Given two point patterns, find the distance between them based on optimal point matching.
pppdist(X, Y, type = "spa", cutoff = 1, q = 1, matching = TRUE,
ccode = TRUE, auction = TRUE, precision = NULL, approximation = 10,
show.rprimal = FALSE, timelag = 0)
Normally an object of class pppmatching
that contains detailed
information about the parameters used and the resulting distance.
See pppmatching.object
for details.
If matching = FALSE
, only the numerical value of the distance
is returned.
Two point patterns (objects of class "ppp"
).
A character string giving the type of distance to be computed.
One of "spa"
(default), "ace"
or "mat"
, indicating
whether the algorithm should find the optimal matching based on
“subpattern assignment”,
“assignment only if cardinalities are equal”
or “mass transfer”. See Details.
The value \(> 0\) at which interpoint distances are cut off.
The order of the average that is applied to the interpoint distances.
May be Inf
, in which case the maximum of the interpoint
distances is taken.
Logical. Whether to return the optimal matching or only the associated distance.
Logical. If FALSE
, R code is used which allows for higher
precision, but is much slower.
Logical. By default a version of Bertsekas' auction algorithm
is used to compute an optimal point matching if type
is
either "spa"
or "ace"
.
If auction
is FALSE
(or type
is "mat"
)
a specialized primal-dual algorithm is used instead.
This was the standard in earlier versions
of spatstat, but is several orders of magnitudes slower.
Index controlling accuracy of algorithm. The q
-th powers of
interpoint distances will be rounded to the nearest multiple of
10^(-precision)
. There is a sensible default which depends
on ccode
.
If q = Inf
, compute distance based on the optimal matching for the
corresponding distance of order approximation
. Can be
Inf
, but this makes computations extremely slow.
Logical. Whether to plot the progress of the primal-dual
algorithm. If TRUE
, slow primal-dual R code is used,
regardless of the arguments ccode
and auction
.
Time lag, in seconds, between successive displays of the iterative solution of the restricted primal problem.
Dominic Schuhmacher dominic.schuhmacher@mathematik.uni-goettingen.de, URL http://dominic.schuhmacher.name/
.
Computes the distance between point patterns X
and Y
based
on finding the matching between them which minimizes the average of
the distances between matched points
(if q=1
), the maximum distance between matched points
(if q=Inf
), and in general the q
-th order average
(i.e. the 1/q
th power of the sum of
the q
th powers) of the distances between matched points.
Distances between matched points are Euclidean distances cut off at
the value of cutoff
.
The parameter type
controls the behaviour of the algorithm if
the cardinalities of the point patterns are different. For the type
"spa"
(subpattern assignment) the subpattern of the point pattern
with the larger cardinality \(n\) that is closest to the point pattern
with the smaller cardinality \(m\) is determined; then the q
-th order
average is taken over \(n\) values: the \(m\) distances of matched points
and \(n-m\) "penalty distances" of value cutoff
for
the unmatched points. For the type "ace"
(assignment only if
cardinalities equal) the matching is empty and the distance returned is equal
to cutoff
if the cardinalities differ. For the
type "mat"
(mass transfer) each point pattern is assumed
to have total mass \(m\) (= the smaller cardinality) distributed evenly
among its points; the algorithm finds then the "mass transfer plan" that
minimizes the q
-th order weighted average of the distances, where
the weights are given by the transferred mass divided by \(m\). The
result is a fractional matching (each match of two points has a weight
in \((0,1]\)) with the minimized quantity as the associated distance.
The central problem to be solved is the assignment problem (for types
"spa"
and "ace"
) or the more general transport problem
(for type "mat"
). Both are well-known problems in discrete
optimization, see e.g. Luenberger (2003).
For the assignment problem pppdist
uses by default the
forward/backward version of Bertsekas' auction algorithm with
automated epsilon scaling; see Bertsekas (1992). The implemented
version gives good overall performance and can handle point patterns
with several thousand points.
For the transport problem a specialized primal-dual algorithm is
employed; see Luenberger (2003), Section 5.9. The C implementation
used by default can handle patterns with a few hundreds of points, but
should not be used with thousands of points. By setting
show.rprimal = TRUE
, some insight in the working of the
algorithm can be gained.
For a broader selection of optimal transport algorithms that are not restricted to spatial point patterns and allow for additional fine tuning, we recommend the R package transport.
For moderate and large values of q
there can be numerical
issues based on the fact that the q
-th powers of distances are
taken and some positive values enter the optimization algorithm as
zeroes because they are too small in comparison with the larger
values. In this case the number of zeroes introduced is given in a
warning message, and it is possible then that the matching obtained is
not optimal and the associated distance is only a strict upper bound
of the true distance. As a general guideline (which can be very wrong
in special situations) a small number of zeroes (up to about 50% of
the smaller point pattern cardinality \(m\)) usually still results
in the right matching, and the number can even be quite a bit higher
and usually still provides a highly accurate upper bound for the
distance. These numerical problems can be reduced by enforcing (much
slower) R code via the argument ccode = FALSE
.
For q = Inf
there is no fast algorithm available, which is why
approximation is normally used: for finding the optimal matching,
q
is set to the value of approximation
. The
resulting distance is still given as the maximum rather than the
q
-th order average in the corresponding distance computation.
If approximation = Inf
, approximation is suppressed and a very
inefficient exhaustive search for the best matching is performed.
The value of precision
should normally not be supplied by the
user. If ccode = TRUE
, this value is preset to the highest
exponent of 10 that the C code still can handle (usually \(9\)). If
ccode = FALSE
, the value is preset according to q
(usually \(15\) if q
is small), which can sometimes be
changed to obtain less severe warning messages.
Bertsekas, D.P. (1992). Auction algorithms for network flow problems: a tutorial introduction. Computational Optimization and Applications 1, 7-66.
Luenberger, D.G. (2003). Linear and nonlinear programming. Second edition. Kluwer.
Schuhmacher, D. (2014). transport: optimal transport in various forms. R package version 0.6-2 (or later)
Schuhmacher, D. and Xia, A. (2008). A new metric between distributions of point processes. Advances in Applied Probability 40, 651--672
Schuhmacher, D., Vo, B.-T. and Vo, B.-N. (2008). A consistent metric for performance evaluation of multi-object filters. IEEE Transactions on Signal Processing 56, 3447--3457.
pppmatching.object
, matchingdist
,
plot.pppmatching
# equal cardinalities
set.seed(140627)
X <- runifrect(500)
Y <- runifrect(500)
m <- pppdist(X, Y)
m
if(interactive()) {
plot(m)}
# differing cardinalities
X <- runifrect(14)
Y <- runifrect(10)
m1 <- pppdist(X, Y, type="spa")
m2 <- pppdist(X, Y, type="ace")
m3 <- pppdist(X, Y, type="mat", auction=FALSE)
summary(m1)
summary(m2)
summary(m3)
if(interactive()) {
m1$matrix
m2$matrix
m3$matrix}
# q = Inf
X <- runifrect(10)
Y <- runifrect(10)
mx1 <- pppdist(X, Y, q=Inf, matching=FALSE)
mx2 <- pppdist(X, Y, q=Inf, matching=FALSE, ccode=FALSE, approximation=50)
mx3 <- pppdist(X, Y, q=Inf, matching=FALSE, approximation=Inf)
all.equal(mx1,mx2,mx3)
# sometimes TRUE
all.equal(mx2,mx3)
# very often TRUE
Run the code above in your browser using DataLab