---
title: Projected Process Ensemble in Quantum Chaos
url: https://www.emergentmind.com/topics/projected-process-ensemble-ppe
type: topic
---

# Projected Process Ensemble in Quantum Chaos

Searching arXiv for recent papers on “Projected Process Ensemble” and closely related projected-ensemble literature.
The **Projected Process Ensemble (PPE)** is an ensemble-valued construction derived from a **pure process tensor** for a quantum many-body system under repeated local interventions. Fixing an orthogonal basis of multi-time interventions and projecting the process tensor onto that basis yields an ensemble of **pure conditional output states** on the remainder/output Hilbert space. In this formulation, the PPE organizes diagnostics of quantum chaos by statistical moment: the **first moment** reproduces standard process-level quantities such as quantum dynamical entropy and spatiotemporal entanglement, while **higher moments** resolve finer structure in the distribution of conditional output states and more sharply separate chaotic, integrable, free, and many-body localized dynamics [2502.13930].

## 1. Definition within the process-tensor formalism

The PPE is defined for a quantum stochastic process subjected to repeated local interventions at times $\vec t=\{t_1,\dots,t_{n_B}\}$. The process is encoded by a **pure process tensor**
\[
|\Upsilon\rangle=\sum_{\vec x} |\ups_{R|\vec x}\rangle\otimes |\vec x\rangle,
\]
where $|\vec x\rangle$ is a basis vector in the **butterfly space** $\mathcal H_B$ associated with an intervention sequence $\vec x=(x_1,\dots,x_{n_B})$, and $|\ups_{R|\vec x}\rangle$ is the corresponding unnormalized output state on the remainder/output space $R$ [2502.13930].

The intervention operators are taken to be local on the system,
\[
A_x=A_x^{(S)}\otimes \id^{(E)},
\]
with orthogonality
\[
\operatorname{tr}(A_x^\dagger A_y)=\delta_{xy},
\]
and completeness
\[
\sum_{x=1}^r A_x^\dagger A_x=\id_R.
\]
For a sequence $\vec x$, the output state is
\[
|\ups_{R|\vec x}\rangle=A_{x_{n_B}}(t_{n_B})\cdots A_{x_1}(t_1)|\psi_R\rangle,
\]
with Heisenberg-picture operators $A_{x_i}(t_i)=U^\dagger(t_i)A_{x_i}U(t_i)$ [2502.13930].

Projecting onto a fixed orthogonal basis of interventions $\mathbf X$ produces probabilities
\[
p_{\vec x}=\langle \ups_{R|\vec x}|\ups_{R|\vec x}\rangle
\]
and normalized conditional states
\[
|\tilde\ups_{R|\vec x}\rangle=\frac{1}{\sqrt{p_{\vec x}}}\,|\ups_{R|\vec x}\rangle.
\]
The **Projected Process Ensemble** is therefore
\[
\mathcal E=\{p_{\vec x},|\tilde\ups_{R|\vec x}\rangle\}_{|\vec x\rangle\in\mathbf X}.
\]

Its $k$-th moment is
\[
\ups_R^{(k)}=\sum_{\vec x}p_{\vec x}\left(|\tilde\ups_{R|\vec x}\rangle\langle \tilde\ups_{R|\vec x}|\right)^{\otimes k}.
\]
The first moment is the reduced state on $R$,
\[
\ups_R=\sum_{\vec x}p_{\vec x}\,|\tilde\ups_{R|\vec x}\rangle\langle \tilde\ups_{R|\vec x}|,
\]
while higher moments encode fluctuations and replica-resolved structure not visible in the reduced state alone [2502.13930].

## 2. First moment and established chaos diagnostics

A central result of the PPE framework is that several previously studied chaos quantifiers are already contained in the **first PPE moment**. Writing $\ups_B$ for the butterfly-space reduction of the pure process tensor, the **quantum dynamical entropy** used in the paper is the Rényi-2 entanglement across the $B:R$ cut,
\[
E(B:R)=S^{(2)}(\ups_B)=-\log\!\big[\operatorname{tr}(\ups_B^2)\big].
\]
Because the full process tensor is pure, one also has
\[
E(B:R)=S^{(2)}(\ups_R)=S^{(2)}(\ups_R^{(1)}).
\]
In this sense, the first PPE moment supplies the common reduced-state structure underlying the paper’s quantum dynamical entropy, butterfly flutter fidelity, and spatiotemporal entanglement [2502.13930].

The butterfly interpretation is operational. If orthogonal intervention histories produce nearly orthogonal future output states, then the process is highly sensitive to perturbations in its intervention history. In the paper’s language, large quantum dynamical entropy indicates that local perturbations in the past are amplified into distinguishable outputs.

The same first moment also generates **spatiotemporal entanglement** for mixed partitions of butterfly and output degrees of freedom. For cuts of the form $BR_1:R_2$,
\[
E(BR_1:R_2)=S^{(2)}\!\left(\operatorname{tr}_{R_1}[\ups_R^{(1)}]\right).
\]
This places quantum dynamical entropy and spatiotemporal entanglement inside a single moment hierarchy. The relation to the original **Alicki–Fannes dynamical entropy** is, however, not an identity: the latter uses von Neumann entropy, a supremum over partitions of unity, and an equilibrium-state setting, whereas the PPE construction fixes a local orthogonal intervention basis and uses Rényi-2 entropy. The connection is therefore a unification at the level of process-tensor structure rather than exact equivalence [2502.13930].

## 3. Higher moments and the entanglement structure of chaos

The principal novelty of the PPE is that it treats chaos as a property of the **full distribution of conditional output states**, not only of their average. For a bipartition $R=R_1R_2$, the paper studies the Rényi-2 entanglement of each conditional output state,
\[
S_{\vec x}=-\log\!\left[\operatorname{tr}(\ups_{R_2|\vec x}^2)\right],
\qquad
\ups_{R_2|\vec x}=\operatorname{tr}_{R_1}\!\left[|\tilde\ups_{R|\vec x}\rangle\langle\tilde\ups_{R|\vec x}|\right].
\]
The PPE mean entanglement is
\[
\langle S_{\vec x}\rangle_{\mathcal E}=\sum_{\vec x}p_{\vec x}S_{\vec x},
\]
and its standard deviation is
\[
\Delta_{\mathcal E}S_{\vec x}=
\sqrt{\langle S_{\vec x}^2\rangle_{\mathcal E}-\langle S_{\vec x}\rangle_{\mathcal E}^2}.
\]
Because the Rényi-2 entropy of a conditional output state requires two copies, the ensemble-averaged entanglement probes the **second moment** of the PPE, while the variance probes the **fourth moment** [2502.13930].

This hierarchy resolves a limitation of first-moment probes. The paper emphasizes that quantum dynamical entropy can become large even in non-chaotic settings, including free fermions and Lindblad–Bernoulli shifts, and that spatiotemporal entanglement, while more discriminating, still does not cleanly separate chaotic from interacting-integrable dynamics in large systems. Higher moments instead test whether **almost every intervention history** generates output states with nearly maximal entanglement, and whether the distribution over histories narrows with system size.

Numerically, the paper finds a characteristic hierarchy. Under deterministic interventions, **chaotic models** show mean entanglement close to a Haar/Page-like value together with small variance. **Interacting integrable** models show somewhat lower mean entanglement and larger variance. **Free-fermion** and **many-body localized** models show strongly lower mean entanglement and much broader distributions. Most notably, finite-size scaling reveals that the standard deviation, or coefficient of variation, of PPE entanglement **decreases exponentially with system size** in chaotic systems—especially clearly in the chaotic XXZ chain—while this narrowing is absent or much weaker in integrable, free, and localized models [2502.13930].

## 4. Dynamical regimes and models studied

The PPE was introduced and tested in a set of one-dimensional many-body models preserving $U(1)$ particle number and initialized in the Néel state
\[
|\psi_0\rangle=|01\rangle^{\otimes L/2}.
\]
The numerical study covers an **interacting-integrable XXZ chain**
\[
H_{\mathrm{XXZ}}=J\sum_{i=1}^L(\sigma_i^x\sigma_{i+1}^x+\sigma_i^y\sigma_{i+1}^y+\Delta\sigma_i^z\sigma_{i+1}^z),
\]
with $J=1$ and $\Delta=0.55$; a **chaotic XXZ variant** with next-nearest-neighbor perturbation and an edge field, using $h=0.6$ and $g=0.1$; an **interacting Aubry–André** model at weak quasiperiodic potential $\lambda=1$; the same model in a **many-body localized** regime at $\lambda=5$; and a **free-fermion** model with nearest-neighbor hopping [2502.13930].

Within these models, the first PPE moment reproduces the familiar pattern that chaotic systems approach the Haar benchmark for quantum dynamical entropy,
\[
\langle E(B:R)\rangle_{\mathrm{Haar}}=n_B\log d_S
\qquad
(n_B\ll \log d_R),
\]
while interacting-integrable systems can approach the same growth with increasing size, obscuring a sharp distinction. For spatiotemporal entanglement the relevant Haar estimate is
\[
\langle E(BR_1:R_2)\rangle_{\mathrm{Haar}}
\approx n_B\log d_S+\log d_{R_1},
\]
which suppresses noninteracting and localized behavior more effectively but still does not decisively isolate chaos.

The higher-moment PPE observables provide that separation. This suggests that chaos, in the paper’s sense, is not merely rapid growth of process entropy but concentration of the conditional output-state ensemble around highly entangled, Haar-like outputs. A plausible implication is that the PPE detects a stronger notion of spatiotemporal complexity than first-moment process entropies alone.

## 5. Intervention basis, Haar benchmark, and computational structure

The higher moments of the PPE are explicitly **basis dependent**. The paper studies two intervention bases. The first consists of **deterministic interventions**, proportional to local unitaries; for spins these are
\[
\left\{\frac{\id_1}{\sqrt 2},\frac{\sigma_1^z}{\sqrt 2}\right\}.
\]
The second consists of **non-deterministic interventions**, namely local projectors,
\[
\{|0\rangle\langle 0|,\ |1\rangle\langle 1|\}.
\]
Under deterministic interventions all trajectories have equal weight,
\[
p_{\vec x}=1/d_S^{n_B},
\]
which makes comparison to Haar-random pure-state ensembles especially direct. Under non-deterministic interventions the $p_{\vec x}$ are genuine outcome probabilities and measurement backaction alters the output states. The paper reports reduced mean entanglement, larger variance, weaker agreement with Haar, and stronger non-ergodic signatures in this measurement-like basis, connecting the PPE to monitored many-body dynamics and measurement-induced phase-transition phenomenology [2502.13930].

A central analytical benchmark is the independently Haar-random process. For such dynamics, the ensemble-averaged $k$-th PPE moment is
\[
\langle \ups_R^{(k)}\rangle_{\mathrm{Haar}}
=
\frac{\sum_{\pi\in S_k}P_\pi}
{d_R(d_R+1)\cdots(d_R+k-1)},
\]
with $P_\pi$ the permutation representation of $S_k$. This is precisely the moment structure of the Haar ensemble of random pure states. In effect, projection onto a fixed intervention basis followed by independent Haar averaging at each time step erases process-specific causal structure and leaves an ordinary random-state ensemble benchmark [2502.13930].

Operationally, PPE computation proceeds by fixing an initial state, choosing intervention times and an orthogonal local intervention basis, evaluating the conditional output states and their probabilities for every sequence, constructing $\ups_R^{(k)}$, and then computing entanglement observables of the conditional states. The study uses exact diagonalization in a fixed $U(1)$ sector. The paper also notes an experimental limitation: brute-force sampling of the full PPE requires exponentially many trajectories in the number of intervention times $n_B$ because of postselection over intervention sequences [2502.13930].

## 6. Relation to projected-state ensembles, adjacent usages, and limitations

The PPE belongs to a broader projected-ensemble program in quantum many-body dynamics, but it is distinct from the **single-time projected ensemble of subsystem states** developed in the context of deep thermalization. In that literature, measuring the complement of a subsystem yields an ensemble of conditional pure states whose higher moments diagnose approach to Haar statistics or more general state-design structure. Dual-unitary circuits can produce exact state designs for all moments at the same time, whereas more generic random circuits display design times with logarithmic $k$-dependence; these results motivate the PPE’s emphasis on moment hierarchies, but remain state-ensemble rather than process-ensemble statements [2204.13657; 2402.16939].

The PPE of process tensors is therefore new in a specific sense: it is **multi-time and spatiotemporal**, it is obtained by projecting in the **butterfly/intervention space** rather than directly on output subsystems, and it is designed to interpolate between established first-moment process diagnostics and new higher-moment probes of conditional-output complexity [2502.13930].

Terminology is not uniform across contemporary literature. In the quantum-scrambling literature, **PPE** can instead denote the **Partial Projected Ensemble**, an ensemble of mixed subsystem states obtained by measuring only part of the complement and tracing out the rest; that construction addresses spatially resolved scrambling and erasure rather than multi-time process tensors [2508.05632]. Outside quantum many-body theory, **PPE** is also used for **Perturbed Parameter Ensembles** in climate-model emulation [2410.00931]. The acronym is therefore context sensitive.

Several limitations are explicit in the projected-process formulation. The construction in [2502.13930] assumes a **pure initial state** and **unitary dynamics**; more general mixed and non-unitary process tensors are not the focus. The numerics are confined to **finite-dimensional** spin and fermion chains. Higher PPE moments depend on the chosen **local intervention basis**. The analysis uses **Rényi-2** rather than von Neumann entropy, so its quantum dynamical entropy is not exactly the original Alicki–Fannes quantity. The evidence is also based on **finite-size numerics**, and the full experimental reconstruction of the PPE is exponentially costly in the number of intervention times [2502.13930].

Within those limits, the PPE provides a unified language in which a quantum process is treated as an ensemble of conditional output states, with the first moment recovering established dynamical-entropy observables and higher moments exposing the distributional entanglement structure that most clearly distinguishes chaotic from non-chaotic many-body dynamics.

Source: https://www.emergentmind.com/topics/projected-process-ensemble-ppe