| name | non-markovian-kerr-feedback-qrc |
| description | Proves unbounded computational superiority of Kerr nonlinear feedback over Gaussian linear reservoirs in continuous-variable quantum reservoir computing. Single Kerr mode with feedback depth D replaces up to ~100 linear modes. Use when: CV-QRC design, non-Markovian reservoir computing, Kerr nonlinear optics, cross-time nonlinear correlations, quantum reservoir capacity analysis, Gaussian limitations, time-delay feedback. |
Overview
This methodology proves that a single Kerr nonlinear element in a time-delayed feedback loop can outperform arbitrarily many linear optical modes in continuous-variable quantum reservoir computing (QRC). It establishes a fundamental limit on Gaussian reservoirs and shows how one nonlinear mode replaces ~100 linear ones.
Paper: Soh (2026). "Computational Superiority of Non-Markovian Kerr Feedback in Continuous-Variable Quantum Reservoir Computing." arXiv:2606.06689 [quant-ph, math-ph].
Core Problem: The Gaussian Ceiling
Why Linear Optical Reservoirs Fail
A linear optical medium can delay, mix, and superpose light pulses — but cannot multiply. Cross-time nonlinear correlations (products of input at different past times) are essential for many temporal computations, yet Gaussian reservoirs cannot form genuine products within the reservoir. They can only:
- Store each past input separately
- Multiply them in the readout layer
- This forces an exponentially harder high-order measurement
The Hardware Ceiling
Theorem (Cross-Time Nonlinear Rank):
- N-mode Gaussian reservoir: maximum cross-time nonlinear rank = 2N (hard ceiling)
- Single Kerr mode with feedback depth D: rank = D (no ceiling)
The Kerr Solution
Time-Delayed Feedback Architecture
Input → Kerr Element → Feedback Loop → Readout
↑ ↓
└── Delay ──────┘
Mechanism:
- Kerr effect: Phase depends on intensity → true multiplication inside the medium
- Feedback: Light revisits the Kerr element repeatedly → one mode mixes its own history against itself once per round-trip
- "Feedback turns time into space": D passes through one nonlinear mode replace D parallel linear modes
Unbounded Resource Separation
| Architecture | Cross-Time Nonlinear Rank | Hardware Cost |
|---|
| N-mode Gaussian reservoir | ≤ 2N (hard ceiling) | N modes |
| Single Kerr + depth D | = D (unbounded) | 1 mode |
For every N, one Kerr mode performs computations no N-mode linear reservoir can.
The Counterintuitive Role of Loss
Loss is the enabler, not the enemy:
- Each round-trip dims the light
- So the nonlinear phase differs pass to pass
- Giving every echo its own fingerprint
- Without loss, the passes would be redundant
Practical Numbers
- Achievable feedback depth D: 30 to 230 on integrated platforms
- Equivalence: One nonlinear mode replaces up to ~100 linear modes
- Tradeoff: Measurement time increases with feedback depth
Validation
- Confirmed on exact open-system simulation
- Grounded in nonlinear channel equalization benchmark
- Theoretical proof (Theorem 3, Corollary 2) with resource separation
Implementation Guide
Kerr Element Design
feedback_depth = D
loss_per_round = η
kerr_strength = χ
Time-Delay Mapping
t=0: Input pulse enters Kerr medium → phase φ₀ = χ|α₀|²
t=1: Feedback + new input → phase φ₁ = χ|α₀e^{-γ} + α₁|²
t=2: Feedback again → phase φ₂ = χ|α₀e^{-2γ} + α₁e^{-γ} + α₂|²
...
The cross-terms α₀·α₁, α₀·α₂, etc. emerge inside the medium, not in the readout.
Cross-Time Nonlinear Rank Calculation
def gaussian_rank(N_modes):
"""Maximum cross-time nonlinear rank for N-mode Gaussian reservoir."""
return 2 * N_modes
def kerr_rank(feedback_depth):
"""Cross-time nonlinear rank for single Kerr mode with feedback depth D."""
return feedback_depth
Design Principles for CV-QRC
- Gaussian is not enough: Linear optics alone cannot compute cross-time nonlinear correlations efficiently
- One nonlinearity + feedback > many linear modes: Kerr element is a universal resource amplifier
- Loss enables computation: Counterintuitively, dissipation creates distinct fingerprints per round-trip
- Time-as-space: Feedback converts temporal depth into computational capacity
- Integrated platform advantage: D=30-230 achievable on silicon photonics
Applications
- Nonlinear channel equalization: Signal processing with cross-time correlations
- Temporal sequence prediction: Time series with nonlinear history dependence
- Quantum signal processing: Analog quantum computing on optical platforms
- Photonic neural networks: Energy-efficient temporal processing
Related Skills
quantum-reservoir-computing — QRC framework overview
amplitude-encoded-quantum-reservoir-protocol — Online QRC with amplitude encoding
quantum-reservoir-computing-risk-bounds — Rademacher complexity bounds for QRC
quantum-reservoir-operating-band — Transferable operating region for QRC
quantum-photonic-neural-networks — Photonic QNN architectures
arXiv: 2606.06689 | Date: June 4, 2026 | Authors: Daniel Soh