arXiv subjects

Jean-Pierre Dussault

Publications and source records attributed to Jean-Pierre Dussault.

5 recordsLinked to original sources

arXiv2026

Quasi-Newton and Krylov Methods for the Solution of Nonconvex Trust-Region Subproblems

We study the solution of symmetric positive-definite linear systems by way of families of full- and limited-memory methods. Our contributions are threefold. We first derive new relationships between the conjugate-gradient method (CG) and quasi-Newton methods of the Broyden class that refine existing results, and clarify when those methods generate the same iterates and enjoy quadratic termination. We extend this perspective to the limited-memory BFGS (LBFGS) method. Next, we examine how DIOM, a limited-memory variant of the full orthogonalization Krylov method (FOM), is akin to LBFGS in that it provides a memory lever that is critical in practical performance. Finally, we generalize LBFGS and DIOM to the computation of trust-region steps for unconstrained, potentially nonconvex, optimization. We report numerical experience on positive-definite linear systems and unconstrained optimization problems. The results show that memory is a key algorithmic lever: LBFGS and DIOM are consistently more robust than CG and often achieve comparable accuracy with fewer Hessian-vector products. They emerge as viable alternatives to CG when high accuracy is desirable or when operations with the Hessian are at a premium. The limited-memory SR1 (LSR1) method can be competitive in full-memory form, but its limited-memory variant suffers from discarded curvature information.

math.OC↗

arXiv2023

A rewriting of the relation between the acolinearity of annihilation photons and their energy in the context of positron emission tomography

Acolinearity of the annihilation photons observed in Positron Emission Tomography (PET) is described as following a Gaussian distribution. However, it is never explicitly said if it refers to the amplitude of the acolinearity angle or its 2D distribution relative to the case without acolinearity (herein defined as the acolinearity deviation). Since the former is obtained by integrating the latter, a wrong interpretation would lead to very different results. The paper of Shibuya et al. (2007), differs from the previous studies since it is based on the precise measurement of the energy of the annihilation photons. They also show that acolinearity follows a Gaussian distribution in the context of PET. However, their notation, which relies on being on the plane where the two annihilation photons travel, could mean that their observation refers to the amplitude of the acolinearity angle. If that understanding is correct, it would mean that acolinearity deviation follows a 2D Gaussian distribution divided by the norm of its argument. Thus, we revisited the proof presented in Shibuya et al. (2007) by using an explicit description of the acolinearity in the 3D unit sphere.

physics.med-ph↗

arXiv2023

Parameterizing the Angular Distribution of Emission: A Model for TOF-PET Low-Count Reconstruction

Low-count reconstruction remains a challenge for Positron Emission Tomography (PET) even with the recent progress in time-of-flight (TOF) resolution. In this context, the bias between the acquired histogram, consisting of low values or zeros, and the expected histogram, obtained from the forward projector, is propagated to the image, resulting in a biased reconstruction. This situation could be exacerbated with finer resolution of the TOF information, which further sparsifies the acquired histogram. We propose a new approach to circumvent this limitation of the classical reconstruction model. It consists of extending the description of the reconstruction scheme to also explicitly include the projection domain by Parameterizing the Angular Distribution of Emission (PADE). This parametrization has greater degrees of freedom than the log-likelihood model, which can not be harnessed in classical circumstances. We hypothesize that with ultra-fast TOF this new approach would not only be viable for low-count reconstruction, but also more adequate than the classical reconstruction model. As a proof of concept, an implementation of this approach is compared to the log-likelihood model by investigating two-dimensional simulations of a hot spots phantom. The proposed model achieves similar contrast recovery coefficients as MLEM except for the smallest structures where the low-count nature of the simulations makes it difficult to draw conclusions. However, this new model seems to converge toward less noisy solutions than MLEM. These results suggest that the PADE approach has potential for low-count reconstruction with ultra-fast TOF.

physics.med-ph↗

arXiv2021

Scalable adaptive cubic regularization methods

Adaptive cubic regularization (ARC) methods for unconstrained optimization compute steps from linear systems involving a shifted Hessian in the spirit of the Levenberg-Marquardt and trust-region methods. The standard approach consists in performing an iterative search for the shift akin to solving the secular equation in trust-region methods. Such search requires computing the Cholesky factorization of a tentative shifted Hessian at each iteration, which limits the size of problems that can be reasonably considered. We propose a scalable implementation of ARC named ARCqK in which we solve a set of shifted systems concurrently by way of an appropriate modification of the Lanczos formulation of the conjugate gradient (CG) method. At each iteration of ARCqK to solve a problem with n variables, a range of m << n shift parameters is selected. The computational overhead in CG beyond the Lanczos process is thirteen scalar operations to update five vectors of length m and two n-vector updates for each value of the shift. The CG variant only requires one Hessian-vector product and one dot product per iteration, independently of the number of shift parameters. Solves corresponding to inadequate shift parameters are interrupted early. All shifted systems are solved inexactly. Such modest cost makes our implementation scalable and appropriate for large-scale problems. We provide a new analysis of the inexact ARC method including its worst case evaluation complexity, global and asymptotic convergence. We describe our implementation and provide preliminary numerical observations that confirm that for problems of size at least 100, our implementation of ARCqK is more efficient than a classic Steihaug-Toint trust region method. Finally, we generalize our convergence results to inexact Hessians and nonlinear least-squares problems.

math.OC↗

arXiv2014

Systematic Ensemble Learning for Regression

The motivation of this work is to improve the performance of standard stacking approaches or ensembles, which are composed of simple, heterogeneous base models, through the integration of the generation and selection stages for regression problems. We propose two extensions to the standard stacking approach. In the first extension we combine a set of standard stacking approaches into an ensemble of ensembles using a two-step ensemble learning in the regression setting. The second extension consists of two parts. In the initial part a diversity mechanism is injected into the original training data set, systematically generating different training subsets or partitions, and corresponding ensembles of ensembles. In the final part after measuring the quality of the different partitions or ensembles, a max-min rule-based selection algorithm is used to select the most appropriate ensemble/partition on which to make the final prediction. We show, based on experiments over a broad range of data sets, that the second extension performs better than the best of the standard stacking approaches, and is as good as the oracle of databases, which has the best base model selected by cross-validation for each data set. In addition to that, the second extension performs better than two state-of-the-art ensemble methods for regression, and it is as good as a third state-of-the-art ensemble method.

stat.ML↗