Fourier Transform Interview Questions

Q1. What is the Fourier transform and what information does it provide about a signal?

The Fourier transform decomposes a time-domain signal into its constituent sinusoidal frequency components, producing a complex-valued frequency-domain spectrum X(jω) = ∫x(t)e^(−jωt)dt that contains both amplitude and phase information at each frequency. Applying the Fourier transform to a 440 Hz tuning fork recording reveals a dominant peak at 440 Hz with harmonics at 880 Hz and 1320 Hz whose relative amplitudes characterize the instrument's timbre. The spectrum shows which frequencies carry signal energy and at what phase, enabling filter design, spectral analysis, and modulation analysis.

Follow-up: What information is lost when you look only at the magnitude spectrum |X(jω)| and discard the phase spectrum?

Q2. What is the difference between the Fourier Series and the Fourier Transform?

The Fourier Series represents a periodic signal as a discrete sum of harmonically related complex exponentials with coefficients computed over one period, while the Fourier Transform represents a non-periodic signal as a continuous integral over all frequencies. A 50 Hz square wave from a function generator has a Fourier Series with odd harmonics at 50, 150, 250 Hz..., whereas a single square pulse has a Fourier Transform giving a continuous sinc-shaped spectrum. As the period of a periodic signal approaches infinity, the Fourier Series coefficient envelope converges to the Fourier Transform, connecting both tools.

Follow-up: What is the relationship between the Fourier Series coefficients and the Fourier Transform of one period of the signal?

Q3. State and explain the convolution theorem of the Fourier transform.

The convolution theorem states that convolution in the time domain corresponds to multiplication in the frequency domain: if y(t) = x(t) * h(t), then Y(jω) = X(jω) × H(jω). Filtering a noisy sensor signal by convolving with a 100-tap FIR filter is equivalent to multiplying the signal's FFT by the filter's frequency response in the frequency domain. This theorem makes frequency-domain filter design and analysis far more intuitive than time-domain convolution and justifies the overlap-add/overlap-save fast convolution algorithms used in audio DSPs.

Follow-up: What is the dual of the convolution theorem — what does multiplication in the time domain correspond to in the frequency domain?

Q4. What is the DFT and how does it relate to the continuous Fourier transform?

The Discrete Fourier Transform (DFT) computes the frequency spectrum of a finite-length discrete-time sequence x[n] of N samples, producing N complex spectral values X[k] at frequencies k×(fs/N), where fs is the sampling rate. A 1024-point DFT of audio sampled at 44.1 kHz computes spectral values at 44100/1024 ≈ 43 Hz intervals, from 0 to 44100 Hz. The DFT assumes the N-sample block is one period of an infinitely periodic signal, which introduces spectral leakage when the signal frequency does not fall exactly on a DFT bin boundary.

Follow-up: Why does the DFT spectrum of a pure sine wave show energy spread across multiple bins when the frequency is between two bins?

Q5. What is the FFT and how does it reduce the computational complexity of the DFT?

The Fast Fourier Transform (FFT) is an efficient algorithm that computes the DFT in O(N log N) operations instead of the O(N²) operations of direct DFT computation, by recursively splitting the N-point DFT into smaller DFTs and reusing intermediate results. Computing a 1024-point DFT directly requires about 1,048,576 multiplications, while the Cooley-Tukey radix-2 FFT requires only about 5,120 complex multiplications. The ARM CMSIS-DSP library's arm_cfft_f32 function implements the Cooley-Tukey FFT and is used in Cortex-M4-based audio spectrum analyzers in real-time.

Follow-up: What constraint on N does the radix-2 Cooley-Tukey FFT algorithm impose?

Q6. What is spectral leakage and how does windowing reduce it?

Spectral leakage occurs when the DFT assumes the N-sample block repeats periodically and the signal's frequency does not align exactly with a DFT bin, causing the true spectral peak to spread energy across many adjacent bins with side lobes. Multiplying the input signal by a Hann window (w[n] = 0.5 − 0.5×cos(2πn/N)) before the FFT smoothly tapers the block edges to zero, reducing the side lobe amplitude from −13 dB (rectangular window) to −31 dB. In an STM32-based vibration spectrum analyzer, a Hann window is applied to each 512-sample frame before FFT to prevent leakage from masking small fault-frequency peaks near large structural resonance peaks.

Follow-up: What is the tradeoff between main lobe width and side lobe level when choosing between a Hann and a Blackman window?

Q7. What is the time-shifting property of the Fourier transform?

A time delay of t0 in the time domain multiplies the Fourier transform by the complex exponential e^(−jωt0): FT{x(t − t0)} = e^(−jωt0) × X(jω). This means a pure time shift changes the phase spectrum linearly with frequency but leaves the magnitude spectrum completely unchanged. In radar signal processing, the time delay between transmitted and received pulses appears as a linear phase slope in the FFT of the received signal, from which target range is calculated as R = c × t0/2.

Follow-up: What does a linear phase response of a filter mean in terms of its effect on the signal waveform?

Q8. What is the Fourier transform of a rect function and why is it important in communications?

The Fourier transform of a rectangular pulse of width T is a sinc function: X(f) = T × sinc(fT) = T × sin(πfT)/(πfT), with the first zero at f = 1/T. A NRZ data pulse of duration 1/9600 seconds (for a 9600 baud UART signal) has its first spectral null at 9600 Hz, which determines the minimum RF bandwidth required to transmit UART data without intersymbol interference. The sinc spectrum of rectangular pulses is why practical communication systems use raised cosine pulse shaping — to limit bandwidth while controlling intersymbol interference at the receiver.

Follow-up: What is the relationship between pulse width and bandwidth, and how does it illustrate the time-bandwidth product principle?

Q9. What is Parseval's theorem and what does it state about signal energy?

Parseval's theorem states that the total energy of a signal computed in the time domain equals the total energy computed from the frequency domain: ∫|x(t)|²dt = (1/2π)∫|X(jω)|²dω. This means energy is conserved under the Fourier transform, and |X(jω)|²/(2π) is the energy spectral density describing how signal energy is distributed across frequency. In audio mastering, the energy distribution from Parseval's theorem is used to verify that an equalizer preserves total signal loudness while reshaping the spectral balance.

Follow-up: What is the difference between energy spectral density and power spectral density, and to which class of signals does each apply?

Q10. What is the frequency-shifting property (modulation theorem) of the Fourier transform?

Multiplying a signal x(t) by a complex exponential e^(jω0t) shifts its Fourier transform by ω0: FT{x(t)e^(jω0t)} = X(j(ω − ω0)). For a real signal, multiplying by cos(ω0t) creates two copies of X(jω) shifted to ±ω0, which is exactly the amplitude modulation (AM) process that moves an audio baseband signal to an RF carrier. This property explains why a mixer in a superheterodyne receiver multiplies the RF signal by the local oscillator to shift the desired channel to the intermediate frequency (IF).

Follow-up: How does the frequency-shifting property explain the creation of upper and lower sidebands in AM modulation?

Q11. What is the Fourier transform of an impulse train (comb function)?

The Fourier transform of an impulse train with period T0, Σδ(t − nT0), is another impulse train in frequency with period 1/T0 and amplitudes scaled by 1/T0. This result directly proves the sampling theorem: sampling (multiplying by an impulse train at fs) creates copies of the signal spectrum centered at multiples of fs. In a 44.1 kHz audio ADC, the sampling impulse train's spectrum has copies at 0, ±44.1 kHz, ±88.2 kHz, and the anti-aliasing filter must remove all signal content above 22.05 kHz before sampling to prevent these copies from overlapping.

Follow-up: What is the mathematical connection between the Fourier transform of a sampled signal and aliasing?

Q12. What is the 2D Fourier transform and where is it used?

The 2D Fourier transform extends the 1D concept to functions of two spatial variables f(x, y), decomposing an image into 2D spatial frequency components where low frequencies represent coarse structures and high frequencies represent fine detail and edges. Applying the 2D FFT to a 512×512 grayscale image produces a complex spectrum whose magnitude plot shows the dominant spatial frequencies of the image texture. JPEG compression uses the discrete cosine transform (DCT), a variant of the 2D Fourier transform, to convert 8×8 pixel blocks into frequency coefficients that are then quantized to achieve compression.

Follow-up: Why do sharp edges in an image produce high-amplitude high-frequency components in the 2D Fourier spectrum?

Q13. What is zero-padding in DFT and what does it do to the frequency resolution?

Zero-padding appends zeros to a finite-length signal sequence before computing the DFT, increasing the number of DFT points N and therefore the density of the spectral output samples without collecting additional data. Padding a 256-sample signal with 768 zeros to compute a 1024-point DFT produces spectral samples at 44100/1024 ≈ 43 Hz intervals instead of 44100/256 ≈ 172 Hz intervals, giving a smoother-looking spectrum. Zero-padding interpolates the spectrum but does not improve true frequency resolution, which is determined solely by the observation window length (number of actual data samples).

Follow-up: If zero-padding does not improve frequency resolution, why is it routinely used in practice?

Q14. What is the relationship between the Fourier transform and the Laplace transform?

The Fourier transform is a special case of the bilateral Laplace transform evaluated on the imaginary axis: X(jω) = X(s)|{s=jω}, valid only when the imaginary axis (jω axis) lies within the region of convergence of the Laplace transform. For a causal decaying exponential x(t) = e^(−at)u(t) with a > 0, the Laplace transform X(s) = 1/(s+a) has ROC Re{s} > −a, which includes the jω axis, so its Fourier transform X(jω) = 1/(jω+a) is obtained by direct substitution. For systems with poles in the right half-plane, the Fourier transform does not exist but the Laplace transform does, making Laplace analysis more general for unstable or marginally stable systems.

Follow-up: Why does the Fourier transform not exist for the growing exponential x(t) = e^(at)u(t) with a > 0?

Q15. What is the short-time Fourier transform (STFT) and what problem does it solve?

The STFT computes the DFT over a sliding windowed segment of a signal, producing a 2D time-frequency representation (spectrogram) that shows how the signal's frequency content changes over time — something the standard Fourier transform cannot reveal. A 25 ms Hann-windowed STFT with 10 ms hop size applied to a speech signal produces a spectrogram where voiced phonemes show harmonic formant bands and unvoiced fricatives show broadband noise energy, enabling speech recognition feature extraction. The STFT is the foundation of audio codec window transforms (MP3, AAC) and is implemented on DSPs like the ADSP-21469 for real-time audio processing.

Follow-up: What is the Heisenberg uncertainty principle of the STFT and what tradeoff does it impose on window length selection?