--- title: Crouzeix's Conjecture in Matrix Analysis url: https://www.emergentmind.com/topics/crouzeix-s-conjecture type: topic --- # Crouzeix's Conjecture in Matrix Analysis Crouzeix’s conjecture is the assertion that the numerical range of a matrix controls its holomorphic functional calculus with the sharp universal constant \(2\). For a square complex matrix \(A\), with numerical range \[ W(A)=\{\langle Ax,x\rangle:\|x\|=1\}, \] the conjecture states that \[ \|p(A)\|\le 2\max_{z\in W(A)}|p(z)| \] for every polynomial \(p\). It lies at the intersection of numerical range theory, spectral sets, operator model theory, and matrix function estimates. The best general theorem currently available is the Crouzeix–Palencia bound \(1+\sqrt2\), while the sharp constant \(2\) is known only in special classes and asymptotic regimes [2006.04901]. ## 1. Statement, variants, and basic framework The conjecture is most commonly formulated for polynomials, but the same inequality is equivalent to corresponding statements for functions analytic on a neighborhood of \(W(A)\), and also for functions in \(\mathbf A(W(A))\), the algebra of functions analytic on \(W(A)^\circ\) and continuous on \(W(A)\). If it holds for matrices, then it extends to bounded operators on Hilbert space; if it holds for polynomials, then Mergelyan approximation extends it to suitable analytic functions on \(W(A)\) [2006.04901]. A standard spectral-set reformulation introduces, for an open set \(\Omega\subset\mathbb C\) containing \(\sigma(A)\), \[ C(\Omega,A):=\sup\left\{\frac{\|f(A)\|}{\|f\|_\Omega}: f \text{ analytic on }\Omega \text{ and continuous on }\partial\Omega\right\}, \qquad \|f\|_\Omega=\sup_{z\in\Omega}|f(z)|. \] Crouzeix’s conjecture is precisely the claim that \[ C(W(A),A)\le 2. \] In this language, \(W(A)\) is conjectured to be a \(2\)-spectral set for every matrix. The strongest general theorem presently cited in the literature is \[ C(W(A),A)\le 1+\sqrt2, \] the Crouzeix–Palencia bound [2303.02042]. A stronger, completely bounded version replaces scalar polynomials by matrix-valued polynomials \(P\) and asks for \[ \|P(A)\|\le 2\max_{z\in W(A)}\|P(z)\|. \] Special-case proofs often establish this stronger form. In the same circle of ideas, an equivalent numerical-radius formulation is available: the conjectured constant \(2\) is equivalent to the sharp bound \[ w(p(A))\le \frac54 \sup_{z\in W(A)}|p(z)|, \] where \(w(\cdot)\) is the numerical radius [1701.01365]. ## 2. Spectral-set machinery and extremal structure A central analytic tool is the Cauchy integral representation \[ f(A)=\frac{1}{2\pi i}\int_{\partial\Omega} f(\sigma)(\sigma I-A)^{-1}\,d\sigma, \] valid when \(\Omega\) contains \(\sigma(A)\). In the Crouzeix–Palencia framework one also defines a companion transform \[ g(A)=\frac{1}{2\pi i}\int_{\partial\Omega}\overline{f(\sigma)}(\sigma I-A)^{-1}\,d\sigma, \] and, after arclength parametrization \(\sigma(s)\), the Hermitian kernel \[ \mu(s,A):=\left[\frac{\sigma'(s)}{2\pi i}(\sigma(s)I-A)^{-1}\right] + \left[\frac{\sigma'(s)}{2\pi i}(\sigma(s)I-A)^{-1}\right]^*. \] For \(\Omega=W(A)\), positivity properties of \(\mu\) yield the universal \(1+\sqrt2\) estimate. For regions strictly inside \(W(A)\), positivity may fail, and modified estimates introduce scalar correction terms built from the smallest eigenvalue of \(\mu(s,A)\) [2303.02042]. The geometry becomes especially transparent for disks. If \(\Omega=\mathcal D(c,r)\) contains \(W(A)\), then the associated scalar transform collapses to a constant, \[ g(z)=\overline{f(c)}, \] and the Caldwell–Greenbaum–Li singular-vector argument yields the exact disk bound \[ \|f(A)\|\le 2\sup_{z\in\Omega}|f(z)|. \] This disk case is a recurring local model in perturbative arguments for “nearly circular” numerical ranges [2303.02042]. The extremal viewpoint is equally important. For a simply connected domain \(\Omega\), an extremal function for \((A,\Omega)\) is a function \(f\in H^\infty_1(\Omega)\) maximizing \(\|f(A)\|\). Such extremals have the form \[ f=B_A\circ \phi, \] where \(\phi:\Omega\to\mathbb D\) is conformal and \(B_A\) is a finite Blaschke product of degree at most \(n-1\). Associated extremal vectors satisfy strong orthogonality identities; for example, if \(\|f(A)\|>1\), then an associated extremal unit vector \(x\) satisfies \[ \langle f(A)x,x\rangle=0. \] There is also a representation theorem: for an extremal pair \((f,x)\), there exists a unique probability measure \(\mu\) on \(\partial\Omega\) such that \[ \langle h(A)x,x\rangle=\int_{\partial\Omega} h\,d\mu \qquad (h\in \mathcal A(\Omega)), \] with additional cancellation identities when \(\|f(A)\|>1\). These results clarify why extremal Blaschke products, conformal maps, and singular vectors are structurally central to the conjecture [2006.04901]. ## 3. Verified classes and sharp extremal mechanisms A substantial literature verifies the conjectured constant \(2\) for specific matrix classes. The following families are explicitly established in the cited works. | Matrix class | Hypothesis | Conclusion | |---|---|---| | \(2\times2\) matrices | none | conjecture holds | | \(3\times3\) tridiagonal constant-diagonal matrices | equivalently, elliptic numerical range centered at an eigenvalue | complete \(2\)-spectral set | | Contractions with separated eigenvalues | \(\delta_A\ge 2\sqrt2/3\) | conjecture holds | | Cyclic weighted shifts \(P_d\operatorname{diag}(a_1,\dots,a_d)\) | none | complete \(2\)-spectral set | | Some unicritical compressed shifts | degree \(4\) or \(5\), explicit parameter ranges | conjecture holds | For contractions with distinct eigenvalues in \(\mathbb D\), the relevant separation parameter is the pseudohyperbolic separation constant \[ \delta_A=\inf_j\prod_{k\ne j}\left|\frac{z_j-z_k}{1-\overline{z_k}z_j}\right|, \] and if \(\delta_A\ge 2\sqrt2/3\), then one obtains \[ \|f(A)\|\le 2\max_j|f(z_j)| \] for every \(f\in H^\infty\), hence the conjecture follows a fortiori since \(\sigma(A)\subseteq W(A)\) [2006.04901]. For \(3\times3\) matrices with elliptic numerical range centered at an eigenvalue, equivalently tridiagonal matrices with constant diagonal, \[ \begin{bmatrix} a&b_1&0\\ c_1&a&b_2\\ 0&c_2&a \end{bmatrix}, \] the conjecture is proved in the strong completely bounded form by combining an explicit conformal map from the ellipse to the disk with carefully constructed similarity transforms of condition number \(2\) [1701.01365]. For cyclic weighted shifts \[ M(a_1,\dots,a_d)=P_d\,\mathrm{diag}(a_1,\dots,a_d), \] a recent alternative proof shows \[ \Psi(M)=\Psi_{cb}(M)\le 2, \] so \(W(M)\) is a complete \(2\)-spectral set. Moreover, if \(a_1\cdots a_d\neq 0\), then the constant is actually strict: \[ \Psi(M)=\Psi_{cb}(M)<2. \] The proof exploits rotational symmetry of \(W(M)\), the identity \(\phi(M)=cM\) for the normalized Riemann map \(\phi\), and an extremal-pair argument within the Crouzeix–Palencia decomposition [2508.12768]. The sharpness problem is illuminated by the structure of half-radial matrices, defined by \[ \|A\|=2r(A), \] where \(r(A)\) is the numerical radius. These are exactly the equality cases for the identity polynomial \(p(z)=z\). They are characterized by a rigid block form \[ A\sim_u (\|A\|I_m\otimes J)\oplus B, \qquad J=\begin{bmatrix}0&1\\0&0\end{bmatrix}, \] with \(\|B\|<\|A\|\) and \(r(B)\le \|A\|/2\). For monomials \(p(z)=z^k\), equality \[ \|A^k\|=2\max_{\zeta\in W(A)}|\zeta^k| \] holds exactly when \(A^k\) is half-radial and \(r(A^k)=r(A)^k\). In dimension \(n+1\), equality for \(p(z)=z^n\) characterizes the Crabb–Choi–Crouzeix matrix, and more generally monomial extremality forces a Crabb–Choi–Crouzeix block. This identifies a concrete extremal mechanism behind the sharp constant \(2\) [1804.10407]. ## 4. Reductions, inheritance principles, and operator models Several recent results reorganize the conjecture by showing that many matrix constructions preserve validity of the \(2\)-bound. If \(A\) satisfies the conjecture, then so do all affine images \(\mu A+\lambda I\), the transpose \(A^T\), and the adjoint \(A^*\). The class is also closed under direct sums, and every rank-one matrix satisfies the conjecture. Tensoring with a normal matrix preserves the property as well: if \(N\) is normal and \(A\) satisfies the conjecture, then both \(N\otimes A\) and \(A\otimes N\) satisfy it [2306.12183]. These closure properties feed into a reduction to cyclic matrices. A vector \(y\) is cyclic for \(A\) if \[ \operatorname{span}\{y,Ay,A^2y,\dots\}=\mathbb C^n, \] and \(A\) is cyclic if it has such a vector. The key reduction is that the full conjecture holds if and only if it holds for cyclic matrices. More precisely, if it holds for all cyclic matrices, then it holds for all matrices. In dimension \(3\), every non-cyclic matrix already satisfies the conjecture, because the extremal norm behavior compresses to a \(1\)- or \(2\)-dimensional subspace where the conjecture is known [2306.12183]. A functional-model reformulation makes the cyclic case more specific. If \(A\) is cyclic and \(v\) is cyclic for \(A^*\), then one defines \[ \widetilde u(z)=\langle e^{zA}u,v\rangle, \qquad E_{A,v}=\{\widetilde u:u\in\mathbb C^n\}, \] with \(\|\widetilde u\|_{E_{A,v}}=\|u\|_{\mathbb C^n}\), and introduces the differentiation operator \[ \mathcal D_{A,v}(\widetilde u)(z)=\langle e^{zA}Au,v\rangle. \] For cyclic \(A\), this operator is unitarily equivalent to \(A\). Consequently, Crouzeix’s conjecture holds in full generality if and only if it holds for these differentiation operators on finite-dimensional spaces of entire functions [2306.12183]. A further operator-model reduction appears for symmetric matrices. In dimension \(3\), the conjecture for symmetric matrices is equivalent to the conjecture for truncated Toeplitz operators on three-dimensional model spaces, and even to the analytic-symbol subclass. This ties one finite-dimensional corner of the conjecture directly to model-space operator theory [2306.12183]. ## 5. Compressions of shifts, Blaschke products, and level-set variants For a finite Blaschke product \[ \Theta(z)=\lambda\prod_{j=1}^n \frac{z-a_j}{1-\overline{a_j}z}, \] the model space \[ K_\Theta=H^2\ominus \Theta H^2 \] has dimension \(n\), and the compressed shift \[ S_\Theta=P_\Theta M_z|_{K_\Theta} \] provides a canonical finite-dimensional model operator. These compressions are central because they realize the completely non-unitary defect-one contractions and offer a highly structured testing ground for Crouzeix-type inequalities [2112.06321]. A weaker but natural specialization is the level set Crouzeix conjecture (LSC). If \(B\) and \(\Theta\) are finite Blaschke products with \(\deg B<\deg \Theta\), then \(\|B(S_\Theta)\|=1\); if the full conjecture held for \(S_\Theta\), it would force \[ \max_{z\in W(S_\Theta)}|B(z)|\ge \frac12. \] Equivalently, with \[ \Omega_{1/2}^B=\{z\in\mathbb C:|B(z)|<1/2\}, \] the level-set formulation is \[ W(S_\Theta)\not\subseteq \Omega_{1/2}^B. \] A decisive sufficient condition is pseudohyperbolic-disk containment: if \[ D_\rho\!\left(z_0,\left(\frac12\right)^{1/(n-1)}\right)\subseteq W(S_\Theta), \] then LSC holds for every \(B\) of degree \(1\), LSC holds for all \(t\) sufficiently close to \(1\), and explicit thresholds are given for \(m=2,\dots,7\) [2312.04537]. The same circle of ideas also produces full, not merely level-set, Crouzeix results for some nilpotent matrices associated with repeated-zero model operators. For instance, certain \(4\times4\) KMS-type nilpotent matrices satisfy the conjecture on explicit parameter intervals such as \(t\in[0,0.363]\), and related families \[ A_{t,m}= \begin{pmatrix} 0&1&t&t^m\\ 0&0&1&t\\ 0&0&0&1\\ 0&0&0&0 \end{pmatrix} \] satisfy the conjecture for \(m=2,3,4\) on explicit \(t\)-ranges obtained via similarity-to-Jordan-block estimates [2312.04537]. ## 6. Random matrices, algorithms, computation, and open directions A recent asymptotic result verifies a restricted form of the conjecture for complex Ginibre matrices. If \(G_N\) is an \(N\times N\) complex Ginibre matrix normalized so that the circular law fills the unit disk, then \[ C(W(G_N),G_N)\to 2 \qquad\text{almost surely as }N\to\infty. \] Equivalently, for every \(\varepsilon>0\), \[ C(W(G_N),G_N)\le 2+\varepsilon \] eventually almost surely. This is not a proof of the full conjecture: it is restricted to a random ensemble, asymptotic in \(N\), probabilistic, and relies on the fact that \(W(G_N)\) becomes nearly circular. The proof combines asymptotic circularity \[ d_H(W(G_N),\mathcal D(0,\sqrt2))\to 0, \] uniform resolvent bounds on annuli, and a perturbative disk argument in the Crouzeix–Palencia/Caldwell–Greenbaum–Li framework [2303.02042]. The same work shows how this asymptotic \(2\)-spectral-set behavior enters GMRES. For systems \[ (I+\sigma G_N)x=b,\qquad \sigma\in(0,1), \] the ideal-GMRES residual satisfies, almost surely as \(N\to\infty\), \[ \frac{\|r^{(k)}\|}{\|b\|}\le 2(\sqrt2\,\sigma)^k, \] with an alternative pseudospectral bound \[ \frac{\|r^{(k)}\|}{\|b\|}\le \min_{\eta>1}\ \eta\,\mathfrak e(\eta)^{-1/2}(\sigma\eta)^k. \] Here the numerical range yields the asymptotically optimal constant \(2\), while pseudospectra provide explicit spectral sets with usually worse constants [2303.02042]. Crouzeix-type bounds also appear in optimization. In nonlinear acceleration of multistep methods, including momentum and primal-dual schemes, the local linearized iteration operator \(G\) is generally non-symmetric. The RNA analysis reduces performance to norms of matrix polynomials \(p(G)\), and Crouzeix’s inequality converts this into the planar approximation problem \[ \|p(G)\|_2 \le c\max_{z\in W(G)}|p(z)|. \] Consequently, acceleration performance is controlled by a Chebyshev problem on the numerical range of a nonsymmetric operator. This is the mechanism that extends numerical-range-based acceleration theory beyond symmetric fixed-point maps [1810.04539]. Computational work on the Crouzeix ratio \[ f(p,A)=\frac{\max_{z\in W(A)}|p(z)|}{\|p(A)\|_2} \] provides another perspective. Extensive nonsmooth optimization experiments, totaling nearly half a million runs and about \(250\) million evaluations, found many locally minimal values strictly between \(0.5\) and \(1\), but none below \(0.5\). The same locally minimal values frequently appear in both real and complex formulations, and approximate nonsmooth stationarity was verified numerically using near-active local maximizers on \(\partial W(A)\). These computations strongly support the conjecture that the global minimum of the Crouzeix ratio is \(0.5\), equivalently that the optimal universal constant is \(2\) [2105.14176]. A broader extension replaces the classical numerical range by the scaled \(q\)-numerical range \[ \Omega_q=\frac{W_q(A)}{q}, \qquad W_q(A)=\{\langle Ax,y\rangle:\|x\|=\|y\|=1,\ \langle x,y\rangle=q\}, \] and develops spectral-set bounds through a generalized Crouzeix–Palencia framework. In this setting one proposed \(q\)-analogue is \[ \|p(A)\| \le \max\!\left(1,\frac{2|q|}{1+\sqrt{1-|q|^2}}\right)\sup_{z\in\Omega_q}|p(z)|, \] which recovers the classical conjecture at \(q=1\) [2603.15536]. The central open problem remains the original universal \(2\)-bound for arbitrary matrices. Several subsidiary questions sharpen the landscape. One asks whether the class of matrices satisfying the conjecture is closed under addition; a positive answer would imply the full conjecture, since rank-one matrices already satisfy it. Another asks whether extremal behavior can always be compressed to dimension \(2\), a mechanism that would force the conjecture by reduction to the \(2\times2\) case. In the model-operator direction, open problems include characterizing extremal Blaschke-product degrees and the relation between norm-extremal and numerical-radius-extremal functions. In the random-matrix direction, it remains unknown whether large Ginibre matrices might asymptotically satisfy a better constant \(K<2\); numerics in that setting suggest the possibility of \(\sqrt2\), but no proof is known [2306.12183]. Source: https://www.emergentmind.com/topics/crouzeix-s-conjecture