--- title: 'Operator Capacity: Variational Approaches & Applications' url: https://www.emergentmind.com/topics/operator-capacity type: topic --- # Operator Capacity: Variational Approaches & Applications Searching arXiv for recent and foundational papers related to “operator capacity” and closely related usages across operator scaling and quantum information. Operator capacity is a polysemous technical term. In the arXiv literature, it denotes several distinct variational constructions attached to operators, currents, channels, or operator-induced dynamics. These include determinant-based capacity for completely positive operators in operator scaling, capacity inequalities for linear operators acting on real-stable polynomials, set capacities generated by complex Hessian or fractional dissipative operators, entanglement-sensitive capacities associated with local or unitary operators, and channel- or network-capacity notions in which the relevant object is an operator system, a linear operator channel, or a service operator [1511.03730] [1504.03519] [2106.00228] [2305.17636]. This suggests a shared structural theme—variational measurement of what an operator can preserve, create, regularize, or transmit—while also making clear that the term does not refer to a single universal invariant. ## 1. Taxonomic scope A common source of confusion is terminological. In one body of work, operator capacity is a determinant-infimum on positive matrices; in another, it is a Choquet-type set function; in quantum-information settings it may be the variance of a modular Hamiltonian or the distance of a unitary from local unitaries; and in communications it may refer to capacities controlled by network operators or to channels defined by random linear operators. The underlying optimization variables, admissible objects, and operational meanings are therefore context-dependent. | Domain | Basic object | Representative capacity | |---|---|---| | Operator scaling | Completely positive operator \(\mathcal T\) | \(\operatorname{cap}(\mathcal T)\) via determinants | | Stable-polynomial theory | Linear operator \(T\) on polynomials | \(\operatorname{cpc}_\alpha(P)=\inf_{x>0}P(x)/x^\alpha\) | | Nonlinear potential theory | Hessian or dissipative operator | \(\operatorname{cap}_{m,T}\), \(\operatorname{cap}_{m,T}^{\phi}\), \(C_{p,q}^{(\alpha)}\), \(\operatorname{Cap}_{V,\infty}\) | | Quantum operators | Local operator or unitary \(U\) | \(C_E(\rho_A)\), \(C(U)\), \(C_E(U)\) | | Quantum channels | Operator system or TRO-channel | \(C_0(\Phi)\) bounds, \(Q(\Phi)\), \(P(\Phi)\), \(C(\Phi)\) | | Communications networks | Linear operator channel or service operator | \(C\), \(C_{SS}\), leasing/sharing capacities | Representative formulations of these usages appear in operator scaling and real-stability theory [1511.03730] [1804.04351], in pluripotential and parabolic analysis [1504.03519] [1212.0744], in quantum-information treatments of local operators and unitaries [2106.00228] [2305.17636], and in communications and network-operations models [1108.4257] [1205.1196]. ## 2. Determinant-based capacity in operator scaling For a completely positive operator \(\mathcal T:\mathbb C^{n\times n}\to\mathbb C^{m\times m}\) with Kraus form \(\mathcal T(X)=\sum_{i=1}^N A_i X A_i^*\), capacity is defined by \[ \operatorname{cap}(\mathcal T)=\inf_{X\succ 0}\frac{\det(\mathcal T(X))^{1/m}}{\det(X)^{1/n}}. \] In the square case \(T:M_n(\mathbb C)\to M_n(\mathbb C)\), the normalization used in operator scaling is \[ \operatorname{cap}(T)=\inf\{\det(T(X)):X\succ0,\ \det(X)=1\}. \] This quantity is the central potential function in Gurvits’s operator-scaling algorithm, where alternating left and right scalings drive row- and column-marginals toward the identity [2508.02118] [1511.03730]. Several structural properties are fundamental. If \[ T_{B,C}(X)=B\,T(CXC^\dagger)\,B^\dagger \] with \(B,C\in GL_n(\mathbb C)\), then \[ \operatorname{cap}(T_{B,C})=|\det B|^2\,|\det C|^2\,\operatorname{cap}(T). \] If \(T\) is trace-preserving or \(T^*\) is trace-preserving, then \(\operatorname{cap}(T)\le 1\). The quantity \[ ds(T)=\|T(I)-I\|_2^2+\|T^*(I)-I\|_2^2 \] measures distance to double stochasticity; when \(ds(T)\) is small, one proves that \(T\) is rank-non-decreasing and that capacity is close to \(1\). In the algorithmic analysis of alternating normalization, the progress estimate \[ ds(T)>\epsilon' \;\Longrightarrow\; \operatorname{cap}(T_{\mathrm{next}})\ge \exp(\epsilon'/(36n))\,\operatorname{cap}(T) \] is the key monotonicity statement [1511.03730]. The determinant-infimum also has direct complexity-theoretic content. For the square completely positive operator \(T(X)=\sum_i A_i X A_i^\dagger\), one has \(\operatorname{cap}(T)>0\) if and only if \(T\) does not decrease rank on any psd \(X\), and Gurvits’s theorem identifies this with non-commutative invertibility of \(\sum x_i A_i\). This is the bridge from capacity to deterministic polynomial-time algorithms for the non-commutative singularity problem and non-commutative polynomial identity testing [1511.03730]. A later regularity theorem substantially strengthens continuity information. There exists an exponent \(\alpha\in(0,1)\), depending only on \((m,n)\), such that on every compact \(\mathcal K\subset\mathcal T_{n,m}\) there is \(C<\infty\) with \[ |\operatorname{cap}(\mathcal T)-\operatorname{cap}(\mathcal T')|\le C\|\mathcal T-\mathcal T'\|^\alpha \] for all \(\mathcal T,\mathcal T'\in\mathcal K\). The proof reduces capacity to an infimum of weighted sums of exponentials and then uses the Bennett–Bez–Buschenhenke–Cowling–Flock theorem together with Lipschitz dependence of the determinant coefficients \(d_j(\mathcal T)\) on \(\mathcal T\) [2508.02118]. A plausible implication is that perturbative analyses of operator scaling can be formulated with genuinely Hölder, rather than inverse-logarithmic, stability. ## 3. Capacity-preserving operators in real-stable polynomial theory A different notion of capacity appears for polynomials with nonnegative coefficients. If \[ P(x)=\sum_{\mu\in\mathbb N^n}P_\mu x^\mu \] and \(\alpha\in\mathbb R_+^n\), the \(\alpha\)-capacity is \[ \operatorname{cpc}_\alpha(P):=\inf_{x>0}\frac{P(x)}{x^\alpha}. \] This capacity is finite exactly when \(\alpha\) lies in the Newton polytope of \(P\), and it behaves naturally under products, scaling, diagonalization, and external fields [1804.04351]. The central question is how much a linear operator \(T\) that preserves real stability can shrink capacity. In the bounded-degree setting, if \(T:\mathbb R_+^\lambda[x]\to\mathbb R_+^\gamma[x]\) has symbol \[ \operatorname{Symb}^\lambda(T)(z,x):=T_x[(1+\langle x,z\rangle)^\lambda] \] with nonnegative coefficients and real stability in \(z\), then for every real-stable \(P\in\mathbb R_+^\lambda[x]\) and nonnegative vectors \(\alpha,\beta\), \[ \operatorname{cpc}_\beta(T(P)) \ge \Bigl[\alpha^\alpha(\lambda-\alpha)^{\lambda-\alpha}/\lambda^\lambda\Bigr]\, \operatorname{cpc}_{(\alpha,\beta)}(\operatorname{Symb}^\lambda(T))\, \operatorname{cpc}_\alpha(P). \] The unbounded-degree analogue replaces the prefactor by \([e^{-\alpha}\alpha^\alpha]\) and uses the transcendental symbol \(\operatorname{Symb}^\infty(T)\). In both cases the bounds are tight [1804.04351]. This framework unifies several lower-bound arguments in combinatorics. By choosing \(P\) to be a permanent or matching polynomial and \(T\) to be a differentiation operator extracting the relevant coefficient, one recovers the Van der Waerden bound, Schrijver’s inequality, and Csikvári’s lower bound settling Friedland’s lower matching conjecture. The significance of the capacity-preservation theorem is therefore methodological: it converts coefficient extraction into a symbol-capacity computation, reducing many counting lower bounds to a common analytic template [1804.04351]. ## 4. Capacities induced by analytic operators In pluripotential theory, Dhouib and Elkhadhra introduced the relative \(m\)-capacity associated to a closed \(m\)-positive current \(T\) of bidimension \((p,p)\), \(1\le p0\), no such solution exists. A density theorem shows that \(C_c^2(\Omega\setminus K)\) is dense in \(C^2(\overline\Omega)\) whenever \(\operatorname{Cap}_{V,\infty}(K)=0\) [1812.04061]. ## 5. Entanglement-sensitive capacities of operators In the study of local operator excitations, the capacity of entanglement is defined for a reduced density matrix \(\rho_A\) by the variance of the modular Hamiltonian \(\mathcal K_A=-\ln\rho_A\): \[ C_E(\rho_A)=\langle \mathcal K_A^2\rangle-\langle \mathcal K_A\rangle^2 = \operatorname{tr}\big[\rho_A(-\ln\rho_A)^2\big] - \big[\operatorname{tr}(\rho_A\ln\rho_A)\big]^2. \] Equivalently, \[ C_E(\rho_A) = \lim_{m\to 1}m^2\partial_m^2\ln[\operatorname{tr}\rho_A^m]. \] For locally excited states, one studies the excess quantity \[ \Delta C_E=\lim_{m\to1}m^2\partial_m^2[(1-m)\Delta S_A^{(m)}]. \] In free, massless fermionic theory and free \(U(N)\) Yang–Mills theory in four spacetime dimensions, \(\Delta C_E(t)\) vanishes for \(t0\) if and only if the operator Schmidt rank of \(U\) exceeds \(1\), and \(C_E(U)\le \sqrt{1-1/r}\) for Schmidt rank \(r\). For generalized control operators \[ U(\alpha_i\otimes\beta)=\alpha_i\otimes V_i\beta, \] maximal entanglement generation occurs if and only if there is some \(\beta\) such that the vectors \(\{V_i\beta\}\) are mutually orthogonal. The capacity \(C(U)\) is invariant under left and right multiplication by local unitaries, satisfies \(C(U)=C(U^\dagger)\), is subadditive under composition, and is continuous in the unitary metric [2305.17636]. ## 6. Operator systems, channel capacities, and superactivation Quantum channels provide another setting in which operator-structured capacity questions arise. Given a channel \(\Phi:M_n\to M_k\) with Kraus operators \(\{A_i\}\), its non-commutative confusability graph is the operator system \[ S_\Phi=\operatorname{span}\{A_i^*A_j:1\le i,j\le m\}\subset M_n. \] The quantum complexity of an operator system \(S\subseteq M_n\) is \[ \gamma(S)=\min\{k\in\mathbb N:\exists\ \text{quantum channel }\Phi:M_n\to M_k\ \text{with }S_\Phi=S\}, \] and the sub-complexity is \[ \beta(S)=\min\{k:\exists\ \text{quantum channel }\Phi:M_n\to M_k\ \text{with }S_\Phi\subseteq S\}. \] These quantities bound zero-error communication: \[ \alpha(\Phi)=\alpha(S_\Phi)\le \Theta(\Phi)\le \beta(S_\Phi)\le \gamma(S_\Phi). \] For suitable families, \(\beta\) and \(\gamma\) outperform the previously known quantum Lovász-\(\vartheta\) upper bound [1710.06456]. For finite-dimensional TRO-channels, \[ \Phi=\bigoplus_i(\operatorname{id}_{n_i}\otimes \operatorname{tr}_{m_i}), \] one has single-letter formulas \[ Q^{(1)}(\Phi)=P^{(1)}(\Phi)=\log(\max_i n_i),\qquad \chi(\Phi)=\log\sum_i n_i, \] and therefore \[ Q(\Phi)=P(\Phi)=\log\max_i n_i,\qquad C(\Phi)=\log\sum_i n_i. \] If \(f\) is a symbol perturbation, the perturbed channel \(\Phi_f\) satisfies \[ Q(\Phi)\le Q(\Phi_f)\le Q(\Phi)+\tau(f\log f), \] with analogous estimates for \(P(\Phi)\), \(C(\Phi)\), and the strong converse capacities. The proofs use operator-space and complex-interpolation methods on the Stinespring TRO [1609.08594]. Zero-error capacity also interacts sharply with non-commutative operator graphs. Shirokov’s graph \(G(\theta)\subset \mathrm{Mat}_4(\mathbb C)\), generated by matrices \(I,X,Y,Z\), yields channels \(\Phi_\theta\) whose one-shot zero-error capacity vanishes when \(\theta\neq \pm 1\), and in fact \[ C_0^{(n)}(\Phi_\theta)=0 \quad \text{for all finite }n. \] However, pairing inverse parameters produces superactivation: \[ C_0^{(1)}(\Phi_\theta\otimes \Phi_{\theta^{-1}})\ge \log 2>0. \] At \(\theta=\pm 1\), the graph becomes abelian and the associated algebra degenerates to the direct sum of four one-dimensional irreducible representations of the Klein group, changing the zero-error behavior qualitatively [1512.02096]. A plausible implication is that algebraic non-commutativity can suppress single-channel zero-error transmission even when tensor products restore a commutative coding structure. ## 7. Linear operator channels and network-operator capacity management In finite-field network coding, a linear operator channel (LOC) is \[ Y=XH, \] with \(X\in\mathbb F_q^{T\times M}\), \(H\in\mathbb F_q^{M\times N}\), and \(Y\in\mathbb F_q^{T\times N}\). Its Shannon capacity is \[ C(H,T)=\max_{p_X}I(X;Y), \] while the subspace-coding capacity \(C_{SS}\) is obtained by encoding and decoding through column spaces. The central structural question is when \(C=C_{SS}\). Sufficient and necessary conditions are formulated through unique subspace degradation, row-space symmetry, and Markov constraints such as \[ \mathrm{Row}(X)\to \mathrm{rk}(X)\to \mathrm{rk}(Y)\to \mathrm{Row}(Y). \] If the LOC is degraded, then \(C=C_{SS}\); if the relevant Markov conditions fail, \(C_{SS}\) can be strictly less than \(C\) [1108.4257]. The broadcast analogue is the constant-dimension multiplicative linear operator broadcast channel (CMLOBC), whose input alphabet is the Grassmannian \(\mathcal P(\mathbb F_q^m,l)\) and whose output law is determined by rank-deficiency distributions \(\epsilon^{(1)}\) and \(\epsilon^{(2)}\). Degradation holds if and only if \[ \sum_{j=0}^i\epsilon_j^{(1)}\le \sum_{j=0}^i\epsilon_j^{(2)} \quad\text{for every } i=0,1,\dots,l. \] Even in the degraded case, time sharing need not exhaust the boundary of the capacity region. In the smallest nontrivial example \((q=2,m=3,l=2)\), the superposition boundary is strictly concave if \[ \epsilon_1^{(1)}\epsilon_2^{(2)}>\epsilon_1^{(2)}\epsilon_2^{(1)}, \] strictly convex if the inequality is reversed, and linear when equality holds [1012.0744]. In engineering usage, the phrase can also refer to capacity controlled by communication-service operators. In a femtocell market model, the macrocell operator chooses a wholesale price \(p_M\) and leasing amount \(K\), while the femtocell operator chooses a retail price \(p_F\) and actual lease \(L\). The equilibrium exhibits two regimes separated by a unique threshold \(S_{th}\approx 4.77\): when \(S\le S_{th}\), the macrocell leases positive spectrum \(K^*(S)>0\); when \(S>S_{th}\), it withholds all spectrum, so \(K^*(S)=0\) [1205.1196]. This directly formalizes the tradeoff between efficiency gains from femtocells and competitive cannibalization. A stochastic capacity-sharing model for public-transit operators treats first-stage committed pooled capacity \(b_f\) and second-stage recourse flows under disruption scenarios. Applied to the Randstad network, the grand coalition of four operators reduces the baseline expected passenger-minutes cost from \(193.95\) million p-min to \(65.66\) million p-min, a \(66\%\) improvement in overall network performance over not having any risk-pooling contract in place. Cooperative-game allocations such as the Shapley value, nucleolus, and \(T\)-value then quantify bargaining power as a function of contributed capacity and network structure [2006.14518]. Across these communications examples, operator capacity is operational rather than purely variational: it measures achievable rates, region boundaries, or allocable infrastructure resources. This contrasts with determinant-based and potential-theoretic usages, but preserves the same formal pattern of optimization under structural constraints. Source: https://www.emergentmind.com/topics/operator-capacity