Parseval's Identity

How it works

Parseval's identity has three equivalent statements depending on whether you are working with the continuous Fourier transform, a Fourier series, or a discrete Fourier transform — but all three say the same thing: signal energy (the squared L²-norm) is preserved when you switch from time domain to frequency domain.

Fourier transform:  ∫_ℝ |f(t)|² dt  =  (1/2π) ∫_ℝ |F(ω)|² dω
Fourier series:     (1/2π) ∫_{-π}^{π} |f(t)|² dt  =  Σ_{n=-∞}^{∞} |cₙ|²
DFT (length N):     Σ_{n=0}^{N-1} |x[n]|²  =  (1/N) Σ_{k=0}^{N-1} |X[k]|²

The polarization identity (4⟨f, g⟩ = ‖f + g‖² − ‖f − g‖² + i‖f + ig‖² − i‖f − ig‖²) immediately upgrades each of these to its inner-product form: ⟨f, g⟩ = (1/2π) ⟨F(f), F(g)⟩ in L²(ℝ); ⟨f, g⟩ = Σ cₙ(f) conjugate(cₙ(g)) for Fourier series. The Fourier transform preserves all inner products, not just norms — that's the strongest form.

The reason is that complex exponentials {e^(iωt)} are generalized orthonormal functions: integrating e^(iωt) against e^(iω′t) gives 2π δ(ω − ω′) — a Dirac delta peak. The Fourier transform is just the change-of-coordinates map from the time basis to this exponential basis. Change-of-orthonormal-basis is an isometry — that's all Parseval says, restated in the language of Hilbert space.

Proof sketch

For functions f, g ∈ L¹ ∩ L² (so all integrals converge absolutely), use the inverse-transform formula and Fubini's theorem to swap orders of integration:

⟨f, g⟩ = ∫ f(t) conjugate(g(t)) dt
       = ∫ f(t) conjugate{(1/2π) ∫ F(g)(ω) e^(iωt) dω} dt
       = (1/2π) ∫∫ f(t) conjugate(F(g)(ω)) e^(-iωt) dω dt
       = (1/2π) ∫ conjugate(F(g)(ω)) [∫ f(t) e^(-iωt) dt] dω
       = (1/2π) ∫ conjugate(F(g)(ω)) F(f)(ω) dω
       = (1/2π) ⟨F(f), F(g)⟩

Setting f = g recovers ‖f‖² = (1/2π) ‖F(f)‖² — Parseval's identity. The L¹ ∩ L² hypothesis is then relaxed to all of L² by density: L¹ ∩ L² is dense in L², the Fourier transform is uniformly continuous in L², and the identity extends to L² by completion. This is exactly Plancherel's theorem — extending Fourier from a smooth domain to all of L² as a unitary operator.

For Fourier series, the proof is the same but with sums replacing the inverse-transform integral. For the DFT, both sides are finite sums of N complex numbers and the identity reduces to the orthogonality of the rows of the DFT matrix (W*W = N · I).

Variants and special cases

• Continuous Fourier transform on ℝ. ∫|f|² dt = (1/2π) ∫|F|² dω. Normalization constants depend on the Fourier convention; the symmetric version F(f)(ω) = (1/√(2π)) ∫ f(t) e^(−iωt) dt makes the identity ∫|f|² = ∫|F|² without the 2π factor.
• Fourier series. For 2π-periodic f, (1/2π) ∫_{−π}^π |f|² = Σ |cₙ|². The classical "Parseval's theorem" in older textbooks.
• Discrete Fourier transform. For length-N x, Σ|x[n]|² = (1/N) Σ|X[k]|². Normalization shifts to (1/√N) per DFT in the symmetric convention.
• Multivariate transforms. On ℝⁿ, ∫|f|² dx = (1/(2π)ⁿ) ∫|F|² dξ.
• Wavelets. For an orthonormal wavelet basis {ψ_{j, k}}, ‖f‖² = Σ_{j, k} |⟨f, ψ_{j, k}⟩|². Used in compression: coefficients with small magnitude can be discarded with bounded reconstruction error in L².
• Spherical harmonics on S². For f ∈ L²(S²), ‖f‖² = Σ_{l, m} |a_{lm}|² where a_{lm} = ⟨f, Y_{lm}⟩. Used in cosmology (CMB power spectrum) and geodesy.
• Hermite, Legendre, Laguerre expansions. Each orthonormal-polynomial basis carries its own Parseval. Used in quantum mechanics (Hermite for harmonic oscillator), numerical analysis (Legendre for Gauss-Legendre quadrature), and statistical thermodynamics.
• Pontryagin duality on general LCAGs. For locally compact abelian groups G, the Fourier transform L²(G) → L²(Ĝ) is unitary up to Haar measure normalization. Specialized to G = ℝ, ℤ, ℝ/2πℤ this recovers all earlier cases.

Numerical examples

Example 1 — sine wave. Let f(t) = A · sin(ω₀ t) on [0, T] where T is a multiple of 2π/ω₀. Time-domain energy:

E_t = ∫₀^T A² sin²(ω₀ t) dt = A²T/2

Fourier-side: the spectrum has two spikes (at +ω₀ and −ω₀) each with amplitude A/2 (in the symmetric convention). After applying Parseval with appropriate normalization, the frequency-domain energy reconstructs to A²T/2 exactly — energy lives in two narrow spikes, each carrying half. Visualizing as bars: one bar of height A²T/2 in time-energy = two bars of height A²T/4 in frequency-energy, summing to the same total.

Example 2 — Gaussian. Let f(t) = e^(−αt²). Time-domain L²-norm:

∫_ℝ e^(−2αt²) dt = √(π / (2α))

Fourier transform: F(f)(ω) = √(π/α) e^(−ω²/(4α)). Frequency-domain L²-norm:

(1/2π) ∫_ℝ (π/α) e^(−ω²/(2α)) dω = (1/2π) (π/α) √(2πα) = √(π/(2α))

Both equal — Parseval verified. The Gaussian's energy is the same in time and frequency. Notice the Gaussian's "width" is √(1/(2α)) in time and √(2α) in frequency — wider in time = narrower in frequency, and vice versa (the uncertainty principle). But the squared-norm integral is invariant.

Example 3 — DFT energy preservation. For x = [1, 2, 3, 4]:

‖x‖² = 1 + 4 + 9 + 16 = 30
DFT: X = [10, -2+2i, -2, -2-2i]
‖X‖² = 100 + 8 + 4 + 8 = 120
‖X‖² / N = 120 / 4 = 30   ✓

The Parseval identity Σ|x|² = (1/N) Σ|X|² is verified — both sides equal 30. This is the canonical sanity check for any DFT implementation.

Common pitfalls

• Wrong normalization. Some libraries scale FFT by 1, 1/√N, or 1/N. Parseval looks different in each: Σ|x|² = (1/N) Σ|X|² in one convention, but = (1/√N) ‖X‖² · √N in another. Always check which factor your library uses.
• Treating Fourier coefficients as the spectrum. For a periodic signal, the Fourier series coefficients cₙ are scalars; the "spectrum" is the discrete set {|cₙ|}. The squared sum equals the time-averaged energy — not the total energy if T < ∞.
• Forgetting positive- and negative-frequency contributions. A real-valued sinusoid sin(ω₀ t) has Fourier coefficients at both +ω₀ and −ω₀. Energy lives in both spikes; computing only positive frequencies misses half.
• Truncating DFT and expecting exact preservation. Parseval relates the full N-point DFT to the length-N time signal. Truncating high-frequency bins (e.g., for compression) reduces frequency-domain energy by exactly the discarded coefficients' squared magnitudes — and exactly that much energy is lost from the reconstruction.
• Confusing power spectral density with Fourier transform. The PSD is the Fourier transform of the autocorrelation, by the Wiener-Khinchin theorem. Parseval relates |F|² (square of magnitude) to the time energy; the integrated PSD gives the same thing for wide-sense-stationary signals.
• Applying to non-L² signals. Pure sinusoids on the infinite line are not in L²(ℝ) — their integral of magnitude squared is infinite. Parseval still applies in a distributional sense (the spectrum is a sum of delta functions, the inner product is finite) but care is needed.

Where it shows up

• Signal energy bounds. Total signal power can be bounded by integrating the magnitude-squared spectrum. Used in receiver design (matched filter, SNR analysis), error control coding, and analog-to-digital converter design.
• Power spectral density. The Wiener-Khinchin theorem says PSD is the Fourier transform of the autocorrelation. Parseval guarantees integrated PSD equals signal variance — the energy is the same.
• Quantum mechanics. ψ(x) and ψ̃(p) = F(ψ)(p) are the position and momentum wave functions. Parseval guarantees ∫|ψ(x)|² dx = ∫|ψ̃(p)|² dp = 1 — probability conservation when changing representation.
• Denoising and compression. In transform-domain denoising (wavelet shrinkage, DCT thresholding), bounding reconstruction error in L² reduces to bounding the discarded coefficient energy — Parseval makes the budget explicit.
• Statistical signal processing. Maximum-likelihood detection in white Gaussian noise reduces to inner products of signal templates with received signals. Parseval lets you compute the inner product in time or frequency, whichever is cheaper.
• Heisenberg uncertainty. The standard proof of the position-momentum uncertainty principle uses Cauchy-Schwarz on inner products of ψ and Pψ; Parseval is implicit in switching between position-derivative and momentum-multiplication forms of the momentum operator.
• Functional analysis. Parseval is the prototypical example of an orthonormal-basis expansion in a separable Hilbert space — every Hilbert-space theorem about ONB convergence generalizes its Fourier version.
• CMB cosmology. The cosmic microwave background is decomposed on the sphere into spherical-harmonic coefficients a_{lm}. The angular power spectrum Cₗ = ⟨|a_{lm}|²⟩ summarizes the energy in each multipole — Parseval guarantees total CMB variance equals Σ(2l+1)Cₗ/(4π).