TIMES-ADAPT: Quantum & Time-Series Adaptation

• TIMES-ADAPT is a multifaceted term referring to a fixed-depth variational quantum algorithm for evolving low-energy subspaces and an online adaptation mechanism in time-series forecasting.
• The quantum variant employs a learned basis-change unitary with TEPID-ADAPT to apply time-dependent phases while avoiding Trotterization errors and depth growth.
• In time-series applications, TIMES-ADAPT integrates Fourier-domain forecasting and exponential weighting to deliver rapid, utility-aware updates under latency and budget constraints.

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” (Sambasivam et al., 2 Mar 2026). 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 (Lee et al., 18 Feb 2025). 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 (Sambasivam et al., 2 Mar 2026). 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 $H_S = P_S H P_S$ (Sambasivam et al., 2 Mar 2026).

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 (Lee et al., 18 Feb 2025). 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 (Sreeram et al., 5 Feb 2026). 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 (Sambasivam et al., 2 Mar 2026). The time-series uses instead concern online adaptation, forecast combination, or utility-aware evaluation during deployment (Lee et al., 18 Feb 2025, Sreeram et al., 5 Feb 2026).

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,

$H$0

Once this change of basis has been learned, evolution inside $H$1 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 (Sambasivam et al., 2 Mar 2026).

Training uses TEPID-ADAPT as a subroutine. Instead of directly minimizing off-diagonal terms, TEPID-ADAPT prepares a low-temperature Gibbs state

$H$2

using a blocked variational ansatz containing the adaptive block $H$3 and a static preparation $H$4. Upon convergence, the procedure yields both the basis-change unitary and subspace energy differences through

$H$5

This avoids the per-eigenstate measurements required in concurrent VQE-style workflows (Sambasivam et al., 2 Mar 2026).

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 $H$6 into a diagonal unitary, so depth does not scale with $H$7. 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 (Sambasivam et al., 2 Mar 2026).

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 $H$8, then apply $H$9
TIMES-ADAPT-II	Computational-basis input	Compile $t$0 and apply it directly

For TIMES-ADAPT-I, the initial state is assumed to have a known truncated eigenbasis expansion,

$t$1

The algorithm prepares

$t$2

then applies $t$3 to obtain

$t$4

Only energy differences are required, and the paper gives a state-preparation circuit based on modified Givens rotations (Sambasivam et al., 2 Mar 2026).

For TIMES-ADAPT-II, the initial state is specified in the computational basis. The method constructs a diagonal unitary

$t$5

with $t$6 for $t$7 and $t$8 otherwise, and then compiles

$t$9

This removes the need to know the coefficients $P_S e^{-iHt} P_S$0 in advance, at the price of applying both $P_S e^{-iHt} P_S$1 and $P_S e^{-iHt} P_S$2 (Sambasivam et al., 2 Mar 2026).

The error behavior differs accordingly. If the initial state has support outside $P_S e^{-iHt} P_S$3, TIMES-ADAPT-I evolves the projection onto $P_S e^{-iHt} P_S$4 and its fidelity is time-independent,

$P_S e^{-iHt} P_S$5

TIMES-ADAPT-II instead produces an oscillatory fidelity,

$P_S e^{-iHt} P_S$6

with frequencies determined by energies outside the chosen subspace (Sambasivam et al., 2 Mar 2026).

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

The adaptive block $P_S e^{-iHt} P_S$7 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 $P_S e^{-iHt} P_S$8, 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 (Sambasivam et al., 2 Mar 2026).

Resource estimates are given explicitly. For TIMES-ADAPT-I, the total two-qubit gate depth is

$P_S e^{-iHt} P_S$9

where the first term comes from the preparation of $H_S = P_S H P_S$0 and the second from $H_S = P_S H P_S$1. The paper gives the example $H_S = P_S H P_S$2 and $H_S = P_S H P_S$3, yielding depth $H_S = P_S H P_S$4 (Sambasivam et al., 2 Mar 2026). For TIMES-ADAPT-II, the total two-qubit gate depth is

$H_S = P_S H P_S$5

with example depth $H_S = P_S H P_S$6 for the seven-qubit setting reported in the paper (Sambasivam et al., 2 Mar 2026).

These estimates are paired with a compilation claim: any $H_S = P_S H P_S$7-controlled $H_S = P_S H P_S$8 can be implemented with CNOT depth $H_S = P_S H P_S$9 for $H$0, and QEB double-excitation operators have CNOT depth $H$1 per operator layer. The resulting circuits are intended for NISQ settings where coherence limits disfavor methods whose depth grows with evolution time (Sambasivam et al., 2 Mar 2026).

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 (Sambasivam et al., 2 Mar 2026).

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$2

the paper studies $H$3 and a low-energy subspace spanned by the $H$4 lowest eigenstates. TEPID-ADAPT is run with $H$5 to train $H$6 and extract $H$7 (Sambasivam et al., 2 Mar 2026).

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 $H$8. The text reports that both variants outperform product-formula behavior at long times in the qualitative comparison given there (Sambasivam et al., 2 Mar 2026).

Two further applications are highlighted. The first is wave-packet evolution in a seven-qubit longitudinal-field XXZ model with open boundary conditions,

$H$9

where the symmetry sector is the single-magnon subspace. The observable is the local magnetization $S$0. 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 (Sambasivam et al., 2 Mar 2026).

The second is energy transport in a seven-qubit staggered-field XXZ model with periodic boundary conditions. The observable is the local energy density

$S$1

In this case, TIMES-ADAPT-II is used, and the reported long-time infidelity remains near $S$2 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 $S$3 (Sambasivam et al., 2 Mar 2026).

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 (Lee et al., 18 Feb 2025).

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

$S$4

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 $S$5, $S$6, $S$7, $S$8, $S$9, and $V_A$0, and the method is described as adding only approximately $V_A$1 seconds per update on $V_A$2 CPUs relative to the FM alone (Lee et al., 18 Feb 2025).

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 $V_A$3, $V_A$4, and $V_A$5. On ImageNet-C across $V_A$6 temporal evaluations, the paper reports rank instability: ETA, described there as the conventional offline winner, falls short in $V_A$7 of evaluations (Sreeram et al., 5 Feb 2026).

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 (Lee et al., 18 Feb 2025, Sreeram et al., 5 Feb 2026).

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 $V_A$8, 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 $V_A$9, and that residual non-diagonal terms in $S$0 can introduce phase errors in observables (Sambasivam et al., 2 Mar 2026).

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 (Sambasivam et al., 2 Mar 2026).

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 (Lee et al., 18 Feb 2025, Sreeram et al., 5 Feb 2026). 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.