⚛️ quantum physics

Scaling Laws for Parameter Setting in Digitized Counterdiabatic QAOA

This paper proposes empirical scaling laws for parameter setting in digitized counterdiabatic QAOA that enable direct, optimization-free angle assignment based on graph properties, significantly outperforming fixed-angle QAOA in simulations and improving approximation ratios in both ideal and hardware experiments.

Original author: Yaocheng Chen, Jiapei Zhuang

Published 2026-11-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original author: Yaocheng Chen, Jiapei Zhuang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast landscape of modern computing, some of the most stubborn challenges involve finding the single best arrangement among a dizzying number of possibilities. Whether it is routing delivery trucks through a city, scheduling factory machines, or settling financial transactions, these are problems of combinatorial optimization. For decades, scientists have looked to quantum computers—machines that use the strange rules of the subatomic world—to solve these puzzles faster than classical machines ever could. One of the most promising tools for this job is an algorithm called the Quantum Approximate Optimization Algorithm, or QAOA. Think of it as a guided search: the computer prepares a quantum state, applies a series of operations controlled by specific numbers called parameters, and then measures the result to see how close it came to the perfect solution. The catch is that for every new problem, the computer usually needs to run a long, slow process to figure out the best numbers to use. This "tuning" step eats up time and resources, often canceling out the speed advantage the quantum machine was supposed to provide.

Researchers Yaocheng Chen and Jiapei Zhuang have found a way to bypass this bottleneck for a specific type of quantum algorithm designed to handle weighted problems, where different connections between items have different strengths. They discovered that the best settings for the algorithm's controls are not random or unique to every single problem. Instead, these settings follow a predictable pattern based on two simple facts about the problem itself: how many connections the items have on average, and how strong those connections are. By studying thousands of different problem setups in simulation, the team built a set of rules that can predict the optimal settings for a new problem just by looking at its basic statistics. This means a quantum computer can be handed a fresh, unsolved problem and immediately given the right instructions to start searching, without needing to spend time figuring out what those instructions should be.

The team focused on a sophisticated version of the standard algorithm known as digitized counterdiabatic QAOA. While the standard version uses two types of operations to shuffle through possible answers, this enhanced version adds a third type of operation designed to keep the system from getting stuck or making mistakes as it evolves. This extra step makes the algorithm more powerful but also more complex, requiring three sets of numbers to be tuned for every layer of the process. The researchers began by running massive offline simulations on a wide variety of graph structures—mathematical maps of connections—ranging from simple rings to complex, random networks. They optimized the settings for each of these thousands of graphs to find the absolute best performance.

What emerged from this mountain of data was a striking regularity. When the researchers adjusted for the size of the connections and the average number of links per item, the optimal settings for all the different graphs collapsed into a tight cluster. It was as if the specific details of each graph didn't matter as much as the overall scale of the problem. The team realized they could describe the best settings using a simple formula that takes the average number of connections and the strength of the links as inputs. This formula acts like a universal key: once you know the basic statistics of a new problem, you can plug them in and get the exact numbers needed to run the quantum circuit. The result is a "frozen" prescription, a fixed set of rules that assigns the correct parameters instantly, skipping the lengthy search process entirely.

To test if this shortcut actually worked, the researchers ran simulations on 1,050 different test graphs that the algorithm had never seen before. They compared their formula-based approach against a standard method that uses a fixed set of numbers for everyone. In every case, from shallow circuits with just one layer of operations to deeper ones with four layers, the formula-based method yielded a higher mean probability of sampling the optimal solution than the standard fixed approach. At a depth of four layers, the mean optimal-solution probability for the new method reached 19.36%, compared to 13.11% for the standard fixed approach. This advantage held true across different types of networks and different ways of assigning connection strengths. The researchers even tested the method on a real quantum processor, the IBM ibm_fez, using small, uniform networks. While the noise in the real hardware complicated the results at deeper levels, the formula-based method still showed a clear advantage at the shallowest level, proving that the rules derived from simulations could translate to physical machines.

The study also explored whether these pre-calculated settings could serve as a good starting point for further refinement. In a separate experiment, the researchers took the settings provided by their formula and used them to initialize a standard optimization search on a small set of graphs. They found that starting with the formula's numbers allowed the search to reach a better solution much faster than starting from a random guess. In fact, the formula-based start achieved a higher average quality of solution than a random search could reach even after 500 attempts. This suggests that the formula doesn't just replace the search; it provides a high-quality launchpad that makes any subsequent fine-tuning more effective.

However, the researchers were careful to note where their method might stumble. The rules were derived from graphs with a moderate number of connections, and when they tested the formula on graphs with very few connections, the performance dropped, especially in deeper circuits. This indicates that while the scaling laws are powerful, they have limits and should not be applied blindly to every possible type of network. Furthermore, the comparison was made layer-for-layer, meaning the enhanced algorithm used more physical operations per step than the standard one. The researchers acknowledged that a fairer comparison would need to account for the total number of operations, a question that remains open for future study.

Ultimately, this work demonstrates that the behavior of complex quantum algorithms can be tamed by understanding the simple geometry of the problems they solve. By recognizing that the best settings depend on the scale of the network rather than its specific details, the researchers have turned a difficult, instance-by-instance tuning problem into a straightforward calculation. This shift allows quantum computers to spend their precious time solving problems rather than learning how to solve them, bringing the promise of quantum optimization one step closer to practical reality. The findings suggest that for a wide range of combinatorial problems, the path to the best answer is not a mystery to be solved from scratch each time, but a pattern that can be read directly from the problem's own structure.

Technical Summary: Scaling Laws for Parameter Setting in Digitized Counterdiabatic QAOA

Problem Statement
Variational Quantum Algorithms (VQAs), such as the Quantum Approximate Optimization Algorithm (QAOA), typically require optimizing variational parameters for every new problem instance. This process involves repeated evaluations of the quantum objective function, creating a significant bottleneck that limits the practical reuse of these algorithms across different graphs. While Digitized Counterdiabatic QAOA (DC-QAOA) offers improved solution quality over standard QAOA at fixed circuit depths by incorporating counterdiabatic driving terms, it introduces additional parameters (counterdiabatic angles $\alpha_l$ ) that further complicate the optimization landscape. The authors address the challenge of transferring optimized parameters across different weighted MaxCut instances without performing target-instance variational optimization.

Methodology
The authors propose a "frozen prescription" for setting DC-QAOA parameters based on empirical scaling laws derived from extensive offline optimization. The methodology involves three main components:

Circuit Definition: The study utilizes a $p$ -layer DC-QAOA circuit for weighted MaxCut. Each layer $l$ consists of a cost unitary ( $e^{-i\gamma_l H_C}$ ), a counterdiabatic unitary ( $e^{-i\alpha_l H_{CD}}$ ), and a mixing unitary ( $e^{-i\beta_l H_B}$ ). The counterdiabatic term $H_{CD}$ is constructed from commutators of the cost and mixer Hamiltonians, implemented as ordered gate products of $Y_iZ_j$ and $Z_iY_j$ terms.
Scaling Laws and Parameter Assignment: The core contribution is a formula-based assignment rule that determines the angles $\gamma_l, \beta_l, \alpha_l$ $γ_{l}, β_{l}, α_{l}$ using only classical graph statistics: the mean degree $D(G)$ $D (G)$ and the root-mean-square (RMS) edge weight $s_J(G)$ $s_{J} (G)$ .
- Weight Scaling: The prescription maintains exact covariance under uniform edge-weight rescaling. Angles scale inversely with the RMS weight ( $1/s_J$ ).
- Degree Scaling: The authors fit empirical power laws to the cost ( $\gamma$ ) and counterdiabatic ( $\alpha$ ) angles based on offline optimization data. The angles follow the form:
  $\theta_l(G) = A_{l,p} \frac{s_J(G)}{(D(G)/D_0)^{b_{l,p}}}$
  where $A_{l,p}$ and $b_{l,p}$ are depth- and layer-specific coefficients fitted from training data, and $D_0$ is a reference degree.
- Mixer Angles: The mixer angles $\beta_l$ are retained as fixed values from the offline schedule and are not rescaled.
Training and Validation: The coefficients were derived from 7,000 offline-optimized DC-QAOA vectors across various graph families (regular, Erdős–Rényi, tree, bipartite, etc.) and weight distributions. The resulting "frozen" coefficients are applied to a disjoint test set of 1,050 graphs (spanning 10–22 qubits) without any further optimization on the target instances.

Key Contributions

Parameter Scaling Laws: The authors combine exact covariance under uniform edge-weight rescaling with empirical, depth-specific degree power laws for cost and counterdiabatic angles.
Compact Frozen Prescription: They demonstrate that after removing weight and degree scales, optimized angles from diverse graphs cluster around common amplitudes. This allows for a compact table of coefficients that assigns all three angle families ( $\gamma, \beta, \alpha$ ) to a new graph using only its mean degree and RMS weight, eliminating the need for target-instance variational optimization.
Comprehensive Evaluation: The study evaluates this direct assignment strategy against fixed-angle QAOA comparators and variational optimization baselines across ideal-circuit simulations, hardware execution, and energy-based refinement experiments.

Results

Ideal-Circuit Simulations: On a test set of 1,050 graphs at depths $p=1$ to $4$, the frozen power-law DC-QAOA prescription achieved a higher mean optimal-solution probability than validation-selected fixed-angle QAOA at every depth. At $p=4$ , the mean probability was 19.36% for DC-QAOA versus 13.11% for QAOA. This advantage persisted and increased at $p=6$ on a subset of 600 graphs (36.81% vs. 26.69%).
Hardware Execution: On the IBM Quantum processor ibm_fez using 21 unit-weight 3-regular graphs, the power-law DC-QAOA showed a mean probability advantage at $p=1$ (1.61% vs. 0.76%). However, at $p=2$ , the ordering reversed (1.25% vs. 1.61%), which the authors attribute to the higher two-qubit gate count in DC-QAOA layers exacerbating hardware noise.
Optimization Cost and Refinement: In energy-based experiments, the formula-assigned angles achieved a mean approximation ratio of 0.8921 without any target-instance objective evaluations. This outperformed random-start variational optimization (L-BFGS-B, COBYLA, Adam) within a budget of 500 evaluations, where the best random-start ratios remained below 0.71. Furthermore, using the formula angles as an initialization for local refinement improved the mean approximation ratio to 0.9010 with minimal additional computational cost.
Limitations: The advantage varies by graph structure and weight distribution. While positive on average, specific instances (e.g., bipartite graphs at high depth) showed losses. Additionally, the scaling law, fitted on graphs with mean degree $D > 1$ , showed negative performance differences for low-degree graphs ( $D \le 1$ ) at higher depths.

Significance and Claims
The paper claims that a scaling-based representation of optimized angles can effectively guide parameter reuse across graph instances. The primary significance lies in enabling direct parameter assignment without the computational overhead of variational optimization on the target instance, while still providing high-quality initial angles for local refinement. The authors position this as a practical step toward making DC-QAOA more scalable and reusable. They modestly note that the findings are specific to the tested graph families and depths, and that performance at matched gate budgets (rather than matched layer counts) and on broader hardware configurations requires further evaluation. The work does not claim to solve the general parameter transfer problem for all quantum algorithms but provides a specific, empirically validated prescription for DC-QAOA on weighted MaxCut.

More like this