A mathematical problem is well-conditioned when small relative changes in its input generally produce small relative changes in its solution. It is ill-conditioned when small input changes may produce much larger changes in the output. Conditioning is an inherent property of the problem rather than simply a property of the algorithm used to solve it. For example, evaluating a function near a point where its derivative is very large may be sensitive to small input perturbations. The condition number is a useful measure of sensitivity in many mathematical settings. Understanding conditioning helps distinguish errors caused by the problem itself from errors introduced by a numerical algorithm.
Numerical Conditioning
Numerical conditioning describes how sensitive the solution of a mathematical problem is to small changes in its input data.
κ ≈ |(Δy/y) / (Δx/x)|
Explanation
Variables
| Symbol | Meaning |
|---|---|
κ | A relative condition number or sensitivity measure. |
x | Original input. |
Δx | Small perturbation in the input. |
y | Original output or solution. |
Δy | Change in the output caused by the perturbation. |
Example or application
Subtracting two nearly equal numbers can be sensitive to small changes in their values because the result may be much smaller than either input.
Definition
Numerical conditioning is the sensitivity of a mathematical problem's solution to small perturbations in its input data.
Purpose
To explain how sensitive mathematical problems are to changes in input data and how that sensitivity influences numerical reliability.
Importance
A poorly conditioned problem can produce unreliable results even when a numerical algorithm is implemented correctly. Conditioning analysis helps determine how many digits of an input or computed solution can be trusted.
Learning Objectives
Define numerical conditioning.
Distinguish conditioning from algorithmic stability.
Explain well-conditioned and ill-conditioned problems.
Interpret a relative condition number.
Recognize numerical sensitivity in subtraction and equation solving.
Key Concepts
Conditioning
Sensitivity
Condition number
Input perturbation
Output variation
Well-conditioned problem
Ill-conditioned problem
Working Principle
Conditioning is studied by examining how small perturbations in input data affect the mathematical solution. Relative input and output changes are compared to estimate the problem's sensitivity.
Process
Define the mathematical problem and its input.
Identify the solution or output quantity.
Introduce a small perturbation in the input.
Determine the resulting change in the output.
Compare relative changes.
Assess whether the problem is well-conditioned or ill-conditioned.
Applications
Numerical linear algebra
Scientific measurement
Parameter estimation
Inverse problems
Engineering computation
Numerical optimization
Worked Example
Problem
Consider y = 1000x and examine a small relative input change.
Steps
Let x = 2, so y = 2000.
Increase x by 0.1%, giving x = 2.002.
The new output is y = 2002.
The relative input change is 0.1%.
The relative output change is also 0.1%.
Answer
The relative sensitivity is 1 in this example; the output changes proportionally to the input.
Common Misconceptions
Conditioning is a property of the mathematical problem, whereas stability is a property of a numerical method and its behaviour.
An accurate algorithm cannot remove all sensitivity inherent in an ill-conditioned problem.
A large condition number indicates sensitivity, not necessarily an implementation error.
Condition numbers depend on how the problem and its variables are scaled.
Practice Questions
| Question | Answer |
|---|---|
What is a well-conditioned problem? | A problem in which small input changes generally produce small changes in the solution. |
What does a large condition number usually indicate? | High sensitivity of the solution to input perturbations. |
Is conditioning the same as stability? | No. Conditioning describes problem sensitivity, while stability describes the behaviour of a numerical algorithm. |
Key Takeaway
Conditioning measures the sensitivity of a mathematical problem to input changes and helps explain limits on the accuracy of computed solutions.
Prerequisites
SCI-MAT-014-011
SCI-MAT-014-012
SCI-MAT-014-013
Functions and relative error
Reference
Higham, N. J. Accuracy and Stability of Numerical Algorithms.
Trefethen, L. N., and Bau, D. Numerical Linear Algebra.
Related Topics
SCI-MAT-014-022
SCI-MAT-014-013
SCI-MAT-014-024
SCI-MAT-014-101
Sources and verification
Sources are shown as provided in the project data. Not every source has yet been independently verified. Please report errors or missing citations.
- Source verification pending; check an authoritative textbook, university, standards body, or peer-reviewed source.