An irrational rotation returns arbitrarily close to its starting point, but recurrence alone does not tell us how often it visits each region. Equidistribution asks for the correct long-run frequency of every observation. Fourier characters reduce this problem of infinitely many tests to exponential sums; a differencing inequality then handles phases that are nonlinear in the index.
Equidistribution and Weyl's criterion
Write , with normalized Haar measure .
Definition 1.9.1
(Equidistributed sequence)
A sequence is equidistributed if for every continuous ,
In particular, continuous functions with zero average must have sample average tending to zero. The definition does not claim a uniform rate over all sequences or all test functions.
Concept lensAnalytic
Empirical measures converging to Haar measure
Let , the probability measure with mass at each of the first points. Then (1) says exactly that for every continuous : the empirical measures converge weakly to Haar measure. Their Fourier coefficients are , while for . Read this way, the criterion below is the Haar-measure case of a general fact: for probability measures on the torus, weak convergence is equivalent to convergence of every Fourier coefficient.
Theorem 1.9.2
(Weyl's Fourier criterion)
The sequence is equidistributed if and only if, for every nonzero ,
Consequently, for every product of intervals in , the fraction of points in that box converges to its Haar measure. The same holds for a one-dimensional interval, with any choice of whether to include its endpoints.
Proof
(From characters to all continuous tests)
Each nonconstant character has integral zero, so (1) implies (2). Conversely, (2) and the constant character establish (1) for every finite trigonometric polynomial. Such polynomials are uniformly dense in : convolve a continuous function with the product of one-dimensional Fejér kernels, which is a trigonometric polynomial and converges uniformly by the approximate-identity property. If is within of uniformly, both the sample average and the integral change by at most . Passing first to the limit for and then letting proves (1).
For an interval, approximate its indicator from below and above by continuous periodic functions with integrals arbitrarily close to the interval's length. A finite product of such nonnegative approximants handles a box; its boundary has Haar measure zero. Applying (1) and squeezing gives the frequency statement. Endpoint choices do not affect the limit because they are included in arbitrarily small boundary neighborhoods.
Worked example 1.9.3
(A rotation in several coordinates)
Let in . The geometric-series formula gives, for ,
whenever . Thus the orbit is equidistributed precisely when for every nonzero integer , equivalently when are linearly independent over . If some nonzero has integral dot product, its character is constant and (2) fails.
Worked example 1.9.4
(How fast one character average decays)
For a one-dimensional rotation with , the formula above and give
Take . For , and the bound is about . For , lies within of an integer, , and the bound is about . For , lies within of , and the bound is about . Each fixed frequency decays like , but the constant grows whenever comes close to an integer; by Dirichlet's approximation theorem the distance is less than for infinitely many , so these bounds have no constant uniform in . The edge case gives a zero denominator: then every term equals one, the average is for every , and this is how the criterion detects a rational relation.
Counterexample mode
Uniform coordinates need not give a uniform pair
Let be irrational and . Each coordinate is equidistributed on the circle by the one-dimensional rotation criterion. But modulo one for every . The nonzero character is identically one, so the pairs are not equidistributed on . Coordinatewise tests do not replace mixed Fourier characters.
Differencing a nonlinear phase
The next elementary inequality replaces the average of a sequence by correlations between shifted copies. Its strength is that the phase difference of a polynomial has lower degree.
Theorem 1.9.5
(Van der Corput differencing)
For complex numbers with and each integer ,
The sum is empty for . In particular, if each fixed positive-shift correlation tends to zero, the original averages tend to zero.
Proof
(Average shifted sums and expand the square)
Fix and extend by zero outside . Every term appears times in
Cauchy–Schwarz bounds the square of the left side by times the sum of squared inner sums. Expanding these squares and using bounds that sum by
where an empty inner sum is zero. Divide by , bound each real part by its absolute value, and take with fixed. This gives (3). If every fixed-shift correlation tends to zero, the limsup is at most for every , hence zero.
Theorem 1.9.6
(Polynomial equidistribution)
For a real polynomial , the sequence is equidistributed on if and only if at least one nonconstant coefficient is irrational.
Proof
(Induction, then remove rational higher terms)
First suppose the leading coefficient is irrational. For every nonzero integer , prove the average of tends to zero by induction on . At , this is the geometric-series formula because . For , a fixed positive shift gives
The difference has degree and irrational leading coefficient ; multiplication by nonzero integer preserves irrationality. The induction hypothesis makes every fixed-h correlation tend to zero, so (3) proves the desired exponential-sum limit. Weyl's criterion finishes the leading-irrational case.
Now suppose the highest irrational nonconstant coefficient occurs at degree , but higher coefficients may be rational. Let be the sum of the terms of degrees above . Choose a positive common denominator for Q's coefficients. The binomial theorem shows is an integer, so its phase is constant on each residue class modulo . On , the polynomial has degree and irrational leading coefficient . The leading-irrational case applies to each residue class; the finitely many class averages, whose lengths differ by at most one, then give the full exponential-sum limit for every nonzero .
Conversely, if every nonconstant coefficient is rational, choose a positive integer clearing their denominators. Then for all integers . Its average cannot tend to zero, so (2) fails. An irrational constant alone cannot make the sequence equidistributed.
Proof X-ray
Where irrationality survives the differencing
The induction runs on the degree, and the van der Corput inequality is the only tool that lowers it. What must survive each step is the irrationality of the leading coefficient: differencing with shift multiplies the top coefficient by the nonzero integer , and a nonzero integer multiple of an irrational number is irrational. The base case is the geometric series, so the argument works one frequency and one shift at a time; it produces limits, never a rate uniform in , or the coefficients. Rational higher terms are not differenced away but frozen: on each residue class modulo their phase is constant, which leaves a polynomial in with irrational leading coefficient . The converse makes the opposite move, multiplying by so that every nonconstant term becomes an integer phase.
Worked example 1.9.7
(A rational quadratic term does not hide an irrational linear term)
For , the leading coefficient is rational, so the first induction step alone is not applicable. The rational quadratic phase is periodic modulo one; the residue-class step above shows is nevertheless equidistributed. Thus, for example, the proportion of indices with fractional part in tends to .
By contrast, has rational nonconstant coefficients. It takes only the two values and modulo one, and it is not equidistributed.
Common mistakes
Common mistake
(Dense is not equidistributed)
A sequence can come arbitrarily close to every point and still have the wrong frequencies. Let when is a perfect square, and otherwise. The polynomial theorem makes equidistributed, hence dense, so is dense in . But fewer than of the first indices are squares, so for continuous the averages in (1) tend to , not to . Equidistribution is a statement about frequencies, which density does not control.
Practice with solutions
Quick check
(Show directly that spends asymptotic proportion in .)
Solution · Answer
Every nonzero integer multiple of is nonintegral, so the geometric sum tends to zero. Weyl's criterion gives equidistribution, and the interval statement gives proportion .
Quick check
(Prove that is equidistributed modulo one.)
Solution · Answer
Its leading coefficient is irrational. For each nonzero integer and fixed , the phase difference of is linear with irrational coefficient . Its average tends to zero geometrically; differencing and Weyl's criterion give the result.
Quick check
(Are equidistributed on ?)
Solution · Answer
No. Their difference is the integer modulo one. The character is identically one, so the two-dimensional criterion fails, although each coordinate separately is equidistributed.
Quick check
(Show that is not equidistributed modulo one.)
Solution · Answer
The nonconstant denominators are cleared by 12. Thus for all , a nonzero constant. Weyl's criterion fails at frequency 12.
Quick check
(For a given , choose an irrational making the first terms of have average of magnitude greater than .)
Solution · Answer
Take . For , the angle lies between 0 and . Every real part exceeds , so their average has magnitude greater than . For this fixed irrational , the average still tends to zero as the observation length subsequently tends to infinity.