GNU Linear Programming Kit: Reference Manual

M L
GNU Linear Programming Kit
Reference Manual
for GLPK Version 4.64
(DRAFT, November 2017)
J K
The GLPK package is part of the GNU Project released under the aegis of GNU.
Copyright ⃝c 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2013, 2014,
2015, 2016, 2017 Andrew Makhorin, Department for Applied Informatics, Moscow Aviation Insti-
tute, Moscow, Russia. All rights reserved.
Free Software Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
Permission is granted to make and distribute verbatim copies of this manual provided the copyright
notice and this permission notice are preserved on all copies.
Permission is granted to copy and distribute modified versions of this manual under the conditions
for verbatim copying, provided also that the entire resulting derived work is distributed under the
terms of a permission notice identical to this one.
Permission is granted to copy and distribute translations of this manual into another language,
under the above conditions for modified versions.
2
Contents
1 Introduction 9
1.1 LP problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 MIP problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Using the package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.1 Brief example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.2 Compiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.3 Linking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Basic API Routines 14

2.1 General conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.1 Library header . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.2 Error handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.3 Thread safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.4 Array indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Problem object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.1 Problem segment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.2 Basis segment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.3 Interior-point segment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.4 MIP segment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Problem creating and modifying routines . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.1 glp create prob — create problem object . . . . . . . . . . . . . . . . . . . . . 18
2.3.2 glp set prob name — assign (change) problem name . . . . . . . . . . . . . . 18
2.3.3 glp set obj name — assign (change) objective function name . . . . . . . . . 18
2.3.4 glp set obj dir — set (change) optimization direction flag . . . . . . . . . . . 19
2.3.5 glp add rows — add new rows to problem object . . . . . . . . . . . . . . . . 19
2.3.6 glp add cols — add new columns to problem object . . . . . . . . . . . . . . 19
2.3.7 glp set row name — assign (change) row name . . . . . . . . . . . . . . . . . 20
2.3.8 glp set col name — assign (change) column name . . . . . . . . . . . . . . . . 20
2.3.9 glp set row bnds — set (change) row bounds . . . . . . . . . . . . . . . . . . 20
2.3.10 glp set col bnds — set (change) column bounds . . . . . . . . . . . . . . . . . 21
2.3.11 glp set obj coef — set (change) objective coefficient or constant term . . . . . 21
2.3.12 glp set mat row — set (replace) row of the constraint matrix . . . . . . . . . 22
2.3.13 glp set mat col — set (replace) column of the constraint matrix . . . . . . . . 22
2.3.14 glp load matrix — load (replace) the whole constraint matrix . . . . . . . . . 23
2.3.15 glp check dup — check for duplicate elements in sparse matrix . . . . . . . . 23
2.3.16 glp sort matrix — sort elements of the constraint matrix . . . . . . . . . . . 24
2.3.17 glp del rows — delete rows from problem object . . . . . . . . . . . . . . . . 24
3
2.3.18 glp del cols — delete columns from problem object . . . . . . . . . . . . . . . 24
2.3.19 glp copy prob — copy problem object content . . . . . . . . . . . . . . . . . . 25
2.3.20 glp erase prob — erase problem object content . . . . . . . . . . . . . . . . . 25
2.3.21 glp delete prob — delete problem object . . . . . . . . . . . . . . . . . . . . . 25
2.4 Problem retrieving routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.1 glp get prob name — retrieve problem name . . . . . . . . . . . . . . . . . . 26
2.4.2 glp get obj name — retrieve objective function name . . . . . . . . . . . . . . 26
2.4.3 glp get obj dir — retrieve optimization direction flag . . . . . . . . . . . . . . 26
2.4.4 glp get num rows — retrieve number of rows . . . . . . . . . . . . . . . . . . 26
2.4.5 glp get num cols — retrieve number of columns . . . . . . . . . . . . . . . . . 27
2.4.6 glp get row name — retrieve row name . . . . . . . . . . . . . . . . . . . . . 27
2.4.7 glp get col name — retrieve column name . . . . . . . . . . . . . . . . . . . . 27
2.4.8 glp get row type — retrieve row type . . . . . . . . . . . . . . . . . . . . . . 27
2.4.9 glp get row lb — retrieve row lower bound . . . . . . . . . . . . . . . . . . . 28
2.4.10 glp get row ub — retrieve row upper bound . . . . . . . . . . . . . . . . . . . 28
2.4.11 glp get col type — retrieve column type . . . . . . . . . . . . . . . . . . . . . 28
2.4.12 glp get col lb — retrieve column lower bound . . . . . . . . . . . . . . . . . . 28
2.4.13 glp get col ub — retrieve column upper bound . . . . . . . . . . . . . . . . . 29
2.4.14 glp get obj coef — retrieve objective coefficient or constant term . . . . . . . 29
2.4.15 glp get num nz — retrieve number of constraint coefficients . . . . . . . . . . 29
2.4.16 glp get mat row — retrieve row of the constraint matrix . . . . . . . . . . . . 29
2.4.17 glp get mat col — retrieve column of the constraint matrix . . . . . . . . . . 30
2.5 Row and column searching routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.1 glp create index — create the name index . . . . . . . . . . . . . . . . . . . . 31
2.5.2 glp find row — find row by its name . . . . . . . . . . . . . . . . . . . . . . . 32
2.5.3 glp find col — find column by its name . . . . . . . . . . . . . . . . . . . . . 32
2.5.4 glp delete index — delete the name index . . . . . . . . . . . . . . . . . . . . 32
2.6 Problem scaling routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.6.2 glp set rii — set (change) row scale factor . . . . . . . . . . . . . . . . . . . . 33
2.6.3 glp set sjj — set (change) column scale factor . . . . . . . . . . . . . . . . . . 33
2.6.4 glp get rii — retrieve row scale factor . . . . . . . . . . . . . . . . . . . . . . 34
2.6.5 glp get sjj — retrieve column scale factor . . . . . . . . . . . . . . . . . . . . 34
2.6.6 glp scale prob — scale problem data . . . . . . . . . . . . . . . . . . . . . . . 34
2.6.7 glp unscale prob — unscale problem data . . . . . . . . . . . . . . . . . . . . 34
2.7 LP basis constructing routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.7.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.7.2 glp set row stat — set (change) row status . . . . . . . . . . . . . . . . . . . 35
2.7.3 glp set col stat — set (change) column status . . . . . . . . . . . . . . . . . . 36
2.7.4 glp std basis — construct standard initial LP basis . . . . . . . . . . . . . . . 36
2.7.5 glp adv basis — construct advanced initial LP basis . . . . . . . . . . . . . . 36
2.7.6 glp cpx basis — construct Bixby’s initial LP basis . . . . . . . . . . . . . . . 37
2.8 Simplex method routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.8.1 glp simplex — solve LP problem with the primal or dual simplex method . . 39
2.8.2 glp exact — solve LP problem in exact arithmetic . . . . . . . . . . . . . . . 45
2.8.3 glp init smcp — initialize simplex solver control parameters . . . . . . . . . . 46
2.8.4 glp get status — determine generic status of basic solution . . . . . . . . . . 46
2.8.5 glp get prim stat — retrieve status of primal basic solution . . . . . . . . . . 46
4
2.8.6 glp get dual stat — retrieve status of dual basic solution . . . . . . . . . . . . 47
2.8.7 glp get obj val — retrieve objective value . . . . . . . . . . . . . . . . . . . . 47
2.8.8 glp get row stat — retrieve row status . . . . . . . . . . . . . . . . . . . . . . 47
2.8.9 glp get row prim — retrieve row primal value . . . . . . . . . . . . . . . . . . 47
2.8.10 glp get row dual — retrieve row dual value . . . . . . . . . . . . . . . . . . . 48
2.8.11 glp get col stat — retrieve column status . . . . . . . . . . . . . . . . . . . . 48
2.8.12 glp get col prim — retrieve column primal value . . . . . . . . . . . . . . . . 48
2.8.13 glp get col dual — retrieve column dual value . . . . . . . . . . . . . . . . . . 48
2.8.14 glp get unbnd ray — determine variable causing unboundedness . . . . . . . 49
2.9 Interior-point method routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.9.1 glp interior — solve LP problem with the interior-point method . . . . . . . . 51
2.9.2 glp init iptcp — initialize interior-point solver control parameters . . . . . . . 55
2.9.3 glp ipt status — determine solution status . . . . . . . . . . . . . . . . . . . . 55
2.9.4 glp ipt obj val — retrieve objective value . . . . . . . . . . . . . . . . . . . . 55
2.9.5 glp ipt row prim — retrieve row primal value . . . . . . . . . . . . . . . . . . 55
2.9.6 glp ipt row dual — retrieve row dual value . . . . . . . . . . . . . . . . . . . 56
2.9.7 glp ipt col prim — retrieve column primal value . . . . . . . . . . . . . . . . 56
2.9.8 glp ipt col dual — retrieve column dual value . . . . . . . . . . . . . . . . . . 56
2.10 Mixed integer programming routines . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.10.1 glp set col kind — set (change) column kind . . . . . . . . . . . . . . . . . . 57
2.10.2 glp get col kind — retrieve column kind . . . . . . . . . . . . . . . . . . . . . 57
2.10.3 glp get num int — retrieve number of integer columns . . . . . . . . . . . . . 57
2.10.4 glp get num bin — retrieve number of binary columns . . . . . . . . . . . . . 58
2.10.5 glp intopt — solve MIP problem with the branch-and-cut method . . . . . . 58
2.10.6 glp init iocp — initialize integer optimizer control parameters . . . . . . . . . 63
2.10.7 glp mip status — determine status of MIP solution . . . . . . . . . . . . . . . 63
2.10.8 glp mip obj val — retrieve objective value . . . . . . . . . . . . . . . . . . . . 63
2.10.9 glp mip row val — retrieve row value . . . . . . . . . . . . . . . . . . . . . . 64
2.10.10 glp mip col val — retrieve column value . . . . . . . . . . . . . . . . . . . . . 64
2.11 Additional routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.11.1 glp check kkt — check feasibility/optimality conditions . . . . . . . . . . . . 65
3 Utility API routines 69

3.1 Problem data reading/writing routines . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.1.1 glp read mps — read problem data in MPS format . . . . . . . . . . . . . . . 69
3.1.2 glp write mps — write problem data in MPS format . . . . . . . . . . . . . . 70
3.1.3 glp read lp — read problem data in CPLEX LP format . . . . . . . . . . . . 70
3.1.4 glp write lp — write problem data in CPLEX LP format . . . . . . . . . . . 71
3.1.5 glp read prob — read problem data in GLPK format . . . . . . . . . . . . . . 71
3.1.6 glp write prob — write problem data in GLPK format . . . . . . . . . . . . . 76
3.2 Routines for processing MathProg models . . . . . . . . . . . . . . . . . . . . . . . . 77
3.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.2.2 glp mpl alloc wksp — allocate the translator workspace . . . . . . . . . . . . 80
3.2.3 glp mpl init rand — initialize pseudo-random number generator . . . . . . . 80
3.2.4 glp mpl read model — read and translate model section . . . . . . . . . . . . 80
3.2.5 glp mpl read data — read and translate data section . . . . . . . . . . . . . . 81
3.2.6 glp mpl generate — generate the model . . . . . . . . . . . . . . . . . . . . . 81
3.2.7 glp mpl build prob — build problem instance from the model . . . . . . . . . 81
5
3.2.8 glp mpl postsolve — postsolve the model . . . . . . . . . . . . . . . . . . . . 82
3.2.9 glp mpl free wksp — free the translator workspace . . . . . . . . . . . . . . . 82
3.3 Problem solution reading/writing routines . . . . . . . . . . . . . . . . . . . . . . . . 83
3.3.1 glp print sol — write basic solution in printable format . . . . . . . . . . . . 83
3.3.2 glp read sol — read basic solution in GLPK format . . . . . . . . . . . . . . 83
3.3.3 glp write sol — write basic solution in GLPK format . . . . . . . . . . . . . . 86
3.3.4 glp print ipt — write interior-point solution in printable format . . . . . . . . 86
3.3.5 glp read ipt — read interior-point solution in GLPK format . . . . . . . . . . 86
3.3.6 glp write ipt — write interior-point solution in GLPK format . . . . . . . . . 88
3.3.7 glp print mip — write MIP solution in printable format . . . . . . . . . . . . 89
3.3.8 glp read mip — read MIP solution in GLPK format . . . . . . . . . . . . . . 89
3.3.9 glp write mip — write MIP solution in GLPK format . . . . . . . . . . . . . 91
3.4 Post-optimal analysis routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.4.1 glp print ranges — print sensitivity analysis report . . . . . . . . . . . . . . . 92
4 Advanced API Routines 98

4.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.2 LP basis routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2.1 glp bf exists — check if the basis factorization exists . . . . . . . . . . . . . . 104
4.2.2 glp factorize — compute the basis factorization . . . . . . . . . . . . . . . . . 104
4.2.3 glp bf updated — check if the basis factorization has been updated . . . . . . 105
4.2.4 glp get bfcp — retrieve basis factorization control parameters . . . . . . . . . 105
4.2.5 glp set bfcp — change basis factorization control parameters . . . . . . . . . 106
4.2.6 glp get bhead — retrieve the basis header information . . . . . . . . . . . . . 108
4.2.7 glp get row bind — retrieve row index in the basis header . . . . . . . . . . . 109
4.2.8 glp get col bind — retrieve column index in the basis header . . . . . . . . . 109
4.2.9 glp ftran — perform forward transformation . . . . . . . . . . . . . . . . . . . 109
4.2.10 glp btran — perform backward transformation . . . . . . . . . . . . . . . . . 110
4.2.11 glp warm up — “warm up” LP basis . . . . . . . . . . . . . . . . . . . . . . . 110
4.3 Simplex tableau routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.3.1 glp eval tab row — compute row of the tableau . . . . . . . . . . . . . . . . . 111
4.3.2 glp eval tab col — compute column of the tableau . . . . . . . . . . . . . . . 112
4.3.3 glp transform row — transform explicitly specified row . . . . . . . . . . . . 113
4.3.4 glp transform col — transform explicitly specified column . . . . . . . . . . . 114
4.3.5 glp prim rtest — perform primal ratio test . . . . . . . . . . . . . . . . . . . 115
4.3.6 glp dual rtest — perform dual ratio test . . . . . . . . . . . . . . . . . . . . . 116
4.4 Post-optimal analysis routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
4.4.1 glp analyze bound — analyze active bound of non-basic variable . . . . . . . 117
4.4.2 glp analyze coef — analyze objective coefficient at basic variable . . . . . . . 117
5 Branch-and-Cut API Routines 119

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.1.1 Using the callback routine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.1.2 Branch-and-cut algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.1.3 The search tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.1.4 Current subproblem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.1.5 The cut pool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.1.6 Reasons for calling the callback routine . . . . . . . . . . . . . . . . . . . . . 123
6
5.2 Basic routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.2.1 glp ios reason — determine reason for calling the callback routine . . . . . . 126
5.2.2 glp ios get prob — access the problem object . . . . . . . . . . . . . . . . . . 126
5.2.3 glp ios row attr — determine additional row attributes . . . . . . . . . . . . . 127
5.2.4 glp ios mip gap — compute relative MIP gap . . . . . . . . . . . . . . . . . . 127
5.2.5 glp ios node data — access application-specific data . . . . . . . . . . . . . . 128
5.2.6 glp ios select node — select subproblem to continue the search . . . . . . . . 128
5.2.7 glp ios heur sol — provide solution found by heuristic . . . . . . . . . . . . . 129
5.2.8 glp ios can branch — check if can branch upon specified variable . . . . . . . 129
5.2.9 glp ios branch upon — choose variable to branch upon . . . . . . . . . . . . . 129
5.2.10 glp ios terminate — terminate the solution process . . . . . . . . . . . . . . . 130
5.3 The search tree exploring routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.3.1 glp ios tree size — determine size of the search tree . . . . . . . . . . . . . . 131
5.3.2 glp ios curr node — determine current active subproblem . . . . . . . . . . . 131
5.3.3 glp ios next node — determine next active subproblem . . . . . . . . . . . . 131
5.3.4 glp ios prev node — determine previous active subproblem . . . . . . . . . . 132
5.3.5 glp ios up node — determine parent subproblem . . . . . . . . . . . . . . . . 132
5.3.6 glp ios node level — determine subproblem level . . . . . . . . . . . . . . . . 132
5.3.7 glp ios node bound — determine subproblem local bound . . . . . . . . . . . 132
5.3.8 glp ios best node — find active subproblem with best local bound . . . . . . 133
5.4 The cut pool routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.4.1 glp ios pool size — determine current size of the cut pool . . . . . . . . . . . 134
5.4.2 glp ios add row — add constraint to the cut pool . . . . . . . . . . . . . . . . 134
5.4.3 glp ios del row — remove constraint from the cut pool . . . . . . . . . . . . . 135
5.4.4 glp ios clear pool — remove all constraints from the cut pool . . . . . . . . . 136
6 Miscellaneous API Routines 137

6.1 GLPK environment routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.1.1 glp init env — initialize GLPK environment . . . . . . . . . . . . . . . . . . . 137
6.1.2 glp version — determine library version . . . . . . . . . . . . . . . . . . . . . 137
6.1.3 glp free env — free GLPK environment . . . . . . . . . . . . . . . . . . . . . 138
6.1.4 glp printf — write formatted output to terminal . . . . . . . . . . . . . . . . 138
6.1.5 glp vprintf — write formatted output to terminal . . . . . . . . . . . . . . . . 138
6.1.6 glp term out — enable/disable terminal output . . . . . . . . . . . . . . . . . 139
6.1.7 glp term hook — intercept terminal output . . . . . . . . . . . . . . . . . . . 139
6.1.8 glp open tee — start copying terminal output . . . . . . . . . . . . . . . . . . 140
6.1.9 glp close tee — stop copying terminal output . . . . . . . . . . . . . . . . . . 140
6.1.10 glp error — display error message and terminate execution . . . . . . . . . . 140
6.1.11 glp at error — check for error state . . . . . . . . . . . . . . . . . . . . . . . . 141
6.1.12 glp assert — check logical condition . . . . . . . . . . . . . . . . . . . . . . . 141
6.1.13 glp error hook — install hook to intercept abnormal termination . . . . . . . 141
6.1.14 glp alloc — allocate memory block . . . . . . . . . . . . . . . . . . . . . . . . 142
6.1.15 glp realloc — reallocate memory block . . . . . . . . . . . . . . . . . . . . . . 142
6.1.16 glp free — free memory block . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.1.17 glp mem usage — get memory usage information . . . . . . . . . . . . . . . . 143
6.1.18 glp mem limit — set memory usage limit . . . . . . . . . . . . . . . . . . . . 143
A Installing GLPK on Your Computer 144
7
A.1 Downloading the distribution tarball . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
A.2 Unpacking the distribution tarball . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
A.3 Configuring the package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
A.4 Compiling the package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
A.5 Checking the package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
A.6 Installing the package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
A.7 Uninstalling the package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
B MPS Format 148

B.1 Fixed MPS Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
B.2 Free MPS Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
B.3 NAME indicator card . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
B.4 ROWS section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
B.5 COLUMNS section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
B.6 RHS section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
B.7 RANGES section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
B.8 BOUNDS section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
B.9 ENDATA indicator card . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
B.10 Specifying objective function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
B.11 Example of MPS file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
B.12 MIP features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
C CPLEX LP Format 158

C.1 Prelude . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
C.2 Objective function definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
C.3 Constraints section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
C.4 Bounds section . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
C.5 General, integer, and binary sections . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
C.6 End keyword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
C.7 Example of CPLEX LP file . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
D Stand-alone LP/MIP Solver 164
E External Software Used In GLPK 167

E.1 AMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
E.2 COLAMD/SYMAMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
E.3 MiniSat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
E.4 zlib . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
GNU General Public License 170
8
Chapter 1
Introduction
GLPK (GNU Linear Programming Kit) is a set of routines written in the ANSI C program-
ming language and organized in the form of a callable library. It is intended for solving linear
programming (LP), mixed integer programming (MIP), and other related problems.
1.1 LP problem
GLPK assumes the following formulation of the linear programming (LP) problem:
minimize (or maximize)
z = c1 xm+1 + c2 xm+2 + . . . + cn xm+n + c0 (1.1)
subject to linear constraints
x1 = a11 xm+1 + a12 xm+2 + . . . + a1n xm+n

x2 = a21 xm+1 + a22 xm+2 + . . . + a2n xm+n
(1.2)
. . . . . . . . . . . . . .
xm = am1 xm+1 + am2 xm+2 + . . . + amn xm+n
and bounds of variables

l1 ≤ x1 ≤ u1
l2 ≤ x2 ≤ u2
(1.3)
. . . . .
lm+n ≤ xm+n ≤ um+n
where: x1 , x2 , . . . , xm are auxiliary variables; xm+1 , xm+2 , . . . , xm+n are structural variables; z is
the objective function; c1 , c2 , . . . , cn are objective coefficients; c0 is the constant term (“shift”) of
the objective function; a11 , a12 , . . . , amn are constraint coefficients; l1 , l2 , . . . , lm+n are lower bounds
of variables; u1 , u2 , . . . , um+n are upper bounds of variables.
Auxiliary variables are also called rows, because they correspond to rows of the constraint
matrix (i.e. a matrix built of the constraint coefficients). Similarly, structural variables are also
called columns, because they correspond to columns of the constraint matrix.
Bounds of variables can be finite as well as infinite. Besides, lower and upper bounds can be
equal to each other. Thus, the following types of variables are possible:
9
Bounds of variable Type of variable
−∞ < xk < +∞ Free (unbounded) variable
lk ≤ xk < +∞ Variable with lower bound
−∞ < xk ≤ uk Variable with upper bound
lk ≤ xk ≤ uk Double-bounded variable
lk = xk = uk Fixed variable
Note that the types of variables shown above are applicable to structural as well as to auxiliary
variables.
To solve the LP problem (1.1)—(1.3) is to find such values of all structural and auxiliary
variables, which:
— satisfy to all the linear constraints (1.2), and
— are within their bounds (1.3), and
— provide smallest (in case of minimization) or largest (in case of maximization) value of the
objective function (1.1).
1.2 MIP problem
Mixed integer linear programming (MIP) problem is an LP problem in which some variables are
additionally required to be integer.
GLPK assumes that MIP problem has the same formulation as ordinary (pure) LP problem
(1.1)—(1.3), i.e. includes auxiliary and structural variables, which may have lower and/or upper
bounds. However, in case of MIP problem some variables may be required to be integer. This
additional constraint means that a value of each integer variable must be only integer number.
(Should note that GLPK allows only structural variables to be of integer kind.)
1.3 Using the package
1.3.1 Brief example
In order to understand what GLPK is from the user’s standpoint, consider the following simple
LP problem:
maximize
z = 10x1 + 6x2 + 4x3
subject to
x1 + x2 + x3 ≤ 100
10x1 + 4x2 + 5x3 ≤ 600
2x1 + 2x2 + 6x3 ≤ 300
where all variables are non-negative
x1 ≥ 0, x2 ≥ 0, x3 ≥ 0
At first, this LP problem should be transformed to the standard form (1.1)—(1.3). This can be
easily done by introducing auxiliary variables, by one for each original inequality constraint. Thus,
the problem can be reformulated as follows:
10
maximize
z = 10x1 + 6x2 + 4x3
subject to
p = x1 + x2 + x3
q = 10x1 + 4x2 + 5x3
r = 2x1 + 2x2 + 6x3
−∞ < p ≤ 100 0 ≤ x1 < +∞
−∞ < q ≤ 600 0 ≤ x2 < +∞
−∞ < r ≤ 300 0 ≤ x3 < +∞
where p, q, r are auxiliary variables (rows), and x1 , x2 , x3 are structural variables (columns).
The example C program shown below uses GLPK API routines in order to solve this LP prob-
lem.1
/* sample.c */
#include <stdio.h>
#include <stdlib.h>
#include <glpk.h>
int main(void)
{ glp_prob *lp;
int ia[1+1000], ja[1+1000];
double ar[1+1000], z, x1, x2, x3;
s1: lp = glp_create_prob();
s2: glp_set_prob_name(lp, "sample");
s3: glp_set_obj_dir(lp, GLP_MAX);
s4: glp_add_rows(lp, 3);
s5: glp_set_row_name(lp, 1, "p");
s6: glp_set_row_bnds(lp, 1, GLP_UP, 0.0, 100.0);
s7: glp_set_row_name(lp, 2, "q");
s9: glp_set_row_name(lp, 3, "r");
s11: glp_add_cols(lp, 3);
s12: glp_set_col_name(lp, 1, "x1");
s13: glp_set_col_bnds(lp, 1, GLP_LO, 0.0, 0.0);
s14: glp_set_obj_coef(lp, 1, 10.0);
s21: ia[1] = 1, ja[1] = 1, ar[1] = 1.0; /* a[1,1] = 1 */
s22: ia[2] = 1, ja[2] = 2, ar[2] = 1.0; /* a[1,2] = 1 */
s23: ia[3] = 1, ja[3] = 3, ar[3] = 1.0; /* a[1,3] = 1 */
s24: ia[4] = 2, ja[4] = 1, ar[4] = 10.0; /* a[2,1] = 10 */
s25: ia[5] = 3, ja[5] = 1, ar[5] = 2.0; /* a[3,1] = 2 */
s26: ia[6] = 2, ja[6] = 2, ar[6] = 4.0; /* a[2,2] = 4 */
1
If you just need to solve LP or MIP instance, you may write it in MPS or CPLEX LP format and then use the
GLPK stand-alone solver to obtain a solution. This is much less time-consuming than programming in C with GLPK
API routines.
11
s27: ia[7] = 3, ja[7] = 2, ar[7] = 2.0; /* a[3,2] = 2 */
s28: ia[8] = 2, ja[8] = 3, ar[8] = 5.0; /* a[2,3] = 5 */
s29: ia[9] = 3, ja[9] = 3, ar[9] = 6.0; /* a[3,3] = 6 */
s30: glp_load_matrix(lp, 9, ia, ja, ar);
s31: glp_simplex(lp, NULL);
s32: z = glp_get_obj_val(lp);
s33: x1 = glp_get_col_prim(lp, 1);
s36: printf("\nz = %g; x1 = %g; x2 = %g; x3 = %g\n",
z, x1, x2, x3);
s37: glp_delete_prob(lp);
return 0;
}
/* eof */
The statement s1 creates a problem object. Being created the object is initially empty. The
statement s2 assigns a symbolic name to the problem object.
The statement s3 calls the routine glp_set_obj_dir in order to set the optimization direction
flag, where GLP_MAX means maximization.
The statement s4 adds three rows to the problem object.
The statement s5 assigns the symbolic name ‘p’ to the first row, and the statement s6 sets the
type and bounds of the first row, where GLP_UP means that the row has an upper bound. The
statements s7, s8, s9, s10 are used in the same way in order to assign the symbolic names ‘q’ and
‘r’ to the second and third rows and set their types and bounds.
The statement s11 adds three columns to the problem object.
The statement s12 assigns the symbolic name ‘x1’ to the first column, the statement s13 sets
the type and bounds of the first column, where GLP_LO means that the column has an lower bound,
and the statement s14 sets the objective coefficient for the first column. The statements s15—s20
are used in the same way in order to assign the symbolic names ‘x2’ and ‘x3’ to the second and
third columns and set their types, bounds, and objective coefficients.
The statements s21—s29 prepare non-zero elements of the constraint matrix (i.e. constraint
coefficients). Row indices of each element are stored in the array ia, column indices are stored in
the array ja, and numerical values of corresponding elements are stored in the array ar. Then the
statement s30 calls the routine glp_load_matrix, which loads information from these three arrays
into the problem object.
Now all data have been entered into the problem object, and therefore the statement s31 calls
the routine glp_simplex, which is a driver to the simplex method, in order to solve the LP problem.
This routine finds an optimal solution and stores all relevant information back into the problem
object.
The statement s32 obtains a computed value of the objective function, and the statements s33—
s35 obtain computed values of structural variables (columns), which correspond to the optimal basic
solution found by the solver.
The statement s36 writes the optimal solution to the standard output. The printout may look
like follows:
12
* 0: objval = 0.000000000e+00 infeas = 0.000000000e+00 (0)
* 2: objval = 7.333333333e+02 infeas = 0.000000000e+00 (0)
OPTIMAL SOLUTION FOUND
z = 733.333; x1 = 33.3333; x2 = 66.6667; x3 = 0
Finally, the statement s37 calls the routine glp_delete_prob, which frees all the memory
allocated to the problem object.
1.3.2 Compiling
The GLPK package has the only header file glpk.h, which should be available on compiling a
C (or C++) program using GLPK API routines.
If the header file is installed in the default location /usr/local/include, the following typical
command may be used to compile, say, the example C program described above with the GNU C
compiler:
$ gcc -c sample.c
If glpk.h is not in the default location, the corresponding directory containing it should be
made known to the C compiler through -I option, for example:
$ gcc -I/foo/bar/glpk-4.15/include -c sample.c
In any case the compilation results in an object file sample.o.
1.3.3 Linking
The GLPK library is a single file libglpk.a. (On systems which support shared libraries there
may be also a shared version of the library libglpk.so.)
If the library is installed in the default location /usr/local/lib, the following typical command
may be used to link, say, the example C program described above against with the library:
$ gcc sample.o -lglpk -lm
If the GLPK library is not in the default location, the corresponding directory containing it
should be made known to the linker through -L option, for example:
$ gcc -L/foo/bar/glpk-4.15 sample.o -lglpk -lm
Depending on configuration of the package linking against with the GLPK library may require
optional libraries, in which case these libraries should be also made known to the linker, for example:
$ gcc sample.o -lglpk -lgmp -lm
For more details about configuration options of the GLPK package see Appendix A, page 144.
13
Chapter 2
Basic API Routines
2.1 General conventions
2.1.1 Library header
All GLPK API data types and routines are defined in the header file glpk.h. It should be
included in all source files which use GLPK API, either directly or indirectly through some other
header file as follows:
#include <glpk.h>
2.1.2 Error handling
If some GLPK API routine detects erroneous or incorrect data passed by the application pro-
gram, it writes appropriate diagnostic messages to the terminal and then abnormally terminates
the application program. In most practical cases this allows to simplify programming by avoiding
numerous checks of return codes. Thus, in order to prevent crashing the application program should
check all data, which are suspected to be incorrect, before calling GLPK API routines.
Should note that this kind of error handling is used only in cases of incorrect data passed by
the application program. If, for example, the application program calls some GLPK API routine
to read data from an input file and these data are incorrect, the GLPK API routine reports about
error in the usual way by means of the return code.
2.1.3 Thread safety
The standard version of GLPK API is not thread safe and therefore should not be used in
multi-threaded programs.
2.1.4 Array indexing
Normally all GLPK API routines start array indexing from 1, not from 0 (except the specially
stipulated cases). This means, for example, that if some vector x of the length n is passed as an
14
array to some GLPK API routine, the latter expects vector components to be placed in locations
x[1], x[2], . . . , x[n], and the location x[0] normally is not used.
To avoid indexing errors it is most convenient and most reliable to declare the array x as follows:
double x[1+n];
or to allocate it as follows:
double *x;
. . .
x = calloc(1+n, sizeof(double));
. . .
In both cases one extra location x[0] is reserved that allows passing the array to GLPK routines
in a usual way.
2.2 Problem object
All GLPK API routines deal with so called problem object, which is a program object of type
glp_prob and intended to represent a particular LP or MIP instance.
The type glp_prob is a data structure declared in the header file glpk.h as follows:
typedef struct glp_prob glp_prob;
Problem objects (i.e. program objects of the glp_prob type) are allocated and managed in-
ternally by the GLPK API routines. The application program should never use any members
of the glp_prob structure directly and should deal only with pointers to these objects (that is,
glp_prob * values).
The problem object consists of the following segments:
— problem segment,
— basis segment,
— interior-point segment, and
— MIP segment.
2.2.1 Problem segment
The problem segment contains original LP/MIP data, which corresponds to the problem formu-
lation (1.1)—(1.3) (see Section 1.1, page 9). This segment includes the following components:
— rows (auxiliary variables),
— columns (structural variables),
— objective function, and
— constraint matrix.
Rows and columns have the same set of the following attributes:
— ordinal number,
— symbolic name (1 up to 255 arbitrary graphic characters),
15
— type (free, lower bound, upper bound, double bound, fixed),
— numerical values of lower and upper bounds,
— scale factor.
Ordinal numbers are intended for referencing rows and columns. Row ordinal numbers are
integers 1, 2, . . . , m, and column ordinal numbers are integers 1, 2, . . . , n, where m and n are, re-
spectively, the current number of rows and columns in the problem object.
Symbolic names are intended for informational purposes. They also can be used for referencing
rows and columns.
Types and bounds of rows (auxiliary variables) and columns (structural variables) are explained
above (see Section 1.1, page 9).
Scale factors are used internally for scaling rows and columns of the constraint matrix.
Information about the objective function includes numerical values of objective coefficients and
a flag, which defines the optimization direction (i.e. minimization or maximization).
The constraint matrix is a m × n rectangular matrix built of constraint coefficients aij , which
defines the system of linear constraints (1.2) (see Section 1.1, page 9). This matrix is stored in the
problem object in both row-wise and column-wise sparse formats.
Once the problem object has been created, the application program can access and modify any
components of the problem segment in arbitrary order.
2.2.2 Basis segment
The basis segment of the problem object keeps information related to the current basic solution.
It includes:
— row and column statuses,
— basic solution statuses,
— factorization of the current basis matrix, and
— basic solution components.
The row and column statuses define which rows and columns are basic and which are non-basic.
These statuses may be assigned either by the application program of by some API routines. Note
that these statuses are always defined independently on whether the corresponding basis is valid
or not.
The basic solution statuses include the primal status and the dual status, which are set by the
simplex-based solver once the problem has been solved. The primal status shows whether a primal
basic solution is feasible, infeasible, or undefined. The dual status shows the same for a dual basic
solution.
The factorization of the basis matrix is some factorized form (like LU-factorization) of the
current basis matrix (defined by the current row and column statuses). The factorization is used
by simplex-based solvers and kept when the solver terminates the search. This feature allows
efficiently reoptimizing the problem after some modifications (for example, after changing some
bounds or objective coefficients). It also allows performing the post-optimal analysis (for example,
computing components of the simplex table, etc.).
16
The basic solution components include primal and dual values of all auxiliary and structural
variables for the most recently obtained basic solution.
2.2.3 Interior-point segment
The interior-point segment contains interior-point solution components, which include the solu-
tion status, and primal and dual values of all auxiliary and structural variables.
2.2.4 MIP segment
The MIP segment is used only for MIP problems. This segment includes:
— column kinds,
— MIP solution status, and
— MIP solution components.
The column kinds define which columns (i.e. structural variables) are integer and which are
continuous.
The MIP solution status is set by the MIP solver and shows whether a MIP solution is integer
optimal, integer feasible (non-optimal), or undefined.
The MIP solution components are computed by the MIP solver and include primal values of all
auxiliary and structural variables for the most recently obtained MIP solution.
Note that in case of MIP problem the basis segment corresponds to the optimal solution of LP
relaxation, which is also available to the application program.
Currently the search tree is not kept in the MIP segment, so if the search has been terminated,
it cannot be continued.
17
2.3 Problem creating and modifying routines
2.3.1 glp create prob — create problem object
Synopsis
glp_prob *glp_create_prob(void);
Description
The routine glp_create_prob creates a new problem object, which initially is “empty”, i.e.
has no rows and columns.
Returns
The routine returns a pointer to the created object, which should be used in any subsequent
operations on this object.
2.3.2 glp set prob name — assign (change) problem name
Synopsis
void glp_set_prob_name(glp_prob *P, const char *name);
Description
The routine glp_set_prob_name assigns a given symbolic name (1 up to 255 characters) to the
specified problem object.
If the parameter name is NULL or empty string, the routine erases an existing symbolic name of
the problem object.
2.3.3 glp set obj name — assign (change) objective function name
Synopsis
void glp_set_obj_name(glp_prob *P, const char *name);
Description
The routine glp_set_obj_name assigns a given symbolic name (1 up to 255 characters) to the
objective function of the specified problem object.
If the parameter name is NULL or empty string, the routine erases an existing symbolic name of
the objective function.
18
2.3.4 glp set obj dir — set (change) optimization direction flag
Synopsis
void glp_set_obj_dir(glp_prob *P, int dir);
Description
The routine glp_set_obj_dir sets (changes) the optimization direction flag (i.e. “sense” of the
objective function) as specified by the parameter dir:
GLP_MIN means minimization;
GLP_MAX means maximization.
Note that by default the problem is minimization.
2.3.5 glp add rows — add new rows to problem object
Synopsis
int glp_add_rows(glp_prob *P, int nrs);
Description
The routine glp_add_rows adds nrs rows (constraints) to the specified problem object. New
rows are always added to the end of the row list, so the ordinal numbers of existing rows are not
changed.
Being added each new row is initially free (unbounded) and has empty list of the constraint
coefficients.
Each new row becomes a non-active (non-binding) constraint, i.e. the corresponding auxiliary
variable is marked as basic.
If the basis factorization exists, adding row(s) invalidates it.
Returns
The routine glp_add_rows returns the ordinal number of the first new row added to the problem
object.
2.3.6 glp add cols — add new columns to problem object
Synopsis
int glp_add_cols(glp_prob *P, int ncs);
Description
The routine glp_add_cols adds ncs columns (structural variables) to the specified problem
object. New columns are always added to the end of the column list, so the ordinal numbers of
existing columns are not changed.
Being added each new column is initially fixed at zero and has empty list of the constraint
coefficients.
Each new column is marked as non-basic, i.e. zero value of the corresponding structural variable
becomes an active (binding) bound.
19
If the basis factorization exists, it remains valid.
Returns
The routine glp_add_cols returns the ordinal number of the first new column added to the
problem object.
2.3.7 glp set row name — assign (change) row name
Synopsis
void glp_set_row_name(glp_prob *P, int i, const char *name);
Description
The routine glp_set_row_name assigns a given symbolic name (1 up to 255 characters) to i-th
row (auxiliary variable) of the specified problem object.
If the parameter name is NULL or empty string, the routine erases an existing name of i-th row.
2.3.8 glp set col name — assign (change) column name
Synopsis
void glp_set_col_name(glp_prob *P, int j, const char *name);
Description
The routine glp_set_col_name assigns a given symbolic name (1 up to 255 characters) to j-th
column (structural variable) of the specified problem object.
If the parameter name is NULL or empty string, the routine erases an existing name of j-th
column.
2.3.9 glp set row bnds — set (change) row bounds
Synopsis
void glp set row bnds(glp prob *P, int i, int type, double lb, double ub);
Description
The routine glp_set_row_bnds sets (changes) the type and bounds of i-th row (auxiliary
variable) of the specified problem object.
The parameters type, lb, and ub specify the type, lower bound, and upper bound, respectively,
as follows:
Type Bounds Comment
GLP_FR −∞ < x < +∞ Free (unbounded) variable
GLP_LO lb ≤ x < +∞ Variable with lower bound
GLP_UP −∞ < x ≤ ub Variable with upper bound
GLP_DB lb ≤ x ≤ ub Double-bounded variable
GLP_FX lb = x = ub Fixed variable
where x is the auxiliary variable associated with i-th row.
20
If the row has no lower bound, the parameter lb is ignored. If the row has no upper bound,
the parameter ub is ignored. If the row is an equality constraint (i.e. the corresponding auxiliary
variable is of fixed type), only the parameter lb is used while the parameter ub is ignored.
Being added to the problem object each row is initially free, i.e. its type is GLP_FR.
2.3.10 glp set col bnds — set (change) column bounds
Synopsis
void glp set col bnds(glp prob *P, int j, int type, double lb, double ub);
Description
The routine glp_set_col_bnds sets (changes) the type and bounds of j-th column (structural
variable) of the specified problem object.
The parameters type, lb, and ub specify the type, lower bound, and upper bound, respectively,
as follows:
Type Bounds Comment
GLP_FR −∞ < x < +∞ Free (unbounded) variable
GLP_LO lb ≤ x < +∞ Variable with lower bound
GLP_UP −∞ < x ≤ ub Variable with upper bound
GLP_DB lb ≤ x ≤ ub Double-bounded variable
GLP_FX lb = x = ub Fixed variable
where x is the structural variable associated with j-th column.
If the column has no lower bound, the parameter lb is ignored. If the column has no upper
bound, the parameter ub is ignored. If the column is of fixed type, only the parameter lb is used
while the parameter ub is ignored.
Being added to the problem object each column is initially fixed at zero, i.e. its type is GLP_FX
and both bounds are 0.
2.3.11 glp set obj coef — set (change) objective coefficient or constant term
Synopsis
void glp_set_obj_coef(glp_prob *P, int j, double coef);
Description
The routine glp_set_obj_coef sets (changes) the objective coefficient at j-th column (struc-
tural variable). A new value of the objective coefficient is specified by the parameter coef.
If the parameter j is 0, the routine sets (changes) the constant term (“shift”) of the objective
function.
21
2.3.12 glp set mat row — set (replace) row of the constraint matrix
Synopsis
void glp_set_mat_row(glp_prob *P, int i, int len, const int ind[],
const double val[]);
Description
The routine glp_set_mat_row stores (replaces) the contents of i-th row of the constraint matrix
of the specified problem object.
Column indices and numerical values of new row elements should be placed in locations
ind[1], . . . , ind[len] and val[1], . . . , val[len], respectively, where 0 ≤ len ≤ n is the new
length of i-th row, n is the current number of columns in the problem object. Elements with
identical column indices are not allowed. Zero elements are allowed, but they are not stored in the
constraint matrix.
If the parameter len is 0, the parameters ind and/or val can be specified as NULL.
Note
If the basis factorization exists and changing the row changes coefficients at basic column(s),
the factorization is invalidated.
2.3.13 glp set mat col — set (replace) column of the constraint matrix
Synopsis
void glp_set_mat_col(glp_prob *P, int j, int len, const int ind[],
const double val[]);
Description
The routine glp_set_mat_col stores (replaces) the contents of j-th column of the constraint
matrix of the specified problem object.
Row indices and numerical values of new column elements should be placed in locations
ind[1], . . . , ind[len] and val[1], . . . , val[len], respectively, where 0 ≤ len ≤ m is the new
length of j-th column, m is the current number of rows in the problem object. Elements with
identical row indices are not allowed. Zero elements are allowed, but they are not stored in the
constraint matrix.
If the parameter len is 0, the parameters ind and/or val can be specified as NULL.
Note
If the basis factorization exists, changing the column corresponding to a basic structural variable
invalidates it.
22
2.3.14 glp load matrix — load (replace) the whole constraint matrix
Synopsis
void glp_load_matrix(glp_prob *P, int ne, const int ia[],
const int ja[], const double ar[]);
Description
The routine glp_load_matrix loads the constraint matrix passed in the arrays ia, ja, and ar
into the specified problem object. Before loading the current contents of the constraint matrix is
destroyed.
Constraint coefficients (elements of the constraint matrix) should be specified as triplets (ia[k],
ja[k], ar[k]) for k = 1, . . . , ne, where ia[k] is the row index, ja[k] is the column index, and
ar[k] is a numeric value of corresponding constraint coefficient. The parameter ne specifies the
total number of (non-zero) elements in the matrix to be loaded. Coefficients with identical indices
are not allowed. Zero coefficients are allowed, however, they are not stored in the constraint matrix.
If the parameter ne is 0, the parameters ia, ja, and/or ar can be specified as NULL.
Note
If the basis factorization exists, this operation invalidates it.
2.3.15 glp check dup — check for duplicate elements in sparse matrix
Synopsis
int glp check dup(int m, int n, int ne, const int ia[], const int ja[]);
Description
The routine glp_check_dup checks for duplicate elements (that is, elements with identical
indices) in a sparse matrix specified in the coordinate format.
The parameters m and n specifies, respectively, the number of rows and columns in the matrix,
m ≥ 0, n ≥ 0.
The parameter ne specifies the number of (structurally) non-zero elements in the matrix,
ne ≥ 0.
Elements of the matrix are specified as doublets (ia[k], ja[k]) for k = 1, . . . , ne, where ia[k] is a
row index, ja[k] is a column index.
The routine glp_check_dup can be used prior to a call to the routine glp_load_matrix to
check that the constraint matrix to be loaded has no duplicate elements.
Returns
0 the matrix representation is correct;
−k indices ia[k] or/and ja[k] are out of range;
+k element (ia[k], ja[k]) is duplicate.
23
2.3.16 glp sort matrix — sort elements of the constraint matrix
Synopsis
void glp_sort_matrix(glp_prob *P);
Description
The routine glp_sort_matrix sorts elements of the constraint matrix by rebuilding its row and
column linked lists.
On exit from the routine the constraint matrix is not changed, however, elements in the row
linked lists become ordered by ascending column indices, and the elements in the column linked
lists become ordered by ascending row indices.
2.3.17 glp del rows — delete rows from problem object
Synopsis
void glp_del_rows(glp_prob *P, int nrs, const int num[]);
Description
The routine glp_del_rows deletes rows from the specified problem object. Ordinal numbers of
rows to be deleted should be placed in locations num[1], . . . , num[nrs], where nrs > 0.
Note that deleting rows involves changing ordinal numbers of other rows remaining in the
problem object. New ordinal numbers of the remaining rows are assigned under the assumption
that the original order of rows is not changed. Let, for example, before deletion there be five rows
a, b, c, d, e with ordinal numbers 1, 2, 3, 4, 5, and let rows b and d have been deleted. Then after
deletion the remaining rows a, c, e are assigned new oridinal numbers 1, 2, 3.
If the basis factorization exists, deleting active (binding) rows, i.e. whose auxiliary variables are
marked as non-basic, invalidates it.
2.3.18 glp del cols — delete columns from problem object
Synopsis
void glp_del_cols(glp_prob *P, int ncs, const int num[]);
Description
The routine glp_del_cols deletes columns from the specified problem object. Ordinal numbers
of columns to be deleted should be placed in locations num[1], . . . , num[ncs], where ncs > 0.
Note that deleting columns involves changing ordinal numbers of other columns remaining in
the problem object. New ordinal numbers of the remaining columns are assigned under the as-
sumption that the original order of columns is not changed. Let, for example, before deletion there
be six columns p, q, r, s, t, u with ordinal numbers 1, 2, 3, 4, 5, 6, and let columns p, q, s have
been deleted. Then after deletion the remaining columns r, t, u are assigned new ordinal numbers
1, 2, 3.
If the basis factorization exists, deleting basic columns invalidates it.
24
2.3.19 glp copy prob — copy problem object content
Synopsis
void glp_copy_prob(glp_prob *dest, glp_prob *prob, int names);
Description
The routine glp_copy_prob copies the content of the problem object prob to the problem object
dest.
The parameter names is a flag. If it is GLP_ON, the routine also copies all symbolic names;
otherwise, if it is GLP_OFF, no symbolic names are copied.
2.3.20 glp erase prob — erase problem object content
Synopsis
void glp_erase_prob(glp_prob *P);
Description
The routine glp_erase_prob erases the content of the specified problem object. The ef-
fect of this operation is the same as if the problem object would be deleted with the routine
glp_delete_prob and then created anew with the routine glp_create_prob, with the only excep-
tion that the pointer to the problem object remains valid.
2.3.21 glp delete prob — delete problem object
Synopsis
void glp_delete_prob(glp_prob *P);
Description
The routine glp_delete_prob deletes a problem object, which the parameter lp points to,
freeing all the memory allocated to this object.
25
2.4 Problem retrieving routines
2.4.1 glp get prob name — retrieve problem name
Synopsis
const char *glp_get_prob_name(glp_prob *P);
Returns
The routine glp_get_prob_name returns a pointer to an internal buffer, which contains symbolic
name of the problem. However, if the problem has no assigned name, the routine returns NULL.
2.4.2 glp get obj name — retrieve objective function name
Synopsis
const char *glp_get_obj_name(glp_prob *P);
Returns
The routine glp_get_obj_name returns a pointer to an internal buffer, which contains symbolic
name assigned to the objective function. However, if the objective function has no assigned name,
the routine returns NULL.
2.4.3 glp get obj dir — retrieve optimization direction flag
Synopsis
int glp_get_obj_dir(glp_prob *P);
Returns
The routine glp_get_obj_dir returns the optimization direction flag (i.e. “sense” of the ob-
jective function):
GLP_MIN means minimization;
GLP_MAX means maximization.
2.4.4 glp get num rows — retrieve number of rows
Synopsis
int glp_get_num_rows(glp_prob *P);
Returns
The routine glp_get_num_rows returns the current number of rows in the specified problem
object.
26
2.4.5 glp get num cols — retrieve number of columns
Synopsis
int glp_get_num_cols(glp_prob *P);
Returns
The routine glp_get_num_cols returns the current number of columns in the specified problem
object.
2.4.6 glp get row name — retrieve row name
Synopsis
const char *glp_get_row_name(glp_prob *P, int i);
Returns
The routine glp_get_row_name returns a pointer to an internal buffer, which contains a symbolic
name assigned to i-th row. However, if the row has no assigned name, the routine returns NULL.
2.4.7 glp get col name — retrieve column name
Synopsis
const char *glp_get_col_name(glp_prob *P, int j);
Returns
The routine glp_get_col_name returns a pointer to an internal buffer, which contains a symbolic
name assigned to j-th column. However, if the column has no assigned name, the routine returns
NULL.
2.4.8 glp get row type — retrieve row type
Synopsis
int glp_get_row_type(glp_prob *P, int i);
Returns
The routine glp_get_row_type returns the type of i-th row, i.e. the type of corresponding
auxiliary variable, as follows:
GLP_FR — free (unbounded) variable;
GLP_LO — variable with lower bound;
GLP_UP — variable with upper bound;
GLP_DB — double-bounded variable;
GLP_FX — fixed variable.
27
2.4.9 glp get row lb — retrieve row lower bound
Synopsis
double glp_get_row_lb(glp_prob *P, int i);
Returns
The routine glp_get_row_lb returns the lower bound of i-th row, i.e. the lower bound of
corresponding auxiliary variable. However, if the row has no lower bound, the routine returns
-DBL_MAX.
2.4.10 glp get row ub — retrieve row upper bound
Synopsis
double glp_get_row_ub(glp_prob *P, int i);
Returns
The routine glp_get_row_ub returns the upper bound of i-th row, i.e. the upper bound of
corresponding auxiliary variable. However, if the row has no upper bound, the routine returns
+DBL_MAX.
2.4.11 glp get col type — retrieve column type
Synopsis
int glp_get_col_type(glp_prob *P, int j);
Returns
The routine glp_get_col_type returns the type of j-th column, i.e. the type of corresponding
structural variable, as follows:
GLP_FR — free (unbounded) variable;
GLP_LO — variable with lower bound;
GLP_UP — variable with upper bound;
GLP_DB — double-bounded variable;
GLP_FX — fixed variable.
2.4.12 glp get col lb — retrieve column lower bound
Synopsis
double glp_get_col_lb(glp_prob *P, int j);
Returns
The routine glp_get_col_lb returns the lower bound of j-th column, i.e. the lower bound of
corresponding structural variable. However, if the column has no lower bound, the routine returns
-DBL_MAX.
28
2.4.13 glp get col ub — retrieve column upper bound
Synopsis
double glp_get_col_ub(glp_prob *P, int j);
Returns
The routine glp_get_col_ub returns the upper bound of j-th column, i.e. the upper bound of
corresponding structural variable. However, if the column has no upper bound, the routine returns
+DBL_MAX.
2.4.14 glp get obj coef — retrieve objective coefficient or constant term
Synopsis
double glp_get_obj_coef(glp_prob *P, int j);
Returns
The routine glp_get_obj_coef returns the objective coefficient at j-th structural variable (col-
umn).
If the parameter j is 0, the routine returns the constant term (“shift”) of the objective function.
2.4.15 glp get num nz — retrieve number of constraint coefficients
Synopsis
int glp_get_num_nz(glp_prob *P);
Returns
The routine glp_get_num_nz returns the number of non-zero elements in the constraint matrix
of the specified problem object.
2.4.16 glp get mat row — retrieve row of the constraint matrix
Synopsis
int glp_get_mat_row(glp_prob *P, int i, int ind[], double val[]);
Description
The routine glp_get_mat_row scans (non-zero) elements of i-th row of the constraint matrix
of the specified problem object and stores their column indices and numeric values to locations
ind[1], . . . , ind[len] and val[1], . . . , val[len], respectively, where 0 ≤ len ≤ n is the number
of elements in i-th row, n is the number of columns.
The parameter ind and/or val can be specified as NULL, in which case corresponding information
is not stored.
Returns
The routine glp_get_mat_row returns the length len, i.e. the number of (non-zero) elements
in i-th row.
29
2.4.17 glp get mat col — retrieve column of the constraint matrix
Synopsis
int glp_get_mat_col(glp_prob *P, int j, int ind[], double val[]);
Description
The routine glp_get_mat_col scans (non-zero) elements of j-th column of the constraint ma-
trix of the specified problem object and stores their row indices and numeric values to locations
ind[1], . . . , ind[len] and val[1], . . . , val[len], respectively, where 0 ≤ len ≤ m is the number
of elements in j-th column, m is the number of rows.
The parameter ind and/or val can be specified as NULL, in which case corresponding information
is not stored.
Returns
The routine glp_get_mat_col returns the length len, i.e. the number of (non-zero) elements
in j-th column.
30
2.5 Row and column searching routines
Sometimes it may be necessary to find rows and/or columns by their names (assigned with the
routines glp_set_row_name and glp_set_col_name). Though a particular row/column can be
found by its name using simple enumeration of all rows/columns, in case of large instances such a
linear search may take too long time.
To significantly reduce the search time the application program may create the row/column
name index, which is an auxiliary data structure implementing a binary search. Even in worst
cases the search takes logarithmic time, i.e. the time needed to find a particular row (or column)
by its name is O(log2 m) (or O(log2 n)), where m and n are, resp., the number of rows and columns
in the problem object.
It is important to note that:
1. On creating the problem object with the routine glp_create_prob the name index is not
created.
2. The name index can be created (destroyed) at any time with the routine glp_create_index
(glp_delete_index). Having been created the name index becomes part of the corresponding
problem object.
3. The time taken to create the name index is O[(m + n) log2 (m + n)], so it is recommended to
create the index only once, for example, just after the problem object was created.
4. If the name index exists, it is automatically updated every time the name of a row/column
is assigned/changed. The update operation takes logarithmic time.
5. If the name index does not exist, the application should not call the routines glp_find_row
and glp_find_col. Otherwise, an error message will be issued and abnormal program termination
will occur.
6. On destroying the problem object with the routine glp_delete_prob, the name index, if
exists, is automatically destroyed.
2.5.1 glp create index — create the name index
Synopsis
void glp_create_index(glp_prob *P);
Description
The routine glp_create_index creates the name index for the specified problem object. The
name index is an auxiliary data structure, which is intended to quickly (i.e. for logarithmic time)
find rows and columns by their names.
This routine can be called at any time. If the name index already exists, the routine does
nothing.
31
2.5.2 glp find row — find row by its name
Synopsis
int glp_find_row(glp_prob *P, const char *name);
Returns
The routine glp_find_row returns the ordinal number of a row, which is assigned the specified
symbolic name. If no such row exists, the routine returns 0.
2.5.3 glp find col — find column by its name
Synopsis
int glp_find_col(glp_prob *P, const char *name);
Returns
The routine glp_find_col returns the ordinal number of a column, which is assigned the
specified symbolic name. If no such column exists, the routine returns 0.
2.5.4 glp delete index — delete the name index
Synopsis
void glp_delete_index(glp_prob *P);
Description
The routine glp_delete_index deletes the name index previously created by the routine
glp_create_index and frees the memory allocated to this auxiliary data structure.
This routine can be called at any time. If the name index does not exist, the routine does
nothing.
32
2.6 Problem scaling routines
2.6.1 Background
In GLPK the scaling means a linear transformation applied to the constraint matrix to improve
its numerical properties.1
The main equality is the following:
e = RAS,
A (2.1)
where A = (aij ) is the original constraint matrix, R = (rii ) > 0 is a diagonal matrix used to scale
rows (constraints), S = (sjj ) > 0 is a diagonal matrix used to scale columns (variables), A e is the
scaled constraint matrix.
From (2.1) it follows that in the scaled problem instance each original constraint coefficient aij
is replaced by corresponding scaled constraint coefficient:
e
aij = rii aij sjj . (2.2)
Note that the scaling is performed internally and therefore transparently to the user. This
means that on API level the user always deal with unscaled data.
Scale factors rii and sjj can be set or changed at any time either directly by the application
program in a problem specific way (with the routines glp_set_rii and glp_set_sjj), or by some
API routines intended for automatic scaling.
2.6.2 glp set rii — set (change) row scale factor
Synopsis
void glp_set_rii(glp_prob *P, int i, double rii);
Description
The routine glp_set_rii sets (changes) the scale factor rii for i-th row of the specified problem
object.
2.6.3 glp set sjj — set (change) column scale factor
Synopsis
void glp_set_sjj(glp_prob *P, int j, double sjj);
Description
The routine glp_set_sjj sets (changes) the scale factor sjj for j-th column of the specified
problem object.
1
In many cases a proper scaling allows making the constraint matrix to be better conditioned, i.e. decreasing its
condition number, that makes computations numerically more stable.
33
2.6.4 glp get rii — retrieve row scale factor
Synopsis
double glp_get_rii(glp_prob *P, int i);
Returns
The routine glp_get_rii returns current scale factor rii for i-th row of the specified problem
object.
2.6.5 glp get sjj — retrieve column scale factor

Synopsis
double glp_get_sjj(glp_prob *P, int j);
Returns
The routine glp_get_sjj returns current scale factor sjj for j-th column of the specified problem
object.
2.6.6 glp scale prob — scale problem data

Synopsis
void glp_scale_prob(glp_prob *P, int flags);
Description
The routine glp_scale_prob performs automatic scaling of problem data for the specified
problem object.
The parameter flags specifies scaling options used by the routine. The options can be combined
with the bitwise OR operator and may be the following:
GLP_SF_GM — perform geometric mean scaling;
GLP_SF_EQ — perform equilibration scaling;
GLP_SF_2N — round scale factors to nearest power of two;
GLP_SF_SKIP — skip scaling, if the problem is well scaled.
The parameter flags may be also specified as GLP_SF_AUTO, in which case the routine chooses
the scaling options automatically.
2.6.7 glp unscale prob — unscale problem data

Synopsis
void glp_unscale_prob(glp_prob *P);
The routine glp_unscale_prob performs unscaling of problem data for the specified problem
object.
“Unscaling” means replacing the current scaling matrices R and S by unity matrices that cancels
the scaling effect.
34
2.7 LP basis constructing routines
2.7.1 Background
To start the search the simplex method needs a valid initial basis. In GLPK the basis is
completely defined by a set of statuses assigned to all (auxiliary and structural) variables, where
the status may be one of the following:
GLP_BS — basic variable;
GLP_NL — non-basic variable having active lower bound;
GLP_NU — non-basic variable having active upper bound;
GLP_NF — non-basic free variable;
GLP_NS — non-basic fixed variable.
The basis is valid, if the basis matrix, which is a matrix built of columns of the augmented
constraint matrix (I | − A) corresponding to basic variables, is non-singular. This, in particular,
means that the number of basic variables must be the same as the number of rows in the problem
object. (For more details see Section 4.2, page 104.)
Any initial basis may be constructed (or restored) with the API routines glp_set_row_stat and
glp_set_col_stat by assigning appropriate statuses to auxiliary and structural variables. Another
way to construct an initial basis is to use API routines like glp_adv_basis, which implement so
called crashing.2 Note that on normal exit the simplex solver remains the basis valid, so in case of
re-optimization there is no need to construct an initial basis from scratch.
2.7.2 glp set row stat — set (change) row status
Synopsis
void glp_set_row_stat(glp_prob *P, int i, int stat);
Description
The routine glp_set_row_stat sets (changes) the current status of i-th row (auxiliary variable)
as specified by the parameter stat:
GLP_BS — make the row basic (make the constraint inactive);
GLP_NL — make the row non-basic (make the constraint active);
GLP_NU — make the row non-basic and set it to the upper bound; if the row is not double-
bounded, this status is equivalent to GLP_NL (only in case of this routine);
GLP_NF — the same as GLP_NL (only in case of this routine);
GLP_NS — the same as GLP_NL (only in case of this routine).
2
This term is from early linear programming systems and means a heuristic to construct a valid initial basis.
35
2.7.3 glp set col stat — set (change) column status
Synopsis
void glp_set_col_stat(glp_prob *P, int j, int stat);
Description
The routine glp_set_col_stat sets (changes) the current status of j-th column (structural
variable) as specified by the parameter stat:
GLP_BS — make the column basic;
GLP_NL — make the column non-basic;
GLP_NU — make the column non-basic and set it to the upper bound; if the column is not
double-bounded, this status is equivalent to GLP_NL (only in case of this routine);
GLP_NF — the same as GLP_NL (only in case of this routine);
GLP_NS — the same as GLP_NL (only in case of this routine).
2.7.4 glp std basis — construct standard initial LP basis
Synopsis
void glp_std_basis(glp_prob *P);
Description
The routine glp_std_basis constructs the “standard” (trivial) initial LP basis for the specified
problem object.
In the “standard” LP basis all auxiliary variables (rows) are basic, and all structural variables
(columns) are non-basic (so the corresponding basis matrix is unity).
2.7.5 glp adv basis — construct advanced initial LP basis
Synopsis
void glp_adv_basis(glp_prob *P, int flags);
Description
The routine glp_adv_basis constructs an advanced initial LP basis for the specified problem
object.
The parameter flags is reserved for use in the future and must be specified as zero.
In order to construct the advanced initial LP basis the routine does the following:
1) includes in the basis all non-fixed auxiliary variables;
2) includes in the basis as many non-fixed structural variables as possible keeping the triangular
form of the basis matrix;
3) includes in the basis appropriate (fixed) auxiliary variables to complete the basis.
As a result the initial LP basis has as few fixed variables as possible and the corresponding basis
matrix is triangular.
36
2.7.6 glp cpx basis — construct Bixby’s initial LP basis
Synopsis
void glp_cpx_basis(glp_prob *P);
Description
The routine glp_cpx_basis constructs an initial basis for the specified problem object with the
algorithm proposed by R. Bixby.3
3
Robert E. Bixby, “Implementing the Simplex Method: The Initial Basis.” ORSA Journal on Computing, Vol. 4,
No. 3, 1992, pp. 267-84.
37
2.8 Simplex method routines
The simplex method is a well known efficient numerical procedure to solve LP problems.
On each iteration the simplex method transforms the original system of equaility constraints
(1.2) resolving them through different sets of variables to an equivalent system called the simplex
table (or sometimes the simplex tableau), which has the following form:
z = d1 (xN )1 + d2 (xN )2 + . . . + dn (xN )n
(xB )1 = ξ11 (xN )1 + ξ12 (xN )2 + . . . + ξ1n (xN )n
(xB )2 = ξ21 (xN )1 + ξ22 (xN )2 + . . . + ξ2n (xN )n (2.3)
. . . . . . . . . . . . . .
(xB )m = ξm1 (xN )1 + ξm2 (xN )2 + . . . + ξmn (xN )n
where: (xB )1 , (xB )2 , . . . , (xB )m are basic variables; (xN )1 , (xN )2 , . . . , (xN )n are non-basic variables;
d1 , d2 , . . . , dn are reduced costs; ξ11 , ξ12 , . . . , ξmn are coefficients of the simplex table. (May note
that the original LP problem (1.1)—(1.3) also has the form of a simplex table, where all equalities
are resolved through auxiliary variables.)
From the linear programming theory it is known that if an optimal solution of the LP problem
(1.1)—(1.3) exists, it can always be written in the form (2.3), where non-basic variables are set
on their bounds while values of the objective function and basic variables are determined by the
corresponding equalities of the simplex table.
A set of values of all basic and non-basic variables determined by the simplex table is called basic
solution. If all basic variables are within their bounds, the basic solution is called (primal) feasible,
otherwise it is called (primal) infeasible. A feasible basic solution, which provides a smallest (in
case of minimization) or a largest (in case of maximization) value of the objective function is called
optimal. Therefore, for solving LP problem the simplex method tries to find its optimal basic
solution.
Primal feasibility of some basic solution may be stated by simple checking if all basic variables
are within their bounds. Basic solution is optimal if additionally the following optimality conditions
are satisfied for all non-basic variables:
Status of (xN )j Minimization Maximization
(xN )j is free dj = 0 dj = 0
(xN )j is on its lower bound dj ≥ 0 dj ≤ 0
(xN )j is on its upper bound dj ≤ 0 dj ≥ 0
In other words, basic solution is optimal if there is no non-basic variable, which changing in the
feasible direction (i.e. increasing if it is free or on its lower bound, or decreasing if it is free or
on its upper bound) can improve (i.e. decrease in case of minimization or increase in case of
maximization) the objective function.
If all non-basic variables satisfy to the optimality conditions shown above (independently on
whether basic variables are within their bounds or not), the basic solution is called dual feasible,
otherwise it is called dual infeasible.
It may happen that some LP problem has no primal feasible solution due to incorrect
formulation — this means that its constraints conflict with each other. It also may happen that
some LP problem has unbounded solution again due to incorrect formulation — this means that
some non-basic variable can improve the objective function, i.e. the optimality conditions are vi-
olated, and at the same time this variable can infinitely change in the feasible direction meeting
38
no resistance from basic variables. (May note that in the latter case the LP problem has no dual
feasible solution.)
2.8.1 glp simplex — solve LP problem with the primal or dual simplex method
Synopsis
int glp_simplex(glp_prob *P, const glp_smcp *parm);
Description
The routine glp_simplex is a driver to the LP solver based on the simplex method. This
routine retrieves problem data from the specified problem object, calls the solver to solve the
problem instance, and stores results of computations back into the problem object.
The simplex solver has a set of control parameters. Values of the control parameters can be
passed in the structure glp_smcp, which the parameter parm points to. For detailed description
of this structure see paragraph “Control parameters” below. Before specifying some control pa-
rameters the application program should initialize the structure glp_smcp by default values of all
control parameters using the routine glp_init_smcp (see the next subsection). This is needed for
backward compatibility, because in the future there may appear new members in the structure
glp_smcp.
The parameter parm can be specified as NULL, in which case the solver uses default settings.
Returns
0 The LP problem instance has been successfully solved. (This code does not nec-
essarily mean that the solver has found optimal solution. It only means that the
solution process was successful.)
GLP_EBADB Unable to start the search, because the initial basis specified in the problem object
is invalid—the number of basic (auxiliary and structural) variables is not the same
as the number of rows in the problem object.
GLP_ESING Unable to start the search, because the basis matrix corresponding to the initial
basis is singular within the working precision.
GLP_ECOND Unable to start the search, because the basis matrix corresponding to the initial
basis is ill-conditioned, i.e. its condition number is too large.
GLP_EBOUND Unable to start the search, because some double-bounded (auxiliary or structural)
variables have incorrect bounds.
GLP_EFAIL The search was prematurely terminated due to the solver failure.
GLP_EOBJLL The search was prematurely terminated, because the objective function being max-
imized has reached its lower limit and continues decreasing (the dual simplex only).
GLP_EOBJUL The search was prematurely terminated, because the objective function being mini-
mized has reached its upper limit and continues increasing (the dual simplex only).
GLP_EITLIM The search was prematurely terminated, because the simplex iteration limit has
been exceeded.
GLP_ETMLIM The search was prematurely terminated, because the time limit has been exceeded.
39
GLP_ENOPFS The LP problem instance has no primal feasible solution (only if the LP presolver
is used).
GLP_ENODFS The LP problem instance has no dual feasible solution (only if the LP presolver is
used).
Built-in LP presolver
The simplex solver has built-in LP presolver. It is a subprogram that transforms the original
LP problem specified in the problem object to an equivalent LP problem, which may be easier for
solving with the simplex method than the original one. This is attained mainly due to reducing
the problem size and improving its numeric properties (for example, by removing some inactive
constraints or by fixing some non-basic variables). Once the transformed LP problem has been
solved, the presolver transforms its basic solution back to the corresponding basic solution of the
original problem.
Presolving is an optional feature of the routine glp_simplex, and by default it is disabled. In
order to enable the LP presolver the control parameter presolve should be set to GLP_ON (see
paragraph “Control parameters” below). Presolving may be used when the problem instance is
solved for the first time. However, on performing re-optimization the presolver should be disabled.
The presolving procedure is transparent to the API user in the sense that all necessary processing
is performed internally, and a basic solution of the original problem recovered by the presolver is
the same as if it were computed directly, i.e. without presolving.
Note that the presolver is able to recover only optimal solutions. If a computed solution is
infeasible or non-optimal, the corresponding solution of the original problem cannot be recovered
and therefore remains undefined. If you need to know a basic solution even if it is infeasible or
non-optimal, the presolver should be disabled.
Terminal output
Solving large problem instances may take a long time, so the solver reports some information
about the current basic solution, which is sent to the terminal. This information has the following
format:
nnn: obj = xxx infeas = yyy (num) cnt
where: ‘nnn’ is the iteration number, ‘xxx’ is the current value of the objective function (it is
unscaled and has correct sign); ‘yyy’ is the current sum of primal or dual infeasibilities (it is scaled
and therefore may be used only for visual estimating), ‘num’ is the current number of primal or dual
infeasibilities (phase I) or non-optimalities (phase II), ‘cnt’ is the number of basis factorizations
since the last terminal output.
The symbol preceding the iteration number indicates which phase of the simplex method is in
effect:
Blank means that the solver is searching for primal feasible solution using the primal simplex
or for dual feasible solution using the dual simplex;
Asterisk (*) means that the solver is searching for optimal solution using the primal simplex;
Hash (#) means that the solver is searching for optimal solution using the dual simplex.
40
Control parameters
This paragraph describes all control parameters currently used in the simplex solver. Symbolic
names of control parameters are names of corresponding members in the structure glp_smcp.
int msg lev (default: GLP MSG ALL)

Message level for terminal output:
GLP_MSG_OFF — no output;
GLP_MSG_ERR — error and warning messages only;
GLP_MSG_ON — normal output;
GLP_MSG_ALL — full output (including informational messages).
int meth (default: GLP PRIMAL)

Simplex method option:
GLP_PRIMAL — use two-phase primal simplex;
GLP_DUAL — use two-phase dual simplex;
GLP_DUALP — use two-phase dual simplex, and if it fails, switch to the primal simplex.
int pricing (default: GLP PT PSE)

Pricing technique:
GLP_PT_STD — standard (“textbook”);
GLP_PT_PSE — projected steepest edge.
int r test (default: GLP RT HAR)

Ratio test technique:
GLP_RT_STD — standard (“textbook”);
GLP_RT_HAR — Harris’ two-pass ratio test.
double tol bnd (default: 1e-7)

Tolerance used to check if the basic solution is primal feasible. (Do not change this parameter
without detailed understanding its purpose.)
double tol dj (default: 1e-7)

Tolerance used to check if the basic solution is dual feasible. (Do not change this parameter
without detailed understanding its purpose.)
double tol piv (default: 1e-9)

Tolerance used to choose eligble pivotal elements of the simplex table. (Do not change this
parameter without detailed understanding its purpose.)
41
double obj ll (default: -DBL MAX)
Lower limit of the objective function. If the objective function reaches this limit and continues
decreasing, the solver terminates the search. (Used in the dual simplex only.)
double obj ul (default: +DBL MAX)

Upper limit of the objective function. If the objective function reaches this limit and continues
increasing, the solver terminates the search. (Used in the dual simplex only.)
int it lim (default: INT MAX)

Simplex iteration limit.
int tm lim (default: INT MAX)

Searching time limit, in milliseconds.
int out frq (default: 500)

Output frequency, in iterations. This parameter specifies how frequently the solver sends infor-
mation about the solution process to the terminal.
int out dly (default: 0)

Output delay, in milliseconds. This parameter specifies how long the solver should delay sending
information about the solution process to the terminal.
int presolve (default: GLP OFF)

LP presolver option:
GLP_ON — enable using the LP presolver;
GLP_OFF — disable using the LP presolver.
42
Example 1
The following example main program reads LP problem instance in fixed MPS format from
file 25fv47.mps,4 constructs an advanced initial basis, solves the instance with the primal simplex
method (by default), and writes the solution to file 25fv47.txt.
/* spxsamp1.c */
#include <stdio.h>
#include <stdlib.h>
#include <glpk.h>
int main(void)
{ glp_prob *P;
P = glp_create_prob();
glp_read_mps(P, GLP_MPS_DECK, NULL, "25fv47.mps");
glp_adv_basis(P, 0);
glp_simplex(P, NULL);
glp_print_sol(P, "25fv47.txt");
glp_delete_prob(P);
return 0;
}
/* eof */
Below here is shown the terminal output from this example program.
Reading problem data from ’25fv47.mps’...
Problem: 25FV47
Objective: R0000
822 rows, 1571 columns, 11127 non-zeros
6919 records were read
One free row was removed
Constructing initial basis...
Size of triangular part is 812
GLPK Simplex Optimizer, v4.57
0: obj = 7.131703290e+03 inf = 2.145e+05 (204)
500: obj = 1.886711682e+04 inf = 8.273e+02 (36) 4
741: obj = 1.846047936e+04 inf = 5.575e-14 (0) 2
* 1000: obj = 9.220063473e+03 inf = 2.423e-14 (432) 2
* 1500: obj = 6.187659664e+03 inf = 1.019e-13 (368) 4
* 2000: obj = 5.503442062e+03 inf = 0.000e+00 (33) 5
* 2052: obj = 5.501845888e+03 inf = 0.000e+00 (0)
OPTIMAL LP SOLUTION FOUND
Writing basic solution to ’25fv47.txt’...
4
This instance in fixed MPS format can be found in the Netlib LP collection; see ftp://ftp.netlib.org/lp/data/.
43
Example 2
The following example main program solves the same LP problem instance as in Example 1
above, however, it uses the dual simplex method, which starts from the standard initial basis.
/* spxsamp2.c */
#include <stdio.h>
#include <stdlib.h>
#include <glpk.h>
int main(void)
{ glp_prob *P;
glp_smcp parm;
glp_init_smcp(&parm);
parm.meth = GLP_DUAL;
glp_simplex(P, &parm);
glp_print_sol(P, "25fv47.txt");
glp_delete_prob(P);
return 0;
}
/* eof */
Reading problem data from ’25fv47.mps’...
Problem: 25FV47
Objective: R0000
One free row was removed
GLPK Simplex Optimizer, v4.57
0: inf = 1.223e+02 (41)
258: inf = 3.091e-16 (0) 2
# 500: obj = -5.071287080e+03 inf = 2.947e-15 (292) 2
# 1000: obj = -1.352843873e+03 inf = 8.452e-15 (302) 5
# 1500: obj = 7.985859737e+02 inf = 1.127e-14 (263) 5
# 2000: obj = 3.059023029e+03 inf = 6.290e-11 (197) 4
# 2500: obj = 5.354770966e+03 inf = 7.172e-13 (130) 5
# 2673: obj = 5.501845888e+03 inf = 3.802e-16 (0) 2
OPTIMAL LP SOLUTION FOUND
Writing basic solution to ’25fv47.txt’...
44
2.8.2 glp exact — solve LP problem in exact arithmetic
Synopsis
int glp_exact(glp_prob *P, const glp_smcp *parm);
Description
The routine glp_exact is a tentative implementation of the primal two-phase simplex method
based on exact (rational) arithmetic. It is similar to the routine glp_simplex, however, for all
internal computations it uses arithmetic of rational numbers, which is exact in mathematical sense,
i.e. free of round-off errors unlike floating-point arithmetic.
Note that the routine glp_exact uses only three control parameters passed in the structure
glp_smcp, namely, msg_lev, it_lim, and tm_lim.
Returns
GLP_EBADB Unable to start the search, because the initial basis specified in the problem object
is invalid—the number of basic (auxiliary and structural) variables is not the same
as the number of rows in the problem object.
GLP_ESING Unable to start the search, because the basis matrix corresponding to the initial
basis is exactly singular.
GLP_EBOUND Unable to start the search, because some double-bounded (auxiliary or structural)
variables have incorrect bounds.
GLP_EFAIL The problem instance has no rows/columns.
GLP_EITLIM The search was prematurely terminated, because the simplex iteration limit has
been exceeded.
Note
Computations in exact arithmetic are very time-consuming, so solving LP problem with the
routine glp_exact from the very beginning is not a good idea. It is much better at first to find an
optimal basis with the routine glp_simplex and only then to call glp_exact, in which case only
a few simplex iterations need to be performed in exact arithmetic.
45
2.8.3 glp init smcp — initialize simplex solver control parameters
Synopsis
int glp_init_smcp(glp_smcp *parm);
Description
The routine glp_init_smcp initializes control parameters, which are used by the simplex solver,
with default values.
Default values of the control parameters are stored in a glp_smcp structure, which the parameter
parm points to.
2.8.4 glp get status — determine generic status of basic solution
Synopsis
int glp_get_status(glp_prob *P);
Returns
The routine glp_get_status reports the generic status of the current basic solution for the
specified problem object as follows:
GLP_OPT — solution is optimal;
GLP_FEAS — solution is feasible;
GLP_INFEAS — solution is infeasible;
GLP_NOFEAS — problem has no feasible solution;
GLP_UNBND — problem has unbounded solution;
GLP_UNDEF — solution is undefined.
More detailed information about the status of basic solution can be retrieved with the routines
glp_get_prim_stat and glp_get_dual_stat.
2.8.5 glp get prim stat — retrieve status of primal basic solution
Synopsis
int glp_get_prim_stat(glp_prob *P);
Returns
The routine glp_get_prim_stat reports the status of the primal basic solution for the specified
problem object as follows:
GLP_UNDEF — primal solution is undefined;
GLP_FEAS — primal solution is feasible;
GLP_INFEAS — primal solution is infeasible;
GLP_NOFEAS — no primal feasible solution exists.
46
2.8.6 glp get dual stat — retrieve status of dual basic solution
Synopsis
int glp_get_dual_stat(glp_prob *P);
Returns
The routine glp_get_dual_stat reports the status of the dual basic solution for the specified
problem object as follows:
GLP_UNDEF — dual solution is undefined;
GLP_FEAS — dual solution is feasible;
GLP_INFEAS — dual solution is infeasible;
GLP_NOFEAS — no dual feasible solution exists.
2.8.7 glp get obj val — retrieve objective value
Synopsis
double glp_get_obj_val(glp_prob *P);
Returns
The routine glp_get_obj_val returns current value of the objective function.
2.8.8 glp get row stat — retrieve row status
Synopsis
int glp_get_row_stat(glp_prob *P, int i);
Returns
The routine glp_get_row_stat returns current status assigned to the auxiliary variable asso-
ciated with i-th row as follows:
GLP_NL — non-basic variable on its lower bound;
GLP_NU — non-basic variable on its upper bound;
GLP_NF — non-basic free (unbounded) variable;
2.8.9 glp get row prim — retrieve row primal value
Synopsis
double glp_get_row_prim(glp_prob *P, int i);
Returns
The routine glp_get_row_prim returns primal value of the auxiliary variable associated with
i-th row.
47
2.8.10 glp get row dual — retrieve row dual value
Synopsis
double glp_get_row_dual(glp_prob *P, int i);
Returns
The routine glp_get_row_dual returns dual value (i.e. reduced cost) of the auxiliary variable
associated with i-th row.
2.8.11 glp get col stat — retrieve column status
Synopsis
int glp_get_col_stat(glp_prob *P, int j);
Returns
The routine glp_get_col_stat returns current status assigned to the structural variable asso-
ciated with j-th column as follows:
GLP_NL — non-basic variable on its lower bound;
GLP_NU — non-basic variable on its upper bound;
GLP_NF — non-basic free (unbounded) variable;
2.8.12 glp get col prim — retrieve column primal value
Synopsis
double glp_get_col_prim(glp_prob *P, int j);
Returns
The routine glp_get_col_prim returns primal value of the structural variable associated with
j-th column.
2.8.13 glp get col dual — retrieve column dual value
Synopsis
double glp_get_col_dual(glp_prob *P, int j);
Returns
The routine glp_get_col_dual returns dual value (i.e. reduced cost) of the structural variable
associated with j-th column.
48
2.8.14 glp get unbnd ray — determine variable causing unboundedness
Synopsis
int glp_get_unbnd_ray(glp_prob *P);
Returns
The routine glp_get_unbnd_ray returns the number k of a variable, which causes primal or
dual unboundedness. If 1 ≤ k ≤ m, it is k-th auxiliary variable, and if m + 1 ≤ k ≤ m + n, it is
(k − m)-th structural variable, where m is the number of rows, n is the number of columns in the
problem object. If such variable is not defined, the routine returns 0.
Note
If it is not exactly known which version of the simplex solver detected unboundedness, i.e.
whether the unboundedness is primal or dual, it is sufficient to check the status of the variable with
the routine glp_get_row_stat or glp_get_col_stat. If the variable is non-basic, the unbounded-
ness is primal, otherwise, if the variable is basic, the unboundedness is dual (the latter case means
that the problem has no primal feasible dolution).
49
2.9 Interior-point method routines
Interior-point methods (also known as barrier methods) are more modern and powerful numerical
methods for large-scale linear programming. Such methods are especially efficient for very sparse
LP problems and allow solving such problems much faster than the simplex method.
In brief, the GLPK interior-point solver works as follows.
At first, the solver transforms the original LP to a working LP in the standard format:
minimize
z = c1 xm+1 + c2 xm+2 + . . . + cn xm+n + c0 (2.4)
a11 xm+1 + a12 xm+2 + . . . + a1n xm+n = b1

a21 xm+1 + a22 xm+2 + . . . + a2n xm+n = b2
(2.5)
. . . . . . . . . . . . . .
am1 xm+1 + am2 xm+2 + . . . + amn xm+n = bm
and non-negative variables
x1 ≥ 0, x2 ≥ 0, . . . , xn ≥ 0 (2.6)
where: z is the objective function; x1 , . . . , xn are variables; c1 , . . . , cn are objective coefficients;

c0 is a constant term of the objective function; a11 , . . . , amn are constraint coefficients; b1 , . . . , bm
are right-hand sides.
Using vector and matrix notations the working LP (2.4)—(2.6) can be written as follows:
z = cT x + c0 → min, (2.7)
Ax = b, (2.8)
x ≥ 0, (2.9)
where: x = (xj ) is n-vector of variables, c = (cj ) is n-vector of objective coefficients, A = (aij ) is
m × n-matrix of constraint coefficients, and b = (bi ) is m-vector of right-hand sides.
Karush–Kuhn–Tucker optimality conditions for LP (2.7)—(2.9) are the following:
Ax = b, (2.10)
AT π + λ = c, (2.11)
λT x = 0, (2.12)
x ≥ 0, λ ≥ 0, (2.13)
where: π is m-vector of Lagrange multipliers (dual variables) for equality constraints (2.8),
λ is n-vector of Lagrange multipliers (dual variables) for non-negativity constraints (2.9),
(2.10) is the primal feasibility condition, (2.11) is the dual feasibility condition, (2.12) is the primal-
dual complementarity condition, and (2.13) is the non-negativity conditions.
The main idea of the primal-dual interior-point method is based on finding a point in the
primal-dual space (i.e. in the space of all primal and dual variables x, π, and λ), which satisfies
50
to all optimality conditions (2.10)—(2.13). Obviously, x-component of such point then provides an
optimal solution to the working LP (2.7)—(2.9).
To find the optimal point (x∗ , π ∗ , λ∗ ) the interior-point method attempts to solve the system
of equations (2.10)—(2.12), which is closed in the sense that the number of variables xj , πi , and
λj and the number equations are the same and equal to m + 2n. Due to condition (2.12) this
system of equations is non-linear, so it can be solved with a version of Newton’s method provided
with additional rules to keep the current point within the positive orthant as required by the
non-negativity conditions (2.13).
Finally, once the optimal point (x∗ , π ∗ , λ∗ ) has been found, the solver performs inverse trans-
formations to recover corresponding solution to the original LP passed to the solver from the
application program.
2.9.1 glp interior — solve LP problem with the interior-point method
Synopsis
int glp_interior(glp_prob *P, const glp_iptcp *parm);
Description
The routine glp_interior is a driver to the LP solver based on the primal-dual interior-point
method. This routine retrieves problem data from the specified problem object, calls the solver to
solve the problem instance, and stores results of computations back into the problem object.
The interior-point solver has a set of control parameters. Values of the control parameters
can be passed in the structure glp_iptcp, which the parameter parm points to. For detailed
description of this structure see paragraph “Control parameters” below. Before specifying some
control parameters the application program should initialize the structure glp_iptcp by default
values of all control parameters using the routine glp_init_iptcp (see the next subsection). This
is needed for backward compatibility, because in the future there may appear new members in the
structure glp_iptcp.
Returns
GLP_EFAIL The problem has no rows/columns.
GLP_ENOCVG Very slow convergence or divergence.
GLP_EITLIM Iteration limit exceeded.
GLP_EINSTAB Numerical instability on solving Newtonian system.
51
Comments
The routine glp_interior implements an easy version of the primal-dual interior-point method
based on Mehrotra’s technique.5
Note that currently the GLPK interior-point solver does not include many important features,
in particular:
— it is not able to process dense columns. Thus, if the constraint matrix of the LP problem
has dense columns, the solving process may be inefficient;
— it has no features against numerical instability. For some LP problems premature termination
may happen if the matrix ADAT becomes singular or ill-conditioned;
— it is not able to identify the optimal basis, which corresponds to the interior-point solution
found.
Terminal output
Solving large LP problems may take a long time, so the solver reports some information about
every interior-point iteration,6 which is sent to the terminal. This information has the following
format:
nnn: obj = fff; rpi = ppp; rdi = ddd; gap = ggg
where: nnn is iteration number, fff is the current value of the objective function (in the case of
maximization it has wrong sign), ppp is the current relative primal infeasibility (cf. (2.10)):
∥Ax(k) − b∥
, (2.14)
1 + ∥b∥
ddd is the current relative dual infeasibility (cf. (2.11)):
∥AT π (k) + λ(k) − c∥

, (2.15)
1 + ∥c∥
ggg is the current primal-dual gap (cf. (2.12)):
|cT x(k) − bT π (k) |

, (2.16)
1 + |cT x(k) |
and [x(k) , π (k) , λ(k) ] is the current point on k-th iteration, k = 0, 1, 2, . . . . Note that all solution
components are internally scaled, so information sent to the terminal is suitable only for visual
inspection.
5
S. Mehrotra. On the implementation of a primal-dual interior point method. SIAM J. on Optim., 2(4), pp.
575-601, 1992.
6
Unlike the simplex method the interior point method usually needs 30—50 iterations (independently on the
problem size) in order to find an optimal solution.
52
Control parameters
This paragraph describes all control parameters currently used in the interior-point solver. Sym-
bolic names of control parameters are names of corresponding members in the structure glp_iptcp.

GLP_MSG_OFF—no output;
GLP_MSG_ERR—error and warning messages only;
GLP_MSG_ON —normal output;
GLP_MSG_ALL—full output (including informational messages).
int ord alg (default: GLP ORD AMD)

Ordering algorithm used prior to Cholesky factorization:
GLP_ORD_NONE —use natural (original) ordering;
GLP_ORD_QMD —quotient minimum degree (QMD);
GLP_ORD_AMD —approximate minimum degree (AMD);
GLP_ORD_SYMAMD—approximate minimum degree (SYMAMD).
Example
The following main program reads LP problem instance in fixed MPS format from file
25fv47.mps,7 solves it with the interior-point solver, and writes the solution to file 25fv47.txt.
/* iptsamp.c */
#include <stdio.h>
#include <stdlib.h>
#include <glpk.h>
int main(void)
{ glp_prob *P;
glp_interior(P, NULL);
glp_print_ipt(P, "25fv47.txt");
glp_delete_prob(P);
return 0;
}
/* eof */
7
This instance in fixed MPS format can be found in the Netlib LP collection; see ftp://ftp.netlib.org/lp/data/.
53
Reading problem data from ‘25fv47.mps’...
Problem: 25FV47
Objective: R0000
Original LP has 822 row(s), 1571 column(s), and 11127 non-zero(s)
Working LP has 821 row(s), 1876 column(s), and 10705 non-zero(s)
Matrix A has 10705 non-zeros
Matrix S = A*A’ has 11895 non-zeros (upper triangle)
Minimal degree ordering...
Computing Cholesky factorization S = L’*L...
Matrix L has 35411 non-zeros
Guessing initial point...
Optimization begins...
0: obj = 1.823377629e+05; rpi = 1.3e+01; rdi = 1.4e+01; gap = 9.3e-01
1: obj = 9.260045192e+04; rpi = 5.3e+00; rdi = 5.6e+00; gap = 6.8e+00
2: obj = 3.596999742e+04; rpi = 1.5e+00; rdi = 1.2e+00; gap = 1.8e+01
3: obj = 1.989627568e+04; rpi = 4.7e-01; rdi = 3.0e-01; gap = 1.9e+01
9: obj = 7.122676293e+03; rpi = 1.2e-03; rdi = 1.5e-04; gap = 6.6e-01
OPTIMAL SOLUTION FOUND
Writing interior-point solution to ‘25fv47.txt’...
54
2.9.2 glp init iptcp — initialize interior-point solver control parameters
Synopsis
int glp_init_iptcp(glp_iptcp *parm);
Description
The routine glp_init_iptcp initializes control parameters, which are used by the interior-point
solver, with default values.
Default values of the control parameters are stored in the structure glp_iptcp, which the
parameter parm points to.
2.9.3 glp ipt status — determine solution status
Synopsis
int glp_ipt_status(glp_prob *P);
Returns
The routine glp_ipt_status reports the status of a solution found by the interior-point solver
as follows:
GLP_UNDEF — interior-point solution is undefined;
GLP_OPT — interior-point solution is optimal;
GLP_INFEAS — interior-point solution is infeasible;
GLP_NOFEAS — no feasible primal-dual solution exists.
2.9.4 glp ipt obj val — retrieve objective value
Synopsis
double glp_ipt_obj_val(glp_prob *P);
Returns
The routine glp_ipt_obj_val returns value of the objective function for interior-point solution.
2.9.5 glp ipt row prim — retrieve row primal value
Synopsis
double glp_ipt_row_prim(glp_prob *P, int i);
Returns
The routine glp_ipt_row_prim returns primal value of the auxiliary variable associated with
i-th row.
55
2.9.6 glp ipt row dual — retrieve row dual value
Synopsis
double glp_ipt_row_dual(glp_prob *P, int i);
Returns
The routine glp_ipt_row_dual returns dual value (i.e. reduced cost) of the auxiliary variable
associated with i-th row.
2.9.7 glp ipt col prim — retrieve column primal value
Synopsis
double glp_ipt_col_prim(glp_prob *P, int j);
Returns
The routine glp_ipt_col_prim returns primal value of the structural variable associated with
j-th column.
2.9.8 glp ipt col dual — retrieve column dual value
Synopsis
double glp_ipt_col_dual(glp_prob *P, int j);
Returns
The routine glp_ipt_col_dual returns dual value (i.e. reduced cost) of the structural variable
associated with j-th column.
56
2.10 Mixed integer programming routines
2.10.1 glp set col kind — set (change) column kind
Synopsis
void glp_set_col_kind(glp_prob *P, int j, int kind);
Description
The routine glp_set_col_kind sets (changes) the kind of j-th column (structural variable) as
specified by the parameter kind:
GLP_CV — continuous variable;
GLP_IV — integer variable;
GLP_BV — binary variable.
Setting a column to GLP_BV has the same effect as if it were set to GLP_IV, its lower bound were
set 0, and its upper bound were set to 1.
2.10.2 glp get col kind — retrieve column kind
Synopsis
int glp_get_col_kind(glp_prob *P, int j);
Returns
The routine glp_get_col_kind returns the kind of j-th column (structural variable) as follows:
GLP_CV — continuous variable;
GLP_IV — integer variable;
GLP_BV — binary variable.
2.10.3 glp get num int — retrieve number of integer columns
Synopsis
int glp_get_num_int(glp_prob *P);
Returns
The routine glp_get_num_int returns the number of columns (structural variables), which are
marked as integer. Note that this number does include binary columns.
57
2.10.4 glp get num bin — retrieve number of binary columns
Synopsis
int glp_get_num_bin(glp_prob *P);
Returns
The routine glp_get_num_bin returns the number of columns (structural variables), which are
marked as integer and whose lower bound is zero and upper bound is one.
2.10.5 glp intopt — solve MIP problem with the branch-and-cut method
Synopsis
int glp_intopt(glp_prob *P, const glp_iocp *parm);
Description
The routine glp_intopt is a driver to the MIP solver based on the branch-and-cut method,
which is a hybrid of branch-and-bound and cutting plane methods.
If the presolver is disabled (see paragraph “Control parameters” below), on entry to the routine
glp_intopt the problem object, which the parameter mip points to, should contain optimal solution
to LP relaxation (it can be obtained, for example, with the routine glp_simplex). Otherwise, if
the presolver is enabled, it is not necessary.
The MIP solver has a set of control parameters. Values of the control parameters can be passed
in the structure glp_iocp, which the parameter parm points to. For detailed description of this
structure see paragraph “Control parameters” below. Before specifying some control parameters
the application program should initialize the structure glp_iocp by default values of all control
parameters using the routine glp_init_iocp (see the next subsection). This is needed for backward
compatibility, because in the future there may appear new members in the structure glp_iocp.
Note that the GLPK branch-and-cut solver is not perfect, so it is unable to solve hard or very
large scale MIP instances for a reasonable time.
Returns
0 The MIP problem instance has been successfully solved. (This code does not nec-
GLP_EBOUND Unable to start the search, because some double-bounded variables have incorrect
bounds or some integer variables have non-integer (fractional) bounds.
GLP_EROOT Unable to start the search, because optimal basis for initial LP relaxation is not
provided. (This code may appear only if the presolver is disabled.)
GLP_ENOPFS Unable to start the search, because LP relaxation of the MIP problem instance has
no primal feasible solution. (This code may appear only if the presolver is enabled.)
58
GLP_ENODFS Unable to start the search, because LP relaxation of the MIP problem instance has
no dual feasible solution. In other word, this code means that if the LP relaxation
has at least one primal feasible solution, its optimal solution is unbounded, so if the
MIP problem has at least one integer feasible solution, its (integer) optimal solution
is also unbounded. (This code may appear only if the presolver is enabled.)
GLP_EFAIL The search was prematurely terminated due to the solver failure.
GLP_EMIPGAP The search was prematurely terminated, because the relative mip gap tolerance has
been reached.
GLP_ESTOP The search was prematurely terminated by application. (This code may appear only
if the advanced solver interface is used.)
Built-in MIP presolver
The branch-and-cut solver has built-in MIP presolver. It is a subprogram that transforms the
original MIP problem specified in the problem object to an equivalent MIP problem, which may be
easier for solving with the branch-and-cut method than the original one. For example, the presolver
can remove redundant constraints and variables, whose optimal values are known, perform bound
and coefficient reduction, etc. Once the transformed MIP problem has been solved, the presolver
transforms its solution back to corresponding solution of the original problem.
Presolving is an optional feature of the routine glp_intopt, and by default it is disabled. In
order to enable the MIP presolver, the control parameter presolve should be set to GLP_ON (see
paragraph “Control parameters” below).
Advanced solver interface
The routine glp_intopt allows the user to control the branch-and-cut search by passing to
the solver a user-defined callback routine. For more details see Chapter “Branch-and-Cut API
Routines”.
Terminal output
Solving a MIP problem may take a long time, so the solver reports some information about best
known solutions, which is sent to the terminal. This information has the following format:
+nnn: mip = xxx <rho> yyy gap (ppp; qqq)
where: ‘nnn’ is the simplex iteration number; ‘xxx’ is a value of the objective function for the
best known integer feasible solution (if no integer feasible solution has been found yet, ‘xxx’ is
the text ‘not found yet’); ‘rho’ is the string ‘>=’ (in case of minimization) or ‘<=’ (in case of
maximization); ‘yyy’ is a global bound for exact integer optimum (i.e. the exact integer optimum
is always in the range from ‘xxx’ to ‘yyy’); ‘gap’ is the relative mip gap, in percents, computed as
gap = |xxx − yyy|/(|xxx| + DBL EPSILON) · 100% (if gap is greater than 999.9%, it is not printed);
‘ppp’ is the number of subproblems in the active list, ‘qqq’ is the number of subproblems which
have been already fathomed and therefore removed from the branch-and-bound search tree.
59
Control parameters
This paragraph describes all control parameters currently used in the MIP solver. Symbolic
names of control parameters are names of corresponding members in the structure glp_iocp.

GLP_MSG_OFF — no output;
GLP_MSG_ERR — error and warning messages only;
GLP_MSG_ON — normal output;
GLP_MSG_ALL — full output (including informational messages).
int br tech (default: GLP BR DTH)

Branching technique option:
GLP_BR_FFV — first fractional variable;
GLP_BR_LFV — last fractional variable;
GLP_BR_MFV — most fractional variable;
GLP_BR_DTH — heuristic by Driebeck and Tomlin;
GLP_BR_PCH — hybrid pseudo-cost heuristic.
int bt tech (default: GLP BT BLB)

Backtracking technique option:
GLP_BT_DFS — depth first search;
GLP_BT_BFS — breadth first search;
GLP_BT_BLB — best local bound;
GLP_BT_BPH — best projection heuristic.
int pp tech (default: GLP PP ALL)

Preprocessing technique option:
GLP_PP_NONE — disable preprocessing;
GLP_PP_ROOT — perform preprocessing only on the root level;
GLP_PP_ALL — perform preprocessing on all levels.
int sr heur (default: GLP ON)

Simple rounding heuristic option:
GLP_ON — enable applying the simple rounding heuristic;
GLP_OFF — disable applying the simple rounding heuristic.
60
int fp heur (default: GLP OFF)
Feasibility pump heuristic option:
GLP_ON — enable applying the feasibility pump heuristic;
GLP_OFF — disable applying the feasibility pump heuristic.
int ps heur (default: GLP OFF)

Proximity search heuristic8 option:
GLP_ON — enable applying the proximity search heuristic;
GLP_OFF — disable applying the proximity search pump heuristic.
int ps tm lim (default: 60000)

Time limit, in milliseconds, for the proximity search heuristic (see above).
int gmi cuts (default: GLP OFF)

Gomory’s mixed integer cut option:
GLP_ON — enable generating Gomory’s cuts;
GLP_OFF — disable generating Gomory’s cuts.
int mir cuts (default: GLP OFF)

Mixed integer rounding (MIR) cut option:
GLP_ON — enable generating MIR cuts;
GLP_OFF — disable generating MIR cuts.
int cov cuts (default: GLP OFF)

Mixed cover cut option:
GLP_ON — enable generating mixed cover cuts;
GLP_OFF — disable generating mixed cover cuts.
int clq cuts (default: GLP OFF)

Clique cut option:
GLP_ON — enable generating clique cuts;
GLP_OFF — disable generating clique cuts.
8
The Fischetti–Monaci Proximity Search (a.k.a. Proxy) heuristic. This algorithm is often capable of rapidly
improving a feasible solution of a MIP problem with binary variables. It allows to quickly obtain suboptimal solutions
in some problems which take too long time to be solved to optimality.
61
double tol int (default: 1e-5)
Absolute tolerance used to check if optimal solution to the current LP relaxation is integer
feasible. (Do not change this parameter without detailed understanding its purpose.)
double tol obj (default: 1e-7)

Relative tolerance used to check if the objective value in optimal solution to the current LP
relaxation is not better than in the best known integer feasible solution. (Do not change this
parameter without detailed understanding its purpose.)
double mip gap (default: 0.0)

The relative mip gap tolerance. If the relative mip gap for currently known best integer feasible
solution falls below this tolerance, the solver terminates the search. This allows obtainig suboptimal
integer feasible solutions if solving the problem to optimality takes too long time.
int tm lim (default: INT MAX)

Searching time limit, in milliseconds.
int out frq (default: 5000)

Output frequency, in milliseconds. This parameter specifies how frequently the solver sends
information about the solution process to the terminal.
int out dly (default: 10000)

Output delay, in milliseconds. This parameter specifies how long the solver should delay sending
information about solution of the current LP relaxation with the simplex method to the terminal.
void (*cb func)(glp tree *tree, void *info) (default: NULL)

Entry point to the user-defined callback routine. NULL means the advanced solver interface is
not used. For more details see Chapter “Branch-and-Cut API Routines”.
void *cb info (default: NULL)

Transit pointer passed to the routine cb_func (see above).
int cb size (default: 0)

The number of extra (up to 256) bytes allocated for each node of the branch-and-bound tree to
store application-specific data. On creating a node these bytes are initialized by binary zeros.
int presolve (default: GLP OFF)

MIP presolver option:
GLP_ON — enable using the MIP presolver;
GLP_OFF — disable using the MIP presolver.
62
int binarize (default: GLP OFF)
Binarization option (used only if the presolver is enabled):
GLP_ON — replace general integer variables by binary ones;
GLP_OFF — do not use binarization.
2.10.6 glp init iocp — initialize integer optimizer control parameters
Synopsis
void glp_init_iocp(glp_iocp *parm);
Description
The routine glp_init_iocp initializes control parameters, which are used by the branch-and-cut
solver, with default values.
Default values of the control parameters are stored in a glp_iocp structure, which the parameter
parm points to.
2.10.7 glp mip status — determine status of MIP solution
Synopsis
int glp_mip_status(glp_prob *P);
Returns
The routine glp_mip_status reports the status of a MIP solution found by the MIP solver as
follows:
GLP_UNDEF — MIP solution is undefined;
GLP_OPT — MIP solution is integer optimal;
GLP_FEAS — MIP solution is integer feasible, however, its optimality (or non-optimality) has
not been proven, perhaps due to premature termination of the search;
GLP_NOFEAS — problem has no integer feasible solution (proven by the solver).
2.10.8 glp mip obj val — retrieve objective value
Synopsis
double glp_mip_obj_val(glp_prob *P);
Returns
The routine glp_mip_obj_val returns value of the objective function for MIP solution.
63
2.10.9 glp mip row val — retrieve row value
Synopsis
double glp_mip_row_val(glp_prob *P, int i);
Returns
The routine glp_mip_row_val returns value of the auxiliary variable associated with i-th row
for MIP solution.
2.10.10 glp mip col val — retrieve column value
Synopsis
double glp_mip_col_val(glp_prob *P, int j);
Returns
The routine glp_mip_col_val returns value of the structural variable associated with j-th
column for MIP solution.
64
2.11 Additional routines
2.11.1 glp check kkt — check feasibility/optimality conditions
Synopsis
void glp check kkt(glp prob *P, int sol, int cond, double *ae max, int *ae ind,
double *re max, int *re ind);
Description
The routine glp_check_kkt allows to check feasibility/optimality conditions for the current so-
lution stored in the specified problem object. (For basic and interior-point solutions these conditions
are known as Karush–Kuhn–Tucker optimality conditions.)
The parameter sol specifies which solution should be checked:
GLP_SOL — basic solution;
GLP_IPT — interior-point solution;
GLP_MIP — mixed integer solution.
The parameter cond specifies which condition should be checked:
GLP_KKT_PE — check primal equality constraints (KKT.PE);
GLP_KKT_PB — check primal bound constraints (KKT.PB);
GLP_KKT_DE — check dual equality constraints (KKT.DE). This conditions can be checked only
for basic or interior-point solution;
GLP_KKT_DB — check dual bound constraints (KKT.DB). This conditions can be checked only
for basic or interior-point solution.
Detailed explanations of these conditions are given below in paragraph “Background”.
On exit the routine stores the following information to locations specified by parameters ae_max,
ae_ind, re_max, and re_ind (if some parameter is a null pointer, corresponding information is not
stored):
ae_max — largest absolute error;
ae_ind — number of row (KKT.PE), column (KKT.DE), or variable (KKT.PB, KKT.DB) with
the largest absolute error;
re_max — largest relative error;
re_ind — number of row (KKT.PE), column (KKT.DE), or variable (KKT.PB, KKT.DB) with
the largest relative error.
Row (auxiliary variable) numbers are in the range 1 to m, where m is the number of rows in
the problem object. Column (structural variable) numbers are in the range 1 to n, where n is the
number of columns in the problem object. Variable numbers are in the range 1 to m + n, where
variables with numbers 1 to m correspond to rows, and variables with numbers m + 1 to m + n
correspond to columns. If the error reported is exact zero, corresponding row, column or variable
number is set to zero.
65
Background
The first condition checked by the routine is the following:
xR − AxS = 0, (KKT.PE)
where xR is the subvector of auxiliary variables (rows), xS is the subvector of structural variables
(columns), A is the constraint matrix. This condition expresses the requirement that all primal
variables should satisfy to the system of equality constraints of the original LP problem. In case of
exact arithmetic this condition would be satisfied for any basic solution; however, in case of inexact
(floating-point) arithmetic, this condition shows how accurate the primal solution is, that depends
on accuracy of a representation of the basis matrix used by the simplex method, or on accuracy
provided by the interior-point method.
To check the condition (KKT.PE) the routine computes the vector of residuals:
g = xR − AxS ,
and determines component of this vector that correspond to largest absolute and relative errors:
ae max = max |gi |,

1≤i≤m
|gi |
re max = max .
1≤i≤m 1 + |(xR )i |
The second condition checked by the routine is the following:
lk ≤ xk ≤ uk for all k = 1, . . . , m + n, (KKT.PB)
where xk is auxiliary (1 ≤ k ≤ m) or structural (m + 1 ≤ k ≤ m + n) variable, lk and uk are,

respectively, lower and upper bounds of the variable xk (including cases of infinite bounds). This
condition expresses the requirement that all primal variables shoudl satisfy to bound constraints of
the original LP problem. In case of basic solution all non-basic variables are placed on their active
bounds, so actually the condition (KKT.PB) needs to be checked for basic variables only. If the
primal solution has sufficient accuracy, this condition shows its primal feasibility.
To check the condition (KKT.PB) the routine computes a vector of residuals:


 0, if lk ≤ xk ≤ uk


hk = xk − lk , if xk < lk



 x −u ,
k k if xk > uk
for all k = 1, . . . , m + n, and determines components of this vector that correspond to largest
absolute and relative errors:
ae max = max |hk |,
1≤k≤m+n
|hk |
re max = max .
1≤k≤m+n 1 + |xk |
66
The third condition checked by the routine is:
grad Z = c = (Ã)T π + d,
where Z is the objective function, c is the vector of objective coefficients, (Ã)T is a matrix transposed
to the expanded constraint matrix Ã = (I|−A), π is a vector of Lagrange multipliers that correspond
to equality constraints of the original LP problem, d is a vector of Lagrange multipliers that
correspond to bound constraints for all (auxiliary and structural) variables of the original LP
problem. Geometrically the third condition expresses the requirement that the gradient of the
objective function should belong to the orthogonal complement of a linear subspace defined by the
equality and active bound constraints, i.e. that the gradient is a linear combination of normals
to the constraint hyperplanes, where Lagrange multipliers π and d are coefficients of that linear
combination.
To eliminate the vector π rewrite the third condition as:
     
I d c
  π =  R +  R ,
−AT dS cS
or, equivalently, 
 π + dR = c R ,
 −AT π + d = c .
S S
Then substituting the vector π from the first equation into the second we finally have:
AT (dR − cR ) + (dS − cS ) = 0, (KKT.DE)
where dR is the subvector of reduced costs of auxiliary variables (rows), dS is the subvector of
reduced costs of structural variables (columns), cR and cS are subvectors of objective coefficients
at, respectively, auxiliary and structural variables, AT is a matrix transposed to the constraint
matrix of the original LP problem. In case of exact arithmetic this condition would be satisfied
for any basic solution; however, in case of inexact (floating-point) arithmetic, this condition shows
how accurate the dual solution is, that depends on accuracy of a representation of the basis matrix
used by the simplex method, or on accuracy provided by the interior-point method.
To check the condition (KKT.DE) the routine computes a vector of residuals:
u = AT (dR − cR ) + (dS − cS ),
and determines components of this vector that correspond to largest absolute and relative errors:
ae max = max |uj |,

1≤j≤n
|uj |
re max = max .
1≤j≤n 1 + |(dS )j − (cS )j |
67
The fourth condition checked by the routine is the following:


 if −∞ < xk < +∞, then dk =0





 if lk ≤ xk < +∞, then dk ≥ 0 (minimization)




 dk ≤ 0 (maximization)
(KKT.DB)

 if −∞ < xk ≤ uk , then dk ≤ 0 (minimization)





 ≥ 0 (maximization)

 dk



if lk ≤ xk ≤ uk , then dk is of any sign
for all k = 1, . . . , m+n, where dk is a reduced cost (Lagrange multiplier) of auxiliary (1 ≤ k ≤ m) or

structural (m + 1 ≤ k ≤ m + n) variable xk . Geometrically this condition expresses the requirement
that constraints of the original problem must “hold” the point preventing its movement along the
anti-gradient (in case of minimization) or the gradient (in case of maximization) of the objective
function. In case of basic solution reduced costs of all basic variables are placed on their active
(zero) bounds, so actually the condition (KKT.DB) needs to be checked for non-basic variables
only. If the dual solution has sufficient accuracy, this condition shows the dual feasibility of the
solution.
To check the condition (KKT.DB) the routine computes a vector of residuals:

 0, if dk has correct sign
vk =
 |d |, if d has wrong sign
k k
for all k = 1, . . . , m + n, and determines components of this vector that correspond to largest
absolute and relative errors:
ae max = max |vk |,
1≤k≤m+n
|vk |
re max = max .
1≤k≤m+n 1 + |dk − ck |
Note that the complete set of Karush-Kuhn-Tucker optimality conditions also includes the fifth,
so called complementary slackness condition, which expresses the requirement that at least either
a primal variable xk or its dual counterpart dk should be on its bound for all k = 1, . . . , m + n.
Currently checking this condition is not implemented yet.
68
Chapter 3
Utility API routines
3.1 Problem data reading/writing routines
3.1.1 glp read mps — read problem data in MPS format
Synopsis
int glp_read_mps(glp_prob *P, int fmt, const glp_mpscp *parm,
const char *fname);
Description
The routine glp_read_mps reads problem data in MPS format from a text file. (The MPS
format is described in Appendix B, page 148.)
The parameter fmt specifies the MPS format version as follows:
GLP_MPS_DECK — fixed (ancient) MPS format;
GLP_MPS_FILE — free (modern) MPS format.
The parameter parm is reserved for use in the future and should be specified as NULL.
The character string fname specifies a name of the text file to be read in. (If the file name ends
with suffix ‘.gz’, the file is assumed to be compressed, in which case the routine glp_read_mps
decompresses it “on the fly”.)
Note that before reading data the current content of the problem object is completely erased
with the routine glp_erase_prob.
Returns
If the operation was successful, the routine glp_read_mps returns zero. Otherwise, it prints an
error message and returns non-zero.
69
3.1.2 glp write mps — write problem data in MPS format
Synopsis
int glp_write_mps(glp_prob *P, int fmt, const glp_mpscp *parm,
const char *fname);
Description
The routine glp_write_mps writes problem data in MPS format to a text file. (The MPS
format is described in Appendix B, page 148.)
The parameter fmt specifies the MPS format version as follows:
GLP_MPS_DECK — fixed (ancient) MPS format;
GLP_MPS_FILE — free (modern) MPS format.
The character string fname specifies a name of the text file to be written out. (If the file name
ends with suffix ‘.gz’, the file is assumed to be compressed, in which case the routine glp_write_mps
performs automatic compression on writing it.)
Returns
If the operation was successful, the routine glp_write_mps returns zero. Otherwise, it prints
an error message and returns non-zero.
3.1.3 glp read lp — read problem data in CPLEX LP format
Synopsis
int glp read lp(glp prob *P, const glp cpxcp *parm, const char *fname);
Description
The routine glp_read_lp reads problem data in CPLEX LP format from a text file. (The
CPLEX LP format is described in Appendix C, page 158.)
The character string fname specifies a name of the text file to be read in. (If the file name
ends with suffix ‘.gz’, the file is assumed to be compressed, in which case the routine glp_read_lp
Returns
If the operation was successful, the routine glp_read_lp returns zero. Otherwise, it prints an
70
3.1.4 glp write lp — write problem data in CPLEX LP format
Synopsis
int glp write lp(glp prob *P, const glp cpxcp *parm, const char *fname);
Description
The routine glp_write_lp writes problem data in CPLEX LP format to a text file. (The
CPLEX LP format is described in Appendix C, page 158.)
The character string fname specifies a name of the text file to be written out. (If the file name
ends with suffix ‘.gz’, the file is assumed to be compressed, in which case the routine glp_write_lp
performs automatic compression on writing it.)
Returns
If the operation was successful, the routine glp_write_lp returns zero. Otherwise, it prints an
3.1.5 glp read prob — read problem data in GLPK format
Synopsis
int glp_read_prob(glp_prob *P, int flags, const char *fname);
Description
The routine glp_read_prob reads problem data in the GLPK LP/MIP format from a text file.
(For description of the GLPK LP/MIP format see below.)
The parameter flags is reserved for use in the future and should be specified as zero.
The character string fname specifies a name of the text file to be read in. (If the file name ends
with suffix ‘.gz’, the file is assumed to be compressed, in which case the routine glp_read_prob
Returns
If the operation was successful, the routine glp_read_prob returns zero. Otherwise, it prints
GLPK LP/MIP format
The GLPK LP/MIP format is a DIMACS-like format.1 The file in this format is a plain ASCII
text file containing lines of several types described below. A line is terminated with the end-of-line
character. Fields in each line are separated by at least one blank space. Each line begins with a
one-character designator to identify the line type.
1
The DIMACS formats were developed by the Center for Discrete Mathematics and Theoretical Computer Science
(DIMACS) to facilitate exchange of problem data. For details see: <http://dimacs.rutgers.edu/Challenges/>.
71
The first line of the data file must be the problem line (except optional comment lines, which
may precede the problem line). The last line of the data file must be the end line. Other lines may
follow in arbitrary order, however, duplicate lines are not allowed.
Comment lines. Comment lines give human-readable information about the data file and are
ignored by GLPK routines. Comment lines can appear anywhere in the data file. Each comment
line begins with the lower-case character c.
c This is an example of comment line
Problem line. There must be exactly one problem line in the data file. This line must appear
before any other lines except comment lines and has the following format:
p CLASS DIR ROWS COLS NONZ
The lower-case letter p specifies that this is the problem line.
The CLASS field defines the problem class and can contain either the keyword lp (that means
linear programming problem) or mip (that means mixed integer programming problem).
The DIR field defines the optimization direction (that is, the objective function sense) and can
contain either the keyword min (that means minimization) or max (that means maximization).
The ROWS, COLS, and NONZ fields contain non-negative integer values specifying, respectively,
the number of rows (constraints), columns (variables), and non-zero constraint coefficients in the
problem instance. Note that NONZ value does not account objective coefficients.
Row descriptors. There must be at most one row descriptor line in the data file for each row
(constraint). This line has one of the following formats:
i ROW f
i ROW l RHS
i ROW u RHS
i ROW d RHS1 RHS2
i ROW s RHS
The lower-case letter i specifies that this is the row descriptor line.
The ROW field specifies the row ordinal number, an integer between 1 and m, where m is the
number of rows in the problem instance.
The next lower-case letter specifies the row type as follows:
∑
f — free (unbounded) row: −∞ < aj xj < +∞;
∑
l — inequality constraint of ‘≥’ type: aj xj ≥ b;
∑
u — inequality constraint of ‘≤’ type: aj xj ≤ b;
∑
d — double-sided inequality constraint: b1 ≤ aj xj ≤ b2 ;
∑
s — equality constraint: aj xj = b.
The RHS field contains a floaing-point value specifying the row right-hand side. The RHS1 and
RHS2 fields contain floating-point values specifying, respectively, the lower and upper right-hand
sides for the double-sided row.
If for some row its descriptor line does not appear in the data file, by default that row is assumed
to be an equality constraint with zero right-hand side.
72
Column descriptors. There must be at most one column descriptor line in the data file for
each column (variable). This line has one of the following formats depending on the problem class
specified in the problem line:
LP class MIP class
j COL f j COL KIND f
j COL l BND j COL KIND l BND
j COL u BND j COL KIND u BND
j COL d BND1 BND2 j COL KIND d BND1 BND2
j COL s BND j COL KIND s BND
The lower-case letter j specifies that this is the column descriptor line.
The COL field specifies the column ordinal number, an integer between 1 and n, where n is the
number of columns in the problem instance.
The KIND field is used only for MIP problems and specifies the column kind as follows:
c — continuous column;
i — integer column;
b — binary column (in this case all remaining fields must be omitted).
The next lower-case letter specifies the column type as follows:
f — free (unbounded) column: −∞ < x < +∞;
l — column with lower bound: x ≥ l;
u — column with upper bound: x ≤ u;
d — double-bounded column: l ≤ x ≤ u;
s — fixed column: x = s.
The BND field contains a floating-point value that specifies the column bound. The BND1 and
BND2 fields contain floating-point values specifying, respectively, the lower and upper bounds for
the double-bounded column.
If for some column its descriptor line does not appear in the file, by default that column is
assumed to be non-negative (in case of LP class) or binary (in case of MIP class).
Coefficient descriptors. There must be exactly one coefficient descriptor line in the data file for
each non-zero objective or constraint coefficient. This line has the following format:
a ROW COL VAL
The lower-case letter a specifies that this is the coefficient descriptor line.
For objective coefficients the ROW field must contain 0. For constraint coefficients the ROW field
specifies the row ordinal number, an integer between 1 and m, where m is the number of rows in
the problem instance.
If both the ROW and COL fields contain 0, the line specifies the constant term (“shift”) of the
objective function rather than objective coefficient.
The VAL field contains a floating-point coefficient value (it is allowed to specify zero value in
this field).
73
The number of constraint coefficient descriptor lines must be exactly the same as specified in
the field NONZ of the problem line.
Symbolic name descriptors. There must be at most one symbolic name descriptor line for the
problem instance, objective function, each row (constraint), and each column (variable). This line
has one of the following formats:
n p NAME
n z NAME
n i ROW NAME
n j COL NAME
The lower-case letter n specifies that this is the symbolic name descriptor line.
The next lower-case letter specifies which object should be assigned a symbolic name:
p — problem instance;
z — objective function;
i — row (constraint);
j — column (variable).
The NAME field contains the symbolic name, a sequence from 1 to 255 arbitrary graphic ASCII
characters, assigned to corresponding object.
End line. There must be exactly one end line in the data file. This line must appear last in the
file and has the following format:
e
The lower-case letter e specifies that this is the end line. Anything that follows the end line is
ignored by GLPK routines.
74
Example of data file in GLPK LP/MIP format
The following example of a data file in GLPK LP/MIP format specifies the same LP problem
as in Subsection “Example of MPS file”.
p lp min 8 7 48 a 1 7 0.38
n p PLAN a 2 1 1
n z VALUE a 2 2 1
i 1 f a 2 3 1
n i 1 VALUE a 2 4 1
i 2 s 2000 a 2 5 1
n i 2 YIELD a 2 6 1
i 3 u 60 a 2 7 1
n i 3 FE a 3 1 0.15
i 4 u 100 a 3 2 0.04
n i 4 CU a 3 3 0.02
i 5 u 40 a 3 4 0.04
n i 5 MN a 3 5 0.02
i 6 u 30 a 3 6 0.01
n i 6 MG a 3 7 0.03
i 7 l 1500 a 4 1 0.03
n i 7 AL a 4 2 0.05
i 8 d 250 300 a 4 3 0.08
n i 8 SI a 4 4 0.02
j 1 d 0 200 a 4 5 0.06
n j 1 BIN1 a 4 6 0.01
j 2 d 0 2500 a 5 1 0.02
n j 2 BIN2 a 5 2 0.04
j 3 d 400 800 a 5 3 0.01
n j 3 BIN3 a 5 4 0.02
j 4 d 100 700 a 5 5 0.02
n j 4 BIN4 a 6 1 0.02
j 5 d 0 1500 a 6 2 0.03
n j 5 BIN5 a 6 5 0.01
n j 6 ALUM a 7 1 0.7
n j 7 SILICON a 7 2 0.75
a 0 1 0.03 a 7 3 0.8
a 0 2 0.08 a 7 4 0.75
a 0 3 0.17 a 7 5 0.8
a 0 4 0.12 a 7 6 0.97
a 0 5 0.15 a 8 1 0.02
a 0 6 0.21 a 8 2 0.06
a 0 7 0.38 a 8 3 0.08
a 1 1 0.03 a 8 4 0.12
a 1 2 0.08 a 8 5 0.02
a 1 3 0.17 a 8 6 0.01
a 1 4 0.12 a 8 7 0.97
a 1 5 0.15 e o f
a 1 6 0.21
75
3.1.6 glp write prob — write problem data in GLPK format
Synopsis
int glp_write_prob(glp_prob *P, int flags, const char *fname);
Description
The routine glp_write_prob writes problem data in the GLPK LP/MIP format to a text
file. (For description of the GLPK LP/MIP format see Subsection “Read problem data in GLPK
format”.)
The parameter flags is reserved for use in the future and should be specified as zero.
The character string fname specifies a name of the text file to be written out. (If the file
name ends with suffix ‘.gz’, the file is assumed to be compressed, in which case the routine
glp_write_prob performs automatic compression on writing it.)
Returns
If the operation was successful, the routine glp_read_prob returns zero. Otherwise, it prints
76
3.2 Routines for processing MathProg models
3.2.1 Introduction
GLPK supports the GNU MathProg modeling language.2 As a rule, models written in MathProg
are solved with the GLPK LP/MIP stand-alone solver glpsol (see Appendix D) and do not need
any programming with API routines. However, for various reasons the user may need to process
MathProg models directly in his/her application program, in which case he/she may use API
routines described in this section. These routines provide an interface to the MathProg translator,
a component of GLPK, which translates MathProg models into an internal code and then interprets
(executes) this code.
The processing of a model written in GNU MathProg includes several steps, which should be
performed in the following order:
1. Allocating the workspace. The translator allocates the workspace, an internal data structure
used on all subsequent steps.
2. Reading model section. The translator reads model section and, optionally, data section
from a specified text file and translates them into the internal code. If necessary, on this step data
section may be ignored.
3. Reading data section(s). The translator reads one or more data sections from specified text
file(s) and translates them into the internal code.
4. Generating the model. The translator executes the internal code to evaluate the content of
the model objects such as sets, parameters, variables, constraints, and objectives. On this step the
execution is suspended at the solve statement.
5. Building the problem object. The translator obtains all necessary information from the
workspace and builds the standard problem object (that is, the program object of type glp_prob).
6. Solving the problem. On this step the problem object built on the previous step is passed to
a solver, which solves the problem instance and stores its solution back to the problem object.
7. Postsolving the model. The translator copies the solution from the problem object to the
workspace and then executes the internal code from the solve statement to the end of the model.
(If model has no solve statement, the translator does nothing on this step.)
8. Freeing the workspace. The translator frees all the memory allocated to the workspace.
Note that the MathProg translator performs no error correction, so if any of steps 2 to 7
fails (due to errors in the model), the application program should terminate processing and go to
step 8.
2
The GNU MathProg modeling language is a subset of the AMPL language. For its detailed description see the
document “Modeling Language GNU MathProg: Language Reference” included in the GLPK distribution.
77
Example 1
In this example the program reads model and data sections from input file egypt.mod3 and
writes the model to output file egypt.mps in free MPS format (see Appendix B). No solution is
performed.
/* mplsamp1.c */
#include <stdio.h>
#include <stdlib.h>
#include <glpk.h>
int main(void)
{ glp_prob *lp;
glp_tran *tran;
int ret;
lp = glp_create_prob();
tran = glp_mpl_alloc_wksp();
ret = glp_mpl_read_model(tran, "egypt.mod", 0);
if (ret != 0)
{ fprintf(stderr, "Error on translating model\n");
goto skip;
}
ret = glp_mpl_generate(tran, NULL);
if (ret != 0)
{ fprintf(stderr, "Error on generating model\n");
goto skip;
}
glp_mpl_build_prob(tran, lp);
ret = glp_write_mps(lp, GLP_MPS_FILE, NULL, "egypt.mps");
if (ret != 0)
fprintf(stderr, "Error on writing MPS file\n");
skip: glp_mpl_free_wksp(tran);
glp_delete_prob(lp);
return 0;
}
/* eof */
3
This is an example model included in the GLPK distribution.
78
Example 2
In this example the program reads model section from file sudoku.mod4 ignoring data section
in this file, reads alternative data section from file sudoku.dat, solves the problem instance and
passes the solution found back to the model.
/* mplsamp2.c */
#include <stdio.h>
#include <stdlib.h>
#include <glpk.h>
int main(void)
{ glp_prob *mip;
glp_tran *tran;
int ret;
mip = glp_create_prob();
tran = glp_mpl_alloc_wksp();
ret = glp_mpl_read_model(tran, "sudoku.mod", 1);
if (ret != 0)
{ fprintf(stderr, "Error on translating model\n");
goto skip;
}
ret = glp_mpl_read_data(tran, "sudoku.dat");
if (ret != 0)
{ fprintf(stderr, "Error on translating data\n");
goto skip;
}
ret = glp_mpl_generate(tran, NULL);
if (ret != 0)
{ fprintf(stderr, "Error on generating model\n");
goto skip;
}
glp_mpl_build_prob(tran, mip);
glp_simplex(mip, NULL);
glp_intopt(mip, NULL);
ret = glp_mpl_postsolve(tran, mip, GLP_MIP);
if (ret != 0)
fprintf(stderr, "Error on postsolving model\n");
skip: glp_mpl_free_wksp(tran);
glp_delete_prob(mip);
return 0;
}
/* eof */
4
This is an example model which is included in the GLPK distribution along with alternative data file sudoku.dat.
79
3.2.2 glp mpl alloc wksp — allocate the translator workspace
Synopsis
glp_tran *glp_mpl_alloc_wksp(void);
Description
The routine glp_mpl_alloc_wksp allocates the MathProg translator workspace. (Note that
multiple instances of the workspace may be allocated, if necessary.)
Returns
The routine returns a pointer to the workspace, which should be used in all subsequent opera-
tions.
3.2.3 glp mpl init rand — initialize pseudo-random number generator
Synopsis
void glp_mpl_init_rand(glp_tran *tran, int seed);
Description
The routine glp_mpl_init_rand initializes a pseudo-random number generator used by the
MathProg translator, where the parameter seed may be any integer number.
A call to the routine glp_mpl_init_rand should immediately follow the call to the routine
glp_mpl_alloc_wksp. However, using of this routine is optional. If it is not called, the effect is
the same as if it were called with seed equal to 1.
3.2.4 glp mpl read model — read and translate model section
Synopsis
int glp_mpl_read_model(glp_tran *tran, const char *fname, int skip);
Description
The routine glp_mpl_read_model reads model section and, optionally, data section, which may
follow the model section, from a text file, whose name is the character string fname, performs
translation of model statements and data blocks, and stores all the information in the workspace.
The parameter skip is a flag. If the input file contains the data section and this flag is non-zero,
the data section is not read as if there were no data section and a warning message is printed. This
allows reading data section(s) from other file(s).
Returns
If the operation is successful, the routine returns zero. Otherwise the routine prints an error
message and returns non-zero.
80
3.2.5 glp mpl read data — read and translate data section
Synopsis
int glp_mpl_read_data(glp_tran *tran, const char *fname);
Description
The routine glp_mpl_read_data reads data section from a text file, whose name is the char-
acter string fname, performs translation of data blocks, and stores the data read in the translator
workspace. If necessary, this routine may be called more than once.
Returns
3.2.6 glp mpl generate — generate the model
Synopsis
int glp_mpl_generate(glp_tran *tran, const char *fname);
Description
The routine glp_mpl_generate generates the model using its description stored in the translator
workspace. This operation means generating all variables, constraints, and objectives, executing
check and display statements, which precede the solve statement (if it is presented).
The character string fname specifies the name of an output text file, to which output produced
by display statements should be written. If fname is NULL, the output is sent to the terminal.
Returns
3.2.7 glp mpl build prob — build problem instance from the model
Synopsis
void glp_mpl_build_prob(glp_tran *tran, glp_prob *P);
Description
The routine glp_mpl_build_prob obtains all necessary information from the translator work-
space and stores it in the specified problem object P. Note that before building the current content
of the problem object is erased with the routine glp_erase_prob.
81
3.2.8 glp mpl postsolve — postsolve the model
Synopsis
int glp_mpl_postsolve(glp_tran *tran, glp_prob *P, int sol);
Description
The routine glp_mpl_postsolve copies the solution from the specified problem object prob to
the translator workspace and then executes all the remaining model statements, which follow the
solve statement.
The parameter sol specifies which solution should be copied from the problem object to the
workspace as follows:
GLP_SOL — basic solution;
GLP_IPT — interior-point solution;
GLP_MIP — mixed integer solution.
Returns
3.2.9 glp mpl free wksp — free the translator workspace
Synopsis
void glp_mpl_free_wksp(glp_tran *tran);
Description
The routine glp_mpl_free_wksp frees all the memory allocated to the translator workspace. It
also frees all other resources, which are still used by the translator.
82
3.3 Problem solution reading/writing routines
3.3.1 glp print sol — write basic solution in printable format
Synopsis
int glp_print_sol(glp_prob *P, const char *fname);
Description
The routine glp_print_sol writes the current basic solution to an LP problem, which is
specified by the pointer P, to a text file, whose name is the character string fname, in printable
format.
Information reported by the routine glp_print_sol is intended mainly for visual analysis.
Returns
If no errors occurred, the routine returns zero. Otherwise the routine prints an error message
and returns non-zero.
3.3.2 glp read sol — read basic solution in GLPK format
Synopsis
int glp_read_sol(glp_prob *P, const char *fname);
Description
The routine glp_read_sol reads basic solution from a text file in the GLPK format. (For
description of the format see below.)
The character string fname specifies the name of the text file to be read in. (If the file name ends
with suffix ‘.gz’, the file is assumed to be compressed, in which case the routine glp_read_sol
decompresses it ”on the fly”.)
Returns
If the operation was successful, the routine glp_read_sol returns zero. Otherwise, it prints an
GLPK basic solution format
The GLPK basic solution format is a DIMACS-like format.5 The file in this format is a plain
ASCII text file containing lines of several types described below. A line is terminated with the
end-of-line character. Fields in each line are separated by at least one blank space. Each line
begins with a one-character designator to identify the line type.
The first line of the solution file must be the solution line (except optional comment lines, which
5
83
Comment lines. Comment lines give human-readable information about the solution file and are
Solution line. There must be exactly one solution line in the solution file. This line must appear
s bas ROWS COLS PST DST OBJ
The lower-case letter s specifies that this is the solution line.
The three-character solution designator bas identifies the file as containing a basic solution to
the LP problem instance.
The ROWS and COLS fields contain non-negative integer values that specify the number of rows
(constraints) and columns (variables), resp., in the LP problem instance.
The PST and DST fields contain lower-case letters that specify the primal and dual solution
status, resp., as follows:
u — solution is undefined;
f — solution is feasible;
i — solution is infeasible;
n — no feasible solution exists.
The OBJ field contains a floating-point number that specifies the objective function value in the
basic solution.
Row solution descriptors. There must be exactly one row solution descriptor line in the solution
file for each row (constraint). This line has the following format:
i ROW ST PRIM DUAL
The lower-case letter i specifies that this is the row solution descriptor line.
The ST field contains one of the following lower-case letters that specifies the row status in the
basic solution:6
b — inactive constraint;
l — inequality constraint active on its lower bound;
u — inequality constraint active on its upper bound;
f — active free (unounded) row;
s — active equality constraint.
The PRIM field contains a floating-point number that specifies the row primal value (the value
of the corresponding linear form).
The DUAL field contains a floating-point number that specifies the row dual value (the Lagrangian
multiplier for active bound).
6
The row status is the status of the associated auxiliary variable.
84
Column solution descriptors. There must be exactly one column solution descriptor line in the
solution file for each column (variable). This line has the following format:
j COL ST PRIM DUAL
The lower-case letter j specifies that this is the column solution descriptor line.
The ST field contains one of the following lower-case letters that specifies the column status in
the basic solution:
b — basic variable;
l — non-basic variable having its lower bound active;
u — non-basic variable having its upper bound active;
f — non-basic free (unounded) variable;
s — non-basic fixed variable.
The PRIM field contains a floating-point number that specifies the column primal value.
The DUAL field contains a floating-point number that specifies the column dual value (the La-
grangian multiplier for active bound).
End line. There must be exactly one end line in the solution file. This line must appear last in
the file and has the following format:
e
Example of solution file in GLPK basic solution format
The following example of a solution file in GLPK basic solution format specifies the optimal
basic solution to the LP problem instance from Subsection “Example of MPS file”.
s bas 7 7 f f 296.216606498195 j 1 l 0 0.253624548736462

i 1 s 2000 -0.0135956678700369 j 2 b 665.342960288809 0
i 2 u 60 -2.56823104693141 j 3 b 490.252707581226 0
i 3 b 83.9675090252707 0 j 4 b 424.187725631769 0
i 4 u 40 -0.544404332129962 j 5 l 0 0.0145559566787004
i 5 b 19.9602888086643 0 j 6 b 299.638989169676 0
i 6 l 1500 0.251985559566788 j 7 b 120.57761732852 0
i 7 l 250 0.48519855595668 e o f
85
3.3.3 glp write sol — write basic solution in GLPK format
Synopsis
int glp_write_sol(glp_prob *P, const char *fname);
Description
The routine glp_write_sol writes the current basic solution to a text file in the GLPK format.
(For description of the GLPK basic solution format see Subsection “Read basic solution in GLPK
format.”)
The character string fname specifies the name of the text file to be written. (If the file name
ends with suffix ‘.gz’, the routine glp_write_sol compresses it ”on the fly”.)
Returns
If the operation was successful, the routine glp_write_sol returns zero. Otherwise, it prints
3.3.4 glp print ipt — write interior-point solution in printable format
Synopsis
int glp_print_ipt(glp_prob *P, const char *fname);
Description
The routine glp_print_ipt writes the current interior point solution to an LP problem, which
the parameter P points to, to a text file, whose name is the character string fname, in printable
format.
Information reported by the routine glp_print_ipt is intended mainly for visual analysis.
Returns
3.3.5 glp read ipt — read interior-point solution in GLPK format
Synopsis
int glp_read_ipt(glp_prob *P, const char *fname);
Description
The routine glp_read_ipt reads interior-point solution from a text file in the GLPK format.
(For description of the format see below.)
with suffix ‘.gz’, the file is assumed to be compressed, in which case the routine glp_read_ipt
Returns
If the operation was successful, the routine glp_read_ipt returns zero. Otherwise, it prints an
86
GLPK interior-point solution format
The GLPK interior-point solution format is a DIMACS-like format.7 The file in this format is
a plain ASCII text file containing lines of several types described below. A line is terminated with
the end-of-line character. Fields in each line are separated by at least one blank space. Each line
s ipt ROWS COLS SST OBJ
The three-character solution designator ipt identifies the file as containing an interior-point
solution to the LP problem instance.
The SST field contains one of the following lower-case letters that specifies the solution status:
o — solution is optimal;
i — solution is infeasible;
n — no feasible solution exists;
u — solution is undefined.
interior-point solution.
i ROW PRIM DUAL
The PRIM field contains a floating-point number that specifies the row primal value (the value
of the corresponding linear form).
The DUAL field contains a floating-point number that specifies the row dual value (the Lagrangian
multiplier for active bound).
7
87
j COL PRIM DUAL
The PRIM field contains a floating-point number that specifies the column primal value.
The DUAL field contains a floating-point number that specifies the column dual value (the La-
grangian multiplier for active bound).
e
Example of solution file in GLPK interior-point solution format
The following example of a solution file in GLPK interior-point solution format specifies the
optimal interior-point solution to the LP problem instance from Subsection “Example of MPS file”.
s ipt 7 7 o 296.216606851403 j 1 3.3482079213784e-07 0.253624547432525

i 1 2000.00000290369 -0.0135956757623443 j 2 665.342955760768 6.04613825351601e-11
i 2 60.0000001302903 -2.568231024875 j 3 490.25271366802 5.8488360240978e-10
i 3 83.9675094251819 -8.85591445202383e-10 j 4 424.187729774275 -2.54144550490434e-11
i 4 39.9999999985064 -0.544404310443766 j 5 1.46067738492801e-06 0.0145559574770786
i 5 19.9602886941262 -2.24817803513554e-09 j 6 299.638985053112 1.49359074902927e-10
i 6 1500.00000199013 0.251985567257828 j 7 120.577616852015 3.50297708781545e-10
i 7 250.000000244896 0.48519856304979 e o f
3.3.6 glp write ipt — write interior-point solution in GLPK format
Synopsis
int glp_write_ipt(glp_prob *P, const char *fname);
Description
The routine glp_write_ipt writes the current interior-point solution to a text file in the GLPK
format. (For description of the GLPK interior-point solution format see Subsection “Read interior-
point solution in GLPK format.”)
ends with suffix ‘.gz’, the routine glp_write_ipt compresses it ”on the fly”.)
Returns
If the operation was successful, the routine glp_write_ipt returns zero. Otherwise, it prints
88
3.3.7 glp print mip — write MIP solution in printable format
Synopsis
int glp_print_mip(glp_prob *P, const char *fname);
Description
The routine glp_print_mip writes the current solution to a MIP problem, which is specified
by the pointer P, to a text file, whose name is the character string fname, in printable format.
Information reported by the routine glp_print_mip is intended mainly for visual analysis.
Returns
3.3.8 glp read mip — read MIP solution in GLPK format
Synopsis
int glp_read_mip(glp_prob *P, const char *fname);
Description
The routine glp_read_mip reads MIP solution from a text file in the GLPK format. (For
description of the format see below.)
with suffix ‘.gz’, the file is assumed to be compressed, in which case the routine glp_read_mip
Returns
If the operation was successful, the routine glp_read_mip returns zero. Otherwise, it prints an
GLPK MIP solution format
The GLPK MIP solution format is a DIMACS-like format.8 The file in this format is a plain
ASCII text file containing lines of several types described below. A line is terminated with the
end-of-line character. Fields in each line are separated by at least one blank space. Each line
8
89
s mip ROWS COLS SST OBJ
The three-character solution designator mip identifies the file as containing a solution to the
MIP problem instance.
The SST field contains one of the following lower-case letters that specifies the solution status:
o — solution is integer optimal;
f — solution is integer feasible (non-optimal);
n — no integer feasible solution exists;
u — solution is undefined.
MIP solution.
i ROW VAL
The VAL field contains a floating-point number that specifies the row value (the value of the
corresponding linear form).
j COL VAL
The VAL field contains a floating-point number that specifies the column value.
e
Example of solution file in GLPK MIP solution format
The following example of a solution file in GLPK MIP solution format specifies an optimal
solution to a MIP problem instance.
90
s mip 8 8 o 1201500 j 1 60
i 1 60 j 2 6
i 2 8400 j 3 0
i 3 -1200 j 4 60
i 4 0 j 5 6
i 5 9000 j 6 600
i 6 -600 j 7 60
i 7 0 j 8 16
i 8 8000 e o f
3.3.9 glp write mip — write MIP solution in GLPK format
Synopsis
int glp_write_mip(glp_prob *P, const char *fname);
Description
The routine glp_write_mip writes the current MIP solution to a text file in the GLPK format.
(For description of the GLPK MIP solution format see Subsection “Read MIP solution in GLPK
format.”)
ends with suffix ‘.gz’, the routine glp_write_mip compresses it ”on the fly”.)
Returns
If the operation was successful, the routine glp_write_mip returns zero. Otherwise, it prints
91
3.4 Post-optimal analysis routines
3.4.1 glp print ranges — print sensitivity analysis report
Synopsis
int glp print ranges(glp prob *P, int len, const int list[], int flags,
const char *fname);
Description
The routine glp_print_ranges performs sensitivity analysis of current optimal basic solution
and writes the analysis report in human-readable format to a text file, whose name is the character
string fname. (Detailed description of the report structure is given below.)
The parameter len specifies the length of the row/column list.
The array list specifies ordinal number of rows and columns to be analyzed. The ordinal numbers
should be passed in locations list[1], list[2], . . . , list[len]. Ordinal numbers from 1 to m refer to
rows, and ordinal numbers from m + 1 to m + n refer to columns, where m and n are, resp., the
total number of rows and columns in the problem object. Rows and columns appear in the analysis
report in the same order as they follow in the array list.
It is allowed to specify len = 0, in which case the array list is not used (so it can be specified as
NULL), and the routine performs analysis for all rows and columns of the problem object.
The parameter flags is reserved for use in the future and must be specified as zero.
On entry to the routine glp_print_ranges the current basic solution must be optimal and
the basis factorization must exist. The application program can check that with the routine
glp_bf_exists, and if the factorization does not exist, compute it with the routine glp_factorize.
Note that if the LP preprocessor is not used, on normal exit from the simplex solver routine
glp_simplex the basis factorization always exists.
Returns
If the operation was successful, the routine glp_print_ranges returns zero. Otherwise, it prints
Analysis report example
An example of the sensitivity analysis report is shown on the next two pages. This example
corresponds to the example of LP problem described in Subsection “Example of MPS file”.
Structure of the analysis report
For each row and column specified in the array list the routine prints two lines containing generic
information and analysis information, which depends on the status of corresponding row or column.
Note
∑ that analysis of a row is analysis of its auxiliary variable, which is equal to the row linear
form aj xj , and analysis of a column is analysis of corresponding structural variable. Therefore,
formally, on performing the sensitivity analysis there is no difference between rows and columns.
92
GLPK 4.42 - SENSITIVITY ANALYSIS REPORT Page 1
Problem: PLAN
Objective: VALUE = 296.2166065 (MINimum)
No. Row name St Activity Slack Lower bound Activity Obj coef Obj value at Limiting
Marginal Upper bound range range break point variable
------ ------------ -- ------------- ------------- ------------- ------------- ------------- ------------- ------------
1 VALUE BS 296.21661 -296.21661 -Inf 299.25255 -1.00000 . MN
. +Inf 296.21661 +Inf +Inf
2 YIELD NS 2000.00000 . 2000.00000 1995.06864 -Inf 296.28365 BIN3

-.01360 2000.00000 2014.03479 +Inf 296.02579 CU
3 FE NU 60.00000 . -Inf 55.89016 -Inf 306.77162 BIN4

-2.56823 60.00000 62.69978 2.56823 289.28294 BIN3
4 CU BS 83.96751 16.03249 -Inf 93.88467 -.30613 270.51157 MN

. 100.00000 79.98213 .21474 314.24798 BIN5
5 MN NU 40.00000 . -Inf 34.42336 -Inf 299.25255 BIN4
93
-.54440 40.00000 41.68691 .54440 295.29825 BIN3
6 MG BS 19.96029 10.03971 -Inf 24.74427 -1.79618 260.36433 BIN1

. 30.00000 9.40292 .28757 301.95652 MN
7 AL NL 1500.00000 . 1500.00000 1485.78425 -.25199 292.63444 CU

.25199 +Inf 1504.92126 +Inf 297.45669 BIN3
8 SI NL 250.00000 50.00000 250.00000 235.32871 -.48520 289.09812 CU

.48520 300.00000 255.06073 +Inf 298.67206 BIN3
GLPK 4.42 - SENSITIVITY ANALYSIS REPORT Page 2
Problem: PLAN
Objective: VALUE = 296.2166065 (MINimum)
No. Column name St Activity Obj coef Lower bound Activity Obj coef Obj value at Limiting
Marginal Upper bound range range break point variable
------ ------------ -- ------------- ------------- ------------- ------------- ------------- ------------- ------------
1 BIN1 NL . .03000 . -28.82475 -.22362 288.90594 BIN4
.25362 200.00000 33.88040 +Inf 304.80951 BIN4
2 BIN2 BS 665.34296 .08000 . 802.22222 .01722 254.44822 BIN1

. 2500.00000 313.43066 .08863 301.95652 MN
3 BIN3 BS 490.25271 .17000 400.00000 788.61314 .15982 291.22807 MN

. 800.00000 -347.42857 .17948 300.86548 BIN5
4 BIN4 BS 424.18773 .12000 100.00000 710.52632 .10899 291.54745 MN

. 700.00000 -256.15524 .14651 307.46010 BIN1
5 BIN5 NL . .15000 . -201.78739 .13544 293.27940 BIN3
94
.01456 1500.00000 58.79586 +Inf 297.07244 BIN3
6 ALUM BS 299.63899 .21000 . 358.26772 .18885 289.87879 AL

. +Inf 112.40876 .22622 301.07527 MN
7 SILICON BS 120.57762 .38000 . 124.27093 .14828 268.27586 BIN5

. +Inf 85.54745 .46667 306.66667 MN
End of report
Generic information
No. is the row or column ordinal number in the problem object. Rows are numbered from 1
to m, and columns are numbered from 1 to n, where m and n are, resp., the total number of rows
and columns in the problem object.
Row name is the symbolic name assigned to the row. If the row has no name assigned, this field
contains blanks.
Column name is the symbolic name assigned to the column. If the column has no name assigned,
this field contains blanks.
St is the status of the row or column in the optimal solution:
BS — non-active constraint (row), basic column;
NL — inequality constraint having its lower right-hand side active (row), non-basic column
having its lower bound active;
NU — inequality constraint having its upper right-hand side active (row), non-basic column
having its upper bound active;
NS — active equality constraint (row), non-basic fixed column.
NF — active free row, non-basic free (unbounded) column. (This case means that the optimal
solution is dual degenerate.)
Activity is the (primal) value of the auxiliary variable (row) or structural variable (column)
in the optimal solution.
Slack is the (primal) value of the row slack variable.
Obj coef is the objective coefficient of the column (structural variable).
Marginal is the reduced cost (dual activity) of the auxiliary variable (row) or structural variable
(column).
Lower bound is the lower right-hand side (row) or lower bound (column). If the row or column
has no lower bound, this field contains -Inf.
Upper bound is the upper right-hand side (row) or upper bound (column). If the row or column
has no upper bound, this field contains +Inf.
Sensitivity analysis of active bounds
The sensitivity analysis of active bounds is performed only for rows, which are active constraints,
and only for non-basic columns, because inactive constraints and basic columns have no active
bounds.
For every auxiliary (row) or structural (column) non-basic variable the routine starts changing
its active bound in both direction. The first of the two lines in the report corresponds to decreasing,
and the second line corresponds to increasing of the active bound. Since the variable being analyzed
is non-basic, its activity, which is equal to its active bound, also starts changing. This changing
leads to changing of basic (auxiliary and structural) variables, which depend on the non-basic
variable. The current basis remains primal feasible and therefore optimal while values of all basic
variables are primal feasible, i.e. are within their bounds. Therefore, if some basic variable called
the limiting variable reaches its (lower or upper) bound first, before any other basic variables, it
thereby limits further changing of the non-basic variable, because otherwise the current basis would
become primal infeasible. The point, at which this happens, is called the break point. Note that
95
there are two break points: the lower break point, which corresponds to decreasing of the non-basic
variable, and the upper break point, which corresponds to increasing of the non-basic variable.
In the analysis report values of the non-basic variable (i.e. of its active bound) being analyzed
at both lower and upper break points are printed in the field ‘Activity range’. Corresponding
values of the objective function are printed in the field ‘Obj value at break point’, and symbolic
names of corresponding limiting basic variables are printed in the field ‘Limiting variable’. If
the active bound can decrease or/and increase unlimitedly, the field ‘Activity range’ contains
-Inf or/and +Inf, resp.
For example (see the example report above), row SI is a double-sided constraint, which is active
on its lower bound (right-hand side), and its activity in the optimal solution being equal to the
lower bound is 250. The activity range for this row is [235.32871, 255.06073]. This means that the
basis remains optimal while the lower bound is increasing up to 255.06073, and further increasing
is limited by (structural) variable BIN3. If the lower bound reaches this upper break point, the
objective value becomes equal to 298.67206.
Note that if the basis does not change, the objective function depends on the non-basic variable
linearly, and the per-unit change of the objective function is the reduced cost (marginal value) of
the non-basic variable.
Sensitivity analysis of objective coefficients at non-basic variables
The sensitivity analysis of the objective coefficient at a non-basic variable is quite simple, because
in this case change in the objective coefficient leads to equivalent change in the reduced cost
(marginal value).
For every auxiliary (row) or structural (column) non-basic variable the routine starts changing
its objective coefficient in both direction. (Note that auxiliary variables are not included in the
objective function and therefore always have zero objective coefficients.) The first of the two lines in
the report corresponds to decreasing, and the second line corresponds to increasing of the objective
coefficient. This changing leads to changing of the reduced cost of the non-basic variable to be
analyzed and does affect reduced costs of all other non-basic variables. The current basis remains
dual feasible and therefore optimal while the reduced cost keeps its sign. Therefore, if the reduced
cost reaches zero, it limits further changing of the objective coefficient (if only the non-basic variable
is non-fixed).
In the analysis report minimal and maximal values of the objective coefficient, on which the
basis remains optimal, are printed in the field ‘Obj coef range’. If the objective coefficient can
decrease or/and increase unlimitedly, this field contains -Inf or/and +Inf, resp.
For example (see the example report above), column BIN5 is non-basic having its lower bound
active. Its objective coefficient is 0.15, and reduced cost in the optimal solution 0.01456. The
column lower bound remains active while the column reduced cost remains non-negative, thus,
minimal value of the objective coefficient, on which the current basis still remains optimal, is
0.15 − 0.01456 = 0.13644, that is indicated in the field ‘Obj coef range’.
Sensitivity analysis of objective coefficients at basic variables
To perform sensitivity analysis for every auxiliary (row) or structural (column) variable the
routine starts changing its objective coefficient in both direction. (Note that auxiliary variables are
not included in the objective function and therefore always have zero objective coefficients.) The
first of the two lines in the report corresponds to decreasing, and the second line corresponds to
increasing of the objective coefficient. This changing leads to changing of reduced costs of non-basic
96
variables. The current basis remains dual feasible and therefore optimal while reduced costs of all
non-basic variables (except fixed variables) keep their signs. Therefore, if the reduced cost of some
non-basic non-fixed variable called the limiting variable reaches zero first, before reduced cost of any
other non-basic non-fixed variable, it thereby limits further changing of the objective coefficient,
because otherwise the current basis would become dual infeasible (non-optimal). The point, at
which this happens, is called the break point. Note that there are two break points: the lower break
point, which corresponds to decreasing of the objective coefficient, and the upper break point,
which corresponds to increasing of the objective coefficient. Let the objective coefficient reach its
limit value and continue changing a bit further in the same direction that makes the current basis
dual infeasible (non-optimal). Then the reduced cost of the non-basic limiting variable becomes “a
bit” dual infeasible that forces the limiting variable to enter the basis replacing there some basic
variable, which leaves the basis to keep its primal feasibility. It should be understood that if we
change the current basis in this way exactly at the break point, both the current and adjacent bases
will be optimal with the same objective value, because at the break point the limiting variable has
zero reduced cost. On the other hand, in the adjacent basis the value of the limiting variable
changes, because there it becomes basic, that leads to changing of the value of the basic variable
being analyzed. Note that on determining the adjacent basis the bounds of the analyzed basic
variable are ignored as if it were a free (unbounded) variable, so it cannot leave the current basis.
In the analysis report lower and upper limits of the objective coefficient at the basic variable
being analyzed, when the basis remains optimal, are printed in the field ‘Obj coef range’. Corre-
sponding values of the objective function at both lower and upper break points are printed in the
field ‘Obj value at break point’, symbolic names of corresponding non-basic limiting variables
are printed in the field ‘Limiting variable’, and values of the basic variable, which it would take
on in the adjacent bases (as was explained above) are printed in the field ‘Activity range’. If the
objective coefficient can increase or/and decrease unlimitedly, the field ‘Obj coef range’ contains
-Inf and/or +Inf, resp. It also may happen that no dual feasible adjacent basis exists (i.e. on
entering the basis the limiting variable can increase or decrease unlimitedly), in which case the field
‘Activity range’ contains -Inf and/or +Inf.
For example (see the example report above), structural variable (column) BIN3 is basic, its
optimal value is 490.25271, and its objective coefficient is 0.17. The objective coefficient range for
this column is [0.15982, 0.17948]. This means that the basis remains optimal while the objective
coefficient is decreasing down to 0.15982, and further decreasing is limited by (auxiliary) variable
MN. If we make the objective coefficient a bit less than 0.15982, the limiting variable MN will
enter the basis, and in that adjacent basis the structural variable BIN3 will take on new optimal
value 788.61314. At the lower break point, where the objective coefficient is exactly 0.15982, the
objective function takes on the value 291.22807 in both the current and adjacent bases.
Note that if the basis does not change, the objective function depends on the objective coefficient
at the basic variable linearly, and the per-unit change of the objective function is the value of the
basic variable.
97
Chapter 4
Advanced API Routines
4.1 Background
Using vector and matrix notations the LP problem (1.1)—(1.3) (see Section 1.1, page 9) can be
stated as follows:
z = cT xS + c0 (3.1)
xR = AxS (3.2)
lR ≤ xR ≤ uR
(3.3)
lS ≤ xS ≤ uS
where:
xR = (x1 , . . . , xm ) is the vector of auxiliary variables;
xS = (xm+1 , . . . , xm+n ) is the vector of structural variables;
z is the objective function;
c = (c1 , . . . , cn ) is the vector of objective coefficients;
c0 is the constant term (“shift”) of the objective function;
A = (a11 , . . . , amn ) is the constraint matrix;
lR = (l1 , . . . , lm ) is the vector of lower bounds of auxiliary variables;
uR = (u1 , . . . , um ) is the vector of upper bounds of auxiliary variables;
lS = (lm+1 , . . . , lm+n ) is the vector of lower bounds of structural variables;
uS = (um+1 , . . . , um+n ) is the vector of upper bounds of structural variables.
From the simplex method’s standpoint there is no difference between auxiliary and structural
variables. This allows combining all these variables into one vector that leads to the following
problem statement:
98
z = (0 | c)T x + c0 (3.4)
(I | −A)x = 0 (3.5)
l≤x≤u (3.6)
where:
x = (xR | xS ) is the (m + n)-vector of (all) variables;
(0 | c) is the (m + n)-vector of objective coefficients;1
(I | −A) is the augmented constraint m × (m + n)-matrix;2
l = (lR | lS ) is the (m + n)-vector of lower bounds of (all) variables;
u = (uR | uS ) is the (m + n)-vector of upper bounds of (all) variables.
By definition an LP basic solution geometrically is a point in the space of all variables, which
is the intersection of hyperplanes corresponding to active constraints3 . The space of all variables
has the dimension m + n, therefore, to define some basic solution we have to define m + n active
constraints. Note that m constraints (3.5) being linearly independent equalities are always active,
so remaining n active constraints can be chosen only from bound constraints (3.6).
A variable is called non-basic, if its (lower or upper) bound is active, otherwise it is called basic.
Since, as was said above, exactly n bound constraints must be active, in any basic solution there
are always n non-basic variables and m basic variables. (Note that a free variable also can be
non-basic. Although such variable has no bounds, we can think it as the difference between two
non-negative variables, which both are non-basic in this case.)
Now consider how to determine numeric values of all variables for a given basic solution.
Let Π be an appropriate permutation matrix of the order (m + n). Then we can write:
( ) ( )
xB x
= Π R = Πx, (3.7)
xN xS
where xB is the vector of basic variables, xN is the vector of non-basic variables, x = (xR | xS ) is
the vector of all variables in the original order. In this case the system of linear constraints (3.5)
can be rewritten as follows:
( )
xB
(I | −A)Π Πx = 0 ⇒ (B | N )
T
= 0, (3.8)
xN
where
(B | N ) = (I | −A)ΠT . (3.9)
Matrix B is a square non-singular m × m-matrix, which is composed from columns of the

augmented constraint matrix corresponding to basic variables. It is called the basis matrix or
1
Subvector 0 corresponds to objective coefficients at auxiliary variables.
2
Note that due to auxiliary variables matrix (I | −A) contains the unity submatrix and therefore has full rank.
This means, in particular, that the system (3.5) has no linearly dependent constraints.
3
A constraint is called active if at a given point it is satisfied as equality, otherwise it is called inactive.
99
simply the basis. Matrix N is a rectangular m × n-matrix, which is composed from columns of the
augmented constraint matrix corresponding to non-basic variables.
From (3.8) it follows that:
BxB + N xN = 0, (3.10)
therefore,
xB = −B −1 N xN . (3.11)
Thus, the formula (3.11) shows how to determine numeric values of basic variables xB assuming
that non-basic variables xN are fixed on their active bounds.
The m × n-matrix
Ξ = −B −1 N, (3.12)
which appears in (3.11), is called the simplex tableau.4 It shows how basic variables depend on
non-basic variables:
xB = ΞxN . (3.13)
The system (3.13) is equivalent to the system (3.5) in the sense that they both define the same
set of points in the space of (primal) variables, which satisfy to these systems. If, moreover, values
of all basic variables satisfy to their bound constraints (3.3), the corresponding basic solution is
called (primal) feasible, otherwise (primal) infeasible. It is understood that any (primal) feasible
basic solution satisfy to all constraints (3.2) and (3.3).
The LP theory says that if LP has optimal solution, it has (at least one) basic feasible solution,
which corresponds to the optimum. And the most natural way to determine whether a given basic
solution is optimal or not is to use the Karush—Kuhn—Tucker optimality conditions.
For the problem statement (3.4)—(3.6) the optimality conditions are the following:5
(I | −A)x = 0 (3.14)
(I | −A)T π + λl + λu = ∇z = (0 | c)T (3.15)

l≤x≤u (3.16)
λl ≥ 0, λu ≤ 0 (minimization) (3.17)
λl ≤ 0, λu ≥ 0 (maximization) (3.18)
(λl )k (xk − lk ) = 0, (λu )k (xk − uk ) = 0, k = 1, 2, . . . , m + n (3.19)
where:
π = (π1 , . . . , πm ) is a m-vector of Lagrange multipliers for equality constraints (3.5);
λl = [(λl )1 , . . . , (λl )n ] is a n-vector of Lagrange multipliers for lower bound constraints (3.6);
λu = [(λu )1 , . . . , (λu )n ] is a n-vector of Lagrange multipliers for upper bound constraints (3.6).
Condition (3.14) is the primal (original) system of equality constraints (3.5).
Condition (3.15) is the dual system of equality constraints. It requires the gradient of the
objective function to be a linear combination of normals to the planes defined by constraints of the
original problem.
4
This definition corresponds to the GLPK implementation.
5
These conditions can be appiled to any solution, not only to a basic solution.
100
Condition (3.16) is the primal (original) system of bound constraints (3.6).
Condition (3.17) (or (3.18) in case of maximization) is the dual system of bound constraints.
Condition (3.19) is the complementary slackness condition. It requires, for each original (auxil-
iary or structural) variable xk , that either its (lower or upper) bound must be active, or zero bound
of the corresponding Lagrange multiplier ((λl )k or (λu )k ) must be active.
In GLPK two multipliers (λl )k and (λu )k for each primal variable xk , k = 1, . . . , m + n, are
combined into one multiplier:
λk = (λl )k + (λu )k , (3.20)
which is called a dual variable for xk . This cannot lead to an ambiguity, because both lower and
upper bounds of xk cannot be active at the same time,6 so at least one of (λl )k and (λu )k must be
equal to zero, and because these multipliers have different signs, the combined multiplier, which is
their sum, uniquely defines each of them.
Using dual variables λk the dual system of bound constraints (3.17) and (3.18) can be written
in the form of so called “rule of signs” as follows:
Original bound Minimization Maximization

constraint (λl )k (λu )k (λl )k + (λu )k (λl )k (λu )k (λl )k + (λu )k
−∞ < xk < +∞ =0 =0 λk = 0 =0 =0 λk = 0
xk ≥ lk ≥0 =0 λk ≥ 0 ≤0 =0 λk ≤ 0
xk ≤ uk =0 ≤0 λk ≤ 0 =0 ≥0 λk ≥ 0
lk ≤ xk ≤ uk ≥0 ≤ 0 −∞ < λk < +∞ ≤0 ≥ 0 −∞ < λk < +∞
xk = lk = uk ≥0 ≤ 0 −∞ < λk < +∞ ≤0 ≥ 0 −∞ < λk < +∞
May note that each primal variable xk has its dual counterpart λk and vice versa. This allows
applying the same partition for the vector of dual variables as (3.7):
( )
λB
= Πλ, (3.21)
λN
where λB is a vector of dual variables for basic variables xB , λN is a vector of dual variables for
non-basic variables xN .
By definition, bounds of basic variables are inactive constraints, so in any basic solution λB = 0.
Corresponding values of dual variables λN for non-basic variables xN can be determined in the
following way. From the dual system (3.15) we have:
(I | −A)T π + λ = (0 | c)T , (3.22)
so multiplying both sides of (3.22) by matrix Π gives:
Π(I | −A)T π + Πλ = Π(0 | c)T . (3.23)
From (3.9) it follows that
Π(I | −A)T = [(I | −A)ΠT ]T = (B | N )T . (3.24)
Further, we can apply the partition (3.7) also to the vector of objective coefficients (see (3.4)):
( ) ( )
cB 0
=Π , (3.25)
cN c
6
If xk is a fixed variable, we can think it as double-bounded variable lk ≤ xk ≤ uk , where lk = uk .
101
where cB is a vector of objective coefficients at basic variables, cN is a vector of objective coefficients
at non-basic variables. Now, substituting (3.24), (3.21), and (3.25) into (3.23), leads to:
(B | N )T π + (λB | λN )T = (cB | cN )T , (3.26)
and transposing both sides of (3.26) gives the system:

( T) ( ) ( )
B λB c
π+ = B , (3.27)
NT λN cT
which can be written as follows: {

B T π + λB = cB
(3.28)
N T π + λN = cN
Lagrange multipliers π = (πi ) correspond to equality constraints (3.5) and therefore can have any
sign. This allows resolving the first subsystem of (3.28) as follows:7
π = B −T (cB − λB ) = −B −T λB + B −T cB , (3.29)
and substitution of π from (3.29) into the second subsystem of (3.28) gives:
λN = −N T π + cN = N T B −T λB + (cN − N T B −T cB ). (3.30)
The latter system can be written in the following final form:
λN = −ΞT λB + d, (3.31)
where Ξ is the simplex tableau (see (3.12)), and
d = cN − N T B −T cB = cN + ΞT cB (3.32)
is the vector of so called reduced costs of non-basic variables.

Above it was said that in any basic solution λB = 0, so λN = d as it follows from (3.31).
The system (3.31) is equivalent to the system (3.15) in the sense that they both define the same
set of points in the space of dual variables λ, which satisfy to these systems. If, moreover, values
of all dual variables λN (i.e. reduced costs d) satisfy to their bound constraints (i.e. to the “rule
of signs”; see the table above), the corresponding basic solution is called dual feasible, otherwise
dual infeasible. It is understood that any dual feasible solution satisfy to all constraints (3.15) and
(3.17) (or (3.18) in case of maximization).
It can be easily shown that the complementary slackness condition (3.19) is always satisfied
for any basic solution.8 Therefore, a basic solution9 is optimal if and only if it is primal and dual
feasible, because in this case it satifies to all the optimality conditions (3.14)—(3.19).
The meaning of reduced costs d = (dj ) of non-basic variables can be explained in the following
way. From (3.4), (3.7), and (3.25) it follows that:
z = cTB xB + cTN xN + c0 . (3.33)

7
B −T means (B T )−1 = (B −1 )T .
8
Until double-bounded variables appear.
9
It is assumed that a complete basic solution has the form (x, λ), i.e. it includes primal as well as dual variables.
102
Substituting xB from (3.11) into (3.33) we can eliminate basic variables and express the objective
only through non-basic variables:
z = cTB (−B −1 N xN ) + cTN xN + c0 =

= (cTN − cTB B −1 N )xN + c0 =
(3.34)
= (cN − N T B −T cB )T xN + c0 =
= dT xN + c0 .
From (3.34) it is seen that reduced cost dj shows how the objective function z depends on non-
basic variable (xN )j in the neighborhood of the current basic solution, i.e. while the current basis
remains unchanged.
103
4.2 LP basis routines
4.2.1 glp bf exists — check if the basis factorization exists
Synopsis
int glp_bf_exists(glp_prob *P);
Returns
If the basis factorization for the current basis associated with the specified problem object
exists and therefore is available for computations, the routine glp_bf_exists returns non-zero.
Otherwise the routine returns zero.
Comments
Let the problem object have m rows and n columns. In GLPK the basis matrix B is a square
non-singular matrix of the order m, whose columns correspond to basic (auxiliary and/or structural)
variables. It is defined by the following main equality:10
(B | N ) = (I | −A)ΠT ,
where I is the unity matrix of the order m, whose columns correspond to auxiliary variables; A is
the original constraint m × n-matrix, whose columns correspond to structural variables; (I | −A)
is the augmented constraint m × (m + n)-matrix, whose columns correspond to all (auxiliary and
structural) variables following in the original order; Π is a permutation matrix of the order m + n;
and N is a rectangular m × n-matrix, whose columns correspond to non-basic (auxiliary and/or
structural) variables.
For various reasons it may be necessary to solve linear systems with matrix B. To provide this
possibility the GLPK implementation maintains an invertable form of B (that is, some representa-
tion of B −1 ) called the basis factorization, which is an internal component of the problem object.
Typically, the basis factorization is computed by the simplex solver, which keeps it in the problem
object to be available for other computations.
Should note that any changes in the problem object, which affects the basis matrix (e.g. changing
the status of a row or column, changing a basic column of the constraint matrix, removing an active
constraint, etc.), invalidates the basis factorization. So before calling any API routine, which uses
the basis factorization, the application program must make sure (using the routine glp_bf_exists)
that the factorization exists and therefore available for computations.
4.2.2 glp factorize — compute the basis factorization
Synopsis
int glp_factorize(glp_prob *P);
Description
The routine glp_factorize computes the basis factorization for the current basis associated
with the specified problem object.11
10
For more details see Subsection 4.1, page 98.
11
The current basis is defined by the current statuses of rows (auxiliary variables) and columns (structural variables).
104
The basis factorization is computed from “scratch” even if it exists, so the application program
may use the routine glp_bf_exists, and, if the basis factorization already exists, not to call the
routine glp_factorize to prevent an extra work.
The routine glp_factorize does not compute components of the basic solution (i.e. primal
and dual values).
Returns
0 The basis factorization has been successfully computed.
GLP_EBADB The basis matrix is invalid, because the number of basic (auxiliary and structural)
variables is not the same as the number of rows in the problem object.
GLP_ESING The basis matrix is singular within the working precision.
GLP_ECOND The basis matrix is ill-conditioned, i.e. its condition number is too large.
4.2.3 glp bf updated — check if the basis factorization has been updated
Synopsis
int glp_bf_updated(glp_prob *P);
Returns
If the basis factorization has been just computed from “scratch”, the routine glp_bf_updated
returns zero. Otherwise, if the factorization has been updated at least once, the routine returns
non-zero.
Comments
Updating the basis factorization means recomputing it to reflect changes in the basis matrix.
For example, on every iteration of the simplex method some column of the current basis matrix is
replaced by a new column that gives a new basis matrix corresponding to the adjacent basis. In
this case computing the basis factorization for the adjacent basis from “scratch” (as the routine
glp_factorize does) would be too time-consuming.
On the other hand, since the basis factorization update is a numeric computational procedure,
applying it many times may lead to accumulating round-off errors. Therefore the basis is peri-
odically refactorized (reinverted) from “scratch” (with the routine glp_factorize) that allows
improving its numerical properties.
The routine glp_bf_updated allows determining if the basis factorization has been updated at
least once since it was computed from “scratch”.
4.2.4 glp get bfcp — retrieve basis factorization control parameters
Synopsis
void glp_get_bfcp(glp_prob *P, glp_bfcp *parm);
Description
The routine glp_get_bfcp retrieves control parameters, which are used on computing and
updating the basis factorization associated with the specified problem object.
105
Current values of the control parameters are stored in a glp_bfcp structure, which the parameter
parm points to. For a detailed description of the structure glp_bfcp see comments to the routine
glp_set_bfcp in the next subsection.
Comments
The purpose of the routine glp_get_bfcp is two-fold. First, it allows the application program
obtaining current values of control parameters used by internal GLPK routines, which compute
and update the basis factorization.
The second purpose of this routine is to provide proper values for all fields of the structure
glp_bfcp in the case when the application program needs to change some control parameters.
4.2.5 glp set bfcp — change basis factorization control parameters
Synopsis
void glp_set_bfcp(glp_prob *P, const glp_bfcp *parm);
Description
The routine glp_set_bfcp changes control parameters, which are used by internal GLPK rou-
tines on computing and updating the basis factorization associated with the specified problem
object.
New values of the control parameters should be passed in a structure glp_bfcp, which the
parameter parm points to. For a detailed description of the structure glp_bfcp see paragraph
“Control parameters” below.
The parameter parm can be specified as NULL, in which case all control parameters are reset to
their default values.
Comments
Before changing some control parameters with the routine glp_set_bfcp the application pro-
gram should retrieve current values of all control parameters with the routine glp_get_bfcp. This
is needed for backward compatibility, because in the future there may appear new members in the
structure glp_bfcp.
Note that new values of control parameters come into effect on a next computation of the basis
factorization, not immediately.
Example
glp_prob *lp;
glp_bfcp parm;
. . .
/* retrieve current values of control parameters */
glp_get_bfcp(lp, &parm);
/* change the threshold pivoting tolerance */
parm.piv_tol = 0.05;
/* set new values of control parameters */
glp_set_bfcp(lp, &parm);
. . .
106
Control parameters
This paragraph describes all basis factorization control parameters currently used in the pack-
age. Symbolic names of control parameters are names of corresponding members in the structure
glp_bfcp.
int type (default: GLP BF LUF + GLP BF FT)

Basis factorization type:
GLP_BF_LUF + GLP_BF_FT — LU F , Forrest–Tomlin update;
GLP_BF_LUF + GLP_BF_BG — LU F , Schur complement, Bartels–Golub update;
GLP_BF_LUF + GLP_BF_GR — LU F , Schur complement, Givens rotation update;
GLP_BF_BTF + GLP_BF_BG — BT F , Schur complement, Bartels–Golub update;
GLP_BF_BTF + GLP_BF_GR — BT F , Schur complement, Givens rotation update.
In case of GLP_BF_FT the update is applied to matrix U , while in cases of GLP_BF_BG and
GLP_BF_GR the update is applied to the Schur complement.
double piv tol (default: 0.10)

Threshold pivoting (Markowitz) tolerance, 0 < piv_tol < 1, used on computing LU -factoriza-
tion of the basis matrix. Element uij of the active submatrix of factor U fits to be pivot if it satisfies
to the stability criterion |uij | >= piv tol · max |ui∗ |, i.e. if it is not very small in the magnitude
among other elements in the same row. Decreasing this parameter may lead to better sparsity at
the expense of numerical accuracy, and vice versa.
int piv lim (default: 4)

This parameter is used on computing LU -factorization of the basis matrix and specifies how
many pivot candidates needs to be considered on choosing a pivot element, piv_lim ≥ 1. If piv_lim
candidates have been considered, the pivoting routine prematurely terminates the search with the
best candidate found.
int suhl (default: GLP ON)

This parameter is used on computing LU -factorization of the basis matrix. Being set to GLP ON it
enables applying the following heuristic proposed by Uwe Suhl: if a column of the active submatrix
has no eligible pivot candidates, it is no more considered until it becomes a column singleton. In
many cases this allows reducing the time needed for pivot searching. To disable this heuristic the
parameter suhl should be set to GLP OFF.
double eps tol (default: 1e-15)

Epsilon tolerance, eps_tol ≥ 0, used on computing LU -factorization of the basis matrix. If an
element of the active submatrix of factor U is less than eps_tol in the magnitude, it is replaced
by exact zero.
int nfs max (default: 100)

Maximal number of additional row-like factors (entries of the eta file), nfs_max ≥ 1, which can
be added to LU -factorization of the basis matrix on updating it with the Forrest–Tomlin technique.
This parameter is used only once, before LU -factorization is computed for the first time, to allocate
107
working arrays. As a rule, each update adds one new factor (however, some updates may need no
addition), so this parameter limits the number of updates between refactorizations.
double upd tol (default: 1e-6)

Update tolerance, 0 < upd_tol < 1, used on updating LU -factorization of the basis matrix with
the Forrest–Tomlin technique. If after updating the magnitude of some diagonal element ukk of
factor U becomes less than upd tol · max(|uk∗ |, |u∗k |), the factorization is considered as inaccurate.
int nrs max (default: 100)

Maximal number of additional rows and columns, nrs_max ≥ 1, which can be added to LU -
factorization of the basis matrix on updating it with the Schur complement technique. This param-
eter is used only once, before LU -factorization is computed for the first time, to allocate working
arrays. As a rule, each update adds one new row and column (however, some updates may need
no addition), so this parameter limits the number of updates between refactorizations.
4.2.6 glp get bhead — retrieve the basis header information
Synopsis
int glp_get_bhead(glp_prob *P, int k);
Description
The routine glp_get_bhead returns the basis header information for the current basis associated
with the specified problem object.
Returns
If basic variable (xB )k , 1 ≤ k ≤ m, is i-th auxiliary variable (1 ≤ i ≤ m), the routine returns i.
Otherwise, if (xB )k is j-th structural variable (1 ≤ j ≤ n), the routine returns m + j. Here m is
the number of rows and n is the number of columns in the problem object.
Comments
Sometimes the application program may need to know which original (auxiliary and structural)
variable correspond to a given basic variable, or, that is the same, which column of the augmented
constraint matrix (I | −A) correspond to a given column of the basis matrix B.
The correspondence is defined as follows:12
( ) ( ) ( ) ( )
xB xR xR T xB
=Π ⇔ =Π ,
xN xS xS xN
where xB is the vector of basic variables, xN is the vector of non-basic variables, xR is the vector
of auxiliary variables following in their original order,13 xS is the vector of structural variables
following in their original order, Π is a permutation matrix (which is a component of the basis
factorization).
Thus, if (xB )k = (xR )i is i-th auxiliary variable, the routine returns i, and if (xB )k = (xS )j is
j-th structural variable, the routine returns m + j, where m is the number of rows in the problem
object.
12
For more details see Subsection 4.1, page 98.
13
The original order of auxiliary and structural variables is defined by the ordinal numbers of corresponding rows
and columns in the problem object.
108
4.2.7 glp get row bind — retrieve row index in the basis header
Synopsis
int glp_get_row_bind(glp_prob *P, int i);
Returns
The routine glp_get_row_bind returns the index k of basic variable (xB )k , 1 ≤ k ≤ m, which
is i-th auxiliary variable (that is, the auxiliary variable corresponding to i-th row), 1 ≤ i ≤ m,
in the current basis associated with the specified problem object, where m is the number of rows.
However, if i-th auxiliary variable is non-basic, the routine returns zero.
Comments
The routine glp_get_row_bind is an inversion of the routine glp_get_bhead; that is, if
glp_get_bhead(P, k) returns i, glp_get_row_bind(P, i) returns k, and vice versa.
4.2.8 glp get col bind — retrieve column index in the basis header
Synopsis
int glp_get_col_bind(glp_prob *P, int j);
Returns
The routine glp_get_col_bind returns the index k of basic variable (xB )k , 1 ≤ k ≤ m, which is
j-th structural variable (that is, the structural variable corresponding to j-th column), 1 ≤ j ≤ n,
in the current basis associated with the specified problem object, where m is the number of rows,
n is the number of columns. However, if j-th structural variable is non-basic, the routine returns
zero.
Comments
The routine glp_get_col_bind is an inversion of the routine glp_get_bhead; that is, if
glp_get_bhead(P, k) returns m + j, glp_get_col_bind(P, j) returns k, and vice versa.
4.2.9 glp ftran — perform forward transformation
Synopsis
void glp_ftran(glp_prob *P, double x[]);
Description
The routine glp_ftran performs forward transformation (FTRAN), i.e. it solves the system
Bx = b, where B is the basis matrix associated with the specified problem object, x is the vector
of unknowns to be computed, b is the vector of right-hand sides.
On entry to the routine elements of the vector b should be stored in locations x[1], . . . , x[m],
where m is the number of rows. On exit the routine stores elements of the vector x in the same
locations.
109
4.2.10 glp btran — perform backward transformation
Synopsis
void glp_btran(glp_prob *P, double x[]);
Description
The routine glp_btran performs backward transformation (BTRAN), i.e. it solves the system
B T x = b, where B T is a matrix transposed to the basis matrix B associated with the specified
problem object, x is the vector of unknowns to be computed, b is the vector of right-hand sides.
On entry to the routine elements of the vector b should be stored in locations x[1], . . . , x[m],
where m is the number of rows. On exit the routine stores elements of the vector x in the same
locations.
4.2.11 glp warm up — “warm up” LP basis
Synopsis
int glp_warm_up(glp_prob *P);
Description
The routine glp_warm_up “warms up” the LP basis for the specified problem object using
current statuses assigned to rows and columns (that is, to auxiliary and structural variables).
This operation includes computing factorization of the basis matrix (if it does not exist), com-
puting primal and dual components of basic solution, and determining the solution status.
Returns
0 The operation has been successfully performed.
GLP_EBADB The basis matrix is invalid, because the number of basic (auxiliary and structural)
variables is not the same as the number of rows in the problem object.
GLP_ESING The basis matrix is singular within the working precision.
GLP_ECOND The basis matrix is ill-conditioned, i.e. its condition number is too large.
110
4.3 Simplex tableau routines
4.3.1 glp eval tab row — compute row of the tableau
Synopsis
int glp_eval_tab_row(glp_prob *P, int k, int ind[], double val[]);
Description
The routine glp_eval_tab_row computes a row of the current simplex tableau (see Subsection
3.1.1, formula (3.12)), which (row) corresponds to some basic variable specified by the parameter
k as follows: if 1 ≤ k ≤ m, the basic variable is k-th auxiliary variable, and if m + 1 ≤ k ≤ m + n,
the basic variable is (k − m)-th structural variable, where m is the number of rows and n is the
number of columns in the specified problem object. The basis factorization must exist.
The computed row shows how the specified basic variable depends on non-basic variables:
xk = (xB )i = ξi1 (xN )1 + ξi2 (xN )2 + . . . + ξin (xN )n ,
where ξi1 , ξi2 , . . . , ξin are elements of the simplex table row, (xN )1 , (xN )2 , . . . , (xN )n are non-basic
(auxiliary and structural) variables.
The routine stores column indices and corresponding numeric values of non-zero elements of
the computed row in unordered sparse format in locations ind[1], . . . , ind[len] and val[1], . . . ,
val[len], respectively, where 0 ≤ len ≤ n is the number of non-zero elements in the row returned
on exit.
Element indices stored in the array ind have the same sense as index k, i.e. indices 1 to m denote
auxiliary variables while indices m + 1 to m + n denote structural variables (all these variables are
obviously non-basic by definition).
Returns
The routine glp_eval_tab_row returns len, which is the number of non-zero elements in the
simplex table row stored in the arrays ind and val.
Comments
A row of the simplex table is computed as follows. At first, the routine checks that the specified
variable xk is basic and uses the permutation matrix Π (3.7) to determine index i of basic variable
(xB )i , which corresponds to xk .
The row to be computed is i-th row of the matrix Ξ (3.12), therefore:
ξi = eTi Ξ = −eTi B −1 N = −(B −T ei )T N,
where ei is i-th unity vector. So the routine performs BTRAN to obtain i-th row of the inverse
B −1 :
ϱi = B −T ei ,
and then computes elements of the simplex table row as inner products:
ξij = −ϱTi Nj , j = 1, 2, . . . , n,
where Nj is j-th column of matrix N (3.9), which (column) corresponds to non-basic variable (xN )j .
The permutation matrix Π is used again to convert indices j of non-basic columns to original ordinal
numbers of auxiliary and structural variables.
111
4.3.2 glp eval tab col — compute column of the tableau
Synopsis
int glp_eval_tab_col(glp_prob *P, int k, int ind[], double val[]);
Description
The routine glp_eval_tab_col computes a column of the current simplex tableau (see Sub-
section 3.1.1, formula (3.12)), which (column) corresponds to some non-basic variable specified
by the parameter k: if 1 ≤ k ≤ m, the non-basic variable is k-th auxiliary variable, and if
m + 1 ≤ k ≤ m + n, the non-basic variable is (k − m)-th structural variable, where m is the
number of rows and n is the number of columns in the specified problem object. The basis factor-
ization must exist.
The computed column shows how basic variables depends on the specified non-basic variable
xk = (xN )j :
(xB )1 = . . . + ξ1j (xN )j + . . .
(xB )2 = . . . + ξ2j (xN )j + . . .
. . . . . . . . . .
(xB )m = . . . + ξmj (xN )j + . . .
where ξ1j , ξ2j , . . . , ξmj are elements of the simplex table column, (xB )1 , (xB )2 , . . . , (xB )m are
basic (auxiliary and structural) variables.
The routine stores row indices and corresponding numeric values of non-zero elements of the
computed column in unordered sparse format in locations ind[1], . . . , ind[len] and val[1], . . . ,
val[len], respectively, where 0 ≤ len ≤ m is the number of non-zero elements in the column
returned on exit.
Element indices stored in the array ind have the same sense as index k, i.e. indices 1 to m denote
auxiliary variables while indices m + 1 to m + n denote structural variables (all these variables are
obviously basic by definition).
Returns
The routine glp_eval_tab_col returns len, which is the number of non-zero elements in the
simplex table column stored in the arrays ind and val.
Comments
A column of the simplex table is computed as follows. At first, the routine checks that the
specified variable xk is non-basic and uses the permutation matrix Π (3.7) to determine index j of
non-basic variable (xN )j , which corresponds to xk .
The column to be computed is j-th column of the matrix Ξ (3.12), therefore:
Ξj = Ξej = −B −1 N ej = −B −1 Nj ,
where ej is j-th unity vector, Nj is j-th column of matrix N (3.9). So the routine performs
FTRAN to transform Nj to the simplex table column Ξj = (ξij ) and uses the permutation matrix
Π to convert row indices i to original ordinal numbers of auxiliary and structural variables.
112
4.3.3 glp transform row — transform explicitly specified row
Synopsis
int glp_transform_row(glp_prob *P, int len, int ind[], double val[]);
Description
The routine glp_transform_row performs the same operation as the routine glp_eval_tab_row
with exception that the row to be transformed is specified explicitly as a sparse vector.
The explicitly specified row may be thought as a linear form:
x = a1 xm+1 + a2 xm+2 + . . . + an xm+n ,
where x is an auxiliary variable for this row, aj are coefficients of the linear form, xm+j are structural
variables.
On entry column indices and numerical values of non-zero coefficients aj of the specified row
should be placed in locations ind[1], . . . , ind[len] and val[1], . . . , val[len], where len is
number of non-zero coefficients.
This routine uses the system of equality constraints and the current basis in order to express
the auxiliary variable x through the current non-basic variables (as if the transformed row were
added to the problem object and the auxiliary variable x were basic), i.e. the resultant row has the
form:
x = ξ1 (xN )1 + ξ2 (xN )2 + . . . + ξn (xN )n ,
where ξj are influence coefficients, (xN )j are non-basic (auxiliary and structural) variables, n is the
number of columns in the problem object.
On exit the routine stores indices and numerical values of non-zero coefficients ξj of the resultant
row in locations ind[1], . . . , ind[len’] and val[1], . . . , val[len’], where 0 ≤ len′ ≤ n is the
number of non-zero coefficients in the resultant row returned by the routine. Note that indices
of non-basic variables stored in the array ind correspond to original ordinal numbers of variables:
indices 1 to m mean auxiliary variables and indices m + 1 to m + n mean structural ones.
Returns
The routine glp_transform_row returns len’, the number of non-zero coefficients in the resul-
tant row stored in the arrays ind and val.
113
4.3.4 glp transform col — transform explicitly specified column
Synopsis
int glp_transform_col(glp_prob *P, int len, int ind[], double val[]);
Description
The routine glp_transform_col performs the same operation as the routine glp_eval_tab_col
with exception that the column to be transformed is specified explicitly as a sparse vector.
The explicitly specified column may be thought as it were added to the original system of
equality constraints:
x1 = a11 xm+1 + . . . + a1n xm+n + a1 x
x2 = a21 xm+1 + . . . + a2n xm+n + a2 x
. . . . . . . . . . . . . .
xm = am1 xm+1 + . . . + amn xm+n + am x
where xi are auxiliary variables, xm+j are structural variables (presented in the problem object),
x is a structural variable for the explicitly specified column, ai are constraint coefficients at x.
On entry row indices and numerical values of non-zero coefficients ai of the specified column
should be placed in locations ind[1], . . . , ind[len] and val[1], . . . , val[len], where len is
number of non-zero coefficients.
This routine uses the system of equality constraints and the current basis in order to express
the current basic variables through the structural variable x (as if the transformed column were
added to the problem object and the variable x were non-basic):
(xB )1 = . . . + ξ1 x
(xB )2 = . . . + ξ2 x
. . . . . .
(xB )m = . . . + ξm x
where ξi are influence coefficients, xB are basic (auxiliary and structural) variables, m is the number
of rows in the problem object.
On exit the routine stores indices and numerical values of non-zero coefficients ξi of the resultant
column in locations ind[1], . . . , ind[len’] and val[1], . . . , val[len’], where 0 ≤ len′ ≤ m
is the number of non-zero coefficients in the resultant column returned by the routine. Note that
indices of basic variables stored in the array ind correspond to original ordinal numbers of variables,
i.e. indices 1 to m mean auxiliary variables, indices m + 1 to m + n mean structural ones.
Returns
The routine glp_transform_col returns len’, the number of non-zero coefficients in the resul-
tant column stored in the arrays ind and val.
114
4.3.5 glp prim rtest — perform primal ratio test
Synopsis
int glp_prim_rtest(glp_prob *P, int len, const int ind[], const double val[],
int dir, double eps);
Description
The routine glp_prim_rtest performs the primal ratio test using an explicitly specified column
of the simplex table.
The current basic solution associated with the LP problem object must be primal feasible.
The explicitly specified column of the simplex table shows how the basic variables xB depend
on some non-basic variable x (which is not necessarily presented in the problem object):
(xB )1 = . . . + ξ1 x
(xB )2 = . . . + ξ2 x
. . . . . .
(xB )m = . . . + ξm x
The column is specifed on entry to the routine in sparse format. Ordinal numbers of basic
variables (xB )i should be placed in locations ind[1], . . . , ind[len], where ordinal number 1 to m
denote auxiliary variables, and ordinal numbers m + 1 to m + n denote structural variables. The
corresponding non-zero coefficients ξi should be placed in locations val[1], . . . , val[len]. The
arrays ind and val are not changed by the routine.
The parameter dir specifies direction in which the variable x changes on entering the basis: +1
means increasing, −1 means decreasing.
The parameter eps is an absolute tolerance (small positive number, say, 10−9 ) used by the
routine to skip ξi ’s whose magnitude is less than eps.
The routine determines which basic variable (among those specified in ind[1], . . . , ind[len])
reaches its (lower or upper) bound first before any other basic variables do, and which, therefore,
should leave the basis in order to keep primal feasibility.
Returns
The routine glp_prim_rtest returns the index, piv, in the arrays ind and val corresponding
to the pivot element chosen, 1 ≤ piv ≤ len. If the adjacent basic solution is primal unbounded,
and therefore the choice cannot be made, the routine returns zero.
Comments
If the non-basic variable x is presented in the LP problem object, the input column can be
computed with the routine glp_eval_tab_col; otherwise, it can be computed with the routine
glp_transform_col.
115
4.3.6 glp dual rtest — perform dual ratio test
Synopsis
int glp_dual_rtest(glp_prob *P, int len, const int ind[], const double val[],
int dir, double eps);
Description
The routine glp_dual_rtest performs the dual ratio test using an explicitly specified row of
the simplex table.
The current basic solution associated with the LP problem object must be dual feasible.
The explicitly specified row of the simplex table is a linear form that shows how some basic
variable x (which is not necessarily presented in the problem object) depends on non-basic variables
xN :
x = ξ1 (xN )1 + ξ2 (xN )2 + . . . + ξn (xN )n .
The row is specified on entry to the routine in sparse format. Ordinal numbers of non-basic
variables (xN )j should be placed in locations ind[1], . . . , ind[len], where ordinal numbers 1 to
m denote auxiliary variables, and ordinal numbers m + 1 to m + n denote structural variables. The
corresponding non-zero coefficients ξj should be placed in locations val[1], . . . , val[len]. The
arrays ind and val are not changed by the routine.
The parameter dir specifies direction in which the variable x changes on leaving the basis: +1
means that x goes on its lower bound, so its reduced cost (dual variable) is increasing (minimization)
or decreasing (maximization); −1 means that x goes on its upper bound, so its reduced cost is
decreasing (minimization) or increasing (maximization).
The parameter eps is an absolute tolerance (small positive number, say, 10−9 ) used by the
routine to skip ξj ’s whose magnitude is less than eps.
The routine determines which non-basic variable (among those specified in ind[1], . . . ,
ind[len]) should enter the basis in order to keep dual feasibility, because its reduced cost reaches
the (zero) bound first before this occurs for any other non-basic variables.
Returns
The routine glp_dual_rtest returns the index, piv, in the arrays ind and val corresponding
to the pivot element chosen, 1 ≤ piv ≤ len. If the adjacent basic solution is dual unbounded, and
therefore the choice cannot be made, the routine returns zero.
Comments
If the basic variable x is presented in the LP problem object, the input row can be computed
with the routine glp_eval_tab_row; otherwise, it can be computed with the routine
glp_transform_row.
116
4.4 Post-optimal analysis routines
4.4.1 glp analyze bound — analyze active bound of non-basic variable
Synopsis
void glp_analyze_bound(glp_prob *P, int k, double *limit1, int *var1,
double *limit2, int *var2);
Description
The routine glp_analyze_bound analyzes the effect of varying the active bound of specified
non-basic variable.
The non-basic variable is specified by the parameter k, where 1 ≤ k ≤ m means auxiliary
variable of corresponding row, and m + 1 ≤ k ≤ m + n means structural variable (column).
Note that the current basic solution must be optimal, and the basis factorization must exist.
Results of the analysis have the following meaning.
value1 is the minimal value of the active bound, at which the basis still remains primal feasible
and thus optimal. -DBL_MAX means that the active bound has no lower limit.
var1 is the ordinal number of an auxiliary (1 to m) or structural (m + 1 to m + n) basic variable,
which reaches its bound first and thereby limits further decreasing the active bound being analyzed.
if value1 = -DBL_MAX, var1 is set to 0.
value2 is the maximal value of the active bound, at which the basis still remains primal feasible
and thus optimal. +DBL_MAX means that the active bound has no upper limit.
var2 is the ordinal number of an auxiliary (1 to m) or structural (m + 1 to m + n) basic variable,
which reaches its bound first and thereby limits further increasing the active bound being analyzed.
if value2 = +DBL_MAX, var2 is set to 0.
The parameters value1, var1, value2, var2 can be specified as NULL, in which case correspond-
ing information is not stored.
4.4.2 glp analyze coef — analyze objective coefficient at basic variable
Synopsis
void glp_analyze_coef(glp_prob *P, int k,
double *coef1, int *var1, double *value1,
double *coef2, int *var2, double *value2);
Description
The routine glp_analyze_coef analyzes the effect of varying the objective coefficient at speci-
fied basic variable.
The basic variable is specified by the parameter k, where 1 ≤ k ≤ m means auxiliary variable
of corresponding row, and m + 1 ≤ k ≤ m + n means structural variable (column).
Note that the current basic solution must be optimal, and the basis factorization must exist.
Results of the analysis have the following meaning.
117
coef1 is the minimal value of the objective coefficient, at which the basis still remains dual
feasible and thus optimal. -DBL_MAX means that the objective coefficient has no lower limit.
var1 is the ordinal number of an auxiliary (1 to m) or structural (m + 1 to m + n) non-basic
variable, whose reduced cost reaches its zero bound first and thereby limits further decreasing the
objective coefficient being analyzed. If coef1 = -DBL_MAX, var1 is set to 0.
value1 is value of the basic variable being analyzed in an adjacent basis, which is defined as
follows. Let the objective coefficient reach its minimal value (coef1) and continue decreasing. Then
the reduced cost of the limiting non-basic variable (var1) becomes dual infeasible and the current
basis becomes non-optimal that forces the limiting non-basic variable to enter the basis replacing
there some basic variable that leaves the basis to keep primal feasibility. Should note that on
determining the adjacent basis current bounds of the basic variable being analyzed are ignored as if
it were free (unbounded) variable, so it cannot leave the basis. It may happen that no dual feasible
adjacent basis exists, in which case value1 is set to -DBL_MAX or +DBL_MAX.
coef2 is the maximal value of the objective coefficient, at which the basis still remains dual
feasible and thus optimal. +DBL_MAX means that the objective coefficient has no upper limit.
var2 is the ordinal number of an auxiliary (1 to m) or structural (m + 1 to m + n) non-basic
variable, whose reduced cost reaches its zero bound first and thereby limits further increasing the
objective coefficient being analyzed. If coef2 = +DBL_MAX, var2 is set to 0.
value2 is value of the basic variable being analyzed in an adjacent basis, which is defined exactly
in the same way as value1 above with exception that now the objective coefficient is increasing.
The parameters coef1, var1, value1, coef2, var2, value2 can be specified as NULL, in which
case corresponding information is not stored.
118
Chapter 5
Branch-and-Cut API Routines
5.1 Introduction
5.1.1 Using the callback routine
The GLPK MIP solver based on the branch-and-cut method allows the application program to
control the solution process. This is attained by means of the user-defined callback routine, which
is called by the solver at various points of the branch-and-cut algorithm.
The callback routine passed to the MIP solver should be written by the user and has the
following specification:1
void foo_bar(glp_tree *T, void *info);
where tree is a pointer to the data structure glp_tree, which should be used on subsequent calls to
branch-and-cut interface routines, and info is a transit pointer passed to the routine glp_intopt,
which may be used by the application program to pass some external data to the callback routine.
The callback routine is passed to the MIP solver through the control parameter structure
glp_iocp (see Chapter “Basic API Routines”, Section “Mixed integer programming routines”,
Subsection “Solve MIP problem with the branch-and-cut method”) as follows:
glp_prob *mip;
glp_iocp parm;
. . .
glp_init_iocp(&parm);
. . .
parm.cb_func = foo_bar;
parm.cb_info = ... ;
ret = glp_intopt(mip, &parm);
. . .
To determine why it is being called by the MIP solver the callback routine should use the routine
glp_ios_reason (described in this section below), which returns a code indicating the reason for
calling. Depending on the reason the callback routine may perform necessary actions to control the
solution process.
1
The name foo bar used here is a placeholder for the callback routine name.
119
The reason codes, which correspond to various point of the branch-and-cut algorithm imple-
mented in the MIP solver, are described in Subsection “Reasons for calling the callback routine”
below.
To ignore calls for reasons, which are not processed by the callback routine, it should simply
return to the MIP solver doing nothing. For example:
void foo_bar(glp_tree *T, void *info)
{ . . .
switch (glp_ios_reason(T))
{ case GLP_IBRANCH:
. . .
break;
case GLP_ISELECT:
. . .
break;
default:
/* ignore call for other reasons */
break;
}
return;
}
To control the solution process as well as to obtain necessary information the callback routine
may use the branch-and-cut API routines described in this chapter. Names of all these routines
begin with ‘glp_ios_’.
5.1.2 Branch-and-cut algorithm
This section gives a schematic description of the branch-and-cut algorithm as it is implemented

in the GLPK MIP solver.
1. Initialization
Set L := {P0 }, where L is the active list (i.e. the list of active subproblems), P0 is the original
MIP problem to be solved.
Set z best := +∞ (in case of minimization) or z best := −∞ (in case of maximization), where z best
is incumbent value, i.e. an upper (minimization) or lower (maximization) global bound for z opt ,
the optimal objective value for P 0 .
2. Subproblem selection
If L = ∅ then GO TO 9.
Select P ∈ L, i.e. make active subproblem P current.
3. Solving LP relaxation
Solve P LP , which is LP relaxation of P .
If P LP has no primal feasible solution then GO TO 8.
Let z LP be the optimal objective value for P LP .
If z LP ≥ z best (minimization) or z LP ≤ z best (), GO TO 8.
120
4. Adding “lazy” constraints
Let xLP be the optimal solution to P LP .
If there are “lazy” constraints (i.e. essential constraints not included in the original MIP problem
P0 ), which are violated at the optimal point xLP , add them to P , and GO TO 3.
5. Check for integrality
Let xj be a variable, which is required to be integer, and let xLP
j ∈ xLP be its value in the
optimal solution to P LP .
If xLP
j are integral for all integer variables, then a better integer feasible solution is found. Store
its components, set z best := z LP , and GO TO 8.
6. Adding cutting planes
If there are cutting planes (i.e. valid constraints for P ), which are violated at the optimal point
xLP , add them to P , and GO TO 3.
7. Branching
Select branching variable xj , i.e. a variable, which is required to be integer, and whose value
xLP
j ∈ xLP is fractional in the optimal solution to P LP .
Create new subproblem P D (so called down branch), which is identical to the current subproblem
P with exception that the upper bound of xj is replaced by ⌊xLPj ⌋. (For example, if xj
LP = 3.14,
the new upper bound of xj in the down branch will be ⌊3.14⌋ = 3.)
Create new subproblem P U (so called up branch), which is identical to the current subproblem
P with exception that the lower bound of xj is replaced by ⌈xLP
j ⌉. (For example, if xj
LP = 3.14,
the new lower bound of xj in the up branch will be ⌈3.14⌉ = 4.)

Set L := (L\{P }) ∪ {P D , P U }, i.e. remove the current subproblem P from the active list L and
add two new subproblems P D and P U to it. Then GO TO 2.
8. Pruning
Remove from the active list L all subproblems (including the current one), whose local bound
ze is not better than the global bound z best , i.e. set L := L\{P } for all P , where ze ≥ z best (in case
of minimization) or ze ≤ z best (in case of maximization), and then GO TO 2.
The local bound ze for subproblem P is an lower (minimization) or upper (maximization) bound
for integer optimal solution to this subproblem (not to the original problem). This bound is local
in the sense that only subproblems in the subtree rooted at node P cannot have better integer
feasible solutions. Note that the local bound is not necessarily the optimal objective value to LP
relaxation P LP .
9. Termination
If z best = +∞ (in case of minimization) or z best = −∞ (in case of maximization), the original
problem P0 has no integer feasible solution. Otherwise, the last integer feasible solution stored on
step 5 is the integer optimal solution to the original problem P0 with z opt = z best . STOP.
121
@ABC
GFED
89:;
?>=<
iiiiii A 88UUUUUUU
iiii 88 UUUU
iiiiiii 88 UUUU
UUUU
iii 8 UUUU
i i ii
GFED
@ABC
?>=<
89:; @ABC
GFED
89:;
?>=< GFED
@ABC
×

B9
99 ss
C 88QQQQQ
99 ss 888 QQ QQQ
ss
99 s ss 88 QQQ
QQQ
ss Q
@ABC
GFED
× GFED
@ABC
D GFED
@ABC
?>=<
89:;
E GFED
@ABC
89:;
?>=<
F GFED
@ABC
G
88 88 88
888 888 888
88 88 88

GFED
@ABC
× @ABC
GFED
× GFED
@ABC
× H @ABC
GFED
I @ABC
GFED
× @ABC
GFED
J
Current GFED
@ABC Active
@ABC
GFED
89:;
?>=< Non-active GFED
@ABC
× Fathomed
Fig. 1. An example of the search tree.
5.1.3 The search tree
On the branching step of the branch-and-cut algorithm the current subproblem is divided into
two2 new subproblems, so the set of all subproblems can be represented in the form of a rooted
tree, which is called the search or branch-and-bound tree. An example of the search tree is shown
on Fig. 1. Each node of the search tree corresponds to a subproblem, so the terms ‘node’ and
‘subproblem’ may be used synonymously.
In GLPK each node may have one of the following four statuses:
— current node is the active node currently being processed;
— active node is a leaf node, which still has to be processed;
— non-active node is a node, which has been processed, but not fathomed;
— fathomed node is a node, which has been processed and fathomed.
In the data structure representing the search tree GLPK keeps only current, active, and non-
active nodes. Once a node has been fathomed, it is removed from the tree data structure.
Being created each node of the search tree is assigned a distinct positive integer called the
subproblem reference number, which may be used by the application program to specify a particular
node of the tree. The root node corresponding to the original problem to be solved is always assigned
the reference number 1.
5.1.4 Current subproblem
The current subproblem is a MIP problem corresponding to the current node of the search tree.
It is represented as the GLPK problem object (glp_prob) that allows the application program
using API routines to access its content in the standard way. If the MIP presolver is not used, it is
2
In more general cases the current subproblem may be divided into more than two subproblems. However, currently
such feature is not used in GLPK.
122
the original problem object passed to the routine glp_intopt; otherwise, it is an internal problem
object built by the MIP presolver.
Note that the problem object is used by the MIP solver itself during the solution process
for various purposes (to solve LP relaxations, to perfom branching, etc.), and even if the MIP
presolver is not used, the current content of the problem object may differ from its original content.
For example, it may have additional rows, bounds of some rows and columns may be changed,
etc. In particular, LP segment of the problem object corresponds to LP relaxation of the current
subproblem. However, on exit from the MIP solver the content of the problem object is restored
to its original state.
To obtain information from the problem object the application program may use any API
routines, which do not change the object. Using API routines, which change the problem object,
is restricted to stipulated cases.
5.1.5 The cut pool
The cut pool is a set of cutting plane constraints maintained by the MIP solver. It is used by
the GLPK cut generation routines and may be used by the application program in the same way,
i.e. rather than to add cutting plane constraints directly to the problem object the application
program may store them to the cut pool. In the latter case the solver looks through the cut pool,
selects efficient constraints, and adds them to the problem object.
5.1.6 Reasons for calling the callback routine
The callback routine may be called by the MIP solver for the following reasons.
Request for subproblem selection
The callback routine is called with the reason code GLP_ISELECT if the current subproblem has
been fathomed and therefore there is no current subproblem.
In response the callback routine may select some subproblem from the active list and pass its
reference number to the solver using the routine glp_ios_select_node, in which case the solver
continues the search from the specified active subproblem. If no selection is made by the callback
routine, the solver uses a backtracking technique specified by the control parameter bt_tech.
To explore the active list (i.e. active nodes of the branch-and-bound tree) the callback routine
may use the routines glp_ios_next_node and glp_ios_prev_node.
Request for preprocessing
The callback routine is called with the reason code GLP_IPREPRO if the current subproblem has
just been selected from the active list and its LP relaxation is not solved yet.
In response the callback routine may perform some preprocessing of the current subproblem like
tightening bounds of some variables or removing bounds of some redundant constraints.
Request for row generation
The callback routine is called with the reason code GLP_IROWGEN if LP relaxation of the current
subproblem has just been solved to optimality and its objective value is better than the best known
integer feasible solution.
123
In response the callback routine may add one or more “lazy” constraints (rows), which are
violated by the current optimal solution of LP relaxation, using API routines glp_add_rows,
glp_set_row_name, glp_set_row_bnds, and glp_set_mat_row, in which case the solver will per-
form re-optimization of LP relaxation. If there are no violated constraints, the callback routine
should just return.
Note that components of optimal solution to LP relaxation can be obtained with API
routines glp_get_obj_val, glp_get_row_prim, glp_get_row_dual, glp_get_col_prim, and
glp_get_col_dual.
Request for heuristic solution
The callback routine is called with the reason code GLP_IHEUR if LP relaxation of the current
subproblem being solved to optimality is integer infeasible (i.e. values of some structural variables
of integer kind are fractional), though its objective value is better than the best known integer
feasible solution.
In response the callback routine may try applying a primal heuristic to find an integer feasible
solution,3 which is better than the best known one. In case of success the callback routine may
store such better solution in the problem object using the routine glp_ios_heur_sol.
Request for cut generation
The callback routine is called with the reason code GLP_ICUTGEN if LP relaxation of the current
feasible solution.
In response the callback routine may reformulate the current subproblem (before it will be
splitted up due to branching) by adding to the problem object one or more cutting plane constraints,
which cut off the fractional optimal point from the MIP polytope.4
Adding cutting plane constraints may be performed in two ways. One way is the same as
for the reason code GLP_IROWGEN (see above), in which case the callback routine adds new rows
corresponding to cutting plane constraints directly to the current subproblem.
The other way is to add cutting plane constraints to the cut pool, a set of cutting plane con-
straints maintained by the solver, rather than directly to the current subproblem. In this case
after return from the callback routine the solver looks through the cut pool, selects efficient cutting
plane constraints, adds them to the current subproblem, drops other constraints, and then performs
re-optimization.
Request for branching
The callback routine is called with the reason code GLP_IBRANCH if LP relaxation of the current
feasible solution.
In response the callback routine may choose some variable suitable for branching (i.e. integer
variable, whose value in optimal solution to LP relaxation of the current subproblem is fractional)
and pass its ordinal number to the solver using the routine glp_ios_branch_upon, in which case
the solver splits the current subproblem in two new subproblems and continues the search. If no
3
Integer feasible to the original MIP problem, not to the current subproblem.
4
Since these constraints are added to the current subproblem, they may be globally as well as locally valid.
124
choice is made by the callback routine, the solver uses a branching technique specified by the control
parameter br_tech.
Better integer solution found
The callback routine is called with the reason code GLP_IBINGO if LP relaxation of the current
subproblem being solved to optimality is integer feasible (i.e. values of all structural variables of
integer kind are integral within the working precision) and its objective value is better than the
best known integer feasible solution.
Optimal solution components for LP relaxation can be obtained in the same way as for the
reason code GLP_IROWGEN (see above).
Components of the new MIP solution can be obtained with API routines glp_mip_obj_val,
glp_mip_row_val, and glp_mip_col_val. Note, however, that due to row/cut generation there
may be additional rows in the problem object.
The difference between optimal solution to LP relaxation and corresponding MIP solution is
that in the former case some structural variables of integer kind (namely, basic variables) may have
values, which are close to nearest integers within the working precision, while in the latter case all
such variables have exact integral values.
The reason GLP_IBINGO is intended only for informational purposes, so the callback routine
should not modify the problem object in this case.
125
5.2 Basic routines
5.2.1 glp ios reason — determine reason for calling the callback routine
Synopsis
int glp_ios_reason(glp_tree *T);
Returns
The routine glp_ios_reason returns a code, which indicates why the user-defined callback
routine is being called:
GLP_ISELECT — request for subproblem selection;
GLP_IPREPRO — request for preprocessing;
GLP_IROWGEN — request for row generation;
GLP_IHEUR — request for heuristic solution;
GLP_ICUTGEN — request for cut generation;
GLP_IBRANCH — request for branching;
GLP_IBINGO — better integer solution found.
5.2.2 glp ios get prob — access the problem object
Synopsis
glp_prob *glp_ios_get_prob(glp_tree *T);
Description
The routine glp_ios_get_prob can be called from the user-defined callback routine to access
the problem object, which is used by the MIP solver. It is the original problem object passed to
the routine glp_intopt if the MIP presolver is not used; otherwise it is an internal problem object
built by the presolver.
Returns
The routine glp_ios_get_prob returns a pointer to the problem object used by the MIP solver.
Comments
To obtain various information about the problem instance the callback routine can access the
problem object (i.e. the object of type glp_prob) using the routine glp_ios_get_prob. It is
the original problem object passed to the routine glp_intopt if the MIP presolver is not used;
otherwise it is an internal problem object built by the presolver.
126
5.2.3 glp ios row attr — determine additional row attributes
Synopsis
void glp_ios_row_attr(glp_tree *T, int i, glp_attr *attr);
Description
The routine glp_ios_row_attr retrieves additional attributes of i-th row of the current sub-
problem and stores them in the structure glp_attr, which the parameter attr points to.
The structure glp_attr has the following fields:
int level
Subproblem level at which the row was created. (If level = 0, the row was added either to the
original problem object passed to the routine glp_intopt or to the root subproblem on generating
“lazy” or/and cutting plane constraints.)
int origin
The row origin flag:
GLP_RF_REG — regular constraint;
GLP_RF_LAZY — “lazy” constraint;
GLP_RF_CUT — cutting plane constraint.
int klass
The row class descriptor, which is a number passed to the routine glp_ios_add_row as its third
parameter. If the row is a cutting plane constraint generated by the solver, its class may be the
following:
GLP_RF_GMI — Gomory’s mixed integer cut;
GLP_RF_MIR — mixed integer rounding cut;
GLP_RF_COV — mixed cover cut;
GLP_RF_CLQ — clique cut.
5.2.4 glp ios mip gap — compute relative MIP gap
Synopsis
double glp_ios_mip_gap(glp_tree *T);
Description
The routine glp_ios_mip_gap computes the relative MIP gap (also called duality gap) with the
following formula:
|best mip − best bnd|
gap =
|best mip| + DBL EPSILON
where best_mip is the best integer feasible solution found so far, best_bnd is the best (global)
bound. If no integer feasible solution has been found yet, gap is set to DBL_MAX.
127
Returns
The routine glp_ios_mip_gap returns the relative MIP gap.
Comments
The relative MIP gap is used to measure the quality of the best integer feasible solution found
so far, because the optimal solution value z ∗ for the original MIP problem always lies in the range
best bnd ≤ z ∗ ≤ best mip
in case of minimization, or in the range
best mip ≤ z ∗ ≤ best bnd
in case of maximization.
To express the relative MIP gap in percents the value returned by the routine glp_ios_mip_gap
should be multiplied by 100%.
5.2.5 glp ios node data — access application-specific data
Synopsis
void *glp_ios_node_data(glp_tree *T, int p);
Description
The routine glp_ios_node_data allows the application accessing a memory block allocated for
the subproblem (which may be active or inactive), whose reference number is p.
The size of the block is defined by the control parameter cb_size passed to the routine
glp_intopt. The block is initialized by binary zeros on creating corresponding subproblem, and
its contents is kept until the subproblem will be removed from the tree.
The application may use these memory blocks to store specific data for each subproblem.
Returns
The routine glp_ios_node_data returns a pointer to the memory block for the specified sub-
problem. Note that if cb_size = 0, the routine returns a null pointer.
5.2.6 glp ios select node — select subproblem to continue the search
Synopsis
void glp_ios_select_node(glp_tree *T, int p);
Description
The routine glp_ios_select_node can be called from the user-defined callback routine in
response to the reason GLP_ISELECT to select an active subproblem, whose reference number
is p. The search will be continued from the subproblem selected.
128
5.2.7 glp ios heur sol — provide solution found by heuristic
Synopsis
int glp_ios_heur_sol(glp_tree *T, const double x[]);
Description
The routine glp_ios_heur_sol can be called from the user-defined callback routine in response
to the reason GLP_IHEUR to provide an integer feasible solution found by a primal heuristic.
Primal values of all variables (columns) found by the heuristic should be placed in locations
x[1], . . . , x[n], where n is the number of columns in the original problem object. Note that the
routine glp_ios_heur_sol does not check primal feasibility of the solution provided.
Using the solution passed in the array x the routine computes value of the objective function.
If the objective value is better than the best known integer feasible solution, the routine computes
values of auxiliary variables (rows) and stores all solution components in the problem object.
Returns
If the provided solution is accepted, the routine glp_ios_heur_sol returns zero. Otherwise, if
the provided solution is rejected, the routine returns non-zero.
5.2.8 glp ios can branch — check if can branch upon specified variable
Synopsis
int glp_ios_can_branch(glp_tree *T, int j);
Returns
If j-th variable (column) can be used to branch upon, the routine returns non-zero, otherwise
zero.
5.2.9 glp ios branch upon — choose variable to branch upon
Synopsis
void glp_ios_branch_upon(glp_tree *T, int j, int sel);
Description
The routine glp_ios_branch_upon can be called from the user-defined callback routine in re-
sponse to the reason GLP_IBRANCH to choose a branching variable, whose ordinal number
is j. Should note that only variables, for which the routine glp_ios_can_branch returns non-
zero, can be used to branch upon.
The parameter sel is a flag that indicates which branch (subproblem) should be selected next
to continue the search:
GLP_DN_BRNCH — select down-branch;
GLP_UP_BRNCH — select up-branch;
GLP_NO_BRNCH — use general selection technique.
129
Comments
On branching the solver removes the current active subproblem from the active list and creates
two new subproblems (down- and up-branches), which are added to the end of the active list. Note
that the down-branch is created before the up-branch, so the last active subproblem will be the
up-branch.
The down- and up-branches are identical to the current subproblem with exception that in
the down-branch the upper bound of xj , the variable chosen to branch upon, is replaced by ⌊x∗j ⌋,
while in the up-branch the lower bound of xj is replaced by ⌈x∗j ⌉, where x∗j is the value of xj in
optimal solution to LP relaxation of the current subproblem. For example, if x∗j = 3.14, the new
upper bound of xj in the down-branch is ⌊3.14⌋ = 3, and the new lower bound in the up-branch is
⌈3.14⌉ = 4.)
Additionally the callback routine may select either down- or up-branch, from which the solver
will continue the search. If none of the branches is selected, a general selection technique will be
used.
5.2.10 glp ios terminate — terminate the solution process
Synopsis
void glp_ios_terminate(glp_tree *T);
Description
The routine glp_ios_terminate sets a flag indicating that the MIP solver should prematurely
terminate the search.
130
5.3 The search tree exploring routines
5.3.1 glp ios tree size — determine size of the search tree
Synopsis
void glp_ios_tree_size(glp_tree *T, int *a_cnt, int *n_cnt, int *t_cnt);
Description
The routine glp_ios_tree_size stores the following three counts which characterize the current
size of the search tree:
a_cnt is the current number of active nodes, i.e. the current size of the active list;
n_cnt is the current number of all (active and inactive) nodes;
t_cnt is the total number of nodes including those which have been already removed from the
tree. This count is increased whenever a new node appears in the tree and never decreased.
If some of the parameters a_cnt, n_cnt, t_cnt is a null pointer, the corresponding count is not
stored.
5.3.2 glp ios curr node — determine current active subproblem
Synopsis
int glp_ios_curr_node(glp_tree *T);
Returns
The routine glp_ios_curr_node returns the reference number of the current active subproblem.
However, if the current subproblem does not exist, the routine returns zero.
5.3.3 glp ios next node — determine next active subproblem
Synopsis
int glp_ios_next_node(glp_tree *T, int p);
Returns
If the parameter p is zero, the routine glp_ios_next_node returns the reference number of the
first active subproblem. However, if the tree is empty, zero is returned.
If the parameter p is not zero, it must specify the reference number of some active subproblem,
in which case the routine returns the reference number of the next active subproblem. However, if
there is no next active subproblem in the list, zero is returned.
All subproblems in the active list are ordered chronologically, i.e. subproblem A precedes
subproblem B if A was created before B.
131
5.3.4 glp ios prev node — determine previous active subproblem
Synopsis
int glp_ios_prev_node(glp_tree *T, int p);
Returns
If the parameter p is zero, the routine glp_ios_prev_node returns the reference number of the
last active subproblem. However, if the tree is empty, zero is returned.
If the parameter p is not zero, it must specify the reference number of some active subproblem,
in which case the routine returns the reference number of the previous active subproblem. However,
if there is no previous active subproblem in the list, zero is returned.
All subproblems in the active list are ordered chronologically, i.e. subproblem A precedes
subproblem B if A was created before B.
5.3.5 glp ios up node — determine parent subproblem
Synopsis
int glp_ios_up_node(glp_tree *T, int p);
Returns
The parameter p must specify the reference number of some (active or inactive) subproblem, in
which case the routine iet_get_up_node returns the reference number of its parent subproblem.
However, if the specified subproblem is the root of the tree and, therefore, has no parent, the routine
returns zero.
5.3.6 glp ios node level — determine subproblem level
Synopsis
int glp_ios_node_level(glp_tree *T, int p);
Returns
The routine glp_ios_node_level returns the level of the subproblem, whose reference number
is p, in the branch-and-bound tree. (The root subproblem has level 0, and the level of any other
subproblem is the level of its parent plus one.)
5.3.7 glp ios node bound — determine subproblem local bound
Synopsis
double glp_ios_node_bound(glp_tree *T, int p);
Returns
The routine glp_ios_node_bound returns the local bound for (active or inactive) subproblem,
whose reference number is p.
132
Comments
The local bound for subproblem p is an lower (minimization) or upper (maximization) bound
for integer optimal solution to this subproblem (not to the original problem). This bound is local in
the sense that only subproblems in the subtree rooted at node p cannot have better integer feasible
solutions.
On creating a subproblem (due to the branching step) its local bound is inherited from its
parent and then may get only stronger (never weaker). For the root subproblem its local bound
is initially set to -DBL_MAX (minimization) or +DBL_MAX (maximization) and then improved as the
root LP relaxation has been solved.
Note that the local bound is not necessarily the optimal objective value to corresponding LP
relaxation.
5.3.8 glp ios best node — find active subproblem with best local bound
Synopsis
int glp_ios_best_node(glp_tree *T);
Returns
The routine glp_ios_best_node returns the reference number of the active subproblem, whose
local bound is best (i.e. smallest in case of minimization or largest in case of maximization).
However, if the tree is empty, the routine returns zero.
Comments
The best local bound is an lower (minimization) or upper (maximization) bound for integer
optimal solution to the original MIP problem.
133
5.4 The cut pool routines
5.4.1 glp ios pool size — determine current size of the cut pool
Synopsis
int glp_ios_pool_size(glp_tree *T);
Returns
The routine glp_ios_pool_size returns the current size of the cut pool, that is, the number
of cutting plane constraints currently added to it.
5.4.2 glp ios add row — add constraint to the cut pool
Synopsis
int glp_ios_add_row(glp_tree *T, const char *name, int klass, int flags,
int len, const int ind[], const double val[], int type, double rhs);
Description
The routine glp_ios_add_row adds specified row (cutting plane constraint) to the cut pool.
The cutting plane constraint should have the following format:
∑ { }
≥
aj xj b,
≤
j∈J
where J is a set of indices (ordinal numbers) of structural variables, aj are constraint coefficients,
xj are structural variables, b is the right-hand side.
The parameter name specifies a symbolic name assigned to the constraint (1 up to 255 charac-
ters). If it is NULL or an empty string, no name is assigned.
The parameter klass specifies the constraint class, which must be either zero or a number in
the range from 101 to 200. The application may use this attribute to distinguish between cutting
plane constraints of different classes.5
The parameter flags currently is not used and must be zero.
Ordinal numbers of structural variables (i.e. column indices) j ∈ J and numerical values of
corresponding constraint coefficients aj should be placed in locations ind[1], . . . , ind[len] and
val[1], . . . , val[len], respectively, where len = |J| is the number of constraint coefficients,
0 ≤ len ≤ n, and n is the number of columns in the problem object. Coefficients with identical
column indices are not allowed. Zero coefficients are allowed, however, they are ignored.
The parameter type specifies the constraint type as follows:
GLP_LO means inequality constraint Σaj xj ≥ b;
GLP_UP means inequality constraint Σaj xj ≤ b;
5
Constraint classes numbered from 1 to 100 are reserved for GLPK cutting plane generators.
134
The parameter rhs specifies the right-hand side b.
All cutting plane constraints in the cut pool are identified by their ordinal numbers 1, 2, . . . ,
size, where size is the current size of the cut pool. New constraints are always added to the end
of the cut pool, thus, ordinal numbers of previously added constraints are not changed.
Returns
The routine glp_ios_add_row returns the ordinal number of the cutting plane constraint added,
which is the new size of the cut pool.
Example
/* generate triangle cutting plane:
x[i] + x[j] + x[k] <= 1 */
. . .
/* add the constraint to the cut pool */
ind[1] = i, val[1] = 1.0;
ind[2] = j, val[2] = 1.0;
ind[3] = k, val[3] = 1.0;
glp_ios_add_row(tree, NULL, TRIANGLE_CUT, 0, 3, ind, val, GLP_UP, 1.0);
Comments
Cutting plane constraints added to the cut pool are intended to be then added only to the
current subproblem, so these constraints can be globally as well as locally valid. However, adding
a constraint to the cut pool does not mean that it will be added to the current subproblem—it
depends on the solver’s decision: if the constraint seems to be efficient, it is moved from the pool
to the current subproblem, otherwise it is simply dropped.6
Normally, every time the callback routine is called for cut generation, the cut pool is empty.
On the other hand, the solver itself can generate cutting plane constraints (like Gomory’s or mixed
integer rounding cuts), in which case the cut pool may be non-empty.
5.4.3 glp ios del row — remove constraint from the cut pool
Synopsis
void glp_ios_del_row(glp_tree *T, int i);
Description
The routine glp_ios_del_row deletes i-th row (cutting plane constraint) from the cut pool,
where 1 ≤ i ≤ size is the ordinal number of the constraint in the pool, size is the current size of
the cut pool.
Note that deleting a constraint from the cut pool leads to changing ordinal numbers of other
constraints remaining in the pool. New ordinal numbers of the remaining constraints are assigned
under assumption that the original order of constraints is not changed. Let, for example, there be
four constraints a, b, c and d in the cut pool, which have ordinal numbers 1, 2, 3 and 4, respectively,
and let constraint b have been deleted. Then after deletion the remaining constraint a, c and d are
assigned new ordinal numbers 1, 2 and 3, respectively.
6
Globally valid constraints could be saved and then re-used for other subproblems, but currently such feature is
not implemented.
135
To find the constraint to be deleted the routine glp_ios_del_row uses “smart” linear search,
so it is recommended to remove constraints in a natural or reverse order and avoid removing them
in a random order.
Example
/* keep first 10 constraints in the cut pool and remove other
constraints */
while (glp_ios_pool_size(tree) > 10)
glp_ios_del_row(tree, glp_ios_pool_size(tree));
5.4.4 glp ios clear pool — remove all constraints from the cut pool
Synopsis
void glp_ios_clear_pool(glp_tree *T);
Description
The routine glp_ios_clear_pool makes the cut pool empty deleting all existing rows (cutting
plane constraints) from it.
136
Chapter 6
Miscellaneous API Routines
6.1 GLPK environment routines
6.1.1 glp init env — initialize GLPK environment
Synopsis
int glp_init_env(void);
Description
The routine glp_init_env initializes the GLPK environment. Normally the application pro-
gram does not need to call this routine, because it is called automatically on the first call to any
API routine.
Returns
0 initialization successful;
1 environment is already initialized;
2 initialization failed (insufficient memory);
3 initialization failed (unsupported programming model).
6.1.2 glp version — determine library version
Synopsis
const char *glp_version(void);
Returns
The routine glp_version returns a pointer to a null-terminated character string, which specifies
the version of the GLPK library in the form "X.Y", where ‘X’ is the major version number, and ‘Y’
is the minor version number, for example, "4.16".
137
6.1.3 glp free env — free GLPK environment
Synopsis
int glp_free_env(void);
Description
The routine glp_free_env frees all resources used by GLPK routines (memory blocks, etc.)
which are currently still in use.
Normally the application program does not need to call this routine, because GLPK routines
always free all unused resources. However, if the application program even has deleted all problem
objects, there will be several memory blocks still allocated for the internal library needs. For some
reasons the application program may want GLPK to free this memory, in which case it should call
glp_free_env.
Note that a call to glp_free_env invalidates all problem objects which still exist.
Returns
0 termination successful;
1 environment is inactive (was not initialized).
6.1.4 glp printf — write formatted output to terminal
Synopsis
void glp_printf(const char *fmt, ...);
Description
The routine glp_printf uses the format control string fmt to format its parameters and writes
the formatted output to the terminal.
This routine is a replacement of the standard C function printf and used by all GLPK routines
to perform terminal output. The application program may use glp_printf for the same purpose
that allows controlling its terminal output with the routines glp_term_out and glp_term_hook.
6.1.5 glp vprintf — write formatted output to terminal
Synopsis
void glp_vprintf(const char *fmt, va_list arg);
Description
The routine glp_vprintf uses the format control string fmt to format its parameters specified
by the list arg and writes the formatted output to the terminal.
This routine is a replacement of the standard C function vprintf and used by all GLPK routines
to perform terminal output. The application program may use glp_vprintf for the same purpose
that allows controlling its terminal output with the routines glp_term_out and glp_term_hook.
138
6.1.6 glp term out — enable/disable terminal output
Synopsis
int glp_term_out(int flag);
Description
Depending on the parameter flag the routine glp_term_out enables or disables terminal output
performed by glpk routines:
GLP_ON — enable terminal output;
GLP_OFF — disable terminal output.
Returns
The routine glp_term_out returns the previous value of the terminal output flag.
6.1.7 glp term hook — intercept terminal output
Synopsis
void glp_term_hook(int (*func)(void *info, const char *s), void *info);
Description
The routine glp_term_hook installs the user-defined hook routine to intercept all terminal
output performed by GLPK routines.
The parameter func specifies the user-defined hook routine. It is called from an internal printing
routine, which passes to it two parameters: info and s. The parameter info is a transit pointer
specified in corresponding call to the routine glp_term_hook; it may be used to pass some additional
information to the hook routine. The parameter s is a pointer to the null terminated character
string, which is intended to be written to the terminal. If the hook routine returns zero, the printing
routine writes the string s to the terminal in a usual way; otherwise, if the hook routine returns
non-zero, no terminal output is performed.
To uninstall the hook routine both parameters func and info should be specified as NULL.
Example
static int hook(void *info, const char *s)
{ FILE *foo = info;
fputs(s, foo);
return 1;
}
int main(void)
{ FILE *foo;
. . .
glp_term_hook(hook, foo); /* redirect terminal output */
. . .
glp_term_hook(NULL, NULL); /* resume terminal output */
. . .
}
139
6.1.8 glp open tee — start copying terminal output
Synopsis
int glp_open_tee(const char *fname);
Description
The routine glp_open_tee starts copying all the terminal output to an output text file, whose
name is specified by the character string fname.
Returns
0 operation successful;
1 copying terminal output is already active;
2 unable to create output file.
6.1.9 glp close tee — stop copying terminal output
Synopsis
int glp_close_tee(void);
Description
The routine glp_close_tee stops copying the terminal output to the output text file previously
open by the routine glp_open_tee closing that file.
Returns
0 operation successful;
1 copying terminal output was not started.
6.1.10 glp error — display error message and terminate execution
Synopsis
void glp_error(const char *fmt, ...);
Description
The routine glp_error (implemented as a macro) formats its parameters using the format
control string fmt, writes the formatted message to the terminal, and then abnormally terminates
the program.
140
6.1.11 glp at error — check for error state
Synopsis
int glp_at_error(void);
Description
The routine glp_at_error checks if the GLPK environment is at error state, i. e. if the call to
the routine is (indirectly) made from the glp_error routine via an user-defined hook routine.
This routine can be used, for example, by a custom output handler (installed with the routine
glp_term_hook) to determine whether or not the message to be displayed is an error message.
Returns
If the GLPK environment is at error state, the routine returns non-zero, otherwise zero.
6.1.12 glp assert — check logical condition
Synopsis
void glp_assert(int expr);
Description
The routine glp_assert (implemented as a macro) checks a logical condition specified by the
expression expr. If the condition is true (non-zero), the routine does nothing; otherwise, if the
condition is false (zero), the routine prints an error message and abnormally terminates the program.
This routine is a replacement of the standard C function assert and used by all GLPK routines
to check program logic. The application program may use glp_assert for the same purpose.
6.1.13 glp error hook — install hook to intercept abnormal termination
Synopsis
void glp_error_hook(void (*func)(void *info), void *info);
Description
The routine glp_error_hook installs a user-defined hook routine to intercept abnormal termi-
nation.
The parameter func specifies the user-defined hook routine. It is called from the routine
glp_error before the latter calls the abort function to abnormally terminate the application pro-
gram because of fatal error. The parameter info is a transit pointer, specified in the corresponding
call to the routine glp_error_hook; it may be used to pass some information to the hook routine.
To uninstall the hook routine the parameters func and info should be specified as NULL.
If the hook routine returns, the application program is abnormally terminated. To prevent
abnormal termnation the hook routine may perform a global jump using the standard function
longjmp, in which case the application program must call the routine glp_free_env.
141
6.1.14 glp alloc — allocate memory block
Synopsis
void *glp_alloc(int n, int size);
Description
The routine glp_alloc dynamically allocates a memory block of n×size bytes long. Note that:
1) the parameters n and size must be positive;
2) having been allocated the memory block contains arbitrary data, that is, it is not initialized
by binary zeros;
3) if the block cannot be allocated due to insufficient memory, the routine prints an error message
and abnormally terminates the program.
This routine is a replacement of the standard C function malloc and used by GLPK routines for
dynamic memory allocation. The application program may use glp_alloc for the same purpose.
Returns
The routine glp_alloc returns a pointer to the memory block allocated. To free this block the
routine glp_free (not the standard C function free!) should be used.
6.1.15 glp realloc — reallocate memory block
Synopsis
void *glp_realloc(void *ptr, int n, int size);
Description
The routine glp_realloc dynamically reallocates a memory block pointed to by ptr, which
was previously allocated by the routine glp_alloc or reallocated by this routine. Note that the
pointer ptr must be valid and must not be NULL. The new size of the memory block is n×size
bytes long. Note that:
1) both parameters n and size must be positive;
2) if the block cannot be reallocated due to insufficient memory, the routine prints an error
message and abnormally terminates the program.
This routine is a replacement of the standard C function realloc and used by GLPK routines
for dynamic memory allocation. The application program may use glp_realloc for the same
purpose.
Returns
The routine glp_realloc returns a pointer to the memory block reallocated. To free this block
the routine glp_free (not the standard C function free!) should be used.
142
6.1.16 glp free — free memory block
Synopsis
void glp_free(void *ptr);
Description
The routine glp_free deallocates a memory block pointed to by ptr, which was previously
allocated by the routine glp_malloc or reallocated by the routine glp_realloc. Note that the
pointer ptr must be valid and must not be NULL.
This routine is a replacement of the standard C function free and used by GLPK routines for
dynamic memory allocation. The application program may use glp_free for the same purpose.
6.1.17 glp mem usage — get memory usage information
Synopsis
void glp_mem_usage(int *count, int *cpeak, size_t *total, size_t *tpeak);
Description
The routine glp_mem_usage reports some information about utilization of the memory by the
routines glp_malloc, glp_calloc, and glp_free. Information is stored to locations specified by
corresponding parameters (see below). Any parameter can be specified as NULL, in which case
corresponding information is not stored.
*count is the number of currently allocated memory blocks.
*cpeak is the peak value of *count reached since the initialization of the GLPK library envi-
ronment.
*total is the total amount, in bytes, of currently allocated memory blocks.
*tpeak is the peak value of *total reached since the initialization of the GLPK library enviri-
onment.
Example
glp_mem_usage(&count, NULL, NULL, NULL);
printf("%d memory block(s) are still allocated\n", count);
6.1.18 glp mem limit — set memory usage limit
Synopsis
void glp_mem_limit(int limit);
Description
The routine glp_mem_limit limits the amount of memory available for dynamic allocation (with
the routines glp_malloc and glp_calloc) to limit megabytes.
143
Appendix A
Installing GLPK on Your Computer
A.1 Downloading the distribution tarball
The distribution tarball of the most recent version of the GLPK package can be found on
http://ftp.gnu.org/gnu/glpk/ [via http] and ftp://ftp.gnu.org/gnu/glpk/ [via FTP]. It can
also be found on one of the FTP mirrors; see http://www.gnu.org/prep/ftp.html. Please use a
mirror if possible.
To make sure that the GLPK distribution tarball you have downloaded is intact you need to
download the corresponding ‘.sig’ file and run a command like this:
gpg --verify glpk-4.38.tar.gz.sig
If that command fails because you do not have the required public key, run the following command
to import it:
gpg --keyserver keys.gnupg.net --recv-keys 5981E818
and then re-run the previous command.
A.2 Unpacking the distribution tarball
The GLPK package (like all other GNU software) is distributed in the form of packed archive.
This is one file named glpk-X.Y.tar.gz, where X is the major version number and Y is the minor
version number.
In order to prepare the distribution for installation you need to copy the GLPK distribution file
to a working subdirectory and then unpack and unarchive the distribution file with the following
command:
tar zx < glpk-X.Y.tar
144
A.3 Configuring the package
After unpacking and unarchiving the GLPK distribution you should configure the package,
i.e. automatically tune it for your platform.
Normally, you should just cd to the subdirectory glpk-X.Y and run the configure script, e.g.
./configure
The ‘configure’ shell script attempts to guess correct values for various system-dependent
variables used during compilation. It uses those values to create a ‘Makefile’ in each directory
of the package. It also creates file ‘config.h’ containing platform-dependent definitions. Finally,
it creates a shell script ‘config.status’ that you can run in the future to recreate the current
configuration, a file ‘config.cache’ that saves the results of its tests to speed up reconfiguring,
and a file ‘config.log’ containing compiler output (useful mainly for debugging ‘configure’).
Running ‘configure’ takes about a minute. While it is running, it displays some informational
messages that tell you what it is doing. If you don’t want to see these messages, run ‘configure’
with its standard output redirected to ‘dev/null’; for example, ‘./configure > /dev/null’.
By default both static and shared versions of the GLPK library will be compiled. Compilation of
the shared librariy can be turned off by specifying the ‘--disable-shared’ option to ‘configure’:
./configure --disable-shared
If you encounter problems building the library try using the above option, because some platforms
do not support shared libraries.
The GLPK package has some optional features listed below. By default all these features are
disabled. To enable a feature the corresponding option should be passed to the configure script.
--with-gmp Enable using the GNU MP bignum library
This feature allows the exact simplex solver to use the GNU MP bignum library. If it is disabled,
the exact simplex solver uses the GLPK bignum module, which provides the same functionality as
GNU MP, however, it is much less efficient.
For details about the GNU MP bignum library see its web page at http://gmplib.org/.
--enable-dl The same as ‘--enable-dl=ltdl’
--enable-dl=ltdl Enable shared library support (GNU)
--enable-dl=dlfcn Enable shared library support (POSIX)
Currently this feature is only needed to provide dynamic linking to ODBC and MySQL shared
libraries (see below).
For details about the GNU shared library support see the manual at http://www.gnu.org/
software/libtool/manual/.
--enable-odbc Enable using ODBC table driver (libiodbc)
--enable-odbc=unix Enable using ODBC table driver (libodbc)
This feature allows transmitting data between MathProg model objects and relational databases
accessed through ODBC.
For more details about this feature see the supplement “Using Data Tables in the GNU Math-
Prog Modeling Language” (doc/tables.pdf).
145
--enable-mysql Enable using MySQL table driver (libmysql)
This feature allows transmitting data between MathProg model objects and MySQL relational
databases.
For more details about this feature see the supplement “Using Data Tables in the GNU Math-
Prog Modeling Language” (doc/tables.pdf).
A.4 Compiling the package
Normally, you can compile (build) the package by typing the command:
make
It reads ‘Makefile’ generated by ‘configure’ and performs all necessary jobs.
If you want, you can override the ‘make’ variables CFLAGS and LDFLAGS like this:
make CFLAGS=-O2 LDFLAGS=-s
To compile the package in a different directory from the one containing the source code, you
must use a version of ‘make’ that supports ‘VPATH’ variable, such as GNU ‘make’. ‘cd’ to the
directory where you want the object files and executables to go and run the ‘configure’ script.
‘configure’ automatically checks for the source code in the directory that ‘configure’ is in and
in ‘..’. If for some reason ‘configure’ is not in the source code directory that you are configuring,
then it will report that it can’t find the source code. In that case, run ‘configure’ with the option
‘--srcdir=DIR’, where DIR is the directory that contains the source code.
Some systems require unusual options for compilation or linking that the ‘configure’ script
does not know about. You can give ‘configure’ initial values for variables by setting them in the
environment. Using a Bourne-compatible shell, you can do that on the command line like this:
CC=c89 CFLAGS=-O2 LIBS=-lposix ./configure
Or on systems that have the ‘env’ program, you can do it like this:
env CPPFLAGS=-I/usr/local/include LDFLAGS=-s ./configure
Here are the ‘make’ variables that you might want to override with environment variables when
running ‘configure’.
For these variables, any value given in the environment overrides the value that ‘configure’
would choose:
CC C compiler program. The default is ‘cc’.
INSTALL Program used to install files. The default value is ‘install’ if you have it,
otherwise ‘cp’.
For these variables, any value given in the environment is added to the value that ‘configure’
chooses:
DEFS Configuration options, in the form ‘-Dfoo -Dbar . . . ’.
LIBS Libraries to link with, in the form ‘-lfoo -lbar . . . ’.
146
A.5 Checking the package
To check the package, i.e. to run some tests included in the package, you can use the following
command:
make check
A.6 Installing the package
Normally, to install the GLPK package you should type the following command:
make install
By default, ‘make install’ will install the package’s files in ‘usr/local/bin’, ‘usr/local/lib’,
etc. You can specify an installation prefix other than ‘/usr/local’ by giving ‘configure’ the
option ‘--prefix=PATH’. Alternately, you can do so by consistently giving a value for the ‘prefix’
variable when you run ‘make’, e.g.
make prefix=/usr/gnu
make prefix=/usr/gnu install
After installing you can remove the program binaries and object files from the source directory by
typing ‘make clean’. To remove all files that ‘configure’ created (‘Makefile’, ‘config.status’,
etc.), just type ‘make distclean’.
The file ‘configure.ac’ is used to create ‘configure’ by a program called ‘autoconf’. You
only need it if you want to remake ‘configure’ using a newer version of ‘autoconf’.
A.7 Uninstalling the package
To uninstall the GLPK package, i.e. to remove all the package’s files from the system places,
you can use the following command:
make uninstall
147
Appendix B
MPS Format
B.1 Fixed MPS Format
The MPS format1 is intended for coding LP/MIP problem data. This format assumes the
formulation of LP/MIP problem (1.1)—(1.3) (see Section 1.1, page 9).
MPS file is a text file, which contains two types of cards2 : indicator cards and data cards.
Indicator cards determine a kind of succeeding data. Each indicator card has one word in
uppercase letters beginning in column 1.
Data cards contain problem data. Each data card is divided into six fixed fields:
Field 1 Field 2 Field 3 Field 4 Field 5 Field 6
Columns 2—3 5—12 15—22 25—36 40—47 50—61
Contents Code Name Name Number Name Number
On a particular data card some fields may be optional.
Names are used to identify rows, columns, and some vectors (see below).
Aligning the indicator code in the field 1 to the left margin is optional.
All names specified in the fields 2, 3, and 5 should contain from 1 up to 8 arbitrary characters
(except control characters). If a name is placed in the field 3 or 5, its first character should not be
the dollar sign ‘$’. If a name contains spaces, the spaces are ignored.
All numerical values in the fields 4 and 6 should be coded in the form sxxEsyy, where s is the
plus ‘+’ or the minus ‘-’ sign, xx is a real number with optional decimal point, yy is an integer
decimal exponent. Any number should contain up to 12 characters. If the sign s is omitted, the
plus sign is assumed. The exponent part is optional. If a number contains spaces, the spaces are
ignored.
If a card has the asterisk ‘*’ in the column 1, this card is considered as a comment and ignored.
Besides, if the first character in the field 3 or 5 is the dollar sign ‘$’, all characters from the dollar
sign to the end of card are considered as a comment and ignored.
1
The MPS format was developed in 1960’s by IBM as input format for their mathematical programming system
MPS/360. Today the MPS format is a most widely used format understood by most mathematical programming
packages. This appendix describes only the features of the MPS format, which are implemented in the GLPK package.
2
In 1960’s MPS file was a deck of 80-column punched cards, so the author decided to keep the word “card”, which
may be understood as “line of text file”.
148
MPS file should contain cards in the following order:
— NAME indicator card;
— ROWS indicator card;
— data cards specifying rows (constraints);
— COLUMNS indicator card;
— data cards specifying columns (structural variables) and constraint coefficients;
— RHS indicator card;
— data cards specifying right-hand sides of constraints;
— RANGES indicator card;
— data cards specifying ranges for double-bounded constraints;
— BOUNDS indicator card;
— data cards specifying types and bounds of structural variables;
— ENDATA indicator card.
Section is a group of cards consisting of an indicator card and data cards succeeding this indi-
cator card. For example, the ROWS section consists of the ROWS indicator card and data cards
specifying rows.
The sections RHS, RANGES, and BOUNDS are optional and may be omitted.
B.2 Free MPS Format
Free MPS format is an improved version of the standard (fixed) MPS format described above.3
Note that all changes in free MPS format concern only the coding of data while the structure of
data is the same for both fixed and free versions of the MPS format.
In free MPS format indicator and data records4 may have arbitrary length not limited to 80
characters. Fields of data records have no predefined positions, i.e. the fields may begin in any
position, except position 1, which must be blank, and must be separated from each other by one
or more blanks. However, the fields must appear in the same order as in fixed MPS format.
Symbolic names in fields 2, 3, and 5 may be longer than 8 characters5 and must not contain
embedded blanks.
Numeric values in fields 4 and 6 are limited to 12 characters and must not contain embedded
blanks.
Only six fields on each data record are used. Any other fields are ignored.
If the first character of any field (not necessarily fields 3 and 5) is the dollar sign ($), all
characters from the dollar sign to the end of record are considered as a comment and ignored.
3
This format was developed in the beginning of 1990’s by IBM as an alternative to the standard fixed MPS format
for Optimization Subroutine Library (OSL).
4
Record in free MPS format has the same meaning as card in fixed MPS format.
5
GLPK allows symbolic names having up to 255 characters.
149
B.3 NAME indicator card
The NAME indicator card should be the first card in the MPS file (except optional comment
cards, which may precede the NAME card). This card should contain the word NAME in the columns
1—4 and the problem name in the field 3. The problem name is optional and may be omitted.
B.4 ROWS section
The ROWS section should start with the indicator card, which contains the word ROWS in the
columns 1—4.
Each data card in the ROWS section specifies one row (constraint) of the problem. All these
data cards have the following format.
‘N’ in the field 1 means that the row is free (unbounded):
−∞ < xi = ai1 xm+1 + ai2 xm+2 + . . . + ain xm+n < +∞;
‘L’ in the field 1 means that the row is of “less than or equal to” type:
−∞ < xi = ai1 xm+1 + ai2 xm+2 + . . . + ain xm+n ≤ bi ;
‘G’ in the field 1 means that the row is of “greater than or equal to” type:
bi ≤ xi = ai1 xm+1 + ai2 xm+2 + . . . + ain xm+n < +∞;
‘E’ in the field 1 means that the row is of “equal to” type:
xi = ai1 xm+1 + ai2 xm+2 + . . . + ain xm+n ≤ bi ,
where bi is a right-hand side. Note that each constraint has a corresponding implictly defined
auxiliary variable (xi above), whose value is a value of the corresponding linear form, therefore row
bounds can be considered as bounds of such auxiliary variable.
The filed 2 specifies a row name (which is considered as the name of the corresponding auxiliary
variable).
The fields 3, 4, 5, and 6 are not used and should be empty.
Numerical values of all non-zero right-hand sides bi should be specified in the RHS section (see
below). All double-bounded (ranged) constraints should be specified in the RANGES section (see
below).
B.5 COLUMNS section
The COLUMNS section should start with the indicator card, which contains the word COLUMNS
in the columns 1—7.
Each data card in the COLUMNS section specifies one or two constraint coefficients aij and
also introduces names of columns, i.e. names of structural variables. All these data cards have the
following format.
150
The field 1 is not used and should be empty.
The field 2 specifies a column name. If this field is empty, the column name from the immediately
preceeding data card is assumed.
The field 3 specifies a row name defined in the ROWS section.
The field 4 specifies a numerical value of the constraint coefficient aij , which is placed in the
corresponding row and column.
The fields 5 and 6 are optional. If they are used, they should contain a second pair “row
name—constraint coefficient” for the same column.
Elements of the constraint matrix (i.e. constraint coefficients) should be enumerated in the
column wise manner: all elements for the current column should be specified before elements for
the next column. However, the order of rows in the COLUMNS section may differ from the order
of rows in the ROWS section.
Constraint coefficients not specified in the COLUMNS section are considered as zeros. Therefore
zero coefficients may be omitted, although it is allowed to explicitly specify them.
B.6 RHS section
The RHS section should start with the indicator card, which contains the word RHS in the
columns 1—3.
Each data card in the RHS section specifies one or two right-hand sides bi (see Section B.4,
page 150). All these data cards have the following format.
The field 2 specifies a name of the right-hand side (RHS) vector6 . If this field is empty, the RHS
vector name from the immediately preceeding data card is assumed.
The field 4 specifies a right-hand side bi for the row, whose name is specified in the field 3.
Depending on the row type bi is a lower bound (for the row of G type), an upper bound (for the
row of L type), or a fixed value (for the row of E type).7
name—right-hand side” for the same RHS vector.
All right-hand sides for the current RHS vector should be specified before right-hand sides for
the next RHS vector. However, the order of rows in the RHS section may differ from the order of
rows in the ROWS section.
Right-hand sides not specified in the RHS section are considered as zeros. Therefore zero right-
hand sides may be omitted, although it is allowed to explicitly specify them.
6
This feature allows the user to specify several RHS vectors in the same MPS file. However, before solving the
problem a particular RHS vector should be chosen.
7
If the row is of N type, bi is considered as a constant term of the corresponding linear form. Should note, however,
this convention is non-standard.
151
B.7 RANGES section
The RANGES section should start with the indicator card, which contains the word RANGES in
the columns 1—6.
Each data card in the RANGES section specifies one or two ranges for double-side constraints,
i.e. for constraints that are of the types L and G at the same time:
li ≤ xi = ai1 xm+1 + ai2 xm+2 + . . . + ain xm+n ≤ ui ,
where li is a lower bound, ui is an upper bound. All these data cards have the following format.
The field 2 specifies a name of the range vector8 . If this field is empty, the range vector name
from the immediately preceeding data card is assumed.
The field 4 specifies a range value ri (see the table below) for the row, whose name is specified
in the field 3.
name—range value” for the same range vector.
All range values for the current range vector should be specified before range values for the next
range vector. However, the order of rows in the RANGES section may differ from the order of rows
in the ROWS section.
For each double-side constraint specified in the RANGES section its lower and upper bounds
are determined as follows:
Row type Sign of ri Lower bound Upper bound
G + or − bi bi + |ri |
L + or − bi − |ri | bi
E + bi bi + |ri |
E − bi − |ri | bi
where bi is a right-hand side specified in the RHS section (if bi is not specified, it is considered as
zero), ri is a range value specified in the RANGES section.
B.8 BOUNDS section
The BOUNDS section should start with the indicator card, which contains the word BOUNDS in
the columns 1—6.
Each data card in the BOUNDS section specifies one (lower or upper) bound for one structural
variable (column). All these data cards have the following format.
The indicator in the field 1 specifies the bound type:
LO — lower bound;
UP — upper bound;
8
This feature allows the user to specify several range vectors in the same MPS file. However, before solving the
problem a particular range vector should be chosen.
152
FX — fixed variable (lower and upper bounds are equal);
FR — free variable (no bounds);
MI — no lower bound (lower bound is “minus infinity”);
PL — no upper bound (upper bound is “plus infinity”).
The field 2 specifies a name of the bound vector9 . If this field is empty, the bound vector name
from the immediately preceeding data card is assumed.
The field 3 specifies a column name defined in the COLUMNS section.
The field 4 specifies a bound value. If the bound type in the field 1 differs from LO, UP, and FX,
the value in the field 4 is ignored and may be omitted.
The fields 5 and 6 are not used and should be empty.
All bound values for the current bound vector should be specified before bound values for the
next bound vector. However, the order of columns in the BOUNDS section may differ from the order
of columns in the COLUMNS section. Specification of a lower bound should precede specification of
an upper bound for the same column (if both the lower and upper bounds are explicitly specified).
By default, all columns (structural variables) are non-negative, i.e. have zero lower bound and
no upper bound. Lower (lj ) and upper (uj ) bounds of some column (structural variable xj ) are set
in the following way, where sj is a corresponding bound value explicitly specified in the BOUNDS
section:
LO sets lj to sj ;
UP sets uj to sj ;
FX sets both lj and uj to sj ;
FR sets lj to −∞ and uj to +∞;
MI sets lj to −∞;
PL sets uj to +∞.
B.9 ENDATA indicator card
The ENDATA indicator card should be the last card of MPS file (except optional comment
cards, which may follow the ENDATA card). This card should contain the word ENDATA in the
columns 1—6.
B.10 Specifying objective function
It is impossible to explicitly specify the objective function and optimization direction in the MPS
file. However, the following implicit rule is used by default: the first row of N type is considered as
a row of the objective function (i.e. the objective function is the corresponding auxiliary variable),
which should be minimized.
9
This feature allows the user to specify several bound vectors in the same MPS file. However, before solving the
problem a particular bound vector should be chosen.
153
GLPK also allows specifying a constant term of the objective function as a right-hand side of
the corresponding row in the RHS section.
B.11 Example of MPS file
To illustrate what the MPS format is, consider the following example of LP problem:
minimize
value = .03 bin1 + .08 bin2 + .17 bin3 + .12 bin4 + .15 bin5 + .21 al + .38 si
yield = bin1 + bin2 + bin3 + bin4 + bin5 + al + si

F E = .15 bin1 + .04 bin2 + .02 bin3 + .04 bin4 + .02 bin5 + .01 al + .03 si
CU = .03 bin1 + .05 bin2 + .08 bin3 + .02 bin4 + .06 bin5 + .01 al
M N = .02 bin1 + .04 bin2 + .01 bin3 + .02 bin4 + .02 bin5
M G = .02 bin1 + .03 bin2 + .01 bin5
AL = .70 bin1 + .75 bin2 + .80 bin3 + .75 bin4 + .80 bin5 + .97 al
SI = .02 bin1 + .06 bin2 + .08 bin3 + .12 bin4 + .02 bin5 + .01 al + .97 si
and bounds of (auxiliary and structural) variables
yield = 2000 0 ≤ bin1 ≤ 200

−∞ < FE ≤ 60 0 ≤ bin2 ≤ 2500
−∞ < CU ≤ 100 400 ≤ bin3 ≤ 800
−∞ < MN ≤ 40 100 ≤ bin4 ≤ 700
−∞ < MG ≤ 30 0 ≤ bin5 ≤ 1500
1500 ≤ AL < +∞ 0 ≤ al < +∞
250 ≤ SI ≤ 300 0 ≤ si < +∞
A complete MPS file which specifies data for this example is shown below (the first two comment
lines show card positions).
154
*000000001111111111222222222233333333334444444444555555555566
*234567890123456789012345678901234567890123456789012345678901
NAME PLAN
ROWS
N VALUE
E YIELD
L FE
L CU
L MN
L MG
G AL
L SI
COLUMNS
BIN1 VALUE .03000 YIELD 1.00000
FE .15000 CU .03000
MN .02000 MG .02000
AL .70000 SI .02000
FE .04000 CU .05000
MN .04000 MG .03000
AL .75000 SI .06000
FE .02000 CU .08000
MN .01000 AL .80000
SI .08000
FE .04000 CU .02000
MN .02000 AL .75000
SI .12000
FE .02000 CU .06000
MN .02000 MG .01000
AL .80000 SI .02000
ALUM VALUE .21000 YIELD 1.00000
FE .01000 CU .01000
AL .97000 SI .01000
SILICON VALUE .38000 YIELD 1.00000
FE .03000 SI .97000
RHS
RHS1 YIELD 2000.00000 FE 60.00000
CU 100.00000 MN 40.00000
SI 300.00000
MG 30.00000 AL 1500.00000
RANGES
RNG1 SI 50.00000
BOUNDS
UP BND1 BIN1 200.00000
UP BIN2 2500.00000
LO BIN3 400.00000
UP BIN3 800.00000
LO BIN4 100.00000
UP BIN4 700.00000
UP BIN5 1500.00000
ENDATA
155
B.12 MIP features
The MPS format provides two ways for introducing integer variables into the problem.
The first way is most general and based on using special marker cards INTORG and INTEND.
These marker cards are placed in the COLUMNS section. The INTORG card indicates the start
of a group of integer variables (columns), and the card INTEND indicates the end of the group.
The MPS file may contain arbitrary number of the marker cards.
The marker cards have the same format as the data cards (see Section B.1, page 148).
The fields 1, 2, and 6 are not used and should be empty.
The field 2 should contain a marker name. This name may be arbitrary.
The field 3 should contain the word ’MARKER’ (including apostrophes).
The field 5 should contain either the word ’INTORG’ (including apostrophes) for the marker
card, which begins a group of integer columns, or the word ’INTEND’ (including apostrophes) for
the marker card, which ends the group.
The second way is less general but more convenient in some cases. It allows the user declaring
integer columns using three additional types of bounds, which are specified in the field 1 of data
cards in the BOUNDS section (see Section B.8, page 152):
LI — lower integer. This bound type specifies that the corresponding column (structural vari-
able), whose name is specified in field 3, is of integer kind. In this case an lower bound of the
column should be specified in field 4 (like in the case of LO bound type).
UI — upper integer. This bound type specifies that the corresponding column (structural
variable), whose name is specified in field 3, is of integer kind. In this case an upper bound of the
column should be specified in field 4 (like in the case of UP bound type).
BV — binary variable. This bound type specifies that the corresponding column (structural
variable), whose name is specified in the field 3, is of integer kind, its lower bound is zero, and its
upper bound is one (thus, such variable being of integer kind can have only two values zero and
one). In this case a numeric value specified in the field 4 is ignored and may be omitted.
Consider the following example of MIP problem:
minimize
Z = 3x1 + 7x2 − x3 + x4
r1 = 2x1 − x2 + x3 − x4
r2 = x1 − x2 − 6x3 + 4x4
r3 = 5x1 + 3x2 + x4
and bound of variables

1 ≤ r1 < +∞ 0 ≤ x1 ≤4 (continuous)
8 ≤ r2 < +∞ 2 ≤ x2 ≤5 (integer)
5 ≤ r3 < +∞ 0 ≤ x3 ≤1 (integer)
3 ≤ x4 ≤8 (continuous)
The corresponding MPS file may look like follows:
156
NAME SAMP1
ROWS
N Z
G R1
G R2
G R3
COLUMNS
X1 R1 2.0 R2 1.0
X1 R3 5.0 Z 3.0
MARK0001 ’MARKER’ ’INTORG’
X2 R1 -1.0 R2 -1.0
X2 R3 3.0 Z 7.0
X3 R1 1.0 R2 -6.0
X3 Z -1.0
MARK0002 ’MARKER’ ’INTEND’
X4 R1 -1.0 R2 4.0
X4 R3 1.0 Z 1.0
RHS
RHS1 R1 1.0
RHS1 R2 8.0
RHS1 R3 5.0
BOUNDS
UP BND1 X1 4.0
LO BND1 X2 2.0
UP BND1 X2 5.0
UP BND1 X3 1.0
LO BND1 X4 3.0
UP BND1 X4 8.0
ENDATA
The same example may be coded without INTORG/INTEND markers using the bound type
UI for the variable x2 and the bound type BV for the variable x3 :
NAME SAMP2
ROWS
N Z
G R1
G R2
G R3
COLUMNS
X1 R1 2.0 R2 1.0
X1 R3 5.0 Z 3.0
X2 R1 -1.0 R2 -1.0
X2 R3 3.0 Z 7.0
X3 R1 1.0 R2 -6.0
X3 Z -1.0
X4 R1 -1.0 R2 4.0
X4 R3 1.0 Z 1.0
RHS
RHS1 R1 1.0
RHS1 R2 8.0
RHS1 R3 5.0
BOUNDS
UP BND1 X1 4.0
LO BND1 X2 2.0
UI BND1 X2 5.0
BV BND1 X3
LO BND1 X4 3.0
UP BND1 X4 8.0
ENDATA
157
Appendix C
CPLEX LP Format
C.1 Prelude
The CPLEX LP format1 is intended for coding LP/MIP problem data. It is a row-oriented
format that assumes the formulation of LP/MIP problem (1.1)—(1.3) (see Section 1.1, page 9).
CPLEX LP file is a plain text file written in CPLEX LP format. Each text line of this file may
contain up to 255 characters2 . Blank lines are ignored. If a line contains the backslash character
(\), this character and everything that follows it until the end of line are considered as a comment
and also ignored.
An LP file is coded by the user using the following elements: keywords, symbolic names, numeric
constants, delimiters, and blanks.
Keywords which may be used in the LP file are the following:
minimize minimum min
maximize maximum max
subject to such that s.t. st. st
bounds bound
general generals gen
integer integers int
binary binaries bin
infinity inf
free
end
All the keywords are case insensitive. Keywords given above on the same line are equivalent. Any
keyword (except infinity, inf, and free) being used in the LP file must start at the beginning
of a text line.
1
The CPLEX LP format was developed in the end of 1980’s by CPLEX Optimization, Inc. as an input format for
the CPLEX linear programming system. Although the CPLEX LP format is not as widely used as the MPS format,
being row-oriented it is more convenient for coding mathematical programming models by human. This appendix
describes only the features of the CPLEX LP format which are implemented in the GLPK package.
2
GLPK allows text lines of arbitrary length.
158
Symbolic names are used to identify the objective function, constraints (rows), and variables
(columns). All symbolic names are case sensitive and may contain up to 16 alphanumeric charac-
ters3 (a, . . . , z, A, . . . , Z, 0, . . . , 9) as well as the following characters:
! " # $ % & ( ) / , . ; ? @ _ ‘ ’ { } | ~
with exception that no symbolic name can begin with a digit or a period.
Numeric constants are used to denote constraint and objective coefficients, right-hand sides of
constraints, and bounds of variables. They are coded in the standard form xxEsyy, where xx is a
real number with optional decimal point, s is a sign (+ or -), yy is an integer decimal exponent.
Numeric constants may contain arbitrary number of characters. The exponent part is optional.
The letter ‘E’ can be coded as ‘e’. If the sign s is omitted, plus is assumed.
Delimiters that may be used in the LP file are the following:
:
+
-
< <= =<
> >= =>
=
Delimiters given above on the same line are equivalent. The meaning of the delimiters will be
explained below.
Blanks are non-significant characters. They may be used freely to improve readability of the
LP file. Besides, blanks should be used to separate elements from each other if there is no other
way to do that (for example, to separate a keyword from a following symbolic name).
The order of an LP file is the following:
— objective function definition;
— constraints section;
— bounds section;
— general, integer, and binary sections (can appear in arbitrary order);
— end keyword.
These components are discussed in following sections.
C.2 Objective function definition
The objective function definition must appear first in the LP file. It defines the objective
function and specifies the optimization direction.
The objective function definition has the following form:
{ }
minimize
f : s c x ... s c x
maximize
where f is a symbolic name of the objective function, s is a sign + or -, c is a numeric constant
that denotes an objective coefficient, x is a symbolic name of a variable.
3
GLPK allows symbolic names having up to 255 characters.
159
If necessary, the objective function definition can be continued on as many text lines as desired.
The name of the objective function is optional and may be omitted (together with the semicolon
that follows it). In this case the default name ‘obj’ is assigned to the objective function.
If the very first sign s is omitted, the sign plus is assumed. Other signs cannot be omitted.
If some objective coefficient c is omitted, 1 is assumed.
Symbolic names x used to denote variables are recognized by context and therefore needn’t to
be declared somewhere else.
Here is an example of the objective function definition:
Minimize Z : - x1 + 2 x2 - 3.5 x3 + 4.997e3x(4) + x5 + x6 +
x7 - .01x8
C.3 Constraints section
The constraints section must follow the objective function definition. It defines a system of
equality and/or inequality constraints.
The constraint section has the following form:
subject to
constraint1
...
constraintm
where constrainti , i = 1, . . . , m, is a particular constraint definition.
Each constraint definition can be continued on as many text lines as desired. However, each
constraint definition must begin on a new line except the very first constraint definition which can
begin on the same line as the keyword ‘subject to’.
Constraint definitions have the following form:
 
<=
r : s c x . . . s c x >= b
 
=
where r is a symbolic name of a constraint, s is a sign + or -, c is a numeric constant that denotes
a constraint coefficient, x is a symbolic name of a variable, b is a right-hand side.
The name r of a constraint (which is the name of the corresponding auxiliary variable) is
optional and may be omitted together with the semicolon that follows it. In the latter case the
default names like ‘r.nnn’ are assigned to unnamed constraints.
The linear form s c x . . . s c x in the left-hand side of a constraint definition has exactly the
same meaning as in the case of the objective function definition (see above).
After the linear form one of the following delimiters that indicates the constraint sense must be
specified:
<= means ‘less than or equal to’
>= means ‘greater than or equal to’
= means ‘equal to’
160
The right hand side b is a numeric constant with an optional sign.
Here is an example of the constraints section:
Subject To
one: y1 + 3 a1 - a2 - b >= 1.5
y2 + 2 a3 + 2
a4 - b >= -1.5
two : y4 + 3 a1 + 4 a5 - b <= +1
.20y5 + 5 a2 - b = 0
1.7 y6 - a6 + 5 a777 - b >= 1
Should note that it is impossible to express ranged constraints in the CPLEX LP format. Each
a ranged constraint can be coded as two constraints with identical linear forms in the left-hand side,
one of which specifies a lower bound and other does an upper one of the original ranged constraint.
Another way is to introduce a slack double-bounded variable; for example, the constraint
10 ≤ x + 2y + 3z ≤ 50
can be written as follows:

x + 2y + 3z + t = 50,
where 0 ≤ t ≤ 40 is a slack variable.
C.4 Bounds section
The bounds section is intended to define bounds of variables. This section is optional; if it is
specified, it must follow the constraints section. If the bound section is omitted, all variables are
assumed to be non-negative (i.e. that they have zero lower bound and no upper bound).
The bounds section has the following form:
bounds
definition1
...
definitionp
where definitionk , k = 1, . . . , p, is a particular bound definition.
Each bound definition must begin on a new line4 except the very first bound definition which
can begin on the same line as the keyword ‘bounds’.
Syntactically constraint definitions can have one of the following six forms:
x >= l specifies a lower bound
l <= x specifies a lower bound
x <= u specifies an upper bound
l <= x <= u specifies both lower and upper bounds
x=t specifies a fixed value
x free specifies free variable
where x is a symbolic name of a variable, l is a numeric constant with an optional sign that defines a
lower bound of the variable or -inf that means that the variable has no lower bound, u is a numeric
4
The GLPK implementation allows several bound definitions to be placed on the same line.
161
constant with an optional sign that defines an upper bound of the variable or +inf that means that
the variable has no upper bound, t is a numeric constant with an optional sign that defines a fixed
value of the variable.
By default all variables are non-negative, i.e. have zero lower bound and no upper bound.
Therefore definitions of these default bounds can be omitted in the bounds section.
Here is an example of the bounds section:
Bounds
-inf <= a1 <= 100
-100 <= a2
b <= 100
x2 = +123.456
x3 free
C.5 General, integer, and binary sections
The general, integer, and binary sections are intended to define some variables as integer or
binary. All these sections are optional and needed only in case of MIP problems. If they are
specified, they must follow the bounds section or, if the latter is omitted, the constraints section.
All the general, integer, and binary sections have the same form as follows:
 
general
integer
 
binary
x1
...
xq
where xk is a symbolic name of variable, k = 1, . . . , q.
Each symbolic name must begin on a new line5 except the very first symbolic name which can
begin on the same line as the keyword ‘general’, ‘integer’, or ‘binary’.
If a variable appears in the general or the integer section, it is assumed to be general integer
variable. If a variable appears in the binary section, it is assumed to be binary variable, i.e. an
integer variable whose lower bound is zero and upper bound is one. (Note that if bounds of a
variable are specified in the bounds section and then the variable appears in the binary section, its
previously specified bounds are ignored.)
Here is an example of the integer section:
Integer
z12
z22
z35
5
The GLPK implementation allows several symbolic names to be placed on the same line.
162
C.6 End keyword
The keyword ‘end’ is intended to end the LP file. It must begin on a separate line and no other
elements (except comments and blank lines) must follow it. Although this keyword is optional, it
is strongly recommended to include it in the LP file.
C.7 Example of CPLEX LP file
Here is a complete example of CPLEX LP file that corresponds to the example given in Section
B.11, page 154.
\* plan.lp *\
Minimize
value: .03 bin1 + .08 bin2 + .17 bin3 + .12 bin4 + .15 bin5 +
.21 alum + .38 silicon
Subject To
yield: bin1 + bin2 + bin3 + bin4 + bin5 +
alum + silicon = 2000
fe: .15 bin1 + .04 bin2 + .02 bin3 + .04 bin4 + .02 bin5 +
.01 alum + .03 silicon <= 60
cu: .03 bin1 + .05 bin2 + .08 bin3 + .02 bin4 + .06 bin5 +
.01 alum <= 100
mn: .02 bin1 + .04 bin2 + .01 bin3 + .02 bin4 + .02 bin5 <= 40
mg: .02 bin1 + .03 bin2 + .01 bin5 <= 30
al: .70 bin1 + .75 bin2 + .80 bin3 + .75 bin4 + .80 bin5 +
.97 alum >= 1500
si1: .02 bin1 + .06 bin2 + .08 bin3 + .12 bin4 + .02 bin5 +
.01 alum + .97 silicon >= 250
si2: .02 bin1 + .06 bin2 + .08 bin3 + .12 bin4 + .02 bin5 +
.01 alum + .97 silicon <= 300
Bounds
bin1 <= 200
bin2 <= 2500
400 <= bin3 <= 800
100 <= bin4 <= 700
bin5 <= 1500
End
\* eof *\
163
Appendix D
Stand-alone LP/MIP Solver
The GLPK package includes the program glpsol, which is a stand-alone LP/MIP solver. This
program can be invoked from the command line to read LP/MIP problem data in any format
supported by GLPK, solve the problem, and write its solution to an output text file.
Usage
glpsol [options. . . ] [filename]
General options
--mps read LP/MIP problem in fixed MPS format
--freemps read LP/MIP problem in free MPS format (default)
--lp read LP/MIP problem in CPLEX LP format
--glp read LP/MIP problem in GLPK format
--math read LP/MIP model written in GNU MathProg modeling
language
-m filename, --model filename
read model section and optional data section from
filename (same as --math)
-d filename, --data filename
read data section from filename (for --math only);
if model file also has data section, it is ignored
-y filename, --display filename
send display output to filename (for --math only);
by default the output is sent to terminal
--seed value initialize pseudo-random number generator used in
MathProg model with specified seed (any integer);
if seed value is ?, some random seed will be used
--mincost read min-cost flow problem in DIMACS format
--maxflow read maximum flow problem in DIMACS format
--cnf read CNF-SAT problem in DIMACS format
--simplex use simplex method (default)
--interior use interior point method (LP only)
-r filename, --read filename
read solution from filename rather to find it with
the solver
164
--min minimization
--max maximization
--scale scale problem (default)
--noscale do not scale problem
-o filename, --output filename
write solution to filename in printable format
-w filename, --write filename
write solution to filename in plain text format
--ranges filename
write sensitivity analysis report to filename in
printable format (simplex only)
--tmlim nnn limit solution time to nnn seconds
--memlim nnn limit available memory to nnn megabytes
--check do not solve problem, check input data only
--name probname change problem name to probname
--wmps filename write problem to filename in fixed MPS format
--wfreemps filename
write problem to filename in free MPS format
--wlp filename write problem to filename in CPLEX LP format
--wglp filename write problem to filename in GLPK format
--wcnf filename write problem to filename in DIMACS CNF-SAT format
--log filename write copy of terminal output to filename
-h, --help display this help information and exit
-v, --version display program version and exit
LP basis factorization options
--luf plain LU factorization (default)
--btf block triangular LU factorization
--ft Forrest-Tomlin update (requires --luf; default)
--cbg Schur complement + Bartels-Golub update
--cgr Schur complement + Givens rotation update
Options specific to the simplex solver
--primal use primal simplex (default)
--dual use dual simplex
--std use standard initial basis of all slacks
--adv use advanced initial basis (default)
--bib use Bixby’s initial basis
--ini filename use as initial basis previously saved with -w
(disables LP presolver)
--steep use steepest edge technique (default)
--nosteep use standard "textbook" pricing
--relax use Harris’ two-pass ratio test (default)
--norelax use standard "textbook" ratio test
--presol use presolver (default; assumes --scale and --adv)
--nopresol do not use presolver
--exact use simplex method based on exact arithmetic
--xcheck check final basis using exact arithmetic
165
Options specific to the interior-point solver
--nord use natural (original) ordering
--qmd use quotient minimum degree ordering
--amd use approximate minimum degree ordering (default)
--symamd use approximate minimum degree ordering
Options specific to the MIP solver
--nomip consider all integer variables as continuous
(allows solving MIP as pure LP)
--first branch on first integer variable
--last branch on last integer variable
--mostf branch on most fractional variable
--drtom branch using heuristic by Driebeck and Tomlin
(default)
--pcost branch using hybrid pseudocost heuristic (may be
useful for hard instances)
--dfs backtrack using depth first search
--bfs backtrack using breadth first search
--bestp backtrack using the best projection heuristic
--bestb backtrack using node with best local bound
(default)
--intopt use MIP presolver (default)
--nointopt do not use MIP presolver
--binarize replace general integer variables by binary ones
(assumes --intopt)
--fpump apply feasibility pump heuristic
--proxy [nnn] apply proximity search heuristic (nnn is time limit
in seconds; default is 60)
--gomory generate Gomory’s mixed integer cuts
--mir generate MIR (mixed integer rounding) cuts
--cover generate mixed cover cuts
--clique generate clique cuts
--cuts generate all cuts above
--mipgap tol set relative mip gap tolerance to tol
--minisat translate integer feasibility problem to CNF-SAT
and solve it with MiniSat solver
--objbnd bound add inequality obj <= bound (minimization) or
obj >= bound (maximization) to integer feasibility
problem (assumes --minisat)
For description of the MPS format see Appendix B, page 148.
For description of the CPLEX LP format see Appendix C, page 158.
For description of the modeling language see the document “Modeling Language GNU Math-
Prog: Language Reference” included in the GLPK distribution.
For description of the DIMACS min-cost flow problem format and DIMACS maximum flow
problem format see the document “GLPK: Graph and Network Routines” included in the GLPK
distribution.
166
Appendix E
External Software Used In GLPK
In the GLPK package there are used some external software listed in this Appendix. Note that
these software are not part of GLPK, but are used with GLPK and included in the distribution.
E.1 AMD
AMD Version 2.2, Copyright ⃝

c 2007 by Timothy A. Davis, Patrick R. Amestoy, and Iain S. Duff.
All Rights Reserved.
Description
AMD is a set of routines for pre-ordering sparse matrices prior to Cholesky or LU factorization,
using the approximate minimum degree ordering algorithm.
License
This library is free software; you can redistribute it and/or modify it under the terms of the
GNU Lesser General Public License as published by the Free Software Foundation; either version
2.1 of the License, or (at your option) any later version.
This library 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 Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public License along with this
library; if not, write to the Free Software Foundation, Inc., 51 Franklin St, Fifth Floor, Boston,
MA 02110-1301 USA.
Permission is hereby granted to use or copy this program under the terms of the GNU LGPL,
provided that the Copyright, this License, and the Availability of the original version is retained on
all copies. User documentation of any code that uses this code or any modified version of this code
must cite the Copyright, this License, the Availability note, and “Used by permission.” Permission
to modify the code and to distribute modified code is granted, provided the Copyright, this License,
and the Availability note are retained, and a notice that the code was modified is included.
AMD is available under alternate licences; contact T. Davis for details.
Availability
http://www.cise.ufl.edu/research/sparse/amd
167
E.2 COLAMD/SYMAMD
COLAMD/SYMAMD Version 2.7, Copyright ⃝

c 1998-2007, Timothy A. Davis, All Rights Reserved.
Description
colamd: an approximate minimum degree column ordering algorithm, for LU factorization of
symmetric or unsymmetric matrices, QR factorization, least squares, interior point methods for
linear programming problems, and other related problems.
symamd: an approximate minimum degree ordering algorithm for Cholesky factorization of
symmetric matrices.
Authors
The authors of the code itself are Stefan I. Larimore and Timothy A. Davis (davis at cise.ufl.edu),
University of Florida. The algorithm was developed in collaboration with John Gilbert, Xerox
PARC, and Esmond Ng, Oak Ridge National Laboratory.
License
This library is free software; you can redistribute it and/or modify it under the terms of the
GNU Lesser General Public License as published by the Free Software Foundation; either version
2.1 of the License, or (at your option) any later version.
This library 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 Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public License along with this
library; if not, write to the Free Software Foundation, Inc., 51 Franklin St, Fifth Floor, Boston,
MA 02110-1301 USA.
Permission is hereby granted to use or copy this program under the terms of the GNU LGPL,
provided that the Copyright, this License, and the Availability of the original version is retained on
all copies. User documentation of any code that uses this code or any modified version of this code
must cite the Copyright, this License, the Availability note, and “Used by permission.” Permission
to modify the code and to distribute modified code is granted, provided the Copyright, this License,
and the Availability note are retained, and a notice that the code was modified is included.
COLAMD is also available under alternate licenses, contact T. Davis for details.
Availability
http://www.cise.ufl.edu/research/sparse/colamd
E.3 MiniSat
MiniSat-C v1.14.1, Copyright ⃝

c 2005, Niklas Sorensson.
Description
MiniSat is a minimalistic implementation of a Chaff-like SAT solver based on the two-literal
watch scheme for fast BCP and clause learning by conflict analysis.
168
License
Permission is hereby granted, free of charge, to any person obtaining a copy of this software
and associated documentation files (the ”Software”), to deal in the Software without restriction,
including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense,
and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do
so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or sub-
stantial portions of the Software.
THE SOFTWARE IS PROVIDED ”AS IS”, WITHOUT WARRANTY OF ANY KIND, EX-
PRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MER-
CHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFT-
WARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
Availability
http://www.cs.chalmers.se/Cs/Research/FormalMethods/MiniSat
E.4 zlib
zlib version 1.2.5, Copyright ⃝

c 1995–2010 Jean-loup Gailly and Mark Adler.
Description
zlib is a general purpose data compression library. All the code is thread safe. The data format
used by the zlib library is described by RFCs (Request for Comments) 1950 to 1952 in the files
rfc1950.txt (zlib format), rfc1951.txt (deflate format) and rfc1952.txt (gzip format).
License
This software is provided ’as-is’, without any express or implied warranty. In no event will the
authors be held liable for any damages arising from the use of this software.
Permission is granted to anyone to use this software for any purpose, including commercial
applications, and to alter it and redistribute it freely, subject to the following restrictions:
1. The origin of this software must not be misrepresented; you must not claim that you wrote
the original software. If you use this software in a product, an acknowledgment in the product
documentation would be appreciated but is not required.
2. Altered source versions must be plainly marked as such, and must not be misrepresented as
being the original software.
3. This notice may not be removed or altered from any source distribution.
Jean-loup Gailly
Mark Adler
Availability
http://www.zlib.net/
169
GNU General Public License
Version 3, 29 June 2007
Copyright ⃝
c 2007 Free Software Foundation, Inc. <http://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it
is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for software and other kinds of works.
The licenses for most software and other practical works are designed to take away your freedom to share and
change the works. By contrast, the GNU General Public License is intended to guarantee your freedom to share
and change all versions of a program–to make sure it remains free software for all its users. We, the Free Software
Foundation, use the GNU General Public License for most of our software; it applies also to any other work released
this way by its authors. You can apply it to your programs, too.
When we speak of free software, we are referring to freedom, not price. Our General Public Licenses are designed
to make sure that you have the freedom to distribute copies of free software (and charge for them if you wish), that
you receive source code or can get it if you want it, that you can change the software or use pieces of it in new free
programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you these rights or asking you to surrender
the rights. Therefore, you have certain responsibilities if you distribute copies of the software, or if you modify it:
responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether gratis or for a fee, you must pass on to the
recipients the same freedoms that you received. You must make sure that they, too, receive or can get the source
code. And you must show them these terms so they know their rights.
Developers that use the GNU GPL protect your rights with two steps: (1) assert copyright on the software, and
(2) offer you this License giving you legal permission to copy, distribute and/or modify it.
For the developers’ and authors’ protection, the GPL clearly explains that there is no warranty for this free
software. For both users’ and authors’ sake, the GPL requires that modified versions be marked as changed, so that
their problems will not be attributed erroneously to authors of previous versions.
Some devices are designed to deny users access to install or run modified versions of the software inside them,
although the manufacturer can do so. This is fundamentally incompatible with the aim of protecting users’ freedom
to change the software. The systematic pattern of such abuse occurs in the area of products for individuals to use,
which is precisely where it is most unacceptable. Therefore, we have designed this version of the GPL to prohibit
the practice for those products. If such problems arise substantially in other domains, we stand ready to extend this
provision to those domains in future versions of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents. States should not allow patents to restrict
development and use of software on general-purpose computers, but in those that do, we wish to avoid the special
danger that patents applied to a free program could make it effectively proprietary. To prevent this, the GPL assures
that patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and modification follow.
170
TERMS AND CONDITIONS
0. Definitions.
“This License” refers to version 3 of the GNU General Public License.
“Copyright” also means copyright-like laws that apply to other kinds of works, such as semiconductor masks.
“The Program” refers to any copyrightable work licensed under this License. Each licensee is addressed as “you”.
“Licensees” and “recipients” may be individuals or organizations.
To “modify” a work means to copy from or adapt all or part of the work in a fashion requiring copyright
permission, other than the making of an exact copy. The resulting work is called a “modified version” of the earlier
work or a work “based on” the earlier work.
A “covered work” means either the unmodified Program or a work based on the Program.
To “propagate” a work means to do anything with it that, without permission, would make you directly or
secondarily liable for infringement under applicable copyright law, except executing it on a computer or modifying
a private copy. Propagation includes copying, distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To “convey” a work means any kind of propagation that enables other parties to make or receive copies. Mere
interaction with a user through a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays “Appropriate Legal Notices” to the extent that it includes a convenient and
prominently visible feature that (1) displays an appropriate copyright notice, and (2) tells the user that there is no
warranty for the work (except to the extent that warranties are provided), that licensees may convey the work under
this License, and how to view a copy of this License. If the interface presents a list of user commands or options,
such as a menu, a prominent item in the list meets this criterion.
1. Source Code.
The “source code” for a work means the preferred form of the work for making modifications to it. “Object code”
means any non-source form of a work.
A “Standard Interface” means an interface that either is an official standard defined by a recognized standards
body, or, in the case of interfaces specified for a particular programming language, one that is widely used among
developers working in that language.
The “System Libraries” of an executable work include anything, other than the work as a whole, that (a) is
included in the normal form of packaging a Major Component, but which is not part of that Major Component, and
(b) serves only to enable use of the work with that Major Component, or to implement a Standard Interface for which
an implementation is available to the public in source code form. A “Major Component”, in this context, means a
major essential component (kernel, window system, and so on) of the specific operating system (if any) on which the
executable work runs, or a compiler used to produce the work, or an object code interpreter used to run it.
The “Corresponding Source” for a work in object code form means all the source code needed to generate,
install, and (for an executable work) run the object code and to modify the work, including scripts to control those
activities. However, it does not include the work’s System Libraries, or general-purpose tools or generally available
free programs which are used unmodified in performing those activities but which are not part of the work. For
example, Corresponding Source includes interface definition files associated with source files for the work, and the
source code for shared libraries and dynamically linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those subprograms and other parts of the work.
The Corresponding Source need not include anything that users can regenerate automatically from other parts
of the Corresponding Source.
The Corresponding Source for a work in source code form is that same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of copyright on the Program, and are irrevocable
provided the stated conditions are met. This License explicitly affirms your unlimited permission to run the unmodi-
fied Program. The output from running a covered work is covered by this License only if the output, given its content,
constitutes a covered work. This License acknowledges your rights of fair use or other equivalent, as provided by
copyright law.
171
You may make, run and propagate covered works that you do not convey, without conditions so long as your
license otherwise remains in force. You may convey covered works to others for the sole purpose of having them make
modifications exclusively for you, or provide you with facilities for running those works, provided that you comply
with the terms of this License in conveying all material for which you do not control copyright. Those thus making
or running the covered works for you must do so exclusively on your behalf, under your direction and control, on
terms that prohibit them from making any copies of your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under the conditions stated below. Sublicensing is
not allowed; section 10 makes it unnecessary.
3. Protecting Users’ Legal Rights From Anti-Circumvention Law.

No covered work shall be deemed part of an effective technological measure under any applicable law fulfilling
obligations under article 11 of the WIPO copyright treaty adopted on 20 December 1996, or similar laws prohibiting
or restricting circumvention of such measures.
When you convey a covered work, you waive any legal power to forbid circumvention of technological measures
to the extent such circumvention is effected by exercising rights under this License with respect to the covered work,
and you disclaim any intention to limit operation or modification of the work as a means of enforcing, against the
work’s users, your or third parties’ legal rights to forbid circumvention of technological measures.
4. Conveying Verbatim Copies.

You may convey verbatim copies of the Program’s source code as you receive it, in any medium, provided that
you conspicuously and appropriately publish on each copy an appropriate copyright notice; keep intact all notices
stating that this License and any non-permissive terms added in accord with section 7 apply to the code; keep intact
all notices of the absence of any warranty; and give all recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey, and you may offer support or warranty
protection for a fee.
5. Conveying Modified Source Versions.

You may convey a work based on the Program, or the modifications to produce it from the Program, in the form
of source code under the terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified it, and giving a relevant date.
b) The work must carry prominent notices stating that it is released under this License and any conditions added
under section 7. This requirement modifies the requirement in section 4 to “keep intact all notices”.
c) You must license the entire work, as a whole, under this License to anyone who comes into possession of a
copy. This License will therefore apply, along with any applicable section 7 additional terms, to the whole of the
work, and all its parts, regardless of how they are packaged. This License gives no permission to license the work in
any other way, but it does not invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display Appropriate Legal Notices; however, if the
Program has interactive interfaces that do not display Appropriate Legal Notices, your work need not make them do
so.
A compilation of a covered work with other separate and independent works, which are not by their nature
extensions of the covered work, and which are not combined with it such as to form a larger program, in or on a
volume of a storage or distribution medium, is called an “aggregate” if the compilation and its resulting copyright
are not used to limit the access or legal rights of the compilation’s users beyond what the individual works permit.
Inclusion of a covered work in an aggregate does not cause this License to apply to the other parts of the aggregate.
6. Conveying Non-Source Forms.

You may convey a covered work in object code form under the terms of sections 4 and 5, provided that you also
convey the machine-readable Corresponding Source under the terms of this License, in one of these ways:
a) Convey the object code in, or embodied in, a physical product (including a physical distribution medium),
accompanied by the Corresponding Source fixed on a durable physical medium customarily used for software inter-
change.
b) Convey the object code in, or embodied in, a physical product (including a physical distribution medium),
accompanied by a written offer, valid for at least three years and valid for as long as you offer spare parts or customer
172
support for that product model, to give anyone who possesses the object code either (1) a copy of the Corresponding
Source for all the software in the product that is covered by this License, on a durable physical medium customarily
used for software interchange, for a price no more than your reasonable cost of physically performing this conveying
of source, or (2) access to copy the Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the written offer to provide the Corresponding
Source. This alternative is allowed only occasionally and noncommercially, and only if you received the object code
with such an offer, in accord with subsection 6b.
d) Convey the object code by offering access from a designated place (gratis or for a charge), and offer equivalent
access to the Corresponding Source in the same way through the same place at no further charge. You need not
require recipients to copy the Corresponding Source along with the object code. If the place to copy the object code
is a network server, the Corresponding Source may be on a different server (operated by you or a third party) that
supports equivalent copying facilities, provided you maintain clear directions next to the object code saying where to
find the Corresponding Source. Regardless of what server hosts the Corresponding Source, you remain obligated to
ensure that it is available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided you inform other peers where the object
code and Corresponding Source of the work are being offered to the general public at no charge under subsection 6d.
A separable portion of the object code, whose source code is excluded from the Corresponding Source as a System
Library, need not be included in conveying the object code work.
A “User Product” is either (1) a “consumer product”, which means any tangible personal property which is
normally used for personal, family, or household purposes, or (2) anything designed or sold for incorporation into
a dwelling. In determining whether a product is a consumer product, doubtful cases shall be resolved in favor of
coverage. For a particular product received by a particular user, “normally used” refers to a typical or common use of
that class of product, regardless of the status of the particular user or of the way in which the particular user actually
uses, or expects or is expected to use, the product. A product is a consumer product regardless of whether the
product has substantial commercial, industrial or non-consumer uses, unless such uses represent the only significant
mode of use of the product.
“Installation Information” for a User Product means any methods, procedures, authorization keys, or other
information required to install and execute modified versions of a covered work in that User Product from a modified
version of its Corresponding Source. The information must suffice to ensure that the continued functioning of the
modified object code is in no case prevented or interfered with solely because modification has been made.
If you convey an object code work under this section in, or with, or specifically for use in, a User Product, and the
conveying occurs as part of a transaction in which the right of possession and use of the User Product is transferred to
the recipient in perpetuity or for a fixed term (regardless of how the transaction is characterized), the Corresponding
Source conveyed under this section must be accompanied by the Installation Information. But this requirement does
not apply if neither you nor any third party retains the ability to install modified object code on the User Product
(for example, the work has been installed in ROM).
The requirement to provide Installation Information does not include a requirement to continue to provide support
service, warranty, or updates for a work that has been modified or installed by the recipient, or for the User Product
in which it has been modified or installed. Access to a network may be denied when the modification itself materially
and adversely affects the operation of the network or violates the rules and protocols for communication across the
network.
Corresponding Source conveyed, and Installation Information provided, in accord with this section must be in a
format that is publicly documented (and with an implementation available to the public in source code form), and
must require no special password or key for unpacking, reading or copying.
7. Additional Terms.
“Additional permissions” are terms that supplement the terms of this License by making exceptions from one or
more of its conditions. Additional permissions that are applicable to the entire Program shall be treated as though
they were included in this License, to the extent that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately under those permissions, but the entire Program
remains governed by this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option remove any additional permissions from that
copy, or from any part of it. (Additional permissions may be written to require their own removal in certain cases
when you modify the work.) You may place additional permissions on material, added by you to a covered work, for
173
which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you add to a covered work, you may (if authorized
by the copyright holders of that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or author attributions in that material or in the
Appropriate Legal Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or requiring that modified versions of such material
be marked in reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or authors of the material; or
e) Declining to grant rights under trademark law for use of some trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that material by anyone who conveys the material (or
modified versions of it) with contractual assumptions of liability to the recipient, for any liability that these contractual
assumptions directly impose on those licensors and authors.
All other non-permissive additional terms are considered “further restrictions” within the meaning of section 10.
If the Program as you received it, or any part of it, contains a notice stating that it is governed by this License
along with a term that is a further restriction, you may remove that term. If a license document contains a further
restriction but permits relicensing or conveying under this License, you may add to a covered work material governed
by the terms of that license document, provided that the further restriction does not survive such relicensing or
conveying.
If you add terms to a covered work in accord with this section, you must place, in the relevant source files, a
statement of the additional terms that apply to those files, or a notice indicating where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the form of a separately written license, or
stated as exceptions; the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly provided under this License. Any attempt
otherwise to propagate or modify it is void, and will automatically terminate your rights under this License (including
any patent licenses granted under the third paragraph of section 11).
However, if you cease all violation of this License, then your license from a particular copyright holder is rein-
stated (a) provisionally, unless and until the copyright holder explicitly and finally terminates your license, and (b)
permanently, if the copyright holder fails to notify you of the violation by some reasonable means prior to 60 days
after the cessation.
Moreover, your license from a particular copyright holder is reinstated permanently if the copyright holder notifies
you of the violation by some reasonable means, this is the first time you have received notice of violation of this License
(for any work) from that copyright holder, and you cure the violation prior to 30 days after your receipt of the notice.
Termination of your rights under this section does not terminate the licenses of parties who have received copies
or rights from you under this License. If your rights have been terminated and not permanently reinstated, you do
not qualify to receive new licenses for the same material under section 10.
9. Acceptance Not Required for Having Copies.

You are not required to accept this License in order to receive or run a copy of the Program. Ancillary propagation
of a covered work occurring solely as a consequence of using peer-to-peer transmission to receive a copy likewise does
not require acceptance. However, nothing other than this License grants you permission to propagate or modify
any covered work. These actions infringe copyright if you do not accept this License. Therefore, by modifying or
propagating a covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.

Each time you convey a covered work, the recipient automatically receives a license from the original licensors,
to run, modify and propagate that work, subject to this License. You are not responsible for enforcing compliance
by third parties with this License.
An “entity transaction” is a transaction transferring control of an organization, or substantially all assets of one,
174
or subdividing an organization, or merging organizations. If propagation of a covered work results from an entity
transaction, each party to that transaction who receives a copy of the work also receives whatever licenses to the
work the party’s predecessor in interest had or could give under the previous paragraph, plus a right to possession
of the Corresponding Source of the work from the predecessor in interest, if the predecessor has it or can get it with
reasonable efforts.
You may not impose any further restrictions on the exercise of the rights granted or affirmed under this License.
For example, you may not impose a license fee, royalty, or other charge for exercise of rights granted under this
License, and you may not initiate litigation (including a cross-claim or counterclaim in a lawsuit) alleging that any
patent claim is infringed by making, using, selling, offering for sale, or importing the Program or any portion of it.
11. Patents.
A “contributor” is a copyright holder who authorizes use under this License of the Program or a work on which
the Program is based. The work thus licensed is called the contributor’s “contributor version”.
A contributor’s “essential patent claims” are all patent claims owned or controlled by the contributor, whether
already acquired or hereafter acquired, that would be infringed by some manner, permitted by this License, of making,
using, or selling its contributor version, but do not include claims that would be infringed only as a consequence of
further modification of the contributor version. For purposes of this definition, “control” includes the right to grant
patent sublicenses in a manner consistent with the requirements of this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free patent license under the contributor’s es-
sential patent claims, to make, use, sell, offer for sale, import and otherwise run, modify and propagate the contents
of its contributor version.
In the following three paragraphs, a “patent license” is any express agreement or commitment, however denomi-
nated, not to enforce a patent (such as an express permission to practice a patent or covenant not to sue for patent
infringement). To “grant” such a patent license to a party means to make such an agreement or commitment not to
enforce a patent against the party.
If you convey a covered work, knowingly relying on a patent license, and the Corresponding Source of the work
is not available for anyone to copy, free of charge and under the terms of this License, through a publicly available
network server or other readily accessible means, then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the patent license for this particular work, or (3) arrange,
in a manner consistent with the requirements of this License, to extend the patent license to downstream recipients.
“Knowingly relying” means you have actual knowledge that, but for the patent license, your conveying the covered
work in a country, or your recipient’s use of the covered work in a country, would infringe one or more identifiable
patents in that country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or arrangement, you convey, or propagate by procuring
conveyance of, a covered work, and grant a patent license to some of the parties receiving the covered work authorizing
them to use, propagate, modify or convey a specific copy of the covered work, then the patent license you grant is
automatically extended to all recipients of the covered work and works based on it.
A patent license is “discriminatory” if it does not include within the scope of its coverage, prohibits the exercise
of, or is conditioned on the non-exercise of one or more of the rights that are specifically granted under this License.
You may not convey a covered work if you are a party to an arrangement with a third party that is in the business
of distributing software, under which you make payment to the third party based on the extent of your activity of
conveying the work, and under which the third party grants, to any of the parties who would receive the covered
work from you, a discriminatory patent license (a) in connection with copies of the covered work conveyed by you (or
copies made from those copies), or (b) primarily for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement, or that patent license was granted, prior to 28
March 2007.
Nothing in this License shall be construed as excluding or limiting any implied license or other defenses to
infringement that may otherwise be available to you under applicable patent law.
12. No Surrender of Others’ Freedom.

If conditions are imposed on you (whether by court order, agreement or otherwise) that contradict the conditions
of this License, they do not excuse you from the conditions of this License. If you cannot convey a covered work
so as to satisfy simultaneously your obligations under this License and any other pertinent obligations, then as a
consequence you may not convey it at all. For example, if you agree to terms that obligate you to collect a royalty
175
for further conveying from those to whom you convey the Program, the only way you could satisfy both those terms
and this License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.

Notwithstanding any other provision of this License, you have permission to link or combine any covered work
with a work licensed under version 3 of the GNU Affero General Public License into a single combined work, and to
convey the resulting work. The terms of this License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License, section 13, concerning interaction through a
network will apply to the combination as such.
14. Revised Versions of this License.

The Free Software Foundation may publish revised and/or new versions of the GNU General Public License from
time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address
new problems or concerns.
Each version is given a distinguishing version number. If the Program specifies that a certain numbered version
of the GNU General Public License “or any later version” applies to it, you have the option of following the terms
and conditions either of that numbered version or of any later version published by the Free Software Foundation.
If the Program does not specify a version number of the GNU General Public License, you may choose any version
ever published by the Free Software Foundation.
If the Program specifies that a proxy can decide which future versions of the GNU General Public License can
be used, that proxy’s public statement of acceptance of a version permanently authorizes you to choose that version
for the Program.
Later license versions may give you additional or different permissions. However, no additional obligations are
imposed on any author or copyright holder as a result of your choosing to follow a later version.
15. Disclaimer of Warranty.

THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY APPLICABLE
LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER
PARTIES PROVIDE THE PROGRAM “AS IS” WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED
OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY
AND FITNESS FOR A PARTICULAR PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFOR-
MANCE OF THE PROGRAM IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME
THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.

IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING WILL ANY
COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS THE PROGRAM
AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY GENERAL, SPECIAL,
INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE THE
PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF DATA OR DATA BEING RENDERED INAC-
CURATE OR LOSSES SUSTAINED BY YOU OR THIRD PARTIES OR A FAILURE OF THE PROGRAM TO
OPERATE WITH ANY OTHER PROGRAMS), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN
ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.

If the disclaimer of warranty and limitation of liability provided above cannot be given local legal effect according
to their terms, reviewing courts shall apply local law that most closely approximates an absolute waiver of all civil
liability in connection with the Program, unless a warranty or assumption of liability accompanies a copy of the
Program in return for a fee.
END OF TERMS AND CONDITIONS
176
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest possible use to the public, the best way to
achieve this is to make it free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest to attach them to the start of each source file to
most effectively state the exclusion of warranty; and each file should have at least the “copyright” line and a pointer
to where the full notice is found.
<one line to give the program’s name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program 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.
This program 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 this program. If not, see <http://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short notice like this when it starts in an interactive
mode:
<program> Copyright (C) <year> <name of author>
This program comes with ABSOLUTELY NO WARRANTY; for details type ‘show w’.
This is free software, and you are welcome to redistribute it
under certain conditions; type ‘show c’ for details.
The hypothetical commands ‘show w’ and ‘show c’ should show the appropriate parts of the General Public License.
Of course, your program’s commands might be different; for a GUI interface, you would use an “about box”.
You should also get your employer (if you work as a programmer) or school, if any, to sign a “copyright disclaimer”
for the program, if necessary. For more information on this, and how to apply and follow the GNU GPL, see
<http://www.gnu.org/licenses/>.
The GNU General Public License does not permit incorporating your program into proprietary programs. If your
program is a subroutine library, you may consider it more useful to permit linking proprietary applications with the
library. If this is what you want to do, use the GNU Lesser General Public License instead of this License. But first,
please read <http://www.gnu.org/philosophy/why-not-lgpl.html>.
177

GNU Linear Programming Kit: Reference Manual

Uploaded by

Copyright:

Available Formats

GNU Linear Programming Kit: Reference Manual

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

GNU Linear Programming Kit: Reference Manual

Uploaded by

Copyright:

Available Formats

M L

GNU Linear Programming Kit

for GLPK Version 4.64

(DRAFT, November 2017)

2 Basic API Routines 14

3 Utility API routines 69

4 Advanced API Routines 98

5 Branch-and-Cut API Routines 119

6 Miscellaneous API Routines 137

A Installing GLPK on Your Computer 144

B MPS Format 148

C CPLEX LP Format 158

D Stand-alone LP/MIP Solver 164

E External Software Used In GLPK 167

GNU General Public License 170

z = c1 xm+1 + c2 xm+2 + . . . + cn xm+n + c0 (1.1)

subject to linear constraints

x1 = a11 xm+1 + a12 xm+2 + . . . + a1n xm+n

and bounds of variables

1.2 MIP problem

1.3 Using the package

1.3.1 Brief example

z = 733.333; x1 = 33.3333; x2 = 66.6667; x3 = 0

Basic API Routines

2.1 General conventions

2.1.1 Library header

2.1.2 Error handling

2.1.3 Thread safety

2.1.4 Array indexing

2.2 Problem object

2.2.1 Problem segment

2.2.2 Basis segment

2.2.3 Interior-point segment

2.2.4 MIP segment

2.3.1 glp create prob — create problem object

2.3.2 glp set prob name — assign (change) problem name

2.3.5 glp add rows — add new rows to problem object

2.3.6 glp add cols — add new columns to problem object

2.3.7 glp set row name — assign (change) row name

2.3.8 glp set col name — assign (change) column name

2.3.9 glp set row bnds — set (change) row bounds

2.3.10 glp set col bnds — set (change) column bounds

2.3.17 glp del rows — delete rows from problem object

2.3.18 glp del cols — delete columns from problem object

2.3.20 glp erase prob — erase problem object content

2.3.21 glp delete prob — delete problem object

2.4.1 glp get prob name — retrieve problem name

2.4.2 glp get obj name — retrieve objective function name

2.4.3 glp get obj dir — retrieve optimization direction flag

2.4.4 glp get num rows — retrieve number of rows

2.4.6 glp get row name — retrieve row name

2.4.7 glp get col name — retrieve column name

2.4.8 glp get row type — retrieve row type

2.4.10 glp get row ub — retrieve row upper bound

2.4.11 glp get col type — retrieve column type

2.4.12 glp get col lb — retrieve column lower bound

2.4.15 glp get num nz — retrieve number of constraint coeﬃcients

2.5.1 glp create index — create the name index

2.5.3 glp find col — find column by its name

void (cb func)(glp tree tree, void *info) (default: NULL)