betaFeedback

WorksPhysics and Astronomy

Quantum machine learning geospatial time-series by seasonal encoding

Abstract

Quantum data reuploading models can increase the expressive capacity of variational quantum models, but do not necessarily improve predictive performance. In the present work, we investigate this behavior for seasonal time series forecasting and show that data encoding determines how well temporal structures are represented as circuit depth increases. We develop a fidelity-based quantum geometric description in which repeated encodings contribute to an effective input generator, so that reuploading depth changes the quantum geometry through which approximately recurrent inputs are distinguished. We construct fixed and adaptive seasonal encodings that represent observation magnitude and seasonal phase in distinct Bloch-sphere coordinates. We evaluate these encodings for monthly precipitation derived from ERA5-Land, a global reanalysis dataset, from a representative semi-arid, drought-prone tropical region in East Africa, using a 14-month lag window and reuploading depths LL= 1-6. We find that in ideal simulations, standard value-based encodings degrade markedly with depth, whereas adaptive seasonal encodings preserve temporal correlation, spectral structure, and forecasting accuracy. Noisy simulations show that this encoding hierarchy persists at shallow and intermediate depths, while larger depths are limited by circuit noise and optimization difficulty. The seasonal quantum models remain competitive with established classical forecasting methods (ARIMA, random forests, and LSTMs). These results identify structure-aware quantum data encoding as a general strategy for building deeper quantum reuploading models.

Introduction

Remote sensing is now central to observing environmental changes across spatial and temporal scales. Optical, radar, thermal, and LiDAR sensors provide repeated measurements of, for instance, land cover and land-use change, forest structure, surface temperature, water availability, and precipitation [1, 2, 3]. These measurements are important not only for characterizing the present state of the Earth’s environment, but also for understanding how it evolves. Precipitation time-series data are especially critical for understanding agricultural productivity, water resource availability, drought and flood conditions, and food security monitoring [4, 5, 6]. However, the diversity of data modalities is growing and the resulting data volumes are increasing. Thus, developing novel methods for the extraction and forecasting of meaningful spatio-temporal patterns is an increasingly important problem in geospatial analysis.

Time series forecasting differs fundamentally from analyzing unordered observations, since predictive information is encoded not only in the values themselves, but also in their temporal organization. Environmental signals often exhibit recurring seasonal patterns, trends, and correlations across time scales, yet they are generally not perfectly periodic. Classical methods, from autoregressive models to random forests to recurrent neural networks, exploit these dependencies in different ways. Across these methods, performance depends on model capacity as well as on whether the representation preserves the relevant temporal relationships. As environmental forecasting problems become increasingly multivariate and spatiotemporal, the relevant predictive structure can involve correlations across time, and also across variables, spatial locations, and observational modalities. Therefore, a central challenge and opportunity is to determine whether new computational paradigms can offer representations that better preserve these predictive structures.

Quantum machine learning (QML) provides a novel framework for constructing such representations by embedding classical data into quantum states and processing them with trainable quantum circuits [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]. The ability of quantum feature maps and entangling operations to construct correlated representations makes QML particularly interesting for learning problems in which multiple interacting sources of structure must be represented simultaneously. Whether this representational flexibility translates into useful predictive performance remains an open question and should first be established in controlled settings where the relevant structure can be identified and interpreted. In variational quantum models, the choice of data encoding determines how relationships in the classical input space are translated into the geometry of Hilbert space and can therefore strongly influence what the model is able to learn [18, 19, 20, 21, 22]. Data reuploading architectures extend this idea by repeatedly inserting the classical input between trainable variational transformations, allowing increasingly rich nonlinear functions to be represented as the number of reuploading layers grows [23, 24, 25]. Increasing reuploading depth, however, does not necessarily lead to improved predictive performance and can be accompanied by challenges optimizing the model as the behavior of the objective function varies significantly [26, 27, 28, 29]. Since the input is repeatedly re-encoded after successive quantum transformations, increasing depth also changes how the circuit represents relationships between different inputs.

Understanding this depth-dependent representation is therefore particularly important for structured data such as time series, where predictive information depends on preserving meaningful relationships between temporally related observations. Before extending such models to multivariate and spatiotemporal settings, though, it is essential to establish how they behave in a controlled forecasting problem where the relevant temporal organization can be isolated. A univariate seasonal time series provides such a setting: the dominant recurrence seasonal structure can be identified directly, the effect of quantum data representation can be examined independently of cross-variable interactions, and forecasting performance can be compared against classical approaches.

Most commonly used quantum angle encodings map numerical input features directly through their magnitude into single-qubit rotations [20, 30, 31, 32, 33, 34] while carrying no explicit information about where that observation occurs within a recurring temporal cycle. For a seasonal signal, these two pieces of information are qualitatively distinct. Similar precipitation values observed at different times of the year can have different temporal meanings, while observations separated by one annual cycle can occupy corresponding seasonal positions even when their magnitudes differ. A value-based encoding therefore requires the variational circuit to infer this organization indirectly from the repeated sequence of observations. This distinction becomes especially relevant in data reuploading models where repeated encodings inserted between variational layers are known to modify the function classes accessible to a variational quantum model [23, 25]. Recent work in quantum machine learning has emphasized the value of incorporating known structure directly into quantum representations rather than relying only on unconstrained model capacity [35, 36, 37]. This motivates a broader design question for quantum time series learning: when a physically or statistically meaningful temporal structure is already known, should that structure be incorporated directly into the quantum data encoding, or left for the variational circuit to infer?

In this work, we address this question by directly linking temporal recurrence to the geometry induced by a quantum data reuploading circuit. We consider input windows separated by a characteristic recurrence period and use their quantum-state fidelity to quantify how closely recurring observations are represented as reuploading depth increases. A local geometric analysis relates the distinguishability of these inputs to a Fubini-Study metric [38, 39, 40, 41] generated by contributions from each reuploading layer, showing explicitly that circuit depth changes the geometry through which pre-existing temporal differences in the data are resolved. As a result, we construct season-aware quantum encodings that represent observation magnitude and seasonal phase via separate coordinates on the Bloch sphere. Precipitation magnitude determines the polar coordinate, while position within the annual cycle or calendar month determines the azimuthal coordinate. We consider a fixed seasonal representation and its adaptive variants in which this reference geometry can be deformed during training, allowing the model to learn from the data without discarding the known temporal organization of the signal.

We demonstrate these ideas using monthly precipitation forecasting for Napak, Uganda, with a hybrid quantum data reuploading model across reuploading depths L=1,…,6L = 1,\ldots,6. To this end, we compare conventional value-based angle encodings with the fixed and adaptive seasonal encodings in ideal simulations, and subsequently examine the seasonal models under noisy simulations. We further benchmark the resulting forecasts against classical statistical and machine learning approaches: ARIMA (autoregressive integrated moving average) [42], random forest [43], and LSTM (long short-term memory) [44].

As a main result we find that in ideal simulations, the value-based encodings progressively lose forecasting performance and temporal structure as reuploading depth increases, whereas the adaptive seasonal encodings remain substantially more stable across depth. This hierarchy is also retained in noisy simulations at shallow and intermediate depths, but circuit noise and optimization difficulty begin to dominate at larger LL. Taken together, these results show that increasing circuit depth alone is insufficient to improve quantum time-series learning. Rather, the structure by which the data is encoded can play a central role in determining whether predictive temporal information remains accessible as the model deepens.

Hybrid quantum-classical machine learning

We begin by describing the hybrid quantum-classical forecasting model used throughout this work. Given a temporally ordered sequence of observations, the forecasting task is formulated as a one-step-ahead prediction from a lag window. We implement this map using a quantum data reuploading architecture in which the same input is repeatedly encoded between variational circuit blocks [20, 25, 46, 48] [47]. This construction allows us to vary the reuploading depth while keeping the underlying forecasting task and readout architecture fixed.

Consider a univariate time series {y0,y1,…,yt,…}\{y_0,y_1,\ldots,y_t,\ldots\}. To predict the observation at time t+1t+1, we construct a lag window containing the ww preceding observations

xt=(yt−w+1,…,yt−1,yt)T∈Rw,(1)\mathbf{x}_t=(y_{t-w+1},\ldots,y_{t-1},y_t)^{T}\in\mathbb{R}^{w}, \tag*{(1)}

and learn the one-step forecasting map

fθ:xt⟼y^t+1,(2)f_{\boldsymbol{\theta}}:\mathbf{x}_t\longmapsto\widehat{y}_{t+1}, \tag*{(2)}

where y^t+1\widehat{y}_{t+1} is the prediction of the subsequent observation and θ\boldsymbol{\theta} denotes the complete set of trainable parameters of the hybrid quantum-classical model. We realize this map using a quantum data reuploading architecture. Each component of the lag window is associated with one qubit, resulting in a ww-qubit register initialized in ∣0⟩⊗w|0\rangle^{\otimes w}. We separate these parameters according to their role: θ=(θenc,θvar,θhead)\boldsymbol{\theta}=(\boldsymbol{\theta}^{\mathrm{enc}},\boldsymbol{\theta}^{\mathrm{var}},\boldsymbol{\theta}^{\mathrm{head}}). Here, θenc\boldsymbol{\theta}^{\mathrm{enc}} contains parameters of a trainable encoding block, θvar\boldsymbol{\theta}^{\mathrm{var}} contains the parameters of the variational block, and θhead\boldsymbol{\theta}^{\mathrm{head}} contains the parameters of the classical output head.

The quantum circuit consists of LL successive data reuploading layers, cf. Figure 1 for a sketch. At layer ℓ\ell, the input window is first mapped onto the quantum register via a data encoding block Sℓ(xt,θℓenc)S_\ell(\mathbf{x}_t,\boldsymbol{\theta}^{\mathrm{enc}}_\ell) followed by a trainable variational circuit block Vℓ(θℓvar)V_\ell(\boldsymbol{\theta}^{\mathrm{var}}_\ell). For our present purposes, we implement VℓV_\ell using a hardware-efficient variational ansatz consisting of a trainable three-parameter rotation on every qubit followed by a nearest-neighbor ring of CNOT gates, where the three rotation angles on each qubit are independently trainable. The raw data values are suitably transformed via classical pre-processing such that they can be encoded into SℓS_\ell. One complete reuploading block is, therefore,

Figure 1. The hybrid quantum-classical machine learning model for time-series forecasting. Within each reupload layer, classically preprocessed data enters the quantum circuit through a data encoding block, followed by a variational block. This is repeated LL times followed by Pauli-ZZ measurement on all qubits, which is then passed onto a classical output head to obtain the prediction after classical post-processing. This circuit includes trainable parameters in the encoding step (θenc\theta^{\mathrm{enc}}), which parameter is omitted in some of the encodings. Figure generated using ⟨q∣pic⟩\langle q|\mathrm{pic}\rangle [45].

Aℓ(xt,θℓenc,θℓvar)=Vℓ(θℓvar)Sℓ(xt,θℓenc),(3)A_\ell(\mathbf{x}_t,\boldsymbol{\theta}^{\mathrm{enc}}_\ell,\boldsymbol{\theta}^{\mathrm{var}}_\ell)=V_\ell(\boldsymbol{\theta}^{\mathrm{var}}_\ell)S_\ell(\mathbf{x}_t,\boldsymbol{\theta}^{\mathrm{enc}}_\ell), \tag*{(3)}

and the depth-LL circuit becomes

UL(xt,θenc,θvar)=ALAL−1⋯A1.(4)U_L(\mathbf{x}_t,\boldsymbol{\theta}^{\mathrm{enc}},\boldsymbol{\theta}^{\mathrm{var}})=A_LA_{L-1}\cdots A_1. \tag*{(4)}

Throughout this work, the reuploading depth LL denotes the number of complete encoding-variational blocks. For fixed data encodings, SℓS_\ell contains no trainable encoding parameters, whereas for adaptive encodings, the parameters θenc\boldsymbol{\theta}^{\mathrm{enc}} are optimized jointly with the variational circuit.

The quantum state representing the lag window after LL reuploading layers then reads

∣ψL(xt)⟩=UL(xt)∣0⟩,(5)|\psi_L(\mathbf{x}_t)\rangle=U_L(\mathbf{x}_t)|0\rangle, \tag*{(5)}

where the dependence on trainable parameters has been suppressed for ease of notation. The quantum state is converted into a finite-dimensional classical vector by performing a Pauli-ZZ measurement on all qubits {Zq}q=1w\{Z_q\}_{q=1}^{w}, which is the output of the quantum model,

zt(L)={⟨ψL(xt)|Zq|ψL(xt)⟩}q=1w.(6)\mathbf{z}_t^{(L)}=\left\{\left\langle\psi_L(\mathbf{x}_t)\middle|Z_q\middle|\psi_L(\mathbf{x}_t)\right\rangle\right\}_{q=1}^{w}. \tag*{(6)}

These measurements are passed to a classical output head,

y^t+1=hθhead(zt(L))=tanh⁡(wTzt(L)+b),(7)\widehat{y}_{t+1}=h_{\boldsymbol{\theta}^{\mathrm{head}}}\left(\mathbf{z}_t^{(L)}\right)=\tanh\left(\mathbf{w}^{T}\mathbf{z}_t^{(L)}+b\right), \tag*{(7)}

where y^t+1\widehat{y}_{t+1} denotes the forecast in the transformed target space, which is then transformed back into raw data units using classical post-processing. The forecasting model is trained by minimizing the empirical prediction loss over the training windows.

With the forecasting architecture established, we now turn to the precipitation time series on which it is evaluated. The temporal structure present in the data will provide the basis for examining how recurrent information is represented by the quantum model.

Dataset – precipation time-series

We study a monthly precipitation time series for Napak, Uganda (2.2∘N2.2^\circ\mathrm{N}, 34.2∘E34.2^\circ\mathrm{E}), spanning January 1981 through November 2025. Napak is located in the drought-vulnerable Karamoja sub-region, an agro-pastoral area characterized by high climate variability and frequent, acute rainfall deficits [49]. Demonstrating the efficacy of quantum machine learning algorithms in a challenging region with such acute environmental dynamics establishes a robust baseline for extending the proposed architecture to similar climate-vulnerable regions globally. The monthly precipitation record was extracted from the ECMWF ERA5-Land Monthly Aggregated dataset (ECMWF/ERA5_LAND/MONTHLY_AGGR) via Google Earth Engine [6]. We retrieved the accumulated monthly precipitation (total_precipitation_sum) and converted the values from meters to millimeters.

The resulting record contains 539 consecutive monthly precipitation observations. We use a lag window of length w=14w = 14 months to predict precipitation in the subsequent month. Thus, each supervised example consists of 14 consecutive monthly observations followed by a one-month-ahead target. Figure 2(a) shows the complete time series chronologically divided into 350 training targets spanning March 1982–April 2011, 70 validation targets spanning May 2011–February 2017 and 105 test targets spanning March 2017–November 2025. This chronological split preserves the temporal ordering of the forecasting problem while preventing leakage from future intervals into model training.

Dataset and seasonal structure

Figure 2. Dataset and seasonal structure (a) Monthly precipitation time series for Napak, partitioned chronologically into training, validation, and test intervals. (b) Normalized Fourier power of the training data only, computed after mean subtraction. The Fourier spectrum has a peak at approximately 12 months, indicating annual recurrence of the seasonal dataset.

The signal exhibits substantial month-to-month and year-to-year variability while retaining a recurrent annual structure. To characterize this recurrence, we compute the Fourier spectrum from the training data after subtracting the mean. As shown in Figure 2(b), the data exhibits a dominant annual frequency corresponding to a period of approximately 12 months. The data, however, is not perfectly periodic; lag windows separated by one annual cycle are generally similar but not identical. This motivates the central question considered next: how does a quantum data reuploading circuit distinguish such approximately recurrent input windows, and how does that distinction change as the reuploading depth increases?

Theory – ideal quantum circuit

Having shown that the time series has a characteristic recurrence period T0=12T_0 = 12 months and is approximately periodic, we ask how the quantum model represents two lag windows separated by one such period, xt\mathbf{x}_t and xt+T0\mathbf{x}_{t+T_0}. This comparison isolates the effect of repeated data encoding on temporally related inputs and provides the basis for analyzing how depth modifies the representation of seasonal structure.

The similarity between two such inputs is measured by the fidelity,

FL(t,t+T0)=∣⟨ψL(xt)∣ψL(xt+T0)⟩∣2,(8)F_L(t,t+T_0)=\left|\langle\psi_L(\mathbf{x}_t)|\psi_L(\mathbf{x}_{t+T_0})\rangle\right|^2, \tag*{(8)}

with 0≤FL≤10 \le F_L \le1. A value FL=1F_L = 1 corresponds to identical quantum states up to a global phase, whereas smaller values indicate greater distinguishability in the quantum representation. Suppose that the time series is perfectly periodic with period T0T_0, that is, yn+T0=yn∀ny_{n+T_0} = y_n \forall n. Every component of the lag vector then repeats xt+T0=xt\mathbf{x}_{t+T_0} = \mathbf{x}_t. Since the trained circuit is a deterministic function of the input and trained parameters, UL(xt+T0)=UL(xt)U_L(\mathbf{x}_{t+T_0}) = U_L(\mathbf{x}_t) for arbitrary reuploading depth and arbitrary variational parameters. Consequently,

FL(t,t+T0)=1∀L.(9)F_L(t,t+T_0)=1 \quad\forall L. \tag*{(9)}

Thus, repeated reuploading of the same data cannot distinguish two identical windows. Any depth-dependent change in the representation of seasonally related inputs must therefore originate from the fact that real seasonal signals are only approximately recurrent.

We describe this imperfect recurrence by writing

xt+T0=xt+δt,(10)\mathbf{x}_{t+T_0}=\mathbf{x}_t+\boldsymbol{\delta}_t, \tag*{(10)}

where δt=(δt,1,…,δt,w)T\boldsymbol{\delta}_t=(\delta_{t,1},\ldots,\delta_{t,w})^T contains the cycle-to-cycle differences between corresponding components of the lag window. The displacement δt\boldsymbol{\delta}_t is a property of the time series and exists before any input is processed by the quantum model. The role of the quantum circuit is therefore not to create this seasonal mismatch, but to determine how strongly this mismatch is represented in Hilbert space.

To make this explicit, consider sufficiently small δt\boldsymbol{\delta}_t. The state associated with the displaced window can be expanded around xt\mathbf{x}_t as

∣ψL(xt+δt)⟩=∣ψL(xt)⟩+∑j=1wδt,j∂j∣ψL(xt)⟩+O(∥δt∥2).(11)\left|\psi_L(\mathbf{x}_t+\boldsymbol{\delta}_t)\right\rangle=\left|\psi_L(\mathbf{x}_t)\right\rangle+\sum_{j=1}^{w}\delta_{t,j}\partial_j\left|\psi_L(\mathbf{x}_t)\right\rangle+O(\lVert\boldsymbol{\delta}_t\rVert^2). \tag*{(11)}

The local distinguishability of two seasonally related windows is therefore controlled by the derivatives of the depth LL quantum state with respect to the input coordinates. Determining these derivatives requires accounting for the contribution of every reupload and the circuit transformations that precede it. Assume that the encoding is differentiable with respect to the chosen continuous input coordinates. For the jthj^{\text{th}} component of the lag window at reuploading layer ℓ\ell, we define the encoding generator as

Gj,ℓ=iSℓ†(xt)∂jSℓ(xt),(12)G_{j,\ell}=iS_{\ell}^{\dagger}(\mathbf{x}_t)\partial_jS_{\ell}(\mathbf{x}_t), \tag*{(12)}

so that

∂jSℓ(xt)=−iSℓ(xt)Gj,ℓ.(13)\partial_jS_{\ell}(\mathbf{x}_t)=-iS_{\ell}(\mathbf{x}_t)G_{j,\ell}. \tag*{(13)}

This generator describes the sensitivity of a single encoding layer to a change in input coordinate xjx_j. In a data-reuploading circuit, however, the same coordinate enters at every encoding layer, so the derivative of the complete circuit contains one contribution from each reupload. Define the partial circuits

Bℓ=AℓAℓ−1⋯A1withB0=I,(14)B_{\ell}=A_{\ell}A_{\ell-1}\cdots A_1 \quad\text{with} \quad B_0=I, \tag*{(14)}

so that Bℓ−1B_{\ell-1} contains all operations applied before the ℓth\ell^{\text{th}} reupload. The resulting effective generator is

Gj(L)=iUL†∂jUL=∑ℓ=1LBℓ−1†Gj,ℓBℓ−1.(15)G_j^{(L)}=iU_L^{\dagger}\partial_jU_L=\sum_{\ell=1}^{L}B_{\ell-1}^{\dagger}G_{j,\ell}B_{\ell-1}. \tag*{(15)}

A straightforward calculation using Eq. (11) and Eq. (15) shows that the corresponding fidelity admits the local expansion

FL(t,t+T0)=1−∑j,k=1wgjk(L)δt,jδt,k+O(∥δt∥3),(16)F_L(t,t+T_0)=1-\sum_{j,k=1}^{w}g_{jk}^{(L)}\delta_{t,j}\delta_{t,k}+O(\lVert\boldsymbol{\delta}_t\rVert^3), \tag*{(16)}

where g(L)g^{(L)} is the Fubini-Study metric [38, 40] induced by the trained depth-LL circuit in input space. The metric can be written as the symmetrized covariance

gjk(L)=12⟨{ΔGj(L),ΔGk(L)}⟩0,(17)g_{jk}^{(L)}=\frac{1}{2}\left\langle\left\{\Delta G_j^{(L)},\Delta G_k^{(L)}\right\}\right\rangle_0, \tag*{(17)}

where ΔGj(L)=Gj(L)−⟨Gj(L)⟩0\Delta G_j^{(L)}=G_j^{(L)}-\langle G_j^{(L)}\rangle_0 and ⟨O⟩0=⟨0∣O∣0⟩\langle O\rangle_0=\langle0|O|0\rangle. The structure is particularly transparent when projected directly onto the seasonal displacement δt\boldsymbol{\delta}_t. Define

Gδ,ℓ=∑j=1wδt,jGj,ℓ.(18)G_{\boldsymbol{\delta},\ell}=\sum_{j=1}^{w}\delta_{t,j}G_{j,\ell}. \tag*{(18)}

Equation (16) becomes

1−FL(t,t+T0)≃Var⁡0(∑ℓ=1LBℓ−1†Gδ,ℓBℓ−1),(19)1-F_L(t,t+T_0)\simeq\operatorname{Var}_0\left(\sum_{\ell=1}^{L}B_{\ell-1}^{\dagger}G_{\boldsymbol{\delta},\ell}B_{\ell-1}\right), \tag*{(19)}

where equality is understood to leading quadratic order in δt\boldsymbol{\delta}_t. Equation (16) separates two conceptually different contributions. The displacement δt\boldsymbol{\delta}_t is a property of the data and exists before the quantum model is applied. The metric g(L)g^{(L)}, in contrast, describes how strongly the trained depth-LL circuit resolves that displacement. The details of the derivation are presented in the Supplemental Material.

The role of encoding is particularly transparent at a single reuploading layer. At L=1L=1 the fidelity becomes

F1(t,t+T0)=∣⟨0∣S1†(xt)V1†V1S1(xt+T0)∣0⟩∣2=∣⟨0∣S1†(xt)S1(xt+T0)∣0⟩∣2.(20)\begin{aligned} F_1(t,t+T_0)=\left|\langle0|S_1^{\dagger}(\mathbf{x}_t)V_1^{\dagger}V_1S_1(\mathbf{x}_{t+T_0})|0\rangle\right|^2 \\ &=\left|\langle0|S_1^{\dagger}(\mathbf{x}_t)S_1(\mathbf{x}_{t+T_0})|0\rangle\right|^2. \tag*{(20)} \end{aligned}

The single reupload layer fidelity is therefore determined entirely by the encoding map. Consequently, F1<1F_1<1 reflects the imperfect recurrence δt\boldsymbol{\delta}_t already present in the data in the Hilbert space. At higher depths, intermediate variational blocks occur between repeated encoding operations and rotate the layerwise generators appearing in Eq. (15), thereby changing the geometry of the temporal representation.

For L≫1L\gg1, FL(t,t+T0)<F1(t,t+T0)F_L(t,t+T_0)<F_1(t,t+T_0) implies that the effective generator, which is transformed by all preceding encoding and variational layers through BℓB_{\ell}, leads to degradation in the temporal representation in the quantum circuit. However, in general, multiple reupload depth may suppress, amplify, or redirect the pre-existing differences in the real time series. Importantly, Eqs. (19) and (16) do not imply that larger fidelity leads to better forecasting. A useful predictor must remain sensitive to genuine changes that contain predictive information. The result instead establishes that reuploading depth changes the geometry with which the circuit distinguishes temporally related inputs. Whether this change is beneficial depends on its alignment with the data’s predictive structure.

This observation provides the central motivation for the encoding strategies considered in this work. An encoding based only on the instantaneous value of an observation has no explicit information about where the observation occurs within a recurring temporal cycle. This suggests a general design principle for the quantum time series models: when a known temporal structure is predictive, it can be advantageous to encode the structure explicitly rather than requiring the variational circuit to infer it solely from repeated value-based reuploads. In the following sections, we apply this idea to construct quantum encodings that progressively incorporate and adapt to its seasonal organization.

Quantum data encodings

The analysis in Sec. IV showed that the geometry of approximately recurring inputs depends on the data encoding map, while Sec. III identified a dominant annual frequency with T0=12T_0 = 12 months. Therefore, we will compare a sequence of quantum encodings that progressively incorporate this temporal structure.

The first two encodings depend only on precipitation magnitude and serve as value-based baselines. The remaining three encodings explicitly account for the position of each observation within the annual cycle, with the last two allowing this seasonal geometry to adapt during training. In all five cases, the lag windows, variational circuit, quantum readout, and classical prediction head are kept fixed, so that the differences between models arise from data representation itself.

For an input window xt=(yt−w+1,…,yt)T\mathbf{x}_t = (y_{t-w+1}, \ldots, y_t)^{\mathsf{T}} qubit q∈{1,…,w}q \in\{1, \ldots, w\} represents the observation at the index

rtq=t−w+1+(q−1).(21)r_{tq} = t - w + 1 + (q - 1). \tag*{(21)}

Since the time series has a characteristic period T0=12T_0 = 12, we associate each observation with the seasonal index

mtq=rtqmodT0.(22)m_{tq} = r_{tq} \mathop{\mathrm{mod}} T_0. \tag*{(22)}

This implies that

mtq∈{0,…,T0−1}.(23)m_{tq} \in\{0, \ldots, T_0 - 1\}. \tag*{(23)}

Thus, rtqr_{tq} identifies the observation within the full time series, whereas mtqm_{tq} identifies its position within the recurring annual cycle.

At reuploading layer ℓ\ell, the complete encoding operation is constructed from single-qubit operations as

⨂q=0w−1Sℓ,q(yrtq,mtq).(24)\bigotimes_{q=0}^{w-1} S_{\ell,q}\left(y_{r_{tq}},m_{tq}\right). \tag*{(24)}

The dependence of mtqm_{tq} is absent for value-based encodings introduced below and becomes explicit only for seasonal encodings.

Value based encodings

We first consider two fixed encodings that depend only on the precipitation magnitude. Let y~r∈[−1,1]\widetilde{y}_r \in[-1,1] denote the linearly scaled precipitation value corresponding to the physical observation yry_r. Define the associated rotation angle πy~r\pi\widetilde{y}_r. Since these encodings contain no trainable encoding parameters and are identical at every reuploading layer, we suppress the layer index ℓ\ell.

Ry encoding

The simplest encoding maps each scaled precipitation value to a rotation about the yy axis. The complete window encoder is, therefore,

S(Ry)(xt)=⨂q=0w−1Ry(πy~rtq).(25)S^{(R_y)}(\mathbf{x}_t) = \bigotimes_{q=0}^{w-1} R_y\left(\pi\widetilde{y}_{r_{tq}}\right). \tag*{(25)}

The encoded state is restricted to a single great circle around the Bloch sphere or, in other words, on a single Bloch sphere meridian as shown in Figure 3(a). The magnitude of precipitation determines the position along this one-dimensional trajectory and contains no information about the seasonal position of the observation.

Bloch-sphere geometry of the quantum data encodings

Figure 3. Bloch-sphere geometry of the quantum data encodings Each point represents one of the 364 monthly observations in the training set (or equivalently, 350 training targets) and is colored according to calendar month. (a) The RyR_y encoding maps the scaled precipitation values onto a single meridian. (b) The RxRyRzR_xR_yR_z encoding applies the same data-dependent angle about all three rotation axes, producing a more intricate trajectory over the sphere. (c) The fixed seasonal encoding maps the empirical training data’s cumulative distribution to the polar coordinate and assigns the calendar month to the azimuthal coordinate. Overall, it maps the data onto 12 equally spaced seasonal meridians on the Bloch sphere. (d) The learned polar encoding starts from the fixed seasonal encoding as the reference but later introduces month- and layer-dependent deformations of the polar mapping while retaining the fixed seasonal azimuths. (e) The learned polar + azimuthal encoding additionally permits bounded, month- and layer-dependent azimuthal adaptations in addition to the trainable polar coordinate. It also starts from the fixed seasonal encoding as the reference. For (d)(e), the embeddings shown here are representative of a single layer from a single trained reuploading model; the learned polar mapping generally changes from layer to layer and from model to model.

RxRyRz encoding

The second encoding applies the same data-dependent angle about all three rotation axes. The corresponding encoding operator is

S(RxRyRz)(xt)=⨂q=0w−1Rz(πy~rtq)Ry(πy~rtq)Rx(πy~rtq).(26)S^{(R_xR_yR_z)}(\mathbf{x}_t) = \bigotimes_{q=0}^{w-1} R_z\left(\pi\widetilde{y}_{r_{tq}}\right)R_y\left(\pi\widetilde{y}_{r_{tq}}\right)R_x\left(\pi\widetilde{y}_{r_{tq}}\right). \tag*{(26)}

This construction generates a more intricate trajectory than the RyR_y encoding, but the same scalar πy~rtq\pi\widetilde{y}_{r_{tq}} controls all three rotations as shown in Figure 3(b). This keeps the representation one-dimensional with respect to the information supplied by each observation.

Fixed seasonal CDF encoding

The RyR_y and RxRyRzR_xR_yR_z encodings introduced above distinguish observations only through their precipitation magnitude. To incorporate seasonal recurrence, we construct a two-dimensional representation in which the precipitation magnitude and the seasonal phase control two independent angular coordinates on the Bloch sphere. We first map precipitation magnitude using the empirical cumulative distribution function derived from the training data. Let the empirical training CDF (cumulative distribution function) be denoted by CDF⁡(y)\operatorname{CDF}(y) as defined in Eq. (74) in the Supplemental Material. For the observation encoded on qubit qq, define

utq=CDF⁡(yrtq)andζtq=2utq−1,(27)u_{tq} = \operatorname{CDF}(y_{rtq}) \quad\text{and} \quad\zeta_{tq} = 2u_{tq} - 1, \tag*{(27)}

where ζtq∈[−1,1]\zeta_{tq} \in[-1,1]. The corresponding polar angle is

αtqCDF⁡=arccos⁡(ζtq).(28)\alpha_{tq}^{\operatorname{CDF}} = \arccos(\zeta_{tq}). \tag*{(28)}

The polar coordinate is determined by the relative position of the precipitation value within the training distribution rather than by its absolute magnitude. The seasonal coordinate is determined by the position of the same observation within the annual cycle.

Using the seasonal index mtqm_{tq} defined above, we assign the fixed azimuthal angle,

ϕtq=2πmtqT0.(29)\phi_{tq} = \frac{2\pi m_{tq}}{T_0}. \tag*{(29)}

For T0=12T_0 = 12, the annual cycle is therefore partitioned into 12 equally spaced azimuthal directions, or in other words, the data are encoded on 12 seasonal Bloch sphere meridians as shown in Figure 3(c). Correspondingly, the encoding operator becomes

S(CDF⁡)(xt)=⨂q=0w−1Rz(ϕtq) Ry(αtqCDF⁡).(30)S^{(\operatorname{CDF})}(\mathbf{x}_t) = \bigotimes_{q=0}^{w-1} R_z(\phi_{tq})\,R_y\left(\alpha_{tq}^{\operatorname{CDF}}\right). \tag*{(30)}

The precipitation controls the polar coordinate, while the calendar month controls the azimuthal coordinate. This construction embeds the known annual structure into the quantum representation before any variational transformation is applied. In contrast to RyR_y and RxRyRzR_xR_yR_z encodings, two observations with similar precipitation magnitude but different calendar months are no longer represented identically, and observations occurring at the same calendar month in different years lie on the same seasonal Bloch sphere meridians even when their magnitudes differ.

Adaptive seasonal encodings

The fixed seasonal CDF encoding imposes a physically interpretable geometry in which precipitation magnitude determines the polar coordinate and calendar month determines the azimuthal angle. Next, we allow this geometry to adapt during training while retaining the annual recurrence structure as the reference configuration. In both adaptive encodings, the trainable encoding parameters depend on the seasonal index mtqm_{tq} and on the reuploading layer ℓ\ell, so that different phases of the annual cycle may be represented differently at different reupload depths.

Learned polar seasonal encoding

First, we allow the polar coordinate to vary while keeping the seasonal azimuthal coordinate fixed. The CDF derived reference polar angle αtq(CDF)\alpha_{t_q}^{(\mathrm{CDF})} is transformed through a month and layer dependent map αtq,ℓ(LP)=Amtq,ℓ(αtq(CDF))\alpha_{t_q,\ell}^{(\mathrm{LP})}=A_{m_{t_q},\ell}\left(\alpha_{t_q}^{(\mathrm{CDF})}\right) where Am,ℓA_{m,\ell} denotes a trainable deformation of the polar coordinate. The seasonal azimuthal angle remains fixed as ϕtq\phi_{t_q} so that the seasonal information is preserved.

The encoding at reuploading layer ℓ\ell is, therefore,

Sℓ(LP)(xt)=⨂q=0w−1Rz(ϕtq)Ry(αtq,ℓ(LP)).(31)S_{\ell}^{(\mathrm{LP})}(\mathbf{x}_t)=\bigotimes_{q=0}^{w-1}R_z\left(\phi_{t_q}\right)R_y\left(\alpha_{t_q,\ell}^{(\mathrm{LP})}\right). \tag*{(31)}

The trainable map is initialized to the identity, so it reduces to the fixed seasonal CDF encoding before training. Optimization subsequently allows the model to reshape the precipitation-dependent polar coordinate independently for each seasonal phase and reuploading layer, while preserving the fixed cyclic organization in azimuthal coordinate as shown in Figure 3(d).

Learned polar + azimuthal seasonal encoding

The final encoding also allows the azimuthal coordinate to adapt by a trainable month- and layer-dependent map ϕtq+Δϕmtq,ℓ\phi_{t_q}+\Delta\phi_{m_{t_q},\ell}. The polar coordinate is adapted the same way as in the learned polar seasonal encoding. The corresponding encoding operator is

Sℓ(LPA)(xt)=⨂q=0w−1Rz(ϕtq,ℓ(LPA))Ry(αtq,ℓ(LPA)).(32)S_{\ell}^{(\mathrm{LPA})}(\mathbf{x}_t)=\bigotimes_{q=0}^{w-1}R_z\left(\phi_{t_q,\ell}^{(\mathrm{LPA})}\right)R_y\left(\alpha_{t_q,\ell}^{(\mathrm{LPA})}\right). \tag*{(32)}

The azimuthal displacement Δϕm,ℓ\Delta\phi_{m,\ell} is bounded so that the model can locally adjust the seasonal phase associated with each month without arbitrarily destroying the cyclic ordering of calendar months as shown in Figure 3(e). The trainable map is initialized such that the encoding reduces exactly to the fixed seasonal CDF encoding. The three encodings consequently form a nested family. The fixed seasonal CDF encoding imposes the reference geometry.

The value-based encodings in Figure 3(a),(Figure 3b) map the time-series data exclusively according to their precipitation magnitude and completely ignore their seasonal structure. The seasonal encodings in Figure 3(c),(Figure 3d),(Figure 3e) take the seasonality into account and map the data onto 12 different meridians on the Bloch sphere. The adaptive variants deform this seasonal geometry through their learned encoding parameters. In both adaptive encodings, the encoding parameters are optimized jointly with other trainable parameters of the forecasting model. The details of the various encoding schemes are presented in the Supplemental Material.

Results

We evaluate the five encodings within the same 14-qubit data reuploading forecasting architecture, varying the number of reuploading layers over L=1,…,6L=1,\ldots,6. For each encoding and depth, five independent models are trained, and model selection is performed exclusively on the validation set. The checkpoint with the lowest validation loss is retained for each run and evaluated once on the test set.

Forecasting performance is assessed using four complementary quantities: the test root mean squared error (RMSE) (Eq. (100)), which measures absolute prediction error in millimeters, the Pearson correlation (Eq. (106)), which measures temporal alignment between the observed and predicted series, the Fourier power cosine similarity CFFTC_{\mathrm{FFT}} (Eq. (105)), which compares their normalized frequency domain structure; and the Amplitude variance ratio (Eq. (107)), which measures recovery of the observed forecast amplitude. The detailed definitions of these metrics are given in the Supplemental Material.

Ideal simulations

The ideal simulations were performed using the default.qubit state-vector simulator and automatic differentiation framework by PennyLane [50]. The complete hybrid quantum-classical learning pipeline was implemented using PyTorch [51].

Representation of seasonal recurrence with depth

We first examine how the different encodings represent inputs separated by the dominant seasonal period. This tests the quantum geometric picture developed in Sec. IV before considering forecasting performance.

For a depth-LL circuit, we define the mean seasonal fidelity over all valid test set window pairs separated by T0T_0 as

F‾L(T0)=1NT0,test∑t∣⟨ψL(xt)∣ψL(xt+T0)⟩∣2,(33)\overline{F}_L(T_0)=\frac{1}{N_{T_0,\mathrm{test}}}\sum_t\left|\langle\psi_L(\mathbf{x}_t)|\psi_L(\mathbf{x}_{t+T_0})\rangle\right|^2, \tag*{(33)}

where NT0,testN_{T_0,\mathrm{test}} is the number of valid pairs in the test data. For our time series, T0=12T_0=12 months. Thus, F‾L(T0)\overline{F}_L(T_0) measures the average similarity between quantum representations of lag windows separated by one annual cycle.

Figure 4(a) shows that the mean seasonal fidelity decreases with reuploading depth for all five encodings. The rate and scale of this decrease, however, depend strongly on how the data are represented. The RyR_y and RxRyRzR_xR_yR_z encodings begin with relatively small seasonal overlaps and decrease rapidly with depth. By contrast, the three seasonal encodings produce substantially larger overlaps at shallow depth but also exhibit a pronounced decline with LL.

Same-season quantum fidelity as a function of reuploading depth $L$

Figure 4. Same-season quantum fidelity as a function of reuploading depth LL. (a) Mean seasonal fidelity F‾L(12)\overline{F}_L(12) measures the similarity between two quantum representations of two lag windows separated by 12 months, averaged over all possible pairs of lag windows in the test set. (b) The same fidelity normalized by expected fidelity between Haar-random states, highlighting similarity compared to a random state reference. Curves show the mean across five independently trained seeds, with shaded region indicating ±\pm standard deviation.

The single-layer behavior is consistent with the geometric analysis of Sec. IV. For fixed encodings, the variational block cancels exactly from the fidelity at L=1L = 1, so that the overlap is determined entirely by the encoding map. Correspondingly, the fixed seasonal CDF representation gives a substantially larger single-layer seasonal overlap than either value-based encoding. The adaptive encodings are initialized from the same seasonal geometry but subsequently optimize their encoding parameters during training, allowing the learned representation itself to change.

To interpret the magnitude of these fidelities in context, we compare them with the characteristic fidelity of two independent Haar-random pure states [52, 53] in a Hilbert space of dimension d=214d = 2^{14}. For such states, E[FHaar]=1/d\mathbb{E}[F_{\mathrm{Haar}}] = 1/d. We consider the normalized quantity F‾L(12)/E[FHaar]=dF‾L(12)\overline{F}_L(12)/\mathbb{E}[F_{\mathrm{Haar}}] = d\overline{F}_L(12) shown in Figure 4(b). At L=6L = 6, the RyR_y and RxRyRzR_xR_yR_z encodings have decreased to the FHaarF_{\mathrm{Haar}} scale. The fixed seasonal CDF representation remains above FHaarF_{\mathrm{Haar}}, better than the value-based encodings. In contrast, the two adaptive seasonal encodings remain roughly two orders of magnitude above the Haar reference even at L=6L = 6. These results show that repeated data reuploading progressively separates the quantum representations of input windows related to different years, but the extent of this separation depends strongly on the encoding geometry.

Importantly, F‾L(12)\overline{F}_L(12) is a diagnostic of representation geometry rather than a forecasting metric. A reduction in fidelity does not by itself imply a loss of predictive information, since the seasonal differences between xt\mathbf{x}_t and xt+T0\mathbf{x}_{t+T_0} may themselves be informative. We therefore next examine whether the depth-dependent changes in quantum geometry are accompanied by corresponding changes in forecasting performance.

Forecasting performance

Figure 5 summarizes the test data predictions as a function of reploading depth LL. Figure 5(a) shows a clear separation between the seasonal and value-based encodings in terms of prediction error. The RyR_y and RxRyRzR_xR_yR_z encodings attain their lowest RMSE at shallow depth and generally deteriorate as additional reuploading layers are introduced. The fixed seasonal CDF encoding substantially improves upon these baselines at small depth, but its error also increases progressively with LL. In contrast, both learned seasonal encodings remain comparatively stable across the full depth range. The lowest test error is obtained by the learned polar encoding at L=2L = 2, Test RMSE =52.81±2.55= 52.81 \pm2.55 mm, whereas the best value-based encodings remain near 63 mm.

A similar hierarchy appears in the frequency-domain results of Figure 5(b). The learned seasonal encodings preserve an almost unchanged spectral structure across the complete depth range, with CFFT≃0.96C_{\mathrm{FFT}} \simeq0.96–0.970.97. The fixed seasonal CDF encoding also retains strong spectral agreement at shallow and intermediate depth, although its similarity decreases at the largest LL. By comparison, both value-based encodings exhibit substantially stronger depth-dependent fluctuations and ultimately lose much of their agreement with the observed spectrum.

Four-panel plot of forecasting metrics as a function of reuploading depth across five encodings

Figure 5. Forecasting metrics shown as a function of reuploading depth LL across the five different encodings. (a) Test root mean square error (RMSE) measures the magnitude of pointwise forecast error in precipitation units (mm). Lower values indicate better predictive accuracy. (b) Fourier cosine similarity measures the similarity between the normalized Fourier power spectra of the predicted and observed time series, thereby probing how well the forecast preserves dominant temporal frequencies. (c) Pearson correlation measures the linear correlation between the predicted and observed time series and captures whether temporal fluctuations are tracked correctly, independent of the absolute scale. (d) Amplitude variance ratio assesses whether the forecast reproduces the observed variability in amplitude. A value of 1 indicates matched variance. Curves denote the mean over five independently trained seeds, and shaded regions indicate ±1\pm1 standard deviation.

The Pearson correlations in Figure 5(c) show an even clearer difference in depth dependence. Both learned seasonal encodings maintain strong and comparatively stable temporal alignment throughout the entire range of depths. The learned polar encodings reach their largest correlation 0.634±0.0440.634 \pm0.044 at L=2L = 2, and remain close to this level at larger depths. The learned polar + azimuthal encoding exhibits a similar trend. The fixed seasonal encoding shows a more gradual loss of correlation with depth, whereas the value-based encodings deteriorate much more strongly and approach negligible temporal correlation by L=6L = 6. Both learnable seasonal models show a slight increase in performance at L=2L = 2 that is not seen in any other model. This highlights the distinction between predictive model performance and the seasonal fidelity of Eq. (33) and indicates that the optimization is able to effectively utilize some of the enhanced expressive capacity from data reuploading.

The amplitude variance ratios in Figure 5(d) reveal a complementary limitation. All five models underestimate the amplitude of the observed precipitation fluctuations throughout the depth range. The learned seasonal models nevertheless retain substantially more variability than the value-based encodings and exhibit much weaker amplitude collapse as the circuit becomes deeper. The largest amplitude recovery occurs for the learned polar model at L=1L = 1.

Taken together, the four metrics show that the learned seasonal encodings are not only distinguished by a lower pointwise error, but also preserve temporal correlation, normalized spectral structure, and forecast amplitude much more consistently as reuploading depth increases. The fixed seasonal CDF representation captures a substantial fraction of this improvement at shallow depth, but its performance degrades more strongly with LL, indicating that explicitly encoding the annual cycle is beneficial while allowing the corresponding geometry to adapt provides additional robustness.

Observed monthly precipitation and forecasts for the $R_y$, fixed seasonal, and learned polar encodings

Figure 6. Observed monthly precipitation is compared with the mean forecast across five independently trained seeds for the RyR_y, fixed seasonal, and learned polar encodings over the first 36 months of the test data. (a) Shallow reuploading circuit with L=1L=1. (b) Deepest reuploading circuit with L=6L=6. At L=6L=6, the learned polar encoding continues to retain good forecasts, the fixed seasonal encoding performs slightly better, and the RyR_y encoding shows the poorest forecast quality.

The comparison with the fidelity results further shows that forecasting performance cannot be identified with preservation of seasonal recurrence alone. The seasonal fidelity of the learned encodings decreases substantially with depth, yet their Pearson correlation and Fourier similarity remain nearly unchanged. The learned seasonal representations therefore continue to retain forecast-relevant temporal structure even as annually separated input windows become increasingly distinguishable in Hilbert space. By contrast, for the value-based encodings, the reduction in seasonal fidelity is accompanied by a significant loss of temporal and spectral forecasting structure.

Finally, allowing the azimuthal coordinate to become trainable does not produce a consistent improvement over adapting the polar coordinate alone. The learned polar and learned polar + azimuthal encodings exhibit very similar behavior across all four forecasting metrics, while the learned polar encoding attains the overall minimum RMSE and maximum Pearson correlation. For this time series, the additional azimuthal degrees of freedom therefore provide no predictive advantage beyond the learned polar representation and that the evenly-spaced seasonal encoding is near optimal.

Noisy Simulations

The ideal simulations show that the seasonal encodings remain comparatively robust as the number of reuploading layers increases. They further establish a clear hierarchy among the seasonal encodings, with the learned seasonal models generally outperforming the fixed seasonal CDF representation and remaining substantially more robust as reuploading depth increases. We next examine if this behavior persists in the presence of noise. For the noisy simulations, we evaluated the three seasonal encoding strategies: Fixed seasonal CDF, learned polar, and learned polar + azimuthal across L=1,…,6L=1,\ldots,6.

Noisy circuit evaluations were performed with Qiskit Aer using the IBM FakeAachen noise model [54]. Circuit executions used 512 measurement shots. The logical circuits were compiled onto a fixed 14-qubit physical layout as shown in Figure 7. The entire forecasting model was trained using a hybrid optimization procedure in which quantum circuit parameters were updated using simultaneous perturbation stochastic approximation (SPSA) [55], while classical parameters were optimized with Adam [56]. To make the noisy simulations computationally tractable, the circuit evaluations were parallelized across multiple CPU cores on an HPC cluster. The details of the exact implementation are presented in the Supplemental Material.

Topology of IBM FakeAachen used for noisy simulations

Figure 7. Topology of IBM FakeAachen used for noisy simulations. The 14-qubit subset selected for implementing the quantum reuploading circuit is highlighted, while the remaining qubits and couplings are shown in gray.

Figure 8 compares the noisy simulations with their ideal counterparts over reuploading depths L=1,…,6L=1,\ldots,6. A central observation is that, at shallow and intermediate depths L=1,2,3L=1,2,3, the hierarchy of encodings established in the ideal simulations is preserved under noise. In particular, the adaptive seasonal encodings retain their advantage over the fixed seasonal CDF encoding across all the forecasting metrics. Thus, the benefit of learning the seasonal encoding geometry is not restricted to ideal state-vector simulations, but is also retained in the presence of hardware-derived noise. At these depths, hardware noise degrades the absolute forecasting performance relative to the ideal case, but it does not immediately erase the relative advantage produced by the encoding choice.

Comparison of ideal and noisy quantum simulations across reuploading depth $L$ for the three seasonal encoding schemes

Figure 8. Comparison of ideal and noisy quantum simulations across reuploading depth LL for the three seasonal encoding schemes. Both regimes are shown for a single matched initialization, rather than as an ensemble average across seeds. Solid lines with filled markers denote ideal simulations, while dashed lines with open markers denote simulations using the IBM FakeAachen noise model.

The situation changes as additional reuploading layers are introduced. We again observe that all adaptive seasonal encoding models perform optimally at L=2L = 2, an observation that is not true of the fixed seasonal encodings. Beyond L=3L = 3, the forecasting metrics deteriorate substantially for all three encodings and the differences between them become progressively compressed. The test RMSE increases and eventually enters a similar high-error regime for the three seasonal models, while the Pearson correlation and forecast amplitude decrease more strongly with depth. By the deepest circuits, the predictions retain little temporal correlation with the observed precipitation series and exhibit severe amplitude suppression and spectral dissimilarity in the Fourier domain. These trends indicate that the advantages associated with the different seasonal geometries become increasingly difficult to exploit as the depth and noise grow.

The degradation in performance at higher depths should also be interpreted in the context of the hybrid optimization procedure. During the noisy simulations, we find that convergence is sensitive to the SPSA hyperparameters, particularly the parameter controlling the gradient updates [57]. This sensitivity becomes increasingly relevant as the circuit depth grows and the stochastic optimization landscape becomes more difficult to navigate. Consequently, the deterioration observed at large LL need not arise solely from the accumulation of hardware noise; it may also reflect the increasing difficulty of jointly optimizing the noisy quantum circuit and classical prediction head. The performance of the noisy models could also potentially be improved through error mitigation strategies [58, 59, 60]. We therefore leave a combined systematic study of the encoding design, depth-dependent optimizer tuning, and error mitigation as an important direction for future work.

Classical Baselines

Having established that the seasonal quantum encodings substantially outperform the standard encodings as reuploading depth increases and this robustness persists under noise, we finally compare the resulting quantum models with conventional classical forecasting approaches. We consider three complementary baselines: a random forest (RF), representing a nonlinear machine learning model; an autoregressive integrated moving average model, ARIMA(7,0,6), representing a statistical time series model; and a stacked long short-term memory (LSTM), representing a recurrent neural network with a comparable notion of depth. All classical models are trained and evaluated over the same chronological training, validation, and test data splits used on the same chronological training, validation, and test data splits as the quantum models. The implementation details are presented in the Supplemental Material.

The RF and ARIMA models provide depth-independent reference points for the forecasting task. For a more controlled comparison with the depth-dependent quantum models, we constructed stacked LSTMs with recurrent depth L=1,…,6L = 1,\ldots,6 and hidden dimension h=3h = 3. The LSTM architecture was intentionally not chosen as a high-capacity, state-of-the-art forecasting model. Instead, its parameters were chosen so that it could serve as a classically comparable model to our learned quantum models, with a parameter count that grows approximately in step with that of the quantum model with equivalent reuploading depth.

Compared with these classical baseline models, the learned polar seasonal-encoding quantum model occupies the same overall forecasting regime, as shown in Figure 9. In ideal simulations, the learnable polar achieved best test RMSEs of 52.81±2.5552.81 \pm2.55. For comparison, the RF achieved 55.28±0.2255.28 \pm0.22 mm, ARIMA(7,0,6) achieved 55.5055.50 mm, and the best mean LSTM result was 52.84 ± 0.82 mm. The same overall picture is reflected in the correlation amplitude and spectral metrics: no classical baseline uniformly dominates across all forecast metrics, and the learned polar quantum model achieves the same performance. In the presence of noise, the performance of quantum model decreased, but still achieved an RMSE of 56.54 mm and retained strong temporal and spectral structure. Together, these results show that the robustness introduced by seasonal quantum encoding is not at the expense of forecasting accuracy; rather, the resulting models remain competitive with established statistical and machine learning methods.

Comparison of the Learned Polar quantum data reuploading model with classical forecasting baselines across depth

Figure 9. Comparison of the Learned Polar quantum data reuploading model with classical forecasting baselines across depth LL. For the quantum model, LL denotes the number of repeated data-encoding and variational-circuit layers; for the LSTM, it denotes the number of stacked recurrent layers. The learned polar quantum model and LSTM are shown as depth-dependent trajectories, whereas random forest and ARIMA serve as depth-independent reference baselines. Shaded regions indicate ±1\pm1 standard deviation across five independent realizations for the learned polar quantum model, LSTM, and ARIMA.

Discussion

In this work, we have demonstrated that the choice of encoding classical data into a quantum circuit can play a central role in determining how a data reuploading model behaves as circuit depth increases. For the rainfall time series considered here, standard value-based encodings progressively lose both seasonal structure and forecasting performance as additional reuploading layers are introduced. Explicitly incorporating the known seasonal recurrence pattern into the quantum feature map substantially changes this behavior. The fixed seasonal encoding shows improved performance at shallow depth, while the adaptive seasonal encodings remain far more robust across the full depth range. All adaptive seasonal encoding showed modest performance improvents at a depth of L=2L = 2 before declining at larger depths. These results suggest that increasing circuit depth alone is not sufficient to improve a quantum time series model (in many cases it is detrimental to performance); the geometry through which temporally structured data are repeatedly presented to the circuit is equally (or more) important.

The geometric analysis provides a framework for interpreting this behavior. Two input windows separated by one seasonal period differ by a small displacement that is already present in the underlying time series, which prevents the time series from being perfectly periodic. The quantum circuit does not create this displacement but determines how strongly it is resolved by the input-space geometry induced by the model. Locally, this relation is captured by the Fubini-Study metric generated by the complete data reuploading circuit via a quantum geometric fidelity analysis. Since the effective input generators receive contributions from every reuploading layer, increasing depth can alter how the same underlying temporal differences are represented. However, preservation of seasonal fidelity should not be interpreted as the forecasting objective itself. Rather, the useful representation is one that preserves alignment of the quantumly encoded data with known seasonal structure. In this sense, the fidelity analysis provides a diagnostic of how temporal relationships are represented, rather than a direct measure of prediction quality. These observations suggest a broader design principle for quantum learning of structured time series. When a physically or statistically meaningful temporal organization is already known, it can be advantageous to incorporate that structure directly into the quantum data encoding rather than requiring the variational circuit to infer it solely through repeated value-based encodings. In the present construction, precipitation magnitude determines the polar coordinate of the encoded qubit state while seasonal phase or calendar month determines the azimuthal coordinate, providing an explicit geometric representation of the annual cycle. The adaptive variants then deform this structured reference geometry during training. The fact that learning the polar deformation provides most of the observed improvement, while additional azimuthal flexibility gives no consistent advantage, further suggests that useful learning need not require unconstrained modification of every geometric degree of freedom. More generally, analogous constructions could incorporate other known temporal periods, phase variables, spatial coordinates, or physical symmetries into quantum representations.

The noisy simulations further show that the benefit of structured encoding is not limited to ideal state-vector evolution. At shallow and intermediate reuploading depths, the ordering established in the ideal simulations is retained under the IBM FakeAachen noise model, with the adaptive seasonal encodings continuing to outperform the fixed seasonal representation. At larger depths, however, the performance of all three seasonal models deteriorates, and their differences become increasingly compressed. This behavior highlights a practical trade-off among representational depth, hardware noise, and the increasing difficulty of hybrid SPSA optimization in a noisy landscape. A systematic separation of these effects, together with noise-aware optimization and error mitigation strategies, remains an important direction for future work.

In the presence of noise, the quantum performance decreases but remains close to the classical reference models at the best shallow and intermediate depths. These results demonstrate that the improved depth robustness achieved through structured quantum encoding can coexist with forecasting performance comparable to that of conventional classical approaches.

Finally, the classical benchmarks place the quantum results in the context of established forecasting methods. Random forests, ARIMA, and LSTM represent complementary standard statistical and machine learning approaches to the same prediction task. The best ideal seasonal quantum models achieve forecasting errors in family with these classical baselines. While it remains an open question whether QML methods can exhibit any quantum advantage, it is generally expected that quantum methods might be superior in analyzing correlated data sets [9]. The present comparison of the quantum algorithm with classical methods gives reason to be cautiously optimistic about the realizability of more intricate quantum machine learning protocols. Thus, an important next step is to extend our paradigm beyond univariate forecasting to multivariate time series, where temporal structure coexists with intervariable dependencies [61, 62]. In this setting, structure-aware quantum encodings could be used to investigate how temporal and cross-variable relationships are jointly represented, and how entangling operations provide useful representational structure [63, 64, 65, 66].

A complementary direction is to evaluate these models on real quantum hardware [67, 68, 69, 70, 71, 72] and to develop more closely matched classical baselines that enable controlled comparisons with the corresponding quantum architectures [73]. Together, these extensions would test how broadly the encoding principles identified here apply as quantum time-series models move toward richer data and experimentally realized circuits.

Data Availability

The code for numerical simulations is available can be found at [74].

Acknowledgments

S.D. acknowledges support from the John Templeton Foundation under Grant No. 63626. During the course of the project, OpenAI’s ChatGPT models were used to help understand Pennylane with PyTorch and Qiskit, to optimize code for HPCs and to orchestrate the project workflow. They were also used to help writing the plotting scripts for the figures.

References

Derivation of the local fidelity geometry

Here, we present the details of the derivation for Eq. (19) in the main text: the relation between the fidelity of two approximately recurrent input windows and the effective generators of a depth LL data reuploading circuit. Consider two lag windows separated by the characteristic period T0T_0 of the time series,

xt+T0=xt+δt,(34)\mathbf{x}_{t+T_0}=\mathbf{x}_t+\boldsymbol{\delta}_t, \tag*{(34)}

where δt=(δt,1,…,δt,w)T\boldsymbol{\delta}_t=(\delta_{t,1},\ldots,\delta_{t,w})^{T} quantifies the displacement between the two windows. Throughout the derivation, all derivatives are taken with respect to the continuous input coordinates entering the encoding operation, while the trained circuit parameters are held fixed. For compactness, we write ∣ψ⟩≡∣ψL(xt)⟩|\psi\rangle\equiv|\psi_L(\mathbf{x}_t)\rangle and evaluate all derivatives at x=xt\mathbf{x}=\mathbf{x}_t unless otherwise stated. For sufficiently small δt\boldsymbol{\delta}_t, the state associated with the displaced window has the Taylor expansion

∣ψL(xt+δt)⟩=∣ψ⟩+∑j=1wδt,j∣∂jψ⟩+12∑j,k=1wδt,jδt,k∣∂j∂kψ⟩+O(∥δt∥3),|\psi_L(\mathbf{x}_t+\boldsymbol{\delta}_t)\rangle=|\psi\rangle+\sum_{j=1}^{w}\delta_{t,j}|\partial_j\psi\rangle+\frac{1}{2}\sum_{j,k=1}^{w}\delta_{t,j}\delta_{t,k}|\partial_j\partial_k\psi\rangle+O(\|\boldsymbol{\delta}_t\|^3),

where ∂j≡∂∂xj\partial_j\equiv\frac{\partial}{\partial x_j}.

Local expansion of the fidelity

The fidelity between the two approximately recurrent windows is

FL(t,t+T0)=∣⟨ψL(xt)∣ψL(xt+δt)⟩∣2.(35)F_L(t,t+T_0)=|\langle\psi_L(\mathbf{x}_t)|\psi_L(\mathbf{x}_t+\boldsymbol{\delta}_t)\rangle|^2. \tag*{(35)}

The corresponding overlap amplitude is

A≡⟨ψ∣ψL(xt+δt)⟩=1+∑jδt,j⟨ψ∣∂jψ⟩+12∑j,kδt,jδt,k⟨ψ∣∂j∂kψ⟩+O(∥δt∥3).\mathcal{A}\equiv\langle\psi|\psi_L(\mathbf{x}_t+\boldsymbol{\delta}_t)\rangle=1+\sum_j\delta_{t,j}\langle\psi|\partial_j\psi\rangle+\frac{1}{2}\sum_{j,k}\delta_{t,j}\delta_{t,k}\langle\psi|\partial_j\partial_k\psi\rangle+O(\|\boldsymbol{\delta}_t\|^3).

For notational convenience, define cj≡⟨ψ∣∂jψ⟩c_j\equiv\langle\psi|\partial_j\psi\rangle, djk≡⟨ψ∣∂j∂kψ⟩d_{jk}\equiv\langle\psi|\partial_j\partial_k\psi\rangle. Then,

A=1+∑jδt,jcj+12∑j,kδt,jδt,kdjk+O(∥δt∥3).(36)\mathcal{A}=1+\sum_j\delta_{t,j}c_j+\frac{1}{2}\sum_{j,k}\delta_{t,j}\delta_{t,k}d_{jk}+O(\|\boldsymbol{\delta}_t\|^3). \tag*{(36)}

The fidelity is obtained from FL=AA∗F_L=\mathcal{A}\mathcal{A}^*. Keeping terms up to second order gives

FL(t,t+T0)=1+∑jδt,j(cj+cj∗)+∑j,kδt,jδt,k[cjck∗+12(djk+djk∗)]+O(∥δt∥3).(37)F_L(t,t+T_0)=1+\sum_j\delta_{t,j}(c_j+c_j^*)+\sum_{j,k}\delta_{t,j}\delta_{t,k}\left[c_jc_k^*+\frac{1}{2}(d_{jk}+d_{jk}^*)\right]+O(\|\boldsymbol{\delta}_t\|^3). \tag*{(37)}

Using ⟨ψ∣ψ⟩=1\langle\psi|\psi\rangle=1, we can write ⟨∂jψ∣ψ⟩+⟨ψ∣∂jψ⟩=0\langle\partial_j\psi|\psi\rangle+\langle\psi|\partial_j\psi\rangle=0 or equivalently cj+cj∗=0c_j+c_j^*=0. Thus, the fidelity contains no term linear in δt\boldsymbol{\delta}_t. Differentiating the normalization condition twice yields

0=∂j∂k⟨ψ∣ψ⟩=⟨∂j∂kψ∣ψ⟩+⟨∂jψ∣∂kψ⟩+⟨∂kψ∣∂jψ⟩+⟨ψ∣∂j∂kψ⟩.(38)\begin{aligned} 0&=\partial_j\partial_k\langle\psi|\psi\rangle\\ &=\langle\partial_j\partial_k\psi|\psi\rangle+\langle\partial_j\psi|\partial_k\psi\rangle+\langle\partial_k\psi|\partial_j\psi\rangle+\langle\psi|\partial_j\partial_k\psi\rangle. \tag*{(38)} \end{aligned}

Taking the real part gives

Re⁡(djk)=−Re⁡⟨∂jψ∣∂kψ⟩.(39)\operatorname{Re}(d_{jk})=-\operatorname{Re}\langle\partial_j\psi|\partial_k\psi\rangle. \tag*{(39)}

Substituting these identities into Eq. (37), and using the symmetry of δt,jδt,k\delta_{t,j}\delta_{t,k} under interchange of jj and kk, gives

FL(t,t+T0)=1−∑j,kδt,jδt,kRe⁡[⟨∂jψ∣∂kψ⟩−⟨∂jψ∣ψ⟩⟨ψ∣∂kψ⟩]+O(∥δt∥3).(40)F_L(t,t+T_0)=1-\sum_{j,k}\delta_{t,j}\delta_{t,k}\operatorname{Re}\left[\langle\partial_j\psi|\partial_k\psi\rangle-\langle\partial_j\psi|\psi\rangle\langle\psi|\partial_k\psi\rangle\right]+O(\|\boldsymbol{\delta}_t\|^3). \tag*{(40)}

We therefore identify the input-space Fubini–Study metric induced by the reuploading depth-LL circuit as

gjk(L)=Re⁡[⟨∂jψ∣∂kψ⟩−⟨∂jψ∣ψ⟩⟨ψ∣∂kψ⟩].(41)g_{jk}^{(L)}=\operatorname{Re}\left[\langle\partial_j\psi|\partial_k\psi\rangle-\langle\partial_j\psi|\psi\rangle\langle\psi|\partial_k\psi\rangle\right]. \tag*{(41)}

Then Eq. (40) becomes

FL(t,t+T0)=1−∑j,k=1wgjk(L)δt,jδt,k+O(∥δt∥3),(42)F_L(t,t+T_0)=1-\sum_{j,k=1}^{w}g^{(L)}_{jk}\delta_{t,j}\delta_{t,k}+O(\lVert\delta_t\rVert^3), \tag*{(42)}

or equivalently,

1−FL(t,t+T0)=δtTg(L)δt+O(∥δt∥3).(43)1-F_L(t,t+T_0)=\delta_t^{T}g^{(L)}\delta_t+O(\lVert\delta_t\rVert^3). \tag*{(43)}

Equation (A10) separates the cycle-to-cycle displacement of the data, δt\delta_t, from the local geometry induced by the quantum model, g(L)g^{(L)}. We next express this metric directly in terms of the generators of the data-reuploading circuit.

Effective input generator of a reuploading circuit

The depth-LL circuit is UL(x)=AL(x)AL−1(x)⋯A1(x)U_L(\mathbf{x})=A_L(\mathbf{x})A_{L-1}(\mathbf{x})\cdots A_1(\mathbf{x}) where each reuploading block is Aℓ(x)=VℓSℓ(x)A_\ell(\mathbf{x})=V_\ell S_\ell(\mathbf{x}). The dependence on the trained circuit parameters is suppressed here for notational clarity. The variational operation VℓV_\ell does not depend explicitly on the input coordinates. For each input coordinate xjx_j and encoding layer ℓ\ell, define the local encoding generator

Gj,ℓ≡iSℓ†(x)∂jSℓ(x).(44)G_{j,\ell}\equiv iS_\ell^\dagger(\mathbf{x})\partial_jS_\ell(\mathbf{x}). \tag*{(44)}

Since SℓS_\ell is unitary, Sℓ†Sℓ=IS_\ell^\dagger S_\ell=I. Differentiating this identity gives

(∂jSℓ†)Sℓ+Sℓ†(∂jSℓ)=0,(45)(\partial_jS_\ell^\dagger)S_\ell+S_\ell^\dagger(\partial_jS_\ell)=0, \tag*{(45)}

which shows that Sℓ†∂jSℓS_\ell^\dagger\partial_jS_\ell is anti-Hermitian. Hence, Gj,ℓG_{j,\ell} defined in Eq. (44) is Hermitian. Multiplying Eq. (44) by SℓS_\ell from the left gives ∂jSℓ=−iSℓGj,ℓ\partial_jS_\ell=-iS_\ell G_{j,\ell}. Because VℓV_\ell is independent of x\mathbf{x},

∂jAℓ=Vℓ∂jSℓ=−iVℓSℓGj,ℓ=−iAℓGj,ℓ.(46)\partial_jA_\ell=V_\ell\partial_jS_\ell=-iV_\ell S_\ell G_{j,\ell}=-iA_\ell G_{j,\ell}. \tag*{(46)}

The same input coordinate is present in every reuploading layer. Therefore,

∂jUL=∑ℓ=1LAL⋯Aℓ+1(∂jAℓ)Aℓ−1⋯A1.(47)\partial_jU_L=\sum_{\ell=1}^{L}A_L\cdots A_{\ell+1}(\partial_jA_\ell)A_{\ell-1}\cdots A_1. \tag*{(47)}

Substituting Eq. (46) gives

∂jUL=−i∑ℓ=1LAL⋯Aℓ+1AℓGj,ℓAℓ−1⋯A1.(48)\partial_jU_L=-i\sum_{\ell=1}^{L}A_L\cdots A_{\ell+1}A_\ell G_{j,\ell}A_{\ell-1}\cdots A_1. \tag*{(48)}

We now define the effective generator of the complete depth-LL circuit,

Gj(L)≡iUL†∂jUL.(49)\mathcal{G}^{(L)}_j\equiv iU_L^\dagger\partial_jU_L. \tag*{(49)}

To evaluate it, define the partial circuit Bℓ=AℓAℓ−1⋯A1B_\ell=A_\ell A_{\ell-1}\cdots A_1, B0=IB_0=I. Using UL†=A1†A2†⋯AL†U_L^\dagger=A_1^\dagger A_2^\dagger\cdots A_L^\dagger, the contribution from reuploading layer ℓ\ell becomes

iUL†[−iAL⋯Aℓ+1AℓGj,ℓAℓ−1⋯A1]=A1†⋯AL†AL⋯Aℓ+1AℓGj,ℓAℓ−1⋯A1=A1†⋯Aℓ−1†Gj,ℓAℓ−1⋯A1=Bℓ−1†Gj,ℓBℓ−1.(50)\begin{aligned} iU_L^\dagger[-iA_L\cdots A_{\ell+1}A_\ell G_{j,\ell}A_{\ell-1}\cdots A_1] &=A_1^\dagger\cdots A_L^\dagger A_L\cdots A_{\ell+1}A_\ell G_{j,\ell}A_{\ell-1}\cdots A_1\\ &=A_1^\dagger\cdots A_{\ell-1}^\dagger G_{j,\ell}A_{\ell-1}\cdots A_1\\ &=B_{\ell-1}^\dagger G_{j,\ell}B_{\ell-1}. \tag*{(50)} \end{aligned}

Summing over all reuploading layers gives

Gj(L)=∑ℓ=1LBℓ−1†Gj,ℓBℓ−1.(51)\mathcal{G}^{(L)}_j=\sum_{\ell=1}^{L}B_{\ell-1}^\dagger G_{j,\ell}B_{\ell-1}. \tag*{(51)}

Thus, each reupload contributes a generator associated with the same input coordinate, but the contribution from layer ℓ\ell is conjugated by all circuit operations that precede that layer. Equation (A16) also implies

∂jUL=−iULGj(L).(52)\partial_j U_L=-iU_LG_j^{(L)}. \tag*{(52)}

The depth-LL state is ∣ψL(x)⟩=UL(x)∣0⟩|\psi_L(\mathbf{x})\rangle=U_L(\mathbf{x})|0\rangle, and therefore,

∣∂jψ⟩=(∂jUL)∣0⟩=−iULGj(L)∣0⟩.(53)|\partial_j\psi\rangle=(\partial_jU_L)|0\rangle=-iU_LG_j^{(L)}|0\rangle. \tag*{(53)}

Fubini-Study metric as a generator covariance

We can now substitute Eq. (A20) directly into the Fubini-Study metric of Eq. (A8). First,

⟨∂jψ∣∂kψ⟩=⟨0∣Gj(L)UL†ULGk(L)∣0⟩=⟨0∣Gj(L)Gk(L)∣0⟩.(54)\langle\partial_j\psi|\partial_k\psi\rangle=\langle0|G_j^{(L)}U_L^\dagger U_LG_k^{(L)}|0\rangle=\langle0|G_j^{(L)}G_k^{(L)}|0\rangle. \tag*{(54)}

Similarly, ⟨ψ∣∂jψ⟩=−i⟨0∣Gj(L)∣0⟩\langle\psi|\partial_j\psi\rangle=-i\langle0|G_j^{(L)}|0\rangle and hence,

⟨∂jψ∣ψ⟩⟨ψ∣∂kψ⟩=⟨Gj(L)⟩0⟨Gk(L)⟩0,(55)\langle\partial_j\psi|\psi\rangle\langle\psi|\partial_k\psi\rangle=\langle G_j^{(L)}\rangle_0\langle G_k^{(L)}\rangle_0, \tag*{(55)}

where ⟨O⟩0≡⟨0∣O∣0⟩\langle O\rangle_0\equiv\langle0|O|0\rangle. It follows that,

gjk(L)=Re⁡⟨Gj(L)Gk(L)⟩0−⟨Gj(L)⟩0⟨Gk(L)⟩0.(56)g_{jk}^{(L)}=\operatorname{Re}\langle G_j^{(L)}G_k^{(L)}\rangle_0-\langle G_j^{(L)}\rangle_0\langle G_k^{(L)}\rangle_0. \tag*{(56)}

Since the effective generators are Hermitian,

Re⁡⟨Gj(L)Gk(L)⟩0=12⟨{Gj(L),Gk(L)}⟩0.(57)\operatorname{Re}\langle G_j^{(L)}G_k^{(L)}\rangle_0=\frac{1}{2}\left\langle\left\{G_j^{(L)},G_k^{(L)}\right\}\right\rangle_0. \tag*{(57)}

Defining the centered generators ΔGj(L)=Gj(L)−⟨Gj(L)⟩0\Delta G_j^{(L)}=G_j^{(L)}-\langle G_j^{(L)}\rangle_0 the metric takes the symmetrized covariance form

gjk(L)=12⟨{ΔGj(L),ΔGk(L)}⟩0.(58)g_{jk}^{(L)}=\frac{1}{2}\left\langle\left\{\Delta G_j^{(L)},\Delta G_k^{(L)}\right\}\right\rangle_0. \tag*{(58)}

Substituting Eq. (A25) into Eq. (A10) gives

1−FL(t,t+T0)=12∑j,kδtjδtk⟨{ΔGj(L),ΔGk(L)}⟩0+O(∥δt∥3).(59)1-F_L(t,t+T_0)=\frac{1}{2}\sum_{j,k}\delta t_j\delta t_k\left\langle\left\{\Delta G_j^{(L)},\Delta G_k^{(L)}\right\}\right\rangle_0+O(\|\delta t\|^3). \tag*{(59)}

To expose the structure along the actual seasonal displacement, define the directional bare generator at layer ℓ\ell,

Gδ,ℓ≡∑j=1wδtjGj,ℓ.(60)G_{\delta,\ell}\equiv\sum_{j=1}^{w}\delta t_jG_{j,\ell}. \tag*{(60)}

Similarly, define the effective generator along δt\delta t,

Gδ(L)≡∑j=1wδtjGj(L).(61)G_{\delta}^{(L)}\equiv\sum_{j=1}^{w}\delta t_jG_j^{(L)}. \tag*{(61)}

Using Eq. (A18),

Gδ(L)=∑jδtj∑ℓ=1LBℓ−1†Gj,ℓBℓ−1=∑ℓ=1LBℓ−1†(∑jδtjGj,ℓ)Bℓ−1=∑ℓ=1LBℓ−1†Gδ,ℓBℓ−1.(62)\begin{aligned} G_{\delta}^{(L)}&=\sum_j\delta t_j\sum_{\ell=1}^{L}B_{\ell-1}^{\dagger}G_{j,\ell}B_{\ell-1}\\ &=\sum_{\ell=1}^{L}B_{\ell-1}^{\dagger}\left(\sum_j\delta t_jG_{j,\ell}\right)B_{\ell-1}\\ &=\sum_{\ell=1}^{L}B_{\ell-1}^{\dagger}G_{\delta,\ell}B_{\ell-1}. \tag*{(62)} \end{aligned}

Because δt\delta_t is real, Eq. (59) can now be contracted directly:

∑j,kδt,jδt,kgjk(L)=12⟨{ΔGδ(L),ΔGδ(L)}⟩0=⟨(ΔGδ(L))2⟩0=Var⁡0(Gδ(L)).(63)\sum_{j,k}\delta_{t,j}\delta_{t,k}g_{jk}^{(L)}=\frac{1}{2}\left\langle\left\{\Delta\mathcal{G}_{\delta}^{(L)},\Delta\mathcal{G}_{\delta}^{(L)}\right\}\right\rangle_0=\left\langle\left(\Delta\mathcal{G}_{\delta}^{(L)}\right)^2\right\rangle_0=\operatorname{Var}_0\left(\mathcal{G}_{\delta}^{(L)}\right). \tag*{(63)}

The leading deviation from unit fidelity is therefore

1−FL(t,t+T0)=Var⁡0[∑ℓ=1LBℓ−1†Gδ,ℓBℓ−1]+O(∥δt∥3).(64)1-F_L(t,t+T_0)=\operatorname{Var}_0\left[\sum_{\ell=1}^{L}\mathcal{B}_{\ell-1}^{\dagger}\mathcal{G}_{\delta,\ell}\mathcal{B}_{\ell-1}\right]+O\left(\lVert\delta_t\rVert^3\right). \tag*{(64)}

Quantum data encodings

This Appendix gives the explicit parameterizations used for the five quantum data encodings considered in § 1.4 of the main text. All encodings act independently on the w=14w=14 components of the lag window before the variational block at each data-reuploading layer. The value-based encodings depend solely on precipitation magnitude, whereas the three seasonal encodings additionally depend on the position of each observation within the annual cycle. For a lag window

xt=(yt−w+1,…,yt)T,(65)\mathbf{x}_t=(y_{t-w+1},\ldots,y_t)^{T}, \tag*{(65)}

the observation encoded on qubit q∈{0,…,w−1}q\in\{0,\ldots,w-1\} has source index

rtq=t−w+1+q.(66)r_{tq}=t-w+1+q. \tag*{(66)}

For the present data set, w=14w=14. The seasonal index associated with the same observation is

mtq=rtq mod T0,mtq∈{0,…,T0−1},(67)m_{tq}=r_{tq}\bmod T_0,\qquad m_{tq}\in\{0,\ldots,T_0-1\}, \tag*{(67)}

with seasonal period T0=12T_0=12. For the two value-based encodings, precipitation is linearly mapped from the physical interval y∈[0,350]y\in[0,350] mm to

y~=2y350−1,y~∈[−1,1].(68)\widetilde{y}=2\frac{y}{350}-1,\qquad\widetilde{y}\in[-1,1]. \tag*{(68)}

The corresponding rotation angle is ϑ(y)=πy~\vartheta(y)=\pi\widetilde{y}. At reuploading layer ℓ\ell, the complete encoding operation has the product form

Sℓ=⨂q=0w−1Sℓ,q,(69)S_\ell=\bigotimes_{q=0}^{w-1}S_{\ell,q}, \tag*{(69)}

where the single-qubit map Sℓ,qS_{\ell,q} depends on the encoding strategy described below.

RyR_y encoding

The simplest encoding uses the scaled precipitation value as a rotation angle about the yy axis,

Sq(Ry)(yrtq)=Ry[πy~rtq](70)S_q^{(R_y)}(y_{r_{tq}})=R_y\left[\pi\widetilde{y}_{r_{tq}}\right] \tag*{(70)}

The complete encoding layer is, therefore,

Sℓ(Ry)(xt)=⨂q=0w−1Ry[πy~rtq](71)S_\ell^{(R_y)}(\mathbf{x}_t)=\bigotimes_{q=0}^{w-1}R_y\left[\pi\widetilde{y}_{r_{tq}}\right] \tag*{(71)}

This encoding contains no trainable parameters and is identical at every reuploading layer ℓ\ell. Each observation is represented along a single Bloch sphere meridian.

Same-scalar RxRyRzR_xR_yR_z encoding

The second value-based encoding applies the same scalar angle to rotations about all three axes. For a single observation,

Sq(RxRyRz)(yrtq)=Rz[πy~rtq]Ry[πy~rtq]Rx[πy~rtq].(72)S_q^{(R_xR_yR_z)}(y_{rt_q}) = R_z\left[\pi\widetilde{y}_{rt_q}\right]R_y\left[\pi\widetilde{y}_{rt_q}\right]R_x\left[\pi\widetilde{y}_{rt_q}\right]. \tag*{(72)}

Thus,

Sℓ(RxRyRz)(xt)=⨂q=0w−1Sq(RxRyRz)(yrtq).(73)S_\ell^{(R_xR_yR_z)}(\mathbf{x}_t)=\bigotimes_{q=0}^{w-1}S_q^{(R_xR_yR_z)}(y_{rt_q}). \tag*{(73)}

Although this map generates a more intricate Bloch sphere trajectory than the RyR_y encoding, the same value controls all three rotations. Consequently, the information supplied by each observation contains no explicit seasonal coordinate.

Fixed seasonal CDF encoding

Precipitation magnitude is first transformed using the empirical cumulative distribution function constructed exclusively from the raw training observations. The training reference contains the 364 observations with indices r=0,…,363r=0,\ldots,363, and the right-continuous empirical CDF is

CDF⁡(y)=1364∑r=0363I(yr≤y).(74)\operatorname{CDF}(y)=\frac{1}{364}\sum_{r=0}^{363}\mathbb{I}(y_r\le y). \tag*{(74)}

The empirical rank is mapped to the interval [−1,1][-1,1] as

ζ(y)=2CDF⁡(y)−1,(75)\zeta(y)=2\operatorname{CDF}(y)-1, \tag*{(75)}

and subsequently to the polar angle

αCDF⁡(y)=arccos⁡[ζ(y)].(76)\alpha^{\operatorname{CDF}}(y)=\arccos\left[\zeta(y)\right]. \tag*{(76)}

For the observation on qubit qq, αtqCDF⁡=αCDF⁡(yrtq)\alpha_{tq}^{\operatorname{CDF}}=\alpha^{\operatorname{CDF}}(y_{rt_q}). The seasonal phase is assigned to the azimuthal coordinate,

ϕtq=2πmtqT0=2πmtq12.(77)\phi_{tq}=\frac{2\pi m_{tq}}{T_0}=\frac{2\pi m_{tq}}{12}. \tag*{(77)}

The resulting single-qubit encoding is

Sq(CDF⁡)=Rz(ϕtq)Ry(αtqCDF⁡),(78)S_q^{(\operatorname{CDF})}=R_z(\phi_{tq})R_y\left(\alpha_{tq}^{\operatorname{CDF}}\right), \tag*{(78)}

and the complete encoding layer is

Sℓ(CDF⁡)(xt)=⨂q=0w−1Rz(ϕtq)Ry(αtqCDF⁡).(79)S_\ell^{(\operatorname{CDF})}(\mathbf{x}_t)=\bigotimes_{q=0}^{w-1}R_z(\phi_{tq})R_y\left(\alpha_{tq}^{\operatorname{CDF}}\right). \tag*{(79)}

This encoding is fixed and therefore does not depend explicitly on ℓ\ell. The corresponding single-qubit Bloch vector is

rtq=(1−ζtq2cos⁡ϕtq,1−ζtq2sin⁡ϕtq,ζtq)T,(80)\mathbf{r}_{tq}=\left(\sqrt{1-\zeta_{tq}^{2}}\cos\phi_{tq},\sqrt{1-\zeta_{tq}^{2}}\sin\phi_{tq},\zeta_{tq}\right)^{T}, \tag*{(80)}

Thus, empirical precipitation rank controls the polar coordinate, whereas the location within the annual cycle or calendar month controls the azimuthal coordinate.

Learned polar seasonal CDF encoding

The fixed seasonal encoder applies the same encoding at every reuploading layer. To allow the representation to adapt to both seasonal phase and circuit depth, the learned polar encoding introduces two trainable parameters for every month-layer pair (m,ℓ)(m,\ell),

amℓ,cmℓ,m=0,…,11,ℓ=1,…,L.(81)a_{m\ell},\quad c_{m\ell},\qquad m=0,\ldots,11,\qquad\ell=1,\ldots,L. \tag*{(81)}

The first parameter determines a bounded multiplicative strength,

smℓ=1+tanh⁡(amℓ),0<smℓ<2,(82)s_{m\ell}=1+\tanh\left(a_{m\ell}\right),\qquad0<s_{m\ell}<2, \tag*{(82)}

while the second determines a bounded polar displacement,

bmℓ=π2tanh⁡(cmℓ),∣bmℓ∣<π2.(83)b_{m\ell}=\frac{\pi}{2}\tanh\left(c_{m\ell}\right),\qquad\lvert b_{m\ell}\rvert<\frac{\pi}{2}. \tag*{(83)}

For the observation encoded on qubit qq, the resulting month- and layer-dependent polar angle is

αtq,ℓ(LP)=π2+bmtqℓ+smtqℓ[αtqCDF−π2].(84)\alpha_{t_q,\ell}^{(\mathrm{LP})}=\frac{\pi}{2}+b_{m_{t_q}\ell}+s_{m_{t_q}\ell}\left[\alpha_{t_q}^{\mathrm{CDF}}-\frac{\pi}{2}\right]. \tag*{(84)}

The fixed seasonal azimuthal coordinate is retained

ϕtq(LP)=ϕtq=2πmtq12.(85)\phi_{t_q}^{(\mathrm{LP})}=\phi_{t_q}=\frac{2\pi m_{t_q}}{12}. \tag*{(85)}

The single qubit encoding at layer ℓ\ell is, therefore,

Sℓ,q(LP)=Rz(ϕtq(LP))Ry(αtq,ℓ(LP)),(86)S_{\ell,q}^{(\mathrm{LP})}=R_z\left(\phi_{t_q}^{(\mathrm{LP})}\right)R_y\left(\alpha_{t_q,\ell}^{(\mathrm{LP})}\right), \tag*{(86)}

and the complete layer is

Sℓ(LP)=⨂q=0w−1Sℓ,q(LP).(87)S_{\ell}^{(\mathrm{LP})}=\bigotimes_{q=0}^{w-1}S_{\ell,q}^{(\mathrm{LP})}. \tag*{(87)}

All raw encoding parameters are initialized at zero, amℓ=cmℓ=0a_{m\ell}=c_{m\ell}=0. Consequently, smℓ=1s_{m\ell}=1, bmℓ=0b_{m\ell}=0, and Eq. (84) reduces exactly to

αtq,ℓ(LP)=αtqCDF.(88)\alpha_{t_q,\ell}^{(\mathrm{LP})}=\alpha_{t_q}^{\mathrm{CDF}}. \tag*{(88)}

The learned-polar model therefore begins from the fixed seasonal CDF geometry and subsequently learns month- and layer-dependent deformations of its polar coordinate.

Learned polar + azimuthal seasonal CDF encoding

The final encoding extends the learned-polar model by also allowing the seasonal azimuthal coordinate to adapt during training. The polar angle is identical to Eq. (84). In addition, each month-layer pair is assigned a third raw trainable parameter dmℓd_{m\ell} which determines a bounded azimuthal displacement,

Δϕmℓ=π12tanh⁡(dmℓ),∣Δϕmℓ∣<π12.(89)\Delta\phi_{m\ell}=\frac{\pi}{12}\tanh\left(d_{m\ell}\right),\qquad\lvert\Delta\phi_{m\ell}\rvert<\frac{\pi}{12}. \tag*{(89)}

The month- and layer-dependent azimuthal coordinate is

ϕtq,ℓ(LPA)=2πmtq12+Δϕmtqℓ.(90)\phi_{t_q,\ell}^{(\mathrm{LPA})}=\frac{2\pi m_{t_q}}{12}+\Delta\phi_{m_{t_q}\ell}. \tag*{(90)}

The corresponding single-qubit encoding is

Sℓ,q(LPA)=Rz(ϕtq,ℓ(LPA))Ry(αtq,ℓ(LP)),(91)S_{\ell,q}^{(\mathrm{LPA})}=R_z\left(\phi_{t_q,\ell}^{(\mathrm{LPA})}\right)R_y\left(\alpha_{t_q,\ell}^{(\mathrm{LP})}\right), \tag*{(91)}
Sℓ(LPA)=⨂q=0w−1Sℓ,q(LPA).(92)S_{\ell}^{(\mathrm{LPA})} = \bigotimes_{q=0}^{w-1} S_{\ell,q}^{(\mathrm{LPA})}. \tag*{(92)}

Adjacent fixed seasonal sectors are separated by π/6\pi/6. Since the learned displacement satisfies ∣Δϕmℓ∣<π/12|\Delta\phi_{m\ell}| < \pi/12, its maximum magnitude is half the angular separation between adjacent month sectors. The azimuthal coordinate can therefore adapt locally while remaining bounded around the corresponding fixed seasonal direction.

All three raw encoding parameters are initialized at zero, amℓ=cmℓ=dmℓ=0a_{m\ell}=c_{m\ell}=d_{m\ell}=0, which gives smℓ=1s_{m\ell}=1, bmℓ=0b_{m\ell}=0, Δϕmℓ=0\Delta\phi_{m\ell}=0. Thus, the learned polar + azimuthal model also reduces exactly to the fixed seasonal CDF encoding at initialization.

Noisy Simulation Details

The finite-shot noisy simulations are carried out with the IBM FakeAachen through the PennyLane-Qiskit interface. Each expectation value is estimated from S=512S=512 measurement shots.

All circuits are transpiled to the native FakeAachen basis using transpiler optimization level 1 and SABRE routing. This implementation uses the fixed physical qubit layout

[33,32,31,18,11,10,9,8,7,17,27,28,29,38].[33,32,31,18,11,10,9,8,7,17,27,28,29,38].

The closing logical ring CNOT interaction is therefore routed through the available physical connectivity whenever the corresponding physical qubits are not directly coupled. Simulator truncation is enabled so that inactive physical qubits do not contribute to the simulated state space.

Hybrid SPSA-Adam optimization

Let θ∈RPq\theta\in\mathbb{R}^{P_q} denote the complete set of trainable quantum parameters. This vector contains the variational circuit parameters and, for the adaptive encodings, the trainable encoding parameters. We denote the parameters of the classical prediction head by ϕ=(wh,bh)\phi=(w_h,b_h). The quantum parameters are optimized using simultaneous perturbation stochastic approximation (SPSA), while the classical prediction head is optimized using Adam. At SPSA iteration kk, the learning rate and perturbation schedules are

ak=a(A+k)α,ck=ckγ,(93)a_k = \frac{a}{(A+k)^\alpha}, \qquad c_k = \frac{c}{k^\gamma}, \tag*{(93)}

with

a=0.40,c=0.05,A=50,α=0.602,γ=0.101.(94)a=0.40,\qquad c=0.05,\qquad A=50,\qquad\alpha=0.602,\qquad\gamma=0.101. \tag*{(94)}

At each iteration, a perturbation vector Δk∈{−1,+1}Pq\Delta_k \in\{-1,+1\}^{P_q} is drawn with independent Rademacher components. Two perturbed quantum-parameter vectors are then constructed,

θk±=θk±ckΔk.(95)\theta_k^{\pm}=\theta_k\pm c_k\Delta_k. \tag*{(95)}

For a mini-batch Bk\mathcal{B}_k, the noisy quantum circuit is evaluated once at each perturbed parameter vector. The corresponding losses are

Lk±=1∣Bk∣∑t∈Bk[hϕk(Qθk±(xt))−y~t+1]2.(96)\mathcal{L}_k^{\pm}=\frac{1}{|\mathcal{B}_k|}\sum_{t\in\mathcal{B}_k}\left[h_{\phi_k}\left(Q_{\theta_k^{\pm}}(x_t)\right)-\widetilde{y}_{t+1}\right]^2. \tag*{(96)}

where Qθ(x)Q_{\theta}(x) denotes the finite-shot vector of quantum features after measurement, hϕh_{\phi} is the classical prediction head, and y~t+1\widetilde{y}_{t+1} denotes the scaled/transformed target used during training. The SPSA estimator of the quantum gradient is

g^θ,k=(Lk+−Lk−2ck)Δk−1.(97)\widehat{g}_{\theta,k}=\left(\frac{\mathcal{L}_k^{+}-\mathcal{L}_k^{-}}{2c_k}\right)\Delta_k^{-1}. \tag*{(97)}

where the inverse is componentwise. Because each component of Δk\Delta_k is either +1+1 or −1-1, Δk−1=Δk\Delta_k^{-1}=\Delta_k. The quantum parameters are then updated as

θk+1=θk−akg^θ,k.(98)\theta_{k+1} = \theta_k - a_k \widehat{g}_{\theta,k}. \tag*{(98)}

The classical-head loss is defined as the symmetric average of the two perturbed losses,

Lhead,k=12(Lk++Lk−).(99)\mathcal{L}_{\mathrm{head},k} = \frac{1}{2}\left(\mathcal{L}_k^{+}+\mathcal{L}_k^{-}\right). \tag*{(99)}

The quantum features are detached from the PyTorch computational graph before the classical backward pass, so automatic differentiation acts only on ϕ\phi. The classical parameters are updated with Adam using learning rate ηAdam=0.005\eta_{\mathrm{Adam}}=0.005, and the norm of the classical gradient is clipped to 1.0 before each update.

Both the SPSA estimator and the Adam gradient are constructed from the same two quantum evaluations at θk+\theta_k^{+} and θk−\theta_k^{-}. The complete training procedure is summarized below.

Algorithm: Finite-shot hybrid SPSA-Adam training

Require: Training set Dtr={(xt,yt+1)}\mathcal{D}_{\mathrm{tr}}=\{(x_t,y_{t+1})\}

Require: Initial quantum parameters θ1\theta_1

Require: Initial classical head parameters ϕ1\phi_1

Require: SPSA constants a,c,A,α,γa,c,A,\alpha,\gamma

Require: Adam learning rate ηAdam\eta_{\mathrm{Adam}}

  1. for e=1,…,Ee=1,\ldots,E do

  2. Shuffle the training windows

  3. for each mini-batch Bk\mathcal{B}_k do

  4. ak←a/(A+k)αa_k \leftarrow a/(A+k)^\alpha

  5. ck←c/kγc_k \leftarrow c/k^\gamma

  6. Sample Δk∈{−1,+1}Pq\Delta_k \in\{-1,+1\}^{P_q}

  7. θk+←θk+ckΔk\theta_k^{+} \leftarrow\theta_k+c_k\Delta_k

  8. θk−←θk−ckΔk\theta_k^{-} \leftarrow\theta_k-c_k\Delta_k

  9. Evaluate Qθk+(xt)Q_{\theta_k^{+}}(x_t) for all t∈Bkt \in\mathcal{B}_k

  10. Evaluate Qθk−(xt)Q_{\theta_k^{-}}(x_t) for all t∈Bkt \in\mathcal{B}_k

  11. Compute Lk+\mathcal{L}_k^{+} and Lk−\mathcal{L}_k^{-} using Eq. (96)

  12. Lhead,k←(Lk++Lk−)/2\mathcal{L}_{\mathrm{head},k} \leftarrow(\mathcal{L}_k^{+}+\mathcal{L}_k^{-})/2

  13. Detach the quantum features from the PyTorch graph

  14. Compute ∇ϕLhead,k\nabla_\phi\mathcal{L}_{\mathrm{head},k} by automatic differentiation

  15. Clip ∥∇ϕLhead,k∥2\lVert\nabla_\phi\mathcal{L}_{\mathrm{head},k}\rVert_2 to 1.0

  16. Update ϕk+1\phi_{k+1} from ϕk\phi_k using Adam

  17. Compute g^θ,k\widehat{g}_{\theta,k} using Eq. (97)

  18. θk+1←θk−akg^θ,k\theta_{k+1} \leftarrow\theta_k-a_k\widehat{g}_{\theta,k}

  19. Increment kk

  20. end for

  21. if e=1e=1, e=Ee=E, or ee is a multiple of 5 then

  22. Evaluate the center-point validation loss Lval(θk,ϕk)\mathcal{L}_{\mathrm{val}}(\theta_k,\phi_k)

  23. if the validation loss has improved then

  24. Save the current model as the best checkpoint

  25. end if

  26. end if

  27. Save the resumable training checkpoint

  28. end for

  29. Restore the checkpoint with the lowest validation loss

  30. Evaluate the restored model once on the test set

Finite-shot quantum evaluations are distributed across Nworker=32N_{\mathrm{worker}}=32 worker processes. Each worker constructs and retains its own PennyLane QNode and Qiskit Aer backend. For a mini batch of 16 training windows, the positive and negative SPSA perturbations require 2∣Bk∣=322|\mathcal{B}_k|=32 quantum circuit evaluations. These evaluations are submitted concurrently to the 32 CPU workers. The transpiler seed is identical across workers, ensuring that all workers use the same compiled logical circuit. In contrast, the simulator seed assigned to worker jj is sj=sbase+js_j=s_{\mathrm{base}}+j, j=j= 0, …\ldots, Nworker−1N_{\mathrm{worker}} - 1, which provides independent finite-shot and noise realizations across workers. With 350 training windows and mini batch size 16, each epoch contains Nstep=⌊35016⌋=22N_{\mathrm{step}} = \left\lfloor\frac{350}{16} \right\rfloor= 22 SPSA updates. Because each update evaluates two perturbed quantum circuits, training requires 700 quantum evaluations per epoch and 70 000 training evaluations over 100 epochs for each independent run.

CategoryParameterValue or specification
DataTime-series length539 monthly observations
DataLag-window lengthw=14w = 14
DataTraining windows350
DataValidation windows70
DataTest windows105
DataTarget range used for scaling[0, 350] mm
DataScaled target range[−1-1, 1]
DataForecast horizonOne month
Quantum modelNumber of logical qubits14
Quantum modelReuploading depthsL∈{1,…,6}L \in\{1, \ldots, 6\}
Quantum modelAnsatz layers per reupload1
Quantum modelSingle-qubit ansatzRot⁡(θ(1),θ(2),θ(3))\operatorname{Rot}(\theta^{(1)}, \theta^{(2)}, \theta^{(3)}) on every qubit
Quantum modelEntangling patternNearest-neighbor CNOT chain with closing interaction CNOT13,0_{13,0}
Quantum modelMeasurementIndividual Pauli-Z expectation values on all 14 qubits
Quantum modelClassical headLinear(14, 1) followed by tanh
Quantum modelVariational initializationIndependent zero-centered initialization with scale 0.01
Quantum modelLearned-encoding initializationRaw strength, polar-shift, and azimuth-shift parameters initialized to zero
Quantum modelQuantum parameters42L42L for fixed encoding, 66L66L for learned polar, and 78L78L for learned polar plus azimuth
Quantum modelClassical parameters15
Noisy simulationBackend modeQiskit Aer finite-shot noisy simulation
Noisy simulationNoise sourceIBM FakeAachen
Noisy simulationPhysical backend size156 qubits
Noisy simulationPhysical layout[33, 32, 31, 18, 11, 10, 9, 8, 7, 17, 27, 28, 29, 38]
Noisy simulationRouting methodSABRE
Noisy simulationTranspiler optimization level1
Noisy simulationShots512 per expectation-value estimate
Noisy simulationAer methodautomatic
Noisy simulationSimulator truncationEnabled
TrainingIndependent seeds42
TrainingEpochs100
TrainingMini-batch size16
TrainingUpdates per epoch22
TrainingSPSA updates per run2200
TrainingLossMean-squared error
TrainingQuantum optimizerSPSA
TrainingSPSA coefficient aa0.40
TrainingSPSA perturbation coefficient cc0.05
TrainingSPSA stability constant AA50
TrainingSPSA decay exponent α\alpha0.602
TrainingSPSA perturbation exponent γ\gamma0.101
TrainingClassical optimizerAdam
TrainingAdam learning rate0.005
TrainingClassical gradient clippingMaximum Euclidean norm 1.0
TrainingValidation scheduleEpoch 1, every fifth epoch, and the final epoch
TrainingValidation repeats1
TrainingValidation batch size32
TrainingCheckpoint intervalEvery epoch
TrainingModel selectionLowest validation mean-squared error

Table I. Noisy Simulation Details

Forecasting metrics

Forecasting performance was evaluated on the test set using four complementary metrics. Let yty_t and y^t\hat{y}_t denote the observed and predicted precipitation, respectively, for t=0,…,Ntest−1t = 0, \ldots, N_{\mathrm{test}} - 1, where Ntest=105N_{\mathrm{test}} = 105. The test root-mean-square error is defined as

Test RMSE⁡=1Ntest∑t=0Ntest−1(y^t−yt)2.(100)\operatorname{Test\ RMSE} = \sqrt{\frac{1}{N_{\mathrm{test}}}\sum_{t=0}^{N_{\mathrm{test}}-1}(\hat{y}_t-y_t)^2}. \tag*{(100)}

RMSE was computed after restoring model predictions to the original precipitation scale and is therefore reported in millimeters.

To quantify similarity in frequency domain structure, the observed and predicted test sequences were first centered independently by subtracting their respective means,

ytc=yt−yˉ,y^tc=y^t−y^‾.(101)y_t^c = y_t - \bar{y}, \qquad\hat{y}_t^c = \hat{y}_t - \overline{\hat{y}}. \tag*{(101)}

Their discrete Fourier transforms were computed as

y~k=∑t=0Ntest−1ytcexp⁡(−2πiktNtest),(102)\tilde{y}_k = \sum_{t=0}^{N_{\mathrm{test}}-1} y_t^c \exp\left(-\frac{2\pi i k t}{N_{\mathrm{test}}}\right), \tag*{(102)}
y^~k=∑t=0Ntest−1y^tcexp⁡(−2πiktNtest).(103)\tilde{\hat{y}}_k = \sum_{t=0}^{N_{\mathrm{test}}-1} \hat{y}_t^c \exp\left(-\frac{2\pi i k t}{N_{\mathrm{test}}}\right). \tag*{(103)}

The corresponding Fourier power spectra are

Pk=∣y~k∣2,P^k=∣y^~k∣2.(104)P_k = \lvert\tilde{y}_k\rvert^2, \qquad\hat{P}_k = \lvert\tilde{\hat{y}}_k\rvert^2. \tag*{(104)}

The Fourier cosine similarity was evaluated over the non-zero positive frequency bins K+\mathcal{K}_+,

CFFT=∑k∈K+PkP^k∑k∈K+Pk2∑k∈K+P^k2.(105)C_{\mathrm{FFT}} = \frac{\sum_{k\in\mathcal{K}_+} P_k\hat{P}_k}{\sqrt{\sum_{k\in\mathcal{K}_+}P_k^2}\sqrt{\sum_{k\in\mathcal{K}_+}\hat{P}_k^2}}. \tag*{(105)}

Since the Fourier power is nonnegative, 0≤CFFT≤10 \le C_{\mathrm{FFT}} \le1, with CFFT=1C_{\mathrm{FFT}} = 1 indicating identical normalized spectral shapes. This quantity therefore measures whether the forecast reproduces the relative distribution of power across temporal frequencies. Because it is a cosine similarity, it is insensitive to an overall multiplicative rescaling of the predicted power spectrum. This means that high spectral similarity does not by itself imply agreement in forecast amplitude.

Temporal alignment between the observed and predicted sequences was quantified using the Pearson correlation coefficient,

rP=∑t=0Ntest−1(yt−yˉ)(y^t−y^‾)∑t=0Ntest−1(yt−yˉ)2∑t=0Ntest−1(y^t−y^‾)2.(106)r_P = \frac{\sum_{t=0}^{N_{\mathrm{test}}-1}(y_t-\bar{y})(\hat{y}_t-\overline{\hat{y}})}{\sqrt{\sum_{t=0}^{N_{\mathrm{test}}-1}(y_t-\bar{y})^2}\sqrt{\sum_{t=0}^{N_{\mathrm{test}}-1}(\hat{y}_t-\overline{\hat{y}})^2}}. \tag*{(106)}

The coefficient satisfies −1≤rP≤1-1 \le r_P \le1, with rP=1r_P = 1 corresponding to perfect positive linear alignment between the observed and predicted temporal fluctuations. Because the means and overall scales of the two sequences are removed in the correlation calculation, rPr_P primarily measures whether the forecast tracks the temporal variation of the observations rather than their absolute magnitude.

Finally, recovery of the observed forecast amplitude was quantified using the amplitude variance ratio,

Rvar=Var⁡(y^)Var⁡(y).(107)R_{\mathrm{var}} = \frac{\operatorname{Var}(\hat{y})}{\operatorname{Var}(y)}. \tag*{(107)}

where the variances were evaluated over the complete test interval using Var⁡(y)=1Ntest∑t=0Ntest−1(yt−yˉ)2\operatorname{Var}(y) = \frac{1}{N_{\mathrm{test}}}\sum_{t=0}^{N_{\mathrm{test}}-1}(y_t-\bar{y})^2 and analogously for y^\hat{y}. The ideal value is Rvar=1R_{\mathrm{var}} = 1. Values below 1 indicate suppression of the observed variability and therefore forecasts with reduced amplitude, whereas values above unity indicate variability larger than that of the observed precipitation series.

These four metrics probe distinct aspects of forecast quality: Test RMSE measures absolute pointwise error, CFFTC_{\mathrm{FFT}} measures normalized frequency-domain structure, rPr_{\mathrm{P}} measures temporal alignment, and RvarR_{\mathrm{var}} measures recovery of the observed variation in amplitude.

Classical forecasting baselines

To compare the performance of our quantum models, we evaluated three complementary classical baselines: random forest (RF), autoregressive integrated moving average (ARIMA), and long short-term memory (LSTM). The classical models were evaluated over the same chronological training, validation, and test splits used for the quantum models. No test observations were used for hyperparameter or model selection. For the random forest and LSTM models, the precipitation series was represented using the same supervised lag window construction as the quantum models. Each input contained the preceding w=14w = 14 monthly precipitation observations,

xt=(yt−13,…,yt)T,(108)\mathbf{x}_t = (y_{t-13}, \ldots, y_t)^{T}, \tag*{(108)}

with the subsequent observation yt+1y_{t+1} used as the one-month-ahead target. This resulted in 350 training, 70 validation, and 105 test forecasting examples.

Precipitation values were transformed to the interval [−1,1][-1,1] according to

y~=2y350−1,(109)\tilde{y} = 2\frac{y}{350} - 1, \tag*{(109)}

where yy is measured in millimeters. Predictions were transformed back to physical units using y=175(y~+1)y = 175(\tilde{y}+1). Random forest and LSTM are stochastic models and were evaluated over five independent runs using random seeds 42, 43, 44, 45, and 46. The source of stochasticity is model-dependent: for the LSTM, the random state determines the initialization of the trainable parameters and the shuffled mini-batch sequence during training. For the random forest, it determines the stochastic bootstrap of the tree ensemble. The ARIMA model is deterministic for a fixed model specification and was consequently evaluated using a single fit. All classical models were evaluated using the same set of forecasting metrics used for the quantum models.

Random forest

The random-forest baseline was implemented using the RandomForestRegressor in scikit-learn. Each 14-month lag window was supplied to the model as a 14-dimensional feature vector. Individual regression trees used the squared-error criterion and bootstrap sampling. Model hyperparameters were selected using a grid search over the validation set, with the parameter search including

nestimators∈{100,300,500,1000},(110)n_{\mathrm{estimators}} \in\{100,300,500,1000\}, \tag*{(110)}
dmax⁡∈{3,5,8,10,15,None},(111)d_{\max} \in\{3,5,8,10,15,\mathrm{None}\}, \tag*{(111)}
nleafmin⁡∈{1,2,4,8,12,16},(112)n_{\mathrm{leaf}}^{\min} \in\{1,2,4,8,12,16\}, \tag*{(112)}
ffeat∈{0.25,0.50,0.75,1.0},(113)f_{\mathrm{feat}} \in\{0.25,0.50,0.75,1.0\}, \tag*{(113)}

where nestimatorsn_{\mathrm{estimators}} denotes the number of trees, dmax⁡d_{\max} the maximum tree depth, nleafmin⁡n_{\mathrm{leaf}}^{\min} the minimum number of samples in a terminal leaf, and ffeatf_{\mathrm{feat}} the fraction of input features available when determining a split. The complete grid therefore contained 576 hyperparameter configurations. Each configuration was trained independently for the five random states 42, . . . ,46. The selected configuration, according to the mean validation mean-squared error, was

nestimators=500,dmax⁡=5,nleafmin⁡=2,ffeat=1.0.(114)n_{\mathrm{estimators}} = 500,\qquad d_{\max} = 5,\qquad n_{\mathrm{leaf}}^{\min} = 2,\qquad f_{\mathrm{feat}} = 1.0. \tag*{(114)}

LSTM

Each 14 month input window was represented as a sequence of length 14 with one scalar feature per time step, so that the input to the recurrent network had shape 14 × 1. The input dimension and hidden dimension were fixed to nin=1n_{\mathrm{in}} = 1, h=3h = 3, while the number of stacked recurrent layers was varied over L=1,…,6L = 1,\ldots,6. All LSTMs were unidirectional and used no dropout. After processing the complete 14 month sequence, the final hidden state was passed to a linear output layer, followed by a tanh activation to produce the forecast in the scaled target space. The models were trained by minimizing mean squared error using Adam with learning rate η=0.005\eta= 0.005. For every depth LL, five independent runs were performed using random states 42,…,4642,\ldots,46. The best model was chosen using minimum validation loss during training.

DepthLSTMFixed SeasonalLearned PolarLearned Polar + Azimuthal
176578193
217299147171
3268141213249
4364183279327
5460225345405
6556267411483

Table II. Total trainable parameter counts for the LSTM and the three seasonal encoding models across model depth. The LSTM hidden dimension was fixed at h=3h = 3. LSTM depth and quantum reuploading depth are not architecturally equivalent and are aligned here only for a scaling comparison of parameter counts.

The hidden size h=3h = 3 was intentionally chosen to provide a comparable recurrent baseline whose parameter count increases with recurrent depth on a scale comparable to the quantum models.

ARIMA

The ARIMA baseline was applied directly to the chronologically ordered precipitation series rather than to fixed 14 month input windows. Model order was selected using the training and validation periods while preserving their temporal ordering. We considered the grid

p∈{0,…,12},d∈{0,1},q∈{0,…,6}.(115)p \in\{0,\ldots,12\}, \qquad d \in\{0,1\}, \qquad q \in\{0,\ldots,6\}. \tag*{(115)}

Stationarity and invertibility constraints were enforced during fitting. Models were estimated using the state space implementation in statsmodels, with a maximum of 500 optimizer iterations. For each model in the grid search, the model was initially fit using the training portion of the raw precipitation record, comprising observations with indices 0,…,3630,\ldots,363. Validation performance was then evaluated over the subsequent 70 observations using rolling one-step-ahead forecasts. After each forecast, the corresponding observed precipitation value was incorporated into the ARIMA state before forecasting the next month, while the fitted model coefficients remained fixed. Models were ranked according to their validation RMSE.

This procedure selected ARIMA(7, 0, 6). After model-order selection, the selected ARIMA specification was refit using all 434 observations preceding the test interval. Test predictions were then generated sequentially over indices 434, …\ldots, 538 using the same rolling one-step-ahead procedure. Following each test forecast, the corresponding observed precipitation value was incorporated into the model state without re-estimating the ARIMA coefficients.

Consequently, the ARIMA baseline differs from the fixed-window models in that each prediction can use the complete observed history available up to the forecast time, rather than only the preceding 14 months. No seasonal ARIMA terms were included.

ModelParameterValue
Random forestTrees500
Maximum depth5
Minimum samples per leaf2
Maximum feature fraction1.0
Split criterionSquared error
Bootstrap samplingYes
Independent runs5
ARIMASelected order(7, 0, 6)
Candidate orders182
Seasonal termsNone
Forecasting protocolRolling one-step
Final fit observations434
LSTMInput sequence14×114 \times1
Hidden dimension3
Recurrent depthD=1,…,6D = 1, \ldots, 6
OptimizerAdam
Learning rate0.005
Batch size16
Epochs100
Gradient clipping1.0
Independent runs5

Table III. Summary of the classical baseline configurations.

References

  1. [1]M. C. Hansen, P. V. Potapov, R. Moore, M. Hancher, S. a. Turubanova, A. Tyukavina, D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini, C. O. Justice, and J. R. G. Townshend, High-resolution global maps of 21st-century forest cover change., Science 342, 850 (2013).
  2. [2]Z. Wan, P. Wang, and X. Li, Using MODIS Land Surface Temperature and Normalized Difference Vegetation Index products for monitoring drought in the southern Great Plains, USA, Int. J. Remote Sens. 25, 61 (2004).DOI
  3. [3]R. Dubayah, J. B. Blair, S. Goetz, L. Fatoyinbo, M. Hansen, S. Healey, M. Hofton, G. Hurtt, J. Kellner, and S. Luthcke, The global ecosystem dynamics investigation: High-resolution laser ranging of the Earth’s forests and topography, Sci. Remote Sens. 1, 100002 (2020).DOI
  4. [4]A. D. Adebayo and C. Nakalembe, Sub-seasonal forecasting of cropland productivity anomalies using satellite soil moisture in water-limited environments, Remote Sens. Environ. 347, 115645 (2026).DOI
  5. [5]S. Shukla, G. Husak, W. Turner, F. Davenport, C. Funk, L. Harrison, and N. Krell, A slow rainy season onset is a reliable harbinger of drought in most food insecure regions in Sub-Saharan Africa, PLoS ONE 16, 1 (2021).DOI
  6. [6]J. Muñoz-Sabater, E. Dutra, A. Agustí-Panareda, C. Albergel, G. Arduini, G. Balsamo, S. Boussetta, M. Choulga, S. Harrigan, H. Hersbach, B. Martens, D. G. Miralles, M. Piles, N. J. Rodríguez-Fernández, E. Zsoter, C. Buontempo, and J.-N. Thépaut, ERA5-Land: A state-of-the-art global reanalysis dataset for land applications, Earth Syst. Sci. Data 13, 4349 (2021).
  7. [7]M. Schuld and N. Killoran, Quantum machine learning in feature Hilbert spaces, Phys. Rev. Lett. 122, 040504 (2019).DOI
  8. [8]M. Benedetti, E. Lloyd, S. Sack, and M. Fiorentini, Parameterized quantum circuits as machine learning models, Quantum Sci. Technol. 4, 043001 (2019).DOI
  9. [9]M. Schuld and N. Killoran, Is quantum advantage the right goal for quantum machine learning?, PRX Quantum 3, 030101 (2022).DOI
  10. [10]N. Schetakis, D. Aghamalyan, P. Griffin, and M. Boguslavsky, Review of some existing QML frameworks and novel hybrid classical-quantum neural networks realising binary classification for the noisy datasets, Sci. Rep. 12, 11927 (2022).DOI
  11. [11]S. H. Kim, J. Mei, C. Girotto, M. Yamada, and M. Roetteler, Quantum large language model fine-tuning (2025), arXiv:2504.08732 [quant-ph].arxiv.org/abs/2504.08732
  12. [12]J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, Quantum machine learning, Nature 549, 195 (2017).DOI
  13. [13]H.-Y. Huang, M. Broughton, M. Mohseni, R. Babbush, S. Boixo, H. Neven, and J. R. McClean, Power of data in quantum machine learning, Nat. Commun. 12, 2631 (2021).
  14. [14]W. Guan, G. Perdue, A. Pesah, M. Schuld, K. Terashi, S. Vallecorsa, and J.-R. Vlimant, Quantum machine learning in high energy physics, Mach. Learn.: Sci. Technol. 2, 011003 (2021).DOI
  15. [15]M. Larocca, F. Sauvage, F. M. Sbahi, G. Verdon, P. J. Coles, and M. Cerezo, Group-invariant quantum machine learning, PRX Quantum 3, 030341 (2022).DOI
  16. [16]M. C. Caro, H.-Y. Huang, M. Cerezo, K. Sharma, A. Sornborger, L. Cincio, and P. J. Coles, Generalization in quantum machine learning from few training data, Nat. Commun. 13, 4919 (2022).DOI
  17. [17]P. Bermejo, P. Braccia, M. S. Rudolph, Z. Holmes, L. Cincio, and M. Cerezo, Quantum convolutional neural networks are effectively classically simulable, PRX Quantum 7, 020304 (2026).arxiv.org/abs/2408.12739
  18. [18]V. Havlíček, A. D. Córcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, Supervised learning with quantum-enhanced feature spaces, Nature 567, 209–212 (2019).DOI
  19. [19]M. Cerezo, G. Verdon, H.-Y. Huang, L. Cincio, and P. J. Coles, Challenges and opportunities in quantum machine learning, Nat. Comput. Sci. 2, 567–576 (2022).DOI
  20. [20]N. Schetakis, P. Bonfini, N. Alisoltani, K. Blazakis, S. I. Tsintzos, A. Askitopoulos, D. Aghamalyan, P. Fafoutellis, and E. I. Vlahogianni, Quantum neural networks with data re-uploading for urban traffic time series forecasting, Sci. Rep. 15, 19400 (2025).DOI
  21. [21]T. Hubregtsen, D. Wierichs, E. Gil-Fuster, P.-J. H. S. Derks, P. K. Faehrmann, and J. J. Meyer, Training quantum embedding kernels on near-term quantum computers, Phys. Rev. A 106, 042431 (2022).DOI
  22. [22]J. Jäger, P. Elsässer, and E. Torabian, Quantum feature-map learning with reduced resource overhead, Phys. Rev. Res. 8, 023247 (2026).DOI
  23. [23]M. Schuld, R. Sweke, and J. J. Meyer, The effect of data encoding on the expressive power of variational quantum machine learning models, Phys. Rev. A 103, 032430 (2021).DOI
  24. [24]A. Pérez-Salinas, D. López-Núñez, A. García-Sáez, P. Forn-Díaz, and J. I. Latorre, One qubit as a universal approximant, Phys. Rev. A 104, 012405 (2021).
  25. [25]A. Pérez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre, Data re-uploading for a universal quantum classifier, Quantum 4, 226 (2020).
  26. [26]J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Barren plateaus in quantum neural network training landscapes, Nat. Commun. 9, 4812 (2018).
  27. [27]M. Cerezo, A. Sone, T. Volkovich, L. Cincio, and P. J. Coles, Cost function dependent barren plateaus in shallow parametrized quantum circuits, Nat. Commun. 12, 1791 (2021).DOI
  28. [28]S. Sim, P. D. Johnson, and A. Aspuru-Guzik, Expressibility and entangling capability of parameterized quantum circuits for hybrid quantum-classical algorithms, Adv. Quantum Technol. 2, 1900070 (2019).DOI
  29. [29]Y. Du, T. Huang, S. You, M.-H. Hsieh, and D. Tao, Quantum circuit architecture search for variational quantum algorithms, npj Quantum Inf. 8, 62 (2022).DOI
  30. [30]S. Y.-C. Chen, S. Yoo, and Y.-L. L. Fang, Quantum long short-term memory (2020), arXiv:2009.01783 [quant-ph].arxiv.org/abs/2009.01783
  31. [31]S. Z. Khan, N. Muzammil, S. Ghafoor, H. Khan, S. M. H. Zaidi, A. J. Aljohani, and I. Aziz, Quantum long short-term memory (QLSTM) vs. classical LSTM in time series forecasting: a comparative study in solar power forecasting, Front. Phys. 12, 1439180 (2024).DOI
  32. [32]K.-H. Moon, S.-G. Jeong, and W.-J. Hwang, QSegRNN: Quantum segment recurrent neural network for time series forecasting, EPJ Quantum Technol. 12, 32 (2025).
  33. [33]S. Cao, W. Zhang, J. Tilly, A. Agarwal, M. Bakr, G. Campanaro, S. D Facciati, J. Wills, B. Shteynas, V. Chambarama, P. Leek, and I. Rungger, Encoding optimization for quantum machine learning demonstrated on a superconducting transmon qutrit, Quantum Sci. Technol. 9, 045037 (2024).arxiv.org/abs/2309.13036
  34. [34]Z. Yu, Q. Chen, Y. Jiao, Y. Li, X. Lu, X. Wang, and J. Z. Yang, Non-asymptotic approximation error bounds of parameterized quantum circuits, in Advances in Neural Information Processing Systems, Vol. 37, edited by A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (Curran Associates, Inc., 2024) pp. 99089–99127.arxiv.org/abs/2310.07528
  35. [35]M. T. West, J. Heredge, M. Sevior, and M. Usman, Provably trainable rotationally equivariant quantum machine learning, PRX Quantum 5, 030202 (2024).DOI
  36. [36]Z. Li, L. Nagano, and K. Terashi, Enforcing exact permutation and rotational symmetries in the application of quantum neural networks on point cloud datasets, Phys. Rev. Res. 6, 043028 (2024).arxiv.org/abs/2405.11150
  37. [37]J. Heredge, C. Hill, L. Hollenberg, and M. Sevior, Permutation invariant encodings for quantum machine learning with point cloud data, Quantum Mach. Intell. 6, 25 (2024).DOI
  38. [38]S. L. Braunstein and C. M. Caves, Statistical distance and the geometry of quantum states, Phys. Rev. Lett. 72, 3439 (1994).pubmed.ncbi.nlm.nih.gov/10056200
  39. [39]J. Stokes, J. Izaac, K. Killoran, and G. Carleo, Quantum natural gradient, Quantum 4, 269 (2020).DOI
  40. [40]D. Spehner, Bures geodesics and quantum metrology, Quantum 9, 1715 (2025).DOI
  41. [41]J. P. Provost and G. Vallee, Riemannian structure on manifolds of quantum states, Commun. Math. Phys. 76, 289 (1980).DOI
  42. [42]G. Box and G. Jenkins, Time Series Analysis: Forecasting and Control (Holden-Day, 1970).
  43. [43]L. Breiman, Random forests, Mach. Learn. 45, 5 (2001).
  44. [44]S. Hochreiter and J. Schmidhuber, Long short-term memory, Neural Comput. 9, 1735 (1997).
  45. [45]T. Draper and S. Kutin, 〈qpic〉: Quantum circuit diagrams in LaTeX. (2016), accessed October 2026.
  46. [46]N. L. Wach, M. S. Rudolph, F. Jendrejewski, and S. Schmitt, Data re-uploading with a single qudit, X. Wang, H.-X. Tao, and R.-B. Wu, Predictive performance of deep quantum data re-uploading models (2025), arXiv:2505.20337 [quant-ph].
  47. [48]A. Barthe and A. Pérez-Salinas, Gradients and frequency profiles of quantum re-uploading models, Quantum 8, 1523 (2024).DOI
  48. [49]C. Nakalembe, Characterizing agricultural drought in the Karamoja subregion of Uganda with meteorological and satellite-based indices, Nat. Hazards 91, 837 (2018).DOI
  49. [50]V. Bergholm, J. Izaac, M. S. Schuld, C. Gogolin, S. Ahmed, V. Ajith, M. S. Alam, G. Alonso-Linaje, B. AkashNarayanan, A. Asadi, J. M. Arrazola, U. Azad, S. Ban n ing, C. Blank, T. R. Bromley, B. A. Cordier, J. Ceroni, A. Delgado, O. D. Matteo, A. Dusko, T. Garg, D. Guala, A. Hayes, R. Hill, A. Ijaz, T. Isacsson, D. Ittah, S. Jahangiri, P. Jain, E. Jiang, A. Khandelwal, K. Kottmann, R. A. Lang, C. Lee, T. Loke, A. Lowe, K. McKiernan, J. J. Meyer, J. A. Montañez-Barrera, R. Moyard, Z. Niu, L. J. O’Riordan, S. Oud, A. Panigrahi, C.-Y. Park, D. Polatajko, N. Quesada, C. Roberts, N. Sá, I. Schoch, B. Shi, S. Shu, S. Sim, A. Singh, I. Strandberg, J. Soni, A. Száva, S. Thabet, R. A. Vargas-Hernández, T. Vincent, N. Vitucci, M. Weber, D. Wierichs, R. Wiersema, M. Willmann, V. Wong, S. Zhang, and N. Killoran, PennyLane: Automatic differentiation of hybrid quantum-classical computations (2022), arXiv:1811.04968 [quant-ph].
  50. [51]A. Paszke, S. Gross, F. Massa, A. Ler er, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, PyTorch: An imperative style, high-performance deep learning library, in Advances in Neural Information Processing Systems, Vol. 32, edited by H. Wallach, H. Larochelle, A. Beygelzimer, F. d’Alché-Buc, E. Fox, and R. Garnett (Curran Associates, Inc., 2019).arxiv.org/abs/1912.01703
  51. [52]K. Życzkowski and H.-J. Sommers, Average fidelity between random quantum states, Phys. Rev. A 71, 032313 (2005).DOI
  52. [53]S. Sim, P. D. Johnson, and A. Aspuru-Guzik, Expressibility and entangling capability of parameterized quantum circuits for hybrid quantum-classical algorithms, Adv. Quantum Technol. 2, 12 (2019).DOI
  53. [54]A. Javadi-Abhari, M. Treinish, K. Krsulich, C. J. Wood, J. Lishman, J. Gacon, S. Martiel, P. D. Nation, L. S. Bishop, A. W. Cross, B. R. Johnson, and J. M. Gambetta, Quantum computing with Qiskit (2024), arXiv:2405.08810 [quant-ph].
  54. [55]J. Spall, Multivariate stochastic approximation using a simultaneous perturbation gradient approximation, IEEE Trans. Autom. Control 37, 332 (1992).DOI
  55. [56]D. P. Kingma and J. Ba, Adam: A method for stochastic optimization, in International Conference on Learning Representations (2015).arxiv.org/abs/1412.6980
  56. [57]J. Spall, Implementation of the simultaneous perturbation algorithm for stochastic optimization, IEEE Trans. Aerosp. Electron. Syst. 34, 817 (1998).DOI
  57. [58]K. Temme, S. Bravyi, and J. M. Gambetta, Error mitigation for short-depth quantum circuits, Phys. Rev. Lett. 119, 180509 (2017).DOI
  58. [59]Y. Li and S. C. Benjamin, Efficient variational quantum simulator incorporating active error minimization, Phys. Rev X 7, 021050 (2017).DOI
  59. [60]Z. Cai, R. Babbu sh, S. C. Benjamin, S. Endo, W. J. Huggins, Y. Li, J. R. McClean, and T. E. O’Brien, Quantum error mitigation, Rev. Mod. Phys. 95, 045005 (2023).DOI
  60. [61]C. Sharma and S. K. Sood, Variational quantum reservoir networks for multivariate time-series prediction, Nonlinear Dyn. 114, 818 (2026).DOI
  61. [62]S. Ranilla-Cortina, D. A. Aranda, J. Ballesteros, J. Bonilla, N. Monr io, and E. F. Combarro, Quantum-enhanced architectures for multivariate time-series forecasting, J. Supercomput. 82, 155 (2026).
  62. [63]M. R. Laskar and R. Goel, Shallow entangled circuits for quantum time series prediction on IBM devices, Sci. Rep. 15, 43727 (2025).DOI
  63. [64]I. Panadero, Y. Ban, H. Espinós, R. Puebla, J. Casanova, and E. Torrontegui, Regressions on quantum neural networks at maximal expressivity, Sci. Rep. 14, 31669 (2024).DOI
  64. [65]A. Kutvonen, K. Fujii, and T. Sagawa, Optimizing a quantum reservoir computer for time series prediction, Sci. Rep. 10, 14687 (2020).DOI
  65. [66]K. Kobayashi, K. Fujii, and N. Yamamoto, Feedback-driven quantum reservoir computing for time-series analysis, PRX Quantum 5, 040325 (2024).DOI
  66. [67]Y. Hou, J. Hua, Z. Wu, W. Xia, Y. Chen, X. Li, Z. Li, X. Peng, and J. Du, High-accuracy temporal prediction via experimental quantum reservoir computing in correlated spins, Phys. Rev. Lett. 136, 120602 (2026).DOI
  67. [68]O. Ahmed, F. Tennie, and L. Magri, Optimal training of finitely sampled quantum reservoir computers for forecasting of chaotic dynamics, Quantum Mach. Intell. 7, 31 (2025).arxiv.org/abs/2409.01394
  68. [69]A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, J. M. Chow, and J. M. Gambetta, Hardware-efficient variational quantum eigensolver for small molecules and quantum magnets, Nature 549, 242 (2017).DOI
  69. [70]L. Candelori, S. Majumder, A. Mezzacapo, J. R. Moreno, K. Musaeilan, S. Nagarajan, S. Pinnamaneni, K. Sharma, and D. Villani, Shallow-circuit supervised learning on a quantum processor (2026), arXiv:2601.03235 [quant-ph].arxiv.org/abs/2601.03235
  70. [71]N. Mathur, P. K. Barkoutsos, M. Yamada, M. Roetteler, and I. Kerenidis, Scalable on-hardware training of quantum neural networks and application to clinical data imputation (2026), arXiv:2606.03517 [quant-ph].arxiv.org/abs/2606.03517
  71. [72]M. Kornjača, H.-Y. Hu, C. Zhao, J. Wurtz, P. Weinberg, M. Hamdan, A. Zhdanov, S. H. Cantu, H. Zhou, R. A. Bravo, K. Bagnall, J. I. Basham, J. Campo, A. Choukri, R. DeAngelo, P. Frederick, D. Haines, J. Hammett, N. Hsu, M.-G. Hu, F. Huber, P. N. Jepsen, N. Jia, T. Karol yshyn, M. Kwon, J. Long, J. Lopatin, A. Lukin, T. Macri, O. Marković, L. A. Martínez-Martínez, X. Meng, E. Ostroumov, D. Paquette, J. Robinson, P. S. Rodriguez, A. Singh, N. Sinha, H. Thoreen, N. Wan, D. Waxman-Lenz, T. Wong, K.-H. Wu, P. L. S. Lopes, Y. Boger, N. Gemelke, T. Kitagawa, A. Keesling, X. Gao, A. Bylinskii, S. F. Yelin, F. Liu, and S.-T. Wang, Large-scale quantum reservoir learning with an analog quantum computer (2024), arXiv:2407.02553 [quant-ph].arxiv.org/abs/2407.02553
  72. [73]O. Knitter, J. Mei, S. H. Kim, C. Chen, M. Yamada, and M. Roetteler, Neural fourier surrogates for data reuploading quantum neural networks (2026), arXiv:2610.00841 [quant-ph].arxiv.org/abs/2610.00841
  73. [74]D. Tripathy, R. Robertson, J. Stevens, C. L. Nakalembe, and S. Deffner, GitHub Repo: Quantum Learning for Geospatial Time Series (2026), accessed October 2026.

Paper details

vlnc:W_96ygpq8r
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

OpenAI’s ChatGPT models

View statement
Looking for related papers…
Contents