Ferkans — Interactive Telecom Tutor

Why Embedded Pilots Work

In Chapter 4 we established that the DD channel is a sparse 2D convolution: the channel's action on any input DD grid is to produce $P$ shifted-and-scaled copies of it. This convolutional structure is also the key to pilot design. Place a single isolated impulse on the DD grid — a "pilot impulse" — and surround it with a region of zeros. After the channel, the zero region receives exactly the impulse response of the channel: $P$ scaled copies of the pilot at the path locations $(\ell_i, k_i)$ . Reading off these copies gives the complete channel state.

The point is that a single pilot carries information about all $P$ paths simultaneously. This is the single-pilot channel-estimation advantage of OTFS over OFDM, where each resolvable coherence cell needs its own pilot.

Definition:
Embedded Pilot Structure

An OTFS frame with an embedded pilot partitions the DD grid into three disjoint regions:

A single pilot cell at position $(\ell_p, k_p)$ with $X_{DD}[\ell_p, k_p] = \alpha$ (a known complex amplitude).
A guard region $\mathcal{G}$ of cells around the pilot, all set to zero: $X_{DD}[\ell, k] = 0$ for $(\ell, k) \in \mathcal{G}$ .
A data region $\mathcal{D}$ containing the remaining cells, carrying QAM data symbols.

The guard region must be large enough to prevent data from bleeding into the pilot's response through the channel's convolution: $\mathcal{G} \;\supseteq\; \{(\ell_p + \ell,\,k_p + k) : 0 \leq \ell \leq l_{\max},\,-k_{\max} \leq k \leq k_{\max}\},$ where $l_{\max}, k_{\max}$ are the maximum delay and Doppler indices of the physical channel (Chapter 4).

Theorem: Pilot Response Recovers All Path Parameters

With an embedded pilot of amplitude $\alpha$ at $(\ell_p, k_p)$ and a guard region $\mathcal{G}$ sized for the channel's effective support, the received DD grid satisfies $Y_{DD}[\ell_p + \ell_i,\,k_p + k_i] \;=\; \alpha\,h_i\,e^{j\phi_i} \;+\; W_{DD}[\cdot], \quad i = 1, \ldots, P,$ for some known phase $\phi_i$ . All other cells in $\mathcal{G}$ receive only noise. Consequently, the $P$ path parameters $\{(h_i, \ell_i, k_i)\}_{i=1}^P$ are identifiable from the guard-region observations.

Each path reproduces the pilot at its own $(\ell_i, k_i)$ offset, scaled by $h_i$ . The $P$ nonzero observations in the guard region are the channel's delay-Doppler signature. The receiver reads them off like entries in a table.

Proof

Start from the DD input-output relation

From Theorem TDiscrete DD Input-Output Relation (Integer Doppler), $Y_{DD}[\ell, k] = \sum_i h_i\,e^{j\alpha_i}\,X_{DD}[(\ell - \ell_i)\bmod M, (k - k_i)\bmod N] + W_{DD}[\ell, k]$ .

Pilot structure

$X_{DD}$ has exactly one non-zero in and near $\mathcal{G}$ : the pilot at $(\ell_p, k_p)$ . Data is far enough away that its shifted copies do not reach $\mathcal{G}$ (guard condition).

Evaluate at guard cells

At cell $(\ell_p + \ell, k_p + k) \in \mathcal{G}$ : $Y_{DD}[\ell_p + \ell, k_p + k] = \sum_i h_i\,e^{j\alpha_i}\,X_{DD}[(\ell_p + \ell - \ell_i)\bmod M, (k_p + k - k_i)\bmod N] + W$ . The sum has a non-zero contribution only when $(\ell_p + \ell - \ell_i, k_p + k - k_i) = (\ell_p, k_p)$ , i.e., when $(\ell, k) = (\ell_i, k_i)$ .

Read off the path

At each guard-cell offset $(\ell, k)$ that matches one of the $P$ path indices, the observed value is $\alpha\,h_i\,e^{j\alpha_i}$ plus noise. Otherwise it is pure noise. The support of the non-noise-only cells reveals $(\ell_i, k_i)$ ; the amplitude reveals $h_i$ . $\blacksquare$

Key Takeaway

One pilot, $P$ paths. A single DD-grid impulse, surrounded by a guard region, suffices to identify the complete $P$ -path channel after one OTFS frame. This is the pilot-overhead advantage of OTFS over OFDM — an order-of-magnitude reduction in the number of pilot resource elements. It is enabled entirely by the sparsity of the DD channel and the convolution structure of the input-output relation.

Embedded Pilot: DD Grid Layout Before and After the Channel

Configure the pilot position, guard-region size, and channel complexity. Left panel: transmit DD grid (pilot, guard region shown, data region). Right panel: received DD grid, with path copies of the pilot appearing in the guard region. Verify that the guard region is large enough to contain all path copies — if not, data leakage corrupts the pilot response. This is a live demo of the theorem above.

Parameters

Delay bins

M

32

Doppler bins

N

16

Number of paths

P

3

l_{\max}

4

k_{\max}

1

Pilot SNR (dB)25

Seed3

Definition:
Minimum Guard Region

The minimum guard region for a channel with maximum delay index $l_{\max}$ and maximum Doppler index $k_{\max}$ is $|\mathcal{G}_{\min}| \;=\; (l_{\max} + 1)(2 k_{\max} + 1).$ This is the product of (one-sided) delay span and two-sided Doppler span — Doppler can be positive or negative because reflectors may approach or recede from the receiver.

Guard cells carry no data, so the pilot overhead is $\eta_{\text{pilot}} \;=\; \frac{|\mathcal{G}_{\min}| + 1}{MN} \;=\; \frac{(l_{\max} + 1)(2 k_{\max} + 1) + 1}{MN}.$

Example: Overhead for an Urban Vehicular Channel

An OTFS system has $(M, N) = (512, 32)$ with $W = 10$ MHz and $T = 4$ ms. The channel has $\tau_{\max} = 5\,\mu$ s and $f_D = 500$ Hz. Compute $(l_{\max}, k_{\max})$ , the minimum guard-region size, and the pilot overhead.

Solution

Grid resolutions

$\Delta\tau = 1/W = 100$ ns. $\Delta\nu = 1/T = 250$ Hz.

Channel support indices

$l_{\max} = \lceil \tau_{\max}/\Delta\tau\rceil = \lceil 5000/100\rceil = 50$ . $k_{\max} = \lceil f_D/\Delta\nu\rceil = \lceil 500/250\rceil = 2$ .

Guard region

$|\mathcal{G}_{\min}| = 51 \cdot 5 = 255$ cells. Plus the pilot: 256.

Overhead

$\eta_{\text{pilot}} = 256/(512 \cdot 32) = 256/16{,}384 = 1.56\%$ . Compare with 5G NR DMRS: $\sim 7$ – $14\%$ depending on numerology. The embedded pilot saves 5–12 percentage points of resource elements — directly converted to extra throughput.

Embedded Pilot DD-Grid Layout — Schematic of the embedded pilot structure. The DD grid $(M \times N)$ is partitioned into three regions: a single pilot cell (blue), a guard region of size $(l_{\max} + 1) \times (2 k_{\max} + 1)$ with zeros (gray), and a data region with QAM symbols (green). The receiver reads path parameters from the guard region, then detects data from the data region. The dashed rectangle marks the minimum guard region that prevents data-pilot interference.

Pilot Impulse Creates $P$ Copies in the Guard Region

Animation of a single pilot impulse (blue) passing through a three-path DD channel. At the receiver (guard region), three copies of the pilot appear at

(\ell_p + \ell_i, k_p + k_i)

with amplitudes

\alpha\,h_i

. The data region remains empty throughout the pilot transmission. The three observations in the guard region are the channel's complete fingerprint.

⚠️Engineering Note

Pilot Power Boosting

The pilot's SNR determines the fidelity of the channel estimate. In practice, the pilot cell carries higher power than a data cell: a typical choice is $|\alpha|^2 = 10 \cdot \mathbb{E}|X_{\text{data}}|^2$ (10 dB pilot boost). This costs 10 dB of effective pilot power but keeps the detector threshold comfortably above the noise floor for reliable path detection (Section 2).

In terms of frame-level power budget: $P_{\text{total}} = |\alpha|^2 + |\mathcal{D}| \cdot \mathbb{E}|X_{\text{data}}|^2$ , with $|\mathcal{D}| = MN - |\mathcal{G}_{\min}| - 1$ . For a typical frame $|\mathcal{D}| \approx 0.98\,MN$ , so the pilot boost costs a negligible fraction of total power — a favorable trade for better channel estimates.

Practical Constraints

•
Typical pilot boost: 6–10 dB above average data power
•
Pilot fraction of frame energy: $\sim 10^{-3}$ for standard parameters
•
Higher boost improves MSE but increases PAPR of the OTFS waveform

📋 Ref: 3GPP TS 38.211, §6.4

Common Mistake: Undersized Guard Region Corrupts Everything

Mistake:

Using a guard region smaller than the channel's effective support: $|\mathcal{G}| < (l_{\max} + 1)(2 k_{\max} + 1)$ . Data from outside the guard region leaks into the pilot response, corrupting the path detection.

Correction:

Always size the guard region for the worst-case channel in the deployment. Use the 99th percentile of $\tau_{\max}$ and $f_D$ from measurement campaigns (e.g., 3GPP TR 38.901). In unknown deployments, use 2× the expected values. Under-allocation causes the detector to report ghost paths where data leakage mimics a path — a silent failure mode that degrades BER without obvious diagnostic signals.

Why This Matters: Connection to Sparse Recovery

The embedded-pilot problem is a textbook instance of sparse signal recovery: a noisy measurement vector (the guard-region observations) is the linear response of a $P$ -sparse signal (the path taps) through a known dictionary (the delta-like pilot). In the RF Imaging book (Chapter 13), we studied sparse-recovery algorithms like OMP, LASSO, and AMP in detail. Any of these apply here — but the special structure (single pilot, integer-indexed paths) makes threshold-based detection of Section 2 the most efficient option.

Embedded Pilot Structure and Guard Regions