wai-quantum 0.3.38: neural quantum states held in their symmetry sector #180

Merged

dcharlot merged 2 commits from wai-quantum-nqs-symmetry into main 2026-10-08 10:24:11 -04:00

Conversation 0 Commits 2 Files changed 11 +506 -50

dcharlot commented 2026-10-07 09:44:02 -04:00

Owner

Copy link

What this adds

Symmetry projects a neural quantum state onto a sector, Ψ(s) = Σ_g χ(g) ψ(g·s):

• over a lattice's translations (Symmetry::ring, Symmetry::torus), at momentum 0 or π along each direction;
• optionally over the flip of every spin (with_spin_flip(odd)), with even or odd parity.

Ratios and log-derivatives become weighted averages over the transformed configurations. They are held in a per-configuration cache that accepted Metropolis moves keep exact. A configuration the symmetry forbids, whose projected sum is below the rounding of its own terms, is treated as exactly zero, since its log-derivative would be noise over noise.

Biases now start random, like the weights. With zero biases a network is exactly even under the global flip, and training keeps it even. That is what held 10-site rings away from their odd singlet ground state in 0.3.37 (documented there).

SR falls back to an eigendecomposition (eigenvalues below 10⁻¹² of the largest dropped) when Cholesky loses a pivot. A projected network started near even has log-derivatives up to 10⁸ and a metric diagonal of 10¹². Against an absolute shift of 10⁻³ the factorisation produced NaN. Runs that factor cleanly keep their bits.

Checked

• The projection equals the explicit sum over the group, ratios and log-derivatives included. A translation multiplies Ψ by its character and a global flip by its parity. The cache stays equal to a fresh one through accepted moves.
• Rings in their sector, α = 1, every configuration enumerated:

ring (sector)	relative error
6 sites, Heisenberg (π, odd), 150 steps	< 10⁻⁸
10 sites, Heisenberg / Majumdar–Ghosh / J₁–J₂ 0.3 (π), 300 steps	7·10⁻¹⁴ / 3·10⁻¹⁵ / 3·10⁻¹⁴
12 sites, the same three (0, even), 400 steps	7·10⁻¹⁶ / 4·10⁻¹⁶ / 1·10⁻¹⁵

• The frustrated 4×4 torus at J₂ = 0.5 (exact −8.45792335), a complex α = 1 network projected onto momentum 0 and even parity, 1000 samples a step:
  • 1.7 % after 600 steps;
  • 7.7·10⁻³ ± 1.2·10⁻³ after 2000 steps;
  • unprojected, every network stayed about 3 % above it at 600 steps.
• The NQS pin covers a projected machine running through the fallback. All 22 pins pass natively and under wasm32-wasip2.
• nqs --project [--kpi] [--odd] is byte-identical under WASI.
• The full suite passes in debug (449 library tests), along with clippy and the prose checks.

Boundaries

• Translations and the global flip only: rotations and reflections are not in the group, and characters are ±1.
• Projection multiplies every amplitude's cost by the size of the group.

🤖 Generated with Claude Code

## What this adds **`Symmetry`** projects a neural quantum state onto a sector, `Ψ(s) = Σ_g χ(g) ψ(g·s)`: - over a lattice's translations (`Symmetry::ring`, `Symmetry::torus`), at momentum 0 or π along each direction; - optionally over the flip of every spin (`with_spin_flip(odd)`), with even or odd parity. Ratios and log-derivatives become weighted averages over the transformed configurations. They are held in a per-configuration cache that accepted Metropolis moves keep exact. A configuration the symmetry forbids, whose projected sum is below the rounding of its own terms, is treated as exactly zero, since its log-derivative would be noise over noise. **Biases now start random**, like the weights. With zero biases a network is exactly even under the global flip, and training keeps it even. That is what held 10-site rings away from their odd singlet ground state in 0.3.37 (documented there). **SR falls back to an eigendecomposition** (eigenvalues below `10⁻¹²` of the largest dropped) when Cholesky loses a pivot. A projected network started near even has log-derivatives up to `10⁸` and a metric diagonal of `10¹²`. Against an absolute shift of `10⁻³` the factorisation produced NaN. Runs that factor cleanly keep their bits. ## Checked - The projection equals the explicit sum over the group, ratios and log-derivatives included. A translation multiplies `Ψ` by its character and a global flip by its parity. The cache stays equal to a fresh one through accepted moves. - Rings in their sector, `α = 1`, every configuration enumerated: | ring (sector) | relative error | |---|---| | 6 sites, Heisenberg (π, odd), 150 steps | `< 10⁻⁸` | | 10 sites, Heisenberg / Majumdar–Ghosh / J₁–J₂ 0.3 (π), 300 steps | 7·10⁻¹⁴ / 3·10⁻¹⁵ / 3·10⁻¹⁴ | | 12 sites, the same three (0, even), 400 steps | 7·10⁻¹⁶ / 4·10⁻¹⁶ / 1·10⁻¹⁵ | - The frustrated 4×4 torus at J₂ = 0.5 (exact −8.45792335), a complex `α = 1` network projected onto momentum 0 and even parity, 1000 samples a step: - 1.7 % after 600 steps; - `7.7·10⁻³ ± 1.2·10⁻³` after 2000 steps; - unprojected, every network stayed about 3 % above it at 600 steps. - The NQS pin covers a projected machine running through the fallback. All 22 pins pass natively and under `wasm32-wasip2`. - `nqs --project [--kpi] [--odd]` is byte-identical under WASI. - The full suite passes in debug (449 library tests), along with clippy and the prose checks. ## Boundaries - Translations and the global flip only: rotations and reflections are not in the group, and characters are ±1. - Projection multiplies every amplitude's cost by the size of the group. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

dcharlot added 1 commit 2026-10-07 09:44:02 -04:00

wai-quantum 0.3.38: neural quantum states held in their symmetry sector ... e59cea7814

Symmetry projects a network, Psi(s) = sum_g chi(g) psi(g.s), over a lattice's
translations (ring or torus, momentum 0 or pi along each direction) and
optionally the global spin flip (even or odd parity).
- Ratios and log-derivatives become weighted averages over the transformed
  configurations, held in a cache that Metropolis moves keep exact.
- A configuration the symmetry forbids (its sum below the rounding of its
  terms) is treated as zero.

Biases now start random as well as weights. With zero biases the network is
exactly even under the global flip, and training keeps it even, which is
what trapped 10-site rings in 0.3.37.

SR falls back to an eigendecomposition (cutoff 1e-12) when Cholesky loses a
pivot. A projected network started near even drives its metric to 1e12, and
the factorisation produced NaN there.

Rings in their sector reach exact energies to rounding: 10 sites at pi,
7e-14; 12 sites at 0, even, 1e-15. The frustrated 4x4 torus at J2 = 0.5
comes within 7.7e-3 (2000 steps; 3.6% unprojected).

The NQS pin covers a projected machine through the fallback; all 22 pins
pass natively and under wasm32-wasip2. The nqs CLI gains --project, --kpi
and --odd, byte-identical under WASI.

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

dcharlot added 1 commit 2026-10-07 14:28:44 -04:00

wai-pack: lock wai-quantum 0.3.38 ... 0ffdd67bd3

wai-pack reaches wai-quantum through wai-rs, so its own Cargo.lock
pins the version. It still said 0.3.36, which --locked refuses; the
workflow only runs when wai-rs or wai-pack changes, so the stale
entry went unseen until now.

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

dcharlot merged commit fc11ff7307 into main 2026-10-08 10:24:11 -04:00

dcharlot referenced this pull request from a commit 2026-10-08 10:24:11 -04:00

Merge pull request 'wai-quantum 0.3.38: neural quantum states held in their symmetry sector' (#180) from wai-quantum-nqs-symmetry into main

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!180

No description provided.