Electronics

Fourier Transform Reference

Continuous and discrete Fourier transform definitions, properties, and common pairs.

Definitions

CTFTX(f) = ∫ x(t) · e^{−j2πft} dt
DTFTX(e^{jω}) = Σ x[n] · e^{−jωn}
DFTX[k] = Σ_{n=0}^{N−1} x[n] · e^{−j2πkn/N}
IDFTx[n] = (1/N) Σ_{k=0}^{N−1} X[k] · e^{j2πkn/N}
FFTO(N log N) algorithm for computing the DFT

Properties

Propertyx(t) ↔ X(f)
Linearitya·x + b·y ↔ a·X + b·Y
Time shiftx(t − t₀) ↔ X(f) · e^{−j2πf t₀}
Frequency shiftx(t) · e^{j2πf₀t} ↔ X(f − f₀)
Time scalingx(a·t) ↔ (1/|a|) X(f/a)
Convolution ↔ productx ⊛ y ↔ X · Y
Product ↔ convolutionx · y ↔ X ⊛ Y
Parseval (energy)∫|x(t)|² dt = ∫|X(f)|² df
Dualityx(t) ↔ X(f) implies X(t) ↔ x(−f)

Common pairs

x(t)X(f)
δ(t) (impulse)1
1δ(f)
cos(2πf₀t)(δ(f − f₀) + δ(f + f₀)) / 2
rect(t)sinc(f)
sinc(t)rect(f)
e^{−at}u(t) (a > 0)1 / (a + j2πf)
e^{−πt²}e^{−πf²}
u(t) (step)1 / (j2πf) + ½δ(f)

Worked example: a 4-point DFT

Take x = [1, 2, 3, 4] with N = 4. The twiddle factor e−j2πkn/4 steps through 1, −j, −1, +j for k = 1. Then X[0] = 1 + 2 + 3 + 4 = 10 (the DC term, N times the mean). X[1] = 1 + 2(−j) + 3(−1) + 4(+j) = −2 + 2j. X[2] = 1 − 2 + 3 − 4 = −2. X[3] = −2 − 2j, the complex conjugate of X[1], as expected for a real input.

Parseval check: the time-domain energy is 1 + 4 + 9 + 16 = 30. In the frequency domain, (1/N) × (|10|² + |−2+2j|² + |−2|² + |−2−2j|²) = (100 + 8 + 4 + 8) / 4 = 30. The two agree, which is a quick way to catch a missing 1/N in your own code.

Practical points

  • The FFT is the same answer, faster. A direct DFT takes about N² operations; the FFT takes about N log2 N. For N = 1,000,000 that is 1012 against roughly 2 × 107, about 50,000 times fewer.
  • Check the convention. Libraries differ on where the 1/N goes (inverse only, both transforms as 1/√N, or forward) and on whether the exponent is e−jωt or e+jωt. The magnitude is the same but the scale and phase sign may not be.
  • Pulse width sets bandwidth. A rectangular pulse of width T has a sinc spectrum with its first null at 1/T, so a 1 ms pulse has its first null at 1 kHz. Shorter pulses spread over more frequency, and no signal can be limited in both time and frequency.
  • Phase wraps. Phase is only defined modulo 2π, and is meaningless where the magnitude is near zero. Unwrap it only where the signal has real content.
Was this article helpful?