avoid the 'for' loop in matlab script by permuting - matlab

I have a code as follows:
Ne = 100;
H = rand(Ne,Ne);
g = zeros(Ne,1);
for e =1:Ne
hue = H(:,e);
ss1 = bsxfun(#times, hue', hue) .* M; % M is a Ne*Ne matrix
g(e) = sum(ss1(:));
end
when Ne > 1000, it runs very slowly.
I read the matlab documents, and find permute function is a possible way to speed up.
But I tried a whole day and failed.
Here is my code, and I do not know what is wrong.
C = permute(bsxfun(#times, permute(H, [1 3 2]), permute(H', [1 3 2])), [1 3 2]);
g = sum(sum(C))


If you do the math, you'll see that all you have to do is this:
g = sum(H) .^ 2;
Run speed: 0.000681 seconds, with Ne = 1000 (the original code took 3.047315 seconds).
EDIT:
Now, for your edited code, all you have to do is this:
g = diag(H.' * M * H);
Run speed: 0.072273 seconds, with Ne = 1000.
A speedup can be obtained if you notice that if you rearrange the terms, you can avoid a second matrix multiplication (which changes to a dot product) and all you have to do is sum the columns, like this:
g = sum(M.' * H .* H);
Run speed: 0.044190 seconds, with Ne = 1000.
It's always a good idea to do the math. We spend some time, but the code gains a good speedup. :)
NOTE: Run speeds were measured by averaging the time of a hundred runs.


For your edited code, this would work -
H1 = permute(H,[1 3 2]);
H2 = permute(H,[3 1 2]);
p1 = bsxfun(#times,H2,H1);
p1 = bsxfun(#times,p1,M);
g = sum(reshape(p1,Ne*Ne,[]),1)';
There isn't a great performance improvement with this though, as it's a bit faster only on a limited range of small datasizes.

Related

Gradient descent for small and large values of multidimensional matrix

I have a matrix R which I want to estimate using gradient descent method. The code is working good for small dimension matrix (e.g. 100x1) but it gives NaN values as I increase dimension (e.g. 10x3). Please help!
R =[196,242,3;186,302,3;22,377,1;244,51,2;166,346,1;298,474,4;115,265,2;253,465,5;305,451,3;6,86,3];
N = length(R);
M = size(R,2);
K = 3;
P = rand(N,K);
Q = rand(M,K);
alpha = 0.002;
beta = 0.02;
for iter = 1 : 5000
e = R - P*Q';
P_new = P + 2*alpha*(e*Q-beta*P);
Q_new= Q + 2*alpha*(e'*P-beta*Q);
mse2(iter) = norm(R - P*Q')/norm(R);
P=P_new;
Q=Q_new;
end
R_est = P*Q';

Your problem lies in the for loop. The largest double matlab can store is 1.7977e+308, after 7 runs of the loop your values for P and Q are on the order of 10^101 and after 8 runs inf. I'm not sure what you mean by estimate matrix but your nanvalues are coming from the way P an Q grow so rapidly. One other things of note, the mse2 variable is not to used. What is its purpose?

Custom-made linspace and logspace in MATLAB

I decided to take a look at two functions linspace and logspace. Below I give two examples, one using MATLAB's built-in linspace and one for logspace along with their hand made implementation. In the first case both the built-in function linspace and the handmade code give the same results. However, this is not true when examining the logspace function. Could you please help me to found the error in the handmade code?
a = 1; b = 5; n = 7;
y = linspace(1,5,7);
yy = zeros(1,n); yy(1) = a;
for i=2:n
yy(i) = yy(i-1) + (b-a)/(n-1);
end
x = logspace(1,5,7);
xx = zeros(1,n); xx(1) = 10^a;
for i=2:n
xx(i) = xx(i-1) + (10^b-10^a)/(n-1);
end
Thank you!

The only difference between linspace and logspace is that they go one step further and take the power of 10 for every element in the linspace array.
As such, you'd simply take your equation for linspace you generated, take the result and raise it to the power of 10. However, with your code, you are relying on the previous result and that is already raised to the power of 10. Therefore, you'll need to take the anti-log to convert the previous result back to a linear form, then use the same logic was used to generate the linspace, then raise it back to the power of 10. Therefore, the relationship is:
xx[n] = 10^(log10(xx[n-1]) + ((b-a)/(n-1)))
You can certainly simplify this, taking advantage of the fact that 10^(log10(z)) = z, as long as z > 0. We can also split up the terms in the power using the property that 10^(m + n) = (10^m) * (10^n). Therefore:
xx[n] = xx[n-1] * (10^((b-a)/(n-1)))
As such, simply take your previous result multiply with 10^((b-a)/(n-1))
a = 1; b = 5; n = 7;
x = logspace(1,5,7);
xx = zeros(1,n); xx(1) = 10^a;
for i=2:n
xx(i) = xx(i-1)*(10^((b-a)/(n-1))); %// Change
end
We get for both x and xx:
>> format long g;
>> x
x =
Columns 1 through 4
10 46.4158883361278 215.443469003188 1000
Columns 5 through 7
4641.58883361278 21544.3469003189 100000
>> xx
xx =
Columns 1 through 4
10 46.4158883361278 215.443469003188 1000
Columns 5 through 7
4641.58883361278 21544.3469003188 100000

Get binomial coefficients

In an attempt to vectorize a particular piece of Matlab code, I could not find a straightforward function to generate a list of the binomial coefficients. The best I could find was nchoosek, but for some inexplicable reason this function only accepts integers (not vectors of integers). My current solution looks like this:
mybinom = #(n) arrayfun(#nchoosek, n*ones(1,n), 1:n)
This generates the set of binomial coefficients for a given value of n. However, since the binomial coefficients are always symmetric, I know that I am doing twice as much work as necessary. I'm sure that I could create a solution that exploits the symmetry, but I'm sure that it would be at the expense of readability.
Is there a more elegant solution than this, perhaps using a Matlab function that I am not aware of? Note that I am not interested in using the symbolic toolbox.

If you want to minimize operations you can go along these lines:
n = 6;
k = 1:n;
result = [1 cumprod((n-k+1)./k)]
>> result
result =
1 6 15 20 15 6 1
This requires very few operations per coefficient, because each cofficient is obtained exploiting the previously computed one.
You can reduce the number of operations by approximately half if you take into account the symmetry:
m1 = floor(n/2);
m2 = ceil(n/2);
k = 1:m2;
result = [1 cumprod((n-k+1)./k)];
result(n+1:-1:m1+2) = result(1:m2);

What about a modified version of Luis Mendo's solution - but in logarithms:
n = 1e4;
m1 = floor(n/2);
m2 = ceil(n/2);
k = 1:m2;
% Attempt to compute real value
out0 = [1 cumprod((n-k+1)./k)];
out0(n+1:-1:m1+2) = out0(1:m2);
% In logarithms
out1 = [0 cumsum((log(n-k+1)) - log(k))];
out1(n+1:-1:m1+2) = out1(1:m2);
plot(log(out0) - out1, 'o-')
The advantage of working with logarithms is that you can set n = 1e4; and still obtain a good approximation of the real value (nchoosek(1e4, 5e3) returns Inf and this is not a good approximation at all!).
EDIT following horchler's comment
You can use the gammaln function to obtain the same result but it's not faster. The two approximations seem to be quite different:
n = 1e7;
m1 = floor(n/2);
m2 = ceil(n/2);
k = 1:m2;
% In logarithms
tic
out1 = [0 cumsum((log(n-k+1)) - log(k))];
out1(n+1:-1:m1+2) = out1(1:m2);
toc
% Elapsed time is 0.912649 seconds.
tic
k = 0:m2;
out2 = gammaln(n + 1) - gammaln(k + 1) - gammaln(n - k + 1);
out2(n+1:-1:m1+2) = out2(1:m2);
toc
% Elapsed time is 1.020188 seconds.
tmp = out2 - out1;
plot(tmp, '.')
prctile(tmp, [0 2.5 25 50 75 97.5 100])
% 1.0e-006 *
% -0.2217 -0.1462 -0.0373 0.0363 0.1225 0.2943 0.3846
Is adding three gammaln worse than adding n logarithms? Or viceversa?

This works for Octave only
You can use bincoeff function.
Example: bincoeff(5, 0:5)
EDIT :
Only improvement I can think of goes like this. Maybe you already thought this trivial solution and didn't like it.
# Calculate only the first half
mybinomhalf = #(n) arrayfun(#nchoosek, n*ones(1,n/2+1), 0:n/2)
# pad your array symmetrically
mybinom = #(n) padarray(mybinomhalf(n), [0 n/2], 'symmetric', 'post')
# I couldn't test it and this line may not work

Optimizing repetitive estimation (currently a loop) in MATLAB

I've found myself needing to do a least-squares (or similar matrix-based operation) for every pixel in an image. Every pixel has a set of numbers associated with it, and so it can be arranged as a 3D matrix.
(This next bit can be skipped)
Quick explanation of what I mean by least-squares estimation :
Let's say we have some quadratic system that is modeled by Y = Ax^2 + Bx + C and we're looking for those A,B,C coefficients. With a few samples (at least 3) of X and the corresponding Y, we can estimate them by:
Arrange the (lets say 10) X samples into a matrix like X = [x(:).^2 x(:) ones(10,1)];
Arrange the Y samples into a similar matrix: Y = y(:);
Estimate the coefficients A,B,C by solving: coeffs = (X'*X)^(-1)*X'*Y;
Try this on your own if you want:
A = 5; B = 2; C = 1;
x = 1:10;
y = A*x(:).^2 + B*x(:) + C + .25*randn(10,1); % added some noise here
X = [x(:).^2 x(:) ones(10,1)];
Y = y(:);
coeffs = (X'*X)^-1*X'*Y
coeffs =
5.0040
1.9818
0.9241
START PAYING ATTENTION AGAIN IF I LOST YOU THERE
*MAJOR REWRITE*I've modified to bring it as close to the real problem that I have and still make it a minimum working example.
Problem Setup
%// Setup
xdim = 500;
ydim = 500;
ncoils = 8;
nshots = 4;
%// matrix size for each pixel is ncoils x nshots (an overdetermined system)
%// each pixel has a matrix stored in the 3rd and 4rth dimensions
regressor = randn(xdim,ydim, ncoils,nshots);
regressand = randn(xdim, ydim,ncoils);
So my problem is that I have to do a (X'*X)^-1*X'*Y (least-squares or similar) operation for every pixel in an image. While that itself is vectorized/matrixized the only way that I have to do it for every pixel is in a for loop, like:
Original code style
%// Actual work
tic
estimate = zeros(xdim,ydim);
for col=1:size(regressor,2)
for row=1:size(regressor,1)
X = squeeze(regressor(row,col,:,:));
Y = squeeze(regressand(row,col,:));
B = X\Y;
% B = (X'*X)^(-1)*X'*Y; %// equivalently
estimate(row,col) = B(1);
end
end
toc
Elapsed time = 27.6 seconds
EDITS in reponse to comments and other ideas
I tried some things:
1. Reshaped into a long vector and removed the double for loop. This saved some time.
2. Removed the squeeze (and in-line transposing) by permute-ing the picture before hand: This save alot more time.
Current example:
%// Actual work
tic
estimate2 = zeros(xdim*ydim,1);
regressor_mod = permute(regressor,[3 4 1 2]);
regressor_mod = reshape(regressor_mod,[ncoils,nshots,xdim*ydim]);
regressand_mod = permute(regressand,[3 1 2]);
regressand_mod = reshape(regressand_mod,[ncoils,xdim*ydim]);
for ind=1:size(regressor_mod,3) % for every pixel
X = regressor_mod(:,:,ind);
Y = regressand_mod(:,ind);
B = X\Y;
estimate2(ind) = B(1);
end
estimate2 = reshape(estimate2,[xdim,ydim]);
toc
Elapsed time = 2.30 seconds (avg of 10)
isequal(estimate2,estimate) == 1;
Rody Oldenhuis's way
N = xdim*ydim*ncoils; %// number of columns
M = xdim*ydim*nshots; %// number of rows
ii = repmat(reshape(1:N,[ncoils,xdim*ydim]),[nshots 1]); %//column indicies
jj = repmat(1:M,[ncoils 1]); %//row indicies
X = sparse(ii(:),jj(:),regressor_mod(:));
Y = regressand_mod(:);
B = X\Y;
B = reshape(B(1:nshots:end),[xdim ydim]);
Elapsed time = 2.26 seconds (avg of 10)
or 2.18 seconds (if you don't include the definition of N,M,ii,jj)
SO THE QUESTION IS:
Is there an (even) faster way?
(I don't think so.)

You can achieve a ~factor of 2 speed up by precomputing the transposition of X. i.e.
for x=1:size(picture,2) % second dimension b/c already transposed
X = picture(:,x);
XX = X';
Y = randn(n_timepoints,1);
%B = (X'*X)^-1*X'*Y; ;
B = (XX*X)^-1*XX*Y;
est(x) = B(1);
end
Before: Elapsed time is 2.520944 seconds.
After: Elapsed time is 1.134081 seconds.
EDIT:
Your code, as it stands in your latest edit, can be replaced by the following
tic
xdim = 500;
ydim = 500;
n_timepoints = 10; % for example
% Actual work
picture = randn(xdim,ydim,n_timepoints);
picture = reshape(picture, [xdim*ydim,n_timepoints])'; % note transpose
YR = randn(n_timepoints,size(picture,2));
% (XX*X).^-1 = sum(picture.*picture).^-1;
% XX*Y = sum(picture.*YR);
est = sum(picture.*picture).^-1 .* sum(picture.*YR);
est = reshape(est,[xdim,ydim]);
toc
Elapsed time is 0.127014 seconds.
This is an order of magnitude speed up on the latest edit, and the results are all but identical to the previous method.
EDIT2:
Okay, so if X is a matrix, not a vector, things are a little more complicated. We basically want to precompute as much as possible outside of the for-loop to keep our costs down. We can also get a significant speed-up by computing XT*X manually - since the result will always be a symmetric matrix, we can cut a few corners to speed things up. First, the symmetric multiplication function:
function XTX = sym_mult(X) % X is a 3-d matrix
n = size(X,2);
XTX = zeros(n,n,size(X,3));
for i=1:n
for j=i:n
XTX(i,j,:) = sum(X(:,i,:).*X(:,j,:));
if i~=j
XTX(j,i,:) = XTX(i,j,:);
end
end
end
Now the actual computation script
xdim = 500;
ydim = 500;
n_timepoints = 10; % for example
Y = randn(10,xdim*ydim);
picture = randn(xdim,ydim,n_timepoints); % 500x500x10
% Actual work
tic % start timing
picture = reshape(picture, [xdim*ydim,n_timepoints])';
% Here we precompute the (XT*Y) calculation to speed things up later
picture_y = [sum(Y);sum(Y.*picture)];
% initialize
est = zeros(size(picture,2),1);
picture = permute(picture,[1,3,2]);
XTX = cat(2,ones(n_timepoints,1,size(picture,3)),picture);
XTX = sym_mult(XTX); % precompute (XT*X) for speed
X = zeros(2,2); % preallocate for speed
XY = zeros(2,1);
for x=1:size(picture,2) % second dimension b/c already transposed
%For some reason this is a lot faster than X = XTX(:,:,x);
X(1,1) = XTX(1,1,x);
X(2,1) = XTX(2,1,x);
X(1,2) = XTX(1,2,x);
X(2,2) = XTX(2,2,x);
XY(1) = picture_y(1,x);
XY(2) = picture_y(2,x);
% Here we utilise the fact that A\B is faster than inv(A)*B
% We also use the fact that (A*B)*C = A*(B*C) to speed things up
B = X\XY;
est(x) = B(1);
end
est = reshape(est,[xdim,ydim]);
toc % end timing
Before: Elapsed time is 4.56 seconds.
After: Elapsed time is 2.24 seconds.
This is a speed up of about a factor of 2. This code should be extensible to X being any dimensions you want. For instance, in the case where X = [1 x x^2], you would change picture_y to the following
picture_y = [sum(Y);sum(Y.*picture);sum(Y.*picture.^2)];
and change XTX to
XTX = cat(2,ones(n_timepoints,1,size(picture,3)),picture,picture.^2);
You would also change a lot of 2s to 3s in the code, and add XY(3) = picture_y(3,x) to the loop. It should be fairly straight-forward, I believe.

Results
I sped up your original version, since your edit 3 was actually not working (and also does something different).
So, on my PC:
Your (original) version: 8.428473 seconds.
My obfuscated one-liner given below: 0.964589 seconds.
First, for no other reason than to impress, I'll give it as I wrote it:
%%// Some example data
xdim = 500;
ydim = 500;
n_timepoints = 10; % for example
estimate = zeros(xdim,ydim); %// initialization with explicit size
picture = randn(xdim,ydim,n_timepoints);
%%// Your original solution
%// (slightly altered to make my version's results agree with yours)
tic
Y = randn(n_timepoints,xdim*ydim);
ii = 1;
for x = 1:xdim
for y = 1:ydim
X = squeeze(picture(x,y,:)); %// or similar creation of X matrix
B = (X'*X)^(-1)*X' * Y(:,ii);
ii = ii+1;
%// sometimes you keep everything and do
%// estimate(x,y,:) = B(:);
%// sometimes just the first element is important and you do
estimate(x,y) = B(1);
end
end
toc
%%// My version
tic
%// UNLEASH THE FURY!!
estimate2 = reshape(sparse(1:xdim*ydim*n_timepoints, ...
builtin('_paren', ones(n_timepoints,1)*(1:xdim*ydim),:), ...
builtin('_paren', permute(picture, [3 2 1]),:))\Y(:), ydim,xdim).'; %'
toc
%%// Check for equality
max(abs(estimate(:)-estimate2(:))) % (always less than ~1e-14)
Breakdown
First, here's the version that you should actually use:
%// Construct sparse block-diagonal matrix
%// (Type "help sparse" for more information)
N = xdim*ydim; %// number of columns
M = N*n_timepoints; %// number of rows
ii = 1:N;
jj = ones(n_timepoints,1)*(1:N);
s = permute(picture, [3 2 1]);
X = sparse(ii,jj(:), s(:));
%// Compute ALL the estimates at once
estimates = X\Y(:);
%// You loop through the *second* dimension first, so to make everything
%// agree, we have to extract elements in the "wrong" order, and transpose:
estimate2 = reshape(estimates, ydim,xdim).'; %'
Here's an example of what picture and the corresponding matrix X looks like for xdim = ydim = n_timepoints = 2:
>> clc, picture, full(X)
picture(:,:,1) =
-0.5643 -2.0504
-0.1656 0.4497
picture(:,:,2) =
0.6397 0.7782
0.5830 -0.3138
ans =
-0.5643 0 0 0
0.6397 0 0 0
0 -2.0504 0 0
0 0.7782 0 0
0 0 -0.1656 0
0 0 0.5830 0
0 0 0 0.4497
0 0 0 -0.3138
You can see why sparse is necessary -- it's mostly zeros, but will grow large quickly. The full matrix would quickly consume all your RAM, while the sparse one will not consume much more than the original picture matrix does.
With this matrix X, the new problem
X·b = Y
now contains all the problems
X1 · b1 = Y1
X2 · b2 = Y2
...
where
b = [b1; b2; b3; ...]
Y = [Y1; Y2; Y3; ...]
so, the single command
X\Y
will solve all your systems at once.
This offloads all the hard work to a set of highly specialized, compiled to machine-specific code, optimized-in-every-way algorithms, rather than the interpreted, generic, always-two-steps-away from the hardware loops in MATLAB.
It should be straightforward to convert this to a version where X is a matrix; you'll end up with something like what blkdiag does, which can also be used by mldivide in exactly the same way as above.

I had a wee play around with an idea, and I decided to stick it as a separate answer, as its a completely different approach to my other idea, and I don't actually condone what I'm about to do. I think this is the fastest approach so far:
Orignal (unoptimised): 13.507176 seconds.
Fast Cholesky-decomposition method: 0.424464 seconds
First, we've got a function to quickly do the X'*X multiplication. We can speed things up here because the result will always be symmetric.
function XX = sym_mult(X)
n = size(X,2);
XX = zeros(n,n,size(X,3));
for i=1:n
for j=i:n
XX(i,j,:) = sum(X(:,i,:).*X(:,j,:));
if i~=j
XX(j,i,:) = XX(i,j,:);
end
end
end
The we have a function to do LDL Cholesky decomposition of a 3D matrix (we can do this because the (X'*X) matrix will always be symmetric) and then do forward and backwards substitution to solve the LDL inversion equation
function Y = fast_chol(X,XY)
n=size(X,2);
L = zeros(n,n,size(X,3));
D = zeros(n,n,size(X,3));
B = zeros(n,1,size(X,3));
Y = zeros(n,1,size(X,3));
% These loops compute the LDL decomposition of the 3D matrix
for i=1:n
D(i,i,:) = X(i,i,:);
L(i,i,:) = 1;
for j=1:i-1
L(i,j,:) = X(i,j,:);
for k=1:(j-1)
L(i,j,:) = L(i,j,:) - L(i,k,:).*L(j,k,:).*D(k,k,:);
end
D(i,j,:) = L(i,j,:);
L(i,j,:) = L(i,j,:)./D(j,j,:);
if i~=j
D(i,i,:) = D(i,i,:) - L(i,j,:).^2.*D(j,j,:);
end
end
end
for i=1:n
B(i,1,:) = XY(i,:);
for j=1:(i-1)
B(i,1,:) = B(i,1,:)-D(i,j,:).*B(j,1,:);
end
B(i,1,:) = B(i,1,:)./D(i,i,:);
end
for i=n:-1:1
Y(i,1,:) = B(i,1,:);
for j=n:-1:(i+1)
Y(i,1,:) = Y(i,1,:)-L(j,i,:).*Y(j,1,:);
end
end
Finally, we have the main script which calls all of this
xdim = 500;
ydim = 500;
n_timepoints = 10; % for example
Y = randn(10,xdim*ydim);
picture = randn(xdim,ydim,n_timepoints); % 500x500x10
tic % start timing
picture = reshape(pr, [xdim*ydim,n_timepoints])';
% Here we precompute the (XT*Y) calculation
picture_y = [sum(Y);sum(Y.*picture)];
% initialize
est2 = zeros(size(picture,2),1);
picture = permute(picture,[1,3,2]);
% Now we calculate the X'*X matrix
XTX = cat(2,ones(n_timepoints,1,size(picture,3)),picture);
XTX = sym_mult(XTX);
% Call our fast Cholesky decomposition routine
B = fast_chol(XTX,picture_y);
est2 = B(1,:);
est2 = reshape(est2,[xdim,ydim]);
toc
Again, this should work equally well for a Nx3 X matrix, or however big you want.

I use octave, thus I can't say anything about the resulting performance in Matlab, but would expect this code to be slightly faster:
pictureT=picture'
est=arrayfun(#(x)( (pictureT(x,:)*picture(:,x))^-1*pictureT(x,:)*randn(n_ti
mepoints,1)),1:size(picture,2));

Multiply a 3D matrix with a 2D matrix

Suppose I have an AxBxC matrix X and a BxD matrix Y.
Is there a non-loop method by which I can multiply each of the C AxB matrices with Y?

As a personal preference, I like my code to be as succinct and readable as possible.
Here's what I would have done, though it doesn't meet your 'no-loops' requirement:
for m = 1:C
Z(:,:,m) = X(:,:,m)*Y;
end
This results in an A x D x C matrix Z.
And of course, you can always pre-allocate Z to speed things up by using Z = zeros(A,D,C);.

You can do this in one line using the functions NUM2CELL to break the matrix X into a cell array and CELLFUN to operate across the cells:
Z = cellfun(#(x) x*Y,num2cell(X,[1 2]),'UniformOutput',false);
The result Z is a 1-by-C cell array where each cell contains an A-by-D matrix. If you want Z to be an A-by-D-by-C matrix, you can use the CAT function:
Z = cat(3,Z{:});
NOTE: My old solution used MAT2CELL instead of NUM2CELL, which wasn't as succinct:
[A,B,C] = size(X);
Z = cellfun(#(x) x*Y,mat2cell(X,A,B,ones(1,C)),'UniformOutput',false);

Here's a one-line solution (two if you want to split into 3rd dimension):
A = 2;
B = 3;
C = 4;
D = 5;
X = rand(A,B,C);
Y = rand(B,D);
%# calculate result in one big matrix
Z = reshape(reshape(permute(X, [2 1 3]), [A B*C]), [B A*C])' * Y;
%'# split into third dimension
Z = permute(reshape(Z',[D A C]),[2 1 3]);
Hence now: Z(:,:,i) contains the result of X(:,:,i) * Y
Explanation:
The above may look confusing, but the idea is simple.
First I start by take the third dimension of X and do a vertical concatenation along the first dim:
XX = cat(1, X(:,:,1), X(:,:,2), ..., X(:,:,C))
... the difficulty was that C is a variable, hence you can't generalize that expression using cat or vertcat. Next we multiply this by Y:
ZZ = XX * Y;
Finally I split it back into the third dimension:
Z(:,:,1) = ZZ(1:2, :);
Z(:,:,2) = ZZ(3:4, :);
Z(:,:,3) = ZZ(5:6, :);
Z(:,:,4) = ZZ(7:8, :);
So you can see it only requires one matrix multiplication, but you have to reshape the matrix before and after.

I'm approaching the exact same issue, with an eye for the most efficient method. There are roughly three approaches that i see around, short of using outside libraries (i.e., mtimesx):
Loop through slices of the 3D matrix
repmat-and-permute wizardry
cellfun multiplication
I recently compared all three methods to see which was quickest. My intuition was that (2) would be the winner. Here's the code:
% generate data
A = 20;
B = 30;
C = 40;
D = 50;
X = rand(A,B,C);
Y = rand(B,D);
% ------ Approach 1: Loop (via #Zaid)
tic
Z1 = zeros(A,D,C);
for m = 1:C
Z1(:,:,m) = X(:,:,m)*Y;
end
toc
% ------ Approach 2: Reshape+Permute (via #Amro)
tic
Z2 = reshape(reshape(permute(X, [2 1 3]), [A B*C]), [B A*C])' * Y;
Z2 = permute(reshape(Z2',[D A C]),[2 1 3]);
toc
% ------ Approach 3: cellfun (via #gnovice)
tic
Z3 = cellfun(#(x) x*Y,num2cell(X,[1 2]),'UniformOutput',false);
Z3 = cat(3,Z3{:});
toc
All three approaches produced the same output (phew!), but, surprisingly, the loop was the fastest:
Elapsed time is 0.000418 seconds.
Elapsed time is 0.000887 seconds.
Elapsed time is 0.001841 seconds.
Note that the times can vary quite a lot from one trial to another, and sometimes (2) comes out the slowest. These differences become more dramatic with larger data. But with much bigger data, (3) beats (2). The loop method is still best.
% pretty big data...
A = 200;
B = 300;
C = 400;
D = 500;
Elapsed time is 0.373831 seconds.
Elapsed time is 0.638041 seconds.
Elapsed time is 0.724581 seconds.
% even bigger....
A = 200;
B = 200;
C = 400;
D = 5000;
Elapsed time is 4.314076 seconds.
Elapsed time is 11.553289 seconds.
Elapsed time is 5.233725 seconds.
But the loop method can be slower than (2), if the looped dimension is much larger than the others.
A = 2;
B = 3;
C = 400000;
D = 5;
Elapsed time is 0.780933 seconds.
Elapsed time is 0.073189 seconds.
Elapsed time is 2.590697 seconds.
So (2) wins by a big factor, in this (maybe extreme) case. There may not be an approach that is optimal in all cases, but the loop is still pretty good, and best in many cases. It is also best in terms of readability. Loop away!

Nope. There are several ways, but it always comes out in a loop, direct or indirect.
Just to please my curiosity, why would you want that anyway ?

To answer the question, and for readability, please see:
ndmult, by ajuanpi (Juan Pablo Carbajal), 2013, GNU GPL
Input
2 arrays
dim
Example
nT = 100;
t = 2*pi*linspace (0,1,nT)’;
# 2 experiments measuring 3 signals at nT timestamps
signals = zeros(nT,3,2);
signals(:,:,1) = [sin(2*t) cos(2*t) sin(4*t).^2];
signals(:,:,2) = [sin(2*t+pi/4) cos(2*t+pi/4) sin(4*t+pi/6).^2];
sT(:,:,1) = signals(:,:,1)’;
sT(:,:,2) = signals(:,:,2)’;
G = ndmult (signals,sT,[1 2]);
Source
Original source. I added inline comments.
function M = ndmult (A,B,dim)
dA = dim(1);
dB = dim(2);
# reshape A into 2d
sA = size (A);
nA = length (sA);
perA = [1:(dA-1) (dA+1):(nA-1) nA dA](1:nA);
Ap = permute (A, perA);
Ap = reshape (Ap, prod (sA(perA(1:end-1))), sA(perA(end)));
# reshape B into 2d
sB = size (B);
nB = length (sB);
perB = [dB 1:(dB-1) (dB+1):(nB-1) nB](1:nB);
Bp = permute (B, perB);
Bp = reshape (Bp, sB(perB(1)), prod (sB(perB(2:end))));
# multiply
M = Ap * Bp;
# reshape back to original format
s = [sA(perA(1:end-1)) sB(perB(2:end))];
M = squeeze (reshape (M, s));
endfunction

I highly recommend you use the MMX toolbox of matlab. It can multiply n-dimensional matrices as fast as possible.
The advantages of MMX are:
It is easy to use.
Multiply n-dimensional matrices (actually it can multiply arrays of 2-D matrices)
It performs other matrix operations (transpose, Quadratic Multiply, Chol decomposition and more)
It uses C compiler and multi-thread computation for speed up.
For this problem, you just need to write this command:
C=mmx('mul',X,Y);
here is a benchmark for all possible methods. For more detail refer to this question.
1.6571 # FOR-loop
4.3110 # ARRAYFUN
3.3731 # NUM2CELL/FOR-loop/CELL2MAT
2.9820 # NUM2CELL/CELLFUN/CELL2MAT
0.0244 # Loop Unrolling
0.0221 # MMX toolbox <===================

I would like to share my answer to the problems of:
1) making the tensor product of two tensors (of any valence);
2) making the contraction of two tensors along any dimension.
Here are my subroutines for the first and second tasks:
1) tensor product:
function [C] = tensor(A,B)
C = squeeze( reshape( repmat(A(:), 1, numel(B)).*B(:).' , [size(A),size(B)] ) );
end
2) contraction:
Here A and B are the tensors to be contracted along the dimesions i and j respectively. The lengths of these dimensions should be equal, of course. There's no check for this (this would obscure the code) but apart from this it works well.
function [C] = tensorcontraction(A,B, i,j)
sa = size(A);
La = length(sa);
ia = 1:La;
ia(i) = [];
ia = [ia i];
sb = size(B);
Lb = length(sb);
ib = 1:Lb;
ib(j) = [];
ib = [j ib];
% making the i-th dimension the last in A
A1 = permute(A, ia);
% making the j-th dimension the first in B
B1 = permute(B, ib);
% making both A and B 2D-matrices to make use of the
% matrix multiplication along the second dimension of A
% and the first dimension of B
A2 = reshape(A1, [],sa(i));
B2 = reshape(B1, sb(j),[]);
% here's the implicit implication that sa(i) == sb(j),
% otherwise - crash
C2 = A2*B2;
% back to the original shape with the exception
% of dimensions along which we've just contracted
sa(i) = [];
sb(j) = [];
C = squeeze( reshape( C2, [sa,sb] ) );
end
Any critics?

I would think recursion, but that's the only other non- loop method you can do

You could "unroll" the loop, ie write out all the multiplications sequentially that would occur in the loop