Is every positive-definite matrix invertible?

Yes, in the real symmetric finite-dimensional setting used here.

Is every positive-semidefinite matrix positive definite?

No. A semidefinite matrix may have zero eigenvalues.

Does positive definiteness require all entries to be positive?

No. It concerns quadratic values, not entrywise signs.

AI Glossary / Mathematics and optimization / Advanced

Positive definite (pd)

Definition

A real symmetric matrix is positive definite, abbreviated pd, when its quadratic form is strictly positive for every nonzero vector. In symbols, Q is positive definite if xᵀQx > 0 whenever x ≠ 0. This means the matrix assigns positive quadratic size or curvature to every nonzero direction, not that every matrix entry is positive.

What it means

The expression xᵀQx combines a direction x with a matrix Q to produce one number. Positive definiteness says this number cannot be zero or negative unless x itself is zero. It is stronger than positive semidefiniteness, which permits zero values in nonzero directions.

This entry uses the symmetric real-matrix convention in the supplied glossary. For complex matrices, the corresponding setting uses Hermitian matrices and the conjugate transpose. Keeping the matrix class explicit prevents ambiguous tests and inappropriate use of numerical routines.

How it works

For a real symmetric matrix, positive definiteness is equivalent to all eigenvalues being strictly positive. Consequently the matrix is invertible. Its quadratic form also defines a squared norm, and a quadratic objective with that matrix as its Hessian has strictly positive curvature in every direction.

The Cholesky factorization expresses such a matrix as LLᵀ, where L is triangular with positive diagonal entries. It provides an efficient route for solving related linear systems. In numerical work, however, mathematical definiteness and computational stability are separate concerns: a very small positive eigenvalue can make the matrix nearly singular.

Do not rely on one implementation's successful factorization to verify every input property. Some routines use only one triangle of the matrix and assume symmetry rather than checking it. Symmetry, scaling, and the accuracy of the supplied data need explicit attention.

Three practical examples

1. Quadratic optimization: a positive-definite Hessian in a quadratic objective ensures a unique unconstrained minimizer.

2. Gaussian modeling: a positive-definite covariance matrix describes nonzero variance in every nonzero linear direction and supports a nonsingular multivariate Gaussian density.

3. Regularized fitting: adding a positive multiple of the identity to a positive-semidefinite Gram matrix can make the resulting system positive definite.

Think of it this way

Think of a bowl that rises away from its bottom in every direction. No direction is completely flat. This captures a positive-definite quadratic form. The analogy does not imply that a general nonlinear surface is bowl-shaped everywhere merely because its Hessian is positive definite at one point.

Worked example

Let Q have rows (2, −1) and (−1, 2). For x = (a, b), its quadratic form is 2a² − 2ab + 2b².

Rewrite it as a² + b² + (a − b)². This is strictly positive whenever at least one of a and b is nonzero, so Q is positive definite. Its negative off-diagonal entries do not prevent that property.

Two eigenvector directions make the geometry visible. Q(1, 1) = (1, 1), giving eigenvalue 1. Q(1, −1) = 3(1, −1), giving eigenvalue 3. Both eigenvalues are positive, but the quadratic form grows more rapidly in the second direction for vectors of equal Euclidean length.

Now minimize F(x) = xᵀQx/2 − cᵀx with c = (1, 0). The stationary equation Qx = c gives 2a − b = 1 and −a + 2b = 0. Solving yields a = 2/3 and b = 1/3.

At this point, xᵀQx = xᵀc = 2/3, so F = 1/3 − 2/3 = −1/3. For any displacement h, the objective increases by hᵀQh/2, which is positive for h ≠ 0. Thus the stationary point is the unique global minimum of this quadratic. The calculations are original and hand-checked.

Illustrative case study

Fictional case: Harbor Sensors models jointly varying instrument readings with a covariance matrix. After adding a nearly redundant channel, its numerical system becomes sensitive to tiny changes even though a mathematical analysis suggests positive eigenvalues.

The team checks units, symmetry, and the smallest eigenvalues relative to the matrix scale. It investigates whether redundancy or estimation noise explains the instability before adding a documented regularization adjustment. A successful factorization is treated as one numerical check, not proof that the covariance model captures the real process. No measured accuracy benefit is claimed.

Common misconceptions and mistakes

Positive diagonal entries are necessary but not sufficient. The symmetric matrix with rows (1, 2) and (2, 1) has positive diagonal entries, yet the direction (1, −1) gives a negative quadratic value.

A positive determinant alone is also insufficient in general dimensions. Positive definiteness concerns every direction. Conversely, a positive-definite matrix can contain negative off-diagonal values, as the worked example shows.

Positive definite does not mean well-conditioned. A diagonal matrix with entries 1 and 0.00000001 is positive definite, but its eigenvalue ratio is 100 million, indicating very different scales across directions.

Business use

The property helps determine which optimization and linear-algebra tools are appropriate. Teams should document whether definiteness follows from the model construction or is only inferred numerically. Distinguish genuine negative curvature from small roundoff effects using scale-aware checks.

If regularization is added to ensure numerical definiteness, record its magnitude and impact. It changes the system being solved, and an apparently convenient repair can materially alter an uncertainty estimate or fitted solution.

Check your understanding

A symmetric matrix has eigenvalues 2, 5, and 0. Is it positive definite?

Answer: No. The zero eigenvalue gives a nonzero direction with zero quadratic value. It is positive semidefinite.

Related terms

Eigenvalue
Hessian
Covariance matrix
Condition number
positive semidefinite
quadratic form
Cholesky factorization

Learn next

compare positive-definite and singular semidefinite matrices, then study why conditioning matters even when an inverse exists mathematically.

Sources

Primary sources: Aalto Dictionary of Machine Learning, positive definite: https://aaltodictionaryofml.github.io/ADictML_Main.pdf ; NumPy's Cholesky documentation, including its symmetry assumption: https://numpy.org/doc/stable/reference/generated/numpy.linalg.cholesky.html . Relevant passages were read. The examples are original and are not reported as executed factorization tests.

Previous: Positive class Next: Positive semi-definite (psd)