def
MP(C, D) = {h ∈ P (C, D) | h 6= 0 and h 6= f g for any f ∈ P (C, D), deg g ≥ 1}.
Lemma 1. If C, D ⊂ {0, 1}n , then the canonical form of their union is given by the set of minimal
reduced products from their canonical forms: CF(JC∪D ) = M P (C, D).
Proof. First, we show MP(C, D) ⊆ JC∪D . For any h ∈ MP(C, D), there is some fi ∈ CF(JC ) and
gj ∈ CF(JD ) so h = (fi gj )R . In particular, h ∈ JC as h is a multiple of fi , and h ∈ JD as it is a
multiple of gj . Thus h(c) = 0 for all c ∈ C ∪ D, so h ∈ JC∪D .
Suppose h ∈ CF(JC∪D ). Then as JC∪D ⊂ JC , there is some fi ∈ CF(JC ) so that h = h1 fi , and
likewise there is some gj ∈ CF(JD ) so h = h2 gj where h1 , h2 are pseudo-monomials. Thus h is a
multiple of (fi gj )R and hence is a multiple of some element of MP(C, D). But as every element
of MP(C, D) is an element of JC∪D , and h ∈ CF(JC∪D ), this means h itself must actually be in
MP(C, D). Thus, CF(JC∪D ) ⊆ MP(C, D). For the reverse containment, suppose h ∈ MP(C, D);
by the above, h ∈ JC∪D It is thus the multiple of some f ∈ CF(JC∪D ). But we have shown that
f ∈ MP(C, D), which contains no multiples. So h = f is in CF(JC∪D ).
Proof of Algorithm. Note that if c ∈ C, then L = CF(JC ), so the algorithm ends immediately
and outputs CF(JC ); we will generally assume c ∈
/ C.
To show that the algorithm produces the correct canonical form, we apply Lemma 1; it suffices
to show that the set L ∪ N is exactly MP(C, {c}). This requires that all products are considered,
and that we remove exactly those which are multiples or other elements, or zeros. Note that
CF(J{c} ) = {xi − ci | i ∈ [n]}.
To see that all products are considered we will look at L and M separately. Let g ∈ L. Since
g(c) = 0, we know (g · (xi − ci ))R = g for at least one i. So g ∈ MP(C, {c}). Any other product
(g · (xj − cj ))R will either be 0, g, or a multiple of g, and hence will not appear in MP(C, {c}).
Thus, all products of linear terms with elements of L are considered, and all multiples or zeros are
removed. It is impossible for elements of L to be multiples of one another, as L ⊂ CF(C).
We also consider all products of elements of M with the linear elements of CF(J{c} ). We discard
them if their reduction would be 0, or if they are a multiple of anything in L. If neither holds, we
add them to N . So it only remains to show that no element of N can be a multiple of any other
element in N , and no element of N can be a multiple of anything in L, and thus that we have
removed all possible multiples. First, no element of N may be a multiple of an element of L, since
if g ∈ L, f · (xi − ci ) ∈ N , and f · (xi − ci ) · p = g for some pseudo-monomial p, then f g. But this is
impossible as f, g are both in CF(JC ). Now, suppose f · (xi − ci ) = h · g · (xj − cj ) for f, g ∈ CF(JC )
and f · (xi − ci ), g · (xj − cj ) ∈ N , and h a pseudo-monomial. Then as f 6 |g and g 6 |f , we have i 6= j,
and so (xj − cj ) f . But this means f · (xj − cj ) = f and therefore f ∈ L, which is a contradiction.
So no elements of N may be multiples of one another.
8
E. PETERSEN, N. YOUNGS, R. KRUSE, D. MIYATA, R. GARCIA, AND L. D. GARCÍA PUENTE
References
[1] Curto, C., Gross, E., Jeffries, J., Morrison, K., Omar, M., Rosen, Z., Shiu, A. and Youngs, N. “What makes a
neural code convex?” arXiv preprint arXiv:1508.00150 (2015).
[2] Curto, C., Itskov, V., Veliz-Cuba, A., Youngs, N. “The Neural Ring: An Algebraic Tool for Analyzing the
Intrinsic Structure of Neural Codes.” Bulletin of Mathematical Biology, Volume 75, Issue 9, pp. 1571-111, 2013
[3] Curto, C., and Youngs, N. “Neural ring homomorphisms and maps between neural codes.” arXiv preprint
arXiv:1511.00255 (2015).
[4] O’Keefe, J., and Dostrovsky, J. “The hippocampus as a spatial map. Preliminary evidence from unit activity in
the freely-moving rat.” Brain research 34, 171-175. 1971.
[5] Sage Mathematics Software (Version 7.2.0), The Sage Developers, 2016, http://www.sagemath.org.
[6] Youngs,
N.
Neural
ideal:
a
Matlab
package
for
computing
canonical
forms.
http://github.com/nebneuron/neural-ideal, 2015.
(E. Petersen) Department of Mathematics, Rose-Hulman Institute of Technology, Terre Haute, IN
47803
E-mail address: [email protected]
(N. Youngs) Department of Mathematics and Statistics, Colby College, Waterville, ME 04901
E-mail address: [email protected]
(R. Kruse) Mathematics Department, Central College, Pella, IA 50219
E-mail address: [email protected]
(D. Miyata) Mathematics Department, Willamette University, Salem, OR 97301
E-mail address: [email protected]
(R. Garcia and L. D. Garcı́a Puente) Department of Mathematics and Statistics, Sam Houston State
University, Huntsville, TX 77341-2206
E-mail address: [email protected]
E-mail address: [email protected]
| 0math.AC
|
Under review as a workshop contribution at ICLR 2015
AUDIO S OURCE S EPARATION U SING A D EEP AUTOEN CODER
arXiv:1412.7193v1 [cs.SD] 22 Dec 2014
Giljin Jang
School of Electronics Engineering
Kyungpook National University
Daegu, Republic of Korea
[email protected]
Han-Gyu Kim & Yung-Hwan Oh
Computer Science Department
Korea Advanced Institute of Science and Technology (KAIST)
Daejeon, Republic of Korea
{hgkim,yhoh}@cs.kaist.ac.kr
A BSTRACT
This paper proposes a novel framework for unsupervised audio source separation
using a deep autoencoder. The characteristics of unknown source signals mixed
in the mixed input is automatically by properly configured autoencoders implemented by a network with many layers, and separated by clustering the coefficient vectors in the code layer. By investigating the weight vectors to the final
target, representation layer, the primitive components of the audio signals in the
frequency domain are observed. By clustering the activation coefficients in the
code layer, the previously unknown source signals are segregated. The original
source sounds are then separated and reconstructed by using code vectors which
belong to different clusters. The restored sounds are not perfect but yield promising results for the possibility in the success of many practical applications.
1
I NTRODUCTION
In audio analysis applications such as speech recognition and audio-text alignment, the clean target sound helps improve the quality of analysis while noisy input may disturb the analysis process. In a situation that we only have recordings from a single sensor, the problem is very difficult
and hence no general solution has been found. Masking in the spectro-temporal region was proposed in Hu & Wang (2004) where the separation mask was constructed by estimated speech pitch.
Such method worked well in extracting speech from noisy environment, but the performance is
not guaranteed if the target source is not a speech signal. Several separation methods based on
non-negative matrix factorization (NMF) was proposed to solve the monaural source separation
problem (Raj et al., 2010). NMF utilized the redundancy of the sound spectrogram for source separation. NMF-based methods works for various types of sources. However, the mixing is assumed to
be a non-negative linear mixing, and it cannot handle complex sound sources.
In this paper, we propose applying a deep autoencoder (Hinton & Salakhutdinov, 2006) to the singlechannel, source separation problem. The autoencoders constructed by networks with many hidden
layers have been successfully adopted in many applications, such as image and audio denoising,
speech recognition, data compression, and so on. Unlike other applications, our proposed method
tries to apply the autoencoder to solve the problem of unsupervised audio source separation. The
characteristics of sound sources in the input mixture are learned and stored in a deep autoencoder,
and an appropriate source separation algorithm based on unsupervised learning is proposed. Experimental results on mixtures of 5 types of music and 2 types of speech signals suggest that the coefficients in the encoding layer is useful in distinguishing and separating the unknown target sources.
The detailed descriptions are in the following sections.
1
Under review as a workshop contribution at ICLR 2015
2
S OURCE
SEPARATION USING AUTOENCODER
In order to learn an autoencoder for audio source mixtures, we first apply short-time Fourier analysis
to the time-domain audio signal, resulting magnitude spectrum matrix denoted by Xc,m , where c is
the frequency channel index and m the temporal frame index. For the input to the autoencoder, we
construct a rectangular window with consecutive frequency channel and time index, such that
Wi,j = {Xc,m |i ≤ c < i + h, j ≤ m < j + l},
(1)
where Wi,j is the window starting from frequency channel i and frame j of the magnitude spectra,
where h and l are the height and length of the windows, respectively. The rectangular windows are
then unrolled into supervectors for the autoencoder training. The data set is generated by shifting the
window one frame in time axis or one frame in frequency axis from all the input spectrum matrix.
We designed an autoencoder with 5 hidden layers, which have 50, 18, 6, 18, and 50 units, respectively, as shown in Figure 1.
Figure 1: The structure of autoencoder for source separation.
Because there is no labeled training data for the unknown sources, a k-means clustering is performed
on the 6-dimensional coefficient vectors of the middle layer. The whole process of the clustering is
shown in Figure 2. In this method, we consider the feature vectors of the windows extracted using
the autoencoder from different clusters for different sound sources. According to the clustering
result of the feature vectors, the windows are also classified into the clusters, and original audio
sources are reconstructed using the vectors in the given cluster only.
#! $% & ' (
(!& %) " (
!"
* + & !
%!"
* + % !"
.+ )/ "" " !" 0
$! $% ! 0 -12 %
,) " ( ' % -
.+ )/ "" " !" 0
$! $% ! 0 345 %
,) " ( ' % 3
777
777
.+ )/ "" " !" 0
$! $% ! 0 26 %
,) " ( ' %
Figure 2: The block diagram of the clustering-based source separation method.
3
E XPERIMENTAL
RESULTS
To evaluate the performance of the proposed method, source separation experiments were carried
out on a mixtures of 5 different music sounds and 2 different speech sources. The speech sources are
2
Under review as a workshop contribution at ICLR 2015
selected from TIMIT speech corpus, and music sounds are jazz, drum, acoustic guitar, electric guitar
and piano. These speech and music sounds were mixed together, generating 2 × 5 = 10 mixtures in
total. The audio sounds are sampled at 8 kHz, and only 8 seconds are used in training the respective
autoencoders. The spectrogram matrix of the mixtures are obtained by short-time Fourier analysis
with frame length 40 ms and shift size 10 ms. For window generation in Equation 1, h = 30 and
l = 5 were used, which were decided empirically. Figure 3 shows 10 selected examples from the
total 50 weight vectors connecting all the nodes in the last hidden layer to one of the nodes in the
final output layer. The input is male speech and jazz music mixture. Four of them represents the
change of the spectral peaks over time: third, fourth, fifth, and seventh, which reflect the temporal
change of harmonics of speech signals. The remaining six vectors are composed of straight lines
over time, modeling the stationary frequency components of music signals. Figure 4 represents the
mask constructed by clustering all the code vectors of dimension 6 in the middle layer. The first and
the second clusters are jazz music, with slowly changing spectral peaks, and the third and the fourth
clusters are mostly speech sounds. These results show that the validity of the proposed method.
Figure 3: Example windows constructed by the weights to the last target layer for jazz and male
speech mixture.
2
0
0
2
4
6
time (seconds)
8
C3
0.5
2
0
0
2
4
6
time (seconds)
8
C4
freq (kHz)
0.5
freq (kHz)
C2
freq (kHz)
freq (kHz)
C1
0.5
2
0
0
2
4
6
time (seconds)
8
0.5
2
0
0
2
4
6
time (seconds)
8
Figure 4: Separation masks constructed by reconstructed vectors in each of 4 clusters for jazz and
male speech mixture.
4
C ONCLUSIONS
An application of autoencoders to audio source separation is proposed. The spectrum matrix obtained by short-time Fourier analysis is used for the initial representation of the mixed source signal,
and a deep autoencoder is used to represent the spatio-temporal local parts of the spectrum matrix.
The code coefficients in the middle layer of the autoencoder is used as a feature for distinguishing
audio sources. The main contribution of the proposed method is that the characteristics of unknown
sources are extracted from the mixed signals. Although the experimental results are not complete
yet, ongoing efforts are being made to improve the proposed method.
ACKNOWLEDGMENTS
This work was supported by the Basic Science Research Program through the National Research
Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (no.
NRF-2010-0025642).
R EFERENCES
Hinton, Geoffrey E and Salakhutdinov, Ruslan R. Reducing the dimensionality of data with neural
networks. Science, 313(5786):504–507, 2006.
Hu, Guoning and Wang, DeLiang. Monaural speech segregation based on pitch tracking and amplitude modulation. IEEE Transactions on Neural Networks, 15(5):1135–1150, 2004.
Raj, Bhiksha, Virtanen, Tuomas, Chaudhuri, Sourish, and Singh, Rita. Non-negative matrix factorization based compensation of music for automatic speech recognition. In Proc. INTERSPEECH,
pp. 717–720, 2010.
3
| 9cs.NE
|
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
Multiple model approach and experimental validation of
a residential air to air heat pump.
François GARDE Research Engineer, Electricité de France / University of Reunion Island .*
Harry BOYER, Assistant Professor, University of Reunion Island*
Florence PIGNOLET, Researcher, University of Reunion Island*
Franck LUCAS, Researcher, University of Reunion Island.*
Jean BRAU, Professor, INSA de Lyon, CETHIL/ Thermique du Bâtiment†
Abstract :
The beginning of this work is the achievement of a design tool, which is a multiple model software
called « CODYRUN », suitable for professionnals and usable by researchers. The original aspect of this
software is that the designer has at his disposal a wide panel of choices between different heat transfer models
More precisely, it consists in a multizone software integrating both natural ventilation and moisture tranfers .
This software is developed on PC micro computer and gets advantage of the Microsoft WINDOWS front-end.
Most of time, HVAC systems and specially domestic air conditioners, are taken into account in a very
simplified way, or in a elaborated one. On one side,they are just supposed to supply the demand of cooling
loads with an ideal control loop (no delay between the sollicitations and the time response of the system), The
available outputs are initially the hourly cooling and heating consumptions without integrating the real
caracteristics of the HVAC system This paper is also following the same multiple model approach than for the
building modelling by defining different modelling levels for the air conditionning systems, from a very
simplified one to a detailled one.
An experimental validation is achieved in order to compare the sensitivity of each defined model and
to point out the interaction between the thermal behaviour of the envelop and the electrical system
consumption. For validation purposes, we will describe the data acquisition system. and the used real size test
cell located in the University of Reunion island, Indian Ocean.
1. INTRODUCTION
The main problem of Demand Side Management in islands such as Reunion island is
that the means of electrical production are restricted. The demand of power supply is not so
important to justify a nuclear power plant, and the means of production are more expensive
than the nuclear. In that case, cooling of buildings and interaction between the envelop and the
cooling system is a growing problem because air conditioning systems consume a large amount
of energy and generate investments in power supply. In addition, building envelops suffer from
a bad thermal conception that generates high cooling loads. The large population increase in
the tropical islands, the rise in the living standards and the decreasing costs of air-conditioning
units constitute a real energetic problem.
The modeling and simulation of HVAC systems , and the interaction with the building
is also necessary for the evaluation of means for reducing energy use.
We focuse in our studies on the modelling of a residential air-to-air split-system heatpump, which is the more widespread kind of air conditioning system in tropical islands. The
problem in heat-pump modelling is the level accuracy of the model. Most of time, HVAC
systems and specially domestic air conditioners, are taken into account in a very simplified
way, or in a elaborated one. The question is : Untill what kind of degree of complexity the
heat-pump can be modelled for a further integration in a thermal and airflow simulation
software. According to Irving [15], there is an important number of simulations softwares of
varying degrees of complexity available in the industry. These programs range from very
*University of Reunion Island, Faculté des Sciences, Laboratoire de Génie Industriel, BP 7151, 15 av. René
Cassin, 97 715 Saint-Denis Messag. Cedex 9, France – e mail : [email protected]
†INSA Lyon, CETHIL,Thermique du Bâtiment, Bât 307, 20 av. Albert Einstein, 69621 Villeurbanne cedex
1
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
simple steady-state heat or cooling loads calculation running on programmable calculators, to
very detailed simulation programs which require large amounts of computer power. It is
obvious that no single program can satisfy the many needs of the engineer at every stage of his
design.
Another point to take into account is the time step of the calculation programs. It is
often of one hour, what is too long to take into account the HVAC systems control, because
their time constant are shorter than the building ones. It is apparent that optimal cooling
control strategies require a dynamic model for predicting the performance characteristics of the
overall system [10]. At last, the cooling coil of the indoor unit of the split-system is one of the
key component in air conditionning system. The main difficulties are located in wet and/or
transient regimes and the performance of the whole system (and thus the accuracy of the
model) depends on a good model of the cooling coil .
2. MODELLING REVIEW
Concerning heat-pump simulation, it appears that the different models can be broadly
classified as steady and dynamic state simulations. Weslby [25] reviewed with particular
emphasis on their bases and end-uses various mathematical models on mechanical vapourcompression heat-pumps since from 1975 to 1988.
It appears that detailed steady state models are often used to study variations in system
or component configuration in order to identify optimal values of parameters that dictate the
performance of the system. They are very helpful to optimize the design of air conditioning
equipments. The component approach is often used by researchers because of the strong
coupling between them. [2]. In a general way, authors have developped a numerical model to
determine the steady-state performance of heat-pumps for specified source and sink conditions
and specified components. The model is based on simple models of the components of the heat
pump. Considering the coupling with a thermal simulation software, Loveday [16] has
developped a heat-pump simulation programme in order to relate the hourly thermal building
loads to the corresponding electricity inputs to the heat-pump, which is a function of its COP.
The hourly heating energy to be supplied to the building is obtained from ESP [9]. The heatpump model is a steady-state model which relates heat-pump COP to external ambient
temperature, and accounts for motor/compressor inefficiencies, thermal/mechanical losses of
the heat transport system, and defrost performance.
These steady-state models are most of time too detailed for the use we need. Other
investigators have worked on steady-state models based on experimental data or manufactured
data [1], [12], [24]. From the knowing of differents set points and different experimental
conditions, a model based on polynomial laws relating the cooling capacity or the power
consumptions to the outdoor and indoor temperature can be elaborated.
Transient perturbations and dynamic regimes could be simulated to analyse the start-up
process, defrost-cycle and rapidly varying operating conditions. The aim of these models is to
study the variation of working fluid mass and thermodynamic state distributions within the
system from known initial conditions, when the system operating conditions are changed.
Some reseachers have developped rigourous dynamic models [8], [19], [17], [23]. These
models have ranged from lumped parameters [19], [23] to mathematical models based on a
multi-node or distributed approach [17]. These models tend to be complex, require large
computers to use, and do not readily yield physical insight into the variables affecting system
2
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
performance. However, they do allow the user to track detailed information (temperatures,
pressures, etc.) in the system during the transient start-up process.
Other investigators [18] have suggested a two-time constant model to capture the
physical phenomena responsible for strat-up losses. One time constant would capture capacity
delay due to the mass of the heat exchanger. The second time constant would capture a very
high, but rapidly decaying, initial capacity that is produced by the time lag for the refrigerant to
be pumped from the evaporator to the rest of the system. This model should provide a better
fit to the experimental data during the start-up process.
Other investigators have hypothized that during start-up, the system capacity could be
modeled as a first-order (single-time constant) system [7], [13], [20], [21], [22]. Experimental
data from Murphy [20] showed good agreement between a single time constant model and
data from a heat-pump. We also verify that hypothesis by ourselves and have found time
constant of 2 minutes. Murphy found time constants for cooling mode operations ranging from
0.32 to 0.47 minutes for a heat-pump and an air-conditionner respectively. O’Neal [22] found
time constant of more than 2 minutes for a heat-pump operating in cooling mode. Another
important assumption made by Glosdsmith [13] and O'Neal [21] was the constant power upon
star-up. This assumption seems to be reasonable. While there is an initial surge in power during
the few seconds after start-up, the power is relatively constant during start-up. We also verify
that point by our own experiments and didn't notice a high surge in power after start-up of the
split-system. Finally, because the aim of our study was to take into account the exponential
increasing of cooling capacity and not specially the mecanisms affecting cycling, the choice was
made to use a single-time constant model.
Considering the coupling effects of cooling and deshumidifying on a cooling coil,
several authors have proposed different models of coil in steady-state and/or transient
conditions. Hirsch [14] considers that the moisture condensation on cooling coils is simulates
by characterizing the coils by their bypass factors and solving the bypass relation with the
system moisture balance in steady-state conditions. Xin Ding [26] has worked on different
models of cooling coil in steady-state and transient conditions, and in dry or wet regimes.
These models are based on the NTU method and on the determination of the effectiveness of
the coil in wet and dry regimes. The dynamic model of the coil is supposed to be a first order
model.
3. CODYRUN AND THE
INTEGRATED EQUATIONS FOR HVAC SYSTEMS
3.1 An overview of CODYRUN software :
CODYRUN is a thermal multizone software integrating both natural ventilation and
moisture transfers. Its main characteristics is to be a multiple model structure, allowing the
choice between a wide range of models of heat transfer and meteorological data reconstitution.
More information can be obtained about the software in [3] concerning multiple model aspect,
building description in [5], thermal model constitution in [4] and preliminary software
validation in [11].
Concerning the calculation, the main parts are the airflow model (pressure model taking
into account wind, thermal buoyancy and large openings), the thermal model and the moisture
transfer model. For the study presented in this paper, the two last models has been modified by
keeping the same multiple model concept, allowing finally different levels of modelisation for
the air cooling system.
3
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
3.2 Quick thermal and moisture transfer models constitution :
By considering usual assumptions as mono-dimensional heat conduction in walls, well
mixed air volumes, and linearized superficial exchanges, nodal analysis (or lumped capacities
analysis) leads to an electrical network. To simplify our discussion, we'll suppose that the heat
conduction is treated with the help of a model constituted of a thermal resistance and 2
capacitors, said “R2C” model (leading to no internal nodes in walls). Then, the thermal model
of the building is obtained as a set of equations of type from 1 to 4, traducing thermal balance
of inside (Tsi) and outside nodes (Tse) according to boundary conditions, thermo-convective
balance equation of dry-bulb air nodes (Tai) and radiative balance equation of the inside mean
radiant temperature nodes (Trm).
d Tsi
dt
d Tse
Cse
dt
d Tai
Cai
dt
Csi
hci (Tai Tsi ) hri (Trm Tsi ) K (Tse Tsi ) swi
(1)
hce (Tae Tse ) hre (Tae Tse ) K (Tsi Tse ) swe
( 2)
0
Nw
h
(Tai Tsi ( j ) c Q (Tae Tai )
h
A j (Tsi ( j ) Trm )
ci
j 1
Nw
ri
Q sens
(3)
( 4)
j 1
Calculating separatly moisture transfers (induced by air motion between indoor zones
and outdoor, but without latent storage in walls and room furniture), for a zone of specific
humidity rs , moisture balance leads to a linear system of equations, each being as
C
drs
in lv rs in m
out lv rs Q lat
m
dt
(5)
Q sens and Q lat are respectively is the sensible and latent capacities injected in the zone, even by
internal loads (lighting, occupants) or HVAC systems.
3.3 Integration of air cooling system :
At each time step, depending on indoor and outdoor conditions, in case of air cooling,
the HVAC model has to calculate Q sens and Q lat for the considered zones, values to be injected
in equations (3) and (5).
4. MODELS DEVELOPMENT
4.1 Model n°0
This model is just supposed to supply the demand of cooling loads with an ideal control
loop (no delay between the sollicitations and the time response of the system), The available
outputs are initially the hourly cooling consumptions without integrating the real caracteristics
of the HVAC system. We assume that sensible latent cooling rates are not dependant. The real
physical phenomenon are different beacuse both cooling and deshumidifying occur at the same
time.
4
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
The HVAC system must reach a temperature set point and a humidity set point. It
allows the user to determine the hourly sensible and coolings loads of the building, but does
not depend on some specific characteristics of the system.
In order to arrive at electric consumption, the system is modelled by its cooling Coefficient of
Performance (COP), which is the ratio between the cooling rate and the electric power. The
COP is taken constant in our case.
The COP is also a first comparison quality criteria for heat pump systems . It allows to
detemine the electric consumption of the system all along the simulation period.
4.2 Model n °1:
This model is more detailled than the first one. The time step was reduced to one
minute in order to take into account the on-off cycling and the control of the system.
Steady -state conditions :
The capacity rate is the nominal capacity of the system operating in steaty state and
standard conditions. (27°C dry bulb temperature, 19°C wet bulb temperature, 35°C outdoor
temperature). The system is defined in steady state conditions by its SHF (sensible heat factor)
which is the ratio between the sensible cooling rate and the total cooling rate. The SHF gives
also the sensible and latent capacities which will be integrated at each time step in the solving
equations of CODYRUN. These capacities are held constant each time the split is on.
Dynamic conditions :
As seen in the modelling review part, the system capacity is modeled as a first order
model, in order to take into account the start-up of the split-system [21], [22].
Figure 1 shows the idealized capacity during start-up and shutdown of a residential airconditioner.
The instantaneous cooling capacity, Qcyc
is given by :
Qss
(1e t / )
Qcyc
t
: steady state capacity
: time constant for the system
: time after start-up
Qcy/Qss
where Qcyc
(6)
1
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Time (minute)
Fig 1 : Dynamic behaviour during start-up condition.
5
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
System control :
An on/off controller with a dead zone has been developped. This kind of controller are
always used for small sized residential unit. Some experiments allowed us to find the correct
dead zone of the controller, which was found equal to 1°C.
Electric power :
The electric power of the system is assumed to be constant, even during start-up
conditions [13], [21], [22]. The user enters the cooling COP in nominal conditions in order to
calculate the electric consumption of the system.
4.3 Model n°2 :
The performances of an air-to-air heat-pump (total cooling capacity, sensible and latent
capacities, electric power) are strongly influenced by parameters such as the indoor dry air
temperature, the indoor relative humidity and the outdoor air temperature. In model 1, the
performances are supposed to be constant, but figure 1 points out that they depend on external
parameters. The aims of these models are to take into account these effects in steady-state
conditions.
Steady-state conditions :
The method is based upon manufacturer data given in Table 1. These data were
mesured during steady-state conditions. The indoor dry-bulb temperature was held constant.
Only parameters such as wet bulb temperature and outdoor temperature were changing. We
simply add to table 1 the values of the indoor air specific enthalpy.
CARRIER Indoor unit : 42 HWA - Outdoor unit : 38 SDF012
Indoor dry bulb temperature : 27°C
By Pass factor = 0.04
Wet bulb temp (°C)
16
18
Outdoor Indoor air specific
45
55
temp (°C) enthalpy (kJ/kg)
Total capacity
3.46
3.65
21
Sensible capacity
3.26
2.88
Power (kW)
0.99
1.00
Total capacity
3.34
3.54
25
Sensible capacity
3.20
2.84
Power (kW)
1.06
1.07
Total capacity
3.16
3.37
30
Sensible capacity
3.10
2.78
Power (kW)
1.14
1.15
Total capacity
2.95
3.18
35
Sensible capacity
2.95
2.70
Power (kW)kW
1.23
1.25
Total capacity
2.67
2.77
40
Sensible capacity
2.67
2.52
Power (kW)
1.33
1.34
Total capacity
2.49
2.56
45
Sensible capacity
2.49
2.43
Power (kW)
1.40
1.40
19
58
3.72
2.67
1.00
3.64
2.65
1.07
3.48
2.59
1.16
3.30
2.52
1.25
2.95
2.38
1.35
2.69
2.28
1.41
Air flow rate : 110 l/s
20
22
61.2
65
3.78
2.45
1.01
3.72
2.44
1.08
3.59
2.41
1.17
3.40
2.34
1.27
3.18
2.25
1.36
2.93
2.16
1.42
3.91
2.04
1.02
3.84
2.03
1.09
3.79
2.03
1.19
3.63
1.98
1.28
3.44
1.91
1.38
3.3
1.86
1.45
6
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
Table 1 : Manufacturer data for a residential air-to-air split-system heat pump- Source : CARRIER Corp.
Professionnal catalogue for residential units, 1995-1996. (The nominal values are noticed in italic and heavy
fonts)
We assume that for a fixed outdoor temperature, the sensible and total capacities are
following linear laws. These equations can be written as following :
where :
(27 C) a0 a1 hent
Qtot
(7)
Qsens
(27 C) b0 b1 hent
(8)
hent is the air entering specific enthalpy at the evaporator.
a0, a1, b0, b1 are the correlation coefficients of Qtot
and Qsens
respectively.
4
TOTAL COOLING CAPACITY
3.8
OUTDOOR AIR
TEMPERATURE
COOLING CAPACITY (kW)
3.6
TOT (21°C)
3.4
SENS (21°C)
3.2
TOT (25°C)
SENS (25°C)
3
TOT (30°C)
2.8
SENS (30°C)
2.6
TOT (35°C)
SENS (35°C)
2.4
SENSIBLE COOLING CAPACITY
2.2
2
40
45
50
55
60
65
70
SPECIFIC ENTHALPY (kJ/kg)
Fig 2 : Evolution of total cooling capacity and sensible cooling capacity according to the indoor air specific
enthalpy and for different outdoor air temperatures. Indoor air temperature = 27°C
These coefficients depend on the outdoor air temperature. We determined the values of
these coefficients for outdoor temperatures of 21°C, 25°C, 30°C and 35°C. We didn’t take
into account the values of 40°C et 45°C because on the one hand we are outside the operating
range of the split-system, and on the other hand the evolution of the capacities tend to becom
non linear.
These coefficients are also approximated by a linear law, so they can be expressed as :
a0 = c1 + c2 . Text
a1 = c3 + c4 . Text
b0 = c5+ c6 . Text
b1 = c7+ c8 . Text
(9)
(10)
(11)
(12)
7
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
where Text is the outdoor air temperature.
c1, c2, c3, c4, c5, c6, c7, c8 are the calculated correlation coefficients coming from the
linear equations (9), (10), (11), (12).
Thanks to these correlations, we can expressed the total capacity and the sensible
capacity by the following equations for an indoor air temperature of 27°C :
(27C) = c1 + c2 . Text +(c3 + c4 . Text).hent
Qtot
Qsens
(27C) = c5+ c6 . Text + (c7+ c8 . Text).hent
The manufacturer data have allowed us to estimate the cooling capacities, both total
and sensible, as function of the entering specific enthalpy and the outdoor air temperature for
the set temperature of 27°C. Now, we have to determine the cooling capacities at any
operating point. Thus, we introduce the coil bypass factor concept [6], [14].
The coil bypass factor (BF) model characterizes the air exiting the coil as being
composed of two major streams : the air which has not been influenced by the coil and the air
which leaves at the coil surface temperature (apparatus dew point). The coil bypass factor is
the fraction of air which is unaffected by the coil. Thus, we have relations for the exit dry-bulb
temperature and humidity ratio in terms of entering conditions and the coil bypass factor.
BF
with
Tout Tadp
Tent Tadp
hout hadp
hent hadp
wout wadp
(15)
went wadp
Tadp : apparatus dew point temperature or temperature of the coil surface
hadp : specific enthalpy on the coil surface.
Humidity ratio
hent
Qtot
hout
Q lat
Q sens
E
went
E’
hadp
Dry-bulb temperature (°C)
wout
wadp
Tadp T T’
out
out
Tent
T’ent
Fig 3 : Cooling coil performance and determination of cooling capacities
8
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
The coil bypass factor is a function of both physical and operation parameters of the
coil : these parameters are the coil exchange surface A (when A is rising, BF is decreasing), and
the coil air velocity v (when v is creasing, BF is creasing). We assume in our model that these
parameters are constant (the air velocity is supposed to stay in the high position), so the bypass
factor is held constant too.
Figure 3 shows an illustration of the problem. We made one important assumption in
developping this model :
The total cooling capacity rate was constant for two operating points E and E' having the same
entering specific enthalpy. That means that :
( E ) Qtot
( E ') = m
.(hout hent )
Qtot
(16)
where m is the mass flow rate of the air through the coil (kg.s-1)
The apparatus dew point is by consequence the same for the two entering points :
hadp(E)=hadp(E')
The sensible and latent capacities are changing between the two points, but the total
capacity remains constant. Considering the point E', the equations (13) and (14) give the value
for the same specific enthalpy of the total capacity and the sensible capacity for a drybulb
temperature of 27°C.
In that case :
Qsens
m (hout hB ) m c pm (Tout TE) m c pm (1 BF)(Tadp TE)
'
Qsens'
m (hout
hE' ) m c pm (T'out TE' ) m c pm (1 BF)(Tadp TE' )
Qsens'
m c pm (1 BF)(Tadp TE ) m c pm (1 BF)(TE TE' )
then
where
Qsens' Qsens
V
c (1 BF)(27 TE' )
v pm
(17)
is the sensible capacity of the heat-pump for the E' entering point.
Qsens'
is the sensible capacity of the heat-pump for the E entering point (TE =
Qsens
27°C) given by the expression (14).
V is the air flow rate. It is held constant and equal to 0.136 m3.s-1
v is the specific volume of the air (m3.kg-1).
cpm is the specific heat of the mixing of dry air and water vapour. It is
assumed to be constant : cpm=1.02kJ.kg-1.
We can determine the latent loads thanks to the expressions (16) and (17):
( E ) Qtot
( E ')
Qtot
V
Qsens'
Qsens
c pm (1 BF)(27 TE' )
v
' Qtot
Qsens
Qlat
'
(18)
9
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
are integrated at each time step in equations (3) and (5).
Then, the values of Qsens'
and Qlat'
Dynamic state conditions :
The taking into account of the dynamic effects upon start-up is the same than in model
n°1 (single time constant).
System control :
The system control is an on/off controller with a dead zone of 1°C.
Electric power :
We assumed that the electric power is a linear dependant fonction of the outdoor
temperature. Figure 4 shows a good validation of this assumption. We assume that there is no
dynamic effects during the few time steps after start-up.
1.5
POWER (kW)
1.4
1.3
1.2
y = 0.0176x + 0.6366
R2 = 0.9981
1.1
1
0.9
0.8
20
25
30
35
40
45
OUTDOOR AIR TEMPERATURE (°C)
Fig 4 : Electric power of the split-system
5. EXPERIMENTAL SET-UP AND TEST PROCEDURE
A single-speed air-to air heat pump was instrumented and tested in a real size test cell
located at the University of Reunion Island. The experiments were conducted in natural
climatic conditions.
5.1 Test split-system :
The test unit is a R22 air-to-air split-system residential heat pump that is commercially
available. This split-system is operating only in cooling mode, and has a 3.3kW nominal
cooling capacity. Both indoor and outdoor heat exchangers are of tube-and-plate-fin
construction. The compressor is a hermetic rotative compressor and the expansion device is a
capillary tube.
Monitoring included the temperatures of the refrigerant fluid from both sides of each
components (indoor and outdoor units, compressor and expansion device), the R22 volume
flow rate by means of a liter-meter, high and low pressures, power and energy supply (active,
apparant and reactive) by means of an electronic watt-meter. Even the temperatures of the
10
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
return air and air outlet of the indoor unit the mass flow rate of the water coming from the
indoor unit are mesured. Thus we can determine the total cooling capacity thanks to the
thermodynamic cycle of the R22, and both sensible and latent capacities with the
measurements of return and exit air of the indoor unit, and the measured mass flow rate of the
water respectively.
5.2 Test cell
The size of the test-cell is 3.0 x 3.0 x 2.30m ( see fig. 5).
The test cell is made of sandwich panels consisting of two 7 mm layers of ciment-fiber boards
with 6 cm of polyurethane foam between them. A layer of 5 cm of styrofoam is put between
the floor panel and the concrete. On the roof, there is an extra sandwich panel, made of
aluminium sheets and polyurethane, about 5 cm thick. The inner floor is made of 5cm concrete
paving stones .
Fig 5 : Vue of the Test cell and the outdoor unit of the split-system.
Indoor air temperatures are mesured by type T (copper-constantan) thermocouples
shielded against radiation, at three different levels (0.30m, 1.20m, 1.80m). External and
internal temperature are mesured with type T thermocouple. The sensors are glued to the walls
and painted with the same colours as the walls (i.e for the roof).
The indoor relative humidity is also recorded thanks to a temperatue and relative
humidity probe.
5.3 Weather data acquisition :
A weather station was used close to the test cell (see fig.5 ) to measure the outdoor air
temperature, the outdoor relative humidity, wind speed and direction and the global and diffuse
horizontal solar radiations. Thanks to these data, the meteorological file of will be able to be
monitored in vue to compare measurements and modelisation.
5.4 Data acquisition system and computer processing.
All the data are mesured and collected thanks to dataloggers that are controlled and
programmed by a PC computer via RS232 links. The time step of the data acquisition can be
modified function of our needs. For instance, we use a time step of one hour for the validation
of the test cell without HVAC system, but the time step is reduced to one minute when the
expriments are conducted with active systems.
11
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
30
10.00
28
9.00
8.00
26
7.00
24
6.00
22
5.00
20
4.00
Mesurements
18
3.00
Predictions
2.00
16
Residuals (°C)
Test cell air temperature (°C)
5.5 Experimental procedure
The experiments were conducted at different steps. The first step consisted in a
comparison between the simulation and the mesurements on the test cell without airconditioning system in order to optimize the modelisation of the envelop. Figure 6 shows a
passive period of two weeks. The measured and simulated air temperatures are very similar,
with deviations smaller than 1°C.
In a second time, a lot of experiments were conducted in order to get accurate
informations about the behaviour of the split-system unit (determination of the time constant
for the model n°1 and 2, values of the dead zone band of the controller, etc...). Thanks to these
preliminary tests we determined the time constant of the unit, which is equal to 2 minutes, and
the dead zone band which is in a range of + 0.5°C from the set temperature.
1.00
14
0.00
12
10
252
-1.00
Résiduals
-2.00
254
256
258
260
262
264
266
Julian day
Fig 6 : Comparison between mesurements and predictions - Passive period
On the third time, experiments were conducted with the split-system in the operating
mode for different indoor drybulb set temperatures. The time step of the experiments was
taken to 1 minute. The test periods lasted generally for seven days . The first results are
detailed in the following part
6. RESULTS
The temperature of the heat-pump was set to 23°C during the whole period of the
measurements. The air renewal was neglected. There were no internal loads inside the test cell.
Figures 7 to 10 present measured data and simulation results of models 1 and 2 for a
time step of one minute. The presented outputs in figures 7, 8, 9,10 were the indoor air
temperature, the electric power, the sensible coling capacity and the total cooling capacity
respectively The selected day belongs to a test period of seven days. (Oct. 20-27, 1996). We
have pulled out the best day of the test period.
12
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
The model n°0 requires a time step of one hour. In view to compare the model n°0 to
the simulations of models n°1 and 2 and to measurements, the results for a time step of one
minute were averaged to one hour. Thus, Figures 11 to 13 presents the measured and
simulation results for a time step of one hour. The presented outputs are the average indoor air
temperature, the total cooling capacity and the electric power.
13
MEASUREMENTS
25
24
23
22
21
20
500
1000
1500
Time (Hour)
MODEL N°1
25
0
2000
500
2000
MODEL N°1
1.4
1.2
1
0.8
0.6
0.4
0.2
0
24
1000
1500
Time (Hour)
Power (kw)
Indoor air temp(°C)
0
23
22
21
20
0
500
1000
1500
Time (Hour)
MODEL N°2
25
0
2000
500
2000
MODEL N°2
1.4
1.2
1
0.8
0.6
0.4
0.2
0
24
1000
1500
Time (Hour)
Power (kw)
Indoor air temp(°C)
MEASUREMENTS
1.4
1.2
1
0.8
0.6
0.4
0.2
0
Power (kw)
Indoor air temp(°C)
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
23
22
21
20
0
500
1000
1500
Time (Hour)
2000
Fig 7 : Indoor air temperature - time step =1 min
0
500
1000
1500
Time (Hour)
2000
Fig 8 : Electric power - time step = 1 min
14
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
MEASUREMENTS
500
1000
1500
0
-0.2
-0.4
-0.6
-0.8
-1
-1.2
-1.4
-1.6
-1.8
MEASUREMENTS
2000
0
Time (Hour)
500
500
1000
1500
MODEL 1
2000
Time (Hour)
0
500
1000
1500
1500
2000
Time (Hour)
MODEL 2
2000
Time (Hour)
Fig 9 : Sensible capacity - time step = 1 min
0
Total capacity
(kW)
Sensible
capacity (kW)
0
-0.2
-0.4
-0.6
-0.8
-1
-1.2
-1.4
-1.6
-1.8
-2
-2.2
500
1000
0
-0.5
-1
-1.5
-2
-2.5
-3
-3.5
-4
MODEL N°2
0
2000
Time (Hour)
Total capacity
(kW)
Sensible
capacity (kW)
0
1500
0
-0.5
-1
-1.5
-2
-2.5
-3
-3.5
-4
MODEL N°1
0
-0.2
-0.4
-0.6
-0.8
-1
-1.2
-1.4
-1.6
-1.8
1000
Total capacity
(kW)
Sensible
capacity (kW)
0
0
-0.5
-1
-1.5
-2
-2.5
-3
-3.5
-4
500
1000
1500
2000
Time (Hour)
Fig 10 : Total cooling capacity - time step = 1 min
15
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
25
24.5
Indoor air temp (°C)
24
23.5
23
22.5
22
21.5
Mesurements
Model n°1
21
Model n°2
20.5
Model n°0
20
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time (Hour)
Fig 11 : Indoor air temperature - Comparison of measurements and simulations - time step = 1 hour
Cooling capacity (kw)
23
21
19
17
-0.05
15
11
9
7
5
3
1
0
13
Time (Hour)
-0.1
-0.15
-0.2
-0.25
Model n°0
Model n°1
-0.3
Model n°2
-0.35
Fig 12 : Total cooling capacity - Comparison of measurements and simulations - time step = 1 hour
mesurements
Model n°0
Model n°1
Model n°2
0.25
0.2
0.15
0.1
23
21
19
17
15
9
7
5
3
1
0
13
0.05
11
Electric Power (kW)
0.3
Time (Hour)
Fig 13 : Electric power - Comparison of measurements and simulations - time step = 1 hour
16
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
Visual inspection and comparison of the various graphs yields the following observations.
Short time step (one minute):
The comparison of the measured and simulated indoor air temperatures are very
similar. The time of the first start-up of the split system is earlier for the models
(approximatively around 7h30 a.m) than for the experiment (around 8h00). This is due to the
0.5°C gap between models and measurements during the passive period. Yet, the number of
on/off cycling, the daily variations of the indoor air temperature and the last shut down (aroud
17h30) are well modellised.
One can see the same trends for the predicted power demand. The model n°2 seems to
be more accurate than the model n°1 where the electric power is constant (it is supposed te be
equal to the nominal power). Then power demand of Model n°1 is systematically bigger than
the measured one.
The predictions of the sensible cooling capacity (see figure 9) show the correct general
tendancies when compared to the measured values. On the contrary, there is a large gap
between measurements values and the two models for the total cooling capacity (see figure
10). This gap is due to the fact that the time delay of one on-off cycle is shorter than the
measured one. The modelled capacity has not time enough to reach the steady-state value.
One hour time step :
Indoor air temperature predictions exibit the same trends as the air temperature. This is
quite obvious because these values are just the average of the one-minute-time-step values.
The maximum gap is around 0.5°C.
The results given by figures 12 and 13 are quite interesting. The predictions of total
cooling capacities for models 1 and 2 are larger than the one obtained for the model 0. These
discrepancies can be attributed to the simulation of the controller which is taken into account
in models n°1 and 2 (and supposed to be ideal in model n°0). The same trends are pointed out
in figure 13 for the predictions of electric power demand. The model 2 seems to be the more
accurate with an accuraccy of less than 10%. On the contrary, model n°0 underestimates the
power demand in a very important way. We can also see the limits of hourly simulations to
predict systems consumptions such as heat-pump systems with specific control strategies.
7. CONCLUSIONS
Three modelling levels of air-to-air residential heat-pumps have been defined and been
integrated in a thermal building simulation software. The first model is just supposed to supply
the demand of hourly cooling loads and electric demand with an ideal control loop (no delay
between the sollicitations and the time response of the system) . The second model is taking
into account the on/off cycle and the control processing. The third one is based on linear model
determined from manufacturer data. Dynamic effects are taken into account in the last two
models thanks to a single time constant model. The time step is reduced to one minute. The
ouputs are the total, sensible and latent cooling capacities at each time step.
The first comparisons between measurements and simulations point out that a dynamic
simulation with shorter time steps than an hour give better results on the estimation of electric
consumption. This is mostly due to the kind of control process ( on/off controller ) of these airconditionner systems. Netherveless, improvments have to be made concerning the estimation
of the total cooling capacity. We have to face physical problem such as non homogeneity of the
air inside the test cell and air infiltrations, what implies a time-gap between measures and
17
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
models. Future research will focus on the introduction of a time delay due to theses physical
phenomenon.
8. REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
Allen J.J, Hamilton J.F. 1983. Steady-state reciprocating water chillers models.
ASHRAE Transactions, Vol. 84, part 2, p 398-407.
Armand J.L, Mondot M., Molle N., Habershill P., Lallemand M., Component based
modelling of refrigerating compression cycle. In proceedings of System Simulation in
Buildings, 1990.
Boyer H., Brau J., Gatina J.C., Multiple model software for airflow and thermal
building simulation. A case study under tropical humid climate, in Réunion Island. In
Proceedings of Building Simulation '93, (IBPSA, Adelaide, Aug.), 111-117.
Boyer H., Chabriat J.P., Grondin-Perez B., Tourrand C., Brau J., Thermal Building
Simulation and Computer Generation of Nodal Models. Building and Environment, Vol.
31, n°3 (1996) 207-214.
Boyer H., Gatina J.C., Pignolet-Tardan F., Brau J., Modelisation methods and data
structuration induced for building thermal simulation codes. In proceedings of the
Conference of international Simulation Societies, (Zurich, Switzerland, August 22-25,
1994), p 729-733.
Brau J. , Théorie du conditionnement d’air, Institut National des Sciences Appliquées
de Lyon, Laboratoire Equipement de l’Habitat, 1981, 81p.
Bullock C.E., Wroblewski D.E, Groff, G.C, A dynamic simulation model for
residential air-to-water heat pump system, in proceedings tome V of the XVIth
International congress of Refrigeration, Paris, 1983.
Chi J., Didion D. , A simulation model of the transient performance of a heat pump.
International Journal of Refrigeration, Vol 5, n°3 (1982) p176-184.
Clarke J.A. Energy simulation in building design. Glasgow (U.K). Adam Hilger Ltd,
1985, 383p. ISBN 0-8574-797-7.
Crawfford R.R., 1987. Dynamic modelling of a residential heat pump from actual
system performance data. ASHRAE Transactions, Vol.93, Part 2, p 1179-1190.
Garde F., Boyer H., Brau J., Gatina J.C., Validation expérimentale d’un code de
modélisation thermique de bâtiments (CODYRUN). Une application en climat tropical
humide. In proceedings of 2ème colloque interuniversitaire franco-québecois, Thermique
des systèmes à température modérée. Sherbrooke, Montréal, Canada, p. 197-202, 1995.
Gluck R., Pollack E., 1978. Design optimisation of air-conditioning systems. ASHRAE
Transactions, Vol84, Part 2, p304-314.
Goldsmith V.W., Hart G.H., Reiner R.C., 1980. A note on the degradation
coefficient of a field tested heat pump coolind and heating mode. ASHRAE
Transactions, Vol. 86, Part 2.
Hirsch J.J, Simulation of HVAC Equipment in the DOE-2 Program. In proceedings of
System simulation in Buildings, Liège, 1982, p 89-107.
Irving S.J., APACHE - an integrated apporoach to thermal and HVAC systems
analysis. International Journal of Ambiant Energy, Vol. 7, n°3 (1986) p. 129-136.
Loveday D.L., Ewers R.A., The cost effectiveness of heat pumps operated by a BEMS
- A comparison of smart and standard control using dynamic simulation, in proceeding
of System Simulation in Buildings, 1990.
18
___________________________________________________________________________
CLIMA 2000, Bruxelles, Août 1997
___________________________________________________________________________
[17] MacArthur J.M., Grald E.W, 1987. Prediction of cyclic heat pump performance with
a fully distributed model and a comparison with experimental data. ASHRAE
Transactions, Vol. 93, Part 2, p. 1159-1178.
[18] Mulroy W.J. Didion D.A. 1985. Refrigerant migration in a split-unit air conditioner.
ASHRAE Transactions, Vol. 91, Part 1A, p. 193-206.
[19] Mulroy W.J. 1986. The effect of short cyling and fan delay on the efficiency of a
modified residential heat pump. ASHRAE Transactions, SF-86-17 No1, p. 813-826.
[20] Murphy W.E., Goldsmith V.W, 1979. The degradation coefficient coefficient of a
field tested self-contained3-ton air conditionner. ASHRAE Transactions, Vol. 85, Part
1, p. 839-849
[21] O’Neal D.L, Katipamula S., 1987. Performance degradation during on-off cycling of
single-speed air conditioners and heat pumps : Model development and analysis,
ASHRAE Transactions, Vol. 97, Part 2, p. 316-323.
[22] O’Neal D.L, Katipamula S. Development of nondimensionnal cycling model for
estimating the seasonal performance of Air conditionners. Journal of Solar Energy
Engineering Vol. 115 (1993), p. 176-181.
[23] Sami S.M., Duong T.N., Mercadier Y., Galanis N., 1987. Prediction of the transient
response of heat pumps, ASHRAE Transactions, Vol. 93, Part 2, p. 471-490.
[24] Stan S., Maîrise et calcul des consommations des installations de climatisation. Phd
Thesis, Ecole Nationale Supérieure des Mines de Paris, 1993.
[25] Welsby P., Devotta S., Diggory P.J. Steady- and Dynamic-State Simulations of HeatPumps. Part I : Literature Review, Applied Energy 31 (1988) 189-203.
[26] Xin Ding, Eppe Jean-Pol, Lebrun Jean, Wasacz Marian, Previous models and new
proposals of cooling coil in transient and/or wet regimes. Theorical and experimental
validation . In proceeding of System Simulation in buildings, 1990.
19
| 5cs.CE
|
Enslaving the Algorithm: From a “right to an
explanation” to a “right to better decisions”?
Lilian Edwards, University of Strathclyde
Michael Veale, University College London
arXiv:1803.07540v1 [cs.AI] 20 Mar 2018
IEEE Security & Privacy, forthcoming (2018)
As concerns about unfairness and discrimination in “black box” machine learning systems rise, a legal “right to an explanation” has emerged as a compellingly
attractive approach for challenge and redress. We outline recent debates on the
limited provisions in European data protection law, and introduce and analyze
newer explanation rights in French administrative law and the draft modernized
Council of Europe Convention 108. While individual rights can be useful, in privacy law they have historically unreasonably burdened the average data subject.
“Meaningful information” about algorithmic logics is more technically possible
than commonly thought, but this exacerbates a new “transparency fallacy”—an
illusion of remedy rather than anything substantively helpful. While rights-based
approaches deserve a firm place in the toolbox, other forms of governance, such
as impact assessments, “soft law,” judicial review, and model repositories deserve more attention, alongside catalyzing agencies acting for users to control
algorithmic system design.
1 Introduction
Businesses and governments are increasingly deploying machine learning (ML) systems to
make and support decisions that have a crucial impact on everyday life: decisions about
(inter alia) criminal sentencing and release on bail, medical treatment, eligibility for welfare
benefits, what entertainment we see and can access, the price and availability of goods and
services delivered online, and the political information to which we are exposed. These ML
systems—colloquially entering public consciousness as just algorithms, or even just AI—have
been extensively criticized in the past few years as a result of a number of well-known “war
stories” that have revealed patterns of discrimination embedded but invisible to casual users
in such systems.1
Because algorithms are trained on historical data, they risk replicating unwanted historical patterns of unfairness and/or discrimination. For example, in hiring systems, a lack of
women being hired in the past may mean the systems fail to recognize the worth of female
applicants, or even outright discriminate against them. Luxury goods may be advertised to
1
people with certain profiles on social media and not to others, creating a consumer “under
class.”
A severe obstacle to challenging such systems is that outputs, which translate with or
without human intervention to decisions, are made not by humans or even human-legible
rules, but by less scrutable mathematical techniques. A loan applicant denied credit by a
credit-scoring ML algorithm cannot easily understand if her data was wrongly entered, or
what she can do to have a greater chance of acceptance in the future, let alone prove the
system is illegally discriminating against her (perhaps based on race, sex, or age). This
opacity has been described as creating a “black box” society.2
2 Enter the Right to an Explanation
Since the 1990s, the law in Europe has been concerned with this kind of opaque and difficultto-challenge decision making by automated systems. In consequence, the Data Protection
Directive (DPD), a measure that harmonized relevant law across EU member states in 1995,
provided that a “significant” decision could not be on based solely on automated data processing (article 15). Some EU members interpreted this as a strict prohibition, others as
giving citizens a right to challenge such a decision and ask for a “human in the loop.” A
second right, embedded within article 12, which generally gives users rights to obtain information about whether and how their particular personal data was processed, gave users
the specific right to obtain “knowledge of the logic involved in any automatic processing” of
their data. Both these provisions, but especially the latter, were not much noticed, even by
lawyers, and scarcely ever litigated, but have revived in significance in the latest iteration of
EU data protection (DP) law within the General Data Protection Regulation (GDPR), which
passed in 2016 and will come into operation across Europe in 2018.
In the GDPR, article 15 has been transformed into Article 22 and has arguably created
what the media and some technical press have portrayed as a new “right to an explanation”
of algorithms. The former article 12 has also been revamped to a new article 15 and now
includes a right to access to “meaningful information about the logic involved, as well as the
significance and the envisaged consequences of such processing” (article 15(1)(h)). This
provision, notably, applies only in the context of “automated decision making in the context
of” Article 22. This leaves it unclear if all the constraints on Article 22 (discussed below) are
ported into article 15 (though our view is that it does not). Sadly, all this adds up to a reality
considerably foggier than the media portrayal.
Several factors undermine the idea that Article 22 contains a right to an explanation.
Primarily, Article 22 does not in its main thrust even contain a right to an explanation, but
is merely a right to stop processing unless a human is introduced to review the decision on
challenge. However, Article 22 does refer at points to a requirement of “safeguards,” both
where the right to prevent processing (paradoxically) does not operate, and where it does
but sensitive personal data is processed. In relation to the first case, safeguards are partly
listed in Article 22(3), but in the second case, the only guidance is in Recital 71. (“Sensitive”
personal data in DP law refers to a restricted list of factors regarded as particularly important
such as health, race, sex, sexuality, and religious beliefs.)
2
It is important to note that, in European legislation, the articles in the main text are binding on member states but are accompanied by “recitals,” which are designed to help states
interpret the articles and understand their purpose. Recitals are usually regarded as helpful
rather than binding, but this is contested and differs among states. Unfortunately, in relation
to Article 22, Recital 71 mentions some key matters not included in the main text. Article
22(3) mandates that safeguards include “at least the right to obtain human intervention on
the part of the controller, to express his or her point of view and to contest the decision,” but
the safeguards listed in Recital 71 “should include specific information to the data subject
and the right to obtain human intervention, to express his or her point of view, to obtain an
explanation of the decision reached after such assessment and to challenge the decision” (italics
added).
This strange mishmash of texts thus cannot firmly be said to mandate a right to explanation
in all or indeed any circumstances and may not be interpreted the same way from state to
state.
This is a serious, but not the only, problem with Article 22.
• Article 22 applies only to systems where decisions are made in a “solely” automated
way—that is, there is no human in the loop—and there are very few of these and fewer
that are “significant” (see below). How meaningful this input has to be is subject to
recent regulatory guidance,3 but remains unclear and untested.
• What is a “decision”? The GDPR gives us no help with this at all other than that it includes a “measure” (Recital 71). Is sending a targeted ad to a user using an algorithmic
system a decision? It produces no binding effect; the advert may be ignored; and in
many cases, it is hard to see what action causally flows from it. Yet as in the wellpublicized Latanya Sweeney example,4 sending adverts promoting help with criminal
arrests solely to “black-sounding” names was worrying and offensive—and potentially
dangerous if these characterizations were inherited by systems selecting individuals
for stop and search or airport screening. Although a single advert delivery decision
might not have a significant effect on an individual’s life, the cumulative effect on an
entire group or class may be worrying. Such group privacy impacts are not dealt with
well by DP law—an area based on individualistic human rights—and are exacerbated
by a continuing lack of provision for class actions in EU states.
• Article 22 applies only to a decision that produces legal or other “significant” effects.
This is vague in the extreme. Some would argue this could only apply to systems
that make important, binding decisions on things like criminal justice, risk assessment,
credit scoring, education applications, or employment. Yet such systems are rarely
if ever entirely automated, even if the human’s involvement is often nominal. Furthermore, some commercial decisions may seem trivial as a one-off, but are significant in aggregate. Mendoza and Bygrave argue that advertising decisions can never
be significant,5 while European regulators recently produced guidance indicating the
opposite.3 Might systems recommending buying choices or targeting adverts not limit a
user’s worldview or choices, or disseminate “fake news” via algorithmic filter bubbles?
Arguably, such phenomena are becoming deeply and significantly destructive to our
3
democracy. We have an obvious link here to the issue of to whom a decision needs to
be significant: the individual in question or society as a whole?
Turning to the new article 15 of the GDPR (right to information), this right to “meaningful
information about the logic involved” in any decision-making system may be more useful
than Article 22. It is (arguably) not directly as restrictive as Article 22 is to solely automated
decisions having significant or legal effects. But there is an unresolved doubt about whether
it only applies to information available before the system makes a decision about a particular
data subject (see Wachter and colleagues’ work6 ). In ML parlance, that means it is uncertain
if the right is only to a general explanation of the model of the system as a whole (modelbased explanation), rather than an explanation of how a decision was made based on that
particular data subject’s particular facts (subject-based explanation).1
But even if we agree that Article 15 may give us some kind of functioning right to an
explanation, we still have huge problems. The GDPR can apply only where decisions are
made based on personal data. Personal data is defined in article 4(1), as “any information
relating to an identified or identifiable natural person” and is certainly wider than data that
has the name of a data subject attached to it. According to Recital 26:
To determine whether a natural person is identifiable, account should be taken of all the
means reasonably likely to be used, such as singling out, either by the controller or by another
person to identify the natural person directly or indirectly.
Even allowing for this broad (and still controversial) approach to defining personal data,
some data will remain outside the scope of the GDPR. Yet algorithmic decisions that affect people may not involve personal data. Take self-driving cars. They may kill people—
passengers or pedestrians—as a result of algorithmic processing, yet the data involved in
that decision may be entirely related to traffic, road conditions, and other non-personal matters. Other circumstances may involve data that was once personal but has been allegedly
anonymized. This is very common, for example, with profiles made from personal data
collected by social networks and used to generate targeted marketing.
Lastly, a final restriction to these rights comes in the form of a carve-out in recital 63
(though not main text) for intellectual property (IP) and trade secrets. Explanations of how
an algorithm works might reveal a firm’s competitive advantage—its notorious “secret sauce.”
Although this should not result in a “refusal to provide all information to the data subject,”
the lack of clear-cut provisions will likely continue to further muddy the waters. In these
cases, proposed explanations of systems based on modeling slices of a proprietary model
without access to its innards may prove a useful middle ground.7,8
2.1 Improving the Right to an Explanation after the GDPR
Above, we have seen how the GDPR’s right to an explanation appears far from ideal. This
is hardly surprising given that the provisions remain largely modeled on the 1995 directive,
which effectively predated the Internet and modern algorithm design and accompanying
harms.
More modern laws exist: one example is the French loi pour une RÃl’publique numÃl’rique
(Digital Republic Act, law no. 2016-1321), This gives a right to an explanation for adminis-
4
trative algorithmic decisions made about individuals. The new law provides that in the case
of “a decision taken on the basis of an algorithmic treatment” (author translation), the rules
that define that treatment and its “principal characteristics” must be communicated upon
request. Further details were added by decree in March 2017 (R311-3-1-2) elaborating that
the administration shall provide information about:
1. the degree and the mode of contribution of the algorithmic processing to the decision
making;
2. the data processed and its source;
3. the treatment parameters and, where appropriate, their weighting, applied to the situation of the person concerned; and
4. the operations carried out by the treatment.
Some areas of government, such as national security and defense, are excluded.
The French approach has important advantages over the GDPR. First, considering point 1,
true decision-support systems are not automatically excluded from being explained.
Secondly, point 3 provides that, where appropriate, the weightings of factors in a system can be disclosed. This seems to imply the explanation must be of a particular decision
(subject-based explanation) rather than a vague overview of a complex model (model-based
explanation), contradicting the interpretation by Wachter and colleagues.6 Extracting estimates of the weightings within a complex algorithm is increasingly possible, particularly if
only the area “local” to the query is being considered,7 which unlike the complex innards of
the entire network, might display recognizable patterns.9
Weights may help explain systems, but they are by no means a complete fast track to
interpretability. There are at least two occasions when a court might say that weights are
not useful for explaining a decision to a human user, and therefore, it is not appropriate
to order disclosure. These are when the weighted inputs do not map to any real-world
features the user will find intelligible and, in older or restricted systems, where retrofitting
an explanation system is infeasible.
On the downside, the new French right applies only to administrative decisions. This
makes its advances more comprehensible, for several reasons. First, the number of discretionary decisions currently made by governmental algorithmic systems is relatively small
compared to the increasing amount of ML profiling in the private/commercial sector. Second,
there is a long-established constitutional expectation that democratic governments will, to
some extent, be transparent—via, for example, freedom of information requests. By contrast,
the private sector is usually not required to disclose its secrets except on limited occasions
such as financial disclosures.
Another new instrument in the planning is the modernization of the Council of Europe
Convention 108 for the Protection of Individuals with Regard to Automatic Processing of
Personal Data (CoE 108). CoE 108 is an international treaty relating to DP, which regardless
of its name, can be signed by any state in the world. Its membership, however, remains
relatively limited and, for instance, does not currently include either the US or China.
5
In a recently circulated draft, it appeared that the CoE was considering as one alternative,
a version of the right to an explanation for all automated decisions, without the restriction
of GDPR Article 22. This would be an exciting development. While the outcomes of the CoE
negotiations are far from settled, interestingly it seems this potentially expanded CoE right
has already been ported to the draft UK transposition of the GDPR, the Data Protection Bill
2017 (see https://perma.cc/X7X6-TXW9). The bill is oddly drafted but, in brief, the first
two parts of the bill reflect EU law since they transpose both the GDPR and the new EU
directive relating to DP in policing matters (Directive (EU) 2016/680). The third part of the
bill, however, applies DP law to UK intelligence services. These are outside of EU law, but
the UK has chosen to make these provisions compliant not with the GDPR but with CoE 108.
As a result, the draft bill’s s96 currently contains one of the most advanced provisions on
a right to explanation in the world! Before too much excitement is generated, however, it
should be noted that any disclosures will still be controlled by overarching exemptions for
national security in draft s108 and s109, and so are in fact never likely to be exercised.
3 Is a Right to an Explanation Our Best Remedy?
The right to an explanation is only one tool for scrutinizing, challenging, and restraining
algorithmic decision making. While it has rhetorical strength in demanding transparency to
enable user challenge, it has serious practical and conceptual flaws.
First, reliance on an explanation to bolster individual rights places a primary and heavy
onus on users to challenge bad decisions. Even ordinary DP subject access requests (SARs)
demand an enormous amount of time and persistence and, in reality, are mainly used effectively only by journalists and insiders who know how the company in question organizes its
data processing systems. Very few ordinary users historically make use of SARs, and still
fewer will probably use the right to an explanation. These issues are abetted by the common
problems of consumer access to justice, including a general lack of access to legal aid and,
in the EU, to class actions or collective redress. As noted below, users are supposed to be
represented by their state data protection authorities (DPAs) in Europe—but in reality, this
support may be lacking due to a dearth of personpower and expertise in these bodies.
Second, even if obtained, an explanation may not be helpful in mounting a challenge.
This may not be just because of the well-known difficulties about expressing machine logic
in human-comprehensible form,1 despite the progress that is being made in overcoming the
hurdles to producing “meaningful information” about algorithmic logic. But algorithmic
models, inputs, and weightings, however disclosed, may still not show that a system has
been designed to be biased, unfair, or deceptive. Most algorithms will display inadvertent
bias rather than explicitly coded-in bias. (Designers will not want to be sued or prosecuted
for illegal action even if their own ethics do not forbid it.) Researchers are discovering now
how difficult many of these problematic but non-obvious issues can be to spot even when
they have the whole dataset to hand. Biased or discriminatory behavior may only become
apparent looking at the corpus of users as a whole—something that will not happen through
individual user challenges.
It might be possible to better understand these aspects of a model as a whole if many in-
6
dividuals could utilize their individual rights to explanation at once. This would require far
better legal and technical mechanisms for collective action and challenge than we have now.
At the moment, even gathering information about the collective impact of algorithmic systems on users is difficult and unusual. A good example is the sending of “dark ads” to voters
during recent political campaigns via political profiling of voters on social media platforms.
Because these adverts were personally targeted to users, outside agencies could not even
know what ads were being deployed, less still count them or their influence. Dark ads during the 2016 British general election were tracked to some extent by volunteers who installed
and browsed the Internet with the WhoTargetsMe tool, but naturally this counting was very
partial and non-representative. Given the already weighty burden on the individual, also
requiring them to coordinate collective action to expose unfair algorithms seems unlikely to
succeed.
In short, a legal right to an explanation may be a good place to start, but it is by no means
the end of the story. Rights become dangerous things if they are unreasonably hard to exercise or ineffective in results, because they give the illusion that something has been done
while in fact things are no better. It is instructive here to compare the history of consent
to sharing of data, which has moved in the online world from a real bulwark of privacy to
something most often described as meaningless or illusory. Consent is usually given via privacy policies that are largely never read—and, if read, not understood—and they cannot be
negotiated and change from time to time. Thus, consent has become a formality validating
the actions of the data controller rather than something empowering the user. This is sometimes known as the notice and choice fallacy. It would be worrying and dangerous to see the
right to an explanation become a similarly empty formality. This “transparency fallacy” is
something we should guard against and that should spur us on, as in the next section, to
look at alternative and supplementary ways to build better systems.
In particular, any remedy given by a right to an explanation will often come too late for
that specific user. The current investigations into automated recidivism decisions in the US
biased against black data subjects show very clearly a history of discrimination for years,
which is only now becoming apparent. In many cases, we would rather the system was
never broken in first place, or at very least that the individual decision concerned swung in
a more just direction. This takes us to a range of ex ante as well as ex post governance tools.
In the next sections, we begin to consider what other regulatory tools have been created
in the GDPR and elsewhere that might be pressed into use to try to ensure, audit, or instigate
the creation of algorithms that are fairer, less discriminatory, and ideally, less opaque.
4 Investigating before Deployment or Decision-Making
When algorithmic systems make choices in real time, such as who can access a service or
who is selected for security screening, redress or transparency long after the choice is little
help. For this, we must consider the governance tools that have impacts upstream, while
systems are being designed or at least before they are deployed.
7
4.1 Privacy by Design, Data Protection by Design, and Impact Assessments
The GDPR introduces a number of new provisions that, rather radically, do not confer individual rights, but rather attempt to create an environment in which less “toxic” automated
systems will be built in future. These ideas come out of the long evolution of privacy by
design (PbD) engineering as a way to build privacy-aware or privacy-friendly systems, generally in a voluntary rather than mandated way. They recognize that a regulator cannot
do everything by top-down control, but that controllers must themselves be involved in the
design of less privacy-invasive systems. These provisions include the following requirements:
• Controllers must, at the time systems are developed as well as at the time of actual
processing, implement “appropriate technical and organisational measures” to protect the rights of data subjects (GDPR, article 25). In particular, “data protection
by design/default” is required so that only personal data necessary for processing is
gathered. Suggestions for PbD include making use of pseudonymization and data minimization.
• When a type of processing using new technologies is “likely to result in a high risk” to
the rights of data subjects, then there must be a prior data protection impact assessment (DPIA) (article 35), and under some conditions, consultation with the regulator.
• Every public authority, every “large scale” private-sector controller, and any controller
who processes the special categories of data under article 9 (sensitive personal data)
must appoint a data protection officer (DPO) (article 37).
DPIAs especially have tremendous implications for ML design. Impact assessments are
tools used in many domains to assess or estimate impacts of particular interventions or
courses of action. GDPR article 35 notes that:
Where a type of processing in particular using new technologies . . . is likely to result in
a high risk to the rights and freedoms of natural persons, the controller shall, prior to the
processing, carry out an assessment of the impact of the envisaged processing operations on
the protection of personal data.
Where a DPIA “indicates that the processing would result in a high risk in the absence of
measures taken by the controller to mitigate the risk,” the controller is obliged to “consult
[the data protection authority] prior to processing” (article 36(1)) with a view to putting in
measurers to mitigate the risk. Realistically, this seems only likely in cases of highly novel
technologies or the use of existing technology in a new context; DPIAs are not intended as a
tool to stop processing, but rather as a way to refine, or provide points of accountability for
the future operation of, complex systems.
The Article 29 Working Party (A29 WP; a body made up of national DP supervisory authorities that gives authoritative but not binding recommendations on how to interpret DP
laws—soon to be revamped as the European Data Protection Board) has issued draft guidance (17/EN WP 248) further elaborating the conditions under which a data controller must
carry out a DPIA. These include if two or more of the following conditions are met as part
of processing:
8
• evaluation or scoring of individuals;
• automated decision making with legal or similar significant effect;
• systematic monitoring, including of a public area;
• sensitive data processing, as defined in the GDPR;
• data processed on a large scale;
• matched or combined datasets, particularly if data subjects might have had different
expectations about their use;
• data concerning vulnerable data subjects;
• innovative use of technological or organizational solutions;
• data transfer outside the EU; and
• processing that prevents rights being exercised, such as in a public area people cannot
avoid, or as a necessary prerequisite to service provision.
Taken together with the GDPR, this guidance indicates that a DPIA will be an obligatory
precursor for many ML systems with sizable anticipated risks or consequences for individuals
or groups.
DPIAs are not a replacement for explanations of algorithmic systems, not least because
they are aimed at helping builders and regulators, not, directly, users (although article 35(9)
does provide that “where appropriate, the controller shall seek the views of data subjects”).
DPIAs are also not required to be public documents, although it is considered good practice
by the A29 WP to do so at least in part. However, they may be of considerable value in
leading to the building of better systems overall.
Articles 35 and 36 do not specifically require a DPIA to combat potential discrimination:
early drafts made a DPIA mandatory where there was a “risk of discrimination being embedded in or reinforced by the operation.”10 This amendment in the final text was relegated to
recitals (see recitals 71 and 75), which as discussed, have a murky status in EU law. However, in its draft guidance, the A29 WP explicitly clarified that “rights and freedoms” in article
35(1) “may also involve other fundamental rights such as . . . prohibition of discrimination.”
The UK’s Information Commissioner’s Office (ICO), in its influential guidance report on big
data, artificial intelligence, machine learning, and data protection,11 notes firmly that “potential privacy risks” have already been identified with “the use of inferred data and predictive
analytics” and goes on to provide a draft DPIA for big data analytics (Annex 1). It seems clear
that, despite the uncertainty of the “high risk” threshold, DPIAs are quite likely to become
the required norm for algorithmic systems, especially where sensitive personal data, such as
race or political opinion, is processed on a “large scale” (GDPR, article 35(3)(b)).
Impact assessments that deal with risks of discrimination do already exist. The UK has
extensive experience with equality impact assessments (EqIAs), which used to be required for
every new governmental policy and were considered one of the main ways of documenting
9
fulfilment of the Public Sector Equality Duty, which was brought into law by the Equality
Act 2010. The Public Sector Equality Duty requires due regard to be given to impacts on
protected classes before a policy is finalized or implemented, and it is primarily accessed
through judicial review. This requirement does not always need to be met via an EqIA (see R
(Brown) v Sec of State for Work and Pensions 2008 EWHC 3158 (Admin)) but must be carried
out with rigor before a policy is implemented, with documentation to show the process if it
is not otherwise clear. Arguably, where new public sector decision support systems are built,
an EqIA would be highly appropriate and could be combined with the DPIA process. Recent
proposals for “algorithmic impact assessments”, while an immature field awaiting concrete
proposals, would appear to straddle EqIAs and DPIAs in form.
4.2 Certification Systems
The GDPR also introduces the idea of voluntary certification for ML systems. Article 42 proposes voluntary certification of controllers and processors to demonstrate compliance with
the regulation, with “certification mechanisms” and the development of “seals and marks”
to be encouraged by EU member states. In the UK, a tender has already been advertised by
the ICO for a certification authority to run a UK privacy seal, although progress has been
interrupted by the vote to exit the European Union and the subsequent political turmoil.
Taken together, these provisions offer exciting opportunities to operationalize what in the
US have been called “big data due process” rights.12,13 Certification could be applied to two
main aspects of algorithmic systems:
• certification of the algorithm as a software object by (a) directly specifying either
its design specifications or the process of its design, such as the expertise involved
(technology-based standards) and/or (b) specifying output-related requirements that
can be monitored and evaluated (performance-based standards); and
• certification of the whole person or process using the system (system controller) to
make decisions, which would consider algorithms as situated in the context of their
use.
One notable advantage is that certification standards could be set on a per-sector basis.
This is already very common in other sociotechnical areas, such as environmental sustainability standards.
The downside of what seems an exciting approach is that the history in the privacy domain
of self-regulation of the private sector by seal and certificates is dispiriting. Essentially this
involves privatization of regulation and scrutiny. Certification scheme and trust seals have
to make money to survive, which can only be obtained by asking fees from members. Given
this self-interest, it is hard to punish members too hard when they breach the rules of the
seal or certificate, for fear they will leave, either altogether or for a less demanding trust seal
(in a plural market, which is generally what is envisaged). This in turn tends to diminish the
value of the seal or certificate as a guarantee of trustworthiness. There is also little proof
users regard seals and certificates as indicators of trust, which may mean organizations are
10
unwilling to pay or make an effort to belong to them, unless by doing so they can avoid more
stringent top-down regulation.
A clear example here is the US–EU Safe Harbor agreement for the export of personal data
to the US, which came ignominiously to an end in Schrems v DPC of Ireland in the Court of
Justice of the EU in 2015 (Case C-362/14). The US as a country had been deemed not to
have adequate protection in its law for personal data, and so in principal, EU data could not
be exported to the US as far back as the DPD in 1995. A solution was found, however, in the
Safe Harbor agreement whereby US companies could receive EU data if they joined a trust
seal, membership of which guaranteed they were meeting adequate privacy standards. One
of the largest and most prominent seals used in this way was TrustE. Yet as Charlesworth
showed back in 2000,14 TrustE had a long history of overlooking major data breaches by
members or imposing only desultory sanctions. This, as well as the US’s cavalier attitude to
covert state surveillance exposed in the Snowden revelations, led to the agreement’s judicial
annulling.
5 Enabling Review and Challenge Without Individual Burden
Injustices in decision-making can also emerge even from algorithmic systems thought to
be well designed, and in those cases, reactive provisions are appropriate and needed. But
how can these be achieved whilst avoiding burdening individuals with the responsibility of
understanding, investigating, and identifying points of action in arcane and often shadowy
technological deployments?
5.1 Representation Bodies for Data Subjects
The aim of a right to an explanation is fundamentally to enable challenge to poor or wongly
made decisions by black box systems. Yet, as we have seen, individual users typically The to
assert their rights, especially in complex and opaque areas such as ML systems. Moreover,
if we are talking about a harm such as embedded systemic discrimination, which typically
affects a whole class of people, then it may seem far more appropriate for a representative
body to accept and mount challenges than each individual user. This is a common problem
in consumer law, and was long anticipated in DP law by the creation of state DPAs, which
already have a role to investigate user complaints and enforce breaches against data controllers. One historic problem here has been the low level of sanctions a DPA could dish out:
this has famously been met in the GDPR by the creation of a maximum fine of up to EUR 20
million or 4 percent of global annual turnover. This welcome change does not alter the fact,
however, that the volume of DP breaches both deliberate and inadvertent is now so huge
that one agency alone cannot combat it. Furthermore, state DPAs throughout the EU are
wildly underfunded, since in their nature (and by law), they have to be seen to be independent of both state and the commercial sector. A final major problem germane to ML-related
complaints is that DPAs are typically staffed by civil servants and/or lawyers and rarely have
much technical understanding or capacity.
The GDPR tries to help here in two main ways. Article 80(1) provides for all member
11
states that a data subject can mandate a third body to lodge a complaint, exercise the right
to judicial remedy, and receive compensation on his or her behalf. This is a useful step,
particularly if civil society bodies can find an exemplary case to support, but it still requires a
data subject to notice a breach and have the time and effort to reach out to a third body and
enlist its help. Given a third body can receive compensation on behalf of a data subject, this
might also result in a dubious “ambulance-chasing” industry (similar to the UK furor over
wrongly paid payment protection insurance, or PPI).
Article 80(2), GDPR, by contrast, permits member states to allow third-party bodies to
take up complaints, for example, against a data controller, without being mandated by a data
subject. Through this provision, civil society bodies could monitor sectors and controllers for
breaches and pursue suspected infringements of their own accord through judicial remedies.
Article 80(1) is mandatory for GDPR-implementing states to put into law, whereas article 80(2) is not. While some countries such as Germany will implement this, in the UK
the Government seems unwilling to, having only conceded to review the functionality after
significant parliamentary pressure. This seems odd, given the UK has already nominated
“super-complaint” non-governmental organizations (NGOs) that can, for example, raise consumer or financial rights issues to regulators on behalf of groups they perceive as affected.
Bodies like this might become effective watchdogs in particular areas or sectors, but they
will find it difficult to do this without having the capability for some access in the round to
training data, input data, outputs, and models of algorithmic systems in order to establish
whether breaches are occurring. A large problem here is that courts have typically been
reluctant to order access to source code for decision-making systems even in relation to traditional non-algorithmic systems. This is of course partly because of the issue of proprietary
IP rights in code already noted above (see, for instance, Viacom v YouTube Civ. 2103 (LL) in
the US), but other issues may also be implicated.
In the UK, there appears to be no reported case where a court has ordered disclosure
of the source code of a decision support system to litigants, even in the surprisingly high
number of disputes involving public sector systems, where issues of copyright might have
been thought to be less prominent. It is interesting though that in at least one case, Northern
Metco Estates Ltd v Perth & Kinross District Council 1993 S.L.T. (Lands Tribunal) 28, the output
of a conventional though complex automated decision support system was doubted in respect
to its value without more information as to how it was generated. The system in question
calculated one factor (economic rent) to feed into a compensation valuation in cases of
compulsory acquisition of land. The court was disturbed at the lack of evidence it received
as to exactly how this calculation had been done but, interestingly, did not seem interested
to find out more but rather to exclude it from influence. It seems quite likely that courts
will be reluctant to become activists about disclosures of source code, let alone algorithmic
training sets and models, until they feel more confident of their ability to comprehend and
use such evidence—which may take some time.
This likely judicial reluctance to become involved in unravelling algorithmic systems is
unfortunate because, at least in the public sector, a very valuable tool for transparency might
be found in the institution of judicial review. In the UK, courts have the power on petition
to review if administrative acts and discretion were legal and reasonable. In principle, there
seems no reason why the court’s powers in such review might not extend to demanding
12
access to data and models. However, as noted above, there is as yet no record of this having
been done. This may be due to a combination of proprietary code issues and technophobia,
but there may also be worries about exposing sensitive personal data to the public eye (court
documents are usually public).
One way forward here may be drawn from developing practice around sensitive public
sector data and access to it by non-governmental data analysts. In recent years, much work
has been put into building methods by which sensitive data can be accessed and analyzed
securely, proportionally, and safely. National statistics offices have been working on ways
to enable access to these rich troves of information by external data analysts without compromising security and privacy. Although several methods have been used here, the most
interesting for our purposes is the development of secure environments for access to sensitive data. These locations are usually disconnected from a network and may be physically
located in a public agency, private company, or increasingly, academia.
We imagine that such systems might scale to become depositories for the scrutiny of the
models and data associated with public sector algorithmic systems. Challenges to private
sector systems might also benefit from such systems, albeit with IP issues to overcome.
Archival or specialist libraries may yet become homes for infrastructure to publicly scrutinize
models and code, just as they already often act as code escrow agencies.
As things stand, the right to an explanation found in the GDPR, though beguiling, is uncertain, convoluted, rife with technical difficulties, and likely to be interpreted differently in
different member states. There are key restrictions in the remedies it gives, which are likely
to exclude the decision support systems where transparency would be most welcomed: in
crucial sectors such as criminal justice, policing, and child protection where decisions are not
solely based on automated processing. As a rights-based remedy, it places a large burden
on individual users to not only seek their own explanations but follow up with challenges.
It also does not work well as a remedy for harms experienced in aggregate by groups or
protected classes. There is a danger that the hype around the right to an explanation may
create the belief that there is no need for further legal and non-legal solutions. Some newer
legal instruments, such as the new French law, offer refinements to the GDPR but may also
add their own restrictions.
Accordingly, we proposed a wider sweep of attention toward other legal and paralegal remedies that may also have scope to impel the creation of better and more scrutable algorithmic
systems. These include privacy by design, data protection impact assessments, and certification or seal schemes, all of which might offer scrutiny during the process of building systems
rather than merely after they cause harm. We also suggested that redress for users after
a system has caused harm could be improved by taking the opportunities the GDPR gives
for representation and collective action by non-governmental bodies. Finally, we noted that
access to source code and data of algorithms will remain a practical issue for NGOs as well
as in contested legal cases for as long as courts are reluctant to order its disclosure, and
suggested secure model depositories might help in this area.
13
Acknowledgments
Lilian Edwards was supported in part by the Arts and Humanities Research Council (AHRC)
Centre CREATe (grant number AH/K000179/1), and the EPSRC Digital Economy Hub Horizon at the University of Nottingham (grant number EP/ G065802/1). Michael Veale acknowledges funding from the Engineering and Physical Sciences Research Council (EPSRC)
(grant number EP/ M507970/1).
References
1. L. Edwards and M. Veale, “Slave to the Algorithm? Why a “Right to an Explanation”
Is Probably Not the Remedy You Are Looking For,” Duke Law & Technology Review, vol.
16, 2017, pp. 18–84; DOI: 10.2139/ssrn.2972855.
2. F. Pasquale, The Black Box Society: The Secret Algorithms That Control Money and Information, Harvard University Press, 2015.
3. M. Veale and L. Edwards, “Clarity, Surprises, and Further Questions in the Article 29
Working Party Draft Guidance on Automated Decision-Making and Profiling,” Computer Law & Security Review, vol. 34, no. 2, 2018; DOI: 10.1016/j.clsr.2017.12.002.
4. L. Sweeney, “Discrimination in Online Ad Delivery,” Queue, vol. 11, no. 3, 2013, p.
10.
5. I. Mendoza and L.A. Bygrave, “The Right Not to Be Subject to Automated Decisions
Based on Profiling,” EU Internet Law: Regulation and Enforcement, T.-E. Synodinou et
al., eds., Springer, 2017.
6. S. Wachter et al., “Why a Right to Explanation of Automated Decision-Making Does
Not Exist in the General Data Protection Regulation,” International Data Privacy Law,
vol. 7, no. 2, 2017, pp. 76–99; DOI: 10.1093/idpl/ipx005.
7. M.T. Ribeiro et al., “Why Should I Trust You?: Explaining the Predictions of Any Classifier,” Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, 2016, pp. 1135–1144.
8. A.B. Tickle et al., “The Truth Will Come to Light: Directions and Challenges in Extracting the Knowledge Embedded within Trained Artificial Neural Networks,” IEEE Trans.
Neural Netw., vol. 9, no. 6, 1998, pp. 1057–1068.
9. G. Montavon et al., “Explaining Nonlinear Classification Decisions with Deep Taylor
Decomposition,” Pattern Recognit., vol. 65, 2017, pp. 211–222.
10. R. Binns, “Data Protection Impact Assessments: A Meta-Regulatory Approach,” International Data Privacy Law, vol. 7, no. 1, 2017, pp. 22–35; DOI: 10.1093/idpl/ipw027.
14
11. “Big Data, Artificial Intelligence, Machine Learning and Data Protection,” Information
Commissioner’s Office, 2017.
12. D.K. Citron, “Technological Due Process,” Washington University Law Review, vol. 85,
2008, pp. 1249–1313.
13. K. Crawford and J. Schultz, “Big Data and Due Process: Toward a Framework to Redress Predictive Privacy Harms,” Boston College Law Review, vol. 55, 2014, pp. 93–128.
14. A. Charlesworth, “Clash of the Data Titans—US and EU Data Privacy Regulation,” Eur.
Pub. L., vol. 6, 2000, p. 253.
15
| 2cs.AI
|
Learning Binary Residual Representations for Domain-specific Video Streaming
Yi-Hsuan Tsai1
Ming-Yu Liu2
arXiv:1712.05087v1 [cs.CV] 14 Dec 2017
1
Deqing Sun2
University of California, Merced
Abstract
We study domain-specific video streaming. Specifically, we
target a streaming setting where the videos to be streamed
from a server to a client are all in the same domain and
they have to be compressed to a small size for low-latency
transmission. Several popular video streaming services, such
as the video game streaming services of GeForce Now and
Twitch, fall in this category. While conventional video compression standards such as H.264 are commonly used for this
task, we hypothesize that one can leverage the property that
the videos are all in the same domain to achieve better video
quality. Based on this hypothesis, we propose a novel video
compression pipeline. Specifically, we first apply H.264 to
compress domain-specific videos. We then train a novel
binary autoencoder to encode the leftover domain-specific
residual information frame-by-frame into binary representations. These binary representations are then compressed and
sent to the client together with the H.264 stream. In our experiments, we show that our pipeline yields consistent gains
over standard H.264 compression across several benchmark
datasets while using the same channel bandwidth.
Introduction
Video streaming services, such as Netflix and YouTube, are
popular methods of viewing entertainment content nowadays. Due to large video sizes and limited network bandwidth, video compression is required for streaming video
content from a server to a client. While video compression
can reduce the size of a video, it often comes with undesired compression artifacts, such as image blocking effects
and blurry effects.
Decades of efforts were made towards delivering the best
possible video quality under bandwidth constraint. Stateof-the-art video compression methods such as MPEG-4 (Li
2001), H.264 (Wiegand et al. 2003), and HEVC (Sullivan
et al. 2012) combine various classical techniques including
image transform coding, predictive coding, source coding,
and motion estimation in a carefully-engineered framework.
These methods are general and can be applied to various
video domains for effectively compressing most of the information in a video. However, the residual information, which
is the difference between the uncompressed and compressed
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Ming-Hsuan Yang12
2
Jan Kautz2
NVIDIA
videos, is difficult to compress because it contains highly
non-linear patterns. Neither linear predictive coding nor linear transform coding can effectively compress the residual.
In this paper, we hypothesize that one can effectively
compress the residual information if one is willing to limit
the use of the video compression method to a specific domain. In other words, we no longer wish to have a video
compression method that works universally well on all
videos. We only wish to have a method that works particularly well in one specific domain. Although this setting
may appear inconvenient at a first glance as one needs to
have a video compressor for each domain, it does fit well
with several major video streaming applications, such as
video game streaming and sports streaming. For example,
for video game streaming services such as GeForce Now
and Twitch, the gamer chooses a game title to play, and
the video content is rendered on the server and delivered to
client’s mobile console. During the game playing period, all
the video content in the stream is in the same video game
domain and a domain-specific video compression method
is entirely appropriate. The setting also fits other user cases
such as streaming sports, which are often limited to a particular discipline, as well as things like compressing dash cam
videos, as all the videos are about street scenes.
To verify our hypothesis, we leverage deep learning models, which have been established as powerful non-linear
function approximators, to encode the highly nonlinear
residual information. In our video compression pipeline, we
first apply H.264 to compress videos in a specific domain
and train a novel binary autoencoder to encode the resulting
residual information frame-by-frame into a binary representation. We then apply Huffman coding (Cover and Thomas
2006) to compress the binary representations in a lossless
manner. The compressed binary representations can be sent
to the client in the meta data field in the H.264 streaming
packet. This way, our method can be integrated into the existing video streaming standard. We illustrate our method in
Figure 1. We show that with our proposed binary residual
representation, one can achieve a better video quality (at a
chosen bandwidth) by sending the video stream using H.264
at a smaller bandwidth and utilizing the saved bandwidth to
transmit the learned binary residual representation.
We conduct extensive experiments to verify our hypothesis that we can improve state-of-the-art video compression
methods by learning to encode the domain-specific residual
information in a binary form. We also compare various ways
of training the proposed binary autoencoder for encoding the
residual information. On the KITTI (Geiger, Lenz, and Urtasun 2012) and three games video datasets, we show that our
method consistently outperforms H.264 both quantitatively
and qualitatively. For example, our PSNR score is 1.7dB better than H.264 on average under the bandwidth at 5Mbps.
Related Work
We review the related works by their categories.
Image/Video Compression
Transform coding using the discrete/integer cosine transform (DCT/ICT) has been widely used in image and video
coding standards, such as JPEG (Wallace 1991), MPEG-4
(Li 2001), H.264 (Wiegand et al. 2003), and HEVC (Sullivan et al. 2012). The encoder divides an image/video frame
into non-overlapping blocks, applies DCT/ICT to each individual block, and quantizes the coefficients. Because of the
energy concentration ability of DCT/ICT, the compressed
images are of good quality at moderate to high bit rates.
However, real-time video stream requires very low bit-rate
compression. As a result, the compressed images often suffer from blocking artifacts due to the block-wise processing.
Another problem with existing coding standards is that they
have been designed to be universal coders and cannot be tailored to specific domains.
Machine learning-based techniques have been developed to aid these compression standards. A colorization
model (Cheng and Vishwanathan 2007) is utilized to learn
the representative color pixels to store for better reconstruction. Another work adopts a dictionary-based approach (Skretting and Engan 2011) to learn sparse representations for image compression. Our method is also based on
learning, in which we leverage state-of-the-art deep learning models to compress the residual information. Recently,
a reinforcement learning scheme is adopted to adaptively select the bit rate for video streaming for minimizing the latency (Mao, Netravali, and Alizadeh 2017).
Deep Learning-based Image Compression
Instead of engineering every step in the coding system,
numerous learning-based approaches as discussed in Jiang
et al. (Jiang 1999) have been developed to compress the
data in the holistic manner. Recently, autoencoders (Hinton and Salakhutdinov 2006) have been widely used in extracting abstract representations of images through learning to reconstruct the input signals (Vincent et al. 2008;
Pathak et al. 2016; Tsai et al. 2017). While these methods
use a bottleneck layer to learn a compact representation,
each dimension of the compact representation is continuous,
which needs to be further quantized for compression.
Various approaches are utilized in the bottleneck layer to
compress the data. Quantization methods (Ballé, Laparra,
and Simoncelli 2017; Theis et al. 2017) estimate the entropy
rate of the quantized data while optimizing used number of
bits. In addition, an adversarial training scheme (Rippel and
Bourdev 2017) is developed to produce sharper and visually
pleasing results. On the other hand, variants of the autoencoder architecture with recurrent networks show the ability to directly learn binary representations in the bottleneck
layer (Toderici et al. 2016; 2017).
While theses methods only focus on still image compression, our algorithm is designed to improve video quality
in streaming applications, especially in the domain-specific
scenarios such as game or street view videos. To this end,
we integrate existing video compression standards that can
effectively exploit temporal information and learning-based
methods that can efficiently transmit binary residual representations. As a result of the integration, our system can be
adaptively applied to existing video compression platforms
and improve their performance.
Post-processing to Remove Compression Artifacts
Post-processing at the decoder end aims at reducing the
coding artifacts without introducing additional bit rates at
the encoder end. Earlier approaches (Reeve and Lim 1983;
Chen, Wu, and Qiu 2001) manually use smoothing filters to reduce blocking effects caused by DCT in the expense of more blurry image outputs. Recently, learningbased methods (Chang, Ng, and Zeng 2014; Choi et al. 2015;
Liu et al. 2015; Mao, Shen, and Yang 2016) are developed
to model different compression artifacts. For example, the
ARCNN method (Dong et al. 2015) uses end-to-end training for removing various JPEG artifacts. Furthermore, the
D3 scheme (Wang et al. 2016) employs the JPEG priors to
improve the reconstruction results. Deeper models such as
DDCN (Guo and Chao 2016) are also developed to eliminate
artifacts, and a recent work (Guo and Chao 2017) combines
the perceptual loss to generate visually pleasing outputs.
We note that these post-processing methods require extensive computation on the client side, which is not well-suited
for embedded devices. In contrast, our method encodes the
residual information in a binary form on the server side and
send it to the client. Utilizing the encoded residual information, the client can better recover the original video using
much less computational power.
Domain-specific Video Compression with
Residual Autoencoder
Here, we first introduce our video compression pipeline for
streaming domain-specific videos. We then present the details of our autoencoder for encoding the residual information into binary forms and the training methods.
Video Streaming Pipeline
The proposed pipeline consists of two modules: a video
compression module and a deep learning-based autoencoder, as shown in Figure 1. The video compression module adopts the H.264 standard, which has demonstrated good
performance in compressing temporally smooth visual signals, while our residual autoencoder assists to recover the
lost information during compression on the client side by
leveraging the property that the videos to be compressed are
from the same domain. Using such a hybrid approach, not
Figure 1: Overview of our proposed video streaming pipeline. It consists of two modules: a conventional H.264 module and our
proposed residual autoencoder. The input to our residual module is the difference between the original and compressed videos.
The difference is encoded and binarized to generate binary representations. We utilize Huffman coding to further compress the
binary representations into a bit stream in a lossless manner. On the client side, we reconstruct the output video by adding back
the decoded difference to the compressed video.
only can we improve the output quality by spending a small
amount of effort, but also the system can adapt to existing
compression platforms and train for specific domains by exploiting large-scale data. Note that, although we use H.264
in our pipeline, other video compression standards such as
MPEG4 and HEVC can be used as well.
Given an input video X, we obtain the compressed video
Y by applying H.264. The difference between the two
videos is called the residual information R = X − Y. The
larger the residual information, the poorer the compressed
video quality. We also note that R is not included in Y because it consists of highly non-linear patterns, which can not
be compressed effectively with conventional approaches.
We argue that by limiting the video domain, we could
leverage a novel autoencoder to effectively compress the
residual information. The autoencoder consists of a pair of
functions (E, D), where the encoder E maps R to a binary
map and the decoder D recovers R from the binary map
on the client side. The recovered residual information is referred to as R̂ and the final output video Y + R̂ has a better
visual quality than Y. We note that the binary map is further
mapped to a bit stream by using the Huffman coding algorithm (Cover and Thomas 2006), which is asymptotically
optimal, to reduce its bandwidth usage.
Sending the bit stream of the residual information requires
additional bandwidth. However, we can train an autoencoder
that only requires a much smaller bandwidth to compress
the residual information than H.264. Therefore, we can run
the H.264 standard in a higher compression rate, which uses
a smaller bandwidth but results in a larger residual signal.
We then apply our autoencoder to compress the residual signal into a small bit stream. Considering a scenario where
the bandwidth for a video stream is 5Mbps, we can apply
the proposed pipeline to compress the video in 4Mbps using H.264 and utilize the remaining 1Mbps for sending the
residual signal. Because our autoencoder is more efficient
than H.264 in compressing the residual signal, our system
achieve better performance than a baseline system that allocates all the 5Mbps for H.264.
One may wonder why not completely replacing the H.264
standard with the proposed residual autoencoder. We argue
that our residual autoencoder is only more efficient than
H.264 in compressing the residual signal. The carefullyengineered H.264 is more efficient in compressing the core
video. By marrying the strength of H.264 and the proposed
autoencoder, our hybrid system achieves better performance.
Moreover, our pipeline can be easily integrated into the existing H.264 standard since the residual information can
be attached in the meta field of a H.264 streaming packet.
Hence we can enjoy the popularity of the H.264 standard
and various hardware accelerators implemented for H.264.
We note that in the proposed domain-specific video
streaming pipeline, one needs to send the parameters of D
and the Huffman decoder table to the client for the streaming service, which requires an additional bandwidth. However, the parameters can be sent before the streaming starts.
Once the parameters are sent, the user can enjoy the low latency and high video quality features of the proposed video
streaming pipeline.
Binary Residual Autoencoder
We design an autoencoder that consists of three components:
encoder E, binarizer B (introduced in the next section) and
decoder D. For the encoder, the goal is to learn to extract
compact feature representations for the following binarizer.
We use L convolutional layers in our encoder, in which each
layer has the same channel number C and a stride of two
that down-samples feature maps. The binarizer converts the
output from the last convolutional layer into a binary map.
For the decoder, we aim to up-sample the binary map back to
the original input. Our decoder has L convolutional layers.
At the end of each convolutional layer, a sub-pixel layer (Shi
Figure 2: Architecture of the proposed binary residual autoencoder. It process the video frame by frame. The autoencoder
consists of an encoder, a binarizer and a decoder. The encoder/decoder, each has L convolutional layers and each layer contains
C/4C channels. For simplicity, Huffman coding modules are not illustrated here.
et al. 2016) is used for up-sampling. Since we aim for upsampling the resolution of the feature map by two on each
spatial dimension, the channel number of each convolutional
layer in the decoder is set to 4 × C due to the use of the subpixel layer. To facilitate the learning process, batch normalization (Ioffe and Szegedy 2015) and ReLU layers are used.
The architecture of the autoencoder are given in Figure 2.
We encode and decode the residual signal R frame by
frame. Let {ri } be a set of residual frames computed by applying H.264 at a target bit rate. We train our autoencoder
by solving the following optimization problem:
X
min
||ri − D(B(E(ri )))||22 .
(1)
D,E
i
We care a lot about the bandwidth required for transmitting
binary representations, which is determined by two factors:
1 the number of layers L in the encoder, and 2) the number
of channels C. Let W and H be the width and height of the
×H
input image, the binary map size is given by C×W
.A
22L
large number of L and a smaller number of C would result
in a smaller size of the encoder output and hence a smaller
binary map. However, a smaller encoder output makes training of the autoencoder difficult. In our experiments we will
discuss the results with different numbers of L and C.
Training of the Binary Residual Autoencoder
Binarizing feature maps in neural networks have been studied in several earlier works (Rastegari et al. 2016; Courbariaux et al. 2016; Tang, Hua, and Wang 2017) for the
purpose of reducing memory footprint in mobile devices. It
is also used for image compression (Toderici et al. 2016;
2017), while our work is different in that we binarize feature
maps for video streaming. We will discuss several binarization methods here and compare their advantages and drawbacks when used in our pipeline in the experiment section.
Formulation. Let the output feature map of E be ei = E(ri ).
Our binarizer B aims to produce the binary output {−1, 1}1 .
To generate such binary outputs, the process B consists of
1
We have found that our network requires negative responses to
achieve reasonable results. Instead of producing the binary output
{0, 1}, we aim to generate the discrete output {−1, 1}.
two parts: 1) map each element of the encoder output ei to
the interval [−1, 1], and 2) discretize it to {−1, 1} given by:
B(ei ) = b(σ(ei )),
(2)
where σ and b are the activation and discretization functions, respectively. In the following, we discuss different
functions for activation (i.e., tanh, hardtanh, sigmoid) and
various methodologies for binarization (i.e., stochastic regularization (Raiko et al. 2015) , Gumbel noise (Jang, Gu, and
Poole 2017; Maddison, Mnih, and Teh 2017)).
Tanh/Hardtanh Activation. tanh is a common activation
to project feature values to [−1, 1], so as the approximation version hardtanh function. Here, we define the binarized output as:
1,
if z ≥ 0
b(z) =
(3)
−1, if z < 0,
where z = σ(ei ) and σ can be the tanh or hardtanh function. However, since binarization is not a differentiable function, we can not train the proposed autoencoder using backpropagation. To avoid the issue, inspired by the recent binarization work (Courbariaux et al. 2016), we adopt a piecewise function bbp during back-propagation:
if z > 1
1,
bbp (z) = z,
(4)
if −1 ≤ z ≤ 1
−1, if z < −1.
By using the straight-through estimator (Courbariaux et al.
2016), we can compute the gradient of bbp (z) and pass gradients through b unchanged as:
1, if −1 ≤ z ≤ 1
b0bp (z) =
(5)
0, otherwise.
Sigmoid Activation. The sigmoid function outputs a value
in [0, 1]. We convert the output to [−1, 1] by applying z =
2(σ(ei ) − 0.5). We can then use the approach discussed in
the previous paragraph for binarization and training.
Stochastic Regularization. Following Toderici et
al. (Toderici et al. 2016), we incorporate a stochastic
process into (3) using the tanh activation:
b(z) = z + ,
Table 1: Compression ratio versus number of bits in a group
computed from our KITTI dataset experiments.
# of bits in a group
Compression ratio
8
1.05
16
1.47
32
3.03
Videos
Frames
64
6.67
where is a randomized quantization noise, resulting in:
1,
with probability 1+z
2
(6)
b(z) =
−1, with probability 1−z
2 .
For back-propagation, similarly we can pass unchanged gradients through b (Toderici et al. 2016).
Gumbel Noise. The Gumbel-Softmax distributions have
been utilized for learning categorical/discrete outputs (Jang,
Gu, and Poole 2017). We adopt a similar approach by adding
Gumbel noise in the sigmoid function, which we refer as
Gumbel-Sigmoid. Since sigmoid can be viewed as a special
2-class case (ei and 0 in our case) of softmax, we derive the
Gumbel-Sigmoid as:
σ(ei ) =
exp((ei + gk )/τ )
,
exp((ei + gk )/τ ) + exp(gl /τ )
Table 2: Number of videos and frames on the datasets.
(7)
where g is the sample from the Gumbel noise and τ is the
temperature that controls how closely the Gumbel-Sigmoid
distribution approximates the binary distribution. From (7),
we can further simplify it as:
KITTI
50
19,057
Assassins Creed
50
34,448
Table 3: Impact of L and C
Group Mbps (C, L)
(8, 3)
1
∼0.24
(32, 4)
(16, 3)
2
∼0.48
(64, 4)
(8, 2)
3
∼0.96
(32, 3)
Skyrim
9
9,337
Borderlands
19
8,752
on PSNR/SSIM.
PSNR SSIM
29.28 0.8571
29.39 0.8577
29.83 0.8681
29.81 0.8652
30.35 0.8855
30.67 0.8901
We report PSNR and SSIM scores at different bit rates
for quantitative comparisons. These two metrics are popular
performance metrics for benchmarking video compression
algorithms (Wang et al. 2004). We first conduct an ablation
study using the KITTI dataset where we discuss the impact
of layer and channel numbers in our design. We then compare various methods for training the autoencoder. Finally,
we report the performance with comparisons to H.264 and a
deep learning baseline.
(8)
Implementation Details. Throughout the paper, we use
Adam (Kingma and Ba 2015) to train our binary residual
autoencoder. The learning rate is set to 10−3 and then decreased by half for every 5 epochs. The momentums are set
to 0.9 and 0.999. The batch size is 10, and we train the model
for 50 epochs. The implementation is based on PyTorch.
where sigm is the sigmoid function. Since σ(ei ) is still in
the range of [0, 1], we adopt the same approach as introduced
before to shift the value via z = 2(σ(ei ) − 0.5). Following
(Jang, Gu, and Poole 2017), we start with a high temperature
and gradually anneal it to a small but non-zero temperature.
Runtime Analysis. On the server side, our encoder and the
binarizer takes about 0.001 seconds to compress the residual image with a resolution of 360 × 1200 using a Titan X
GPU. The decoder on the client side takes 0.001 seconds to
reconstruct the residual image.
1
1 + exp(−(ei + gk − gl )/τ )
= sigm((ei + gk − gl )/τ ),
σ(ei ) =
Lossless Compression. Once we generate the binary feature
map, we further use Huffman coding (Cover and Thomas
2006) to reduce the size of the binary representation. These
coding schemes are asymptotically optimal, which means
that the more the number of bits we group together as a
symbol, the better the compression ratio. Such a behavior
is illustrated in Table 1 using the KITTI dataset. We note
that other source coding methods such as arithmetic coding
(Marpe, Schwarz, and Wiegand 2003) can be used as well.
Experimental Results
We evaluate our pipeline using the KITTI (Geiger, Lenz, and
Urtasun 2012) dataset, which consists of various driving sequences of street scenes, and three popular video games:
Assassins Creed, Skyrim and Borderlands. The details of
the datasets are shown in Table 2. We use the tracking
benchmark on the KITTI dataset that contains 50 street-view
videos. We randomly select 42 videos for training and 8
videos for testing. The images in the videos are resized to
360 × 1200. The game video resolutions are 720 × 1280.
Ablation Study
Depth and Breadth of the Model. We analyze the impact
of layer numbers L and channel numbers C on the video
compression performance of our pipeline. For this study, the
bandwidth of H.264 is set to 5Mbps for generating training
and testing videos. We then train the residual autoencoder
with different L and C utilizing the hardtanh activation for
binarization. These two parameters determine the size of the
binary map, which has impacts on the compression rate as
well as the training difficulty. Intuitively, the smaller the binary map, the easier the compression task but the harder the
training task. The results are shown in Table 3. Based on the
bit rate, we divide the settings into three groups. The ones
in the same group have a very similar bit rate after applying
the Huffman coding. We then use the best setting in each bit
rate group for the rest of the experiments.
Binarization. We compare various methods discussed earlier for training the binary residual autoencoder. We fix
L = 3 and C = 32 and train our model with different
(a) KITTI
(b) Assassins Creed
(c) Skyrim
(d) Borderlands
Figure 3: PSNR comparisons on four datasets at different bandwidths. We compare our pipeline with H.264 and an artifactremoval method based on (Kim, Lee, and Lee 2016; Zhang et al. 2017).
(a) KITTI
(b) Assassins Creed
(c) Skyrim
(d) Borderlands
Figure 4: SSIM comparisons on four datasets at different bandwidths. We compare our pipeline with H.264 and an artifactremoval method based on (Kim, Lee, and Lee 2016; Zhang et al. 2017).
Figure 5: Example results on the KITTI and video game datasets. We compare our pipeline with H.264 and an artifact-removal
method. The corresponding bit rate and PSNR are shown next to the images. Best viewed with enlarged images.
Table 4: Comparisons of different
the KITTI dataset using PSNR.
H.264 @ Mbps
1M
Hardtanh
28.95
Tanh
28.95
Sigmoid
28.89
Stochastic
28.26
Gumbel-Sigmoid 26.66
binarization methods on
2M
29.97
29.97
29.88
29.21
27.82
5M
30.67
30.67
30.61
29.78
28.66
H.264 compression rates. The results are reported in Table 4 and 5. We find that the models trained using hardtanh and tanh activations consistently outperform the others.
The model trained with the stochastic regularization does
not perform well possibly due to the training difficulty. It is
known that the stochastic noise increases the gradient variance. In addition, we empirically find that the model using Gumbel-Sigmoid performs much worse for the residual signal compression task. (We note that (Jang, Gu, and
Poole 2017) uses the Gumbel noise for reconstructing discrete/categorical signals but not continuous signals of images in our task.). Hence, we use hardtanh activations in the
rest of the experiments for its superior performance.
Table 5: Comparisons of different
the KITTI dataset using SSIM.
H.264 @ Mbps
1M
Hardtanh
0.8575
Tanh
0.8577
Sigmoid
0.8538
Stochastic
0.8297
Gumbel-Sigmoid 0.7683
binarization methods on
2M
0.8773
0.8770
0.8722
0.8470
0.8128
5M
0.8901
0.8882
0.8861
0.8617
0.8464
To validate the importance of modeling the residual, we
also carry out experiments by directly compressing the raw
video using the same autoencoder architecture. However, it
results in a worse performance compared to H.264 (0.52dB
drop in PSNR) since this method does not leverage any motion information for video compression. Overall, our results
show that by propagating the extracted binary residual representations from the server to the client, the quality of reconstructed videos can be largely improved. It outperforms the
artifact removal network, which aims for solving a challenging inverse problem and does not leverage any prior knowledge about the compression process. In addition, the runtime
of our binary residual autoencoder (decoding on the client
side) is two times faster than the artifact removal network.
Video Compression Performance
We compare our video compression pipeline to the H.264
standard and an artifact removal network, which is a popular
approach to reduce distortions. Following recent deep learning works in image enhancement and denoising (Kim, Lee,
and Lee 2016; Zhang et al. 2017), we utilize a neural network with 8 convolutional layers (no strides) to remove the
artifacts. The network takes the compressed H.264 images
as inputs in a frame-by-frame fashion and outputs enhanced
images. We then train an artifact removal network for each
video domain for fair comparisons.
We note that the proposed pipeline requires bandwidth for
both H.264 and the binary map. We account both in the bit
rate calculation for fair comparisons with the baseline methods. In other words, the bit rate of the proposed pipeline is
the sum of the bit rate from the H.264 stream and the Huffman code of the binary map. Again, we do not take transmitting the network parameters into account since it can be sent
before the streaming starts. In the streaming services, the
main goal is to have high quality videos with low latency.
We report PSNR and SSIM scores on the KITTI benchmark and 3 video games in Figures 3 and 4. We find our
pipeline consistently outperforms the baseline methods at
all bit rates. Our pipeline achieves a PSNR of 33.26dB at
5Mbps averaged on four datasets, which is 1.71dB better
than H.264 and 0.84dB better than the artifact-removal network. Similarly, our pipeline performs better in SSIM, e.g.,
5.3% and 4.1% improvements over H.264 and the artifactremoval network at 2Mbps, respectively. In Figure 5, we
present some qualitative results, showing our method preserves more details and textured contents (e.g., tree, rock and
water) in reconstructed images using a smaller bandwidth2 .
2
The website link http://research.nvidia.com/publication/2018-
Conclusions
In this paper, we propose a video streaming system that integrates H.264 and a binary residual autoencoder to encode
non-linear compression errors for domain-specific video
streaming. We analyze various network design choices and
methods for obtaining binary representations of the residual information. The binary representations are further compressed and transmitted from the server to the client. On
the KITTI benchmark dataset and three popular video game
datasets, the proposed algorithm generates better reconstructed videos than H.264 and artifact-removal methods
while using a smaller bandwidth.
Acknowledgment The authors would like to thank Sam H.
Azar and Zheng Yuan for their helpful discussions.
References
Ballé, J.; Laparra, V.; and Simoncelli, E. P. 2017. End-to-end
optimized image compression. In ICLR.
Chang, H.; Ng, M. K.; and Zeng, T. 2014. Reducing artifacts in jpeg decompression via a learned dictionary. IEEE
Transactions on Signal Processing 62(3):718–728.
Chen, T.; Wu, H. R.; and Qiu, B. 2001. Adaptive postfiltering of transform coefficients for the reduction of blocking
artifacts. IEEE Transactions on Circuits and Systems for
Video Technology 11(5):594602.
Cheng, L., and Vishwanathan, S. V. N. 2007. Learning to
compress images and videos. In ICML.
Choi, I.; Kim, S.; Brown, M. S.; and Tai, Y.-W. 2015. A
learning-based approach to reduce jpeg artifacts in image
matting. In ICCV.
02 LearningBinaryResidual contains more visualization results.
Courbariaux, M.; Hubara, I.; Soudry, D.; El-Yaniv, R.; and
Bengio, Y. 2016. Binarynet: Training deep neural networks
with weights and activations constrained to +1 or -1. CoRR
abs/1602.02830.
Cover, T. M., and Thomas, J. A. 2006. Elements of Information Theory. Wiley-Interscience.
Dong, C.; Deng, Y.; Loy, C. C.; and Tang, X. 2015. Compression artifacts reduction by a deep convolutional network. In ICCV.
Geiger, A.; Lenz, P.; and Urtasun, R. 2012. Are we ready
for autonomous driving? the kitti vision benchmark suite. In
CVPR.
Guo, J., and Chao, H. 2016. Building dual-domain representations for compression artifacts reduction. In ECCV.
Guo, J., and Chao, H. 2017. One-to-many network for visually pleasing compression artifacts reduction. In CVPR.
Hinton, G., and Salakhutdinov, R. 2006. Reducing the
dimensionality of data with neural networks. Science
313(5786):504–507.
Ioffe, S., and Szegedy, C. 2015. Batch normalization: Accelerating deep network training by reducing internal covariate
shift. In ICML.
Jang, E.; Gu, S.; and Poole, B. 2017. Categorical reparameterization with gumbel-softmax. In ICLR.
Jiang, J. 1999. Image compression with neural networks a
survey. In Signal Processing: Image Communication, 737–
760.
Kim, J.; Lee, J. K.; and Lee, K. M. 2016. Accurate image
super-resolution using very deep convolutional networks. In
CVPR.
Kingma, D. P., and Ba, J. 2015. Adam: A method for
stochastic optimization. In ICLR.
Li, W. 2001. Overview of fine granularity scalability in
mpeg-4 video standard. IEEE Transactions on Circuits and
Systems for Video Technology 11(3):301–317.
Liu, X.; Wu, X.; Zhou, J.; and Zhao, D. 2015. Datadriven sparsity-based restoration of jpeg-compressed images
in dual transform-pixel domain. In CVPR.
Maddison, C. J.; Mnih, A.; and Teh, Y. W. 2017. The concrete distribution: A continuous relaxation of discrete random variables. In ICLR.
Mao, H.; Netravali, R.; and Alizadeh, M. 2017. Neural
adaptive video streaming with pensieve. In Proceedings of
the ACM SIGCOMM.
Mao, X.-J.; Shen, C.; and Yang, Y.-B. 2016. Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections. In NIPS.
Marpe, D.; Schwarz, H.; and Wiegand, T. 2003. Contextbased adaptive binary arithmetic coding in the h.264/avc
video compression standard. IEEE Transactions on Circuits
and Systems for Video Technology 13(7):620–636.
Pathak, D.; Krähenbühl, P.; Donahue, J.; Darrell, T.; and
Efros, A. A. 2016. Context encoders : Feature learning by
inpainting. In CVPR.
Raiko, T.; Berglund, M.; Alain, G.; and Dinh, L. 2015. Techniques for learning binary stochastic feedforward neural networks. In ICLR.
Rastegari, M.; Ordonez, V.; Redmon, J.; and Farhadi, A.
2016. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV.
Reeve, H., and Lim, J. 1983. Reduction of blocking effect
in image coding. In ICASSP, volume 8, 1212–1215.
Rippel, O., and Bourdev, L. 2017. Real-time adaptive image
compression. In ICML.
Shi, W.; Caballero, J.; Huszár, F.; Totz, J.; Aitken, A. P.;
Bishop, R.; Rueckert, D.; and Wang, Z. 2016. Real-time
single image and video super-resolution using an efficient
sub-pixel convolutional neural network. In CVPR.
Skretting, K., and Engan, K. 2011. Image compression using
learned dictionaries by rls-dla and compared with k-svd. In
ICASSP.
Sullivan, G. J.; Ohm, J.; Han, W.-J.; and Wiegand, T. 2012.
Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video
Technology 22(12):1649–1668.
Tang, W.; Hua, G.; and Wang, L. 2017. How to train a compact binary neural network with high accuracy? In AAAI.
Theis, L.; Shi, W.; Cunningham, A.; and Huszár, F. 2017.
Lossy image compression with compressive autoencoders.
In ICLR.
Toderici, G.; O’Malley, S. M.; Hwang, S. J.; Vincent, D.;
Minnen, D.; Baluja, S.; Covell, M.; and Sukthankar, R.
2016. Variable rate image compression with recurrent neural
networks. In ICLR.
Toderici, G.; Vincent, D.; Johnston, N.; Hwang, S. J.; Minnen, D.; Shor, J.; and Covell, M. 2017. Full resolution image
compression with recurrent neural networks. In CVPR.
Tsai, Y.-H.; Shen, X.; Lin, Z.; Sunkavalli, K.; Lu, X.; and
Yang, M.-H. 2017. Deep image harmonization. In CVPR.
Vincent, P.; Larochelle, H.; Bengio, Y.; and Manzagol, P.A. 2008. Extracting and composing robust features with
denoising autoencoders. In ICML.
Wallace, G. K. 1991. The jpeg still picture compression
standard. Communications of the ACM 30–44.
Wang, Z.; Bovik, A. C.; Sheikh, H. R.; and Simoncelli, E. P.
2004. Image quality assessment: from error visibility to
structural similarity. IEEE Transactions on Image Processing 13(4):600–612.
Wang, Z.; Liu, D.; Chang, S.; Ling, Q.; Yang, Y.; and Huang,
T. S. 2016. D3: Deep dual-domain based fast restoration of
jpeg-compressed images. In CVPR.
Wiegand, T.; Sullivan, G. J.; Bjontegaard, G.; and Luthra,
A. 2003. Overview of the h.264/avc video coding standard.
IEEE Transactions on Circuits and Systems for Video Technology 13(7):560–576.
Zhang, K.; Zuo, W.; Chen, Y.; Meng, D.; and Zhang, L.
2017. Beyond a gaussian denoiser: Residual learning of
deep cnn for image denoising. IEEE Transactions on Image Processing 26(7):3142–3155.
| 2cs.AI
|
arXiv:1607.04877v4 [math.AC] 3 Nov 2016
GABRIEL LOCALIZATION THEORY AND ITS
APPLICATIONS
ABOLFAZL TARIZADEH
Abstract. In this article first we develop the Gabriel localizations (abbreviated as G-localizations) for commutative rings, specially some new results in this direction are proven. Then, as an
application, it is shown that a ring map is a flat epimorphism if
and only if it corresponds to a kind of the G-localizations. As
a by-product of this study, a characterization for the flatness of
the quotient rings is given. The exactness of the G-localization
functor are characterized. The structure of prime ideals in the Glocalization rings are also studied. Finally, it is shown that the
Gabriel localization theory is a natural generalization of the usual
localization theory.
1. Introduction
Our main aim in this article is to analyze and then understand the
structure of flat epimorphisms of rings more deeply. Indeed, we have
successfully applied the Gabriel localization theory and thereby we have
found a new and simple proof to the fact that “a ring map is a flat epimorphism if and only if it corresponds to a type of the G-localizations”,
see Corollary 3.3 and Theorem 3.6. We should mention that, according to the knowledge of the author, there are some different proofs of
this theorem or a part of it in the literature. E.g. see [1], [5], [6],
[7] and [8, Chap. XI, Theorem 2.1]. These proofs are based upon
other enormous results and therefore it is practically very hard to follow and fully understand the proof. In fact, they use the theory of
Grothendieck categories and Giraud subcategories and also they use
the fact that “the set of idempotent topologizing systems on the ring
R is bijectively corresponding to the set of Giraud subcategories of the
category of R−modules”, see [8, Chap. X, Theorem 2.1]. While the
latter in turn is a huge result. Our approach, unlike their methods, is
based upon some simple observations. More precisely, after developing
the Gabriel localization theory for commutative rings we then obtain
0
2010 Mathematics Subject Classification: 13A99, 13B10, 13B30, 13C99.
Key words and phrases: G-localization, negligible module, flat epimorphism.
1
2
ABOLFAZL TARIZADEH
more general and new results, namely Theorems 2.14, 2.16 and 3.2.
These results pave the way in order to reach to a natural and simple
proof of the fact. Using these results then we are also able to characterize the exactness of the G-localization functor, see Theorem 4.4. Note
that the G-localization functor, in general, is left exact. Moreover, as
a by-product of this study, a characterization for the flatness of the
quotient rings is given, Corollary 3.4.
The localization theory, in particular the local rings, play a major role in commutative algebra, algebraic and arithmetic geometry,
number and valuation theories. The notion of the G-localization with
respect to an idempotent topologizing system, which is in turn a natural generalization of the usual localization theory, first appeared in the
Gabriel thesis [3, Chap V, §2] which was conducted under the supervision of Grothendieck. In the literature, an idempotent topologizing system is also called a Gabriel topology. The Gabriel localization theory
provides a very general method of localization which is even applicable
in noncommutative situations. In commutative algebra a large number of important constructions are special cases of the G-localizations.
The idea of the G-localization is essentially due to Grothendieck and it
appeared in more general setting in SGA 4, tomme 1, exposé II. In this
article, our presentation of the theory will closely follow the Gabriel’s
approach in Bourbaki [2, Les exercices 16 à 25].
The article is organized as follows. In Section 2, as mentioned in the
above, the Gabriel localization theory is developed for commutative
rings. The content of the third section was completely described in the
first paragraph. Regarding to this section, we should mention that this
expression of flat epimorphisms (Corollary 3.3) has some important applications. For example, using this result then one can show that every
injective flat epimorphism of rings which is also of finite type then it
is of finite presentation. The latter is also a highly non-trivial result
in commutative algebra. In the final section, in addition to the characterizing the exactness of the G-localization functor, the structure of
prime ideals in the G-localization rings are also studied. Theorems 4.4
and 4.10 are the main results of this section.
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
3
2. G-localizations
Throughout the article, all of the rings which we consider are commutative.
Definition 2.1. A topologizing system on the ring R is a non-empty
family F of ideals of R satisfying in the following conditions.
(a) Every ideal of R containing an ideal I ∈ F belongs to F .
(b) The family F is stable under finite intersections.
Example 2.2. Let R be a ring, let S be a multiplicative subset of R
and let P be a family of prime ideals of R. Examples of topologizing
systems are the single-point set {R}, the set of ideals of R containing
a fixed ideal, the set of ideals I of R such that I + J = R where J is a
fixed ideal of R, the set of ideals of R which meeting S, the set of ideals
I of R such that P ∩ V (I) = ∅, the set of ideals I of R such that V (I)
is contained in V (J) where J is a fixed ideal of R. See also Theorem
3.2 and [8, Chap. X, Theorem 2.1]. By V (I) we mean the set of prime
ideals p of R such that I ⊆ p.
Definition 2.3. Let F be a topologizing system on the ring R. An
R−module M is called F −negligible if the annihilator of every element
of M belongs to F . Clearly submodules, quotients, localizations and
finite direct sums of F −negligible modules are F −negligible.
Let F be a topologizing system on the ring R. For every R−module
M, the set F (M) = {m ∈ M : AnnR (m) ∈ F } is a R−submodule
of M because for every elements m, m′ ∈ F (M) and for each r ∈ R
we have Ann(m) ∩ Ann(m′ ) ⊆ Ann(m + m′ ) and Ann(m) ⊆ Ann(rm).
Clearly it is the greatest F −negligible submodule of M. Each R−linear
map u : M → N induces a map F (M) → F (N) given by m
u(m)
which we denote it by F (u). In fact F (−) is a left exact functor from
the category of R−modules to itself.
Let F be a topologizing system on the ring R. The family F with
the relation I < J if J is a proper subset of I, is a directed poset.
If I ≤ J then for each R−module M the canonical injection J ⊆ I
induces the R−linear map uI,J : HomR (I,
M) → Hom(J, M) given
by f
f |J . Clearly HomR (I, M), uI,J is an inductive system of
R−modules and R−homomorphisms over the directed poset (F , <).
4
ABOLFAZL TARIZADEH
We shall denote by M(F ) the inductive limit (colimit) of the system.
Therefore M(F ) = colimI∈F HomR (I, M). For each I ∈ F the canonical map HomR (I, M) → M(F ) , if there is no confusion, is denoted
≃
/ M(F ) is also
/ HomR (R, M)
by [ ]. The composed map M
denoted by δM . Clearly Ker(δM ) = F (M). Moreover Coker(δM ) is
F −negligible. Because let f : I → M be a R−linear map where
I ∈ F . For each a ∈ I, (δm )|I = a.f where m = f (a) and δm : R
→M
is given by r
rm. This means that I ⊆ AnnR [f ] + Im(δM ) .
Every R−linear map u : M → N, by the universal property of the
colimits, induces a unique R−linear map u(F ) : M(F ) → N(F ) such
that for each I ∈ F the following diagram is commutative
HomR (I, M)
/
u(F)
M(F )
HomR (I, N)
/
N(F )
where the columns are the canonical maps and the top row map is
given by f
u ◦ f . Therefore u(F ) is defined as [f ]
[u ◦ f ].
Lemma 2.4. Let F be a topologizing system on the ring R. If
v /
/ M′ u / M
M ′′ is an exact sequence of R−modules then
0
the sequence 0
/
′
M(F
)
u(F)
/
M(F )
v(F)
/
′′
M(F
) is exact.
Proof. It is an easy exercise.
Let F and G be two topologizing systems on the ring R. We denote
by F .G the set of ideals I of R such that there exists some J ∈ G containing I in which J/I is F −negligible. Clearly F .G is a topologizing
system on the ring R. Also IJ ∈ F .G for all I ∈ F and all J ∈ G . In
particular, F and G are contained in F .G . A topologizing system F
is called idempotent if F .F = F . All of the topologizing systems in
Example 2.2, except the second one, are idempotent.
Proposition 2.5. Let F and G be two topologizing systems on the
ring R and let M be a R−module. Then M is F .G −negligible if and
only if there exists a F −negligible submodule M ′ of M such that M/M ′
is G −negligible. In particular, if H is a third topologizing system on
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
5
the ring R then we have (F .G ).H = F .(G .H ).
Proof. If M is F .G −negligible then take M ′ = F (M) and the remaining assertions are straightforward. The converse is also routine.
Lemma 2.6. Let R be a ring and let F be a non-empty family of ideals
of R. Then F is an idempotent topologizing system on the ring R if
and only if it satisfies in the following conditions.
(i) Every ideal of R containing an ideal I ∈ F belongs to F .
(ii) If I ∈ F and J is an ideal of R such that J : a ∈ F for all a ∈ I,
then J ∈ F .
Proof. Suppose F is an idempotent topologizing system on the
ring R. Let I ∈ F and let J be an ideal of R such that J : a ∈ F
for all a ∈ I. It suffices to show that J ∈ F .F . But it is clear since
J ′ = I + J ∈ F and for each a ∈ I, AnnR (a + J) = J : a ∈ F .
Conversely, suppose I, J ∈ F . We show that I ∩ J ∈ F . But
J ⊆ IJ : a for all a ∈ I. Therefore IJ ∈ F and so I ∩ J ∈ F .
Hence, F is a topologizing system. Take J ∈ F .F . Thus there exists
some I ∈ F containing J such that I/J is F −negligible. Therefore
J : a ∈ F for all a ∈ I. Thus J ∈ F .
From now onwards, if it is not stated, F is always an idempotent
topologizing system on the ring R.
Here some new rings and modules are introduced. The basic set-up
is as follows. Let M be a R−module. For each I, J ∈ F and for
each f ∈ HomR (I, R) we
have f −1 (J) ∈ F . Because for each a ∈ I,
J ⊆ AnnR a + f −1 (J) . Therefore I/f −1 (J) is F −negligible and so
f −1 (J) ∈ F .F = F . The map f induces a map f −1 (J) → J which we
denote it as usual by f |f −1 (J) . Now
we define the pairing R(F ) ×M(F ) →
M(F ) as [f ].[g] = [g ◦ f |f −1 (J) ] where J is the domain of g. Note
that it is well-defined. Because let [f1 ] = [f2 ] and [g1 ] = [g2 ] where
fk ∈ HomR (Ik , R) and gk ∈ HomR (Jk , M) for some Ik , Jk ∈ F with
k = 1, 2. There exist L′ , L′′ ∈ F such that Ik ≤ L′ , Jk ≤ L′′ , (f1 )|L′ =
−1
′′
′′
′
(f2 )|L′ and (g1 )|L′′ = (g2 )|L′′ . Take L = f1−1 (L
) ∩ f2 (L ) ∩ L which
belongs to F and we have g1 ◦ (f1 )|f1−1 (J1 ) |L = g2 ◦ (f2 )|f2−1 (J2 ) |L .
The pairing is also R−bilinear (details omitted).
6
ABOLFAZL TARIZADEH
In particular, the binary operation R(F ) × R(F ) → R(F ) , as multiplicative, puts a commutative ring structure on the R−module R(F ) .
Its commutativity implies from this simple observation that for every R−linear maps f, g : I → R where
I is an ideal
of R and for
′
′
′
every elements a, a ∈ I then g f (aa ) = f g(aa ) . The canonical map δR : R → R(F ) is a ring homomorphism. Moreover, the
map R(F ) × M(F ) → M(F ) puts a R(F ) −module structure over M(F ) .
For every R−linear map u : M → N then u(F ) : M(F ) → N(F ) is
R(F ) −linear. Indeed, M
M(F ) is a left exact functor from the category of R−modules to the category of R(F ) −modules.
Definition 2.7. Let M be a R−module. The R(F ) −module
M/F (M) (F )
is called the Gabriel localization of M with respect to the system F and
it is denoted by MF . Therefore MF = colimI∈F HomR I, M/F (M) .
π
δ
/ M/F (M)
/ MF is denoted by jM .
The composed map M
Therefore for each m ∈ M, jM (m) = [δm ] where m = m + F (M) and
δm : R → M/F (M) is defined as r
rm + F (M).
Lemma 2.8. Let M be a R−module. Then Ker(jM ) = F (M) and
Coker(jM ) is F −negligible.
Proof. First note that
F
M/F
(M)
= 0 because if m = m +
F (M) ∈ F M/F (M) then Ann(m) : a = Ann(am) ∈ F for all
a ∈ AnnR (m). Therefore, by Lemma 2.6, Ann(m) ∈ F and so m =
0.
−1
We have Ker(jM ) = jM (0) = π −1 Ker(δ) = π −1 F M/F (M) =
Ker(π) = F (M). We also have Coker(jM ) = Coker(δ) therefore
Coker(jM ) is F −negligible.
Lemma 2.9. Let M be a R−module. Then M is F −negligible if and
only if MF = 0.
Proof. If M is F −negligible then clearly MF = 0 because the inductive limit of the zero system is zero. Conversely, if MF = 0 then
M = Ker(jM ) = F (M).
Now we prove a very useful result:
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
7
Proposition 2.10. Let f : M → N and g : M → P be R−linear
maps such that Ker(g) and Coker(g) are F −negligible. Then there is
a unique R−linear map h : P → NF which making commutative the
following diagram
M
f
/
N
g
jN
h
P
/
NF .
Proof. For each x ∈ P let J = AnnR x + Im(g) which belongs to F since Coker(g) is F −negligible. Then consider the map
hx : J → N/F (N) which maps each a ∈ J into f (m) + F (N) where
m is an element of M such that g(m) = ax. The map hx is welldefined since Ker(g) is F−negligible. It is also R−linear. Therefore
hx ∈ HomR J, N/F (N) . We define h : P → NF as x
[hx ].
Clearly the map h is R−linear and the completed diagram is commutative. Suppose ψ : P → NF is another R−linear map which making
commutative the foregoing diagram. Take x ∈ P , and let ψ(x) = [h′ ]
where h′ : I → N/F (N) is a R−linear map for some I ∈ F . It suf
fices to show that h′ |I∩J = (hx )|I∩J where J = AnnR x + Im(g) .
For each b ∈
I ∩ J there
is some m ∈ M such that g(m) = bx.
But ψ g(m) = jN f (m) . Therefore there is an ideal L ∈ F contained in I such that for each c ∈ L, ch′ (b) = cf (m) + F (N). Let
h′ (b) = n+ F (N) then we observe that L ⊆ AnnR f (m) −n+ F (N) .
Thus f (m) − n + F (N) ∈ F N/F (N) = 0.
Corollary 2.11. For each R−linear map u : M → N then there is a
unique R−linear map uF : MF → NF such that the following diagram
is commutative
u
/N
M
jN
jM
MF
uF
/
NF .
In particular, (jM )F = jMF .
Proof. By Lemma 2.8, Ker(jM ) and Coker(jM ) are F −negligible
therefore the first part of the assertion is an immediate consequence
of Proposition 2.10. Take N = MF and u = jM then the second part
8
ABOLFAZL TARIZADEH
implies.
For every R−linear map u : M → N and for each [f ] ∈ MF , we have
uF ([f ]) = [u ◦ f ] where u : M/F (M) → N/F (N) is induced by u.
Lemma 2.12. If 0
/
M′
u
/
v
M
R−modules then the sequence 0
act.
/
MF′
/
M ′′ is an exact sequence of
uF
/
MF
vF
/
MF′′ is ex-
Proof. Let [f ] ∈ Ker(vF ) where f : I → M/F (M) is a R−linear
map for some I ∈ F . Thus there is an ideal J ∈ F contained in I
−1
Im(u) ∈ F . Because
such that v ◦ f |J = 0. We have
L = J ∩f
for each b ∈ J, AnnR v(m) ∈ F where f (b) = m + F (M). But
AnnR v(m) ⊆ AnnR b + L and so J/L is F −negligible. Hence
′
′
L ∈ F .F = F . Now consider the map g : L → M
/F (M ) given
′
′
′
′
by b
m + F (M ) where f (b) = u m + F (M ) . The map g is
well-defined since u is injective. It is also R−linear. Thus [g] ∈ MF′ .
Clearly f |L = u ◦ g and so [f ] ∈ Im(uF ).
Lemma 2.13. Let M be a R−module. Then the canonical map (jM )F =
jMF : MF → (MF )F is bijective.
Proof. By Corollary 2.11, (jM )F = jMF . By Lemma 2.12, from
the exact sequence 0
/
F (M)
i
/
M
jM
iF
/
MF we obtain the fol(jM )F
/ (F (M))F
/ MF
/ (MF )F . By
lowing exact sequence 0
Lemma 2.9, F (M) F = 0 hence jMF is injective thus F (MF ) =
0. Therefore (MF )F = colimI∈F HomR (I, MF ). Take [f ] ∈ (MF )F
where f : I → MF is a R−linear map and I ∈ F . We claim that
L = f −1 Im jM ∈ F . For each a ∈ I there is an ideal J ∈ F and
also there is a R−linear map h : J → M/F (M) such that f (a) = [h].
To prove the claim, first we show that J ⊆ AnnR (a + L). Let b ∈ J and
let h(b) = m + F (M) where m ∈ M. We have f (ab) = bf (a) = [b.h].
But b.h = (δm )|J therefore f (ab) = [δm ] ∈ Im jM . This implies that
I/L is F −negligible and so L ∈ F .F = F . This establishes the
claim. Now consider the map g : L → M/F (M) which maps each
a ∈ L into m + F (M) where f (a) = jM (m). The map g is well-defined
since Ker(jM ) = F (M). It is also R−linear. Hence [g] ∈ MF . Clearly
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
9
f |L = jM ◦ g thus [f ] = jMF ([g]) and so jMF is surjective.
Now we are ready to prove the first main result of this article:
Theorem 2.14. Let ϕ : M → N be a R−linear map such that Ker ϕ
and Coker ϕ are F −negligible. Then ϕF : MF → NF is bijective.
Proof. By Proposition 2.10, there is a (unique) R−linear map
h : N → MF such that h ◦ ϕ = jM . By Lemma 2.13, jMF is bi−1
jective. We shall prove that jM
◦ hF is the inverse of ϕF . We have
F
−1
−1
jM
◦
h
◦
ϕ
=
j
◦
(h
◦
ϕ)
F
F
F = IdMF . To conclude the proof, by
MF
F
−1
Corollary 2.11, it suffices to show that ϕF ◦ jM
◦ hF ◦ jN = jN . But
F
−1
−1
ϕF ◦ jMF ◦ hF ◦ jN = ϕF ◦ jMF ◦ jMF ◦ h = ϕF ◦ h. For each x ∈ N,
from the proof of Proposition 2.10, we know that h(x) = [hx ] where
hx : J = AnnR (x + Im ϕ) → M/F (M) is a R−linear map which maps
each a ∈ J into m + F (M) such that ϕ(m) = ax. Thus ϕ ◦ hx = (δx )|J
where x = x + F (N). This means that ϕF ◦ h = jN .
Now the G-localization rings are introduced. Let M be a R−module.
For each [f ] ∈ RF and for each [g] ∈ MF where f : I → R/F (R)
and g : J → M/F (M) are R−linear maps and I, J ∈ F , we have
f −1 (J) ∈ F where J = J + F (R)/F (R). Moreover g induces a
map J → R/F (R) given by b + F (R)
g(b) which we denote it
by g. The map g is clearly well-defined. Now
we define the pairing
RF × MF → MF as [f ].[g] = [g ◦ f |f −1 (J) ]. It is easy to see that it
is R−bilinear (details omitted).
In particular, the binary operation RF × RF → RF , as multiplicative, puts a commutative ring structure on the R−module RF . Its
commutativity implies from the fact that for every R−linear maps
f, g : I → R/F (R) where I is an ideal
of R and for every elements a, b ∈ I then g f (ab) = f g(ab) . The unit element of the
ring RF is [π] where π : R → R/F (R) is the canonical map. The
map jR : R → RF is a ring homomorphism. Moreover, the pairing
RF × MF → MF puts a RF −module structure on MF . For every
R−linear map u : M → N then uF : MF → NF is RF −linear. In fact
M
MF is a left exact functor from the category of R−modules into
the category of RF −modules, see Corollary 2.11 and Lemma 2.12. It
is called the Gabriel localization (G-localization) functor with respect
10
ABOLFAZL TARIZADEH
to the system F .
Proposition 2.15. Let J be an ideal of R such that JRF = RF . Then
J ∈ F.
Proof. We may write 1 =
n
P
zi jR (ai ) where ai ∈ J and zi ∈ RF for
i=1
all i. Let f : I → N/F (N) be a R−linear map where N = R/J and
n
P
I ∈ F . Then we have [f ] =
zi .[ai .f ] = 0 since a.f = 0 for all a ∈ J.
i=1
Therefore R/J is F −negligible and so J ∈ F .F = F .
Note that the converse of Proposition 2.15 does not necessarily hold.
In fact, the condition “IRF = RF for all I ∈ F ”, as we shall observe
in the article, is a crucial point of the G-localization rings. Many interesting and major facts are equivalent or imply from this condition,
see for example, Corollary 3.7 and Theorem 4.4.
The second main result of this article is the following.
Theorem 2.16. Let F be an idempotent topologizing system on the
ring R and let ϕ : R → S be a ring map such that IS = S for all
I ∈ F . Then the following conditions hold.
(i) For each S−module M, the canonical map jM : M → MF is bijective.
(ii) The map ψ = jS−1 ◦ ϕF is the only ring homomorphism from RF
into S such that ϕ = ψ ◦ jR .
Proof. (i) : If m
∈ F . Thus
P ∈′ F (M) then I = AnnR (m)
we may write 1 =
si ϕ(ai ) where ai ∈ I and s′i ∈ S. Therei
P ′
fore m =
si ϕ(ai )m = 0 hence jS is injective. Let f : J → M
i
be a R−linear map where J ∈ F . We may write 1 =
n
P
sj ϕ(bj )
j=1
where bj ∈ J and sj ∈ S for all j. For each c ∈ J we have f (c) =
n
n
n
P
P
P
sj ϕ(bj )f (c) =
sj f (bj c) =
sj f (bj ) ϕ(c). Therefore (δm )|J =
j=1
f where m =
j=1
n
P
j=1
sj f (bj ) and δm : R → M which maps each r ∈ R into
j=1
r.m = ϕ(r)m. This means that jM (m) = [f ] and so jM is surjective.
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
11
(ii) : First we show that the map ψ = jS−1 ◦ ϕF is actually a ring
homomorphism. Clearly it is additive and transforms the unit element of RF to the unit of S. To prove that it is multiplicative take
two elements z, z ′ ∈ RF with representations f and g, i.e., z = [f ]
and z ′ = [g]. By passing to the restrictions, if it is necessary, therefore we may assume that the R−linear maps f and g have the same
domain. Namely f, g : J → R/F (R) where J ∈ F . We know
that f −1 (J) P
∈ F where J = J + F (R)/F (R). Therefore we may
write 1 =
ϕ(bj )sj where bj ∈ f −1 (J ) ⊆ J and sj ∈ S. We
j
P
have then ψ(z.z ′ ) =
ϕ(c′j )sj where for each j, f (bj ) = cj + F (R),
j
P
g(cj ) = c′j + F (R) and cj ∈ J. Similarly ψ(z) = ϕ(cj )sj and ψ(z ′ ) =
j
P
ϕ(ei )si where for each i, g(bi ) = ei + F (R). Thus ψ(z)ψ(z ′ ) =
i
P
P
P
ϕ(ei cj )si sj . But for each j we have ϕ(c′j ) =
ϕ(c′j bi )si =
j
i
i P
P
P
P
ϕ c′j bi + F (R) si = ϕ bi g(cj ) si =
ϕ cj g(bi ) si = ϕ ei cj +
i
i
i
i
P
F (R) si =
ϕ(ei cj )si . Therefore ψ(z.z ′ ) = ψ(z)ψ(z ′ ). Finally, supi
pose ψ ′ : RF → S is another ring map such that ψ ′ ◦ jR = ϕ. Clearly
jS ◦ ψ ′ is R−linear and (jS ◦ ψ ′ ) ◦ jR = jS ◦ ϕ. Therefore, by Corollary
2.11, jS ◦ ψ ′ = ϕF and so ψ ′ = jS−1 ◦ ϕF .
3. Flat epimorphisms as G-localizations
By an epimorphism ϕ : R → S we mean it is an epimorphism in the
category of commutative rings. We should mention that the surjective
ring maps are just special cases of the epimorphisms. For example, the
canonical ring map Z → Q is an epimorphism while it is not surjective.
A ring map which is both flat and an epimorphism is called a flat epimorphism. The canonical map R → S −1 R where S is a multiplicative
subset of a ring R is a typical example of flat epimorphisms.
The following lemma is a well-known result but we have provided a
proof for the sake of completeness.
Lemma 3.1. Let ϕ : R → S be a ring homomorphism. Then the following conditions are equivalent.
(i) ϕ is an epimorphism.
(ii) For each s ∈ S, s ⊗ 1 = 1 ⊗ s.
12
ABOLFAZL TARIZADEH
(iii) The map p : S ⊗R S → S defined by s ⊗ s′
ss′ is bijective.
(iv) The map j : S → S ⊗R S defined by s
1 ⊗ s is bijective.
Proof. (i) ⇒ (ii) : We have i ◦ ϕ = j ◦ ϕ where i, j : S → S ⊗R S
are the canonical ring maps which map each s ∈ S into s ⊗ 1 and 1 ⊗ s,
respectively.
(ii) ⇒ (iii) : We have Ker(p) = hs ⊗ 1 − 1 ⊗ s : s ∈ Si because if ss′ = 0
then we may write s ⊗ s′ = 1 ⊗ s′ (s ⊗ 1 − 1 ⊗ s).
(iii) ⇒ (i) : Let f, g : S → T be ring maps such that f ◦ ϕ = g ◦ ϕ. By
the universal property of the pushouts, there is a (unique) ring map
ψ : S ⊗R S → T such that f = ψ ◦ i and g = ψ ◦ j. But i = j since
Ker(p) = 0. Thus f = g.
(ii) ⇒ (iv) : We have s ⊗ s′ = (s ⊗ 1)(1 ⊗ s′ ) = 1 ⊗ ss′ = j(ss′ ).
(iv) ⇒ (iii) : We have p ◦ j = Id. Therefore p is bijective.
The third main result of this article is the following.
Theorem 3.2. Let ϕ : R → S be a ring map and let F be the set of
ideals of R whose extensions under ϕ are equal to S. Then the following conditions hold.
(i) The family F is an idempotent toplogizing system on the ring R.
(ii) If ϕ is a flat epimorphism then for each R−module M, the map
ηF : MF → (S ⊗R M)F induced by the canonical map η : M → S ⊗R M
given by m
1 ⊗ m is an isomorphism. In particular the map ϕF is
bijective.
Proof. (i) : Clearly F is non-empty since R ∈ F . Suppose I ∈ F
and J is an ideal of R such that J : a ∈ F for all a ∈ I. We may
n
P
sj ϕ(aj ) where aj ∈ I and sj ∈ S. For each j, we may also
write 1 =
write 1 =
then 1 =
j=1
kj
P
ij =1
k
1
P
sij ,j ϕ(bij ,j ) where bij ,j ∈ J : aj and sij ,j ∈ S. We have
...
i1 =1
n
kn P
P
in =1 j=1
si1 ,1 ...sin ,n sj ϕ(bi1 ,1 ...bin ,n aj ) and all the elements
bi1 ,1 ...bin ,n aj belong to J and so J ∈ F . Therefore, by Lemma 2.6, F
is an idempotent topologizing system.
(ii) : To prove the assertion, by Theorem 2.14, it suffices to show
that Ker η and Coker η are F −negligible. Let m ∈ Ker η then the
map λ : S ⊗R N → S ⊗R M induced by the canonical injection
N = Rm → M is injective since S is R−flat. But Im λ = 0 because
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
13
1 ⊗ m = 0. This implies that AnnR (m)S = S and so AnnR (m) ∈ F .
Thus Ker η is F −negligible.
By applying the right exact functor S ⊗R − to the exact sequence
M
η
/
S ⊗R M
/
Coker η
exact sequence S ⊗R M
1⊗η
/
/
0 we then obtain the following
S ⊗R (S ⊗R M)
The map 1 ⊗ η factors as S ⊗R M
j⊗1
/
/
S ⊗R Coker η
(S ⊗R S) ⊗R M
≃
/
/
S ⊗R (S ⊗R M). By Lemma 3.1, j ⊗ 1 is bijective, hence so is 1 ⊗ η.
In particular it is surjective and so S ⊗R Coker η = 0. For each
x ∈ Coker η, the map S ⊗R Rx → S ⊗R Coker η induced by the canonical injection Rx → Coker η is injective since S is R−flat. This implies
that AnnR (x)S = S and so AnnR (x) ∈ F . Therefore Coker η is also
F −negligible. Finally, the map ϕ factors as R
hence ϕF is bijective.
η
/
R ⊗R S
≃
/
S
Corollary 3.3. Let ϕ : R → S be a ring map and let F be the set of
ideals I of R such that IS = S. If ϕ is a flat epimorphism then there
is a unique isomorphism of rings ψ : RF → S such that ϕ = ψ ◦ jR .
Proof. By Theorem 2.16, the map ψ = jS−1 ◦ ϕF is the only ring
homomorphism such that ϕ = ψ ◦ jR . By Theorem 3.2, ϕF is bijective,
therefore ψ is an isomorphism.
Corollary 3.4. Let J be an ideal of a ring R. Then R/J is R−flat if
and only if AnnR (a) + J = R for all a ∈ J.
Proof. Suppose R/J is R−flat. Let F be the set of ideals I of
R such that I + J = R. By Corollary 3.3, there is a (unique) isomorphism of rings ψ : R/J → RF such that jR = ψ ◦ π where
π : R → R/J is the canonical map. We have F (R) = Ker(jR ) =
jR−1 (0) = π −1 ψ −1 (0) = Ker(π) = J. Conversely, let f : M → N be an
injective R−linear map. To prove the assertion it suffices to show that
the induced map M/JM → N/JN given by m + JM
f (m) + JN
s
P
ai ni where
is injective. If f (m) ∈ JN then we may write f (m) =
i=1
ai ∈ J and ni ∈ N for all i. By the hypothesis, there are elements
bi ∈ AnnR (ai ) and ci ∈ J such that 1 = bi + ci . It follows that
1 = (b1 + c1 )(b2 + c2 )...(bs + cs ) = b + c where b = b1 b2 ...bs and c ∈ J.
0.
14
ABOLFAZL TARIZADEH
Thus f (m) = bf (m) + cf (m) = f (cm). Therefore m = cm ∈ JM.
Corollary 3.4, in particular, tells us that if the ideal J has a generating set S such that each a ∈ S can be written as a = a2 b for some
b ∈ R then R/J is R−flat. As another application, if J is an ideal of a
domain R such that R/J is R−flat then we have either J = 0 or J = R.
Lemma 3.5. Let F be an idempotent topologizing system on the ring
R such that IRF = RF for all I ∈ F . Then jR is an epimorphism.
Proof. Let f, g : RF → S be two ring maps such that f ◦jR = g ◦jR .
If we consider the ring map ϕ = f ◦ jR : R → S then IS = S for
all I ∈ F . Therefore, by Theorem 2.16, there is a unique ring map
ψ : RF → S such that ϕ = ψ ◦ jR . Thus f = g.
The converse of Corollary 3.3 also holds even under a mild hypothesis which is another main result of this article:
Theorem 3.6. Let F be an idempotent topologizing system on the ring
R and let ϕ : R → S be a ring map such that IS = S for all I ∈ F .
If there is an isomorphism of rings ψ : RF → S such that ϕ = ψ ◦ jR
then ϕ : R → S is a flat epimorphism and ψ = jS−1 ◦ ϕF .
Proof. We have IRF = RF for all I ∈ F . Therefore, by Lemma
3.5, jR and so ϕ are epimorphisms. By Theorem 2.16, ψ = jS−1 ◦ ϕF .
Therefore ϕF is bijective. This, in particular, implies that jS (s) ∈
Im ϕF for all s ∈ S. Thus there is an ideal I ∈ F such that I ⊆
AnnR (s+Im ϕ). Therefore Coker ϕ and so any finite direct sum of it are
F −negligible. To prove that S is R−flat, by [4, Theorem 7.7], it suffices to show that for each ideal J of R then the canonical map J ⊗R S →
n
n
n
P
P
P
ϕ(ai )si = 0
ϕ(ai )si is injective. Suppose
ai ⊗ si
S given by
i=1
i=1
i=1
where ai ∈ J and si ∈ S for all i. We have AnnR (x) ∈ F where
x = (si +Im ϕ)ni=1 ∈ (Coker ϕ)n since (Coker ϕ)n is F −negligible. Thus
there are elements b1 , ..., bm ∈ AnnR (x) and also elements s′1 , ..., s′m ∈ S
m
P
such that 1 =
ϕ(bj )s′j . Moreover there are elements ri,j ∈ R such
j=1
that si ϕ(bj ) = ϕ(ri,j ) for all i, j. Thus cj =
n
P
i=1
ai ri,j ∈ Ker ϕ = F (R)
′
′
for all j. Hence for each j there are elements rj,1
, ..., rj,N
∈ AnnR (cj )
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
and also elements s′′j,1 , ..., s′′j,N ∈ S such that 1 =
we have
n
P
ai ⊗ si =
i=1
i=1
cj ⊗ s′j = cj ⊗
n
P
N
P
k=1
ai ⊗
m
P
j=1
si ϕ(bj )s′j
′
ϕ(rj,k
)s′j s′′j,k =
N
P
k=1
=
m
P
j=1
N
P
k=1
15
′
ϕ(rj,k
)s′′j,k . Now
cj ⊗ s′j . For each j,
′
rj,k
cj ⊗ s′j s′′j,k = 0.
Corollary 3.7. Let F be an idempotent topologizing system on the
ring R such that IRF = RF for all I ∈ F . Then jR is a flat epimorphism.
Proof. It is an immediate consequence of Theorem 3.6.
It is natural to ask whether the converse of Corollary 3.7 holds.
4. G-localizations of finite type systems
Definition 4.1. An R−module M is said to be F −closed if the canonical map jM is bijective. It is called strongly F −closed if the canonical
map M → HomR (I, M) given by m
(δm )|I is bijective for all I ∈ F .
For each R−module M then MF , by Lemma 2.13, is F −closed.
Clearly each strongly F −closed module is F −closed. By the category
of F −closed modules we mean a full subcategory of the category of
R−modules whose objects are the F −closed modules.
Lemma 4.2. Let F be an idempotent topologizing system on the ring
R such that IRF = RF for all I ∈ F . Then every F −closed module
is strongly F −closed and the G-localization functor with respect to F
is an equivalence between the category of F −closed modules and the
category of RF −modules.
Proof. Let N be a F −closed module and let I ∈ F . Then the
canonical map N → HomR (I, N) is injective since F (N) = 0. Let
n
P
f : I → N be a R−linear map. We may write 1 =
jR (ai )zi
i=1
where ai ∈ I and zi ∈ RF for all i. There exists an element x ∈ N
n
P
such that jN (x) =
zi jN f (ai ) . Now, for each a ∈ I, we have
i=1
16
ABOLFAZL TARIZADEH
jN (ax) = jN f (a) and so ax = f (a). Thus (δx )|I = f . Therefore
N is strongly F −closed. Moreover, for each R−module M then the
map HomR (M, N) → HomRF (MF , NF ) given by u
uF is bijective.
′
′
Because suppose uF = uF where u, u : M → N are R−linear maps.
−1
We have u = jN
◦ uF ◦ jM = u′. Let ϕ : MF → NF be a R−linear
−1
map. Then, by Corollary 2.11, uF = ϕ where u = jN
◦ ϕ ◦ jM . Finally, we show that the G-localization functor is essentially surjective.
Let L be a RF −module. By Theorem 2.16, jL is bijective. Thus L
as R−module is F −closed. It remains to show that jL is RF −linear.
We have jL = σL ◦ η where η : L → RF ⊗R L is the canonical map.
The map σL is already RF −linear. By Lemma 3.5, jR is an epimorphism. Note that for any epimorphism of rings ϕ : R → S and for any
S−modules M and N then the two S−module structures on M ⊗R N
defined on pure tensors by s.(m⊗n) = sm⊗n and s∗(m⊗n) = m⊗sn
are the same since s ⊗ 1 = 1 ⊗ s for all s ∈ S and so the canonical map
η : N → S ⊗R N is S−linear. Therefore jL is RF −linear.
Definition 4.3. An idempotent topologizing system is called of finite
type if every element of it containing a finitely generated ideal belonging to the system. For example, if ϕ : R → S is a ring map then the
system {I ⊆ R : IS = S} is of finite type.
In the next result we shall use, for each R−module M, the canonical
RF −linear map σM : RF ⊗R M → MF which maps each pure tensor
a ⊗ m into a.jM (m) for all a ∈ RF and all m ∈ M. Note that σR is
bijective and for each R−linear map ϕ : M → N then the following
diagram is commutative
RF ⊗R M
1⊗ϕ
/
RF ⊗R N
σM
MF
σN
ϕF
/
NF .
The following is another main result of this article:
Theorem 4.4. Let F be an idempotent topologizing system on the ring
R. Then the following conditions are equivalent.
(i) For each R−module M, the canonical map σM : RF ⊗R M → MF
is bijective.
(ii) For every R−module M, F (M) is the kernel of the canonical map
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
17
η : M → RF ⊗R M.
(iii) For each I ∈ F , IRF = RF .
(iv) The G-localization functor with respect to F is exact and preserves direct sums.
(v) The G-localization functor with respect to F is exact and the system F is of finite type.
(vi) The G-localization functor with respect to F is essentially surjective.
(vii) For each RF −module N, F (N) = 0.
Proof. (i) ⇒ (ii): We have jM = σM ◦ η. Thus F (M) = Ker jM =
−1
−1
jM
(0) = η −1 σM
(0) = Ker η.
(ii) ⇒ (iii) : If I ∈ F then R/I is F −negligible. Thus Ker η = R/I
where η : R/I → RF ⊗R R/I is the canonical map. Therefore Im η = 0.
This implies
that RF ⊗R R/I = 0 since z ⊗ (r + I) = z ⊗ (1 + I) 1 ⊗
(r + I) = 0. It follows that IRF = RF .
(iii) ⇒ (i) : By Corollary 3.7, jR is a flat epimorphism. Moreover, by
Proposition 2.15, F = {I ⊆ R : IRF = RF }. Therefore, by the proof
of Theorem 3.2, Ker η and Coker η are F −negligible where η : M →
RF ⊗R M is the canonical map. Thus, by Proposition 2.10, there is
a unique R−linear map h : RF ⊗R M → MF such that jM = h ◦ η.
By Theorem 2.16, jN is bijective where N = RF ⊗R M. Moreover,
−1
by Theorem 3.2, ηF is bijective. We also have jM = (ηF
◦ jN ) ◦ η.
−1
Therefore σM = ηF ◦ jN and so σM is bijective.
(i) ⇒ (iv) and (v) : Easy.
(iv) ⇒ (i) : Let L be a free R−module and let B =L
{xi : i ∈ I} be a
basis of L. Then there is a bijective map ψ : L →
R which maps
i∈I
P
each x ∈ L into (ri )i∈I where x =
ri xi . We denote the converse of
i∈I L
ψ by θ and for a given R−module M,
M is denoted by M ⊕I . We
i∈I
claim that the map σL factors as
RF ⊗R L
≃
/
(RF ⊗R R)⊕I
(σR )I
/
(RF )⊕I
τ (I)
/
(R⊕I )F
θF
/
LF
P
where τ (I) (zi )i∈I =
(τi )F (zi ) and for each i, τi : R → R⊕I =
i∈I
L
R is the canonical map which is defined as r
(rδi,j )j∈I . In
j∈I
fact, for each pure tensor z ⊗ x of RF ⊗RPL we have z ⊗ x
(z ⊗
ri )i∈I
(ri .z)i∈I
ξ
θF (ξ) where ξ = (τi )F (ri .z). But θF (ξ) =
i∈I
18
ABOLFAZL TARIZADEH
P
P
P
zθF
(τi )F ◦ jR (ri ) = z θF ◦ j(R⊕I ) ◦ τi (ri ) = z jL ◦ θ ◦ τi (ri ) =
i∈I
i∈I
i∈IP
zjL θ
τi (ri ) = zjL (x) = σL (z ⊗ x). This establishes the claim.
i∈I
The map τ (I) is bijective since the G-localization functor preserves direct sums. Therefore σL is bijective. Now, by applying the tensor
functor RF ⊗R − and the G-localization functor which is, by the hy/L
/M
/0
potheses, right exact to the exact sequence L′
′
where L and L are free R−modules, and also by using a special case
of the five lemma we conclude that σM is bijective:
RF ⊗R L′
/
RF ⊗R L
σL′
/
RF ⊗R M
σL
/
0
σM
L′F
/
LF
/
MF
/
0.
(v) ⇒ (iii) : The G-localization functor, by Corollary 2.11, is additive.
In every abelian category, each split and exact sequence is left split
and exact by an additive functor. This, in particular, implies that
the G-localization functor preserves finite direct sums. Specially, τ (n) :
(RF )n → (Rn )F is an isomorphism as RF −modules for all natural
numbers n. Therefore all of the factors in the decomposition of σRn are
bijective. Let I be a f.g. ideal of R. Now, by applying the tensor functor
RF ⊗R − and the G-localization functor which is, by the hypotheses,
/R
/ R/I
/ 0 and
right exact to the exact sequence Rn
also by using a special case of the five lemma we obtain that σR/I is
bijective:
RF ⊗R Rn
σRn
/
RF ⊗R R
/
RF ⊗R R/I
(Rn )F
/
0
σR/I
σR
/
RF
/
(R/I)F
/
0.
If moreover I ∈ F then 0 = (R/I)F ≃ RF ⊗R R/I ≃ RF /IRF .
Therefore IRF = RF . In fact the latter holds for each element of the
system F since it is of finite type.
(iii) ⇒ (vi) : See Lemma 4.2.
(vi) ⇒ (vii) : Let N be a RF −module. By the hypothesis, there exists
a R−module M such that N ≃ MF . Therefore F (N) ≃ F (MF ) = 0.
(vii)
⇒ (iii) : If I ∈ F then RF ⊗R R/I = 0 since I ⊆ AnnR z ⊗ (r +
I) thus z ⊗ (r + I) ∈ F (RF ⊗R R/I) = 0. Therefore IRF = RF for
all I ∈ F .
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
19
Proposition 4.5. Let ϕ : R → S be a ring map, let F be an of finite
type system on the ring R and let G be the set of ideals J of S such
that S/J as R−module is F −negligible. Then the following conditions
hold.
(i) An ideal J of S belongs to G if and only if IS ⊆ J for some I ∈ F .
(ii) The family G is an idempotent topologizing system on the ring S
which is also of finite type.
(iii) For each S−module M, F (M) = G (M) and there is a canonical
′
isomorphism of R−modules η : MG → MF such that jM = η ◦jM
where
′
jM : M → MG is the canonical map.
(iv) The map ϕF : RF → SF is a ring homomorphism when we identify SF with SG via η.
Proof. (i) : Suppose J is an ideal of S and IS ⊆ J for some
I ∈ F . Let s ∈ S we have then I ⊆ AnnR (s + J). Therefore S/J is
F −negligible. Conversely, suppose J ∈ G . Then I = AnnR (1S + J) ∈
F and clearly IS ⊆ J.
(ii) : Clearly G is a topologizing system on the ring S, it is also of finite type once we have verified that it is idempotent. Let J ∈ G .G
then there exists an ideal J ′ ∈ G containing J such that J ′ /J is
G −negligible. By the hypothesis, there is a finitely generated ideal
I = ha1 , ..., an i of R such that I ∈ F and IS ⊆ J ′ . For each j, there
is also an ideal Ij ∈ F such that (Ij aj )S ⊆ J. Let I ′ = I1 I2 ...In then
clearly II ′ ∈ F and II ′ S ⊆ J. This means that J ∈ G .
(iii) : If m ∈ F (M) then I = AnnR (m) ∈ F and IS ⊆ AnnS (m)
therefore m ∈ G (M). Conversely, if m ∈ G (M) then there exists an
ideal I ∈ F such that IS ⊆ AnnS (m). Thus I ⊆ AnnR (m) and so
m ∈ F (M). Therefore F (M) = G (M).
Let f : J → M/G (M) be a S−linear map where J ∈ G ; there exists
some I ∈ F such that IS ⊆ J. Consider the map η : MG → MF given
by [f ]
[f ◦ ϕ|I ]. Note that this definition is independent of choosing
such I. We shall prove that η is bijective. Clearly it is injective. Let
f : I → M/F (M) be an R−linear map for some I ∈ F . Then conn
P
sider the map f ∗ : IS → M/G (M) which maps each
sj ϕ(aj ) into
j=1
n
P
j=1
m
P
i=1
sj f (aj ). The map f ∗ is well-defined. Because, suppose
n
P
sj ϕ(aj ) =
j=1
s′i ϕ(bi ) where aj , bi ∈ I and sj , s′i ∈ S. Let f (aj ) = mj + F (M)
and f (bi ) = m′i + F (M) for all i and j. Set m =
n
P
j=1
sj mj and
20
ABOLFAZL TARIZADEH
m′ =
m
P
i=1
s′i m′i . We have IS ⊆ AnnS m − m′ + G (M) . Because for
m
n
P
P
sj ϕ(c)f (aj ) − s′i ϕ(c)f (bi ) =
each c ∈ I, ϕ(c) m − m′ + G (M) =
i=1
j=1
n
P
m
P
m
P
sj ϕ(aj ) −
s′i ϕ(bi ) f (c) = 0. Therej=1
i=1
j=1
i=1
fore m − m′ + G (M) ∈ G M/G (M) = 0. Thus f ∗ is well-defined.
Clearly it is S−linear and η([f ∗ ]) = [f ] and so η is surjective.
(iv): One can easily verify that η −1 ◦ ϕF : RF → SG is actually a ring
homomorphism.
sj f (aj c) −
s′i f (bi c) =
n
P
Lemma 4.6. Let F be an idempotent topologizing system on the ring
R such that the zero ideal of R does not belong to F . If R is an integral
domain, then so is RF .
Proof. Clearly RF is a non-trivial ring if and only if 0 ∈
/ F . We
′
also have F (R) = 0. Let z, z ∈ RF with representations f and g,
i.e., z = [f ] and z ′ = [g] such that z.z ′ = 0. Suppose z, z ′ 6= 0.
We may assume that the R−linear maps f and g have the same domain I ∈ F , i.e., f, g : I → R. There exists an ideal J ∈ F such
that J ⊆ f −1 (I) and g f (a) = 0 for all a ∈ J. There are also ele′
ments b, c ∈ J such that
f (b), g(c) 6= 0 since z, z 6= 0. But we have
0 = g f (bc) = g cf (b) = f (b)g(c) which is a contradiction. Therefore we have either z = 0 or that z ′ = 0.
Remark 4.7. If M ′ is a R−submodule of M then iF : MF′ → MF is
injective where i : M ′ → M is the canonical injection. By abuse of the
notation, the image of iF is also denoted by MF′ .
Remark 4.8. For each ideal I of R we have IRF ⊆ IF . Moreover,
IF = RF for all I ∈ F because R/I is F −negligible then apply
Lemma 2.12. As a second proof, we observe that π|I = i ◦ f where
π : R → R/F (R) is the canonical map, f : I → I/F (I) which maps
each a ∈ I into a + F (I) and i : I → R is the canonical injection.
Thus 1RF = [π] = [π|I ] ∈ IF .
Lemma 4.9. Let M be a R−module and let N be a R−submodule
of
−1
−1
MF . Then jM (N) F = jMF (NF ). If moreover F MF /N = 0, then
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
21
−1
N = jM
(NF ).
F
−1
−1
Proof. Let N ′ = jM
(N). First we show that NF′ ⊆ jM
(NF ). Let
F
′
′
′
x = [i ◦ f ] ∈ NF where f : I → N /F (N ) is an R−linear map and
I ∈ F . Consider the map g : I → N which maps each a ∈ I into
jM (z) where f (a) = z + F (N ′ ). Note that the map g is well-defined
because if there exists another z ′ ∈ N ′ such that f (a) = z ′ + F (N ′ )
then jM (z − z ′ ) ∈ F (N) ⊆ F (MF ) = 0. Clearly it is also R−linear.
Therefore [g] ∈ NF . We have (δx )|I = i ◦ g where i : N → MF is
the canonical injection. This implies that jMF (x) ∈ NF . To prove the
−1
converse inclusion we act as follows. Let x ∈ jM
(NF ). Thus there
F
exists an element y ∈ NF such that jMF (x) = iF (y). We claim that
Coker(ψ) = N/jM (N ′ ) is F −negligible where ψ = (jM )|N ′ : N′ → N.
For each x = [f ] ∈ N, we show that I ⊆ AnnR x + jM (N ′ ) where
f : I → M/F (M) is an R−linear map and I ∈ F . For each a ∈ I
there is some m ∈ M such that f (a) = m = m + F (M). Clearly
(δm )|I = a.f thus jM (m) = a.x ∈ N and so m ∈ N ′ . This establishes
the claim. We also have Ker(ψ) = F (N ′ ). Therefore, by Theorem 2.14,
ψF is bijective. Thus there is an element z ∈ NF′ such that ψF (z) = y.
Therefore x = i′F (z) ∈ Im i′F = NF′ because jMF ◦ i′F = iF ◦ ψF where
i′ : N ′ → M is the canonical injection. Now we prove the last part of
the assertion. For each x ∈ N, δx = i ◦ g where the map g : R → N is
−1
−1
defined as r
r.x. Thus N ⊆ jM
(NF ). Conversely, let x ∈ jM
(NF )
F
F
then there exists an ideal
I
∈
F
such
that
I
⊆
Ann
(x
+
N).
ThereR
fore x + N ∈ F MF /N = 0.
Now we prove the final main result of this article:
Theorem 4.10. Let F be an of finite type system on the ring R and
let F ′ be the set of ideals J of RF such that RF /J as R−module is
F −negligible. Then the map p
pF is a bijection between the set of
prime ideals of R which do not belong to F and the set of prime ideals
of RF which do not belong to F ′ .
Proof. First of all we show that if p is a prime ideal of R such that
p∈
/ F then pF is a prime ideal of RF and that pF ∈
/ F ′ . Suppose
′
′
1RF = [π ] ∈ pF where π : R → R/F (R) is the canonical map. Thus
there exists an ideal I ∈ F such that for each a ∈ I there is some
xa ∈ p in which AnnR (a − xa ) ∈ F . But I is not contained in p since
p∈
/ F . Take b ∈ I \ p then clearly AnnR (b − xb ) ⊆ p. This implies that
22
ABOLFAZL TARIZADEH
p ∈ F which is a contradiction therefore pF is a proper ideal of RF .
We shall apply Proposition 4.5 to the canonical ring map π : R → R/p.
If 0 ∈ G then R/p will be F −negligible. By applying Lemma 2.12 to
/p
/ R π / R/p we get that pF = RF
the exact sequence 0
which is a contradiction since pF is a proper ideal of RF . Thus 0 ∈
/G
and so by Lemma 4.6, (R/p)G ≃ (R/p)F is an integral domain. Suppose xy ∈ pF for some elements x, y ∈ RF . Thus 0 = πF (xy) =
πF (x)πF (y). Therefore we have either πF (x) = 0 or that πF (y) = 0.
Hence pF is a prime ideal.
Suppose pF ∈ F ′ then there exists an ideal I ∈ F such that IRF ⊆
pF . This implies that I ⊆ p. Because for each a ∈ I there exists an
ideal J ∈ F such that for each b ∈ J there is some yb ∈ p for which
AnnR (ab − yb ) ∈ F . Take some b ∈ J \ p then AnnR (ab − yb) is not
contained in p. It follows that a ∈ p. But I ⊆ p is a contradiction since
p∈
/ F , therefore pF ∈
/ F ′.
The map p
pF is injective between the foregoing sets. Because suppose pF = p′F where p and p′ are prime ideals of R with p, p′ ∈
/ F . For
′
each a ∈ p, jp (a) ∈ pF thus there exists an ideal J ∈ F such that for
each b ∈ J there is some xb ∈ p′ in which AnnR (ab − xb ) ∈ F . Now
take some b ∈ J \ p′ also take some c ∈ AnnR (ab − xb ) \ p′ , then we
have c(ab − xb ) = 0 ∈ p′ and so a ∈ p′ ; symmetrically we have p′ ⊆ p,
thus p = p′ .
Finally, we show that the map p
pF is surjective. Let q be a prime
ideal of RF which does not belong to F ′ . We have p = jR−1 (q) ∈
/ F
since pRF ⊆ q. We also have F (RF /q) = 0 because pick x + q ∈
F (RF /q), if x ∈
/ q then AnnR (x + q)RF ⊆ q it follows that q ∈ F ′
which is a contradiction. Therefore, by Lemma 4.9, pF = q.
The usual localization theory is just a special case of the G-localizations:
Proposition 4.11. Let S be a multiplicative subset in the ring R
and let F be the set of ideals of R which meeting S. Then for each
R−module M, there is a unique R−linear map ϕ : S −1 M → MF such
that jM = ϕ ◦ π where π : M → S −1 M is the canonical map. Moreover
ϕ is bijective. In particular, the R−algebras S −1 R and RF are canonically isomorphic.
Proof. Clearly Ker(π) and Coker(π) are F −negligible. Therefore,
by Proposition 2.10, the first part of the assertion is realized. Now for
GABRIEL LOCALIZATION THEORY AND ITS APPLICATIONS
23
each I ∈ F , consider the map λI : HomR I, M/F (M) → S −1 M defined by f
m/s where f (s) = m+F (M) for some s ∈ I∩S. The map
λI is well-defined. Because suppose there is another element s′ ∈ I ∩ S
such that f (s′ ) = m′ + F (M). We have s′ f (s) = f (ss′ ) = sf (s′ ). It
implies that S ∩ AnnR (sm′ − s′ m) 6= ∅. The map λI is also R−linear
and clearly λI = λJ ◦ uI,J for every I, J ∈ F with I ≤ J. Therefore, by the universal property of the colimits, there is a (unique)
R−linear map λ : MF → S −1 M such that the map λI factors as
/ MF λ / S −1 M for all I ∈ F . The map
HomR I, M/F (M)
λ is injective because suppose λ([f ]) = 0 for some element [f ] ∈ MF
where f : I → M/F (M) is a R−linear map and I ∈ F . We know
that I ∩ S 6= ∅. Take some s ∈ I ∩ S then clearly Rs ∈ F and
f |Rs = 0. The map λ is also surjective. Because for each element
m/s ∈ S −1 M then consider the R−linear map g : Rs → M/F (M) defined by rs
rm + F (M). Note that the map g is well-defined since
′
if rs = r s then r − r ′ ∈ F (R) and so (r − r ′ )m ∈ F (R)M ⊆ F (M).
It is also R−linear and clearly λ([g]) = m/s. We have λ−1 = ϕ.
References
[1] T. Akiba, Remarks on generalized rings of quotients, III, J. Math. Kyoto Univ,
Vol. 9, Number 2 (1969), 205-212.
[2] N. Bourbaki, Algèbre commutative, chapitres 1 à 4, Springer 2006.
[3] P. Gabriel, Des catégories abeliennes. Bull. Soc. Math. France. 90 (1962), 323448.
[4] H. Matsumura, Commutative ring theory, Cambridge University Press, 1989.
[5] C. Nastasescu, N. Popescu, On the localization ring of ring, J.Algebra, 41-56
(1970).
[6] J. Olivier, Anneaux absolument plats universels et épimorphismes à buts
réduits, Séminaire Samuel. Algèbre commutative, tomme 2 (1967-1968).
[7] N. Popescu, T. Spircu, Quelques observations sur les épimorphismes plats (à
gauche) d’anneaux, J. Algebra, 40-59 (1970).
[8] B. Stenström, Rings of quotients: an introduction to the methods of ring
theory, Springer 1975.
Department of Mathematics, Faculty of Basic Sciences, University
of Maragheh, P. O. Box 55181-83111, Maragheh, Iran.
E-mail address: [email protected]
| 0math.AC
|
Label Distribution Learning Forests
Wei Shen1,2 , Kai Zhao1 , Yilu Guo1 , Alan Yuille2
Key Laboratory of Specialty Fiber Optics and Optical Access Networks,
Shanghai Institute for Advanced Communication and Data Science,
School of Communication and Information Engineering, Shanghai University
2
Department of Computer Science, Johns Hopkins University
arXiv:1702.06086v4 [cs.LG] 16 Oct 2017
1
{shenwei1231,zhaok1206,gyl.luan0,alan.l.yuille}@gmail.com
Abstract
Label distribution learning (LDL) is a general learning framework, which assigns
to an instance a distribution over a set of labels rather than a single label or multiple
labels. Current LDL methods have either restricted assumptions on the expression
form of the label distribution or limitations in representation learning, e.g., to
learn deep features in an end-to-end manner. This paper presents label distribution
learning forests (LDLFs) - a novel label distribution learning algorithm based on
differentiable decision trees, which have several advantages: 1) Decision trees
have the potential to model any general form of label distributions by a mixture
of leaf node predictions. 2) The learning of differentiable decision trees can be
combined with representation learning. We define a distribution-based loss function
for a forest, enabling all the trees to be learned jointly, and show that an update
function for leaf node predictions, which guarantees a strict decrease of the loss
function, can be derived by variational bounding. The effectiveness of the proposed
LDLFs is verified on several LDL tasks and a computer vision application, showing
significant improvements to the state-of-the-art LDL methods.
1
Introduction
Label distribution learning (LDL) [6, 11] is a learning framework to deal with problems of label
ambiguity. Unlike single-label learning (SLL) and multi-label learning (MLL) [26], which assume an
instance is assigned to a single label or multiple labels, LDL aims at learning the relative importance
of each label involved in the description of an instance, i.e., a distribution over the set of labels. Such
a learning strategy is suitable for many real-world problems, which have label ambiguity. An example
is facial age estimation [8]. Even humans cannot predict the precise age from a single facial image.
They may say that the person is probably in one age group and less likely to be in another. Hence it is
more natural to assign a distribution of age labels to each facial image (Fig. 1(a)) instead of using a
single age label. Another example is movie rating prediction [7]. Many famous movie review web
sites, such as Netflix, IMDb and Douban, provide a crowd opinion for each movie specified by the
distribution of ratings collected from their users (Fig. 1(b)). If a system could precisely predict such a
rating distribution for every movie before it is released, movie producers can reduce their investment
risk and the audience can better choose which movies to watch.
Many LDL methods assume the label distribution can be represented by a maximum entropy model [2]
and learn it by optimizing an energy function based on the model [8, 11, 28, 6]. But, the exponential
part of this model restricts the generality of the distribution form, e.g., it has difficulty in representing
mixture distributions. Some other LDL methods extend the existing learning algorithms, e.g, by
boosting and support vector regression, to deal with label distributions [7, 27], which avoid making
this assumption, but have limitations in representation learning, e.g., they do not learn deep features
in an end-to-end manner.
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
Figure 1: The real-world data which are suitable to be modeled by label distribution learning. (a)
Estimated facial ages (a unimodal distribution). (b) Rating distribution of crowd opinion on a movie
(a multimodal distribution).
In this paper, we present label distribution learning forests (LDLFs) - a novel label distribution
learning algorithm inspired by differentiable decision trees [20]. Extending differentiable decision
trees to deal with the LDL task has two advantages. One is that decision trees have the potential
to model any general form of label distributions by mixture of the leaf node predictions, which
avoid making strong assumption on the form of the label distributions. The second is that the split
node parameters in differentiable decision trees can be learned by back-propagation, which enables
a combination of tree learning and representation learning in an end-to-end manner. We define a
distribution-based loss function for a tree by the Kullback-Leibler divergence (K-L) between the
ground truth label distribution and the distribution predicted by the tree. By fixing split nodes, we
show that the optimization of leaf node predictions to minimize the loss function of the tree can
be addressed by variational bounding [19, 29], in which the original loss function to be minimized
gets iteratively replaced by a decreasing sequence of upper bounds. Following this optimization
strategy, we derive a discrete iterative function to update the leaf node predictions. To learn a forest,
we average the losses of all the individual trees to be the loss for the forest and allow the split nodes
from different trees to be connected to the same output unit of the feature learning function. In this
way, the split node parameters of all the individual trees can be learned jointly. Our LDLFs can be
used as a (shallow) stand-alone model, and can also be integrated with any deep networks, i.e., the
feature learning function can be a linear transformation and a deep network, respectively. Fig. 2
illustrates a sketch chart of our LDLFs, where a forest consists of two trees is shown.
We verify the effectiveness of our model on several LDL tasks, such as crowd opinion prediction on
movies and disease prediction based on human genes, as well as one computer vision application, i.e.,
facial age estimation, showing significant improvements to the state-of-the-art LDL methods. The
label distributions for these tasks include both unimodal distributions (e.g., the age distribution in
Fig. 1(a)) and mixture distributions (the rating distribution on a movie in Fig. 1(b)). The superiority
of our model on both of them verifies its ability to model any general form of label distributions
Figure 2: Illustration of a label distribution learning forest. The top circles denote the output units
of the function f parameterized by Θ, which can be a feature vector or a fully-connected layer of
a deep network. The blue and green circles are split nodes and leaf nodes, respectively. Two index
function ϕ1 and ϕ2 are assigned to these two trees respectively. The black dash arrows indicate the
correspondence between the split nodes of these two trees and the output units of function f . Note
that, one output unit may correspond to the split nodes belonging to different trees. Each tree has
independent leaf node predictions q (denoted by histograms in leaf nodes). The output of the forest
is a mixture of the tree predictions. f (·; Θ) and q are learned jointly in an end-to-end manner.
2
2
Related Work
Since our LDL algorithm is inspired by differentiable decision trees, it is necessary to first review
some typical techniques of decision trees. Then, we discuss current LDL methods.
Decision trees. Random forests or randomized decision trees [16, 1, 3, 4], are a popular ensemble
predictive model suitable for many machine learning tasks. In the past, learning of a decision tree was
based on heuristics such as a greedy algorithm where locally-optimal hard decisions are made at each
split node [1], and thus, cannot be integrated into in a deep learning framework, i.e., be combined
with representation learning in an end-to-end manner.
The newly proposed deep neural decision forests (dNDFs) [20] overcomes this problem by introducing
a soft differentiable decision function at the split nodes and a global loss function defined on a tree.
This ensures that the split node parameters can be learned by back-propagation and leaf node
predictions can be updated by a discrete iterative function.
Our method extends dNDFs to address LDL problems, but this extension is non-trivial, because
learning leaf node predictions is a constrained convex optimization problem. Although a step-size
free update function was given in dNDFs to update leaf node predictions, it was only proved to
converge for a classification loss. Consequently, it was unclear how to obtain such an update function
for other losses. We observed, however, that the update function in dNDFs can be derived from
variational bounding, which allows us to extend it to our LDL loss. In addition, the strategies used in
LDLFs and dNDFs to learning the ensemble of multiple trees (forests) are different: 1) we explicitly
define a loss function for forests, while only the loss function for a single tree was defined in dNDFs;
2) we allow the split nodes from different trees to be connected to the same output unit of the feature
learning function, while dNDFs did not; 3) all trees in LDLFs can be learned jointly, while trees in
dNDFs were learned alternatively. These changes in the ensemble learning are important, because as
shown in our experiments (Sec. 4.4), LDLFs can get better results by using more trees, but by using
the ensemble strategy proposed in dNDFs, the results of forests are even worse than those for a single
tree.
To sum up, w.r.t. dNDFs [20], the contributions of LDLFs are: first, we extend from classification [20]
to distribution learning by proposing a distribution-based loss for the forests and derive the gradient to
learn splits nodes w.r.t. this loss; second, we derived the update function for leaf nodes by variational
bounding (having observed that the update function in [20] was a special case of variational
bounding); last but not the least, we propose above three strategies to learning the ensemble of
multiple trees, which are different from [20], but we show are effective.
Label distribution learning. A number of specialized algorithms have been proposed to address the
LDL task, and have shown their effectiveness in many computer vision applications, such as facial
age estimation [8, 11, 28], expression recognition [30] and hand orientation estimation [10].
Geng et al. [8] defined the label distribution for an instance as a vector containing the probabilities
of the instance having each label. They also gave a strategy to assign a proper label distribution
to an instance with a single label, i.e., assigning a Gaussian or Triangle distribution whose peak
is the single label, and proposed an algorithm called IIS-LLD, which is an iterative optimization
process based on a two-layer energy based model. Yang et al. [28] then defined a three-layer energy
based model, called SCE-LDL, in which the ability to perform feature learning is improved by
adding the extra hidden layer and sparsity constraints are also incorporated to ameliorate the model.
Geng [6] developed an accelerated version of IIS-LLD, called BFGS-LDL, by using quasi-Newton
optimization. All the above LDL methods assume that the label distribution can be represented by a
maximum entropy model [2], but the exponential part of this model restricts the generality of the
distribution form.
Another way to address the LDL task, is to extend existing learning algorithms to deal with label
distributions. Geng and Hou [7] proposed LDSVR, a LDL method by extending support vector
regressor, which fit a sigmoid function to each component of the distribution simultaneously by a
support vector machine. Xing et al. [27] then extended boosting to address the LDL task by additive
weighted regressors. They showed that using the vector tree model as the weak regressor can lead to
better performance and named this method AOSO-LDLLogitBoost. As the learning of this tree model
is based on locally-optimal hard data partition functions at each split node, AOSO-LDLLogitBoost is
unable to be combined with representation learning. Extending current deep learning algorithms to
3
address the LDL task is an interesting topic. But, the existing such a method, called DLDL [5], still
focuses on maximum entropy model based LDL.
Our method, LDLFs, extends differentiable decision trees to address LDL tasks, in which the predicted
label distribution for a sample can be expressed by a linear combination of the label distributions
of the training data, and thus have no restrictions on the distributions (e.g., no requirement of the
maximum entropy model). In addition, thanks to the introduction of differentiable decision functions,
LDLFs can be combined with representation learning, e.g., to learn deep features in an end-to-end
manner.
3
Label Distribution Learning Forests
A forest is an ensemble of decision trees. We first introduce how to learn a single decision tree by
label distribution learning, then describe the learning of a forest.
3.1
Problem Formulation
Let X = Rm denote the input space and Y = {y1 , y2 , . . . , yC } denote the complete set of labels,
where C is the number of possible label values. We consider a label distribution learning (LDL)
problem, where for each input sample x ∈ X , there is a label distribution d = (dxy1 , dyx2 , . . . , dyxC )> ∈
RC . Here dyxc expresses the probability of the sample x having the c-th label yc and thus has the
PC
constraints that dyxc ∈ [0, 1] and c=1 dyxc = 1. The goal of the LDL problem is to learn a mapping
function g : x → d between an input sample x and its corresponding label distribution d.
Here, we want to learn the mapping function g(x) by a decision tree based model T . A decision
tree consists of a set of split nodes N and a set of leaf nodes L. Each split node n ∈ N defines
a split function sn (·; Θ) : X → [0, 1] parameterized by Θ to determine whether a sample is sent
to the left or right subtree. Each leaf node ` ∈ L holds a distribution q` = (q`1 , q`2 , . . . , q`C )>
PC
over Y, i.e, q`c ∈ [0, 1] and c=1 q`c = 1. To build a differentiable decision tree, following [20],
we use a probabilistic split function sn (x; Θ) = σ(fϕ(n) (x; Θ)), where σ(·) is a sigmoid function,
ϕ(·) is an index function to bring the ϕ(n)-th output of function f (x; Θ) in correspondence with
split node n, and f : x → RM is a real-valued feature learning function depending on the sample x
and the parameter Θ, and can take any form. For a simple form, it can be a linear transformation
of x, where Θ is the transformation matrix; For a complex form, it can be a deep network to
perform representation learning in an end-to-end manner, then Θ is the network parameter. The
correspondence between the split nodes and the output units of function f , indicated by ϕ(·) that is
randomly generated before tree learning, i.e., which output units from “f ” are used for constructing a
tree is determined randomly. An example to demonstrate ϕ(·) is shown in Fig. 2. Then, the probability
of the sample x falling into leaf node ` is given by
Y
l
r
p(`|x; Θ) =
sn (x; Θ)1(`∈Ln ) (1 − sn (x; Θ))1(`∈Ln ) ,
(1)
n∈N
where 1(·) is an indicator function and Lln and Lrn denote the sets of leaf nodes held by the left and
right subtrees of node n, Tnl and Tnr , respectively. The output of the tree T w.r.t. x, i.e., the mapping
function g, is defined by
X
g(x; Θ, T ) =
p(`|x; Θ)q` .
(2)
`∈L
3.2
Tree Optimization
Given a training set S = {(xi , di )}N
i=1 , our goal is to learn a decision tree T described in Sec. 3.1
which can output a distribution g(xi ; Θ, T ) similar to di for each sample xi . To this end, a
straightforward way is to minimize the Kullback-Leibler (K-L) divergence between each g(xi ; Θ, T )
and di , or equivalently to minimize the following cross-entropy loss:
R(q, Θ; S) = −
N C
N C
X
1 X X yc
1 X X yc
dxi log(gc (xi ; Θ, T )) = −
dxi log
p(`|xi ; Θ)q`c , (3)
N i=1 c=1
N i=1 c=1
`∈L
4
where q denote the distributions held by all the leaf nodes L and gc (xi ; Θ, T ) is the c-th output unit
of g(xi ; Θ, T ). Learning the tree T requires the estimation of two parameters: 1) the split node
parameter Θ and 2) the distributions q held by the leaf nodes. The best parameters (Θ∗ , q∗ ) are
determined by
(Θ∗ , q∗ ) = arg min R(q, Θ; S).
(4)
Θ,q
To solve Eqn. 4, we consider an alternating optimization strategy: First, we fix q and optimize
Θ; Then, we fix Θ and optimize q. These two learning steps are alternatively performed, until
convergence or a maximum number of iterations is reached (defined in the experiments).
3.2.1
Learning Split Nodes
In this section, we describe how to learn the parameter Θ for split nodes, when the distributions held
by the leaf nodes q are fixed. We compute the gradient of the loss R(q, Θ; S) w.r.t. Θ by the chain
rule:
N
∂R(q, Θ; S) X X ∂R(q, Θ; S) ∂fϕ(n) (xi ; Θ)
=
,
(5)
∂Θ
∂fϕ(n) (xi ; Θ)
∂Θ
i=1
n∈N
where only the first term depends on the tree and the second term depends on the specific type of the
function fϕ(n) . The first term is given by
C
gc (xi ; Θ, Tnl )
gc (xi ; Θ, Tnr )
1 X yc
∂R(q, Θ; S)
=
dxi sn (xi ; Θ)
− 1 − sn (xi ; Θ)
, (6)
∂fϕ(n) (xi ; Θ)
N c=1
gc (xi ; Θ, T )
gc (xi ; Θ, T )
P
P
where gc (xi ; Θ, Tnl ) = `∈Lln p(`|xi ; Θ)q`c and g c (xi ; Θ, Tnr ) = `∈Lrn p(`|xi ; Θ)q`c . Note that,
let Tn be the tree rooted at the node n, then we have gc (xi ; Θ, Tn ) = gc (xi ; Θ, Tnl ) + gc (xi ; Θ, Tnr ).
This means the gradient computation in Eqn. 6 can be started at the leaf nodes and carried out in a
bottom up manner. Thus, the split node parameters can be learned by standard back-propagation.
3.2.2
Learning Leaf Nodes
Now, fixing the parameter Θ, we show how to learn the distributions held by the leaf nodes q, which
is a constrained optimization problem:
min R(q, Θ; S), s.t., ∀`,
q
C
X
q`c = 1.
(7)
c=1
Here, we propose to address this constrained convex optimization problem by variational bounding [19, 29], which leads to a step-size free and fast-converged update rule for q. In variational
bounding, an original objective function to be minimized gets replaced by its bound in an iterative
manner. A upper bound for the loss function R(q, Θ; S) can be obtained by Jensen’s inequality:
R(q, Θ; S) = −
N C
X
1 X X yc
dxi log
p(`|xi ; Θ)q`c
N i=1 c=1
`∈L
≤−
where ξ` (q`c , xi ) =
1
N
N X
C
X
i=1 c=1
p(`|xi ;Θ)q`c
gc (xi ;Θ,T )
φ(q, q̄) = −
dyxci
X
ξ` (q̄`c , xi ) log
`∈L
p(`|x ; Θ)q
i
`c
,
ξ` (q̄`c , xi )
(8)
. We define
N C
p(`|x ; Θ)q
1 X X yc X
i
`c
dxi
ξ` (q̄`c , xi ) log
.
N i=1 c=1
ξ` (q̄`c , xi )
(9)
`∈L
Then φ(q, q̄) is an upper bound for R(q, Θ; S), which has the property that for any q and q̄,
φ(q, q̄) ≥ R(q, Θ; S), and φ(q, q) = R(q, Θ; S). Assume that we are at a point q(t) corresponding
to the t-th iteration, then φ(q, q(t) ) is an upper bound for R(q, Θ; S). In the next iteration, q(t+1)
is chosen such that φ(q(t+1) , q) ≤ R(q(t) , Θ; S), which implies R(q(t+1) , Θ; S) ≤ R(q(t) , Θ; S).
5
Consequently, we can minimize φ(q, q̄) instead of R(q, Θ; S) after ensuring that R(q(t) , Θ; S) =
φ(q(t) , q̄), i.e., q̄ = q(t) . So we have
q(t+1) = arg min φ(q, q(t) ), s.t., ∀`,
q
C
X
q`c = 1,
(10)
c=1
which leads to minimizing the Lagrangian defined by
ϕ(q, q(t) ) = φ(q, q(t) ) +
X
λ` (
`∈L
where λ` is the Lagrange multiplier. By setting
λ` =
(t+1)
Note that, q`c
(t+1)
∈ [0, 1] and
distributions held by the leaf nodes. The starting
(0)
distribution: q`c = C1 .
3.3
q`c − 1),
(11)
c=1
∂ϕ(q,q(t) )
∂q`c
N C
1 X X yc
(t)
(t+1)
d ξ` (q`c , xi ) and q`c
N i=1 c=1 xi
satisfies that q`c
C
X
= 0, we have
PN yc
(t)
dxi ξ` (q`c , xi )
.
= PC i=1
PN yc
(t)
ξ
(q
,
x
)
d
x
`
i
i
c=1
i=1
`c
(12)
(t+1)
= 1. Eqn. 12 is the update scheme for
c=1 q`c
(0)
point q` can be simply initialized by the uniform
PC
Learning a Forest
A forest is an ensemble of decision trees F = {T1 , . . . , TK }. In the training stage, all trees in the
forest F use the same parameters Θ for feature learning function f (·; Θ) (but correspond to different
output units of f assigned by ϕ, see Fig. 2), but each tree has independent leaf node predictions
q. The loss function for a forest is given by averaging the loss functions for all individual trees:
PK
1
RF = K
k=1 RTk , where RTk is the loss function for tree Tk defined by Eqn. 3. To learn Θ by
fixing the leaf node predictions q of all the trees in the forest F, based on the derivation in Sec. 3.2
and referring to Fig. 2, we have
N K
∂fϕk (n) (xi ; Θ)
1 XX X
∂RF
∂RTk
=
,
∂Θ
K i=1
∂fϕk (n) (xi ; Θ)
∂Θ
(13)
k=1 n∈Nk
where Nk and ϕk (·) are the split node set and the index function of Tk , respectively. Note that,
the index function ϕk (·) for each tree is randomly assigned before tree learning, and thus split
nodes correspond to a subset of output units of f . This strategy is similar to the random subspace
method [17], which increases the randomness in training to reduce the risk of overfitting.
As for q, since each tree in the forest F has its own leaf node predictions q, we can update them
independently by Eqn. 12, given by Θ. For implementational convenience, we do not conduct this
update scheme on the whole dataset S but on a set of mini-batches B. The training procedure of a
LDLF is shown in Algorithm. 1.
Algorithm 1 The training procedure of a LDLF.
Require: S: training set, nB : the number of mini-batches to update q
Initialize Θ randomly and q uniformly, set B = {∅}
while Not converge do
while |B| < nB do
Randomly select a mini-batch B from S
UpdateS
Θ by computing gradient (Eqn. 13) on B
B=B B
end while
Update q by iterating Eqn. 12 on B
B = {∅}
end while
In the testing stage, the output of the forest F is given by averaging the predictions from all the
PK
1
individual trees: g(x; Θ, F) = K
k=1 g(x; Θ, Tk ).
6
4
Experimental Results
Our realization of LDLFs is based on “Caffe” [18]. It is modular and implemented as a standard
neural network layer. We can either use it as a shallow stand-alone model (sLDLFs) or integrate it
with any deep networks (dLDLFs). We evaluate sLDLFs on different LDL tasks and compare it with
other stand-alone LDL methods. As dLDLFs can be learned from raw image data in an end-to-end
manner, we verify dLDLFs on a computer vision application, i.e., facial age estimation. The default
settings for the parameters of our forests are: tree number (5), tree depth (7), output unit number of
the feature learning function (64), iteration times to update leaf node predictions (20), the number of
mini-batches to update leaf node predictions (100), maximum iteration (25000).
4.1
Comparison of sLDLFs to Stand-alone LDL Methods
We compare our shallow model sLDLFs with other state-of-the-art stand-alone LDL methods.
For sLDLFs, the feature learning function f (x, Θ) is a linear transformation of x, i.e., the i-th
output unit fi (x, θi ) = θi> x, where θi is the i-th column of the transformation matrix Θ. We
used 3 popular LDL datasets in [6], Movie, Human Gene and Natural Scene1 . The samples
in these 3 datasets are represented by numerical descriptors, and the ground truths for them are
the rating distributions of crowd opinion on movies, the diseases distributions related to human
genes and label distributions on scenes, such as plant, sky and cloud, respectively. The label
distributions of these 3 datasets are mixture distributions, such as the rating distribution shown in
Fig. 1(b). Following [7, 27], we use 6 measures to evaluate the performances of LDL methods,
which compute the average similarity/distance between the predicted rating distributions and the real
rating distributions, including 4 distance measures (K-L, Euclidean, Sφrensen, Squared χ2 ) and two
similarity measures (Fidelity, Intersection).
We evaluate our shallow model sLDLFs on these 3 datasets and compare it with other state-of-the-art
stand-alone LDL methods. The results of sLDLFs and the competitors are summarized in Table 1.
For Movie we quote the results reported in [27], as the code of [27] is not publicly available. For the
results of the others two, we run code that the authors had made available. In all case, following [27, 6],
we split each dataset into 10 fixed folds and do standard ten-fold cross validation, which represents
the result by “mean±standard deviation” and matters less how training and testing data get divided.
As can be seen from Table 1, sLDLFs perform best on all of the six measures.
Table 1: Comparison results on three LDL datasets [6]. “↑” and “↓” indicate the larger and the smaller
the better, respectively.
Dataset
Method
K-L ↓
Euclidean ↓
Sφrensen ↓
Squared χ2 ↓
Fidelity ↑
Intersection ↑
Movie
sLDLF (ours)
AOSO-LDLogitBoost [27]
LDLogitBoost [27]
LDSVR [7]
BFGS-LDL [6]
IIS-LDL [11]
0.073±0.005
0.086±0.004
0.090±0.004
0.092±0.005
0.099±0.004
0.129±0.007
0.133±0.003
0.155±0.003
0.159±0.003
0.158±0.004
0.167±0.004
0.187±0.004
0.130±0.003
0.152±0.003
0.155±0.003
0.156±0.004
0.164±0.003
0.183±0.004
0.070±0.004
0.084±0.003
0.088±0.003
0.088±0.004
0.096±0.004
0.120±0.005
0.981±0.001
0.978±0.001
0.977±0.001
0.977±0.001
0.974±0.001
0.967±0.001
0.870±0.003
0.848±0.003
0.845±0.003
0.844±0.004
0.836±0.003
0.817±0.004
sLDLF (ours)
LDSVR [7]
BFGS-LDL [6]
IIS-LDL [11]
0.228±0.006
0.245±0.019
0.231±0.021
0.239±0.018
0.085±0.002
0.099±0.005
0.076±0.006
0.089±0.006
0.212±0.002
0.229±0.015
0.231±0.012
0.253±0.009
0.179±0.004
0.189±0.021
0.211±0.018
0.205±0.012
0.948±0.001
0.940±0.006
0.938±0.008
0.944±0.003
0.788±0.002
0.771±0.015
0.769±0.012
0.747±0.009
sLDLF (ours)
LDSVR [7]
BFGS-LDL [6]
IIS-LDL [11]
0.534±0.013
0.852±0.023
0.856±0.061
0.879±0.023
0.317±0.014
0.511±0.021
0.475±0.029
0.458±0.014
0.336±0.010
0.492±0.016
0.508±0.026
0.539±0.011
0.448±0.017
0.595±0.026
0.716±0.041
0.792±0.019
0.824±0.008
0.813±0.008
0.722±0.021
0.686±0.009
0.664±0.010
0.509±0.016
0.492±0.026
0.461±0.011
Human Gene
Natural Scene
4.2
Evaluation of dLDLFs on Facial Age Estimation
In some literature [8, 11, 28, 15, 5], age estimation is formulated as a LDL problem. We conduct
facial age estimation experiments on Morph [24], which contains more than 50,000 facial images
from about 13,000 people of different races. Each facial image is annotated with a chronological age.
To generate an age distribution for each face image, we follow the same strategy used in [8, 28, 5],
which uses a Gaussian distribution whose mean is the chronological age of the face image (Fig. 1(a)).
The predicted age for a face image is simply the age having the highest probability in the predicted
1
We download these datasets from http://cse.seu.edu.cn/people/xgeng/LDL/index.htm.
7
label distribution. The performance of age estimation is evaluated by the mean absolute error (MAE)
between predicted ages and chronological ages. As the current state-of-the-art result on Morph
is obtain by fine-tuning DLDL [5] on VGG-Face [23], we also build a dLDLF on VGG-Face, by
replacing the softmax layer in VGGNet by a LDLF. Following [5], we do standard 10 ten-fold cross
validation and the results are summarized in Table. 2, which shows dLDLF achieve the state-of-the-art
performance on Morph. Note that, the significant performance gain between deep LDL models
(DLDL and dLDLF) and non-deep LDL models (IIS-LDL, CPNN, BFGS-LDL) and the superiority
of dLDLF compared with DLDL verifies the effectiveness of end-to-end learning and our tree-based
model for LDL, respectively.
Table 2: MAE of age estimation comparison on Morph [24].
Method
IIS-LDL [11]
CPNN [11]
BFGS-LDL [6]
DLDL+VGG-Face [5]
dLDLF+VGG-Face (ours)
MAE
5.67±0.15
4.87±0.31
3.94±0.05
2.42±0.01
2.24±0.02
As the distribution of gender and ethnicity is very unbalanced in Morph, many age estimation methods [13, 14, 15] are evaluated on a subset of Morph, called Morph_Sub for short, which consists of
20,160 selected facial images to avoid the influence of unbalanced distribution. The best performance
reported on Morph_Sub is given by D2LDL [15], a data-dependent LDL method. As D2LDL used
the output of the “fc7” layer in AlexNet [21] as the face image features, here we integrate a LDLF
with AlexNet. Following the experiment setting used in D2LDL, we evaluate our dLDLF and the
competitors, including both SLL and LDL based methods, under six different training set ratios (10%
to 60%). All of the competitors are trained on the same deep features used by D2LDL. As can be
seen from Table 3, our dLDLFs significantly outperform others for all training set ratios.
Note that, the generated age distri- Figure 3: MAE of age estimation comparison on
butions are unimodal distributions Morph_Sub.
and the label distributions used in
Training set ratio
Method
Sec. 4.1 are mixture distributions.
10%
20%
30%
40%
50%
60%
The proposed method LDLFs achieve
AAS [22]
4.9081
4.7616
4.6507
4.5553
4.4690
4.4061
the state-of-the-art results on both of
LARR [12]
4.7501
4.6112
4.5131
4.4273
4.3500
4.2949
IIS-ALDL [9]
4.1791
4.1683
4.1228
4.1107
4.1024
4.0902
them, which verifies that our model
D2LDL [15]
4.1080
3.9857
3.9204
3.8712
3.8560
3.8385
has the ability to model any general
dLDLF (ours)
3.8495
3.6220
3.3991
3.2401
3.1917
3.1224
form of label distributions.
4.3
Time Complexity
Let h and sB be the tree depth and the
batch size, respectively. Each tree has 2h−1 − 1 split nodes and 2h−1 leaf nodes. Let D = 2h−1 − 1.
For one tree and one sample, the complexity of a forward pass and a backward pass are O(D +
D + 1×C) = O(D×C) and O(D + 1×C + D×C) = O(D×C), respectively. So for K trees and
nB batches, the complexity of a forward and backward pass is O(D×C×K×nB ×sB ). The complexity of an iteration to update leaf nodes are O(nB ×sB ×K×C×D + 1) = O(D×C×K×nB ×sB ).
Thus, the complexity for the training procedure (one epoch, nB batches) and the testing procedure
(one sample) are O(D×C×K×nB ×sB ) and O(D×C×K), respectively. LDLFs are efficient: On
Morph_Sub (12636 training images, 8424 testing images), our model only takes 5250s for training
(25000 iterations) and 8s for testing all 8424 images.
4.4
Parameter Discussion
Now we discuss the influence of parameter settings on performance. We report the results of rating
prediction on Movie (measured by K-L) and age estimation on Morph_Sub with 60% training set
ratio (measured by MAE) for different parameter settings in this section.
Tree number. As a forest is an ensemble model, it is necessary to investigate how performances
change by varying the tree number used in a forest. Note that, as we discussed in Sec. 2, the
ensemble strategy to learn a forest proposed in dNDFs [20] is different from ours. Therefore, it is
necessary to see which ensemble strategy is better to learn a forest. Towards this end, we replace our
ensemble strategy in dLDLFs by the one used in dNDFs, and name this method dNDFs-LDL. The
corresponding shallow model is named by sNDFs-LDL. We fix other parameters, i.e., tree depth and
8
output unit number of the feature learning function, as the default setting. As shown in Fig. 4 (a), our
ensemble strategy can improve the performance by using more trees, while the one used in dNDFs
even leads to a worse performance than one for a single tree.
Observed from Fig. 4, the performance of LDLFs can be improved by using more trees, but the
improvement becomes increasingly smaller and smaller. Therefore, using much larger ensembles
does not yield a big improvement (On Movie, the number of trees K = 100: K-L = 0.070 vs K = 20:
K-L = 0.071). Note that, not all random forests based methods use a large number of trees, e.g.,
Shotton et al. [25] obtained very good pose estimation results from depth images by only 3 decision
trees.
Tree depth. Tree depth is another important parameter for decision trees. In LDLFs, there is an
implicit constraint between tree depth h and output unit number of the feature learning function τ :
τ ≥ 2h−1 − 1. To discuss the influence of tree depth to the performance of dLDLFs, we set τ = 2h−1
and fix tree number K = 1, and the performance change by varying tree depth is shown in Fig. 4 (b).
We see that the performance first improves then decreases with the increase of the tree depth. The
reason is as the tree depth increases, the dimension of learned features increases exponentially, which
greatly increases the training difficulty. So using much larger depths may lead to bad performance
(On Movie, tree depth h = 18: K-L = 0.1162 vs h = 9: K-L = 0.0831).
Figure 4: The performance change of age estimation on Morph_Sub and rating prediction on Movie
by varying (a) tree number and (b) tree depth. Our approach (dLDLFs/sLDLFs) can improve the
performance by using more trees, while using the ensemble strategy proposed in dNDFs (dNDFsLDL/sNDFs-LDL) even leads to a worse performance than one for a single tree.
5
Conclusion
We present label distribution learning forests, a novel label distribution learning algorithm inspired by
differentiable decision trees. We defined a distribution-based loss function for the forests and found
that the leaf node predictions can be optimized via variational bounding, which enables all the trees
and the feature they use to be learned jointly in an end-to-end manner. Experimental results showed
the superiority of our algorithm for several LDL tasks and a related computer vision application, and
verified our model has the ability to model any general form of label distributions.
Acknowledgement. This work was supported in part by the National Natural Science Foundation of
China No. 61672336, in part by “Chen Guang” project supported by Shanghai Municipal Education
Commission and Shanghai Education Development Foundation No. 15CG43 and in part by ONR
N00014-15-1-2356.
References
[1] Y. Amit and D. Geman. Shape quantization and recognition with randomized trees. Neural Computation,
9(7):1545–1588, 1997.
[2] A. L. Berger, S. D. Pietra, and V. J. D. Pietra. A maximum entropy approach to natural language processing.
Computational Linguistics, 22(1):39–71, 1996.
[3] L. Breiman. Random forests. Machine Learning, 45(1):5–32, 2001.
[4] A. Criminisi and J. Shotton. Decision Forests for Computer Vision and Medical Image Analysis. Springer,
2013.
[5] B.-B. Gao, C. Xing, C.-W. Xie, J. Wu, and X. Geng. Deep label distribution learning with label ambiguity.
arXiv:1611.01731, 2017.
[6] X. Geng. Label distribution learning. IEEE Trans. Knowl. Data Eng., 28(7):1734–1748, 2016.
9
[7] X. Geng and P. Hou. Pre-release prediction of crowd opinion on movies by label distribution learning. In
Pro. IJCAI, pages 3511–3517, 2015.
[8] X. Geng, K. Smith-Miles, and Z. Zhou. Facial age estimation by learning from label distributions. In Proc.
AAAI, 2010.
[9] X. Geng, Q. Wang, and Y. Xia. Facial age estimation by adaptive label distribution learning. In Proc.
ICPR, pages 4465–4470, 2014.
[10] X. Geng and Y. Xia. Head pose estimation based on multivariate label distribution. In Proc. CVPR, pages
1837–1842, 2014.
[11] X. Geng, C. Yin, and Z. Zhou. Facial age estimation by learning from label distributions. IEEE Trans.
Pattern Anal. Mach. Intell., 35(10):2401–2412, 2013.
[12] G. Guo, Y. Fu, C. R. Dyer, and T. S. Huang. Image-based human age estimation by manifold learning and
locally adjusted robust regression. IEEE Trans. Image Processing, 17(7):1178–1188, 2008.
[13] G. Guo and G. Mu. Human age estimation: What is the influence across race and gender? In CVPR
Workshops, pages 71–78, 2010.
[14] G. Guo and C. Zhang. A study on cross-population age estimation. In Proc. CVPR, pages 4257–4263,
2014.
[15] Z. He, X. Li, Z. Zhang, F. Wu, X. Geng, Y. Zhang, M.-H. Yang, and Y. Zhuang. Data-dependent label
distribution learning for age estimation. IEEE Trans. on Image Processing, 2017.
[16] T. K. Ho. Random decision forests. In Proc. ICDAR, pages 278–282, 1995.
[17] T. K. Ho. The random subspace method for constructing decision forests. IEEE Trans. Pattern Anal. Mach.
Intell., 20(8):832–844, 1998.
[18] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe:
Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014.
[19] M. I. Jordan, Z. Ghahramani, T. S. Jaakkola, and L. K. Saul. An introduction to variational methods for
graphical models. Machine Learning, 37(2):183–233, 1999.
[20] P. Kontschieder, M. Fiterau, A. Criminisi, and S. R. Bulò. Deep neural decision forests. In Proc. ICCV,
pages 1467–1475, 2015.
[21] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural
networks. In Proc. NIPS, pages 1106–1114, 2012.
[22] A. Lanitis, C. Draganova, and C. Christodoulou. Comparing different classifiers for automatic age
estimation. IEEE Trans. on Cybernetics,, 34(1):621–628, 2004.
[23] O. M. Parkhi, A. Vedaldi, and A. Zisserman. Deep face recognition. In Proc. BMVC, pages 41.1–41.12,
2015.
[24] K. Ricanek and T. Tesafaye. MORPH: A longitudinal image database of normal adult age-progression. In
Proc. FG, pages 341–345, 2006.
[25] J. Shotton, A. W. Fitzgibbon, M. Cook, T. Sharp, M. Finocchio, R. Moore, A. Kipman, and A. Blake.
Real-time human pose recognition in parts from single depth images. In Proc. CVPR, pages 1297–1304,
2011.
[26] G. Tsoumakas and I. Katakis. Multi-label classification: An overview. International Journal of Data
Warehousing and Mining, 3(3):1–13, 2007.
[27] C. Xing, X. Geng, and H. Xue. Logistic boosting regression for label distribution learning. In Proc. CVPR,
pages 4489–4497, 2016.
[28] X. Yang, X. Geng, and D. Zhou. Sparsity conditional energy label distribution learning for age estimation.
In Proc. IJCAI, pages 2259–2265, 2016.
[29] A. L. Yuille and A. Rangarajan. The concave-convex procedure. Neural Computation, 15(4):915–936,
2003.
[30] Y. Zhou, H. Xue, and X. Geng. Emotion distribution recognition from facial expressions. In Proc. MM,
pages 1247–1250, 2015.
10
| 1cs.CV
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.