Blind Combination of Hadronic Cross-Section Measurements by Secret Sharing

The largest uncertainty in the Standard Model prediction for the muon's magnetic moment comes from one quantity, called hadronic vacuum polarization. It is computed by integrating measured rates of electron-positron collisions that produce hadrons, $R(s)$, against a known weight function that heavily favors low collision energies. Several colliders measure these rates: BABAR, KLOE, BESIII and CMD-3. Their results disagree with one another. The combined data-driven answer also disagrees with lattice QCD, the approach that simulates the strong force on a space-time grid. In the current workflow, each collaboration publishes its measurement once it is final, and phenomenologists then combine the published tables centrally. So by the time anyone sees whether a new measurement agrees with the others, the analysis is already unblinded. Its authors have seen where the answer lands.

We propose a protocol that lets collaborations do the combination while their results are still blind. Each experiment secret-shares its binned cross sections and error model. The shares are random-looking pieces that reveal nothing individually. The parties jointly compute the combined dispersive integral and a small set of tension statistics, and nothing else is ever revealed. The claim is narrow and testable. The heavy part of a correlated least-squares combination can be rewritten as a sum of per-experiment terms, so a simple secure sum plus a small secure linear solve reproduces the centrally computed answer well below experimental precision. The real research problem is therefore not cryptographic speed. It is deciding which released numbers leak too much about any one experiment's unpublished spectrum.

Where this sits relative to prior work

On the physics side, Phenomenological Estimate of Isospin Breaking in Hadronic Vacuum Polarization motivates the problem. It combines exclusive channels from different colliders through dispersion relations and splits the result into Euclidean-time windows to compare with lattice QCD. It shows that isospin-breaking corrections largely cancel and cannot explain the lattice-versus-data gap, so cross-experiment disagreement remains central. That analysis is done centrally on published tables. The BABAR ratio-method measurement of the two-pion channel (Precise Measurement of the e+e- → π+π-(γ) Cross Section with the Initial State Radiation Method at BABAR) and the BESII scans (R value measurements at 2.60, 3.07 and 3.65 GeV; Continuum R_uds between 3.650 and 3.872 GeV) supply the kinds of inputs we will re-enact as if they were private. The first uses a continuous spectrum; the second uses discrete points, the later scan at 68 energies. F. De Mori et al. (2006) constrains the design: sparse scans carry interpolation error. Each party must therefore map its data onto a common energy grid locally, with that interpolation uncertainty included in its own error model.

On the privacy side, coded computing encodes data into polynomials, so that any large enough subset of workers can recover a product while small colluding groups learn nothing. The basic scheme is Secure Distributed Computing With Straggling Servers Using Polynomial Codes. Later schemes lower worker counts, such as CAT and DOG, or add flexibility, such as Generalized Lagrange Coded Computing. All of these target large matrix products over finite fields and treat slow workers as the main cost. Coded Secure Multi-Party Computation for Massive Matrices with Adversarial Nodes adds tolerance of cheating workers through interactive verification. Analog Lagrange Coded Computing moves the encoding to complex floating-point numbers to avoid overflow. It reports relative errors of $10^{-4}$ to $10^{-5}$ on synthetic $X^TX$ products.

Our problem differs from all of these. It has only a handful of parties, small matrices, and an accuracy requirement set by physics rather than by throughput. We borrow two things from this literature. First, we use its secret-sharing primitives. Second, we borrow the Freivalds-style randomized check of Secure Private and Adaptive Matrix Multiplication Beyond the Singleton Bound to verify the one non-trivial step, the linear solve. We drop straggler coding, because it is not the bottleneck here.

Core Claim

Write the combination as a fit of common-grid values $x$ (the cross section in each energy bin) together with shared nuisance parameters $\theta$. The nuisance parameters are quantities, such as common radiative-correction uncertainties, that shift several experiments at once. Treating cross-experiment correlations this way, instead of through one joint covariance matrix, makes the fit's normal equations additive:

$$\Big(\sum_i F_i\Big)\,(x,\theta) = \sum_i b_i ,$$

Here $F_i$ and $b_i$ are experiment $i$'s local information matrix and data vector, built from its own measurements, covariance and projection onto the grid. Each party computes $F_i$ and $b_i$ privately and secret-shares them. The parties add the shares, solve this small system on the shares, and verify the solve with a randomized matrix-vector check. They then release only a few scalars:

  • the integral $a_\mu^{\text{HVP}} = \sum_j K_j x_j$, restricted to the measured range, where $K_j$ is the dispersion weight integrated over bin $j$;
  • its short-distance, intermediate and long-distance window pieces;
  • the global fit quality;
  • one pull per experiment.

The novelty lies in the problem decomposition and in a leakage analysis of the released outputs. This analysis matters because with only two experiments covering a bin, releasing combined bin values would let each side back out the other's numbers. We make no claim to a new code construction.

Research Program

The work has three parts.

Reformulation and error budget. We derive the additive nuisance-parameter form for correlated systematics as they appear in ratio-method ISR measurements and in direct scans. We then check against a plaintext joint-covariance fit that the two forms agree exactly when the nuisances are Gaussian.

Arithmetic. We implement the secure sum and secure solve in two ways. The first is fixed-point secret sharing over a 128-bit ring. The second is real-valued masking in the style of Analog Lagrange Coded Computing, with evaluation points on the unit circle. We measure the arithmetic error of each against the plaintext answer. The dispersion weight rises steeply toward threshold, so we report the error per energy window, not only in aggregate.

Output leakage. Treat one party's spectrum as unknown to an adversary who holds everyone else's inputs and all released scalars. We measure how well the adversary can reconstruct that spectrum, in two ways: analytically, as the rank and conditioning of the map from the hidden spectrum to the outputs, and empirically, with the best linear reconstruction. We then search over release sets, such as more or fewer windows, or pulls with and without the fit-quality statistic. The goal is the richest set that keeps reconstruction error above that party's own statistical uncertainty.

Baselines

Plaintext central combination. The reference point is a trusted third party that receives all inputs and runs the ordinary correlated fit. This is how Phenomenological Estimate of Isospin Breaking in Hadronic Vacuum Polarization-style analyses proceed. It defines the target answer and has zero cryptographic cost. The secure protocol "wins" only by matching it. We require the deviation in every window to stay below one percent of that window's experimental uncertainty. We run it on identical inputs and the identical common grid.

Do nothing. The naive reference is separate publication followed by later combination. Its cost is not numerical. It forfeits the pre-unblinding consistency check, and we can only argue for that check, not measure it.

Fixed-point additive secret sharing. The strongest existing competitor is standard multi-party computation: additive shares over a large ring, with a generic secure linear solve. No reported result found on Lacuna for fixed-point MPC applied to cross-section combination. We give it the same input encoding, the same grid, and the same verification step. We expect it to be a serious contender, possibly the winner, on accuracy.

Analog coded masking. The coded-computing competitor is Analog Lagrange Coded Computing. Its reported relative error is $10^{-4}$ to $10^{-5}$ on synthetic products (Analog Lagrange Coded Computing). Against a sub-percent physics error this looks sufficient, but only if it holds after a linear solve. We run it with the same noise-to-signal leakage level as the ring scheme.

Ablation. The ablation removes the shared nuisance parameters and treats experiments as uncorrelated. This isolates whether the additive reformulation, rather than simple averaging, is what makes the secure answer agree with the plaintext correlated fit.

Why not just…?

Why not just use the published tables, since the data become public anyway? For final results, this is the right approach and we do not compete with it. The protocol only has value before unblinding. Nothing in the research file shows collaborations asking for such a tool. This is the proposal's weakest point. It is a question of demand, and our experiments cannot settle it. What we can show is that a blind check costs little and leaks provably little, which is the precondition for anyone adopting it.

Why not just compute the integral on one laptop? Speed is not the issue, and we say so. The computation here really is small. That is exactly why we drop straggler-resilient coding and keep only secret sharing and verification. The measured wall-clock cost is reported only to show it is negligible.

Why not just use standard secure computation instead of coded computing? This threat is real. Given the additive structure, a secure sum over a 128-bit ring plus a small secure solve is probably enough. If fixed-point sharing matches the plaintext answer and analog masking does not, we will report that as a negative result: in this regime analog coded computing adds error without adding anything useful. The contribution then rests on the reformulation and the leakage analysis, which do not depend on the arithmetic choice.

Why not add differential-privacy noise to the outputs? Any noise large enough to hide a single experiment's spectrum would swamp the permille-level differences that the tension is about. The leakage study instead controls what is released rather than blurring it.

Proposed Experiments

We re-enact a blind combination using published inputs treated as private. The two-pion ISR spectrum from BABAR ( Precise Measurement of the e + e − → π + π − ( γ ) Cross Section with the Initial State Radiation Method atBABAR) serves as one party. The BESII points at 2.60, 3.07 and 3.65 GeV and the 68-point scan serve as others. The 2–5 GeV BES data (J.Z. Bai et al. (2002)), with uncertainties of about 6%, provide the high-energy tail. Where published covariances are coarse, we build synthetic error models with realistic shared normalization and radiative-correction nuisances. We also inject a deliberate 1% shift into one party to test whether the released pulls detect a planted tension.

Three checks are decisive:

  1. Precision. Does the secure combination reproduce the plaintext window values to better than one percent of their uncertainties under both arithmetics?
  2. Planted tension. Does the per-experiment pull flag the planted shift at the same significance as the plaintext fit?
  3. Leakage. With the chosen release set, does the best reconstruction of any single party's spectrum from the other parties' inputs and the outputs have error at least as large as that party's own statistical errors?

If the first check fails for fixed-point sharing, the approach is dead. If the third check fails for every useful release set, the right conclusion is that a blind combination with this few experiments cannot be made private, and we will publish that bound.

References

  1. Phenomenological Estimate of Isospin Breaking in Hadronic Vacuum Polarization
  2. Precise Measurement of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo stretchy="false">(</mml:mo><mml:mi>γ</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>Cross Section with the Initial State Radiation Method atBABAR
  3. R value measurements for <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math> annihilation at 2.60, 3.07 and 3.65 GeV
  4. Measurements of the Continuum<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>R</mml:mi><mml:mrow><mml:mi>u</mml:mi><mml:mi>d</mml:mi><mml:mi>s</mml:mi></mml:mrow></mml:msub></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>R</mml:mi></mml:math>Values in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math>Annihilation in the Energy Region between 3.650 and 3.872 GeV
  5. Secure Distributed Computing With Straggling Servers Using Polynomial Codes
  6. CAT and DOG: Improved Codes for Private Distributed Matrix Multiplication
  7. Generalized Lagrange Coded Computing: A Flexible Computation-Communication Tradeoff for Resilient, Secure, and Private Computation
  8. Coded Secure Multi-Party Computation for Massive Matrices with Adversarial Nodes
  9. Analog Lagrange Coded Computing
  10. Secure Private and Adaptive Matrix Multiplication Beyond the Singleton Bound
  11. Precise Measurement of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo stretchy="false">(</mml:mo><mml:mi>γ</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>Cross Section with the Initial State Radiation Method atBABAR
  12. Measurements of the Cross Section for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">e</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">e</mml:mi></mml:mrow><mml:mrow><mml:mo>−</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mi/><mml:mo>→</mml:mo><mml:mi/><mml:mi>Hadrons</mml:mi></mml:math>at Center-of-Mass Energies from 2 to 5 GeV