feat: BUG integrator - #468
Conversation
21e8388 to
f8a54a2
Compare
Codecov Report❌ Patch coverage is
... and 1 file with indirect coverage changes 🚀 New features to boost your workflow:
|
35d07d8 to
6b058d1
Compare
Renormalization after each step is now controlled by an explicit `normalize` kwarg on `timestep`/`time_evolve` (default `false`) instead of being tied to `imaginary_evolution`. By default the norm is preserved, so it retains useful information (accumulated truncation error in real time, decaying weight in imaginary time); pass `normalize = true` to renormalize, e.g. for ground-state or thermal-state search. Threaded through the driver and both TDVP/TDVP2 integrators; test norm/imaginary-time expectations updated accordingly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add the `BUG` time-evolution algorithm: a symmetric second-order Basis-Update & Galerkin integrator for finite MPS. Unlike `TDVP` it has no backward-in-time substep (stable for imaginary-time evolution); passing a truncating `trscheme` enables rank-adaptivity, growing and shrinking the bond dimension to track the entanglement. Wires up the export, include, docs (`@docs` entry, changelog, Ceruti & Lubich 2022 reference). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add tests for the sequential `BUG` integrator: energy conservation, eigenstate phase, agreement with TDVP, second-order convergence vs dense propagation, imaginary-time cooling, LazySum/time-dependent operators, rank-adaptivity, and symmetric tensors (U(1)/Z2/SU(2), incl. rank-adaptive-symmetric). Plus unit tests for the `_bug_augment_*` basis-augmentation helpers. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Remove unused `tolgauge`/`gaugemaxiter` fields from the `BUG` struct and constructor: they are read only for `InfiniteMPS` gauge reconstruction in TDVP, and BUG is finite-only, so they were silently ignored. - Clarify the `BUG` docstring (single-site, symmetric second-order, finite MPS) and correct the `finalize` signature docs (`finalize(t, …)`, not `iter`) in BUG, TDVP and TDVP2. - Document the `normalize` behavior change under changelog "Changed": it is now an explicit kwarg (default false) decoupled from `imaginary_evolution`; imaginary-time evolution no longer renormalizes unless `normalize = true`. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Jutho <Jutho@users.noreply.github.com> small docstring update
The BUG work added a lot of tests that were written to debug the implementation as it was developed. Now that it is settled, keep BUG tested at the level TDVP/TDVP2 were: parametrized into the existing shared testsets, plus one testset for its distinguishing feature. Removed: - `test/algorithms/bug_augment.jl` — unit tests for the inlined basis-augmentation step (isometry, old-first per sector, rank growth) - `second-order convergence (BUG)` — log-log temporal-order fit against a dense reference - `BUG agrees with TDVP` — subsumed by the `Finite BUG` testset, which runs the same system and asserts energy plus the eigenstate phase - four of the five `BUG with truncation` testsets, and the Z2/SU(2)/ second-sector/imaginary-time repeats of the symmetric-tensor block Retained beyond the shared testsets: one truncated-accuracy testset against a dense reference, one imaginary-time testset (BUG has no backward substep), and one U(1) testset covering energy conservation, eigenstate phase and boundary-charge preservation — symmetric coverage the finite integrators did not have before. Every `BUG` in the file is now constructed with a `trunc`: uncapped, the augmentation doubles the bond dimension on every half-sweep with nothing cutting it back, which is neither realistic nor cheap. Added bond-dimension bounds where the cap should bite. 84 tests pass; the file runs in 573s, down from 879s. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
6b058d1 to
d7251aa
Compare
|
Alright, after a round of review and additional discussions, (thanks @Jutho for the insights!), I think I've now addressed some of the issues that were here and added some more convenience parts to the implementation. The first major thing is that I now no longer truncate the augmented basis, which is indeed incorrect. From the small tests I ran this definitely did improve the accuracy, so this was definitely a needed step. To consider where the truncation actually has to happen, I agree that the original formulation of the algorithm would mean to first evolve left-to-right, and then truncate right-to-left. Since I kind of wanted to keep some of the left-right symmetry here, I made a small alteration that I think is both more accurate as well as more efficient, but I would be happy to get a second opinion on this. The train of thought goes something like this: imagine that on the reverse pass, instead of first truncating right to left, then regauging to the right, and then starting the new evolution run from right-to-left again, I truncate site L-1, and then immediately start the evolution, as if no truncation happened I also added some timeroutput support to be able to track where the time is actually spent (almost entirely in the exponentiator, of course), but this can be useful for debugging. |
2 participants
This PR adds the BUG integrator as an alternative to TDVP. It's a consolidation of part of #454.
The main purpose is to have a stable integrator to be able to handle (large) imaginary or complex time steps in a way that is more stable than standard TDVP since it does not involve a backwards integration.
@Jutho I would be very happy to hear your opinions on this, I'm still not completely sure I've done everything correctly and many parts still confuse me, so a second set of eyes would be greatly appreciated. 🙃