Lecture 22: Introduction to Partition Matrix or Block Matrix | GATE DA Linear Algebra Module-2

Summary

The lecture introduces partition matrices, also called block matrices or partitioned matrices. A matrix is divided into smaller submatrices, or blocks, to make its structure easier to see and sometimes simplify operations. The instructor notes that partition matrices and their properties are explicitly included in the GATE DA syllabus.

1. How to partition a matrix

• A partition is made by drawing horizontal and/or vertical lines through a matrix.
• Each cut must run continuously across the matrix; zigzag partitions are not valid.
• A matrix can be partitioned in many different valid ways.
• Any valid partition is allowed, but some are more useful than others. For example, isolating blocks of zeros or an identity matrix can simplify later calculations.
• Blocks can be labeled (A_{11}, A_{12}, A_{21}, A_{22}), and so on.

The purpose of a useful partition is not merely to divide the matrix, but to reveal structure that can help with computations.

2. Basic block-matrix operations

• Addition: Two partitioned matrices can be added block by block when their partitions align and corresponding blocks have the same dimensions. They must be partitioned in the same way.
• Scalar multiplication: Multiply every block by the scalar.
• Matrix multiplication: Treat the blocks like entries in the usual multiplication pattern, while checking that each block product is dimensionally valid.

For (AB), the key alignment rule is that the column partitions of (A) must match the row partitions of (B). Equivalently, each block product in the usual block-multiplication pattern must have compatible inner dimensions.

A partition can be valid even if it does not allow a particular multiplication. Not every way of partitioning (A) and (B) is suitable for computing (AB).

3. Choosing partitions for multiplication

To ensure a block partition works for (AB):

1. Choose the column cuts in (A).
2. Put corresponding row cuts in (B) at the same positions in the shared dimension.

Row cuts in (A) can be chosen independently, as can column cuts in (B). The shared dimension is what matters: the columns of (A) must align with the rows of (B). Extra row partitions in (A), or extra column partitions in (B), do not by themselves prevent multiplication.

4. Block multiplication and other viewpoints

The lecture connects block multiplication to familiar ways of writing a matrix product, including expressing a product as a sum of column–row outer products. For example, if (A) has columns (a_1,\ldots,a_n) and (v) is a column vector with entries (v_1,\ldots,v_n), then

[ Av = v_1a_1+\cdots+v_na_n. ]

The instructor relates this viewpoint to material previously discussed in the context of SVD.

5. Why block structure can help

In large systems, standard computations may be inefficient. Partitioning can help exploit a matrix’s internal structure, especially when it contains many zeros. If a zero block appears in a product, its contribution may vanish, reducing the work. The lecture therefore emphasizes choosing partitions that make useful structure visible.

6. Determinants of block matrices

For a block matrix of the form

[ \begin{bmatrix} A & B\ C & D \end{bmatrix}, ]

the determinant is especially simple when an off-diagonal block is zero. In the illustrated case, it reduces to the product of the determinants of the diagonal blocks.

For the general case, the instructor presents formulas that require an invertible diagonal block:

• If (A) is invertible,

[ \det\begin{bmatrix} A & B\ C & D \end{bmatrix} =\det(A)\det(D-CA^{-1}B). ]

• If (D) is invertible,

[ \det\begin{bmatrix} A & B\ C & D \end{bmatrix} =\det(D)\det(A-BD^{-1}C). ]

The lecture also discusses special commuting conditions that can simplify these expressions. The main caution is that the simple product of the diagonal-block determinants does not apply automatically to every block matrix.

7. Topics deferred to the next lecture

The instructor postpones block-matrix eigenvalues to the next class, noting that related questions have appeared in GATE CS. The instructor also says that other linear-algebra topics—such as quadratic forms and positive-(semi)definite matrices—will be covered later when needed for machine learning or calculus.

Speakers and sources

• Speaker: An unnamed instructor or lecturer. Ajay and Anand are greeted, but the subtitles do not establish them as speakers.
• Sources mentioned: The GATE DA syllabus and prior GATE CS questions on block-matrix eigenvalues.