About four months of gap (missing GFG), a new post. Given an M x N matrix, transpose the matrix without auxiliary memory.It is easy to transpose matrix using an auxiliary array. If the matrix is symmetric in size, we can transpose the matrix inplace by mirroring the 2D array across it’s diagonal (try yourself). How to transpose an arbitrary size matrix inplace? See the following matrix,
a b c a d g j d e f ==> b e h k g h i c f i l j k l
As per 2D numbering in C/C++, corresponding location mapping looks like,
Org element New 0 a 0 1 b 4 2 c 8 3 d 1 4 e 5 5 f 9 6 g 2 7 h 6 8 i 10 9 j 3 10 k 7 11 l 11
Note that the first and last elements stay in their original location. We can easily see the transformation forms few permutation cycles.
- 1->4->5->9->3->1 – Total 5 elements form the cycle
- 2->8->10->7->6->2 – Another 5 elements form the cycle
- 0 – Self cycle
- 11 – Self cycle
From the above example, we can easily devise an algorithm to move the elements along these cycles. How can we generate permutation cycles? Number of elements in both the matrices are constant, given by N = R * C, where R is row count and C is column count. An element at location ol (old location in R x C matrix), moved to nl (new location in C x R matrix). We need to establish relation between ol, nl, R and C. Assume ol = A[or][oc]. In C/C++ we can calculate the element address as,
ol = or x C + oc (ignore base reference for simplicity)
It is to be moved to new location nl in the transposed matrix, say nl = A[nr][nc], or in C/C++ terms
nl = nr x R + nc (R - column count, C is row count as the matrix is transposed)
Observe, nr = oc and nc = or, so replacing these for nl,
nl = oc x R + or -----> [eq 1]
after solving for relation between ol and nl, we get
ol = or x C + oc ol x R = or x C x R + oc x R = or x N + oc x R (from the fact R * C = N) = or x N + (nl - or) --- from [eq 1] = or x (N-1) + nl
nl = ol x R - or x (N-1)
Note that the values of nl and ol never go beyond N-1, so considering modulo division on both the sides by (N-1), we get the following based on properties of congruence,
nl mod (N-1) = (ol x R - or x (N-1)) mod (N-1) = (ol x R) mod (N-1) - or x (N-1) mod(N-1) = ol x R mod (N-1), since second term evaluates to zero nl = (ol x R) mod (N-1), since nl is always less than N-1
A curious reader might have observed the significance of above relation. Every location is scaled by a factor of R (row size). It is obvious from the matrix that every location is displaced by scaled factor of R. The actual multiplier depends on congruence class of (N-1), i.e. the multiplier can be both -ve and +ve value of the congruent class.Hence every location transformation is simple modulo division. These modulo divisions form cyclic permutations. We need some book keeping information to keep track of already moved elements. Here is code for inplace matrix transformation,
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 1 7 13 19 25 2 8 14 20 26 3 9 15 21 27 4 10 16 22 28 5 11 17 23 29 6 12 18 24 30
Extension: 17 – March – 2013 Some readers identified similarity between the matrix transpose and string transformation. Without much theory I am presenting the problem and solution. In given array of elements like [a1b2c3d4e5f6g7h8i9j1k2l3m4]. Convert it to [abcdefghijklm1234567891234]. The program should run inplace. What we need is an inplace transpose. Given below is code.
Update 09-July-2016: Notes on space complexity and storage order.
After long time, it happened to review this post. Some readers pointed valid questions on how can it be in-place (?) when we are using bitset as marker (hash in code). Apologies for incorrect perception by looking at the article heading or content. While preparing the initial content, I was thinking of naive implementation using auxiliary space of atleast O(MN) needed to transpose rectangualr matrix. The program presented above is using constant space as bitset size is fixed at compile time. However, to support arbitrary size of matrices we need bitset size atleast O(MN) size. One can use a HashMap (amortized O(1) complexity) for marking finished locations, yet HashMap’s worst case complexity can be O(N) or O(log N) based on implementation. HashMap space cost also increases based on items inserted. Please note that in-place was used w.r.t. matrix space.
Also, it was assumed that the matrix will be stored in row major ordering (contigueous locations in memory). The reader can derive the formulae, if the matrix is represented in column major order by the programming language (e.g. Fortran/Julia).
Thanks to the readers who pointed these two gaps.
The post is incomplete without mentioning two links.
1. Aashish covered good theory behind cycle leader algorithm. See his post on string transformation.
— Venki. Please write comments if you find anything incorrect, or you want to share more information about the topic discussed above.
- C# | Check if the ArrayList has a fixed size
- C# | Check if ListDictionary has a fixed size
- C# | Check if HybridDictionary has fixed size
- C Program to find transpose of a matrix
- Program to find transpose of a matrix
- C# | Check if a SortedList object has a fixed size
- Transpose a matrix in Single line in Python
- Inplace rotate square matrix by 90 degrees | Set 1
- Java Program to find transpose of a matrix
- Python Program to find transpose of a matrix
- Count frequency of k in a matrix of size n where matrix(i, j) = i+j
- Sorted subsequence of size 3 in linear time using constant space
- Print n x n spiral matrix using O(1) extra space
- Rotate a matrix by 90 degree without using any extra space | Set 2
- Printing Matrix Chain Multiplication (A Space Optimized Solution)