## Polynomial approximation algorithms for belief matrix maintenance in identity management (2004)

Venue: | In 43rd IEEE Conference on Decision and Control |

Citations: | 10 - 2 self |

### BibTeX

@INPROCEEDINGS{Balakrishnan04polynomialapproximation,

author = {Hamsa Balakrishnan and Inseok Hwang and Claire J. Tomlin},

title = {Polynomial approximation algorithms for belief matrix maintenance in identity management},

booktitle = {In 43rd IEEE Conference on Decision and Control},

year = {2004},

pages = {4874--4879}

}

### Years of Citing Articles

### OpenURL

### Abstract

Abstract — Updating probabilistic belief matrices as new observations arrive, in the presence of noise, is a critical part of many algorithms for target tracking in sensor networks. These updates have to be carried out while preserving sum constraints, arising for example, from probabilities. This paper addresses the problem of updating belief matrices to satisfy sum constraints using scaling algorithms. We show that the convergence behavior of the Sinkhorn scaling process, used for scaling belief matrices, can vary dramatically depending on whether the prior unscaled matrix is exactly scalable or only almost scalable. We give an efficient polynomial-time algorithm based on the maximum-flow algorithm that determines whether a given matrix is exactly scalable, thus determining the convergence properties of the Sinkhorn scaling process. We prove that the Sinkhorn scaling process always provides a solution to the problem of minimizing the Kullback-Leibler distance of the physically feasible scaled matrix from the prior constraint-violating matrix, even when the matrices are not exactly scalable. We pose the scaling process as a linearly constrained convex optimization problem, and solve it using an interior-point method. We prove that even in cases in which the matrices are not exactly scalable, the problem can be solved to ɛ−optimality in strongly polynomial time, improving the best known bound for the problem of scaling arbitrary nonnegative rectangular matrices to prescribed row and column sums. I.

### Citations

8603 |
Elements of information theory
- Cover, Thomas
- 1991
(Show Context)
Citation Context ...rupulous care is taken not to make any assumptions not presented by A. Bearing all this in mind, a suitable measure is the Kullback-Leibler measure (also known as the KL-distance or the cross-entropy =-=[5]-=-) given by: n� m� I(B : A) = bij log j=1 i=1 bij (1) aij This is sometimes also called the directed divergence (since it measures the divergence of distribution B from distribution A), and is denoted ... |

3703 |
Convex optimization
- Boyd, Vandenberghe
- 2004
(Show Context)
Citation Context ...in the relative interior of the feasible set. We have already seen how the maximum-flow formulation can be used to compute a feasible point x (0) in polynomial time. Using the techniques presented in =-=[4]-=- and [16], we can derive the following theorem: Theorem 4: The complexity of scaling an m×n matrix to specified row and column sums using the proposed interiorpoint method with a logarithmic barrier a... |

1415 |
Network flows: theory, algorithms, and applications
- Ahuja, Magnanti, et al.
- 1993
(Show Context)
Citation Context ...uch that bij ≥ ɛ∀{(i, j)|aij �= 0} if and only if the maximum flow equals K. The problem of finding the maximum flow in a network with arc lower bounds is as hard as solving two maximum flow problems =-=[1]-=- (Section 6.7, Flows with lower bounds). Both maximum flow problems are solved on bipartite graphs (one on the original graph, and one on a transformed graph); therefore, the complexity of finding an ... |

514 | A new approach to the maximum flow problem
- Goldberg, Tarjan
- 1986
(Show Context)
Citation Context ...solution to the Sinkhorn scaling process. The maximum flow problem in bipartite networks can be solved in O(pq log(q2 /p)), wherepis the number of nonzero elements in A =(aij), andqis min{n, m} ([2], =-=[9]-=-). 4877 Source r1 r2 rm−1 rm 1 2 m−1 b11 1 2 n−1 m bmn n Fig. 1. Equivalence of feasibility problem to maximum-flow problem B. Infeasibility of an exactly scaled solution We consider the case in which... |

362 |
AMPL: A Modeling Language for
- Fourer, Gay, et al.
- 1993
(Show Context)
Citation Context ...an ɛ of 10−8 .A MATLAB implementation of the Sinkhorn scaling process takes 1.718 seconds (and 7105 iterations) to converge to a solution, while an implementation of the interior-point method in AMPL =-=[6]-=- using the MINOS [15] solver for the centering takes 0.0156 seconds to reach the same solution. 4878sFinally, we present a 100 trial Monte Carlo simulation over a range of matrix sizes, for two differ... |

140 |
A relationship between arbitrary positive matrices and doubly stochastic matrices
- Sinkhorn
- 1964
(Show Context)
Citation Context ...e class of square matrices, the algorithm based on the barrier method has lower complexity. In Section V, we present some examples. II. SINKHORN SCALING The Sinkhorn scaling procedure was proposed in =-=[19]-=- as a method for scaling positive matrices to doubly stochastic matrices. Since then, there have been several extensions to treat the case of nonnegative matrices [21], to scaling positive rectangular... |

57 | A distributed algorithm for managing multi-target identities in wireless ad-hoc sensor networks
- Shin, Guibas, et al.
- 2003
(Show Context)
Citation Context ...h as sensor networks. Our motivation for this work stems from distributed identity management algorithms in air traffic control and sensor networks [12]. The identity or belief matrix was proposed in =-=[18]-=- as a possible method of integrating information available in the system with external information which might be available sporadically. The belief matrix is a matrix B, in which elements bij represe... |

55 | A deterministic strongly polynomial algorithm for matrix scaling and approximate permanents
- Linial, Samorodnitsky, et al.
(Show Context)
Citation Context ...ich attains ɛ−optimality with complexity O(n6 log(n/ɛ)) for an n × n matrix. Our approach to the problem is different from the only other strongly polynomial approximation scheme for matrix balancing =-=[13]-=-, which proposes a modified Sinkhorn algorithm; we also compare the complexity of the two schemes and show that for the class of square matrices, the algorithm based on the barrier method has lower co... |

50 |
MINOS 5.5 user’s guide
- Murtagh, Saunders
- 1983
(Show Context)
Citation Context ...AB implementation of the Sinkhorn scaling process takes 1.718 seconds (and 7105 iterations) to converge to a solution, while an implementation of the interior-point method in AMPL [6] using the MINOS =-=[15]-=- solver for the centering takes 0.0156 seconds to reach the same solution. 4878sFinally, we present a 100 trial Monte Carlo simulation over a range of matrix sizes, for two different cases - Sinkhorn ... |

41 | Improved algorithms for bipartite network flow
- Ahuja, Orlin, et al.
- 1994
(Show Context)
Citation Context ...ible solution to the Sinkhorn scaling process. The maximum flow problem in bipartite networks can be solved in O(pq log(q2 /p)), wherepis the number of nonzero elements in A =(aij), andqis min{n, m} (=-=[2]-=-, [9]). 4877 Source r1 r2 rm−1 rm 1 2 m−1 b11 1 2 n−1 m bmn n Fig. 1. Equivalence of feasibility problem to maximum-flow problem B. Infeasibility of an exactly scaled solution We consider the case in ... |

39 | Concerning nonnegative matrices and doubly stochastic matrices - Sinkhorn, Knopp - 1967 |

28 |
Biproportional Matrices and Input-Output Change. Number 16 in
- Bacharach
- 1970
(Show Context)
Citation Context ...tic. This method was shown to converge for different classes of matrices, in [21], [20] and [17]. Its properties were analyzed, for the special case of matrices known as exactly scalable matrices, in =-=[3]-=-, [17] and [8]. This technique was analyzed further and applied to the problem of identity management for Air Traffic Control in [10], [11]. The convergence behavior of the Sinkhorn scaling process fo... |

23 |
Recovering information from incomplete or partial multisectoral economic data
- Golan, Judge, et al.
- 1994
(Show Context)
Citation Context ...od was shown to converge for different classes of matrices, in [21], [20] and [17]. Its properties were analyzed, for the special case of matrices known as exactly scalable matrices, in [3], [17] and =-=[8]-=-. This technique was analyzed further and applied to the problem of identity management for Air Traffic Control in [10], [11]. The convergence behavior of the Sinkhorn scaling process for a nonnegativ... |

19 | Scaling of matrices to achieve specified row and column sums - Marshall, Olkin - 1968 |

19 |
Scalings of matrices which have prespecified row sums and column sums via optimisation, Linear Algebra Appl
- Rothblum, Schneider
- 1989
(Show Context)
Citation Context ...s paper. Sinkhorn iterations were first proposed as a method of scaling matrices to make them doubly-stochastic. This method was shown to converge for different classes of matrices, in [21], [20] and =-=[17]-=-. Its properties were analyzed, for the special case of matrices known as exactly scalable matrices, in [3], [17] and [8]. This technique was analyzed further and applied to the problem of identity ma... |

18 |
Diagonal equivalence to matrices with prescribed row and column sums
- Sinkhorn
(Show Context)
Citation Context ...lumn sums. A natural way is to simply normalize alternately the rows and columns until the constraints are met. This method of scaling by repeated normalization is called the Sinkhorn scaling process =-=[20]-=-. However, it is not obvious that such a process would always converge; and if it does, that it would converge in a reasonable amount of time. It is also not clear what the quality of the resulting so... |

13 | Multiple-target Tracking and Identity Management in clutter with application to aircraft tracking
- Hwang, Balakrishnan, et al.
- 2004
(Show Context)
Citation Context ...ecial case of matrices known as exactly scalable matrices, in [3], [17] and [8]. This technique was analyzed further and applied to the problem of identity management for Air Traffic Control in [10], =-=[11]-=-. The convergence behavior of the Sinkhorn scaling process for a nonnegative matrix depends greatly on the sparsity structure of the matrix, and can fall into one of two regimes. Most studies, such as... |

8 |
On the scaling of multidimensional matrices. Linear Algebra and its Applications
- Franklin, Lorenz
- 1989
(Show Context)
Citation Context ...column sums. While Sinkhorn and others ([20], [21], [14], [17]) proved that the iterative procedure converges for appropriate matrices, they did not study the rate of convergence. Franklin and Lorenz =-=[7]-=- showed that each iteration of Sinkhorn scaling for an exactly scalable matrix is a contraction map in the Hilbert projective metric, and they concluded that the number of iterations is bounded by O �... |

5 |
Multiple-target Tracking and Identity Management algorithm for Air Traffic Control
- Hwang, Balakrishnan, et al.
- 2003
(Show Context)
Citation Context ...the special case of matrices known as exactly scalable matrices, in [3], [17] and [8]. This technique was analyzed further and applied to the problem of identity management for Air Traffic Control in =-=[10]-=-, [11]. The convergence behavior of the Sinkhorn scaling process for a nonnegative matrix depends greatly on the sparsity structure of the matrix, and can fall into one of two regimes. Most studies, s... |

1 |
Distributed multiple-target tracking and identity management in sensor networks
- Hwang, Roy, et al.
- 2004
(Show Context)
Citation Context ...h the advent of large scale networks of systems, such as sensor networks. Our motivation for this work stems from distributed identity management algorithms in air traffic control and sensor networks =-=[12]-=-. The identity or belief matrix was proposed in [18] as a possible method of integrating information available in the system with external information which might be available sporadically. The belief... |