racket/collects/math/private/matrix/matrix-operator-norm.rkt
Neil Toronto f42cc6f14a Fixed major performance issue with matrix arithmetic; please merge to 5.3.2
The fix consists of three parts:

1. Rewriting `inline-matrix*'. The material change here is that the
   expansion now contains only direct applications of `+' and `*'.
   TR's optimizer replaces them with `unsafe-fx+' and `unsafe-fx*',
   which keeps intermediate flonum values from being boxed.

2. Making the types of all functions that operate on (Matrix Number)
   values more precise. Now TR can prove that matrix operations preserve
   inexactness. For example, matrix-conjugate : (Matrix Flonum) ->
   (Matrix Flonum) and three other cases for Real, Float-Complex, and
   Number.

3. Changing the return types of some functions that used to return
   things like (Matrix (U A 0)). Now that we worry about preserving
   inexactness, we can't have `matrix-upper-triangle' always return a
   matrix that contains exact zeros. It now accepts an optional `zero'
   argument of type A.
2013-01-21 22:04:04 -07:00

155 lines
6.6 KiB
Racket

#lang typed/racket/base
#|
Two of the functions defined here currently just raise an error: `matrix-op-2norm' and
`matrix-op-angle'. They need to compute, respectively, the maximum and minimum singular values of
their matrix argument.
See "How to Measure Errors" in the LAPACK manual for more details:
http://www.netlib.org/lapack/lug/node75.html
http://www.netlib.org/lapack/lug/node76.html
|#
(require racket/list
racket/fixnum
math/flonum
"matrix-types.rkt"
"matrix-arithmetic.rkt"
"matrix-constructors.rkt"
"matrix-basic.rkt"
"utils.rkt"
"../array/array-struct.rkt"
"../array/array-pointwise.rkt"
"../array/array-fold.rkt"
)
(provide
;; Operator norms
matrix-op-1norm
matrix-op-2norm
matrix-op-inf-norm
matrix-basis-cos-angle
matrix-basis-angle
;; Error measurement
matrix-error-norm
matrix-absolute-error
matrix-relative-error
;; Approximate predicates
matrix-zero?
matrix-identity?
matrix-orthonormal?
)
(: matrix-op-1norm (case-> ((Matrix Flonum) -> Nonnegative-Flonum)
((Matrix Real) -> Nonnegative-Real)
((Matrix Float-Complex) -> Nonnegative-Flonum)
((Matrix Number) -> Nonnegative-Real)))
;; When M is a column matrix, this is equivalent to matrix-1norm
(define (matrix-op-1norm M)
(parameterize ([array-strictness #f])
(assert (apply max (map matrix-1norm (matrix-cols M))) nonnegative?)))
(: matrix-op-2norm (case-> ((Matrix Flonum) -> Nonnegative-Flonum)
((Matrix Real) -> Nonnegative-Real)
((Matrix Float-Complex) -> Nonnegative-Flonum)
((Matrix Number) -> Nonnegative-Real)))
;; When M is a column matrix, this is equivalent to matrix-2norm
(define (matrix-op-2norm M)
;(matrix-max-singular-value M)
;(sqrt (matrix-max-eigenvalue M))
(error 'unimplemented))
(: matrix-op-inf-norm (case-> ((Matrix Flonum) -> Nonnegative-Flonum)
((Matrix Real) -> Nonnegative-Real)
((Matrix Float-Complex) -> Nonnegative-Flonum)
((Matrix Number) -> Nonnegative-Real)))
;; When M is a column matrix, this is equivalent to matrix-inf-norm
(define (matrix-op-inf-norm M)
(parameterize ([array-strictness #f])
(assert (apply max (map matrix-1norm (matrix-rows M))) nonnegative?)))
(: matrix-basis-cos-angle (case-> ((Matrix Flonum) (Matrix Flonum) -> Flonum)
((Matrix Real) (Matrix Real) -> Real)
((Matrix Float-Complex) (Matrix Float-Complex) -> Float-Complex)
((Matrix Number) (Matrix Number) -> Number)))
;; Returns the angle between the two subspaces spanned by the two given sets of column vectors
(define (matrix-basis-cos-angle M R)
;(matrix-min-singular-value (matrix* (matrix-hermitian M) R))
(error 'unimplemented))
(: matrix-basis-angle (case-> ((Matrix Flonum) (Matrix Flonum) -> Flonum)
((Matrix Real) (Matrix Real) -> Real)
((Matrix Float-Complex) (Matrix Float-Complex) -> Float-Complex)
((Matrix Number) (Matrix Number) -> Number)))
;; Returns the angle between the two subspaces spanned by the two given sets of column vectors
(define (matrix-basis-angle M R)
(acos (matrix-basis-cos-angle M R)))
;; ===================================================================================================
;; Error measurement
(: matrix-error-norm (Parameterof ((Matrix Number) -> Nonnegative-Real)))
(define matrix-error-norm (make-parameter matrix-op-inf-norm))
(: matrix-absolute-error
(case-> ((Matrix Number) (Matrix Number) -> Nonnegative-Real)
((Matrix Number) (Matrix Number) ((Matrix Number) -> Nonnegative-Real)
-> Nonnegative-Real)))
(define (matrix-absolute-error M R [norm (matrix-error-norm)])
(parameterize ([array-strictness #f])
(define-values (m n) (matrix-shapes 'matrix-absolute-error M R))
(array-strict! M)
(array-strict! R)
(cond [(array-all-and (inline-array-map eqv? M R)) 0]
[(and (array-all-and (inline-array-map number-rational? M))
(array-all-and (inline-array-map number-rational? R)))
(norm (matrix- (inline-array-map inexact->exact M)
(inline-array-map inexact->exact R)))]
[else +inf.0])))
(: matrix-relative-error
(case-> ((Matrix Number) (Matrix Number) -> Nonnegative-Real)
((Matrix Number) (Matrix Number) ((Matrix Number) -> Nonnegative-Real)
-> Nonnegative-Real)))
(define (matrix-relative-error M R [norm (matrix-error-norm)])
(parameterize ([array-strictness #f])
(define-values (m n) (matrix-shapes 'matrix-relative-error M R))
(array-strict! M)
(array-strict! R)
(cond [(array-all-and (inline-array-map eqv? M R)) 0]
[(and (array-all-and (inline-array-map number-rational? M))
(array-all-and (inline-array-map number-rational? R)))
(define num (norm (matrix- M R)))
(define den (norm R))
(cond [(and (zero? num) (zero? den)) 0]
[(zero? den) +inf.0]
[else (assert (/ num den) nonnegative?)])]
[else +inf.0])))
;; ===================================================================================================
;; Approximate predicates
(: matrix-zero? (case-> ((Matrix Number) -> Boolean)
((Matrix Number) Real -> Boolean)))
(define (matrix-zero? M [eps (* 10 epsilon.0)])
(cond [(eps . < . 0) (raise-argument-error 'matrix-identity? "Nonnegative-Real" 1 M eps)]
[else
(define-values (m n) (matrix-shape M))
(<= (matrix-absolute-error M (make-matrix m n 0)) eps)]))
(: matrix-identity? (case-> ((Matrix Number) -> Boolean)
((Matrix Number) Real -> Boolean)))
(define (matrix-identity? M [eps (* 10 epsilon.0)])
(cond [(eps . < . 0) (raise-argument-error 'matrix-identity? "Nonnegative-Real" 1 M eps)]
[else (and (square-matrix? M)
(<= (matrix-relative-error M (identity-matrix (square-matrix-size M))) eps))]))
(: matrix-orthonormal? (case-> ((Matrix Number) -> Boolean)
((Matrix Number) Real -> Boolean)))
(define (matrix-orthonormal? M [eps (* 10 epsilon.0)])
(cond [(eps . < . 0) (raise-argument-error 'matrix-orthonormal? "Nonnegative-Real" 1 M eps)]
[else (and (square-matrix? M)
(matrix-identity? (matrix* M (matrix-hermitian M)) eps))]))