⚛️ quantum physics

Higher moment theory and learnability of bosonic states

This paper presents a sample- and time-efficient algorithm to learn arbitrary Gaussian-transformed bosonic Fock states, resolving a key open question in Fock state BosonSampling, while also establishing a hierarchy of higher-moment-based state classes and deriving a complete set of Gaussian unitary invariants to characterize state equivalence.

Original authors: Joseph T. Iosue, Yu-Xin Wang, Ishaun Datta, Soumik Ghosh, Changhun Oh, Bill Fefferman, Alexey V. Gorshkov

Published 2026-08-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joseph T. Iosue, Yu-Xin Wang, Ishaun Datta, Soumik Ghosh, Changhun Oh, Bill Fefferman, Alexey V. Gorshkov

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 authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quantum world, light is not just a wave or a stream of particles; it is a complex field that can be shaped into specific, fragile patterns called quantum states. Scientists often study these patterns by looking at their "moments," which are essentially statistical snapshots of how the light behaves. The simplest snapshots, known as first and second moments, describe the average position and spread of the light, and they are sufficient to fully describe a common type of quantum state called a Gaussian state. However, many of the most interesting and powerful quantum states used in advanced experiments do not fit this simple description. These more complex states, which can be created by starting with a specific arrangement of photons and then passing them through optical devices, have been notoriously difficult to map out. Traditional methods for figuring out what these states look like require so many measurements that the task becomes impossible as the system grows larger, much like trying to map a vast, shifting landscape by measuring every single grain of sand.

A team of researchers has now developed a new method to efficiently learn the structure of these complex light states, solving a problem that had remained open for some time. The researchers focused on a specific class of states created by taking a precise arrangement of photons, known as a Fock state, and passing them through a device that performs a Gaussian transformation. While the transformation changes the state, the researchers discovered that the resulting pattern is still fully defined by a surprisingly small amount of information: just the first four statistical moments. By showing that these four snapshots contain all the necessary details to reconstruct the entire state, they proved that the complex, high-dimensional object could be learned with a manageable number of samples and a reasonable amount of computer time.

The team designed an algorithm that acts like a decoder. Instead of trying to measure every possible detail of the light field, which would take an exponential amount of time, the algorithm uses the measured moments to mathematically reverse-engineer the transformation that created the state. They demonstrated that for the specific case of light states used in BosonSampling experiments—a setup designed to show quantum advantage—the algorithm can efficiently determine the exact configuration of the light. This is a significant departure from previous approaches, which often relied on the assumption that the states could be easily simulated by classical computers. The new method works even when the states are too complex for classical simulation, proving that efficient learning is possible even in regimes where classical computers struggle.

To ensure their method works in practice, the researchers also analyzed how many measurements are needed to get a reliable result. They found that the number of samples required grows at a manageable rate relative to the size of the system, specifically scaling with the eighth power of the number of light modes in the worst-case theoretical bound. However, when they ran numerical simulations to test the algorithm on random scenarios, they observed that the actual number of samples needed was even lower, scaling closer to the sixth power. This suggests that while the theoretical limits are strict, the method performs very well in realistic situations. The algorithm relies on standard linear algebra techniques, such as breaking down matrices into their fundamental components, making it computationally efficient and practical to run on current hardware.

This work resolves a specific question posed by other scientists regarding whether these particular quantum states could be learned efficiently. By proving that the states are defined by a finite set of higher-order moments, the researchers provided a necessary condition for two states to be related by a specific type of transformation. They also identified a new set of mathematical quantities, called symplectic invariants, that remain unchanged regardless of how the light is transformed. These invariants act as a fingerprint for the state, allowing scientists to determine if two different-looking light patterns are actually the same underlying object viewed through different transformations. This capability is crucial for verifying quantum experiments and understanding the fundamental properties of light.

The implications of this discovery extend beyond just learning the state of light. The framework developed by the researchers offers a new way to characterize and learn quantum states that are not Gaussian, a category that includes many states of interest for future quantum technologies. The authors suggest that their approach could be extended to other classes of states and even to different physical systems, such as fermions. They also noted that while their current method is efficient, there is still room to improve the bounds on how many samples are needed, particularly by using adaptive measurement techniques. The work stands as a concrete demonstration that complex quantum systems, previously thought to be too difficult to map, can be understood through the careful analysis of their statistical moments, opening the door to more efficient verification and control of quantum devices.

Problem Statement

Quantum state tomography generally suffers from intrinsic exponential sample and time complexity as the system size scales. While efficient learning algorithms exist for specific classes of states (e.g., stabilizer states, low-depth circuits, or Gaussian states), a significant open question remained regarding Fock states acted upon by arbitrary Gaussian unitaries. This class of states is central to BosonSampling experiments, which are designed to demonstrate quantum advantage.

Previous results showed that fermionic Fock states acted upon by Gaussian unitaries are Gaussian states and thus fully specified by their second moments, allowing for efficient learning. However, bosonic Fock states acted upon by Gaussian unitaries are not Gaussian states. Consequently, existing efficient learning algorithms relying on second moments (covariance matrices) or classical simulability do not apply. The paper addresses the question: Does there exist a sample- and time-efficient algorithm to learn any bosonic Fock state acted upon by an arbitrary Gaussian unitary?

Methodology and Theoretical Framework

1. Higher Moment Theory and $G_t$ States
The authors introduce a hierarchy of state classes defined by their moments. They define a $G_t$ state as a bosonic mixed state that is the thermal state of a Hamiltonian with degree $\le t$ in the quadrature operators (or the ground state of such a Hamiltonian for pure states).

Key Theoretical Insight: While Gaussian states ( $t=2$ ) are fully specified by first and second moments, the authors prove that Fock states acted upon by arbitrary Gaussian unitaries are fully specified by their first four moments (specifically, the second and fourth moments, as odd moments vanish for Fock states).
Reconstruction: They establish that if a state is promised to be a $G_t$ state, knowing its first $t$ moments allows for the complete reconstruction of the state (via maximum entropy principles for mixed states or ground state uniqueness for pure states).

2. Symplectic Invariants
To characterize the relationship between states under Gaussian unitaries, the authors develop a spectrum of symplectic invariants.

They define moment tensors $\Sigma^{(t)}$ and reshaped operators $\Lambda^{(s)}$ .
They construct a collection of invariants $\lambda(\rho)$ derived from the eigenvalues of specific operators involving the symplectic form $\Omega$ and permutation operators.
Result: They show that any polynomial invariant under Gaussian unitaries can be expressed as a function of these symplectic invariants. This provides necessary conditions for two states to be related by a Gaussian unitary (Gaussian convertibility). They demonstrate that second-moment invariants (symplectic eigenvalues of the covariance matrix) are insufficient to distinguish certain non-Gaussian states, necessitating higher moments.

3. Learning Algorithm
The core contribution is an explicit algorithm to learn the unknown Gaussian unitary $\mathcal{U}_S$ and the unknown initial Fock state $|f\rangle$ given sample access to the state $|\psi\rangle = \mathcal{U}_S |f\rangle$ .

Step 1: Moment Estimation. The algorithm estimates the first and second moment matrices ( $\Lambda^{(1)}$ and $\Lambda^{(2)}$ ) using heterodyne measurements.
Step 2: Fock State Identification. By performing a Williamson decomposition on the real part of the first moment matrix (the covariance matrix), the algorithm extracts the symplectic eigenvalues. These eigenvalues uniquely determine the photon numbers in the initial Fock state $|f\rangle$ (up to permutation).
Step 3: Unitary Recovery. Once the initial state structure is identified, the algorithm "undoes" the active (squeezing) part of the Gaussian unitary using the symplectic diagonalizing matrix. This reduces the problem to learning a passive Gaussian unitary (linear optics) acting on a known Fock state.
Step 4: Passive Unitary Learning. For the passive case, the algorithm utilizes the fourth moment matrix. It constructs a specific matrix $A$ from the fourth moments and the known Fock state parameters. By diagonalizing $A$ and performing a simultaneous diagonalization of derived matrices, the algorithm recovers the unknown unitary matrix $W$ (up to permutation and global phases).

Key Results

1. Efficient Learning Algorithm
The paper presents Theorem 2 (and its detailed versions in the Supplemental Material, Theorems S1–S3), which proves that an unknown Fock state $|f\rangle$ acted upon by an unknown Gaussian unitary $\mathcal{U}_S$ can be learned efficiently.

Sample Complexity: The number of heterodyne samples required scales polynomially with the number of modes $n$ , the maximum photon number $f_{\max}$ , and the squeezing parameter $e^s$ . Specifically, for the standard BosonSampling case (passive unitaries, $f_{\max}=1$ ), the sample complexity is $\tilde{O}(n^8)$ .
Time Complexity: The classical post-processing time is polynomial in $n$ , $f_{\max}$ , and $e^s$ . The algorithm relies on standard linear algebra routines: matrix diagonalization, singular value decomposition (SVD), and Williamson decomposition.

2. Resolution of an Open Question
The work resolves the open question posed by Aaronson and Grewal regarding the learnability of BosonSampling states. While previous algorithms for "t-doped" Gaussian states or classical shadows existed, they either had exponential time complexity or did not apply to this specific non-Gaussian class. This work provides the first sample- and time-efficient algorithm for this class.

3. Numerical Evidence
Numerical simulations on random unitaries suggest that the practical scaling of the required number of samples is closer to $O(n^6)$ , which is better than the rigorous worst-case bound of $O(n^8 \log n)$ .

Significance and Claims

The authors claim that their work establishes a new framework for learning quantum states based on higher moments, extending beyond the limitations of Gaussian state tomography.

Complementarity: The results are complementary to other learning algorithms (e.g., those for $t$ -doped states or classical shadows) that apply to different classes of non-Gaussian states.
Invariants: The derived symplectic invariants provide a tool for determining Gaussian convertibility without a Fock space cutoff, a problem where previous methods were computationally demanding or restricted.
Future Directions: The paper modestly suggests that these moment-based methods could be extended to learn other $G_t$ states (e.g., $t=6$ ) and that the invariants could find use in resource theories where Gaussian operations are "free." It also notes that the definition of $G_t$ states and the moment techniques could be generalized to fermionic systems.

The paper does not propose new experimental setups but rather provides a theoretical and algorithmic solution to the problem of characterizing states generated in existing BosonSampling architectures, demonstrating that despite their non-Gaussian nature, they remain efficiently learnable.

Problem Statement

Methodology and Theoretical Framework

Key Results

Significance and Claims

More like this