Ferkans — Interactive Telecom Tutor

Sparsity Is a Resource, Not a Curiosity

We have said repeatedly that the DD channel is "sparse." This section turns that qualitative observation into quantitative advantages — concrete reductions in pilot overhead, detection complexity, and estimation error — that motivate the detailed transceiver design of Chapters 6-8.

The point is that sparsity is not merely an aesthetic property. It determines whether OTFS can be deployed in 5G/6G: a channel with $P$ parameters can be estimated from $O(P)$ pilots instead of $O(MN)$ , detected with $O(PMN)$ operations instead of $O(MN)^2$ , and tracked robustly under noise that would wash out a dense $MN$ -parameter estimate.

Definition:
Effective Channel Support

The effective support of the DD channel on the $(M, N)$ grid is the rectangle $\mathcal{S} \;=\; \{(\ell, k) : 0 \leq \ell \leq l_{\max},\,-k_{\max} \leq k \leq k_{\max}\},$ with $l_{\max} = \lceil \tau_{\max} W \rceil$ and $k_{\max} = \lceil f_D T \rceil$ . The actual $P$ paths lie within $\mathcal{S}$ , often covering only a small fraction of it. The effective support cardinality $|\mathcal{S}| = (l_{\max} + 1)(2 k_{\max} + 1)$ is the natural upper bound for the number of "active" channel taps on the grid.

Theorem: Minimum Pilot Overhead for DD Channel Estimation

Consider a single pilot impulse placed at $(\ell_p, k_p)$ on the DD grid, surrounded by a guard region $\mathcal{G} = \{(\ell_p + \ell, k_p + k) : 0 \leq \ell \leq l_{\max},\,-k_{\max} \leq k \leq k_{\max}\}$ in which no data is transmitted. After passing through a $P$ -path channel with effective support within $\mathcal{S}$ , the received DD grid contains in $\mathcal{G}$ a scaled-and-shifted copy of the pilot with amplitudes $\{h_i\}_{i=1}^P$ at positions $\{(\ell_i, k_i)\}$ . All $P$ parameters can be recovered from the $|\mathcal{G}|$ guard-region observations provided $\mathrm{SNR}_{\text{pilot}} \cdot |\mathcal{G}| \;\geq\; P,$ i.e., the pilot SNR is large enough that each path's peak stands out above the noise floor.

The pilot overhead is $|\mathcal{G}|/(MN) = (l_{\max} + 1)(2k_{\max} + 1)/(MN)$ .

The pilot is "delta-like" in DD; the channel's action is to produce $P$ copies of it in the guard region. Each copy reveals one $(h_i, \ell_i, k_i)$ triple — all the information needed to reconstruct $h(\tau, \nu)$ . The guard region must be large enough to contain all copies without collision with data cells; the pilot SNR must be high enough for reliable peak detection.

Proof

Received guard region

Let the pilot be $X_p[\ell_p, k_p] = \alpha$ , $0$ elsewhere in its region. Applying Theorem TDiscrete DD Input-Output Relation (Integer Doppler), $Y[\ell_p + \ell, k_p + k] = \sum_i h_i\,\alpha\,\mathbf{1}_{(\ell, k) = (\ell_i, k_i)}$ (ignoring phase for clarity). The guard region contains exactly the $P$ path impulses.

Identifiability

Each $h_i$ is read off at position $(\ell_p + \ell_i, k_p + k_i)$ . Distinct paths (integer indices) give distinct positions, hence no ambiguity. This requires the guard region to fit all paths without overlap: $\mathcal{G} \supseteq \{(\ell_p + \ell_i, k_p + k_i)\}_i$ , equivalent to $\mathcal{S} \subseteq \mathcal{G}$ .

SNR requirement

Each observed amplitude is $h_i \alpha$ ; noise has variance $\sigma^2$ . Detection requires $|h_i|^2 \alpha^2 / \sigma^2 \gtrsim \text{threshold}$ . With $\mathrm{SNR}_{\text{pilot}} = \alpha^2 / \sigma^2$ and the channel normalized so $\sum |h_i|^2 = 1$ , the condition is equivalent to the stated inequality. $\blacksquare$

Embedded Pilot in a Guard Region (Preview)

Place a single pilot impulse on the DD grid surrounded by a guard region sized for the channel's maximum delay and Doppler. After passing through the channel, the guard region contains $P$ copies of the pilot at the path coordinates — directly readable as $(h_i, \ell_i, k_i)$ . Adjust pilot power, guard size, and channel complexity to see how the peaks stand out against the noise floor. This is a preview of the embedded pilot estimator of Chapter 7.

Parameters

Delay bins

M

32

Doppler bins

N

16

P

3

l_{\max}

4

k_{\max}

1

Pilot SNR (dB)25

Seed5

Example: Pilot Overhead at 5G Numerology

At 5G NR numerology-1 ( $\Delta f = 30$ kHz, $T_s = 33.3\,\mu$ s), an OTFS frame uses $M = 512$ , $N = 16$ . The channel has $\tau_{\max} = 2\,\mu$ s, $f_D = 500$ Hz. Compute the pilot overhead using an embedded pilot with guard region of minimum size.

Solution

Grid resolutions

$W = M \Delta f = 15.36$ MHz, so $\Delta\tau = 1/W = 65$ ns. $T = N T_s = 533\,\mu$ s, so $\Delta\nu = 1/T = 1.88$ kHz.

Channel support indices

$l_{\max} = \lceil 2000/65 \rceil = 31$ , $k_{\max} = \lceil 500/1880 \rceil = 1$ .

Guard region size

$|\mathcal{G}| = (l_{\max} + 1)(2 k_{\max} + 1) = 32 \cdot 3 = 96$ cells. Grid total: $MN = 8192$ . Pilot overhead: $|\mathcal{G}|/(MN) = 96/8192 = 1.2\%$ .

Comparison with OFDM

Standard 5G NR allocates roughly 1/14 of resource elements to demodulation reference signals — about 7.1% overhead. The DD embedded pilot halves this (to 1.2% for similar channel, for a specific numerology), and does so with a single pilot identifying all $P$ paths — no per-subcarrier pilot as in OFDM. At this rate, the embedded pilot advantage translates to roughly 6% extra throughput.

Theorem: MSE Scaling of Least-Squares DD Channel Estimation

Given a pilot of power $P_{\text{pil}}$ and a guard region large enough to fit all paths, the least-squares estimate $\hat{h}_i$ of each path gain has variance $\mathrm{Var}(\hat{h}_i - h_i) \;=\; \frac{\sigma^2}{P_{\text{pil}}}.$ The total MSE of the channel estimate is $\mathrm{MSE} = P \cdot \sigma^2/P_{\text{pil}}$ , independent of the grid size $MN$ .

Because the DD channel has only $P$ unknown parameters, the estimation error scales with $P$ , not with $MN$ . In the TF domain, estimating $H(f, t)$ at the Nyquist rate requires $O(\tau_{\max} f_D T W)$ pilot samples; the MSE scales with this larger number. Sparsity is the reason OTFS estimation is far cleaner than OFDM estimation at high mobility.

Proof

Per-path estimator

At position $(\ell_p + \ell_i, k_p + k_i)$ , the observation is $y_i = h_i \alpha + w_i$ with $w_i \sim \mathcal{CN}(0, \sigma^2)$ . The LS estimate is $\hat{h}_i = y_i / \alpha$ .

Error variance

$\hat{h}_i - h_i = w_i / \alpha$ . Variance: $\sigma^2/\alpha^2 = \sigma^2/P_{\text{pil}}$ .

Total MSE

Sum over $P$ independent estimates: $\mathrm{MSE} = P \cdot \sigma^2/P_{\text{pil}}$ . Independent of $MN$ — the sparsity renders the estimation problem essentially $P$ -dimensional. $\blacksquare$

Key Takeaway

Sparsity converts an $MN$ -dimensional channel estimation problem into a $P$ -dimensional one. The consequences are: (i) pilot overhead drops from $\sim 10\%$ (OFDM) to $\sim 1\%$ (OTFS) under similar conditions; (ii) estimation MSE scales with $P$ , not $MN$ ; (iii) detection complexity scales with $P \cdot MN$ rather than $(MN)^2$ . All three concrete advantages flow from the single structural fact that $h(\tau, \nu)$ has support of cardinality $P$ on the physical channel, and the DD grid resolves that support cleanly.

⚠️Engineering Note

When Sparsity Breaks Down

The sparsity argument requires $P$ to be small and the paths to be resolvable. Three regimes stress this assumption:

Rich scattering (dense urban NLOS): when the number of significant paths $P \sim 50$ or more (rare but possible in certain deep-scatter environments), the "sparsity" becomes marginal. The DD channel has many entries but each is weak, and the advantage over OFDM shrinks. In the extreme limit $P \to \infty$ , OTFS has no essential advantage over OFDM.
Unresolvable clusters: when multiple physical reflectors are closer than $\Delta\tau = 1/W$ in delay or $\Delta\nu = 1/T$ in Doppler, they merge into a single grid cell. The effective sparsity is determined by the number of resolvable clusters, not raw reflectors. For narrowband systems ( $W < 1$ MHz), the delay resolution can be too coarse to resolve close scatterers.
Diffuse scattering: some channel models (3GPP TR 38.901 indoor hotspot) include a diffuse component beyond the discrete paths. On the DD grid this manifests as a low-level "background" across many cells, reducing effective sparsity.

For realistic terrestrial mobile channels (urban macro, vehicular), measurements (COST 2100, METIS) consistently find $P \leq 20$ with good cluster resolution — the regime where OTFS sparsity is decisive.

Practical Constraints

•
Discrete-path model valid for $P \leq 20$ typical
•
Resolution requires $W \geq 1/\Delta\tau_{\min}$ (bandwidth covers tap separation)
•
Diffuse component adds $\sim 5\%$ noise floor on DD cells

📋 Ref: 3GPP TR 38.901

Why This Matters: Embedded Pilot Design in Chapter 7

The pilot-overhead result of this section is the motivation for the embedded pilot scheme developed in Chapter 7. There we show (following the CommIT cell-free OTFS work) that a single pilot impulse with an appropriately sized guard region achieves the minimum overhead established here, with MSE matching the theoretical lower bound. The superimposed pilot approach — also treated in Chapter 7 — removes the guard region at the cost of additional receiver processing.

Sparsity and Its Consequences