21.06.2014 Views

Numerical Methods Contents - SAM

Numerical Methods Contents - SAM

Numerical Methods Contents - SAM

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

numerically equivalent ˆ= same result when executed with the same machine arithmetic<br />

This introduces a new and important aspect in the study of numerical algorithms!<br />

The above statement means that whenever we study the impact of roundoff errors on LUfactorization<br />

it is safe to consider only the basic version without pivoting, because we can<br />

always assume that row swaps have been conducted beforehand.<br />

“Computers use floating point numbers (scientific notation)”<br />

△<br />

Example 2.4.1 (Decimal floating point numbers).<br />

3-digit normalized decimal floating point numbers:<br />

2.4 Supplement: Machine Arithmetic<br />

valid: 0.723 · 10 2 , 0.100 · 10 −20 , −0.801 · 10 5<br />

invalid: 0.033 · 10 2 , 1.333 · 10 −4 , −0.002 · 10 3<br />

A very detailed exposition and in-depth discussion of the all the material in this section can be found<br />

in [24]:<br />

• [24, Ch. 1]: excellent collection of examples concerning the impact of roundoff errors.<br />

• [24, Ch. 2]: floating point arithmetic, see Def. 2.4.1 below and the remarks following it.<br />

Ôº½¼ ¾º<br />

General form of m-digit normalized decimal floating point number:<br />

never = 0 !<br />

x = ± 0 . 1 1 1 1 1 ... 1 1 · 10<br />

E<br />

} {{ }<br />

m digits of mantissa exponent ∈ Z<br />

Ôº½¼ ¾º<br />

✸<br />

✗<br />

✖<br />

✔✗<br />

Computer = finite automaton ➢ can handle only finitely many numbers, not R<br />

✕✖<br />

✕<br />

machine numbers, set M<br />

Essential property:<br />

M is a discrete subset of R<br />

M not closed under elementary arithmetic operations +, −, ·,/.<br />

roundoff errors (ger.: Rundungsfehler) are inevitable<br />

The impact of roundoff means that mathematical identities may not carry over to the computational<br />

realm. Putting it bluntly,<br />

✗<br />

✖<br />

Computers cannot compute “properly” !<br />

✔<br />

✕<br />

✔<br />

Of course,<br />

computers are restricted to a finite range of exponents.<br />

Definition 2.4.1 (Machine numbers).<br />

Given ☞ basis B ∈ N \ {1},<br />

☞ exponent range {e min , ...,e max }, e min ,e max ∈ Z, e min < e max ,<br />

☞ number m ∈ N of digits (for mantissa),<br />

the corresponding set of machine numbers is<br />

M := {d · B E : d = i · B −m , i = B m−1 ,...,B m − 1, E ∈ {e min ,...,e max }}<br />

never = 0 !<br />

1 1 . .. 1 1<br />

} {{ }<br />

digits for exponent<br />

machine number ∈ M : x = ± 0 . 1 1 1 1 1 . .. 1 1 · B<br />

} {{ }<br />

m digits for mantissa<br />

✬<br />

✫<br />

numerical computations<br />

≠<br />

analysis<br />

linear algebra<br />

✩<br />

✪<br />

Ôº½¼ ¾º<br />

Largest machine number (in modulus) : x max = (1 − B −m ) · B e max<br />

Smallest machine number (in modulus) : x min = B −1 · B e min<br />

Distribution of machine numbers:<br />

Ôº½¼ ¾º

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!