Missing Data Matrix: Models & Methods

• Missing data matrix is a partially observed matrix with an associated mask that differentiates observed from missing entries in various statistical and random matrix theory formulations.
• It encompasses diverse formulations such as Bernoulli-masked samples, deterministic Hadamard products, and imputation techniques used in covariance and graphical-model estimation.
• Analyses focus on identifiability, regularization, and computational trade-offs, driving advances in high-dimensional inference and efficient matrix recovery methods.

A missing data matrix is a matrix-valued object with unobserved entries and an associated observation pattern, mask, or sampling operator that specifies which entries are available. In the literature, this notion appears in several distinct but related forms: an incomplete data matrix $Y_{\mathrm{obs}} = Y \odot \Omega$ in missing-data statistics, a Bernoulli-masked observation matrix in covariance estimation, and a Hadamard-masked random matrix $D_n \odot X_n$ in random matrix theory (Sportisse et al., 2018, Lounici, 2012, Adhikari et al., 14 Aug 2025). This suggests that the term denotes not one canonical model but a family of formulations whose common feature is entrywise partial observation, with consequences for identifiability, regularization, asymptotic spectral behavior, and computation.

1. Core representations of a missing data matrix

Across the cited work, the basic object is always a matrix together with a mask or observation set indicating which entries are observed and which are missing.

Formulation	Observed object	Missingness representation
Statistical incomplete matrix	$Y_{\mathrm{obs}} = Y \odot \Omega$	$\Omega_{ij}=1$ observed, $\Omega_{ij}=0$ missing
Bernoulli-masked covariance samples	$Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$	$\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$
Random-matrix missing data matrix	$D_n \odot X_n$	deterministic $D_n\in\{0,1\}^{n\times n}$

In high-dimensional covariance estimation with missing observations, one observes i.i.d. vectors $X_1,\dots,X_n\in\mathbb{R}^p$ only through entrywise Bernoulli masks, so that $D_n \odot X_n$0, and the observed matrix is the original data matrix with missing entries replaced by $D_n \odot X_n$1 (Lounici, 2012). In general missing-data modeling, the same structure is written as $D_n \odot X_n$2, where $D_n \odot X_n$3 is the missingness mask (Sportisse et al., 2018). In SEGAN, the incomplete multivariate sample matrix $D_n \odot X_n$4 is paired with a mask matrix $D_n \odot X_n$5, where $D_n \odot X_n$6 for observed values and $D_n \odot X_n$7 for missing values (Pan et al., 2024). In additive NMF for missing attributes, the same role is played by a binary mask $D_n \odot X_n$8 with $D_n \odot X_n$9 if $Y_{\mathrm{obs}} = Y \odot \Omega$0 is observed and $Y_{\mathrm{obs}} = Y \odot \Omega$1 otherwise (Gupta, 2010).

A distinct formulation arises in random matrix theory. There, a missing data matrix is strictly $Y_{\mathrm{obs}} = Y \odot \Omega$2, where $Y_{\mathrm{obs}} = Y \odot \Omega$3 is a deterministic $Y_{\mathrm{obs}} = Y \odot \Omega$4–$Y_{\mathrm{obs}} = Y \odot \Omega$5 matrix and $Y_{\mathrm{obs}} = Y \odot \Omega$6 denotes the Hadamard product (Adhikari et al., 14 Aug 2025). That framework is explicitly differentiated from classical statistical missing-data models such as MCAR, MAR, and MNAR, because the mask is deterministic rather than probabilistic (Adhikari et al., 14 Aug 2025).

2. Missingness mechanisms, identifiability, and observation models

The cited literature treats several missingness mechanisms. In the Bernoulli masking model for covariance estimation, masks are i.i.d. and independent of the data, which is exactly a missing completely at random setting (Lounici, 2012). SEGAN also states that its theoretical analysis is predicated on relative independence between the mask and the data, meaning that the multivariate data are Missing Completely At Random (Pan et al., 2024).

Other models relax MCAR. In the Censored Graphical Horseshoe framework for Gaussian graphical models, missing data are assumed Missing At Random, so the probability that $Y_{\mathrm{obs}} = Y \odot \Omega$7 is missing can depend on the observed data but not on the unobserved value itself; this makes the missingness mechanism ignorable for likelihood-based inference (Mai et al., 10 Jan 2026). In survey matrix completion, the response probability is modeled as

$Y_{\mathrm{obs}} = Y \odot \Omega$8

with a column-specific logistic model, again placing the problem under MAR conditional on fully observed covariates (Mao et al., 2019).

Missing Not At Random is addressed explicitly in low-rank matrix completion with informative missingness. There, the paper uses a self-masked MNAR mechanism in which each indicator depends only on the corresponding cell value,

$Y_{\mathrm{obs}} = Y \odot \Omega$9

and models $\Omega_{ij}=1$0 by logistic regression (Sportisse et al., 2018). A different informative-missingness construction appears in matrix completion under a low-rank missing mechanism, where the observation probabilities are collected in a matrix $\Omega_{ij}=1$1 with $\Omega_{ij}=1$2 and a low-rank linear predictor $\Omega_{ij}=1$3 (Mao et al., 2018).

Identifiability depends on the mechanism and the structural model. In covariance estimation with Bernoulli masking, the masked covariance satisfies

$\Omega_{ij}=1$4

and this map is invertible for every $\Omega_{ij}=1$5, yielding

$\Omega_{ij}=1$6

Accordingly, $\Omega_{ij}=1$7 remains identifiable despite missingness (Lounici, 2012). Under deterministic sampling, identifiability is governed by the isomeric condition: if a matrix is not $\Omega_{ij}=1$8-isomeric, then there exist infinitely many distinct matrices with rank no larger than $\Omega_{ij}=1$9 that agree with the observed entries, so completion is not identifiable (Liu et al., 2018).

3. Covariance and graphical-model estimation with incomplete observations

One major line of work treats a missing data matrix as an obstacle to second-order estimation rather than entrywise imputation. In high-dimensional covariance estimation with missing observations, the empirical masked covariance

$\Omega_{ij}=0$0

is biased for $\Omega_{ij}=0$1, but the algebraic inversion above produces the unbiased estimator

$\Omega_{ij}=0$2

The proposed estimator then solves

$\Omega_{ij}=0$3

over positive semidefinite matrices, with $\Omega_{ij}=0$4 denoting the nuclear norm. The solution is spectral soft-thresholding of the eigenvalues of $\Omega_{ij}=0$5, and the paper establishes non-asymptotic oracle inequalities in Frobenius and spectral norms together with minimax lower bounds that are minimax optimal up to a logarithmic factor (Lounici, 2012).

The corresponding high-dimensional quantity is the effective rank

$\Omega_{ij}=0$6

which can be much smaller than $\Omega_{ij}=0$7 even when $\Omega_{ij}=0$8. The rates depend on $\Omega_{ij}=0$9, $Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$0, and $Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$1, and the dependence $Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$2 is shown to be sharp (Lounici, 2012). The method avoids explicit imputation altogether; missingness is handled analytically through re-normalization.

In sparse precision-matrix estimation, the Censored Graphical Horseshoe extends the Graphical Horseshoe prior of Li et al. to censored and arbitrarily missing Gaussian data by introducing a latent complete Gaussian matrix $Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$3 and sampling missing entries from conditional Gaussian distributions. For missing data, if $Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$4 and $Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$5,

$Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$6

with $Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$7 and $Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$8. A Gibbs sampler alternates latent-data augmentation with nodewise Horseshoe regressions, and the paper proves fractional-posterior contraction at rate $Y_i^{(j)}=\delta_{i,j}X_i^{(j)}$9 under both censoring and missingness (Mai et al., 10 Jan 2026). Here the missing data matrix is treated as partial information about an underlying latent Gaussian matrix, and posterior uncertainty propagates through the imputation step into graph uncertainty (Mai et al., 10 Jan 2026).

4. Matrix completion under deterministic, informative, high-rank, and nonlinear structure

Classical low-rank matrix completion is only one instance of missing-data recovery. Under deterministic sampling, recoverability is characterized by the isomeric condition and the relative condition number

$\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$0

For the full sampling pattern, $\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$1 quantifies relative well-conditionedness. The paper proves that if $\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$2 is $\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$3-isomeric and $\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$4, then nuclear norm minimization exactly and uniquely recovers $\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$5. It further studies the Schatten quasi-norm induced nonconvex method called isomeric dictionary pursuit,

$\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$6

and shows that exact solutions are strict local minima under $\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$7-isomerism and $\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$8 (Liu et al., 2018).

When missingness is informative rather than uniform, one can model the missingness matrix itself. In matrix completion under a low-rank missing mechanism, observation probabilities $\delta_{i,j}\sim \mathrm{Bernoulli}(\delta)$9 are estimated from the mask $D_n \odot X_n$0 by low-rank Bernoulli generalized linear modeling and then used in inverse probability weighting. The completed matrix is obtained from the weighted objective

$D_n \odot X_n$1

with constrained re-estimation of $D_n \odot X_n$2 to avoid instability from extremely small observation probabilities (Mao et al., 2018).

Other completion models replace global low rank by richer structure. In survey data prediction with multivariate MAR missingness, the population matrix is decomposed as

$D_n \odot X_n$3

with a low-rank residual $D_n \odot X_n$4 orthogonal to the covariate space. A projection strategy yields uniquely identified transformed parameters, and penalized estimators with nuclear-norm regularization produce a doubly robust estimator of population means with smaller mean squared error than competing methods in the reported simulations (Mao et al., 2019). In high-rank matrix completion and subspace clustering with missing data, the columns are assumed to lie in a union of at most $D_n \odot X_n$5 subspaces of rank at most $D_n \odot X_n$6, and the main theorem shows that each column can be perfectly recovered with high probability provided at least $D_n \odot X_n$7 entries are observed uniformly at random (Eriksson et al., 2011). In polynomial matrix completion, the ambient matrix may be high-rank or full-rank, but a polynomial feature map $D_n \odot X_n$8 is low-rank; the proposed method minimizes the rank of $D_n \odot X_n$9 using the kernel trick and Schatten-type relaxations, and extends to multiple nonlinear manifolds (Fan et al., 2019).

5. Imputation architectures, factorization models, and non-Gaussian completion

A second major line treats the missing data matrix as the direct input to an imputation architecture. SEGAN formulates multivariate imputation with three modules—generator, discriminator, and classifier—and defines the completed matrix by

$D_n\in\{0,1\}^{n\times n}$0

Its missingness hint matrix,

$D_n\in\{0,1\}^{n\times n}$1

partially reveals the mask to the discriminator. Under MCAR and idealized conditions, the paper proves that the hint mechanism uniquely determines the equilibrium and yields $D_n\in\{0,1\}^{n\times n}$2, while the classifier enforces $D_n\in\{0,1\}^{n\times n}$3 at Nash equilibrium (Pan et al., 2024).

Nonnegative factorization provides a different route. Additive NMF for missing data treats missing attributes as optimization variables and, for a single missing entry $D_n\in\{0,1\}^{n\times n}$4, solves

$D_n\in\{0,1\}^{n\times n}$5

The updates

$D_n\in\{0,1\}^{n\times n}$6

yield monotonic convergence of the squared reconstruction error (Gupta, 2010). For KL-based rank-1 NMF with missing data, A1GM exploits a grid-like mask, transforms the problem into non-negative multiple matrix factorization, and uses an analytical closed-formula for the best rank-1 solution; empirically it is more efficient than a gradient method with competitive reconstruction errors (Ghalamkari et al., 2021).

Recommendation-style missing data matrices motivate weighted factorization. Fast matrix factorization with non-uniform weights on missing data uses the whole-matrix loss

$D_n\in\{0,1\}^{n\times n}$7

allows arbitrary weighting on the missing data, and factorizes the missing-entry weight matrix as $D_n\in\{0,1\}^{n\times n}$8. The resulting fast eALS algorithm has per-iteration complexity

$D_n\in\{0,1\}^{n\times n}$9

so the time complexity is determined by the number of observed entries rather than the matrix size (He et al., 2018).

For count-valued incomplete matrices, negative binomial matrix completion replaces Poisson noise by the negative binomial model

$X_1,\dots,X_n\in\mathbb{R}^p$0

and solves the nuclear-norm-regularized objective

$X_1,\dots,X_n\in\mathbb{R}^p$1

by proximal gradient descent with singular value thresholding. In the reported experiments, the negative binomial model outperforms Poisson matrix completion under overdispersed count noise and high missingness (Lu et al., 2024).

6. Asymptotics, dependence, and computational limits

The phrase missing data matrix also has a precise asymptotic meaning in random matrix theory. For deterministic masks $X_1,\dots,X_n\in\mathbb{R}^p$2, the object $X_1,\dots,X_n\in\mathbb{R}^p$3 preserves classical free-probability limits if and only if the density of ones tends to $X_1,\dots,X_n\in\mathbb{R}^p$4: $X_1,\dots,X_n\in\mathbb{R}^p$5 in the square case, or

$X_1,\dots,X_n\in\mathbb{R}^p$6

in the rectangular covariance case. Under these necessary and sufficient conditions, iid, elliptic, and covariance missing data matrices converge to the same circular, elliptic, and Marčenko–Pastur limits as their fully observed counterparts, and independent families remain asymptotically free (Adhikari et al., 14 Aug 2025). The same paper explicitly distinguishes this deterministic-masking setting from statistical MCAR, MAR, and MNAR models (Adhikari et al., 14 Aug 2025).

Time dependence introduces another layer. In sparse transition matrix estimation for sub-Gaussian VAR processes with missing observations, the latent dynamics satisfy

$X_1,\dots,X_n\in\mathbb{R}^p$7

but each coordinate is only partially observed at each time point. The estimator corrects for the masking bias and solves a nonconvex Lasso-type program of the form

$X_1,\dots,X_n\in\mathbb{R}^p$8

together with an $X_1,\dots,X_n\in\mathbb{R}^p$9-norm constraint. The analysis relies on the convex concentration property for innovations and introduces a new transfer-function quantity $D_n \odot X_n$00, which characterizes the interaction of the time-varying observation process with the underlying dynamical system (Jalali et al., 2018).

Finally, worst-case computation is provably difficult. Weighted low-rank approximation

$D_n \odot X_n$01

specializes to low-rank approximation with missing data when $D_n \odot X_n$02. The paper proves that it is NP-hard to find an approximate solution even when a rank-one approximation is sought, both for strictly positive weights and for binary weights corresponding to missing entries (Gillis et al., 2010). This places a formal barrier beneath the wide range of alternating minimization, proximal, spectral-thresholding, and adversarial methods used throughout the missing-data literature.

The cumulative picture is technically heterogeneous but conceptually coherent. A missing data matrix may be an incomplete statistical dataset, a masked covariance sample, a partially observed time series design matrix, or a deterministic variance-profile random matrix. The central questions—what is identifiable, what structure replaces full observation, and what computational surrogate is tractable—are answered differently by low-rank geometry, nuclear or Schatten regularization, latent Gaussian augmentation, adversarial imputation, weighted factorization, and free-probability asymptotics (Lounici, 2012, Liu et al., 2018, Mai et al., 10 Jan 2026, Fan et al., 2019, Adhikari et al., 14 Aug 2025).