---
title: 'TIMES-ADAPT: Quantum & Time-Series Adaptation'
url: https://www.emergentmind.com/topics/times-adapt
type: topic
---

# TIMES-ADAPT: Quantum & Time-Series Adaptation

TIMES-ADAPT most specifically denotes a variational quantum algorithm for real-time evolution in low-energy or symmetric subspaces of a time-independent Hamiltonian using fixed-depth circuits, introduced in “TIMES-ADAPT: A Quantum algorithm for real-time evolution in low-energy subspaces using fixed-depth circuits” [2603.02305]. In the supplied literature, however, the label is also reused in a broader, interpretive sense for time-sensitive or time-series adaptation mechanisms at deployment, especially in online forecasting and test-time adaptation settings [2502.12920]. The term therefore has a domain-dependent meaning: in quantum computing it names a concrete algorithm, whereas in time-series ML it can function as an umbrella label for online adaptation under temporal shift.

## 1. Terminological scope and domain-specific meanings

In its formal, paper-title sense, TIMES-ADAPT is the quantum algorithm of [2603.02305]. Its purpose is to prepare time-evolved states inside a chosen subspace $S$ of a time-independent Hamiltonian $H$, with circuit depth that does not grow with the simulated time $t$. The target restricted evolution is written as $P_S e^{-iHt} P_S$, equivalently evolution under $H_S = P_S H P_S$ [2603.02305].

In the supplied time-series literature, the same string is also used less formally. The overview of AdapTS explicitly states that it “equate[s] it with ‘TIMES-ADAPT’,” using the label for a lightweight mechanism that adapts time-series foundation model forecasts online through a Fourier-domain forecaster and an exponential-weighting module [2502.12920]. A separate synthesis of Tempora states that “TIMES-ADAPT is not explicitly mentioned in the paper” and then interprets the phrase as time-sensitive test-time adaptation under latency, deadline, and budget constraints [2602.06136]. This suggests that the label has become polysemous across at least two technical communities.

That polysemy matters because the underlying objects differ substantially. The quantum TIMES-ADAPT is a fixed-depth circuit construction built around a trained basis-change unitary and diagonal phase application [2603.02305]. The time-series uses instead concern online adaptation, forecast combination, or utility-aware evaluation during deployment [2502.12920], [2602.06136].

## 2. Quantum TIMES-ADAPT: subspace evolution by learned diagonalization

The quantum algorithm addresses real-time evolution under a time-independent Hamiltonian $H$ restricted to a low-energy or symmetry subspace $S$. Its central idea is to train a unitary $V_A$ that maps a chosen computational-basis subspace to eigenstates spanning $S$:
$$
V_A \equiv \sum_{k\in S} |\psi_k\rangle\langle c_k| + V_\perp,
$$
with approximate diagonalization inside the subspace,
$$
V_A^\dagger H V_A \approx D_S.
$$
Once this change of basis has been learned, evolution inside $S$ reduces to phase application in the diagonal basis, so time enters only through gate parameters rather than through circuit repetition or product-formula depth growth [2603.02305].

Training uses TEPID-ADAPT as a subroutine. Instead of directly minimizing off-diagonal terms, TEPID-ADAPT prepares a low-temperature Gibbs state
$$
\rho_G = e^{-\beta H_S}/Z_S,
$$
using a blocked variational ansatz containing the adaptive block $V_A$ and a static preparation $U_m(\mu)$. Upon convergence, the procedure yields both the basis-change unitary and subspace energy differences through
$$
\Delta E_k = \frac{1}{\beta}\log(\mu_1/\mu_k).
$$
This avoids the per-eigenstate measurements required in concurrent VQE-style workflows [2603.02305].

The fixed-depth property follows from this structure. After training, the circuit architecture is frozen. Real-time evolution is implemented by inserting time-dependent phases such as $e^{-iE_kt}$ into a diagonal unitary, so depth does not scale with $t$. The paper positions this as a direct contrast to Trotterization and related time-discretized methods, whose depth grows with target time and whose errors accumulate correspondingly [2603.02305].

## 3. The two algorithmic variants

TIMES-ADAPT has two versions, distinguished by how the initial state is specified.

| Variant | Initial-state specification | Core construction |
|---|---|---|
| TIMES-ADAPT-I | Energy eigenbasis coefficients known | Prepare $\sum_{k\in S}\alpha_k e^{-iE_kt}|c_k\rangle$, then apply $V_A$ |
| TIMES-ADAPT-II | Computational-basis input | Compile $\tilde U(t)=V_A D(t) V_A^\dagger$ and apply it directly |

For TIMES-ADAPT-I, the initial state is assumed to have a known truncated eigenbasis expansion,
$$
|\psi(0)\rangle = \sum_{k\in S}\alpha_k |\psi_k\rangle.
$$
The algorithm prepares
$$
|\chi(t)\rangle = \sum_{k\in S}\alpha_k e^{-iE_kt}|c_k\rangle,
$$
then applies $V_A$ to obtain
$$
|\psi(t)\rangle = V_A|\chi(t)\rangle.
$$
Only energy differences are required, and the paper gives a state-preparation circuit based on modified Givens rotations [2603.02305].

For TIMES-ADAPT-II, the initial state is specified in the computational basis. The method constructs a diagonal unitary
$$
D(t)=\sum_{k=1}^{2^{N_s}} e^{-i\Lambda_k t}|c_k\rangle\langle c_k|,
$$
with $\Lambda_k = E_k$ for $k\le m$ and $\Lambda_k=0$ otherwise, and then compiles
$$
\tilde U(t)=V_A D(t) V_A^\dagger.
$$
This removes the need to know the coefficients $\alpha_k$ in advance, at the price of applying both $V_A$ and $V_A^\dagger$ [2603.02305].

The error behavior differs accordingly. If the initial state has support outside $S$, TIMES-ADAPT-I evolves the projection onto $S$ and its fidelity is time-independent,
$$
F_I(t)=\left(\sum_{k\in S}|\alpha_k|^2\right)^2.
$$
TIMES-ADAPT-II instead produces an oscillatory fidelity,
$$
F_{II}(t)=\left|\sum_{k\in S}|\alpha_k|^2+\sum_{k\in(S'\setminus S)}|\alpha_k|^2e^{-iE_kt}\right|^2,
$$
with frequencies determined by energies outside the chosen subspace [2603.02305].

## 4. Circuit construction, resource estimates, and relation to ADAPT methods

The adaptive block $V_A$ is built layer by layer from an operator pool in ADAPT style, but the cost function is the free energy of a Gibbs state rather than a single-state ground-energy objective. For symmetry sectors such as fixed $S^z$, the operator pool can be chosen to preserve the symmetry; the paper cites QEB and TETRIS-ADAPT-VQE as relevant constructions for symmetry preservation and depth reduction [2603.02305].

Resource estimates are given explicitly. For TIMES-ADAPT-I, the total two-qubit gate depth is
$$
8m(N_s-1)+11N_{\text{adapt}},
$$
where the first term comes from the preparation of $|\chi(t)\rangle$ and the second from $V_A$. The paper gives the example $N_s=m=7$ and $N_{\text{adapt}}=7$, yielding depth $\le 413$ [2603.02305]. For TIMES-ADAPT-II, the total two-qubit gate depth is
$$
8m(N_s-1)+2\times 11N_{\text{adapt}},
$$
with example depth $\le 1534$ for the seven-qubit setting reported in the paper [2603.02305].

These estimates are paired with a compilation claim: any $n$-controlled $SU(2)$ can be implemented with CNOT depth $8n$ for $n\ge 6$, and QEB double-excitation operators have CNOT depth $11$ per operator layer. The resulting circuits are intended for NISQ settings where coherence limits disfavor methods whose depth grows with evolution time [2603.02305].

The comparison to other quantum simulation paradigms is structural. Product-formula approaches such as Trotter and qDRIFT scale depth with time and may break symmetries; qubitization and LCU are resource-heavy; prior variational fast-forward methods may begin from trotterized approximations. TIMES-ADAPT is presented as avoiding trotterization altogether and therefore avoiding time-discretization error accumulation [2603.02305].

## 5. Benchmarks and applications in spin dynamics

The benchmark Hamiltonian family is based on Heisenberg XXZ variants. For the six-qubit antiferromagnetic XXZ model,
$$
H_{XXZ}(J_z)=\sum_{k=1}^{5}\left(\sigma_k^x\sigma_{k+1}^x+\sigma_k^y\sigma_{k+1}^y+J_z\sigma_k^z\sigma_{k+1}^z\right),
$$
the paper studies $J_z=1.5$ and a low-energy subspace spanned by the $m=5$ lowest eigenstates. TEPID-ADAPT is run with $\beta=2.0$ to train $V_A$ and extract $\Delta E_k$ [2603.02305].

In this setting, TIMES-ADAPT-I exhibits constant-in-time fidelities equal to the squared norm of the projection onto the chosen subspace, while TIMES-ADAPT-II shows oscillatory fidelity when $m<5$. The text reports that both variants outperform product-formula behavior at long times in the qualitative comparison given there [2603.02305].

Two further applications are highlighted. The first is wave-packet evolution in a seven-qubit longitudinal-field XXZ model with open boundary conditions,
$$
H_{LFXXZ}(J_z,h_z)=\sum_{k=1}^{N_s-1}\left(\sigma_k^x\sigma_{k+1}^x+\sigma_k^y\sigma_{k+1}^y+J_z\sigma_k^z\sigma_{k+1}^z\right)+h_z\sum_{k=1}^{N_s}\sigma_k^z,
$$
where the symmetry sector is the single-magnon subspace. The observable is the local magnetization $1-\langle \sigma_j^z\rangle(t)$. At matched two-qubit gate depth, the paper reports that TIMES-ADAPT-I maintains low infidelity for long times while first-order Trotter error grows [2603.02305].

The second is energy transport in a seven-qubit staggered-field XXZ model with periodic boundary conditions. The observable is the local energy density
$$
H_k=\sigma_k^x\sigma_{k+1}^x+\sigma_k^y\sigma_{k+1}^y+J_z\sigma_k^z\sigma_{k+1}^z+(-1)^k h_z\sigma_k^z.
$$
In this case, TIMES-ADAPT-II is used, and the reported long-time infidelity remains near $10^{-3}$ while Trotter infidelity grows by an order of magnitude; the text also states that TIMES-ADAPT-II tracks exact observables closely and can be improved further by increasing $m$ [2603.02305].

## 6. Broader use of the label in time-series adaptation

A broader reading in the supplied corpus treats TIMES-ADAPT as a label for online adaptation in time-series ML rather than for a specific quantum circuit family. The clearest explicit example is AdapTS, described as a lightweight, FM-agnostic mechanism for online adaptation of foundation model forecasts and explicitly equated with TIMES-ADAPT in the supplied overview [2502.12920].

In that usage, the system has two components: an AdapTS-Forecaster, which is a linear multi-horizon forecaster fit in the Fourier domain with the regularized closed-form solution
$$
W=(\tilde X^*\tilde X+\lambda I)^{-1}\tilde X^*\tilde Y,
$$
and an AdapTS-Weighter, which dynamically combines the foundation-model forecast and the online forecaster through exponential weighting. The implementation uses low-pass compression, Woodbury updates, instance normalization, Welford-based online scaling, and a fast/slow/merge weighting scheme. The reported default configuration includes $L=520$, $M=200$, $\alpha=0.9$, $\lambda=20$, $\eta=0.5$, and $B=5$, and the method is described as adding only approximately $0.38$ seconds per update on $2$ CPUs relative to the FM alone [2502.12920].

A second, explicitly interpretive use appears in Tempora. The supplied synthesis states that TIMES-ADAPT is not named in the paper, but that Tempora provides the formalism needed to evaluate online TTA under temporal pressure. Tempora defines three scenarios—hard deadlines in asynchronous streams, latency-decaying value in interactive settings, and budget-constrained deployments—and corresponding utilities $U_{\text{discrete}}$, $U_{\text{continuous}}$, and $U_{\text{amortised}}$. On ImageNet-C across $240$ temporal evaluations, the paper reports rank instability: ETA, described there as the conventional offline winner, falls short in $41.2\%$ of evaluations [2602.06136].

This suggests a substantive distinction between “TIMES-ADAPT” as an algorithm name in quantum computing and “TIMES-ADAPT” as an umbrella concept for temporally constrained adaptation in time-series and deployment-centric ML. In the latter sense, the label is tied less to a single mechanism than to recurring requirements: online updates, source-free operation, temporal feedback, and explicit management of adaptation under shift or latency constraints [2502.12920], [2602.06136].

## 7. Limitations and future directions

For the quantum algorithm, the principal limitations arise from subspace quality and diagonalization accuracy. If the initial state has substantial support outside the chosen subspace $S$, then TIMES-ADAPT-I is fidelity-limited by the projected norm, while TIMES-ADAPT-II exhibits oscillatory fidelity determined by out-of-subspace energies. The paper also notes that highly degenerate or complicated subspaces with weak energy separation may hinder clean diagonalization or require larger $m$, and that residual non-diagonal terms in $V_A^\dagger H V_A$ can introduce phase errors in observables [2603.02305].

The stated extensions are correspondingly structural: more complex symmetry sectors, larger subspaces, multi-wave-packet scattering, time-dependent Hamiltonians, and operator pools engineered for domain-specific symmetries. These are natural continuations of the fixed-depth subspace-evolution program because the current construction already separates basis learning from time-parameterized phase application [2603.02305].

In the broader time-series usage, the open questions are different. AdapTS emphasizes efficient online feedback usage without touching foundation-model weights, whereas Tempora emphasizes that method selection should be utility-driven rather than accuracy-driven under deadlines, latency decay, and compute budgets [2502.12920], [2602.06136]. A plausible implication is that the shared label “TIMES-ADAPT” now spans two distinct research agendas: fixed-depth variational quantum simulation in low-energy subspaces, and temporally aware online adaptation in sequence modeling.

Source: https://www.emergentmind.com/topics/times-adapt