Hybrid Angular Spectrum (kWASM)

• Hybrid Angular Spectrum (kWASM) is a framework that retains the angular spectrum representation while adding model constraints, adaptive sampling, or basis transformations to address key limitations.
• It integrates methodologies like Mahalanobis geometry in massive MIMO, energy-retention in diffraction, and intrinsic coordinate conversions in vectorial optics to enhance performance.
• The approach clarifies common misconceptions by distinguishing its hybrid enhancements from kernelized or Wasserstein methods, ensuring precise application in optics, communications, and cosmology.

Searching arXiv for recent and relevant papers on “Hybrid Angular Spectrum (kWASM)” and adjacent uses of “angular spectrum” across optics, cosmology, and communications. Searching for the exact arXiv identifier and title provided, plus related “angular spectrum” papers. “Hybrid Angular Spectrum (kWASM)” is not a standardized term in the cited literature. Instead, the available work uses “angular spectrum” to denote several distinct objects: the optical angular spectrum in diffraction propagation, the angular power spectrum in massive MIMO channel inference, the vectorial $k$-space spectrum of electromagnetic beams, and the CMB angular power spectrum in loop quantum cosmology. The acronym kWASM does not appear in the cited works on diffraction, massive MIMO APS estimation, or vectorial beam propagation, and the loop-quantum-cosmology study is relevant to a hybrid angular-spectrum analysis only indirectly through its treatment of perturbative mass ambiguities (Wang et al., 2022, Cavalcante et al., 2020, Muys, 2011, Mohammadi et al., 23 May 2025). This suggests that “Hybrid Angular Spectrum” is best understood as a contextual umbrella for methods that retain an angular-spectrum representation while adding an additional structure such as model constraints, learned metrics, adaptive spectral support control, or intrinsic polarization coordinates.

1. Terminology and cross-domain scope

Across the cited works, “hybrid” refers to different kinds of augmentation rather than to a single universal algorithm. In massive MIMO, hybridity is explicit and denotes the fusion of model-based covariance consistency with data-driven statistical selection. In diffraction, the method remains an angular-spectrum propagator but adds energy-based spectral truncation and nonuniform FFT sampling. In vectorial optics, the central move is a change of basis from Cartesian laboratory coordinates to an intrinsic $(\mathbf{s},\mathbf{p},\mathbf{m})$ frame tied to each wavevector $\mathbf{k}$. In loop quantum cosmology, the relevant issue is not a hybrid propagator in the optical sense, but the sensitivity of observable angular spectra to the perturbative mass prescription used in the Mukhanov–Sasaki sector (Cavalcante et al., 2020, Wang et al., 2022, Muys, 2011, Mohammadi et al., 23 May 2025).

Domain	Spectrum object	Added structure
Massive MIMO	Angular power spectrum $\rho$	Model consistency $A\rho=r$ plus learned Mahalanobis geometry (Cavalcante et al., 2020)
Diffraction	$A(f_x,f_y)$ in ASM	Energy-retention boundary $f_{CE}$, transfer-function sampling control, type-3 NUFFT (Wang et al., 2022)
Vectorial optics	TE/TM spectral amplitudes	Intrinsic basis $(\mathbf{s},\mathbf{p},\mathbf{m})$ and exact nonparaxial propagation (Muys, 2011)
Loop quantum cosmology	Primordial and CMB angular power spectra	Alternative polymerized effective mass in the scalar perturbation sector (Mohammadi et al., 23 May 2025)

A persistent source of confusion is the temptation to interpret “kWASM” as implying kernelization or Wasserstein geometry. The massive-MIMO paper is explicit that there is no explicit Wasserstein distance and no RKHS-based kernel estimator in the method itself; the learned geometry is instead a Hilbertian metric induced by a Mahalanobis matrix $M_\alpha=(C+\alpha I)^{-1}$ (Cavalcante et al., 2020).

2. Hybrid APS estimation in massive MIMO

The clearest formal use of “hybrid” occurs in the estimation of angular power spectra from channel covariance matrices in massive MIMO. The forward model is linear: the covariance vector $r$ is related to a nonnegative APS $(\mathbf{s},\mathbf{p},\mathbf{m})$0 by $(\mathbf{s},\mathbf{p},\mathbf{m})$1 in the continuous setting and $(\mathbf{s},\mathbf{p},\mathbf{m})$2 after discretization on a grid of $(\mathbf{s},\mathbf{p},\mathbf{m})$3 angular samples. Nonnegativity is enforced by the cone $(\mathbf{s},\mathbf{p},\mathbf{m})$4, and the measured covariance induces an affine consistency set $(\mathbf{s},\mathbf{p},\mathbf{m})$5. The physically admissible candidates are therefore the elements of $(\mathbf{s},\mathbf{p},\mathbf{m})$6, but this set is generally nonunique because the inverse problem is ill-posed and $(\mathbf{s},\mathbf{p},\mathbf{m})$7 has nontrivial null space (Cavalcante et al., 2020).

The hybrid construction adds a dataset $(\mathbf{s},\mathbf{p},\mathbf{m})$8 of APS samples. From it, the empirical mean $(\mathbf{s},\mathbf{p},\mathbf{m})$9 and covariance $\mathbf{k}$0 are estimated, and the matrix

$\mathbf{k}$1

defines the Hilbertian metric

$\mathbf{k}$2

The lower-level problem remains model driven,

$\mathbf{k}$3

while the upper level selects, among its minimizers, the APS nearest to $\mathbf{k}$4 in the learned Mahalanobis geometry. The resulting hierarchical problem is

$\mathbf{k}$5

which is precisely the paper’s fusion of model-based and data-driven information (Cavalcante et al., 2020).

Two algorithms are proposed. The first is an iterative fixed-point method based on Haugazeau’s iteration,

$\mathbf{k}$6

with each proximal step reduced to a nonnegative least-squares problem. The second solves a single-shot regularized problem,

$\mathbf{k}$7

which also reduces to one NNLS solve. Algorithm 1 realizes the hierarchical projection exactly; Algorithm 2 is computationally cheaper but depends strongly on the tradeoff parameter $\mathbf{k}$8 (Cavalcante et al., 2020).

The paper also proves a structural nonuniqueness result. Under feasibility $\mathbf{k}$9 and an array assumption implying that $\rho$0 is a row of $\rho$1, the solution sets of three standard model-only formulations coincide: feasibility in $\rho$2, nonnegative least squares, and $\rho$3-minimization over $\rho$4. This removes the possibility that nonnegativity or sparsity alone can select a unique APS, and it motivates the learned second level (Cavalcante et al., 2020).

In simulations with a ULA, $\rho$5, carrier frequency $\rho$6 GHz, $\rho$7, noise variance $\rho$8, and 1,000 APS samples for estimating $\rho$9 and $A\rho=r$0, the hybrid methods outperform the pure model-based EAPM when training and test distributions match. Algorithm 2 with $A\rho=r$1 performs similarly to Algorithm 1, whereas $A\rho=r$2 performs noticeably worse. A neural network trained on 110,200 samples performs worse still and degrades dramatically under train/test distribution mismatch. Increasing $A\rho=r$3 toward $A\rho=r$4 makes the hybrid methods behave more like the model-only method and improves robustness under mismatch (Cavalcante et al., 2020).

3. Energy-controlled angular spectrum propagation

In diffraction theory, the “Controllable Energy Angular Spectrum Method” retains the standard angular-spectrum propagator but hybridizes it numerically through joint control of spectral support and spectral sampling. The propagated field is written in the usual form

$A\rho=r$5

with transfer function

$A\rho=r$6

The numerical difficulty is transfer-function undersampling at long propagation distance. After double zero-padding to $A\rho=r$7 samples, the frequency pitch is $A\rho=r$8, and the phase-sampling condition yields the critical distance

$A\rho=r$9

For $A(f_x,f_y)$0, standard ASM becomes undersampled (Wang et al., 2022).

The paper reviews two valid frequency boundaries. Band-limited ASM uses

$A(f_x,f_y)$1

while band-extended ASM uses

$A(f_x,f_y)$2

CEASM introduces an intermediate boundary $A(f_x,f_y)$3, constrained by

$A(f_x,f_y)$4

and chosen by an energy-retention criterion. If $A(f_x,f_y)$5 is the energy spectral density, then

$A(f_x,f_y)$6

must satisfy

$A(f_x,f_y)$7

where $A(f_x,f_y)$8. A near-field variant replaces $A(f_x,f_y)$9 by $f_{CE}$0 and can even push $f_{CE}$1 below $f_{CE}$2 (Wang et al., 2022).

Once $f_{CE}$3 is selected, CEASM determines the minimum sample count from the transfer-function sampling condition,

$f_{CE}$4

and, under $f_{CE}$5,

$f_{CE}$6

Because the new spectral support and sample count are generally not FFT-aligned, the implementation uses type-3 NUFFT for both forward and inverse transforms: $f_{CE}$7 The method is not presented as a physical hybridization of distinct wave equations. Rather, it combines ASM transfer-function propagation, band-validity constraints, source-spectrum energy adaptation, and NUFFT-based resampling (Wang et al., 2022).

The numerical results quantify the speed–accuracy tradeoff. In a 1D aperture experiment with $f_{CE}$8, $f_{CE}$9, $(\mathbf{s},\mathbf{p},\mathbf{m})$0, 758 aperture samples, and propagation from $(\mathbf{s},\mathbf{p},\mathbf{m})$1 to $(\mathbf{s},\mathbf{p},\mathbf{m})$2, CEASM and band-extended ASM remain highly consistent as distance increases, while adaptive-sampling ASM degrades at long range. CEASM has the same sample count as adaptive-sampling ASM up to about $(\mathbf{s},\mathbf{p},\mathbf{m})$3 for $(\mathbf{s},\mathbf{p},\mathbf{m})$4 because the energy already inside $(\mathbf{s},\mathbf{p},\mathbf{m})$5 exceeds $(\mathbf{s},\mathbf{p},\mathbf{m})$6 (Wang et al., 2022).

In a 2D triangle diffraction experiment at $(\mathbf{s},\mathbf{p},\mathbf{m})$7 and $(\mathbf{s},\mathbf{p},\mathbf{m})$8, Table I reports: adaptive-sampling ASM, SNR $(\mathbf{s},\mathbf{p},\mathbf{m})$9 dB, time $M_\alpha=(C+\alpha I)^{-1}$0 s, sampling pixels $M_\alpha=(C+\alpha I)^{-1}$1; band-extended ASM, SNR $M_\alpha=(C+\alpha I)^{-1}$2 dB, time $M_\alpha=(C+\alpha I)^{-1}$3 s, sampling pixels $M_\alpha=(C+\alpha I)^{-1}$4; CEASM, SNR $M_\alpha=(C+\alpha I)^{-1}$5 dB, time $M_\alpha=(C+\alpha I)^{-1}$6 s, sampling pixels $M_\alpha=(C+\alpha I)^{-1}$7. At $M_\alpha=(C+\alpha I)^{-1}$8 and $M_\alpha=(C+\alpha I)^{-1}$9, adaptive-sampling ASM gives SNR $r$0 dB with $r$1 samples, band-extended ASM gives $r$2 dB with $r$3, and CEASM gives $r$4 dB with $r$5. The paper therefore characterizes CEASM as achieving essentially the same accuracy as band-extended ASM with markedly fewer spectral samples (Wang et al., 2022).

4. Intrinsic-coordinate vectorial angular spectra

In vectorial beam theory, the relevant hybridizing move is a basis transformation in $r$6-space. The wavevector is decomposed as

$r$7

and the paper replaces the Cartesian basis by an intrinsic right-handed orthonormal system $r$8 defined by

$r$9

Equivalent definitions are

$(\mathbf{s},\mathbf{p},\mathbf{m})$00

This basis makes TE/TM structure explicit and automatically respects transversality, since both $(\mathbf{s},\mathbf{p},\mathbf{m})$01 and $(\mathbf{s},\mathbf{p},\mathbf{m})$02 are orthogonal to $(\mathbf{s},\mathbf{p},\mathbf{m})$03 (Muys, 2011).

The paper derives the vectorial angular spectrum from a scalar Hertz vector polarized along $(\mathbf{s},\mathbf{p},\mathbf{m})$04,

$(\mathbf{s},\mathbf{p},\mathbf{m})$05

whose boundary field at $(\mathbf{s},\mathbf{p},\mathbf{m})$06 has 2D Fourier transform

$(\mathbf{s},\mathbf{p},\mathbf{m})$07

Propagation in the half-space $(\mathbf{s},\mathbf{p},\mathbf{m})$08 is then given by the standard spectral factor $(\mathbf{s},\mathbf{p},\mathbf{m})$09,

$(\mathbf{s},\mathbf{p},\mathbf{m})$10

When reconstructed in intrinsic coordinates, the electromagnetic fields take a strikingly simple form: the electric field is proportional to $(\mathbf{s},\mathbf{p},\mathbf{m})$11 and the magnetic field to $(\mathbf{s},\mathbf{p},\mathbf{m})$12, which the paper identifies as a TM field. The general vectorial angular spectrum can therefore be written as

$(\mathbf{s},\mathbf{p},\mathbf{m})$13

with corresponding field decompositions

$(\mathbf{s},\mathbf{p},\mathbf{m})$14

$(\mathbf{s},\mathbf{p},\mathbf{m})$15

The relation $(\mathbf{s},\mathbf{p},\mathbf{m})$16 links the scalar Hertz spectrum directly to the TM angular spectrum (Muys, 2011).

For axially symmetric vector beams, the 2D Fourier integrals reduce to Bessel/Hankel forms. The TE field is purely azimuthal in real space, whereas the TM field has radial and longitudinal components weighted by $(\mathbf{s},\mathbf{p},\mathbf{m})$17 and $(\mathbf{s},\mathbf{p},\mathbf{m})$18. The paper applies this formalism to a nondiffracting vector Bessel beam,

$(\mathbf{s},\mathbf{p},\mathbf{m})$19

and emphasizes that this non-diffracting result is obtained without invoking the paraxial approximation. The method is also stated to be relevant for nonparaxial resonators, such as microresonators (Muys, 2011).

The paper does not itself supply a complete hybrid propagation algorithm. A plausible implication is that the intrinsic $(\mathbf{s},\mathbf{p},\mathbf{m})$20 basis provides a natural vectorial branch for any angular-spectrum scheme that wishes to combine scalar and exact nonparaxial propagation within the same $(\mathbf{s},\mathbf{p},\mathbf{m})$21-space framework (Muys, 2011).

5. Angular spectra in loop quantum cosmology

A distinct use of “angular spectrum” occurs in loop quantum cosmology, where the relevant observables are the primordial scalar power spectrum and the CMB temperature angular power spectrum. The paper studies a spatially flat FLRW universe in effective LQC with a Starobinsky potential and standard effective background dynamics,

$(\mathbf{s},\mathbf{p},\mathbf{m})$22

with the bounce at $(\mathbf{s},\mathbf{p},\mathbf{m})$23. The novelty is in the perturbation sector: the classical Mukhanov–Sasaki equation in the comoving gauge,

$(\mathbf{s},\mathbf{p},\mathbf{m})$24

is polymerized to generate an alternative effective mass function for scalar perturbations (Mohammadi et al., 23 May 2025).

The paper is explicit that this effective mass is distinct from the dressed metric mass, and that there is no choice of the polymerization function $(\mathbf{s},\mathbf{p},\mathbf{m})$25 that reproduces either the dressed metric or the hybrid effective mass function. The modified mass has the form

$(\mathbf{s},\mathbf{p},\mathbf{m})$26

with additional correction terms $(\mathbf{s},\mathbf{p},\mathbf{m})$27 relative to the dressed-metric mass, including a $(\mathbf{s},\mathbf{p},\mathbf{m})$28 contribution that can survive even in a kinetic-energy-dominated bounce (Mohammadi et al., 23 May 2025).

The specific polymerization ansatz is

$(\mathbf{s},\mathbf{p},\mathbf{m})$29

so the parameter $(\mathbf{s},\mathbf{p},\mathbf{m})$30 controls the ambiguity in the inverse-Hubble polymerization. In the kinetic-energy-dominated approximation,

$(\mathbf{s},\mathbf{p},\mathbf{m})$31

and at the bounce,

$(\mathbf{s},\mathbf{p},\mathbf{m})$32

The sign of the mass at the bounce therefore changes with $(\mathbf{s},\mathbf{p},\mathbf{m})$33, which the authors use to explain the sensitivity of the perturbation dynamics and the final spectra (Mohammadi et al., 23 May 2025).

For numerical evolution, the perturbations are initialized in the contracting branch in the Bunch–Davies vacuum, the primordial spectrum

$(\mathbf{s},\mathbf{p},\mathbf{m})$34

is evaluated at the end of inflation, and the resulting numerical $(\mathbf{s},\mathbf{p},\mathbf{m})$35 is supplied to CAMB to compute the temperature angular power spectrum. The representative choices are $(\mathbf{s},\mathbf{p},\mathbf{m})$36, $(\mathbf{s},\mathbf{p},\mathbf{m})$37, $(\mathbf{s},\mathbf{p},\mathbf{m})$38, $(\mathbf{s},\mathbf{p},\mathbf{m})$39, $(\mathbf{s},\mathbf{p},\mathbf{m})$40, and $(\mathbf{s},\mathbf{p},\mathbf{m})$41 (Mohammadi et al., 23 May 2025).

The primordial spectrum exhibits three regimes already familiar from dressed metric and hybrid calculations: suppressed infrared power for $(\mathbf{s},\mathbf{p},\mathbf{m})$42, an amplified oscillatory regime for $(\mathbf{s},\mathbf{p},\mathbf{m})$43, and an almost scale-invariant regime for $(\mathbf{s},\mathbf{p},\mathbf{m})$44. The distinctive new feature is a wave-packet structure in the region preceding the almost scale-invariant regime. It does not appear for negative $(\mathbf{s},\mathbf{p},\mathbf{m})$45 and small positive $(\mathbf{s},\mathbf{p},\mathbf{m})$46, begins to emerge around $(\mathbf{s},\mathbf{p},\mathbf{m})$47, and becomes higher, wider, and slightly left-shifted in $(\mathbf{s},\mathbf{p},\mathbf{m})$48-space as $(\mathbf{s},\mathbf{p},\mathbf{m})$49 increases. For very large $(\mathbf{s},\mathbf{p},\mathbf{m})$50, such as $(\mathbf{s},\mathbf{p},\mathbf{m})$51, the points accumulate more strongly in the upper part of the packet, amplifying the averaged power. The authors connect this to the near-bounce structure of $(\mathbf{s},\mathbf{p},\mathbf{m})$52: for $(\mathbf{s},\mathbf{p},\mathbf{m})$53 and $(\mathbf{s},\mathbf{p},\mathbf{m})$54 the bounce is a local minimum of the effective mass, while for $(\mathbf{s},\mathbf{p},\mathbf{m})$55 it becomes a local maximum with two nearby local minima (Mohammadi et al., 23 May 2025).

At the CMB level, all models considered agree very well with $(\mathbf{s},\mathbf{p},\mathbf{m})$56CDM for $(\mathbf{s},\mathbf{p},\mathbf{m})$57, while differences arise mainly at $(\mathbf{s},\mathbf{p},\mathbf{m})$58. For positive $(\mathbf{s},\mathbf{p},\mathbf{m})$59, the paper compares $(\mathbf{s},\mathbf{p},\mathbf{m})$60 and reports that $(\mathbf{s},\mathbf{p},\mathbf{m})$61 gives the best agreement with observations and the best-fit $(\mathbf{s},\mathbf{p},\mathbf{m})$62CDM curve. For $(\mathbf{s},\mathbf{p},\mathbf{m})$63, the $(\mathbf{s},\mathbf{p},\mathbf{m})$64 curve lies below the best-fit $(\mathbf{s},\mathbf{p},\mathbf{m})$65CDM curve and crosses near $(\mathbf{s},\mathbf{p},\mathbf{m})$66. Negative values $(\mathbf{s},\mathbf{p},\mathbf{m})$67 remain acceptable for $(\mathbf{s},\mathbf{p},\mathbf{m})$68 but grow above $(\mathbf{s},\mathbf{p},\mathbf{m})$69CDM at low $(\mathbf{s},\mathbf{p},\mathbf{m})$70, which is disfavored. The dependence on $(\mathbf{s},\mathbf{p},\mathbf{m})$71 is stated to be nonlinear, with observationally viable values of order $(\mathbf{s},\mathbf{p},\mathbf{m})$72 (Mohammadi et al., 23 May 2025).

Several caveats qualify the result. The analysis uses the Starobinsky potential only, initializes perturbations in the contracting-branch Bunch–Davies vacuum, studies the scalar sector and temperature spectrum only, and does not provide a full likelihood or MCMC analysis. The statement that $(\mathbf{s},\mathbf{p},\mathbf{m})$73 is “best-fit” is therefore a phenomenological comparison of plotted spectra rather than a reported $(\mathbf{s},\mathbf{p},\mathbf{m})$74, Bayesian evidence, or posterior constraint. The paper also does not provide a direct side-by-side hybrid $(\mathbf{s},\mathbf{p},\mathbf{m})$75 computation with the same numerical pipeline (Mohammadi et al., 23 May 2025).

6. Conceptual synthesis, misconceptions, and limits

The cited literature does not support treating “Hybrid Angular Spectrum (kWASM)” as a single canonically defined framework. Instead, the term spans at least four technical patterns. First, a hybrid angular spectrum can mean a model-constrained inverse estimator whose solution is selected by a data-learned geometry, as in massive MIMO APS reconstruction. Second, it can mean an adaptive spectral-support controller that preserves the ASM propagator while changing the retained bandwidth and sample count according to source energy. Third, it can denote a basis-hybridized vectorial formulation in which the angular spectrum is represented in TE/TM-like intrinsic coordinates rather than Cartesian components. Fourth, it can refer, more loosely, to the fact that observable angular spectra may be sensitive to a hidden modeling choice in the perturbation sector even when the background dynamics and inflationary potential are held fixed (Cavalcante et al., 2020, Wang et al., 2022, Muys, 2011, Mohammadi et al., 23 May 2025).

Three misconceptions are explicitly ruled out by the source material. The first is that kWASM denotes a kernelized or Wasserstein APS estimator: the massive-MIMO paper says otherwise and uses a Mahalanobis/Hilbertian metric instead (Cavalcante et al., 2020). The second is that CEASM is a hybrid propagator in the sense of switching between different wave equations or between spatial-domain and $(\mathbf{s},\mathbf{p},\mathbf{m})$76-space corrected solvers: it is still an ASM method whose novelty lies in energy-aware spectral truncation and type-3 NUFFT sampling (Wang et al., 2022). The third is that the vectorial intrinsic-coordinate paper already delivers a full hybrid algorithm: it supplies an exact free-space vectorial angular-spectrum formulation and emphasizes nonparaxial resonator relevance, but not a switching rule, anti-aliasing strategy, or multilayer propagation scheme (Muys, 2011).

In this literature, the most defensible use of the label is therefore methodological rather than taxonomic. A hybrid angular spectrum method is one that preserves an angular-spectrum core while introducing an additional constraint or representation that resolves a limitation of the baseline formalism: nonuniqueness in inverse APS estimation, oversampling or undersampling in diffraction propagation, loss of polarization structure in scalar optics, or insensitivity to perturbative ambiguities in cosmological angular observables. This suggests that any use of the acronym kWASM should be accompanied by an explicit statement of the hybridization mechanism being invoked, since the cited works do not establish a single accepted meaning for the term (Cavalcante et al., 2020, Wang et al., 2022, Muys, 2011, Mohammadi et al., 23 May 2025).