Fermionic Non-Gaussianity

• Fermionic non-Gaussianity is defined as the deviation of many-body fermionic states from the Gaussian (quasi-free) manifold characterized by two-point correlators and Wick factorization.
• Multiple quantifiers, including relative entropy, antiflatness, and Gaussian fidelity, are employed to measure non-Gaussian contributions and indicate resource costs in simulation and state preparation.
• The framework connects covariance structure, random-state typicality, and many-body phase behavior, offering actionable experimental protocols such as Bell-sampling for state certification.

Searching arXiv for papers on fermionic non-Gaussianity and closely related resource-theoretic and random-state results. Fermionic non-Gaussianity is the deviation of a fermionic many-body state from the manifold of fermionic Gaussian, or quasi-free, states. In the fermionic setting, Gaussian states are exactly those fully characterized by two-point correlators and obeying Wick factorization; they are the natural free objects for matchgate and fermionic linear-optics dynamics, and they admit polynomial classical descriptions through covariance matrices. Non-Gaussianity therefore serves simultaneously as a structural notion—failure of Wick factorization, spectral non-flatness of the covariance matrix, or distance from Gaussianization—and as a resource-theoretic notion associated with computational complexity beyond free fermions. Recent work has developed several inequivalent but closely related quantifiers, clarified their scaling in random states and many-body systems, and introduced measurement protocols and algorithmic monotones that make the concept experimentally and computationally actionable (Ares et al., 18 May 2026).

1. Gaussian manifold, covariance structure, and Gaussianization

A standard starting point is a system of fermionic modes or, equivalently via the Jordan–Wigner map, qubits equipped with Majorana operators. For complex fermions $c_i,c_i^\dagger$, one defines

$$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$

with $\{\gamma_i,\gamma_j\}=2\delta_{ij}$. The covariance matrix is the real antisymmetric matrix built from Majorana two-point functions, for example

$$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$

or, in an equivalent sign convention for pure states,

$$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$

Pure fermionic Gaussian states are characterized by covariance flatness, such as $M^TM=\mathds{1}$ or $\Gamma^2=-I_{2n}$, and all higher-order correlators reduce to Pfaffians of covariance submatrices by Wick’s theorem (Falcão et al., 30 Jan 2026, Sierant et al., 30 May 2025, Haug et al., 25 May 2026).

Several papers formulate Gaussianity in the Nambu basis. For a subsystem $A$ of size $\ell$, with number-conserving correlator $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$ and pairing matrix $$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$0, the $$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$1 covariance matrix can be written

$$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$2

The corresponding Gaussianized state is the maximum-entropy fermionic Gaussian state reproducing the same one- and two-point correlators,

$$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$3

with $$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$4. In the $$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$5-symmetric case, $$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$6, while in the general paired case one keeps both $$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$7 and $$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$8 terms in the quadratic exponent (Ares et al., 18 May 2026, Ares et al., 17 Mar 2026).

This covariance-centric formulation is the basis for both structure theorems and practical computation. Gaussian states are the efficiently simulable free set because their dynamics under quadratic Hamiltonians reduce to orthogonal transformations of the covariance matrix, and number measurements admit closed update rules within the same formalism (Dias et al., 2023). A plausible implication is that fermionic non-Gaussianity is best understood not as a single scalar quantity but as a family of diagnostics built from how and where covariance data fail to provide a complete description.

2. Quantifiers of fermionic non-Gaussianity

One major family of measures is distance-based. For a subsystem reduced state $$\gamma_{2j-1}=c_j^\dagger+c_j,\qquad \gamma_{2j}=i(c_j^\dagger-c_j),$$9, the relative entropy of non-Gaussianity is defined by

$\{\gamma_i,\gamma_j\}=2\delta_{ij}$0

and, using $\{\gamma_i,\gamma_j\}=2\delta_{ij}$1, one obtains

$\{\gamma_i,\gamma_j\}=2\delta_{ij}$2

The same paper emphasizes that $\{\gamma_i,\gamma_j\}=2\delta_{ij}$3 is the unique Gaussian state with the same correlation matrix, so the measure captures entropy generated by higher-order, non-Wick data beyond the Gaussian reconstruction (Ares et al., 18 May 2026). Related work gives a lower bound on this relative-entropy non-Gaussianity in terms of the Shannon entropy $\{\gamma_i,\gamma_j\}=2\delta_{ij}$4 of the total particle-number distribution, establishing a direct link to particle-number asymmetry (Ares et al., 17 Mar 2026).

A second family is covariance-spectral. The fermionic antiflatness (FAF) of a pure state is defined by

$\{\gamma_i,\gamma_j\}=2\delta_{ij}$5

or, in equivalent notation,

$\{\gamma_i,\gamma_j\}=2\delta_{ij}$6

It is faithful for pure states, invariant under Gaussian unitaries, and directly expressible through two-point Majorana correlators. For $\{\gamma_i,\gamma_j\}=2\delta_{ij}$7,

$\{\gamma_i,\gamma_j\}=2\delta_{ij}$8

This measure is repeatedly interpreted as quantifying deviation from covariance flatness, hence the term antiflatness, and as a witness of failure of Wick factorization and departure from classically tractable free-fermion descriptions (Falcão et al., 30 Jan 2026, Sierant et al., 30 May 2025, Haug et al., 25 May 2026).

A third line is algorithmic. The fermionic Gaussian fidelity

$\{\gamma_i,\gamma_j\}=2\delta_{ij}$9

and the fermionic Gaussian extent

$$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$0

quantify, respectively, maximal Gaussian overlap and the cost of decomposing a state into Gaussian components. In this framework, classical simulation costs for fermionic linear optics with non-Gaussian inputs scale polynomially in system size and linearly in $$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$1, up to accuracy factors (Dias et al., 2023).

Recent work also introduced computable covariance-derived monotones for pure states. The occupation-number entropy family,

$$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$2

is defined from the Williamson eigenvalues $$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$3, and the $$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$4 member equals the entropy of the Gaussian state with the same covariance matrix. That quantity is a strong monotone under pure-state Gaussian protocols and lower-bounds the number of SWAP gates needed for preparation (Tarabunga et al., 2 Jul 2026). Separately, Bell-sampling methods yield the bridge degree, defined from the largest occupied eigensector of the operator $$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$5; this bridge degree is non-increasing under post-selected Gaussian protocols and yields stronger exact-conversion no-go theorems than previously known monotones (Tarabunga, 3 Jun 2026).

A fourth construction uses fermionic convolution. For even states $$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$6, the balanced convolution

$$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$7

damps higher-order Grassmann cumulants while preserving second moments. Iterated self-convolution converges to the Gaussian state with the same covariance, and the associated non-Gaussian entropy $$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$8 is faithful, additive on pure even states, and converges under iteration to the relative entropy of fermionic non-Gaussianity (Lyu et al., 2024).

3. Typical behavior in random states

The large-system scaling of fermionic non-Gaussianity has been worked out analytically for Haar-random states. For a global pure state on $$G_{ij}=i\,\operatorname{Tr}(\rho\,\gamma_i\gamma_j)$$9 qubits and a subsystem of size $$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$0, the average relative-entropy non-Gaussianity obeys a sharp two-regime law. Without symmetries,

$$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$1

in the thermodynamic limit. The mechanism for the vanishing branch is decoupling: for $$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$2, typical reduced density matrices are exponentially close to the maximally mixed state, which is itself Gaussian. At $$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$3, the non-Gaussianity is continuous but nonanalytic because the derivative with respect to $$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$4 jumps at the Page point (Ares et al., 18 May 2026).

With a global $$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$5 symmetry at fixed filling $$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$6, the scaling changes to

$$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$7

with $$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$8. Thus the symmetryless case has vanishing non-Gaussianity below half system size, whereas the $$M_{mn}=-\frac{i}{2}\,\langle\psi|[\gamma_m,\gamma_n]|\psi\rangle.$$9-symmetric case retains an $M^TM=\mathds{1}$0, filling-independent contribution for $M^TM=\mathds{1}$1 and an extensive volume law with coefficient $M^TM=\mathds{1}$2 for $M^TM=\mathds{1}$3 (Ares et al., 18 May 2026).

A complementary perspective studies full many-body pure states rather than subsystems. Haar-random pure states have nearly maximal antiflatness, and recent work on one-body purity and non-Gaussianity in random superpositions of Gaussian states sharpens this distinction. Random superpositions of polynomially many Gaussian states exhibit one-body purity above its lower bound and non-Gaussianity below $M^TM=\mathds{1}$4, whereas random superpositions of exponentially many Gaussian states drive the one-body purity close to its lower bound and approach the maximal volume-law coefficient $M^TM=\mathds{1}$5 (Świętek et al., 1 Jul 2026). This suggests that “Gaussian rank” of the superposition is a useful organizing principle for distinguishing integrable from nonintegrable typicality.

4. Many-body phases, localization, and dynamical growth

In interacting many-body systems, fermionic non-Gaussianity has emerged as a phase-sensitive probe. In the disordered XXZ chain, the FAF density $M^TM=\mathds{1}$6 approaches its typical-state value $M^TM=\mathds{1}$7 in the ergodic regime, corresponding to a volume law $M^TM=\mathds{1}$8. Deep in the many-body-localized regime, the paper reports

$M^TM=\mathds{1}$9

with $\Gamma^2=-I_{2n}$0 in the XXZ chain and $\Gamma^2=-I_{2n}$1 in an impurity model. Consequently, MBL suppresses fermionic non-Gaussianity, but the scaling differs sharply between extensive and local interactions: extensive in XXZ, area-law-like in the impurity setting (Falcão et al., 30 Jan 2026).

The same study shows that rare long-range catlike eigenstates strongly enhance FAF. For ideal resonant cat segments, the covariance matrix vanishes on the resonant support and the FAF equals the number of spins participating in the resonance. Numerically, long-range catlike eigenstates display $\Gamma^2=-I_{2n}$2 matching the fluctuating segment length, and short-range resonances produce a pronounced peak near $\Gamma^2=-I_{2n}$3 at the ergodic–MBL crossover (Falcão et al., 30 Jan 2026). This makes FAF a sensitive diagnostic of rare-event mechanisms proposed to destabilize MBL.

Equilibrium results in other interacting systems support the same interpretation. In the ANNNI model and impurity-deformed Ising chains, FAF detects phase transitions, exhibits logarithmic anomalies at criticality under open boundary conditions, and vanishes or becomes boundary-localized at special solvable points such as the Peschel–Emery line (Sierant et al., 30 May 2025). In pure-gauge flux ladders, the fermionic antiflatness density $\Gamma^2=-I_{2n}$4 is extensive in weak-coupling SU(2) ladders but decreases approximately as $\Gamma^2=-I_{2n}$5 in the Abelian $\Gamma^2=-I_{2n}$6 case, consistent with the latter remaining close to a free-fermion description after Jordan–Wigner mapping (Santra et al., 8 Oct 2025).

Dynamics reveal similarly sharp contrasts. In the MBL regime, starting from the Néel state, the disorder-averaged FAF grows anomalously slowly and saturates through a power law,

$\Gamma^2=-I_{2n}$7

reflecting slow relaxation of two-point Majorana correlators (Falcão et al., 30 Jan 2026). By contrast, in ergodic random circuits and interacting Hamiltonian dynamics, FAF becomes extensive rapidly, with random circuits exhibiting exponential saturation in time and local Hamiltonians exhibiting much slower, conservation-law-constrained relaxation (Sierant et al., 30 May 2025).

5. Detection, witnesses, and Bell-sampling algorithms

The covariance-matrix character of fermionic non-Gaussianity enables practical tests. For pure-state testing, a two-copy Bell-measurement protocol estimates $\Gamma^2=-I_{2n}$8 through the observable

$\Gamma^2=-I_{2n}$9

with

$A$0

The associated one-sided Gaussianity tester distinguishes pure Gaussian states from states $A$1-far from the Gaussian set using

$A$2

two-copy Bell measurements, while a single-copy commuting-matchings protocol yields a two-sided tester with

$A$3

samples (Haug et al., 25 May 2026).

For mixed states, the purity-corrected witness

$A$4

satisfies $A$5 is not Gaussian. Under global depolarizing noise $A$6, the witness obeys

$A$7

so any pure non-Gaussian state remains certified for every $A$8 (Haug et al., 25 May 2026). That same work reports an implementation on the IQM Garnet quantum computer and observes that noise can either enhance or suppress the witness depending on circuit parameters.

Bell sampling has also led to new monotones and algorithms. Measuring the spectrum of

$A$9

on two copies produces the bridge spectral distribution $\ell$0. The bridge degree $\ell$1, defined by the largest occupied eigensector, is faithful, additive, and non-increasing under post-selected Gaussian protocols. It therefore yields exact-conversion no-go theorems and lower bounds on non-Gaussian gate complexity. The same Bell-sampling primitive gives a two-copy Gaussianity test with perfect completeness that is optimal among two-copy tests sharing that property, and a test for the state $\ell$2-design property of matchgate-invariant ensembles (Tarabunga, 3 Jun 2026).

An earlier three-copy alternative is based on fermionic convolution. For a pure even state $\ell$3, prepare $\ell$4 from two copies and perform a swap test against a third copy. Acceptance probability equals

$\ell$5

and $\ell$6 if and only if $\ell$7 is Gaussian (Lyu et al., 2024).

6. Complexity, asymmetry, and broader implications

The resource-theoretic meaning of fermionic non-Gaussianity is most explicit in simulation and state-preparation results. For fermionic linear optics with non-Gaussian initial states, exact strong simulation scales polynomially in mode number and quadratically in the Gaussian rank, while approximate weak simulation scales linearly in the Gaussian extent up to polynomial factors (Dias et al., 2023). In the covariance-based resource theory, occupation-number entropy $\ell$8 lower-bounds the number of SWAP gates required for preparation, and natural-orbital participation entropies upper-bound the classical simulation cost in an orthonormal Gaussian basis (Tarabunga et al., 2 Jul 2026).

Bridge-degree methods make these complexity statements sharper in the exact-conversion regime. Since $\ell$9 is additive and monotone under post-selected Gaussian protocols, any $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$0-doped Gaussian protocol using four-Majorana non-Gaussian gates satisfies $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$1; states with $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$2 therefore require $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$3 non-Gaussian gates. The same framework implies irreversibility of the exact resource theory and yields lower bounds for constructing approximate state $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$4-designs (Tarabunga, 3 Jun 2026).

Cross-resource relations have also been established. Particle-number asymmetry, quantified for pure states by the Shannon entropy $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$5 of the number distribution, lower-bounds the relative entropy of non-Gaussianity through a concentration inequality for Gaussian number statistics. Large $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$6-asymmetry therefore forces large non-Gaussianity because mixed fermionic Gaussian states have concentrated particle-number distributions and at most $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$7 number entropy (Ares et al., 17 Mar 2026). This does not identify the two resources, but it does show that symmetry-breaking number fluctuations can serve as a practical certificate of non-Gaussianity.

A consistent picture emerges across these developments. Fermionic non-Gaussianity can vanish for typical subsystems of Haar-random states below the Page point, remain finite under global constraints such as $C_{ij}=\operatorname{Tr}(\rho_A c_i^\dagger c_j)$8, become extensive in ergodic eigenstates, be parametrically suppressed in localized regimes, and admit experimentally accessible Bell-sampling, covariance, and convolution-based diagnostics. This suggests that the subject now links random-state typicality, many-body phase structure, and resource-theoretic hardness within a common covariance-based language (Ares et al., 18 May 2026, Falcão et al., 30 Jan 2026, Haug et al., 25 May 2026).