Online Adaptive Channel Denoising

• Online adaptive channel denoising is a set of techniques that dynamically suppress noise in channel estimates by leveraging current observations and meta-learning for rapid adjustment.
• These methods employ few-shot meta-learning, reinforcement learning, and diffusion models to enhance channel estimation under low SNR, limited pilots, and dynamic interference.
• They provide practical benefits by reducing pilot overhead and computation while ensuring real-time, robust performance in non-stationary wireless environments.

Online adaptive channel denoising denotes a class of methods that suppress noise in channel estimates, equalized observations, or channel-dependent latent representations while adjusting to time-varying operating conditions during deployment. In the literature, the term spans few-shot adaptation for massive MIMO channel prediction, successive denoising for MIMO-OFDM, continuously retrained OFDM receivers, diffusion-based post-equalization, data-aided online label generation for denoising networks, and multi-snapshot restoration in OTFS. The unifying premise is that fixed offline denoisers are mismatched to dynamic channels, limited pilots, low-SNR operation, and unforeseen interference, whereas online or near-online adaptation can exploit current observations, recovered labels, or recent temporal context to maintain fidelity under distribution shift (Kim et al., 2022, Oh et al., 2021, Fischer et al., 2022, Wu et al., 2023, Ha et al., 13 Aug 2025, Gehlot et al., 28 May 2026).

1. Problem setting and system assumptions

In wireless communications, accurate channel knowledge is critical in massive multiple-input multiple-output, reliable communication via MIMO-OFDM requires accurate channel estimation at the receiver, and channels in communication systems are dynamic and change with time. Several papers therefore pose denoising not as a one-time preprocessing step but as an adaptive inference problem executed under changing channel statistics, limited training samples, and, in some cases, absent ground-truth CSI (Kim et al., 2022, Oh et al., 2021, Owfi et al., 3 Jan 2025).

Existing limitations recur across the corpus. One line of work states that denoising methods for channel estimation largely depend on either channel analysis in the time-domain with prior channel knowledge or supervised learning techniques which require large pre-labeled datasets for training. Another emphasizes that conventional channel autoencoders are designed based on the assumption of having access to a large number of pilot signals, while practical real-time systems must operate in few-shot learning scenarios. A third shows that online adaptation can be performed solely based on recovered labels from an outer forward error correction code without any additional piloting overhead (Oh et al., 2021, Owfi et al., 3 Jan 2025, Fischer et al., 2022).

The operational manifestations differ by waveform and receiver architecture.

Setting	Online signal used for adaptation	Reported challenge
Massive MIMO channel prediction	Few fine-tuning samples and noisy LS estimates	New environments and low SNR
MIMO-OFDM channel estimation	LS estimates and channel curvature across adjacent subcarriers	No channel statistics and no pre-labeled data
OFDM deep receivers	Recovered labels from the outer FEC code	Unforeseen conditions and no additional piloting overhead
OTFS channel estimation	Multiple DD snapshots within the geometric coherence time	Low pilot SNR and fractional delay / fractional Doppler

This suggests that “online” in this area is not restricted to sample-by-sample recursion. It also includes episodic few-shot updates, causal-window conditioning, post-equalization denoising, and asynchronous retraining triggered only when mismatch is detected (Uzlaner et al., 2024).

2. Few-shot meta-learning and denoising priors

A central approach is to learn an initialization that can be rapidly specialized to a new channel. In massive MIMO channel prediction, model-agnostic meta-learning is used to learn an initialization of deep network parameters such that measurements from a new environment or new UE can be handled with only a small number of channel samples. The meta-training stage alternates task-specific inner updates and an outer meta-update,

$$\boldsymbol{\Omega}_{\mathrm{Tr},t} \leftarrow \boldsymbol{\Omega}_{\mathrm{Tr},t} - \alpha \nabla_{\boldsymbol{\Omega}_{\mathrm{Tr},t}} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Sup}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$

$$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$

followed by meta-adaptation on a new environment with a few adaptation samples. The same work augments this with deep image prior denoising on stacked LS channel estimates,

$$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$

using only the noisy measurements and no pre-training. Numerical results report that the MAML-based predictor improves prediction accuracy with only a few fine-tuning samples, that even with as few as 20 adaptation samples it significantly outperforms conventional MLP predictors, and that “MAML-DIP” achieves a substantial NMSE improvement, with approximately $4$ dB gain at SNR $= 0$ dB over MAML alone (Kim et al., 2022).

The same few-shot logic appears in end-to-end physical layer optimization. The Online Meta Learning channel AE framework adapts a channel autoencoder to dynamic channels using only a few pilots, with MAML inner and outer updates

$$\theta'_i = \theta - \alpha \nabla_{\theta} \mathcal{L}_{T_i}(f_\theta), \qquad \theta \leftarrow \theta - \beta \nabla_{\theta} \sum_{T_i \sim p(T)} \mathcal{L}_{T_i}(f_{\theta'_i}),$$

and an online task buffer for continual refinement. Under autoregressive Rayleigh fading, the reported findings state that OML-CAE outperformed all baselines under both low and moderate SNR conditions, and that conventional CAE required approximately 3.8 times more pilots to reach the same SER as OML-CAE (Owfi et al., 3 Jan 2025).

A closely related development replaces unavailable ground-truth CFRs with data-aided estimates generated online from adjacent detected data symbols used as virtual reference signals. For each target sub-CFR, the receiver forms a time-frequency neighborhood and computes

$$\tilde{\mathbf{H}}^{(i,j)} = {\mathbf{Y}_{P,Q}[n_i,k_j]\, \hat{\mathbf{X}}_{P,Q}^{\sf H}[n_i,k_j]} \big(\hat{\mathbf{X}}_{P,Q}[n_i,k_j] \hat{\mathbf{X}}_{P,Q}^{\sf H}[n_i,k_j]\big)^{-1},$$

which serves as a surrogate label for online transfer learning or online meta learning. The paper also derives an optimization over neighborhood size $(P,Q)$ to minimize the resulting MSE, and reports that meta learning is especially effective in out-of-distribution settings when only a few online samples or gradient steps are available (Ha et al., 13 Aug 2025).

A common misconception is that adaptive denoising necessarily requires either true CSI labels or large pilot overhead. The literature above gives counterexamples: few-shot adaptation from a meta-learned initialization, unsupervised DIP denoising from noisy LS estimates alone, and standard-compatible online label generation through virtual pilots (Kim et al., 2022, Ha et al., 13 Aug 2025).

3. Reinforcement learning and sequential decision formulations

Another major line formulates denoising as sequential control. In MIMO-OFDM, successive denoising begins from the observation that the true channel response across adjacent subcarriers is smooth, whereas additive noise induces abrupt spikes or outliers in LS channel estimates. The method computes the channel curvature

$$\widehat{C}_{qp}^{(k)} = \widehat{H}_{qp}^{(k+1)} - 2\widehat{H}_{qp}^{(k)} + \widehat{H}_{qp}^{(k-1)},$$

uses an analytically derived curvature magnitude threshold $\widetilde{C}$ to identify unreliable estimates, and casts the denoising order and location as a Markov decision process. When a subcarrier is denoised, its value is updated by a geometric projection rule,

$$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$0

with $$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$1. Q-learning then updates the state-action table via

$$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$2

The reported performance is a significant reduction in MSE compared to LS, up to $$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$3 dB gain, with performance closely approaching ideal LMMSE with perfect knowledge of channel statistics (Oh et al., 2021).

Policy-gradient RL has also been used for adaptive filtering in dynamic, non-stationary environments. A PPO agent controls FIR filter parameters in real time, with reward

$$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$4

balancing SNR improvement, MSE reduction, and residual smoothness. The policy is trained with the clipped PPO objective

$$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$5

Experiments on synthetic signals with Gaussian, Laplacian, impulse, pink, brown, and uniform noise report higher output SNR than LMS, RLS, Wiener, and Kalman baselines, robust generalization beyond the training distribution, and real-time inference taking less than $$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$6 ms per time step (Bereketoglu, 29 May 2025).

These RL-based systems make the adaptive dimension explicit: the denoiser is not only estimating a clean channel or signal, but also deciding where to denoise, in what order, or how aggressively to alter filter parameters. A plausible implication is that this control-theoretic perspective becomes increasingly relevant when denoising actions themselves consume limited latency or compute budgets.

4. Diffusion and score-based channel denoisers

Diffusion models introduce a different formulation: denoising is posed as reverse stochastic generation conditioned on channel information or temporal history. Channel Denoising Diffusion Models are inserted as a new physical layer module after channel equalization, learning the distribution of the channel input signal and then using that learned knowledge to remove residual channel noise. After MMSE equalization and normalization, the forward process is aligned with the channel through

$$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$7

and training minimizes

$$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$8

At inference, the reverse process starts from the actual equalized signal and denoises it over $$\boldsymbol{\Omega} \leftarrow \boldsymbol{\Omega} - \beta \nabla_{\boldsymbol{\Omega}} \sum_{t=1}^{V} \mathrm{Loss}_{\mathbb{D}_{\mathrm{Que}(t)}}(\boldsymbol{\Omega}_{\mathrm{Tr},t}),$$9 steps. Reported gains include up to $$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$0 dB MSE improvement over MMSE at high SNR $$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$1 dB, up to $$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$2 dB at $$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$3 dB, and a $$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$4 dB PSNR gain for joint CDDM and JSCC over JSCC at $$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$5 dB in Rayleigh fading (Wu et al., 2023).

For non-stationary urban microcell channels, conditional prior diffusion learns a history-conditioned score and uses a temporal encoder with cross-time attention, temporal self-conditioning, and feature-wise linear modulation. The forward diffusion follows

$$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$6

with an SNR-matched initialization

$$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$7

and a shortened, geometrically spaced reverse schedule. On a 3GPP UMi benchmark with 8 users, 52 subcarriers, 3.5 GHz, and 80 MHz bandwidth, the reported result is $$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$8 dB mean NMSE, $$\min_{\boldsymbol{\Phi}} \left\| \boldsymbol{\mathcal{H}}^{\mathrm{LS}} - g_{\boldsymbol{\Phi}}(\mathbf{Z}_1) \right\|^2,$$9 dB better than the best baseline, and within $4$0-$4$1 dB of the oracle at high SNR (Mohsin et al., 18 Sep 2025).

Diffusion has also been coupled to semantic communications through joint source-channel noise adding with adaptive denoising. There, source and channel noise are intentionally incorporated into the forward diffusion, and the received image is partitioned into patches whose denoising steps are allocated according to a vision transformer self-attention map:

$4$2

At Rayleigh channel SNR $4$3 dB on ImageNet-256, the reported PSNR is $4$4 dB for JSCNA-AD, compared with $4$5 for JSCNA without adaptive denoising, $4$6 for WITT, $4$7 for JPEG2000+LDPC, and $4$8 for DeepJSCC, with inference time $4$9 ms (Liang et al., 11 May 2025).

A recurring misconception is that diffusion-based denoising is inherently too costly for online use. The evidence here is narrower: some systems reduce computation by SNR-matched initialization and geometric truncation, and some allocate denoising steps only where semantics indicate importance (Mohsin et al., 18 Sep 2025, Liang et al., 11 May 2025).

5. Self-supervision, recovered labels, and continual online adaptation

Online adaptive channel denoising does not always rely on explicit denoising networks trained against clean labels. In adaptive neural network-based OFDM receivers, the retraining target is recovered during operation from the outer decoder. The receiver outputs soft LLRs, the BP LDPC decoder produces corrected codeword estimates, and those recovered labels are then used to retrain the receiver using the binary cross-entropy loss

$= 0$0

The procedure uses no additional pilot overhead, filters labels by decoder quality, and retrains periodically on collected batches. The reported gains include approximately $= 0$1 dB in a static corner case, substantial recovery in an out-of-spec channel with 16 taps and $= 0$2 km/h where LMMSE IEDD fails dramatically, and approximately $= 0$3 dB improvement over the universal RNN under localized interference on four edge subcarriers (Fischer et al., 2022).

A different self-supervised strategy learns CSI denoising and feedback without supervision. An autoencoder is trained at the base station using only noisy uplink CSI, minimizing the empirical MSE between input and reconstruction,

$= 0$4

and then half of the autoencoder is offloaded to the mobile terminal to encode noisy downlink CSI for feedback. The encoder compresses a $= 0$5 real-valued input to a 256-dimensional codeword, a compression factor of $= 0$6. Simulations with QuaDRiGa 2.2 report strong uplink-to-downlink generalization, substantial NMSE reduction over IDFT and CsiNet baselines, and multi-user rate close to that with perfect CSI (Rizzello et al., 2021).

These methods establish that online adaptation need not be identical to gradient-based denoising of channel estimates. It may instead arise from decoder-corrected pseudo-labels, uplink-only unsupervised pretraining, or virtual pilots generated from detected data (Fischer et al., 2022, Rizzello et al., 2021, Ha et al., 13 Aug 2025).

6. Structural invariance, compute-aware adaptation, and selective retraining

Several works exploit structural regularity in the channel itself to make online denoising lighter and more reliable. In OTFS, CSI recovery is formulated as an image restoration problem in the delay-Doppler domain. Because the DD-domain channel remains approximately invariant over a geometric coherence time, $= 0$7 consecutive OTFS frames provide independent noisy snapshots of the same effective channel:

$= 0$8

Each snapshot is denoised by a lightweight NAFNet,

$= 0$9

and then averaged,

$$\theta'_i = \theta - \alpha \nabla_{\theta} \mathcal{L}_{T_i}(f_\theta), \qquad \theta \leftarrow \theta - \beta \nabla_{\theta} \sum_{T_i \sim p(T)} \mathcal{L}_{T_i}(f_{\theta'_i}),$$0

The reported NAFNet-based denoiser uses only $$\theta'_i = \theta - \alpha \nabla_{\theta} \mathcal{L}_{T_i}(f_\theta), \qquad \theta \leftarrow \theta - \beta \nabla_{\theta} \sum_{T_i \sim p(T)} \mathcal{L}_{T_i}(f_{\theta'_i}),$$1 parameters and $$\theta'_i = \theta - \alpha \nabla_{\theta} \mathcal{L}_{T_i}(f_\theta), \qquad \theta \leftarrow \theta - \beta \nabla_{\theta} \sum_{T_i \sim p(T)} \mathcal{L}_{T_i}(f_{\theta'_i}),$$2 FLOPs per snapshot, while improving NMSE and BER under low pilot SNR and fractional delay / fractional Doppler effects (Gehlot et al., 28 May 2026).

A complementary line addresses not how to denoise, but when and which components to adapt. Asynchronous online adaptation via modular drift detection studies deep receivers composed of sub-modules and attaches a drift detector to each one. Retraining is then triggered only when necessary and only for the sub-modules that exhibit drift. Soft-output detectors based on empirical posteriors and a Hotelling-type statistic are proposed, and the expected per-block modular cost is written as

$$\theta'_i = \theta - \alpha \nabla_{\theta} \mathcal{L}_{T_i}(f_\theta), \qquad \theta \leftarrow \theta - \beta \nabla_{\theta} \sum_{T_i \sim p(T)} \mathcal{L}_{T_i}(f_{\theta'_i}),$$3

Numerical studies report that asynchronous adaptation can match the BER of always retrain with $$\theta'_i = \theta - \alpha \nabla_{\theta} \mathcal{L}_{T_i}(f_\theta), \qquad \theta \leftarrow \theta - \beta \nabla_{\theta} \sum_{T_i \sim p(T)} \mathcal{L}_{T_i}(f_{\theta'_i}),$$4-$$\theta'_i = \theta - \alpha \nabla_{\theta} \mathcal{L}_{T_i}(f_\theta), \qquad \theta \leftarrow \theta - \beta \nabla_{\theta} \sum_{T_i \sim p(T)} \mathcal{L}_{T_i}(f_{\theta'_i}),$$5 fewer retraining events, and that modular adaptation can yield up to $$\theta'_i = \theta - \alpha \nabla_{\theta} \mathcal{L}_{T_i}(f_\theta), \qquad \theta \leftarrow \theta - \beta \nabla_{\theta} \sum_{T_i \sim p(T)} \mathcal{L}_{T_i}(f_{\theta'_i}),$$6 reduction in the number of retraining operations, with little compromise on performance (Uzlaner et al., 2024).

Two broader clarifications follow from these results. First, online adaptive channel denoising does not imply continuous retraining of an entire receiver; selective or asynchronous updates may be sufficient (Uzlaner et al., 2024). Second, online adaptation is not synonymous with abandoning model structure. Curvature thresholds, DD-domain invariance, channel-aligned diffusion schedules, and causal temporal encoders all encode strong priors; what changes online is the operating point, the conditional context, or the subset of parameters engaged at inference (Oh et al., 2021, Wu et al., 2023, Mohsin et al., 18 Sep 2025, Gehlot et al., 28 May 2026).