WorksPhysics and Astronomy
Classical training of conditional instantaneous quantum polynomial-time Born machines
Abstract
Instantaneous quantum polynomial-time (IQP) Born machines are attractive quantum generative models, as certain IQP circuit families are conjectured to be hard to sample, yet they can be trained classically. Previously proposed conditional IQP Born machines make the parameters of the diagonal layer condition dependent. We introduce conditional IQP Born machines, in which the condition is encoded in the input state and a single IQP unitary is shared across conditions. Efficient sampling and amplitude-ratio evaluation for the reference state entering the diagonal layer suffice to obtain unbiased classical estimates of a conditional maximum mean discrepancy objective and its gradients. We show that input conditioning can represent conditional families that no diagonally conditioned IQP model with the same number of qubits can reproduce. For the sixteen-qubit Curie-Weiss spin model, a single conditional model captures the magnetization transition at held-out temperatures. Its mean total variation distance is six times lower than that of diagonal conditioning under the same protocol and comparable to that of classical generative models with up to seven times as many parameters. We also train conditional models classically with up to 2048 qubits, which reproduce the mean squared magnetization of the target away from the crossover. These results extend classically trained IQP Born machines to conditional generation.
Introduction
Conditional generative models approximate a family of target distributions with a single model. They capture how a distribution depends on one or more conditioning variables, such as the temperature of a physical system or a class label in image generation. Learning across conditions aims to capture structure shared by the family, which in turn supports generation at conditions not encountered during training. In classical machine learning, examples include conditional variational autoencoders [1, 2, 3], generative adversarial networks [4, 5, 6], and diffusion models [7, 8]. Among quantum generative models, a widely studied architecture is the quantum circuit Born machine (QCBM), which represents a probability distribution through the measurement statistics of a parameterized circuit [9, 10]. QCBMs have been extended to conditional generation through input-state preparation, conditional registers, and postselection [11, 12, 13, 14, 15].
Training QCBMs at scale requires quantum hardware, as the cost of classical simulation generally grows exponentially. However, training on hardware is also costly. Gradients are typically estimated with the parameter-shift rule, which requires a number of circuit evaluations that grows linearly with the number of parameters [16]. Circuits with commuting generators reduce this overhead to backpropagation-like scaling [17], but training on noisy intermediate-scale quantum hardware [18] still faces noise-induced optimization issues, such as barren plateaus [19].
Instantaneous quantum polynomial-time (IQP) circuits [20] offer a way around some of these limitations. They are attractive architectures for QCBMs [21], because sampling from certain IQP circuit families is conjectured to be classically hard under standard complexity-theoretic assumptions [22, 23, 24, 25], yet their Pauli- expectation values can be estimated classically to additive accuracy [26]. Objectives expressed through these expectation values, such as the maximum mean discrepancy (MMD) with a Gaussian kernel [27, 28], can therefore be estimated and optimized classically for circuits far beyond the reach of exact simulation [28, 29]. Recio-Armengol et al. combined these ingredients into a framework in which IQP Born machines are trained classically and require quantum hardware only for sampling [29], an approach since extended to non-binary discrete data [30].
In this framework, the IQP unitary acts on a fixed input state, typically , and represents a single distribution. Existing conditional IQP Born machines make part of the diagonal layer condition dependent. Liu et al. train a circuit core at a reference condition, freeze it, and then adapt a block of low-order diagonal parameters sequentially at each further condition, interpolating the adapted blocks to unseen conditions [31, 32]. ConQuER instead fits a separate controller layer of diagonal gates, added to a pretrained generator, to data filtered by each target condition [33]. In both approaches, no parameters are trained jointly across conditions, and the dependence on the condition is not learned from the family as a whole. Outside the IQP setting, Kiss et al. conditioned a QCBM through its input state and trained a single circuit across conditions with an MMD loss and gradients estimated with the parameter-shift rule from circuit samples [12]. Conditioning the input state while training a shared IQP unitary jointly across conditions and estimating the objective and its gradients classically has not been studied. In this work, we encode the condition in the input state of an IQP Born machine and share a single IQP unitary across conditions, so that the dependence on the condition is learned from the whole family (Figure 1). This extends the framework of Recio-Armengol et al. to condition-dependent inputs while retaining classical training [29]. The underlying estimators require only sampling and amplitude ratios of the reference state, the input after the first Hadamard layer of the IQP unitary [26, 34], so the encoder may include non-diagonal operations while the trainable block remains an IQP unitary (§ 2 A). Matrix product and permutation-symmetric states, for example, support both operations efficiently. Input conditioning also enlarges the conditional families that the model can represent. An IQP unitary cannot
Figure 1. Proposed framework for classical training and quantum inference of conditional instantaneous quantum polynomial-time (IQP) Born machines. An encoder maps each condition to an input state . (a) Classical training. Reference samples and amplitude ratios from are combined with phase differences and derivatives from the diagonal layer to estimate conditional Pauli- expectation values and their gradients. These estimates update the parameters by minimizing the squared maximum mean discrepancy () averaged over the training conditions. (b) Quantum deployment. After training, the parameters define the quantum circuit. Applying the trained IQP unitary, shared across all conditions, to and measuring in the computational basis yields samples from the learned conditional distribution .
change the output’s distribution in the basis, which is therefore fixed by the input state. As a result, input conditioning can represent families that diagonal conditioning on a fixed input cannot reproduce with the same number of qubits (§ 2 C, Proposition 1), and it can impose output symmetries, such as global spin-flip invariance, on every member of the family (§ 2 D). We test the framework at sixteen qubits on two target families, with diagonal conditioning as the main baseline (Secs. IV–§ 3). On a family of Hamming-weight distributions with tunable mean and width, a model trained with each coordinate varied separately generalizes to their combinations, with about one third of the total variation distance (TVD) of diagonal conditioning. On the Curie-Weiss spin model, a single model captures the transition from a unimodal to a bimodal magnetization profile at held-out temperatures, and the encoder matters: an encoder that conditions only through input-state phases already lowers the TVD 2.7 times relative to diagonal conditioning, and adding a one-axis-twisting step lowers it more than twofold further, to one sixth of diagonal conditioning. This shared model is, on average, as accurate as circuits trained at each evaluation temperature and comparable to classical generative models with up to seven times as many parameters.
Because Curie–Weiss observables remain classically tractable at any size, we also train conditional models classically with up to 2048 qubits and check that they reproduce the mean squared magnetization away from the crossover. Finally, we sample models with up to 30 qubits on real quantum hardware (Appendix B § 5.0.4).
Conditional IQP Born machines
Consider a target family of distributions over -bit strings, which we model using qubits. A conditional IQP Born machine consists of two components: an encoder that maps each condition to an input state , and a parameterized IQP unitary with parameters . The model represents the distribution family as
Training aims to find a single unitary that maps each input state to a distribution that closely approximates the target for all .
The IQP unitary has the form
where the diagonal layer
is parameterized by a phase for each generator in a set of polynomially many Pauli- strings of length . Because the Hadamard layers exchange the and bases, is diagonal in the basis. The diagonal layer acts on the reference state , whose amplitudes we write . We include the initial Hadamard layer in so that the input recovers the standard IQP Born machine.
Sampling-and-ratio interface
Classical training of a standard IQP Born machine relies on estimating the Pauli- moments of its output distribution,
for bit masks [26, 29]. Van den Nest’s sparse-observable method estimates such moments for computationally tractable reference states [26], and Havlíček showed that two operations suffice: sampling the computational-basis distribution and evaluating ratios of amplitudes [34]. For the standard reference state , the samples are uniform bitstrings and all ratios equal one, which gives the cosine estimator of IQP moments [26, 29, 35].
Because the estimator accesses the reference state only through these two operations, any condition-dependent reference state that supports them efficiently admits the same estimation; this is what lets us move the condition from the diagonal layer into the input state. Examples of reference states that support both operations efficiently include product states, matrix product states of polynomial bond dimension [26], and permutation-symmetric states, whose amplitudes depend only on the Hamming weight. For each condition , we require exact samples , and, for a sampled and a mask , the amplitude ratio
where denotes bitwise addition modulo two. We call these two operations the sampling-and-ratio interface of the reference state. Replacing by in Eq. (4) defines the conditional moments . With the sampling-and-ratio interface, these moments, the MMD objective of Sec. III built from them, and their gradients can be estimated to additive accuracy in time polynomial in and (Appendix A 2).
Classical estimation of Pauli-Z moments
The diagonal layer multiplies each computational-basis state by a phase, , where
and each generator is identified with its support , so that is its eigenvalue on . Each moment is an expectation over the reference distribution,
where and is the ratio of Eq. (5). This holds because the Hadamard layers turn into the bit flip , which pairs each string with . Since strings with are never sampled, the ratio is needed only where it is defined. Equation (7) is Havlíček’s amplitude-ratio estimator [34 Thm. 5], built on Van den Nest’s sparse-observable method [26 Thm. 3], applied to the one-sparse operator , whose expectation in is . The ratios do not depend on the normalization of .
Since the moment is real and only the phase factor depends on ,
are unbiased single-sample estimators of and . Here, the derivative of with respect to , , equals when the supports of and overlap in an odd number of positions and zero otherwise. Averaging samples gives mean-square errors of at most and , without any bound on individual ratios [34 Sec. 3.4], because and
Encoders beyond diagonal conditioning
Because is diagonal in the basis, it multiplies each reference amplitude by a phase and leaves its magnitude unchanged. For every parameter vector, the output’s -basis distribution therefore equals the reference distribution,
The encoder thus influences the output in two ways: through the phases of the reference amplitudes, which a diagonal layer can also modify, and through their magnitudes, which no diagonal layer can alter. Diagonal conditioning on a fixed input changes the output only through the phases and keeps the same -basis distribution at every condition.
Although the model samples in the computational basis, this conservation constrains its output, because the computational-basis and -basis distributions of a pure state are not independent. For example, has a deterministic computational-basis outcome and a uniform -basis distribution, and the reverse. We call an -basis distribution compatible with an output distribution if some pure state has both, and say that a family has a common compatible -basis distribution if one distribution is compatible with all its members.
Proposition 1. Consider IQP Born machines on qubits without ancillas. (i) Diagonal conditioning on a fixed pure input can reproduce a family exactly only if its reference distribution , the -basis distribution of its input, is compatible with every member. (ii) For every and every fixed IQP unitary, there are families with no common compatible -basis distribution, which by (i) no fixed input reproduces, but which condition-dependent inputs with an efficient sampling-and-ratio interface reproduce exactly.
Appendix A.1 proves both parts. Part (i) holds for any generator set, and its condition is also sufficient when the diagonal layer contains all Pauli- strings, a layer with exponentially many generators that is no longer an IQP circuit. Part (ii) uses a two-member family consisting of a deterministic outcome and a uniform distribution over three strings; for , the latter is the distribution with which Kurkin et al. showed that two-qubit IQP Born machines are not universal [36 Lemma 11]. Since the IQP unitary cannot change the -basis distribution, the separation comes from the input, independent of the diagonal layer.
We call an encoder diagonal-equivalent if its reference magnitudes do not depend on the condition. Then for a fixed state and a condition-dependent diagonal unitary , and the model is diagonal conditioning on a fixed input with the diagonal layer . Hence, only encoders that are not diagonal-equivalent can reach families beyond diagonal conditioning. Moreover, if the magnitudes are uniform at every condition, then , and the model is a standard IQP Born machine on with diagonal layer . When is built from the ansatz generators, it merges with the learned diagonal layer,
and the encoder reduces to a condition-dependent shift of the parameters.
Whether an encoder is diagonal-equivalent cannot be read off the gates that prepare the input, because the classification concerns the reference state, which includes the initial Hadamard layer of . The input , prepared by a non-diagonal gate, has -dependent magnitudes, but the Hadamard layer turns them into phases. Then, its reference state has uniform magnitudes, and Eq. (11) absorbs it as . Conversely, the input , prepared by a diagonal gate, is not diagonal-equivalent, because its reference state has magnitudes that depend on . This classification applies to any input; classical training additionally requires the sampling-and-ratio interface of Sec. II A, which every encoder in this work provides.
For the input , part (i) recovers the known non-universality of IQP Born machines [29, 36]. Kurkin et al. restored universality by tracing out hidden qubits, at the cost of exponentially many phases [36]. Hidden qubits also lift the obstruction of part (i): with a single one, diagonal conditioning reproduces every two-qubit distribution [36 Lemma 12], and hence both members of the family of part (ii), which involve only two qubits. Input conditioning reaches such families without additional qubits, which suggests a trade-off between the number of hidden qubits and the complexity of the circuit that prepares the input state.
Symmetries imposed by the encoder
Many target families have symmetries known in advance, and models that respect them by construction need not learn them [37]. Encoding a symmetry in the input state imposes it on every member of the family while leaving the IQP unitary unrestricted.
For a bit mask , the bit-flip operator commutes with every IQP unitary, since commutes with the diagonal layer. If the input is an eigenstate of at every condition, so is the output, whose amplitudes at and differ at most by a sign, and hence for every parameter vector and condition. Recio-Armengol et al. introduced this construction for unconditional models, using GHZ inputs to impose global bit-flip symmetry [29 Sec. 6].
Equivalently, the reference state is an eigenstate of , that is, all strings in its support have the same parity , and the condition may change the amplitudes within that sector freely. The estimator is unaffected, since samples never leave the sector. In particular, even-parity reference states give , with the bitwise complement of ; the parity encoders of § 3 B use this to impose global spin-flip symmetry on the Curie–Weiss models at every temperature.
Conditional MMD objective
We train conditional IQP Born machines by minimizing the squared maximum mean discrepancy (MMD) across training conditions. With a Gaussian kernel, this objective depends on the model only through its Pauli-Z moments, which can be estimated classically as described in § 2 B. For distributions and on and a positive-definite kernel , the population squared MMD is [27]
where the samples in each expectation are drawn independently.
For binary strings, equals the Hamming distance , and the Gaussian kernel with bandwidth becomes . It depends only on the number of differing bits and treats all bits equally.
Let denote the set of training conditions and the set of kernel bandwidths. We define the condition-balanced objective as
Here, denotes the MMD computed with kernel . Each discrepancy compares the target and the model at the same condition, and the objective contains no term that couples different conditions; learning across conditions enters only through the shared parameters .
To express the objective through Pauli-Z moments, we apply the expansion of the Gaussian kernel in binary characters used in earlier work on QCBM and classical IQP training [28, 29] separately at each condition. Let be the law of a random mask with independent coordinates
The Gaussian kernel is then the average of the characters of Eq. (6) over this law,
since each bit in which and differ contributes a factor . With the Pauli-Z moments of § 2 A, , and the target moments defined in the same way, the squared MMD becomes
The squared MMD is thus a weighted squared distance between the Walsh–Fourier spectra of the two distributions, with weights . Since , every mask has positive weight, and the MMD vanishes if and only if . The weights of high-order masks can, however, be exponentially small, so we also evaluate the TVD and physical observables.
The bandwidth sets which correlation orders the objective emphasizes. The Hamming weight , which is the order of , follows , and we call its mean the average bodyness, adapting the term of Rudolph et al. [28]. Larger bandwidths emphasize lower orders, and as ; averaging over several bandwidths in Eq. (13) combines these regimes. Emphasizing low orders also mitigates barren plateaus [38, 39, 40], at the cost of sensitivity to higher-order differences [28]; dedicated mitigation strategies for IQP Born machines [36, 41, 42] are outside our scope.
Estimating the objective
Each optimizer update draws a training condition and a bandwidth uniformly, a minibatch of target samples with replacement from a training pool at , independent reference samples from , and masks, with . All masks are evaluated on the same two batches, which correlates the per-mask estimates but leaves each unbiased for the MMD between the model and the empirical distribution of the pool. For each mask , the estimator of Recio-Armengol et al. [29 Prop. 3], extended to condition-dependent reference states, is
where and are single-sample estimates of the target and model moments, and bars denote sample means. The estimator is unbiased for , where is the moment of the training pool, because its pair sums exclude self-products [27] and distinct samples are independent; it can take negative values. Its gradient,
is also unbiased, with the single-sample gradient estimates of Eq. (8) and their mean. Each mask and gradient component costs operations once the ratios and phase differences are available, since the pair sums reduce to sums and sums of squares. Since the identity mask contributes nothing, we draw the masks from and multiply their average by ,
which is unbiased for Eq. (16) with replaced by the empirical distribution of the pool. We quote throughout, although the sampled masks have mean order . The variances of these estimators are bounded independently of the number of qubits (Appendix A 2).
Target distribution families
Both target families assign equal probability to bitstrings with the same Hamming weight and are invariant under permutations of the bits. This symmetry matches the Hamming-distance kernel of the MMD objective. The two families play complementary roles. The first has two independent conditioning coordinates and tests whether the model generalizes to their combinations. In the second, the Curie–Weiss model, the magnetization profile changes from unimodal to bimodal as the temperature decreases, which requires a single model to capture a qualitative change in the shape of the distribution across conditions. Its observables remain classically tractable at any size, which lets us validate models trained with up to 2048 qubits. Figure 2 shows representative members of both families.
Figure 2. Target families. (a) Hamming-weight profiles of the two-dimensional family: the reference , a shift of the mean, a change of the width, and both combined. (b) Magnetization profiles of the sixteen-spin Curie–Weiss model across the crossover.
Hamming-weight family with tunable mean and width
The first family is indexed by a two-dimensional condition , where controls the location and the width of a Hamming-weight distribution. We deform the binomial Hamming-weight profile by a quadratic exponential factor. For , the target Hamming-weight profile is
normalized over , with , and the binomial width . The second term cancels the Gaussian shape of the binomial coefficient to leading order, so the profile is close to a Gaussian with center and width ; and thus control its location and width without setting its moments exactly. We then assign equal probability to bitstrings with the same Hamming weight ,
Along , the two quadratic terms in the exponent cancel, leaving a term linear in , and the targets reduce to product distributions with binomial Hamming-weight profiles. At the origin, the target is uniform over bitstrings, and its Hamming-weight profile is binomial.
Curie–Weiss model
The second family is the zero-field Curie–Weiss spin model [43], conditioned on the inverse temperature . For spins, we represent each configuration by a bitstring , with . The magnetization per spin is then . With unit interaction strength, the energy is , and the target is the Gibbs distribution at inverse temperature ,
where is the normalization constant. In addition to permutation invariance, the target exhibits global spin-flip symmetry. The crossover from the disordered to the ordered regime occurs near and becomes sharper as increases.
We characterize this crossover using two observables. The mean absolute magnetization measures the magnetic order; we use the absolute value because spin-flip symmetry makes the signed mean vanish. The per-spin energy-fluctuation statistic
equals the heat capacity per spin for the Gibbs target, and its peak indicates the crossover. Since the target depends on only through , its observables follow exactly from sums over Hamming weights at any size, which allows trained models to be validated far beyond exact simulation (§ 3 D).
Numerical experiments
We test the framework in four experiments. We train all IQP models by minimizing the sampled squared MMD of Sec. III. Within each experiment, all models, including the baselines, are selected by the same hyperparameter-optimization protocol on development conditions and evaluated on separate conditions outside the training set. The per- circuits of § 3 B are the exception, since they serve as a reference trained directly at the evaluation (Appendix B). At sixteen qubits, our accuracy metric is the exact total variation distance,
which we compute from exact circuit probabilities, average over evaluation conditions within a fit, and report as the mean and standard deviation across independent fits, which differ in their random seed. For Hamming-weight profiles, we also use the coarser profile TVD, , the TVD between the Hamming-weight distributions and of the two distributions, which never exceeds the full TVD.
Generalization to unseen condition combinations
The first experiment tests whether a conditional model trained with one condition coordinate varied at a time generalizes to combinations of both coordinates, on the Hamming-weight family of Sec. IV A with sixteen qubits. We encode the two coordinates in the phases of the uniform superposition (Figure 3(a)),
Figure 3. Encoder circuits, drawn for four qubits with the angles of the sixteen-qubit models, where and connected dots denote with the angle shown above each group. (a) Hamming-weight encoder of Eq. (25). (b) Twisted parity encoder of Eq. (27); without the dashed block, which contains the twist and the rotation, it reduces to the parity encoder of Eq. (26) with angle .
where is the collective spin. Since , the two rotations give phases linear and quadratic in the Hamming weight , matched to the location and the width of the target. Both are diagonal and leave the input probabilities uniform. The reference magnitudes nonetheless depend on , so the encoder is not diagonal-equivalent (§ 2 C), and the IQP unitary must convert the encoded phases into the target probabilities. At the origin, the input is , which the IQP unitary preserves up to a global phase, and the model matches the uniform target by construction. The amplitudes depend only on the Hamming weight, which makes sampling and amplitude ratios efficient (Appendix B2).
We use an IQP unitary with 40 trainable phases, one on each qubit and one on each of the 24 nearest-neighbor pairs of a square lattice. We train at nine conditions that vary one coordinate at a time with the other fixed at zero, which span the diamond . We select models on eight further development conditions on the axes and report their accuracy at sixteen combinations of both coordinates, inside and outside this convex hull, which enter neither training nor selection.
The diagonal-conditioning baseline applies the same lattice circuit to the fixed input and makes its 40 phases an affine function of the condition, , with 120 parameters in total. With reference state at every condition, it is diagonal-equivalent (§ 2 C). We include no term in : every training condition has or , so such a term would vanish throughout training and could not be learned.
Figure 4(a) shows that the conditional model reproduces the target family across the evaluated condition square, although it was trained only on the two axes. At the sixteen mixed conditions, its TVD is (Table IV in Appendix B 2), with at the four conditions inside the training hull and at the twelve outside it. Across the condition square, its error correlates more strongly with each target’s TVD from the uniform distribution (0.91) than with its distance to the nearest training condition (0.72) (Appendix B 2).
Figure 4. Joint conditional generation of the Hamming-weight family, (a) with input conditioning and (b) with diagonal conditioning of the same lattice circuit acting on . Colors show the mean full-distribution TVD across ten seeds on a grid of conditions, on a shared logarithmic scale. Diamonds mark the nine training conditions, which lie on the two axes, and the dashed line marks their convex hull. At the origin, the target and the input-conditioned model are uniform by construction.
| Model | TVD |
| Encoder only | 0.387 |
| Input conditioning (40 parameters) | 0.0216 0.0017 |
| Diagonal conditioning (120 parameters) | 0.069 0.002 |
| Diagonal, quadratic map (200 parameters) | 0.078 0.002 |
| Diagonal, all-to-all layer (408 parameters) | 0.103 0.004 |
| Diagonal, all-to-all layer, quadratic map (680 parameters) | 0.109 0.005 |
Table IV. Two-coordinate errors after axis training, evaluated on the sixteen mixed conditions. Input conditioning is the conditional IQP Born machine; diagonal conditioning makes the phases of the same lattice circuit, acting on , an affine function of the condition, and its variants use a quadratic condition map, an all-to-all diagonal layer, or both. Values are means and standard deviations of the per-seed mean full-distribution TVD across ten seeds; the deterministic encoder-only row has no spread. Lower is better.
Diagonal conditioning reaches a TVD of at the mixed conditions under the same protocol, about three times that of input conditioning. The gap is already present at the training conditions other than the origin, with mean TVD against , which rules out a purely extrapolation effect. It comes mainly from the width coordinate. The diagonally conditioned model follows the shift of the mean closely but changes the width by only 3–4% of the target change, so its error is largest at the corners of the square, where it reaches 0.085–0.109 against about 0.035 for input conditioning (Figure 4). More flexible diagonal variants, with quadratic condition dependence or all-to-all connectivity, did not improve on the affine map under this protocol (Appendix B 2). With three times as many parameters, diagonal conditioning thus fails to learn a condition dependence that input conditioning captures.
Conditional modeling of the Curie–Weiss model
The second experiment tests whether a single sixteen-qubit model learns how the Curie–Weiss distribution of Sec. IV B changes with temperature, and how much the encoder matters.
Both encoders we use treat one qubit, which we call the root (qubit 16), differently from the other fifteen, which we call the leaves and treat identically; is the collective spin of the leaves. The parity encoder applies rotations between the root and every leaf,
with angle . Because its reference state is supported on even-parity strings, the output is invariant under a global spin flip (§ 2 D). Its input probabilities are uniform, but it is not diagonal-equivalent, because its reference magnitudes depend on . Every Hamming weight enters with its binomial weight, and the IQP unitary must produce the entire temperature dependence of the profile from phases.
The twisted parity encoder adds two leaf operations before the root–leaf rotations (Figure 3(b)),
Together, the two leaf operations form the one-axis-twisting protocol for spin squeezing [44], and the twist generator is proportional to the Curie–Weiss energy of the leaves. The twisted parity encoder thus builds the form of the target interaction into the input, whereas the parity encoder imposes only the spin-flip symmetry. The twist narrows the transverse spread of the collective spin, initially along , and the rotation about turns this narrowed direction toward , so the leaf Hamming-weight distribution becomes narrower than binomial and the input probabilities depend on temperature. We fix the rotation angle at , close to the angle that narrows this distribution most for our twist strengths (Appendix B § 5). Since rotations are invariant under the global flip and commutes with , every layer commutes with , and both encoders impose for every parameter vector. Every layer is also invariant under permutations of the leaves, which makes sampling and amplitude ratios efficient (Appendix B § 5).
The angles of the twisted parity encoder follow the fixed schedule
with . The root–leaf angle rises from 0 and saturates at . This mirrors the target, which concentrates on the two fully magnetized states as grows, and keeps inputs beyond the largest training temperature close to those seen in training. The twist strength vanishes at and as , so the encoder twists mainly around the crossover. With , the rotation acts on as a global phase, and the encoder reduces to the parity encoder with angle . At , the input is and the model again matches the uniform target by construction. The constant and the rotation angle are fixed, and is selected by the hyper-parameter search (Appendix B § 5). None of the encoder parameters are updated by the optimizer.
We use an IQP unitary with all sixteen one-body and 120 two-body generators, giving 136 trainable phases per circuit. We train the conditional models at and evaluate them at , the last slightly beyond the largest training temperature, and on a separate grid of 101 temperatures in , computing the TVD and the two observables of Sec. IV B.
We write C-IQP for a conditional IQP model that shares one circuit across conditions and add its encoder in parentheses; our model is C-IQP (TP), with TP standing for twisted parity. Table I lists the IQP-based alternatives. C-IQP (P) uses the parity encoder with the linear angle . The per- circuits are trained directly at the evaluation , so they measure how well a single circuit fits each target rather than generalization, and test whether sharing one circuit across temperatures costs accuracy.
The main baseline is diagonal conditioning, the approach of previous conditional IQP models, which we call diagonally conditioned IQP (D-IQP). Whereas previous approaches fit the phases separately at each condition and interpolate between them [31, 32, 33], D-IQP is trained jointly across temperatures like our model. It uses one circuit with the same 136 generators, whose phases are a cubic polynomial in the condition, , with Chebyshev polynomials . Since four parameter vectors determine a unique cubic, it can represent any set of per-temperature fits at our four training temperatures, interpolated by a cubic. Its fixed input, the parity encoder at , imposes the same spin-flip symmetry.
We compare C-IQP (TP) with D-IQP, C-IQP (P), and the per- circuits. Figure 5 shows that C-IQP (TP) captures the change from a central magnetization peak in the hot regime () to two peaks at large absolute magnetization in the cold regime (). Near the crossover at , it follows the broad target profile, with only a shallow dip at , whereas C-IQP (P) places too much probability at large and too little near . D-IQP follows the envelope of the target profiles only roughly, distributing probability unevenly between neighboring Hamming weights. Much of its error lies within the Hamming-weight shells, among strings a profile does not distinguish.
Figure 5. Magnetization profiles of the sixteen-qubit Curie–Weiss models at the evaluation , for the conditional models with the twisted parity encoder, C-IQP (TP), and with the parity encoder, C-IQP (P), and for diagonal conditioning, D-IQP. Magnetization per spin is , where is Hamming weight. Colored curves show exact probability masses averaged across ten independent fits, and shading shows one standard deviation above and below the mean. Dashed black curves show the targets. The D-IQP band is drawn lighter for readability. It is wide because most D-IQP fits alternate between high and low probability on neighboring Hamming weights, each with a different pattern, which the mean smooths out. The displayed values of are excluded from training and model selection. Panels use separate vertical scales.
Figure 6(a) compares the errors. Averaged over the evaluation grid, the TVD is for C-IQP (TP), for C-IQP (P), and for D-IQP.

Figure 6. Sixteen-qubit Curie–Weiss models across temperature: (a) exact full-distribution TVD and (b) per-spin energy fluctuations on the 101-point evaluation grid. Curves show the models of Figure 5, C-IQP (TP) (solid), C-IQP (P) (dashed), and D-IQP (dotted), with the target dashed black in (b); curves and bands show means and one standard deviation over ten fits. Markers show the per- IQP (TP) (circles) and per- IQP (squares) circuits, trained at the evaluation . The inset in (b) zooms into the range of the target fluctuations.
| Model | Input and conditioning | Parameters | Difference from C-IQP (TP) |
| C-IQP (TP) | twisted parity, Eq. (28) | 136 | (proposed model) |
| C-IQP (P) | parity, | 136 | no twist, and a linear angle instead of the schedule of Eq. (28) |
| C-IQP (TP), | parity, Eq. (28) | 136 | twist removed (ablation, Appendix B § 5) |
| D-IQP | parity at , cubic phases | 544 | fixed input, with the condition in the diagonal-layer phases |
| Per- IQP (TP) | twisted parity at each | same encoder, with one circuit trained at each evaluation | |
| Per- IQP | one circuit per evaluation , on the standard input without the spin-flip symmetry |
Table 1. Sixteen-qubit Curie–Weiss models. All circuits use the same 136 one- and two-body generators. The shared models use one circuit for all , and the per- models one circuit per evaluation . Parameters count all phases or weights updated by the optimizer.
The twisted parity encoder lowers the mean TVD relative to the parity encoder at every nonzero grid point and in all ten seeds, which share training data and random streams between the two models. Its error grows through the hot regime and stays near its maximum between and , whereas C-IQP (P) has its largest errors at the crossover. An ablation that removes only the twist raises the averaged TVD to , above C-IQP (P), in all ten seeds, so the improvement comes from the twist (Appendix B § 5). Diagonal conditioning, although its 544 parameters can fit each training temperature separately, learns this target far less accurately than either conditional model under the same protocol. This already holds for C-IQP (P), whose encoder shares the spin-flip symmetry of the D-IQP input and adds only a linear angle, with a TVD 2.7 times lower. The gain over diagonal conditioning therefore does not depend on the target-specific twist, which lowers the TVD by a further factor of two.
Averaged over the evaluation , sharing one circuit across temperatures produces no appreciable loss in TVD. The per- IQP (TP) circuits, trained directly at these temperatures with three times the total parameters and optimizer updates across the three-circuit family, reach a TVD of , against for C-IQP (TP), which never sees these temperatures in training. The shared model is more accurate at and less accurate at and . The per- IQP circuits, whose standard input imposes no spin-flip symmetry, have substantially higher TVD at every evaluation and vary strongly across fits at .
Figure 6(b) shows the energy fluctuations. C-IQP (TP) reproduces the fluctuation peak in both height and position, at against the target at . Since , the fluctuation statistic depends on the fourth moment of the magnetization and is therefore sensitive to the tails of the profile, where similar-looking profiles can differ. At , for example, C-IQP (P) has 1.7 times the full TVD of C-IQP (TP) ( against ) but seven times its error in normalized by the target peak ( against ). Table V in Appendix B3e compares the observable errors of the main models at .
| Model | ||
| C-IQP (TP) | 0.007 | 0.010 |
| C-IQP (P) | 0.028 | 0.058 |
| D-IQP | 0.056 | 0.32 |
| Per- IQP (TP) | 0.005 | 0.018 |
| Large Transformer | 0.006 | 0.014 |
Table V. Observable errors of the sixteen-qubit Curie–Weiss models, averaged over and over ten fits: is the absolute error in and the absolute error in divided by the maximum target value, Eq. (37).
Comparison with classical models
The third experiment tests whether conditional IQP Born machines are competitive with classical generative models on the sixteen-qubit Curie–Weiss task.
We train three classical conditional models: a restricted Boltzmann machine (RBM) with condition-dependent biases [45, 46], an autoregressive Transformer [47], and a conditional variational autoencoder (VAE) [1, 48]. Each model receives as the conditioning variable, and we average the probabilities each model assigns to and , during both training and evaluation, to impose the same symmetry that the IQP encoders impose. The RBM and Transformer train on the exact likelihood objective, and the VAE on its evidence lower bound, and parameter counts include all parameters learned in training, among them the VAE encoder. We use two sizes of each model: small models with a number of learned parameters comparable to C-IQP, and large models with between 950 and 986. All models follow the hyperparameter-optimization protocol of the IQP models, including the training temperatures and target samples (Appendix B4).
Figure 7 shows their errors across the evaluation grid, and Table VI in Appendix B4 lists the aggregate scores. At comparable learned-parameter counts, C-IQP (TP) achieves lower mean TVD than the three tested classical models, in every seed; the closest small model, the Transformer, reaches against . Among the large models, with about seven times as many parameters, only the Transformer is comparable to C-IQP (TP). It reaches , higher than C-IQP (TP) in eight of ten seeds, whereas the large RBM and VAE reach and . It reproduces the mean absolute magnetization slightly more accurately, with an error of against , but the normalized fluctuations less accurately, with against (Table V). With the parity encoder, C-IQP (P) falls behind the small Transformer and the large RBM and Transformer.
Figure 7. Exact full-distribution TVD of C-IQP (TP) and C-IQP (P), each with 136 learned parameters, and of classical conditional models on the 101-point Curie–Weiss evaluation grid: (a) small models with 134 to 140 and (b) large models with 950 to 986 learned parameters. Curves and bands show means and one standard deviation over ten fits.
| Model | Parameters | Mean TVD | ||
| C-IQP (TP) | 136 | 0.060 ± 0.004 | 0.046 ± 0.004 | 0.0237 ± 0.0011 |
| C-IQP (P) | 136 | 0.128 ± 0.003 | 0.103 ± 0.004 | 0.086 ± 0.009 |
| Small RBM | 140 | 0.16 ± 0.06 | 0.13 ± 0.05 | 0.10 ± 0.07 |
| Small Transformer | 137 | 0.10 ± 0.02 | 0.068 ± 0.015 | 0.050 ± 0.016 |
| Small VAE | 134 (92) | 0.20 ± 0.02 | 0.138 ± 0.012 | 0.169 ± 0.018 |
| Large RBM | 986 | 0.076 ± 0.009 | 0.061 ± 0.007 | 0.030 ± 0.011 |
| Large Transformer | 971 | 0.070 ± 0.008 | 0.052 ± 0.005 | 0.022 ± 0.006 |
| Large VAE | 950 (548) | 0.14 ± 0.04 | 0.10 ± 0.03 | 0.12 ± 0.04 |
Table VI. Comparison of C-IQP models with classical conditional models on the 101-point Curie–Weiss evaluation grid. All models share the training temperatures, target samples, update budget, hyperparameter search, and selection rule. C-IQP models train on the sampled MMD objective, whereas the classical models train on their native likelihood or evidence lower bound. Parameters count all learned parameters used in training; for the VAEs, parentheses give the generator parameters alone. Mean TVD is the temperature-averaged full-distribution TVD, and Appendix B § 5 defines the scores and . Values are means ± standard deviations over ten independent fits. Lower is better, and bold marks the lowest mean in each column.
Three aspects of the training setup should be kept in mind when interpreting this comparison. The classical models train on likelihood-based objectives, exact for the RBM and Transformer at this size and an evidence lower bound for the VAE, whereas the C-IQP models minimize a sampled MMD estimate. Conversely, the twisted parity encoder applies an angle schedule chosen for this target, whereas the classical models learn their own dependence on . Finally, the twisted parity encoder builds the form of the Curie–Weiss interaction into an input that is symmetric under permutations of fifteen of the sixteen qubits, whereas the classical models must learn this structure from data. Neither model class imposes the full permutation symmetry of the target.
Large-scale training
The fourth experiment scales training to qubits, beyond the reach of exact simulation, and validates the models by comparing classically estimated observables with exact values (Sec. IV B).
These models use the parity encoder, which needs only a tree of nearest-neighbor CNOT gates, whereas the twist couples every pair of leaves (Sec. VI). Its angle rises with and saturates, for the same reason as in § 3 B, and its crossover can sharpen with through a steepness exponent selected from data (Appendix B § 5.0.3).
The IQP unitary has trainable phases only on each qubit and each nearest-neighbor pair of an open rectangular lattice. An lattice thus has parameters, against with all-to-all connectivity. The trained models have between 176 parameters at 64 qubits and 6048 at 2048 qubits (Appendix B § 5.0.3), which keeps training efficient at these sizes. Since exact probabilities are unavailable at these sizes, we select hyperparameters and checkpoints by the classically estimated MMD at development temperatures, with searches at 64 and 256 qubits whose selected configuration is used at every size, and we validate the models through the mean squared magnetization (Appendix B § 5.0.3).
At every size, the classically estimated test decreases within the first 500 updates and then stays nearly constant, from 0.042 to 0.0015 at 64 qubits and to 0.0043 at 2048 qubits (Figure 8(a)). Because the kernels emphasize low-order moments [28] and their bandwidths are rescaled with , these values are not comparable across sizes, so we validate them with the mean squared magnetization . It is a two-body quantity, , where is the mean of over distinct pairs, and the estimators of § 2 B give it classically at any size (Appendix B § 5.0.3).
Figure 8. Classical training of conditional IQP Born machines with 64 to 2048 qubits. All models use the parity encoder with the saturating angle of Eq. (39) and an open rectangular lattice of trainable one- and two-body phases. (a) Classically estimated test MMD during training, averaged over the test temperatures and two kernel bandwidths whose average bodyness, two and four, is kept fixed as the size varies; logarithmic vertical axis. (b) Error in the classically estimated mean squared magnetization, the model estimate at the selected checkpoint of each fit minus the exact target, at 41 equally spaced temperatures in . Gray vertical lines mark the training temperatures. Lines and bands show the mean and one standard deviation across three fits per size.
Figure 8(b) shows that the models reproduce away from the crossover at every size, within 0.016 of the target for and within 0.039 for , beyond the training temperatures. The long-range order of the cold phase is therefore learned and extrapolated up to 2048 qubits. Near the crossover, the learned transition is broader than that of the target, whose finite-size width shrinks as grows. At , the model gives between 0.21 and 0.24 at every size, whereas the target decreases from 0.142 at 64 qubits to 0.026 at 2048 qubits, and the error grows from 0.097 to 0.189. The crossover error also appears at the training temperature , where the model exceeds the target by 0.056 to 0.090. A 64-qubit model with all-to-all two-body connectivity, selected by a separate search with the same estimator budget, reduces the error at from 0.097 to 0.013 but underestimates beyond the development temperatures (Appendix B § 5.0.3). Although it also selected different encoder settings, this suggests that the lattice restriction contributes substantially to the crossover error.
We also chose the lattice ansatz because it maps directly onto superconducting processors with square-lattice connectivity. Sampling 16- and 30-qubit models on IQM’s Emerald quantum computer reproduces the unimodal-to-bimodal transition qualitatively, with substantial errors in the ordered regime, where device noise contracts the ordered peaks and the profile TVD at reaches 0.55 and 0.76 (Appendix B § 5.0.4).
Discussion
Proposition 1 shows that, at a fixed number of qubits, input conditioning can represent conditional families that no diagonally conditioned IQP model reproduces, because the IQP unitary cannot change the output’s X-basis distribution. On the two target families studied here, models conditioned through the input learned the family more accurately than diagonal conditioning under matched protocols, even when the diagonal baseline had more parameters. These targets need not lie outside the reach of diagonal conditioning; rather, the experiments indicate that carrying the condition in the input also eases learning in practice. Together, these results extend classically trained IQP Born machines from single distributions to conditional families, with a training cost set by the sampling-and-ratio interface of the encoder.
Within input conditioning, the encoder strongly affects performance. On the Curie–Weiss model, two encoders that impose the same symmetry differ by more than a factor of two in mean TVD, and removing the twist from the selected configuration makes it less accurate than the simpler parity encoder. The encoder also constrains the model before training. The IQP unitary leaves the distribution of outcomes in the X basis unchanged, so at every condition this distribution equals the reference distribution (Eq. (10)). The output state must therefore have the target as its computational-basis distribution and as its X-basis distribution. When no -qubit state has this pair of distributions, as for the family of Proposition 1 with any fixed input, no choice of phases reproduces the target at that condition. This compatibility is necessary but not sufficient, since one shared IQP unitary must still produce the right phases at every condition. How to test compatibility efficiently, and how to construct tractable encoders that satisfy it for a given target family, remain open questions.
At sixteen qubits, we select models on exact TVD at development conditions, so the reported accuracies characterize the tested configurations rather than a sample-based selection procedure. In a separate sixteen-qubit study, the development score built from the classically estimated MMD had a Spearman rank correlation of 0.97 with the exact TVD across the checkpoints of 120 configurations, against 0.83 for the MMD at the training temperatures (Appendix B § 5.0.3), which indicates that sample-based selection can track the exact TVD closely. At large sizes, we validate models through the low-order observable , which, like the low-order emphasis of the MMD kernel, does not certify the full distribution. Both target families are permutation invariant, which makes exact observables tractable at any size but also makes the targets easy to sample classically, so the experiments test representation and generalization rather than a quantum advantage; targets without this symmetry are a natural next step.
Quantum deployment raises separate questions. The sampling-and-ratio interface guarantees efficient classical access to the reference state, not an efficient circuit for preparing on hardware. The twist of the twisted parity encoder, like the Curie–Weiss interaction itself, couples every pair of leaves through two-qubit rotations, 105 at sixteen qubits, which compiles to many gates on processors without all-to-all connectivity. The large-scale and hardware experiments therefore use the parity encoder, which needs at most CNOT gates along a tree of nearest-neighbor couplers. The most accurate sixteen-qubit models and the large-scale demonstrations thus rely on different encoders, which reflects a trade-off between accuracy and preparation cost.
Classical trainability also raises the question of whether the trained models remain hard to sample. When the encoder can be absorbed into the diagonal layer, conditioning reparameterizes a standard IQP circuit, and sampling-hardness results for that circuit family carry over under the same assumptions [22, 23], although they do not certify that a particular trained circuit is hard to sample. Other encoders, including diagonal-equivalent ones with non-uniform reference magnitudes, need not admit this reduction, and the separation shown here concerns expressivity, not sampling complexity. Moreover, efficient estimation of the objective does not guarantee successful optimization: generative models whose output distributions anticoncentrate, a property that hardness arguments typically use, suffer from loss concentration and are not trainable on average [49]. Characterizing which encoders preserve sampling hardness, and whether training reaches hard instances, remains open.
Conclusions and outlook
Conditional IQP Born machines encode the condition in the input state and share one IQP unitary across conditions, and they can be trained classically whenever the input state supports efficient sampling and amplitude ratios. Encoding the condition in the input lets the model represent conditional families that no diagonally conditioned IQP model with the same number of qubits reproduces (Proposition 1). In our sixteen-qubit experiments, input conditioning was also more accurate than diagonal conditioning under the same training protocol, with a six times lower mean TVD on the Curie–Weiss model.
We trained conditional IQP Born machines for the Curie–Weiss model with up to 2048 qubits. The classical estimator used in training also enabled us to estimate physical observables of the trained circuits, such as the mean squared magnetization, at every size without quantum hardware. Because the Curie–Weiss model is permutation invariant, its exact observables reduce to sums over the values of the magnetization and remain computable at these sizes, so we could check what the models learned against the target. Both sides of this comparison are classically accessible, which makes the Curie–Weiss model a case where the workflow can be verified today, at a scale beyond state-vector simulation. The most interesting cases are those that cannot be checked classically, such as physical systems characterized only through experimental measurements. For such systems, a conditional IQP Born machine could still be trained classically at scale on the measured settings, such as temperatures, and then sampled to generate new measurements at other settings, including ones never probed in the experiment. This last step would need a fault-tolerant quantum computer of this size.
Recio-Armengol et al. [29] anticipated extending classical training to broader input states § 5.0.4 and 9.5; the sampling-and-ratio interface realizes this extension and lets the input depend on the condition. Three directions follow. The first is to characterize which conditional families a class of tractable encoders can represent, how this compares with adding hidden qubits [36], and how the encoder shapes generalization, as it does for supervised quantum models [50]. The second is to learn the encoder jointly with the IQP unitary, which requires extending the classical estimators to derivatives with respect to the input state; such trainable inputs would also benefit unconditional IQP Born machines. The third is to use conditional IQP Born machines as classically trainable components of generative architectures that reduce to sequences of condition-dependent sampling tasks, such as quantum denoising diffusion models [51, 52, 53] and autoregressive models.
Data availability
The code and data supporting this study are available on GitHub [54].
Author contributions
O.B. conceived the study, developed the formulation, carried out the experiments, and wrote the manuscript. The idea emerged from discussions between O.B. and E.T., whose subsequent input shaped the study and the manuscript. L.D. critically reviewed the manuscript, in particular its theoretical results. F.V. and A.G. supervised the research throughout the project, providing scientific guidance and oversight. All authors reviewed the final manuscript.
Large language models were used to assist with implementing the experiment and plotting code and with checking the grammar of the manuscript. All scientific content was conceived, verified, and approved by the authors, who take full responsibility for it.
Acknowledgments
O.B. is a fellow of Eurecat’s “Vicente López” PhD grant program. This work was partially supported by grants PID2024-157778OB-I00 and AIA2025-163435-C43 from the Spanish Ministry of Science and Innovation, the European Union, and the Agencia Estatal de Investigación, and grant No. TSI-100929-2023-2, Cátedra ENIA UAB-Cruïlla, from the Spanish Ministry of Economic Affairs and Digital Transformation and the European Union NextGenerationEU.
References
Proof of Proposition 1 and estimator variance
This appendix contains the proof of Proposition 1 (§ 2 C) and a bound on the variance of the training estimator, which supports the estimation claims of § 2 A and III A.
Proof of Proposition 1
Part (i). By Eq. (10), every output of diagonal conditioning on a fixed input has -basis distribution , the reference distribution of , whatever the parameters and generators. A member is therefore reproduced only if some pure state has computational-basis distribution and -basis distribution , that is, only if the two are compatible. Conversely, a diagonal layer with all Pauli- strings can multiply each entry of by any phase and hence reach every such state, so the condition is then also sufficient.
Part (ii). Fix and an IQP unitary , and write for the string with a single one at position . Consider a family of two distributions, indexed by : the first always outputs , and the second outputs , , or with probability each. We show that input conditioning reproduces this family with any fixed IQP unitary, whereas no fixed input does. The two distributions are the computational-basis distributions of
Input conditioning reproduces the family. Choose the encoder so that the reference state at condition is
The final Hadamard layer in the IQP unitary gives the output . Since only multiplies each amplitude by a phase, , where . Because is a two-qubit state tensored with , is a two-qubit state tensored with . These reference states therefore admit the sampling-and-ratio interface. Sampling the reference distribution amounts to drawing the first two bits from this two-qubit state and the remaining bits uniformly, and the amplitude ratios factorize as
where the fraction depends only on the first two bits of and , and the phase factor is the one that the estimator already evaluates, in operations, Eq. (7).
No fixed input reproduces it. By part (i), a fixed input would need an -basis distribution compatible with both members. The only state with the first distribution is , up to a global phase, and its -basis distribution is uniform (§ 2 C). It therefore suffices to show that no state with the second distribution has a uniform -basis distribution. Such a state has the form
If its -basis distribution were uniform, every expectation value with would vanish, because equals and for equally many strings . Since flips the first bit, , and as has no component along , . The same computation for and gives
The global phase of does not affect these values, so we may take real and positive. The first two then vanish only if and are purely imaginary, which makes real and nonzero, so the third cannot vanish. No -basis distribution is compatible with both members, and no fixed input reproduces the family. The same argument applies to any distribution supported on three strings; for , the second member is the distribution of Ref. [36 Lemma 11].
Variance of the estimator
Section II A states that the objective and its gradients can be estimated to additive accuracy in time polynomial in and , and Sec. III A that the variances of the estimators are bounded independently of . Recio-Armentgol et al. bound the variance of a single sampled expectation value by the inverse number of samples [29 Sec. 4.1], and § 2 B gives the same bounds for the ratio estimators. The complete estimator of Eq. (19) combines such terms on shared samples, and the same argument bounds it directly.
For each condition and mask, write for the three terms of Eq. (17). Since , we have and . Since distinct samples are independent, and for , and averaging preserves these bounds, so and the mixed pair mean of Eq. (18) is bounded by 2. The triangle inequality then gives
for every . Since , averaging over conditions, bandwidths, and masks preserves these bounds. Averaging independent repetitions of the complete estimator therefore gives mean-square error at most for the objective and for each gradient component, and by Chebyshev’s inequality, repetitions give additive error at fixed confidence.
Training and experiment details
This appendix gives the settings behind the results of § 3. Appendix B1 lists the settings common to the IQP experiments, and each following subsection gives the details of one experiment, titled after the corresponding section, followed by the deployment on quantum hardware. The accompanying repository contains the code, the frozen protocol file of every experiment, with the complete search spaces, condition grids, estimator budgets, and random streams, and the result bundles behind every figure and table [54]. Running reproduce.py retrains every reported model from the selected configurations and regenerates all figures and tables; with --full, it also reruns the hyperparameter searches. The hardware samples are read from the stored data, as device noise and calibration drift prevent exact reproduction.
Common training settings
All IQP models are trained with Adam [55], with and no weight decay. Each update draws one training condition and one kernel bandwidth uniformly. For an average bodyness , we set the bandwidth by inverting with Eq. (14), which gives
where counts the identity mask, as in Sec. III A. Table II lists the training settings and search spaces of each experiment, and Table III the selected hyperparameters. At sixteen qubits, we select checkpoints and hyperparameters on the exact TVD at development conditions outside the training set; the per-temperature circuits of § 3 B are selected at the temperatures at which they are trained. The large-scale and hardware models are selected on the classically estimated MMD (Appendix B5).
| Hamming-weight family | Curie–Weiss, | Large-scale training | |
| Updates (checkpoint interval) | 2000 (100) | 2000 (100) | 5000 (250) |
| Training conditions | nine, on the two axes | ||
| Target pool; minibatch | 1024; 256 | 1024; 256 | 10 000; 450 |
| Masks; reference samples | 1024; 1024 | 2048; 1024 | searched |
| Average bodyness | 2, , | 2, , | 2, 4 |
| Search | 60 random + 60 TPE trials | 60 random + 60 TPE trials | 60 random trials per schedule |
| Learning rate | , log | , log | , log |
| Schedule | constant or cosine | constant or cosine | cosine, final ratio , log |
| , log | , log | ||
| Initialization width | , log | , log | , log |
| One-body mean | 0 | 0 | |
| Kernel | — | ||
| Masks; references at | — | — | ; |
| Budget growth | — | — | , |
| Encoder constants | — | or , log | , log, , |
Table II. Training settings and hyperparameter search spaces. Searched hyperparameters list their range; “log” marks a log-uniform prior. Cosine schedules decay to 0.1 times the initial learning rate unless the final ratio is searched.
| Model | Learning rate | Schedule | Init. width | Other | |||
| Hamming-weight, input conditioning | 0.0012 | cosine | 0.011 | 0.83 | 0.0011 | 5, 7 | |
| Hamming-weight, diagonal conditioning | 0.00070 | cosine | 0.019 | 0.89 | 0.014 | 4.5, 7.99 | |
| C-IQP (TP) | 0.0055 | cosine | 0.076 | 0.94 | 0.0013 | 5, 7 | |
| C-IQP (P) | 0.0034 | cosine | 0.022 | 0.80 | 0.0080 | 4, 7.5 | |
| D-IQP | 0.0091 | constant | 0.10 | 0.85 | 0.0027 | 4, 5 | — |
| Per- IQP (TP) | 0.0051 | constant | 0.073 | 0.82 | 0.028 | 5, 7 | |
| Per- IQP | 0.090 | cosine | 0.033 | 0.95 | 0.0012 | 5, 5.5 | — |
| Small RBM | 0.073 | cosine | 1.2 | 0.82 | 0.048 | — | — |
| Small Transformer | 0.040 | cosine | 0.39 | 0.95 | 0.010 | — | — |
| Small VAE | 0.041 | cosine | 0.13 | 0.83 | 0.0022 | — | — |
| Large RBM | 0.045 | constant | 0.12 | 0.86 | 0.0026 | — | — |
| Large Transformer | 0.0031 | cosine | 0.22 | 0.94 | 0.036 | — | — |
| Large VAE | 0.039 | cosine | 1.8 | 0.88 | 0.0053 | — | — |
| Large-scale, all-to-all () | 0.0017 | cosine | 0.013 | 0.93 | 0.018 | — | , , , |
| Large-scale, lattice | 0.041 | cosine | 0.082 | 0.90 | 0.043 | — | , , , |
Table III. Selected hyperparameters, rounded to two significant figures; the protocol files in the repository give the exact values. The large-scale configuration is used at every size and for the hardware models. For the classical models, the initialization width is the multiplier of the Xavier initialization.
Details for generalization to unseen condition combinations
Conditions. Training uses the nine conditions at and at . Development uses at and at . The sixteen mixed conditions , with both coordinates nonzero, are used for reporting; four lie inside the training convex hull , and twelve lie outside.
Training and selection. All models use the settings of Table II and are selected by the mean exact TVD over the development conditions. Table III gives the selected configurations.
Encoder evaluation. The encoded input and its Hadamard transform are invariant under qubit permutations, so the input and reference amplitudes depend only on Hamming weight. The reference amplitudes follow from the input amplitudes by an linear transform over Hamming weights (a Krawtchouk transform), computed The large-scale configuration uses a final learning-rate ratio of 0.09, 512 masks and 1024 reference samples at , and budget growth . The all-to-all configuration uses a final learning-rate ratio of 0.34 and the same 512 masks and 1024 reference samples.
once per condition in operations. A reference bitstring is sampled by drawing its weight and then a uniformly random string of that weight, and each amplitude ratio is a quotient of two table entries, costing .
Diagonal conditioning. The baseline applies the same lattice circuit to , with phases . We initialize and draw as the phases of the conditional model. Three variants use the quadratic map on the lattice (200 parameters), and the affine and quadratic maps on all 136 one- and two-body products (408 and 680 parameters). All four follow the same protocol.
Results. Table IV lists the TVDs at the mixed conditions. At the origin, the affine diagonal baseline has TVD 0.017 0.002. Averaged over the ten fits, it reproduces between 0.95 and 1.02 of the target change in mean along , for , and between 0.03 and 0.04 of the target change in width along , for . Across the grid of Figure 4(a), the TVD of the conditional model has correlation 0.91 with the target TVD between each target and the uniform distribution, and 0.72 with the distance from each condition to the nearest training condition.
Details for conditional modeling of the Curie–Weiss model
All models use the same IQP ansatz, with sixteen one-body and 120 two-body phases and no ancillary qubits, and the encoders of Eqs. (27) and (26), or for per- IQP.
Twisted parity encoder
Every layer of the twisted parity encoder is invariant under permutations of the fifteen leaves, so the input and reference amplitudes depend only on the root bit and the leaf Hamming weight. The leaf state stays in the 16-dimensional symmetric subspace spanned by Dicke states, where the twist is diagonal and the collective rotation follows from the collective spin operators, so a matrix exponential gives the twisted leaf state. The Hamming-weight transform of Appendix B 2 then gives the table of reference amplitudes in operations per temperature.
Sampling draws the root bit and leaf weight from this table and then a uniformly random set of leaves; since a mask changes the leaf weight to and the root bit to , where and are the leaf parts of the sampled string and the mask, each amplitude ratio is a quotient of two table entries and costs .
To prepare the encoder, we apply a Hadamard gate to every qubit, an rotation to every pair of leaves, an rotation to every leaf, and an rotation between each leaf and the root. The 105 leaf-pair rotations commute, so the preparation is exact and needs no Trotterization.
We fix and the rotation angle , and is searched in (Appendix B § 5). For one-axis twisting, the rotation angle that narrows the leaf Hamming-weight distribution most tends to as the twisting strength vanishes [44]; for the selected schedule, it lies between and over .
Parity encoder
For a reference bitstring , with leaf bits and root bit , the parity encoder in Eq. (26) with angle gives the amplitude
C-IQP (P) uses . To sample this state, we draw the first fifteen bits independently with probability of obtaining one, and set the root bit to their parity. For an even-parity mask, the amplitude ratio is the product of the corresponding single-bit amplitude ratios over the leaves; for an odd-parity mask, it vanishes. Both operations require arithmetic operations. We prepare the reference state by applying to each leaf and a CNOT from each leaf to the root. A final Hadamard layer prepares the encoded input.
Training and selection
We fit the shared models at . Each per-temperature family contains three circuits fitted directly at , with separate learned phases and one jointly selected configuration.
We select checkpoints and configurations using a full-distribution TVD score and a physical-observable score . For a temperature grid and a nonnegative error curve , define
where is the trapezoidal integral over the grid divided by its temperature span. Averages over the evaluation grid use this normalized trapezoidal rule, and averages over the three evaluation temperatures are arithmetic means. We compute from exact probabilities over all bitstrings, and
with , where is the maximum target value of for $0\leq\beta\leq 2 (Appendix B § 5). Shared models are scored on a grid of seventeen temperatures in that excludes the evaluation temperatures but contains a point within of each, and each per-temperature circuit is scored at the temperature at which it is trained. Within each fit, we select the checkpoint with the smallest among those whose lies within of the minimum.
The TPE sampler receives only . Confirmation refits the top three or four configurations with five fits of three initializations each, ranking shared models on together with twenty interleaved temperatures, and the final stage performs ten fits with three initializations per circuit. All families share the random streams for target pools and stochastic estimates. The evaluation grid contains 101 temperatures in ; neither this grid nor enters the training or selection of the shared models.
Diagonal conditioning
The D-IQP baseline uses the fixed input , the parity encoder at , whose reference state is uniform on even-parity strings, so the output is invariant under global spin flips for every parameter vector. Its phases , with , map to the Chebyshev interval . We initialize from the paired initialization streams of the other models and set for , so the initial model is a constant-parameter IQP circuit. It follows the protocol of the other shared models. Over the ten final fits, is , the temperature-averaged TVD is , and the mean TVD lies between and at every grid point.
Physical observables
For a distribution , let denote its exact Hamming-weight profile and set with . We compute its moments as and its mean absolute magnetization as . With in units , Eq. (23) becomes
We compute each observable within each fit before averaging.
Twist ablation
To isolate the effect of the twist, we refit the selected C-IQP (TP) configuration with , which removes the twist and leaves every other setting unchanged.
Protocol. With , the twist strength vanishes, the rotation acts trivially, and the encoder becomes the parity encoder with the schedule . We follow the final-stage protocol, with ten fits paired with the reported fits, and keep the optimizer settings selected for without a new search. We repeat the refit at the two ends of the search range of .
Results. At the selected , removing the twist raises the temperature-averaged TVD from 0.046 ± 0.004 to 0.132 ± 0.004, above the 0.103 ± 0.004 of C-IQP (P). Every fit has a higher and a higher temperature-averaged TVD than its paired C-IQP (P) fit, and hence than its paired C-IQP (TP) fit. At and 0.7, the TVD is 0.299 ± 0.006 and 0.148 ± 0.009. Under the tested configurations, the improvement of C-IQP (TP) therefore comes from the twist, and the schedule alone does not reproduce it.
Details for comparison with classical models
All classical models take the inverse temperature as their condition. We symmetrize each model as , where flips every bit, and use in training, evaluation, and sampling, so every model satisfies ; none imposes permutation symmetry.
The RBM has hidden units and visible and hidden biases that depend affinely on , with for the small model and for the large one, and we compute its partition function exactly over the visible configurations. The Transformer is a decoder-only model with one block and one attention head that generates the bits autoregressively, with an affine embedding of and model and feed-forward widths 3 and 7 for the small model and 8 and 34 for the large one. The VAE has a one-dimensional latent variable with a standard normal prior and single-hidden-layer encoder and decoder networks that receive by concatenation, with widths 2 and 4 for the small model and 20 and 28 for the large one. We symmetrize its Bernoulli decoder likelihood as the equal mixture of the probabilities of and and train on the evidence lower bound of this symmetrized decoder. Parameter counts include the VAE encoder, which is needed only for training; the generators alone have 92 and 548 parameters. The RBM and Transformer minimize the exact negative log-likelihood, and we evaluate their exact probabilities. For the VAE, we integrate the decoder over the latent prior by Gauss–Hermite quadrature, refined until the TVD changes by at most .
The classical models reuse the training temperatures, target pools, minibatch size, update budget, checkpoint schedule, Adam optimizer, stages, scores, and tolerance of the IQP models, and fits with the same index share their target samples, which allows the paired comparisons of Sec. VC. Each model family has its own search of 60 random and 60 TPE trials over the learning rate, schedule, and Adam moments, with the Curie–Weiss ranges of Table II, and over a multiplier in , log-uniform, that scales the Xavier initialization of the weights; biases start at zero. Table III gives the selected configurations.
Details for large-scale training
Encoder and architecture. We use the parity encoder in Eq. (26), generalized to qubits, with the saturating angle
where is the logistic function, the saturation angle, the inflection point, and a steepness whose growth with is set by . The angle vanishes at , so the input there is , and unlike the linear angle of C-IQP (P), it stops growing beyond the training temperatures. The IQP unitary has one-body phases on every qubit and two-body phases on the nearest-neighbor pairs of open rectangular lattices, from at 64 qubits to at 2048 qubits. At size , the target is .
Training. We train three fits per size, with the settings of Table II. With the selected budget growth, the masks and reference samples per update grow from 512 and 1024 at 64 qubits to 2880 and 5824 at 2048 qubits. For the parity encoder and the lattice ansatz, an update with reference samples, target samples, and masks of average bodyness requires operations, since each amplitude ratio, phase difference, and target character involves only the qubits and generators touched by the mask. The arithmetic cost per update therefore grows no faster than .
Development score and search. We select checkpoints on the development grid , which shares no temperature with training or with the MMD test temperatures . Let denote the squared-MMD estimate at , averaged over two independent estimation banks of target, mask, and reference draws and over both bandwidths. We score a checkpoint by
where the integral uses the trapezoidal rule on , and we retain negative unbiased estimates. In a sixteen-qubit study on the lattice, where exact probabilities are available, this score had a Spearman correlation of 0.97 with the exact TVD at 40 test temperatures, aggregated by the same rule, across the 2520 checkpoints of 120 random configurations, against 0.83 for the squared MMD at the training temperatures. The search trains each configuration at 64 and 256 qubits and ranks it by the larger of its two scores. We confirm the three leading saturating configurations with three independent fits per size and select the one with the lowest score; Table III gives its constants.
Evaluation. We draw five MMD test banks independently of training and selection and, every 500 updates, average the estimated MMD over the banks, test temperatures, and bandwidths (Figure 8(a)). The initial phases are small, so the value at update zero approximates that of the encoder alone. For , we use the selected checkpoint of each fit. At each of 41 equally spaced temperatures in , we draw five independent banks of 256 distinct pairs , estimate each with 4096 reference draws, and average the resulting unbiased estimates of over the banks. The target value follows exactly from the one-dimensional magnetization distribution.
All-to-all connectivity at 64 qubits. To test the lattice restriction, we ran a separate search of 60 saturating configurations with the all-to-all two-body ansatz at 64 qubits, with the estimator budget of the lattice model, and trained three fits of the selected configuration, using the seeds, data, and evaluation of the lattice fits. The all-to-all model removes most of the crossover error and lowers the test MMD, but underestimates in the ordered regime beyond the development grid (Table VII).
TABLE VII. Lattice and all-to-all models at 64 qubits, each with its own selected configuration. Test MMD is the held-out value at the selected checkpoint, and is the classically estimated model value minus the exact target. Values are means over three fits. Lower absolute values are better.
| Quantity | lattice | All-to-all |
| Trainable parameters | 176 | 2080 |
| Test MMD | 0.00143 | 0.00097 |
| at | +0.056 | −0.027 |
| at | +0.097 | −0.013 |
| at | +0.045 | +0.004 |
| Mean , | 0.065 | 0.012 |
| at | −0.010 | −0.106 |
| at | −0.008 | −0.184 |
Table VII. Lattice and all-to-all models at 64 qubits, each with its own selected configuration. Test MMD is the held-out value at the selected checkpoint, and is the classically estimated model value minus the exact target. Values are means over three fits. Lower absolute values are better.
Deployment on quantum hardware
We train conditional IQP Born machines at 16 and 30 qubits, on and lattices, and deploy them on IQM Emerald, a 54-qubit superconducting processor with square-lattice topology. The lattice is the largest rectangle of our ansatz that fits on the couplers calibrated on the day of the experiment, and we place each lattice to maximize the product of the calibrated fidelities of the CZ gates and readouts it uses. Each two-body phase compiles to two CZ gates, and the parity encoder to a tree of CNOT gates along lattice couplers, so the circuits need no routing. More generally, the lattice ansatz maps onto the couplers of square-lattice processors [56, 57, 58], whereas earlier hardware deployments of IQP Born machines used heavy-hexagonal connectivity [59, 60, 61]. The compiled 16- and 30-qubit circuits contain 74 and 152 CZ gates, with native-gate depths 62 and 92. We sample the training temperatures and , the crossover , and the extrapolated temperature , with 3000 shots each. All runs took place on 2026-10-16, and the repository contains the calibration data of that day [54].
Figure 9 shows that the hardware reproduces the transition from a unimodal to a bimodal magnetization profile at both sizes. The exact profiles of the noiseless models, obtained by state-vector simulation, agree with the classical estimates of in Table VIII to within 0.002. At , the hardware profiles have TVD 0.14 to the target at both sizes. At the crossover , they are closer to the target than the noiseless models, with TVD 0.14 and 0.20 against 0.20 and 0.27 for 16 and 30 qubits, because noise partly fills the dip of the lattice models near (§ 3 D). In the ordered regime, noise moves weight from the ordered peaks toward , and the TVD at grows to 0.55 and 0.76, against 0.14 for the noiseless models. For , the hardware retains about 60% of the model’s excess correlation at 16 qubits and about 47% at 30 qubits, nearly independently of temperature, a deviation more than an order of magnitude above shot noise. A suppression that is nearly uniform in and stronger for the circuit with about twice as many CZ gates is consistent with gate errors and decoherence. Readout mitigation with M3 [62] changes the profile TVDs by at most 0.02, so we report unmitigated results.
Figure 9. Magnetization profiles sampled from classically trained conditional IQP Born machines on IQM Emerald, without error mitigation. Bars show the distribution of over 3000 shots at each inverse temperature, dashed lines the exact target, and solid lines the exact profile of the noiseless model. Each panel reports the total variation distance between the hardware and target profiles.
| Hardware | Model | Target | Profile TVD | ||
| 16 | 0.6 | 0.105 0.002 | 0.116 0.001 | 0.131 | 0.139 |
| 16 | 1 | 0.217 0.004 | 0.306 0.003 | 0.278 | 0.139 |
| 16 | 1.4 | 0.406 0.005 | 0.615 0.001 | 0.574 | 0.269 |
| 16 | 2 | 0.533 0.005 | 0.834 0.002 | 0.888 | 0.551 |
| 30 | 0.6 | 0.058 0.001 | 0.0712 0.0003 | 0.075 | 0.139 |
| 30 | 1 | 0.135 0.002 | 0.247 0.001 | 0.205 | 0.195 |
| 30 | 1.4 | 0.296 0.004 | 0.588 0.001 | 0.593 | 0.497 |
| 30 | 2 | 0.424 0.005 | 0.879 0.001 | 0.904 | 0.765 |
Table VIII. Mean squared magnetization on IQM Emerald. Hardware values are sample means over 3000 shots, with one standard error from shot noise. Model values are classical estimates for the noiseless model (§ 3 D), with one standard error over the five estimation banks of correlator pairs. The last column gives the TVD between the hardware and target magnetization profiles.
References
- [1]K. Sohn, H. Lee, and X. Yan, Learning structured output representation using deep conditional generative models, in Advances in Neural Information Processing Systems, Vol. 28 (2015).
- [2]J. Walker, C. Doersch, A. Gupta, and M. Hebert, An uncertain future: Forecasting from static images using variational autoencoders, in Computer Vision – ECCV 2016, Lecture Notes in Computer Science, Vol. 9911 (2016) pp. 835–851, arXiv:1606.07873.arxiv.org/abs/1606.07873
- [3]X. Yan, J. Yang, K. Sohn, and H. Lee, Attribute2Image: Conditional image generation from visual attributes, in Computer Vision – ECCV 2016, Lecture Notes in Computer Science, Vol. 9908 (2016) pp. 776–791, arXiv:1512.00570.arxiv.org/abs/1512.00570
- [4]M. Mirza and S. Osindero, Conditional generative adversarial nets (2014), arXiv:1411.1784 [cs.LG].arxiv.org/abs/1411.1784
- [5]T. Miyato and T. Koyama, cGANs with projection discriminator, in International Conference on Learning Representations (2018) arXiv:1802.05637.
- [6]P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, Image-to-image translation with conditional adversarial networks, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017) arXiv:1611.07004.arxiv.org/abs/1611.07004
- [7]J. Ho and T. Salimans, Classifier-free diffusion guidance (2022), arXiv:2207.12598 [cs.LG].arxiv.org/abs/2207.12598
- [8]P. Dhariwal and A. Nichol, Diffusion models beat GANs on image synthesis, in Advances in Neural Information Processing Systems, Vol. 34 (2021) arXiv:2105.05233.arxiv.org/abs/2105.05233
- [9]M. Benedetti, D. Garcia-Pintos, O. Perdomo, V. Leyton-Ortega, Y. Nam, and A. Perdomo-Ortiz, A generative modeling approach for benchmarking and training shallow quantum circuits, npj Quantum Information 5, 45 (2019).DOI
- [10]J.-G. Liu and L. Wang, Differentiable learning of quantum circuit Born machine, Physical Review A 98, 062324 (2018).DOI
- [11]Z. He, L. Li, S. Zheng, Z. Huang, and H. Situ, A conditional generative model based on quantum circuit and classical optimization, International Journal of Theoretical Physics 58, 1138 (2019).DOI
- [12]O. Kiss, M. Grossi, E. Kajomovitz, and S. Vallecorsa, Conditional Born machine for Monte Carlo event generation, Physical Review A 106, 022612 (2022).DOI
- [13]Q.-W. Zeng, H.-Y. Ge, C. Gong, and N.-R. Zhou, Conditional quantum circuit Born machine based on a hybrid quantum–classical framework, Physica A: Statistical Mechanics and its Applications 618, 128693 (2023).
- [14]J. Zeng, Y. Wu, J.-G. Liu, L. Wang, and J. Hu, Learning and inference on generative adversarial quantum circuits, Physical Review A 99, 052306 (2019), arXiv:1808.03425 [quant-ph].DOI
- [15]S. Certo, A. Pham, N. Robles, and A. Vlasic, Conditional generative models for learning stochastic processes, Quantum Machine Intelligence 5, 42 (2023), arXiv:2304.10382 [quant-ph].DOI
- [16]D. Wierichs, J. Isaac, C. Wang, and C. Y.-Y. Lin, General parameter-shift rules for quantum gradients, Quantum 6, 677 (2022), arXiv:2107.12390 [quant-ph].DOI
- [17]J. Bowles, D. Wierichs, and C.-Y. Park, Backpropagation scaling in parameterised quantum circuits, Quantum 9, 1873 (2025), arXiv:2306.14962 [quant-ph].DOI
- [18]J. Preskill, Quantum computing in the NISQ era and beyond, Quantum 2, 79 (2018), arXiv:1801.00862.DOI
- [19]S. Wang, E. Fontana, M. Cerezo, K. Sharma, A. Sone, L. Cincio, and P. J. Coles, Noise-induced barren plateaus in variational quantum algorithms, Nature Communications 12, 6961 (2021), arXiv:2007.14384 [quant-ph].DOI
- [20]D. Shepherd and M. J. Bremner, Temporally unstructured quantum computation, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 465, 1413 (2009), arXiv:0809.0847 [quant-ph].
- [21]B. Coyle, D. Mills, V. Danos, and E. Kashefi, The Born supremacy: Quantum advantage and training of an Ising Born machine, npj Quantum Information 6, 60 (2020).DOI
- [22]M. J. Bremner, R. Jozsa, and D. J. Shepherd, Classical simulation of commuting quantum computations implies collapse of the polynomial hierarchy, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 467, 459 (2011).DOI
- [23]M. J. Bremner, A. Montanaro, and D. J. Shepherd, Average-case complexity versus approximate simulation of commuting quantum computations, Physical Review Letters 117, 080501 (2016).DOI
- [24]M. J. Bremner, A. Montanaro, and D. J. Shepherd, Achieving quantum supremacy with sparse and noisy commuting quantum computations, Quantum 1, 8 (2017).DOI
- [25]S. C. Marshall, S. Aaronson, and V. Dunjko, Improved separation between quantum and classical computers for sampling and functional tasks, in 40th Computational Complexity Conference (CCC 2025), Leibniz International Proceedings in Informatics (LIPIcs), Vol. 339 (Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2025) pp. 5:1–5:14, arXiv:2410.20935 [quant-ph].arxiv.org/abs/2410.20935
- [26]M. Van den Nest, Simulating quantum computers with probabilistic methods, Quantum Information and Computation 11, 784 (2011), arXiv:0911.1624 [quant-ph].arxiv.org/abs/0911.1624
- [27]A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola, A kernel two-sample test, Journal of Machine Learning Research 13, 723 (2012).DOI
- [28]M. S. Rudolph, S. Lerch, S. Thanasilp, O. Kiss, O. Shaya, S. Vallecorsa, M. Grossi, and Z. Holmes, Trainability barriers and opportunities in quantum generative modeling, npj Quantum Information 10, 116 (2024).DOI
- [29]E. Recio-Armengol, S. Ahmed, and J. Bowles, Train on classical, deploy on quantum: Scaling generative quantum machine learning to a thousand qubits (2026), arXiv:2503.02934v2 [quant-ph].
- [30]A. Huang, W. Maxwell, V. Belis, E. Peters, J. Pye, S. Jahangiri, and J. Bowles, Spectral Born machines: Classi- cally trainable quantum generative models for discrete data (2026), arXiv:2607.06675 [quant-ph].
- [31]C.-Y. Liu, L. Placidi, E. Brunner, and E. Rinaldi, Toward generative quantum utility via correlation-complexity map (2026), arXiv:2603.06440v2 [cs.LG].arxiv.org/abs/2603.06440
- [32]C.-Y. Liu, L. Placidi, M. Ballarin, and E. Rinaldi, Generative IQP circuit learning with physics-informed latent initialization (2026), arXiv:2607.28866v1 [quant-ph].arxiv.org/abs/2607.28866
- [33]X. Zou, S. Duan, C. Fleming, G. Liu, R. N. Kompella, S. Ren, and X. Xu, ConQuER: Modular architectures for control and bias mitigation in IQP quantum generative models (2025), arXiv:2509.22551v2 [quant-ph].arxiv.org/abs/2509.22551
- [34]V. Havlíček, Amplitude ratios and neural network quantum states, Quantum 7, 938 (2023), arXiv:2201.09128 [quant-ph].DOI
- [35]E. Armengol and J. Bowles, IQPOpt: Fast optimization of instantaneous quantum polynomial circuits in JAX (2026), arXiv:2501.04776v2 [quant-ph].arxiv.org/abs/2501.04776
- [36]A. Kurkin, K. Shen, S. Pielawa, H. Wang, and V. Dunjko, Universality and kernel-adaptive training for classically trained, quantum-deployed generative models (2025), arXiv:2510.08476 [quant-ph].arxiv.org/abs/2510.08476
- [37]J. J. Meyer, M. Mularski, E. Gil-Fuster, A. A. Mele, F. Arzani, A. Wilms, and J. Eisert, Exploiting symmetry in variational quantum machine learning, PRX Quantum 4, 010328 (2023).DOI
- [38]J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Barren plateaus in quantum neural network training landscapes, Nature Communications 9, 4812 (2018).
- [39]M. Larocca, S. Thanasilp, S. Wang, K. Sharma, J. Biamonte, P. J. Coles, L. Cincio, J. R. McClean, Z. Holmes, and M. Cerezo, Barren plateaus in variational quantum computing, Nature Reviews Physics 7, 174 (2025).DOI
- [40]M. Cerezo, A. Sone, T. Volkoff, L. Cincio, and P. J. Coles, Cost function dependent barren plateaus in shallow parametrized quantum circuits, Nature Communications 12, 1791 (2021).DOI
- [41]K. Shen, S. Pielawa, V. Dunjko, and H. Wang, Characterizing trainability of instantaneous quantum polynomial circuit Born machines (2026), arXiv:2602.11042v2 [quant-ph].arxiv.org/abs/2602.11042
- [42]S. Lerch, J. Bowles, R. Puig, E. Armengol, Z. Holmes, and S. Thanasilp, IQP Born machines under data-dependent and agnostic initialization strategies (2026), arXiv:2603.14576 [quant-ph].arxiv.org/abs/2603.14576
- [43]R. S. Ellis, Entropy, Large Deviations, and Statistical Mechanics, Classics in Mathematics (Springer, Berlin, Heidelberg, 2006).
- [44]M. Kitagawa and M. Ueda, Squeezed spin states, Physical Review A 47, 5138 (1993).
- [45]G. E. Hinton, Training products of experts by minimizing contrastive divergence, Neural Computation 14, 1771 (2002).pubmed.ncbi.nlm.nih.gov/12180402
- [46]G. W. Taylor, G. E. Hinton, and S. T. Roweis, Modeling human motion using binary latent variables, in Advances in Neural Information Processing Systems, Vol. 19 (2007) pp. 1345–1352.DOI
- [47]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, Attention is all you need, in Advances in Neural Information Processing Systems, Vol. 30 (2017).arxiv.org/abs/1706.03762
- [48]D. P. Kingma and M. Welling, Auto-encoding variational Bayes, in International Conference on Learning Representations (2014) arXiv:1312.6114arxiv.org/abs/1312.6114
- [49]S. Herbst, I. Brandić, and A. Pérez-Salinas, Limits of quantum generative models with classical sampling hardness (2025), arXiv:2512.24801 [quant-ph].arxiv.org/abs/2512.24801
- [50]M. C. Caro, E. Gil-Fuster, J. J. Meyer, J. Eisert, and R. Sweke, Encoding-dependent generalization bounds for parametrized quantum circuits, Quantum 5, 582 (2021), arXiv:2106.03880 [quant-ph].DOI
- [51]A. Acioppo, L. Colantonio, S. Bordoni, and S. Giagu, Quantum diffusion models (2023), arXiv:2311.15444 [quant-ph].
- [52]D. Quinn, L. Buffoni, S. Gherardini, and G. De Chiara, Conditioning in generative quantum denoising diffusion models, Quantum Science and Technology 11, 035047 (2026), arXiv:2509.17569 [quant-ph].DOI
- [53]C. Chen, Q. Zhao, M. Zhou, D. Niyato, Z. He, and H. Situ, Overcoming dimensional factorization limits in discrete diffusion models through quantum joint distribution learning, npj Quantum Information 12, 49 (2026).arxiv.org/abs/2505.05151
- [54]O. Balló-Gimbernat, Conditional generation with IQP circuits: Code and data (2026), https://github.com/uriballo/conditional-generation-iqp.
- [55]D. P. Kingma and J. Ba, Adam: A method for stochastic optimization, in International Conference on Learning Representations (2015) arXiv:1412.6980 [cs.LG].arxiv.org/abs/1412.6980
- [56]F. Arute, K. Arya, R. Babbush, et al., Quantum supremacy using a programmable superconducting processor, Nature 574, 505 (2019).DOI
- [57]L. Abdurakhimov et al., Technology and performance benchmarks of IQM’s 20-qubit quantum computer (2024), arXiv:2408.12433 [quant-ph].arxiv.org/abs/2408.12433
- [58]IBM, IBM delivers new quantum processors, software, and algorithm breakthroughs on path to advantage and fault tolerance, IBM Newsroom (2025), https://newsroom.ibm.com/2025-11-12-ibm-delivers-new-quantum-processors,-software,-and-algorithm-breakthroughs-on-path-to-advantage-and-fault-tolerance.
- [59]O. Balló-Gimbernat, M. Arroyo-Sánchez, P. García-Molina, A. Garriga, and F. Vilariño, Shallow instantaneous quantum polynomial-time circuits for generative modeling on noisy intermediate-scale quantum hardware, Physical Review A 113, 042617 (2026).
- [60]J. Slim, S. Monaco, F. Rehm, D. Krücker, and K. Borras, An IQP Born machine for calorimeter image generation at 64 qubits with compiled-IQP deployment (2026), arXiv:2605.27735 [quant-ph].arxiv.org/abs/2605.27735
- [61]J.-Y. Ryu, W. Song, K. Bae, W. Lee, and I. Sohn, Deployment of large IQP circuit Born machines on qubit-limited hardware using noise injection (2026), arXiv:2609.35163 [quant-ph].arxiv.org/abs/2609.35163
- [62]P. D. Nation, H. Kang, N. Sundaresan, and J. M. Gambetta, Scalable mitigation of measurement errors on quantum computers, PRX Quantum 2, 040326 (2021).DOI







