E-Statistics: Multivariate Inference via the Energy of Data

E-statistics (energy) tests and statistics for multivariate and univariate inference, including distance correlation, one-sample, two-sample, and multi-sample tests for comparing multivariate distributions, are implemented. Measuring and testing multivariate independence based on distance correlation, partial distance correlation, multivariate goodness-of-fit tests, k-groups and hierarchical clustering based on energy distance, testing for multivariate normality, distance components (disco) for non-parametric analysis of structured data, and other energy statistics/methods are implemented.

energy package for R

The energy package for R implements several methods in multivariate analysis and multivariate inference based on the energy distance, which characterizes equality of distributions.

Distance correlation (multivariate independence), disco (nonparametric extension of ANOVA), and goodness-of-fit tests are examples of some of the methods included.

energy is named based on the analogy with potential energy in physics. See the references in the manual for more details.

[ CRAN_Status_Badge ]https://cran.r-project.org/package=energy)

News

energy 1.7-5

User level changes:
- kgroups: (new) implements energy clustering for a specified number k classes by energy distance criterion, analogous to the k classes of the k-means algorithm.
- dcov2d and dcor2d: (new) O(n log n) methods to compute the squared U or V statistics for real x and y
- sortrank() function added (a utility)
Internal changes:
- B-tree.cpp: Btree_sum and other internal functions implement binary tree search for faster O(n log n) calculation of paired distances in dcov2d
- kgroups.cpp: Rcpp implementation of k-groups algorithm
- energy.hclust implementation: replaced C++ code with call to stats::hclust; since R > 3.0.3 it is now equivalent for alpha = 1 with method = "ward.D". Input and return value unchanged except heights from hclust are half.

energy 1.7-4

User level changes
- disco: handle the case when the user argument x is dist with conflicting argument distance=FALSE
- dcor.t and dcor.ttest: handle the cases when class of argument x or y conflicts with the distance argument
- Split manual page of dcovU into two files.
- indep.etest and indep.e removed now Defunct (were Deprecated since Version 1.1-0, 2008-04-07; replaced by indep.test).
Internal changes
- BCDCOR: handle the cases when class of argument x or y conflicts with the distance argument

energy 1.7-2

User level changes
- Provided new dcor.test function, similar to dcov.test but using the distance correlation as the test statistic.
- Number of replicates R for Monte Carlo and permutation tests now matches the argument of the boot::boot function (no default value, user must specify).
- If user runs a test with 0 replicates, p-value printed is NA
Internal changes
- energy_init.c added for registering routines

energy 1.7-0

Partial Distance Correlation statistics and tests added
- pdcov, pdcor, pdcov.test, pdcor.test
- dcovU: unbiased estimator of distance covariance
- bcdcor: bias corrected distance correlation
- Ucenter, Dcenter, U_center, D_center: double-centering and U-centering utilities
- U_product: inner product in U-centered Hilbert space
updated NAMESPACE and DESCRIPTION imports, etc.
revised package Title and Description in DESCRIPTION
package now links to Rcpp
mvnorm c code ported to c++ (mvnorm.cpp); corresponding changes in Emvnorm.R
syntax for bcdcor: "distance" argument removed, now argument can optionally be a dist object
syntax for energy.hclust: first argument must now be a dist object
default number of replicates R in tests: for all tests, R now defaults to 0 or R has no default value.

energy 1.6.2

inserted GetRNGstate() .. PutRNGState around repl. loop in dcov.c.

energy 1.6.1

replace Depends with Imports in DESCRIPTION file

energy 1.6.0

implementation of high-dim distance correlation t-test introduced in JMVA Volume 117, pp. 193-213 (2013).
new functions dcor.t, dcor.ttest in dcorT.R
minor changes to tidy other code in dcov.R
removed unused internal function .dcov.test

energy 1.5.0

NAMESPACE: insert UseDynLib; remove zzz.R, .First.Lib()

energy 1.4-0

NAMESPACE added.
(dcov.c, Eindep.c) Unused N was removed.
(dcov.c) In case dcov=0, bypass the unnecessary loop that generates replicates (in dCOVtest and dCovTest). In this case dcor=0 and test is not significant. (dcov=0 if one of the samples is constant.)
(Eqdist.R) in eqdist.e and eqdist.etest, method="disco" is replaced by two options: "discoB" (between sample components) and "discoF" (disco F ratio).
(disco.R) Added disco.between and internal functions that compute the disco between-sample component and corresponding test.
(utilities.c) In permute function replaced rand_unif with runif.
(energy.c) In ksampleEtest the pval computation changed from ek/B to (ek+1)/(B+1) as it should be for a permutation test, and unneeded int* n removed.

energy 1.3-0

In distance correlation, distance covariance functions (dcov, dcor, DCOR) and dcov.test, arguments x and y can now optionally be distance objects (result of dist function or as.dist). Matrices x and y will always be treated as data.
Functions in dcov.c and utilities.c were modified to support arguments that are distances rather than data. In utilities.c the index_distance function changed. In dcov.c there are many changes. Most importantly for the exported objects, there is now an extra required parameter in the dims argument passed from R. In dCOVtest dims must be a vector c(n, p, q, dst, R) where n is sample size, p and q are dimensions of x and y, dst is logical (TRUE if distances) and R is number of replicates. For dCOV dims must be c(n, p, q, dst).

energy 1.2-0

disco (distance components) added for one-way layout.
A method argument was added to ksample.e, eqdist.e, and eqdist.etest, method = c("original", "disco").
A method argument was added to edist, which summarizes cluster distances in a table: method = c("cluster","discoB","discoF"))