wai-quantum 0.3.37: complex neural quantum states, and referees that cannot stop on an excited state #176

Merged

dcharlot merged 3 commits from wai-quantum-nqs-complex into main 2026-10-07 09:13:46 -04:00

Conversation 0 Commits 3 Files changed 13 +614 -248

dcharlot commented 2026-10-06 17:16:10 -04:00

Owner

Copy link

Fix: the referees could stop on an excited state (0.3.36)

0.3.36 made Lanczos stop a restart as soon as its lowest Ritz pair converged. From DMRG's warm starts (the current state) that is safe. From the referees' cold starts, the lowest Ritz pair can be an excited state that converges first: exact diagonalisation of the 8-site Majumdar–Ghosh ring returned −2.595 instead of −3. A variational energy then came out below the referee, which is how it surfaced.

Early stopping is now opt-in and used by the DMRG sweeps only. Exact diagonalisation, full CI and the spin-model referee build every Krylov vector again. The DMRG and NQS pins were re-taken.

New: complex neural quantum states

• Rbm<T>, Vmc<T> and stochastic reconfiguration are generic over Amplitude, which is real (f64) or complex. The real path keeps its exact bits (the NQS pin did not move under the refactor).
• The complex path adds exp, ln cosh and tanh, built on the crate's reproducible exp, ln, sin/cos and atan2. Reconfiguration becomes the Hermitian step (S = ⟨O*O⟩ − …, F = ⟨O*E⟩ − …), still solved in whichever of parameter or sample space is smaller, with a Hermitian Cholesky.
• SpinModel::j1j2, chain_next_nearest, square_diagonals.

Checked

• A complex network with real parameters is the real one, to 10⁻¹³. Its complex ratios equal amplitude ratios, holomorphic derivatives agree with real-step differences, and both spaces take the same complex step.
• Exact diagonalisation gives Majumdar–Ghosh rings exactly −3n/8, at 8 and 12 sites.
• A sign nobody supplied. On 8-site rings (Heisenberg and Majumdar–Ghosh, no sign rule given), a real network stays more than 10 % off; a complex one comes within 2·10⁻³. The CLI solves the 8-site Majumdar–Ghosh ring to 3·10⁻¹⁵, with variance 5·10⁻¹⁷.
• 12-site rings, 400 steps: Majumdar–Ghosh 2·10⁻⁶, J₁–J₂ (0.3) 3.5·10⁻⁴, Heisenberg 2·10⁻³.
• The NQS pin now includes a complex machine. All 22 pins pass natively and under wasm32-wasip2, and nqs --complex is byte-identical under WASI.
• The full suite passes in debug (446 library tests).

Measured, and documented

• 10-site rings settle about 9 % high, on an eigenstate with zero variance. (Corrected in this PR: the first version blamed momentum, which was wrong.) The cause is spin-flip parity. With zero biases the network is exactly even under a global spin flip, and training keeps it even. A singlet is odd when half the ring is odd, so on 10-site rings the ground state is out of reach. Drawn biases reach it: 3·10⁻⁴, 1·10⁻⁶ and 1.4·10⁻⁵ in 400 steps. On 12-site rings, where the ground state is even, the even start does better. Choosing the sector explicitly is the next PR.
• The 4×4 torus at J₂ = 0.5: all three networks at α = 1 and 600 steps stay about 3 % above exact (−8.45792335). Recorded as a boundary.

Also

• nqs_ground now declares required-features (a featureless --examples build failed on it since 0.3.36), and a stray scratch example is removed.

CI note

The release build of staged_resources, the one wall-clock step left over from #164's replication, now passes at low load (0.78 s). Every macOS step has now passed on #164's head.

🤖 Generated with Claude Code

## Fix: the referees could stop on an excited state (0.3.36) 0.3.36 made Lanczos stop a restart as soon as its lowest Ritz pair converged. From DMRG's warm starts (the current state) that is safe. From the referees' cold starts, the lowest Ritz pair can be an excited state that converges first: exact diagonalisation of the 8-site Majumdar–Ghosh ring returned **−2.595 instead of −3**. A variational energy then came out below the referee, which is how it surfaced. Early stopping is now opt-in and used by the DMRG sweeps only. Exact diagonalisation, full CI and the spin-model referee build every Krylov vector again. The DMRG and NQS pins were re-taken. ## New: complex neural quantum states - `Rbm<T>`, `Vmc<T>` and stochastic reconfiguration are generic over `Amplitude`, which is real (`f64`) or complex. The real path keeps its exact bits (the NQS pin did not move under the refactor). - The complex path adds exp, ln cosh and tanh, built on the crate's reproducible exp, ln, sin/cos and atan2. Reconfiguration becomes the Hermitian step (`S = ⟨O*O⟩ − …`, `F = ⟨O*E⟩ − …`), still solved in whichever of parameter or sample space is smaller, with a Hermitian Cholesky. - `SpinModel::j1j2`, `chain_next_nearest`, `square_diagonals`. ### Checked - A complex network with real parameters is the real one, to `10⁻¹³`. Its complex ratios equal amplitude ratios, holomorphic derivatives agree with real-step differences, and both spaces take the same complex step. - Exact diagonalisation gives Majumdar–Ghosh rings exactly `−3n/8`, at 8 and 12 sites. - **A sign nobody supplied.** On 8-site rings (Heisenberg and Majumdar–Ghosh, no sign rule given), a real network stays more than 10 % off; a complex one comes within `2·10⁻³`. The CLI solves the 8-site Majumdar–Ghosh ring to `3·10⁻¹⁵`, with variance `5·10⁻¹⁷`. - 12-site rings, 400 steps: Majumdar–Ghosh `2·10⁻⁶`, J₁–J₂ (0.3) `3.5·10⁻⁴`, Heisenberg `2·10⁻³`. - The NQS pin now includes a complex machine. All 22 pins pass natively and under `wasm32-wasip2`, and `nqs --complex` is byte-identical under WASI. - The full suite passes in debug (446 library tests). ### Measured, and documented - **10-site rings** settle about 9 % high, on an eigenstate with zero variance. *(Corrected in this PR: the first version blamed momentum, which was wrong.)* The cause is **spin-flip parity**. With zero biases the network is exactly even under a global spin flip, and training keeps it even. A singlet is odd when half the ring is odd, so on 10-site rings the ground state is out of reach. Drawn biases reach it: `3·10⁻⁴`, `1·10⁻⁶` and `1.4·10⁻⁵` in 400 steps. On 12-site rings, where the ground state is even, the even start does better. Choosing the sector explicitly is the next PR. - **The 4×4 torus at J₂ = 0.5**: all three networks at α = 1 and 600 steps stay about 3 % above exact (−8.45792335). Recorded as a boundary. ## Also - `nqs_ground` now declares `required-features` (a featureless `--examples` build failed on it since 0.3.36), and a stray scratch example is removed. ## CI note The release build of `staged_resources`, the one wall-clock step left over from #164's replication, now passes at low load (0.78 s). Every macOS step has now passed on #164's head. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

dcharlot added 1 commit 2026-10-06 17:16:10 -04:00

wai-quantum 0.3.37: complex neural quantum states, and referees that cannot stop on an excited state ... 5cd36795cd

Fix. In 0.3.36, Lanczos stopped a restart as soon as its lowest Ritz pair
converged. That is safe from DMRG's warm starts, but from the referees' cold
ones that pair can be an excited state. Exact diagonalisation of the 8-site
Majumdar-Ghosh ring returned -2.595 instead of -3, and a variational energy
came out below it. Early stopping is now opt-in, used by the DMRG sweeps only;
exact diagonalisation, full CI and the spin-model referee build every Krylov
vector. The DMRG and NQS pins were re-taken.

New: the RBM, VMC and stochastic reconfiguration are generic over real and
complex parameters, and the real path keeps its bits.
- The complex path adds exp, ln cosh and tanh, the Hermitian reconfiguration
  step in either space, and a Hermitian Cholesky.
- New lattices: J1-J2 models, plus next-nearest bonds on chains and diagonal
  bonds on square lattices.
- A complex network learns sign structures nobody supplied. On 8-site rings it
  gets within 2e-3 where a real one stays >10% off, and the CLI solves the
  8-site Majumdar-Ghosh ring to rounding. On 12-site rings: Majumdar-Ghosh
  2e-6, J1-J2 (0.3) 3.5e-4.
- Measured and documented:
  - 10-site rings settle on a momentum-zero excited eigenstate, because the
    ground state has momentum pi.
  - At J2 = 0.5 on the 4x4 torus, alpha = 1 stays about 3% above exact.

The NQS pin now covers a complex machine. All 22 pins pass natively and
under wasm32-wasip2. The nqs CLI gains --complex, --model j1j2 and
--periodic, byte-identical under WASI.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot added 1 commit 2026-10-06 17:27:37 -04:00

wai-quantum 0.3.37: the 10-site trap is spin-flip parity, not momentum; drop a stray probe ... defa8a8a94

With zero biases the network is exactly even under a global spin flip, and
training keeps it even. A singlet is odd when half the ring is odd, so on
10-site rings the ground state was out of reach. Drawn biases reach it, and
the docs now say so, with both starts measured. The earlier text blamed
momentum, which is wrong.

Also removes examples/nqs_probe.rs, a scratch probe that the squash swept in.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot added 1 commit 2026-10-06 17:28:21 -04:00

wai-quantum 0.3.37: nqs_ground declares its feature ... a6636aa3dc

Without it a build of every example with default features failed on it
(since 0.3.36).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot merged commit ab9650d521 into main 2026-10-07 09:13:46 -04:00

dcharlot referenced this pull request from a commit 2026-10-07 09:13:48 -04:00

Merge pull request 'wai-quantum 0.3.37: complex neural quantum states, and referees that cannot stop on an excited state' (#176) from wai-quantum-nqs-complex into main

dcharlot referenced this pull request from a commit 2026-10-07 14:29:11 -04:00

wai-pack: the lock still named wai-quantum 0.3.36

dcharlot referenced this pull request 2026-10-07 14:29:13 -04:00

[obsolete] wai-pack: the lock still named wai-quantum 0.3.36 #182

dcharlot referenced this pull request from a commit 2026-10-08 01:50:20 -04:00

Merge main (#175 merged as ade3f855, #176) into wai/strict-ed25519

dcharlot referenced this pull request from a commit 2026-10-08 01:50:20 -04:00

wai-sig: the strict Ed25519 rule as its own crate, and wai-quantum on it

dcharlot referenced this pull request from a commit 2026-10-08 09:58:10 -04:00

Merge origin/main (HTTP playlists #175, wai-quantum 0.3.37 #176) into wai/w5-4-measure-web

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

Transaction-Science/open-standards!176

No description provided.