Numerical Analysis/Truncation Errors

Definition

Truncation errors are defined as the errors that result from using an approximation in place of an exact mathematical procedure.

There are two ways to measure the errors:

1. Local Truncation Error (LTE): the error, $\tau$ , introduced by the approximation method at each step.
2. Global Truncation Error (GTE): the error, $e$ , is the absolute difference between the correct value and the approximate value.

Assume that our methods take the form:

 Let yn+1 and yn be approximation values. Then $y_{n+1}=y_{n}+h\cdot A(t_{n},y_{n},h,f)$ , where $h$ is the time step, equal to $t_{n+1}-t_{n}$ , and $A$ is an increment function and is some algorithm for approximating the average slope ${\frac {y_{n+1}-y_{n}}{h}}$ . Three important examples of $A$ are: Euler’s method: $A(t_{n},y_{n},h,f)=f(t_{n},y_{n})$ . Modified Euler's method: $A(t_{n},y_{n},h,f)={\frac {1}{2}}(A_{1}+A_{2})$ , where {\begin{aligned}A_{1}&=f(y_{n},y_{n}){\text{, and}}\\A_{2}&=f(t_{n}+h,y_{n}+h\cdot A_{1}){\text{.}}\end{aligned}} Runge-Kutta method: $A(t_{n},y_{n},h,f)={\frac {1}{6}}(A_{1}+2A_{2}+2A_{3}+A_{4})$ , where {\begin{aligned}A_{1}&=f(t_{n},y_{n}){\text{,}}\\A_{2}&=f(t_{n}+{\frac {1}{2}}h,y_{n}+{\frac {1}{2}}h\cdot A_{1}){\text{,}}\\A_{3}&=f(t_{n}+{\frac {1}{2}}h,y_{n}+{\frac {1}{2}}h\cdot A_{2}){\text{, and}}\\A_{4}&=f(t_{n}+h,y_{n}+h\cdot A_{3}){\text{.}}\end{aligned}} Why do we care about truncation errors?

In the case of one-step methods, the local truncation error provides us a measure to determine how the solution to the differential equation fails to solve the difference equation. The local truncation error for multistep methods is similar to that of one-step methods.

A one-step method with local truncation error $\tau _{n}(h)$ at the nth step:

• This method is consistent with the differential equation it approximates if
$\lim _{h\to 0}\max _{1\leq n\leq N}|\tau _{n}(h)|=0.$ Note that here we assume that the approximation values are exactly equal to the true solution at every step.

• The method is convergent with respect to the differential equation it approximates if
$\lim _{h\to 0}\max _{1\leq n\leq N}|y_{n}-y(t_{n})|=0,$ where $y_{n}$ denotes the approximation obtained from the method at the nth step, and $y(t_{n})$ the exact value of the solution of the differential equation.

How do we avoid truncation errors?

The truncation error generally increases as the step size increases, while the roundoff error decreases as the step size increases.

Relationship Between Local Truncation Error and Global Truncation Error

The global truncation error (GTE) is one order lower than the local truncation error (LTE).
That is,

if $\tau _{n}(h)=O(h^{p+1})$ , then $e_{n}(h)=O(h^{p})$ .

Proof

We assume that perfect knowledge of the true solution at the initial time step.
Let ${\tilde {y}}(t)$ be the exact solution of

{\Big \{}{\begin{aligned}y'&=f(t,y){\text{, and}}\\y(t_{n})&=y_{n}\,.\end{aligned}} The truncation error at step n+1 is defined as $\tau _{n+1}(h)={\tilde {y}}(t_{n+1})-y_{n+1}.$ Also, the global errors are defined as

{\begin{aligned}e_{0}(h)&=0\\e_{n+1}(h)&=y(t_{n+1})-y_{n+1}\\&=[y(t_{n+1})-{\tilde {y}}(t_{n+1})]+[{\tilde {y}}(t_{n+1})-y_{n+1}]\,.\end{aligned}} According to the w:Triangle inequality, we obtain thatString Module Error: function rep expects a number as second parameter, received "

"
$|e_{n+1}(h)|\leq |y(t_{n+1})-{\tilde {y}}(t_{n+1})|+|{\tilde {y}}(t_{n+1})-y_{n+1}|.$ (1)

String Module Error: function rep expects a number as second parameter, received "

"

The second term on the right-hand side of (1 ) is the truncation error $\tau _{n+1}(h)$ .

Here we assumeString Module Error: function rep expects a number as second parameter, received "

"
$\tau _{n+1}(h)={\tilde {y}}(t_{n+1})-y_{n+1}=O(h^{p+1}).$ (2)

String Module Error: function rep expects a number as second parameter, received "

" Thus,String Module Error: function rep expects a number as second parameter, received "

"
$|{\tilde {y}}(t_{n+1})-y_{n+1}|\leq Ch^{p+1}\,.$ (3)

String Module Error: function rep expects a number as second parameter, received "

"

The first term on the right-hand side of (1 ) is the difference between two exact solutions.

Both $y(t)$ and ${\tilde {y}}(t)$ satisfy $y'=f(t,y)$ so

{\Big \{}{\begin{aligned}y'(t)&=f(t,y){\text{, and}}\\{\tilde {y}}(t)&=f(t,{\tilde {y}})\,.\end{aligned}} By subtracting one equation from the other, we can get that

{\begin{aligned}y'(t)-{\tilde {y}}(t)&=f(t,y)-f(t,{\tilde {y}})\quad {\text{so}}\\|y'(t)-{\tilde {y}}'(t)|&=|f(t,y)-f(t,{\tilde {y}})|.\end{aligned}} Since $f$ is w:Lipschitz continuous, then

$|y'(t)-{\tilde {y}}'(t)|\leq L|y(t)-{\tilde {y}}(t)|,$ where $t>t_{n}.$ {\begin{aligned}|y(t)-{\tilde {y}}(t)|&\leq |y(t_{n})-{\tilde {y}}(t_{n})|\exp \left(\int _{t_{n}}^{t}Lds\right)\\&=e^{L(t-t_{n})}|y(t_{n})-{\tilde {y}}(t_{n})|,\end{aligned}} where $t\in [t_{n},t_{n+1}].$ Setting $t=t_{n+1}$ , we have thatString Module Error: function rep expects a number as second parameter, received "

"
{\begin{aligned}|y(t_{n+1})-{\tilde {y}}(t_{n+1})|&\leq e^{L(t_{n+1}-t_{n})}|y(t_{n})-{\tilde {y}}(t_{n})|\\&=e^{Lh}|y(t_{n})-{\tilde {y}}(t_{n})|\\&=e^{Lh}|y(t_{n})-y_{n})|\\&=e^{Lh}|e_{n}(h)|\,.\end{aligned}} (4)

String Module Error: function rep expects a number as second parameter, received "

"

Plugging equation (3 ) and (4 ) into (1

), we can get thatString Module Error: function rep expects a number as second parameter, received "

"
$|e_{n+1}(h)|\leq e^{Lh}|e_{n}(h)|+Ch^{p+1}\,.$ (5)

String Module Error: function rep expects a number as second parameter, received "

"

Note that equation (5 ) is a recursive inequality valid for all values of $n$ .

Next, we are trying to use it to estimate $|e_{N}(h)|,$ where we assume $Nh=T$ .

Let $\alpha =e^{Lh}.$ Dividing both sides of (4 ) by $\alpha ^{n+1},$ we get that

${\frac {|e_{n+1}(h)|}{\alpha ^{n+1}}}\leq {\frac {e_{n}(h)}{\alpha ^{n}}}+Ch^{p+1}{\frac {1}{\alpha ^{n+1}}}\,.$ Summing over n = 0,1, 2,…, N-1,

${\frac {|e_{1}(h)|}{\alpha ^{1}}}\leq {\frac {e_{0}(h)}{\alpha ^{0}}}+Ch^{p+1}{\frac {1}{\alpha ^{1}}}$ ,
${\frac {|e_{2}(h)|}{\alpha ^{2}}}\leq {\frac {e_{1}(h)}{\alpha ^{1}}}+Ch^{p+1}{\frac {1}{\alpha ^{2}}}$ ,
$\vdots$ and

${\frac {|e_{N}(h)|}{\alpha ^{N}}}\leq {\frac {e_{N-1}(h)}{\alpha ^{N-1}}}+Ch^{p+1}{\frac {1}{\alpha ^{N}}}\,.$ Then we obtain

{\begin{aligned}{\frac {|e_{N}(h)|}{\alpha ^{N}}}&\leq {\frac {e_{0}(h)}{\alpha ^{0}}}+Ch^{p+1}({\frac {1}{\alpha ^{1}}}+{\frac {1}{\alpha ^{2}}}+\cdots +{\frac {1}{\alpha ^{N}}})\\&={\frac {e_{0}(h)}{\alpha ^{0}}}+Ch^{p+1}[{\frac {1}{\alpha ^{N}}}(1+\alpha +\alpha ^{2}+\cdots +\alpha ^{N-1})]\\&={\frac {e_{0}(h)}{\alpha ^{0}}}+Ch^{p+1}[{\frac {1}{\alpha ^{N}}}({\frac {\alpha ^{N}-1}{\alpha -1}})]\,.\end{aligned}} Since $e_{0}(h)=0,$ we have

{\begin{aligned}|e_{N}(h)|&\leq Ch^{p+1}[{\frac {1}{\alpha ^{N}}}({\frac {\alpha ^{N}-1}{\alpha -1}})]\\&\leq Ch^{p+1}({\frac {\alpha ^{N}-1}{\alpha -1}}),{\text{since }}\alpha ^{N}>1\,.\end{aligned}} Using the inequality $e^{x}-1\geq x,$ we get

{\begin{aligned}\alpha -1&=e^{Lh}-1\geq Lh\quad {\text{and}}\\\alpha ^{N}-1&=e^{LNh}-1=e^{LT}-1\,.\end{aligned}} Therefore, we can obtain that

$|e_{N}(h)|\leq Ch^{p+1}({\frac {e^{LT}-1}{Lh}})=C({\frac {e^{LT}-1}{L}})h^{p}.$ That is,String Module Error: function rep expects a number as second parameter, received "

"
$|e_{N}(h)|=O(h^{p}).$ (6)

String Module Error: function rep expects a number as second parameter, received "

"

From equation (2 ) and (6 ),

$\tau _{n+1}(h)=O(h^{p+1})\quad {\text{and}}$ $|e_{N}(h)|=O(h^{p}),$ so we can conclude that the global truncation error is one order lower than the local truncation error.

Graph

File:W LTE and GTE.jpg
Relationship between LTE and GTE

In this graph, $c=a+{\frac {b-a}{2}}.$ The red line is the true value, the green line is the first step, and the blue line is the second step.

${\overline {AB}}$ is the local truncation error at step 1, $\tau _{1}=e_{1}$ , equal to ${\overline {CD}}.$ ${\overline {DE}}$ is separation because after the first step we are on the wrong solution of the ODE.
${\overline {EF}}$ is $\tau _{2}.$ Thus, ${\overline {CF}}$ is the global truncation error at step 2, $e_{2}.$ We can see from this,

$e_{n+1}=e_{n}+h[A(t_{n}),y(t_{n}),h,f)-A(t_{n},y_{n},h,f)]+\tau _{n+1}.$ Then,

$e_{2}=e_{1}+h[A(t_{1}),y(t_{1}),h,f)-A(t_{1},y_{1},h,f)]+\tau _{2}.$ $e_{2}={\overline {AB}}+{\overline {DE}}+{\overline {CF}}.$ Exercise

Find the order of the 2-steps Adams-Bashforth method. You need to show the order of truncation error.