GRIN: Gradient-Index and Beyond

• GRIN is a paradigm where a spatially varying refractive index manipulates wave trajectories in media such as optical fibers, acoustic devices, and metamaterials.
• GRIN devices often implement continuous profiles like Luneburg or discretized solutions using unit-cell libraries to achieve precise wave control.
• GRIN principles enable applications ranging from enhanced radar and quantum sensing to gradient-based computational methods and advanced library infrastructures.

GRIN denotes several distinct technical constructs, but across the cited wave-physics literature it most commonly refers to gradient-index or graded-index media: materials and devices in which refractive index, effective permittivity, or an analogous wave-speed parameter varies spatially so as to bend, collimate, focus, split, or otherwise control wave propagation. In this sense, GRIN spans acoustic devices, lens antennas, phononic crystals, multimode fibers, and thermally loaded laser media, with realizations ranging from continuous laws such as the Luneburg profile $n(r)=\sqrt{2-(r/R)^2}$ to discretized unit-cell libraries and 3D-printed metamaterials (Xie et al., 2018, Bagheri et al., 19 Jan 2026, Garcia et al., 2020, Wang et al., 2023). The acronym also appears in unrelated contexts, including GRadient-INformed MoE, Zero-Shot Metric Depth with Pixel-Level Diffusion, the Google Return Interface, and the “grin” of a Quantum Cheshire Cat (Liu et al., 2024, Guizilini et al., 2024, Daly et al., 14 Nov 2025, Das et al., 2019).

1. Gradient-index media and their physical principle

In GRIN optics and acoustics, the defining principle is a spatially varying refractive index $n(\mathbf r)$, or equivalently a spatially varying effective permittivity $\varepsilon_r(\mathbf r)$ when $\mu_r=1$, so that rays or wavefronts are continuously bent rather than redirected only at discrete interfaces. The modified spherical Luneburg-lens work states this explicitly: “Gradient‐index (GRIN) optics relies on a spatially varying refractive index $n(\mathbf r)$ to continuously bend rays,” while the acoustic-design work describes GRIN acoustic devices as having a “spatially inhomogeneous refractive index profile” that allows “flexible control of the propagation of acoustic waves” (Bagheri et al., 19 Jan 2026, Xie et al., 2018).

The same design logic appears in several geometries. The canonical Luneburg lens uses

$n(r)=\sqrt{\,2-(r/R)^2\,},$

or equivalently $\varepsilon_{\rm LL}(r)=2-(r/R)^2$, so that rays from any point on the surface are refracted to a corresponding point on the opposite surface (Bagheri et al., 19 Jan 2026). Conformal gradient-index phononic crystal lenses instead use a hyperbolic-secant profile,

$n(r)=n_0\,\operatorname{sech}(\alpha r),$

chosen for aberration-free focusing of meridional rays and then mapped onto cylindrical or conical shells (Danawe et al., 2023). In acoustic CAD, the target index may be an arbitrary analytic or tabulated function over a design volume $\Omega\subset\mathbb R^3$, for example the Gaussian bump lens

$$n(x,y,z)=1+0.7\exp\!\left[-\frac{(x-x_c)^2+(y-y_c)^2}{r_0^2}\right],$$

with $n(\mathbf r)$0, $n(\mathbf r)$1, and $n(\mathbf r)$2 (Xie et al., 2018).

A common misconception is that GRIN implies a single material class or a single wave type. The cited literature instead treats GRIN as a wave-control paradigm instantiated in airborne acoustics, millimeter-wave antennas, elastic-wave phononic crystals, optical propagation models, and fiber nonlinear dynamics (Xie et al., 2018, Bagheri et al., 19 Jan 2026, Danawe et al., 2023, Arabí et al., 2017).

2. Mathematical formulations across acoustics, optics, fibers, and thermal media

The mathematical representation of GRIN depends on the platform, but in each case the core object is a smooth or piecewise-smooth spatial profile that modifies phase accumulation and ray trajectories. In flat wideband lens antennas, the GRIN principle is written as a radial phase-delay law,

$n(\mathbf r)$3

or, if the index varies along $n(\mathbf r)$4,

$n(\mathbf r)$5

with the lens designed so that the exit phase front is planar (Garcia et al., 2020). In linear and quadratic optical GRIN media, the paraxial equation is

$n(\mathbf r)$6

with $n(\mathbf r)$7 for a linear GRIN and $n(\mathbf r)$8 for a quadratic GRIN (Asenjo et al., 2020).

In graded-index multimode fibers, the core index decreases quadratically from the center toward the cladding:

$n(\mathbf r)$9

With $\varepsilon_r(\mathbf r)$0, the self-imaging distance is

$\varepsilon_r(\mathbf r)$1

a consequence of equally spaced modal propagation constants in the parabolic index well (Arabí et al., 2017). In thermally induced GRIN laser media, pump absorption generates an axially varying radial parabola,

$\varepsilon_r(\mathbf r)$2

and the corresponding paraxial ray equation becomes

$\varepsilon_r(\mathbf r)$3

which admits a Bessel-equation reduction and a closed-form ABCD matrix in terms of Bessel and Neumann functions (Kalantarifard et al., 23 Mar 2026).

These formulations show that “GRIN” is not restricted to one canonical profile. It includes radial laws such as Luneburg and hyperbolic-secant lenses, transverse parabolas in fibers and thermally loaded crystals, and fully three-dimensional analytic or tabulated functions in acoustic design (Bagheri et al., 19 Jan 2026, Danawe et al., 2023, Arabí et al., 2017, Xie et al., 2018).

3. Discretization, CAD automation, and fabrication constraints

A second defining feature of modern GRIN research is the replacement of ideal continuous profiles by subwavelength discretizations. Xie and Cummer’s GRIPP, “GRadient Index Pick-and-Place,” formalizes this for 3D-printable acoustic devices: the algorithm receives a spatial distribution of refractive index and a pre-defined library of gradient-index unit cells, discretizes the design volume into an $\varepsilon_r(\mathbf r)$4 Cartesian grid of cubic cells of side $\varepsilon_r(\mathbf r)$5, assigns each cell the library geometry whose effective index $\varepsilon_r(\mathbf r)$6 minimizes $\varepsilon_r(\mathbf r)$7, performs Boolean union, and exports STL. The stated complexity is $\varepsilon_r(\mathbf r)$8 nearest-neighbor searches plus $\varepsilon_r(\mathbf r)$9 geometry-placement operations, with the mapping summarized as “continuous refractive-index profile $\mu_r=1$0 discretized grid $\mu_r=1$1 discrete cell selection $\mu_r=1$2 composite 3D geometry $\mu_r=1$3 STL” (Xie et al., 2018).

Millimeter-wave GRIN lens antennas use analogous but platform-specific discretizations. One wideband flat-lens method constructs a library of intrinsically matched unit-cells containing an input matching taper, a phase-delay core, and an output matching taper; each ring of the lens is then assigned an optimal unit-cell according to the local incident angle $\mu_r=1$4 and required phase correction $\mu_r=1$5 (Garcia et al., 2020). A related ultra-wideband taper-design method treats practical uniform-thickness layers as a non-commensurate transmission-line problem and derives

$\mu_r=1$6

to equalize the electrical length of the ideal commensurate taper and the fabricated non-commensurate taper (Wang et al., 2021).

A common misconception is that GRIN devices require a physically continuous gradient. The cited implementations instead rely on cubic cells, concentric rings, drilled hexagonal lattices, gyroid voxels, or other meta-atoms, provided the unit-cell period remains electrically small (Xie et al., 2018, Garcia et al., 2020, Wang et al., 2023). The high-frequency-limit study makes this condition explicit for 3D-printed Luneburg antennas: four 10 cm lenses with gyroid unit-cells of 12.5, 10, 7.5, and 5 mm exhibited maximum frequencies of 20, 25, 33, and $\mu_r=1$7 GHz, leading to the rule of thumb that the smallest repeating sub-unit should satisfy “sub-unit period $\mu_r=1$8” (Wang et al., 2023).

Fabrication limits follow directly from this discretization strategy. GRIPP notes that the index range is limited by the unit-cell library, that library granularity $\mu_r=1$9 sets discretization error, and that Boolean union can become computationally heavy for very large $n(\mathbf r)$0 (Xie et al., 2018). The mm-wave modified Luneburg lens likewise reports that a gyroid voxel size of 2 mm, about $n(\mathbf r)$1 at 60 GHz, retains effective-medium behavior up to about 62.8 GHz but introduces phase errors that lower realized gain by about 6 dB versus a continuous gradient (Bagheri et al., 19 Jan 2026).

4. Representative devices and measured performance

The range of GRIN applications in the cited literature is unusually broad. In 3D acoustics, GRIPP-generated structured lenses were compared against an “ideal” continuous-index lens at 40 kHz. Both produced the same “bent” focusing behavior; the focal spot location agreed within $n(\mathbf r)$2, the beamwidth was matched to within $n(\mathbf r)$3, and impedance mismatch caused only minor ripple $n(\mathbf r)$4 in transmission (Xie et al., 2018). This supports the claim that automated discretization can reproduce the performance of an analytically specified GRIN device when the unit-cell library is adequate.

In healthcare radar, the modified Multi–Radar Modified Luneburg Lens integrates five Infineon BGT60TR13C bistatic FMCW radar modules operating in the 58–63 GHz band around a 100 mm-diameter spherical lens. The rod-based perturbation transforms the classical single focal point into a distributed focal arc, producing five fixed beams spaced by $n(\mathbf r)$5, a half-power beamwidth of approximately $n(\mathbf r)$6, and $n(\mathbf r)$7 total angular coverage. Measured TX power shows about +12 dB enhancement across 58–63 GHz, realized gain peaks near 28 dB for a single radar and 27 dB for concurrent five-radar operation, sidelobes are below $n(\mathbf r)$8 dB, and inter-sector isolation exceeds 15 dB. In assisted-living fall-detection experiments, the system achieved continuous detection across all five zones, near-zero false alarms $n(\mathbf r)$9 in repeated trials), and reliable fall detection within 10 s of signal loss (Bagheri et al., 19 Jan 2026).

In quantum sensing, a 3D-printed Luneburg-type metamaterial GRIN lens was used to enhance a Cesium-vapor Rydberg RF receiver. Anechoic-chamber characterization gave a peak focusing gain of 8.42 dB at 3.6 GHz, while the Rydberg measurement showed Autler–Townes peak splittings increasing from $n(r)=\sqrt{\,2-(r/R)^2\,},$0 to $n(r)=\sqrt{\,2-(r/R)^2\,},$1 MHz at 2.2 GHz and from $n(r)=\sqrt{\,2-(r/R)^2\,},$2 to $n(r)=\sqrt{\,2-(r/R)^2\,},$3 MHz at 3.6 GHz, corresponding to about 6 dB gain and a factor-of-two improvement in minimum detectable field (Tishchenko et al., 3 Dec 2025).

Integrated and structural-wave platforms show the same design logic in different geometries. GRIN-lens-based waveguide splitters based on the generalized Maxwell’s fisheye produced $n(r)=\sqrt{\,2-(r/R)^2\,},$4, $n(r)=\sqrt{\,2-(r/R)^2\,},$5, and $n(r)=\sqrt{\,2-(r/R)^2\,},$6 branching angles, while an Eaton-lens design yielded a 90$n(r)=\sqrt{\,2-(r/R)^2\,},$7 splitter; full-wave results gave branch efficiencies around 29–35% for the fisheye cases and about 41% for the truncated Eaton splitter (Badri et al., 2019). Danawe and Tol’s conformal GRIN-PC lenses on steel pipes and polymer cones combined hyperbolic-secant profiles, coordinate mappings, and anisotropy-corrected ray tracing; the conformal theory predicted focal location within $n(r)=\sqrt{\,2-(r/R)^2\,},$8 of simulation, and experiments on 3D-printed cones showed spot widths matching simulations within $n(r)=\sqrt{\,2-(r/R)^2\,},$9 (Danawe et al., 2023).

5. Propagation dynamics, instability, and reconfigurability

Beyond static lensing, GRIN media support distinctive propagation dynamics. In homogeneous GRIN multimode fibers, equally spaced modal propagation constants cause periodic self-imaging at $\varepsilon_{\rm LL}(r)=2-(r/R)^2$0 (Arabí et al., 2017). When the core radius is weakly modulated as

$\varepsilon_{\rm LL}(r)=2-(r/R)^2$1

with the modulation period close to the self-imaging distance, the cited work reports a “Moiré-like” beating of the fast self-imaging period and the slow modulation. This modifies geometric parametric instability, splitting each principal GPI band into a comb of sub-bands indexed by $\varepsilon_{\rm LL}(r)=2-(r/R)^2$2 (Arabí et al., 2017).

The optical-propagation study on linear and quadratic GRIN media gives a different but related perspective. By exact operator factorization, propagation in a linear GRIN can be written as free-space evolution followed by a phase and a transverse shift, and propagation in a quadratic GRIN as free-space evolution by an effective “time” $\varepsilon_{\rm LL}(r)=2-(r/R)^2$3 followed by a squeeze and a quadratic phase. This is the basis for the paper’s statement that one may “export” free-space solutions into these GRIN media. The Bohm potential,

$\varepsilon_{\rm LL}(r)=2-(r/R)^2$4

then modifies the effective optical potential according to $\varepsilon_{\rm LL}(r)=2-(r/R)^2$5 (Asenjo et al., 2020). This suggests a wave-optical interpretation of GRIN propagation in which diffraction itself renormalizes the material lensing.

Thermally induced GRIN media introduce axial nonuniformity rather than periodic modulation. With Beer–Lambert absorption, the GRIN strength decays as $\varepsilon_{\rm LL}(r)=2-(r/R)^2$6, and the exact ABCD matrix recovers both the constant-GRIN limit,

$\varepsilon_{\rm LL}(r)=2-(r/R)^2$7

and the uniform-medium limit $\varepsilon_{\rm LL}(r)=2-(r/R)^2$8, $\varepsilon_{\rm LL}(r)=2-(r/R)^2$9, $n(r)=n_0\,\operatorname{sech}(\alpha r),$0, $n(r)=n_0\,\operatorname{sech}(\alpha r),$1 (Kalantarifard et al., 23 Mar 2026). The corresponding Bessel–Gaussian beam analysis shows that larger axial decay confines GRIN-induced refocusing and self-imaging to the entrance region of the crystal.

A separate line of work treats GRIN not as fixed structure but as a tunable basis expansion. The mechanically reconfigurable lens concept expresses the transverse permittivity profile as a superposition of Chebyshev-polynomial components across laterally shiftable corrugated layers. In the reported three-shift scenarios, geometrical optics predicted focal positions of about 134 mm, 125 mm, and 119 mm, while full-wave simulations located peaks at about 110 mm, 87 mm, and 79 mm; the discrepancy was attributed to modeling approximations and structural granularity (Kaboutari et al., 26 Jun 2025).

6. Other technical meanings of GRIN

Outside wave physics, GRIN is an acronym with several unrelated meanings. In large-language-model training, GRIN denotes GRadient-INformed MoE, a sparse Mixture-of-Experts training method that combines sparse gradient estimation for discrete routing with a parallelism configuration that avoids token dropping. The reported top-2 $n(r)=n_0\,\operatorname{sech}(\alpha r),$2B MoE activates only 6.6B parameters per token, outperforms a 7B dense model, matches the performance of a 14B dense model trained on the same data, and achieves 79.4 on MMLU, 83.7 on HellaSwag, 74.4 on HumanEval, and 58.9 on MATH (Liu et al., 2024).

In monocular 3D reconstruction, GRIN names a diffusion-based depth estimator that is explicitly designed to ingest sparse, unstructured supervision. It uses image features with 3D geometric positional encodings, a RIN backbone with local and global conditioning tokens, and metric-depth supervision without post hoc scale alignment when intrinsics are available. Across eight indoor and outdoor datasets, the paper reports a new state of the art in 22 of 24 zero-shot metric depth metrics; on KITTI, for example, AbsRel is 0.046, RMSE is 2.25 m, and $n(r)=n_0\,\operatorname{sech}(\alpha r),$3 is 0.983 (Guizilini et al., 2024).

In library infrastructure, GRIN denotes the Google Return Interface, the platform through which partner libraries retrieve scanned images, OCR text, and metadata from Google Books. The technical report on GRIN Transfer, developed by the Institutional Data Initiative at Harvard Law School Library, describes authenticated URLs under https://books.google.com/libraries/{LIBRARY_DIRECTORY}/…, a global rate limit of $n(r)=n_0\,\operatorname{sech}(\alpha r),$4 requests/sec, and a maximum conversion queue size of $n(r)=n_0\,\operatorname{sech}(\alpha r),$5 pending barcodes. GRIN Transfer is implemented in Python 3.12+, uses a SQLite database and a TokenBucket rate limiter, and adds ETag-based idempotency, resumable sync, and support for both local filesystems and S3-compatible storage (Daly et al., 14 Nov 2025).

The word “grin” also appears literally, rather than acronymically, in quantum foundations. Das and Pati’s “Teleporting Grin of a Quantum Cheshire Cat without cat” uses “grin” for the circular polarization degree of freedom of a photon under suitable pre- and post-selection. In the stated weak-value analysis, the photon is found in the left arm, with $n(r)=n_0\,\operatorname{sech}(\alpha r),$6 and $n(r)=n_0\,\operatorname{sech}(\alpha r),$7, while the polarization is found in the right arm, with $n(r)=n_0\,\operatorname{sech}(\alpha r),$8 and $n(r)=n_0\,\operatorname{sech}(\alpha r),$9. The resulting disembodied teleportation protocol succeeds on the postselected subensemble with probability $\Omega\subset\mathbb R^3$0 (Das et al., 2019).

The broad lesson of this acronymic dispersion is simply that GRIN is not a univocal term. In arXiv-adjacent technical usage it most often names gradient-index media and devices, but it also labels specific architectures, software systems, and conceptual protocols whose only commonality is the four-letter string itself (Liu et al., 2024, Daly et al., 14 Nov 2025, Das et al., 2019).