Floating point error in Matrix.invert() #1
Labels
Comments
hal7df
added a commit
that referenced
this issue
Sep 2, 2017
- All tests are passing (with current coded-in exception for issue #1) - Don't be wishy-washy about using Matrix.multiply with a scalar value (either provided as a primitive JavaScript value or as a 1x1 matrix). Matrix.multiply will only accept arguments that have conforming dimensions.
hal7df
added a commit
that referenced
this issue
Sep 6, 2017
Matrix.rref now chooses scale or divide based on the magnitude of the unscaled leading value. Hopefully should cut down on floating point error (issue #1), but tests don't seem to be indicating that...
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Both of the (valid) unit tests for Matrix.invert() require an epsilon scale factor of 1e10 (i.e. only accurate to 5-6 decimal places), which is undesirable. This should be investigated to find a solution for increasing the accuracy of the inversion operation.
Curiously, the rref() method seems to be accurate to 14 decimal places, yet the inverse it produces is less accurate. Is this a side effect of the algorithm/is there a better matrix inversion algorithm to use?
Here's an example from one of the test cases:
Actual:
Computed:
Observed in Chromium 60.0.3112.78 on Ubuntu 16.04
The text was updated successfully, but these errors were encountered: