An operator space is a linear family of matrices or operators. It is called hyperreflexive when a simple test controls membership in it quantitatively: whenever every vector $x$ is sent close to the set of places the family can send $x$, the operator must itself be close to the family. The constant in that implication is the hyperreflexivity constant. The known results about how this constant behaves near nilpotent matrices (matrices some power of which is zero) are almost all qualitative. Drahovský and Zając show that in finite dimensions an operator is "hyperreflexive" exactly when it is similar to a diagonal matrix, so any nonzero nilpotent fails. Bessonov, Bračič and Zając build a direct sum of $2\times 2$ diagonalizable blocks whose constants run off to infinity. Neither result says how fast the constant grows as a diagonalizable matrix is deformed toward a Jordan block, which is the shifting, non-diagonalizable normal form of a nilpotent. Neither says which measurable property of the matrix controls that growth.
This note proposes to measure that rate and to prove a two-sided bound for it. The bound would be stated in terms of the eigenvector condition number, the quantity that measures how nearly parallel a matrix's eigenvectors are. The long-range payoff is a line of attack on an open question posed by Drahovský and Zając: must a hyperreflexive quasinilpotent operator on an infinite-dimensional space be zero? The attack works through finite truncations of the prototypical quasinilpotent, the Volterra integration operator. There, the exact compactness numbers computed by Bakşi, Khan, Lang and Musil serve both as a calibration target and as a candidate controlling parameter. Algebraic complexity plays a smaller, supporting role: it certifies the exact Jordan structure of the matrices being studied.
Core Claim
For a matrix $T$, let $\mathcal{W}(T)$ be the algebra of polynomials in $T$. The hyperreflexivity constant $\kappa(\mathcal{S})$ of an operator space $\mathcal{S}$ is the smallest $\kappa$ with
$$\operatorname{dist}(A,\mathcal{S}) \le \kappa \sup_{\|x\|=1}\operatorname{dist}(Ax,\mathcal{S}x)\quad\text{for all } A.$$
Here the left side is operator-norm distance, and the right side, called the Arveson distance, is the worst-case vector-wise distance. For a nonderogatory $T$, meaning one eigenvector per eigenvalue (which covers both distinct-eigenvalue matrices and single Jordan blocks), the commutant $\{T\}'$ equals $\mathcal{W}(T)$. The algebra and the commutant questions therefore coincide on exactly the family we study.
Write $T=SDS^{-1}$ with $D$ diagonal and distinct eigenvalues, and let $\operatorname{cond}(T)=\inf\|S\|\|S^{-1}\|$. We conjecture that there are an exponent $\alpha_k>0$ and constants depending only on dimension such that
$$c_n\,\operatorname{cond}(T)^{\alpha_k}\;\le\;\kappa(\mathcal{W}(T))\;\le\;C\,\operatorname{cond}(T)^{2}.$$
In this bound, $k$ is the size of the largest Jordan block that $T$ is approaching. The upper bound should follow from the similarity-invariance estimate in Proposition 1.1 of Bessonov et al.. It transports the bounded constant of a diagonal algebra through the similarity $S$. Our reading is that it costs about $\|S\|^2\|S^{-1}\|^2$, and we will re-derive the exponent rather than assume it.
The new content is the lower bound: a proof that $\kappa$ must grow at a definite polynomial rate in $\operatorname{cond}(T)$, together with the dependence of $\alpha_k$ on $k$. This turns Bessonov et al.'s Proposition 1.12, a single blow-up example with $2\times 2$ blocks, into a law. It also identifies what a quasinilpotent operator would have to avoid in order to be hyperreflexive.
One point in the source material must be settled before any of this means anything. Under the metric definition, every reflexive space in finite dimensions has a finite constant. The reason is that both sides of the inequality above are seminorms on a finite-dimensional quotient that vanish on the same set, so they are equivalent. The Drahovský–Zając statement that a $3\times3$ nilpotent is "reflexive but not hyperreflexive" therefore cannot refer to the metric constant of $\mathcal{W}(T)$. Given their companion work on hyperinvariant subspaces, it most likely refers to a notion built on the lattice of hyperinvariant subspaces. We treat the metric constant as the object of study. We treat the Drahovský–Zając dichotomy as its qualitative shadow at $\operatorname{cond}=\infty$, and the first experiment checks this reading directly.
Research Program
The first stage is exact analysis in dimension two. There the Arveson supremum is an optimization over a circle and $\mathcal{W}(T)$ is two-dimensional, so $\kappa$ should admit a closed form in the angle between the eigenvectors. This recovers and sharpens the blow-up of Proposition 1.12 in Bessonov et al. and pins down $\alpha_2$.
The second stage is certified numerics for $n$ up to about 8. We study three deformation families, each approaching a Jordan block of size $k$ at rate $\varepsilon$: $T_\varepsilon = J_k + \varepsilon D$, and two others. The operator-norm distance to $\mathcal{W}(T)$ is a semidefinite program, which gives exact upper values of the left side. The Arveson supremum is a nonconvex polynomial optimization over the unit sphere. We lower-bound it by optimizing from many starting points, and upper-bound it by evaluating on a net whose size is set using the covering-number bounds for real algebraic varieties of Zhang and Kileel. Combined with a Lipschitz estimate, that net turns sampling into a certified interval for $\kappa$. Their paper supplies the net-size method, and we apply it to a new target.
The third stage is proof. The numerics will suggest the extremal pair: the operator $A$ and vector $x$ that nearly achieve $\kappa$. We expect $A$ to be built from the "missing" upper-triangular entries that a limiting Jordan block would allow. From that pair we aim to prove the lower bound and the formula for $\alpha_k$.
The fourth stage moves to infinite dimensions through finite sections. We take $n\times n$ truncations of Volterra-type operators, namely strictly lower-triangular summation matrices, which are nilpotent. We then take their best diagonalizable approximants at a fixed operator-norm distance $\varepsilon$ and track $\kappa$ as $n$ grows. The hypothesis to test is that the approximation numbers staying at $1/2$ in Bakşi et al. witness a non-compact, Jordan-like core. On that core, the condition number of any nearby diagonalizable matrix must grow with $n$, which would force $\kappa$ to diverge. Their decaying isomorphism numbers would, under this hypothesis, not control $\kappa$.
If the hypothesis holds uniformly in $n$, it gives a concrete route to proving that $V$ is not hyperreflexive in the metric sense. Such a proof would be a partial answer to the Drahovský–Zając question for a natural class of quasinilpotents. It would not settle the question for all of them, since Müller and Zając exhibit quasinilpotents with reflexive commutants. Their example, with $\|T^n\|\to 0$ arbitrarily fast, is the natural stress test for the conjecture.
The complexity thread is modest and honest about its role. Every experiment depends on knowing the exact Jordan type of the limit matrix, and floating-point arithmetic cannot certify it. The number of distinct eigenvalues is the degree of the squarefree part, or radical, of the minimal polynomial. That is the same parameter that controls factor size in the closure theorem of Dutta, Saxena and Sinhababu. For symbolic families $T(x_1,\dots,x_m)$, deciding whether the generic member is diagonalizable amounts to asking whether a discriminant is identically zero. That is a polynomial identity test, a check of whether a polynomial given as a circuit is identically zero. Low-degree factors of that discriminant, which describe where the family degenerates, can be extracted deterministically by the reduction of Dutta, Sinhababu and Thierauf. This lets us state results about whole families rather than sampled members. We do not claim it advances identity testing itself.
Baselines
The naive reference point is randomized sampling of the Arveson supremum: draw unit vectors, take the maximum of $\operatorname{dist}(Ax,\mathcal{W}(T)x)$, and combine it with the SDP value. This gives only an upper estimate of $\kappa$, because an underestimated supremum inflates the ratio. It is what one would do without the proposal, and our certified intervals must be strictly tighter at equal compute. The comparison uses the same matrices, the same SDP solver, and the same number of function evaluations. We beat it if the certified interval width falls below the naive method's spread across seeds.
The strongest existing tool is the similarity-invariance bound of Proposition 1.1 in Bessonov et al.. It is the only quantitative control on $\kappa$ for non-normal matrices in the file. We evaluate it on every test matrix using the computed $\operatorname{cond}(T)$ and compare it to the certified $\kappa$. No reported numerical values of hyperreflexivity constants were found on Lacuna for any family, so the comparison is against the bound's formula, not a published number. The proposal beats this baseline in two ways. It must supply a lower bound, which Proposition 1.1 does not give, and show the upper exponent is not tight whenever the measured slope of $\log\kappa$ against $\log\operatorname{cond}$ is strictly below 2.
For the compactness arm, the reference values are exact. Bakşi et al. report $a_n(\sigma)=c_n(\sigma)=d_n(\sigma)=1/2$ and $i_n(\sigma)=m_n(\sigma)=1/(2n-1)$ for all $n\ge 2$ for the discrete summation operator $\sigma:\ell^1\to c$. Our finite-section code must reproduce these values to solver tolerance before any correlation with $\kappa$ is trusted.
The key ablation replaces $\operatorname{cond}(T)$ with competing predictors of $\kappa$ fitted on the same data. These are the minimum eigenvalue gap, the departure from normality $\|T^*T-TT^*\|$, the decay of $\|T^j\|$, and the approximation numbers. This isolates whether eigenvector conditioning is really the controlling quantity.
Why not just…?
Why not just use the similarity bound and stop? Proposition 1.1 only says $\kappa$ cannot grow faster than a power of the condition number. It is silent on whether $\kappa$ must grow at all, and that is the question relevant to quasinilpotents. The blow-up example of Bessonov et al. shows growth can happen but not that it must. Our deformation families will show directly whether the measured slope is positive and whether it matches the upper exponent.
Why not treat near-nilpotent matrices as small perturbations of diagonal ones and read $\kappa$ off by continuity? The file's evidence is that hyperreflexivity behavior is discontinuous at Jordan structure. The finite-dimensional dichotomy of Drahovský and Zając and the unbounded constants of Proposition 1.12 both point this way. Continuity arguments give the right answer everywhere except the regime we care about. Measuring the divergence rate is precisely what perturbation theory cannot provide.
Why not just brute-force $\kappa$ with a generic nonconvex optimizer? It would often find the right number in small dimensions, and this is a real threat to the value of the certification machinery for $n\le 4$. The difference is that the optimizer yields no certificate. It also becomes unreliable exactly when $\operatorname{cond}$ is large, because the extremal vector concentrates near a nearly degenerate eigenvector pair. We will run both methods and report where they disagree. If they never disagree up to $n=8$, we will say so and keep the net-based bound only for the proof-supporting cases.
Why not use randomized identity testing, or plain exact rational arithmetic, for the Jordan-type certification? For individual numeric matrices, exact rational arithmetic is enough, and we will use it. The randomized test of Bläser and Pandey is also entirely adequate in practice. The deterministic route matters only if we want unconditional statements about symbolic families. We therefore present it as optional and do not rest any main claim on it.
Proposed Experiments
The first decisive check tests the definitional reading. We compute the certified metric $\kappa(\mathcal{W}(T))$ for the Drahovský–Zając $3\times3$ example and for a $3\times3$ Jordan block. Both values should be finite. If either is infinite, our reading of the metric notion is wrong and the program must be rebuilt around their notion.
The second check fits the slope of $\log\kappa$ against $\log\operatorname{cond}(T)$ along $T_\varepsilon=J_k+\varepsilon D$ and two other deformation families for $k=2,\dots,6$. The conjecture predicts a positive slope $\alpha_k$ that is the same across families for each $k$ and changes monotonically with $k$. A family where $\operatorname{cond}\to\infty$ while $\kappa$ stays bounded would falsify the conjecture outright. The ablation would then show which predictor should replace the condition number.
The third check runs the finite-section Volterra study for $n$ up to about 200. It is calibrated against the exact $s$-numbers of Bakşi et al., and it asks whether $\kappa$ of the $\varepsilon$-nearest diagonalizable approximants grows without bound in $n$ at fixed $\varepsilon$. Growth tracking the non-decaying approximation numbers would support the proposed mechanism for the quasinilpotent question. Bounded $\kappa$ would suggest that Volterra-type operators may escape it, and would redirect the effort toward constructing a nonzero hyperreflexive quasinilpotent, a counterexample in the spirit of Müller and Zając.
References
- Hyperreflexive operators on finite dimensional Hilbert spaces
- Non-hyperreflexive reflexive spaces of operators
- Strict 𝑆-numbers of the Volterra operator
- Hyperinvariant subspaces of operators on Hilbert spaces
- Covering Number of Real Algebraic Varieties and Beyond: Improved Bounds and Applications
- A quasi-nilpotent operator with reflexive commutant, II
- Discovering the Roots: Uniform Closure Results for Algebraic Classes Under Factoring
- Derandomizing Multivariate Polynomial Factoring for Low Degree Factors
- Polynomial Identity Testing for Low Degree Polynomials with Optimal Randomness