Field-Response Channel Model

• Field-response channel models are defined as operators that map excitation fields or source distributions to observed responses, offering a unified view of propagation.
• They apply across diverse domains such as electromagnetic propagation, XL-MIMO, RIS surfaces, molecular communication, and site-trained shadowing, integrating geometric and physical principles.
• These models enable inversion, sparse recovery, and learning-based estimation while addressing challenges like mixed far-/near-field effects and model mismatches.

A field-response channel model represents propagation as a mapping from an excitation field, source distribution, or latent environmental field to an observed response. In electromagnetic propagation, this mapping appears as a spatial kernel $h(\mathbf{r},\mathbf{s})$ relating a source distribution $j(\mathbf{s})$ to a received field $e(\mathbf{r})$; in XL-MIMO it appears as a superposition of far-field and near-field responses or as an exact element-wise line-of-sight matrix plus sparse non-line-of-sight components; in synaptic molecular communication it appears as a diffusion-reaction field whose response is the channel impulse response $h(t)$; and in site-trained shadowing prediction it appears as a spatial loss field whose line integral or weighted sampling yields link attenuation (Pizzo et al., 2021, Wei et al., 2021, Lotter et al., 2019, Wang et al., 2023).

1. Conceptual scope and canonical forms

Across the literature, the defining feature is not a single physical medium but a common operator viewpoint: the channel is specified by a field and by a response functional acting on that field. This perspective is continuous-space in some settings, discrete-grid in others, and may be either deterministic or stochastic. The “field” may be an electromagnetic wavefield, a molecular concentration field, or a latent shadowing field; the “response” may be a received complex baseband coefficient, a bound-receptor count, or a predicted path loss.

Setting	Field variable	Representative response relation
Electromagnetic random channels	source $j(\mathbf{s})$, kernel $h(\mathbf{r},\mathbf{s})$	$$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$
XL-MIMO hybrid-field	angle-domain and polar-domain coefficients	$$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$
Synaptic molecular communication	concentration field $C(x,t)$	$y(t)=(h*s)(t)$
Site-trained shadowing prediction	loss field $j(\mathbf{s})$0	$j(\mathbf{s})$1

This commonality is technically significant because it separates geometry and physics from inference. Once a field-response operator is specified, channel estimation, control, sparse recovery, or end-to-end learning can be posed as inversion, approximation, or optimization of that operator. The same abstraction supports both exact kernels, such as Green’s-function-based propagation, and latent-field regressors, such as Bayesian loss-field estimation (Pizzo et al., 2021, Wang et al., 2023).

2. Electromagnetic operator formulations

In the electromagnetic setting, the field-response formulation begins from the scalar Helmholtz equation and treats propagation as a linear mapping from source distribution to received field. For a homogeneous medium, the channel is linear and space-invariant, with Green’s function $j(\mathbf{s})$2, whereas in a scattered environment it is linear but space-variant, with kernel $j(\mathbf{s})$3 depending on transmit and receive coordinates separately. The deterministic channel can then be expressed as a four-dimensional plane-wave integral,

$j(\mathbf{s})$4

where $j(\mathbf{s})$5 is an angular response matrix and $j(\mathbf{s})$6 is the plane-wave array response. This formulation is explicitly stated to remain valid in the radiative near-field and even the reactive near-field, because the Weyl decomposition is exact rather than asymptotic (Pizzo et al., 2021).

A central distinction is between propagating and evanescent components. Propagating waves satisfy $j(\mathbf{s})$7; evanescent waves satisfy $j(\mathbf{s})$8 and decay exponentially with distance. The stochastic stationary model retains only propagating modes, thereby preserving spatial stationarity in the radiative near-field while excluding reactive mechanisms confined in close proximity to the source. This directly connects the channel model to the Fourier spectral representation of a stationary spatial random field, with the channel power spectral density supported on the double sphere of radius $j(\mathbf{s})$9 at transmitter and receiver (Pizzo et al., 2021).

Within RIS modeling, this field-level viewpoint appears as a hierarchy of abstractions. Physics/EM-based models, near-field versus far-field formulations, and clustered statistical models are presented as different layers of abstraction of a field-response description. At the element level, a single RIS element is modeled as an EM field-to-field mapping through S-parameters,

$e(\mathbf{r})$0

which yields

$e(\mathbf{r})$1

This decomposes the reflected response into a structural mode and an antenna mode, with the latter tunable through the equivalent reflection coefficient $e(\mathbf{r})$2. In the broader RIS literature surveyed for wireless communications, far-field plane-wave models, near-field element-wise formulations, and EM-consistent impedance models are all treated as valid channel-modeling layers, provided their operating assumptions are respected (Chian et al., 2023, Yildirim et al., 2022).

3. Extremely large-scale MIMO and hybrid-field representations

The XL-MIMO literature uses the field-response concept to formalize the breakdown of purely far-field plane-wave models. For a ULA with aperture $e(\mathbf{r})$3, the Rayleigh distance

$e(\mathbf{r})$4

separates far-field and near-field regions. Far-field paths are represented by planar-wave steering vectors $e(\mathbf{r})$5, while near-field paths are represented by spherical-wave steering vectors $e(\mathbf{r})$6 whose phase depends on both angle and distance through per-element path lengths. In practical XL-MIMO environments, some scatterers lie beyond $e(\mathbf{r})$7 and some within it, leading to a hybrid-field channel in which the total response is the superposition of far-field and near-field contributions,

$e(\mathbf{r})$8

In transformed form this becomes

$e(\mathbf{r})$9

with $h(t)$0 the DFT angle dictionary and $h(t)$1 the polar-domain angle-distance dictionary. The point of the model is that $h(t)$2 is sparse in angle for far-field paths, whereas $h(t)$3 is sparse in angle-distance for near-field paths; neither domain is globally sparse for a true hybrid-field channel, which is why pure far-field or pure near-field estimators are mismatched (Wei et al., 2021).

A related but more geometrically explicit formulation addresses mixed LoS/NLoS near-field XL-MIMO. There, the LoS component is not approximated as a product of transmit and receive array responses. Instead, each entry is modeled by geometric free-space propagation between individual transmit and receive antennas,

$h(t)$4

with $h(t)$5 determined by array-center distance $h(t)$6, departure angle $h(t)$7, relative orientation $h(t)$8, and element offsets. NLoS components remain representable by near-field array response vectors and a sparse polar-domain coefficient matrix,

$h(t)$9

The resulting mixed model is

$j(\mathbf{s})$0

This leads to two distance scales: the MIMO Rayleigh distance

$j(\mathbf{s})$1

which separates far-field and near-field, and the MIMO advanced Rayleigh distance

$j(\mathbf{s})$2

which separates the regime where exact element-wise LoS modeling is needed from the regime where a two-array-response near-field approximation is adequate. A recurring misconception in this literature is that “near-field” can be handled by a single spherical-wave codebook; the mixed LoS/NLoS analysis shows that, in very near-field XL-MIMO, LoS and NLoS require structurally different field-response representations (Lu et al., 2022).

4. Structured wireless field models: RIS surfaces and site-trained loss fields

RIS channel modeling exposes a different aspect of the same idea: the channel can be expressed as the response of an engineered surface to an incident field. In a single RIS element, the outgoing field is decomposed into a structural mode $j(\mathbf{s})$3 and an antenna mode $j(\mathbf{s})$4. At link level, the total scalar contribution of an RIS element is written as

$j(\mathbf{s})$5

where $j(\mathbf{s})$6 is the direct path, $j(\mathbf{s})$7 is the antenna-mode contribution, and $j(\mathbf{s})$8 is the structural-mode contribution. For an $j(\mathbf{s})$9-element array,

$h(\mathbf{r},\mathbf{s})$0

The model is polarization-aware through $h(\mathbf{r},\mathbf{s})$1 transformations such as the direction inversion matrix $h(\mathbf{r},\mathbf{s})$2, the antenna-mode matrix $h(\mathbf{r},\mathbf{s})$3, and the structural-mode matrix $h(\mathbf{r},\mathbf{s})$4. An important practical point is that the switch component is not phase-only: the equivalent reflection coefficient

$h(\mathbf{r},\mathbf{s})$5

exhibits phase-dependent attenuation, so idealized models of the form $h(\mathbf{r},\mathbf{s})$6 omit a measured impairment. This is why the paper emphasizes polarization and switch impairments, and why a blind controlling algorithm and tracking mechanism are proposed for practical control (Chian et al., 2023).

A complementary wireless example replaces explicit EM geometry by a latent site-specific field. CELF models shadowing as a discretized loss field $h(\mathbf{r},\mathbf{s})$7, with link shadowing on link $h(\mathbf{r},\mathbf{s})$8 given by

$h(\mathbf{r},\mathbf{s})$9

The prior is Gaussian,

$$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$0

and the posterior mean is obtained by Bayesian linear regression or the regularized estimator

$$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$1

In this model, the field-response relation is not an EM boundary-value problem but a path-integral-like linear functional of a learned attenuation field. The reported result is that CELF lowers the variance of channel estimates by up to $$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$2 and outperforms Random Forest, SVR, and MLP-ANN in variance reduction on every dataset considered. This suggests that field-response modeling need not be tied to explicit physics solvers; it can also denote a structured latent-field representation whose response operator is learned from measurements (Wang et al., 2023).

5. Diffusion, reaction, and externally forced molecular channels

In molecular communication, the field-response model is literal: the channel is a spatiotemporal concentration field governed by transport equations, and the observable output is an arrival or binding process. For synaptic diffusive molecular communication, the synaptic cleft is reduced to a one-dimensional diffusion domain $$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$3 with aggregated concentration $$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$4 satisfying

$$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$5

The presynaptic boundary implements re-uptake through the radiating condition

$$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$6

and the postsynaptic boundary implements reversible binding through

$$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$7

Here $$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$8 is the channel impulse response, defined as the expected number of molecules bound to postsynaptic receptors at time $$e(\mathbf{r})=\int_{\mathbb{R}^3} j(\mathbf{s})\, h(\mathbf{r},\mathbf{s})\, d\mathbf{s}$$9. The analytical result is a modal expansion,

$$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$0

and for arbitrary release pattern $$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$1 the expected receptor-binding signal is

$$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$2

Within this framework, re-uptake shortens the impulse response and reversible binding extends its tail, making the field-response model directly useful for inter-symbol interference analysis (Lotter et al., 2019).

Field-assisted molecular communication extends the same architecture to externally controlled drift. Particle motion is modeled by the Itô SDE

$$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$3

and the Cameron-Martin-Girsanov theorem is used to express the drifted process as a change of measure relative to pure diffusion. For passive spherical receivers, the endpoint density under arbitrary time-varying drift is exactly

$$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$4

leading to a closed-form sensing probability $$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$5. For fully absorbing spherical receivers, a time-averaged effective drift

$$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$6

is inserted into a constant-drift hitting solution to produce an analytically tractable approximation to the CIR. On top of this, the paper develops dynamic waveform design, and under a MAP-DFE framework states that the first-slot received probability serves as the primary determinant of the bit error probability, while inter-symbol interference manifests as higher-order corrections. This motivates the low-complexity Maximize Received Probability algorithm and its specializations MHP and MSP (Chou et al., 29 Mar 2026).

6. Inference, learning, and open modeling issues

Field-response models are useful only insofar as they can be inverted or embedded into estimation pipelines. One line of work treats the unknown field response as a sparse Green’s-function vector. With multiple sources transmitting random probes simultaneously, each receiver observes

$$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$7

where $$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$8 concatenates all source-receiver channel responses and $$\mathbf{y}=\mathbf{P}\mathbf{F}\mathbf{h}_A+\mathbf{P}\mathbf{W}\mathbf{h}_P+\mathbf{n}$$9 is a structured random convolution matrix. Recovery is posed via $C(x,t)$0-constrained least squares,

$C(x,t)$1

and the analysis gives measurement-length bounds of the form

$C(x,t)$2

in expectation and

$C(x,t)$3

with high probability for restricted isometry. In this setting the field-response model is a sparse linear inverse problem for multiple Green’s functions rather than a geometric path model (Romberg et al., 2010).

A different methodological challenge appears in end-to-end learned communication systems. There the channel model must be differentiable, because transmitter training requires gradients through the channel block. The surveyed remedies are two-phase training, model-free or reinforcement learning-based training, generative surrogate channels such as conditional GANs, and single-stage stochastic gradient approximation over the real channel. The recurring issue is model mismatch: training on AWGN or other simplified differentiable surrogates biases learned weights, while measurement-driven surrogates improve realism at the cost of higher complexity and larger sample requirements. For field-response channels, this creates a tension between physics-based differentiability, black-box realism, and sample efficiency (Ahmad et al., 2022).

The literature also makes clear that field-response models are rarely assumption-free. Reported limitations include narrowband formulations, ULA-only geometries, single-user models, the requirement that the proportion parameter $C(x,t)$4 be known in HF-OMP, static or slowly varying latent fields in site-trained loss modeling, and the need to exclude reactive propagation mechanisms if spatial stationarity is to be preserved in stochastic electromagnetic models (Wei et al., 2021, Wang et al., 2023, Pizzo et al., 2021). A plausible implication is that future field-response channel models will continue to move in two directions at once: toward finer physical fidelity, especially in near-field and dynamically controlled environments, and toward reduced-order or learned surrogates that preserve the operator structure while remaining estimable from finite data.