Routing Polarization in Photonics and Networks

• Routing polarization is a multifaceted phenomenon that exploits polarization-dependent dispersion, birefringence, and interference to route photons, plasmons, and quantum signals.
• It encompasses techniques that either use polarization as the routing variable or preserve it as an unaltered payload, enabling tunable delays and precise output discrimination.
• Beyond photonics, the term also describes uneven resource allocation in GPU clusters and capsule networks, underscoring challenges in balanced routing across various systems.

Routing polarization denotes a family of routing phenomena in which polarization is either the control variable, the preserved information carrier, the routed output label, or—outside photonics—a metaphor for route imbalance. In quantum and classical photonics, the term usually refers to directing photons, plasmons, guided modes, or multipolar emission into distinct spatial or temporal channels by exploiting polarization-dependent dispersion, birefringence, chirality, topology, or interference; in several recent works it also denotes routing while leaving an arbitrary polarization qubit unchanged. In other literatures, the same phrase names symmetry-constrained concentration of vacuum correlations into one polarization sector, uneven use of routing resources in optical-circuit-switched GPU fabrics, or the sharpening of capsule-to-capsule couplings in neural routing algorithms (Maisch et al., 2019, Wang et al., 18 Feb 2025, EL-Amrani et al., 29 May 2026, Han et al., 30 Mar 2026, Paik et al., 2019).

1. Terminological scope

The phrase is not used with a single invariant meaning. In photonics, “routing polarization” most often means that polarization determines where radiation goes, or that a device routes signals while preserving polarization. In some many-body settings it means that a symmetry confines correlations to one polarization sector. In network systems and machine learning, it is a descriptive term for route concentration rather than an optical polarization effect.

Domain	Routed quantity	Meaning of “polarization”
Quantum and nanophotonics	photons, plasmons, guided modes, emission channels	polarization selects path, delay, mode, or output channel
Polarization-preserving quantum routing	single-photon or entangled-photon paths	path changes while polarization is maintained
Ultrastrong-coupling cavity QED	vacuum correlations, squeezing, entanglement	correlations are routed into one polarization sector
OCS GPU clusters	Pod-to-Pod logical bandwidth	route allocation becomes uneven across spine switches
Capsule networks	coupling coefficients $c_{ij}$	routing iterations polarize link strengths

A common misconception is to treat all usages as equivalent. The photonic literature distinguishes at least three technically different operations: polarization-dependent routing, where polarization selects a path; polarization-maintaining routing, where the path changes but polarization is ideally unchanged; and polarization-channel engineering, where different physical processes are emitted into different polarization outputs. A further misconception is that “routing polarization” necessarily concerns real-space directionality. Several works instead route arrival time, quantum correlations, or multipolar content into polarization-resolved channels (Maisch et al., 2019, Wang et al., 18 Feb 2025, EL-Amrani et al., 29 May 2026, Rikers et al., 25 Feb 2026).

2. Polarization-dependent routing of single photons

A concrete single-photon realization combines a resonantly excited In(Ga)As/GaAs quantum dot, a 25 cm hot cesium vapor cell on the $D_1$ line, and polarization optics consisting of a polarizer before the cell and a quarter-wave plate plus polarizing beam splitter after the cell. In that system, a linearly polarized photon is a superposition of $\sigma^+$ and $\sigma^-$; inside the magnetized vapor, the two circular components experience different refractive indices and dispersions, hence different group velocities and delays. The analyzer maps $\sigma^\pm$ onto spatially separated outputs, so the photon’s output path and arrival time are controlled by its polarization state via the polarization-dependent dispersion of a magnetized cesium vapor (Maisch et al., 2019).

The delay mechanism is standard slow light. For a given polarization mode,

$$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$

At approximately $130^\circ$C, in the narrow transmission window between the cesium $D_1$ hyperfine resonances, the experiment measured an overall delay of approximately $25\ \text{ns}$ through the $25\ \text{cm}$ cell, with fine-tuning of more than $D_1$0 by changing the magnetic field and thereby the relative delay between polarization components (Maisch et al., 2019).

The routing functionality is controlled by the sign and magnitude of the longitudinal magnetic field. For $D_1$1, the $D_1$2 and $D_1$3 components see the same dispersion and their arrival times coincide. At $D_1$4, the arrival-time histograms at the two detectors are shifted in opposite orders, and reversing the sign of $D_1$5 swaps which output is early and which is late. The theoretical model reproduces the timing data by computing the complex refractive index including Doppler broadening, Zeeman splitting, and spectral diffusion of the quantum-dot photons (Maisch et al., 2019).

This implementation is significant because it makes routing and buffering inseparable: the same polarization-dependent medium produces both output-port discrimination and a tunable temporal offset. A plausible implication is that polarization routing in dispersive media is not only a switching primitive but also a spectro-temporal interface between solid-state emitters and atomic systems.

3. Polarization as payload: routers that preserve the polarization state

A distinct line of work treats polarization not as the selector of the route but as the quantum information that must survive routing unchanged. In an interferometer-based telecom router, the desired operation is

$D_1$6

namely an identity map on polarization together with a controllable map on path. The implementation uses a free-space Mach–Zehnder interferometer with two 50:50 non-polarizing beam splitters, a semi-common-path geometry, low angle of incidence optics, and electro-optic phase modulators built from cross-aligned rubidium titanyl phosphate crystals so that static and electro-optic birefringence cancel. The reported performance is a 1.3% loss, a $D_1$7 switching extinction ratio, and $D_1$8 polarization process fidelity to ideal identity operation, together with routing of two-photon N00N-type entangled states with a highly maintained interference visibility of $D_1$9 (Wang et al., 18 Feb 2025).

The same preserve-while-route objective appears in a fiber-optical Sagnac interferometer in which a pair of fast electro-optical telecom phase modulators act on orthogonal polarization components of the single photons. The output state factorizes into a path part controlled by the applied phase $\sigma^+$0 and a polarization part independent of $\sigma^+$1, so the routing probabilities are

$\sigma^+$2

independent of the input polarization amplitudes. The experiment obtained an average extinction ratio of more than $\sigma^+$3 between both outputs of the switch, and the total loss from the Sagnac input to the detectors was approximately $\sigma^+$4, dominated by the insertion loss of the phase modulators (Alarcón et al., 2020).

On-chip, the rotated polarization directional coupler provides yet another preserve-and-project architecture. There, a double-track femtosecond-laser-written waveguide acquires an arbitrary birefringent optical axis, and two such waveguides in a strong-coupling regime form a coupler that routes orthogonal linear polarizations in a rotated basis into different outputs. Demonstrated devices at $\sigma^+$5 and $\sigma^+$6 yielded average extinction ratios of about $\sigma^+$7 and $\sigma^+$8 for the corresponding orthogonal polarizations, and average Stokes-vector reconstruction fidelities up to $\sigma^+$9 and $\sigma^-$0 for the perfectly initialized states in the $\sigma^-$1 and $\sigma^-$2 devices, respectively (Wang et al., 2018).

Taken together, these works suggest a useful taxonomy. In one class, polarization is the routing variable. In another, polarization is the protected payload and routing must approximate $\sigma^-$3. Conflating these classes obscures a central design difference: the first requires strong polarization dependence, whereas the second requires its systematic cancellation.

4. Nanophotonic and metasurface implementations

At subwavelength scales, routing polarization is often implemented through spin–orbit coupling, anisotropic scattering, nonlinear interference, or polarization-engineered local density of states. A mirror-symmetric periodic metasurface on gold can route surface plasmon polaritons because the SPP field

$\sigma^-$4

carries a transverse spin locked to propagation direction, while oblique circularly polarized illumination provides a longitudinal spin with a projection onto that transverse spin proportional to $\sigma^-$5. Switching between right- and left-handed circular polarization reverses the preferential SPP direction, even though the unit cell itself is mirror symmetric (Revah et al., 2018).

A related nanoscale route selection is achieved with structured illumination of an isotropic scatterer supporting electric and magnetic dipolar resonances. By tightly focusing radial or azimuthal beams, one can realize two Huygens-dipole configurations: longitudinal electric with transverse magnetic dipole moments, or longitudinal magnetic with transverse electric dipole moments. The first yields directional emission only in the TM channel, the second only in the TE channel, with experimentally measured scattering directivities of around $\sigma^-$6 and $\sigma^-$7 in TM and TE modes, respectively (Nechayev et al., 2019).

On an integrated silicon nitride slab waveguide, a plasmonic grating made of gold nanobars rotated by $\sigma^-$8 uses circular polarization to select both guided-mode family and propagation direction. Right-handed circular polarization excites a TE mode propagating only to the left and a TM mode propagating only to the right; left-handed circular polarization reverses those directions. The routing efficiency for normally incident light reaches up to $\sigma^-$9, and in the reciprocal out-coupling regime the degree of circular polarization reaches up to $\sigma^\pm$0 (Fradkin et al., 2023).

Nonlinear interferometry extends the same idea to frequency-converted light. In an AlGaAs metasurface, third-harmonic generation and sum-frequency generation both produce light at $\sigma^\pm$1, but with orthogonal local polarizations at selected diffraction orders. Varying the relative phase between the $\sigma^\pm$2 and $\sigma^\pm$3 pumps moves the $\sigma^\pm$4 output continuously between linear and nearly circular states, with a degree of circular polarization up to approximately $\sigma^\pm$5. At the $\sigma^\pm$6 and $\sigma^\pm$7 diffraction orders, opposite circular handedness is routed into opposite sides of $\sigma^\pm$8-space, and a phase shift of $\sigma^\pm$9 swaps the handedness between the two directions (Luan et al., 2024).

An anisotropic dielectric metasurface can also route multipolar order into polarization channels. For Eu$$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$0 coupled to a square array of elliptical a-Si:H dimers, the electric-dipole transition $$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$1 at approximately $$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$2 is preferentially emitted into the $$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$3-polarized channel, while the magnetic-dipole transition $$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$4 at approximately $$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$5 is preferentially emitted into the $$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$6-polarized channel. On the metasurface, the ratio $$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$7 becomes $$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$8 in the $$n_g(\omega)=n'(\omega)+\omega\frac{\partial n'(\omega)}{\partial \omega},\qquad v_g=\frac{c}{n_g},\qquad \tau=\frac{L n_g}{c}.$$9-polarized channel and $130^\circ$0 in the $130^\circ$1-polarized channel, compared with $130^\circ$2 on the substrate for both polarizations (Rikers et al., 25 Feb 2026).

In these nanophotonic realizations, polarization routing does not reduce to a single mechanism. The routed object may be a plasmonic beam, a TE/TM waveguide mode, a nonlinear diffraction order, or an electric- versus magnetic-dipole emission line. What is shared is the deliberate conversion of polarization information into channel selectivity.

5. Symmetry-, topology-, and correlation-mediated routing

A more abstract usage appears in ultrastrong-coupling cavity QED, where chirality routes vacuum correlations rather than photons along spatial paths. In a multimode Hopfield model for Landau polaritons, the exact chiral charge

$130^\circ$3

commutes with the Hamiltonian. This symmetry implies selection rules such as

$130^\circ$4

so the spectroscopically bright polarization $130^\circ$5 supports the polariton branches, while the dominant anomalous correlations, squeezing, and cavity–matter entanglement are routed into the opposite, spectroscopically dark polarization $130^\circ$6. The computed logarithmic negativities satisfy $130^\circ$7 over the full magnetic-field range, whereas $130^\circ$8 and similarly for magnetoplasmon modes (EL-Amrani et al., 29 May 2026).

Topological photonics provides a spatial routing counterpart. A Floquet-engineered microring lattice on a silicon nitride platform supports complementary trivial and topological band gaps for orthogonal eigenpolarizations. At telecom wavelengths, TE modes propagate via a topological edge state while TM modes are suppressed by a trivial gap; at shorter wavelengths the behavior reverses. The measured extinction ratios are $130^\circ$9–$D_1$0 for the protected port and $D_1$1–$D_1$2 for the non-protected port, with insertion loss of $D_1$3 at long wavelengths. The same lattice also exhibits spectral regions in which both polarizations are topological, enabling a Floquet dual-polarization topological regime (Afzal et al., 21 Apr 2026).

A distinct topological construction works directly in Stokes space. Using passive optical components together with Mach–Zehnder interferometers, one can generate a polarization circle $D_1$4, then sweep it into a polarization torus $D_1$5, and the same framework also admits Möbius strips, Hopf links, and topological Dirac bosons with a bulk-edge correspondence. Here the routed entity is the trajectory of polarization states under controlled phase shifts and rotations, rather than a beam in real space (Saito, 2023).

Ferroelectric surface ferrons provide a further collective-wave example. The high-frequency branch has locked circular polarization and momentum, while the low-frequency branch is in-plane polarized normal to its wave vector. Because that lower branch is strongly anisotropic, a focused laser beam can induce directional emissions of electric polarization and chiral near fields, allowing optical routing in ferroelectric devices (Zhou et al., 2022).

A recurring misconception is that the spectroscopically bright channel is necessarily the channel that carries the decisive routed quantity. The Landau-polariton result shows the opposite: spectroscopic brightness and correlation-bearing capacity can be orthogonalized by symmetry. This suggests that “routing polarization” can designate the organization of hidden sectors as much as visible transport.

6. Divergent non-photonic usages

Outside photonics, the phrase names concentration phenomena in routing variables themselves. In OCS-based GPU clusters, routing polarization refers to the scenario where the bandwidth requirements between specific pairs of Pods are unevenly fulfilled through links among different spine switches. The consequence is traffic contention and bottlenecks on specific leaf-to-spine links, even when aggregate Pod-to-Pod connectivity is sufficient. A leaf-centric logical-topology design enforces balanced distribution of traffic originating from the same leaf across multiple spines. For $D_1$6, the paper establishes a sufficient condition for avoiding routing polarization and proposes a polynomial-time algorithm; large-scale simulations show up to $D_1$7 throughput improvement and a $D_1$8 reduction in logical-topology computation overhead compared to Mixed Integer Programming-based methods (Han et al., 30 Mar 2026).

In capsule networks, routing polarization denotes the sharpening of coupling coefficients $D_1$9 produced by iterative routing algorithms. The empirical analysis of five routing algorithms found that, in most cases, the routing algorithms do not change the classification result but polarize the link strengths, and the polarization can be extreme when they continue to repeat without stopping. The study also reported that routing algorithms often produce results that are worse than simple baseline algorithms that assign the connection strengths uniformly or randomly (Paik et al., 2019).

These non-photonic usages are not merely accidental homonyms. In both cases, “polarization” denotes concentration of routing degrees of freedom onto a small subset of available channels. That is the opposite of the balanced or deliberately engineered multiplexing sought in photonic routing devices. A plausible implication is that the term now spans two almost inverse design goals: in photonics, one often engineers polarization selectivity to create function, whereas in networking and capsule routing one seeks to prevent uncontrolled polarization of resource allocation or couplings.