Eliminate a block, not an entry
Elimination can solve a system in groups of variables. Its remaining block records how the eliminated variables affect the others. We derive that block and use it for determinants and real positive-definite matrices.
Let be a field and let . Consider compatible blocks
Definition 6.1.1
(Schur complement)
Assume is invertible. The Schur complement of in is . No commutativity of the blocks is assumed: the order of multiplication is essential.
Theorem 6.1.2
(Block factorization)
With invertible,
In particular, is invertible exactly when is invertible.
Proof
(Multiply out and use block-triangular determinants)
Multiplying the two block matrices gives the original four blocks. The lower factor has determinant one and the upper block-triangular factor has determinant . Since over a field, exactly when .
For a right-hand side , the first block equation gives . Substitution into the second yields
If is invertible, this determines uniquely and then . In computation one generally solves systems with rather than explicitly forming an inverse; the algebraic notation records the map being applied.
Positive definiteness
Now work over and suppose is symmetric, so and are symmetric. Assume is positive definite. Then is invertible and is symmetric.
Theorem 6.1.3
(A positive-definite block test)
Under these real symmetric hypotheses, is positive definite if and only if is positive definite.
Proof
(Completing the square in block form)
For and , put . Expanding gives
If is positive definite, the right side is positive for every nonzero : if , then ; otherwise the second term is positive. Conversely, if is positive definite, choose for any nonzero . Then and .
Proof X-ray
Where the two directions use the identity
The one non-obvious move is the substitution : it is the block analogue of completing the square, and it is chosen so that the cross term is absorbed into , leaving a quadratic form in alone with matrix . Positive definiteness of is used twice: to make exist, and to make with equality only for . For the direction from to , that equality case is exactly what is needed when , because then and the first term already forces positivity. For the direction from to , the proof does not estimate anything: it picks the particular that switches the first term off, so evaluated at equals . Symmetry of gives , which is what allows the cross terms and to be combined into in the expansion.
Worked examples
Worked example 6.1.4
(A three-variable system)
Take
Then and . Both and are positive definite, hence so is . For right-hand side , , the reduced equation gives , so and . Substitution verifies all three original equations.
Worked example 6.1.5
(A singular Schur complement)
Take with . Then , so , which agrees with . The factorization still holds, but is not invertible, so is not invertible and the reduced equation is solvable only when .
Worked example 6.1.6
(Testing a tridiagonal matrix by blocks)
Let
The matrix is positive definite: its leading entry is and . Its inverse is , so and . By the block test, is positive definite, and , which agrees with direct cofactor expansion.
Counterexample mode
Positive diagonal blocks do not make a matrix positive definite
A tempting claim is that a symmetric block matrix with positive definite diagonal blocks and is positive definite. Take , and , so . Here . Concretely, for one computes , so is not positive definite. The correct test constrains , which contains the cross term through , not .
Common mistake
(A singular pivot block does not make the matrix singular)
The factorization needs invertible, but may be invertible when is not. For the upper-left block is singular while . In such a case the formula with simply does not apply; another pivot or ordering of the blocks must be used.
Exercises
Work over throughout these exercises.
Quick check
(For , take and find and .)
Use .
Solution · Answer
, so .
Quick check
(Using the same matrix, solve by block elimination.)
Solve for first.
Solution · Answer
, so and . Check: .
Quick check
(For the real symmetric matrix , apply the positive-definite block test.)
Check first, then .
Solution · Answer
and , so the matrix is positive definite.
Quick check
(Why does invertibility of not imply invertibility of its upper-left block?)
Think of a permutation matrix.
Solution · Answer
The matrix is invertible but has a zero upper-left entry. The formula requires an invertible pivot block independently of invertibility of the full matrix.