Lepton energy scale and resolution corrections based on the minimization of an analytical likelihood: IJazZ2.0

This paper introduces IJazZ2.0, a novel analytical likelihood-based method implemented in the IJazZ software that enables computationally efficient, unbiased, and robust simultaneous extraction of lepton (and photon) energy scale and resolution corrections across multiple categories by leveraging exact smearing treatments and automatic differentiation.

The Gist

Imagine you are a master chef trying to perfect a recipe for a giant, complex cake (the Large Hadron Collider, or LHC). To make sure your cake tastes exactly right, you need to measure the ingredients with extreme precision. But here's the problem: your kitchen scales (the particle detectors) aren't perfect. Sometimes they weigh a gram too much, sometimes a gram too little, and sometimes the numbers just jitter a bit.

If you don't fix these scale errors, your cake (the scientific data) will be ruined, and you might think you discovered a new flavor (a new particle) when it was just a measurement error.

This paper introduces a new, super-smart way to fix those kitchen scales using a tool called IJazZ2.0. Here is how it works, broken down into simple concepts:

1. The Problem: The "Fuzzy" Photo

In particle physics, scientists look at particles called leptons (like electrons and muons) that come from the decay of a Z boson. Think of the Z boson as a perfectly round, heavy ball that always weighs exactly the same amount (like a standard 1kg weight).

When this ball splits into two smaller balls (leptons), we know exactly how heavy they should be combined. However, our detectors are a bit "fuzzy."

• Scale Error: The detector might consistently say "1.01kg" when it's actually "1.00kg."
• Resolution Error: The detector might say "1.00kg" one second and "0.99kg" the next, just because of random noise.

Traditionally, to fix this, scientists would run thousands of computer simulations, randomly adding "noise" to the data millions of times to see what the average looks like. It's like trying to find the center of a dartboard by throwing darts blindfolded a million times. It works, but it takes forever and uses a lot of computer power.

2. The Solution: The "Magic Formula" (Analytical Likelihood)

The authors of this paper said, "Why throw darts blindly when we can use math to calculate exactly where the center is?"

They developed a mathematical formula (an analytical likelihood) that describes exactly how the fuzziness affects the data.

• The Old Way (Random Smearing): Imagine trying to smooth out a bumpy road by randomly filling in potholes with sand, one by one, and checking the road after every handful.
• The New Way (IJazZ2.0): Imagine having a blueprint that tells you exactly how much sand is needed to fill every pothole perfectly in one go.

Because this formula is "fully differentiable" (a fancy math term), it allows computers to use automatic differentiation. Think of this as a GPS that doesn't just tell you "you're lost," but instantly calculates the exact direction and speed you need to drive to get back on track. This makes the calculation hundreds or thousands of times faster than the old random method.

3. The "Relative Speed" Trick

One tricky part of the recipe is that the "fuzziness" changes depending on how fast the particles are moving.

• The Problem: If you group particles by their absolute speed (e.g., "all particles moving between 45 and 50 mph"), you run into a problem. Particles near the edge of that group might get misclassified due to the fuzziness, causing a "migration" bias. It's like sorting marbles by size, but the ruler is slightly wobbly, so small marbles get counted as big ones and vice versa.
• The Fix: The authors suggest sorting by relative speed (speed compared to the total weight of the pair). Instead of saying "45 mph," they say "45% of the total speed." This is like sorting marbles by how big they are relative to the box they are in. This prevents the "migration" errors and gives a much cleaner result.

4. Extending to Light (Photons)

The method isn't just for heavy particles; it also works for photons (light particles).

• The Challenge: In a specific type of decay (Z → µµγ), there is a photon involved. The math gets a bit trickier because the photon doesn't carry the whole "weight" of the system; it's just a fraction.
• The Innovation: The authors created a new variable (called $V_{DY}$) that acts like a "special ruler" just for this situation. It allows them to apply the same "Magic Formula" to light particles, ensuring the energy scale is perfectly calibrated even when the photon is just a small part of the puzzle.

5. Why This Matters

• Speed: What used to take days on a supercomputer now takes minutes on a laptop.
• Precision: It removes the "random noise" of the old simulation methods, giving scientists a clearer picture of reality.
• Stability: It's less likely to crash or give weird answers when the data is messy.

The Bottom Line

This paper presents a new, faster, and smarter way to calibrate the "scales" of the world's most powerful particle detectors. By replacing a slow, random guessing game with a precise mathematical blueprint, scientists can now measure the properties of the universe (like the mass of the Higgs boson) with greater confidence and less computing power. It's the difference between trying to find a needle in a haystack by shaking the hay, versus using a magnet that knows exactly where the needle is.

Technical Summary

1. Problem Statement

Precision measurements at high-energy collider experiments (such as the LHC) rely heavily on the accurate calibration of photon and lepton energy scales. Small biases in these scales directly impact critical measurements, such as the Higgs boson mass in channels like $H \to \gamma\gamma$ and $H \to ZZ^* \to 4\ell$.

Current calibration methods typically compare reconstructed dilepton invariant mass distributions (from $Z \to \ell\ell$ decays) between data and Monte Carlo (MC) simulations. However, existing approaches face significant challenges:

• Correlations: Scale and resolution parameters are highly correlated, complicating their simultaneous extraction.
• Computational Cost: Traditional methods often rely on "random smearing" techniques (convolving the MC distribution with random numbers to mimic detector resolution). This is computationally expensive and prevents the use of efficient gradient-based optimization algorithms because the likelihood function is not differentiable with respect to the parameters.
• Bias in Categorization: When categorizing events by lepton transverse momentum ($p_T$), kinematic correlations and category migration can introduce biases in the measured scale and resolution.

2. Methodology: The IJazZ2.0 Approach

The authors propose IJazZ2.0 (I Just AnalyZe the Z), a software tool that replaces random smearing with a fully analytical likelihood maximization.

Core Principles

• Analytical Smearing: Instead of generating random numbers for every event, the method models the energy smearing as an exact analytical Gaussian convolution. The probability of a simulated event falling into a specific mass bin is calculated using the error function ($\text{Erf}$).
• Differentiable Likelihood: Because the smearing is analytical, the negative log-likelihood (NLL) function is fully differentiable with respect to all calibration parameters (scale $r$ and resolution $\sigma$). This enables the use of automatic differentiation (via TensorFlow), allowing for exact gradient computation and rapid convergence.
• Parameter Definition:
  • $r_\ell(\vec{X})$: The relative energy scale correction applied to simulation to match data ($E_{corr} = r \times E_{raw}$).
  • $\sigma_\ell(\vec{X})$: The additional smearing required to degrade the simulation's resolution to match data.
• Binning Strategy:
  • Adaptive Binning: To avoid zero-probability bins and bias from low-statistics tails, the invariant mass spectrum is binned adaptively (equal number of events per bin) based on the Freedman–Diaconis rule.
  • Fine Simulation Binning: The simulation is binned very finely ($<0.2$ GeV) to accurately model event migration between bins due to scaling and smearing.

Handling $p_T$ Categorization

The paper identifies that categorizing by absolute lepton $p_T$ introduces biases due to the threshold effects on the invariant mass distribution (creating double-peak structures).

• Solution: A relative-$p_T$ strategy is introduced, using the variable $r_{pT} = p_T / m_{\ell\ell}$.
• Procedure:
  1. Fit for scale parameters using relative $p_T$ bins.
  2. Correct the data/simulation energies.
  3. Transform the relative bins back to absolute $p_T$ bins for the final resolution fit.
     This mitigates kinematic correlations and category migration biases.

Extension to Photons ($Z \to \mu\mu\gamma$)

The method is extended to photon energy scale measurement using $Z \to \mu^+\mu^-\gamma$ decays.

• A new variable, $V_{DY}$, is defined to represent the relative mass carried by the photon.
• $V_{DY}$ is sensitive to residual photon energy miscalibration. The analytical likelihood is adapted to fit the $V_{DY}$ distribution instead of the dilepton mass.
• An iterative fitting procedure is employed to correct for non-linearities in the $V_{DY}$ definition, ensuring rapid convergence to the true miscalibration value.

3. Key Contributions

1. Analytical Likelihood Formulation: The development of a closed-form analytical expression for the smeared invariant mass distribution, eliminating the need for stochastic random-number generation.
2. Computational Efficiency: By enabling automatic differentiation, the method achieves a CPU-time gain of 500x to 5000x compared to random smearing methods. Fits that take days with random smearing can be completed in minutes on a laptop or seconds on a GPU.
3. Robust Uncertainty Estimation: The framework includes an analytical method to propagate uncertainties arising from the finite size of the MC simulation sample, in addition to data statistical uncertainties.
4. Bias Mitigation Strategy: The introduction of the relative-$p_T$ categorization strategy effectively removes biases induced by kinematic correlations in $p_T$-dependent calibration.
5. Open-Source Tool: The implementation is released as IJazZ2.0 via PyPI, making the method accessible to the broader physics community.

4. Results and Validation

The method was validated using both "toy" Monte Carlo simulations and realistic Pythia-based Drell–Yan simulations:

• Parameter Recovery: The method successfully recovers injected scale and resolution parameters with zero bias within statistical precision.
• Uncertainty Accuracy: The predicted statistical uncertainties (including MC fluctuations) match the observed standard deviations from repeated simulations.
• Performance:
  • A fit with 100 parameters and $2 \times 10^7$ events takes 20–30 minutes on a laptop (or <1 minute on a GPU), whereas the random smearing equivalent would take several days.
  • The analytical prediction of the smeared distribution is indistinguishable from a Monte Carlo simulation with 10,000 random trials per event.
• Photon Calibration: The extension to $Z \to \mu\mu\gamma$ accurately retrieves injected energy shifts ($\delta r_\gamma$) and resolution smearing ($\sigma_\gamma$) with an accuracy better than $10^{-4}$.
• Limitations Analysis: The paper identifies and quantifies potential biases arising from large resolution mismodeling and steep $p_T$ spectra (migration effects). It provides analytical estimates for these biases and suggests mitigation strategies (e.g., iterative smearing of the simulation).

5. Significance

The IJazZ2.0 method represents a paradigm shift in collider calibration techniques. By moving from stochastic, non-differentiable methods to a fully analytical, differentiable framework, it offers:

• Scalability: Essential for the high-luminosity LHC era where calibration tasks involve massive datasets and complex, multi-dimensional categorizations.
• Precision: Enhanced numerical stability and the ability to use advanced optimization algorithms lead to more precise extraction of calibration constants.
• Versatility: The framework is general enough to handle both lepton and photon calibration, and adaptable to various detector conditions and kinematic variables.

This work provides a robust, efficient, and open-source solution for the critical task of energy scale and resolution calibration, directly supporting the precision physics goals of current and future collider experiments.