Motivation
A Fourier coefficient compares a function with an oscillation extending across the entire period. Sometimes the location of a change is as important as its frequency. A wavelet coefficient compares behavior at a particular location and scale. The Haar system is the simplest setting in which both ideas can be calculated exactly: each wavelet is positive on one half of a dyadic interval and negative on the other.
The construction is not merely a collection of useful functions. We will prove that its finite levels describe every dyadic step function, then show that increasingly fine dyadic steps approximate every function. This gives a complete orthonormal basis and an energy identity. The conclusion is in the norm; it must not be silently strengthened to recovery of every chosen point value.
Definitions
We work on with Lebesgue measure. Functions in are equivalence classes under equality almost everywhere. Use the inner product linear in its first argument:
The intervals below are half-open. Their treatment of the endpoint affects only a null set and does not affect any statement.
Definition 3.1.1
(Dyadic intervals and Haar functions)
For integers and , let
and write and for its left and right halves. Define the scaling function and the Haar wavelet
The wavelet vanishes outside , takes opposite constant values on its two halves, and has integral zero. The factor makes its squared norm one: the interval has length .
Quick check
(Compute without integrating a product formula.)
Locate the support of inside .
Solution · Answer
The support of is , the right half of , where is the constant . Hence , because every Haar wavelet has integral zero. This is the nesting argument used in the proof below.
The coefficient of is
These integrals are finite: on a finite-measure interval, Cauchy–Schwarz gives . A coefficient is a scaled difference of the two half-interval integrals; it is not generally the unscaled difference of their average values. It vanishes when the two integrals agree, even if is not constant on either half.
Definition 3.1.2
(Dyadic step spaces and finite approximations)
Let be the space of functions constant almost everywhere on each , . It has dimension , with the interval indicators as a basis. For define
where the sum is empty at . The endpoint convention does not change this element of .
Theorems: orthogonality and completeness
Theorem 3.1.3
(The finite-level Haar basis)
For every , the family
is an orthonormal basis of . Consequently is the orthogonal projection of onto and, on each , is equal almost everywhere to the interval average
Theorem 3.1.4
(Haar completeness and Parseval)
The countable family is an orthonormal basis of . More explicitly, for every ,
The corresponding Haar series converges to in when summed by levels. It also converges in in any enumeration of the basis. These statements identify an equivalence class, not values at every point.
Why the Haar system is complete
Proof
(Orthogonality and dimension at a fixed level)
Each has norm one and mean zero, so it is orthogonal to . Distinct intervals at the same level are disjoint. For different levels, two dyadic intervals are either disjoint or the smaller is contained in the larger. In the latter case the finer interval lies entirely in one half of the coarser interval. The coarser wavelet is constant there, while the finer wavelet has integral zero; their inner product is zero. Endpoint overlaps, if a different representative is chosen, have measure zero.
For , each Haar function is constant on every level- cell, so the family lies in . Its number of members is
An orthonormal set is linearly independent, so it is a basis of . Orthogonal projection onto this basis yields the displayed formula for . Alternatively, the interval-average function belongs to and its residual has integral zero on every cell; it is therefore orthogonal to every member of . Uniqueness of the orthogonal projection proves the average formula.
Proof
(Dyadic steps are dense in )
First let be continuous on . Uniform continuity gives a modulus tending to zero with . On a level- cell the difference between and its cell average is at most . Thus
Here is the essential supremum, and the first inequality uses that the total interval has measure one. The cell-average representative may be assigned arbitrary endpoint values without changing the estimate.
Continuous functions are dense in . One route is to approximate functions by simple functions and then approximate measurable-set indicators in by continuous functions, using regularity of Lebesgue measure. Given and , choose continuous with . Orthogonal projection has norm at most one, since the Pythagorean identity gives . Using and the triangle inequality,
First choose , then take sufficiently large. This proves convergence. The finite orthogonal identity
now yields Parseval as . It also proves completeness: a function orthogonal to every Haar function has every , so the limit gives in .
For an arbitrary enumeration, the th partial sum is the orthogonal projection onto the span of its first functions. Every finite level eventually occurs in the enumeration. Once all basis elements through level have appeared, that span contains ; best approximation gives residual no greater than . The spans only grow, so the same bound holds for every later partial sum. Since , every enumeration converges to in . This says nothing by itself about pointwise convergence of an arbitrary enumeration.
Proof X-ray
Why finite-level counting is not yet completeness
The two proofs use different kinds of fact. The first is pure counting. Nesting of dyadic intervals makes the coarser wavelet constant on the support of a finer one, and the mean-zero property of the finer wavelet then kills the inner product; the count upgrades an orthonormal set inside to a basis of . That half alone cannot give completeness, since an orthonormal family can miss a whole direction: delete and the remaining family is still orthonormal, while is orthogonal to all of it.
The second proof is where analysis enters. Uniform continuity gives the cell-oscillation bound , density of continuous functions carries it to every class, and the bound prevents the approximation error from being amplified. The argument closes with the finite Pythagorean identity: once the residual tends to zero, the finite energy identity becomes Parseval, and a function orthogonal to the whole family has for every , hence is zero in .
The projection formula also gives a concrete two-scale energy calculation. Let and be the averages of on the halves of , and write . Their common parent average is , while the Haar coefficient is
On the two halves, has values and has value . Consequently the extra squared norm contributed on is
This identity holds for complex averages as well. It explains what one squared coefficient measures: the energy gained by resolving that cell into two averages. A zero coefficient says those two averages agree; it does not say that is constant on either half.
Concept lensAlgorithmic
Coefficients by repeated averaging and differencing
The two-scale identities compute every coefficient below level from the cell averages of at level . Pair neighboring cells: each pair inside yields the parent average and the coefficient . Repeat on the parent averages until a single average, , remains. There are pairings in all, so passing between the cell-average and the Haar descriptions of costs a number of operations proportional to .
A Lipschitz condition gives a quantitative example of how regularity controls coefficients. Suppose a representative satisfies . Write , where , and pair each point in its left half with in its right half. Then
There are coefficients at level . Combining their squared bounds with Parseval yields
The next example attains equality for and . This is a sufficient estimate for Lipschitz functions, not a characterization of every function with similarly decaying coefficients.
Worked examples
Worked example 3.1.5
(A function already captured at level one)
Let . Its constant coefficient is and its first wavelet coefficient is . Every finer wavelet is supported within a region on which is constant, apart from endpoints, so its coefficient is zero. Therefore
At , under our half-open convention, so the displayed representative on the right has value while the indicator on the left is zero. The equality is still valid in : one endpoint is a null set. This small bookkeeping point is essential when functions are equivalence classes.
Worked example 3.1.6
(A linear function and the energy at each scale)
Let . Its mean is . On a level- interval of length , direct integration shows that the left integral minus the right integral is , independently of the interval's location. Hence
Summing over levels gives , so Parseval reads , matching . The discarded energy after level is
Thus the approximation error in norm is . The squared error and the norm have different powers; taking the square root matters.
Worked example 3.1.7
(A local coefficient detects an imbalance)
Suppose on the left half of , on its right half, and elsewhere. Each half has length , and , so
A different coefficient can also be nonzero: the wavelet on a larger interval may detect that the whole support has more mass on one side. Localization of a basis function does not mean a local feature has only one nonzero coefficient.
Counterexample mode
No basis can recover arbitrary null-set values
Let be zero everywhere except . Every Haar coefficient is zero, as is every partial sum, although the displayed representative has value one at zero. There is no contradiction: this function and zero represent the same element. The basis theorem determines the equivalence class, not values that can be changed without changing any integral.
Common mistakes
Common mistake
(Forgetting the normalization)
Without , the squared norm of the difference of the two half-interval indicators is , not one. Orthonormal projection and Parseval require the normalized functions actually stated. Likewise, the coefficient is an inner product with the normalized wavelet, not just an informal contrast of the two sides.
Common mistake
(Confusing convergence in norm and at a point)
The theorem proves and permits changes on null sets. For continuous , assigning the last cell average to the endpoint gives representatives that converge uniformly, by the modulus estimate above. The literal displayed sums need not do so at : for every wavelet vanishes at and the sum there is . Conversely, for , the dyadic cell containing has average or at every level, because the fractional part of alternates between those values. On that cell the essential supremum error is . Thus uniform convergence fails for this discontinuous function even after changing endpoint values. General pointwise results require separate hypotheses.
Summary
Haar wavelets are normalized local half-interval contrasts. Dyadic nesting proves orthogonality; finite-dimensional counting identifies each level with a step space; density of dyadic steps gives completeness. The coefficient squares divide a function's energy among the constant part and the local scales, while point values remain a separate question.
Exercises
Quick check
(Compute explicitly on its two halves and verify its mean and squared norm.)
Solution · Answer
On , ; on it is ; elsewhere it is zero. The two half-integrals cancel, and the squared norm is .
Quick check
(For , find its nonzero Haar coefficients and verify Parseval.)
Solution · Answer
The constant coefficient is and ; all finer coefficients vanish. Parseval gives .
Quick check
(For , calculate on each half of the interval and compute .)
Solution · Answer
The cell averages are on and on . Each half contributes after translating coordinates, so , also matching the energy tail .
Quick check
(Suppose is orthogonal to and every . Prove that almost everywhere, and explain why this does not imply that every representative has value zero everywhere.)
Solution · Answer
Each is zero. Completeness gives , which implies almost everywhere. A representative may still be changed at, for example, one point without changing its class or coefficients.