Inverting Block-Diagonal Matrices

Reading a Block-Diagonal Matrix

A large matrix sometimes acts on separate groups of coordinates without mixing them. In that case, it can be written in blocks: square matrices appear on the diagonal, and zero matrices appear everywhere else. This structure can simplify inversion. Rather than row-reducing the entire matrix, we can invert each diagonal block separately and place those inverses in the corresponding positions. The key is to check the block sizes and verify the proposed inverse by multiplication.

Throughout, the matrices have real entries. For a matrix \(\mathbf{A}\in\mathbb{R}^{n\times n}\), an inverse is a matrix \(\mathbf{B}\in\mathbb{R}^{n\times n}\) satisfying both \(\mathbf{A}\mathbf{B}=\mathbf{I}_n\) and \(\mathbf{B}\mathbf{A}=\mathbf{I}_n\). In that case, \(\mathbf{A}\) is invertible and \(\mathbf{B}=\mathbf{A}^{-1}\). A block is itself a matrix, and its size determines which other blocks can be multiplied with it.

For example, with two blocks the structure is

where each zero block has the dimensions required by its position. Block multiplication follows the usual row-by-column rule, treating each block as a matrix. In particular, when the products are defined, multiplying two block-diagonal matrices with the same block sizes multiplies corresponding diagonal blocks and leaves zero blocks off the diagonal:

The off-diagonal products are zero because, for instance, the upper-right block is \(\mathbf{A}_1\mathbf{O}+\mathbf{O}\mathbf{B}_2=\mathbf{O}\). This simple multiplication rule is the reason the inverse formula works.

The Inverse Formula and Its Proof

Proof. First suppose every diagonal block \(\mathbf{A}_i\) is invertible. Define \(\mathbf{D}=\operatorname{diag}(\mathbf{A}_1^{-1},\dots,\mathbf{A}_k^{-1})\). In the product \(\mathbf{A}\mathbf{D}\), the \(i\)-th diagonal block is \(\mathbf{A}_i\mathbf{A}_i^{-1}=\mathbf{I}_{n_i}\), and every off-diagonal block is zero by block multiplication. Thus \(\mathbf{A}\mathbf{D}=\mathbf{I}_{n_1+\cdots+n_k}\). In the reverse order, the \(i\)-th diagonal block is \(\mathbf{A}_i^{-1}\mathbf{A}_i=\mathbf{I}_{n_i}\), and the off-diagonal blocks are again zero. Thus \(\mathbf{D}\mathbf{A}=\mathbf{I}_{n_1+\cdots+n_k}\), so \(\mathbf{D}\) is the inverse of \(\mathbf{A}\). Conversely, suppose \(\mathbf{A}\) is invertible and write \(\mathbf{X}=\mathbf{A}^{-1}\) in blocks \(\mathbf{X}_{ij}\), where \(\mathbf{X}_{ij}\) has \(n_i\) rows and \(n_j\) columns. From \(\mathbf{A}\mathbf{X}=\mathbf{I}\), the \(i\)-th diagonal block gives \(\mathbf{A}_i\mathbf{X}_{ii}=\mathbf{I}_{n_i}\). From \(\mathbf{X}\mathbf{A}=\mathbf{I}\), the same diagonal position gives \(\mathbf{X}_{ii}\mathbf{A}_i=\mathbf{I}_{n_i}\). Therefore \(\mathbf{X}_{ii}\) is a two-sided inverse of \(\mathbf{A}_i\), so every diagonal block is invertible. This proves both directions and the stated formula. \(\square\)

The two product checks in the proof matter: an inverse is defined by both orders. The theorem also gives a useful exact condition. If even one diagonal block is not invertible, the whole block-diagonal matrix cannot be invertible.

Worked Inversions

Common Mistake: Keep the Blocks in Their Places

A frequent error is to invert one diagonal block but leave another unchanged. Using the first example, a mistaken candidate is \(\operatorname{diag}(\mathbf{B}^{-1},\mathbf{C})\), rather than \(\operatorname{diag}(\mathbf{B}^{-1},\mathbf{C}^{-1})\). Multiplying the second block by itself gives

So the mistaken candidate fails the inverse check. The correction is to invert each diagonal block, not to copy the original blocks into the proposed inverse. Then multiply in both orders: each corresponding block product must be an identity matrix, and every off-diagonal block must be zero.

Block structure also gives a practical advantage. If the matrix has many blocks, each inverse can be computed and checked separately; no multiplication between different diagonal blocks is needed. This is useful when the blocks are small even if the full matrix is large. The method does not bypass the need to establish that each block is invertible, and it does not allow blocks of incompatible sizes to be placed or multiplied arbitrarily. First identify the block partition, then find the inverse of each square block, then assemble and verify.