Mitigating photon loss in linear optical quantum circuits

This paper introduces a novel postprocessing framework using "recycled probabilities" to mitigate photon loss in linear optical quantum circuits, demonstrating that it can outperform the standard postselection method by achieving lower combined bias and statistical errors, whereas zero-noise extrapolation fails to offer similar improvements.

The Gist

Here is an explanation of the paper "Mitigating photon loss in linear optical quantum circuits," translated into simple, everyday language with creative analogies.

The Big Picture: The Leaky Bucket Problem

Imagine you are trying to build a complex machine out of water balloons (photons). You have a machine (a quantum circuit) that shuffles these balloons around through a maze of tubes (mirrors and beam splitters) to perform a calculation.

The problem? The tubes are leaky. As the balloons travel through the maze, some pop or fall out. In the world of quantum computing, this is called photon loss.

If you lose even one balloon, the calculation is ruined. Traditionally, scientists have dealt with this by using a strategy called Postselection.

• The Postselection Strategy: Imagine you run the experiment 1,000 times. In 999 of those runs, a balloon popped. You throw those 999 results in the trash. You only keep the one run where all balloons made it to the end.
• The Problem: As the machine gets bigger (more balloons, more tubes), the chance of all balloons surviving drops to near zero. You might have to run the experiment a billion times just to get one good result. It's incredibly slow and expensive.

The New Solution: "Recycling" the Trash

The authors of this paper propose a new family of techniques called Recycling Mitigation. Instead of throwing away the "failed" experiments (where balloons popped), they say: "Let's look at the trash!"

They realized that even the "broken" results contain hidden clues about what the "perfect" result would have been. By mathematically combining the data from the runs where 1 balloon was lost, 2 were lost, and so on, they can reconstruct the answer to the perfect experiment.

Think of it like this:

• Old Way (Postselection): You try to guess the recipe for a perfect cake by only eating the cakes that came out of the oven perfectly. If the oven is broken, you rarely get a perfect cake, so you starve.
• New Way (Recycling): You look at the burnt cakes, the undercooked cakes, and the ones with missing ingredients. You use a special mathematical "decoder" to figure out, "Okay, this burnt cake tells me the oven was 20 degrees too hot, and this missing-ingredient cake tells me we used too much flour." By combining these clues, you can reconstruct the recipe for the perfect cake, even if you never actually baked one.

How It Works: The "Recycled Probabilities"

The paper introduces a specific method to do this reconstruction:

1. Collect the Leaky Data: Instead of discarding the runs where photons were lost, they group them. They look at all the runs where exactly 1 photon was lost, all the runs where 2 were lost, etc.
2. Build "Recycled" Estimates: They take these groups and mathematically "stitch" them back together. They create a new set of numbers called Recycled Probabilities. These numbers are a mix of the real signal (what they wanted) and some "noise" (interference from the missing pieces).
3. The Post-Processing (The Magic Trick): They use classical computers to clean up this noisy data. They have two main tools:
   • Linear Solving: They treat the noise as a predictable pattern and subtract it out, like tuning a radio to remove static.
   • Extrapolation: They look at how the signal fades as more photons are lost. If they know the signal fades in a specific curve (like an exponential decay), they can mathematically "draw the line back" to where it would have started (zero loss).

Why Is This Better?

The paper proves two main things:

1. It's Faster: For a certain range of "leakiness" (which is common in real-world quantum devices), this recycling method gives you a more accurate answer much faster than waiting for a perfect run. You don't have to wait for the millionth trial to get a result; you can get a good result from the first thousand trials by using the "trash" data.
2. It's Smarter than "Zero-Noise Extrapolation" (ZNE): There is another popular method called ZNE, which tries to guess the answer by running the experiment with even more noise and then guessing backwards. The authors show that for photon loss, ZNE is actually a bad idea. It's like trying to guess the shape of a circle by looking at a square that's been stretched; it doesn't work well. Their "Recycling" method is the superior tool for this specific job.

The Catch (Bias vs. Variance)

There is a small trade-off.

• Postselection is perfectly accurate (unbiased) but takes forever (high variance).
• Recycling is slightly "fuzzy" (biased) because it's an estimate, but it's very precise because it uses so much data (low variance).

The authors show that for a long time (up to a massive number of samples), the "fuzziness" of the recycling method is actually better than the "waiting time" of the postselection method. It's like getting a slightly blurry photo that you can see right now, versus waiting 10 years to get a perfect photo that you might never get.

The Bottom Line

This paper is a guide for how to stop wasting data in quantum computing. Instead of throwing away the "failed" experiments because of photon loss, we can use clever math to recycle that information. This allows us to run larger, more complex quantum calculations on current, imperfect hardware, bringing us one step closer to the era of useful quantum computers.

In short: Don't throw away the broken pieces; use them to build the whole picture.

Technical Summary

Here is a detailed technical summary of the paper "Mitigating photon loss in linear optical quantum circuits" by James Mills and Rawad Mezher.

1. Problem Statement

Discrete Variable Linear Optical Quantum Computing (DVLOQC) is a promising framework for quantum advantage (e.g., Boson Sampling) and near-term applications. However, the scalability of these devices is severely limited by photon loss.

• Current Standard: The standard method to handle photon loss is postselection, where experimental runs with any lost photons are discarded, and only runs where all $n$ photons are detected are kept.
• Limitation of Postselection: As the circuit depth or number of modes ($m$) increases, the probability of losing at least one photon approaches 1 exponentially. Consequently, the sampling cost to obtain a single valid postselected sample scales exponentially with the circuit size, rendering large-scale computations infeasible.
• Goal: The authors aim to develop classical postprocessing techniques that utilize "lossy" data (runs where photons were lost) to reconstruct ideal (lossless) output probabilities or expectation values, outperforming the sampling efficiency of postselection without requiring additional quantum resources.

2. Methodology: Recycling Mitigation

The authors propose a family of techniques collectively called Recycling Mitigation. The core idea is to construct "recycled probabilities" from output statistics where $k$ photons were lost ($k > 0$) and use them to estimate the ideal $n$-photon distribution.

A. Construction of Recycled Probabilities

For a target ideal bit string $s_l^n$ (representing $n$ photons), the authors define a mapping to a set of lossy bit strings $s_i^{n-k}$ (representing $n-k$ photons).

• Mapping: An ideal output $s_l^n$ can map to multiple lossy outputs $s_i^{n-k}$ depending on which $k$ photons were lost.
• Recycled Probability ($p_R^k$): The recycled probability for a specific ideal string is constructed by summing the probabilities of all corresponding lossy strings, normalized by a combinatorial factor.
  $$p_R^k(s_l^n) = \frac{1}{\binom{m-n+k}{k}} \sum_{s_i^{n-k} \in L(s_l^n)} p(s_i^{n-k})$$
• Decomposition: Theoretical analysis shows that $p_R^k(s_l^n)$ can be decomposed into:
  1. A signal term proportional to the ideal probability $p(s_l^n)$.
  2. An interference term consisting of a mixture of other ideal probabilities.
     Crucially, the signal of the ideal probability is amplified relative to the interference term by a factor dependent on $k$.

B. Classical Postprocessing Techniques

To extract the ideal probability from the recycled probabilities, the authors introduce two main classical postprocessing methods:

1. Linear Solving:

   • Assumes the interference term behaves like a uniform distribution (or a linear function of the ideal probability).
   • Substitutes the interference term with its expected value (derived analytically) and solves the resulting linear equation for the ideal probability.
   • Includes a "dependency" variant that estimates the correlation between the interference term and the ideal probability to improve accuracy.

2. Extrapolation (Linear and Exponential):

   • Utilizes recycled probabilities constructed at multiple values of $k$ (e.g., $k=1, 2, \dots$).
   • Observes that the ideal signal decays as $k$ increases (as the distribution converges toward uniformity).
   • Fits the decay of the recycled probabilities to a model function (linear or exponential) and extrapolates back to $k=0$ to estimate the ideal probability.
   • Exponential Extrapolation is found to be superior in numerical simulations, as the decay of the ideal signal follows an exponential curve more closely than a linear one.

3. Key Contributions

• Recycling Mitigation Framework: A novel protocol that repurposes discarded lossy data to improve estimation accuracy, avoiding the exponential sampling overhead of postselection.
• Analytical Error Bounds:
  • The authors prove that these methods introduce a bias error (systematic error) in addition to statistical error.
  • They derive tight upper bounds on this bias error, showing it scales as $O((n/m)^n)$ for large $m$ and $n$, which is exponentially small in the system size for the Boson Sampling regime ($m \in \Omega(n^2)$).
  • They prove that for photon loss rates $\eta$ above a certain threshold ($\eta_{th}$), the combined error (bias + statistical) of recycling mitigation is lower than the statistical error of postselection for a wide range of sample sizes.
• Performance Comparison with ZNE:
  • The paper rigorously analyzes Zero-Noise Extrapolation (ZNE), specifically Richardson extrapolation, applied to photon loss.
  • Result: The authors prove that ZNE requires inverting a Vandermonde matrix. They demonstrate that the error amplification from inverting this matrix (due to statistical noise) is always larger than the error of simple postselection. Thus, ZNE offers no advantage over postselection for DVLOQC photon loss mitigation.
• Computational Hardness:
  • The authors argue that even with the introduced bias, the mitigated probabilities remain classically hard to compute (related to the $\#P$-hardness of estimating permanents of Gaussian matrices). This suggests that the quantum advantage is preserved despite the bias.

4. Results

• Numerical Simulations:
  • Simulations were performed on random unitary circuits with $m=20$ modes and $n=4$ photons.
  • Linear Solving: Outperformed postselection up to $\approx 1.1 \times 10^6$ samples (for $\eta=0.8$). With the dependency factor, this extended to $\approx 3.1 \times 10^6$.
  • Exponential Extrapolation: Outperformed postselection up to $\approx 3.0 \times 10^6$ samples.
  • The advantage holds for loss rates $\eta > 0.5$, which is typical for current photonic hardware.
  • The sample size where postselection eventually overtakes recycling mitigation scales as $\binom{m}{n}^2$, which is extremely large, making recycling mitigation practical for near-term devices.
• Bias Analysis:
  • Numerical evidence suggests the bias error is significantly smaller than the analytical upper bounds, often scaling as $O((n/m)^n)$.
  • The bias is small enough that the mitigated distribution retains the complexity properties of the ideal distribution.

5. Significance and Impact

• Practical Utility: This work provides a concrete, implementable method to extend the computational reach of current noisy photonic quantum devices. By utilizing data that would otherwise be discarded, it reduces the total number of experimental runs required to achieve a target precision.
• Benchmarking: It establishes a new benchmark for error mitigation in photonic systems. The authors demonstrate that simply discarding lossy data (postselection) is suboptimal and that classical postprocessing can significantly improve performance.
• Theoretical Insight: The paper clarifies the limitations of ZNE in the context of photon loss, distinguishing it from other error types where ZNE is effective. It also provides a deeper understanding of the statistical properties of lossy Boson Sampling.
• Applications: The techniques are directly applicable to various photonic quantum tasks, including Boson Sampling, Variational Quantum Eigensolvers (VQE), Quantum Machine Learning (QCBM), and solving graph problems.

In summary, the paper introduces Recycling Mitigation as a superior alternative to postselection for linear optical quantum circuits, offering a rigorous theoretical framework and strong numerical evidence that it can significantly reduce the sampling overhead required to observe quantum advantage in the presence of photon loss.