SpectralLens: Spectrum-Centric Sensing

• SpectralLens is defined as methods that treat spectral structure—such as wavelength-dependent focusing and phase shifts—as the primary coordinate for sensing and imaging.
• It integrates tailored optical hardware and computational inversion techniques to extract multispectral and angular-spectral information with high resolution.
• Applications range from compact hyperspectral cameras and astronomical imaging to machine learning diagnostics that leverage spectral aggregation of activations and gradients.

SpectralLens is a recurrent label in recent research for systems that treat spectral structure as a primary degree of freedom rather than as a secondary attribute of imaging or inference. In optics, the term denotes devices or architectures in which wavelength-dependent focusing, retardation, or blur is deliberately engineered so that spatial and spectral information can be acquired or separated within the same hardware stack. In computational sensing, it denotes calibrated forward operators whose spectral fingerprints are inverted to recover spectra or multispectral image cubes. In astronomy and plasma lensing, it names spectral inversion frameworks tied to physically known propagation laws. In machine learning, it has been repurposed for spectral aggregation of explanations and for activation/gradient-spectrum diagnostics of optimization (Zhou et al., 2023, Liang et al., 2024, Kar et al., 2019, Turyshev et al., 2022, Mehrpanah et al., 14 Aug 2025, Liu et al., 7 May 2026).

1. Scope and unifying idea

The literature does not define SpectralLens as a single standardized apparatus. Instead, the name is applied to several technically distinct constructions that all use a spectrum-dependent transfer function as the main operator of sensing, focusing, inversion, or diagnosis.

Research domain	Core mechanism	Representative papers
Planar and multifunctional optics	Wavelength-dependent phase or transmission in a compact optic	(Zhou et al., 2023, Britton et al., 2020, Saragadam et al., 2021, Qadri et al., 1 Dec 2025)
Computational spectral sensing	Calibrated spectral fingerprints inverted by optimization or learning	(Liang et al., 2024, Wang et al., 2017, Kar et al., 2019, Oktem et al., 2020, Thomas et al., 14 Aug 2025)
Astronomical, mathematical, and ML diagnostics	Spectral laws used for inversion or internal-state analysis	(Turyshev et al., 2022, Tuntsov et al., 2015, Fan et al., 2022, Mehrpanah et al., 14 Aug 2025, Liu et al., 7 May 2026)

A common misconception is that SpectralLens denotes only a compact hyperspectral camera. The papers are broader. Some implementations are single-shot, some require voltage-sequenced or focus-diverse measurements, some rely on explicit scanning in an image plane, and some are not optical devices at all. This suggests that the term functions best as an umbrella label for methods that make a spectrum—optical, radio, Fourier, or singular-value—the principal coordinate of the problem.

2. Planar, dual-band, and programmable optical spectral lenses

One major line of work uses planar optics to co-locate focusing and spectral selection. The liquid-crystal spectral lens in "Spectral lens enables a minimalist framework for hyperspectral imaging" replaces the conventional imaging lens of a camera with a 2.12 mm × 2.12 mm planar LC lens and a standard monochrome CMOS sensor. Its azimuth $\theta$ encodes the geometric phase $\phi(x,y,\lambda)=2\theta$, while the polar angle $\alpha$ controls a voltage-dependent retardation and therefore a spectral response $SR(V_i,\lambda)$. The forward model is

$$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$

Experimentally, the system produced 500×500-pixel hyperspectral cubes with 29 spectral channels across 550–700 nm and object-space spatial resolution of 31 μm, while the PSF remained highly correlated across voltage settings, with Pearson correlation coefficients $>0.9$ for most voltages (Zhou et al., 2023).

A related but static implementation appears in "Compact dual-band multi-focal diffractive lenses". There, a sub-wavelength-thickness hydrogenated amorphous silicon microlens simultaneously focuses $632.8$ nm and $808$ nm to $80$ μm and $110$ μm, respectively, using a $\phi(x,y,\lambda)=2\theta$0 μm aperture, a minimum feature size of $\phi(x,y,\lambda)=2\theta$1 nm, and a fabricated thickness of $\phi(x,y,\lambda)=2\theta$2 nm. Measured focusing efficiencies were $\phi(x,y,\lambda)=2\theta$3 for the visible focus and $\phi(x,y,\lambda)=2\theta$4 for the near-IR focus, and broadband imaging showed that the two focal planes can encode depth and spectral content separately (Britton et al., 2020).

Programmability can be pushed further with phase spatial light modulators. In "Programmable Spectral Filter Arrays using Phase Spatial Light Modulator", a crossed-polarizer LC-SLM behaves as a spatially varying spectral filter array with per-pixel transmission

$\phi(x,y,\lambda)=2\theta$5

That system operated over 400–1000 nm, achieved 1024×1024 spatial sampling with 53 spectral bands over 420–940 nm, and reported laser-measured spectral FWHM of 23.9 nm at 532 nm, 31.9 nm at 635 nm, and 69.0 nm at 850 nm. Because unintended phase modulation introduced aberrations, the work also derived “good patterns” that bound phase gradients and used a learned restoration network to correct residual blur and vignetting (Saragadam et al., 2021).

These planar systems show that spectral-lens behavior does not require lateral dispersion by a grating or prism. Spectral information can instead be routed to different axial planes, encoded into voltage-dependent amplitude modulation, or imposed as a programmable spatially varying transmission law.

3. Computational spectral encoding and inverse reconstruction

A second major lineage treats SpectralLens as a calibrated forward model plus inverse solver. In the PhyspeNet lensless speckle reconstructive spectrometer, a BaSO$\phi(x,y,\lambda)=2\theta$6-coated integrating sphere generates wavelength-specific speckle fingerprints, calibration produces a transmission matrix $\phi(x,y,\lambda)=2\theta$7, and reconstruction solves

$\phi(x,y,\lambda)=2\theta$8

without pre-training. The experimental system used a 3 mm output port and a 5 cm integrating-sphere–camera distance, resolved dual peaks separated by 2 pm, verified 1 pm single-line stepping, and achieved a 40 nm operating range from 1525–1565 nm with average $\phi(x,y,\lambda)=2\theta$9 and average reconstructed SNR $\alpha$0 dB (Liang et al., 2024). The paper is explicit that training-free does not mean calibration-free: the stability of $\alpha$1 is the central requirement.

Single-frame angular–spectral multiplexing appears in "Computational snapshot angular-spectral lensless imaging". There, a transmissive diffractive element is placed directly in front of an image sensor, and each $\alpha$2 pair produces a distinctive angular–spectral point-spread function. The discrete model is

$\alpha$3

The proof-of-concept recovered one-dimensional angle and wavelength from a single frame with resolutions of $\alpha$4 and 6 nm, using $\alpha$5 angular samples from $\alpha$6 to $\alpha$7 and $\alpha$8 wavelength samples from 460–640 nm (Wang et al., 2017).

Diffractive compressive imagers push the same idea into multispectral imaging. In "Compressive Spectral Imaging with Diffractive Lenses" and "High-resolution Multi-spectral Imaging with Diffractive Lenses and Learned Reconstruction", a photon sieve or related diffractive optic supplies wavelength-dependent focusing via

$\alpha$9

For visible spectral imaging, $SR(V_i,\lambda)$0 monochrome measurements reconstructed $SR(V_i,\lambda)$1 bands over 410–710 nm with PSNR up to 34.19 dB, SSIM up to 0.88, and SAM down to $SR(V_i,\lambda)$2. In the EUV regime, a photon sieve with $SR(V_i,\lambda)$3 mm and $SR(V_i,\lambda)$4 μm separated the lines at 33.28, 33.42, and 33.54 nm, and an unrolled learned reconstructor reached 35.77 dB PSNR and 0.92 SSIM at 30 dB input SNR (Kar et al., 2019, Oktem et al., 2020).

The same forward-operator logic extends to refractive optics. For Earth observation, "Compressive Spectral Imaging in View of Earth Observation Applications" proposes replacing the diffractive lens with a conventional refractive lens and exploiting longitudinal chromatic aberration, with a sensing model $SR(V_i,\lambda)$5 and basis-pursuit recovery solved by Douglas–Rachford. In simulation, $SR(V_i,\lambda)$6 measurements reconstructed $SR(V_i,\lambda)$7 bands from 420–700 nm, corresponding to $SR(V_i,\lambda)$8, and the RGB reconstruction reached 35.46 dB PSNR (Thomas et al., 14 Aug 2025).

Across these papers, SpectralLens denotes not merely a lens element but an end-to-end operator: encoded measurement diversity, a calibrated spectral response, and an inversion procedure that recovers the latent spectral quantity of interest.

4. Broadband, nonlinear, and adaptive implementations

At the broadband extreme, SpectralLens denotes monolithic optics designed to suppress chromatic variation over unusually large spans. "Visible to Longwave-infrared imaging via an inverse-designed monolithic lens" reports a potassium bromide optic whose microrelief phase is

$SR(V_i,\lambda)$9

The lens has diameter 19 mm, focal length approximately 22.5 mm, thickness approximately 3 mm, and field of view approximately $$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$0, and it provides near-achromatic focusing from 0.45–14 μm, corresponding to a fractional bandwidth of approximately 1.9. One device supported 46 channels from 400–850 nm in 10-nm steps without refocus, 12 additional SWIR–MWIR–LWIR channels, simultaneous visible–LWIR and SWIR–LWIR imaging, 6 mrad angular resolution at 81 m, and LWIR imaging of aircraft engines at $$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$1 km (Qadri et al., 1 Dec 2025).

A different route replaces the physical lens with a nonlinear virtual one. In "Nonlinear virtual lens for programmable and multispectral infrared upconversion imaging", a phase-only SLM writes a Fresnel-zone-lens phase onto a 1550 nm pump; the pump and signal fields interact in an ultrathin lithium-niobate film of thickness approximately 600 nm; and the sum-frequency field inherits the programmed phase. The focal law is

$$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$2

With $$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$3 μm, $$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$4 mm, $$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$5 mm, and $$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$6 px, the measured SFG focal length changed from approximately 12.2 mm at $$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$7 nm to approximately 9.3 mm at $$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$8 nm, while pump-induced SHG at 775 nm focused at approximately 4.1 mm (Zheng et al., 8 Jan 2026).

Fluidic lenses supply an adaptive refractive counterpart. "Spectral Characterization of Optical Aberrations in Fluidic Lenses" used a Shack–Hartmann wavefront sensor to reconstruct wavelength-dependent aberrations of PDMS-based circular and elliptical fluidic lenses over 400–650 nm. The analysis combined a nonlinear membrane model, Snell-law ray tracing, and Zernike decomposition, and found dominant residual content in tilt, astigmatism, and trefoil-like modes, with amplitudes varying approximately as $$I_i(x,y)=\int_{\Lambda} SR(V_i,\lambda)\,[O(Mx,My,\lambda)\ast PSF(x,y,\lambda)]\,d\lambda.$$9 and increasing asymmetry-driven aberrations as ellipticity increased (Puentes et al., 2023).

In these broadband, nonlinear, and adaptive variants, the spectral lens is not simply a compact hyperspectral camera component. It can be a near-achromatic monolithic optic, a wavelength-dependent upconversion focus generator, or a tunable refractive element whose spectral aberrations are themselves the object of characterization.

5. Astronomical and lensing uses

SpectralLens also appears in contexts where the “lens” is gravitational or plasma-mediated rather than fabricated. "Spectrally resolved imaging with the solar gravitational lens" develops an end-to-end wave-optical model of the Solar Gravitational Lens convolved with a telescope pupil, including photon spectral density and per-pixel SNR. The focal region begins at

$>0.9$0

the on-axis gain is

$>0.9$1

and at 650 AU the effective-aperture intuition gives

$>0.9$2

The analysis assigns a central role to the external occulter: unlike an internal coronagraph with a Lyot stop and transmittance reduced to $>0.9$3, a starshade allows small telescopes such as $>0.9$4 m over optical to mid-IR wavelengths, with broadband SNR in 1 s of $>0.9$5 for $>0.9$6 m and effective mid-IR SNR $>0.9$7 in 300 s for $>0.9$8 m (Turyshev et al., 2022).

In radio astronomy, Tuntsov, Walker, and collaborators use SpectralLens as an inversion pipeline for interstellar plasma lenses. The method exploits the $>0.9$9 scaling of plasma refraction to map characteristic curves in a dynamic spectrum $632.8$0 to positions on the lens. Applied to the PKS 1939-315 extreme scattering event, the high-frequency 4.2–10.8 GHz data were fitted with demerit approximately $632.8$1 Jy$632.8$2 and r.m.s. flux residual approximately 7.1–7.2 mJy per 100 MHz channel, whereas extrapolation to 1.6–3.1 GHz failed because multiple imaging, source structure, and wave optics became important (Tuntsov et al., 2015).

A mathematically unrelated spectral viewpoint appears in "Spectral Analysis of the Kohn Laplacian on Lens Spaces". There the relevant “lens” is a CR lens space $632.8$3, not an imaging element, and the main result is an analog of Weyl’s law: $632.8$4 In three dimensions, two lens spaces with equal prime-order fundamental groups are Kohn-isospectral if and only if they are CR isometric (Fan et al., 2022).

These uses make the semantic breadth of SpectralLens especially clear. In astronomy it can denote a physical strategy for spectrally resolved imaging through the Solar Gravitational Lens, a radio inversion method for plasma lenses, or a spectral-analysis viewpoint on geometric lens spaces.

6. Spectral diagnostics in machine learning

In machine learning, SpectralLens has been reinterpreted as a diagnostic operator on model behavior. "On Spectral Properties of Gradient-based Explanation Methods" analyzes attribution methods in Fourier space and shows that gradient operators act as high-pass filters while input perturbations act as low-pass filters, yielding an effective band-pass bias. The proposed SpectralLens aggregates squared-gradient explanations across perturbation scales,

$632.8$5

and pairs this with ArgLens, the perturbation scale maximizing $632.8$6 per pixel. On ImageNet with ResNet-50, SL$632.8$7 achieved insertion $632.8$8 and deletion $632.8$9; on ViT-B/16 it achieved insertion $808$0 and deletion $808$1. The central claim is not that spectral bias disappears, but that aggregating across perturbation scales mitigates the inconsistency of fixed-$808$2 explanations and that squared gradients avoid the phase and sign ambiguities of raw gradients (Mehrpanah et al., 14 Aug 2025).

A second non-optical use appears in "Spectral Lens: Activation and Gradient Spectra as Diagnostics of LLM Optimization". There, SpectralLens measures trace-normalized activation covariance eigenvalues and per-sample gradient SVD spectra in decoder-only LLMs derived from the modded NanoGPT codebase. Across 12-, 36-, and 48-layer tiers, matched-loss runs with different batch sizes converged to systematically distinct activation spectra; early tail exponents $808$3 predicted downstream token efficiency with mean within-family Spearman $808$4 on d12 variants, pooled $808$5, and mean within-family $808$6 on d36/d48 runs; and joint activation–gradient spectra separated learning-side architectural gains from primarily execution-side throughput gains (Liu et al., 7 May 2026).

These machine-learning usages are sometimes treated as purely metaphorical. That characterization is too weak. Both define explicit spectral objects and operational procedures: one over Fourier spectra of explanations, the other over activation covariance and gradient singular values. What they do not share with optical SpectralLens systems is a wavelength axis. Their connection to the optical literature is methodological rather than physical: spectral decomposition becomes the lens through which hidden structure is made observable.