Matlab matrix-multiplication and transpose precision - matlab

The next script should print out a 7x7 all-ones matrix, because the equation is satisfied.
A = rand(5,7);
B = rand(5,7);
C = (A' * B)';
D = B' * A;
C == D
Instead of this kind of answer:
ans =
1 1 1 1 0 1 1
1 1 1 1 0 1 0
1 1 1 1 1 1 1
1 1 1 1 0 0 0
1 0 1 1 1 1 1
0 0 1 1 1 1 1
0 1 1 0 1 1 1
I think this is a floating-point precision problem, because with format long the numbers differ in C and D.
What do I do wrong?
Where does it go wrong?
How can I avoid it?

You don't do anything wrong - the computer has finite precision, and your calculation reveals it - just like 1e6 + 0.1 - 1e6 (try it in Matlab). One way to avoid it is to use some library for arbitrary precision - but it won't 'solve' it, just push the problem towards smaller and smaller numbers.
See these links for some more info:
http://floating-point-gui.de/errors/comparison/
http://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html
By the way, format long has nothing to do with the actual precision, it just sets the way the numbers are formatted for displaying.

Related

Combinations of zeros and ones of vector size n [duplicate]

I have an algorith that the number of possibles combinations of 0 and 1, can reach the number 2^39. Let's say i have n=2 situations, or n1=2^2=4 combinations of 0 and 1: 00,01,10,11.From that i can create an array a=zeros(n,n1) and fill the columns with the possible combinations? That means first column has 00,second 01,third 10,last 11.I want this to be dynamic that means that n can be 1,2,3...,39, show the array will be a=zeros(n,2^n).Thanks for any response!
Just for general understanding: why do you think you need an array of all combinations of all integers from 0 to 2³⁹? That array would consume 39×2³⁹/1000⁴ ≈ 21TB of RAM...last time I checked, only the world's most advanced supercomputers have such resources, and most people working with those machines consider generating arrays like this quite wasteful...
Anyway, for completeness, for any N, this is the simplest solution:
P = dec2bin(0:(2^N)-1)-'0'
But, a little piece of advice: dec2bin outputs character arrays. If you want numerical arrays, you can subtract the character '0', however, that gives you an array of doubles according to the rules of MATLAB:
>> P = dec2bin(0:(2^3)-1)-'0';
>> whos P
Name Size Bytes Class Attributes
P 8x3 192 double
If you want to minimize your memory consumption, generate a logical array instead:
>> P = dec2bin(0:(2^3)-1)=='1';
>> whos P
Name Size Bytes Class Attributes
P 8x3 24 logical
If you want to also speed up the execution, use the standard algorithm directly:
%// if you like cryptic one-liners
B1 = rem(floor((0:pow2(N)-1).' * pow2(1-N:0)), 2) == 1;
%// If you like readability
B = false(N,pow2(N));
V = 0:pow2(N)-1;
for ii = 1:N
B(ii,:) = rem(V,2)==1;
V = (V-B(ii,:))/2;
end
That last one (the loop) is fastest of all solutions for any N (at least on R2010b and R2013a), and it has the smallest peak memory (only 1/Nth of the cryptic one-liner).
So I'd go for that one :)
But, that's just me.
Using ndgrid with a comma-separated list as output (see also here):
[c{1:N}] = ndgrid(logical([0 1]));
c = cat(N+1,c{N:-1:1});
c = reshape(c,[],N);
Example: N=4 gives
c =
0 0 0 0
0 0 0 1
0 0 1 0
0 0 1 1
0 1 0 0
0 1 0 1
0 1 1 0
0 1 1 1
1 0 0 0
1 0 0 1
1 0 1 0
1 0 1 1
1 1 0 0
1 1 0 1
1 1 1 0
1 1 1 1

Fastest way to copy values between two matrix

I am looking for a fastest way to copy some special value of a matrix to other matrix. Assume that I have matrix A such as
A =[4 1 5 4 4
-2 -1 1 2 2
3 -1 1 7 3
5 3 -1 1 -2
6 4 4 -1 1]
My aim is that copy element that have value 1 and -1 to matrix B. The expected matrix B such as
B =[ 0 1 0 0 0
0 -1 1 0 0
0 -1 1 0 0
0 0 -1 1 0
0 0 0 -1 1]
I performed two way to create matrix B. However, I think that my way is still not fastest way if size of matrix A becomes larger. I know that the forum has many expert matlab guy. Could you suggest to me another way?
This is my code
%%First way:
tic;B=((A==1)|(A==-1)).*A;toc
Elapsed time is 0.000026 seconds.
%%Second way:
tic;idx1=find(A==1);idx2=find(A==-1);B=zeros(size(A));B(idx1)=1; B(idx2)=-1;toc;B
Elapsed time is 0.000034 seconds.
here's somthing on par with #thewaywewalk
B=A.*reshape(abs(A(:))==1,size(A));
This is how I test these:
A=randi(10,1000,1000)-7;
B1=#() ((A==1)|(A==-1)).*A;
B2=#() (abs(A) == 1).*A;
B3=#() A.*reshape(abs(A(:))==1,size(A));
timeit(B1)
ans =
0.0136
timeit(B2)
ans =
0.0080
timeit(B3)
ans =
0.0079
These will change from run to run, but the methods are on par...
here's the same test on a range of matrix sizes:
The only thing which comes to my mind, which could be faster:
B = (abs(A) == 1).*A;

How to replace non-zero elements randomly with zero?

I have a matrix including 1 and 0 elements like below which is used as a network adjacency matrix.
A =
0 1 1 1
1 1 0 1
1 1 0 1
1 1 1 0
I want to simulate an attack on the network, so I must replace some specific percent of 1 elements randomly with 0. How can I do this in MATLAB?
I know how to replace a percentage of elements randomly with zeros, but I must be sure that the element that is replaced randomly, is one of the 1 elements of matrix not zeros.
If you want to change each 1 with a certain probability:
p = 0.1%; % desired probability of change
A_ones = find(A); % linear index of ones in A
A_ones_change = A_ones(rand(size(A_ones))<=p); % entries to be changed
A(A_ones_change) = 0; % apply changes in those entries
If you want to randomly change a fixed fraction of the 1 entries:
f = 0.1; % desired fraction
A_ones = find(A);
n = round(f*length(A_ones));
A_ones_change = randsample(A_ones,n);
A(A_ones_change) = 0;
Note that in this case the resulting fraction may be different to that intended, because of the need to round to an integer number of entries.
#horchler's point is a good one. However, if we keep it simple, then you can just multiple your input matrix to a mask matrix.
>> a1=randint(5,5,[0 1]) #before replacing 1->0
a1 =
1 1 1 0 1
0 1 1 1 0
0 1 0 0 1
0 0 1 0 1
1 0 1 0 1
>> a2=random('unif',0,1,5,5) #Assuming frequency distribution is uniform ('unif')
a2 =
0.7889 0.3200 0.2679 0.8392 0.6299
0.4387 0.9601 0.4399 0.6288 0.3705
0.4983 0.7266 0.9334 0.1338 0.5751
0.2140 0.4120 0.6833 0.2071 0.4514
0.6435 0.7446 0.2126 0.6072 0.0439
>> a1.*(a2>0.1) #And the replacement prob. is 0.1
ans =
1 1 1 0 1
0 1 1 1 0
0 1 0 0 1
0 0 1 0 1
1 0 1 0 0
And other trick can be added to the mask matrix (a2). Such as a different freq. distribution, or a structure (e.g. once a cell is replaced, the adjacent cells become less likely to be replaced and so on.)
Cheers.
The function find is your friend:
indices = find(A);
This will return an array of the indices of 1 elements in your matrix A and you can use your method of replacing a percent of elements with zero on a subset of this array. Then,
A(subsetIndices) = 0;
will replace the remaining indices of A with zero.

Effiicient ways to count a streak of consecutive integers in MATLAB

Say I have a vector containing only logical values, such as
V = [1 0 1 0 1 1 1 1 0 0]
I would like to write a function in MATLAB which returns a 'streak' vector S for V, where S(i) represents the number of consecutive 1s in V up to but not including V(i). For the example above, the streak vector would be
S = [0 1 0 1 0 1 2 3 4 0]
Given that I have to do this for a very large matrix, I would very much appreciate any solution that is vectorized / efficient.
This should do the trick:
S = zeros(size(V));
for i=2:length(V)
if(V(i-1)==1)
S(i) = 1 + S(i-1);
end
end
The complexity is only O(n), which I guess should be good enough.
For your sample input:
V = [1 0 1 0 1 1 1 1 0 0];
S = zeros(size(V));
for i=2:length(V)
if(V(i-1)==1)
S(i) = 1 + S(i-1);
end
end
display(V);
display(S);
The result would be:
V =
1 0 1 0 1 1 1 1 0 0
S =
0 1 0 1 0 1 2 3 4 0
You could also do it completely vectorized with a couple intermediate steps:
V = [1 0 1 0 1 1 1 1 0 0];
Sall = cumsum(V);
stopidx = find(diff(V)==-1)+1;
V2=V;
V2(stopidx) = -Sall(stopidx)+[0 Sall(stopidx(1:end-1))];
S2 = cumsum(V2);
S = [0 S2(1:end-1)];
Afaik the only thing that can take a while is the find call; you can't use logical indexing everywhere and bypass the find call, because you need the absolute indices.
It's outside the box - but have you considered using text functions? Since strings are just vectors for Matlab it should be easy to use them.
Regexp contains some nice functions for finding repeated values.

Inverse lookup table

This is a follow up question from a previous SO question. Now I have a bit which I have spread it into 8 bits. I have use Amro's solution to spread the bit to 8 bits. Now I want an inverse way to convert the 8bits back to the single bit.
I have only managed to implement the inverse using for loop which take alot of time in the application.
Is there a faster way of doing it?
Since you are using the solution I suggested last time, lets say you have a matrix N-by-8 of these 'bits' where each row represent one 8-bit binary number. To convert to decimal in a vectorized way, its as simple as:
» M = randi([0 1], [5 8]) %# 5 random 8-bit numbers
M =
1 0 1 0 1 0 1 1
0 1 1 0 1 1 1 0
1 1 0 1 1 0 1 1
1 0 0 0 0 1 1 0
1 0 0 1 0 1 1 0
» d = bin2dec( num2str(M) )
d =
171
110
219
134
150
An alternative solution:
d = sum( bsxfun(#times, M, power(2,7:-1:0)), 2)