Signals: Matched Filtering for Detection

Detecting a known signal buried in noise is a fundamental challenge in communications, radar, and sonar. Matched filtering is the optimal linear filter for maximizing the signal-to-noise ratio (SNR) at a specific instant in time, making it the cornerstone of reliable detection systems. Understanding it moves you from simply processing signals to strategically designing receivers that can pull faint, predictable waveforms out of the chaos of random interference.

The Core Principle: Maximizing Signal-to-Noise Ratio

The primary goal of a matched filter is to give you the best possible chance of deciding whether a known signal is present in a noisy observation. It doesn't necessarily make the output "look like" the input signal; instead, it optimizes for a single, critical metric: the peak instantaneous signal-to-noise ratio (SNR). Imagine trying to hear a specific whistle in a crowded, noisy room. Your brain effectively becomes "matched" to that whistle's pattern, amplifying it relative to the background babble. The matched filter does this mathematically.

Formally, for a known, deterministic signal $s(t)$ embedded in additive white noise, the matched filter's impulse response $h(t)$ is proportional to the time-reversed, conjugated version of the signal. If the signal ends at time $T$, the filter is $h(t)=k{s}^{\ast }(T-t)$, where $k$ is a constant and $\ast$ denotes complex conjugation. When the received signal $r(t)=s(t)+n(t)$ passes through this filter, the output peaks at time $t=T$. At that precise moment, the ratio of the signal power to the average noise power at the filter's output is maximized. This peak SNR is given by $2E/N_0$, where $E$ is the energy of the signal $s(t)$ and $N_0/2$ is the two-sided power spectral density of the white noise. Notice the result depends only on the signal's energy, not its shape—a longer or higher-amplitude pulse provides more energy and a better SNR.

Derivation from the Schwarz Inequality

Why is the time-reversed conjugate the optimal form? This can be proven rigorously using the Schwarz inequality (also called the Cauchy-Schwarz inequality). This inequality states that for any two complex-valued functions $x(t)$ and $y(t)$, the following holds:

$${\left|\int x(t)y^{\ast }(t)dt\right|}^2\le \int |x(t){|}^{2}dt\cdot \int |y(t){|}^{2}dt$$

Equality is achieved if and only if $x(t)=c\cdot y(t)$, where $c$ is a constant. In our detection problem, we model the filter output at the sampling time $t=T$ as $y(T)=\int r(\tau )h(T-\tau )d\tau$. We want to maximize the SNR at $T$, which is the ratio of the squared magnitude of the signal component to the variance of the noise component. By assuming the noise is white with power spectral density $N_0/2$, the noise variance at the output is $(N_0/2)\int |h(t){|}^{2}dt$. The signal component, if present, is $\int s(\tau )h(T-\tau )d\tau$.

Setting $x(\tau )=s(\tau )$ and $y(\tau )=h(T-\tau )$ in the Schwarz inequality leads directly to the conclusion that SNR is maximized when $h(T-\tau )$ is proportional to $s^{\ast }(\tau )$, or equivalently, $h(t)\propto s^{\ast }(T-t)$. Plugging this condition back in yields the maximum possible SNR: $SN{R}_{max}=(2/N_0)\int |s(t){|}^{2}dt=2E/N_0$. This derivation is powerful because it doesn't just give you the answer; it shows why no other linear filter can do better.

Implementing a Pulse Detection System

Let's walk through a concrete example of implementing a matched filter for a simple rectangular pulse. Suppose the known signal $s(t)$ is a unit-amplitude pulse of duration $T$ seconds: $s(t)=1$ for $0\le t\le T$, and zero elsewhere. Its matched filter impulse response is $h(t)=s(T-t)$, which is also a rectangular pulse from $0$ to $T$. This filter is essentially a moving integrator.

The operation of the filter, $y(t)=\int r(\tau )h(t-\tau )d\tau$, becomes $y(t)={\int }_{t-T}^{t}r(\tau )d\tau$ for $t\ge T$. The filter sums (integrates) the input over the last $T$ seconds. When the signal $s(t)$ is perfectly aligned in the integrator, the output rises linearly to a maximum value equal to the pulse energy $E=T$ at time $t=T$. The noise, being zero-mean, integrates toward a much smaller value on average. You would sample the filter's output at $t=T$ and compare this value to a pre-set threshold to decide "signal present" (if $y(T)>threshold$) or "noise only" (if $y(T)\le threshold$). This system forms the basic heart of a digital receiver for a known bit sequence.

Analyzing Detection Performance: Probability vs. False Alarm

The matched filter gives you the best possible data (the peak SNR) to feed into your detector, but you must still make a decision. This leads to the critical trade-off between probability of detection ($P_D$) and probability of false alarm ($P_{FA}$). $P_D$ is the chance you correctly say "signal present" when it is. $P_{FA}$ is the chance you incorrectly say "signal present" when only noise exists.

After matched filtering, the sampled output is a random variable. Under the common assumption of additive white Gaussian noise, this output is Gaussian distributed. If only noise is present, the sample has a mean of zero and a certain variance. If signal-plus-noise is present, the mean shifts to the signal's amplitude (related to $\sqrt{2E/N_0}$). You set a threshold $\gamma$: samples above $\gamma$ are declared detections.

• Raising $\gamma$ makes the detector more conservative: $P_{FA}$ drops, but so does $P_D$.
• Lowering $\gamma$ makes it more sensitive: $P_D$ rises, but so does P_{FA.

For Gaussian noise, these probabilities are related to the Q-function (the tail probability of the standard normal distribution). If $d=\sqrt{2E/N_0}$ is the peak voltage SNR, then $P_D=Q\left(Q^{-1}(P_{FA})-d\right)$. This equation quantifies the trade-off: for a fixed acceptable $P_{FA}$, your $P_D$ improves dramatically as the SNR $d$ increases. The matched filter's role is to maximize $d$, thereby pushing this performance curve to its absolute limit for a given signal energy.

Application to Radar Systems

Radar is a quintessential application of matched filtering. A radar transmitter sends out a known pulse, often a chirp or coded sequence, and listens for its reflection from a target. The received echo is an attenuated, time-delayed, and possibly Doppler-shifted copy of the transmitted pulse, buried in noise. The matched filter is implemented in the receiver, tuned to the transmitted pulse shape.

When the reflected pulse passes through its matched filter, it produces a sharp peak at the output. The time delay of this peak relative to the transmission time directly gives the target's range: $range=c\cdot \Delta t/2$, where $c$ is the speed of light. The matched filter provides two key benefits here: 1) It maximizes the SNR of the received echo, allowing detection of fainter, more distant targets. 2) For pulses with a broad bandwidth (like chirps), the matched filter output pulse is much narrower than the transmitted pulse. This "pulse compression" improves range resolution, allowing the radar to distinguish between two closely spaced targets. The entire process elegantly solves the dual problem of detection and precise parameter estimation.

Common Pitfalls

1. Ignoring the Noise Assumption: The matched filter is optimal for additive white Gaussian noise (AWGN). If the noise is colored (non-white), you must first whiten the noise with a pre-whitening filter, then apply a filter matched to the whitened version of your signal. Applying a standard matched filter in colored noise severely degrades performance.
2. Mismatching the Signal: The filter must be matched to the exact expected signal, including its shape, timing, and carrier phase (for passband signals). A mismatch—such as using a filter for a rectangular pulse when the received pulse is rounded—results in a loss in SNR. In radar, if the target is moving, the Doppler shift can cause a mismatch, requiring a bank of filters matched to different frequency shifts.
3. Confusing the Output Shape: Students often expect the filter output to resemble the input signal. It does not. For example, a rectangular pulse input yields a triangular output from its matched filter (the integrator). The important feature is the timing and amplitude of the peak, not the overall waveform.
4. Misunderstanding the Threshold: The matched filter optimizes the SNR but does not set the decision threshold. Choosing the threshold requires knowledge of the desired $P_{FA}$ and the noise statistics. Setting it arbitrarily, without analyzing the $P_D$/$P_{FA}$ trade-off, can render an optimal filter useless in practice.

Summary

• A matched filter is the optimal linear filter for maximizing the output signal-to-noise ratio at a specific time for a known signal in additive white noise. Its impulse response is the time-reversed, complex-conjugated version of the expected signal.
• The optimality is proven using the Schwarz inequality, which leads to the maximum achievable SNR of $2E/N_0$, dependent only on the signal energy $E$ and noise power density $N_0$.
• In implementation, such as for pulse detection, the filter output is sampled and compared to a threshold. The classic example is using an integrator matched to a rectangular pulse.
• System performance is characterized by the trade-off between probability of detection ($P_D$) and probability of false alarm ($P_{FA}$), a relationship formalized using the Q-function for Gaussian noise. The matched filter's high output SNR makes favorable trade-offs possible.
• Radar systems heavily rely on matched filtering not only for superior detection capability but also for accurate range estimation and resolution through the pulse compression effect.