WorksEngineering
Temporal Diffusion Models for Joint Signal-Channel Estimation under Gaussian and -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 -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 -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 -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 -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 -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 -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, -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 -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
for some scalar , where and are independent copies of and is the stability index. The case recovers the Gaussian law. In this work, we say RV is Symmetric -Stable (SS) and write , , whenever
where and are its location and scale, respectively. The index controls the tail asymptotics: a smaller produces heavier power-law tails of order . Moreover, , whenever , so the variance is infinite for .
In addition, a random vector is said to be “circular” SS with law SS if and only if
It admits the Gaussian scale-mixture representation
where is independent of —a positive -stable variable with scale equal to . Consequently, given that is equal to
Reconstruction Metrics
For we use the Normalized Mean Squared-Error
for Gaussian data, and
& , (4)
for -stable ones. Note that 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 transmit antennas and receive antennas, define
where is the -th transmitted symbol. We adopt a block-fading model where the channel matrix is assumed constant over these symbols. For each , the received symbol is
Stacking the received symbols we write
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 -stable law. For the Gaussian channel, the “vectorized” channel matrix:
where .
For the symmetric -stable channel,
where , preserving the stationary marginal . We restrict to guarantee a finite mean, in accordance with the LIM model [9].
The additive noise in (4) are independent and chosen from the corresponding family: Gaussian noise for the Gaussian channel and SS 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 . The channel coefficients are processed jointly as a vector, giving the vectorized observation model
The objective is to estimate the temporally-varying channel vector 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 . For , no additional pilots are transmitted, and both the current channel and transmitted image are unknown. Let be an MNIST image [15]. A fixed encoder maps it to
where is the symbol vector transmitted during channel use . As previously stated, the channel is assumed constant over these channel uses.
The receiver must jointly estimate and the information-bearing encoded signal , or equivalently the underlying image . 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 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 -dimensional . In DDPM, predicts the Gaussian perturbation, whereas in the continuous formulation, predicts the score.
Let . The canonical forward marginals are given by for DDPM and for the score-based model. The same networks are adapted, without retraining, to and the perturbed marginals are
At every reverse step, let
The canonical inputs are
and the network outputs are transported as
Note that the DDPM transport is stated for the MSE-optimal noise predictor. It is established using the joint Gaussianity of and the fact that the MSE-optimal noise predictor is . Applying the Gaussian conditional-mean formula together with 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 (see Appendix B).
Canonical Alpha-Stable Adaptation
For each -stable formulation, one diffusion model is trained on . In DLPM, predicts the stable perturbation, whereas in LIM, predicts the fractional score defined in [9, 16]. We use the logarithmic error defined in (4) in the noise-prediction models, with in the experiments.
The trained models are adapted without retraining to
For DLPM, whose canonical forward marginal is , where , the experiments were conducted with , and
For LIM, whose canonical forward marginal is ,
For DLPM with , the transport follows directly by substituting 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 , hence the factor (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
where each unknown has a separately trained diffusion prior and reverse state . Matrix-valued observations are understood through vectorization.
For Gaussian noise , the reverse models provide clean estimates , and the likelihood gradient for the -th unknown is
A weighted version of is added to the corresponding reverse update. Thus, separately trained priors are coupled only through the shared likelihood.
When the Gaussian observation is linear,
the posterior is Gaussian with
This permits exact Gaussian conditional sampling in the channel-only setting. For noise , we use the Gaussian scale-mixture representation in (1). Conditioned on the positive-stable variable ,
Using samples , the marginal likelihood is approximated by
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,
where , .
If is treated as exact, this reduces to
For the channel, uncertainty propagation is approximated by
where ,
This is an approximation because the distribution of the previous estimation error is generally not -stable. We therefore match a zero-location 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 , yielding
Experiments and Results
Channel-Only Estimation
Using pilot signals as in (7), we jointly process three channel coefficients, with .
Gaussian Channel
As an example, we consider channels’ coefficients with correlated stationary distribution
and . The independent method uses the stationary prior, whereas the temporal methods use the predictive prior in (14). At , , and the posterior covariance is propagated recursively for .
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 diffusion samples , the guided estimate is selected as
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 SS : stability index , location , and unit stationary scale. The additive -stable noise is SS . These -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,
For equal signal and noise stability indices and scales, the exact independent absolute-error estimator is
For the temporal reference, the MAE-optimal estimator is the componentwise median of , approximated numerically using samples from the known temporal model weighted by the observation likelihood [19]. Performance is evaluated using defined in (3).
Results: We use sequences. Figure 1 shows that the gain of temporal over independent reconstruction increases with , measured by a reduction in NMSE for Gaussian channels with DDPM and in MAE for -stable channels with DLPM. For the Gaussian case, this behavior is observed under both guided and conditional reconstruction. At , the temporal and independent methods coincide, as expected. Similar trends are observed for the continuous score-based and LIM models in Figure 6.

Figure 1. Reconstruction loss versus for independent and temporal channel reconstruction. Top: Gaussian channels using DDPM. Bottom: -stable channels using DPLM with .

Figure 6. Reconstruction loss versus for independent and temporal channel reconstruction. Top: Gaussian channels using the score-based model. Bottom: -stable channels using LIM with . 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 symbols in (5). The channel is with an ”initial” matrix that is known from the training stage; for , 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 and , with .
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 and the entries of are independent 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 , we select
Alpha-Stable Channel
The channel models are trained on the 16-dimensional , for and the additive -stable noise is . Since -stable noise has infinite variance, its scale is selected according to
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 -stable measurement likelihood,
Independent and temporal reconstruction results: Each sequence contains MNIST images. Figure 2 shows that temporal inference improves image MS-SSIM, channel NMSE (Gaussian) and (-stable) over independent inference as increases. At , 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 increases for both Gaussian and -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.

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: -stable channels with using DLPM.

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

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: -stable channels with using LIM. This corresponds to the joint channel–image MIMO experiment in § 5-B.

Figure 10. Sent images and the corresponding independent and temporal joint MIMO reconstructions. Top: Gaussian channel using the score-based model at SNR dB with . Bottom: -stable channel using LIM with at SNR dB with . These results correspond to the joint channel–image MIMO experiment in § 5-B.
Model mismatch: Matched -stable channel priors are compared with Gaussian channel priors trained and evaluated on the same -stable channel data with . 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.

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

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

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

Figure 8. Sent image on -stable channel with joint MIMO reconstructions obtained using matched LIM and mismatched Gaussian score-based model for and SNR 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 -stable models.
The channel-model comparison separately shows the importance of matching the diffusion noise family with the underlying data distribution. For Gaussian and -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 -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
whose marginal, with , is
The exact reverse posterior conditioned on is
where
The learned reverse transition replaces the unknown perturbation by such that
with
The corresponding clean estimate is
Starting from the terminal marginal, reverse sampling uses
where and .
Canonical Gaussian Adaptation
We train a single DDPM on the canonical distribution and adapt its noise predictor to without retraining. From (18),
Let be its Cholesky factorization. The transformed input
satisfies .
Lemma 1. Let be the MSE-optimal noise predictor trained on the canonical distribution . Then, the MSE-optimal noise predictor for is obtained without retraining as
Proof. Under the squared-error objective, the optimal predictor is the conditional expectation . Since and are independent Gaussian vectors,
By (18), is a linear transformation of and is therefore jointly Gaussian. Its cross-covariance is
where independence makes the first term zero.
For jointly Gaussian vectors, the conditional expectation is
Using , , , and , we obtain
For the canonical model, and . Hence, its MSE-optimal predictor satisfies
Evaluating it at gives
which establishes (8).
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 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 and . For the independent conditional method, 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 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 , , and . 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 , , , and . 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 , the predicted clean state is
where denotes the adapted noise prediction. The observation likelihood gradients are then evaluated as
The corresponding guided reverse update, performed at every time step, is
where is obtained from (20), is the guidance strength, and .
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 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
where is a standard Wiener process. We consider the linear Gaussian perturbation kernel associated with this SDE,
Let denote the marginal density of . The corresponding reverse-time SDE is
which is solved from to . The network approximates the score and is substituted into (25) during generation.
For the perturbation kernel in (24), the clean estimate is
Starting from the terminal marginal, the reverse SDE is evaluated on a decreasing numerical grid of intervals, .
Canonical Gaussian Adaptation
We train a single score-based model on the canonical distribution and adapt its score to without retraining. From (23),
Let be its Cholesky factorization. The transformed state
has the same canonical marginal as .
Lemma 2. Let be the score trained on the canonical state . Then, the score for is obtained without retraining as
Proof. Let and denote the densities of and . Since and , the change-of-variables formula gives
Hence,
Because , applying the chain rule therefore gives
For the canonical model , the score function is . Evaluating it at gives
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 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 and . For the independent conditional method, in (12) and (13) is the stationary channel prior in (5); for the temporal conditional method, it is the predictive prior in (14).
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 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 , , and . 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 , , , and . 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 , the predicted clean state is
where denotes the adapted score estimate. The observation likelihood gradients are then evaluated as
The corresponding guided score, computed at every reverse time , is
Algorithm 4 Score-Based Likelihood-Guided Reconstruction
Require: , Gaussian likelihood , score-based priors, , SDE coefficients, reverse grid,
for do
Draw from their terminal marginals
for do
for do
Compute adapted prior score
Compute clean estimate using (25)
end for
for do
Compute likelihood gradient using (26)
Form guided score using (27)
Advance to using (28)
end for
end for
end for
return
where is the guidance strength. The guided reverse SDE is
which is integrated backward from 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 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 and let be a noise schedule. Define
A DLPM defines the forward process
Equivalently,
Using the Gaussian scale-mixture representation in (1), each perturbation is written as
where are independent positive -stable variables with skewness , scale , and location . Conditioned on , define
Then
The exact reverse posterior conditioned on and is
where
The learned reverse transition replaces the unknown marginal perturbation by . Let
The reverse transition is
with
The corresponding clean estimate is
At the beginning of each reverse trajectory, one independent path is drawn and (30) is evaluated. Starting from the terminal marginal, reverse sampling uses
where .
Canonical Alpha-Stable Adaptation
We train one DLPM on the canonical distribution . The DLPM experiments use unit scale ; accordingly, the following adaptation changes only the location. For the unit-scale target distribution
the forward marginal in (29) gives
Lemma 3. Let be the stable-noise predictor trained on the canonical distribution. The canonical input and transported prediction are
Proof. Substituting into (29) gives
Therefore, has the same distribution as the canonical state . Evaluating the canonical predictor at proves (31).
Likelihood-Guided Reconstruction
We consider the general observation model (10), where each unknown has a separately trained diffusion prior. This formulation includes both channel-only estimation in (7) and joint MIMO reconstruction in (4).
For -stable channel-only estimation, we set , , and . The independent method uses the stationary -stable channel prior in (6), whereas the temporal method uses the predictive prior in (16).
For blind joint channel–image MIMO reconstruction, we set , , , and . The channel and image have separately trained diffusion priors, coupled through the MIMO likelihood. The independent method uses the stationary -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),
Using the Gaussian scale-mixture representation in (1) gives
where is positive -stable. Using independent samples , the likelihood is approximated by
Samples are used only to approximate the likelihood of measurement and are independent of the DLPM trajectory variables .
For each unknown DLPM and at each time step , the predicted clean state is
The observation likelihood gradients are then evaluated as
The corresponding guided reverse update, performed at every time step, is
where is the guidance weight and , with .
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 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: , approximated likelihood , diffusion priors, , diffusion schedules, ,
Draw the measurement-likelihood samples
for do
Draw an independent for each DLPM prior
Compute the corresponding using (30)
Draw from their terminal marginals
for do
for do
Compute adapted prediction
Compute clean estimate using (32) for DLPM
end for
for do
Compute likelihood gradient using (33)
Advance to using the guided reverse update of its prior
end for
end for
end for
if channel-only reconstruction then
return
else
return
end if
LIM
Forward and Reverse Processes
A continuous LIM defines the forward SDE
where is a symmetric -stable Lévy process. We consider the linear -stable perturbation kernel associated with this SDE,
Let denote the marginal density of . Its fractional score is
The corresponding reverse-time dynamics used by LIM are
which are solved from to . The network approximates the fractional score and is substituted into (37) during generation.
For the perturbation kernel in (35), the clean estimate is
Starting from the terminal marginal, the reverse dynamics are evaluated on a decreasing numerical grid of intervals, .
Canonical Alpha-Stable Adaptation
We train one LIM on the canonical distribution and adapt its fractional score to without retraining. From (35),
Thus, has the same canonical marginal as .
Lemma 4. Let be the fractional score trained on the canonical state . The fractional score for is obtained without retraining as
Proof. Let and denote the densities of and . From the affine transformation, . The density transformation is
Using ,
By the fractional-score definition in (36),
Therefore,
The Fourier scaling-and-translation property also gives
Hence,
Taking the inverse Fourier transform gives
Dividing by gives
and hence
establishes (38).
Likelihood-Guided Reconstruction
We consider the general observation model (11), where each unknown has a separately trained diffusion prior. This formulation includes both channel-only estimation in (8) and joint MIMO reconstruction in (5).
For -stable channel-only estimation, we set , , and . The independent method uses the stationary -stable channel prior in (7), whereas the temporal method uses the predictive prior in (17).
For blind joint channel–image MIMO reconstruction, we set , , , and . The channel and image have separately trained diffusion priors, coupled through the MIMO likelihood. The independent method uses the stationary -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),
Using the Gaussian scale-mixture representation in (1) gives
where is positive -stable. Using independent samples , the likelihood is approximated by
For each LIM unknown and at every reverse time , the predicted clean state is
The observation likelihood gradients are then evaluated as
At every reverse time, the prior probability-flow step from to is first evaluated using (36). The likelihood correction is then applied as
where is the guidance strength.
Algorithm 6 LIM Likelihood-Guided Reconstruction
Require: , approximated likelihood , diffusion priors, , reverse grid, ,
Draw the measurement-likelihood samples
for do
Draw from their terminal marginals
for do
for do
Compute adapted fractional score and perturbation estimate
Compute clean estimate using (39)
end for
for do
Compute likelihood gradient using (40)
Advance to using (41)
end for
end for
end for
if channel-only reconstruction then
return
else
return
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 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]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]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]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]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]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]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]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]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]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]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]B. V. Gnedenko and A. N. Kolmogorov, Limit Distributions for Sums of Independent Random Variables. Reading, MA, USA: Addison-Wesley Publishing Company, 1968.
- [12]G. Samorodnitsky and M. S. Taqqu, Stable Non-Gaussian Random Processes. Chapman & Hall, 1994.
- [13]J. Fahs, “On performance analysis in heavy-tailed (in)dependent wireless channels,” IEEE Commun. Lett., vol. 30, pp. 2278–2282, 2026.DOI
- [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]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]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]J. H. McCulloch, “Simple consistent estimators of stable distribution parameters,” Commun. Stat. Simul. Comput., vol. 15, no. 4, pp. 1109–1136, 1986.DOI
- [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]A. Doucet, N. de Freitas, and N. Gordon, Eds., Sequential Monte Carlo Methods in Practice. New York, NY, USA: Springer, 2001.