Error Continued..... : Department of Mathematics IIT Guwahati

Error continued.....
Department of Mathematics
IIT Guwahati
SUDARSHAN K. K. MA311 (2018)

Rude behavior of roundoff error!!
Cancellation error: how to avoid it in practice?
Using this formula, computing the expression in 5 digit decimal arithmetic

yields 1.5811 × 10−3 .(Excersise) ( Exact value is 1.58113487 × 10−3 )

Underflow and Overflow: Consider a floating point system with 4

decimal digits and 2 exponent digits ( i.e. −99 ≤ e ≤ 99,). Compute
p
c = (a2 + b2 ) for a = 1060 and b = 1.
f l(a) = 1.000 × 1060 , f l(b) = 1.000 × 100 .

f l(a) × f l(a) = 1.000 × 10120 which cannot be represented in this
floating system as exponent is three digit number.

Underflow and Overflow: Consider a floating point system with 4

decimal digits and 2 exponent digits ( i.e. −99 ≤ e ≤ 99,). Compute
p
c = (a2 + b2 ) for a = 1060 and b = 1.
f l(a) = 1.000 × 1060 , f l(b) = 1.000 × 100 .

f l(a) × f l(a) = 1.000 × 10120 which cannot be represented in this
floating system as exponent is three digit number.
How to avoid it? Rewrite the expression as
p
c = s (a/s)2 + (b/s)2 for any s 6= 0.
Thus if we use s = a = 1060 there will be an underflow in (b/s)2 ,

f l((b/s)2 ) = 1.000 × 10−120 , this will be set as zero. Finally we get the
correct answer up to the precision of floating point system.

Accumulation of roundoff error: x, y and z are machine numbers in 32
bit system.

Accumulation of roundoff error: x, y and z are machine numbers in 32
bit system.
Theorem
(On the relative roundoff error in adding) Let x0 , x1 , · · · , xn be positive
machine numbers in a computer whose unit roundoff error is η. Then the
relative roundoff error in computing
n
X
xi
i=0
in the usual way is at most (1 + η)n − 1 ≈ nη.

10 -8
6
4
roundoff error
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t
Figure: Error in sampling exp(−t)(sin(2πt) + 2) in single precision.

Root of quadratic polynomial: Example. Fining roots of
x2 + 62.10x + 1 = 0 in 4 digits decimal floating point system. Exact
roots are found to be

[Compare with ax2 + bx + c = 0] The effect is due to the fact that

√
b2 4ac and b2 − 4ac ≈ |b|. To reduce the error we use the formula
x1 x2 = 1 We compute x2 as it is and
x1 = 1/x2 ; f l(x1 ) = 1.000/(−62.0) = −0.01610
| − 0.01610723 + 0.01610|
≈ 4.4 × 10−4 .
| − 0.01610723|




Nested arithmetic:
pm (x) = c0 +c1 x+· · ·+cn xn → pn (x) = (. . . ((cn xn +cn−1 )x+cn−2 )x.....)x+c0 .

Nested arithmetic:
pm (x) = c0 +c1 x+· · ·+cn xn → pn (x) = (. . . ((cn xn +cn−1 )x+cn−2 )x.....)x+c0 .


You save number of operations! → the number of operations comes

down to O(n) from O(n2 ) when we calculate the value of a nth order
polynomial.;

Catastrophe due to roundoff error!! :(
• Patriot missile failure in Dhahran, Saudi Arabia, on February 25,

1991, which resulted in 28 deaths.pause
• Intel Pentium flaw (1994):
http://en.wikipedia.org/wiki/Pentium FDIV bug
A = 4195835.0, B = 3145727.0; A − (A/B) ∗ B =? (five digit
arithmetic)

Catastrophe due to roundoff error!! :(
• Patriot missile failure in Dhahran, Saudi Arabia, on February 25,

1991, which resulted in 28 deaths.pause
• Intel Pentium flaw (1994):
http://en.wikipedia.org/wiki/Pentium FDIV bug
A = 4195835.0, B = 3145727.0; A − (A/B) ∗ B =? (five digit
arithmetic)
• The Explosion of the Ariane 5, June 4- 1996.

Numerical instability
To compute ( 13 )15 .

1 13 4
x0 = 1, x1 = ,··· xn = xn−1 − xn−2 .
3 3 3
Thus x15 = ( 31 )15 can be computed successively using the sequence
above.

1 13 4
x0 = 1, x1 = ,··· xn = xn−1 − xn−2 .
3 3 3
above.
If there is no roundoff error in the computer we will get the exact value of
x15 .

1 13 4
x0 = 1, x1 = ,··· xn = xn−1 − xn−2 .
3 3 3
above.
If there is no roundoff error in the computer we will get the exact value of
x15 .
What happens if there is roundoff error?


Speaking informally, we say that a numerical process is unstable if small

errors made at one stage of the process are magnified in subsequent
stages and seriously degrade the accuracy of the overall calculation.

Error Continued..... : Department of Mathematics IIT Guwahati

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Error Continued..... : Department of Mathematics IIT Guwahati

Uploaded by

Copyright:

Available Formats

Error continued.....

SUDARSHAN K. K. MA311 (2018)

Cancellation error: how to avoid it in practice?

Using this formula, computing the expression in 5 digit decimal arithmetic

SUDARSHAN K. K. MA311 (2018)

Underflow and Overflow: Consider a floating point system with 4

f l(a) = 1.000 × 1060 , f l(b) = 1.000 × 100 .

SUDARSHAN K. K. MA311 (2018)

Underflow and Overflow: Consider a floating point system with 4

f l(a) = 1.000 × 1060 , f l(b) = 1.000 × 100 .

Thus if we use s = a = 1060 there will be an underflow in (b/s)2 ,

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

in the usual way is at most (1 + η)n − 1 ≈ nη.

Figure: Error in sampling exp(−t)(sin(2πt) + 2) in single precision.

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

[Compare with ax2 + bx + c = 0] The effect is due to the fact that

x1 = 1/x2 ; f l(x1 ) = 1.000/(−62.0) = −0.01610

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

You save number of operations! → the number of operations comes

SUDARSHAN K. K. MA311 (2018)

• Patriot missile failure in Dhahran, Saudi Arabia, on February 25,

SUDARSHAN K. K. MA311 (2018)

• Patriot missile failure in Dhahran, Saudi Arabia, on February 25,

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

SUDARSHAN K. K. MA311 (2018)

Speaking informally, we say that a numerical process is unstable if small

You might also like