Full-Spectrum Modeling Overview

• Full-spectrum modeling is a methodological approach that retains the entire spectrum—whether optical, gamma-ray, graph eigenspectra, or astronomical—rather than reducing data to a few handcrafted features.
• It employs structured compressions and advanced algebraic formulations to preserve distributed spectral information, enabling effective inference across varied applications.
• The approach has demonstrated success in reducing errors in refractive-index sensing, enhancing gamma-ray spectrometry, and enabling full-shape analyses in cosmology while managing trade-offs inherent in high-dimensional data.

Searching arXiv for papers on “full-spectrum modeling” and closely related uses of the term. I’m going to look up relevant arXiv papers on “full-spectrum modeling” to ground the article in current usage across domains. Full-spectrum modeling denotes a family of methods that take an entire spectral object, full response curve, or complete spectral basis as the primitive quantity of inference, rather than reducing data to a small set of manually chosen features. In contemporary arXiv usage, the term spans full optical spectra in sensing, full pulse-height spectra in gamma-ray spectrometry, full paired Laplacian spectra in graph learning, native-resolution astronomical spectra, and full broadband clustering shapes in cosmology (Aalizadeh et al., 8 Apr 2025, Breitenmoser et al., 21 Jun 2025, Wang et al., 7 May 2026, Islam et al., 21 Jan 2026, Rodriguez-Meza et al., 2023). The common thread is not a single algorithm, but a commitment to preserving distributed spectral information that would be discarded by peak-picking, windowing, or low-dimensional summary statistics.

1. Semantic range and domain-specific meanings

The phrase “full-spectrum” is used across multiple technical literatures, but its referent changes with the object being modeled.

Domain	Full object being modeled	Representative form
Optical and chemical sensing	Entire absorption or transmission spectrum	PCA or multi-wavelength regression
Gamma-ray spectrometry	Entire pulse-height spectrum	Template matching with IRF-based forward models
Graph learning	Full paired eigenspectrum	Bivariate spectral filters over $U \otimes U$
Astronomy and cosmology	Native-resolution spectra or broadband power-spectrum shape	Variable-length spectral encoders or full-shape likelihoods
Wave and diffraction physics	Full spatial-frequency spectrum	Vector angular spectrum propagation

In refractive-index sensing, full-spectrum modeling means using the entire absorption spectrum of a metasurface sensor rather than a single wavelength intensity or a single resonance shift (Aalizadeh et al., 8 Apr 2025). In mobile gamma-ray spectrometry, it means predicting and matching the entire measured pulse-height spectrum, including photopeaks, Compton continua, escape features, and low-energy tails (Breitenmoser et al., 21 Jun 2025). In the “Full-Spectrum Graph Neural Network,” it means lifting signals from the node domain to the node-pair domain and filtering with a bivariate response $H(\lambda_i,\lambda_j)$ over the full paired spectrum (Wang et al., 7 May 2026). In astronomy, it means learning directly from full native-resolution spectra without truncation, resampling, or interpolation (Islam et al., 21 Jan 2026). In cosmology, the closely related expression “full-shape” refers to fitting the full broadband and BAO structure of the galaxy power spectrum rather than compressing the signal into $f\sigma_8$ and distance-ratio summaries (Rodriguez-Meza et al., 2023).

A plausible implication is that “full-spectrum modeling” is best understood as a methodological umbrella rather than a single formalism. What is preserved in full varies by field: wavelength-resolved intensities, pulse-height counts, eigenvalue-pair interactions, or broadband two-point statistics.

2. Recurrent mathematical structures

Despite this semantic diversity, several recurrent constructions appear across the literature. One is the replacement of a scalar feature by a structured high-dimensional representation. In refractive-index sensing, each spectrum is treated as a row of a centered matrix $X_c = X - 1\mu^T$, reduced by SVD, projected to $Z = X_cV_k$, and regressed via $y = Z\beta + \epsilon$ (Aalizadeh et al., 8 Apr 2025). In full-spectrum correlated $k$-distribution radiative transfer, the spectrum is reordered into a cumulative variable $g \in [0,1]$ and transmittance is written as

$T(a)=\int_0^1 \exp[-k(g)a]\,dg,$

with the Simple FSCK MLP predicting both $k(g)$ and the FSCK-3 product $H(\lambda_i,\lambda_j)$0 directly (Wang et al., 2024). In mobile gamma-ray spectrometry, the expected spectrum is assembled from directional flux banks and dynamic anisotropic instrument response functions,

$H(\lambda_i,\lambda_j)$1

so the entire spectral template becomes the forward-model output (Breitenmoser et al., 21 Jun 2025). In graph learning, full-spectrum lifting produces

$H(\lambda_i,\lambda_j)$2

which generalizes classical univariate spectral filters to the paired domain (Wang et al., 7 May 2026).

Another recurrent pattern is that “full-spectrum” rarely means “use every raw channel with no structure.” The dominant implementations instead combine full-spectrum retention with a model of spectral organization: PCA in sensing, $H(\lambda_i,\lambda_j)$3-space quadrature in radiative transfer, low-rank or polynomial parameterizations in graph operators, adaptive patching in astronomy, and explicit IRF factorization in spectrometry. This suggests that the central issue is not dimensional maximalism, but preserving the spectral degrees of freedom that carry task-relevant information while imposing a computationally tractable structure.

3. Optical, colorimetric, telluric, and dielectric spectroscopy

In optical sensing, full-spectrum modeling is often motivated by the inadequacy of single-feature readouts. For titanium and silicon meta-gratings used in refractive-index sensing, the absorption spectrum was compressed to $H(\lambda_i,\lambda_j)$4 principal components and mapped to refractive index by ordinary least squares under five-fold cross-validation. The resulting full-spectrum PCR mean squared errors were $H(\lambda_i,\lambda_j)$5 for Ti–TM, $H(\lambda_i,\lambda_j)$6 for Ti–TE, $H(\lambda_i,\lambda_j)$7 for Si–TM, and $H(\lambda_i,\lambda_j)$8 for Si–TE. Relative to the best single-feature baselines, the improvement factors were $H(\lambda_i,\lambda_j)$9 for Ti–TM, $f\sigma_8$0 for Ti–TE, $f\sigma_8$1 for Si–TM, and $f\sigma_8$2 for Si–TE (Aalizadeh et al., 8 Apr 2025). The underlying physical distinction is explicit in the paper: titanium spectra undergo broadband intensity modulation and form a near-linear low-dimensional manifold, whereas silicon spectra are dominated by narrow Mie resonances whose global evolution with refractive index is nonlinear.

A closely related result appears in colorimetric sensing. There, conventional single-wavelength analysis gave a best ten-fold cross-validated MSE of $f\sigma_8$3 at $f\sigma_8$4 nm, whereas forward feature selection on normalized transmission spectra reduced the error to $f\sigma_8$5 using twelve wavelengths, corresponding to a more than $f\sigma_8$6-fold reduction in MSE and an RMSE drop from $f\sigma_8$7 to $f\sigma_8$8 (Aalizadeh et al., 3 Sep 2025). The important methodological point is that the improvement came from modeling rather than hardware changes: the optical setup was unchanged.

In astronomical spectroscopy, full-spectrum fitting has long been used to avoid isolated-feature corrections. A telluric model computed with LBLRTM and HITRAN can be embedded directly into a joint full-spectrum fit of stellar spectrum, atmospheric transmission, line-spread function, and wavelength solution, eliminating the need for an A/B-type telluric standard star (Husser et al., 2013). The forward model multiplies stellar and telluric components, convolves them with the instrument LSF, and fits them directly to the observed science exposure. This is a model-based use of full-spectrum information: the whole spectrum, including telluric bands, becomes a calibration source rather than a nuisance to be masked.

Full-spectrum modeling can also refer to broadband material parameterization rather than inference from a dataset. For liquid water, a single causality-consistent dielectric model was assembled across microwave, IR, visible/UV, and soft X-ray energies up to $f\sigma_8$9 eV. At $X_c = X - 1\mu^T$0C the parameterization uses two Debye oscillators, seven IR oscillators, and twelve UV/X-ray oscillators; at $X_c = X - 1\mu^T$1C it uses two Debye oscillators, five IR oscillators, and eleven UV/X-ray oscillators, with the imaginary-axis response $X_c = X - 1\mu^T$2 constrained by Kramers–Kronig integration (Fiedler et al., 2019). Here “full-spectrum” means that the model covers the full electromagnetic response rather than a restricted band.

4. Gamma-ray spectrometry and radiative-transfer formulations

Gamma-ray spectrometry provides one of the clearest forward-modeling uses of the term. In mobile gamma-ray spectrometry, full-spectrum analysis writes the expected count-rate spectrum as a nonnegative superposition of templates,

$X_c = X - 1\mu^T$3

and matches measured counts using weighted least squares or a Poisson log-likelihood (Breitenmoser et al., 21 Jun 2025). The full-spectrum template is generated by convolving a dynamic, anisotropic instrument response function with a double-differential gamma-ray flux bank. Benchmarked against brute-force Monte Carlo, the method achieved an effective speedup of $X_c = X - 1\mu^T$4 with median spectral deviations below $X_c = X - 1\mu^T$5.

The broader airborne gamma-ray spectrometry program develops the same idea at larger scale. A Monte Carlo forward model predicts the full pulse-height spectrum for arbitrary gamma-ray fields, includes non-proportional scintillation physics, aircraft geometry, atmospheric transport, and detector response, and is accelerated by a PCE–PCA surrogate within a Bayesian inversion framework (Breitenmoser, 2024). The reported performance includes median relative deviation below $X_c = X - 1\mu^T$6 across the spectral band and Bayesian activity estimates with relative deviations below $X_c = X - 1\mu^T$7 at $X_c = X - 1\mu^T$8 s measurement time. In this usage, full-spectrum modeling replaces window-based analysis with a full detector-response operator.

In situ outdoor gamma monitoring during rainfall extends the same logic to transient backgrounds. A NaI(Tl) detector’s full spectrum is decomposed into Monte Carlo-derived templates for terrestrial backgrounds, cosmic continuum, positron annihilation, and precipitation-deposited Pb-214 and Bi-214. Activities are inferred every $X_c = X - 1\mu^T$9 s by Poisson maximum likelihood, and then regularized by a decay-chain temporal model. The static pre-rain fit showed deviance residuals typically within $Z = X_cV_k$0 Gaussian $Z = X_cV_k$1-equivalents across most of the spectrum; during rain events, deposited Pb-214 and Bi-214 activities were of order $Z = X_cV_k$2 kBq/m$Z = X_cV_k$3, and the on-ground ratio $Z = X_cV_k$4 approached the transient-equilibrium value $Z = X_cV_k$5 after rainfall ceased (Bandstra et al., 19 Aug 2025).

Gas radiative transfer uses a different but structurally related form. In the full-spectrum correlated $Z = X_cV_k$6-distribution method, the nongray spectrum is reordered into a cumulative variable $Z = X_cV_k$7, and radiative transfer is solved over a small quadrature in that reordered space. The Simple FSCK MLP predicts both $Z = X_cV_k$8 and $Z = X_cV_k$9 for a fixed pressure, reducing storage from more than $y = Z\beta + \epsilon$0 GB for the look-up table to about $y = Z\beta + \epsilon$1 MB per pressure. For generating $y = Z\beta + \epsilon$2 $y = Z\beta + \epsilon$3-distributions, the reported times were $y = Z\beta + \epsilon$4 s for the table, $y = Z\beta + \epsilon$5 s for the traditional FSCK MLP, and $y = Z\beta + \epsilon$6 s for the simple model; slab-test errors were below $y = Z\beta + \epsilon$7 (Wang et al., 2024). The distinctive point is that full-spectrum information is preserved after reordering rather than through direct wavelength-wise processing.

5. Frequency-aware machine learning, operators, and wave physics

In machine learning, “full-spectrum” has acquired several specialized meanings. In graph learning, the Full-Spectrum GNN lifts signals from nodes to node pairs and replaces univariate eigenvalue filters with bivariate filters over eigenvalue pairs. Classical spectral GNNs appear as a diagonal special case, while the full model can universally approximate node-pair signals under the stated conditions, is at most as expressive as Local 2-GNN for polynomial filters, and is particularly beneficial for heterophilic graph learning because it enables off-diagonal spectral interactions that diagonal spectral filters cannot realize (Wang et al., 7 May 2026). Here the “spectrum” is not a measurement spectrum but the full paired Laplacian eigenspectrum.

For learning-based full-waveform inversion, full-spectrum modeling is framed as protection against frequency entanglement. SPAMoE introduces a Spectral-Preserving DINO Encoder, concentric Gaussian spectral masks, adaptive frequency routing, and a mixture of FNO, MNO, and LNO experts. The encoder is paired with a theorem giving a lower bound on the high-to-low energy ratio of the prediction,

$y = Z\beta + \epsilon$8

under assumptions A1–A3 (Wang et al., 8 Apr 2026). On the ten OpenFWI sub-datasets, SPAMoE reduced the average MAE by $y = Z\beta + \epsilon$9 relative to the best officially reported OpenFWI baseline.

In privacy-preserving image learning, full-spectrum modeling is used almost adversarially: unlearnable perturbations should remain effective after arbitrary spectral filtering. Existing unlearnable examples were shown to fail under low-pass filtering, with CIFAR-10 ResNet-18 test accuracy rebounding to $k$0–$k$1, while FUSE held accuracy at $k$2 under the same filter (Cai et al., 25 Jun 2026). The method combines Random Spectral Masking with Cross-Band Guidance, and defines spectral equalization through entropy of the perturbation power spectrum. In this context, “full-spectrum” means spectrum-agnostic robustness rather than direct use of a measured spectrum.

Wave and diffraction physics offers yet another interpretation. In the full-vector angular spectrum method, any field is decomposed into its complete spatial-frequency spectrum, but each $k$3-mode is projected onto transverse $k$4 polarizations so that propagation obeys Maxwell transversality. The compact projector

$k$5

makes explicit that longitudinal-to-transverse projection is part of the full-vector treatment (Song et al., 20 May 2025). Interfaces enter as $k$6-space Fresnel filters, and rotationally symmetric systems can be evaluated in a few seconds. A related all-order use of “full spectrum” appears in medium-induced gluon radiation, where the full $k$7 and $k$8 spectrum is computed with realistic interaction potentials rather than truncated opacity or Gaussian approximations; the resulting calculation shows that GLV $k$9 overshoots the low-$g \in [0,1]$0 region while the harmonic-oscillator approximation misses the high-$g \in [0,1]$1 tail (Andres et al., 2021).

6. Native-resolution spectra, stellar full-spectrum fitting, and cosmological full-shape analysis

In astronomy, full-spectrum modeling often means that the observed spectrum itself is the learning object, at its original resolution and sampling. OmniSpectra implements this idea with adaptive patching of native-resolution spectra, sinusoidal global wavelength encoding, local positional embeddings through depthwise convolution, and validity-aware self-attention masks. It pretrains jointly on eight surveys, including DESI EDR, SDSS Legacy, BOSS, eBOSS, SEGUE-1, SEGUE-2, APOGEE DR17, and VIPERS (Islam et al., 21 Jan 2026). The backbone uses patch size $g \in [0,1]$2 with overlap $g \in [0,1]$3, embedding dimension $g \in [0,1]$4, six encoder layers, and six heads. Ablation results show validation reconstruction MSE $g \in [0,1]$5 for the full model, degrading to $g \in [0,1]$6 without wavelength embedding, which identifies physically grounded wavelength encoding as the dominant component.

Three-dimensional integral-field spectroscopy extends the same philosophy into the spatial dimension. In 3D full spectrum fitting, PNKR and a spatially correlated Bayes-LOSVD variant fit many spectra simultaneously rather than spaxel-by-spaxel. On mock IFS data over SNR $g \in [0,1]$7–$g \in [0,1]$8, spatially correlated Bayes-LOSVD improved LOSVD accuracy by about $g \in [0,1]$9 at SNR $T(a)=\int_0^1 \exp[-k(g)a]\,dg,$0 and about $T(a)=\int_0^1 \exp[-k(g)a]\,dg,$1 at SNR $T(a)=\int_0^1 \exp[-k(g)a]\,dg,$2, while PNKR recovered the most accurate kinematics overall for SNR $T(a)=\int_0^1 \exp[-k(g)a]\,dg,$3 but showed a metallicity bias of approximately $T(a)=\int_0^1 \exp[-k(g)a]\,dg,$4–$T(a)=\int_0^1 \exp[-k(g)a]\,dg,$5 dex (Jethwa et al., 5 Nov 2025). The broader methodological point is that “full-spectrum” in stellar population work typically implies pixel-space fitting of the full spectral energy distribution, often coupled to nonparametric LOSVD recovery.

In cosmology, the closely related term “full-shape” denotes direct modeling of the full anisotropic galaxy power spectrum or correlation function rather than compressed BAO+RSD summaries. The fkPT framework retains the scale-dependent growth rate $T(a)=\int_0^1 \exp[-k(g)a]\,dg,$6 explicitly and computes the redshift-space galaxy power spectrum in about $T(a)=\int_0^1 \exp[-k(g)a]\,dg,$7 s, enabling MCMC in modified-gravity models with scale-dependent linear growth (Rodriguez-Meza et al., 2023). Validation against MG-GLAM showed an F5 detection at about $T(a)=\int_0^1 \exp[-k(g)a]\,dg,$8 for effective volume $T(a)=\int_0^1 \exp[-k(g)a]\,dg,$9 and about $k(g)$0 for $k(g)$1, with baseline $k(g)$2.

DESI full-modeling studies in Fourier and configuration space make the same point operational. With Velocileptors in Fourier space, Full-Modeling and ShapeFit were found consistent within standard $k(g)$3CDM, with baseline $k(g)$4 for $k(g)$5 and $k(g)$6 on single-box covariances (Maus et al., 2024). In configuration space, the EFT-GSM pipeline fit the complete anisotropic correlation signal and recovered mock cosmologies within $k(g)$7 for a DESI Year 1-like volume, using baseline cuts $k(g)$8 and $k(g)$9 for $H(\lambda_i,\lambda_j)$00 and $H(\lambda_i,\lambda_j)$01 (Ramirez-Solano et al., 2024). In this literature, “full-spectrum” is effectively synonymous with retaining the broadband cosmological shape rather than only oscillatory or compressed summary information.

7. Limits, trade-offs, and recurring misconceptions

A common misconception is that full-spectrum modeling is automatically superior because it uses more data channels. The literature is more conditional. In refractive-index sensing, titanium meta-gratings showed dramatic gains because the spectrum changed smoothly and globally in amplitude, whereas silicon meta-gratings showed only modest gains because narrow resonances made the full spectral manifold globally nonlinear for a linear model (Aalizadeh et al., 8 Apr 2025). In graph learning, full-spectrum bivariate filtering expands the space of spectral interactions, but for degree-$H(\lambda_i,\lambda_j)$02 polynomial filters its distinguishing power is still upper-bounded by $H(\lambda_i,\lambda_j)$03-step Local 2-GNN refinement (Wang et al., 7 May 2026). In DESI cosmology, pushing to too-small scales produces parameter shifts because finite-order perturbation theory and nuisance models cease to be adequate; the recommended cuts remain conservative (Maus et al., 2024, Ramirez-Solano et al., 2024).

Another misconception is that “full-spectrum” means “no compression.” Many of the most effective implementations are highly structured compressions of the full object: $H(\lambda_i,\lambda_j)$04 principal components in metasurface sensing, twelve selected wavelengths in colorimetry, eight $H(\lambda_i,\lambda_j)$05-quadrature nodes in FSCK, validity-aware tokenization of arbitrary-length spectra in OmniSpectra, and low-rank or polynomial parameterizations in full-spectrum graph convolution (Aalizadeh et al., 8 Apr 2025, Aalizadeh et al., 3 Sep 2025, Wang et al., 2024, Islam et al., 21 Jan 2026, Wang et al., 7 May 2026). What they avoid is not compression per se, but lossy reduction to a single handcrafted feature that fixes the wrong inductive bias.

A third misconception is that the term always refers to physical spectra. In image protection it means robustness across the full spatial-frequency spectrum; in full-waveform inversion it means preserving both low- and high-frequency content of the recovered subsurface model; in full-vector diffraction it means treating all spatial-frequency modes with the correct vector constraints (Cai et al., 25 Jun 2026, Wang et al., 8 Apr 2026, Song et al., 20 May 2025). The phrase therefore denotes a commitment to whole-spectrum structure, not a specific sensor modality.

The broader record suggests that full-spectrum modeling is most effective when task-relevant information is distributed across the spectrum and when the model class respects the underlying physics or operator geometry. Where those conditions fail, the same literature repeatedly turns to hybrids: nonlinear regressors for resonance-dominated sensing, low-rank approximations for graph spectral operators, richer IRF state coverage for mobile gamma spectrometry, or conservative scale cuts and EFT counterterms in cosmological full-shape inference (Aalizadeh et al., 8 Apr 2025, Breitenmoser et al., 21 Jun 2025, Maus et al., 2024). This suggests that the enduring significance of full-spectrum modeling lies less in raw dimensionality than in the disciplined retention of spectral structure.