Chebyshev approximation, Chebyshev–Gauss sampling, and aliasing on an interval
Turn a nonperiodic interval into a cosine series, prove continuous and discrete orthogonality, and analyze Chebyshev interpolation, reflected aliasing, endpoint singularities, and noise.
§6.1Estimated reading time: 40 min
Motivation
Throughout this note, functions, polynomial coefficients, sample values and measurement errors are real-valued.
A Fourier method on an interval first identifies the two endpoints. If the endpoint values or derivatives of the function disagree, the periodic extension acquires a jump or a kink that the function never had, and the coefficients decay slowly for a reason unrelated to the interior behavior. This note avoids that defect by changing the variable instead of the function. Writing x=cosθ turns a function f on [−1,1] into g(θ)=f(cosθ), which is 2π-periodic and even for every f. Its values at θ=±π both equal f(−1), so there is no endpoint compatibility condition to check.
The cosine series of g, rewritten in x, is a series of polynomials. Its sampled version again aliases, but the folding is a reflection with a change of sign rather than a shift. We separate what the samples justify on their own from what needs an extra hypothesis about f, and end with two limitations the change of variables does not remove: genuine endpoint singularities and noise amplification.
Chebyshev polynomials
Definition 6.1.1
(Chebyshev polynomials)
Set T0(x)=1, T1(x)=x and, for k≥1,
Tk+1(x)=2xTk(x)−Tk−1(x).
The polynomial Tk is the Chebyshev polynomial of degree k. For every real θ and every integer k≥0,
Tk(cosθ)=coskθ.
The identity follows by induction from cos((k+1)θ)+cos((k−1)θ)=2cosθcoskθ, which is the recurrence evaluated at x=cosθ. The same induction shows that Tk has degree exactly k and, for k≥1, leading coefficient 2k−1: multiplying by 2x raises the degree by one and doubles the leading coefficient, while subtracting Tk−1 cannot cancel the new top term. The first few are T2=2x2−1, T3=4x3−3x, T4=8x4−8x2+1 and T5=16x5−20x3+5x.
Three consequences are used repeatedly.
Bounds and endpoint values. Every x∈[−1,1] equals some cosθ, so ∣Tk(x)∣≤1 there. Moreover Tk(1)=1 and Tk(−1)=(−1)k.
A basis. The degrees are exactly 0,1,…,N, so T0,…,TN form a basis of the space PN of real polynomials of degree at most N.
A product formula.2TjTk=Tj+k+T∣j−k∣ is the polynomial form of 2cosjθcoskθ=cos(j+k)θ+cos(j−k)θ. It holds on [−1,1], hence everywhere, because two polynomials that agree on an interval are equal.
Continuous orthogonality and the weight
For continuous functions u,v on [−1,1] define
⟨u,v⟩w=∫−11u(x)v(x)1−x2dx,∥u∥w=⟨u,u⟩w1/2.
The integral is improper at both endpoints but converges absolutely: the integrand is bounded by ∥u∥∞∥v∥∞(1−x2)−1/2, and the substitution below shows that the weight has total integral π. In particular ∥u∥w≤π∥u∥∞, which converts uniform estimates into weighted ones.
Theorem 6.1.2
(Continuous orthogonality)
For integers j,k≥0,
⟨Tj,Tk⟩w=⎩⎨⎧0,π/2,π,j=k,j=k≥1,j=k=0.
Proof
(Substitute x=cosθ)
On (0,π) the map θ↦cosθ is a strictly decreasing bijection onto (−1,1). With x=cosθ we have dx=−sinθdθ and 1−x2=sinθ, because sinθ>0 on (0,π). Reversing the orientation cancels the minus sign, so for continuous u,
∫−11u(x)1−x2dx=∫0πu(cosθ)dθ.
The right-hand side is an ordinary integral of a continuous function over a compact interval, so the improper integral on the left converges. With u=TjTk and Tj(cosθ)Tk(cosθ)=cosjθcoskθ,
⟨Tj,Tk⟩w=21∫0π(cos(j+k)θ+cos(j−k)θ)dθ.
For an integer m, ∫0πcosmθdθ equals π if m=0 and sin(mπ)/m=0 otherwise. If j=k, both j+k and j−k are nonzero and the integral vanishes. If j=k≥1, only the second term contributes, giving π/2. If j=k=0, both contribute, giving π.
Proof X-ray
Where the weight comes from
The factor (1−x2)−1/2 is not chosen for convenience. It is exactly the reciprocal of the Jacobian factor sinθ produced by dx=−sinθdθ. With this factor, the x-integral becomes the plain θ-integral, in which cosines are orthogonal. Without it, the substitution leaves an extra factor sinθ, and orthogonality fails. For instance, ∫−11T0T2dx=∫−11(2x2−1)dx=−2/3=0. The weight is large near ±1, so errors near the endpoints count more heavily in ∥⋅∥w than in the unweighted norm.
The Chebyshev series is a mapped cosine series
Definition 6.1.3
(Chebyshev coefficients and truncation)
For continuous f on [−1,1] and k≥0, the Chebyshev coefficients are
The constant term is halved because ∥T0∥w2=π is twice ∥Tk∥w2=π/2. With this normalization, ⟨ΠNf,Tk⟩w equals akπ/2 for 1≤k≤N and (a0/2)π for k=0, which is ⟨f,Tk⟩w in both cases. Hence f−ΠNf is w-orthogonal to PN, and for every q∈PN Pythagoras' theorem gives
∥f−q∥w2=∥f−ΠNf∥w2+∥ΠNf−q∥w2.
Thus ΠNf is the unique best approximation to f from PN in the weighted norm.
The second formula for ak explains the section title. Since g(θ)=f(cosθ) is even and 2π-periodic, its sine coefficients vanish and its cosine coefficients are π−1∫−ππg(θ)coskθdθ=ak. The Chebyshev series of f is the Fourier cosine series of g, read through x=cosθ. If f is continuous, so is g; differentiability of f passes to g by the chain rule, with no condition linking f(−1) to f(1).
Theorem 6.1.4
(Coefficient decay without endpoint matching)
Suppose f is twice continuously differentiable on an open interval containing [−1,1]. Then for every k≥1,
∣ak∣≤k22(∥f′∥∞+∥f′′∥∞),
where the norms are taken over [−1,1]. Consequently k∑∣ak∣<∞, and f(x)=a0/2+k≥1∑akTk(x) uniformly on [−1,1].
Proof
(Two integrations by parts, then Fejér)
The function g=f∘cos is twice continuously differentiable on R, with
This boundary term vanishes because g′(0)=g′(π)=0. Therefore ∣ak∣≤(2/(πk2))⋅π∥g′′∥∞, which is the stated bound.
For convergence: ∑∣ak∣<∞ and ∣Tk∣≤1, so the series converges uniformly to a continuous F. Term-by-term integration shows that F has the same Chebyshev coefficients as f. Then h(θ)=g(θ)−F(cosθ) is continuous, even and 2π-periodic, with all cosine coefficients zero and, by evenness, all sine coefficients zero. Its Fejér means are therefore identically zero, and by the Fejér theorem for continuous periodic functions they converge uniformly to h. Hence h=0 and F=f.
It is worth isolating exactly where the hypotheses on f enter.
Proof X-ray
Why no endpoint condition is needed
Each integration by parts produces a boundary term, and both vanish for structural reasons. The first contains sinkθ, which is zero at 0 and π for every integer k. The second contains g′(θ)=−sinθf′(cosθ), which is zero there because of the Jacobian factor sinθ, whatever the values of f′(±1). No hypothesis on f(±1) or f′(±1) was used; differentiability near [−1,1] only makes g twice continuously differentiable on R. The convergence half uses a different idea: absolute summability gives a uniform limit, and Fejér's theorem identifies it with f.
The same pattern continues under further integrations by parts. An even-order derivative of g meets sinkθ at the boundary, and an odd-order derivative of the even 2π-periodic g is odd and 2π-periodic, so it vanishes at 0 and π. Consequently, if f is m times continuously differentiable near [−1,1], then ∣ak∣≤2∥g(m)∥∞k−m, and ∥g(m)∥∞ is bounded in terms of f′,…,f(m).
Chebyshev–Gauss nodes and discrete orthogonality
Definition 6.1.5
(Chebyshev–Gauss nodes)
For an integer n≥1 and j=0,1,…,n−1, put
θj=2n(2j+1)π,xj=cosθj.
The angles θj are the midpoints of the n equal subintervals of [0,π], so the nodes xj are distinct and lie strictly inside (−1,1). Since nθj=(2j+1)π/2, Tn(xj)=cosnθj=0; a polynomial of degree n has at most n zeros, so the nodes are exactly the zeros of Tn. Equally spaced angles give unequally spaced nodes: near the center consecutive nodes are about π/n apart, while 1−x0=2sin2(π/(4n))≤π2/(8n2) by sint≤t. The endpoints themselves are never sampled.
The discrete analogue of the orthogonality integral is a sum over these angles, and it rests on one exact computation.
Theorem 6.1.6
(Cosine sums over the Gauss angles)
For every integer m, let Sn(m)=j=0∑n−1cosmθj. Then Sn(m)=0 unless m is a multiple of 2n, and Sn(2nq)=(−1)qn for every integer q.
Proof
(A geometric sum with a half-step phase)
If m=2nq, then mθj=q(2j+1)π, so cosmθj=(−1)q(2j+1)=(−1)q for every j and the sum is (−1)qn.
Otherwise ω=eimπ/n=1, and Sn(m) is the real part of
By the product formula, the sum equals 21(Sn(k+l)+Sn(k−l)). The hypotheses give 0≤k+l≤2n−2 and ∣k−l∣≤n−1, and the only multiple of 2n in these ranges is zero. So Sn(k−l) equals n if k=l and 0 otherwise, while Sn(k+l) equals n if k=l=0 and 0 otherwise. Combining the cases gives the three values.
Quick check
(With n=3, the node values of T1 are 3/2,0,−3/2 and those of T5 are −3/2,0,3/2. Why is j∑T1(xj)T5(xj) not zero?)
Locate k+l relative to 2n.
Solution · Answer
Directly, the sum is −43+0−43=−23. By the lemma it is 21(S3(6)+S3(−4))=21(−3+0)=−23, because 6=2n with q=1. Discrete orthogonality assumes l≤n−1, and l=5 violates it. Discrete orthogonality holds only on a finite block of indices determined by the grid; the failure outside that block is exactly aliasing.
Discrete coefficients and the interpolation theorem
Definition 6.1.8
(Discrete Chebyshev coefficients and the interpolant)
For n≥1 and a function f defined at the nodes, set
The formula is the midpoint rule with n subintervals of width π/n applied to ak=(2/π)∫0πg(θ)coskθdθ: the midpoints are the θj, and (2/π)(π/n)=2/n. Up to normalization, the map from samples to (a0,…,an−1) is a type-II discrete cosine transform, for which fast algorithms exist.
Theorem 6.1.9
(Exactness and interpolation)
Let n≥1.
If q∈Pn−1, then Inq=q.
For every function f defined at the nodes, Inf(xj)=f(xj) for j=0,…,n−1, and Inf is the only polynomial in Pn−1 with these values.
Proof
(Exactness by orthogonality, interpolation by dimension count)
(1) Write q=b0/2+l=1∑n−1blTl. For 1≤k≤n−1, linearity and discrete orthogonality give
For k=0 only the l=0 term survives, and a0=(2/n)(b0/2)n=b0. Hence Inq=q.
(2) The evaluation map q↦(q(x0),…,q(xn−1)) from Pn−1 to Rn is linear and injective, because a nonzero polynomial of degree at most n−1 cannot vanish at n distinct points. Both spaces have dimension n, so it is bijective, and exactly one q∈Pn−1 has q(xj)=f(xj) for all j. The coefficients ak depend only on the samples, which are the same for f and q, so Inf=Inq=q by part (1).
The hypothesis n≥1 makes the grid nonempty. For n=1 the single node is x0=cos(π/2)=0, a0=2f(0), and I1f is the constant f(0), as the theorem predicts. Part (2) is a uniqueness statement inside Pn−1 only; it says nothing about which function produced the data.
Aliasing as a reflection
Fourier aliasing shifts a frequency by a multiple of the number of samples. On the Chebyshev–Gauss grid the key identity is 2nθj=(2j+1)π, an odd multiple of π, and it produces a reflection instead of a shift.
Theorem 6.1.10
(Folding of Chebyshev indices)
Let n≥1 and m≥0, and write m=2nq+r with integers q≥0 and 0≤r≤2n−1. At every node xj,
Consequently InTm equals (−1)qTr, 0 or (−1)q+1T2n−r in the three cases, and ∣InTm(x)∣≤1 on [−1,1]. In particular, InT2n−k=−Tk for 1≤k≤n−1.
Proof
(Use 2nθj=(2j+1)π)
Since 2nqθj=q(2j+1)π and cos(α+pπ)=(−1)pcosα for every integer p,
cosmθj=(−1)q(2j+1)cosrθj=(−1)qcosrθj.
This is the first case. If r=n, then cosnθj=0. If n+1≤r≤2n−1, put s=2n−r∈{1,…,n−1}; then rθj=(2j+1)π−sθj, so cosrθj=−cossθj. In each case the right-hand side is a polynomial in Pn−1 with the same node values as Tm, so by uniqueness in the interpolation theorem it is InTm, and its absolute value is at most one on [−1,1]. The final statement is the case q=0, r=2n−k.
Quick check
(With n=4, which polynomial in P3 has the same node values as T6?)
Write 6 as 2n−k.
Solution · Answer
Here 6=2⋅4−2, so q=0, r=6≥n+1 and the reflected index is 2n−r=2. The folding theorem gives I4T6=−T2=1−2x2.
Worked example 6.1.11
(Sampling T5 on three nodes)
Take n=3. The angles are π/6,π/2,5π/6, and the nodes are 3/2,0,−3/2. With k=1 the reflected index is 2n−k=5, so the theorem predicts T5(xj)=−xj.
We check this from T5(x)=16x5−20x3+5x. At x=3/2 we have x3=33/8 and x5=93/32, so
T5(23)=293−2153+253=−23.
Since T5 is odd, T5(−3/2)=3/2, and T5(0)=0. These are the values of −x at the nodes, so the three-point interpolant of T5 is −T1: a polynomial of a different degree with the opposite sign. The content at index 5 is reported at index 1, reflected about n=3.
The example shows two different polynomials with the same three samples. This is not a peculiarity of T5.
Counterexample mode
Equal samples, different functions
Claim to test: a function is determined by its values at the n Chebyshev–Gauss nodes. Let h=T5+T1. By the product formula, h=2T3T2, and since T3 vanishes at the three nodes, h(xj)=0 for j=0,1,2. Explicitly h(x)=16x5−20x3+6x=2x(4x2−3)(2x2−1), which vanishes at 0,±3/2 and at ±1/2, yet h(1)=2. So T5 and −T1 have identical samples but differ by 2 at the endpoint.
More generally, f and f+Tnu share all their samples for any function u. The samples determine Inf; they determine f only if f is known in advance to lie in a class, such as Pn−1, in which no two distinct members agree at all the nodes.
The folding theorem becomes a coefficient identity once f is represented by its Chebyshev series at the nodes. We state this representation as a hypothesis, because the samples cannot verify it.
Theorem 6.1.12
(Aliasing formula)
Assume (H): f is continuous on [−1,1], m≥0∑∣am∣<∞, and f(x)=a0/2+m≥1∑amTm(x) for every x∈[−1,1]. Then for n≥1 and 1≤k≤n−1,
By the folding theorem, either Tm vanishes at all nodes, or Tm(xj)=σTs(xj) for a sign σ and a folded index 0≤s≤n−1. Discrete orthogonality then gives, for k≥1, cm,k=σ if s=k and 0 otherwise; for k=0, cm,0=2σ if s=0 and 0 otherwise.
For k≥1, the indices folding to k are m=2nq+k with sign (−1)q and m=2n(q+1)−k with sign (−1)q+1, for q≥0; renaming q+1 as q in the second family gives the first formula. For k=0, only m=2nq folds to zero, with sign (−1)q: the term m=0 contributes (a0/2)⋅2=a0, and each q≥1 contributes 2(−1)qa2nq.
Thus each index m≥n contributes to at most one retained coefficient of Inf, with coefficient ±1 in a0/2+k≥1∑akTk, and indices with m≡n(mod2n) contribute nowhere. Without (H), the right-hand sides need not converge, or need not describe the function that was actually sampled.
A fully calculated example with n=3
Worked example 6.1.13
(Discrete coefficients of x4 on three nodes)
Let f(x)=x4 and n=3. The samples at 3/2,0,−3/2 are 9/16,0,9/16. For j=0,1,2, the values of coskθj are 1,1,1 for k=0; 3/2,0,−3/2 for k=1; and 1/2,−1,1/2 for k=2. As a check of discrete orthogonality, j∑cos2θj=3/2=n/2 and j∑cosθjcos2θj=3/4+0−3/4=0. The discrete coefficients are
which equals 9/16 at ±3/2 and 0 at 0, as interpolation requires.
To see where the coefficients come from, substitute T2 and T4 to check 8x4=3T0+4T2+T4. So a0=3/4, a2=1/2, a4=1/8 and ak=0 otherwise. The aliasing formula with n=3 predicts a2=a2−a4=3/8, a1=a1−a5=0 and a0=a0=3/4, matching the direct computation. Index 4 has been reflected about 3 onto index 2, with a minus sign.
The interpolant is a different polynomial from the truncation of the same degree, and the difference is measurable in both norms.
Worked example 6.1.14
(Truncation versus interpolation for x4)
The truncation is Π2f=3/8+21T2=x2−1/8, and the two errors are
The aliased component doubles the squared weighted error, consistent with the optimality of Π2f. In the uniform norm, ∣T4/8∣≤1/8 with equality at 0, ±1/2 and ±1. For the interpolation error put s=x2∈[0,1]: s2−43s has minimum −9/64 at s=3/8 and maximum 1/4 at s=1. Hence ∥f−I3f∥∞=1/4, attained at the unsampled endpoints x=±1.
Error estimates under absolutely summable coefficients
Theorem 6.1.15
(Truncation and interpolation errors)
Assume (H). Then for N≥0 and n≥1:
∥f−ΠNf∥w2=2πk>N∑ak2;
∥f−ΠNf∥∞≤k>N∑∣ak∣;
∥f−Inf∥∞≤2k≥n∑∣ak∣;
∥f−Inf∥w≥∥f−Πn−1f∥w.
Proof
(Discard the tail, then fold it back)
Put RN=f−ΠNf=k>N∑akTk. By (H) and ∣Tk∣≤1 this series converges uniformly, which gives (2).
For (1), fix M>N and let RN,M=k=N+1∑MakTk. By continuous orthogonality, ∥RN,M∥w2=2πk=N+1∑Mak2, and ∥RN−RN,M∥w≤πk>M∑∣ak∣→0. Letting M→∞ gives (1); the sum of squares converges because ak2≤(lsup∣al∣)∣ak∣.
For (3), linearity of In and exactness on Pn−1 give Inf=Πn−1f+InRn−1. The coefficients of InRn−1 are finite combinations of the samples Rn−1(xj)=m≥n∑amTm(xj), so InRn−1=m≥n∑amInTm, and ∣InTm∣≤1 by the folding theorem. Therefore
Statement (4) is the Pythagorean identity with q=Inf∈Pn−1.
Interpolation costs at most twice the uniform truncation bound: once for discarding the tail and once for folding it back. The factor 2 cannot be improved in general, as the two-node exercise with f(x)=x3 below shows. Combined with coefficient decay, the theorem gives rates: if ∣ak∣≤C/k2, then k≥n∑k−2≤n−2+∫n∞t−2dt≤2/n gives ∥f−Inf∥∞≤4C/n. This is an upper bound, not a rate every function attains.
Degree, grid and stability are different parameters
The degree (N for ΠN, n−1 for In), the grid where f is observed, and the propagation of observation errors are separate matters. The interpolant ties degree to grid, but nothing forces this. Computing ak on a finer grid of n′ nodes and keeping only 0≤k≤N, with N≤n′−1, gives Q=a0/2+k=1∑NakTk, and under (H)
∥f−Q∥∞≤k>N∑∣ak∣+m≥2n′−N∑∣am∣.
Indeed, an index m≤n′−1 folds to itself and is either kept exactly or discarded, while an index m≥n′ folding onto a retained k≤N is 2n′q±k with q≥1, hence at least 2n′−N. Refinement pushes the aliasing bound into the tail and leaves the truncation term unchanged, but the actual aliasing error need not decrease at every step: with N=0 and f=T2+T4, the sampled constant is 0 for one node and −1 for two nodes, whereas the exact constant coefficient is zero.
For data errors, suppose each sample is perturbed by εj with ∣εj∣≤ε, and let e be any function with e(xj)=εj. The interpolant changes by Ine, whose coefficients have absolute value at most (2/n)⋅nε=2ε, so
∥Ine∥∞≤ε+2(n−1)ε=(2n−1)ε.
Derivatives at the endpoints behave worse. Choose εj=εTn−1(xj); by exactness Ine=εTn−1. Differentiating Tk(cosθ)=coskθ gives Tk′(cosθ)=ksinkθ/sinθ for 0<θ<π, and letting θ→0 gives Tk′(1)=k2. Hence
(Ine)′(1)=ε(n−1)2,
whereas at an interior point ∣Tk′(x)∣≤k/1−x2. So worst-case endpoint derivative amplification is at least (n−1)2; a constant perturbation, by contrast, has zero derivative.
Endpoint singularities
The change of variables removes the artificial jump of a periodic extension. It cannot remove a singularity that f genuinely has at an endpoint.
Worked example 6.1.16
(Chebyshev coefficients of 1+x)
The function f(x)=1+x is continuous on [−1,1] with an unbounded derivative at −1. Using 1+cosθ=2cos2(θ/2) and cos(θ/2)≥0 on [0,π],
g(θ)=2cos2θ,
which behaves like ∣π−θ∣/2 near θ=π: the even periodic function g has a corner there. For the coefficients, use cos(θ/2)coskθ=21(cos(k+21)θ+cos(k−21)θ) and sin((k±21)π)=±(−1)k:
These coefficients are absolutely summable, and g is continuous, even and periodic, so the Fejér argument used earlier shows that (H) holds. However, they decay only like k−2, however smooth f is elsewhere.
The last exercise below shows that the uniform truncation error for this f is exactly 22/(π(2N+1)), attained at x=−1. If the type of endpoint singularity is known, one may subtract it or change variables before applying the method; without such information, adding nodes improves the approximation only slowly.
Common mistakes and subtle points
Common mistake
(Treating the interpolant as the truncation)
For f(x)=x4 and n=3, the discrete coefficient a2=3/8 absorbed the reflected coefficient a4, while a2=1/2. So I3f and Π2f are different polynomials with different errors. Reading discrete coefficients as continuous ones ignores exactly this reflected tail.
Common mistake
(Expecting exactness or uniqueness beyond Pn−1)
Every sample of Tn is zero, so InTn=0 although ∥Tn∥∞=1. Uniqueness in the interpolation theorem is uniqueness inside Pn−1, not among all functions with the given samples, as T5+T1 on three nodes shows.
Common mistake
(Confusing endpoint mismatch with endpoint singularity)
The cosine map makes every continuous f produce a continuous periodic g, so mismatched values f(−1)=f(1) cost nothing. A square-root or other singular behavior at ±1 still becomes a lack of smoothness of g at π or 0. Likewise, the weighted norm weights the endpoints heavily but is not a uniform norm; claims about values or derivatives at ±1 need uniform or derivative estimates.
Exercises
Exercise 6.1.17
(Let n=2 and f(x)=x3. Compute I2f, confirm it with the aliasing formula, and compare ∥f−I2f∥∞ with the bound 2k≥2∑∣ak∣.)
Core
Solution(4)
1
Sample and compute the discrete coefficients
The angles are π/4,3π/4, the nodes ±2/2 and the samples ±2/4. Then
Because: ∣ak∣=42/(π(4k2−1)) and 4/(4k2−1)=2/(2k−1)−2/(2k+1).
2
Evaluate the error at x=−1
Key step
f(−1)−ΠNf(−1)=k>N∑ak(−1)k=−π(2N+1)22.
This reaches the upper bound in statement (2) of the error theorem, so ∥f−ΠNf∥∞=22/(π(2N+1)) exactly.
Because: (H) holds and Tk(−1)=(−1)k, so ak(−1)k=−42/(π(4k2−1)).
3
Solve for N
The condition is 2N+1≥20002/π≈900.32, that is, N≥449.66. The smallest integer is N=450; for N=449, 2N+1=899 is too small. The cause is the corner of g at θ=π, not any interior feature of f.
Related notes
This note builds on Fourier truncation, sampling, and the Gibbs phenomenon: its Fejér theorem identifies a Chebyshev series with the function it came from, and its aliasing formula is the shifted counterpart of the reflected formula proved here. Nodes that include the endpoints, spectral differentiation matrices and sharper decay for functions analytic near the interval need separate arguments and are not covered here.