New calculus rules #10
Conversation
Codecov Report
@@ Coverage Diff @@
## master #10 +/- ##
==========================================
- Coverage 84.27% 83.81% -0.47%
==========================================
Files 45 49 +4
Lines 1475 1643 +168
==========================================
+ Hits 1243 1377 +134
- Misses 232 266 +34
Continue to review full report at Codecov.
|
src/calculus/HadamardProd.jl
Outdated
|
|
||
| Compose opeators such that: | ||
|
|
||
| `(Ax).*(Bx)` |
There was a problem hiding this comment.
This passage is not really clear. Maybe mention explicitly state that P*x = A*x .* B*x where P is the operator being constructed.
Also, this operator seems to in fact reduce to the usual Hadamard product between matrices in the linear case: is the nonlinear case a well known generalization? Just asking out of curiosity, to make sure there’s no confusion with the naming
There was a problem hiding this comment.
I'll make the docs more clear, thanks for pointing that out!
Not sure how much the nonlinear generalization is known, however it is, as usual, the chain rule applied 😄
Anyway the name could be perhaps improved with something like HadamardCompose, but it is very lengthy.
By the way my idea is to deprecate the other calc rule Hadamard since HadamardProd generalises it. So definitively it won't be confused with that one.
(updated first comment ⬆️ )
|
Yes I agree there’s a high level of redundancy... probably all the Ax_mul_Bx variants can be merged into a single one! But I still don’t have a clear idea on how! Maybe a new calculus rule that transposes the output would do the job... future work though! |
Ax_mul_Bx--> GeneralizesNonLinearComposeAxt_mul_BxAx_mul_BxtHadamardProd--> GeneralizesHadamardHadamard&NonLinearComposewill be deprecated in future version of AbstractOperators.