Ferkans — Interactive Telecom Tutor

Closing the Loop

Classical communication systems separate sensing and comms: a dedicated radar (or a CSI-estimation sub-frame) produces channel estimates; the comms transmitter uses them for precoding. The estimates are always of the past — by the time they arrive at the precoder, the channel has already changed. At vehicular speeds, this is catastrophic: a 100-km/h vehicle at 28 GHz has coherence time $\sim 1$ ms, but CSI feedback and precoder application take longer. The DD-domain ISAC framework of Chapters 12-13 enables a cleaner approach: the sensing and comms functions share the same waveform, and the same target-scene estimate yields a short-horizon forecast of the channel. This chapter makes that forecast explicit and quantifies its benefits.

,

Definition:
Sensing-Assisted Communication (SAC)

Sensing-assisted communication (SAC) is a paradigm in which outputs of a sensing subsystem — namely target position, velocity, and scattering coefficients — are used as inputs to the communication subsystem. Concretely:

Sensing estimates $\hat{\Theta}^{(t)} = \{(\hat\tau_i, \hat\nu_i, \hat\theta_i, \hat\phi_i, \hat a_i)\}_i$ from the ISAC waveform.
These estimates predict the comms channel $\hat{\mathbf{h}}^{(t+1)}$ at the next frame.
The predicted channel drives the precoder, resource allocation, and beam management.

The complementary paradigm — comms-assisted sensing (CAS) — uses data detection results to refine sensing estimates (the joint EM iteration of Ch. 12). Together, SAC and CAS close the ISAC loop.

,

SAC Architecture at a Glance

The SAC pipeline at a BS or UE runs the following loop per frame:

Receive current-frame waveform; run joint estimation-detection (Ch. 12 algorithm).
Update target-scene estimate $\hat\Theta^{(t)}$ and decoded data.
Propagate $\hat\Theta^{(t)}$ through the state-evolution model to get $\hat\Theta^{(t+1|t)}$ .
Map $\hat\Theta^{(t+1|t)}$ to predicted channel $\hat{\mathbf{h}}^{(t+1)}$ .
Design precoder, beamformer, resource allocation based on $\hat{\mathbf{h}}^{(t+1)}$ .

The key point is that step 4 — mapping scene $\to$ channel — is deterministic given the geometry. No separate channel estimation is needed. The pilot overhead required by classical CSI-feedback systems is replaced by sensing.

Theorem: Scene-to-Channel Map

Given the target scene $\Theta = \{(\tau_i, \nu_i, \theta_i, \phi_i, a_i)\}_{i=1}^{P}$ , the DD-domain MIMO channel matrix is $\mathbf{H}_{DD}(\Theta)[\ell, k] \;=\; \sum_{i=1}^{P} a_i\, \mathbf{a}_r(\theta_i) \mathbf{a}_t(\phi_i)^H \,\delta[\ell - \ell_i, k - k_i] \,e^{-j2\pi \nu_i \ell_i / (MN)}.$ Consequence. If $\hat\Theta$ is an unbiased estimate with covariance $\mathrm{CRB}(\Theta)$ , then the channel estimate $\hat{\mathbf{h}} = \mathbf{H}_{DD}(\hat\Theta)$ has MSE $\mathrm{MSE}(\hat{\mathbf{h}}) \;\sim\; \mathbf{J}^T \,\mathrm{CRB}(\Theta)\, \mathbf{J},$ where $\mathbf{J} = \partial \mathbf{h}/\partial \Theta$ is the Jacobian of the scene-to-channel map. For small errors in the target parameters, the channel estimate degrades smoothly — no additional noise is introduced beyond the sensing CRB.

The point is that channel estimation and sensing are the same estimation problem in the DD domain. Once the sensing CRB is bounded, the channel MSE is bounded too. The gain over classical pilot-based estimation comes from the fact that the sensing observation uses all the DD cells, not just the pilot cells — so the effective sample size is much larger.

Proof

Linearization

$\hat{\mathbf{h}} - \mathbf{h} = \mathbf{J}(\hat\Theta - \Theta) + \mathcal{O}(\|\hat\Theta - \Theta\|^2)$ , where the linearization is around the true scene.

Covariance

Taking expectation of the outer product: $\mathrm{MSE}(\hat{\mathbf{h}}) = \mathbb{E}[(\hat{\mathbf{h}} - \mathbf{h})(\hat{\mathbf{h}} - \mathbf{h})^H] = \mathbf{J}\,\mathrm{CRB}(\Theta)\,\mathbf{J}^T + o(1)$ .

Interpretation

The sensing CRB propagates through the geometry (Jacobian $\mathbf{J}$ ) to give the comms channel MSE. No additional estimation noise. $\blacksquare$

,

Key Takeaway

Channel estimation is geometry estimation in the DD domain. In the DD-sparse representation, estimating the channel means estimating the $P$ -path geometry $\Theta = \{(\tau_i, \nu_i, \theta_i, \phi_i, a_i)\}$ . A good sensing algorithm, which estimates $\Theta$ with CRB-level accuracy, is a good channel estimator. The pilot overhead becomes redundant — except for covering the gain parameters $a_i$ , which can be estimated from a single reference symbol per frame.

Definition:
Prediction Horizon

The prediction horizon $T_{\text{pred}}$ is the number of frames over which $\hat\Theta^{(t)}$ remains predictive of the channel with acceptable MSE. It depends on:

Maneuver dynamics: higher acceleration shortens $T_{\text{pred}}$ .
Doppler spread: wider spectrum (faster motion) shortens it.
CRB: tighter sensing widens it (more accurate initial state).
Frame duration: shorter frames provide more updates per second.

Typical values for automotive mmWave:

Pedestrian ( $v \leq 2$ m/s): $T_{\text{pred}} \approx 100$ frames (= 1 second at 100 Hz frame rate).
Vehicle ( $v = 20$ - $30$ m/s): $T_{\text{pred}} \approx 10$ - $30$ frames.
LEO ( $v = 7$ km/s, long-distance): $T_{\text{pred}} \leq 1$ frame (per-frame prediction only).

Theorem: Prediction Horizon vs Coherence Time

For a target with acceleration variance $\sigma_a^2$ and sensing CRB on radial velocity $\mathrm{CRB}_v$ , the prediction horizon $T_{\text{pred}}$ at which the predicted Doppler error equals the coherence tolerance $\nu_{\text{tol}}$ satisfies $T_{\text{pred}} \;\approx\; \frac{\nu_{\text{tol}}}{\sigma_a \lambda^{-1}} \cdot \frac{1}{\sqrt{\mathrm{CRB}_v}}.$ Numerical: for $\sigma_a = 1$ m/s $^2$ , $\lambda = 4$ mm (77 GHz), $\nu_{\text{tol}} = 10$ Hz, $\mathrm{CRB}_v = 10^{-4}$ (m/s) $^2$ : $T_{\text{pred}} \approx 10$ ms $\approx 1$ - $10$ frames at 100 Hz frame rate.

Consequence. For automotive scenarios, the prediction horizon is 1-10 frames — long enough to skip pilots for $\geq 90\%$ of frames. The SAC paradigm saves $\sim 90\%$ of pilot overhead.

A target's velocity does not change much in a few frames unless it is accelerating hard. If the sensing CRB on velocity is tight, the predicted channel is accurate for multiple frames. Only at acceleration events does the horizon shrink; otherwise, the BS can coast on its sensing-derived channel estimate for many frames without pilot refresh.

Proof

Error growth model

State evolves as $\mathbf{s}^{(t+1)} = \mathbf{A}\mathbf{s}^{(t)} + \mathbf{u}$ with process noise $\mathbf{Q}$ . Error accumulates linearly: $\sigma_{\text{err}}^2(T) = \sigma_0^2 + T \sigma_u^2 = \mathrm{CRB} + T \sigma_a^2 T^2 = \mathrm{CRB} + T^3 \sigma_a^2$ (dominated by the process noise at large $T$ ).

Horizon criterion

Set $\sqrt{\sigma_{\text{err}}^2(T)} \leq \sigma_{\text{tol}}$ : $T \leq \sigma_{\text{tol}}/\sigma_a^{1/3}$ for cubic growth. Simplified linear scaling at smaller $T$ : $T \approx \sigma_{\text{tol}} / \sigma_a$ .

Doppler tolerance

Doppler error propagates through velocity error as $\nu_{\text{err}} = (2/\lambda) \cdot v_{\text{err}}$ . Set $\nu_{\text{err}} \leq \nu_{\text{tol}}$ : $T \approx \nu_{\text{tol}} \lambda / (2 \sigma_a)$ .

Numerical

For values above: $T \approx 10 \cdot 0.004 / 2 = 0.02$ s = 20 ms. Sensing CRB term adds minor contribution. $\blacksquare$

Example: SAC Pilot Savings for Vehicular Link

A BS serves a vehicle at 30 m/s on a highway (77 GHz, frame rate 100 Hz). Classical pilot-based comms requires a pilot every frame (10% of resources). SAC uses sensing to extrapolate the channel. Assume sensing CRB on velocity is $10^{-4}$ (m/s) $^2$ and target acceleration variance is 1 m/s $^2$ .

(a) Compute the prediction horizon. (b) Determine the minimum pilot refresh rate. (c) Quantify pilot overhead savings.

Solution

Prediction horizon

From Thm. 14.5: $T_{\text{pred}} \approx \nu_{\text{tol}} \lambda / (2 \sigma_a) = 10 \cdot 0.004 / 2 = 0.02$ s = 2 frames at 100 Hz.

Pilot refresh

Refresh every 2-3 frames. Effective pilot rate: $\sim 30\%$ of baseline (refresh 30/100 = 30 frames per second).

Savings

Pilot overhead: $10\% \times 30\% = 3\%$ (from 10% baseline). Rate savings: $\sim 7\%$ absolute. At 100 Mbps link: 7 Mbps more usable throughput.

Summary

SAC effectively eliminates $\sim 70\%$ of pilot resources in vehicular scenarios. Combined with the DD-domain sparsity (Ch. 4), this is the quantitative case for SAC in 6G.

SAC vs Pilot-Based Rate for Varying Mobility

Plot effective comms rate vs vehicle velocity (0-300 km/h) for three schemes: pilot-based (fixed 10% overhead), SAC (variable overhead based on $T_{\text{pred}}$ ), and perfect CSI (no overhead). Sliders: frame rate, BS $N_t$ .

Parameters

Frame rate (Hz)100

N_t

16

Sensing SNR (dB)15

🎓CommIT Contribution(2023)

Sensing-Assisted Communication Framework

F. Liu, G. Caire, Z. Fei, Y. Cui — IEEE Trans. Wireless Communications

The CommIT contribution on SAC establishes the architectural framework used in this chapter: sensing outputs drive comms operations (precoder, beamformer, scheduler). Three foundational results:

Scene-to-channel map: Derived closed-form expression for the channel prediction from target parameters (Theorem 14.3).
Prediction horizon: Quantitative bounds on the duration over which sensing predicts the channel (Theorem 14.5).
Pilot savings: For automotive mmWave, $\sim 70\%$ of pilot resources can be eliminated, yielding $\sim 30\%$ spectral efficiency gain.

This paper is the pivot from "ISAC as efficient resource sharing" to "ISAC as active comms performance enhancement". The CommIT framework unifies SAC and CAS (Chapter 12 joint estimation- detection) into a single feedback loop. Subsequent work (Chapter 13 predictive tracking, Chapter 17 cell-free OTFS) builds on this foundation.

commitsacisac

Common Mistake: Don't Predict Beyond the Horizon

Mistake:

Assuming that a single sensing estimate holds indefinitely. Using $\hat\Theta^{(t)}$ to predict channels at $t + 100$ frames (1 second ahead at 100 Hz) when maneuver dynamics say the prediction horizon is 2-3 frames.

Correction:

Implement a pilot refresh policy: schedule a pilot when either (a) the predicted MSE has grown beyond a threshold, or (b) a fixed number of frames has elapsed since the last pilot. Typical: refresh every 5-10 frames, plus adaptive refresh when sensing detects an acceleration event. The refresh saves pilot overhead without compromising link reliability.

Why This Matters: From ISAC Beamforming (Ch 13) to SAC (Ch 14)

Chapter 13 designed the MIMO-OTFS-ISAC transmitter assuming the target scene was known. This chapter makes explicit the logic flow of that assumption: the scene is always being sensed by the waveform, so the precoder uses the most recent estimate. SAC is thus not an optional add-on but the natural way to run an ISAC system — every frame computes a new scene estimate, and the precoder tracks it. The DD-domain framework makes this feedback loop coherent and stable.

The Sensing-Assisted Communication Paradigm