Multi normal derivatives

[spinkney]

Summary

This shows that multi_normal should just take the cholesky decomposition and call multi_normal_cholesky.

I just copied all the multi_normal_cholesky_lpdf code over. I wasn't sure how to just call it. That should be updated. The unit tests that call multi_normal_log fail.

Question: Would it be better to form the cholesky decomposition using ldlt instead of cholesky_decompose()?

Tests

No new tests.

Here's a benchmark vs the current multi_normal_lpdf

Side Effects

None

Release notes

multi_normal_lpdf now takes the Cholesky decomposition of the covariance matrix and calls multi_normal_cholesky_lpdf under the hood

Checklist

• [ ] Math issue #2544

• [x] Copyright holder: Sean Pinkney

  The copyright holder is typically you or your assignee, such as a university or company. By submitting this pull request, the copyright holder is agreeing to the license the submitted work under the following licenses: - Code: BSD 3-clause (https://opensource.org/licenses/BSD-3-Clause) - Documentation: CC-BY 4.0 (https://creativecommons.org/licenses/by/4.0/)

• [x] the basic tests are passing

  • unit tests pass (to run, use: ./runTests.py test/unit)
  • header checks pass, (make test-headers)
  • dependencies checks pass, (make test-math-dependencies)
  • docs build, (make doxygen)
  • code passes the built in C++ standards checks (make cpplint)

• [x] the code is written in idiomatic C++ and changes are documented in the doxygen

• [x] the new changes are tested

[spinkney]

@SteveBronder or @bob-carpenter , everything is working well except that the expression_tests fail for multi_normal_log.

I assume that if the expression_tests got to multi_normal_lpdf it would fail also. However, if I replace those with the test in multi_normal_cholesky_lpdf (but calling `multi_normal_lpdf' at the end of the test) then the test pass.

I think the issue is that there are some legacy signatures in multi_normal_log which just calls multi_normal_lpdf. So should we add those to `multi_normal_cholesky'? Or can we remove them?

[spinkney]

The one issue is the failing expression test. The issue is that the multi normal is testing the expression vector, array[] vector, matrix whereas the cholesky version is testing array[] vector, array[] vector, matrix. So now that the multi_normal calls the cholesky version, it is failing.

@rok-cesnovar is this a stanc required change?

[rok-cesnovar]

The expressions are now handled properly. @bob-carpenter would you have time to review this? You definitely know more about these multi normal lpdf than me.

[stan-buildbot]

---

Name	Old Result	New Result	Ratio	Performance change( 1 - new / old )
gp_pois_regr/gp_pois_regr.stan	3.47	3.39	1.02	2.23% faster
low_dim_corr_gauss/low_dim_corr_gauss.stan	0.02	0.02	1.0	0.25% faster
eight_schools/eight_schools.stan	0.09	0.09	0.99	-1.39% slower
gp_regr/gp_regr.stan	0.14	0.14	0.99	-0.98% slower
irt_2pl/irt_2pl.stan	5.11	5.13	1.0	-0.4% slower
performance.compilation	91.01	89.05	1.02	2.16% faster
low_dim_gauss_mix_collapse/low_dim_gauss_mix_collapse.stan	8.11	8.16	0.99	-0.52% slower
pkpd/one_comp_mm_elim_abs.stan	31.92	29.9	1.07	6.34% faster
sir/sir.stan	118.97	120.89	0.98	-1.61% slower
gp_regr/gen_gp_data.stan	0.03	0.03	0.99	-1.34% slower
low_dim_gauss_mix/low_dim_gauss_mix.stan	3.21	3.01	1.07	6.19% faster
pkpd/sim_one_comp_mm_elim_abs.stan	0.37	0.39	0.95	-5.15% slower
arK/arK.stan	2.03	2.03	1.0	-0.14% slower
arma/arma.stan	0.25	0.25	1.0	-0.1% slower
garch/garch.stan	0.61	0.6	1.02	1.75% faster
Mean result: 1.00571737881

---

Jenkins Console Log Blue Ocean Commit hash: 18f82c41c53586e354649e8b21e9324e13c3cbed

---

Machine information

ProductName: Mac OS X ProductVersion: 10.11.6 BuildVersion: 15G22010

CPU: Intel(R) Xeon(R) CPU E5-1680 v2 @ 3.00GHz

G++: Configured with: --prefix=/Applications/Xcode.app/Contents/Developer/usr --with-gxx-include-dir=/usr/include/c++/4.2.1 Apple LLVM version 7.0.2 (clang-700.1.81) Target: x86_64-apple-darwin15.6.0 Thread model: posix

Clang: Apple LLVM version 7.0.2 (clang-700.1.81) Target: x86_64-apple-darwin15.6.0 Thread model: posix

[spinkney]

This is ready, @bob-carpenter will you have time to look prior to the feature freeze?

[bob-carpenter]

I'm heading out of town on vacation for a few days, so I won't be able to review until next Tuesday at the earliest.

[spinkney]

No worries, enjoy your vacation!

[rok-cesnovar]

@spinkney mind merging in develop here? We should get this in for the next release.

[spinkney]

@rok-cesnovar can you restart the CI/jenkins process? Once that finishes I can merge in

[spinkney]

@rok-cesnovar (or @bob-carpenter or @bgoodri) I updated this to use the ldlt decom rather than the cholesky as it's more numerically stable. I attempted to keep the ldlt factor type and use it in the mdivide_*_ldlt for the derivatives but I was getting a bunch of type errors. I resorted to creating the cholesky factor from the L and sqrt(D) multiplication.

[bgoodri]

Taking the square root is where most of the precision gained from using LDLT gets lost. There should be a way to differentiate with respect to the diagonal elements of D directly.

[spinkney]

Taking the square root is where most of the precision gained from using LDLT gets lost. There should be a way to differentiate with respect to the diagonal elements of D directly.

Ok, then I may as well have kept it simple by taking the cholesky decomposition and passing that to multi_normal_cholesky_lpdf. The issue with that is the numerical precision of the cholesky decomp is sometimes going to fail and users will have to add a jitter to the diagonal.

The way I think this should be done is to create a multi_normal_ldlt_lpdf() that doesn't get exposed outside of stan-math. Autodiff through the ldlt then pass that into the multi_normal_ldlt function and utilize the LDLT functions to get the derivatives. My one worry is that we need the inverse of the cholesky factor L. Can we just use mdivide_left_ldlt(Sigma_ldlt) for this? Or will this introduce the sqrt problem all over again?

[syclik]

@spinkney, I didn't realize this was a draft. Let me know what you think.

[spinkney]

The issue is that the cholesky decomposition is less stable than the ldlt so I'm waiting until tuples get in the language or once someone implements ldlt with derivatives.