WorksPhysics and Astronomy
Phenomenology of neutrino mass ordering synergies in reactor, accelerator, and atmospheric neutrino experiments, and the role of the detector response
Abstract
The neutrino mass ordering is one of the main open questions of neutrino physics, addressed by reactor, accelerator, and atmospheric experiments and by their combination. When the data are fitted under the wrong ordering, each experiment returns its own value of , its wrong-ordering solution. We show that this solution is a weighted average of the values preferred by the individual bins of the measured event spectra, in reconstructed energy and, for atmospheric neutrinos, also zenith angle, with weights set by their statistical precision. Both the values and the weights depend on the detector response. In the plane of fitted under the two orderings, the solutions lie on lines of slope close to , so that each experiment is described by one intercept. The weighted bin average extends the mass ordering sum rule to any oscillation experiment, including atmospheric neutrinos in Earth matter. It recovers the known closed forms for short-baseline reactors and for accelerators in vacuum, and for JUNO it equals a closed form evaluated at one effective energy, which depends on the energy resolution. Above the MSW resonance the frequency of atmospheric disappearance becomes nearly independent of , and the disappearance solution moves mainly with . The atmospheric intercept depends on the energy and direction resolutions, the energy window, and the event classes, and differs between detectors. The synergy of two experiments is determined by the separation of their intercepts relative to their precisions on . With JUNO, this separation depends strongly on for an accelerator and is nearly independent of and for IceCube DeepCore using all channels.
Introduction
Reactor, accelerator, and atmospheric experiments determine the atmospheric mass splitting, , mainly from the disappearance of at reactors and of and in the other two [1, 2, 3, 4, 5, 6, 7, 8]. Each class of experiments is sensitive to the neutrino mass ordering (NMO) in its own way. JUNO [9, 10, 11] is designed to determine it from the interference of the two fast oscillations in disappearance at a medium baseline [12, 13, 14]. Accelerator experiments are sensitive through the matter effect in appearance [3, 4, 15], and atmospheric detectors through the Earth matter effect across the Mikheyev–Smirnov–Wolfenstein (MSW) resonance [16, 17, 18, 19].
Combining experiments adds a further source of information. Each experiment measures through the oscillation frequencies of its channels. These frequencies depend on the mass ordering in a way set by the channel, baseline, energy, and matter density of each experiment [18, 20, 21, 22]. An ordering analysis fits the data under both orderings and compares the two minima of the , and the fit under the wrong ordering measures how strongly the data disfavor it. In this fit each experiment therefore returns a different value of , which we call its wrong-ordering solution. In a joint fit under the wrong ordering a single value of has to describe all experiments, and their disagreement disfavors the wrong ordering beyond the intrinsic sensitivities of the individual experiments. This synergy in the determination of the NMO is known for reactors combined with accelerator [9, 14, 21, 23, 24, 25] and atmospheric experiments [26, 27, 28], where it has been evaluated both in phenomenological studies and in joint sensitivity studies of the experiments. Accelerator and atmospheric data have also been analyzed jointly [29]. Reactor data have been combined with accelerator data [1, 3, 30] and with atmospheric data [30, 31], and global fits combine all of them [32, 33, 34, 35].
We take the true ordering to be normal (NO), so that the wrong ordering is the inverted one (IO). Two experiments and that fit NO data under the IO hypothesis return their own solutions, and , the fitted values of , with uncertainties and . The ordering sensitivity of their combination is then [26]
The first two terms are the intrinsic sensitivities of the individual experiments. The third is the synergy term, which depends only on the two wrong-ordering solutions and on their precisions.
For reactor and accelerator disappearance in vacuum each channel has one fast oscillation, whose frequency is a combination of and that depends on the channel, and the wrong-ordering solution is known in closed form as a function of the splitting of the true ordering [20, 21, 23, 24, 25, 36, 37]. The disappearance of the accelerator experiments T2K [38] and NOvA [39] and the disappearance of the medium-baseline reactor JUNO were combined into a mass ordering sum rule [24]. The rule states that if the two channels measure the same splitting under the true ordering, their splittings under the wrong ordering differ by a quantity set by , the mixing angles, and , with the CP-violating phase. On the other hand, atmospheric neutrinos are observed over many baselines and energies, across the MSW resonance, where Earth matter effects are large. A single frequency does not describe these data, and the closed forms do not apply. For example, for the proposed atmospheric detector PINGU, numerical studies found that its wrong-ordering solution differs from the vacuum closed forms by an amount comparable to its precision on [26], and that Earth matter enlarges this difference [27].
In this work we describe the wrong-ordering solution of any oscillation experiment in a single way, as a weighted average of the values of preferred by the individual bins of the measured event spectra, in reconstructed energy and, for atmospheric neutrinos, also zenith angle, with weights set by the statistical precision of each bin. Both the values and the weights depend on the detector response. This weighted bin average recovers the closed forms and extends the sum rule to any oscillation experiment, including atmospheric neutrinos in Earth matter. Our main application is to atmospheric neutrinos. We show how Earth matter moves the atmospheric solution away from the closed forms, how this solution depends on the unknown and on the atmospheric mixing angle , and how it changes with the detector, with IceCube DeepCore [40] and KM3NeT/ORCA [41] as examples. For a medium-baseline reactor the same description gives the dependence of the solution on the energy resolution. We then compare the solutions of present reactor, accelerator, and atmospheric experiments, which set the synergy of their combinations.
Section 2 derives the weighted bin average, section 3 recovers the reactor and accelerator closed forms and gives the dependence of JUNO on its energy resolution, section 4 treats atmospheric neutrinos with DeepCore as the main example, compares it with the first six detection units of KM3NeT/ORCA (ORCA6), and tests it against a full fit, section 5 compares the solutions of current experiments with each other and with their published best fits, and section 6 discusses the synergy of combinations. Section 7 summarizes the results. The inputs of each experiment, namely exposures, spectra, and detector responses, are described in appendix A, and the derivations and numerical checks in the other appendices. The code and the inputs that reproduce all figures and numbers of this paper are public [42].
The wrong-ordering solution as a weighted bin average
Consider an experiment that measures event counts in bins of its observables, for example the reconstructed energy for a reactor or accelerator experiment, or the reconstructed energy and zenith angle for an atmospheric detector. Each bin has an unoscillated expectation and an oscillation probability , averaged over the true energies and directions of the events reconstructed in the bin, for the mass ordering , so that its expected count is . The ordering is determined by fitting the same data under both hypotheses and comparing the two minima of the . To evaluate sensitivities we use Asimov data [43], the expected event counts without statistical fluctuations. On NO Asimov data the NO fit reproduces the data exactly, with at the true parameters, so the ordering sensitivity is the minimum of the fit under the IO hypothesis, which we call the IO fit and on which we focus.
Unless stated otherwise, the true values of the oscillation parameters are the NuFIT 6.0 best fit for normal ordering [32], listed in appendix A.1, and the unknown and are varied over their ranges where indicated. The solar parameters and and the reactor angle are fixed to these values in both orderings and in all experiments. Without superscript, and denote the true values. The IO fit adjusts and, where stated, also its own mixing angle and CP phase, and . Where they are not adjusted, equals the true and is the NuFIT 6.0 IO best fit, .
For NO Asimov data at the true splitting , the IO fit minimizes
with and the inverse Poisson variance of the measured probability . The IO fit returns the value of whose predicted counts are closest to the NO data, in the sense of eq. (2). Each term equals , twice the Poisson log-likelihood ratio to second order in [44, 45]. Although the fit searches for a common minimizing the total , we define to be the value “preferred” by bin , that is, the IO splitting that reproduces the NO data of that bin when is expanded linearly around the solution ,
It equals the exact solution of when is close to linear between and , that is, when the bin is not near an oscillation maximum or minimum, where vanishes. With this definition the condition becomes
The wrong-ordering solution is therefore the average of the values preferred by the individual bins, each weighted by the inverse variance with which it determines the splitting. We call this average the weighted bin average. Bins in which varies most steeply with , that is, with the largest , dominate, while bins at oscillation maxima or minima, where vanishes, do not contribute. Since and are evaluated at , eq. (4) is a condition that satisfies at the minimum, exact for the of eq. (2). It does not replace the fit. It splits the solution into the contributions of the bins, so that the solution for a subset of the bins, or for the same bins with other weights, follows to first order from the same without a new fit. It has the structure of any one-parameter least-squares minimum [46], written so that the contribution of each bin is explicit, and separates what each bin prefers, , which is set by the oscillation physics and by the averaging of the probability over the bin, from its weight in the average, , which also depends on the detector and the event rate.
The solution depends on the true splitting . In vacuum the oscillation frequency measured in each channel lies below in the NO and above in the IO, by a term proportional to that does not depend on [20, 21]. The IO fit matches these frequencies, so it returns , with independent of and positive.
Other effects, such as Earth matter effects, make depend slightly on , but it stays positive for every experiment considered here. We therefore look for solutions of the form
lines of slope close to in the plane of the two splittings, and describe each solution by its intercept . Since is negative, . Close to the measured value of each experiment is therefore described by a single number, its intercept . We assign to each bin the intercept , and the intercept of the experiment is their weighted average,
When all coincide the weights drop out of the average and is determined by the oscillation parameters alone, as for short-baseline reactors and accelerators in vacuum, whose solutions are known in closed form, as discussed in § 3. When the bins prefer different values, as for a medium-baseline reactor and for atmospheric neutrinos in Earth matter, depends on how the detector distributes the weights.
Eq. (4) holds for bins of any channel, and for bins that contain events of several channels. An appearance bin enters the average in the same way as a disappearance bin, with its own and . In current accelerator experiments the disappearance bins carry most of the weight and set the solution, while the appearance bins mainly determine the CP phase of the IO fit. In atmospheric detectors a bin contains and charged-current and neutral-current events together with the charged-current events. These events add to the expected count of the bin, which changes its and and can move the solution.
When the IO fit adjusts only , we obtain the solution by solving eq. (4) iteratively, because and depend on . This iteration is the Gauss–Newton minimization of the of eq. (2), so that the weighted bin average and the minimum of the coincide. When the IO fit also adjusts other parameters, such as its mixing angle and CP phase, eq. (4) holds at their fitted values. Appendix § 11 gives the derivation of eq. (4), its form with the orderings exchanged and with nuisance parameters, and the minimization.
The detector response, that is, how the detector maps the true energies and directions of the neutrinos onto the analysis bins, enters eq. (4) in two ways. First, the probability is averaged over the true energies and directions of the events reconstructed in the bin, so the energy and direction resolutions determine which oscillation phases are averaged together. Second, the weights depend on the event rate and on how sharply each bin resolves the oscillation pattern.
Figure 1 shows an example, the weights and the values of DeepCore, computed with the detector response of its public data release, described in § 4. About two thirds of the weight lies between 20 and 50 GeV for up-going neutrinos, in bins that prefer nearly the same value. Below 20 GeV, the bins whose neutrinos cross the core or the deep mantle carry about a sixth of the weight and prefer a much smaller , that is, a larger bin intercept . The intercept of the experiment therefore lies above the intercepts of the bins that carry most of the weight.
Figure 1. The weighted bin average for IceCube DeepCore at the NuFIT 6.0 best fit, on the response of the public data release ( disappearance, with and the CP phases fixed at their NuFIT 6.0 values). (a) Share of the total statistical weight carried by each reconstructed bin. (b) The value preferred by each reconstructed bin, eq. (3), where a smaller is a larger bin intercept (bins with negligible weight in gray, color scale centered on the solution and saturated). Dashed and dotted reference lines indicate the core–mantle boundary () and the horizon ().
The of eq. (2) is a simplified version of the one an experiment minimizes. It fits only the oscillation parameters and contains no systematic uncertainties. Evaluating it on Asimov data is the standard choice for sensitivity studies, and the obtained in this way approximates the median outcome of real data [47, 48, 49]. Appendix § 14 compares this reduced with a full fit of the DeepCore public data. The weighted bin average keeps its form when systematic uncertainties are included through nuisance parameters that are profiled in the fit [50]. As derived in appendix § 11.§ 4, to first order the values preferred by the bins do not change, and the nuisance parameters only redistribute the weight among the bins. They therefore move the solution only through the spread of the , which vanishes for short-baseline reactors and accelerators in vacuum and is set by the solar phase for JUNO and by Earth matter for atmospheric detectors. If , fixed in this work, were fitted with an external prior, such as that of Daya Bay, it would act in the same way.
Reactor and accelerator experiments: the closed forms as limits
Values preferred by the bins in vacuum
By eq. (6) the intercept of an experiment is the weighted average of the bin intercepts . For reactor and accelerator experiments matter effects are small, so we first derive in vacuum. Where all coincide, the weights drop out of the average and a closed form follows. In vacuum the survival probability of flavor can be written with one fast oscillation [21, 25, 36, 37],
where are the squared moduli of the elements of row of the lepton mixing matrix , , and the sign is for the normal and for the inverted ordering. Three quantities describe the fast oscillation. Its frequency is the effective splitting
with phase . Its amplitude , given in appendix § 12, is close to one when the solar oscillation has not developed. Its phase is shifted by the solar phase , which enters with opposite signs in the two orderings and grows with [36, 51]. The derivations of this section are in appendix § 12.
For a bin described by a single energy , the IO probability equals the NO one when the phases of the fast oscillation coincide [20, 21, 22, 36, 52], , where we define . With and , solving this equality of the phases for gives the value preferred by each bin in vacuum,
where is evaluated with and of each ordering. This value depends only on the oscillation probability of the bin and not on its weight. The bin intercept is a term common to all bins, set by the mixing, minus the solar-phase term of the bin. Inserting eq. (9) into eq. (4) gives the solution of every vacuum case. When is negligible in every bin, all coincide, the weighted average returns their common value for any choice of weights, and the detector does not affect the solution.
Accelerators
For accelerator disappearance () at long baselines (LBL), is negligible [21]. All coincide, and eq. (4) reduces to
where carries the leading dependence on the CP phase [21, 24], with and . We call , the leading part of the intercept that depends on the CP phases, the CP term. From eq. (8), the exact coefficient is
and the second relation in eq. (10) is its expansion to first order in , which makes the dependence on explicit. All numerical results use the exact form. With the NuFIT 6.0 values, and in both orderings, the intercept of eq. (10) is
where is the true phase and that of the IO fit.
The left panel of Figure 2 compares the weighted bin average with this closed form for an 810 km beam with equal neutrino and antineutrino event samples, against the true . The two agree over the full range of . The band shows the range of the intercept over all values of , which disappearance alone does not determine. A change of changes , the IO fit compensates it with a change of , and every value of gives nearly the same . The is therefore nearly constant along a curve in the plane of and , which we call the flat valley. A fit of the NOvA disappearance sample alone indeed gives a flat posterior of in both orderings [54]. The intercepts of the left panel include the matter effect of the Earth’s crust, which shifts the splitting measured in disappearance slightly [55]. The matter potential changes sign between neutrinos and antineutrinos, and we find that it shifts the accelerator solution in opposite directions for the two. With equal neutrino and antineutrino event samples the two shifts cancel in the weighted average. With neutrinos alone they do not cancel, and the solution moves from the closed form by a few at the baseline of NOvA and by less at that of T2K. Beams that also run with antineutrinos give smaller shifts. All these shifts are small compared with the range spanned by .
Figure 2. Left: accelerator disappearance (810 km, , Earth’s crust, equal and event samples). The intercept from eq. (4) (points) against the closed form (10) (line), as a function of the true with at its NuFIT 6.0 value. The points and the line agree to . The shaded range covers all values of , which give nearly equally good solutions in disappearance. Right: JUNO, eight reactor cores, in matter. The intercept from eq. (4) as a function of the stochastic term of the energy resolution, with the constant and noise terms scaled with it to the values of the JUNO simulation at [53] (open diamond). The short-baseline intercept and eq. (14) with the solar-phase term evaluated at 4 MeV in place of its weighted average (dashed) are shown for reference. Oscillation parameters from NuFIT 6.0, with .
Reactors
For reactor disappearance (), in both orderings and does not depend on .
At short baselines, such as those of Daya Bay [2], the solar oscillation has not developed, the solar phase is negligible, and all coincide. The solution of a short-baseline reactor (SBR),
then follows from eq. (4) for any choice of weights.
At a medium baseline the solar oscillation is developed and depends on energy [12, 13, 56]. The bins prefer different values, as found for the energy ranges of JUNO in ref. [57], and eq. (4) gives the solution of a medium-baseline reactor (MBR),
The medium-baseline solution is the short-baseline one minus the solar-phase term averaged with the weights of the bins, so that it depends on the energy resolution and on the energy spectrum of the detected events.
The right panel of Figure 2 shows the intercept of JUNO [11, 53, 58] from eq. (4) as a function of the stochastic term of the energy resolution, together with the short-baseline intercept and a closed form with the solar phase at a single energy. Our model of JUNO includes its eight reactor cores and the small matter effects of the crust [59, 60]. The solar-phase term grows as the solar oscillation develops [36, 51], and its exact form is given in appendix § 12. At the baseline of JUNO it is largest near 2.5 MeV, close to the maximum of the solar oscillation at about 3 MeV, and it falls steeply at higher energies, where the solar oscillation is less developed. At low energy the fast oscillation also has the shortest period in energy. A worse energy resolution averages the fast oscillation in these bins first, which reduces their derivative and therefore their weight . The bins at higher energies, where the solar-phase term is small, then carry a larger share of the weight, and the solution moves toward the short-baseline one. At the energy resolution predicted by the full simulation of JUNO [53] and used in its ordering sensitivity study [11], , which JUNO quotes as 2.95% at 1 MeV [53], the weighted bin average gives . Between and 4% the intercept changes by about . The ordering sensitivity of JUNO is known to depend on the energy resolution [11, 14, 52], and so does its wrong-ordering solution, by small amounts compared with its separation from the accelerator and atmospheric solutions. A combined study of JUNO and ORCA also found that the JUNO best fits in the two orderings change little with the resolution [28].
Closed forms in the literature describe the solar-phase term by a single shift of the fitted splitting, for example of or of , the effective splitting of eq. (8) for , evaluated at a representative energy of about 4 MeV [23, 52] or obtained by matching the spectra of the two orderings [24, 36, 61]. Above the maximum of the solar-phase term, decreases with energy, so the weighted average equals its value at one effective energy, which depends on the resolution. It is about 4 MeV at , so a closed form at 4 MeV reproduces eq. (14) only at the resolution for which the effective energy is 4 MeV. The width of the energy bins matters little as long as the bins are narrower than the resolution, as for the 20 keV bins used here.
In a full fit, systematic uncertainties such as the nonlinearity of the energy scale and the shape of the reactor flux are mildly correlated with [11, 58]. They redistribute the weight among the bins, and their effect on the intercept can be evaluated with the full systematic model of the experiment [52, 61], but it is not included in this work.
In summary, the solutions of a short-baseline reactor and of an accelerator in vacuum are determined by the oscillation parameters alone, while the solution of a medium-baseline reactor also depends on its energy resolution. The difference between the intercepts of the accelerator and of the reactor is the difference between the wrong-ordering splittings of the two channels in the mass ordering sum rule [24].
Atmospheric neutrinos: matter effects and detector response
Atmospheric neutrinos reach the detector from all directions, along baselines from a few tens of kilometers for down-going neutrinos to the diameter of the Earth for up-going ones, with energies from a few GeV to hundreds of GeV. Along the paths that cross the Earth, neutrinos in the NO and antineutrinos in the IO undergo the MSW resonance of , where the mixing angle in matter becomes maximal, at about 6–10 GeV in the mantle. These matter effects differ between the two orderings and give atmospheric detectors their own sensitivity to the NMO [8, 18, 62, 63, 64, 65, 66, 67, 68, 69].
All atmospheric probabilities are computed with three flavors through the Preliminary Reference Earth Model (PREM) [70], with the algorithm of NuFast-Earth [71]. Unless stated otherwise, § 4.1 to § 4.3 use disappearance with an IO fit that adjusts only , which isolates the effect of matter on the frequency. In this reference configuration the intercept of DeepCore is . The descriptions of DeepCore and ORCA6 are in appendix § 9, the derivations of this section in appendix § 13, and the comparison with the full fit in appendix § 14.
Matter shift of the values preferred by the bins
Without matter effects, each bin of an atmospheric sample behaves like an accelerator measurement at its own baseline and energy, as described by eq. (9). The solar phase is negligible at the energies that carry the weight in DeepCore, above about 10 GeV, so that all coincide and eq. (4) returns the accelerator closed form of eq. (10). With matter the bins prefer different values. We define the remainder of each bin through
where is the vacuum value without the CP term, to first order in . The solution is then this vacuum value plus the weighted average of the remainders,
The remainder contains the matter effects, the CP term, the solar phase, and, when they are included, the contribution of the channels other than disappearance. In Figure 1(b) the values preferred by the DeepCore bins spread over a wide range. This is the spread of the , and their weighted average raises the intercept above .
Above the resonance, has a closed form. In matter the component of the mass eigenstates is suppressed above the resonance, but a small admixture remains. This admixture shifts the effective oscillation frequency of at second order in [72, 73, 74], in opposite directions for neutrinos and antineutrinos and for the two orderings. Matching the shifted frequencies of the two orderings gives the remainder of each bin, which we derive in appendix § 13.§ 1,
with for neutrinos and for antineutrinos, to leading order in . Here is the Fermi constant, and and are a true neutrino energy and the electron density averaged along its path. A reconstructed bin collects many true energies and directions, and its combines their values of eq. (16), each weighted by how strongly it changes the bin count with the splitting. The second term, the matter term, is proportional to and grows toward the resonance, where is close to one and the expansion no longer holds.
A detector that does not distinguish neutrinos from antineutrinos averages the two shifts, weighted by the shares of neutrinos and antineutrinos in each bin. The average does not vanish, because the atmospheric flux and the cross section are larger for neutrinos than for antineutrinos. The same neutrino excess gives such a detector its own ordering sensitivity [18, 64]. Averaged with the weights of DeepCore and evaluated with the exact eigenvalues of appendix § 13.§ 1 in place of the expansion of eq. (16), the frequency shift accounts for about 70% of , and most of the rest comes from the amplitude of the oscillation near the resonance. Eq. (16) shows how scales with and the energy, and gives the sign of its dependence on the density. The matter term of eq. (16) depends on through its prefactor and through , so depends slightly on , and the slope of the solution differs from by about 1% for DeepCore.
Dependence on and
Two sets of parameters enter the atmospheric wrong-ordering solution. The true and define the NO data and are unknown, while and belong to the IO fit, which adjusts them or keeps them fixed, as stated in each case. Besides shifting the values preferred by the bins, the matter effect changes how the solution depends on . Figure 3 shows the intercept of DeepCore against the true , computed with and without Earth matter, and compares it with the accelerator closed form. Without matter the intercepts (open points) lie on the closed form and follow , as for accelerator experiments. With matter (filled points) they lie higher, as found in ref. [27], and are nearly flat. For the same reason the band over all values of is narrow.
Figure 3. Intercept of the IceCube DeepCore wrong-ordering solution against the true , computed with Earth matter (filled points) and without (open points), together with the accelerator closed form (dashed line). Points and line use at its NuFIT 6.0 value, and the band covers all values of , with matter. In both orderings is fixed at its true value, and the other parameters are at NuFIT 6.0, with the detector response of the IceCube DeepCore public data release [78] for disappearance. Dotted: , the vacuum value without the CP term.
The explanation lies in the decoupling of above the MSW resonance. In vacuum, oscillates with the effective frequency of eq. (8), set by , which depends on . A change of therefore changes the value preferred by every bin by the same amount. Above the MSW resonance decouples [18, 72, 73], and oscillates with a single frequency of the – system. This frequency is a difference of eigenvalues of the Hamiltonian in matter, which do not depend on , because can be removed by a rephasing that commutes with the matter potential [74, 75, 76, 77]. Above the resonance, the value preferred by each bin is then , where takes the form of eq. (16), which does not contain . To first order in , the phase enters only the amplitude of the oscillation, through a term proportional to that vanishes at maximal mixing, [18, 77]. This holds for samples whose statistical weight lies mostly above the MSW resonance, as for DeepCore. The atmospheric solution therefore depends little on , while the accelerator solution follows . As a result, the distance between the atmospheric and the accelerator intercepts, which sets the synergy of an atmospheric and an accelerator experiment, changes with the true , as the distance between the filled points and the line in Figure 3 shows.
For the situation is reversed: the accelerator solution depends weakly on it, and the atmospheric disappearance solution strongly. In disappearance sets the amplitude of the oscillation, with . It enters the frequency only through the small CP term of eq. (10). In vacuum a change of cannot compensate a change of frequency, so the IO fit keeps the amplitude of the true oscillation. The true and its octant mirror, , give the same amplitude, and the IO fit returns one of the two.
In Earth matter, near the MSW resonance, part of the convert to . This conversion grows with , and it occurs in neutrinos for the normal ordering and in antineutrinos for the inverted one. It gives atmospheric detectors their sensitivity to the octant of [79, 80]. The IO fit reduces the mismatch of this conversion between the orderings with a value of different from the true . The pull on comes mainly from the low-energy bins whose neutrinos cross the core or the deep mantle, and it moves toward maximal mixing, which deepens the oscillation. The bins that carry most of the weight lie above the energy of the first oscillation minimum, where both a deeper oscillation and a larger reduce the expected counts, so the IO fit compensates the deeper oscillation with a smaller . When the IO fit adjusts and , the intercept of DeepCore disappearance therefore rises, by an amount that depends on the true , so the solution moves with it. § 5 shows this for DeepCore: against the true the intercept changes little up to maximal mixing and grows in the upper octant. When all channels are included, the and charged-current events constrain , and the solution depends much less on the true .
Detector response
Besides the oscillation parameters, the atmospheric solution depends on the detector response. The values and the weights , and therefore , depend on the energy and direction resolutions, which determine the oscillation phases averaged in each bin, and on the energy window. The ordering sensitivity of atmospheric detectors is known to depend on their energy and direction resolutions [18, 19, 63, 64], and the same holds for their wrong-ordering solution, by amounts small compared with its separation from a reactor experiment [26]. Our reference description of DeepCore is built from the simulated events of its public data release [78]. From these events we tabulate the probability that a neutrino of a given flavor, energy, and direction is reconstructed in each analysis bin, and call this the tabulated response. Appendix § 14 compares it with a full fit of the DeepCore public data, with all event classes and the systematic parameters of the release, and on Asimov data the two solutions agree within . To separate the roles of the resolution and of the energy window, we keep the DeepCore acceptance, that is, the expected event rate against the true energy and direction, and replace the migration from the true to the reconstructed variables with Gaussian kernels in energy and zenith.
Figure 5. Intercept of the wrong-ordering solution of IceCube DeepCore (tabulated response, all channels) and KM3NeT/ORCA6 (public binned response [81], all classes and high-purity tracks) against the true , from the weighted bin average with the IO fit at its global minimum in and . Other oscillation parameters from NuFIT 6.0.
Figure 4 shows the resulting intercept against the energy resolution , the width of the Gaussian kernel in (left, with the value of the tabulated response as a dashed line), and against the lower and the upper edge of the energy window (right, with the value for the full window as a dotted line). With energy bins that follow the resolution, the intercept changes by up to about with the resolution, and not monotonically, because the resolution changes both which oscillation phases each bin averages and which bins carry the weight. The energy window changes the intercept through the bins just above the MSW resonance, which have the largest . Raising the lower edge of the window removes them and lowers the intercept, while lowering the upper edge to a few tens of GeV increases their share of the weight and raises it. A worse direction resolution, which Figure 4 does not vary, lowers the intercept by less than the energy resolution moves it, as quantified in appendix § 9.§ 4.
Figure 4. Intercept of IceCube DeepCore for the acceptance of its public data release with Gaussian kernels in energy and zenith in place of the tabulated migration. Left: against the resolution , with energy bins one standard deviation wide over the energy window of DeepCore, 6.31–158.5 GeV. Dashed line: the tabulated response. Right: at the energy resolution of DeepCore extracted from the simulation of its public release, – without the width of the analysis bins, varying only the upper edge of the window (circles, lower edge at 6.31 GeV) or only the lower edge (squares, upper edge at 158.5 GeV). Dotted line: the full window. disappearance, oscillation parameters from NuFIT 6.0, .
Comparison of two atmospheric detectors
We then apply the weighted bin average to a second atmospheric detector, KM3NeT/ORCA [41], whose atmospheric neutrinos have the same oscillation probabilities as those of DeepCore, so that only the detector differs. ORCA is located in sea water, its energy window starts at a lower energy, and it has its own energy and direction resolutions and three event classes (high-purity tracks, low-purity tracks, and showers). The binned response of its first six detection units (ORCA6) is public [7, 81, 82], and we evaluate the weighted bin average of eq. (4) on it with the same oscillation probabilities.
Figure 5 shows the intercepts of DeepCore with all channels, of ORCA6 with all event classes, and of the ORCA6 high-purity tracks alone, against the true , with and at the global minimum of the IO fit. At the NuFIT 6.0 value of the two detectors differ by about , and the high-purity tracks alone lie close to DeepCore. This difference comes mainly from the showers, which carry about 10% of the ORCA6 weight and lower the intercept in two ways. They prefer bin intercepts far below those of the track bins, and they move the IO fit to a larger . The intercepts of the two detectors change in opposite directions with the true splitting, and the slope of the ORCA6 solution stays within a few percent of . Two detectors of atmospheric neutrinos therefore have different wrong-ordering solutions, set by their energy and direction resolutions, their energy window, and their event classes.
In summary, Earth matter makes the atmospheric bins prefer different values. The atmospheric solution therefore lies above the accelerator one, depends little on , moves with in disappearance, and depends on the energy and direction resolutions, the energy window, and the event classes of the detector.
Wrong-ordering solutions of current experiments
In this section we compare the wrong-ordering solutions of current experiments. We compute the solution of every experiment from eq. (4) as a function of the true . This gives the closed form of eq. (13) for the short-baseline reactor Daya Bay and the weighted bin average of § 3 at for the medium-baseline reactor JUNO. For the accelerators NOvA and T2K, and for DeepCore, we evaluate the weighted bin average on their spectra or detector responses. The accelerator spectra are built from public information [83, 84, 85], with the exposures and backgrounds of the experiments and statistical uncertainties only, as described in appendix § 9.§ 3.
The plane of the two splittings
Besides , the IO fit can adjust and . For every experiment we take the solution at the global minimum of the IO over , in both octants, and . Where the stays within 0.01 of its minimum over a range of , we treat the minimum as degenerate and keep the whole range. In accelerator disappearance this range is the flat valley, which covers most or all of the circle [21, 23]. For DeepCore, also with all channels, it covers part or all of the circle. With appearance the accelerator minimum is instead a single point.
Figure 6 shows the solutions in the plane of against . Daya Bay and JUNO give one line each, since their solutions do not depend on or . The accelerators and DeepCore give bands, whose width comes from the unknown true values of and and, in accelerator disappearance, also from the flat valley of . All lines and bands have slope close to , and the bands of the accelerators and of DeepCore partly overlap. From the lowest to the highest intercept, the solutions are those of JUNO, Daya Bay, the accelerators, and DeepCore. JUNO lies below Daya Bay by the solar-phase term. The accelerators lie higher, since in disappearance, while . DeepCore lies higher still, mainly because of the matter shift contained in .
The true values of and are assumed unknown, and we vary them over their NuFIT 6.0 ranges for normal ordering [32]. The two ranges are varied independently, so the bands also contain combinations outside the joint region. The bands describe what is expected before the data are known. In nature and have a single value, so each experiment has one solution inside its band. For accelerator disappearance alone this solution is itself an interval, because of the flat valley of .
Ranges over and
Figure 7 shows, for each experiment at the NuFIT 6.0 value of , three ranges of the intercept: when only the true is varied over its range, when only the true is varied, and when both are varied. In disappearance all three bars of an accelerator experiment contain the flat valley of , which spreads the solution even when the true and are known. The bar for an unknown is wide because of this valley, and itself adds little, while the unknown true widens the bar further. The DeepCore solution is moved mainly by , and with all channels DeepCore is narrow in both parameters.
Figure 7. Intercepts at . For each sample, three bars give the range of when only the true is unknown (true at its NuFIT 6.0 value), when only the true is unknown (true at its NuFIT 6.0 value), and when both are unknown. The widths in are given at the right. The same IO fit, at its global minimum, is used in all three.
In disappearance, NOvA and T2K have nearly the same intercept, since does not depend on the baseline, and their bands differ by a few . With appearance their bands are also close, but their solutions at the same true values differ by up to about , mainly because the returned by the IO fit of T2K, where the matter effect is smaller, changes with the true , while that of NOvA stays close to . Adding appearance to the IO fit constrains , but not the unknown true . The accelerator bands in the right panel of Figure 6 and in the rows of Figure 7 with appearance are therefore narrower, while the position of these bands is still determined mainly by disappearance. With appearance the IO fit cannot reproduce the NO data of disappearance and of appearance at once, and its minimum is larger than zero. This minimum is the intrinsic ordering sensitivity of the accelerator, the first term of eq. (1), and does not enter the synergy term.
Figure 6. Wrong-ordering solutions in the plane of against , computed in this work around the NuFIT 6.0 value of . Lines: Daya Bay (short-baseline reactor) and JUNO (). Bands: range of the solution over the unknown true and , for NOvA (filled), T2K (dashed lines), and IceCube DeepCore (solid lines). Left: disappearance only. Right: with appearance for the accelerators and all channels for DeepCore. The range of the true covers nearly the whole circle, and that of is 0.430–0.596.
In the following we show NOvA as an example. Its longer baseline gives a larger matter effect than at T2K, so it is the case where the full calculation differs most from the closed form of eq. (10). Figure 8 shows how the intercepts of NOvA and DeepCore move with the true (left) and the true (right). For the accelerator the dotted curves of Figure 8 show the closed form [23], which follows from eq. (10) as
Figure 8. Intercept at against the true (left, true ) and the true over its NuFIT 6.0 range (right, true ). The IO fit is at its global minimum in and . NOvA disappearance (band): the flat valley of , where the is within 0.01 of its minimum. Lines: NOvA with appearance (dark gold), DeepCore with disappearance (magenta) and with all channels (violet), at the global minimum. For DeepCore the intercept changes little over the range of within 0.01 of the minimum, and this range is not drawn. Only NOvA is shown among the accelerators, and T2K has nearly the same disappearance band. All curves use statistical uncertainties only. Dotted: the accelerator closed form, eq. (10) with the exact [23], with over its flat valley and the octant mirror of the true value, as returned by the fit to the NOvA spectrum.
where and are the true values and and those of the IO fit. The dotted curves use the exact and, under the IO, the octant mirror of returned by the fit. In disappearance alone spans the flat valley, and the closed form gives a band, whose center follows and whose width is set by the range of [21, 23]. With appearance the fit fixes , and the band reduces to a line.
The NOvA solution stays close to the closed form of eq. (17). In disappearance the true and its octant mirror fit the NOvA data almost equally well. Matter effects change the effective in opposite directions for neutrinos and antineutrinos, and the IO fit returns the mirror, which describes both samples best. The same matter effects lower the effective of neutrinos under the IO, and with it the term in , so the NOvA band has nearly the same center as the closed form and is slightly narrower. With appearance, under the IO matter effects lower the appearance rate of neutrinos and raise that of antineutrinos, and the value returned by NOvA compensates this change most. Then , and the solution lies in the middle of the disappearance band. Against the true the accelerator band changes slowly, because enters it only through the small CP term, while the DeepCore disappearance solution moves with it, as described in § 4.2. With all channels DeepCore stays nearly constant in both scans.
Comparison with published best fits
Most experiments publish their best fits under both orderings. If the true ordering is normal, the best fit of an experiment under the IO is its wrong-ordering solution at its measured and at the true and , up to statistical fluctuations, which move the two best fits nearly together along a line of slope close to . Each experiment therefore gives one point in the plane of Figure 6. Figure 9 shows the published best fits together with the lines and bands computed in this work. Daya Bay [2] lies on its line. JUNO’s measurement with 207 days of data [1] gives , consistent with the line computed for the assumed resolution.1 NOvA [3], T2K [86], and the joint NOvA and T2K analysis [15] lie inside the NOvA band with appearance, and so does the global NuFIT 6.0 point [32], which precedes the JUNO measurement and whose is dominated by the accelerators.
Figure 9. Predicted wrong-ordering solutions (reactor lines and the bands of Figure 6, right: NOvA with appearance and IceCube DeepCore with all channels) against the published best fits of under the normal and the inverted ordering, : Daya Bay [2], JUNO after 207 days [1], NOvA [3] (frequentist best fit), the 2024 analysis of T2K [86], the joint NOvA and T2K analysis [15] (Bayesian, highest-posterior values), and the NuFIT 6.0 global fit [32]. Where is quoted, we use with the NuFIT 6.0 value of , and for IceCube DeepCore 9.3 years with the value used in its fit. Pentagon: IceCube DeepCore, 9.3 years [87], with the inverted-ordering best fit and interval taken from the released likelihood map [88], profiled over . Open circle: our fit to the IceCube DeepCore public data [78], , with the oscillation parameters of the release, as described in appendix E.4.
Our fit of both orderings to the DeepCore public data, described in appendix E.4, lies inside the DeepCore all-channel band. Its position depends on how the detector corrections of the release, which are tabulated against the splitting of the normal-ordering simulation, are read in the IO fit. The 9.3-year DeepCore sample reconstructed with convolutional neural networks [87], whose likelihood is released in both orderings [88], lies higher than our all-channel band. This sample has its own selection and reconstruction, with about seven times more events (150 257 against 21 402), so its position is not a test of our modeling. The other published best fits lie on or inside the computed lines and bands. The published uncertainty of each splitting is larger than the distance between the accelerator and DeepCore bands, so these points cannot yet tell the two bands apart, and their agreement with the computed solutions is a check of the consistency of our inputs.
In summary, the published best fits agree with the computed solutions. The solution of JUNO lies far below those of the accelerators and of DeepCore, whose bands partly overlap. The position of the accelerator band is uncertain mainly through the unknown , and that of DeepCore through in disappearance, while with all channels DeepCore changes little.
Mass ordering synergy in combinations
We now use the wrong-ordering solutions to evaluate the ordering sensitivity of a combination of two experiments. On NO Asimov data this sensitivity is the evaluated at the minimum of the joint IO fit. It contains the minima of the two individual IO fits, and , and a term that arises because the two experiments return different values of . Under the NO hypothesis both experiments return the true splitting , and under the IO hypothesis their own solutions and . A joint IO fit has a single , which lies between them. Close to its minimum, the IO profile of each experiment is well approximated by a parabola in , with widths and , respectively.2 Minimizing the sum of the two parabolas gives eq. (1), where the approximate sign refers to the parabolic form of the profiles. Appendix F extends eq. (1) to oscillation parameters shared by the two experiments and to real data, where the two experiments can measure different values of . The synergy term of eq. (1) depends only on the difference of the intercepts and on the precisions, . Figure 10 shows it for the pairs formed by JUNO, NOvA, and DeepCore, as a function of the relative precision , with the true NO splitting, common to the two experiments of a pair, with disappearance only (left) and with appearance and all channels (right). The true and are properties of nature and are the same for the two experiments of a pair, so each band is the range of the synergy term over the unknown true values of and . In Figure 10 each experiment keeps its own . Where its IO fit keeps a range of , as the flat valley of an accelerator in disappearance or the range of DeepCore, the joint IO fit takes the value that brings the two solutions closest, that is, the smallest .
Figure 10. The synergy term of eq. (1) for pairs of JUNO (), NOvA, and IceCube DeepCore, as a function of the relative precision , common to the two experiments of a pair, at . Left: disappearance only. Right: NOvA with appearance and DeepCore with all channels. Bands: range over the unknown true and , the same for the two experiments of a pair. Where the IO fit keeps a range of , the flat valley of NOvA in disappearance and the range of DeepCore in both panels, the joint IO fit returns the value that brings the two solutions closest. In the right panel the JUNO–NOvA band is hatched. The precisions are illustrative.
The intercept of JUNO lies far below those of NOvA and DeepCore, so the pairs with JUNO give the largest synergy, in line with the large synergies reported in reactor–accelerator and reactor–atmospheric studies [21, 23, 24, 26, 27, 28, 89, 90, 91].3 The NOvA–
DeepCore pair gives the smallest synergy, depending on the unknown and . For this pair eq. (1) is an approximation, because the two experiments also share and . The synergy term of eq. (1) treats the IO fits of the two experiments as independent. JUNO does not depend on and , so in a pair with JUNO the two experiments share only , apart from and the solar parameters. Eq. (1) then gives the joint result, provided that the solution and the precision of the partner are taken with and profiled. NOvA and DeepCore both depend on and , and a joint IO fit shares these parameters between them. The synergy term then takes the form of eq. (F.2), derived in appendix F, where the two solutions differ also in and and all differences are weighted by the inverse of the sum of the two covariance matrices. A joint fit of NOvA and DeepCore that shares and gives a larger synergy at almost all true values of and , because the two IO fits prefer different values of these parameters also when they return the same octant, so the NOvA–DeepCore bands of Figure 10 are lower estimates nearly everywhere.
Figure 11 shows the synergy term of JUNO with DeepCore (all channels, left) and with NOvA ( disappearance and appearance, right) for several assumed precisions of the two experiments. For comparison, the present uncertainties on or are about 1.0% for JUNO with 207 days of data [1], 1.4% for NOvA [3], and 1.9% for DeepCore with 9.3 years of data [87], a sample different from the one modeled here. Each ring corresponds to one pair of precisions, that of JUNO on the vertical axis and that of the other experiment on the horizontal axis. Its inner and outer radii give the smallest and the largest synergy over the unknown true and , and the two values are reported below the ring. The width of the rings shows how much the synergy depends on the unknown and . For JUNO with DeepCore the ring is narrow. For a given detector response the synergy depends little on the unknown and , and it depends instead on the response, which has to be modeled for each sample. For JUNO with NOvA the ring is wide, and depending on the true the same precisions give a synergy that changes by about a factor of two. At equal precisions the synergy with DeepCore is close to the largest value reached with NOvA.
Figure 11. Synergy term for JUNO with IceCube DeepCore (all channels, left) and with NOvA ( disappearance and appearance, right), for different sets of precisions of the two experiments, relative to the true , with JUNO between 0.5% and its present precision of 1.0% [1]. Each ring spans the range of over the unknown true and (inner dashed: smallest, outer: largest), with the joint IO fit as in Figure 10. The area enclosed by each circle is proportional to , and the rings of NOvA are hatched. The axis labels , , and are the uncertainties and of eq. (1.1) for JUNO, DeepCore, and NOvA. The precisions are illustrative.
Combinations with JUNO already add to the ordering sensitivity in data. With constraints on from long-baseline accelerator experiments, the first 207 days of JUNO data favor the normal ordering with a one-sided significance of at least for any value of [1]. Under the inverted ordering, the atmospheric data of Super-Kamiokande, IceCube DeepCore, and ORCA6 and the first 207 days of JUNO data prefer different values of , and their combination raises the preference for the normal ordering from , which comes mainly from the intrinsic sensitivity of the atmospheric data, to 15.1 [31].
In summary, the pairs with JUNO give the largest synergy. With an accelerator it depends strongly on the unknown , and with an atmospheric experiment using all channels it is nearly fixed by the measured oscillation parameters and the detector response.
Conclusions
The sensitivity of a combination of experiments to the neutrino mass ordering depends on how far apart their wrong-ordering solutions lie, that is, the values of they return when NO data are fitted under the IO hypothesis. We have shown how the wrong-ordering solution of any oscillation experiment is determined. It is the weighted average of the values preferred by the individual bins, with weights set by the statistical precision of each bin. Both the values and the weights depend on the detector response. In the plane of against every solution is a line of slope close to . The slope is exactly for short-baseline reactors and accelerators in vacuum, and it differs from by about 1% for JUNO and DeepCore and by a few percent for ORCA6. Each experiment is therefore described by one number, its intercept , for atmospheric neutrinos in Earth matter as for reactors and accelerators, where matter effects are small. All intercepts of this work are computed with the solar parameters and fixed and with statistical uncertainties only.
For short-baseline reactors, and for accelerators in vacuum, all bins prefer the same value, and the weighted bin average returns the known closed forms [20, 21, 23, 24, 25, 36, 37]. The difference of two intercepts is the difference of the wrong-ordering splittings of the mass ordering sum rule [24], so the weighted bin average extends the sum rule to any oscillation experiment. Systematic uncertainties act only on the weights of the bins, so to first order they cannot move the solution where all bins prefer the same value. At a medium baseline the solar phase varies across the spectrum, so the intercept of JUNO depends on its energy resolution. The accelerator solution depends on the unknown , and adding appearance narrows its range but leaves its position set by disappearance. For atmospheric neutrinos, Earth matter makes the bins prefer different values. The solution then depends on the energy and direction resolutions, the energy window, and the event classes of the detector, and it differs between DeepCore and ORCA6. Above the MSW resonance the frequency of disappearance is nearly independent of , so the atmospheric solution moves mainly with .
The weighted bin average shows which features of an experiment set its wrong-ordering solution. The intercepts of the individual experiments, computed in this way or read from their published best fits in the two orderings, give the synergy of a combination before a joint fit, through eq. (1) or, for measured values, through its form of appendix F. For combinations, the intercept quantifies the mismatch of the wrong-ordering solutions that earlier studies identified as the origin of the large synergies of reactors with accelerator and atmospheric experiments, and the intercept of JUNO lies far below those of both. Our results also show that the synergy of JUNO with an accelerator depends strongly on the unknown , while with IceCube DeepCore using all channels it is nearly independent of and but depends on the detector response, which therefore has to be modeled for each experiment.
As the precisions of JUNO and of the accelerator and atmospheric experiments improve, the separation of their intercepts sets how much their combination adds to the determination of the neutrino mass ordering.
Code and data availability. The code that reproduces all figures and numbers of this paper is public [42], and its documentation lists the version and the checksum of every input. All inputs are public: the IceCube DeepCore data releases [78, 88], the KM3NeT/ORCA6 open-data release [81], the NOvA data and flux releases [83, 84], the T2K flux release [85], the GENIE cross sections [92, 93, 94], the atmospheric neutrino fluxes of ref. [95], and the published configurations and results of the experiments cited in the text.
Acknowledgments
We thank the IceCube, KM3NeT, NOvA, and T2K collaborations and their funding agencies for making their data publicly available. During the preparation of this work the authors used AI tools to help edit the text, to assist with coding tasks, and to cross-check numerical calculations. The analysis, the physics content, and the style and content of the figures were defined by the authors, who verified all derivations, numerical results, and statements and take full responsibility for the content of the paper.
Inputs and experimental configurations
This appendix lists the oscillation parameters and describes how we model each experiment. These inputs enter all results of sections 2 to 6.
Notation and oscillation parameters
The true values of the oscillation parameters are the NuFIT 6.0 NO best fit without the Super-Kamiokande atmospheric data [32],
We call the true and the value of given in § 2 the NuFIT phases, and all results use them unless a range is shown. Two results use other values. The validation of appendix § 14 uses the same in both orderings, and the fits to the observed data of the DeepCore release in appendices E.3 and E.4 use the fixed oscillation parameters of the release. Intercepts are quoted in units of . For reference, and .
Reactor experiments
For the short-baseline reactor Daya Bay the solution is the closed form of eq. (13), which does not depend on the detector, so it needs no further input.
For the medium-baseline reactor JUNO we use its published configuration [58], with eight cores, six of 2.9 GW at 52.1–52.8 km and two of 4.6 GW at 52.6–52.8 km, each weighted by its thermal power over . The flux is based on the Huber–Mueller model [96, 97], with the fission fractions of a commercial pressurized-water reactor (U 0.58, Pu 0.30, U 0.07, Pu 0.05), and we use the inverse-beta-decay cross section [98]. The bins span 1.81–9 MeV in reconstructed neutrino energy, in 20 keV steps, and the energy resolution is evaluated at the prompt energy MeV,
where the stochastic term is , the constant term , and the noise term unless varied. These are the values predicted by the full simulation of the detector [53] and used in its ordering sensitivity study [11]. When we vary , we scale and by the same factor, so that the resolution is perfect at . The survival probability is the exact three-flavor one at constant density, computed with the NuFast algorithm [99]. We use g cm, with and the average density g cm that JUNO estimates for the path of the antineutrinos through the crust and the sediment [58]. This description contains statistical uncertainties only, without backgrounds. Within it the intercept is set by the baselines, the reactor spectrum, and the energy resolution.
Accelerator experiments
For all accelerators the oscillation probabilities are the exact three-flavor ones in matter of constant crust density, computed with the NuFast algorithm [99], with the densities and the electron fraction listed with the other accelerator configurations in Table 1.
Table 1. Accelerator configurations, with densities in g cm and an electron fraction of 0.5. FHC and RHC are the forward and the reverse horn current, that is, the neutrino and the antineutrino mode.
The closed-form checks of § 3 use a generic beam, with a cross section rising like , events in 0.1 GeV bins over 0.5–5 GeV, and .
The realistic configurations of NOvA and T2K enter the bands of § 5, with disappearance alone and together with appearance. We build their far-detector spectra from public information, separately for the neutrino and the antineutrino modes of the beam, with forward (FHC) and reverse (RHC) horn current. Every sample is normalized to the published predicted event numbers, at the NOvA best fit for NOvA (table II of ref. [3]) and at for T2K (table XI of ref. [101]). The cross sections therefore enter only through the ratio of the to rates for NOvA and through the energy dependence of the rates for T2K. We use the cross sections of GENIE 2.12.10 [92] on water for T2K and on for NOvA, built from its tabulated cross sections on oxygen, carbon, and hydrogen. For NOvA the -like samples are the no-oscillation prediction of the data release in reconstructed energy, split into and by the ratio of their charged-current rates, each computed as the public near-detector flux [84] times the cross section. The probability is averaged over each reconstructed-energy bin without energy migration. In disappearance the solution does not depend on the response, by eq. (10), up to the small matter shifts of appendix § 12.§ 2. In the -like samples the appeared are the no-oscillation spectrum times , the backgrounds have the shape of the no-oscillation spectrum, and the beam have the shape of the flux times the cross section. For T2K we use the flux at Super-Kamiokande and one efficiency per signal or background component, which sets the relative size of the components in each sample. The efficiencies are calibrated to the event numbers of the Hyper-Kamiokande design report [102], which has the same beam line and detector technology, and the total of each sample is then normalized to the T2K prediction as stated above. The T2K spectra are folded with . The -like count of a bin sums the surviving and , and the -like count adds the appeared and the surviving beam and . The backgrounds (neutral current and, for NOvA, cosmic rays and the other backgrounds of its release) are not oscillated.
DeepCore
For DeepCore and ORCA6 we use the three-flavor probabilities through PREM of § 4, computed with the algorithm of NuFast-Earth [71]. As in NuFast-Earth, each path is cut at the PREM discontinuities, and we divide each piece into layers of at most 500 km, each at its mean density. The eigenvalues and eigenvectors of each layer are obtained with the NuFast algorithm [99], and the amplitudes of the layers are multiplied along the path. Each path starts at a production height of 20 km in the atmosphere, treated as vacuum, and crosses the ice above DeepCore or the sea water above ORCA6.
We use the detector response tabulated from the public data release [78], which contains the high-purity “golden” event sample of ref. [6]. The response is normalized to the 21 402 neutrino events of this sample, the 21 914 observed events minus the 512 atmospheric muons expected from the simulation of the release. We use the azimuth-averaged Honda flux at the South Pole at solar minimum [95] and the probabilities described above.
The release provides the simulated events of the analysis. For each event it gives the flavor and charge, the true energy and direction, the analysis bin in which it is reconstructed, and a weight in , which gives the expected number of events when multiplied by the flux and the livetime. We use the events reconstructed inside the published analysis bins, nine uniform in from 6.31 to 88.2 GeV and one from 88.2 to 158.5 GeV, and ten uniform in from to 0.1, with the event classes merged. From them we build three tables, separately for each channel (, , and charged current, and neutral current by flavor) and for neutrinos and antineutrinos. The acceptance is the summed event weight per unit and over 1–1000 GeV and the full sky. It contains the cross section, the target mass, the trigger, and the selection. Each analysis bin is labeled by its reconstructed-energy bin and its reconstructed-zenith bin . The energy migration gives the fraction of events of a true-energy cell reconstructed in each analysis energy bin, and the zenith migration the same for the analysis zenith bins. The migration tables are binned in coarse cells of true energy and zenith and filled with the event weight times the Honda flux of the event’s own flavor, so that inside each coarse cell the migration follows the falling flux of the events. The expected count of the reconstructed bin is a sum over a finer grid of true cells , each of which uses the tables of the coarse cell that contains it,
with the flux of each initial flavor and the probability of oscillation into the flavor of channel evaluated on the grid, with , , and the tables of channel . The sum over runs over charged current alone or over all channels, and the counts are normalized to the events of the sample given above. The grid sizes are given in the public code [42], and a five times finer grid leaves the intercepts unchanged at the precision quoted in this paper. We make three approximations: the migration is factorized in energy and zenith, the event classes are merged, and the atmospheric-muon background is not included. Their combined effect on the intercept is evaluated in appendix § 14.§ 2. The release gives only the bin of each reconstructed event, not its reconstructed value, so the tabulated migration exists only for the published bins. Other bins and windows, as in Figure 4, therefore need a model of the migration. For this model we extract the energy and zenith resolutions of the release from its charged-current events, as half the central 68% interval of and of in bins of true energy. The reconstructed values are the centers of the analysis bins, so these widths include the bin sizes. The Gaussian kernels are integrated over the bin edges, and we therefore remove from each extracted width, in quadrature, the variance of the analysis bins of width .
For Figure 4 we keep the tabulated acceptance and model the migration with Gaussian kernels in and , integrated over the bin edges. In the resolution scan the zenith resolution is the one extracted from the simulation of the release. In the window scans the published bins are cut at the new edge, and the resolutions extracted from the simulation are used. Scaling the zenith width of the Gaussian kernels from half to twice that of the release, with the energy response and the bins of the release, lowers the intercept by for disappearance and by with all channels. At the resolutions of the release and over the full window, the Gaussian kernels and the tabulated response, two models of the same detector, give intercepts apart.
KM3NeT/ORCA6
The ORCA6 open-data release [81, 82] (433 kt yr, the sample of ref. [7]) gives the expected weight of each true cell in each reconstructed bin, for every neutrino flavor, charge, and interaction. The reconstructed bins are 10 up-going bins in each of three classes, namely high-purity tracks, low-purity tracks, and showers. In energy the two track classes have 14 bins between 2 and 100 GeV, and the showers have one more bin, from 100 to 1000 GeV [7]. The count of a reconstructed bin is the sum of these weights times the oscillated flux of the initial flavors, plus the atmospheric-muon background of the release. We use the Honda flux [95] at the Fréjus site (solar minimum, no mountain), as prescribed by the release [81], and the probabilities of appendix A.4, averaged over each true cell on a finer grid. Without any nuisance parameter the predicted total agrees with the observed one to 0.7%. The high-purity tracks, the low-purity tracks, and the showers carry 65%, 25%, and 10% of the weight, respectively. Adding the low-purity tracks to the high-purity ones changes the intercept by less than . Adding the showers lowers it by , and the fitted rises by about 0.01. The high-purity tracks alone therefore lie within a few of DeepCore.
Derivation of the weighted bin average
This appendix derives the weighted bin average of § 2, its symmetry under the exchange of the orderings, its iterative solution, and its form with profiled nuisance parameters.
Bin observable, Poisson weight, and stationarity
The expected count in a bin of reconstructed variables, under ordering and for a given value of , is
where the integrals run over the true energy and direction, is the flux, the cross section, the efficiency, the exposure, and the probability that an event at the true point is reconstructed in bin . Only depends on the splitting, so the count factorizes exactly, , with the unoscillated count . The bin observable is the probability averaged over the true energies and directions that the bin collects. For a bin that collects several channels, with any fixed reference count , since and do not depend on it. For Asimov data the of eq. (2) is twice the Poisson log-likelihood ratio, expanded to second order in , with the inverse variance of the measured fraction .
Expanding around a point gives the Newton step
with . It approximates the exact root of , where one exists, with an error of order . Taken at as an equality, it is the definition of in eq. (3). The minimum of eq. (2) satisfies
By the definition of , , and eq. (B.3) becomes , which is eq. (4). The relation is exact for the of eq. (2).
Exchanging the two orderings
For a true inverted ordering with splitting , the fit under the NO hypothesis returns . Each bin prefers the value given by eq. (B.2) with the two orderings exchanged, provided that and the CP phase of each ordering are exchanged as well. Where all bins agree, as for short-baseline reactors and accelerators in vacuum, a true inverted ordering gives with the same intercept. In matter the two orderings differ in amplitude as well as in phase, so both the bin intercepts and the weights change for a true inverted ordering, and the two changes partly cancel. As a check for DeepCore with fixed, we take the IO solution as the true splitting of an inverted ordering and fit it under the NO. The fit returns the original to about for disappearance and to within with all channels. The synergy term of eq. (1) is therefore symmetric under the exchange of the two orderings when, besides the intercepts, the widths and of the IO fits are the same for the two true orderings.
Gauss–Newton iteration
Inserting eq. (B.2) at a trial value into eq. (4) gives the Gauss–Newton step for the of eq. (2), whose fixed point satisfies eq. (B.3). For the atmospheric evaluation we use the same step written in counts, with in place of and in place of . The iteration starts from the closed forms of § 3. All intercepts with fixed and are obtained in this way. The bands, which take the global minimum over and , use the same , minimized as follows. As shown in appendix § 13.§ 2, the Hamiltonian in matter depends on and only through a rotation of the flavor basis, so at fixed every expected count is a trigonometric polynomial of degree two in and in , for any matter profile [75, 76]. A polynomial of degree two in an angle has five terms, a constant and the cosines and sines of the angle and of twice the angle, so 25 coefficients describe each count. We compute them once and minimize the in , in separately in each octant, and in . The grids, the tolerances, and the sampling of the true and are given in the public code [42].
Profiled nuisance parameters
The validation of appendix § 14 compares the weighted bin average with a full fit that profiles the systematic uncertainties. This requires the weighted bin average with profiled nuisance parameters, which we derive here. Let the IO fit also vary nuisance parameters , with Jacobian , constrained by pull terms [50] with prior covariance . A parameter without a prior, such as , has a vanishing row and column in . From here on we work in counts. We write , which is times the derivative of § 2, and we redefine , the weight of § 2 divided by , so that keeps its value. With the residual and the diagonal matrix of these weights, the linearized is
Minimizing over leaves , with . This matrix satisfies , with the profiled derivative
and by the definition of . The minimum in is therefore at
the weighted bin average with the weights . The profiled derivative is the part of that no combination of nuisance parameters can reproduce within their priors. A bin whose derivative with respect to the splitting is mimicked by a nuisance parameter loses weight, and the sum of the weights is the inverse variance of the splitting with the nuisance parameters profiled. The sum, , is positive, while a single can be negative in bins that a nuisance parameter correlates strongly with others, so the are the weights of a linear estimator rather than probabilities. For nuisance parameters that stay close to their prior centers the do not change and only the weights do. The fitted values of and can move far from the true ones, as § 4.2 and § 5 show, so we evaluate , the , and the at these fitted values. The is stationary there, and eq. (4) holds exactly with .
Disappearance in vacuum: reactor and accelerator experiments
This appendix gives the effective-frequency form of the survival probability used in § 3, the matter effects in accelerator disappearance, and the dependence of the JUNO solution on the energy resolution.
Effective-frequency form of the survival probability
In vacuum, with and , the two fast terms of the survival probability combine into one oscillation, eq. (7) [21, 25, 36, 37], where is the phase of the frequency of eq. (8). The amplitude and the solar phase are
For equal mixing parameters in the two orderings, the coefficients and are the same, so the ordering enters only through the sign in front of . When the solar oscillation has not developed, is close to one. Matching the two cosines at one energy, with , , and , gives eq. (9). This phase-matching value equals of eq. (3) when is close to linear between and , and, for , when the two orderings have the same CP phase. With different phases the terms of that oscillate with differ between the orderings and add to an offset of order , below for DeepCore without matter. For at short baselines [37] is negligible, since at 4 MeV at 1.65 km, i.e., the flux-averaged baseline of the far hall of Daya Bay [2], which gives eq. (13).
Accelerator disappearance and matter effects
For accelerators the solar phase is negligible, since at 810 km and 2 GeV. For in vacuum eq. (4) returns eq. (10) for any binning and resolution. Matter makes the bins differ only slightly, in line with the small size of matter effects in long-baseline disappearance [55], and the intercept depends on the share of antineutrino events. The matter effects shift the solution in opposite directions for neutrinos and antineutrinos, so the shift cancels for equal shares of the two, as in the generic beam of Figure 2, and each experiment is evaluated with its own share. For a beam of neutrinos alone the intercept lies 4 to below the closed form at 810 km, depending on the true , and less than away from it at 295 km, with the flux of the generic beam at 810 km and a flux peaked at 0.6 GeV at 295 km, and fixed at the same value in both orderings. The same matter effects change the width of the flat valley. Under the IO they reduce the effective of neutrinos, roughly as with at 2 GeV, and increase it for antineutrinos. The term of in is proportional to , so the NOvA sample, which consists mostly of neutrinos, depends less on , and its disappearance band in Figure 8 is about 11% narrower than the closed form, with its center close to that of the closed form.
Medium-baseline reactor: solar phase and energy resolution
At the baseline of JUNO is of order one, and the average of eq. (14) depends on the weights. We define the effective energy as the energy at which equals its weighted average . Figure 12 shows , the weights of the bins for resolutions from to 8%, and the average for each resolution placed at its effective energy. A worse resolution moves the weights, and with them , to higher energies, from 3.5 MeV for a perfect resolution to 4.1 MeV at . A closed form with the solar phase at 4 MeV and km, , therefore equals eq. (14) only at the resolution for which MeV, . At the weighted average of eq. (14) reproduces the full weighted bin average to within , while the closed form at 4 MeV lies below it. Bins of 10 keV instead of 20 keV change the intercept by , while bins of 100 keV, about twice the resolution near 3 MeV, raise it by . All values include matter effects.
Figure 12. JUNO, eight cores, in matter: the solar-phase term (black line, left axis), the weight of each reconstructed-energy bin for six values of from 0 to 8% (shaded, right axis, smoothed, each with unit area, darker shades for worse resolutions), and the weighted average of for each resolution placed at its effective energy (points, with in the legend). The star is the closed form at 4 MeV.
Atmospheric neutrinos in Earth matter
This appendix supports § 4. It derives the matter shift above the MSW resonance and the suppression of the CP dependence. In the flavor basis the Hamiltonian is [72]
where is the rotation by in the – plane and for neutrinos ( for antineutrinos).
Matter shift above the resonance
The remainder of eq. (15) contains four parts. The first is the matter shift of the atmospheric frequency. The second is the vacuum term
which is the CP term together with a small term of second order in . The third is the solar-phase term , and the fourth, when they are included, is the contribution of the other channels in the bin. The solar-phase term is negligible for DeepCore. Matter makes the differ between bins.
Above the resonance the -like level lies far from the two levels of the – system. It shifts them only through its small mixing, at second order in perturbation theory [72, 73, 74]. In the -rotated basis the -like level, near , couples to the third level, at , with strength , and pushes it by , to leading order in . Its coupling to the second level is proportional to and negligible. For neutrinos in the normal ordering the third level is pushed down, and the frequency of the – system, the difference of its two levels, decreases in size by . In the inverted ordering the third level, at about , is pushed away from the -like level, and increases by , to leading order with . Matching between the orderings gives the remainder of eq. (16), with + for neutrinos and − for antineutrinos. For antineutrinos changes sign, and in eq. (16) denotes in both cases.
The first term of eq. (16), , is the difference between the solar term of the frequency above the resonance and , and does not depend on the matter density. The second term, the matter term, falls as at high energy. At a fixed energy it grows as far above the resonance. The expansion assumes that the -like level is far from the third one, and it fails at the resonance, where the second-order denominator vanishes. In the configuration of Figure 1, the bins between 20 and 50 GeV for up-going neutrinos, above the resonance, prefer between and , slightly above the solution, and the bins drawn in gray carry less than 0.05% of the weight each.
To follow the shift through the resonance, where the expansion of eq. (16) fails, we use the exact eigenvalues of the two-level system formed by the -like and the third level, with set to zero, at the path-averaged PREM density of each cell of true energy and direction, and average the shift with the weights of the detector. At the NuFIT phases this reproduces about 70% of , the value quoted in § 4.1, and at , where the CP amplitude term of eq. (D.4) vanishes, 72–77%. The rest is an amplitude effect. Near the resonance the effective in matter approaches one, in neutrinos for one ordering and in antineutrinos for the other, so the two orderings differ also in the depth of the oscillation, which the IO fit compensates with the frequency. The one-mass-scale probability in matter [72], that is, the probability with set to zero, which keeps both effects, reproduces 92–95% of at . At the NuFIT phases a further – comes from the CP amplitude term of eq. (D.4).
Suppression of the CP-phase dependence in matter
The suppression of the CP dependence in matter follows from the structure of the Hamiltonian given above. The matter potential commutes with and with , and commutes with , so with independent of [77]. Two consequences follow, in vacuum and in matter alike [74, 75].
The eigenvalues of , and hence all oscillation frequencies, do not depend on .
enters only through the projections of on the eigenstates, , with the eigenvectors of .
Above the MSW resonance, , the -like eigenstate decouples [72, 73], with a residual admixture of order . The state then lives in a two-state subspace, and restricted to the second and third states of the -rotated basis is
This two-state system has a single frequency. The frequency is an eigenvalue difference and therefore does not depend on , . Matching between the two orderings gives, for every bin,
the common part of eq. (15), without any CP term.
The CP phase survives only in the amplitude of the oscillation. The small off-diagonal term of the two-state Hamiltonian rotates its eigenstates by a small angle, of order times . Through the phase in the projection of , this rotation changes the effective mixing angle of by an amount proportional to , and the probability becomes
with () for the normal (inverted) ordering, for neutrinos and antineutrinos alike. The depth of the oscillation then changes with by an amount proportional to . This change vanishes at maximal mixing, changes sign with the octant, and does not move the frequency. At finite the decoupled state contributes a third frequency with a weight of order , and through it a small dependence of the frequency on returns. The same admixture shifts in opposite directions in the two orderings, which is the matter shift of appendix § 13.§ 1.
Validation against the IceCube DeepCore data release
This appendix describes the full fit of the DeepCore public data, the validation of the weighted bin average against it, the reproduction of the published measurement, and the fit of both orderings to the observed data.
Full-fit configuration
Following the DeepCore oscillation analyses [6, 87, 103], we build a standalone fitter for the DeepCore public data release [78], which provides the simulated events, the detector systematic corrections, and the observed data. The full fit uses the expectation of the release for every flavor and interaction, both event classes of the release, the mixed and the track class, and the atmospheric-muon background. The three-flavor oscillation probabilities are weighted by the flux of the initial flavor of each event. The fit varies the thirteen parameters of Table 2. The detector parameters enter through the hypersurfaces of the release, which parameterize the change of each bin with each detector parameter. The is the modified Pearson form , with the expected and the observed counts of bin , and the statistical variance of from the finite number of simulated events. For the validation of appendix § 14.§ 2 we build Asimov data at 25 values of in –, with the NuFIT 6.0 oscillation parameters of eq. (A.1) and the nuisance parameters at the best fit of the data of the release. The inverted ordering is fitted with all thirteen parameters, with in both orderings and the pull terms centered at the true values. The hypersurfaces of the release depend on the splitting and are tabulated against for NO oscillations. On Asimov data we evaluate them at the true splitting in both orderings, so that the detector correction is common to the data and the fit, as in the weighted bin average.
| Parameter | Prior |
| , | none |
| Optical efficiency | Gaussian, relative width 10% |
| Hole ice (two parameters) | none |
| Bulk-ice absorption and scattering | bounded scale factors, 0.90–1.10 and 0.95–1.15 |
| Effective-area normalization | none |
| Atmospheric-muon normalization | none |
| Spectral index | Gaussian, absolute width 0.10 |
| Axial mass (two parameters) | Gaussian, width set by the parameterization of the release |
| Neutral-current normalization | Gaussian, relative width 20% |
Table 2. Parameters of the DeepCore full fit and their priors. The optical efficiency, the hole ice, and the bulk ice are the five detector parameters of the release.
Comparison with the weighted bin average
We fit the IO hypothesis to NO Asimov data at each true . As in the DeepCore analyses, is profiled, while , to which DeepCore is hardly sensitive, is fixed. The comparison only requires the same phase in the fit and in the weighted bin average. This configuration differs from the reference configuration of § 4, so the intercepts differ from the value given there. Figure 13 compares the intercept of this full fit (stars) with three evaluations of the weighted bin average. The first uses the tabulated response with all channels and profiled (line and circles). The other two use the expected counts of the full fitter, which include the detector model of the release, first with only fitted (open diamonds) and then with and all nuisance parameters profiled together with the splitting, by iterating eq. (B.6) (filled diamonds). The lower panel shows the difference of each evaluation from the full fit. Over the range of the three evaluations compare with the full fit as follows. (i) The tabulated response with all channels and profiled lies above it by –, the difference quoted in § 4.3. (ii) The weighted bin average on the expected counts of the full fitter with only fitted lies above it by at most . The effect of and of the systematic uncertainties is therefore about a tenth of . (iii) With and all nuisance parameters profiled, the iteration reaches the minimum of the full fit.
Figure 13. IceCube DeepCore, NuFIT 6.0, in both orderings. Weighted bin average, tabulated response, all channels. Full fit (all channels, nuisance parameters). Weighted bin average, full-fit model, and nuisance parameters profiled. Weighted bin average, full-fit model, only fitted (lower panel). Figure 13. Validation of the weighted bin average for IceCube DeepCore, NuFIT 6.0 with in both orderings. Top: intercept against the true from the full fit (stars), from the weighted bin average of eq. (4) on the tabulated response with all channels and profiled (line and circles), and on the expected counts of the full fitter with and all nuisance parameters profiled (filled diamonds, which coincide with the stars). Bottom: difference of each evaluation from the full fit, including the weighted bin average on the expected counts of the full fitter with only fitted (open diamonds).
The nuisance parameters account for – of the difference (i) over the range of , and the approximations of the tabulated response for the rest.
Reproduction of the published DeepCore measurement
We fit the observed events of the release in normal ordering and compare the result with the IceCube DeepCore measurement on the same sample [6]. The fit uses the full expectation described above, at the value of ref. [6], and the parameters of Table 2 with their priors. Figure 14 shows our 90% contour and best fit together with the published ones, both contours assuming Wilks’ theorem [104], while the published one-dimensional 68% intervals are Feldman–Cousins. Our fit gives and , against the published 0.51 ± 0.05 and . The best fit and the 90% contour agree with the published ones to a small fraction of a standard deviation, and our uncertainties, from the curvature of the at its minimum, are within about 20% of the published Feldman–Cousins intervals.
Figure 14. The IceCube DeepCore measurement reproduced from the public release. Solid: the 90% contour of our fit to the public data, with every nuisance parameter the release supports profiled with the priors of IceCube DeepCore. Dashed: the published 90% contour of IceCube DeepCore on the same sample, digitized from figure 26 of ref. [6]. Dot: our best fit. Star: the published best fit, digitized from the same figure.
Fit of both orderings to the data
The detector hypersurfaces of appendix § 14.§ 1 are tabulated against the splitting of the NO simulation, so for the IO fit the splitting at which they are read is a choice. We read them at , the splitting whose NO muon-disappearance pattern is closest to that of the IO model according to eq. (10) without the CP term, so that the detector correction follows the oscillation pattern rather than the label of the splitting. On Asimov data at the NuFIT 6.0 best fit this reading reproduces the intercept obtained with the hypersurfaces at the true splitting to , while reading them at lowers it by .
We apply the same fitter to the data of the release, with the fixed oscillation parameters of the release (, , ) and this reading. It returns under the normal and under the inverted ordering. The intercept is . The two splittings move together under statistical fluctuations, so is determined better than either splitting.
Combination of experiments
This appendix derives the synergy term of eq. (1), its form for real data, where the two experiments measure different splittings, and its form with shared oscillation parameters, used in § 6.
The synergy term
For two independent experiments the combined IO is the sum of the individual ones, minimized over the common parameters. The minimum of a sum is not smaller than the sum of the minima. For Asimov data the NO minima of both experiments and of their combination vanish, so . The equality holds only when both experiments return the same IO parameters. The excess is the information carried by their disagreement, and it equals the parameter goodness-of-fit statistic of refs. [105, 106], evaluated under the IO hypothesis.
For Asimov data the NO solutions of the two experiments coincide at the true splitting. Near its minimum the IO of experiment is . Here is the width of the IO fit in , which is close to the NO measurement error but not equal to it, and and are values of the same parameter, . The sum is minimized over the common at , which leaves , the synergy term of eq. (1). At the common true splitting exactly, by the definition . In the joint fit each experiment is moved away from its own solution by , and the parabolic form holds as long as this shift stays within the range where its IO is parabolic.
With real data the two experiments measure different values and under the true ordering, with widths and , and their IO solutions are . Minimizing the combined in each ordering in the same approximation gives
This expression reduces to eq. (1) for . A difference of the measured splittings that moves the two IO best fits further apart enlarges the synergy term , and one that moves them closer reduces it. The synergy term is negative when the two experiments agree better under the IO than under the NO. In this case the combination prefers the normal ordering less than the two experiments taken separately.
Shared oscillation parameters
When the two experiments also fit common parameters, such as or , the IO solution of each experiment is a vector with covariance , twice the inverse of the Hessian of its IO at the minimum, and the same minimization gives
The difference in then counts together with the differences in the other parameters, all weighted by the inverse of . In the flat valley of the variance of an experiment in is unbounded, the component of along the flat valley drops out, and the joint fit takes the value of that brings the two splittings closest, as in Figure 10.
In the left panel of Figure 10 the upper edge of the NOvA–DeepCore band comes from true values of of 0.53 or more, where the IO fit of NOvA returns the octant mirror of the true and that of DeepCore stays in the upper octant. In both panels, a joint fit of the two samples, which shares and , gives a larger synergy than eq. (1) at almost all true values of and , and its largest values come from a true in the lower octant, where eq. (1) gives nearly zero.
References
- [1]JUNO collaboration, Measurement of Neutrino Oscillation Parameters at JUNO and Indication of the Neutrino Mass Ordering, 2609.37895.
- [2]DAYA BAY collaboration, Precision Measurement of Reactor Antineutrino Oscillation at Kilometer-Scale Baselines by Daya Bay, Phys. Rev. Lett. 130 (2023) 161802 [2211.14988].DOI
- [3]NOvA collaboration, Precision Measurement of Neutrino Oscillation Parameters with 10 Years of Data from the NOvA Experiment, Phys. Rev. Lett. 136 (2026) 011802 [2509.04361].DOI
- [4]T2K collaboration, Measurements of neutrino oscillation parameters from the T2K experiment using 3.6 × 10²¹ protons on target, Eur. Phys. J. C 83 (2023) 782 [2303.03222].DOI
- [5]T2K collaboration, Results from the T2K Experiment on Neutrino Mixing Including a New Far Detector μ-like Sample, Phys. Rev. Lett. 135 (2025) 261801 [2506.05889].DOI
- [6]IceCube collaboration, Measurement of atmospheric neutrino mixing with improved IceCube DeepCore calibration and data processing, Phys. Rev. D 108 (2023) 012014 [2304.12236].
- [7]KM3NeT collaboration, Measurement of neutrino oscillation parameters with the first six detection units of KM3NeT/ORCA, JHEP 10 (2024) 206 [2408.07015].
- [8]Super-Kamiokande collaboration, Atmospheric neutrino oscillation analysis with neutron tagging and an expanded fiducial volume in Super-Kamiokande I–V, Phys. Rev. D 109 (2024) 072014 [2311.05105].
- [9]JUNO collaboration, Neutrino Physics with JUNO, J. Phys. G 43 (2016) 030401 [1507.05613].DOI
- [10]JUNO collaboration, JUNO physics and detector, Prog. Part. Nucl. Phys. 123 (2022) 103927 [2104.02565].
- [11]JUNO collaboration, Potential to identify neutrino mass ordering with reactor antineutrinos at JUNO, Chin. Phys. C 49 (2025) 033104 [2405.18008].
- [12]S.T. Petcov and M. Piai, The LMA MSW solution of the solar neutrino problem, inverted neutrino mass hierarchy and reactor neutrino experiments, Phys. Lett. B 533 (2002) 94 [hep-ph/0112074].DOI
- [13]J. Learned, S.T. Dye, S. Pakvasa and R.C. Svoboda, Determination of neutrino mass hierarchy and θ₁₃ with a remote detector of reactor antineutrinos, Phys. Rev. D 78 (2008) 071302 [hep-ex/0612022].DOI
- [14]Y.-F. Li, J. Cao, Y. Wang and L. Zhan, Unambiguous Determination of the Neutrino Mass Hierarchy Using Reactor Neutrinos, Phys. Rev. D 88 (2013) 013008 [1303.6733].DOI
- [15]T2K and NOvA collaborations, Joint neutrino oscillation analysis from the T2K and NOvA experiments, Nature 646 (2025) 818 [2510.19888].
- [16]L. Wolfenstein, Neutrino Oscillations in Matter, Phys. Rev. D 17 (1978) 2369.
- [17]S.P. Mikheyev and A.Y. Smirnov, Resonance Amplification of Oscillations in Matter and Spectroscopy of Solar Neutrinos, Sov. J. Nucl. Phys. 42 (1985) 913.
- [18]E.K. Akhmedov, S. Razzаque and A.Y. Smirnov, Mass hierarchy, 2-3 mixing and CP-phase with Huge Atmospheric Neutrino Detectors, JHEP 02 (2013) 082 [Erratum ibid. 07 (2013) 026] [1205.7071].
- [19]IceCube collaboration, Sensitivity of the IceCube Upgrade to Atmospheric Neutrino Oscillations, PoS ICRC2023 (2023) 1036 [2307.15295].
- [20]A. de Gouvêa, J. Jenkins and B. Kayser, Neutrino mass hierarchy, vacuum oscillations, and vanishing |Ue3|, Phys. Rev. D 71 (2005) 113009 [hep-ph/0503079].DOI
- [21]H. Nunokawa, S.J. Parke and R. Zukanovich Funchal, Another possible way to determine the neutrino mass hierarchy, Phys. Rev. D 72 (2005) 013009 [hep-ph/0503283].DOI
- [22]X. Qian and P. Vogel, Neutrino Mass Hierarchy, Prog. Part. Nucl. Phys. 83 (2015) 1 [1505.01891].DOI
- [23]A. Cabrera et al., Synergies and prospects for early resolution of the neutrino mass ordering, Sci. Rep. 12 (2022) 5393 [2008.11280].DOI
- [24]S.J. Parke and R. Zukanovich Funchal, Mass ordering sum rule for the neutrino disappearance channels in T2K, NOvA, and JUNO, Phys. Rev. D 111 (2025) 013008 [2404.08733].DOI
- [25]H. Minakata, H. Nunokawa, S.J. Parke and R. Zukanovich Funchal, Determining neutrino mass hierarchy by precision measurements in electron and muon neutrino disappearance experiments, Phys. Rev. D 74 (2006) 053008 [hep-ph/0607284].DOI
- [26]M. Blennow and T. Schwetz, Determination of the neutrino mass ordering by combining PINGU and Daya Bay II, JHEP 09 (2013) 089 [1306.3988].DOI
- [27]IceCube-Gen2 collaboration, Combined sensitivity to the neutrino mass ordering with JUNO, the IceCube Upgrade, and PINGU, Phys. Rev. D 101 (2020) 032006 [1911.06745].DOI
- [28]KM3NeT and JUNO collaborations, Combined sensitivity of JUNO and KM3NeT/ORCA to the neutrino mass ordering, JHEP 03 (2022) 055 [2108.06293].
- [29]T2K and Super-Kamiokande collaborations, First Joint Oscillation Analysis of Super-Kamiokande Atmospheric and T2K Accelerator Neutrino Data, Phys. Rev. Lett. 134 (2025) 011801 [2405.12488].
- [30]P.A.N. Machado and J.P. Pinheiro, Where does the neutrino Mass Ordering preference come from?, 2611.13629.arxiv.org/abs/2611.13629
- [31]M. Jin, C.A. Argüelles, P. Fernández-Menéndez and I. Martínez-Soler, Evidence for Normal Neutrino Ordering with Combined Atmospheric Neutrino Experiments, 2610.12208.arxiv.org/abs/2610.12208
- [32]I. Esteban, M.C. Gonzalez-Garcia, M. Maltoni, I. Martinez-Soler, J.P. Pinheiro and T. Schwetz, NuFit-6.0: updated global analysis of three-flavor neutrino oscillations, JHEP 12 (2024) 216 [2410.05380].DOI
- [33]F. Capozzi, W. Giarè, E. Lisi, A. Marrone, A. Melchiorri and A. Palazzo, Neutrino masses and mixing: Entering the era of subpercent precision, Phys. Rev. D 111 (2025) 093006 [2503.07752].arxiv.org/abs/2503.07752
- [34]P.F. de Salas, D.V. Forero, S. Gariazzo, P. Martínez-Miravé, O. Mena, C.A. Ternes et al., 2020 global reassessment of the neutrino oscillation picture, JHEP 02 (2021) 071 [2006.11237].DOI
- [35]I. Esteban, M.C. Gonzalez-Garcia, M. Maltoni, I. Martinez-Soler, J.P. Pinheiro and T. Schwetz, Lessons from the first JUNO results, JHEP 04 (2026) 089 [2601.09791].DOI
- [36]H. Minakata, H. Nunokawa, S.J. Parke and R. Zukanovich Funchal, Determination of the Neutrino Mass Hierarchy via the Phase of the Disappearance Oscillation Probability with a Monochromatic ν̄e Source, Phys. Rev. D 76 (2007) 053004 [Erratum ibid. 76 (2007) 079901] [hep-ph/0701151].arxiv.org/abs/hep-ph/0701151
- [37]S. Parke, What is Δm²ee ?, Phys. Rev. D 93 (2016) 053008 [1601.07464].
- [38]T2K collaboration, The T2K Experiment, Nucl. Instrum. Meth. A 659 (2011) 106 [1106.1238].
- [39]NOvA collaboration, Improved measurement of neutrino oscillation parameters by the NOvA experiment, Phys. Rev. D 106 (2022) 032004 [2108.08219].DOI
- [40]IceCube collaboration, The Design and Performance of IceCube DeepCore, Astropart. Phys. 35 (2012) 615 [1109.6096].DOI
- [41]KM3NeT collaboration, Determining the neutrino mass ordering and oscillation parameters with KM3NeT/ORCA, Eur. Phys. J. C 82 (2022) 26 [2103.09885].DOI
- [42]V. Cerrone and M. Grassi, “nmo-synergies: wrong-ordering solutions of reactor, accelerator, and atmospheric neutrino experiments as weighted bin averages.” https://github.com/vanessacerrone/nmo-synergies, 2026.
- [43]G. Cowan, K. Cranmer, E. Gross and O. Vitells, Asymptotic formulae for likelihood-based tests of new physics, Eur. Phys. J. C 71 (2011) 1554 [Erratum ibid. 73 (2013) 2501] [1007.1727].DOI
- [44]S. Baker and R.D. Cousins, Clarification of the Use of Chi Square and Likelihood Functions in Fits to Histograms, Nucl. Instrum. Meth. 221 (1984) 437.DOI
- [45]X. Ji, W. Gu, X. Wang, X. Qian and T. Xin, Combined Neyman-Pearson chi-square: An improved approximation to the Poisson-likelihood chi-square, Nucl. Instrum. Meth. A 961 (2020) 163677 [1903.07185].DOI
- [46]Particle Data Group collaboration, Review of Particle Physics, Int. J. Mod. Phys. A 41 (2026) 2630011.
- [47]M. Blennow, P. Coloma, P. Huber and T. Schwetz, Quantifying the sensitivity of oscillation experiments to the neutrino mass ordering, JHEP 03 (2014) 028 [1311.1822].DOI
- [48]X. Qian, A. Tan, W. Wang, J.J. Ling, R.D. McKeown and C. Zhang, Statistical Evaluation of Experimental Determinations of Neutrino Mass Hierarchy, Phys. Rev. D 86 (2012) 113011 [1210.3651].DOI
- [49]E. Ciuffoli, J. Evslin and X. Zhang, Confidence in a neutrino mass hierarchy determination, JHEP 01 (2014) 095 [1305.5150].DOI
- [50]G.L. Fogli, E. Lisi, A. Marrone, D. Montanino and A. Palazzo, Getting the most from the statistical analysis of solar neutrino oscillations, Phys. Rev. D 66 (2002) 053010 [hep-ph/0206162].DOI
- [51]S.J. Parke, H. Minakata, H. Nunokawa and R. Zukanovich Funchal, Mass Hierarchy via Mössbauer and Reactor Neutrinos, Nucl. Phys. B Proc. Suppl. 188 (2009) 115 [0812.1879].DOI
- [52]X. Qian, D.A. Dwyer, R.D. McKeown, P. Vogel, W. Wang and C. Zhang, Mass Hierarchy Resolution in Reactor Anti-neutrino Experiments: Parameter Degeneracies and Detector Energy Response, Phys. Rev. D 87 (2013) 033005 [1208.1551].DOI
- [53]JUNO collaboration, Prediction of Energy Resolution in the JUNO Experiment, Chin. Phys. C 49 (2025) 013003 [2405.17860].
- [54]NOvA collaboration, Z. Vallari, “Probing the 3-Flavor paradigm with NOvA.” Talk at Neutrino 2026, XXXII International Conference on Neutrino Physics and Astrophysics, Irvine, USA, 23 June 2026, slide 27, https://indico.global/event/15740/contributions/155585, 2026.
- [55]P.B. Denton and S.J. Parke, Smallness of matter effects in long-baseline muon neutrino disappearance, Phys. Rev. D 109 (2024) 053002 [2401.10326].DOI
- [56]L. Zhan, Y. Wang, J. Cao and L. Wen, Determination of the Neutrino Mass Hierarchy at an Intermediate Baseline, Phys. Rev. D 78 (2008) 111103 [0807.3203].DOI
- [57]S. Choubey, M. Ghosh and D. Raikwal, Neutrino mass ordering: Circumventing the challenges using synergy between T2HK and JUNO, Phys. Rev. D 106 (2022) 115013 [2207.04784].DOI
- [58]JUNO collaboration, Sub-percent precision measurement of neutrino oscillation parameters with JUNO, Chin. Phys. C 46 (2022) 123001 [2204.13249].
- [59]Y.-F. Li, Y. Wang and Z.-z. Xing, Terrestrial matter effects on reactor antineutrino oscillations at JUNO or RENO-50: how small is small?, Chin. Phys. C 40 (2016) 091001 [1605.00900].DOI
- [60]A.N. Khan, H. Nunokawa and S.J. Parke, Why matter effects matter for JUNO, Phys. Lett. B 803 (2020) 135354 [1910.12900].DOI
- [61]D.V. Forero, S.J. Parke, C.A. Ternes and R. Zukanovich Funchal, JUNO’s prospects for determining the neutrino mass ordering, Phys. Rev. D 104 (2021) 113004 [2107.12410].DOI
- [62]E.K. Akhmedov, M. Maltoni and A.Y. Smirnov, 1-3 leptonic mixing and the neutrino oscillograms of the Earth, JHEP 05 (2007) 077 [hep-ph/0612285].DOI
- [63]W. Winter, Neutrino mass hierarchy determination with IceCube-PINGU, Phys. Rev. D 88 (2013) 013013 [1305.5539].DOI
- [64]F. Capozzi, E. Lisi and A. Marrone, PINGU and the neutrino mass hierarchy: Statistical and systematic aspects, Phys. Rev. D 91 (2015) 073011 [1503.01999].DOI
- [65]M.C. Banuls, G. Barenboim and J. Bernabeu, Medium effects for terrestrial and atmospheric neutrino oscillations, Phys. Lett. B 513 (2001) 391 [hep-ph/0102184].DOI
- [66]J. Bernabeu, S. Palomares Ruiz and S.T. Petcov, Atmospheric neutrino oscillations, θ₁₃ and neutrino mass hierarchy, Nucl. Phys. B 669 (2003) 255 [hep-ph/0305152].DOI
- [67]S.T. Petcov and T. Schwetz, Determining the neutrino mass hierarchy with atmospheric neutrinos, Nucl. Phys. B 740 (2006) 1 [hep-ph/0511277].DOI
- [68]R. Gandhi, P. Ghoshal, S. Goswami, P. Mehta, S.U. Sankar and S. Shalgar, Mass Hierarchy Determination via future Atmospheric Neutrino Detectors, Phys. Rev. D 76 (2007) 073012 [0707.1723].DOI
- [69]O. Mena, I. Mocioiu and S. Razzaque, Neutrino mass hierarchy extraction using atmospheric neutrinos in ice, Phys. Rev. D 78 (2008) 093003 [0803.3044].DOI
- [70]A.M. Dziewonski and D.L. Anderson, Preliminary reference earth model, Phys. Earth Planet. Inter. 25 (1981) 297.
- [71]P.B. Denton and S.J. Parke, Efficient atmospheric, solar, and supernova neutrino propagation through the Earth, Phys. Rev. D 113 (2026) 053002 [2511.04735].DOI
- [72]E.K. Akhmedov, R. Johansson, M. Lindner, T. Ohlsson and T. Schwetz, Series expansions for three flavor neutrino oscillation probabilities in matter, JHEP 04 (2004) 078 [hep-ph/0402175].
- [73]H. Minakata and S.J. Parke, Simple and Compact Expressions for Neutrino Oscillation Probabilities in Matter, JHEP 01 (2016) 180 [1505.01826].DOI
- [74]P.B. Denton, H. Minakata and S.J. Parke, Compact Perturbative Expressions For Neutrino Oscillations in Matter, JHEP 06 (2016) 051 [1604.08167].
- [75]K. Kimura, A. Takamura and H. Yokomakura, Exact formulas and simple CP dependence of neutrino oscillation probabilities in matter with constant density, Phys. Rev. D 66 (2002) 073005 [hep-ph/0205295].DOI
- [76]H. Yokomakura, K. Kimura and A. Takamura, Overall feature of CP dependence for neutrino oscillation probability in arbitrary matter profile, Phys. Lett. B 544 (2002) 286 [hep-ph/0207174].DOI
- [77]E.K. Akhmedov, M. Maltoni and A.Y. Smirnov, Neutrino oscillograms of the Earth: Effects of 1-2 mixing and CP-violation, JHEP 06 (2008) 072 [0804.1466].DOI
- [78]IceCube collaboration, “Replication Data for: Measurement of atmospheric neutrino mixing with improved IceCube DeepCore calibration and data processing.” Harvard Dataverse, version 1.0, doi:10.7910/DVN/B4RITM, 2025.DOI
- [79]S. Choubey and A. Ghosh, Determining the Octant of θ₂₃ with PINGU, T2K, NOνA and Reactor Data, JHEP 11 (2013) 166 [1309.5760].DOI
- [80]S. Choubey and P. Roy, Probing the deviation from maximal mixing of atmospheric neutrinos, Phys. Rev. D 73 (2006) 013006 [hep-ph/0509197].DOI
- [81]KM3NeT collaboration, “Oscillation Analysis ORCA433kt-y / ORCA 6.” KM3NeT Open Data Centre, dataset version 2.3, data file v0.5, https://opendata.km3net.de/dataset.xhtml?persistentId=doi:10.5072/FK2/YOUXVW, 2025.
- [82]KM3NeT collaboration, Addendum: Measurement of neutrino oscillation parameters with the first six detection units of KM3NeT/ORCA, JHEP 10 (2025) 041.DOI
- [83]NOvA collaboration, “NOvA 2024 official data release (26.61E20 neutrino + 12.5E20 antineutrino).” FERMILAB-DATA-2025-05, Zenodo, version of 9 July 2026, doi:10.5281/zenodo.21284589, 2026.DOI
- [84]NOvA collaboration, “Flux Histograms for 2017 Analysis.” FERMILAB-DATA-2017-01, Fermilab public document 8, version 1, doi:10.15484/1959359, 2017.
- [85]T2K collaboration, “2020 T2K flux prediction file release.” https://t2k-experiment.org/results/neutrino-beam-flux-prediction-2020/, 2021.
- [86]T2K collaboration, G. Žarnecki, “Overview of the T2K experiment.” Talk at the 9th International Conference on High Energy Particle and Nuclear Physics in the LHC Era (HEPNP 2025), Valparaíso, Chile, 8 January 2025, https://indico.global/event/8670/contributions/82766, 2025.
- [87]IceCube collaboration, Measurement of Atmospheric Neutrino Oscillation Parameters Using Convolutional Neural Networks with 9.3 Years of Data in IceCube DeepCore, Phys. Rev. Lett. 134 (2025) 091801 [2405.02163].arxiv.org/abs/2405.02163
- [88]IceCube collaboration, “Data release for neutrino oscillation parameters using convolutional neural networks with 9.3 years of data in IceCube DeepCore.” Harvard Dataverse, version 2.0, doi:10.7910/DVN/U20MMB, 2024.DOI
- [89]D. Raikwal, S. Choubey and M. Ghosh, Determining neutrino mass ordering with ICAL, JUNO and T2HK, Eur. Phys. J. Plus 138 (2023) 110 [Erratum ibid. 138 (2023) 485] [2207.06798].DOI
- [90]NOvA collaboration, Determining Neutrino Mass Ordering with NOvA and Upcoming JUNO Measurements, 2606.14121.arxiv.org/abs/2606.14121
- [91]G.F.S. Alves, H. Nunokawa and R. Zukanovich Funchal, Impact of new physics on the JUNO-long-baseline synergy in the neutrino mass ordering determination, Phys. Rev. D 113 (2026) 115059 [2603.17181].arxiv.org/abs/2603.17181
- [92]C. Andreopoulos et al., The GENIE Neutrino Monte Carlo Generator, Nucl. Instrum. Meth. A 614 (2010) 87 [0905.2517].
- [93]J. Nieves, I. Ruiz Simo and M.J. Vicente Vacas, Inclusive Charged-Current Neutrino–Nucleus Reactions, Phys. Rev. C 83 (2011) 045501 [1102.2777].DOI
- [94]J. Schwehr, D. Cherdack and R. Gran, GENIE implementation of IFIC Valencia model for QE-like 2p2h neutrino-nucleus cross section, 1601.02038.arxiv.org/abs/1601.02038
- [95]M. Honda, M. Sajjad Athar, T. Kajita, K. Kasahara and S. Midorikawa, Atmospheric neutrino flux calculation using the NRLMSISE-00 atmospheric model, Phys. Rev. D 92 (2015) 023004 [1502.03916].DOI
- [96]P. Huber, On the determination of anti-neutrino spectra from nuclear reactors, Phys. Rev. C 84 (2011) 024617 [Erratum ibid. 85 (2012) 029901] [1106.0687].DOI
- [97]T.A. Mueller et al., Improved Predictions of Reactor Antineutrino Spectra, Phys. Rev. C 83 (2011) 054615 [1101.2663].DOI
- [98]A. Strumia and F. Vissani, Precise quasielastic neutrino/nucleon cross-section, Phys. Lett. B 564 (2003) 42 [astro-ph/0302055].DOI
- [99]P.B. Denton and S.J. Parke, Fast and accurate algorithm for calculating long-baseline neutrino oscillation probabilities with matter effects, Phys. Rev. D 110 (2024) 073005 [2405.02400].DOI
- [100]NOvA collaboration, Constraints on Oscillation Parameters from νₑ Appearance and νμ Disappearance in NOvA, Phys. Rev. Lett. 118 (2017) 231801 [1703.03328].DOI
- [101]T2K collaboration, Improved constraints on neutrino mixing from the T2K experiment with 3.13 × 10²¹ protons on target, Phys. Rev. D 103 (2021) 112008 [2101.03779].
- [102]Hyper-Kamiokande collaboration, Hyper-Kamiokande Design Report, 1805.04163.arxiv.org/abs/1805.04163
- [103]IceCube collaboration, Measurement of Atmospheric Neutrino Oscillations at 6–56 GeV with IceCube DeepCore, Phys. Rev. Lett. 120 (2018) 071801 [1707.07081].DOI
- [104]S.S. Wilks, The Large-Sample Distribution of the Likelihood Ratio for Testing Composite Hypotheses, Annals Math. Statist. 9 (1938) 60.DOI
- [105]M. Maltoni, T. Schwetz, M.A. Tortola and J.W.F. Valle, Ruling out four neutrino oscillation interpretations of the LSND anomaly?, Nucl. Phys. B 643 (2002) 321 [hep-ph/0207157].DOI
- [106]M. Maltoni and T. Schwetz, Testing the statistical compatibility of independent data sets, Phys. Rev. D 68 (2003) 033020 [hep-ph/0304176].DOI













