Perplexed by logical condition result

I have a matrix that is a mixture of numbers and NaNs by construction. I looking through answers as to how to use functions like GEOMEAN and MEAN on these type of matrices while ignoring the NaNs in these calculations. I found a response indicating I could use the following structure:

condition=isnan(A)
B=geomean(A(~condition))

The "condition" assignment statement works like a charm. In my case it creates an 87x181 logical matrix which is the same dimension as A. However, the statement "A(~condition)" creates a column vector that stacks all my numerical data from A on top of each other instead of preserving the 87x181 structure. This of course completely screws up the GEOMEAN calculation since it returns a single number instead of 181 different numbers.

Appreciate any help,

JK

Answer 1

[Star Strider]

If you have geomean, you have the Statistics Toolbox. If you want to take the geometric mean of your array (with NaN values included, use nanmean of the logs of the values:

A = randi(10, 15, 20);
A(A == 5) = NaN;
B = exp(nanmean(log(A)));

A bit indirect, but it gets you the result you want, with the dimensions you want.

Answer 2

[John D'Errico]

Well, what do you expect?

A = [1 2;NaN 4];

isnan(A)

ans =
   0     0
   1     0

Suppose I tried A(~isnan(A)) out? Would you expect to see a resulting matrix with one element in the first column, and two elements in the second column? Surely you understand that MATLAB does not allow this. A 2-d MATRIX in MATLAB is a square or rectangular array of numbers.

As such, the only result that can be is what MATLAB does do:

A(~isnan(A))
ans =
   1
   2
   4

There are tools of course that can handle the computations you apparently wish to do, such as nanmean.