Static Pruning in Dense Retrieval using Matrix Decomposition

Fabrizio Silvestri; Federico Siciliano; Francesca Pezzuti; Nicola Tonellotto

REVIEW 3 major objections 5 minor 21 references

Static Pruning in Dense Retrieval using Matrix Decomposition

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Static PCA pruning can remove more than half of the dimensions from dense retrieval embeddings while keeping ranking quality almost intact.

desk verdict Simple offline PCA pruning mostly works, but the paper's central 'under 5% loss' claim is contradicted by its own out-of-domain results. read the letter →

arxiv 2412.09983 v1 pith:65XNJQF7 submitted 2024-12-13 cs.IR

Federico Siciliano , Francesca Pezzuti , Nicola Tonellotto , Fabrizio Silvestri This is my paper · ORCID

classification cs.IR

keywords denseretrievalstaticpruningprincipalcomponentanalysisdimensionalityreductionembeddingquery-independentmatrixdecompositionefficiency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the dimensions of dense document embeddings can be pruned statically, before any query arrives, using Principal Components Analysis, and that this pruning needs no query-specific information. The authors show that dropping more than half of the embedding dimensions leaves ranking quality almost intact, with at most a 5% reduction in NDCG@10 across the dense retrieval models and query sets they test. If true, this gives a simple offline way to halve the storage footprint of an embedding index and to cut the dominant per-query dot-product cost, complementing earlier pruning methods that remove whole documents or that decide which dimensions to keep only after seeing the query. The practical payoff would be cheaper and faster dense retrieval systems whose effectiveness stays close to the unpruned system.

What carries the argument

The central object is the PCA eigendecomposition of the document embedding matrix, written as $D^T D = W \Lambda W^T$, with the eigenvalues of $\Lambda$ sorted decreasingly. The ordering induced by the eigenvalues is what makes pruning static: the first $m$ columns of $W$ define a projection $W_m$ that is computed offline once and then applied to any query at inference time, lowering the query-time cost from $O(dn)$ to $O(dm + mn)$ and the index size from $O(dn)$ to $O(mn + md)$. The mechanism therefore treats the variance explained by each dimension as a proxy for its contribution to ranking, which is the assumption that carries the whole method.

What would settle it

A concrete check is to run the same 50% pruning experiment while comparing against an oracle that keeps the $m$ dimensions with the highest empirical correlation to relevance judgments; if the oracle substantially outperforms PCA pruning on the same query sets, then variance-based ordering is the wrong criterion. An even sharper test is to construct a synthetic embedding matrix where the top-variance directions are orthogonal to the directions that actually determine relevance and observe a large NDCG@10 drop under PCA pruning.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that projecting the document embedding matrix onto its top principal components is an effective static pruning strategy: the eigendecomposition $D^T D = W \Lambda W^T$ orders the dimensions of the transformed matrix $T = D W$ by variance, and keeping only the first $m$ columns of $T$ yields a reduced index $\hat D = D W_m$. Queries are transformed with the same offline projection, $\hat q = W_m^T q$, so no per-query dimension selection is needed. The empirical claim is that at a 50% pruning cutoff, the best encoder in the tests shows no statistically significant loss in effectiveness, and the other encoders lose at most a few points of NDCG@10; even at a 75% cutoff one encoder remains statistically unchanged. The same transformation transfers to a different corpus with effectiveness on par with in-domain pruning, and computing the PCA from as few as $10^3$ documents is nearly as good as using $10^5$.

Load-bearing premise

The load-bearing premise is that the embedding dimensions with the largest variance, as selected by the PCA eigenvalue ordering, are the dimensions that matter most for ranking; if variance and relevance do not align, cutting the low-variance dimensions will remove useful signal and the reported robustness will not generalize.

Editorial extensions

If this is right

Halving the embedding dimensionality ($c = 0.5$) gives a 2x reduction in embedding index storage with at most a 5% NDCG@10 drop, and for the strongest tested encoder no statistically significant effectiveness loss.
Query processing becomes faster by a factor $O(d/m)$ in the dominant term, because the dot product runs in $m$ dimensions after a cheap $O(dm)$ query transformation.
Because the projection is query-independent and offline, the same precomputed $W_m$ can be applied to new corpora; the out-of-domain experiments show effectiveness on par with in-domain pruning.
Using only $10^3$ documents to estimate the PCA is enough to match $10^5$, so the offline computation can be fit to small samples without observable quality loss.
The paper argues the approach is orthogonal to content-dependent static document pruning, so dimension-level savings can be layered on top of methods that prune whole embeddings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: an obvious untested combination is PCA dimension pruning applied together with document-level static pruning; since the two mechanisms remove different resources, their space savings should compound.
Beyond the paper: the variance-as-relevance assumption could be checked directly by comparing the top-$m$ variance dimensions with the top-$m$ dimensions by empirical correlation to relevance labels; the paper does not run this control.
Beyond the paper: the method is a natural pre-processing step for approximate nearest neighbour indexes, since shorter vectors reduce storage and scan cost, but the paper does not measure end-to-end ANN latency.
Beyond the paper: the out-of-domain transfer results suggest that a single projection matrix per encoder could be published and reused across collections, avoiding per-corpus PCA computation; cross-encoder transfer is not explored.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. First, the idea is clean and genuinely different from the query-dependent pruning in Faggioli et al.: compute a PCA projection offline, keep the top m dimensions of document embeddings, transform queries with the same projection. Second, the headline claim that this costs at most 5% nDCG@10 'across all models and query sets' is false as written. Table 2 shows Contriever out-of-domain at a 50% cutoff dropping 6.27% on DEV SMALL, 5.67% on DL HARD, and 6.02% on COVID. Those are statistically significant per their own test. The 5% bound only holds for the in-domain Table 1.

What's good: the paper runs three encoders over five query sets, reports multiple cutoffs, and runs significance tests. The in-domain results are genuinely useful: pruning half the dimensions keeps TAS-B statistically unchanged and keeps Contriever within about 4.6%. The RQ3 result that 10^3 documents suffice for the transformation is a nice robustness finding.

The soft spots beyond the overstated bound. Efficiency is argued from asymptotic complexity, never measured. For a paper whose motivation is latency and memory, that's a gap. There is no comparison against Acquavia, Chang, or Faggioli, so we don't know whether this beats the existing static/dynamic pruning trade-offs. The out-of-domain protocol is vague: we are not told which corpus provides the transformation matrix for the COVID results. The PCA is applied to D^T D without centering, which is not textbook PCA, though the projection still makes sense. And the multiple significance tests are not corrected.

The central effectiveness claim for in-domain pruning holds up; the generalization claim does not. This deserves a serious referee because the method is simple, reproducible, and likely useful, but the authors need to fix the bound, add latency/memory figures, compare to the cited pruning baselines, and specify the out-of-domain setup. I'd be happy to see it in a workshop or a main IR conference after that.

Referee Report

3 major / 5 minor

Summary. The paper proposes a static, query-independent dimensionality reduction method for dense retrieval document embeddings, based on PCA. The document embedding matrix is projected onto its top principal components offline, queries are transformed with the same projection, and ranking is performed in the reduced space. Experiments with TAS-B, Contriever, and ANCE on MS MARCO and BEIR COVID report AP, MRR@10, and NDCG@10 at 25%, 50%, and 75% pruning cutoffs, for both in-domain and out-of-domain PCA. The central claim is that pruning more than 50% of dimensions costs at most 5% NDCG@10 across all models and query sets, while yielding a 2x space reduction and a corresponding speed-up.

Significance. If the claim is correct, the method is attractive because it is simple, fully offline, and does not require query-dependent computation, unlike prior dynamic dimension-pruning approaches. The experimental setup uses standard collections and metrics, and the paper applies paired significance testing. The robustness study with varying numbers of PCA training documents (RQ3) is a useful addition. However, the headline quantitative bound is contradicted by the paper's own out-of-domain results, and the efficiency claim rests only on asymptotic complexity rather than measured latency or index size. The absence of comparison to alternative dimension-pruning baselines further limits the strength of the contribution.

major comments (3)

[Abstract and Section 1; Table 2] The statement that dimension reduction by over 50% causes 'at most a 5% reduction in nDCG@10 across all models and query sets' is contradicted by Table 2 for out-of-domain PCA at c=50%. For Contriever, the baseline-to-pruned NDCG@10 values are DEV SMALL .4070 to .3815 (6.27% loss), DL HARD .3772 to .3558 (5.67% loss), and COVID .4963 to .4664 (6.02% loss); several of these differences are marked with a dagger for significance. The 5% bound appears to hold only for the in-domain results in Table 1, where the largest loss is about 4.6% (Contriever on DEV SMALL). The authors should either restrict the claim to in-domain PCA or report the actual out-of-domain bound; as written, the paper's central claim is numerically false.
[Section 2 and Section 4] The efficiency benefit is argued only from asymptotic complexity (O(d/m) speed-up and O(m/d) space reduction), but the abstract and introduction claim a 'significant boost in dense retrieval efficiency' and a 2x space efficiency. No wall-clock latency, query throughput, memory footprint, or index-size measurements are reported in the experimental section. Since the motivation of the paper is efficiency, the absence of any direct efficiency measurement leaves the practical impact unverified. The authors should add such measurements or explicitly restrict the claims to theoretical complexity.
[Section 4] The experiments compare only against the full-dimensionality baseline; there is no comparison to alternative dimension-pruning baselines, such as keeping a random subset of dimensions, keeping the first m original dimensions, or adapting the dynamic dimension-importance method of Faggioli et al. [8] to a static setting. Without such baselines, the specific benefit of the PCA-induced ordering over arbitrary dimension pruning is not established, and the reader cannot tell whether the observed effectiveness is due to the eigenvalue ordering or simply to the reduced dimensionality.

minor comments (5)

[Section 4 and Table 1/2 captions] The notation '105' and '103' in the text and figure captions should read 10^5 and 10^3; the superscript formatting is missing.
[References] Reference [1] gives an incomplete venue ('In Proc. Eng. 1–10') and reference [3] uses 'InCoCo@NIPS', which appears to be an incorrect venue abbreviation; these should be corrected.
[Throughout] There are several typos and grammatical slips, including 'efficient' for 'efficient' in Section 1, 'In study' at the start of Section 5, and 'A key advantages' in Section 5; the paper would benefit from a careful proofreading pass.
[Figures 1 and 2] The figure legends state 'Baseline' as if it were a curve, but the baseline is a horizontal reference; this should be clarified in the caption.
[Section 2] The footnote marker '1' appears as a superscript after 'O(n+klogk)' but the footnote text is placed at the bottom of the page; the formatting should be fixed so the marker and footnote are visually linked.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PCA pruning is an unsupervised projection evaluated against external relevance judgments, with no fitted parameter or self-citation chain carrying the central claim.

full rationale

The paper's derivation chain is self-contained and non-circular. The method computes the eigendecomposition of the embedding covariance matrix, keeps the top-m principal components, and transforms queries with the corresponding projection matrix. The cutoff m is a user-controlled experimental variable explored at 25%, 50%, and 75%, not a parameter fitted to the evaluation data. Ranking effectiveness is measured with standard external benchmarks (TREC DL 19, DL 20, DL HARD, MS MARCO Dev Small, BEIR COVID) and baseline comparisons, so the claim that pruning preserves nDCG@10 is an empirical finding rather than an identity. The only conceptual premise, that high-variance dimensions are useful for ranking, is an explicit modeling assumption, not a result derived from the evaluation metric. Self-citations to prior work by co-author Tonellotto ([1], [8]) appear only as related-work context and are not load-bearing for the central result. The discrepancy between the abstract's 'at most 5% reduction' claim and the out-of-domain numbers in Table 2 is a correctness/consistency issue, not circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities. Its free parameters are the pruning cutoff and the PCA sample size, both controlled by the experimenter. The key domain assumption is that variance ordering aligns with ranking importance, which is plausible but unproven.

free parameters (2)

m, number of retained dimensions = Varied as cutoffs 25%, 50%, 75% of original dimensions
The pruning cutoff is chosen by the experimenter, not learned from data. It directly controls the trade-off between efficiency and effectiveness.
N, number of documents used for PCA computation = 10^3, 10^4, or 10^5
The covariance matrix is estimated on a subsample of documents. The paper varies this in RQ3 and reports negligible differences, but the choice is an experimental hyperparameter.

assumptions (4)

standard math Eigendecomposition of the covariance matrix orders dimensions by explained variance.
Invoked in Section 2 when defining T = D W and using the eigenvalue ordering to prune dimensions.
domain assumption The relevance score between a query and a document is the dot product of their embeddings.
Used in Section 2 as the scoring function for dense retrieval, following standard bi-encoder practice.
domain assumption Dimensions with higher variance are more important for ranking than dimensions with lower variance.
This is the load-bearing assumption behind pruning based on PCA eigenvalues. It is stated implicitly in Section 2 and tested only empirically, without comparison to other dimension importance criteria.
domain assumption The query embedding and document embeddings live in the same latent vector space.
Needed for the projection W_m to be applicable to queries, as done in Section 2.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Static Pruning in Dense Retrieval using Matrix Decomposition." pith.science (2026). https://pith.science/paper/65XNJQF7

@misc{pith2026241209983,
  author       = {Pith},
  title        = {Pith review of: Static Pruning in Dense Retrieval using Matrix Decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/65XNJQF7}},
  note         = {Machine review of arXiv:2412.09983}
}

read the original abstract

In the era of dense retrieval, document indexing and retrieval is largely based on encoding models that transform text documents into embeddings. The efficiency of retrieval is directly proportional to the number of documents and the size of the embeddings. Recent studies have shown that it is possible to reduce embedding size without sacrificing - and in some cases improving - the retrieval effectiveness. However, the methods introduced by these studies are query-dependent, so they can't be applied offline and require additional computations during query processing, thus negatively impacting the retrieval efficiency. In this paper, we present a novel static pruning method for reducing the dimensionality of embeddings using Principal Components Analysis. This approach is query-independent and can be executed offline, leading to a significant boost in dense retrieval efficiency with a negligible impact on the system effectiveness. Our experiments show that our proposed method reduces the dimensionality of document representations by over 50% with up to a 5% reduction in NDCG@10, for different dense retrieval models.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 19 canonical work pages

[8]

Guglielmo Faggioli, Nicola Ferro, Raffaele Perego, and Nicola Tonellotto. 2024. Dimension importance estimation for dense information retrieval. In Proc. SIGIR. 1318–1328

work page 2024
[1]

Antonio Acquavia, Craig Macdonald, and Nicola Tonellotto. 2023. Static Pruning for Multi-Representation Dense Retrieval. In Proc. Eng. 1–10

work page 2023
[2]

Negar Arabzadeh, Xinyi Yan, and Charles LA. Clarke. 2021. Predicting efficiency/effectiveness trade-offs for dense vs. sparse retrieval strategy selection. In Proc. CIKM. 2862–2866

work page 2021
[3]

Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNa- mara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Wang Tong. 2016. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. In InCoCo@NIPS

work page 2016
[4]

Xuejun Chang, Debabrata Mishra, Craig Macdonald, and Sean MacAvaney. 2024. Neural Passage Quality Estimation for Static Pruning. In Proc. SIGIR. 174–185

2024
[5]

Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the TREC 2020 deep learning track. In Proc. TREC

work page 2021
[6]

Voorhees

Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M. Voorhees. 2020. Overview of the TREC 2019 deep learning track. In Proc. TREC

work page 2020
[7]

Dumais, George W

Susan T. Dumais, George W. Furnas, Thomas K. Landauer, Scott Deerwester, and Richard Harshman. 1988. Using latent semantic analysis to improve access to textual information. In Proc. SIGCHI. 281–285

work page 1988

Show all 21 references

[9]

Yan Fang, Jingtao Zhan, Qingyao Ai, Jiaxin Mao, Weihang Su, Jia Chen, and Yiqun Liu. 2024. Scaling laws for dense retrieval. In Proc. SIGIR. 1339–1349

2024
[10]

Lin, and Allan Hanbury

Sebastian Hofstätter, Sheng-Chieh Lin, Jheng-Hong Yang, Jimmy J. Lin, and Allan Hanbury. 2021. Efficiently Teaching an Effective Dense Retriever with Balanced Topic Aware Sampling. In Proc. SIGIR. 113–122

2021
[11]

Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised Dense Information Retrieval with Contrastive Learning. Proc. TMLR (2022)

2022
[12]

Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2021. Billion-Scale Similarity Search with GPUs. Trans. Big Data 7, 3 (2021), 535–547

2021
[13]

Yubin Kim. 2022. Applications and future of dense retrieval in industry. In Proc. SIGIR. 3373–3374

2022
[14]

MacAvaney, Sean and Macdonald, Craig and Ounis, Iadh. 2022. Streamlining Evaluation with ir-measures. In Proc. ECIR. 305–310

2022
[15]

Iain Mackie, Jeffery Dalton, and Andrew Yates. 2021. How Deep is your Learning: the DL-HARD Annotated Deep Learning Dataset. In Proc. TREC

2021
[16]

Conrado Martınez. 2004. Partial quicksort. In Proc. 6th ACMSIAM Workshop on Algorithm Engineering and Experiments and 1st ACM-SIAM Workshop on Analytic Algorithmics and Combinatorics . 224–228

2004
[17]

Karl Pearson. 1901. Principal components analysis. The Philosophical Magazine and Journal of Science 6, 2 (1901), 559

1901
[18]

Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models. In Proc. NeurIPS nd Benchmarks Track

2021
[19]

Bennett, Junaid Ahmed, and Arnold Overwijk

Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk
[20]

Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji-Rong Wen. 2024. Dense text retrieval based on pretrained language models: A survey. Trans. TOIS 42, 4 (2024), 1–60

2024
[2021]

Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval. In Proc. ICLR

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [8]

Guglielmo Faggioli, Nicola Ferro, Raffaele Perego, and Nicola Tonellotto. 2024. Dimension importance estimation for dense information retrieval. In Proc. SIGIR. 1318–1328

work page 2024

[2] [1]

Antonio Acquavia, Craig Macdonald, and Nicola Tonellotto. 2023. Static Pruning for Multi-Representation Dense Retrieval. In Proc. Eng. 1–10

work page 2023

[3] [2]

Negar Arabzadeh, Xinyi Yan, and Charles LA. Clarke. 2021. Predicting efficiency/effectiveness trade-offs for dense vs. sparse retrieval strategy selection. In Proc. CIKM. 2862–2866

work page 2021

[4] [3]

Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNa- mara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Wang Tong. 2016. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. In InCoCo@NIPS

work page 2016

[5] [4]

Xuejun Chang, Debabrata Mishra, Craig Macdonald, and Sean MacAvaney. 2024. Neural Passage Quality Estimation for Static Pruning. In Proc. SIGIR. 174–185

2024

[6] [5]

Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the TREC 2020 deep learning track. In Proc. TREC

work page 2021

[7] [6]

Voorhees

Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M. Voorhees. 2020. Overview of the TREC 2019 deep learning track. In Proc. TREC

work page 2020

[8] [7]

Dumais, George W

Susan T. Dumais, George W. Furnas, Thomas K. Landauer, Scott Deerwester, and Richard Harshman. 1988. Using latent semantic analysis to improve access to textual information. In Proc. SIGCHI. 281–285

work page 1988