Motivation
A narrow graph often has a broad Fourier transform, but a drawing does not say what “narrow” means or which constant is sharp. We will measure spread by a second moment, weighted by the squared magnitude of the function. In that language the trade-off has an exact inequality, and Gaussian functions are precisely the equality cases under the hypotheses below.
The phase matters as well as the magnitude. Multiplying a function by a linear phase shifts its frequency center without changing its intrinsic spread. Multiplying by a quadratic phase can increase the frequency spread even though the spatial magnitude stays exactly the same. Distinguishing these operations prevents an attractive but false rule: every Gaussian-shaped magnitude must be an uncertainty minimizer.
This unit develops the one-dimensional theorem on Schwartz functions, using the Fourier convention and Plancherel theorem already established in the transform chapter. It does not silently extend differentiation identities to arbitrary functions.
Definitions
Definition 2.3.1
(Transform convention and admissible class)
For , use the symmetric angular-frequency transform
Schwartz functions are smooth and all their derivatives decay faster than any inverse power after multiplication by any polynomial. Both and are Schwartz. We use , linear in the first argument, and write for a nonzero , so .
The transform is unitary on under this normalization. In particular . All the integrals defining moments below are absolutely finite in the chosen class. The word “frequency” refers to in ; a convention with in the exponential has a different numerical uncertainty constant.
Definition 2.3.2
(Centers and standard deviations)
Define the spatial and frequency centers by
The variances and standard deviations are
These use squared magnitudes as weights, not the possibly complex values of itself.
The center minimizes the corresponding second moment. Expanding the square and using gives
for every real , and likewise in frequency. A large moment about zero may therefore reflect a translated center rather than a broad shape.
Theorem/Proposition
Theorem 2.3.3
(Differentiation turns frequency moments into an norm)
For every Schwartz function,
The first identity is an integration-by-parts identity; the second follows from Plancherel applied to .
Theorem 2.3.4
(Sharp uncentered uncertainty inequality)
For nonzero ,
Equality holds exactly when for a real and a complex constant .
Theorem 2.3.5
(Centered uncertainty and all its extremizers)
With the centers and variances defined above,
Equality holds exactly for functions of the form
Their centers are and . A nonzero quadratic phase is not part of this equality family.
The theorem is about second moments. It is not a statement that a transform has a particular compact support, that it vanishes outside a chosen frequency band, or that its tallest peak determines its variance. Such conclusions need different definitions and separate arguments.
Proof sketch or proof idea
Proof
(Integrate by parts on a bounded interval)
For the differentiation rule, integrate on and let . The boundary contribution tends to zero, and absolute integrability justifies the limit, yielding . Since is Schwartz, Plancherel applies and gives the norm identity.
Proof
(The sharp constant and its equality equation)
Schwartz decay makes at both infinities. Integrating the derivative of this expression and then passing to the limit gives
The mixed integral is absolutely integrable by Cauchy–Schwarz. Therefore
Neither nor is zero for a nonzero Schwartz function. Equality in Cauchy–Schwarz requires almost everywhere for a constant . Equality in the preceding real-part bound requires the inner product to be real and negative. Since it equals , the constant must be with real .
The functions in that differential identity are continuous, so their almost-everywhere equality is equality everywhere. Solving gives , with . Conversely these functions are Schwartz and make both inequalities equalities. This proves sharpness and exhausts the equality cases.
Proof X-ray
Why the equality condition controls phase
Cauchy–Schwarz alone permits a complex proportionality constant. The real-part identity forces that constant to be negative and real, which produces a real-width Gaussian rather than a Gaussian with a quadratic phase. Tracking both equality steps explains the extremizers; simply observing a Gaussian in one example would not prove that the list is complete.
Proof
(Remove translation and modulation before measuring spread)
Let be the actual centers of , and define
Translation and modulation preserve the Schwartz class and norm. The transform rules give
Consequently and . Applying the uncentered inequality to and dividing by proves the centered one.
Equality holds precisely when , by the equality statement already proved. Undo the two operations and absorb a constant phase into to obtain the stated translated, modulated Gaussian. Conversely its squared magnitudes are symmetric about and , and its centered widths have product .
Worked examples
Worked example 2.3.6
(A normalized Gaussian attains the bound at every scale)
For , take
The Gaussian integral gives . In the stated convention,
For completeness, the transform of has value at zero. Differentiating its transform in and using the differentiation rule gives , hence . This fixes the normalization as well as the shape.
Both centers are zero, and Gaussian second moments give and . Their product in standard deviations is . Increasing narrows the spatial squared-magnitude distribution while broadening the frequency one. The bound is attained throughout this family, not merely at one preferred width.
Worked example 2.3.7
(A shifted center is different from additional spread)
Let . Then
The centers are and , while the variances remain and . The centered bound is still attained. In contrast, the uncentered normalized moments are and . They can be large simply because a center moved far from zero.
For example, at , and , the product of the uncentered root-mean-square widths is , although the centered standard-deviation product remains . The displayed second moments themselves have product . Translating a signal does not create wider spatial uncertainty merely by changing where the coordinate origin lies.
Worked example 2.3.8
(A quadratic phase increases the frequency variance)
Normalize
Its spatial squared magnitude is identical to that of . Both frequency and spatial squared magnitudes are even: is even, so its transform is even. Their centers are zero. Also , so Plancherel yields
This is strictly larger than when . The derivative computation avoids having to choose a complex square-root branch in an explicit transform formula. Equal Gaussian-shaped magnitudes in space can therefore have different frequency variances.
Common mistakes
Counterexample mode
An almost-everywhere classical derivative cannot replace the hypothesis
The normalized box belongs to and has
Its ordinary derivative is zero away from its two jumps. Substituting that almost-everywhere derivative into the Schwartz differentiation theorem would falsely predict a zero frequency moment. The function is not Schwartz or absolutely continuous across its jumps. Integrating by parts on its separate smooth pieces leaves jump boundary terms; they cannot be discarded. The box does not contradict the theorem, and a symmetric principal-value convention cannot turn its divergent moments into the required Lebesgue integrals.
The inequality concerns finite second moments on the stated class. It is not a sharp-support theorem. In particular, the equality classification does show that a nonzero smooth compactly supported function cannot attain the bound, since it is not a Gaussian; it does not by itself prove every possible assertion about compact support in both domains.
Common mistake
(A proof of the inequality does not classify equality)
A proof of the inequality is not automatically a proof of its equality classification. Both the Cauchy–Schwarz condition and the real-part condition must be met. The quadratic-phase example satisfies a complex proportionality relation but fails the real negative proportionality required for the sharp product.
Summary
Schwartz regularity permits integration by parts and the Fourier derivative rule. Plancherel turns a frequency second moment into a derivative norm; the real-part identity then gives the sharp uncertainty constant. Equality forces a Gaussian differential equation. Centering removes translations and linear modulation, while quadratic phase can genuinely enlarge the frequency spread. The hypotheses and normalization are part of the result.
Exercises
Quick check
(Normalize with . Find its spatial and frequency variances, and decide whether it attains the uncertainty bound.)
The Gaussian transform and the identity identify its transform.
Solution · The first odd Gaussian mode is not an extremizer
Since , the constant is . The function is odd and , so both squared magnitudes are even and both centers are zero. Gaussian fourth and second moments give
Thus . The extra factor changes the shape; it is not in the translated-modulated Gaussian equality family.
Quick check
(For define . Prove that the norm and the centered uncertainty product are unchanged. How should be chosen to make the two standard deviations equal?)
Track the means as well as the second moments under the change of variable.
Solution · A dilation redistributes the two widths
A change of variable gives and . The centers become and . The variances become and , so their standard-deviation product is unchanged. Equality of the two new widths requires , or . Both widths are positive for a nonzero Schwartz function by the uncertainty inequality.
Quick check
(Define . Derive the uncertainty constant when frequency spread is measured with rather than with the angular-frequency transform.)
First compare with , including the normalization factor.
Solution · Cycles and angular frequency have different constants
The relation is . Changing variables gives , the frequency center , and standard deviation . Hence
The different constant reflects the unit used to measure frequency, not a different trade-off or a failure of sharpness. The same Gaussian equality family attains it.