Ferkans — Interactive Telecom Tutor

Why Channel Estimation Matters in OFDM

The per-subcarrier equalisation $\hat{X}[k] = Y[k]/H[k]$ requires knowledge of the channel frequency response $H[k]$ at every data subcarrier. In practice, the channel is unknown and time-varying, so it must be estimated from known pilot symbols inserted into the OFDM time-frequency grid. The accuracy of channel estimation directly limits system performance.

Definition:
Pilot Placement Strategies

Pilots are known reference symbols $X_p[k]$ inserted at specific subcarrier-symbol positions in the OFDM time-frequency grid. The three main placement strategies are:

Block-type: Pilots occupy all subcarriers in certain OFDM symbols, with data-only symbols in between. Best suited for slow fading channels where the channel is approximately constant over several symbols.
Comb-type: Pilots are placed on evenly spaced subcarriers within every OFDM symbol, with data on the remaining subcarriers. Best suited for fast fading channels, as every symbol contains pilot information.
Scattered (lattice): Pilots are placed on a regular lattice in both time and frequency, with the pattern shifted across successive symbols. Provides tracking in both dimensions and is used in LTE and DVB-T.

The pilot density is governed by the sampling theorem: pilots must be spaced no further than $1/(2 f_{D,\max})$ in time and $1/(2 \tau_{\max})$ in frequency.

Definition:
Comb-Type Pilot Arrangement

In a comb-type arrangement, every $D_f$ -th subcarrier carries a pilot in every OFDM symbol. The pilot subcarrier indices are $\mathcal{P} = \{0, D_f, 2D_f, \ldots\}$ . The channel is estimated at pilot positions and then interpolated to data subcarrier positions.

The pilot spacing must satisfy the Nyquist condition:

$D_f \leq \frac{N}{\tau_{\max} / T_s} = \frac{N}{L}$

where $L$ is the number of resolvable channel taps.

Definition:
Block-Type Pilot Arrangement

In a block-type arrangement, certain OFDM symbols are all-pilot symbols (every subcarrier carries a known pilot). The channel is estimated at these pilot symbols and then interpolated in time to the intervening data symbols.

The pilot symbol spacing must satisfy:

$D_t \leq \frac{1}{2 f_{D,\max} T_{\text{total}}}$

where $f_{D,\max}$ is the maximum Doppler frequency and $T_{\text{total}}$ is the total OFDM symbol duration.

Theorem: Least-Squares Channel Estimation

At pilot positions $\mathcal{P}$ , the received signal is $Y_p[k] = H[k] X_p[k] + W[k]$ . The least-squares (LS) channel estimate at each pilot subcarrier is:

$\hat{H}_{\text{LS}}[k] = \frac{Y_p[k]}{X_p[k]}, \qquad k \in \mathcal{P}$

The LS estimator is unbiased with variance $\text{Var}(\hat{H}_{\text{LS}}[k]) = \sigma^2 / |X_p[k]|^2$ . For constant-modulus pilots ( $|X_p[k]|^2 = E_p$ ), this simplifies to $\sigma^2 / E_p$ .

The LS estimator simply divides out the known pilot and attributes whatever remains to the channel (plus noise). It requires no knowledge of channel statistics but is noise-enhanced.

Proof

Derivation

The observation at pilot $k$ is $Y_p[k] = H[k] X_p[k] + W[k]$ . The LS estimate minimises $|Y_p[k] - H[k] X_p[k]|^2$ over $H[k]$ , yielding $\hat{H}_{\text{LS}}[k] = Y_p[k] / X_p[k]$ .

Bias and variance

$\mathbb{E}[\hat{H}_{\text{LS}}[k]] = H[k] + \mathbb{E}[W[k]/X_p[k]] = H[k]$ (unbiased, since $X_p[k]$ is deterministic).

$\text{Var}(\hat{H}_{\text{LS}}[k]) = |1/X_p[k]|^2 \sigma^2 = \sigma^2/|X_p[k]|^2$ . $\blacksquare$

,

Theorem: MMSE Channel Estimation

The MMSE channel estimator exploits knowledge of the channel correlation to improve upon the LS estimate. Given the LS estimates $\hat{\mathbf{H}}_{\text{LS}}$ at pilot positions, the MMSE estimate of the full channel vector $\mathbf{H} = [H[0], \ldots, H[N-1]]^T$ is:

$\hat{\mathbf{H}}_{\text{MMSE}} = \mathbf{R}_{H\hat{H}} \mathbf{R}_{\hat{H}\hat{H}}^{-1} \hat{\mathbf{H}}_{\text{LS}}$

where:

$\mathbf{R}_{H\hat{H}} = \mathbb{E}[\mathbf{H} \hat{\mathbf{H}}_{\text{LS}}^H]$ is the cross-correlation
$\mathbf{R}_{\hat{H}\hat{H}} = \mathbf{R}_{HH} + (\sigma^2/E_p)\mathbf{I}$ is the auto-correlation of the LS estimates
$\mathbf{R}_{HH} = \mathbb{E}[\mathbf{H}\mathbf{H}^{H}]$ is the channel frequency correlation matrix

The MMSE estimator has lower MSE than LS but requires knowledge of $\mathbf{R}_{HH}$ and the noise variance $\sigma^2$ .

The MMSE estimator acts as a Wiener filter in the frequency domain, smoothing the noisy LS estimates by exploiting the fact that nearby subcarriers have correlated channel responses.

OFDM Channel Estimation

Compare LS and MMSE channel estimates on an OFDM system with comb-type pilots. Observe how the MMSE estimator produces smoother estimates by exploiting frequency-domain channel correlation. Increase the SNR to see both estimators converge to the true channel.

Parameters

Number of subcarriers

N

64

Pilot spacing

D_f

4

SNR (dB)15

Channel length

L

6

Estimator type

Example: LS Channel Estimation with Comb Pilots

An OFDM system has $N = 8$ subcarriers with comb-type pilots on subcarriers $k \in \{0, 2, 4, 6\}$ (pilot spacing $D_f = 2$ ). The pilot symbols are all $X_p[k] = 1$ . The received pilot values are:

$Y_p[0] = 0.8 - j0.3, \quad Y_p[2] = 0.6 + j0.5$ $Y_p[4] = -0.2 + j0.9, \quad Y_p[6] = 0.4 + j0.1$

(a) Compute the LS channel estimates at the pilot positions.

(b) Use linear interpolation to estimate $H[1]$ and $H[3]$ .

Solution

LS estimates at pilot positions

Since $X_p[k] = 1$ for all pilots:

$\hat{H}_{\text{LS}}[0] = Y_p[0] / 1 = 0.8 - j0.3$ $\hat{H}_{\text{LS}}[2] = Y_p[2] / 1 = 0.6 + j0.5$ $\hat{H}_{\text{LS}}[4] = Y_p[4] / 1 = -0.2 + j0.9$ $\hat{H}_{\text{LS}}[6] = Y_p[6] / 1 = 0.4 + j0.1$

Linear interpolation

For $H[1]$ (between pilots at $k=0$ and $k=2$ ):

$\hat{H}[1] = \frac{1}{2}(\hat{H}[0] + \hat{H}[2]) = \frac{1}{2}((0.8-j0.3)+(0.6+j0.5))$ $= 0.7 + j0.1$

For $H[3]$ (between pilots at $k=2$ and $k=4$ ):

$\hat{H}[3] = \frac{1}{2}(\hat{H}[2] + \hat{H}[4]) = \frac{1}{2}((0.6+j0.5)+(-0.2+j0.9))$ $= 0.2 + j0.7$

$\blacksquare$

LS vs. MMSE Channel Estimation

Property	LS Estimator	MMSE Estimator
Formula	$\hat{H}[k] = Y_p[k]/X_p[k]$	$\hat{\mathbf{H}} = \mathbf{R}_{H\hat{H}} \mathbf{R}_{\hat{H}\hat{H}}^{-1} \hat{\mathbf{H}}_{\text{LS}}$
Complexity	$O(N_p)$ — one division per pilot	$O(N_p^2)$ — matrix inversion required
Prior knowledge	None required	Channel correlation $\mathbf{R}_{HH}$ , noise variance $\sigma^2$
Bias	Unbiased	Biased (but lower MSE)
MSE performance	Higher MSE, especially at low SNR	Lower MSE by exploiting correlation
Robustness	Robust to model mismatch	Sensitive to incorrect $\mathbf{R}_{HH}$
Practical use	Initial estimate, low-complexity systems	Refined estimate in advanced receivers

Quick Check

In a comb-type pilot arrangement, what determines the maximum allowable pilot spacing $D_f$ in the frequency domain?

The maximum Doppler spread of the channel

The maximum delay spread (number of channel taps $L$ )

The number of subcarriers $N$

The modulation order of the data subcarriers

Correction:

The maximum delay spread (number of channel taps

L

)

By the sampling theorem, the pilot spacing must satisfy $D_f \leq N/L$ to avoid aliasing in the frequency-domain channel estimate. Larger delay spread requires denser pilots in frequency.

Common Mistake: Pilot Spacing Too Wide

Mistake:

Placing pilots too far apart in frequency ( $D_f > N/L$ ) to maximise data throughput, assuming interpolation will fill in the gaps.

Correction:

If $D_f > N/L$ , the channel frequency response is undersampled at the pilot positions, causing aliasing in the time-domain representation. No amount of interpolation can recover the lost information. The pilot spacing must satisfy the Nyquist condition $D_f \leq N/L$ . In practice, use $D_f$ slightly less than $N/L$ to account for leakage and windowing effects.

Pilot Symbol

A known reference symbol transmitted on a specific subcarrier and OFDM symbol position, used by the receiver to estimate the channel frequency response.

LS Channel Estimator

Least-squares estimator for the channel at pilot subcarriers: $\hat{H}[k] = Y_p[k]/X_p[k]$ . Simple and unbiased but noise-enhanced compared to MMSE.

Channel Estimation in OFDM