Numerical Analysis: 3. Approximations and Round-Off Errors
Solution
- analytical solution
- numerical methods
Error Types
- Round-Off Error
- computers can represent quantities with a finite number of digits
- Truncation Error
- approximate solution
Significant Digits
- is the number of certain digits plus one estimated digit.
- Zeros are not always significant figures. They are just to locate a decimal point.
- 0.00001845 -> four
- Four signicant figures may be acceptable.
- Omission of the remaining significant figures is called round-off error.
Error
- True error;
- True percent relative error;
- Approximation error
- In real-world, we will obviously not know the true answer.
- continue approximation until e < prespecified acceptable level.
- (Scarborough, 1966) We can be assured that the result is correct to at least n significant figures if the approximation error falls below:
Floating-Point
Representation
- m b^e
- m : matissa
- b : base of the number system
- e : exponent
- example: 7 bits
- [0] : sign of number
- [1] : sign of exponent
- [2-3] : magnitude of exp
- [4-6] : magnitude of mantissa
- number range : 0011001 (1e-3) ~ 1111111 (7e+3)
Approximation
- Rounding
- nearby value
- Chopping
- discard remains
Error
- Quantizing errors
- Overflow
- Underflow
Quantizing Errors
- machine epsilon : ε = b^(1-t)
- b : number base
- t : the number of mantissa bits
Subtractive Cancellation
- round-off induced when subtracting two nearly equal floating-point numbers
- subtractive cancellation occurs where b^2 >> 4ac