Reduced States and the Partial Trace
Statement
Given a joint state \(\rho_{AB}\) on a tensor-product Hilbert space \(\mathcal{H}_A \otimes \mathcal{H}_B\), there exists a unique operator \(\rho_A\) on \(\mathcal{H}_A\) such that \(\operatorname{Tr}_A[M_A\,\rho_A] = \operatorname{Tr}_{AB}\!\big[(M_A \otimes I_B)\,\rho_{AB}\big]\) for every local observable \(M_A\). That operator is the partial trace \(\rho_A = \operatorname{Tr}_B[\rho_{AB}]\), and it reproduces the Born-rule statistics of every measurement performed on subsystem \(A\) alone.
Why it matters
Once two systems are entangled, no wavefunction of \(A\) by itself exists — yet an experimenter with access only to \(A\) must still get definite probabilities for every measurement they can do. The partial trace is exactly the object that packages those local probabilities. It is the bridge between the global pure-state description and the local, generally mixed, description that any real, spatially bounded laboratory actually sees.
It is also the engine of decoherence and of entanglement entropy: tracing out an unobserved environment turns coherent superpositions into apparent classical mixtures, and the entropy of \(\rho_A\) measures how entangled the two halves are. Almost every statement about open quantum systems begins here.
Assumptions
Derivation
Result
Reading. The reduced density matrix is obtained by "summing over" — tracing out — everything belonging to \(B\). It is the one and only operator on \(\mathcal{H}_A\) that reproduces the Born-rule statistics of every local observable \(M_A\otimes I_B\). For an entangled \(\rho_{AB}\) it is generically a mixed state even when \(\rho_{AB}\) is pure: entanglement, viewed locally, looks like classical uncertainty.
Units check. Density operators and the trace are dimensionless (\(\operatorname{Tr}\rho=1\), a pure number). Contracting the \(B\) indices removes the \(\dim\mathcal{H}_B\) factors of the joint operator, leaving an operator on \(\mathcal{H}_A\) with \(\operatorname{Tr}_A\rho_A=1\) — dimensionally and normalisation-wise a valid state. Both sides of \(\operatorname{Tr}_A[M_A\rho_A]=\operatorname{Tr}_{AB}[(M_A\otimes I_B)\rho_{AB}]\) carry the units of \(M_A\).
Limiting cases
- Product state \(\rho_{AB}=\rho_A\otimes\rho_B\): \(\operatorname{Tr}_B[\rho_A\otimes\rho_B]=\rho_A\operatorname{Tr}[\rho_B]=\rho_A\). No information is lost by reduction.
- Pure product \(|\psi\rangle_A\otimes|\chi\rangle_B\): \(\rho_A=|\psi\rangle\langle\psi|\) stays pure — zero entanglement.
- Maximally entangled (Bell) state: \(\rho_A=I_A/d\), maximally mixed; local measurements are completely random.
- Trivial \(B\) (\(\dim\mathcal{H}_B=1\)): \(\rho_A=\rho_{AB}\); the partial trace acts as the identity.
- Schmidt form \(|\psi\rangle=\sum_k\sqrt{p_k}\,|a_k\rangle|b_k\rangle\): \(\rho_A=\sum_k p_k|a_k\rangle\langle a_k|\), a mixture with weights equal to the squared Schmidt coefficients.
Breaks when
- Non-local observables. \(\rho_A\) reproduces \(\langle M_A\otimes I_B\rangle\) but says nothing about correlators \(\langle M_A\otimes N_B\rangle\). Two very different joint states can share the same \(\rho_A\); the partial trace deliberately throws away all cross-system correlation.
- No tensor factorisation. Identical particles (symmetrised/antisymmetrised states), fermionic modes, or systems with superselection rules do not split as \(\mathcal{H}_A\otimes\mathcal{H}_B\). Without a chosen factorisation "\(\operatorname{Tr}_B\)" is ambiguous or ill-defined.
- Non-trace-class operators. In infinite-dimensional \(\mathcal{H}_B\) (e.g. a field mode) the sum \(\sum_j\langle j|\rho_{AB}|j\rangle\) may diverge; the construction and its uniqueness proof presuppose a convergent trace.
Failure modes
- "Pure stays pure." Assuming \(\rho_{AB}\) pure forces \(\rho_A\) pure — false: entangled pure states have mixed reductions.
- Full trace instead of partial. Writing \(\operatorname{Tr}[\rho_{AB}]=1\) (a number) where an operator \(\operatorname{Tr}_B[\rho_{AB}]\) on \(\mathcal{H}_A\) is required.
- Contracting the wrong index. Sandwiching \(\langle j|_B\) but letting it act on the \(A\) factor, or tracing out \(A\) when \(B\) was intended.
- Basis anxiety. Believing \(\rho_A\) depends on the chosen \(\{|j\rangle_B\}\); it does not — cyclicity makes the partial trace basis-independent.
- Correlation amnesia in reverse. Treating \(\rho_A\) as the whole story and trying to predict \(A\)–\(B\) correlations from it alone.
- Improper normalisation. Forgetting the cross terms \(\langle j'|j\rangle=\delta_{j'j}\) and mis-summing off-diagonal \(B\) elements.
Discussion
The deepest content of this result is operational: it is the statement that the local physics of \(A\) is completely encoded in one Hermitian operator. Any two global states indistinguishable by local measurements on \(A\) are mapped to the same \(\rho_A\); conversely, everything a local observer can ever learn is a functional of \(\rho_A\). This is why the reduced density matrix, not the global wavefunction, is the correct notion of "the state of a subsystem".
Entanglement makes reduction irreversible. For a pure entangled \(|\psi\rangle_{AB}\), the reduced state is mixed and its von Neumann entropy \(S(\rho_A)=-\operatorname{Tr}[\rho_A\log\rho_A]\) — the entanglement entropy — quantifies the correlation with \(B\). It equals \(S(\rho_B)\) by the symmetry of the Schmidt decomposition, and vanishes exactly when the state factorises. Decoherence is this mechanism in motion: coupling to an unobserved environment \(B\) and tracing it out suppresses the off-diagonal (coherence) elements of \(\rho_A\), producing the appearance of a classical statistical mixture.
A subtler consequence is no-signalling. Any operation Bob performs on \(B\) — a unitary, or even a measurement whose outcome is not communicated — leaves \(\rho_A=\operatorname{Tr}_B[\rho_{AB}]\) unchanged, because the partial trace is invariant under \(I_A\otimes U_B\) conjugation and under coarse-graining over measurement outcomes. Hence no local statistic on \(A\) can reveal what was done at \(B\), and entanglement cannot transmit information faster than light. The partial trace is thus not a computational convenience but the formal guarantor of relativistic causality in quantum theory. It also sits inside the general framework of completely positive trace-preserving maps: \(\operatorname{Tr}_B\) is a CPTP channel, and every such channel arises as a unitary on a larger space followed by a partial trace (Stinespring dilation), which is the reverse operation to purification.
Common misconceptions. A "proper" mixture (a classical ensemble of pure states Alice actually prepared) and an "improper" mixture (the reduced state of an entangled pair) are described by the same \(\rho_A\) and are locally indistinguishable — but they are not the same physical situation, since the improper mixture retains correlations with \(B\) that no experiment on \(A\) alone can access. Reading \(\rho_A\) as "\(A\) is secretly in one definite pure state we just don't know" is the classic error.
Worked examples
Example 1 — reduced state of a Bell pair.
Reading. Locally, Alice's qubit is completely random: every measurement direction gives \(50/50\). The purity \(\tfrac12<1\) and the one bit of entropy signal maximal entanglement, even though the global state is perfectly pure.
Units check. Dimensionless; \(\operatorname{Tr}\rho_A=\tfrac12+\tfrac12=1\), and \(0\le\operatorname{Tr}\rho_A^2\le1\) as required for a density matrix.
Example 2 — a partially entangled state and a cross-check.
Reading. Weaker entanglement than the Bell pair: purity \(\tfrac58\) and \(0.811\) bits of entropy, both strictly between the product-state and maximally-entangled limits. The local expectation \(\langle\sigma_z\rangle=\tfrac12\) matches the direct joint computation, confirming \(\rho_A\) carries all local statistics.
Units check. Dimensionless; \(\operatorname{Tr}\rho_A=\tfrac34+\tfrac14=1\); \(S\) in bits (\(\log_2\)), bounded by \(0\le S\le\log_2 2=1\).
Problems
- Compute the reduced state of Alice's qubit for the singlet \(|\Psi^{-}\rangle=\tfrac{1}{\sqrt2}(|01\rangle-|10\rangle)\).
Solution
\(\rho_{AB}=\tfrac12(|01\rangle-|10\rangle)(\langle01|-\langle10|)\). Tracing out \(B\): \(\operatorname{Tr}_B|01\rangle\langle01|=|0\rangle\langle0|\), \(\operatorname{Tr}_B|10\rangle\langle10|=|1\rangle\langle1|\), and the cross terms \(\operatorname{Tr}_B|01\rangle\langle10|=\langle0|1\rangle|0\rangle\langle1|=0\). Hence \(\rho_A=\tfrac12(|0\rangle\langle0|+|1\rangle\langle1|)=\tfrac12 I\). Maximally mixed, exactly as for \(|\Phi^{+}\rangle\): all four Bell states have the same maximally-mixed reductions. - For the mixed joint state \(\rho_{AB}=p\,|\Phi^{+}\rangle\langle\Phi^{+}|+(1-p)\,|00\rangle\langle00|\), find \(\rho_A\) as a function of \(p\).
Solution
By linearity of the partial trace, \(\rho_A=p\,\operatorname{Tr}_B|\Phi^{+}\rangle\langle\Phi^{+}|+(1-p)\,\operatorname{Tr}_B|00\rangle\langle00|\). From Example 1 the first term gives \(p\cdot\tfrac12 I=\tfrac{p}{2}(|0\rangle\langle0|+|1\rangle\langle1|)\); the second gives \((1-p)\,|0\rangle\langle0|\). Adding, \(\rho_A=\big(1-\tfrac{p}{2}\big)|0\rangle\langle0|+\tfrac{p}{2}\,|1\rangle\langle1|\). Trace \(=1-\tfrac{p}{2}+\tfrac{p}{2}=1\). At \(p=0\), \(\rho_A=|0\rangle\langle0|\) (pure); at \(p=1\), \(\rho_A=\tfrac12 I\) (maximally mixed). - Show directly from the definition that \(\operatorname{Tr}_B[\rho_A\otimes\rho_B]=\rho_A\) for any \(\rho_B\) with \(\operatorname{Tr}\rho_B=1\).
Solution
Using an orthonormal basis \(\{|j\rangle_B\}\): \(\operatorname{Tr}_B[\rho_A\otimes\rho_B]=\sum_j(I_A\otimes\langle j|)(\rho_A\otimes\rho_B)(I_A\otimes|j\rangle)=\rho_A\sum_j\langle j|\rho_B|j\rangle=\rho_A\,\operatorname{Tr}[\rho_B]=\rho_A\). The \(A\)-factor \(\rho_A\) passes through untouched because \(\langle j|_B\) acts only on the \(B\)-factor, and \(\sum_j\langle j|\rho_B|j\rangle\) is just the definition of \(\operatorname{Tr}\rho_B=1\). - For the GHZ state \(|\mathrm{GHZ}\rangle=\tfrac{1}{\sqrt2}(|000\rangle+|111\rangle)\) on qubits \(A,B,C\), find the single-qubit reduced state \(\rho_A=\operatorname{Tr}_{BC}|\mathrm{GHZ}\rangle\langle\mathrm{GHZ}|\).
Solution
\(\rho_{ABC}=\tfrac12(|000\rangle+|111\rangle)(\langle000|+\langle111|)\). Tracing out \(C\) first: \(\operatorname{Tr}_C|000\rangle\langle111|=\langle1|0\rangle_C|00\rangle\langle11|=0\), so the coherences die and \(\operatorname{Tr}_C\rho_{ABC}=\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|)\). Tracing out \(B\): \(\operatorname{Tr}_B|00\rangle\langle00|=|0\rangle\langle0|\), \(\operatorname{Tr}_B|11\rangle\langle11|=|1\rangle\langle1|\), giving \(\rho_A=\tfrac12(|0\rangle\langle0|+|1\rangle\langle1|)=\tfrac12 I\). Each qubit of a GHZ state looks maximally mixed on its own. - Prove the no-signalling property: for any unitary \(U\) on \(B\), the reduced state \(\rho_A'=\operatorname{Tr}_B[(I_A\otimes U)\rho_{AB}(I_A\otimes U^{\dagger})]\) equals \(\rho_A=\operatorname{Tr}_B[\rho_{AB}]\).
Solution
Write the \(B\)-trace in the basis \(\{|j\rangle_B\}\): \(\rho_A'=\sum_j(I_A\otimes\langle j|)(I_A\otimes U)\rho_{AB}(I_A\otimes U^{\dagger})(I_A\otimes|j\rangle)=\sum_j(I_A\otimes\langle j|U)\rho_{AB}(I_A\otimes U^{\dagger}|j\rangle)\). Since \(U\) is unitary, \(\{U^{\dagger}|j\rangle\}\equiv\{|k\rangle\}\) is another orthonormal basis of \(\mathcal{H}_B\), and \(\langle j|U=\langle k|\) with the sum over \(j\) equal to the sum over \(k\). Therefore \(\rho_A'=\sum_k(I_A\otimes\langle k|)\rho_{AB}(I_A\otimes|k\rangle)=\operatorname{Tr}_B[\rho_{AB}]=\rho_A\). The partial trace is invariant under any local operation on \(B\), so nothing Bob does to \(B\) can change Alice's local statistics — the formal statement of no-signalling.