---
title: Quantum State Preparation (QSP)
url: https://www.emergentmind.com/topics/quantum-state-preparation-qsp
type: topic
---

# Quantum State Preparation (QSP)

Searching arXiv for recent quantum state preparation papers to ground the article in fresh literature.
Quantum state preparation (QSP) is the task of converting a fixed fiducial input, usually \(|0\rangle^{\otimes n}\), into a target \(n\)-qubit state whose amplitudes encode classical or physically derived data. In exact amplitude encoding, one seeks a unitary \(U\) such that
\[
U|0\ldots 0\rangle=\sum_{k=0}^{2^n-1} d_k |k\rangle,
\qquad \|d\|_2=1,
\]
while in sparse quantum state preparation (SQSP) the target has only \(d\) nonzero amplitudes, for example
\[
|\phi(n,d,S)\rangle=\sum_{i=0}^{d-1}\alpha_i|q_i\rangle
\]
for a list \(S=\{(\alpha_i,q_i)\}\) [2408.05435, 2508.21346]. QSP is a key component in many quantum algorithms, but for a general \(n\)-qubit state exact amplitude encoding typically requires \(O(2^n)\) CNOT gates and \(O(2^n)\) circuit depth, making worst-case loading impractical on near-term hardware and driving a large literature on sparse, low-rank, transform-based, ancilla-assisted, variational, and measurement-based alternatives [2512.01798].

## 1. Formal problem and worst-case complexity

The canonical formulation of QSP is to prepare
\[
|\Psi\rangle=\sum_{k=0}^{2^n-1} x_k |k\rangle
\]
from a normalized classical vector \(x\in\mathbb C^{2^n}\). In the sparse setting, the input is not the full \(2^n\)-component vector but the sparsity \(d\), a list of nonzero amplitudes, and the associated basis strings \(q_i\in\{0,1\}^n\) [2512.01798, 2508.21346]. This distinction is operationally decisive: generic exact amplitude encoding scales exponentially with \(n\), whereas SQSP and other structured regimes admit lower-depth constructions when the target has compressed classical structure.

Several papers make the worst-case barrier explicit. Exact amplitude encoding for a general \(n\)-qubit state requires on the order of \(O(2^n)\) CNOT gates and \(O(2^n)\) circuit depth in standard decompositions such as Mottonen-Iten-style schemes, because the amplitudes must be synthesized by cascades of controlled rotations over the computational basis [2512.01798]. A related phase-estimation-based line of work associates state preparation efficiency with the decomposition of diagonal unitaries \(U_\alpha^{(k)}\), \(U_1\), and \(U_\theta\), and states that efficient decomposition of the corresponding diagonal unitary operators is a sufficient condition for efficient state preparation [1912.05335].

Ancilla qubits change the depth landscape. Any \(n\)-qubit state can be prepared with a \(\Theta(n)\)-depth circuit using only single- and two-qubit gates, but this construction uses \(O(2^n)\) ancillary qubits [2201.11495]. For sparse states with \(d\ge 2\) nonzero entries, the same work gives depth \(\Theta(\log(nd))\) with \(O(nd\log d)\) ancillary qubits and argues optimality through light-cone and fan-out style lower bounds [2201.11495]. This establishes a recurring theme of the QSP literature: depth reduction is routinely purchased by ancillas, classical preprocessing, or stronger hardware assumptions.

## 2. Exact structured preparation: sparsity, factorization, and compressed classical descriptions

SQSP has become a central exact regime because many targets of practical interest have small support in the computational basis. A 2025 construction gives two sparse-state algorithms with total size \(O(dn)\) and \(O(d)\) ancilla qubits: a fully unitary algorithm of depth \(O(n\log d)\), and a second algorithm of depth \(O(n)\) that uses mid-circuit measurement and feedforward [2508.21346]. The unitary version is organized into four conceptual steps: dense GQSP on \(\lceil \log d\rceil\) qubits, one-hot encoding, permutation into the full \(n\)-qubit support, and garbage elimination. The measurement-assisted version replaces copy trees and OR-controlled gates by constant-depth measurement-based fan-out constructions, using the MaF primitive of Bäumer and Woerner and the equivalence between OR-controlled \(X\) and parity-controlled \(X\) on one-hot inputs [2508.21346].

That same sparse-state work compares its resource profile with several prior proposals. Relative to Sun et al. for \(m=O(d)\), it reports lower depth; relative to Zhang–Li–Yuan it matches \(\Theta(\log d\, n)\)-type depth with exponentially fewer ancillas \(O(d)\); relative to CCQSP it keeps \(O(d)\) ancillas instead of \(O(d n/\log d)\); and relative to constant-depth MaF preparation it reduces size and ancilla count to \(O(dn)\) and \(O(d)\), at the cost of linear depth [2508.21346]. The practical limitations are equally explicit: ancilla count \(O(d)\) may still be large for intermediate \(d\approx n\), the MaF version requires low-latency classical feedforward and mid-circuit measurements, and all-to-all connectivity is assumed for large-fanout and multi-Toffoli gates [2508.21346].

Another exact line of work exploits compressed classical data structures rather than sparsity alone. Weighted free binary decision diagrams (FBDDs) encode a state by assigning complex weights to graph edges, and any state represented by a weighted FBDD with \(N\) nodes can be prepared by an \(O(N)\)-sized, \(O(N)\)-depth circuit using \(N-2\) ancillas [2407.01671]. The paper emphasizes that FBDDs strictly generalize OBDDs and provides examples where a state with \(\Theta(2^n)\) support has an \(O(\mathrm{poly}(n))\)-node FBDD, yielding exponential improvement over OBDD-based QSP [2407.01671].

LimTDD-based QSP similarly uses compressed graph structure, but with local invertible map tensor decision diagrams. It proposes algorithms for no ancilla qubits, one ancilla qubit, and many ancilla qubits, with complexities governed by the number of reduced paths \(p\) or the number of nodes \(m\) in the LimTDD [2507.17170]. In the best-case “tower” form, the no-ancilla algorithm uses \(2n\) single-qubit plus \(\tfrac{n(n-1)}{2}\) two-qubit gates, and the many-ancilla algorithm runs in \(O(m)\) time. Benchmarks on random Clifford+T states up to \(n=15\) report substantial post-transpile gate-count reductions relative to Qiskit, QuICT, ADD-based, and FBDD-based baselines [2507.17170].

Exact preparation can also be simplified when the target factors. A multiplexer simplification method detects tensor-product structure by reshaping amplitudes into a \(2^p\times 2^{n-p}\) matrix and checking whether the matrix rank is \(1\) [2409.05618]. When repeated blocks occur in the multiplexer tree, controls can be eliminated, replacing worst-case depth and CNOT count \(\Theta(2^n)\) by \(\Theta(2^{n_e})\), where \(n_e\) is the size of the largest entangled component after factorization [2409.05618]. This does not address approximate factorization, but it yields substantial compilation-time and circuit-size gains for exactly separable or partially disentangled states.

## 3. Compression, low-rank structure, tensor networks, and multivariate functions

A major branch of QSP shifts complexity into classical compression or factorization. An ancilla-free hybrid classical-quantum framework first applies a reversible transform \(T\), such as the discrete Fourier transform or the discrete Haar wavelet transform, sparsifies the transformed vector by thresholding, prepares the resulting \(d\)-sparse state, and then applies the inverse quantum transform \(U_T^\dagger\) [2512.01798]. For compressible data this replaces exponential quantum cost by \(O(\mathrm{poly}(n))\) quantum gates. The paper reports, for example, a multi-frequency periodic benchmark with \(n=8\), \(d=4\), \(CR=64{:}1\), where the hybrid method uses 82 CNOTs, depth 74, and fidelity \(1.0\), compared with \(\sim 240\) CNOTs and depth \(\sim 550\) for exact amplitude encoding; for a piecewise-constant benchmark with \(n=10\), \(d=8\), \(CR=128{:}1\), it reports 126 CNOTs, depth 46, and fidelity \(1.0\), compared with \(\sim 1024\) CNOTs and depth \(\sim 2000\) for exact loading [2512.01798]. The same work states that worst-case scaling remains exponential if \(d=O(2^n)\), so the polynomial regime is conditional on classical compressibility [2512.01798].

Low-rank quantum state preparation uses Schmidt truncation. For a bipartition \(A|B\),
\[
|\psi\rangle=\sum_{i=1}^{k}\sigma_i |i_A\rangle\otimes |i_B\rangle,
\]
and the target can be approximated by retaining only the top \(r\) Schmidt terms [2111.03132]. The quantum circuit first prepares Schmidt coefficients on \(m=\lceil \log_2 r\rceil\) qubits, then uses CNOT entangling and two isometries \(U\) and \(V^T\). If \(m<n_A\), the CNOT count and depth become \(O(2^{m+n_B})\), interpolating between highly compressed and worst-case \(O(2^n)\) preparation [2111.03132]. On discretized probability distributions over 7 qubits, allowing fidelity loss \(\ell_{\max}=0.02\), the bounded-approximation-error wrapper reduced CNOT counts from 28–109 to 3–6, and on a random complex 3-qubit vector the LRSP circuit achieved tomography fidelity \(\approx 0.92\) versus \(\approx 0.78\) for Qiskit’s standard amplitude encoding under the FakeCairo noise model [2111.03132].

Tensor-network structure yields another compressed regime. For smooth differentiable probability densities, an MPS representation with small bond dimension can be exploited by iterative MPS circuit loading [2303.01562]. For normal distributions, the paper uses an Irwin–Hall approximation whose piecewise-polynomial structure admits an exact MPS of bond dimension \(\chi=n(n+1)\), and then applies Ran et al.’s iterative \(\chi'=2\) circuit-loading scheme [2303.01562]. The circuits are ancilla-free, the two-qubit gate count is approximately \(2(N-1)D\) CNOTs, and experiments on IonQ Aria reached up to 20 qubits. For \(D=1\), the Kolmogorov–Smirnov statistic is reported as \(\approx 0.09\) for 10 qubits and \(\approx 0.15\) for 20 qubits with \(s=10\,000\) shots [2303.01562].

Multivariate function loading can also be phrased as linear combination of block-encodings. A Fourier/Chebyshev approach prepares the normalized discretization of \(f:[-1,1]^D\to\mathbb R\) by block-encoding basis functions and combining them with LCU, without arithmetic circuits, QFTs, or multivariate quantum signal processing [2405.21058]. For the Fourier version, the asymptotic resource statement is \(O(d^D + D n\log d)\) two-qubit gates and depth, with \(D\,\lceil \log_2(2d+1)\rceil\) ancillas; for the Chebyshev version, the two-qubit gate count is \(O(d^D + D d n\log n)\) [2405.21058]. On Quantinuum H2-1, the method prepared bivariate Gaussian distributions on a \(512\times 512\) grid using 24 qubits and up to 237 two-qubit gates, with overall fidelity \(F=0.986\) for the uncorrelated case and \(F=0.918\) for the correlated case [2405.21058].

## 4. Approximate series methods, variational loaders, and learned circuit synthesis

Approximate QSP is often attractive when fidelity can be traded for shallower circuits. The Walsh Series Loader (WSL) targets states defined by real-valued functions of a single real variable and uses a truncated Walsh expansion implemented through diagonal Walsh rotations plus a Repeat-Until-Success interference step [2307.08384]. In the dense truncated-series version, the truncation order is \(M=O(1/\sqrt{\epsilon})\), the circuit depth is \(O(1/\sqrt{\epsilon})\), the size is \(O(n+1/\sqrt{\epsilon})\), and only one ancilla qubit is needed [2307.08384]. In the sparse-Walsh version, if the series has sparsity \(s\) and maximal Walsh-index Hamming weight \(k\), the circuit approximates the target up to error \(\epsilon\) with depth \(O(sk)\), size \(O(n+sk)\), and one ancilla qubit [2307.08384]. The protocol is probabilistic, with success probability \(P=\Theta(\epsilon)\) in the main setting, so its averaged total time becomes \(O(1/\epsilon^{3/2})\) for WSL and \(O(sk/\epsilon)\) for sparse WSL [2307.08384].

Variational and adaptive ansätze approach QSP by circuit compression rather than direct exact synthesis. In the context of strongly correlated chemistry states up to 28 qubits, Overlap-ADAPT-VQE builds a circuit by greedily appending operators that maximize the gradient of the overlap with a target sparse CI state [2311.03347]. For the 28-qubit H\(_{14}\) chain, the paper reports that 50% overlap is reached at \(\sim 100\) CNOTs, 90% at \(\sim 500\), 95% at \(\sim 1\,200\), and 99% at \(\sim 2\,400\), while exact CVO-QRAM loading of the full ground state requires \(\sim 2\times 10^5\) CNOTs [2311.03347]. The authors conclude that Overlap-ADAPT-VQE offers the most advantageous performance for near-term applications [2311.03347].

A different approximate paradigm treats QSP as a learned compilation problem. SuperEncoder uses a pre-trained fully connected MLP with two hidden layers, each of width 512 in the prototype, to map a target amplitude vector directly to the parameters of a hardware-efficient PQC [2408.05435]. The final circuit depth is \(3L\), with empirical settings \(n=4\to L=8\to D=24\), \(n=6\to L=20\to D=60\), and \(n=8\to L=40\to D=120\) [2408.05435]. On a 4-qubit ideal simulator, the reported per-sample runtimes are \(\sim 0.015\) s for exact amplitude encoding, \(\sim 5\) s for iterative approximate amplitude encoding, and \(\sim 0.014\) s for SuperEncoder, with fidelities \(1.0000\), \(0.9994\), and \(0.9310\), respectively [2408.05435]. This suggests a specific trade-off: online optimization is removed, but fidelity is systematically lower than in per-instance variational refinement.

## 5. Measurement, feedforward, and control-theoretic preparation

Mid-circuit measurement and classical feedback have become explicit algorithmic resources in QSP. In sparse preparation, measurement-assisted fan-out reduces the depth from \(O(n\log d)\) to \(O(n)\) by replacing copy trees and multi-control constructions with MaF-based constant-depth primitives, but the same paper notes that on current hardware measurement latency in the \(\mu\)s–ms range may dominate gate delays in the ns range and negate the nominal depth advantage [2508.21346]. The resulting advantage is therefore architecture-dependent rather than purely asymptotic.

Measurement-induced steering provides a conceptually different use of feedback. A system register \(S\) is repeatedly entangled with an ancilla \(A\), the ancilla is measured, and \(A\) is actively reset, while \(S\) is never measured or reset [2302.13518]. The induced CPTP map has Kraus operators \(K_m=\langle m|U_{SA}|\psi_A\rangle_A\), and for suitable \(U_{SA}\) the steering inequality
\[
\langle \psi_\oplus|\rho_S^{n+1}|\psi_\oplus\rangle \ge \langle \psi_\oplus|\rho_S^n|\psi_\oplus\rangle
\]
holds at every step [2302.13518]. On IBM superconducting processors, blind passive qubit steering to \(|+\rangle\) reached \(F=0.954\pm 0.026\) after \(N\approx 4\) rounds, the average fidelity over six stabilizer targets was \(0.925\pm 0.012\), and active feedback reduced the mean number of rounds from \(\approx 3.8\) to \(\approx 1.6\) to reach \(F>0.9\) [2302.13518]. The same framework was extended to qutrit steering, where \(F=0.80\pm 0.09\) was obtained after six rounds [2302.13518].

Control-theoretic formulations cast state preparation as policy learning. In semiconductor double quantum dots, arbitrary-to-arbitrary pure-state preparation was mapped to a discrete-time Markov decision process and solved with a Deep Q-Network whose input concatenates informationally complete POVM statistics for the current and target states [2407.16368]. For the single-qubit system \(H(t)=J(t)\sigma_z+\sigma_x\), the action set is \(J(t)\in\{0,1,2,3,4\}\) with \(dt=\pi/5\); for the two-qubit capacitively coupled system the action space is \((J_1,J_2)\in\{1,2,3,4,5\}^2\) with \(dt=\pi/4\) [2407.16368]. Average fidelities over randomized test sets are reported as \(\bar F=0.9864\) for single-qubit AQSP and \(\bar F=0.9556\) for two-qubit AQSP, with robustness against moderate charge and nuclear noise [2407.16368].

Ground-state preparation constitutes a specialized but important subdomain. A deterministic protocol based on a Power-Cosine quantum signal processing filter applies repeated controlled time evolution and ancilla measurement/reset so that the effective non-unitary filter is
\[
\mathcal F^{(d)}=\left(\frac{I+e^{-iH\tau}}{2}\right)^d,
\]
whose envelope is \(F_d(x)=\cos^d(x/2)\) [2602.19556]. The paper derives exponential suppression of excited components and a depth scaling \(O(\Delta^{-2}\log(1/\epsilon))\), with one ancilla qubit and mid-circuit measurement/reset [2602.19556]. On the 1D Heisenberg XYZ model, it reports infidelity \(1-\mathcal F\approx 10^{-6}\) at QSP depth \(d=20\), compared with Trotterized adiabatic state preparation infidelity \(\sim 10^{-2}\) at the same cost [2602.19556].

## 6. Applications, recurring trade-offs, and limiting conditions

QSP is tightly coupled to downstream algorithm design. Low-depth state preparation has been used to reduce oracle depth in Hamiltonian simulation, linear-system solving, and QRAM, with explicit exponential depth savings in sparse regimes [2201.11495]. Distribution-loading methods are motivated by amplitude estimation, HHL, quantum machine learning, and initialization for Hamiltonian simulation [2303.01562]. Function-based loaders target finance, physics, and chemistry simulations, including Student’s \(t\)-distributions, Ricker wavelets, and electron wavefunctions in Coulomb potentials [2405.21058]. Sparse chemistry-state preparation is motivated by quantum phase estimation and ground-state algorithms for strongly correlated systems [2311.03347].

Across these subliteratures, three constraints recur. First, low depth is rarely free: arbitrary-state \(\Theta(n)\)-depth circuits require \(O(2^n)\) ancillas, sparse log-depth circuits often need \(O(nd\log d)\) or \(O(d)\) ancillas, FBDD and LimTDD methods scale with diagram size, and measurement-assisted schemes require mid-circuit measurement and feedforward [2201.11495, 2508.21346, 2407.01671]. Second, polynomial quantum cost usually presupposes classical structure, such as sparsity, low Schmidt rank, small bond dimension, transform-domain compressibility, separability, or compact decision diagrams; absent such structure, worst-case scaling remains exponential [2512.01798, 2111.03132, 2409.05618]. Third, compilation and hardware realism matter as much as asymptotics: multi-controlled gates remain expensive after transpilation, all-to-all connectivity is often assumed, measurement latency can offset feedforward gains, and noise can favor approximate or compressed circuits over exact ones [2507.17170, 2508.21346, 2111.03132].

A common misconception is that “state preparation” denotes a single primitive with a single optimal implementation. The literature instead presents a heterogeneous design space: exact arithmetic-style synthesis, sparse-state loaders, diagram-based compilers, low-rank and tensor-network methods, transform-domain compression, Repeat-Until-Success series approximations, adaptive variational ansätze, neural one-shot compilers, steering protocols, and control-learning formulations all instantiate QSP under different assumptions [1912.05335, 2307.08384, 2408.05435]. Another misconception is that ancilla-free schemes are automatically preferable. Ancilla-free methods can be highly attractive, as in transform-based hybrid loading and MPS circuits, but ancilla-assisted schemes sometimes achieve much lower depth, and measurement-assisted schemes can outperform unitary ones only when classical feedforward is fast enough [2512.01798, 2303.01562, 2508.21346].

Taken together, the modern theory of QSP is best understood as a theory of resource conversion. Classical structure is exchanged for lower quantum depth; ancillas are exchanged for fan-out, routing, or isometries; measurements and resets are exchanged for shallower coherent circuits; and approximation error is exchanged for improved fidelity on noisy devices. This suggests that no single asymptotic bound captures the practical frontier of QSP: the dominant question is which representation of the target state makes the relevant quantum resource—depth, width, two-qubit count, measurement latency, or classical preprocessing—least costly for the hardware and algorithmic context at hand.

Source: https://www.emergentmind.com/topics/quantum-state-preparation-qsp