Divided differences

In mathematics, divided differences is an algorithm, historically used for computing tables of logarithms and trigonometric functions. Charles Babbage's Difference engine, an early mechanical calculator, was designed to use this algorithm in its operation.[1]

Divided differences is a recursive division process. The method can be used to calculate the coefficients in the interpolation polynomial in the Newton form.

Definition

Given k+1 data points

(x_0, y_0),\ldots,(x_{k}, y_{k})

The forward divided differences are defined as:

[y_\nu] := y_\nu, \qquad \nu \in \{ 0,\ldots,k\}
[y_\nu,\ldots,y_{\nu+j}] := \frac{[y_{\nu+1},\ldots , y_{\nu+j}] - [y_{\nu},\ldots , y_{\nu+j-1}]}{x_{\nu+j}-x_\nu}, \qquad \nu\in\{0,\ldots,k-j\},\ j\in\{1,\ldots,k\}.

The backward divided differences are defined as:

[y_\nu] := y_{\nu},\qquad \nu \in \{ 0,\ldots,k\}
[y_\nu,\ldots,y_{\nu-j}] := \frac{[y_\nu,\ldots , y_{\nu-j+1}] - [y_{\nu-1},\ldots , y_{\nu-j}]}{x_\nu - x_{\nu-j}}, \qquad \nu\in\{j,\ldots,k\},\ j\in\{1,\ldots,k\}.

Notation

If the data points are given as a function ƒ,

(x_0, f(x_0)),\ldots,(x_{k}, f(x_{k}))

one sometimes writes

f[x_\nu] := f(x_{\nu}), \qquad \nu \in \{ 0,\ldots,k \}
f[x_\nu,\ldots,x_{\nu+j}] := \frac{f[x_{\nu+1},\ldots , x_{\nu+j}] - f[x_\nu,\ldots , x_{\nu+j-1}]}{x_{\nu+j}-x_\nu}, \qquad \nu\in\{0,\ldots,k-j\},\ j\in\{1,\ldots,k\}.

Several notations for the divided difference of the function ƒ on the nodes x0, ..., xn are used:

[x_0,\ldots,x_n]f,
[x_0,\ldots,x_n;f],
D[x_0,\ldots,x_n]f

etc.

Example

For the first few values of \nu, this yields:


\begin{align}
  \mathopen[y_0] &= y_0 \\
  \mathopen[y_0,y_1] &= \frac{y_1-y_0}{x_1-x_0} \\
  \mathopen[y_0,y_1,y_2]
&= \frac{\mathopen[y_1,y_2]-\mathopen[y_0,y_1]}{x_2-x_0}
 =  \frac{\frac{y_2-y_1}{x_2-x_1}-\frac{y_1-y_0}{x_1-x_0}}{x_2-x_0}
 = \frac{y_2-y_1}{(x_2-x_1)(x_2-x_0)}-\frac{y_1-y_0}{(x_1-x_0)(x_2-x_0)}
\\
  \mathopen[y_0,y_1,y_2,y_3] &= \frac{\mathopen[y_1,y_2,y_3]-\mathopen[y_0,y_1,y_2]}{x_3-x_0}
\end{align}

To make the recursive process more clear, the divided differences can be put in a tabular form:


\begin{matrix}
x_0 & y_0 = [y_0] &           &               & \\
        &       & [y_0,y_1] &               & \\
x_1 & y_1 = [y_1] &           & [y_0,y_1,y_2] & \\
        &       & [y_1,y_2] &               & [y_0,y_1,y_2,y_3]\\
x_2 & y_2 = [y_2] &           & [y_1,y_2,y_3] & \\
        &       & [y_2,y_3] &               & \\
x_3 & y_3 = [y_3] &           &               & \\
\end{matrix}

Properties

(f+g)[x_0,\dots,x_n] = f[x_0,\dots,x_n] + g[x_0,\dots,x_n]
(\lambda\cdot f)[x_0,\dots,x_n] = \lambda\cdot f[x_0,\dots,x_n]
(f\cdot g)[x_0,\dots,x_n] = f[x_0]\cdot g[x_0,\dots,x_n] + f[x_0,x_1]\cdot g[x_1,\dots,x_n] + \dots + f[x_0,\dots,x_n]\cdot g[x_n]
f[x_0, \dots, x_n] = f[x_{\sigma(0)}, \dots, x_{\sigma(n)}]
f[x_0,\dots,x_n] = \frac{f^{(n)}(\xi)}{n!}

Matrix form

The divided difference scheme can be put into an upper triangular matrix. Let T_f(x_0,\dots,x_n)=
\begin{pmatrix}
f[x_0] & f[x_0,x_1] & f[x_0,x_1,x_2] & \ldots & f[x_0,\dots,x_n] \\
0 & f[x_1] & f[x_1,x_2] & \ldots & f[x_1,\dots,x_n] \\
\vdots & \vdots & \vdots & \ddots & \vdots \\
0 & 0 & 0 & \ldots & f[x_n]
\end{pmatrix}.

Then it holds

This follows from the Leibniz rule. It means that multiplication of such matrices is commutative. Summarised, the matrices of divided difference schemes with respect to the same set of nodes form a commutative ring.
\delta_\xi(t) = \begin{cases}1 &: t=\xi , \\0 &: \mbox{else}.\end{cases}
Obviously f\cdot \delta_\xi = f(\xi)\cdot \delta_\xi, thus \delta_\xi is an eigenfunction of the pointwise function multiplication. That is T_{\delta_{x_i}} x is somehow an "eigenmatrix" of T_f x:  T_f x \cdot T_{\delta_{x_i}} x = f(x_i) \cdot T_{\delta_{x_i}} x . However, all columns of T_{\delta_{x_i}} x are multiples of each other, the matrix rank of T_{\delta_{x_i}} x is 1. So you can compose the matrix of all eigenvectors from the i-th column of each T_{\delta_{x_i}} x. Denote the matrix of eigenvectors with U x. Example
 U(x_0,x_1,x_2,x_3) = \begin{pmatrix}
1 & \frac{1}{(x_1-x_0)} & \frac{1}{(x_2-x_0)\cdot(x_2-x_1)} & \frac{1}{(x_3-x_0)\cdot(x_3-x_1)\cdot(x_3-x_2)} \\
0 & 1 & \frac{1}{(x_2-x_1)} & \frac{1}{(x_3-x_1)\cdot(x_3-x_2)} \\
0 & 0 & 1 & \frac{1}{(x_3-x_2)} \\
0 & 0 & 0 & 1
\end{pmatrix}
The diagonalization of  T_f x can be written as
 U x \cdot \operatorname{diag}(f(x_0),\dots,f(x_n)) = T_f x \cdot U x .

Alternative definitions

Expanded form


\begin{align}
f[x_0] &= f(x_0) \\
f[x_0,x_1] &= \frac{f(x_0)}{(x_0-x_1)} + \frac{f(x_1)}{(x_1-x_0)} \\
f[x_0,x_1,x_2] &= \frac{f(x_0)}{(x_0-x_1)\cdot(x_0-x_2)} + \frac{f(x_1)}{(x_1-x_0)\cdot(x_1-x_2)} + \frac{f(x_2)}{(x_2-x_0)\cdot(x_2-x_1)} \\
f[x_0,x_1,x_2,x_3] &= \frac{f(x_0)}{(x_0-x_1)\cdot(x_0-x_2)\cdot(x_0-x_3)} + \frac{f(x_1)}{(x_1-x_0)\cdot(x_1-x_2)\cdot(x_1-x_3)} + \frac{f(x_2)}{(x_2-x_0)\cdot(x_2-x_1)\cdot(x_2-x_3)} +\\&\quad\quad \frac{f(x_3)}{(x_3-x_0)\cdot(x_3-x_1)\cdot(x_3-x_2)} \\
f[x_0,\dots,x_n] &=
\sum_{j=0}^{n} \frac{f(x_j)}{\prod_{k\in\{0,\dots,n\}\setminus\{j\}} (x_j-x_k)}
\end{align}

With the help of a polynomial function q with q(\xi) = (\xi-x_0) \cdots (\xi-x_n) this can be written as


f[x_0,\dots,x_n] = \sum_{j=0}^{n} \frac{f(x_j)}{q'(x_j)}.

Alternatively, we can allow counting backwards from the start of the sequence by defining x_k = x_{k + n + 1} = x_{k - (n + 1)} whenever k < 0 or  n < k . This definition allows x_{-1} to be interpreted as x_{n}, x_{-2} to be interpreted as x_{n-1}, x_{-n} to be interpreted as x_{0}, etc. The expanded form of the divided difference thus becomes


f[x_0,\dots,x_n] = \sum_{j=0}^{n} \frac{f(x_j)}{\prod\limits_{k=j-n}^{j-1} (x_j - x_k)} + \sum_{j=0}^{n} \frac{f(x_j)}{\prod\limits_{k=j+1}^{j+n} (x_j - x_k)}

Yet another characterization utilizes limits:


f[x_0,\dots,x_n] = \sum_{j=0}^{n} \lim_{x \rightarrow x_j} \left[ \frac{f(x_j)(x - x_j)}{\prod\limits_{k=0}^{n} (x - x_k)} \right]

Partial fractions

You can represent partial fractions using the expanded form of divided differences. (This does not simplify computation, but is interesting in itself.) If p and q are polynomial functions, where \mathrm{deg}\ p < \mathrm{deg}\ q and q is given in terms of linear factors by q(\xi) = (\xi-x_1)\cdot \dots \cdot(\xi-x_n), then it follows from partial fraction decomposition that

\frac{p(\xi)}{q(\xi)} = \left(t\to\frac{p(t)}{\xi-t}\right)[x_1,\dots,x_n].

If limits of the divided differences are accepted, then this connection does also hold, if some of the x_j coincide.

If f is a polynomial function with arbitrary degree and it is decomposed by f(x) = p(x) + q(x)\cdot d(x) using polynomial division of f by q, then

\frac{p(\xi)}{q(\xi)} = \left(t\to\frac{f(t)}{\xi-t}\right)[x_1,\dots,x_n].

Peano form

The divided differences can be expressed as

f[x_0,\ldots,x_n] = \frac{1}{n!} \int_{x_0}^{x_n} f^{(n)}(t)B_{n-1}(t) \, dt

where B_{n-1} is a B-spline of degree n-1 for the data points x_0,\dots,x_n and f^{(n)} is the n-th derivative of the function f.

This is called the Peano form of the divided differences and B_{n-1} is called the Peano kernel for the divided differences, both named after Giuseppe Peano.

Taylor form

First order

If nodes are cumulated, then the numerical computation of the divided differences is inaccurate, because you divide almost two zeros, each of which with a high relative error due to differences of similar values. However we know, that difference quotients approximate the derivative and vice versa:

\frac{f(y)-f(x)}{y-x} \approx f'(x) for x \approx y

This approximation can be turned into an identity whenever Taylor's theorem applies.

f(y) = f(x) + f'(x)\cdot(y-x) + f''(x)\cdot\frac{(y-x)^2}{2!} + f'''(x)\cdot\frac{(y-x)^3}{3!} + \dots
\Rightarrow \frac{f(y) - f(x)}{y-x} = f'(x) + f''(x)\cdot\frac{y-x}{2!} + f'''(x)\cdot\frac{(y-x)^2}{3!} + \dots

You can eliminate the odd powers of y-x by expanding the Taylor series at the center between x and y:

x = m-h, y=m+h, that is m = \frac{x+y}{2}, h=\frac{y-x}{2}
f(m+h) = f(m) + f'(m)\cdot h + f''(m)\cdot\frac{h^2}{2!} + f'''(m)\cdot\frac{h^3}{3!} + \dots
f(m-h) = f(m) - f'(m)\cdot h + f''(m)\cdot\frac{h^2}{2!} - f'''(m)\cdot\frac{h^3}{3!} + \dots
\frac{f(y) - f(x)}{y-x} = \frac{f(m+h) - f(m-h)}{2\cdot h} =
 f'(m) + f'''(m)\cdot\frac{h^2}{3!} + \dots

Higher order

The Taylor series or any other representation with function series can in principle be used to approximate divided differences. Taylor series are infinite sums of power functions. The mapping from a function f to a divided difference f[x_0,\dots,x_n] is a linear functional. We can as well apply this functional to the function summands.

Express power notation with an ordinary function: p_n(x) = x^n.

Regular Taylor series is a weighted sum of power functions: f = f(0)\cdot p_0 + f'(0)\cdot p_1 + \frac{f''(0)}{2!}\cdot p_2 + \frac{f'''(0)}{3!}\cdot p_3 + \dots

Taylor series for divided differences: f[x_0,\dots,x_n] = f(0)\cdot p_0[x_0,\dots,x_n] + f'(0)\cdot p_1[x_0,\dots,x_n] + \frac{f''(0)}{2!}\cdot p_2[x_0,\dots,x_n] + \frac{f'''(0)}{3!}\cdot p_3[x_0,\dots,x_n] + \dots

We know that the first n terms vanish, because we have a higher difference order than polynomial order, and in the following term the divided difference is one:


\begin{array}{llcl}
\forall j<n & p_j[x_0,\dots,x_n] &=& 0 \\
 & p_n[x_0,\dots,x_n] &=& 1 \\
 & p_{n+1}[x_0,\dots,x_n] &=& x_0 + \dots + x_n \\
 & p_{n+m}[x_0,\dots,x_n] &=& \sum_{a\in\{0,\dots,n\}^m \text{ with } a_1 \le a_2 \le \dots \le a_m} \prod_{j\in a} x_j. \\
\end{array}

It follows that the Taylor series for the divided difference essentially starts with \frac{f^{(n)}(0)}{n!} which is also a simple approximation of the divided difference, according to the mean value theorem for divided differences.

If we would have to compute the divided differences for the power functions in the usual way, we would encounter the same numerical problems that we had when computing the divided difference of f. The nice thing is, that there is a simpler way. It holds


t^n = (1 - x_0\cdot t) \dots \cdot (1 - x_n\cdot t) \cdot
(p_0[x_0,\dots,x_n] + p_1[x_0,\dots,x_n]\cdot t + p_2[x_0,\dots,x_n]\cdot t^2 + \dots) .

Consequently we can compute the divided differences of p_n by a division of formal power series. See how this reduces to the successive computation of powers when we compute p_n[h] for several n.

If you need to compute a whole divided difference scheme with respect to a Taylor series, see the section about divided differences of power series.

Polynomials and power series

Divided differences of polynomials are particularly interesting, because they can benefit from the Leibniz rule. The matrix J with


J=
\begin{pmatrix}
x_0 & 1 & 0 & 0 & \cdots & 0 \\
0 & x_1 & 1 & 0 & \cdots & 0 \\
0 & 0 & x_2 & 1 &        & 0 \\
\vdots & \vdots & & \ddots & \ddots & \\
0 & 0 & 0 & 0 &          & x_n
\end{pmatrix}

contains the divided difference scheme for the identity function with respect to the nodes x_0,\dots,x_n, thus J^n contains the divided differences for the power function with exponent n. Consequently you can obtain the divided differences for a polynomial function \varphi(p) with respect to the polynomial p by applying p (more precisely: its corresponding matrix polynomial function \varphi_{\mathrm{M}}(p)) to the matrix J.

\varphi(p)(\xi) = a_0 + a_1\cdot \xi + \dots + a_n\cdot \xi^n
\varphi_{\mathrm{M}}(p)(J) = a_0 + a_1\cdot J + \dots + a_n\cdot J^n
= \begin{pmatrix}
\varphi(p)[x_0] & \varphi(p)[x_0,x_1] & \varphi(p)[x_0,x_1,x_2] & \ldots & \varphi(p)[x_0,\dots,x_n] \\
0 & \varphi(p)[x_1] & \varphi(p)[x_1,x_2] & \ldots & \varphi(p)[x_1,\dots,x_n] \\
\vdots & \ddots & \ddots & \ddots & \vdots \\
0 & \ldots & 0 & 0 & \varphi(p)[x_n]
\end{pmatrix}

This is known as Opitz' formula.[2] [3]

Now consider increasing the degree of p to infinity, i.e. turn the Taylor polynomial to a Taylor series. Let f be a function which corresponds to a power series. You can compute a divided difference scheme by computing the according matrix series applied to J. If the nodes x_0,\dots,x_n are all equal, then J is a Jordan block and computation boils down to generalizing a scalar function to a matrix function using Jordan decomposition.

Forward differences

For more details on this topic, see Finite difference.

When the data points are equidistantly distributed we get the special case called forward differences. They are easier to calculate than the more general divided differences.

Note that the "divided portion" from forward divided difference must still be computed, to recover the forward divided difference from the forward difference.

Definition

Given n data points

(x_0, y_0),\ldots,(x_{n-1}, y_{n-1})

with

x_{\nu} = x_0 + \nu h,\ h > 0,\ \nu=0,\ldots,n-1

the divided differences can be calculated via forward differences defined as

\triangle^{(0)} y_i := y_i
\triangle^{(k)}y_i := \triangle^{(k-1)}y_{i+1} - \triangle^{(k-1)}y_i,\ k \ge 1.

The relationship between divided differences and forward differences is[4]

f[x_0, x_1, \ldots , x_k] = \frac{1}{k!h^k}\triangle^{(k)}f(x_0).

Example


\begin{matrix}
y_0 &               &                   &                  \\
    & \triangle y_0 &                   &                  \\
y_1 &               & \triangle^2 y_0 &                  \\
    & \triangle y_1 &                   & \triangle^3 y_0\\
y_2 &               & \triangle^2 y_1 &                  \\
    & \triangle y_2 &                   &                  \\
y_3 &               &                   &                  \\
\end{matrix}

See also

References

  1. Isaacson, Walter (2014). The Innovators. Simon & Schuster. p. 20. ISBN 978-1-4767-0869-0.
  2. de Boor, Carl, Divided Differences, Surv. Approx. Theory 1 (2005), 46–69,
  3. Opitz, G. Steigungsmatrizen, Z. Angew. Math. Mech. (1964), 44, T52–T54
  4. Burden, Richard L.; Faires, J. Douglas (2011). Numerical Analysis (9th ed.). p. 129.

External links

This article is issued from Wikipedia - version of the Thursday, April 28, 2016. The text is available under the Creative Commons Attribution/Share Alike but additional terms may apply for the media files.