The Heisenberg Uncertainty Principle for Functions

The precise statement

Fix the analyst's Fourier convention f̂(ξ) = ∫_ℝ f(x) e^(−2πixξ) dx, so that Plancherel gives ‖f̂‖₂ = ‖f‖₂. Let f ∈ L²(ℝ) with ‖f‖₂ = 1, and suppose the relevant integrals below are finite (i.e. x·f ∈ L² and ξ·f̂ ∈ L²). Then for every pair a, b ∈ ℝ,

(∫_ℝ (x−a)² |f(x)|² dx) · (∫_ℝ (ξ−b)² |f̂(ξ)|² dξ) ≥ 1/(16π²).

The two factors are the variances of the probability densities |f|² and |f̂|² (both integrate to 1 by Plancherel). Writing V_a(f) and V_b(f̂) for these, and optimizing over a, b, the minima are attained at the means a* = ∫x|f|²dx and b* = ∫ξ|f̂|²dξ. So the sharpest reading is: the product of variances of a unit-mass function's position and momentum densities is bounded below by 1/(16π²) ≈ 0.00633. Equivalently, the standard deviations satisfy σ_x · σ_ξ ≥ 1/(4π).

The picture: you cannot cheat both domains

A function and its Fourier transform are two views of the same object, linked by Plancherel's unitarity. Scaling reveals the tension immediately: if f_λ(x) = √λ · f(λx), then f̂_λ(ξ) = λ^(−1/2) f̂(ξ/λ). Squeezing f horizontally (large λ) shrinks its position variance by λ² but inflates the frequency variance by exactly λ², so the product is scale-invariant — you can move the trade-off around but never beat it.

Intuitively, resolving fine detail in x requires high frequencies; a spike needs a broad band to synthesize it (the δ-function has flat spectrum), while a pure tone e^(2πibx) is perfectly localized in ξ but spread over all of x. The inequality says these are not accidents but the extremes of a single conserved quantity. The Gaussian sits precisely at the balance point: it is its own Fourier transform (up to the scaling constant), so it distributes 'blur' symmetrically between the two domains and achieves equality.

Key idea of the proof: integration by parts meets Cauchy–Schwarz

By translating (x ↦ x−a) and modulating (f ↦ e^(−2πibx)f, which shifts f̂ by b), we may assume a = b = 0; these operations preserve ‖f‖₂ and the two variances. The engine is the identity that Fourier turns differentiation into multiplication: the transform of f′ is 2πiξ·f̂. Hence ∫ ξ²|f̂|² dξ = (1/4π²) ∫ |f′|² dx by Plancherel.

Now integrate by parts the real quantity ∫ x·(f̄ f′ + f f̄′) dx = ∫ x·(|f|²)′ dx = −∫ |f|² dx = −1 (the boundary terms vanish since x|f|² → 0, forced by x·f ∈ L²). So |∫ x·2Re(f̄ f′) dx| = 1, giving 1 ≤ 2 ∫ |x f| · |f′| dx. Apply Cauchy–Schwarz:

1 ≤ 2 (∫ x²|f|² dx)^(1/2) (∫ |f′|² dx)^(1/2) = 4π (∫ x²|f|² dx)^(1/2) (∫ ξ²|f̂|² dξ)^(1/2).

Squaring yields the constant 1/(16π²). Equality in Cauchy–Schwarz forces f′ = −c·x·f for a constant c > 0, i.e. f is a Gaussian.

Worked example: the Gaussian family attains equality

Take f(x) = (2α)^(1/4) e^(−παx²), α > 0, normalized so ‖f‖₂ = 1. A direct computation gives the position variance ∫ x²|f|² dx = 1/(4πα). Its Fourier transform is again Gaussian: f̂(ξ) = (2/α)^(1/4) e^(−πξ²/α), with frequency variance ∫ ξ²|f̂|² dξ = α/(4π).

The product is (1/(4πα))·(α/(4π)) = 1/(16π²) — equality, for every α. Choosing α tunes the width in x versus ξ but never the product. This confirms the extremizer analysis: the full equality set is exactly the four-parameter family c·e^(−α(x−a)²)·e^(2πibx) with |c| fixed by normalization, obtained from the standard Gaussian by scaling (α), translation (a), and modulation (b). Physically these are the coherent states / minimum-uncertainty wave packets: a particle described by such a state saturates ΔX·ΔP = ℏ/2. No other functions do.

Why the hypotheses matter — and sharper cousins

The finiteness assumptions x·f ∈ L² and ξ·f̂ ∈ L² are not decoration: if either variance is infinite the inequality holds trivially (∞ ≥ constant) but the extremizer characterization fails, and the integration-by-parts boundary term x|f|² must genuinely vanish, which requires the decay these conditions encode. Unit mass ‖f‖₂ = 1 is a normalization; for general f both sides scale by ‖f‖₂⁴, so the correct homogeneous form is V(f)·V(f̂) ≥ ‖f‖₂⁴/(16π²).

The variance form is only one face of a deeper phenomenon. Benedicks' theorem: if f and f̂ are both supported on sets of finite measure, then f ≡ 0 — you cannot compactly support both. Hardy's theorem quantifies the Gaussian barrier: |f| ≲ e^(−παx²) and |f̂| ≲ e^(−πβξ²) with αβ > 1 forces f ≡ 0, and αβ = 1 forces f to be a Gaussian. These are qualitative uncertainty principles, invisible to the variance inequality but expressing the same rigidity.

Why it matters: from spectroscopy to operator theory

In quantum mechanics the inequality is Heisenberg's principle: for position operator X and momentum P = −iℏ d/dx = (ℏ/i) d/dx, the commutator [X,P] = iℏ forces σ_X σ_P ≥ ℏ/2 via the Robertson–Schrödinger inequality, of which the L² statement is the free-particle case. It bounds ground-state energies (the stability of matter), Gabor/wavelet time-frequency resolution (you cannot build a basis both time- and frequency-localized — the Balian–Low theorem), and the diffraction limit in optics.

Mathematically it is the prototype for a whole industry: the Amrein–Berthier quantitative form, entropic uncertainty (Hirschman–Beckner, sharpened by the Babenko–Beckner Hausdorff–Young constant), Donoho–Stark discrete/finite uncertainty underpinning compressed sensing, and analogues on ℝⁿ, tori, finite abelian groups, and the Heisenberg group. Each says the same thing: a function and its spectrum cannot both be concentrated, and the Gaussian (or its group analogue) marks the boundary of what is possible.

Uncertainty inequalities: hypotheses, quantities controlled, and extremizers

Inequality	Statement (‖f‖₂=1)	Extremizer / sharpness
Heisenberg (variance form)	V(f)·V(f̂) ≥ 1/(16π²), where V(f)=∫(x−a)²|f|²dx	Gaussians e^(−αx²) (mod translation/modulation/scaling)
Hardy's uncertainty	If |f(x)|≤Ce^(−παx²) and |f̂(ξ)|≤Ce^(−πβξ²) with αβ>1, then f≡0	αβ=1 forces f = c·e^(−παx²) (Gaussian)
Benedicks / Amrein–Berthier	If supp f and supp f̂ both have finite Lebesgue measure, then f≡0	Qualitative — no nonzero extremizer
Entropic (Hirschman–Beckner)	h(|f|²)+h(|f̂|²) ≥ log(e/2)	Gaussians; implies Heisenberg via Jensen
Higher-dim ℝⁿ	V(f)·V(f̂) ≥ n²/(16π²)	Radial Gaussians e^(−α|x|²)