56 views (last 30 days)

Show older comments

Here is a profile of the call n = 2*10^3; M = DStochMat02(n,ones(n)./n);

More specifically, can the hot-spot, statement 14, be reduced?

time calls line

1 function M = DStochMat02(n,c)

2 % Generate a random doubly stochastic matrix using

3 % Theorem (Birkhoff [1946], von Neumann [1953])

4 % Any doubly stochastic matrix M can be written as a convex combination

5 % of permutation matrices P1,...,Pk (i.e. M = c1*P1+...+ ck*Pk

6 % for nonnegative c1,...,ck with c1+...+ck = 1).

7 % Complexity: O(n^2)

8 % USE: M = DStochMat02(4,[1/2 1/8 1/8 1/4])

9 % Derek O'Connor, Oct 2006, Nov 2012. derekroconnor@eircom.net

0.02 1 10 M = zeros(n,n);

< 0.01 1 11 I = eye(n);

< 0.01 1 12 for k = 1:n

1.64 2000 13 pidx = GRPdur(n); % Random Permutation

107.72 2000 14 P = I(pidx,:); % Random P matrix

41.09 2000 15 M = M + c(k)*P;

< 0.01 2000 16 end

%--------------------------------------------------------------

function p = GRPdur(n)

% -------------------------------------------------------------

% Generate a random permutation p(1:n) using Durstenfeld's

% O(n) Shuffle Algorithm, CACM, 1964.

% See Knuth, Section 3.4.2, TAOCP, Vol 2, 3rd Ed.

% Complexity: O(n)

% USE: p = GRPdur(10^7);

% Derek O'Connor, 8 Dec 2010. derekroconnor@eircom.net

% -------------------------------------------------------------

p = 1:n; % Start with Identity permutation

for k = n:-1:2

r = 1+floor(rand*k); % random integer between 1 and k

t = p(k);

p(k) = p(r); % Swap(p(r),p(k)).

p(r) = t;

end

return % GRPdur

Matt Fig
on 17 Nov 2012

Edited: Matt Fig
on 17 Nov 2012

This looks promising. I have kept some notes in the comments so you can see what I was doing...

There are three lines of code that are commented out. If you uncomment them, and comment the line after the second one, you can check that these give the same results. Of course the timing is not fair under those circumstances, but with the code commented the timings show a fantastic speed difference for large values....

M = zeros(n,n);

I = eye(n);

tic % Original method.

for k = 1:n

pidx = GRPdur(n); % Random Permutation

% F{k} = pidx; % Used to check same results below....

P = I(pidx,:); % Random P matrix

M = M + c(k)*P;

end

toc

M2 = zeros(n,n);

tic % proposed alternative

for k = 1:n

% [~,idx] = sort(F{k}); % Used to compare same results.

[~,idx] = sort(GRPdur(n)); % Or just idx = GRPdur(n); ??

idx = idx + (0:n-1)*n;

M2(idx) = M2(idx) + c(k);

end

toc

% max(abs(M2(:)-M(:))) % Used to compare same results.

Matt Fig
on 18 Nov 2012

Hi Derek, I am glad you found the results useful. What I did was basically take the operations from matrix operations to vector operations. Since you are multiplying a scalar (c(k)) by an identity matrix, there is no need to keep all the zeros. I simply take the single vector of values (there are n c(k) values) and index that directly into only those elements of M that are actually updating - thus avoiding the unnecessary addition of

K = n^2 - n

zeros each time through the loop. There is no need to add K zeros to M each time through! In addition, my approach avoids multiplying K+n numbers by c(k) each time through the loop, when we don't even need to do a single such multiplication. When n is large these operations takes a significant amount of time and memory acrobatics....

Also, note that I only kept the call to SORT in there in order to make sure that the codes produced the exact same result (when the appropriate lines are uncommented of course). After hundreds of runs, my code does produce the exact same result. Thus I think you could avoid the call to SORT all together unless you have some statistical connection to GRPdur you are trying to preserve.

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!