Many wireless links do not need to decide about the channel instantly. Block transmissions, retransmission scheduling and buffered telemetry all allow a short look-ahead before the receiver commits. That look-ahead could substantially improve estimates of when a link was "good" or "bad". A finite-state Markov channel (FSMC) is a model that sorts the signal-to-noise ratio (SNR) into a few discrete levels and lets the link hop between them according to a Markov chain. Over such a channel, using future observations is an old idea: the forward-backward algorithm does it exactly. Machine learning offers a different route: bidirectional networks, which read a sequence forward and backward and combine the two readings. These networks are more flexible, but their backward half is either an independent second network or a weight-shared copy of the forward one. Neither choice is correct when the physics has a direction. A train entering a tunnel does not fade like a train leaving it, and a satellite approaching a ground station does not behave like one receding.
We propose a bidirectional channel smoother whose backward pass contains no parameters of its own. A small network reads side information such as position, speed or elevation and outputs a time-varying transition matrix and an emission model. Exact forward-backward recursions are then run on that output, so the backward message is fully determined by the forward physics. Both recursions are linear in their message vectors, so they run as parallel scans with the same O(T) cost as modern bidirectional linear recurrent models. Because the matrices are stochastic, they are stable by construction. The question is empirical and sharp: on direction-dependent channels, does tying the backward pass to the forward kernel beat unconstrained bidirectional networks? The test cases are opposite direction of travel and unseen speeds, and the comparison also covers classical location-binned FSMCs that need data in every bin.
Core Claim
When the hidden channel process is non-reversible and spatially inhomogeneous, a bidirectional model should share the transition kernel across directions, not share weights or learn separate backward weights. In exact smoothing of a Markov chain with time-varying transition matrices $P_t$, the backward message obeys
$$\beta_t = P_{t+1}\,(b_{t+1} \odot \beta_{t+1}),$$
where $\beta_t(i)$ is the likelihood of future observations given state $i$ at time $t$, $b_{t+1}$ is the vector of observation likelihoods at $t+1$, and $\odot$ is elementwise product. The backward pass therefore reuses $P_{t+1}$ exactly as the forward pass does. Any asymmetry in the physics is carried automatically into the backward direction. By contrast, an independent backward LSTM has to learn that asymmetry from data. Weight-sharing schemes in the style of time reversal assume it away.
We claim that this structural tie buys three things: better smoothed state estimates in the low-data regime, robust transfer when the direction of travel or the speed changes, and a principled choice of look-ahead length. For the last point, the influence of an observation $D$ steps ahead on today's state posterior decays roughly as $|\lambda_2|^D$. Here $\lambda_2$ is the second-largest eigenvalue of the local transition matrix, which sets the chain's mixing rate. The learned kernel therefore tells the system how much buffering is worth paying for.
Relation to Prior Work
The FSMC literature supplies the regime and the classical baseline. Wang et al. build location-aware FSMCs for subway tunnels. They place SNR state boundaries with Lloyd-Max quantization and find Nakagami fading fits best. They report that coarsening position bins from 5 m to 100 m sharply raises error, and that moving from 4 to 8 SNR states lowers it. This motivates our work: the right model is fine-grained and inhomogeneous. It also points to the cost, since every bin needs its own data. Our kernel network replaces independent bins with a smooth function of position and speed. Ma et al. model low-Earth-orbit links with FSMCs over time-varying distance, which is a second naturally asymmetric regime. Impicciatore et al. and Cheng and Lau show that exploiting Markov transition structure helps estimation and power-control tracking. Both use causal filters only, and we ask what buffered look-ahead adds.
Forward-backward machinery over finite-state channels supplies the method. Finite-State Semi-Markov Channels for Nanopore Sequencing runs generalized forward-backward recursions on a hand-built 1,024-state graph with known, fixed transition priors. Ruiz et al. avoid exponential trellis smoothing with particle Gibbs sampling, which is accurate but batch-iterative. Neither learns how transitions change with side information. We keep exact recursions on a modest state count and learn the kernel.
Bidirectional sequence modeling supplies the competitors and the tension. BLUR runs linear recurrences forward and backward with parallel scans, constrains eigenvalues for stability, and reports beating Informer and PatchTST on long-horizon forecasting. Its backward recurrence has its own parameters. Osogami et al. tie forward and time-reversed models through shared parameters and report roughly tenfold RMSE reductions on a sawtooth task. They acknowledge that this backward model is not the exact reversal of a non-reversible process, and that admission is precisely the gap we target. Brakel et al. show truncated bidirectional recurrence denoises well on matched noise (9.59% word error rate on Aurora2 Test Set A) but degrades on mismatched noise. That is a cautionary result: unconstrained bidirectional models may overfit the conditions seen in training, which is exactly what our transfer tests probe. Zhou et al. exploit forward/reverse symmetry for adaptation in sleep staging. That approach relies on reversibility and would be wrong for a train leaving a tunnel.
Research Program
The model has three parts. The first is a kernel network, a small MLP that maps side information at step $t$ (track position and speed, or satellite elevation and range-rate) to a $K \times K$ stochastic matrix $P_t$ through a row-wise softmax. The second is an emission model giving $b_t(i)$, the likelihood of the observed SNR, or of raw pilot-free received samples, under state $i$. It is parameterized either as Nakagami per state, following Wang et al., or as a small density network. The third is exact forward-backward computed by parallel scans in log-space. Training maximizes the marginal likelihood of observation sequences. Because the forward-backward pass is differentiable, the gradient flows end to end, and no state labels are needed. Fixed-lag variants with look-ahead $D$ truncate the backward scan.
The analysis component estimates $\lambda_2(P_t)$ along trajectories. It then checks whether the empirical gain from look-ahead $D$ follows the predicted $|\lambda_2|^D$ saturation. If it does, operators can pick a buffer length from the learned kernel alone. The analysis also measures the learned asymmetry directly, by comparing $P_t$ on approach and departure at the same position.
Baselines
The naive reference is a causal filter: the same learned kernel and emission model, run forward only. Any gain we report must come from look-ahead, not from the kernel network. The do-nothing reference for state estimation is a memoryless per-sample SNR quantizer, which is what an i.i.d. channel view implies.
The strongest classical competitor is a location-binned FSMC fit in the manner of Wang et al.: Lloyd-Max state boundaries, a separate transition matrix per 5 m bin estimated by counting or Baum-Welch, and exact forward-backward smoothing. It is the strongest competitor because it is already exact and already inhomogeneous. If our model cannot beat it, the kernel network is not worth having. It will see the same training trajectories and the same $K$, and its bin width will be tuned over the same validation budget. No reported result found on Lacuna for smoothed state-estimation error on the Changping tunnel traces; Wang et al. report the effect of bin size and state count only qualitatively. Beating it means lower held-out per-step negative log-likelihood and lower smoothed SNR-state error at matched data. The decisive regime is when training data is scarce or the test speed lies outside the training range.
The strongest neural competitor is BLUR (Liu et al.) with a classification head over the $K$ states, together with a BiLSTM. Both receive the same side-information inputs, the same look-ahead window and a matched parameter count. Their hyperparameter searches use the same number of trials as ours. They are trained either on likelihood-free state targets taken from a genie quantizer, or on next-SNR prediction. We report both and use whichever is stronger. Beating them means equal-or-better accuracy on matched conditions and clearly smaller degradation under direction reversal and speed shift.
The ablation that isolates our key ingredient keeps the kernel network and changes only the backward pass. It compares three variants: exact reuse of $P_t$ (ours), an independently learned backward kernel network, and an Osogami-style shared-weight reversal that uses the transposed and renormalized $P_t$ (Osogami et al.). If the second variant matches ours, the "respect the arrow of time" claim fails.
Why not just…?
Why not just use the classical location-binned FSMC with forward-backward? This is the real threat. On a well-sampled, fixed-speed corridor it may match us. Its weakness is that each bin is estimated alone, so error grows as bins shrink and data thins. Wang et al.'s own coarse-versus-fine trade-off is this tension. It also cannot represent a speed it never saw. The data-scaling curve and the held-out-speed split will settle it. If the binned model matches us on both, we will report that the kernel network adds nothing and reframe the contribution as the look-ahead analysis alone.
Why not just train a BiLSTM or BLUR end to end? These models can in principle learn asymmetry. They have no reason to transfer it when the direction of travel flips, and Brakel et al. report bidirectional recurrent models losing accuracy under mismatched conditions. Our opposite-direction test, which trains on inbound runs and tests on outbound, measures this directly. If BLUR transfers as well as our model, the structural prior is unnecessary and we will say so.
Why not just filter causally and skip buffering? Buffering costs latency, and for slowly mixing channels the future may add little. The $\lambda_2$ analysis predicts where look-ahead pays. The fixed-lag sweep from $D=0$ upward measures the actual gain, so we can identify which channels justify any delay at all.
Decisive Experiments
The primary testbed is a tunnel channel simulator built to Wang et al.'s reported setup: a 2.412 GHz rectangular tunnel, Nakagami fading, and location-dependent mean SNR with distinct entry and exit profiles. Trains run at several speeds. The Changping field traces will be used if they can be obtained. A second testbed is LEO passes following the time-varying-distance model of Ma et al.. These would be generated with the Sionna link-level simulator used by Tuononen et al. when raw-sample emissions are needed.
There are four evaluation splits. The first is matched (same direction, same speeds). The second is reversed direction. The third is interpolated and extrapolated speed. The fourth is a data-scaling sweep from 1% to 100% of training trajectories. Metrics are held-out per-step negative log-likelihood of observations, smoothed state accuracy against the generating state, and SNR estimation error in dB. A downstream metric is goodput of a retransmission-aware rate adapter that commits rate decisions after a $D$-step buffer.
The decisive outcomes are three. First, our model should beat the independent-backward ablation on the reversed and extrapolated splits; if not, the core claim is false. Second, it should beat the binned FSMC at low data and match it at full data. Third, look-ahead gains should saturate at the lag predicted by $|\lambda_2|^D$.
References
- Finite-state Markov modeling of tunnel channels in communication-based train control (CBTC) systems
- Effect of Strong Time-Varying Transmission Distance on LEO Satellite-Terrestrial Deliveries
- Secure state estimation over Markov wireless communication channels
- Distributive power control algorithm for multicarrier interference network over time-varying fading channels: tracking performance analysis and optimization
- Finite-State Semi-Markov Channels for Nanopore Sequencing
- Infinite Factorial Finite State Machine for Blind Multiuser Channel Estimation
- Bidirectional linear recurrent models for sequence-level multisource fusion
- Bidirectional Learning for Time-series Models with Hidden Units
- Bidirectional truncated recurrent neural networks for efficient speech denoising
- Personalized Sleep Staging Leveraging Source-free Unsupervised Domain Adaptation
- Interpreting Deep Neural Network-Based Receiver Under Varying Signal-To-Noise Ratios