An algorithm to detect permutations of Hankel matrices - algorithm

I am trying to write code to detect if a matrix is permutation of a Hankel matrix but I can't think of an efficient solution other than very slow brute force. Here is the spec.
Input: An n by n matrix M whose entries are 1 or 0.
Input format: Space separated rows. One row per line. For example
0 1 1 1
0 1 0 1
0 1 0 0
1 0 1 1
Output: A permutation of the rows and columns of M so that M is a Hankel matrix if that is possible. A Hankel matrix has constant skew-diagonals (positive sloping diagonals).
When I say a permutation, I mean we can apply one permutation to the order of the rows and a possibly different one to the columns.
I would be very grateful for any ideas.

Without Loss of Generality, we will assume that there are fewer 0's than 1's. We can then find the possible diagonals in a Hankel Matrix that could be 0's to give us the appropriate number of 0's in the entire matrix. And, this will give us the possible Hankel matrices. From there, you can count the number of 0's in each column, and compare it to the number of 0's in the columns of the original matrix. Once you have done this, you have a much smaller space in which to perform a brute force search: permuting on columns and rows that have the right number of 0's.
Example: OP's suggested a 4x4 matrix with 7 0's. We need to partition this using the set {4,3,3,2,2,1,1}. So, or partitions would be:
{4,3}
{4,2,1} (2 of these matrices)
{3,3,1}
{3,2,2}
{3,2,1,1} (2 of these matrices)
And this gives us the Hankel Matrices (excluding symmetries)
1 1 0 0 1 1 1 0 0 1 1 0 1 1 0 1
1 0 0 1 1 1 0 1 1 1 0 1 1 0 1 0
0 0 1 1 1 0 1 0 1 0 1 0 0 1 0 1
0 1 1 1 0 1 0 0 0 1 0 1 1 0 1 0
1 0 0 1 0 1 1 1 0 1 0 1
0 0 1 1 1 1 1 0 1 0 1 1
0 1 1 0 1 1 0 0 0 1 1 0
1 1 0 1 1 0 0 0 1 1 0 0
The original matrix had columns with 3, 1, 2, and 1 0's in its four columns. Comparing this to the 7 possible Hankel matrices gives us 2 possibilities
1 1 1 0 0 1 1 1
1 1 0 1 1 1 1 0
1 0 1 0 1 1 0 0
0 1 0 0 1 0 0 0
Now, there are only 4 possible permutations that could map the original matrix to each of these: we have only 1 choice based on the columns with 2 and 3 0's, but 2 choices for the columns with 1 0's, and also 2 choices for the rows with 1 0's. Checking those permutations, we see that the following Hankel matrix is a permutation of the original
0 1 1 1
1 1 1 0
1 1 0 0
1 0 0 0

The one thing which the first answer to this question got right is that permuting the rows and columns doesn't change the row sums or column sums.
Another easy observation is that in a Hankel matrix, the difference in row sum between two consecutive rows is -1, 0, or 1, and each case gives us a constraint on the rows. If the difference is 0 then the entering variable is equal to the exiting variable; otherwise we know which is 0 and which is 1.
0 1 1 1
0 1 0 1
0 1 0 0
1 0 1 1
has row sums 3, 2, 1, 3. The orders which respect the difference requirement are 1 2 3 3 and 3 3 2 1, and wlog we can discard reversals because reversing the row and column permutations just rotates the matrix by 180 degrees. Therefore we reduce to considering four permuted matrices (two possible orderings of the 3s in the row sums, and two in the column sums):
0 0 1 0 0 0 1 0 0 0 0 1 0 0 0 1
0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1
0 1 1 1 1 1 0 1 0 1 1 1 1 1 1 0
1 1 0 1 0 1 1 1 1 1 1 0 0 1 1 1
We could actually have taken the analysis further by observing that by forcing the initial rows to have sums 1 and 2 we constrain the order of the columns with sum 3, since
0 0 1 0
0 0 1 1
is not a valid initial two rows of a Hankel matrix. Whether or not this kind of reasoning is easy to implement depends on your programming paradigm.
Note that in the worst case this kind of reasoning still doesn't leave a polynomial number of cases to brute force through.

Here are some ideas.
1)
Row and Column permutations preserve the row and column sums:
1 0 1 0 - 2
0 0 0 1 - 1 row sums
1 0 0 0 - 1
1 1 1 0 - 3
| | | |
3 1 2 1
column sums
Whichever way you permute the rows, the row sums will still be {2, 1, 1, 3} in some permutation; the column sums will be unchanged. And vice versa. Hankel matrices and their permutations will always have the same set of row sums as column sums. This gives you a quick test to rule out a set of non-viable matrices.
2)
I posit that Hankel matrices can always be permuted in such a way that their row and column sums are in ascending order, and the result is still a Hankel matrix:
0 1 1 0 - 2 0 0 0 1 - 1
1 1 0 0 - 2 0 0 1 1 - 2
1 0 1 1 - 3 --> 0 1 1 0 - 2
0 0 1 0 - 1 1 1 0 1 - 3
| | | | | | | |
2 2 3 1 1 2 2 3
Therefore if a matrix can be permuted into a Hankel matrix, then it can also be permuted into a Hankel matrix of ascending row and column sum. That is, we can reduce the number of permutations needed to test by only testing permutations where the row and column sums are in ascending order.
3)
I posit further that for any Hankel matrix where two or more rows have the same sum, every permutation of columns has a matching permutation of rows that also produces a Hankel matrix. That is, if a Hankel matrix exists for one permutation of columns, then it exists for every permutation of columns - since we can simply apply that same permutation to the corresponding rows and achieve a symmetrical result.
The upshot is that we only need to test permutations of rows or columns, not rows and columns.
Applied to the original example:
1 0 1 0 - 2 0 0 0 1 0 1 0 0 - 1 0 0 0 1
0 0 0 1 - 1 1 0 0 0 0 0 0 1 - 1 0 1 0 0
1 0 0 0 - 1 --> 1 0 1 0 --> 0 0 1 1 - 2 --> 0 0 1 1 = Hankel!
1 1 1 0 - 3 1 1 1 0 1 0 1 1 - 3 1 0 1 1
| | | |
3 1 2 1 permute rows into| ditto | try swapping
ascending order | for columns | top 2 rows
4)
I posit, finally, that every Hankel matrix where there are multiple rows and columns with the same sum can be permuted into another Hankel matrix with the property that those rows and columns are in increasing order when read as binary numbers - reading left-to-right for rows and top-to-bottom for columns. That is:
0 1 1 0 0 1 0 1 0 0 1 1
1 0 0 1 0 1 1 0 0 1 0 1 New
1 0 1 0 --> 1 0 0 1 --> 1 0 1 0 Hankel
0 1 0 1 1 0 1 0 1 1 0 0
Original rows columns
Hankel ascending ascending
If this is true (and I'm still undecided), then we only ever need to create and test one permutation of any given input matrix. That permutation puts both the rows and columns in order of ascending sum, and in the case of equal sums, orders them by their binary number interpretations. If this resultant matrix is not Hankel, then there is no permutation that will make it Hankel.
Hope that gets you on the way to an algorithm!
Addendum: Counterexamples?
Trying #orlp's example:
0 0 1 0 0 0 1 0 0 0 0 1
0 1 0 1 0 1 0 1 0 1 1 0
1 0 1 1 --> 0 1 1 1 --> 0 1 1 1
0 1 1 1 1 0 1 1 1 0 1 1
(A) (B) (C)
A: Original Hankel. Row sums are 1, 2, 3, 3; Rows 3 and 4 are not in binary order.
B: Swap rows 3 and 4. Columns 3 and 4 are not in binary order.
C: Swap columns 3 and 4. Result is Hankel and satisfies all the properties.
Trying #Degustaf's example:
1 1 0 1 0 1 0 0 0 0 1 0
1 0 1 0 1 0 0 1 0 1 0 1
0 1 0 0 --> 1 0 1 0 --> 1 0 0 1
1 0 0 1 1 1 0 1 0 1 1 1
(A) (B) (C)
A: Original Hankel matrix. Row sums are 3, 2, 1, 2.
B: Rearrange so that the row sums are 1, 2, 2, 3, and the rows of sum 2 are in ascending binary order (i.e. 1001, 1010)
C: Rearrange column sums to 1, 2, 2, 3, with the two columns of sum 2 in order (0101, 1001). Result is Hankel and satisfies all the properties. Note also that the permutation on the columns matches the permutation on the rows: the new column order from the old one is {3, 4, 2, 1}, the same operation to get from A to B.
Note: I suggest the binary order (#4) only for tiebreak situations on the row or column sum, not as a replacement for the sort in (#2).

Related

Finding the largest submatrix contain only one zero and 1's in a matrix

I need to find a submatrix that only contain a 0,but others are 1.
Here's an example:
5*4 matrix
1 0 1 1 1
1 1 1 0 0
1 1 1 0 1
0 1 0 1 1
And the answer will be
3*3
1 0 1
1 1 1
1 1 1
I tried to use other similar questions' methods.But so far I can't figure out how to calculate it without using brutal solution.

Find the largest chessboard surface in 2D Matrix

I'm having a bit of difficulty by solving an exercise, hope I can get some help from you.
Given a 2D Array with N rows and M columns, and the array has only elements with the value 0 and 1.
Find the largest surface that is similar to a chessboard(where 0 are the white squares and 1 the black squares) and print the size of the surface(number of squares).
Constraints:
2<=N<=1000
2<=M<=1000
Example:
N=4, M=5
0 1 1 0 1
1 0 1 0 1
0 0 1 1 0
1 1 0 1 1
Output: Number of squares=5 (row 2-from column 1 to column 5)
Example:
N=3, M=4
0 0 1 0
1 1 0 0
1 0 1 0
Output: Number of squares=6 (from row 1 to row 3- from column 2 to column 3)
Flip every second cell in a checkerboard pattern
Find the largest rectangle containing only 0s or only 1s.
See Find largest rectangle containing only zeros in an N×N binary matrix for help on the 2nd part.
For your second example:
0 0 1 0
1 1 0 0
1 0 1 0
flipping every 2nd cell produces:
0 1 1 1
0 1 1 0
1 1 1 1
And you can see the rectangle of 1s that you're looking for.

Sorting rows and columns of adjacency matrix to reveal cliques

I'm looking for a reordering technique to group connected components of an adjacency matrix together.
For example, I've made an illustration with two groups, blue and green. Initially the '1's entries are distributed across the rows and columns of the matrix. By reordering the rows and columns, all '1''s can be located in two contiguous sections of the matrix, revealing the blue and green components more clearly.
I can't remember what this reordering technique is called. I've searched for many combinations of adjacency matrix, clique, sorting, and reordering.
The closest hits I've found are
symrcm moves the elements closer to the diagonal, but does not make groups.
Is there a way to reorder the rows and columns of matrix to create a dense corner, in R? which focuses on removing completely empty rows and columns
Please either provide the common name for this technique so that I can google more effectively, or point me in the direction of a Matlab function.
I don't know whether there is a better alternative which should give you direct results, but here is one approach which may serve your purpose.
Your input:
>> A
A =
0 1 1 0 1
1 0 0 1 0
0 1 1 0 1
1 0 0 1 0
0 1 1 0 1
Method 1
Taking first row and first column as Column-Mask(maskCol) and
Row-Mask(maskRow) respectively.
Get the mask of which values contains ones in both first row, and first column
maskRow = A(:,1)==1;
maskCol = A(1,:)~=1;
Rearrange the Rows (according to the Row-mask)
out = [A(maskRow,:);A(~maskRow,:)];
Gives something like this:
out =
1 0 0 1 0
1 0 0 1 0
0 1 1 0 1
0 1 1 0 1
0 1 1 0 1
Rearrange columns (according to the column-mask)
out = [out(:,maskCol),out(:,~maskCol)]
Gives the desired results:
out =
1 1 0 0 0
1 1 0 0 0
0 0 1 1 1
0 0 1 1 1
0 0 1 1 1
Just a check whether the indices are where they are supposed to be or if you want the corresponding re-arranged indices ;)
Before Re-arranging:
idx = reshape(1:25,5,[])
idx =
1 6 11 16 21
2 7 12 17 22
3 8 13 18 23
4 9 14 19 24
5 10 15 20 25
After re-arranging (same process we did before)
outidx = [idx(maskRow,:);idx(~maskRow,:)];
outidx = [outidx(:,maskCol),outidx(:,~maskCol)]
Output:
outidx =
2 17 7 12 22
4 19 9 14 24
1 16 6 11 21
3 18 8 13 23
5 20 10 15 25
Method 2
For Generic case, if you don't know the matrix beforehand, here is the procedure to find the maskRow and maskCol
Logic used:
Take first row. Consider it as column mask (maskCol).
For 2nd row to last row, the following process are repeated.
Compare the current row with maskCol.
If any one value matches with the maskCol, then find the element
wise logical OR and update it as new maskCol
Repeat this process till the last row.
Same process for finding maskRow while the column are used for
iterations instead.
Code:
%// If you have a square matrix, you can combine both these loops into a single loop.
maskCol = A(1,:);
for ii = 2:size(A,1)
if sum(A(ii,:) & maskCol)>0
maskCol = maskCol | A(ii,:);
end
end
maskCol = ~maskCol;
maskRow = A(:,1);
for ii = 2:size(A,2)
if sum(A(:,ii) & maskRow)>0
maskRow = maskRow | A(:,ii);
end
end
Here is an example to try that:
%// Here I removed some 'ones' from first, last rows and columns.
%// Compare it with the original example.
A = [0 0 1 0 1
0 0 0 1 0
0 1 1 0 0
1 0 0 1 0
0 1 0 0 1];
Then, repeat the procedure you followed before:
out = [A(maskRow,:);A(~maskRow,:)]; %// same code used
out = [out(:,maskCol),out(:,~maskCol)]; %// same code used
Here is the result:
>> out
out =
0 1 0 0 0
1 1 0 0 0
0 0 0 1 1
0 0 1 1 0
0 0 1 0 1
Note: This approach may work for most of the cases but still may fail for some rare cases.
Here, is an example:
%// this works well.
A = [0 0 1 0 1 0
1 0 0 1 0 0
0 1 0 0 0 1
1 0 0 1 0 0
0 0 1 0 1 0
0 1 0 0 1 1];
%// This may not
%// Second col, last row changed to zero from one
A = [0 0 1 0 1 0
1 0 0 1 0 0
0 1 0 0 0 1
1 0 0 1 0 0
0 0 1 0 1 0
0 0 0 0 1 1];
Why does it fail?
As we loop through each row (to find the column mask), for eg, when we move to 3rd row, none of the cols match the first row (current maskCol). So the only information carried by 3rd row (2nd element) is lost.
This may be the rare case because some other row might still contain the same information. See the first example. There also none of the elements of third row matches with 1st row but since the last row has the same information (1 at the 2nd element), it gave correct results. Only in rare cases, similar to this might happen. Still it is good to know this disadvantage.
Method 3
This one is Brute-force Alternative. Could be applied if you think the previous case might fail. Here, we use while loop to run the previous code (finding row and col mask) number of times with updated maskCol, so that it finds the correct mask.
Procedure:
maskCol = A(1,:);
count = 1;
while(count<3)
for ii = 2:size(A,1)
if sum(A(ii,:) & maskCol)>0
maskCol = maskCol | A(ii,:);
end
end
count = count+1;
end
Previous example is taken (where the previous method fails) and is run with and without while-loop
Without Brute force:
>> out
out =
1 0 1 0 0 0
1 0 1 0 0 0
0 0 0 1 1 0
0 1 0 0 0 1
0 0 0 1 1 0
0 0 0 0 1 1
With Brute-Forcing while loop:
>> out
out =
1 1 0 0 0 0
1 1 0 0 0 0
0 0 0 1 1 0
0 0 1 0 0 1
0 0 0 1 1 0
0 0 0 0 1 1
The number of iterations required to get the correct results may vary. But it is safe to have a good number.
Good Luck!

How to create a symmetric matrix of 1's and 0's with constant row and column sum

I'm trying to find an elegant algorithm for creating an N x N matrix of 1's and 0's, under the restrictions:
each row and each column must sum to Q (to be picked freely)
the diagonal must be 0's
the matrix must be symmetrical.
It is not strictly necessary for the matrix to be random (both random and non-random solutions are interesting, however), so for Q even, simply making each row a circular shift of the vector
[0 1 1 0 ... 0 0 0 ... 0 1 1] (for Q=4)
is a valid solution.
However, how to do this for Q odd? Or how to do it for Q even, but in a random fashion?
For those curious, I'm trying to test some phenomena on abstract networks.
I apologize if this has already been answered before, but none of the questions I could find had the symmetric restriction, which seems to make it much more complicated. I don't have a proof that such a matrix always exists, but I do assume so.
The object that you're trying to construct is known more canonically as an undirected d-regular graph (where d = Q). By the handshaking theorem, N and Q cannot both be odd. If Q is even, then connect vertex v to v + k modulo N for k in {-Q/2, -Q/2 + 1, ..., -1, 1, ..., Q/2 - 1, Q/2}. If Q is odd, then N is even. Construct a (Q - 1)-regular graph as before and then add connections from v to v + N/2 modulo N.
If you want randomness, there's a Markov chain whose limiting distribution is uniform on d-regular graphs. You start with any d-regular graph. Repeatedly pick vertices v, w, x, y at random. Whenever the induced subgraph looks like
v----w
x----y ,
flip it to
v w
| |
x y .
You can perhaps always follow your circular shift algorithm, when possible.
The only condition you need to follow while using the circular shift algorithm is to maintain the symmetric nature in the first row.
i.e. keeping Q 1's in the first row so that Q[0,1] to Q[0,N-1] {Assuming 0 indexed rows and cols, Q[0,0] is 0.} is symmetric, a simple example being 110010011.
Hence, N = 10, Q = 5, you can get many possible arrangements such as:
0 1 0 0 1 1 1 0 0 1
1 0 1 0 0 1 1 1 0 0
0 1 0 1 0 0 1 1 1 0
0 0 1 0 1 0 0 1 1 1
1 0 0 1 0 1 0 0 1 1
1 1 0 0 1 0 1 0 0 1
1 1 1 0 0 1 0 1 0 0
0 1 1 1 0 0 1 0 1 0
0 0 1 1 1 0 0 1 0 1
1 0 0 1 1 1 0 0 1 0
or
0 1 1 0 0 1 0 0 1 1
1 0 1 1 0 0 1 0 0 1
1 1 0 1 1 0 0 1 0 0
0 1 1 0 1 1 0 0 1 0
0 0 1 1 0 1 1 0 0 1
1 0 0 1 1 0 1 1 0 0
0 1 0 0 1 1 0 1 1 0
0 0 1 0 0 1 1 0 1 1
1 0 0 1 0 0 1 1 0 1
1 1 0 0 1 0 0 1 1 0
But as you can see for odd N(that means even N-1) and odd Q there can't be any such symmetric distribution.. Hope it helped.

How can I find a solution of binary matrix equation AX = B?

Given an m*n binary matrix A, m*p binary matrix B, where n > m what is an efficient algorithm to compute X such that AX=B?
For example:
A =
1 1 0 0 1 1 0 1 0 0
1 1 0 0 1 0 1 0 0 1
0 1 1 0 1 0 1 0 1 0
1 1 1 1 1 0 0 1 1 0
0 1 1 0 1 0 1 1 1 0
B =
0 1 0 1 1 0 1 1 0 1 0 0 1 0
0 0 1 0 1 1 0 0 0 1 0 1 0 0
0 1 1 0 0 0 1 1 0 0 1 1 0 0
0 0 1 1 1 1 0 0 0 1 1 0 0 0
1 0 0 1 0 0 1 0 1 0 0 1 1 0
Note, when I say binary matrix I mean matrix defined over the field Z_2, that is, where all arithmetic is mod 2.
If it is of any interest, this is a problem I am facing in generating suitable matrices for a random error correction code.
You can do it with row reduction: Place B to the right of A, and then swap rows (in the whole thing) to get a 1 in row 0, col 0; then xor that row to any other row that has a '1' in column 0, so you have only a single 1 in column 0. Then move to the next column; if [1,1] is zero then swap row 1 with a later row that has a 1 there, then xor rows to make it the only 1 in the column. Assuming 'A' is a square matrix and a solution exists, then you eventually have converted A to unity, and B is replaced with the solution to Ax=B.
If n > m, you have a system with more unknowns than equations, so you can solve for some of the unknowns, and set the others to zero. During the row reduction, if there are no values in a column which have a '1' to use (below the rows already reduced) you can skip that column and make the corresponding unknown zero (you can do this at most n-m times).

Resources