Probabilistic Iterative Methods for Linear Systems

Authors: Jon Cockayne, Ilse C.F. Ipsen, Chris J. Oates, Tim W. Reid

JMLR 2021 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conclude with an empirical illustration that highlights the potential for probabilistic iterative methods to provide insight into solution uncertainty. [...] The aim of this section is to empirically assess our proposed probabilistic iterative methods.
Researcher Affiliation Academia Jon Cockayne EMAIL Alan Turing Institute 96 Euston Road London, NW1 2DB, UK; Ilse C.F. Ipsen EMAIL Department of Mathematics North Carolina State University Raleigh, NC 27695-8205, USA; Chris J. Oates EMAIL School of Mathematics and Statistics Newcastle University Newcastle-upon-Tyne, NE1 7RU, UK; Tim W. Reid EMAIL Department of Mathematics North Carolina State University Raleigh, NC 27695-8205, USA
Pseudocode Yes Example 3 (Conjugate gradient method). In CG, for a symmetric positive-definite matrix A the iteration is of the form xm = xm 1 + αmsm αm = s mrm s m Asm sm+1 = rm + βmsm βm = r mrm r m 1rm 1 where the initial direction s0 is taken to be the initial residual r0, and we recall that recall that rm = b Axm.
Open Source Code Yes The code to reproduce these results is available on Git Hub https://github.com/jcockayne/probabilistic_iterative_methods_code
Open Datasets No A data set of size d = 520 was generated, with (zi)i=1,...,d consisting of 60 evenly spaced points in [0, 0.1], 400 evenly spaced points in [0.2, 0.8] and 60 evenly spaced points in [0.9, 1], and yi = f(zi) where f(z) = 1z<0.5 sin(2πz)+1z 0.5 sin(4πz). The paper describes generating a dataset but does not provide access information (link, citation, or repository) for public availability.
Dataset Splits No The paper defines a generated dataset for an interpolation problem. There is no mention of splitting this data into training, validation, or test sets; the entire dataset is used to form the linear system to be solved.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used to run the experiments.
Software Dependencies No The paper does not mention any specific software dependencies or their version numbers (e.g., Python, PyTorch, specific libraries).
Experiment Setup Yes The step size ω was set to either the optimal value in Fig. 1b, ω = 2/(λmin(A) + λmax(A)), that minimises the spectral radius of G, or a default value ω = 2/3 in Fig. 1a. [...] The parameter ℓ= 0.0012 was used, which produces a system for which a direct solver can be used. [...] For the initial distribution µ0 several candidates were considered. Firstly a default choice given by µ0 = N(0, Id). [...] Secondly the natural choice µ0 = N(0, A 1). [...] The third initial distribution we consider is applicable only in settings where a small number of ansatz solutions (i.e. guesses) are provided, perhaps obtained by expert knowledge of the system at hand. Let xi, i = 1, . . . , N, be these ansatz solutions; we use these to estimate the scaling parameter ν2 for an initial distribution µ0 = N(0, ν2Σ0) where Σ0 is fixed. [...] We used N = 5 ansatz solutions. [...] For the second order method, we opted to use the rich initial distribution for x1. [...] For each method, m = 10 iterations were performed.