Ferkans — Interactive Telecom Tutor

The Pilot Cost Is Quadratic in Antennas

On a block-fading channel, the transmitter does not know $\mathbf{h}_k$ a priori. To steer beams it must estimate them — and estimation costs channel uses. The standard protocol: each coherence block begins with a pilot phase of $\tau$ channel uses where users send known training symbols, followed by a data phase of $T_c - \tau$ channel uses where the transmitter uses the estimated channel to precode.

The catch: to estimate $K$ user channels each of dimension $L$ , one typically needs $\tau \geq L$ pilots. As $L$ grows — say in a massive MIMO regime — the pilot phase consumes a growing fraction of the coherence block. At $\tau = L \approx T_c$ , there is essentially no time left for data. This is the pilot wall: effective spatial DoF degrades as $L(1 - L/T_c)_+$ , peaking at $L = T_c/2$ .

Coded caching doesn't face this problem. Cache contents are pre-placed; the delivery phase uses a mix of cached bits and XOR messages that don't require real-time CSIT. The effective caching gain $t$ is pilot-free.

Theorem: Effective DoF with Pilot Overhead

For the cache-aided fading BC with coherence block $T_c$ , antennas $L$ , and pilot allocation $\tau = L$ (the minimum for full CSIT), the effective DoF per coherence block is $\mathrm{DoF}_{\text{eff}}(T_c) \;=\; t + L\left(1 - \frac{L}{T_c}\right)_+ .$ This DoF is maximized over $L$ by choosing $L^* = T_c/2$ , yielding spatial DoF $T_c/4$ .

Fraction $\tau/T_c = L/T_c$ of each coherence block is consumed by pilots; the remaining fraction $(1 - L/T_c)$ carries data. The spatial multiplexing gain $L$ is effective only during data. Coded caching uses all $T_c$ channel uses (placement is off-coherence-block; cached bits are always usable).

Proof

Data fraction

Per coherence block, data channel uses = $T_c - \tau = T_c - L$ (for $\tau = L$ ). Data fraction: $1 - L/T_c$ .

Spatial DoF during data

During data transmission, with perfect CSIT, $L$ streams can be zero-force-beamed. Instantaneous DoF = $L$ during data phase.

Time average

Average DoF = (data fraction) × (instantaneous DoF) = $(1 - L/T_c) \cdot L$ .

Add caching gain

Caching gain $t$ is available over the full block (pilot-free): $\mathrm{DoF}_{\text{eff}} = t + L(1 - L/T_c)_+$ , capped at $K$ .

Optimize $L$

$\partial_L[L(1 - L/T_c)] = 1 - 2L/T_c = 0$ at $L^* = T_c/2$ . Max spatial DoF = $T_c/4$ . Beyond this, more antennas hurt via pilot overhead. $\blacksquare$

,

Effective DoF vs Coherence Block Length

Plot the effective DoF as a function of the coherence block length $T_c$ , for fixed $K$ , $L$ , and memory ratio. Three curves: (1) blue cache + MIMO with pilot cost; (2) red dashed pure MIMO ( $t = 0$ ); (3) green dotted pure caching (CSIT-free). At small $T_c$ (high mobility, mmWave), the cache-aided curve approaches the green (CSIT-free) floor; at large $T_c$ (quasi-static), it approaches the full $t + L$ .

Parameters

Users K20

Antennas L4

Memory ratio M/N0.2

Example: mmWave vs Sub-6 GHz DoF

Compare the effective DoF for two 5G-NR-like scenarios: (a) Sub-6 GHz, $T_c = 1000$ , $L = 8$ , $K = 50$ , $\mu = 0.1$ . (b) mmWave, $T_c = 100$ , $L = 32$ , $K = 50$ , $\mu = 0.1$ . Both have the same caching gain $t = KM/N = 5$ .

Solution

(a) Sub-6 GHz

Pilot fraction $\tau/T_c = 8/1000 = 0.008$ . Effective spatial DoF: $L(1 - L/T_c) = 8(1 - 0.008) = 7.94$ . $\mathrm{DoF}_{\text{eff}} = t + 7.94 = 12.94$ .

(b) mmWave

Pilot fraction $\tau/T_c = 32/100 = 0.32$ . Effective spatial DoF: $L(1 - L/T_c) = 32 \cdot 0.68 = 21.76$ . $\mathrm{DoF}_{\text{eff}} = t + 21.76 = 26.76$ . But saturation is capped at $K = 50$ ; still within bound.

Per-user GDoF at different SNRs

Sub-6 GHz at 20 dB: per-user $\approx 12.94 \cdot \log 100 / 50 \approx 1.72$ bits/use. mmWave at 15 dB (lower SNR but much wider BW): per-user $\approx 26.76 \cdot \log 31.6/50 \approx 2.65$ bits/use.

Interpretation

mmWave's per-coherence-block spatial DoF is larger despite the pilot overhead, because $L$ is much larger. But caching gain is identical in both. Where caching makes the biggest relative impact is when pilot overhead dominates — e.g., $L = T_c$ , or very high mobility, where spatial DoF collapses.

Crossover

If we set $L = T_c/2$ (pilot-optimal), the spatial DoF peaks at $T_c/4$ . At $T_c = 100$ , max spatial DoF = 25. Adding caching gain $t = 5$ yields 30. Without caching, pure MIMO peaks at 25. The caching gain remains a +5 additive boost across all regimes.

Implication for Massive MIMO

The pilot-overhead analysis bears on the "more antennas = more gain" narrative of massive MIMO. For a fixed coherence block $T_c$ , adding antennas beyond $L^* = T_c/2$ hurts spatial DoF. This is a hard limit of TDD operation with pilot-based estimation.

But the picture changes when coded caching is added. Each extra user $K$ (with cache) contributes to the aggregate cache $KM/N$ and hence to the caching gain $t$ . This gain is not subject to pilot overhead. If the deployment is cache-rich, adding users can compensate for the pilot wall on the spatial side. This is a subtle design point: caching lets us decouple antenna count from CSIT overhead.

Common Mistake: Do Not Confuse Coherence Block with Coherence Time

Mistake:

Using $T_c$ in "channel uses" interchangeably with $T_{\text{coh}}$ in "seconds".

Correction:

$T_c$ (coherence block length) is measured in channel uses and equals $T_{\text{coh}} \cdot W_s$ where $T_{\text{coh}}$ is coherence time (seconds) and $W_s$ is the symbol rate (symbols per second). Similarly, $T_c$ can be interpreted as $B_c / \Delta f$ for wideband systems. The formulas of this chapter use $T_c$ in channel uses.

A 10 ms coherence time at 100 kBaud is $T_c = 10^3$ ; the same 10 ms at 10 MBaud is $T_c = 10^5$ . These are very different regimes for pilot-overhead analysis.

⚠️Engineering Note

Pilot Design in 5G NR

In 5G NR, pilot design is a nuanced tradeoff:

DMRS (Demodulation Reference Signal). User-specific pilots for coherent demodulation. Overhead: 1-2 OFDM symbols per slot.
SRS (Sounding Reference Signal). Uplink pilots for CSIT acquisition (TDD reciprocity). Periodic; 10-160 ms intervals.
CSI-RS. Downlink CSI measurement in FDD; feedback to BS via PUCCH.
Massive MIMO constraints. Pilot contamination (Marzetta 2010) when nearby cells reuse pilots; bounds per-user rate.

For cache-aided systems, pilot design must balance the usual MU-MIMO tradeoffs with the caching gain. A common design: reserve a smaller pilot allocation than the fully SU-optimal choice, trading a small spatial DoF loss for reduced overhead. The cache-aided Lampiris- Caire scheme tolerates this well because the caching component $t$ is pilot-insensitive.

Production 5G gNBs handle 4-8 DMRS ports per slot; mmWave mMIMO systems (e.g., 64+ ports) use hybrid beamforming to reduce effective pilot dimensionality.

Practical Constraints

•
5G NR DMRS: 1-2 OFDM symbols per 14-symbol slot (7-14% overhead)
•
Type II CSI feedback: up to 64 bits per reporting instance
•
SRS periodicity: 5-160 ms (vs coherence time of 1-10 ms at 100 km/h, 2 GHz)
•
Pilot contamination limits per-cell effective L to ~30 even with 100+ physical antennas

CSIT Acquisition and the Pilot Overhead Penalty