Ferkans — Interactive Telecom Tutor

The Pilot Overhead Problem

In narrowband massive MIMO (Chapters 3–4), we need $K$ orthogonal pilot sequences per coherence interval to estimate $\mathbf{H}$ . In wideband massive MIMO-OFDM, we must estimate $\mathbf{H}[k]$ at every subcarrier $k = 0, \ldots, N-1$ . If done naively — $K$ pilots at each of $N$ subcarriers — the overhead would consume most of the coherence block. The saving grace is the channel's finite delay spread: the $N$ frequency-domain channel vectors are determined by only $L \ll N$ delay-domain taps. This structure enables pilot designs that are dramatically more efficient than the naive approach.

Definition:
Pilot Subcarrier Grid

Let $\mathcal{P} \subset \{0, 1, \ldots, N-1\}$ denote the set of pilot subcarrier indices with $|\mathcal{P}| = N_p$ . At each pilot subcarrier $k \in \mathcal{P}$ , user $j$ transmits a known pilot symbol $\phi_j[k]$ . The pilot sequences must satisfy orthogonality:

$\sum_{k \in \mathcal{P}} \phi_i^*[k] \, \phi_j[k] = N_p \cdot \delta_{ij}$

for all user pairs $(i, j)$ , $i, j = 1, \ldots, K$ . This requires $N_p \geq K$ .

Common designs:

Comb-type pilots: $\mathcal{P} = \{0, D, 2D, \ldots\}$ with spacing $D = \lfloor N/N_p \rfloor$
Block-type pilots: One entire OFDM symbol dedicated to pilots
Scattered pilots: Staggered positions across time and frequency (used in 5G NR)

Pilot Overhead

The fraction of time-frequency resources consumed by pilot symbols rather than data. In a coherence block of $T_c / T_s$ OFDM symbols and $N$ subcarriers, if $N_p$ subcarriers in $\tau_p$ OFDM symbols carry pilots, the overhead is $\tau_p N_p / (T_c N / T_s)$ . Reducing pilot overhead while maintaining estimation quality is a central design challenge.

Theorem: Nyquist Pilot Spacing in Frequency

For a channel with $L$ delay taps and an OFDM system with $N$ subcarriers, the minimum number of uniformly spaced pilot subcarriers needed to perfectly reconstruct the frequency-domain channel (in the absence of noise) is

$N_p \geq L$

Equivalently, the maximum pilot spacing in frequency is $D \leq \lfloor N / L \rfloor$ . If $D > N/L$ , the channel cannot be uniquely recovered from the pilots (aliasing in the delay domain).

The channel frequency response is a trigonometric polynomial of degree $L-1$ (a sum of $L$ complex exponentials). By the Nyquist sampling theorem, $L$ uniformly spaced samples in frequency suffice to determine all $L$ coefficients uniquely.

Proof

Channel as trigonometric polynomial

The frequency response at subcarrier $k$ is

$\mathbf{h}_j[k] = \sum_{\ell=0}^{L-1} \mathbf{h}_{j,\ell} \, e^{-j 2\pi k \ell / N}$

This is the $N$ -point DFT of the zero-padded tap vector $[\mathbf{h}_{j,0}, \ldots, \mathbf{h}_{j,L-1}, \mathbf{0}, \ldots, \mathbf{0}]$ .

Sampling theorem in frequency

The tap vector has support on $\{0, 1, \ldots, L-1\}$ (the "bandwidth" in the delay domain is $L$ ). By the DFT sampling theorem, $L$ uniformly spaced frequency-domain samples uniquely determine the $L$ unknown taps.

Interpolation formula

Given the channel at pilot subcarriers $\{kD : k = 0, \ldots, N_p - 1\}$ with $N_p \geq L$ , the channel at any subcarrier $n$ is reconstructed via

$\mathbf{h}_j[n] = \sum_{m=0}^{N_p - 1} \mathbf{h}_j[mD] \cdot \frac{\sin(\pi(n - mD)N_p/N)}{N_p \sin(\pi(n-mD)/N)}$

This is the Dirichlet interpolation kernel. $\blacksquare$

,

Definition:
LS and MMSE Channel Estimation at Pilot Subcarriers

At pilot subcarrier $k \in \mathcal{P}$ , the base station receives (uplink training):

$\mathbf{y}[k] = \mathbf{H}[k] \, \boldsymbol{\Phi}[k] + \mathbf{w}[k]$

where $\boldsymbol{\Phi}[k] = \text{diag}(\phi_1[k], \ldots, \phi_{K}[k])$ is the diagonal pilot matrix and $\mathbf{w}[k] \sim \mathcal{CN}(\mathbf{0}, \sigma^2\mathbf{I})$ .

LS estimate: $\hat{\mathbf{H}}_{\text{LS}}[k] = \mathbf{y}[k] \, \boldsymbol{\Phi}^{-1}[k]$

MMSE estimate (assuming known channel covariance $\mathbf{R}_k = \mathbb{E}[\text{vec}(\mathbf{H}[k])\text{vec}(\mathbf{H}[k])^H]$ ):

$\hat{\mathbf{H}}_{\text{MMSE}}[k] = \mathbf{R}_k (\mathbf{R}_k + \sigma^2\mathbf{I})^{-1} \hat{\mathbf{H}}_{\text{LS}}[k]$

The MMSE estimator has lower MSE but requires knowledge of the second-order statistics.

,

Example: Pilot Overhead in a 5G NR Massive MIMO System

Consider a 5G NR system with: $N_t = 64$ , $K = 16$ , $N = 3276$ subcarriers, $\Delta f = 30\,\text{kHz}$ , channel delay spread $L = 31$ taps, and coherence time $T_c = 5\,\text{ms}$ (corresponding to $\sim 140$ OFDM symbols).

(a) What is the minimum pilot overhead using comb-type pilots?

(b) How does this compare to block-type pilots (one full OFDM symbol of pilots)?

(c) What fraction of the coherence block is available for data?

Solution

Comb-type pilot overhead

We need $N_p \geq \max(L, K) = \max(31, 16) = 31$ pilot subcarriers per user per OFDM symbol (since we need frequency-domain Nyquist sampling). With $K = 16$ users sharing pilots via orthogonal codes in time, we need $\tau_p = \lceil K / 1 \rceil = 16$ OFDM symbols for uplink training (one per user, sequentially) with $N_p = 31$ pilot subcarriers each.

Pilot overhead: $\frac{16 \times 31}{140 \times 3276} = \frac{496}{458{,}640} \approx 0.11\%$

This is extremely small — comb-type pilots are very efficient when $L \ll N$ .

Block-type pilot overhead

With block-type pilots, we dedicate $\tau_p = 16$ full OFDM symbols (one per user):

Pilot overhead: $\frac{16}{140} \approx 11.4\%$

Block-type pilots use all $N$ subcarriers per pilot symbol, which is wasteful when the channel has only $L = 31$ degrees of freedom in frequency.

Data fraction

With comb-type pilots: data fraction $\approx 99.9\%$ of resources. With block-type pilots: data fraction $\approx 88.6\%$ of resources.

The massive bandwidth inefficiency of block-type pilots motivates the comb/scattered designs used in 5G NR.

Definition:
Interpolation-Based Channel Estimation

Rather than estimating $\mathbf{H}[k]$ independently at every subcarrier, interpolation-based estimation proceeds in two steps:

Step 1 — Estimate at pilots: Obtain $\hat{\mathbf{H}}[k]$ at pilot subcarriers $k \in \mathcal{P}$ using LS or MMSE estimation.

Step 2 — Interpolate to data subcarriers: For $k \notin \mathcal{P}$ , reconstruct

$\hat{\mathbf{H}}[k] = \sum_{m \in \mathcal{P}} \hat{\mathbf{H}}[m] \cdot w_{m,k}$

where $\{w_{m,k}\}$ are interpolation weights. Common choices:

Linear interpolation: Uses the two nearest pilot subcarriers.
DFT-based interpolation: Transforms to the delay domain, truncates to $L$ taps, transforms back — equivalent to ideal sinc interpolation.
Wiener interpolation (MMSE): $w_{m,k}$ chosen to minimize the MSE, using the channel frequency correlation $r[\Delta k]$ .

Pilot Density vs. Channel Estimation MSE

Explore the tradeoff between pilot density (number of pilot subcarriers $N_p$ ) and the normalized channel estimation MSE, for LS, MMSE, and DFT-based interpolation.

Parameters

N_t

64

N

512

L

(delay taps)16

\text{SNR}

(dB)10

Estimation method

Interpolation Quality vs. Subcarrier Spacing

Visualize how different interpolation methods (linear, DFT-based, Wiener) reconstruct the channel between pilot subcarriers. Compare the true channel frequency response with the interpolated estimate.

Parameters

N

256

L

(delay taps)8

D

(pilot spacing)8

\text{SNR}

(dB)15

Interpolation method

Theorem: Optimal Wiener Interpolation for OFDM Channel Estimation

Given noisy LS channel estimates $\hat{\mathbf{H}}_{\text{LS}}[m]$ at pilot subcarriers $m \in \mathcal{P}$ , the MMSE interpolated estimate at data subcarrier $k \notin \mathcal{P}$ for user $j$ is

$\hat{\mathbf{h}}_j^{\text{MMSE}}[k] = \mathbf{r}_{k,\mathcal{P}}^H \bigl(\mathbf{R}_{\mathcal{P},\mathcal{P}} + \sigma^2\mathbf{I}\bigr)^{-1} \hat{\mathbf{h}}_j^{\text{LS}}[\mathcal{P}]$

where $[\mathbf{r}_{k,\mathcal{P}}]_m = r_j[k - m]$ is the cross-correlation vector between subcarrier $k$ and the pilot subcarriers, and $[\mathbf{R}_{\mathcal{P},\mathcal{P}}]_{m,n} = r_j[m - n]$ is the frequency correlation matrix among pilot subcarriers.

The resulting MSE is

$\text{MSE}_j[k] = r_j[0] - \mathbf{r}_{k,\mathcal{P}}^H (\mathbf{R}_{\mathcal{P},\mathcal{P}} + \sigma^2\mathbf{I})^{-1} \mathbf{r}_{k,\mathcal{P}}$

This is the Wiener filter applied to the frequency-domain interpolation problem. It exploits the known frequency correlation structure (determined by the power delay profile) to optimally combine the noisy pilot observations.

Proof

MMSE estimation setup

We want to estimate $\mathbf{h}_j[k]$ from the observation vector $\hat{\mathbf{h}}_j^{\text{LS}}[\mathcal{P}] = \mathbf{h}_j[\mathcal{P}] + \tilde{\mathbf{w}}[\mathcal{P}]$ where $\tilde{\mathbf{w}}[\mathcal{P}] \sim \mathcal{CN}(\mathbf{0}, \sigma^2\mathbf{I})$ .

Apply the LMMSE formula

The LMMSE estimate is $\hat{\mathbf{h}}_j[k] = \boldsymbol{\Sigma}_{h[k], \hat{h}[\mathcal{P}]} \boldsymbol{\Sigma}_{\hat{h}[\mathcal{P}]}^{-1} \hat{\mathbf{h}}_j^{\text{LS}}[\mathcal{P}]$ .

The cross-covariance is $\boldsymbol{\Sigma}_{h[k], \hat{h}[\mathcal{P}]} = \mathbf{r}_{k,\mathcal{P}}^H$ and the observation covariance is $\boldsymbol{\Sigma}_{\hat{h}[\mathcal{P}]} = \mathbf{R}_{\mathcal{P},\mathcal{P}} + \sigma^2\mathbf{I}$ .

MSE expression

The MSE follows from the standard LMMSE error formula: $\text{MSE} = r_j[0] - \mathbf{r}_{k,\mathcal{P}}^H (\mathbf{R}_{\mathcal{P},\mathcal{P}} + \sigma^2\mathbf{I})^{-1} \mathbf{r}_{k,\mathcal{P}}$ . $\blacksquare$

,

DFT-Based Channel Interpolation

Complexity:

O(N_p \log N_p + N \log N)

using FFT

Input: LS channel estimates

\hat{\mathbf{H}}_{\text{LS}}[k]

at pilot subcarriers

k \in \mathcal{P}

,

number of delay taps

L

, total subcarriers

N

Output: Interpolated channel

\hat{\mathbf{H}}[k]

for

k = 0, \ldots, N-1

1. Compute the

N_p

-point IDFT of the pilot-domain estimates:

\hat{\mathbf{g}}_j[\ell] = \frac{1}{N_p} \sum_{m=0}^{N_p-1} \hat{\mathbf{h}}_j^{\text{LS}}[m D] \, e^{j 2\pi m \ell / N_p}

,

\ell = 0, \ldots, N_p - 1

2. Truncate to the first

L

taps: set

\hat{\mathbf{g}}_j[\ell] = \mathbf{0}

for

\ell \geq L

3. Zero-pad to length

N

:

\tilde{\mathbf{g}}_j[\ell] = \hat{\mathbf{g}}_j[\ell]

for

\ell < L

,

\mathbf{0}

otherwise

4. Compute the

N

-point DFT:

\hat{\mathbf{h}}_j[k] = \sum_{\ell=0}^{L-1} \tilde{\mathbf{g}}_j[\ell] \, e^{-j 2\pi k \ell / N}

,

k = 0, \ldots, N-1

The truncation in step 2 acts as a low-pass filter in the delay domain, suppressing noise at delays beyond the channel support. This is the key advantage over simple linear interpolation.

Common Mistake: Pilot Contamination Is Worse in Wideband

Mistake:

Assuming that the pilot contamination analysis from Chapter 3 (narrowband) directly carries over to wideband MIMO-OFDM without modification.

Correction:

In wideband systems, pilot contamination occurs at each pilot subcarrier independently. If users in adjacent cells share the same pilot subcarrier positions, the contamination depends on the channel frequency response at those specific subcarriers. Frequency-domain pilot hopping — assigning different pilot positions to users in adjacent cells — can partially mitigate contamination. However, the total number of orthogonal pilot dimensions is $N_p \times \tau_p$ , which must accommodate all users in the cell and its neighbors.

Pilot Contamination

Interference caused when users in different cells transmit the same pilot sequences, causing the base station to estimate a superposition of desired and interfering channels. In massive MIMO, pilot contamination is the fundamental performance-limiting factor that does not vanish with increasing $N_t$ .

Related: Pilot Overhead

Why This Matters: Connection to OFDM in Telecom Book

The OFDM system model and time-frequency resource grid used here are developed in detail in Book 1 (Telecom), Chapter 24. There, the focus is on single-user and small-MIMO OFDM. This chapter extends the treatment to massive MIMO, where the spatial dimension introduces both new opportunities (per-subcarrier beamforming, spatial multiplexing gain) and new challenges (per-subcarrier CSI estimation, computational complexity scaling with $N_t \times N$ ).

Historical Note: Evolution of Pilot Design for MIMO-OFDM

2004–2018

Early MIMO-OFDM systems (IEEE 802.11n, LTE) used block-type pilots: entire OFDM symbols dedicated to training. This was acceptable because $N_t \leq 8$ and bandwidths were modest (20–40 MHz). With massive MIMO and bandwidths up to 400 MHz, Marzetta's 2010 paper showed that pilot overhead must scale with $K$ , not $N_t$ — TDD reciprocity is the key. The comb-type and scattered pilot designs in 5G NR (Release 15) were specifically designed for massive MIMO, exploiting the finite delay spread to minimize overhead.

,

⚠️Engineering Note

Pilot Power Boosting in 5G NR

In 5G NR, the SRS (Sounding Reference Signal) used for uplink channel estimation can be power-boosted relative to data symbols to improve estimation SNR. The specification allows up to 3 dB of power boosting for SRS. However, excessive boosting creates near-far problems with adjacent-cell users. The pilot power must be jointly optimized with the pilot density and the target estimation MSE.

Practical Constraints

•
SRS power boosting limited to 3 dB in 5G NR (3GPP TS 38.211)
•
SRS bandwidth can be configured from 4 to 272 resource blocks
•
SRS periodicity ranges from 1 slot to 2560 slots

📋 Ref: 3GPP TS 38.211, Section 6.4.1.4

Quick Check

A channel has $L = 20$ delay taps and the OFDM system uses $N = 1024$ subcarriers. What is the maximum pilot spacing $D$ that avoids frequency-domain aliasing?

$D = 20$

$D = 51$

$D = 1024$

$D = 5$

Correction:

D = 51

Correct: $D = \lfloor N/L \rfloor = \lfloor 1024/20 \rfloor = 51$ . This gives exactly $\lceil N/D \rceil = 21 \geq L$ pilot subcarriers.

Pilot Design and Channel Estimation

The Pilot Overhead Problem

Definition: Pilot Subcarrier Grid

Pilot Overhead

Theorem: Nyquist Pilot Spacing in Frequency

Channel as trigonometric polynomial

Sampling theorem in frequency

Interpolation formula

Definition: LS and MMSE Channel Estimation at Pilot Subcarriers

Example: Pilot Overhead in a 5G NR Massive MIMO System

Comb-type pilot overhead

Block-type pilot overhead

Data fraction

Definition: Interpolation-Based Channel Estimation

Pilot Density vs. Channel Estimation MSE

Parameters

Interpolation Quality vs. Subcarrier Spacing

Parameters

Theorem: Optimal Wiener Interpolation for OFDM Channel Estimation

MMSE estimation setup

Apply the LMMSE formula

MSE expression

DFT-Based Channel Interpolation

Common Mistake: Pilot Contamination Is Worse in Wideband

Pilot Contamination

Why This Matters: Connection to OFDM in Telecom Book

Historical Note: Evolution of Pilot Design for MIMO-OFDM

Pilot Power Boosting in 5G NR

Quick Check

Definition:
Pilot Subcarrier Grid

Definition:
LS and MMSE Channel Estimation at Pilot Subcarriers

Definition:
Interpolation-Based Channel Estimation