Why are composite systems described by tensor products?

The question in perspective

In standard textbook presentations of quantum mechanics—such as Section 4.2.3 of the Quantum Computing course—the rule that composite systems are described by the tensor product of their component Hilbert spaces,

\[\mathcal{H}_{AB} = \mathcal{H}_A \otimes \mathcal{H}_B,\]

is introduced as an independent postulate—the fourth postulate in the Dirac–von Neumann framework. In classical mechanics, the state space of two distinct particles is the Cartesian product of their individual phase spaces, \(M_{AB} = M_A \times M_B\). A composite pure state is simply an ordered pair of points \((x_A, x_B)\). When moving to statistical mechanics, probability distributions combine through the tensor product of function spaces: \(L^1(M_A \times M_B) \cong L^1(M_A) \otimes L^1(M_B)\).

In quantum theory, pure states are unit rays in a Hilbert space \(\mathcal{H}\). If composite systems were combined by the Cartesian product \(\mathcal{H}_A \times \mathcal{H}_B\), the state space of two qubits would have only four complex dimensions and could not support linear superpositions of product states. If they were combined by the direct sum \(\mathcal{H}_A \oplus \mathcal{H}_B\), the composite would behave as a single particle with four mutually exclusive choices rather than two independent, coexisting subsystems. The tensor product allows linear combinations of product states, \(\ket{\psi} = \sum_{ij} c_{ij} \ket{i}_A \ket{j}_B\), which produces entanglement, violations of Bell inequalities, and an exponential expansion of dimension: \(\dim(\mathcal{H}_{AB}) = \dim(\mathcal{H}_A) \dim(\mathcal{H}_B)\).

The foundational question is whether the tensor product is an arbitrary convention or whether it is mathematically forced by deeper physical and operational principles. Modern axiomatic reconstructions and algebraic quantum field theory show that the tensor product is indeed uniquely singled out under natural assumptions—most notably local tomography and no-signalling (causality)—while also revealing precisely where and why the tensor product fails in infinite-dimensional and relativistic settings.

The uniqueness proofs: Operational reconstructions and algebraic factorisation

Two distinct mathematical programs investigate the necessity of the tensor product: the operational reconstruction program (deriving quantum theory from informational axioms) and the algebraic program (deriving spatial tensor products from commuting observable algebras).

1. Operational Derivation via Local Tomography

In operational quantum foundations, physical systems are defined through preparations, transformations, and measurements. Let two separated observers, Alice and Bob, choose measurement settings \(x\) and \(y\) and observe outcomes \(a\) and \(b\). Two core operational principles govern composite systems:

Subsystem Independence (Causality / No-Signalling): Alice's choice of measurement cannot alter Bob's marginal outcome distribution, and vice versa:

\[\sum_b p(a, b \mid x, y) = p(a \mid x), \qquad \sum_a p(a, b \mid x, y) = p(b \mid y).\]

This condition guarantees that probability assignments are bilinear in local measurement effects: \(p(a, b \mid x, y) = \operatorname{tr}\bigl(\rho_{AB} (E_{a|x} \otimes F_{b|y})\bigr)\).

Local Tomography (Tomographic Locality): The global state of a composite system is uniquely characterized by the joint statistics of measurements performed independently on its components. If two global states \(\rho_{AB}\) and \(\sigma_{AB}\) satisfy

\[\operatorname{tr}\bigl(\rho_{AB} (A_i \otimes B_j)\bigr) = \operatorname{tr}\bigl(\sigma_{AB} (A_i \otimes B_j)\bigr)\]

for all local observables \(A_i\) and \(B_j\), then \(\rho_{AB} = \sigma_{AB}\). There are no holistic degrees of freedom that evade detection through correlated local measurements.

Let \(K(d)\) denote the number of independent real parameters required to specify an arbitrary state of a \(d\)-dimensional system. Local tomography imposes a strict product condition on state-space dimensions:

\[K(d_{AB}) = K(d_A) \times K(d_B).\]

For an isolated quantum system on a \(d\)-dimensional complex Hilbert space \(\mathcal{H}\), a state is described by a \(d \times d\) Hermitian matrix with \(\operatorname{tr}(\rho) = 1\). The real vector space of Hermitian matrices has dimension \(d^2\). Under local tomography,

\[(d_{AB})^2 = d_A^2 \times d_B^2 = (d_A d_B)^2 \implies d_{AB} = d_A d_B.\]

Because the composite Hilbert space dimension must equal \(d_A d_B\), and the mapping \((\ket{\psi}_A, \ket{\phi}_B) \mapsto \ket{\psi, \phi}_{AB}\) must be bilinear to preserve convex combinations of preparations, the composite state space is uniquely the Hilbert space tensor product \(\mathcal{H}_{AB} \cong \mathcal{H}_A \otimes \mathcal{H}_B\).

Why the Choice of Number Field Matters

The uniqueness of the tensor product depends critically on the field of complex numbers \(\mathbb{C}\):

Real Quantum Mechanics (\(\mathbb{R}\)): If quantum states are defined over real Hilbert spaces, the density operators are \(d \times d\) real symmetric matrices. The parameter count is \(K_\mathbb{R}(d) = d(d+1)/2\). For two real qubits (\(d_A = d_B = 2\)),

\[K_\mathbb{R}(2) = \frac{2 \times 3}{2} = 3, \qquad K_\mathbb{R}(2) \times K_\mathbb{R}(2) = 3 \times 3 = 9.\]

However, the composite of two real qubits has Hilbert space dimension \(d_{AB} = 4\), whose space of real symmetric matrices has dimension

\[K_\mathbb{R}(4) = \frac{4 \times 5}{2} = 10 > 9.\]

Real quantum mechanics violates local tomography: there is a tenth global parameter that cannot be determined by correlating local measurements. This parameter is supported on the operator \(\sigma_y \otimes \sigma_y\). Although \(\sigma_y\) is antisymmetric and therefore not an observable on a single real qubit, the tensor product \(\sigma_y \otimes \sigma_y\) is real and symmetric. Local measurements on real qubits can never detect this degree of freedom.

Quaternionic Quantum Mechanics (\(\mathbb{H}\)): Over the quaternions, the non-commutative division algebra prevents the definition of a canonical bilinear tensor product of quaternionic modules altogether. Composite systems cannot be formed in the standard way.

2. The Algebraic View: Commuting Observables vs. Tensor Products

In relativistic quantum mechanics and algebraic quantum field theory (AQFT), one does not postulate a spatial tensor product of Hilbert spaces. Instead, the physical requirement of spatial separation is encoded as microcausality (Einstein causality): observables localized in spacelike-separated regions commute on a common global Hilbert space \(\mathcal{H}\):

\[[A, B] = 0 \quad \text{for all } A \in \mathcal{A}_A, \, B \in \mathcal{A}_B.\]

Does the existence of commuting observable algebras imply a tensor product factorisation of the Hilbert space?

In finite dimensions: Yes. By the representation theory of \(C^*\)-algebras (and the Artin–Wedderburn theorem), if \(\mathcal{A}_A\) and \(\mathcal{A}_B\) commute, generate \(\mathcal{B}(\mathcal{H})\), and each factor is a type I von Neumann algebra, there exists a unitary isomorphism \(\mathcal{H} \cong \mathcal{H}_A \otimes \mathcal{H}_B\) under which \(A \in \mathcal{A}_A\) acts as \(A \otimes I\) and \(B \in \mathcal{A}_B\) acts as \(I \otimes B\). In finite dimensions, microcausality mathematically derives the tensor product.

In infinite dimensions: No. In infinite-dimensional Hilbert spaces, the commuting-operator model and the tensor-product model diverge. This discrepancy was the subject of Tsirelson’s problem, which asked whether every correlation achievable by commuting observables in an infinite-dimensional field theory could be approximated by tensor-product states. The resolution of Tsirelson's problem via \(\mathrm{MIP}^* = \mathrm{RE}\) proved that commuting-operator correlations are strictly richer than tensor-product correlations.

What is established

Hardy’s operational reconstruction established that five operational axioms—most notably continuous reversible transformations and local tomography—single out quantum theory over complex Hilbert spaces. In particular, local tomography separates standard complex quantum mechanics from real quantum mechanics and higher-order theories.

Chiribella, D’Ariano, and Perinotti’s informational derivation proved that causality (the impossibility of signalling from the future or across space without interaction) and local tomography uniquely require the state space of composite systems to be the tensor product of the component state spaces.

The resolution of Tsirelson’s problem via \(\mathrm{MIP}^* = \mathrm{RE}\) demonstrated that in infinite dimensions, the commuting-operator model produces non-local correlations that cannot be approximated by any tensor-product state. For infinite-dimensional systems and relativistic quantum fields, the commuting-operator framework is strictly more general than the tensor product.

What remains unresolved

In gauge theories (such as quantum electrodynamics and lattice gauge theory) and for systems of identical fermions, the global physical Hilbert space does not admit a tensor product factorisation into local spatial subregions due to Gauss-law constraints and anti-commutation relations. Characterizing entanglement across spatial boundaries in such systems requires decomposing operator algebras into center elements and superselection sectors rather than partitioning Hilbert space factors.

Furthermore, within Generalized Probabilistic Theories (GPTs), the choice of tensor product is not unique: between the minimal tensor product (which admits only separable states) and the maximal tensor product (which admits all non-signalling states, including Popescu–Rohrlich boxes), there exists an entire continuum of valid state spaces. Identifying the precise physical principle that singles out the quantum state cone remains an active open question.

A concrete next investigation

Reproduce: count the real dimensions of state spaces for systems of dimensions \(d_A\) and \(d_B\) over \(\mathbb{R}\), \(\mathbb{C}\), and \(\mathbb{H}\). Verify that the local tomography condition \(K(d_A d_B) = K(d_A) K(d_B)\) holds identically for all dimensions if and only if the field is \(\mathbb{C}\).

Check: verify that \(\sigma_y \otimes \sigma_y\) is real and symmetric, and show that its Hilbert–Schmidt inner product with every product observable \(A \otimes B\) (where \(A, B \in \{I, \sigma_x, \sigma_z\}\)) is zero, proving that it is invisible to local measurements in real quantum mechanics.

Extend: analyze how gauge constraints (such as a one-dimensional \(\mathbb{Z}_2\) gauge theory on a small lattice) obstruct spatial tensor product factorisation, and compute the resulting center algebra for a chosen subsystem bipartition.

Selected references

1. Quantum Theory From Five Reasonable Axioms

   Lucien Hardy. 2001. Preprint

   Reconstructs quantum theory from five operational postulates, demonstrating that continuous reversibility and local tomography isolate the tensor product rule.

2. Informational derivation of Quantum Theory

   Giulio Chiribella, Giacomo Mauro D'Ariano, Paolo Perinotti. 2010. Paper

   Derives the mathematical formalism of quantum mechanics from operational axioms, proving that causality and local tomography uniquely specify the tensor product.

3. MIP* = RE

   Zhengfeng Ji, Anand Natarajan, Thomas Vidick, John Wright, Henry Yuen. 2020. Paper

   Resolves Tsirelson's problem and Connes' embedding conjecture, proving that commuting-operator correlations strictly exceed tensor-product correlations in infinite dimensions.

Years refer to first arXiv submission. Publication details are available in the linked records.