Questions: Wiener Filter for Optimal Estimation

Question 1

At frequency ω₀, a signal has PSD S_ss(ω₀) = 100 and noise has PSD S_nn(ω₀) = 900 (signal and noise are uncorrelated). What is the Wiener filter gain H_opt(ω₀) at this frequency?

• A. 0.1 — the filter heavily suppresses this frequency because noise power is 9× greater than signal power
• B. 1.0 — the Wiener filter always passes all frequencies to avoid distorting the signal
• C. 0.5 — the filter treats signal and noise symmetrically by averaging them equally
• D. 10 — the filter amplifies the signal to overcome the noise floor

Answer: (A)

With uncorrelated signal and noise, H_opt(ω) = S_ss/(S_ss + S_nn) = 100/(100+900) = 0.1. The filter passes only 10% of input at this frequency because noise accounts for 90% of the total power there. This is frequency-by-frequency SNR weighting in action: where noise dominates, the filter suppresses heavily; where signal dominates, H → 1. Option 1 reflects a misconception about distortion — the goal is not to avoid distortion but to minimize mean-square error, which sometimes means heavily attenuating a frequency band.

Question 2

Why is the non-causal Wiener filter unsuitable for real-time signal processing applications?

• A. Its impulse response extends to negative times, meaning the filter output at time t requires future input values not yet available
• B. It amplifies noise at most frequencies, making it worse than a simple low-pass filter in practice
• C. It can only process signals whose noise is white (spectrally flat), limiting its applicability
• D. It requires the signal and noise to be perfectly uncorrelated, a condition that never holds in real systems

Answer: (A)

The non-causal Wiener filter's impulse response h(t) is nonzero for t < 0, so computing the output at time t requires input values at times t+τ for τ > 0 — future samples that are not yet available in real-time processing. This is a fundamental physical causality constraint, not a computational difficulty. For offline processing (recorded audio, medical images, seismic data), the non-causal solution is optimal and fully realizable. For real-time use, a causal approximation is required, at the cost of some MSE performance.

Question 3

The Wiener filter achieves the minimum possible mean-square error among all linear time-invariant filters for estimating a signal from noisy observations, given known signal and noise statistics.

• T. True
• F. False

Answer: True

The Wiener filter is derived by minimizing E[(s(t) − ŝ(t))²] over all LTI filters, using the orthogonality principle (the estimation error must be uncorrelated with the observation). The resulting filter is provably optimal within the class of linear estimators. The 'LTI' qualifier matters: nonlinear estimators may achieve lower MSE in some cases, but the Wiener filter is the best linear solution. This optimality is why it serves as the target that adaptive filters (LMS, RLS) converge toward — the Wiener solution is the fixed point of adaptation.

Question 4

A conventional low-pass filter with an optimally chosen cutoff frequency achieves essentially the same noise reduction performance as the Wiener filter for typical signals.

• T. True
• F. False

Answer: False

A low-pass filter applies a uniform cutoff: it passes frequencies below the cutoff equally and suppresses those above. The Wiener filter applies continuously varying frequency-specific weights based on the actual SNR at each frequency. When signal and noise occupy overlapping frequency ranges — common in speech, biomedical signals, and communications — the low-pass filter must either pass noise (if the cutoff is set too high) or remove signal (if set too low). The Wiener filter navigates this overlap optimally by applying a smooth, SNR-matched gain at every frequency point, achieving lower MSE than any fixed cutoff can.

Question 5

Explain how the Wiener filter differs from a conventional low-pass filter, and why this difference matters when signal and noise spectra overlap.

Model answer: A low-pass filter applies a binary-ish gain: approximately 1 below the cutoff and approximately 0 above, with one transition frequency chosen to balance signal preservation against noise rejection. The Wiener filter computes H_opt(ω) = S_ss(ω)/(S_ss(ω) + S_nn(ω)) independently at every frequency — the ratio of signal power to total power at that point. Where signal dominates (SNR >> 1), H → 1; where noise dominates (SNR << 1), H → 0. The gain varies smoothly and continuously, shaped to the spectral profiles of both signal and noise. When their spectra overlap, the Wiener filter fractionally passes each frequency in exact proportion to how much of the power there is signal, achieving minimum MSE without forcing a binary pass/reject decision.

The practical advantage is largest when signal and noise have overlapping spectra — for example, speech (dominated by frequencies below ~4 kHz) contaminated by colored noise that peaks in a similar range. A low-pass filter cannot simultaneously preserve the speech and reject the noise; the Wiener filter applies different gain to each frequency subband depending on local SNR, achieving a much better tradeoff. This same principle, generalized to time-varying statistics, underlies modern audio noise suppression, MRI reconstruction, and adaptive array processing.