bottleneck_distance() returns the bottleneck distance between two
persistence diagrams, one value per homology dimension. It takes two
persistent_homology() results, or two data frames of persistence
pairs. For the finite pairs, the distance is
$$W_\infty(D_1, D_2) = \inf_{\gamma} \sup_{p \in D_1} \|p - \gamma(p)\|_\infty,$$
where \(\gamma\) ranges over bijections \(D_1 \cup \Delta \to
D_2 \cup \Delta\) and \(\Delta = \{(x,x)\}\) is the diagonal. Each
point is matched to a point of the other diagram or to its projection
onto the diagonal, at cost \(|d - b|/2\).
The distance is the largest shift needed to match the two diagrams'
features: a small value means every feature of either diagram has a close
counterpart in the other or is short-lived enough to sit near the
diagonal. It is a distance between two observed diagrams and carries no
sampling uncertainty of its own; to compare diagrams on a common grid,
or to average several, use persistence_landscape().
Arguments
- d1, d2
The two diagrams:
persistent_homology()results, or data frames with the columnsdimension,birthanddeath(other columns are ignored). Any other input, or a data frame missing one of those columns, raises an error.- dimension
The homology dimensions to compare, as whole numbers. Default
NULL: every dimension that appears in either diagram. A dimension absent from both gives 0.- tol
The tolerance added to each candidate cost when checking whether a matching exists, so that floating-point error does not miss an exact match. Default
.Machine$double.eps ^ 0.5.
Value
A named numeric vector with one distance per dimension, named
"dim_<k>" (empty when there is no dimension to compare). Inf marks
a structural mismatch: different numbers of essential classes in that
dimension. The distance of a diagram to itself is 0.
Details
The smallest feasible \(\varepsilon\) is found by binary search over the candidate costs (every pairwise \(L_\infty\) distance and every diagonal projection), checking each with a Kuhn bipartite matching.
Essential classes (death = Inf in VR mode, death = 0 with
birth > 0 in clique mode, as persistence_landscape() reads them; a
(0, 0) pair is a zero-length finite point) are matched one-to-one within
each
dimension by their sorted births. When the two diagrams have different
numbers of essential classes in a dimension, the distance for that
dimension is Inf.