Gradient and NaN

What does the function gradient does with nan?

Answer 1

[Matt Fig]

You can see what MATLAB does by looking at a plot. Look at the difference between the two figures:

v = -2:0.5:2;
[x,y] = meshgrid(v);
z = x .* exp(-x.^2 - y.^2);
[px,py] = gradient(z,.5,.5);
quiver(v,v,px,py),
z(5,5) = nan;  % Set the central data to nan.
[px2,py2] = gradient(z,.2,.2);
figure
quiver(v,v,px2,py2)

Answer 2

[Walter Roberson]

>> [fx,fy] = gradient([1 nan 3;5 9 15])
fx =
   NaN     1   NaN
     4     5     6
fy =
     4   NaN    12
     4   NaN    12

In other words, any location whose value would involve using the nan in the calculation, gets calculated as nan. Which is the result you would expect if the routine did simple arithmetic operations without any checking for nan.