Technical Note · Non-Hermitian Photonics · Physical Reservoir Computing

A Coexistence Criterion for Computation in Driven Disordered Optical Cavities

Why high modal dimensionality is necessary but not sufficient, and the conditions a substrate must meet simultaneously.

STATUS — PRE-EXPERIMENTAL LICENSE — CC0 / PUBLIC DOMAIN FORM — SHORT NOTE

A passive optical cavity filled with linear media implements a linear input–output map, however elaborate its engineered disorder. Such a map can be of arbitrarily high rank and still computes no nonlinear function of its input history. The conditions under which a driven, open, disordered cavity becomes a genuine computational substrate are therefore conditions on three coupled physical observables — dimensionality, memory, and a nonlinearity internal to the field evolution — that all enter through the same complex spectrum. This note states those conditions and reduces the computation question to a measurable criterion on the Volterra kernels of the input–output map.

§ 1

Setup: the open cavity as a non-Hermitian operator

Consider a resonator with a spatially structured permittivity \(\varepsilon(\mathbf r)\) — a plano-hemispherical cavity seeded with subwavelength dielectric scatterers. In the frequency domain the passive field obeys

(1) \[ \nabla\times\nabla\times\mathbf E(\mathbf r,\omega)-\frac{\omega^{2}}{c^{2}}\,\varepsilon(\mathbf r)\,\mathbf E(\mathbf r,\omega)=\mathbf 0. \]

Because the cavity is open — radiative leakage through the curved wall — the relevant operator is non-Hermitian, and its natural modes are quasinormal modes with complex eigenfrequencies

(2) \[ \tilde\omega_{m}=\omega_{m}-i\,\gamma_{m},\qquad \gamma_{m}>0. \]

The real part fixes the spectral position; the imaginary part fixes the linewidth and the photon lifetime. The scatterers break integrability and mix these modes.

§ 2

The linearity constraint

For linear media, Eq. (1) is linear in \(\mathbf E\). The map from drive to field is therefore a linear operator \(A(\omega)\),

(3) \[ \mathbf E_{\text{out}}(\omega)=A(\omega)\,\mathbf E_{\text{in}}(\omega). \]

Equivalently, expanding the field in quasinormal modes, \(\mathbf E(\mathbf r,t)=\sum_m a_m(t)\,\boldsymbol\psi_m(\mathbf r)\), the modal amplitudes obey a coupled-mode equation

(4) \[ \frac{da_{m}}{dt}=(i\omega_{m}-\gamma_{m})\,a_{m}+\sum_{n}\kappa_{mn}\,a_{n}+N_{m}\!\left[\{a\}\right]+s_{m}(t), \]

where \(\kappa_{mn}\) is the disorder-induced linear mode coupling, \(s_m\) the projected drive, and \(N_m\) the coupling generated by any field-dependence of \(\varepsilon\). In the passive linear limit \(N_m=0\), and the solution is a convolution,

(5) \[ a_{m}(t)=\int_{0}^{\infty}\! G_{m}(t-t')\,s(t')\,dt'. \]

This is the constraint in one line: the internal state is a linear functional of the drive history. No matter how many modes participate or how complex \(A(\omega)\) is, the state never depends on products of its own past. A high-rank linear map is a random projection, not a dynamical system that computes.

§ 3

Three observables

Dimensionality. Let \(M\) be the state matrix sampled by the readout, with singular values \(\sigma_i\). The effective dimensionality is the participation ratio of its singular spectrum,

(6) \[ r_{\text{eff}}=\frac{\left(\sum_i \sigma_i\right)^{2}}{\sum_i \sigma_i^{2}}, \]

set by the eigenfunction statistics (participation ratio, IPR) of the mixed mode field — not by the detector count.

Memory. The fading-memory timescale is the decay-rate spectrum measured against the drive period: \(\tau_m=1/\gamma_m\), with the cavity scale \(\tau\simeq \mathcal F L/\pi c\) for finesse \(\mathcal F\). Modal individuality is governed by the modal-overlap (Thouless) parameter

(7) \[ \delta=\frac{\bar\gamma}{\Delta\omega}, \]

with \(\Delta\omega\) the mean level spacing. Resolved modes require \(\delta\lesssim1\); for \(\delta\gg1\) resonances overlap and individuality is lost.

Nonlinearity. A genuine activation enters only through a field-dependent permittivity,

(8) \[ \varepsilon(\mathbf r)\;\to\;\varepsilon(\mathbf r)+\chi_{\text{eff}}\!\left(\mathbf r,|\mathbf E|^{2}\right), \]

for example saturable gain \(\chi_{\text{eff}}=\chi_0\,(1+I/I_{\text{sat}})^{-1}\) or a Kerr shift \(\Delta n=n_2 I\). What matters is not only its magnitude but its response time relative to \(\tau_m\): a nonlinearity slow compared with the photon lifetime acts as a quasi-static background, not a per-symbol activation.

§ 4

The decisive criterion

The computation question, free of any reservoir-computing vocabulary, is a statement about the Volterra expansion of the input–output map:

(9) \[ y(t)=\sum_{k\ge1}\;\int\!\cdots\!\int h_{k}(\tau_1,\dots,\tau_k)\prod_{i=1}^{k}u(t-\tau_i)\,d\tau_i. \]

A linear system has only a first-order kernel — \(h_1=G\) and \(h_{k\ge2}\equiv0\). Nonlinearity in the state evolution appears as a nonvanishing higher-order kernel, and the physically meaningful requirement is that it be sourced by the intracavity field-dependence rather than by a nonlinear detector, with temporal support set by the photon lifetime:

(10) \[ h_{2}(\tau_1,\tau_2)\neq0,\qquad \operatorname{supp} h_2 \sim \tau_{\text{photon}}. \]

This collapses three fuzzy properties into one measurable quantity: drive the system with impulses and impulse pairs and read off whether \(h_2\) survives over the memory window.

A nonvanishing \(h_2\) is a necessary criterion, not a sufficient one. A substrate can be nonlinear and still compute poorly if its memory capacity or readout conditioning is weak; the sufficient figure of merit is task-level — memory capacity, or accuracy on a benchmark that demands nonlinear memory (NARMA, parity-\(N\)) measured against a linear-readout control.

§ 5

The coexistence problem

By coexistence we mean the simultaneous satisfaction of the conditions below — not phase or modal coexistence, nor bistability. The difficulty is that the observables are not independent — disorder, loss, and the nonlinearity all enter through the same complex spectrum \(\{\omega_m-i\gamma_m\}\) and the same susceptibility. Disorder raises \(r_{\text{eff}}\) but adds radiative loss, which shortens memory and drives \(\delta\) past unity; a nonlinearity strong enough to make \(h_2\neq0\) — gain saturation near threshold, say — tends to induce mode competition that collapses the field onto a few modes and sends \(r_{\text{eff}}\to1\). The same non-Hermiticity exerts one further pull: the cavity's modes are non-orthogonal, and that non-orthogonality amplifies quantum-limited noise by the Petermann factor \(K\), which diverges near the exceptional points that strong mixing drives toward — so the high-\(r_{\text{eff}}\) regime is intrinsically the high-noise regime. The open question is whether a single operating point satisfies all of the following at once:

Simultaneous conditions — does such a regime exist?
  • Nonlinearity in the recurrence\(h_2\neq0\), intracavity-sourced, with \(\operatorname{supp}h_2\sim\tau_{\text{photon}}\).
  • Dimensionality\(r_{\text{eff}}=\mathcal O(N_{\text{modes}})\).
  • Memory matched to drivespread\(\{\gamma_m\}\) comparable to the symbol rate.
  • Modal individuality\(\delta=\bar\gamma/\Delta\omega=\mathcal O(1)\).
  • Readout above the excess-noise floorper-mode SNR degraded by \(\sqrt{K}\) (Petermann factor), worst exactly where mode mixing is strongest.

The linear-disorder half — the eigenfunction statistics and decay spectrum — is well-trodden open-wave-chaos physics. The unsettled half is the multimode nonlinear regime: whether a field-dependence strong enough to register in \(h_2\) on the photon-lifetime timescale can coexist with high modal participation, or whether it generically destroys it. Whether the favorable region is nonempty is the whole problem.