qr-methods: QR Decomposition - S4 Methods and Generic

Description Usage Arguments Methods See Also Examples

Description

The Matrix package provides methods for the QR decomposition of special classes of matrices. There is a generic function which uses qr as default, but methods defined in this package can take extra arguments. In particular there is an option for determining a fill-reducing permutation of the columns of a sparse, rectangular matrix.

Usage

1
2
qr(x, ...)
qrR(qr, complete=FALSE, backPermute=TRUE, row.names = TRUE)

Arguments

x

a numeric or complex matrix whose QR decomposition is to be computed. Logical matrices are coerced to numeric.

qr

a QR decomposition of the type computed by qr.

complete

logical indicating whether the \bold{R} matrix is to be completed by binding zero-value rows beneath the square upper triangle.

backPermute

logical indicating if the rows of the \bold{R} matrix should be back permuted such that qrR()'s result can be used directly to reconstruct the original matrix \bold{X}.

row.names

logical indicating if rownames should propagated to the result.

...

further arguments passed to or from other methods

Methods

x = "dgCMatrix"

QR decomposition of a general sparse double-precision matrix with nrow(x) >= ncol(x). Returns an object of class "sparseQR".

x = "sparseMatrix"

works via "dgCMatrix".

See Also

qr; then, the class documentations, mainly sparseQR, and also dgCMatrix.

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
##------------- example of pivoting -- from base'  qraux.Rd -------------
X <- cbind(int = 1,
           b1=rep(1:0, each=3), b2=rep(0:1, each=3),
           c1=rep(c(1,0,0), 2), c2=rep(c(0,1,0), 2), c3=rep(c(0,0,1),2))
rownames(X) <- paste0("r", seq_len(nrow(X)))
dnX <- dimnames(X)
bX <- X
X <- as(bX, "sparseMatrix")
X # is singular, columns "b2" and "c3" are "extra"
stopifnot(identical(dimnames(X), dnX))# some versions changed X's dimnames!
c(rankMatrix(X)) # = 4 (not 6)

##----- regular case ------------------------------------------
Xr <- X[ , -c(3,6)] # the "regular" (non-singular) version of X
stopifnot(rankMatrix(Xr) == ncol(Xr))
Y <- cbind(y <- setNames(1:6, paste0("y", 1:6)))
## regular case:
m <- function(.) as(., "matrix")
qXr   <- qr(  Xr)
qxr   <- qr(m(Xr))
qxrLA <- qr(m(Xr), LAPACK=TRUE) # => qr.fitted(), qr.resid() not supported
qcfXy <- qr.coef (qXr, y) # vector
qcfXY <- qr.coef (qXr, Y) # 4x1 dgeMatrix
cf <- c(int=6, b1=-3, c1=-2, c2=-1)
stopifnot(
    all.equal(qr.coef(qxr,  y),   cf, tol=1e-15)
  , getRversion() <= "3.4.1" ||
    all.equal(qr.coef(qxrLA,y),   cf, tol=1e-15)
  , all.equal(qr.coef(qxr,  Y), m(cf), tol=1e-15)
  , all.equal(  qcfXy,    cf, tol=1e-15)
  , all.equal(m(qcfXY), m(cf), tol=1e-15)
  , all.equal(y, qr.fitted(qxr, y), tol=2e-15)
  , all.equal(y, qr.fitted(qXr, y), tol=2e-15)
  , all.equal(m(qr.fitted(qXr, Y)), qr.fitted(qxr, Y), tol=1e-15)
  , all.equal(  qr.resid (qXr, y),  qr.resid (qxr, y), tol=1e-15)
  , all.equal(m(qr.resid (qXr, Y)), qr.resid (qxr, Y), tol=1e-15)
)

##----- rank-deficient ("singular") case ------------------------------------

(qX <- qr(  X))           # both @p and @q are non-trivial permutations
 qx <- qr(m(X)) ; str(qx) # $pivot is non-trivial, too

drop0(R. <- qr.R(qX), tol=1e-15) # columns *permuted*: c3 b1 ..
Q. <- qr.Q(qX)
qI <- sort.list(qX@q) # the inverse 'q' permutation
(X. <- drop0(Q. %*% R.[, qI], tol=1e-15))## just = X, incl. correct colnames
stopifnot(all(X - X.) < 8*.Machine$double.eps,
          ## qrR(.) returns R already "back permuted" (as with qI):
          identical(R.[, qI], qrR(qX)) )
##
## In this sense, classical qr.coef() is fine:
cfqx <- qr.coef(qx, y) # quite different from
nna <- !is.na(cfqx)
stopifnot(all.equal(unname(qr.fitted(qx,y)),
                    as.numeric(X[,nna] %*% cfqx[nna])))
## FIXME: do these make *any* sense? --- should give warnings !
qr.coef(qX, y)
qr.coef(qX, Y)
rm(m)

Matrix documentation built on Aug. 16, 2017, 9:03 a.m.