All-to-All Brownian Circuit Ensembles

• The paper shows that disorder averaging transforms replicated quantum evolution into an effective statistical-mechanics problem, enabling exact analytical results.
• It demonstrates that shallow, all-to-all circuit ensembles rapidly scramble dynamics and approach Haar randomness, forming unitary k-designs.
• Exact formulas for output moments and benchmarks are derived, clarifying noise sensitivity and sample complexity for near-term quantum hardware.

All-to-all Brownian circuit ensembles are continuous-time random quantum circuit models with time-dependent all-to-all interactions whose couplings are Gaussian white-noise variables. In these ensembles, every allowed interaction term can act between any pair of sites, or in fermionic variants between any allowed tuple of sites, while the disorder is renewed at each instant. The central technical advantage is that the disorder average over replicated evolution can be mapped to an effective statistical-mechanics problem, making the ensembles analytically tractable. In recent arXiv work, they serve both as solvable models for the approach to Haar randomness and unitary $k$-design formation, and as analytically controlled proxies for shallow fully connected random quantum circuits used in benchmarking noisy intermediate-scale quantum hardware (Jian et al., 2022, Bentsen et al., 21 May 2026).

1. Model definition and Brownian limit

For shallow, fully connected random quantum circuits, the all-to-all Brownian circuit ensemble is defined by a product of short-time unitary steps,

$$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$

with Gaussian white-noise couplings obeying

$$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$

The circuit depth is $\beta=Kdt$, and the continuum Brownian limit is taken as $dt\to 0$ with $K\to\infty$ at fixed $\beta$ (Bentsen et al., 21 May 2026).

A closely related formulation appears in the Brownian spin-cluster model, where

$$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$

with

$$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$

This is described as a continuously driven random circuit with time-dependent all-to-all interactions. The same paper also studies a Brownian SYK fermion cluster,

$$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$

with

$$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$0

which is again all-to-all, now with 4-body interactions among Majoranas (Jian et al., 2022).

The phrase “all-to-all” is therefore structural rather than merely geometric: every allowed interaction can connect any sites, and the coupling graph is effectively renewed at each time slice. In the benchmarking application, this structure is crucial because it produces rapid scrambling and mean-field behavior at large $$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$1, while in the design-theoretic application it yields analytically tractable dynamics.

2. Replica formulation and effective statistical mechanics

The key simplification is that Brownian disorder averaging converts a random-unitary problem into an effective statistical-mechanics problem. For the shallow all-to-all ensemble, the replicated disorder average takes the form

$$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$2

with effective Hamiltonian

$$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$3

where $$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$4 and $$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$5 denote forward and time-reversed replicas, and time reversal is implemented by

$$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$6

Because the couplings are all-to-all, the model is permutation-symmetric and amenable to large-$$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$7 saddle-point analysis. The effective Hamiltonian has a global $$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$8 symmetry and a replica symmetry $$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$9, which makes the spectrum tractable (Bentsen et al., 21 May 2026).

In the frame-potential analysis, the same logic yields

$$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$0

for the Brownian spin model. A spin coherent-state path integral then leads to bilocal Green’s functions

$$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$1

and the corresponding large-$$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$2 saddle-point problem. For Brownian SYK, the frame potential is written as a functional integral over $$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$3 and $$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$4, with the long-time behavior controlled by saddle points corresponding to pairings between forward and backward replicas (Jian et al., 2022).

The replicated theory is therefore not a secondary technicality but the central organizational principle of the subject. It is what makes closed-form output moments, asymptotic frame potentials, and benchmark statistics accessible.

3. Finite-depth output statistics and Porter–Thomas crossover

For a fixed bitstring $$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$5, the shallow-circuit analysis computes the probability moments

$$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$6

by a Choi–Jamiołkowski / replica rewrite followed by a saddle-point evaluation. The resulting exact formula is

$$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$7

with

$$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$8

where $$\mathbb{E}\!\left[J_{jk}^{\alpha\beta}(t)J_{j'k'}^{\alpha'\beta'}(t')\right] =\frac{J}{n\,dt}\, \delta_{jj'}\delta_{kk'}\delta^{\alpha\alpha'}\delta^{\beta\beta'}\delta_{tt'}.$$9 is the Hamming weight of $\beta=Kdt$0. In the deep-circuit limit $\beta=Kdt$1, these moments reduce to the Porter–Thomas moments, $\beta=Kdt$2 (Bentsen et al., 21 May 2026).

Using the moment-generating function and inverse Laplace transform, the same analysis reconstructs the depth-dependent probability distribution,

$\beta=Kdt$3

so the finite-depth ensemble behaves like a Porter–Thomas law with an effective dimension rescaling set by $\beta=Kdt$4. Averaging over bitstrings gives a depth-dependent family $\beta=Kdt$5 that approaches Porter–Thomas as $\beta=Kdt$6 (Bentsen et al., 21 May 2026).

This finite-depth crossover matters because shallow all-to-all circuits are not simply “miniature Haar” circuits. The outputs do not anticoncentrate at very shallow depth, and benchmark statistics are correspondingly depth dependent. At the same time, the same non-anticoncentration is used in the paper to argue robustness of nonlinear benchmarks against certain noise-induced spoofing strategies.

4. Nonlinear cross-entropy as a benchmark

Random circuit sampling remains a competitive framework for demonstrating quantum advantage on NISQ hardware, but existing benchmarks such as linear cross-entropy have been classically spoofed due to noise. The 2026 analysis addresses the regime of shallow-depth random quantum circuits, where sampling is described as plausibly classically intractable but inadequately characterized by earlier benchmarks (Bentsen et al., 21 May 2026).

The nonlinear benchmark is written as

$\beta=Kdt$7

where $\beta=Kdt$8 is the sampler’s actual output distribution and $\beta=Kdt$9 is the ideal probability. The replica identity

$dt\to 0$0

reduces the calculation to an analytic continuation of the computed moments. For a clean quantum sampler $dt\to 0$1, the result is

$dt\to 0$2

with $dt\to 0$3 and $dt\to 0$4 Euler’s constant. With single-qubit depolarizing noise at rate $dt\to 0$5,

$dt\to 0$6

and the replica continuation yields

$dt\to 0$7

Its variance is also obtained exactly: $dt\to 0$8

These expressions isolate a noise-sensitive signal term $dt\to 0$9. The paper emphasizes a hierarchy,

$K\to\infty$0

with the separation controlled by that term. In the shallow-depth regime, parametrized by

$K\to\infty$1

the signal scales as

$K\to\infty$2

while the estimated sample complexity obeys

$K\to\infty$3

Accordingly, nonlinear XEB is described as sample-efficient in the shallow all-to-all regime $K\to\infty$4, including constant depth and logarithmic depth, but inefficient at polynomial or linear depths where the signal is exponentially suppressed (Bentsen et al., 21 May 2026).

5. Heavy-output generation and binary classification

The same Brownian formalism is used to construct a heavy-output-generation classifier. The score is defined as

$K\to\infty$5

and its distribution $K\to\infty$6 is derived for the clean sampler and for the classical Harvard spoofer: $K\to\infty$7 The clean-versus-spoofer gap is

$K\to\infty$8

A threshold $K\to\infty$9 is then chosen so that scores above $\beta$0 are more likely from the quantum sampler than the spoofer (Bentsen et al., 21 May 2026).

If $\beta$1 independent samples are collected and one counts how many scores exceed threshold, standard concentration gives

$\beta$2

where

$\beta$3

From this, the paper concludes that $\beta$4 suffices to drive the failure probability below $\beta$5 (Bentsen et al., 21 May 2026).

The operational significance is that the classifier exploits a structural property of shallow all-to-all chaotic circuits: clean circuits preferentially generate a small set of heavy strings, whereas the classical spoofer does not. This does not replace the nonlinear cross-entropy benchmark; rather, it complements it with a binary decision procedure whose sample complexity is logarithmic at short depth.

6. Frame potential, $\beta$6-design formation, and complexity growth

The design-theoretic analysis centers on the $\beta$7th frame potential,

$\beta$8

For the Haar ensemble,

$\beta$9

which is the minimum possible value for ordinary unitary ensembles. The associated moment map is

$$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$0

and an ensemble is an approximate $$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$1-design if

$$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$2

A key inequality relates this to the frame potential: $$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$3 For fermionic models with parity symmetry, the corresponding minimum is modified to

$$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$4

These relations make the frame potential an operational diagnostic of Haar randomness, not merely a formal one (Jian et al., 2022).

For the Brownian spin model, the long-time behavior is controlled by permutation pairings between the $$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$5 forward and $$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$6 backward replicas. There are $$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$7 such pairings, and the spectrum around each pairing has a finite gap

$$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$8

independent of $$H(t)=\sum_{i<j,\alpha\beta}J_{ij\alpha\beta}(t)\,\sigma_{i\alpha}\sigma_{j\beta},$$9 at large $$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$0. The asymptotic frame potential behaves as

$$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$1

and a crude estimate gives

$$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$2

Combining this with the diamond-norm bound yields an $$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$3-approximate $$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$4-design time of the form

$$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$5

with an explicit bound

$$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$6

For Brownian SYK, the asymptotics are

$$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$7

and the derived approximate-design time is

$$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$8

The same paper also treats time-independent Hamiltonians perturbed by weak Brownian noise,

$$\mathbb E\!\left[J_{ij\alpha\beta}(t)J_{i'j'\alpha'\beta'}(0)\right] =\delta_{ii'}\delta_{jj'}\delta_{\alpha\alpha'}\delta_{\beta\beta'}\,\delta(t)\,\frac{J}{N}.$$9

which leads to the effective non-Hermitian $$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$0-replica Hamiltonian

$$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$1

Under the assumption of quantum chaos, the only long-lived dark states are again replica pairings, and approximate $$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$2-design formation occurs on a timescale

$$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$3

The broader conclusion is that all-to-all Brownian models furnish explicit analytically tractable examples in which the circuit approaches Haar behavior on the $$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$4-copy sector after a time linear in $$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$5, and this is the sense in which they realize linear growth of circuit complexity (Jian et al., 2022).

7. Numerical correspondence, limitations, and neighboring Brownian frameworks

The shallow-circuit benchmarking paper explicitly treats the all-to-all Brownian ensemble as an analytically solvable proxy rather than a literal discrete circuit instance. Its appendix compares Brownian predictions with numerics for discrete all-to-all Haar-random two-qubit circuits and reports qualitative agreement and, at sufficiently large depth, quantitative agreement. In particular, the Brownian formulas

$$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$6

capture the $$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$7-scaling and noise dependence for depths $$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$8, whereas shallower depths $$H(t)=\sum_{i<j<k<l} J_{ijkl}(t)\,\psi_i\psi_j\psi_k\psi_l,$$9 show deviations because they are not yet fully in the $$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$00 regime. The comparison with 1D brickwork circuits shows larger deviations, which the paper uses to underscore that all-to-all connectivity is essential for close agreement between Brownian analysis and discrete-circuit numerics (Bentsen et al., 21 May 2026).

A common misconception is that Brownian methods merely provide asymptotic Haar results. The finite-depth moment formulas, the depth-dependent $$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$01, the explicit $$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$02 signal term, and the short-depth heavy-output classifier all show that the Brownian ensemble retains nontrivial shallow-depth structure. Another possible misconception is that “Brownian” denotes a single research direction. A distinct example is the study of integrable fishnet circuits with stochastic time rescaling, where Brownian behavior appears as Brownian soliton dynamics in a one-dimensional classical lattice model with commuting propagators and exact conservation laws. That framework is classical, integrable, and one-dimensional, rather than an all-to-all random quantum circuit ensemble (Krajnik et al., 2024).

Within quantum information and many-body theory, the specific role of all-to-all Brownian circuit ensembles is therefore sharply defined. They provide a solvable bridge between random circuit dynamics and measurable statistics: exact or asymptotic frame potentials for design formation, closed-form output moments for finite-depth sampling distributions, exact replica-trick expressions for nonlinear cross-entropy and its variance, and threshold-based heavy-output classification. Their usefulness is strongest precisely where all-to-all scrambling, permutation symmetry, and large-$$U=\prod_t U_t=\prod_t \exp\!\left[-i\sum_{j<k}\sum_{\alpha,\beta} J_{jk}^{\alpha\beta}(t)\,\sigma_j^\alpha\sigma_k^\beta\,dt\right],$$03 saddle structure coexist.