wai-quantum 0.3.40: rectangle space groups, a sector referee on its own basis, and a sector larger than the network #194

Merged

dcharlot merged 3 commits from wai-quantum-nqs-rectangle into main 2026-10-09 03:48:28 -04:00

Conversation 0 Commits 3 Files changed 9 +370 -81

dcharlot commented 2026-10-09 02:41:26 -04:00

Owner

Copy link

What

wai-quantum 0.3.40. A symmetry sector larger than the network, held to its exact answer.

• Symmetry::rectangle(lx, ly, kx_pi, ky_pi, irrep): the translations of an lx × ly torus, each combined with the identity, the reflections x → −x and y → −y, or both (the half turn). It projects onto one of their four one-dimensional representations (RectIrrep::{A1, A2, B1, B2}), at momentum 0 or π along each direction.
• SpinModel::exact_ground_energy_in now works on the sector's own basis: one state per orbit of the group, Σ_g χ(g)|g·r⟩ normalised for each least configuration r that no element of character −1 fixes. It reaches 24 spins.
  • The projector referee it replaces (Lanczos on P H P) stays in the tests as an independent cross-check.
• The CLI and nqs_ground take --reflections a1|a2|b1|b2. A projected run of up to 24 spins is held to its sector alone. Previously, past 20 spins, it first computed a DMRG reference and then discarded it: 136 s on the 6 × 4 torus, for an answer 3 % off.

Checked

• The two referees agree to 10⁻⁹ in all 24 sectors of the 4 × 3 J₁–J₂ torus: four representations, both momenta along x, and no flip, an even flip or an odd flip.
• They also agree on a 10-site ring, and on an Ising ring in a transverse field, where the magnetisation is not conserved.
• With the identity alone, the orbit referee is the plain diagonalisation. With the flip alone, one half holds the ground state and the other lies above it.
• Every rectangle group on 4 × 3 and 6 × 4 closes and multiplies its characters.

Measured

Periodic 6 × 4 J₁–J₂ at J₂ = J₁/2: the full 24-spin diagonalisation gives −12.5405984488, and the orbit referee places it in the sector even under the half turn and both reflections, at momentum zero, with even parity.

That sector holds 15 578 configurations up to symmetry, counted by character sums; the same count gives 107 for the 4 × 4 sector of 0.3.39. A complex α = 1 network has 624 parameters.

Results after 600 steps of 1000 samples, with energies over every configuration:

projection, even parity	guide	relative error
translations	0.1	1.06·10⁻³
rectangle group, A₁	none	8.32·10⁻³
rectangle group, A₁	0.1	5.57·10⁻⁴

On a sector larger than the network, the guide is worth a factor of fifteen and the reflections a factor of two. Unlike the small 4 × 4 sector, this one is not reached to rounding: what remains is the network's.

On the 6 × 6 open Heisenberg lattice (unprojected, α = 2, 800 steps) the guide doubled the error, from 3.1·10⁻³ to 6.5·10⁻³. Unprojected networks keep sampling |ψ|², and the docs say why.

Gates

• cargo test --release --features full --lib quantum_nqs: 22 passed.
• cargo test --release --features full --test bit_identity: 22 passed, digests unchanged.
• cargo test --locked --features full --lib --tests --no-run (debug) compiles. Clippy is clean in the module and the example.
• --locked resolves in all five lockfiles that pin wai-quantum.
• scripts/check-comparative-prose.sh and scripts/check-no-vanity-metrics.sh are clean.

🤖 Generated with Claude Code

## What wai-quantum 0.3.40. A symmetry sector larger than the network, held to its exact answer. - **`Symmetry::rectangle(lx, ly, kx_pi, ky_pi, irrep)`**: the translations of an `lx × ly` torus, each combined with the identity, the reflections `x → −x` and `y → −y`, or both (the half turn). It projects onto one of their four one-dimensional representations (`RectIrrep::{A1, A2, B1, B2}`), at momentum 0 or π along each direction. - **`SpinModel::exact_ground_energy_in` now works on the sector's own basis**: one state per orbit of the group, `Σ_g χ(g)|g·r⟩` normalised for each least configuration `r` that no element of character −1 fixes. It reaches 24 spins. - The projector referee it replaces (Lanczos on `P H P`) stays in the tests as an independent cross-check. - **The CLI and `nqs_ground`** take `--reflections a1|a2|b1|b2`. A projected run of up to 24 spins is held to its sector alone. Previously, past 20 spins, it first computed a DMRG reference and then discarded it: 136 s on the 6 × 4 torus, for an answer 3 % off. ## Checked - The two referees agree to `10⁻⁹` in all 24 sectors of the 4 × 3 J₁–J₂ torus: four representations, both momenta along x, and no flip, an even flip or an odd flip. - They also agree on a 10-site ring, and on an Ising ring in a transverse field, where the magnetisation is not conserved. - With the identity alone, the orbit referee is the plain diagonalisation. With the flip alone, one half holds the ground state and the other lies above it. - Every rectangle group on 4 × 3 and 6 × 4 closes and multiplies its characters. ## Measured Periodic 6 × 4 J₁–J₂ at `J₂ = J₁/2`: the full 24-spin diagonalisation gives `−12.5405984488`, and the orbit referee places it in the sector even under the half turn and both reflections, at momentum zero, with even parity. That sector holds 15 578 configurations up to symmetry, counted by character sums; the same count gives 107 for the 4 × 4 sector of 0.3.39. A complex `α = 1` network has 624 parameters. Results after 600 steps of 1000 samples, with energies over every configuration: | projection, even parity | guide | relative error | |---|---|---| | translations | 0.1 | 1.06·10⁻³ | | rectangle group, A₁ | none | 8.32·10⁻³ | | rectangle group, A₁ | 0.1 | 5.57·10⁻⁴ | On a sector larger than the network, the guide is worth a factor of fifteen and the reflections a factor of two. Unlike the small 4 × 4 sector, this one is not reached to rounding: what remains is the network's. On the 6 × 6 open Heisenberg lattice (unprojected, `α = 2`, 800 steps) the guide doubled the error, from 3.1·10⁻³ to 6.5·10⁻³. Unprojected networks keep sampling `|ψ|²`, and the docs say why. ## Gates - `cargo test --release --features full --lib quantum_nqs`: 22 passed. - `cargo test --release --features full --test bit_identity`: 22 passed, digests unchanged. - `cargo test --locked --features full --lib --tests --no-run` (debug) compiles. Clippy is clean in the module and the example. - `--locked` resolves in all five lockfiles that pin wai-quantum. - `scripts/check-comparative-prose.sh` and `scripts/check-no-vanity-metrics.sh` are clean. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

dcharlot added 3 commits 2026-10-09 02:41:26 -04:00

wai-quantum: rectangle space groups, and the sector referee on the sector's own basis ... 3f7b7f347c

Symmetry::rectangle projects onto the translations of an lx × ly torus
with its half turn and axis reflections, in one of their four
one-dimensional representations (RectIrrep), at momentum 0 or π along
each direction.

SpinModel::exact_ground_energy_in now works on one state per orbit of
the group, Σ_g χ(g)|g·r⟩ normalised for each least configuration r
that no element of character −1 fixes, up to 24 spins. The projector
referee it replaces is kept in the tests, and the two agree to 1e-9 in
all 24 sectors of the 4 × 3 J1-J2 torus, on a 10-site ring, and on an
Ising ring in a transverse field. The identity alone gives the plain
diagonalisation; the flip alone splits it in two.

The CLI and nqs_ground take --reflections a1|a2|b1|b2, and hold a
projected network to its sector up to 24 spins.

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

nqs_ground: a projected network up to 24 spins is held to its sector alone ... e7d4380eba

The full reference was computed first and then replaced by the
sector's; past 20 spins that first step was a DMRG run (136 s on the
6 × 4 torus, and 3 % off the exact energy) whose answer was thrown
away.

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

wai-quantum 0.3.40: a sector larger than the network, held to its exact answer ... 890ed855c6

On the periodic 6 × 4 J1-J2 lattice at J2 = J1/2 the orbit referee
places the ground state (−12.54059845, equal to the full 24-spin
diagonalisation) in the sector even under the half turn and both
reflections at momentum zero, with even parity: 15 578 configurations
up to symmetry, against a complex α = 1 network's 624 parameters.
After 600 steps, energies over every configuration: translations
guided 1.06e-3; rectangle group unguided 8.32e-3; rectangle group
guided 5.57e-4.

On the 6 × 6 open Heisenberg lattice (unprojected, α = 2, 800 steps)
the guide doubled the error, so unprojected networks keep sampling
|ψ|²; the docs say so.

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

dcharlot merged commit 412a08c5d6 into main 2026-10-09 03:48:28 -04:00

dcharlot referenced this pull request from a commit 2026-10-09 03:48:28 -04:00

Merge pull request 'wai-quantum 0.3.40: rectangle space groups, a sector referee on its own basis, and a sector larger than the network' (#194) from wai-quantum-nqs-rectangle 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!194

No description provided.