Low-Rank Matrix Completion With Missing Omics Blocks

From Missing Blocks to a Reconstruction Problem

In a multi-omics study, some subjects may have measurements from one modality but not another. After arranging subjects in rows and features from all modalities in columns, a modality that is unavailable for a subject appears as a block of missing entries in that row. The central question here is how to formulate a low-rank reconstruction using only the entries that were actually measured.

The earlier material on modality-availability indicators provides a way to record which blocks are observed; preprocessing and scaling determine how measurements from different features are placed on comparable scales. The matrix-completion objective uses the observed values and that availability information. It does not require filling missing cells with zeros before fitting.

Definition: Let \(X\) be an \(n \times p\) data matrix, with subjects in rows and omics features in columns. Let \(\Omega\) be the set of index pairs \((i,j)\) for which \(X_{ij}\) is observed. Define the observation mask \(M\) by \(M_{ij}=1\) if \((i,j)\in\Omega\), and \(M_{ij}=0\) otherwise. A completed matrix \(Z\) agrees with the observed data when its entries \(Z_{ij}\), for \((i,j)\in\Omega\), approximate the measured values \(X_{ij}\).

A low-rank model says that the full subject-by-feature matrix can be represented by a relatively small number of shared patterns. In a rank-\(r\) factorization, \(Z=UV^\mathsf{T}\), where \(U\) is \(n \times r\) and \(V\) is \(p \times r\). Each subject and each feature has an \(r\)-dimensional representation, and their inner product gives the reconstructed value. Rank is a structural restriction, not a guarantee that the missing entries can be recovered accurately.

Observed-Entry Loss and Rank Constraints

For numerical measurements with squared-error loss, the basic objective sums discrepancies only over the observed entries. A rank-constrained formulation is:

\widehat{Z} = \arg\min_{Z:\,\operatorname{rank}(Z)\le r} \sum_{(i,j)\in\Omega}(X_{ij}-Z_{ij})^2.

Equivalently, the objective can be written as \(\|M\odot(X-Z)\|_F^2\), where \(\odot\) denotes entrywise multiplication and the Frobenius norm squares and sums all matrix entries. Missing cells contribute zero to this masked residual because they are excluded from the sum; this is not the same as asserting that their measured values equal zero.

The rank constraint states how complex the reconstructed matrix may be. For fixed \(r\), writing \(Z=UV^\mathsf{T}\) gives an optimization over \(U\) and \(V\). This factorized problem is generally nonconvex: different initial values or optimization procedures can lead to different solutions. An alternative is to penalize the nuclear norm, \(\|Z\|_*\), defined as the sum of the singular values of \(Z\):

\widehat{Z} = \arg\min_Z \left\{ \frac{1}{2}\sum_{(i,j)\in\Omega}(X_{ij}-Z_{ij})^2 +\lambda\|Z\|_* \right\}, \qquad \lambda\ge 0.

The penalty discourages complex matrices without fixing their rank in advance. The nuclear norm is a convex surrogate for rank, so this formulation is convex in \(Z\), although the selected reconstruction still depends on the value of \(\lambda\). Larger penalties generally favor lower-complexity solutions and may also pull fitted values toward zero. Choosing the rank or penalty well is a separate practical task.

Model interpretation: The rank-constrained objective is a least-squares fit to observed entries among matrices of rank at most \(r\). It can also be motivated by independent, mean-zero Gaussian measurement errors with common variance on observed entries, conditional on the observation pattern. That likelihood interpretation does not by itself justify ignoring a missingness process that depends on unobserved values.

Worked Example: Writing a Block-Missing Mask

Suppose four subjects have two features from modality A and two from modality B. Subjects 1 and 2 have both modalities measured, subject 3 has A only, and subject 4 has B only. The following hypothetical matrix uses a dash to mark an unobserved entry:

\(X=\begin{bmatrix}2&1&4&3\\3&2&5&4\\4&3&-&-\\-&-&2&1\end{bmatrix}\). The feature columns are ordered as A1, A2, B1, B2.

The observed set contains the four entries in each of rows 1 and 2, the first two entries of row 3, and the last two entries of row 4. Thus there are \(4+4+2+2=12\) observed entries out of 16. For example, a proposed reconstruction \(Z\) is assessed by summing squared residuals at those 12 positions only. Its values in row 3's B block and row 4's A block do not enter this training loss.

This distinction matters: if a missing cell were replaced by zero and then included in the loss, the procedure would be asked to fit a value that was never measured. The mask preserves the difference between an observed zero and an unavailable measurement.

Why the Squared-Error Objective Has This Form

The link between observed-entry loss and Gaussian errors follows directly from the likelihood. Suppose a fixed underlying matrix \(Z\) generates each observed value as \(X_{ij}=Z_{ij}+\varepsilon_{ij}\), where the errors for \((i,j)\in\Omega\) are independent and \(\varepsilon_{ij}\sim N(0,\sigma^2)\). Conditional on the observed index set, the likelihood is

L(Z) = \prod_{(i,j)\in\Omega} \frac{1}{\sqrt{2\pi\sigma^2}} \exp\left\{-\frac{(X_{ij}-Z_{ij})^2}{2\sigma^2}\right\}.

Taking negative logarithms gives a constant plus \(\frac{1}{2\sigma^2}\sum_{(i,j)\in\Omega}(X_{ij}-Z_{ij})^2\). The constant does not depend on \(Z\), and multiplying the objective by a positive constant does not change its minimizer. This yields the observed-entry squared-error objective. A rank restriction can then be imposed to express the low-complexity signal assumption.

This argument also clarifies the scope of the objective. It describes measurement errors in observed cells; it does not establish that the observation pattern is unrelated to the unseen measurements. If that relationship is important, a reconstruction objective alone does not account for it.

Worked Example: Completing a Missing Omics Block Under Rank One

Consider hypothetical measurements with one feature from modality A and two features from modality B. Two subjects have all three features observed, while a third subject has A measured and the entire B block missing:

\(X=\begin{bmatrix}2&1&2\\4&2&4\\6&-&-\end{bmatrix}\). Assume for illustration that the completed matrix has rank one and fits the observed entries exactly.

The first two observed columns have the relationship \(X_{i2}=0.5X_{i1}\) for the first two subjects. The third observed column satisfies \(X_{i3}=X_{i1}\) for those subjects. A rank-one pattern consistent with these observations is therefore \(Z_{i2}=0.5Z_{i1}\) and \(Z_{i3}=Z_{i1}\). For subject 3, the observed A value is 6, so these relationships give \(Z_{32}=0.5(6)=3\) and \(Z_{33}=6\).

The resulting candidate is \(Z=\begin{bmatrix}2&1&2\\4&2&4\\6&3&6\end{bmatrix}\). Its observed-entry loss is

$$ (2-2)^2+(1-1)^2+(2-2)^2 +(4-4)^2+(2-2)^2+(4-4)^2+(6-6)^2=0. $$

The completed matrix has rank one because every row is a multiple of \((1,0.5,1)\). This is a constructed illustration: it shows how a low-rank pattern can supply values for an unobserved block, but a zero observed loss does not prove that those inferred values are correct in a real data set.

What the Optimization Can and Cannot Say

An observed-entry loss evaluates agreement with measured data, not accuracy in unmeasured cells. Many completed matrices may have similar observed loss, particularly when whole modality blocks are absent for some subjects. A low-rank constraint narrows the candidates but does not automatically make the missing values identifiable. The later material on identifiability assumptions addresses what observation patterns and structural conditions are needed for stronger recovery claims.

The method is most interpretable when the features have been put on scales appropriate for a shared low-rank representation and when a low-rank signal is scientifically plausible. If one high-variance feature dominates the squared loss, or if features have incompatible scales, the fit can mainly reflect those scale differences. Preprocessing and block weighting therefore affect what the objective considers a good fit; the loss does not neutralize poor scaling.

A useful diagnostic is to examine observed residuals \(X_{ij}-\widehat{Z}_{ij}\) by feature, modality, subject, and availability pattern. Systematic residuals or substantially different residual spreads across modalities suggest that a common squared-error model may not describe all observed measurements well. Also check whether optimization results are sensitive to initialization in a factorized rank-constrained fit. For an initial predictive check, some observed entries can be held out, the model fitted to the remaining entries, and predictions compared with the withheld values. Such a check assesses prediction only in the evaluated observed-data setting; it cannot directly verify accuracy in a modality block that is never observed for comparable subjects.

Worked Example: Comparing Two Reconstructions by Observed Loss

Suppose the hypothetical observed matrix is \(X=\begin{bmatrix}2&1\\1&-\end{bmatrix}\), where the lower-right entry is missing. Compare two candidate completions: \(Z_A=\begin{bmatrix}2&1\\1&0\end{bmatrix}\) and \(Z_B=\begin{bmatrix}2&1\\1&0.5\end{bmatrix}\). Both candidates agree exactly with every observed entry.

For either candidate, the observed-entry loss is \((2-2)^2+(1-1)^2+(1-1)^2=0\). The candidate's value for the missing cell does not affect this loss. In fact, the observed-entry objective alone cannot prefer 0 to 0.5 here.

The ranks differ: \(Z_A\) has determinant \(-1\), so it has rank two, while \(Z_B\) has determinant \(2(0.5)-1(1)=0\), so it has rank one. A rank-one constraint would prefer \(Z_B\) among these two candidates, but that preference comes from the structural constraint, not from additional evidence in the observed residuals. The example makes clear why a good observed fit is not, by itself, a validation of the missing entry.

Rank Constraints, Penalties, and Practical Alternatives

Approach	What it imposes	Practical distinction
Rank-constrained least squares	A maximum rank \(r\)	Directly limits matrix complexity, but optimization is generally nonconvex.
Nuclear-norm-penalized least squares	A cost proportional to the sum of singular values	Convex in the completed matrix; complexity is controlled by the penalty rather than a fixed rank.
Complete-case analysis	Uses subjects with all required modalities observed	Avoids reconstructing cells but may discard substantial data and change the analyzed population.
Conditional Gaussian imputation	Uses a specified joint distribution to predict missing values	Can represent distributional uncertainty and dependence, but relies on a distributional model.

Low-rank completion is attractive when shared structure across features and subjects is a reasonable approximation and retaining partially observed subjects matters. Its limitations follow from the same simplicity: it represents signal through a small number of patterns, and a single squared-error objective may not suit every feature or modality. It also returns a fitted completion, not automatically a measure of uncertainty for each missing value or a solution to bias from a data-dependent observation process.

Common Mistakes and Checks

Filling missing cells with zero before fitting. This treats unmeasured values as observed zeros and changes the loss. Keep a mask and sum over observed indices only.
Interpreting a low training loss as proof of accurate completion. The loss contains no direct information about entries that were never observed. State the low-rank assumption and evaluate predictions where observed values can be held out.
Assuming low rank resolves every missing-block pattern. A rank restriction reduces the set of candidate matrices, but it does not guarantee a unique or accurate completion for every observation pattern.
Ignoring scale and modality-specific residuals. Squared error gives greater influence to discrepancies that are large on the chosen scale. Review preprocessing and residuals before treating a single common loss as appropriate.

Key takeaway: Matrix completion with missing omics blocks minimizes discrepancies over observed entries while restricting the completed matrix to be low complexity. The mask defines what is fitted; the rank or penalty defines what patterns are permitted. Neither alone guarantees recovery of every missing measurement.

Check Your Understanding

Answer each question, then compare with the solutions below.

In a matrix with 20 observed entries and 5 missing entries, which entries contribute to the observed-entry squared-error loss?
For observed values \(X_{11}=3\) and \(X_{12}=2\), a candidate has \(Z_{11}=2\) and \(Z_{12}=2\). What is its contribution to the observed loss from these two cells?
What does a rank constraint add to the observed-entry least-squares problem?
Two completions have identical fitted values on every observed cell but different values in a missing block. Can observed-entry loss distinguish them?
Name one diagnostic that could reveal a poor fit to the observed data, and state one limitation of that diagnostic.

Solutions. (1) Only the 20 observed entries; the five missing entries are excluded. (2) \((3-2)^2+(2-2)^2=1+0=1\). (3) It restricts the candidate completed matrices to rank at most \(r\), limiting their complexity. (4) No. Their observed-entry losses are equal because missing-cell values do not appear in that loss. (5) Residuals grouped by modality can reveal systematic patterns or differing spreads; however, observed residuals do not directly establish accuracy for a block that remains unobserved.