Optical Bilinear Interaction

• Optical Bilinear Interaction is a mapping that multiplies two operand sets while remaining linear in each operand when the other is fixed.
• It appears in various contexts including passive interferometric circuits, pump-linearized photon-phonon couplings, and inverse-problem formulations, each with distinct physical implementations.
• The framework distinguishes true bilinear operators from other multiplicative processes like Kerr-induced mixing, emphasizing precise algebraic structure and modeling techniques.

Optical bilinear interaction denotes a class of optical or optically mediated mappings in which two operand sets enter multiplicatively while the mapping remains linear in either operand when the other is fixed. In contemporary literature, the term spans several distinct technical settings: phase-domain modular inner products in passive interferometric circuits, pump-linearized photon-phonon couplings in Brillouin ring-resonator arrays, standard bilinear free-electron/optical-field coupling to nonlinear cavity states, and inverse problems in which an optical image and a forward-operator correction interact multiplicatively. By contrast, several multiplication-like optical processes are not bilinear in this strict sense, including cubic Kerr-induced multimode mixing, quadratic square-law detection, and strong-field spectra that mimic even-order response without realizing a genuine bilinear operator (Pavlichin et al., 2014, Vovchenko et al., 17 Dec 2025, Yang et al., 2023, Hakula et al., 19 Mar 2026, Teğin et al., 2020, Villari et al., 2022).

1. Terminological scope and formal distinctions

In the cited literature, optical bilinearity appears in at least four nonidentical forms. One is algebraic bilinearity, where an output such as $\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$ is linear in a stored phase vector and in a selector vector separately. A second is operator bilinearity, where the interaction Hamiltonian is bilinear in bosonic operators, as in beam-splitter-like photon-phonon terms or two-mode-squeezing-like sideband couplings. A third is measurement-model bilinearity, where two unknowns, such as an optical image $x$ and forward-model coefficients $y$, enter through products $y_iV_ix$. A fourth is effective bilinear response in nonlinear optics, where even-order signals are interpreted as field-bilinear only when symmetry and perturbative structure justify that interpretation.

Setting	Bilinear structure	Necessary distinction
Passive interferometric circuits	$\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$	Output is a phase modulo $2\pi$, not a conventional analog amplitude MAC
Pump-linearized interaction media	Bilinear operator terms such as $\hat b^\dagger \hat c$	The implemented computation can still be linear in encoded inputs
Optical inverse problems	$b=A_0x+\sum_i y_iV_ix$	Bilinearity is in estimation, not in physical wave mixing
Kerr multimode processors or strong-field spectra	Multiplication-like observables may appear	Cubic propagation, quadratic detection, or spectral overlap are not bilinear maps

This suggests a useful technical distinction: the phrase is most precise when reserved for operators that are linear in each operand separately, and less precise when applied to any optical process exhibiting multiplicative structure. Much of the recent literature is concerned precisely with enforcing this distinction.

2. Bilinear interaction as a physical coupling mechanism

In spontaneous-Brillouin optical coprocessing, the underlying physical interaction is a pump-enhanced photon-phonon process that is microscopically three-wave mixing, but after the undepleted pump approximation it reduces to a bilinear interaction for the sidebands. The Hamiltonian then contains anti-Stokes terms of beam-splitter type, $$\hat b_{k,a}^\dagger \hat c_{k,q_k}+\hat c_{k,q_k}^\dagger \hat b_{k,a}$$, and Stokes terms of two-mode-squeezing type, $$\hat b_{k,s}^\dagger \hat c_{k,q_k}^\dagger+\hat c_{k,q_k}\hat b_{k,s}$$. In the weak-coupling regime, the stationary anti-Stokes occupancy becomes a weighted sum of thermal phonon occupations,

$x$0

The input vector is therefore the thermal phonon occupation vector, the effective matrix is set by Brillouin rates and ring-waveguide couplings, and the output is a stationary anti-Stokes occupancy. The physical interaction is bilinear in operators, whereas the implemented map from encoded input to measured output is linear (Vovchenko et al., 17 Dec 2025).

An analogous but conceptually separate instance appears in free-electron optics. There, the electron-light interaction remains the standard bilinear free-electron/optical-field coupling, even when the optical field sampled by the electron is generated by nonlinear Kerr dynamics inside a microresonator. The interaction is summarized by the coupling parameter

$x$1

which determines sideband amplitudes

$x$2

The novelty in that setting is not a nonlinear electron-photon Hamiltonian, but bilinear electron-light coupling to optical states produced by the Lugiato–Lefever dynamics of a Kerr microresonator, including Turing patterns, chaotic modulation instability, and dissipative Kerr solitons. The interaction observable remains bilinear in the conventional PINEM/Ramsey sense, while the sampled optical state is nonlinear and driven-dissipative (Yang et al., 2023).

3. Interferometric phase-domain bilinearity

A distinct lineage treats optical bilinear interaction as a passive interferometric realization of inner-product-like algebra. In "Optical modular arithmetic" the basic architecture is a cascade of Mach–Zehnder interferometers built from 50/50 beamsplitters, phase shifters, and a coherent input field propagating in two rails. Memory is stored in phase shifts $x$3, while binary control is encoded through phases $x$4. The exact arithmetic primitive is a modular inner product,

$x$5

with arithmetic modulo $x$6 for the optical phase and modulo $x$7 for the control synthesis. The selector relation is mediated by the lower-triangular matrix $x$8, and the final tail phase satisfies $x$9 so that the light exits the desired output port (Pavlichin et al., 2014).

The physical mechanism is routing and phase accumulation rather than multiplication of two propagating optical signals. For a length-$y$0 vector, the base selector uses $y$1 50/50 beamsplitters, $y$2 memory phase shifters, $y$3 control phase shifters, and one extra tail phase. Parallel replication yields matrix-vector and matrix-matrix products in the same phase-modular algebra:

$y$4

A feedback architecture simplifies the selector relation to $y$5 in the binary case, while a modified feedback circuit yields an approximately weighted readout

$y$6

so that for small memory phases the phase-domain output is approximately linear in $y$7 with a control-dependent weight. The important limitation is that exact arbitrary analog weights are not available in the base architecture; analog weighting is approximate and small-signal.

This line of work is therefore best interpreted as an interferometric implementation of bilinear algebraic structure, not as nonlinear optical field-field multiplication. One operand is stored in hardware phases, the other is programmed through control settings, and the coherent field serves as a carrier of accumulated phase.

4. Nonlinear multimode mixing and why it is not bilinear

The distinction between bilinear interaction and more general nonlinear optical mixing is especially clear in the multimode-fiber "Scalable Optical Learning Operator". In that system, input data are encoded onto the phase of a short optical pulse with a spatial light modulator, launched into a graded-index multimode fiber, propagated through dispersion, linear intermodal coupling, and nonlinear Kerr-type intermodal coupling, then measured as output intensity and passed to a trained digital single-layer regression or classification head. The optical transformation is fixed and untrained; only the output decision layer is trained (Teğin et al., 2020).

The central modal evolution equation is

$y$8

The first four terms are linear in the field amplitudes and implement only linear mode mixing. The last term,

$y$9

is the true nonlinear optical engine. It is cubic, or trilinear in modal field amplitudes, and physically corresponds to Kerr $y_iV_ix$0 interactions encompassing self-phase modulation, cross-phase modulation, and four-wave-mixing-like intermodal processes. After propagation, the measured observable is output intensity, effectively a square-law dependence on the optical field. The resulting feature map is therefore a composition of linear mode mixing, cubic field dynamics, and quadratic detection rather than a clean bilinear operator.

The paper is explicit that the system is not an engineered two-input map $y_iV_ix$1. There is a single encoded optical field, and the multiplicative structure arises among its modal components. This is why the architecture is closer to extreme learning machines, optical random projections, and reservoir-style front ends than to a trainable optical bilinear layer. The empirical evidence for the usefulness of nonlinearity comes from power and length dependence: for the Sinc benchmark, performance is poor at low peak power where transmission is “very nearly a linear transformation,” improves near an optimum around $y_iV_ix$2, and later degrades when Raman beam cleanup dominates; in COVID-19 classification, experimental accuracy rises from $y_iV_ix$3 at $y_iV_ix$4 to $y_iV_ix$5 at $y_iV_ix$6. These results show that nonlinear multimode interactions improve the learned feature map, but they do not convert the processor into a bilinear optical module.

5. Bilinear interaction in optical inverse problems

In diffuse optical tomography, the phrase “bilinear inverse problem” refers not to nonlinear wave mixing but to multiplicative coupling between two unknowns in the measurement model. The neonatal atlas-based DOT formulation starts from a linear inverse problem,

$y_iV_ix$7

where $y_iV_ix$8 is the absorption image and $y_iV_ix$9 is the Jacobian. Because the forward operator is uncertain, the unknown true Jacobian is modeled by a PCA expansion around the mean,

$\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$0

Substituting this into the forward model gives

$\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$1

which is linear in $\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$2 for fixed $\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$3 and linear in $\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$4 for fixed $\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$5, but not jointly linear. This is the bilinear structure: the unknown optical image and the unknown forward-operator correction interact multiplicatively (Hakula et al., 19 Mar 2026).

The tensor notation

$\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$6

makes the same point in compact form. The Bayesian model uses independent Gaussian priors for noise, operator coefficients, and image:

$\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$7

leading to the MAP objective

$\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$8

Because of the bilinear term, $\mu_{\text{out}}=\vec{\mu}\cdot\vec{s}$9 is not convex, and multiple local minima may arise.

The application context is neonatal atlas-based frequency-domain DOT with $2\pi$0 and $2\pi$1 measurements from a high-density layout. The study uses 215 segmented neonatal head models spanning 29.3–44.3 weeks combined gestational and chronological age, retains the first ten PCs for each measurement in the row-wise PCA model, and evaluates reconstructions using contrast-to-noise ratio and RMSE. Three computational strategies are developed: block coordinate descent, Gauss–Newton, and Gibbs sampling. Gauss–Newton was the fastest practical method, with reconstruction time about 35 s at 1 mm and about 0.7 s at 2 mm in a laptop MATLAB implementation; block coordinate descent was much slower, requiring roughly 100x more iterations; Gibbs sampling produced the best reconstruction quality among the three methods, but at much higher computational cost. The broader significance is that bilinearity here is an estimation device for compensating atlas mismatch rather than a physical optical interaction.

6. Bilinear optics, even-order response, and interpretive pitfalls

A different use of bilinear language arises in nonlinear optical response, where second-order processes are conventionally identified with field-bilinear current or polarization terms. In twisted bilayer graphene, the nonperturbative Dirac–Bloch-equation framework does not explicitly derive a standard $2\pi$2 tensor, but it shows that inversion-symmetry breaking enables even-order optical response, including a second-harmonic-like feature, and that the effect is strongly controlled by Berry phase and twist-angle-dependent band geometry. The model introduces a monolayer gap $2\pi$3, which effectively breaks inversion symmetry through valley inequivalence; in inversion-symmetric twisted bilayer graphene only odd harmonics are found, whereas in the gapped case even harmonics appear (Villari et al., 2022).

The time-dependent dynamics are built from instantaneous eigenstates, with Berry phase

$2\pi$4

and microscopic polarization governed by equations in which the Berry-phase difference $2\pi$5 enters explicitly. The emitted current is decomposed into intraband and interband parts, and the harmonic amplitudes are extracted from the nonlinear current spectrum. The paper’s central claim for bilinear optics is interpretive rather than tensorial: true even-order response is associated with inversion breaking and is substantially enhanced by Berry phase, especially at smaller twist angles where the moiré bands are flatter.

The same work also formulates a caution that is broadly relevant to optical bilinear interaction: an observed peak at $2\pi$6 does not necessarily imply a genuine second-order process. In the extreme nonlinear-optics regime, using pulses with $2\pi$7 and intensity $2\pi$8, the authors find “odd harmonics in disguise of even harmonics.” These peaks result from spectral overlap and interference of broadened odd-harmonic envelopes rather than from a true field-bilinear susceptibility. In the genuine inversion-broken case, the second harmonic is paired with a zeroth-order peak because $2\pi$9 and $\hat b^\dagger \hat c$0 channels occur together; disguised harmonics do not show that paired structure. A plausible implication is that symmetry arguments, perturbative scaling, and spectral context must all be checked before classifying an optical signal as bilinear.

Optical bilinear interaction is therefore not a single physical primitive but a family of related structures. In some settings it is an exact algebraic form implemented interferometrically; in others it is a bilinear operator coupling after pump linearization; in still others it is a modeling device for joint estimation. The main technical boundary runs between operators that are truly linear in each operand separately and broader nonlinear optical processes that merely contain multiplicative terms or even-frequency observables.