photonic-variational-trainability

name

photonic-variational-trainability

description

Pre-asymptotic trainability analysis for photonic variational quantum circuits under postselection. Covers barren plateau dynamics in passive linear-optical circuits, Lie algebra dimension scaling, postselection-induced gradient concentration (allow-bunching, collision-free, dual-rail), and design guidance for near-term photonic variational architectures. Use when: analyzing photonic QNN trainability, designing variational photonic circuits, understanding gradient concentration under postselection, or comparing postselection regimes for optical quantum computing. Trigger keywords: photonic barren plateau, variational photonic circuits, postselection gradient concentration, linear optical quantum computing, dual-rail postselection, collision-free filtering, photonic QNN trainability.

Photonic Variational Trainability

From arXiv:2605.11879 "Pre-Asymptotic Trainability in Photonic Variational Circuits under Postselection" (Xie, Notton, Senellart, 2026).

Core Problem

Barren plateaus in variational quantum circuits cause gradient variance to vanish exponentially with system size. Passive photonic circuits challenge this picture: although their Hilbert space is exponentially large, dynamics are constrained to a Lie algebra of dimension O(m²) where m = number of modes.

Key Finding: Postselection Determines Trainability

Gradient concentration is governed not by Hilbert-space dimension but by how postselection reshapes the effective observable.

Three Postselection Regimes

Regime	Gradient Scaling	Trainability
Allow-bunching	Polynomial decay	✅ Trainable
Collision-free filtering	Polynomial decay	✅ Trainable
Dual-rail postselection	Exponential concentration	❌ Barren plateau

Mechanism

1. Allow-bunching: No postselection filtering → full Lie algebra access

   • Gradient variance ~ O(1/poly(m))
   • Remains trainable for moderate system sizes

2. Collision-free filtering: Post-select on no two photons in same mode

   • Restricts accessible subspace but preserves polynomial scaling
   • Trainability maintained across initialization ensembles

3. Dual-rail postselection: Encode qubits in dual-rail basis, post-select

   • Induces exponential gradient concentration beyond moderate m
   • Robust across three initialization ensembles tested
   • Critical insight: dual-rail encoding + postselection = barren plateau

Design Guidelines for Photonic Variational Architectures

When Trainability is Expected

• Use allow-bunching or collision-free regimes
• Keep observable support localized (few-body operators)
• Avoid dual-rail postselection for variational training

When to Expect Barren Plateaus

• Dual-rail encoding with postselection beyond moderate system sizes
• Global observables spanning many modes
• Deep circuits with extensive mode mixing

Practical Recommendations

1. Initialization matters: Results robust across three ensembles, but initialization choice affects pre-asymptotic behavior
2. Observable choice: Local observables preserve trainability longer
3. Mode count: Polynomial regime extends to moderate m (tested up to ~20 modes)
4. Postselection geometry: The shape of the postselected subspace, not just its dimension, determines gradient behavior

Pitfalls

• Assuming photonic circuits are immune to barren plateaus — dual-rail postselection DOES cause exponential concentration
• Extrapolating small-system behavior: polynomial scaling eventually breaks
• Ignoring observable structure: global observables accelerate concentration

Related Patterns

• See quantum-neural-barren-plateau for general QNN barren plateau mitigation
• See photonic-qnn-algorithmic-advantage for photonic QNN expressivity analysis