Schmidt Decomposition

Write a two-particle quantum state in fixed local bases and it is a full rectangular array of amplitudes — a $d_A \times d_B$ matrix $C$ , opaque about what is correlated with what. The Schmidt decomposition is the theorem that this complexity is an illusion of the chosen coordinates. For any bipartite pure state there exist orthonormal local bases $\{|i\rangle_A\}$ and $\{|i\rangle_B\}$ — one on each side, chosen with foresight — in which the whole state collapses to a single diagonal sum $|\psi\rangle = \sum_i \lambda_i |i\rangle_A|i\rangle_B$ , with real, positive $\lambda_i$ obeying $\sum_i \lambda_i^2 = 1$ . A generic $d_A d_B$ -dimensional amplitude array is reduced to at most $\min(d_A,d_B)$ numbers. The engine underneath is a workhorse of linear algebra: the singular value decomposition. The amplitude matrix always factors as $C = U\Sigma V^\dagger$ , and the Schmidt coefficients are nothing but its singular values. Because the SVD exists for every complex matrix, the Schmidt form exists for every bipartite pure state — no exceptions. And because a matrix and its singular values are tightly constrained, so is entanglement: the squared coefficients $\lambda_i^2$ are simultaneously the eigenvalues of both reduced density matrices, forcing the two halves of an entangled pair to be spectrally identical no matter how lopsided their dimensions. This one diagonal line is why a single number — the entanglement entropy — captures everything a bipartite pure state hides, why local operations can never manufacture entanglement, and why matrix product states compress a many-body wavefunction. It is also intrinsically two-party: three or more subsystems admit no simultaneous Schmidt basis, and that failure is where multipartite entanglement gets hard.

Key takeaways

Every bipartite pure state can be written as one diagonal sum $|\psi\rangle=\sum_i \lambda_i|i\rangle_A|i\rangle_B$ over orthonormal local bases, with real $\lambda_i>0$ and $\sum_i\lambda_i^2=1$ — its Schmidt decomposition.
The $\lambda_i$ are the singular values of the amplitude matrix $C$ ; equivalently $\lambda_i^2$ are the eigenvalues of either reduced density matrix, so both reduced states share one nonzero spectrum even when $d_A\neq d_B$ .
The Schmidt rank $r$ (the count of nonzero $\lambda_i$ ) equals $1$ exactly for product states; $r\ge 2$ certifies entanglement, with no further computation needed.
Every pure-state entanglement measure is a function of $\{\lambda_i^2\}$ alone: the entanglement entropy is $-\sum_i\lambda_i^2\log_2\lambda_i^2$ , and concurrence, negativity, and LOCC convertibility all follow from the same vector.
The construction is bipartite by nature — three or more parties generally have no simultaneous Schmidt basis, which is what makes multipartite entanglement classification genuinely harder.
Key idea: the Schmidt decomposition is the singular value decomposition of a state's amplitude matrix — the singular values are the Schmidt coefficients, their squares are the common eigenvalues of both reduced density matrices, and the number of nonzero ones decides product versus entangled.

The core insight

The Schmidt decomposition is the singular value decomposition of a bipartite pure state's amplitude matrix: the singular values are the Schmidt coefficients $\lambda_i$ , their squares $\lambda_i^2$ are the shared eigenvalues of both reduced density matrices, and the count of nonzero ones — the Schmidt rank — decides product versus entangled.

Where this lives today

The Schmidt decomposition is the daily arithmetic of quantum many-body simulation. Every matrix product state, and the density-matrix renormalization group (DMRG) that made one-dimensional systems numerically tractable, is built by repeatedly Schmidt-decomposing a wavefunction across each bond and keeping only the largest $\chi$ coefficients — the optimal low-rank truncation (Eckart–Young), with error given by the discarded weight $\sum_{i>\chi}\lambda_i^2$ . Ground states of gapped one-dimensional systems obey an area law: their Schmidt spectra decay exponentially, so a modest $\chi$ captures the state to machine precision, which is why DMRG is effectively exact and why the same spectra detect quantum phase transitions through entanglement-entropy scaling. The identical idea governs classical simulation of circuits — the bond dimension a circuit generates across a cut is its Schmidt rank, and that rank must grow exponentially before the circuit is hard to simulate, making Schmidt rank a working measure of quantum advantage. A 1907 theorem about integral kernels is now the compression engine of computational quantum physics.

The physical picture

Lay the amplitudes of $|\psi\rangle$ out as a rectangular grid: rows indexed by $A$ , columns by $B$ , each cell holding the amplitude $C_{jk}$ for finding $A$ in $|j\rangle$ and $B$ in $|k\rangle$ . Entanglement is a statement about the shape of this grid. A product state is the outer product of one column pattern with one row pattern — a rank-one grid, a single stripe of correlation. An entangled state needs several independent stripes stacked together, and no local relabelling can merge them into one. Diagonalizing entanglement means rotating the row frame (a unitary on $A$ ) and the column frame (a unitary on $B$ ) independently until the grid is diagonal with nonnegative entries. That rotation-to-diagonal is precisely the singular value decomposition, and the surviving diagonal entries — the singular values — are the Schmidt coefficients. How many are nonzero is the Schmidt rank: how "wide" the entanglement is.

Figure

Diagonalizing entanglement: the amplitude matrix C of the state (1/√3)(00 + 01 + 10) is rotated by a local unitary on A rows and a local unitary on B columns into the diagonal Schmidt form Σ, giving singular values λ₁ = 0.934 and λ₂ = 0.357 with λ₁² + λ₂² = 1, so the Schmidt rank is 2 and the state is entangled

Now trace out one half. Because the Schmidt form is diagonal, the partial trace is immediate: $\rho_A$ is diagonal in the Schmidt basis with weights $\lambda_i^2$ , and $\rho_B$ carries the very same weights on its own basis. The two reduced states are spectral twins — identical ignorance about their partners, even if one lives in a two-dimensional space and the other in a hundred-dimensional one. This is the exact sense in which an EPR pair holds no local information yet perfect correlation. The second figure makes it concrete for a qutrit entangled with a qubit — unequal dimensions, one shared nonzero spectrum, one entropy.

Figure

Spectral twins across unequal dimensions: a qutrit A of dimension 3 entangled with a qubit B of dimension 2 has reduced states whose nonzero eigenvalues coincide at λ₁² = 0.75 and λ₂² = 0.25. The 3×3 matrix ρ_A simply carries an extra zero eigenvalue, and both reduced states have the same von Neumann entropy S = 0.811 ebit

Finally, note what a local unitary can and cannot do. Acting on $A$ alone rotates the frame $\{|i\rangle_A\}$ ; acting on $B$ alone rotates $\{|i\rangle_B\}$ . Neither touches the coefficients $\lambda_i$ . So the sorted Schmidt vector is a gauge-invariant fingerprint — the part of $|\psi\rangle$ that survives every local manoeuvre without communication. Two states are interconvertible by local unitaries alone if and only if they share this fingerprint, and every clean statement about bipartite pure-state entanglement is, in the end, about this vector.

The core equations

The central object is the decomposition itself: a single-index, bi-orthogonal sum. Its content is that a double sum over two independent local indices always collapses to one.

|\psi\rangle_{AB}=\sum_{i=1}^{r}\lambda_i\,|i\rangle_A\otimes|i\rangle_B \qquad (1)

where: $\lambda_i$ — Schmidt coefficients, real with $\lambda_i>0$ ; $r$ — Schmidt rank, $1\le r\le\min(d_A,d_B)$ ; $\{|i\rangle_A\},\{|i\rangle_B\}$ — orthonormal Schmidt bases in $\mathcal H_A,\mathcal H_B$ . Entanglement pairs one $A$ -state with one $B$ -state at a time. Normalization turns the squared coefficients into a probability distribution — the one you sample by measuring either subsystem in its Schmidt basis.

\sum_{i=1}^{r}\lambda_i^2=1,\qquad \lambda_i>0 \qquad (2)

where: $\lambda_i^2$ — the Schmidt weight, equal to the probability of the $i$ -th outcome in either reduced state. The whole existence claim rests on one linear-algebra fact: collect the amplitudes into a matrix and apply the singular value decomposition, and the Schmidt coefficients are its singular values.

C=U\Sigma V^\dagger,\qquad \Sigma=\mathrm{diag}(\lambda_1,\dots,\lambda_r) \qquad (3)

where: $C$ — the $d_A\times d_B$ amplitude matrix, $C_{jk}=\langle j|_A\langle k|_B|\psi\rangle$ ; $U,V$ — unitaries whose columns become the Schmidt bases; $\Sigma$ — diagonal matrix of singular values, identically the Schmidt coefficients. Because every complex matrix has an SVD, every bipartite pure state has a Schmidt form. Tracing out $B$ leaves a reduced state already diagonal in the Schmidt basis — no further diagonalization needed.

\rho_A=\mathrm{Tr}_B\,|\psi\rangle\langle\psi|=\sum_{i=1}^{r}\lambda_i^2\,|i\rangle_A\langle i|_A \qquad (4)

where: $\rho_A$ — reduced density matrix of $A$ ; $\lambda_i^2$ — its eigenvalue for eigenvector $|i\rangle_A$ , so the Schmidt basis is simultaneously the eigenbasis of $\rho_A$ . Because $\rho_A=CC^\dagger$ and $\rho_B=C^\dagger C$ are built from the same singular values, their nonzero spectra must agree — the spectral-twins relation.

\mathrm{spec}^{\neq 0}(\rho_A)=\mathrm{spec}^{\neq 0}(\rho_B)=\{\lambda_i^2\} \qquad (5)

where: $\mathrm{spec}^{\neq 0}$ — the set of nonzero eigenvalues. The larger space merely pads its spectrum with extra zeros; the shared entanglement content is unchanged. The unique pure-state entanglement measure is the von Neumann entropy of that shared spectrum — the entanglement entropy.

E(\psi)=S(\rho_A)=-\sum_{i=1}^{r}\lambda_i^2\log_2\lambda_i^2 \qquad (6)

where: $S$ — von Neumann entropy (base 2 gives ebits); $E(\psi)$ — entanglement entropy, equal for $A$ and $B$ , governing asymptotic distillation and dilution rates. The simplest thing the rank says is whether there is any entanglement at all: a pure state is a product exactly when its rank is one.

r=1\iff |\psi\rangle=|a\rangle_A|b\rangle_B\ (\text{product}) \qquad (7)

where: $r$ — Schmidt rank; $r\ge 2$ certifies entanglement with no threshold to tune — the cleanest pure-state separability test in quantum information. For two qubits the Schmidt spectrum is two numbers, and a second standard measure — the concurrence — is a simple product of them.

C(\psi)=2\lambda_1\lambda_2\qquad(\text{two qubits}) \qquad (8)

Derivation AFrom the singular value decomposition to the Schmidt form

The goal is to prove existence constructively: starting from an arbitrary bipartite pure state in any fixed product basis, we show that a single application of the singular value decomposition to its amplitude matrix produces orthonormal local bases in which the state is diagonal, with strictly positive coefficients that square to a probability distribution. Nothing is assumed beyond purity and finite dimension; the argument is an identity, not an approximation, and it makes the phrase "the Schmidt decomposition is the SVD of a quantum state" literal rather than metaphorical.

Step 1. Expand the state in arbitrary orthonormal product bases, $|\psi\rangle=\sum_{jk}C_{jk}\,|j\rangle_A|k\rangle_B$ , and read off the amplitude matrix $C_{jk}=\langle j|_A\langle k|_B|\psi\rangle$ , a generic $d_A\times d_B$ complex matrix. Why: everything about a pure bipartite state is stored in this array of amplitudes, so any structural theorem about the state must be a theorem about the matrix.

Step 2. Apply the singular value decomposition $C=U\Sigma V^\dagger$ , which exists for every complex matrix, with $U,V$ unitary and $\Sigma$ diagonal with entries $\lambda_i\ge 0$ . Exactly $r=\mathrm{rank}\,C$ of the $\lambda_i$ are strictly positive; the rest vanish. Why: the SVD is precisely the operation that finds the two rotations — one for rows, one for columns — that diagonalize a rectangular matrix, which is what "one basis per subsystem" demands.

Step 3. Define the rotated local vectors $|i\rangle_A=\sum_j U_{ji}|j\rangle_A$ and $|i\rangle_B=\sum_k V^*_{ki}|k\rangle_B$ . Unitarity gives $\langle i|i'\rangle_A=\sum_j U^*_{ji}U_{ji'}=\delta_{ii'}$ , and likewise on $B$ , so both new sets are orthonormal. Why: the Schmidt bases are not postulated — they are the columns of the SVD unitaries, and their orthonormality is inherited for free from $U^\dagger U=V^\dagger V=\mathbb 1$ .

Step 4. Substitute and reorder the sums: $|\psi\rangle=\sum_{jk}\big(\sum_i U_{ji}\lambda_i V^*_{ki}\big)|j\rangle|k\rangle=\sum_i \lambda_i|i\rangle_A|i\rangle_B$ . The double sum over $j,k$ collapses to a single sum over $i$ , with only the $r$ nonzero terms surviving. Why: this is the payoff — two free indices become one, and the state is diagonal in the paired Schmidt bases.

Step 5. Normalize: $1=\langle\psi|\psi\rangle=\sum_{i,i'}\lambda_i\lambda_{i'}\langle i|i'\rangle_A\langle i|i'\rangle_B=\sum_i\lambda_i^2$ , using orthonormality on both sides to kill cross terms. Why: the squared coefficients therefore form a genuine probability distribution, which is what lets them double as reduced-state eigenvalues and feed every entropy formula.

Go deeper · Derivation B · 1 minEqual reduced spectra and the entropy formula

Derivation BEqual reduced spectra and the entropy formula

The goal is to compute both reduced density matrices directly from the Schmidt form, prove they carry the same nonzero eigenvalues, and identify the von Neumann entropy of that shared spectrum as the entanglement entropy. This is the technical heart of why one number quantifies bipartite pure-state entanglement, and why the two halves of an entangled pair must be equally mixed even when their Hilbert spaces differ wildly in dimension.

Step 1. Form the global projector $|\psi\rangle\langle\psi|=\sum_{i,i'}\lambda_i\lambda_{i'}\,|i\rangle_A\langle i'|_A\otimes|i\rangle_B\langle i'|_B$ by taking the outer product of the Schmidt sum with itself. Why: the partial trace acts on this double sum, and writing it out exposes exactly which terms will survive tracing.

Step 2. Trace out $B$ using $\mathrm{Tr}_B\big(|i\rangle_B\langle i'|_B\big)=\langle i'|i\rangle_B=\delta_{ii'}$ , which collapses the double sum to $\rho_A=\sum_i\lambda_i^2|i\rangle_A\langle i|_A$ . Why: orthonormality of the $B$ -side Schmidt vectors annihilates every off-diagonal term, so $\rho_A$ emerges already diagonal — the Schmidt basis is its eigenbasis and the $\lambda_i^2$ are its eigenvalues, no diagonalization required.

Step 3. Trace out $A$ instead and find $\rho_B=\sum_i\lambda_i^2|i\rangle_B\langle i|_B$ by the identical argument. The two reduced states are built from the same weights $\lambda_i^2$ ; any dimensional mismatch between $A$ and $B$ shows up only as extra zero eigenvalues in the larger space. Why: this is the spectral-twins theorem — the nonzero spectra of $\rho_A$ and $\rho_B$ coincide, so all spectral quantities (entropy, purity, rank) agree across the cut.

Step 4. Evaluate the von Neumann entropy of a diagonal density matrix, which is just the Shannon entropy of its eigenvalues: $S(\rho_A)=-\sum_i\lambda_i^2\log_2\lambda_i^2=S(\rho_B)\equiv E(\psi)$ . Why: because the spectra are equal, $S(\rho_A)=S(\rho_B)$ automatically, and this common value is the entanglement entropy — the unique measure of pure-state entanglement, reported in ebits.

Go deeper · Derivation C · 1 minThe Schmidt rank as a separability certificate

Derivation CThe Schmidt rank as a separability certificate

The goal is to prove the sharpest structural statement the decomposition offers: a bipartite pure state is a product state if and only if its Schmidt rank equals one, with reduced-state purity giving a third equivalent form of the certificate. Unlike mixed-state separability, which is computationally hard, pure-state separability is decided by a single integer read off the singular values — the base case all entanglement detection builds on, turning "is this state entangled?" into "does its amplitude matrix have rank above one?".

Step 1. Forward direction: a product state $|\psi\rangle=|a\rangle_A|b\rangle_B$ has amplitude matrix $C_{jk}=a_j b_k$ , i.e. $C=ab^\top$ , an outer product. An outer product has exactly one nonzero singular value, so $r=1$ . Why: a product state is a single correlation stripe — one row pattern times one column pattern — which is the very definition of a rank-one matrix.

Step 2. Reverse direction: if $r=1$ the Schmidt sum has a single term, $|\psi\rangle=\lambda_1|1\rangle_A|1\rangle_B$ , and normalization forces $\lambda_1=1$ , which is manifestly a product state with factors $|1\rangle_A$ and $|1\rangle_B$ . Why: rank one leaves no room for correlation — the state literally factorizes, so there is nothing to entangle.

Step 3. Spectral restatement: $r=1$ is equivalent to $\rho_A=|1\rangle_A\langle 1|$ being a pure state, which is equivalent to $S(\rho_A)=0$ . Why: a pure reduced state means the subsystem is in a definite state on its own — no ignorance, no entanglement — so vanishing entanglement entropy is the same certificate wearing a thermodynamic hat.

Limiting cases

The extreme values of the Schmidt spectrum pin down the qualitative entanglement structure, from none to maximal to the tripartite breakdown:

Case	Result	Meaning
Product state	$r=1,\ \lambda_1=1$	A single Schmidt term $\|\psi\rangle=\|a\rangle_A\|b\rangle_B$ ; both reduced states are pure; entanglement entropy zero. The separable case for pure states.
Maximally entangled	$\lambda_i=1/\sqrt{d}\ \forall i,\ r=d$	Flat spectrum on $d=\min(d_A,d_B)$ terms; reduced states are maximally mixed $\rho_A=\mathbb 1/d$ ; entropy $\log_2 d$ ebits. Bell states are the $d=2$ instance.
Generic two qubits	$\lambda_1^2+\lambda_2^2=1,\ \lambda_1\ge\lambda_2\ge 0$	Two coefficients fix the state up to local unitaries; concurrence $2\lambda_1\lambda_2$ , entropy $H(\lambda_1^2)$ with $H$ the binary entropy.
Tripartite breakdown	$\|\mathrm{GHZ}\rangle=\tfrac{1}{\sqrt 2}(\|000\rangle+\|111\rangle)$	No three-party basis diagonalizes a generic tripartite state; a Schmidt vector exists only across a chosen cut such as $A$ versus $BC$ .

Common trapThinking the Schmidt decomposition works for any state

The decomposition is guaranteed only for bipartite pure states, and misapplying it is a common error. For a mixed state $\rho$ there is no Schmidt form — diagonalize $\rho$ by its eigendecomposition instead, and remember a mixed state can be entangled even with full-rank marginals, so no single rank certifies it. For three or more subsystems there is in general no choice of local bases that simultaneously diagonalizes the state: you can Schmidt-decompose across a bipartition (say $A$ versus $BC$ ) and get a valid spectrum for that cut, but the three cuts of a GHZ or W state cannot all be diagonalized together — which is exactly why multipartite entanglement splits into inequivalent classes. And even in the bipartite pure case the Schmidt bases are non-unique when singular values are degenerate, though the multiset of coefficients is always unique and is the true invariant.

Worked examples

1 · Full Schmidt decomposition of an unnormalized-looking state (fully worked)

Decompose $|\psi\rangle=\tfrac{1}{\sqrt 3}(|00\rangle+|01\rangle+|10\rangle)$ . Find its Schmidt coefficients, rank, entanglement entropy, and concurrence.

Step 1. Build the amplitude matrix with rows indexed by the first qubit ( $A$ ) and columns by the second ( $B$ ). The amplitudes are $C_{00}=C_{01}=C_{10}=1/\sqrt 3$ and $C_{11}=0$ , so $C=\tfrac{1}{\sqrt 3}\begin{pmatrix}1&1\\1&0\end{pmatrix}$ . This matrix is not rank one — its determinant is $\tfrac{1}{3}(1\cdot 0-1\cdot 1)=-\tfrac{1}{3}\neq 0$ — so we already expect entanglement.

Step 2. Form the reduced state $\rho_A=CC^\dagger=\tfrac{1}{3}\begin{pmatrix}2&1\\1&1\end{pmatrix}$ . Its eigenvalues solve $\mu^2-\mu+\tfrac{1}{9}=0$ , giving $\mu_\pm=\tfrac{3\pm\sqrt 5}{6}$ , i.e. $\lambda_1^2=\mathbf{0.8727}$ and $\lambda_2^2=\mathbf{0.1273}$ . They sum to $1$ , as a reduced spectrum must.

Step 3. The Schmidt coefficients are the square roots: $\lambda_1=\mathbf{0.934}$ and $\lambda_2=\mathbf{0.357}$ . Both are nonzero, so the Schmidt rank is 2 — the state is entangled, confirming Step 1.

Step 4. Entanglement entropy: $E=-0.8727\log_2 0.8727-0.1273\log_2 0.1273=\mathbf{0.550\ \text{ebit}}$ . This sits well below the maximal $1$ ebit of a Bell state — the state is entangled but only partially.

Step 5. Concurrence: $C(\psi)=2\lambda_1\lambda_2=2(0.934)(0.357)=\mathbf{0.667}=2/3$ . Cross-check with the two-qubit shortcut $2|ad-bc|$ with $a=b=c=1/\sqrt 3,\ d=0$ : $2|0-\tfrac13|=2/3$ . The two routes agree, verifying the decomposition end to end.

2 · Reading everything off a diagonal state, and Nielsen convertibility

Take $|\psi\rangle=\tfrac{1}{2}(\sqrt 3\,|00\rangle+|11\rangle)$ . Give its Schmidt data, and decide whether LOCC can turn it into a Bell state.

Step 1. The state is already in Schmidt form: the coefficients are $\lambda_1=\sqrt 3/2\approx\mathbf{0.866}$ and $\lambda_2=1/2$ , with $\lambda_1^2+\lambda_2^2=\tfrac34+\tfrac14=1$ . The Schmidt weights are $\{0.75,0.25\}$ and the rank is $2$ .

Step 2. The reduced state is $\rho_A=\mathrm{diag}(3/4,1/4)$ , so the entanglement entropy is the binary entropy $H(3/4)=-\tfrac34\log_2\tfrac34-\tfrac14\log_2\tfrac14=\mathbf{0.811\ \text{ebit}}$ , and the concurrence is $2\lambda_1\lambda_2=2\cdot\tfrac{\sqrt3}{2}\cdot\tfrac12=\sqrt3/2\approx\mathbf{0.866}$ .

Step 3. Nielsen's theorem: $|\psi\rangle\to|\phi\rangle$ deterministically by LOCC iff the Schmidt-weight vector of $|\psi\rangle$ is majorized by that of $|\phi\rangle$ . Our vector is $(0.75,0.25)$ ; the Bell target is $(0.5,0.5)$ . The leading partial sums give $0.75\not\le 0.5$ , so $(0.75,0.25)$ is not majorized by $(0.5,0.5)$ — LOCC cannot produce a Bell state from this state.

Step 4. The reverse works: $(0.5,0.5)$ is majorized by $(0.75,0.25)$ (partial sums $0.5\le 0.75$ ), so a shared Bell pair can be deterministically converted into $|\psi\rangle$ . Maximal entanglement is the universal source — it flows down to any state of the same dimension, never up, which is the operational meaning of "maximally entangled".

Physics of the portrait

The decomposition carries the name of a mathematician who never thought about qubits. Erhard Schmidt introduced the bi-orthogonal expansion in 1907 as a tool for integral equations; it took the arrival of quantum entanglement, three decades later, to reveal what he had really found.

Origins · 1 minErhard Schmidt and how this was discovered

Erhard Schmidt

1876–1959 · functional analysis and the theory of integral equations

Erhard Schmidt was a German mathematician, a student of David Hilbert at Göttingen, and one of the founders of modern functional analysis — the Gram–Schmidt orthogonalization process and Hilbert–Schmidt operators both bear his name. In his 1907 work on linear integral equations he studied kernels $K(x,y)$ and proved that any such kernel admits an expansion in a pair of orthonormal function systems weighted by non-negative numbers — the singular values — which is exactly the singular value decomposition in an infinite-dimensional setting, and exactly the structure later called the Schmidt decomposition. He had no notion of quantum mechanics; his motivation was the spectral theory of operators. The physical significance surfaced only in 1935, when Erwin Schrödinger, analyzing the "entanglement" (his coined Verschränkung) of two quantum systems, used precisely Schmidt's bi-orthogonal form to argue that measuring one system instantly fixes the description of the other. What Schmidt built as a theorem about kernels became the mathematical skeleton of quantum correlation — the right structure discovered decades before the phenomenon it describes.

History

1907

— Erhard Schmidt gives the bi-orthogonal expansion of an integral-equation kernel by its singular values

1935

— Schrödinger uses the same bi-orthogonal form to analyze entangled (verschränkt) states

1935

— Einstein, Podolsky and Rosen sharpen the paradox the decomposition later tames

1996

— Bennett, Bernstein, Popescu and Schumacher fix the reduced entropy as the pure-state entanglement measure

1999

— Nielsen recasts LOCC state conversion as majorization of Schmidt vectors

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §2.5 (Schmidt decomposition and purifications) · Preskill, Quantum Information lecture notes, Ch. 4 (entanglement and the Schmidt form) · Watrous, The Theory of Quantum Information, §2.2 (operator–vector correspondence and the SVD) · Wilde, Quantum Information Theory, §3.6 (Schmidt decomposition) · Kitaev, Shen & Vyalyi, Classical and Quantum Computation, §11 (entanglement) · Mermin, Quantum Computer Science, Ch. 1. The two-qubit concurrence $2\lambda_1\lambda_2$ is Wootters' formula specialized to pure states. Key numerical results: the golden-ratio spectrum $\{(3\pm\sqrt5)/6\}=\{0.8727,0.1273\}$ with entropy $0.550$ ebit and concurrence $2/3$ ; the diagonal state with weights $\{0.75,0.25\}$ , entropy $H(3/4)=0.811$ ebit, and concurrence $\sqrt3/2$ ; the Bell value $S=1$ ebit; and the Nielsen majorization verdicts.

Prerequisites. Separable versus Entangled States · Bell States and EPR Pairs · Density Matrix and Partial Trace. Leads to. Entanglement Entropy of Pure States · Concurrence and Wootters' Formula · Entanglement of Formation · PPT Criterion and the Peres–Horodecki Test.

Ask the tutor: Give me any two-qubit or qudit state and I will walk you through its amplitude matrix, its Schmidt coefficients, and exactly what its rank and entropy say about how entangled it is.

Is the Schmidt decomposition really just the SVD?

Yes — literally, not by analogy. Write the bipartite pure state in any fixed product basis and its amplitudes form a rectangular matrix $C$ . The singular value decomposition $C=U\Sigma V^\dagger$ rotates the row basis (a unitary on $A$ ) and column basis (a unitary on $B$ ) until the matrix is diagonal with non-negative entries. Those entries are the Schmidt coefficients, the columns of $U$ and $V$ are the Schmidt bases, and the number of nonzero singular values is the Schmidt rank. Every property — existence, uniqueness of the coefficients, invariance under local unitaries — is inherited directly from the SVD. If you can compute an SVD, you can Schmidt-decompose a state.

Why must the two halves of an entangled pair be equally mixed?

Because both reduced density matrices are built from the same Schmidt weights. Tracing out $B$ gives $\rho_A=\sum_i\lambda_i^2|i\rangle_A\langle i|$ , and tracing out $A$ gives $\rho_B=\sum_i\lambda_i^2|i\rangle_B\langle i|$ — identical eigenvalues $\lambda_i^2$ , just on different bases. Their entropies are therefore automatically equal, so $A$ is exactly as uncertain about its own state as $B$ is about its. This holds even when the subsystems differ in size: if $A$ is a qutrit and $B$ a qubit, $\rho_A$ is $3\times 3$ but simply carries an extra zero eigenvalue, leaving the nonzero spectrum — and the entanglement — identical. It is the precise sense in which entanglement is a symmetric, shared property of the pair.

What is the Schmidt rank good for beyond detecting entanglement?

It is a full resource counter, not just a yes/no flag. Rank $1$ means product; rank $r$ (the "Schmidt number") means the state is at least as entangled as a maximally entangled state of dimension $r$ . In quantum communication the Schmidt rank of a shared state bounds how much high-dimensional entanglement two parties hold. In classical simulation of circuits it is the bond dimension across a cut — a circuit is efficiently simulable while its rank across every cut stays polynomial, and becomes hard when the rank grows exponentially. In tensor networks it is the number of parameters you must keep on a bond to represent the state faithfully.

Can I Schmidt-decompose a three-qubit GHZ state?

Not across all three parties at once. You can always decompose across a bipartition — group the three qubits into two blocks, say $A$ versus $BC$ , and take the Schmidt form of that cut; for GHZ $\tfrac{1}{\sqrt2}(|000\rangle+|111\rangle)$ every such cut gives coefficients $\{1/\sqrt2,1/\sqrt2\}$ . What you cannot do is find three local bases that diagonalize the state simultaneously — a genuine three-party Schmidt form with one index linking all subsystems does not exist for generic tripartite states. That obstruction is exactly why multipartite entanglement fragments into inequivalent classes (GHZ, W, and beyond) that no single vector of coefficients can capture.

When are the Schmidt bases non-unique, and what stays fixed?

The coefficients $\{\lambda_i\}$ are always unique as a sorted multiset — they are the singular values, which any matrix determines uniquely. The bases can fail to be unique when singular values are degenerate: if two coefficients are equal, you may apply the same rotation inside their degenerate subspace on both sides without changing the state, giving a continuous family of valid Schmidt bases. The extreme case is a maximally entangled state, where all coefficients are equal and any orthonormal basis on $A$ (with the matched basis on $B$ ) works. This never touches the entanglement content, since every measure depends only on the invariant coefficients.

How does the Schmidt spectrum relate to purification?

Purification runs the decomposition backwards. Given any mixed state $\rho_A=\sum_i p_i|i\rangle\langle i|$ on $A$ , define a pure state $|\psi\rangle_{AB}=\sum_i\sqrt{p_i}\,|i\rangle_A|i\rangle_B$ on an enlarged system; tracing out $B$ recovers $\rho_A$ exactly. This $|\psi\rangle$ is already in Schmidt form with coefficients $\lambda_i=\sqrt{p_i}$ , so the eigenvalues of the mixed state you started from are the squared Schmidt coefficients of its purification, and the minimum ancilla dimension is the rank of $\rho_A$ . Purifications built this way underlie Stinespring dilation of channels, entanglement of formation, and the modelling of decoherence as entanglement with the environment.