Ferkans — Interactive Telecom Tutor

Channel Prediction from Geometry

The power of SAC comes from a simple observation: in the DD domain, the channel is determined by the scatterers' geometry (positions, velocities) and the array manifolds. Once the scatterers are tracked, the next frame's channel is predicted deterministically — provided the kinematic model is accurate and the array manifolds are calibrated. This section derives the channel prediction machinery, bounds its accuracy, and shows how it interacts with the sensing CRB.

Definition:
Predicted Channel from Sensing

At frame $t$ , the sensor produces an estimate $\hat\Theta^{(t)}$ with covariance $\hat{\mathbf{P}}^{(t|t)}$ (from the Kalman filter of Ch. 13 §4). The one-step-ahead predicted scene is $\hat\Theta^{(t+1|t)} \;=\; \mathbf{A}\, \hat\Theta^{(t)}, \qquad \hat{\mathbf{P}}^{(t+1|t)} \;=\; \mathbf{A}\, \hat{\mathbf{P}}^{(t|t)}\, \mathbf{A}^T + \mathbf{Q},$ where $\mathbf{A}$ is the state-transition matrix and $\mathbf{Q}$ the process noise covariance.

The predicted DD channel at frame $t+1$ is $\hat{\mathbf{H}}_{DD}^{(t+1|t)}[\ell, k] \;=\; \mathbf{H}_{DD}(\hat\Theta^{(t+1|t)}),$ with uncertainty propagated via $\mathrm{MSE}(\hat{\mathbf{H}}_{DD}^{(t+1|t)}) \;=\; \mathbf{J}\, \hat{\mathbf{P}}^{(t+1|t)}\, \mathbf{J}^H,$ where $\mathbf{J}$ is the scene-to-channel Jacobian.

Theorem: Channel Prediction Accuracy

Let $\hat{\mathbf{h}}^{(t+1|t)} = \mathbf{H}_{DD}(\hat\Theta^{(t+1|t)})$ be the predicted channel at frame $t+1$ . Under the linearized scene-to-channel map, the prediction MSE is $\mathrm{MSE}(\hat{\mathbf{h}}^{(t+1|t)}) \;=\; \mathrm{CRB}(\Theta^{(t+1|t)}) \,+\, \|\mathbf{J}\|^2 T_{\text{fr}}^2 \sigma_a^2,$ where the first term is the sensing contribution and the second is the process-noise contribution over the prediction step of duration $T_{\text{fr}}$ .

Consequence. If the sensing CRB is $\sigma_s^2$ and process noise is $\sigma_p^2 T_{\text{fr}}^2$ , the predicted-channel MSE is bounded by $\sigma_s^2 + \sigma_p^2 T_{\text{fr}}^2$ . At typical operating points ( $\sigma_s^2 \ll \sigma_p^2 T_{\text{fr}}^2$ ), the prediction is dominated by process noise — and the sensing CRB does not need to be tight, just good enough.

This is the quantitative form of the design principle "sense enough to distinguish regimes, predict from dynamics". The steady-state prediction error is limited by how erratically the target moves, not by how precisely the initial measurement was taken. If the process is near-deterministic (pedestrian walking straight), prediction accuracy can approach the sensing CRB. If the process is erratic (vehicle in dense traffic), process noise dominates.

Proof

Decomposition

$\hat{\mathbf{h}}^{(t+1|t)} - \mathbf{h}^{(t+1)} = \mathbf{J} (\hat\Theta^{(t+1|t)} - \Theta^{(t+1)}) = \mathbf{J}[\mathbf{A} (\hat\Theta^{(t)} - \Theta^{(t)}) - \mathbf{u}^{(t)}]$ . The first term is the sensing error at time $t$ ; the second is the process noise over one step.

Covariance

Squared and averaged: $\mathrm{MSE} = \mathbf{J}\mathbf{A} \mathrm{CRB}(\Theta^{(t)})\mathbf{A}^T\mathbf{J}^T + \mathbf{J}\mathbf{Q}\mathbf{J}^T$ . First term: sensing error propagated to $t+1$ . Second term: process noise contribution.

Simplification

For a constant-velocity model, $\mathbf{A}$ is close to identity over one frame, so $\mathrm{CRB}(\Theta^{(t+1|t)}) \approx \mathrm{CRB}(\Theta^{(t)})$ . $\blacksquare$

Definition:
Kinematic Prediction Models

Four classes of kinematic models for prediction:

Constant-Velocity (CV): state = (position, velocity), evolves linearly. Good for predictable trajectories (freeway cruise). Process noise: $\sigma_a = 0.5$ m/s $^2$ (typical highway).

Constant-Acceleration (CA): state = (position, velocity, acceleration). Accounts for sustained acceleration. Used when vehicle maneuvers are smooth (acceleration changes slowly). Process noise: $\sigma_j = 0.1$ m/s $^3$ (jerk).

Coordinated-Turn (CT): state = (position, velocity, turn rate). For circular motion (roundabouts, lane changes). Process noise on turn rate.

Interacting Multiple Model (IMM): Maintains $k$ parallel filters (CV, CA, CT) with Markov transition probabilities. Switches model based on observed maneuvers. Robust to model mismatch.

Theorem: IMM Prediction Robustness

The Interacting Multiple Model (IMM) filter with $k$ candidate models and Markov transition probabilities matched to the observed maneuver rate achieves prediction MSE $\mathrm{MSE}_{\text{IMM}} \;\leq\; \min_i \mathrm{MSE}_i + \epsilon$ where $\mathrm{MSE}_i$ is the MSE of the $i$ -th single-model filter and $\epsilon$ is a small model-switching cost.

Consequence. IMM adapts to maneuver changes without the overshoot/lag of single-model filters. Cost: $k\times$ compute. For automotive, $k = 3$ (CV + CA + CT) is typical.

A single kinematic model always gives up some performance: too simple, it misses maneuvers; too complex, it adds noise. IMM hedges by running multiple filters in parallel and weighting them by how well each matches recent observations. The multiplicative compute cost is modest; the robustness gain is substantial, especially for automotive and UAV applications.

Proof

Filter ensemble

Each of $k$ models propagates a posterior state distribution. Weights $\mu_i^{(t)}$ track model-likelihood given observations.

Mixing

At each step, mixed initial conditions $\bar{\mathbf{s}}^{(t-1|t-1)}_i = \sum_j \mathbf{s}^{(t-1|t-1)}_j \mu_{ij}$ reset each filter.

Fusion

Final estimate is the weighted average: $\hat{\mathbf{s}}^{(t|t)} = \sum_i \mathbf{s}^{(t|t)}_i \mu_i^{(t)}$ .

Performance

Proved by Blom and Bar-Shalom (1988): $\mathrm{MSE}_{\text{IMM}} \leq \min_i \mathrm{MSE}_i + \epsilon$ , with $\epsilon \to 0$ as the transition model matches the true maneuver rate. $\blacksquare$

Channel Prediction MSE vs Prediction Horizon

Plot the predicted-channel MSE as a function of horizon $T_{\text{pred}}$ (frames), comparing: constant-velocity model, constant-acceleration, and IMM. Sliders: target acceleration variance, sensing SNR, frame rate.

Parameters

\sigma_a

(m/s²)1

Sensing SNR (dB)15

Frame rate (Hz)100

Example: Highway SAC: 5-Frame Prediction

A BS tracks a vehicle at 30 m/s on a highway. Sensing CRB on velocity is $10^{-4}$ (m/s) $^2$ (tight). Acceleration variance is 0.5 m/s $^2$ (cruise mode, minor throttle adjustments). Frame rate 100 Hz.

Compute the predicted channel MSE at 1, 2, 5, 10 frames ahead and state the maximum pilot-free interval.

Solution

MSE formula

$\sigma_h^2(T) \approx \mathrm{CRB}_v + T \cdot T_{\text{fr}} \cdot \sigma_a^2$ . CRB term: $10^{-4}$ . Process: $T \cdot 0.01 \cdot 0.25 = 2.5 \times 10^{-3} T$ .

Per-horizon values

$T = 1$ : $\sigma_h^2 = 2.6 \times 10^{-3}$ . Tiny. $T = 2$ : $5.1 \times 10^{-3}$ . Minor. $T = 5$ : $1.3 \times 10^{-2}$ . Noticeable. $T = 10$ : $2.5 \times 10^{-2}$ . Approaching tolerance.

Tolerance threshold

For 10% comms SINR loss: $\sigma_h^2 \leq 0.1 \cdot \|\mathbf{h}\|^2$ . Assuming $\|\mathbf{h}\|^2 = 0.1$ : $\sigma_h^2 \leq 0.01$ . $T \leq 4$ frames.

Pilot-free interval

4 frames between pilots at 100 Hz → refresh every 40 ms. Classical baseline: 10 ms. SAC savings: $\sim 75\%$ .

⚠️Engineering Note

Prediction Quality Monitor

Deploy a prediction-quality monitor to protect the link:

Consistency check: Compare the predicted channel with the next frame's implicit channel (from data detection residuals). Large discrepancy $\Rightarrow$ acceleration event or lost track.
Pilot-refresh trigger: When the monitor detects a discrepancy above threshold $\gamma$ , insert an emergency pilot on the next frame. Cost: occasional pilot overhead; benefit: link reliability.
Track quality: Feedback to the sensing subsystem — low track quality (high innovation norm) $\Rightarrow$ increase sensing SNR (more power to sensing-oriented directions in the joint beamformer).

Typical automotive design: refresh trigger every 50 ms worst-case, 5 ms in sustained acceleration events. Falls back gracefully to classical pilot-based comms if the sensing subsystem fails (single-point of failure avoidance).

Practical Constraints

•
Consistency-check against data-detection residual
•
Pilot refresh on threshold breach + periodic safety refresh
•
Graceful fallback if sensing fails

Common Mistake: Kinematic Model Mismatch

Mistake:

Using a constant-velocity (CV) model exclusively. A vehicle accelerating to pass another vehicle violates the CV assumption, and the predicted channel lags the true one by a growing margin. Pilot-refresh triggers fire but cannot recover the baseline rapidly.

Correction:

Deploy IMM or CA model for automotive; switch to CV only for stable cruise. Detect maneuvers via acceleration detection (second- derivative of state) and trigger model switch. This adds $\sim 2\times$ compute but eliminates the maneuver-induced prediction failure mode.

The SAC Feedback Loop: Sense $\to$ Predict $\to$ Precode

Animation of the four-step SAC loop: sense target, predict channel from

\hat\Theta

, precode using the prediction, transmit. Each frame repeats the loop, with the target-scene estimate driving the precoder without additional pilot overhead.

Historical Note: The Kalman Legacy

Rudolph Kalman's 1960 paper "A New Approach to Linear Filtering and Prediction Problems" (J. Basic Eng.) established the framework used in this section: a state estimator that is optimal for linear Gaussian systems and nearly optimal (extended/unscented variants) for nonlinear ones. Kalman's key insight was the recursive update form — $\mathbf{K}^{(t)} = \mathbf{P}^{(t|t-1)} \mathbf{H}^T \mathbf{S}^{-1}$ — which separates the prediction step from the update step.

Classical radar-tracking textbooks (Bar-Shalom, Li, Blackman) applied the Kalman filter to target tracking. The extension to SAC here uses the same framework but treats the channel (not the target alone) as the quantity to be tracked. The DD-domain representation makes the scene-to-channel map explicit, unifying two distinct tracking problems into one.

Channel Prediction from Radar Estimates