Quantum-Enhanced Single-Parameter Phase Estimation with Adaptive NOON States

This paper presents an end-to-end differentiable quantum-optical framework that optimizes adaptive NOON states for single-parameter phase estimation, demonstrating that the previously established experimental working point is significantly suboptimal for $N \geq 3$ and achieving up to a 1598% improvement in classical Fisher information alongside an 8- to 133-fold increase in useful measurement events.

The Gist

The Big Picture: Tuning a Quantum Radio

Imagine you are trying to tune a very sensitive radio to hear a faint signal. In the world of quantum physics, scientists use light particles (photons) to measure tiny changes, like a slight shift in a mirror's position. The more photons you use, the clearer the signal should be.

However, there's a catch. To get the best possible clarity (called the "Heisenberg limit"), you need to arrange these photons in a very specific, tricky pattern called a NOON state. Think of this like a choir where everyone sings the exact same note at the exact same time to create a perfect harmony. If even one person is slightly off, the harmony breaks, and the signal gets noisy.

For years, scientists have used a specific "recipe" (developed by researchers named Afek et al.) to create these quantum choirs. They thought this recipe was pretty good. But this new paper asks a simple question: "Is this recipe actually the best one, or is it just a convenient starting point?"

The authors used a computer program to act like a "smart tuner" that automatically adjusts the recipe to find a much better signal.

The Setup: The Quantum Kitchen

To make these quantum states, the researchers use a "kitchen" with two main ingredients:

1. Coherent Light: Like a steady, calm stream of water (a laser).
2. Squeezed Light: Like water that has been squished into a weird, wobbly shape to make it more sensitive.

They mix these two ingredients in a machine with two main mixing bowls (beamsplitters) and a few knobs to turn. The goal is to mix them perfectly so that when they come out the other side, they form that perfect "NOON state" choir.

The Problem: The Old Recipe Was "Good Enough"

The old recipe (the Afek method) set the knobs to specific positions based on math calculations done years ago. It worked, but it had two big problems:

1. It was too quiet: You had to wait a very long time to hear the signal because the "volume" (the number of successful measurements) was very low.
2. It wasn't perfect: The signal wasn't as clear as it theoretically could be.

For small groups of photons (2 or 3), the old recipe was okay. But as they tried to use bigger groups (4 or 5 photons), the recipe became very inefficient. It was like trying to bake a cake with a recipe that works for a cupcake but fails miserably for a wedding cake.

The Solution: The "Smart Tuner" (AI)

The authors built a computer model that can "learn." They didn't just guess new settings; they used a method called gradient descent (think of it as a hiker feeling their way down a mountain to find the lowest valley).

They let the computer tweak all eight knobs in their machine simultaneously. The computer's goal was simple: Maximize the information we get from every single photon.

The Results: A Massive Upgrade

When the "Smart Tuner" finished its work, the results were shocking:

• For 2 Photons: The signal got about 1.5 times louder. The old recipe was already pretty close to perfect, so there wasn't much room to improve.
• For 3 Photons: The signal got 8 to 9 times louder.
• For 4 Photons: The signal got 8 to 16 times louder.
• For 5 Photons: The signal got almost 18 times louder.

The "Volume" Analogy:
Imagine you are trying to hear a whisper in a noisy room.

• The Old Method: You have to stand there for 22 hours to be sure you heard the whisper correctly.
• The New Method: You only need to stand there for 22 minutes.

The computer found that by slightly changing how the light was mixed, they could get a much stronger signal without needing any new hardware. They just needed better settings.

The "Trade-Off" Surprise

There was one interesting twist.

• At 2 Photons: Improving the signal for one type of measurement made another type slightly worse. It was like turning up the bass on a stereo made the treble sound a bit muddy. The computer had to choose which one to prioritize.
• At 3, 4, and 5 Photons: The computer found a "sweet spot" where everything got better at the same time. It turned up the volume on all the channels simultaneously. This means that for larger experiments, you don't have to sacrifice one thing to get another; you can have it all.

Why This Matters (According to the Paper)

The paper claims that the old way of doing this (the Afek method) was significantly "suboptimal" (not the best possible) for larger groups of photons. By using this new, computer-optimized approach:

1. Experiments become practical: What used to take days of waiting in a lab can now be done in minutes.
2. Better sensitivity: The measurements are much more precise, getting closer to the theoretical limit of how good a measurement can possibly be.
3. It's real quantum magic: The authors checked the "Wigner function" (a way to map the shape of the quantum state) and confirmed that the improvements weren't just a trick of the math; the light itself became more "quantum" and strange, which is exactly what makes these measurements so powerful.

In Summary

The authors took a known method for creating super-sensitive quantum measurements, realized it was far from perfect, and used a computer to "re-tune" the machine. They found that for larger measurements, the old settings were holding the experiment back. By letting the computer find the perfect settings, they made the experiment 10 to 30 times faster and significantly more accurate, proving that the old "standard recipe" was just a starting point, not the finish line.

Technical Summary

Technical Summary: Quantum-Enhanced Single-Parameter Phase Estimation with Adaptive NOON States

Problem Statement
Quantum metrology aims to surpass the classical shot-noise limit ($\delta\phi \sim 1/\sqrt{N}$) in phase estimation, potentially reaching the Heisenberg limit ($\delta\phi \sim 1/N$) using maximally path-entangled NOON states. While the theoretical framework for NOON states is established, their experimental generation at high photon numbers ($N$) remains challenging. A practical approach proposed by Afek et al. utilizes linear optics to mix coherent and squeezed vacuum states, post-selecting on specific coincidence patterns to generate NOON-like interference fringes up to $N=5$. However, the parameters used in the Afek protocol (squeezing strength, coherent amplitude ratio, and beam splitter angles) were derived analytically for single-mode optimality at each $N$. It remains an open question whether this specific initialization represents a global optimum when considering the joint optimization of all circuit parameters, post-selection rates, and multi-channel sensitivity simultaneously. The primary experimental bottleneck in high-$N$ quantum metrology is the low post-selection rate, which necessitates prohibitively long integration times for statistically significant measurements.

Methodology
The authors present an end-to-end differentiable quantum-optical framework implemented using the Strawberry Fields photonic simulator with a TensorFlow backend. This approach enables the use of automatic differentiation to optimize circuit parameters via gradient descent.

• Circuit Architecture: The study models a two-mode linear-optical circuit comprising:

  1. State Preparation: A coherent state $|\alpha\rangle$ on mode 0 and a squeezed vacuum $|r\rangle$ on mode 1, with $\alpha = \sqrt{\gamma r}$.
  2. Input Phase Rotations: Rotations $R(d_{coh})$ and $R(d_{sq})$ applied before the first beam splitter.
  3. Probe Generation: A beam splitter $BS(\theta_1, \phi_1)$ creates the entangled probe state via Hong-Ou-Mandel interference.
  4. Phase Encoding: A phase shift $R(\phi_{est})$ applied to mode 0.
  5. Measurement Basis: A second beam splitter $BS(\theta_2, \phi_2)$ followed by photon-number-resolving detection.

  The system treats all eight parameters ($r, \log \gamma, d_{coh}, d_{sq}, \theta_1, \phi_1, \theta_2, \phi_2$) as trainable variables, whereas the original Afek protocol fixed most of these.

• Optimization Objective: The loss function maximizes the total Classical Fisher Information (CFI) across all coincidence channels $(N_1, N_2)$. The CFI is estimated using a differentiable spectral derivative method during training and validated against a high-accuracy ground-truth estimator (400-point numerical gradient). The optimization employs the Adam optimizer over 100 steps for $N = 2, 3, 4, 5$.

• Validation: The framework was validated by reproducing the Afek et al. coincidence fringes with a maximum normalized error $< 3 \times 10^{-4}$. Wigner function analysis was used to characterize the non-classicality of the probe states.

Key Contributions

1. Fully Validated Differentiable Model: The authors provide a numerically rigorous forward model for all Afek $N=2$–$5$ coincidence patterns, verified against exact Fock-space simulations.
2. Systematic Gradient-Based Optimization: This is the first study to simultaneously optimize all eight circuit parameters for $N=2$ through $5$, moving beyond the analytical single-mode constraints of previous work.
3. Discovery of Suboptimality: The study demonstrates that the Afek initialization point is significantly suboptimal for $N \ge 3$, with raw CFI improvements scaling dramatically with photon number.
4. Characterization of Inter-Channel Trade-offs: The work identifies and characterizes the trade-off structure between different coincidence channels, noting that while $N=2$ exhibits a strong trade-off, this weakens at higher $N$.

Results

• CFI Improvements: Gradient-based optimization yields substantial gains in raw CFI, which scale monotonically with $N$:
  • $N=2$: +153% (for the $|1,1\rangle$ channel).
  • $N=3$: +834% to +956%.
  • $N=4$: +829% to +1598%.
  • $N=5$: +1775%.
• Post-Selection Rates: The optimized parameters significantly increase the probability of detecting useful coincidence events. Improvements range from +153% to +3269% (specifically for the $|3,0\rangle$ channel at $N=3$). At $N=5$, the rate increases by a factor of $\sim 32$, transforming an experiment requiring hours of integration into one feasible within minutes.
• Fringe Quality and Heisenberg Limit:
  • At $N=2$, the optimized state reaches 101% of the normalized Heisenberg limit ($F^{norm}_{peak}/N^2 \approx 1.01$) for the $|1,1\rangle$ channel.
  • At $N=5$, the normalized CFI reaches 51.5% of the Heisenberg limit. While this is lower than at $N=2$, it represents a significant improvement over the Afek initialization (which was 36% at $N=5$).
  • The authors note that for $N \ge 3$, the optimizer often trades a small amount of fringe contrast (quality) for a massive gain in post-selection rate, a favorable exchange for experimental feasibility.
• Quantum Origin: Wigner negativity analysis confirms that the improvements are quantum in origin. For $N \ge 3$, the Wigner negativity volume of the probe state increases upon optimization, indicating a redistribution of the quantum state further from the classical region.
• Measurement Efficiency: The combined figure of merit (useful events per pulse, $\eta_\Sigma \times P_{sel}$) improves by factors of 8x ($N=2$) to 133x ($N=4$).

Significance
The paper claims that these results provide numerically rigorous benchmarks for adaptive single-parameter quantum sensing. The primary significance lies in demonstrating that the "convenient" parameters chosen in the original Afek protocol are far from optimal for $N \ge 3$. By utilizing variational methods to optimize the full parameter space, the study shows that quantum-enhanced sensing at higher photon numbers ($N \ge 3$) can be made experimentally practical by drastically reducing the integration time required for data acquisition. The work establishes that the inter-channel trade-off observed at low $N$ weakens at higher $N$, allowing for simultaneous improvements in sensitivity and detection rates across all channels. This paves the way for future experimental implementations of high-$N$ NOON-state metrology using optimized, rather than analytically derived, circuit parameters.