/* glpini01.c */ /*********************************************************************** * This code is part of GLPK (GNU Linear Programming Kit). * * Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, * 2009, 2010 Andrew Makhorin, Department for Applied Informatics, * Moscow Aviation Institute, Moscow, Russia. All rights reserved. * E-mail: . * * GLPK is free software: you can redistribute it and/or modify it * under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * GLPK is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY * or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public * License for more details. * * You should have received a copy of the GNU General Public License * along with GLPK. If not, see . ***********************************************************************/ #include "glpapi.h" /*---------------------------------------------------------------------- -- triang - find maximal triangular part of a rectangular matrix. -- -- *Synopsis* -- -- int triang(int m, int n, -- void *info, int (*mat)(void *info, int k, int ndx[]), -- int rn[], int cn[]); -- -- *Description* -- -- For a given rectangular (sparse) matrix A with m rows and n columns -- the routine triang tries to find such permutation matrices P and Q -- that the first rows and columns of the matrix B = P*A*Q form a lower -- triangular submatrix of as greatest size as possible: -- -- 1 n -- 1 * . . . . . . x x x x x x -- * * . . . . . x x x x x x -- * * * . . . . x x x x x x -- * * * * . . . x x x x x x -- B = P*A*Q = * * * * * . . x x x x x x -- * * * * * * . x x x x x x -- * * * * * * * x x x x x x -- x x x x x x x x x x x x x -- x x x x x x x x x x x x x -- m x x x x x x x x x x x x x -- -- where: '*' - elements of the lower triangular part, '.' - structural -- zeros, 'x' - other (either non-zero or zero) elements. -- -- The parameter info is a transit pointer passed to the formal routine -- mat (see below). -- -- The formal routine mat specifies the given matrix A in both row- and -- column-wise formats. In order to obtain an i-th row of the matrix A -- the routine triang calls the routine mat with the parameter k = +i, -- 1 <= i <= m. In response the routine mat should store column indices -- of (non-zero) elements of the i-th row to the locations ndx[1], ..., -- ndx[len], where len is number of non-zeros in the i-th row returned -- on exit. Analogously, in order to obtain a j-th column of the matrix -- A, the routine mat is called with the parameter k = -j, 1 <= j <= n, -- and should return pattern of the j-th column in the same way as for -- row patterns. Note that the routine mat may be called more than once -- for the same rows and columns. -- -- On exit the routine computes two resultant arrays rn and cn, which -- define the permutation matrices P and Q, respectively. The array rn -- should have at least 1+m locations, where rn[i] = i' (1 <= i <= m) -- means that i-th row of the original matrix A corresponds to i'-th row -- of the matrix B = P*A*Q. Similarly, the array cn should have at least -- 1+n locations, where cn[j] = j' (1 <= j <= n) means that j-th column -- of the matrix A corresponds to j'-th column of the matrix B. -- -- *Returns* -- -- The routine triang returns the size of the lower tringular part of -- the matrix B = P*A*Q (see the figure above). -- -- *Complexity* -- -- The time complexity of the routine triang is O(nnz), where nnz is -- number of non-zeros in the given matrix A. -- -- *Algorithm* -- -- The routine triang starts from the matrix B = P*Q*A, where P and Q -- are unity matrices, so initially B = A. -- -- Before the next iteration B = (B1 | B2 | B3), where B1 is partially -- built a lower triangular submatrix, B2 is the active submatrix, and -- B3 is a submatrix that contains rejected columns. Thus, the current -- matrix B looks like follows (initially k1 = 1 and k2 = n): -- -- 1 k1 k2 n -- 1 x . . . . . . . . . . . . . # # # -- x x . . . . . . . . . . . . # # # -- x x x . . . . . . . . . . # # # # -- x x x x . . . . . . . . . # # # # -- x x x x x . . . . . . . # # # # # -- k1 x x x x x * * * * * * * # # # # # -- x x x x x * * * * * * * # # # # # -- x x x x x * * * * * * * # # # # # -- x x x x x * * * * * * * # # # # # -- m x x x x x * * * * * * * # # # # # -- <--B1---> <----B2-----> <---B3--> -- -- On each iteartion the routine looks for a singleton row, i.e. some -- row that has the only non-zero in the active submatrix B2. If such -- row exists and the corresponding non-zero is b[i,j], where (by the -- definition) k1 <= i <= m and k1 <= j <= k2, the routine permutes -- k1-th and i-th rows and k1-th and j-th columns of the matrix B (in -- order to place the element in the position b[k1,k1]), removes the -- k1-th column from the active submatrix B2, and adds this column to -- the submatrix B1. If no row singletons exist, but B2 is not empty -- yet, the routine chooses a j-th column, which has maximal number of -- non-zeros among other columns of B2, removes this column from B2 and -- adds it to the submatrix B3 in the hope that new row singletons will -- appear in the active submatrix. */ static int triang(int m, int n, void *info, int (*mat)(void *info, int k, int ndx[]), int rn[], int cn[]) { int *ndx; /* int ndx[1+max(m,n)]; */ /* this array is used for querying row and column patterns of the given matrix A (the third parameter to the routine mat) */ int *rs_len; /* int rs_len[1+m]; */ /* rs_len[0] is not used; rs_len[i], 1 <= i <= m, is number of non-zeros in the i-th row of the matrix A, which (non-zeros) belong to the current active submatrix */ int *rs_head; /* int rs_head[1+n]; */ /* rs_head[len], 0 <= len <= n, is the number i of the first row of the matrix A, for which rs_len[i] = len */ int *rs_prev; /* int rs_prev[1+m]; */ /* rs_prev[0] is not used; rs_prev[i], 1 <= i <= m, is a number i' of the previous row of the matrix A, for which rs_len[i] = rs_len[i'] (zero marks the end of this linked list) */ int *rs_next; /* int rs_next[1+m]; */ /* rs_next[0] is not used; rs_next[i], 1 <= i <= m, is a number i' of the next row of the matrix A, for which rs_len[i] = rs_len[i'] (zero marks the end this linked list) */ int cs_head; /* is a number j of the first column of the matrix A, which has maximal number of non-zeros among other columns */ int *cs_prev; /* cs_prev[1+n]; */ /* cs_prev[0] is not used; cs_prev[j], 1 <= j <= n, is a number of the previous column of the matrix A with the same or greater number of non-zeros than in the j-th column (zero marks the end of this linked list) */ int *cs_next; /* cs_next[1+n]; */ /* cs_next[0] is not used; cs_next[j], 1 <= j <= n, is a number of the next column of the matrix A with the same or lesser number of non-zeros than in the j-th column (zero marks the end of this linked list) */ int i, j, ii, jj, k1, k2, len, t, size = 0; int *head, *rn_inv, *cn_inv; if (!(m > 0 && n > 0)) xerror("triang: m = %d; n = %d; invalid dimension\n", m, n); /* allocate working arrays */ ndx = xcalloc(1+(m >= n ? m : n), sizeof(int)); rs_len = xcalloc(1+m, sizeof(int)); rs_head = xcalloc(1+n, sizeof(int)); rs_prev = xcalloc(1+m, sizeof(int)); rs_next = xcalloc(1+m, sizeof(int)); cs_prev = xcalloc(1+n, sizeof(int)); cs_next = xcalloc(1+n, sizeof(int)); /* build linked lists of columns of the matrix A with the same number of non-zeros */ head = rs_len; /* currently rs_len is used as working array */ for (len = 0; len <= m; len ++) head[len] = 0; for (j = 1; j <= n; j++) { /* obtain length of the j-th column */ len = mat(info, -j, ndx); xassert(0 <= len && len <= m); /* include the j-th column in the corresponding linked list */ cs_prev[j] = head[len]; head[len] = j; } /* merge all linked lists of columns in one linked list, where columns are ordered by descending of their lengths */ cs_head = 0; for (len = 0; len <= m; len++) { for (j = head[len]; j != 0; j = cs_prev[j]) { cs_next[j] = cs_head; cs_head = j; } } jj = 0; for (j = cs_head; j != 0; j = cs_next[j]) { cs_prev[j] = jj; jj = j; } /* build initial doubly linked lists of rows of the matrix A with the same number of non-zeros */ for (len = 0; len <= n; len++) rs_head[len] = 0; for (i = 1; i <= m; i++) { /* obtain length of the i-th row */ rs_len[i] = len = mat(info, +i, ndx); xassert(0 <= len && len <= n); /* include the i-th row in the correspondng linked list */ rs_prev[i] = 0; rs_next[i] = rs_head[len]; if (rs_next[i] != 0) rs_prev[rs_next[i]] = i; rs_head[len] = i; } /* initially all rows and columns of the matrix A are active */ for (i = 1; i <= m; i++) rn[i] = 0; for (j = 1; j <= n; j++) cn[j] = 0; /* set initial bounds of the active submatrix */ k1 = 1, k2 = n; /* main loop starts here */ while (k1 <= k2) { i = rs_head[1]; if (i != 0) { /* the i-th row of the matrix A is a row singleton, since it has the only non-zero in the active submatrix */ xassert(rs_len[i] == 1); /* determine the number j of an active column of the matrix A, in which this non-zero is placed */ j = 0; t = mat(info, +i, ndx); xassert(0 <= t && t <= n); for (t = t; t >= 1; t--) { jj = ndx[t]; xassert(1 <= jj && jj <= n); if (cn[jj] == 0) { xassert(j == 0); j = jj; } } xassert(j != 0); /* the singleton is a[i,j]; move a[i,j] to the position b[k1,k1] of the matrix B */ rn[i] = cn[j] = k1; /* shift the left bound of the active submatrix */ k1++; /* increase the size of the lower triangular part */ size++; } else { /* the current active submatrix has no row singletons */ /* remove an active column with maximal number of non-zeros from the active submatrix */ j = cs_head; xassert(j != 0); cn[j] = k2; /* shift the right bound of the active submatrix */ k2--; } /* the j-th column of the matrix A has been removed from the active submatrix */ /* remove the j-th column from the linked list */ if (cs_prev[j] == 0) cs_head = cs_next[j]; else cs_next[cs_prev[j]] = cs_next[j]; if (cs_next[j] == 0) /* nop */; else cs_prev[cs_next[j]] = cs_prev[j]; /* go through non-zeros of the j-th columns and update active lengths of the corresponding rows */ t = mat(info, -j, ndx); xassert(0 <= t && t <= m); for (t = t; t >= 1; t--) { i = ndx[t]; xassert(1 <= i && i <= m); /* the non-zero a[i,j] has left the active submatrix */ len = rs_len[i]; xassert(len >= 1); /* remove the i-th row from the linked list of rows with active length len */ if (rs_prev[i] == 0) rs_head[len] = rs_next[i]; else rs_next[rs_prev[i]] = rs_next[i]; if (rs_next[i] == 0) /* nop */; else rs_prev[rs_next[i]] = rs_prev[i]; /* decrease the active length of the i-th row */ rs_len[i] = --len; /* return the i-th row to the corresponding linked list */ rs_prev[i] = 0; rs_next[i] = rs_head[len]; if (rs_next[i] != 0) rs_prev[rs_next[i]] = i; rs_head[len] = i; } } /* other rows of the matrix A, which are still active, correspond to rows k1, ..., m of the matrix B (in arbitrary order) */ for (i = 1; i <= m; i++) if (rn[i] == 0) rn[i] = k1++; /* but for columns this is not needed, because now the submatrix B2 has no columns */ for (j = 1; j <= n; j++) xassert(cn[j] != 0); /* perform some optional checks */ /* make sure that rn is a permutation of {1, ..., m} and cn is a permutation of {1, ..., n} */ rn_inv = rs_len; /* used as working array */ for (ii = 1; ii <= m; ii++) rn_inv[ii] = 0; for (i = 1; i <= m; i++) { ii = rn[i]; xassert(1 <= ii && ii <= m); xassert(rn_inv[ii] == 0); rn_inv[ii] = i; } cn_inv = rs_head; /* used as working array */ for (jj = 1; jj <= n; jj++) cn_inv[jj] = 0; for (j = 1; j <= n; j++) { jj = cn[j]; xassert(1 <= jj && jj <= n); xassert(cn_inv[jj] == 0); cn_inv[jj] = j; } /* make sure that the matrix B = P*A*Q really has the form, which was declared */ for (ii = 1; ii <= size; ii++) { int diag = 0; i = rn_inv[ii]; t = mat(info, +i, ndx); xassert(0 <= t && t <= n); for (t = t; t >= 1; t--) { j = ndx[t]; xassert(1 <= j && j <= n); jj = cn[j]; if (jj <= size) xassert(jj <= ii); if (jj == ii) { xassert(!diag); diag = 1; } } xassert(diag); } /* free working arrays */ xfree(ndx); xfree(rs_len); xfree(rs_head); xfree(rs_prev); xfree(rs_next); xfree(cs_prev); xfree(cs_next); /* return to the calling program */ return size; } /*---------------------------------------------------------------------- -- adv_basis - construct advanced initial LP basis. -- -- *Synopsis* -- -- #include "glpini.h" -- void adv_basis(glp_prob *lp); -- -- *Description* -- -- The routine adv_basis constructs an advanced initial basis for an LP -- problem object, which the parameter lp points to. -- -- In order to build the initial basis the routine does the following: -- -- 1) includes in the basis all non-fixed auxiliary variables; -- -- 2) includes in the basis as many as possible non-fixed structural -- variables preserving triangular form of the basis matrix; -- -- 3) includes in the basis appropriate (fixed) auxiliary variables -- in order to complete the basis. -- -- As a result the initial basis has minimum of fixed variables and the -- corresponding basis matrix is triangular. */ static int mat(void *info, int k, int ndx[]) { /* this auxiliary routine returns the pattern of a given row or a given column of the augmented constraint matrix A~ = (I|-A), in which columns of fixed variables are implicitly cleared */ LPX *lp = info; int m = lpx_get_num_rows(lp); int n = lpx_get_num_cols(lp); int typx, i, j, lll, len = 0; if (k > 0) { /* the pattern of the i-th row is required */ i = +k; xassert(1 <= i && i <= m); #if 0 /* 22/XII-2003 */ /* if the auxiliary variable x[i] is non-fixed, include its element (placed in the i-th column) in the pattern */ lpx_get_row_bnds(lp, i, &typx, NULL, NULL); if (typx != LPX_FX) ndx[++len] = i; /* include in the pattern elements placed in columns, which correspond to non-fixed structural varables */ i_beg = aa_ptr[i]; i_end = i_beg + aa_len[i] - 1; for (i_ptr = i_beg; i_ptr <= i_end; i_ptr++) { j = m + sv_ndx[i_ptr]; lpx_get_col_bnds(lp, j-m, &typx, NULL, NULL); if (typx != LPX_FX) ndx[++len] = j; } #else lll = lpx_get_mat_row(lp, i, ndx, NULL); for (k = 1; k <= lll; k++) { lpx_get_col_bnds(lp, ndx[k], &typx, NULL, NULL); if (typx != LPX_FX) ndx[++len] = m + ndx[k]; } lpx_get_row_bnds(lp, i, &typx, NULL, NULL); if (typx != LPX_FX) ndx[++len] = i; #endif } else { /* the pattern of the j-th column is required */ j = -k; xassert(1 <= j && j <= m+n); /* if the (auxiliary or structural) variable x[j] is fixed, the pattern of its column is empty */ if (j <= m) lpx_get_row_bnds(lp, j, &typx, NULL, NULL); else lpx_get_col_bnds(lp, j-m, &typx, NULL, NULL); if (typx != LPX_FX) { if (j <= m) { /* x[j] is non-fixed auxiliary variable */ ndx[++len] = j; } else { /* x[j] is non-fixed structural variables */ #if 0 /* 22/XII-2003 */ j_beg = aa_ptr[j]; j_end = j_beg + aa_len[j] - 1; for (j_ptr = j_beg; j_ptr <= j_end; j_ptr++) ndx[++len] = sv_ndx[j_ptr]; #else len = lpx_get_mat_col(lp, j-m, ndx, NULL); #endif } } } /* return the length of the row/column pattern */ return len; } static void adv_basis(glp_prob *lp) { int m = lpx_get_num_rows(lp); int n = lpx_get_num_cols(lp); int i, j, jj, k, size; int *rn, *cn, *rn_inv, *cn_inv; int typx, *tagx = xcalloc(1+m+n, sizeof(int)); double lb, ub; xprintf("Constructing initial basis...\n"); #if 0 /* 13/V-2009 */ if (m == 0) xerror("glp_adv_basis: problem has no rows\n"); if (n == 0) xerror("glp_adv_basis: problem has no columns\n"); #else if (m == 0 || n == 0) { glp_std_basis(lp); return; } #endif /* use the routine triang (see above) to find maximal triangular part of the augmented constraint matrix A~ = (I|-A); in order to prevent columns of fixed variables to be included in the triangular part, such columns are implictly removed from the matrix A~ by the routine adv_mat */ rn = xcalloc(1+m, sizeof(int)); cn = xcalloc(1+m+n, sizeof(int)); size = triang(m, m+n, lp, mat, rn, cn); if (lpx_get_int_parm(lp, LPX_K_MSGLEV) >= 3) xprintf("Size of triangular part = %d\n", size); /* the first size rows and columns of the matrix P*A~*Q (where P and Q are permutation matrices defined by the arrays rn and cn) form a lower triangular matrix; build the arrays (rn_inv and cn_inv), which define the matrices inv(P) and inv(Q) */ rn_inv = xcalloc(1+m, sizeof(int)); cn_inv = xcalloc(1+m+n, sizeof(int)); for (i = 1; i <= m; i++) rn_inv[rn[i]] = i; for (j = 1; j <= m+n; j++) cn_inv[cn[j]] = j; /* include the columns of the matrix A~, which correspond to the first size columns of the matrix P*A~*Q, in the basis */ for (k = 1; k <= m+n; k++) tagx[k] = -1; for (jj = 1; jj <= size; jj++) { j = cn_inv[jj]; /* the j-th column of A~ is the jj-th column of P*A~*Q */ tagx[j] = LPX_BS; } /* if size < m, we need to add appropriate columns of auxiliary variables to the basis */ for (jj = size + 1; jj <= m; jj++) { /* the jj-th column of P*A~*Q should be replaced by the column of the auxiliary variable, for which the only unity element is placed in the position [jj,jj] */ i = rn_inv[jj]; /* the jj-th row of P*A~*Q is the i-th row of A~, but in the i-th row of A~ the unity element belongs to the i-th column of A~; therefore the disired column corresponds to the i-th auxiliary variable (note that this column doesn't belong to the triangular part found by the routine triang) */ xassert(1 <= i && i <= m); xassert(cn[i] > size); tagx[i] = LPX_BS; } /* free working arrays */ xfree(rn); xfree(cn); xfree(rn_inv); xfree(cn_inv); /* build tags of non-basic variables */ for (k = 1; k <= m+n; k++) { if (tagx[k] != LPX_BS) { if (k <= m) lpx_get_row_bnds(lp, k, &typx, &lb, &ub); else lpx_get_col_bnds(lp, k-m, &typx, &lb, &ub); switch (typx) { case LPX_FR: tagx[k] = LPX_NF; break; case LPX_LO: tagx[k] = LPX_NL; break; case LPX_UP: tagx[k] = LPX_NU; break; case LPX_DB: tagx[k] = (fabs(lb) <= fabs(ub) ? LPX_NL : LPX_NU); break; case LPX_FX: tagx[k] = LPX_NS; break; default: xassert(typx != typx); } } } for (k = 1; k <= m+n; k++) { if (k <= m) lpx_set_row_stat(lp, k, tagx[k]); else lpx_set_col_stat(lp, k-m, tagx[k]); } xfree(tagx); return; } /*********************************************************************** * NAME * * glp_adv_basis - construct advanced initial LP basis * * SYNOPSIS * * void glp_adv_basis(glp_prob *lp, int flags); * * DESCRIPTION * * The routine glp_adv_basis constructs an advanced initial basis for * the specified problem object. * * The parameter flags is reserved for use in the future and must be * specified as zero. */ void glp_adv_basis(glp_prob *lp, int flags) { if (flags != 0) xerror("glp_adv_basis: flags = %d; invalid flags\n", flags); if (lp->m == 0 || lp->n == 0) glp_std_basis(lp); else adv_basis(lp); return; } /* eof */