WorksPhysics and Astronomy
Entanglement in Collective Neutrino Oscillations on a Quantum Computer
Abstract
Simulations on a quantum computer have been propounded as a next-generation tool for studying neutrino flavour oscillations in dense environments like supernovae or neutron star mergers. However, quantum advantage hinges on the presence of quantum entanglement in the neutrino ensemble, and it has been argued that such entanglement becomes negligible in the limit of large neutrino number density. Here, we address this question by explicitly simulating the multiparticle dynamics of entangled neutrino ensembles with a varying number of particles, , on both classical and quantum hardware, and comparing our results to those obtained with conventional ``mean-field'' methods based on Boltzmann-type kinetic equations. For simple toy systems, we demonstrate both analytically and numerically that, indeed, the predictions of full multiparticle simulations agree with those of mean-field methods in the large- limit. Along the way, we develop novel algorithms for the efficient simulation of entangled neutrino gases on both classical and quantum hardware. We then discuss physical situations in which an entangled neutrino system cannot be described through mean-field dynamics alone and quantum entanglement therefore becomes phenomenologically relevant. We comment on the conditions under which these scenarios may be realized in supernova cores or neutron star mergers.
Introduction
Neutrinos are an ideal laboratory for studying the intricacies of quantum mechanics, given that their extremely weak interactions allow them to maintain quantum properties like superpositions and entanglement for extended periods of time. An interesting question in this context is under which circumstances the flavours of different neutrinos can become entangled, and whether this entanglement can have observable consequences. In fact, it has been argued that ultradense environments like supernova cores, neutron star mergers, or the early Universe may provide conditions where such entanglement occurs [1, 2, 3, 4, 5, 6, 7, 8]. In these environments, neutrino densities are so high that neutrino–neutrino coherent forward scattering [9, 10], wherein the neutrino flavour may change while momenta remain unaltered, becomes non-negligible [11, 12, 13, 14, 15, 16]. This leads to a plethora of novel, “collective,” effects [17, 18, 19, 20], an understanding of which is vital in the context of supernova neutrino observations whose goal is to infer the conditions in the supernova core (before oscillations) from the neutrino spectra, “light” curves (luminosity as a function of time), and the flavour composition observed upon reaching Earth [21, 22]. In addition, the neutrino flavour distribution plays a pivotal role in the supernova explosion mechanism itself and in the associated nucleosynthesis processes [23, 24, 25]. To this day, however, our understanding of how neutrinos change flavour on their way out of a supernova is fragmented. The problem is usually formulated using Boltzmann-type quantum kinetic equations (“mean-field approach”) [26], but solving these equations for systems that go beyond simple toy models has proven to be a formidable challenge. To describe conditions in a real supernova, one is facing a multiscale problem in six phase space dimensions, and with a background of ordinary matter that is itself highly asymmetric and turbulent. Consequently, a complete understanding of supernova neutrino flavour evolution remains elusive.
Quantum entanglement, which is not captured by quantum kinetic equations, would add a further layer of complications, and would pose a formidable computational challenge, given that the size of the relevant Hilbert space scales exponentially with the number of simulated neutrinos. Collectively oscillating neutrino systems with entanglement may therefore call for simulations on a quantum computer, which would reduce the computational complexity to polynomial in . Pioneering work on this topic has been carried out in refs. [27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39]. However, while there is a clear consensus that entanglement is phenomenologically important in the few-particle toy models that can be studied on current quantum hardware, there are arguments suggesting that it becomes phenomenologically irrelevant at the very high neutrino densities, , realized in supernovae and similar environments [40, 41, 42, 43]. For some toy systems, this can be shown explicitly (see e.g. Refs. [40, 41] and § 1.2.0.1 below). Moreover, in a three-dimensional system like a supernova or neutron star merger, with neutrinos streaming in all directions, the probability that any pair of neutrinos gets more than one chance to interact is negligible. This “once-in-a-lifetime encounter” argument [44, 45] leads to the conjecture that entanglement cannot build up efficiently in realistic environments. Both arguments would suggest that quantum kinetic equations provide a sufficiently accurate description in practice. Overall, the quantum kinetic (Boltzmann) approach to collective neutrino oscillations has not been proven wrong, but possible shortcomings due to its inability to capture entanglement have not been ruled out either.
The goals of the present work are twofold: first, we aim to study explicitly the emergence of Boltzmann-like “mean-field” behaviour in fully entangled systems of self-interacting neutrinos in the large- limit. We will do so using analytical calculations, multiparticle simulations on classical hardware, and simulations on superconducting quantum processors. Second, we will introduce neutrino systems in which the preceding arguments do not hold and quantum entanglement among neutrinos becomes phenomenologically relevant. We will discuss under which conditions these systems may be realized in nature.
The paper is structured as follows: in Section 2 we first review the quantum kinetic approach to collective neutrino oscillations (Section 2.1). For specific systems consisting of just two neutrino modes with similar numbers of neutrinos in each (Section 2.2), we then argue analytically that quantum kinetic behaviour emerges from a full multiparticle treatment including entanglement in the limit of large system size (Section 2.3). In Section 3, we describe the quantum simulations we have carried out, both on conventional hardware and on superconducting quantum processors. In this context, we set out the general formalism (Section 3.1), develop new, highly optimized algorithms based on our previous work in Ref. [38] (Sections 3.2 and 3.3), and we present the results of our simulation (Section 3.4). Finally, in Section 4, we introduce two more toy systems, inspired by the actual dynamics inside a supernova, to which the preceding arguments do not apply, and in which quantum entanglement may therefore be phenomenologically relevant. Specifically, we consider a linear stream of neutrinos with overlapping wave packets in Section 4.1, as well as systems in which multiple once-in-a-lifetime encounters lead to the build-up of entanglement in Section 4.2. We conclude in Section 5.
Quantum Kinetic Equations vs. Multiparticle Entanglement
The Quantum Kinetic Equations for Collective Neutrino Oscillations
We begin by briefly reviewing the quantum kinetic approach to collective neutrino oscillations (sometimes called the “mean-field approach”), based on pioneering work by Sigl and Raffelt [26]. In this formalism, the von Neumann equation describing vacuum oscillations, Mikheyev–Smirnov–Wolfenstein (MSW) matter effects [9, 10], neutrino–neutrino interactions, and neutrino–antineutrino interactions takes the form [17]
where () is the neutrino (antineutrino) density matrix at position , momentum , and time in the flavour basis . For simplicity, we work here in the two-flavour approximation, which is motivated by the fact that and behave similarly in supernovae, neutron star mergers, and the early Universe (see, however, Refs. [46, 47, 48, 49, 50]). The two terms on the left-hand side of Eq. (1) describe time evolution and spatial transport, respectively. The first term in the Hamiltonian on the right-hand side encapsulates vacuum oscillations and MSW matter effects. It is given by
where is the neutrino energy, the leptonic mixing matrix, the mass squared difference, the Fermi constant, and the net electron number density in the background matter, which controls the MSW potential. The second term on the right-hand side of Eq. (1) describes neutrino self-interactions. Note in particular the angular dependence: the self-interaction term depends on the angle between the unit vectors and .
To solve Eq. (1), and are discretized, leading to coupled equations, one for each mode . The main assumption that justifies working with separate density matrices instead of a single matrix is “molecular chaos”, that is, the assumption that the duration of each interaction is short compared to the time a neutrino takes to travel the mean free path between interactions, and that any quantum correlations built up during an interaction is “forgotten” before the next interaction happens [26, 43, 44]. Equation (1) is not valid when a description of the system in terms of individual particles breaks down, for instance because the size of the neutrino wave packets is larger than the mean free path, or when quantum entanglement builds up over multiple interactions.
The condition for the validity of the quantum kinetic approach can be recast into the requirement that is not too small [40, 41]. The apparent contradiction to the above statement that the mean free path should be large is resolved as follows: to leave the collective forward scattering interaction strength invariant as is varied, a change in needs to be accompanied by an opposite change in the effective coupling constant . Therefore, as , increasing indeed increases the mean free path.
Our goal in the following will be to explicitly demonstrate the emergence of quantum kinetic behaviour in toy systems as is increased.
multiparticule treatments. The solution of the quantum-kinetic equations, Eq. (1), is shown as a black curve in the top panel of Figure 1. The other curves in this plot, showing how the full multiparticle solution converges to the quantum kinetic one in the large- limit, will be discussed in § 1.2.1.
Figure 1. Comparison of the quantum kinetic mean-field treatment of collective neutrino oscillations (black curves, see § 1.1.1) to the results of full multiparticle simulations (see § 1.2.1) on classical hardware (purple curves) and on a superconducting transmon quantum processor (blue data points). The blue dashed line corresponds to a noiseless simulation of the Trotterized quantum circuit. Top: a two-mode system with initially electron neutrinos and muon neutrinos, exhibiting both slow and fast oscillations. Middle: bipolar oscillations in a system with neutrinos and antineutrinos. Bottom: root-mean-square residuals between the quantum kinetic mean-field results and the multiparticle simulation as a function of the number of neutrinos per mode.
Example: Two-mode oscillations
As a first example, we consider a toy system with initially electron neutrinos and muon neutrinos. This is one of the systems for which ref. [40] has demonstrated analytically the equivalence between the mean-field and
The second system we consider consists of neutrinos that interact with antineutrinos. Flavour conversion in such a system can exhibit highly anharmonic behaviour (so-called “bipolar” oscillations) [17, 51, 52], as shown in the middle panel of Figure 1. This behaviour can be understood in the “flavour pendulum” picture from refs. [51, 52], where the neutrino and antineutrino density matrices, averaged over and , are represented by a flavour polarization vector in three dimensions via and , with the Pauli matrices . In this picture, an initial state consisting of roughly equal numbers of and corresponds to and pointing in opposite directions, forming a “pendulum” in an unstable equilibrium configuration. The small perturbation of this equilibrium configuration due to the vacuum oscillation term in the Hamiltonian leads to the sudden, anharmonic, flavour flips observed in Figure 1.
In the following, we demonstrate explicitly how the quantum kinetic description of these two-mode systems emerges from the full multiparticle dynamics. We begin with analytic arguments, which we then confirm using quantum simulations in § 1.2.1.
Emergence of quantum kinetic behaviour in the large- limit
Following ref. [40], we show analytically how quantum kinetic behaviour emerges in the large- limit for bipolar oscillations between neutrinos and antineutrinos. (The analogous calculation for the – system from the top panel of Figure 1 can be found in ref. [40].) We assume that a beam of neutrinos are initially in a state
while a beam of antineutrinos are in a state
An ensemble of neutrinos that is originally in a flavour eigenstate will eventually evolve into such a superposition due to vacuum oscillations. We also introduce the states orthogonal to and , which are given by
and satisfy . It is useful to transform into the basis according to
Coherent flavour changes in this system can occur either through ordinary oscillations in matter or through coherent forward scattering. The latter can be mediated by -channel interactions of the form , and -channel interactions like and . -channel processes (e.g. ) cannot cause coherent interference due to the final state containing an antineutrino in the neutrino beam and vice-versa. The same is true for the additional -channel processes of the form , .
Dropping the non-coherent processes, the neutrino interaction Hamiltonian in the basis is
Since terms proportional to the identity matrix only add a global phase under time evolution and therefore do not affect physical quantities, we can equivalently write
Dropping the contribution is equivalent to neglecting the -channel contributions of the form and , which are always flavour-conserving as they are mediated by the boson.
In one time step governed by the Hamiltonian in Eq. (8), a system consisting initially of and evolves according to6
Here, we have used the fact that acts non-trivially only on the component of the initial state, while leaving the component invariant. We have moreover defined .
We now generalize the two-particle interaction from Eq. (10) to the full -particle system. To simplify the notation, we write in the following, and analogously for , , . The evolution of the full system during an infinitesimal time step is given by
where
This expression sums the -channel interaction from Eq. (10) over all neutrino–antineutrino combinations. We now express and in the basis:
When computing flavour transition probabilities, we are ultimately interested in expectation values of the form , where is the operator that counts the number of in the neutrino ensemble (). It is given by
noting that projects the -th neutrino onto the flavour eigenstate . There is a similar operator for antineutrinos. Using Eqs. (11) and (14), we find
where
Dropping terms that are not coherently enhanced – that is, terms that are not proportional to at least four powers of or – we see that the last three terms in Eq. (17) are negligible and can be removed. Therefore, for practical purposes, the last term in Eq. (14) can also be dropped, which significantly simplifies . Notably, we can now express the final state after an infinitesimal time step as a product state of the form
up to first order in , where
This shows that, in the limit of large and , the system can be written as a product state without entanglement after a single time step. Relabelling and for the next time step, we see that factorization continues to hold after an arbitrary number of time steps. Consequently, the system never develops significant entanglement, which justifies a description using quantum kinetic equations.
One may wonder whether the picture changes if entanglement is already present in the initial state. (For the moment, we leave aside the question of how such entanglement could have been created in a realistic astrophysical or cosmological environment.) As an example, we introduce an entangled state of the form
where parameterizes the amount of entanglement. The interaction between the neutrino and antineutrino beams becomes
where is given by Eq. (14) and reads, in analogy to Eq. (14),
As in Eq. (14), the last term is not amplified coherently and can be dropped. We can then once more express the final state as a sum of product states
where
Therefore, despite the initial state being entangled, no additional entanglement builds up over time. In the bipolar system, the quantum kinetic approach is hence justified in the large and limit even in the case where entanglement is initially present. We expect that this conclusion generalizes to more complex systems.
Quantum Simulation of Multiparticle Dynamics
Having demonstrated analytically the emergence of the mean-field limit at large and in simple toy systems, we now confirm this result using full multiparticle simulations with entanglement on both classical and quantum hardware.
Quantum Circuits for General Systems
In the simplest quantum circuits for collective neutrino oscillations in the two-flavour approximation, the flavour eigenstates and of each neutrino are mapped onto qubit states according to , . The Hamiltonian of the system is
The first term, describing vacuum oscillations and MSW interactions with ordinary matter, has the form
where is the vector of Pauli matrices acting on the -th qubit and , where is the oscillation frequency of the -th neutrino. is the mixing angle and the direction of in vacuum is only dependent on the mixing between lighter and heavier neutrino eigenstates. Both and are understood to be taken at their effective values in matter. The second term, describing self-interactions, is
where is the self-interaction strength and is the angle between the momentum vectors of modes and . In this paper, we will not concern ourselves with angular effects and simply treat as a free coupling constant. For neutrino–antineutrino interactions, the self-interaction term is replaced by
Since , it is possible to unify the description of , , and interactions by changing the mapping between flavour eigenstates and qubit states for antineutrinos to , and replacing for antineutrinos. After these changes, .
The -qubit quantum state is evolved in time using a second-order Suzuki–Trotter decomposition:
with
where is the number of time steps.
Since the Hamiltonian in Eq. (29) is invariant under permutations of neutrinos that share the same 4-momentum, i.e. the same and , we can use the correspondence between qubits and spin- particles to rewrite it as [38]
with the total spin vector. If the system is initially in an eigenstate of and , a so-called Dicke state, the second term can be dropped as is a conserved quantity. In order to describe multiple neutrino modes with different energies and propagation directions, and/or a system with mixtures of neutrinos and antineutrinos, Eq. (35) can be generalized to
where parameterizes vacuum oscillations and MSW matter effects with ordinary matter of the -th mode, is the self-interaction strength among neutrinos in mode , and is the interaction strength between modes and . If the system starts out in an eigenstate of the , the self-interaction terms can be dropped.
To encode Dicke states in quantum registers, we split the mode comprised of neutrinos into multiple new modes which are eigenstates of the corresponding with quantum numbers and . The Hilbert space of the -th mode is restricted by permutation symmetry to a subspace spanned by states with . This state can be encoded as a binary number in a register consisting of qubits via the computational-basis mapping
with , and prepared from a state in which all qubits are in the state by applying a series of Pauli gates. To evolve the system with time according to the Hamiltonian in Eq. (36), we also generalize the , , , , and gates to Dicke states. The Dicke approach, which exploits the symmetry of the Hamiltonian and relates it to algebra, is computationally advantageous, since the complexity of the Hilbert space on which Eq. (36) acts grows exponentially with the number of modes, while the complexity of the original Hamiltonian, Eq. (29), is exponential in the number of neutrinos. More details about conventional circuits for collective neutrino oscillations, Dicke states, and the encoding of Dicke states in quantum registers are given in ref. [38].
Optimized Multiparticle Quantum Simulation of Bipolar Oscillations on Classical Hardware
The two-mode oscillation scenarios introduced in § 1.1 feature a Hamiltonian structure that allows us to evolve the full multiparticle system on classical hardware even for . (Quantum-kinetic mean-field behaviour begins to emerge at – as we will see shortly.) In the following, we outline our algorithm to achieve this.
Consider for definiteness a system with neutrinos and antineutrinos, described according to the conventions outlined above as two collective spin vectors , with opposite vacuum oscillation terms, i.e., for . (For neutrinos and no antineutrinos, this would become ; the derivation remains otherwise unchanged.) From Eq. (36), the Hamiltonian of the bipolar system is
with and . We have again dropped the self-interaction terms, assuming the system to be in an eigenstate of and . These terms commute with , so assuming that the initial state is an eigenstate of with eigenvalue , the system will never leave the subspace corresponding to that eigenvalue.
In the limit of large self-coupling, where vacuum oscillations can be neglected (), the first (vacuum) term in Eq. (38) is negligible. In this case, the total flavour isospin along the direction, is conserved in the bipolar scenario [6, 38].
When the vacuum mixing angle , it is no longer the flavour isospin along the axis that is conserved, but instead the flavour isospin along a tilted axis. Namely,
Defining
the conserved quantity is given by the operator
The new quantization axis corresponds to working in the mass basis rather than the flavour basis. In the flavour basis, an spin oriented along the positive -axis () corresponds to a , while in the mass basis indicates the mass eigenstate. (We assume an inverted mass ordering here.)
In the unprimed basis, the initial state of the bipolar system with and corresponds to the neutrinos being in the eigenstate of , and the antineutrinos in the eigenstate of . In other words, each mode is in the Dicke state of maximal weight along the flavour axis. After a rotation into the primed basis, it remains a spin coherent state [53] with the highest-weight state rotated by the angle between the unprimed and primed axes. It remains a product state of identical spins, but it is not an eigenstate of . In the primed basis, a maximal-weight Dicke state expands to [54]
with
which is a binomial distribution. In the absence of vacuum oscillations (), the old and new -axes would be identical, and only the term in Eq. (42) would contribute.
As shown in Section A, the Hamiltonian in the primed basis has a tridiagonal structure, with the diagonal elements given by
and the off-diagonal ones by
On classical hardware, a sizable computational optimization is gained from evolving only the most relevant eigenstates of in time and dropping those which have only a small overlap with the initial state. We now briefly explain this procedure, with more details deferred to Section B.
The first step of the algorithm involves transforming the initial state from the flavour basis into the mass basis following Eq. (42). Then, if vacuum oscillations are subdominant compared to a strong interaction term, a limited set of subspaces carries most of the initial state amplitude, since the initial state is close to the highest-weight state of . Therefore, in performing the transformation in Eq. (42), we restrict the sum over to a contiguous set of sectors capturing all but a preset tolerance of the initial wave function norm. Only these are evolved in time using the Chebyshev expansion method [55], under the Hamiltonian defined by Eqs. (44) and (45). The components of the initial state that fall outside this range are dropped. The smaller the vacuum term, the fewer subspaces are needed to simulate the system.
A further optimization is possible during time-evolution by restricting the simulation in each sector to only the most relevant values states. Further details on this are given in Section B.
Simulating two-mode systems with neutrinos in each mode then becomes possible, and the resulting oscillation probabilities are shown in Figure 1. We will discuss these results in more detail in § 1.5.0.1 below. A total of 1 856 CPU hours was used to produce the results shown in the top panel of Figure 1. 9 548 CPU hours were used to produce the result in the middle panel. The neutrino–antineutrino case took noticeable longer to compute since the stronger interaction in that scenario required finer time steps. Additionally, a wider range of states needs to be tracked in this case: 94% of all states are “active” in the – system, compared to only 57% for the - system. The bottom panel of Figure 1 required 12 239 CPU hours for the – curves and 19 068 CPU hours for the – curve.
Quantum Algorithms with Efficient Encoding
On quantum hardware, a truncation scheme like the one described in § 1.2.1.2 would be difficult to implement, but it is also unnecessary. Instead, we need an optimization strategy that minimizes the number of (noisy) 2-qubit entangling gates and keeps the number of required qubits small. To achieve the second goal, we use the encoding from Eq. (37) again. For a two-mode system with particles, this leads to two quantum registers consisting of physical qubits each and providing a computational basis with dimension . Each register can represent the flavor state of physical neutrinos, encoded as a binary number, where the number corresponds to a state with all electron neutrinos (all muon antineutrinos), and the number 0 corresponds to all muon neutrinos (all electron antineutrinos). Expressed in terms of Dicke states, the neutrinos are represented by a spin vector of length , with , and register encodes the integer
where is the magnetic quantum number of mode . For both toy systems from § 1.1, register 1 is initialized to , corresponding to electron neutrinos, while the initial state of register 2 is corresponding to muon neutrinos or electron antineutrinos. We write this initial state as and a general basis state as . We use the convention that the -th qubit of a register carries the -th bit of the number , least significant digit first. In the mass basis (i.e., the primed frame from § 1.2.1.2), the total flavour quantum number , and hence is conserved.
Based on ref. [38], we introduce in the following an efficient algorithm for evolving the bipolar system in time using the above encoding. While tailored towards oscillations in two-mode scenarios, the algorithm remains generalizable to systems with more than two modes. A schematic diagram of the circuit is presented in Figure 2.
Figure 2. (a) Schematic circuit diagram of the quantum algorithm introduced in § 1.2.1.3. performs the initial state preparation. is a Draper adder for transforming the second register according to Eq. (47). denotes the time evolution under (even bonds) for a single Trotter step, with a diagonal matrix removed. Similarly, denotes the time evolution under with its diagonals removed in the same way. includes the decrement and increment of register 1. The – sequence is repeated times, where is the number of Trotter steps. is the single deferred diagonal operator. implements the measurement basis and performs single qubit rotations to measure in the , , or basis. The shown circuit applies a second order Trotter decomposition and subsequently merges adjacent half-step even blocks, leaving only the first and last half-step blocks. (b) Structure of the sub-circuit for a single Trotter step on the even bonds for with three qubits in each register. A pending diagonal followed by the exact multiplexer of Eq. (58) is equivalent to a modified multiplexer followed by a new pending diagonal . This replacement is applied recursively across subsequent layers to defer the diagonal operators until the final closing diagonal.
a Sum register. The fact that is conserved can be exploited by applying the transformation [56]
with
Afterwards, time-evolution affects only the first register. The inverse transformation, applied at the end of the circuit, is given by . It is exact since . The transformation can be implemented using the QFT-based adder of ref. [57], which applies a quantum Fourier transform to one register, a sequence of controlled phase rotations conditioned on the bits of the other register, and an inverse quantum Fourier transform. It does not require ancillary qubits.
b Hamiltonian. Unlike in § 1.2.1.2, where we worked with eigenstates of the total angular momentum operator , it is more convenient here to retain the individual quantum numbers under and . We write the Hamiltonian of the bipolar system, Eq. (38), as
with the ladder operators . (For the other toy example from § 1.1, namely a system with two neutrino modes rather than one neutrino mode and one antineutrino mode, only the sign in front of changes. In the following, we will work with the bipolar neutrino–antineutrino system for definiteness, which is computationally the more challenging of the two examples.) In the basis, the diagonal entries of this Hamiltonian are
where again . Similarly, the off-diagonal elements are
with
Here is the ladder-operator matrix element which can be expressed in terms of . Since , the off-diagonal element vanishes for or , which prevents transitions into unphysical states.
Time Evolution
To evolve the system under the Hamiltonian defined by Eqs. (50) and (51), we follow an idea from ref. [56], namely to split the Hamiltonian into operations that couple pairs of states which differ only in the least significant qubit of register 1, (“even bonds”), and those which affect pairs of the form (“odd bonds”). This has the advantage that, for even bonds, transitions between the two states and affect only a single qubit and are therefore easy to implement. Consequently, we split the Hamiltonian into , where contains the operations on even bonds as well as the diagonal components, and contains the operations on odd bonds.
In each Trotter step, we first apply . In the subspace spanned by the states and , that operator is a matrix . In terms of the diagonal elements from Eq. (50), we define their mean and half-difference,
We also use the shorthand notation
for the off-diagonal elements of from Eq. (51). Note that and . Defining finally , the time evolution over a time interval in the subspace is described by the operator
where and are Pauli matrices acting in the subspace. The full time evolution under is then
The sum over accounts for the (decoupled) sectors into which the initial state decomposes via Eq. (42).
As the two states and differ only in their least significant qubit, we can implement as a uniformly controlled gate following ref. [58]. A uniformly controlled gate implements a rotation on the least significant qubit, with the parameters of that rotation depending on all other qubits in the register [58]. Such an operation is sometimes also referred to as a multiplexer.
We then apply time evolution under the odd half of the Hamiltonian by first carrying out a decrement on register 1 (subtraction of 1 mod ), which maps each odd bond onto an even bond . We can then use the operator from Eq. (58) again with
where now and . Finally, a cyclic increment is applied to register 1 to undo the relabelling. We implement the cyclic increment as an ancilla-free cascade of multi-controlled gates [59]. It proceeds from the most significant digit down while flipping each qubit conditioned on all less significant qubits in the state, followed by a final gate on the least significant qubit. The decrement is the inverse gate.
Deferred diagonal block
The implementation of a uniformly controlled gate from Ref. [58] decomposes as , where is diagonal in the -dimensional Hilbert space and has a fixed structure of only CNOT gates and single qubit gates. The implementation of typically involves both single-qubit and two-qubit gates. It is straightforward to show that , where is also diagonal, though in general . Similarly, one can show that can be commuted past the cyclic increment / decrement operations, past the transform from Eq. (47), and past any single-qubit gate (at the expense of changing that single-qubit gate). Therefore, all diagonal operators appearing in the circuit can be commuted past the other operators to collect them in a single place, as a single diagonal operator. This leads to an important reduction in circuit depth.
Deployment on quantum hardware
We have deployed the algorithm described above on the IBM Kingston QPU, which uses the Heron R2 architecture. We have chosen the parameters for the self-coupling, for the vacuum mixing angle, and for the Trotter step size. only sets the overall energy scale and, as long as all other dimensional parameters are chosen relative to , does not enter the circuit. For producing the data points in the top panel of Figure 1, we have run the circuit with fixed for different total times, sampling the result on the time grid . For the middle panel of Figure 1, we have used and we have sampled the result at for the first 14 steps, and with twice the spacing for later times. Each run requires a patch of four adjacent qubits for placing the two registers with 3 neutrinos each. To pick the optimal patch based on real-time calibration data, we consider all qubits with readout error below 5% and all connections with CZ error below 1%. Among this set, we choose the patch with the lowest estimated error rate for the deepest circuit under consideration. For error mitigation, we have used the Qiskit Runtime Estimator class at resilience level 2. This includes Pauli twirling for gates and measurements [54, 60], zero-noise extrapolation at noise factors 1, 1.5, and 2 with a linear fit [61, 62], and XY4 dynamic decoupling [63]. For each noise factor, single-qubit pre-measurement rotations are applied to all qubits to measure the expectation value along each Pauli basis (, , and ) with 2048 shots each. Each time point therefore uses shots in total. Over
, the root-mean-square error of the mitigated observable deviates from the exact solution by 0.053 in and 0.068 in . (Without zero-noise extrapolation, these errors increase to 0.077 and 0.098.)
To also quantify the Trotterization error, we compare a noiseless implementation of the quantum circuit on classical hardware to the exact solution, obtained by directly exponentiating the full Hamiltonian. We find that after Trotter steps, the largest Trotterization error on the survival probability across the two modes is 7.3 × , 0.043, and 0.33 for , respectively. Reducing the Trotter step size by using and results in reduced relative errors of 2.1 × , 7.4 × , and 7.1 × .
We finally emphasize that the main advantage of the algorithm presented here is not necessarily an extremely low number of two-qubit gates. In this regard, it is still a significant improvement compared to the algorithms presented in [38], and requires a similar number of two-qubit gates as a naive implementation where each neutrino is represented by one qubit. The main strength of our new algorithm is that it requires only a small number of qubits to simulate a large number of neutrinos. It is therefore best suited for QPU architectures with low noise, but only few qubits. In that sense, deployment on IBM’s transmon QPUs, which have the opposite characteristic, was not optimal.
Generalizations. We briefly mentioned above that the algorithm introduced here is generalizable to systems with more than two modes. To achieve this, it is no longer sufficient to apply the transform, which reduces the interaction between the two modes to an operation on a single register, only once at the beginning of the circuit. Rather, each interaction between two modes needs to be wrapped between and operations that combine the registers of those two modes. For a system with modes, this implies that each Trotter step contains pairs of and transformations. The full circuit would correspond to the gate sequence in Figure 2 being applied between all register pairs in every Trotter step.
Results: Emergence of Quantum Kinetic Behavior in the Large- Limit
We are now ready to return to Figure 1 and compare the time evolution of the full multiparticle system to the predictions of the quantum kinetic (mean field) approach. In the figure, the solid black line corresponds to the mean-field prediction obtained by solving Eq. (1). Lines colored in different shades of purple were computed using multi-particle dynamics on classical hardware, along the lines of § 1.2.1.2. The results of the quantum circuits from § 1.2.1.3, finally, are plotted in blue, with data points corresponding to runs on the IBM Kingston quantum processor, and the blue line showing results from the same circuit on a noiseless simulator on classical hardware.
As expected from our analytical results in § 1.2.0.1, we find that the mean-field and multiparticle approaches agree in the large- limit, while showing significant differences for small systems. For the neutrino–neutrino toy system shown in the top panel of Figure 1, we see that is already sufficient to achieve excellent agreement. For the neutrino–antineutrino system in the middle panel, on the other hand, the same level of convergence requires . The main reason for this is the different choice of the self-interaction strength, parameterized here as . We can understand this from the analytical results in § 1.2.0.1: without vacuum oscillations, the states and from Eqs. (3) and (4) coincide with flavour eigenstates. All the flavour-changing terms in Eq. (17) that are coherently enhanced (the second term in the first line and all the terms on the second line) then come with small coefficient, suppressed by factors of the form and/or . This makes it easier for the last, non-coherent term to compete. Only after vacuum oscillations have driven and away from the flavour eigenbasis, coherent enhancement comes into play.
Note that, besides the smaller interaction strength, we have also chosen for the – system in the top panel of Figure 1. Without this small asymmetry, we would have found essentially vacuum oscillations. This can be understood in the flavour polarization vector formalism (see § 1.2): the and initial state correspond to polarization vectors of equal length and pointing in opposite directions, so the net polarization would be zero. If the two polarization vectors precess around at the same speed (i.e., ), this will never change. Only for , they can become imbalanced, leading to an enhanced flavour transition amplitude.
Comparing results derived on classical hardware to those obtained on quantum hardware, we see that the noisy quantum computation faithfully reproduces the classical result for about 10 Trotter steps before noise become visible. Results remain meaningful for up to 20 Trotter steps, beyond which the simulation becomes entirely noise-dominated. We have checked that when simulating a larger system ( instead of ), the number of meaningful Trotter steps reduces to .
The take-home message from Figure 1 is that full multiparticle simulations of collectively oscillating neutrino systems reproduce the predictions of the quantum-kinetic mean-field approach in the large- limit. We have explicitly demonstrated this for simple toy systems, and in the process have taken significant steps forward in developing new algorithms for large multiparticle systems on both classical and quantum hardware. It stands to reason that the conclusions derived by our two-mode toy systems generalize to more complex systems with many modes, as long as there are all-to-all interactions. We will investigate in the next section how our conclusions change for systems in which interactions are not all-to-all.
Phenomenologically Relevant Entanglement
Do the arguments given above show conclusively that quantum entanglement is irrelevant to collective neutrino oscillations in realistic astrophysical and cosmological environments? We now argue that they do not. We recall that the main arguments against phenomenologically relevant quantum entanglement are:
In a realistic environment, neutrinos are streaming in different directions, therefore each pair of neutrinos sees each other at most once. This prevents the build-up of significant entanglement.
Even for toy systems with persistent all-to-all interactions, we have seen that, in the large limit, a quantum kinetic description is sufficient since only processes that do not generate entanglement get coherently enhanced.
Neutrino Streams
To see entanglement, we are therefore looking for a system in which only a small number of neutrinos get entangled, and in which these neutrinos interact with each other for an extended period of time. Consider an ensemble of neutrinos traveling in a particular direction. If the neutrino density is extremely high, and the size of each neutrino wave packet is larger than the separation between neutrinos, interactions become possible. In particular, each neutrino will interact with its nearest neighbours (that is, the neutrinos directly ahead of it and directly behind it). Longer-range interactions will be suppressed exponentially for Gaussian wave packets or as a power law for Lorentzian wave packets. The scenario is schematically depicted in Figure 3. As all neutrinos travel at the same speed, these interactions are active for an extended period of time. They only cease when minor kinematic differences eventually drive the neutrinos apart.
Figure 3. Setup for a quantum simulation of a stream of neutrinos leaving a supernova, with partially overlapping wave packets. Each qubit , , . . . (black dots) encodes the flavour state of one neutrino. Neutrinos have sizeable interactions with their nearest neighbours and weaker interactions with their next-to-nearest and next-to-next-to-nearest neighbours (blue arcs).
To gauge whether this scenario can be realized in nature, for instance in neutrinos streaming away from a supernova or a neutron star merger, we need to examine the width of a typical supernova neutrino wave packet, which has been estimated to be of order cm [64] (see also ref. [65] for a very similar estimate). This is comparable to the mean distance between neutrinos at typical densities around the supernova neutrinosphere, where implies a mean separation of order cm. However, neighbouring neutrinos will not typically stream in the same direction. Therefore, according to these estimates, the time during which the interactions depicted in Figure 3 are active will be extremely short, leaving insufficient time to build up entanglement.
This conclusion can be avoided if neutrino wave packets are significantly larger than estimated in Refs. [64, 65]. We speculate here that this could indeed be the case thanks to the presence of strong magnetic fields in supernova cores and neutron star mergers. In a magnetic field, charged particles experience Landau quantization, so it stands to reason that nucleons and charged leptons are delocalized across their Landau levels. The spatial extent of a Landau level is comparable to the cyclotron radius, which for a particle of charge and momentum in a magnetic field of strength is given by . Neutrinos produced from such a system will inherit this delocalization. At the aforementioned densities of order , each neutrino wave packet of size will then overlap with other neutrinos. The likelihood that some of these neutrinos have momentum vectors sufficiently close to the one of the test neutrino to retain wave function overlap over macroscopic distances is now of order one. There is then sufficient time to build up phenomenologically relevant quantum entanglement among neutrinos, and the one-dimensional neutrino stream from Figure 3 can serve as a good proxy to understand the physical consequences.
These considerations motivate us to push ahead with the simulation of one-dimensional neutrino streams and to set up a quantum simulation of the system shown in Figure 3. Each black dot in the diagram now represents a qubit encoding the flavour of one neutrino. The interactions between qubits have the same structure as in Eq. (31), but the coupling constants are chosen such that the -th qubit has the strongest interactions with qubits , (representing the neutrinos just ahead of it and just behind it), and progressively weaker interactions with more distant qubits. Specifically, we choose the interaction strength between qubits and as
where imposes a cutoff on interactions between qubits that are too far apart. This cutoff is introduced to reduce the number of two-qubit gates required in a quantum simulation.
In Figure 4, we compare the average survival probability of electron and muon neutrinos in a neutrino chain over time, calculated using the quantum kinetic mean-field approach, a noiseless quantum simulation on classical hardware, and a noisy quantum simulation on an IBM Heron r3 QPU, specifically the IBM Aachen device. For the quantum simulations, the system’s time evolution is described by the Hamiltonian in Eqs. (30) and (31), where we set for all neutrinos, and , with . In panel (a) of Figure 4, we have chosen (interactions up to the next-to-nearest neighbour) to keep noise manageable; in panel (b), we have used (interactions up to the next-to-next-to-nearest neighbour). We choose an initial state in which the sixth, seventh and eighth neutrino along the chain are muon neutrinos and the rest are electron neutrinos. In the quantum kinetic approach (dashed curves in Figure 4), we solve Eq. (1), neglecting the spatial transport term , and replacing the integral by a sum , with the coupling constants given by Eq. (60). For the quantum simulations, we use the approach described in § 1.2.1.1, which consists of evolving the Hamiltonian in Eqs. (30) and (31) in time using the second-order Suzuki–Trotter decomposition in Eq. (33). In the noiseless simulation, we use shot-based sampling with 1 024 shots and the AerSimulator() class implemented in Qiskit [66] with the matrix product state (MPS) method.7 For the noisy simulation on superconducting quantum hardware, we computed our results using the Estimator primitive and used a Heron r3 IBM QPU with error mitigation through Pauli twirling [60], dynamical decoupling [63] with an XpXm pulse sequence, zero-noise extrapolation [61, 62] and twirled readout error extinction [54].

Figure 4. Average survival probability of electron (orange) and muon (purple) neutrinos in a one-dimensional neutrino stream (cf. Figure 3) as a function of time. In panel (a) we consider an 8-neutrino stream and compare the solution from the quantum kinetic mean-field approach, where in the self-interaction term in Eq. (1), the integral is replaced by a weighted sum over just the nearest and next-to-nearest neighbours (dashed curves), a noiseless simulation on classical hardware (solid curves), and a noisy simulation on an IBM Heron r3 QPU, specifically the IBM Aachen device (data points with error bars). The three rightmost neutrinos in the chain are initially , the rest are initialized as . Panel (b) corresponds to a 30-neutrino system with interactions up to next-to-next-to-nearest neighbours, for which we only show the mean-field and noiseless simulations as simulations on real hardware become noise-dominated too quickly. Again, three neutrinos are initially taken in the state, the rest in the state, as shown in the diagram inset into the plot. We see that, even with a larger number of neutrinos, the quantum simulation retains features which are lost in the quantum kinetic approach. Panel (c) shows the entanglement entropy as a function of time for the third neutrino in the stream in the noiseless simulation, proving the build-up of significant entanglement. Panel (d) shows the maximal Lyapunov exponent as a function of time for the quantum kinetic solution, where a positive value points towards the emergence of chaotic behaviour, which we indeed observe in panel (b).
The accuracy of the simulation on the QPU is limited for this system due to the large number of noisy two-qubit operations necessary to account for interactions between neutrinos up to the next-to-nearest neighbour. We can decrease the number of two-qubit gates by limiting the number of qubits and by increasing the size of the Trotter steps, albeit the latter increases the Trotterization error. In panel (a) of Figure 4, we have balanced the number of two-qubit gates against the size of the Trotter steps such that the noisy simulation resolves the first oscillation period. At later times, though, the simulation reaches two-qubit gates, causing the system to become noise-dominated and all oscillation probabilities to converge towards 0.5.
As the number of neutrinos increases, such as in panel (b) of Figure 4 for a 30-neutrino system, we find that the quantum simulation (which in this case is only feasible using the MPS simulator on classical hardware) still retains features which are lost in the quantum kinetic approach. This shows that the quantum kinetic approach is indeed not able to model collective neutrino oscillations in a neutrino stream, providing evidence for the arguments motivating quantum simulations above. To further corroborate this conclusion, we compute the entanglement entropy of the system over time. The entanglement entropy for a single neutrino is given by
where are the eigenvalues of the reduced density matrix of the neutrino under consideration, and where the degrees of freedom of all other neutrinos have been traced out. The result, shown in Figure 4 (c) for the third neutrino in the chain, demonstrates that entanglement does indeed build up. We have checked that the entanglement entropy behaves similarly for the other neutrinos except those at the edges.
Another notable point is the possible emergence of chaotic behaviour in the solution of the quantum kinetic equations. Indeed, we see in the right part of Figure 4 (b) that, after , the solution becomes irregular. We have checked that the solution does not stabilize as we vary the tolerances and step sizes of the ODE solver. To see whether the behaviour is indeed chaotic, we have calculated the maximal Lyapunov exponent, , using a two-trajectory Benettin algorithm and found that converges to a positive value, which points towards a chaotic solution [67], see Figure 4 (d). The algorithm consists of constructing a physical perturbation, in this case a small rotation of the density matrix of one of the neutrinos in the stream, and measuring the separation between the perturbed and unperturbed (reference) solutions. The reference and perturbed trajectories are evolved independently and the separation is accumulated over time steps until a chosen end time, , by which the Lyapunov exponent should have converged towards a final value. The maximal Lyapunov exponent is thus given by
where a positive means that initially nearby trajectories diverge exponentially, the hallmark of a chaotic solution.
“Once-in-a-Lifetime Encounter” Argument
We have argued that phenomenologically relevant entanglement may be realized in the special case of a neutrino stream being radiated in a particular direction. However, in the case where neutrinos are streaming in different directions, they will not interact with each other more than once, and therefore may not become strongly entangled. We now show the limitations of this “once-in-a-lifetime encounter” argument.
We consider first a toy system of three neutrinos (, , ) as depicted in the inset of Figure 5, where is initially a muon neutrino and the rest are electron neutrinos, and each pair interacts only once. The system’s density matrix is evolved in time by solving its von Neumann equation with the Hamiltonian (cf. Eqs. (29) to (31))

Figure 5. Entanglement entropy as a function of time for a toy system of three neutrinos, where each pair of neutrinos only interacts once, at times . The entanglement build-up increases with the coupling strength . Consequently, in the realistic case of small couplings, entanglement effects are negligible.
where
The Gaussian in the interaction term in Eq. (63) accounts for the limited duration of each encounter by restricting each interaction to a time window of width centered at a time . For the numerical evaluation, we choose the parameters , , and . After tracing out and , we calculate the entanglement entropy of , , where are the eigenvalues of the reduced density matrix of . We calculate the entanglement entropies of and analogously.
Our findings are summarized in Figure 5. For an initial state consisting of two electron neutrinos and one muon neutrino, we find that the entanglement entropy of all three neutrinos increases over time. This means that a build-up of entanglement is possible. However, as we decrease the coupling strength , the build-up becomes much less prominent. Therefore, for realistic couplings between neutrinos in a supernova environment, we do not expect multiple once-in-a-lifetime encounters to lead to phenomenologically relevant degrees of entanglement.
We also note that interactions are sparser in a three-dimensional system than in our planar toy system, since in three dimensions, most pairs of neutrino trajectories never cross.
As a generalization of this discussion, we increase the size of the system to up to ten neutrinos, leaving the number of pairwise interactions as a free parameter, and choosing the participating neutrino pairs at random with the timing of the interactions evenly spaced between and . We forbid repeated interactions of the same neutrino pair. In Figure 6, we show the average entanglement entropy over all neutrinos in the system after a given number of pairwise interactions has been completed. Unsurprisingly, a larger number of interactions leads to a larger entanglement entropy. Large entanglement entropies are easier to reach in small systems than in large ones because, for a fixed total number of interactions, each neutrino experiences more interactions in a small system than in a large one. Extrapolating these trends towards supernova conditions, i.e., a system with a large number of neutrinos that allows for a large number of pairwise interactions, we find that the build-up of entanglement from the many total interactions would not counterbalance the small entanglement entropy of each individual neutrino. In other words, we expect the larger number of neutrinos in a supernova to strongly hinder the build-up of entanglement.

Figure 6. Average entanglement entropy over all neutrinos in an system with once-in-a-lifetime interactions as a function of the system size (number of neutrinos) and the number of pairwise interactions. We set and choose an initial state consisting of half electron and half muon neutrinos. We see that few neutrinos or many pairwise interactions lead to a larger entanglement entropy. The grey curves represent different fractions of the maximum number of pairwise interactions.
Conclusions
To recapitulate, we have investigated the possible phenomenological impact of quantum entanglement among the flavours of different neutrinos in a dense neutrino gas, with supernovae, neutron star mergers, and the early Universe in mind. Our results can be summarized as follows.
In -neutrino toy systems with all-to-all interactions, the dynamics converge to the mean-field limit at large . We have demonstrated this explicitly for a system consisting initially of similar numbers of and , as well as for a system consisting of equal numbers of and . To corroborate our analytic results, we have carried out multiparticle simulations with up to neutrinos on classical hardware and up to 6 neutrinos on a superconducting quantum processor. Our results are summarized in Figure 1. In both cases, we have developed highly optimized algorithms that allow us to go to large and, in the case of the simulation on noisy hardware, simulate more than 10 Trotter steps before becoming noise-dominated.
Quantum entanglement can become phenomenologically relevant in systems where interactions have a short range (i.e. no all-to-all connectivity), but still a long duration. As an example, we have analyzed a linear stream of neutrinos using again both classical and quantum hardware. Our results, presented in Figure 4, clearly reveal discrepancies between the full multiparticle treatment and the computationally cheaper mean-field approximation. We have contemplated the possibility that such a system may be realized in environments with large magnetic fields if the neutrinos’ parent particles are delocalized over Landau levels.
If neutrinos only experience short-duration interactions, with each pair meeting at most once (“once-in-a-lifetime encounters”), entanglement is unlikely to build up.
Our results show that answering the question of whether multiparticle dynamics is phenomenologically relevant in astrophysical systems, and therefore whether quantum computing may be needed to understand these systems, requires a more detailed understanding of the relevant neutrino production processes and, in particular, of the wave packet properties of the resulting neutrino states.
Acknowledgments
It is a pleasure to thank Ting Cheng, Basudeb Dasgupta, Enrique Rico Ortega, and Cristina Volpe for insightful and illuminating discussions. We are moreover grateful to CERN, where crucial parts of this work were carried out, for hospitality and for access to IBM QPUs through the CERN Quantum Technology Initiative. We moreover acknowledge use of the Fir cluster at Simon Fraser University as part of the Digital Research Alliance of Canada. KB and JK have received support from the Cluster of Excellence “Precision Physics, Fundamental Interactions and Structure of Matter”, PRISMA++ (EXC 2118/2, Project ID 390831469) funded by the German Research Foundation (DFG). UR would like to acknowledge support from the Department of Atomic Energy, Government of India (Project Identification Number RTI 4002) and from the J. C. Bose Grant of the Anusandhan National Research Foundation (ANRF), Government of India (ANRF/JBG/2025/000265/PS).
Hamiltonian Matrix Elements of the Bipolar System
In this appendix, we derive the matrix elements of the Hamiltonian describing a bipolar system, Eqs. (44) and (45). Our starting point is Eq. (36), in which we drop the self-interaction term as before, so that the Hamiltonian becomes
In a slight generalization of the discussion in § 1.2.1.2, we allow and to differ in length and orientation and write . The following results are therefore applicable to both systems shown in Figure 1: the neutrino-neutrino system as well as the neutrino–antineutrino system. After transforming to the primed basis from § 1.2.1.2 in which the -axis is aligned with and defining
Eq. (A1) becomes
with
Note that for the special case relevant for bipolar oscillations.
We will work in a basis of and eigenstates, , where , , , and are eigenvalues of , , , and , respectively. Here and in the remainder of this appendix, we denote the eigenvalues under simply as (without a prime) to make the notation less cumbersome. It is, however, understood that the quantization axis is always the primed one. We will moreover use the shorthand notation . We will see that is diagonal in this basis, while leaves only invariant but can change . We discuss first.
Interaction Term
We can rewrite from Eq. (68) as
which is evidently diagonal in the basis. The diagonal matrix elements are
Vacuum Term – Diagonal Part
Looking at Eq. (68), it is evident that the first term in , namely , is diagonal in the basis, with matrix elements
The second term, , on the other hand, has both diagonal and off-diagonal entries. To evaluate the matrix elements of this term, we first apply the Wigner–Eckart theorem, which states that [68, 69]
thereby defining the reduced matrix elements . The fact that the Wigner–Eckart theorem can be applied to the components of follows from the fact that transforms as a vector operator, which can be formally shown by demonstrating that [69]. We see from Eq. (72) that the matrix elements are non-zero only when and . This shows that decouples sectors with different quantum number, and has a tridiagonal structure within each sector, changing by at most one unit.
The contribution to the diagonal elements of the Hamiltonian follows from a corollary of the Wigner–Eckart theorem, the projection theorem, which asserts that
Using , Eq. (73) leads to
Note that in the specific case , this contribution vanishes.
Vacuum Term – Off-Diagonal Part
Concerning the off-diagonal term , we first note that, since , it follows that . Therefore,
To evaluate this further, we decompose and into their component spins using the Clebsch–Gordan coefficients:
We now employ the recursion relation for the -symbols [70],
Using the symmetry properties of the -symbols to permute the columns, setting , , , and , and isolating the term containing an prefactor, we obtain
We multiply this expression by , sum over , , and use the orthogonality properties of the -symbols:
In the last term on the right-hand side, we can relabel and and use the symmetry properties of the -symbols again to see that this term is identical, up to a sign, to the left-hand side and can therefore be combined with the latter. Our final result is then
Summary
Putting together the results derived in this section, in particular Eqs. (70), (71), (74) and (80), the Hamiltonian matrix elements take the form
For the bipolar system with equal number of neutrinos and antineutrinos ( and , and therefore ), these expressions reduce to Eqs. (44) and (45).
Adaptive Classical Algorithm
The algorithm in § 1.2.1.2 achieves high performance by truncating the initial state to the most relevant sectors (keeping in mind that is a constant of motion), which introduces a constant error. Additional truncations are possible during time evolution at the cost of introducing additional errors in each step of the computation. This section describes the algorithm in more detail.
Within each block, the algorithm operates on a set of states that form a contiguous chain in , , which is a subset of the full chain. At , and in each block are chosen in the following way: the algorithm ranks the overlap of the initial state with each state. The state with the largest overlap is selected first, followed by the one with the next-to-largest overlap, and so on. This is repeated until the total amplitude of the left-out states is at most at a configured tolerance . Then, within each block, the retained states are closed into one contiguous interval by choosing and as the lowest and highest retained values, respectively, and keeping all states in between.
A runtime error estimation is used to dynamically expand the active window within each of the retained blocks. The windows are only ever enlarged and never shrunk. Since is tridiagonal in within each block, amplitude can leave an active window only through its two edges, via the off-diagonal matrix elements. The error is therefore monitored via the amplitude flux through the two edges. Let denote the truncated state that the algorithm actually propagates, with its basis being the states within the active window. Denote the amplitudes of a particular basis state as . During a time step , the amplitude transferred out of the upper edge of the active window to the first excluded state is approximately
where the matrix element is given by Eq. (45). The lower edge is treated in the same way. Before each time step, any window whose amplitude flux across the edge exceeds a preset threshold ( for the results in this paper) is extended on that side. The added states are initialized with zero amplitudes. Note that the window expansion threshold is not an error tolerance setting, as global error budget applies and will be explained below.
For a full error estimation, let be the projector onto the active window of all retained blocks, and . We compare the exact state with the truncated state . The latter starts from the normalized projection of the initial state onto the retained sectors and
windows and is evolved under . For the unitary time evolution, over any interval over which is fixed,
The error due to leakage of amplitude outside the active window, , is supported on the two omitted states adjacent to each window. Therefore,
The solver can then accumulates the leakage error by an endpoint-trapezoid estimation. In this work, for to , the error remained below for both toy scenarios. In the two largest runs we carried out, we obtained an accumulated error of for the – system at with an expansion threshold of , and for the neutrino–antineutrino system at with .
References
References
- [1]N. F. Bell, A. A. Rawlinson, and R. F. Sawyer, Speedup through Entanglement: Many Body Effects in Neutrino Processes, Phys. Lett. B 573, 86 (2003), arXiv:hep-ph/0304082.DOI
- [2]M. J. Cervia, A. V. Patwardhan, A. B. Balantekin, ‡. S. N. Coppermith, and C. W. Johnson, Entanglement and collective flavor oscillations in a dense neutrino gas, Phys. Rev. D 100, 083001 (2019), arXiv:1908.03511 [hep-ph].DOI
- [3]E. Rrapaj, Exact solution for multiangle quantum many-body collective neutrino-flavor oscillations, Phys. Rev. C 101, 065801 (2020), arXiv:1905.13335 [hep-ph].
- [4]A. Roggero, Entanglement and many-body effects in collective neutrino oscillations, Phys. Rev. D 104, 103016 (2021), arXiv:2102.10188 [hep-ph].DOI
- [5]A. V. Patwardhan, M. J. Cervia, and A. B. Balantekin, Spectral splits and entanglement entropy in collective neutrino oscillations, Phys. Rev. D 104, 123035 (2021), arXiv:2109.08995 [hep-ph].DOI
- [6]Z. Xiong, Many-body effects of collective neutrino oscillations, Phys. Rev. D 105, 103002 (2022), arXiv:2111.00437 [astro-ph.HE].DOI
- [7]A. Roggero, E. Rrapaj, and Z. Xiong, Entanglement and correlations in fast collective neutrino flavor oscillations, Phys. Rev. D 106, 043022 (2022), arXiv:2203.02783 [astro-ph.HE].DOI
- [8]J. D. Martin, A. Roggero, H. Duan, and J. Carlson, Many-body neutrino flavor entanglement in a simple dynamic model, (2023), arXiv:2301.07049 [hep-ph].arxiv.org/abs/2301.07049
- [9]S. P. Mikheyev and A. Y. Smirnov, Resonance Amplification of Oscillations in Matter and Spectroscopy of Solar Neutrinos, Sov. J. Nucl. Phys. 42, 913 (1985).
- [10]L. Wolfenstein, Neutrino Oscillations in Matter, Phys.Rev. D17, 2369 (1978).
- [11]J. T. Pantaleone, Neutrino oscillations at high densities, Phys. Lett. B 287, 128 (1992).DOI
- [12]J. T. Pantaleone, Dirac neutrinos in dense matter, Phys. Rev. D 46, 510 (1992).pubmed.ncbi.nlm.nih.gov/10014964
- [13]Y. Z. Qian and G. M. Fuller, Neutrino-neutrino scattering and matter enhanced neutrino flavor transformation in Supernovae, Phys. Rev. D 51, 1479 (1995), arXiv:astro-ph/9406073.DOI
- [14]S. Samuel, Bimodal coherence in dense selfinteracting neutrino gases, Phys. Rev. D 53, 5382 (1996), arXiv:hep-ph/9603414.
- [15]S. Pastor and G. Raffelt, Flavor oscillations in the supernova hot bubble region: Nonlinear effects of neutrino background, Phys. Rev. Lett. 89, 191101 (2002), arXiv:astro-ph/0207281.DOI
- [16]A. B. Balantekin and H. Yuksel, Neutrino mixing and nucleosynthesis in core-collapse supernovae, New J. Phys. 7, 51 (2005), arXiv:astro-ph/0411159.DOI
- [17]H. Duan, G. M. Fuller, and Y.-Z. Qian, Collective Neutrino Oscillations, Ann. Rev. Nucl. Part. Sci. 60, 569 (2010), arXiv:1001.2799 [hep-ph].DOI
- [18]A. Mirizzi, I. Tamborra, H.-T. Janka, N. Saviano, K. Scholberg, R. Bollig, L. Hudepohl, and S. Chakraborty, Super- nova Neutrinos: Production Oscillations and Detection, Riv. Nuovo Cim. 39, 1 (2016), arXiv:1508.00785 [astro-ph.HE].
- [19]M. Sen, Supernova Neutrinos: Flavour Conversion Mechanisms and New Physics Scenarios, Universe 10, 238 (2024), arXiv:2405.20432 [hep-ph].DOI
- [20]L. Johns, S. Richers, and M.-R. Wu, Neutrino Oscillations in Core-Collapse Supernovae and Neutron Star Mergers, Ann. Rev. Nucl. Part. Sci. 75, 399 (2025), arXiv:2503.05959 [astro-ph.HE].DOI
- [21]Y.-Z. Qian and G. M. Fuller, Matter enhanced antineutrino flavor transformation and supernova nucleosynthesis, Phys. Rev. D 52, 656 (1995), arXiv:astro-ph/9502080.DOI
- [22]Y.-Z. Qian, G. M. Fuller, G. J. Mathews, R. Mayle, J. R. Wilson, and S. E. Woosley, A Connection between flavor mixing of cosmologically significant neutrinos and heavy element nucleosynthesis in supernovae, Phys. Rev. Lett. 71, 1965 (1993).pubmed.ncbi.nlm.nih.gov/10054549
- [23]J. Pruet, R. D. Hoffman, S. E. Woosley, H. T. Janka, and R. B. Buras, Nucleosynthesis in early supernova winds. 2. the role of neutrinos, Astrophys. J. 644, 1028 (2006), arXiv:astro-ph/0511194.DOI
- [24]A. Arcones, H.-T. Janka, and L. Scheck, Nucleosynthesis-relevant conditions in neutrino-driven supernova outflows. 1. Spherically symmetric hydrodynamic simulations, Astron. Astrophys. 467, 1227 (2007), arXiv:astro-ph/0612582.
- [25]G. M. Fuller and B. S. Meyer, Neutrino capture and supernova nucleosynthesis, Astrophys. J. 453, 792 (1995).
- [26]G. Sigl and G. Raffelt, General kinetic description of relativistic mixed neutrinos, Nucl. Phys. B 406, 423 (1993).
- [27]B. Hall, A. Roggero, A. Baroni, and J. Carlson, Simulation of collective neutrino oscillations on a quantum computer, Phys. Rev. D 104, 063009 (2021), arXiv:2102.12556 [quant-ph].DOI
- [28]K. Yeter-Aydeniz, S. Bangar, G. Siopsis, and R. C. Pooser, Collective neutrino oscillations on a quantum computer, Quant. Inf. Proc. 21, 84 (2022), arXiv:2104.03273 [quant-ph].
- [29]A. K. Jha and A. Chatla, Quantum studies of neutrinos on IBMQ processors, Eur. Phys. J. ST 231, 141 (2022).DOI
- [30]V. Amitrano, A. Roggero, P. Luchi, F. Turro, L. Vespucchi, and F. Pederiva, Trapped-ion quantum simulation of collective neutrino oscillations, Phys. Rev. D 107, 023007 (2023), arXiv:2207.03189 [quant-ph].DOI
- [31]A. B. Balantekin, M. J. Cervia, A. V. Patwardhan, E. Rrapaj, and P. Siwach, Quantum information and quantum simulation of neutrino physics, Eur. Phys. J. A 59, 186 (2023), arXiv:2305.01150 [nucl-th].DOI
- [32]P. Siwach, K. Harrison, and A. B. Balantekin, Collective neutrino oscillations on a quantum computer with hybrid quantum-classical algorithm, Phys. Rev. D 108, 083039 (2023), arXiv:2308.09123 [quant-ph].arxiv.org/abs/2308.09123
- [33]F. Turro, I. A. Chernyشev, R. Bhaskar, and M. Illa, Qutrit and qubit circuits for three-flavor collective neutrino oscillations, Phys. Rev. D 111, 043038 (2025), arXiv:2407.13914 [quant-ph].DOI
- [34]I. Chernyshев, C. E. P. Robin, and M. J. Savage, Quantum magic and computational complexity in the neutrino sector, Phys. Rev. Res. 7, 023228 (2025), arXiv:2411.04203 [quant-ph].arxiv.org/abs/2411.04203
- [35]G. Singh, Arvind, and K. Dorai, Simulating three-flavor neutrino oscillations on a nuclear magnetic resonance quantum processor, Phys. Scripta 100, 085106 (2025), arXiv:2412.15617 [quant-ph].
- [36]L. Spagnoli et al., Collective neutrino oscillations in three flavors on qubit and qutrit processors, Phys. Rev. D 111, 103054 (2025), arXiv:2503.00607 [quant-ph].DOI
- [37]S. Tripathi, S. Joshi, G. Rajpoot, and P. Shukla, Quantum Simulation of Collective Neutrino Oscillations in Dense Neutrino Environment, (2025), arXiv:2508.11610 [quant-ph].arxiv.org/abs/2508.11610
- [38]K. Bleau, N. Ilic, J. Kopp, U. Rahaman, and X. Y. Yu, Quantum Simulation of Collective Neutrino Oscillations using Dicke States, (2026), arXiv:2604.07452 [quant-ph].arxiv.org/abs/2604.07452
- [39]D. J. Heimsoтh, P. Siwach, and A. B. Balantekin, Three-flavor supernova neutrino simulation using a hybrid quantum-classical algorithm with qutrits, Phys. Rev. D 114, 043055 (2026), arXiv:2605.01099 [hep-ph].
- [40]A. Friedland and C. Lunardini, Neutrino flavor conversion in a neutrino background: Single particle versus multi-particle description, Phys. Rev. D 68, 013007 (2003), arXiv:hep-ph/0304055.
- [41]A. Friedland and C. Lunardini, Do many particle neutrino interactions cause a novel coherent effect?, JHEP 10, 043, arXiv:hep-ph/0307140.
- [42]S. Shalgar and I. Tamborra, Do we have enough evidence to invalidate the mean-field approximation adopted to model collective neutrino oscillations?, Phys. Rev. D 107, 123004 (2023), arXiv:2304.13050 [astro-ph.HE].DOI
- [43]L. Johns, Neutrino many-body correlations, Int. J. Mod. Phys. A 39, 2450122 (2024), arXiv:2305.04916 [hep-ph].DOI
- [44]A. Kost, L. Johns, and H. Duan, Once-in-a-lifetime encounter models for neutrino media: From coherent oscillations to flavor equilibration, Phys. Rev. D 109, 103037 (2024), arXiv:2402.05022 [hep-ph].DOI
- [45]A. Kost, L. Johns, and H. Duan, Once-in-a-lifetime encounter models for neutrino media. II. Quasisteady states and msicdynamc flavor evolution, Phys. Rev. D 112, 103004 (2025), arXiv:2508.14331 [hep-ph].DOI
- [46]A. Friedland, Self-refraction of supernova neutrinos: mixed spectra and three-flavor instabilities, Phys. Rev. Lett. 104, 191102 (2010), arXiv:1001.0996 [hep-ph].DOI
- [47]B. Dasgupta, A. Mirizzi, I. Tamborra, and R. Tomas, Neutrino mass hierarchy and three-flavor spectral splits of supernova neutrinos, Phys. Rev. D 81, 093008 (2010), arXiv:1002.2943 [hep-ph].DOI
- [48]A. Esteban-Pretel, S. Pastor, R. Tomas, G. G. Raffelt, and G. Sigl, µ-τ neutrino refraction and collective three-flavor transformations in supernovae, Phys. Rev. D 77, 065024 (2008), arXiv:0712.1137 [astro-ph].
- [49]F. Capozzi, M. Chakraborty, S. Chakraborty, and M. Sen, Fast flavor conversions in supernovae: the rise of mu-tau neutrinos, Phys. Rev. Lett. 125, 251801 (2020), arXiv:2005.12404 [hep-ph].
- [50]F. Capozzi, M. Chakraborty, S. Chakraborty, and M. Sen, Supernova fast flavor conversions in 1+1D: Influence of mu-tau neutrinos, Phys. Rev. D 106, 083011 (2022), arXiv:2205.06272 [hep-ph].DOI
- [51]S. Hannestad, G. G. Raffelt, G. Sigl, and Y. Y. Y. Wong, Self-induced conversion in dense neutrino gases: Pendulum in flavour space, Phys. Rev. D 74, 105010 (2006), [Erratum: Phys.Rev.D 76, 029901 (2007)], arXiv:astro-ph/0608695.arxiv.org/abs/astro-ph/0608695
- [52]G. G. Raffelt and G. Sigl, Self-induced decoherence in dense neutrino gases, Phys. Rev. D 75, 083002 (2007), arXiv:hep-ph/0701182.
- [53]F. T. Arecchi, E. Courtens, R. Gilmore, and H. Thomas, Atomic coherent states in quantum optics, Phys. Rev. A 6, 2211 (1972).
- [54]E. van den Berg, Z. K. Minev, and K. Temme, Model-free readout-error mitigation for quantum expectation values, Phys. Rev. A 105, 032620 (2022).DOI
- [55]H. Tal-Elzer and R. Kosloff, An accurate and efficient scheme for propagating the time dependent Schrödinger equation, J. Chem. Phys. 81, 3967 (1984).
- [56]J. R. Stryker, Shearing approach to gauge-invariant Trotterization, Phys. Rev. D 112, 014508 (2025), arXiv:2105.11548 [hep-lat].
- [57]T. G. Draper, Addition on a Quantum Computer, (2000), arXiv:quant-ph/0008033.DOI
- [58]V. Bergholm, J. J. Vartiainen, M. Möttönen, and M. M. Salomaa, Quantum circuits with uniformly controlled one-qubit gates, Phys. Rev. A 71, 052330 (2005), arXiv:quant-ph/0410066.DOI
- [59]B. L. Douglas and J. B. Wang, Efficient quantum circuit implementation of quantum walks, Phys. Rev. A 79, 052335 (2009), arXiv:0706.0304 [quant-ph].arxiv.org/abs/0706.0304
- [60]J. J. Wallman and J. Emerson, Noise tailoring for scalable quantum computation via randomized compiling, Phys. Rev. A 94, 052325 (2016), arXiv:1512.01098 [quant-ph].DOI
- [61]K. Temme, S. Bravyi, and J. M. Gambetta, Error Mitigation for Short-Depth Quantum Circuits, Phys. Rev. Lett. 119, 180509 (2017), arXiv:1612.02058 [quant-ph].DOI
- [62]T. Giurgica-Tiron, Y. Hindy, R. LaRose, A. Mari, and W. J. Zeng, Digital zero noise extrapolation for quantum error mitigation, in 2020 IEEE International Conference on Quantum Computing and Engineering (QCE) (2020) pp. 306–316.
- [63]L. Viola, E. Knill, and S. Lloyd, Dynamical decoupling of open quantum systems, Phys. Rev. Lett. 82, 2417 (1999).DOI
- [64]E. Akhmedov, J. Kopp, and M. Lindner, Collective neutrino oscillations and neutrino wave packets, JCAP 09, 017, arXiv:1702.08338 [hep-ph].arxiv.org/abs/1702.08338
- [65]J. Kersten and A. Y. Smirnov, Decoherence and oscillations of supernova neutrinos, Eur. Phys. J. C 76, 339 (2016), arXiv:1512.09068 [hep-ph].DOI
- [66]A. Javadi-Abhari, M. Treinish, K. Trulish, C. J. Wood, J. Lishman, J. Gacon, S. Martiel, P. D. Nation, L. S. Bishop, A. W. Cross, B. R. Johnson, and J. M. Gambetta, Quantum computing with qiskit (2024), arXiv:2405.08810 [quant-ph].
- [67]R. S. Hansen and S. Hannestad, Chaotic flavor evolution in an interacting neutrino gas, Phys. Rev. D 90, 025009 (2014), arXiv:1404.3833 [hep-ph].DOI
- [68]D. A. Varshalovich, A. N. Moskalev, and V. K. Khersonskii, Quantum Theory of Angular Momentum: Irreducible Tensors, Spherical Harmonics, Vector Coupling Coefficients, 3nj Symbols (World Scientific Publishing Company, 1988).
- [69]J. J. Sakurai and J. J. Napolitano, Modern Quantum Mechanics, 3rd ed., Quantum physics, quantum information and quantum computation (Cambridge University Press, 2020).
- [70]DLMF, NIST Digital Library of Mathematical Functions, https://dlmf.nist.gov/, Release 1.2.8 of 2026-09-15, f. W. J. Olver, A. B. Olde Daalhuis, D. W. Lozier, B. I. Schneider, R. F. Boisvert, C. W. Clark, B. R. Miller, B. V. Saunders, H. S. Cohl, and M. A. McClain, eds.


