Crouzeix’s Conjecture: Why the Constant Is Two

On August 4, 2026, Emiel Lorist and Felix L. Schwenninger (Lorist and Schwenninger 2026) posted a proof of Crouzeix’s conjecture. Shanmu Jin (Jin 2026) had posted an independent proof shortly before. The conjecture asks how much a polynomial can amplify a matrix, and its sharp constant, , can already be seen in a simple example.

Let be a complex square matrix. If , we define We measure the size of by its operator norm, the largest length of over all unit vectors .

For a normal matrix, such as a Hermitian matrix, the spectral theorem gives a simple answer: is the largest value of over the eigenvalues of . For a general matrix, the eigenvalues alone may tell us very little. For example, the matrix has both eigenvalues equal to zero, yet .

Crouzeix’s idea was to use a larger set, the numerical range: where denotes the conjugate transpose of . This is a compact convex subset of the complex plane containing every eigenvalue of . For the matrix above, a unit vector gives . Its modulus is at most , and equality is attained when . Thus, for , the norm of is exactly twice the maximum of on .

Could the ratio ever be larger than ? In 2004, Michel Crouzeix (Crouzeix 2004) conjectured that it could not. He proved a universal bound of in 2007 (Crouzeix 2007). In 2017, Crouzeix and Palencia (Crouzeix and Palencia 2017) lowered it to . The 2026 proofs closed the remaining gap.

Theorem. For every complex square matrix and every polynomial with complex coefficients, The constant is best possible.

Lorist and Schwenninger proved the inequality more generally for bounded operators on complex Hilbert spaces, with the maximum replaced by a supremum. Their theorem also allows rational functions whose poles lie outside the closure of the numerical range.

The key idea in their proof is to use all the powers of a function at once. An earlier integral formula compares with a weighted average of the values of on the boundary of a convex region surrounding . There is an error term, and estimating it directly loses something in the constant. Applying the formula to instead gives a sequence of bounded errors with a useful commutation property. A telescoping argument then forces the sharp bound.

There is a natural further question. Suppose is a matrix whose entries are polynomials. Evaluating each entry at gives a block matrix . Does the same bound hold regardless of how many rows and columns has? This is the complete Crouzeix conjecture. OpenAI’s October release gave a proof (OpenAI 2026), and the complete inequality has also been formalized in Lean.

References

Crouzeix, Michel. 2004. “Bounds for Analytical Functions of Matrices.” Integral Equations and Operator Theory 48 (4): 461–77. https://doi.org/10.1007/s00020-002-1188-6.
Crouzeix, Michel. 2007. “Numerical Range and Functional Calculus in Hilbert Space.” Journal of Functional Analysis 244 (2): 668–90. https://doi.org/10.1016/j.jfa.2006.10.013.
Crouzeix, Michel, and Cesar Palencia. 2017. “The Numerical Range Is a -Spectral Set.” SIAM Journal on Matrix Analysis and Applications 38 (2): 649–55. https://doi.org/10.1137/17M1116672.
Jin, Shanmu. 2026. The Numerical Range Is a -Spectral Set. https://doi.org/10.20944/preprints202607.1919.v4.
Lorist, Emiel, and Felix L. Schwenninger. 2026. A Solution to Crouzeix’s Conjecture. https://arxiv.org/abs/2608.03841v2.
OpenAI. 2026. A Direct Proof of the Complete Crouzeix Inequality. https://github.com/openai/math/blob/main/preprints/A-direct-proof-of-the-complete-Crouzeix-inequality-September-26-2026/paper.pdf.

No comment found.

Add a comment

You must log in to post a comment.