Haar-Random Measurement Bases

• Haar-random measurement bases are orthonormal bases obtained by applying Haar-distributed unitaries to a fixed reference, ensuring uniform outcome probabilities.
• They enable advanced quantum protocols like Bayesian estimation, classical shadows, and randomized benchmarking through precise statistical moment analyses.
• Pseudorandom designs and circuit realizations serve as practical approximations to Haar randomness, facilitating efficient state discrimination and quantum geometry estimation.

Haar-random measurement bases are orthonormal bases obtained by applying a Haar-distributed unitary $U \sim \mathrm{Haar}(U(d))$ to a fixed reference basis of a $d$-dimensional Hilbert space. Equivalently, one may apply $U$ to the state and measure in the computational basis, or measure directly in the rotated basis $\{U|k\rangle\}_{k=1}^d$; for an input state $|\psi\rangle$, the outcome probabilities are $p_U(x)=|\langle x|U|\psi\rangle|^2$ (Quadeer, 2023, Brakerski et al., 2024). This construction is a standard symmetry-driven model for randomized measurements, and it appears in Bayesian state estimation, classical shadows, entropic uncertainty, randomized benchmarking, distinguishability problems, fermionic linear optics, and quantum-geometry estimation (Yang et al., 18 Jun 2026, Adamczak et al., 2014, Lu et al., 10 Sep 2025).

1. Definition on $U(d)$ and basic measurement equivalences

Let $U(d)$ be the group of $d\times d$ unitary matrices. The Haar measure $dU$ is the unique probability measure on $d$0 that is invariant under left and right multiplication: for any fixed $d$1, $d$2 for all integrable $d$3 (Tang et al., 2021). A Haar-random measurement basis is therefore the orthonormal basis formed by the columns of a Haar-random unitary, and the associated projective POVM is

$d$4

Measuring $d$5 with $d$6 gives

$d$7

(Quadeer, 2023).

A central first-moment identity is the Haar twirl

$d$8

which implies uniform expected output probabilities, $d$9, for pure inputs (Tang et al., 2021). In the language of randomized measurements, measuring after $U$0 in the computational basis is equivalent to measuring in the basis $U$1, so Haar-random measurement bases can be realized either by random basis rotation before measurement or by random evolution followed by a fixed measurement (Mehrani et al., 2024).

The same formalism underlies Haar-random measure-and-prepare channels. For $U$2 shots, with $U$3, the channel

$U$4

dephases the rotated state in the product computational basis, and this induced classical record is the starting point for several asymptotic discrimination problems (Marcinkowska et al., 29 May 2026).

2. Statistical laws and moment structure

The first and second moments of Haar-random measurement outcomes admit closed forms. In the tensor-network formulation, for a fixed state $U$5 and computational-basis projector $U$6, the outcome probability is $U$7. Using the first- and second-moment Haar twirls, one obtains

$U$8

$U$9

$\{U|k\rangle\}_{k=1}^d$0

and

$\{U|k\rangle\}_{k=1}^d$1

For a normalized pure state, these reduce to $\{U|k\rangle\}_{k=1}^d$2, $\{U|k\rangle\}_{k=1}^d$3, and $\{U|k\rangle\}_{k=1}^d$4 (Fukuda et al., 2019).

At the level of full outcome vectors, Haar averaging produces a Dirichlet–multinomial structure. In one discrimination model, if $\{U|k\rangle\}_{k=1}^d$5, then $\{U|k\rangle\}_{k=1}^d$6, and the aggregate histogram $\{U|k\rangle\}_{k=1}^d$7 of $\{U|k\rangle\}_{k=1}^d$8 measurements is uniformly distributed over weak compositions of $\{U|k\rangle\}_{k=1}^d$9 into $|\psi\rangle$0 parts: $|\psi\rangle$1 for every $|\psi\rangle$2 with $|\psi\rangle$3 (Marcinkowska et al., 29 May 2026). In a chaos-oriented formulation, measurement statistics in Haar-random bases obey the Porter–Thomas distribution,

$|\psi\rangle$4

with mean $|\psi\rangle$5 and variance $|\psi\rangle$6 (Cusumano et al., 31 Mar 2026).

A separate observable-based verification framework emphasizes a Dirichlet description of Haar-random expectation values. In that convention, for a Haar-random pure state in a fixed basis the rank-1 outcome probabilities are stated to follow a marginal

$|\psi\rangle$7

with mean $|\psi\rangle$8, and the rescaled variable $|\psi\rangle$9 approaches an exponential Porter–Thomas law at large $p_U(x)=|\langle x|U|\psi\rangle|^2$0 (Bonet-Monroig et al., 2024). This suggests that moment-based diagnostics depend not only on Haar invariance itself but also on the precise representation and normalization convention adopted in a given verification framework.

3. Designs, pseudorandomness, and approximate Haar behavior

A unitary $p_U(x)=|\langle x|U|\psi\rangle|^2$1-design is an ensemble whose $p_U(x)=|\langle x|U|\psi\rangle|^2$2-th moments match the Haar moments: $p_U(x)=|\langle x|U|\psi\rangle|^2$3 Analogously, a state $p_U(x)=|\langle x|U|\psi\rangle|^2$4-design matches Haar moments on complex projective space (Yang et al., 18 Jun 2026). The minimal requirement for uniform first moments is a unitary $p_U(x)=|\langle x|U|\psi\rangle|^2$5-design; in one photonic implementation this is described as “1-pad Haar-uniform randomness” (Tang et al., 2021).

The distinction between exact Haar randomness and lower-order designs is operationally significant. For shadow estimation, state $p_U(x)=|\langle x|U|\psi\rangle|^2$6-designs achieve the optimal worst-case shadow norm bound $p_U(x)=|\langle x|U|\psi\rangle|^2$7, whereas any state $p_U(x)=|\langle x|U|\psi\rangle|^2$8-design already suffices for average-case optimality, with mean squared shadow norm bounded by a universal constant and only $p_U(x)=|\langle x|U|\psi\rangle|^2$9 measurement bases required, in contrast to $U(d)$0 for worst-case optimality (Yang et al., 18 Jun 2026). For Bayesian estimation of mixed qubits, measurements defined via unitary $U(d)$1-designs closely approximate Haar-random measurements, while the Pauli group as a unitary $U(d)$2-design yields only a weak lower bound (Quadeer, 2023). A common misconception is therefore that exact Haar sampling is uniformly necessary; the cited results indicate that this is false for many average-case tasks, although higher-moment control remains necessary in worst-case settings.

Pseudorandom substitutes can reproduce selected Haar-random measurement statistics under restricted input models. One real-valued construction studies

$U(d)$3

where $U(d)$4 and $U(d)$5 are diagonal binary-phase unitaries and $U(d)$6 is a computational-basis permutation. Assuming quantum-secure one-way functions, this family is secure for non-adaptive orthogonal-inputs pseudorandomness, with trace-distance error

$U(d)$7

relative to Haar action on polynomially many orthogonal inputs with polynomial multiplicity (Brakerski et al., 2024). The simpler

$U(d)$8

suffices for orthogonal $U(d)$9-flat inputs. The same paper also records a limitation due to Haug, Bharti, and Koh: real-valued unitaries cannot be fully pseudorandom against entangled queries, so these constructions do not supply unrestricted Haar replacement (Brakerski et al., 2024).

4. Roles in estimation, shadows, and quantum geometry

In Bayesian state estimation, Haar-random bases provide a symmetry-adapted projective measurement model. For a uniform pure-state prior and one Haar-random basis measurement, the average fidelity of the Bayesian mean estimator is

$U(d)$0

and for $U(d)$1 IID Haar-random basis measurements the paper derives the upper-bound form

$U(d)$2

For mixed qubit ensembles, unitary $U(d)$3-designs numerically track Haar performance closely, whereas Pauli measurements are substantially weaker (Quadeer, 2023).

In classical shadows, Haar-random bases define a canonical randomized-measurement benchmark, but the optimal number of bases depends on the performance criterion. Worst-case optimal shadow estimation requires $U(d)$4 orthonormal bases, while average-case optimality requires only $U(d)$5 bases; any state $U(d)$6-design suffices for the latter (Yang et al., 18 Jun 2026). In particular, for a state $U(d)$7-design $U(d)$8 and a unitary $U(d)$9-design orbit, the mean squared shadow norm obeys

$d\times d$0

and for Haar-random pure-state fidelity estimation the sample complexity is $d\times d$1 (Yang et al., 18 Jun 2026).

Haar-random bases also connect classical and quantum Fisher information. For pure states in $d\times d$2, if $d\times d$3 is the classical Fisher information matrix obtained from projective measurement in a Haar-random basis and $d\times d$4 is the quantum Fisher information matrix, then

$d\times d$5

The elementwise variance is $d\times d$6, and the concentration bounds decay as $d\times d$7, supporting the use of a small number of random bases for QFIM approximation in high-dimensional settings (Lu et al., 10 Sep 2025).

5. Physical generation and circuit realizations

One route to approximate Haar-random measurement bases is stochastic photonic evolution. A two-dimensional stochastic quantum walk on an integrated photonic chip implements a piecewise random Hamiltonian

$d\times d$8

with unitary evolution

$d\times d$9

For a $dU$0 lattice ($dU$1), with $dU$2–$dU$3 cm, $dU$4 mm, and detuning amplitude $dU$5, the average output distribution converges toward uniformity, and a $dU$6D array outperforms a $dU$7D array of the same number of waveguides in convergence speed (Tang et al., 2021). Since measuring after the chip unitary is equivalent to measuring in the basis $dU$8, this realizes approximate Haar-random measurement bases at the level of first moments.

For small qubit systems, a fixed-depth randomized benchmarking protocol constructs Haar-random unitaries directly from native gates rather than from compiled $dU$9-design elements. For one qubit, the fixed-length universal circuit

$d$00

is sampled using

$d$01

For two qubits, a three-CZ template is combined with Weyl-distributed eigenvalue phases. Applying the compiled $d$02 immediately before computational-basis measurement realizes Haar-random measurement bases; one-qubit Bloch-sphere uniformity and two-qubit CUE-like eigenvalue statistics are used as validation (Mehrani et al., 2024).

In fermionic linear optics, Haar-random measurement bases can be generated exactly within active and passive FLO. The cited constructions produce random active and passive FLO circuits with optimal $d$03 depth and $d$04 gate count, using only $d$05 classical overhead. Active FLO samples Haar on $d$06; passive FLO samples Haar on $d$07 embedded as $d$08. These circuits furnish randomized Gaussian measurement bases relevant to fermionic classical shadows and benchmarking (Braccia et al., 30 May 2025).

6. Verification, distinguishability, and asymptotic phenomena

Observable-based verification can test whether a set of states or bases is compatible with Haar $d$09-moments. For a dataset $d$10 and observable $d$11, the average-randomness metric is

$d$12

where $d$13 is the Haar-predicted $d$14-th moment. The framework also gives a shot requirement

$d$15

for $d$16-verification, and extends the test by permutation and unitary conjugation of observables so that small $d$17 across the family is compatible with a $d$18-design (Bonet-Monroig et al., 2024).

At the level of induced classical data, Haar-random measurement bases can remain distinguishable depending on whether the same random basis is reused collectively or sampled independently across blocks. For aggregate histograms, the collective model gives the uniform composition law

$d$19

while the independent-block law is

$d$20

Their total variation distance obeys explicit asymptotics. For fixed $d$21 and large $d$22,

$d$23

while in the critical scaling regime $d$24 it converges to a Poisson-collision expression (Marcinkowska et al., 29 May 2026). This shows that “Haar-random measurement” is not a single observational object: distinguishability depends on whether one studies the underlying quantum channel, block-resolved classical records, or coarse-grained histograms.

Haar-random bases also exhibit strong asymptotic uncertainty behavior. For two generic orthogonal measurements related by a Haar-random unitary $d$25, the maximal entry satisfies

$d$26

and the paper proves that with high probability

$d$27

which is asymptotically stronger than the Maassen–Uffink bound in the Haar-typical regime (Adamczak et al., 2014). For many independent Haar bases, the average entropy lower bound scales as $d$28, establishing the Wehner–Winter conjecture $d$29 (Adamczak et al., 2014).

Finally, deviations from Haar moments can be diagnosed dynamically. In chaos probes, Clifford circuits reproduce Haar moments up through third order but differ at fourth order; T-doped random circuits interpolate from stabilizer bases toward Haar-random bases, and $d$30 doping layers suffice empirically for fourth-moment probes to approach Haar values (Cusumano et al., 31 Mar 2026). This reinforces a general point: agreement with Haar-random measurement statistics is task-dependent and moment-dependent, so “Haar-like” may mean first-moment uniformity, low-order design matching, or full higher-moment convergence depending on the application.