MétaCan
Menu
← Back to cohort
Record W7094350435

A Split Fast Fourier Transform Algorithm for Block Toeplitz Matrix-Vector Multiplication

2024· other· fr· W7094350435 on OpenAlexfundno aff

Bibliographic record

VenuePolyPublie (École Polytechnique de Montréal) · 2024
Typeother
Languagefr
Field
Topic
Canadian institutionsnot available
FundersInstitut de Valorisation des DonnéesCanada First Research Excellence Fund
KeywordsToeplitz matrixMultiplication (music)Matrix multiplicationBlock (permutation group theory)Fast Fourier transform
DOInot available

Abstract

fetched live from OpenAlex

RÉSUMÉ: La modélisation des phénomènes physiques invariants par translation, dans une base vectorielle spatiale, entraine des systèmes d’équations qui peuvent être représentés par des matrices de Toeplitz par bloc multi-niveaux. C’est notamment le cas des phénomènes liés à l’électromagnétisme ou l’acoustique. La résolution d’équations impliquant ces matrices requiert généralement une méthode de résolution itérative, nécessitant un grand nombre de multiplications de matrice de Toeplitz par bloc avec des vecteurs. Ainsi, pour simuler des problèmes physiques de grande taille, il est indispensable d’améliorer l’efficacité de ces opérations. Il existe un grand nombre de méthodes réalisant ces multiplications, mais chacune présente des limites majeures. Les plus prometteuses de ces méthodes sont celle par insertion de matrices circulantes et celle par décomposition tensorielle en paquet. Malheureusement, ces méthodes présentent des temps de calculs et des besoins en mémoire qui augmentent rapidement avec la taille du système ou le rang du tenseur. Dans le cas de vecteurs sources quelconques, le calcul de la décomposition tensorielle est plus coûteux que la multiplication directe. Aussi, lorsque l’on étend l’insertion de matrices circulantes aux matrices de Toeplitz par bloc en d dimensions, l’insertion entraine l’introduction d’un grand nombre de coefficients redondants. Dans ce cas, seul une proportion de 1/2d du vecteur traité dans le calcul contient des informations utiles. Cette redondance conduit rapidement à une surcharge de mémoire et à une complexité opérationnelle élevée. L’algorithme présenté dans ce mémoire a pour objectif de répondre à ce problème d’inefficacité en reportant au plus tard les insertions et en avançant au plus tôt les projections. Cette stratégie permet de calculer les transformées de Fourier sur un nombre réduit de coefficients, diminuant ainsi la complexité par un facteur de d/ 2 − 2−d+1 et la demande en mémoire maximale de 2/ (d + 1)2−d+1. Pour une multiplication matrice-vecteur en trois dimensions, le ratio entre la complexité théorique de la méthode standard par insertion de matrices circulantes et celle de la version améliorée de cette méthode tend vers 12/7 lorsque la taille de la matrice de Toeplitz par bloc tend vers l’infini pour chaque niveau. De plus, l’algorithme utilise une propriété de la transformée de Fourier rapide afin de séparer le vecteur transformé en deux vecteurs de coefficients d’indices paires et impaires respectivement, et ce pour chaque transformée de Fourier. Cela résulte en une structure arborescente des tâches de l’algorithme. ABSTRACT: Modeling translationally invariant physics, such as electromagnetism, in a spatial basis typ-ically results in systems of equations involving multi-level block-Toeplitz matrices. These matrices often necessitate iterative solution methods, which in turn require numerous block-Toeplitz matrix-vector multiplications. Therefore, enhancing the efficiency of these multi-plications is essential for simulating a wide range of large-scale physical systems. While various methods for performing these multiplications currently exist, each comes with major drawbacks. Specifically, the most promising methods—circulant embedding and tensor train decomposition—suffer from poor scalability with respect to system size or tensor rank. For general source vectors, computing the tensor decomposition is more computationally inten-sive than direct multiplication. When extended to d-dimensional block-Toeplitz matrices, embedding leads to the introduction of a large number of redundant coefficients, so that only 1/2d of the vector treated in computation contains useful information. This scaling quickly leads to memory overload and high operational complexity. The algorithm introduced in this thesis aims to address this latter inefficiency through lazy embeddings and eager projections, deferring embeddings as late as possible and performing projections as early as possible. This strategy allows for the computation of Fourier trans-forms on a reduced number of coefficients, lowering complexity by a factor of d/ 2 − 2−d+1and peak memory usage by a factor of 2/ (d + 1)2−d+1. For three-dimensional matrix-vector multiplications, the theoretical complexity ratio of the standard circulant embedding method over the enhanced version approaches 12/7 as the size of the block-Toeplitz matrix tends to infinity for each level. Additionally, the algorithm exploits a property of the fast Fourier transform (FFT) to divide the transformed vector into even and odd coefficients, fol-lowing a tree branch structure. This structure offers greater flexibility in managing memory and computational speed through various parallelization strategies along each branch using several GPUs. Running these branches independently on separate GPUs can boost compu-tational speed at the expense of increased memory usage. On a single GPU, the memory usage ratio between the standard and enhanced methods tends to 4/3 for three-dimensional matrices and vectors as the sizes of each level increase. In cases where the Toeplitz data is entirely symmetric or anti-symmetric—meaning each block at each level is a symmetric or anti-symmetric Toeplitz matrix–—a clever embedding can reduce memory consumption. The theoretical memory ratio between the two methods then becomes (2d +1)/(d+2).

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.

How this classification was reachedexpand

Full frame machine prediction

Teacher imitation

Not calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.

metaresearch head score (Codex)0.001
metaresearch head score (Gemma)0.002
Version: metacan-v3-hybrid-931329e0061cValidation status: machine_predicted_unvalidated
Candidate categoriesnone
Consensus categoriesnone
DomainCandidate signal: none · Consensus signal: none
Study designCandidate signal: Theoretical or conceptual · Consensus signal: none
GenreCandidate signal: Methods · Consensus signal: Methods
Teacher disagreement score0.014
Threshold uncertainty score0.046

Distilled classifier scores by category (both heads)

CategoryCodexGemma
Metaresearch0.0010.002
Meta-epidemiology (narrow)0.0010.000
Meta-epidemiology (broad)0.0010.001
Bibliometrics0.0010.001
Science and technology studies0.0000.000
Scholarly communication0.0010.001
Open science0.0010.001
Research integrity0.0010.001
Insufficient payload (model declined to judge)0.0140.005

Machine scores (provisional)

The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.

Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.

Opus teacher head0.010
GPT teacher head0.253
Teacher spread0.243 · how far apart the two teachers sit on this one work
Validation statusscore_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it

Classification

machine, unvalidated

Machine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.

The models applied no category: nothing in the taxonomy fit this work.
Study designTheoretical or conceptual
Domainnot available
GenreMethods

How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".

Quick stats

Citations0
Published2024
Admission routes1
Has abstractyes

Explore more

Same venuePolyPublie (École Polytechnique de Montréal)→French-language works237,207→