⚛️ quantum physics

New directions in dynamical expectation estimation

This paper introduces a novel sweep algorithm with coupled loss functions that jointly optimizes state and observable approximations to significantly reduce errors in dynamical expectation value estimation, achieving two to three orders of magnitude higher accuracy than variational state compression in 30-qubit random circuits.

Original authors: Panjin Kim, Kyung Chul Jeong, Yun-Tak Oh

Published 2026-09-17

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Panjin Kim, Kyung Chul Jeong, Yun-Tak Oh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the realm of quantum physics, scientists often need to predict the outcome of a measurement after a system has evolved over time. This is a task known as calculating an expectation value, which essentially asks: if we look at a quantum system after it has changed, what average result will we see? To answer this, researchers must track two things simultaneously: how the state of the system changes as it moves forward in time, and how the question being asked about the system changes if we were to look at it from the end of the process backward. For decades, the standard way to handle these calculations on classical computers has been to simplify the state and the question separately. Scientists would compress the description of the system to make it manageable, and they would simplify the description of the measurement tool independently, assuming that doing both well enough would lead to a good answer. However, this approach treats the two parts as if they exist in isolation, ignoring how their individual simplifications might interact to create a larger error in the final result.

A team of researchers from the Aﬃliated Institute of ETRI in Korea has proposed a different way to think about this problem. They argue that the best way to simplify a quantum calculation is to guide the simplification of the state and the measurement tool together, using the final answer as a compass. Instead of trying to make the state look as close as possible to the original, or the measurement tool as close as possible to the original, their new method asks: "Does this simplified version still give us the right answer for the specific question we are trying to solve?" They developed a computer algorithm that moves through a quantum circuit step by step, first moving forward to update the state and then moving backward to update the measurement tool. At each step, the algorithm adjusts both the state and the tool at the same time, ensuring that any small errors made in one part are immediately corrected by the other, specifically to keep the final calculated value accurate.

The researchers tested this new approach on complex simulations involving thirty-qubit quantum circuits, which are systems large enough to be difficult for classical computers to handle exactly. They compared their method against two established techniques: one that compresses the state alone and another that simplifies the measurement tool alone. The results showed a dramatic improvement. In these simulations, the new method produced errors that were two to three orders of magnitude smaller than the traditional methods, even when all methods were using the same amount of computer memory to store the data. This means the new approach was hundreds of times more accurate. The team found that by aligning the simplification of the state with the simplification of the observable, they could discard details that did not matter for the final answer while keeping the details that did.

To understand why this works, consider how errors usually build up. In the old methods, a small mistake in describing the state and a small mistake in describing the measurement tool are calculated separately. When these two are combined at the end, the errors can add up or interfere in ways that degrade the result. The new algorithm uses a specific mathematical rule, or loss function, that penalizes the system if the combination of the state and the tool produces a wrong answer. It does this by checking the result at every single step of the process. If the state is simplified in a way that makes the final answer drift, the algorithm corrects it immediately, even if the state itself still looks somewhat different from the original. This ensures that the most important features for the final calculation are preserved, while less relevant details are allowed to fade away.

The study demonstrates that this joint approach is highly effective for the specific types of quantum circuits they tested, which are known as barren plateau circuits. These are random circuits designed to be particularly challenging for classical computers. The researchers ran five hundred different versions of these circuits, varying the number of steps in the process, and consistently found that their new method outperformed the standard techniques. They noted that while their specific algorithm is not claimed to be the perfect solution, the underlying idea of guiding approximations by the final goal is a powerful new direction. The work suggests that future improvements in quantum simulation will likely come from methods that treat the state and the observable as a single, interconnected system rather than two separate problems to be solved in isolation.

Technical Summary: New Directions in Dynamical Expectation Estimation

Problem Statement
Computing expectation values of observables is a foundational task in physical sciences and quantum information. However, classically simulating the dynamics of quantum systems is computationally hard, necessitating controlled approximations. Current representative methods, such as Matrix Product States (MPS) for state evolution and Pauli propagation or Heisenberg-picture DMRG for observable evolution, rely on truncation criteria optimized for the state or the observable in isolation. These methods minimize the distance between the approximate and target states (or observables) without accounting for how their respective errors combine in the final expectation value expression. Consequently, these approximations may preserve features of the state or observable that are irrelevant to the specific expectation value being computed, leading to suboptimal accuracy for the target quantity.

Methodology
The authors propose an alternative approach where approximations of both the state and the observable are guided jointly by how their errors combine in the final expectation value. The core of the method is a sweep algorithm (Algorithm 1) that optimizes forward state updates and backward observable updates using coupled loss functions.

The algorithm operates on a quantum circuit $C = U_L \cdots U_1$ with an initial state $\rho_0$ and observable $O$ . It defines forward evolved states $\rho_t$ and backward evolved observables $O_t$ . The method iterates through time slices, updating the state approximation $\tilde{\rho}_t$ and the observable approximation $\tilde{O}_t$ to minimize specific loss functions:

Forward Loss ( $\ell_F$ ): Minimizes the error in the state approximation while penalizing the error in the expectation value.
$\ell_F(\rho; \rho^*, \bar{O}) = \gamma |\text{Tr}[(\rho - \rho^*)\bar{O}]|^2 + \mu \|\rho - \rho^*\|^2$
Here, $\rho^*$ is the uncompressed target state, $\bar{O}$ is the backward-propagated observable held fixed, and $\|\cdot\|$ denotes the Hilbert-Schmidt norm. The parameter $\gamma$ controls the weight of the expectation value error, while $\mu$ controls the conventional state-only compression error.
Backward Loss ( $\ell_B$ ): Analogously minimizes the error in the observable approximation, penalizing the deviation in the expectation value using the forward-propagated state.
$\ell_B(O; O^*, \bar{\rho}) = \gamma |\text{Tr}[\bar{\rho}(O - O^*)]|^2 + \mu \|O - O^*\|^2$

The implementation utilizes tensor networks (MPS for states and MPO/Pauli strings for observables). During the forward sweep, the algorithm updates the MPS tensors to minimize $\ell_F$ , requiring the fixed backward observable. Conversely, the backward sweep updates the observable tensors to minimize $\ell_B$ , requiring the fixed forward state. The $\mu$ -term acts as a regularizer for directions not constrained by the $\gamma$ -term, while the $\gamma$ -term specifically targets the contraction relevant to the expectation value.

Key Contributions and Theoretical Rationale
The paper provides an analytical rationale for why this joint optimization reduces errors compared to conventional single-sided truncation. The total error $E - b^{(i)}$ (where $E$ is the exact value and $b^{(i)}$ is the estimate) is decomposed into two remainder terms, $R_1$ and $R_2$ :

$R_1$ (Local Expectation Error): Represents the sum of local errors in the expectation value at each time slice. The $\gamma$ -terms in the loss functions directly penalize these local changes, suppressing the magnitude of $R_1$ in the proposed method ( $i=1$ ) compared to the conventional method ( $i=0$ ).
$R_2$ (Combined Approximation Error): Represents the product of state approximation errors and backward compression errors. The authors argue that $R_2$ scales as $O(L\epsilon^2)$ when both errors are $O(\epsilon)$ , whereas the dominant error term in conventional methods ( $R_1^{(0)}$ ) scales as $O(L\epsilon)$ . Thus, for sufficiently small errors, the joint approach yields a smaller total error.

The loss functions are presented as heuristic choices; the paper notes that the optimal form of these objectives remains an open question.

Numerical Results
The proposed algorithm was benchmarked against Variational State Compression (VSC) and Pauli Propagation (PP) on 30-qubit random circuits drawn from the "barren plateau" ensemble. The circuits consisted of $L$ layers of single-qubit rotations and nearest-neighbor controlled-Z gates, with depths ranging from 5 to 100.

Performance: At equal bond dimensions ( $\chi_\rho = \chi_O \leq 30$ or $50$), the proposed algorithm achieved expectation value errors ( $\langle Z_1 Z_2 \rangle$ ) that were two to three orders of magnitude smaller than those produced by variational state compression.
Error Analysis: Numerical tests on 20-qubit instances confirmed the theoretical decomposition: the proposed method significantly suppressed the $|R_1|$ term (local expectation error) relative to the conventional method, while both $|R_2|$ terms remained smaller than the conventional $|R_1|$ term.

Significance and Claims
The paper claims that the central message is that state and observable approximations should be guided by their combined effect on the final expectation value rather than being optimized in isolation. While the authors do not claim their sweep algorithm is the optimal realization of this principle, they demonstrate that it provides a concrete analytical rationale and significant empirical improvements in accuracy. The results motivate further exploration of joint state and observable approximation for dynamical expectation value estimation, suggesting that aligning the approximation criteria with the target quantity can prioritize the most relevant features of the quantum system. The work remains modest regarding the specific form of the loss functions, acknowledging them as examples rather than uniquely preferred objectives.

Technical Summary: New Directions in Dynamical Expectation Estimation

More like this