betaFeedback

WorksPhysics and Astronomy

Witnessing Magic (and other Quantum Resources)

Abstract

Motivated by measurements of quantum magic, we study the sampling cost of resource robustness —a quantifier for a broad class of quantum resources— by bounding the variance of optimal robustness witnesses. We first lower bound the samples required to estimate a state's resource robustness using Cramér-Rao bounds. We find that the number of measurements scales at least quadratically in the robustness for highly mixed states, and at least linearly in the generalized robustness for broad families of pure states, making it difficult to measure either robustness to low absolute error when the states are highly resourceful. We then derive two complementary bounds in terms of the geometry of witnesses, one of which is tight for a class of resource theories including coherence and bipartite entanglement. An additional hurdle occurs for tracking the onset of resource robustness along state-space trajectories: the integrated witness variance along a path is at least the squared resource gain over Fisher length, indicating that a state evolution with exponential resource gain necessarily goes through states for which the robustness cannot be measured efficiently. This hurdle does not occur when estimating robustness to fixed relative, rather than absolute, error. These results suggest that the ideal witness-based resource estimation task, for general resource theories, is the estimation of generalized robustness to low relative error on highly resourceful pure states.

Introduction

Quantum computers offer a route to communication, encryption, and materials simulations beyond the reach of classical computation. Understanding our progress towards this goal, from noisy intermediate-scale devices to fault-tolerant architectures, calls for benchmarks that quantify the resources available for quantum information processing. Quantum resource theories provide a unified framework for this task, encompassing magic, entanglement, coherence, non-Gaussianity, and other properties of interest [1, 2]. Magic is particularly relevant to quantum advantage, as its quantification directly describes the cost of simulating states classically under stabilizer techniques [3, 4, 5]. Measuring these resources can inform the assessment of quantum capabilities, and guide the design of the next generation of quantum devices.

Protocols for measuring magic have been demonstrated on quantum processors [6, 7, 8], while tensor-network methods enable its calculation in weakly-entangled systems [9]. Nevertheless, hardness results establish substantial barriers to estimating magic in general [10, 11]. When analyzing the utility of a resource measure, a core question is under what conditions it is accessible within the available computational and experimental resources. Robustness measures are natural candidates because they quantify broad classes of resources and admit a formulation as the maximal expectation value of a resource witness, making them directly measurable. In particular, robustness of magic bounds the non-Clifford resources required for state preparation [3].

This witness formulation separates two potential challenges to resource estimation: finding a suitable witness, and performing enough measurements to estimate its expectation value to bounded error. Finding or even verifying a magic witness can be computationally difficult [12], though it may be easier for other resource theories. Here we focus on the second challenge, and address the sampling cost that remains once an optimal witness is known.

Given an optimal witness WW, the required number of shots to estimate a state’s robustness to fixed error ϵ\epsilon goes as Var⁡(W)/ϵ2\operatorname{Var}(W)/\epsilon^2, so the shot complexity of resource robustness is inherently related to the variance of the optimal witness. A high-variance optimal witness can require prohibitively many shots to estimate, making resource estimation inefficient. Witness variance is then a necessary consideration when determining the utility of resource measures towards benchmarking.

Here, we derive bounds on witness variance for standard and generalized robustness for general quantum resource theories, with magic as principal motivation. We provide background on resource theories, Cramér-Rao bounds, and Danskin’s theorem in § 2. Building on quantum estimation approaches to entanglement [13], we derive Cramér-Rao bounds for mixed states and for the generalized robustness of pure states in Section III. In Section IV we study the geometric nature of pure-state witness variance. Finally, we relate the integrated witness variance along a state-space path to its total robustness gain in § 2.4, and derive a no-go theorem for efficiently tracking large changes in robustness. We discuss the applicability of the results to benchmarking and directions for future research in Section VI.

Background

Resource Theories

Resource theories are an approach to quantifying the capabilities of quantum states and devices with a broad range of applications [1]. They divide state space into resourceful states and non-resourceful free states, and, roughly, describe the resourcefulness of a state in terms of its distance from the set of free states.

Robustness measures

We consider robustness measures to quantify the resources in a quantum system [1, 5, 14, 15, 16, 17]. Robustness measures quantify, e.g., resource-enhanced advantages in channel discrimination tasks [18, 19], and the rates at which resourceful states can be distilled [2, 20]. Given a convex set F\mathcal{F} of free states, the resource robustness is defined by

R(ρ):=1+2⋅min⁡γ∈F{s≥0  |  ρ+sγ1+s∈F}.(1)\mathcal{R}(\rho) := 1 + 2 \cdot\min_{\gamma\in\mathcal{F}} \left\{s \ge0 \;\middle|\; \frac{\rho+ s\gamma}{1+s} \in\mathcal{F}\right\}. \tag*{(1)}

Intuitively, R(ρ)\mathcal{R}(\rho) is a measure of the distance of a given state ρ\rho from F\mathcal{F}, or, equivalently, it quantifies how much mixing with an arbitrary non-resourceful state γ\gamma a resource state ρ\rho can withstand while still being resourceful. Sufficient mixing eventually removes all resources from a state. It follows that states ρ\rho contained in the free set F\mathcal{F} have R(ρ)=1\mathcal{R}(\rho)=1, while states outside of F\mathcal{F} have R(ρ)>1\mathcal{R}(\rho)>1. If the set of free states is convex, the robustness cannot increase under operations that transform free states into free states; i.e., they are monotones under free operations [1]. For brevity, we will write R\mathcal{R} instead of R(ρ)\mathcal{R}(\rho) when the state is unambiguous or irrelevant.

The definition of the free set F\mathcal{F} characterizes the quantum resource of interest. For instance, if FSep\mathcal{F}_{\mathrm{Sep}} is defined as the set of separable states, the corresponding REnt\mathcal{R}_{\mathrm{Ent}} quantifies a state’s entanglement. If FIncoh\mathcal{F}_{\mathrm{Incoh}} includes all states without coherence in a given basis, then (a generalization of) RCoh\mathcal{R}_{\mathrm{Coh}} quantifies coherence in said basis [17]. And most commonly, if FStab\mathcal{F}_{\mathrm{Stab}} is the set of stabilizer states, then RMag\mathcal{R}_{\mathrm{Mag}} quantifies a state’s magic.

An operator WW is called a resource witness for the free set F\mathcal{F} if

∣Tr⁡(Wϕ)∣≤1∀ϕ∈F.(2)\left|\operatorname{Tr}(W\phi)\right| \le1 \quad\forall\phi\in\mathcal{F}. \tag*{(2)}

The primal definition of robustness admits a dual definition in terms of an optimization problem [3, 5, 21, 22]:

R(ρ)=max⁡W†=W{Tr⁡(Wρ)  |  ∣Tr⁡(Wϕ)∣≤1 ∀ϕ∈F}.(3)\mathcal{R}(\rho)=\max_{W^\dagger=W}\left\{\operatorname{Tr}(W\rho)\;\middle|\;\left|\operatorname{Tr}(W\phi)\right|\le1\ \forall\phi\in\mathcal{F}\right\}. \tag*{(3)}

This definition is operationally useful, as it highlights that R\mathcal{R} is measurable, as the expectation value of any optimal witness maximizing the right-hand side of the equation. While the optimal witness set in general depends on state ρ\rho, Danskin’s theorem (discussed in § 2 C) guarantees that in every direction xx in state space, there exists an optimal WW such that ∂x⟨W⟩=Tr⁡(W∂xρ)\partial_x\langle W\rangle=\operatorname{Tr}(W\partial_x\rho) to first order. For this WW, small changes in the expectation value entirely reflect changes in ρ\rho, rather than changes in the witness, so WW acts as an estimator of R\mathcal{R} which is unbiased in xx. One could alternatively probe robustness

Robustness measures

with any witness, with the resulting expectation value acting as a lower bound on robustness. A generalization of robustness, called generalized robustness, is defined by restricting the set of witnesses to positive semidefinite operators [5]

R+(ρ)=max⁡W⪰0{Tr⁡(Wρ)  |  Tr⁡(Wϕ)≤1 ∀ϕ∈F}.(4)\mathcal{R}^{+}(\rho)=\max_{W\succeq0}\left\{\operatorname{Tr}(W\rho)\;\middle|\;\operatorname{Tr}(W\phi)\le1\ \forall\phi\in\mathcal{F}\right\}. \tag*{(4)}

in which case the optimal witness can be found via a semidefinite optimization problem [23]. Since the set of semidefinite witnesses is a subset of the set of Hermitian witnesses, it holds that R+≤R\mathcal{R}^{+}\le\mathcal{R}. Additionally, all robustness results derived here also hold for generalized robustness, since they come from properties of the witness set. The primal definition of generalized robustness is equivalent to Eq. (1), but γ\gamma is allowed to be an arbitrary state and the factor of 2 is dropped [5]. In a sense, it quantifies robustness against even resourceful noise.

Free-state fidelity

The stabilizer fidelity measure [4, 5], traditionally used in the context of magic, can be extended to a free-state fidelity defined for arbitrary resource theories. The free-state fidelity of a pure state ∣ψ⟩\lvert\psi\rangle is then defined as

F(ψ):=max⁡ϕ∈F⟨ψ∣ϕ⟩.(5)F(\psi):=\max_{\phi\in\mathcal{F}}\langle\psi\lvert\phi\rangle. \tag*{(5)}

where we use the pure state notation ψ:=∣ψ⟩⟨ψ∣\psi:=\lvert\psi\rangle\langle\psi\rvert. The smaller the free-state fidelity F(ψ)F(\psi), the higher the resource in ∣ψ⟩\lvert\psi\rangle. To see how to relate FF to R\mathcal{R}, notice that Wψ:=ψ/F(ψ)W_\psi:=\psi/F(\psi) is a feasible, typically sub-optimal, witness for the generalized robustness [5]. Since ⟨Wψ⟩ψ=1/F(ψ)\langle W_\psi\rangle_\psi=1/F(\psi), the hierarchy of resource measures is 1/F≤R+≤R1/F\le\mathcal{R}^{+}\le\mathcal{R}, with equality of the first two terms when WψW_\psi is an optimal witness.

Cramér-Rao Bound

The Cramér-Rao (CR) bound lower-bounds the variance of estimators of parameters or functions. Traditionally, the bound is applied to unbiased estimators of unknown parameters determining a state, but it applies just as well to functions of the state. Let R=⟨M⟩R=\langle M\rangle be some function of a state which can be written as the expectation value of some observable MM, where MM has no first-order dependence on ρ\rho. The quantum CR bound tells us that for any parameter xx of the state,

Var⁡(M)≥∣∂xR∣2Ix,(6)\operatorname{Var}(M)\ge\frac{\left|\partial_x R\right|^2}{I_x}, \tag*{(6)}

for variance Var⁡(A)=⟨δA2⟩\operatorname{Var}(A)=\langle\delta A^2\rangle with δA:=A−⟨A⟩\delta A:=A-\langle A\rangle [24, 25]. When Lx\mathcal{L}_x is a log-derivative of the state with respect to xx, Ix:=Var⁡(Lx)I_x:=\operatorname{Var}(\mathcal{L}_x) is the quantum Fisher Information about xx, which quantifies the rate at which the system’s state changes with changes in xx. Usually one takes a symmetric log-derivative satisfying ∂xρ=12{Lx,ρ}\partial_x\rho= \frac{1}{2}\{L_x,\rho\}, but the bound also holds for left-log-derivative satisfying ∂xρ=LxLρ\partial_x\rho= L_x^{L}\rho, where in this case IxL=⟨LxL†LxL⟩I_x^{L} = \langle L_x^{L\dagger}L_x^{L}\rangle. The bound is most commonly used for the case R=x^R = \hat{x}, an unbiased estimator of parameter xx, so that ∂xR=1\partial_xR = 1, but holds in any case where we can calculate (or lower bound) ∂xR\partial_xR. The CR bound is straightforward to derive:

∂x⟨M⟩=Tr⁡(M∂xρ)=12Tr⁡(M{Lx,ρ})=12⟨{M,Lx}⟩=12⟨{δM,δLx}⟩,(7)\partial_x\langle M\rangle= \operatorname{Tr}(M\partial_x\rho) = \frac{1}{2}\operatorname{Tr}(M\{L_x,\rho\}) = \frac{1}{2}\langle\{M,L_x\}\rangle= \frac{1}{2}\langle\{\delta M,\delta L_x\}\rangle, \tag*{(7)}

where ⟨Lx⟩=Tr⁡(∂xρ)=0\langle L_x\rangle= \operatorname{Tr}(\partial_x\rho) = 0. The Cauchy–Schwarz inequality guarantees that

∣∂xR∣2=∣12⟨{δM,δLx}⟩∣2=∣Cov⁡(M,Lx)∣2≤Var⁡(M)Var⁡(Lx),(8)|\partial_xR|^2 = \left|\frac{1}{2}\langle\{\delta M,\delta L_x\}\rangle\right|^2 = |\operatorname{Cov}(M,L_x)|^2 \le\operatorname{Var}(M)\operatorname{Var}(L_x), \tag*{(8)}

for covariance Cov⁡(A,B):=⟨{A,B}⟩/2−⟨A⟩⟨B⟩=⟨{δA,δB}⟩/2\operatorname{Cov}(A,B) := \langle\{A,B\}\rangle/2 - \langle A\rangle\langle B\rangle= \langle\{\delta A,\delta B\}\rangle/2. This can be rearranged to get the lower bound on Var⁡(M)\operatorname{Var}(M) in Eq. (6).

When multiple MM satisfy R=⟨M⟩R = \langle M\rangle, the CR bound applies to them all. Moreover, the CR bound is saturable by picking MM such that δM∝Lx\delta M \propto L_x. There is guaranteed to be a satisfying MM which has no first-order dependence on xx; however it can potentially have dependence on other parameters or on the state itself.

Danskin's Theorem

Danskin’s theorem gives a formula for the derivative of a function which is a maximum over a compact set, such as resource robustness. Let Z⊂RmZ \subset\mathbb{R}^{m} be a compact subset, g:Rm×Z→Rg : \mathbb{R}^{m} \times Z \to\mathbb{R} continuous, and g(⋅,z):Rm→Rg(\cdot,z) : \mathbb{R}^{m} \to\mathbb{R} convex for every z∈Zz \in Z. Define the function f:Rm→Rf : \mathbb{R}^{m} \to\mathbb{R}

f(x)=max⁡z∈Zg(x,z),(9)f(x) = \max_{z\in Z} g(x,z), \tag*{(9)}

and ZxZ_x the set of maximizing points

Zx={z∣g(x,z)=f(x)}.(10)Z_x = \{z \mid g(x,z) = f(x)\}. \tag*{(10)}

Then Danskin’s theorem [26] says that

  1. If ∣Zx∣=1|Z_x| = 1 so that the optimizing z∗z^* is unique, then the derivative of ff is well-defined as the derivative of gg in its first argument. That is,

∇f(x)=∇xg(x,z∗).(11)\nabla f(x) = \nabla_x g(x,z^*). \tag*{(11)}
  1. If ∣Zx∣>1|Z_x| > 1, then ff is differentiable iff ∇xg(x,z∗)\nabla_xg(x,z^*) is the same for every z∗∈Zxz^* \in Z_x. When ff is not differentiable, it instead has a one-sided directional derivative in each direction yy, given by

∂yf(x)=max⁡z∈Zx∂yg(x,z),(12)\partial_y f(x) = \max_{z\in Z_x}\partial_y g(x,z), \tag*{(12)}

where ∂yg(x,z)\partial_y g(x,z) is the directional derivative of g(⋅,z)g(\cdot,z) in the direction yy.

In the contexts used here, ff will be either standard or generalized resource robustness, both of which can be expressed in the form necessary for Danskin’s theorem, either by Hermitian matrices in Cn2×2n2\mathbb{C}^{n^2\times2n^2}, or in terms of a vector of real parameters in R4n2\mathbb{R}^{4n^2}, for example their real and imaginary coefficients. In the case of robustness Z≃WZ \simeq\mathcal{W}, i.e., ZZ is (isomorphic to) the convex set of witnesses.

Danskin's Theorem for Robustness Witnesses

In the context of robustness used in this paper, we take ρx\rho_x defined in terms of a single parameter xx and W\mathcal{W} defined in terms of its coefficients, so that g(x,W)=Tr⁡(ρxW)g(x,W) = \operatorname{Tr}(\rho_xW) and f(x)=R(ρx)f(x) = \mathcal{R}(\rho_x), for either standard or generalized robustness Eq. (3) or Eq. (4). Let Wx⊆W\mathcal{W}_x \subseteq\mathcal{W} be the set of W∈WW \in\mathcal{W} maximizing Tr⁡(Wρx)\operatorname{Tr}(W\rho_x), i.e., the set of optimal witnesses at xx. Then Danskin’s theorem guarantees that when R(ρx)\mathcal{R}(\rho_x) is a differentiable function of xx, for example when ∣Wx∣=1|\mathcal{W}_x| = 1,

ddxR(ρx)=Tr⁡(Wx∂xρx)∀Wx∈Wx,(13)\frac{d}{dx}\mathcal{R}(\rho_x) = \operatorname{Tr}(W_x\partial_x\rho_x) \qquad\forall W_x \in\mathcal{W}_x, \tag*{(13)}

i.e., the derivative can be written equivalently in terms of any optimal witness at xx. While Danskin’s theorem gives other results when R(ρx)\mathcal{R}(\rho_x) is non-differentiable, we will only use the differentiable result here.

Compactness of W\mathcal{W}, required for Danskin’s theorem, is guaranteed whenever robustness has a finite upper bound Rmax⁡\mathcal{R}_{\max} for the relevant finite-dimensional state space, as in this case we can bound ∥W∥≤Rmax⁡ ∀W∈W\lVert W\rVert\le\mathcal{R}_{\max}\ \forall W \in\mathcal{W}. Let λmax⁡,λmin⁡\lambda_{\max},\lambda_{\min} be the largest and smallest eigenvalues of WW, with eigenstates ψmax⁡,ψmin⁡\psi_{\max},\psi_{\min}. Then λmax⁡=Tr⁡(Wψmax⁡)≤R(ψmax⁡)≤Rmax⁡\lambda_{\max} = \operatorname{Tr}(W\psi_{\max}) \le\mathcal{R}(\psi_{\max}) \le\mathcal{R}_{\max}. For generalized robustness, W≥0W \ge0, so ∥W∥=λmax⁡\lVert W\rVert= \lambda_{\max}. For standard robustness, −W-W is also a valid witness, so −λmin⁡≤Tr⁡(−Wψmin⁡)≤Rmax⁡-\lambda_{\min} \le\operatorname{Tr}(-W\psi_{\min}) \le\mathcal{R}_{\max}. In either case, ∥W∥=max⁡{λmax⁡,−λmin⁡}≤Rmax⁡\lVert W\rVert= \max\{\lambda_{\max},-\lambda_{\min}\} \le\mathcal{R}_{\max}. W\mathcal{W} is also guaranteed compact when free states span the full operator space, or, for generalized robustness, whenever F\mathcal{F} contains any full rank state (such as the maximally mixed state).

Resource Cramér-Rao Bounds

Say our goal is to estimate some quantity, in this case resource robustness R\mathcal{R}, which can be written as the expectation value of an observable R=⟨W⟩\mathcal{R} = \langle W\rangle. For an unbiased estimate R^\hat{\mathcal{R}} resulting from averaging mm independent measurements of WW, the mean squared error ϵ2\epsilon^2 of the estimate is related to the variance of R^\hat{\mathcal{R}} [27] by

ϵ2=E[(R^−R)2]=Var⁡(R^)=Var⁡(W)m.(14)\epsilon^2 = \mathbb{E}[(\hat{\mathcal{R}}-\mathcal{R})^2] = \operatorname{Var}(\hat{\mathcal{R}}) = \frac{\operatorname{Var}(W)}{m}. \tag*{(14)}

To estimate R\mathcal{R} to target error ϵ\epsilon then requires m=Var⁡(W)/ϵ2m = \operatorname{Var}(W)/\epsilon^2 samples. When Var⁡(W)∼exp⁡(n)\operatorname{Var}(W) \sim\exp(n), we require m∼exp⁡(n)m \sim\exp(n) samples in order to accurately estimate the true value R\mathcal{R}, making its measurement inefficient. If our goal is instead to estimate R\mathcal{R} to low multiplicative error ϵ′=ϵ/R\epsilon'=\epsilon/\mathcal{R}, the required number of samples goes as m=(Var⁡(W)/R2)/ϵ′2m=(\operatorname{Var}(W)/\mathcal{R}^{2})/\epsilon'^{2}. That is, it now depends on the relative variance Var⁡(W)/R2\operatorname{Var}(W)/\mathcal{R}^{2}. By using Cramér-Rao (CR) bounds to lower bound Var⁡(W)\operatorname{Var}(W), we can then determine cases where Var⁡(W)\operatorname{Var}(W) or Var⁡(W)/R2\operatorname{Var}(W)/\mathcal{R}^{2} are large, making R\mathcal{R} inefficient to estimate to low absolute or multiplicative error from standard measurements, respectively.

Resource Robustness measures are a natural context for CR bounds, as their dual definition allows us to write them as the expectation value R=⟨W⟩\mathcal{R}=\langle W\rangle for a (generally non-unique) optimal resource witness WW. We will mainly consider perturbative CR bounds constructed as follows. We will pick some initial state ρ\rho with no parameter dependence, and perturb it to ρc:=ρ+cδρ\rho_c:=\rho+c\delta\rho, where Tr⁡(δρ)=0\operatorname{Tr}(\delta\rho)=0, then calculate the CR bound in the limit c→0c\to0. While ∂cR(ρc)∣c=0≠∂cRW(ρc)∣c=0\left.\partial_c\mathcal{R}(\rho_c)\right|_{c=0}\ne\left.\partial_c\mathcal{R}_W(\rho_c)\right|_{c=0} in general, we have RW(ρc)∣c=0=R(ρ)\left.\mathcal{R}_W(\rho_c)\right|_{c=0}=\mathcal{R}(\rho), and calculating ∂cRW(ρc)∣c=0\left.\partial_c\mathcal{R}_W(\rho_c)\right|_{c=0} will turn out to be sufficient to place a CR bound on the variance of WW in the state ρ\rho.

As an alternative approach, Danskin’s theorem § 2 C guarantees that ∂cR(ρc)∣c=0=Tr⁡(W∂cρc)∣c=0\left.\partial_c\mathcal{R}(\rho_c)\right|_{c=0}=\left.\operatorname{Tr}(W\partial_c\rho_c)\right|_{c=0} for any optimal WW, as long as R(ρc)\mathcal{R}(\rho_c) is differentiable at c=0c=0. The differentiability condition occurs, for example, when WW is the unique optimal witness for ρ\rho. The bounds in this section are framed in terms of ∂cRW\partial_c\mathcal{R}_W for generality, but hold equally well for ∂cR\partial_c\mathcal{R} when R\mathcal{R} is differentiable. In the latter case, they lower bound the variance of any unbiased estimator of R\mathcal{R}, not just that of an optimal witness.

In this section, we will use this technique to bound the variance of arbitrary optimal witnesses in two cases: robustness and generalized robustness for mixed states, and generalized robustness of pure states. In Appendix A 2, we also bound the variance of observables which can be used to estimate moments, including stabilizer, purity, and entanglement moments. The bounds presented in this section are not necessarily saturable, but are instead intended to be general and interpretable. Tighter CR bounds can be derived in terms of different state perturbations, but incur dependence on the specific WW or on diagonalization of ρ\rho, which limits their utility.

Robustness of Mixed States

The first CR bound we find is on the variance of an optimal witness of robustness or generalized robustness of sufficiently mixed states. Specifically, this holds for mixed states for which B(Vρ)∩F≠∅\mathcal{B}(\mathcal{V}_{\rho})\cap\mathcal{F}\ne\varnothing, where Vρ\mathcal{V}_{\rho} is the subspace spanned by the eigenvectors of ρ\rho with nonzero eigenvalues. This class of states includes full-rank states as a subset.

To place bounds on the variance of a robustness estimator at ρ\rho, we consider a derivative of the perturbed state ρc=ρ+c(ρ−ϕ)\rho_c=\rho+c(\rho-\phi) in the limit c→0c\to0, where ϕ∈B(Vρ)∩F≠∅\phi\in\mathcal{B}(\mathcal{V}_{\rho})\cap\mathcal{F}\ne\varnothing can be any free state composed of eigenvectors of ρ\rho, and consider the fixed witness “robustness” RW(ρc)\mathcal{R}_W(\rho_c).

Let R\mathcal{R} denote either standard or generalized robustness, and let WW be an optimal witness so that R(ρ)=Tr⁡(Wρ)=RW(ρ)\mathcal{R}(\rho)=\operatorname{Tr}(W\rho)=\mathcal{R}_W(\rho) and ∣Tr⁡(Wϕ′)∣≤1|\operatorname{Tr}(W\phi')|\le1 for all free states ϕ′\phi'. Then, it holds that

∣∂cRW(ρc)∣c=0∣=∣Tr⁡(W∂cρc)∣c=0=∣R(ρ)−Tr⁡(Wϕ)∣≥∣R(ρ)−1∣,(15)\left|\left.\partial_c\mathcal{R}_W(\rho_c)\right|_{c=0}\right|=\left|\operatorname{Tr}(W\partial_c\rho_c)\right|_{c=0}=\left|\mathcal{R}(\rho)-\operatorname{Tr}(W\phi)\right| \ge\left|\mathcal{R}(\rho)-1\right|, \tag*{(15)}

since R(ρ)−Tr⁡(Wϕ)≥R(ρ)−1≥0\mathcal{R}(\rho)-\operatorname{Tr}(W\phi)\ge\mathcal{R}(\rho)-1\ge0.

In theory, one could find a symmetric log-derivative generating ρc\rho_c, but its form would depend explicitly on the diagonalization of ρ\rho and not be particularly enlightening. Instead, we consider ρc\rho_c as being generated by the left log-derivative LcL=(ρ−ϕ)ρ−1\mathcal{L}^{L}_{c}=(\rho-\phi)\rho^{-1}, where ρ−1\rho^{-1} is the inverse of ρ\rho within Vρ\mathcal{V}_{\rho}, which is always well-defined. Then the left Fisher information in the c→0c\to0 limit is

IcL∣c=0=Var⁡(LcL)∣c=0=⟨LcL†LcL⟩∣c=0=Tr⁡(ρ−1(ρ−ϕ)2)=Tr⁡(ρ−1ϕ2)−1.(16)\left.\mathcal{I}^{L}_{c}\right|_{c=0}=\left.\operatorname{Var}(\mathcal{L}^{L}_{c})\right|_{c=0}=\left.\langle\mathcal{L}^{L\dagger}_{c}\mathcal{L}^{L}_{c}\rangle\right|_{c=0}=\operatorname{Tr}\left(\rho^{-1}(\rho-\phi)^{2}\right) =\operatorname{Tr}\left(\rho^{-1}\phi^{2}\right)-1. \tag*{(16)}

Combining these terms with the CR formula in Eq. (6), we obtain:

Theorem 1. For any state ρ\rho and any optimal witness WW of its standard or generalized robustness R\mathcal{R}, the variance of WW obeys

Var⁡(W)≥(R−1)2Tr⁡(ρ−1ϕ2)−1(17)\operatorname{Var}(W)\ge\frac{(\mathcal{R}-1)^{2}}{\operatorname{Tr}(\rho^{-1}\phi^{2})-1} \tag*{(17)}

for every free state ϕ\phi within the span of the eigenvectors of ρ\rho with nonzero eigenvalues.

The denominator can also be written as ∥ρ−1/2(ρ−ϕ)∥22\lVert\rho^{-1/2}(\rho-\phi)\rVert_{2}^{2}, highlighting that it is always non-negative, and only zero when ρ=ϕ\rho=\phi. When multiple valid ϕ\phi exist, the tightest bound comes from picking the ϕ\phi which minimizes Tr⁡(ρ−1ϕ2)\operatorname{Tr}(\rho^{-1}\phi^{2}), which will tend to be one with more support on the eigenspaces of ρ\rho corresponding to larger eigenvalues.

As the bound’s explicitly dependence on R\mathcal{R} is quadratic, the lower bound on the relative variance, Var⁡(W)/R2\operatorname{Var}(W)/\mathcal{R}^{2}, has no direct dependence on R\mathcal{R} and only on the overlap term Tr⁡(ρ−1ϕ2)\operatorname{Tr}(\rho^{-1}\phi^{2}). Additionally, since any resourceful state has R>1\mathcal{R}>1, this bound shows that the variance cannot be zero except in the limit where the denominator approaches infinity, which is not possible when ϕ∈B(Vρ)\phi\in\mathcal{B}(\mathcal{V}_{\rho}). Thus, a necessary condition for measuring mixed-state robustness with zero variance is that B(Vρ)∩F=∅\mathcal{B}(\mathcal{V}_{\rho})\cap\mathcal{F}=\varnothing, i.e. there is no free state within the span of the nonzero eigenvectors of ρ\rho, so that the bound breaks down. This occurs for any resourceful pure state, so we consider these separately in the next subsection.

This bound is greater for highly mixed states. Consider a nearly-pure state ρ=(1−ϵ)∣ψ0⟩⟨ψ0∣+ϵ∣ψ1⟩⟨ψ1∣\rho=(1-\epsilon)|\psi_0\rangle\langle\psi_0|+\epsilon|\psi_1\rangle\langle\psi_1|, and some ϕ=∣ϕ⟩⟨ϕ∣\phi=|\phi\rangle\langle\phi| in the relevant subspace with ⟨ϕ∣ψ1⟩≠0\langle\phi|\psi_1\rangle\ne0. In this case the denominator becomes (1−ϵ)−1∣⟨ψ0∣ϕ⟩∣2+(1-\epsilon)^{-1}|\langle\psi_0|\phi\rangle|^{2}+ ϵ−1∣⟨ψ1∣ϕ⟩∣2−1∈Ω(ϵ−1)\epsilon^{-1}|\langle\psi_1|\phi\rangle|^2-1\in\Omega(\epsilon^{-1}), which becomes infinite as ϵ→0\epsilon\to0, causing the lower bound to approach zero when R\mathcal{R} is held fixed. In contrast, the bound is tight for a number of classes of mixed states, with the example of robustness of entanglement for isotropic states considered in Appendix B1.

Generalized Robustness of Pure States

Let us consider the generalized robustness R+\mathcal{R}^{+} of a pure state ∣ψ⟩|\psi\rangle. In this case, we consider the perturbation ∣ψc⟩=∣ψ⟩+c∣ψ′⟩|\psi_c\rangle=|\psi\rangle+c|\psi'\rangle, where ∣ψ′⟩=(F−ϕ)∣ψ⟩|\psi'\rangle=(F-\phi)|\psi\rangle, FF is the free-state fidelity of ∣ψ⟩|\psi\rangle, and ϕ\phi is a (potentially mixed) free state with maximal fidelity to ∣ψ⟩|\psi\rangle, so that F=⟨ψ∣ϕ∣ψ⟩F=\langle\psi|\phi|\psi\rangle. The CR bound derivation is slightly more involved than the mixed state case, but no restriction must be placed on ∣ψ⟩|\psi\rangle – the bound holds for every resourceful pure state. Under this construction, in Appendix A1 we show the following result:

Theorem 2. For any pure state with free-state fidelity FF and generalized robustness R+\mathcal{R}^{+}, the variance of any optimal witness WW of R+\mathcal{R}^{+} obeys

Var⁡(W)≥R+(R+F−1)21−F.(18)\operatorname{Var}(W)\ge\frac{\mathcal{R}^{+}\left(\sqrt{\mathcal{R}^{+}F}-1\right)^2}{1-F}. \tag*{(18)}

Unlike the mixed-state case, this bound can be zero for even resourceful states, in the case where R+=1/F\mathcal{R}^{+}=1/F. This case describes states which are proportional to their own optimal witness, and can thus be measured with zero variance by measuring the state itself. A common example is the ∣T⟩|T\rangle state for generalized robustness of magic. Upper bounding 1−F≤11-F\le1 and inspecting the relative variance reveals

Var⁡(W)R+2≥(F−1R+)2,(19)\frac{\operatorname{Var}(W)}{\mathcal{R}^{+2}}\ge\left(\sqrt{F}-\frac{1}{\sqrt{\mathcal{R}^{+}}}\right)^2, \tag*{(19)}

which grows larger the further R+\mathcal{R}^{+} is from its lower bound 1/F1/F. Increasing both R+\mathcal{R}^{+} and 1/F1/F by some factor cc reduces the relative variance bound by a factor of 1/c1/c, suggesting that generalized robustness will tend to be harder to measure with low relative error for less resourceful pure states. In particular, if R+\mathcal{R}^{+} and 1/F1/F are both exponential in system size while maintaining their ratio, the relative error bound is exponentially small in system size. Theorem 1 does not have the same behavior, suggesting that generalized robustness measures on highly resourceful pure states may be better candidates for estimating the capabilities of quantum devices.

A comparison of the bound in Theorem 2 and those in the next section for low- and high-resource states is considered in Appendix B2, including a family of states for which all bounds are tight.

In Appendix A2, we extend the present approach based on CR bounds to the context of moments, which includes stabilizer, purity, and entanglement moments as example cases. While nonlinear in the state, these quantities can be estimated from an observable on kk copies of a state, of which we bound the variance.

Upper Bounds

When either notion of robustness takes a maximum value Rmax⁡\mathcal{R}_{\max} over the state space, we can also place a loose upper bound on the variance of any witness. As mentioned in Section III C1, in this case ∥W∥≤Rmax⁡\lVert W\rVert\le\mathcal{R}_{\max} for any witness WW. For standard robustness, this implies −Rmax⁡I≤W≤Rmax⁡I-\mathcal{R}_{\max}\mathbb{I}\le W\le\mathcal{R}_{\max}\mathbb{I}, so

Var⁡(W)=⟨W2⟩−⟨W⟩2≤Rmax⁡2−R2,(20)\operatorname{Var}(W)=\langle W^2\rangle-\langle W\rangle^2\le\mathcal{R}_{\max}^2-\mathcal{R}^2, \tag*{(20)}

while for generalized robustness, 0≤W≤Rmax⁡+I0\le W\le\mathcal{R}_{\max}^{+}\mathbb{I}, and

Var⁡(W)≤R+(Rmax⁡+−R+).(21)\operatorname{Var}(W)\le\mathcal{R}^{+}(\mathcal{R}_{\max}^{+}-\mathcal{R}^{+}). \tag*{(21)}

So as R→Rmax⁡\mathcal{R}\to\mathcal{R}_{\max} in either case, the witness variance necessarily approaches zero. Moreover, in both cases the relative variance upper bounds depend only on Rmax⁡/R\mathcal{R}_{\max}/\mathcal{R}, and has no inherent scaling with R\mathcal{R}, unlike the absolute variance.

Estimating Robustness with Bounded Variance

The derivations of Theorems 1 and 2 do not require witness optimality; they apply equally well for any witness WW, with the caveat that Theorem 2 requires ⟨W⟩≥1/F(ψ)\langle W\rangle\ge1/F(\psi). As we will see shortly, this imposes no restriction here. Measurement of sub-optimal witnesses yields a lower bound on robustness, so this motivates asking how good an optimal witness can be achieved within a given variance threshold. Equivalently, how good an estimate of robustness can be achieved with a fixed number of measurements. Substituting R→⟨W⟩\mathcal{R}\to\langle W\rangle into Theorem 1 and solving for it gives the immediate bound

⟨W⟩≤min⁡{1+αVar⁡(W),R(ρ)},(22)\langle W\rangle\le\min\left\{1+\alpha\sqrt{\operatorname{Var}(W)},\mathcal{R}(\rho)\right\}, \tag*{(22)}

where α=Tr⁡(ρ−1ϕ2)−1\alpha=\sqrt{\operatorname{Tr}(\rho^{-1}\phi^2)-1} depends on ρ\rho but not on WW, and R\mathcal{R} stands for either standard or generalized robustness. When relative variance VrelV_{\mathrm{rel}} is held fixed instead, the resulting bound instead goes as

⟨W⟩≤min⁡{11−αVrel,R(ρ)}.(23)\langle W\rangle\le\min\left\{\frac{1}{1-\alpha\sqrt{V_{\mathrm{rel}}}},\mathcal{R}(\rho)\right\}. \tag*{(23)}

for αVrel<1\alpha\sqrt{V_{\mathrm{rel}}}<1, and the trivial ⟨W⟩≤R(ρ)\langle W\rangle\le\mathcal{R}(\rho) otherwise. This behaves similarly to the absolute variance bound for small VrelV_{\mathrm{rel}} but yields much higher allowable ⟨W⟩\langle W\rangle for large VrelV_{\mathrm{rel}}, achieving the true value ⟨W⟩=R\langle W\rangle=\mathcal{R} for some Vrel∗≤α−1/2V_{\mathrm{rel}}^{*}\le\alpha^{-1/2}. For generalized robustness of a pure state, the sub-optimal witness Wψ=ψ/F(ψ)W_{\psi}=\psi/F(\psi) can always be measured with zero variance, and gives ⟨Wψ⟩=1/F(ψ)\langle W_\psi\rangle= 1/F(\psi), giving a somewhat trivial lower bound. To improve on this, we then only need to consider witnesses with ⟨W⟩>1/F(ψ)\langle W\rangle> 1/F(\psi), which satisfies the requirement in Theorem 2. This allows us to substitute R+→⟨W⟩R^+ \to\langle W\rangle in the associated bound. The bound can be solved for R+R^+, but the resulting form is opaque. It is more transparent to instead solve for the best achievable ⟨W⟩\langle W\rangle with fixed relative variance, which gives

⟨W⟩≤min⁡{1F(1−βVrel)2,R+(ψ)}(24)\langle W\rangle\le\min\left\{\frac{1}{F\left(1-\beta\sqrt{V_{\mathrm{rel}}}\right)^2},R^+(\psi)\right\} \tag*{(24)}

when βVrel<1\beta\sqrt{V_{\mathrm{rel}}}<1, where β=1−FF\beta=\sqrt{\frac{1-F}{F}} is independent of WW. At fixed FF this offers a quadratic improvement over the mixed-state bound. For small relative variance, we see that the max achievable robustness bound goes as ⟨W⟩≲1+αVrel\langle W\rangle\lesssim1+\alpha\sqrt{V_{\mathrm{rel}}} while for generalized robustness it goes as ⟨W⟩≲1F+2βFVrel\langle W\rangle\lesssim\frac{1}{F}+\frac{2\beta}{F}\sqrt{V_{\mathrm{rel}}}.

Generalized Robustness Witness Geometry

In the previous section we saw that the pure-state variance of an optimal generalized robustness witness depends on the relationship between the state’s generalized robustness R+R^+ and free-state fidelity FF. In fact, for pure states, the variance can be written exactly in terms of the angle between the target state and a witness state, making the variance an inherently geometric quantity. In this section, we will use this geometric nature to get a different lower bound on witness variance. Its geometric origin makes it easier to interpret and potentially to generalize, but at present only holds for states with high free-state fidelity, typically those with low resource.

For a pure state ∣ψ⟩|\psi\rangle, there always exists an optimal generalized robustness witness which is proportional to a pure state projector [18, 21, 28]:

W=∣ω⟩⟨ω∣F(ω)(25)W=\frac{|\omega\rangle\langle\omega|}{F(\omega)} \tag*{(25)}

where ∣ω⟩|\omega\rangle is a normalized state and F(ω)F(\omega) is the free-state fidelity of state ∣ω⟩|\omega\rangle.

One can also construct such a pure optimal witness WW from any optimal witness W′W' by setting

W=W′∣ψ⟩⟨ψ∣W′R+(ψ).(26)W=\frac{W'|\psi\rangle\langle\psi|W'}{R^+(\psi)}. \tag*{(26)}

The resulting WW obeys ⟨ϕ∣W∣ϕ⟩=∣⟨ϕ∣W′∣ψ⟩∣2/R+≤⟨ϕ∣W′∣ϕ⟩≤1 ∀ϕ∈F\langle\phi|W|\phi\rangle=|\langle\phi|W'|\psi\rangle|^2/R^+\le\langle\phi|W'|\phi\rangle\le1\ \forall\phi\in\mathcal{F} by Cauchy-Schwarz, verifying that it is a witness. Moreover, since W∣ψ⟩=W′∣ψ⟩W|\psi\rangle=W'|\psi\rangle, it is both optimal and has Var⁡(W)=Var⁡(W′)\operatorname{Var}(W)=\operatorname{Var}(W'). Thus, while we work with pure optimal witnesses in this section, the variance results hold for any optimal witness. However, unlike the previous section, they do not extend to other unbiased estimators of R+R^+.

For pure WW, the variance of the witness operator takes a geometric form:

Var⁡(W)=⟨W2⟩−⟨W⟩2=∣⟨ψ∣ω⟩∣2F(ω)2−∣⟨ψ∣ω⟩∣4F(ω)2=(R+)2(1−∣⟨ψ∣ω⟩∣2∣⟨ψ∣ω⟩∣2).(27)\begin{aligned} \operatorname{Var}(W)=\langle W^2\rangle-\langle W\rangle^2=\frac{|\langle\psi|\omega\rangle|^2}{F(\omega)^2}-\frac{|\langle\psi|\omega\rangle|^4}{F(\omega)^2} \\ &=(R^+)^2\left(\frac{1-|\langle\psi|\omega\rangle|^2}{|\langle\psi|\omega\rangle|^2}\right). \tag*{(27)} \end{aligned}

Make the substitution θψω=cos⁡−1(∣⟨ψ∣ω⟩∣)\theta_{\psi\omega}=\cos^{-1}(|\langle\psi|\omega\rangle|) representing the state-space angle between states ∣ψ⟩|\psi\rangle and ∣ω⟩|\omega\rangle. Since sin⁡2(θψω)=1−∣⟨ψ∣ω⟩∣2\sin^2(\theta_{\psi\omega})=1-|\langle\psi|\omega\rangle|^2, the variance trigonometrically simplifies to

Var⁡(W)=(R+)2tan⁡2(θψω).(28)\operatorname{Var}(W)=(R^+)^2\tan^2(\theta_{\psi\omega}). \tag*{(28)}

For pure states, the variance of the optimal witness is quadratic in generalized robustness but also depends on the alignment between the original and witness states. If one can pin down the set of potential witness states, one can lower bound the variance in terms of the distance between the original state and this set. The tighter the set of potential witnesses, the greater the minimum possible angle θψω\theta_{\psi\omega}, and the tighter the variance lower bound.

Witness Free-State Fidelity

As a first step towards characterizing this set, consider the witness free-state fidelity F(ω)F(\omega). We can upper bound F(ω)F(\omega) in terms of F(ψ)F(\psi) as

F(ψ)≥1R+(ψ)=F(ω)∣⟨ψ∣ω⟩∣2≥F(ω).(29)F(\psi)\ge\frac{1}{R^+(\psi)}=\frac{F(\omega)}{|\langle\psi|\omega\rangle|^2}\ge F(\omega). \tag*{(29)}

with the two having equality iff ∣ψ⟩=∣ω⟩|\psi\rangle=|\omega\rangle, which also implies R+(ψ)=1/F(ψ)R^+(\psi)=1/F(\psi).

When F\mathcal{F} is the convex hull of pure free states Fpure\mathcal{F}_{\mathrm{pure}} and ∣ψ⟩|\psi\rangle is a pure state, R+(ψ)R^+(\psi) admits a primal definition as the pure-state extent [4, 5, 21]

R+(ψ)=ξ(ψ)=min⁡c⃗{∥c⃗∥12 | ∣ψ⟩=∑i=1rci∣ϕi⟩}(30)R^+(\psi)=\xi(\psi)=\min_{\vec{c}}\left\{\lVert\vec{c}\rVert_1^2\ \middle|\ |\psi\rangle=\sum_{i=1}^{r}c_i|\phi_i\rangle\right\} \tag*{(30)}

where ci∈Cc_i\in\mathbb{C} are nonzero, ∣ϕi⟩∈Fpure|\phi_i\rangle\in\mathcal{F}_{\mathrm{pure}}, and ∥c⃗∥12=(∑i∣ci∣)2\lVert\vec{c}\rVert_1^2=\left(\sum_i|c_i|\right)^2.

For a given decomposition, call a free state “active” if it has a nonzero coefficient. When ∣ψ⟩|\psi\rangle admits an optimal decomposition with minimum angle cos⁡(θf)=max⁡i≠j∣⟨ϕi∣ϕj⟩∣\cos(\theta_f)=\max_{i\ne j}|\langle\phi_i|\phi_j\rangle| between active free states, we can also obtain a θf\theta_f-independent upper bound on F(ω)F(\omega). This is guaranteed to exist in resource theories with a minimal angle between pure free states, such as magic (θf=π4\theta_f=\frac{\pi}{4} [4]) and coherence (θf=π2\theta_f=\frac{\pi}{2}). It also holds for entanglement with θf=π2\theta_f=\frac{\pi}{2}, as in this case the Schmidt decomposition is optimal, even though general free states have no minimal separating angle.

Lemma 1. When a state admits an optimal free-state extent decomposition with rr active free states and minimum angle θf\theta_f between active free states, any pure optimal witness state ∣ω⟩\lvert\omega\rangle must have

F(ω)≤Fr:=1+(r−1)cos⁡(θf)r.(31)F(\omega) \le F_r := \frac{1+(r-1)\cos(\theta_f)}{r}. \tag*{(31)}

For any resourceful state, r≥2r \ge2, so F(ω)≤F2=(1+cos⁡(θf))/2=cos⁡2(θf/2)F(\omega) \le F_2=(1+\cos(\theta_f))/2=\cos^2(\theta_f/2).

Proof. Plugging an optimal free-state decomposition of ∣ψ⟩\lvert\psi\rangle into the dual definition R+(ψ)=∣⟨ψ∣ω⟩∣2/F(ω)R^+(\psi)=\lvert\langle\psi\vert\omega\rangle\rvert^2/F(\omega) and equating ξ(ψ)=R+(ψ)\xi(\psi)=R^+(\psi) reveals that ∣⟨ω∣ϕi⟩∣2=F(ω)\lvert\langle\omega\vert\phi_i\rangle\rvert^2=F(\omega) for every free state ∣ϕi⟩\lvert\phi_i\rangle in the sum [5].

Pick the phase of each ∣ϕi⟩\lvert\phi_i\rangle so that ⟨ω∣ϕi⟩=F(ω)\langle\omega\vert\phi_i\rangle=\sqrt{F(\omega)}, and set ∣v⟩=∑i∣ϕi⟩\lvert v\rangle=\sum_i\lvert\phi_i\rangle. Then ⟨ω∣v⟩=rF(ω)\langle\omega\vert v\rangle=r\sqrt{F(\omega)}. Since ⟨ϕi∣ϕj⟩≤cos⁡(θf)\langle\phi_i\vert\phi_j\rangle\le\cos(\theta_f), ⟨v∣ϕi⟩≤1+(r−1)cos⁡(θf)\langle v\vert\phi_i\rangle\le1+(r-1)\cos(\theta_f), and we can bound

r2F(ω)=⟨ω∣v⟩2≤⟨v∣v⟩=∑i⟨v∣ϕi⟩≤r(1+(r−1)cos⁡(θf)).(32)r^2F(\omega)=\langle\omega\vert v\rangle^2\le\langle v\vert v\rangle=\sum_i\langle v\vert\phi_i\rangle\le r(1+(r-1)\cos(\theta_f)). \tag*{(32)}

Solving for F(ω)F(\omega) then shows Eq. (31).

Note that rr can be lower-bounded by the state’s free-state rank. Plugging Fr≥F(ω)F_r\ge F(\omega) into the second line of Eq. (27) gives:

Theorem 3. For a pure state ∣ψ⟩\lvert\psi\rangle, the variance of any optimal generalized robustness witness WW satisfies

Var⁡(W)≥R+(1Fr−R+),(33)\operatorname{Var}(W)\ge R^+\left(\frac{1}{F_r}-R^+\right), \tag*{(33)}

where Fr:=(1+(r−1)cos⁡(θf))/r≤cos⁡2(θf/2)F_r:=(1+(r-1)\cos(\theta_f))/r\le\cos^2(\theta_f/2), rr is the number of active free states in the optimal free-state extent decomposition of ∣ψ⟩\lvert\psi\rangle, and θf\theta_f is the minimal angle between active free states.

In the r=2r=2 case, the geometric interpretation of the lemma is particularly simple: ∣ω⟩\lvert\omega\rangle is equidistant to ∣ϕ1⟩,∣ϕ2⟩\lvert\phi_1\rangle,\lvert\phi_2\rangle, so θωϕ1=θωϕ2\theta_{\omega\phi_1}=\theta_{\omega\phi_2}. As the angle between the two free states is at least θf\theta_f, the triangle inequality gives

θf≤θϕ1ϕ2≤θωϕ1+θωϕ2=2θωϕ1.(34)\theta_f\le\theta_{\phi_1\phi_2}\le\theta_{\omega\phi_1}+\theta_{\omega\phi_2}=2\theta_{\omega\phi_1}. \tag*{(34)}

Thus, F(ω)=cos⁡2(θωϕ1)≤cos⁡2(θf/2)F(\omega)=\cos^2(\theta_{\omega\phi_1})\le\cos^2(\theta_f/2), matching the general bound in the r=2r=2 case. For r>2r>2, ∣ω⟩\lvert\omega\rangle lies at the center of an rr-vertex polytope.

When θf=π2\theta_f=\frac{\pi}{2} and the witness W∗=∣v⟩⟨v∣=∑ij∣ϕi⟩⟨ϕj∣W^*=\lvert v\rangle\langle v\rvert=\sum_{ij}\lvert\phi_i\rangle\langle\phi_j\rvert is feasible, W∗W^* is both optimal and saturates the variance bound in Theorem 3. This is guaranteed when all pure free states are orthogonal, such as for coherence, where they are basis states [28]. It also occurs for bipartite entanglement, where the active free states are the product states resulting from the Schmidt decomposition, which can be shown to obtain the optimal pure-state extent [28]. Feasibility of W∗W^* then follows from a Cauchy–Schwarz argument showing ∣⟨v∣a,b⟩∣≤1\lvert\langle v\vert a,b\rangle\rvert\le1 for any product state ∣a,b⟩\lvert a,b\rangle; see, e.g., Eq. (B.2).

In contrast, when cos⁡(θf)>0\cos(\theta_f)>0, one can lower bound Fr≥cos⁡(θf)F_r\ge\cos(\theta_f), so Theorem 3 is only nontrivial when R+<sec⁡(θf)R^+<\sec(\theta_f). For example, in the case of magic, the bound is only of interest for R+<2R^+<\sqrt{2}. Within this regime it is still useful for bounding the relative variance of optimal witnesses. For example, for a nearly-free state with R+=1+ϵR^+=1+\epsilon and r=2r=2, F2=cos⁡2(θf/2)F_2=\cos^2(\theta_f/2), and the bound yields

Var⁡(W)(R+)2≥1FrR+−1=sec⁡2(θf2)(1−ϵ)−1+O(ϵ2)=tan⁡2(θf2)−sec⁡2(θf2)ϵ+O(ϵ2).(35)\begin{aligned} \frac{\operatorname{Var}(W)}{(R^+)^2}\ge\frac{1}{F_rR^+}-1=\sec^2\left(\frac{\theta_f}{2}\right)(1-\epsilon)-1+O(\epsilon^2) \\ &=\tan^2\left(\frac{\theta_f}{2}\right)-\sec^2\left(\frac{\theta_f}{2}\right)\epsilon+O(\epsilon^2). \tag*{(35)} \end{aligned}

This highlights that for a sufficiently low-resource non-free state, the optimal witness relative variance can only be zero if θf∼O(ϵ)\theta_f\sim O(\sqrt{\epsilon}), necessitating that F\mathcal{F} be continuous as ϵ→0\epsilon\to0.

Geometric Variance Bound

We can use Lemma 1 and Eq. (28) to lower bound the variance of the optimal witness when F(ψ)≥FrF(\psi)\ge F_r.

Theorem 4. For a pure state ∣ψ⟩\lvert\psi\rangle with F:=F(ψ)>Fr:=(1+(r−1)cos⁡(θf))/rF:=F(\psi)>F_r:=(1+(r-1)\cos(\theta_f))/r, the variance of any optimal witness WW is lower bounded by

Var⁡(W)≥(R+)2(F−FrF(1−F)+Fr(1−Fr))2,(36)\operatorname{Var}(W)\ge(R^+)^2\left(\frac{F-F_r}{\sqrt{F(1-F)}+\sqrt{F_r(1-F_r)}}\right)^2, \tag*{(36)}

where rr is the number of active free states in the optimal pure state extent decomposition of ∣ψ⟩\lvert\psi\rangle, and θf\theta_f is the minimal angle between active free states.

Proof. We will lower bound θψω\theta_{\psi\omega}, then use Var⁡(W)=(R+)2tan⁡2(θψω)\operatorname{Var}(W)=(R^+)^2\tan^2(\theta_{\psi\omega}) and several trig identities. Let ∣ϕ⟩\lvert\phi\rangle be any free state maximizing overlap with ∣ψ⟩\lvert\psi\rangle, so that F(ψ)=cos⁡2(θϕψ)F(\psi)=\cos^2(\theta_{\phi\psi}). Lemma 1 guarantees that θϕω:=arccos⁡Fr≤min⁡ϕ′θϕ′ω≤θϕω\theta_{\phi\omega}:=\arccos\sqrt{F_r}\le\min_{\phi'}\theta_{\phi'\omega}\le\theta_{\phi\omega}. Combining this with the triangle inequality gives

θr≤θϕω≤θϕψ+θψω.(37)\theta_r\le\theta_{\phi\omega}\le\theta_{\phi\psi}+\theta_{\psi\omega}. \tag*{(37)}

We can then rearrange terms to solve for θψω\theta_{\psi\omega} to get θψω≥θr−θϕψ\theta_{\psi\omega}\ge\theta_r-\theta_{\phi\psi}. Plugging this in as a lower bound on θψω\theta_{\psi\omega} in Eq. (28) then gives

Var⁡(W)≥(R+)2tan⁡2(θr−θϕψ)\operatorname{Var}(W)\ge(R^+)^2\tan^2(\theta_r-\theta_{\phi\psi})
=(R+)2tan⁡2(arccos⁡Fr−arccos⁡F).(38)=(R^+)^2\tan^2\left(\arccos\sqrt{F_r}-\arccos\sqrt{F}\right). \tag*{(38)}

One can then use the trigonometric identities for difference of angles to expand sin⁡(θr−θϕψ)\sin(\theta_r-\theta_{\phi\psi}) and cos⁡(θr−θϕψ)\cos(\theta_r-\theta_{\phi\psi}), then substitute sin⁡(θ)→1−cos⁡2(θ)\sin(\theta)\to\sqrt{1-\cos^2(\theta)}, leading to a cancellation of all arccos terms. The resulting form is

lengthy, but multiplying the resulting numerator and denominator by F(1−Fr)+Fr(1−F)\sqrt{F(1-F_r)}+\sqrt{F_r(1-F)} simplifies it to

tan⁡(θr−θϕψ)=F−FrF(1−F)+Fr(1−Fr).(39)\tan(\theta_r-\theta_{\phi\psi})=\frac{F-F_r}{\sqrt{F(1-F)}+\sqrt{F_r(1-F_r)}}. \tag*{(39)}

Plugging this into Eq. (38) then proves the theorem. □

The denominator in Theorem 4 is maximized for F=Fr=12F=F_r=\frac{1}{2}, so it can also be upper bounded by 1 to get a simpler but looser bound:

Var⁡(W)≥(R+)2(F−Fr)2.(40)\operatorname{Var}(W)\geq(\mathcal{R}^+)^2(F-F_r)^2. \tag*{(40)}

Theorem 4 highlights the approximately-quadratic dependence of relative variance on F−FrF-F_r, the difference between the state’s fidelity and the maximal fidelity allowed for any optimal witness state. F−FrF-F_r can be interpreted as the fidelity-distance of ∣ψ⟩|\psi\rangle from the set of possible optimal witness states. When ∣ψ⟩|\psi\rangle is close to being a free state, the fidelity part of the bound in Theorem 4 has a positive relationship with F(ψ)F(\psi). That is, the relative variance lower bound increases as the state becomes less resourceful, highlighting the difficulty of detecting resources in low-resource states. This mirrors the result from the previous section, where relative variance was found to decrease with R+\mathcal{R}^+, 1/F1/F, as long as the ratio of the two is held constant.

In Appendix B2 we provide a comparison of the bounds in this section and the bound in Theorem 2 for low and high-resource states, and discuss a family of states for which all bounds are tight. Unlike in the previous section, the bounds in this section rely on witness optimality, so they cannot be immediately used to determine the best allowed ⟨W⟩\langle W\rangle at fixed variance.

Sampling hardness of tracking the onset of resource robustness

An additional challenge emerges if we want to track the onset of resource robustness. That is, if we want to evolve some state and determine when it transitions from being low-resource to exponentially high resource. In this case, even if the start and endpoints of evolution can both be measured with low variance, we show that there always exist parts of the trajectory for which resource robustness cannot be efficiently estimated. While the previous two sections established conditional lower bounds on the variance of optimal witnesses, it is hard to determine when these bounds become exponentially large without additional information (ϕ\phi for mixed states, or FF for pure states). Such resource-specific information is not needed here; we instead establish global bounds along paths from low to high resource states, for which the integrated variance is guaranteed to be large as long as the path has a local generator.

Consider a system whose state ρx\rho_x depends on a parameter xx, ranging from zero to one (without loss of

generality), for which R(ρx)\mathcal{R}(\rho_x) is an absolutely continuous function of xx, so that R(ρx)\mathcal{R}(\rho_x) is differentiable almost everywhere. At each point xx, let WxW_x be any estimator of R(ρx)\mathcal{R}(\rho_x), which is locally unbiased in xx so that ∂xR(ρx)=Tr⁡(Wx∂xρx)\partial_x\mathcal{R}(\rho_x)=\operatorname{Tr}(W_x\partial_x\rho_x). This includes any optimal witness for ρx\rho_x, as shown by Danskin’s theorem [Section II C]. The total change ΔR:=R(ρ1)−R(ρ0)\Delta\mathcal{R}:=\mathcal{R}(\rho_1)-\mathcal{R}(\rho_0) satisfies

∣ΔR∣≤∫01∣∂xR(ρx)∣ dx=∫01∣Tr⁡(Wx∂xρx)∣ dx=∫01∣Cov⁡(Lx,Wx)∣ dx≤∫01Var⁡(Lx)Var⁡(Wx) dx≤∫01Var⁡(Lx) dx∫01Var⁡(Wx) dx.(41)\begin{aligned} |\Delta\mathcal{R}| &\leq\int_0^1|\partial_x\mathcal{R}(\rho_x)|\,dx \\ &= \int_0^1|\operatorname{Tr}(W_x\partial_x\rho_x)|\,dx \\ &= \int_0^1|\operatorname{Cov}(\mathcal{L}_x,W_x)|\,dx \\ &\leq\int_0^1\sqrt{\operatorname{Var}(\mathcal{L}_x)\operatorname{Var}(W_x)}\,dx \\ &\leq\sqrt{\int_0^1\operatorname{Var}(\mathcal{L}_x)\,dx}\sqrt{\int_0^1\operatorname{Var}(W_x)\,dx}. \tag*{(41)} \end{aligned}

The third line comes from expanding ∂xρx\partial_x\rho_x in terms of the log-derivative, while the last two lines follow from the Cauchy–Schwarz inequality for operators and functions, respectively. We exclude the measure-zero set of points where R(ρx)\mathcal{R}(\rho_x) is not differentiable from the integral, as this does not change its total value.

Equation (41) then relates the integrated noise on the estimator to the change in the robustness:

Theorem 5. For a path ρ0→ρ1\rho_0\to\rho_1 over which R(ρx)\mathcal{R}(\rho_x) is an absolutely continuous function of xx, and where WxW_x is any unbiased estimator of robustness for ρx\rho_x such as an optimal witness, the integrated variance obeys

∫01Var⁡(Wx) dx≥(ΔR)2∫01Ix dx,(42)\int_0^1\operatorname{Var}(W_x)\,dx\geq\frac{(\Delta\mathcal{R})^2}{\int_0^1 I_x\,dx}, \tag*{(42)}

for quantum Fisher information Ix=Var⁡(Lx)I_x=\operatorname{Var}(\mathcal{L}_x) and total change in robustness ΔR=R(ρ1)−R(ρ0)\Delta\mathcal{R}=\mathcal{R}(\rho_1)-\mathcal{R}(\rho_0). Consider probing the resource generation, in an experimentally measurable manner, of an nn-qubit system evolving from ρ0\rho_0 to ρ1\rho_1 with ΔR∈exp⁡(n)\Delta R \in\exp(n). When Ix∈O(poly⁡(n))I_x \in O(\operatorname{poly}(n)) for xx along the path, we are guaranteed the integrated variance in Eq. (42) is Ω(exp⁡(n))\Omega(\exp(n)). The latter condition occurs, for example, any time the generator Lx\mathcal{L}_x of changes in xx is a sum of geometrically local terms over k∈O(1)k \in O(1) qubits, in which case Ix∈O(n)I_x \in O(n) for states with finite correlation length, or O(n2)O(n^2) for globally correlated states. Another case with polynomial Fisher information is real or imaginary time evolution generated by a local Hamiltonian in time T∈poly⁡(n)T \in\operatorname{poly}(n), for which Ix≤4T2Var⁡(Hx)∈O(poly⁡(n))I_x \leq4T^2\operatorname{Var}(H_x) \in O(\operatorname{poly}(n)), where the factor of T2T^2 comes from the rescaling x=t/Tx=t/T [25]. For some non-measure-zero subset of the path from ρ0\rho_0 to ρ1\rho_1, any estimator has exponential variance, and so robustness cannot be measured efficiently on this subset. This holds even when both endpoints possess an easily measurable, low-variance witness, in which case the high-variance subset separates the two states while not including them. Figure 1 illustrates such a no-go result for efficiently and continuously tracking the creation of a quantum resource by robustness measures.

Diagram showing a path between free and resourceful states

Figure 1. Consider any free state ρ0∈F\rho_0\in\mathcal{F} and exponentially-resourceful state ρ1\rho_1 with R(ρ1)∼exp⁡(n)\mathcal{R}(\rho_1)\sim\exp(n). Any absolutely continuous path {ρx∣0≤x≤1}\{\rho_x\mid0\leq x\leq1\} of polynomial Fisher length which connects ρ0\rho_0 to ρ1\rho_1, must pass through a region within which every optimal witness has variance exponential in system size (red). This holds even if path endpoints admit low-variance optimal witnesses (blue regions). Thus, robustness cannot be efficiently measured on a subset of the path.

Consider two regions AA and BB of state space between which there is an exponentially large robustness gap, i.e., min⁡ρA∈A,ρB∈B∣R(ρA)−R(ρB)∣∈exp⁡(n)\min_{\rho_A\in A,\rho_B\in B}|R(\rho_A)-R(\rho_B)|\in\exp(n). Then every path of sub-exponential Fisher length ∫01Ix dx\int_0^1 I_x\,dx from a state in AA to one in BB must cross through the set of states for which the robustness cannot be efficiently estimated. Moreover, any two regions A,BA,B with R(ρA)∼poly⁡(n)ebAn ∀ρA∈AR(\rho_A)\sim\operatorname{poly}(n)e^{b_A n}\ \forall\rho_A\in A and R(ρB)∼poly⁡(n)ebBn ∀ρB∈BR(\rho_B)\sim\operatorname{poly}(n)e^{b_B n}\ \forall\rho_B\in B for bA≠bBb_A\ne b_B will have ΔR∼exp⁡(n)\Delta R\sim\exp(n). This implies state space can be partitioned into regions of differing exponential robustness growth rate, between which any path either has exponential Fisher length or contains states for which robustness cannot be efficiently estimated.

If we instead consider the relative variance, the picture is much more positive. A slight modification of the previous bound bounds the relative variance in terms of ratio R(ρ1)/R(ρ0)R(\rho_1)/R(\rho_0):

∣log⁡(R(ρ1)R(ρ0))∣≤∫01∣∂xlog⁡R(ρx)∣dx=∫011R(ρx)∣∂xR(ρx)∣dx≤∫011R(ρx)Var⁡(Lx)Var⁡(Wx) dx≤∫01Ix dx∫01Var⁡(Wx)R(ρx)2 dx.(43)\begin{aligned} \left|\log\left(\frac{R(\rho_1)}{R(\rho_0)}\right)\right| \leq\int_0^1 \left|\partial_x\log R(\rho_x)\right|dx \\ &= \int_0^1 \frac{1}{R(\rho_x)}\left|\partial_xR(\rho_x)\right|dx \\ &\leq\int_0^1 \frac{1}{R(\rho_x)}\sqrt{\operatorname{Var}(\mathcal{L}_x)\operatorname{Var}(W_x)}\,dx \\ &\leq\sqrt{\int_0^1 I_x\,dx}\sqrt{\int_0^1\frac{\operatorname{Var}(W_x)}{R(\rho_x)^2}\,dx}. \tag*{(43)} \end{aligned}

where we get from the second to third line by taking similar steps as in Eq. (41). Then the resulting bound on the relative variance is

∫01Var⁡(Wx)R(ρx)2 dx≥∣log⁡(R(ρ1)R(ρ0))∣2∫01Ix dx.(44)\int_0^1\frac{\operatorname{Var}(W_x)}{R(\rho_x)^2}\,dx\geq\frac{\left|\log\left(\frac{R(\rho_1)}{R(\rho_0)}\right)\right|^2}{\int_0^1 I_x\,dx}. \tag*{(44)}

Under our convention R≥1R\geq1, and R∈O(exp⁡(n))R\in O(\exp(n)) for many resources of interest, such as magic, entanglement, generalized robustness of coherence, and for well-behaved resources on product states. In such cases, as long as Ix∈O(1)I_x\in O(1) (non-decreasing in system size), the lower bound in Eq. (44) goes as O(n2)O(n^2) allowing the possibility of estimating robustness to low relative error in O(n2)O(n^2) measurements.

To compare the two lower bounds, consider the generalized robustness of magic R+\mathcal{R}^+ along a path of states of the form ∣ψxTn−1⟩|\psi_xT^{n-1}\rangle, with ∣ψ0⟩=∣+⟩|\psi_0\rangle=|+\rangle and ∣ψ1⟩=∣T⟩|\psi_1\rangle=|T\rangle, where ∣T⟩|T\rangle is defined as

∣T⟩:=T∣+⟩=(∣0⟩+eiπ/4∣1⟩)/2.(45)|T\rangle:= T|+\rangle=\left(|0\rangle+e^{i\pi/4}|1\rangle\right)/\sqrt{2}. \tag*{(45)}

Recall that for magic, ∣+⟩|+\rangle is a stabilizer (free) state, while ∣T⟩|T\rangle is resourceful. ∣T⟩|T\rangle can be generated by applying the single-qubit Hamiltonian H=−π4∣1⟩⟨1∣H=-\frac{\pi}{4}|1\rangle\langle1| to ∣+⟩|+\rangle for time t=1t=1, and so ∫01Ixdx∈O(1)\int_0^1I_xdx\in O(1). ∣+⟩|+\rangle, ∣T⟩|T\rangle, and in fact all states along the path between them, share a common optimal generalized robustness witness WT=∣T⟩⟨T∣/F(T)W_T=|T\rangle\langle T|/F(T), proportional to a projector onto the ∣T⟩|T\rangle state [5]. The generalized robustness for the ∣T⟩|T\rangle state can then be measured with zero variance, since the state is being measured in its own eigenbasis.

Generalized robustness of magic is multiplicative for product single-qubit pure states, so in this case [5]

R+(ρ0)=R+(T)n−1R+(+)=R+(T)n−1\mathcal{R}^+(\rho_0)=\mathcal{R}^+(T)^{n-1}\mathcal{R}^+(+)=\mathcal{R}^+(T)^{n-1}
R+(ρ1)=R+(T)n\mathcal{R}^+(\rho_1)=\mathcal{R}^+(T)^n
ΔR+=R+(T)n−1(R+(T)−1)∈Θ(exp⁡(n))\Delta R^+=\mathcal{R}^+(T)^{n-1}\left(\mathcal{R}^+(T)-1\right)\in\Theta(\exp(n))
log⁡(R+(ρ1)R+(ρ0))=log⁡R+(T)∈O(1).(46)\log\left(\frac{\mathcal{R}^+(\rho_1)}{\mathcal{R}^+(\rho_0)}\right)=\log\mathcal{R}^+(T)\in O(1). \tag*{(46)}

This also implies the optimal witness is a product of the single state optimal witnesses, in this case WT⊗n=∣Tn⟩⟨Tn∣/F(T)nW_T^{\otimes n}=|T^n\rangle\langle T^n|/F(T)^n, and R+(T)n=1/F(T)n\mathcal{R}^+(T)^n=1/F(T)^n.

The optimal witness variance at each point xx is

Var⁡x(WT⊗n)=1F(T)2n(⟨T∣ψx⟩2−⟨T∣ψx⟩4)=R+(T)2n−2Var⁡ψx(WT).(47)\begin{aligned} \operatorname{Var}_x(W_T^{\otimes n})=\frac{1}{F(T)^{2n}}\left(\langle T|\psi_x\rangle^2-\langle T|\psi_x\rangle^4\right) \\ &=\mathcal{R}^+(T)^{2n-2}\operatorname{Var}_{\psi_x}(W_T). \tag*{(47)} \end{aligned}

The scaling captured by both bounds is tight, as Var⁡x(WT⊗n)∼exp⁡(n)\operatorname{Var}_x(W_T^{\otimes n})\sim\exp(n) while Var⁡x(WT⊗n)/R+(ρx)2=Var⁡ψx(T)/R+(ψx)2∈O(1)\operatorname{Var}_x(W_T^{\otimes n})/\mathcal{R}^+(\rho_x)^2=\operatorname{Var}_{\psi_x}(T)/\mathcal{R}^+(\psi_x)^2\in O(1) at each point. However, if we interpret the nn-qubit system as nn independent qubits, we see that the path can be interpreted as creating only one qubit’s worth of resource, so the exponential scaling in ΔR+\Delta R^+ is somewhat misleading. Similarly, witness measurements on ρx\rho_x can be interpreted as n−1n-1 independent zero-variance measurements and one with nonzero variance (for x<1x<1), so the exponential scaling of absolute variance is somewhat misleading, since it comes from the scale of WT⊗nW_T^{\otimes n} alone. The relative variance does not have such shortcomings, highlighting why estimating R+R^+ to low multiplicative error can be qualitatively easier than to low absolute error.

Discussion

We obtain bounds on the sampling cost of estimating standard and generalized robustness that apply across resource theories. These bounds expose statistical limitations that remain even when an optimal witness is known, and connect experimental precision to the target state’s resource content. Our results indicate potential hurdles to resource estimation on mixed states, estimation of standard robustness, and resource estimation to low absolute error. This suggests that the estimation of generalized robustness to low relative error on highly resourceful pure states is a preferable candidate resource-estimation task, since it avoids many of the hurdles.

Treating the optimal witness as an observable allows one to place Cramér-Rao bounds on the number of measurements required to estimate robustness for a range of resources (Theorems 1, 2). For mixed states, this shot complexity generically scales at least quadratically in the resource robustness, making resource estimation of generic highly resourceful states inefficient. For generalized robustness of pure states, the shot complexity lower bound obeys different scaling – linear in robustness for state families with fixed robustness-fidelity ratio. When the task is estimating the generalized robustness to fixed relative error, the bound has a decreasing relationship with resource robustness, indicating the task has the potential to be easy for highly resourceful states.

Our geometric bounds highlight that shot complexity can remain large even for low-resource states, depending on the structure of the free set. In particular, the bound in Theorem 3 is tight when active free states are orthogonal, which is the case for the generalized robustness of coherence and bipartite entanglement. In these cases, one can analytically find a witness which both is optimal and achieves minimal shot complexity. A stronger obstruction arises when tracking resource generation along a trajectory in state space (Theorem 5). A large gain in robustness necessitates a path which crosses through intermediate states with high witness variance, even when the path endpoints are easily measurable. This establishes a no-go theorem in tracking the onset of resource robustness in state-space evolutions with fixed absolute error. The corresponding bound on relative error is significantly weaker, scaling squared-logarithmically in resource gain. As we saw from the example, high absolute error and low relative error can coexist for local evolutions on highly resourceful states. This distinction highlights the importance of the specific estimation task when assessing the feasibility of a benchmark.

An open question is to what extent the Cramér-Rao bounds in Section III can be improved without requiring prior knowledge about the optimal witness. A perturbation in the direction of an optimal witness yields a saturable Cramér-Rao bound, but the resulting expression is witness-dependent, limiting its practicality. A useful next step is to find perturbations determined by accessible state properties which yield tighter bounds. The geometric bounds suggest a complementary approach: restricting the allowed set of potential optimal witnesses for a given state would let one better bound its distance from its optimal witness, resulting in tighter geometric bounds on shot complexity.

The location and extent of the unmeasurable subset observed along resource-generating paths described in § 2.4 also remain unclear. Relative variance bounds and the absolute variance upper bound imply typical unmeasurable subsets consist of highly-but-not-maximally resourceful states. However, it remains open whether the unmeasurable subset is inherently path-dependent, whether there exists a boundary region of unmeasurable states separating measurable regions, or if efficiently measurable states make up small islands in state space amid a sea of unmeasurable states. This question motivates an experimental approach where one evolves a state from zero to exponentially high resource, while measuring resource at each step to track when the witness variance becomes prohibitively high. Generalized robustness of coherence is a particularly clean choice for this, as the optimal witness can be found analytically, and potentially in practice for all states along the trajectory. Comparisons of shot complexity and computational complexity of finding the optimal witness would also be revealing – categorizing the set of states which are low in both complexity metrics could guide experiments towards efficient demonstration of quantum capabilities.

Acknowledgments

This work was supported by the Quantum Science Center (QSC), a National Quantum Information Science Research Center of the U.S. Department of Energy (DOE). B.B. acknowledges support from the DOE, Office of Science, Accelerated Research in Quantum Computing, Fundamental Algorithmic Research toward Quantum Utility (FAR-Qu). L.P.G.P. acknowledges support from the DOE, Office of Science, Basic Energy Sciences program (award No. DE-SCLO000157).

ChatGPT-6 Astra was used for editing, literature search, and to ensure the accuracy and novelty of results. It assisted in the derivations of Eq. (26), Lemma 1, and Theorem 4. It derived the results in Appendix B, helped shorten Appendix A1, and made Figure 1.

APPENDIX

Resource Cramér-Rao Bounds

This section contains derivations of the perturbative Cramér-Rao (CR) bounds on generalized robustness and on moment functions used in Section III.

Generalized Robustness of Pure States

Here we derive a perturbative CR bound on the variance of any optimal witness of generalized robustness of pure states. For pure state ∣ψ⟩|\psi\rangle, consider the perturbed state ∣ψc⟩=∣ψ⟩+c∣ψ′⟩|\psi_c\rangle= |\psi\rangle+ c|\psi'\rangle in the limit c→0c \to0, where $\langle\psi|\psi'\rangle = 0 and c∈Rc \in\mathbb{R}. We use the perturbation ∣ψ′⟩=(F−ϕ)∣ψ⟩|\psi'\rangle= (F-\phi)|\psi\rangle, where FF is the free-state fidelity of ∣ψ⟩|\psi\rangle and ϕ\phi is the (potentially mixed) free state maximizing ⟨ψ∣ϕ∣ψ⟩\langle\psi|\phi|\psi\rangle, so that F=⟨ψ∣ϕ∣ψ⟩F=\langle\psi|\phi|\psi\rangle. At c=0c=0, the symmetric log-derivative is Lc∣c=0=2(FI−ϕ)\mathcal{L}_c|_{c=0}=2(F\mathbb{I}-\phi), making the quantum Fisher information

Ic∣c=0=⟨Lc2⟩∣c=0=4(⟨ϕ2⟩−F2)≤4F(1−F).(48)I_c\big|_{c=0}=\langle\mathcal{L}_c^2\rangle\big|_{c=0}=4(\langle\phi^2\rangle-F^2)\leq4F(1-F). \tag*{(48)}

Turning now to the numerator of the CR bound, we have

∂cRW+(ψc)∣c=0=2Re⁡⟨ψ∣W∣ψ⟩=2(R+F−Re⁡⟨ψ∣Wϕ∣ψ⟩)(49)\partial_c\mathcal{R}^{+}_{W}(\psi_c)\big|_{c=0}=2\operatorname{Re}\langle\psi|W|\psi\rangle=2(\mathcal{R}^{+}F-\operatorname{Re}\langle\psi|W\phi|\psi\rangle) \tag*{(49)}

Cauchy-Schwarz gives ∣⟨ψ∣Wϕ∣ψ⟩∣2≤⟨ψ∣W∣ψ⟩⟨ψ∣ϕWϕ∣ψ⟩=R+Tr⁡(Wϕψϕ)|\langle\psi|W\phi|\psi\rangle|^2\leq\langle\psi|W|\psi\rangle\langle\psi|\phi W\phi|\psi\rangle=\mathcal{R}^{+}\operatorname{Tr}(W\phi\psi\phi) for ψ:=∣ψ⟩⟨ψ∣\psi:=|\psi\rangle\langle\psi|, while trace-norm submultiplicativity gives

Tr⁡(Wϕψϕ)≤Tr⁡(∣Wϕψϕ∣)≤Tr⁡(∣ϕWϕ∣)Tr⁡(∣ϕψϕ∣)=Tr⁡(Wϕ)Tr⁡(ψϕ)≤F.(50)\operatorname{Tr}(W\phi\psi\phi)\leq\operatorname{Tr}(|W\phi\psi\phi|)\leq\operatorname{Tr}\left(\left|\sqrt{\phi}W\sqrt{\phi}\right|\right)\operatorname{Tr}\left(\left|\sqrt{\phi}\psi\sqrt{\phi}\right|\right)=\operatorname{Tr}(W\phi)\operatorname{Tr}(\psi\phi)\leq F. \tag*{(50)}

Thus, ⟨ψ∣ϕWϕ∣ψ⟩≤R+F\langle\psi|\phi W\phi|\psi\rangle\leq\sqrt{\mathcal{R}^{+}F}. We can then bound

∣∂cRW+(ψc)∣2∣c=0≥4(R+F−Re⁡⟨ψ∣Wϕ∣ψ⟩)2≥4(R+F−∣⟨ψ∣Wϕ∣ψ⟩∣)2≥4(R+F−R+F)2=4R+F(R+F−1)2,(51)\begin{aligned} \left|\partial_c\mathcal{R}^{+}_{W}(\psi_c)\right|^2\big|_{c=0}&\geq4(\mathcal{R}^{+}F-\operatorname{Re}\langle\psi|W\phi|\psi\rangle)^2\geq4(\mathcal{R}^{+}F-|\langle\psi|W\phi|\psi\rangle|)^2\\ &\geq4(\mathcal{R}^{+}F-\sqrt{\mathcal{R}^{+}F})^2=4\mathcal{R}^{+}F(\sqrt{\mathcal{R}^{+}F}-1)^2, \tag*{(51)} \end{aligned}

since R+F≥1\mathcal{R}^{+}F\geq1 implies R+F−Re⁡⟨ψ∣Wϕ∣ψ⟩≥R+F−R+F≥0\mathcal{R}^{+}F-\operatorname{Re}\langle\psi|W\phi|\psi\rangle\geq\mathcal{R}^{+}F-\sqrt{\mathcal{R}^{+}F}\geq0. Plugging this and the previous Fisher information result into the CR bound in Eq. (6) yields

Var⁡(W)≥4R+F(R+F−1)24F(1−F),(52)\operatorname{Var}(W)\geq\frac{4\mathcal{R}^{+}F(\sqrt{\mathcal{R}^{+}F}-1)^2}{4F(1-F)}, \tag*{(52)}

which after canceling terms gives the final form in Theorem 2. While this is not the tightest CR bound which can be placed on Var⁡(W)\operatorname{Var}(W) (which would come from optimizing over path-derivatives), it is the one we found which can be expressed in terms of standard, index-free quantities of ∣ψ⟩|\psi\rangle.

Moments

A number of resource measures can be described as the kthk^{\mathrm{th}} moment of some distribution, including purity, entanglement purity, and stabilizer moments. Moments are interesting in their own right, but are also useful as building blocks of entropy measures, such as stabilizer Rényi entropy [29, 30]. While they are not linear in state like robustness is, they can be written as the expectation value of a joint observable acting on kk state copies [31]. For standard purity PkP_k (entanglement purity EkE_k), the relevant measurement operator is the SWAP (partial SWAP) gate acting on k=2k=2 copies, while the kthk^{\mathrm{th}} stabilizer moment AkA_k is instead written in terms of Γ=2−n∑p⃗σp⃗⊗2k\Gamma=2^{-n}\sum_{\vec{p}}\sigma_{\vec{p}}^{\otimes2k} acting on 2k2k state copies [8, 30]. As these measurements do not necessarily estimate the quantity with minimal error, we turn to the CR bound once more to investigate if better measurements on kk state copies exist.

To approach this, we first define a formalized moment function FkF_k, acting on state ρ\rho and a (potentially ρ\rho-dependent) set of observables {Mj}j\{M_j\}_j as

Fk({Mj}j)=∑j⟨Mj⟩k.(53)F_k(\{M_j\}_j)=\sum_j\langle M_j\rangle^k. \tag*{(53)}

The previously mentioned examples are special cases of this function: the purity moment Pk=Tr⁡(ρk)=Fk({∣ψj⟩⟨ψj∣})P_k=\operatorname{Tr}(\rho^k)=F_k(\{|\psi_j\rangle\langle\psi_j|\}), where {∣ψj⟩}\{|\psi_j\rangle\} are eigenvectors of ρ\rho, the entanglement purity moment Ek=∑jλjk=Fk({∣uj⟩⟨uj∣⊗I})E_k=\sum_j\lambda_j^k=F_k(\{|u_j\rangle\langle u_j|\otimes\mathbb{I}\}), for Schmidt decomposition ∣ψ⟩=∑jλj ∣ujvj⟩|\psi\rangle=\sum_j\sqrt{\lambda_j}\,|u_jv_j\rangle when ρ\rho is pure, and the stabilizer moment Ak=2−n∑p⟨σp⟩2k=2−nF2k({σp})A_k=2^{-n}\sum_p\langle\sigma_p\rangle^{2k}=2^{-n}F_{2k}(\{\sigma_p\}), for nn-qubit Pauli strings σp\sigma_p. Thus by bounding the variance of an estimator of FkF_k, we can determine the difficulty of estimating these functions.

Unlike in previous cases, we start from the symmetric log-derivative. Consider a state ρc\rho_c perturbed by the log-derivative Lc=2(Gk−Fk)\mathcal{L}_c=2(G_k-F_k), so that ∂cρc∣c=0={Gk,ρc}−2Fkρc\partial_c\rho_c|_{c=0}=\{G_k,\rho_c\}-2F_k\rho_c and Var⁡(Lc)=4Var⁡(Gk)\operatorname{Var}(\mathcal{L}_c)=4\operatorname{Var}(G_k). For a given set of observables {Mj}\{M_j\}, which we will hold fixed, define Gk=∑j⟨Mj⟩k−1MjG_k=\sum_j\langle M_j\rangle^{k-1}M_j, so that Fk=⟨Gk⟩F_k=\langle G_k\rangle. Then

∂cFk({Mj}j)∣c=0=k∑j⟨Mj⟩k−1Tr⁡(Mj∂cρc)∣c=0=kTr⁡(Gk∂cρc)∣c=0=kTr⁡(Gk{Gk,ρ}−2GkρFk)=2k(⟨Gk2⟩−Fk2)=2kVar⁡(Gk).(54)\begin{aligned} \partial_cF_k(\{M_j\}_j)\big|_{c=0}&=k\sum_j\langle M_j\rangle^{k-1}\operatorname{Tr}(M_j\partial_c\rho_c)\big|_{c=0}=k\operatorname{Tr}(G_k\partial_c\rho_c)\big|_{c=0}\\ &=k\operatorname{Tr}\left(G_k\{G_k,\rho\}-2G_k\rho F_k\right)=2k(\langle G_k^2\rangle-F_k^2)=2k\operatorname{Var}(G_k). \tag*{(54)} \end{aligned}

When considering kk state copies ρc⊗k\rho_c^{\otimes k}, it follows from the product rule that the full log-derivative is Lc,k=∑iLc,Si\mathcal{L}_{c,k}=\sum_i\mathcal{L}_{c,S_i}, the sum of log-derivatives acting on each system copy SiS_i. Given that ⟨Lc,SiLc,Sj⟩=⟨Lc2⟩δij\langle\mathcal{L}_{c,S_i}\mathcal{L}_{c,S_j}\rangle=\langle\mathcal{L}_c^2\rangle\delta_{ij}, we find that

Var⁡(Lc,k)=∑ij⟨Lc,SiLc,Sj⟩=kVar⁡(Lc)=4kVar⁡(Gk).(55)\operatorname{Var}(\mathcal{L}_{c,k})=\sum_{ij}\langle\mathcal{L}_{c,S_i}\mathcal{L}_{c,S_j}\rangle=k\operatorname{Var}(\mathcal{L}_c)=4k\operatorname{Var}(G_k). \tag*{(55)}

Then for some operator MM acting on kk state copies with Fk=Tr⁡(Mρc⊗k)F_k=\operatorname{Tr}(M\rho_c^{\otimes k}), CR tells us that

Var⁡(M)≥∣∂cFk∣2Var⁡(Lc,k)=4k2Var⁡(Gk)24kVar⁡(Gk)=kVar⁡(Gk).(56)\operatorname{Var}(M)\ge\frac{\lvert\partial_cF_k\rvert^2}{\operatorname{Var}(\mathcal{L}_{c,k})}=\frac{4k^2\operatorname{Var}(G_k)^2}{4k\operatorname{Var}(G_k)}=k\operatorname{Var}(G_k). \tag*{(56)}

While we can write Var⁡(Gk)=⟨Gk2⟩−Fk2=∑ij⟨Mi⟩k−1⟨Mj⟩k−1Cov⁡(Mi,Mj)\operatorname{Var}(G_k)=\langle G_k^2\rangle-F_k^2=\sum_{ij}\langle M_i\rangle^{k-1}\langle M_j\rangle^{k-1}\operatorname{Cov}(M_i,M_j), it is hard to cleanly solve for the ⟨Gk2⟩\langle G_k^2\rangle term in general. When 0≤Mi≤I0\le M_i\le\mathbb{I} we can upper bound ∣⟨MiMj⟩∣≤1\lvert\langle M_iM_j\rangle\rvert\le1 and know ∣⟨Mi⟩∣=⟨Mi⟩\lvert\langle M_i\rangle\rvert=\langle M_i\rangle, so ⟨Gk2⟩≤Fk2k−1\langle G_k^2\rangle\le F_k^{2k-1} by Cauchy-Schwarz. Moreover, when {Mi}\{M_i\} are orthogonal operators so that Tr⁡(MiMj)=δij\operatorname{Tr}(M_iM_j)=\delta_{ij}, we can bound ⟨Gk2⟩≤Tr⁡(Gk2)=F2k−2\langle G_k^2\rangle\le\operatorname{Tr}(G_k^2)=F_{2k-2}. This is the relevant case for AkA_k, up to a factor of 2n2^n since Tr⁡(σpσp′)=2nδpp′\operatorname{Tr}(\sigma_p\sigma_{p'})=2^n\delta_{pp'}. In particular, when {Mi}\{M_i\} form a set of orthogonal projectors (as they do for PkP_k, EkE_k), ⟨MiMj⟩=⟨Mi⟩δij\langle M_iM_j\rangle=\langle M_i\rangle\delta_{ij}, ⟨Gk2⟩=F2k−1\langle G_k^2\rangle=F_{2k-1} exactly, and Var⁡(Gk)=F2k−1−Fk2\operatorname{Var}(G_k)=F_{2k-1}-F_k^2. When ∣ψ⟩\lvert\psi\rangle is an eigenvector of GkG_k with maximal eigenvalue, we have Gk∣ψ⟩=Fk∣ψ⟩G_k\lvert\psi\rangle=F_k\lvert\psi\rangle. In this case Var⁡(Gk)=⟨Gk2⟩−Fk2=0\operatorname{Var}(G_k)=\langle G_k^2\rangle-F_k^2=0, so FkF_k can be measured with zero variance by measuring GkG_k.

Bound Tightness Examples

Several families of states exist for which the optimal witness can be found explicitly, allowing us to test the tightness of the bounds in Sections III and IV.

Mixed State Robustness

While the bound in Theorem 1 approaches triviality for nearly pure states, it is tight in various cases, such as for robustness of entanglement of Werner states, dephased maximally entangled states, and isotropic states. We will focus on the third example here. Isotropic states are defined as [32]:

ρf=fP+1−fd2−1(I−P)P=∣Φ⟩⟨Φ∣,(57)\rho_f=fP+\frac{1-f}{d^2-1}(\mathbb{I}-P)\qquad P=\lvert\Phi\rangle\langle\Phi\rvert, \tag*{(57)}

for maximally entangled state ∣Φ⟩=1d∑i∣ii⟩\lvert\Phi\rangle=\frac{1}{\sqrt{d}}\sum_i\lvert ii\rangle, dd the dimension of a single subsystem, and 1d<f<1\frac{1}{d}<f<1. With our normalization, the robustness of bipartite entanglement is R(ρf)=2df−1\mathcal{R}(\rho_f)=2df-1 [32]. Any separable (i.e., free) state ϕ\phi can be decomposed as ϕ=∑ipi∣aibi⟩⟨aibi∣\phi=\sum_i p_i\lvert a_i b_i\rangle\langle a_i b_i\rvert for ∑ipi=1\sum_i p_i=1, so one has

Tr⁡(Pϕ)=∑ipid∣∑j⟨jj∣aibi⟩∣2≤∑ipid∑j∣⟨j∣ai⟩∣2∑k∣⟨k∣bi⟩∣2=1d∑ipi=1d,(58)\operatorname{Tr}(P\phi)=\sum_i\frac{p_i}{d}\left\lvert\sum_j\langle jj\vert a_i b_i\rangle\right\rvert^2\le\sum_i\frac{p_i}{d}\sum_j\lvert\langle j\vert a_i\rangle\rvert^2\sum_k\lvert\langle k\vert b_i\rangle\rvert^2=\frac{1}{d}\sum_i p_i=\frac{1}{d}, \tag*{(58)}

where the inequality comes from Cauchy-Schwarz, and the following equality from the fact that ∑i∣i⟩⟨i∣=I\sum_i\lvert i\rangle\langle i\rvert=\mathbb{I}. Thus, the witness W∗=2dP−IW^*=2dP-\mathbb{I} is feasible, as ∣Tr⁡(W∗ϕ)∣≤1\lvert\operatorname{Tr}(W^*\phi)\rvert\le1 for separable ϕ\phi. Moreover, one can verify that Tr⁡(W∗ρf)=2df−1\operatorname{Tr}(W^*\rho_f)=2df-1, so W∗W^* is optimal. One can then explicitly calculate Var⁡(W∗)=4d2f(1−f)\operatorname{Var}(W^*)=4d^2f(1-f). Turning to the bound in Theorem 1, we can use the fact that PP, I−P\mathbb{I}-P are orthogonal projectors and ρf\rho_f is full rank to calculate ρf−1\rho_f^{-1}, and pick the optimal separable state ϕ=ρ1/d\phi=\rho_{1/d}. Substituting this in and simplifying reveals that the bound in Theorem 1 is equal to Var⁡(W∗)\operatorname{Var}(W^*). Thus, for this class of states the bound is tight.

Pure-state Generalized Robustness

We consider the case of generalized robustness of coherence for simplicity. Equivalently, the results hold for generalized robustness of bipartite entanglement, which has similar properties since the Schmidt decomposition acts as an orthonormal basis (as opposed to standard robustness of coherence and entanglement, which do not behave similarly). As previously mentioned, Theorem 3 yields a tight lower bound in this case, since a target state ∣ψ⟩=∑i=1rci∣ϕi⟩\lvert\psi\rangle=\sum_{i=1}^r c_i\lvert\phi_i\rangle for ∣ϕi⟩∈F\lvert\phi_i\rangle\in\mathcal{F} admits the variance-minimizing optimal witness W∗=∣v⟩⟨v∣W^*=\lvert v\rangle\langle v\rvert, with ∣v⟩=∑i=1r∣ϕi⟩\lvert v\rangle=\sum_{i=1}^r\lvert\phi_i\rangle. This witness has variance Var⁡(W∗)=R+(r−R+)\operatorname{Var}(W^*)=\mathcal{R}^{+}(r-\mathcal{R}^{+}), equal to the bound since θf=π2\theta_f=\frac{\pi}{2} and Fr=1rF_r=\frac{1}{r} for coherence and bipartite entanglement. The other generalized robustness bounds in this paper are not always tight for coherence, but are worth comparing since they perform better in other regimes, such as for magic where the bound in Theorem 3 becomes trivial whenever R+>2\mathcal{R}^{+}>\sqrt{2}.

Low-resource state

Let ∣ψϵ−⟩|\psi^-_\epsilon\rangle be a nearly free state, defined as ∣ψϵ−⟩=1−ϵ2∣ϕ1⟩+ϵ∑i=2rci∣ϕi⟩|\psi^-_\epsilon\rangle=\sqrt{1-\epsilon^2}|\phi_1\rangle+\epsilon\sum_{i=2}^{r}c_i|\phi_i\rangle, where we take ci>0c_i>0 and ∑i=2rci2=1\sum_{i=2}^{r}c_i^2=1 without loss of generality. Note that orthogonality of {∣ϕi⟩}i\{|\phi_i\rangle\}_i means the decomposition is unique, and thus optimal for pure-state extent ξ\xi [Eq. (30)]. Let A=∑i=2rci≤r−1A=\sum_{i=2}^{r}c_i\leq\sqrt{r-1}, with equality when all ci=1/r−1c_i=1/\sqrt{r-1}. For states ∣ψϵ−⟩|\psi^-_\epsilon\rangle, R+=ξ=(1−ϵ2+Aϵ)2\mathcal{R}^{+}=\xi=(\sqrt{1-\epsilon^2}+A\epsilon)^2 and F=1−ϵ2F=1-\epsilon^2. Let CR2\mathrm{CR}_2, Geo1\mathrm{Geo}_1, and Geo2\mathrm{Geo}_2 denote the variance bounds in Theorems 2, 3, and 4, respectively, and Vmin⁡V_{\min} the minimum achievable variance by an optimal witness. Plugging in the resolution of R+\mathcal{R}^{+}, FF, FrF_r and expanding to first order in ϵ\epsilon yields

Vmin⁡=Geo1=(r−1)+2A(r−2)ϵ+O(ϵ2),Geo2=(r−1)+2r−1(2Ar−1−r)ϵ+O(ϵ2),CR2=A2+2A(A2−1)ϵ+O(ϵ2).(59)\begin{aligned} V_{\min}&=\mathrm{Geo}_1=(r-1)+2A(r-2)\epsilon+O(\epsilon^2),\\ \mathrm{Geo}_2&=(r-1)+2\sqrt{r-1}(2A\sqrt{r-1}-r)\epsilon+O(\epsilon^2),\\ \mathrm{CR}_2&=A^2+2A(A^2-1)\epsilon+O(\epsilon^2). \tag*{(59)} \end{aligned}

Geo2\mathrm{Geo}_2 is tight in the limit ϵ→0\epsilon\to0, and differs by of Vmin⁡−Geo2=2r(r−1−A)ϵ+O(ϵ2)V_{\min}-\mathrm{Geo}_2=2r(\sqrt{r-1}-A)\epsilon+O(\epsilon^2) otherwise. Meanwhile, CR2\mathrm{CR}_2 has an O(1)O(1) difference in general, and has an optimality ratio of CR2/Vmin⁡=A2/(r−1)+O(ϵ)\mathrm{CR}_2/V_{\min}=A^2/(r-1)+O(\epsilon). This ratio goes to one when ci=1/r−1 ∀ic_i=1/\sqrt{r-1}\ \forall i, in which case CR2\mathrm{CR}_2 is tight for all ϵ\epsilon. Intuitively, the O(1)O(1) difference in CR2\mathrm{CR}_2 comes from the fact that we took a CR bound in the direction ∣ϕ⟩−∣ψϵ−⟩|\phi\rangle-|\psi^-_\epsilon\rangle rather than the optimal ∣ω⟩−∣ψϵ−⟩|\omega\rangle-|\psi^-_\epsilon\rangle for optimal witness state ∣ω⟩|\omega\rangle. Taking the latter in fact yields Eq. (27), which is tight but ∣ω⟩|\omega\rangle-dependent.

High-resource state

Let ∣ψϵ+⟩|\psi^+_\epsilon\rangle be a nearly-maximally-resourceful state, defined (for coherence, or bipartite entanglement via the Schmidt decomposition) as ∣ψϵ+⟩=1−ϵ2∣+r⟩+ϵ∣u⟩|\psi^+_\epsilon\rangle=\sqrt{1-\epsilon^2}|+_r\rangle+\epsilon|u\rangle, where ∣+r⟩=1r∑i=1r∣ϕi⟩|+_r\rangle=\frac{1}{\sqrt{r}}\sum_{i=1}^{r}|\phi_i\rangle and ∣u⟩|u\rangle is some normalized vector obeying ⟨+r∣u⟩=0\langle+_r|u\rangle=0. Let umax⁡=max⁡iRe⁡[ui]u_{\max}=\max_i\operatorname{Re}[u_i] be the maximal real coefficient of ∣u⟩|u\rangle in the free state basis, and usum=∑i=1rRe⁡[ui]2u_{\mathrm{sum}}=\sum_{i=1}^{r}\operatorname{Re}[u_i]^2 for convenience. In this case we have R+=ξ=r(1−usumϵ2)+O(ϵ3)\mathcal{R}^{+}=\xi=r(1-u_{\mathrm{sum}}\epsilon^2)+O(\epsilon^3) and F=1r+2umax⁡rϵ+O(ϵ2)F=\frac{1}{r}+\frac{2u_{\max}}{\sqrt{r}}\epsilon+O(\epsilon^2). Plugging these into the variance bounds gives

Vmin⁡=Geo1=r2usumϵ2+O(ϵ3),Geo2=r3umax⁡2r−1ϵ2+O(ϵ3),CR2=r3umax⁡2r−1ϵ2+O(ϵ3).(60)\begin{aligned} V_{\min}&=\mathrm{Geo}_1=r^2u_{\mathrm{sum}}\epsilon^2+O(\epsilon^3),\\ \mathrm{Geo}_2&=\frac{r^3u_{\max}^2}{r-1}\epsilon^2+O(\epsilon^3),\\ \mathrm{CR}_2&=\frac{r^3u_{\max}^2}{r-1}\epsilon^2+O(\epsilon^3). \tag*{(60)} \end{aligned}

where CR2\mathrm{CR}_2 and Geo2\mathrm{Geo}_2 differ by O(ϵ3)O(\epsilon^3) terms. Evidently, both have an optimality ratio of rumax⁡2(r−1)usum+O(ϵ)\frac{ru_{\max}^2}{(r-1)u_{\mathrm{sum}}}+O(\epsilon), which achieves one when umax⁡u_{\max} takes its largest possible value, umax⁡=r−1ru_{\max}=\sqrt{\frac{r-1}{r}}.

State family with tight bounds

In both cases, the limit where CR2=Geo2=Vmin⁡\mathrm{CR}_2=\mathrm{Geo}_2=V_{\min} occurs when ∣ψϵ−⟩|\psi^-_\epsilon\rangle and ∣ψϵ+⟩|\psi^+_\epsilon\rangle approach a family of states for which all bounds considered here are equal and optimal. That family of states is given by ∣ψf⟩=f∣ϕ1⟩+1−f2∣+2,r⟩|\psi_f\rangle=f|\phi_1\rangle+\sqrt{1-f^2}|+_{2,r}\rangle for ∣+2,r⟩:=1r−1∑i=2r∣ϕi⟩|+_{2,r}\rangle:=\frac{1}{\sqrt{r-1}}\sum_{i=2}^{r}|\phi_i\rangle and f≥1/rf\geq1/\sqrt{r}. In this case R+=(f+(r−1)(1−f2))2\mathcal{R}^{+}=(f+\sqrt{(r-1)(1-f^2)})^2 and F=f2F=f^2. Calculating all bounds eventually yields the same result for this family, though the result itself is unenlightening.

References

  1. [1]E. Chitambar and G. Gour, Quantum resource theories, Rev. Mod. Phys. 91, 025001 (2019).DOI
  2. [2]F. G. S. L. Brandão and G. Gour, Reversible framework for quantum resource theories, Phys. Rev. Lett. 115, 070503 (2015).
  3. [3]M. Howard and E. Campbell, Application of a resource theory for magic states to fault-tolerant quantum computing, Phys. Rev. Lett. 118, 090501 (2017).DOI
  4. [4]S. Bravyi, D. Browne, P. Calpin, E. Campbell, D. Gosset, and M. Howard, Simulation of quantum circuits by low-rank stabilizer decompositions, Quantum 3, 181 (2019).
  5. [5]J. R. Seddon, B. Regula, H. Pashayan, Y. Ouyang, and E. T. Campbell, Quantifying quantum speedups: Improved classical simulation from tighter magic monotones, PRX Quantum 2, 010345 (2021).DOI
  6. [6]S. F. Oliviero, L. Leone, A. Hamma, and S. Lloyd, Measuring magic on a quantum processor, npj Quantum Inf. 8, 148 (2022).DOI
  7. [7]T. Haug and M. S. Kim, Scalable measures of magic for quantum computers, PRX Quantum 4, 010301 (2023).DOI
  8. [8]T. Haug, S. Lee, and M. S. Kim, Efficient quantum algorithms for stabilizer entropies, Phys. Rev. Lett. 132, 240602 (2024).DOI
  9. [9]P. S. Tarabunga, E. Tirrito, M. C. Ba\u nuls, and M. Dalmonte, Nonstabilizerness via matrix product states in the pauli basis, Phys. Rev. Lett. 133, 010601 (2024).DOI
  10. [10]R. J. Garcia, G. Bhole, K. Bu, L. Chen, H. Arthanari, and A. Jaffe, Hardness of measuring magic resource, Phys. Rev. Res. 7, 033271 (2025).DOI
  11. [11]A. Gu, L. Leone, S. Ghosh, J. Eisert, S. F. Yelin, and Y. Quek, Pseudomagic quantum states, Phys. Rev. Lett. 132, 210602 (2024).DOI
  12. [12]L. Leone, J. Eisert, and S. F. E. Oliviero, The unbearable hardness of deciding about magic (2026), arXiv:2602.22330 [quant-ph].arxiv.org/abs/2602.22330
  13. [13]M. G. Genoni, P. Giorda, and M. G. A. Paris, Optimal estimation of entanglement, Phys. Rev. A 78, 032303 (2008).DOI
  14. [14]G. Vidal and R. Tarrach, Robustness of entanglement, Phys. Rev. A 59, 141 (1999).DOI
  15. [15]M. Steiner, Generalized robustness of entanglement, Phys. Rev. A 67, 054305 (2003).DOI
  16. [16]D. Cavalcanti and P. Skrzypczyk, Quantitative relations between measurement incompatibility, quantum steering, and nonlocality, Phys. Rev. A 93, 052112 (2016).DOI
  17. [17]C. Napoli, T. R. Bromley, M. Cianciaruso, M. Piani, N. Johnston, and G. Adesso, Robustness of coherence: An operational and observable measure of quantum coherence, Phys. Rev. Lett. 116, 150502 (2016).DOI
  18. [18]R. Takagi, B. Regula, K. Bu, Z.-W. Liu, and G. Adesso, Operational advantage of quantum resources in subchannel discrimination, Phys. Rev. Lett. 122, 140402 (2019).DOI
  19. [19]R. Takagi and B. Regula, General resource theories in quantum mechanics and beyond: Operational characterization via discrimination tasks, Phys. Rev. X 9, 031053 (2019).
  20. [20]B. Regula, K. Bu, R. Takagi, and Z.-W. Liu, Benchmarking one-shot distillation in general quantum resource theories, Phys. Rev. A 101, 062315 (2020).DOI
  21. [21]B. Regula, Convex geometry of quantum resource quantification, J. Phys. A: Math. Theor. 51, 045303 (2018).DOI
  22. [22]Z.-W. Liu and A. Winter, Many-body quantum magic, PRX Quantum 3, 020333 (2022).DOI
  23. [23]F. G. S. L. Brandão, Quantifying entanglement with witness operators, Phys. Rev. A 72, 022310 (2005).DOI
  24. [24]S. L. Braunstein and C. M. Caves, Statistical distance and the geometry of quantum states, Phys. Rev. Lett. 72, 3439 (1994).pubmed.ncbi.nlm.nih.gov/10056200
  25. [25]J. Liu, H. Yuan, X.-M. Lu, and X. Wang, Quantum Fisher information matrix and multiparameter estimation, J. Phys. A: Math. Theor. 53, 023001 (2019).DOI
  26. [26]D. Bertsekas, Nonlinear Programming, Athena Scientific, optimization and computation series (Athena Scientific, 1999) p. 717.
  27. [27]M. Tsang, F. Albarelli, and A. Datta, Quantum semiparametric estimation, Phys. Rev. X 10, 031023 (2020).DOI
  28. [28]N. Johnston, C.-K. Li, S. Plosker, Y.-T. Poon, and B. Regula, Evaluating the robustness of k-coherence and k-entanglement, Phys. Rev. A 98, 022328 (2018).DOI
  29. [29]E. Tirrito, P. S. Tarabunga, G. Lami, T. Chanda, L. Leone, S. F. E. Oliviero, M. Dalmonte, M. Collura, and A. Hamma, Quantifying nonstabilizerness through entanglement spectrum flatness, Phys. Rev. A 109, L040401 (2024).arxiv.org/abs/2304.01175
  30. [30]L. Leone, S. F. E. Oliviero, and A. Hamma, Stabilizer rényi entropy, Phys. Rev. Lett. 128, 050402 (2022).DOI
  31. [31]A. K. Ekert, C. M. Alves, D. K. L. Oi, M. Horodecki, P. Horodecki, and L. C. Kwek, Direct estimations of linear and nonlinear functionals of a quantum state, Phys. Rev. Lett. 88, 217901 (2002).pubmed.ncbi.nlm.nih.gov/12059503
  32. [32]M. A. Jafаrizadeh, M. Mirzaee, and M. Rezaee, Exact calculation of robustness of entanglement via convex semidefinite programming, Int. J. Quantum Inf. 3, 511 (2005).arxiv.org/abs/quant-ph/0608225

Paper details

vlnc:W_zd25umde
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

ChatGPT-6 Astra

View statement
Looking for related papers…
Contents