Fermionic Gaussian States: Theory & Applications

• Fermionic Gaussian States are states of many-body systems defined by quadratic fermionic operators and fully characterized by two-point correlation data.
• They underpin efficient classical simulations, matchgate circuits, and variational methods in condensed-matter and quantum information studies.
• Their structure enables precise analytical treatment of entanglement, thermal dynamics, and resource theories of non-Gaussianity.

Searching arXiv for recent and foundational papers on fermionic Gaussian states to ground the article. Fermionic Gaussian states (FGS) are states of fermionic many-body systems whose density operators are exponential of quadratic fermionic forms, or, in the pure case, states obtained from the vacuum by fermionic Gaussian unitaries. They are the natural states of free or non-interacting fermions, obey Wick’s theorem, and are completely specified by two-point correlation data. Because their many-body structure reduces to covariance-matrix linear algebra, they are central in condensed-matter physics, Jordan–Wigner treatments of spin systems, matchgate and fermionic linear-optics simulation, and contemporary resource-theoretic analyses of non-Gaussianity and magic (Ares et al., 17 Mar 2026, Lyu et al., 2024).

1. Algebraic definitions and equivalent formulations

An $N$-mode fermionic system is generated by annihilation and creation operators $c_i,c_i^\dagger$ satisfying

$$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$

A mixed, full-rank fermionic Gaussian state can be written as

$\rho=\frac{1}{Z}e^{-K},$

with $K$ quadratic in the fermions. In Nambu form, with

$$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$

one writes $K=\mathbf c^\dagger M_K \mathbf c$, where $M_K$ has the standard block structure determined by matrices $A$ and $B$. The same class is equivalently described in a Majorana representation by

$c_i,c_i^\dagger$0

where $c_i,c_i^\dagger$1 is real antisymmetric and the Majorana operators satisfy the Clifford relations. Ground states and thermal states of quadratic fermionic Hamiltonians are therefore Gaussian, and pure Gaussian states arise by acting with a fermionic Gaussian unitary on the vacuum (Ares et al., 17 Mar 2026, Lyu et al., 2024).

The Majorana operators are commonly introduced as

$c_i,c_i^\dagger$2

or equivalently with $c_i,c_i^\dagger$3, and obey

$c_i,c_i^\dagger$4

This representation makes manifest that Gaussian unitaries act linearly on Majoranas by orthogonal transformations. In the particle-number-conserving subclass, often called $c_i,c_i^\dagger$5-FGS, a pure state can be written as a filled Fermi sea

$c_i,c_i^\dagger$6

after diagonalizing the one-body correlation matrix; this is the Slater-determinant form emphasized in correlation-matrix algorithms and in several many-body applications (Liu et al., 2024, Langer et al., 5 Mar 2026).

Parity superselection is structurally important. Physical fermionic states commute with the parity operator, and the even sector is the operationally relevant one in several formulations of fermionic Gaussianity, convolution, and separability. A useful misconception to exclude is that fermionic Gaussianity is synonymous with number conservation or with ordinary Slater determinants alone: pairing terms $c_i,c_i^\dagger$7 and $c_i,c_i^\dagger$8 are part of the general definition, so superconducting and BCS-type states are included on equal footing (Lyu et al., 2024, Negari et al., 25 Aug 2025, Boutin et al., 2021).

2. Covariance matrices, canonical forms, and geometry

FGS are characterized by two-point functions. In the Nambu formulation one uses

$c_i,c_i^\dagger$9

In the Majorana formulation one uses real antisymmetric covariance matrices such as

$$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$0

or

$$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$1

Physicality is encoded by the bona fide condition

$$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$2

equivalently by the requirement that the spectrum of $$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$3 lies in $$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$4. For pure Gaussian states, the covariance matrix saturates the corresponding constraint: in different conventions this appears as $$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$5, or as $$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$6, or as the statement that the covariance matrix is orthogonal and antisymmetric (Negari et al., 25 Aug 2025, Sierant et al., 30 May 2025, Langer et al., 5 Mar 2026).

Wick’s theorem is the defining operational feature. For Gaussian states, higher-order correlators are fixed by two-point data; in Majorana language, even correlators are Pfaffians of covariance submatrices and odd correlators vanish. The entropy of a Gaussian state is correspondingly reducible to the paired eigenvalues $$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$7 of the correlation matrix: $$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$8 This covariance-matrix closure is the basis for both analytical formulas and scalable numerics (Ares et al., 17 Mar 2026, Carollo et al., 2019).

Across bipartitions, pure FGS admit a canonical BCS-like form under local fermionic Gaussian unitaries,

$$\{c_i,c_j^\dagger\}=\delta_{ij},\qquad \{c_i,c_j\}=\{c_i^\dagger,c_j^\dagger\}=0.$$9

up to permutations within the subsystems. This is the fermionic analogue of a Schmidt decomposition: entanglement decomposes into independent two-mode pairs and inert vacuum modes. In a more geometric language, pure fermionic Gaussian states correspond to compatible Kähler triples $\rho=\frac{1}{Z}e^{-K},$0 with

$\rho=\frac{1}{Z}e^{-K},$1

whereas mixed Gaussian states obey the same structural relations but generally have $\rho=\frac{1}{Z}e^{-K},$2. The manifold of pure FGS is

$\rho=\frac{1}{Z}e^{-K},$3

of dimension $\rho=\frac{1}{Z}e^{-K},$4 (Iannotti et al., 29 Apr 2026, Hackl et al., 2020).

3. Simulation, circuit representation, and efficient classical descriptions

FGS are classically tractable because they are fully described by $\rho=\frac{1}{Z}e^{-K},$5 covariance data rather than a $\rho=\frac{1}{Z}e^{-K},$6-component wavefunction. Gaussian unitaries act by orthogonal rotations on Majorana space,

$\rho=\frac{1}{Z}e^{-K},$7

so dynamics, reduced states, and many observables reduce to matrix manipulations. This is the sense in which FGS are the “free” or classically simulable states of fermionic linear optics and matchgate computation (Sierant et al., 30 May 2025, Lyu et al., 2024).

Pure FGS admit an exact matchgate-circuit representation,

$\rho=\frac{1}{Z}e^{-K},$8

where $\rho=\frac{1}{Z}e^{-K},$9 is a matchgate circuit and $K$0 a computational basis state. Under such a circuit,

$K$1

with $K$2, and the covariance matrix evolves as $K$3. This structure underlies optimal synthesis results for pure FGS, including lower bounds on nearest-neighbor gate counts in terms of cumulative Schmidt-rank growth, exact optimality statements for right-standard-form matchgate circuits, and constructive preparation algorithms based on symmetric Euler decompositions and Givens rotations. The same framework yields criteria for shallow preparability: $K$4-banded covariance matrices are equivalent, up to constants, to preparation by matchgate circuits of depth $K$5 (Langer et al., 5 Mar 2026).

Correlation-matrix methods also connect FGS to tensor-network descriptions. For particle-number-conserving pure FGS with $K$6, recursive Schmidt decomposition and Gaussian projection rules yield an efficient conversion to matrix product states. In translationally invariant infinite-cylinder settings, the resulting iMPS transfer matrix can be used to extract minimally entangled states and anyon sectors in chiral spin-liquid constructions, together with entanglement spectra and, in principle, modular data (Liu et al., 2024). Practical numerical introductions additionally exploit covariance-matrix diagonalization, imaginary-time evolution, and Gaussian bond-dimension reduction to build MPS-like compression schemes for one-dimensional and sectorized two-dimensional free-fermion states (Surace et al., 2021).

High-temperature interacting fermionic Gibbs states exhibit an additional structural simplification. For bounded-degree local fermionic Hamiltonians, sufficiently small inverse temperature implies that

$K$7

where each $K$8 is a fermionic Gaussian state. The structural theorem holds for

$K$9

and a corresponding classical sampling algorithm is established in the more restrictive regime

$$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$0

with runtime polynomial in $$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$1 and $$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$2. This places the convex hull of Gaussian states, not merely Gaussian states themselves, at the center of high-temperature classical tractability for local fermions (Ramkumar et al., 14 May 2025).

4. Analytical, variational, and dynamical applications

Because Gaussian structure survives under quadratic evolution and supports closed formulas for observables, FGS are a standard analytical laboratory for many-body dynamics, thermodynamics, and metrology. A closed-form symmetric logarithmic derivative has been derived for arbitrary FGS, with the central covariance-matrix equation

$$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$3

This makes the quantum Fisher information of thermal states and non-equilibrium steady states of quadratic fermionic systems directly computable from covariance data rather than from full many-body operator algebra (Carollo et al., 2019).

FGS also function as variational manifolds for interacting systems. In generalized Hartree–Fock treatments of the long-range transverse-field Ising model, the interacting spin problem is mapped to Majoranas and its ground state is approximated by a pure Gaussian covariance matrix $$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$4 satisfying $$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$5. The variational energy contains Pfaffians of covariance submatrices, and self-consistency can be implemented either by imaginary-time evolution,

$$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$6

or by zero-temperature fixed-point updates. In the weak long-range regime $$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$7, this Gaussian treatment shows excellent agreement with DMRG and linked-cluster benchmarks for phase boundaries and entanglement scaling, while for $$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$8 it still captures the logarithmic growth of entanglement entropy away from criticality (Kaicher et al., 2023).

A related but richer ansatz is the coherent superposition of non-orthogonal fermionic Gaussian states,

$$\mathbf c=(c_1,\dots,c_N,c_1^\dagger,\dots,c_N^\dagger),$$9

This SGS construction is used as a practical impurity solver for Anderson and two-channel Kondo models. Each component state remains individually Gaussian, overlaps and matrix elements are handled through Gram matrices and Pfaffians, and the covariance matrices obey a closed, purity-preserving evolution equation. The ansatz has $K=\mathbf c^\dagger M_K \mathbf c$0 parameters, its update cycle costs $K=\mathbf c^\dagger M_K \mathbf c$1, and it captures screening clouds, impurity Rényi entropies, and non-Fermi-liquid signatures in geometries where standard tensor-network methods are less convenient (Boutin et al., 2021).

FGS also support nontrivial measurement-induced phenomena. For two identical copies of a half-filled, particle-number-conserving Gaussian state,

$K=\mathbf c^\dagger M_K \mathbf c$2

post-selected Bell measurements on half of the corresponding interlayer rungs force the unmeasured half into a tensor product of Bell pairs, independent of the microscopic details of the original Gaussian state. In the half-filled case $K=\mathbf c^\dagger M_K \mathbf c$3, the post-measurement state factorizes as

$K=\mathbf c^\dagger M_K \mathbf c$4

and the remaining subsystem has maximal interlayer entanglement $K=\mathbf c^\dagger M_K \mathbf c$5. The result is traced to the determinant structure of Slater amplitudes together with fermionic statistics (Fang et al., 17 Dec 2025).

5. Entanglement, partial transpose, and extendibility

Entanglement theory for FGS departs sharply from the bosonic Gaussian case at the level of partial transposition. For a fermionic Gaussian state with covariance matrix $K=\mathbf c^\dagger M_K \mathbf c$6, the partial transpose is generally not Gaussian. In a suitable Majorana basis, however, it takes the exact form

$K=\mathbf c^\dagger M_K \mathbf c$7

where $K=\mathbf c^\dagger M_K \mathbf c$8 are Gaussian operators with covariance matrices

$K=\mathbf c^\dagger M_K \mathbf c$9

For pure bipartite Gaussian states, $M_K$0 and $M_K$1 commute, and the logarithmic negativity reduces to

$M_K$2

with $M_K$3 the eigenvalues of the reduced covariance matrix. For mixed states the non-Gaussianity of $M_K$4 obstructs such a direct formula, but the two-Gaussian decomposition remains the starting point for efficient evaluation of integer moments and for lower bounds consistent with conformal-field-theory predictions in the Ising chain (Eisler et al., 2015).

This non-Gaussian partial-transpose structure motivates rigorous negativity bounds. One lower bound is obtained by local pinching or twirling to decouple the covariance matrix into $M_K$5 mode-pair blocks. Two upper bounds are available: one from semidefinite programming based on the Lagrangian theory of fermionic Gaussian maps, and another from the Gaussian-operator decomposition of $M_K$6. These bounds are efficiently computable from covariance data, remain rigorous, and reproduce the expected adjacent-interval scaling $M_K$7 in the homogeneous critical chain, while also detecting topological edge-state contributions in dimerized SSH-type systems (Eisert et al., 2016).

Extendibility and separability impose another rigid layer of Gaussian structure. For a bipartite fermionic Gaussian state with covariance matrix

$M_K$8

$M_K$9-extendibility is equivalent to the existence of a fermionic Gaussian extension. The corresponding covariance-matrix criterion is an explicit semidefinite program whose size scales linearly with the number of modes. This leads to finite de Finetti–type bounds in trace norm, relative entropy of entanglement, and squashed entanglement, with two-mode asymptotics of order $A$0. In the Gaussian sector, extendibility is especially rigid: $A$1

and product, arbitrary extendibility, and physical separability coincide (Negari et al., 25 Aug 2025).

A related rigorous approximation program for general interacting fermions uses fermionic extendibility and evenization to show when site-product or Gaussian-product approximations are accurate in energy density. On graphs with large coordination, odd fermionic correlations satisfy monogamy bounds of order $A$2, which in turn imply graph-dependent guarantees for product-state approximations to ground states. This result is not a statement that all interacting fermionic ground states are close to Gaussian states; rather, it identifies explicit structural regimes in which fermionic mean-field or Hartree–Fock-like descriptions become provably accurate (Krumnow et al., 2024).

6. Gaussianity, non-Gaussianity, and magic resources

In the resource theory of fermionic non-Gaussianity, FGS form the free set. For a general state $A$3, the relative entropy of non-Gaussianity is

$A$4

where $A$5 is the Gaussianification of $A$6, the Gaussian state with the same covariance matrix. For pure states, $A$7. Fermionic convolution provides an operational construction of this Gaussianification: iterative self-convolution

$A$8

suppresses higher cumulants and converges to the Gaussian state with the same covariance matrix. This yields an efficient three-copy test for pure even Gaussianity and the finite-shot non-Gaussian entropy

$A$9

which vanishes exactly on Gaussian states and approaches the relative-entropy measure under iteration (Lyu et al., 2024).

Non-Gaussianity can also be lower-bounded from particle-number statistics. With $B$0, particle-number distribution $B$1, and Shannon entropy

$B$2

one has

$B$3

with equality for pure states. The key structural input is a concentration inequality showing that Gaussian states have sharply concentrated particle-number distributions. Consequently, a broad charge distribution certifies distance from the Gaussian manifold, leading to an explicit lower bound on $B$4 in terms of $B$5. The bound becomes informative for sufficiently large particle-number entropy and provides a practical witness because $B$6 is often efficiently computable or experimentally measurable (Ares et al., 17 Mar 2026).

A different but related construction is fermionic antiflatness,

$B$7

where $B$8 are Williamson eigenvalues of the Majorana covariance matrix. For pure states, $B$9 if and only if the state is Gaussian. Because

$c_i,c_i^\dagger$00

the measure has a direct interpretation in terms of two-point Majorana correlations and is used to detect phase transitions, critical scaling, excited-state complexity, and dynamical growth of fermionic non-Gaussianity in interacting systems (Sierant et al., 30 May 2025).

A final conceptual distinction is essential. Fermionic Gaussianity does not imply low stabilizer magic or low nonstabilizerness. Random pure FGS have filtered stabilizer Rényi entropies with extensive leading behavior close to Haar-random states and logarithmic subleading corrections, even though they remain free-fermionic and classically simulable (Collura et al., 2024). Likewise, the fermionic non-local magic of a pure FGS across a bipartition is exactly computable from the positive eigenvalues $c_i,c_i^\dagger$01 of the reduced Majorana covariance matrix,

$c_i,c_i^\dagger$02

with, for $c_i,c_i^\dagger$03,

$c_i,c_i^\dagger$04

This admits polynomial-time evaluation, supports Page-like formulas for random Gaussian states, logarithmic critical scaling in the XY chain, quasiparticle growth after quenches, and estimation by fermionic shadow tomography (Iannotti et al., 29 Apr 2026). The resulting picture is that Gaussianity is a statement about fermionic cumulant structure and simulability, not about the absence of all quantum resources.