BLAS-like kernel for symmetric updates in LDL factorization without pivoting

[poulson]

There is a long history of proposals to include simple modifications of ?syrk and ?herk to support C += alpha A (D A)^T, where D is diagonal. This kernel turns out to be important for Interior Point Methods, which often make use of LDL factorizations (without pivoting) of modified saddle-point systems.

Does such a routine already exist in BLIS? If not, is there a good place to start for adding support? Or a preferred name/convention?

[chriscoey]

Seconding this, but I would suggest further generalizing from diagonal D to symmetric/hermitian.

[fgvanzee]

@poulson @chriscoey Can you elaborate more on what you have in mind, since I'm unfamiliar with this operation?

Specifically, I'm interested in knowing the structure of A in C += alpha A (D A)^T. (Also, what is an LDL factorization? It's not clear to me how this relates to your request, or if it is strictly unnecessary background / context.)

[rvdg]

An LDL factorization is like a Cholesky factorization, but for a matrix that is symmetric but not necessarily positive definite. L is unit lower triangular and D is diagonal.

…

On Apr 26, 2019, at 12:35 PM, Field G. Van Zee ***@***.***> wrote: @poulson <https://github.com/poulson> @chriscoey <https://github.com/chriscoey> Can you elaborate more on what you have in mind, since I'm unfamiliar with this operation? Specifically, I'm interested in knowing the structure of A in C += alpha A (D A)^T. (Also, what is an LDL factorization? It's not clear to me how this relates to your request, or if it is strictly unnecessary background / context.) — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#31 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABLLYJY3726HZCG6KVBG6ULPSM4MXANCNFSM4BM5AJUA>.

[fgvanzee]

@rvdg Thanks, Robert. So is a real domain LDL factorization specified as A = L D L^T?

[devinamatthews]

@fgvanzee yes. The updates in question are the A -= L D L^T updates for A11, A21, and A22 (depending on variant).

@poulson I implemented C = A D B multiplications (with D diagonal) in TBLIS (interface and implementation). C = A B D is also possible. It should be possible to port this to BLIS and also extend to ?syrk.

@chriscoey we had an undergraduate working on general 3-matrix multiplication. @rvdg did that end up as a usable product?

[poulson]

@fgvanzee Yes, Robert (@rvdg) already clarified. Such factorizations are especially useful for symmetric quasi-(semi)definite matrices, which are, up to permutation, of the form [A, B; B', -C], where A, and C are Hermitian positive-(semi)definite. In the definite case, one can show bounds (e.g., via Saunders' analysis) on the stability of such factorizations. In practice, Interior Point Methods are built on this type of factorization (with the semidefinite cases modified to be definite).

[fgvanzee]

I still haven't seen anyone relate C += alpha A (D A)^T and A -= L D L^T. Even after you nix the alpha, harmonize the +/-, and assume A (in the first expression) is unit lower triangular, those don't look like the same operation to me.

[fgvanzee]

Nevermind, I see now that they are equivalent.

EDIT: No, I take that back. I'm still in doubt.

[fgvanzee]

Let A = L. I don't see how (D A)^T = D L^T. The lhs applies the diagonal elements to the columns of A^T (because (D A)^T = A^T D) while the rhs applies the diagonal elements to the rows of L^T.

This may seem like minutae, but I like to understand the operation before I think about implementing it.

Note: While I appreciate that there are sometimes rich stories to tell about the applications that employ these operations, I actually don't need to know how the operation is used (it just clutters my brain and doesn't stick anyway). I happily leave such knowledge-keeping to people like @poulson and @devinamatthews. :)

[poulson]

Hi Field,

Think of an $LDL^H$ factorization as a square-root free Cholesky factorization, where the pivots are stored in the diagonal matrix $D$. Each rank-one update of a right-looking / greedy / variant 3 algorithm are Hermitian and of the form $a_{2,1} \delta_{1,1} a_{2,1}^H$. In the blocked algorithm, they become $A_{2,1} D_{1, 1} A_{2,1}^H$, where $D_{1,1}$ is the diagonal matrix of pivots from the active diagonal block, $A_{1,1} = L_{1,1} D_{1,1} L_{1,1}^H$.

[fgvanzee]

@poulson I've never heard of a pivoted Cholesky factorization, but I get the idea!

[poulson]

@fgvanzee The diagonal entries you divide by are sometimes called pivots even if you didn't permute. Per the other question: diagonally-pivoted Cholesky is useful for computing a rank-revealing factorization of a Hermitian positive semi-definite matrix.

[fgvanzee]

@poulson Ah, thanks for that clarification. I'd never encountered this alternate instance/usage of "pivot." Good to know. (The President might say I am a fake linear algebraist. Although I think he would struggle to even pronounce "algebraist" correctly.)

[pauldj]

The D of an LDL' factorization is actually NOT diagonal.
It is block-diagonal, with blocks of size 1x1 or 2x2.

An LDL factorization is like a Cholesky factorization, but for a matrix that is symmetric but not necessarily positive definite. L is unit lower triangular and D is diagonal.

[poulson]

@pauldj Quasi-diagonal 'D" -- which I find useful to refer to as 'B', as in LBL -- is only needed for the general symmetric/Hermitian-indefinite case. It is provably not needed for symmetric quasi-definite matrices. Indeed, most Interior Point Methods have highly indefinite, highly ill-conditioned sparse systems that they factor with LDL with diagonal D.