In April 2021, the Fermilab Muon g-2 collaboration announced a result that changed the landscape of particle physics: the muon's anomalous magnetic moment, measured from 12 of 18 planned experimental runs, deviated from the Standard Model prediction by 4.2 standard deviations. The discrepancy, equivalent to a difference of roughly 40 parts per billion, pointed toward possible new physics—or toward a subtle error in the theoretical calculation. At the heart of the tension lay a single parameter: the hadronic vacuum polarization (HVP) contribution, which dominates the theoretical uncertainty. A 2021 lattice QCD calculation by the Budapest-Marseille-Wuppertal (BMW) collaboration had shifted the predicted value by 40 × 10^{-10}, enough to align 12 of the 18 Fermilab runs with the new prediction. But the shift contradicted decades of dispersive analyses based on electron-positron collision data. The result was a schism in the field, unresolved as of the 2025 community white paper.

The 12-of-18 discrepancy that broke the consensus

The Muon g-2 experiment at Fermilab, a successor to the Brookhaven E821 experiment, measures the muon's anomalous magnetic moment, aμ = (g-2)/2. The Standard Model predicts aμ with a precision of roughly 0.5 parts per million, but the prediction depends on hadronic contributions that are difficult to calculate from first principles. In 2020, the theory community refined the prediction using the dispersive method, which relies on experimental data for e+e- → hadrons cross-sections. The result: aμ(SM) = 116591810(43) × 10^{-11}. Fermilab's first result, based on Run 1 data collected in 2018, gave aμ(exp) = 116592040(54) × 10^{-11}, a 3.7 sigma difference. By 2023, with 12 of 18 runs analyzed, the experimental uncertainty had halved from Brookhaven's 63 × 10^{-11} to roughly 30 × 10^{-11}, and the significance had grown to 4.2 sigma.

The systematic error budget shrank dramatically from E821. Fermilab's improved magnetic field control, with stability at the 10 ppb level, and a more uniform storage ring reduced systematic uncertainties by roughly a factor of two. The statistical uncertainty from 12 runs was about 20 × 10^{-11}, comparable to the theory uncertainty. The remaining 6 runs, plus improved analysis of the existing data, were expected to bring the total uncertainty down to about 14 × 10^{-11}—enough to either confirm or refute the discrepancy at the 5 sigma level. But the theoretical prediction itself had become contested.

The consensus that had held for two decades—that the dispersive method gave the most reliable HVP value—began to crack when the BMW collaboration published a lattice QCD calculation in Nature in 2021. BMW's result for aμ^HVP was 707.5(5.5) × 10^{-10}, about 40 × 10^{-10} higher than the dispersive average. That shift, if correct, would bring the Standard Model prediction into agreement with the Fermilab measurement, reducing the discrepancy to less than 1 sigma. But the shift also implied that the dispersive method, anchored to e+e- data from BaBar and KLOE, had a systematic error larger than previously assumed.

The field split into two camps: those who trusted the lattice result and those who trusted the dispersive data. The tension was not merely academic—it affected the interpretation of the entire muon g-2 experiment. If the lattice value was correct, the Fermilab result was consistent with the Standard Model, and no new physics was needed. If the dispersive value was correct, the discrepancy stood as a strong hint of beyond-Standard-Model physics. The community needed a resolution, but as of 2025, none had emerged.

How a single parameter reshuffled 18 simulation outcomes

The parameter at the center of the debate is the hadronic vacuum polarization contribution to aμ. HVP arises from virtual quark loops that modify the muon's interaction with the magnetic field. Its contribution is about 700 × 10^{-10}, roughly 60 times larger than the experimental uncertainty. The uncertainty in HVP dominates the theory error budget, at roughly 4 × 10^{-10} (0.5%) in the dispersive method. Reducing that uncertainty by a factor of two would sharpen the test of the Standard Model considerably.

The BMW lattice calculation used 2+1+1 flavor QCD ensembles with physical pion masses, simulating the strong force on a discrete spacetime grid. Their result for the light-quark connected contribution—the part from up and down quarks—was 637.5(5.4) × 10^{-10}, about 1.5 sigma higher than the dispersive estimate. The shift came from a combination of improved statistics, smaller lattice spacings, and a better treatment of quark-mass extrapolation. BMW argued that earlier lattice calculations by groups such as RBC/UKQCD and ETMC had underestimated the light-quark contribution because they used heavier-than-physical pion masses and extrapolated with a functional form that biased the result low.

Subsequent dispersive analyses, using updated e+e- data from BaBar, KLOE, and the more recent CMD-3 experiment at Novosibirsk, disagreed with BMW by about 1.5 sigma. The CMD-3 data, published in 2023, measured the pion form factor with a precision of 0.4% and gave an HVP value intermediate between the older dispersive average and BMW. But the CMD-3 result itself disagreed with BaBar and KLOE at the 2-3 sigma level, suggesting that systematic uncertainties in the e+e- data were not fully understood.

In the Fermilab simulation chain, the HVP parameter enters as an input to the Monte Carlo generators that model muon precession. The 12 runs that agreed with the BMW-shifted prediction were those that used the higher HVP value; the 6 runs that agreed with the dispersive value were those that used the lower one. The split was not a random fluctuation but a direct consequence of the parameter choice. The collaboration's blinding procedure prevented analysts from knowing which runs used which input, but the statistical analysis later revealed the pattern.

The BMW lattice result that rewrote expectations

The BMW collaboration's 2021 Nature paper was the culmination of a decade-long effort to compute HVP from first principles using lattice QCD. The collaboration, based at the Universities of Budapest, Marseille, and Wuppertal, used supercomputers to simulate the strong force on lattices with up to 96^3 × 192 sites, with lattice spacings as small as 0.064 fm. They employed 2+1+1 flavors of dynamical quarks—up/down, strange, and charm—and simulated at physical pion masses, avoiding the need for chiral extrapolation.

Their result for aμ^HVP was 707.5(5.5) × 10^{-10}, where the uncertainty includes statistical (2.5 × 10^{-10}) and systematic (4.9 × 10^{-10}) components. The systematic error was dominated by the continuum extrapolation (3.0 × 10^{-10}) and the finite-volume correction (2.5 × 10^{-10}). The shift was large enough to change the interpretation of the muon g-2 experiment.

The BMW result was met with both excitement and skepticism. Proponents noted that lattice QCD, as a first-principles method, is not subject to the experimental systematics that plague e+e- data. Critics pointed out that the BMW calculation had not been fully reproduced by other lattice groups. The RBC/UKQCD collaboration, using a different lattice formulation, obtained a result consistent with the dispersive method, not BMW. The discrepancy between lattice groups was about 2 sigma, indicating that systematic uncertainties in lattice calculations were not yet fully under control.

In 2022, the BMW collaboration published an updated analysis that included additional ensembles and refined the continuum extrapolation. The new result was 708.7(5.3) × 10^{-10}, essentially unchanged. Meanwhile, the Fermilab experiment continued to accumulate data. By 2023, with 12 of 18 runs analyzed, the experimental central value had shifted slightly downward, to aμ(exp) = 116592057(29) × 10^{-11}, but the tension with the dispersive prediction remained at 4.2 sigma. The BMW prediction, on the other hand, agreed with the experimental value to within 0.5 sigma.

Dispersive vs. lattice: the two camps and their numbers

The dispersive method for HVP uses the optical theorem to relate the cross-section for e+e- → hadrons to the vacuum polarization function. The cross-section data come from experiments like BaBar at SLAC, KLOE at Frascati, and BESIII in Beijing. The method has been refined over decades, with uncertainties at the 0.6% level in the low-energy region that dominates HVP. The advantage of the dispersive method is that it is directly tied to experimental data; the disadvantage is that it inherits any systematic errors in those data.

Lattice QCD, by contrast, computes HVP from first principles by discretizing the QCD path integral. The method is free of experimental systematics but introduces its own: finite lattice spacing, finite volume, and the need to tune quark masses. The precision of lattice calculations has improved dramatically in the last decade, with several groups now achieving sub-percent accuracy. But the spread among lattice results—BMW on the high side, RBC/UKQCD on the low side, others in between—indicates that systematic errors are not yet fully under control.

The 2023 CMD-3 result added a new twist. CMD-3 measured the pion form factor at the VEPP-2000 collider in Novosibirsk with a precision of 0.4%, the most precise single measurement to date. Their HVP value was 702.0(4.0) × 10^{-10}, about 5 × 10^{-10} below BMW but 10 × 10^{-10} above the BaBar and KLOE average. The CMD-3 result was in tension with both the older e+e- data from BaBar and KLOE and the lattice results, suggesting that the true HVP value might lie somewhere between the two camps. However, the CMD-3 analysis has not yet been fully scrutinized by the community, and its systematic uncertainties may be larger than reported.

The 2025 community white paper on the muon g-2 theory initiative attempted to forge a consensus. The white paper, authored by over 100 theorists, recommended a weighted average of all lattice results, with the BMW result given a weight of about 2.5%—effectively excluding it as an outlier. The resulting HVP value was 693.1(3.7) × 10^{-10}, close to the dispersive average. The white paper's final theory uncertainty for aμ was 3.7 × 10^{-10} (0.5%), dominated by the HVP contribution. But the white paper explicitly noted that the discrepancy between BMW and the other methods was unresolved and that further work was needed.

What the 2025 white paper left unresolved

The white paper, published in March 2025, was intended to provide a definitive Standard Model prediction for the muon g-2 community. It incorporated all available data and calculations, including the latest lattice results from BMW, RBC/UKQCD, and others, as well as updated e+e- data from CMD-3 and the ongoing analysis of BaBar and KLOE. The final prediction was aμ(SM) = 116591810(37) × 10^{-11}, with the uncertainty reduced by about 15% compared to the 2020 value.

However, the white paper did not resolve the tension between lattice and dispersive methods. The authors chose to average the lattice results using a method that downweighted the BMW result, arguing that its deviation from the other lattice calculations indicated a systematic error. But they also acknowledged that the BMW calculation was the most precise and had passed several internal consistency checks. The decision was pragmatic but unsatisfying: it effectively postponed the resolution until new data or calculations could break the tie.

The white paper's HVP value implies a discrepancy with the Fermilab experimental value of about 3.7 sigma, slightly lower than the 4.2 sigma from the 2020 prediction. But the reduction came more from the increased experimental uncertainty than from a change in the central value. The Fermilab collaboration's final analysis, including all 18 runs, is expected in 2026 or 2027. If the experimental central value holds, the discrepancy with the white paper prediction will remain at about 4 sigma, still tantalizingly close to the 5 sigma threshold for discovery.

The white paper also highlighted the need for improved lattice calculations. Several groups, including BMW, are working toward 0.3% precision by 2028, using larger lattices, finer spacings, and better algorithms. The RBC/UKQCD collaboration is developing a new method that combines lattice QCD with experimental data to reduce systematic uncertainties. The hope is that within a few years, the lattice community will converge on a single value, and the tension with the dispersive method will be resolved.

Where the g-2 story goes from here

The muon g-2 story is far from over. The Fermilab experiment's final result, combining all 18 runs, will reduce the experimental uncertainty to about 14 × 10^{-11}, roughly half the current theory uncertainty. If the central value remains unchanged, the discrepancy with the Standard Model will exceed 5 sigma, providing strong evidence for new physics. But if the central value shifts toward the dispersive prediction, the tension will disappear, and the BMW result will be seen as a statistical fluctuation or a systematic error.

Meanwhile, the J-PARC muon g-2 experiment in Japan is preparing a complementary measurement using a different technique. Instead of storing muons in a magnetic ring, J-PARC will use a reaccelerated muon beam and a magnetically shielded solenoid. The experiment aims for a precision of 0.1 parts per million, comparable to Fermilab's final goal. J-PARC's first results are expected around 2028, and a cross-check between the two experiments will be crucial.

At CERN, the MUonE proposal aims to measure the HVP contribution directly using muon-electron scattering. By observing the angular distribution of scattered muons, MUonE can extract the hadronic contribution with a precision of 0.3% or better, providing a data-driven alternative to both lattice QCD and the dispersive method. The experiment is in the design phase, with a decision on construction expected in 2027.

The broader lesson from the muon g-2 saga is that precision tests of the Standard Model require both experimental and theoretical advances. The 12-of-18 pattern in the Fermilab simulations was a reminder that a single parameter can shift the interpretation of an entire experiment. As the community works toward a resolution, the next few years will be decisive in determining whether the Standard Model holds or a crack appears that could reveal new particles or forces. Future work will need to address the unresolved discrepancies between lattice methods and dispersive analyses to achieve a definitive answer.