Skip to content

Partial trace extends from finite matrices to a norm-one map between trace-class spaces. Its rigorous characterization is preservation of all bounded local-observable expectations. The canonical Partial Trace article supplies the subsystem interpretation and index calculations; this page establishes existence, convergence, and the infinite-dimensional limitations of those formulas.

Helpful background. Trace-class operators supplies ideal estimates; trace-norm duality supplies uniqueness and continuity.

Let HA,HB\mathcal H_A,\mathcal H_B be separable Hilbert spaces. There is a unique bounded linear map

Tr⁡B:S1(HA⊗HB)⟶S1(HA)\operatorname{Tr}_B: \mathcal S_1(\mathcal H_A\otimes\mathcal H_B) \longrightarrow\mathcal S_1(\mathcal H_A)

such that, for all bounded AA,

Tr⁡(ATr⁡BT)=Tr⁡((A⊗IB)T).\operatorname{Tr}\bigl(A\operatorname{Tr}_BT\bigr) =\operatorname{Tr}\bigl((A\otimes I_B)T\bigr).

It is positive, trace preserving, completely positive, and contractive in trace norm. Hence it sends density operators to density operators. Complete positivity follows by applying the same positive construction with an additional untouched tensor factor; it is stronger than positivity alone.

The identity IBI_B is bounded even when HB\mathcal H_B is infinite-dimensional. It need not be trace class: only the global operator TT must lie in the trace ideal in this formula.

Choose an orthonormal basis (bj)(b_j) of HB\mathcal H_B and let Vj:HA→HA⊗HBV_j:\mathcal H_A\to\mathcal H_A\otimes\mathcal H_B be the isometry Vjx=x⊗bjV_jx=x\otimes b_j. For T≥0T\geq0 trace class, define

RN=∑j=1NVj∗TVj.R_N=\sum_{j=1}^{N}V_j^*TV_j.

Every term is positive trace class. Using a product orthonormal basis and nonnegative sums,

∑jTr⁡(Vj∗TVj)=Tr⁡T.\sum_j\operatorname{Tr}(V_j^*TV_j)=\operatorname{Tr}T.

Therefore RNR_N is Cauchy in trace norm: the norm of any positive tail is its trace. Its limit RR is positive and has trace Tr⁡T\operatorname{Tr}T.

For arbitrary trace-class TT, decompose its real and imaginary parts into differences of positive trace-class operators. The same construction extends linearly. In fact the series ∑jVj∗TVj\sum_jV_j^*TV_j is absolutely convergent in trace norm, as follows by bounding its terms using those four positive parts.

To verify the characterizing identity, pair the finite sums with bounded AA. The sum on the global space is (A⊗PN)T(A\otimes P_N)T, where PNP_N projects onto the first NN vectors bjb_j. Since A⊗PNA\otimes P_N converges strongly and stays uniformly bounded, its trace pairing with TT converges. Thus RR satisfies the defining formula.

That formula determines RR uniquely by trace duality, so the construction is independent of the basis chosen in HB\mathcal H_B. It is the trace-class convergence, rather than a purely symbolic contraction of indices, that justifies the infinite sum.

Contractivity and local indistinguishability

Section titled “Contractivity and local indistinguishability”

Trace duality and ∥A⊗IB∥=∥A∥\|A\otimes I_B\|=\|A\| give

∥Tr⁡BT∥1=sup⁡∥A∥≤1∣Tr⁡((A⊗IB)T)∣≤∥T∥1.\begin{aligned} \|\operatorname{Tr}_BT\|_1 &=\sup_{\|A\|\leq1} |\operatorname{Tr}((A\otimes I_B)T)|\\ &\leq\|T\|_1. \end{aligned}

Thus tracing out a subsystem cannot increase the trace distance of two states. Any local measurement is also a global measurement with effect EA⊗IBE_A\otimes I_B, so this inequality matches operational intuition. It does not say that partial trace preserves all distinctions: orthogonal global states can have identical reduced states.

For a trace-class product T=A1⊗B1T=A_1\otimes B_1, the result is A1Tr⁡B1A_1\operatorname{Tr}B_1. By contrast, if HB\mathcal H_B is infinite-dimensional, A1⊗IBA_1\otimes I_B is generally not trace class and has no finite partial trace. One cannot assign Tr⁡BIAB\operatorname{Tr}_B I_{AB} a finite operator value by analogy with finite dimension.

Let (an)(a_n) and (bn)(b_n) be orthonormal families and pn≥0p_n\geq0 with ∑npn=1\sum_np_n=1. The vector

∣Ψ⟩=∑npn ∣an⟩⊗∣bn⟩|\Psi\rangle=\sum_n\sqrt{p_n}\,|a_n\rangle\otimes|b_n\rangle

converges in Hilbert-space norm. Applying the construction to its rank-one projector gives

Tr⁡B∣Ψ⟩⟨Ψ∣=∑npn∣an⟩⟨an∣\operatorname{Tr}_B|\Psi\rangle\langle\Psi| =\sum_np_n|a_n\rangle\langle a_n|

in trace norm. This is an infinite-rank reduced state even though the global projector has rank one. Choosing a slowly decreasing summable pnp_n can produce infinite reduced entropy; partial trace does not ensure entropy finiteness.

For an unbounded local observable HA≥0H_A\geq0, apply the bounded identity to spectral truncations and take a monotone limit. The global and reduced energy expectations agree as extended nonnegative numbers, possibly infinity. For a signed observable, require absolute integrability before claiming a finite equality.

  1. Let ρAB\rho_{AB} and σAB\sigma_{AB} have trace distance at most ϵ\epsilon. Bound the probability difference for an arbitrary local effect EAE_A.
Solution

The global effect EA⊗IBE_A\otimes I_B lies between zero and identity. Its probability difference is at most ϵ\epsilon. Equivalently, partial-trace contractivity bounds the reduced trace distance by ϵ\epsilon. This connects the theorem to local measurement statistics.

  1. Let A≥0A\geq0 be a nonzero trace-class operator and let BB have infinite dimension. Explain why the basis partial sums for A⊗IBA\otimes I_B diverge.
Solution

Every contraction Vj∗(A⊗IB)VjV_j^*(A\otimes I_B)V_j equals AA, so the first NN terms sum to NANA. Their trace norms equal NTr⁡A→∞N\operatorname{Tr}A\to\infty. The trace-class hypothesis of the theorem fails.

  • Holevo, Alexander S., Quantum Systems, Channels, Information: A Mathematical Introduction, 2nd revised and expanded ed., De Gruyter, 2019, Chapters 3 and 11 — tensor products, reduced states, and infinite systems. Publisher.
  • Simon, Barry, Trace Ideals and Their Applications, 2nd ed., Mathematical Surveys and Monographs 120, American Mathematical Society, 2005 — trace ideal estimates and convergence used in the construction. Publisher.