$W = \begin{pmatrix}A & B &B &\cdots&B\\ B& A & B &\cdots &B\\ \vdots & \vdots & \vdots & \ddots &\vdots &\\ B& B & B &\cdots &A \end{pmatrix} $ where $A$ and $B$ are symmetric matrices of appropriate order. In fact $B = aJ$, where $J$ is the matrix of all ones. Is there any compact form of $W^{-1}$ ?. Thanks

  • $\begingroup$ Is $A$ known to be invertible? ($B=aJ$ is a rank-one matrix and is not invertible.) $\endgroup$ – Semiclassical May 1 '17 at 12:25
  • $\begingroup$ Yes, $A$ is invertible. $\endgroup$ – Satya Prakash May 1 '17 at 12:28
  • $\begingroup$ @SatyaPrakash use \ddots to get diagonal dots. just so you can apply this in the future!! $\endgroup$ – Alexander Day May 1 '17 at 15:22

To make the answer below easier to read, I will write the block matrix of interest as $\mathbf{W}$; furthermore, I'll denote elements of block matrices using indices $I,J$ (for regular matrices I'll just use $i,j$). For specificity, I will also assume that $\mathbf{W}$ is block $n$-by-$n$ and $A,B$ are $k$-by-$k$. (If you had other symbols in mind for these dimensions, let me know.) Finally, for later convenience I will define $u_n$ to be the column $n$-vector of ones and $J_n$ the $n$-by-$n$ matrix of ones.

Using the above conventions, the block matrix of interest has elements $$(\mathbf{W})_{IJ}=A \delta_{IJ}+B(1-\delta_{IJ})=(A-B)\delta_{IJ}+B$$ where $\delta_{IJ}$ is the Kronecker delta. If we recall the definition of the Kronecker product and consider the form of $B=aJ_k$, these two terms can be compactly represented as $$\mathbf{W}=I_n\otimes (A-B)+J_n\otimes B=I_n\otimes (A-a J_k)+a J_n\otimes J_k.$$ Furthermore, since $J_n=u_nu_n ^T$ we can write $$J_n\otimes J_k=u_nu_n^T\otimes u_ku_k^T = (u_n\otimes u_k)(u_k^T\otimes u_k^T)=\mathbf{u}\mathbf{u}^T$$ where $\mathbf{u}$ is the column $nk$-vector of ones. But this is a rank-one matrix, so we can express $\mathbf{W}^{-1}$ via the Sherman-Morrison formula: \begin{align} \mathbf{W}^{-1} &=\left[I_n\otimes (A-a J)+a \mathbf{u}\mathbf{u}^T\right]^{-1}\\ &=[I_n\otimes (A-a J_k)]^{-1}-\frac{[I_n\otimes (A-a J_k)]^{-1}a\mathbf{u}\mathbf{u}^T[I_n\otimes (A-a J_k)]^{-1}}{1+a\mathbf{u}^T (I_n\otimes (A-a J_k)^{-1})\mathbf{u}}.\\ &=I_n\otimes (A-a J_k)^{-1}-\frac{(I_n\otimes (A-a J_k)^{-1})(J_n\otimes aJ_k)(I_n\otimes (A-a J_k)^{-1})}{1+a(u_n^T\otimes u_k^T) (I_n\otimes (A-a J_k)^{-1})(u_n\otimes u_k)}.\\ &=I_n\otimes (A-a J_k)^{-1}-\frac{J_n\otimes[(A-a J_k)^{-1}J_k(A-a J_k)^{-1}]}{1+na\cdot u_k^T(A-aJ_k)^{-1}u_k}. \end{align} Since $J_k=u_ku_k^T$ is itself a rank-one matrix, we can again appeal to Sherman-Morrison to get $$(A-aJ_k)^{-1}=A^{-1}+A^{-1}\frac{au_ku_k^T}{1-a u_k^TA^{-1}u_k}A^{-1}.$$ This simplifies the product in the denominator to

$$u_k^T(A-aJ_k)^{-1}u_k=u_k^T A^{-1}u_k+(u_k^T A^{-1}u_k)\frac{a}{1-a u_k^TA^{-1}u_k}(u_k^T A^{-1}u_k)=\frac{u_k^T A^{-1}u_k}{1-a u_k^TA^{-1}u_k}.$$ It's possible one could simplify this further but for the time being I'll call it quits here. However, note that $\mathbf{W}^{-1}$ been expressed entirely in terms of $u$ and the matrix inverse $A^{-1}$. (In particular, $u_k^T A^{-1}u_k$ is just the sum of all elements of $A^{-1}$.) So if we already knew $A^{-1}$ then we could compute $\mathbf{W}^{-1}$ directly from the above.

  • $\begingroup$ Note: I've tried to ensure that the signs and such are right, but there's always the chance I missed something. So the reader is advised to check the algebra themself. $\endgroup$ – Semiclassical May 1 '17 at 14:50
  • $\begingroup$ Thanks a lot. It's a great help. $\endgroup$ – Satya Prakash May 1 '17 at 14:52
  • $\begingroup$ Hi, I am not sure about the third equality in $W^{-1}$. Since the product of $I_{n}$ and $uu^{T}$ not defined. They have different sizes. Please let me know, I think I am missing something. $\endgroup$ – Satya Prakash May 1 '17 at 15:22
  • $\begingroup$ You're right. But I think the source of the problem is earlier: In writing $\mathbf{B}=aJ\otimes J$, I proceeded as though each entry were of the same dimension. But if $\mathbf{B}$ is block $k$-by-$k$, then one of those should really be a $n$-by-$n$ matrix of ones (not $J$, which is $k$-by-$k$). So I'll need to rewrite the answer to remove the ambiguity. $\endgroup$ – Semiclassical May 1 '17 at 15:35
  • $\begingroup$ Okay, it should be consistent now. In particular, one has $I_n u_n u_n^T=u_n u_n^T$ as expected for the product with an identity matrix. $\endgroup$ – Semiclassical May 1 '17 at 16:38

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.