Use the covariance matrix for fitting error propagation

[Roman Tolchenov]

At the moment to estimate the errors of the calculated values we are using the formula:

err = sqrt( sigma² * J * J^T )

where sigma are standard deviations of the parameters and J is the Jacobian. This doesn't take into account correlation between the parameters and may lead to overestimated errors. I want to try to use a more accurate formula:

err = sqrt( J * C * J^T )

where C is the covariance matrix.

[Nick Draper]

Could you add a bit more description please.

[Roman Tolchenov]

Re #9030. Removed unused classes.

Changeset: 65a9f64fe4620062059b0f83f9f30e7a3fb586fc

[Roman Tolchenov]

Re #9030. Use covariance matrix to estimate errors of calc. values

Changeset: 390e8021a205792408b1f86517aa1df87aee4270

[Roman Tolchenov]

Re #9030. Added a unit test

Changeset: 18b7c64594ced719140c932ff96103f198f96f6d

[Roman Tolchenov]

Re #9030. Changed getCovarianceMatrix to return a const Matrix

Changeset: a09765134950053c58031b555cae637822450238

[Roman Tolchenov]

Re #9030. Use chi squared in error calculation

Changeset: db9ab977afc71f1788826529a0943b73fe432cd7

[Gesner Passos]

(In #9011) The entire fit will change the way it calculates error. It will have impact on CalculateTransmission.

[Roman Tolchenov]

Re #9030. Removed unused classes.

Changeset: f4e931b6b86783b92b2008dcead357a8465cca4e

[Roman Tolchenov]

Re #9030. Use covariance matrix to estimate errors of calc. values

Changeset: ed25aa030a67b2ca375cc4dcfd12f96bcd979fce

[Gesner Passos]

re #9030: New reference results after changing fit error calculation

Changeset: 2e8562e590c618e48cec80657cae71a139a51b23

systemtest: ​https://github.com/mantidproject/systemtests/commit/2e8562e590c618e48cec80657cae71a139a51b23

[Gesner Passos]

The last was not merged to develop, it is just a branch created in case Roman will need to merge them to develop later. The compare_errors.py was the file I used to compare the current results of SystemTests and the new ones when applying this ticket. Testers may be interested in using something like.

[Gesner Passos]

File that highlight the systemtests that will change with the new error propagation

[Roman Tolchenov]

Merge branch 'master' into feature/9030_fitting_error_propagation

Conflicts:

Code/Mantid/Framework/API/inc/MantidAPI/IDomainCreator.h Code/Mantid/Framework/CurveFitting/inc/MantidCurveFitting/FitMW.h Code/Mantid/Framework/CurveFitting/inc/MantidCurveFitting/MultiDomainCreator.h Code/Mantid/Framework/CurveFitting/src/FitMW.cpp Code/Mantid/Framework/MDEvents/inc/MantidMDEvents/FitMD.h Code/Mantid/Framework/MDEvents/src/FitMD.cpp

Re #9030

Changeset: 939b68caa756a22acd24c5a783d0dad9af34a6c5

[Roman Tolchenov]

Re #9030. Fixed an error after a merge with master.

Changeset: df6aaf38132040ec98e993c059fb66c40aa98bde

[Roman Tolchenov]

Re #9030. Fix a doxygen warning

Changeset: 34c1e2e3e6891da4ae1e75b72128ee30101cadc8

[Roman Tolchenov]

Re #9030. Replaced files to avoid merge conflicts.

Changeset: 5d448583305f341145c55f8f6ebacb2dc5784238

[Roman Tolchenov]

Re #9030. Updated ref files due to change in fitting errors.

Changeset: fa399288d2fa4758e84b4b2337bd65b2b4741aff

[Roman Tolchenov]

To test:

1. Load MUSR00015189 from AutoTestData.
2. Fit its 1st spectrum with a Polynomial(n=5) on interval 0.3-30.
3. Plot the calculated spectrum with error bars.
4. The error bars should be reasonable.

[Samuel Jackson]

Merge remote-tracking branch 'origin/feature/9030_fitting_error_propagation'

Full changeset: ​74c7ae4bcdaf3aae0402e19e604f141b68c7129b

[Samuel Jackson]

Merge remote-tracking branch 'origin/feature/9030_fitting_error_propagation'

Full changeset: ​38c7de6d6b91630fbc36d06052a87fd20d5f87b6

[Stuart Campbell]

This ticket has been transferred to github issue ​9873