Transfer Learning for Matrix Completion

Dali Liu; Haolei Weng

REVIEW 3 major objections 7 minor 12 references

Transfer Learning for Matrix Completion

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Pooling source data with a debiasing step makes matrix completion minimax optimal and removes the logarithmic factor from the convergence rate.

desk verdict Solid transfer-learning paper with a genuine, probably repairable gap in the main upper-bound proof. read the letter →

arxiv 2507.02248 v1 pith:4RHBS3FO submitted 2025-07-03 stat.ML cs.LG

Dali Liu , Haolei Weng This is my paper

classification stat.MLcs.LG MSC 62J0762C2062H12

keywords transferlearningmatrixcompletionminimaxoptimalitynuclearnormpenalizationpoolinganddebiasingsourceselectionsharpconcentrationinequalitieslog-freerates

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 claims that transfer learning can make noisy matrix completion minimax optimal without the logarithmic-dimensional factor that appears in single-task rates. The proposed estimator, TransMC, pools target and source samples in a first nuclear-norm penalized regression and then debiases with a second such regression, exploiting the assumption that source matrices are close to the target in nuclear norm. Under sampling-regularity conditions, its average squared error is of order $\frac{\mu^2\max(a^2,v^2) r M}{N} + \frac{h^2}{m_1 m_2}$, and a matching lower bound says no estimator can do better up to constants. The result matters because it characterizes when borrowing source data improves low-rank recovery and comes with a selection rule, S-TransMC, that identifies informative sources when their usefulness is unknown.

What carries the argument

The mechanism is a two-step pool-and-debias construction. Step one solves a nuclear-norm penalized least-squares problem on all $N$ samples; its population optimum is the weighted average $\bar A=\sum_{k=0}^K \alpha_k A_k$, and because $\|\bar A - A_0\|_*\le h$, the pooled estimate is nearly low rank and close to $\bar A$. Step two solves a second nuclear-norm penalized problem on target data for the correction $D=A_0-\bar A$, debiasing the pooled estimate. The proof is carried by sharp spectral-norm concentration bounds for sums of independent random matrices, which control the noise term and the empirical-process term without paying a $\log d$ factor. The sampling-heterogeneity assumption (8), $\|\sum_{k=0}^K \alpha_k P^{(k)}\odot(A_k-\bar A)\|\le L_4\sqrt{a^2/(Nm)}$, keeps the pooling bias small.

What would settle it

Construct a target matrix and source matrices all within nuclear-norm distance $h$, but let every source dataset sample from a different disjoint block of entries while the target samples uniformly. If the left side of condition (8) exceeds $L_4\sqrt{a^2/(Nm)}$ and the empirical error of TransMC still respects the bound (13), condition (8) is not necessary; if the error exceeds it, the condition is what held the rate up.

Watch

Extended reading notes

Core claim

The central claim is Theorem 4.1: under Assumptions 1–5, the TransMC estimator satisfies $\|\hat A_T - A_0\|_F^2/(m_1 m_2) \lesssim \mu^2\max(a^2,v^2) r M/N + C_h \wedge h^2/(m_1 m_2)$, with $C_h=\sqrt{h^2/(m_1 m_2)}\,\mu\sqrt{\max(a^2,v^2)M/n_0}$, with probability at least $1-1/d$. Theorem 4.2 proves a matching minimax lower bound by Fano's method, so the rate is optimal up to constants. The standard $\log d$ factor is removed through sharp concentration inequalities for sums of independent random matrices, and the removal is essential: the lower bound cannot be matched if a logarithmic term remains. When the source matrices are close enough that $h^2/(m_1 m_2)\lesssim r^2 M/n_0$ and the pooled sample size $N$ dominates $n_0$, TransMC beats the single-target minimax rate $rM/n_0$.

Load-bearing premise

The load-bearing premise is condition (8): the combined effect of different sampling patterns and source-target differences must be no larger than the noise floor, and if the sampling distributions differ substantially across datasets it can fail, so the claimed rate no longer follows.

Editorial extensions

If this is right

When $\mu,a,v$ are constants and $h^2/(m_1m_2)\ll r^2M/n_0$ with $N\gg n_0$, TransMC strictly improves on the single-target rate $rM/n_0$.
The rate separates into an oracle term $\mu^2\max(a^2,v^2)rM/N$ from pooled samples and a bias term $C_h\wedge h^2/(m_1m_2)$ from source-target dissimilarity; when all sources equal the target ($h=0$), only the oracle term remains.
The matching lower bound shows the logarithmic-factor-free rate is the true minimax rate in this transfer setting, not a proof artifact.
S-TransMC selects exactly the informative sources with probability at least $1-K/d-K/n_0$ under Assumptions 6–7, so the practitioner need not know a priori which sources are useful.

Reading between the lines

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

Extending the paper's logic, the same concentration machinery should yield log-free minimax rates for collective matrix completion and dynamic matrix recovery, where multiple related matrices are observed; the paper does not state this application.
A practical diagnostic suggested by the theory is to estimate the left-hand side of condition (8) from sampling histograms; if it is large, weighting or discarding sources before pooling may restore the rate, an adaptive step the paper leaves open.
The lower bound effectively says the usefulness of a source is governed primarily by its nuclear-norm gap $h$; in applications like the TEC maps, choosing temporally close sources should therefore be the main practical lever.
Since condition (8) holds automatically only when all sampling distributions coincide, a natural test is to compare per-source row and column sampling frequencies and flag sources that deviate strongly; the paper does not propose such a test.

Share X Bluesky LinkedIn Reddit HN

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

The key thing to know: this is a well-designed transfer matrix completion paper, but the central upper bound has a gap that is probably fixable. The contribution—pooling plus debiasing under nuclear-norm closeness, with a matching minimax lower bound and a source-selection rule—is real and timely. The sharp concentration framework from Brailovskaya and van Handel is the right tool, and removing the log d factor is a genuine step beyond the single-task rate.\n\nThe core claim is Theorem 4.1: TransMC achieves error roughly rM/N + min(Ch, h²/(m1m2)). The lower bound in Theorem 4.2 matches it in the regimes that matter, and I found no issue with the Fano constructions. The simulations and the TEC application are standard but honest.\n\nThe soft spot is in §8.1.2, right after Eq (32). The proof drops the term h a µ sqrt(M/n0)/(m1m2) (up to their notation) without bounding it. The stress-test note and I independently landed on the same point. There is also an inequality in that paragraph that runs the wrong way: when n0 ≤ N, which is the usual transfer setting, sqrt(h² M/n0) is larger than sqrt(h² M/N), not smaller. These are genuine proof errors in the main theorem, not cosmetic typos. That said, they look repairable: the dropped term is absorbed by h²/(m1m2) when h is large, and by the rM/N term when h is small, using condition (12) and m ≥ r log d. The repair case analysis is simply absent from the paper.\n\nTwo smaller things. Assumption 5(8) is strong: it controls the combined effect of heterogeneous sampling and source-target deviations, and the paper only states it holds automatically when all sampling distributions are identical. That is a real limitation, but it is explicitly assumed, so it is not a hidden flaw. Also, the proof mislabels condition (12) as condition (8) in the paragraph following Eq (32), which will confuse careful readers.\n\nBottom line: the paper deserves a serious referee, and I would send it out with a request to repair the Theorem 4.1 proof and to restate Assumption 5 more carefully. If the repair goes through, I would cite it. For now, I would not rely on the central theorem as written.

Referee Report

3 major / 7 minor

Summary. The paper studies transfer learning for noisy matrix completion. It proposes TransMC, a two-step estimator that first pools all source and target data in a nuclear-norm penalized regression and then debiases the pooled estimate using the target data alone. Under assumptions that the source matrices are close to the target matrix in nuclear norm and that sampling distributions are sufficiently regular, Theorem 4.1 gives a high-probability Frobenius error bound with no logarithmic factor, and Theorem 4.2 gives a matching minimax lower bound. When the set of informative sources is unknown, the paper proposes S-TransMC, a cross-validation-based source selection procedure, and proves selection consistency in Theorem 4.3. Simulations and a total electron content (TEC) data application illustrate the empirical behavior of both algorithms.

Significance. If Theorem 4.1 and Theorem 4.2 are correct, the paper makes a valuable contribution: it establishes minimax-optimal transfer learning for matrix completion, removes the logarithmic dimensional factor that typically appears in matrix completion rates, and provides a principled source-selection procedure with consistency guarantees. The proof strategy is well matched to the claimed results: it builds on the sharp concentration inequalities of Brailovskaya and van Handel (2024), supplies explicit debiasing and pooling analyses, and verifies the lower bound by a Fano-method construction. The numerical studies and the TEC application are consistent with the theoretical rates and add credibility to the practical claims. The main issue is a gap in the proof of the central upper bound, which is identified below; the core idea appears repairable, but the paper as written does not fully establish the headline rate.

major comments (3)

[§8.1.2, Eq. (32)] The proof of Theorem 4.1 drops the term (h√(m1m2))·√(µ²a²/(n0m)) = h a µ √(M/n0) without bounding it. This term is not dominated by either Ch = sqrt(h²/(m1m2)·µ² max(a²,v²)M/n0) or h²/(m1m2) under the stated conditions; in fact it is a factor √(m1m2) larger than Ch. The subsequent sentence 'under ... (32) can be written as' therefore does not follow from the displayed inequality. A repair is plausible by splitting on the relative sizes of h and aµ√(M/n0), but the argument is absent and must be written out before the upper bound can be accepted.
[§8.1.2, paragraph after Eq. (32)] In the same paragraph the inequality sqrt(h²/(m1m2)·µ² max(a²,v²)M/n0) ≤ sqrt(h²/(m1m2)·µ² max(a²,v²)M/N) is stated, but whenever n0 ≤ N, which is the typical transfer regime, the inequality is reversed. The final displayed rate in the paragraph appears to use the correct direction (1/n0 ≥ 1/N), so this seems to be a local slip; nevertheless the direction must be corrected because it affects the simplification of the central bound.
[Assumption 5, condition (8)] Condition (8) is a substantive restriction: it controls the combined effect of heterogeneous sampling distributions and source-target deviations, and it is used directly in Lemma 8.2 to bound the pooling bias. The paper only verifies that it holds automatically when all sampling distributions are identical. Since Theorem 4.1 and the minimax claim are predicated on this condition, the scope of the result is narrower than the general non-uniform sampling setup of Assumption 3 may suggest. The authors should clarify this limitation in the main text and, ideally, provide a nontrivial sufficient condition or a concrete example with heterogeneous P^(k) under which (8) is satisfied.

minor comments (7)

[Abstract] The phrase 'out method' should be 'our method'.
[Theorem 4.1 statement] The sentence 'Assume that and N satisfies (10)' has a grammatical error; it should be 'Assume that N satisfies (10)'.
[§8.1.2, after Eq. (32)] The condition n0/N ≥ a² log d / (max(a²,v²) rM) is cited as 'condition (8)', but it is condition (12) in the statement of Theorem 4.1; condition (8) is the first inequality of Assumption 5.
[Section 6] The estimator name is misspelled as 'TrasnMC'; it should be 'TransMC'.
[Lemma 8.2 proof and Section 5.1] There are minor typos: 'Cathy-Schwartz' should be 'Cauchy-Schwarz', and 'aross' should be 'across'.
[Assumption 6] Assumption 6 introduces a constant '˜C1' but the selection threshold in Algorithm 2 and the proof use '˜C0'; the notation should be aligned.
[Proof of Lemma 8.4] After showing that the failure probability is at most 1/d, the text says 'with probability at least 1/d'; this should be 'with probability at least 1 − 1/d'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the minimax upper and lower bounds are derived from the stated assumptions and external concentration inequalities; the only self-citations are contextual and not load-bearing.

full rationale

The paper's central upper bound (Theorem 4.1) is obtained by a self-contained two-step analysis of the pooling estimator (6) and the debiasing estimator (7) in Sections 8.1.1-8.1.2. The stochastic terms are controlled via Lemmas 8.1, 8.2, and 8.4, where Lemma 8.1 applies the external sharp concentration inequalities of Brailovskaya and van Handel (2024) and Lemma 8.2 uses Assumption 5(8) exactly as an explicit modeling condition. The dependence on h, N, and n0 arises from algebraic bounds on the bias terms, not from a quantity fitted to the target data. The minimax lower bound (Theorem 4.2) is proved with Fano's inequality and the external packing lemma of Negahban and Wainwright (2012); the shared definition of Ch in the upper and lower bounds reflects the same parameter space Omega_h, not a fitted value. The self-citations are to Liu and Weng (2024) for the single-task baseline rate and to He et al. (2024) in related-work discussion; neither is used to prove the paper's theorems. Assumption 5(8) may be restrictive, but it is a stated assumption rather than a prediction renamed as a result. A possible proof gap in the simplification after Eq. (32) would be a correctness issue, not circularity. No load-bearing step reduces by construction to its own input, and no fitted parameter is relabeled as a prediction.

Assumptions & free parameters 0 free parameters · 9 assumptions · 0 invented entities

All statistical assumptions are stated explicitly (Assumptions 1-7). The external concentration inequalities are cited and not re-derived. No new physical or mathematical entities are introduced; no parameters are fitted to data in the theory. The proof gap discussed in soundness is separate from the ledger.

assumptions (9)

domain assumption Assumption 1: ∥Ak − A0∥∗ ≤ h for all k
Defines informative sources by nuclear-norm closeness; the central rate depends on h.
domain assumption Assumption 2: rank(A0) ≤ r and ∥Ak∥∞ ≤ a
Standard low-rank and bounded-entry conditions needed for the nuclear norm penalty.
domain assumption Assumption 3: sampling distributions satisfy lower/upper bounds and no concentration
Conditions (1)-(3) on P^(k)_jl are used in the empirical-process and concentration lemmas.
domain assumption Assumption 4: sub-gaussian noise with variance v²
Used in Lemmas 8.1 and the selection proof.
domain assumption Assumption 5: sampling heterogeneity bounds (8) and (9)
Load-bearing for the pooling step; ensures the pooled population minimizer is close to the weighted average of source matrices.
domain assumption Assumption 6: separation ∥Ak − A0∥²_F/(m1m2) > 2 C̃0 ε0 for uninformative sources
Makes detection of informative sources possible in Theorem 4.3.
domain assumption Assumption 7: Γk, Υk, Γ0 = o(1)
Small-error conditions for selection consistency; all terms are controlled rates.
standard math Brailovskaya and van Handel (2024) sharp matrix concentration inequalities (Theorem 8.7)
External published result used to remove a logarithmic factor in Lemmas 8.1 and 8.2.
standard math Lemma 8.5 (Negahban and Wainwright 2012) construction of δ-separated rank-r matrices
Used in the minimax lower bound via Fano's method.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Transfer Learning for Matrix Completion." pith.science (2026). https://pith.science/paper/4RHBS3FO

@misc{pith2026250702248,
  author       = {Pith},
  title        = {Pith review of: Transfer Learning for Matrix Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4RHBS3FO}},
  note         = {Machine review of arXiv:2507.02248}
}

read the original abstract

In this paper, we explore the knowledge transfer under the setting of matrix completion, which aims to enhance the estimation of a low-rank target matrix with auxiliary data available. We propose a transfer learning procedure given prior information on which source datasets are favorable. We study its convergence rates and prove its minimax optimality. Our analysis reveals that with the source matrices close enough to the target matrix, out method outperforms the traditional method using the single target data. In particular, we leverage the advanced sharp concentration inequalities introduced in \cite{brailovskaya2024universality} to eliminate a logarithmic factor in the convergence rate, which is crucial for proving the minimax optimality. When the relevance of source datasets is unknown, we develop an efficient detection procedure to identify informative sources and establish its selection consistency. Simulations and real data analysis are conducted to support the validity of our methodology.

Figures

Figures reproduced from arXiv: 2507.02248 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 9 canonical work pages

[3]

E. J. Cand` es and T. Tao. The power of convex relaxation: Near-optimal matrix completion. IEEE transactions on information theory , 56(5):2053–2080,

work page
[4]

K. Choi, G. Fazekas, M. Sandler, and K. Cho. Transfer learning for music classification and regression tasks. arXiv preprint arXiv:1703.09179 ,

work page Pith review arXiv
[7]

J. Liu, M. Shah, B. Kuipers, and S. Savarese. Cross-view action recognition via view knowledge transfer. In CVPR 2011 , pages 3209–3216. IEEE,

work page 2011
[9]

C. Tan, F. Sun, T. Kong, W. Zhang, C. Yang, and C. Liu. A survey on deep transfer learning. In Artificial Neural Networks and Machine Learning–ICANN 2018: 27th International Conference on Artificial Neural Networks, Rhodes, Greece, October 4-7, 2018, Proceedings, Part III 27 , pages 270–279. Springer,

work page 2018
[2010]

T. T. Cai and H. Pu. Transfer learning for nonparametric regression: Non-asymptotic minimax analysis and adaptive procedure. arXiv preprint arXiv:2401.12272 ,

arXiv
[2012]

Yan and S

H. Yan and S. X. Chen. Transfer learning with general estimating equations. arXiv preprint arXiv:2410.04398,

arXiv
[2013]

Optimal Transfer Learning for Missing Not-at-Random Matrix Completion

A. Jalan, Y. Jedra, A. Mazumdar, S. S. Mukherjee, and P. Sarkar. Optimal transfer learning for missing not-at-random matrix completion. arXiv preprint arXiv:2503.00174 ,

work page Pith review arXiv
[2016]

Weng and X

Z. Weng and X. Wang. Low-rank matrix completion for array signal processing. In 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 2697–2700. IEEE,

work page 2012

Show all 12 references

[2020]

S. Zou, A. J. Ridley, M. B. Moldwin, M. J. Nicolls, A. J. Coster, E. G. Thomas, and J. M. Ruo- honiemi. Multi-instrument observations of sed during 24–25 october 2011 storm: Implications for sed formation processes. Journal of Geophysical Research: Space Physics, 118(12):7798–7809,

2011
[2021]

Y. He, Z. Li, D. Liu, K. Qin, and J. Xie. Representational transfer learning for matrix completion. arXiv preprint arXiv:2412.06233 ,

arXiv
[2022]

McGrath, C

S. McGrath, C. Zhu, M. Guo, and R. Duan. Learner: A transfer learning method for low-rank matrix estimation. arXiv preprint arXiv:2412.20605 ,

arXiv
[2024]

C. Cai, T. T. Cai, and H. Li. Transfer learning for contextual multi-armed bandits. The Annals of Statistics, 52(1):207–232, 2024a. J.-F. Cai, E. J. Cand` es, and Z. Shen. A singular value thresholding algorithm for matrix completion. SIAM Journal on optimization , 20(4):1956–1982,

1956

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [3]

E. J. Cand` es and T. Tao. The power of convex relaxation: Near-optimal matrix completion. IEEE transactions on information theory , 56(5):2053–2080,

work page

[2] [4]

K. Choi, G. Fazekas, M. Sandler, and K. Cho. Transfer learning for music classification and regression tasks. arXiv preprint arXiv:1703.09179 ,

work page Pith review arXiv

[3] [7]

J. Liu, M. Shah, B. Kuipers, and S. Savarese. Cross-view action recognition via view knowledge transfer. In CVPR 2011 , pages 3209–3216. IEEE,

work page 2011

[4] [9]

C. Tan, F. Sun, T. Kong, W. Zhang, C. Yang, and C. Liu. A survey on deep transfer learning. In Artificial Neural Networks and Machine Learning–ICANN 2018: 27th International Conference on Artificial Neural Networks, Rhodes, Greece, October 4-7, 2018, Proceedings, Part III 27 , pages 270–279. Springer,

work page 2018

[5] [2010]

T. T. Cai and H. Pu. Transfer learning for nonparametric regression: Non-asymptotic minimax analysis and adaptive procedure. arXiv preprint arXiv:2401.12272 ,

arXiv

[6] [2012]

Yan and S

H. Yan and S. X. Chen. Transfer learning with general estimating equations. arXiv preprint arXiv:2410.04398,

arXiv

[7] [2013]

Optimal Transfer Learning for Missing Not-at-Random Matrix Completion

A. Jalan, Y. Jedra, A. Mazumdar, S. S. Mukherjee, and P. Sarkar. Optimal transfer learning for missing not-at-random matrix completion. arXiv preprint arXiv:2503.00174 ,

work page Pith review arXiv

[8] [2016]

Weng and X

Z. Weng and X. Wang. Low-rank matrix completion for array signal processing. In 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 2697–2700. IEEE,

work page 2012