Accepted Manuscript

A new and sharper bound for Legendre expansion of differentiable

functions

Haiyong Wang

PII: S0893-9659(18)30164-2
DOI: https://doi.org/10.1016/j.aml.2018.05.022
Reference: AML 5529

To appear in: Applied Mathematics Letters

Received date : 9 April 2018

Revised date : 24 May 2018
Accepted date : 24 May 2018

Please cite this article as: H. Wang, A new and sharper bound for Legendre expansion of
differentiable functions, Appl. Math. Lett. (2018), https://doi.org/10.1016/j.aml.2018.05.022

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
*Manuscript
Click here to view linked References

A new and sharper bound for Legendre expansion of

differentiable functions
Haiyong Wang†‡
May 24, 2018

Abstract
In this paper, we present a new and sharper bound for the Legendre coefficients of
differentiable functions and then provide a new bound on the approximation error of the
truncated Legendre series in the uniform norm. An illustrative example is provided to
demonstrate the sharpness of our new results.
Keywords: Legendre coefficient, differentiable functions, sharp bound.
AMS classifications: 41A25, 41A10

1 Introduction
The set of Legendre polynomials {P0 (x), P1 (x), · · · } form a system of polynomials orthogonal on
the interval [−1, 1] with respect to the weight function ω(x) = 1 and
Z 1
Pn (x)Pm (x)dx = hn δmn , (1.1)
−1

where δmn is the Kronecker delta and

µ ¶−1
1
hn = n+ . (1.2)
2
The Legendre expansion of a function f := [−1, 1] → R is defined by
∞
X Z 1
f (x) = an Pn (x), an = h−1
n f (x)Pn (x)dx. (1.3)
n=0 −1

The problem of estimating the magnitude of the Legendre coefficients an is of particular interest
both from the theoretical and numerical point of view. Indeed, it is useful not only in under-
standing the rate of convergence of Legendre expansion but useful also in estimating the degree of
the Legendre polynomial approximation to f (x) within a given accuracy. When f (x) is analytic
in a neighborhood of the interval [−1, 1], we note that the estimate of the Legendre coefficients,
or more generally, the Gegenbauer and Jacobi coefficients, has been studied in [6, 7, 8, 9].
† School of Mathematics and Statistics, Huazhong University of Science and Technology, Wuhan 430074, P. R.

China. E-mail: haiyongwang@hust.edu.cn

‡ Hubei Key Laboratory of Engineering Modeling and Scientific Computing, Huazhong University of Science

and Technology, Wuhan 430074, China.

1
 In this work, we are interested in the case where f (x) is a differentiable function. We first
establish a new bound for the Legendre coefficients and the key ingredient here is that the Leg-
endre polynomial satisfies a sharp Bernstein-type inequality. An illustrative example is provided
to show that our new result is sharper than the result given in [6, Theorem 2.1]. Furthermore,
we derive a new bound on the approximation error of the truncated Legendre expansion in the
uniform norm. Our main results are stated in the next section.
Remark 1.1. Throughout the paper, the derivative of a function f (x) is defined in the distri-
butional sense. If f (x) is continuously differentiable, then the distributional derivative of f (x)
coincides with its classical derivative. We refer to [4, Chapter 7] for more discussions.

2 A new and sharper bound for Legendre coefficients of

differentiable functions
In this section we state a new and explicit bound for the Legendre coefficients of differentiable
functions. Before proceeding, we first define the weighted semi-norm
Z 1
|f 0 (x)|
kf k := dx. (2.1)
2 1
−1 (1 − x ) 4

The following Bernstein-type inequality of Legendre polynomials will be useful.

Lemma 2.1. For x ∈ [−1, 1] and n ≥ 0, we have
r µ ¶− 21
1 2 1
(1 − x2 ) 4 |Pn (x)| < n+ . (2.2)
π 2
1
Moreover, the bound is optimal in the sense that the factor (n + 12 )− 2 can not be improved to
1 p
(n + 12 + ²)− 2 for any ² > 0 and the constant 2/π is best possible.
Proof. See [1, 2].

Figure 1: The ratio of the term on the left-hand side to the term on the right-and side of (2.2)
for n = 2 (left), n = 6 (middle) and n = 18 (right).

Figure 1 shows the ratio of the term on the left-hand side to the term on the right-hand side
of (2.2) for three values of n. One can see that the maximum value of the ratio is very close to
one, which shows that the Bernstein-type inequality (2.2) is sharp.
We are now ready to state a new and sharper bound for the Legendre coefficients.

2
Theorem 2.2. Assume that f, f 0 , . . . , f (m−1) are absolutely continuous and the mth derivative
f (m) (x) is of bounded variation. Furthermore, assume that Vm = kf (m) k < ∞. Then, for
n ≥ m + 1,
m
Y
2Vm
|an | ≤ p hn−k . (2.3)
π(2n − 2m − 1) k=1

where hn is defined as in (1.2) and the product is assumed to be one when m = 0.

Proof. The basic idea of our proof is to employ integration by parts and the Bernstein-type
inequality in Lemma 2.1. By combining [3, Equation 18.9.7] and [3, Eqation 18.9.19], we have
Pn+1 0 (x) − Pn−1 0 (x)
Pn (x) = , n ≥ 1. (2.4)
2h−1
n

Applying integration by part once, we obtain

Z 1
an = h−1
n f (x)Pn (x)dx
−1
Z 1
Pn+1 0 (x) − Pn−1 0 (x)
= f (x) dx
−1 2
· ¸1 Z 1
Pn+1 (x) − Pn−1 (x) Pn−1 (x) − Pn+1 (x)
= f (x) + f 0 (x) dx. (2.5)
2 −1 −1 2

Furthermore, by making use of Pn (±1) = (±1)n for each n ≥ 0, it is easy to see that the first
term in the last equation vanishes and therefore
Z 1
Pn−1 (x) − Pn+1 (x)
an = f 0 (x) dx. (2.6)
−1 2
This together with the result of Lemma 2.1 gives
Z 1 0 " 1 1
#
|f (x)| 2(1 − x2 )− 4 2(1 − x2 )− 4
|an | ≤ p +p dx
−1 2 π(2n − 1) π(2n + 3)
Z 1
2 |f 0 (x)|
≤p 1 dx
π(2n − 1) −1 (1 − x2 ) 4
2V0
=p .
π(2n − 1)
This proves the case m = 0.
When m = 1, integrating by part to (2.6) again, we get
Z 1 0 " #
f (x) Pn 0 (x) − Pn−2 0 (x) Pn+2 0 (x) − Pn 0 (x)
an = − dx
−1 2 2h−1
n−1 2h−1
n+1
" #1 " #1
0 Pn (x) − Pn−2 (x) 0 Pn+2 (x) − Pn (x)
= f (x) − f (x)
4h−1
n−1 4h−1
n+1
−1 −1
Z 1 " #
00 Pn−2 (x) Pn (x) Pn (x) Pn+2 (x)
+ f (x) − −1 − −1 + dx. (2.7)
−1 4h−1
n−1 4hn−1 4hn+1 4h−1
n+1

3
We see that the first two terms in the last equation vanish and therefore
Z 1 " #
00 Pn−2 (x) Pn (x) Pn (x) Pn+2 (x)
an = f (x) − −1 − −1 + dx. (2.8)
−1 4h−1
n−1 4hn−1 4hn+1 4h−1
n+1

By using the Bernstein-type inequality in Lemma 2.1 again, we obtain

Z 1 "
|f 00 (x)| h hn−1
|an | ≤ 1 p n−1 + p
2 2 π(2n − 3) 2 π(2n + 1)
−1 (1 − x ) 4
#
hn+1 hn+1
+ p + p dx
2 π(2n + 1) 2 π(2n + 5)
Z 1
2hn−1 |f 00 (x)|
≤p 1 dx
π(2n − 3) −1 (1 − x2 ) 4
2V1
=p hn−1 , (2.9)
π(2n − 3)
where we have used the property that hn is strictly decreasing with respect to n in the second
step and this proves the case m = 1.
When m ≥ 2, we may continue the above process and this brings in higher derivatives of f
and corresponding higher variations up to Vm . Hence we can obtain the desired result.
How sharp is Theorem 2.2? We consider the following example
f (x) = |x − t|, (2.10)
where t ∈ (−1, 1). It is easy to see that this function is absolutely continuous on [−1, 1] and its
derivative in the distributional sense is given by
(
1, x ≥ t,
0
f (x) = 2H(x − t) − 1 =
−1, x < t.
where H(x) is the Heaviside step function. Furthermore, the second order derivative of f (x) in
the distributional sense is given by
f 00 (x) = 2δ(x − t),
where δ(x) is the Dirac delta function. It is easily verified that f 0 (x) is of bounded variation and
Z 1
2δ(x − t) 2 − 41
V1 = kf 0 k = 1 dx = 2(1 − t ) < ∞.
2
−1 (1 − x ) 4

Hence, all assumptions of Theorem 2.2 are clearly satisfied with m = 1 and we then obtain
µ ¶−1 1 µ ¶−1
2V1 1 4(1 − t2 )− 4 1
|an | ≤ p n− =p n− , n ≥ 2. (2.11)
π(2n − 3) 2 π(2n − 3) 2
Let B1 (n) denote the bound on the right-hand side of (2.11). We compare B1 (n) with the absolute
values of the Legendre coefficients |an | and numerical results are illustrated in Figure 2 for two
values of t. In the case of t = 0, i.e., f (x) = |x|, we shall only consider the case that n is even
since an = 0 when n is odd. We can see clearly that our bound B1 (n) is almost indistinguishable
with |an | as n increases. In fact, when n = 400, we have B1 (n) ≈ 0.0002000963242 and |an | ≈
0.0001991004306, which shows that our bound is rather sharp.

4
 6
Figure 2: Log-log plot of the bound B1 (n) (line) and |an | (dots) for t = 0 (left) and t = 7 (right).
Here n ranges from 5 to 400.

Remark 2.3. In [6, Theorem 2.1], an upper bound of the Legendre coefficients was given by
r
Vbm π
|an | ≤ , (2.12)
(n − 12 )(n − 32 ) · · · (n − m + 12 ) 2(n − m − 1)
where n ≥ m + 2 and Z 1
|f (m+1) (x)|
Vbm = 1 dx.
−1 (1 − x2 ) 2
Comparing the new bound on the right-hand side of (2.3) with the bound on the right-hand side
of (2.12) we see that, for n ≥ m + 2, the ratio of the former to the latter is
r
2Vm 2n − 2m − 2 2Vm
< .
π Vbm 2n − 2m − 1 π Vbm
Moreover, it is easy to see that Vm ≤ Vbm and thus the ratio is always less than 2/π ≈ 0.64. This
means that the new bound in (2.3) is at least 36% smaller than the bound in (2.12).
We now consider the Legendre polynomial approximation which is defined by truncating the
first N terms of (1.3), i.e.,
N
X −1
fN (x) = an Pn (x). (2.13)
n=0

We are ready to establish a new bound on the approximation error of fN (x) in the uniform norm.
To this end, let kf k∞ = max−1≤x≤1 |f (x)|. The following theorem is a corollary of Theorem 2.2.
Theorem 2.4. Under the assumptions of Theorem 2.2 and assume that m ≥ 1.
• When m = 1, then for each N ≥ 3,
4V1
kf (x) − fN (x)k∞ ≤ p . (2.14)
π(2N − 5)

5
 • When m ≥ 2, then for each N ≥ m + 1,
Ym
2Vm
kf (x) − fN (x)k∞ ≤ p hN −k . (2.15)
(m − 1) π(2N − 2m − 1) k=2

Proof. Recall the inequality |Pn (x)| ≤ 1 for x ∈ [−1, 1], we have
∞
X
kf (x) − fN (x)k∞ ≤ |an |. (2.16)
n=N

We first consider the case m = 1. By using Theorem 2.2, we obtain

∞
X ∞
2Vm hn−1 2Vm X 1
kf (x) − fN (x)k∞ ≤ p =√ q . (2.17)
n=N
π(2n − 3) 2π n=N (n − ) n − 3
1
2 2

Note that
∞
X ∞
X Z ∞ µ ¶− 32
1 1 3 2
q ≤ 3 32
≤ x− dx = q .
(n − 21 ) n − 3 (n − 2) N −1 2 N − 52
n=N 2 n=N

This together with (2.17) gives the desired result.

Next, we consider the case m ≥ 2. Combining (2.16) and Theorem 2.2, we obtain
∞
X Ym
2Vm
kf (x) − fN (x)k∞ ≤ p hn−k
n=N
π(2n − 2m − 1) k=1
X∞ Y m
2Vm
≤ p hn−k . (2.18)
π(2N − 2m − 1) n=N k=1

By using the following decomposition

m
"m m−1
#
Y 1 Y Y
hn−k = hn−k − hn−k ,
m−1
k=1 k=2 k=1

we get
∞ Y
m ∞
" m m−1
# m
X 1 X Y Y 1 Y
hn−k = hn−k − hn−k = hN −k . (2.19)
m−1 m−1
n=N k=1 n=N k=2 k=1 k=2

Substituting (2.19) into (2.18) gives the desired result. This completes the proof.
An interesting question is: What is the error bound of Legendre approximation to the function
f (x) = |x|? Note that the analysis of Chebyshev polynomial approximations to this function
has been discussed in [5, Chapter 7]. Here we provide a corresponding result for the truncated
Legendre series.
Corollary 2.5. Let f (x) = |x| and let fN (x) be the truncated Legendre series of f (x). Then,
for each N ≥ 3,
8
kf (x) − fN (x)k∞ ≤ p . (2.20)
π(2N − 5)

6
 10 -1

10 -2
Error

10 -3

10 1 10 2 10 3
N

Figure 3: Log-log plot of N −1 (line) and kf − fN k∞ (dots) for f (x) = |x|. Here N ranges from
10 to 1000.

Proof. Note that m = 1 and V1 = 2 for this function. The bound follows immediately from
Theorem 2.4.

Remark 2.6. The result in Corollary 2.5 is actually overestimated. In fact, numerical experiments
show that kf (x) − fN (x)k∞ = O(N −1 ) as N → ∞; see Figure 3. This rate is the same as that
of the truncated Chebyshev series and the Chebyshev interpolant (see [5, Page 53]). However, a
rigorous proof for this observation is still open.

3 Conclusions
In this paper, we have presented a new and sharper bound for the Legendre coefficients of
differentiable functions. As a direct consequence, we further obtain a new error bound on the
truncated Legendre series in the uniform norm.
Finally, we remark that it is possible to extend the result of Theorem 2.2 to more general
cases. Indeed, from [3, Equation 18.14.7] we see that the Gegenbauer polynomial also satisfies a
Bernstein-type inequality, e.g.,

λ 21−λ
(1 − x2 ) 2 |Cnλ (x)| < (n + λ)λ−1 , n ≥ 0,
Γ(λ)

where x ∈ [−1, 1] and 0 < λ < 1 and Cnλ (x) denotes the Gegenbauer polynomial of degree n.
Therefore, one can expect that a sharp bound for the Gegenbauer coefficients of differentiable
functions can be obtained in a similar way.

Acknowledgement
This work was supported by the National Natural Science Foundation of China under grant
11671160. The author would like to thank anonymous referees for their helpful comments.

7
References
[1] V. A. Antonov and K. V. Holsevnikov, An estimate of the remainder in the expansion of
the generating function for the Legendre polynomials (Generalization and improvement of
Bernstein’s inequality), Vestnik Leningrad University Mathematics, 13, 163–166, 1981.
[2] L. Lorch, Alternative proof of a sharpened form of Bernstein’s inequality for Legendre
polynomials, Applicable Analysis, 14, 237–240, 1983.
[3] F. W. J. Olver, D. W. Lozier, R. F. Boisvert and C. W. Clark, NIST Handbook of Mathe-
matical Functions, Cambridge University Press, 2010.
[4] B. D. Reddy, Introductory Functional Analysis: With applications to boundary value prob-
lems and finite elements, Texts in Applied Mathematics 27, Springer-Verlag, New York,
1998.
[5] L. N. Trefethen, Approximation Theory and Approximation Practice, SIAM, 2013.
[6] H. Wang and S. Xiang, On the convergence rates of Legendre approximation, Mathematics
of Computation, 81(278), 861–877, 2012.
[7] H. Wang, On the optimal estimates and comparison of Gegenbauer expansion coefficients,
SIAM Journal on Numerical Aanlysis, 54(3), 1557–1581, 2016.
[8] S. Xiang, On error bounds for orthogonal polynomial expansions and Gauss-type quadra-
ture, SIAM Journal on Numerical Analysis, 50(3), 1240–1263, 2012.
[9] X. Zhao, L. Wang and Z. Xie, Sharp error bounds for Jacobi expansions and Gegenbauer–
Gauss quadrature of analytic functions, SIAM Journal on Numerical Analysis, 519(3),
1443–1469, 2013.