Error Propagation
Error Propagation
\
|
~ A
e.g., if z =x
n
, then :
x
x
nz
x
x
nx x nx x
dx
dz
z
n n
A
=
A
= A = A
|
.
|
\
|
~ A
1
Hence, for this particular function,
the percent (or fractional) error in z is :
|
.
|
\
|
A
=
|
.
|
\
|
A
x
x
n
z
z
or...... just n times the percent error in x
Error Propagation using Calculus
Functions of one variable
4
y
y
z
x
x
z
z A
|
|
.
|
\
|
c
c
+ A
|
.
|
\
|
c
c
~ A
Suppose uncertainties in two measured quantities x and y are : x and y ,
what is the uncertainty in some derived quantity z(x,y) ?
For such functions of 2 variables
we use partial differentiation
But, combining errors ALWAYS INCREASES total
error - so make sure terms add with the same sign :
y
y
z
x
x
z
z A
|
|
.
|
\
|
c
c
+ A
|
.
|
\
|
c
c
~ A
( ) ( )
2
2
2
2
y
y
z
x
x
z
z A |
.
|
\
|
c
c
+ A
|
.
|
\
|
c
c
~ A
Later we will show that it turns out to be
better to add in quadrature
i.e. the root of the sum of the squares :
( ) ( )
2
2
2
2
y
y
z
x
x
z
A
|
|
.
|
\
|
c
c
+ A
|
.
|
\
|
c
c
=
or ....
We can usually always handle error propagation in this way by calculus,
(see Lab Handbook, Section 6, for details)
but next we give some short cuts for common cases...
Error Propagation using Calculus
Functions of more than one variable
5
Instead of differentiating cz/cx, cz/cy etc, a simpler approach is also acceptable :
1. In the derived quantity z, replace x by x + x, say
2. Evaluate z in the approximation that x is small
( ) ( ) x z x x z z A + ~ A
x z
x z a x x z z
A = A
A + = + A + = A + ) (
( )
x b z
x b z x b bx x x b z z
A = A
A + = A + = A + = A +
Simplified Error Propagation
A short-cut avoiding calculus!
Ex. 1 : z =x +a , where a = constant
Ex. 2 : z =bx , where b = constant
( ) ( )
x
x
z z
x
x
bx x x x x b x x b z z
A
+ =
|
.
|
\
|
A
+ ~ A + A + = A + = A +
2 2
1 2
2 2 2
2
Ex. 3 : z =bx
2
, where b = constant
x
x
z
z
x
x
z z
A
~
A
A
~ A 2
2
(same as obtained earlier
for z = x
n
, with n=2)
6
( ) ( )
y x z
y x y x y y x x z z
A + A ~ A
A + A + + = A + + A + ~ A + ) ( ) (
Ex. 4 : z =x +y
+
Ax
Ay
=
Az
x y y x +
P
r
o
b
.
But in many cases the actual probability distribution is not rectangular but Normal ( or Gaussian)
Then, the above method would overestimate Az
P
r
o
b
.
x
Ax
+
y
Ay
=
y x +
Az
What is Az for such
prob. distributions?
( ) ( )
2 2
y x z A + A = A
y x z A + A = A
Answer : the errors add
in quadrature :
Ay
Ax
Az
Ax
Ay
i.e., this would mean :
Simplified Error Propagation
Functions of more than one variable
7
Mnemonic :
Ex. 5 : z =x - y
( ) ( )
y
y
x
x
z
z
y
y
x
x
xy
y x x y y x xy y y x x z z
A
+
A
~
A
A
+
A
+ ~
A A + A + A + = A + A + ~ A +
1
2
2
|
|
.
|
\
| A
+
|
.
|
\
|
A
=
A
y
y
x
x
z
z
For a more accurate estimate, use the quadrature formula :
Aside : For complicated expressions for z, as a last resort,
you can always try the Min-Max method to get Az : given any errors in x, y, ....etc
1. Calculate max. possible value z can take: e.g. z
max
= z(x+Ax, y-Ay,....)
2. Similarly, calc. min. poss. value z can take, z
min
= z(x-Ax, y+Ay,....)
3. Then error can be estimated as : Az ~ (z
max
z
min
)
Remember : Even a crude estimate of an error is always better than no estimate at all !
i.e. percent error in z ~
sum of percent errors in x & y
Exercise : Show that for z = x/y then (try using the min-max method for this example)
y
y
x
x
z
z
A
+
A
~
A
Simplified Error Propagation
Functions of more than one variable
8