Quantum Entanglement Revisited

Consider two quantum systems $A$ and $B$ with Hilbert spaces $\mathcal{H}_A$ (dimension $d_A$) and $\mathcal{H}_B$ (dimension $d_B$). A joint pure state lives in the tensor product space $\mathcal{H}_A \otimes \mathcal{H}_B$. We call $|\psi\rangle \in \mathcal{H}_A \otimes \mathcal{H}_B$ a product state (or separable pure state) if there exist $|a\rangle \in \mathcal{H}_A$ and $|b\rangle \in \mathcal{H}_B$ such that

Any pure state that cannot be written in this form is called entangled. This is a statement about the pure-state case; mixed states have a richer notion of separability (convex combinations of products) that we set aside here, focusing on pure bipartite states as the clean foundation. The subtlety is that a state can look complicated in one basis and be a product in another, or vice versa — so "entangled" must be a basis-independent property, and indeed it is: no local change of basis on $A$ alone or $B$ alone can turn a product state into an entangled one or vice versa, because local unitaries act as $U_A \otimes U_B$ and clearly preserve the product form $|a\rangle\otimes|b\rangle \mapsto (U_A|a\rangle)\otimes(U_B|b\rangle)$. This invariance under local unitary transformations is exactly what the Schmidt decomposition exploits: it finds the one basis (well, one canonical pair of orthonormal bases, one for each subsystem) in which the state's structure is maximally transparent. The Schmidt decomposition theorem states that every $|\psi\rangle \in \mathcal{H}_A \otimes \mathcal{H}_B$ can be written as

where $\{|u_i\rangle\}$ are orthonormal in $\mathcal{H}_A$, $\{|v_i\rangle\}$ are orthonormal in $\mathcal{H}_B$, the coefficients $\lambda_i > 0$ are real, satisfy $\sum_i \lambda_i^2 = 1$, and $r \le \min(d_A, d_B)$ is called the Schmidt rank. The state is a product state exactly when $r = 1$; it is entangled precisely when $r \ge 2$. This single integer — computable, basis-independent, and cheap to obtain via linear algebra — is the formal dividing line between "classically describable correlations of separate parts" and "genuine quantum entanglement." Among all bipartite states of two qubits, four occupy a special place: the Bell states,

They form an orthonormal basis for the two-qubit Hilbert space $\mathbb{C}^2 \otimes \mathbb{C}^2$ — the Bell basis — and each has Schmidt rank 2 with equal coefficients $\lambda_1 = \lambda_2 = 1/\sqrt2$, the maximum possible entanglement for a two-qubit pure state.

Step 1: set up coordinates. Write the bipartite state's amplitudes as a matrix, not a vector. If

collect the coefficients into a $d_A \times d_B$ matrix $C = (c_{jk})$. This reshaping is the crucial trick: entanglement structure, invisible in the flat coefficient list, becomes visible as the rank structure of a matrix.

Step 2: take the singular value decomposition. Every matrix $C$ admits an SVD

with $U$ a $d_A \times d_A$ unitary, $V$ a $d_B \times d_B$ unitary, and $\Sigma$ a $d_A \times d_B$ diagonal matrix of non-negative real numbers $\sigma_1 \ge \sigma_2 \ge \dots \ge 0$ (the singular values). This decomposition always exists and is computed by standard, efficient numerical linear algebra.

Step 3: read off the Schmidt decomposition. Define $|u_i\rangle = \sum_j U_{ji} |j\rangle_A$ (the $i$-th column of $U$) and $|v_i\rangle = \sum_k V^*_{ki} |k\rangle_B$ (the $i$-th column of $V$, conjugated appropriately for the bra-ket convention). Then

and the Schmidt coefficients are exactly the nonzero singular values (normalized so $\sum_i \sigma_i^2 = 1$, guaranteed automatically since $|\psi\rangle$ is normalized). The Schmidt rank $r$ is the number of nonzero singular values, i.e. $r = \operatorname{rank}(C)$.

Step 4: classify. If $r = 1$, the state is a product state — stop, you can write $|\psi\rangle = |u_1\rangle\otimes|v_1\rangle$ directly. If $r \ge 2$, the state is entangled, and the multiset $\{\sigma_i\}$ quantifies exactly how entangled it is (e.g. via the entanglement entropy $-\sum_i \sigma_i^2 \log_2 \sigma_i^2$, foreshadowing later measures of entanglement as a resource). ``` FUNCTION SchmidtDecompose(psi, dA, dB):

# psi is given as a flat list of dA*dB complex amplitudes,

# indexed psi[j][k] for basis state |j>_A ⊗ |k>_B

# Step 1: reshape into a dA x dB matrix C = Matrix(dA, dB) FOR j in 0 .. dA-1: FOR k in 0 .. dB-1: C[j][k] = psi[j][k]

# Step 2: singular value decomposition U, Sigma, V_dagger = SVD(C) # C = U Sigma V_dagger

# Sigma sorted descending, entries >= 0

# Step 3: extract nonzero singular values and vectors schmidt_terms = [] FOR i in 0 .. min(dA, dB) - 1: lambda_i = Sigma[i][i] IF lambda_i > EPSILON: # numerical tolerance u_i = Column(U, i) # orthonormal basis vector in H_A v_i = ConjugateColumn(V_dagger, i) # orthonormal basis vector in H_B schmidt_terms.append( (lambda_i, u_i, v_i) ) r = length(schmidt_terms) # Schmidt rank RETURN schmidt_terms, r FUNCTION ClassifyEntanglement(psi, dA, dB): terms, r = SchmidtDecompose(psi, dA, dB) IF r == 1: RETURN "PRODUCT STATE", terms ELSE: RETURN "ENTANGLED (rank " + r + ")", terms ``` The Bell basis itself is generated from the computational basis by a fixed two-qubit circuit, which is worth internalizing as a mechanism (it is the standard entangling gadget reused throughout quantum protocols): ``` FUNCTION PrepareBellState(a, b):

# a, b are classical bits selecting one of the four Bell states

# start with |a> ⊗ |b> qubit1 = |a> qubit2 = |b> ApplyHadamard(qubit1) # |a> -> (|0> + (-1)^a |1>)/sqrt(2) ApplyCNOT(control=qubit1, target=qubit2) RETURN qubit1, qubit2

# (a,b) = (0,0) -> |Phi+>, (0,1) -> |Psi+>

# (a,b) = (1,0) -> |Phi->, (1,1) -> |Psi-> ```

Read this diagram as a state-preparation pipeline: the Hadamard gate alone still leaves a product state $\left(\frac{|0\rangle+(-1)^a|1\rangle}{\sqrt2}\right)\otimes|b\rangle$ — no entanglement yet. Only after the CNOT gate correlates the second qubit's value with the first does the joint state become inseparable, since after the CNOT the amplitude matrix $C$ acquires two nonzero singular values instead of one.

Computing the Schmidt decomposition. For a bipartite state on $d_A \times d_B$ dimensions, forming the coefficient matrix $C$ takes $O(d_A d_B)$ time (just a reshape, no arithmetic). Computing the SVD of a $d_A \times d_B$ matrix by standard algorithms (e.g. Golub–Kahan) costs $O(d_A d_B \min(d_A, d_B))$ time using classical dense linear algebra, and $O(d_A d_B)$ space to store the matrix and its factors. For a system of $n$ qubits split into subsystem $A$ (size $n_A$ qubits, $d_A = 2^{n_A}$) and $B$ (size $n_B = n - n_A$ qubits), this is exponential in $n$ in the worst case — but that exponential cost is a fact about classically representing a general quantum state, not a flaw in the Schmidt method itself; the method is the fastest known classical way to certify separability given the state vector explicitly.

Correctness of the Schmidt decomposition — proof sketch. The claim rests directly on the SVD theorem from linear algebra: every complex matrix $C \in \mathbb{C}^{d_A \times d_B}$ factors as $C = U\Sigma V^\dagger$ with $U, V$ unitary and $\Sigma$ non-negative diagonal. This is proven by diagonalizing the Hermitian positive-semidefinite matrix $C^\dagger C$ (or $CC^\dagger$), whose eigenvectors give the columns of $V$ (or $U$) and whose eigenvalues' square roots give the singular values. Translating this matrix identity back into Dirac notation term by term, using $|j\rangle_A \otimes |k\rangle_B$ as the fixed tensor basis, produces exactly

with orthonormality of $\{|u_i\rangle\}$ and $\{|v_i\rangle\}$ inherited from unitarity of $U$ and $V$. Normalization $\sum_i \sigma_i^2 = 1$ follows from $\langle \psi|\psi\rangle = \operatorname{Tr}(C^\dagger C) = \operatorname{Tr}(\Sigma^2) = \sum_i \sigma_i^2 = 1$.

Correctness of the rank test. If $|\psi\rangle$ is a product state $|a\rangle\otimes|b\rangle$, its coefficient matrix is the outer product $C = a b^T$ (up to conjugation conventions), which has rank exactly 1 whenever $a, b \ne 0$ — a basic fact about outer products. Conversely, if $\operatorname{rank}(C) = 1$, the SVD guarantees a single nonzero singular value, so $|\psi\rangle = \sigma_1 |u_1\rangle\otimes|v_1\rangle$ is manifestly a product state (absorb the scalar $\sigma_1$ into either factor). Hence rank 1 if and only if product — this biconditional is the formal definition of entanglement made computable.

Uniqueness caveat. The Schmidt coefficients $\{\sigma_i\}$ are always unique (they are eigenvalues of a fixed operator, the reduced density matrix, discussed below). The bases $\{|u_i\rangle\}, \{|v_i\rangle\}$ are unique only when all nonzero $\sigma_i$ are distinct; degenerate singular values permit a change of basis within the degenerate subspace (an SVD ambiguity, not a flaw in the theorem). This is worth flagging because naive implementations sometimes assume a unique output basis when the true invariant is only the multiset of coefficients.

Connection to reduced density matrices. Tracing out subsystem $B$ from $|\psi\rangle\langle\psi|$ gives $\rho_A = \sum_i \sigma_i^2 |u_i\rangle\langle u_i|$ — the Schmidt vectors are exactly the eigenvectors of the reduced density matrix, and the $\sigma_i^2$ are its eigenvalues. This is why $\rho_A$ is mixed (has more than one nonzero eigenvalue) precisely when $|\psi\rangle$ is entangled, giving a second, equivalent, and often more practical entanglement test: check whether the reduced state of either subsystem is pure or mixed.

Tradeoffs. The Schmidt/SVD approach is exact and efficient given the amplitudes, but obtaining those amplitudes for a physical system generally requires either full state tomography (exponentially many measurements) or already knowing the state from how it was prepared. In practice, entanglement in the lab is certified via entanglement witnesses or Bell-inequality violation statistics rather than by directly computing an SVD — the Schmidt decomposition is the theoretical gold standard, not the experimental protocol.

Let's trace the algorithm on a state that is not symmetric and not obviously either type at a glance:

Step 1 — build the coefficient matrix. Using basis order $\{|0\rangle,|1\rangle\}$ for both qubits,

Step 2 — compute $C^\dagger C$ (real here, so $C^\dagger = C^T$):

Step 3 — eigenvalues of $C^T C$. Solve $\det(C^TC - \mu I) = 0$:

So $\mu_1 \approx 0.9714$, $\mu_2 \approx 0.0286$, giving singular values

Check normalization: $\sigma_1^2 + \sigma_2^2 = 0.9714 + 0.0286 = 1.000$ ✓ — matches $\langle\psi|\psi\rangle = (4+1+1+0)/6 = 1$.

Step 4 — classify. Both singular values are strictly positive, so $\operatorname{rank}(C) = 2$: the Schmidt rank is 2, so $|\psi\rangle$ is entangled, despite its coefficients looking like an "almost separable" list at first glance (the $|11\rangle$ term is entirely absent, which might mislead a hasty eye into guessing "product state").

Step 5 — extract the Schmidt vectors (sketch). Solving $(C^TC - \mu_1 I)v_1 = 0$ gives the eigenvector direction $v_1 \propto (2, 1)$ up to normalization — i.e. $|v_1\rangle \approx 0.894|0\rangle + 0.447|1\rangle$ for subsystem $B$; the corresponding $|u_1\rangle$ for subsystem $A$ follows from $u_1 = Cv_1/\sigma_1$. We omit the full numeric normalization here, but the essential output is already established: two nonzero terms, i.e. genuine entanglement, with a dominant $\sigma_1 \approx 0.986$ term and a small correction $\sigma_2 \approx 0.169$ — this state is "close to" a product state but not exactly one. Contrast this with the Bell state $|\Phi^+\rangle = \frac{1}{\sqrt2}(|00\rangle+|11\rangle)$, whose coefficient matrix is $C = \frac{1}{\sqrt2}\begin{pmatrix}1&0\\0&1\end{pmatrix}$, already diagonal with $\sigma_1=\sigma_2=1/\sqrt2$ — maximal, perfectly balanced entanglement, the opposite extreme from the lopsided example above.

The idea that a composite quantum state need not decompose into independent parts was recognized as deeply strange from the very inception of quantum mechanics, and it took decades for the community to develop the precise linear-algebraic language — Schmidt rank, Bell basis — that this article uses.

Every protocol built on shared entangled pairs — quantum key distribution (e.g. the Ekert protocol), superdense coding, and Quantum Teleportation — depends on the sender and receiver holding a state with Schmidt rank exactly 2 and equal coefficients, i.e. a Bell state, or being able to certify that they do. Real photonic and superconducting-qubit experiments that claim to "distribute entanglement" over fiber or through the air are, at bottom, claiming to prepare and verify high-fidelity Bell states, and the Schmidt-rank criterion (via reduced-state purity, or via witnessed CHSH violation as a proxy) is the formal standard against which fidelity is measured. In quantum error correction, stabilizer codes such as the Shor code and the surface code protect logical qubits by encoding them into highly entangled multi-qubit states; while the bipartite Schmidt formalism developed here is the simplest case, the underlying intuition — that information "smeared" across entangled degrees of freedom resists local corruption — is exactly what motivates those larger constructions. In computational complexity, the amount of entanglement generated during a quantum circuit's execution is a leading candidate explanation for when and why quantum algorithms outperform classical ones; classical simulation methods such as tensor-network and matrix-product-state techniques exploit low Schmidt rank across cuts of a circuit to simulate weakly entangled quantum computations efficiently, while circuits that build up large Schmidt rank resist such simulation — a direct, practical descendant of the very rank test introduced in this article.

Ask the tutor: How does the Schmidt decomposition of a state change when you apply a local unitary to just one subsystem?