0
$\begingroup$

I am trying to create a sequence of functions and have it properly memoize the results. The recursive operation is simply convolution, so it possible there is a better way to do this (obviously, if I could find a closed form expression for the n-th convolution, that would be ideal, but I'm not certain I can do that). Here is my current code

f[x_, c2_] := (
 Exp[-c2^2/2 + c2 Sqrt[-Log[2 π] - 2 Log[x]]] + 
 Exp[-c2^2/2 - c2 Sqrt[-Log[2 π] - 2 Log[x]]]
) / Sqrt[(-Log[2 π] - 2 Log[x])]

LogExpect[x_, shift_] := Piecewise[{{f[Exp[x], shift] Exp[x], x < -1/2 Log[2 π]}}, 0]

LogExpectN[t_, n_] := Module[{xx}, 
  LogExpectN[t, n] = Convolve[LogExpectN[xx, n - 1], LogExpect[xx, 0], xx, t]
]
LogExpectN[t_, 1] = LogExpect[t, 0]

In general this works, as it will properly calculate LogExpectN[t,5] for example, and it will memorize that specific result, but if I ask it for LogExpectN[2 t,5] or LogExpectN[x,5] it has to redo the copmutation completely, instead of just plugging 2t or y into the function it already found and thus there is no speed up from the memoization. Is there a way to get the behavior I want?

Edit: By changing the the recursive definition to use a pattern, I have improved things

LogExpectN[t_, n_] := Module[{xx}, LogExpectN[t_, n] = Convolve[LogExpectN[xx, n - 1], LogExpect[xx, 0], xx, t]]

However, This only works if I evaluate consecutive values of n in order. FOr example, after setting the base case for n=1, I can calculate for n=2 and then n=3, but if I try to calculate n=3 directly, it will give me an expression in terms of the internal variable xx$<numbers>$. I am not really sure what is going on here.

$\endgroup$
  • $\begingroup$ Related, possibly duplicate: (53) $\endgroup$ – Mr.Wizard Jul 3 '18 at 12:15
2
$\begingroup$

Memoize a worker function.

f[x_] := f[x] = (Print["Memoized ", x]; f[x - 1] + 1);
f[0] = 0;
{f[3], f[3]}
(*the first f[3] prints trice, the second does not print*)

Then, use the worker function to memoize your main function.

SetAttributes[g, HoldFirst]
g[x_] := g[x] = f[x];
{g[2 2], g[2 2 + 1]}
(*g[2 2] prints 4, g[2 2 + 1] prints 5*)

Check that the downvalues of the main function have your required behavior.

DownValues[g]
(*there are entries for 2 2 and 2 2 + 1*)

Note that you cannot combine the worker function and the main function in one step because HoldFirst will not allow to get to the stopping h[0].

SetAttributes[h, HoldFirst]; 
h[x_] := h[x] = (Print["Memoized ", x]; h[x - 1] + 1);
h[0] = 0;
h[3]
(*Recursion depth of 1024 exceeded*)

UPDATE

Consider

LogExpectN[t, n] = worker[n][t]

Because there are two variables, there are two memoizations involved. One to compute the n-th worker function and another to avoid replacing t in that worker function. The code before this update takes care of second type of memoization. The first memoization is trickier because you are computing a function.

f[x_, c2_] := (Exp[-c2^2/2 + c2 Sqrt[-Log[2 π] - 2 Log[x]]] + 
 Exp[-c2^2/2 - c2 Sqrt[-Log[2 π] - 2 Log[x]]])/
   Sqrt[(-Log[2 π] - 2 Log[x])];

LogExpectN[t_, n_] := LogExpectN[t, n] = (Print["Memoizing t=", t]; worker[n][t]);

worker[n_] := 
  worker[n] = 
   Module[{xx, tt, integral}, Print["Memoizing n=", n]; 
    integral = Convolve[worker[n - 1][xx], stoppper[0][xx], xx, tt]; 
     Function @@ {tt, integral}];

worker[1] = stoppper[0];

stoppper[shift_] := 
  Function[{x}, 
   Piecewise[{{f[Exp[x], shift] Exp[x], x < -1/2 Log[2 π]}}, 0]];

LogExpectN[t, 2]
(*Prints t=t and n=2*)

LogExpectN[t, 3]
(*Prints t=t and n=3*)

LogExpectN[3 t, 2]
(*Prints t=3 t only*)
$\endgroup$
  • $\begingroup$ Thanks. When I manage to apply this to my code I will accept the answer. Still trying to figure out the how and why of this. $\endgroup$ – irh Jul 3 '18 at 12:44
  • $\begingroup$ After futzing around, I can't get this approach to work. $\endgroup$ – irh Jul 3 '18 at 14:08
  • $\begingroup$ @irh Maybe, your problem is needing to evaluate the integral before memoizing the worker function. In the update, check the usage of Function @@. $\endgroup$ – Hector Jul 13 '18 at 16:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.