Simple Algorithm
Quantum Computing · Rydberg State · 26 min

The Schmidt Decomposition

The canonical form of every bipartite pure state.

Any pure state of two systems, however you wrote it down, has a canonical form: a single sum of perfectly paired orthonormal terms. That form, the Schmidt decomposition, is just the singular value decomposition, and it settles almost every question about bipartite pure-state entanglement.

What you’ll learn

  • State and prove the Schmidt decomposition via the SVD, including the uniqueness of the coefficients.
  • Compute Schmidt coefficients, Schmidt bases and the entanglement entropy of a concrete state.
  • Prove that two bipartite pure states are related by local unitaries if and only if they have the same Schmidt coefficients.
  • Explain why no Schmidt decomposition exists for three or more parties in general.

The theorem and its proof

DerivationProof via the singular value decomposition
  1. ∣ψ⟩=∑j,kCjk∣j⟩∣k⟩\ket\psi = \sum_{j,k}C_{jk}\ket{j}\ket{k}

    Fix orthonormal bases and form the coefficient matrix C∈Cm×pC\in\C^{m\times p}, Cjk=(⟨j∣⊗⟨k∣)∣ψ⟩C_{jk} = (\bra j\otimes\bra k)\ket\psi:

  2. Cjk=∑iUji σi Vki‾C_{jk} = \sum_iU_{ji}\,\sigma_i\,\overline{V_{ki}}

    Take an SVD C=UΣV†C = U\Sigma V^\dagger, with UU, VV unitary and Σ\Sigma the m×pm\times p diagonal matrix of singular values σ1≥σ2≥⋯≥0\sigma_1\ge\sigma_2\ge\cdots\ge0:

  3. ∣ψ⟩=∑iσi(∑jUji∣j⟩)⊗(∑kVki‾∣k⟩)≡∑iσi ∣ai⟩⊗∣bi⟩\ket\psi = \sum_i\sigma_i\Big(\sum_jU_{ji}\ket j\Big)\otimes\Big(\sum_k\overline{V_{ki}}\ket k\Big)\equiv\sum_i\sigma_i\,\ket{a_i}\otimes\ket{b_i}

    Substitute and regroup the sums:

  4. ⟨ai∣ai′⟩=(U†U)ii′=δii′,⟨bi∣bi′⟩=(V†V)ii′‾=δii′\braket{a_i|a_{i'}} = (U^\dagger U)_{ii'} = \delta_{ii'},\qquad \braket{b_i|b_{i'}} = \overline{(V^\dagger V)_{ii'}} = \delta_{ii'}

    Orthonormality is inherited from unitarity. Keep only the terms with σi>0\sigma_i>0: then r=rank⁡Cr = \rank C and λi=σi\lambda_i = \sigma_i.

  5. ρA=CC†=∑iλi2∣ai⟩ ⁣⟨ai∣,ρB=CTC‾=∑iλi2∣bi⟩ ⁣⟨bi∣\rho_A = CC^\dagger = \sum_i\lambda_i^2\dyad{a_i},\qquad \rho_B = C^{T}\overline{C} = \sum_i\lambda_i^2\dyad{b_i}

    Uniqueness: from any decomposition of this form, ρA=∑iλi2∣ai⟩ ⁣⟨ai∣\rho_A = \sum_i\lambda_i^2\dyad{a_i}, so the λi2\lambda_i^2 are the nonzero eigenvalues of ρA\rho_A (with multiplicity), which depend on no choices. Normalization: ∑iλi2=Tr⁡CC†=⟨ψ∣ψ⟩\sum_i\lambda_i^2 = \Tr CC^\dagger = \braket{\psi|\psi}.

A worked example

Take ∣ψ⟩=13(∣00⟩+∣01⟩+∣10⟩)\ket\psi = \tfrac{1}{\sqrt3}(\ket{00}+\ket{01}+\ket{10}), so C=13(1110)C = \tfrac{1}{\sqrt3}\begin{pmatrix}1&1\\1&0\end{pmatrix}. Its concurrence is 2∣det⁡C∣=232\abs{\det C} = \tfrac23. How much entanglement is that in ebits?

Predict before you look

Estimate the entanglement entropy S(ρA)S(\rho_A) of this state, in ebits. (A Bell pair has 1, a product state 0.)

Show the answer

0.55 ebit

S=h(3+56)≈0.550S = h\big(\tfrac{3+\sqrt5}{6}\big)\approx0.550, where hh is the binary entropy. Many guess higher because the concurrence is 23\tfrac23; entropy and concurrence are different scales, related monotonically by S=h(1+1−C22)S = h\big(\tfrac{1+\sqrt{1-\mathcal C^2}}{2}\big).

DerivationSchmidt data of the golden-ratio state
  1. λ1=φ3≈0.934,λ2=1φ3≈0.357\lambda_1 = \frac{\varphi}{\sqrt3}\approx0.934,\qquad \lambda_2 = \frac{1}{\varphi\sqrt3}\approx0.357

    CC is real symmetric with eigenvalues 13⋅1±52\tfrac{1}{\sqrt3}\cdot\tfrac{1\pm\sqrt5}{2}; the singular values are their absolute values. The golden ratio φ=1+52\varphi = \tfrac{1+\sqrt5}{2} appears because (1110)\begin{pmatrix}1&1\\1&0\end{pmatrix} is the Fibonacci matrix. Check: λ12+λ22=13(φ2+φ−2)=1\lambda_1^2+\lambda_2^2 = \tfrac13(\varphi^2+\varphi^{-2}) = 1.

  2. ∣a1⟩=∣b1⟩∝φ∣0⟩+∣1⟩,∣a2⟩∝∣0⟩−φ∣1⟩,∣b2⟩=−∣a2⟩\ket{a_1} = \ket{b_1}\propto\varphi\ket0+\ket1,\qquad \ket{a_2}\propto\ket0-\varphi\ket1,\qquad \ket{b_2} = -\ket{a_2}

    For a real symmetric C=∑iμi viviTC = \sum_i\mu_i\,v_iv_i^{T}, take ∣ai⟩=vi\ket{a_i} = v_i and ∣bi⟩=sign⁡(μi) vi\ket{b_i} = \operatorname{sign}(\mu_i)\,v_i: the negative eigenvalue −1/(φ3)-1/(\varphi\sqrt3) is absorbed into ∣b2⟩\ket{b_2}.

  3. λ1,22=3±56≈0.873, 0.127\lambda_{1,2}^2 = \frac{3\pm\sqrt5}{6}\approx0.873,\ 0.127

    The Schmidt probabilities are the eigenvalues of ρA=13(2111)\rho_A = \tfrac13\begin{pmatrix}2&1\\1&1\end{pmatrix}, as they must be:

  4. S(ρA)=−∑iλi2log⁡2λi2≈0.550 ebit,C=2λ1λ2=23S(\rho_A) = -\sum_i\lambda_i^2\log_2\lambda_i^2\approx0.550\ \text{ebit},\qquad \mathcal{C} = 2\lambda_1\lambda_2 = \tfrac23

    About 55% of a Bell pair, by the entropy measure; the product of the coefficients reproduces the concurrence.

What the Schmidt coefficients decide

So for bipartite pure states the Schmidt vector λ⃗\vec\lambda, equivalently the spectrum of ρA\rho_A, is a complete invariant; everything else is a choice of local bases. What it decides:

QuestionAnswer in Schmidt terms
Is ∣ψ⟩\ket\psi entangled?Iff r≥2r\ge2.
Do the two sides look alike?ρA\rho_A and ρB\rho_B always share the nonzero eigenvalues λi2\lambda_i^2, so S(ρA)=S(ρB)S(\rho_A) = S(\rho_B) even when dim⁡HA≠dim⁡HB\dim\Hil_A\neq\dim\Hil_B.
Is it maximally entangled?Iff r=d=min⁡(m,p)r = d = \min(m,p) and every λi=1/d\lambda_i = 1/\sqrt d; for m≤pm\le p this means ρA=I/m\rho_A = I/m.
How much entanglement?Entanglement entropy −∑iλi2log⁡2λi2-\sum_i\lambda_i^2\log_2\lambda_i^2; for two qubits also the concurrence 2λ1λ22\lambda_1\lambda_2.
Can LOCC turn ψ\psi into ϕ\phi?Iff (λi2)ψ(\lambda_i^2)_\psi is majorized by (λi2)ϕ(\lambda_i^2)_\phi: Nielsen's theorem, in LOCC Transformations & Majorization.
Check your understanding

A pure state of a qubit (A) and a qutrit (B) has Schmidt rank at most:

  1. 11
  2. 33
  3. 66
  4. 22
Show the answer

D. 22

r=rank⁡C≤min⁡(2,3)=2r = \rank C\le\min(2,3) = 2. The qutrit can only use a two-dimensional subspace, the support of ρB\rho_B, to be entangled with the qubit.

Beyond two parties, beyond finite dimensions

Does every tripartite state have a form ∑iλi∣ai⟩∣bi⟩∣ci⟩\sum_i\lambda_i\ket{a_i}\ket{b_i}\ket{c_i} with three orthonormal sets? No: such "tri-orthogonal" decompositions exist only for special states (Peres 1995). GHZ already has one; W cannot.

DerivationThe W state has no tripartite Schmidt form
  1. ρAB=∑iλi2 ∣ai⟩ ⁣⟨ai∣⊗∣bi⟩ ⁣⟨bi∣\rho_{AB} = \sum_i\lambda_i^2\,\dyad{a_i}\otimes\dyad{b_i}

    Suppose ∣W⟩=13(∣001⟩+∣010⟩+∣100⟩)=∑iλi∣ai⟩∣bi⟩∣ci⟩\ket W = \tfrac{1}{\sqrt3}(\ket{001}+\ket{010}+\ket{100}) = \sum_i\lambda_i\ket{a_i}\ket{b_i}\ket{c_i} with orthonormal {∣ci⟩}\{\ket{c_i}\}. Tracing out the third qubit kills all cross terms:

  2. ρAB=13∣00⟩ ⁣⟨00∣+23∣Ψ+⟩ ⁣⟨Ψ+∣\rho_{AB} = \tfrac13\dyad{00}+\tfrac23\dyad{\Psi^+}

    That is a mixture of product states, hence separable. But the actual two-qubit marginal of ∣W⟩\ket W is

  3. C(ρAB)=23>0 ⟹ no tri-orthogonal form\mathcal{C}(\rho_{AB}) = \tfrac23>0\ \Longrightarrow\ \text{no tri-orthogonal form}

    which is entangled: its concurrence is 23\tfrac23 (see Quantifying Entanglement). Contradiction.

In infinite dimensions the bipartite theorem survives: a normalized ∣ψ⟩\ket\psi defines a Hilbert–Schmidt operator, whose singular value decomposition gives a countable Schmidt sum with ∑iλi2=1\sum_i\lambda_i^2 = 1. The two-mode squeezed vacuum of quantum optics comes already in Schmidt form,

∣TMSV⟩=1cosh⁡s∑n=0∞(tanh⁡s)n∣n⟩∣n⟩,\ket{\mathrm{TMSV}} = \frac{1}{\cosh s}\sum_{n=0}^{\infty}(\tanh s)^n\ket{n}\ket{n},

with geometric Schmidt probabilities and an entanglement entropy that grows without bound as the squeezing s→∞s\to\infty.

Key takeaways

  • Every bipartite pure state is ∑iλi∣ai⟩∣bi⟩\sum_i\lambda_i\ket{a_i}\ket{b_i} with orthonormal sets; the λi\lambda_i are the singular values of the coefficient matrix, unique and basis-independent.
  • ρA\rho_A and ρB\rho_B share the spectrum {λi2}\{\lambda_i^2\}, so pure-state entanglement is symmetric between the two parties.
  • The Schmidt vector is a complete invariant under local unitaries and decides entanglement, its amount, and LOCC convertibility.
  • No tripartite analogue exists in general (W is a counterexample), and tensor rank, the multipartite generalization of Schmidt rank, is NP-hard.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

With ∣ψ⟩=∑iλi∣ai⟩∣bi⟩\ket\psi = \sum_i\lambda_i\ket{a_i}\ket{b_i} (λi>0\lambda_i>0, orthonormal sets), the λi\lambda_i are:

  1. the eigenvalues of the coefficient matrix CC
  2. the eigenvalues of ρA\rho_A
  3. the singular values of the coefficient matrix CC
  4. the diagonal entries of CC
Show the answer

C. the singular values of the coefficient matrix CC

Schmidt coefficients are the singular values of CC; their squares are the eigenvalues of ρA=CC†\rho_A = CC^\dagger.

Question 2

Two bipartite pure states are related by a local unitary UA⊗UBU_A\otimes U_B if and only if they have the same:

  1. entanglement entropy
  2. Schmidt coefficients
  3. Schmidt rank
  4. reduced state ρA\rho_A, as a matrix
Show the answer

B. Schmidt coefficients

The Schmidt vector is a complete local-unitary invariant, as proved above.

Question 3

Which of these is also a Schmidt decomposition of ∣Φ+⟩\ket{\Phi^+}?

  1. 12(∣+−⟩+∣−+⟩)\tfrac{1}{\sqrt2}(\ket{+-}+\ket{-+})
  2. 12(∣+i⟩∣+i⟩+∣−i⟩∣−i⟩)\tfrac{1}{\sqrt2}(\ket{+i}\ket{+i}+\ket{-i}\ket{-i})
  3. 12(∣0⟩∣+⟩+∣1⟩∣−⟩)\tfrac{1}{\sqrt2}(\ket{0}\ket{+}+\ket{1}\ket{-})
  4. 12(∣++⟩+∣−−⟩)\tfrac{1}{\sqrt2}(\ket{++}+\ket{--})
Show the answer

D. 12(∣++⟩+∣−−⟩)\tfrac{1}{\sqrt2}(\ket{++}+\ket{--})

∣Φ+⟩=12∑i∣ai⟩∣ai‾⟩\ket{\Phi^+} = \tfrac{1}{\sqrt2}\sum_i\ket{a_i}\ket{\overline{a_i}} for any orthonormal basis {∣ai⟩}\{\ket{a_i}\}. The states ∣±⟩\ket\pm are real, so 12(∣++⟩+∣−−⟩)=∣Φ+⟩\tfrac{1}{\sqrt2}(\ket{++}+\ket{--}) = \ket{\Phi^+}.

Question 4

Why does the W state admit no decomposition ∑iλi∣ai⟩∣bi⟩∣ci⟩\sum_i\lambda_i\ket{a_i}\ket{b_i}\ket{c_i} with orthonormal {∣ci⟩}\{\ket{c_i}\}?

  1. Such a form would make its two-qubit marginal separable, but W's marginals are entangled
  2. W has Schmidt rank 3 across every cut
  3. W is not normalized
  4. Three-qubit states never have such a form
Show the answer

A. Such a form would make its two-qubit marginal separable, but W's marginals are entangled

Tracing out C in a tri-orthogonal form leaves ∑iλi2∣ai⟩ ⁣⟨ai∣⊗∣bi⟩ ⁣⟨bi∣\sum_i\lambda_i^2\dyad{a_i}\otimes\dyad{b_i}, which is separable, whereas W's two-qubit marginals have concurrence 23\tfrac23.

Further reading

  • E. Schmidt, "Zur Theorie der linearen und nichtlinearen Integralgleichungen", Math. Ann. 63, 433 (1907).
  • A. Ekert & P. L. Knight, "Entangled quantum systems and the Schmidt decomposition", Am. J. Phys. 63, 415 (1995).
  • M. A. Nielsen & I. L. Chuang, Quantum Computation and Quantum Information, §2.5 (Cambridge, 2010).
  • A. Peres, "Higher order Schmidt decompositions", Phys. Lett. A 202, 16 (1995).
  • A. Acín, A. Andrianov, L. Costa, E. Jané, J. I. Latorre & R. Tarrach, "Generalized Schmidt decomposition and classification of three-quantum-bit states", Phys. Rev. Lett. 85, 1560 (2000).