Package 'ADMM' reference manual

Title:	Algorithms using Alternating Direction Method of Multipliers
Description:	Provides algorithms to solve popular optimization problems in statistics such as regression or denoising based on Alternating Direction Method of Multipliers (ADMM). See Boyd et al (2010) <doi:10.1561/2200000016> for complete introduction to the method.
Authors:	Kisung You [aut, cre] , Xiaozhi Zhu [aut]
Maintainer:	Kisung You <[email protected]>
License:	GPL (>= 3)
Version:	0.3.3
Built:	2024-11-05 06:41:34 UTC
Source:	CRAN

ADMM : Algorithms using Alternating Direction Method of Multipliers

Description

An introduction of Alternating Direction Method of Multipliers (ADMM) method has been a breakthrough in solving complex and non-convex optimization problems in a reasonably stable as well as scalable fashion. Our package aims at providing handy tools for fast computation on well-known problems using the method. For interested users/readers, please visit Prof. Stephen Boyd's website entirely devoted to the topic.

Basis Pursuit

Description

For an underdetermined system, Basis Pursuit aims to find a sparse solution that solves

$\textrm{min}_x ~ \|x\|_1 \quad \textrm{s.t} \quad Ax=b$

which is a relaxed version of strict non-zero support finding problem. The implementation is borrowed from Stephen Boyd's MATLAB code.

Usage

admm.bp(
  A,
  b,
  xinit = NA,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)
admm.bp(
  A,
  b,
  xinit = NA,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)

Arguments

`A`	an $(m \times n)$ regressor matrix
`b`	a length- $m$ response vector
`xinit`	a length- $n$ vector for initial value
`rho`	an augmented Lagrangian parameter
`alpha`	an overrelaxation parameter in [1,2]
`abstol`	absolute tolerance stopping criterion
`reltol`	relative tolerance stopping criterion
`maxiter`	maximum number of iterations

Value

a named list containing

x: a length- $n$ solution vector
history: dataframe recording iteration numerics. See the section for more details.

Iteration History

When you run the algorithm, output returns not only the solution, but also the iteration history recording following fields over iterates,

objval: object (cost) function value
r_norm: norm of primal residual
s_norm: norm of dual residual
eps_pri: feasibility tolerance for primal feasibility condition
eps_dual: feasibility tolerance for dual feasibility condition

In accordance with the paper, iteration stops when both r_norm and s_norm values become smaller than eps_pri and eps_dual, respectively.

Examples

## generate sample data
n = 30
m = 10

A = matrix(rnorm(n*m), nrow=m)    # design matrix
x = c(stats::rnorm(3),rep(0,n-3)) # coefficient
x = base::sample(x)
b = as.vector(A%*%x)              # response

## run example
output  = admm.bp(A, b)
niter   = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)

## generate sample data
n = 30
m = 10

A = matrix(rnorm(n*m), nrow=m)    # design matrix
x = c(stats::rnorm(3),rep(0,n-3)) # coefficient
x = base::sample(x)
b = as.vector(A%*%x)              # response

## run example
output  = admm.bp(A, b)
niter   = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)

Elastic Net Regularization

Description

Elastic Net regularization is a combination of $\ell_2$ stability and $\ell_1$ sparsity constraint simulatenously solving the following,

$\textrm{min}_x ~ \frac{1}{2}\|Ax-b\|_2^2 + \lambda_1 \|x\|_1 + \lambda_2 \|x\|_2^2$

with nonnegative constraints $\lambda_1$ and $\lambda_2$ . Note that if both lambda values are 0, it reduces to least-squares solution.

Usage

admm.enet(
  A,
  b,
  lambda1 = 1,
  lambda2 = 1,
  rho = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)
admm.enet(
  A,
  b,
  lambda1 = 1,
  lambda2 = 1,
  rho = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)

Arguments

`A`	an $(m\times n)$ regressor matrix
`b`	a length- $m$ response vector
`lambda1`	a regularization parameter for $\ell_1$ term
`lambda2`	a regularization parameter for $\ell_2$ term
`rho`	an augmented Lagrangian parameter
`abstol`	absolute tolerance stopping criterion
`reltol`	relative tolerance stopping criterion
`maxiter`	maximum number of iterations

Value

a named list containing

x: a length- $n$ solution vector
history: dataframe recording iteration numerics. See the section for more details.

Iteration History

When you run the algorithm, output returns not only the solution, but also the iteration history recording following fields over iterates,

objval: object (cost) function value
r_norm: norm of primal residual
s_norm: norm of dual residual
eps_pri: feasibility tolerance for primal feasibility condition
eps_dual: feasibility tolerance for dual feasibility condition

In accordance with the paper, iteration stops when both r_norm and s_norm values become smaller than eps_pri and eps_dual, respectively.

Author(s)

Xiaozhi Zhu

References

Zou H, Hastie T (2005). “Regularization and variable selection via the elastic net.” Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2), 301–320. ISSN 1369-7412, 1467-9868, doi:10.1111/j.1467-9868.2005.00503.x.

Examples

## generate underdetermined design matrix
m = 50
n = 100
p = 0.1   # percentange of non-zero elements

x0 = matrix(Matrix::rsparsematrix(n,1,p))
A  = matrix(rnorm(m*n),nrow=m)
for (i in 1:ncol(A)){
  A[,i] = A[,i]/sqrt(sum(A[,i]*A[,i]))
}
b = A%*%x0 + sqrt(0.001)*matrix(rnorm(m))

## run example with both regularization values = 1
output = admm.enet(A, b, lambda1=1, lambda2=1)
niter  = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)

## generate underdetermined design matrix
m = 50
n = 100
p = 0.1   # percentange of non-zero elements

x0 = matrix(Matrix::rsparsematrix(n,1,p))
A  = matrix(rnorm(m*n),nrow=m)
for (i in 1:ncol(A)){
  A[,i] = A[,i]/sqrt(sum(A[,i]*A[,i]))
}
b = A%*%x0 + sqrt(0.001)*matrix(rnorm(m))

## run example with both regularization values = 1
output = admm.enet(A, b, lambda1=1, lambda2=1)
niter  = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)

Generalized LASSO

Description

Generalized LASSO is solving the following equation,

$\textrm{min}_x ~ \frac{1}{2}\|Ax-b\|_2^2 + \lambda \|Dx\|_1$

where the choice of regularization matrix $D$ leads to different problem formulations.

Usage

admm.genlasso(
  A,
  b,
  D = diag(length(b)),
  lambda = 1,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)
admm.genlasso(
  A,
  b,
  D = diag(length(b)),
  lambda = 1,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)

Arguments

`A`	an $(m\times n)$ regressor matrix
`b`	a length- $m$ response vector
`D`	a regularization matrix of $n$ columns
`lambda`	a regularization parameter
`rho`	an augmented Lagrangian parameter
`alpha`	an overrelaxation parameter in [1,2]
`abstol`	absolute tolerance stopping criterion
`reltol`	relative tolerance stopping criterion
`maxiter`	maximum number of iterations

Value

a named list containing

x: a length- $n$ solution vector
history: dataframe recording iteration numerics. See the section for more details.

Iteration History

When you run the algorithm, output returns not only the solution, but also the iteration history recording following fields over iterates,

objval: object (cost) function value
r_norm: norm of primal residual
s_norm: norm of dual residual
eps_pri: feasibility tolerance for primal feasibility condition
eps_dual: feasibility tolerance for dual feasibility condition

In accordance with the paper, iteration stops when both r_norm and s_norm values become smaller than eps_pri and eps_dual, respectively.

Author(s)

Xiaozhi Zhu

References

Tibshirani RJ, Taylor J (2011). “The solution path of the generalized lasso.” The Annals of Statistics, 39(3), 1335–1371. ISSN 0090-5364, doi:10.1214/11-AOS878.

Zhu Y (2017). “An Augmented ADMM Algorithm With Application to the Generalized Lasso Problem.” Journal of Computational and Graphical Statistics, 26(1), 195–204. ISSN 1061-8600, 1537-2715, doi:10.1080/10618600.2015.1114491.

Examples

## generate sample data
m = 100
n = 200
p = 0.1   # percentange of non-zero elements

x0 = matrix(Matrix::rsparsematrix(n,1,p))
A  = matrix(rnorm(m*n),nrow=m)
for (i in 1:ncol(A)){
  A[,i] = A[,i]/sqrt(sum(A[,i]*A[,i]))
}
b = A%*%x0 + sqrt(0.001)*matrix(rnorm(m))
D = diag(n);

## set regularization lambda value
regval = 0.1*Matrix::norm(t(A)%*%b, 'I')

## solve LASSO via reducing from Generalized LASSO
output  = admm.genlasso(A,b,D,lambda=regval) # set D as identity matrix
niter   = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)

## generate sample data
m = 100
n = 200
p = 0.1   # percentange of non-zero elements

x0 = matrix(Matrix::rsparsematrix(n,1,p))
A  = matrix(rnorm(m*n),nrow=m)
for (i in 1:ncol(A)){
  A[,i] = A[,i]/sqrt(sum(A[,i]*A[,i]))
}
b = A%*%x0 + sqrt(0.001)*matrix(rnorm(m))
D = diag(n);

## set regularization lambda value
regval = 0.1*Matrix::norm(t(A)%*%b, 'I')

## solve LASSO via reducing from Generalized LASSO
output  = admm.genlasso(A,b,D,lambda=regval) # set D as identity matrix
niter   = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)

Least Absolute Deviations

Description

Least Absolute Deviations (LAD) is an alternative to traditional Least Sqaures by using cost function

$\textrm{min}_x ~ \|Ax-b\|_1$

to use $\ell_1$ norm instead of square loss for robust estimation of coefficient.

Usage

admm.lad(
  A,
  b,
  xinit = NA,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)
admm.lad(
  A,
  b,
  xinit = NA,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)

Arguments

`A`	an $(m\times n)$ regressor matrix
`b`	a length- $m$ response vector
`xinit`	a length- $n$ vector for initial value
`rho`	an augmented Lagrangian parameter
`alpha`	an overrelaxation parameter in [1,2]
`abstol`	absolute tolerance stopping criterion
`reltol`	relative tolerance stopping criterion
`maxiter`	maximum number of iterations

Value

a named list containing

x: a length- $n$ solution vector
history: dataframe recording iteration numerics. See the section for more details.

Iteration History

When you run the algorithm, output returns not only the solution, but also the iteration history recording following fields over iterates,

objval: object (cost) function value
r_norm: norm of primal residual
s_norm: norm of dual residual
eps_pri: feasibility tolerance for primal feasibility condition
eps_dual: feasibility tolerance for dual feasibility condition

In accordance with the paper, iteration stops when both r_norm and s_norm values become smaller than eps_pri and eps_dual, respectively.

Examples


## generate data
m = 1000
n = 100
A = matrix(rnorm(m*n),nrow=m)
x = 10*matrix(rnorm(n))
b = A%*%x

## add impulsive noise to 10% of positions
idx = sample(1:m, round(m/10))
b[idx] = b[idx] + 100*rnorm(length(idx))

## run the code
output  = admm.lad(A,b)
niter   = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)


## generate data
m = 1000
n = 100
A = matrix(rnorm(m*n),nrow=m)
x = 10*matrix(rnorm(n))
b = A%*%x

## add impulsive noise to 10% of positions
idx = sample(1:m, round(m/10))
b[idx] = b[idx] + 100*rnorm(length(idx))

## run the code
output  = admm.lad(A,b)
niter   = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)

Least Absolute Shrinkage and Selection Operator

Description

LASSO, or L1-regularized regression, is an optimization problem to solve

$\textrm{min}_x ~ \frac{1}{2}\|Ax-b\|_2^2 + \lambda \|x\|_1$

for sparsifying the coefficient vector $x$ . The implementation is borrowed from Stephen Boyd's MATLAB code.

Usage

admm.lasso(
  A,
  b,
  lambda = 1,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)
admm.lasso(
  A,
  b,
  lambda = 1,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)

Arguments

`A`	an $(m\times n)$ regressor matrix
`b`	a length- $m$ response vector
`lambda`	a regularization parameter
`rho`	an augmented Lagrangian parameter
`alpha`	an overrelaxation parameter in [1,2]
`abstol`	absolute tolerance stopping criterion
`reltol`	relative tolerance stopping criterion
`maxiter`	maximum number of iterations

Value

a named list containing

x: a length- $n$ solution vector
history: dataframe recording iteration numerics. See the section for more details.

Iteration History

When you run the algorithm, output returns not only the solution, but also the iteration history recording following fields over iterates,

objval: object (cost) function value
r_norm: norm of primal residual
s_norm: norm of dual residual
eps_pri: feasibility tolerance for primal feasibility condition
eps_dual: feasibility tolerance for dual feasibility condition

In accordance with the paper, iteration stops when both r_norm and s_norm values become smaller than eps_pri and eps_dual, respectively.

References

Tibshirani R (1996). “Regression Shrinkage and Selection via the Lasso.” Journal of the Royal Statistical Society. Series B (Methodological), 58(1), 267–288. ISSN 00359246.

Examples


## generate sample data
m = 50
n = 100
p = 0.1   # percentange of non-zero elements

x0 = matrix(Matrix::rsparsematrix(n,1,p))
A  = matrix(rnorm(m*n),nrow=m)
for (i in 1:ncol(A)){
  A[,i] = A[,i]/sqrt(sum(A[,i]*A[,i]))
}
b = A%*%x0 + sqrt(0.001)*matrix(rnorm(m))

## set regularization lambda value
lambda = 0.1*base::norm(t(A)%*%b, "F")

## run example
output  = admm.lasso(A, b, lambda)
niter   = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)


## generate sample data
m = 50
n = 100
p = 0.1   # percentange of non-zero elements

x0 = matrix(Matrix::rsparsematrix(n,1,p))
A  = matrix(rnorm(m*n),nrow=m)
for (i in 1:ncol(A)){
  A[,i] = A[,i]/sqrt(sum(A[,i]*A[,i]))
}
b = A%*%x0 + sqrt(0.001)*matrix(rnorm(m))

## set regularization lambda value
lambda = 0.1*base::norm(t(A)%*%b, "F")

## run example
output  = admm.lasso(A, b, lambda)
niter   = length(output$history$s_norm)
history = output$history

## report convergence plot
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
plot(1:niter, history$objval, "b", main="cost function")
plot(1:niter, history$r_norm, "b", main="primal residual")
plot(1:niter, history$s_norm, "b", main="dual residual")
par(opar)

Robust Principal Component Analysis

Description

Given a data matrix $M$ , it finds a decomposition

$\textrm{min}~\|L\|_*+\lambda \|S\|_1\quad \textrm{s.t.}\quad L+S=M$

where $\|L\|_*$ represents a nuclear norm for a matrix $L$ and $\|S\|_1 = \sum |S_{i,j}|$ , and $\lambda$ a balancing/regularization parameter. The choice of such norms leads to impose low-rank property for $L$ and sparsity on $S$ .

Usage

admm.rpca(
  M,
  lambda = 1/sqrt(max(nrow(M), ncol(M))),
  mu = 1,
  tol = 1e-07,
  maxiter = 1000
)
admm.rpca(
  M,
  lambda = 1/sqrt(max(nrow(M), ncol(M))),
  mu = 1,
  tol = 1e-07,
  maxiter = 1000
)

Arguments

`M`	an $(m\times n)$ data matrix
`lambda`	a regularization parameter
`mu`	an augmented Lagrangian parameter
`tol`	relative tolerance stopping criterion
`maxiter`	maximum number of iterations

Value

a named list containing

L: an $(m\times n)$ low-rank matrix
S: an $(m\times n)$ sparse matrix
history: dataframe recording iteration numerics. See the section for more details.

Iteration History

For RPCA implementation, we chose a very simple stopping criterion

$\|M-(L_k+S_k)\|_F \le tol*\|M\|_F$

for each iteration step $k$ . So for this method, we provide a vector of only relative errors,

error: relative error computed

References

Candès EJ, Li X, Ma Y, Wright J (2011). “Robust principal component analysis?” Journal of the ACM, 58(3), 1–37. ISSN 00045411, doi:10.1145/1970392.1970395.

Examples

## generate data matrix from standard normal
X = matrix(rnorm(20*5),nrow=5)

## try different regularization values
out1 = admm.rpca(X, lambda=0.01)
out2 = admm.rpca(X, lambda=0.1)
out3 = admm.rpca(X, lambda=1)

## visualize sparsity
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
image(out1$S, main="lambda=0.01")
image(out2$S, main="lambda=0.1")
image(out3$S, main="lambda=1")
par(opar)

## generate data matrix from standard normal
X = matrix(rnorm(20*5),nrow=5)

## try different regularization values
out1 = admm.rpca(X, lambda=0.01)
out2 = admm.rpca(X, lambda=0.1)
out3 = admm.rpca(X, lambda=1)

## visualize sparsity
opar <- par(no.readonly=TRUE)
par(mfrow=c(1,3))
image(out1$S, main="lambda=0.01")
image(out2$S, main="lambda=0.1")
image(out3$S, main="lambda=1")
par(opar)

Semidefinite Programming

Description

We solve the following standard semidefinite programming (SDP) problem

$\textrm{min}_X ~ \textrm{tr}(CX)$

$\textrm{s.t.} A(X)=b, ~ X \geq 0$

with $A(X)_i = \textrm{tr}(A_i^\top X) = b_i$ for $i=1,\ldots,m$ and $X \geq 0$ stands for positive-definiteness of the matrix $X$ . In the standard form, matrices $C, A_1,A_2,\ldots,A_m$ are symmetric and solution $X$ would be symmetric and positive semidefinite. This function implements alternating direction augmented Lagrangian methods.

Usage

admm.sdp(
  C,
  A,
  b,
  mu = 1,
  rho = 1,
  abstol = 1e-10,
  maxiter = 496,
  print.progress = FALSE
)
admm.sdp(
  C,
  A,
  b,
  mu = 1,
  rho = 1,
  abstol = 1e-10,
  maxiter = 496,
  print.progress = FALSE
)

Arguments

`C`	an $(n\times n)$ symmetric matrix for cost.
`A`	a length- $m$ list of $(n\times n)$ symmetric matrices for constraint.
`b`	a length- $m$ vector for equality condition.
`mu`	penalty parameter; positive real number.
`rho`	step size for updating in $(0, \frac{1+\sqrt{5}}{2})$ .
`abstol`	absolute tolerance stopping criterion.
`maxiter`	maximum number of iterations.
`print.progress`	a logical; `TRUE` to show the progress, `FALSE` to go silent.

Value

a named list containing

x: a length- $n$ solution vector
history: dataframe recording iteration numerics. See the section for more details.

Iteration History

When you run the algorithm, output returns not only the solution, but also the iteration history recording following fields over iterates,

objval: object (cost) function value
eps_pri: feasibility tolerance for primal feasibility condition
eps_dual: feasibility tolerance for dual feasibility condition
gap: gap between primal and dual cost function.

We use the stopping criterion which breaks the iteration when all eps_pri,eps_dual, and gap become smaller than abstol.

Author(s)

Kisung You

References

Wen Z, Goldfarb D, Yin W (2010). “Alternating direction augmented Lagrangian methods for semidefinite programming.” Mathematical Programming Computation, 2(3-4), 203–230. ISSN 1867-2949, 1867-2957, doi:10.1007/s12532-010-0017-1.

Examples

## a toy example
#  generate parameters
C  = matrix(c(1,2,3,2,9,0,3,0,7),nrow=3,byrow=TRUE)
A1 = matrix(c(1,0,1,0,3,7,1,7,5),nrow=3,byrow=TRUE)
A2 = matrix(c(0,2,8,2,6,0,8,0,4),nrow=3,byrow=TRUE)

A  = list(A1, A2)
b  = c(11, 19)

# run the algorithm
run = admm.sdp(C,A,b)
hst = run$history

# visualize
opar <- par(no.readonly=TRUE)
par(mfrow=c(2,2))
plot(hst$objval,   type="b", cex=0.25, main="objective value")
plot(hst$eps_pri,  type="b", cex=0.25, main="primal feasibility")
plot(hst$eps_dual, type="b", cex=0.25, main="dual feasibility")
plot(hst$gap,      type="b", cex=0.25, main="primal-dual gap")
par(opar)

## Not run: 
## comparison with CVXR's result
require(CVXR)

#  problems definition
X = Variable(3,3,PSD=TRUE)
myobj = Minimize(sum_entries(C*X)) # objective
mycon = list(                      # constraint
  sum_entries(A[[1]]*X) == b[1],
  sum_entries(A[[2]]*X) == b[2]
)
myp = Problem(myobj, mycon)        # problem

# run and visualize
res  = solve(myp)
Xsol = res$getValue(X)

opar = par(no.readonly=TRUE)
par(mfrow=c(1,2), pty="s")
image(run$X, axes=FALSE, main="ADMM result")
image(Xsol,  axes=FALSE, main="CVXR result")
par(opar)

## End(Not run)

## a toy example
#  generate parameters
C  = matrix(c(1,2,3,2,9,0,3,0,7),nrow=3,byrow=TRUE)
A1 = matrix(c(1,0,1,0,3,7,1,7,5),nrow=3,byrow=TRUE)
A2 = matrix(c(0,2,8,2,6,0,8,0,4),nrow=3,byrow=TRUE)

A  = list(A1, A2)
b  = c(11, 19)

# run the algorithm
run = admm.sdp(C,A,b)
hst = run$history

# visualize
opar <- par(no.readonly=TRUE)
par(mfrow=c(2,2))
plot(hst$objval,   type="b", cex=0.25, main="objective value")
plot(hst$eps_pri,  type="b", cex=0.25, main="primal feasibility")
plot(hst$eps_dual, type="b", cex=0.25, main="dual feasibility")
plot(hst$gap,      type="b", cex=0.25, main="primal-dual gap")
par(opar)

## Not run: 
## comparison with CVXR's result
require(CVXR)

#  problems definition
X = Variable(3,3,PSD=TRUE)
myobj = Minimize(sum_entries(C*X)) # objective
mycon = list(                      # constraint
  sum_entries(A[[1]]*X) == b[1],
  sum_entries(A[[2]]*X) == b[2]
)
myp = Problem(myobj, mycon)        # problem

# run and visualize
res  = solve(myp)
Xsol = res$getValue(X)

opar = par(no.readonly=TRUE)
par(mfrow=c(1,2), pty="s")
image(run$X, axes=FALSE, main="ADMM result")
image(Xsol,  axes=FALSE, main="CVXR result")
par(opar)

## End(Not run)

Sparse PCA

Description

Sparse Principal Component Analysis aims at finding a sparse vector by solving

$\textrm{max}_x~x^T\Sigma x \quad \textrm{s.t.} \quad \|x\|_2\le 1,~\|x\|_0\le K$

where $\|x\|_0$ is the number of non-zero elements in a vector $x$ . A convex relaxation of this problem was proposed to solve the following problem,

$\textrm{max}_X~<\Sigma,X> ~\textrm{s.t.} \quad Tr(X)=1,~\|X\|_0 \le K^2, ~X\ge 0,~\textrm{rank}(X)=1$

where $X=xx^T$ is a $(p\times p)$ matrix that is outer product of a vector $x$ by itself, and $X\ge 0$ means the matrix $X$ is positive semidefinite. With the rank condition dropped, it can be restated as

$\textrm{max}_X~ <\Sigma,X>-\rho\|X\|_1 \quad \textrm{s.t.}\quad Tr(X)=1,X\ge 0.$

After acquiring each principal component vector, an iterative step based on Schur complement deflation method is applied to regress out the impact of previously-computed projection vectors. It should be noted that those sparse basis may not be orthonormal.

Usage

admm.spca(
  Sigma,
  numpc,
  mu = 1,
  rho = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)
admm.spca(
  Sigma,
  numpc,
  mu = 1,
  rho = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)

Arguments

`Sigma`	a $(p\times p)$ (sample) covariance matrix.
`numpc`	number of principal components to be extracted.
`mu`	an augmented Lagrangian parameter.
`rho`	a regularization parameter for sparsity.
`abstol`	absolute tolerance stopping criterion.
`reltol`	relative tolerance stopping criterion.
`maxiter`	maximum number of iterations.

Value

a named list containing

basis: a $(p\times numpc)$ matrix whose columns are sparse principal components.
history: a length-numpc list of dataframes recording iteration numerics. See the section for more details.

Iteration History

For SPCA implementation, main computation is sequentially performed for each projection vector. The history field is a list of length numpc, where each element is a data frame containing iteration history recording following fields over iterates,

r_norm: norm of primal residual
s_norm: norm of dual residual
eps_pri: feasibility tolerance for primal feasibility condition
eps_dual: feasibility tolerance for dual feasibility condition

In accordance with the paper, iteration stops when both r_norm and s_norm values become smaller than eps_pri and eps_dual, respectively.

References

Ma S (2013). “Alternating Direction Method of Multipliers for Sparse Principal Component Analysis.” Journal of the Operations Research Society of China, 1(2), 253–274. ISSN 2194-668X, 2194-6698, doi:10.1007/s40305-013-0016-9.

Examples

## generate a random matrix and compute its sample covariance
X    = matrix(rnorm(1000*5),nrow=1000)
covX = stats::cov(X)

## compute 3 sparse basis
output = admm.spca(covX, 3)

## generate a random matrix and compute its sample covariance
X    = matrix(rnorm(1000*5),nrow=1000)
covX = stats::cov(X)

## compute 3 sparse basis
output = admm.spca(covX, 3)

Total Variation Minimization

Description

1-dimensional total variation minimization - also known as signal denoising - is to solve the following

$\textrm{min}_x ~ \frac{1}{2}\|x-b\|_2^2 + \lambda \sum_i |x_{i+1}-x_i|$

for a given signal $b$ . The implementation is borrowed from Stephen Boyd's MATLAB code.

Usage

admm.tv(
  b,
  lambda = 1,
  xinit = NA,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)
admm.tv(
  b,
  lambda = 1,
  xinit = NA,
  rho = 1,
  alpha = 1,
  abstol = 1e-04,
  reltol = 0.01,
  maxiter = 1000
)

Arguments

`b`	a length- $m$ response vector
`lambda`	regularization parameter
`xinit`	a length- $m$ vector for initial value
`rho`	an augmented Lagrangian parameter
`alpha`	an overrelaxation parameter in $[1,2]$
`abstol`	absolute tolerance stopping criterion
`reltol`	relative tolerance stopping criterion
`maxiter`	maximum number of iterations

Value

a named list containing

x: a length- $m$ solution vector
history: dataframe recording iteration numerics. See the section for more details.

Iteration History

When you run the algorithm, output returns not only the solution, but also the iteration history recording following fields over iterates,

objval: object (cost) function value
r_norm: norm of primal residual
s_norm: norm of dual residual
eps_pri: feasibility tolerance for primal feasibility condition
eps_dual: feasibility tolerance for dual feasibility condition

In accordance with the paper, iteration stops when both r_norm and s_norm values become smaller than eps_pri and eps_dual, respectively.

Examples

## generate sample data
x1 = as.vector(sin(1:100)+0.1*rnorm(100))
x2 = as.vector(cos(1:100)+0.1*rnorm(100)+5)
x3 = as.vector(sin(1:100)+0.1*rnorm(100)+2.5)
xsignal = c(x1,x2,x3)

## run example
output  = admm.tv(xsignal)

## visualize
opar <- par(no.readonly=TRUE)
plot(1:300, xsignal, type="l", main="TV Regularization")
lines(1:300, output$x, col="red", lwd=2)
par(opar)

## generate sample data
x1 = as.vector(sin(1:100)+0.1*rnorm(100))
x2 = as.vector(cos(1:100)+0.1*rnorm(100)+5)
x3 = as.vector(sin(1:100)+0.1*rnorm(100)+2.5)
xsignal = c(x1,x2,x3)

## run example
output  = admm.tv(xsignal)

## visualize
opar <- par(no.readonly=TRUE)
plot(1:300, xsignal, type="l", main="TV Regularization")
lines(1:300, output$x, col="red", lwd=2)
par(opar)

Package 'ADMM'

Help Index

ADMM : Algorithms using Alternating Direction Method of Multipliers

Description

Basis Pursuit

Description

Usage

Arguments

Value

Iteration History

Examples

Elastic Net Regularization

Description

Usage

Arguments

Value

Iteration History

Author(s)

References

See Also

Examples

Generalized LASSO

Description

Usage

Arguments

Value

Iteration History

Author(s)

References

Examples

Least Absolute Deviations

Description

Usage

Arguments

Value

Iteration History

Examples

Least Absolute Shrinkage and Selection Operator

Description

Usage

Arguments

Value

Iteration History

References

Examples

Robust Principal Component Analysis

Description

Usage

Arguments

Value

Iteration History

References

Examples

Semidefinite Programming

Description

Usage

Arguments

Value

Iteration History

Author(s)

References

Examples

Sparse PCA

Description

Usage

Arguments

Value

Iteration History

References

Examples

Total Variation Minimization

Description

Usage

Arguments

Value

Iteration History

Examples