JOURNAL OF APPROXIMATION THEORY 30, 59-70 (1980)

Convergence of Numerical Differentiation

R. A. LORENTZ

Gesellschaft f&r Mathematik und Datencerarbeitung,

Postfach 1240, 5205 St. Augustin I, German Federal Republic

Communicated by Oved Shisha

Received April 16, 1979

Let D, be the functional given by D,,f = f’(0) on C’( - I, I). Let n, be the set
of polynomials of degree not exceeding n and let M, be the polynomial inter-
polation to fat a given set of points x1, x2 ,..., x,, We approximate DOf by
D&J This is called a numerical differentiation formula. We study the point-
wise convergence of DJvf,, to D, for two choices of the set of points: for equi-
spaced points and for the extrema of the Chebycheff polynomials.

Let C = C[- 1, l] be the Banach space of continuous real-valued func-

tions on the interval [-I, I] with the Chebycheff norm and let Ci denote the
space of i-times continuously differentiable functions. By numerical differen-
tiation we mean the replacement of the derivative evaluated at the point zero

by the approximate expression

where -- I -.:<x1 < x2 < ... < x, .< 1 and the 3(i are real. This will be called
a differentiation formula. The justification of the term “approximate” lies in
our demand that DJ = Ff for all polynomials f of degree not exceeding
I? ~ I
FP == D,P, P ELITLpl.

Given the points xi , this determines the coefficients myiuniquely. In fact,

if K is any functional defined for all polynomials and if xi, i =-- l,..., n, are
given, then the equality

0021-9045/80:090059-12$02.00i0
CoppriEht G’ 1980 by Academic Press, Inc.
All rights of reproduction in any form reserved.
60 Ii. A. LORENTZ

for all P t n,_, determines the coefficients uniquely. Let L,, ,,,(s, ..__..I-,?: .v)
be the Lagrange interpolating polynomial:

L,(Xl ,.-.’ s,, ; Xi) s,, , i, ,j =- 1 )..., II.

Then
ai2 -2 f&(x, ,...) x, ; .). (1)

We will consider two choices of the nodes {x,1. The first is the easiest
choice made using information neither on the functional to be approximated
nor of the set on which the approximation is exact. This is

xi = i/2m, i 0, iI ,..., +?I. (2)

The second choice is the set of extremal points of the Chebycheff polynomial
of degree 2m - I, TznL+l .
If we let

Tz,,Lm,(x) m= cos((2m ~ 1) cosr’x;, -I :< x > I.

and
n(i + m)
-xi -= -- cos 2m-I ’
-yo 0, (3)
i7(i $ m - 1)
x, ~~ -~ cos i ~~ 1, 2,. . ., m.
2m-1 ’

then the xi, i -F 0, are the extrema of T2,,P1 and

Tz,,,+,(xi) = (-l)ilm~l, i = -myI,..., -- 1,

EL,,, l(xi) -7 (- I)i ’ “l) i =- I,..., m.

In either case, the functional

(4)

where the Lzt are the point evaluation fimctionals, which reproduces D,,
exactly on Il,, , has coefficients given by (1) if we replace K by D,
As a functional on C,
i,,
 NUMERICAL DIFFERENTIATION 61

It turns out that the choice of the Chebycheff extrema minimizes the norm of
F,, among all choices of nodes. As a consequence, this choice yields a differen-
tiation formula which is the least sensitive to uncertainty in the given informa-
tion f(xi). For more detail on this aspect see Pallaschke [4].
It is another aspect, however, which is of primary interest to us, namely,
the convergence of FZm(f) to f’(o) as m increases.
Pallaschke, [3], has proved that the differentiation formula with the
Chebycheff nodes (3) converges tof’(o) for anyfwhich is twice continuously
differentiable. We will show here that this formula even converges for once
differentiable functions. We then compare this behavior with that of the
differentiation formula with equi-spaced nodes (2) which are worse behaved.
It is shown that for equi-spaced nodes the formula also converges for three
times continuously differentiable functions but diverges for at least one func-
tion not in this class.
The following three theorems represent our main results.

THEOREM I. Let

where the xi are the Chebycheff nodes (3). Then for all f E C,[- 1, I],

L(f) -+f’(O>
for m + 00.

THEOREM 2. Let FZmbe given as above, however, 11Yth

xi = i/2m, i = -m, -m -1 I,..., m.

Then
&m(f) -f’(O)
for any f satisfying the Dini-Lipschitz condition for its derivative:

~(f’, n-l) log n + 0.

THEOREM 3. Let FZ, be as in Theorem 2. Then there is somef E C,[-- 1, 11,

so that
~‘df) +f’(O>.
One may resume these results in saying that the numerical differentiation
formula for equi-spaced nodes is really worse than that for Chebycheff nodes
but not so bad after all since it converges for functions satisfying a relatively
weak condition on the derivative.
62 R. A. LORENTZ

We begin a general analysis of these differentiation formulas following,

in part, Pallaschke [3]. In the equi-spaced case,
d

for i # 0, and 01~= 0.

Setting in the values for the nodes

For the case of Chebycheff nodes, we use the following representation of

Lagrange interpolant to the function f at the Chebycheff nodes, which can be
found in Rivlin [5].

Thus
 NUMERICAL DIFFERENTIATION 63

Since

Now we use the fact that LBmml,i is a polynomial of degree 2m - 1 itself

and so

Since

and since

we get

117 --I- 2,..., - 1,

k, -- (2m - l))l(- I)i+l~;z, i 7: I, 2,..., m - 1,

(7)
*,,i - ;(2177~ I)-](--- 1)',1?'.
Now using (3). we have an explicit formula for F2,,, :

F,,,,(f) :: (2m - I)-' ;(-I)"",f(-l) + s(-l)""+'f(J)

1
-t j, (-l)i [cos ;~2-r”l)]-L+OS~~]

Note that we have interpolated from 172,,Lm1 because L;,,,-,,,(O) == L&,,i(0)

and so the numerical differentiation formulas from n,,, and II,,,~ 1 are the
same.
In order to prove Theorems 1 and 3, we consider not the functionals
F.,,,, but rather the composition F& of F?,,, with the operator of integration

We do this because of the following lemma:

64 K. A. LOKm’lz

LEMMA -1. Let 7-,,lx> (I .wqurnc~ offiwtcfional~~ on C[ I. 11

Then
T,L(.f) + ,f’(()) (8)
Jbr ull f E C, if cd otll)2if
T,P -+ P’(0) (9)
for my polynotrticit i” cm/ fhcre csists utl M 0 such thut
/ ‘TJ 1’ =: M. t1 E, 2,... (IO)
PSOOJ Necessity: Let f E C. Then SJ’E C, Thus
T,,,(Sf) + (Sj)‘(W f(O).
An M satisfying (10) exists since point-wise boundedness implies uniform
boundedness. Obviously (9) is satisfied also.
Sufficiency: Assume (9) and (10) and let P ,SQ, \\here Q is a polynomial.
Then

By the Banach-Steinhaus theorem again, we may conclude that

(T,r,,Sk p s(O)
for any g ;I C. since the poiynomials are dense in C. But anyf’t C, may bc
written as j‘ S<qfor sotne g E C. so that
r,,J‘ -+ g(O) f“(O).
in considering the convergence of difltreniiation formulas. we can thus, by
using Lemtna 4. restict ourselves to the functionals TJ. Note also that this
lemma is independent of the nodes.
Now, also for both our choices of nodes
 NlJhlEKICAL DlFFERE:NTIATION 65

and so

Now since the I;‘-?,,are exact on n.,,,, ,

From (6) and (7), we see that

Because

We want to continue with exact formulas for 1:F2,,,S I/. To do this, we show
that

First we observe that for Chebycheff nodes,

and that the ‘P, alternate in sign. This is clear from (6). For the equi-spaced
case. we consider

Since both factors of the right hand side are smaller than one, here also the .x,
decrease in absolute value with increasingj and alternate in sign.
By re-grouping,
66

one sees that

and

Thus we can write

We are now ready for the

Proof of Thorem 3
Now xj 1;2m so that

The last term is -+!,or z~,,-~depending on which ever index is odd.

Note that this formula is exact. It enables us to obtain upper and lower
bounds on 11E1,,:,,,,S
‘I.
Obviously
 NUMERICAL DIFFERENTIATION 67

and because

Upper and lower bounds for 11E2,,&S~/are thus

From the explicit formula for aii ,

m vl(tl- l)...(m-i-+ 1)
~_ (-l)i'l
'%
i (m +1)(/n A~ 2) ... (m ~- i) ’

it follows that for fixed i,

as m ---f 03. It follows immediately that

From Lemma 4, we may conclude that there exists anfE C, for which

Fmf +f ‘0
which concludes the proof.

Proof of Theorem 1
Using (13) and (14),

i nz (Xi- Xi-,) )
<2) (22 1,x-+;* (2m- 1)x3

640!30/1-6
 R. A. LORENTZ

Hence

Now

x, - (-0s ~---m7r 57
2m -- I s’n >jzrn - 1) .

For 0 b .Y .. 4,

sin .Y : ; .Y - x3, 6

from which

Thus

which together with Lemma 4 proves Theorem 1.

To prove Theorem 2, we find an upper bound for ,i ki,, ,,. We can use the
previous formulas for xi in the case of equi-spaced nodes to obtain

and

I m’)
, . i --. I.
i m
 NUMERICAL DIFFERENTIATION 69

Thus

Now we may use the triangle inequality in the usual way

Thus

I E2w.f - .f’ .< {II F*,S I! -L- I ) . $lf, l:.f’ - P2,>,~1

-G E,,-,(.f’)(log m t- I).

1f.f’ satisfies the Dini-Lipschitz condition, then Fzqllf converges tof’.

In concluding, it is perhaps of interest to compare these results, which show
convergence at a point, with a recent result of Haverkamp [I] who analysed
the uniform convergence of the differentiation formula for Chebycheff nodes.
He obtained

THEOREM 5. Let Prim(f) be the polynomial in 112,,,which interpolates f at

the ChebwhqJf’nodes. Then for f E Cl

.f’ ~- P~llL(.f)i; 5’ 4[ln 2171 i 21 E,,, ,(f’),

i.s thr degree of approximation to ffrom Il’, and jl . /I is the uniform norm,
If we know thatf’ E Lip O(for some (Y ‘: 0, then

E,(f’) :< M/P

for some iM 0. Thus Pim(f) converges to f’ uniformly, If, however,

nothing is known about f’ other than that is continuous, Pi,(f) need not
converge to f’ since I&&‘) may decrease arbitrarily slowly. Theorem 1 shows
that even for arbitrary functions in C, , convergence at a point is possible.
70 R. A. LORENTZ

REFERENCES

1. R. HAVERKAMP,Eine Formel zur numerischen Differentiation. J. A/q~o.\-. T/~c’orjs23

(1978), 261-266.
2. G. G. LOREUTZ.“Approximation of Functions.” Holt, Rinehart & Winston. New York.
1966.
3. D. PALLAKHKL‘, Konvergenz optimaler Differentiationsformeln, Nu~r. Murh. 27 (1977).
421-426.
4. D. PALLASCHKE, Optimale Differentiations- und lntegrationsformeln in C,(rr, /I), NU?U~I
Murh. 26 (1976), 201-210.
5. T. J. RIVLIN, “The Chebyshev Polynomials,” Wiley, New York, 1974.
6. T. J. RIVLIN, Optimally stable numerical differentiation, SIAM J. .Yunw. Ami. 12
(1975). 712-725.