Prerequisites are symmetric matrices, determinants, block multiplication, orthogonal diagonalization and the real spectral theorem. This section treats symmetric matrices that may be indefinite.
Why allow negative pivots?
Definition 7.2.1
(Definite and indefinite symmetric matrices)
A real symmetric matrix defines the quadratic form . The matrix is positive definite if for every nonzero , negative definite if for every nonzero , and indefinite if takes both positive and negative values.
The real spectral theorem writes , so the signs of the eigenvalues determine these directions. But one can often read the same sign information from a triangular factorization, without finding eigenvalues.
Definition 7.2.2
(Leading minors and pivots)
Let be the determinant of the upper-left block of , the -th leading principal minor, and set . When every , the numbers are the pivots of for elimination without row exchanges.
Theorem 7.2.3
( factorization without pivoting)
Let be a real symmetric matrix with for . Then
The factorization is unique.
The can be negative; their nonzero assumption concerns the elimination route, not positive definiteness. For positive definite matrices all are positive, but the formula also handles many indefinite matrices.
Constructing the factor without assuming positivity
Proof
(Induction through the Schur complement)
We prove existence and uniqueness by induction. The case is immediate. For , partition
Here , so the division is legitimate. Multiplication gives
For each leading block, the triangular factors have determinant one. Therefore its determinant is times the determinant of the leading block of . Those determinants of are , all nonzero. By induction , with unit lower triangular and diagonal. Substituting yields
Uniqueness follows because the first pivot and first lower column are forced, as is ; induction then forces the remaining factors. For each , the leading block of is , so and the displayed ratios follow. No square root of a negative pivot is taken.
Proof X-ray
The move that makes the induction work
The factorization is peeled off one row and column at a time. The hypothesis is used exactly where a division occurs: by at the first step, and, inductively, by the leading minors of . The non-obvious step is the minor identity: the two outer triangular factors in the three-factor product have determinant one, so . This is what lets the induction hypothesis apply to with no new assumption. Uniqueness holds because , and are forced by , and the argument closes because every step shrinks the matrix by one. If the very first division fails, and shows that then no factorization of this form need exist.
What the signs mean
Definition 7.2.4
(Inertia)
The inertia of a nonsingular real symmetric matrix is the pair , where and are the numbers of positive and negative eigenvalues, counted with multiplicity. Nonsingularity gives .
The eigenvalue form and the pivot form are both congruences: and . They produce different diagonal entries, because is orthogonal while is merely triangular. What every such diagonal form shares is the count of positive and negative entries, that is, the inertia. The individual diagonal entries are not invariants of the quadratic form.
Theorem 7.2.5
(Pivot signs and inertia)
Let be as in the factorization theorem, with , and put .
- . Hence is positive definite exactly when every , negative definite exactly when every , and indefinite exactly when both signs occur.
- The number of positive equals and the number of negative equals .
Proof
(Rewriting the form in the coordinates y)
Since is invertible, ranges over all of , and
Thus is positive definite exactly when every , negative definite exactly when every , and indefinite when both signs occur. This is a congruence transformation, not a similarity: the numbers need not equal eigenvalues.
Proof
(Counting dimensions of definite subspaces)
Their counts nevertheless match the counts of positive and negative eigenvalues. For example, let be the number of positive and the number of positive eigenvalues. The -dimensional subspace obtained by setting all negative- coordinates of to zero has positive quadratic form on every nonzero vector. If , it intersects the spectral subspace for nonpositive eigenvalues nontrivially, because their dimensions sum to more than . That vector would have both positive and nonpositive quadratic form, a contradiction. Reversing the roles of the two diagonal descriptions rules out . Hence ; similarly the negative counts agree. Under our nonzero-minor hypothesis no eigenvalue or pivot is zero. This argument explains the inertia count without claiming equality of individual pivots and eigenvalues.
Proof X-ray
How the dimension count forces equal signs
The two diagonalizations describe the same form , but in different coordinates: is not orthonormal, whereas the eigenbasis is. So the numbers and differ, and the argument never compares them. It compares dimensions. The subspace where the negative- coordinates vanish has dimension and a positive form; the spectral subspace of nonpositive eigenvalues has dimension and a nonpositive form. Two subspaces of whose dimensions add to more than meet in a nonzero vector, and that vector would give a form value both positive and nonpositive. Hence , and the symmetric argument gives . Nonsingularity is needed so that no or is zero; otherwise "positive" and "nonpositive" would not split the coordinates cleanly.
Worked examples
Worked example 7.2.6
(Two positive pivots and one negative)
Factor
as , determine its inertia, and exhibit one positive and one negative direction.
Solution. The leading determinants are , , . The pivots are , , , and direct multiplication verifies
There are two positive and one negative pivots, so there are two positive and one negative eigenvalues. For instance , while . Thus is indefinite. The pivots are not asserted to be its eigenvalues.
Worked example 7.2.7
(Pivots differ from eigenvalues but share signs)
Let . Then and , so and , with and . Multiplication confirms . One pivot is positive and one negative, so is indefinite with inertia . The eigenvalues are and (trace , determinant ): different numbers, with one of each sign. For the vector we have and , while directly .
Worked example 7.2.8
(A zero pivot repaired by a symmetric permutation)
Let . Then but , so the scalar-pivot formula stops after one step, although . Exchange the second and third coordinates, that is, replace by where swaps the indices and :
Now , and , so the pivots are , and
The third row of is , as required. The forms and take the same values (replace by ), so has two positive and one negative eigenvalue. This agrees with the eigenvalues and of .
When a pivot vanishes
Counterexample mode
An invertible symmetric matrix with no scalar-pivot factorization
Claim: every invertible real symmetric matrix can be written as with unit lower triangular and diagonal. Take , for which , so is invertible. If with and , then
The entry forces , and then the entry is , a contradiction. So the hypothesis cannot be dropped. The matrix still has inertia : the orthogonal matrix gives . A symmetric permutation does not help, because swapping both coordinates leaves unchanged; practical symmetric elimination continues with a block pivot. What fails is the pivoting route, not invertibility.
Common mistakes
Common mistake
(Reading pivots as eigenvalues)
In the first worked example the pivots are , with sum , while the trace of the matrix is , which is the sum of the eigenvalues. The pivots are not the eigenvalues. Only the number of positive and of negative entries agrees, and the products agree: .
Common mistake
(Taking the square root of a negative pivot)
For a positive definite matrix one may write . If some this would need the real number , which does not exist. For we have , and there is no real with : such a factorization would give , yet at .
Exercises
Quick check
(Compute the pivots of and decide whether is positive definite.)
Solution · Answer
and , so , and . Both pivots are positive, so is positive definite. Check: .
Quick check
(Use the pivots to show that is indefinite, and exhibit one vector with and one with .)
Solution · Answer
and , so and have opposite signs and is indefinite. Here , so gives and gives .
Quick check
(A real symmetric matrix has leading minors , , . How many of its eigenvalues are positive, and how many negative?)
Solution · Answer
The pivots are , and . Two pivots are positive and one is negative, so has two positive eigenvalues and one negative eigenvalue (inertia ).
Quick check
(If a real symmetric matrix has , can we conclude that is singular?)
Solution · Answer
No. The matrix has but . A vanishing first pivot only blocks the scalar-pivot formula; it says nothing about invertibility.