Simple Algorithm
Quantum Computing · Rydberg State · 28 min

The Partial Trace, Rigorously

Reduced states, and the only consistent way to ignore a subsystem.

Every time you look at only part of a quantum system, you are taking a partial trace. It is not one choice among many: it is the unique map that predicts every local measurement correctly, and its properties (complete positivity, no-signalling, contractivity) are exactly what make "subsystem" a meaningful word.

What you’ll learn

  • Define the partial trace, prove it is basis-independent, and prove it is the unique map reproducing all local expectation values.
  • Show that Tr⁡B\Tr_B is a quantum channel with Kraus operators I⊗⟨j∣I\otimes\bra{j}, and derive no-signalling and contractivity from it.
  • Compute reduced states of two-qubit states, including ∣r⃗A∣=1−C2\abs{\vec r_A} = \sqrt{1-\mathcal{C}^2} for pure states.
  • State the unitary freedom of purifications and its consequence, remote steering of ensembles.

Definition and uniqueness

DerivationThe formula does not depend on the basis
  1. ∑j′(I⊗⟨j′∣)M(I⊗∣j′⟩)=∑k,l(∑j′Ulj′Ukj′‾)(I⊗⟨k∣)M(I⊗∣l⟩)\sum_{j'}\big(I\otimes\bra{j'}\big)M\big(I\otimes\ket{j'}\big) = \sum_{k,l}\Big(\sum_{j'}U_{lj'}\overline{U_{kj'}}\Big)\big(I\otimes\bra{k}\big)M\big(I\otimes\ket{l}\big)

    Let ∣j′⟩=∑lUlj′∣l⟩\ket{j'} = \sum_lU_{lj'}\ket{l} be another orthonormal basis, with UU unitary. Substitute:

  2. =∑k(I⊗⟨k∣)M(I⊗∣k⟩)= \sum_k\big(I\otimes\bra k\big)M\big(I\otimes\ket k\big)

    The bracket is (UU†)lk=δlk(UU^\dagger)_{lk} = \delta_{lk}:

  3. ∑j(I⊗⟨j∣)(X⊗Y)(I⊗∣j⟩)=X∑j⟨j|Y|j⟩=X Tr⁡Y\sum_j\big(I\otimes\bra j\big)(X\otimes Y)\big(I\otimes\ket j\big) = X\sum_j\mel{j}{Y}{j} = X\,\Tr Y

    On product operators the formula gives XTr⁡YX\Tr Y, so it is the map defined above, and linearity fixes it everywhere.

A channel, and no-signalling

Write Kj=I⊗⟨j∣K_j = I\otimes\bra{j}, an operator from HA⊗HB\Hil_A\otimes\Hil_B to HA\Hil_A. Then Tr⁡BM=∑jKjMKj†\Tr_BM = \sum_jK_jMK_j^\dagger and ∑jKj†Kj=I⊗∑j∣j⟩ ⁣⟨j∣=I\sum_jK_j^\dagger K_j = I\otimes\sum_j\dyad{j} = I. So the partial trace is a quantum channel in Kraus form: completely positive and trace preserving.

DerivationProof
  1. Tr⁡B[(id⊗Φ)(X⊗Y)]=Tr⁡B[X⊗Φ(Y)]=X Tr⁡ Φ(Y)\Tr_B\big[(\mathrm{id}\otimes\Phi)(X\otimes Y)\big] = \Tr_B\big[X\otimes\Phi(Y)\big] = X\,\Tr\,\Phi(Y)

    Product operators X⊗YX\otimes Y span L(HA⊗HB)\mathcal{L}(\Hil_A\otimes\Hil_B), and both sides are linear, so check them there:

  2. =X Tr⁡Y=Tr⁡B(X⊗Y)= X\,\Tr Y = \Tr_B(X\otimes Y)

    Φ\Phi preserves the trace:

  3. Only trace preservation of Φ\Phi was used. The statement therefore holds for every trace-preserving linear map on B; complete positivity is needed only for Φ\Phi to be physical in the first place.

Check your understanding

ρAB=∣Φ+⟩ ⁣⟨Φ+∣\rho_{AB} = \dyad{\Phi^+}. Bob measures his qubit in the XX basis and does not tell Alice the result. What is Alice's state now?

  1. ∣+⟩ ⁣⟨+∣\dyad{+}
  2. I/2I/2
  3. ∣0⟩ ⁣⟨0∣\dyad{0}
  4. A state that depends on Bob's outcome, so Alice can detect that he measured
Show the answer

B. I/2I/2

Measuring and forgetting is a channel on B, so ρA=Tr⁡BρAB=I/2\rho_A = \Tr_B\rho_{AB} = I/2 is unchanged. Conditioned on Bob's outcome, Alice holds ∣+⟩\ket+ or ∣−⟩\ket- with probability ½ each, but without the outcome that ensemble is again I/2I/2.

Reduced states in practice

For two qubits, expand in Paulis: ρAB=14(I⊗I+r⃗⋅σ⃗⊗I+I⊗s⃗⋅σ⃗+∑ijTij σi⊗σj)\rho_{AB} = \tfrac14\big(I\otimes I+\vec r\cdot\vec\sigma\otimes I+I\otimes\vec s\cdot\vec\sigma+\sum_{ij}T_{ij}\,\sigma_i\otimes\sigma_j\big). Since Tr⁡σj=0\Tr\sigma_j = 0, the partial trace keeps exactly the terms with nothing on B:

Tr⁡B ρAB=12(I+r⃗⋅σ⃗).\Tr_B\,\rho_{AB} = \tfrac12\big(I+\vec r\cdot\vec\sigma\big).

The local Bloch vectors r⃗\vec r and s⃗\vec s are what each party sees on their own; the correlation tensor TT is invisible to both separately.

DerivationPure states: how far the Bloch vector shrinks
  1. ρA=λ12∣a1⟩ ⁣⟨a1∣+λ22∣a2⟩ ⁣⟨a2∣,∣r⃗A∣=∣λ12−λ22∣\rho_A = \lambda_1^2\dyad{a_1}+\lambda_2^2\dyad{a_2},\qquad \abs{\vec r_A} = \abs{\lambda_1^2-\lambda_2^2}

    Take a pure two-qubit state in Schmidt form, λ1∣a1⟩∣b1⟩+λ2∣a2⟩∣b2⟩\lambda_1\ket{a_1}\ket{b_1}+\lambda_2\ket{a_2}\ket{b_2} (proved in The Schmidt Decomposition). Its marginal is diagonal in the orthonormal basis ∣a1⟩,∣a2⟩\ket{a_1},\ket{a_2}:

  2. ∣r⃗A∣2=(λ12+λ22)2−4λ12λ22=1−C2\abs{\vec r_A}^2 = (\lambda_1^2+\lambda_2^2)^2-4\lambda_1^2\lambda_2^2 = 1-\mathcal{C}^2

    Use λ12+λ22=1\lambda_1^2+\lambda_2^2 = 1, and write the concurrence as C=2λ1λ2=2∣det⁡C∣\mathcal{C} = 2\lambda_1\lambda_2 = 2\abs{\det C} (the singular values of the coefficient matrix are the Schmidt coefficients):

  3. ∣r⃗A∣=∣r⃗B∣=1−C2\boxed{\abs{\vec r_A} = \abs{\vec r_B} = \sqrt{1-\mathcal{C}^2}}

    Both local Bloch vectors shrink by the same amount, to zero exactly for maximally entangled states. This is the "short arrow means entangled" rule of the circuit widget's Bloch spheres.

Predict before you look

What is the Bloch-vector length of either qubit of 13(∣00⟩+∣01⟩+∣10⟩)\tfrac{1}{\sqrt3}(\ket{00}+\ket{01}+\ket{10})?

Show the answer

0.75

det⁡C=13(1⋅0−1⋅1)=−13\det C = \tfrac13(1\cdot0-1\cdot1) = -\tfrac13, so C=23\mathcal C = \tfrac23 and ∣r⃗∣=1−49=53≈0.745\abs{\vec r} = \sqrt{1-\tfrac49} = \tfrac{\sqrt5}{3}\approx0.745. Directly: ρA=13(2111)\rho_A = \tfrac13\begin{pmatrix}2&1\\1&1\end{pmatrix} has r⃗=(23,0,13)\vec r = (\tfrac23,0,\tfrac13).

Purifications and steering

A striking consequence, noticed by Schrödinger in 1935–36 and made precise by Hughston, Jozsa and Wootters (HJW): if Alice holds A and Bob holds a purification R of ρA\rho_A, then for every ensemble decomposition ρA=∑kqk∣ϕk⟩ ⁣⟨ϕk∣\rho_A = \sum_kq_k\dyad{\phi_k} there is a measurement on R (in general a POVM) whose outcome kk occurs with probability qkq_k and leaves Alice in ∣ϕk⟩\ket{\phi_k}. Bob can steer which ensemble Alice holds, but not which member she gets, and her average state stays ρA\rho_A, as no-signalling demands. (Ensembles and the HJW theorem are developed fully in Ensembles, Purification & the HJW Theorem.)

Key takeaways

  • Tr⁡B\Tr_B is the linear map with Tr⁡B(X⊗Y)=X Tr⁡Y\Tr_B(X\otimes Y) = X\,\Tr Y; it is basis-independent and the unique map satisfying Tr⁡[M Tr⁡Bρ]=Tr⁡[(M⊗I)ρ]\Tr[M\,\Tr_B\rho] = \Tr[(M\otimes I)\rho] for all observables MM.
  • It is a channel with Kraus operators I⊗⟨j∣I\otimes\bra j; consequently nothing done on B alone changes ρA\rho_A (no-signalling), and trace distance can only decrease.
  • For two qubits the partial trace keeps the local Bloch vector and discards the correlation tensor; for pure states ∣r⃗A∣=1−C2\abs{\vec r_A} = \sqrt{1-\mathcal C^2}.
  • Purifications exist and are unique up to a unitary on the purifier; holding a purification lets Bob steer Alice's ensemble, never her average state.

Check yourself

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

Question 1

What is Tr⁡B(∣0⟩ ⁣⟨1∣⊗∣+⟩ ⁣⟨−∣)\Tr_B\big(\ketbra{0}{1}\otimes\ketbra{+}{-}\big)?

  1. ∣0⟩ ⁣⟨1∣\ketbra01
  2. 12∣0⟩ ⁣⟨1∣\tfrac12\ketbra01
  3. I/2I/2
  4. 00
Show the answer

D. 00

Tr⁡B(X⊗Y)=X Tr⁡Y\Tr_B(X\otimes Y) = X\,\Tr Y, and Tr⁡∣+⟩ ⁣⟨−∣=⟨−∣+⟩=0\Tr\ketbra{+}{-} = \braket{-|+} = 0.

Question 2

For a pure state ∣ψ⟩AB\ket\psi_{AB}, the reduced state ρA\rho_A is pure if and only if:

  1. ∣ψ⟩\ket\psi is a Bell state
  2. ∣ψ⟩\ket\psi is a product state
  3. dim⁡HA=dim⁡HB\dim\Hil_A = \dim\Hil_B
  4. ρB\rho_B is maximally mixed
Show the answer

B. ∣ψ⟩\ket\psi is a product state

ρA\rho_A is pure iff it has a single nonzero eigenvalue, iff the Schmidt rank is 1, iff ∣ψ⟩=∣a⟩⊗∣b⟩\ket\psi = \ket a\otimes\ket b.

Question 3

∣ψ⟩\ket\psi and ∣ψ′⟩\ket{\psi'} in HA⊗HR\Hil_A\otimes\Hil_R purify the same ρA\rho_A. Then necessarily:

  1. ∣ψ′⟩=(U⊗I)∣ψ⟩\ket{\psi'} = (U\otimes I)\ket\psi for a unitary UU on A
  2. ∣ψ′⟩=∣ψ⟩\ket{\psi'} = \ket\psi
  3. ∣ψ′⟩=(I⊗U)∣ψ⟩\ket{\psi'} = (I\otimes U)\ket\psi for a unitary UU on R
  4. ∣ψ′⟩=(U⊗U‾)∣ψ⟩\ket{\psi'} = (U\otimes\overline{U})\ket\psi for some unitary UU
Show the answer

C. ∣ψ′⟩=(I⊗U)∣ψ⟩\ket{\psi'} = (I\otimes U)\ket\psi for a unitary UU on R

All the freedom sits on the purifying system: fixing an eigenbasis of ρA\rho_A fixes the A side of both decompositions, and a unitary on R maps one set of partner vectors to the other.

Question 4

The trace distance between ρAB\rho_{AB} and σAB\sigma_{AB} is 0.30.3. What is the largest possible trace distance between ρA\rho_A and σA\sigma_A?

  1. 0.150.15
  2. 0.60.6
  3. 11
  4. 0.30.3
Show the answer

D. 0.30.3

The partial trace is a channel, and trace distance is contractive under channels, so D(ρA,σA)≤0.3D(\rho_A,\sigma_A)\le0.3. Equality occurs, for example, for ρAB=ρA⊗τ\rho_{AB} = \rho_A\otimes\tau and σAB=σA⊗τ\sigma_{AB} = \sigma_A\otimes\tau.

Further reading

  • M. A. Nielsen & I. L. Chuang, Quantum Computation and Quantum Information, §2.4.3 and §2.5 (Cambridge, 2010).
  • L. P. Hughston, R. Jozsa & W. K. Wootters, "A complete classification of quantum ensembles having a given density matrix", Phys. Lett. A 183, 14 (1993).
  • W. F. Stinespring, "Positive functions on C-algebras", Proc. Amer. Math. Soc.* 6, 211 (1955).
  • J. Watrous, The Theory of Quantum Information (Cambridge, 2018).
  • Z. Ji, A. Natarajan, T. Vidick, J. Wright & H. Yuen, "MIP* = RE", arXiv:2001.04383 (2020).