[challenge]: Sector-crossing global updates for the triangular-lattice bond-SSH quantum spin liquid

Released by

Zihong Liu, Fudan University

Contact email

liuzihong@fudan.edu.cn

Method

Quantum Monte Carlo

Challenge issue

Difficulty: ★★

Background

The bond Su–Schrieffer–Heeger (SSH) model on the triangular lattice,

H = Σ_⟨ij⟩ [ −(t + g X̂_ij) B̂_ij + P̂²_ij/(2M) + (K/2) X̂²_ij ],   B̂_ij = Σ_σ (c†_iσ c_jσ + h.c.),

with optical phonons living on nearest-neighbour bonds, dimensionless coupling λ ≡ zg²/(KW)
(z = 6, W = 9t) and phonon frequency ω₀ = √(K/M), is sign-problem-free at arbitrary
filling: positivity is enforced by (Majorana) time-reversal symmetry (Wu and Zhang, PRB 71,
155115; Li, Jiang and Yao, PRL 117, 267002), not by particle–hole symmetry at half filling, so
doping does not reintroduce a sign.
Cai, Han, Li, Kivelson and Yao (PNAS 2025, arXiv:2408.04002) recently showed by projective DQMC at
half filling that it hosts a fully gapped Z₂ quantum spin liquid in an intermediate window — at λ = 2.16 the
QSL occupies 0.7 ≲ ω₀/t ≲ 1.2, bounded below by a staggered-VBS transition and above by a
superconducting one. This is the first QSL established in a non-engineered electron–phonon model.

The physics is settled; the sampling is not. The authors state plainly that these simulations are
"computationally more demanding than purely electronic models due to the longer auto-correlation
times," and that despite adopting "various strategies to reduce the autocorrelation time ... heavy
computational resources have still been required to achieve reliable results with high accuracy."

The obstruction is discrete, and saying so precisely matters because it dictates which algorithms
can possibly help. In the strong-coupling limit the same paper shows the low-energy manifold
collapses onto hard-core dimer coverings of the triangular lattice (⟨X_ij⟩ = 2g/K on covered
bonds, 0 elsewhere), governed by a quantum dimer model with a diagonal repulsion V and a
plaquette-flip kinetic term J; the QSL is the Moessner–Sondhi Z₂ liquid near J ≈ V.
Configuration space therefore fragments into near-degenerate dimer sectors whose connections require
plaquette flips or loop moves. Continuous-field dynamics — local Metropolis, and equally the
exact-force HMC that SmoQyDQMC.jl uses — equilibrates within a sector but crosses between sectors
with exponentially small probability. Note this is not the unbounded-force pathology of Beyl,
Goth and Assaad (PRB 97, 085144): that work identifies SSH-type actions as precisely the
singular-free class where HMC does work well.

Self-learning Monte Carlo (SLMC; Liu, Qi, Meng, Fu, PRB 95, 041101(R)) is the natural remedy, and
the corresponding problem for on-site (Holstein) coupling on bipartite lattices is already
solved: Chen, Xu, Liu, Batrouni, Scalettar and Meng (PRB 98, 041102(R)) fit a symmetry-enforced
local effective boson action to the DQMC weights, propose Wolff cluster moves in that effective
model, and accept them against the exact determinant weight — cutting autocorrelation times enough
to locate the CDW critical point on large lattices; the same machinery delivered the honeycomb
Holstein phase diagram (PRL 122, 077601).

Whether this transfers to a frustrated, non-bipartite lattice with bond-centred phonons is open, and
it is exactly where the difficulty lies. The nearest-neighbour bonds of the triangular lattice form
a kagome lattice, so the effective model to be learned lives on kagome sites × imaginary time
and is geometrically frustrated.

Research objective

Build a self-learning, sector-crossing global-update layer for the triangular-lattice bond-SSH
model, and use it to reproduce the QSL window of Fig. 1 of arXiv:2408.04002 at materially lower cost.

1. Learn. Fit a local, few-parameter, symmetry-enforced effective classical action
   S_eff({X_{ij,τ}}) to the DQMC weights sampled in and around the QSL window, i.e. a multivariate
   regression of −ln W on a symmetry-adapted local operator basis over the sampled phonon
   configurations {X_{ij,τ}}. Constraints: the triangular point group, lattice translations,
   imaginary-time translation, and X → −X where the bond parity permits. Keeping the model local
   and few-parameter is what allows couplings fitted at L = 6, 8 to be reused at L = 12, 16 —
   this transferability is the whole economic point. The quality of this fit is itself the first
   deliverable and is achievable independently of everything below; see Gate 0.
2. Move. Design global moves for S_eff that actually change dimer sector — loop/pocket moves
   in the dimer representation, plaquette-flip clusters, or Wolff-type clustering on a
   coarse-grained bond-occupation variable, whichever the fitted couplings support — and accept each
   against the exact determinant ratio, so the algorithm remains unbiased by construction.
3. Measure honestly. Quantify the gain against the strongest available baseline (the HMC +
   reflection + swap + radial updates already shipped in SmoQyDQMC.jl), not against naive local
   updates, and always in cost-normalised units.
4. Reproduce. Recover the Fig. 2 crossings that bound the QSL window.

Stretch: replace the hand-designed ansatz with a generative model of the phonon configurations,
using the fitted S_eff as the baseline to beat.

Verification plan

Gate 0 — the fit itself (cheap, standalone, and predictive). Before any update machinery is
written, sample configurations with the existing updates, evaluate −ln W exactly for each, and
regress it on the symmetry-adapted local basis. Report the reduced χ² and the out-of-sample
residual δ ≡ (−ln W) − S_eff. Success has two parts, and the second is the one that matters:

• reduced χ² of order unity, i.e. the residual scatter is consistent with the intrinsic Monte Carlo
  error on −ln W rather than with a misspecified basis; and
• the residual variance per site does not grow with volume, measured across L = 6, 8, 12 at
  fixed λ = 2.16, ω₀/t = 1.0. This is the decisive test, because the acceptance rate of any
  global move goes as exp(−Δδ): an intensive residual supports a finite acceptance rate at any
  volume, whereas an extensive one guarantees exponential decay and thereby proves that no local
  S_eff can work here — a clean negative result, obtained at a small fraction of the cost of
  discovering the same thing through failed cluster updates.

A controlled χ² with an intensive residual is a reportable result on its own: it says a local
effective phonon model exists for bond coupling on a frustrated lattice, which is precisely
what is unknown today. It also fixes which operators the basis needs, so it is the correct thing to
do first regardless of how far the rest gets.

Success gate — three further tests, all untameable, in order:

1. Unbiasedness (must pass first). SLMC is exact by construction, so any systematic offset is a
   bug in the acceptance ratio and cannot be tuned away. On L = 6 and L = 8 at λ = 2.16,
   ω₀/t ∈ {0.6, 1.0, 1.4}, every measured observable — total energy, kinetic and phonon energy,
   ⟨X²⟩, staggered-VBS structure factor S(Q), and the Q = 0 pair correlation — must agree with
   a long reference run using only the pre-existing updates, within 2σ of the combined error.
2. Efficiency. Report the integrated autocorrelation time τ_int of the two slowest observables
   (S_sVBS(Q) and the total phonon action) multiplied by wall-clock cost per sweep. Success =
   ≥ 5× reduction in cost-normalised τ_int at L = 12, λ = 2.16, ω₀/t = 1.0, at a βt
   large enough to resolve the Fig. 3(a) gaps, using effective-model couplings fitted at L = 8
   — the transfer must be demonstrated, not assumed.
3. Physics. Reproduce Fig. 2 of arXiv:2408.04002 at λ = 2.16 for L = 8, 12, 16: the R^c_SC
   crossing at ω₀ ≈ 1.2 and the sVBS Binder crossing at ω₀ ≈ 0.7, with error bars no larger than
   published, at ≤ 1/3 of the baseline cost. Both crossings are published numbers that cannot be
   tuned towards without the algorithm actually working.

Hope signal — continue; the concept is alive but the ansatz needs work. Gate 0 passes
(controlled χ², intensive residual) and sector-crossing moves are exact but accepted at only a few
percent, giving < 2× speedup; or the gain is clear at L = 8 but degrades by L = 16. Response:
enlarge the operator basis, or restrict moves to finite spatial patches with a fixed acceptance
floor rather than whole-lattice clusters.

Pivot signal — the local-effective-model route is dead; change approach. Either Gate 0 fails on
its second clause (the residual variance per site grows with L), or acceptance falls exponentially
in L (or β) for every sector-crossing move attempted — both mean the determinant weight and
the effective-model weight decorrelate as the volume grows, and no local S_eff captures the
fermionic non-locality here. Pivot to (a) spatially local, fixed-size cluster sector moves with a
guaranteed acceptance floor, or (b) a generative model trained directly on DQMC configurations.
Because Gate 0 delivers this verdict from fits alone, it should be checked before any update
machinery is written.

Why this may lead to research output

1. The open questions in arXiv:2408.04002 are statistics-limited, not understanding-limited.
   The QSL–SC exponent (measured η ≈ 1.45 ± 0.08 against the XY* value 1.49), the phase
   boundaries at larger λ, and the lightly-doped QSL that the authors propose as a high-Tc route
   are all bounded by autocorrelation time. Crucially, the sign-free property here is
   symmetry-protected and survives at any filling, so the lightly-doped QSL — which the authors
   name as ongoing work — is gated by autocorrelation time and nothing else. An algorithm that buys
   an order of magnitude opens a doping axis that is otherwise fully accessible in principle and
   unaffordable in practice.
2. Whether SLMC survives frustration is itself an unanswered methodological question. Every SLMC
   success in fermionic QMC to date is on bipartite lattices with site-centred coupling. A clean
   negative result here — with the exponential-acceptance mechanism identified and localized — is
   publishable and directly useful to anyone attempting the same transfer.
3. The learned model has an analytically known limit to check against. The strong-coupling
   quantum dimer model with its V and J (Eqs. 4–6 of arXiv:2408.04002; Moessner–Sondhi,
   PRL 86, 1881) predicts what the fitted couplings should become as λ → ∞, ω₀ → 0. This is a
   rare case where a learned effective model can be validated against theory rather than only
   against its own predictions — and a rare, concrete handle on what SLMC actually learns.

Suggested route and remarks

• Code. SmoQyDQMC.jl (SciPost Phys. Codebases 29; arXiv:2311.09395) supports triangular
  geometries via LatticeUtilities.jl and SSH-type bond coupling, and already ships hmc_update!,
  reflection_update!, swap_update! and radial_update! — so the baseline comes for free and the
  SLMC layer can be added as a new updater. Caveat: it is a finite-temperature code whereas
  arXiv:2408.04002 used projective DQMC, so βt must be large enough to resolve the Fig. 3(a) gaps;
  verify against those gaps before trusting any T = 0 claim. The successor SmoQyElPhQMC.jl
  (arXiv:2606.14425) is faster and worth evaluating.
• Warm up on the solved case before touching the hard one. Reproduce the square-lattice Holstein
  SLMC benchmark of PRB 98, 041102(R). It is small, well documented, and it validates the whole
  fit → propose → accept pipeline end to end.
• Then transfer the bipartite ansatz unmodified, and watch it fail. The failure mode — which
  correlators the fit misses, how the residual scales with L — is the actual signal for which
  operators the kagome-geometry ansatz needs. Budget for this diagnostic step rather than skipping
  straight to a hand-designed frustrated ansatz.
• Get Gate 0 out of the way before writing a single line of update code. It needs only
  configurations from the existing samplers, exact −ln W per configuration, and a linear
  least-squares solve; it is the cheapest item here and it decides whether the rest is worth
  attempting. Sweep the operator basis from minimal (on-site X², nearest-neighbour X_iX_j on the
  kagome geometry, imaginary-time gradient (X_τ − X_{τ+1})²) upward, and watch the reduced χ² and
  the per-site residual variance as terms are added — the point at which χ² stops improving names
  the physics the fit is missing.
• Insist on locality and few parameters. A large-basis fit that cannot be extrapolated from
  L = 8 to L = 16 defeats the purpose, however good its R².
• Always report τ_int in cost-normalised units. A cluster update 10× more expensive per sweep
  and 10× less autocorrelated is worth nothing.

References

1. X. Cai, Z. Han, Z.-X. Li, S. A. Kivelson, H. Yao, Quantum spin liquid from electron–phonon
   coupling, PNAS 122 (2025); arXiv:2408.04002.
2. C. Chen, X. Y. Xu, J. Liu, G. Batrouni, R. Scalettar, Z. Y. Meng, Symmetry-enforced self-learning
   Monte Carlo method applied to the Holstein model, Phys. Rev. B 98, 041102(R) (2018);
   arXiv:1802.06177.
3. J. Liu, Y. Qi, Z. Y. Meng, L. Fu, Self-learning Monte Carlo method, Phys. Rev. B 95,
   041101(R) (2017); arXiv:1610.03137.
4. C. Chen, X. Y. Xu, Z. Y. Meng, M. Hohenadler, Charge-density-wave transitions of Dirac fermions
   coupled to phonons, Phys. Rev. Lett. 122, 077601 (2019); arXiv:1809.07903.
5. S. Beyl, F. Goth, F. F. Assaad, Revisiting the hybrid quantum Monte Carlo method for Hubbard and
   electron-phonon models, Phys. Rev. B 97, 085144 (2018); arXiv:1708.03661.
6. C. Wu, S.-C. Zhang, Sufficient condition for absence of the sign problem in the fermionic
   quantum Monte Carlo algorithm, Phys. Rev. B 71, 155115 (2005).
7. Z.-X. Li, Y.-F. Jiang, H. Yao, Majorana-time-reversal symmetries: a fundamental principle for
   sign-problem-free quantum Monte Carlo simulations, Phys. Rev. Lett. 117, 267002 (2016).
8. R. Moessner, S. L. Sondhi, Resonating valence bond phase in the triangular lattice quantum dimer
   model, Phys. Rev. Lett. 86, 1881 (2001).
9. B. Cohen-Stead et al., SmoQyDQMC.jl: A flexible implementation of determinant quantum Monte Carlo
   for Hubbard and electron-phonon interactions, SciPost Phys. Codebases 29 (2024);
   arXiv:2311.09395.

Start with SmoQyDQMC.jl's existing hmc_update!, reflection_update!, swap_update!, and radial_update! implementations, then reproduce the square-lattice Holstein SLMC benchmark before the triangular SSH case. Run Gate 0 at L = 6, 8, and 12 using the stated parameters and report reduced χ² and residual variance per site. Done requires an unbiased sector-crossing updater, a cost-normalised speedup, and the Fig. 2 crossings.