Uneven Cell Density and Trajectory Geometry

Why Uneven Cell Density Matters

An inferred trajectory is built from observed cells, not from every possible state along a biological process. If many cells occupy one region of a state path and few occupy another, the geometry of the observations can therefore differ markedly from the geometry of the underlying path. A dense region may look finely resolved, while a sparse region may be represented by a few widely separated points. This matters for principal curves, minimum spanning trees (MSTs) and diffusion distances, but it does not affect all three in the same way.

The central distinction is between the path’s geometry and the distribution of sampled cells along that path. Density is not itself evidence of a branch, a state transition or a biological pause. Conversely, density can influence an estimator: an empirical fit or graph is constructed from the cells that were sampled, and those cells are not necessarily distributed uniformly along the path.

This distinction is useful but not absolute. The population distribution of cells can be part of the target of some methods, including the usual self-consistency definition of a principal curve. Thus, changing density while holding the support fixed may leave the underlying path unchanged but can still change a distribution-dependent fitted curve.

Local Spacing and the Principal-Curve Fit

Suppose cells are sampled independently along a smooth path, with density \(f(s)\), and consider a small region away from the ends of the path. A local approximation treats \(f(s)\) as nearly constant. In a short interval of arclength \(h\), the expected number of sampled cells is approximately \(n f(s)h\), where \(n\) is the total number of cells. The typical local spacing is therefore on the order of \(1/[n f(s)]\): it is smaller in dense regions and larger in sparse ones.

For a more precise justification in the uniform case, take \(n\) independent sample positions on a path of length \(L\). These positions divide the path into \(n+1\) intervals, including the two end intervals. By symmetry, the expected length of each interval is \(L/(n+1)\), since their lengths sum to \(L\). In a small region with approximately constant density \(f(s)\), the local rate of points per unit arclength is approximately \(n f(s)\); the corresponding local spacing is approximately its reciprocal. This is a local approximation, not an exact spacing formula for arbitrary density or near a path endpoint.

A principal curve is commonly described as a smooth curve that is self-consistent: at each location along the curve, the mean of the data points projecting to that location lies on the curve. In practice, fitting algorithms approximate this condition using the observed cells and their projections. Because the observations contribute to the fit, densely sampled regions contribute more terms to empirical summaries and fitting criteria than sparsely sampled regions unless an explicit weighting scheme changes their influence.

This does not mean that density always bends a fitted curve toward the densest region. For example, if all cells lie exactly on a straight line, changing their density along that line does not change the line’s geometry. The risk arises when there is noise, curvature, alternative plausible fits or limited data: then the empirical distribution can affect which curve best represents the observations. The earlier material on principal curves and branch failures provides the relevant geometric context; here the concern is how uneven sampling changes the evidence available to the fit.

Density Effects on Minimum Spanning Trees

An MST connects the observed cells with minimum total edge weight, without cycles. For a cell graph whose edge weights are Euclidean distances, the tree is shaped by the point cloud and by the chosen candidate edges. Uneven density changes the scale of local distances: dense regions usually contain many short candidate edges, while sparse regions are more likely to require longer edges to connect their cells to the rest of the graph.

A long MST edge is therefore not, by itself, proof of a biological gap. It may simply bridge a sparsely sampled stretch of a continuous path. On the other hand, sparse sampling also gives the graph fewer observations with which to establish the route through that stretch. If different portions of a curved path are close in the measured space, a short edge between nearby but nonconsecutive portions may compete with the longer edge that follows the actual path. The resulting tree can take a shortcut or attach cells in an unintended order.

The edge weights and candidate-neighbor construction matter. A graph built from a restricted neighborhood may omit a potentially misleading long edge; a graph that permits all pairwise edges may include it. The earlier material on choosing MST edge weights and pruning spurious branches is relevant to those design decisions. Neither choice removes the basic limitation that a sparse region contains less information about the path between its observed cells.

Density and Diffusion Distance

A diffusion distance compares cells by how different their transition-probability profiles are after a specified number of steps. Its behavior depends both on the affinity graph and on how the transition operator is normalized. For a connected, reversible Markov chain with transition matrix \(P\) and stationary distribution \(\pi\), the diffusion distance at integer time \(t\) can be written

Here \(P^t(i,k)\) is the probability of reaching cell \(k\) from cell \(i\) in \(t\) steps, and \(\pi_k\) is the stationary probability of cell \(k\). The inverse-\(\pi_k\) factor means that differences in transition probability toward low-stationary-mass cells can contribute substantially. In a graph whose stationary mass reflects local cell density, dense regions can have a different influence from sparse ones.

The exact density effect is not universal. It depends on the affinity kernel, the neighborhood scale and the normalization used to construct \(P\). Row normalization makes each row sum to one, but does not generally make the stationary distribution uniform. Other normalizations can alter how strongly sampling density enters the operator. Thus, one should not interpret a diffusion distance as a density-free measure of physical or biological path length. It is a measure of separation under a specified transition model.

Diagnostics and Method Comparisons

Density effects are most concerning when a method’s apparent geometry changes substantially under reasonable changes to sampling or construction choices. Useful checks include plotting cells along the inferred ordering, examining local neighbor distances, and comparing results across neighborhood scales or density-aware weighting choices. For a principal curve, inspect fit residuals and whether dense regions dominate the fitted path. For an MST, inspect unusually long edges and possible shortcuts between nearby parts of a curved path. For diffusion distances, record the kernel, normalization and diffusion time; compare the inferred structure across defensible settings rather than treating one distance as an intrinsic quantity.

Each method also has a different limitation. A principal curve summarizes data with a smooth curve and can be sensitive to the distribution it summarizes. An MST gives a tree connecting observed cells but does not fill in unobserved states or establish that each edge follows a biological process. A diffusion distance summarizes connectivity through a chosen transition operator, not direct arclength along a unique path. These are complementary representations, not interchangeable estimates of one guaranteed geometric truth.

A common mistake is to interpret a dense cluster as a biological state in which cells must remain longer, or to interpret a long sparse-region edge as evidence of a true trajectory break. The correction is to separate what the data directly show—cell locations and their sampling frequency—from what the inferred geometry assumes. Density can reflect biological abundance, sampling design, capture efficiency or other factors; geometry alone does not identify which explanation applies.