Innovation-Gated State Space Models for Low-Pilot Channel Tracking

A base station has to know the channel to every user: how each antenna-to-antenna path changes the amplitude and phase of each frequency. It measures this with known reference signals called pilots, and those pilots consume bandwidth that could carry data. Learned "channel deduction" cuts this cost. It combines the user's recent channel history with a crude estimate taken from a handful of current pilots, and in doing so avoids both the full pilot cost of estimation and the drift of pure forecasting.

This note proposes building channel deduction on a selective state space model. Such a model carries a compact memory forward one time slot at a time, at constant cost per slot, and it can decide from its current input how much of that memory to keep. The specific change we propose is about what drives that decision. The model's forget-and-overwrite rate should depend on the innovation: the mismatch between what its memory predicts at the pilot positions and what the pilots actually show. When the prediction matches, the model coasts on its history. When the user turns or accelerates and the mismatch grows, the model discards stale history and rebuilds its estimate from the pilots. This behaves like a learned version of the gain in a Kalman filter, which sets how strongly a new measurement corrects the running estimate. It also gives a natural trigger for asking for pilots only when they are needed.

Core Claim

A selective state space model, often called by its best-known instance, Mamba, updates a hidden state as $h_t = \exp(\Delta_t A)\,h_{t-1} + \Delta_t B_t x_t$. Here $h_t$ is the memory at slot $t$, $x_t$ is the slot's input, $A$ is a fixed stable decay matrix, $B_t$ is an input-dependent write matrix, and $\Delta_t > 0$ is an input-dependent step size. A large $\Delta_t$ shrinks the old state and writes the new input strongly; a small $\Delta_t$ preserves memory. We set

$$\Delta_t = \mathrm{softplus}\big(W[\,x_t;\; m_t \odot (y_t - \hat{y}_{t|t-1})\,]\big),$$

where $y_t$ is the sparse pilot observation at slot $t$, $\hat{y}_{t|t-1}$ is the model's prediction of that observation from $h_{t-1}$, $m_t$ is a mask that is zero on slots or positions with no pilot, and $W$ is learned.

The claim has two parts. First, making the step size depend on the pilot innovation, rather than only on the raw input, gives faster recovery after abrupt motion changes than an ungated model that sees the same inputs. Second, it keeps tracking stable over thousands of slots at low pilot density. Because one recurrence handles slots with pilots and slots without them, the same model also supports adaptive pilot scheduling: a small uncertainty head requests pilots when predicted error crosses a threshold.

Positioning Against Prior Work

The supplying and motivating work is Channel Deduction. It fuses outdated channel samples with a coarse current estimate using a complex-valued mixer front-end followed by an LSTM (RCDNet) or self-attention (ACDNet). It reports up to 88.9% pilot reduction and a 13 dB NMSE gain over traditional estimation at 2×2 pilots, and it names lower complexity for base-station hardware as future work. We keep its problem formulation and front-end but change the fusion step. Its networks concatenate past and present information and learn to combine them; they have no explicit innovation signal and no pilot-free slots.

MambaCSP shows that a selective state space model beats LLM-based channel predictors on QuaDRiGa urban channels, with 9% lower NMSE in time-division duplexing, 12% lower in frequency-division duplexing, 3.0× throughput and 2.6× less memory. It is a pure predictor, though: it uses no current pilots and lists joint pilot design as open. We take it as evidence that state space backbones suit channel time series, and as the "prediction-only" baseline. We do not reuse its frozen language-model backbone.

The mechanism differs from SSM-based in-context equalization, which prompts Mamba with pilots and received symbols to decode data in block-fading channels. That work produces no explicit channel matrix and does no tracking over time.

Three papers constrain the design. Is Mamba Capable of In-Context Learning? reports a U-shaped error curve when Mamba runs past its training context. That is a direct threat to continuous tracking, and we treat it as a hypothesis to test: pilot corrections should keep the state anchored and remove the curve. Reverse Ordering Techniques shows that recency-anchored indexing gives length-invariant attention predictors. Our recurrence is recency-anchored by construction, so their Transformer-RPE becomes a natural comparator. Normalization-Free Knowledge Distillation shows that z-score normalization strips path-loss information and can hurt channel models. Accordingly, we normalize only by a per-user running power estimate that is passed to the model as a side input, and we compare this against standard layer normalization.

Research Program

The model has four parts. First, a complex-to-real token front-end in the style of the channel-deduction mixer maps each slot's sparse pilot observation and mask to a feature vector. Second, a stack of selective state space layers runs over slots. Only the first layer receives the innovation term; deeper layers use standard selectivity. Third, a readout head maps the state to the full antenna-by-subcarrier channel. The same head, evaluated at pilot positions, gives $\hat{y}_{t|t-1}$. Fourth, an uncertainty head predicts the next-slot NMSE.

Training unrolls the scan in parallel over windows of 64 slots. Pilot density is randomized, including fully pilot-free stretches, so the model learns to both coast and correct. The adaptive-pilot variant adds a penalty on pilot count, with a thresholded uncertainty head deciding requests at test time.

Beyond accuracy, we run a mechanism analysis. We measure the correlation between the learned $\Delta_t$ and the true channel change between slots, and we look at whether $\Delta_t$ spikes at injected trajectory breaks. If $\Delta_t$ does not respond to the innovation, the claimed mechanism is not what drives any gains.

Baselines

The naive reference is least-squares estimation at the sparse pilots followed by interpolation across antennas and subcarriers, run at each slot with no history. It sets the pilot-only floor, and its gap to the learned methods shows how much history buys.

The strongest classical competitor is a Kalman filter on a multivariate autoregressive channel model. Its parameters are fitted on the training split, and its measurement update uses the same sparse pilots. This baseline embodies exactly the coasting-versus-correction logic we claim to learn. Beating it at matched pilot density, especially after trajectory breaks where its fixed statistics are wrong, is the minimum bar.

The strongest learned competitor is CDNet in both its ACDNet and RCDNet forms from Channel Deduction. We re-implement it on identical data, pilot patterns and splits, with the same hyperparameter-search budget in GPU-hours. The file reports its gains relative to traditional estimation (13 dB at 2×2 pilots, up to 88.9% pilot reduction) but gives no absolute NMSE: no reported absolute result was found on Lacuna for CDNet on DeepMIMO O1. We therefore rerun it rather than quote a target. Beating it means equal or lower NMSE at every pilot density, at lower per-slot FLOPs and latency, and a smaller error spike after trajectory breaks.

The prediction-only comparator is a MambaCSP-style predictor without current pilots. It quantifies how much pilots add.

The key ablation is an ungated Mamba deduction model: the same architecture and the same inputs, with pilots concatenated into $x_t$, but no innovation term in $\Delta_t$. A second ablation feeds the innovation into a GRU's gates instead. That one separates "innovation gating helps" from "state space models help."

Why not just…?

Why not just use a Kalman filter, since deduction is filtering? It is the right conceptual baseline, and in stationary, slowly varying channels it may come close. Its weakness is that it needs known Doppler and second-order statistics. Reverse Ordering Techniques notes these are unavailable in fast-changing channels, and Pruning the Pilots finds learned estimators beat traditional ones even when the traditional ones are given perfect statistics. We report results separately for stationary segments and for segments after trajectory breaks. If the Kalman filter matches us on both, the learned model is not justified.

Why not just keep CDNet with a short attention window, which is cheap anyway? This is a real threat. If deduction needs only a few slots of history, quadratic cost is irrelevant and our efficiency argument evaporates. The history-length sweep settles it. If ACDNet saturates at about 8 slots and matches our accuracy, the contribution reduces to innovation gating and pilot scheduling, and we will say so. We expect longer history to matter for low pilot density and long pilot-free stretches, but that is an empirical question.

Why not just put the innovation into a GRU? It is also linear-time, and gating on innovation is not specific to state space models. The GRU ablation tests this directly. If it matches accuracy, the remaining advantage of the state space version is parallel training over long windows and throughput, as reported in MambaCSP. We would report that advantage as an engineering benefit, not an accuracy one.

Proposed Experiments

We evaluate on two datasets. The first is DeepMIMO O1, using the Channel Deduction configuration for direct comparability. The second is QuaDRiGa 3GPP urban NLOS at 10 to 100 km/h, following MambaCSP. On QuaDRiGa we add injected trajectory breaks (turns, stops, accelerations) to create non-stationarity. Metrics are NMSE in dB against pilot overhead, NMSE against slot index, spectral efficiency under matched-filter beamforming, and per-slot FLOPs and latency.

Four checks are decisive. The first is post-break recovery: the number of slots to return within 1 dB of steady-state NMSE, compared between gated and ungated models. The second is long-horizon stability: train on 64-slot windows, run for 5000 slots, and look for any U-shaped drift. The third is the pilot-budget curve: adaptive scheduling against fixed periodic pilots at equal average overhead. The fourth is the correlation between $\Delta_t$ and channel change described above.

The proposal fails if the ungated model or the innovation-gated GRU matches the gated state space model on recovery and stability. It also fails if a Kalman filter matches it after trajectory breaks.