PUBLICATIONS

P.01 // PREPRINT

Correlation Kernels Determine Local Symmetry and Projective Dynamics

AUTHOR
Michael Chamerski
AFFILIATION
Aathra Research LLC
DATE
August 9, 2026
KEYWORDS
multipartite tensor-product structure · correlation kernel · invariant connection · holonomy · Lie generation · quantum control
FORMAT
HTML · PDF (8 PP)

ABSTRACT //

A multipartite factorization separates traceless observables into one-site and correlation sectors. We solve the inverse problem of recognizing the correlation sector without assuming the factorization. Let be an arbitrary real subspace, set , and let be its Lie normalizer in . The subspace is the joint kernel of all one-site partial traces for a factorization of type exactly when and . The simple ideals of then recover the local matrix algebras. The same kernel is the kernel of the linear marginal map and the direction space of every full-rank equal-marginal state fibre. It also gives the tangent space to factorization space and the horizontal sector of its unique invariant connection. That connection has full local projective holonomy. The skew-Hermitian kernel generates at bracket step two. For every pure ray, fully product factorizations of the fixed type form a measure-zero submanifold. No physical control law or spacetime claim is derived.

1 Introduction

A tensor-product structure specifies local observable algebras, partial traces, reduced states, product operations, and entanglement predicates. A bare finite-dimensional Hilbert space selects none of these. Existing constructions start from accessible commuting algebras, an operator partition, a Hamiltonian or its spectrum, or a trial family of factorizations [1, 9, 10, 15, 20, 22, 23]. Embedded local Lie algebras and their commutants also have established roles in quantum control [24]. Pure-state separability is consequently a relation between a ray and a subsystem structure [6].

Recent work represents a tensor-product structure by its one-local operator subspace and measures unitary displacement by Hilbert–Schmidt projector distance [3]. Their zero-distance classification already gives the full stabilizer as local unitaries composed with permutations among equal-dimensional factors. Orthogonal complementation therefore gives the same full stabilizer for the correlation kernel. The projected interaction component also defines the infinitesimal tensor-product-structure metric [21]. Related horizontal distributions occur in geometric quantum control and in the problem on compact symmetric spaces [14, 18].

The homogeneous-space and holonomy ingredients are standard [2, 13, 17]. Lie generation is standard [4, 11, 12, 19]. The result not contained in the cited constructions is an intrinsic recognition criterion. It begins with an arbitrary global operator subspace, computes its Lie normalizer, and decides whether the subspace is exactly a multipartite one-site marginal kernel. Under the equivalent recognition conditions, the same reductive pair gives the tensor-product-structure tangent bundle, invariant connection, curvature, holonomy, and step-two projective Lie closure. Recent subsystem geometry formulated through Segre stabilizers and operator partitions does not give this normalizer-and-orthogonality criterion [7, 16].

Projective dynamics means Lie closure in . It does not mean that the recognized horizontal directions are physically available controls.

2 Intrinsic correlation data

Let with . Scalar unitary phase acts trivially by conjugation, so the global transformation group isIts Lie algebra iswhere denotes the traceless skew-Hermitian operators. This convention matters because Hermitian operators are multiplied by before they are treated as infinitesimal unitary generators. For , the adjoint action is , where is any unitary representative of the projective class . The result is independent of the representative.

denotes the real vector space of traceless Hermitian operators on . It carries the Hilbert–Schmidt inner productThe corresponding positive inner product on is . For a subspace of either ambient real inner-product space, the superscript denotes its orthogonal complement in that ambient space.

Let be an arbitrary real subspace. This is the input to the inverse problem. It is not assumed to arise from subsystems, partial traces, or a tensor product. Write the same subspace in skew-Hermitian generator form asMultiplication by converts Hermitian operators into infinitesimal unitary generators. The Lie normalizer contains exactly those global generators whose commutators preserve . Explicitly, is the set of commutators with . The space is only a real vector subspace at this stage. It is not assumed to be a Lie algebra. Equation (1) extracts the continuous symmetry algebra of the supplied subspace without presupposing what that symmetry represents.

3 Recognition of correlation kernels

3.1 Kernel of a known factorization

Fix integers and with . A supplied factorization of this type is a unitary identificationWrite and let denote the complementary set of sites. The operators and are the identities on and on . We write for the partial trace over the complementary factors, so its output is an operator on .

For , define its embedding into the global operator algebra byWe write . Operators with different site labels commute, and their ordinary product is the corresponding tensor-product operator.

We call the joint kernel of these one-site marginal maps the correlation kernel of the factorization:It contains exactly the traceless operator directions invisible to all one-site marginals. Its orthogonal complement is the one-site sectorConsequently

For two qubits, the split has the concrete formHere are the Pauli matrices, and the indices and range independently over . The six one-site Pauli operators and the nine two-site Pauli products exhaust the fifteen-dimensional real space . The second subspace is therefore not an added class of operators. It is the full orthogonal complement of the first.

The factorization was used only to define the target object. The inverse problem starts with the arbitrary subspace from Section 2 and asks whether it is the correlation kernel of any factorization of the stated type.

3.2 Intrinsic criterion

For a unitary identification and a subspace , the notation means the conjugated subspace .

Theorem 3.1For the fixed factor type , the following statements are equivalent.

  1. (a) There is a unitary identification for which is the joint kernel of the one-site partial traces on .

  2. (b) The intrinsic spaces in Eq. (1) satisfyas compact real Lie algebras.

When these conditions hold, the simple ideals of generate the local full matrix algebras. The factorization is therefore recovered from up to local unitary frames and permutations of equal-dimensional factors.

Condition (a) is a concrete tensor-product statement. Condition (b) uses only the supplied subspace, the Hilbert–Schmidt metric, and its Lie normalizer. In Eq. (3), the orthogonal complement is taken inside . The equality says that the supplied subspace exhausts every direction orthogonal to its own normalizer. The Lie-algebra isomorphism says that this normalizer has exactly the continuous algebra of an -part local unitary group.

Proof Forward implication. Assume statement (a) and use to identify the tensor factors. Under this identification, . Setfor every nonempty set of sites . A tensor product of local Hermitian bases gives the following orthogonal decomposition by operator support:The set records the factors on which an operator is traceless. A term survives a one-site partial trace exactly when its support is that one site. This proves Eq. (2) directly.

Set . Equations (1) and (2) give . Local conjugations preserve operator support, so .

Let belong to this normalizer and define . This linear map is skew-adjoint for the Hilbert–Schmidt inner product. Since it preserves , it also preserves . Its restriction to the semisimple algebra is a derivation and is therefore inner. There is such thatThe difference centralizes every local matrix algebra. These algebras generate , so their commutant contains only scalars. No nonzero scalar lies in . Hence , andThis proves statement (b).

Reverse implication. Assume statement (b), without assuming a factorization. If an element of centralizes , it preserves and therefore belongs to . The center of the semisimple algebra is zero, soThe representation of on is consequently irreducible. Otherwise an invariant orthogonal projection would make a nonzero element of the centralizer.

Write according to its simple ideals. An irreducible complex module for a direct sum of compact simple Lie algebras is an outer tensor productHere carries the irreducible action of the th simple ideal, the other ideals act trivially on , and the full algebra acts factorwise on the displayed tensor product. Every acts nontrivially because the embedding in is faithful. Its least nontrivial complex representation has dimension [5]. Thus for every . The equality forces factor by factor.

Each is equivalent either to the standard -dimensional representation of or to its dual. In either case, the image of generates as a unital complex algebra. The tensor product above therefore supplies a unitary factorization of type whose one-site traceless sector is . Since , Eq. (2) identifies with the joint one-site marginal kernel. This proves statement (a).

The decomposition of a semisimple Lie algebra into simple ideals is unique up to permutation. Let denote the smallest unital complex associative subalgebra of containing the represented ideal . Each recovered ideal generatesThus every simple ideal recovers one commuting full matrix algebra. The algebras commute, and multiplication givesOnly local changes of basis and permutations among equal-dimensional simple ideals remain. ◻

4 Marginals and correlation fibres

4.1 Linear marginal quotient

Fix a factorization recovered by Theorem 3.1 and write . Define the linear marginal map byIts kernel is . The map is onto. To see this, put traceless and form the global operatorThe th summand has th marginal . Its other one-site marginals vanish because is traceless. Hence . The resulting exact sequence isTwo traceless global operators have the same one-site data exactly when their difference lies in . Thus the exact sequence says in one line that the marginal map is onto and loses precisely the correlation kernel.

The dimension of this lost sector can be read either from the exact sequence or from the support decomposition:It is positive for every and .

4.2 Equal-marginal states

Let be the density-operator space. The same partial-trace formula defines the affine state mapIt is onto because the tensor product of any tuple of one-site states realizes that tuple. The fibre through isPositive definiteness is open in the trace-one affine space. Hence the fibre through every full-rank state is locally an affine manifold with direction space and dimension given by Eq. (5).

The two-qubit case gives an explicit witness. For , the statesare positive definite and distinct. Both have reduced state at each site. Their difference lies entirely in the nine-dimensional correlation kernel.

4.3 Local state insufficiency

The embedding defined in Section 3.1 turns a local observable into a global operator. A state defines the global functional and the one-site restrictions . A factorization is locally state-sufficient only if is injective. No nontrivial recognized factorization has this property. The tuple of restrictions fixes the component of in and leaves its component undetermined. The missing state information is exactly the correlation kernel reconstructed in Theorem 3.1.

5 Cut-relative separability

Let denote the projective Hilbert space, whose points are one-dimensional complex subspaces of . A nonzero vector defines the ray . This distinction removes normalization and global phase from the pure-state description. A cut means an ordered tensor-product factorization of with the fixed factor dimensions .

5.1 Product states at a fixed cut

A density operator is fully product relative to the recovered factorization when . Equivalently, its global functional is multiplicative across the local embeddings:for all . Elementary tensor operators span , so equality of the two functionals on these operators forces . Productness is therefore defined only after the embeddings have been fixed.

5.2 Product cuts for a fixed ray

The projective local-unitary group of the reference factorization isThe group changes local frames without changing the ordered factorization. We use as a descriptive symbol for the space of ordered tensor-product structures of type . Let be the identity of and let be the reference cut. A general cut is denoted by . It is obtained by conjugating the reference factor algebras by . Permuting equal-dimensional factors can produce a different ordered point with the same unlabeled kernel. Such permutations preserve full productness, so they do not affect the dimension or measure statements below.

Let be the fully product pure-ray locus of the reference cut and writeFor a general cut , defineThe notation uses the projective action induced by any unitary representative of . The set contains exactly the pure rays that are fully product relative to the cut . For a fixed ray , its product-cut locus isChoose a reference product ray and with . Define the ray stabilizer byEach local projective unitary group acts transitively on its local pure rays, so acts transitively on . A coset belongs to exactly when differs from by a ray stabilizer on the left and a local frame transformation on the right. ThereforeThe orbit of a ray is , with real dimension . Since , its stabilizer has dimensionAt a product ray, the stabilizer inside the th local factor has dimension . HenceSubtracting these dimensions from givesThe inequality follows by expanding . The compact orbit is therefore a closed embedded submanifold of positive codimension. It has zero measure for the normalized -invariant measure on factorization space.

Transitivity of maps a reference product ray to any chosen ray. It also maps a reference nonproduct ray to that same ray. HenceEvery pure ray is product for some cut, no pure ray is product for every cut, and almost every ordered cut makes any fixed ray nonproduct. Full pure-state separability is therefore a relation between a ray and a factorization. It is not a nonconstant -invariant unary property of the ray.

The factor algebras remain distinct commuting subalgebras. Standard mixed-state separability remains the convex hull of cut-relative product states. These results neither prove nor invalidate Bell local causality and predict no Bell violation.

6 Correlation geometry

6.1 Tangent directions

For the reference cut, multiplication by gives the orthogonal reductive splittingThe vertical directions change local frames while leaving the cut fixed. The complementary directions change the cut itself. The letter is the conventional notation for a reductive complement on a general homogeneous space. No symmetric-space relation is assumed.

At the cut , the correlation sector is . This definition is independent of the representative because preserves the reference kernel. The associated bundle records a cut together with a correlation operator expressed in a local frame. Its equivalence relation is for and .

Equip with the normal homogeneous metric induced by on , and transport the Hilbert–Schmidt metric on to every fibre of the associated bundle. Write for the full tangent bundle of factorization space. With these metrics, the mapis a -equivariant isometric vector-bundle isomorphism. The tangent space at the reference cut is . The splitting in Eq. (12) identifies this quotient with , and left translation gives every other fibre. Thus correlation operators represent infinitesimal changes of factorization, not additional local observables.

6.2 Kernel stabilizer

The kernel also determines its local symmetry group. For a subspace , its stabilizer under conjugation is . The group normalizer of is . Let be the symmetric group on the factor labels, and defineA superscript denotes the connected component containing the identity. The stabilizer and its identity component satisfyConjugation preserves the Hilbert–Schmidt inner product. Hence an element of preserves exactly when it preserves the orthogonal complement . Preserving is equivalent to normalizing the connected subgroup with Lie algebra . A normalizer element permutes the simple ideals of . Their dimensions allow only permutations among equal-dimensional factors. After such a permutation, every remaining automorphism of a local full matrix algebra is inner. ThereforeThe first arrow includes the local group, and the last map records the permutation induced on equal-dimensional factors. Exactness says that the elements inducing no permutation are precisely the local transformations. The sequence splits after unitary identifications between equal-dimensional factor spaces have been chosen. The orbit of the unlabeled kernel is , while retains the ordering of equal-dimensional factors.

6.3 Invariant connection

An invariant connection on is determined by an -invariant horizontal complement of at the identity. Every complement is the graph over of a linear map . This means that every vector in the complement has a unique form with . Invariance requires to intertwine the actions. Consider a support sector and a one-site ideal . If , choose . If , choose any . The factor is the th factor of . It acts through its adjoint representation on and trivially on . Its adjoint representation has no invariant covector, so every -equivariant map is zero. Summing over the support sectors and one-site ideals gives the following result. Here denotes the real vector space of -equivariant linear maps from to .The absence of an invariant covector means that the adjoint representation has no nonzero invariant real linear functional. The complement is the unique -invariant complement.

By Wang’s classification, invariant principal connections on correspond to equivariant projections onto [17]. The vanishing intertwiner space leaves one such projection. Let be the Hilbert–Schmidt orthogonal projection defined by Eq. (12). For , the left Maurer–Cartan form sends to . The connection and its horizontal space areHere is left multiplication and is its differential at the identity. This is the only principal -connection invariant under . A curve is horizontal exactly when , where this expression denotes left trivialization. Equivalently, under right trivialization, at the cut .

6.4 Curvature and local holonomy

Let denote the curvature two-form of . The Maurer–Cartan equation gives its value at the identity :The value is the local-frame component produced by commuting the horizontal directions and . The relevant curvature values come from pairs of two-site correlation directions. Fix a site , choose , and let be a Hilbert–Schmidt orthonormal basis of . Its quadratic Casimir identity isThe sum commutes with every unitary on , so it is scalar. Taking its trace fixes the displayed coefficient.

Choose and define two correlation directions for each basis element:Both lie in . Summing their commutators removes the helper site and leaves a generator acting only at site :Commutators of traceless Hermitian matrices span . Varying in Eq. (17) provesEquation (18) states explicitly that real linear combinations of curvature values exhaust the local Lie algebra. The connection is therefore nonflat, and its curvature spans every local projective generator.

The Ambrose–Singer theorem identifies the restricted holonomy algebra with the algebra generated by curvature values along horizontal lifts [2]. Equation (18) spans all of at the base point. Write for the holonomy group based at the point in the fibre over the reference cut . Since is connected,The chosen point identifies the holonomy group as a concrete subgroup of . Every is the endpoint of the horizontal lift, starting at , of a contractible piecewise-smooth loop in factorization space.

6.5 Projective Lie closure

Let denote the real linear span of all commutators with . Let denote the smallest real Lie subalgebra of containing . Equation (17) shows that . Combining this inclusion with the reductive splitting givesThus every projective unitary direction is either a correlation direction or a real linear combination of commutators of two correlation directions. This is the meaning of bracket generation at step two. The first bracket span itself has one exceptional form:In both cases, the Lie algebra generated by is all of .

Let be the real Lie algebra of all skew-Hermitian matrices. Adding the scalar direction givesScalar phase is the only unitary Lie-algebra direction outside the closure generated by the correlation kernel.

7 Mathematical scope

The inverse datum is an unlabeled subspace . Equation (3) recognizes whether that datum is a multipartite correlation kernel without using a prior factorization in the converse. It does not select from a bare Hilbert space. Equal-dimensional factor labels are not encoded in the subspace.

Equations (10) and (11) exclude intrinsic full pure-state separability. They do not deny product rays relative to a recognized cut. The maximally mixed state is product for every factorization, so the empty-intersection statement does not extend to mixed states.

The unique -invariant connection is nonflat, and every element of is its holonomy around a contractible loop. Hence no -invariant principal connection identifies local frames by endpoints alone. This does not exclude noninvariant connections or define physical transport.

The term correlation kernel is algebraic. For a Hamiltonian , the one-site marginals of need not vanish. Equation (20) implies algebraic controllability only for a system with arbitrary signed piecewise-constant generators taking values in . One fixed interaction need not have the same Lie closure. The connection supplies no Hamiltonian, intervention class, signaling protocol, propagation law, or spacetime interpretation. Correlation-induced holonomies in other state and control bundles are distinct constructions [8, 12].

A Exact first-bracket span

Two qubits. For operators on the first factor and on the second, write for the anticommutator. ThenThe anticommutator of two traceless Hermitian matrices is scalar. A commutator of two two-site correlation directions therefore contains only one-site terms. This gives . The reverse inclusion follows from Eq. (17), so equality holds.

Other bipartite dimensions. For every bipartite type other than , one factor has dimension . On that factor chooseThis operator is traceless and its square is not scalar. DefineThen and . Let label the other factor and choose noncommuting . The two correlation directionssatisfyThe first term is local and already lies in the first-bracket span by Eq. (17). Subtracting it leaves the nonzero correlation term .

In the bipartite case, is the external tensor product of the two real adjoint representations on and . Each adjoint representation is irreducible, and their external tensor product is irreducible under . The bracket span is -invariant. Once it contains one nonzero correlation direction, it therefore contains all of . Together with its local part, it is all of .

Three or more factors. Let . Identity factors are suppressed in the site-embedded operator products. Fix a support set with and choose distinct sites . Choose and . For every , choose . DefineBoth and have support on at least two sites, so they belong to . Their commutator isCommutators of and span the traceless skew-Hermitian operators on site . Varying all displayed operators therefore spans every sector with .

It remains to obtain a sector supported on two sites and . Choose a third site . Let be a Hilbert–Schmidt orthonormal basis of , and take and . DefineBoth are correlation directions. The Casimir identity on site givesThese brackets span every support-two sector. Hence . Combining this with from Eq. (17) proves Eq. (21) for every multipartite type.

Author contributions

The author conceived the work, formulated and verified the theorem, and assumes full responsibility. Large language models assisted literature discovery, adversarial proof checking, LaTeX drafting, and language editing.

Data and code availability

No data or computational results were used.

Competing interests

The author declares no competing interests.

References

[1]
A. Adil, M. S. Rudolph, A. Arrasmith, Z. Holmes, A. Albrecht, and A. Sornborger, “Search for classical subsystems in quantum worlds,” Physical Review D, 113, 103535 (2026), doi:10.1103/vj4x-96fk.
[2]
W. Ambrose and I. M. Singer, “A theorem on holonomy,” Transactions of the American Mathematical Society, 75, 428–443 (1953), doi:10.1090/S0002-9947-1953-0063739-1.
[3]
F. Andreadakis and P. Zanardi, “Tensor product structure geometry under unitary channels,” Quantum, 9, 1668 (2025), doi:10.22331/q-2025-03-25-1668.
[4]
W.-L. Chow, “Über systeme von linearen partiellen differentialgleichungen erster ordnung,” Mathematische Annalen, 117, 98–105 (1939), doi:10.1007/BF01450011.
[5]
W. Fulton and J. Harris, Representation theory: A first course, Graduate texts in mathematics, 129, Springer (1991), doi:10.1007/978-1-4612-0979-9.
[6]
N. L. Harshman and K. S. Ranade, “Observables can be tailored to change the entanglement of any pure state,” Physical Review A, 84, 012303 (2011), doi:10.1103/PhysRevA.84.012303.
[7]
K. Ikeda, “Quantum entanglement geometry on Severi–Brauer schemes: Subsystem reductions of Azumaya algebras,” arXiv:2601.13764 [math.AG] (2026), doi:10.48550/arXiv.2601.13764.
[8]
M. Johansson, M. Ericsson, K. Singh, E. Sjöqvist, and M. S. Williamson, “Correlation-induced non-abelian quantum holonomies,” Journal of Physics A: Mathematical and Theoretical, 44, 145301 (2011), doi:10.1088/1751-8113/44/14/145301.
[9]
X. F. Liu and C. P. Sun, “On the relative quantum entanglement with respect to tensor product structure,” arXiv:quant-ph/0410245 (2004), doi:10.48550/arXiv.quant-ph/0410245.
[10]
N. Loizeau and D. Sels, “Quantum mereology and subsystems from the spectrum,” Foundations of Physics, 55, 3 (2025), doi:10.1007/s10701-024-00813-2.
[11]
D. Lucarelli, “Chow’s theorem and universal holonomic quantum computation,” Journal of Physics A: Mathematical and General, 35, 5107–5116 (2002), doi:10.1088/0305-4470/35/24/309.
[12]
D. Lucarelli, “Control aspects of holonomic quantum computation,” Journal of Mathematical Physics, 46, 052103 (2005), doi:10.1063/1.1888028.
[13]
K. Nomizu, “Invariant affine connections on homogeneous spaces,” American Journal of Mathematics, 76, 33–65 (1954), doi:10.2307/2372398.
[14]
E. Perrier and C. S. Jackson, “Solving the problem with the global Cartan decomposition,” arXiv:2404.02358 (2024), doi:10.48550/arXiv.2404.02358.
[15]
A. Soulas, G. Franzmann, and A. Di Biagio, “On the emergence of preferred structures in quantum theory,” arXiv:2512.07468 [quant-ph] (2025), doi:10.48550/arXiv.2512.07468.
[16]
A. Vanrietvelde, O. Mestoudjian, and P. Arrighi, “Partitions in quantum theory,” arXiv:2506.22218 (2025), doi:10.48550/arXiv.2506.22218.
[17]
H.-C. Wang, “On invariant connections over a principal fibre bundle,” Nagoya Mathematical Journal, 13, 1–19 (1958), doi:10.1017/S0027763000023461.
[18]
R. Wiersema, A. F. Kemper, B. N. Bakalov, and N. Killoran, “Geometric quantum machine learning with horizontal quantum gates,” Physical Review Research, 7, 013148 (2025), doi:10.1103/PhysRevResearch.7.013148.
[19]
R. Wiersema, E. Kökcü, A. F. Kemper, and B. N. Bakalov, “Classification of dynamical lie algebras of 2-local spin systems on linear, circular and fully connected topologies,” npj Quantum Information, 10, 110 (2024), doi:10.1038/s41534-024-00900-2.
[20]
P. Zanardi, “Virtual quantum subsystems,” Physical Review Letters, 87, 077901 (2001), doi:10.1103/PhysRevLett.87.077901.
[21]
P. Zanardi, E. Dallas, and F. Andreadakis, “Mereological quantum phase transitions,” Physical Review A, 113, 042201 (2026), doi:10.1103/r193-5s7w.
[22]
P. Zanardi, E. Dallas, F. Andreadakis, and S. Lloyd, “Operational quantum mereology and minimal scrambling,” Quantum, 8, 1406 (2024), doi:10.22331/q-2024-07-11-1406.
[23]
P. Zanardi, D. A. Lidar, and S. Lloyd, “Quantum tensor product structures are observable-induced,” Physical Review Letters, 92, 060402 (2004), doi:10.1103/PhysRevLett.92.060402.
[24]
R. Zeier and T. Schulte-Herbrüggen, “Symmetry principles in quantum systems theory,” Journal of Mathematical Physics, 52, 113510 (2011), doi:10.1063/1.3657939.