Prior-Channel Decompositions

• Prior-channel decompositions are a family of strategies that split a channel or its representation into parts, one capturing the channel law and the other incorporating fixed prior information.
• They are applied across diverse fields—information geometry, quantum information, imaging, and communications—with each domain adapting the method to its specific challenges.
• Advanced techniques include KL projection hierarchies, convex and barycentric decompositions, iterative scaling, and tensor factorization to enhance inference accuracy and efficiency.

Searching arXiv for the cited works and closely related uses of “prior-channel decomposition” to ground the article in current arXiv-indexed sources. Prior-channel decompositions denote a heterogeneous family of decomposition strategies in which a channel, a channel-like map, or a channel-derived representation is split into components whose meaning is fixed by a prior, a conditional support restriction, or a selected channel construction. The term is not standardized across disciplines. In information geometry it refers to projections of Markov kernels under a fixed input prior; in quantum information it appears in convex, barycentric, and support-restricted decompositions of channels; in imaging it refers to derived image channels such as dark, bright, or green channels that make a latent transmission or radiance decomposition tractable; and in communications it denotes estimation procedures that separate observation consistency from prior-driven denoising, score matching, or multilinear factorization (Perrone et al., 2015, Perrone et al., 2016, Łuczyński et al., 2018, Qiu et al., 8 Jul 2025).

1. Conceptual scope and recurring structure

Taken together, these works suggest a common pattern: a complicated inference or restoration problem is rewritten so that one part captures a channel law or a forward operator, while another part carries prior information in a form that is easier to manipulate. The “prior” may be an input distribution $p(x)$, a handcrafted image-channel statistic, a location-conditioned channel density, a low-rank tensor model, or a pre-/post-selected support constraint. The “channel” may be a Markov kernel $k(y\mid x)$, a CPTP map, an atmospheric-scattering model, a wireless propagation matrix, or a multichannel mixing operator (Perrone et al., 2015, Wang, 2015, Liang et al., 2021, Fesl et al., 2024).

This broad usage produces several non-equivalent meanings of decomposition. Some papers decompose a channel itself into extreme or generalized extreme components. Others decompose an estimation objective into a data-fidelity block and a prior block. Still others construct a special “prior channel” from raw observables, such as the dark channel, bright channel, or green channel, and then infer hidden variables from that derived representation. The literature therefore treats “prior-channel decomposition” less as a single theorem than as a recurrent design principle connecting prior information to a chosen channel representation (Cai et al., 2019, Kong et al., 2024, Qiu et al., 8 Jul 2025).

2. Prior-dependent geometry of channels

In the information-geometric formulation, a multi-input channel is a Markov kernel $k\in K(X;Y)$, and a prior $p\in P(X)$ induces the joint law $(pk)(x,y)=p(x)k(x;y)$. The relevant divergence between channels is the prior-weighted conditional KL divergence

$$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$

This makes the decomposition explicitly prior-dependent: changing $p$ changes the geometry, the projections, and the resulting interaction terms (Perrone et al., 2015).

The core construction is a nested hierarchy of exponential families of channels

$$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$

where $\mathcal E_i$ contains channels whose dependence on $Y$ can be generated using interactions among at most $k(y\mid x)$0 input variables. If $k(y\mid x)$1 denotes the KL projection under $k(y\mid x)$2, then the Pythagorean relation yields

$k(y\mid x)$3

Here $k(y\mid x)$4 is the amount of genuine $k(y\mid x)$5-th order interaction beyond lower orders. For two inputs, $k(y\mid x)$6 plays the role of pairwise synergy; for parity/XOR-type examples, all mutual information can concentrate in a single higher-order term (Perrone et al., 2015).

The companion algorithmic development extends iterative scaling from distributions to channels. For a fixed prior $k(y\mid x)$7, channel marginals such as $k(y\mid x)$8 define mixture families, while channel exponential families encode allowed interaction terms in $k(y\mid x)$9. The normalized channel-scaling update

$k\in K(X;Y)$0

iteratively projects onto intersections of such families and converges to the desired I-projection; by duality, this yields the rI-projection of a target channel onto a structured exponential family (Perrone et al., 2016). This supplies a practical route for computing the prior-dependent projections that underlie hierarchical interaction decompositions.

3. Quantum-channel decompositions

In finite-dimensional quantum information, one prominent decomposition problem is convex-geometric. A dimension-altering quantum channel is a CPTP map

$k\in K(X;Y)$1

with Kraus form

$k\in K(X;Y)$2

The set $k\in K(X;Y)$3 is convex, and Choi’s theorem characterizes extreme channels by linear independence of $k\in K(X;Y)$4. Since every extreme channel has Kraus rank at most $k\in K(X;Y)$5, the paper introduces generalized extreme channels $k\in K(X;Y)$6 and studies the conjectural decomposition

$k\in K(X;Y)$7

The work does not prove this decomposition in general, but develops circuit parameterizations and numerical approximations up to dimension four. For example, for $k\in K(X;Y)$8, Ansatz I and II use 65 parameters and achieve $k\in K(X;Y)$9 trace-distance error on Choi states, while Ansatz III uses 41 parameters and achieves $p\in P(X)$0 (Wang, 2015).

A distinct but related result is barycentric decomposition. For channels $p\in P(X)$1 with $p\in P(X)$2 separable and $p\in P(X)$3 finite-dimensional, every channel is the barycenter of a probability measure supported on extreme channels: $p\in P(X)$4 The paper is explicit that this is not a theory of prior channels, preprocessing orders, or compatibility; it is a Choquet-type decomposition over the extreme boundary of a convex set (Pellonpää et al., 2023).

A more directly prior-informed variant appears in channel decomposition with pre- and post-selection. Instead of decomposing an $p\in P(X)$5-qubit unitary channel on the full Hilbert space, the target is restricted to

$p\in P(X)$6

where $p\in P(X)$7 and $p\in P(X)$8 project onto the relevant input and output sectors. If $p\in P(X)$9 and $(pk)(x,y)=p(x)k(x;y)$0, the effective problem has dimension

$(pk)(x,y)=p(x)k(x;y)$1

so the basis size scales like $(pk)(x,y)=p(x)k(x;y)$2 rather than $(pk)(x,y)=p(x)k(x;y)$3. In the HHL example, a 5-qubit conditional target is reduced to a 1-qubit effective decomposition with only three nonzero product-channel terms and sampling overhead $(pk)(x,y)=p(x)k(x;y)$4 (Nagai et al., 2023). This is a literal decomposition induced by prior information about admissible support.

4. Image restoration as decomposition by prior channels

In image restoration, “channel” usually means a color channel or a derived channel representation rather than a communication channel. The classical example is the dark channel prior. Under the atmospheric model

$(pk)(x,y)=p(x)k(x;y)$5

the dark channel

$(pk)(x,y)=p(x)k(x;y)$6

acts as a handcrafted prior channel used to estimate transmission and atmospheric light. The underwater adaptation of this framework is especially instructive. Starting from

$(pk)(x,y)=p(x)k(x;y)$7

with attenuation

$(pk)(x,y)=p(x)k(x;y)$8

forward scattering

$(pk)(x,y)=p(x)k(x;y)$9

and backscatter

$$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$0

the main claim is that underwater DCP variants had modified the wrong part of the decomposition. The paper argues that the standard transmission decomposition

$$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$1

remains structurally valid underwater; the crucial adaptation should target estimation of the ambient/background light term $$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$2, not the core dark-channel machinery. Its final pipeline uses underwater white balance, atmospheric-light estimation on the white-balanced image, standard DCP transmission estimation, bilateral-filter refinement, restoration, and brightening (Łuczyński et al., 2018).

Subsequent work generalizes this logic across scales. Pyramid Fusion Dark Channel Prior keeps the DCP prior but applies it on a multi-level pyramid, using a fixed $$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$3 patch at every level and fusing the transmission maps. Atmospheric light at each level is estimated from the top $$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$4 dark-channel pixels, and the final transmission is fused with empirical lower:upper weights $$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$5 for indoor scenes and $$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$6 for outdoor scenes. On RESIDE SOTS, PF-DCP reports $$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$7 PSNR/SSIM, versus $$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$8 for baseline DCP, with about 30\% computational overhead relative to DCP (Liang et al., 2021).

Another line keeps the DCP decomposition but learns a correction layer on top of its latent variables. The multiple-linear-regression model rewrites

$$D_p(k\|m) := \sum_{x,y} p(x)\,k(x;y)\log\frac{k(x;y)}{m(x;y)}.$$9

as

$p$0

The dark channel, atmospheric light estimate, and rough transmission still come from DCP; only the recombination of these terms is learned. On RESIDE SOTS outdoor images, the reported performance is PSNR $p$1, SSIM $p$2, compared with DCP at PSNR $p$3, SSIM $p$4. In downstream hazy object detection, the paper reports $p$5 mAP for the MLDCP-preprocessed system, versus $p$6 for the DCP-preprocessed variant and $p$7 for the Mask R-CNN baseline (Li et al., 2021).

The same prior-channel idea also appears outside dehazing. ECPeNet for dynamic scene deblurring decomposes feature processing into a standard feature stream $p$8, a dark-channel-constrained branch $p$9, and a bright-channel-constrained branch $$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$0, with a loss

$$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$1

This is an implicit feature-level prior-channel decomposition rather than an explicit image-layer factorization (Cai et al., 2019). GCP-ID for denoising adopts a different asymmetry: the green channel is privileged because RGGB sensing provides twice the green sampling density. The method uses green-guided patch search, rewrites RGB patches into RGGB arrays, and applies t-SVD/PCA collaborative filtering. On SIDD validation it reports $$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$2 PSNR/SSIM for raw denoising with GCP-ID + CNN, and its ablation shows $$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$3 when green-guided search and RGGB representation are combined, versus $$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$4 or $$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$5 when each is used alone (Kong et al., 2024).

5. Communications and signal processing

In wireless estimation, prior-channel decomposition often means splitting a channel-estimation objective into an observation-consistency term and a prior term. For XL-MIMO, the regularized MAP objective

$$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$6

is rewritten with an auxiliary variable $$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$7 so that the $$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$8-update handles only the quadratic pilot-consistency term and the $$\mathcal E_0 \subset \mathcal E_1 \subset \cdots \subset \mathcal E_N = K(X;Y),$$9-update handles only the prior. The local prior is provided by a location-indexed channel knowledge map

$\mathcal E_i$0

implemented as a channel score function map whose denoisers approximate the score through Tweedie’s formula. The resulting PnP updates separate linear inversion from prior correction, and the experimental evidence shows average denoiser gains of $\mathcal E_i$1 dB for grid size $\mathcal E_i$2 m, $\mathcal E_i$3 dB for $\mathcal E_i$4 m, and $\mathcal E_i$5 dB for $\mathcal E_i$6 m. In the no-pilot regime, LMMSE reduces to the mean channel and gives NMSE $\mathcal E_i$7 dB, whereas CSFM-NN gives $\mathcal E_i$8 dB; the proposed method improves further in that regime (Qiu et al., 8 Jul 2025).

A different prior-driven split appears in diffusion-based MIMO estimation. With unitary pilots and full pilot observations, the measurement is first reduced by LS decorrelation to

$\mathcal E_i$9

then normalized and transformed into the angular domain,

$Y$0

where the channel is sparse or highly compressible. A lightweight diffusion model is trained in this domain and used as a deterministic denoiser, initialized at the reverse-diffusion step whose SNR best matches the observation. The online estimator is

$Y$1

The model uses $Y$2 parameters, much smaller than the cited score model at $Y$3, and reports up to about $Y$4 dB SNR gain over the score-based baseline on QuaDRiGa (Fesl et al., 2024).

Tensor methods produce another decomposition idiom. For hybrid analog/digital receivers with too few RF chains to observe the full antenna-domain channel, the wideband time-varying channel is written as a low-rank third-order tensor

$Y$5

and the compressed measurements satisfy

$Y$6

After CP decomposition, the spatial covariance is reconstructed as

$Y$7

The practical point is that the useful prior for hybrid beamforming is derived from a latent multilinear decomposition of space, frequency, and time, rather than from direct covariance observation; the paper reports that this tensor approach outperforms CS-based and MUSIC-based covariance estimation, especially in the low-SNR regime (Park et al., 2019).

At a more statistical level, the microscopic decoupling principle for large random linear vector channels shows that a high-dimensional channel with prior $Y$8 asymptotically decomposes into a bank of independent scalar Gaussian channels followed by scalar Bayesian inference with a prior-modulated posterior. The effective scalar law

$Y$9

summarizes the channel side, while the scalar posterior

$k(y\mid x)$00

carries the prior side (0801.4198). Multichannel nonstationary signal decomposition uses yet another decomposition strategy: the eigenvectors of the autocorrelation matrix

$k(y\mid x)$01

span the component subspace, and individual components are recovered by searching for linear combinations with minimum time-frequency concentration. In the statistical study of a 9-component noisy example, successful reconstruction required approximately $k(y\mid x)$02 sensors for $k(y\mid x)$03 and $k(y\mid x)$04 for $k(y\mid x)$05 (Stankovic et al., 2019).

6. Related constructions, boundaries, and limitations

Several closely related decomposition theories are structurally relevant but are not themselves prior-channel decompositions in the strict sense. The barycentric decomposition of quantum instruments is one example: it decomposes channels and instruments into extreme points of a convex set, but explicitly does not study prior channels, preprocessing orders, or compatibility (Pellonpää et al., 2023). Generalized pseudoskeleton decompositions are another. They characterize exact matrix and tensor factorizations such as

$k(y\mid x)$06

over arbitrary fields using generalized inverses, and are best viewed as algebraic analogues of selected-substructure decompositions rather than probabilistic prior-channel factorizations (Hamm, 2022).

Other decomposition theories sit at the interface between state and channel structure. A separable state of the form

$k(y\mid x)$07

with $k(y\mid x)$08 having independent images admits a canonical one-sided filtering

$k(y\mid x)$09

and belongs to the class precisely when the blocks $k(y\mid x)$10 are normal and mutually commute. In that case the decomposition with $k(y\mid x)$11 and distinct $k(y\mid x)$12 is unique, and under the Choi isomorphism this includes QC and CQ channels (Alfsen et al., 2012). By contrast, irreducible decompositions of quantum channels into recurrent subspaces and minimal enclosures classify stationary structure rather than prior-conditioned structure (Carbone et al., 2015). For continuous-variable Gaussian channels, a divergence-free Kraus construction based on finitely entangled two-mode squeezed states regularizes the Choi method, but its goal is operator-sum realization rather than prior-channel separation (Sabapathy et al., 2017).

Across these domains, the main limitations are likewise heterogeneous. In information geometry, the decomposition is explicitly prior-dependent and changes with $k(y\mid x)$13 (Perrone et al., 2015). In quantum convex decompositions, exact low-cardinality expansions remain conjectural or are only numerically supported in low dimensions (Wang, 2015). In imaging, dark-, bright-, and green-channel priors inherit failure modes tied to scene statistics, spectral bias, or sensor assumptions (Łuczyński et al., 2018, Kong et al., 2024). In wireless inference, location-specific or generative priors depend on representative training data and, in some cases, accurate side information such as position or SNR (Qiu et al., 8 Jul 2025, Fesl et al., 2024). Taken together, these limitations show that “prior-channel decomposition” is best understood as a family of domain-specific strategies for organizing inference around a prior-compatible channel representation, rather than as a single invariant formalism.