This is a rather common issue that comes up with many numerical functions (FindRoot, NIntegrate, FindMaximum, NMaximize, etc.) It is also explained in this Wolfram Knowledge Base article. Sometimes you want to pass these functions an expression that has a symbolic parameter, and compute the result for different values of that parameter.
Example:
fun[a_] := Block[{b}, b /. NMinimize[(a^2 + b)^2, {b}][[2]]]
This will work nicely if you call it with a numeric argument: fun[3]. But it will cause an error in NMinimize if you call it with a symbolic parameter: fun[a] (for obvious reasons).
The solution is:
Clear[fun]
fun[a_?NumericQ] := Block[{b}, b /. NMinimize[(a^2 + b)^2, {b}][[2]]]
NMaximize[fun[a], {a}]
(Be sure to evaluate Clear[...] to get rid of the previous definition of fun!)
This ensures that fun will only evaluate for numerical arguments, i.e. fun[a] won't evaluate inside NMaximize before NMaximize actually substitutes a number for a.
And this is also the answer to your specific question: make the inner NMinimize expression a separate function, and make sure it only evaluates for numerical arguments.
Requested edit
An important related point is: how can we match only numerical quantities using a pattern? One might think of using _Real (as in the comment below). The problem with this is that it will only match numbers whose Head is Real. This excludes integers (such as 1,2,3), rationals (2/3, 4/5), constants (such as Pi or E), or expressions like Sqrt[2].
The only robust solution is using NumericQ[] (x_ ? NumericQ in a pattern). NumericQ will return True for anything that gives a number when N[] is applied to it.
There's another related function, NumberQ[], which gives True only for objects with Integer, Rational, Real or Complex, but not for constant or expressions (Pi or Sin[3]).
fun[a_Real]seems to work as well in my toy example (I hoped to make it faster that way, but it didn't have any significant effect). Are there cases forNMinimizewherea_Realwould fail, buta_?NumericQwouldn't? – celtschk Jan 20 '12 at 10:13NMinimize. It's clear to me that the function then will not work if given symbolic or exact arguments; however, it's not clear to me whetherNMinimizewill ever pass anything not matching_Realto the function. – celtschk Jan 20 '12 at 10:29NumericQis well optimized. I'll dig up a benchmark later. – Szabolcs Jan 20 '12 at 10:32NMinimize/NMaximize: this is simply not an issue here because checking the pattern is much faster than (negligible compared to) theNMinimizeinside the function. – Szabolcs Jan 20 '12 at 11:41