use index1 / index2, etc, not X/Y or COL/ROW to denote indices.

[CommonLispStat.git] / linalg.lsp

;;; -*- mode: lisp -*-
;;; Copyright (c) 2005--2007, by A.J. Rossini <blindglobe@gmail.com>
;;; See COPYRIGHT file for any additional restrictions (BSD license).
;;; Since 1991, ANSI was finally finished.  Edited for ANSI Common Lisp.
;;;
;;; what this should do:
;;; #1 - use CFFI (and possibly Verazanno) to import C/C++.
;;; #2 - what to do for Fortran?  Possibly: C <-> bridge, or CLapack? 
;;;      problem: would be better to have access to Fortran.  For
;;;      example, think of Doug Bates comment on reverse-calls (as
;;;      distinct from callbacks).  It would be difficult if we don't
;;;      -- however, has anyone run Lapack or similar through F2CL?
;;;      Answer: yes, Matlisp does this.
;;;
;;; #3 - Use a lisp-based matrix system drop-in?  (matlisp, femlisp, clem, ...?)
;;;


;;;; linalg -- Lisp-Stat interface to basic linear algebra routines.
;;;; 
;;;; Copyright (c) 1991, by Luke Tierney. Permission is granted for
;;;; unrestricted use.

;;;
;;; Package Setup
;;;

(in-package :cl-user)

(defpackage :lisp-stat-linalg
  (:use :common-lisp
        :cffi
        :lisp-stat-ffi-int
        :lisp-stat-math
        :lisp-stat-types
        :lisp-stat-float
        :lisp-stat-compound-data
        :lisp-stat-linalg-data
        :lisp-stat-matrix)
  (:shadowing-import-from :lisp-stat-math
          expt + - * / ** mod rem abs 1+ 1- log exp sqrt sin cos tan
          asin acos atan sinh cosh tanh asinh acosh atanh float random
          truncate floor ceiling round minusp zerop plusp evenp oddp 
          < <= = /= >= > complex conjugate realpart imagpart phase
          min max logand logior logxor lognot ffloor fceiling
          ftruncate fround signum cis)
  (:export chol-decomp lu-decomp lu-solve determinant inverse
           sv-decomp qr-decomp rcondest make-rotation spline
           kernel-dens kernel-smooth 
           fft make-sweep-matrix sweep-operator ax+y eigen

           check-real ;; for optimize

           covariance-matrix matrix print-matrix solve
           backsolve eigenvalues eigenvectors accumulate cumsum combine
           lowess))

(in-package #:lisp-stat-linalg)

;;; CFFI Support

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;;
;;;             Lisp Interfaces to Linear Algebra Routines
;;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;;;
;;; Cholesky Decomposition
;;;

(cffi:defcfun ("ccl_chol_decomp_front" ccl-chol-decomp-front)
    :int (x :pointer) (y :int) (z :pointer))
(defun chol-decomp-front (x y z) 
  (ccl-chol-decomp-front x y z))

;;;;
;;;; LU Decomposition
;;;;

(cffi:defcfun ("ccl_lu_decomp_front" ccl-lu-decomp-front)
  :int (x :pointer) (y :int) (z :pointer) (u :int) (v :pointer))
(defun lu-decomp-front (x y z u v) 
(ccl-lu-decomp-front x y z u v))

(cffi:defcfun ("ccl_lu_solve_front" ccl-lu-solve-front)
    :int (x :pointer) (y :int) (z :pointer) (u :pointer) (v :int))
(defun lu-solve-front (x y z u v) 
  (ccl-lu-solve-front x y z u v))

(cffi:defcfun ("ccl_lu_inverse_front" ccl-lu-inverse-front)
    :int (x :pointer) (y :int) (z :pointer) (u :pointer) (v :int) (w :pointer))
(defun lu-inverse-front (x y z u v w) 
  (ccl-lu-inverse-front x y z u v w))

;;;;
;;;; SV Decomposition
;;;;

(cffi:defcfun ("ccl_sv_decomp_front" ccl-sv-decomp-front)
    :int (x :pointer) (y :int) (z :int) (u :pointer) (v :pointer))
(defun sv-decomp-front (x y z u v)
  (ccl-sv-decomp-front x y z u v))

;;;;
;;;; QR Decomposition
;;;;

(cffi:defcfun ("ccl_qr_decomp_front" ccl-qr-decomp-front)
    :int (x :pointer) (y :int) (z :int) (u :pointer) (v :pointer) (w :int))
(defun qr-decomp-front (x y z u v w) 
  (ccl-qr-decomp-front x y z u v w))

;;;
;;; Estimate of Condition Number for Lower Triangular Matrix
;;;

(cffi:defcfun ("ccl_rcondest_front" ccl-rcondest-front)
    :double (x :pointer) (y :int))
(defun rcondest-front (x y) 
  (ccl-rcondest-front x y))


#|
(defun rcondest-lisp-front (mat)
  "Lisp-only version of rcondest."
  (if (and (check-square-matrix mat)
           (not (is-complex mat))) ;; Complex condition estimate not available
      (if (= 0.0 (diag mat))
          0.0
          (let ((est (aref mat 0 0)))
            (dotimes (j 0 n)

  /* Set est to reciprocal of L1 matrix norm of A */
  est = fabs(a[0][0]);
  for (j = 1; j < n; j++) {
    for (i = 0, temp = fabs(a[j][j]); i < j; i++)
      temp += fabs(a[i][j]);
    est = Max(est, temp);
  }
  est = 1.0 / est;
  
  /* Solve A^Tx = e, selecting e as you proceed */
  x[0] = 1.0 / a[0][0];
  for (i = 1; i < n; i++) p[i] = a[0][i] * x[0];
  for (j = 1; j < n; j++) {
    /* select ej and calculate x[j] */
    xp = ( 1.0 - p[j]) / a[j][j];
    xm = (-1.0 - p[j]) / a[j][j];
    temp = fabs(xp);
    tempm = fabs(xm);
    for (i = j + 1; i < n; i++) {
      pm[i] = p[i] + a[j][i] * xm;
      tempm += fabs(pm[i] / a[i][i]);
      p[i] += a[j][i] * xp;
      temp += fabs(p[i] / a[i][i]);
    }
    if (temp >= tempm) x[j] = xp;
    else {
      x[j] = xm;
      for (i = j + 1; i < n; i++) p[i] = pm[i];
    }
  }
  
  for (j = 0, xnorm = 0.0; j < n; j++) xnorm += fabs(x[j]);
  est = est * xnorm;
  backsolve(a, x, n, RE);
  for (j = 0, xnorm = 0.0; j < n; j++) xnorm += fabs(x[j]);
  if (xnorm > 0) est = est / xnorm;
  
|#


;;;;
;;;; Make Rotation Matrix
;;;;

(cffi:defcfun ("ccl_make_rotation_front" ccl-make-rotation-front)
    :int (x :int) (y :pointer) (z :pointer) (u :pointer) (v :int) (w :double))
(defun make-rotation-front (x y z u v w)
  (ccl-make-rotation-front x y z u v (float w 1d0)))

;;;;
;;;; Eigenvalues and Eigenvectors
;;;;

(cffi:defcfun ("ccl_eigen_front" ccl-eigen-front)
    :int (x :pointer) (y :int) (z :pointer) (u :pointer) (v :pointer))
(defun eigen-front (x y z u v) 
  (ccl-eigen-front x y z u v))

;;;;
;;;; Spline Interpolation
;;;;

(cffi:defcfun ("ccl_range_to_rseq" ccl-range-to-rseq)
    :int (x :int) (y :pointer) (z :int) (u :pointer))
(defun la-range-to-rseq (x y z u)
  (ccl-range-to-rseq x y z u))

(cffi:defcfun ("ccl_spline_front" ccl-spline-front)
    :int (x :int) (y :pointer) (z :pointer) (u :int) (v :pointer) (w :pointer) (a :pointer))
(defun spline-front (x y z u v w a) 
  (ccl-spline-front x y z u v w a))

;;;;
;;;; Kernel Density Estimators and Smoothers
;;;;

(cffi:defcfun ("ccl_kernel_dens_front" ccl-kernel-dens-front)
    :int (x :pointer) (y :int) (z :double) (u :pointer) (v :pointer) (w :int) (a :int))
(defun kernel-dens-front (x y z u v w a)
  (ccl-kernel-dens-front x y (float z 1d0) u v w a))

(cffi:defcfun ("ccl_kernel_smooth_front" ccl-kernel-smooth-front)
    :int (x :pointer) (y :pointer) (z :int) (u :double) (v :pointer) (w :pointer) (a :int) (b :int))
(defun kernel-smooth-front (x y z u v w a b)
  (ccl-kernel-smooth-front x y z (float u 1d0) v w a b))

;;;;
;;;; Lowess Smoother Interface
;;;;

(cffi:defcfun ("ccl_base_lowess_front" ccl-base-lowess-front)
  :int (x :pointer) (y :pointer) (z :int) (u :double) (v :int) (w :double) (a :pointer) (b :pointer) (c :pointer))
(defun base-lowess-front (x y z u v w a b c)
  (ccl-base-lowess-front x y z (float u 1d0) v (float w 1d0) a b c))

;;;;
;;;; FFT
;;;;

(cffi:defcfun ("ccl_fft_front" ccl-fft-front)
    :int (x :int) (y :pointer) (z :pointer) (u :int))
(defun fft-front (x y z u) 
  (ccl-fft-front x y z u))




;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;;;
;;;;                 Lisp to C number conversion and checking
;;;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;;;;
;;;; Lisp to/from C sequence and matrix conversion and checking
;;;;

(defun is-cons (a)
  "FIXME:AJR this is not used anywhere?"
  (if (consp a) 1 0))

(defun check-fixnum (a)
  (if (/= 0 (la-data-mode a)) (error "not an integer sequence - ~s" a)))

(defun check-real (data)
  (let ((data (compound-data-seq data)))
    (cond 
     ((vectorp data)
      (let ((n (length data)))
        (declare (fixnum n))
        (dotimes (i n)
                 (declare (fixnum i))
                 (check-one-real (aref data i)))))
     ((consp data) (dolist (x data) (check-one-real x)))
     (t (error "bad sequence - ~s" data)))))

(defun vec-assign (a i x) (setf (aref a i) x))

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;;;
;;;;             Lisp Interfaces to Linear Algebra Routines
;;;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;;; FIXME: use dpbt[f2|rf], dpbstf, dpot[f2|rf]; dpptrf, zpbstf, zpbt[f2|rf]
;;; remember: factorization = decomposition, depending on training.

(defun chol-decomp (a &optional (maxoffl 0.0))
"Args: (a)
Modified Cholesky decomposition. A should be a square, symmetric matrix.
Computes lower triangular matrix L such that L L^T = A + D where D is a
diagonal matrix. If A is strictly positive definite D will be zero.
Otherwise, D is as small as possible to make A + D numerically strictly
positive definite. Returns a list (L (max D))."
  (check-square-matrix a)
  (check-real a)
  (let* ((n (array-dimension a 0))
         (result (make-array (list n n)))
         (dpars (list maxoffl 0.0)))
    (check-real dpars)
    (let ((mat (la-data-to-matrix a +mode-re+))
          (dp (la-data-to-vector dpars +mode-re+)))
      (unwind-protect
          (progn
            (chol-decomp-front mat n dp)
            (la-matrix-to-data mat n n +mode-re+ result)
            (la-vector-to-data dp 2 +mode-re+ dpars))
        (la-free-matrix mat n)
        (la-free-vector dp)))
    (list result (second dpars))))


;;; REPLACE with
;;;         (matlisp:lu M)
;;; i.e. result use by:
;;;         (setf (values (lu-out1 lu-out2 lu-out3)) (matlisp:lu my-matrix))
;;; for solution, ...
;;; for lu-solve:
;;;         (matlisp:gesv a b &opt ipivot)

(defun lu-decomp (a)
"Args: (a)
A is a square matrix of numbers (real or complex). Computes the LU
decomposition of A and returns a list of the form (LU IV D FLAG), where
LU is a matrix with the L part in the lower triangle, the U part in the 
upper triangle (the diagonal entries of L are taken to be 1), IV is a vector
describing the row permutation used, D is 1 if the number of permutations
is odd, -1 if even, and FLAG is T if A is numerically singular, NIL otherwise.
Used bu LU-SOLVE."
  (check-square-matrix a)
  (let* ((n (array-dimension a 0))
         (mode (max +mode-re+ (la-data-mode a)))
         (result (list (make-array (list n n)) (make-array n) nil nil)))
    (let ((mat (la-data-to-matrix a mode))
          (iv (la-vector n +mode-in+))
          (d (la-vector 1 +mode-re+))
          (singular 0))
      (unwind-protect
          (progn
            (setf singular (lu-decomp-front mat n iv mode d))
            (la-matrix-to-data mat n n mode (first result))
            (la-vector-to-data iv n +mode-in+ (second result))
            (setf (third result) (la-get-double d 0))
            (setf (fourth result) (if (= singular 0.0) nil t)))
        (la-free-matrix mat n)
        (la-free-vector iv)
        (la-free-vector d)))
    result))

(defun lu-solve (lu lb)
"Args: (lu b)
LU is the result of (LU-DECOMP A) for a square matrix A, B is a sequence.
Returns the solution to the equation Ax = B. Signals an error if A is
singular."
  (let ((la (first lu))
        (lidx (second lu)))
    (check-square-matrix la)
    (check-sequence lidx)
    (check-sequence lb)
    (check-fixnum lidx)
    (let* ((n (num-rows la))
           (result (make-sequence (if (consp lb) 'list 'vector) n))
           (a-mode (la-data-mode la))
           (b-mode (la-data-mode lb)))
      (if (/= n (length lidx)) (error "index sequence is wrong length"))
      (if (/= n (length lb)) (error "right hand side is wrong length"))
      (let* ((mode (max +mode-re+ a-mode b-mode))
             (a (la-data-to-matrix la mode))
             (indx (la-data-to-vector lidx +mode-in+))
             (b (la-data-to-vector lb mode))
             (singular 0))
        (unwind-protect
            (progn
              (setf singular (lu-solve-front a n indx b mode))
              (la-vector-to-data b n mode result))
          (la-free-matrix a n)
          (la-free-vector indx)
          (la-free-vector b))
        (if (/= 0.0 singular) (error "matrix is (numerically) singular"))
        result))))

(defun determinant (a)
"Args: (m)
Returns the determinant of the square matrix M."
  (let* ((lu (lu-decomp a))
         (la (first lu))
         (n (num-rows a))
         (d1 (third lu))
         (d2 0.d0))
    (declare (fixnum n))
    (flet ((fabs (x) (float (abs x) 0.d0)))
      (dotimes (i n (* d1 (exp d2)))
        (declare (fixnum i))
        (let* ((x (aref la i i))
               (magn (fabs x)))
          (if (= 0.0 magn) (return 0.d0))
          (setf d1 (* d1 (/ x magn)))
          (setf d2 (+ d2 (log magn))))))))
          
(defun inverse (a)
"Args: (m)
Returns the inverse of the the square matrix M; signals an error if M is ill
conditioned or singular"
  (check-square-matrix a)
  (let ((n (num-rows a))
        (mode (max +mode-re+ (la-data-mode a))))
    (declare (fixnum n))
    (let ((result (make-array (list n n) :initial-element 0)))
      (dotimes (i n)
               (declare (fixnum i))
               (setf (aref result i i) 1))
      (let ((mat (la-data-to-matrix a mode))
            (inv (la-data-to-matrix result mode))
            (iv (la-vector n +mode-in+))
            (v (la-vector n mode))
            (singular 0))
        (unwind-protect
            (progn
              (setf singular (lu-inverse-front mat n iv v mode inv))
              (la-matrix-to-data inv n n mode result))
          (la-free-matrix mat n)
          (la-free-matrix inv n)
          (la-free-vector iv)
          (la-free-vector v))
        (if (/= singular 0) (error "matrix is (numerically) singular"))
        result))))

;;;;
;;;; SV Decomposition
;;;;

(defun sv-decomp (a)
"Args: (a)
A is a matrix of real numbers with at least as many rows as columns.
Computes the singular value decomposition of A and returns a list of the form
(U W V FLAG) where U and V are matrices whose columns are the left and right
singular vectors of A and W is the sequence of singular values of A. FLAG is T
if the algorithm converged, NIL otherwise."
  (check-matrix a)
  (let* ((m (num-rows a))
         (n (num-cols a))
         (mode (max +mode-re+ (la-data-mode a)))
         (result (list (make-array (list m n)) 
                       (make-array n)
                       (make-array (list n n))
                       nil)))
    (if (< m n) (error "number of rows less than number of columns"))
    (if (= mode +mode-cx+) (error "complex SVD not available yet"))
    (let ((mat (la-data-to-matrix a mode))
          (w (la-vector n +mode-re+))
          (v (la-matrix n n +mode-re+))
          (converged 0))
      (unwind-protect
          (progn
            (setf converged (sv-decomp-front mat m n w v))
            (la-matrix-to-data mat m n mode (first result))
            (la-vector-to-data w n mode (second result))
            (la-matrix-to-data v n n mode (third result))
            (setf (fourth result) (if (/= 0.0 converged) t nil)))
        (la-free-matrix mat m)
        (la-free-vector w)
        (la-free-matrix v n))
      result)))


;;;;
;;;; QR Decomposition
;;;;

(defun qr-decomp (a &optional pivot)
"Args: (a &optional pivot)
A is a matrix of real numbers with at least as many rows as columns. Computes
the QR factorization of A and returns the result in a list of the form (Q R).
If PIVOT is true the columns of X are first permuted to place insure the
absolute values of the diagonal elements of R are nonincreasing. In this case
the result includes a third element, a list of the indices of the columns in
the order in which they were used."
  (check-matrix a)
  (let* ((m (num-rows a))
         (n (num-cols a))
         (mode (max +mode-re+ (la-data-mode a)))
         (p (if pivot 1 0))
         (result (if pivot
                     (list (make-array (list m n))
                           (make-array (list n n))
                           (make-array n))
                     (list (make-array (list m n)) (make-array (list n n))))))
    (if (< m n) (error "number of rows less than number of columns"))
    (if (= mode +mode-cx+) (error "complex QR decomposition not available yet"))
    (let ((mat (la-data-to-matrix a mode))
          (v (la-matrix n n +mode-re+))
          (jpvt (la-vector n +mode-in+)))
      (unwind-protect
          (progn
            (qr-decomp-front mat m n v jpvt p)
            (la-matrix-to-data mat m n mode (first result))
            (la-matrix-to-data v n n mode (second result))
            (if pivot (la-vector-to-data jpvt n +mode-in+ (third result))))
        (la-free-matrix mat m)
        (la-free-matrix v n)
        (la-free-vector jpvt))
      result)))

;;;;
;;;; Estimate of Condition Number for Lower Triangular Matrix
;;;;

(defun rcondest (a)
"Args: (a)
Returns an estimate of the reciprocal of the L1 condition number of an upper
triangular matrix a."
  (check-square-matrix a)
  (let ((mode (max +mode-re+ (la-data-mode a)))
        (n (num-rows a)))
    (if (= mode +mode-cx+)
        (error "complex condition estimate not available yet"))
    (let ((mat (la-data-to-matrix a mode))
          (est 0.0))
      (unwind-protect
          (setf est (rcondest-front mat n))
        (la-free-matrix mat n))
      est)))
  
;;;;
;;;; Make Rotation Matrix
;;;;

(defun make-rotation (x y &optional alpha)
"Args: (x y &optional alpha)
Returns a rotation matrix for rotating from X to Y, or from X toward Y 
by angle ALPHA, in radians. X and Y are sequences of the same length."
  (check-sequence x)
  (check-sequence y)
  (if alpha (check-one-real alpha))
  (let* ((n (length x))
         (mode (max +mode-re+ (la-data-mode x) (la-data-mode y)))
         (use-angle (if alpha 1 0))
         (angle (if alpha (float alpha 0.0) 0.0))
         (result (make-array (list n n))))
    (if (/= n (length y)) (error "sequences not the same length"))
    (if (= mode +mode-cx+) (error "complex data not supported yet"))
    (let ((px (la-data-to-vector x +mode-re+))
          (py (la-data-to-vector y +mode-re+))
          (rot (la-matrix n n +mode-re+)))
      (unwind-protect
          (progn
            (make-rotation-front n rot px py use-angle angle)
            (la-matrix-to-data rot n n +mode-re+ result))
        (la-free-vector px)
        (la-free-vector py)
        (la-free-matrix rot n))
      result)))

;;;;
;;;; Eigenvalues and Vectors
;;;;

(defun eigen (a)
"Args: (a)
Returns list of list of eigenvalues and list of eigenvectors of square,
symmetric matrix A. Third element of result is NIL if algorithm converges.
If the algorithm does not converge, the third element is an integer I.
In this case the eigenvalues 0, ..., I are not reliable."
  (check-square-matrix a)
  (let ((mode (max +mode-re+ (la-data-mode a)))
        (n (num-rows a)))
    (if (= mode +mode-cx+) (error "matrix must be real and symmetric"))
    (let ((evals (make-array n))
          (evecs (make-list (* n n)))
          (pa (la-data-to-vector (compound-data-seq a) +mode-re+))
          (w (la-vector n +mode-re+))
          (z (la-vector (* n n) +mode-re+))
          (fv1 (la-vector n +mode-re+))
          (ierr 0))
      (unwind-protect
          (progn
            (setf ierr (eigen-front pa n w z fv1))
            (la-vector-to-data w n +mode-re+ evals)
            (la-vector-to-data z (* n n) +mode-re+ evecs))
        (la-free-vector pa)
        (la-free-vector z)
        (la-free-vector w)
        (la-free-vector fv1))
      (list (nreverse evals)
            (nreverse (mapcar #'(lambda (x) (coerce x 'vector))
                              (split-list evecs n)))
            (if (/= 0 ierr) (- n ierr))))))

;;;;
;;;; Spline Interpolation
;;;;

(defun make-smoother-args (x y xvals)
  (check-sequence x)
  (check-real x)
  (when y
        (check-sequence y)
        (check-real y))
  (unless (integerp xvals)
          (check-sequence xvals)
          (check-real xvals))
  (let* ((n (length x))
         (ns (if (integerp xvals) xvals (length xvals)))
         (result (list (make-list ns) (make-list ns))))
    (if (and y (/= n (length y))) (error "sequences not the same length"))
    (list x y n (if (integerp xvals) 0 1) ns xvals result)))

(defun get-smoother-result (args) (seventh args))

(defmacro with-smoother-data ((x y xvals is-reg) &rest body)
  `(progn 
     (check-sequence ,x)
     (check-real ,x)
     (when ,is-reg
           (check-sequence ,y)
           (check-real ,y))
     (unless (integerp ,xvals)
             (check-sequence ,xvals)
             (check-real ,xvals))
     (let* ((supplied (not (integerp ,xvals)))
            (n (length ,x))
            (ns (if supplied (length ,xvals) ,xvals))
            (result (list (make-list ns) (make-list ns))))
       (if (and ,is-reg (/= n (length ,y)))
           (error "sequences not the same length"))
       (if (and (not supplied) (< ns 2))
           (error "too few points for interpolation"))
       (let* ((px (la-data-to-vector ,x +mode-re+))
              (py (if ,is-reg (la-data-to-vector ,y +mode-re+)))
              (pxs (if supplied 
                       (la-data-to-vector ,xvals +mode-re+)
                       (la-vector ns +mode-re+)))
              (pys (la-vector ns +mode-re+)))
         (unless supplied (la-range-to-rseq n px ns pxs))
         (unwind-protect
             (progn ,@body
                    (la-vector-to-data pxs ns +mode-re+ (first result))
                    (la-vector-to-data pys ns +mode-re+ (second result)))
           (la-free-vector px)
           (if ,is-reg (la-free-vector py))
           (la-free-vector pxs)
           (la-free-vector pys))
         result))))

(defun spline (x y &key (xvals 30))
"Args: (x y &key xvals)
Returns list of x and y values of natural cubic spline interpolation of (X,Y).
X must be strictly increasing. XVALS can be an integer, the number of equally
spaced points to use in the range of X, or it can be a sequence of points at 
which to interpolate."
  (with-smoother-data (x y xvals t)
    (let ((work (la-vector (* 2 n) +mode-re+))
          (error 0))
      (unwind-protect
          (setf error (spline-front n px py ns pxs pys work))
        (la-free-vector work))
      (if (/= error 0) (error "bad data for splines")))))

;;;;
;;;; Kernel Density Estimators and Smoothers
;;;;

(defun kernel-type-code (type)
  (cond ((eq type 'u) 0)
        ((eq type 't) 1)
        ((eq type 'g) 2)
        (t 3)))

(defun kernel-dens (x &key (type 'b) (width -1.0) (xvals 30))
"Args: (x &key xvals width type)
Returns list of x and y values of kernel density estimate of X. XVALS can be an
integer, the number of equally spaced points to use in the range of X, or it
can be a sequence of points at which to interpolate. WIDTH specifies the
window width. TYPE specifies the lernel and should be one of the symbols G, T,
U or B for gaussian, triangular, uniform or bisquare. The default is B."
  (check-one-real width)
  (with-smoother-data (x nil xvals nil) ;; warning about deleting unreachable code is TRUE -- 2nd arg=nil!
    (let ((code (kernel-type-code type))
          (error 0))
      (setf error (kernel-dens-front px n width pxs pys ns code))
      (if (/= 0 error) (error "bad kernel density data")))))

(defun kernel-smooth (x y &key (type 'b) (width -1.0) (xvals 30))
"Args: (x y &key xvals width type)
Returns list of x and y values of kernel smooth of (X,Y). XVALS can be an
integer, the number of equally spaced points to use in the range of X, or it
can be a sequence of points at which to interpolate. WIDTH specifies the
window width. TYPE specifies the lernel and should be one of the symbols G, T,
U or B for Gaussian, triangular, uniform or bisquare. The default is B."
  (check-one-real width)
  (with-smoother-data (x y xvals t)
    (let ((code (kernel-type-code type))
          (error 0))
      (kernel-smooth-front px py n width pxs pys ns code)
      ;; if we get the Lisp version ported from C, uncomment below and
      ;; comment above.  (thanks to Carlos Ungil for the initial CFFI
      ;; work).
      ;;(kernel-smooth-Cport px py n width pxs pys ns code)
      (if (/= 0 error) (error "bad kernel density data")))))


#|
(defun kernel-smooth-Cport (px py n width ;;wts wds ;; see above for mismatch?
                            xs ys ns ktype)
  "Port of kernel_smooth (Lib/kernel.c) to Lisp.
FIXME:kernel-smooth-Cport :  This is broken.
Until this is fixed, we are using Luke's C code and CFFI as glue."
  (declare (ignore width xs))
  (cond ((< n 1) 1.0)
        ((and (< n 2) (<= width 0)) 1.0)
        (t (let* ((xmin (min px))
                  (xmax (max px))
                  (width (/ (- xmax xmin) (+ 1.0 (log n)))))
             (dotimes  (i (- ns 1))
               (setf (aref ys i)
                     (let ((wsum 0.0)
                           (ysum 0.0))
                       (dotimes (j (- n 1))   )
;;;possible nasty errors...
;;                     (let* 
;;                           ((lwidth (if wds (* width (aref wds j)) width))
;;                            (lwt (* (kernel-Cport (aref xs i) (aref px j) lwidth ktype) ;; px?
;;                                    (if wts (aref wts j) 1.0))))
;;                         (setf wsum (+ wsum lwt))
;;                         (setf ysum (if py (+ ysum (* lwt (aref py j)))))) ;; py? y?
;;
;;; end of errors
                       (if py
                           (if (> wsum 0.0) 
                               (/ ysum wsum)
                               0.0)
                           (/ wsum n)))))
             (values ys)))))
|#


(defun kernel-Cport (x y w ktype)
  "Port of kernel() (Lib/kernel.c) to Lisp.
x,y,w are doubles, type is an integer"
  (if (<= w 0.0)
      0.0
      (let ((z (- x y)))
        (cond ((eq ktype "B") 
               (let* ((w (* w 2.0))
                      (z (* z 0.5)))
                 (if (and (> z -0.5)
                          (< z 0.5))
                     (/ (/ (* 15.0 (* (- 1.0 (* 4 z z))  ;; k/w
                                      (- 1.0 (* 4 z z)))) ;; k/w
                           8.0)
                        w)
                     0)))
              ((eq ktype "G")
               (let* ((w (* w 0.25))
                      (z (* z 4.0))
                      (k (/ (exp (* -0.5 z z))
                            (sqrt (* 2 PI)))))
                 (/ k w)))
              ((eq ktype "U")
               (let* ((w (* 1.5 w))
                      (z (* z 0.75))
                      (k (if (< (abs z) 0.5)
                             1.0
                             0.0)))
                 (/ k w)))
              ((eq ktype "T")
               (cond ((and (> z -1.0)
                           (< z 0.0))
                      (+ 1.0 z))  ;; k
                     ((and (> z 0.0)
                           (< z 1.0))
                      (- 1.0 z))  ;; k
                     (t 0.0)))
              (t (values 0.0))))))
                

;;;;
;;;; Lowess Smoother Interface
;;;;

(defun |base-lowess| (s1 s2 f nsteps delta)
  (check-sequence s1)
  (check-sequence s2)
  (check-real s1)
  (check-real s2)
  (check-one-real f)
  (check-one-fixnum nsteps)
  (check-one-real delta)
  (let* ((n (length s1))
         (result (make-list n)))
    (if (/= n (length s2)) (error "sequences not the same length"))
    (let ((x (la-data-to-vector s1 +mode-re+))
          (y (la-data-to-vector s2 +mode-re+))
          (ys (la-vector n +mode-re+))
          (rw (la-vector n +mode-re+))
          (res (la-vector n +mode-re+))
          (error 0))
      (unwind-protect
          (progn
            (setf error (base-lowess-front x y n f nsteps delta ys rw res))
            (la-vector-to-data ys n +mode-re+ result))
        (la-free-vector x)
        (la-free-vector y)
        (la-free-vector ys)
        (la-free-vector rw)
        (la-free-vector res))
      (if (/= error 0) (error "bad data for lowess"))
      result)))

#|
static LVAL add_contour_point(i, j, k, l, x, y, z, v, result)
        int i, j, k, l;
        RVector x, y;
        RMatrix z;
        double v;
        LVAL result;
{
  LVAL pt;
  double p, q;
  
  if ((z[i][j] <= v && v < z[k][l]) || (z[k][l] <= v && v < z[i][j])) {
    xlsave(pt);
    pt = mklist(2, NIL);
    p = (v - z[i][j]) / (z[k][l] - z[i][j]);
    q = 1.0 - p;
    rplaca(pt, cvflonum((FLOTYPE) (q * x[i] + p * x[k])));
    rplaca(cdr(pt), cvflonum((FLOTYPE) (q * y[j] + p * y[l])));
    result = cons(pt, result);
    xlpop();
  }
  return(result);
}

LVAL xssurface_contour()
{
  LVAL s1, s2, mat, result;
  RVector x, y;
  RMatrix z;
  double v;
  int i, j, n, m;
  
  s1 = xsgetsequence();
  s2 = xsgetsequence();
  mat = xsgetmatrix();
  v = makedouble(xlgetarg());
  xllastarg();
    
  n = seqlen(s1); m = seqlen(s2);
  if (n != numrows(mat) || m != numcols(mat)) xlfail("dimensions do not match");
  if (data_mode(s1) == CX || data_mode(s2) == CX || data_mode(mat) == CX)
    xlfail("data must be real");
  
  x = (RVector) data_to_vector(s1, RE);
  y = (RVector) data_to_vector(s2, RE);
  z = (RMatrix) data_to_matrix(mat, RE);
  
  xlsave1(result);
  result = NIL;
  for (i = 0; i < n - 1; i++) {
        for (j = 0; j < m - 1; j++) {
          result = add_contour_point(i, j, i, j+1, x, y, z, v, result);
          result = add_contour_point(i, j+1, i+1, j+1, x, y, z, v, result);
          result = add_contour_point(i+1, j+1, i+1, j, x, y, z, v, result);
          result = add_contour_point(i+1, j, i, j, x, y, z, v, result);
        }
  }
  xlpop();
  
  free_vector(x);
  free_vector(y);
  free_matrix(z, n);
  
  return(result);
}
|#

;;;
;;; FFT
;;;
;;; FIXME:ajr
;;; ??replace with matlisp:fft  and matlisp:ifft (the latter for inverse mapping)
;;;
(defun fft (x &optional inverse)
"Args: (x &optional inverse)
Returns unnormalized Fourier transform of X, or inverse transform if INVERSE
is true."
  (check-sequence x)
  (let* ((n (length x))
         (mode (la-data-mode x))
         (isign (if inverse -1 1))
         (result (if (consp x) (make-list n) (make-array n))))
    (let ((px (la-data-to-vector x +mode-cx+))
          (work (la-vector (+ (* 4 n) 15) +mode-re+)))
      (unwind-protect
          (progn
            (fft-front n px work isign)
            (la-vector-to-data px n +mode-cx+ result))
        (la-free-vector px)
        (la-free-vector work))
      result)))

;;;
;;; SWEEP Operator: FIXME: use matlisp
;;;

(defun make-sweep-front (x y w n p mode has_w x_mean result)
  (declare (fixnum n p mode has_w))
  (let ((x_data nil)
        (result_data nil)
        (val 0.0)
        (dxi 0.0)
        (dyi 0.0)
        (dv 0.0)
        (dw 0.0)
        (sum_w 0.0)
        (dxik 0.0)
        (dxjk 0.0)
        (dyj 0.0)
        (dx_meani 0.0)
        (dx_meanj 0.0)
        (dy_mean 0.0)
        (has-w (if (/= 0 has_w) t nil))
        (RE 1))
    (declare (long-float val dxi dyi dv dw sum_w dxik dxjk dyj
                    dx_meani dx_meanj dy_mean))
;;    (declare-double val dxi dyi dv dw sum_w dxik dxjk dyj
;;                  dx_meani dx_meanj dy_mean)
  
    (if (> mode RE) (error "not supported for complex data yet"))

    (setf x_data (compound-data-seq x))
    (setf result_data (compound-data-seq result))
  
    ;; find the mean of y
    (setf val 0.0)
    (setf sum_w 0.0)
    (dotimes (i n)
      (declare (fixnum i))
      (setf dyi (makedouble (aref y i)))
      (when has-w
        (setf dw (makedouble (aref w i)))
        (incf sum_w dw)
        (setf dyi (* dyi dw)))
      (incf val dyi))
    (if (not has-w) (setf sum_w (float n 0.0)))
    (if (<= sum_w 0.0) (error "non positive sum of weights"))
    (setf dy_mean (/ val sum_w))
  
    ;; find the column means
    (dotimes (j p)
      (declare (fixnum j))
      (setf val 0.0)
      (dotimes (i n)
        (declare (fixnum i))
        (setf dxi (makedouble (aref x_data (+ (* p i) j))))
        (when has-w
          (setf dw (makedouble (aref w i)))
          (setf dxi (* dxi dw)))
        (incf val dxi))
      (setf (aref x_mean j) (/ val sum_w)))
  
    ;; put 1/sum_w in topleft, means on left, minus means on top
    (setf (aref result_data 0) (/ 1.0 sum_w))
    (dotimes (i p)
      (declare (fixnum i))
      (setf dxi (makedouble (aref x_mean i)))
      (setf (aref result_data (+ i 1)) (- dxi))
      (setf (aref result_data (* (+ i 1) (+ p 2))) dxi))
    (setf (aref result_data (+ p 1)) (- dy_mean))
    (setf (aref result_data (* (+ p 1) (+ p 2))) dy_mean)
  
    ;; put sums of adjusted cross products in body
    (dotimes (i p)
      (declare (fixnum i))
      (dotimes (j p)
        (declare (fixnum j))
        (setf val 0.0)
        (dotimes (k n)
          (declare (fixnum k))
          (setf dxik (makedouble (aref x_data (+ (* p k) i))))
          (setf dxjk (makedouble (aref x_data (+ (* p k) j))))
          (setf dx_meani (makedouble (aref x_mean i)))
          (setf dx_meanj (makedouble (aref x_mean j)))
          (setf dv (* (- dxik dx_meani) (- dxjk dx_meanj)))
          (when has-w
            (setf dw (makedouble (aref w k)))
            (setf dv (* dv dw)))
          (incf val dv))
        (setf (aref result_data (+ (* (+ i 1) (+ p 2)) (+ j 1))) val)
        (setf (aref result_data (+ (* (+ j 1) (+ p 2)) (+ i 1))) val))
      (setf val 0.0)
      (dotimes (j n)
        (declare (fixnum j))
        (setf dxik (makedouble (aref x_data (+ (* p j) i))))
        (setf dyj (makedouble (aref y j)))
        (setf dx_meani (makedouble (aref x_mean i)))
        (setf dv (* (- dxik dx_meani) (- dyj dy_mean)))
        (when has-w
          (setf dw (makedouble (aref w j)))
          (setf dv (* dv dw)))
        (incf val dv))
      (setf (aref result_data (+ (* (+ i 1) (+ p 2)) (+ p 1))) val)
      (setf (aref result_data (+ (* (+ p 1) (+ p 2)) (+ i 1))) val))
    (setf val 0.0)
    (dotimes (j n)
      (declare (fixnum j))
      (setf dyj (makedouble (aref y j)))
      (setf dv (* (- dyj dy_mean) (- dyj dy_mean)))
      (when has-w
        (setf dw (makedouble (aref w j)))
        (setf dv (* dv dw)))
      (incf val dv))
    (setf (aref result_data (+ (* (+ p 1) (+ p 2)) (+ p 1))) val)))

;;; FIXME: use matlisp
(defun sweep-in-place-front (a rows cols mode k tol)
  "Sweep algorithm for linear regression."
  (declare (long-float tol))
  (declare (fixnum rows cols mode k))
  (let ((data nil)
        (pivot 0.0)
        (aij 0.0)
        (aik 0.0)
        (akj 0.0)
        (akk 0.0)
        (RE 1))
    (declare (long-float pivot aij aik akj akk))
  
    (if (> mode RE) (error "not supported for complex data yet"))
    (if (or (< k 0) (>= k rows) (>= k cols)) (error "index out of range"))
  
    (setf tol (max tol machine-epsilon))
    (setf data (compound-data-seq a))

    (setf pivot (makedouble (aref data (+ (* cols k) k))))
  
    (cond
     ((or (> pivot tol) (< pivot (- tol)))
      (dotimes (i rows)
        (declare (fixnum i))
        (dotimes (j cols)
          (declare (fixnum j))
          (when (and (/= i k) (/= j k))
            (setf aij (makedouble (aref data (+ (* cols i) j))))
            (setf aik (makedouble (aref data (+ (* cols i) k))))
            (setf akj (makedouble (aref data (+ (* cols k) j))))
            (setf aij (- aij (/ (* aik akj) pivot)))
            (setf (aref data (+ (* cols i) j)) aij))))

      (dotimes (i rows)
        (declare (fixnum i))
        (setf aik (makedouble (aref data (+ (* cols i) k))))
        (when (/= i k)
          (setf aik (/ aik pivot))
          (setf (aref data (+ (* cols i) k)) aik)))

      (dotimes (j cols)
        (declare (fixnum j))
        (setf akj (makedouble (aref data (+ (* cols k) j))))
        (when (/= j k)
          (setf akj (- (/ akj pivot)))
          (setf (aref data (+ (* cols k) j)) akj)))

      (setf akk (/ 1.0 pivot))
      (setf (aref data (+ (* cols k) k)) akk)
      1)
     (t 0))))

;; FIXME: use matlisp
(defun make-sweep-matrix (x y &optional w)
  "Args: (x y &optional weights)
 X is matrix, Y and WEIGHTS are sequences. Returns the sweep matrix of the
 (weighted) regression of Y on X"
   (check-matrix x)
   (check-sequence y)
   (if w (check-sequence w))
   (let ((n (num-rows x))
        (p (num-cols x)))
      (if (/= n (length y)) (error "dimensions do not match"))
    (if (and w (/= n (length w))) (error "dimensions do not match"))
    (let ((mode (max (la-data-mode x) 
                     (la-data-mode x) 
                     (if w (la-data-mode w) 0)))
          (result (make-array (list (+ p 2) (+ p 2))))
          (x-mean (make-array p))
          (y (coerce y 'vector))
          (w (if w (coerce w 'vector)))
          (has-w (if w 1 0)))
      (make-sweep-front x y w n p mode has-w x-mean result)
      result)))

 (defun sweep-in-place (a k tol)
  (check-matrix a)
  (check-one-fixnum k)
  (check-one-real tol)
  (let ((rows (num-rows a))
        (cols (num-cols a))
        (mode (la-data-mode a)))
    (let ((swept (sweep-in-place-front a rows cols mode k tol)))
      (if (/= 0 swept) t nil))))

(defun sweep-operator (a columns &optional tolerances)
"Args: (a indices &optional tolerances)
A is a matrix, INDICES a sequence of the column indices to be swept. Returns
a list of the swept result and the list of the columns actually swept. (See
MULTREG documentation.) If supplied, TOLERANCES should be a list of real
numbers the same length as INDICES. An index will only be swept if its pivot
element is larger than the corresponding element of TOLERANCES."
  (check-matrix a)
  (check-sequence columns)
  (if tolerances (check-sequence tolerances))
  (check-real a)
  (check-fixnum columns)
  (if tolerances (check-real tolerances))
  (do ((tol .0000001)
       (result (copy-array a))
       (swept-columns nil)
       (columns (coerce columns 'list) (cdr columns))
       (tolerances (if (consp tolerances) (coerce tolerances 'list))
                   (if (consp tolerances) (cdr tolerances))))
      ((null columns) (list result swept-columns))
      (let ((col (first columns))
            (tol (if (consp tolerances) (first tolerances) tol)))
        (if (sweep-in-place result col tol)
            (setf swept-columns (cons col swept-columns))))))


;;;
;;; AX+Y
;;;

;;; matlisp:axpy
;;;
(defun ax+y (a x y &optional lower)
"Args (a x y &optional lower)
Returns (+ (matmult A X) Y). If LOWER is not nil, A is taken to be lower
triangular.
This could probably be made more efficient."
  (check-square-matrix a)
  (check-sequence x)
  (check-sequence y)
  (check-real a)
  (check-real x)
  (check-real y)
  (let* ((n (num-rows a))
         (result (make-list n))
         (a (compound-data-seq a)))
    (declare (fixnum n))
    (if (or (/= n (length x)) (/= n (length y)))
        (error "dimensions do not match"))
    (do* ((tx (make-next-element x) (make-next-element x))
          (ty (make-next-element y))
          (tr (make-next-element result))
          (i 0 (+ i 1))
          (start 0 (+ start n))
          (end (if lower (+ i 1) n) (if lower (+ i 1) n)))
         ((<= n i) result)
       (declare (fixnum i start end))
       (let ((val (get-next-element ty i)))
         (dotimes (j end)
           (declare (fixnum j))
           (setf val (+ val (* (get-next-element tx j)
                               (aref a (+ start j))))))
         (set-next-element tr i val)))))





;;;;
;;;; Linear Algebra Functions
;;;;

(defun matrix (dim data)
"Args: (dim data)
returns a matrix of dimensions DIM initialized using sequence DATA
in row major order." 
  (let ((dim (coerce dim 'list))
        (data (coerce data 'list)))
    (make-array dim :initial-contents (split-list data (nth 1 dim)))))

(defun flatsize (x)
  (length x))  ;; FIXME: defined badly!!

(defun print-matrix (a &optional (stream *standard-output*))
"Args: (matrix &optional stream)
Prints MATRIX to STREAM in a nice form that is still machine readable"
  (unless (matrixp a) (error "not a matrix - ~a" a))
  (let ((size (min 15 (max (map-elements #'flatsize a)))))
    (format stream "#2a(~%")
    (dolist (x (row-list a))
            (format stream "    (")
            (let ((n (length x)))
              (dotimes (i n)
                       (let ((y (aref x i)))
                         (cond
                           ((integerp y) (format stream "~vd" size y))
                           ((floatp y) (format stream "~vg" size y))
                           (t (format stream "~va" size y))))
                       (if (< i (- n 1)) (format stream " "))))
            (format stream ")~%"))
    (format stream "   )~%")
    nil))

(defun solve (a b)
"Args: (a b)
Solves A x = B using LU decomposition and backsolving. B can be a sequence
or a matrix."
  (let ((lu (lu-decomp a)))
    (if (matrixp b)
        (apply #'bind-columns 
               (mapcar #'(lambda (x) (lu-solve lu x)) (column-list b)))
        (lu-solve lu b))))
        
(defun backsolve (a b)
"Args: (a b)
Solves A x = B by backsolving, assuming A is upper triangular. B must be a
sequence. For use with qr-decomp."
  (let* ((n (length b))
         (sol (make-array n)))
    (dotimes (i n)
             (let* ((k (- n i 1))
                    (val (elt b k)))
               (dotimes (j i)
                        (let ((l (- n j 1)))
                          (setq val (- val (* (aref sol l) (aref a k l))))))
               (setf (aref sol k) (/ val (aref a k k)))))
    (if (listp b) (coerce sol 'list) sol)))

(defun eigenvalues (a) 
"Args: (a)
Returns list of eigenvalues of square, symmetric matrix A"
  (first (eigen a)))

(defun eigenvectors (a) 
"Args: (a)
Returns list of eigenvectors of square, symmetric matrix A"
  (second (eigen a)))

(defun accumulate (f s)
"Args: (f s)
Accumulates elements of sequence S using binary function F.
(accumulate #'+ x) returns the cumulative sum of x."
  (let* ((result (list (elt s 0)))
         (tail result))
    (flet ((acc (dummy x)
                (rplacd tail (list (funcall f (first tail) x)))
                (setf tail (cdr tail))))
      (reduce #'acc s))
    (if (vectorp s) (coerce result 'vector) result)))

(defun cumsum (x)
  "Args: (x)
Returns the cumulative sum of X."
  (accumulate #'+ x))

(defun combine (&rest args) 
  "Args (&rest args) 
Returns sequence of elements of all arguments."
  (copy-seq (element-seq args)))

(defun lowess (x y &key (f .25) (steps 2) (delta -1) sorted)
"Args: (x y &key (f .25) (steps 2) delta sorted)
Returns (list X YS) with YS the LOWESS fit. F is the fraction of data used for
each point, STEPS is the number of robust iterations. Fits for points within
DELTA of each other are interpolated linearly. If the X values setting SORTED
to T speeds up the computation."
  (let ((x (if sorted x (sort-data x)))
        (y (if sorted y (select y (order x))))
        (delta (if (> delta 0.0) delta (/ (- (max x) (min x)) 50))))
    (list x)));; (|base-lowess| x y f steps delta))))