Questions: Rounding Errors and Error Propagation

5 questions to test your understanding

Score: 0 / 5
Question 1 Multiple Choice

A program computes (1.0000001 − 1.0000000) in 64-bit floating-point. Both inputs are accurate to about 16 significant decimal digits, yet the result has far fewer significant digits. What phenomenon explains this?

AOverflow: the result exceeds the maximum representable floating-point value
BCatastrophic cancellation: subtracting nearly equal values destroys the leading significant digits, leaving only rounding-error residue
CUnderflow: the result is too small to be stored in normalized floating-point form
DTruncation error: the values were stored as integers and rounded before subtraction
Question 2 Multiple Choice

An algorithm evaluates a function f whose condition number κ ≈ 10⁸. The inputs carry relative error ≈ 10⁻¹⁶ (one unit of machine epsilon). What relative error should you expect in the output?

AAbout 10⁻¹⁶, because machine epsilon bounds all floating-point errors regardless of the function
BAbout 10⁻⁸, because the condition number amplifies input relative error into output relative error
CAbout 10⁸, because the condition number is larger than machine epsilon
DExactly zero, because the algorithm uses exact arithmetic internally
Question 3 True / False

A numerically stable algorithm can still produce a result with large forward error if the problem itself is ill-conditioned.

TTrue
FFalse
Question 4 True / False

Floating-point rounding errors from chained operations tend to cancel out on average, so longer computations are generally as accurate as shorter ones.

TTrue
FFalse
Question 5 Short Answer

What does it mean for an algorithm to be 'numerically stable' in terms of backward error analysis? Why is backward error analysis more useful than forward error analysis for judging algorithm quality?

Think about your answer, then reveal below.