Quantum computation with the eigenstate thermalization hypothesis instead of wavefunction preparation

Thomas E. Baker · 2025

Read Quantum computation with the eigenstate thermalization hypothesis instead of wavefunction preparation by Thomas E. Baker (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-19185.txt.

Introduction

Thomas Baker proposes a new way to run quantum computations in a 2025 physics paper. Standard quantum algorithms spend much effort preparing a precise starting state. Baker instead lets the system heat up in a controlled way. A rule of quantum physics says that a hot enough system settles into an even mix of its energy states on its own.

From that even mix, a phase-reading step can pull out functions of the input operator. Those functions solve a family of linear algebra problems, the workhorse of data science. The method skips the hardest setup step of older algorithms. Speed stays in the fast poly-log range that makes quantum methods attractive. The idea trades careful cooking for natural settling.

Catalog record

CREATORThomas E. Baker
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-19185

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Quantum computation with the eigenstate thermalization hypothesis instead of wavefunction preparation Thomas E. Baker Abstract The eigenstate thermalization hypothesis connects a time-average of expectation values for non-integrable systems to the normalized trace of an operator. It is shown here that the hypothesis can be a suitable way to take an expectation value over an equal superposition of eigenstates on the quantum computer. The complexity of taking the expectation value relies on an efficient time-evolution implementation with complexity O ⁡ ( T ) O(T) where T T is the cost of a single time-step going as O ⁡ ( log 2 ​ N ) O(\log_{2}N) and sampling at each time step. The resulting algorithm is demonstrated for the expectation value of the inverse of an operator and then generalized to the gradient of a logarithm-determinant. The method scales as O ⁡ ( τ ​ T ​ M / ε 3 ) O(\tau TM/\varepsilon^{3}) with a factor M M relating to the number of digits of precision, ε \varepsilon , and τ \tau the averaging time. This is independent of the condition number of the input operator and does not require a quantum memory. There is a potential for a many-body localized state–which would increase the thermalization time–but many circuits of practical interest are uniform enough to avoid this, and it is also expected that the eigenstate thermalization hypothesis holds beyond the currently available system sizes that are studied. The boundary where this algorithm would be considered a classical versus quantum algorithm is also considered. Introduction.– A quantum computer fundamentally accesses an N -sized Hilbert-space sized volume of quantum states from only \log_{2}N=n qubits Nielsen and Chuang 2010 . If the number of gates required to construct the correct state in the Hilbert space requires a depth N circuit, then this destroys the quantum advantage that is sought for quantum algorithms Tang 2021 . Considering the entanglement for some states can be very high, generating a superposition of the form |\Psi\rangle=\sum_{p=1}^{N}c_{p}|p\rangle (1) is potentially very expensive. It has been a dream of quantum computing to solve linear algebra problems of the form Ax=y (find a vector x from an input vector y and some matrix A ) in a faster way than classical algorithms can. Solving a 2^{n}\times 2^{n} matrix in poly-logarithmic time ( \sim\log_{2}N ) could create new opportunities for solutions in linear algebra beyond the capabilities of the classical computer. However, many quantum algorithms have hidden costs based on their assumptions. One source of extra cost is the computational complexity requirements to generate the starting state for a given algorithm Tang 2021 . For example, previous algorithms to solve linear algebra problems have relied on careful preparation of a superposition Harrow et al. 2009 ; Clader et al. 2013 ; Childs et al. 2017 ; Wossnig et al. 2018 ; Duan et al. 2020 ; Costa et al. 2022 where c_{p} are not necessarily uniform nor efficiently available Aaronson 2015 . Here it is shown that the explicit generation of the superposition of states can be replaced by thermalizing an initial wavefunction. Time-evolving a random initial wavefunction in a non-integrable system makes use of the eigenstate thermalization hypothesis (ETH) Deutsch 1991 ; Srednicki 1994 ; D’Alessio et al. 2016 ; Deutsch 2018 to sample from an equal superposition of eigenstates. This circumvents the need for a preparation of an equal superposition of eigenstates and allows for a full sum of all states. The resulting algorithm computes expectation values of the inverse in poly-logarithmic time without a quantum memory. Eigenstate thermalization hypothesis.– We can identify that the time-evolution operator of an arbitrary input A on a random vector is e^{-iAt}|r\rangle=\sum_{p=1}^{N}c_{p}e^{-iE_{p}t}|p\rangle=|\Psi(t)\rangle (2) for some eigenbasis \{|p\rangle\} . When we go to take an expectation value of the time evolved vector, \langle\Psi(t)|A|\Psi(t)\rangle=\sum_{p=1}^{N}a_{pp}|c_{p}|^{2} (3) with A=\sum_{pq}a_{pq}|p\rangle\langle q| , then the starting-state dependence seems to always be retained due to the unitarity of quantum physics. This is the case as it is applies to integrable systems. The time-evolution for non-integrable operators is not as straightforward. Imagining that the expectation value of any operator A is no longer integrable, then we have \overline{\langle\Psi(t)|A|\Psi(t)\rangle}\overset{!}{=}\frac{1}{N}\sum_{p=1}^{N}\langle p|A|p\rangle=\frac{1}{N}\mathrm{Tr}(A) (4) from ETH Deutsch 1991 ; Srednicki 1994 ; D’Alessio et al. 2016 ; Deutsch 2018 where an over-bar represents a time-average. The idea is that the eigenstates are allowed to transition between each other and thermalize or erase the initial state dependence in the time-average. ETH is more frequently discussed in the context of rectifying the unitarity of quantum mechanics with the ability for classical systems to exhibit chaos, but there is an opportunity to use ETH to replace wavefunction preparation. Quantum circuits are inherently designed to solve problems that are not integrable because of our desire to solve problems beyond what is possible mathematically or efficiently classically. The time required to average can vary depending on the system’s complexity in the decomposition of the time-evolution operator. For systems that display many-body localization (MBL) Nandkishore and Huse 2015 ; Sierant et al. 2025 , the time can be very long to thermalize. There is some indication that this time is not necessarily infinite for MBL Sels and Polkovnikov 2021 , but for the purposes here the resolution of whether MBL is a phase or merely has a long equilibration time is likely not a hurdle for envisioned application. The time to thermalize does not necessarily go up with system size. The fluctuations can decrease and the time-average can be very short, even in larger systems. ETH has been primarily tested on small systems and the proposal here is envisioned for large-scale quantum computation. If the ETH does not hold for this algorithm, which is a possibility due to a lack of a firm proof of ETH Rigol et al. 2008 , it would open up many questions about the thermalization of operators that has been verified in many previous articles in the literature. Expectation value of inverses.– Consider an operator A which has the relationships \langle p|A|p\rangle=E_{p}\quad\mathrm{and}\quad A|p\rangle=E_{p}|p\rangle (5) between eigenvalues E_{p} and eigenvectors |p\rangle . The inverse B=A^{-1} therefore has the relationships \langle p|B|p\rangle=\frac{1}{E_{p}}\quad\mathrm{and}\quad B|p\rangle=\frac{1}{E_{p}}|p\rangle (6) which can be verified by considering the form that the eigenvalue decomposition of A must take and that AB=I , for the identity I . Thus, B takes the form B=\sum_{p=1}^{N}\frac{1}{E_{p}}|p\rangle\langle p| (7) where A=\sum_{p}E_{p}|p\rangle\langle p| . The form of B in Eq. ( 7 ) has expectation value \langle B\rangle=\langle\Phi|B|\Phi\rangle=\sum_{p=1}^{N}\langle\Phi|\left(\frac{1}{E_{p}}|p\rangle\langle p|\right)|\Phi\rangle (8) for some wavefunction \Phi . Note that \langle B\rangle=\sum_{p=1}^{N}\frac{1}{E_{p}}\langle p|\Phi\rangle\langle\Phi|p\rangle=\sum_{p=1}^{N}\frac{|\langle\Phi|p\rangle|^{2}}{E_{p}} (9) as a more convenient representation. Algorithm 1 ETH- \Sigma 1: Prepare a random state |r\rangle 2: Time evolve the state according to an operator A , \exp(-i A t) 3: Apply a QPE to determine the eigenvalues in an m -qubit auxiliary register 4: Apply \sqrt{\Upsilon} ( vector form ) or \Upsilon ( operator form ) \triangleright Eq. ( 11 ) 5: Undo the QPE on the m -qubit auxiliary register 6: ( Vector form ): Perform a swap test with |\Phi\rangle at each time step and re-prepare the initial state OR ( Operator form ): Measure \Delta time step and re-prepare the initial state 7: Repeat until a number of samples is created 8: return Average many samples Vector

Read the complete text