- Modelling Errors
- Comes from simplifying assumptions (or omitting aspects of the problem)
- Data Errors
- Comes from inaccurate and imprecise empirical measurements
- Can also come from using potentially incorrect results from some previous computational process
- Computational Errors
- Can come from truncating a computational process
- Can come from rounding
- Absolute Error —
| Approx. value - True value |
- Relative Error —
Absolute error ÷ true value × 100%
- Let:
f
be a true functionx
be a true value to be used in a computationf(x)
be the desired result of the true function on the true valueg
be an approximation off
y
be an inexact measurement ofx
g(y)
be the inexact result
- Then:
- Computational Error =
g(y) - f(y)
- Propagated Data Error =
f(y) - f(x)
- Total Error = Computational Data Error + Propagated Error =
g(y) - f(x)
- Computational Error =
- Forward Error
- Discrepancy between the computed and true values
inexact output - exact output
- Quality of results:
|inexact output - exact output|
is the relative magnitude of the forward error
- Backward Error
- Discrepancy between the initial output that produced the discrepancy in the result
inexact input - exact input
- Sensitivity — qualitative statement of propagated data error
- Conditioning — quantitative measure of propagated data error
- A problem is insensitive or well-conditioned if a given relative change in the input data causes a reasonably commensurate relative change in the solution
|input a - input b| ~ |output a - output b|
- Otherwise, it is sensitive or ill-conditioned
|input a - input b| >> or << |output a - output b|
- Condition Number
- Ratio of relative change
| relative forward error | ÷ | relative backward error |
- Ill-conditioned problems have a condition number larger than 1
~ xf'(x) ÷ f(x)