Quantum information · Theory

Near-Optimal Multiparameter Pure-State Metrology with a Constant Number of Bases

Bowen Xiao; Zhiyao Yang; Huangjun Zhu

arXiv:2611.17281v1 ↗PDF ↗Submitted 12 November 2026

Question
To estimate an unknown pure state of dimension \(d\), an experimenter fixes some measurement bases in advance and uses them at random. One basis extracts the full quantum Fisher information in \(d-1\) directions and nothing in the other \(d-1\), and measurement incompatibility caps the fraction that can be retained in every direction at one half. How many fixed bases guarantee a positive fraction, the Fisher floor, at every state and in every direction, independently of \(d\)?
Result
A constant number. With probability at least \(1-e^{-(d+18)/2}\), five independent Haar-random bases have a floor of at least \(c_5>4.8\times10^{-10}\) in every dimension; 71 bases give 1/10 and 271 give 1/4 with failure probability below \(10^{-5}\). About \(12\varepsilon^{-2}\ln(e/\varepsilon)\) bases give a floor of \(1/2-\varepsilon\), and in large dimension at least \(1/(4\varepsilon^2)\) bases of any kind are necessary. By contrast, a complete set of \(d+1\) mutually unbiased bases minimises the average anisotropy among families of its size but has floor exactly \(1/(d+1)\).

The paper in detail

For a pure state with \(2d-2\) real parameters, the quantum Fisher information \(J_\psi\) sets the best possible single-copy sensitivity in each direction, and a measurement extracts classical Fisher information \(I_{\mathcal M,\psi}\le J_\psi\). The normalised matrix \(J_\psi^{-1/2}I_{\mathcal M,\psi}J_\psi^{-1/2}\) has eigenvalues between 0 and 1, the fractions retained in its eigendirections. The Gill–Massar bound makes their sum at most \(d-1\), so on average and in the worst case at most half survives. The Fisher floor \(\eta(\mathcal M)\) is the smallest eigenvalue over all pure states.

A single basis is the extreme opposite of balanced: at a generic state it has \(d-1\) unit eigenvalues and \(d-1\) zeros, keeping full information about the populations and none about the relative phases. Mixing bases averages these projectors. The question is how many bases, fixed before the state is known, make the average bounded below everywhere. A complex projective 3-design would do it, but needs order \(d^2\) bases.

The answer separates average balance from uniform sensitivity. Mutually unbiased bases distribute information perfectly on average: among families of \(L\) bases they minimise the mean variance of the Fisher eigenvalues. Yet the complete family has a direction with almost no information: at \((\ket{\varphi_0}+\ket{\varphi_1})/\sqrt2\) for two vectors of one basis, the tangent \((\ket{\varphi_0}-\ket{\varphi_1})/\sqrt2\) is seen only by that basis, because every other basis is unbiased with respect to both vectors, and the floor equals its weight \(1/(d+1)\).

Haar-random bases are worse on average, with expected mean anisotropy \(1/(4L)\), but have no such blind spot. Five already give a floor bounded away from zero in every dimension, with failure probability falling exponentially in \(d\), and the floor approaches \(1/2\) at rate about \(\sqrt{\ln L/L}\). A converse shows that the quadratic count in \(1/\varepsilon\) is necessary up to the logarithm, and that no finite family reaches exactly \(1/2\).

Open problems are explicit: whether four bases suffice, closing the logarithmic gap, explicit constructions, and computable certificates for the floor of a specific family below \(d+1\) bases. Numerics at \(d=32\) suggest that far fewer bases than the theorems require reach a given accuracy.

Transferable methods

  1. The Fisher operator as an average of tangent projectors

    To compare bases one needs the Fisher information in a form that adds across bases. On the real tangent space at \(\ket\psi\), each basis contributes an orthogonal projector, and the mixture's Fisher operator is their weighted average. The structure of these projectors explains both the half bound and the MUB blind spot.

    For a basis with amplitudes \(a_k=\langle\varphi_k|\psi\rangle\) and a unit tangent \(\ket\nu\perp\ket\psi\), the retained fraction is \[g_{\mathcal B}(\psi,\nu)=\sum_{k:a_k\neq0}\frac{[\mathrm{Re}(\bar a_kv_k)]^2}{\lvert a_k\rvert^2},\qquad v_k=\langle\varphi_k|\nu\rangle,\] the squared length of \(\ket\nu\) projected on the real span of the phase-aligned vectors \((a_k/\lvert a_k\rvert)\ket{\varphi_k}\). Replacing \(\ket\nu\) by \(i\ket\nu\) projects on the complementary span, so \(g(\psi,\nu)+g(\psi,i\nu)\le1\). Averaging over bases, eigenvalues come in pairs \(\lambda,1-\lambda\) at regular states, and a floor of \(1/2-\varepsilon\) is equivalent to every eigenvalue lying within \(\varepsilon\) of \(1/2\).

    For complete MUBs, Cauchy–Schwarz and \(\sum_kx_{\ell k}=0\) give \(\sum_kx_{\ell k}^2/p_{\ell k}\ge2\sum_kx_{\ell k}^2\) with \(x_{\ell k}=\mathrm{Tr}(P_{\ell k}O)\), and the 2-design property gives \(\sum_{\ell,k}x_{\ell k}^2=\mathrm{Tr}\,O^2=2\). Hence every response is at least \(\min_\ell q_\ell\), and the state and tangent built from two vectors of the least-weighted basis attain it.

    Editorial check of Proposition 5 at \(d=5\): at exactly \(\ket{\psi_+}=(\ket0+\ket1)/\sqrt2\), omitting zero-probability outcomes, the spectrum is \(1/6\), \(5/12\) six times, and \(5/6\), as stated. This is a singular state: perturbing it slightly to a regular state turns three of the \(5/12\) into \(7/12\), restoring the pairing and the trace \(d-1\), while the floor \(1/6\) is unchanged. Random sampling of qubit states gives the minimum response \(1/3\) for the three Pauli bases.

    How to reuse it

    Whenever measurement settings are mixed, write each setting's Fisher information as a projector on a common tangent space. Worst-case guarantees then become questions about how well a random or structured family of subspaces covers every direction, a question for concentration and covering arguments rather than state-by-state optimisation.

  2. Mean anisotropy from a two-basis phase integral

    To show that MUBs minimise average imbalance, the authors need the state-averaged overlap of two bases' Fisher projectors. It depends on the bases only through the transition probabilities \(x_{jk}=\lvert\langle\varphi_j|\phi_k\rangle\rvert^2\), via one concave function.

    Lemma S1: \[\int\mathrm{tr}_{\mathbb R}\bigl(\mathsf P_{\psi,\mathcal B}\mathsf P_{\psi,\mathcal C}\bigr)\,d\psi=\frac{d-1}{2}+\frac12\Bigl[\sum_{j,k}f(x_{jk})-1\Bigr],\qquad f(x)=x+(1-x)\ln(1-x).\] It follows from the average of the doubled relative phase of two Gaussian amplitudes, which equals \(\max\{1-\kappa^2,0\}\). Since \(f\) is convex with \(\sum_kx_{jk}=1\), the sum is smallest when all \(x_{jk}=1/d\), which gives Proposition 1, with equality exactly for pairwise unbiased bases.

    Editorial checks by Monte Carlo at \(d=3\): two Haar bases give mean anisotropy \(0.128\pm0.004\) against \(1/(4L)=0.125\), and four MUBs give \(0.0221\) against the equality value \(0.0219\) of Proposition 1. The stated constants are consistent: \(c_5=\tfrac14e^{-6\cdot5^{3/4}}=4.84\times10^{-10}\), and the failure bounds at \(d=2\) are \(4.5\times10^{-5}\), \(6.8\times10^{-7}\) and \(6.1\times10^{-6}\) for 5, 71 and 271 bases.

    How to reuse it

    For any figure of merit that is a state average of products of two settings' contributions, reduce it to the pairwise transition probabilities and use convexity to find the extremal families. Then check the worst case separately: average-optimal families such as MUBs can have measure-zero blind spots.

Source: Eqs. (1)–(12); End Matter (Proposition 5, Eqs. (14)–(17)); SM Secs. S1.1–S1.2 (Eqs. (S2)–(S11)), S2.1 (Eqs. (S12)–(S21)) and S2.4 (Eqs. (S30)–(S32)). Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. Uniform sensitivity for full pure-state estimation does not need a number of settings that grows with dimension: a fixed set of bases, drawn once, keeps the inverse-information benchmark within a constant factor of the optimum, and maximum-likelihood estimation reaches mean infidelity \(O(d/N)\) uniformly. It also shows that the usual criterion for balanced measurements, mutual unbiasedness, optimises the average and not the worst case.

Limitations. The certified constants are tiny for small counts (\(c_5\approx5\times10^{-10}\), against numerical averages above \(7\times10^{-4}\)) and the counts for useful floors are large. The bases are random; explicit constructions with comparable guarantees are open, as is whether four bases suffice. The floor of a given family cannot be computed efficiently: numerical searches only give upper estimates, and the computable frame certificate needs at least \(d+1\) bases. A concurrent preprint by Li and Liu (arXiv:2611.00599) proves related results for large dimensions. The paper has no AI-use statement.

Description based on the full arXiv HTML of v1, read 2026-11-15: the main text, End Matter and Supplemental Sections S1, S2.1 and S2.4; the probabilistic proofs of Theorems 1–3 and Proposition 3 (S3–S8) were not read. The odd-prime MUB spectrum of Proposition 5 at \(d=5\), the qubit MUB floor, the Haar mean anisotropy \(1/(4L)\) and the MUB equality case of Proposition 1 at \(d=3\) (by Monte Carlo), and the stated constants and failure probabilities were recomputed.