I am attempting to solve the system Ax = b in MATLAB, where A is a 30x30 triangular matrix with (nonzero) values ranging from 1e-14 to 0.7, and b is a 30-element column vector with values ranging from 1e-3 to 1e3. When I enter x = A\b, I get an answer and no warning messages, but the answer is not reasonable (looks like just random numbers at the bottom of the vector). I presume this is due to numerical errors.
Message 5 on this page suggests decomposing/scaling the matrix in order to avoid numerical errors, but I haven't been able to figure out how to calculate the scaling matrices.
So the question is: Is this indeed an example of numerical instability, and if so, how can I rescale my matrix A, or change how MATLAB is performing the calculation, to avoid it?
Here is the matrix and vector that are producing the issue:
A =
Columns 1 through 15
0.69 0.4278 0.19893 0.082223 0.031861 0.011852 0.0042866 0.0015187 0.00052965 0.00018243 6.221e-05 2.1038e-05 7.0653e-06 2.3587e-06 7.8344e-07
0 0.4761 0.44277 0.27452 0.14183 0.065953 0.028624 0.011831 0.0047156 0.0018273 0.00069233 0.00025755 9.4356e-05 3.4126e-05 1.2206e-05
0 0 0.32851 0.40735 0.3157 0.19573 0.10618 0.052668 0.02449 0.010846 0.004623 0.0019108 0.00077007 0.00030383 0.00011773
0 0 0 0.22667 0.35134 0.32675 0.23635 0.14653 0.081766 0.042246 0.02058 0.0095696 0.0042851 0.0018597 0.00078615
0 0 0 0 0.1564 0.29091 0.31564 0.26093 0.182 0.11284 0.064129 0.03408 0.017168 0.0082788 0.0038496
0 0 0 0 0 0.10792 0.23418 0.29039 0.27006 0.2093 0.14274 0.088499 0.05095 0.02764 0.014281
0 0 0 0 0 0 0.074464 0.18467 0.25761 0.2662 0.22694 0.16884 0.1134 0.070311 0.040868
0 0 0 0 0 0 0 0.05138 0.14335 0.22219 0.25256 0.23488 0.18931 0.13694 0.090965
0 0 0 0 0 0 0 0 0.035452 0.1099 0.18738 0.23235 0.2341 0.2032 0.15748
0 0 0 0 0 0 0 0 0 0.024462 0.083415 0.15515 0.20842 0.22614 0.21031
0 0 0 0 0 0 0 0 0 0 0.016879 0.062789 0.12652 0.18303 0.21277
0 0 0 0 0 0 0 0 0 0 0 0.011646 0.046935 0.10185 0.15786
0 0 0 0 0 0 0 0 0 0 0 0 0.008036 0.034876 0.081087
0 0 0 0 0 0 0 0 0 0 0 0 0 0.0055448 0.025783
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.0038259
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Columns 16 through 30
2.5906e-07 8.5327e-08 2.8007e-08 9.1646e-09 2.9906e-09 9.7342e-10 3.1613e-10 1.0246e-10 3.3142e-11 1.0702e-11 3.4504e-12 1.1108e-12 3.5709e-13 1.1465e-13 3.6767e-14
4.3246e-06 1.5194e-06 5.2988e-07 1.8359e-07 6.3236e-08 2.1667e-08 7.3883e-09 2.5085e-09 8.4833e-10 2.8585e-10 9.5998e-11 3.214e-11 1.073e-11 3.5726e-12 1.1866e-12
4.492e-05 1.6909e-05 6.2902e-06 2.3156e-06 8.445e-07 3.0543e-07 1.0963e-07 3.9084e-08 1.3847e-08 4.8779e-09 1.7094e-09 5.9615e-10 2.0698e-10 7.1568e-11 2.4651e-11
0.00032494 0.00013173 5.2503e-05 2.0616e-05 7.9887e-06 3.0592e-06 1.1591e-06 4.3497e-07 1.6181e-07 5.9715e-08 2.1877e-08 7.9615e-09 2.8794e-09 1.0354e-09 3.7037e-10
0.0017358 0.00076232 0.00032721 0.00013766 5.69e-05 2.3151e-05 9.2878e-06 3.679e-06 1.4406e-06 5.5824e-07 2.1426e-07 8.1515e-08 3.0763e-08 1.1523e-08 4.2867e-09
0.0070833 0.0033935 0.001578 0.00071495 0.00031662 0.00013741 5.8573e-05 2.4566e-05 1.0154e-05 4.1418e-06 1.6691e-06 6.6527e-07 2.6248e-07 1.0259e-07 3.9755e-08
0.022523 0.01187 0.0060211 0.0029554 0.0014095 0.00065541 0.00029799 0.00013279 5.8116e-05 2.5022e-05 1.0615e-05 4.4423e-06 1.8361e-06 7.5031e-07 3.0339e-07
0.056398 0.033024 0.018428 0.0098671 0.005098 0.0025529 0.0012436 0.00059114 0.00027488 0.00012531 5.6113e-05 2.4719e-05 1.0728e-05 4.5926e-06 1.9414e-06
0.11158 0.073506 0.045573 0.026843 0.01513 0.0082078 0.0043059 0.0021929 0.0010877 0.00052686 0.00024979 0.00011615 5.3064e-05 2.3852e-05 1.0563e-05
0.17385 0.13089 0.091294 0.059747 0.037043 0.021923 0.012459 0.0068335 0.0036315 0.0018763 0.00094518 0.00046536 0.00022441 0.00010618 4.9374e-05
0.21107 0.18539 0.14778 0.10881 0.074955 0.048796 0.030253 0.017976 0.010288 0.0056949 0.0030601 0.0016008 0.00081734 0.00040822 0.00019981
0.19575 0.20632 0.19188 0.16145 0.12513 0.090508 0.061727 0.04001 0.024806 0.014788 0.0085139 0.0047507 0.0025773 0.0013629 0.00070418
0.13406 0.17663 0.19712 0.1935 0.17139 0.13947 0.10569 0.075354 0.050967 0.032916 0.020408 0.012201 0.0070603 0.003967 0.0021702
0.063943 0.11233 0.1567 0.18459 0.19074 0.17739 0.15122 0.1198 0.089133 0.062798 0.042179 0.027157 0.016837 0.010091 0.0058655
0.018977 0.050003 0.093006 0.13695 0.16982 0.18425 0.17952 0.15999 0.13226 0.1025 0.075107 0.052387 0.034978 0.022461 0.013926
0.0026399 0.013912 0.038815 0.076207 0.11812 0.15379 0.17481 0.17806 0.16559 0.14259 0.11493 0.087452 0.063257 0.043745 0.029059
0 0.0018215 0.010164 0.029933 0.061862 0.10068 0.13733 0.16319 0.17345 0.16803 0.15048 0.12595 0.099387 0.074457 0.053266
0 0 0.0012569 0.0074028 0.022949 0.049799 0.084907 0.12108 0.15014 0.16622 0.16747 0.15575 0.13519 0.11049 0.085626
0 0 0 0.00086723 0.0053768 0.017502 0.039787 0.07092 0.10553 0.13631 0.15695 0.16421 0.15837 0.14237 0.12037
0 0 0 0 0.00059839 0.0038955 0.013284 0.031571 0.058722 0.091019 0.12227 0.1462 0.15862 0.15845 0.14736
0 0 0 0 0 0.00041289 0.0028159 0.010039 0.024896 0.048236 0.077756 0.10847 0.1345 0.15115 0.15618
0 0 0 0 0 0 0.00028489 0.0020313 0.0075564 0.019521 0.039334 0.065845 0.095256 0.12234 0.14222
0 0 0 0 0 0 0 0.00019658 0.0014625 0.0056673 0.015226 0.031861 0.05531 0.082873 0.1101
0 0 0 0 0 0 0 0 0.00013564 0.0010512 0.0042363 0.011819 0.025648 0.046115 0.071478
0 0 0 0 0 0 0 0 0 9.359e-05 0.00075433 0.0031569 0.0091339 0.020528 0.038183
0 0 0 0 0 0 0 0 0 0 6.4577e-05 0.00054051 0.0023458 0.0070296 0.016344
0 0 0 0 0 0 0 0 0 0 0 4.4558e-05 0.00038676 0.0017385 0.0053894
0 0 0 0 0 0 0 0 0 0 0 0 3.0745e-05 0.0002764 0.0012852
0 0 0 0 0 0 0 0 0 0 0 0 0 2.1214e-05 0.00019729
0 0 0 0 0 0 0 0 0 0 0 0 0 0 1.4638e-05
b =
3712
246.89
43.304
22.55
14.897
10.066
6.8138
4.6131
3.1232
2.1146
1.4316
0.96927
0.65623
0.44429
0.3008
0.20365
0.13788
0.093351
0.063202
0.04279
0.02897
0.019614
0.013279
0.0089906
0.006087
0.0041211
0.0027902
0.001889
0.0012789
0.00086589
A .mat file with the full-precision variables may be found here.
Here are the results I'm getting on my machine (Matlab R2013a on OS X 10.10.5):
>> x=A\b
x =
5087.6
433.99
64.166
27.995
19.494
14.546
10.934
8.2265
6.1834
4.6933
3.2779
3.8272
-3.5375
23.953
-79.278
254.22
-702.1
1713.2
-3658.2
6822.7
-11046
15412
-18349
18393
-15244
10181
-5273.4
1992.3
-489.85
59.155
Although norm(A*x-b) returns a value on the order of 1e-13, the results are not physically reasonable given the problem I am trying to solve (values in x should be monotonically decreasing, and none should be negative). As an example, here is a similar dataset that returns a correct (looking) solution with the same matrix A:
>> c
c =
5142.1
339.52
22.417
1.4802
0.097731
0.0064529
0.00042607
2.8132e-05
1.8575e-06
1.2265e-07
8.0979e-09
5.3469e-10
3.5304e-11
2.331e-12
1.5391e-13
1.0162e-14
6.7099e-16
4.4304e-17
2.9253e-18
1.9315e-19
1.2753e-20
8.4205e-22
5.5598e-23
3.671e-24
2.4239e-25
1.6004e-26
1.0567e-27
6.9771e-29
4.6068e-30
3.0418e-31
>> x = A\c
x =
7029.1
653.25
60.709
5.642
0.52434
0.04873
0.0045287
0.00042087
3.9114e-05
3.635e-06
3.3782e-07
3.1395e-08
2.9177e-09
2.7116e-10
2.52e-11
2.342e-12
2.1765e-13
2.0227e-14
1.8798e-15
1.747e-16
1.6236e-17
1.5089e-18
1.4023e-19
1.3033e-20
1.21e-21
1.1339e-22
9.9766e-24
1.1858e-24
2.3902e-26
2.078e-26