WorksPhysics and Astronomy
Exploiting the Weyl-Chamber Structure for Hardware-Aware Quantum Circuit Compression
Abstract
Quantum circuit compression based on tensor networks enables long-time simulation of quantum many-body dynamics on near-term hardware. While Riemannian optimization has been successfully employed for this purpose, it typically optimizes over the manifold of generic two-qubit unitaries. Consequently, the resulting gates must still be synthesized into the hardware's native gate set via KAK decomposition, an overhead that partially negates the original compression gain. We show that Riemannian-optimized gates already exhibit a predictable Weyl structure: near- for the Ising model and near-fSim for the Heisenberg XXZ model. Building on this insight, we investigate whether parameterized optimization directly over the native gate set can offer a beneficial trade-off between expressivity and circuit efficiency by reducing the circuit's dimensionality. When comparing the two approaches, we find that the small residual Weyl parameters neglected by the latter are essential for high compression of the Ising propagator, whereas for the Heisenberg propagator, the native XXZ structure can be exploited directly to build a more efficient compressed circuit.
Introduction
Simulating the time evolution of quantum systems, particularly in quantum many-body physics, remains a primary motivation for developing quantum computers [1]. In this work, we consider a time-independent Hamiltonian and aim to implement the time evolution operator . Besides the qubitization framework [2], a natural procedure is using Trotterization-type circuit ansätze [3], which do neither require block-encoding nor additional auxiliary qubits.
Several works have demonstrated that “vanilla” Trotterization circuits based on product formulas can be further optimized to improve the approximation accuracy and to reduce the circuit depth [4, 5, 6, 7, 8, 9, 10, 11, 12, 13]. One prominent pathway combines Riemannian optimization (RO) with tensor networks (TNs) [4, 5, 14, 15, 16] to reduce the number of layers in a Trotterized circuit by optimizing over the manifold of generic unitaries. However, a significant practical bottleneck emerges during hardware execution: RO returns dense unitaries, each of which must be decomposed into the underlying native gate set. This synthesis naturally increases circuit depth, partially undoing the initial compression [17]. The Cartan coordinates of these gates characterize their non-local entangling content and reveal predictable, model-dependent structures, suggesting that some systems may not require the expressivity of the full unitary manifold. This raises the possibility that a lower-dimensional native ansatz can retain the relevant expressivity while reducing synthesis overhead. We therefore ask: can optimization restricted to a hardware-native gate family achieve lower approximation error at a fixed native entangling depth than optimization over generic two-qubit unitaries followed by synthesis?
Motivated by these structural patterns, we suggest an alternative hardware-aware parameterized optimization (PO) method and compare it with RO. While optimizing structured, parameterized ansätze has been explored in classical compression schemes [11, 12, 13], our approach expands on these foundations in three key ways. First, rather than relying on explicit translational invariance to extrapolate from small chains [11], we optimize the full chain directly through matrix product operator (MPO) gradients. Second, we explicitly address the expressivity limits of native gate sets – a question previously left open [12]. Finally, we generate a reusable time-evolution operator capable of reaching longer times, in contrast to methods that compile single-target state-preparation circuits [13].
When comparing PO and RO, we observe that PO struggles to match RO’s generalizability and low loss for the Ising propagator. In contrast, the local XXZ structure in the Heisenberg propagator can be leveraged to construct a superior compressed circuit. These results inform the design of more resource-efficient hardware experiments [17, 18].
Motivation
In order to execute a quantum circuit experimentally, its constituent quantum gates ultimately need to be decomposed into the native gate set of the target hardware [19, 20, 21]. To assess how compression using TN methods affects the two-qubit gate count and circuit depth after decomposition into the target hardware’s native gate set, we first examine the fundamental algebraic structure of two-qubit gates through the KAK decomposition – the Cartan decomposition of the Lie group [22, 23]. This theorem states that any arbitrary two-qubit unitary can be factored into a canonical, multiparameter entangling core flanked by local single-qubit rotations:
Here, are local single-qubit rotations, and are the standard Pauli matrices. The real-valued parameters are the so-called Weyl or Cartan coordinates [24, 25], which fully characterize the non-local, entanglement of up to local equivalence [24, 26].
The local gauge freedom can be fixed by confining these coordinates to a canonical sub-volume known as the Weyl chamber, defined by the ordering (Figure 1). Points in the bulk of the chamber correspond to generic two-qubit gates, with all three coordinates nonzero. Synthesizing such a generic unitary on standard hardware requires, in the worst case, three maximally entangling two-qubit gates (e.g., CNOT, CZ, or an equivalent parametric rotation) interleaved with single-qubit rotations [20]. By contrast, gates on the face of the chamber, such as the fractional iSWAP or the B gate [24], can be synthesized with only two CNOT-equivalent entangling gates, and gates locally equivalent to CNOT itself, , require a single one [21]. This distinction is central to our motivation: for a given native gate set, the entangling-gate cost of a circuit after decomposition is determined not only by the number of two-qubit gates it contains, but also by where in the Weyl chamber those gates lie.

Figure 1. Geometric representation of the Weyl chamber for two-qubit local equivalence classes. The chamber forms a tetrahedron within the reduced bounding region . Distinct high-symmetry operating points correspond to core quantum gates: Identity (I), CNOT, iSWAP, and SWAP define the tetrahedral vertices, while the B gate is positioned strictly on the base edge directly between CNOT and iSWAP. The native gate families used in this work are highlighted: gates lie on the I–CNOT edge (orange), whereas fSim gates covers the face (I–iSWAP–SWAP) (teal).
Figure 2a illustrates this structural overhead for a circuit composed of generic two-qubit unitaries. To address this limitation, we aim to preserve the advantages of the TN optimization framework developed in Ref. [5] while adapting it to a hardware-native, parameterized ansatz.

Figure 2. One repetition of second-order Trotterization under the two splittings (Ising case). Blue and pink boxes act on even and odd bonds, respectively; grey boxes are single-qubit gates. (a) With the splitting used for Riemannian optimization Eq. (4), each generic two-qubit gate must be synthesized into the native gate set, here three gates interleaved with single-qubit rotations; adjacent single-qubit rotations are merged. This triples the two-qubit gate depth. (b) With the hardware-aware splitting Eq. (5), the Trotter step is implemented directly with native gates, yielding a shallower circuit.
Analyzing the Weyl coordinate distributions of RO-optimized circuits reveals a striking structural pattern: despite having the entire generic manifold available for optimization, the resulting matrices naturally adopt simplified structures that mirror the local interactions of the simulated Hamiltonian:
Ising model: For each of the optimized gates, one Weyl coordinate is consistently several orders of magnitude larger than the other two (Figure 3a). Each optimized gate is therefore locally close to a single rotation,1 which requires one native entangling gate instead of the three needed for a generic gate. Most of the synthesis overhead in Figure 2a is thus spent on two small, apparently subleading coordinates; in Section IV A we test whether they can be dropped by optimizing gates directly.

Figure 3. (a) Ising model [Eq. (3)] with , , , . Coordinates are shown as log-space mean , since the three components span several orders of magnitude and a linear violin plot would obscure and . The dominant coordinate shows that nearly every optimized two-qubit gate collapses onto the first axis of the Weyl chamber, i.e. an -like gate, consistent with the pure coupling of the Ising Hamiltonian.
(b) Heisenberg XXZ model [Eq. (11)] with , , , . Violins show the full distribution of each Weyl coordinate over all optimized two-qubit gates (mirrored about the central axis, per the usual violin-plot convention), with horizontal lines marking the mean, median, maximum, and minimum. The near-equality reflects the U(1) symmetry of the XXZ interaction (equal and couplings ), while (dashed line, ) directly encodes the coupling ratio between the and terms of the Hamiltonian.
FIG. 3: Weyl-chamber coordinates () distribution of the RO-optimized two-qubit gates for sites, extracted via the KAK decomposition, for two Trotterized target Hamiltonians. Both circuits use the same initial ansatz: 7 repetitions of a second-order Trotter product formula. The gate content reflects the symmetry of the underlying Hamiltonian: Ising optimization converges to near- gates (a), while Heisenberg XXZ optimization converges to near-fSim gates (b).
Heisenberg XXZ model: A similar phenomenon occurs in the Heisenberg model. After optimization on the arbitrary manifold, the Weyl spectrum demonstrates that the gates strongly favor a local XXZ structure, (Figure 3b). Such gates are locally equivalent to a single fSim gate [29, 30], so in Section IV B we optimize parameterized fSim gates directly, rather than synthesizing each generic gate to an fSim and an gate.
These observations suggest that, for these models, the optimized gates occupy only a small subset of the full 15-parameter manifold. This motivates our core hypothesis: instead of optimizing over the full manifold and then paying the full cost of synthesizing each gate into native operations, we can optimize directly over native gates, with significantly fewer degrees of freedom, without a substantial loss in simulation accuracy. Such a hardware-native ansatz avoids the synthesis overhead altogether. Furthermore, because each hardware-native Trotter step is shallower than its synthesized counterpart (Figure 2b), more steps fit within a fixed native entangling-gate depth. The ansatz can therefore be initialized from a Trotter circuit with a smaller time step, and hence a lower Trotter error.
Methods
We compare two distinct TN-based approaches for compressing 1D time-evolution circuits: standard RO and PO. Both approaches rely on the same underlying computational engine, namely efficient TN contraction to obtain loss gradients, but differ in the initial ansatz (as seen in Figure 2) over which the two-qubit gates are optimized. Given a circuit composed of gates , the goal in both approaches is to minimize the same loss between and a reference time-evolution unitary . We compare the resulting approximation error at fixed native entangling depth, including the synthesis of generic two-qubit gates in the RO circuits.
Riemannian optimization
As a baseline, we use the TN RO framework of Ref. [5], which optimizes generic two-qubit unitaries in a brick-wall circuit directly on the unitary manifold. The reference propagator is represented as an MPO, and the gate-matrix gradient is obtained by contracting the environment of each gate . This Euclidean gradient is then projected onto the tangent space of to give the Riemannian gradient, and the updated gate is mapped back onto the manifold by a retraction, so unitarity is preserved throughout. We use the Riemannian ADAM optimizer with a polar retraction [5]. After optimization, each two-qubit gate is synthesized into the native gate set via its KAK decomposition (Eq. (1)).
Hardware-aware parameterized optimization
Justified by the structural behaviors observed in Sec. II, PO restricts the optimization to the native gate set of the target hardware instead of allowing for generic two-qubit unitaries. To this end, the circuit is initialized in a Trotterization split natively aligned with that gate set. The optimization is performed using a hybrid scheme:
The native two-qubit gates are parameterized directly by their physical control parameters (e.g., the angles of an or fSim gate) and updated via Euclidean gradient descent. The loss gradient with respect to these parameters is computed via the chain rule:
The gradient is the same gate-matrix gradient computed through TN contraction in Ref. [5]; the chain rule reuses this gradient with the derivatives of the parameterized native gate. The product in Eq. (2) denotes contraction over the gate-matrix entries. The key difference is that the optimization proceeds directly in the (typically low-dimensional) Euclidean parameter space of the native gate, rather than being projected onto a Riemannian manifold. Because the resulting gates are already expressed in native form, no subsequent KAK synthesis is required, avoiding the associated blow-up in single- and two-qubit gate count.
Conversely, single-qubit gates are optimized directly as generic unitaries via RO. This allows the TN environment tensor to be constructed efficiently around each single-qubit unitary, rather than separately for its constituent rotations. The optimized unitary can then be readily decomposed into elementary rotations using an Euler-angle decomposition, without introducing additional two-qubit gates.
Analysis and Results
We study the transverse-longitudinal-field Ising model (TLFIM) and the XXZ transverse-field Heisenberg model in one dimension, both on chains of qubits with open boundary conditions. Neither model is integrable in the parameter regimes considered. In the Ising model, the longitudinal field breaks the spin-flip symmetry. In the XXZ model, the transverse field breaks the symmetry associated with spin rotations about the -axis. Because both models obey an area law in their ground states, matrix product states (MPS) represent them efficiently at . During time evolution, entanglement often grows approximately linearly in time. Since an MPS with bond dimension can support an entanglement entropy of at most , representing this growth requires . The required bond dimension therefore grows exponentially in time. This exponential growth in classical memory and runtime motivates compressing the time-evolution circuit for execution on quantum hardware rather than simulating it classically.
Transverse-longitudinal field Ising model
The TLFIM is governed by the Hamiltonian
We consider a chain of sites. As the reference, we use the time-evolution operator obtained from 20 repetitions of a fine fourth-order Suzuki–Trotter decomposition based on the even–odd splitting of Eq. (4) below, which on hardware would require 603 gates.
Circuit ansätze and Trotter splitting
Both methods are initialized from a second-order Trotterized brickwall circuit, but they use different splittings of the Hamiltonian. The splitting determines both the accuracy of the initial (unoptimized) circuit and the number of native layers per Trotter step. For a fair comparison, we use the best available splitting for each method.
For RO, we use the even–odd splitting of Ref. [5], in which the single-qubit field terms are divided equally between the two halves:
Of the splittings we tested, this one gives the smallest Trotter error (Figure 4, blue). It is, however, unsuitable for PO: each bond gate combines the interaction with the non-commuting field, so it is a generic two-qubit unitary rather than an rotation and would itself require synthesis.
Figure 4. Estimated spectral error for different Trotter splittings as a function of the time step (, ). Blue: splitting used for RO (Eq. (4)). Yellow and green: splittings available to PO (Eq. (8) and Eq. (5), respectively). All lines share the same slope; the offset between them grows linearly with system size and interaction strength, reflecting the number of non-commuting terms introduced by each splitting.
For PO, we therefore keep the interaction and field terms in separate layers, , with
For a time step , every two-qubit gate is then a pure rotation , with , and the field term contributes a layer of single-qubit gates. An alternative is to place the field layer between the even and odd bonds, giving
This saves one gate in the second-order Trotter circuit (Appendix, Figure 10), but introduces more non-commuting terms per step and hence a larger Trotter error (Figure 4, yellow vs. green). Because this error grows with system size, we use Eq. (5) throughout.

Figure 10. Two repetitions of second-order Trotterization with the field placed between the even and odd bonds [Eq. (8)]. The even-bond layers of adjacent Trotter steps merge into a single operation, while the half-step field layers stay separated by the odd-bond layers.
A natural question is whether RO could also benefit from the lower cost of the PO splitting by using it as the initial circuit. It cannot: during optimization, the gates evolve into generic two-qubit unitaries, each of which requires three gates after synthesis, so the native gate count would roughly triple.
Degrees of freedom
The degrees of freedom (DOF) of a circuit depend on its gate count and on the manifold dimension per gate: a generic two-qubit gate in has 15 real parameters, and a single-qubit gate in has 3.
For a second-order Trotterization with repetitions, RO produces layers of gates: even-bond layers of gates each and odd-bond layers of gates each. The total gate count is , giving
For PO, the same evolution time and order yields interaction layers ( even and odd layers of gates, with one parameter each) and layers of single-qubit rotations on all sites (three parameters each):
For and , this gives and . The parameterized ansatz therefore reaches the same DOF budget at a larger number of Trotter repetitions, as shown in Figure 5.
Figure 5. Total number of adjustable circuit parameters (degrees of freedom) versus the number of spins for Ising time-evolution circuits initialized from a second-order Trotterization. Circles represent RO; triangles represent PO. Shades indicate the number of Trotter repetitions . Both parameter counts grow linearly with system size. At fixed and , the hardware-aware circuits use fewer parameters, permitting more Trotter repetitions within the same parameter budget.
Results
Figure 6 compares RO and PO as a function of native entangling-gate depth. At low depth, PO benefits
Figure 6. Loss as a function of native entangling-gate depth for the Ising model, . RO is compared with PO. Solid lines show results after optimization; dashed lines show the corresponding initial Trotter circuits before optimization. Lower loss indicates a more accurate approximation of the reference time evolution.
from its Trotter-based initialization, since more Trotter steps fit within a fixed gate budget. However, optimizing its parameters yields only marginal gains over the unoptimized Trotterization. RO, by contrast, overtakes PO once the circuit exceeds 26 layers. This is consistent with Ref. [5], where the Ising model is likewise the case in which RO shows the clearest advantage.
We attribute the gap at higher depth to the limited expressivity of the native two-qubit gate: each gate is restricted to a one-dimensional segment of the Weyl chamber (Figure 1), whereas each generic gate in RO can access the full chamber. The residual Weyl coordinates of Figure 3a, which PO neglects, therefore turn out to be essential for high compression of the Ising propagator. Above 26 layers, PO surpasses RO at only a single point in the tested range; we attribute this to the sensitivity of RO to the quality of its initialization, which leaves it prone to local minima, an effect we revisit in Section IV B.
The XXZ transverse-field Heisenberg model
The XXZ transverse-field Heisenberg model is given by
with , the anisotropic exchange couplings and the transverse-field strength.
Native gate and Trotterization
A continuous fSim gate is native to superconducting circuits with two transmons coupled via a frequency-tunable coupler [30], and has also been proposed via a single set of flux pulses [29]:
As derived in Appendix B, this gate is equivalent (up to local rotations) to the XXZ interaction,
with , .
By combining a continuous fSim gate with an pulse, any arbitrary unitary can be synthesized (up to local gates) using only two entangling operations; we adopt this synthesis for the RO circuits. For PO, by contrast, we leverage the native XXZ (fSim) structure directly at initialization, splitting the Hamiltonian into even-bond, odd-bond, and transverse-field terms, with the constituent sub-Hamiltonians defined as:
The Trotterization for RO mirrors the Ising case, distributing the transverse field terms symmetrically across both the even and odd bonds, where half of the local transverse field terms are absorbed directly into the two-body bond interactions:
The performance of the two compression methods is compared in Figure 7. In this Heisenberg setting, the depth reduction achieved by the hardware-aware approach is less pronounced than for the Ising model, since matching a generic two-qubit unitary now requires two native gates (an fSim and an ) rather than three. For depths up to layers, PO attains fidelities roughly half an order of magnitude better than RO at equal depth. This advantage is inherited from both the initialization and the optimization: the hardware-aware decomposition realizes each second-order Trotter step in fewer native layers, so
Figure 7. Loss as a function of native entangling gate depth for the Heisenberg XXZ model, . Riemannian optimization (RO) is compared with hardware-aware parameterized optimization (PO). Each layer consists of native two-qubit gates: the RO circuits alternate fSim and layers, while the PO circuits use only fSim layers. Solid lines show results after optimization; dashed lines show the corresponding initial Trotter circuits before optimization. Lower loss indicates a more accurate approximation to the reference time evolution.
at a fixed layer budget the PO ansatz encodes a finer Trotter discretization and starts from a lower initial error. Furthermore, the fSim gate is more expressive and entangling than the gate, since it reaches an entire face of the Weyl chamber (Figure 1) rather than a single line.
The RO circuit (up to single-qubit transformations) depends heavily on the quality of its initial state and tends to become trapped in local minima, as reflected in the loss plateau seen in Figure 12. Whether this loss can be further reduced remains open; alternative optimization strategies, e.g. noise injection, may help escape these plateaus. The PO circuit, by contrast, starts from a better initial state, and its gates are expressive enough to reach a higher fidelity early on. Beyond 35 layers, however, our hardware-aware training grows harder and longer, due to a higher number of tensor contractions, as explained in Section D) relative to RO.
Figure 12. Riemannian Adam Loss History for a XXZ Heisenberg system with 50 sites and initialization to 5 repetitions of second-order Trotterization.
Conclusion
We compared two TN-based compression strategies for Trotterized time-evolution circuits on a 50-qubit chain: standard RO over generic unitaries, and the hardware-aware PO method that optimizes directly over the native two-qubit gate set. The comparison was motivated by our observation that RO-optimized gates already collapse into a predictable region of the Weyl chamber: near- gates for the Ising model, and a local XXZ/fSim-like structure for the Heisenberg model. This structural regularity suggested that a shallower, hardware-native ansatz could recover most of RO’s compression benefit without the synthesis overhead of KAK decomposition.
The two methods perform very differently depending on the target Hamiltonian. For the Ising case, RO clearly outperforms PO, particularly beyond layers. We attribute this gap to the expressivity of each two-qubit block: an RO gate can be any element of , realized by three gates on the same pair of qubits, whereas in PO each pair receives a single per layer, restricted to one line of the Weyl chamber (Figure 1). Since a fixed-depth circuit can reach only a subset of the unitaries generated by its gates [31], this per-block restriction directly limits PO. It matters even though the optimized RO gates are close to (Figure 3a), since their small and components cannot be produced by a single . For the Heisenberg XXZ model, the outcome is reversed and, at first glance, surprising: PO reaches a loss roughly half an order of magnitude lower than RO for depths up to layers, despite RO being the more general method. We attribute this to two compounding effects: (i) the hardware-aware Trotter splitting realizes each second-order step in fewer native layers, giving PO a better-conditioned initial circuit at a fixed layer budget, and (ii) the fSim gate, generated by and (up to single-qubit phases), spans a two-parameter family covering an entire face of the Weyl chamber. Each PO block therefore explores a substantially larger subgroup of than an gate, which gives PO enough expressivity to exploit its better starting point. RO’s comparative weakness in this regime appears to stem from sensitivity to initialization and susceptibility to local minima. Beyond approximately 35 layers, PO’s advantage diminishes as optimization yields less improvement over the Trotter initialization. A possible explanation is correlated angle directions in finer Trotter decompositions. Establishing the mechanism requires further analysis of gradients and curvature, which we leave for future work.
Beyond raw fidelity, PO carries a substantial practical advantage in degrees of freedom: for and Trotter repetitions, PO requires roughly a quarter of the parameters of RO (1044 vs. 4050), which in turn permits more repetitions – and thus finer Trotterizations – within the same parameter budget. However, this comes at a computational cost: TN contraction scales as for two-qubit gates, and for single-qubit gates, and since PO retains Riemannian updates for its single-qubit gates, its per-step cost remains higher than RO’s, as detailed in Section D.
Taken together, these results indicate that neither method dominates uniformly: RO’s generality pays off when the native entangling gate covers only a small region of the Weyl chamber, as does in the Ising case, whereas PO’s leaner, hardware-native parametrization is preferable when the native gate already covers the region relevant to the problem, as fSim does in the Heisenberg case, at least up to depths of layers. More broadly, this suggests that exploiting the target hardware’s native gate set – and, in future work, its qubit connectivity – is a promising direction for compressing circuits on both near-term and early fault-tolerant quantum computers.
This work focused on non-integrable models with no conserved observables of interest, and Trotterization itself breaks whatever symmetries the target Hamiltonian possesses. An open direction is to study how RO and PO compare when conserved quantities are relevant to the observables of interest, which likely calls for a different initialization strategy. Extending this comparison to fermionic systems is another natural avenue for future work. Under the Jordan–Wigner transformation, fermionic hopping and density–density interactions map directly onto the fSim gate, which forms the basis of efficient fermionic swap networks [32]. This makes fermionic models a promising testbed for PO.
Acknowledgments
All tensor network operations were implemented using jax.numpy [33]. We thank Chinonso Onah, Roman Firt and Arne-Christian Voigt for valuable discussions over quantum native gates and support while running large scale experiments. The results, opinions and conclusions expressed in this publication are not necessarily those of Volkswagen AG.
Data availability
The implementation used in this project, as well as the numerical results and a script to generate the figures in this publication, are available under https://github.com/ixiaxim/weyl-circuit-compression.
Ising Trotterization
For the second-order Ising Trotterization with the splitting of Eq. (5), the two-qubit depth scales as when the interaction layers are placed outside (Figure 8) and as when the field layers are placed outside (Figure 9). Placing the field between the even and odd bonds instead, as in Eq. (8), gives (Figure 10).

Figure 8. Two repetitions of the Interaction-Outside Trotter sequence for the splitting [Eq. (5)]. In the center, the interaction layers from both Trotter steps merge into a single operation.

Figure 9. Two repetitions of Interaction-Inside (Field-Outside) Trotterization for the same splitting [Eq. (5)]. Note the merged central layer.
The Heisenberg Model
In the following, we prove:

Figure 11. Equivalence with explicit parameter mapping. To implement fSim(, ), one evolves the XXZ Hamiltonian for and , followed by single-qubit rotations (grey).
Proof
To prove that the entangling layer is locally equivalent to the fSim gate, we begin by defining the rotation matrices , , and . These derivations rely on the property that for any Pauli tensor product , . Using the identity , we can express the rotations for an angle .
First, we consider the matrices , , and :
The rotation matrices for these operators are given by:
For the combined interaction, the evolution yields:
Next, we write the evolution explicitly:
Finally, combining these terms, we obtain the full entangling operation:
To transform the derived entangling operation into the standard fSim gate, we apply local single-qubit -rotations to both qubits. We define the local transformation as:
Since -rotations commute with both and interactions, this local operation can be applied either before or after .
a. Step-by-Step Verification Using the standard definition , the local operator is:
Multiplying this local gate with the previously derived matrix :
By factoring out the global phase , we obtain:
This result is identical to the standard fSim gate matrix:
where the parameters are mapped as:
In practical hardware implementations, such as superconducting circuits, these single-qubit rotations are implemented as virtual phase updates at zero cost, making the entangling layer locally equivalent to the fSim gate.
Optimization
Figure 12 shows the loss history of Riemannian Adam for the XXZ Heisenberg model with 50 sites, initialized with 5 repetitions of second-order Trotterization. After a rapid initial decrease, the loss levels off into a plateau. Such plateaus are indicative of local minima of the Riemannian loss function.

Figure 13. Contraction scheme for a two-qubit gate , contracted with a left environment and the corresponding MPO tensors , , , of the sites on which the gate acts.
Contraction cost
In this appendix we detail the computational cost of contracting the two elementary building blocks of a Trotter layer – single- and two-qubit gates – into the TN used to evaluate the loss function and its gradient during optimization. Namely, we are showing the cost of updating a left environment during a sweep with the tensor contraction presented in [5]. Because a PO circuit interleaves single-qubit gates between the two-qubit entangling gates, while an RO circuit is built exclusively from two-qubit gates, the TN underlying a PO layer contains additional tensors to contract relative to an RO layer of the same width, as illustrated in Figure 2. Throughout, denotes the bond dimension of the boundary environment (and of the MPO tensors , , , below), while denotes the local dimension of a gate leg ( for a single qubit);
Two-qubit gate
Figure 13 shows the TN encountered when contracting a two-qubit gate , acting on sites , , against the accumulated left environment and the four neighboring MPO tensors , (upper/bra environment) and , (lower/ket environment). The environment carries the two bond-dimension- legs , inherited from the previously contracted sites.
We contract this network sequentially into an updated environment , absorbing one tensor at a time in the order . Table I lists, for each step, the indices summed over and the resulting cost, computed as the product of the dimensions of every leg touched (summed or left open) in that step.
| Step | Absorb | Contract | Cost |
| 1 | |||
| 2 | |||
| 3 | |||
| 4 | |||
| 5 |
Table I. Cost of each step in the sequential contraction of a two-qubit gate.
Since the gate leg dimension is small compared to the bond dimensions used in this work (), the two steps absorbing and dominate, and contracting a single two-qubit gate into the environment scales as
Single-qubit gate
The corresponding network for a single-qubit gate is shown in Figure 14: the gate acts on a single upper tensor and lower tensor , each connected to by a -bond and to each other by a direct -dimensional bond . Absorbing , then the gate, then (Table II) updates the environment to .

Figure 14. Contraction scheme for a single-qubit gate , contracted with a left environment , an upper tensor , and a lower tensor from the respective upper and lower environments.
| Step | Absorb | Contract | Cost |
| 1 | |||
| 2 | |||
| 3 |
Table 2. Cost of each step in the sequential contraction of a single-qubit gate.
Here the first and last steps dominate, so contracting a single-qubit gate into the environment scales as
These per-gate costs directly account for the runtime gap discussed in Sec. IV: on top of the cost of the two-qubit gates common to both methods, each PO Trotter layer additionally requires the contraction of single-qubit gates, retained as Riemannian updates, which RO does not incur.
References
- [4]simulation, J. Phys. A: Math. Theor. 57, 135303 (2024).
- [5]I. N. M. Le, S. Sun, and C. B. Mendl, Riemannian quantum circuit optimization based on matrix product operators, Quantum 9, 1833 (2025).
- [6]Y. Zhang, R. Wiersema, J. Carrasquilla, L. Cincio, and Y. B. Kim, Scalable quantum dynamics compilation via quantum machine learning, arXiv preprint arXiv:2409.16346 (2024).arxiv.org/abs/2409.16346
- [7]J. Gibbs and L. Cincio, Deep circuit compression for quantum dynamics via tensor networks, Quantum 9, 1789 (2025).
- [8]J. Gibbs and L. Cincio, Learning circuits with infinite tensor networks, arXiv preprint arXiv:2506.02105 (2025).arxiv.org/abs/2506.02105
- [9]Y. Guo and S. Yang, Efficient quantum circuit compilation for near-term quantum advantage, EPJ Quantum Technol. 12, 69 (2025).
- [10]M. Dinca, D. J. Luitz, and M. Debertolis, Quantum process tomography of a compressed time evolution circuit on superconducting quantum processors, arXiv preprint arXiv:2509.25342 (2025).arxiv.org/abs/2509.25342
- [11]R. Mansuroglu, T. Eckstein, L. Nützel, S. A. Wilkinson, and M. J. Hartmann, Variational Hamiltonian simulation for translational invariant systems via classical pre-processing, Quantum Sci. Technol. 8, 025006 (2023).
- [12]C. Mc Keever and M. Lubasch, Classically optimized Hamiltonian simulation, Phys. Rev. Res. 5, 023146 (2023).
- [13]N. Robertson, A. Akhriev, J. Vala, and S. Zhuk, Approximate quantum compiling for quantum simulation: A tensor network based approach, ACM Trans. Quantum Comput. 6, 20 (2025).
- [14]F. Putterer, M. M. Zumpe, I. N. M. Le, Q. Huang, and C. B. Mendl, High-performance contraction of quantum circuits for Riemannian optimization, arXiv preprint arXiv:2506.23775 (2025).arxiv.org/abs/2506.23775
- [15]E. Karacan, I. N. M. Le, M. D’Anna, J. Carasquilla, C. B. Mendl, and I. Rojkov, Near-optimal quantum time evolution circuits via provably convergent compression, arXiv preprint arXiv:2605.17067 (2026).arxiv.org/abs/2605.17067
- [16]I. N. M. Le, R. Wiersema, and C. B. Mendl, Hessian-vector products for tensor networks via recursive tangent-state propagation, arXiv preprint arXiv:2604.20384 (2026).arxiv.org/abs/2604.20384
- [17]H.-T. Hung, I. N. M. Le, J. Knolle, and Y.-J. Kao, Improved Ising meson spectroscopy simulation on a noisy digital quantum device, arXiv preprint arXiv:2512.02516 (2025).arxiv.org/abs/2512.02516
- [18]Y. Chai, J. Gibbs, V. R. Pascuzzi, Z. Holmes, S. Kühn, F. Tacchino, and I. Tavernelli, Resource-efficient simulations of particle scattering on a digital quantum computer, arXiv preprint arXiv:2507.17832 (2025).arxiv.org/abs/2507.17832
- [19]F. Vatan and C. Williams, Optimal quantum circuits for general two-qubit gates, Phys. Rev. A 69, 032315 (2004).DOI
- [20]G. Vidal and C. M. Dawson, Universal quantum circuit for two-qubit transformations with three controlled-NOT gates, Phys. Rev. A 69, 010301(R) (2004).DOI
- [21]V. V. Shende, I. L. Markov, and S. S. Bullock, Minimal universal two-qubit controlled-NOT-based circuits, Phys. Rev. A 69, 062321 (2004).DOI
- [22]B. Kraus and J. I. Cirac, Optimal creation of entanglement using a two-qubit gate, Phys. Rev. A 63, 062309 (2001).DOI
- [23]N. Khaneja and S. J. Glaser, Cartan decomposition of SU(2ⁿ) and control of spin systems, Chem. Phys. 267, 11 (2001).DOI
- [24]J. Zhang, J. Vala, S. Sastry, and K. B. Whaley, Minimum construction of two-qubit quantum operations, Phys. Rev. Lett. 93, 020502 (2004).pubmed.ncbi.nlm.nih.gov/15323888
- [25]A. Wu, J. Leng, and M. Guo, Design the quantum instrument set with the Cartan coordinate analysis framework, arXiv preprint arXiv:2410.04008 (2025).arxiv.org/abs/2410.04008
- [26]Y. Makhlin, Nonlocal properties of two-qubit gates and mixed states, and the optimization of quantum computations, Quantum Inf. Process. 1, 243 (2002).DOI
- [27]E. Y. Egorova, A. S. Kazmina, I. A. Simakov, I. N. Moskalenko, N. N. Abramov, D. A. Kalacheva, V. B. Lusbanov, A. N. Bolgar, N. Maleeva, and I. S. Besedin, Three-mode tunable coupler for superconducting two-qubit gates, Phys. Rev. Appl. 23, 064056 (2025).arxiv.org/abs/2405.10886
- [28]Quantinuum Systems, Native arbitrary angle hardware gates, https://docs.quantinuum.com/systems/trainings/getting_started/arbitrary-angle_2-qubit_gates.html, accessed 2026.
- [29]B. Foxen, C. Neill, A. Dunsworth, P. Roushan, B. Chiaro, A. Megrant, J. Kelly, Z. Chen, K. Satzinger, R. Barends, F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, S. Boixo, D. Buell, B. Burkett, Y. Chen, R. Collins, E. Farhi, A. Fowler, C. Gidney, M. Giustina, R. Graff, M. Harrigan, T. Huang, S. V. Isakov, E. Jeffrey, Z. Jiang, D. Kafri, K. Kecedzhi, P. Klimov, A. Korotkov, F. Kostritsa, D. Landhuis, E. Lucero, J. McClean, M. McEwen, X. Mi, M. Mohseni, J. Y. Mutus, O. Naa-man, M. Neeley, M. Niu, A. Petukhov, C. Quintana, N. Rubin, D. Sank, V. Smelyanskiy, A. Vainsencher, T. C. White, Z. Yao, P. Yeh, A. Zalcman, H. Neven, and J. M. Martinis (Google AI Quantum), Demonstrating a continuous set of two-qubit gates for near-term quantum algorithms, Phys. Rev. Lett. 125, 120504 (2020).
- [30]Z. Jiang and M. H. Ansari, Concurrent fermionic simulation gate, Phys. Rev. Appl. 24, 024050 (2025).
- [31]A. Bärligea, M. L. Sims-Goh, and J. S. Kottmann, Enabling Lie-algebraic classical simulation beyond free fermions, arXiv preprint arXiv:2604.16701 (2026).arxiv.org/abs/2604.16701
- [32]I. D. Kivlichan, J. McClean, N. Wiebe, C. Gidney, A. Aspuru-Guzik, G. K.-L. Chan, and R. Babbush, Quantum simulation of electronic structure with linear depth and connectivity, Phys. Rev. Lett. 120, 110501 (2018).DOI
- [33]J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, Y. Katariya, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne, and Q. Zhang, JAX: composable transformations of Python+NumPy programs (2018).




