Enhancing the reachability of variational quantum algorithms via input-state design

Background & Academic Lineage

The Origin & Academic Lineage

The problem addressed in this paper originates from the burgeoning field of quantum computing, specifically within the context of Noisy Intermediate-Scale Quantum (NISQ) devices. Quantum computers hold immense promise for solving problems intractable for classical machines, but current NISQ hardware is limited by imperfections in initialization, operation, and readout. A central challenge for these near-term devices is to demonstrate a "quantum advantage" – that is, to solve practical problems more efficiently than classical computers.

Variational Quantum Algorithms (VQAs) emerged as a leading approach to tackle this challenge. First introduced around 2014 with the Variational Quantum Eigensolver (VQE) [20, 104, 105], VQAs frame complex computational tasks as optimization problems. The core idea is to approximate a desired quantum state, often the ground state of a physical system's Hamiltonian $H$, by minimizing a cost function $E(\theta)$. This is achieved by preparing a parameterized quantum state, known as an "ansatz state" $|\Psi(\theta)\rangle = U(\theta)|\Psi_0\rangle$. Here, $U(\theta)$ is a quantum circuit with tunable parameters $\theta$, and $|\Psi_0\rangle$ is an initial input state, traditionally a simple product state like $|0\rangle^{\otimes n}$. A classical optimizer then iteratively adjusts $\theta$ to minimize $E(\theta)$.

The fundamental limitation, or "pain point," of previous VQA approaches is a trade-off between the "expressivity" (the range of states a circuit can produce) and "trainability" (how easily the parameters can be optimized). Deeper, more expressive circuits can theoretically reach a wider array of quantum states, but they are highly susceptible to noise accumulation and the notorious "barren plateau" problem, where the optimization landscape becomes extremely flat, making training impossible. Conversely, shallow circuits are more trainable and less prone to noise, but they often have an "insufficient reachable set" – meaning the target quantum state $|\Psi_{\text{tar}}\rangle$ simply isn't among the states that the circuit can produce, no matter how perfectly the parameters $\theta$ are tuned.

Historically, most research efforts to improve VQAs have focused almost exclusively on designing better ansatz circuits $U(\theta)$. Strategies included hardware-efficient ansatz (HEA), Hamiltonian variational ansatz (HVA), and adaptive circuits, among others. However, the role of the initial input state $|\Psi_0\rangle$ in determining the reachable set has recieved comparatively little attention. This oversight meant that even with a well-designed circuit $U(\theta)$, if the target state was outside the reachable set defined by the fixed input state, the algorithm would inevitably converge to a suboptimal approximation. This paper aims to address this by focusing on input-state design, thereby enhancing the reachable set without increasing circuit depth or parameter count, offering a powerful complement to existing circuit design strategies.

Intuitive Domain Terms

1. Variational Quantum Algorithms (VQAs): Imagine you're trying to find the best recipe for a complex dish using a new, somewhat finicky oven. A VQA is like a "smart trial-and-error" process: you start with a basic recipe (quantum circuit with initial settings), bake it, taste the result (measure the cost function), and then adjust the recipe's settings (circuit parameters) based on the taste. You repeat this cycle until the dish is as perfect as your oven allows.
2. Noisy Intermediate-Scale Quantum (NISQ) Devices: These are like early prototypes of a super-fast, complex calculator. They can perform amazing calculations that regular calculators can't, but they're still a bit buggy, make small errors, and can only handle a limited number of steps before the errors pile up. These are the quantum computers we have available today, before we have fully error-corrected, "fault-tolerant" machines.
3. Barren Plateaus: Picture yourself lost in a vast, flat desert, trying to find the lowest point. Everywhere you look, the terrain seems perfectly level, making it impossible to tell which direction leads downhill. In VQAs, this describes an optimization problem where the "landscape" of possible solutions becomes so flat that the classical optimizer can't find any gradient to follow, effectively getting stuck and preventing the algorithm from finding the best parameters.
4. Reachable Set: Think of this as the "menu" of all possible outcomes a specific cooking process can produce, given a fixed set of ingredients and a particular cooking method. If the desired outcome (e.g., a perfectly baked cake) isn't on that menu, you can't make it, no matter how much you tweak the oven's temperature or timing. In VQAs, it's the collection of all quantum states that a given quantum circuit can generate.
5. Ansatz: This is the "template" or "blueprint" for a quantum circuit. It defines the general structure of quantum gates and operations, but with adjustable "knobs" (parameters) that can be tuned. Different ansatz designs are like different types of engines – some are more efficient for certain tasks, some are more powerful, but all have adjustable parts.

Notation Table

Notation	Description
$|\Psi_{\text{tar}}\rangle$	The target quantum state, which the VQA aims to approximate (e.g., the ground state of a Hamiltonian).
$H$	The observable or Hamiltonian of the quantum system, whose expectation value is minimized.
$|\Psi(\theta)\rangle$	The parameterized quantum state generated by the ansatz circuit $U(\theta)$.
$U(\theta)$	The unitary ansatz circuit, a sequence of quantum gates with tunable parameters $\theta$.
$|\Psi_0\rangle$	The initial input state to the ansatz circuit, typically a simple product state like $|0\rangle^{\otimes n}$.
$E(\theta)$	The cost function, which is minimized by a classical optimizer to find optimal parameters.
$\theta$	A vector of tunable parameters for the ansatz circuit $U(\theta)$.
$\theta_{\text{opt}}$	The optimal parameters found by minimizing the cost function $E(\theta)$.
$V(\gamma)$	The encoder circuit, an additional parameterized circuit used to prepare a designed input state.
$|\Psi_0(\gamma)\rangle$	The designed input state, a superposition of candidate states, prepared by the encoder $V(\gamma)$.
$\gamma$	A vector of tunable parameters for the encoder circuit $V(\gamma)$.
$F$	Fidelity, a measure of how close a generated quantum state is to the target state, defined as $F = |\langle\Psi|\Psi_{\text{tar}}\rangle|^2$.
$m$	The number of selected mutually orthogonal states used to construct the designed input state.
$M$	The total number of computational basis states sampled during the pre-selection stage.
$n$	The number of qubits in the quantum system.

Problem Definition & Constraints

Core Problem Formulation & The Dilemma

In the realm of Variational Quantum Algorithms (VQAs), the fundamental problem addressed by this paper revolves around the limited "reachability" of target quantum states.

Input/Current State: The starting point for a VQA is typically a parameterized quantum state $|\Psi(\theta)\rangle = U(\theta)|\Psi_0\rangle$. Here, $U(\theta)$ represents a unitary quantum circuit with tunable parameters $\theta$, and $|\Psi_0\rangle$ is a simple, often fixed, initial input state (e.g., the computational basis state $|0\rangle^{\otimes n}$). The objective is to find an optimal set of parameters $\theta_{\text{opt}}$ by minimizing a cost function, such that the resulting state $|\Psi(\theta_{\text{opt}})\rangle$ closely approximates a desired target quantum state $|\Psi_{\text{tar}}\rangle$. For instance, in the Variational Quantum Eigensolver (VQE), $|\Psi_{\text{tar}}\rangle$ is the ground state of a given Hamiltonian $H$.

Desired Endpoint/Goal State: The ultimate goal is to achieve a high fidelity approximation of the target state $|\Psi_{\text{tar}}\rangle$, meaning that $|\langle\Psi_{\text{tar}}|\Psi(\theta_{\text{opt}})\rangle|^2$ should be close to 1. This implies that the target state must be within the "reachable set" of states that the ansatz $U(\theta)$ can generate from the input state $|\Psi_0\rangle$.

Missing Link/Mathematical Gap: The critical missing link is that the target state $|\Psi_{\text{tar}}\rangle$ may not be contained within the reachable set of $U(\theta)$ when starting from a fixed, simple input state $|\Psi_0\rangle$. If $|\Psi_{\text{tar}}\rangle$ is outside this set, no amount of optimization of the circuit parameters $\theta$ will allow the VQA to reach it, leading to suboptimal results regardless of computational effort. The paper aims to bridge this gap by introducing a designed input state $|\Psi_0(\gamma)\rangle$, prepared by an additional parameterized encoder circuit $V(\gamma)$, such that the reachable set of $U(\theta)|\Psi_0(\gamma)\rangle$ is modified to include $|\Psi_{\text{tar}}\rangle$. Mathematically, the paper seeks to maximize the fidelity $F = |\langle\Psi_{\text{tar}}|U(\theta)V(\gamma)|0\rangle^{\otimes n}|^2$ by jointly optimizing $\theta$ and $\gamma$. Theorem 1 provides a rigorous foundation, showing that a linear superposition of orthogonal input states can systematically enhance the achievable fidelity.

The Dilemma: The core dilemma in VQAs, which has trapped previous researchers, is the painful trade-off between expressivity and trainability.
* Expressivity: Deeper quantum circuits (those with more layers or gates) are generally more expressive, meaning they can, in principle, reach a larger set of quantum states, including potentially the target state.
* Trainability: However, increasing circuit depth often leads to "barren plateaus," where the gradients of the cost function become exponentially small, making classical optimization extremely difficult or impossible. This prevents the algorithm from converging to the optimal parameters.
* Conversely, shallow circuits are more trainable and less susceptible to barren plateaus, but they often have insufficent reachability, meaning their limited expressivity prevents them from approximating complex target states.

Previous efforts primarily focused on designing the unitary circuit $U(\theta)$. This paper proposes to tackle the dilemma by focusing on the input state $|\Psi_0\rangle$ instead, aiming to enhance reachability without significantly increasing circuit depth or suffering from barren plateaus.

Constraints & Failure Modes

The problem of enhancing VQA reachability is insanely difficult due to several harsh, realistic constraints:

1. Hardware Memory Limits & Noise Accumulation (Physical/Computational): Near-term quantum computers (NISQ devices) have limited qubit counts, short coherence times, and are susceptible to noise. Deeper circuits, while potentially more expressive, accumulate more noise, leading to errors and making the output unreliable. This imposes a strict physical limit on the practical depth of quantum circuits that can be executed.

2. Barren Plateaus (Computational/Mathematical): This is a fundamental mathematical constraint where the cost function landscape for deep, randomly initialized VQAs becomes extremely flat. The gradients vanish exponentially with the number of qubits, making classical optimizaton ineffective and preventing convergence to the true minimum. This phenomenon severely limits the trainability of expressive circuits.

3. Limited Reachability of Fixed Input States (Mathematical): For a given ansatz $U(\theta)$ and a simple, fixed input state $|\Psi_0\rangle$ (like $|0\rangle^{\otimes n}$), the set of reachable states may be very small. If the target state $|\Psi_{\text{tar}}\rangle$ lies outside this "reachable set," no amount of parameter tuning for $U(\theta)$ can achieve the desired fidelity. The reachable set of the ansatz stat $|\Psi(\theta)\rangle$ is inherently constrained by the initial state.

4. Computational Cost of Input-State Design (Computational/Data-driven): While the proposed input-state design offers a solution, it introduces its own overheads:

   • Sampling Budget ($M$): The method requires sampling $M$ computational basis states to identify promising candidates for the input superposition. If $M$ is too large (e.g., $2^n$ for $n$ qubits), this pre-selection step becomes classically intractable, especially for larger systems. The paper aims to keep $M$ moderate.
   • Encoder Circuit Depth ($m$): The encoder $V(\gamma)$ adds to the total circuit depth. While designed to be low-depth, increasing $m$ (the number of selected basis states in the superposition) increases the complexity and gate cost of $V(\gamma)$, thus consuming more quantum resources.
   • Classical Optimization Overhead: The joint optimization of both ansatz parameters $\theta$ and encoder parameters $\gamma$ increases the classical computational cost. The paper attempts to mitigate this by restricting the number of optimization iterations for the joint training.

5. Inefficient Parameterization of Hardware-Efficient Ansätze (Algorithmic): Hardware-Efficient Ansätze (HEA) are flexible and adaptable to various hardware platforms, but they often lack problem-specific structure. This can lead to inefficient parameterization, making them susceptible to optimization challenges like barren plateaus, particularly in deeper configurations.

Figure 1. Representative variational quantum ansatz. (a) Hardware-efficient ansatz (HEA). Each layer consists of alternating single-qubit rotations Ry and Rz followed by a chain of CZ gates. The dashed box indicates one circuit layer, which is repeated p times. (b) General Hamiltonian variational ansatz (HVA). Each layer contains a product of unitaries Qq k=1 e−iθkHk, where {Hk} are problem-specific Hamiltonian terms. (c-e) Examples of HVA design for three different models. (c) For the transverse-field Ising model. An initial layer of Hadamard gates H prepares |+⟩⊗n. UZZ(θ) = e−i(θ/2) σz i σz j represents the two-qubit gate for ZZ interaction, while Rx(θ) = e−iθ σx i represents the single- qubit X-rotation. (d) For the cluster-Ising model. UZXZ(θ) = e−i(θ/2) σz i σx j σz k is a three-qubit gate, and UXX(θ) = e−i(θ/2) σx i σx j is a two-qubit gate. (e) For the Fermi-Hubbard model. The upper (lower) register encodes spin-↑(spin- ↓). On-site interactions between the two spins at site i are implemented as UZZ(θ). Hopping terms on odd and even bonds are realized by UXY (θ) = e−i(θ/2) (σx i σx i+1+σy i σy i+1)

Why This Approach

The Inevitability of the Choice

The core problem addressed by this work is the fundamental trade-off in Variational Quantum Algorithms (VQAs) between circuit expressivity and trainablity. Traditional "state-of-the-art" (SOTA) methods primarily focus on designing the unitary circuit $U(\theta)$ itself. However, as the authors clearly articulate, increasing the expressivity of $U(\theta)$ by using deeper circuits leads to significant challenges such as noise accumulation and the notorious barren plateau problem, which hinders classical optimization (Abstract, page 2; Introduction, page 4). Conversely, shallow circuits, while trainable and less susceptible to noise, often have an insufficient "reachable set" of quantum states, meaning the desired target state $|\Psi_{tar}\rangle$ might simply not be accessible, regardless of how well the circuit parameters $\theta$ are optimized (page 4, Figure 2).

The authors realized the insufficiency of these traditional circuit-centric approaches when they observed that even after extensive optimization, existing VQA ansatzes would plateau at a fidelity around 0.95. Beyond this point, merely increasing circuit depth or training iterations yielded little to no improvement due to inherent ansatz limitations or the onset of barren plateaus (page 9). This critical observation highlighted that the problem wasn't solely about how to evolve a state, but from where to start the evolution. If the target state lies outside the initial reachable set of $U(\theta)$ when starting from a standard input like $|0\rangle^{\otimes n}$, then no amount of circuit optimization can ever reach it (page 6). This realization made input-state design the only viable solution for enhancing reachability without resorting to deeper, more problematic circuits or fundamentally altering the ansatz structure. The approach became inevitable as a means to overcome expressivity bottlenecks under fixed circuit depth and structure (page 9).

Comparative Superiority

This method demonstrates qualitative superiority not by replacing existing VQA ansatzes, but by providing a powerful, complimentary framework that enhances their performance. Its structural advantage lies in its ability to reshape the reachable set of any given VQA ansatz $U(\theta)$ by designing a more suitable input state $|\Psi_0(\gamma)\rangle$, rather than solely modifying the ansatz circuit itself. This is achieved via a low-depth "encoder" circuit $V(\gamma)$ that prepares a superposition of candidate states (page 4, Figure 2).

Figure 2. Reachable sets modified through input-state design. For a fixed uni- tary U(θ), a simple input state |Ψ0⟩induces a reachable set (red-shaded) that excludes the target |Ψtar⟩, causing optimization to converge to a suboptimal state |Ψ′(θ)⟩(blue path). By contrast, a designed input state |Ψ0(γ)⟩, pre- pared by the encoder V(γ), produces a different reachable set (green-shaded) that contains |Ψtar⟩, enabling the same U(θ) to reach the target (red path)

The key advantages are:
1. Enhanced Reachability and Accuracy: By starting from a carefully designed input state, the method ensures that the target state becomes accessible to the existing $U(\theta)$ circuit. This leads to consistently higher fidelity and more accurate ground energy estimates compared to standard methods, even at the same gate budget (Abstract, page 2; page 9). For instance, in the 1D transverse-field Ising model, the method acheives 0.99 fidelity with only 8 layers, a 33% reduction in depth compared to the 12 layers required by a conventional Hardware-Efficient Ansatz (HEA) (page 4, page 10). Similar gains are observed for 2D Ising and cluster-Ising models (page 4).
2. Resource Efficiency: The input-state design significantly reduces both quantum and classical resources. It lowers the required gate count (e.g., 112 CNOT gates vs. 144 for HEA to reach 0.99 fidelity) and optimization effort (1100 steps vs. 1500 steps for HEA) (page 11). The encoder $V(\gamma)$ itself is a low-depth circuit whose gate cost is comparable to a single layer of the baseline ansatz, ensuring that the overall overhead is minimal and manageable (page 6, page 10, page 17).
3. Broad Applicability: Since the method modifies the input state and keeps the parameterized circuit $U(\theta)$ unchanged, it is broadly applicable across different ansatz families, including HEA and Hamiltonian Variational Ansatz (HVA) (page 4). This universality makes it a versatile tool for improving VQA performance.
4. Theoretical Foundation: Theorem 1 provides a rigorous mathematical basis, demonstrating that the achievable ground-state fidelity of a linear combination of orthogonal input states relates to the sum of individual fidelities, justifying the strategy of selecting candidate states with larger overlap with the target (page 6-7).

The paper does not explicitly discuss handling high-dimensional noise better or reducing memory complexity from $O(N^2)$ to $O(N)$. Its superiority is primarily in improving expressivity and reachability under strict resource constraints, leading to higher fidelity with fewer quantum and classical resources.

Alignment with Constraints

The chosen input-state design method perfectly aligns with the harsh requirements of near-term quantum computing, particularly those related to Noisy Intermediate-Scale Quantum (NISQ) devices. The "marriage" between the problem's constraints and the solution's unique properties is evident in several ways:

1. Constraint: Limited Circuit Depth and Noise Accumulation: NISQ devices are highly susceptible to noise, which accumulates with increasing circuit depth. Deeper circuits also exacerbate barren plateaus, making optimization difficult.

   • Alignment: This method directly addresses this by avoiding deeper circuits. Instead of increasing the depth of $U(\theta)$, it modifies the input state to shift the reachable set, allowing shallower circuits to achieve higher fidelities. For example, it achieves 0.99 fidelity with 8 layers where conventional HEA needs 12 layers, representing a significant reduction in depth (page 4, page 10). The encoder itself is low-depth, ensuring it only modestly increases overall circuit depth (page 6).

2. Constraint: Fixed Gate Budget and Resource Limitations: Practical NISQ applications demand solutions that operate within a constrained budget of quantum gates and classical optimization steps.

   • Alignment: The input-state design is inherently resource-efficient. It reduces the total gate count and optimization steps required to reach a target fidelity (page 11). The encoder's gate cost is kept comparable to a single layer of the baseline ansatz, ensuring that the additional quantum overhead is minimal and controlled (page 10, page 17). This allows for "scalable and resource-efficient enhancement of fidelity" (page 9).

3. Constraint: Insufficient Reachability of Shallow Circuits: A major limitation of shallow VQAs is that the target state may lie outside the set of states reachable from a standard input state like $|0\rangle^{\otimes n}$.

   • Alignment: This is the central problem the method solves. By preparing a carefully designed input state $|\Psi_0(\gamma)\rangle$ as a superposition of selected basis states, the method reconfigures the reachable set of the fixed ansatz $U(\theta)$ to include the target ground state (page 9, Figure 2). This ensures that even a shallow $U(\theta)$ can now access the desired state, overcoming the expressivity bottleneck without increasing circuit complexity.

In essence, the method respects the "harsh requirements" of NISQ by working within the limitations of shallow circuits and fixed gate budgets, rather than trying to overcome them by demanding more quantum resources. It provides a clever workaround by optimizing the starting point of the quantum evolution.

Rejection of Alternatives

The paper's rejection of alternatives is primarily an implicit one, focusing on the limitations of existing VQA improvement strategies rather than a direct comparison with entirely different quantum algorithms like GANs or Diffusion models for generative tasks. The authors highlight that "most efforts to improve VQAs have focused on circuit design for $U(\theta)$" (page 4). These circuit-centric approaches, while valuable, are shown to be insufficient for the specific problem of enhancing reachability under fixed resource constraints.

The reasoning for rejecting these circuit-only alternatives is as follows:
1. Trade-off between Expressivity and Trainability: Increasing the expressivity of $U(\theta)$ by making circuits deeper (e.g., using more layers in HEA or HVA) leads to noise accumulation and barren plateaus, making the optimization process difficult or impossible (page 4). This means that while deeper circuits could theoretically reach more states, they become untrainable in practice.
2. Limited Reachability of Shallow Circuits: Conversely, keeping circuits shallow to maintain trainability and mitigate noise often results in the target state being outside the reachable set of the ansatz (page 4). This is the fundamental limitation that the input-state design directly addresses.
3. Inefficiency for Fixed Gate Budget: The paper explicitly demonstrates that conventional HEA or HVA, when constrained to the same gate budget (e.g., total layers or CNOT gates), achieve lower fidelity than the proposed input-state design method (page 4, page 10, page 11). This implies that simply optimizing $U(\theta)$ alone, without considering the input state, is less efficient in terms of resource utilization for achieving high fidelity.

Therefore, the paper doesn't argue that other approaches fail entirely, but rather that they are insufficient or suboptimal for improving VQA reachability under the practical constraints of NISQ hardware. The input-state design is presented as a necessary complement to circuit design, addressing an overlooked aspect that traditional methods could not effectively tackle on their own within a fixed resource budget.

Mathematical & Logical Mechanism

The Master Equation

The absolute core equation that powers the enhanced variational quantum algorithm (VQA) proposed in this paper is the objective function for energy minimization, which incorporates both the variational ansatz and the newly introduced input-state encoder. This function is jointly optimized with respect to the parameters of both components:

$$ E(\theta,\gamma) = \langle 0|V^\dagger(\gamma)U^\dagger(\theta) H U(\theta)V(\gamma)|0\rangle $$

Term-by-Term Autopsy

Let's dissect this equation to understand each component's mathematical definition, physical/logical role, and the rationale behind its inclusion and operation.

• $E(\theta,\gamma)$:

  • Mathematical Definition: This term represents the expectation value of the Hamiltonian $H$ with respect to the quantum state $|\Psi(\theta,\gamma)\rangle = U(\theta)V(\gamma)|0\rangle^{\otimes n}$.
  • Physical/Logical Role: This is the cost function of the Variational Quantum Eigensolver (VQE). The primary goal of the algorithm is to minimize this energy expectation value. In quantum mechanics, the minimum expectation value of the Hamiltonian corresponds to the ground state energy of the system, and the state that achieves this minimum is the ground state itself.
  • Why this operator? The expectation value $\langle\Psi|H|\Psi\rangle$ is a fundamental quantity in quantum mechanics for calculating the average energy of a system in a given state $|\Psi\rangle$. Minimizing this value is the standard and most direct approach for VQE to find the ground state.

• $|0\rangle^{\otimes n}$:

  • Mathematical Definition: This denotes the initial computational basis state where all $n$ qubits are set to the $|0\rangle$ state.
  • Physical/Logical Role: This is the standard, easily preparable, and unentangled starting state for most quantum circuits. It serves as a "blank slate" from which all subsequent quantum operations begin.
  • Why this operator? It's the simplest and most common initial state, providing a universal and reproducible starting point for quantum computation.

• $V(\gamma)$:

  • Mathematical Definition: A parameterized unitary operator, referred to as the "encoder" circuit. It takes the initial state $|0\rangle^{\otimes n}$ and transforms it into a designed input state $|\Psi_0(\gamma)\rangle$. The parameters $\gamma$ are a set of tunable classical values that control the specific gates and rotations within this circuit.
  • Physical/Logical Role: This operator is the central innovation of this paper. Its role is to prepare a "smarter" or more favorable initial state for the subsequent VQA ansatz. By modifying the input state, it effectively reshapes the set of states reachable by the main ansatz, making the target state more accessible. It's designed to be a low-depth circuit to maintain efficiency.
  • Why this operator? It's a unitary operator because quantum evolution must be unitary to preserve probability and adhere to quantum mechanical principles. It's parameterized to allow for flexible optimization and adaptation to different problems, enabling the design of an input state that is tailored to the problem at hand.

• $U(\theta)$:

  • Mathematical Definition: A parameterized unitary operator, representing the "ansatz" circuit. It takes the input state $|\Psi_0(\gamma)\rangle$ (or $|\Psi_0\rangle$ in a conventional VQA) and transforms it into the variational state $|\Psi(\theta,\gamma)\rangle$. The parameters $\theta$ are a set of tunable classical values that control the gates within this circuit.
  • Physical/Logical Role: This is the primary variational quantum circuit that attempts to approximate the target quantum state (e.g., the ground state of $H$). Its expressivity dictates the range of quantum states it can generate and explore within the Hilbert space.
  • Why this operator? Like $V(\gamma)$, it must be unitary for physical realizability on a quantum computer. Its parameterization allows for iterative optimization to find the best approximation of the target state. The specific structure of $U(\theta)$ (e.g., Hardware-Efficient Ansatz or Hamiltonian Variational Ansatz) is chosen based on the problem and available hardware.

• $H$:

  • Mathematical Definition: The Hamiltonian operator of the physical system. It is a Hermitian operator, meaning $H = H^\dagger$.
  • Physical/Logical Role: This operator represents the total energy of the quantum system. In the context of VQE, the objective is to find the quantum state that minimizes the expectation value of this Hamiltonian, which corresponds to the system's ground state.
  • Why this operator? The Hamiltonian is the fundamental operator in quantum mechanics that describes the energy of a system and its time evolution. Its expectation value is the quantity that VQE aims to minimize.

• $U^\dagger(\theta)$:

  • Mathematical Definition: The Hermitian conjugate (or adjoint) of the unitary operator $U(\theta)$. Since $U(\theta)$ is unitary, $U^\dagger(\theta) = U^{-1}(\theta)$.
  • Physical/Logical Role: This operator "undoes" the transformation performed by $U(\theta)$. In the expectation value, it acts on the state $H U(\theta)V(\gamma)|0\rangle$ from the left, effectively projecting it back into the space defined by $V(\gamma)|0\rangle$.
  • Why this operator? It is an essential part of forming the bra vector $\langle\Psi|$. If $|\Psi\rangle = U(\theta)V(\gamma)|0\rangle$, then $\langle\Psi| = \langle 0|V^\dagger(\gamma)U^\dagger(\theta)$. This is a mathematical necessity for computing an expectation value.

• $V^\dagger(\gamma)$:

  • Mathematical Definition: The Hermitian conjugate (or adjoint) of the unitary operator $V(\gamma)$. Since $V(\gamma)$ is unitary, $V^\dagger(\gamma) = V^{-1}(\gamma)$.
  • Physical/Logical Role: Similar to $U^\dagger(\theta)$, this operator "undoes" the transformation performed by $V(\gamma)$. It acts on the state $U^\dagger(\theta) H U(\theta)V(\gamma)|0\rangle$ from the left.
  • Why this operator? It completes the formation of the bra vector $\langle\Psi|$, ensuring the correct calculation of the expectation value.

• $\langle 0| \dots |0\rangle$:

  • Mathematical Definition: This represents the inner product of the state $V^\dagger(\gamma)U^\dagger(\theta) H U(\theta)V(\gamma)|0\rangle$ with the initial state $|0\rangle^{\otimes n}$.
  • Physical/Logical Role: The entire expression $\langle 0| \dots |0\rangle$ calculates the expectation value. In quantum mechanics, expectation values of observables are obtained by "sandwiching" the observable between a state and its conjugate transpose. This is how the average energy of the system is measured.
  • Why this operator? The inner product is the mathematical operation used to project one quantum state onto another, or to calculate the probability amplitude of finding one state in another. Here, it is used to compute the expectation value of $H$.

• Why multiplication instead of addition? The operators $V(\gamma)$, $U(\theta)$, $H$, $U^\dagger(\theta)$, and $V^\dagger(\gamma)$ are multiplied because they represent sequential operations on the quantum state. In quantum mechanics, applying multiple gates or operators to a state is represented by multiplying their corresponding matrices in the order of application. First, $V(\gamma)$ acts on $|0\rangle^{\otimes n}$, then $U(\theta)$ acts on the resulting state, then $H$ acts on that state, and finally the bra vector $\langle 0|V^\dagger(\gamma)U^\dagger(\theta)$ acts from the left to compute the expectation value. This sequence of operations is fundamental to how quantum circuits function.

• Why summation instead of integral? The expectation value, when measured on a quantum computer, involves a summation over measurement outcomes. For a discrete set of computational basis states, the expectation value is inherently a sum of probabilities weighted by the observable's eigenvalues. An integral would be used for continuous variables, which are not typically the direct output of VQAs on current hardware.

Step-by-Step Flow

Imagine a single abstract quantum state, initially a pristine $|0\rangle^{\otimes n}$, moving through a sophisticated quantum assembly line to become a highly optimized approximation of a target ground state.

1. Initial State Input: The process begins with the quantum system being prepared in the simple, unentangled computational basis state $|0\rangle^{\otimes n}$. This is our "raw material" entering the first stage of the assembly line.

2. Encoder Pre-processing: This raw $|0\rangle^{\otimes n}$ state first enters the "encoder" circuit, $V(\gamma)$. This circuit, controlled by its parameters $\gamma$, acts like a specialized pre-processor. It applies a series of carefully chosen single- and multi-qubit gates (rotations, entangling operations) to transform the simple $|0\rangle^{\otimes n}$ into a more complex, "designed input state" $|\Psi_0(\gamma)\rangle$. This step is akin to shaping the raw material into a specific, advantageous form before it enters the main manufacturing process. The goal is to ensure this pre-processed state is already "closer" to the desired final product in the vast quantum state space.

3. Ansatz Main Transformation: The designed input state $|\Psi_0(\gamma)\rangle$ then proceeds to the main "ansatz" circuit, $U(\theta)$. This is the core variational engine, parameterized by $\theta$. It applies another sequence of quantum gates, further transforming $|\Psi_0(\gamma)\rangle$ into the final variational state $|\Psi(\theta,\gamma)\rangle = U(\theta)|\Psi_0(\gamma)\rangle$. This is the primary manufacturing stage, where the state is iteratively refined to approximate the target ground state as closely as possible.

4. Energy Measurement (Conceptual): Once the state $|\Psi(\theta,\gamma)\rangle$ is prepared, it's not directly "acted upon" by the Hamiltonian $H$ in a time-evolution sense. Instead, its energy is measured. This involves decomposing the Hamiltonian $H$ into a sum of measurable Pauli terms. For each term, the quantum computer performs measurements on $|\Psi(\theta,\gamma)\rangle$ to obtain its expectation value.

5. Expectation Value Aggregation: The results from these measurements are then classically aggregated. The expectation value $E(\theta,\gamma)$ is computed by summing the measured expectation values of the individual Pauli terms, weighted by their coefficients in the Hamiltonian. This final numerical value, $E(\theta,\gamma)$, represents the "quality score" of the current variational state, indicating how close its energy is to the true ground state energy. This completes one pass through the quantum-classical loop.

Optimization Dynamics

The mechanism learns, updates, and converges through a hybrid quantum-classical optimization loop, aiming to find the optimal parameters $(\theta_{\text{opt}}, \gamma_{\text{opt}})$ that minimize the energy expectation value $E(\theta,\gamma)$.

1. Two-Stage Optimization Strategy: The learning process is structured in two main phases to manage complexity and improve efficiency:

   • Pre-training of the Ansatz: Initially, only the ansatz circuit $U(\theta)$ is optimized. This involves starting with a simple input state (e.g., $|0\rangle^{\otimes n}$) and iteratively adjusting $\theta$ to minimize $E(\theta) = \langle 0|U^\dagger(\theta) H U(\theta)|0\rangle$. A classical optimizer (like gradient descent or a gradient-free method) explores the loss landscape defined by $E(\theta)$. The optimizer calculates (or estimates) the gradients $\nabla_\theta E(\theta)$ to determine the direction of steepest descent, guiding the parameters $\theta$ towards a local minimum. This phase continues until the gradient norm falls below a predefined threshold, indicating a stable point, and yields a set of pre-optimized parameters $\tilde{\theta}_{\text{opt}}$. This step is crucial for establishing a baseline performance for the ansatz.
   • Joint Optimization of Encoder and Ansatz: After pre-training, the input-state design mechanism is engaged. A pool of $M$ computational basis states is sampled, and their energies are estimated using the pre-trained ansatz $U(\tilde{\theta}_{\text{opt}})$. From this pool, $m$ "promising" states are selected (based on energy and including $|0\rangle^{\otimes n}$) to form the basis for the encoder. The encoder circuit $V(\gamma)$ is then constructed to prepare a superposition of these $m$ states as the new input state $|\Psi_0(\gamma)\rangle$. Now, the full circuit $U(\theta)V(\gamma)$ is considered, and both sets of parameters, $\theta$ and $\gamma$, are jointly optimized to minimize $E(\theta,\gamma) = \langle 0|V^\dagger(\gamma)U^\dagger(\theta) H U(\theta)V(\gamma)|0\rangle$. The $\theta$ parameters are initialized with $\tilde{\theta}_{\text{opt}}$, while $\gamma$ parameters are typically initialized randomly.

2. Gradient Behavior and Loss Landscape:

   • During optimization, the classical optimizer iteratively updates $\theta$ and $\gamma$ by computing gradients $\nabla_\theta E(\theta,\gamma)$ and $\nabla_\gamma E(\theta,\gamma)$. These gradients indicate the slope of the energy landscape with respect to each parameter. The parameters are then adjusted in the direction opposite to the gradient to decrease the energy.
   • The loss landscape for VQAs can be highly complex, often characterized by numerous local minima and "barren plateaus" where gradients vanish exponentially with the number of qubits, hindering convergence.
   • The key insight of this paper is that designing the input state via $V(\gamma)$ effectively reshapes this loss landscape. By starting the ansatz $U(\theta)$ from a more favorable initial state $|\Psi_0(\gamma)\rangle$, the reachable set of $U(\theta)$ is shifted. This shift can move the global minimum of the energy landscape into a region that is more accessible to the optimizer, or it can make the landscape "smoother" in the vicinity of the target state, thereby mitigating barren plateaus and improving trainability. The authors intentionally keep the encoder $V(\gamma)$ shallow to avoid introducing new, problematic local minima or increasing training difficulties.

3. Iterative State Updates and Convergence:

   • In each iteration of the joint optimization, the parameters $\theta$ and $\gamma$ are updated based on the calculated gradients. These updated parameters define a new quantum circuit $U(\theta)V(\gamma)$, which prepares a new variational state $|\Psi(\theta,\gamma)\rangle$.
   • The energy $E(\theta,\gamma)$ of this new state is then measured, and the cycle repeats. This iterative process drives the quantum state $|\Psi(\theta,\gamma)\rangle$ closer and closer to the true ground state of the Hamiltonian $H$.
   • Convergence is typically achieved when the energy $E(\theta,\gamma)$ stabilizes or reaches a value very close to the known exact ground state energy, and the fidelity (overlap with the true ground state) approaches 1. The paper demonstrates that this input-state design leads to consistently higher accuracy and faster convergence (fewer layers, fewer iterations) compared to conventional VQAs, indicating a more efficient exploration of the Hilbert space and a better ability to find the optimal state. The optimization is typically restricted to a modest number of iterations (e.g., 200) to control classical overhead.

Results, Limitations & Conclusion

Experimental Design & Baselines

The authors' experimental design was meticulously crafted to ruthlessly validate their core claim: that designing a suitable input state can significantly enhance the reachability and performance of Variational Quantum Algorithms (VQAs) without increasing circuit depth or parameter count. The central idea is to introduce an "encoder" circuit, $V(\gamma)$, which prepares a superposition input state $|\Psi_0(\gamma)\rangle = V(\gamma)|0\rangle^{\otimes n} = \sum_{j=1}^m \alpha_j |\psi_j\rangle$, instead of the conventional product state $|0\rangle^{\otimes n}$ or $|+\rangle^{\otimes n}$. This encoder effectively shifts the reachable set of the subsequent ansatz circuit $U(\theta)$ in the Hilbert space, making the target state more accessible.

The experimental architecture involved a multi-stage optimization process:
1. Pre-training: The standard ansatz circuit $U(\theta)$ (either Hardware-Efficient Ansatz (HEA) or Hamiltonian Variational Ansatz (HVA)) was first optimized using the conventional input state (e.g., $|0\rangle^{\otimes n}$) to obtain an initial set of parameters, $\theta_{\text{opt}}$. This step establishes a baseline performance for the given circuit depth.
2. Candidate State Selection: A pool of $M$ computational basis states $\{|j^{(k)}\rangle\}_{k=1}^M$ was sampled. For each state, its energy expectation value $E_{j^{(k)}} = \langle j^{(k)}|U^\dagger(\theta_{\text{opt}})HU(\theta_{\text{opt}})|j^{(k)}\rangle$ was estimated using quantum measurements. The number of measurements scaled as $N_m = 1/\epsilon^2$ for a target estimation error $\epsilon$. From this pool, $m$ low-energy states (including $|0\rangle^{\otimes n}$) were selected to form the set $A_m$, which would constitute the superposition for the new input state. The choice of $m$ was made to scale linearly with the system size (e.g., $m=6$ for 12 qubits), ensuring the encoder's gate cost remained comparable to a single layer of the ansatz.
3. Joint Optimization: An encoder $V(\gamma)$ was constructed based on the selected $m$ basis states. Then, the parameters of both the encoder $\gamma$ and the ansatz $\theta$ were jointly optimized to minimize the energy $E(\theta, \gamma) = \langle 0|V^\dagger(\gamma)U^\dagger(\theta)HU(\theta)V(\gamma)|0\rangle$. The ansatz parameters were initialized with $\theta_{\text{opt}}$, and encoder parameters $\gamma$ were randomized. This joint optimization was limited to a modest number of iterations (typically $T=200$) to control classical overhead.

The "victims" (baseline models) against which the proposed method was ruthlessly tested were conventional HEA and HVA circuits, initialized with standard product states. The key to proving their mathematical claims was to compare the performance of the enhanced VQA (ansatz + encoder) against these baselines under a matched gate budget or fixed circuit depth. This meant that the total quantum resources (e.g., number of layers, CNOT gates) for the enhanced method (L layers of ansatz + 1 encoder layer) were kept comparable to the baseline (L+1 layers of ansatz). This architectural choice ensured that any observed performance improvement was directly attributable to the input-state design, rather than simply using a deeper or more complex circuit.

The experiments were conducted on several representative quantum many-body models:
* 1D Transverse-Field Ising Model (TFIM): Using a 12-qubit system and HEA as the ansatz.
* 2D TFIM: Using HVA as the ansatz, across various transverse field strengths.
* Cluster-Ising Model: Using HVA as the ansatz.
* 1D Fermi-Hubbard Model: Using HVA as the ansatz, at half-filling and different interaction strengths.

Performance was primarily measured by ground energy and fidelity (F = $|\langle \Psi|\Psi_{\text{tar}}\rangle|^2$), as well as quantum resources (circuit depth, CNOT gates) and classical resources (optimization steps).

What the Evidence Proves

The evidence presented in the paper definitively proves that input-state design is a powerful and broadly applicable tool for enhancing the reachability and performance of VQAs. The core mechanism, as rigorously proven by Theorem 1, is that by constructing an input state as a linear superposition of $m$ orthogonal states, the maximum achievable fidelity $F$ is the sum of the individual fidelities $F_j = |\langle \psi_j|\Psi_{\text{tar}}\rangle|^2$. This allows the reachable set of the ansatz to be effectively modified to include the target state, even for shallow circuits.

Here's the undeniable evidence:

• 1D Transverse-Field Ising Model (HEA): For the 12-qubit 1D TFIM, the enhanced HEA achieved a fidelity of 0.99 with only 8 layers (112 CNOT gates). In stark contrast, the conventional HEA required 12 layers (144 CNOT gates) to reach the same fidelity, representing a substantial 33% reduction in circuit depth (Fig. 4(a)-(b)). Furthermore, the enhanced method required only 1100 optimization steps compared to 1500 for the conventional HEA, demonstrating gains in classical resources too. The training trajectories clearly show a rapid improvement in energy and fidelity once the encoder is introduced (Fig. 4(c)-(d)).

• 2D Transverse-Field Ising Model (HVA): Across various field strengths ($h \in \{0.5, 1.0, 1.5\}$), the enhanced HVA consistently outperformed the conventional HVA. It achieved lower variational energies and higher fidelities at lower circuit depths (Fig. 5). The encoder, built from $m=8$ basis states, had a gate cost comparable to a single HVA layer, ensuring a fair comparison under a matched gate budget.

• Cluster-Ising Model (HVA): For this model, the enhanced HVA reached a fidelity of 0.99 with just 6 layers (450 two-qubit gates), whereas the conventional HVA needed 9 layers (558 two-qubit gates). The classical resource comparison was even more striking: the input-state design method required $C_R = 54550$ optimization steps, while the conventional HVA demanded $C_R = 118800$. This is a clear demonstration of superior efficiency.
• 1D Fermi-Hubbard Model (HVA): This model served as a stringent test for strongly correlated systems. For $U=2$, the enhanced HVA achieved 0.99 fidelity with 5 layers, while the conventional HVA required 9 layers. More importantly, for higher interaction strengths ($U=5$ and $U=10$), where the conventional method often stagnated at fidelities around 0.6 due to initialization sensitivity and barren plateaus, the input-state design consistently pushed the fidelity above 0.99 (Fig. 8). This highlights the encoder's ability to provide a physically relevant and expressive initialization.
• Resource Overhead Analysis: The paper also analyzed the overhead associated with the sampling number $M$ and the encoder size $m$. It showed that increasing $M$ improves accuracy but with diminishing returns, and that useful improvements are obtained with moderate $M$ (Fig. 9). For instance, for 12-qubit systems, reducing $M$ from 2000 to 400 still yielded fidelities of 0.99. The classical cost of pre-selection for $M=400$ candidates was modest, adding approximately USD 1000 to the total classical cost of USD 189600 for the 12-qubit Ising model (as mentioned in the paper, though the table reference is missing, it's likely Table 2). This combined cost of USD 190600 was still significantly lower than the USD 432000 required by the conventional HEA baseline to achieve the same target fidelity.

Figure 9. Impact of sampling size on infidelity for basis-state in pre-selection step. We consider the 12-qubit 1D transverse-field Ising model with a 5-layer HEA as the variational circuit U(θ) and m = 6. The horizontal axis shows the sampling number M, and the vertical axis reports the final infidelity 1 −F obtained after the joint optimization. Increasing M improves the final accuracy by providing a better set of candidate basis states for constructing the encoder input state, while the improvement quickly saturates for larger M, indicating diminishing returns beyond a moderate sampling number

In summary, the consistent improvements in fidelity, reduced circuit depth, and lower classical optimization costs across diverse models and ansatz families, all under a matched gate budget, provide undeniable evidence that the input-state design mechanism effectively enhances the reachability of VQAs.

Limitations & Future Directions

While the input-state design framework offers a compelling solution to the reachability problem in VQAs, the authors acknowledge several limitations and open avenues for future research.

One primary limitation lies in the heuristic nature of the current encoder construction and candidate-selection strategy. The paper explicitly states that "our current encoder construction and candidate-selection strategy is largely heuristic rather than optimal." This implies that while the method works well, there might be more efficient or robust ways to identify the optimal basis states for the superposition and to construct the encoder circuit $V(\gamma)$. The current approach relies on sampling $M$ computational basis states and selecting $m$ low-energy ones, which, while effective, might not always be the most resource-efficient or globally optimal strategy, especially for problems without strong prior physical intuition.

Another point of discussion revolves around the sampling budget $M$. Although the paper demonstrates that useful improvements are achieved with moderate $M$ and that increasing $M$ yields diminishing returns, the fundamental challenge remains: in the absence of any prior knowledge about the Hamiltonian, guaranteeing high fidelity would, in principle, require sampling almost all $2^n$ computational basis states, which is exponentially inefficient. While the authors focus on the regime where $M$ grows polynomially with system size, the trade-off between sampling overhead and target accuracy is still a practical consideration for scaling to larger systems.

Furthermore, while the input-state design addresses the expressivity bottleneck by modifying the reachable set, its relationship to barren plateaus is nuanced. The authors suggest that using highly entangled input states, which cannot be classically simulated, might allow their protocol to achieve quantum advantage without facing the barren plateau no-go theorem that applies to classically simulable input states. However, the encoder itself is intentionally kept shallow to avoid introducing new training problems. This implies that the core barren plateau issue for the ansatz $U(\theta)$ itself is not directly solved but rather circumvented by a better starting point.

Based on these insights, several discussion topics emerge for further development and evolution of these findings:

1. Adaptive and Intelligent Candidate State Selection: How can we move beyond simple energy-based filtering for selecting candidate basis states? Future research could explore advanced machine learning techniques, such as reinforcement learning or active learning, to adaptively identify the most "informative" basis states or even non-computational basis states for the input superposition. This could significantly reduce the sampling budget $M$ and the number of measurements required, making the pre-selection stage more resource-efficient and scalable.
2. Optimal Encoder Architecture and Parameterization: The current encoder $V(\gamma)$ is constructed to prepare a superposition of computational basis states. Can we explore more sophisticated or problem-specific encoder architectures that might prepare more complex, entangled input states more efficiently? This could involve designing encoders that leverage symmetries of the Hamiltonian or incorporating insights from quantum information theory to optimize the $\alpha_j$ coefficients and the gate structure of $V(\gamma)$ itself.
3. Synergistic Barren Plateau Mitigation: While input-state design improves reachability, it doesn't directly solve the barren plateau problem for the ansatz circuit. A crucial future direction is to investigate how input-state design can be synergistically combined with other barren plateau mitigation techniques (e.g., parameter initialization strategies, local cost functions, or problem-inspired ansatzes). Could a carefully chosen entangled input state also "flatten" or "steepen" the loss landscape of the subsequent ansatz optimization, thereby improving trainability?
4. Scalability and Resource Analysis for Larger Systems: The current simulations are limited to up to 12 qubits. A critical question is how the overheads (sampling $M$, encoder gate count, classical optimization steps) scale for much larger quantum systems. Detailed theoretical and numerical analyses are needed to establish the practical limits and optimal trade-offs for $M$ and $m$ as $n$ increases, ensuring the method remains viable for fault-tolerant quantum computing.
5. Generalization to Other VQA Tasks: The current work primarily focuses on ground state preparation. How can this input-state design framework be extended and validated for other VQA applications, such as quantum machine learning, optimization problems (e.g., QAOA), or simulating excited states? Each application might have unique requirements for the target state, necessitating different strategies for input-state design.
6. Noise Resilience and Hardware Implementation: Given the inherent noise in current and near-term quantum hardware, how robust is the input-state design approach to various noise models (e.g., depolarization, dephasing, readout errors)? Future work could explore noise-aware input-state design strategies or error mitigation techniques specifically tailored for the encoder circuit to ensure performance gains are preserved in realistic hardware implementations.

Table 2. Minimum sample size required to reach target fidelity in the transverse- field Ising model. Results are shown for target fidelities F = 0.99 and 0.95 and for n ∈{6, 8, 10, 12} qubits. The number of selected computational-basis states is set to m = 4 for n = 6, 8 and m = 6 for n = 10, 12. We also report the final fidelity achieved by the baseline hardware-efficient ansatz (HEA) without input-state design under a matched quantum-resource budget: the baseline uses (L + 1) HEA layers, whereas our method uses L HEA layers plus one encoder layer Figure 5. Simulation results for the 12-qubit 2D Ising model at h = 0.5, 1, and 1.5. The upper panels (a, c, e) show the ground energy as a function of circuit depth p for h = 0.5, 1, and 1.5, respectively, and the lower panels (b, d, f) show the corresponding fidelity to the exact ground state. The blue curves correspond to the conventional HVA, and the orange curves correspond to the input-state design (enhanced HVA). Each marker represents the mean over 100 random initializations, and the error bars represent standard deviations over these runs. Across all three values of h, the input-state design consistently achieves lower variational energies and higher fidelities under the same depth, and it reaches the 0.99 fidelity threshold with fewer layers than the baseline