Pith. sign in

REVIEW 5 cited by

On the Expressive Power of Self-Attention Matrices

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2106.03764 v2 pith:D5YPUQOI submitted 2021-06-07 cs.LG

On the Expressive Power of Self-Attention Matrices

classification cs.LG
keywords matrixself-attentionsparseapproximatefixedmatricesonlydepending
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Transformer networks are able to capture patterns in data coming from many domains (text, images, videos, proteins, etc.) with little or no change to architecture components. We perform a theoretical analysis of the core component responsible for signal propagation between elements, i.e. the self-attention matrix. In practice, this matrix typically exhibits two properties: (1) it is sparse, meaning that each token only attends to a small subset of other tokens; and (2) it changes dynamically depending on the input to the module. With these considerations in mind, we ask the following question: Can a fixed self-attention module approximate arbitrary sparse patterns depending on the input? How small is the hidden size $d$ required for such approximation? We make progress in answering this question and show that the self-attention matrix can provably approximate sparse matrices, where sparsity is in terms of a bounded number of nonzero elements in each row and column. While the parameters of self-attention are fixed, various sparse matrices can be approximated by only modifying the inputs. Our proof is based on the random projection technique and uses the seminal Johnson-Lindenstrauss lemma. Our proof is constructive, enabling us to propose an algorithm for finding adaptive inputs and fixed self-attention parameters in order to approximate a given matrix. In particular, we show that, in order to approximate any sparse matrix up to a given precision defined in terms of preserving matrix element ratios, $d$ grows only logarithmically with the sequence length $L$ (i.e. $d = O(\log L)$).

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Looped Transformers with Layer Normalization Provably Learn the Power Method

    cs.LG 2026-05 unverdicted novelty 8.0

    Looped linear transformers with LN provably converge via GD to implement the power method on principal component prediction.

  2. Analogies between Transformer Layers and Power Method

    cs.LG 2026-05 unverdicted novelty 6.0

    Transformer layers are analogous to power method steps, tilting tokens toward the principal eigenvector of the output-value weight product, with stronger analytical and empirical alignment in shared-weight models and ...

  3. H$_2$O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models

    cs.LG 2023-06 unverdicted novelty 6.0

    H2O evicts non-heavy-hitter tokens from the KV cache using a dynamic submodular policy, retaining recent and frequent-co-occurrence tokens to reduce memory while preserving accuracy.

  4. Nexusformer: Nonlinear Attention Expansion for Stable and Inheritable Transformer Scaling

    cs.LG 2026-04 unverdicted novelty 5.0

    Nexusformer uses a three-stage nonlinear mapping in attention to enable stable, inheritable scaling of transformers, matching baseline perplexity with up to 41.5% less compute when growing from 240M to 440M parameters.

  5. UNIVID: Unified Vision-Language Model for Video Moderation

    cs.MM 2026-06 unverdicted novelty 4.0

    UNIVID generates policy-aware captions for video moderation, reducing violation leakage by 42.7% and overkill rate by 37.0% while replacing over 1,000 policy-specific models with a single backbone.