Additive Kernel Fusion for Complementary Signals

Why Additive Fusion?

Different omics assays can capture different aspects of similarity between the same patients. For example, a transcriptomic view may make two patients similar because their expression profiles are alike, while a proteomic view may identify similarity for a different pair. An additive fused kernel combines these signals by adding their similarities, so the contribution from each view remains explicit.

This is useful when views are intended to contribute complementary evidence rather than requiring patients to be similar in every view at once. The central choice is not merely how to add matrices: each kernel must refer to the same patients in the same order, the weights must have a meaningful interpretation, and the input kernels must satisfy the conditions that make their sum a valid kernel.

A kernel matrix is positive semidefinite (PSD) if it is symmetric and \(z^\mathsf{T}Kz\geq 0\) for every real vector \(z\) of length \(n\). This condition is the relevant validity guarantee: a PSD matrix can be interpreted as inner products among feature representations, even when those features are not constructed explicitly. The earlier material on positive semidefinite kernels and kernel matrices establishes this basic interpretation.

For additive fusion, the standard assumptions are that all matrices describe the same \(n\) patients in the same order, each input matrix is symmetric and PSD, and each \(\alpha_m\geq 0\). Nonnegative weights are needed for the general PSD guarantee. The kernels need not have equal numerical scales for the sum to be mathematically valid, but their scales matter when interpreting the weights as relative contributions.

Why the Sum Remains a Kernel

The PSD guarantee follows directly from the definition. For any vector \(z\), the quadratic form of the fused matrix is the weighted sum of the input quadratic forms. Since each input quadratic form is nonnegative and each weight is nonnegative, the sum cannot be negative.

The equality uses linearity of matrix multiplication and the scalar product. For each view, \(z^\mathsf{T}K^{(m)}z\geq 0\) by the PSD assumption, so every term \(\alpha_m z^\mathsf{T}K^{(m)}z\) is nonnegative. The sum is therefore nonnegative for every \(z\). Symmetry is also preserved because a weighted sum of symmetric matrices is symmetric. Together these facts show that \(K_{\mathrm{add}}\) is PSD.

There is an equivalent feature-space explanation. If \(K^{(m)}\) has feature vectors \(\phi_m(x_i)\), then its entries are inner products \(K^{(m)}_{ij}=\langle\phi_m(x_i),\phi_m(x_j)\rangle\). The fused feature representation is the concatenation \(\big(\sqrt{\alpha_1}\phi_1(x_i),\ldots,\sqrt{\alpha_M}\phi_M(x_i)\big)\). Its inner product with the corresponding representation for patient \(j\) is exactly \(K_{\mathrm{add},ij}\). Thus addition allows distinct views to contribute coordinates to one larger representation. This argument applies to nonnegative weights; a negative weight cannot be represented by a real square-root scaling and can destroy PSD.

What Weights Do—and Do Not—Mean

A weight controls the multiplier applied to one kernel matrix. If the kernels have comparable scales, larger weights generally give that view more influence on the entries of the fused matrix. But a weight alone does not state the fraction of biological information supplied by a view. A kernel whose entries are numerically much larger can dominate the sum even if its weight is smaller.

For instance, multiplying one kernel by 10 and dividing its weight by 10 leaves its contribution unchanged. Consequently, the meaning of the weights depends on how the kernels are scaled as well as on their numerical values. Kernel centering and normalization are established operations, but deciding how to prepare kernels for fusion is a separate modeling choice. Additive fusion itself does not automatically put heterogeneous kernels on comparable scales.

Nor does a nonnegative weight prove that a view is informative, reliable, or biologically relevant. A zero weight removes a view; a positive weight includes it. Choosing weights from data or imposing particular constraints requires additional reasoning and can affect interpretation. Here the construction is defined conditional on fixed weights, rather than claiming that one set of weights is universally optimal.

Seeing Complementary Signals in Pairwise Similarities

The entrywise decomposition is a practical interpretive advantage of additive fusion. For any patient pair, the fused entry is the sum of the weighted view-specific entries. One view can contribute strongly to one pair and weakly to another, so the overall matrix can retain different patterns from different assays. This does not guarantee that every view contributes unique information; it makes the arithmetic of its contribution inspectable.

Diagnostics, Alternatives, and Limitations

Before interpreting a fused kernel, check the alignment and the arithmetic. A matrix can be PSD and still be paired with the wrong patient ordering; in that case its entries refer to different patient pairs across views, so their sum has no intended interpretation. Check patient identifiers and row and column order, not just matrix dimensions. Also verify symmetry, inspect whether the input matrices are PSD to numerical tolerance, and calculate several fused entries directly from their weighted components.

A common diagnostic is to compare the scale and distribution of entries across views, including diagonal values and off-diagonal ranges. A view with much larger entries may dominate the fused similarity. Conversely, a centered kernel may contain negative entries, so a negative fused off-diagonal value is not by itself evidence that the fused matrix is invalid. PSD concerns quadratic forms, not a requirement that every entry be nonnegative. If a matrix has a materially negative eigenvalue, first investigate construction and numerical error rather than assuming addition will repair it.

Additive kernel fusion differs from Similarity Network Fusion (SNF). Additive fusion combines kernel entries by a weighted sum and has a direct PSD guarantee under the stated assumptions. SNF operates on affinity networks and iteratively exchanges information between them through diffusion; it is not simply the same matrix operation with different notation. Kernel alignment, covered in the earlier material on kernel alignment for comparing omics views, is useful for assessing agreement between views, but agreement is not the same as preserving each view’s weighted contribution in a fused matrix.

For linear kernels, concatenating feature vectors after multiplying view \(m\)'s features by \(\sqrt{\alpha_m}\) produces the same kernel as additive fusion. Direct concatenation without that scaling generally corresponds to particular implicit weights, not necessarily the weights intended by the analyst. A single-view analysis avoids cross-view scaling choices but discards the other views. These alternatives answer different modeling needs; none guarantees that a fused similarity captures a biological mechanism or causes an outcome.

Additive fusion is most useful when the views are aligned, the component kernels are meaningful on their own, and a transparent sum of their similarities matches the scientific question. Its guarantee is mathematical, not biological: PSD fusion does not correct batch effects, missingness, poor feature construction, or a mismatch between the chosen similarity and the scientific target. A fused kernel also compresses the views into one matrix, so retaining and inspecting the component kernels remains important for interpretation.