Inverse
Problems:
What are they, how can
we approach them, and
what can we learn from
them?
Consider the Following:
Johnny works two jobs. This past April he made
$2100 at the first, and an additional $1300 at the
second. How much did Johnny make in total for
the month of April?
Consider the Following:
Johnny works two jobs. This past April he made
$2100 at the first, and an additional $1300 at the
second. How much did Johnny make in total for
the month of April?
Not surprisingly, Johnny has made
$3400 for April.
But How About This:
Susie also works two jobs. If Susie made $4500
for herself during the month of April, how much did
she earn at each job individually?
But How About This:
Susie also works two jobs. If Susie made $4500
for herself during the month of April, how much did
she earn at each job individually?
There isn't a single answer for this question!
Here's a general diagram
of the situations we just had.
Job #1 Salary
Total Income
Job #2 Salary
In Johnny's case, we combined
A & B to get C.
Job #1 Salary
Total Income
Job #2 Salary
But in Susie's case, we only had C
to try finding A & B.
Job #1 Salary
? Total Income
Job #2 Salary
?
This is an example of an inverse problem;
how can we attempt to solve it?
One way is by utilizing contours,
like on this map of Hawaii.
Here's a zoomed-in look at the Big Island.
Note that the contours represent lines
of constant elevation.
We'll use that same idea, but applied
in a different way.
For example, here's the combination A+B,
where we require that C = 75.
All points on the blue contour represent
A+B = 75.
We could also show many contours at once.
Here, the line in red represents the
requirement that A+B = 110.
We could combine A & B any way we want actually;
here are contours for A•B.
So, in the regular
problem, we find a
unique answer C from
our combined
parameters A & B.
However, in the inverse
problem, we find a
contour of solutions for
A & B from a specified
value for C.
Mathematically
speaking, there are
infinitely-many
solutions!
Things start to get even more complex
when we take uncertainties into account.
This is a Normal Curve, also called
a Bell Curve. Data can often be
represented by this type of graph.
Center
Spread
Notice how the graph has both a center (or mean)
& a spread (or standard deviation.)
The numbers at the bottom are a measure of how far
from the center a value is: greater means
farther away, and thus more unlikely to occur.
We can use this idea in our requirement for C...
...with the Chi-square Statistic (pronounced “K-eye.”)
We can look at the value of Chi-square for C
at many locations on our contour plot.
The higher the value, the less likely that combination
of A & B is (given our requirements on C.)
Let's Try it Out:
Suppose we are interested in the combination:
A+B=C
Say we also know that C = 110, with an
uncertainty of 10.
We might guess that we'd get a simple contour of
possible A & B combos, like before.
Is that the case?...
...No! Now we find possible regions for A & B.
The darker regions are more likely combos for A & B,
while lighter regions are more unlikely.
See how the uncertainty in C has spread out the
combinations of A & B from a single contour?
As we can see, there are a LOT of reasonable
possibilities for A & B in this situation.
Consider, Though:
Say we were also interested in the combination:
A–B=D
Suppose we know that D = 30, with an
uncertainty of 5.
We realize now that we'll get a swath, just like with
C from before.
But what will it look like?...
...Hey! This region runs in a different direction.
By itself, this isn't really any more interesting. If we
put both regions together, though...
...Presto! We've shrunk the possible A & B
combinations dramatically.
In such situations, this Chi-square analysis has the
potential to be quite powerful.
In Summary:
Relationship: Parameters {A,B} & Constraint C
{A,B} → C: Unique (Normal Problem)
C → {A,B}: Contour (Inverse Problem)
Uncertainties on C: Contour morphs into Region
Compile C with D: Region can shrink smaller
Apply this idea to tackle real problems!