Reading Diffusion Eigenvalues and Eigenvectors

Why Read the Spectrum?

A diffusion map represents cells using eigenvectors of a transition operator built from a cell-neighborhood graph. The coordinates are useful, but their interpretation is not automatic: some eigenvectors describe broad, smooth differences among cells, while others vary more rapidly or reflect details of the graph. Reading the eigenvalues alongside the eigenvectors helps assess which patterns dominate the diffusion geometry and whether a single coordinate could reasonably summarize cell-state variation.

The key diagnostic is joint rather than numerical alone. A clearly separated leading nontrivial eigenvalue can indicate one dominant mode of variation, but it does not prove that cells lie on one biological path. The corresponding eigenvector must also have a pattern consistent with a smooth ordering, and the result should be checked for instability and graph artifacts.

Eigenvalues and Modes of Variation

Let \(W\) be a symmetric, nonnegative affinity matrix for \(n\) cells, with degree matrix \(D\) whose \(i\)th diagonal entry is \(d_i=\sum_j W_{ij}\). Assume every cell has positive degree. The row-stochastic transition matrix is \(P=D^{-1}W\), and its symmetrically normalized counterpart is \(S=D^{-1/2}WD^{-1/2}\). These matrices have the same eigenvalues. Because the graph is undirected and the chain is reversible, the eigenvalues are real. Write them as \(\lambda_0,\lambda_1,\ldots,\lambda_{n-1}\), ordered by decreasing magnitude when discussing diffusion-time persistence, with \(\lambda_0=1\) for a connected graph.

A right eigenvector \(\psi_\ell\) of \(P\) satisfies \(P\psi_\ell=\lambda_\ell\psi_\ell\). The corresponding eigenvector \(u_\ell\) of \(S\) is related by \(\psi_\ell=D^{-1/2}u_\ell\). The mode \(\psi_0\) is constant for a connected graph; it describes the stationary component rather than a cell ordering. The nontrivial eigenvectors \(\psi_1,\psi_2,\ldots\) describe patterns of variation across cells. In diffusion-map coordinates, a common convention weights the \(\ell\)th mode by \(\lambda_\ell^t\), where \(t\) is the number of diffusion steps.

For nonnegative transition probabilities, eigenvalues lie between \(-1\) and \(1\). A mode with eigenvalue near \(1\) changes slowly under repeated transitions; a mode with smaller absolute eigenvalue decays more quickly. A negative eigenvalue also changes sign from one step to the next, so a large negative eigenvalue can signal oscillatory behavior rather than a smooth, persistent ordering. In many practical graphs the modes of interest have positive eigenvalues, but the spectrum should be checked rather than assumed to be so.

When the relevant eigenvalues are positive and ordered \(\lambda_1\geq\lambda_2\), the gap \(\lambda_1-\lambda_2\) compares the leading nontrivial mode with the next mode. A large gap is evidence that the first mode is more prominent in the diffusion representation. When signs matter, compare magnitudes and inspect negative modes too; an absolute-value gap is \(|\lambda_1|-|\lambda_2|\) when the modes are ordered by magnitude. Neither gap is a universal cutoff or a formal test of one-dimensionality.

Why Eigenvalues Weight the Modes

The connection between eigenvalues and persistence follows directly from repeated transitions. For a reversible operator with a complete eigenbasis, a cell-wise function \(f\) can be expanded in the right eigenvectors. Applying \(P\) repeatedly multiplies each eigenvector component by the corresponding eigenvalue at every step.

Thus, if \(0<\lambda_2<\lambda_1<1\), the relative contribution of mode 2 to mode 1 after \(t\) steps is multiplied by \((\lambda_2/\lambda_1)^t\). This explains why a mode with a larger positive eigenvalue persists longer and why diffusion coordinates downweight less persistent modes. It does not show that the persistent mode is biological: it only describes the transition geometry encoded by the chosen graph.

Read the Eigenvector Pattern, Not Just the Gap

An eigenvector is a vector of values indexed by cells. To evaluate it as a possible ordering coordinate, inspect those values against the inferred neighborhood structure or a proposed path. A plausible one-dimensional mode often changes gradually across connected neighborhoods. If an independent ordering along a candidate path is available, the eigenvector may rise or fall smoothly along it. Its sign is arbitrary: replacing \(\psi_1\) by \(-\psi_1\) gives the same eigenspace and reverses the ordering, not its substantive content.

Patterns that deserve closer inspection include a sharp division between two groups, a mode that distinguishes two arms from their shared region, and rapid alternation between neighboring cells. These can indicate a split, a poorly connected graph, or a mode that does not represent progression along one path. A smooth-looking coordinate is not sufficient evidence either: the graph construction and the data may impose a smooth pattern even when the biology is not a single trajectory.

Near-Ties, Stability, and Diagnostics

If two eigenvalues are equal, their eigenvectors are not individually determined: any orthonormal rotation within their shared eigenspace is also a valid basis. When eigenvalues are close, small changes to the graph or data can substantially rotate the estimated eigenvectors, even when the combined subspace is stable. Consequently, comparing a particular eigenvector across runs can be misleading when its eigenvalue is nearly tied with another. In that situation, assess the span of the modes as a group rather than treating each coordinate as uniquely identified.

Useful diagnostics include examining the ordered eigenvalues and their gaps, plotting or tabulating leading eigenvector values across neighboring cells, and checking for negative eigenvalues with large magnitude. Repeat the inspection under reasonable neighborhood and kernel choices from the earlier material on selecting neighborhood and kernel scales. A claimed one-dimensional structure is more credible if its leading mode remains interpretable across such choices. If a graph has disconnected components, the stationary eigenspace is not just one constant mode; handle connectivity before interpreting a nontrivial spectrum.

These are diagnostics, not a mechanical decision rule. No fixed gap size identifies a trajectory dimension across datasets: the spectrum depends on the graph, its scaling, the number of cells, and sampling variation. A gap can be small even when a useful one-dimensional approximation exists, or large when a graph artifact creates one dominant split. Eigenvector smoothness also cannot establish a temporal direction, a biological mechanism, or causation.

Common Mistake and Practical Comparison

A common mistake is to conclude, “The first eigenvalue is the largest nontrivial one, so the first eigenvector is the trajectory.” The correction is that eigenvalue rank indicates persistence under the chosen transition operator, not biological meaning or one-dimensionality. Inspect the eigenvector pattern, the next modes, the graph’s connectivity, and sensitivity to reasonable graph choices before using it as an ordering coordinate.

The spectrum is a compact diagnostic for diffusion-based geometry. In contrast, a principal curve asks whether cells can be summarized by a smooth curve in a chosen representation, while an MST represents cells using connected paths in a graph. These approaches answer related but different geometric questions. A leading diffusion eigenvector can support or challenge a one-dimensional interpretation, but it is not a substitute for inspecting the inferred curve or graph structure. The introductory material on diffusion maps establishes the coordinate construction; here the emphasis is on how to interpret its modes and limitations.