Why Separate Shared and Block-Specific Variation?
Matched omics blocks measure different features on the same samples. Their sample patterns may align: samples high on one block's molecular profile may also tend to be high on another's. They may also vary in ways that do not align across blocks. Distinguishing these patterns matters because a joint summary can obscure whether a sample separation is supported across assays or mainly reflects one assay.
Here, “shared” refers to statistical structure that is expressed in more than one block across matched samples. “Block-specific” refers to structure present within a block but not reflected in the other blocks under consideration. These are descriptions of covariance patterns, not claims that the same molecules are measured, that a biological mechanism is common, or that one block causes another.
Covariance as Evidence of Shared Sample Structure
Suppose \(n\) matched samples have measurements in \(B\) blocks. Write \(X^{(b)}\) for the \(n \times p_b\) data matrix for block \(b\), with samples in rows and features in columns. Assume the rows are aligned to the same samples in every block. Let \(X^{(b)}\) be column-centered, so each feature has sample mean zero. The sample covariance matrix within block \(b\) is \(S_{bb}\); the cross-covariance matrix between blocks \(b\) and \(c\) is
An entry of \(S_{bc}\) is the sample covariance between one feature in block \(b\) and one feature in block \(c\). A nonzero entry indicates that those feature measurements vary together across the observed samples, in the direction given by its sign. A cross-covariance pattern across several entries can indicate shared sample structure even when no single feature pair is a strong summary.
Within-block covariance alone cannot establish that a pattern is shared. A block can have substantial variation among its own features while its cross-covariances with other blocks are near zero. Conversely, cross-block covariance is not a direct measure of how much total variation each block contains: covariance is affected by the units and scales of the features. The sample covariance matrix also needs interpretation in light of the number of samples, feature counts, and data quality.
Worked Example: Perfectly Aligned Variation in Two Blocks
Consider two centered single-feature blocks measured on four hypothetical matched samples. The feature values are \(x=(-1,-1,1,1)\) and \(z=(-2,-2,2,2)\). Both have mean zero, and the same two samples are low and the same two are high.
The sample variances and covariance are
The sample correlation is \(s_{xz}/(s_xs_z)=1\). Thus, in these four observations, the two features have exactly aligned sample patterns: the larger scale of \(z\) changes its variance and covariance, but not the correlation or the ordering of samples. This is evidence of shared variation for this pair of features. It does not establish why the profiles align or whether the same pattern will hold in a new sample.
A Simple Model and Its Covariance Implication
A useful way to formalize the distinction is to consider a model with one shared sample score. Let \(t\) be an \(n\)-vector of centered sample scores, and let \(a_b\) be a \(p_b\)-vector of feature loadings for block \(b\). Write
The first term gives the part of block \(b\)'s profile aligned with \(t\); \(E^{(b)}\) contains remaining variation. To interpret the first term as shared, assume the same \(t\) applies across the blocks, the residuals have zero cross-block covariance, and the residuals are uncorrelated with \(t\). In a population formulation, these mean that the residual vectors from different blocks have zero cross-covariance and that each residual has zero covariance with the shared score. In a finite sample, the corresponding inner products may be zero by construction or only approximately zero.
Under these assumptions, the cross-covariance between blocks contains the shared contribution. For a population score with variance \(\operatorname{Var}(t)\), expansion of the covariance gives
The terms involving one score and one residual vanish by the uncorrelatedness assumption, and the residual cross-covariance vanishes by assumption. Within a block, by contrast, its covariance includes both the shared term and its own residual covariance:
This calculation explains why a block can contain variation that is not shared: its residual covariance can be substantial even when residuals do not covary with other blocks. It also shows the assumptions doing the work. If residuals are correlated across blocks, some cross-covariance is not represented by the proposed shared score. With multiple shared scores, the shared contribution becomes a sum of such terms.
Worked Example: One Feature With Both Common and Block-Specific Variation
Consider four hypothetical samples with centered score vectors \(t=(-1,-1,1,1)\) and \(e=(-1,1,-1,1)\). Their inner product is \((-1)(-1)+(-1)(1)+(1)(-1)+(1)(1)=0\). Define a feature in block A by \(x=t+e\), and a feature in block B by \(z=2t\). Thus \(x=( -2,0,0,2)\) and \(z=(-2,-2,2,2)\).
The sample variance of \(t\) is \(4/3\), and that of \(e\) is also \(4/3\). Since their sample covariance is zero, the variance of \(x=t+e\) is
The part of \(x\)'s variation aligned with \(t\) has variance \(4/3\); the part aligned with \(e\) has variance \(4/3\). In block B, \(z=2t\), so \(s_z^2=4(4/3)=16/3\). The cross-covariance is
The covariance arises from the shared score \(t\), while the \(e\) component contributes to block A's variance but not to its covariance with block B. This is an exact decomposition for the constructed vectors, not evidence that real data have a uniquely identifiable biological component.
Zero Covariance, Correlation, and Interpretation
Covariance is scale-dependent. If feature \(z\) is recorded in units 100 times larger, its covariance with \(x\) is also multiplied by 100. Correlation standardizes covariance by the two sample standard deviations, so it is useful for judging linear alignment without the original units. However, it does not answer every question: a small covariance can result from weak association, small measurement scales, or noisy estimation, while a high correlation can occur for features with very little absolute variation.
Worked Example: Distinct Sample Patterns With Zero Covariance
Take the centered four-sample vectors \(x=(-1,0,1,0)\) and \(y=(0,-1,0,1)\). Their sample variances are both \(2/3\), because each sum of squared values is 2 and \(n-1=3\). Their covariance is
The two observed profiles therefore have zero sample linear covariance, despite each varying across samples. This supports the description “no linear association in this small sample,” not the stronger claim that the blocks are independent or that no shared structure of any kind exists. With only four observations, nonlinear relationships, outliers, or sampling variation could also make covariance an incomplete guide.
Assumptions, Diagnostics, and Alternatives
Covariance comparisons are most interpretable when blocks refer to the same matched units, features have been handled consistently, and the measurements have been centered appropriately for the question. Missing samples or mismatched row order can create meaningless cross-block covariance. Strong batch effects, clinical group differences, or other sample-level covariates can induce covariance across blocks even when the underlying biological processes are not directly shared. Likewise, unremoved technical artifacts can make two assays appear aligned.
Useful checks include verifying sample identifiers and block sizes, inspecting feature distributions and outliers, and examining whether cross-block patterns persist within relevant groups or after accounting for known covariates. In high-dimensional data, a large cross-covariance matrix can be noisy relative to the number of samples; apparent patterns should be checked for stability rather than treated as certain. Zero empirical covariance is not proof of population zero covariance, especially in a small sample.
The choice of summary depends on the question. Separate within-block analyses describe each block's own variation but do not by themselves establish shared structure. Pairwise covariance or correlation matrices expose feature-level cross-block associations, while CCA and PLS are standard approaches for finding coordinated combinations of features across two blocks. Those approaches target particular kinds of association and do not, by their existence alone, establish a unique common-versus-specific decomposition. A latent-factor model can express such a decomposition, but its interpretation depends on model assumptions and may not be identifiable from observed covariance alone.
| Approach | What it describes | Main caution |
|---|---|---|
| Within-block covariance | Variation and co-variation among features in one block | Cannot alone show whether a pattern is shared across blocks |
| Cross-block covariance or correlation | Linear association between features across matched samples | Scale, confounding, and sampling variability affect interpretation |
| CCA or PLS | Coordinated combinations of features across two blocks | Targets selected associations; does not automatically partition all variation |
The most important limitation is that “common” and “specific” are model-relative descriptions. The observed covariance matrices do not generally tell us uniquely which latent biological processes produced them. A shared sample pattern may reflect a common biological state, a shared technical artifact, or a covariate affecting both assays. A block-specific pattern may be biologically meaningful, measurement noise, or a combination. Therefore, report what the covariance supports and state the assumptions needed for stronger interpretations.
Common Pitfalls
- Calling any within-block variation “shared.” Variation is shared only when there is corresponding structure across blocks; within-block variance alone is insufficient.
- Treating zero covariance as independence. Zero covariance rules out only linear association at the population level under the relevant definition. It does not generally rule out nonlinear dependence.
- Calling covariance biological evidence. Covariance describes how measurements co-vary across samples. Batch, group composition, or other confounding can produce the same pattern.
- Comparing raw covariances across differently scaled features. Their magnitude depends on measurement units. Use scale-aware interpretation and state whether covariance or correlation is being examined.
- Assuming every feature is wholly common or wholly specific. As the decomposition example shows, a feature can contain both components.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- Why does a nonzero cross-block covariance support shared sample structure, and what does it not establish?
- Suppose \(X^{(A)}=t a_A^\mathsf{T}+E^{(A)}\) and \(X^{(B)}=t a_B^\mathsf{T}+E^{(B)}\). Name the assumptions that yield cross-covariance \(a_A\operatorname{Var}(t)a_B^\mathsf{T}\).
- Two centered features have sample covariance 6. If the second feature is converted to units 10 times larger, what is the new covariance? Does their correlation change?
- Can a feature have both common and block-specific variation? Explain using the score-and-residual representation.
- Why is it incorrect to conclude that two blocks are independent because their sample covariance is zero?
Solutions. (1) It indicates linear co-variation across the matched observed samples, but does not identify a biological mechanism or rule out confounding. (2) The residuals must have zero cross-block covariance and be uncorrelated with the shared score; the same centered score \(t\) must apply to both blocks. (3) The covariance becomes \(10(6)=60\). Correlation is unchanged by a positive change of units because the covariance and the second feature's standard deviation are both multiplied by 10. (4) Yes. The score term contributes common variation, while the residual can contribute within-block variation not expressed in the other block. (5) Zero sample covariance is an empirical statement about linear co-variation in the observed sample, not proof of population independence; nonlinear dependence and sampling uncertainty remain possible.