Contents

Appendix C C Programs 1007

C.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1007
C.2 Roundoff Error . . . . . . . . . . . . . . . . . . . . . . . . . . 1011
1 CASSUM . . . . . . . . . . . . . . . . . . . . . . . . . . 1011
2 ROUND . . . . . . . . . . . . . . . . . . . . . . . . . . . 1011
3 CDIV . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1012
4 CABS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1012
5 CSQRT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1012
C.3 Linear Algebraic Equations . . . . . . . . . . . . . . . . . . . 1012
6 GAUELM . . . . . . . . . . . . . . . . . . . . . . . . . . 1012
7 MATINV . . . . . . . . . . . . . . . . . . . . . . . . . . 1013
8 CROUT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1013
9 CROUTH . . . . . . . . . . . . . . . . . . . . . . . . . . 1014
10 CHOLSK . . . . . . . . . . . . . . . . . . . . . . . . . . 1016
11 GAUBND . . . . . . . . . . . . . . . . . . . . . . . . . . 1017
12 SVD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1018
13 SVDEVL . . . . . . . . . . . . . . . . . . . . . . . . . . 1018
C.4 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1019
14 DIVDIF . . . . . . . . . . . . . . . . . . . . . . . . . . . 1019
15 DIVDIF0 . . . . . . . . . . . . . . . . . . . . . . . . . . 1020
16 NEARST . . . . . . . . . . . . . . . . . . . . . . . . . . 1020
17 SPLINE . . . . . . . . . . . . . . . . . . . . . . . . . . . 1021
18 SPLEVL . . . . . . . . . . . . . . . . . . . . . . . . . . 1021
19 SMOOTH . . . . . . . . . . . . . . . . . . . . . . . . . . 1021
20 BSPLIN . . . . . . . . . . . . . . . . . . . . . . . . . . . 1022
21 BSPINT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1022
22 BSPEVL . . . . . . . . . . . . . . . . . . . . . . . . . . 1023
23 RATNAL . . . . . . . . . . . . . . . . . . . . . . . . . . 1024
24 POLY2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1024
25 LINRN . . . . . . . . . . . . . . . . . . . . . . . . . . . 1025
26 LOCATE . . . . . . . . . . . . . . . . . . . . . . . . . . 1025
27 BSPINT2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1025
28 BSPEV2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1026
29 BSPINTN . . . . . . . . . . . . . . . . . . . . . . . . . . 1027
ii Contents

30 BSPEVN . . . . . . . . . . . . . . . . . . . . . . . . . . 1028
31 BSPEVN1 . . . . . . . . . . . . . . . . . . . . . . . . . . 1029
32 BSPEVN2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1029
C.5 Differentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1029
33 DRVT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1029
C.6 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1030
34 SIMSON . . . . . . . . . . . . . . . . . . . . . . . . . . 1030
35 SPLINT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1030
36 BSPQD . . . . . . . . . . . . . . . . . . . . . . . . . . . 1031
37 ROMBRG . . . . . . . . . . . . . . . . . . . . . . . . . . 1031
38 EPSILN . . . . . . . . . . . . . . . . . . . . . . . . . . . 1032
39 GAUSS . . . . . . . . . . . . . . . . . . . . . . . . . . . 1032
40 GAUCBY . . . . . . . . . . . . . . . . . . . . . . . . . . 1033
41 GAUCB1 . . . . . . . . . . . . . . . . . . . . . . . . . . 1033
42 GAUCB2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1034
43 GAUSQ2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1034
44 GAUSQ . . . . . . . . . . . . . . . . . . . . . . . . . . . 1035
45 GAULAG . . . . . . . . . . . . . . . . . . . . . . . . . . 1036
46 LAGURE . . . . . . . . . . . . . . . . . . . . . . . . . . 1037
47 HERMIT . . . . . . . . . . . . . . . . . . . . . . . . . . 1037
48 GAULG2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1038
49 GAULOG . . . . . . . . . . . . . . . . . . . . . . . . . . 1039
50 GAUSRC . . . . . . . . . . . . . . . . . . . . . . . . . . 1039
51 GAULEG . . . . . . . . . . . . . . . . . . . . . . . . . . 1040
52 GAUJAC . . . . . . . . . . . . . . . . . . . . . . . . . . 1040
53 LAGURW . . . . . . . . . . . . . . . . . . . . . . . . . . 1041
54 GAUHER . . . . . . . . . . . . . . . . . . . . . . . . . . 1041
55 GAUSWT . . . . . . . . . . . . . . . . . . . . . . . . . . 1041
56 FILON . . . . . . . . . . . . . . . . . . . . . . . . . . . 1042
57 ADPINT . . . . . . . . . . . . . . . . . . . . . . . . . . 1043
58 KRONRD . . . . . . . . . . . . . . . . . . . . . . . . . . 1044
59 GAUS16 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1044
60 CAUCHY . . . . . . . . . . . . . . . . . . . . . . . . . . 1044
61 EULER . . . . . . . . . . . . . . . . . . . . . . . . . . . 1045
62 BSPQD2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1046
63 BSPQDN . . . . . . . . . . . . . . . . . . . . . . . . . . 1047
64 MULINT . . . . . . . . . . . . . . . . . . . . . . . . . . 1047
65 NGAUSS . . . . . . . . . . . . . . . . . . . . . . . . . . 1048
66 SPHND . . . . . . . . . . . . . . . . . . . . . . . . . . . 1049
67 STRINT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1050
68 STROUD . . . . . . . . . . . . . . . . . . . . . . . . . . 1051
69 MCARLO . . . . . . . . . . . . . . . . . . . . . . . . . . 1051
70 RAN1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1052
71 RANF . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1052
72 EQUIDS . . . . . . . . . . . . . . . . . . . . . . . . . . 1052
Contents iii

C.7 Nonlinear Algebraic Equations . . . . . . . . . . . . . . . . . . 1053

73 BISECT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1053
74 SECANT . . . . . . . . . . . . . . . . . . . . . . . . . . 1053
75 SECANI . . . . . . . . . . . . . . . . . . . . . . . . . . . 1054
76 NEWRAP . . . . . . . . . . . . . . . . . . . . . . . . . . 1055
77 BRENT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1056
78 SEARCH . . . . . . . . . . . . . . . . . . . . . . . . . . 1056
79 ZROOT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1057
80 ZROOT2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1057
81 MULLER . . . . . . . . . . . . . . . . . . . . . . . . . . 1058
82 MULER2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1059
83 POLYR . . . . . . . . . . . . . . . . . . . . . . . . . . . 1060
84 LAGITR . . . . . . . . . . . . . . . . . . . . . . . . . . 1061
85 DAVIDN . . . . . . . . . . . . . . . . . . . . . . . . . . 1061
86 NEWTON . . . . . . . . . . . . . . . . . . . . . . . . . 1063
87 BROYDN . . . . . . . . . . . . . . . . . . . . . . . . . . 1063
C.8 Optimisation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1064
88 BRACKM . . . . . . . . . . . . . . . . . . . . . . . . . . 1064
89 GOLDEN . . . . . . . . . . . . . . . . . . . . . . . . . . 1064
90 BRENTM . . . . . . . . . . . . . . . . . . . . . . . . . . 1064
91 DAVIDM . . . . . . . . . . . . . . . . . . . . . . . . . . 1065
92 BFGS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1066
93 LINMIN . . . . . . . . . . . . . . . . . . . . . . . . . . . 1066
94 FLNM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1067
95 NMINF . . . . . . . . . . . . . . . . . . . . . . . . . . . 1068
96 LINMNF . . . . . . . . . . . . . . . . . . . . . . . . . . 1068
97 FLN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1069
98 SIMPLX . . . . . . . . . . . . . . . . . . . . . . . . . . 1069
99 SIMPX . . . . . . . . . . . . . . . . . . . . . . . . . . . 1070
C.9 Statistical Inferences . . . . . . . . . . . . . . . . . . . . . . . 1071
100 SHSORT . . . . . . . . . . . . . . . . . . . . . . . . . . 1071
101 GAMMAP . . . . . . . . . . . . . . . . . . . . . . . . . 1071
102 BETAP . . . . . . . . . . . . . . . . . . . . . . . . . . . 1071
103 BETSER . . . . . . . . . . . . . . . . . . . . . . . . . . 1072
104 BETCON1 . . . . . . . . . . . . . . . . . . . . . . . . . 1072
105 BETCON . . . . . . . . . . . . . . . . . . . . . . . . . . 1073
106 BETAI . . . . . . . . . . . . . . . . . . . . . . . . . . . 1073
107 FBETA . . . . . . . . . . . . . . . . . . . . . . . . . . . 1073
108 RANGAU . . . . . . . . . . . . . . . . . . . . . . . . . . 1073
109 IRANBIN . . . . . . . . . . . . . . . . . . . . . . . . . . 1073
110 IRANPOI . . . . . . . . . . . . . . . . . . . . . . . . . . 1074
111 PCOR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1074
C.10 Functional Approximations . . . . . . . . . . . . . . . . . . . 1074
112 POLFIT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1074
113 POLEVL . . . . . . . . . . . . . . . . . . . . . . . . . . 1075
iv Contents

114 POLFIT1 . . . . . . . . . . . . . . . . . . . . . . . . . . 1075

115 POLORT . . . . . . . . . . . . . . . . . . . . . . . . . . 1076
116 POLFIT2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1076
117 POLEV2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1077
118 POLFITN . . . . . . . . . . . . . . . . . . . . . . . . . . 1077
119 POLEVN . . . . . . . . . . . . . . . . . . . . . . . . . . 1078
120 POLEVN1 . . . . . . . . . . . . . . . . . . . . . . . . . 1078
121 POLEVN2 . . . . . . . . . . . . . . . . . . . . . . . . . 1079
122 LLSQ . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1079
123 BSPFIT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1080
124 BSPFIT2 . . . . . . . . . . . . . . . . . . . . . . . . . . 1082
125 BSPFITW2 . . . . . . . . . . . . . . . . . . . . . . . . . 1083
126 BSPFITN . . . . . . . . . . . . . . . . . . . . . . . . . . 1084
127 BSPFITWN . . . . . . . . . . . . . . . . . . . . . . . . 1086
128 LINFITXY . . . . . . . . . . . . . . . . . . . . . . . . . 1087
129 NLLSQ . . . . . . . . . . . . . . . . . . . . . . . . . . . 1087
130 DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1088
131 FFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1088
132 FFTR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1089
133 FFTN . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1089
134 LAPINV . . . . . . . . . . . . . . . . . . . . . . . . . . 1090
135 POLD . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1090
136 RMK . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1090
137 RMK1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1091
138 RMKD . . . . . . . . . . . . . . . . . . . . . . . . . . . 1091
139 RMKD1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1091
140 PADE . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1091
141 CHEBCF . . . . . . . . . . . . . . . . . . . . . . . . . . 1092
142 CHEBEX . . . . . . . . . . . . . . . . . . . . . . . . . . 1092
143 CHEBAP . . . . . . . . . . . . . . . . . . . . . . . . . . 1092
144 REMES . . . . . . . . . . . . . . . . . . . . . . . . . . . 1093
145 FM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1095
146 GAMMA . . . . . . . . . . . . . . . . . . . . . . . . . . 1095
147 GAMMAL . . . . . . . . . . . . . . . . . . . . . . . . . 1095
148 ERF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1095
149 ERFC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1096
150 BJ0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1096
151 BJ1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1096
152 BJN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1097
153 BY0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1098
154 BJY0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1098
155 BY1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1098
156 BJY1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1099
157 BYN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1099
158 SPHBJN . . . . . . . . . . . . . . . . . . . . . . . . . . 1100
Contents v

159 BI0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101

160 BI1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101
161 BIN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101
162 BK0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1102
163 BK1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103
164 BKN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103
165 DAWSON . . . . . . . . . . . . . . . . . . . . . . . . . . 1103
166 FERMM05 . . . . . . . . . . . . . . . . . . . . . . . . . 1103
167 FERM05 . . . . . . . . . . . . . . . . . . . . . . . . . . 1104
168 FERM15 . . . . . . . . . . . . . . . . . . . . . . . . . . 1104
169 FERM25 . . . . . . . . . . . . . . . . . . . . . . . . . . 1104
170 PLEG . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1104
171 PLM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1104
172 YLM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1104
173 MINMAX . . . . . . . . . . . . . . . . . . . . . . . . . . 1105
174 POLYL1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105
175 LINL1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1106
176 SIMPL1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1106
C.11 Algebraic Eigenvalue Problem . . . . . . . . . . . . . . . . . . 1107
177 INVIT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1107
178 TRED2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1108
179 TRBAK . . . . . . . . . . . . . . . . . . . . . . . . . . . 1108
180 TQL2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1109
181 TRIDIA . . . . . . . . . . . . . . . . . . . . . . . . . . . 1109
182 STURM . . . . . . . . . . . . . . . . . . . . . . . . . . . 1110
183 TINVIT . . . . . . . . . . . . . . . . . . . . . . . . . . . 1110
184 HEREVP . . . . . . . . . . . . . . . . . . . . . . . . . . 1111
185 BALANC . . . . . . . . . . . . . . . . . . . . . . . . . . 1111
186 BALBAK . . . . . . . . . . . . . . . . . . . . . . . . . . 1112
187 ELMHES . . . . . . . . . . . . . . . . . . . . . . . . . . 1112
188 HQR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1113
C.12 Ordinary Differential Equations . . . . . . . . . . . . . . . . . 1113
189 RKM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1113
190 RK4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1114
191 RK2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1114
192 MSTEP . . . . . . . . . . . . . . . . . . . . . . . . . . . 1114
193 ADAMS . . . . . . . . . . . . . . . . . . . . . . . . . . . 1116
194 STRT4 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1117
195 GEAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1117
196 EXTP . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1118
197 FDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1119
198 GEVP . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1121
199 GAUBLK . . . . . . . . . . . . . . . . . . . . . . . . . . 1125
200 SETMAT . . . . . . . . . . . . . . . . . . . . . . . . . . 1126
201 BSPODE . . . . . . . . . . . . . . . . . . . . . . . . . . 1126
vi Contents

C.13 Integral Equations . . . . . . . . . . . . . . . . . . . . . . . . 1128

202 FRED . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1128
203 FREDCO . . . . . . . . . . . . . . . . . . . . . . . . . . 1130
204 FUNK . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1130
205 RLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1131
206 FORW . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1134
207 VOLT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1135
208 VOLT2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 1135
C.14 Partial Differential Equations . . . . . . . . . . . . . . . . . . 1136
209 CRANK . . . . . . . . . . . . . . . . . . . . . . . . . . . 1136
210 LINES . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1137
211 ADM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1138
212 LAX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1140
213 SOR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1141
214 ADI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1142
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . 1144
Appendix C
C Programs

C.1 Introduction
All functions in this appendix are written in ANSI C, and have been checked
to some extent, but the author cannot guarantee their correctness. Readers are
welcome to use these functions at their own risk. These functions are straight-
forward translation of corresponding Fortran subprograms in Appendix B. It
should be noted that, these functions are not a substitute for standard mathe-
matical software, but merely concrete (and straightforward) examples of imple-
menting numerical algorithms described in this book. They are only suitable for
solving exercises and simple problems encountered in scientific computations.
Readers are expected to use these functions with ‘reasonable’ inputs. Passing
on arbitrary or invalid input parameters may give unpredictable results with
or without any warning. These functions can be effectively used to understand
the working as well as the limitations of various numerical algorithms, on dif-
ferent problems. Apart form the mathematical software libraries a number of
software packages are also available which allow the mathematical problems to
be specified in a convenient form for numerical solution. These software may
also produce incorrect results without any warning. Thus it is advisable to use
only those software where the user is aware of which technique is actually im-
plemented, so that their limitations may be known. The programs in the online
material are also provided in the same spirit. They may not necessarily give the
correct result in all cases, but since the algorithm used is known to the readers,
they can modify these to suit their requirements.
The functions in this Appendix are all written in ANSI C, but may require
some changes before running them on a new system. We leave it to the readers
to figure out the changes needed to run the programs on their machines. Since
these routines are simple translation of Fortran programs, they preserve the
same variable names and arguments in most cases. Nevertheless, since there
are some basic differences between C and Fortran this appendix provides an
independent summary of usage for these functions. In this section we outline
the major differences between the C and Fortran programs.
1008 Appendix C. C Programs

While in Fortran variable names are not case sensitive, C variable names
are case sensitive. Most of the variables in the C functions in Appendix C use
only lower case characters. However, for constants and global variables, which
may be accessible to more than one functions, we have used upper case letters
to avoid any possible conflict with local variables in functions. Nevertheless,
in most of comments in programs and in this Appendix, the variables appear
in upper case. There are two main reasons for this, first use of upper case
clearly identifies that the string of characters is a variable name and helps in
understanding and, second it makes it easier to translate the document from
Fortran to C. We hope that this will not confuse the readers.
In C all subprogram units are treated as functions which can return any
required value. Hence, we have used this facility to return the error parameter
(IER in Fortran subprograms) as far as possible. Only in those functions which
are supposed to return some other value the error parameter occurs in argument
list. But in almost all cases the significance and values returned by C functions
are same as those for IER in the corresponding Fortran subprograms. The argu-
ment IER appears explicitly only in functions BSPEVL, BSPEV2, BSPEVN,
BSPEVN1, BSPEVN2, BSPQD, BSPQD2, BSPQDN, DRVT, MULER2, SE-
CANI, SPLEVL. The argument IER or the returned value of the function is
used as an error flag. After execution of the functions this value should be zero
if no error has occurred. Nonzero values indicate some error condition. Values
of IER (or the function) less than 100, usually (but not necessarily) indicate
an error condition, which is not very serious and the result may still be useful.
However, further checks may be necessary before accepting the result. In most
of the interpolation and integration functions the returned value is set to less
than 100 even when it fails to converge to satisfactory accuracy. This is mainly
because in most cases these functions will still give a reasonable accuracy. But
this may not be the case in exceptional situations, due to singularity, or in case
of integration in large number of dimensions, simply because the number of
points allowed to be used may be completely inadequate. Even a zero value of
error flag does not necessarily imply that the result is correct as all possible
errors are not detected by any of these functions. Thus additional checks are
required in all cases. But in general one can expect that the result is likely to be
correct if error flag is zero. A value greater than 100, usually signifies that the
execution had to be terminated because of some serious error, and the result
is not likely to be correct. Of course, there will be many cases where even the
results in these cases are acceptable, but it will need careful analysis. A few of
the functions (e.g., BISECT, NEWRAP, FRED, SECANI) return a negative
value of error flag under some special circumstances, even though the execution
is successful. It may be noted that some functions call other functions. Hence, if
error flag is nonzero the error may have occurred in the secondary functions. In
such cases, only the final value of error flag may be returned. It may be noted
that if an error condition is detected the variable that is expected to return
the required result may contain some irrelevant value. Hence it is necessary
to check the value of IER after every call to these routines before accepting
C.1. Introduction 1009

the result. The value of IER returned by various functions has been modified
since the first edition was written, to make them more or less unique between
different functions. In the present version, when IER > 0, in most cases the
first digit is determined by the topic covered by the function. Thus the values
1x or 1xx are for linear algebra functions (Chapters 3 and 11), values 2x or
2xx are for interpolation and differentiation (Chapters 4 and 5), 3x or 3xx for
integration (Chapters 6), 4x or 4xx for nonlinear equations (Chapter 7), 5x or
5xx for optimisation (Chapter 8), 6x or 6xx for statistics and approximation
(Chapters 9, 10), 7x or 7xx for differential and integral equations (Chapters 12,
13, 14).
Apart from this C allocates memory dynamically and hence the variables
are not generally preserved between two successive calls to the same function. In
those functions, where the variables need to be preserved the required variables
are declared to be static. Since C also allows additional memory to be allocated
during execution, there is no need to pass on scratch arrays as required by
Fortran 77. Thus the arguments corresponding to scratch arrays may be missing
in C version of the Fortran subprograms. Similarly, in some cases where the
Fortran subprogram uses an array of fixed size, which limits the parameter
range where the solution can be calculated. This restriction is not necessary in
C functions as the array of required size can be allocated at run time.
The main source of confusion between Fortran and C subprograms arises
from the treatment of arrays. While Fortran by default numbers the array el-
ements of array A(N) from 1 to N, in C functions the array index varies from
0 to N − 1. Besides C does not have much facility to treat multi-dimensional
arrays and these are essentially treated as array of array. But the main dif-
ference arises in order in which elements are stored in the computer memory.
Thus an array of array A[3][3] is stored in the order A[0][0], A[0][1], A[0][2],
A[1][0], A[1][1], A[1][2], A[2][0], A[2][1], A[2][2]. Which is different from what
Fortran will use to store a two dimensional array. Basically, in Fortran the
first dimension varies most rapidly, while in C the last dimension varies most
rapidly in memory locations. To compound the problems, C does not allow the
dimensions to be passed on as argument to the function since the dimensions
have to be constant. On the other hand, Fortran allows variable dimensions
in subprograms, provided both the array name as well as the variable defining
the dimension also occur in the argument list. For one-dimensional arrays the
dimension information is not important as it is not required to calculate the
memory location of the required array element. But for two-dimensional arrays
it is not possible to calculate the memory location unless the second dimension
is known. In Fortran we need to know the first dimension. Unfortunately, there
is no straightforward way to pass on this second dimension to a C function.
Thus in all functions in this appendix we treat a multi-dimensional array as one
dimensional array. For example, in the array A[M][N] the element A[i][j] would
be in memory location j + iN with respect to the first element. We can refer
to this element as A[j + iN] if A is simply declared as a one-dimensional array
or just the address of its first element is passed as the argument to function.
1010 Appendix C. C Programs

We have followed this practice in all functions, where the second dimension
N can be variable. To calculate this position we need the actual value of the
second dimension N. While in Fortran we will need the first dimension of array
A(M, N) for the same purpose. As a result, the argument to be passed on to
Fortran program will not be same as that for the corresponding C function. To
circumvent this problem in most cases we have reversed the order of indices in
the arrays, thus an array element A[I][J][K] in C function will correspond to
A(K, J, I) in Fortran subprogram. The main reason for this choice is that in
many functions it is convenient to assume that elements are stored in the order
determined by Fortran subprogram, because the order of indices was chosen
to satisfy this requirement. If a different order is used then we may need fre-
quent copying of elements in another temporary array. Another advantage of
this convention is that the same number has to be passed on to the Fortran
and C programs as array dimension information.
Only exception we have made is for the situation where a two dimensional
array represents a matrix stored in the normal format for a linear algebra prob-
lem. In that case A[i][j] still represents jth column of ith row, which in Fortran
is referred to as A(i, j). For all other arrays the indices are interchanged to fol-
low the order in which memory is allocated. The main reason for this exception
is that the matrix elements are conventionally written in this order. Of course,
any two-dimensional array can be treated as a matrix, but we have broken
our normal convention of interchanging indices only when the two dimensional
array is supposed to represent a matrix arising in linear algebra problem like
solving a system of linear equations or the eigenvalue problem. In the documen-
tation for each function, the order of indices is spelt out, so hopefully, there
would be no confusion on this account.
Another major shortcoming of C is lack of variables of type complex.
Hence all complex variables need to be treated as arrays of length 2 or as struc-
tures and more importantly the arithmetic operations between these variables
need to be defined separately. Most functions in this Appendix treat complex
variables as arrays of length 2, while in some cases a complex array is treated
as two separate arrays containing the real and imaginary parts. The addition,
subtraction and multiplication of complex variables is explicitly coded in the
functions as corresponding operation on real and imaginary parts. But other op-
erations like division, taking absolute value or the square root require more care
to avoid possible problems with roundoff errors or overflows and underflows. For
these operations separate functions CDIV, CABS and CSQRT are provided. A
few of the Fortran subprograms requiring complex arithmetic, namely, CON-
TUR, DELVES, LAGITC, NEWRAC, POLYC, SECANC have not been trans-
lated to C. Apart from CONTUR and DELVES all these are complex versions
of corresponding routines for real functions. Apart from these, complex versions
of other routines like those for solution of a system of linear equations are also
not translated. The main reason for this is that C is probably not designed for
complex arithmetic and these routines are not likely to be required by C pro-
grammers. In principle, it is possible to call Fortran subprograms directly from
C.2. Roundoff Error 1011

a C function and the Fortran version for any of these routines can be used from
a C function. For calling a Fortran subprogram from C function all arguments
must be passed through address. Further, in general the Fortran compiler adds
some prefix or suffix of underline character, so the name has to be modified
accordingly.
All C functions in this appendix assume that variables of type int are 32
bit integers and all real variables are assumed to be of type double which is
expected to be 64 bits. In most functions 16 bit integers may be enough, but
there are a few functions, where this will not be sufficient. Similarly, 32 bit
real variables (type float) may be sufficient in many cases but for uniformity
all functions use real variables of type double. On most modern computers
the execution time is not significantly different between 32 bit and 64 bit real
variables, but the memory requirements will be doubled when 64 bits are used
instead of 32 bits.
Many of the functions have a parameter REPS and/or AEPS, which spec-
ify the required accuracy. Depending on the function this parameter could con-
trol either the relative or the absolute error in the final result. In all cases,
some heuristic convergence test is used to check for convergence of the results.
Consequently, the actual error may not necessarily be smaller than the required
accuracy. On the other hand, in most of the simple cases, where the conver-
gence is very rapid the actual error may be an order of magnitude smaller than
the required accuracy.
In the following sections we give the description of each function included
in the online material and its usage. A brief description of each variable in the
call statement also appears in the program files. Some programs giving actual
examples of usage are also included in the online material. Functions with
an underscore in the names are simple variants of other functions with the
suffix dropped and are not described separately in this Appendix. For example
CASSUM A is a variant of CASSUM. These functions are also included in the
list which can be found in the online material.

C.2 Roundoff Error

1. CASSUM Function to sum N terms of a series using the cascade sum
algorithm. The Ith term is calculated by the Function TERM(I), which must
be supplied by the user. If the number of terms exceeds 2N2MAX (= 230 ), then
true “binary” sum may not be calculated. By a simple change as indicated in
the program, it is possible to use this function to sum the terms in an array
TERM. In that case TERM should be an array containing the terms to be
summed. Function CASSUM A implements this variation.
2. ROUND Function to round a real number X to N digits using base B. It
is assumed that ~BN < 1, since otherwise, the accuracy of computer arithmetic
is not sufficient to give N digits. The returned value of the function should
be the rounded value of X. This function can be used to simulate calculations
1012 Appendix C. C Programs

using any required number of digits by rounding intermediate result after every
arithmetic operation. Since it is fairly expensive to use this function, it should
be used for small problems only.
3. CDIV Function to perform division for complex numbers. C1, C2 and
CR are all arrays of length 2 containing the real and imaginary parts of the
complex numbers. C1 is the numerator, C2 the denominator and CR = C1/C2
is the computed ratio. CR can be same as either C1 or C2, that is the result
can be overwritten on one of the operands. To avoid overflow or underflow in
some cases the calculations are rearranged as explained in Eq. (2.34).
4. CABS Function to calculate absolute value of a complex number. The
calculations are rearranged to avoid overflow or underflow by writing
r 
p x 2  y 2
CABS(x + iy) = x2 + y 2 = |w| + , (C.1)
w w
where w = max(|x|, |y|). CX is an array of length 2 containing the real and
imaginary parts of the complex number and CABS is the calculated absolute
magnitude.
5. CSQRT Function to calculate the square root of a complex number. The
calculations are rearranged as explained above to avoid overflow or underflow
and in addition the correct branch of the square root function has to be selected.
CX and CR are arrays of length 2 containing the real and imaginary parts of the
complex √numbers. CX is the input complex number while CR is the computed
value of CX. CR can be same as CX, that is, the result can be overwritten
on the input number.

C.3 Linear Algebraic Equations

6. GAUELM Function to solve a system of N linear equations using Gaussian
elimination with partial pivoting. N is the number of equations as well as the
number of unknown variables. NUM is the number of different right-hand side
vectors, for which the equations are to be solved. A is an array of length LJ × N
containing the matrix of coefficients. A[I][J] is the coefficient of xj in the Ith
equation. The matrix will be overwritten by the function. Hence, if it is required
afterwards, a separate copy should be saved. X is an array of length LJ × NUM,
containing the right-hand sides of the linear systems. The system will be solved
for each right-hand side vector given by the rows of X. X[J][I] should contain
the Ith component for the Jth right-hand side vector. After execution, the
array X will contain the required solutions. Thus the right hand side vectors
too will be overwritten by the function and if they are required afterwards,
a separate copy should be preserved. DET is an output parameter containing
the value of the determinant. INC is an integer array of length N, which will
contain the pivoting information after the function is executed. LJ is the actual
C.3. Linear Algebraic Equations 1013

value of the second dimension of arrays A and X, as declared in the calling

function (LJ ≥ N). IFLG is an integer parameter which determines the kind of
calculations required. If IFLG ≤ 1, the elimination is performed and IFLG is
set to 2, so that next time the elimination is not performed. For higher values
of IFLG it is assumed that elimination has already been performed, and the
matrix A is overwritten by the triangular factors, while the array INC contains
the information about interchanges. For IFLG ≤ 0, both the elimination as well
as solution for the required right-hand sides is calculated. If IFLG = 1, then
only the elimination is performed and the value of determinant is calculated. If
IFLG ≥ 2, then only the solution for the required right-hand sides is obtained.
It may be noted that the function resets the value of IFLG to 2. Hence, if it
is called again for a different matrix, the value of IFLG must be set to 0 or
1, to perform elimination for the new matrix. Error status can be obtained
from the returned value of GAUELM. A value of 101 indicates that N ≤ 0,
or N > LJ, in which case, no calculations will be performed. A value of 121
denotes that at some stage of elimination process, the pivot turned out to be
zero, in which case, the calculations are terminated at that stage. This failure
can occur only for matrices which are singular or almost singular. This error
can also arise if the second dimension of the matrix is not specified correctly.
Note that order of subscripts in X is reverse of that in the Fortran version, since
X is not considered as a matrix. This is to ensure that same value of LJ can
serve as second dimension for both arrays A and X. To calculate the inverse
of a matrix, the function can be called with NUM = N and array X set to a
unit matrix of order N. In this case, the transposed inverse will be returned in
the array X. Thus after calculation we will need to transpose the array X to
get the inverse matrix. This is required because of the order in which matrices
are assumed to be stored in memory. This function assumes that the matrix is
equilibrated and does not attempt any scaling. The determinant is calculated
in the simple form, which may give overflow or underflow for some matrices. If
the system of equations is ill-conditioned, the results could be unreliable and
the function may not give any indication of the problem. SVD should be used
to check for ill-conditioning.
7. MATINV Function to calculate the inverse of a square matrix using Gaus-
sian elimination, with partial pivoting. N is the order of matrix, IA is the second
dimension of arrays A and AI as specified in the calling function. A is an array
of length IA × N containing the matrix, which must be supplied at the time
of calling. AI is an array of length IA × N which will contain the inverse of A
as calculated by the function. Error status can be obtained from the returned
value of MATINV. Nonzero values may be set by function GAUELM, which is
used to calculate the inverse. Instead of GAUELM it is possible to use CROUT
for calculating the triangular decomposition.
8. CROUT Function to solve a system of N linear equations using Crout’s
algorithm for LU decomposition, with partial pivoting. N is the number of
equations as well as the number of unknown variables. NUM is the number of
1014 Appendix C. C Programs

different right-hand side vectors, for which the equations are to be solved. A is
an array of length LJ × N, containing the matrix of coefficients. A[i][j] is the
coefficient of xj in ith equation. This matrix will be overwritten by the func-
tion. Hence, if it is required afterwards, a separate copy should be saved. After
execution, A will contain the triangular decomposition of the original matrix.
X is an array of length LJ × NUM, containing the right-hand sides of the linear
systems. The system will be solved for each right-hand side vector given by
the rows of X. X[J][I] should contain the Ith component for the Jth right-hand
side vector. After execution, the array X will contain the required solutions.
Thus if the right-hand side vectors are required afterwards, a separate copy
must be preserved. DET and IDET are the output parameters containing the
value of the determinant in a scaled form. The actual value of the determinant
is DET × 2IDET . INC is an integer array of length N, which will contain the
pivoting information after the function is executed. LJ is the actual value of
the second dimension of arrays A and X, as declared in the calling function
(LJ ≥ N). IFLG is an integer parameter which determines the kind of calcula-
tions required. If IFLG ≤ 1, the LU decomposition is performed and IFLG is
set to 2, so that next time the triangular decomposition is not performed. For
higher values of IFLG, it is assumed that LU decomposition has already been
performed, and the matrix A is overwritten by the triangular factors, while the
array INC contains the information about interchanges. For IFLG ≤ 0, both
the LU decomposition as well as the solution for the required right-hand sides
are obtained. If IFLG = 1, then only LU decomposition is performed and the
value of determinant is calculated. If IFLG ≥ 2, then only the solution for the
required right-hand sides are obtained. It may be noted that, the function resets
the value of IFLG to 2. Hence, if it is called again for a different matrix, the
value of IFLG must be set to 0 or 1, to perform LU decomposition for the new
matrix. Error status can be obtained from the returned value of CROUT. A
value of 102 indicates that N ≤ 0, or N > LJ, in which case, no calculations will
be performed. A value of 122 denotes that at some stage of LU decomposition
the pivot turned out to be zero, in which case, the calculations are terminated
at that stage. This failure can occur only for matrices which are singular or
almost singular. To calculate the inverse of a matrix, the function can be called
with NUM = N and array X set to a unit matrix of order N. In this case, the
transposed inverse will be returned in the array X. Thus after calculation we
will need to transpose the array X to get the inverse matrix. This is required
because of the order in which matrices are assumed to be stored in memory.
This function implements implicit scaling of rows as explained in Section 3.3.
If the system of equations is ill-conditioned, the results could be unreliable and
the function may not give any indication of the problem. SVD should be used
to check for ill-conditioning.
9. CROUTH Function to solve a system of N linear equations using the tech-
nique of iterative refinement and the Crout’s algorithm for LU decomposition
with partial pivoting. N is the number of equations as well as the number of
unknown variables. NUM is the number of different right-hand side vectors, for
C.3. Linear Algebraic Equations 1015

which the equations are to be solved. A is an array of length LJ × N, containing

the matrix of coefficients. A[I][J] is the coefficient of xJ in Ith equation. The
matrix will be preserved by the function. B is an array of the same dimen-
sions as A. After execution, B will contain the triangular decomposition of the
original matrix. X is an array of length LJ × NUM, containing the right-hand
sides of the linear systems. The system will be solved for each right-hand side
vector given by the rows of X. X[J][I] should contain the Ith component for
the Jth right-hand side vector. After execution, the array X will contain the
required solutions. DET and IDET are the output parameters containing the
value of the determinant in a scaled form. The actual value of the determi-
nant will be DET × 2IDET . INC is an integer array of length N, which will
contain the pivoting information after the function is executed. LJ is the ac-
tual value of the second dimension of arrays A, B and X, as declared in the
calling function (LJ ≥ N). REPS is the required relative accuracy, to which
the iterative refinement is required. The iteration is terminated when the max-
imum change in any component of the solution is less than REPS times the
maximum component of the solution vector. REPS should be greater than ~
for the arithmetic used. It should be noted that convergence of iteration does
not guarantee that the solution is accurate to the specified level. A reasonable
estimate of error is provided by WK[J] for the Jth right-hand side. However,
for extremely ill-conditioned matrices, it may underestimate the error. IFLG
is an integer parameter which determines the kind of calculations required. If
IFLG ≤ 1, the LU decomposition is performed and IFLG is set to 2, so that
next time the LU decomposition is not performed. For higher values of IFLG,
it is assumed that LU decomposition has already been performed and the array
B contains the triangular factors, while the array INC contains the informa-
tion about interchanges. For IFLG ≤ 0, both the LU decomposition as well
as solution for the required right-hand sides are obtained. If IFLG = 1, then
only the LU decomposition is performed and the value of determinant is cal-
culated. If IFLG ≥ 2, then only the solution for the required right-hand sides
are obtained. It may be noted that the function resets the value of IFLG to 2.
Hence, if it is called again for a different matrix, the value of IFLG must be set
to 0 or 1, to perform LU decomposition for the new matrix. WK is an array of
length NUM. After execution, WK[I] will contain the estimated error for the
Ith right-hand side. Error status can be obtained from the returned value of
CROUTH. A value of 11 implies that the iterative refinement did not converge
to the specified accuracy. This failure can occur if the system of equations is
extremely ill-conditioned, or if the specified value of REPS is too low, or if the
residues are accumulated using the same level of precision as all other calcula-
tions. To avoid this the variable D1 should generally be of higher precision than
other variables. If the compiler supports long double arithmetic, D1 should be
declared so. If the matrix is ill-conditioned, increasing NITR may cause the
iteration to converge, but the result is unlikely to be more reliable. Returned
value of 102 indicates that N ≤ 0 or N > LJ, in which case, no calculations
will be performed. Returned value of 122 denotes that at some stage of LU
1016 Appendix C. C Programs

decomposition, the pivot turned out to be zero, in which case, the calculations
are terminated at that stage. This failure can occur only for matrices which
are singular or almost singular. This function requires CROUT to perform the
LU decomposition and to calculate solution of resulting equations for a given
right-hand side vector. To calculate the inverse of a matrix, the function can
be called with NUM = N and array X set to a unit matrix of order N. In this
case, the transposed inverse will be returned in the array X. If the system of
equations is extremely ill-conditioned, the results could be unreliable and the
function may not give any indication of problem. However, in most cases, WK[I]
will give a reasonable estimate of the actual error. This function should give
correctly rounded solution to the system as represented in the computer, pro-
vided the system is not too ill-conditioned for the precision of arithmetic used
in the calculations. It may be noted that for iterative refinement to converge it
may be necessary to calculate the residuals using higher precision arithmetic as
compared to that used for solution. Otherwise, the solution may not converge.
For this purpose the variable D1 should have higher precision as compared to
other variables.
10. CHOLSK Function to solve a system of N linear equations with a real sym-
metric positive definite matrix using Cholesky decomposition. N is the number
of equations as well as the number of unknown variables. NUM is the number
of different right-hand side vectors, for which the equations are to be solved.
A is an array of length ND × N, containing the matrix of coefficients. A[I][J]
is the coefficient of xJ in Ith equation. Only the lower triangular part of the
matrix may be used by the function, though the memory is reserved for the full
array. This matrix will be overwritten by the function. Hence, if it is required
afterwards, a separate copy should be saved. After execution, A will contain
the Cholesky decomposition of the original matrix in its lower triangular part.
X is an array of length ND×NUM, containing the right-hand sides of the linear
systems. The system will be solved for each right-hand side vector given by the
rows of X. X[J][I] should contain the Ith component for the Jth right-hand side
vector. After execution, the array X will contain the required solutions. DET
is the output parameter containing the value of the determinant. ND is the ac-
tual value of the second dimension of arrays A and X, as declared in the calling
function (ND ≥ N). IFLG is an integer parameter which determines the kind of
calculations required. If IFLG ≤ 1, the triangular decomposition is performed
and IFLG is set to 2, so that next time the triangular decomposition is not
performed. For higher values of IFLG, it is assumed that the Cholesky decom-
position has already been performed, and the matrix A is overwritten by the
triangular factor. For IFLG ≤ 0, both the decomposition as well as the solution
for the required right-hand sides are obtained. If IFLG = 1, then only decom-
position is performed and the value of determinant is calculated. If IFLG ≥ 2,
then only the solution for the required right-hand sides are obtained. It may
be noted that, the function resets the value of IFLG to 2. Hence, if it is called
again for a different matrix, the value of IFLG must be set to 0 or 1, to per-
form Cholesky decomposition for the new matrix. Error status can be obtained
C.3. Linear Algebraic Equations 1017

from the returned value of CHOLSK. A value of 103 indicates that N ≤ 0, or

N > ND, in which case, no calculations will be performed. CHOLSK = 123 de-
notes that at some stage of Cholesky decomposition the pivot turned out to be
zero, in which case, the calculations are terminated at that stage. This failure
can occur if the triangular decomposition does not exist without pivoting or if
the matrix is not positive definite or if it is almost singular. To calculate the
inverse of a matrix, the function can be called with NUM = N and array X set
to a unit matrix of order N. In this case, the transposed inverse will be returned
in the array X. If the system of equations is ill-conditioned, the results could
be unreliable and the function may not give any indication of the problem.
11. GAUBND Function to solve a system of N linear equations using Gaussian
elimination with partial pivoting for a band matrix. The matrix is stored in
the band form as explained below. N is the number of equations as well as
the number of unknown variables. KB is the bandwidth of the matrix, that is,
aij = 0, if |i − j| > KB. NUM is the number of different right-hand side vectors,
for which the equations are to be solved. A is an array of length LJ × (3KB + 1)
containing the matrix of coefficients which is stored in the band form. A[J −
I + KB][I] is the coefficient of xJ in the Ith equation. It may be noted that if
KB is comparable to N, then the band form will require larger storage than
the simple form of storing the matrix and no purpose will be served by using
this function instead of GAUELM. The matrix A will be overwritten by the
function. Hence, if it is required afterwards, a separate copy should be saved.
X is an array of length LJ × NUM, containing the right-hand sides of the linear
systems. The system will be solved for each right-hand side vector given by the
rows of X. X[J][I] should contain the Ith component for the Jth right-hand side
vector. After execution, the array X will contain the required solutions. DET
and IDET are the output parameters containing the value of the determinant
in a scaled form. The actual value of the determinant is DET × 2IDET . INC
is an integer array of length N, which will contain the pivoting information
after the function is executed. LJ is the actual value of the second dimension
of arrays A and X, as declared in the calling function (LJ ≥ N). IFLG is
an integer parameter which determines the kind of calculations required. If
IFLG ≤ 1, elimination is performed and IFLG is set to 2, so that next time
the elimination is not performed. For higher values of IFLG it is assumed that
elimination has already been performed, and the matrix A is overwritten by
the triangular factors, while the array INC contains the information about
interchanges. For IFLG ≤ 0, both the elimination as well as solution for the
required right-hand sides is calculated. If IFLG = 1, then only the elimination
is performed and the value of determinant is calculated. If IFLG ≥ 2, then only
the solution for the required right-hand sides is obtained. For IFLG = −1, both
elimination and solution are calculated without pivoting and IFLG is set to 2.
This option can be used if the matrix is known to be diagonally dominant and
pivoting is not necessary. It may be noted that the function resets the value of
IFLG to 2. Hence, if it is called again for a different matrix, the value of IFLG
must be set to 0 or 1 or −1, to perform elimination for the new matrix. Error
1018 Appendix C. C Programs

status can be obtained from the returned value of GAUBND. A value of 104
indicates that N ≤ 0, or N > LJ or KB > N, in which case, no calculations
will be performed. GAUBND = 124 implies that at some stage of elimination
process, the pivot turned out to be zero, in which case, the calculations are
terminated at that stage. This failure can occur only for matrices which are
singular or almost singular. To calculate the inverse of a matrix, the function
can be called with NUM = N and array X set to a unit matrix of order N. In
this case, the transposed inverse will be returned in the array X. The inverse of
a band matrix will in general not be banded and full storage will be required
to store the inverse. This function assumes that the matrix is equilibrated and
does not attempt any scaling. If the system of equations is ill-conditioned, the
results could be unreliable and the function may not give any indication of the
problem.
12. SVD Function to perform the singular value decomposition of a M × N
matrix, A = U ΣV T (M ≥ N). This function is based on the procedure svd in
Wilkinson and Reinsch (1971). A is an array of length LA × M containing the
matrix. After execution, the matrix U of SVD will be overwritten on A. V is
an array of length LV × N, which will contain the matrix V (not V T ) of SVD.
SIGMA is an array of length N, which will contain the singular values of A, i.e.,
the diagonal elements of the diagonal matrix Σ. The singular values may not be
arranged in a descending order. LA is the integer variable specifying the actual
value of the second dimension of array A in the calling function (LA ≥ N).
Similarly, LV is the actual value of the second dimension of array V in the
calling function (LV ≥ N). The matrices A and V are stored in normal form with
first index denoting the row. In principle, LA and LV could be equal and one of
them would be enough, but for compatibility with Fortran version two different
values are provided. Error status can be obtained from the returned value of
SVD. A value of 12 implies that the QR iteration used for finding SVD of the
bidiagonal form did not converge to the required accuracy. In this case, the last
value is accepted and calculations are continued. This situation will generally
arise only if parameter EPS is too low for the arithmetic used. The specified
value is suitable for 53-bit arithmetic. For use with lower precision arithmetic
EPS may need to be increased. If a reasonable value of EPS is specified, then
it normally requires only two or three iterations for each singular value, as
compared to the maximum number of iterations ITMAX = 30. SVD = 105
denotes that N ≤ 0, or N > LV, or N ≤ 0, or M > LA, or N > M, in which
case, no calculations are performed.
13. SVDEVL Function to evaluate the solution of a system of linear equations
using the SVD (x = V Σ−1 U T b). This function assumes that SVD is already
performed. N is the number of variables in the linear system, while M is the
number of equations. U is an array of size LU × M containing the left-hand
transformation matrix in SVD. V is an array of length LV × N containing the
right-hand transformation matrix in SVD. SIGMA is an array of length N,
containing the singular values. The singular values need not be arranged in
C.4. Interpolation 1019

any definite order. LU and LV are the actual values of the second dimension
of arrays U and V respectively, in the calling function. B is an array of length
M containing the right-hand side vector. After execution, B will contain the
required solution. REPS specifies the required accuracy for zeroing the singular
values. If σi < REPS×max(σj ), then the corresponding σi−1 is set to zero while
calculating the solution. For a square matrix, where none of the singular values
are reduced to zero, this function gives the unique solution. If some of the
singular values are reduced to zero, it gives solution with minimum norm to
which any arbitrary combination of columns of V corresponding to zero σi can
be added, to get the general solution. If any of the singular values are reduced
to zero the routine will give the least squares solution unless the right hand
side is in the range of the matrix. The right hand side is in the range if it is
orthogonal to all columns of U corresponding to singular values which have been
reduced to zero. For over-determined systems (M > N), this function gives the
least squares solution (with minimum norm if some of the singular values are
reduced to zero). Before using this function, the singular value decomposition
should be computed using the function SVD. It should be noted that, the
function SVD overwrites the matrix U on the original matrix A. The returned
value SVDEVL is always zero.

C.4 Interpolation

14. DIVDIF Function for interpolation using the Newton’s divided difference
formula. XB is the x value at which interpolated value is required. NTAB is
the number of entries in the table. Arrays X and F of length NTAB contain
the abscissas and the function values at the corresponding points. The array X
must contain abscissas in either ascending or descending order. NUSE specifies
the maximum number of points to be used for interpolation. If NUSE is larger
than NTAB, or the parameter NMAX in the function, then it will be reduced
to the minimum of these three numbers and the error flag DIVDIF will be set
to 22. If NUSE is less than 1, it will be set to MIN(6, NTAB) and DIVDIF
will be set to 21. In all cases, after execution, NUSE will contain the number
of points actually used. Thus for subsequent calls NUSE must be reset to the
required value before call. FB is the output array of length NUSE + 1 which
will contain the interpolated values. FB[I] gives the interpolated value using
I points, and FB[NUSE] gives the final value. It may be noted that the first
element FB[0] is not used to avoid confusion with the index. AEPS specifies the
required accuracy. Interpolation is continued, until two successive values differ
by less than AEPS. If AEPS ≤ 0, then this convergence criterion will never be
satisfied. DFB and DDFB are output parameters containing the first and second
derivative at the same point XB. Error status can be obtained from the returned
value of DIVDIF. A value of 21 implies that NUSE < 1, in which case it is set
to MIN(6, NTAB). DIVDIF = 22 implies that NUSE > NTAB, or NUSE >
NMAX, in which case it is reduced appropriately. DIVDIF = 23 implies that
1020 Appendix C. C Programs

the interpolation has not converged to the specified accuracy. It should be

noted that the convergence criterion only checks for convergence of interpolated
value and not the derivatives. Hence, the error in computed derivatives may
be much larger. In particular, if the derivative is required at one of the tabular
points, then AEPS should be set to zero, since otherwise the interpolated value
will converge immediately and execution will be terminated with a very crude
estimate for the derivatives. If the derivatives are not required, then function
DIVDIF0 may be used. This function requires the function NEARST.

15. DIVDIF0 Function for interpolation using the Newton’s divided differ-
ence formula. This is a simplified version of DIVDIF which avoids derivative
calculation and has a flag to decide whether the first point for interpolation
should be chosen from the nearest value in table or another point specified by
the user. This flag may be useful if the function is used to calculate interpolation
in higher dimensions. XB is the x value at which interpolated value is required.
NTAB is the number of entries in the table. Arrays X and F of length NTAB
contain the abscissas and the function values at the corresponding points. The
array X must contain abscissas in either ascending or descending order. NUSE
specifies the maximum number of points to be used for interpolation. If NUSE
is larger than NTAB, or the parameter NMAX in the function, then it will be
reduced to the minimum of these three numbers and the error flag DIVDIF0
will be set to 22. If NUSE is less than 1, it will be set to MIN(6, NTAB) and
DIVDIF0 will be set to 21. In all cases, after execution, NUSE will contain the
number of points actually used. Thus for subsequent calls NUSE must be reset
to the required value before call. FB is the output array of length NUSE + 1
which will contain the interpolated values. FB[I] gives the interpolated value us-
ing I points, and FB(NUSE) gives the final value. The first element FB[0] is not
used to avoid confusion. AEPS specifies the required accuracy. Interpolation is
continued, until two successive values differ by less than AEPS. If AEPS ≤ 0,
then this convergence criterion will never be satisfied. IFLG is an integer vari-
able used as a flag to decide the first point to be used for interpolation. If
IFLG = 0 then the nearest point in the table is used to start interpolation.
For other values of IFLG a user supplied value (IF1) is used, provided that it
is admissible (0 ≤ IF1 ≤ NTAB − 1). IF1 specifies the first point to be used
for interpolation when IFLG 6= 0. For IFLG = 0 the function uses the nearest
point in that table and the value of IF1 is set to index of this point, so that
next time we can use the same point if the interpolation in required at same
XB with different F values. This situation arises if this function is used for
interpolation in more than one variables. Error status can be obtained from
the returned value of DIVDIF0. A value of 21 implies that NUSE < 1, in which
case it is set to MIN(6, NTAB). DIVDIF0 = 22 implies that NUSE > NTAB,
or NUSE > NMAX, in which case it is reduced appropriately. DIVDIF0 = 23
implies that the interpolation has not converged to the specified accuracy. This
function requires the function NEARST.
C.4. Interpolation 1021

16. NEARST Function to locate the nearest value in an ordered table using
a bisection algorithm. X is an array containing the table with NTAB entries in
either ascending or descending order. After execution, X[NEARST] will be the
point nearest to XB.
17. SPLINE Function to calculate the coefficients of cubic spline interpolation
with not-a-knot boundary conditions using N tabular points. Arrays X and F
of length N should contain the input data. F[I] should contain the function
value at X[I]. The array X must be in ascending or descending order. After
execution, array C of length 3 × N, will contain the coefficients of cubic spline.
If XB is between X[I] and X[I + 1], (0 ≤ I < N − 1) the interpolant is given
by F[I]+DX*(C[I][0]+DX*(C[I][1]+DX*C[I][2])), where DX = XB − X[I].
Error status can be obtained from the returned value of SPLINE. A value of
201 implies that the number of points is less than 2. If the number of points
is 2, linear interpolation will be used, while for N = 3 quadratic interpolation
will be used. For higher values of N, cubic spline interpolation with not-a-knot
boundary conditions is calculated.
18. SPLEVL Function for evaluating the cubic spline interpolant at XB, us-
ing the coefficients of cubic spline which have been calculated by the function
SPLINE. This function first locates the subinterval containing the required
point XB, using the technique described in Section 4.2. N is the number of
data points. X and F are arrays of length N containing the data points, F[I] is
the tabulated function value at X[I]. C is an array of length 3 × N containing
the coefficients of cubic spline. The array X must be in ascending or descending
order. DFB and DDFB are output parameters containing the first and second
derivatives of the tabulated function at x = XB. IER is the error parameter.
IER = 24 implies that XB is outside the range of table on the higher side, in
which case, the cubic for last subinterval will be used. IER = 25 implies that
XB is outside the range on the lower side, in which case, the cubic for the first
subinterval will be used. It may be noted that spline interpolation is extremely
unreliable outside the range of table and resulting value may have large errors.
For extrapolation, it may be better to use polynomial interpolation with func-
tion DIVDIF, though it also has limitations. IER = 201 implies that N < 2.
SPLEVL will return the interpolated value of the function at XB. Before using
this function, the coefficients of cubic spline must be calculated using function
SPLINE or any other equivalent function.
19. SMOOTH Function to draw a smooth curve passing through a set of data
points using cubic spline interpolation. X and F are arrays of length NTAB
containing the given data points, C is an array of length 3 × NTAB, which will
contain the coefficients of cubic spline. The array X must be in either ascending
or descending order. This function will calculate the interpolated value of the
function at NP uniformly spaced points in the interval spanned by the tabular
points. XP and FP should be arrays of length NP which will contain the X and
F values for the uniformly spaced output table. If NP is sufficiently large, the
output arrays XP and FP can be used to plot a smooth curve through the given
1022 Appendix C. C Programs

points. It may be noted that this function does not actually plot the curve but
only generates data that can be used to get a smooth plot through the data
points. If additional smoothing is required and the curve is not required to
pass through data points then a least squares approximation should be used to
generate the points using function BSPFIT instead of SPLINE. Error status
can be obtained from the returned value of SMOOTH. A value of 202 implies
that NP < 2. Other values may be set by function SPLINE which is called to
calculate the spline coefficients. This function requires functions SPLINE and
SPLEVL.
20. BSPLIN Function to calculate the B-spline basis functions at a specified
point. X is an array of size NX containing the knots. The knots must be in
ascending order and distinct. NX is the number of knots. K is the order of
B-spline, K = 4 implies cubic B-splines, while K = 2 gives linear B-splines, etc.
XB is the point at which B-splines need to be evaluated. NDERIV specifies the
number of derivatives to be calculated, for NDERIV ≤ 0 only the B-splines will
be calculated. For NDERIV = 1, the first derivative of B-splines will also be cal-
culated. For NDERIV > 1, the second derivative will also be calculated. Higher
derivatives are not calculated, but the function can be modified to include cal-
culation of higher derivatives. B, DB and DDB are arrays of length NX + K − 2
which will contain the values of B-splines, its first and second derivative re-
spectively. All NX + K − 2 basis functions and if required the derivatives are
calculated simultaneously using the recurrence relations. Even if derivatives are
not required the arrays with required size must be supplied. This function first
locates the subinterval containing the required point XB, using the technique
described in Section 4.2. LEFT is an output variable which will give the loca-
tion of XB in the table of knots, i.e., X[LEFT] ≤ XB ≤ X[LEFT + 1]. Error
status can be obtained from the returned value of BSPLIN. A value of 26 im-
plies that the required point is outside the table on higher side. BSPLIN = 27
implies that the required point, XB is outside the range of table on lower side.
In both these cases the basis functions will be calculated using the same re-
currence relations, but the results may not be useful. BSPLIN = 203 implies
that NX < 2 or K < 1, in which case no calculations are done. The parameter
KMAX in the function can be trivially increased, but such high order splines
may not be meaningful.
21. BSPINT Function to calculate the coefficients for B-spline interpolation
to a table of values. N is the number of points in the table, X and F are arrays
of length N, specifying the abscissas and function values. The abscissas X must
be unique and in ascending order. F[I] is the tabulated function value at X[I].
K is the order of B-spline interpolation required. K = 4 implies cubic B-splines,
while K = 2 gives linear B-splines, etc. A is an array of length LA × 3K,
which will contain the triangular decomposition of the matrix of equations
that is solved to calculate the coefficients of expansion. Since the matrix is in
band form with bandwidth K − 1, it is stored in band form. This matrix will
be required if another interpolation with same set of knots is required with
C.4. Interpolation 1023

different F. This is useful in B-spline interpolation in higher dimensions. LA

is the second dimension of A as specified in the calling function (LA ≥ N).
C is an array of length N, which will contain the coefficients of B-spline basis
functions in expansion for interpolation. XF is an array of size NO, which will
contain the knots used for B-spline definition. Since the number of B-spline
basis functions with NO knots is NO + K − 2, some tabular points may have to
be dropped to ensure that N = NO + K − 2. This is equivalent to using the not-
a-knot boundary condition in cubic spline. The number of points to be dropped
depends on K. In all cases some points near both ends are dropped. For linear
B-splines, K = 2 all points are used. NO is the number of knots used for B-spline
interpolation, this would be equal to N+2−K. IFLG is an integer variable that
specifies what calculation is to be done. For IFLG ≤ 1 the matrix is calculated
and its triangular decomposition is computed. If execution is successful, IFLG
will be set to 2, so that next time the matrix calculations will be skipped. If
IFLG ≤ 0 the coefficients of expansion are also calculated. If IFLG = −1, the
system of equations will be solved without pivoting. This option may not be
used as the coefficient matrix may not be diagonally dominant and some pivot
may turn out to be zero. If IFLG = 2, only the coefficients of expansion will
be calculated using the old triangular decomposition available in A and the
(hopefully new) function values F. INC is an integer array of length N, which
will contain the pivoting information for solution of system of linear equations.
This array will be required if another interpolation with same knots is needed.
Error status can be obtained from the returned value of BSPINT. A value of
204 implies N < K or K < 2, in which case no calculations are done. Other
values may be set by functions BSPLIN or GAUBND, which are called. This
function requires function BSPLIN to calculate the B-spline basis functions
and function GAUBND to solve the system of linear equations with a band
matrix. The interpolant at any required point can be calculated using function
BSPEVL, using the calculated coefficients of expansion.
22. BSPEVL To calculate the value of function using available coefficients of
B-spline expansion. The calculated value is given by:
N+K−3
X
BSPEVL = WT[i]φi (X0), (C.2)
i=0

where φi are the B-spline basis functions. It can be used to calculate the in-
terpolated value of the function, if the coefficients are already calculated using
BSPINT. N is the number of knots, which may not be the same as the number
of points in the table for interpolation. X is an array of length N containing the
knots for B-splines. X[0] should contain the first knot and the knots must be
distinct and in ascending order. K is the order of B-splines, K = 2 gives linear
B-splines, while K = 4 yields cubic B-splines, etc. NDERIV specifies the num-
ber of derivatives to be calculated. For NDERIV ≤ 0 only the function value
will be calculated. For NDERIV = 1, the first derivative is also calculated. For
NDERIV > 1, the second derivative will also be calculated. WT is the array
1024 Appendix C. C Programs

of length N, containing the coefficients of expansion in terms of B-spline ba-

sis functions. These coefficients must be calculated beforehand using BSPINT
or any other equivalent function for B-spline approximations. X0 is the point
where the function value needs to be calculated. DF is the first derivative of
function at X0, while DDF is the second derivative. The derivatives would be
calculated only if NDERIV has been set appropriately. IER is the error param-
eter, which should be zero after successful execution. Nonzero values of IER
may be set by function BSPLIN which is called to calculate the B-spline basis
functions. This function requires function BSPLIN.
23. RATNAL To calculate rational function interpolation. XB is the x value
at which interpolated value is required. NTAB is the number of entries in the
given table. Arrays X and F of length NTAB contain the abscissas and the func-
tion values at the corresponding points. The array X must contain abscissas
in either ascending or descending order. NUSE specifies the maximum number
of points to be used for interpolation. If NUSE is larger than NTAB or the
parameter NMAX in the function, then it will be reduced to the minimum of
these three numbers and the error flag RATNAL may be set to 22. If NUSE is
less than 1, it will be set to MIN(6, NTAB) and RATNAL will be set to 21.
In all cases, after execution, NUSE will contain the number of points actually
used. Thus NUSE must be reset to required value before subsequent calls to the
function. FB is the output parameter containing the interpolated value. AEPS
is a parameter specifying the required accuracy. Interpolation will be continued
until two successive values differ by less than AEPS. If AEPS ≤ 0, then this
convergence criterion will never be satisfied. Error status can be obtained from
the returned value of RATNAL. A value of 21 implies that NUSE < 1, in which
case it is set to MIN(6, NTAB). RATNAL = 22 implies that NUSE > NTAB,
or NUSE > NMAX, in which case it is reduced appropriately. RATNAL = 23
implies that, the interpolation has not converged to the specified accuracy.
RATNAL = 205 implies that, the execution was terminated because the de-
nominator at some stage was zero. In this case, either the interpolant has a
pole at the requested point or it has a 0/0 form, and the function will return
with the interpolated value obtained in the previous step. This value may or
may not be acceptable. The reliability of this value may be verified by running
this function with successively increasing value of NUSE, until it encounters the
singularity. If these values are converging, then the result may be acceptable.
This function requires function NEARST.
24. POLY2 Function for polynomial interpolation in two dimensions. (XB1,
XB2) is the point at which interpolated value is required. X1 and X2 are arrays
of length N1 and N2 respectively, containing the abscissas. F is an array of
length NDIM × N2 containing the function values F[J][I] = f (X1[I], X2[J]).
NDIM should be set to the actual value of the second dimension of the array
F in the calling function (NDIM ≥ N1). NP1 and NP2 are the number of
points to be used along the two axes for interpolation. The function may adjust
these values if required. FB is the output parameter containing the interpolated
C.4. Interpolation 1025

value. Error status can be obtained from the returned value of POLY2. A value
of 206 implies that N1 > NDIM, in which case no calculations are done. This
function calls DIVDIF0 for one-dimensional interpolation along X1. To improve
efficiency, the derivative calculation has been removed from DIVDIF0 and a
flag has been introduced to avoid locating the nearest point every time. This
function will require functions DIVDIF0 and NEARST.
25. LINRN Function to perform linear interpolation in N variables. XB is an
array of length N containing the coordinates of the point at which interpolated
value is required. X is an array of length NXD × N containing the abscissas.
X[i][j] is the jth point along ith dimension. F is an N-dimensional array of di-
mension F[NDIM[N−1]] . . . [NDIM[1]][NDIM[0]] containing the function values

F[IN−1 ] . . . [I1 ][I0 ] = f (X[0][I0 ], X[1][I1 ], . . . , X[N − 1][IN−1 ]), (C.3)

NP is an integer array of length N, and NP[I] is the number of tabular points

along Ith coordinate. FB is the output parameter which will contain the inter-
polated value. NDIM is an integer array of length N, specifying the dimension
of F as explained above. This array must specify the dimensions of F as used
in the calling function. NXD is the actual value of the second dimension of X,
and it should be greater than the maximum element in NP. As an example,
consider the case, when we want to perform interpolation in four variables, with
4, 5, 6 and 7 points along the four coordinates, then the relevant statements
could be as follows
double f[8][6][5][5],x[5][10],xb[5],fb;
int n=4, nxd=10, np[4]={4, 5, 6, 7},ndim[4]={5, 5, 6, 8};
. . . . . . . . . . . . . . . .
ier= linrn(n,xb,&x[0][0],&f[0][0][0][0],np,&fb,ndim,nxd);
Error status can be obtained from the returned value of LINRN. A value of 207
implies that NP[I] is less than two or greater than NDIM[I] or NXD for some
I, in which case no calculations are performed. This function requires function
LOCATE.
26. LOCATE Function to locate a given point XB between two points of an
ordered table using the bisection algorithm. X is the table with NP entries in
either ascending or descending order. After execution, LOCATE returns a value
such that, XB should be between X[LOCATE] and X[LOCATE+1], unless it is
outside the range of the table. If XB is before the first point, then LOCATE = 0,
while if XB is beyond the last point, then LOCATE = NP − 2.
27. BSPINT2 Function to calculate the coefficients of interpolating B-spline
expansion in two dimensions. NX, NY are the number of points along x and
y, respectively in the table. X, Y are arrays of length NX and NY containing
the abscissas. The abscissas must be distinct and in ascending order. F is an
array of size LA × NY containing the function values. F[J][I] should contain
f (X[I], Y[J]). K is an integer variable specifying the required order of B-splines,
1026 Appendix C. C Programs

K = 4 for cubic B-splines, K = 2 for linear B-splines, etc. For simplicity, it is

assumed that order is the same for expansions along x and y. AX is an array of
length LA × 3K which will contain the triangular decomposition of matrix for
interpolation along X. AY is an array of length LA × 3K which will contain the
triangular decomposition of matrix for interpolation along Y. LA is the second
dimension of arrays AX, AY, C, F as declared in the calling function. For
simplicity, the second dimension is assumed to be the same in all these arrays.
LA has to be greater than the maximum of NX and NY. C is an array of length
LA × NY which will contain the calculated coefficients of expansion. This array
will be required to evaluate the interpolated value at any given point. XF and
YF are arrays of length MX and MY, respectively containing the knots for B-
splines along X and Y. MX, MY are the actual number of knots used along x and
y. These numbers may not be the same as NX, NY since some points may have
to be dropped to match the number of equations and number of B-spline basis
functions. IFLG is a flag to decide the nature of computations. For IFLG ≤ 1
the triangular decomposition of matrices is calculated, for larger values of IFLG
it is assumed that the triangular decomposition and other information is already
available in arrays AX, AY. INTX and INTY are integer arrays of lengths
NX and NY respectively, which will contain the information about pivoting
used to solve the system of equations for determining the coefficients for 1-
dimensional interpolation. Error status can be obtained from the returned value
of BSPINT2. A zero value implies successful execution of the function. Nonzero
values may be set by BSPINT, BSPLIN or GAUBND which are called. This
function requires BSPINT, BSPLIN and GAUBND. The interpolated value
can be calculated using function BSPEV2 with the coefficients computed by
BSPINT2.

28. BSPEV2 Function to calculate the value of B-spline expansion in two

dimensions using available coefficients. The calculated value is given by:

NX+K−3
X NY+K−3
X
BSPEV2 = WT[j][i]φi (X0)ψj (Y0), (C.4)
i=0 j=0

where φi (x) are the B-spline basis functions along x and ψi (y) are the B-spline
basis functions along y. It can be used to calculate the interpolated value of
the function if the coefficients are already calculated using BSPINT2. NX, NY
are the number of knots along X, Y, which may not be the same as the num-
ber of points in the table for interpolation. X, Y are arrays of length NX, NY
containing the knots for B-splines. The knots must be in ascending order with
X[0] and Y[0] containing the first knot along respective directions. K is the
order of B-splines, K = 2 gives linear B-splines, while K = 4 yields cubic B-
splines, etc. For simplicity, the order is assumed to be the same along both axes.
NDERIV specifies the number of derivatives to be calculated. For NDERIV ≤ 0
only the function value will be calculated. For NDERIV = 1, the first deriva-
tives are also calculated. For NDERIV > 1, the second derivatives will also be
C.4. Interpolation 1027

calculated. WT is the array of length IW × (NY + K − 2), containing the coeffi-

cients of expansion in terms of B-spline basis functions. These coefficients must
be calculated beforehand using BSPINT2 or any other equivalent function for
B-spline approximations. X0, Y0 is the point where the function value needs
to be calculated. DFX, DFY are the first derivatives of function with respect
to x, y respectively, at (X0, Y0). while DFXX, DFXY, DFYY are the second
derivatives ∂ 2 f /∂x2 , ∂ 2 f /∂x∂y, ∂ 2 f /∂y 2 . The derivatives would be calculated
only if NDERIV has been set appropriately. IER is the error parameter, which
should be zero after successful execution. Nonzero values of IER may be set by
function BSPLIN which is called to calculate the B-spline basis functions. This
function requires function BSPLIN.
29. BSPINTN Function to calculate the coefficients for B-spline interpolation
to a table of values in N dimensions. N is the number of dimensions. NK is an
integer array of length N, giving the number of tabular points along each dimen-
sion. NK[I] is the number of points along xI . X is an array of length NXD × N
specifying the abscissas along each dimension. X[J][I] is the Ith abscissa along
the Jth dimension. For each dimension, the abscissas must be unique and in
ascending order. NXD is the second dimension of arrays X, XF and INTX as
specified in the calling function, NXD ≥ max(NK[0], NK[1], . . . , NK[N − 1]). F
is an array of length NK[0] × NK[1] × · · · × NK[N − 1] containing the table of
values. Since BSPINTN treats this array as a one dimensional array, the di-
mensions of F in the calling function must exactly match the size of the table so
that there are no gaps in memory allocation. For example, the dimension could
be F[NK[N−1]] . . . [NK[1]][NK[0]]. Alternately, it could be treated as a one di-
mensional array of appropriate length. K is the order of B-spline interpolation
required. K = 4 implies cubic B-splines, while K = 2 gives linear B-splines, etc.
For simplicity, K is assumed to be the same along all dimensions. AX is an array
of length NXD × 3K × N, which will contain the triangular decomposition of
the matrix of equations that is solved to calculate the coefficients of expansion
for each dimension. Since the matrix is in band form with bandwidth K − 1, it
is stored in band form. C is an array of length NK[0] × NK[1] × · · · × NK[N−1],
which will contain the coefficients of B-spline basis functions in expansion for
interpolation. This array is also treated as one-dimensional array just like F
and hence the dimensions in calling function must exactly match the size of
the table. XF is an array of size NXD × N which will contain the knots used
in each dimension. This array will contain the knots used for B-spline inter-
polation in each dimension. Since the number of B-spline basis functions with
MK[I] knots is MK[I] + K − 2, some tabular points may have to be dropped to
ensure that NK[I] = MK[I] + K − 2. This is equivalent to using the not-a-knot
boundary conditions in cubic spline interpolation. The number of points to be
dropped depends on K. In all cases some points near both ends are dropped.
For linear B-splines, K = 2 all points are used. MK is an integer array of length
N containing the number of knots for B-splines in each dimension. MK[I] is
the number of knots used for B-spline interpolation, this would be equal to
NK[I] + 2 − K. INTX is an integer array of length NXD × N, which will contain
1028 Appendix C. C Programs

the pivoting information for solution of system of linear equations. Error status
can be obtained from the returned value of BSPINTN. A zero value implies suc-
cessful execution of function. Nonzero values may be set by functions BSPINT,
BSPLIN or GAUBND, which are called. It may be noted that for simplicity,
this function does not use IFLG to control the calculations as for large num-
ber of dimensions main effort goes in calculating the coefficients rather than
in calculating the triangular decomposition of matrices. This function requires
BSPINT to perform interpolation in one dimension, BSPLIN to calculate the
B-spline basis functions and GAUBND to solve the system of linear equations
with a band matrix. The interpolant at any required point can be calculated
using function BSPEVN, using the calculated coefficients of expansion. If the
first derivative of the function is also required then use BSPEVN1, while for
second derivatives use BSPEVN2.

30. BSPEVN Function to calculate the value of B-spline expansion in n-

dimensions using available coefficients. The calculated value is given by:

NK[0]+K−3 NK[n−1]+K−3
X X
BSPEVN = ··· WT[in−1 ] . . . [i0 ]
i0 =0 in−1 =0 (C.5)
(0) (n−1)
× φi0 (X0[0]) · · · φin−1 (X0[n − 1]),

(j)
where φij are the B-spline basis functions along jth dimension. It can be used
to calculate the interpolated value of the function if the coefficients are already
calculated using BSPINTN. N is the number of dimensions. NK is an integer
array of length N containing the number of knots along each dimension, which
may not be the same as the number of points in the table for interpolation.
NK[I] is the number of knots along the Ith dimension. X is an array of length
NXD × N containing the knots for B-splines along each dimension. The knots
must be distinct and in ascending order with X[J][I] containing Ith knot along
Jth dimension. NXD is the second dimension of array X in the calling function.
NXD must be greater than or equal to the maximum of NK[I]. K is the order of
B-splines, K = 2 gives linear B-splines, while K = 4 yields cubic B-splines, etc.
WT is the array of length (NK[0]+K−2)(NK[1]+K−2) · · · (NK[N−1]+K−2),
containing the coefficients of expansion in terms of B-spline basis functions.
Since BSPEVN treats this array as a one-dimensional array, the dimensions of
this array in the calling function must exactly match the table size, so that
there are no gaps in memory allocation. For example, the dimension could
be WT[NK[N−1]+K−2] . . . [NK[1]+K−2][NK[0]+K−2]. These coefficients WT
must be calculated beforehand using BSPINTN or any other equivalent func-
tion for B-spline approximations in n-dimensions. X0 is an array of length N,
specifying the coordinates of the point where the function value needs to be
calculated. IER is the error parameter, which should be zero after successful
execution. Nonzero value of IER may be set by function BSPLIN which is called
to calculate the B-spline basis functions. This function does not calculate the
C.5. Differentiation 1029

derivatives of expansion. Since calculating derivatives could require consider-

able extra time separate versions of this function BSPEVN1 and BSPEVN2
are provided to calculate first and second derivatives. These functions can be
used instead of BSPEVN when derivatives are required. This function requires
function BSPLIN.
31. BSPEVN1 To calculate the value of function and its first derivatives using
available coefficients of B-spline expansion in n-dimensions. This function is
the version of BSPEVN to calculate the first derivatives in addition to function
values. It can be used when derivatives are also required. The arguments are the
same as those for BSPEVN, except for array DF of length N, which will contain
the calculated derivatives with respect to each of the dimensions. DF[I] will
contain the first derivative with respect to xI at the point X0. If only function
value is required then BSPEVN should be used, while if second derivative is
also required then BSPEVN2 should be used.
32. BSPEVN2 To calculate the value of function and its first and second
derivatives using available coefficients of B-spline expansion in n-dimensions.
This function is the version of BSPEVN to calculate the first and second deriva-
tives in addition to function values. It can be used when derivatives are also
required. The arguments are the same as those for BSPEVN, except for array
DF of length N, which will contain the calculated derivatives with respect to
each of the dimensions. DF[I] will contain the first derivative with respect to xI
at the point X0. DDF is an array of length N × N, which will contain the calcu-
lated second derivatives. DDF[I][J] will contain ∂ 2 f /∂xI ∂xJ at the point X0.
If second derivatives are not required then it will be better to use BSPEVN1
(for first derivatives) or BSPEVN (for only function value).

C.5 Differentiation

33. DRVT Function to calculate derivative of a function which can be evalu-

ated at any required point using a user supplied function F(X). This function
uses h → 0 extrapolation to obtain accurate value of the derivative. It can
evaluate the first, second, third or fourth derivative of the given function. A is
the value of x at which the derivative is to be evaluated. ID is the order of the
derivative required. ID can be 1, 2, 3 or 4. If any other value is specified the
function will exit with the error flag IER set to 208. HH0 is the initial spacing
to be used, which depends on the function and the precision of the arithmetic
used. A value between 0.1 and 1 will usually be sufficient. If HH0 is too large,
then sufficient accuracy may not be achieved by the function, while if it is too
small, then roundoff error will dominate. AEPS and REPS specify the required
absolute and relative accuracy. This function exits when the difference between
successive values is less than max(AEPS, REPS × |DRVT|), or when roundoff
error is dominating. The presence of roundoff error is detected by using a simple
test as explained in Section 5.3. F is the name of the function which calculates
1030 Appendix C. C Programs

the given function at any required point. This function must be supplied by the
user. IER is the error parameter, which will be set to 28 if sufficient accuracy
is not achieved. IER will be set to 29 if roundoff errors start dominating before
adequate convergence is achieved. IER will be set to 208 if ID is outside the
specified range.

C.6 Integration
34. SIMSON To integrate a smooth function over a finite interval using com-
posite Simpson’s rule. RI is the output parameter containing the calculated
value of the integral. XL and XU are the lower and upper limits for the inte-
gral. It is not essential to have XL < XU. REPS and AEPS specify the required
relative and absolute accuracy. The calculations are terminated when two suc-
cessive values differ by less than max(AEPS, REPS × |RI|). DIF is an output
parameter containing an estimate of (absolute) error calculated by the function.
N is an output parameter which will contain the number of function evaluations
used. FUN is the name of the function to calculate the integrand. Error status
can be obtained from the returned value of SIMSON. A value of 30 implies
that the function has failed to converge to the specified accuracy. In most cases
the computed value of the integral may still be approximately correct. Only
in extreme cases, the computed value will be far from actual value. In this
case, RI will contain the best estimate for integral and DIF should contain the
estimated error. The parameter NMAX in the function could be increased, if
larger number of function evaluations are to be allowed. However, in most case,
it will be better to use another technique for evaluating the integral. Function
FUN(X) must be supplied by the user.
35. SPLINT To compute integral of a function supplied in the form of a ta-
ble of values by integrating the interpolating cubic spline. Before calling this
function, the coefficients of cubic spline must be calculated by a call to function
SPLINE. SINT and TINT are the output parameters containing the value of the
integral. SINT is the estimate using cubic spline, while TINT is the trapezoidal
rule estimate, using only the tabular points. The difference between TINT and
SINT may give some estimate of truncation error in the calculations. If the lim-
its are not among the tabular points, then the function value at the end points
are obtained using function SPLEVL. A and B are the lower and upper limits
of the integral. It is essential to ensure that A < B. N is the number of points
in the table of values. X and F are arrays of length N, containing the abscissas
and function values. The abscissas must be supplied in ascending order. C is
an array of length 3 × N containing the coefficients of cubic spline, which may
be calculated using the function SPLINE. Error status can be obtained from
the returned value of SPLINT. A value of 31 implies that the lower limit A is
outside the limits of table, while SPLINT = 32 denotes that the upper limit
B is outside the limits of table. In such cases, the accuracy of integration may
be questionable. SPLINT = 301 implies that A > B or X[0] > X[N − 1]. This
C.6. Integration 1031

function requires function SPLEVL to evaluate the cubic spline, while func-
tion SPLINE will be required to calculate the coefficients of cubic spline before
calling this function.
36. BSPQD Function to compute integral of an expansion in terms of B-spline
basis functions. The expansion may be obtained by interpolating or approxi-
mating a table of values. Before calling this function, the coefficients of B-spline
expansion must be calculated by a call to function BSPINT.
Z XU N+K−3
X
BSPQD = WT[i] φi (x) dx , (C.6)
XL i=0

where φi (x) are the B-spline basis functions of order K. N is the number of
knots, X is an array of length N containing the knots for B-splines. The knots
must be in ascending order with X[0] containing the first knot. K is the order
of B-splines, K = 2 gives linear B-splines, while K = 4 yields cubic B-splines,
etc. WT is an array of length N + K − 2 containing the coefficients of expan-
sion in terms of B-spline basis functions. These coefficients may be calculated
using BSPINT for interpolation in a table of values. XL and XU are the lower
and upper limits of the integral. IER is the error parameter. IER = 31 implies
that the lower limit XL is outside the limits of table, while IER = 32 denotes
that the upper limit XU is outside the limits of table. In such cases, the accu-
racy of integration may be questionable. Other values may be set by function
BSPLIN which is called to calculate B-spline basis functions. This function re-
quires function BSPLIN to evaluate the B-spline basis functions, while function
BSPINT will be required to calculate the coefficients of B-spline before calling
this function.
37. ROMBRG Function to compute integral over a finite interval, using
Romberg integration or the h → 0 extrapolation. RI is the output parame-
ter containing the value of the integral. A and B are the lower and upper limits
of the integral. It is not essential to have A < B. GI is an array of dimension
NMAX (= 13), containing the expected values of successive exponents in the
error expansion for the trapezoidal rule. If GI[I] ≤ 0, then it will be set to 2I+2,
which is the correct value for smooth functions. AEPS and REPS specify the
required absolute and relative accuracy. The calculations are terminated when
the successive values differ by less than max(AEPS, REPS × |RI|). DIF is an
output parameter giving the error as estimated by the function. N specifies the
number of abscissas to be used for the first attempt with trapezoidal rule. After
execution, N will contain the number of abscissas actually used by the function.
Hence, N will have to be reset after every call to the function. If N < 2 it will
be set to a default value of 2. On the other hand, if N > NPT (= 100), then
it will be set to a default value of 2 and the error flag ROMBRG will be set
to 33. In particular, this situation may arise if N is not reset after the previous
call to the function. Error status can be obtained from the returned value of
ROMBRG. A value of 30 implies that the function failed to converge to the
1032 Appendix C. C Programs

specified accuracy. In this case, RI will contain the best estimate for the integral
and DIF should contain the estimated error. The parameter NMAX inside the
function could be increased to allow for larger number of function evaluations,
but it is better to use some other technique for evaluating the integral. FUN is
the name of the function used to calculate the value of the integrand. Function
FUN(X) must be supplied by the user. This function should be used for smooth
functions with GI(I) = 0. It can also handle algebraic singularity at one or both
end points, provided correct values of GI are supplied.
38. EPSILN To compute integral of a function over a finite interval, using -
algorithm to accelerate the convergence. RI is the output parameter containing
the value of the integral. A and B are the lower and upper limits of the integral.
It is not essential to have A < B. AEPS and REPS specify the required absolute
and relative accuracy. The calculations are terminated when the successive
values differ by less than max(AEPS, REPS×|RI|). DIF is an output parameter
giving the error as estimated by the function. N specifies the number of abscissas
to be used for the first attempt with trapezoidal rule. After execution, N will
contain the number of abscissas actually used by the function. Hence, N will
have to be reset after every call to the function. If N < 2, it will be set to
a default value of 2. On the other hand, if N > NPT (= 100), then it will
be set to a default value of 2 and the error flag EPSILN will be set to 33. In
particular, this situation may arise if N is not reset after the previous call to the
function. Error status can be obtained from the returned value of EPSILN. A
value of 30 implies that the function failed to converge to the specified accuracy.
EPSILN = 34 denotes that at some stage while constructing the -table, the
denominator was zero. In this case, the calculations are continued further after
ignoring the corresponding term. This may be justified if this problem occurs
in higher columns which are not converging because of roundoff error. A better
strategy would be to stop calculating the higher columns. EPSILN = 35 denotes
that the roundoff error is dominating and the calculations are terminated, even
though the required accuracy is not achieved. In all these cases, RI will contain
the best estimate for the integral and DIF should contain the estimated error.
The parameter NMAX inside the function could be increased to allow for larger
number of function evaluations, but that is not recommended. FUN is the
name of the function used to calculate the value of the integrand. Function
FUN(X) must be supplied by the user. This function could be used for singular
integrands, if only moderate accuracy is required. It will fail to converge to high
accuracy because of roundoff error.
39. GAUSS Function to compute integral over a finite interval, using com-
posite Gauss-Legendre formulae. RINT is the output parameter containing the
value of the integral. A and B are the lower and upper limits of the integral.
It is not essential to have A < B. NP specifies the formula to be used. This
function uses a composite rule based on NP-point Gauss-Legendre formula.
NP should be 2, 4, 8, 16 or 32, since these are the only formulae for which
the weights and abscissas are incorporated. If NP is not equal to one of these
C.6. Integration 1033

values, then it will be set to a default value of 8, and error flag will be set
to 36. AEPS and REPS specify the required absolute and relative accuracy.
The calculations are terminated when the successive values differ by less than
max(AEPS, REPS × |RINT|). DIF is an output parameter giving the error as
estimated by the function. NPT is an output parameter containing the num-
ber of function evaluations actually used. Error status can be obtained from
the returned value of GAUSS. A value of 30 implies that the function failed
to converge to the specified accuracy. In this case, RINT will contain the best
estimate for the integral and DIF should contain the estimated error. In such
cases, attempt could be made to use higher order Gaussian formula, NP = 32 is
the maximum value allowed by this function. The parameter NMAX inside the
function could be increased to allow for larger number of function evaluations,
but it will be preferable to use some other technique for evaluating the integral.
GAUSS = 36 implies that NP was not 2,4,8,16 or 32, in which case it is set to
8. FUN is the name of the function used to calculate the value of the integrand.
Function FUN(X) must be supplied by the user. This function should be used
for smooth integrands. NP could be set to a low value if only moderate accu-
racy is required, while for high accuracy NP may be increased. The optimum
value of NP will depend on the degree of smoothness of the integrand and the
accuracy required.
40. GAUCBY Function topcompute integral using Gauss-Chebyshev formu-
lae with weight function 1/ (x − A)(B − x). RINT is the output parameter
containing the value of the integral. A and B are the lower and upper limits
of the integral. It is essential to have A < B. AEPS and REPS specify the
required absolute and relative accuracy. The calculations are terminated when
the successive values differ by less than max(AEPS, REPS × |RINT|). DIF is
an output parameter giving the error as estimated by the function. NPT is an
output parameter containing the number of function evaluations actually used.
Error status can be obtained from the returned value of GAUCBY. A value
of 30 implies that the function failed to converge to the specified accuracy. In
this case, RINT will contain the best estimate for the integral and DIF should
contain the estimated error. The parameter NMAX inside the function could
be increased to allow for larger number of function evaluations, but it will be
preferable to use some other technique for evaluating the integral. FUN is the
name
p of the function used to calculate the value of the integrand multiplied by
(x − A)(B − x). Function FUN(X) must be supplied by the user. This func-
tion should be used for integrands with square root singularity at both ends.
It will evaluate the integral
Z B
FUN(x)
p dx . (C.7)
A (x − A)(B − x)

41. GAUCB1 Function p to compute integral using Gauss-Chebyshev formulae

with weight function (x − A)/(B − x). RINT is the output parameter con-
taining the value of the integral. A and B are the lower and upper limits of
1034 Appendix C. C Programs

the integral. It is essential to have A < B. AEPS and REPS specify the re-
quired absolute and relative accuracy. The calculations are terminated when
the successive values differ by less than max(AEPS, REPS × |RINT|). DIF is
an output parameter giving the error as estimated by the function. NPT is an
output parameter containing the number of function evaluations actually used.
Error status can be obtained from the returned value of GAUCB1. A value
of 30 implies that the function failed to converge to the specified accuracy. In
this case, RINT will contain the best estimate for the integral and DIF should
contain the estimated error. The parameter NMAX inside the function could
be increased to allow for larger number of function evaluations, but it will be
preferable to use some other technique for evaluating the integral. FUN is the
namep of the function used to calculate the value of the integrand multiplied
by (B − x)/(x − A). Function FUN(X) must be supplied by the user. This
function should be used for integrands with square root singularity at both
ends. It will evaluate the integral
Z B √
FUN(x) x − A
√ dx . (C.8)
A B−x

42. GAUCB2 Function p to compute integral using Gauss-Chebyshev formulae

with weight function (x − A)(B − x). RINT is the output parameter con-
taining the value of the integral. A and B are the lower and upper limits of
the integral. It is essential to have A < B. AEPS and REPS specify the re-
quired absolute and relative accuracy. The calculations are terminated when
the successive values differ by less than max(AEPS, REPS × |RINT|). DIF is
an output parameter giving the error as estimated by the function. NPT is an
output parameter containing the number of function evaluations actually used.
Error status can be obtained from the returned value of GAUCB2. A value
of 30 implies that the function failed to converge to the specified accuracy. In
this case, RINT will contain the best estimate for the integral and DIF should
contain the estimated error. The parameter NMAX inside the function could
be increased to allow for larger number of function evaluations, but it will be
preferable to use some other technique for evaluating the integral. FUN is the
name
p of the function used to calculate the value of the integrand divided by
(B − x)(x − A). Function FUN(X) must be supplied by the user. This func-
tion should be used for integrands with square root singularity at both ends.
It will evaluate the integral
Z B p
FUN(x) (x − A)(B − x) dx . (C.9)
A

43. GAUSQ2 To compute integral of a function with square root singular-

ity over√(0, a] using a combination of Gauss-Legendre and a Gaussian formula
with 1/ x weight function. RINT is the output parameter containing the value
of the integral. A is the upper limit of the integral. A1 is the point at which
C.6. Integration 1035

the integral is broken, with the integral over [A1, A] being evaluated using
Gauss-Legendre √ formula, while that over (0, A1] is evaluated using a Gaussian
formula with 1/ x weight function. If A < A1, then A1 is set equal to A to
start with. A1 is adjusted by the function to achieve the required accuracy.
After execution, A1 will contain the final value used by the function. A1 may
need to be reset after every call to GAUSQ2. If the next integral is similar to
the previous one, then the value of A1 need not be reset, since it is probably
the required value. AEPS and REPS specify the required absolute and relative
accuracy. The calculations are terminated when the successive values differ by
less than max(AEPS, REPS × |RINT|). Convergence is checked separately for
the two integrals. DIF is an output parameter giving the error as estimated
by the function. NP is the output parameter containing the number of func-
tion evaluations actually used. This function calls the functions GAUSS and
GAUSQ to perform the integration over required subintervals. Error status can
be obtained from the returned value of GAUSQ2. A value of 30 implies that the
function GAUSS failed to converge to the specified accuracy over the interval
[A1, A]. GAUSQ2 = 37 denotes that the function GAUSQ failed to converge to
the specified accuracy, even when A1 was reduced to its minimum permissible
value of AMN (= 0.01). This limit is provided to prevent the function from
getting√ into an infinite loop. This can happen if the integrand is not in a form
f (x)/ x as required for Gaussian formula. If necessary, the lower limit AMN
can be reduced. GAUSQ2 = 38 denotes that both functions failed. In all these
cases, RINT will contain the best estimate for the integral and DIF should
contain the estimated error. F is the name of the function used to calculate
the value of the integrand. F2 is the name of the function used to calculate
F(X)*SQRT(X), as required by the Gaussian formula. Note that the function
will find the integral
Z A Z A Z A1
F2(x)
F(x) dx = F(x) dx + √ dx . (C.10)
0 A1 0 x
Functions F(X), and F2(X) must be supplied by the user. The requirement of
two separate functions is to avoid possible problems caused by singularity in
integrand. If A − A1 < AEPS, then the first integral is not evaluated. This
condition may need to be changed if the integral can make significant contribu-
tion, even though this condition is satisfied, which may happen if the limits are
very small. This function requires functions GAUSS and GAUSQ. It should be
used only for those integrands which have √ square root singularity at x = 0. It
can be used for integrands of form f (x)/ x,√where f (x) is regular
√ at x = 0. It
can also be used for integrands of form f (x) x = (xf (x))/ x.
44. GAUSQ Function to compute integral
Z A
FUN(x)
√ dx , (C.11)
0 x
√
using Gaussian formulae with 1/ x weight function. The function computes
the integral using different formulae, until convergence is achieved or until the
1036 Appendix C. C Programs

table of weights and abscissas is exhausted. RINT is the output parameter

containing the value of the integral. A is the upper limit of the integral. REPS
and AEPS specify the required accuracy, while DIF is an output parameter
giving the error as estimated by the function. The convergence criterion used
is |DIF| < max(REPS|RINT|, AEPS). NPT is an output parameter containing
the number of function evaluations actually used. FUN is the name √ of the
function used to calculate the value of the integrand (multiplied by x). Error
status can be obtained from the returned value of GAUSQ. A value of 30
implies that the function failed to converge to the specified accuracy. In this
case, RINT will contain the best estimate for the integral and DIF should
contain the estimated error. In such cases, we can try to use a lower value
of A. Function FUN(X) must be supplied by the user. This function should
be used
√ for integrands
√ which have a square root singularity (functions of form
f (x)/ x or f (x) x) at x = 0. This function should be preferably used through
GAUSQ2, in which case the upper limit A may be adjusted if the required
accuracy is not achieved.
45. GAULAG Function to compute integral of a function over [a, ∞) using
a combination of Gauss-Laguerre and Gauss-Legendre quadrature formulae.
RINT is the output parameter containing the value of the integral. A is the
lower limit of the integral. A1 is the point at which the integral is broken, with
the integral over [A, A1] being evaluated using Gauss-Legendre formula, while
that over [A1, ∞) is evaluated using the Gauss-Laguerre formula. If A1 < A,
then A1 is set equal to A to start with. A1 is adjusted by the function to achieve
the required accuracy. After execution, A1 will contain the final value used by
the function. A1 may need to be reset after every call to GAULAG. If the next
integral is similar to the previous one, then the value of A1 need not be reset,
since it is probably the required value. AEPS and REPS specify the required
absolute and relative accuracy. The calculations are terminated when the suc-
cessive values differ by less than max(AEPS, REPS × |RINT|). Convergence is
checked separately for the two integrals. DIF is an output parameter giving the
error as estimated by the function. NP is the output parameter containing the
number of function evaluations actually used. This function calls the functions
GAUSS and LAGURE to perform the integration over required subintervals.
Error status can be obtained from the returned value of GAULAG. A value of
30 implies that the function GAUSS failed to converge to the specified accuracy
over the interval [A, A1]. GAULAG = 37 denotes that the function LAGURE
failed to converge to the specified accuracy, even when A1 was raised to its
maximum permissible value of AMAX (= 50). This limit is provided to prevent
the function from getting into an infinite loop. This can happen if the integrand
is not in a form e−x f (x) as required for Gauss-Laguerre formula. If necessary,
the upper limit AMAX can be increased. GAULAG = 38 denotes that both
functions failed. In all these cases, RINT will contain the best estimate for the
integral and DIF should contain the estimated error. F is the name of the func-
tion used to calculate the value of the integrand. F2 is the name of the function
used to calculate F(X)*EXP(X), as required by the Gauss-Laguerre formula.
C.6. Integration 1037

Note that the function will find the integral

Z ∞ Z A1 Z ∞
F(x) dx = F(x) dx + e−x F2(x) dx . (C.12)
A A A1

Both Functions F(X), and F2(X) must be supplied by the user. The requirement
of two separate functions is to avoid the problem of overflow and underflow,
which may occur if F2 is formed by adding the exponential factor to F in-
side the function. If A1 − A < AEPS, then the first integral is not evaluated.
This condition may need to be changed if the integral can make significant
contribution, even though this condition is satisfied, which may happen if the
limits are very small. This function requires functions GAUSS and LAGURE.
It should be used only for those functions which fall off exponentially at large
X. Similar function can be written to handle singularities by combining the
Gauss-Legendre formula with another Gaussian formula for singular weight
functions.
R∞
46. LAGURE Function to compute integral A e−x F(x) dx using Gauss-
Laguerre formulae. The function computes the integral using different formulae,
until convergence is achieved or until the table of weights and abscissas is
exhausted. RINT is the output parameter containing the value of the integral. A
is the lower limit of the integral. REPS and AEPS specify the required accuracy,
while DIF is an output parameter giving the error as estimated by the function.
The convergence criterion used is |DIF| < max(REPS|RINT|, AEPS). NPT is
an output parameter containing the number of function evaluations actually
used. Error status can be obtained from the returned value of LAGURE. A value
of 30 implies that the function failed to converge to the specified accuracy. In
this case, RINT will contain the best estimate for the integral and DIF should
contain the estimated error. In such cases, we can try to use a higher value
of A. It may be noted that, depending on the function being integrated the
higher order formula may have problems, because of underflow and overflow on
computers with short range for the exponents. F is the name of the function
used to calculate the value of the integrand (multiplied by ex ). Function F(X)
must be supplied by the user. This function should be used for integrands which
fall off exponentially at large x. It will be better to use this function through
GAULAG, which adjusts the lower limit A to achieve the required accuracy.
R∞ 2
47. HERMIT Function to compute integral −∞ e−x F(x) dx using Gauss-
Hermite formulae. The function computes the integral using different formulae,
until convergence is achieved or until the table of weights and abscissas is ex-
hausted. RINT is the output parameter containing the value of the integral.
REPS and AEPS specify the required accuracy, while DIF is an output param-
eter giving the error as estimated by the function. The convergence criterion
used is |DIF| < max(REPS|RINT|, AEPS). NPT is an output parameter con-
taining the number of function evaluations actually used. Error status can be
obtained from the returned value of HERMIT. A value of 30 implies that the
function failed to converge to the specified accuracy. In this case, RINT will
1038 Appendix C. C Programs

contain the best estimate for the integral and DIF should contain the estimated
error. It may be noted that, depending on the function being integrated the
higher order formula may have problems, because of underflow and overflow on
computers with short range for the exponents. F is the name of the function
2
used to calculate the value of the integrand (multiplied by ex ). Function F(X)
must be supplied by the user. This function should be used for integrands which
2
fall off exponentially as e−x at large |x|. In this case it is not possible to divide
the range and hence it is not possible to apply any composite formulae or to
use a combination of formulae to improve accuracy.
48. GAULG2 Function to compute integral of a function with logarithmic
singularity over (0, a] using a combination of Gauss-Legendre and a Gaussian
formula with log(1/x) weight function. RINT is the output parameter contain-
ing the value of the integral. A is the upper limit of the integral. A1 is the
point at which the integral is broken, with the integral over [A1, A] being eval-
uated using Gauss-Legendre formula, while that over (0, A1] is evaluated using
a Gaussian formula with log(1/x) weight function. If A < A1, then A1 is set
equal to A to start with. A1 is adjusted by the function to achieve the required
accuracy. After execution, A1 will contain the final value used by the function.
A1 may need to be reset after every call to GAULG2. If the next integral is
similar to the previous one, then the value of A1 need not be reset, since it
is probably the required value. AEPS and REPS specify the required absolute
and relative accuracy. The calculations are terminated when the successive val-
ues differ by less than max(AEPS, REPS × |RINT|). This criterion is applied
separately to each part of the integral. DIF is an output parameter giving the
error as estimated by the function. NP is the output parameter containing the
number of function evaluations actually used. This function calls the functions
GAUSS and GAULOG to perform the integration over required subintervals.
Error status can be obtained from the returned value of GAULG2. A value of
31 implies that the function GAUSS failed to converge to the specified accuracy
over the interval [A1, A]. GAULG2 = 32 denotes that the function GAULOG
failed to converge to the specified accuracy, even when A1 was reduced to its
minimum permissible value of AMN (= 0.01). This limit is provided to prevent
the function from getting into an infinite loop. This can happen if the inte-
grand is not in a form f (x) log x as required for Gaussian formula. If necessary,
the lower limit AMN can be reduced. GAULG2 = 34 denotes that function
GAUSS failed while evaluating the integral over [0,A1]. In case of multiple fail-
ures, the second digit of GAULG2 will be the sum of these values. In all these
cases, RINT will contain the best estimate for the integral and DIF should
contain the estimated error. F is the name of the function used to calculate
the value of the integrand. F2 is the name of the function used to calculate
F(X)/LOG(1/X), as required by the Gaussian formula. Note that the function
will find the integral
Z A Z A Z A1 Z A1
F(x) dx = F(x) dx − log A1 F2(x) dx + F2(x) log(A1/x) dx .
0 A1 0 0
(C.13)
C.6. Integration 1039

Both the functions F(X), and F2(X) must be supplied by the user. The re-
quirement of two separate functions is to avoid possible problems caused by
singularity in integrand. The first two integrals should not have any singularity
and are evaluated using Gauss-Legendre formula, while the last integral is eval-
uated using Gaussian formula with logarithmic singularity. Note that the last
two integrals arise when the range in integral over (0, A1] is transformed to (0,
1]. If A − A1 < AEPS, then the first integral is not evaluated. This condition
may need to be changed if the integral can make significant contribution, even
though this condition is satisfied, which may happen if the limits are very small.
This function requires functions GAUSS and GAULOG. It should be used only
for those functions which have logarithmic singularity at x = 0.
49. GAULOG Function to compute integral
Z A
F(x) log(A/x) dx , (C.14)
0

using Gaussian formulae with log(A/x) weight function. The function computes
the integral using different formulae, until convergence is achieved or until the
table of weights and abscissas is exhausted. RINT is the output parameter
containing the value of the integral. A is the upper limit of the integral. REPS
and AEPS specify the required accuracy, while DIF is an output parameter
giving the error as estimated by the function. The convergence criterion used
is |DIF| < max(REPS|RINT|, AEPS). NPT is an output parameter containing
the number of function evaluations actually used. F is the name of the function
used to calculate the value of the integrand (divided by log A/x). Error status
can be obtained from the returned value of GAULOG. A value of 30 implies that
the function failed to converge to the specified accuracy. In this case, RINT will
contain the best estimate for the integral and DIF should contain the estimated
error. In such cases, we can try to use a lower value of A. Function F(X) must
be supplied by the user. This function should be used for integrands which
have a logarithmic singularity at x = 0. It would be better to use this function
through GAULG2 which adjusts the value of A to achieve required accuracy.
50. GAUSRC Function to calculate weights and abscissas of a Gaussian
quadrature formula, with arbitrary weight function. This function requires the
recurrence relation for the corresponding orthogonal polynomials. The recur-
rence relation is assumed to be in the form

Pj (x) = (aj x + bj )Pj−1 (x) − cj Pj−2 (x), (C.15)

where Pj (x) is the orthogonal polynomial of degree j in x. The coefficients

aj , bj , cj for j = 1, 2, . . . , n must be supplied. Here N is the number of abscissas
in the required Gaussian formula. Output parameters W and AB are arrays
of length N, containing the weights and corresponding abscissas. COF is an
array of length 3 × N containing the coefficients in the recurrence relation for
the polynomials. COF[i−1][0], COF[i−1][1] and COF[i−1][2] are respectively,
1040 Appendix C. C Programs

ai , bi , ci as defined above. These coefficients must be supplied. RI0 is the integral

Rb
a
w(x) dx for the required weight function over the corresponding interval. If
this integral is not known it must be evaluated using appropriate quadrature
formula before using this function to calculate the weights and abscissas. Error
status can be obtained from the returned value of GAUSRC. A value of 302 im-
plies that N ≤ 0, in which case no calculations are performed. GAUSRC = 321
implies that some coefficient becomes imaginary during calculation. This could
happen only if coefficients for recurrence relation are not specified correctly.
Other values may be set by TQL2 which is called to solve the eigenvalue prob-
lem. There is no error check on the output and the accuracy can be tested
by integrating the functions w(x)xj for j = 0, 1, . . . , 2N − 1, using the cal-
culated weights and abscissas. This function is in general better conditioned
than GAUSWT, which uses only the moments. This function requires function
TQL2.
51. GAULEG Function to calculate weights and abscissas of a Gauss-
Legendre quadrature formula. This function uses the recurrence relation for
Legendre polynomials

nPn (x) = (2n − 1)xPn−1 (x) − (n − 1)Pn−2 (x), (C.16)

to calculate the weights and abscissas using function GAUSRC. Here N is the
number of abscissas in the required Gaussian formula. Output parameters W
and A are arrays of length N, containing the weights and corresponding ab-
scissas. Error status can be obtained from the returned value of GAULEG.
Zero value implies successful execution. Nonzero values may be set by function
GAUSRC. There is no error check on the output and the accuracy can be tested
by integrating the functions xj for j = 0, 1, . . . , 2N − 1, using the calculated
weights and abscissas. This function requires functions GAUSRC and TQL2.
52. GAUJAC Function to calculate weights and abscissas of a Gauss-Jacobi
quadrature formula. This function uses the recurrence relation for Jacobi poly-
nomials

2n(n + α + β)(2n − 2 + α + β)Pnα,β (x) =

α,β
− 2(n − 1 + α)(n − 1 + β)(2n + α + β)Pn−2 (x)

α,β
+ (2n − 1 + α + β) α2 − β 2 + (2n + α + β)(2n − 2 + α + β)x Pn−1 (x),
(C.17)
to calculate the weights and abscissas using function GAUSRC. Here N is the
number of abscissas in the required Gaussian formula while ALP and BETA
are the indices α and β in (C.17). The corresponding weight function is w(x) =
(1 − x)α (1 + x)β on the interval (−1, 1). Output parameters W and A are
arrays of length N, containing the weights and corresponding abscissas. Error
status can be obtained from the returned value of GAUJAC. Zero value implies
successful execution. GAUJAC = 313 implies ALP ≤ −1 or BETA ≤ −1. Other
values may be set by function GAUSRC. There is no error check on the output
C.6. Integration 1041

and the accuracy can be tested by integrating the functions (1 − x)α (1 + x)β xj
for j = 0, 1, . . . , 2N − 1, using the calculated weights and abscissas. Setting
ALP = 0 and BETA = 0 in this function should yield the Gauss-Legendre
formula. Similarly, setting ALP = −1/2 and BETA = −1/2 should yield the
Gauss-Chebyshev formula. Using α = 0 one can calculate quadrature formulae
for algebraic singularity of form tβ , by transforming the lower limit to t = 0.
Z 1 Z 1
(1 − x)α (1 + x)β f (x) dx = 2α+β+1 (1 − t)α tβ f (2t − 1) dt . (C.18)
−1 0

Thus for integration over interval (0, 1) the weights W[I] should be divided by
21+α+β and abscissas should be (1+A[I])/2. This function requires functions
GAMMA, GAUSRC and TQL2.
53. LAGURW Function to calculate weights and abscissas of a Gauss-
Laguerre quadrature formula. This function uses the recurrence relation for
associated Laguerre polynomials

nLα

α 2 α
n (x) = (n+α) (2n−1+α)−x Ln−1 (x)−(n−1+α) (n+α)Ln−2 (x), (C.19)

to calculate the weights and abscissas using function GAUSRC. Here N is

the number of abscissas in the required Gaussian formula and ALP is the
index α in (C.19). The corresponding weight function is w(x) = xα e−x on
the interval (0, ∞). Output parameters W and A are arrays of length N, con-
taining the weights and corresponding abscissas. Error status can be obtained
from the returned value of LAGURW. Zero value implies successful execution.
LAGURW = 313 implies ALP ≤ −1. Other values may be set by function
GAUSRC. There is no error check on the output and the accuracy can be
tested by integrating the functions e−x xj+α for j = 0, 1, . . . , 2N − 1, using
the calculated weights and abscissas. For ALP = 0 it will yield the standard
Gauss-Laguerre quadrature formula. This function requires functions GAMMA,
GAUSRC and TQL2.
54. GAUHER Function to calculate weights and abscissas of a Gauss-Hermite
quadrature formula. This function uses the recurrence relation for Hermite
polynomials
Hn (x) = 2xHn−1 (x) − 2(n − 1)Hn−2 (x), (C.20)
to calculate the weights and abscissas using function GAUSRC. Here N is
the number of abscissas in the required Gaussian formula. The corresponding
2
weight function is w(x) = e−x on the interval (−∞, ∞). Output parameters
W and A are arrays of length N, containing the weights and corresponding
abscissas. Error status can be obtained from the returned value of GAUHER.
Zero value implies successful execution. Nonzero values may be set by function
GAUSRC. There is no error check on the output and the accuracy can be tested
2
by integrating the functions e−x xj for j = 0, 1, . . . , 2N−1, using the calculated
weights and abscissas. This function requires functions GAUSRC and TQL2.
1042 Appendix C. C Programs

55. GAUSWT Function to calculate weights and abscissas of a Gaussian

quadrature formula, with arbitrary weight function. This function requires
Rb
the values of moments i.e., the integrals FMOM(m) = a w(x)xm dx for
m = 0, 1, . . . , 2N − 1. Existence of such formulae should be ensured before
trying to use this function. If w(x) > 0 throughout the interval, then such
formulae should exist. Here N is the number of abscissas in the required Gaus-
sian formula. Output parameters W and AB are arrays of length N, containing
the weights and corresponding abscissas. FMOM is the name of the function
which calculates the moments as defined above. Function FMOM(M) must be
supplied by the user. QGAUS is a parameter which specifies the type of for-
mula required. If QGAUS = 1, then a Gaussian formula is found, otherwise
the function calculates the weights for an interpolatory formula with given
weight function. In this case, the array AB must contain the required abscis-
sas. Error status can be obtained from the returned value of GAUSWT. A
value of 303 implies that N ≤ 0, in which case no calculations are performed.
GAUSWT = 322 implies that the function GAUELM failed to find the coeffi-
cients of the required polynomial. GAUSWT = 323 implies that the function
POLYR failed to find zeros of the polynomial. Both these errors can occur only
for QGAUS = 1 and in both these cases, no further calculations are performed.
Similarly, GAUSWT = 324 implies that function GAUELM failed to calculate
the weights. In this case, the abscissas are already calculated, but the values
may not be reliable. GAUSWT = 322, 323, 324 will normally imply that either
the Gaussian formula does not exist or the weights and abscissas cannot be
calculated because of ill-conditioning. In latter case, improving the accuracy of
arithmetic may help. Even if GAUSWT = 0, the results could be unreliable. In
particular, the zeros of polynomial may be complex, but only the real part may
be returned in AB. This failure can occur either because of ill-conditioning, or
because the formula does not exist. This function requires functions POLYR
and LAGITR for solving the required polynomial, and function GAUELM for
solving a system of linear equations. The algorithm used here is rather ill-
conditioned and should not be used for large N, unless very high precision
arithmetic is being used. There is no error check on the output and the accu-
racy can be tested by integrating the functions w(x)xn for n = 0, 1, . . . , 2N − 1,
using the calculated weights and abscissas. If the recurrence relation for the cor-
responding orthogonal polynomial are known then GAUSRC should be used to
calculate the weights and abscissas.
56. FILON Function to evaluate integrals of the form
Z b Z b
f (x) sin kx dx or f (x) cos kx dx , (C.21)
a a

using Filon’s method. RI is an output parameter which will contain the value
of the integral. XL and XU are the lower and upper limits for the integral.
It is not essential to have XL < XU. RK is the coefficient k, multiplying x
in the sine or cosine function. QSIN is a parameter to determine the type of
C.6. Integration 1043

integral, QSIN = 1, implies that the integrand involves sin kx otherwise the
oscillatory factor is cos kx. AEPS and REPS specify the required absolute and
relative accuracy. The calculations are terminated when the successive values
differ by less than max(AEPS, REPS×|RI|). DIF is an output parameter which
should contain an estimate of (absolute) error calculated by the function. N is
an output parameter containing the number of function evaluations used. Error
status can be obtained from the returned value of FILON. A value of 30 implies
that the function has failed to converge to the specified accuracy. In this case,
RI will contain the best estimate for the integral and DIF should contain the
estimated error. The parameter NMAX in the function could be increased if
larger number of function evaluations are to be allowed. FUN is the name of the
function to calculate the non-oscillatory part f (x) in the integrand. Function
FUN(X) must be supplied by the user. The parameter THC is the critical value
of θ, below which the Taylor series expansion is used to evaluate the functions
α(θ), β(θ) and γ(θ). THC should be of the order of (100~)1/6 . The value here
is appropriate for 53-bit arithmetic. If other values are used, the errors could
be larger.
57. ADPINT Function to compute integral using adaptive technique based
on Gauss-Kronrod rule. RINT is the output parameter containing the value of
the integral. XL and XU are the lower and upper limits of the integral. It is not
essential to have XL < XU. REPS and AEPS specify the required accuracy,
while DIF is an output parameter giving the error as estimated by the function.
The function will attempt to ensure that DIF < max(|RINT| × REPS, AEPS).
However, it is not always possible to ensure that DIF is less than the expected
tolerance and in some cases even though the execution is successfully completed
DIF may be somewhat larger than the required accuracy. NPT is an output
parameter which will contain the number of function evaluations actually used.
NMAX is the maximum number of function evaluations that user is prepared
to allow. If NMAX ≤ 0, then NMAX will be set to a default value of MAXPT
(= 100000). Error status can be obtained from the returned value of ADPINT.
A value of 31 implies that the function failed to converge to the specified ac-
curacy on at least one subinterval. ADPINT = 32 denotes that this failure
occurred more than IFMAX (= 5) times. In this case the accuracy requirement
is adjusted by effectively increasing AEPS within the function. The value of pa-
rameter AEPS is not actually changed. This failure could occur if the function
has a very strong singularity at some points, or if the accuracy requirement is
too high. In the former case, increasing IPMAX in the function may help, pro-
vided the singularity is at x = 0, since otherwise it is not possible to subdivide
the interval if it spans two consecutive numbers in the machine representation.
It may be better to shift the singularity to x = 0. ADPINT = 325 denotes that
the function failed to achieve satisfactory accuracy in NMAX function evalua-
tions. This situation may occur either because of singularity, or because very
high accuracy is required. Increasing NMAX or REPS may help in such cases.
In all cases, RI will contain the best estimate for integral and DIF should con-
tain the estimated error. If the integrand has singularities, then this estimate
1044 Appendix C. C Programs

may be unreliable. F is the name of the function used to calculate the value
of the integrand. Function F(X) must be supplied by the user. This function
also needs the function KRONRD (or GAUS16). This function could be used
for mildly singular functions or for functions which vary by several orders of
magnitude over the interval of integration. However, if very large interval is
given and the function is almost constant over most of the interval, spurious
convergence may take place as the function may fail to detect the variation
in function value. In such cases integration should be calculated over smaller
subintervals and sum can be calculated to get the total integral.
58. KRONRD Function to compute integral using Gauss-Kronrod formula.
RI is the output parameter containing the value of the integral. A and B are
the lower and upper limits of the integral. It is not essential to have A < B.
DIF is an output parameter giving the error, as estimated by the function.
The function uses 7-point Gauss-Legendre formula and its 15-point Kronrod
extension to calculate the integral. RI will contain the estimate using 15-point
Kronrod formula, while DIF contains the magnitude of the difference between
the two values. N is an output parameter which will contain the number of
function evaluations actually used. F is the name of the function to calculate
the integrand. Function F(X) must be supplied by the user. The returned value
of KRONRD is always zero. This function is used by ADPINT.
59. GAUS16 Function to compute integral using 16 point Gauss-Legendre
formula. This function can be used instead of KRONRD with ADPINT for
adaptive integration. It calculates the integral using 8 and 16 point Gaussian
formula and the difference between the two estimates is considered as estimated
error. This function will be less efficient as compared to KRONRD as it requires
24 function evaluations as opposed to 15 required by KRONRD. Although, the
16 point Gauss-Legendre formula will have higher accuracy as compared to
15 point Gauss-Kronrod formula, the error estimate is essentially the error in
8-point Gauss-Legendre formula and it will not be much different from that
obtained by KRONRD. Since the adaptive choice of subintervals is determined
by the estimated error, it will be almost same for both these functions. Thus
in general KRONRD should be preferred but this function is provided as an
alternative. RI is the output parameter containing the value of the integral
obtained using 16 point formula. A and B are the lower and upper limits of
the integral. It is not essential to have A < B. DIF is an output parameter
giving the error, as estimated by the function. The function uses 8 and 16-
point Gauss-Legendre formulae to calculate the integral. RI will contain the
estimate using 16-point Gaussian formula, while DIF contains the magnitude
of the difference between the two values. N is an output parameter which will
contain the number of function evaluations actually used. F is the name of the
function to calculate the integrand. Function F(X) must be supplied by the
user. The returned value of GAUS16 is always zero. This function is used by
ADPINT.
C.6. Integration 1045

60. CAUCHY Function to compute the Cauchy principal value of an integral.

It uses ADPINT to perform the integral as explained in Section 6.6.10. RI is the
output parameter containing the value of the integral. A and B are the lower
and upper limits of the integral. It is essential to have A < B. C is the point
inside the interval [A, B] where the integrand is singular. REPS and AEPS
specify the required accuracy, while DIF is an output parameter giving the
error as estimated by the function. The function will attempt to ensure that
DIF < max(|RINT| × REPS, AEPS). F is the name of the function used to
calculate the value of the integrand. FUNP is the name of the function used to
calculate F(C + X) + F(C − X). Error status can be obtained from the returned
value of CAUCHY. A value of 304 implies that A > B, A > C or C > B. In this
case no calculations are done. Other values may be set by function ADPINT,
which is called twice to perform the integration. The returned value is set to
IER1+2IER2, where IER1 and IER2 are the values returned by the two calls to
ADPINT. In these cases DIF will contain the estimated error in the computed
value of RI. NPT is an output parameter which will contain the number of
function evaluations actually used. This function needs functions ADPINT and
KRONRD. Functions F(X) and FUNP(X) must be supplied by the user. The
value of C is available through global variable CAU CC. There is no provision
to pass on the name of the function F(X) to FUNP(X) and if needed it will
have to be put explicitly. If F(C + X) + F(C − X) can be simplified to remove
singularity, the roundoff errors will be reduced.
61. EULER Function to compute the sum of an alternating series using Eu-
ler’s transformation.
N−1
X
SUM = A0 × (−1)i TERM(i). (C.22)
i=0

The series may be finite or infinite. Before applying the Euler transformation,
it is advisable to sum the first few terms (and last few terms, if the series is
finite) separately as it helps in improving the convergence. This function sums
the first M1 terms and the last M2 terms separately and applies the Euler
transform to the remaining terms from M1 to N − M2 − 1. The differences
are calculated until the sum of each part converges to the specified accuracy.
For finite series the contribution from the upper end is calculated separately.
At most NMAX differences are calculated. If the sum does not converge then
another attempt may be made by increasing M1 or M2. Although NMAX can
be trivially increased, but if the convergence is slow the error estimate will
be unreliable and it will be better to improve convergence, which is generally
achieved by increasing M1 or M2. Increasing NMAX may not help, as in most
cases the roundoff error will dominate in higher order differences. N is the
number of terms to be summed, while M1, M2 are the number of terms at the
two ends which need to be summed separately. For summing an infinite series
use N = 0, in which case M2 is ignored. A0 specifies the sign of the first term.
The sum assuming the first term to be positive is multiplied by A0. REPS and
1046 Appendix C. C Programs

AEPS specify the required relative and absolute accuracy. The estimated error
should be less than max(AEPS, REPS × |SUM|). DIF is the output parameter
which gives the estimated error in the computed value of the sum. In most
cases, actual error will be somewhat larger than DIF and can be estimated by
repeating the calculations with different values of M1 and M2. N1, N2 are the
number of terms from both ends actually used by the function. M2 and N2 are
relevant only for finite series. SUM is the calculated value of the sum. TERM
is the name of the function to calculate the terms of the series. Error status
can be obtained from the returned value of EULER. A value of 31 implies that
M1 + M2 + 2NMAX ≥ N, in which case M1 is set to N and the series is summed
directly. In this case although Euler transform is not used the sum should be
exact apart from roundoff errors. EULER = 32 implies that the series obtained
after Euler transform did not converge to required accuracy at the lower end.
EULER = 34 implies that the series obtained after Euler transform did not
converge to required accuracy at the upper end. This is relevant only for finite
series. EULER = 36 implies that the series obtained after Euler transform did
not converge to required accuracy at both ends. Function TERM(I) must be
supplied by the user to calculate the Ith term of the series apart from the sign
as in Eq. (C.22). Note that in keeping with the C convention the terms are
numbered from 0 to N − 1. This must be accounted for in writing function
TERM(I).
62. BSPQD2 Function to compute integral of an expansion in terms of B-
spline basis functions in two dimensions over a rectangular region. The expan-
sion may be obtained by interpolating or approximating a table of values. Before
calling this function, the coefficients of B-spline expansion must be calculated
by a call to the function BSPINT2.
Z XU Z YU NX+K−3
X NY+K−3
X
BSPQD2 = dx dy WT[j][i] φi (x)ψj (y),
XL YL i=0 j=0
(C.23)
where φi (x) are the B-spline basis functions of order K along x, and ψj (y) are
those along y. The integral is evaluated recursively. First the integral along
x is evaluated for each value of j and then the integral over y is calculated.
It may be noted that the function BSPQD to evaluate the integral in one
dimension is not called recursively. NX, NY are the number of knots along
x and y respectively. X and Y are arrays of length NX and NY containing
the knots for B-splines. The knots must be in ascending order with X[0], Y[0]
containing the first knot in respective directions. K is the order of B-splines,
K = 2 gives linear B-splines, while K = 4 yields cubic B-splines, etc. WT is
an array of length IW × (NY + K − 2) containing the coefficients of expansion
in terms of B-spline basis functions. These coefficients may be calculated using
BSPINT2 for interpolation in a table of values. IW is the second dimension
of array WT as declared in the calling function. XL and XU are the lower
and upper limits of the integral along x. YL and YU are the lower and upper
limits of the integral along y. IER is the error parameter. IER = 0 implies
C.6. Integration 1047

successful execution of the function. Nonzero values may be set by BSPQD

which is called to perform integration in one dimension. This function requires
function BSPLIN to evaluate the B-spline basis functions, and function BSPQD
for integration in one dimension, while function BSPINT2 will be required to
calculate the coefficients of B-spline before calling this function.
63. BSPQDN Function to compute integral of an expansion in terms of B-
spline basis functions in n dimensions over a hyper-rectangular region. The
expansion may be obtained by interpolating or approximating a table of val-
ues. Before calling this function, the coefficients of B-spline expansion must
be calculated by a call to the function BSPINTN. This function evaluates the
integral
Z XU[0] Z XU[N−1] NK[0]+K−3 NK[N−1]+K−3
X X
dx0 · · · dxN −1 ···
XL[0] XL[N−1] i0 =0 iN −1 =0 (C.24)
(0) (N −1)
WT[iN−1 ] . . . [i0 ] φi0 (x0 ) · · · φiN −1 (xN −1 )
(j)
where φi (x) are the B-spline basis functions of order K along xj . The integral
is evaluated recursively. First the integral along x0 is evaluated for each value
of other indices and then the integral over x1 is calculated and so on. It may
be noted that the function BSPQD to evaluate the integral in one dimension is
not called recursively. N is the number of dimensions. NK is an integer array of
length N containing the number of knots along each direction. NK[I] is the num-
ber of knots along Ith dimension. X is an array of length NXD × N containing
the knots for B-splines along each dimension. The knots must be in ascending
order. X[J][I] is the Ith knot along Jth dimension. It may be noted that the
function BSPINTN gives the list of knots in the same format and that array
can be directly used in this function. NXD is the second dimension of the array
X as specified in the calling function, NXD ≥ max(NK[i]). K is the order of B-
splines, K = 2 gives linear B-splines, while K = 4 yields cubic B-splines, etc. WT
is an array of length (NK[0] + K − 2)(NK[1] + K − 2) · · · (NK[N − 1] + K − 2)
containing the coefficients of expansion in terms of B-spline basis functions.
These coefficients may be calculated using BSPINTN for interpolation in a
table of values. The coefficients are assumed to be stored in natural Fortran
order with no gaps in data. In the calling function the array should have dimen-
sion WT[NK[N−1]+K−2]. . .[NK[1]+K−2][NK[0]+K−2]. Alternately, it can be
treated as a one dimensional array of required length. XL and XU are arrays of
length N containing the lower and upper limits of the integral along each direc-
tion. XL[I] is the lower limit and XU[I] is the upper limit for integration along
Ith dimension. IER is the error parameter. IER = 0 implies successful execu-
tion of the function. Nonzero values may be set by BSPQD which is called to
perform integration in one dimension. This function requires function BSPLIN
to evaluate the B-spline basis functions, and function BSPQD for integration
in one dimension, while function BSPINTN will be required to calculate the
coefficients of B-splines before calling this function.
1048 Appendix C. C Programs

64. MULINT Function for integration over a hyper-rectangle in N dimensions

using a series of product Gauss-Legendre formulae. A and B are arrays of length
N, with A[I] and B[I] specifying the lower and upper limits for the Ith variable.
It is not essential to have A[I] < B[I]. N is an integer variable specifying the
number of dimensions. There is no limit on the value of N as far as this func-
tion is concerned, but the number of function evaluations required to evaluate
the integral for large N will put some limit. M is an integer array of length
N, specifying which Gaussian formula is to be used along each axis in the first
approximation. The function will attempt to use M[I]-point Gauss-Legendre
formula along the Ith axis. Formulae with 2, 4, 8, 16 and 32 points are incorpo-
rated in the function. If any other value is specified for M[I], then M[I] will be
set to the default value of 2. IND is an integer array of length N, specifying the
number of subdivisions to be used for the first approximation along each axis. If
IND[I] < 1, then IND[I] is set to a default value of 1. After execution, the arrays
M and IND will contain the final values of the formula and number of subinter-
vals used along each axis. M[I] and IND[I] are increased to achieve the required
accuracy or until the maximum number of function evaluations is reached. F
is the name of the function to calculate the value of integrand at any given
point. Function F(N, X) must be supplied by the user. (Here the first argument
specifies the number of dimensions, while the second argument is an array of
length N, specifying the coordinates of the point at which the integrand is to
be evaluated.) RINT is the output parameter which will contain the value of
the integral. REPS and AEPS specify the accuracy required by user while DIF
is an output parameter containing the estimated (absolute) error in RINT. The
function will attempt to ensure that DIF < max(|RINT|×REPS, AEPS). NUM
is an output parameter containing the number of function evaluations actually
used. MAXPT is the maximum limit on the number of function evaluations
which the user is prepared for. If MAXPT < 1, then it is set to a default value
of MAXPTS (= 1100000). The value of MAXPT will depend on the computer
time that is available and the accuracy required. Error status can be obtained
from the returned value of MULINT. A value of 39 implies that the integral
failed to converge to the required accuracy in MAXPT function evaluations. In
this case, RINT will contain the best approximation to the integral and DIF
will contain the error estimate. The error estimate may not be very reliable
as the number of points may not be sufficient to check for convergence along
each dimension. If a second attempt is to be made with larger MAXPT, then
the values of M[I] and IND[I] should not be reset, since that will save some
function evaluations in computing initial approximations, which are known to
be unsatisfactory. MULINT = 305 implies that N < 1. MULINT = 307 implies
that the number of points exceeded MAXPT in the first attempt itself. In such
cases, no approximation for the integral will be available. This function requires
function NGAUSS.
65. NGAUSS Function for integration over a hyper-rectangle in N dimensions
using a product Gauss-Legendre formula. A and B are arrays of length N,
with A[I] and B[I] specifying the lower and upper limits for the Ith variable.
C.6. Integration 1049

It is not essential to have A[I] < B[I]. N is an integer variable specifying the
number of dimensions. M is an integer array of length N, specifying the Gaussian
formula to be used along each axis. The function attempts to use M[I]-point
Gauss-Legendre formula along the Ith axis. Formulae with 2, 4, 8, 16 and 32
points are incorporated in the function. If any other value is specified for M[I],
then no calculations will be attempted. IND is an integer array of length N,
specifying the number of subdivisions to be used along each axis. If IND[I] <
1, then no calculations will be performed. F is the name of the function to
calculate the value of integrand at any given point. The Function F(N, X)
must be supplied by the user. (Here the first argument specifies the number
of dimensions, while the second argument is an array of length N specifying
the coordinates of the point at which the integrand is to be evaluated.) RI is
the output parameter which will contain the value of the integral. NUM is an
output parameter containing the number of function evaluations actually used.
MAXPT is the maximum limit on the number of function evaluations which
the user is prepared for. If MAXPT < 1, then it is set to a default value of
MAXPTS (= 1100000). The value of MAXPT will depend on the computer
time that is available and the accuracy required. Error status can be obtained
from the returned value of NGAUSS. A value of 305 implies that N < 1.
NGAUSS = 306 implies that either M[J] is not 2, 4, 8, 16 or 32 or INT[J] <
1 for some J. NGAUSS = 307 implies that the number of points exceeded
MAXPT. In all these cases no calculations are performed. This function is
called by MULINT which adjusts the number of abscissas to be used along
each dimension depending on the required accuracy. It can also be used directly
to calculate the integral using quadrature formulae with specified number of
points in each dimension.

66. SPHND Function to transform the coordinates from hyper-spherical to

Cartesian in n-dimensions. It can be used for integration over hyper-spherical
shell in n-dimensions using function MULINT, STRINT, MCARLO or EQUIDS
when the function is known in terms of Cartesian coordinates. For this purpose
SPHND should be passed on as the name of the function to calculate the
integrand. The limits A[I] and B[I] of integration along Ith dimension should
be [0, π] for I = 1, . . . , N − 2, [0, 2π] for I = N − 1 and [rl , ru ] for I = 0. Here
rl and ru are the radial coordinates of the hyper-spherical shell over which
integration is required. For integration over a hyper-sphere of radius R, rl = 0
and ru = R. This function gives the transformation between hyper-spherical
coordinates xi and Cartesian coordinates yi

y0 = x0 cos(x1 ),
y1 = x0 sin(x1 ) cos(x2 ),
y2 = x0 sin(x1 ) sin(x2 ) cos(x3 ), (C.25)
...
yn−2 = x0 sin(x1 ) sin(x2 ) · · · sin(xn−2 ) cos(xn−1 ),
yn−1 = x0 sin(x1 ) sin(x2 ) · · · sin(xn−2 ) sin(xn−1 ).
1050 Appendix C. C Programs

The integral can be written as

Z
f (x) dV
Sn
Z ru Z π Z π Z 2π n−2
Y
= dx0 dx1 · · · dxn−2 dxn−1 f (x)xn−1
0 sinn−i−1 xi .
rl 0 0 0 i=1
(C.26)
The function f (x) is calculated after transformation to Cartesian coordinates
and the volume element as given above is multiplied. N is the number of dimen-
sions, X and Y are arrays of length N containing the coordinates of required
point. X[I] are the input hyper-spherical coordinates, while Y[I] are the Carte-
sian coordinates. Function FUNSPH(N, Y) is called to calculate the required
function in Cartesian coordinates. Since there is no provision to pass on the
name of this function, it has to be the same as what appears in the routine.
67. STRINT Function for integration over a hyper-rectangle in N dimensions
using compound monomial rules of degree 1, 3 or 5. A and B are arrays of length
N, with A[I] and B[I] specifying the lower and upper limits for the Ith variable.
It is not essential to have A[I] < B[I]. N is an integer variable specifying the
number of dimensions. There is no limit on the value of N as far as this function
is concerned, but the function STROUD which is invoked to actually evaluate
the integrals will have some limit. If this limit is exceeded, error flag will be set.
Although a large limit on N is allowed the result is unlikely to be reliable for very
large values of N. M is an integer parameter specifying the formula to be used on
each subdivision. The allowed values of M are 1, 3 and 5. M = 1 selects the one-
point formula of degree one (essentially a generalisation of the midpoint rule).
M = 3 selects the 2N-point formula of degree 3 due to Stroud. M = 5 selects the
(2N2 + 1)-point formula of degree 5. If M is not one of these values, it will be
set to the default value of 3. IND is an integer array of length N, specifying the
number of subdivisions to be used for the first approximation along each axis.
If IND[I] < 1, then it will be set to a default value of 1. After execution, the
array IND will contain the final values of the number of subintervals used along
each axis. F is the name of the function to calculate the value of the integrand
at any given point. Function F(N, X) must be supplied by the user. (Here the
first argument specifies the number of dimensions, while the second argument
is an array of length N, specifying the coordinates of the point at which the
integrand is to be evaluated.) RINT is the output parameter which will contain
the value of the integral. REPS and AEPS specify the accuracy required by
user, while DIF is an output parameter containing the estimated (absolute)
error in RINT. The function will attempt to ensure that DIF < max(|RINT| ×
REPS, AEPS). NUM is an output parameter containing the number of function
evaluations actually used. MAXPT is the maximum limit on the number of
function evaluations which the user is prepared for. If MAXPT < 1, then it
is set to a default value of MAXPTS (= 1000000). The value of MAXPT
will depend on the computer time that is available and the accuracy required.
C.6. Integration 1051

Error status can be obtained from the returned value of STRINT. A value of 39
implies that the integral failed to converge to the required accuracy in MAXPT
function evaluations. In this case, RINT will contain the best approximation to
the integral and DIF will contain the error estimate. The error estimate may
not be very reliable as the number of points may not be sufficient to check
for convergence along each dimension. If a second attempt is to be made with
larger MAXPT, then the value of IND[I] should not be reset, since that will
save some function evaluations in computing initial approximations, which are
known to be unsatisfactory. STRINT = 308 implies that the number of points
exceeded MAXPT in the first attempt itself. In such cases, no approximation
for the integral will be available. STRINT = 309 implies that N is outside the
limits accepted by STROUD. This function requires function STROUD.
68. STROUD Function for integration over a hyper-rectangle in N dimensions
using a compound monomial rule of degree 1, 3 or 5. A and B are arrays of
length N, with A[I] and B[I] specifying the lower and upper limits for the Ith
variable. It is not essential to have A[I] < B[I]. N is an integer variable specifying
the number of dimensions. The value of N must be between 1 and 50. If N is
outside these limits error flag will be set. M is an integer specifying the formula
to be used on each subdivision. The allowed values of M are 1, 3 and 5. M = 1
selects the one-point formula of degree one (essentially a generalisation of the
midpoint rule). M = 3 selects the 2N-point formula of degree 3 due to Stroud.
M = 5 selects the (2N2 + 1)-point formula of degree 5. If any other value is
specified for M, then no calculations will be attempted. IND is an integer array
of length N, specifying the number of subdivisions to be used along each axis.
If IND[I] < 1, then no calculations will be performed. F is the name of the
function to calculate the value of integrand at any given point. The function
F(N, X) must be supplied by the user. (Here the first argument specifies the
number of dimensions, while second argument is an array of length N specifying
the coordinates of the point at which the integrand is to be evaluated.) RI is
the output parameter which will contain the value of the integral. NUM is an
output parameter containing the number of function evaluations actually used.
MAXPT is the maximum limit on the number of function evaluations which
the user is prepared for. If MAXPT < 1, then it is set to a default value of
MAXPTS (= 1000000). The value of MAXPT will depend on the computer
time that is available and the accuracy required. Error status can be obtained
from the returned value of STROUD. A value of 308 implies that the number of
points exceeded MAXPT. STROUD = 309 implies that N < 1 or N > NMAX.
STROUD = 310 implies that M is not 1, 3 or 5, or IND[I] < 1 for some I.
In all these cases and no calculations are performed. This function is called by
STRINT which adjusts the number of abscissas to be used along each dimension
depending on the required accuracy.
69. MCARLO Function for integration over a hyper-rectangle in N dimen-
sions using Monte Carlo method. A and B are arrays of length N, with A[I]
and B[I] specifying the lower and upper limits for the Ith variable. It is not
1052 Appendix C. C Programs

essential to have A[I] < B[I]. N is an integer variable specifying the number of
dimensions. NPT is the maximum number of function evaluations to be used
for integration. F is the name of the function to calculate the value of the inte-
grand at any given point. The function F(N, X) must be supplied by the user.
(Here the first argument specifies the number of dimensions, while the second
argument is an array of length N specifying the coordinates of the point at
which the integrand is to be evaluated.) RI is an output parameter, which will
contain the value of the integral. REPS and AEPS specify the desired accuracy.
ERR is an output parameter containing the (absolute) error estimate for the
integral. This error estimate is obtained
√ by using the variance σ as defined in
Section 6.11. ERR is set to 2.576σ/ NP corresponding to a confidence level of
99%. The function attempts to ensure that ERR < max(|RI| × REPS, AEPS).
NP is the number of function evaluations actually used. Error status can be
obtained from the returned value of MCARLO. A value of 39 implies that the
function failed to converge to the specified accuracy. In this case, RI will contain
the best estimate and ERR the corresponding error estimate. MCARLO = 311
implies that N < 1 and no calculations are done. This function requires function
RANF(ISEED) to generate random numbers. If a different function is used for
this purpose, then the value of seed may have to be changed.
70. RAN1 Function to generate a sequence of random numbers using a simple
linear congruential method. The constants AM, A and AC will give overflow
in integer arithmetic using 32-bits and hence these must be treated as double
precision variables. SEED could be set to any positive value less than AM,
before first call to the function. After the first call, this variable should not
be changed in any other function. It may be noted that this function requires
a real number as seed as opposed to RANF, which needs an integer seed as
argument. Further, in this case the seed has to be positive, while for RANF it
has to be negative for initialisation. This has to be kept in mind while changing
the random number generator in any function.
71. RANF Function to generate a sequence of random numbers using a com-
bination of three linear congruential generators. This function is based on the
function RAN1 in Press et al. (2007). The constants m, a and c for each of these
generators are chosen, such that there should be no overflow on a machine with
32-bit word length. ISEED could be set to any negative value, before first call
to the function. After the first call, this variable should not be changed in any
other function.
72. EQUIDS Function for integration over a hyper-rectangle in N dimensions
using the method based on equidistributed sequences. A and B are arrays of
length N, with A[I] and B[I] specifying the lower and upper limits for the Ith
variable. It is not essential to have A[I] < B[I]. N is an integer variable specifying
the number of dimensions. N should be between 1 and NMAX (= 21). If N is
outside these limits, then error flag will be set. NPT is the maximum number
of function evaluations to be used in computing the integral. F is the name
of the function to calculate the value of the integrand at any given point.
C.7. Nonlinear Algebraic Equations 1053

The function F(N, X) must be supplied by the user. (Here the first argument
specifies the number of dimensions, while the second argument is an array of
length N specifying the coordinates of the point at which the integrand is to be
evaluated.) S1 and S2 are the output parameters which will contain the value of
the integral. These values provide the two approximations using equidistributed
sequences. If the function is sufficiently smooth, S2 is expected to be a better
approximation and could in general be used as the best approximation to the
integral. REPS and AEPS specify the accuracy required by the user, while DIF
is an output parameter containing the estimated (absolute) error in S2. The
function attempts to ensure that DIF < max(|S2| × REPS, AEPS). NP is an
output parameter containing the number of function evaluations actually used.
Error status can be obtained from the returned value of EQUIDS. A value of
39 implies that the integral failed to converge to the required accuracy in NPT
function evaluations. In this case, S2 (or S1) will contain the best approximation
to the integral and DIF will contain the error estimate. Evaluation of the sum
in EQUIDS involves large roundoff errors and it is recommended to use this
function with double precision. EQUIDS = 312 implies that N < 1 or N >
NMAX.

C.7 Nonlinear Algebraic Equations

73. BISECT Function to find a real zero of a continuous function using the
method of bisection. XL and XU bracket the interval containing the zero. The
function must have opposite signs at these two points. This interval will be
refined by bisection, and after execution XL and XU will contain the refined
estimate for the interval containing the zero. It is not necessary to have XL <
XU. X is an output parameter which will contain the interpolated value of
the zero, using the last estimate at the end points XL and XU. NB is an
input parameter specifying the number of bisections to be performed. F is
the name of the function routine to calculate the function. Error status can
be obtained from the returned value of BISECT. A value of 401 implies that
NB ≤ 0, in which case, no calculations are performed. BISECT = 421 implies
that the function has the same sign at both the end points and hence bisection
cannot be performed. BISECT = −1 implies that the function vanishes at
one of the points. Hence, the required number of bisections have not been
performed, but in this case, X will contain the “zero”. Before invoking this
function, it must be ensured that the function is continuous in the interval.
In particular, the sign change should not be due to a singularity within the
interval. The number of bisections to be performed should be carefully chosen,
since performing unnecessary bisections will not improve the accuracy of the
computed zero. The function F(X) must be supplied by the user.
74. SECANT To calculate a real zero of a given function using the secant
iteration. XL and XU specify the limits within which the zero is expected. It is
essential to have XL < XU i.e., XL is the lower limit and XU is the upper limit.
1054 Appendix C. C Programs

It is not necessary that the function value should have opposite sign at these two
points. The limits are only used to terminate the iteration, if it is going astray.
X0 is the initial guess for the zero and it is essential to have XL ≤ X0 ≤ XU. X
is an output parameter containing the computed value of the zero. REPS and
AEPS specify the relative and absolute convergence criterion. The zero should
be determined with an accuracy of max(AEPS, REPS|X|). FUN is the name of
the external routine to calculate the required function. Function FUN(X) must
be supplied by the user. Error status can be obtained from the returned value of
SECANT. A value of 40 implies that the calculated values of the function at the
two most recent points was equal, and it is not possible to continue the iteration
further. This situation can arise if the zero is multiple, or if the convergence
criterion is too stringent, or sometimes just by coincidence. In the first two
cases, the computed value of the zero may be reasonable. SECANT = 402
implies that the starting value was outside the given interval, in which case,
no calculations are performed. SECANT = 422 implies that the iteration has
gone outside the prescribed limits, and was terminated at that stage. This
situation may arise, because the zero is outside or close to the specified limits.
However, this is not essential and the iteration may tend to go outside for
other reasons. SECANT = 423 implies that the iteration failed to converge to
the specified accuracy. This failure could be due to various reasons, like very
stringent convergence criterion, or excessive roundoff error in calculating the
function, or the iteration simply failed to detect the zero. This function attempts
to make a reasonable guess for the second starting value, using the limiting
interval and the convergence criterion. However, if the root is too large, or too
small, or if the bounding interval is very large this choice may not be good. In
such cases, the iteration may not converge. This problem can be rectified by
choosing the proper increment DX at the first step, or by making the bounding
interval smaller. If the function is calculated in the form FUN × 2JF to avoid
overflows and underflows, then use the function SECAN 2 instead of SECANT.
The call statement is identical for the two functions.
75. SECANI To calculate a real zero of a given function using the secant
iteration. The function is calculated in the form F × 2JF to avoid overflows
and underflows. It also uses the reverse communication technique for passing
function values. This function will return control to the calling function when
it needs a function evaluation. In that case, the error parameter IER will be
set to a negative value and the function should be evaluated at X. The value of
the function should be returned in the variables F and JF. During the function
evaluation variables other than F and JF in the call statement should not be
disturbed. Before the first call, IER should be set to zero. IER ≥ 0 implies that
the execution is complete and no more function evaluations are required. In
that case the value of X will give the root (if IER = 0) or the final value of X
where the function was evaluated. XL and XU specify the limits within which
the zero is expected. It is essential to have XL < XU i.e., XL is the lower limit
and XU is the upper limit. It is not necessary that the function value should
have opposite sign at these two points. The limits are only used to terminate
C.7. Nonlinear Algebraic Equations 1055

the iteration, if it is going astray. X0 is the initial guess for the zero and it is
essential to have XL ≤ X0 ≤ XU. X is an output parameter containing the
value of x at which the function value is required, when IER < 0. For other
values of IER it will contain the computed root or a failed approximation to it.
The calculated function value is to be passed through the variables F and JF.
REPS and AEPS specify the relative and absolute convergence criterion. The
zero should be determined with an accuracy of max(AEPS, REPS|X|). IER is
the error parameter. If IER < 0, then a fresh function evaluation is required.
IER = 0 implies that function execution has been successfully completed and
X should contain the calculated root. IER = 40 implies that the calculated
values of the function at the two most recent points was equal, and it is not
possible to continue the iteration further. This situation can arise if the zero
is multiple, or if the convergence criterion is too stringent, or sometimes just
by coincidence. In the first two cases, the computed value of the zero may be
reasonable. IER = 402 implies that the starting value was outside the given
interval, in which case, no calculations are performed. IER = 422 implies that
the iteration has gone outside the prescribed limits, and was terminated at
that stage. This situation may arise, because the zero is outside or close to
the specified limits. However, this is not essential and the iteration may tend
to go outside for other reasons. IER = 423 implies that the iteration failed
to converge to the specified accuracy. This failure could be due to various
reasons, like very stringent convergence criterion, or excessive roundoff error in
calculating the function, or the iteration simply failed to detect the zero. This
function attempts to make a reasonable guess for the second starting value,
using the limiting interval and the convergence criterion. However, if the root
is too large, or too small, or if the bounding interval is very large this choice
may not be good. In such cases, the iteration may not converge. This problem
can be rectified by choosing the proper increment DX at the first step, or by
making the bounding interval smaller.
76. NEWRAP To calculate a real zero of a given function using the Newton-
Raphson method. This function tries to estimate the multiplicity of the zero by
looking at the convergence rate and then uses this estimate to accelerate the
convergence. XL and XU specify the limits within which the zero is expected.
It is essential to have XL < XU i.e., XL is the lower limit and XU is the
upper limit. It is not necessary that the function should have opposite signs
at these two points. The limits are only used to terminate the iteration, if it
is going astray. X0 is the initial guess for the zero and it is essential to have
XL ≤ X0 ≤ XU. X is an output parameter containing the computed value of the
zero. REPS and AEPS specify the relative and absolute convergence criterion.
The zero should be calculated with an accuracy of max(AEPS, REPS|X|). FUN
is the name of the external routine to calculate the required function and its
derivative. Function FUN(X, DF) must be supplied by the user, where DF
is the first derivative of FUN at X. Error status can be obtained from the
returned value of NEWRAP. A value of 403 implies that the starting value
was outside the given range, in which case, no calculations are performed.
1056 Appendix C. C Programs

NEWRAP = 424 implies that the iteration has gone outside the prescribed
limits, and was terminated at that stage. This situation may arise, because the
zero is outside or close to the specified limits. However, this is not essential and
iteration may tend to go outside for other reasons. NEWRAP = 425 implies
that the iteration failed to converge to the specified accuracy. This failure could
be due to various reasons, like very stringent convergence criterion, or excessive
roundoff error in calculating the function, or the iteration simply failed to detect
the zero. NEWRAP = 426 implies that the calculated value of the derivative
at the last point is zero, and it is not possible to continue the iteration further.
In this case the zero is detected to be simple. This can happen by coincidence.
This situation can happen if the zero is multiple, in which case NEWRAP is
set to 126 − 100 × multiplicity. In this case, the computed value of the zero may
be reasonable. On successful completion of the function, the returned value of
NEWRAP will be zero if the root is simple and NEWRAP = −k, if the root
is detected to be multiple with multiplicity k. Function FUN(X, DF) must be
supplied by the user.
77. BRENT To calculate a real zero of a given function using the Brent’s
method. This function is based on the procedure given by Brent (1973). A and
B specify the limits within which the zero is located. It is essential that the
function has opposite signs at these two points. The values of A and B will
be updated by the function to locate the zero with required accuracy. X is
the output parameter containing the computed value of the zero. REPS and
AEPS specify the relative and absolute convergence criterion. The zero should
be calculated with an accuracy of max(AEPS, REPS|X|). F is the name of the
external routine to calculate the required function. Function F(X) must be
supplied by the user. Error status can be obtained from the returned value of
BRENT. A value of 427 implies that the function has the same sign at both the
end points and hence the Brent’s method cannot be applied. BRENT = 428
implies that the iteration failed to converge to the specified accuracy. This fail-
ure could be due to the fact that the convergence criterion is too stringent.
Increasing the parameter NIT in the function may allow convergence, but be-
fore doing that it must be ensured that it will be meaningful to do so. Since
convergence to any arbitrary accuracy does not ensure that the zero is correct
to that accuracy. If the accuracy requirement cannot be satisfied within the
available floating-point arithmetic, then the iteration may never converge to
the specified accuracy. Before invoking this function, it must be ensured that
the function is continuous in the interval. In particular, the sign change should
not be due to a singularity within the interval.
78. SEARCH Function to search for complex zeros by looking for sign changes
in the real and imaginary parts of the function. It will output an array of values
in the two-dimensional plane giving the quadrant value of the function at that
point. Using this array the zero may be located as explained in Section 7.7. The
zeros will be searched in a rectangular region bounded by RX1 and RX2 along
the real axis and by RY1 and RY2 along the imaginary axis. It is not really
C.7. Nonlinear Algebraic Equations 1057

essential to have RX1 < RX2 or RY1 < RY2. NX and NY are the number
of points along the real and imaginary axes respectively, where the function
value is calculated. NX and NY should be greater than 1, if not, a default
value of 21 will be used. Further, NX ≤ IMAX (= 41), otherwise it will be
reduced to the maximum permissible value of IMAX. CFUN is the name of
the function routine to calculate the function of complex variable. Function
CFUN(Z,F) must be supplied by the user. Here Z and F are both arrays of
length 2 containing the real and imaginary parts of the complex values. F is
the function value at Z. The output is displayed on the screen. The returned
value of the function SEARCH is always zero.
79. ZROOT Function to calculate complex zeros of a given function using
Muller’s method. This function uses deflation to remove the known zeros. De-
flation is carried out by explicitly dividing the function by factors of the form
z−zi , where zi are the known zeros. This function calls MULLER (or MULER2)
to find the zeros. N specifies the number of zeros to be determined. CX is an
array of length 2N, containing the real and imaginary parts of the N complex
starting values for the iteration. NZ is the number of known zeros of the given
function. CZERO is a array of length 2(NZ + N) containing the real and imagi-
nary parts of zeros. At the time of calling, the first 2NZ elements should contain
the known zeros of the function. All zeros found by the function will be added
to this list and the value of NZ will be increased accordingly. REPS and AEPS
specify the relative and absolute convergence criterion. The zeros should be
calculated with an accuracy of max(AEPS, REPS|X|). RMAX is a parameter
to specify the approximate range of z values, where zeros are required. If the
iteration goes outside the region |Z| ≤ RMAX, then it will be terminated. This
parameter may be used to ensure that iteration does not stray into a region
which will yield overflow or some other problem. CF is the name of the exter-
nal routine to calculate the required function. Function CF(Z, F) (or Function
CF(Z, F, JF) for MULER2) must be supplied by the user. Here both F and Z
are assumed to be arrays of length 2 containing the real and imaginary parts of
the complex variables. F (or F × 2JF ) is the value of function at Z. Error status
can be obtained from the returned value of ZROOT. A value of 41 implies that
the Muller’s iteration did not converge to the required accuracy for at least one
of the zeros, but the computed root may still be acceptable at lower accuracy.
This means that the function MULLER has returned a nonzero value, which
is less than 100. ZROOT = 429 implies that the iteration failed to converge
for at least one of the zeros. The number of zeros successfully found can be
obtained from NZ, which will give the number of known zeros including those
which were known before the function was called. Apart from CF, the function
also requires functions MULLER or MULER2, CABS, CSQRT, CDIV. To use
MULER2 instead of MULLER, change the call statement as indicated in the
comments. ZROOT2 is the version of ZROOT for use with MULER2.
80. ZROOT2 To calculate complex zeros of a given function using Muller’s
method. This function uses deflation to remove the known zeros. Deflation is
1058 Appendix C. C Programs

carried out by explicitly dividing the function by factors of the form z − zi ,

where zi are the known zeros. This function calls MULER2 (or MULLER) to
find the zeros. N specifies the number of zeros to be determined. CX is an
array of length 2N, containing the real and imaginary parts of the N starting
values for the iteration. NZ is the number of known zeros of the given function.
CZERO is an array of length 2(NZ+N) containing the real and imaginary parts
of zeros. At the time of calling, the first 2NZ elements should contain the known
zeros of the function. All zeros found by the function will be added to this list
and the value of NZ will be increased accordingly. REPS and AEPS specify
the relative and absolute convergence criterion. The zeros should be calculated
with an accuracy of max(AEPS, REPS|X|). RMAX is a parameter to specify
the approximate range of z values, where zeros are required. If the iteration goes
outside the region |Z| ≤ RMAX, then it will be terminated. This parameter
may be used to ensure that iteration does not stray into a region which will
yield overflow or some other problem. CF is the name of the external routine to
calculate the required function. Function CF(Z, F, IX) must be supplied by the
user. Here both F and Z are assumed to be arrays of length 2 containing the
real and imaginary parts of the complex argument, and IX is an integer variable
such that the function value is given by f (Z) = F × 2IX . This form is useful for
functions which will otherwise result in overflow or underflow. Error status can
be obtained from the returned value of ZROOT2. A value of 41 implies that the
Muller’s iteration did not converge to the required accuracy for at least one of
the zeros, but the computed root may still be acceptable at lower accuracy. This
means that the function MULER2 has returned with a nonzero value, which is
less than 100. ZROOT2 = 429 implies that the iteration failed to converge for at
least one of the zeros. The number of zeros successfully found can be obtained
from NZ, which will give the number of known zeros including those which were
known before the function was called. Apart from CF, the function also requires
functions MULER2 (or MULLER), CABS, CSQRT, CDIV. Function MULER2
uses reverse communication technique to pass the function value. This function
returns the control to the calling function whenever a function evaluation is
required. This technique is useful when a large number of parameters, including
some external function names, are required for the function evaluation. At the
first call to MULER2 the error parameter IER should be set to zero. When
the control is returned to the calling function, it should check the value of
IER to decide the action. If IER < 0, then a function evaluation is required
and MULER2 should be called once again after calculating the function. It is
important to ensure that none of the other parameters in the call statements,
e.g., CX1, CX2, CX3 and IER are changed in between. Nonnegative values of
IER signify that the execution of MULER2 is complete and it should not be
called again. ZROOT is the version of ZROOT2 for use with MULLER.
81. MULLER To calculate a complex zero of a given function using Muller’s
method. CX1, CX2 and CX3 are complex variables specifying the three start-
ing values required for Muller’s method. All these variables are treated as
arrays of length 2 containing the real and imaginary parts. After execution,
C.7. Nonlinear Algebraic Equations 1059

CX3 will contain the computed value of the zero, while CX2 and CX1 will
contain the previous iterates. REPS and AEPS specify the relative and abso-
lute convergence criterion. The root should be calculated with an accuracy of
max(AEPS, REPS|CX3|). CF is the name of the external routine to calculate
the required function. Function CF(Z,F) must be supplied by the user. Here F
and Z are both arrays of length 2 containing the real and imaginary parts of
complex arguments and F is the function value at Z. NZ is the number of known
zeros of the function. CZERO is an array of length 2NZ containing the real and
imaginary parts of the known complex zeros. RMAX is a parameter to specify
the approximate range of z values, where zeros are required. If the iteration
gives |Z| > RMAX, then it will be terminated. This parameter may be used to
ensure that the iteration does not stray into a region which will yield overflows
or other problems. Error status can be obtained from the returned value of
MULLER. A value of 42 implies that the iteration converged to a moderate
accuracy specified by the parameter REPS0 (= 10−4 ), but after some stage
the difference between successive iterates started increasing and the iteration
was terminated. This will usually imply that specified accuracy is too stringent,
or that there is a significant roundoff error in evaluating the function. In such
cases, the computed value of the zero should be within the domain of indeter-
minacy for the function and may be acceptable. The difference |CX2 − CX3|
should give an estimate of accuracy achieved. MULLER = 43 implies that the
moderate convergence criterion was satisfied, but the iteration failed to con-
verge to the required accuracy in specified maximum number of iterations NIT
(= 50). This failure could be due to slow convergence and increasing the value
of NIT, or a second attempt with the new estimate as the starting value may
yield a better result. Further, the difference between CX1, CX2 and CX3 will
give an estimate of the expected error in the computed zero. MULLER = 404
implies that the three starting values are not distinct, in which case no calcu-
lations are performed. MULLER = 431 implies that iteration has gone outside
the specified limits i.e., |CX3| > RMAX. MULLER = 432 implies that the
iteration has failed to converge to any reasonable accuracy. MULLER = 433
implies that the iteration cannot be continued further, because the denomina-
tor in the iteration function vanishes. In this case, CX1, CX2 and CX3 will
contain the last three iterates. This failure can occur at multiple zeros after the
iteration has entered into the domain of indeterminacy, in which case, the root
may be acceptable. Apart from CF this function also requires CABS, CDIV
and CSQRT to handle complex arithmetic.
82. MULER2 Function to calculate a complex zero of a given function using
Muller’s method. This function is essentially identical to the function MULLER
described earlier, except for the fact that here the given function is calculated
in the form f (CX) = CF × 2IX , which is suitable for those functions which will
normally overflow or underflow on the computer. For example, this function is
very useful to calculate the zeros of determinants arising out of finite difference
approximation to differential equations. Apart from this, it uses the reverse
communication technique for passing function values. This function will return
1060 Appendix C. C Programs

control to the calling function when it needs a function evaluation. In that

case, the error parameter IER will be set to a negative value and the function
should be evaluated at z = CX. The value of the function should be returned
in the variables CF and IX. Here CX and CF are arrays of length 2 containing
the real and imaginary parts of complex variables while IX is an integer. The
function value is expressed in the form explained above. During the function
evaluation variables other than CF and IX in the call statement should not
be disturbed. Before the first call IER should be set to zero. IER ≥ 0 imply
that the execution is complete and no more function evaluations are required.
Apart from CF, CX and IX other arguments as well as error exits are identical
to those for MULLER, except that in this case the parameter IER gives the
error parameter instead of the returned value of function. IER = 42 implies
that the iteration converged to a moderate accuracy specified by the param-
eter REPS0 (= 10−4 ), but after some stage the difference between successive
iterates started increasing and the iteration was terminated. This will usually
imply that specified accuracy is too stringent, or that there is a significant
roundoff error in evaluating the function. In such cases, the computed value of
the zero should be within the domain of indeterminacy for the function and
may be acceptable. The difference |CX2 − CX3| should give an estimate of
accuracy achieved. IER = 43 implies that the moderate convergence criterion
was satisfied, but the iteration failed to converge to the required accuracy in
specified maximum number of iterations NIT (= 50). This failure could be due
to slow convergence and increasing the value of NIT, or a second attempt with
the new estimate as the starting value may yield a better result. Further, the
difference between CX1, CX2 and CX3 will give an estimate of the expected
error in the computed zero. IER = 404 implies that the three starting values
are not distinct, in which case no calculations are performed. IER = 431 im-
plies that iteration has gone outside the specified limits i.e., |CX3| > RMAX.
IER = 432 implies that the iteration has failed to converge to any reasonable
accuracy. IER = 433 implies that the iteration cannot be continued further,
because the denominator in the iteration function vanishes. In this case, CX1,
CX2 and CX3 will contain the last three iterates. This failure can occur at
multiple zeros after the iteration has entered into the domain of indeterminacy,
in which case, the root may be acceptable. Negative values of IER imply that
the execution is not over and MULER2 should be called back after calculating
the function value at the required point. Variables other than CF and IX should
not be modified during this function calculations. For an example of the usage
of this function see function ZROOT2 above. Apart from CF this function also
requires CABS, CDIV and CSQRT to handle complex arithmetic.
83. POLYR Function to calculate all zeros of a polynomial of degree N, with
real coefficients, using Laguerre’s method. N is the degree of polynomial, A is
an array of length N+1 containing the coefficients of the polynomial. A[I] is the
coefficient of xI in the polynomial i.e., A[0] is the constant term while A[N] is the
coefficient of xN . A[N] should be nonzero. RR is an array of length N, which will
contain the real part of zeros of the polynomial after the execution is complete.
C.7. Nonlinear Algebraic Equations 1061

The roots are sorted in the order of increasing real part. However, if the function
is terminated abnormally, then the roots may not be sorted. RI is an array of
length N, which will contain the imaginary part of zeros of the polynomial after
the execution is complete. QREFIN is a parameter to specify if refinement of
roots is needed. QREFIN = 1 specifies that the computed zeros should be
“refined” by iterating with the original polynomial, otherwise refinement is not
required. This parameter may be useful, if there is some doubt that the iteration
during refinement has converged to some other root, thereby giving rise to an
apparent double root. Error status can be obtained from the returned value
of POLYR. A value of 406 implies that N ≤ 0, while POLYR = 408 implies
that A[N] = 0. In such cases, no calculations are performed. POLYR = 430
implies that the Laguerre’s iteration failed to converge at some stage. In this
case, the roots already located until then, will be available in arrays RR and
RI. POLYR = n × 11 implies that the iteration for refining the roots did not
converge for n of the roots. The parameter EPS inside the function should
be of the order of ~ for the arithmetic being used. This value is used only to
decide whether the root is real or complex. If z = x + iy is the computed root
and |y| ≤ 10EPS × |x|, then the root is assumed to be real. Depending on
the outcome of this test, the function performs deflation for a real root, or a
pair of complex conjugate roots. This parameter EPS should be changed to
appropriate value, but the exact value may not be crucial for working of the
function, unless there are roots with very small, but nonzero imaginary parts.
This function requires function LAGITR to perform the Laguerre’s iteration,
and CABS, CDIV and CSQRT to handle complex arithmetic.
84. LAGITR Function to find one root of a polynomial of degree N with real
coefficients, using Laguerre’s method. N is the degree of polynomial and A is
the array of length N + 1 containing the coefficients. A[I] is the coefficient of
xI . CXI is an array of length 2 containing the real and imaginary parts of the
starting value for the Laguerre’s iteration. After execution, CXI will contain
the computed root. Error status can be obtained from the returned value of
LAGITR. A value of 439 implies that the iteration has failed to converge to
any reasonable accuracy. LAGITR = 438 implies that the denominator in the
iteration function vanishes and the iteration cannot be continued. In practice,
it is rare for this iteration to fail, unless the polynomial is too ill-conditioned
for the precision of the arithmetic used. This function requires CABS, CDIV
and CSQRT to handle complex arithmetic.
85. DAVIDN Function to solve a system of nonlinear equations using Davi-
denko’s method. This method can be coupled with any function for solving a
system of nonlinear equations to improve its chances of convergence. It can be
used with function NEWTON for Newton’s method, when the Jacobian can
be easily calculated, or with function BROYDN for Broyden’s method, when
the Jacobian is too complicated to be calculated explicitly. FCN is the name
of the function to calculate the vector function, i.e., the left-hand sides of the
equations (the right-hand sides are assumed to be zero). This function should
1062 Appendix C. C Programs

also calculate the Jacobian if Newton’s method is to be used. The parame-

ters THETA, X0[NMAXD], F0[NMAXD] are global variables used to pass on
parameters to function FCN. Here, X0 and F0 are arrays of length NMAXD
containing the initial guess and corresponding function values, which may be
required to calculate the function required for the Davidenko’s method. The
function can be parametrised in any convenient way, but THETA = θ0 should
have the solution X0 and THETA = 0 should correspond to the required func-
tion. The function FCN should be of the form
double THETA, X0[NMAXD], F0[NMAXD];
void fcn(int np, double x[], double f[], double *df)
{
f[0]=...; ... ; F[NP-1]=...;
.......................
for(i=0; i<np; ++i) f[i]=f[i]-THETA*F0[i];
}
Here NP is the number of variables, X and F are the vectors x and f , DF is the
Jacobian. This form is for use with function NEWTON. For use with BROYDN,
the variable DF should be removed from the list of arguments. The last state-
ment parametrise the function for Davidenko’s method. Other parametrisation
mentioned in Section 7.16 can also be used, but if X[I] also occurs with THETA,
then the Jacobian will also need to be modified (if Newton’s method is to be
used). If the function has some natural parameter, then it can be used instead
of artificial parametrisation, provided the solution is known for some value of
the parameter and θ = 0 corresponds to the required solution. The last require-
ment can be removed by a trivial change in the function. The function FCN
with above specifications must be supplied by the user.
NP is the number of variables which must be equal to the number of
equations. X is an array of length NP containing the starting values. After
execution, the solution will be returned in the same array. It is not essential
to supply the starting values, since these values can be supplied at the time
of execution through the scanf statement. Apart from the starting values, the
function will also ask for successive values of THETA, the parameter introduced
into the equations. The value of THETA should be changed from θ0 to zero
gradually, where θ0 is the value for which X0 is the exact solution of the system
of equations. The value of θ0 is generally assumed to be 1. Steps in which the
value of THETA should change will depend on the functions involved. If at
any stage the iteration fails to converge, a second attempt could be made with
THETA changing in smaller steps, or with a different starting value. F is an
array of length NP, which will contain the value of the vector function f at
the point specified by the array X. REPS and AEPS specify the relative and
absolute convergence criterion. All components of the root should be calculated
to an accuracy of max(AEPS, REPS|X[I]|). The function initially uses a more
modest criterion, which may need to be changed if some of the components of
the root are very small, since an absolute criterion with AEPS0 = 10−4 may
be too large in that case. For such cases, change the value of AEPS0 suitably.
C.7. Nonlinear Algebraic Equations 1063

AEPS0 may be set equal to AEPS if needed. Error status can be obtained from
the returned value of DAVIDN. A value of 407 implies that the value of NP is
greater than NMAXD (= 200), or NP < 1. In this case, the value of NMAXD
can be increased. DAVIDN = 440 implies that THETA did not reduce to zero,
even after 100 steps. In this case, a second attempt could be made with the last
value of X as the starting value. Apart from this, other values may be set by the
function NEWTON or BROYDN. This function requires function NEWTON
or BROYDN to solve the system of nonlinear equations and function GAUELM
to solve intermediate systems of linear equations. To use BROYDN instead of
NEWTON, the call statement should be changed as done in DAVIDN B, which
is the version for use with BROYDN.
86. NEWTON Function to solve a system of nonlinear equations using New-
ton’s method. This method requires the calculation of the full Jacobian at every
step. This function can be used directly, if a good approximation to the root
is known. Otherwise, it may be better to use it through function DAVIDN to
improve the chances of convergence. NP is the number of equations in the sys-
tem. X is an array of length NP containing the initial approximation to the
solution. After execution, the computed solution will be returned in the same
array. F is an array of length NP containing the value of vector function f at
the point specified by the array X. FCN is the name of the function used to
calculate the left-hand sides of the system of equations (the right-hand sides are
assumed to be zero). The function FCN(NP, X, F, DF) must be supplied by the
user. (Here NP is the number of equations in the system. X and F are arrays
of length NP containing the values of x and the vector function f , respectively.
DF is a two-dimensional array of length NP × NP containing the Jacobian with
DF[J][I] = ∂F[I]/∂X[J]. It should be noted that in function FCN, the second di-
mension of array DF must be equal to NP itself.) REPS and AEPS specify the
relative and absolute convergence criterion. All components of the root should
be calculated to an accuracy of max(AEPS, REPS|X[I]|). Error status can be
obtained from the returned value of NEWTON. A value of 442 implies that the
iteration did not converge to the specified accuracy. NEWTON = 441 implies
that the function GAUELM used to solve the intermediate systems of linear
equation has failed. This failure can occur if the value of NP is unacceptable,
or if the Jacobian matrix is nearly singular.
87. BROYDN Function to solve a system of nonlinear equations using Broy-
den’s method. This method does not require the calculation of derivatives. This
function can be used directly if a good approximation to the root is known. Oth-
erwise, it may be better to use it through function DAVIDN B to improve the
chances of convergence. NP is the number of equations in the system. X is an
array of length NP containing the initial approximation to the solution. After
execution, the computed solution will be returned in the same array. F is an
array of length NP containing the value of vector function f at the point spec-
ified by the array X. FCN is the name of the function used to calculate the
left-hand sides of the system of equations (the right-hand sides are assumed to
1064 Appendix C. C Programs

be zero). The function FCN(NP, X, F) must be supplied by the user. (Here NP

is the number of equations in the system. X and F are arrays of length NP con-
taining the values of x and the vector function f (x), respectively.) REPS and
AEPS specify the relative and absolute convergence criterion. All components
of the root should be calculated to an accuracy of max(AEPS, REPS|X[I]|).
Error status can be obtained from the returned value of BROYDN. A value
of 442 implies that the iteration did not converge to the specified accuracy.
BROYDN = 441 implies that the function GAUELM used to solve the inter-
mediate systems of linear equation has failed. This failure can occur if the value
of NP is unacceptable, or if the matrix is nearly singular.

C.8 Optimisation

88. BRACKM Function to bracket a minimum in one dimension. At the

time of calling A and B should contain the two starting values for search.
After execution, the triplet (A, X, B) should bracket the minimum with
F(X) < min(F(A), F(B)) and X in between A and B. F is the name of the
function routine used to calculate the function to be minimised. Function F(X)
must be supplied by the user. Error status can be obtained from the returned
value of BRACKM. A value of 501 implies that A = B and no calculations
are performed. BRACKM = 521 implies that the function failed to locate a
minimum. This failure can occur, if there is no minimum, or if the initial step
|B − A| is too large, which causes the function to jump over the minimum, or
if the minimum is too far off as compared to the initial step size.
89. GOLDEN To minimise a function in one dimension using the golden sec-
tion search. Before calling the function, the minimum should be bracketed by
the triplet (A, X, B) with X in between A and B and F(X) < min(F(A), F(B)).
It is not necessary to have B > A, but X must be between A and B. After
execution, the final bracket will be overwritten on (A, X, B) with the cen-
tral value X giving the best approximation to the minimiser. FX will give the
value of the function at X. The parameters REPS and AEPS specify the rel-
ative and absolute convergence criterion. The interval will be subdivided until
|B − A| < max(REPS|X|, AEPS). F is the name of the function routine used
to calculate the function to be minimised. Function F(X) must be supplied by
the user. Error status can be obtained from the returned value of GOLDEN. A
value of 50 implies that the process of subdivision was terminated, because the
function value is equal at all the three points. This is usually due to roundoff
error as explained in Section 8.1. GOLDEN = 51 implies that the required
accuracy was not achieved, even after NIT (= 100) subdivisions. This problem
can arise if very high accuracy is required, or if the initial bracket is too large.
GOLDEN = 522 implies that the input values of A, B, X are not consistent,
i.e., either they do not bracket a minimum or X is not between A and B. In
this case, no calculations will be performed.
C.8. Optimisation 1065

90. BRENTM To minimise a function in one dimension using the Brent’s

method. This function is based on the procedure given by Brent (1973). Before
calling this function, the minimum should be bracketed by the triplet (A, X,
B) with F(X) < min(F(A), F(B)) and X between A and B. After execution, the
final bracket will be overwritten on (A, X, B) with the central value X giving
the best approximation to the minimiser. FX will give the value of the function
at X. The parameters REPS and AEPS specify the relative and absolute con-
vergence criterion. The minimiser X should be calculated with an accuracy of
max(REPS|X|, AEPS). F is the name of the function routine used to calculate
the function to be minimised. Function F(X) must be supplied by the user.
Error status can be obtained from the returned value of BRENTM. A value of
51 implies that the required accuracy was not achieved, even after NIT (= 75)
iterations. This failure can occur if very high accuracy is required or if the ini-
tial bracket is too large. BRENTM = 523 implies that the input values of A,
B, X are not consistent, i.e., either they do not bracket a minimum or X is not
between A and B. In this case, no calculations will be performed. No test for
roundoff error is performed in this function and the result may not be accurate
to the specified accuracy, even if the interval has been reduced to the required
size.

91. DAVIDM To minimise a function in one dimension using the Hermite

cubic interpolation. This method requires the calculation of the first derivative
in addition to the function value. X1 and X2 specify the two distinct start-
ing values. These values need not bracket the minimum. After execution, X1
and X2 will contain the last two iterates with X2 giving the best approxima-
tion to the minimiser. F2 will give the value of the function at X2, while D2F
gives the estimate for second derivative, which is used to distinguish between
a minimum and a maximum. If D2F > 0 then X2 should be a minimiser, oth-
erwise X2 should be a maximiser. Since D2F is merely an estimate of second
derivative, if it is close to zero the nature of extremum will be difficult to de-
termine. The parameters REPS and AEPS specify the relative and absolute
convergence criterion. The minimiser X2 should be calculated with an accu-
racy of max(REPS|X2|, AEPS). F is the name of the function routine used to
calculate the function to be minimised and its derivative. Error status can be
obtained from the returned value of DAVIDM. A value of 52 implies that the
iteration has converged to a maximiser, rather than a minimiser. This distinc-
tion is made on the basis of the sign of the estimated second derivative and
hence may not necessarily be correct if roundoff error is significant. In particu-
lar, a point of inflection can be passed on as either a minimiser or a maximiser.
DAVIDM = 502 implies that X1 = X2, in which case, no calculations will be
performed. DAVIDM = 524 implies that the iteration cannot be continued fur-
ther, since the Hermite cubic does not have a minimum and the corresponding
parabolic interpolation yields a zero denominator. DAVIDM = 525 implies that
the required accuracy was not achieved, even after NIT (= 75) iterations. The
convergence failure can occur if very high accuracy is required, or if the starting
1066 Appendix C. C Programs

values are far from the minimiser. Function F(X, DF) must be supplied by the
user. Here DF is the first derivative of F at X.
92. BFGS To find minimum of a function of N variables using the quasi-
Newton method, with BFGS formula for updating the Hessian matrix. This
method requires the calculation of gradient vector in addition to the function
value. X is an array of length N, containing the starting values for the inde-
pendent variables. After execution, the minimiser will be returned in the same
array X. F gives the value of the function at X. G is an array of length N,
which will contain the value of the gradient vector at X (G[I] = ∂F ∂X[I] ). H is
a two-dimensional array of length at least N × N, which will contain an ap-
proximation to the inverse of the Hessian matrix at X. The second dimension
of H must be equal to N. F, G and H are output parameters, which need not
be initialised at the time of calling. NUM is an output parameter containing
the number of function evaluations required. The parameters REPS and AEPS
specify the relative and absolute convergence criterion. Each component of the
minimiser X should be calculated with an accuracy of max(REPS|X[I]|, AEPS).
FCN is the name of the function which calculates the function F and the gra-
dient vector G. Error status can be obtained from the returned value of BFGS.
A value of 53 implies that the Hessian matrix is probably singular at the final
point. In this case, the iteration could have converged to a saddle point and
further investigation may be required to determine the nature of the stationary
point. This is detected by considering the norm of the matrix H as explained
in Section 8.5. This test is not very reliable, particularly if the required accu-
racy is too low or too high. BFGS = 503 implies that N < 1, in which case
no calculations are performed. BFGS = 526 implies that the iteration failed to
converge to a satisfactory accuracy. Other values may be set by the function
LINMIN, which is called to perform the line searches. Function FCN(N, X, F,
G) must be supplied by the user. Here N is the number of variables, F is the
function value, while X and G are arrays of length N, containing the coordi-
nates of the point X, where the function F and the gradient vector G are to
be evaluated. Apart from FCN this function requires the function LINMIN to
perform the line searches and the function FLNM to calculate the function and
its derivative along the given line, as required by LINMIN.
93. LINMIN Function to perform line search as required by the function
BFGS. It implements a crude, but reasonably efficient algorithm to find an
acceptable minimum of a function of N variables along a given direction. Any
point which satisfies the conditions (8.37) is acceptable. This function requires
the first derivative of the function in addition to the function value. This func-
tion should only be used for line search as required by quasi-Newton methods
and not for any other purpose, since the criterion for acceptance will not suffice
for other purposes. X1 is the starting point from where the line search is to
be performed. X2 is the first guess for the minimum. After execution, X1, F1
and DF1 will respectively contain the accepted point, the value of the function
and its first derivative along the given direction at that point. The parameters
C.8. Optimisation 1067

REPS and AEPS specify the relative and absolute convergence criterion for
the function BFGS. These parameters are used here only to terminate the line
search, once the interval has been reduced sufficiently. This situation should not
arise normally, but occasionally because of roundoff error, or error in coding the
derivatives, the function may fail to find acceptable points. The parameters F,
V, XI, N and NUM are not used by this function, but are simply passed on to
the function FLNM, for calculating the function value and the first derivative.
F is the name of the function used to calculate the function of N variables, V
is an array of length 3N, first N elements of which contain the direction along
which the line search is to be carried out. After execution of FLNM the next
N elements will contain the coordinates of the last point at which the function
is evaluated, while the last N elements of this array will contain the gradient
vector at the last point. XI is an array of length N, containing the coordinates of
the starting point for the line search. NUM is an integer variable to keep count
of the number of function evaluations. Error status can be obtained from the
returned value of LINMIN. A value of 527 implies that the LINMIN has failed
to find any point, where the function value is less than or equal to that at the
starting point. This failure could be due to roundoff error, or more likely due
to some error in coding the derivatives. LINMIN = 55 or 528 implies that the
function failed to find an acceptable point, even though the interval has been
reduced to specified tolerance. In this case, the last point is accepted if the
function value is less than or equal to that at starting point (LINMIN = 55),
otherwise LINMIN = 528. LINMIN = 54 implies that the function failed to
find an acceptable point in NIT (= 15) iterations, but the function value at
the last approximation is less than or equal to the value at the starting point
and this point is accepted. The last two conditions are often encountered, when
the iteration is close to convergence and the accuracy requirement is too high.
The final point may be acceptable in such circumstances. This can be checked
by verifying the value of the gradient vector in function BFGS. This function
requires the function FLNM and the function F to calculate the function. Func-
tion F(N, X, FX, G) must be supplied by the user. Here N is the number of
variables, FX is the function value, while X and G are arrays of length N, con-
taining the coordinates of the point X, where the function FX and the gradient
vector G are to be evaluated.
94. FLNM To calculate the function as required for the line search routine
LINMIN. FCN is the name of the function used to calculate the function of N
variables. X is a parameter along the line which specifies the point at which the
function is to be evaluated. DF is the first derivative along the line. V is an array
of length 3N, the first N elements of which specify the direction along which
the search is being carried out. Next N elements will contain the coordinates
of the point at which function is evaluated, while the last N elements contain
the gradient vector at this point. X0 is an array of length N, containing the
coordinates of the starting point. NUM is an integer variable which keeps a
count of function evaluations. The function will be evaluated at the point with
coordinates XI[I] = X0[I] + X × V[I]. These coordinates are returned in N to
1068 Appendix C. C Programs

2N − 1 element of array V. The Function FCN(N, XI, F, G) must be supplied

by the user. Here XI is an array of length N, containing the coordinates of the
required point. F is the function value at this point. G is an array of length N
containing the gradient vector at XI.
95. NMINF To find minimum of a function of N variables, using a direction
set method. This method does not require calculation of the derivatives. X is an
array of length N, containing the starting values for the independent variables.
After execution, the minimiser will be returned in the same array X. F gives the
value of the function at X. NUM is an output parameter containing the number
of function evaluations required. The parameters REPS and AEPS specify the
relative and absolute convergence criterion. The iteration is continued until
the change in function value δF during one complete cycle consisting of N
iterations, satisfies |δF | < max(REPS|F|, AEPS). It may be noted that unlike
function BFGS, here the convergence test is applied on function value, rather
than the minimiser. This must be kept in mind while specifying REPS and
AEPS. In general, relative
√ change by REPS in function value will translate to
relative variation by REPS in coordinates of minimiser. FCN is the name of
the function which calculates the function F. Error status can be obtained from
the returned value of NMINF. A value of 504 implies that N ≤ 1, in which case,
no calculations are performed. NMINF = 529 implies that the iteration failed
to converge to a satisfactory accuracy. Other values may be set by the function
LINMNF, which is called to perform the line searches. Function FCN(N, X,
F) must be supplied by the user. Here N is the number of variables, X is an
array of length N containing the parameter values at which the function value
is required. F is the calculated value of the function at X. Apart from FCN, this
function requires function LINMNF to perform the line searches and function
FLN to calculate the function as required by LINMNF, and function SVD to
perform the singular value decomposition.
96. LINMNF Function to perform line search as required by the function
NMINF. It implements a crude, but reasonably efficient algorithm to find an
acceptable minimum of a function of N variables along a given direction. This
function does not require any derivative. Any point, where the function value is
less than that at the starting point is acceptable. This function should only be
used for line search as required by direction set methods. Because of the crude
criterion for acceptance it is not suitable for a general minimisation along a
given direction. X0 is the starting point, from where the line search is to be
performed. X1 is the first guess for the minimum. After execution, X0 and
F0 will contain the accepted point and the corresponding function value. The
parameters REPS and AEPS specify the relative and absolute convergence
criterion for the function NMINF. These parameters are used here only to
terminate the line search, once the interval has been reduced sufficiently. This
situation should not arise normally, but occasionally because of roundoff error,
the function may fail to find an acceptable point. The parameters F, V, XI,
N and NUM are not used by this function, but are simply passed on to the
C.8. Optimisation 1069

function FLN, to calculate the function value. F is the name of the function used
to calculate the function of N variables, V is an array of length N, containing
the direction along which the line search is to be carried out. XI is an array
of length 2N, the first N elements contain the coordinates of the starting point
for the line search. After execution of FLN, the next N elements will contain
the coordinates of the last point at which the function is evaluated. NUM is
an integer variable to keep count of the number of function evaluations. Error
status can be obtained from the returned value of LINMNF. A value of 56
implies that the function failed to find an acceptable point and in this case, the
starting point itself is accepted. The function LINMNF requires the functions
FLN and F to calculate the function. Function F(N, X, FX) must be supplied
by the user. Here N is the number of variables, X is an array of length N
containing the parameter values at which the function value is required. FX is
the calculated value of the function at X.
97. FLN To calculate the function as required for the line search routine
LINMNF. FCN is the name of the function used to calculate the function of N
variables. X is a parameter along the line which specifies the point at which the
function is to be evaluated. V is an array of length N, containing the direction
along which the search is being carried out. X0 is an array of length 2N, the
first N elements of which specify the coordinates of the starting point. NUM is
an integer variable which keeps a count of function evaluations. The function
will be evaluated at a point with coordinates XI(I) = X0(I) + X × V(I). The last
N elements of X0 are used to store XI. The Function FCN(N, XI, F) must be
supplied by the user. Here XI is an array of length N, containing the coordinates
of the required point and F is the function value at this point.
98. SIMPLX Function to solve a linear programming problem using the sim-
plex method. N is the number of variables, each of which is constrained to be
nonnegative. M1 is the number of constraints of the form aT x ≤ bi ≥ 0, M2 is
the number of constraints of the form aT x ≥ bi ≥ 0, M3 is the number of con-
straints of the form aT x = bi ≥ 0. A is an array of length IA × (N + M2 + 1).
IA specifies the second dimension of A, exactly as in the calling function
(IA ≥ M1 + M2 + M3 + 2). The array A contains the tableau as explained
in Section 8.7. It may be noted that in the C function the dimensions of ar-
ray A have been interchanges so effectively, the rows are treated as columns
and vice versa. At input the first column of A contains the cost coefficients
A[i + 1][0] = ci , where c0 x0 + c1 x1 + . . . + cN −1 xN −1 is the function to be min-
imised. The next M = M1 + M2 + M3 columns contain the coefficients for the
constraints, with the columns 1 to M1 containing the constraints of the first
type, columns M1 + 1 to M1 + M2 containing the constraints of the second
type and the remaining columns containing the constraints of the third type.
A[0][i] = bi and A[j + 1][i] = aj for the ith constraint. If M2 + M3 > 0, then
the column M + 1 will be used by the function to specify the cost coefficients
for the auxiliary objective function to be minimised for finding a basic feasible
vector. If only constraints of the first type are present, then this column will not
1070 Appendix C. C Programs

be required. After the execution is complete, the optimal feasible vector will be
returned in the array X of length N. F is an output parameter containing the
optimum value of the objective function. Error status can be obtained from the
returned value of SIMPLX. A value of 57 implies that the objective function
is unbounded from below and no optimal feasible vector exists. SIMPLX = 58
implies that, there is no basic feasible vector. Hence, the constraints are incon-
sistent or the problem is not properly formulated. SIMPLX = 505 implies that
at least one of the variables N, M1, M2 and M3 is negative or IA < M + 2.
SIMPLX = 506 implies that at least one of the right-hand side coefficients in
the constraints is negative. In both these cases, no calculations are performed.
SIMPLX = 531 implies that the simplex algorithm failed to find the opti-
mal vector in a reasonable number of iterations. This failure can occur either
because the number of iterations required are very large, or because the sim-
plex iteration has gone into a nonterminating loop because of degeneracy. In
principle, this failure can happen for the auxiliary problem for finding a basic
feasible vector also. SIMPLX = 532 implies that a basic feasible vector could
not be found to start the simplex iteration. This situation can occur if the
solution to the auxiliary problem is degenerate and the artificial variable on
the left-hand side cannot be exchanged. This situation should not arise, unless
the constraints are not linearly independent, or roundoff error is significant.
AEPS is a parameter to control roundoff error, any quantity less than AEPS
in magnitude is assumed to be zero. This function requires SIMPX to actually
perform the simplex minimisation.

99. SIMPX Function to solve linear programming problems using the simplex
method. The problem is assumed to be in the standard form and the initial
tableau is supplied in the array A of length IA × (N − M + 1) with IA ≥ M + 2.
IA is the second dimension of A, exactly as specified in the calling function. N
is the total number of variables in the given problem, including any slack or
artificial variables that have been introduced. M is the number of constraints
in the problem. NV is the number of variables excluding the artificial variables
(if any). This parameter is used only for the auxiliary function, to check if any
artificial variable is remaining on the left-hand side. QF is a parameter which
determines the column from which cost coefficients are used. If QF = 1, then
the main objective function specified by the first column of the tableau is to
be minimised. Otherwise, the auxiliary objective function specified by the last
column of A is to be minimised for finding the initial basic feasible vector. In
that case, attempt is made to remove the artificial variables from the set of
left-hand side variables, before accepting the solution. ID and IV are integer
arrays of length M + 1 and N − M + 1 respectively, used to store permutations
of the original variables. AEPS is a parameter used to control roundoff error.
Any quantity less than AEPS in magnitude is assumed to be zero. Error status
can be obtained from the returned value of SIMPX. A value of 57 implies
that the objective function is unbounded from below and the optimal feasible
vector does not exist. SIMPX = 531 implies that the simplex iteration has not
C.9. Statistical Inferences 1071

converged in a reasonable number of iterations. This failure may be due to

degeneracy, since that is not accounted for in this function.

C.9 Statistical Inferences

100. SHSORT Function to sort an array of length N in ascending order, using

shell sort algorithm. A is a real array of length N, which is to be sorted. After
execution the sorted elements will be returned in the same array. N is the
number of elements of array A to be sorted.
101. GAMMAP Function to calculate the incomplete Gamma function
Z x
γ(a, x) 1
P (a, x) = = e−t ta−1 dt . (C.27)
Γ(a) Γ(a) 0

It may be noted that sometimes the incomplete Gamma function is defined

by γ(a, x). This should be accounted for while using the function routine. It
is difficult to approximate this function over entire range of a, x values and
it is possible that for some combination the value may not be reliable. A is
the argument for the complete Gamma function and X is the upper limit of
integration in the above equation. The function is not defined for A ≤ 0 or
X < 0 and in these cases a value of −1 will be returned without any warning.
For x < 3 it uses the power series (DiDonato & Morris 1986) to approximate
∞
!
xa X (−x)n
P (a, x) = 1+a . (C.28)
Γ(a + 1) n=1
(a + n)n!

For x < 1.2a a continued fraction approximation for complementary function

e−x xa
Γ(a)
Q(a, x) = 1 − P (a, x) = (C.29)
1−a
x+
1
1+
2−a
x+ 2
1 + x+···

is used. For other values of parameter the power series

∞
!
e−x xa X xn
P (a, x) = 1+ (C.30)
Γ(a + 1) n=1
(a + 1)(a + 2) · · · (a + n)

is used. This function requires Functions GAMMA and GAMMAL to calculate

the (complete) Gamma function or its logarithm.
1072 Appendix C. C Programs

102. BETAP Function to calculate the incomplete Beta function

Z x
Bx (a, b) 1
Ix (a, b) = = ta−1 (1 − t)b−1 dt . (C.31)
B(a, b) B(a, b) 0
It may be noted that sometimes the incomplete Beta function is defined by
Bx (a, b). This should be accounted for while using the function routine. It is
difficult to approximate this function over entire range of a, b, x values and it is
possible that for some combination the value may not be reliable. A and B are
the arguments for the complete Beta function and X is the upper limit of inte-
gration in the above equation. Here A − 1 is the exponent of t and B − 1 is that
of 1 − t in the integral. The function is not define for A ≤ 0 or B ≤ 0 or X < 0
or X > 1 and in these cases a value of −1 will be returned without any warn-
ing. These function requires Functions BETSER, BETCON, BETCON1 and
BETAI for calculating the function using different approximations depending
on the arguments. Apart from these Function GAMMAL is required to calcu-
late logarithm of Gamma function and subroutines ADPINT, KRONRD and
FBETA to calculate the integral for some argument range.
103. BETSER Function routine to calculate the incomplete Beta function
(C.31) using the infinite series (DiDonato & Morris 1992)
 
a ∞
x X (1 − b)(2 − b) · · · (j − b) j 
Ix (a, b) = 1 + a x (C.32)
aB(a, b) j=1
j!(a + j)

This series is useful for small values of x. This function is called by BETAP
for some range of arguments. This routine should not be used to calculate the
function for arbitrary values of its arguments as the approximation may not
be valid. A and B are the arguments for the complete Beta function and X
is the upper limit of integration in Eq. (C.31). This function needs Function
GAMMAL to calculate the logarithm of Gamma function.
104. BETCON1 Function routine to calculate the incomplete Beta function
(C.31) using a continued fraction (DiDonato & Morris 1992) approximation

xa (1 − x)b
aB(a, b)
Ix (a, b) = (C.33)
d1
1+
d2
1+
1 + ···
with
n(b − n)x (a + n)(a + b + n)x
d2n = , d2n+1 = − . (C.34)
(a + 2n − 1)(a + 2n) (a + 2n)(a + 2n + 1)
This function is called by BETAP for some range of arguments. This routine
should not be used to calculate the function for arbitrary values of its arguments
C.9. Statistical Inferences 1073

as the approximation may not be valid. A and B are the arguments for the
complete Beta function and X is the upper limit of integration in Eq. (C.31).
This function needs Function GAMMAL to calculate the logarithm of Gamma
function.
105. BETCON Function routine to calculate the incomplete Beta function
(C.31) using a continued fraction (DiDonato & Morris 1992) approximation

xa (1 − x)b
a1
aB(a, b)
Ix (a, b) = a2 (C.35)
b1 + a3
b2 +
b3 + · · ·
where, coefficients ai , bi are defined using Eq. (C.34)

a1 = 1, b1 = 1+d1 , an+1 = −d2n−1 d2n , bn+1 = 1+d2n +d2n+1 , (n > 1).

(C.36)
This function is called by BETAP for some range of arguments. This routine
should not be used to calculate the function for arbitrary values of its arguments
as the approximation may not be valid. A and B are the arguments for the
complete Beta function and X is the upper limit of integration in Eq. (C.31).
This function needs Function GAMMAL to calculate the logarithm of Gamma
function.
106. BETAI Function routine to calculate the incomplete Beta function
(C.31) by directly evaluating the integral. This function is called by BETAP
for some range of arguments. This routine should not be used to calculate the
function for arbitrary values of its arguments as it may not be very efficient for
all values of the arguments. A and B are the arguments for the complete Beta
function and X is the upper limit of integration in Eq. (C.31). This function
needs Function GAMMAL to calculate the logarithm of Gamma function, sub-
routines ADPINT and KRONRD to evaluate the integral. Further, the function
FBETA is needed to define the integrand for the integral.
107. FBETA Function routine to calculate the integrand for calculating the
incomplete beta function. This is used by Function BETAI.
108. RANGAU Function to generate a sequence of random numbers with
Normal distribution of probability, with zero mean and unit variance. This
function uses algorithm given by Knuth. SEED could be set to any positive
value less than AN before first call to the function. After the first call, this
variable should not be changed in any other function, unless an independent
sequence of random numbers is required.
109. IRANBIN Function routine to generate a sequence of random numbers
with binomial distribution. SEED is the seed for generating random numbers.
It should be negative for the first call to function and should not be changed
in any other program, unless an independent sequence of random number with
1074 Appendix C. C Programs

different n or p is required. It should be noted that although the random number

is an integer in this case, the SEED is of type Real. N is the number of trials
in the binomial distribution and P is the probability of the event in one trial.
C is a real array of length N, which is used to store the cumulative probability
table for use in calculations. This array should not be modified in any other
program. This is used only if the mean (np) is less than RMAX, in which case
C[I] will be the probability of having I or less events.

110. IRANPOI Function routine to generate a sequence of random numbers

with Poisson distribution. SEED is the seed for generating random numbers.
It should be negative for the first call to function and should not be changed
in any other program, unless an independent sequence of random number with
different mean is required. It should be noted that although the random number
is an integer in this case, the SEED is of type Real. RMU is the mean of Poisson
distribution. P is a real array of length NMAX (=200), which is used to store
the cumulative probability table for use in calculations. This array should not
be modified in any other program. This is used only if the mean (RMU) is less
than some critical value, in which case P[I] will be the probability of having I
or less counts.

111. PCOR Function routine to calculate the probability that two uncorre-
lated sequences of length (n + 2) will give a correlation coefficient exceeding
|x|. For large even values of n the series can give large roundoff errors in which
case the distribution is approximate by normal distribution. N is the number
of degrees of freedom, i.e., N + 2 is the length of the sequences and XX is the
value of correlation coefficient. Since the probability distribution of correlation
coefficient for uncorrelated data is symmetric about x = 0, the probability
is calculated for |x|. This function requires Function GAMMAL to calculate
the logarithm of Gamma function and Function ERF to calculate the Error
function.

C.10 Functional Approximations

112. POLFIT Function to perform least squares polynomial fit using orthog-
onal polynomials. N is the number of data points, M is the degree of polynomial
to be fitted. X, F and SIG are arrays of length N containing the data points.
F[I] is the value of function at X[I] and SIG[I] is the corresponding error. If error
estimates are not available then all SIG[I] can be set to one. A is an array of
length M + 1, which will contain the coefficients of the orthogonal polynomials
in the calculated fit. ALP and BETA are arrays of length M + 1, which will
contain the coefficients αi and βi as defined in recurrence relation for orthogo-
nal polynomials. ALP[I] = αI+1 and BETA[I] = βI . Y is an array of length N,
Y[I] will contain the calculated value of function at X[I] using the least squares
fit. H is an array of length M + 1 containing the χ2 values. H[I] will contain the
C.10. Functional Approximations 1075

χ2 using polynomial of degree I,

N−1
X 2
Y[j] − F[j]
H[M] = . (C.37)
j=0
SIG[j]

As explained in Section 10.2.2, this estimate of H[I] may have a significant

roundoff error. GAM is an array of length M+1, which will contain the quantity
γi for the orthogonal polynomials as defined in Section 10.2.2. Error status can
be obtained from the returned value of POLFIT. A value of 601 implies that
M ≥ N or M < 0, in which case, no calculations are performed. POLFIT =
621 implies that one of the γi = 0, which can happen if the points X[I] are
not distinct. The fitted polynomial can be calculated at any point using the
coefficients A, ALP and BETA by function POLEVL.
113. POLEVL Function to evaluate the approximating polynomial, and its
derivatives using Clenshaw’s recurrences. This function can be used to calcu-
late the value of approximating polynomial at any point X, after the required
coefficients aj , αj and βj have been calculated by function POLFIT. M (> 0)
is the degree of polynomial. A, ALP and BETA are arrays of length M + 1, with
A[j] = aj , ALP[j] = αj+1 and BETA[j] = βj . These coefficients must be cal-
culated before calling the function POLEVL. X is the value of the independent
variable, at which the approximation has to be evaluated. F, DF and DDF are
output parameters containing the calculated values of the function and its first
and second derivatives, respectively. The returned value of POLEVL is always
zero.
114. POLFIT1 Function to perform least squares polynomial fit using orthog-
onal polynomials. This is a version of POLFIT which can handle multiple fits,
i.e., multiple sets of function values over the same set of abscissas and errors.
This is useful for recursive use in multiple dimensions. N is the number of data
points, M is the degree of polynomial to be fitted. NUM is the number of dif-
ferent right hand sides (function values) to be fitted. X and SIG are arrays of
length N containing the data points and errors. F is an array of length N×NUM
containing the function values for each set of points. F[J][I] is the value of func-
tion in Jth set at X[I] and SIG[I] is the corresponding error. If error estimates
are not available then all SIG[I] can be set to one. The second dimension of F
must be N in the calling program. A is an array of length (M+1)×NUM, which
will contain the coefficients of the orthogonal polynomials in the calculated fit
for each set. The second dimension of A must be M + 1 in the calling program.
ALP and BETA are arrays of length M + 1, which will contain the coefficients
αi and βi as defined in recurrence relation for orthogonal polynomials. GAM is
an array of length M + 1, which will contain the quantity γi for the orthogonal
polynomials as defined in Section 10.2.2. Error status can be obtained from the
returned value of POLFIT1. A value of 601 implies that M ≥ N or M < 0, in
which case, no calculations are performed. POLFIT1 = 621 implies that one
of the γi = 0, which can happen if the points X[I] are not distinct. The fitted
1076 Appendix C. C Programs

polynomial can be calculated at any point using the coefficients A, ALP and
BETA by function POLEVL.
115. POLORT Function to evaluate the orthogonal polynomials, and its
derivatives. It may be noted that this function calculates the value of orthogo-
nal polynomials at the given point as opposed to POLEVL which calculates the
value of fitted function using the coefficients of these polynomials. This function
can be used to calculate the value of orthogonal polynomial basis functions at
any point X, after the required coefficients αj and βj have been calculated by
function POLFIT or POLFIT1. M (> 0) is the degree of polynomial. ALP and
BETA are arrays of length M + 1, with ALP[j] = αj+1 and BETA[j] = βj .
These coefficients must be calculated before calling the function POLORT. X
is the value of the independent variable, at which the polynomials have to be
evaluated. F, DF and DDF are arrays of length M + 1, which will contain the
calculated values of the M + 1 orthogonal polynomials and its first and second
derivatives, respectively at X.
116. POLFIT2 Function to perform least squares polynomial fit using or-
thogonal polynomials in two dimensions. The data values must be available at
a rectangular grid of points and weights are assumed to be equal for all points.
NX is the number of data points along X axis, NY is the number of data points
along Y axis. X is an array of length NX containing the points along X axis,
while Y is an array of length NY containing the points along the Y axis. F is
an array of length LA × NY containing the function values. F[J][I] is the value
at X[I], Y[J]. AX is an array of length IC × 3 containing information about the
fit along X direction. AX[0][I], AX[1][I], AX[2][I] will respectively contain the
coefficients αi+1 , βi , γi for the orthogonal polynomials. AY is an array of length
IC×3 containing information about the fit along Y direction. AY[0][I], AY[1][I],
AY[2][I] will respectively contain the coefficients αi+1 , βi , γi for the orthogonal
polynomials. The arrays AX and AY will be calculated by the function. LA is
the second dimension of arrays F and FY as declared in the calling function,
LA ≥ max(NX, NY). C is an array of length IC × (MY + 1) containing the fit-
ted coefficients for the polynomial fit in two dimensions. The fitted polynomial
would be
MX X
X MY
C[j][i]φi (x)ψj (y), (C.38)
i=0 j=0

where φi (x) and ψj (y) are the orthogonal polynomials in x and y respectively.
IC is the second dimension of arrays AX, AY and C as declared in the calling
function, IC > max(MX, MY). MX is the required degree of polynomial in
X. MY is the required degree of polynomial in Y. FY is an array of length
LA × NY containing the fitted values of the function at each of the tabular
points. CHISQ is calculated value of χ2 for the fit. Error status can be obtained
from the returned value of POLFIT2. A value of 602 implies that IC < MX + 1
or LA < max(NX, NY). POLFIT2 = 603 implies that NX ≤ MX, NY ≤ MY,
MX < 0 or MY < 0. In all these cases, no calculations are performed. The
C.10. Functional Approximations 1077

fitted polynomial can be calculated at any point using the coefficients AX, AY,
C by function POLEV2. This function requires functions POLFIT1, POLEV2
and POLORT.

117. POLEV2 Function to evaluate the approximating polynomial, and its

derivatives using expansion in orthogonal polynomials in 2 dimensions. This
function can be used to calculate the value of approximating polynomial at any
point, after the required coefficients cij as well as other auxiliary coefficients
needed to define the orthogonal polynomials have been calculated by function
POLFIT2. NX is the degree of polynomial in X, NY is the degree of poly-
nomial in Y. AX is an array of length LA × 2 containing the coefficients αi
and βi for orthogonal polynomials in X. AX[0][I] contains αi+1 and AX[1][I]
contains βi . AY is an array of length LA × 2 containing the coefficients αi
and βi for orthogonal polynomials in Y. AY[0][I] contains αi+1 and AY[1][I]
contains βi . These coefficients must be calculated before calling the function
POLEV2 using POLFIT2. LA is the second dimension of arrays AX, AY and
WT in the calling function, LA > max(NX, NY). WT is an array of length
LA × (NY + 1), containing the coefficients of the fit. X0, Y0 are the coordinates
of the point at which function value needs to be calculated. F is the output
parameter containing the calculated values of the function at (X0, Y0). Output
parameters, DFX and DFY are respectively, ∂F/∂x and ∂F/∂y, while DFXX,
DFXY, DFYY are the second derivative ∂ 2 F/∂x2 , ∂ 2 F/∂x∂y, ∂ 2 F/∂y 2 . Error
status can be obtained from the returned value of POLEV2. A value of 604
implies that max(NX, NY) ≥ LA, in which case no calculations are done. The
function requires POLORT to calculate the orthogonal basis functions in one
dimensions.

118. POLFITN Function to perform least squares polynomial fit using or-
thogonal polynomials in N dimensions. The data values must be available at
a hyper-rectangular grid of points and weights are assumed to be equal for all
points. N is the number of dimensions. NK is an integer array of length N con-
taining the number of data points along each axis. NK[I] is the number of points
along Ith axis. X is an array of length LA × N containing the points along each
axis, X[J][I] is the Ith point along the Jth dimension. F is an array of length
NK[0]×NK[1]×· · ·×NK[N − 1] containing the function values. F is treated as a
one dimensional array in this function and hence the dimensions of array in the
calling function must exactly match the number of points in each dimension,
e.g., F[NK[N−1]]. . .[NK[1]][NK[0]]. AX is an array of length LA × (3N + 3) con-
taining information about the fit along each direction. AX[3J][I], AX[3J+1][I],
AX[3J+2][I] will respectively contain the coefficients αi+1 , βi , γi for the orthog-
onal polynomials along Jth dimension. The rest of the array is used as scratch
space while calculating fits in one dimension. LA is the second dimension of
arrays X and AX as declared in the calling function, LA ≥ max(NK[I]). C is
an array of length (MK[0] + 1)(MK[1] + 1) · · · (MK[N − 1] + 1) containing the
fitted coefficients for the polynomial fit in N dimensions. The fitted polynomial
1078 Appendix C. C Programs

would be
MK[0] MK[1] MK[N−1]
X X X
··· C[iN −1 ] . . . [i1 ][i0 ]φi0 (x0 )φi1 (x1 ) · · · φiN −1 (xN −1 ),
i0 =0 i1 =0 iN −1 =0
(C.39)
where φij (xj ) are the orthogonal polynomials along jth dimension. C is treated
as a one dimensional array in this function and hence the dimensions of array
in the calling function must exactly match the number of polynomials in each
dimension, e.g. C[MK[N−1]+1]. . .[MK[1]+1][MK[0]+1]. MK is an integer array
of length N containing the required degree of polynomial in each direction. FY
is an array of the same length and shape as F which will contain the fitted value
of the function at each of the tabular points. CHISQ is calculated value of χ2
for the fit. Error status can be obtained from the returned value of POLFITN.
A value of 605 implies that LA < max(NK[I]). In this case, no calculations
are performed. The fitted polynomial can be calculated at any point using the
coefficients C and AX by function POLEVN, POLEVN1 or POLEVN2. This
function requires functions POLFIT1, POLEVN, POLORT.
119. POLEVN Function to evaluate the approximating polynomial using ex-
pansion in orthogonal polynomials in N dimensions. This function can be used
to calculate the value of approximating polynomial at any point, after the
required coefficients as well as other auxiliary coefficients needed to define
the orthogonal polynomials have been calculated by function POLFITN. N
is the number of dimensions. NK is an integer array of length N contain-
ing the degree of polynomial in each dimension. AX is an array of length
LA × (3N + 3) containing the coefficients αi and βi for orthogonal poly-
nomials in X. AX[3J][I] contains αI+1 and AX[3J+1][I] contains βI for or-
thogonal polynomials along Jth dimension. These coefficients must be calcu-
lated before calling the function POLEVN using POLFITN. LA is the sec-
ond dimension of array AX in the calling function. This must be the same
as what was used in call to POLFITN while calculating the coefficients. WT
is an array of length (MK[0] + 1)(MK[1] + 1) · · · (MK[N − 1] + 1), containing
the coefficients of the fit. This array is treated as one-dimensional array in
the function and hence its dimensions in the calling function must exactly
match the number of orthogonal polynomials in each dimension, for example,
WT[MK[N−1]+1]. . .[MK[1]+1][MK[0]+1]. X0 is an array of length N contain-
ing the coordinates of the point at which function value needs to be calculated.
F is the output parameter containing the calculated values of the function at
X0. The returned value is always zero. This function requires POLORT to cal-
culate the orthogonal basis functions in one dimensions. This function does not
calculate the derivatives of F. If first derivatives are required then one can use
POLEVN1, while for second derivatives use POLEVN2.
120. POLEVN1 Function to evaluate the approximating polynomial and its
first derivative using expansion in orthogonal polynomials in N dimensions.
This function can be used to calculate the value of approximating polynomial
C.10. Functional Approximations 1079

at any point, after the required coefficients as well as other auxiliary coeffi-
cients needed to define the orthogonal polynomials have been calculated by
function POLFITN. This is the version of POLEVN which also calculates the
first derivatives. The arguments are the same as those for POLEVN, except for
array DF of length N, which will contain the first derivatives of F at X0. DF[i]
will contain ∂F/∂xi . This function does not calculate the second derivatives of
F. If second derivatives are required then one can use POLEVN2, while if no
derivatives are required then use POLEVN.
121. POLEVN2 Function to evaluate the approximating polynomial and its
first and second derivatives using expansion in orthogonal polynomials in N
dimensions. This function can be used to calculate the value of approximating
polynomial at any point, after the required coefficients as well as other auxiliary
coefficients needed to define the orthogonal polynomials have been calculated
by function POLFITN. This is the version of POLEVN which also calculates
the first and second derivatives. The arguments are same as those of POLEVN,
except for arrays DF and DDF. DF is an array of length N which will contain
the first derivatives of F at X0. DF[i] will contain ∂F/∂xi . DDF is an array
of length N2 which will contain the second derivatives of F at X0. DDF[j][i]
will contain ∂ 2 F/∂xi ∂xj . The second dimension of DDF in the calling function
must be equal to N. This function calculates the first and second derivatives of
F. If second derivatives are not required then one can use POLEVN1, while if
no derivatives are required then use POLEVN.
122. LLSQ Function to calculate a general linear least squares fit in K di-
mensions. It uses singular value decomposition (SVD) to solve the system of
equations. The tabular points could have arbitrary distribution in K-space and
the basis functions also are arbitrary functions to be defined by the user. N
is the number of tabular points, M is the number of basis functions, K is the
number of dimensions. X is an array of length IX×N containing the coordinates
of tabular points. X[J][I] is the Ith coordinate of Jth tabular point. IX is the
second dimension of X in the calling function (IX ≥ K). For fitting in one di-
mension, IX can be set to 1 and array X can be passed on as a one dimensional
array of length N. F is an array of length N, containing the function values.
F[I] is the function value at (X[I][0], X[I][1], . . ., X[I][K−1]). This is treated as
one dimensional array in the function and hence should not have any gaps in
storage. This allows the function to be used for a general distribution of points
not necessarily along a hyper-rectangular mesh. EF is an array of length N, con-
taining the estimated error in F. This is only used for determining the weights
associated with each points. Thus, it solves the following system of equations
M−1
X A[i] F[j]
φi (X[j][0], . . . , X[j][K − 1]) = , j = 0, 1, . . . , N − 1; (C.40)
i=0
EF[j] EF[j]

using SVD. A is an array of length N, which will contain the fitted coefficients.
Although, there are only M coefficients the array must have a length of at least
1080 Appendix C. C Programs

N, since the remaining elements are used as scratch space. U is an array of

length IU × N which will contain the matrix U from SVD of the design matrix.
V is an array of length IV × M which will contain the matrix V from SVD
of the design matrix, (G = U ΣV T ). IU is the second dimension of U in the
calling function (IU ≥ M). IV is the second dimension of V and COV in the
calling function (IV ≥ M). SIGMA is an array of length M, which will contain
the singular values of the design matrix. Y is an array of length N which will
contain the fitted values of the function at the tabular points. PHI is the name
of the function to calculate the basis functions at a given point. REPS is the
required accuracy for the solution of equations. All singular values less than
REPS times the largest singular value will be set to zero during solution. This
parameter can be used to eliminate the linear combinations of basis functions
that contribute little to the fit. CHISQ is the minimum value of χ2 defined by
 2
N−1 M−1
X F[i] X A[j]
χ2 =  − φj (X[i][0], . . . , X[i][K − 1]) . (C.41)
i=0
EF[i] j=0
EF[i]

COV is an array of length IV × M which will contain the covariance matrix

between fitted parameters. COV[I][J] is the covariance between A[I] and A[J].
The diagonal elements are the variance in fitted parameters. Error status can
be obtained from the returned value of LLSQ. A value of 606 implies that
M > N, M ≤ 0, N ≤ 0 or K > IX. LLSQ = 607 implies that EF[I] are not all
positive. In both these cases no calculations are done. The function PHI(M, X,
FX) must be supplied by the user to calculate the required basis functions at
a given point. Here M is the number of basis functions, X is an array of length
K containing the coordinates of the point at which the basis functions need
to be calculated. FX is an array of length M containing the calculated basis
functions at X. FX[I] should give φI (X). For polynomial fits in one dimension
φi (x) = xi , but in that case it may be better to use POLFIT to calculate the
fit. Apart from PHI it also needs functions SVD and SVDEVL.
123. BSPFIT Function to calculate linear least squares fit to B-spline basis
functions in one dimension. It uses singular value decomposition (SVD) to solve
the system of equations and also allows regularisation to be incorporated. N
is the number of tabular points, X is an array of length N containing the co-
ordinates of tabular points. F is an array of length N, containing the function
values. F[I] is the function value at X[I]. EF is an array of length N, contain-
ing the estimated error in F. This is only used for determining the weights
associated with each points. Thus, it solves the following system of equations
NO+K−3
X C[i] F[j]
φi (X[j]) = , j = 0, 1, . . . , N − 1; (C.42)
i=0
EF[j] EF[j]

using SVD. Here φi (x) are the B-spline basis functions. K is the order of B-
splines required, K = 4 for cubic B-splines and K = 2 for linear B-splines,
C.10. Functional Approximations 1081

etc. A is an array of length LA × 2N (when RLM > 0) which will contain the
matrix U from SVD of the design matrix. If RLM ≤ 0 then A can be an array
of length LA × N. V is an array of length IV × (NO + K − 2) which will contain
the matrix V from SVD of the design matrix, (G = U ΣV T ). LA is the second
dimension of A in the calling function, LA ≥ NO + K − 2. IV is the second
dimension of V and COV in the calling function (IV ≥ NO + K − 2). SIGMA
is an array of length NO + K − 2, which will contain the singular values of
the design matrix. C is an array of length 2N, which will contain the fitted
coefficients. Although, there are only NO + K − 2 coefficients the array must
have a length of at least 2N, since the remaining elements are used as scratch
space. XF is an array of length NO containing the knots required to define the
B-spline basis functions. NO is the number of knots for defining B-splines. The
knots must be distinct and in ascending order with XF[0] containing the first
knot. This will yield (NO + K − 2) B-spline basis functions for fitting. Y is an
array of length N which will contain the fitted values of the function at the
tabular points. IFLG is an integer variable that specifies what calculation is
to be done. For IFLG ≤ 1 the matrix is calculated and its SVD is computed.
If execution is successful, IFLG will be set to 2, so that next time the matrix
calculations will be skipped. If IFLG ≤ 0 the coefficients of expansion are also
calculated and the fitted values Y as well as the CHISQ and COV are computed.
If IFLG = 2, only the coefficients of expansion will be calculated using the old
SVD available in arrays A, V and SIGMA and the (hopefully new) function
values F. For IFLG = 2 the fitted values Y, CHISQ and COV are not calculated.
If IFLG = 3, the coefficients of expansion as well as the fitted values Y and
CHISQ are calculated using the old SVD. REPS is the required accuracy for
the solution of equations. All singular values less than REPS times the largest
singular value will be set to zero during solution. This parameter can be used
to eliminate the linear combinations of basis functions that contribute little
to the fit. RLM is the regularisation parameter λ for smoothing. If λ ≤ 0 no
regularisation is applied, while for λ > 0 regularisation is applied using either
first or second derivative. IDE is the integer parameter which specifies the order
of derivative to be used for regularisation. This is used only if λ > 0, in which
case it must be either 1 or 2. For IDE = 1 first derivative smoothing is used,
while for IDE = 2 second derivative smoothing is applied. The regularisation
is applied at all tabular points, making the number of equations 2N. CHISQ is
the minimum value of χ2 defined by
 2
N−1 NO+K−3
X F[i] X A[j]
χ2 =  − φj (X[i]) . (C.43)
i=0
EF[i] j=0
EF[i]

COV is an array of length IV × M which will contain the covariance matrix

between fitted parameters. COV[I][J] is the covariance between C[I] and C[J].
The diagonal elements are the variance in fitted parameters. Error status can
be obtained from the returned value of BSPFIT. A value of 608 implies that
NO + K − 2 > N, or K < 2. BSPFIT = 609 implies that RLM > 0 but IDE is
1082 Appendix C. C Programs

not 1 or 2. BSPFIT = 610 implies that EF[I] are not all positive. In all these
cases no calculations are done. Other values may be set by SVD or BSPLIN.
This function requires functions BSPLIN, BSPEVL, SVD, and SVDEVL.
124. BSPFIT2 Function to calculate linear least squares fit to B-spline basis
functions in two dimensions. It uses singular value decomposition (SVD) to
solve the system of equations and also allows regularisation to be incorporated,
but it is restricted to equal weights for all points. The function solves the
system of equations in one dimension at a time, that is why the weights have to
be equal. For general least squares solution with varying weights users can try
BSPFITW2, which solves the system of equations for 2 dimensions directly and
hence will require much more time. NX is the number of tabular points along
x-axis. NY is the number of tabular points along y-axis. X and Y are arrays of
length NX, NY containing the coordinates of tabular points. F is an array of
length LA × NY, containing the function values. F[J][I] is the function value at
(X[I], Y[J]). K is the order of B-splines required, K = 4 for cubic B-splines and
K = 2 for linear B-splines, etc. For simplicity, the order is assumed to be the
same in both directions. AX is an array of length IV × 2NX which will contain
the matrix U from SVD of the design matrix for fit along x-axis. AY is an array
of length IV × 2NY which will contain the matrix U from SVD of the design
matrix for fit along y-axis. These dimensions are when RLM > 0. For RLM ≤ 0
the required dimensions are IV × NX and IV × NY for AX and AY respectively.
LA is the second dimension of arrays F, C and FY as declared in the calling
function. All these arrays should have the same second dimension. LA must
be at least 2 × max(NX, NY) when RLM > 0, otherwise LA must be at least
max(NX, NY). VX is an array of length IV × (MX + K − 2) which will contain
the matrix V from SVD of the design matrix for fit along the x-axis. VY is
an array of length IV × (MY + K − 2) which will contain the matrix V from
SVD of the design matrix for fit along the y-axis. IV is the second dimension
of AX, AY, VX and VY in the calling function (IV ≥ max(MX, MY) + K − 2).
SIGMAX is an array of length MX + K − 2, which will contain the singular
values of the design matrix for fit along the x-axis. SIGMAY is an array of
length MY + K − 2, which will contain the singular values of the design matrix
for fit along the y-axis. C is an array of length LA × NY, which will contain
the fitted coefficients. Although, there are only (MX + K − 2)(MY + K − 2)
coefficients the array must have a larger length, since the remaining elements
are used as scratch space. XF is an array of length MX containing the knots
required to define the B-spline basis functions along x. YF is an array of length
MY containing the knots required to define the B-spline basis functions along
y. MX is the number of knots for defining B-splines along x-axis. MY is the
number of knots for defining B-splines along y-axis. The knots must be distinct
and in ascending order with XF[0], YF[0] containing the first knot. FY is an
array of length LA × NY which will contain the fitted values of the function at
the tabular points. REPS is the required accuracy for the solution of equations.
All singular values less than REPS times the largest singular value will be set
to zero during solution. This parameter can be used to eliminate the linear
C.10. Functional Approximations 1083

combinations of basis functions that contribute little to the fit. RLM is the
regularisation parameter λ for smoothing. If λ ≤ 0 no regularisation is applied,
while for λ > 0 regularisation is applied using either first or second derivative.
IDE is the integer parameter which specifies the order of derivative to be used
for regularisation. This is used only if λ > 0, in which case it must be either 1
or 2. For IDE = 1 first derivative smoothing is used, while for IDE = 2 second
derivative smoothing is applied. The regularisation is applied at all tabular
points, making the number of equations 2 times larger. CHISQ is the minimum
value of χ2 defined by
 2
NX−1
X NY−1
X MX+K−3
X MY+K−3
X
χ2 = F[j][i] − A[j2][j1]φj1 (X[i])ψj2 (Y[j]) .
i=0 j=0 j1=0 j2=0
(C.44)
Here φj (x) are the basis functions in x and ψj (y) are those in y. Error status
can be obtained from the returned value of BSPFIT2. A value of 608 implies
that MX + K − 2 > NX, MY + K − 2 > NY, or K < 2. BSPFIT2 = 609 implies
that RLM > 0 but IDE is not 1 or 2. In all these cases no calculations are done.
Other values may be set by BSPFIT, SVD or BSPLIN. This function requires
functions BSPFIT, BSPLIN, BSPEVL, BSPEV2, SVD and SVDEVL.
125. BSPFITW2 Function to calculate linear least squares fit to B-spline ba-
sis functions in two dimensions. It uses singular value decomposition (SVD) to
solve the system of equations and also allows regularisation as well as nonuni-
form weights to be incorporated. This function solves the system of equations
directly in two dimensions and hence will require much more time as compared
to BSPFIT2. NX is the number of tabular points along x-axis. NY is the number
of tabular points along y-axis. X and Y are arrays of length NX, NY containing
the coordinates of tabular points. F is an array of length IC × NY, containing
the function values. F[J][I] is the function value at (X[I], Y[J]). EF is an array
of length IC × NY, containing the estimated errors in F. These values are used
to choose the weight for each equation for least squares solution. K is the order
of B-splines required, K = 4 for cubic B-splines and K = 2 for linear B-splines,
etc. For simplicity, order of B-splines is assumed to be the same along each
axes. A is an array of length LA × 3NX × NY which will contain the matrix
U from SVD of the design matrix. If RLM ≤ 0 then the size of array A needs
to be only LA × NX × NY. LA is the second dimension of array A as declared
in the calling function. LA must be at least (MX + K − 2)(MY + K − 2). V
is an array of length IV × (MX + K − 2)(MY + K − 2) which will contain the
matrix V from SVD of the design matrix. IV is the second dimension of V in
the calling function, IV ≥ (MX + K − 2)(MY + K − 2). SIGMA is an array of
length (MX + K−2)(MY + K−2), which will contain the singular values of the
design matrix. C is an array of length IC × (MY + K − 2), which will contain
the fitted coefficients. IC is the second dimension of arrays C, F, EF, FY as
declared in the calling function, IC must be at least NX. XF is an array of
length MX containing the knots required to define the B-spline basis functions
1084 Appendix C. C Programs

along x. YF is an array of length MY containing the knots required to define

the B-spline basis functions along y. MX is the number of knots for defining
B-splines along x-axis. MY is the number of knots for defining B-splines along
y-axis. The knots must be distinct and in ascending order with XF[0], YF[0]
containing the first knot. FY is an array of length IC × NY which will contain
the fitted values of the function at the tabular points. REPS is the required
accuracy for the solution of equations. All singular values less than REPS times
the largest singular value will be set to zero during solution. This parameter can
be used to eliminate the linear combinations of basis functions that contribute
little to the fit. RLM is the regularisation parameter λ for smoothing. If λ ≤ 0
no regularisation is applied, while for λ > 0 regularisation is applied using ei-
ther first or second derivative. IDE is the integer parameter which specifies the
order of derivative to be used for regularisation. This is used only if λ > 0, in
which case it must be either 1 or 2. For IDE = 1 first derivative smoothing is
used, while for IDE = 2 second derivative smoothing is applied. The regulari-
sation is applied at all tabular points, making the number of equations 3 times
larger. CHISQ is the minimum value of χ2 defined by

NX−1 NY−1 PMX+K−3 PMY+K−3 !2

X X F[j][i] − j1=0 j2=0 A[j2][j1]φj1 (X[i])ψj2 (Y[j])
.
i=0 j=0
EF[j][i]
(C.45)
Here φj (x) are the basis functions in x and ψj (y) are those in y. This routine
does not calculate the covariance matrix, but this can be easily added following
BSPFIT. Error status can be obtained from the returned value of BSPFITW2.
A value of 608 implies that MX + K − 2 > NX, MY + K − 2 > NY, or K < 2.
BSPFITW2 = 609 implies that RLM > 0 but IDE is not 1 or 2. BSPFITW2 =
610 implies that EF[J][I] are not all positive. In all these cases no calculations
are done. Other values may be set by SVD or BSPLIN. This function requires
functions BSPLIN, BSPEV2, SVD and SVDEVL.
126. BSPFITN Function to calculate linear least squares fit to B-spline basis
functions in N dimensions. It uses singular value decomposition (SVD) to solve
the system of equations and also allows regularisation to be incorporated, but
it is restricted to equal weights for all points. The function solves the system
of equations in one dimension at a time, that is why the weights have to be
equal. For general least squares solution with varying weights users can try
BSPFITWN, which solves the system of equations for N dimensions directly
and hence will require much more time. N is the number of dimensions. NK
is an integer array of length N containing the number of tabular points along
each axis. X is an array of length LA × N containing the coordinates of tabular
points. X[J][I] is the Ith point along Jth axis. F is an array of length NK[0] ×
NK[1] × · · · × NK[N − 1], containing the function values. F[iN −1 ] . . . [i1 ][i0 ] is
the function value at (X[0][i0 ], X[1][i1 ], . . ., X[N−1][iN −1 ]). This is treated
as a one-dimensional array and hence its dimensions in the calling function
must exactly match the size in each dimension, e.g., the dimensions could be
C.10. Functional Approximations 1085

F[NK[N−1]]. . .[NK[1]][NK[0]]. K is the order of B-splines required, K = 4 for

cubic B-splines and K = 2 for linear B-splines, etc. For simplicity, the order is
assumed to be the same along each dimension. A is an array of length IV ×
LA × N which will contain the matrix U from SVD of the design matrix for
fit along each axis. LA is the second dimension of array X, as declared in the
calling function. The second dimension of A should be LA × IV. LA must
be at least 2 × max(NK[I]) when RLM > 0, otherwise LA must be at least
max(NK[I]). V is an array of length IV2 × N which will contain the matrix
V from SVD of the design matrix for fit along each axis. IV is the second
dimension of XF and SIGMA in the calling function (IV ≥ K − 2 + max MK[I]).
The second dimension of V in calling function is IV2 . SIGMA is an array of
length IV × N, which will contain the singular values of the design matrix for
fit along each axis. C is an array of length NK[0] × NK[1] × · · · × NK[N − 1],
which will contain the fitted coefficients. If RLM > 0 this length must be 2
times larger. Note that although there are only (MK[0] + K − 2)(MK[1] + K −
2) · · · (MK[N − 1] + K − 2) coefficients the array must have a larger length,
since the remaining elements are used as scratch space. This is treated as a
one-dimensional array and hence its dimensions in the calling function must
exactly match the size in each dimension, e.g., the dimensions could be

C[NX][MK[N − 2] + K − 2] . . . [MK[1] + K − 2][MK[0] + K − 2]. (C.46)

Here the first dimension has to be increased suitably to accommodate the

scratch space required. The first dimension NX should be chosen such that
the total size is greater than the required value. If this array is only passed
on to BSPEVN or equivalent functions to calculate the function value at any
required point, then the exact dimensions in calling function are immaterial as
long as the total length is larger than the required value. XF is an array of
length IV × N containing the knots required to define the B-spline basis func-
tions in each dimension. The knots must be distinct and in ascending order.
XF[J][I] is the Ith knot along Jth dimension. MK is an integer array of length
N containing the number of knots for defining B-splines along each axis. FY is
an array of same size and shape as F which will contain the fitted values of the
function at the tabular points. REPS is the required accuracy for the solution
of equations. All singular values less than REPS times the largest singular value
will be set to zero during solution. This parameter can be used to eliminate
the linear combinations of basis functions that contribute little to the fit. RLM
is the regularisation parameter λ for smoothing. If λ ≤ 0 no regularisation is
applied, while for λ > 0 regularisation is applied using either first or second
derivative. IDE is the integer parameter which specifies the order of derivative
to be used for regularisation. This is used only if λ > 0, in which case it must
be either 1 or 2. For IDE = 1 first derivative smoothing is used, while for
IDE = 2 second derivative smoothing is applied. The regularisation is applied
at all tabular points, making the number of equations 2 times larger. CHISQ
is the minimum value of χ2 obtained using the fitted coefficients. Error status
can be obtained from the returned value of BSPFITN. A value of 609 implies
1086 Appendix C. C Programs

that RLM > 0 but IDE is not 1 or 2. In this case no calculations are done.
Other values may be set by BSPFIT, SVD or BSPLIN. This function requires
functions BSPFIT, BSPLIN, BSPEVL, BSPEVN, SVD and SVDEVL.
127. BSPFITWN Function to calculate linear least squares fit to B-spline
basis functions in N dimensions. It uses singular value decomposition (SVD) of
the full set of equations to obtain the fits and also allows regularisation to be
incorporated. It incorporates unequal weights, but can require several orders
of magnitude larger time as compared to BSPFITN for the same size of table.
The memory required is also much larger. Hence, it should be used only if
BSPFITN is not suitable because of highly varying weights. N is the number of
dimensions. NK is an integer array of length N containing the number of tabular
points along each axis. X is an array of length LA×N containing the coordinates
of tabular points. X[J][I] is the Ith point along Jth axis. F is an array of length
NK[0]×NK[1]×· · ·×NK[N−1], containing the function values. F[iN −1 ] . . . [i1 ][i0 ]
is the function value at (X[0][i0 ], X[1][i1 ], . . ., X[N−1][iN −1 ]). This is treated
as an one-dimensional array and hence its dimensions in the calling function
must exactly match the size in each dimension, e.g., the dimensions could be
F[NK[N−1]]. . .[NK[1]][NK[0]]. EF is an array of same size and shape as F,
containing the estimated errors in F. K is the order of B-splines required, K = 4
for cubic B-splines and K = 2 for linear B-splines, etc. For simplicity, the order
is assumed to be the same along each dimension. A is an array of length

NK[0]NK[1] · · · NK[N − 1]
(C.47)
× (MK[0] + K − 2)(MK[1] + K − 2) · · · (MK[N − 1] + K − 2),

which will contain the matrix U from SVD of the design matrix. If RLM > 0
the required size will be N + 1 times larger. LA is the second dimension of array
X, as declared in the calling function. LA must be at least max(NK[I]). V is an
array of length

[(MK[0] + K − 2)(MK[1] + K − 2) · · · (MK[N − 1] + K − 2)]2 , (C.48)

which will contain the matrix V from SVD of the design matrix. IV is the sec-
ond dimension of XF in the calling function (IV ≥ max(MK[I])). SIGMA is an
array of length (MK[0] + K − 2)(MK[1] + K − 2) · · · (MK[N − 1] + K − 2), which
will contain the singular values of the design matrix. C is an array of length
NK[0] × NK[1] × · · · × NK[N − 1], which will contain the fitted coefficients. If
RLM > 0 this length must be N + 1 times larger. Note that although there
are only (MK[0] + K − 2)(MK[1] + K − 2) · · · (MK[N − 1] + K − 2) coefficients
the array must have a larger length, since the remaining elements are used as
scratch space. This is treated as a one-dimensional array and hence its dimen-
sions in the calling function must exactly match the size in each dimension,
e.g., the dimensions could be
C[NX][MK[N − 2] + K − 2] . . . [MK[1] + K − 2][MK[0] + K − 2].
Here the first dimension has to be increased suitably to accommodate the
scratch space required. The first dimension NX should be chosen such that
C.10. Functional Approximations 1087

the total size is greater than the required value. If this array is only passed
on to BSPEVN or equivalent functions to calculate the function value at any
required point, then the exact dimensions in calling function are immaterial as
long as the total length is larger than the required value. XF is an array of
length IV × N containing the knots required to define the B-spline basis func-
tions in each dimension. The knots must be distinct and in ascending order.
XF[J][I] is the Ith knot along Jth dimension. MK is an integer array of length
N containing the number of knots for defining B-splines along each axis. FY is
an array of same size and shape as F which will contain the fitted values of the
function at the tabular points. REPS is the required accuracy for the solution
of equations. All singular values less than REPS times the largest singular value
will be set to zero during solution. This parameter can be used to eliminate
the linear combinations of basis functions that contribute little to the fit. RLM
is the regularisation parameter λ for smoothing. If λ ≤ 0 no regularisation is
applied, while for λ > 0 regularisation is applied using either first or second
derivative. IDE is the integer parameter which specifies the order of deriva-
tive to be used for regularisation. This is used only if λ > 0, in which case it
must be either 1 or 2. For IDE = 1 first derivative smoothing is used, while
for IDE = 2 second derivative smoothing is applied. The regularisation is ap-
plied at all tabular points, making the number of equations N + 1 times larger.
CHISQ is the minimum value of χ2 obtained using the fitted coefficients. Error
status can be obtained from the returned value of BSPFITWN. A value of 608
implies that LA or IV are not large enough to store the required quantities.
BSPFITWN = 609 implies that RLM > 0 but IDE is not 1 or 2. In these cases
no calculations are done. Other values may be set by SVD or BSPLIN. This
function requires functions BSPLIN, BSPEVN, SVD and SVDEVL.
128. LINFITXY Subroutine to calculate the least squares straight line fit
when there is error in both x and y values, using the technique described in
Section 10.2.4. For simplicity it is assumed that all data points have the same
errors and correlation. N is the number of data points. X and Y are real arrays
of length N, specifying the data values. Fit of the form Y(I) = a + b × X(I) is
to be calculated. SIGX and SIGY are the estimated errors in X and Y values,
while RHO is the correlation between the errors in X and Y. XI and YI are
the arrays which will contain the fitted values of X and Y. Since both X and Y
have errors fitted values of both need to be calculated. A and B are the fitted
values of the intercept and slope, respectively. Thus YI(I) = A + B × XI(I).
CHI is the value of χ2 at the fit as defined by Eq. (10.49). IER is the error
parameter. IER = 617 implies that the discriminant of the quadratic (10.59) is
negative and the calculations are aborted.
129. NLLSQ Function to calculate the χ2 function for a nonlinear least
squares fit for use with function BFGS. By suppressing the gradient calcu-
lations it can also be used with function NMINF. This is the function to be
minimised by BFGS. N is the number of parameters to be fitted, A is an array
of length N containing the values of the parameters at which the function is
1088 Appendix C. C Programs

required to be calculated. F is the calculated value of the function. G is an

∂F
array of length N, containing the calculated derivatives. G[i] will contain ∂a i
.
The data points are passed through global variables NNLSQ, FXLSQ, XLSQ,
EFLSQ, FX1LSQ, which must be initialised in the calling function before call-
ing BFGS. Here, XLSQ, FXLSQ, EFLSQ are arrays of length NPL containing
respectively, the values of abscissas, function values and estimated errors in
function values at each point. NNLSQ is the number of data points in the ta-
ble, NNLSQ ≤ NPL. FX1LSQ is an array of length NP which will contain the
fitted value of the function at all points. Although EFLSQ[I] should contain
the estimated error in FXLSQ[I], in many cases it is found that multiplying
all elements of EFLSQ by suitable constant improves the convergence of BFGS
significantly without changing the minimum. Only the value of χ2 will need to
be scaled suitably. This function needs FCN to calculate the function value at
any X for specified values of the parameters. It may be noted that there is no
provision to pass on the name of a function to this function and hence the name
has to be explicitly changed in the function. Function FCN(N, A, X, F, DF)
must be supplied by the user. Here N is the number of parameters, A is an array
of length N containing the values of parameters. X specifies the point where the
function value needs to be calculated. F should contain the calculated function
value F(X, A), while DF is an array of length N containing the derivatives of F.
∂F
DF[I] should give ∂A i
. By suppressing the derivative calculations NLLSQ can
also be used with function NMINF. The corresponding version may be found
in NLLSQ F.
130. DFT Function to calculate the discrete Fourier transform (DFT) using
normal sum. This function is applicable to arbitrary number of points, but it
requires O(N 2 ) arithmetic operations, and hence should be used only when N
is relatively small. N is the number of data points. CG is an array of length 2N,
which should contain the data points. CF is an array of length 2N which will
contain the Fourier transform of CG. Both CF and CG are complex and hence
the arrays should contain the real and imaginary parts stored consecutively.
IFLG is a flag, if IFLG ≥ 0 the DFT is calculated, while if IFLG < 0 the
inverse DFT will be calculated. Error status can be obtained from the returned
value of DFT. A value of 611 implies that N < 2, in which case, no calculations
are performed. It may be noted that when inverse transform is calculated the
result will need to be divided by N to match the original data. The function
does not perform the division. It may be noted that even while calculating the
inverse transform the input should be provided in array CG and calculated
transform will be available in array CF.
131. FFT Function to calculate the discrete Fourier transform (DFT) using a
FFT algorithm. N is the number of data points which must be equal to a power
of 2. CG is an array of length 2N, which should contain the data points when
calling the function. After execution, the DFT will be overwritten on the same
array CG. Hence, if necessary a copy of the original data should be preserved
for later use before calling the function. CG is complex and hence both real and
C.10. Functional Approximations 1089

imaginary parts must be stored in the array in consecutive locations. IFLG is a

flag, if IFLG ≥ 0 the DFT is calculated, while if IFLG < 0 the inverse DFT will
be calculated. Error status can be obtained from the returned value of FFT. A
value of 611 implies that N < 2, in which case, no calculations are performed.
DFT = 631 implies that N is not a power of 2. This test is performed towards
the end of the calculation and hence in this case, the contents of array CG will
be destroyed. It may be noted that when inverse transform is calculated the
result will need to be divided by N to match the original data. The function
does not perform the division.
132. FFTR Function to calculate the discrete Fourier transform (DFT) of real
data using a FFT algorithm. N is the number of data points which must be
equal to a power of 2. CG is an array of length 2 × (N/2), which should contain
the data points when calling the function. In the calling function it may be
treated as a real array of length N stored in normal order. After execution, the
DFT will be overwritten on the same array CG as explained in Section 10.6. The
DFT will be complex and it will contain the real and imaginary parts stored in
consecutive memory locations. If necessary, another copy of the original data
should be preserved for later use before calling the function. IFLG is a flag. If
IFLG ≥ 0 the DFT is calculated, while if IFLG < 0 the inverse DFT will be
calculated. Error status can be obtained from the returned value of FFTR. A
value of 611 implies that N < 4 and no calculations are done. FFTR = 631
implies that N is not a power of 2. Since this test is performed towards the end
of calculations, the contents of array CG will be destroyed. It may be noted
that when inverse transform is calculated the result will need to be divided by
N/2 to match the original data. The function does not perform the division.
This function requires function FFT for calculating the DFT of complex data
and function CDIV to perform complex division.
133. FFTN Function to calculate the discrete Fourier transform (DFT) in
n dimensions using a FFT algorithm. ND is the number of dimensions. NN is
an integer array of length ND. NN[I] is the number of data points along the
Ith coordinate, which must be equal to a power of 2. CG is a complex array of
length NN[0] × NN[1] × . . . × NN[ND − 1], which should contain the data points
when calling the function. The data should be stored in the normal Fortran
order with
CG [j0 + j1 NN[0] + j2 NN[0]NN[1] + · · · + jN D−1 NN[0] . . . NN[ND − 2]]
= gj0 ,j1 ,...,jN D−1 ,
(C.49)
for 0 ≤ jr < NN[r] − 1 and 0 ≤ r < ND. In fact, in the calling func-
tion CG can be treated as a (ND + 1)-dimensional array with dimension
CG[NN[ND−1]]. . .[NN[1]][NN[0]][2]. Both real and imaginary parts of data
needs to be stored in consecutive locations. It may be noted that the dimen-
sions of this ND-dimensional array must be exactly equal to the number of
data points in the corresponding variables. After execution, the DFT will be
overwritten on the same array CG. Hence, if necessary a copy of original data
1090 Appendix C. C Programs

should be preserved for later use, before calling the function. IFLG is a flag. If
IFLG ≥ 0 the DFT is calculated, while if IFLG < 0 the inverse DFT will be cal-
culated. Error status can be obtained from the returned value of FFTN. A value
of 631 implies that at least one of the NN[I], (I = 0, 1, . . . , ND−1) is not a power
of 2. Since this test is performed towards the end of calculations, the contents
of array CG will be destroyed. It may be noted that when inverse transform is
calculated the result will need to be divided by NN[0]×NN[1]×· · ·×NN[ND−1]
to match the original data. The function does not perform the division.
134. LAPINV Function to calculate the inverse Laplace transform of a given
function F (s). N is the number of points at which the value of the inverse
function is required. T is an array of length N containing the points ti , at which
the inverse transform is required. The elements ti need not be in any order, but
the last element tN should be the largest or close to the largest, since this
element is used to control the value of T0 as explained in Section 10.8. F is an
array of length N, which will contain the value of the required function at ti
after execution of the function. CFS is the name of the function routine used
to calculate the function F (s). ALPHA is an estimate for the exponential order
of the function f (t) as explained in Section 10.8. REPS is the convergence
parameter. The results are normally expected to have a relative accuracy of
REPS. However, as explained in the text, this is not guaranteed when the
convergence is slow. The results can be improved by either increasing the value
of the parameter NMAX in the function, or by removing the discontinuity which
is causing the slow convergence. The second alternative will be more effective.
Error status can be obtained from the returned value of LAPINV. A value of
61 implies that the -algorithm failed to converge for at least one of the points.
LAPINV = 62 implies that the -algorithm encountered a zero denominator at
some stage. Since only the last value will be retained, the error flag may be
misleading in those cases, where failure has occurred at more than one points.
Function CFS(CS,CF) must be supplied by the user. Here both CS and CF are
arrays of length 2 containing the real and imaginary parts of the arguments.
CF is f(CS).
135. POLD Function to evaluate a polynomial and its derivatives at any point.
N is the degree of polynomial. A is an array of length N + 1 containing the
coefficients of the polynomial. A[0] should contain the constant term and A[N]
should be the coefficient of XN . X is the point at which polynomial is to be
evaluated. ND is the number of derivatives to be evaluated. It would evaluate
the first ND derivatives. The first derivative is always evaluated, irrespective of
the value of ND. PD is an array of length ND, which will contain the computed
values of the derivatives. PD[I−1] will contain the Ith derivative of polynomial.
The value of polynomial is returned as POLD.
136. RMK To evaluate a rational function at any point. M is the degree
of numerator, while K is the degree of denominator. A is an array of length
M + 1 containing the coefficients of the polynomial in the numerator. A[0]
should contain the constant term and A[M] should be the coefficient of XM .
C.10. Functional Approximations 1091

B is an array of length K + 1 containing the coefficients of the polynomial in

the denominator. B[0] should contain the constant term and B[K] should be
the coefficient of XK . X is the point at which the rational function is to be
evaluated. The value of rational function is returned as RMK.
137. RMK1 To evaluate a rational function at any point. This is the same
as RMK, except that the constant term for polynomial in the denominator is
assumed to be one and hence is not supplied. M is the degree of numerator,
while K is the degree of denominator. A is an array of length M + 1 containing
the coefficients of the polynomial in the numerator. A[0] should contain the
constant term and A[M] should be the coefficient of XM . B is an array of length
K containing the coefficients of the polynomial in the denominator. B[I−1]
should contain the coefficient of XI . X is the point at which the rational function
is to be evaluated. The value of rational function is returned as RMK1.
138. RMKD To evaluate a rational function and its first derivative at any
point. M is the degree of numerator, while K is the degree of denominator.
A is an array of length M + 1 containing the coefficients of the polynomial in
the numerator. A[0] should contain the constant term and A[M] should be the
coefficient of XM . B is an array of length K + 1 containing the coefficients of the
polynomial in the denominator. B[0] should contain the constant term and B[K]
should be the coefficient of XK . X is the point at which the rational function
is to be evaluated. DF will contain the computed value of the derivative. The
value of rational function is returned as RMKD.
139. RMKD1 To evaluate a rational function and its first derivative at any
point. This is the same as RMKD, except that the constant term for polynomial
in the denominator is assumed to be one and hence is not supplied. M is the
degree of numerator, while K is the degree of denominator. A is an array of
length M + 1 containing the coefficients of the polynomial in the numerator.
A[0] should contain the constant term and A[M] should be the coefficient of
XM . B is an array of length K containing the coefficients of the polynomial in
the denominator. B[I−1] should contain the coefficient of XI . X is the point at
which the rational function is to be evaluated. DF will contain the computed
value of the derivative. The value of rational function is returned as RMKD1.
140. PADE Function to calculate the coefficients of Padé approximation
Rmk (x) from the known coefficients of Maclaurin series. M and K are the de-
grees of polynomials in the numerator and the denominator, respectively. A is
an array of length M + K + 1, which will contain the coefficients of the Padé
approximation. A[i], (i = 0, . . . , K − 1) is the coefficient of xi+1 in the denomi-
nator, the constant term being assumed to be unity, A[K + i], (i = 0, 1, . . . , M)
is the coefficient of xi in the numerator, which gives the Padé approximation
A[K] + A[K + 1]x + A[K + 2]x2 + · · · + A[K + M]xM
RMK (x) = , (C.50)
1 + A[0]x + A[1]x2 + · · · + A[K − 1]xK
C is an array of length M + K + 1 containing the coefficients of the Maclaurin
series for the required function. C[i] should contain the coefficient of xi in the
1092 Appendix C. C Programs

Maclaurin series. These coefficients must be supplied by the user. Error status
can be obtained from the returned value of PADE. A value of 612 implies
that either M < 0 or K < 0, in which case no calculations are performed.
Other values may be set by the function GAUELM, which is called to solve the
system of linear equations. This function requires function GAUELM to solve
the system of linear equations.
141. CHEBCF Function to convert a power series into a series of Chebyshev
polynomials and vice versa. N is the degree of polynomial. C and P are arrays of
length N + 2 containing respectively, the coefficients of the Chebyshev and the
power series expansions. C[i] is the coefficient of Ti (x) in Chebyshev expansion,
while P[i] is the coefficient of xi in the power series. It should be noted that
C[0] is the coefficient of T0 (x) and is not doubled as in the normal Chebyshev
expansions. IFLG is the flag which decides the type of conversion required. If
IFLG = 0, then coefficients of Chebyshev expansion will be calculated. In that
case, the power series coefficients must be supplied. After execution, the array
C will contain the coefficients of Chebyshev expansion, while the array P is
unaffected. If IFLG 6= 0, then the coefficients of power series will be calculated.
In that case, the coefficients of Chebyshev expansion must be supplied and after
execution, the array P will contain the coefficients of power series, while the
contents of C will be destroyed, since this array is used as a scratch space by
the function. The returned value of the function is always zero.
142. CHEBEX Function to calculate the coefficients of Chebyshev expansion
of a function that can be evaluated at any required point. It uses orthogonality
of Chebyshev polynomials over a set of discrete points to find the coefficients
and the value will only be approximately correct.
N−1
1 X
FUN(x) = c0 + ci Ti (x). (C.51)
2 i=1

N is the number of coefficients required. This number should be much larger

than the actual number of coefficients needed. The accuracy of computed co-
efficients increases with N. There is no check to test the accuracy and it has
to be ascertained by recomputing the coefficients with larger N (say 2N) and
comparing the two values. C is an array of length N, which will contain the com-
puted coefficients. FUN is the name of the function routine supplied to calculate
the required function. Error status can be obtained from the returned value of
CHEBEX. A value of 613 implies that N < 10, in which case no calculations
are done. Function FUN(X) must be supplied by the user.
143. CHEBAP Function to calculate the coefficients of Rational function
Chebyshev approximations Tmk (x) from the known coefficients of expansion
in Chebyshev polynomials. It should be noted that, this function does not gen-
erate minimax approximations, but if the coefficients of Chebyshev expansion
fall off rapidly, then the approximation will be close to minimax. M and K are
the degrees of polynomials in the numerator and the denominator, respectively.
C.10. Functional Approximations 1093

A is an array of length M + K + 1, which will contain the coefficients of the ra-

tional function approximation. A[i − 1], (i = 1, . . . , K) is the coefficient of Ti (x)
in the denominator, the constant term being assumed to be unity, A[K + i],
(i = 0, 1, . . . , M) is the coefficient of Ti (x) in the numerator, which gives the
approximation

A[K] + A[K + 1]T1 (x) + A[K + 2]T2 (x) + · · · + A[K + M]TM (x)
RMK (x) = .
1 + A[0]T1 (x) + A[1]T2 (x) + · · · + A[K − 1]TK (x)
(C.52)
C is an array of length M + 2K + 1, containing coefficients of the Chebyshev
series for the required function. C[i] should contain the coefficient of Ti (x) in the
Chebyshev series. These coefficients must be supplied by the user. Following
the usual convention of Chebyshev expansion, the coefficient of T0 (x) in the
expansion is C[0]/2. Error status can be obtained from the returned value of
CHEBAP. A value of 612 implies that either M < 0 or K < 0, in which case, no
calculations are performed. Other values may be set by the function GAUELM,
which is called to solve the system of linear equations. This function requires
function GAUELM to solve the system of linear equations.
144. REMES Function to calculate the minimax rational function approxi-
mation for a given function over a finite interval, using the second algorithm of
Remes. M and K are the degrees of polynomials in the numerator and the de-
nominator, respectively. N is the number of points which will be used for initial
scan of extrema in the error curve. This number should be at least 3(M+K+1),
in order to be able to isolate different extrema. XL and XU are respectively, the
lower and upper limits of the interval over which the approximation is required.
A is an array of length M + K + 2, which will contain the coefficients in the
same form as that given by the function PADE. A[i − 1], (i = 1, 2, . . . , K) is
the coefficient of xi in the denominator, the constant term being unity, while
A[K + i], (i = 0, 1, . . . , M) is the coefficient of xi in the numerator. If IFLG = 0,
then the initial guess for these coefficients must be supplied. X and F are arrays
of length N with F[I] containing the value of the function at X[I]. These values
need not be stored before calling the function, since the function itself selects
a uniform mesh and calculates the value of the function at the required points.
EX is an array of length M + K + 5 containing the extrema of the error curve.
If IFLG = 2, then the initial guess for the extrema must be supplied, otherwise
these values are not required. In all cases, after execution, EX will contain the
extrema of the error curve for the final approximation. IE is the number of
extrema, which should be M + K + 2, if the error curve is of the standard form.
Some types of nonstandard error curves can be handled by this function (Ex-
ample 10.11). If IFLG = 2, then the value of IE must be supplied. EMAX is the
maximum error in the final approximation calculated by the function. EPS is
the required tolerance. The iteration for calculating the coefficients of rational
function is continued until the maximum error differs by less than EPS. The
Remes iteration is continued until the difference between different extrema of
error curve is less than 1% of the maximum error. EPSM specifies the tolerance
1094 Appendix C. C Programs

for finding the extrema of the error curve. In general, it is found that a mod-
erate value for EPS and EPSM is enough to find approximations even to very
high accuracy. IFLG is an integer parameter, which specifies the nature of ini-
tial approximation for the Remes algorithm. If IFLG = 0, then the iteration is
started from a known initial approximation. In this case, the coefficients of ini-
tial approximation must be supplied in the array A. If IFLG = 1, then no initial
approximation is required and the first iteration is performed by assuming that
the extrema of error curves are given by those of TM+K+1 (x). This is the most
useful case, if no approximation of the right form and with the correct number
of extrema is known. If IFLG = 2, then iteration is started with an initial ap-
proximation for the extrema of the error curve. In this case, the approximate
location of the extrema must be supplied in the array EX and IE (≥ M + K + 2)
should be set equal to the number of extrema. If the error curve is expected to
be nonstandard, then it will be preferable to supply only M + K + 2 extrema,
where the error is expected to be the largest and is alternating in sign. Error sta-
tus can be obtained from the returned value of REMES. A value of 614 implies
that either M < 0, or K < 0, or XU ≤ XL, or M + K + 2 > NMAXR, in which
case, no calculations are performed. The last requirement arises because of the
dimensions of global array AA, which is used to transfer the coefficients to an
auxiliary function routine for finding extrema of the function. REMES = 632
implies that the Remes iteration failed to converge to the specified accuracy in
NIT (= 30) iterations. This failure could be due to roundoff error, or because
the starting values are not sufficiently close. REMES = 633 implies that at
some stage the error curve does not have the required number of extrema and
hence the iteration cannot proceed further. This condition does not necessarily
imply that the error curve is nonstandard. Apart from these, other values may
be set by the function BRENTM which is called to find the extrema, or the
function GAUELM which is called to solve the system of linear equations. This
function requires the functions BRENTM, GAUELM, FM, FUN and FUND.
Function FUN(X) calculates the required function, while Function FUND(X)
calculates the weight function. REMES generates approximation of the form
FUN(X) ≈ FUND(X)RMK (X). If FUND(x) = 1 and FUN(x) = f (x), then
the function will calculate minimax approximation to f (x) with respect to the
absolute error, while if FUND(x) = 1/f (x) and FUN(x) = 1, then the approx-
imation will be obtained with respect to the relative error, provided f (x) 6= 0
throughout the required interval. It should be noted that the names of the func-
tions FUN and FUND are fixed and cannot be changed. The functions FUN
and FUND must be supplied by the user. The global variables MM, KK, AA,
SI are used to pass on parameters to function FM(X), which calculates the
function to be minimised. Here AA is an array of length NMAXR which con-
tains the coefficients of rational function approximation. SI is a variable which
is set to positive value if minimum is to be found and to negative value when
maximum is to be found. While SI > 10 implies that function value, FUN(X),
is known and need not be calculated. MM and KK are integer variables which
are equal to M and K respectively. The rational function approximation can be
C.10. Functional Approximations 1095

computed at any required value of X, using function RMK1(M, K, &A[K], A,

X).
145. FM Function to be used with REMES. This function is called by
BRENTM to locate the extrema of error curve. It calculates the difference
between the actual function and the rational function approximation weighted
by the required weight function. This difference is multiplied by SI which is set
to a negative value if we want to find a maxima. For initial scan SI ≥ 10, in
which case, function value is not calculated, since it is already known.
146. GAMMA Function to calculate the Gamma function for a real argument
Z ∞
Γ(x) = tx−1 e−t dt , (x > 0). (C.53)
0

The computed value should generally have relative accuracy of 10−15 . The
function first calculates the value of Gamma function at |x|. The values for
negative argument are then calculated using
π
−xΓ(−x)Γ(x) = . (C.54)
sin(πx)
For x > 1000 the Stirling’s formula is used to calculate the function value. For
8 < x ≤ 1000 a rational function approximation of form
1 √ 1
log Γ(x) = (x − ) log x − x + log( 2π) + Rmk (1/x2 ), (C.55)
2 x
is used. For 0 < x ≤ 8 the range is first translated to [2, 3] using Γ(x+1) = xΓ(x)
and then a rational function approximation over [2, 3] is used to approximate the
function value. This function does not check for overflow or invalid arithmetic
operations. Since Γ(x) increases very rapidly with x, the function evaluation will
lead to overflow for x larger than approximately 170 on a 53 bit arithmetic.
Similarly, Γ(x) diverges when x is a negative integer or zero. If overflow is
expected, then it may be better to use GAMMAL instead which gives the
logarithm of |Γ(x)|.
147. GAMMAL Function to calculate the natural logarithm of Gamma func-
tion for a real argument. It uses the same approximations as those used by
Function GAMMA, except that for large arguments (|x| > 8), it directly cal-
culates the logarithm and hence should not give overflow. It gives log(|Γ(X)|)
and hence sign should be accounted separately.
148. ERF Function to calculate the Error function
Z x
2 2
erf(x) = √ e−t dt , (C.56)
π 0
for a real argument. This function uses rational function approximations of the
form
erf(x) ≈ xRmk (x2 ), (0 ≤ x < 2);
−x2 (C.57)
e 0 1
erf(x) ≈ 1 − Rm 0 k0 (
2
), (2 ≤ x < ∞).
x x
1096 Appendix C. C Programs

While for x < 0 it uses erf(−x) = −erf(x) to evaluate the function value. Each
of the rational function approximation has relative accuracy of better than
10−15 . For large x the value may be indistinguishable from 1.
149. ERFC Function to calculate the complementary Error function
Z ∞
2 2
erfc(x) = 1 − erf(x) = √ e−t dt , (C.58)
π x
for a real argument. This function uses the same approximations as those used
by ERF(X). The value should generally have relative accuracy of 10−15 . For
small arguments the value of function may not be distinguishable from unity,
while for large negative arguments it may be indistinguishable from 2.
150. BJ0 Function to calculate the Bessel function of the first kind of order
zero, J0 (x) for a real argument. For |x| < 8 it uses a rational function approxi-
mation to J0 (x) ≈ Rmk (x2 ), while for larger arguments it uses the asymptotic
form
r
2  π π 
J0 (x) = P0 (x) cos(x − ) − Q0 (x) sin(x − ) . (C.59)
πx 4 4
The functions P0 (x) and Q0 (x) are approximated by rational function approx-
imations of form
1
P0 (x) ≈ Rmk (1/x2 ); Q0 (x) ≈ Rm0 k0 (1/x2 ). (C.60)
x
All the rational function approximations have absolute accuracy of better than
10−15 , but the resulting function values may not have the same relative accuracy
at all points. Near the zeros of J0 (x) the relative accuracy would be lower. For
large arguments, |x| > 8, where the asymptotic formula is used there could
be some cancellation between the two terms and consequent loss of significant
figures close to zeros of J0 (x).
151. BJ1 Function to calculate the Bessel function of the first kind of order
one, J1 (x) for a real argument. For |x| < 8 it uses a rational function approxi-
mation to J1 (x) ≈ xRmk (x2 ), while for larger arguments it uses the asymptotic
form
r  
2 3π 3π
J1 (x) = P1 (x) cos(x − ) − Q1 (x) sin(x − ) . (C.61)
πx 4 4

The functions P1 (x) and Q1 (x) are approximated by rational function approx-
imations of form
1
P1 (x) ≈ Rmk (1/x2 ); Q1 (x) ≈ Rm0 k0 (1/x2 ). (C.62)
x
All the rational function approximations have absolute accuracy of better than
10−15 , but the resulting function values may not have the same relative accuracy
C.10. Functional Approximations 1097

at all points. Near the zeros of J1 (x) the relative accuracy would be lower. For
large arguments, |x| > 8, where the asymptotic formula is used there could
be some cancellation between the two terms and consequent loss of significant
figures close to zeros of J1 (x).
152. BJN Function to calculate the Bessel function of the first kind of integral
order, Jn (x) for a real argument. This function calculates Jk (x) for k between 0
and n. Since a recurrence relation is used to calculate these values, all the values
are returned. Here N is the maximum order of Bessel functions required. For
positive N, Jk (x) with positive k = 0, 1, . . . , N will be calculated. For negative
N, Jk (x) with negative k = 0, −1, . . . , −N are calculated. XB is the argument
at which the function values are required. BJ is an array which will contain the
computed values of the Bessel functions. BJ[|i|] = Ji (XB). Since this array is
also used pas scratch space the length of the array must be at least, |N| + 16 +
max(25, 5 |N |). For |N| < |XB| the recurrence relation is stable in forward
direction and hence it is used to calculate Jk (x) in a straightforward manner
using the values of J0 (x) and J1 (x). For smaller XB, the recurrence relation
is unstable in forward direction and hence is used in backward direction as
explained in {2.38} In this case for some suitable n1 , we start using Jn1 = 0
and Jn1 −1 = 1 and calculate lower order functions using the recurrence relation
2k
Jk−1 (x) = Jk (x) − Jk+1 (x), k = n1 − 1, . . . , 2, 1. (C.63)
x
The normalisation is then found using the relation
∞
X
J0 (x) + 2 J2k (x) = 1. (C.64)
k=1

Thus we can divide all values by the calculated sum to obtain the correct values
for Jk (x). To use this technique we need to select a suitably large value of n1
to start the recurrence relation. Accuracy of resulting function values can be
checked by comparing the value of Jn+1 (x) without normalisation. The recip-
rocal of this value will give an estimate of accuracy achieved by this process.
Thus if the computed value of Jn+1 (x) is not large enough an error message is
printed out by the routine. This should not normally happen, but in case for
some combination of N and XB, the chosen value of n1 is not large enough,
the error message will be printed out. In such cases the choice of N1 in the
routine should be increased suitably. In principle, this technique can be used
for all values of x, but n1 has to be larger than x and hence for large x, we
will need lot of computation using this technique. Further, the use of the recur-
rence relation in this form results in overflow even when the required values are
within the range of computer arithmetic. This problem is more severe at low x,
where the values can be easily computed using the series expansion. Thus we
use the series expansion for x ≤ 4 for computing Jn (x) and Jn−1 (x). After that
the recurrence relation is used in backward direction to get other values. The
overflow problem becomes very acute when single precision arithmetic is used
1098 Appendix C. C Programs

and it is not recommended to use this routine in single precision. The accuracy
attained by this routine will depend on the values of N and XB, but in general
we expect accuracy of order of 10−15 which is the accuracy with which J0 (x)
and J1 (x) are calculated. This routine needs BJ0 and BJ1 to calculate J0 (x)
and J1 (x).
153. BY0 Function to calculate the Bessel function of the second kind of
order zero, Y0 (x) for a real argument. This function is not defined for x ≤ 0,
and the function simply returns a value of zero without any warning or error
flag. User must ensure that x is positive. For x < 8 it uses a rational function
approximation of the form
2
Y0 (x) = J0 (x)(log(x/2) + γ) + x2 Rmk (x2 ) , (C.65)
π
where γ is the Euler’s constant. J0 (x) is computed using a rational function
approximation. For larger arguments the asymptotic form is used:
r
2  π π 
Y0 (x) = P0 (x) sin(x − ) + Q0 (x) cos(x − ) . (C.66)
πx 4 4
The functions P0 (x) and Q0 (x) are approximated by rational function approx-
imations of form
1
P0 (x) ≈ Rmk (1/x2 ); Q0 (x) ≈ Rm0 k0 (1/x2 ). (C.67)
x
All the rational function approximations have absolute accuracy of better than
10−15 , but the resulting function values may not have the same relative accuracy
at all points. Near the zeros of Y0 (x) the relative accuracy would be lower. For
large arguments, x > 8, where the asymptotic formula is used there could be
some cancellation between the two terms and consequent loss of significant
figures close to zeros of Y0 (x).
154. BJY0 Function to calculate the Bessel function of the first and second
kind of order zero, J0 (x), Y0 (x) for a real argument. Since computation of Y0 (x)
involves J0 (x) also, this function is provided to give both values together and
can be used if both functions are required. The function of the second kind is
not defined for x ≤ 0, and the function simply returns a value of zero without
any warning or error flag. The value of J0 (x) should be calculated even in
this case. User must ensure that x is positive if Y0 (x) is also required. For
calculating J0 (x) alone it will be more efficient to use BJ0. This function is
essentially same as BY0, except that in this case the function values have to
appear as arguments while calling the function. The function BY0 can also be
modified to include BJ0 in its argument list, in which case it can return that
value also. In that case the line computing BJ0 should be uncommented. Here
XB is the argument at which the function values are required. BJ0 and BY0
are the computed values of J0 (XB) and Y0 (XB).
C.10. Functional Approximations 1099

155. BY1 Function to calculate the Bessel function of the second kind of
order one, Y1 (x) for a real argument. This function is not defined for x ≤ 0,
and the function simply returns a value of zero without any warning or error
flag. User must ensure that x is positive. For x < 8 it uses a rational function
approximation of the form
 
2 1 2
Y1 (x) = J1 (x)(log(x/2) + γ) − − xRmk (x ) , (C.68)
π x

where γ is the Euler’s constant. J1 (x) is computed using a rational function

approximation. For larger arguments the asymptotic form is used:
r  
2 3π 3π
Y1 (x) = P1 (x) sin(x − ) + Q1 (x) cos(x − ) . (C.69)
πx 4 4

The functions P1 (x) and Q1 (x) are approximated by rational function approx-
imations of form
1
P1 (x) ≈ Rmk (1/x2 ); Q1 (x) ≈ Rm0 k0 (1/x2 ). (C.70)
x
All the rational function approximations have absolute accuracy of better than
10−15 , but the resulting function values may not have the same relative accuracy
at all points. Near the zeros of Y1 (x) the relative accuracy would be lower. For
large arguments, x > 8, where the asymptotic formula is used there could be
some cancellation between the two terms and consequent loss of significant
figures close to zeros of Y1 (x).
156. BJY1 Function to calculate the Bessel function of the first and second
kind of order one, J1 (x), Y1 (x) for a real argument. Since computation of Y1 (x)
involves J1 (x) also, this function is provided to give both values together and
can be used if both functions are required. The function of the second kind is
not defined for x ≤ 0, and the function simply returns a value of zero without
any warning or error flag. The value of J1 (x) should be calculated even in
this case. User must ensure that x is positive if Y1 (x) is also required. For
calculating J1 (x) alone it will be more efficient to use BJ1. This function is
essentially same as BY1, except that in this case the function values have to
appear as arguments while calling the function. The function BY1 can also be
modified to include BJ1 in its argument list, in which case it can return that
value also. In that case the line computing BJ1 should be uncommented. Here
XB is the argument at which the function values are required. BJ1 and BY1
are the computed values of J1 (XB) and Y1 (XB).
157. BYN Function to calculate the Bessel function of the second kind of
integral order, Yn (x) for a real argument. This function calculates Yk (x) for k
between 0 and n. This function is not defined for x ≤ 0, and the function does
not calculate the value and does not give any warning or error flag. In this case
values in array BY will be preserved and may be misinterpreted as calculated
1100 Appendix C. C Programs

value. User must ensure that x is positive. Since a recurrence relation is used to
calculate these values, all the values are returned. Here N is the maximum order
of Bessel functions required. For positive N, Yk (x) with positive k = 0, 1, . . . , N
will be calculated. For negative N, Yk (x) with negative k = 0, −1, . . . , −N are
calculated. XB is the argument at which the function values are required. BY is
an array of length |N| + 1 which will contain the computed values of the Bessel
functions. BY[|i|] = Yi (XB). Since the recurrence relation for Yn (x)
2n
Yn+1 (x) = Yn (x) − Yn−1 (x), (C.71)
x
is stable in the forward direction for all x, these values are computed in a
straightforward manner using the values of Y0 (x) and Y1 (x). This function
needs BY0 and BY1 to calculate Y0 (x) and Y1 (x).
158. SPHBJN Function to calculate the spherical Bessel function of integral
order, r
π
jn (x) = Jn+1/2 (x), (C.72)
2x
for a real argument. This function calculates jk (x) for k between 0 and n. Since a
recurrence relation is used to calculate these values, all the values are returned.
Here N is the maximum order of Bessel functions required. For positive N,
jk (x) with positive k = 0, 1, . . . , N will be calculated. For negative N, jk (x)
with negative k = 0, −1, . . . , −N are calculated. XB is the argument at which
the function values are required. BJ is an array which will contain the computed
values of the Bessel functions. BJ[|i|] = ji (XB). Since this array is also used
p as
scratch space the length of the array must be at least, |N|+16+max(25, 5 |N |).
For |N| < |XB| or for N < 0 the recurrence relation is stable in forward direction
and hence it is used to calculate jk (x) in a straightforward manner using the
values of
sin x sin x cos x cos x
j0 (x) = ; j1 (x) = − or j−1 (x) = . (C.73)
x x2 x x
For smaller XB and N > 0, the recurrence is unstable in forward direction and
hence is used in backward direction as explained in {2.38} In this case for some
suitable n1 , we start using jn1 = 0 and jn1 −1 = 1 and calculate lower order
functions using the recurrence relation
2k + 1
jk−1 (x) = jk (x) − jk+1 (x), k = n1 − 1, . . . , 2, 1. (C.74)
x
The normalisation is then found using the value of j0 (x). Thus we can divide all
values by the calculated ratio S of j0 (x) as computed from recurrence relation
to its actual value, to obtain the correct values for jk (x). To use this technique
we need to select a suitably large value of n1 to start the recurrence relation.
Accuracy of resulting function values can be checked by comparing the value
of jn+1 (x) before normalisation. Its reciprocal will give an estimate of accuracy
C.10. Functional Approximations 1101

achieved by this process. Thus if the computed value of jn+1 (x) is not large
enough an error message is printed out by the routine. This should not normally
happen, but in case for some combination of N and XB, the chosen value of n1 is
not large enough, the error message will be printed out. In such cases the choice
of N1 in the routine should be increased suitably. In principle, this technique
can be used for all values of x, but n1 has to be larger than x and hence for large
x, we will need lot of computation using this technique. Further, the use of the
recurrence relation in this form results in overflow even when the required values
are within the range of computer arithmetic. This problem is more severe at low
x, where the values can be easily computed using the series expansion. Thus we
use the series expansion for x ≤ 4 for computing jn (x) and jn−1 (x). After that
the recurrence relation is used in backward direction to get other values. The
overflow problem becomes very acute when single precision arithmetic is used
and it is not recommended to use this routine in single precision. The accuracy
attained by this routine will depend on the arithmetic used. In general, we
expect to achieve accuracy of order of ~ in these computations, since j0 (x) and
j1 (x) can be expressed in terms of trigonometric functions.
159. BI0 Function to calculate the modified Bessel function of the first kind
of order zero, I0 (x) for a real argument. For |x| < 8 it uses a rational function
approximation to I0 (x) ≈ Rmk (x2 ), while for larger arguments it uses the
asymptotic form
e|x|
I0 (x) = p Rmk (1/|x|). (C.75)
|x|
All the rational function approximations have absolute accuracy of better than
10−15 , but the resulting function values may not have the same relative accuracy
at all points.
160. BI1 Function to calculate the modified Bessel function of the first kind
of order one, I1 (x) for a real argument. For |x| < 8 it uses a rational function
approximation to I1 (x) ≈ xRmk (x2 ), while for larger arguments it uses the
asymptotic form
e|x|
I1 (x) = p Rmk (1/|x|). (C.76)
|x|
All the rational function approximations have absolute accuracy of better than
10−15 , but the resulting function values may not have the same relative accuracy
at all points.
161. BIN Function to calculate the modified Bessel function of the first kind
of positive integral order, In (x) for a real argument. This function calculates
Ik (x) for k between 0 and n. Since a recurrence relation is used to calculate
these values, all the values are returned. Here N is the maximum order of Bessel
functions required, Ik (x) with k = 0, 1, . . . , N will be calculated. N must be
positive, otherwise |N| will be used. XB is the argument at which the function
values are required. BI is an array which will contain the computed values of
the Bessel functions. BI[i] = Ii (XB). Since this array is also used as scratch
1102 Appendix C. C Programs
√
space the length of the array must be at least, N + 16 + max(25, 5 N ). For
N < XB − 10 the recurrence relation is stable in forward direction and hence it
is used to calculate Ik (x) in a straightforward manner using the values of I0 (x)
and I1 (x). For smaller XB, the recurrence is unstable in forward direction and
hence is used in backward direction as explained in {2.38} In this case for some
suitable n1 , we start using In1 = 0 and In1 −1 = 1 and calculate lower order
functions using the recurrence relation
2k
Ik−1 (x) = Ik (x) + Ik+1 (x), k = n1 − 1, . . . , 2, 1. (C.77)
x
The normalisation is then found using the directly computed value of I0 (x).
Thus we can divide all values by the calculated ratio for I0 (x) to obtain the
correct values for Ik (x). To use this technique we need to select a suitably large
value of n1 to start the recurrence relation. Accuracy of resulting function
values can be checked by comparing the value of In+1 (x) before normalisation.
Its reciprocal will give an estimate of accuracy achieved by this process. Thus if
the computed value of In+1 (x) is not large enough an error message is printed
out by the routine. This should not normally happen, but in case for some
combination of N and XB, the chosen value of n1 is not large enough, the error
message will be printed out. In such cases the choice of N1 in the routine should
be increased suitably. In principle, this technique can be used for all values of
x, but n1 has to be larger than x and hence for large x, we will need lot of
computation using this technique. Further, the use of the recurrence relation
in this form results in overflow even when the required values are within the
range of computer arithmetic. This problem is more severe at low x, where
the values can be easily computed using the series expansion. Thus we use
the series expansion for x ≤ 4 for computing In (x) and In−1 (x). After that
the recurrence relation is used in backward direction to get other values. The
overflow problem becomes very acute when single precision arithmetic is used
and it is not recommended to use this routine in single precision. The accuracy
attained by this routine will depend on the values of N and XB, but in general,
we expect accuracy of order of 10−15 which is the accuracy with which I0 (x)
and I1 (x) are calculated. This routine needs BI0 and BI1 to calculate I0 (x) and
I1 (x).
162. BK0 Function to calculate the modified Bessel function of the second
kind of order zero, K0 (x) for a real argument. This function is not defined for
x ≤ 0, and the function simply returns a value of zero without any warning
or error flag. User must ensure that x is positive. For x < 8 it uses a rational
function approximation of the form
K0 (x) = −I0 (x)(γ + log(x/2)) + x2 Rmk (x2 ), (C.78)
where γ is the Euler’s constant and I0 (x) is computed using appropriate rational
function approximation. For larger arguments the asymptotic form is used:
e−x
K0 (x) = √ Rmk (1/x). (C.79)
x
C.10. Functional Approximations 1103

All the rational function approximations have absolute accuracy of better than
10−15 , but the resulting function values may not have the same relative accuracy
at all points when different terms are combined.
163. BK1 Function to calculate the modified Bessel function of the second
kind of order one, K1 (x) for a real argument. This function is not defined for
x ≤ 0, and the function simply returns a value of zero without any warning
or error flag. User must ensure that x is positive. For x < 8 it uses a rational
function approximation of the form

1
K1 (x) = I1 (x)(γ + log(x/2)) + − xRmk (x2 ), (C.80)
x
where γ is the Euler’s constant and I1 (x) is computed using appropriate rational
function approximation. For larger arguments the asymptotic form is used:

e−x
K1 (x) = √ Rmk (1/x). (C.81)
x

All the rational function approximations have absolute accuracy of better than
10−15 , but the resulting function values may not have the same relative accuracy
at all points when different terms are combined.
164. BKN Function to calculate the modified Bessel function of the second
kind of positive integral order, Kn (x) for a real argument. This function cal-
culates Kj (x) for j between 0 and n. This function is not defined for x ≤ 0,
and no calculations are done but no warning is issued. In this case values in
array BK will be preserved and may be misinterpreted as calculated value. User
must ensure that x is positive. Since a recurrence relation is used to calculate
these values, all the values are returned. Here N is the maximum order of Bessel
functions required, Kj (x) with j = 0, 1, . . . , N will be calculated. N must be
positive, otherwise |N| will be used. X is the argument at which the function
values are required. BK is an array of length N + 1 which will contain the
computed values of the Bessel functions. BK[i] = Ki (XB). Since the recurrence
relation is stable in forward direction for all x, it is used in a straightforward
manner using the values of K0 (x) and K1 (x). This function needs BK0 and
BK1 to calculate K0 (x) and K1 (x).
165. DAWSON Function to calculate the Dawson’s integral
Z x
2 2
D(x) = e−x et dt , (C.82)
0

using rational function approximation. The range is split into 4 parts. In

[0, 2.5), [2.5, 4.0), [4.0, 5.5) approximation of the form D(x) = xRmk (x2 ) is used
while for higher values an approximation of the form D(x) = (1/x)Rmk (1/x2 )
is used. All approximations have a relative accuracy of 10−15 .
1104 Appendix C. C Programs

166. FERMM05 Function to calculate the Fermi integrals

∞
tk dt
Z
Fk (x) = , (C.83)
0 et−x+1

for k = −1/2 using rational function approximations. The range is split into
3 parts. For x < 2 (including all negative values) an approximation of the
form Fk (x) = ex Rmk (ex ) is used. For 2 ≤ x < 10 an approximation of form
Fk (x) = Rmk (x) is used, while for higher values of x an approximation of the
form Fk (x) = xk+1 Rmk (1/x2 ) is used. All the approximations have a relative
accuracy of approximately 10−15 .

167. FERM05 Function to calculate the Fermi integrals for k = 1/2 using
approximations similar to those for k = −1/2.

168. FERM15 Function to calculate the Fermi integrals for k = 3/2 using
approximations similar to those for k = −1/2.

169. FERM25 Function to calculate the Fermi integrals for k = 5/2 using
approximations similar to those for k = −1/2.

170. PLEG Function to calculate Legendre polynomial for a specified value of

x. L is the order of polynomial and X is the argument at which the polynomial
value is required. P is an array of length L + 1, which will contain the computed
values of the Legendre polynomials for all orders up to L, P[j] = Pj (X). The
recurrence relation is used to compute the values.

171. PLM Function to calculate associated Legendre functions, Plm (x) for a
specified value of x. L, M define the order of Legendre functions, L > 0 and
|M| ≤ L. No error message is issued if L, M do not satisfy these constraints.
X is the argument at which the function value is required. P is an array of
length L+1, which will contain the computed values of the associated Legendre
functions. P[j] = PjM (X) for j ≥ M. The function first computes the value of
M
PM . For positive m,
m
Y
m
Pm = (2i − 1)(1 − x2 )m/2 . (C.84)
i=1

For negative m there is extra normalisation factor. Using this value the recur-
rence relation

(n − m)Pnm (x) = (2n − 1)xPn−1

m m
(x) − (n − 1 + m)Pn−2 (x), (C.85)

is used to compute the value of PLM .

C.10. Functional Approximations 1105

172. YLM Function to calculate the Spherical harmonic,

Ylm (θ, φ) = (−1)m clm Plm (cos θ)eimφ , (C.86)

where
(2l + 1)(l − m)!
c2lm = , (C.87)
4π(l + m)!
is chosen to ensure that integral of |Ylm |2 over the unit sphere is 1. L is the
degree, M is the azimuthal order. L > 0, |M| ≤ L. THETA and PHI are real
variables specifying the angular coordinates θ and φ respectively. Y is an array
of length 2 containing the real and imaginary parts of spherical harmonic. It
is possible to use the argument x = cos θ instead of θ. This function requires
PLM to compute the associated Legendre functions. YLM X is the version of
YLM with x = cos θ as the argument.
173. MINMAX Function to calculate minimax rational function approxima-
tion of discrete data using a differential correction algorithm. M and K are the
degrees of polynomials in the numerator and the denominator, respectively. N
is the number of data points which should be at least M + K + 2. A is an
array of length M + K + 2, which will contain the coefficients in the same form
as that given by the function PADE. A[i − 1], (i = 1, 2, . . . , K) is the coeffi-
cient of xi in the denominator, the constant term being unity, while A[K + i],
(i = 0, 1, . . . , M) is the coefficient of xi in the numerator. At the time of call-
ing the function, this array should contain the initial approximation for the
coefficients. If K = 0, then the algorithm will converge from arbitrary initial
approximation and hence all elements of A may be set to zero or any other suit-
able value. For K 6= 0 the iteration is unlikely to converge, unless the starting
values are close to the actual values of the coefficients. X and F are arrays of
length N, containing the input data, with F[I] giving the value of the function
at X[I]. These values must be supplied. EMAX is the maximum error in the
final approximation calculated by the function. EPS is the required tolerance,
the iteration is continued until the change in EMAX is less than EPS. Error
status can be obtained from the returned value of MINMAX. A value of 615
implies that either M < 0, or K < 0, or M + K + 2 > N, in which case, no
calculations are performed. MINMAX = 634 implies that the iteration failed
to converge to the specified accuracy. Other values may be set by the function
SIMPX. This function requires the function SIMPX to solve the LP problem.
174. POLYL1 Function to calculate polynomial L1 -approximation of discrete
data. M is the degree of polynomial. N is the number of data points, which
should be at least M + 2. A is an array of length M + 2, which will contain the
coefficients. A[i], (i = 0, 1, 2, . . . , M) is the coefficient of xi . X and F are arrays
of length N, containing the input data, with F[I] containing the value of the
function at X[I]. These values must be supplied by the user. ESUM is the sum
of the magnitude of the error at all points (i.e., the L1 norm of the residual) in
the final approximation calculated by the function. EPS is the expected level of
1106 Appendix C. C Programs

roundoff error, this parameter is passed on to the function SIMPL1 for simplex
iteration. It is used to decide the sign of cost coefficients and pivot elements.
Error status can be obtained from the returned value of POLYL1. A value of
616 implies that either M < 0, or M + 2 > N, in which case no calculations
are performed. Other values may be set by the function SIMPL1. This function
requires the function SIMPL1 to solve the LP problem using a slightly modified
simplex algorithm.
175. LINL1 Function to calculate a general linear L1 -approximation in terms
of arbitrary basis functions for discrete data. Approximation of the form
M−1
X
f (x) ≈ ai φi (x), (C.88)
i=0

are sought, where φi (x) are the basis functions. There is no restriction on the
basis functions, apart from the fact that they should be independent on the set
of points in the table. This function can be used for approximation in multiple
dimensions as points need not be restricted to one dimension. M is the number
of basis functions. N is the number of data points, which should be at least
M + 1. A is an array of length M + 1, which will contain the coefficients of
approximation. A[i], (i = 0, 1, . . . , M − 1) is the coefficient of φi . F is an array
of length N, containing the input data, F[I] should contain the value of the
function at Ith point. These values must be supplied by the user. G is an array
of length IG × N containing the values of basis functions at each point in the
table, G[j][i] = φi (xj ). These values must be supplied by the user. It may be
noted that in this case the coordinates xj are not required by the function
and it could even be a vector for approximation in multiple dimension. It is
users responsibility to calculate the values φi (xj ) accordingly. IG is the second
dimension of G as specified in the calling function (IG ≥ M). ESUM is the sum
of the magnitude of the error at all points (i.e., the L1 norm of the residual) in
the final approximation calculated by the function. EPS is the expected level of
roundoff error, this parameter is passed on to the function SIMPL1 for simplex
iteration. It is used to decide the sign of cost coefficients and pivot elements.
Error status can be obtained from the returned value of LINL1. A value of 616
implies that M ≤ 0, or M + 1 > N, in which case no calculations are performed.
Other values may be set by the function SIMPL1. This function requires the
function SIMPL1 to solve the LP problem using a slightly modified simplex
algorithm.
176. SIMPL1 Function to solve LP problems using a modified version of
the simplex algorithm, specially suitable for LP problems arising in L1 -
approximations. This function is called by function POLYL1 or LINL1 and is
similar to function SIMPX. The LP problem is assumed to be in the standard
form and the initial tableau is supplied in the array A of length IA×(N − M + 1)
with IA ≥ M + 2. IA is the second dimension of A, exactly as specified in the
calling function. N is the total number of variables in the given problem. M is
C.11. Algebraic Eigenvalue Problem 1107

the number of constraints in the problem. ID and IV are integer arrays of length
M + 1 and N − M + 1 respectively, used to store permutations of the original
variables. AEPS is a parameter used to control roundoff error. Any quantity less
than AEPS in magnitude is assumed to be zero. Error status can be obtained
from the returned value of SIMPL1. A value of 63 implies that the objective
function is unbounded from below and the optimal feasible vector does not
exist. SIMPL1 = 635 implies that the simplex iteration has not converged in a
reasonable number of iterations. This failure may be due to degeneracy, since
that is not accounted for in this function. Since the problems arising out of
L1 -approximation have very high degree of degeneracy, this function may fail
to converge in some cases, because of cycling as explained in Section 8.7.

C.11 Algebraic Eigenvalue Problem

177. INVIT Function to find a real eigenvalue and the corresponding eigen-
vector of a general real matrix using inverse iteration. M is the order of the
matrix. A is an array of length IA × M containing the input matrix. IA is
the second dimension of A as specified in the calling function. P is a variable
specifying the initial shift to be used. This shift should be close to the re-
quired eigenvalue. U is an array of length M, which should contain an initial
approximation to the eigenvector. This approximation need not be close to the
required eigenvector, but must be nonzero. The initial vector U need not be
normalised. In most cases setting all components of U to 1 will be enough, but
if the initial vector so chosen is orthogonal to the required eigenvector there
may be some problem in convergence. In that case an arbitrary choice which is
not orthogonal to eigenvector will be required. For multiple eigenvalues if more
than one eigenvectors are required, then initial vector should be chosen to be
orthogonal to all known eigenvectors with same eigenvalue to ensure that an
independent eigenvector is found. After execution, the array U will contain the
required eigenvector, which is normalised such that the maximum component
is unity. IFLG is a flag to decide the kind of iteration required. If IFLG = 0,
then the shift P is kept fixed. If IFLG = 1, then the shift is varied after each
iteration, using the computed Rayleigh quotient. If IFLG = 2, then the shift is
varied using max(vs+1 ) at each iteration. IFLG = 0 should be used when the
eigenvalue is already known accurately, or when iteration is strictly required to
converge to the eigenvalue nearest to P. In other cases, we may use IFLG = 1,
if the matrix is Hermitian and IFLG = 2 otherwise. In practice, it is found
that, even for some non-Hermitian matrices, IFLG = 1 often gives faster con-
vergence, but that cannot be assumed. After execution, EI and ERC give the
estimated eigenvalue (corrected for the shift P). ERC is the estimate using
Rayleigh quotient, while EI is the estimate given by simple inverse iteration.
ERC is relevant only for Hermitian matrices, while EI is applicable to all matri-
ces. REPS is the required (absolute) tolerance. The iteration is continued until
the maximum change in the eigenvalue and the M components of the eigen-
1108 Appendix C. C Programs

vector is less than REPS. NIT is an integer variable specifying the maximum
number of iterations required to be performed. If NIT ≤ 0, then a default value
of NIT0 (= 100) will be used. Error status can be obtained from the returned
value of INVIT. A value of 106 implies that M ≤ 1 or M > IA, in which case,
no calculations are performed. INVIT = 141 implies that the vector is zero at
some stage, which is usually due to either the matrix A or the initial vector
U being zero. If the matrix is nonzero, then using a different initial vector or
changing the shift P may overcome this problem. INVIT = 142 implies that
the inverse iteration has failed to converge, which could be either because the
starting shift P is not sufficiently close to an eigenvalue, or because REPS is too
small, or because the corresponding eigenvalue has a nonlinear divisor. Apart
from these, other values may be set by function GAUELM, which is called to
solve the system of linear equations. In particular, INVIT = 121 implies that
one of pivots during the Gaussian elimination is zero. This problem can usually
be overcome by perturbing the shift slightly. If this problem occurs persistently,
it may be better to modify the Gaussian elimination function, such that zero
pivots are replaced by a suitably chosen small number. To find an eigenvector
of a generalised eigenvalue problem A(λ)x = 0, use P = 0, IFLG = 0, and the
matrix A should be set to A(λ) with λ equal to the known eigenvalue. In this
case the eigenvalue has to be determined before finding the eigenvector. To find
the left eigenvector of the matrix take transpose of the matrix while applying
the shift. This is implemented in INVIT L
178. TRED2 Function to reduce a real symmetric matrix to tridiagonal form
using Householder’s method. This function is based on the procedure tred2
in the Handbook. A is an array of length IA × N containing the matrix. After
execution, the array A will be overwritten by the transformation matrix Q, such
that QT AQ is tridiagonal. This matrix may be required for back-transforming
the eigenvectors of tridiagonal matrix to that of the original input matrix A.
N is the order of the matrix. IA is the second dimension of the array A, as
specified in the calling function, IA ≥ N. D and E are arrays of length N, giving
the diagonal and off-diagonal elements of the reduced tridiagonal matrix with
D[i] = aii and E[i + 1] = ai,i+1 = ai+1,i . REPS is the tolerance which should be
equal to η/~, where η is the smallest positive number that is representable in the
computer and ~ is the machine accuracy. For single precision arithmetic (type
float) use REPS = 10−30 . For double precision arithmetic use REPS = 10−300 .
Error status can be obtained from the returned value of TRED2. A value of 107
implies that N ≤ 1 or N > IA, in which case, no calculations are performed.
179. TRBAK Function to perform back-transformation on eigenvectors of
reduced tridiagonal matrix to obtain the eigenvectors of the original real sym-
metric matrix reduced by TRED2. This back-transformation is not required if
the eigenvectors are calculated using the function TQL2, but will be required
if the eigenvectors are calculated by the function TRIDIA. A is an array of
length IA × N containing the transformation matrix Q generated by the func-
tion TRED2. The last column of A is used as a scratch space by the function
C.11. Algebraic Eigenvalue Problem 1109

and hence its contents will be destroyed during execution. IA is the second di-
mension of A, as specified in the calling function. N is the order of the matrix.
Z is an array of length IZ×N containing the eigenvectors of the reduced tridiag-
onal matrix. After execution, this array will be overwritten by the eigenvectors
of the original matrix A. IZ is the second dimension of array Z, as specified in
the calling function. NZ is the number of eigenvectors. Thus, Z[i][j] should con-
tain the ith component of the jth eigenvector. This function simply performs
the matrix multiplication AZ to obtain the required eigenvectors. The returned
value of the function is always zero.
180. TQL2 Function to find eigenvalues and eigenvectors of ZT Z T using the
QL algorithm, where T is a symmetric tridiagonal matrix and Z is an orthogo-
nal matrix. This function is based on the procedure tql2 in the Handbook. If the
matrix Z is the transformation matrix, which reduces a real symmetric matrix
to tridiagonal form, then this function gives the eigenvectors of the original
matrix. This function can be used to find eigenvalues and eigenvectors of a
real symmetric matrix, after it is reduced to a tridiagonal form using TRED2.
In that case, Z will be the transformation matrix generated by TRED2. To
find eigenvalues and eigenvectors of a symmetric tridiagonal matrix, set the
matrix Z to an identity matrix. Z is an array of length IZ × N, containing the
transformation matrix. After execution, this array will be overwritten by the
eigenvectors of the original matrix ZT Z T , with ith column containing the ith
eigenvector. N is the order of the matrix. IZ is the second dimension of Z, as
specified in the calling function. D is an array of length N, containing the di-
agonal elements of the tridiagonal matrix with D[i] = tii . After execution, the
array D will be overwritten by the eigenvalues of the matrix. The eigenvalues
are sorted in the ascending order. E is an array of length N containing the
off-diagonal elements of the tridiagonal matrix, with E[i + 1] = ti,i+1 = ti+1,i .
E is used as a scratch space by the function and hence its contents will be
destroyed during the execution. REPS is the tolerance which should be equal
to ~, the machine accuracy. Error status can be obtained from the returned
value of TQL2. A value of 108 implies that N ≤ 1 or N > IZ, in which case, no
calculations are performed. TQL2 = 143 implies that the QL algorithm failed
to converge for some eigenvalue, in which case, the calculations are abandoned.
181. TRIDIA Function to find some eigenvalues and eigenvectors of a sym-
metric tridiagonal matrix using the Sturm sequence property, coupled with the
inverse iteration method. N is the order of the matrix. E and D are arrays of
length N containing respectively, the off-diagonal and diagonal elements of the
symmetric tridiagonal matrix, with E[i + 1] = ai,i+1 = ai+1,i and D[i] = aii . It
is assumed that all the off-diagonal elements are nonzero. Otherwise, the ma-
trix should be split into two or more parts and each part should be considered
independently. M1 and M2 are integers specifying which eigenvalues are to be
calculated. If eigenvalues λi are sorted in increasing order, then all eigenval-
ues from λM1 to λM2 will be determined. If M1 > M2, then no calculations
are performed. EI is an array of length M2 − M1 + 1, which will contain the
1110 Appendix C. C Programs

eigenvalues after execution. EPS1 is a parameter, which specifies the accuracy

to which the eigenvalues should be located by bisection. This parameter is re-
quired in some cases, since the inverse iteration with variable shift, which is
used to determine eigenvalues accurately may not always converge to the near-
est eigenvalue. If that happens, then the parameter EPS1 can be decreased. A
moderate value of the order of 0.1 or 0.01 times the typical eigenvalues should
be normally sufficient. Increasing EPS1 will improve the efficiency, provided
the inverse iteration converges. REPS is the required tolerance in the eigenval-
ues and eigenvectors. EV is an array of length IV × N which will contain the
calculated eigenvectors. EV[i][j] will contain the ith component of the jth eigen-
vector. IV is the second dimension of the array EV as specified in the calling
function, IV ≥ (M2 − M1 + 1). Error status can be obtained from the returned
value of TRIDIA. A value of 109 implies that N ≤ 1 or M2 − M1 + 1 > IV or
M1 < 0 or M2 ≥ N, in which case, no calculations are performed. Other values
may be set by function TINVIT, which is called to calculate the eigenvalues
and eigenvectors. Only the last nonzero value of error flag will be retained. This
function requires the function STURM to locate the eigenvalues and function
TINVIT to find the eigenvalues and eigenvectors of a symmetric tridiagonal
matrix and function RAN1 to generate random numbers.
182. STURM Function to locate required eigenvalues of a real symmetric
tridiagonal matrix using the method of bisection on the Sturm sequence. It is
assumed that all the off-diagonal elements are nonzero. Otherwise, the matrix
should be split into two or more parts and each part should be considered
independently. N is the order of the matrix. E and D are arrays of length N
containing respectively, the off-diagonal and diagonal elements of the symmetric
tridiagonal matrix, with E[i + 1] = ai,i+1 = ai+1,i and D[i] = aii . M1 and M2
are integers specifying which eigenvalues are to be located. If eigenvalues λi are
sorted in increasing order, then all eigenvalues from λM1 to λM2 will be located.
If M1 > M2, then no calculations are performed. EL and EU are arrays of
length M2. After execution, the ith eigenvalue should be located in the interval
(EL[i], EU[i]). NUM is an integer variable which will contain the number of
times the Sturm sequence was evaluated for locating the eigenvalues to the
specified accuracy. REPS is the specified accuracy to which the eigenvalues are
to be located. Bisection is continued until |EU[i] − EL[i]| < REPS and the
eigenvalue has been isolated. At least three bisections are performed after the
eigenvalues are isolated, that is EU[i − 1] < EL[i]. If at any stage the interval is
too small to be bisected further, then the iteration is naturally terminated. This
situation can arise when the matrix has very close eigenvalues. Error status can
be obtained from the returned value of STURM. A value of 110 implies that
M1 < 0 or M2 > N − 1, in which case no calculations are done.
183. TINVIT Function to find a specified eigenvalue and eigenvector of a sym-
metric tridiagonal matrix using inverse iteration. N is the order of the matrix. E
and D are arrays of length N containing respectively, the off-diagonal and diago-
nal elements of the symmetric tridiagonal matrix, with E[i+1] = ai,i+1 = ai+1,i
C.11. Algebraic Eigenvalue Problem 1111

and D[i] = aii . EL and EU give the lower and upper limit on the eigenvalue.
The iteration will not go beyond these limits. EI will contain the calculated
eigenvalue. EV is an array of length N, which will contain the calculated eigen-
vector. For IFLG 6= 0 it should contain the previous eigenvector determined
by the function as input, so that the function can choose initial vector or-
thogonal to it. REPS is the required accuracy. The iteration is continued until
either of the following three conditions are satisfied, (1) the relative change
in the eigenvalue as estimated using the Rayleigh quotient is less than REPS,
(2) max(vj+1 )λ > 1/REPS (where λ is an estimate for the eigenvalue), (3)
kuj+1 − uj k1 < REPS. IFLG is an integer variable used as a flag. If IFLG = 0,
then the initial vector is chosen randomly. Otherwise, the initial vector is cho-
sen to be orthogonal to the vector EV. This device can be used to ensure that
an independent vector is found when two or more eigenvalues are very close or
equal. Since only one previous eigenvector is used, this device may not be very
effective for eigenvalues of multiplicity greater than two. NUM is an integer
variable, which will contain the number of iterations required by the function.
Error status can be obtained from the returned value of TINVIT. A value of 144
implies that the inverse iteration failed to converge to the specified accuracy.
This function requires Function RAN1(SEED) to generate random numbers.
If a different function is used for this purpose, the seed should be changed
appropriately.
184. HEREVP Function to find all eigenvalues and eigenvectors of a complex
Hermitian matrix. This function converts the eigenvalue problem for a Hermi-
tian matrix into that for a real symmetric matrix of order 2N and solves this
expanded problem. The expanded eigenvalue problem is solved by reducing
the matrix to tridiagonal form using TRED2 and then solving the eigenvalue
problem for tridiagonal matrix using TQL2. AR and AI are arrays of length
IA × N, containing the real and imaginary parts of the matrix. The function
preserves the contents of these arrays. N is the order of the matrix. IA is the
second dimension of AR, AI as specified in the calling function. EI is an array
of length N, which will contain the eigenvalues of the matrix. The eigenvalues
are sorted in the ascending order. VR and VI are arrays of length IZ × N, which
will contain the real and imaginary parts of eigenvectors of the complex matrix.
VR[I][J] and VI[I][J] will contain the Ith component of Jth eigenvector. IZ is
the second dimension of arrays VR and VI as declared in the calling function.
REPS is the tolerance which should be of the order of ~, the machine accuracy.
Error status can be obtained from the returned value of HEREVP. A value of
111 implies that N ≤ 1 or N > IA or N > IZ, in which case, no calculations
are performed. Other values may be set by TRED2 or TQL2. The function
requires TRED2 and TQL2 to solve the eigenvalue problem for equivalent real
symmetric matrix. Since every eigenvalue of the equivalent real matrix is re-
peated, there may be problem in isolating the eigenvectors when the eigenvalue
of complex matrix is multiple. If there is some problem all eigenvectors of the
2N × 2N real matrix should be preserved so that the right combinations can be
identified. This function picks alternate eigenvectors.
1112 Appendix C. C Programs

185. BALANC Function for reducing the norm of a matrix by exact diagonal
similarity transformations. This function is based on the procedure balance in
the Handbook. A is an array of length IA × N containing the matrix. After
execution, the balanced matrix will be overwritten on the same array. N is the
order of the matrix and IA is the second dimension of array A, as declared in the
calling function. B is the base of floating-point representation in the machine.
For most machines B = 2. LOW and IGH are integer variables, such that in the
balanced matrix aij = 0 if i > j and (j < LOW or i > IGH). This essentially
means that after balancing we have to only consider the sub-matrix in rows
and columns from LOW to IGH, since other eigenvalues are already isolated by
function BALANC. D is an array of length N containing the information about
transformation. The elements D[LOW] to D[IGH] will contain the elements
of diagonal matrix used for balancing, while other elements will contain the
permutations used to isolate eigenvalues. Error status can be obtained from
the returned value of BALANC. A value of 112 implies that N ≤ 1 or N > IA,
in which case, no calculations are performed.
186. BALBAK Function to perform back-transformation of a set of right
eigenvectors from those of the balanced matrix to that for the original matrix.
This function is based on the procedure balbak in the Handbook. This function
only handles real eigenvectors, but since the transformation is real and linear,
the real and imaginary parts of complex eigenvectors can be transformed by
two separate calls to this function. N is the order of the matrix. LOW and
IGH are integer variables obtained by the function BALANC. CZ is an array
of length IZ × M containing the eigenvectors of the balanced matrix. After
execution, the eigenvectors of the original matrix will be overwritten on the
same array. For complex eigenvectors two calls will be needed for transforming
the real and imaginary parts separately. In that case CZ should contain the
real (or imaginary) part of the eigenvectors. M is the number of eigenvectors,
and IZ is the second dimension of array CZ as declared in the calling function.
D is an array of length N, containing the information about transformation
as generated by the function BALANC. The returned value of the function is
always zero. This function can be modified as explained in the file, to back-
transform left eigenvectors and the corresponding version is implemented in
BALBAK L.
187. ELMHES Function for reducing a general real matrix to an upper Hes-
senberg form using real stabilised elementary similarity transformation. This
function is based on the procedure elmhes in the Handbook. A is an array of
length IA × N containing the matrix. After execution, the reduced matrix will
be overwritten on the same array. Information about the transformation (i.e.,
the elements mi,r+1 ) will also be written on the (i, r) element of the same
array. N is the order of the matrix and IA is the second dimension of array
A as declared in the calling function. LOW and IGH are integer variables as
given by the function BALANC while balancing the matrix. This function only
processes the sub-matrix in rows LOW to IGH. If the matrix is not balanced,
C.12. Ordinary Differential Equations 1113

then set LOW = 0 and IGH = N − 1 before calling this function. INC is an
integer array of length N, containing the information about the row and column
interchanges used during the reduction. Error status can be obtained from the
returned value of ELMHES. A value of 113 implies that N ≤ 1 or N > IA, in
which case, no calculations are performed.
188. HQR Function to find eigenvalues of an upper Hessenberg matrix using
QR algorithm. This function is based on the procedure hqr in the Handbook.
It does not keep track of the transformations and hence cannot be used to find
eigenvectors. H is an array of length IH × NN containing the matrix. During
execution, the contents of H will be destroyed. NN is the order of the matrix
and IH is the second dimension of H as specified in the calling function. ER
and EI are arrays of length NN, which will contain the real and imaginary
parts of the eigenvalues. REPS is the tolerance which should be of the order
of ~, the machine accuracy. Error status can be obtained from the returned
value of HQR. A value of 114 implies that N ≤ 1 or N > IH, in which case, no
calculations are performed. HQR = 145 implies that the QR iteration failed to
converge at some stage and the calculations are abandoned. In this case, the
eigenvalues which are already isolated should be available in arrays ER and EI.

C.12 Ordinary Differential Equations

189. RKM Function to solve initial value problems in ordinary differential
equations, using a second or fourth-order Runge-Kutta method with adaptive
step size control. This function accepts a system of first-order differential equa-
tions. N is the number of first-order equations. Y is an array of length N, which
should contain the initial values of the variables at input. After successful execu-
tion, it will contain the solution at the required point. If the execution is aborted
in between, then it will contain the solution at some intermediate point T0, up
to where the integration was successful. DY is an array of length N, containing
the first derivative of the solution. The derivative need not be supplied at the
time of calling. DIF is the name of the external function, which is invoked for
calculating the right-hand side of the differential equation y0 = f (t, y). H is the
initial step size to be tried. The function will adjust the step size if necessary,
using the technique described in Section 12.4. After execution, H will contain
the current value of the step size. T0 is the initial value of the independent vari-
able t, at which the initial conditions are specified. After execution, the value
of T0 will be updated, to the point up to which the integration is successful.
TN is the value of t at which the solution is required. TN need not be greater
than T0, i.e., the solution can be calculated in forward or backward direction.
If the integration is successful, then T0 will be set equal to TN. Thus, if the
solution is required at several points, then the function can be called repeatedly
after changing TN between every call. Other variables need not be changed.
REPS is the required relative accuracy in each component of the solution. This
function only tries to estimate the local truncation error and the computed
1114 Appendix C. C Programs

solution may not be correct to the specified accuracy. NSTEP is the number of
steps used by the Runge-Kutta method, each step may require 10 or 11 func-
tion evaluations with fourth-order Runge-Kutta method. Each step may not be
successful. NMAX is the maximum number of steps that the function is allowed
to use. If NMAX ≤ 0, then a default value of NMX (= 10000) will be used.
Error status can be obtained from the returned value of RKM. A value of 701
implies that N ≤ 0, in which case, no calculations are performed. RKM = 721
implies that step size has become smaller than REPS|TN − T0|. If in a small
region the step size is actually expected to be smaller than this limit, then the
corresponding statement in the function may be modified. RKM = 722 implies
that the step size is too small for the arithmetic used (i.e., H < ~|T0|) and the
calculations are aborted. RKM = 723 implies that the integration could not be
completed in the specified number (NMAX) of steps. The failure of this func-
tion can arise either because of singularity, or stiffness, or because the specified
accuracy is too high. This function requires function RK4 or RK2 to perform
one step of Runge-Kutta integration. Function RK4 uses a fourth-order Runge-
Kutta method, while RK2 uses a second-order method. To use RK2 the call
statements as well as the parameters in the beginning of the function should be
changed as indicated. RKM 2 is the version of RKM for second-order method.
Further, the function DIF(T, N, Y, DY) must be supplied by the user to cal-
culate the right-hand sides of the equations, i.e., the derivatives Y0 . Here T is
the value of independent variable and N is the number of equations. Y is the
array containing the dependent variables at time t = T, where the derivatives
need to be evaluated. DY is the array which will contain the calculated value
of the derivatives. DY[i] should contain Y0 [i] = fi (T, Y).

190. RK4 Function to perform one step of integration using a fourth-order

Runge-Kutta method. N is the number of first-order differential equations in
the system. T is the value of independent variable at the initial point. This value
is not updated by the function. H is the step size to be used. Y0, DY0 and Y1
are arrays of length N. Y0 contains the initial value, while after execution, Y1
will contain the final value of the solution vector at t = T + H. DY0 contains
the first derivative at Y0. Y0 and DY0 must be supplied at the time of calling
the function. DIF is the name of the external function used to calculate the
right-hand sides of the differential equations. The function DIF as described in
the write up for subroutine RKM must be supplied by the user. This function
is called by function RKM, but it can be used independently to integrate an
equation with constant step size. This can be achieved by repeatedly calling
the function after updating the value of T (= T + H), Y0 (= Y1) and DY0
(using DIF).

191. RK2 Function to perform one step of integration using a second-order

Runge-Kutta method. All arguments have the same meaning as that for the
function RK4.
C.12. Ordinary Differential Equations 1115

192. MSTEP Function to solve initial value problems in ordinary differential

equations, using a fourth-order multistep method with adaptive step size con-
trol. This function accepts a system of first-order differential equations. It can
be used with either function ADAMS for Adams-Bashforth-Moulton predictor-
corrector method, or function GEAR for stiffly stable fourth-order method.
The change can be affected by choice of a flag. By setting the flag appropri-
ately, this function can also be used to integrate the equation using a fixed step
size. This function is a crude implementation of multistep methods with simple
techniques for adjusting the step-size which should work on simple problems. A
more sophisticated implementation will be required to handle a larger fraction
of equations efficiently. N is the number of first-order equations. Y is an array
of length 7N, the first N elements of which should contain the initial values
of the solution at input. After execution, it will contain the solution at some
intermediate points. The contents of this array must be preserved, if a second
call to the function is required for continuing the integration further. DY is an
array of length 7N, containing the first derivatives of the solution. The first
derivative need not be supplied at the time of calling. The arrays Y and DY
are used to store the solution at seven most recent points. The first index of the
array is increased in a circular manner, to avoid frequent copying of the array
elements. DIF is the name of the external function, which is invoked for cal-
culating the right-hand sides of the differential equations y0 = f (t, y). H is the
initial step size to be tried. This function generates the starting values using the
fourth-order Runge-Kutta method, which will adjust the step size if necessary.
After the starting values are generated, the step size can be doubled or halved
as explained in Section 12.3. After execution, H will contain the current value
of the step size. T0 is the initial value of the independent variable t, at which
the initial conditions are specified. After execution, the value of T0 will be up-
dated to the point up to which the integration is successful. If the execution is
successful the final value of T0 may be beyond the requested value of TN. The
solution at TN is computed using interpolation between values of solution at
different values of t. TN is the value of t at which the solution is required. If the
solution is required at several points, then the function can be called repeatedly
after changing TN between every call. Other variables (including the scratch
arrays) should not be changed, because the function uses the previous values
of the solution as the starting values for the new problem. YF is an array of
length N, which will contain the final value of the solution at the required point
TN, provided the integration is successful. If the integration is aborted at some
intermediate point after the starting values were calculated, then this array will
contain the solution at the last successful point. REPS is the required relative
accuracy in each component of the solution. This function only tries to estimate
the local truncation error. Hence, the computed solution may not be correct
to the specified accuracy. NSTP is the number of evaluations of the right-hand
side required by the function. NSTP is initialised only during the first call
(IFLG = 0 or 1) and hence it will accumulate the number of calls to DIF from
the first call, until IFLG is reset to 0 or 1. NMAX is the maximum number of
1116 Appendix C. C Programs

function evaluations that the function is allowed to use. If NMAX ≤ 0, then

a default value of NMX (= 100000) will be used. IFLG is an integer variable
used as a flag by the function. If IFLG = 0 or 1, then the integration is started
by generating fresh starting values using function STRT4. For higher values
of IFLG it is assumed that the starting values are already available in the ar-
rays Y and DY. The function itself resets the value of IFLG after generating
the starting values. Hence, this parameter need not be reset between two calls,
unless it is required to generate starting values again. If IFLG = 0 or 2, then
the step size is adjusted according to the required accuracy. For IFLG = 1 and
3, the step size is kept fixed and no attempt is made to check the truncation
error. However, the parameter REPS is used to check for convergence of the
corrector iteration. In this case, if the iteration on corrector fails to converge,
then error flag will be set to 729. IST is an integer parameter to decide which
multistep method is to be used. If IST = 0 the fourth-order Adams-Bashforth-
Moulton predictor-corrector method is used, while for other values of IST, the
fourth-order stiffly stable method due to Gear is used. WK is an array, which is
used as a scratch space by the function. The length of this array should be 2N
for Adams method and 3N + N2 for stiffly stable method. Error status can be
obtained from the returned value of MSTEP. A value of 702 implies that N ≤ 0,
in which case, no calculations are performed. MSTEP = 724 or 725 implies that
function STRT4 failed to generate the required starting values. MSTEP = 726
implies that step size has become smaller than REPS|TN − T0|. If in a small
region the step size is actually expected to be smaller than this limit, then the
corresponding statement in the function may be modified. MSTEP = 727 im-
plies that the step size is too small for the arithmetic used (i.e., H < ~|T0|) and
the calculations are aborted. MSTEP = 728 implies that the integration could
not be completed in the specified number (NMAX) of function evaluations. The
failure of this function can arise either because of singularity, or stiffness (with
ADAMS), or because the specified accuracy is too high. For integration with
fixed step size, error flag may be set to 729 if the corrector fails to converge.
This function requires functions ADAMS, GEAR and GAUELM to perform
the integration and function STRT4 and RK4 to generate the starting values.
In addition the function DIF(T, N, Y, DY) must be supplied by the user to
calculate the right-hand sides of the equations, i.e., the derivatives Y0 . Here T
is the value of independent variable and N is the number of equations. Y is the
array containing the dependent variables at time t = T, where the derivatives
need to be evaluated. DY is the array which will contain the calculated value
of the derivatives. DY[i] should contain Y0 [i] = fi (T, Y).
193. ADAMS Function to perform one step of solution of an initial value
problem, using a fourth-order Adams-Bashforth-Moulton predictor-corrector
method. This function can be easily modified to use any other predictor-
corrector formula. This function is called by function MSTEP to perform one
step of integration. N is the number of first-order differential equations, Y and
DY are arrays of length 7N, as used by function MSTEP. DIF is the name of
the function used to calculate the right-hand sides of the equations. H is the
C.12. Ordinary Differential Equations 1117

step size and T is the value of independent variable at which the solution is
required. REPS is the specified accuracy. This function does not check for the
truncation error, but the parameter REPS is used to check for the convergence
of corrector iteration. NSTP is the number of function evaluations used so far.
This number is updated by the function. IJ, IJM1, IJM2, IJM3 and IJM4 are
respectively, the indices j + 1, j, j − 1, j − 2 and j − 3 in the predictor-corrector
formula. WK is an array of length 2N, which is used to transmit the predicted
value to function MSTEP. Error status can be obtained from the returned value
of ADAMS. A value of 729 implies that the iteration on corrector has failed to
converge. Function DIF as described in the writeup for function MSTEP must
be supplied by the user.
194. STRT4 Function to generate the starting values y1 , y2 and y3 using a
fourth-order Runge-Kutta method. The truncation error is estimated by per-
forming one integration with double step size. This function is called by the
function MSTEP. N is the number of first-order differential equations to be
solved. Y and DY are arrays of length 4N, containing the values of y and y0 at
the four starting points. While calling the function, the initial values y0 must
be stored in the first N elements of Y. Other elements of these arrays will be
calculated by the function. DIF is the name of the function used to calculate the
right-hand sides of the differential equations. H is the initial guess for the step
size. If necessary the step size will be adjusted by the function. T is the value
of the independent variable at the initial point. If the execution is successful,
then T is updated to T + 3H. REPS is the specified relative accuracy. IFLG is
an integer parameter used as a flag. If IFLG = 0, then the step size may be
adjusted while for other values of IFLG the step size will not be adjusted. For
IFLG 6= 0, the value of REPS is redundant, since no attempt is made to check
the truncation error. TSTEP is the size of interval over which the integration
is requested. It is used only for convergence check. NSTP is an integer variable
which contains the number of function evaluations used so far. This value is
updated by the function. WK is an array of length 2N used to pass on the value
for modifier. Error status can be obtained from the returned value of STRT4. A
value of 724 implies that the function failed to find the starting values, because
the step size became too small. STRT4 = 725 implies that the routine failed to
find the starting values in the specified number (NIT = 10) of attempts. This
function requires functions RK4 and DIF. Function DIF as described in the
writeup for function MSTEP must be supplied by the user.
195. GEAR Function to perform one step of solution of initial value problem
using a fourth-order stiffly stable method based on the backward differentiation
formula. This function uses the Broyden’s method to solve the implicit corrector
formula. This function is called by the function MSTEP to perform one step
of integration. N is the number of first-order differential equations, Y and DY
are arrays of length 7N as used by function MSTEP. DIF is the name of the
function used to calculate the right-hand sides of the equations. H is the step
size and T is the value of the independent variable at which the solution is
1118 Appendix C. C Programs

required. REPS is the specified accuracy. This function does not check for the
truncation error, but the parameter REPS is used to check for the convergence
of corrector iteration. NSTP is the number of function evaluations used so far.
This number is updated by the function. IJ, IJM1, IJM2, IJM3 and IJM4 are
respectively, the indices j + 1, j, j − 1, j − 2 and j − 3 in the corrector formula.
IFLAG is an integer variable used as a flag. If IFLAG = 0, initial approximation
to the Jacobian is generated, otherwise the old approximation is used. In any
case, Broyden’s method will be used to update the initial approximation to
the inverse of the Jacobian. If IFLAG = 0, it is set to 1 after calculation of
Jacobian so that the Jacobian is not calculated again next time. WK1 is an
array of length 3N used to pass predicted values. WK is an array of length N2
used to store the inverse of the Jacobian. Error status can be obtained from
the returned value of GEAR. A value of 729 implies that the corrector iteration
fails to converge. Function DIF as described in write-up for MSTEP must be
supplied by the user. This function also requires function GAUELM to solve
systems of linear equations. The parameter CFAC in the function may need to
be reduced if the function MSTEP needs to adjust the step size too often.
196. EXTP Function to solve initial value problems in ordinary differential
equations using extrapolation method. This function accepts a system of first-
order differential equations. N is the number of first-order equations. Y is an
array of length N, which should contain the initial values of solution at input.
After successful execution, it will contain the solution at the required point.
If the execution is aborted in between, then it will contain the solution at
some intermediate point T0, up to where the integration was successful. DY is
an array of length N containing the first derivative of the solution. The first
derivative need not be supplied at the time of calling. DIF is the name of
the external function, which is invoked for calculating the right-hand sides of
the differential equations y0 = f (t, y). H is the initial step size to be tried.
The function will adjust the step size if necessary, using the crude technique
described in Section 12.5. After execution, H will contain the current value of
the step size. T0 is the initial value of the independent variable t, at which the
initial conditions are specified. After execution, the value of T0 will be updated
to the point, up to which the integration is successful. TN is the value of t
at which the solution is required. If the integration is successful, then T0 will
be set equal to TN. Thus, if the solution is required at several points, then
the function can be called repeatedly after changing TN between every call.
Other variables need not be changed. REPS is the required relative accuracy
in each component of the solution. This function only tries to estimate the
local truncation error. Hence, the computed solution may not be correct to
the specified accuracy. NSTEP is the number of function evaluations used by
the function. This number is initialised to zero at every call to EXTP. Thus
if EXTP is called repeatedly, then the total number of function evaluations
used will need to be accumulated separately. NMAX is the maximum number
of function evaluations that the function is allowed to use. If NMAX ≤ 0, then
a default value of NMX (= 100000) will be used. IFLG is an integer variable
C.12. Ordinary Differential Equations 1119

used to decide which type of extrapolation is to be used. If IFLG = 0, then

polynomial extrapolation is used, otherwise rational function extrapolation is
used. Error status can be obtained from the returned value of EXTP. A value of
703 implies that N ≤ 0, in which case, no calculations are performed. EXTP =
730 implies that step size has become smaller than REPS|TN−T0|. If in a small
region the step size is actually expected to be smaller than this limit, then the
corresponding statement in the function may be modified. EXTP = 731 implies
that the step size is too small for the arithmetic used (i.e., H < ~|T0|) and the
calculations are aborted. EXTP = 732 implies that the integration could not be
completed in the specified number (NMAX) of steps. EXTP = 733 implies that
the denominator in the rational function extrapolation is zero. The failure of
this function can arise either because of singularity, or stiffness, or because the
specified accuracy is too high. For rational function extrapolation, the failure
can also be due to the fact that denominator has become zero. In this case,
polynomial extrapolation can be tried. The function DIF(T, N, Y, DY) must be
supplied by the user to calculate the right-hand sides of the equations, i.e., the
derivatives Y0 . Here T is the value of independent variable and N is the number
of equations. Y is the array containing the dependent variables at time t = T,
where the derivatives need to be evaluated. DY is the array which will contain
the calculated value of the derivatives. DY[i] should contain Y0 [i] = fi (T, Y).
197. FDM Function to solve a two-point boundary value problem with sepa-
rable boundary conditions, using finite difference method as explained in Sec-
tion 12.8. If a uniform mesh is used, then it is also possible to estimate the
first-order correction, using the method of deferred correction. N is the number
of mesh points to be used. M is the number of first-order differential equations
in the system. ML is the number of boundary conditions at the first boundary
(i.e., T[0]). PAR is an array which is passed on to the function EQN and BCS
for calculating the required information about the equation and the boundary
conditions. This array can be used to pass any parameters that may be required
by these functions. This array is not used by function FDM, or any other func-
tion called by it except EQN and BCS, and its size can be arbitrary depending
on the requirements. X is an array of length M × N, which should contain the
initial guess for the solution. After execution, it will contain the computed so-
lution. If the problem is linear, then there is no need for initial approximation
and this array could be initialised to zero or any other convenient value. The
function FDM treats it as a two-dimensional array of dimension X[N][M] with
X[J][I] containing the Ith component at Jth mesh point. Unfortunately, there
is no provision to pass the second dimension of the array X, as specified in the
calling function. Hence, to obtain meaningful results, the second dimension of
the array X in the calling function must also be exactly equal to M. This prob-
lem arises because, while solving the finite difference equations, it is convenient
to assume that the array elements are stored in consecutive positions without
any gaps. XC is also an array of length M × N, which will contain the solution
after applying the deferred correction. This array is also stored in the same
format as the array X and the same remarks apply here too. Even if deferred
1120 Appendix C. C Programs

correction is not requested, this array should be provided, since it is also used
as a scratch space during the computations. The difference XC[J][I] − X[J][I]
gives an estimate of the truncation error in X[J][I] provided the deferred cor-
rection is calculated. T is an array of length N, containing the mesh points tj ,
(j = 0, . . . , N − 1), to be used in the finite difference approximation. This array
must be supplied by the user. If deferred correction is required, then the mesh
must be uniformly spaced, otherwise spacing can be arbitrary. In any case, the
mesh points tj must be in either ascending or descending order. EQN is the
name of the function, which defines the differential equations. The differential
equation is assumed to be in the form By0 = f (t, y), where B is a M × M ma-
trix and f is a vector function of length M. The matrix B must be nonsingular.
The elements of B and f at a given point T are calculated by the function
EQN. Apart from this, it must also calculate the Jacobian ∂fi /∂yj (stored in
the array A). BCS is the name of the function, which specifies the boundary
conditions. The boundary conditions are defined by functions gi (t0 , y0 ) = 0
for i = 0, . . . , ML − 1 at the first point t = T[0], and gi (tN −1 , yN−1 ) = 0 for
i = ML, . . . , M − 1 at the last point t = T[N − 1]. Apart from the functions gi ,
this function should also calculate the Jacobian matrix ∂gi /∂yj (stored in the
array BC). The general form of these functions is as follows:
void eqn(int j, int m, int ml, double par[], double *a,
double *b, double y[], double f[], double t)
{
int i,k,m1;
m1=m+ml;
for(i=0; i<m; ++i) {
f[i]=fi (t, par, y);
for(k=0; k<m; ++k) {
a[k+i*m1]=∂fk /∂y[i];
b[k+i*m1]=bki (t, par);
}
}
return;
}
void bcs(int m, int ml, double par[], double *bc, double g[],
double t1, double tn, double y1[], double yn[])
{
int i,k,m1;
m1=m+ml;
for(i=0; i<m; ++i) {
if(i<ml) g[i]=gi (t1, par, y1);
else g[i]=gi (tn, par, yn);
for(k=0; k<m; ++k) bc[i+k*m1]=∂gi /∂y[k];
}
C.12. Ordinary Differential Equations 1121

return;
}
It should be noted that the second dimension of A, B and BC is M+ML instead
of M. Here T, T1 (= T[0]) and TN (= T[N − 1]) are the values of t at relevant
points, while the arrays Y, Y1 and YN contain the values of y at these points.
In function EQN the parameter J identifies the interval at which the functions
are evaluated (i.e., T = (tJ + tJ+1 )/2). This index is useful in cases, where the
coefficients are read in the form of a table.
IFLAG is an integer variable which is used as a flag. If IFLAG = 0 or
1, then the equation is treated as nonlinear and solution will be calculated
iteratively. If IFLAG = 2 or 3, then the equations are treated as linear and
calculation will be terminated after the first iteration. If IFLAG = 0 or 2,
then the first-order correction will be calculated using the method of deferred
correction. For other values of IFLAG, the deferred correction will not be cal-
culated. The deferred correction can be calculated only if the mesh spacing
is uniform and the number of points N is at least five. REPS is the specified
accuracy. This parameter is only used to check for the convergence of Newton’s
method, for solution of finite difference equations (only for nonlinear equations
i.e., IFLAG = 0, 1). This parameter has no effect on truncation error, which
is determined by the mesh spacing. The truncation error can be estimated by
the difference between X and XC, if deferred correction is used. Error status
can be obtained from the returned value of FDM. A value of 704 implies that
N < 3, or M ≤ ML, or ML ≤ 0, in which case, no calculations are performed.
FDM = 734 implies that N ≤ 4 and deferred correction is requested. In this
case, the deferred correction is not calculated. FDM = 735 implies that the
finite difference matrix is singular and the solution cannot be obtained. This
problem may be due to some error in specifying the equation matrix or bound-
ary conditions. FDM = 736 implies that the mesh spacing is not uniform, in
which case, the deferred correction will not be calculated. FDM = 737 implies
that the Newton’s iteration for solving the finite difference equations fails to
converge. This function requires function SETMAT for setting up the matrix of
finite difference equations, and function GAUBLK to solve the system of linear
equations defined by the block matrix as explained in Section 12.8. Apart from
these, the functions EQN and BCS specifying the problem must be supplied by
the user.
198. GEVP Function to solve a generalised eigenvalue problem for a system
of differential equations, with separable boundary conditions, using finite dif-
ference method as explained in Section 12.9. If a uniform mesh is used, then
it is also possible to estimate the first-order correction to the eigenvalue, us-
ing the method of deferred correction. N is the number of mesh points to be
used. M is the number of first-order differential equations in the system. ML
is the number of boundary conditions at the first boundary (i.e., T[0]). PAR
is an array which is passed on to the functions EQN, EQND, BCS and BCSD
for calculating the required information about the equation and the boundary
conditions. The first element of PAR is used to pass the eigenvalue and hence
1122 Appendix C. C Programs

it should not be used for any other purpose. After execution, PAR[0] will con-
tain the calculated eigenvalue without applying the deferred correction. Other
elements of this array can be used to pass any parameter that may be required
by these functions. Other elements of the array are not used by function GEVP
or any other function called by it except EQN, BCS, EQND and BCSD, and
its size can be arbitrary depending on the requirements. X is an array of length
M × N, which will contain the computed eigenfunction. Even if the eigenfunc-
tion is not required, this array should be provided. The function GEVP treats
it as a two-dimensional array of dimension X[N][M], with X[J][I] containing the
Ith component at Jth mesh point. Unfortunately, there is no provision to pass
the second dimension of the array X as specified in the calling function. Hence,
to obtain meaningful results, the second dimension of the array X in the calling
function must also be exactly equal to M. This problem arises, because while
solving the finite difference equations it is convenient to assume that the array
elements are stored in consecutive positions without any gaps. XC is also an
array of length M × N, which will contain the left eigenvector of the finite dif-
ference matrix, which is required for calculating the deferred correction. This
array is also stored in the same format as the array X and the same remarks
apply here too. Even if deferred correction is not requested, this array should
be provided, since it is also used as a scratch space during the computations.
T is an array of length N, containing the mesh points tj for j = 0, . . . , N − 1 to
be used in the finite difference approximation. This array must be supplied by
the user. If deferred correction is required, then the mesh must be uniformly
spaced, otherwise spacing can be arbitrary. In any case, the mesh points tj
must be in either ascending or descending order. E0 is the initial guess for the
eigenvalue. After execution, E0 will contain the corrected eigenvalue, provided
deferred correction is calculated.
EQN is the name of the function which defines the differential equation.
The differential equation is assumed to be in the form By0 = Ay, where A
and B are M × M matrices. The matrix B must be nonsingular. The elements
of A and B which could be arbitrary functions of the eigenvalue λ and t, are
calculated by the function EQN. BCS is the name of the function which specifies
the boundary conditions. The boundary conditions are defined by C1 y0 = 0,
and CN yN−1 = 0, where C1 and CN are ML × M and (M − ML) × M matrices,
respectively. These matrices are combined in one M×M matrix BC, the first ML
rows of which correspond to the boundary conditions at the first boundary (t =
T[0]), while the remaining rows refer to boundary conditions at the last point
(t = T[N − 1]). EQND is the name of the function to calculate the derivatives
of matrices A and B (as defined in EQN) with respect to the eigenvalue λ.
Similarly, BCSD is the name of the function to calculate the derivatives of
matrix BC (as defined in BCS) with respect to the eigenvalue λ. Following the
C convention the indices of arrays A, B and BC are interchanged as compared
to the Fortran subprogram, thus the rows are really columns. The order of index
should be clear from the form of functions given below. If deferred correction is
not required, then the functions EQND and BCSD are not called, but in order
C.12. Ordinary Differential Equations 1123

to satisfy the compiler and the linker, dummy functions may be required. The
general form of these functions is as follows:
void eqn(int j, int m, int ml, double par[], double *a,
double *b, double y[], double f[], double t)
{
int i,k,m1;
m1=m+ml;
for(i=0; i<m; ++i) {
f[i]=0.0;
for(k=0; k<m; ++k) {
a[k+i*m1]=aki (t, par);
b[k+i*m1]=bki (t, par);
}
}
return;
}
void bcs(int m, int ml, double par[], double *bc, double g[],
double t1, double tn, double y1[], double yn[])
{
int i,k,m1;
m1=m+ml;
for(i=0; i<m; ++i) {
g[i]=0.0;
for(k=0; k<m; ++k) {
if(i<ml) bc[i+k*m1]=bcik (t1, par);
else bc[i+k*m1]=bcik (tn, par);
}
}
return;
}
void eqnd(int j, int m, int ml, double par[], double *a,
double *b, double t)
{
int i,k,m1;
m1=m+ml;
for(i=0; i<m; ++i) {
for(k=0; k<m; ++k) {
a[k+i*m1]=∂aki (t, par)/∂par[0];
b[k+i*m1]=∂bki (t, par)/∂par[0];
}
}
return;
1124 Appendix C. C Programs

}
void bcsd(int m, int ml, double par[], double *bc,
double t1, double tn)
{
int i,k,m1;
m1=m+ml;
for(i=0; i<m; ++i) {
for(k=0; k<m; ++k) {
if(i<ml) bc[i+k*m1]=∂bcik (t1, par)/∂par[0];
else bc[i+k*m1]=∂bcik (tn, par)/∂par[0];
}
}
return;
}
It should be noted that the second dimension of A, B and BC is M+ML instead
of M. Here T, T1 (= T[0]) and TN (= T[N − 1]) are the values of t at relevant
points. The arrays Y, Y1, YN, F and G are not really required for eigenvalue
problem, but are provided for consistency. Thus, in EQN and BCS, F[I] and
G[I] should be set to zero in order to be consistent with function FDM, since the
same function SETMAT is used in both cases. In functions EQN and EQND,
the parameter J identifies the interval at which the functions are evaluated (i.e.,
T = (tJ + tJ+1 )/2). This index is useful in cases, where the coefficients are read
in the form of a table.
IFLAG is an integer variable which is used as a flag. If IFLAG = 0, then
only the eigenvalue is calculated. If IFLAG = 1, then eigenvector is also calcu-
lated, using the method of inverse iteration. If IFLAG = 2, then the first-order
correction to the eigenvalue is also calculated, using the method of deferred
correction. The deferred correction can be calculated only if the mesh spac-
ing is uniform and the number of mesh points N is at least five. REPS is the
specified accuracy. This parameter is only used to check for the convergence of
Muller’s method (or secant method) for finding zeros of the determinant, and
the inverse iteration method for finding eigenvectors. This parameter has no
effect on the truncation error, which is determined by the mesh spacing. If de-
ferred correction is used, then the difference E0 − PAR[0] will give an estimate
of the truncation error in PAR[0]. EL and EU respectively, specify the lower
and upper limit on eigenvalue. These parameters are used to specify the region
in which the eigenvalues are to be searched. These parameters are passed on
to function SECANI, (or MULER2) which terminates the iteration if at any
stage it goes outside the specified limits. Error status can be obtained from the
returned value of GEVP. A value of 704 implies that N < 3, or M ≤ ML or
ML ≤ 0. In this case, no calculations are performed. GEVP = 734 implies that
N ≤ 4 and deferred correction is requested. In this case, the deferred correction
is not calculated. GEVP = 735 implies that one of the pivots has vanished dur-
ing the calculation and hence the eigenvector cannot be calculated. In this case,
C.12. Ordinary Differential Equations 1125

the eigenvalue will be calculated. This problem can be avoided by perturbing

the eigenvalue slightly and recalculating the determinant before calculating the
eigenvector. GEVP = 736 implies that the mesh spacing is not uniform, in
which case the deferred correction will not be calculated. GEVP = 738 implies
that the eigenvector vanishes, in which case the calculations are terminated. In
this case also the eigenvalue is already calculated and will be stored in PAR[0].
GEVP = 739 implies that the inverse iteration for calculating the eigenvector
failed to converge. GEVP = 740 implies that the inverse iteration for calculat-
ing the left eigenvector failed to converge. Apart from these, other values may
be set by function SECANI (or MULER2), which is called to find zeros of the
determinant. The function GEVP requires function SETMAT for setting up the
matrix of finite difference equations, and function GAUBLK to solve the sys-
tem of linear equations, or to calculate the determinant of the block matrix as
explained in Section 12.8. Function MULER2, SECANI or any other equivalent
function is required for finding zeros of the determinant. It may be noted that
function MULER2 requires the function to be complex, but if we want to avoid
unnecessary use of complex arithmetic for calculating real eigenvalues, then we
can take the real part of CX and calculate the real function DET and pass it
back to MULER2 in the complex variable CF. In principle, such a procedure
can lead to trouble, if the iterates become complex at some stage and the next
iterate has the same real part. This problem can be avoided if SECANI is used.
Both SECANI and MULER2 use reverse communication technique and hence
the control is passed back to GEVP when a function evaluation is required.
Apart from these, the functions EQN, EQND, BCS and BCSD specifying the
problem must be supplied by the user.
199. GAUBLK Function to solve a system of linear equations involving finite
difference matrix of the form described in Section 12.8. This function uses
Gaussian elimination with partial pivoting. N is the number of mesh points,
M is the number of first-order differential equations in the system. ML is the
number of boundary conditions at the first boundary. A is an array of length
(M + ML) × 2M × N containing the matrix (Section 12.8). IFLG is an integer
variable used as a flag. If IFLG = 0, the function performs elimination as well
as the solution of linear equations. If IFLAG = 1, then only the elimination
is performed. In both these cases, the determinant is calculated and IFLG is
set to 2. If IFLG = 2, then it is assumed that elimination is already performed
and the triangular factors are stored in the same array A. In this case, only
the linear equations are solved. DET is a variable specifying the scaled value
of the determinant, IDET is an integer variable which contains the exponent
part of the determinant. The actual value of the determinant is DET × 2IDET .
INC is an integer array of length M × N, which contains information about
row interchanges used during Gaussian elimination. X is an array of length
M × N, which should contain the right-hand sides of the linear equations. After
execution, the solution will be returned in the same array. Error status can
be obtained from the returned value of GAUBLK. A value of 735 implies that
one of the pivots vanishes during elimination and hence the equations cannot
1126 Appendix C. C Programs

be solved. In this case, the determinant will vanish, but that does not cause
any problem for finding eigenvalues, which are obtained by finding zeros of the
determinant. This function is called by functions FDM and GEVP.

200. SETMAT Function to setup the finite difference matrix for a system
of first-order differential equations with separable boundary conditions, using
a central difference approximation. This function is called by functions FDM
and GEVP. N is the number of mesh points, M is the number of first-order
differential equations in the system, ML is the number of boundary conditions
at the first boundary. A is an array of length (M + ML) × 2M × N, which will
contain the finite difference matrix for the linearised equations. BC is an array
of length (M + ML) × (M + 1), which is used to store information about the
boundary conditions. X and XC are arrays of length M×N. X should contain the
approximation to solution of the boundary value problem, while after execution,
XC will contain the right-hand sides of the finite difference equations. X[J][I]
should contain the Ith component at Jth mesh point. T is an array of length
N containing the mesh points. PAR is an array containing parameters, which
are passed on to the functions EQN and BCS. EQN and BCS are the names
of functions for defining the differential equation and the boundary conditions,
respectively. These functions must be supplied by the user, as explained in the
write up for functions FDM or GEVP. The returned value of the function is
always zero.

201. BSPODE Function to solve a two-point boundary value problem with

separable boundary conditions, using expansion method with B-spline basis
functions as explained in Section 12.10. NK is the number of knots to be used.
K is the order of B-splines to be used. K = 4 corresponds to cubic B-splines,
while K = 2 gives linear B-splines, etc. The order may need to be increased to
get higher accuracy. M is the number of first-order differential equations in the
system. ML is the number of boundary conditions at the first boundary (i.e.,
T[0]). PAR is an array which is passed on to the function EQN and BCS for
calculating the required information about the equation and the boundary con-
ditions. This array can be used to pass any parameters that may be required by
these functions. This array is not used by function BSPODE, or any other func-
tion called by it except EQN and BCS, and its size can be arbitrary depending
on the requirements. X is an array of length M × N, which will contain the
computed solution. X[J][I] will contain the Ith component of solution at TX[J].
The second dimension of X must be equal to M in the calling function. A is an
array of length (NK + K − 2) × M containing the coefficients of expansion in
terms of B-splines. At the time of calling it should contain the initial guess to
the coefficients, while after execution it will contain the calculated coefficients.
If the problem is linear, then there is no need for initial approximation and this
array could be initialised to zero or any other convenient value. The function
BSPODE treats it as a two-dimensional array of dimension A[M][NK + K − 2]
with A[J][I] containing the coefficient of Ith basis function in Jth component of
C.12. Ordinary Differential Equations 1127

the solution.
NK+K−3
X
Xj (t) = A[j][i]φi (t), j = 0, . . . , M − 1. (C.89)
i=0

Unfortunately, there is no provision to pass the second dimension of the ar-

ray X or A, as specified in the calling function. Hence, to obtain meaningful
results, the second dimension of the array A in the calling function must also
be exactly equal to NK + K − 2. This problem arises because, while solving
the system of equations, it is convenient to assume that the array elements are
stored in consecutive positions without any gaps. T is an array of length NK,
containing the knots tj , (j = 0, . . . , NK − 1), to be used for calculating the
B-spline basis functions. This array must be supplied by the user. The knots
must be in ascending order with T[0] and T[NK−1] as the two boundaries. N
is the number of mesh points to be used for obtaining the equations connecting
the coefficients of expansion, N ≥ NK + K − 2. TX is an array of length N con-
taining the mesh points to be used for calculating the coefficients. This array
must be supplied by the user if IFLAG > 1, otherwise the function computes
the elements of this array assuming uniform spacing and covering the interval
T[0] to T[NK−1]. The solution will be calculated at all these points. If the so-
lution is required at any other point then function BSPEVL can be used with
the calculated coefficients. EQN is the name of the function, which defines the
differential equations. The differential equation is assumed to be in the form
By0 = f (t, y), where B is a M×M matrix and f is a vector function of length M.
The matrix B must be nonsingular. The elements of B and f at a given point
T are calculated by the function EQN. Apart from this, it must also calculate
the Jacobian ∂fi /∂yj (stored in the array A). BCS is the name of the function,
which specifies the boundary conditions. The boundary conditions are defined
by functions gi (t0 , y0 ) = 0 for i = 0, . . . , ML − 1 at the first point t = T[0], and
gi (tN K−1 , yNK−1 ) = 0 for i = ML, . . . , M − 1 at the last point t = T[NK − 1].
Apart from the functions gi , this function should also calculate the Jacobian
matrix ∂gi /∂yj (stored in the array BC). The general form of these functions
is as follows:
void eqn(int j, int m, int ml, double par[], double *a,
double *b, double y[], double f[], double t)
{
int i,k;
for(i=0; i<m; ++i) {
f[i]=fi (t, par, y);
for(k=0; k<m; ++k) {
a[k+i*m]=∂fk /∂y[i];
b[k+i*m]=bki (t, par)
}
}
return;
1128 Appendix C. C Programs

}
void bcs(int m, int ml, double par[], double *bc, double g[],
double t1, double tn, double y1[], double yn[])
{
int i,k;
for(i=0; i<m; ++i) {
if(i<ml) g[i]=gi (t1, par, y1);
else g[i]=gi (tn, par, yn);
for(k=0; k<m; ++k) bc[i+k*m]=∂gi /∂y[k];
}
return;
}
It should be noted that these functions are same as those required by FDM
except that the second dimension of arrays A, B and BC is M instead of M+ML.
Here T, T1 (= T[0]) and TN (= T[NK − 1]) are the values of t at relevant
points, while the arrays Y, Y1 and YN contain the values of y at these points.
In function EQN the parameter J identifies the interval at which the functions
are evaluated (i.e., TX[j]). This index is useful in cases, where the coefficients
are read in the form of a table.
IFLAG is an integer variable which is used as a flag. If IFLAG = 0 or
2, then the equation is treated as nonlinear and solution will be calculated
iteratively. If IFLAG = 1 or 3, then the equations are treated as linear and
calculation will be terminated after the first iteration. If IFLAG = 0 or 1, then
the array TX need not be initialised at the time of calling as the function will
compute the elements assuming uniform spacing. For other values of IFLAG,
the array TX must be supplied by the user. REPS is the specified accuracy.
This parameter is only used to check for the convergence of Newton’s method,
for solution of system of equations (only for nonlinear equations i.e., IFLAG =
0, 2). This parameter has no effect on truncation error, which is determined by
the knots and the order of B-splines. Error status can be obtained from the
returned value of BSPODE. A value of 705 implies that N < 3, or M ≤ ML,
or ML ≤ 0, or N < NK + K − 2, in which case, no calculations are performed.
BSPODE = 741 implies that the Newton’s iteration for solving the system of
equations fails to converge. This function requires function BSPLIN, BSPEVL,
SVD and SVDEVL. Apart from these, the functions EQN and BCS specifying
the problem must be supplied by the user.

C.13 Integral Equations

202. FRED Function to solve a linear Fredholm equation as defined by
Eq. (13.1) using quadrature methods. M is the number of abscissas to be used in
the quadrature formula. A and B are variables specifying respectively, the lower
and the upper limit of the integral. WT and X are arrays of length M, which will
C.13. Integral Equations 1129

return the weights and abscissas used by the quadrature formula. F and FC are
arrays of length M, which will contain the calculated solution. F[I] will contain
the computed solution at the abscissa X[I], while if the trapezoidal rule is used,
then FC[I] will contain the value obtained after applying deferred correction,
using the Gregory’s formula including second-order differences. Array FC is not
required for other quadrature formulae. Hence, in that case, a dummy array
of length one may be provided. FG and FKER are the names of the functions
provided to calculate the functions g(x) and K(x, t). For Fredholm equation of
the third kind g(x) is not required, but a dummy function giving some nonzero
value (e.g., g(x) = 1) should be provided. In fact, this function supplies the
starting vector for the inverse iteration and if necessary a reasonable guess for
the eigenfunction may be provided via this function. However, in most cases,
the inverse iteration method does converge to the eigenvalue, even when initial
vector is far from the eigenvector and such a guess is not essential, but it can
be used to improve the efficiency. EI gives the initial guess for the eigenvalue of
the Fredholm equation of the third kind. After execution, EI will contain the
calculated eigenvalue. For equations of other kinds, EI will be ignored.
IQ is an integer variable used to specify the quadrature formula to be
used. If IQ = 1, then trapezoidal rule is used, for IQ = 2, Simpson’s 1/3
rule is used, for IQ = 4, 8, 16, 32 a composite rule using 4, 8, 16 or 32-point
Gauss-Legendre formula is used. If IQ is negative, then it is assumed that
weights and abscissas are supplied in the arrays WT and X. Other values of
IQ will cause an error return. For IQ > 1 if the number of points are not
compatible with the composite quadrature formula, then the next lower value
of M is used. For example, if IQ = 4 and M = 11, then computations will
be performed using only 8 points, but the value of M remains unchanged.
In this case, a warning is issued by setting the returned value of FRED to
−11. IT is an integer variable, which specifies the kind of Fredholm equation
to be solved. IT = 1, 2 and 3 corresponds to the Fredholm equations of the
first, second and third kind, respectively. This function can only find a real
eigenvalue. REPS is a parameter specifying the required (relative) accuracy in
calculating the eigenvalues. This parameter is used only if IT = 3. It is only
passed on to the function INVIT for calculating the eigenvalue. Error status
can be obtained from the returned value of FRED. A value of 706 implies that
IT > 3, or IT ≤ 0, or the number of points M are not sufficient for application
of the specified quadrature formula. In this case, no calculations are performed.
FRED = 707 implies that the value of IQ is not proper, in which case also no
calculations are performed. FRED = −11 implies that the number of points
have been adjusted. In this case, the calculations are performed with a smaller
value of M, which is applicable for the quadrature formula requested. Apart
from these, other values may be set by functions GAUELM and INVIT, which
are called by this function. Function FRED requires function GAUELM to solve
the system of linear equations, and function INVIT (for IT = 3) to calculate the
eigenvalue and the eigenvector. It should be noted that function INVIT is called
with IFLG = 0, which keeps the shift P constant, to ensure convergence to the
1130 Appendix C. C Programs

nearest eigenvalue. Consequently, the convergence could be rather slow in some

cases. It may be more efficient to use other values of IFLG, but in that case,
the inverse iteration will not necessarily converge to the nearest eigenvalue.
Functions FG(X) and FKER(X, T) must be supplied by the user.
203. FREDCO Function to solve a linear Fredholm equation of the first or
second kind using the collocation Pmethod. The function is approximated by an
N−1
expansion of the form f (x) = i=0 ai PHI(i,x). The basis functions PHI(I,
X) must be supplied by the user. N is the number of points (which equals the
number of basis functions) to be used in the collocation method. A and B are
variables specifying respectively, the lower and the upper limit of the integral.
F is an array of length N, which will contain the calculated coefficients of the
expansion. F[I] contains the coefficient (ai ) of PHI(I, X) in the expansion for
the solution. X is an array of length N, which specifies the points to be used for
collocation. This array must be supplied by the user. REPS and AEPS specify
the accuracy with which the integrals are to be calculated. If IQ = 0, then these
parameters are passed on to the function ADPINT, while in other cases, these
parameters are ignored. IQ is an integer variable used to specify the treatment
of integrals required to calculate
Z B
PSI(I, X) = FKER(X, T) PHI(I, T) dT . (C.90)
A

If IQ 6= 0, then it is assumed that the function PSI(I, X) is supplied separately,

while if IQ = 0, then the integrals are calculated using the function ADPINT.
This function uses external function FUNK to calculate the integrand, which
in turn requires the functions PHI(I, X) and FKER(X, T). IT is an integer
variable which specifies the kind of Fredholm equation to be solved. IT = 1
and 2 corresponds to the Fredholm equations of the first and second kind,
respectively. Error status can be obtained from the returned value of FREDCO.
A value of 708 implies that IT > 2, or IT ≤ 0, or N < 1, in which case,
no calculations are performed. Apart from these, other values may be set by
the functions GAUELM or ADPINT, which are called by this function. This
function requires function GAUELM to solve the system of linear equations,
functions ADPINT and KRONRD (for IQ = 0) to evaluate the integrals and
function FUNK(X) to evaluate the required integrand. Further, the functions
FG(X) (= g(x)), PHI(I, X) (= φi (x)) and FKER(X, T) (= K(x, t), for IQ = 0)
or PSI(I, X) (for IQ 6= 0) must be supplied by the user. Depending on value
of IQ, some of these functions may not be required, but even in that case a
dummy function to satisfy the compiler or linker may be needed. The names of
these functions are fixed and cannot be passed on to the function FREDCO,
because the functions FKER and PHI are called by the function FUNK, and
there is no simple way of passing on these names to FUNK. In order to maintain
consistency, the names of other functions are also not passed on to the function.
The global variables XFRED and IFRED are used to pass on the value of x
and i to FUNK for calculating the integrand. This function can also be used
to solve linear Volterra equations, by defining the kernel to be zero for t > x.
C.13. Integral Equations 1131

204. FUNK Function to calculate the integrand for calculating ψi (x) as re-
quired by function FREDCO. Function FKER(X, T) is the kernel K(x, t), while
PHI(I, T) are the basis functions φi (t) used for the expansion. Names of these
functions have to be the same as there is no provision to pass on the names
to FUNK. The values of X and I are passed on through the global variables
XFRED and IFRED.
205. RLS Function to solve a linear inverse problem in one dimension using
Regularised Least Squares (RLS) technique with B-spline basis functions. The
inverse problem is defined as
Z b
di = Ki (t)f (t) dt, i = 0, 1, . . . , NM − 1; (C.91)
a

where the unknown function f (t) is to be obtained using given data di and ker-
nels Ki (t) for all i. The corresponding forward problem where di are calculated
for a given f (t) is solved by the function FORW. The kernel is given in the
form of table of values at a grid of points in t covering the required interval.
Alternately, it is possible to provide directly the coefficients of matrix which
are Z b
aij = Ki (t)φj (t) dt , (C.92)
a

where φj (t) is the jth basis function. This form is more efficient to use and
can be utilised when many options are to be tried with the same data points
and set of knots. The function itself calculates these integrals and hence during
second and subsequent calls this option can be used, provided the data points
and knots have not been modified. Using the coefficients aij instead of kernels
also saves storage space. It may be noted that for this to be useful, di need
not have the same values between different sets, only the set of NM kernels
should be the same between different inversion problems. This option is useful
as the solution will need to be calculated with different values of regularisation
parameter before the optimum value is found. This function can also estimate
the errors in computed solution using Monte Carlo simulation.
NK is the number of knots defining the B-spline basis functions. The num-
ber of basis functions would be NK + K − 2. XO is an array of length at least
NK which should contain the knots for defining the B-splines. The knots must
be in ascending order with XO[0] containing the first knot. K is the order of
B-splines required, K = 2 for linear B-splines while K = 4 for cubic B-splines,
etc. NR is the number of points used for defining the kernels. R is an array
of length NR, which should contain the points ri at which kernel values are
available. This array should be in ascending order. The same set of points are
used to calculate the solution after the coefficients of expansion are obtained.
Hence, this array will be required even if matrix coefficients are directly avail-
able (unless NR = 0). RKER is an array of length IK × NR containing the
kernels for the inverse problem. RKER[j][i] should contain Ki (rj ). This array
must be supplied if IFLG < 2, otherwise it is not required. A dummy array
1132 Appendix C. C Programs

may be required to satisfy the compiler, even if IFLG ≥ 2. The mesh rj need
not be uniformly spaced as the integrals are evaluated using the trapezoidal
rule. Since the accuracy of trapezoidal rule is low and also because in realistic
problems the kernels are generally highly oscillatory, a large number of mesh
points will be required to define the integrals to a good accuracy. IK is the
second dimension of arrays RKER and AC, as specified in the calling function.
IK must be at least NM + NS. AC is an array of length IK × (NK + K − 2)
containing the coefficients of the matrix defining the inversion problem. AC[j][i]
should contain the coefficient aij defined by (C.92). If IFLG < 2, these coeffi-
cients are calculated by evaluating the integrals in (C.92) using the trapezoidal
rule. These calculations require substantial computing and hence on subsequent
attempts the calculations can be suppressed by using IFLG = 2, 3. The matrix
coefficients depend on the kernels (and the number of data points) and the
knots and order of B-splines. Thus if these are not modified between two calls,
it is preferable to use IFLG = 2 to find the solution using pre-calculated coef-
ficients. If either the kernels or basis functions are modified then IFLG should
be reset to 0 to force calculations of the coefficients once again.
NM is the number of data points (and kernels) in the inversion problem.
For each data point, the corresponding kernel must be supplied. NS is the num-
ber of points to be used for calculating the regularisation term. The function
chooses a uniform mesh covering the full interval for this purpose. ALP is the
regularisation parameter, which must be positive. IDE is an integer parameter,
which specifies the order of derivative to be used for regularisation. IDE should
be 1 or 2 for first or second derivative smoothing. DI is an array of length NM,
which should contain the data di , for inversion. DE is an array of length NM,
containing the estimated errors in di . These errors must be positive. DF is an
array of length NM, which will contain the normalised residuals obtained by
fit. If DS[I] is the computed value of DI[I] using the solution, then the resid-
ual DF[I]=(DI[I]−DS[I])/DE[I]. Ideally, these residuals should have a Gaussian
distribution with zero mean and unit variance. A significant trend in the resid-
ual would generally imply that ALP should be reduced or number of knots is
not sufficient to represent the solution. F is an array of length NR which will
contain the calculated solution at R[I]. B is an array of length NM + NS, which
will contain the coefficients of expansion. The computed solution is
NK+K−3
X
f (t) = B[i]φi (t). (C.93)
i=0

These coefficients can be used to compute the solution at any required point us-
ing function BSPEVL. Although, the number of coefficients is only NK + K − 2,
the rest of the array is used as scratch space.
IFLG is an integer parameter, which specifies the type of calculation re-
quired. If IFLG = 0, the matrix coefficients are computed using the kernels.
The system of equations is solved to find the solution and IFLG is set to 4 after
calculations. The calculation of integrals to find matrix coefficients take sig-
nificant time and hence this option should be used only when the coefficients
C.13. Integral Equations 1133

are not available. After the first call to RLS, the coefficients are computed
and returned in array AC. These coefficients should be used again if another
solution is required with same set of kernels but with different regularisation
(ALP or IDE) or with different data points DI for the same set of kernels. If
IFLG = 1, the matrix coefficients are computed using the kernels. The system
of equations is setup and the SVD of the matrix is computed, but the solution
for given DI is not computed. In this case also IFLG is set to 4 after calcula-
tions. Thus during subsequent calls solution for different DI’s can be computed
using SVDEVL. If IFLG = 2, the matrix coefficients are assumed to be known
and available in array AC. These coefficients are used to setup the matrix and
find the solution and IFLG is set to 4 after calculations. This option is more
efficient than IFLG = 0, and should be used when the coefficients have already
been computed. In this case the kernels RKER are not used. If IFLG = 3,
the matrix coefficients are assumed to be known and available in array AC.
These coefficients are used to setup the matrix and its SVD is computed, but
the solution for given DI is not computed. In this case also IFLG is set to 4
after calculations. Thus during subsequent calls solution for different DI’s can
be computed using the function SVDEVL.
If IFLG = 4, the SVD of equation matrix is assumed to be available
in arrays, A, AV and SIGMA. These can be used to compute the solution
for a given data set DI very efficiently. It may be noted that the matrix of
equations depends only on the knots, kernels, regularisation parameters, (ALP
and IDE) and errors DE, but not on DI. So if only DI is changed there is no
need to recompute the SVD. This is very useful, since on many applications
several different data sets have to be inverted. Another application comes in
estimating errors in computed inverse solution. For this purpose we need to
solve the problem with different DI, which differ only in addition of different
realisation of random errors to same basic data. Thus for estimating the errors
in inversion although several different solutions are required, the net effort is
not much since the SVD, which takes much more time, is computed only once.
It should be noted that in all cases the function sets IFLG to 4 so if during
subsequent call the matrix needs to be changed, IFLG must be set to 0 or 2
before calling. This would be needed if either the knots, kernels or errors DE
are changed. If only errors are modified, then IFLG can be set to 2, to avoid
computation of coefficients. While if kernels or knots are modified then IFLG
should be set to 0, so that the coefficients are recomputed. It may be noted
that the set of kernels may change if we are interested in studying the effect of
dropping some of these modes on the computed solution.
REPS is the required accuracy for solution of equations using SVD. All
singular values less than REPS times the largest singular value will be set to
zero during solution. CHISQ is the computed value of χ2 define by
 2
NM−1 NM−1
X  2 NK+K−3
X 1 X
χ2 = DF[i]2 = DI[i] − aij B(j) , (C.94)
i=0 i=0
DE[i] j=0
1134 Appendix C. C Programs

SUMD is the value of regularisation term computed at the solution, for IDE = 2,
it is given by
 NS−1 2
d2 f
 X 
R[NR − 1] − R[0]
SUMD = (ti ) , (C.95)
NS − 1 i=0
dt2

where ti are the points at which smoothing is applied. For IDE = 1 the second
derivative is replaced by first derivative. The values of CHISQ and SUMD for
different values of ALP can be used to infer the optimal value of ALP using
L-curve. A is an array of length IV × (NM + NS) containing the matrix U of
SVD of the matrix of equations. AV is an array of length IV × (NK + K − 2)
containing the matrix V of SVD of the matrix of equations. IV is the second
dimension of arrays A and AV as declared in the calling function. SIGMA is
an array of length (NK + K − 2) containing the singular values of the matrix.
If IFLG < 4, the arrays A, AV and SIGMA will be calculated, otherwise they
must be supplied. NSIM is the number of data sets to be used for Monte Carlo
simulation for estimating the errors in solution. If NSIM < 2, error estimates
are not calculated. FE is an array of length NR which will contain the estimated
errors. FE[I] would contain the estimated error in F[I]. This is calculated only
if NSIM > 1. Error status can be obtained from the returned value of RLS.
A value of 705 implies that NM ≤ NK + K − 2 or IK < NM + NS or IV <
NK + K − 2. RLS = 710 implies that ALP < 0 or IDE < 1 or IDE > 2. In
all these cases no calculations are done. Other values may be set by functions
BSPLIN or SVD which are called. It may be noted that the function allows
solution to be computed even when ALP = 0, although the solution would be
unacceptable in this case. This function requires functions BSPLIN, BSPEVL,
SVD, SVDEVL and RANGAU.
206. FORW To solve the forward problem corresponding to the inverse prob-
lem solved by RLS. This function evaluates the required integrals using the
trapezoidal rule. NP is the number of points used in defining the kernels. NM
is the number of data points, di to be calculated. This should be the same
as the number of kernels that are supplied in RKER. R is an array of length
NP containing the points at which values of kernels are available. RKER is
an array of length IK × NP containing the kernels for the inverse problem,
RKER[j][i] = Ki (R[j]). These kernels are the same as what are used by RLS
for inversion. IK is the second dimension of RKER as declared in the calling
function, IK ≥ NM. DI is an array of length NM, which will contain the calcu-
lated data points di , using the kernels. F is an array of length NP containing the
function values, F[i] = f (R[i]). If IFLG = 0, the function values are calculated
using user supplied function FUN, otherwise, these values must be supplied
while calling the function. FUN is the name of the function routine to calculate
the given function. This is used only if IFLG = 0, otherwise the function values
are to be supplied in the array F. In the latter case a dummy function will still
be required to satisfy the compiler or linker. Error status can be obtained from
the returned value of FORW. A value of 711 implies that IK < NM and no
C.13. Integral Equations 1135

calculations are done. IFLG is an integer parameter that is used as a flag to

decide the type of computation required. If IFLG = 0, then the function values
are calculated using a user supplied function FUN. These values are returned
in the array F and IFLG is set to 1, so that next time the values need not be
calculated. If data points corresponding to different kernels are required then
IFLG need not be reset. If a different FUN is used next time then IFLG must
be reset to 0. For other values of IFLG the function values must be supplied in
the array F. Function FUN(X) must be supplied by the user to calculate f (x).
If the data points DI calculated by this function are supplied to RLS without
adding any errors and using the same kernels with very small RLM, then the
inverted function should match the function FUN used by FORW to a very
good accuracy.
207. VOLT Function to solve linear Volterra equations using quadrature
method based on the trapezoidal rule. N is the number of points at which
the function value needs to be estimated. A is a variable specifying the lower
limit of the integral. It is also the initial point, from which the solution is re-
quired. H is the step size to be used in the quadrature method. The points are
assumed to have a uniform spacing of H. F and X are arrays of length N, which
will contain the calculated value of the solution at a set of uniformly spaced
points. F[I] is the calculated solution at X[I]. FG and FKER are the names of
the functions used to calculate the g(x) and the kernel K(x, t), respectively.
IT is an integer variable specifying the kind of integral equation. If IT = 2,
Volterra equation of the second kind are solved, while for other values of IT
a Volterra equation of the first kind is solved. If IT = −1, then the computed
values are smoothed as explained in Section 13.8. This option can be used only
for equations of the first kind. It should be noted that smoothing cannot be
applied to the first and the last point. Error status can be obtained from the
returned value of VOLT. A value of 712 implies that N < 3, in which case
no calculations are performed. VOLT = 751 implies that the denominator is
zero at some stage, in which case, no further calculations can be performed.
This failure usually occurs for equations of the first kind when either H = 0 or
K(x, x) = 0. Functions FG(X) and FKER(X, T) must be supplied by the user.
208. VOLT2 Function to solve nonlinear Volterra equation of the second kind
Z x
K(x, t, f (t)) dt = f (x) + g(x), (C.96)
A

using quadrature method based on the Simpson’s rule. N is the number of points
at which the function value needs to be estimated. A is the lower limit of the
integral. It is also the initial point from which the solution is required. H is the
step size to be used in the quadrature method. The points are assumed to have
a uniform spacing of H. F and X are arrays of length N, which will contain the
calculated value of the solution at a set of uniformly spaced points. F[I] is the
calculated solution at X[I]. FG and FKER are the names of the functions used to
calculate the g(x) and the kernel K(x, t, f ), respectively. REPS is a parameter,
1136 Appendix C. C Programs

which specifies the (relative) accuracy to which the resulting nonlinear algebraic
equations are solved. This parameter does not control the truncation error,
which is determined by H. It is only used to decide the termination criterion for
the fixed-point iteration, while solving the nonlinear equations for F[I]. Error
status can be obtained from the returned value of VOLT2. A value of 712
implies that N < 3, in which case no calculations are performed. VOLT2 = 752
implies that the fixed-point iteration fails to converge at some point, in which
case, the calculations are aborted at that point. In this case the values already
calculated so far will be available in arrays X and F, but their reliability may
need to be checked. Functions FG(X) and FKER(X, T, F) must be supplied by
the user. It may be noted that FKER is not the kernel in the usual sense since
it also includes the unknown function f (t) as it appears inside the integral in
Eq. (C.96).

C.14 Partial Differential Equations

All functions in this chapter are simple implementations of finite difference
methods. Although these functions may accept equations in somewhat general
form with variable coefficients, the result may not be reliable in all cases. For
example, if the coefficient A of the parabolic equation in CRANK is negative in
some range of (x, t) values, no reliable solution can be computed. The function
does not check for such inconsistencies and no error message will be issued.
Similarly, some functions allow for rather general boundary conditions, but the
accuracy may be only first order in ∆t, if the boundary conditions depend
on time explicitly. Further, for special differential equations, it is possible to
improve the efficiency significantly by writing the difference equations directly
without calculating the coefficients. Similarly, the boundary conditions can be
implemented directly to improve efficiency in some of the functions (e.g., ADI).
On the other hand, those functions which accept boundary condition in sim-
ple Dirichlet form only, can be modified to incorporate more general boundary
conditions. Hence, these functions can only be treated as concrete (and straight-
forward) examples of implementing the algorithms covered in Chapter 14. They
are not expected to be particularly efficient or robust. Users are expected to
modify these functions to suit their problems.
209. CRANK Function to solve a linear parabolic equation of the form

∂u ∂2u ∂u
= A(x, t) 2 + B(x, t) + C(x, t)u + D(x, t), (C.97)
∂t ∂x ∂x
subject to boundary conditions

∂
A0 (t)u(X0, t) + B0 (t) u(X0, t) = F0 (t),
∂x (C.98)
∂
An (t)u(XN, t) + Bn (t) u(XN, t) = Fn (t),
∂x
C.14. Partial Differential Equations 1137

using the Crank-Nicolson difference scheme (14.29). Although the coefficients

in the boundary conditions are functions of time, if B0 = 0 at some time step,
then it must be so at every time step. The same is true for the coefficient Bn .
This essentially means that if boundary condition is of Dirichlet type, it should
remain so at all times.
T is the initial value of time. After execution, it is replaced by the value
of t at the last point. DT is the time step to be used in the computations.
(X0, XN) is the range of x values over which the equation needs to be solved.
NT is the number of time steps each of length DT, over which the equation is to
be integrated. NX is the number of mesh points in the X direction. X is an array
of length NX, which contains the mesh points in X direction. This array need
not be initialised at input, since the function assumes a uniform mesh spacing
and calculates the required mesh points. U is an array of length NX, containing
the solution at some time step. This array can be used to supply the initial
values, while after execution, it will contain the solution at the required time.
COF and BC are the names of the functions used to calculate the coefficients in
the differential equation and the boundary conditions, respectively. Functions
COF(X, T, A, B, C, D) and BC(T, X0, XN, A0, B0, F0, AN, BN, FN) must
be supplied by the user. Function COF should calculate the coefficients A, B,
C, D at a given value of X and T. Function BC should calculate the coefficients
A0, B0, F0, AN, BN and FN for a given value of T at the end points X0 and
XN. FIC is the name of the function, which may be used to calculate the initial
values for the solution, when IFLG = 0. Function FIC(X, T) must be supplied
by the user if IFLG = 0. This function calculates the solution u(x, t) at the
initial time T. For other values of IFLG a dummy function FIC may be required
to satisfy the compiler or the linker. IFLG is an integer variable used as a flag.
If IFLG = 0, the initial values are calculated using the function FIC supplied
by the user. For other values of IFLG the initial values must be supplied in the
array U at the time of calling. In any case, the function sets IFLG to 1, so that
during subsequent calls to the function, the initial values are taken from the
array U. This function can be used to calculate the solution at several values
of T by calling it repeatedly, after changing DT and NT appropriately. Other
variables should not be changed. Error status can be obtained from the returned
value of CRANK. A value of 713 implies that DT = 0, or XN = X0, or NX ≤ 2,
in which case, no calculations are performed. CRANK = 761 implies that the
difference equations are singular and the solution cannot be calculated. This
problem may arise due to some error in specifying the equation or the boundary
conditions.
210. LINES Function to solve a system of nonlinear parabolic equations of
the form
∂u ∂ 2 u
 
∂ui
= fi x, t, u, , 2 , (i = 0, 1, . . . , NELIN − 1); (C.99)
∂t ∂x ∂x
subject to Dirichlet boundary conditions u(X0, t) = g1 (t) and u(XN, t) = g2 (t),
using the method of lines. This function is actually used through the functions
1138 Appendix C. C Programs

MSTEP or RKM of Chapter 12. Since the system of resulting ordinary differen-
tial equations are expected to be stiff, it is preferable to use the function MSTEP
with GEAR. Function LINES specifies the system of ordinary differential equa-
tions, as required by the function MSTEP (function DIF in the arguments for
MSTEP). The parameters for the partial differential equations are passed via
the global variables, DXLIN, X0LIN, XNLIN, XLIN, U0LIN, UNLIN, NXLIN,
NELIN. The relevant variables amongst these must be initialised before calling
the function MSTEP. NXLIN is the number of mesh points in the x direction.
NELIN is the number of parabolic equations in the system. DXLIN is the step
size in x direction. (X0LIN, XNLIN) is the interval over which the solution
is to be calculated. Since the boundary conditions are assumed to be in the
Dirichlet form, the solution at the end points need not be calculated. Hence,
the number of ordinary differential equations N = (NXLIN − 2)NELIN. XLIN
is an array of length NXLIN − 2, containing the mesh points xi , excluding the
end points. This array must be initialised before calling the function MSTEP.
U0LIN and UNLIN are arrays of length NELIN, containing the boundary val-
ues (at X0LIN and XNLIN) of the solution at the current time. These arrays
need not be initialised before calling MSTEP, since the boundary values are cal-
culated by function LINES using a user supplied function BC. The arguments
of LINES have the same definition as specified for function DIF as required
by MSTEP or RKM. T is the value of t at which the derivatives of solution
are required. N is the number of ordinary differential equations (which would
be (NXLIN − 2)NELIN). U and DU are arrays of length N. U will specify the
values of solution at required point and DU will contain the time derivative of
all these functions as calculated by LINES.
This function requires functions FCN and BC to supply the information
about the equations and the boundary conditions. Function FCN(NELIN, X,
T, U, UX, UXX, DU) calculates the derivatives

∂ui
DU[I] = = fi (X, T, U, UX, UXX), (I = 0, . . . , NELIN − 1), (C.100)
∂t

for a given value of X, T, U[J], UX[J], UXX[J], (J = 0, . . . , NELIN − 1), as

specified by the differential equation. Here U, UX, UXX and DU are arrays
of length NELIN, containing u, ∂u/∂x, ∂ 2 u/∂x2 , ∂u/∂t respectively. Function
BC(NELIN, T, X0LIN, XNLIN, U0LIN, UNLIN) calculates the boundary val-
ues U0LIN[I], UNLIN[I], (I = 0, . . . , NELIN − 1) for a given value of T, X0LIN
and XNLIN. The names of the functions FCN and BC are fixed and cannot be
passed on as arguments. The initial values should be supplied to the function
MSTEP at the time of calling. The NELIN × (NXLIN − 2) components of the
initial vector are arranged in the normal Fortran order, with NELIN compo-
nents of u at each of the NXLIN − 2 mesh points stored consecutively. Thus,
uj (xk ) occupies the k ×NELIN+j position in the initial vector. The final result
at the required value of time will also be returned in the same order.
C.14. Partial Differential Equations 1139

211. ADM Function to solve a parabolic equation in two space variables

∂u ∂2u ∂2u ∂u ∂u
= Axx (x, y, t) 2 + Ayy (x, y, t) 2 + Ax (x, y, t) + Ay (x, y, t)
∂t ∂x ∂ y ∂x ∂y
+ Au (x, y, t)u + A0 (x, y, t)
(C.101)
with Dirichlet boundary conditions over a rectangular region, using alternating
direction method. This function uses the difference scheme
n+1/2
ujl − unjl 1 n+1/2 1 n+1/2
= An+1/2 δx2 ujl + An+1/2 δx ujl
1
2 ∆t
(∆x)2 xx 2∆x x
1 1
+ An δ 2 u n + An δy unjl
(∆y)2 yy y jl 2∆y y
1 n+1/2 1 n+1/2
+ (Anu unjl + An+1/2
u ujl ) + (A0 + An0 ),
2 2
n+1/2
un+1
jl − ujl 1 n+1/2 1 n+1/2
= An+1/2 δx2 ujl + An+1/2 δx ujl
1
2 ∆t
(∆x)2 xx 2∆x x
1 1
+ An+1 δ 2 un+1 + An+1 δy un+1
(∆y)2 yy y jl 2∆y y jl

1 n+1/2 1 n+1/2
+ (An+1 un+1 + An+1/2 ujl ) + (A0 + An+1 ),
2 u jl u
2 0
(C.102)
where all coefficients on the right-hand side are evaluated at x = xj and y = yl .
T is the initial value of time, DT is the time step to be used. After execution,
T will be replaced by the current value of time. (X0, XN) and (Y0, YN) are
the intervals along x and y axes, in which the solution is required. NT is the
number of time steps, NX and NY are the number of mesh points in the x and
y directions. X is an array of length NX, which will contain the mesh points xj
in x direction. Y is an array of length NY, which will contain the mesh points
yl in y direction. The arrays X and Y need not be initialised before calling the
function, since the function assumes a uniform spacing and initialises the arrays
before proceeding with the calculations. U is an array of length IU × NY, which
contains the solution u. This array can also be used to pass the initial values of
the solution (if IFLG 6= 0). U[J][I] is the solution u(xI , yJ ). IU (≥ NX) is the
second dimension of U as declared in the calling function. COF is the name of
the function to calculate the coefficients of the equation. Function COF(X, Y,
T, AXX, AYY, AX, AY, AU, A0) should calculate the coefficients AXX, AYY,
AX, AY, AU and A0 at a given value of X, Y and T. BC is the name of the
function to calculate the boundary values. Function BC(IB, X, Y, T) should
calculate the boundary values for the solution. Here IB is an integer variable,
which takes the value 1, 2, 3, 4 corresponding to the four boundary lines. The
boundary values are given by
u(X0, y, t) = BC(1, X0, y, t), u(XN, y, t) = BC(2, XN, y, t),
(C.103)
u(x, Y0, t) = BC(3, x, Y0, t), u(x, YN, t) = BC(4, x, YN, t).
1140 Appendix C. C Programs

If the boundary values depend on time, then the accuracy of difference ap-
proximation may be only O(∆t) instead of O((∆t)2 ). FIC is the name of the
function, which may be used to calculate the initial values for the solution, when
IFLG = 0. Function FIC(X, Y, T) must be supplied by the user, if IFLG = 0.
This function calculates the solution u(x, y, t) at the initial time T. For other
values of IFLG a dummy function FIC may be required to satisfy the com-
piler, or the linker. IFLG is an integer variable used as a flag. If IFLG = 0,
the initial values are calculated using the function FIC supplied by the user.
For other values of IFLG, the initial values must be supplied in the array U
at the time of calling. In any case, the function sets IFLG to 1, so that during
subsequent calls to the function, the initial values are taken from the array U.
This function can be used to calculate the solution at several values of T, by
calling it repeatedly after changing DT and NT appropriately. Other variables
should not be changed. Error status can be obtained from the returned value
of ADM. A value of 714 implies that DT = 0, or XN = X0, or YN = Y0, or
NX ≤ 2, or NY ≤ 2, or IU < NX, in which case, no calculations are performed.
ADM = 762 implies that the difference equations are singular and the solution
cannot be calculated. This problem may arise due to some error in specifying
the equation.

212. LAX Function to solve a system of hyperbolic equations in the con-

servation law form, using Lax-Wendroff difference scheme. The equations are
assumed to be in the form ∂u/∂t + ∂f /∂x = 0, where f (x, t, u) can be a nonlin-
ear function of u. N is the number of equations in the system, T is the initial
value of time. After execution, T will be replaced by the current value of t. DT
is the time step to be used. The time step is kept fixed. (X0, XN) is the interval
along x-axis, on which the solution is required. NT is the number of time steps
over which integration is to be performed, NX is the number of mesh points to
be used along the x-axis. X is an array of length NX, used to store the coor-
dinates of the mesh points xj . This array need not be initialised before calling
the function, since the function assumes a uniform spacing and calculates xj
before performing the calculations. U is an array of length IU × NX, which
contains the solution u. This array can also be used to pass the initial values of
the solution (if IFLG 6= 0). IU[J][I] contains the Ith component of the solution
ui at Jth mesh point xj . IU (≥ N) is the second dimension of U, as declared
in the calling function. FLUX is the name of the function, which is used to
calculate the flux f occurring in the differential equations. Function FLUX(N,
X, T, U, F) should calculate the flux F[I], (I = 0, . . . , N − 1) for a given value of
X, T and U[J], (J = 0, . . . , N − 1). BC is the name of the function to calculate
the boundary values. Function BC(IB, N, X, T, IW, A) should calculate the
coefficients required for the boundary conditions. Here IB is an integer variable
which takes the value 1 or 2 corresponding to X = X0 and XN, respectively.
IW is an integer array of length N, which contains information about the type
of boundary condition on each component of u or f . If IW[I] = 0, then the
boundary condition is of the simple Dirichlet form ai1 (t)ui = ai3 (t), otherwise
C.14. Partial Differential Equations 1141

the boundary condition is applied to the corresponding component of the flux

∂fi
ai1 (t)fi + ai2 (t) = ai3 (t). (C.104)
∂x
The coefficients aij are returned in the array A, which should be dimensioned
as A[3][N] in the function BC. FIC is the name of the function, which may
be used to calculate the initial values for the solution if IFLG = 0. Function
FIC(N, X, T, U) must be supplied by the user, when IFLG = 0. This function
calculates the solution ui (X, T), (i = 0, . . . , N − 1) at the initial time T. For
other values of IFLG a dummy function FIC may be required to satisfy the
compiler or the linker. (In function FIC, U is an array of length N, which
returns the value of the solution ui for i = 0, . . . , N − 1.) IFLG is an integer
variable used as a flag. If IFLG = 0, the initial values are calculated using
the function FIC supplied by the user. For other values of IFLG, the initial
values must be supplied in the array U. In any case, the function sets IFLG to
1, so that during subsequent calls to the function, the initial values are taken
from the array U. This function can be used to calculate the solution at several
values of T by calling it repeatedly, after changing DT and NT appropriately.
Other variables should not be changed. Error status can be obtained from the
returned value of LAX. A value of 715 implies that DT = 0, or XN = X0, or
NX ≤ 2, or IU < N, in which case, no calculations are performed. LAX = 763
implies that the denominator vanishes while calculating the boundary values
and hence the solution cannot be continued. This problem may arise due to
some error in specifying the boundary conditions.
213. SOR Function to solve a linear second-order elliptic equation

∂2u ∂2u ∂2u ∂u

Axx (x, y) 2
+ Axy (x, y) + Ayy (x, y) 2 + Ax (x, y)
∂x ∂x∂y ∂y ∂x
(C.105)
∂u
+ Ay (x, y) + A0 (x, y)u + F (x, y) = 0,
∂y

with Dirichlet boundary conditions on a rectangular region. Although arbitrary

coefficients will be accepted by this function, the result may be meaningful only
when the resulting equation is elliptic, i.e. Axx Ayy − 4A2xy > 0. This function
uses the SOR method to solve the difference equations
2ujl − uj−1,l − uj+1,l uj+1,l−1 + uj−1,l+1 − uj+1,l+1 − uj−1,l−1
Axx 2
+ Axy
(∆x) 4∆x∆y
2ujl − uj,l−1 − uj,l+1 uj−1,l − uj+1,l uj,l−1 − uj,l+1
+ Ayy 2
+ Ax + Ay
(∆y) 2∆x 2∆y
− A0 ujl = F,
(C.106)
where all coefficients are evaluated at (xj , yl ). The points (X0, Y0) and
(XN, YN) define the rectangle over which the solution is required. NX and
1142 Appendix C. C Programs

NY are the number of mesh points along the x and y directions. X and Y are
arrays of lengths NX and NY respectively, which contain the mesh points xj
and yl . These arrays need not be initialised before calling the function, since
the function assumes a uniform spacing and initialises these arrays before pro-
ceeding with the calculations. U is an array of length IU × NY, which contains
the solution. U[J][I] will contain the computed approximation to u(xi , yj ). Be-
fore calling the function, an initial approximation to the solution should be
supplied in this array. IU (≥ NX) is the value of the second dimension of array
U, as declared in the calling function. COF is the name of the function used
to calculate the coefficients in the equation. Function COF(X, Y, AXX, AXY,
AYY, AX, AY, A0, F) should calculate the coefficients AXX, AXY, AYY, AX,
AY, A0 and F at a given value of X and Y. BC is the name of the function
to calculate the boundary values. Function BC(IB, X, Y) should calculate the
boundary values for the solution. Here IB is an integer variable which takes the
value 1, 2, 3, 4, corresponding to the four boundary lines. The boundary values
are given by

u(X0, y) = BC(1, X0, y), u(XN, y) = BC(2, XN, y),

(C.107)
u(x, Y0) = BC(3, x, Y0), u(x, YN) = BC(4, x, YN).

OMEGA is the value of the relaxation parameter ω to be used in the calcula-

tion. For meaningful results 1 < ω < 2, but this condition is not checked by
the function. If OMEGA ≤ 0, then the function uses the optimum value for the
Poisson’s equation, based on the number of mesh points in each direction (Sec-
tion 14.8). AEPS is the convergence parameter, which specifies the accuracy
to which the difference equations are to be solved. The iteration is terminated
when the maximum change in all elements is less than AEPS. It should be
noted that, this parameter does not affect the truncation error, which depends
only on the number of mesh points. For optimum results, AEPS should be
somewhat smaller than the expected accuracy in the solution. Accuracy of the
solution can be checked by repeating the calculations with different number
of mesh points. NIT is an output parameter, which gives the number of SOR
iterations required by the function to achieve the specified accuracy. Error sta-
tus can be obtained from the returned value of SOR. A value of 716 implies
that XN = X0, or YN = Y0, or NX ≤ 2, or NY ≤ 2, or IU < NX, in which
case, no calculations are performed. SOR = 764 implies that the diagonal term
in the difference equation vanishes and the calculations are abandoned. This
failure may arise due to some error in specifying the equation, or the equation
may be such that the diagonal term actually vanishes at some mesh point. In
the latter case, changing mesh spacing along some axis may make the coeffi-
cient nonzero. SOR = 765 implies that the SOR iteration failed to converge in
MAXIT (= 1000) iterations.
C.14. Partial Differential Equations 1143

214. ADI Function to solve a linear second-order elliptic equation

∂2u ∂2u ∂u
Axx (x, y) 2
+ Ayy (x, y) 2 + Ax (x, y)
∂x ∂y ∂x
(C.108)
∂u
+ Ay (x, y) + A0 (x, y)u + F (x, y) = 0,
∂y

on a rectangular region using the Alternating Direction Implicit iterative (ADI)

method. Although arbitrary coefficients will be accepted by this function, the
result may be meaningful only when the resulting equation is elliptic, i.e.
Axx Ayy > 0. This function accepts more general boundary conditions of the
form A0 u+An ∂n u = F , where ∂n u is the normal derivative. The points (X0, Y0)
and (XN, YN) define the rectangle, over which the solution is required. KN is
the parameter k in the ADI algorithm. This function repeats a cycle of 2k it-
erations, until the result converges. The convergence criterion is tested after
each iteration. Hence, the actual number of iterations may not be a multiple
of 2k . Ideally k should be chosen such that convergence takes place in approx-
imately 2k iterations, but the exact choice may not be very crucial. NX and
NY are the number of mesh points along the x and y directions. X and Y are
arrays of lengths NX and NY, respectively containing the mesh points xj and
yl . These arrays need not be initialised before calling the function, since the
function assumes a uniform spacing and initialises these arrays before proceed-
ing with the calculations. U is an array of length IU × NY, which contains the
solution. U[J][I] will contain the computed approximation to u(xi , yj ). Before
calling the function an initial approximation to the solution should be supplied
in this array. IU (≥ NX) is the value of the second dimension of the array
U, as declared in the calling function. COF is the name of the function used
to calculate the coefficients in the equation. Function COF(X, Y, AXX, AYY,
AX, AY, A0, F) should calculate the coefficients AXX, AYY, AX, AY, A0 and
F at a given value of X and Y. BC is the name of the function to calculate
the coefficients for the boundary conditions. Function BC(IB, X, Y, A0, AN,
F) should calculate the coefficients A0, AN and F for the boundary conditions.
Here IB is an integer variable which takes the value 1, 2, 3, 4, corresponding
to the four boundary lines x = X0, x = XN, y = Y0 and y = YN, respectively.
EL and EU are respectively, the lower and upper limit on the eigenvalues of
the partitions X and Y of the finite difference matrix. If EL ≤ 0, then the
value for Poisson’s equation will be used. Similarly, if EU ≤ EL, then the value
for Poisson’s equation will be used. AEPS is the convergence parameter, which
specifies the accuracy to which the difference equations are to be solved. The
iteration is terminated when the maximum change in all elements is less than
AEPS. It should be noted that, this parameter does not affect the truncation
error, which depends only on the number of mesh points. For optimum results,
AEPS should be somewhat smaller than the expected accuracy in the solution.
Accuracy of the solution can be checked by repeating the calculations with
different number of mesh points. NIT is an output parameter, which gives the
1144 Appendix C. C Programs

number of ADI iterations required by the function to achieve the specified ac-
curacy. Error status can be obtained from the returned value of ADI. A value of
717 implies that XN = X0, or YN = Y0, or NX ≤ 2, or NY ≤ 2, or IU < NX, in
which case, no calculations are performed. ADI = 766 implies that the matrix
for ADI iteration is singular and the calculations are abandoned. This problem
may arise due to some error in specifying the equation or the boundary condi-
tions. ADI = 767 implies that the ADI iteration failed to converge in MAXIT
(= 1000) iterations.

Bibliography
Brent, R. P. (1973): Algorithms for Minimization Without Derivatives, Prentice-Hall, Engle-
wood Cliffs, New Jersey.
DiDonato R., Morris, A. H. Jr. (1986) Computation of the Incomplete Gamma Function
Ratios and their Inverse, ACM Transactions on Mathematical Software (TOMS), 12,
p. 377.
DiDonato R., Morris, A. H. Jr. (1992) Algorithm 708: Significant Digit Computation of the
Incomplete Beta Function Ratios, ACM Transactions on Mathematical Software (TOMS),
18, p. 360.
Kernighan, B. W. and Ritchie, D. M. (1988): The C Programming Language, (2nd ed.),
Prentice Hall, New Jersey.
Press, W. H., Teukolsky, S. A., Vetterling, W. T. and Flannery, B. P. (2007): Numerical
Recipes : The Art of Scientific Computing, (3rd ed.) Cambridge University Press, New
York.
Wilkinson, J. H. and Reinsch, C. (1971): Linear Algebra: Handbook for Automatic Compu-
tation, Vol. 2, Springer-Verlag, Berlin.