Asymmetric Low-Rank Matrix Completion

• Asymmetric low-rank matrix completion is the process of recovering a rectangular, rank‐r matrix from partial observations using a factorization M = XYᵀ with distinct row and column factors.
• It leverages spectral initialization and nonconvex gradient descent methods to detect latent rank and achieve linear convergence under various regimes, including noiseless and heavy-tailed noise settings.
• Robust approaches incorporate adaptive Huber loss and leave-one-out analysis to enable reliable recovery and rank inference even with asymmetric, heavy-tailed noise.

Asymmetric low-rank matrix completion studies the recovery of a rectangular rank-$r$ matrix from a subset of its entries, typically through a factorization $M = XY^\top$ with $X$ and $Y$ living in different ambient dimensions. In the cited literature, the topic includes at least three closely related regimes: noiseless completion of a rectangular matrix by unconstrained nonconvex optimization and spectral initialization (Zhang et al., 13 Aug 2025); rank detection and warm-start construction from very sparse observations through the Bethe Hessian in a random matrix setting (Saade et al., 2015); and recovery under highly incomplete observations corrupted by heavy-tailed and possibly asymmetric additive noise via adaptive Huber loss and a balanced Burer–Monteiro factorization (Wang et al., 2022). Together these works frame asymmetric matrix completion as a problem of identifiability, initialization, nonconvex optimization, and robustness.

1. Formal models and problem structure

A standard formulation seeks to recover a ground-truth rank-$r$ matrix $M_\star\in\mathbb{R}^{d_1\times d_2}$ from a subset $\Omega\subset[d_1]\times[d_2]$ of its entries. Writing $X\in\mathbb{R}^{d_1\times r}$ and $Y\in\mathbb{R}^{d_2\times r}$ as low-rank factors, and $P_\Omega$ for the sampling operator defined by

$M = XY^\top$0

the unregularized nonconvex objective is

$M = XY^\top$1

In this setting, asymmetry is realized by the rectangular factorization itself, with distinct row and column factors and no symmetry constraint on $M = XY^\top$2 (Zhang et al., 13 Aug 2025).

A second formulation introduces additive noise on the observed entries. In the “incomplete-and-noisy” model, there is an unknown rank-$M = XY^\top$3 matrix $M = XY^\top$4 and one observes

$M = XY^\top$5

where the errors need only satisfy $M = XY^\top$6 and $M = XY^\top$7, and may be asymmetric with only finite second moment. This setting is explicitly designed to handle heavy-tailed contamination without sub-Gaussian or symmetry assumptions (Wang et al., 2022).

A third formulation emphasizes sparse observation and rank detectability. In the random matrix setting, $M = XY^\top$8 has rank $M = XY^\top$9 and factorization

$X$0

with $X$1 and $X$2, and one observes a subset $X$3 chosen uniformly at random without replacement, with $X$4. The asymptotic regime is $X$5 with $X$6 and fixed finite $X$7, and there is no observation noise (Saade et al., 2015).

Setting	Observation model	Representative method
Noiseless rectangular completion	$X$8	Vanilla GD with spectral initialization
Heavy-tailed noisy completion	$X$9 on $Y$0	Adaptive Huber + balanced Burer–Monteiro
Sparse random-matrix regime	$Y$1, no noise	Bethe Hessian / MaCBetH

These formulations share the same low-rank factorization template, but they differ in what must be proved: exact or near-exact reconstruction, rank detectability, or robustness under non-sub-Gaussian noise. This suggests that asymmetric matrix completion is best viewed as a family of related inverse problems rather than a single algorithmic template.

2. Spectral detectability and initialization mechanisms

Spectral initialization is central across the cited works, but it serves two distinct purposes: producing a basin-of-attraction point for nonconvex optimization, and inferring the latent rank itself.

For vanilla gradient descent in the noiseless asymmetric setting, the initialization is the rank-$Y$2 truncated singular value decomposition of the rescaled sampled matrix

$Y$3

followed by

$Y$4

Under the stated sampling assumptions, one shows with high probability that

$Y$5

together with a row-wise bound for each index $Y$6 (Zhang et al., 13 Aug 2025).

In the sparse random-matrix regime, spectral initialization is coupled to rank estimation through the Bethe Hessian. One builds a weighted bipartite graph $Y$7 with $Y$8 and weighted adjacency matrix

$Y$9

For inverse temperature $r$0, the Bethe Hessian is

$r$1

At the spin-glass threshold $r$2, the number of negative eigenvalues of $r$3 is exactly $r$4, and the corresponding eigenvectors provide approximations to the true row and column factors. Writing the negative-eigenvalue eigenvectors as

$r$5

one gathers

$r$6

and uses them as warm starts for local minimization of

$r$7

where $r$8 is optional and a quasi-Newton solver such as L-BFGS is run from $r$9 (Saade et al., 2015).

The same work derives a detectability threshold from the functions $M_\star\in\mathbb{R}^{d_1\times d_2}$0 and $M_\star\in\mathbb{R}^{d_1\times d_2}$1 defined through population expectations. The spin-glass instability occurs at $M_\star\in\mathbb{R}^{d_1\times d_2}$2, the retrieval instability at $M_\star\in\mathbb{R}^{d_1\times d_2}$3, and a retrieval phase exists iff $M_\star\in\mathbb{R}^{d_1\times d_2}$4, equivalently for $M_\star\in\mathbb{R}^{d_1\times d_2}$5. Numerically one finds

$M_\star\in\mathbb{R}^{d_1\times d_2}$6

so that reliable rank inference is possible once

$M_\star\in\mathbb{R}^{d_1\times d_2}$7

This separates the question of rank detectability from that of final reconstruction accuracy (Saade et al., 2015).

3. Unregularized nonconvex optimization and implicit balancing

The 2025 analysis of asymmetric matrix completion focuses on vanilla gradient descent (VGD) with no explicit regularization term. Given $M_\star\in\mathbb{R}^{d_1\times d_2}$8, the gradients are

$M_\star\in\mathbb{R}^{d_1\times d_2}$9

$\Omega\subset[d_1]\times[d_2]$0

and the updates are

$\Omega\subset[d_1]\times[d_2]$1

$\Omega\subset[d_1]\times[d_2]$2

The admissible step size is of order $\Omega\subset[d_1]\times[d_2]$3, for example

$\Omega\subset[d_1]\times[d_2]$4

The main theorem states that, with probability at least $\Omega\subset[d_1]\times[d_2]$5, for all $\Omega\subset[d_1]\times[d_2]$6,

$\Omega\subset[d_1]\times[d_2]$7

where

$\Omega\subset[d_1]\times[d_2]$8

and

$\Omega\subset[d_1]\times[d_2]$9

This is a global linear convergence result for an unregularized algorithm (Zhang et al., 13 Aug 2025).

A central point is that previous gradient descent approaches typically incorporate regularization terms to guarantee convergence, but numerical experiments and theoretical analysis of the gradient flow demonstrate that eliminating regularization does not adversely affect convergence performance. In the proof, the balancing quantity

$X\in\mathbb{R}^{d_1\times r}$0

stays $X\in\mathbb{R}^{d_1\times r}$1 during the iterations. The paper interprets this as an implicit regularization property of gradient descent: no explicit balancing term is needed, yet the trajectory remains close to balanced (Zhang et al., 13 Aug 2025).

The empirical comparisons are consistent with the theory. VGD and balancing-GD have essentially identical linear rates, while regularized-GD with an $X\in\mathbb{R}^{d_1\times r}$2 penalty converges to a plateau unless the penalty $X\in\mathbb{R}^{d_1\times r}$3. The $X\in\mathbb{R}^{d_1\times r}$4 success contour in the $X\in\mathbb{R}^{d_1\times r}$5-plane coincides for VGD, BGD, and RGD with $X\in\mathbb{R}^{d_1\times r}$6. On the setting $X\in\mathbb{R}^{d_1\times r}$7, VGD took $X\in\mathbb{R}^{d_1\times r}$8 to reach relative error $X\in\mathbb{R}^{d_1\times r}$9, versus $Y\in\mathbb{R}^{d_2\times r}$0 for BGD and $Y\in\mathbb{R}^{d_2\times r}$1 for RGD with $Y\in\mathbb{R}^{d_2\times r}$2; for $Y\in\mathbb{R}^{d_2\times r}$3, VGD remains fastest (Zhang et al., 13 Aug 2025).

4. Robust completion under heavy-tailed and asymmetric noise

When the observed entries are contaminated by heavy-tailed and possibly asymmetric noise, the squared loss is replaced by the adaptive Huber loss

$Y\in\mathbb{R}^{d_2\times r}$4

For typical noise with $Y\in\mathbb{R}^{d_2\times r}$5, the loss remains quadratic; for extreme outliers $Y\in\mathbb{R}^{d_2\times r}$6, the penalty becomes linear and bounds the influence. At the same time, the bias terms $Y\in\mathbb{R}^{d_2\times r}$7 and $Y\in\mathbb{R}^{d_2\times r}$8 remain $Y\in\mathbb{R}^{d_2\times r}$9 via Markov’s inequality. In particular, one sets $P_\Omega$0 so that $P_\Omega$1 but $P_\Omega$2, ensuring

$P_\Omega$3

The informal theorem summarizes the adaptive choice as $P_\Omega$4 (Wang et al., 2022).

The estimator is written in factorized form $P_\Omega$5, $P_\Omega$6, with augmented Huber objective

$P_\Omega$7

The last term enforces a balanced factorization $P_\Omega$8, which is crucial for local strong convexity. Writing

$P_\Omega$9

the gradients are

$M = XY^\top$00

$M = XY^\top$01

Projected gradient descent is then run with step size $M = XY^\top$02:

$M = XY^\top$03

This is a nonconvex algorithm based on balanced low-rank Burer–Monteiro factorization (Wang et al., 2022).

Initialization is also made robust. One forms the Huberized and re-scaled matrix $M = XY^\top$04 with entries

$M = XY^\top$05

where $M = XY^\top$06. One then computes the top-$M = XY^\top$07 SVD $M = XY^\top$08 and sets

$M = XY^\top$09

followed by orthogonal alignment to the true factors. This yields $M = XY^\top$10 close to $M = XY^\top$11 up to $M = XY^\top$12 (Wang et al., 2022).

Under bounded second moment noise, rank-$M = XY^\top$13 structure, incoherence $M = XY^\top$14, singular values in $M = XY^\top$15, and Bernoulli sampling with

$M = XY^\top$16

the paper proves that with probability at least $M = XY^\top$17: the spectral initialization obeys

$M = XY^\top$18

and after $M = XY^\top$19 iterations,

$M = XY^\top$20

Up to logarithmic factors, this matches the minimax lower bound under sub-Gaussian noise even though the analysis assumes only $M = XY^\top$21 and allows asymmetry (Wang et al., 2022).

5. Leave-one-out analysis and the geometry of convergence proofs

A striking commonality between recent analyses is the use of leave-one-out arguments to control the factorized iterates at the level of individual rows or entries.

In the robust heavy-tailed setting, the core of the proof is a leave-one-entry-out induction. For each row or column index $M = XY^\top$22, one defines a “ghost” trajectory $M = XY^\top$23 by re-running gradient descent with the $M = XY^\top$24-th data row or column removed. Because $M = XY^\top$25 is independent of the $M = XY^\top$26-th observations, one can apply matrix-Bernstein bounds to the $M = XY^\top$27-th gradient entry. Pairing the true and ghost iterates gives the recursion

$M = XY^\top$28

with a corresponding initialization bound. Taking a union bound over $M = XY^\top$29 and running $M = XY^\top$30 iterations yields geometric decay of the per-coordinate error down to an $M = XY^\top$31 floor (Wang et al., 2022).

In the analysis of vanilla gradient descent, the leave-one-out construction is slightly different. For each row or column index $M = XY^\top$32, one defines a leave-$M = XY^\top$33-out problem in which the $M = XY^\top$34-th row is fully observed, removing the corresponding randomness, and then runs gradient descent on that modified objective, with a balancing term added only for analysis. The proof propagates several induction statements: small operator-norm error for $M = XY^\top$35, row-wise closeness for $M = XY^\top$36, small $M = XY^\top$37, linear convergence up to iteration $M = XY^\top$38, and small deviation of the optimal alignment $M = XY^\top$39 from $M = XY^\top$40. The update is decomposed into a population gradient step plus a perturbation; the population Hessian is bounded between $M = XY^\top$41 and $M = XY^\top$42, while the perturbation is controlled by RIP-type lemmas and concentration (Zhang et al., 13 Aug 2025).

This recurring proof architecture suggests that leave-one-out analysis has become a principal device for establishing local contractivity and incoherence control in asymmetric factorized models. The suggestion is interpretive, but it is directly motivated by the central role the method plays in both convergence theories.

6. Detectability, recovery regimes, and recurrent misconceptions

The literature distinguishes several regimes that are often conflated. In the Bethe-Hessian analysis, the phase diagram contains an undetectable phase, a detectable but irrecoverable phase, and a perfect-recovery phase for sufficiently large $M = XY^\top$43. Below $M = XY^\top$44, no negative mode appears in $M = XY^\top$45 until $M = XY^\top$46 exceeds the spurious-mode threshold, so rank detection is impossible; above $M = XY^\top$47 one obtains exactly $M = XY^\top$48 negative modes at $M = XY^\top$49. This means that detecting the latent rank and achieving small reconstruction error are related but not identical tasks (Saade et al., 2015).

A second recurring misconception is that explicit regularization is indispensable for gradient-based asymmetric completion. The 2025 convergence theory states the opposite for the noiseless setting studied there: vanilla gradient descent with spectral initialization converges linearly with high probability, and the balancing term remains small throughout the trajectory as an implicit regularization effect (Zhang et al., 13 Aug 2025).

A third misconception is that meaningful theory requires sub-Gaussian or symmetric noise. The robust completion results show that one can work under merely bounded second moments, with heavy-tailed and possibly asymmetric noise, and still obtain geometric decay to a minimax-optimal statistical error up to logarithmic factors by combining adaptive Huberization, robust spectral initialization, balanced factorization, and leave-one-out analysis (Wang et al., 2022).

Empirical evidence in the cited works is aligned with these distinctions. MaCBetH compares favorably to truncated-SVD initialization (OptSpace) and random initialization, and achieves high-probability perfect recovery with RMSE $M = XY^\top$50 at $M = XY^\top$51 barely above $M = XY^\top$52, whereas OptSpace requires a substantially larger $M = XY^\top$53 (Saade et al., 2015). Vanilla GD avoids extra gradient terms and hyperparameter tuning while maintaining comparable completion performance and lower computational cost relative to balancing-GD and regularized-GD (Zhang et al., 13 Aug 2025). In the heavy-tailed setting, the theoretical conclusions are corroborated by simulation studies (Wang et al., 2022).

Taken together, these results portray asymmetric low-rank matrix completion as a domain in which spectral structure, factorized nonconvex geometry, and robustness mechanisms can be analyzed in a unified way. A plausible implication is that the main technical frontier is no longer merely whether factorized methods work, but under which sampling, initialization, and noise conditions they retain fast convergence and statistically sharp error guarantees.