Foundations of quantum mechanics · Theory
Self-testing in Fermionic Networks
arXiv:2611.25074v1 ↗PDF ↗Submitted 18 November 2026
- Question
- Self-testing infers the state and measurements inside untrusted devices from their correlations alone, but the standard theory assumes that the devices hold qudits composed by tensor products. Fermionic modes are different. Information stored in the occupation of a mode, a febit, obeys anticommutation and a parity superselection rule, and in networks febits can produce correlations that no qudit strategy reproduces under the same locality constraints. Does a self-testing statement depend on whether the devices are modelled as qudits or as febits? And can some correlations self-test a genuinely fermionic strategy?
- Result
- Both answers are yes, in a three-party rewiring network. Each party prepares one fermion spread over two modes, keeps one mode and sends the other through a hidden wiring: one of two three-party loops or three two-party loops. On the fixed loop \(ABC\), the correlations self-test a unique qubit strategy. They admit exactly two fermionic realisations, however, that no local parity-preserving isometry relates: the single-mode strategy and its double-mode encoding \(\ket0\mapsto\ket{00}\), \(\ket1\mapsto\ket{11}\). The correlations of all five wirings together self-test the single-mode febit strategy up to local parity-preserving isometries, which is the first genuinely fermionic self-test. The deciding signal is an exchange sign. In a two-party loop the two exchanged fermions turn the conditional state into \(\tfrac1{\sqrt2}(\ket{\tilde0\tilde0}-\ket{\tilde1\tilde1})\), so outcomes of \(\sigma_X\) on both sides are anticorrelated. Any qubit-equivalent strategy makes them correlated.
The paper in detail
A febit is a fermionic mode, empty (\(\ket0\)) or occupied (\(\ket1\)). Two things set febits apart from qubits. First, creation operators anticommute, so the order of modes matters: \(f_2^\dagger f_1^\dagger\ket\Omega=-f_1^\dagger f_2^\dagger\ket\Omega\). Second, the parity superselection rule forbids superpositions of even and odd fermion number, and it restricts each party's local observables to even polynomials in its own creation and annihilation operators. One consequence is that a single febit per party cannot violate the CHSH inequality, because all its allowed observables are diagonal in the occupation basis. Two febits per party can encode a qubit in the even sector, \(\ket{00}\) and \(\ket{11}\), and the standard CHSH self-test carries over, with local isometries required to preserve parity. In centralised scenarios with a single source, the two theories therefore self-test the same strategies.
The scenario the paper uses comes from earlier work by the same group, which found fermion-only correlations in it. Each of \(A\), \(B\) and \(C\) prepares \(\tfrac1{\sqrt2}(f_L^\dagger+f_R^\dagger)\ket\Omega\), a single fermion shared between two modes. Each keeps mode \(R\) and sends mode \(L\) into a network that delivers it according to an undisclosed wiring. The wiring is one of the two directed three-party loops, or one of the three two-party loops in which the third party gets its own mode back. Each party then measures its two modes. The outcomes 0 and 2 count no particle or two particles, and in the one-particle sector the outcomes \(1_\pm\) resolve a basis. In the logical basis \(\ket{\tilde0}=\ket{01}\), \(\ket{\tilde1}=\ket{10}\), \(A\) measures \((\sigma_X\pm\sigma_Z)/\sqrt2\), and \(B\) and \(C\) measure \(\sigma_Z\) or \(\sigma_X\). A party's measurement may depend only on its input and on who sent it the mode, never on the wiring.
On a fixed three-party loop, the coarse outcome "one particle each" happens only when all sources keep their particle or all pass it on, so it heralds a GHZ state on the logical qubits. A cyclic exchange of three fermions is an even permutation, so this state carries no extra sign. Two known results combine to show that in qubit theory the loop's correlations single out one strategy: token counting fixes the sources and coarse measurements, and GHZ self-testing fixes the rest. In fermionic theory one ambiguity survives. The encoding \(\ket0\mapsto\ket{00}\), \(\ket1\mapsto\ket{11}\) reproduces the same correlations, but it changes the relative parity of the two branches, so it is not an allowed local isometry. The single-mode and double-mode strategies are therefore inequivalent realisations of the same data.
A two-party loop breaks the tie. There, the "both pass" branch exchanges two fermions, an odd permutation, and picks up a minus sign that the double-mode encoding cannot produce. Local phases in sources or measurements could try to imitate the sign, but each such phase also enters other wirings that use the same source or local context, and consistency across all five wirings rules this out. The open questions are whether the hidden rewiring is necessary, whether all parity-respecting pure fermionic states can be self-tested in networks, whether a noise-robust version exists, and whether the approach extends to anyons and paraparticles.
Transferable methods
Counting exchange signs around a network
The physical difference between febits and qubits in this scenario is the sign that appears when the parties' modes are put in a canonical order. Which branches survive the conditioning, and in which order their creation operators end up, is fixed by the wiring. A cyclic three-party exchange is an even permutation and gives no sign; a two-party exchange is odd and gives \(-1\). Writing the state with creation operators makes this explicit and shows why the three-party loops cannot distinguish the two realisations while the two-party loops can.
The global state is \(\ket\psi=2^{-3/2}(f^\dagger_{\alpha L}+f^\dagger_{\alpha R})(f^\dagger_{\beta L}+f^\dagger_{\beta R})(f^\dagger_{\gamma L}+f^\dagger_{\gamma R})\ket\Omega\). In the wiring \(BC\), party \(B\) holds the modes \(\beta R,\gamma L\), and \(C\) holds \(\gamma R,\beta L\). Conditioning on one particle at \(B\) and one at \(C\) keeps two terms. "Both keep" gives \(f^\dagger_{\beta R}f^\dagger_{\gamma R}\), already in the order (mode of \(B\), mode of \(C\)). "Both pass" gives \(f^\dagger_{\beta L}f^\dagger_{\gamma L}\), whose first operator belongs to \(C\) and second to \(B\). Reordering it costs a sign (editorial derivation): \[f^\dagger_{\beta L}f^\dagger_{\gamma L}=-f^\dagger_{\gamma L}f^\dagger_{\beta L},\qquad\text{so}\qquad\ket{\psi_{BC}}\propto\ket{\tilde1\tilde1}-\ket{\tilde0\tilde0}.\] A party's measurement operators are even, so \(A\)'s factor plays no role. With both parties measuring \(\sigma_X\), the outcomes are perfectly anticorrelated.
In the loop \(ABC\) the parties hold \((\alpha R,\gamma L)\), \((\beta R,\alpha L)\) and \((\gamma R,\beta L)\). The "all pass" term \(f^\dagger_{\alpha L}f^\dagger_{\beta L}f^\dagger_{\gamma L}\) has to be reordered into \(f^\dagger_{\gamma L}f^\dagger_{\alpha L}f^\dagger_{\beta L}\) (modes of \(A\), \(B\), \(C\)). That reordering is a cyclic permutation, made of two transpositions, so it costs no sign, and the conditional state is \(\tfrac1{\sqrt2}(\ket{\tilde0\tilde0\tilde0}+\ket{\tilde1\tilde1\tilde1})\). A double-mode encoding replaces every occupied mode by an occupied pair. Pairs commute, so the encoding gives \(+\) in every loop. It agrees with the fermionic data on the three-party loops and disagrees on the two-party loops. The editorial simulation found conditional probabilities 1/4 and 1/2 and correlations \(\pm0.707\) and \(-1\), as stated in the summary basis.
How to reuse it
Any certification protocol for devices built from indistinguishable particles, such as fermionic atoms in tweezers or electrons in coupled quantum dots, should ask which isometries count as equivalences. Exchange signs routed through a network are the resource that separates realisations which tensor-product theory identifies. The same bookkeeping, with phases \(e^{i\theta}\) per exchange instead of \(-1\), is the natural starting point for network tests of anyonic statistics.
Consistency across wirings excludes local phase fixes
A qubit-equivalent strategy might try to recover the missing sign with local phases. The locality assumptions forbid tailoring them to the wiring. A source is the same in every wiring, and a party's measurement may depend on who sent it the mode, but not on where its own mode went. Every phase therefore appears in several wirings at once. The paper's Theorem C.1 makes this rigorous with self-testing machinery. A phase-only toy model, given below as an editorial illustration rather than the paper's proof, already shows why the five wirings cannot be satisfied together.
In a qubit-equivalent model, give source \(\theta\) a relative phase \(\phi_\theta\) between its "keep" and "pass" branches. Give party \(i\) a measurement phase \(\mu_i(j)\) that depends on the sender \(j\) of the mode it holds. With \(\sigma_X\)-type measurements, the conditional correlation in a loop is the cosine of the total phase around it. The fermionic data require total phase 0 in the two three-party loops and \(\pi\) in the three two-party loops (editorial model): \[\begin{aligned}ABC:&\ \ \phi_\alpha+\phi_\beta+\phi_\gamma+\mu_A(C)+\mu_B(A)+\mu_C(B)\equiv0,\\ ACB:&\ \ \phi_\alpha+\phi_\beta+\phi_\gamma+\mu_A(B)+\mu_C(A)+\mu_B(C)\equiv0,\\ BC,\,AB,\,AC:&\ \ \phi_\beta+\phi_\gamma+\mu_B(C)+\mu_C(B)\equiv\pi,\ \ \text{and cyclically.}\end{aligned}\]
Adding the two three-party constraints gives \(2\sum_\theta\phi_\theta+\sum\mu\equiv0\). Adding the three two-party constraints gives the same left-hand side, since every source appears twice and every context \(\mu_i(j)\) once, but a right-hand side of \(3\pi\equiv\pi\). No choice of phases satisfies both. Fermions escape because the exchange signs contribute \(+1\) to each three-party loop and \(-1\) to each two-party loop, which supplies exactly the missing \(\pi\) three times. This parity obstruction is the same kind of argument as the Mermin–Peres square or a GHZ paradox, applied to exchange statistics rather than to measurement outcomes.
How to reuse it
Summing constraints around all wirings, or all contexts, to expose an inconsistency mod \(2\pi\) is a cheap first test of whether some class of models can reproduce network data. It applies to checking whether a claimed network self-test survives non-tensor-product adversaries, and to designing tests that distinguish bosonic, fermionic and anyonic exchange phases with untrusted devices.
Context & limitations
Why it matters. Device-independent claims in networks rest on self-testing proofs that silently restrict both devices and adversary to tensor-product quantum theory. The first theorem shows that this restriction carries weight. Correlations that certify a unique qubit strategy can come from inequivalent fermionic ones. That raises the question, posed but not answered here, of whether network protocols such as randomness certification in the triangle remain secure against an adversary who manipulates fermions. Centralised Bell scenarios are unaffected, since there both theories self-test the same strategies.
Limitations. The results assume exact, noise-free correlations. There is no robustness analysis yet, so the self-test cannot be tested experimentally as it stands; the authors suggest fermionic cold atoms. The proof relies on the hidden rewiring and on locality assumptions. Sources do not depend on the wiring, and a measurement depends only on the party's name, its input and the sender of the mode it received. Whether a fixed network suffices is open. Only febits are treated, not anyons or paraparticles. The security concern is a question, not an attack. The authors used AI assistance only to check proofs for gaps and to polish the language. The supplemental proofs of Theorems B.1, B.10 and C.1 were not checked line by line for this summary.