FIR Filter Window Design & Spectral Leakage Calculator

1. The Fundamentals of Finite Impulse Response (FIR) Filter Synthesis

A Finite Impulse Response (FIR) filter is a discrete-time Linear Time-Invariant (LTI) system whose impulse response is nonzero over only a finite duration of $M$ samples. The time-domain output $y[n]$ is computed via direct discrete convolution between the input signal $x[n]$ and the impulse response coefficients $h[n]$:

The corresponding Z-domain system transfer function $H(z)$ contains only numerator polynomial terms:

Because $H(z)$ has no denominator polynomial (other than the trivial pole $z^{M-1}$ at the origin), FIR filters are inherently and unconditionally stable under all operational conditions. Round-off noise or extreme coefficient quantization can never cause parasitic limit-cycle oscillations or catastrophic pole migration into the unstable region ($|z| ge 1.0$).

2. Linear Phase Properties and Filter Types I through IV

The crowning advantage of FIR filters over their IIR counterparts is the ability to achieve strictly linear phase across the entire frequency spectrum. A filter possesses linear phase if its phase response satisfies $theta(omega) = -alpha omega$, which translates to a perfectly constant group delay:

Constant group delay guarantees that all spectral components of a complex waveform propagate through the filter with the exact same time delay, preserving waveform shape without phase distortion. This property is crucial in high-speed digital communications (QAM/OFDM pulse shaping), medical ECG diagnostics, and professional audio mastering.

Linear-phase FIR filters are categorized into four canonical symmetry classes:

• Type I: Odd length $M$ (even order $N = M-1$), symmetric impulse response ($h[n] = h[M-1-n]$). Suitable for all filter responses: Lowpass, Highpass, Bandpass, and Bandstop.
• Type II: Even length $M$, symmetric impulse response. Inherently forces $H(e^{jpi}) = 0$ at the Nyquist frequency, making it unsuitable for highpass or bandstop designs.
• Type III: Odd length $M$, anti-symmetric impulse response ($h[n] = -h[M-1-n]$). Forces $H(e^{j0}) = H(e^{jpi}) = 0$, commonly used for Hilbert transformers and differentiators.
• Type IV: Even length $M$, anti-symmetric impulse response. Forces $H(e^{j0}) = 0$.

3. The Window Design Method & Gibbs Phenomenon

An ideal lowpass brick-wall filter possesses a rectangular frequency response with sharp vertical transitions:

Applying the inverse discrete-time Fourier transform (IDTFT) yields the non-causal, infinitely long ideal sinc impulse response:

Abruptly truncating $h_d[n]$ to $M$ samples (multiplying by a rectangular window) results in the classic Gibbs phenomenon: spectral oscillations with approximately 9% overshoot ($+0.75text{ dB}$) at the discontinuity and a fixed peak side-lobe level of $-13.3text{ dB}$. Increasing the filter length $M$ narrows the oscillation width but never reduces the peak ripple amplitude.

To suppress Gibbs ringing and maximize stopband rejection, the infinite sinc sequence is multiplied by a smooth, tapered window function $w[n]$:

4. Comparative Analysis of Window Functions

Tapering reduces the discontinuity at the window edges, which suppresses spectral side-lobe leakage at the cost of broadening the central main lobe (widening the filter transition band):

• Rectangular: $w[n] = 1$. Peak sidelobe: $-13text{ dB}$. Main lobe width: $4pi / M$. Transition width $Delta f approx 0.9 / M cdot f_s$. Sharpest transition but poorest stopband attenuation.
• Hanning (Hann): $w[n] = 0.5 - 0.5 cos(2pi n / (M-1))$. Peak sidelobe: $-31text{ dB}$. Main lobe width: $8pi / M$. Side-lobe decay rate: $-18text{ dB/octave}$.
• Hamming: $w[n] = 0.54 - 0.46 cos(2pi n / (M-1))$. Optimizes the coefficients to cancel the first side-lobe. Peak sidelobe: $-43text{ dB}$. Stopband attenuation: $approx 53text{ dB}$. Side-lobe decay rate: $-6text{ dB/octave}$. Widely used in speech processing.
• Blackman: $w[n] = 0.42 - 0.5 cos(2pi n / (M-1)) + 0.08 cos(4pi n / (M-1))$. Peak sidelobe: $-58text{ dB}$. Stopband attenuation: $approx 74text{ dB}$. Main lobe width: $12pi / M$.
• Kaiser Window: Based on the modified zero-order Bessel function of the first kind $I_0(x)$:
  w[n] = frac{I_0left(beta sqrt{1 - left(frac{2n}{M-1} - 1right)^2}right)}{I_0(beta)}
  The parameter $beta$ allows continuous adjustment between spectral resolution and side-lobe suppression.

5. Step-by-Step Worked Numerical Calculation: 5-Tap Lowpass Filter

Let us manually calculate the tap weights for a 5-tap ($M=5$, $alpha=2$) lowpass FIR filter with sampling rate $f_s = 8{,}000text{ Hz}$ and cutoff $f_c = 2{,}000text{ Hz}$ using a Hamming window:

1. Normalized Cutoff Frequency: $$omega_c = 2pi frac{2000}{8000} = 0.5pitext{ rad/sample}$$
2. Compute Ideal Sinc Values $h_d[n]$: For $n = 2$ (center tap): $$h_d[2] = frac{omega_c}{pi} = frac{0.5pi}{pi} = 0.5000$$ For $n = 1$ and $n = 3$ ($|n-2| = 1$): $$h_d[1] = h_d[3] = frac{sin(0.5pi cdot (-1))}{pi (-1)} = frac{-1}{-pi} = frac{1}{pi} approx 0.31831$$ For $n = 0$ and $n = 4$ ($|n-2| = 2$): $$h_d[0] = h_d[4] = frac{sin(0.5pi cdot (-2))}{pi (-2)} = frac{0}{-2pi} = 0.0000$$
3. Compute Hamming Window Values $w[n]$: $$w[n] = 0.54 - 0.46 cosleft(frac{2pi n}{4}right)$$ $$w[0] = w[4] = 0.54 - 0.46 cos(0) = 0.0800$$ $$w[1] = w[3] = 0.54 - 0.46 cos(pi/2) = 0.5400$$ $$w[2] = 0.54 - 0.46 cos(pi) = 0.54 - (-0.46) = 1.0000$$
4. Multiply Window by Sinc: $h[n] = h_d[n] cdot w[n]$: $$h[0] = 0.0000 times 0.0800 = 0.0000$$ $$h[1] = 0.31831 times 0.5400 = 0.17189$$ $$h[2] = 0.50000 times 1.0000 = 0.50000$$ $$h[3] = 0.31831 times 0.5400 = 0.17189$$ $$h[4] = 0.0000 times 0.0800 = 0.0000$$
5. Normalize DC Gain to Unity: $$sum_{n=0}^4 h[n] = 0 + 0.17189 + 0.5 + 0.17189 + 0 = 0.84378$$ Dividing by $0.84378$ gives normalized coefficients: $$h_{norm} = [0.0000, 0.2037, 0.5926, 0.2037, 0.0000]$$

6. Computational Architecture & Optimization in Hardware/Software

In embedded digital signal processors (ARM Cortex-M with CMSIS-DSP, TI C6000) and FPGAs, implementing FIR convolution efficiently relies on several structural properties:

• Symmetric Folded Architecture: Because linear-phase Type I filters exhibit even symmetry ($h[n] = h[M-1-n]$), symmetric input samples can be pre-added before multiplication: $$y[n] = h[alpha] x[n-alpha] + sum_{k=0}^{alpha-1} h[k] (x[n-k] + x[n-(M-1-k)])$$ This cuts the required hardware multipliers exactly in half.
• Circular Buffering: To avoid shifting $M$ input samples in memory for every new incoming data sample, software implementations utilize power-of-two circular ring buffers with modulo pointer arithmetic.
• Fast Convolution via FFT (Overlap-Add / Overlap-Save): When the filter length exceeds approximately $M ge 64$ to $128$ taps, direct time-domain convolution ($O(M)$ operations per output sample) becomes computationally inferior to frequency-domain block convolution using the Fast Fourier Transform ($O(log_2 M)$ per sample).