,
Klaus Nordhausen [cre, aut]
, Hannu Oja [aut],
Sara Taskinen [aut] | Title: | Asymptotic Covariance Matrices of Some BSS Mixing and Unmixing Matrix Estimates |
|---|---|
| Description: | Functions to compute the asymptotic covariance matrices of mixing and unmixing matrix estimates of the following blind source separation (BSS) methods: symmetric and squared symmetric FastICA, regular and adaptive deflation-based FastICA, FOBI, JADE, AMUSE and deflation-based and symmetric SOBI. Also functions to estimate these covariances based on data are available. |
| Authors: | Jari Miettinen [aut] ,
Klaus Nordhausen [cre, aut]
, Hannu Oja [aut],
Sara Taskinen [aut] |
| Maintainer: | Klaus Nordhausen <[email protected]> |
| License: | GPL (>= 2) |
| Version: | 1.2-3 |
| Built: | 2025-02-19 04:13:32 UTC |
| Source: | https://github.com/cran/BSSasymp |
Description: Functions to compute the asymptotic covariance matrices of mixing and unmixing matrix estimates of the following blind source separation (BSS) methods: symmetric and squared symmetric FastICA, regular and adaptive deflation-based FastICA, FOBI, JADE, AMUSE and deflation-based and symmetric SOBI. Also functions to estimate these covariances based on data are available.
| Package: | BSSasymp |
| Type: | Package |
| Version: | 1.2-3 |
| Date: | 2021-12-10 |
| License: | GPL (>= 2) |
Jari Miettinen, Klaus Nordhausen, Hannu Oja, Sara Taskinen
Maintainer: Klaus Nordhausen <[email protected]>
Miettinen, J., Nordhausen, K. and Taskinen, S. (2017), Blind Source Separation Based on Joint Diagonalization in R: The Packages JADE and BSSasymp, Journal of Statistical Software, 76, 1-31, <doi:10.18637/jss.v076.i02>.
Deflation-based FastICA solves the blind source separation problem in the case of independent components. These function computes some interesting theoretic quantities related to the deflation-based FastICA unmixing matrix estimate, see details.
alphas(sdf, gs, dgs, name=NULL, supp=NULL,...)alphas(sdf, gs, dgs, name=NULL, supp=NULL,...)
sdf |
a list of density functions of the sources scaled so that the mean is 0 and variance is 1. |
gs |
a list of nonlinearity functions. |
dgs |
the first derivative functions of the nonlinearity functions. |
name |
a list of strings, which labels the nonlinearities. |
supp |
a two column matrix, where each row gives the lower and the upper limit used in numerical integration for the corresponding source component which is done using |
... |
arguments to be passed to |
When the mixing matrix is the identity matrix, the asymptotic variances of the first row elements of the deflation-based FastICA estimate depend only on the corresponding source component and the chosen nonlinearity function g. Furthermore, the asymptotic variances of the off-diagonal elements of the first row are equal, let us call this value alpha. Also the other asymptotic variances depend straightforwardly on alphas corresponding to different components and to (possibly) different nonlinearities. Alphas indicate which nonlinearities should be used and in which order the components should be separated. Reloaded (Nordhausen et al., 2011) and adaptive (Miettinen et al., 2014) deflation-based FastICA estimators are based on the estimation of alphas from the data.
A matrix where the ith row gives the alphas for the ith nonlinearity and the jth column corresponds to the jth density in sdf.
Jari Miettinen
Miettinen, J., Nordhausen, K., Oja, H. and Taskinen, S. (2014), Deflation-based FastICA with adaptive choices of nonlinearities, IEEE Transactions on Signal Processing, 62(21), 5716–5724.
Nordhausen, K., Ilmonen, P., Mandal, A., Oja, H. and Ollila, E. (2011), Deflation-based FastICA reloaded, in Proc. "19th European Signal Processing Conference 2011 (EUSIPCO 2011)", Barcelona, 1854–1858.
# source components have uniform- and exponential(1)- distribution fu <- function(x){1/(sqrt(3)*2)} fe <- function(x){exp(-x-1)} supp <- matrix(c(-sqrt(3),sqrt(3),-1,Inf), nrow=2, ncol=2, byrow=TRUE) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} gs <- c(g_pow3,g_gaus) dgs <- c(dg_pow3,dg_gaus) alphas(sdf=c(fu,fe), gs=c(g_pow3), dgs=c(dg_pow3), supp=supp) alphas(sdf=c(fu,fe), gs=gs, dgs=dgs, supp=supp)# source components have uniform- and exponential(1)- distribution fu <- function(x){1/(sqrt(3)*2)} fe <- function(x){exp(-x-1)} supp <- matrix(c(-sqrt(3),sqrt(3),-1,Inf), nrow=2, ncol=2, byrow=TRUE) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} gs <- c(g_pow3,g_gaus) dgs <- c(dg_pow3,dg_gaus) alphas(sdf=c(fu,fe), gs=c(g_pow3), dgs=c(dg_pow3), supp=supp) alphas(sdf=c(fu,fe), gs=gs, dgs=dgs, supp=supp)
The regular deflation-based FastICA finds the independent components one by one using a nonlinearity function. The adaptive deflation-based FastICA chooses, for each component separately, the best nonlinearity from a set of nonlinearities. This function computes asymptotic covariance matrices of the different deflation-based FastICA mixing and unmixing matrix estimates.
ASCOV_FastICAdefl(sdf, gs, dgs, Gs=NULL, method="adapt", name=NULL, supp=NULL, A=NULL, ...)ASCOV_FastICAdefl(sdf, gs, dgs, Gs=NULL, method="adapt", name=NULL, supp=NULL, A=NULL, ...)
sdf |
a list of density functions of the sources scaled so that the mean is 0 and variance is 1. |
gs |
a list of nonlinearity functions. |
dgs |
the first derivative functions of the nonlinearity functions. |
Gs |
the integral function of the nonlinearity function. Is needed only when |
method |
"adapt", "G" or "regular", see details. |
name |
a list of strings, which labels the nonlinearities. |
supp |
a two column matrix, where each row gives the lower and the upper limit used in numerical integration for the corresponding source component which is done using |
A |
the mixing matrix, identity matrix as default. |
... |
arguments to be passed to |
Depending on the argument method, the function computes the asymptotic covariance matrices for three different extraction orders of the independent components. The choice method="adapt" picks the adaptive deflation-based FastICA estimate, which extracts the components in asymptotically optimal order and uses the best nonlinearity from the set of nonlinearities gs.
The other two methods use only one nonlinearity, and if gs and dgs contain more than one function, the first one is taken. When method="G", the order is based on the deviance from normality measured by Gs. When method="regular", the order is that of sdf.
The signs of the components are fixed so that the sum of the elements of each row of the unmixing matrix is positive.
Since the unmixing matrix has asymptotic normal distribution, we have a connection between the asymptotic variances and the minimum distance index, which is defined as
where is the unmixing matrix estimate, is the mixing matrix, is a permutation matrix and a diagonal matrix with nonzero diagonal entries. If converges to the identity matrix, the limiting expected value of is the sum of the asymptotic variances of the off-diagonal elements of . Here is the sample size and is the number of components.
A list with the following components:
W |
mean of the unmixing matrix estimate. |
COV_W |
asymptotic covariance matrix of the unmixing matrix estimate. |
A |
mean of the mixing matrix estimate. |
COV_A |
asymptotic covariance matrix of the mixing matrix estimate. |
EMD |
the limiting expected value of |
used_gs |
indicates which nonlinearity is used in estimation of each rows of the unmixing matrix. |
Jari Miettinen
Ilmonen, P., Nordhausen, K., Oja, H. and Ollila, E. (2010): A New Performance Index for ICA: Properties, Computation and Asymptotic Analysis. In Vigneron, V., Zarzoso, V., Moreau, E., Gribonval, R. and Vincent, E. (editors) Latent Variable Analysis and Signal Separation, 229–236, Springer.
Miettinen, J., Nordhausen, K., Oja, H. and Taskinen, S. (2014), Deflation-based FastICA with adaptive choices of nonlinearities, IEEE Transactions on Signal Processing, 62(21), 5716–5724.
Nordhausen, K., Ilmonen, P., Mandal, A., Oja, H. and Ollila, E. (2011), Deflation-based FastICA reloaded, in Proc. "19th European Signal Processing Conference 2011 (EUSIPCO 2011)", Barcelona, 1854–1858.
ASCOV_FastICAdefl_est, adapt_fICA, integrate
# source components have uniform- and exponential(1)- distribution fu <- function(x){1/(sqrt(3)*2)} fe <- function(x){exp(-x-1)} supp <- matrix(c(-sqrt(3),sqrt(3),-1,Inf), nrow=2, ncol=2, byrow=TRUE) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} G_pow3 <- function(x){x^4/4} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} gs <- c(g_pow3,g_gaus) dgs <- c(dg_pow3,dg_gaus) A <- matrix(rnorm(4),2,2) res1 <- ASCOV_FastICAdefl(sdf=c(fu,fe), gs=gs, dgs=dgs, supp=supp, A=A) round(res1$COV_W, 2) res1$EMD res1$used_gs res2 <- ASCOV_FastICAdefl(sdf=c(fu,fe), gs=c(g_pow3), dgs=c(dg_pow3), Gs=c(G_pow3), method="G", supp=supp, A=A) res2$EMD# source components have uniform- and exponential(1)- distribution fu <- function(x){1/(sqrt(3)*2)} fe <- function(x){exp(-x-1)} supp <- matrix(c(-sqrt(3),sqrt(3),-1,Inf), nrow=2, ncol=2, byrow=TRUE) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} G_pow3 <- function(x){x^4/4} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} gs <- c(g_pow3,g_gaus) dgs <- c(dg_pow3,dg_gaus) A <- matrix(rnorm(4),2,2) res1 <- ASCOV_FastICAdefl(sdf=c(fu,fe), gs=gs, dgs=dgs, supp=supp, A=A) round(res1$COV_W, 2) res1$EMD res1$used_gs res2 <- ASCOV_FastICAdefl(sdf=c(fu,fe), gs=c(g_pow3), dgs=c(dg_pow3), Gs=c(G_pow3), method="G", supp=supp, A=A) res2$EMD
The regular deflation-based FastICA finds the independent components one by one using a nonlinearity function. The adaptive deflation-based FastICA chooses, for each component separately, the best nonlinearity from a set of nonlinearities. This function computes estimates of the covariance matrices of the different deflation-based FastICA mixing and unmixing matrix estimates.
ASCOV_FastICAdefl_est(X, gs, dgs, Gs=NULL, method="adapt", name=NULL, mixed=TRUE)ASCOV_FastICAdefl_est(X, gs, dgs, Gs=NULL, method="adapt", name=NULL, mixed=TRUE)
X |
a numeric data matrix. |
gs |
a list of nonlinearity functions. |
dgs |
the first derivative functions of the nonlinearity functions. |
Gs |
the integral function of the nonlinearity function. Is needed only when |
method |
"adapt" or "G", see details. |
name |
a list of strings, which labels the nonlinearities. |
mixed |
logical, see details. |
Depending on the argument method, the function computes the asymptotic covariance matrices for two different extraction orders of the independent components. The choice method="adapt" picks the adaptive deflation-based FastICA estimate, which extracts the components in asymptotically optimal order and uses the best nonlinearity from the set of nonlinearities gs.
When method="G", the order is based on the deviance from normality measured by Gs. This method uses only one nonlinearity, and if gs and dgs contain more than one function, the first one is taken.
If mixed is TRUE, then X will be transformed by the adaptive FastICA estimate. The option FALSE can be used, for example, to estimate the covariance when X are source estimates given by some other method than FastICA.
A list with the following components:
W |
estimated mean of the unmixing matrix estimate. |
COV_W |
estimated covariance matrix of the unmixing matrix estimate. |
A |
estimated mean of the mixing matrix estimate. |
COV_A |
estimated covariance matrix of the mixing matrix estimate. |
used_gs |
indicates which nonlinearity is used in estimation of each rows of the unmixing matrix. |
Jari Miettinen
Miettinen, J., Nordhausen, K., Oja, H. and Taskinen, S. (2014), Deflation-based FastICA with adaptive choices of nonlinearities, IEEE Transactions on Signal Processing, 62(21), 5716–5724.
Nordhausen, K., Ilmonen, P., Mandal, A., Oja, H. and Ollila, E. (2011), Deflation-based FastICA reloaded, in Proc. "19th European Signal Processing Conference 2011 (EUSIPCO 2011)", Barcelona, 1854–1858.
# source components have uniform- and exponential(1)- distribution s1 <- runif(1000,-sqrt(3),sqrt(3)) s2 <- rexp(1000) S <- cbind(s1,s2) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} G_pow3 <- function(x){x^4/4} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} gs <- c(g_pow3,g_gaus) dgs <- c(dg_pow3,dg_gaus) A<-matrix(rnorm(4),2,2) X <- S %*% t(A) round(1000*ASCOV_FastICAdefl_est(X, gs=gs, dgs=dgs)$COV_W,2) round(1000*ASCOV_FastICAdefl_est(X, gs=c(g_pow3), dgs=c(dg_pow3), Gs=c(G_pow3), method="G")$COV_W,2)# source components have uniform- and exponential(1)- distribution s1 <- runif(1000,-sqrt(3),sqrt(3)) s2 <- rexp(1000) S <- cbind(s1,s2) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} G_pow3 <- function(x){x^4/4} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} gs <- c(g_pow3,g_gaus) dgs <- c(dg_pow3,dg_gaus) A<-matrix(rnorm(4),2,2) X <- S %*% t(A) round(1000*ASCOV_FastICAdefl_est(X, gs=gs, dgs=dgs)$COV_W,2) round(1000*ASCOV_FastICAdefl_est(X, gs=c(g_pow3), dgs=c(dg_pow3), Gs=c(G_pow3), method="G")$COV_W,2)
Symmetric FastICA estimators solves the blind source separation problem in the case of independent components. These functions compute the asymptotic covariance matrices of the regular and the squared symmetric FastICA mixing and unmixing matrix estimates.
ASCOV_FastICAsym(sdf, G, g, dg, supp=NULL, A=NULL, ...) ASCOV_FastICAsym2(sdf, G, g, dg, supp=NULL, A=NULL, ...)ASCOV_FastICAsym(sdf, G, g, dg, supp=NULL, A=NULL, ...) ASCOV_FastICAsym2(sdf, G, g, dg, supp=NULL, A=NULL, ...)
sdf |
a list of density functions of the sources scaled so that the mean is 0 and variance is 1. |
G |
the integral function of the nonlinearity function, see details. |
g |
the nonlinearity function. |
dg |
the first derivative function of the nonlinearity function. |
supp |
a two column matrix, where each row gives the lower and the upper limit used in numerical integration for the corresponding source component which is done using |
A |
the mixing matrix, identity matrix as default. |
... |
arguments to be passed to |
The signs of the components are fixed so that the sum of the elements of each row of the unmixing matrix is positive.
Since the unmixing matrix has asymptotic normal distribution, we have a connection between the asymptotic variances and the minimum distance index, which is defined as
where is the unmixing matrix estimate, is the mixing matrix, is a permutation matrix and a diagonal matrix with nonzero diagonal entries. If converges to the identity matrix, the limiting expected value of is the sum of the asymptotic variances of the off-diagonal elements of . Here is the sample size and is the number of components.
A list with the following components:
W |
mean of the unmixing matrix estimate. |
COV_W |
asymptotic covariance matrix of the unmixing matrix estimate. |
A |
mean of the mixing matrix estimate. |
COV_A |
asymptotic covariance matrix of the mixing matrix estimate. |
EMD |
the limiting expected value of |
Jari Miettinen
Hyv\"arinen, A. (1999), Fast and robust fixed-point algorithms for independent component analysis, IEEE Transactions on Neural Networks, 10, 626-634.
Wei, T. (2014), The convergence and asymptotic analysis of the generalized symmetric FastICA algorithm, http://arxiv.org/abs/1408.0145.
Miettinen, J., Nordhausen, K., Oja, H., Taskinen, S. and Virta, J. (2015), The squared symmetric FastICA estimator, submitted.
ASCOV_FastICAsym_est, fICA, integrate
# source components have uniform- and exponential(1)- distribution fu <- function(x){1/(sqrt(3)*2)} fe <- function(x){exp(-x-1)} supp <- matrix(c(-sqrt(3),sqrt(3),-1,Inf), nrow=2, ncol=2, byrow=TRUE) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} G_pow3 <- function(x){x^4/4} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} G_gaus <- function(x){-exp(-x^2/2)} A <- matrix(rnorm(4),2,2) res1 <- ASCOV_FastICAsym(sdf=c(fu,fe), G=G_pow3, g=g_pow3, dg=dg_pow3, supp=supp, A=A) res2 <- ASCOV_FastICAsym(sdf=c(fu,fe), G=G_gaus, g=g_gaus, dg=dg_gaus, supp=supp, A=A) round(res1$COV_W, 2) res1$EMD round(res2$COV_W, 2) res2$EMD# source components have uniform- and exponential(1)- distribution fu <- function(x){1/(sqrt(3)*2)} fe <- function(x){exp(-x-1)} supp <- matrix(c(-sqrt(3),sqrt(3),-1,Inf), nrow=2, ncol=2, byrow=TRUE) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} G_pow3 <- function(x){x^4/4} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} G_gaus <- function(x){-exp(-x^2/2)} A <- matrix(rnorm(4),2,2) res1 <- ASCOV_FastICAsym(sdf=c(fu,fe), G=G_pow3, g=g_pow3, dg=dg_pow3, supp=supp, A=A) res2 <- ASCOV_FastICAsym(sdf=c(fu,fe), G=G_gaus, g=g_gaus, dg=dg_gaus, supp=supp, A=A) round(res1$COV_W, 2) res1$EMD round(res2$COV_W, 2) res2$EMD
Symmetric FastICA solves the blind source separation problem in the case of independent components. This function computes an estimate of the covariance matrix of symmetric FastICA mixing and unmixing matrix estimates.
ASCOV_FastICAsym_est(X, G, g, dg, mixed=TRUE) ASCOV_FastICAsym2_est(X, G, g, dg, mixed=TRUE)ASCOV_FastICAsym_est(X, G, g, dg, mixed=TRUE) ASCOV_FastICAsym2_est(X, G, g, dg, mixed=TRUE)
X |
a numeric data matrix. |
G |
the integral function of the nonlinearity function. |
g |
the nonlinearity function. |
dg |
the first derivative function of the nonlinearity function. |
mixed |
logical, see details. |
If mixed is TRUE, then X will be transformed by the symmetric FastICA estimate. The option FALSE can be used, for example, to estimate the covariance when X are source estimates given by some other method than FastICA.
A list with the following components:
W |
estimated mean of the unmixing matrix estimate. |
COV_W |
estimated covariance matrix of the unmixing matrix estimate. |
A |
estimated mean of the mixing matrix estimate. |
COV_A |
estimated covariance matrix of the mixing matrix estimate. |
Jari Miettinen
Hyv\"arinen, A. (1999), Fast and robust fixed-point algorithms for independent component analysis, IEEE Transactions on Neural Networks, 10, 626-634.
Wei, T. (2014), The convergence and asymptotic analysis of the generalized symmetric FastICA algorithm, http://arxiv.org/abs/1408.0145.
# source components have uniform- and exponential(1)- distribution n <- 1000 s1 <- runif(n,-sqrt(3),sqrt(3)) s2 <- rexp(n) S <- cbind(s1,s2) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} G_pow3 <- function(x){x^4/4} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} G_gaus <- function(x){-exp(-x^2/2)} A<-matrix(rnorm(4),2,2) X <- S %*% t(A) round(n*ASCOV_FastICAsym_est(X,G=G_pow3,g=g_pow3,dg=dg_pow3)$COV_W,2) round(n*ASCOV_FastICAsym_est(X,G=G_gaus,g=g_gaus,dg=dg_gaus)$COV_W,2)# source components have uniform- and exponential(1)- distribution n <- 1000 s1 <- runif(n,-sqrt(3),sqrt(3)) s2 <- rexp(n) S <- cbind(s1,s2) # couple of nonlinearities g_pow3 <- function(x){x^3} dg_pow3 <- function(x){3*x^2} G_pow3 <- function(x){x^4/4} g_gaus <- function(x){x*exp(-x^2/2)} dg_gaus <- function(x){exp(-x^2/2)-x^2*exp(-x^2/2)} G_gaus <- function(x){-exp(-x^2/2)} A<-matrix(rnorm(4),2,2) X <- S %*% t(A) round(n*ASCOV_FastICAsym_est(X,G=G_pow3,g=g_pow3,dg=dg_pow3)$COV_W,2) round(n*ASCOV_FastICAsym_est(X,G=G_gaus,g=g_gaus,dg=dg_gaus)$COV_W,2)
JADE solves the blind source separation problem in the case of independent components with at most one component having kurtosis values zero, while FOBI requires distinct kurtosis values. The functions compute the asymptotic covariance matrices of JADE and FOBI estimates for the mixing or the unmixing matrices.
ASCOV_JADE(sdf, supp=NULL, A=NULL, ...) ASCOV_FOBI(sdf, supp=NULL, A=NULL, ...)ASCOV_JADE(sdf, supp=NULL, A=NULL, ...) ASCOV_FOBI(sdf, supp=NULL, A=NULL, ...)
sdf |
a vector of density functions of the sources scaled so that the mean is 0 and variance is 1. |
supp |
a two column matrix, where each row gives the lower and the upper limit used in numerical integration for the corresponding source component which is done using |
A |
the mixing matrix, identity matrix as default. |
... |
arguments to be passed to |
The order of the estimated components is fixed so that their kurtosis values are in a decreasing order. The signs of the components is fixed so that the sum of the elements of each row of the unmixing matrix is positive.
Since the unmixing matrix has asymptotic normal distribution, we have a connection between the asymptotic variances and the minimum distance index, which is defined as
where is the unmixing matrix estimate, is the mixing matrix, is a permutation matrix and a diagonal matrix with nonzero diagonal entries. If converges to the identity matrix, the limiting expected value of is the sum of the asymptotic variances of the off-diagonal elements of . Here is the sample size and is the number of components.
A list with the following components:
W |
mean of the unmixing matrix estimate. |
COV_W |
asymptotic covariance matrix of the unmixing matrix estimate. |
A |
mean of the mixing matrix estimate. |
COV_A |
asymptotic covariance matrix of the mixing matrix estimate. |
EMD |
The limiting expected value of |
Jari Miettinen
Ilmonen, P., Nevalainen, J. and Oja, H. (2010), Characteristics of multivariate distributions and the invariant coordinate system, Statistics and Probability Letters, 80, 1844–1853.
Ilmonen, P., Nordhausen, K., Oja, H. and Ollila, E. (2010), A New Performance Index for ICA: Properties, Computation and Asymptotic Analysis. In Vigneron, V., Zarzoso, V., Moreau, E., Gribonval, R. and Vincent, E. (editors) Latent Variable Analysis and Signal Separation, 229–236, Springer.
Miettinen, J., Taskinen S., Nordhausen, K. and Oja, H. (2015), Fourth Moments and Independent Component Analysis, Statistical Science, 30, 372–390.
ASCOV_JADE_est, ASCOV_FOBI_est, JADE, FOBI
# source components have uniform- and exponential(1)- distribution fu <- function(x){1/(sqrt(3)*2)} fe <- function(x){exp(-x-1)} supp <- matrix(c(-sqrt(3),sqrt(3),-1,Inf), nrow=2, ncol=2, byrow=TRUE) A<-matrix(rnorm(4),2,2) jade <- ASCOV_JADE(sdf=c(fu,fe), supp=supp, A=A) fobi <- ASCOV_FOBI(sdf=c(fu,fe), supp=supp, A=A) round(jade$COV_W,2) round(fobi$COV_W,2) jade$EMD fobi$EMD# source components have uniform- and exponential(1)- distribution fu <- function(x){1/(sqrt(3)*2)} fe <- function(x){exp(-x-1)} supp <- matrix(c(-sqrt(3),sqrt(3),-1,Inf), nrow=2, ncol=2, byrow=TRUE) A<-matrix(rnorm(4),2,2) jade <- ASCOV_JADE(sdf=c(fu,fe), supp=supp, A=A) fobi <- ASCOV_FOBI(sdf=c(fu,fe), supp=supp, A=A) round(jade$COV_W,2) round(fobi$COV_W,2) jade$EMD fobi$EMD
JADE solves the blind source separation problem in the case of independent components with at most one component having kurtosis values zero, while FOBI requires distinct kurtosis values. The functions compute the asymptotic covariance matrices of JADE and FOBI estimates for the mixing or the unmixing matrix.
ASCOV_JADE_est(X, mixed=TRUE) ASCOV_FOBI_est(X, mixed=TRUE)ASCOV_JADE_est(X, mixed=TRUE) ASCOV_FOBI_est(X, mixed=TRUE)
X |
a numeric data matrixx. |
mixed |
logical, see details. |
If mixed is TRUE, then X will be transformed by the corresponding estimate. The option FALSE can be used, for example, to estimate the covariance when X are source estimates given by some other method than JADE or FOBI.
A list with the following components:
W |
estimated mean of the unmixing matrix estimate. |
COV_W |
estimated covariance matrix of the unmixing matrix estimate. |
A |
estimated mean of the mixing matrix estimate. |
COV_A |
estimated covariance matrix of the mixing matrix estimate. |
Jari Miettinen
Ilmonen, P., Nevalainen, J. and Oja, H. (2010), Characteristics of multivariate distributions and the invariant coordinate system, Statistics and Probability Letters, 80, 1844–1853.
Miettinen, J., Taskinen S., Nordhausen, K. and Oja, H. (2015), Fourth Moments and Independent Component Analysis, Statistical Science, 30, 372–390.
ASCOV_JADE, ASCOV_FOBI, JADE, FOBI
# source components have t-10-, uniform- and gaussian distribution s1 <- rt(1000,10)/sqrt(10/8) s2 <- runif(1000,-sqrt(3),sqrt(3)) s3 <- rnorm(1000) S <- cbind(s1,s2,s3) A <- matrix(rnorm(9),3,3) X <- S %*% t(A) round(1000*ASCOV_JADE_est(X)$COV_W,2) round(1000*ASCOV_FOBI_est(X)$COV_W,2)# source components have t-10-, uniform- and gaussian distribution s1 <- rt(1000,10)/sqrt(10/8) s2 <- runif(1000,-sqrt(3),sqrt(3)) s3 <- rnorm(1000) S <- cbind(s1,s2,s3) A <- matrix(rnorm(9),3,3) X <- S %*% t(A) round(1000*ASCOV_JADE_est(X)$COV_W,2) round(1000*ASCOV_FOBI_est(X)$COV_W,2)
The symmetric and deflation-based SOBI methods solve the blind source separation problem in the case of second order stationary time series sources by jointly diagonalizing the covariance matrix and several autocovariance matrices at different lags. The functions compute the asymptotic covariance matrices of a SOBI estimates for the mixing or the unmixing matrices, when the sources are time series. Notice that, since AMUSE method is a special case of SOBI, also the asymptotic covariance matrix of an AMUSE estimate can be computed using these functions.
ASCOV_SOBI(psi, taus, a=2, Beta=NULL, A=NULL) ASCOV_SOBIdefl(psi, taus, Beta=NULL, A=NULL)ASCOV_SOBI(psi, taus, a=2, Beta=NULL, A=NULL) ASCOV_SOBIdefl(psi, taus, Beta=NULL, A=NULL)
psi |
a numeric matrix containing the MA-coefficients of the time series, see details. |
taus |
a vector of integers for the lags. |
a |
numeric, see details. |
Beta |
a matrix of fourth moments of the innovations, see details. |
A |
the mixing matrix, identity matrix as default. |
Naturally, the function can deal with only finite number of coefficients. The larger is the number of the rows of psi, the more accurate is the result, but also the longer is the computation time.
AR or ARMA coefficients can be transformed to MA coefficients by using ARMAtoMA.
The th entry of Beta is , where are the innovations of th source component such that and .
The order of the estimated components is fixed so that the sums of their squared autocovariances over taus are in a decreasing order. The signs of the components are fixed so that the sum of the elements of each row of the unmixing matrix is positive.
Since the unmixing matrix has asymptotic normal distribution, we have a connection between the asymptotic variances and the minimum distance index, which is defined as
where is the unmixing matrix estimate, is the mixing matrix, is a permutation matrix and a diagonal matrix with nonzero diagonal entries. If converges to the identity matrix, the limiting expected value of is the sum of the asymptotic variances of the off-diagonal elements of . Here is the sample size and is the number of components.
The symmetric SOBI estimator maximizes the sum of squares of the diagonal elements of the autocovariance matrices. Different SOBI estimators are obtained when the diagonality of matrices is measured by
with . The diagonality measure can be selected using the argument a.
A list with the following components:
W |
mean of the unmixing matrix estimate. |
COV_W |
asymptotic covariance matrix of the unmixing matrix estimate. |
A |
mean of the mixing matrix estimate. |
COV_A |
asymptotic covariance matrix of the mixing matrix estimate. |
EMD |
The limiting expected value of |
Jari Miettinen
Ilmonen, P., Nordhausen, K., Oja, H. and Ollila, E. (2010): A New Performance Index for ICA: Properties, Computation and Asymptotic Analysis. In Vigneron, V., Zarzoso, V., Moreau, E., Gribonval, R. and Vincent, E. (editors) Latent Variable Analysis and Signal Separation, 229–236, Springer.
Miettinen, J. (2015): Alternative diagonality criteria for SOBI. In Nordhausen, K. and Taskinen, S. (editors) Modern Nonparametric, Robust and Multivariate methods, Festschrift in Honour of Hannu Oja, 455–469, Springer.
Miettinen, J., Nordhausen, K., Oja, H. and Taskinen, S. (2012), Statistical properties of a blind source separation estimator for stationary time series, Statistics and Probability Letters, 82, 1865–1873.
Miettinen, J., Nordhausen, K., Oja, H. and Taskinen, S. (2014), Deflation-based separation of uncorrelated stationary time series, Journal of Multivariate Analysis, 123, 214–227.
Miettinen, J., Illner, K., Nordhausen, K., Oja, H., Taskinen, S. and Theis, F. (2016), Separation of uncorrelated stationary time series using autocovariance matrices, Journal of Time Series Analysis, 37, 337–354.
ASCOV_SOBI_est, SOBI, AMUSE, ARMAtoMA
A<- matrix(rnorm(9),3,3) psi1 <- ARMAtoMA(ar=0.6, lag.max=100) psi2 <- ARMAtoMA(ar=c(0.2,0.3,-0.3),lag.max=100) psi3 <- ARMAtoMA(ar=-0.2, ma=c(0.5,-0.1,0.4), lag.max=100) psi <- cbind(psi1,psi2,psi3) sym <- ASCOV_SOBI(psi=psi, taus=1:10, A=A) defl <- ASCOV_SOBIdefl(psi=psi, taus=1:10, A=A) round(sym$COV_W,2) round(defl$COV_W,2) sym$EMD defl$EMDA<- matrix(rnorm(9),3,3) psi1 <- ARMAtoMA(ar=0.6, lag.max=100) psi2 <- ARMAtoMA(ar=c(0.2,0.3,-0.3),lag.max=100) psi3 <- ARMAtoMA(ar=-0.2, ma=c(0.5,-0.1,0.4), lag.max=100) psi <- cbind(psi1,psi2,psi3) sym <- ASCOV_SOBI(psi=psi, taus=1:10, A=A) defl <- ASCOV_SOBIdefl(psi=psi, taus=1:10, A=A) round(sym$COV_W,2) round(defl$COV_W,2) sym$EMD defl$EMD
The symmetric and deflation-based SOBI methods solve the blind source separation problem in the case of second order stationary time series sources by jointly diagonalizing the covariance matrix and several autocovariance matrices at different lags. The functions compute an estimate of the covariance matrix of a SOBI estimate for the mixing or the unmixing matrix, under the assumption that the sources are time series. Notice that, since AMUSE method is a special case of SOBI, also an estimate of the covariance matrix of an AMUSE estimate can be computed using these functions.
ASCOV_SOBI_estN(X, taus, mixed=TRUE, M=100, a=2) ASCOV_SOBI_est(X, taus, arp=NULL, maq=NULL, mixed=TRUE, M=100, a=2, ...) ASCOV_SOBIdefl_estN(X, taus, mixed=TRUE, M=100) ASCOV_SOBIdefl_est(X, taus, arp=NULL, maq=NULL, mixed=TRUE, M=100, ...)ASCOV_SOBI_estN(X, taus, mixed=TRUE, M=100, a=2) ASCOV_SOBI_est(X, taus, arp=NULL, maq=NULL, mixed=TRUE, M=100, a=2, ...) ASCOV_SOBIdefl_estN(X, taus, mixed=TRUE, M=100) ASCOV_SOBIdefl_est(X, taus, arp=NULL, maq=NULL, mixed=TRUE, M=100, ...)
X |
a numeric data matrix or a multivariate time series object of class |
taus |
a vector of integers for the lags. |
arp |
a vector containing the AR orders used for the estimation of ARMA coefficients. |
maq |
a vector containing the MA orders used for the estimation of ARMA coefficients. |
mixed |
logical, see details. |
M |
the number of autocovariance matrices used for the estimation of the covariance matrices, see details. |
a |
numeric, see details |
... |
arguments to be passed to |
Functions ASCOV_SOBI_estN and ASCOV_SOBIdefl_estN assume that the innovations of the components are gaussian. Therefore, they are faster than ASCOV_SOBI_est and ASCOV_SOBI-defl_est, which estimate the fourth moments of the innovations by estimating the ARMA coefficients of the time series.
Fitting the univariate ARMA coefficients is done using the function arima based on the orders provided by arp and maq.
The estimation is mostly based on autocovariance matrices and all non-zero matrices should be included. On the other hand, too large value of M increases the computation time and it may even reduce the estimation accuracy.
If mixed is TRUE, then X will be transformed by the corresponding SOBI estimate. The option FALSE can be used, for example, to estimate the covariance when X are source estimates given by some other method than SOBI.
The symmetric SOBI estimator maximizes the sum of squares of the diagonal elements of the autocovariance matrices. Different SOBI estimators are obtained when the diagonality of matrices is measured by
with . The diagonality measure can be selected using the argument a.
A list with the following components:
W |
estimated mean of the unmixing matrix estimate. |
COV_W |
estimated covariance matrix of the unmixing matrix estimate. |
A |
estimated mean of the mixing matrix estimate. |
COV_A |
estimated covariance matrix of the mixing matrix estimate. |
Jari Miettinen
Miettinen, J. (2015): Alternative diagonality criteria for SOBI. In Nordhausen, K. and Taskinen, S. (editors), Modern Nonparametric, Robust and Multivariate methods, Festschrift in Honour of Hannu Oja, 455–469, Springer.
Miettinen, J., Nordhausen, K., Oja, H. and Taskinen, S. (2012), Statistical properties of a blind source separation estimator for stationary time series, Statistics and Probability Letters, 82, 1865–1873.
Miettinen, J., Nordhausen, K., Oja, H. and Taskinen, S. (2014), Deflation-based separation of uncorrelated stationary time series, Journal of Multivariate Analysis, 123, 214–227.
Miettinen, J., Illner, K., Nordhausen, K., Oja, H., Taskinen, S. and Theis, F. (2015), Separation of uncorrelated stationary time series using autocovariance matrices, Journal of Time Series Analysis, in print, DOI: 10.1111/jtsa.12159.
ASCOV_SOBI, SOBI, AMUSE, arima
A<- matrix(rnorm(9),3,3) s1 <- arima.sim(list(ar=0.6),1000) s2 <- arima.sim(list(ma=c(0.2,0.3,-0.3)),1000) s3 <- arima.sim(list(ar=-0.2,ma=c(0.5,-0.1,0.4)),1000) S <- cbind(s1,s2,s3) X <- S %*% t(A) round(1000*ASCOV_SOBI_estN(X, taus=1:10)$COV_W,2) round(1000*ASCOV_SOBIdefl_estN(X, taus=1:10)$COV_W,2)A<- matrix(rnorm(9),3,3) s1 <- arima.sim(list(ar=0.6),1000) s2 <- arima.sim(list(ma=c(0.2,0.3,-0.3)),1000) s3 <- arima.sim(list(ar=-0.2,ma=c(0.5,-0.1,0.4)),1000) S <- cbind(s1,s2,s3) X <- S %*% t(A) round(1000*ASCOV_SOBI_estN(X, taus=1:10)$COV_W,2) round(1000*ASCOV_SOBIdefl_estN(X, taus=1:10)$COV_W,2)
The SOBI method solves the blind source separation problem in the case of second order stationary time series sources by jointly diagonalizing the covariance matrix and several autocovariance matrices. In the classical SOBI method, the sum of squares of the diagonal elements is used as diagonality measure. This function computes the SOBI estimate, when a choice from a family of alternative diagonality criteria is used.
aSOBI(X, k=12, a=4, eps=1e-06, maxiter=1000)aSOBI(X, k=12, a=4, eps=1e-06, maxiter=1000)
X |
a numeric data matrix or a multivariate time series object of class |
k |
if a single integer, then the lags 1:k are used, if an integer vector, then these are used as the lags. |
a |
numeric, determines the diagonality criterion, see details. |
eps |
convergence tolerance. |
maxiter |
maximum number of iterations. |
The classical SOBI estimator maximizes the sum of squares of the diagonal elements of the autocovariance matrices. A family of alternative SOBI estimators is obtained when the diagonality of matrices is measured by
with .
A list with class 'bss' containing the following components:
W |
estimated unmixing matrix. |
S |
estimated sources standardized to have mean 0 and unit variances. |
k |
lags used. |
a |
value of the diagonality criterion parameter used. |
Jari Miettinen
Miettinen, J. (2015): Alternative diagonality criteria for SOBI. In Nordhausen, K. and Taskinen, S. (editors), Modern Nonparametric, Robust and Multivariate methods, Festschrift in Honour of Hannu Oja, 455–469, Springer.
A <- matrix(rnorm(9),3,3) s1 <- arima.sim(list(ar=0.6),1000) s2 <- arima.sim(list(ma=c(0.2,0.3,-0.3)),1000) s3 <- arima.sim(list(ar=-0.2,ma=c(0.5,-0.1,0.4)),1000) S <- cbind(s1,s2,s3) X <- S %*% t(A) aSOBI(X, a=3)A <- matrix(rnorm(9),3,3) s1 <- arima.sim(list(ar=0.6),1000) s2 <- arima.sim(list(ma=c(0.2,0.3,-0.3)),1000) s3 <- arima.sim(list(ar=-0.2,ma=c(0.5,-0.1,0.4)),1000) S <- cbind(s1,s2,s3) X <- S %*% t(A) aSOBI(X, a=3)
Cramer-Rao bound for the unmixing matrix estimate in the independent component model.
CRB(sdf,supp=NULL,A=NULL,eps=1e-04,...)CRB(sdf,supp=NULL,A=NULL,eps=1e-04,...)
sdf |
a list of density functions of the sources scaled so that the mean is 0 and variance is 1. |
supp |
a two column matrix, where each row gives the lower and the upper limit used in numerical integration for the corresponding source component which is done using |
A |
the mixing matrix, identity by default. |
eps |
a value which is used when the derivative functions of the density functions are approximated. |
... |
arguments to be passed to |
Let denote an unmixing matrix estimate. If the estimate is affine equivariant, then the matrix does not depend on the mixing matrix and the estimated independent components are , where is the matrix of the true independent components.
A list containing the following components:
CRLB |
A matrix whose elements give the Cramer-Rao lower bounds for the asymptotic variances of the corresponding elements of |
FIM |
The Fisher information matrix. |
EMD |
The sum of the Cramer-Rao lower bounds of the off-diagonal elements of |
Jari Miettinen
Ollila, E., Kim, H. J. and Koivunen, V. (2008), Compact Cramer-Rao bound expression for independent component analysis. IEEE Transactions on Signal Processing, 56(4), 1421–1428.
# source components have t(9)- and Gaussian distribution f1<-function(x) { gamma(5)*(1+(x*sqrt(9/7))^2/9)^(-5)/ (sqrt(9*pi/(9/7))*gamma(9/2)) } f2<-function(x) { exp(-(x)^2/2)/sqrt(2*pi) } CRB(sdf=c(f1,f2))# source components have t(9)- and Gaussian distribution f1<-function(x) { gamma(5)*(1+(x*sqrt(9/7))^2/9)^(-5)/ (sqrt(9*pi/(9/7))*gamma(9/2)) } f2<-function(x) { exp(-(x)^2/2)/sqrt(2*pi) } CRB(sdf=c(f1,f2))
The SOBI method solves the blind source separation problem in the case of second order stationary time series sources by jointly diagonalizing the covariance matrix and several autocovariance matrices. The separation performance depends on the lag set. The efficient SOBI estimator uses asymptotic results to estimate the variances of the elements of the SOBI unmixing matrices obtained by given lag sets. The unmixing matrix corresponding to the lag set which minimizes the sum of the variances is the efficient SOBI estimate.
eSOBI(X, taus=taus_def, M=200, fast=TRUE, eps=1e-06, maxiter=1000)eSOBI(X, taus=taus_def, M=200, fast=TRUE, eps=1e-06, maxiter=1000)
X |
a numeric data matrix or a multivariate time series object of class |
taus |
a list whose components are vectors of integers. The list gives the set of lag sets. The default set is |
M |
the number of autocovariance matrices used for the estimation of the variance estimates, see |
fast |
logical, see details. |
eps |
convergence tolerance. |
maxiter |
maximum number of iterations. |
The function includes two versions of the efficient SOBI algorithm. The faster version uses only one SOBI estimate given by the first component in taus to estimate the sum of the limiting variances for all lag sets. In the other version, which is obtained by fast=FALSE, the sum of the limiting variances of each lag set is estimated using that particular lag set for the initial SOBI estimate. When the length of the time series is sufficient, say 5000 or more, the two versions yield equally good estimates and the use of the faster version is recommended. Otherwise we recommend the use of the slower version. The variance estimates are based on asymptotic results which assume that the time series are multivariate linear processes. Such processes include a wide class of stationary time series, for example all causal ARMA processes. It is also assumed that the innovations are Gaussian. This simplifies the computations and has practically no effect on which lag set is chosen. If the user does not want to make the Gaussianity assumption, the slower version of the function can be easily modified by replacing the function ASCOV_SOBI_estN in the code by ASCOV_SOBI_est. If the SOBI algorithm fails to converge for some lag set, the corresponding value of the estimated sum of variances in sum_var is Inf.
A list with class 'bss' containing the following components:
W |
estimated unmixing matrix. |
S |
estimated sources standardized to have mean 0 and unit variances. |
taus_used |
the lag set which is considered best and used for the estimation of the unmixing matrix. |
sum_var |
estimated sum of variances of the unmixing matrix estimates for all lag sets given in |
Jari Miettinen
Miettinen, J., Illner, K., Nordhausen, K., Oja, H., Taskinen, S. and Theis, F. (2015), Separation of uncorrelated stationary time series using autocovariance matrices, Journal of Time Series Analysis, in print, http://arxiv.org/abs/1405.3388.
Taskinen, S., Miettinen, J. and Nordhausen, K. (2016), Efficient second order blind identification method for separation of uncorrelated stationary time series, submitted.
A <- matrix(rnorm(9),3,3) s1 <- arima.sim(list(ar=0.6),1000) s2 <- arima.sim(list(ma=c(0.2,0.3,-0.3)),1000) s3 <- arima.sim(list(ar=-0.2,ma=c(0.5,-0.1,0.4)),1000) S <- cbind(s1,s2,s3) X <- S %*% t(A) taus <- list(1,1:2,1:5,1:12) eSOBI(X, taus=taus)A <- matrix(rnorm(9),3,3) s1 <- arima.sim(list(ar=0.6),1000) s2 <- arima.sim(list(ma=c(0.2,0.3,-0.3)),1000) s3 <- arima.sim(list(ar=-0.2,ma=c(0.5,-0.1,0.4)),1000) S <- cbind(s1,s2,s3) X <- S %*% t(A) taus <- list(1,1:2,1:5,1:12) eSOBI(X, taus=taus)
The set of lag sets to be used as the default set in function eSOBI.
taus_deftaus_def
The object taus_def contains the following lag sets:
| taus_def[[1] | 1,2,...,12 |
| taus_def[[2]] | 1 |
| taus_def[[3]] | 1,2 |
| taus_def[[4]] | 1,2,3 |
| taus_def[[5]] | 1,2,3,4,5 |
| taus_def[[6]] | 1,2,...,8 |
| taus_def[[7]] | 1,2,...,20 |
| taus_def[[8]] | 1,2,...,50 |
| taus_def[[9]] | 1,2,...,10,12,...,20 |
| taus_def[[10]] | 5,6,...,10,12,...,20,25,...,50 |
| taus_def[[11]] | 2,4,...,20 |
| taus_def[[12]] | 1,2,3,5,7,11,13,17,19 |
Jari Miettinen
Taskinen, S., Miettinen, J. and Nordhausen, K. (2016), Efficient second order blind identification method for separation of uncorrelated stationary time series, submitted.