--- title: Double-Exponential Capacity Measure url: https://www.emergentmind.com/topics/double-exponential-capacity-measure type: topic --- # Double-Exponential Capacity Measure Searching arXiv for recent and directly relevant papers on the term and neighboring capacity notions. A double-exponential capacity measure is not a single standardized invariant across mathematics and theoretical computer science. In the clearest explicit usage, it denotes a capacity-related lower bound whose worst-case dependence on the number of variables is doubly exponential, typically of the form \(\exp(-2^{\Theta(n)})\) or \(2^{-2^{\Theta(n)}}\), rather than singly exponential \(\exp(-\Theta(n))\). That usage is most sharply articulated in the theory of real stable polynomials and strongly Rayleigh distributions, where capacity controls permanents, coefficients, and aggregated-event probabilities, and where earlier bounds used in metric TSP analysis were “doubly exponential in the number of variables” before being replaced by singly exponential gradient-based estimates via productization [2007.08390]. In adjacent literatures, however, the phrase is absent or explicitly rejected as a natural description: biologically plausible associative memories and dynamic energy networks exhibit exponential, not double-exponential, storage capacity in the relevant network size, while measure-theoretic and geometric capacity theories remain in logarithmic, polynomial-scale, or exponential-time regimes [2601.00984] [2510.24965] [2504.16456] [1408.5892]. ## 1. Terminological scope and principal meaning The most direct arXiv-era meaning of the phrase arises from capacity bounds for real stable polynomials. For \(p \in \mathbb{R}_+^n[x_1,\ldots,x_n]\), the capacity at exponent \(\mathbf{1}\) is \[ \cpc_{\mathbf{1}}(p)=\inf_{x_1,\ldots,x_n>0}\frac{p(x)}{x_1\cdots x_n}, \] and more generally \[ \cpc_\kappa(p)=\inf_{x>0}\frac{p(x)}{x^\kappa}, \qquad x^\kappa=x_1^{\kappa_1}\cdots x_n^{\kappa_n}. \] In the strongly Rayleigh and metric-TSP setting, the relevant probabilities are controlled through such a capacity-type quantity, and the older bounds discussed there are described as “doubly exponential in the number of variables,” with asymptotic behavior \(\exp(-2^{\Theta(n)})\). The replacement bounds obtained by Gurvits and Leake are degree-independent and only singly exponential in \(n\), namely \(\exp(-\Theta(n))\), so in this context a “double-exponential capacity measure” means a capacity bound whose guaranteed lower scale deteriorates as a double exponential in the dimension parameter [2007.08390]. This usage is fundamentally different from the way “capacity” is used in associative memory, dynamical systems, or metric geometry. In those areas, capacity may mean the number of stable memories, an upper-capacity dimension of a measure, or a Sobolev-type condenser capacity; none of those papers adopts double-exponential growth as the natural asymptotic regime. This suggests that the expression functions less as a universal definition than as a comparative descriptor for an especially weak asymptotic lower bound. ## 2. Stable-polynomial capacity and the origin of the doubly exponential regime Real stable polynomials are multivariate polynomials \(p \in \mathbb{R}_+^d[x_1,\ldots,x_n]\) such that \[ p(x_1,\ldots,x_n)\neq 0 \quad\text{whenever }x_1,\ldots,x_n \text{ all lie in the complex upper half-plane}. \] They include determinant polynomials and partition functions of negatively dependent distributions, and they are strongly log-concave. Within this class, capacity is tied to normalization, Hall-type obstructions, and coefficient control. In particular, \[ \cpc_{\mathbf{1}}(p)>0 \iff \mathbf{1}\in \Newt(p), \] and for \(p(x)=\prod_{i=1}^n (Ax)_i\) with \(A\ge 0\), vanishing capacity is equivalent to vanishing permanent and to the existence, up to permutation, of a bottom-left \(i\times j\) zero block with \(i+j>n\). If \(p(\mathbf{1})=1\) and \(\nabla p(\mathbf{1})=\alpha\), then \(\cpc_\alpha(p)=1\); in the special case \(\alpha=\mathbf{1}\), the polynomial is called doubly stochastic and satisfies \(\cpc_{\mathbf{1}}(p)=1\). This makes \(\cpc_\kappa(p)\) a multiplicative robustness measure around a target exponent \(\kappa\) [2007.08390]. The phrase “double-exponential capacity measure” becomes relevant when this robustness measure is used to certify probabilities of aggregated events for strongly Rayleigh measures. In that setting, one studies a generating polynomial \(p\) obtained from a strongly Rayleigh distribution after aggregating coordinates into \(n\) blocks, and seeks lower bounds on coefficients \[ \langle x^\kappa\rangle p = \mathbb{P}[A=\kappa]. \] The older bounds invoked in the metric TSP analysis were degree-independent but “doubly exponential in the number of variables,” so the relevant probability or its governing capacity quantity could decay like \[ \exp(-2^{\Theta(n)}). \] The double-exponential feature therefore does not mean that the polynomial itself has doubly exponential degree or state space; it means that the best certified lower bound on a capacity-controlled quantity deteriorates at that rate as a function of the number of aggregated variables. ## 3. Productization and the reduction to singly exponential bounds The decisive change is the productization technique. Its structural statement is that any real stable polynomial with prescribed normalization at \(\mathbf{1}\) can, at any positive point \(u\), be matched by a product of linear or affine forms with the same value and the same gradient at \(\mathbf{1}\). In the homogeneous case, if \(p(\mathbf{1})=1\) and \(\nabla p(\mathbf{1})=\alpha\), then for any \(u\in\mathbb{R}_+^n\) there exists a nonnegative matrix \(A\) with row sums \(1\), column sums \(\alpha\), and \[ p(u)=\prod_{i=1}^d (Au)_i. \] This reduces the minimization of capacity over all real stable polynomials to a minimization over products of linear forms, where convex-geometric and Hall-type arguments are available [2007.08390]. The resulting bounds are explicit. For a homogeneous real stable polynomial of degree \(n\) in \(n\) variables, if \(p(\mathbf{1})=1\) and \[ \|\mathbf{1}-\nabla p(\mathbf{1})\|_1<2, \] then \[ \cpc_{\mathbf{1}}(p)\ge \left(1-\frac{\|\mathbf{1}-\nabla p(\mathbf{1})\|_1}{2}\right)^n. \] For a general real stable polynomial and \(\kappa\in\mathbb{Z}_+^n\), if \(p(\mathbf{1})=1\) and \[ \|\kappa-\nabla p(\mathbf{1})\|_1<1, \] then \[ \cpc_\kappa(p)\ge \left(1-\|\kappa-\nabla p(\mathbf{1})\|_1\right)^n. \] Both bounds are degree-independent and singly exponential in \(n\). The threshold \(\|\mathbf{1}-\alpha\|_1<2\) is best possible in the homogeneous setting because it is equivalent to the Hall-type condition that every matrix in \(\Mat_n(\alpha)\) has strictly positive permanent [2007.08390]. For strongly Rayleigh measures this yields explicit probability bounds. If \(\mu\) is strongly Rayleigh on \(\{0,1\}^m\), \(A_i=\sum_{s\in S_i}X_s\), and \[ \|\mathbb{E}[A]-\kappa\|_1\le 1-\epsilon, \] then \[ \mathbb{P}[A=\kappa] \ge e^{-\|\kappa\|_1}\frac{\kappa^\kappa}{\kappa!}\,\epsilon^n \ge \epsilon^n\prod_{\kappa_i>0}\frac{1}{e\sqrt{\kappa_i}}. \] In the special case \(\kappa=\mathbf{1}\), \[ \mathbb{P}[A=\mathbf{1}] \ge \left(\frac{\epsilon}{e}\right)^n. \] The practical significance is exact: these estimates replace a double-exponential lower scale by a singly exponential one while preserving independence from the polynomial degree. ## 4. Why exponential memory capacity is not a double-exponential capacity measure In associative-memory theory, “capacity” usually means the number of stable fixed points or retrievable memories, not a lower bound on a coefficient or capacity functional. Krotov and Hopfield’s earlier two-layer dense associative memory had winner-takes-all hidden dynamics and capacity at most linear in the number of hidden neurons. The threshold-based two-layer model of 2026 removes that competition: the hidden indicators satisfy \[ s_\mu=\Theta\!\left(\sum_{\nu=1}^{N_h}J_{\mu\nu}s_\nu-\theta\right), \qquad J_{\mu\nu}=\frac{1}{N_v}\sum_{i=1}^{N_v}\xi_{\mu i}\xi_{i\nu}, \] and with \(\theta=1/2\) in the ideal Gaussian-weight regime, all \(2^{N_h}\) binary hidden states become stable fixed points. The resulting capacity is \[ C(N_h;N_v)\simeq 2^{N_h}, \] provided \(N_v\) is sufficiently larger than \(N_h\). The paper states explicitly that its capacity is exponential in \(N_h\) and “not” double-exponential in any standard network-size metric [2601.00984]. A parallel conclusion holds for the Exponential Dynamic Energy Network. EDEN defines capacity through a bitwise retrieval criterion at transition time, \[ C(N,\epsilon,\delta)= \max\Big\{P\in\mathbb{N}: \Pr[v_i(t_e)\xi_i^{(\mu)}\ge 1-\epsilon]\ge 1-\delta\Big\}, \] and derives \[ C_{\text{EDEN}}(N,\epsilon,\delta) = k(\epsilon,\delta) \left( \frac{e^{\alpha r}e^\alpha}{\cosh(\alpha r)\cosh(\alpha)} \right)^{N-1} = O(\gamma^N), \] whereas the reference sequence network has \[ C_{\text{ref}}(N,\epsilon,\delta)= N\frac{\epsilon^2\delta}{\ln N} = O(N). \] The paper does not use the term “double-exponential capacity,” and its own discussion emphasizes that exponential in \(N\) is the relevant scaling for the number of distinct stored patterns in a binary \(N\)-neuron state space [2510.24965]. The contrast matters because a double-exponential capacity measure in the stable-polynomial sense is a pathology of a bound, not a superior memory law. In memory models, by contrast, exponential capacity is already the claimed achievement. ## 5. Measure-theoretic and geometric capacities: neighboring notions without double-exponential growth Morales’ measure-theoretic expansion exponent introduces yet another use of “capacity.” For a measurable map \(T\) on a metric space and a Borel probability measure \(\mu\), \[ E_\mu(T)= \sup\Bigl\{\lambda\in\mathbb{R}\ \Big|\ \exists\,\varepsilon>0: \mu(\{y\in B(x,\varepsilon):d(Tx,Ty)