Averaged Local-Unitary Invariants

• Averaged local-unitary invariants are quantities that remain unchanged under local unitary transformations after an averaging step (e.g., Haar twirling), ensuring basis independence.
• They compress orbit data for multipartite states and quantum channels, enabling hierarchical classification and effective separability detection.
• These invariants mitigate noncanonical choices by averaging over fluctuating inputs, thereby providing robust descriptors for entanglement characterization and LU equivalence.

Searching arXiv for recent and foundational papers on averaged local-unitary invariants. Averaged local-unitary invariants are quantities that remain unchanged under local unitary transformations and are obtained by an additional averaging, twirling, or aggregation step. In the literature, the term covers more than one construction. For multipartite pure states, it includes Haar averages of cumulant-derived polynomials over $SU(2)^n$, producing basis-independent invariants organized into hierarchical families (Mitchison, 2010). For quantum channels, it includes “moments” formed by averaging two-qubit local-unitary invariants over Haar-random product inputs or Haar-random local unitaries acting on maximally entangled inputs, yielding input-independent descriptors of entanglement creation, preservation, and destruction (Shaglel et al., 2 Jun 2026). In both settings, averaging is used to compress orbit data into canonical scalar quantities while preserving entanglement-relevant structure.

1. Conceptual setting

Local-unitary invariants are constant on orbits of the local unitary group and therefore describe the geometry of the orbit space. For multipartite states, this orbit-space viewpoint is standard: a local unitary invariant satisfies

$I(g|\psi\rangle)=I(|\psi\rangle),$

and a fundamental set of such invariants determines orbits and the geometry of the quotient by local unitaries (Mitchison, 2010). For mixed states, the invariant-theoretic formulation is equally explicit: two density operators are in the same local unitary orbit if and only if they agree on polynomial invariants in a Noetherian ring, which implies the existence of a finite complete set of invariants for local unitary equivalence (Turner et al., 2015).

Within that broader theory, “averaged” invariants are distinguished by how they are constructed. Rather than using a single trace polynomial, coefficient-matrix spectrum, or Bloch-tensor contraction, they introduce an additional averaging layer that removes basis or input dependence.

Construction	Averaging object	Resulting invariant
Cumulant twirling	Haar measure on $SU(2)^n$	family $I_{i_1\dots i_n}$
Channel moments	Haar-random inputs or unitaries	$\mathcal{C}^{(t)}$, $\mathcal{P}^{(t)}$
Decomposition averaging	eigenspace or ensemble choices	basis-independent necessary invariants

This organization already indicates that the phrase does not denote a single universal formalism. A plausible implication is that averaged local-unitary invariants are best understood as a subclass of LU invariants in which averaging is the mechanism that enforces canonicity or operational relevance.

2. Haar-twirled cumulant invariants for multipartite states

A foundational construction for pure qubit systems begins from cumulants. The state coefficients are embedded into a commutative algebra with nilpotent generators, so that $\log|\psi\rangle$ admits a finite expansion whose coefficients are interpreted as cumulants. These cumulants detect separability: a state is $T$-separable if and only if the relevant cumulants vanish whenever the index set splits the partition $T$ (Mitchison, 2010).

Because cumulants are basis-dependent, the basis dependence is removed by averaging over the local unitary group. The resulting invariants are defined as Haar averages of the squared modulus of cumulant-derived polynomials. In the two-qubit case, the basic polynomial is

$d_{11}=a_{11}a_{00}-a_{10}a_{01},$

and the associated invariant is

$I(g|\psi\rangle)=I(|\psi\rangle),$0

For three qubits, the lifted family includes $I(g|\psi\rangle)=I(|\psi\rangle),$1, $I(g|\psi\rangle)=I(|\psi\rangle),$2, $I(g|\psi\rangle)=I(|\psi\rangle),$3, and $I(g|\psi\rangle)=I(|\psi\rangle),$4, with $I(g|\psi\rangle)=I(|\psi\rangle),$5, $I(g|\psi\rangle)=I(|\psi\rangle),$6, and $I(g|\psi\rangle)=I(|\psi\rangle),$7 as degree-4 invariants and $I(g|\psi\rangle)=I(|\psi\rangle),$8 as a degree-6 invariant (Mitchison, 2010).

The cumulant family is structurally large but not exhaustive. For pure $I(g|\psi\rangle)=I(|\psi\rangle),$9-qubit states, the orbit-space dimension is

$SU(2)^n$0

while the cumulant-based construction yields $SU(2)^n$1 algebraically independent invariants, which the paper describes as “about half the dimension of orbit space” (Mitchison, 2010). This suggests that Haar-twirled cumulant invariants provide a substantial, highly organized subfamily of LU invariants rather than the entire invariant ring.

3. Hierarchy, lifting, and tracing-out

One of the distinctive features of averaged local-unitary invariants in the cumulant framework is their hierarchical organization. Invariants are related by “lifting”: if $SU(2)^n$2 is an invariant for $SU(2)^n$3 parties, then a systematic procedure produces invariants for $SU(2)^n$4, $SU(2)^n$5, and larger systems by adding zero indices and averaging over the added subsystem. The converse operation is tracing out subsystems. This relation is stated explicitly as

$SU(2)^n$6

The paper identifies this as a simplifying principle: invariants fall into families related by tracing-out, and these families grow exponentially with the number of subsystems (Mitchison, 2010).

The same concern with eliminating noncanonical choices appears in mixed-state LU theory. For arbitrary-dimensional bipartite mixed states with degenerate spectra, decomposition-dependent invariants become ambiguous because eigenvectors are not unique inside degenerate eigenspaces. A reduced, decomposition-independent family is then obtained by averaging or combining over eigenvector choices within the degenerate sector; these averaged invariants are basis-independent and furnish necessary conditions for local unitary equivalence (Zhou et al., 2012). A related construction defines state-decomposition-independent invariants from the characteristic polynomial of

$SU(2)^n$7

and, more generally, from hyperdeterminants of higher-order tensors $SU(2)^n$8; the resulting quantities are independent of the pure-state decomposition and invariant under local unitary transformations (Zhang et al., 2013).

Taken together, these results show that “averaging” can refer either to Haar twirling over a symmetry group or to symmetrization over nonunique decompositions. In both cases, the purpose is the same: to remove auxiliary choices that are not intrinsic to the local-unitary orbit.

4. Channel moments as averaged local-unitary invariants

A recent and explicit use of the term appears in the channel setting, where averaged local-unitary invariants are referred to as moments. The construction begins with two standard LU invariants of a two-qubit state,

$SU(2)^n$9

and then averages them over Haar-random inputs (Shaglel et al., 2 Jun 2026).

For a two-qubit channel $I_{i_1\dots i_n}$0, entanglement-creation moments are defined by

$I_{i_1\dots i_n}$1

$I_{i_1\dots i_n}$2

For a single-qubit channel $I_{i_1\dots i_n}$3, entanglement-breaking or entanglement-preserving moments are defined from maximally entangled inputs: $I_{i_1\dots i_n}$4

$I_{i_1\dots i_n}$5

These moments are invariant under local pre- and post-processing, so they depend only on the local-unitary equivalence class of the channel (Shaglel et al., 2 Jun 2026).

Their operational content is immediate. Any two-qubit separable state satisfies $I_{i_1\dots i_n}$6 for $I_{i_1\dots i_n}$7. Therefore, if $I_{i_1\dots i_n}$8 is non-entangling, then $I_{i_1\dots i_n}$9, while $\mathcal{C}^{(t)}$0 certifies entanglement creation. Likewise, if $\mathcal{C}^{(t)}$1 is entanglement-breaking, then $\mathcal{C}^{(t)}$2, while $\mathcal{C}^{(t)}$3 certifies entanglement preservation (Shaglel et al., 2 Jun 2026).

Higher-order moments refine second-order ones. The paper gives the example that CNOT and the B-gate have the same $\mathcal{C}^{(t)}$4, but different fourth moments,

$\mathcal{C}^{(t)}$5

It also shows that combining moments from different channel families improves the discrimination of locally inequivalent two-qubit unitaries. Thus, averaged local-unitary invariants become computable channel descriptors rather than merely state-classification tools (Shaglel et al., 2 Jun 2026).

5. Relation to ordinary LU invariant theory

Averaged local-unitary invariants sit inside a much larger invariant-theoretic landscape. For multipartite pure states, singular values of coefficient matrices $\mathcal{C}^{(t)}$6 are LU invariants; for multipartite mixed states, rank, eigenvalues, and trace products of coefficient matrices of eigenvectors are LU invariants, and for full-rank nondegenerate mixed states this set is necessary and sufficient for LU equivalence (Zhang et al., 2013). For arbitrary-dimensional bipartite mixed states, complete classification can be expressed through spectral invariants $\mathcal{C}^{(t)}$7 together with trace invariants built from products of reshaped eigenvector matrices $\mathcal{C}^{(t)}$8 and $\mathcal{C}^{(t)}$9 (Zhou et al., 2012). For generic multi-qubit states in generalized Bloch form, complete polynomial sets can be chosen with at most 12 invariants for two qubits and at most 90 for three qubits (Jing et al., 2015).

Against that background, averaged constructions should not be confused with completeness results. Reduced spectra, for example, are LU invariants but do not exhaust LU information. A complementary construction based on reduced-density-matrix eigenvectors forms matrices $\mathcal{P}^{(t)}$0 from tensor unfoldings and uses the coefficients of $\mathcal{P}^{(t)}$1 as LU invariants; these depend on reduced eigenvectors rather than only eigenvalues and can distinguish LU-inequivalent degenerate mixed states (Wang et al., 2014). This suggests that averaging and aggregation typically trade some orbit detail for structural simplicity, decomposition independence, or operational interpretability.

A distinct but related recent direction studies which LU-invariant multi-invariants are monotonic, on average, under LOCC. There the relevant quantities are graph-labeled polynomial invariants, and if the associated graph is connected and edge-convex, then $\mathcal{P}^{(t)}$2 is a pure-state entanglement monotone on average. The current conjectural classification identifies the edge-convex cases with Cayley graphs of finite Coxeter groups and is proved for all but six connected Coxeter types (Gadde et al., 8 Sep 2025). This is not an averaging construction in the same sense as Haar twirling, but it shows that “average” behavior under LOCC has become a parallel criterion for selecting distinguished LU invariants.

6. Terminological distinctions and common misconceptions

A common source of confusion is that not every invariant built from aggregated data is an averaged invariant in the Haar sense. Some coefficient-matrix constructions “aggregate local-unitary information” through singular values or trace products, but the underlying papers explicitly note that they do not literally define an “averaged LU invariant” in the statistical sense (Zhang et al., 2013). Likewise, Bloch-representation methods build invariants from trace powers, determinants, and scalar contractions of coefficient matrices, but do not formulate them as averages over local unitary groups (Cui et al., 2020).

An experimental variant introduces a different kind of averaging. In two-qubit interferometric schemes based on Hong–Ou–Mandel interference, local-unitary invariants are measured as expectation values of multi-copy projectors. The approach uses repeated trials on multiple identical copies of the state, so the relevant “averaging” is over experimental frequencies of coalescence and anti-coalescence rather than over a Haar orbit of local unitaries (Bartkiewicz et al., 2017). This distinction is essential: multi-copy statistical averaging, decomposition averaging, and Haar twirling are mathematically different procedures even when they produce LU-invariant scalars.

Another misconception is that higher-order averaged invariants automatically solve the full LU-classification problem. The channel-moment framework shows that fourth-order moments can distinguish channels beyond second-order moments alone, but it also states that higher-order data still do not fully resolve all channels (Shaglel et al., 2 Jun 2026). Similarly, the cumulant family provides a large algebraically independent hierarchy but not the full orbit-space coordinate ring (Mitchison, 2010). Averaged local-unitary invariants are therefore best viewed as structured, often highly effective summaries of LU orbit data whose strength depends on context: separability detection, channel characterization, hierarchical organization, or decomposition independence.