Derivative Approximation by Finite Differences

David Eberly, Geometric Tools, Redmond WA 98052

https://www.geometrictools.com/
This work is licensed under the Creative Commons Attribution 4.0 International License. To view a copy
of this license, visit http://creativecommons.org/licenses/by/4.0/ or send a letter to Creative Commons,
PO Box 1866, Mountain View, CA 94042, USA.
Created: May 30, 2001
Last Modified: September 21, 2016

Contents

1 Introduction 2

2 Derivatives of Univariate Functions 2

3 Derivatives of Bivariate Functions 7

4 Derivatives of Multivariate Functions 8

1
1 Introduction

This document shows how to approximate derivatives of univariate functions F (x) by finite differences.
Given a small value h > 0, the d-th order derivative satisfies the following equation where the integer order
of error p > 0 may be selected as desired,
iX
hd (d) max

F (x) = Ci F (x + ih) + O(hd+p ) (1)

d! i=i min

for some choice of extreme indices imin and imax and for some choice of coefficients Ci . The equation
becomes an approximation by throwing away the O(hd+p ) term. The vector C = (Cimin , . . . , Cimax ) is called
the template or convolution mask for the approximation. Approximations for the derivatives of multivariate
functions are constructed as tensor products of templates for univariate functions.

2 Derivatives of Univariate Functions

Recall from calculus that the following approximations are valid for the derivative of F (x). A forward
difference approximation is
F (x + h) F (x)
F 0 (x) = + O(h) (2)
h
a backward difference approximation is

F (x) F (x h)
F 0 (x) = + O(h) (3)
h
and a centered difference approximation is

F (x + h) F (x h)
F 0 (x) = + O(h2 ) (4)
2h
The approximations are obtained by throwing away the error terms indicated by the O notation. The order
of the error for each of these approximations is easily seen from formal expansions as Taylor series about the
value x,

0 h2 00 X hn (n)
F (x + h) = F (x) + hF (x) + F (x) + . . . = F (x) (5)
2! n=0
n!
and

0h2 00 X hn
F (x h) = F (x) hF (x) + F (x) + . . . = (1)n F (n) (x) (6)
2! n=0
n!

where F (n) (x) denotes the n-th order derivative of F . Subtracting F (x) from both sides of equation (5 and
then dividing by h leads to the forward difference F 0 (x) = (F (x+h)F (x))/h+O(h). Subtracting F (x) from
both sides of equation (6) and then dividing by h leads to the backward difference F 0 (x) = (F (x) F (x
h))/h + O(h). Both approximations have error O(h). The centered difference is obtained by subtracting
equation (6) from equation (5) and then dividing by 2h to obtain (F (x + h) F (x h))/(2h) + O(h2 ).

2
Higher order approximations to the first derivative can be obtained by using more Taylor series, more terms
in the Taylor series, and appropriately weighting the various expansions in a sum. For example,

X (2h)n (n) X (2h)n (n)
F (x + 2h) = F (x), F (x 2h) = (1)n F (x) (7)
n=0
n! n=0
n!

lead to a forward difference approximation with second order error,

F (x + 2h) + 4F (x + h) 3F (x)
F 0 (x) = + O(h2 ) (8)
2h
to a backward difference approximation with second order error,
3F (x) 4F (x h) + F (x 2h)
F 0 (x) = + O(h2 ) (9)
2h
and to a centered difference approximation with fourth order error,
F (x + 2h) + 8F (x + h) 8F (x h) + F (x 2h)
F 0 (x) = + O(h4 ) (10)
12h
Higher-order derivatives can be approximated in the same way. For example, a forward difference approxi-
mation to F 00 (x) is
F (x + 2h) 2F (x + h) + F (x)
F 00 (x) = + O(h) (11)
h2
and centered difference approximations are
F (x + h) 2F (x) + F (x h)
F 00 (x) = + O(h2 ) (12)
h2
and
F (x + 2h) + 16F (x + h) 30F (x) + 16F (x h) F (x 2h)
F 00 (x) = + O(h4 ) (13)
12h2

Each of these formulas is easily verified by expanding the F (x + ih) terms in a formal Taylor series and
computing the weighted sums on the right-hand sides. However, of greater interest is to select the order of
derivative d and the order of error p and determine the weights Ci for the sum in equation (1). A formal
Taylor series for F (x + ih) is

X hn
F (x + ih) = in F (n) (x) (14)
n=0
n!
Replacing this in equation (1) yields
hd (d)
Pimax P n hn (n)
d! F (x) = i=imin Ci n=0 i n! F (x) + O(hd+p )
 
hn (n)
P Pimax n
= n=0 i=imin i Ci n! F (x) + O(hd+p )
Pd+p1 Pimax n  hn (n) (15)
= n=0 i=i i Ci n! F (x) + O(hd+p )
P min
.
 n
Pd+p1 imax n h (n)
= n=0 i=imin i Ci n! F (x)

Multiplying by d!/hd , the desired approximation is

d+p1 d+p1
iX
! iX
!
(d) d! X max
n hn (n) p . d!
X max
n hn (n)
F (x) = d i Ci F (x) + O(h ) = d i Ci F (x) (16)
h n=0 i=i n! h n=0 i=imin
n!
min

3
Treating the approximation in equation(15) as an equality, the only term in the sum on the right-hand side
of the approximation that contains (hd /d!)F (d) (x) occurs when n = d, so the coefficient of that term must
be 1. The other terms must vanish for there to be equality, so the coefficients of those terms must be 0;
therefore, it is necessary that

iX
max 0, 0 n d + p 1 and n 6= d
in Ci = (17)
1, n = d
i=imin

This is a set of d + p linear equations in imax imin + 1 unknowns. If we constrain the number of unknowns to
be d+p, the linear system has a unique solution. A forward difference approximation occurs if we set imin = 0
and imax = d + p 1. A backward difference approximation occurs if we set imax = 0 and imin = (d + p 1).
A centered difference approximation occurs if we set imax = imin = (d + p 1)/2 where it appears that
d + p is necessarily an odd number. As it turns out, p can be chosen to be even regardless of the parity of d
and imax = b(d + p 1)/2c.
Table 1 indicates the choices for d and p, the type of approximation (forward, backward, centered), and the
corresponding equation number,

Table 1. Choices for the parameters for the example approximations presented previously.

equation d p type imin imax

(2) 1 1 forward 0 1
(3) 1 1 backward -1 0
(4) 1 2 centered -1 1
(8) 1 2 forward 0 2
(9) 1 2 backward -2 0
(10) 1 4 centered -2 2
(11) 2 1 forward 0 2
(12) 2 2 centered -1 1
(13) 2 4 centered -2 2

Example 1. Approximate F (3) (x) with a forward difference with error O(h), so d = 3 and p = 1. We need
imin = 0 and imax = 3. The linear system from equation (17) is

1 1 1 1 C0 0


0 1 2 3 C1 0
=

0 1 4 9 C2 0

0 1 8 27 C3 1

4
and has solution (C0 , C1 , C2 , C3 ) = (1, 3, 3, 1)/6. Equation (1) becomes

F (x) + 3F (x + h) 3F (x + 2h) + F (x + 3h)

F (3) (x) = + O(h)
h3

Approximate F (3) (x) with a centered difference with error O(h2 ), so d = 3 and p = 2. We need imax =
imin = 2. The linear system from equation (17) is

1 1 1 1 1 C2 0

2 1 0 1 2

C1 0

= 0

4 1 0 1 4 C0

8 1 0 1 8

C1 1

16 1 0 1 16 C2 0

and has solution (C2 , C1 , C0 , C1 , C2 ) = (1, 2, 0, 2, 1)/12. Equation (1) becomes

F (x 2h) + 2F (x h) 2F (x + h) + F (x + 2h)
F (3) (x) = + O(h2 )
2h3

Finally, approximate with a centered difference with error O(h4 ), so d = 3 and p = 4. We need imax =
imin = 3. The linear system from equation (17) is

1 1 1 1 1 1 1 C3 0

3 2 1 0 1 2

3 C2 0


9 4 1 0 1 4 9 C 0
1


27 8 1 0 1 8 27 C0 = 1



81 16 1 0 1 16 81 C1 0

243 32 1 0 1 32 243 C2 0


729 64 1 0 1 64 729 C3 0

and has solution (C3 , C2 , C1 , C0 , C1 , C2 , C3 ) = (1, 8, 13, 0, 13, 8, 1)/48. Equation (1) becomes

F (x 3h) 8F (x 2h) + 13F (x h) 13F (x + h) + 8F (x + 2h) F (x + 3h)

F (3) (x) = + O(h4 ).
8h3

Example 2. Approximate F (4) (x) with a forward difference with error O(h), so d = 4 and p = 1. We need

5
imin = 0 and imax = 4. The linear system from equation (17) is

1 1 1 1 1 C 0
0

0 1 2 3 4 C1 0

= 0

0 1 4 9 16 C2


0 1 8 27 64 C3 0

0 1 16 81 256 C4 1
and has solution (C0 , C1 , C2 , C3 , C4 ) = (1, 4, 6, 4, 1)/24. Equation (1) becomes
F (x) 4F (x + h) + 6F (x + 2h) 4F (x + 3h) + F (x + 4h)
F (4) (x) = + O(h)
h4

Approximate F (4) (x) with a centered difference with error O(h2 ), so d = 4 and p = 2. We need imax =
imin = 2. The linear system from equation (17) is

1 1 1 1 1 C2 0

2 1 0 1 2

C1 0

= 0

4 1 0 1 4 C0

8 1 0 1 8

C1 0

16 1 0 1 16 C2 1
and has solution (C2 , C1 , C0 , C1 , C2 ) = (1, 4, 6, 4, 1)/24. Equation (1) becomes
F (x 2h) 4F (x h) + 6F (x) 4F (x + h) + F (x + 2h)
F (4) (x) = + O(h2 )
h4

Finally, approximate with a centered difference with error O(h4 ), so d = 4 and p = 4. We need imax =
imin = 3. The linear system from equation (17) is

1 1 1 1 1 1 1 C3 0

3 2 1 0 1 2

3 C2 0


9 4 1 0 1 4 9 C 0
1


27 8 1 0 1 8 27 C0 = 0



81 16 1 0 1 16 81 C1 1

243 32 1 0 1 32 243 C2 0


729 64 1 0 1 64 729 C3 0
and has solution (C3 , C2 , C1 , C0 , C1 , C2 , C3 ) = (1, 12, 39, 56, 39, 12, 1)/144. Equation (1) becomes

F (x 3h) + 12F (x 2h) 39F (x h) + 56F (x) 39F (x + h) + 12F (x + 2h) F (x + 3h)
F (4) (x) = + O(h4 )
6h4

6
3 Derivatives of Bivariate Functions

For functions with more variables, the partial derivatives can be approximated by grouping together all of
the same variables and applying the univariate approximation for that group. For example, if F (x, y) is our
function, then some partial derivative approximations are
. F (x+h,y)F (xh,y)
fx (x, y) = 2h
. F (x,y+k)F (x,yk)
fy (x, y) = 2k
. F (x+h,y)2F (x,y)+F (xh,y)
(18)
fxx (x, y) = h2
. F (x,y+k)2F (x,y)+F (x,yk)
fyy (x, y) = k2
. F (x+h,y+k)F (x+h,yk)F (xh,y+k)+F (xh,yk)
fxy (x, y) = 4hk

Each of these can be verified in the limit, the x-derivatives by taking the limit as h approaches zero, the
y-derivatives by taking the limit as k approaches zero, and the mixed second-order derivative by taking the
limit as both h and k approach zero.
The derivatives Fx , Fy , Fxx , and Fyy just use the univariate approximation formulas. The mixed derivative
requires slightly more work. The important observation is that the approximation for Fxy is obtained by
applying the x-derivative approximation for Fx , then applying the y-derivative approximation to the previous
approximation. That is,
. F (x+h,y)F (xh,y)
fxy (x, y) = 2h
. F (x+h,y+k)F (xh,y+k)

F (x+h,yk)F (xh,yk)
(19)
= 2h
2k
2h

F (x+h,y+k)F (x+h,yk)F (xh,y+k)+F (xh,yk)

= 4hk

The approximation implied by equation (1) may be written as

imax
hm dm . X (m)
F (x) = Ci F (x + ih) (20)
m! dxm i=i min

The inclusion of the superscript on the C coefficients is to emphasize that those coefficients are constructed
for each order m. For bivariate functions, we can use the natural extension of equation (20) by applying
the approximation in x first, then applying the approximation in y to that approximation, just as in our
example of Fxy ,

kn n hm m . n
n
Pimax (m)
n! y n m! xm F (x, y) = kn! y n i=imin Ci F (x + ih, y)
. P imax P jmax (m) (n)
(21)
= i=imin j=jmin Ci Cj F (x + ih, y + jk)
Pimax Pjmax (m,n)
= i=imin j=jmin Ci,j F (x + ih, y + jk)

where the last equality defines

(m,n) (m) (n)
Ci,j = Ci Cj (22)
The coefficients for the bivariate approximation are just the tensor product of the coefficients for each of the
univariate approximations.

7
4 Derivatives of Multivariate Functions

The approximation concept extends to any number of variables. Let (x1 , . . . , xn ) be those variables and let
F (x1 , . . . , xn ) be the function to approximate. The approximation is
max max
i1 in
hm m1 hm mn
 1 n

1 n . X X (m1 ,...,mn )
m1 F (x1 , . . . , xn ) = C(i1 ,...,i F (x1 + i1 h1 , . . . , xn + in hn ) (23)
m1 ! x1 mn ! xm
1
n n)
i1 =imin
1
in =imin
n

where
(m ,...,m )
1 n (m1 ) (mn )
C(i1 ,...,in)
= Ci1 Cin (24)
is a tensor product of the coefficients of the n univariate approximations.