betaFeedback

WorksEngineering

Temporal Diffusion Models for Joint Signal-Channel Estimation under Gaussian and αα-Stable MIMO Channels

Abstract

Diffusion models have emerged as powerful tools for learning complex data distributions and solving inverse problems. However, most diffusion models rely on the Gaussian distribution, which does not accurately represent heavy-tailed data. In this work, we study diffusion-based reconstruction under Gaussian and symmetric αα-stable fading channels, while also exploring the benefits of using possible temporal dependence in the channel. We consider two variants: pilot-based channel-only estimation and joint image-channel recovery for MIMO systems. Our approach is based on several standard and score-based diffusion models for both Gaussian and αα-stable distributions. We refine these models by guiding the reverse diffusion process using the observation likelihood and show how a single trained diffusion model can be reused for the location-scale family of distributions without retraining. In blind MIMO recovery, separate diffusion models are used for the image and channel, and their estimates are combined through likelihood guidance. For each setting, we compare independent and temporal reconstruction, where information from previous observations is used to improve the estimation. We show that, under a time-dependent channel model, such a temporal reconstruction improves channel estimation and the quality of the reconstructed image. Finally, we show that, for the same heavy-tailed channel data, using an αα-stable diffusion prior provides better channel estimation and image reconstruction than using a Gaussian diffusion prior.

Introduction

Reliable recovery of signals from noisy measurements is a central problem in communication systems and statistical signal processing. In an Integrated Sensing and Communication (ISAC) setting, the receiver must infer both the transmitted information and the unknown propagation effects, often from observations that are incomplete, distorted, or corrupted by noise. Classical Bayesian estimation provides a principled way to combine an observation model with prior information; its performance, however, depends on how faithfully the prior and likelihood represent the underlying physical process.

Diffusion generative models have emerged as flexible tools for learning complex data distributions and incorporating them into inverse problems. Denoising Diffusion Probabilistic Models (DDPM)s construct generative priors through a sequence of Gaussian perturbations and denoising transitions [1]. Their continuous-time score-based counterparts describe related processes through stochastic differential equations [2]. For inverse problems, pretrained diffusion priors can be combined with observation likelihoods through guidance [3]. Piecewise guidance has also been proposed to reduce the computational cost of posterior sampling [4]. In communication systems, diffusion priors have been used to estimate Multiple-Input Multiple-Output (MIMO) channels from known pilots [5, 6] and to jointly estimate the channel and detect unknown transmitted data [7]. More recently, separately-trained channel and source priors have been coupled through the observation model for completely pilot-free blind MIMO semantic communication [8]. These methods provide useful learned priors for reconstruction problems in which an explicit prior distribution is unavailable.

Two characteristics of communication systems motivate this work. First, observations are often collected sequentially, and signals and wireless channels may remain statistically dependent over consecutive time steps. Processing each observation independently therefore ignores potentially useful information from previous observations, whereas a temporal prior can exploit this dependence to improve reconstruction. Second, although Gaussian diffusion models are widely used, they provide poor coverage of heavy-tailed data [9, 10]. Alpha-stable distributions provide a natural extension to the heavy-tailed regime: like the Gaussian distribution, they arise as limits of normalized sums of independent random variables [11], with the Gaussian being a special case [12]. They therefore constitute a canonical model for heavy-tailed wireless fading [13]. In the sequel, stable refers to both the Gaussian and α\alpha-stable cases.

Generative diffusion formulations for heavy-tailed distributions have already been developed, including Lévy-Itô Models (LIM) and Denoising Lévy Probabilistic Models (DLPM) [9, 10]. In this work, we investigate diffusion-based inference for Gaussian and symmetric α\alpha-stable models in two settings: pilot-based channel reconstruction and blind image–channel recovery. We apply DDPM and score-based models for Gaussian distributions, and DLPM and LIM for α\alpha-stable distributions after implementing the necessary changes in terms of training loss, network parameterization, and reverse sampling procedure. Our results show that temporal diffusion priors improve reconstruction when the channel evolves according to a time-dependent model. We also demonstrate that a model trained on a standard distribution can be adapted without retraining to different locations, scales, and Gaussian means and covariances. Finally, we show that, whenever the fading is heavy-tailed, α\alpha-stable diffusion priors provide better channel and image reconstruction than Gaussian diffusion priors.

Background

Diffusion Models Conventions

Throughout the paper, DDPM and score-based diffusion denote discrete and continuous-time Gaussian formulations of [1], [2], based on noise prediction and score estimation, respectively. DLPM and LIM denote their discrete and continuous-time α\alpha-stable counterparts [9], [10], based on stable-noise prediction and fractional-score estimation. A guided diffusion sample denotes the final output of one guided reverse trajectory.

Symmetric Alpha-Stable Distributions

In this section, we define stable distributions and provide some useful properties. For more details, the reader is referred to [12]. Stable Random Variables (RVs) satisfy

aX1+bX2=d(aα+bα)1/αX+r,∀a,b>0,aX_1+bX_2 \mathrel{\stackrel{d}{=}} \left(a^\alpha+b^\alpha\right)^{1/\alpha}X+r,\qquad\forall a,b>0,

for some scalar rr, where X1X_1 and X2X_2 are independent copies of XX and α∈(0,2]\alpha\in(0,2] is the stability index. The case α=2\alpha=2 recovers the Gaussian law. In this work, we say RV XX is Symmetric α\alpha-Stable (Sα\alphaS) and write X∼SαS(δ,γ)X\sim\mathrm{S}\alpha\mathrm{S}(\delta,\gamma), α∈(0,2)\alpha\in(0,2), whenever

E[eiuX]=exp⁡(iδu−γα∣u∣α),u∈R,\mathbb{E}\left[e^{iuX}\right]=\exp\left(i\delta u-\gamma^\alpha|u|^\alpha\right),\qquad u\in\mathbb{R},

where δ∈R\delta\in\mathbb{R} and γ>0\gamma>0 are its location and scale, respectively. The index α\alpha controls the tail asymptotics: a smaller α\alpha produces heavier power-law tails of order ∼1∣x∣1+α\sim\frac{1}{|x|^{1+\alpha}}. Moreover, E[∣X∣p]<∞\mathbb{E}[|X|^p]<\infty, whenever p<αp<\alpha, so the variance is infinite for α<2\alpha<2.

In addition, a random vector X∈RdX\in\mathbb{R}^d is said to be “circular” Sα\alphaS with law Sα\alphaS(δ,γ)(\delta,\gamma) if and only if

E[eiuTX]=exp⁡(iδTu−γα∥u∥α),u∈Rd.\mathbb{E}\left[e^{i\mathbf{u}^{\mathsf{T}}\mathbf{X}}\right]=\exp\left(i\delta^{\mathsf{T}}\mathbf{u}-\gamma^\alpha\|\mathbf{u}\|^\alpha\right),\qquad\mathbf{u}\in\mathbb{R}^d.

It admits the Gaussian scale-mixture representation

X=dδ+γAG,(1)\mathbf{X}\mathrel{\stackrel{d}{=}}\delta+\gamma\sqrt{A}\mathbf{G}, \tag*{(1)}

where G∼N(0,Id)\mathbf{G}\sim\mathcal{N}(0,I_d) is independent of AA—a positive α2\frac{\alpha}{2}-stable variable with scale equal to cos⁡2/α(πα4)\cos^{2/\alpha}\left(\frac{\pi\alpha}{4}\right). Consequently, given that AA is equal to aa

X∣(A=a)∼N(δ,γ2aId).\mathbf{X}\mid(A=a)\sim\mathcal{N}\left(\delta,\gamma^2aI_d\right).

Reconstruction Metrics

For qk,q^k∈Rd\mathbf{q}_k,\hat{\mathbf{q}}_k\in\mathbb{R}^d we use the Normalized Mean Squared-Error

NMSEdB=10log⁡10∑k=1K∥q^k−qk∥22∑k=1K∥qk∥22,(2)\mathrm{NMSE}_{\mathrm{dB}}=10\log_{10}\frac{\sum_{k=1}^{K}\left\|\hat{\mathbf{q}}_k-\mathbf{q}_k\right\|_2^2}{\sum_{k=1}^{K}\left\|\mathbf{q}_k\right\|_2^2}, \tag*{(2)}

for Gaussian data, and

Ep=1Kd∑k=1K∑i=1d∣q^k,i−qk,i∣p,0<p<α,(3)\mathcal{E}_p=\frac{1}{Kd}\sum_{k=1}^{K}\sum_{i=1}^{d}\left|\hat{q}_{k,i}-q_{k,i}\right|^p,\qquad0<p<\alpha, \tag*{(3)}

& Elog⁡,p=1Kd∑k=1K∑i=1dlog⁡(1+∣q^k,i−qk,i∣p)\mathcal{E}_{\log,p}=\frac{1}{Kd}\sum_{k=1}^{K}\sum_{i=1}^{d}\log\left(1+\left|\hat{q}_{k,i}-q_{k,i}\right|^p\right), (4)

for α\alpha-stable ones. Note that E1\mathcal{E}_1 is the mean absolute error. Image reconstruction is evaluated using MS-SSIM [14]. Lower errors and higher MS-SSIM indicate better performance.

Problem Formulation

With a view towards MIMO setups where images are transmitted using NtN_t transmit antennas and NrN_r receive antennas, define

Xk=[Xk,1 ⋯ Xk,L]∈RNt×L,\mathbf{X}_k=\left[\mathbf{X}_{k,1}\ \cdots\ \mathbf{X}_{k,L}\right]\in\mathbb{R}^{N_t\times L},

where Xk,ℓ∈RNt\mathbf{X}_{k,\ell}\in\mathbb{R}^{N_t} is the ℓ\ell-th transmitted symbol. We adopt a block-fading model where the channel matrix Hk∈RNr×Nt\mathbf{H}_k\in\mathbb{R}^{N_r\times N_t} is assumed constant over these LL symbols. For each ℓ=1,…,L\ell=1,\ldots,L, the received symbol is

Yk,ℓ=HkXk,ℓ+Nk,ℓ,Yk,ℓ,Nk,ℓ∈RNr.\mathbf{Y}_{k,\ell}=\mathbf{H}_k\mathbf{X}_{k,\ell}+\mathbf{N}_{k,\ell},\qquad\mathbf{Y}_{k,\ell},\mathbf{N}_{k,\ell}\in\mathbb{R}^{N_r}.

Stacking the LL received symbols we write

Yk=HkXk+Nk,Yk,Nk∈RNr×L.(4)\mathbf{Y}_k=\mathbf{H}_k\mathbf{X}_k+\mathbf{N}_k,\qquad\mathbf{Y}_k,\mathbf{N}_k\in\mathbb{R}^{N_r\times L}. \tag*{(4)}

From one block to another, we model the channel as a temporally-dependent MIMO channel that—while stationary, evolves temporally according to a first-order autoregressive model; it may follow either a Gaussian or a symmetric α\alpha-stable law. For the Gaussian channel, the “vectorized” channel matrix:

Hk≜vec⁡(Hk)∼N(μH,ΣH),\mathbf{H}_k\triangleq\operatorname{vec}(\mathbf{H}_k)\sim\mathcal{N}(\boldsymbol{\mu}_H,\boldsymbol{\Sigma}_H),
Hk=μH+ρH[Hk−1−μH]+1−ρH2Wk(5)\mathbf{H}_k=\boldsymbol{\mu}_H+\rho_H\left[\mathbf{H}_{k-1}-\boldsymbol{\mu}_H\right]+\sqrt{1-\rho_H^2}\mathbf{W}_k \tag*{(5)}

where Wk∼N(0,ΣH)\mathbf{W}_k\sim\mathcal{N}(0,\boldsymbol{\Sigma}_H).

For the symmetric α\alpha-stable channel,

Hk=μH+ρH[Hk−1−μH]+1−∣ρH∣ααWk(6)\mathbf{H}_k=\boldsymbol{\mu}_H+\rho_H\left[\mathbf{H}_{k-1}-\boldsymbol{\mu}_H\right]+\sqrt[\alpha]{1-\left|\rho_H\right|^\alpha}\mathbf{W}_k \tag*{(6)}

where Wk∼SαS(0,γH)\mathbf{W}_k\sim\mathrm{S}\alpha\mathrm{S}(0,\gamma_H), preserving the stationary marginal Hk∼SαS(μH,γH)\mathbf{H}_k\sim\mathrm{S}\alpha\mathrm{S}(\boldsymbol{\mu}_H,\gamma_H). We restrict α>1\alpha>1 to guarantee a finite mean, in accordance with the LIM model [9].

The additive noise {Nk}k\{\mathbf{N}_k\}_k in (4) are independent and chosen from the corresponding family: Gaussian noise for the Gaussian channel and Sα\alphaS noise for the heavy-tailed channel. We study two inference scenarios that differ in what is known at the receiver and in the unknown quantities to be recovered.

Channel-Only Estimation

In the first scenario, a known pilot sequence is used and set, without loss of generality, to Xp=1\mathbf{X}_p=1. The channel coefficients are processed jointly as a vector, giving the vectorized observation model

Yk=Hk+Nk.(7)\mathbf{Y}_k=\mathbf{H}_k+\mathbf{N}_k. \tag*{(7)}

The objective is to estimate the temporally-varying channel vector Hk\mathbf{H}_k from the pilot observations. This setting isolates channel estimation from data detection and allows us to compare independent inference, which ignores past observations, with temporal inference, which uses the previous channel estimate to form a predictive prior. All channel coefficients are reconstructed jointly.

Joint Channel–Image MIMO Estimation

In the second scenario, an initial training stage provides the channel H0H_0. For k>0k > 0, no additional pilots are transmitted, and both the current channel and transmitted image are unknown. Let DkD_k be an MNIST image [15]. A fixed encoder maps it to

Xk=fγ(Dk)=[Xk,1 ⋯ Xk,L]∈RNt×L,X_k = f_\gamma(D_k) = [X_{k,1}\ \cdots\ X_{k,L}] \in\mathbb{R}^{N_t \times L},

where Xk,ℓ∈RNtX_{k,\ell} \in\mathbb{R}^{N_t} is the symbol vector transmitted during channel use ℓ\ell. As previously stated, the channel Hk∈RNr×NtH_k \in\mathbb{R}^{N_r \times N_t} is assumed constant over these LL channel uses.

The receiver must jointly estimate HkH_k and the information-bearing encoded signal Xk=fγ(Dk)X_k = f_\gamma(D_k), or equivalently the underlying image DkD_k. Separate learned priors are used for the channel and image, while the two unknowns are coupled during inference through the MIMO likelihood. The known initial channel H0H_0 anchors the temporal channel prior for subsequent time steps.

Methodology

The two scenarios in Section III use the same inference principle. Diffusion models provide priors for the unknown quantities, the physical observation model supplies the likelihood used to guide the reverse process, and temporal information enters through a predictive channel prior. The diffusion networks are trained once on canonical distributions and adapted during inference without retraining when the location, scale, Gaussian mean, or Gaussian covariance of the prior changes.

Canonical Gaussian Adaptation

For each Gaussian formulation, one diffusion model is trained on the dd-dimensional Z0∼N(0,Id)Z_0 \sim\mathcal{N}(0,I_d). In DDPM, ϵθ\epsilon_\theta predicts the Gaussian perturbation, whereas in the continuous formulation, sθs_\theta predicts the score.

Let ϵt∼N(0,Id)\epsilon_t \sim\mathcal{N}(0,I_d). The canonical forward marginals are given by Zt=αˉtZ0+1−αˉtϵtZ_t = \sqrt{\bar{\alpha}_t}Z_0 + \sqrt{1-\bar{\alpha}_t}\epsilon_t for DDPM and Z(t)=m(t)Z0+σ(t)ϵ(t)Z(t) = m(t)Z_0+\sigma(t)\epsilon(t) for the score-based model. The same networks are adapted, without retraining, to X0∼N(μ,Σ)X_0 \sim\mathcal{N}(\mu,\Sigma) and the perturbed marginals are

Xt∼N(αˉtμ,Ct),Ct=αˉtΣ+(1−αˉt)Id,X_t \sim\mathcal{N}\left(\sqrt{\bar{\alpha}_t}\mu,C_t\right), \qquad C_t = \bar{\alpha}_t\Sigma+ (1-\bar{\alpha}_t)I_d,
X(t)∼N(m(t)μ,C(t)),C(t)=m2(t)Σ+σ2(t)Id.X(t) \sim\mathcal{N}\left(m(t)\mu,C(t)\right), \qquad C(t) = m^2(t)\Sigma+ \sigma^2(t)I_d.

At every reverse step, let

Ct=RtRtT,C(t)=R(t)RT(t)C_t = R_tR_t^{\mathsf{T}}, \qquad C(t) = R(t)R^{\mathsf{T}}(t)

The canonical inputs are

Ut=Rt−1(Xt−αˉtμ),(DDPM),U_t = R_t^{-1}\left(X_t-\sqrt{\bar{\alpha}_t}\mu\right), \qquad\text{(DDPM)},
U(t)=R−1(t)(X(t)−m(t)μ),(score-based),U(t) = R^{-1}(t)\left(X(t)-m(t)\mu\right), \qquad\text{(score-based)},

and the network outputs are transported as

ϵ^μ,Σ(Xt,t)=Rt−Tϵθ(Ut,t),(DDPM).(8)\widehat{\epsilon}_{\mu,\Sigma}(X_t,t) = R_t^{-\mathsf{T}}\epsilon_\theta(U_t,t), \qquad\text{(DDPM)}. \tag*{(8)}
s^μ,Σ(X(t),t)=R−T(t)sθ(U(t),t),(score-based).(9)\widehat{s}_{\mu,\Sigma}(X(t),t) = R^{-\mathsf{T}}(t)s_\theta(U(t),t), \qquad\text{(score-based)}. \tag*{(9)}

Note that the DDPM transport is stated for the MSE-optimal noise predictor. It is established using the joint Gaussianity of (ϵt,Xt)(\epsilon_t,X_t) and the fact that the MSE-optimal noise predictor is E[ϵt∣Xt]\mathbb{E}[\epsilon_t \mid X_t]. Applying the Gaussian conditional-mean formula together with Ct−1=Rt−TRt−1C_t^{-1}=R_t^{-\mathsf{T}}R_t^{-1} gives the stated input and output transformations (see Appendix A). For the score-based formulation, the transport follows directly from the chain rule under the affine transformation U(t)=R−1(t)(X(t)−m(t)μ)U(t)=R^{-1}(t)(X(t)-m(t)\mu) (see Appendix B).

Canonical Alpha-Stable Adaptation

For each α\alpha-stable formulation, one diffusion model is trained on Z0∼SαS⁡(0,1)Z_0 \sim\operatorname{S\alpha S}(0,1). In DLPM, ϵθ\epsilon_\theta predicts the stable perturbation, whereas in LIM, sθ(α)s_\theta^{(\alpha)} predicts the fractional score defined in [9, 16]. We use the logarithmic error Elog⁡,p\mathcal{E}_{\log,p} defined in (4) in the noise-prediction models, with p=1p=1 in the experiments.

The trained models are adapted without retraining to

X0=μ+γZ0∼SαS⁡(μ,γ).X_0=\mu+\gamma Z_0 \sim\operatorname{S\alpha S}(\mu,\gamma).

For DLPM, whose canonical forward marginal is Zt=cˉtZ0+(1−cˉtα)1/αϵtZ_t=\bar{c}_tZ_0+\left(1-\bar{c}_t^\alpha\right)^{1/\alpha}\epsilon_t, where ϵt∼SαS⁡(0,1)\epsilon_t\sim\operatorname{S\alpha S}(0,1), the experiments were conducted with γ=1\gamma=1, and

Ut=Xt−cˉtμ,ϵ^μ,γ(Xt,t)=ϵθ(Ut,t).U_t=X_t-\bar{c}_t\mu, \qquad\widehat{\epsilon}_{\mu,\gamma}(X_t,t)=\epsilon_\theta(U_t,t).

For LIM, whose canonical forward marginal is Z(t)=a(t)Z0+σ(t)ϵ(t)Z(t)=a(t)Z_0+\sigma(t)\epsilon(t),

U(t)=X(t)−a(t)μγ,sμ,γ(α)(X(t),t)=γ1−αsθ(α)(U(t),t).U(t)=\frac{X(t)-a(t)\mu}{\gamma}, \qquad s_{\mu,\gamma}^{(\alpha)}(X(t),t)=\gamma^{1-\alpha}s_\theta^{(\alpha)}(U(t),t).

For DLPM with γ=1\gamma=1, the transport follows directly by substituting X0=μ+Z0X_0=\mu+Z_0 into the forward process (see Appendix C). For LIM, the transport rule follows by applying the Fourier transform to the fractional score and examining its behavior under the location-scale transformation X(t)=a(t)μ+γU(t)X(t)=a(t)\mu+\gamma U(t), hence the factor γ1−α\gamma^{1-\alpha} (see Appendix D).

Observation Conditioning and Joint Guidance

For both scenarios, the diffusion priors are conditioned on the physical observation model. We use the generic vectorized representation

Y=f(ξ0(1),…,ξ0(J))+N,(10)Y=f\left(\xi_0^{(1)},\ldots,\xi_0^{(J)}\right)+N, \tag*{(10)}

where each unknown ξ0(j)\xi_0^{(j)} has a separately trained diffusion prior and reverse state ξt(j)\xi_t^{(j)}. Matrix-valued observations are understood through vectorization.

For Gaussian noise N∼N(0,σn2Id)N\sim\mathcal{N}(0,\sigma_n^2I_d), the reverse models provide clean estimates ξ^0(j)(ξt(j))\widehat{\xi}_0^{(j)}(\xi_t^{(j)}), and the likelihood gradient for the jj-th unknown is

gt(j)=∇ξt(j)log⁡p(y∣ξ^0(1),…,ξ^0(J)).g_t^{(j)}=\nabla_{\xi_t^{(j)}}\log p\left(y\mid\widehat{\xi}_0^{(1)},\ldots,\widehat{\xi}_0^{(J)}\right).

A weighted version of gt(j)g_t^{(j)} is added to the corresponding reverse update. Thus, separately trained priors are coupled only through the shared likelihood.

When the Gaussian observation is linear,

Y=Aξ0+N,ξ0∼N(μ,Σ).(11)Y=A\xi_0+N, \qquad\xi_0\sim\mathcal{N}(\mu,\Sigma). \tag*{(11)}

the posterior is Gaussian with

μpost=μ+ΣAT(AΣAT+σn2Id)−1(y−Aμ),(12)\mu_{\mathrm{post}}=\mu+\Sigma A^{\mathsf T}\left(A\Sigma A^{\mathsf T}+\sigma_n^2 I_d\right)^{-1}(y-A\mu), \tag*{(12)}
Σpost=Σ−ΣAT(AΣAT+σn2Id)−1AΣ.(13)\Sigma_{\mathrm{post}}=\Sigma-\Sigma A^{\mathsf T}\left(A\Sigma A^{\mathsf T}+\sigma_n^2 I_d\right)^{-1}A\Sigma. \tag*{(13)}

This permits exact Gaussian conditional sampling in the channel-only setting. For SαSS\alpha S noise N∼SαS(0,γn)N\sim S\alpha S(0,\gamma_n), we use the Gaussian scale-mixture representation in (1). Conditioned on the positive-stable variable AA,

Y∣ξ0(1),…,ξ0(J),A∼N(f(ξ0(1),…,ξ0(J)),γn2AId).Y\mid\xi_0^{(1)},\ldots,\xi_0^{(J)},A\sim\mathcal{N}\left(f\left(\xi_0^{(1)},\ldots,\xi_0^{(J)}\right),\gamma_n^2 A I_d\right).

Using MMCM_{\mathrm{MC}} samples {A(m)}m=1MMC\{A^{(m)}\}_{m=1}^{M_{\mathrm{MC}}}, the marginal likelihood is approximated by

p^(y∣ξ0(1),…,ξ0(J))=1MMC∑m=1MMCN(y;f(ξ0(1),…,ξ0(J)),γn2A(m)Id).\widehat{p}\left(y\mid\xi_0^{(1)},\ldots,\xi_0^{(J)}\right) =\frac{1}{M_{\mathrm{MC}}}\sum_{m=1}^{M_{\mathrm{MC}}}\mathcal{N}\left(y;f\left(\xi_0^{(1)},\ldots,\xi_0^{(J)}\right),\gamma_n^2 A^{(m)}I_d\right).

The gradient of the logarithm of this approximation is used in the same guided reverse-update construction.

Temporal Channel Priors

Independent inference uses the stationary channel law at every time step. Temporal inference instead uses the previous reconstruction to form a predictive prior for the current channel. The specific uncertainty update depends on whether the preceding channel uncertainty is propagated explicitly or the previous estimate is treated as exact.

For the Gaussian channel, when the preceding posterior covariance is propagated,

Hk∣h^k−1∼N(m‾H,k,P‾H,k),(14)H_k\mid\widehat{h}_{k-1}\sim\mathcal{N}\left(\overline{m}_{H,k},\overline{P}_{H,k}\right), \tag*{(14)}

where m‾H,k=μH+ρH(h^k−1−μH)\overline{m}_{H,k}=\mu_H+\rho_H(\widehat{h}_{k-1}-\mu_H), P‾H,k=ρH2PH,k−1+(1−ρH2)ΣH\overline{P}_{H,k}=\rho_H^2P_{H,k-1}+(1-\rho_H^2)\Sigma_H.

If h^k−1\widehat{h}_{k-1} is treated as exact, this reduces to

Hk∣h^k−1∼N(μH+ρH(h^k−1−μH),(1−ρH2)ΣH).(15)H_k\mid\widehat{h}_{k-1}\sim\mathcal{N}\left(\mu_H+\rho_H(\widehat{h}_{k-1}-\mu_H),(1-\rho_H^2)\Sigma_H\right). \tag*{(15)}

For the SαSS\alpha S channel, uncertainty propagation is approximated by

Hk∣h^k−1≈SαS(μ‾H,k,γ‾H,k),(16)H_k\mid\widehat{h}_{k-1}\approx S\alpha S\left(\overline{\mu}_{H,k},\overline{\gamma}_{H,k}\right), \tag*{(16)}

where μ‾H,k=μH+ρH(h^k−1−μH)\overline{\mu}_{H,k}=\mu_H+\rho_H(\widehat{h}_{k-1}-\mu_H),

γ‾H,k=∣ρH∣αγ‾H,k−1α+(1−∣ρH∣α)γHαα.\overline{\gamma}_{H,k}=\sqrt[\alpha]{|\rho_H|^\alpha\overline{\gamma}_{H,k-1}^\alpha+(1-|\rho_H|^\alpha)\gamma_H^\alpha}.

This is an approximation because the distribution of the previous estimation error is generally not α\alpha-stable. We therefore match a zero-location SαSS\alpha S approximation to the empirical interquartile range of the guided diffusion samples [17].

If the previous estimate is treated as exact, the predictive scale reduces to γH1−∣ρH∣αα\gamma_H\sqrt[\alpha]{1-|\rho_H|^\alpha}, yielding

Hk∣h^k−1∼SαS(μH+ρH(h^k−1−μH),γH1−∣ρH∣αα).(17)H_k\mid\widehat{h}_{k-1}\sim S\alpha S\left(\mu_H+\rho_H(\widehat{h}_{k-1}-\mu_H),\gamma_H\sqrt[\alpha]{1-|\rho_H|^\alpha}\right). \tag*{(17)}

Experiments and Results

Channel-Only Estimation

Using pilot signals as in (7), we jointly process three channel coefficients, with Hk∈R3H_k\in\mathbb{R}^3.

Gaussian Channel

As an example, we consider channels’ coefficients with correlated stationary distribution

Hk∼N([53−2]T,[20.40.20.40.50.10.20.11.5]),H_k\sim\mathcal{N}\left( \begin{bmatrix} 5 & 3 & -2 \end{bmatrix}^{\mathsf T}, \begin{bmatrix} 2 & 0.4 & 0.2\\ 0.4 & 0.5 & 0.1\\ 0.2 & 0.1 & 1.5 \end{bmatrix} \right),

and Nk∼N(0,I3)N_k\sim\mathcal{N}(0,I_3). The independent method uses the stationary prior, whereas the temporal methods use the predictive prior in (14). At k=0k=0, PH,0=(ΣH−1+I3)−1P_{H,0}=(\Sigma_H^{-1}+I_3)^{-1}, and the posterior covariance is propagated recursively for k>0k>0.

We compare four estimators. Independent guided and temporal guided estimators use likelihood guidance, while independent conditional and temporal conditional estimators use exact Gaussian conditioning, both with stationary channel prior (6) and predictive channel prior (15).

From the PP diffusion samples {hk(j)}j=1P\{h_k^{(j)}\}_{j=1}^{P}, the guided estimate is selected as

h^k=hk(jk∗),jk∗=arg⁡min⁡1≤j≤M ∥yk−hk(j)∥2.\widehat{h}_k=h_k^{(j_k^*)},\qquad j_k^*=\underset{1\le j\le M}{\arg\min}\ \left\|y_k-h_k^{(j)}\right\|^2.

The MSE-optimal estimator is the exact Bayesian posterior mean. Performance is evaluated using NMSE defined in (2).

Alpha-Stable Channel

We consider the following example for the Sα\alphaS {Hk}k\{\mathbf{H}_k\}_k: stability index α=1.7\alpha=1.7, location μH=[5  3  −2]T\boldsymbol{\mu}_{H}=[5\ \ 3\ \ -2]^{\mathsf{T}}, and unit stationary scale. The additive α\alpha-stable noise is Sα\alphaS Nk∼SαS(0,1)\mathbf{N}_k\sim\mathrm{S}\alpha\mathrm{S}(0,1). These α\alpha-stable vectors can be generated using appropriate python packages such as SciPy or the AUB-HTP package [18].

Independent inference uses the stationary channel prior (7), whereas temporal inference uses (17). The channel estimate is the componentwise median of the final guided samples,

h^k=median⁡{hk(1),…,hk(P)}.\hat{\mathbf{h}}_k=\operatorname{median}\left\{\mathbf{h}_k^{(1)},\ldots,\mathbf{h}_k^{(P)}\right\}.

For equal signal and noise stability indices and scales, the exact independent absolute-error estimator is

h^kind,opt=yk+μH2.\hat{\mathbf{h}}_k^{\mathrm{ind,opt}}=\frac{\mathbf{y}_k+\boldsymbol{\mu}_H}{2}.

For the temporal reference, the MAE-optimal estimator is the componentwise median of p(Hk∣y0,y1,…,yk)p(\mathbf{H}_k\mid\mathbf{y}_0,\mathbf{y}_1,\ldots,\mathbf{y}_k), approximated numerically using samples from the known temporal model weighted by the observation likelihood [19]. Performance is evaluated using E1\mathcal{E}_1 defined in (3).

Results: We use K=20K=20 sequences. Figure 1 shows that the gain of temporal over independent reconstruction increases with ρH\rho_H, measured by a reduction in NMSE for Gaussian channels with DDPM and in MAE for α\alpha-stable channels with DLPM. For the Gaussian case, this behavior is observed under both guided and conditional reconstruction. At ρH=0\rho_H=0, the temporal and independent methods coincide, as expected. Similar trends are observed for the continuous score-based and LIM models in Figure 6.

Reconstruction loss versus rho_H for independent and temporal channel reconstruction

Figure 1. Reconstruction loss versus ρH\rho_H for independent and temporal channel reconstruction. Top: Gaussian channels using DDPM. Bottom: α\alpha-stable channels using DPLM with α=1.7\alpha=1.7.

Reconstruction loss versus rho_H for independent and temporal channel reconstruction

Figure 6. Reconstruction loss versus ρH\rho_H for independent and temporal channel reconstruction. Top: Gaussian channels using the score-based model. Bottom: α\alpha-stable channels using LIM with α=1.7\alpha=1.7. This corresponds to the channel-only experiment in § 5-A.

Joint Channel–Image MIMO Estimation

In this setup, each sequence contains encoded MNIST images represented as 4×324\times32 symbols {Xk}\{\mathbf{X}_k\} in (5). The channel {Hk}k\{\mathbf{H}_k\}_k is 4×44\times4 with an ”initial” matrix H0\mathbf{H}_0 that is known from the training stage; for k>0k>0, the receiver jointly estimates the current channel and image. The image and channel priors are trained separately and coupled only through the common MIMO likelihood.

In the generic guidance model (11), the two unknowns are ξk(1)=vec⁡(Hk)\boldsymbol{\xi}_k^{(1)}=\operatorname{vec}(\mathbf{H}_k) and ξk(2)=vec⁡(Dk)\boldsymbol{\xi}_k^{(2)}=\operatorname{vec}(\mathbf{D}_k), with f(Hk,Dk)=Hkfγ(Dk)=HkXkf(\mathbf{H}_k,\mathbf{D}_k)=\mathbf{H}_k f_\gamma(\mathbf{D}_k)=\mathbf{H}_k\mathbf{X}_k.

For each channel family, we compare independent joint inference using the stationary channel prior with temporal joint inference using the corresponding predictive prior.

Gaussian Channel

The vectorized channel satisfies Hk=vec⁡(Hk)∼N(0,I16)\mathbf{H}_k=\operatorname{vec}(\mathbf{H}_k)\sim\mathcal{N}(0,\mathbf{I}_{16}) and the entries of Nk\mathbf{N}_k are independent N(0,σN2)\mathcal{N}(0,\sigma_N^2) variables. The temporal method uses (16), treating the previous channel estimate as exact, while the independent method uses the stationary channel prior (6). From the joint candidates {(Hk(j),Dk(j))}j=1M\left\{(\mathbf{H}_k^{(j)},\mathbf{D}_k^{(j)})\right\}_{j=1}^{M}, we select

jk∗=arg⁡min⁡j∥Yk−Hk(j)fγ(Dk(j))∥Frob2.j_k^*=\underset{j}{\arg\min}\left\|\mathbf{Y}_k-\mathbf{H}_k^{(j)}f_\gamma\left(\mathbf{D}_k^{(j)}\right)\right\|_{\mathrm{Frob}}^2.

Alpha-Stable Channel

The channel models are trained on the 16-dimensional Hk=vec⁡(Hk)∼SαS(0,1)\mathbf{H}_k=\operatorname{vec}(\mathbf{H}_k)\sim\mathrm{S}\alpha\mathrm{S}(0,1), for α=1.7\alpha=1.7 and the additive α\alpha-stable noise is SαS Nk∼SαS(0,1)\mathrm{S}\alpha\mathrm{S}\ \mathbf{N}_k\sim\mathrm{S}\alpha\mathrm{S}(0,1). Since α\alpha-stable noise has infinite variance, its scale is selected according to

γN=RMS⁡(Hkfγ(Dk))10SNR/20,RMS⁡(Z)=∥Z∥FrobNrL.\gamma_N=\frac{\operatorname{RMS}(\mathbf{H}_k f_\gamma(\mathbf{D}_k))}{10^{\mathrm{SNR}/20}},\qquad\operatorname{RMS}(\mathbf{Z})=\frac{\left\|\mathbf{Z}\right\|_{\mathrm{Frob}}}{\sqrt{N_rL}}.

The independent method uses the stationary channel prior (7), while the temporal method uses (18). From the joint candidates, we select the one with the largest approximated α\alpha-stable measurement likelihood,

jk∗=arg⁡max⁡1≤j≤M p^(Yk∣Hk(j),Dk(j)).j_k^*=\underset{1\leq j\leq M}{\arg\max}\ \hat{p}\left(\mathbf{Y}_k\mid\mathbf{H}_k^{(j)},\mathbf{D}_k^{(j)}\right).

Independent and temporal reconstruction results: Each sequence contains K=10K=10 MNIST images. Figure 2 shows that temporal inference improves image MS-SSIM, channel NMSE (Gaussian) and Elog⁡,p\mathcal{E}_{\log,p} (α\alpha-stable) over independent inference as ρH\rho_H increases. At ρH=0\rho_H=0, the methods coincide. Figure 3 depicts representative reconstructed images for the two reconstruction methods and shows that temporal reconstruction becomes increasingly beneficial over independent one as ρH\rho_H increases for both Gaussian and α\alpha-stable channels. Similar trends are observed for the score-based and LIM models in Figure 9, with representative continuous-time reconstructions shown in Figure 10.

Image MS-SSIM and channel loss versus SNR for independent and temporal joint channel–image MIMO reconstruction

Figure 2. Image MS-SSIM and channel loss vs. SNR for independent and temporal joint channel–image MIMO reconstruction. Left: Gaussian channels using DDPM. Right: α\alpha-stable channels with α=1.7\alpha=1.7 using DLPM.

Sent images and corresponding independent and temporal joint MIMO reconstructions

Figure 3. Sent images and the corresponding independent and temporal joint MIMO reconstructions for ρH=0.8\rho_H = 0.8. Top: Gaussian channel using DDPM at SNR = 10 dB. Bottom: α\alpha-stable channel using DLPM with α=1.7\alpha= 1.7 at SNR = 5 dB.

Image MS-SSIM and channel loss versus SNR for independent and temporal joint channel–image MIMO reconstruction

Figure 9. Image MS-SSIM and channel loss vs. SNR for independent and temporal joint channel–image MIMO reconstruction. Top: Gaussian channels using the score-based model. Bottom: α\alpha-stable channels with α=1.7\alpha= 1.7 using LIM. This corresponds to the joint channel–image MIMO experiment in § 5-B.

Sent images and independent and temporal joint MIMO reconstructions

Figure 10. Sent images and the corresponding independent and temporal joint MIMO reconstructions. Top: Gaussian channel using the score-based model at SNR =10= 10 dB with ρH=0.4\rho_H = 0.4. Bottom: α\alpha-stable channel using LIM with α=1.7\alpha= 1.7 at SNR =10= 10 dB with ρH=0.8\rho_H = 0.8. These results correspond to the joint channel–image MIMO experiment in § 5-B.

Model mismatch: Matched α\alpha-stable channel priors are compared with Gaussian channel priors trained and evaluated on the same α\alpha-stable channel data with ρ=0\rho=0. DLPM is compared with Gaussian DDPM in Figure 4, while LIM is compared with the Gaussian score-based model. It is apparent that DLPM provides lower channel loss and higher image MS-SSIM than the Gaussian DDPM. Figure 5 shows a representative reconstructed image for the two channel priors. Similar trends are observed when comparing LIM with the Gaussian score-based model in Figure 7, with representative reconstructions shown in Figure 8.

Image MS-SSIM and channel loss versus SNR

Figure 4. Image MS-SSIM and channel loss vs. SNR for α\alpha-stable joint channel–image MIMO reconstruction using matched DLPM and mismatched DDPM for α=1.7\alpha= 1.7.

Sent image with matched and mismatched joint MIMO reconstructions

Figure 5. Sent image on α\alpha-stable channel with joint MIMO reconstructions obtained using matched DLPM and mismatched Gaussian DDPM for α=1.7\alpha= 1.7 and SNR = 15 dB.

Image MS-SSIM and channel loss versus SNR for alpha-stable joint channel-image MIMO reconstruction

Figure 7. Image MS-SSIM and channel loss vs. SNR for α\alpha-stable joint channel–image MIMO reconstruction using matched LIM and mismatched Gaussian score-based model for α=1.7\alpha=1.7. This corresponds to the model-mismatch experiment in § 5-B.

Sent image and joint MIMO reconstructions on an α-stable channel

Figure 8. Sent image on α\alpha-stable channel with joint MIMO reconstructions obtained using matched LIM and mismatched Gaussian score-based model for α=1.7\alpha= 1.7 and SNR =15= 15 dB. These results correspond to the model-mismatch experiment in § 5-B.

Discussion

In this work, we leveraged time-dependence in channel statistics in diffusion-based channel estimation. Propagating previous estimates becomes more informative and improves both signal reconstruction and blind-MIMO recovery. This advantage is common to both Gaussian and α\alpha-stable models.

The channel-model comparison separately shows the importance of matching the diffusion noise family with the underlying data distribution. For Gaussian and α\alpha-stable channel models trained on the same heavy-tailed data and using the same observations and likelihood, the improved channel loss and image quality obtained by DLPM, with similar behavior for LIM, therefore suggest that stable perturbations provide a better-matched prior for α\alpha-stable channels.

In the blind MIMO setting, while the assumption of perfect knowledge of the initial channel estimate is motivated by prior training, this might require a prohibitively long training period. Also, considering the previous channel estimate as exact when constructing the temporal prior is suboptimal. Future work could focus on reducing channel uncertainty propagation over time and the need for a long initial training stage.

Appendix A: DDPM

Forward and Reverse Processes

A DDPM defines the forward process

q(Xt∣Xt−1=xt−1)=N(αtxt−1,βtId),αt=1−βt,q(X_t \mid X_{t-1}=x_{t-1})=\mathcal{N}\left(\sqrt{\alpha_t}x_{t-1},\beta_t I_d\right),\quad\alpha_t=1-\beta_t,

whose marginal, with αˉt=∏s=1tαs\bar{\alpha}_t=\prod_{s=1}^{t}\alpha_s, is

Xt=αˉtX0+1−αˉtϵt,ϵt∼N(0,Id).(18)X_t=\sqrt{\bar{\alpha}_t}X_0+\sqrt{1-\bar{\alpha}_t}\epsilon_t,\quad\epsilon_t\sim\mathcal{N}(0,I_d). \tag*{(18)}

The exact reverse posterior conditioned on X0=x0X_0=x_0 is

q(Xt−1∣Xt=xt,X0=x0)=N(μ~t,β~tId),q(X_{t-1}\mid X_t=x_t,X_0=x_0)=\mathcal{N}\left(\tilde{\mu}_t,\tilde{\beta}_t I_d\right),

where

μ~t=αˉt−1βt1−αˉtx0+αt(1−αˉt−1)1−αˉtxt,\tilde{\mu}_t=\frac{\sqrt{\bar{\alpha}_{t-1}}\beta_t}{1-\bar{\alpha}_t}x_0+\frac{\sqrt{\alpha_t}(1-\bar{\alpha}_{t-1})}{1-\bar{\alpha}_t}x_t,
β~t=1−αˉt−11−αˉtβt.\tilde{\beta}_t=\frac{1-\bar{\alpha}_{t-1}}{1-\bar{\alpha}_t}\beta_t.

The learned reverse transition replaces the unknown perturbation by ϵθ(xt,t)\epsilon_\theta(x_t,t) such that

pθ(Xt−1∣Xt=xt)=N(μθ(xt,t),β~tId),p_\theta(X_{t-1}\mid X_t=x_t)=\mathcal{N}\left(\mu_\theta(x_t,t),\tilde{\beta}_t I_d\right),

with

μθ(xt,t)=1αt(xt−βt1−αˉtϵθ(xt,t)).(19)\mu_\theta(x_t,t)=\frac{1}{\sqrt{\alpha_t}}\left(x_t-\frac{\beta_t}{\sqrt{1-\bar{\alpha}_t}}\epsilon_\theta(x_t,t)\right). \tag*{(19)}

The corresponding clean estimate is

x^0(xt,t)=xt−1−αˉtϵθ(xt,t)αˉt.\hat{x}_0(x_t,t)=\frac{x_t-\sqrt{1-\bar{\alpha}_t}\epsilon_\theta(x_t,t)}{\sqrt{\bar{\alpha}_t}}.

Starting from the terminal marginal, reverse sampling uses

xt−1=μθ(xt,t)+β~tZt,x_{t-1}=\mu_\theta(x_t,t)+\sqrt{\tilde{\beta}_t}Z_t,

where Zt∼N(0,Id)Z_t\sim\mathcal{N}(0,I_d) and Z1=0Z_1=0.

Canonical Gaussian Adaptation

We train a single DDPM on the canonical distribution Z0∼N(0,Id)Z_0\sim\mathcal{N}(0,I_d) and adapt its noise predictor to X0∼N(μ,Σ)X_0\sim\mathcal{N}(\mu,\Sigma) without retraining. From (18),

Xt∼N(αˉtμ,Ct),Ct=αˉtΣ+(1−αˉt)Id.X_t\sim\mathcal{N}\left(\sqrt{\bar{\alpha}_t}\mu,C_t\right),\quad C_t=\bar{\alpha}_t\Sigma+(1-\bar{\alpha}_t)I_d.

Let Ct=RtRtTC_t=R_tR_t^\mathsf{T} be its Cholesky factorization. The transformed input

Ut=Rt−1(Xt−αˉtμ)U_t=R_t^{-1}\left(X_t-\sqrt{\bar{\alpha}_t}\mu\right)

satisfies Ut∼N(0,Id)U_t\sim\mathcal{N}(0,I_d).

Lemma 1. Let ϵθ\epsilon_\theta be the MSE-optimal noise predictor trained on the canonical distribution N(0,Id)\mathcal{N}(0,I_d). Then, the MSE-optimal noise predictor for X0∼N(μ,Σ)X_0\sim\mathcal{N}(\mu,\Sigma) is obtained without retraining as

ϵ^μ,Σ(Xt,t)=Rt−Tϵθ(Ut,t).\hat{\epsilon}_{\mu,\Sigma}(X_t,t)=R_t^{-\mathsf{T}}\epsilon_\theta(U_t,t).

Proof. Under the squared-error objective, the optimal predictor is the conditional expectation E[ϵt∣Xt]\mathbb{E}[\epsilon_t\mid X_t]. Since X0X_0 and ϵt\epsilon_t are independent Gaussian vectors,

[X0ϵt]∼N([μ0],[Σ00Id]).\begin{bmatrix} X_0 \\ \epsilon_t \end{bmatrix}\sim\mathcal{N}\left(\begin{bmatrix}\mu\\0\end{bmatrix},\begin{bmatrix}\Sigma&0\\0&I_d\end{bmatrix}\right).

By (18), (ϵt,Xt)(\epsilon_t,X_t) is a linear transformation of (X0,ϵt)(X_0,\epsilon_t) and is therefore jointly Gaussian. Its cross-covariance is

Cov⁡(ϵt,Xt)=αˉtCov⁡(ϵt,X0)+1−αˉtCov⁡(ϵt)=1−αˉtId,\operatorname{Cov}(\epsilon_t,X_t)=\sqrt{\bar{\alpha}_t}\operatorname{Cov}(\epsilon_t,X_0)+\sqrt{1-\bar{\alpha}_t}\operatorname{Cov}(\epsilon_t) =\sqrt{1-\bar{\alpha}_t}I_d,

where independence makes the first term zero.

For jointly Gaussian vectors, the conditional expectation is

E[ϵt∣Xt]=E[ϵt]+Cov⁡(ϵt,Xt)Cov⁡(Xt)−1(Xt−E[Xt]).\mathbb{E}[\epsilon_t\mid X_t]=\mathbb{E}[\epsilon_t]+\operatorname{Cov}(\epsilon_t,X_t)\operatorname{Cov}(X_t)^{-1}(X_t-\mathbb{E}[X_t]).

Using E[ϵt]=0\mathbb{E}[\epsilon_t]=0, E[Xt]=αˉtμ\mathbb{E}[X_t]=\sqrt{\bar{\alpha}_t}\mu, Cov⁡(ϵt,Xt)=1−αˉtId\operatorname{Cov}(\epsilon_t,X_t)=\sqrt{1-\bar{\alpha}_t}I_d, and Cov⁡(Xt)=Ct\operatorname{Cov}(X_t)=C_t, we obtain

E[ϵt∣Xt]=1−αˉtCt−1(Xt−αˉtμ)=1−αˉtRt−TRt−1(Xt−αˉtμ)=1−αˉtRt−TUt.\mathbb{E}[\epsilon_t\mid X_t]=\sqrt{1-\bar{\alpha}_t}C_t^{-1}(X_t-\sqrt{\bar{\alpha}_t}\mu) =\sqrt{1-\bar{\alpha}_t}R_t^{-\mathsf{T}}R_t^{-1}(X_t-\sqrt{\bar{\alpha}_t}\mu) =\sqrt{1-\bar{\alpha}_t}R_t^{-\mathsf{T}}U_t.

For the canonical model, Zt∼N(0,Id)Z_t\sim\mathcal{N}(0,I_d) and Cov⁡(ϵt,Zt)=1−αˉtId\operatorname{Cov}(\epsilon_t,Z_t)=\sqrt{1-\bar{\alpha}_t}I_d. Hence, its MSE-optimal predictor satisfies

ϵθ(Zt,t)=E[ϵt∣Zt]=1−αˉtZt.\epsilon_\theta(Z_t,t)=\mathbb{E}[\epsilon_t\mid Z_t]=\sqrt{1-\bar{\alpha}_t}Z_t.

Evaluating it at Zt=UtZ_t=U_t gives

Rt−Tϵθ(Ut,t)=1−αˉtRt−TUt=E[ϵt∣Xt],R_t^{-\mathsf{T}}\epsilon_\theta(U_t,t)=\sqrt{1-\bar{\alpha}_t}R_t^{-\mathsf{T}}U_t =\mathbb{E}[\epsilon_t\mid X_t],

which establishes (8). □\square

Conditional Sampling Reconstruction

We consider the linear observation model in (11), with the posterior mean in (12) and posterior covariance in (13). Conditional DDPM sampling incorporates the observation before the reverse process: the Gaussian posterior becomes the target Gaussian law, and the canonical adaptation in (8) is applied using ((12), (13)). Consequently, no likelihood gradient is added during the reverse steps. We generate PP samples from the conditional law and select the one that maximizes the likelihood of Gaussian observation.

In this work, conditional sampling is used only for the Gaussian channel-only experiment, with ξ0=Hk\xi_0=H_k and A=IdA=I_d. For the independent conditional method, (μ,Σ)(\mu,\Sigma) in (12) and (13) is the stationary channel prior in (5); for the temporal conditional method, it is the predictive prior in (14).

Algorithm 1 summarizes this conditional reconstruction procedure.

Algorithm 1 DDPM Conditional Sampling Reconstruction
Require: $y$, $A$, $(\mu,\Sigma)$, $\sigma_n^2$, $\epsilon_\theta$, DDPM schedule, $P$
1: Compute $(\mu_{\mathrm{post}},\Sigma_{\mathrm{post}})$ using (13) and (14)
2: for $p=1,\ldots,P$ do
3: Draw $\xi_T^{(p)}$ from the terminal marginal of the posterior forward process
4: for $t=T,\ldots,1$ do
5: $C_{\mathrm{post},t}\leftarrow\bar{\alpha}_t\Sigma_{\mathrm{post}}+(1-\bar{\alpha}_t)I_d$
6: $R_tR_t^{\mathsf T}\leftarrow C_{\mathrm{post},t}$
7: $U_t^{(p)}\leftarrow R_t^{-1}(\xi_t^{(p)}-\sqrt{\bar{\alpha}_t}\mu_{\mathrm{post}})$
8: $\widehat{\epsilon}_t^{(p)}\leftarrow R_t^{-1}\epsilon_\theta(U_t^{(p)},t)$
9: Compute $\mu_{\theta,t}^{(p)}$ using (20) with $X_t=\xi_t^{(p)}$ and $\epsilon_\theta=\widehat{\epsilon}_t^{(p)}$
10: Advance $\xi_t^{(p)}$ to $\xi_{t-1}^{(p)}\leftarrow\mu_{\theta,t}^{(p)}+\sqrt{\widetilde{\beta}_t}Z_t^{(p)}$ $\{Z_1^{(p)}=0\}$
11: end for
12: end for
13: $p^\star\leftarrow\arg\max_{1\le p\le P}p(y\mid\xi_0^{(p)})$
14: return $\widehat{\xi}_0=\xi_0^{(p^\star)}$
Algorithm 2 DDPM Likelihood-Guided Reconstruction
Require: $y$, Gaussian likelihood $p(y\mid\cdot)$, $J$ DDPM priors, $\{\eta_t^{(j)}\}$, DDPM schedule, $P$
1: for $p=1,\ldots,P$ do
2: Draw $\{\xi_T^{(j,p)}\}_{j=1}^J$ from their terminal marginals
3: for $t=T,\ldots,1$ do
4: for $j=1,\ldots,J$ do
5: Compute adapted noise prediction $\widehat{\epsilon}_t^{(j,p)}$
6: Compute clean estimate $\widehat{\xi}_{0,t}^{(j,p)}$ using (21)
7: end for
8: for $j=1,\ldots,J$ do
9: Compute likelihood gradient $g_t^{(j,p)}$ using (22)
10: Advance $\xi_t^{(j,p)}$ to $\xi_{t-1}^{(j,p)}$ using (23)
11: end for
12: end for
13: end for
14: $p^\star\leftarrow\arg\max_{1\le p\le P}p(y\mid\xi_0^{(1,p)},\ldots,\xi_0^{(J,p)})$
15: return $\{\xi_0^{(j,p^\star)}\}_{j=1}^J$

Likelihood-Guided Reconstruction

We consider the general observation model (11), where each unknown ξ0(j)\xi_0^{(j)} has a separately trained DDPM prior. This formulation includes both channel-only estimation in (8) and joint MIMO reconstruction in (5).

For Gaussian channel-only estimation, we set J=1J=1, ξ0(1)=Hk\xi_0^{(1)}=\mathbf{H}_k, and f(ξ0(1))=ξ0(1)f(\xi_0^{(1)})=\xi_0^{(1)}. The independent method uses the stationary Gaussian channel prior in (6), whereas the temporal method uses the Gaussian predictive prior in (15).

For blind joint channel–image MIMO reconstruction, we set J=2J=2, ξ0(1)=vec⁡(Hk)\xi_0^{(1)}=\operatorname{vec}(\mathbf{H}_k), ξ0(2)=vec⁡(Dk)\xi_0^{(2)}=\operatorname{vec}(\mathbf{D}_k), and f(ξ0(1),ξ0(2))=vec⁡(Hkfγ(Dk))f(\xi_0^{(1)},\xi_0^{(2)})=\operatorname{vec}(\mathbf{H}_k f_\gamma(\mathbf{D}_k)). The channel and image have separately trained DDPM priors, coupled through the MIMO likelihood. The independent method uses the stationary Gaussian channel prior in (6), whereas the temporal method uses the predictive prior in (16), treating the previous channel estimate as exact. The image prior is the same for both methods.

For each unknown and at every time step t=T,…,1t=T,\ldots,1, the predicted clean state is

ξ^0,t(j)=ξt(j)−1−αˉtϵ^t(j)αˉt,(20)\widehat{\xi}_{0,t}^{(j)}=\frac{\xi_t^{(j)}-\sqrt{1-\bar{\alpha}_t}\widehat{\epsilon}_t^{(j)}}{\sqrt{\bar{\alpha}_t}}, \tag*{(20)}

where ϵ^t(j)\widehat{\epsilon}_t^{(j)} denotes the adapted noise prediction. The observation likelihood gradients are then evaluated as

gt(j)=∇ξt(j)log⁡p(y∣ξ^0,t(1),…,ξ^0,t(J)).(21)\mathbf{g}_t^{(j)}=\nabla_{\xi_t^{(j)}}\log p\left(y\mid\widehat{\xi}_{0,t}^{(1)},\ldots,\widehat{\xi}_{0,t}^{(J)}\right). \tag*{(21)}

The corresponding guided reverse update, performed at every time step, is

ξt−1(j)=μθ,t(j)+ηt(j)gt(j)+β~tZt(j),(22)\xi_{t-1}^{(j)}=\mu_{\theta,t}^{(j)}+\eta_t^{(j)}\mathbf{g}_t^{(j)}+\sqrt{\widetilde{\beta}_t}Z_t^{(j)}, \tag*{(22)}

where μθ,t(j)\mu_{\theta,t}^{(j)} is obtained from (20), ηt(j)\eta_t^{(j)} is the guidance strength, and Z1(j)=0Z_1^{(j)}=0.

For channel-only estimation, the likelihood guides the channel reverse state. In blind MIMO reconstruction, the channel and image reverse states are both updated at every diffusion step using gradients of the same joint observation likelihood.

Among PP final candidates, we select the one maximizing the observation likelihood.

Algorithm 2 summarizes this likelihood-guided reconstruction procedure.

Appendix B: Score-Based Diffusion

Forward and Reverse Processes

A continuous score-based model defines the forward SDE

dX(t)=b(X(t),t) dt+g(t) dW(t),d\mathbf{X}(t)=\mathbf{b}(\mathbf{X}(t),t)\,dt+g(t)\,d\mathbf{W}(t),

where W(t)\mathbf{W}(t) is a standard Wiener process. We consider the linear Gaussian perturbation kernel associated with this SDE,

X(t)=m(t)X0+σ(t)ϵ(t),ϵ(t)∼N(0,Id).(23)\mathbf{X}(t)=m(t)\mathbf{X}_0+\sigma(t)\boldsymbol{\epsilon}(t),\qquad\boldsymbol{\epsilon}(t)\sim\mathcal{N}(\mathbf{0},I_d). \tag*{(23)}

Let ptp_t denote the marginal density of X(t)\mathbf{X}(t). The corresponding reverse-time SDE is

dX(t)=[b(X(t),t)−g2(t)∇X(t)log⁡pt(X(t))]dt+g(t) dW‾(t),(24)d\mathbf{X}(t)=\left[\mathbf{b}(\mathbf{X}(t),t)-g^2(t)\nabla_{\mathbf{X}(t)}\log p_t(\mathbf{X}(t))\right]dt+g(t)\,d\overline{\mathbf{W}}(t), \tag*{(24)}

which is solved from t=Tt=T to t=0t=0. The network sθ(X(t),t)s_\theta(\mathbf{X}(t),t) approximates the score ∇X(t)log⁡pt(X(t))\nabla_{\mathbf{X}(t)}\log p_t(\mathbf{X}(t)) and is substituted into (25) during generation.

For the perturbation kernel in (24), the clean estimate is

X^0(X(t),t)=X(t)+σ2(t)sθ(X(t),t)m(t).\widehat{\mathbf{X}}_0(\mathbf{X}(t),t)=\frac{\mathbf{X}(t)+\sigma^2(t)s_\theta(\mathbf{X}(t),t)}{m(t)}.

Starting from the terminal marginal, the reverse SDE is evaluated on a decreasing numerical grid of NgridN_{\mathrm{grid}} intervals, T=tNgrid>⋯>t0=0T=t_{N_{\mathrm{grid}}}>\cdots>t_0=0.

Canonical Gaussian Adaptation

We train a single score-based model on the canonical distribution Z0∼N(0,Id)Z_0 \sim\mathcal{N}(0,I_d) and adapt its score to X0∼N(μ,Σ)X_0 \sim\mathcal{N}(\mu,\Sigma) without retraining. From (23),

X(t)∼N(m(t)μ,C(t)),C(t)=m2(t)Σ+σ2(t)Id.X(t) \sim\mathcal{N}(m(t)\mu,C(t)), \qquad C(t)=m^2(t)\Sigma+\sigma^2(t)I_d.

Let C(t)=R(t)RT(t)C(t)=R(t)R^{\mathrm{T}}(t) be its Cholesky factorization. The transformed state

U(t)=R−1(t)(X(t)−m(t)μ)U(t)=R^{-1}(t)\left(X(t)-m(t)\mu\right)

has the same canonical marginal as Z(t)Z(t).

Lemma 2. Let sθs_\theta be the score trained on the canonical state Z(t)Z(t). Then, the score for X0∼N(μ,Σ)X_0 \sim\mathcal{N}(\mu,\Sigma) is obtained without retraining as

s^μ,Σ(X(t),t)=R−T(t)sθ(U(t),t).\widehat{s}_{\mu,\Sigma}(X(t),t)=R^{-\mathrm{T}}(t)s_\theta(U(t),t).

Proof. Let pX(t)(x(t))p_{X(t)}(x(t)) and pU(t)(u(t))p_{U(t)}(u(t)) denote the densities of X(t)X(t) and U(t)U(t). Since U(t)=R−1(t)(X(t)−m(t)μ)U(t)=R^{-1}(t)\left(X(t)-m(t)\mu\right) and u(t)=R−1(t)(x(t)−m(t)μ)u(t)=R^{-1}(t)\left(x(t)-m(t)\mu\right), the change-of-variables formula gives

pX(t)(x(t))=1∣det⁡R(t)∣pU(t)(u(t)).p_{X(t)}(x(t))=\frac{1}{\lvert\det R(t)\rvert}p_{U(t)}(u(t)).

Hence,

log⁡pX(t)(x(t))=−log⁡∣det⁡R(t)∣+log⁡pU(t)(u(t)).\log p_{X(t)}(x(t))=-\log\lvert\det R(t)\rvert+\log p_{U(t)}(u(t)).

Because ∇x(t)log⁡∣det⁡R(t)∣=0\nabla_{x(t)}\log\lvert\det R(t)\rvert=0, applying the chain rule therefore gives

∇x(t)log⁡pX(t)(x(t))=R−T(t)∇u(t)log⁡pU(t)(u(t)).\nabla_{x(t)}\log p_{X(t)}(x(t))=R^{-\mathrm{T}}(t)\nabla_{u(t)}\log p_{U(t)}(u(t)).

For the canonical model Z(t)∼N(0,Id)Z(t)\sim\mathcal{N}(0,I_d), the score function is sθ(Z(t),t)≈∇u(t)log⁡pU(t)(u(t))∣u(t)=Z(t)s_\theta(Z(t),t)\approx\left.\nabla_{u(t)}\log p_{U(t)}(u(t))\right|_{u(t)=Z(t)}. Evaluating it at Z(t)=U(t)Z(t)=U(t) gives

sθ(U(t),t)=∇u(t)log⁡pU(t)(u(t))∣u(t)=U(t)=R−T(t)∇x(t)log⁡pX(t)(x(t))∣x(t)=X(t)=R−T(t)sθ(X(t),t),\begin{aligned} s_\theta(U(t),t)&=\left.\nabla_{u(t)}\log p_{U(t)}(u(t))\right|_{u(t)=U(t)}\\ &=R^{-\mathrm{T}}(t)\left.\nabla_{x(t)}\log p_{X(t)}(x(t))\right|_{x(t)=X(t)}\\ &=R^{-\mathrm{T}}(t)s_\theta(X(t),t), \end{aligned}

which establishes (9).

Conditional Sampling Reconstruction

We consider the linear observation model in (11), with the posterior mean in (12) and posterior covariance in (13). Conditional score-based sampling incorporates the observation before the reverse process: the Gaussian posterior becomes the target Gaussian law, and the canonical adaptation in (9) is applied using (12),(13). Consequently, no likelihood gradient is added during the reverse steps. We generate PP samples from the conditional law and select the one maximizing the Gaussian observation likelihood.

In this work, conditional sampling is used only for the Gaussian channel-only experiment, with ξ0=Hk\xi_0=H_k and A=IdA=I_d. For the independent conditional method, (μ,Σ)(\mu,\Sigma) in (12) and (13) is the stationary channel prior in (5); for the temporal conditional method, it is the predictive prior in (14).

Algorithm 3Score-Based Conditional Sampling Reconstruction
Require:yy, AA, (μ,Σ)(\mu,\Sigma), σn2\sigma_n^2, sθs_\theta, SDE coefficients, reverse grid, PP
1:Compute (μpost,Σpost)(\mu_{\mathrm{post}},\Sigma_{\mathrm{post}}) using (12) and (13)
2:for p=1,…,Pp=1,\ldots,P do
3:Draw ξ(p)(T)\xi^{(p)}(T) from the terminal marginal of the posterior forward process
4:for ℓ=Ngrid,…,1\ell=N_{\mathrm{grid}},\ldots,1 do
5:t←tℓt\leftarrow t_\ell
6:Cpost(t)←m2(t)Σpost+σ2(t)IdC_{\mathrm{post}}(t)\leftarrow m^2(t)\Sigma_{\mathrm{post}}+\sigma^2(t)I_d
7:R(t)RT(t)←Cpost(t)R(t)R^{\mathrm{T}}(t)\leftarrow C_{\mathrm{post}}(t)
8:U(p)(t)←R−1(t)(ξ(p)(t)−m(t)μpost)U^{(p)}(t)\leftarrow R^{-1}(t)\left(\xi^{(p)}(t)-m(t)\mu_{\mathrm{post}}\right)
9:s^(p)(t)←R−T(t)sθ(U(p)(t),t)\widehat{s}^{(p)}(t)\leftarrow R^{-\mathrm{T}}(t)s_\theta(U^{(p)}(t),t)
10:Advance ξ(p)(t)\xi^{(p)}(t) to ξ(p)(tℓ−1)\xi^{(p)}(t_{\ell-1}) using (24) with score s^(p)(t)\widehat{s}^{(p)}(t)
11:end for
12:end for
13:p⋆←arg max⁡1≤p≤Pp(y∣ξ(p)(0))p^\star\leftarrow\operatorname*{arg\,max}_{1\le p\le P}p\left(y\mid\xi^{(p)}(0)\right)
14:return ξ^0=ξ(p⋆)(0)\widehat{\xi}_0=\xi^{(p^\star)}(0)

Table 3.

(13) is the stationary channel prior in (5); for the temporal conditional method, it is the predictive prior in (14).

Algorithm 3 summarizes this conditional reconstruction procedure.

Likelihood-Guided Reconstruction

We consider the general observation model (10), where each unknown ξ0(j)\xi_0^{(j)} has a separately trained score-based diffusion prior. This formulation includes both channel-only estimation in (7) and joint MIMO reconstruction in (4).

For Gaussian channel-only estimation, we set J=1J=1, ξ0(1)=Hk\xi_0^{(1)}=H_k, and f(ξ0(1))=ξ0(1)f(\xi_0^{(1)})=\xi_0^{(1)}. The independent method uses the stationary Gaussian channel prior in (5), whereas the temporal method uses the Gaussian predictive prior in (14).

For blind joint channel–image MIMO reconstruction, we set J=2J=2, ξ0(1)=vec⁡(Hk)\xi_0^{(1)}=\operatorname{vec}(H_k), ξ0(2)=vec⁡(Dk)\xi_0^{(2)}=\operatorname{vec}(D_k), and f(ξ0(1),ξ0(2))=vec⁡(Hkfγ(Dk))f(\xi_0^{(1)},\xi_0^{(2)})=\operatorname{vec}(H_k f_\gamma(D_k)). The channel and image have separately trained score-based diffusion priors, coupled through the MIMO likelihood. The independent method uses the stationary Gaussian channel prior in (5), whereas the temporal method uses the predictive prior in (15), treating the previous channel estimate as exact. The image prior is the same for both methods.

For each unknown and at every reverse time tt, the predicted clean state is

ξ^0(j)(t)=ξ(j)(t)+σj2(t)s^(j)(ξ(j)(t),t)mj(t).(25)\widehat{\xi}_0^{(j)}(t)=\frac{\xi^{(j)}(t)+\sigma_j^2(t)\widehat{s}^{(j)}(\xi^{(j)}(t),t)}{m_j(t)}. \tag*{(25)}

where s^(j)\widehat{s}^{(j)} denotes the adapted score estimate. The observation likelihood gradients are then evaluated as

g(j)(t)=∇ξ(j)(t)log⁡p(y∣ξ^0(1)(t),…,ξ^0(J)(t)).(26)\mathbf{g}^{(j)}(t)=\nabla_{\xi^{(j)}(t)}\log p\left(y\mid\widehat{\xi}_0^{(1)}(t),\ldots,\widehat{\xi}_0^{(J)}(t)\right). \tag*{(26)}

The corresponding guided score, computed at every reverse time tt, is

sguided(j)(t)=s^(j)(ξ(j)(t),t)+η(j)(t)g(j)(t).(27)\mathbf{s}_{\mathrm{guided}}^{(j)}(t)=\widehat{\mathbf{s}}^{(j)}(\xi^{(j)}(t),t)+\eta^{(j)}(t)\mathbf{g}^{(j)}(t). \tag*{(27)}

Algorithm 4 Score-Based Likelihood-Guided Reconstruction

Require: yy, Gaussian likelihood p(y∣⋅)p(y \mid\cdot), JJ score-based priors, {η(j)(t)}\{\eta^{(j)}(t)\}, SDE coefficients, reverse grid, PP

  1. for p=1,…,Pp = 1,\ldots,P do

  2. Draw {ξ(j,p)(T)}j=1J\{\xi^{(j,p)}(T)\}_{j=1}^{J} from their terminal marginals

  3. for ℓ=Ngrid,…,1\ell= N_{\mathrm{grid}},\ldots,1 do

  4. t←tℓt \leftarrow t_{\ell}

  5. for j=1,…,Jj = 1,\ldots,J do

  6. Compute adapted prior score s^(j,p)(t)\hat{s}^{(j,p)}(t)

  7. Compute clean estimate ξ^0(j,p)(t)\hat{\xi}_{0}^{(j,p)}(t) using (25)

  8. end for

  9. for j=1,…,Jj = 1,\ldots,J do

  10. Compute likelihood gradient g(j,p)(t)g^{(j,p)}(t) using (26)

  11. Form guided score sguided(j,p)(t)s_{\mathrm{guided}}^{(j,p)}(t) using (27)

  12. Advance ξ(j,p)(t)\xi^{(j,p)}(t) to ξ(j,p)(tℓ−1)\xi^{(j,p)}(t_{\ell-1}) using (28)

  13. end for

  14. end for

  15. end for

  16. p⋆←arg⁡max⁡1≤p≤Pp(y∣ξ(1,p)(0),…,ξ(J,p)(0))p^{\star} \leftarrow\arg\max_{1\leq p\leq P} p(y \mid\xi^{(1,p)}(0),\ldots,\xi^{(J,p)}(0))

  17. return {ξ(j,p⋆)(0)}j=1J\{\xi^{(j,p^{\star})}(0)\}_{j=1}^{J}

where η(j)(t)\eta^{(j)}(t) is the guidance strength. The guided reverse SDE is

dξ(j)(t)=[bj(ξ(j)(t),t)−gj2(t)sguided(j)(t)]dt+gj(t)dW‾(j)(t),(28)\mathrm{d}\xi^{(j)}(t)=\left[\mathbf{b}_{j}(\xi^{(j)}(t),t)-g_{j}^{2}(t)s_{\mathrm{guided}}^{(j)}(t)\right]\mathrm{d}t+g_{j}(t)\mathrm{d}\overline{W}^{(j)}(t), \tag*{(28)}

which is integrated backward from TT to 0.

For channel-only estimation, the likelihood guides the channel reverse state. In blind MIMO reconstruction, the channel and image reverse states are both updated at every reverse step using gradients of the same joint observation likelihood.

Among PP final candidates, we select the one maximizing the observation likelihood.

Algorithm 4 summarizes this likelihood-guided reconstruction procedure.

Appendix C: DLPM

Forward and Reverse Processes

Let α∈(1,2)\alpha\in(1,2) and let {βt}t=1T\{\beta_t\}_{t=1}^{T} be a noise schedule. Define

ct=(1−βt)1/α,σt=βt1/α,cˉt=∏s=1tcs.c_t=(1-\beta_t)^{1/\alpha}, \qquad\sigma_t=\beta_t^{1/\alpha}, \qquad\bar{c}_t=\prod_{s=1}^{t}c_s.

A DLPM defines the forward process

q(Xt∣Xt−1=xt−1)=SαS(ctxt−1,σt).q(X_t \mid X_{t-1}=x_{t-1})=S\alpha S(c_tx_{t-1},\sigma_t).

Equivalently,

Xt=cˉtX0+stϵt,st=(1−cˉtα)1/α,ϵt∼SαS(0,1).(29)X_t=\bar{c}_tX_0+s_t\epsilon_t,\qquad s_t=(1-\bar{c}_t^\alpha)^{1/\alpha},\qquad\epsilon_t\sim S\alpha S(0,1). \tag*{(29)}

Using the Gaussian scale-mixture representation in (1), each perturbation is written as

ϵt=dAt Gt,Gt∼N(0,Id),\epsilon_t \overset{\mathrm{d}}{=}\sqrt{A_t}\,G_t,\qquad G_t\sim\mathcal{N}(0,I_d),

where A1,…,AtA_1,\ldots,A_t are independent positive α/2\alpha/2-stable variables with skewness βA=1\beta_A=1, scale γA=cos⁡2/α(πα/4)\gamma_A=\cos^{2/\alpha}(\pi\alpha/4), and location δA=0\delta_A=0. Conditioned on A1:t=(A1,…,At)A_{1:t}=(A_1,\ldots,A_t), define

V0=0,Vt=ct2Vt−1+σt2At.(30)V_0=0,\qquad V_t=c_t^2V_{t-1}+\sigma_t^2A_t. \tag*{(30)}

Then

Xt∣(X0=x0,A1:t)∼N(cˉtx0,VtId).X_t\mid(X_0=x_0,A_{1:t})\sim\mathcal{N}(\bar{c}_tx_0,V_tI_d).

The exact reverse posterior conditioned on X0=x0X_0=x_0 and A1:tA_{1:t} is

q(Xt−1∣Xt=xt,X0=x0,A1:t)=N(μ~t,V~tId),q(X_{t-1}\mid X_t=x_t,X_0=x_0,A_{1:t})=\mathcal{N}(\widetilde{\mu}_t,\widetilde{V}_tI_d),

where

μ~t=ctVt−1Vtxt+cˉt−1σt2AtVtx0,\widetilde{\mu}_t=\frac{c_tV_{t-1}}{V_t}x_t+\frac{\bar{c}_{t-1}\sigma_t^2A_t}{V_t}x_0,
V~t=σt2AtVt−1Vt.\widetilde{V}_t=\frac{\sigma_t^2A_tV_{t-1}}{V_t}.

The learned reverse transition replaces the unknown marginal perturbation by ϵθ(xt,t)\epsilon_\theta(x_t,t). Let

κt=σt2AtVt,rt=κtVt−1.\kappa_t=\frac{\sigma_t^2A_t}{V_t},\qquad r_t=\kappa_tV_{t-1}.

The reverse transition is

pθ(Xt−1∣Xt=xt,A1:T)=N(μθ(xt,t,A1:T),rtId),p_\theta(X_{t-1}\mid X_t=x_t,A_{1:T})=\mathcal{N}(\mu_\theta(x_t,t,A_{1:T}),r_tI_d),

with

μθ(xt,t,A1:T)=xt−κtstϵθ(xt,t)ct.\mu_\theta(x_t,t,A_{1:T})=\frac{x_t-\kappa_ts_t\epsilon_\theta(x_t,t)}{c_t}.

The corresponding clean estimate is

x^0(xt,t)=xt−stϵθ(xt,t)cˉt.\widehat{x}_0(x_t,t)=\frac{x_t-s_t\epsilon_\theta(x_t,t)}{\bar{c}_t}.

At the beginning of each reverse trajectory, one independent path A1:TA_{1:T} is drawn and (30) is evaluated. Starting from the terminal marginal, reverse sampling uses

xt−1=μθ(xt,t,A1:T)+rt G~t,G~t∼N(0,Id),x_{t-1}=\mu_\theta(x_t,t,A_{1:T})+\sqrt{r_t}\,\widetilde{G}_t,\qquad\widetilde{G}_t\sim\mathcal{N}(0,I_d),

where G~1=0\widetilde{G}_1=0.

Canonical Alpha-Stable Adaptation

We train one DLPM on the canonical distribution Z0∼SαS(0,1)Z_0\sim S\alpha S(0,1). The DLPM experiments use unit scale γ=1\gamma=1; accordingly, the following adaptation changes only the location. For the unit-scale target distribution

X0=μ+Z0∼SαS(μ,1),X_0=\mu+Z_0\sim S\alpha S(\mu,1),

the forward marginal in (29) gives

Xt=cˉtμ+Zt.X_t=\bar{c}_t\mu+Z_t.

Lemma 3. Let ϵθ\epsilon_\theta be the stable-noise predictor trained on the canonical distribution. The canonical input and transported prediction are

Ut=Xt−cˉtμ,ϵ^μ,1(Xt,t)=ϵθ(Ut,t).(31)U_t = X_t - \bar{c}_t\mu,\qquad\widehat{\epsilon}_{\mu,1}(X_t,t) = \epsilon_\theta(U_t,t). \tag*{(31)}

Proof. Substituting X0=μ+Z0X_0 = \mu+ Z_0 into (29) gives

Xt=cˉt(μ+Z0)+stϵˉtX_t = \bar{c}_t(\mu+ Z_0) + s_t\bar{\epsilon}_t
=cˉtμ+Zt.= \bar{c}_t\mu+ Z_t.

Therefore, Xt−cˉtμX_t-\bar{c}_t\mu has the same distribution as the canonical state ZtZ_t. Evaluating the canonical predictor at Ut=Xt−cˉtμU_t = X_t-\bar{c}_t\mu proves (31).

Likelihood-Guided Reconstruction

We consider the general observation model (10), where each unknown ξ0(j)\xi_0^{(j)} has a separately trained diffusion prior. This formulation includes both channel-only estimation in (7) and joint MIMO reconstruction in (4).

For α\alpha-stable channel-only estimation, we set J=1J=1, ξ0(1)=Hk\xi_0^{(1)}=\mathbf{H}_k, and f(ξ0(1))=ξ0(1)f(\xi_0^{(1)})=\xi_0^{(1)}. The independent method uses the stationary α\alpha-stable channel prior in (6), whereas the temporal method uses the predictive prior in (16).

For blind joint channel–image MIMO reconstruction, we set J=2J=2, ξ0(1)=vec⁡(Hk)\xi_0^{(1)}=\operatorname{vec}(\mathbf{H}_k), ξ0(2)=vec⁡(Dk)\xi_0^{(2)}=\operatorname{vec}(\mathbf{D}_k), and f(ξ0(1),ξ0(2))=vec⁡(Hkfγ(Dk))f(\xi_0^{(1)},\xi_0^{(2)})=\operatorname{vec}(\mathbf{H}_k f_\gamma(\mathbf{D}_k)). The channel and image have separately trained diffusion priors, coupled through the MIMO likelihood. The independent method uses the stationary α\alpha-stable channel prior in (6), whereas the temporal method uses the predictive prior in (17). The image prior is the same for both methods.

The DLPM experiments use likelihood guidance. For the observation model in (10),

Y=f(ξ0(1),…,ξ0(J))+N,N∼SαS(0,γn),\mathbf{Y}=f\left(\xi_0^{(1)},\ldots,\xi_0^{(J)}\right)+\mathbf{N},\qquad\mathbf{N}\sim S\alpha S(0,\gamma_n),

Using the Gaussian scale-mixture representation in (1) gives

Y∣ξ0(1),…,ξ0(J),B∼N(f(ξ0(1),…,ξ0(J)),γn2BId),\mathbf{Y}\mid\xi_0^{(1)},\ldots,\xi_0^{(J)},B\sim\mathcal{N}\left(f\left(\xi_0^{(1)},\ldots,\xi_0^{(J)}\right),\gamma_n^2B\mathbf{I}_d\right),

where BB is positive α/2\alpha/2-stable. Using MMCM_{\mathrm{MC}} independent samples B(1),…,B(MMC)B^{(1)},\ldots,B^{(M_{\mathrm{MC}})}, the likelihood is approximated by

p^(y∣ξ0(1),…,ξ0(J))=1MMC∑m=1MMCN(y;f(ξ0(1),…,ξ0(J)),γn2B(m)Id).\widehat{p}\left(\mathbf{y}\mid\xi_0^{(1)},\ldots,\xi_0^{(J)}\right) = \frac{1}{M_{\mathrm{MC}}}\sum_{m=1}^{M_{\mathrm{MC}}} \mathcal{N}\left(\mathbf{y};f\left(\xi_0^{(1)},\ldots,\xi_0^{(J)}\right),\gamma_n^2B^{(m)}\mathbf{I}_d\right).

Samples B(m)B^{(m)} are used only to approximate the likelihood of measurement and are independent of the DLPM trajectory variables A1:TA_{1:T}.

For each unknown DLPM and at each time step t=T,…,1t=T,\ldots,1, the predicted clean state is

ξ^0,t(j)=ξt(j)−st(j)ϵ^t(j)cˉt(j).(32)\widehat{\xi}_{0,t}^{(j)}= \frac{\xi_t^{(j)}-s_t^{(j)}\widehat{\epsilon}_t^{(j)}}{\bar{c}_t^{(j)}}. \tag*{(32)}

The observation likelihood gradients are then evaluated as

gt(j)=∇ξt(j)log⁡p^(y∣ξ^0,t(1),…,ξ^0,t(J)).(33)\mathbf{g}_t^{(j)}=\nabla_{\xi_t^{(j)}}\log\widehat{p}\left(\mathbf{y}\mid\widehat{\xi}_{0,t}^{(1)},\ldots,\widehat{\xi}_{0,t}^{(J)}\right). \tag*{(33)}

The corresponding guided reverse update, performed at every time step, is

ξt−1(j)=μθ,t(j)+ηt(j)gt(j)+rt(j)G~t(j),\xi_{t-1}^{(j)}=\mu_{\theta,t}^{(j)}+\eta_t^{(j)}\mathbf{g}_t^{(j)}+\sqrt{r_t^{(j)}}\widetilde{\mathbf{G}}_t^{(j)},

where ηt(j)\eta_t^{(j)} is the guidance weight and G~t(j)∼N(0,Id)\widetilde{\mathbf{G}}_t^{(j)}\sim\mathcal{N}(0,\mathbf{I}_d), with G~1(j)=0\widetilde{\mathbf{G}}_1^{(j)}=0.

For channel-only estimation, the likelihood guides the channel reverse state.

In blind MIMO reconstruction, the channel and image reverse states are both updated at every diffusion step using gradients of the same joint observation likelihood. Among PP final candidates, we select the one maximizing the approximated observation likelihood.

Algorithm 5 summarizes this likelihood-guided reconstruction procedure.

Algorithm 5 DLPM Likelihood-Guided Reconstruction

Require: y\mathbf{y}, approximated likelihood p^(y∣⋅)\widehat{p}(\mathbf{y}\mid\cdot), JJ diffusion priors, {ηt(j)}\{\eta_t^{(j)}\}, diffusion schedules, MMCM_{\mathrm{MC}}, PP

  1. Draw the measurement-likelihood samples {B(m)}m=1MMC\{B^{(m)}\}_{m=1}^{M_{\mathrm{MC}}}

  2. for p=1,…,Pp=1,\ldots,P do

  3. &nbsp;&nbsp;&nbsp;&nbsp;Draw an independent A1:T(j,p)A_{1:T}^{(j,p)} for each DLPM prior

  4. &nbsp;&nbsp;&nbsp;&nbsp;Compute the corresponding {Vt(j,p)}t=1T\{V_t^{(j,p)}\}_{t=1}^{T} using (30)

  5. &nbsp;&nbsp;&nbsp;&nbsp;Draw {ξT(j,p)}j=1J\{\xi_T^{(j,p)}\}_{j=1}^{J} from their terminal marginals

  6. &nbsp;&nbsp;&nbsp;&nbsp;for t=T,…,1t=T,\ldots,1 do

  7. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;for j=1,…,Jj=1,\ldots,J do

  8. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Compute adapted prediction ϵ^t(j,p)\widehat{\epsilon}_t^{(j,p)}

  9. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Compute clean estimate ξ^0,t(j,p)\widehat{\xi}_{0,t}^{(j,p)} using (32) for DLPM

  10. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;end for

  11. &nbsp;&nbsp;&nbsp;&nbsp;for j=1,…,Jj=1,\ldots,J do

  12. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Compute likelihood gradient gt(j,p)\mathbf{g}_t^{(j,p)} using (33)

  13. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;Advance ξt(j,p)\xi_t^{(j,p)} to ξt−1(j,p)\xi_{t-1}^{(j,p)} using the guided reverse update of its prior

  14. &nbsp;&nbsp;&nbsp;&nbsp;end for

  15. &nbsp;&nbsp;&nbsp;&nbsp;end for

  16. end for

  17. if channel-only reconstruction then

  18. &nbsp;&nbsp;&nbsp;&nbsp;ξ^0=median⁡p=1,…,P{ξ0(1,p)}\widehat{\xi}_0=\operatorname{median}_{p=1,\ldots,P}\left\{\xi_0^{(1,p)}\right\}

  19. &nbsp;&nbsp;&nbsp;&nbsp;return ξ^0\widehat{\xi}_0

  20. else

  21. &nbsp;&nbsp;&nbsp;&nbsp;p⋆←arg⁡max⁡1≤p≤Pp^(y∣ξ0(1,p),…,ξ0(J,p))p^\star\leftarrow\arg\max_{1\le p\le P}\widehat{p}\left(\mathbf{y}\mid\xi_0^{(1,p)},\ldots,\xi_0^{(J,p)}\right)

  22. &nbsp;&nbsp;&nbsp;&nbsp;return {ξ0(j,p⋆)}j=1J\left\{\xi_0^{(j,p^\star)}\right\}_{j=1}^{J}

  23. end if

LIM

Forward and Reverse Processes

A continuous LIM defines the forward SDE

dX(t)=b(X(t),t) dt+g(t) dLα(t),\mathrm{d}\mathbf{X}(t)=\mathbf{b}(\mathbf{X}(t),t)\,\mathrm{d}t+g(t)\,\mathrm{d}L^\alpha(t),

where Lα(t)L^\alpha(t) is a symmetric α\alpha-stable Lévy process. We consider the linear α\alpha-stable perturbation kernel associated with this SDE,

X(t)=a(t)X0+σ(t)ϵ(t),ϵ(t)∼SαS(0,1),(34)\mathbf{X}(t)=a(t)\mathbf{X}_0+\sigma(t)\boldsymbol{\epsilon}(t),\qquad\boldsymbol{\epsilon}(t)\sim S\alpha S(0,1), \tag*{(34)}

Let pX(t)p_{\mathbf{X}(t)} denote the marginal density of X(t)\mathbf{X}(t). Its fractional score is

SX(t)(α)(x(t))=(−Δ)α−22∇pX(t)(x(t))pX(t)(x(t)).(35)\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}(\mathbf{x}(t))=\frac{(-\Delta)^{\frac{\alpha-2}{2}}\nabla p_{\mathbf{X}(t)}(\mathbf{x}(t))}{p_{\mathbf{X}(t)}(\mathbf{x}(t))}. \tag*{(35)}

The corresponding reverse-time dynamics used by LIM are

dX(t)=[b(X(t),t)−gα(t)SX(t)(α)(X(t))]dt,(36)\mathrm{d}\mathbf{X}(t)=\left[\mathbf{b}(\mathbf{X}(t),t)-g^{\alpha}(t)\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}(\mathbf{X}(t))\right]\mathrm{d}t, \tag*{(36)}

which are solved from t=Tt=T to t=0t=0. The network Sθ(α)(X(t),t)\mathbf{S}_{\theta}^{(\alpha)}(\mathbf{X}(t),t) approximates the fractional score SX(t)(α)(X(t))\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}(\mathbf{X}(t)) and is substituted into (37) during generation.

For the perturbation kernel in (35), the clean estimate is

X^0(X(t),t)=X(t)−σ(t)ϵ^(X(t),t)a(t).\widehat{\mathbf{X}}_0(\mathbf{X}(t),t)=\frac{\mathbf{X}(t)-\sigma(t)\widehat{\boldsymbol{\epsilon}}(\mathbf{X}(t),t)}{a(t)}.

Starting from the terminal marginal, the reverse dynamics are evaluated on a decreasing numerical grid of NgridN_{\mathrm{grid}} intervals, T=tNgrid>⋯>t0=0T=t_{N_{\mathrm{grid}}}>\cdots>t_0=0.

Canonical Alpha-Stable Adaptation

We train one LIM on the canonical distribution Z0∼SαS⁡(0,1)\mathbf{Z}_0\sim\operatorname{S\alpha S}(0,1) and adapt its fractional score to X0∼SαS⁡(μ,γ)\mathbf{X}_0\sim\operatorname{S\alpha S}(\boldsymbol{\mu},\gamma) without retraining. From (35),

X(t)=a(t)μ+γZ(t),U(t)=X(t)−a(t)μγ.\mathbf{X}(t)=a(t)\boldsymbol{\mu}+\gamma\mathbf{Z}(t),\qquad\mathbf{U}(t)=\frac{\mathbf{X}(t)-a(t)\boldsymbol{\mu}}{\gamma}.

Thus, U(t)\mathbf{U}(t) has the same canonical marginal as Z(t)\mathbf{Z}(t).

Lemma 4. Let Sθ(α)\mathbf{S}_{\theta}^{(\alpha)} be the fractional score trained on the canonical state Z(t)\mathbf{Z}(t). The fractional score for X0∼SαS⁡(μ,γ)\mathbf{X}_0\sim\operatorname{S\alpha S}(\boldsymbol{\mu},\gamma) is obtained without retraining as

S^μ,γ(α)(X(t),t)=γ1−αSθ(α)(U(t),t).(37)\widehat{\mathbf{S}}_{\boldsymbol{\mu},\gamma}^{(\alpha)}(\mathbf{X}(t),t)=\gamma^{1-\alpha}\mathbf{S}_{\theta}^{(\alpha)}(\mathbf{U}(t),t). \tag*{(37)}

Proof. Let pX(t)p_{\mathbf{X}(t)} and pU(t)p_{\mathbf{U}(t)} denote the densities of X(t)\mathbf{X}(t) and U(t)\mathbf{U}(t). From the affine transformation, U(t)=γ−1(X(t)−a(t)μ)\mathbf{U}(t)=\gamma^{-1}(\mathbf{X}(t)-a(t)\boldsymbol{\mu}). The density transformation is

pU(t)(U(t))=γdpX(t)(a(t)μ+γU(t)).p_{\mathbf{U}(t)}(\mathbf{U}(t))=\gamma^{d}p_{\mathbf{X}(t)}(a(t)\boldsymbol{\mu}+\gamma\mathbf{U}(t)).

Using q^(f)=∫q(X(t))e−i2πfTX(t)dX(t)\widehat{q}(\mathbf{f})=\int q(\mathbf{X}(t))e^{-i2\pi\mathbf{f}^{T}\mathbf{X}(t)}\mathrm{d}\mathbf{X}(t),

p^U(t)(f)=ei2πfTa(t)μ/γp^X(t)(fγ).\widehat{p}_{\mathbf{U}(t)}(\mathbf{f})=e^{i2\pi\mathbf{f}^{T}a(t)\boldsymbol{\mu}/\gamma}\widehat{p}_{\mathbf{X}(t)}\left(\frac{\mathbf{f}}{\gamma}\right).

By the fractional-score definition in (36),

F{SX(t)(α)pX(t)}(f)=i2πf(2π∥f∥2)α−2p^X(t)(f).\mathcal{F}\left\{\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}p_{\mathbf{X}(t)}\right\}(\mathbf{f})=i2\pi\mathbf{f}(2\pi\lVert\mathbf{f}\rVert_2)^{\alpha-2}\widehat{p}_{\mathbf{X}(t)}(\mathbf{f}).
F{SX(t)(α)pX(t)}(fγ)=i2πfγ(2π∥fγ∥2)α−2p^X(t)(fγ)=γ1−αi2πf(2π∥f∥2)α−2p^X(t)(fγ).\begin{aligned} \mathcal{F}\left\{\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}p_{\mathbf{X}(t)}\right\}\left(\frac{\mathbf{f}}{\gamma}\right) &=i2\pi\frac{\mathbf{f}}{\gamma}\left(2\pi\left\lVert\frac{\mathbf{f}}{\gamma}\right\rVert_2\right)^{\alpha-2}\widehat{p}_{\mathbf{X}(t)}\left(\frac{\mathbf{f}}{\gamma}\right)\\ &=\gamma^{1-\alpha}i2\pi\mathbf{f}(2\pi\lVert\mathbf{f}\rVert_2)^{\alpha-2}\widehat{p}_{\mathbf{X}(t)}\left(\frac{\mathbf{f}}{\gamma}\right). \end{aligned}

Therefore,

F{SU(t)(α)pU(t)}(f)=i2πf(2π∥f∥2)α−2p^U(t)(f)=γd+α−1ei2πfTa(t)μ/γF{SX(t)(α)pX(t)}(fγ).\begin{aligned} \mathcal{F}\left\{\mathbf{S}_{\mathbf{U}(t)}^{(\alpha)}p_{\mathbf{U}(t)}\right\}(\mathbf{f}) =i2\pi\mathbf{f}(2\pi\lVert\mathbf{f}\rVert_2)^{\alpha-2}\widehat{p}_{\mathbf{U}(t)}(\mathbf{f}) \\ &=\gamma^{d+\alpha-1}e^{i2\pi\mathbf{f}^{T}a(t)\boldsymbol{\mu}/\gamma}\mathcal{F}\left\{\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}p_{\mathbf{X}(t)}\right\}\left(\frac{\mathbf{f}}{\gamma}\right). \end{aligned}

The Fourier scaling-and-translation property also gives

FU(t){SX(t)(α)(a(t)μ+γU(t))pX(t)(a(t)μ+γU(t))}(f)=γ−dei2πfTa(t)μ/γFX(t){SX(t)(α)pX(t)}(fγ).\mathcal{F}_{\mathbf{U}(t)}\left\{\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}(a(t)\boldsymbol{\mu}+\gamma\mathbf{U}(t))p_{\mathbf{X}(t)}(a(t)\boldsymbol{\mu}+\gamma\mathbf{U}(t))\right\}(\mathbf{f}) =\gamma^{-d}e^{i2\pi\mathbf{f}^{T}a(t)\boldsymbol{\mu}/\gamma}\mathcal{F}_{\mathbf{X}(t)}\left\{\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}p_{\mathbf{X}(t)}\right\}\left(\frac{\mathbf{f}}{\gamma}\right).

Hence,

FU(t){SU(t)(α)pU(t)}(f)=γd+α−1FU(t){SX(t)(α)(a(t)μ+γU(t))pX(t)(a(t)μ+γU(t))}(f).\mathcal{F}_{\mathbf{U}(t)}\left\{\mathbf{S}_{\mathbf{U}(t)}^{(\alpha)}p_{\mathbf{U}(t)}\right\}(\mathbf{f}) =\gamma^{d+\alpha-1}\mathcal{F}_{\mathbf{U}(t)}\left\{\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}(a(t)\boldsymbol{\mu}+\gamma\mathbf{U}(t))p_{\mathbf{X}(t)}(a(t)\boldsymbol{\mu}+\gamma\mathbf{U}(t))\right\}(\mathbf{f}).

Taking the inverse Fourier transform gives

SU(t)(α)(U(t))pU(t)(U(t))=γd+α−1SX(t)(α)(a(t)μ+γU(t))pX(t)(a(t)μ+γU(t)).\mathbf{S}_{\mathbf{U}(t)}^{(\alpha)}(\mathbf{U}(t))p_{\mathbf{U}(t)}(\mathbf{U}(t)) =\gamma^{d+\alpha-1}\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}(a(t)\boldsymbol{\mu}+\gamma\mathbf{U}(t))p_{\mathbf{X}(t)}(a(t)\boldsymbol{\mu}+\gamma\mathbf{U}(t)).

Dividing by pU(t)(U(t))=γdpX(t)(a(t)μ+γU(t))p_{\mathbf{U}(t)}(\mathbf{U}(t))=\gamma^{d}p_{\mathbf{X}(t)}(a(t)\boldsymbol{\mu}+\gamma\mathbf{U}(t)) gives

SU(t)(α)(U(t))=γα−1SX(t)(α)(X(t)),\mathbf{S}_{\mathbf{U}(t)}^{(\alpha)}(\mathbf{U}(t))=\gamma^{\alpha-1}\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}(\mathbf{X}(t)),

and hence

SX(t)(α)(X(t))=γ1−αSU(t)(α)(U(t))\mathbf{S}_{\mathbf{X}(t)}^{(\alpha)}(\mathbf{X}(t))=\gamma^{1-\alpha}\mathbf{S}_{\mathbf{U}(t)}^{(\alpha)}(\mathbf{U}(t))

establishes (38).

Likelihood-Guided Reconstruction

We consider the general observation model (11), where each unknown ξ0(j)\boldsymbol{\xi}_0^{(j)} has a separately trained diffusion prior. This formulation includes both channel-only estimation in (8) and joint MIMO reconstruction in (5).

For α\alpha-stable channel-only estimation, we set J=1J=1, ξ0(1)=Hk\boldsymbol{\xi}_0^{(1)}=\mathbf{H}_k, and f(ξ0(1))=ξ0(1)f(\boldsymbol{\xi}_0^{(1)})=\boldsymbol{\xi}_0^{(1)}. The independent method uses the stationary α\alpha-stable channel prior in (7), whereas the temporal method uses the predictive prior in (17).

For blind joint channel–image MIMO reconstruction, we set J=2J=2, ξ0(1)=vec⁡(Hk)\boldsymbol{\xi}_0^{(1)}=\operatorname{vec}(\mathbf{H}_k), ξ0(2)=vec⁡(Dk)\boldsymbol{\xi}_0^{(2)}=\operatorname{vec}(\mathbf{D}_k), and f(ξ0(1),ξ0(2))=vec⁡(Hkfγ(Dk))f(\boldsymbol{\xi}_0^{(1)},\boldsymbol{\xi}_0^{(2)})=\operatorname{vec}(\mathbf{H}_k f_{\gamma}(\mathbf{D}_k)). The channel and image have separately trained diffusion priors, coupled through the MIMO likelihood. The independent method uses the stationary α\alpha-stable channel prior in (7), whereas the temporal method uses the predictive prior in (18). The image prior is the same for both methods.

The LIM experiments use likelihood guidance. For the observation model in (11),

Y=f(ξ0(1),…,ξ0(J))+N,N∼SαS⁡(0,γn),\mathbf{Y}=f\left(\boldsymbol{\xi}_0^{(1)},\ldots,\boldsymbol{\xi}_0^{(J)}\right)+\mathbf{N},\qquad\mathbf{N}\sim\operatorname{S\alpha S}(0,\gamma_n),

Using the Gaussian scale-mixture representation in (1) gives

Y∣ξ0(1),…,ξ0(J),B∼N(f(ξ0(1),…,ξ0(J)),γn2BId),\mathbf{Y}\mid\boldsymbol{\xi}_0^{(1)},\ldots,\boldsymbol{\xi}_0^{(J)},B\sim\mathcal{N}\left(f\left(\boldsymbol{\xi}_0^{(1)},\ldots,\boldsymbol{\xi}_0^{(J)}\right),\gamma_n^2B\mathbf{I}_d\right),

where BB is positive α/2\alpha/2-stable. Using MMCM_{\mathrm{MC}} independent samples B(1),…,B(MMC)B^{(1)}, \ldots, B^{(M_{\mathrm{MC}})}, the likelihood is approximated by

p^(y∣ξ0(1),…,ξ0(J))=1MMC∑m=1MMCN(y;f(ξ0(1),…,ξ0(J)),γn2B(m)∣Id).\widehat{p}\left(y \mid\xi_{0}^{(1)}, \ldots, \xi_{0}^{(J)}\right)=\frac{1}{M_{\mathrm{MC}}}\sum_{m=1}^{M_{\mathrm{MC}}}\mathcal{N}\left(y; f\left(\xi_{0}^{(1)}, \ldots, \xi_{0}^{(J)}\right), \gamma_n^2 B^{(m)} \mid I_d\right).

For each LIM unknown and at every reverse time tt, the predicted clean state is

ξ^0(j)(t)=ξ(j)(t)−σj(t)ϵ^(j)(ξ(j)(t),t)aj(t).(38)\widehat{\xi}_{0}^{(j)}(t)=\frac{\xi^{(j)}(t)-\sigma_j(t)\widehat{\epsilon}^{(j)}\left(\xi^{(j)}(t),t\right)}{a_j(t)}. \tag*{(38)}

The observation likelihood gradients are then evaluated as

g(j)(t)=∇ξ(j)(t)log⁡p^(y∣ξ^0(1)(t),…,ξ^0(J)(t)).(39)\mathbf{g}^{(j)}(t)=\nabla_{\xi^{(j)}(t)}\log\widehat{p}\left(y \mid\widehat{\xi}_{0}^{(1)}(t),\ldots,\widehat{\xi}_{0}^{(J)}(t)\right). \tag*{(39)}

At every reverse time, the prior probability-flow step from tℓt_\ell to tℓ−1t_{\ell-1} is first evaluated using (36). The likelihood correction is then applied as

ξ(j)(tℓ−1)=ξprior(j)(tℓ−1)+η(j)(tℓ)∣tℓ−tℓ−1∣g(j)(tℓ),(40)\xi^{(j)}(t_{\ell-1})=\xi_{\mathrm{prior}}^{(j)}(t_{\ell-1})+\eta^{(j)}(t_\ell)|t_\ell-t_{\ell-1}|\mathbf{g}^{(j)}(t_\ell), \tag*{(40)}

where η(j)(t)\eta^{(j)}(t) is the guidance strength.

Algorithm 6 LIM Likelihood-Guided Reconstruction

Require: yy, approximated likelihood p^(y∣⋅)\widehat{p}(y\mid\cdot), JJ diffusion priors, {η(j)(t)}\{\eta^{(j)}(t)\}, reverse grid, MMCM_{\mathrm{MC}}, PP

  1. Draw the measurement-likelihood samples {B(m)}m=1MMC\{B^{(m)}\}_{m=1}^{M_{\mathrm{MC}}}

  2. for p=1,…,Pp=1,\ldots,P do

  3. Draw {ξ(j,p)(T)}j=1J\{\xi^{(j,p)}(T)\}_{j=1}^{J} from their terminal marginals

  4. for ℓ=Ngrid,…,1\ell=N_{\mathrm{grid}},\ldots,1 do

  5. t←tℓt\leftarrow t_\ell

  6. for j=1,…,Jj=1,\ldots,J do

  7. Compute adapted fractional score and perturbation estimate

  8. Compute clean estimate ξ^0(j,p)(t)\widehat{\xi}_{0}^{(j,p)}(t) using (39)

  9. end for

  10. for j=1,…,Jj=1,\ldots,J do

  11. Compute likelihood gradient g(j,p)(t)\mathbf{g}^{(j,p)}(t) using (40)

  12. Advance ξ(j,p)(t)\xi^{(j,p)}(t) to ξ(j,p)(tℓ−1)\xi^{(j,p)}(t_{\ell-1}) using (41)

  13. end for

  14. end for

  15. end for

  16. if channel-only reconstruction then

  17. ξ^00=median⁡p=1,…,P{ξ0(1,p)}\widehat{\xi}_0^0=\operatorname{median}_{p=1,\ldots,P}\left\{\xi_0^{(1,p)}\right\}

  18. return ξ^00\widehat{\xi}_0^0

  19. else

  20. p⋆←arg max⁡1≤p≤Pp^(y∣ξ0(1,p),…,ξ0(J,p))p^\star\leftarrow\operatorname*{arg\,max}_{1\le p\le P}\widehat{p}\left(y\mid\xi_0^{(1,p)},\ldots,\xi_0^{(J,p)}\right)

  21. return {ξ0(j,p⋆)}j=1J\left\{\xi_0^{(j,p^\star)}\right\}_{j=1}^{J}

  22. end if

For channel-only estimation, the likelihood guides the channel reverse state. In blind MIMO reconstruction, the channel and image reverse states are both updated at every reverse step using gradients of the same joint observation likelihood. Among PP final candidates, we select the one maximizing the approximated observation likelihood. Algorithm 6 summarizes this likelihood-guided reconstruction procedure.

Appendix E: Additional Continuous-Time Results

References

  1. [1]J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems (NeurIPS), 2020.arxiv.org/abs/2006.11239
  2. [2]Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2021.
  3. [3]H. Chung, K. J. Kim, M. T. McCann, M. L. Klasky, and J. C. Ye, “Diffusion posterior sampling for general noisy inverse problems,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2023.arxiv.org/abs/2209.14687
  4. [4]S. Mohseni-Sehdeh, W. Saad, K. Sakaguchi, and T. Yu, “Diffusion models for solving inverse problems via posterior sampling with precise guidance,” Trans. Mach. Learn. Res., 2026.
  5. [5]M. Arvinte and J. J. Tamir, “MIMO channel estimation using score-based generative models,” IEEE Trans. Wireless Commun., vol. 22, no. 6, pp. 3698–3713, Jun. 2023.arxiv.org/abs/2204.07122
  6. [6]B. Fesl, M. Baur, F. Strasser, M. Joham, and W. Utschick, “Diffusion-based generative prior for low-complexity MIMO channel estimation,” IEEE Wireless Commun. Lett., vol. 13, no. 12, pp. 3493–3497, Dec. 2024.arxiv.org/abs/2403.03545
  7. [7]N. Zilberstein, A. Swami, and S. Segarra, “Joint channel estimation and data detection in massive MIMO systems based on diffusion models,” arXiv preprint arXiv:2311.10311, 2023.
  8. [8]H. Jiang, X. Yuan, Y. Huang, and Q. Guo, “Blind MIMO semantic communication via parallel variational diffusion: A completely pilot-free approach,” arXiv preprint arXiv:2510.27043, 2025.
  9. [9]E. B. Yoon, K. Park, S. Kim, and S. Lim, “Score-based generative models with Lévy processes,” in Advances in Neural Information Processing Systems, vol. 36, 2023.DOI
  10. [10]D. Shariatian, U. Şimşekli, and A. O. Durmus, “Heavy-tailed diffusion with denoising Lévy probabilistic models,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2025.arxiv.org/abs/2407.18609
  11. [11]B. V. Gnedenko and A. N. Kolmogorov, Limit Distributions for Sums of Independent Random Variables. Reading, MA, USA: Addison-Wesley Publishing Company, 1968.
  12. [12]G. Samorodnitsky and M. S. Taqqu, Stable Non-Gaussian Random Processes. Chapman & Hall, 1994.
  13. [13]J. Fahs, “On performance analysis in heavy-tailed (in)dependent wireless channels,” IEEE Commun. Lett., vol. 30, pp. 2278–2282, 2026.DOI
  14. [14]Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multiscale structural similarity for image quality assessment,” in Proc. Asilomar Conf. Signals, Syst. Comput., 2003.
  15. [15]Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proc. IEEE, vol. 86, no. 11, pp. 2278–2324, 1998.
  16. [16]J. Fahs and I. Abou-Faycal, “On the linearity of conditional mean estimators of infinite-variance variables,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), 2026, pp. 1–6.
  17. [17]J. H. McCulloch, “Simple consistent estimators of stable distribution parameters,” Commun. Stat. Simul. Comput., vol. 15, no. 4, pp. 1109–1136, 1986.DOI
  18. [18]W. Tabbara, S. Sharafeddine, A. El-Hajj, J. Fahs, and I. Abou-Faycal, “Density functions and random number generators of α-stable distributions,” in Proc. Python in Science Conf. (SciPy’26), 2026. [Online]. Available: https://proceedings.scipy.org/articles/eykk7863arxiv.org/abs/2606.28530
  19. [19]A. Doucet, N. de Freitas, and N. Gordon, Eds., Sequential Monte Carlo Methods in Practice. New York, NY, USA: Springer, 2001.

Paper details

vlnc:W_gcnk3rh8
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents