Weak inter-module coupling and the entanglement/trainability tradeoff in modular quantum circuits

Independent research project. Reproducible numerical evidence that, in modular variational quantum circuits and quantum reservoirs, a weak continuous inter-module coupling g optimizes the tradeoff between entanglement (a quantum resource) and trainability (gradient variance / barren plateaus).

One-sentence claim

In a block-structured circuit (strong intra-block entanglement, tunable inter-block ZZ coupling of continuous angle g), there is a weak-coupling window that keeps roughly half the gradient variance while ~tripling inter-block entanglement — and this window is set by circuit depth, not size: g*(L) ≈ 1.2/√L in the deep regime (N=10, 8 depths, fitted exponent −0.54 ± 0.07, anchors at 10 seeds; pre-registered prediction at L=160 confirmed: predicted 0.098, measured 0.096 ± 0.004; replicated at N=12, exponent −0.59), consistent with a lightcone-percolation picture where integrated inter-block leakage L·g² drives the transition. At fixed depth, g* does not decrease with system size N. The relative advantage of the window over the monolith grows with system size (×1.1 at N=4 → ×2.6 at N=12), suggesting it becomes decisive precisely in the classically-hard regime.

What is here

File	What it does
barren_plateau_v2.py	Monolith vs modular: gradient-variance collapse (slope −0.70 vs −0.01), ratio ×1070 ± 167 at N=12 (5 seeds, see hardening_suite.py)
sweet_spot.py	At fixed N=10: entanglement and trainability vs g — the tradeoff curves cross
scaling_fenetre.py + scaling_extend.py	g*(N) for N=4–14, 5 seeds; the optimum stays at weak coupling at every size (no drift to the monolith)
depth_test.py + depth_test_160.py	Main result: g*(L) at fixed N=10, L=6–160 — crossover ~1/L → 1.2/√L; pre-registered L=160 prediction confirmed
scaling_fenetre_gpu.py	GPU/CPU port (cupy/numpy) to push N=16–28 on free Kaggle/Colab; Rényi-2 entanglement for speed
performance_sweetspot.py	Honest negative result: full variational training at N=6 shows NO performance optimum at weak coupling — expected, since barren plateaus only bite at N≳10–12 (see limitations)
hardening_suite.py	Pre-registered robustness battery: barren ratio on 5 seeds; intermediate depths L=28/56/120 + seeds 6–10 on anchors; metric-weight sensitivity; von Neumann vs Rényi-2 bias; gradient-position invariance
overnight_suite.py	Pre-registered generality battery: g*(N) at fixed depth (depth-not-size test + cut-parity effect), block-size m=2/3/4, ring vs chain, depth-law replication at N=12, N=16
make_fig_barren.py, make_fig_depth.py, make_figs_hardened.py	Regenerate the figures from the JSON results
results_*.json, *_results.json, resultats_*.png	Raw numbers and figures

Reproduce

pip install numpy matplotlib
python scaling_fenetre.py        # ~8 min CPU, produces resultats_scaling_fenetre.png

GPU (Kaggle T4, free): set accelerator to GPU, pip install cupy-cuda12x, run scaling_fenetre_gpu.py (reaches N≈28 in complex128).

Method (honest scope)

• State-vector simulation, blocks of m=2 qubits, intra-block ZZ angle α=π/2 (strong), inter-block ZZ angle g (tunable, chain topology), L=2N layers of RY+RZ.
• Trainability = variance of the parameter-shift gradient of ⟨Z₀⟩ over random parameters.
• Entanglement = half-cut von Neumann entropy (scaling_fenetre.py) or Rényi-2 purity (GPU version, for speed).
• Sweet-spot score = (entanglement fraction) × (gradient-variance fraction); g* = parabolic argmax.

Robustness (overnight pre-registered test battery, June 11–12)

Every test below had its prediction written down before the run (see script docstrings):

Test	Pre-registered	Measured	Verdict
Barren ratio, 5 seeds	≥ ×500 at N=12	×1070 ± 167	✅
Deep-regime exponent (5 depths, 10-seed anchors)	∈ [−0.60, −0.45]	−0.54 ± 0.07 (bootstrap)	✅
Replication at N=12 (L=24/48/96)	same law	exponent −0.59	✅
Gradient position (layer 0 / L/2 / L−1)	≤ ±25%	≤ 8%	✅
Depth-not-size: g*(N) at fixed L=20	no decrease with N	no decrease (N=6→16)	✅
Live-registered cut-parity test (N=16)	boundary-cut class ≈ 0.39 (naive constant: 0.277)	0.449 ± 0.087	✅ (naive disfavored ~2σ)
Metric weight α=0.7	exponent ∈ [−0.65, −0.40]	−0.65	✅ borderline
Metric weight α=0.3	idem	unresolved (g* below grid resolution)	⚠️ open
Rényi-2 vs von Neumann bias	≤ 15%	+2% to +16%	✅ quantified
Block size m=2/3/4 (15 seeds/point)	naive uniform 1/√m	g*·√m = 0.51/0.52/0.37 — 1/√m holds within the same cut class (m=2 vs 3, 2% apart); the m=4 drop coincides with a mid-block half-cut, i.e. the same documented parity artifact as in g*(N)	naive law refuted; unified diagnostic-artifact picture (post-hoc, mechanism independently tested)
Ring vs chain	ring lower ~−10%	−35% (cut crosses 2 links → −29% post-hoc)	direction ✅, magnitude under-predicted

At the deep-regime optimum the window keeps ~50–57% of the gradient variance for ×2.6–3.2 inter-block entanglement (the earlier "~80%" figure was a shallow-depth value — corrected here).

Honest limitations (read before citing)

1. The principle "locality mitigates barren plateaus" is established (Cerezo et al. 2021). The contribution here is the continuous coupling sweep, the depth law g*(L), and the unified VQC+reservoir framing — not the qualitative idea.
2. The performance advantage is structurally beyond classical simulation: barren plateaus only bite the monolith at N ≳ 10–12, so the decisive test needs real hardware.
3. The depth law is established at N=10 and replicated at N=12, one ansatz family (ZZ chain/ring, m=2–4). The exponent is robust (size, gradient position, balanced metric weights); the constant is not universal — it shifts with the half-cut parity of the entanglement diagnostic (boundary vs mid-block cut, documented with a live pre-registered test), with block size (~1/√m), and with topology (ring −35%). An earlier apparent g*(N) trend conflated size and depth (our protocol used L=2N) and a 2-seed "plateau" claim was retracted after a rigorous 5-seed run — documented here deliberately rather than hidden.
4. Open question (the #1 referee objection): does the weak-coupling window escape the Cerezo et al. 2025 result "provable absence of barren plateaus ⇒ classical simulability"? This is the central thing to resolve.

Closest prior work (assumed, cited up front)

• Cerezo et al., Cost-function-dependent barren plateaus, 2021 — locality ↔ trainability.
• Patti et al., Entanglement Devised Barren Plateau Mitigation, PRR 2021.
• Park et al., Multi-Chip Ensembles, arXiv:2505.08782 (2025) — inter-chip coupling, but binary (no continuous g).
• Askari, Kora, Simon, arXiv:2511.04900 (2025) — memory capacity peaks at weak coupling in spin-network QRC.
• Cerezo et al., Absence of BP ⇒ classical simulability?, Nat. Commun. 2025.

Status

Preliminary but stress-tested: adversarial review, prior-art search over 17 papers, pre-registered predictions — one confirmed (L=160: predicted 0.098, measured 0.096 ± 0.004, 10 seeds) — and one earlier claim retracted when a rigorous re-run falsified it. Not a publishable breakthrough as-is — a precise, coherent hypothesis with reproducible POCs, offered to start a conversation with researchers.

Author

Nicolas Berger — independent researcher (France). [email protected], or open an Issue.