JOVANA
Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

Positive operators, square roots, and variational eigenvalues

Specialize the spectral theorem to operators with nonnegative spectrum. You get a unique positive square root, and a way to characterize every eigenvalue by maximizing or minimizing the Rayleigh quotient.

Positivity is a statement about the spectrum

A self-adjoint operator is positive semidefinite when <Ax, x> >= 0 for all x, and positive definite when the inequality is strict for x != 0. Via the spectral theorem this turns into a clean spectral statement: A is positive iff all its eigenvalues are nonnegative. The quadratic form <Ax, x> = sum lambda_i |<x, q_i>|^2 is a nonnegative combination of squares exactly when every lambda_i >= 0. This connects straight back to Vol I's positive definite matrix.

The unique positive square root

Positivity plus functional calculus gives a clean construction: the positive square root. Since every eigenvalue of a positive A is >= 0, the function sqrt is well-defined on the spectrum, so A^{1/2} = sum sqrt(lambda_i) P_i is itself a positive operator with (A^{1/2})^2 = A. And it is unique among positive operators — there is exactly one positive B with B^2 = A.

A = [2, 1; 1, 2]  (positive: eigenvalues 3, 1, both > 0)

  spectral resolution: A = 3 P_1 + 1 P_2
      P_1 = [0.5, 0.5; 0.5, 0.5],   P_2 = [0.5,-0.5;-0.5, 0.5]

  positive square root applies sqrt eigenvalue-wise:
      A^{1/2} = sqrt(3) P_1 + sqrt(1) P_2
              = sqrt(3) [0.5,0.5;0.5,0.5] + [0.5,-0.5;-0.5,0.5]

  verify:  (A^{1/2})^2 = 3 P_1 + 1 P_2 = A     (P_i orthogonal idempotents)

Why unique: any positive B with B^2 = A is diagonal in A's eigenbasis
and must take the NONnegative root on each eigenvalue -> forced to be A^{1/2}.
The positive square root via functional calculus, and why it is unique.

The positive square root is not a curiosity — it is the engine behind the polar decomposition A = U |A| with |A| = (A* A)^{1/2}, and behind whitening transforms in statistics. Whenever you need a 'half' of a positive operator, the spectral theorem hands you exactly one.

Eigenvalues as optimization: Rayleigh and min-max

For a self-adjoint A the Rayleigh quotient R(x) = <Ax, x> / <x, x> is the average stretch of A in direction x. The spectral theorem pins its range exactly: R(x) ranges over [lambda_min, lambda_max], reaching the largest eigenvalue when x is the top eigenvector and the smallest when x is the bottom one. So the extreme eigenvalues are an optimization problem, not just roots of a polynomial.

The interior eigenvalues are captured too by the Courant-Fischer min-max theorem: the k-th eigenvalue is a min over k-dimensional subspaces of the max of the Rayleigh quotient on each (and dually, a max-min). This variational description is what makes eigenvalues *stable*: Weyl's inequalities say that perturbing A by E shifts each eigenvalue by at most ||E||, because each eigenvalue is an optimum of a quotient that itself moves only a little.