⚛️ quantum physics

A machine learning approach to tomographic pattern generation and classification of quantum states of light

This paper presents a deep learning framework utilizing Wasserstein generative adversarial networks to generate and classify optical tomographic patterns of various quantum states of light, enabling direct characterization of state properties like mean photon number without requiring explicit state reconstruction or additional classifiers.

Original authors: Soumyabrata Paul, H. S. Subramania, S. Ramanan, V. Balakrishnan, S. Lakshmibala

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Soumyabrata Paul, H. S. Subramania, S. Ramanan, V. Balakrishnan, S. Lakshmibala

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

Imagine you are trying to identify a mysterious object hidden inside a dark room. In the world of quantum physics, this object is a "state of light" (like a specific arrangement of photons). Traditionally, to figure out what this object looks like, scientists have to take thousands of photos from every possible angle, then use a super-complex mathematical recipe to reconstruct the 3D shape of the object. This process is slow, computationally heavy, and often prone to errors.

This paper proposes a smarter, faster way: Teaching a computer to "dream" the object back into existence.

Here is the breakdown of their approach using simple analogies:

1. The Problem: The "Shadow" Puzzle

Think of a quantum state of light as a complex 3D sculpture. When you shine a light on it from different angles, you get a 2D shadow. In physics, these shadows are called Optical Tomograms.

The Old Way: If you want to know what the sculpture is, you collect all the shadows, feed them into a massive calculator, and try to reverse-engineer the 3D shape. It's like trying to guess the shape of a car by looking at its shadow on a wall, but you have to do the math for every single pixel.
The Challenge: The "sculpture" (the quantum state) exists in a huge, complex space. Reconstructing it perfectly is like trying to solve a jigsaw puzzle where half the pieces are missing and the picture keeps changing.

2. The Solution: The "Art Forgery" Team (WGAN)

The authors used a Machine Learning technique called a Wasserstein Generative Adversarial Network (WGAN). Imagine this as a high-stakes art forgery contest between two AI artists:

The Forger (The Generator): This AI's job is to look at a real "shadow" (a tomogram) of a light state and then try to draw a fake one that looks exactly like the real thing.
The Detective (The Discriminator): This AI's job is to look at the Forger's drawing and the real shadow and say, "Is this real or fake?"

How they learn:

The Forger makes a bad drawing.
The Detective says, "That's obviously fake! The lines are wrong."
The Forger tries again, adjusting its technique based on the Detective's feedback.
They repeat this millions of times. Eventually, the Forger becomes so good at mimicking the shadows that even the Detective can't tell the difference.

3. The Magic Trick: Reading the "Dream"

Usually, when an AI learns to draw, we just look at the picture to see if it's pretty. But the authors wanted to do more. They wanted to know the properties of the object just by looking at the AI's drawing.

They trained the AI on three specific types of light states:

Fock States: Like a light bulb with a specific, fixed number of photons (e.g., exactly 3 photons).
Coherent States: Like a standard laser beam (very orderly).
1-Photon Added States: A laser beam where someone sneaked in one extra photon.

Once the AI (the Forger) mastered drawing these shadows, the scientists didn't need to do the heavy math reconstruction. They simply looked at the AI's generated shadows and calculated the average number of photons and the variance (how "wobbly" the light is) directly from the picture.

The Analogy:
Imagine you want to know the weight of a mystery fruit.

Old Way: You weigh the fruit, then weigh the skin, then weigh the seeds, then do complex math to subtract the skin and seeds to find the fruit's true weight.
This Paper's Way: You teach an AI to draw the fruit so perfectly that when you look at the drawing, you can instantly say, "That looks like a 5-ounce apple," without ever weighing the real fruit.

4. Why This Matters

Speed & Efficiency: They skipped the slow, difficult step of "reconstructing" the full quantum state. They went straight from the shadow (tomogram) to the answer (properties of the state).
Robustness: They tested this with "noisy" data (simulating real-world imperfections like bad detectors). The AI was like a seasoned detective who could still identify the fruit even if the drawing was a bit smudged.
Color Matters: They found that the "color map" used to visualize the shadows mattered. Using a specific type of color gradient helped the AI see the fine details (the "tails" of the probability distribution) better, making the measurements more accurate.

5. The "Real World" Test

The team compared their AI's results with a recent real-life experiment where scientists tried to distinguish between a "standard" laser and a "photon-added" laser.

The Result: The AI successfully mimicked the experimental data. It correctly identified that for certain settings, the two states look very similar (like twins), but by measuring specific "wobbles" (variances) in the AI's generated shadows, it could still tell them apart.

Summary

This paper is about teaching a computer to recognize patterns in quantum light by learning to draw them. Instead of solving a massive, difficult math puzzle to understand the light, they let the AI learn the "shape" of the data. Once the AI can draw the pattern perfectly, we can read the answers (like photon count and energy) directly from the drawing, skipping the heavy lifting of traditional physics reconstruction.

It's a shift from solving the puzzle to learning to recognize the picture.

1. Problem Statement

Quantum State Tomography (QST) is the standard procedure for reconstructing an unknown quantum state of light from experimental data (homodyne measurements). However, traditional QST faces significant challenges:

Computational Complexity: Reconstructing density matrices or Wigner functions, especially for multimode or large Hilbert space systems, is computationally demanding (e.g., requiring Maximum Likelihood Estimation).
Data Requirements: Ideally, an infinite number of tomographic slices (probability density functions at different angles) are needed for perfect reconstruction, which is practically impossible.
Classification Difficulty: Distinguishing between similar states (e.g., amplified coherent states vs. photon-added coherent states) often requires full state reconstruction followed by a separate classifier.

The authors propose bypassing the full state reconstruction step. Instead, they aim to use machine learning to generate optical tomograms directly from input patterns and characterize/classify the quantum states using physical observables extracted directly from these generated tomograms, without needing an auxiliary neural network classifier.

2. Methodology

The core of the approach is a Wasserstein Generative Adversarial Network (WGAN) framework applied to optical tomographic patterns.

A. Data Representation

Input: Optical tomograms $w(X_\theta, \theta)$ , which represent the probability density functions (PDFs) of field quadratures $X_\theta$ at phase angles $\theta$ .
Visualization: Tomograms are treated as 2D images (grids of $128 \times 128$ pixels) where intensity represents probability. The authors tested both RGB image data and raw numerical arrays, finding image-based training sufficient and comparable.
Training States: The model was trained on three classes of single-mode states:
1. Fock States $|n\rangle$ (for $n=0$ to $5$).
2. Coherent States (CS) $|\alpha\rangle$ .
3. Single-Photon Added Coherent States (1-PACS) $|\alpha, 1\rangle$ .

B. Network Architecture

The system utilizes two Convolutional Neural Networks (CNNs):

Generator (G): Takes random noise as input and generates synthetic tomographic patterns.
Discriminator (Critic, D): Distinguishes between real input tomograms and generated ones.

Loss Function: The networks are trained using the Wasserstein distance (Earth Mover's Distance) rather than standard GAN loss. This provides a smooth, non-vanishing gradient even when distributions do not overlap, ensuring stable convergence.
Training Protocol: The discriminator is trained 5 times for every 1 generator update over 25,000 epochs. The Adam optimizer is used with specific hyperparameters ( $\beta_1=0, \beta_2=0.9, lr=0.0001$ ).

C. Error Modeling and Robustness

To simulate experimental imperfections (detector inefficiency, noise), two error models were applied to the training data:

Uniform noise: $[1 + U(\epsilon)]w(X_\theta, \theta)$
Gaussian noise: $[1 + N(0, \epsilon^2)]w(X_\theta, \theta)$
Thresholds of $\epsilon = 0.10$ (marginal) and $0.25$ (realistic) were tested to verify robustness.

D. Characterization without Reconstruction

Crucially, the authors do not reconstruct the density matrix $\hat{\rho}$ or Wigner function. Instead, they extract physical observables directly from the generated PDFs:

Mean Photon Number ( $\langle \hat{n} \rangle$ ): Calculated using a specific integral formula involving PDFs at $\theta = 0, \pi/3, 2\pi/3$ .
Quadrature Variances ( $\Delta \hat{X}^2_\theta$ ): Calculated directly from the reconstructed PDFs.
Nonclassicality Indicator (NI): Derived from the integral of the square root of the variance over all angles.

3. Key Contributions

Direct Tomogram Generation: Demonstrated that WGANs can successfully learn to generate high-fidelity optical tomograms for Fock, Coherent, and 1-PACS states, mimicking the input patterns with high accuracy.
Reconstruction-Free Characterization: Showed that key quantum properties (mean photon number, variances, higher moments) can be computed directly from the generated tomograms with <4% error relative to theoretical values, eliminating the need for complex inverse problems (QST).
State Classification via Observables: Proved that different quantum states can be distinguished solely by comparing their computed observables (e.g., distinguishing 1-PACS from amplified CS via variance/squeezing properties) without a secondary classifier network.
Robustness to Noise: Verified that the trained model remains effective even when trained on noisy data, successfully reproducing trends seen in real experimental data.
Colormap Analysis: Investigated the impact of color mapping on reconstruction accuracy. Found that while nonlinear colormaps with a "regulator" function help suppress tail artifacts, a standard linear colormap is sufficient for most characterizations.

4. Results

Fock States: The generated tomograms correctly reproduced the $\theta$ -independent intensity distribution. The computed mean photon numbers and variances for $n=0$ to $5$ fell within 4% of theoretical values.
Coherent States: The model captured the $\theta$ -dependent rotation of the Gaussian PDFs. The mean photon number matched $|\alpha|^2$ within 4% error, and variances remained close to the vacuum limit (0.5).
1-PACS: Successfully generated states showing the transition from coherent to Fock-like behavior. The model correctly identified the "vertical cut" feature in tomograms for low $\alpha$ and its disappearance for high $\alpha$ .
Experimental Validation: The method was tested against recent experimental data regarding the fidelity between amplified CS, optimal CS, and 1-PACS. The ML-generated observables correctly reproduced the experimental trend: for $\alpha \geq 1$ , the mean photon numbers of these distinct states converge, but their variances (specifically squeezing in 1-PACS) remain distinct, allowing for correct classification.
Nonclassicality: The Nonclassicality Indicator (NI) calculated from generated tomograms correctly identified $n=0$ as classical (NI $\approx$ 0) and $n>0$ as nonclassical (NI > 0).

5. Significance and Outlook

Alternative to QST: This work establishes a viable alternative to traditional Quantum State Tomography. By avoiding the reconstruction of the full density matrix, it significantly reduces computational overhead, making it suitable for systems with large Hilbert spaces (e.g., spin arrays, multipartite systems).
Efficiency: The approach allows for the direct characterization of unknown states using a limited number of tomographic slices, provided the machine is trained on a representative set.
Future Directions: The authors suggest that while the current architecture introduces slight biases (skewness in variance distributions) due to color transitions in the generator, techniques like local padding or super-resolution GANs could further refine the generation process. The method is also poised to act as a denoiser for real experimental data.

In summary, the paper demonstrates that deep learning, specifically WGANs, can effectively learn the "pattern" of quantum optical tomograms, enabling the direct extraction of physical observables and state classification without the computationally expensive step of full state reconstruction.