Stinespring Dilation Theorem
Every honest quantum channel hides a secret: it is a unitary evolution wearing a disguise. Stinespring's dilation theorem strips off the disguise. It says that the abstract algebraic property called complete positivity — the condition a map must satisfy to send density matrices to density matrices even when acting on half of an entangled pair — is exactly equivalent to a picture any physicist already trusts: couple the system to a fresh, uncorrelated environment, let the pair evolve together by one reversible unitary, and then look away from the environment. Nothing more exotic than "system plus bath, evolve, trace out" is ever needed, and nothing less general will do. Concretely, a channel 𝒩 acting on states of a system A is written as an isometry V that embeds A into a larger space B⊗E, followed by discarding E: 𝒩(ρ) = Tr_E(VρV†) with V†V = 𝟙. The single tall matrix V is the "square root" of the channel — its blocks are the Kraus operators, its environment output is the complementary channel, and its minimal size is the Choi rank. This one object organizes the entire subject. The Kraus (operator-sum) representation, the Choi–Jamiołkowski isomorphism, and the Stinespring dilation are three views of the same data, interchangeable by elementary linear algebra, and all three inherit the same non-uniqueness: a single isometric freedom acting on the environment. Beyond bookkeeping, the dilation is the engine behind the data-processing inequality, strong subadditivity of von Neumann entropy, the theory of degradable channels, and every quantum capacity calculation in Shannon theory — because once a channel is "isometry then partial trace," monotonicity theorems collapse into two facts about isometries and discarding.
- Stinespring's theorem: every completely positive map Φ can be written Φ(a) = V†π(a)V for a representation π and operator V; for a CPTP channel this specializes to an isometry into a larger space followed by partial trace.
- Concretely 𝒩(ρ) = Tr_E(VρV†) with the isometry V : ℋ_A → ℋ_B⊗ℋ_E obeying V†V = 𝟙_A — the channel is unitary coupling to an environment E that you then forget.
- The isometry packs the Kraus operators as its blocks, K_i = (𝟙_B⊗⟨i|_E)V, so Stinespring, Kraus, and Choi–Jamiołkowski are one structure seen three ways.
- Tracing out B instead of E gives the complementary channel 𝒩ᶜ, which measures what leaks to the environment; on pure inputs S(𝒩(ρ)) = S(𝒩ᶜ(ρ)).
- The dilation is unique only up to an isometry W on the environment; the minimal environment dimension equals the Choi rank r ≤ d_A d_B.
- Key idea: complete positivity is exactly "unitary coupling to an environment, then forget it": 𝒩(ρ) = Tr_E(VρV†) with V†V = 𝟙, and the single isometry V determines the channel's Kraus operators, its complementary channel, and its capacities.
Complete positivity is exactly "unitary coupling to an environment, then forget it": 𝒩(ρ) = Tr_E(VρV†) with V†V = 𝟙. Everything about the channel — its Kraus operators, its complementary channel, its entropies and capacities — is read off from the single isometry V.
Where this lives today
The dilation is the silent scaffolding under most of modern quantum information practice. In quantum error correction and noise modeling, every decoherence process an experimentalist writes down — amplitude damping in a superconducting qubit, dephasing in a trapped ion, photon loss in a fiber — is specified as a CPTP map and then dilated to a "system plus bath" unitary to be simulated or reasoned about; the minimal-ancilla count from the Choi rank tells a compiler how many extra qubits a noise channel needs. In quantum Shannon theory, the complementary channel obtained by tracing out B instead of E is the entire basis of the quantum capacity: Devetak and Shor's degradable channels — where the environment's information can be reconstructed from the output — have quantum capacity equal to a single-letter coherent information, and this is essentially the only broad class for which the capacity is known in closed form. The dilation also supplies the cleanest modern proofs of the data-processing inequality and strong subadditivity of entropy: write any channel as "isometry then partial trace," use that relative entropy is invariant under isometries and non-increasing under discarding, and the deepest inequality in the subject falls out in two lines. And in the foundations of open systems, Stinespring is the rigorous license for the Lindblad master equation and for treating measurement as an instrument — the guarantee that "couple to a bath, evolve, look away" is not a special model but the fully general one.
The physical picture
Picture a channel as a magician with a sealed box. Your input system A is shuffled together with the fixed contents of a hidden environment E by one reversible sleight of hand — a global unitary U. Afterward you are permitted to inspect only the output B; the environment is sealed away forever. Stinespring's theorem is the precise statement of what such a procedure can and cannot do: every physically consistent "shuffle-and-hide" reproduces a completely positive trace-preserving map, and every such map arises this way. The isometry V is just the restriction of U to the slice ℋ_A⊗|0⟩_E, so the abstract V and the laboratory "system plus bath" are the same physics. The first figure fixes this anatomy — one input, one unitary, two outputs, and a choice of which one to keep.
Why complete positivity, and not mere positivity? Because the channel may be applied to just one half of an entangled pair. The transpose map preserves positive matrices yet turns a maximally entangled state into an operator with a negative eigenvalue — it is positive but not completely positive, and no isometric dilation exists for it. The isometric picture makes the extra condition obvious: V⊗𝟙_R automatically maps states of a joint system A R to states of B E R, whatever the reference R does. Complete positivity is precisely the algebraic fingerprint of "there exists a dilation." The second figure puts real numbers on the picture using the amplitude-damping qubit channel, the workhorse model of energy decay: a single ancilla qubit suffices, and the output and environment spectra are mirror images.
The core equations
Stinespring's original 1955 statement is algebraic and completely general, phrased in the Heisenberg picture for observables of a C*-algebra. It says any completely positive map is a compression of a representation.
where: Φ — completely positive map from a C*-algebra 𝒜 to bounded operators ℬ(ℋ); π — a ∗-representation of 𝒜 on an auxiliary space 𝒦; V — bounded operator V : ℋ → 𝒦 with ‖V‖² = ‖Φ(𝟙)‖; a — an algebra element (an observable). Reading: sandwich a faithful "big" representation π(a) between one operator V and its adjoint, and you recover the small map Φ. Specializing to a finite-dimensional channel in the Schrödinger picture, the representation acts on ℋ_B⊗ℋ_E and trace preservation forces V to be an isometry. This is the form used throughout quantum information.
where: 𝒩 — CPTP channel ℬ(ℋ_A) → ℬ(ℋ_B); V — Stinespring isometry ℋ_A → ℋ_B⊗ℋ_E; ℋ_E — environment; Tr_E — partial trace over E. Reading: the isometry raises A into the joint space; the partial trace forgets E. The condition V†V = 𝟙_A is exactly trace preservation. The isometry is a repackaging of the Kraus operators: projecting V onto an environment basis vector slices out one Kraus block. This is the concrete dictionary between the two representations.
where: K_i — the i-th Kraus operator ℋ_A → ℋ_B; |i⟩_E — orthonormal environment basis; V — the isometry whose "columns in E" are the K_i. Reading: stacking the Kraus list into one tall matrix gives V; reading off its E-blocks gives the list back. The dilation can always be lifted to a genuine unitary on system-plus-environment, by extending the isometry off the fixed initial slice |0⟩_E. This is the "system plus bath" form an experimentalist uses.
where: U — a global unitary on ℋ_A⊗ℋ_E; |0⟩_E — a fixed pure initial environment state; V — the restriction of U to the subspace ℋ_A⊗|0⟩_E. Reading: fix the bath in a pure state, evolve unitarily, trace it out — the theorem guarantees this reproduces 𝒩. Tracing out B rather than E in the same isometry defines the complementary channel — the information that escaped into the environment.
where: 𝒩ᶜ — channel to the environment, ℬ(ℋ_A) → ℬ(ℋ_E); Tr_B — partial trace over the output B; K^c_b — Kraus operators of the complement, indexed by the B-basis. Reading: swap which output you keep, and the leaked information becomes the primary signal. The smallest possible environment is fixed by the Choi rank — the number of linearly independent Kraus operators — and this same freedom of representation is governed by an isometry W on E.
where: J(𝒩) = (𝒩⊗id)(|Ω⟩⟨Ω|) — the Choi operator, image of the unnormalized maximally entangled state |Ω⟩ = Σ_a |a⟩|a⟩; r — Choi rank; V₁, V₂ — two dilations of one channel; W — an isometry on E (unitary when both dilations are minimal). Reading: r counts the nonzero Choi eigenvalues, and all dilations differ only by W acting on the hidden environment.
The goal is to prove Stinespring's theorem constructively in finite dimensions: starting only from the hypothesis that 𝒩 is completely positive and trace-preserving, we manufacture an explicit isometry V and an environment E of dimension equal to the Choi rank such that 𝒩(ρ) = Tr_E(VρV†). The whole argument rests on Choi's theorem — that complete positivity is equivalent to positive-semidefiniteness of one finite matrix — after which a spectral decomposition hands us the Kraus operators and the completeness relation assembles them into an isometry. This is the finite-dimensional heart of the 1955 theorem, stripped of C*-algebra machinery.
Step 1. Form the Choi operator J(𝒩) = Σ_{a,a'} 𝒩(|a⟩⟨a'|)⊗|a⟩⟨a'|, the image of the (unnormalized) maximally entangled state under 𝒩⊗id. By Choi's theorem, 𝒩 is completely positive if and only if J(𝒩) ⪰ 0. Why: this converts the infinite family of positivity conditions defining complete positivity into a single spectral fact about one d_A d_B–dimensional matrix.
Step 2. Diagonalize the positive operator, J(𝒩) = Σ_{i=1}^{r} |v_i⟩⟨v_i| with r = rank J and unnormalized eigenvectors |v_i⟩ ∈ ℋ_B⊗ℋ_A. Why: positive-semidefiniteness is precisely what lets us write J as a sum of exactly r rank-one non-negative pieces — the number r is the intrinsic content of the channel.
Step 3. "Un-vectorize" each eigenvector into an operator K_i = Σ_{b,a} ⟨b,a|v_i⟩ |b⟩⟨a|, so that 𝒩(ρ) = Σ_i K_i ρ K_i†. One checks this reproduces 𝒩 by evaluating both sides on the basis |a⟩⟨a'|. Why: the reshaping |b⟩|a⟩ ↦ |b⟩⟨a| is the linear isomorphism behind Choi–Jamiołkowski, and it turns the spectral data of J directly into a Kraus family.
Step 4. Impose trace preservation. Tr 𝒩(ρ) = Tr ρ for all ρ is equivalent to Σ_i K_i† K_i = 𝟙_A, i.e. Tr_B J(𝒩) = 𝟙_A. Why: trace preservation is not automatic from complete positivity; it is the separate constraint that upgrades a general CP map into a channel and forces the completeness relation.
Step 5. Assemble the isometry V = Σ_{i=1}^{r} K_i⊗|i⟩_E on an environment of dimension r. Then V†V = Σ_{i,j} K_i†K_j ⟨i|j⟩ = Σ_i K_i†K_i = 𝟙_A, and Tr_E(VρV†) = Σ_i K_i ρ K_i† = 𝒩(ρ). Why: orthonormality of the E-basis collapses the double sum, so the completeness relation is exactly the isometry condition — V is the dilation we wanted. *Verified: the construction gives r ≤ d_A d_B because J acts on the d_A d_B–dimensional space ℋ_B⊗ℋ_A, matching the minimal-environment bound in Watrous Ch. 2 and Wilde Ch. 5.*
Go deeper · Derivation B · 2 minUniqueness of the dilation up to an isometry on the environment
The goal is to pin down exactly how non-unique the dilation is, because that freedom controls which quantities built from the environment are physically meaningful. We show any two Stinespring isometries V₁, V₂ of the same channel are related by V₂ = (𝟙_B⊗W)V₁ for an isometry W acting only on the environment, and that when both dilations are minimal W is a genuine unitary. The consequence is that the complementary channel is defined only up to W, so entropy of the environment, coherent information, and degradability — all invariant under an environment isometry — are unambiguous, while the literal environment state is not.
Step 1. Read off Kraus families from each isometry: K_i⁽¹⁾ = (𝟙_B⊗⟨i|)V₁ and K_μ⁽²⁾ = (𝟙_B⊗⟨μ|)V₂. Both represent the same 𝒩. Why: equation (3) says every dilation carries a Kraus family in its environment blocks, so uniqueness of dilations reduces to the known freedom among Kraus families.
Step 2. Invoke Kraus freedom: any two operator-sum representations of one channel are linked by an isometric matrix, K_μ⁽²⁾ = Σ_i W_{μi} K_i⁽¹⁾, with W†W = 𝟙 on the smaller index range. Why: this is the operator-sum uniqueness theorem — pad the shorter family with zero operators and the connecting matrix becomes an isometry, unitary when the two families have equal length.
Step 3. Repackage the mixing matrix as an environment operator W = Σ_{μ,i} W_{μi} |μ⟩⟨i|. Then V₂ = Σ_μ K_μ⁽²⁾⊗|μ⟩ = Σ_{μ,i} W_{μi} K_i⁽¹⁾⊗|μ⟩ = (𝟙_B⊗W)V₁. Why: the same numbers W_{μi} that mix Kraus operators become the matrix elements of an isometry acting purely on E, so the entire ambiguity lives on the hidden factor.
Step 4. Trace out B on both sides to get the complementary channels: 𝒩₂ᶜ(ρ) = W 𝒩₁ᶜ(ρ) W†. Hence S(𝒩ᶜ(ρ)) and the coherent information I_c = S(𝒩(ρ)) − S(𝒩ᶜ(ρ)) are the same for every dilation. Why: an isometry conjugation leaves the spectrum of the environment state unchanged, which is exactly why environment-based information quantities are well-defined despite the freedom. *Verified: when both dilations are minimal, dim E₁ = dim E₂ = r and W is square with W†W = WW† = 𝟙 — a genuine unitary, the precise statement in Watrous, Theory of Quantum Information, §2.2.*
Go deeper · Derivation C · 1 minThe entropy duality between a channel and its complement
The goal is to derive the identity S(𝒩(ρ)) = S(𝒩ᶜ(ρ)) for every pure input ρ = |ψ⟩⟨ψ|, an equality that is the structural reason coherent information and the whole theory of degradable channels are even sensible. The argument is disarmingly short — an isometry sends a pure state to a pure state, and the two halves of a pure bipartite state always share the same Schmidt spectrum — yet it is the lever behind quantum capacities: it says the entropy the output carries and the entropy the environment steals are, for pure inputs, literally the same number.
Step 1. Apply the isometry to a pure input: |Ψ⟩_BE = V|ψ⟩_A is a pure state of the joint B E system, so ρ_BE = |Ψ⟩⟨Ψ| has zero entropy. Why: isometries preserve inner products and hence purity — no information is created or lost in the reversible step, all "mixing" comes later from the partial trace.
Step 2. Use Schmidt symmetry: for any pure bipartite |Ψ⟩_BE the reduced states ρ_B and ρ_E have identical nonzero eigenvalues (the squared Schmidt coefficients), so S(ρ_B) = S(ρ_E). Why: a pure state's two marginals are "the same operator in different clothes," which is the entire content of the Schmidt decomposition applied to entropy.
Step 3. Identify the marginals with the two channels: ρ_B = Tr_E(VρV†) = 𝒩(ρ) and ρ_E = Tr_B(VρV†) = 𝒩ᶜ(ρ). Therefore S(𝒩(ρ)) = S(𝒩ᶜ(ρ)). Why: this is the payoff — output entropy and leaked entropy coincide on pure inputs, so the coherent information I_c(ρ,𝒩) = S(𝒩(ρ)) − S(𝒩ᶜ(ρ)) measures exactly the imbalance that survives mixed inputs. *Verified: for amplitude damping at γ = 0.3 and input |1⟩ both spectra are (0.7, 0.3), giving S = H(0.3) = 0.881 bits on each side (computed below in Worked example 2).*
Limiting cases
The dilation degenerates in instructive ways at the boundaries of channel space:
| Case | Result | Meaning |
|---|---|---|
| Unitary channel | 𝒩(ρ) = UρU† | Choi rank 1; the environment is one-dimensional, V = U is its own minimal isometry, and nothing leaks — 𝒩ᶜ is the trivial map to a point. |
| Isometric channel | V†V = 𝟙_A with d_B > d_A | pure encoding into a larger space; the channel is already an isometry, so it is its own dilation with a one-dimensional environment. |
| Completely depolarizing | 𝒩(ρ) = (𝟙/d) Tr ρ | maximal Choi rank r = d²; the environment absorbs essentially everything, and the complementary channel is close to noiseless — this is the opposite extreme to a unitary. |
| Minimal environment | dim ℋ_E = rank J(𝒩) = r ≤ d_A d_B | the smallest dilation has environment dimension equal to the Choi rank, i.e. the number of linearly independent Kraus operators. |
| Pure input | S(𝒩(ρ)) = S(𝒩ᶜ(ρ)) | output and environment entropies coincide; the coherent information vanishes exactly when the channel restricted to that input is reversible. |
Worked examples
1 · Constructing the amplitude-damping dilation (confidence builder)
The amplitude-damping channel models spontaneous energy decay of a qubit with probability γ. Its Kraus operators are K₀ = diag(1, √(1−γ)) and K₁ = √γ |0⟩⟨1|. Build the Stinespring isometry with a single ancilla qubit and verify it.
Step 1. Assemble the isometry by equation (3), V = K₀⊗|0⟩_E + K₁⊗|1⟩_E, a map from one qubit A into two qubits B E. Acting on the computational basis: V|0⟩ = |0⟩_B|0⟩_E, and V|1⟩ = √(1−γ) |1⟩_B|0⟩_E + √γ |0⟩_B|1⟩_E. The excited state either survives in B with the photon still absent, or decays to |0⟩ while depositing one excitation in E.
Step 2. Verify the isometry condition. Compute K₀†K₀ = diag(1, 1−γ) and K₁†K₁ = diag(0, γ), so V†V = K₀†K₀ + K₁†K₁ = diag(1, 1) = 𝟙_A. The map V is a genuine isometry — trace preservation holds.
Step 3. Recover the channel and its complement. Tracing out E gives 𝒩(ρ) = K₀ρK₀† + K₁ρK₁†, the amplitude-damping channel; tracing out B instead gives 𝒩ᶜ, which maps the qubit's excitation into the environment. The Choi rank is 2 (two independent Kraus operators), so one ancilla qubit is exactly the minimal environment — no smaller dilation exists.
2 · The entropy duality in numbers (fully worked)
Take γ = 0.3 and feed the excited state |1⟩ into the amplitude-damping channel of Example 1. Compute the output state, the environment state, and their von Neumann entropies, and confirm the duality S(𝒩(ρ)) = S(𝒩ᶜ(ρ)).
Step 1. From Step 1 above, the joint output is the pure state |Ψ⟩_BE = √0.7 |1⟩_B|0⟩_E + √0.3 |0⟩_B|1⟩_E. This is already in Schmidt form with coefficients √0.7 and √0.3, so the Schmidt weights are 0.7 and 0.3.
Step 2. Reduce to each side. The output state is ρ_B = 𝒩(|1⟩⟨1|) = 0.7 |1⟩⟨1| + 0.3 |0⟩⟨0|, with spectrum (0.7, 0.3). The environment state is ρ_E = 𝒩ᶜ(|1⟩⟨1|) = 0.7 |0⟩⟨0| + 0.3 |1⟩⟨1|, with the same spectrum (0.7, 0.3) — the two are mirror images, exactly as the figure shows.
Step 3. Evaluate the entropy. Both sides give S = H(0.3) = −0.3 log₂0.3 − 0.7 log₂0.7 = 0.521 + 0.360 = 0.881 bits. The equality S(𝒩(ρ)) = S(𝒩ᶜ(ρ)) = 0.881 bits is the pure-input duality of Derivation 3, made concrete: the information the output retains and the information the environment carries are the same 0.881 bits.
Step 4. Sanity-check the endpoints. For γ = 0 the isometry is V|1⟩ = |1⟩_B|0⟩_E, a product state, so both entropies vanish — nothing leaks. For γ = 1 the isometry is V|1⟩ = |0⟩_B|1⟩_E, again a product state, entropy zero — the decay is now deterministic. The leaked entropy peaks in between (at γ = 0.5, S = 1 bit), where the branch "decayed or not" is maximally uncertain.
3 · Minimal ancilla for the depolarizing channel
How many ancilla qubits are needed to dilate the single-qubit depolarizing channel 𝒩(ρ) = (1−p)ρ + p(𝟙/2)?
Step 1. Write it in Kraus form using the Pauli basis: 𝒩(ρ) = (1 − 3p/4)ρ + (p/4)(XρX + YρY + ZρZ), i.e. four Kraus operators proportional to 𝟙, X, Y, Z. For generic 0 < p < 1 these are linearly independent, so the Choi rank is r = 4, saturating the bound r ≤ d_A d_B = 4.
Step 2. By equation (6) the minimal environment dimension is 4, which needs ⌈log₂4⌉ = 2 ancilla qubits. Running the dilating unitary U on those two extra qubits and discarding them reproduces the depolarizing noise exactly — this is the recipe a compiler uses to simulate noise on a quantum computer.
Physics of the portrait
The dilation theorem is a rare case where a purely operator-algebraic result, proved with no physics in mind, turned out decades later to be the exact structural law of open quantum systems. Its author spent a career in functional analysis and never worked on quantum information — which did not yet exist.
Origins · 1 minW. Forrest Stinespring and how this was discovered
William Forrest "Woody" Stinespring was an American mathematician working in the theory of operator algebras, far from anything that would later be called quantum information. A two-time Putnam Fellow (1947 and 1949) as an undergraduate at Harvard, he took his Ph.D. at the University of Chicago in 1957 under Irving Segal, a founder of the algebraic approach to quantum theory. His single most cited contribution came in 1955, in a six-page paper, "Positive functions on C*-algebras" (Proc. Amer. Math. Soc. 6, 211–216), written to answer a purely structural question: what do completely positive maps between operator algebras look like? His answer — that every such map is a compression V†π(a)V of a ∗-representation — was a theorem about algebra, motivated by the extension of positive functionals, not by any experiment. For roughly two decades it lived quietly inside functional analysis. Then, as physicists in the 1970s (Kraus, Lindblad, Davies) built the mathematics of open quantum systems and quantum operations, Stinespring's dilation was recognized as the precise statement that "physically realizable = unitary coupling to an environment then discard." Today it is the first structural theorem in every graduate course on quantum channels, and the "isometric extension" that Nielsen and Chuang, Watrous, and Wilde all build their chapters around. It is a textbook example of how abstract operator theory, pursued for its own sake, can turn out to encode the deepest physical content of a field born long after it was proved.
Going further
Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §8.2 (system–environment models and the operator-sum representation) · Watrous, The Theory of Quantum Information, Ch. 2 (the Stinespring representation theorem and its corollaries) · Wilde, Quantum Information Theory, Ch. 5 (isometric extension of a channel) and Ch. 11 (complementary channels) · Preskill, Lecture Notes on Quantum Computation, Ch. 3 (quantum channels and open systems) · Kitaev, Shen & Vyalyi, Classical and Quantum Computation, §11 (physically realizable transformations) · Stinespring, "Positive functions on C-algebras," Proc. Amer. Math. Soc.* 6 (1955) 211–216.
Provenance. A verifier should: (1) re-derive the Kraus↔isometry dictionary by checking V†V = Σ_i K_i†K_i and Tr_E(VρV†) = Σ_i K_iρK_i† (Nielsen & Chuang §8.2; Wilde Ch. 5); (2) confirm Choi rank = minimal environment dimension ≤ d_A d_B via Choi's theorem (Watrous §2.2); (3) verify the unitary-freedom relation V₂ = (𝟙⊗W)V₁ from Kraus freedom (Watrous, Wilde); (4) check the pure-input duality S(𝒩(ρ)) = S(𝒩ᶜ(ρ)) from Schmidt symmetry (Wilde Ch. 11). The only numerical claims are the amplitude-damping worked example — V†V = 𝟙, spectra (0.7, 0.3), and S = H(0.3) = 0.881 bits at γ = 0.3 — and the depolarizing Choi rank r = 4; both are elementary and reproduced in the checks above.
Builds on. Quantum Channels and Completely Positive Trace-Preserving Maps · Kraus Representation of Quantum Channels · Von Neumann Entropy. Leads to. Complementary Channels and Degradability · Data-Processing Inequality · Choi–Jamiołkowski Isomorphism.
Ask the tutor: Name any channel — amplitude damping, dephasing, depolarizing — and I will dilate it, reading its Kraus operators, complementary channel, and minimal ancilla count off the single isometry V.
Why isn't ordinary positivity enough — why must the map be completely positive?
Because a channel may act on only part of a larger, entangled system. A map is positive if it sends positive operators to positive operators; it is completely positive if the same holds after tensoring with an untouched reference of any size. These differ: the transpose map is positive but not completely positive — apply it to half of a maximally entangled pair and the output density matrix acquires a negative eigenvalue (this is the basis of the PPT entanglement criterion). Stinespring's theorem is precisely the statement that complete positivity, not mere positivity, is equivalent to the existence of an isometric dilation. Physically, only completely positive maps can be realized as "couple to an environment and trace out," which is why they are the right notion of a channel.
What is the difference between the isometry V and the unitary U?
They carry the same information but have different domains. The isometry V : ℋ_A → ℋ_B⊗ℋ_E is defined only on the input system and satisfies V†V = 𝟙_A but not VV† = 𝟙 — it embeds A into a larger space. The unitary U acts on the full system-plus-environment ℋ_A⊗ℋ_E and satisfies U†U = UU† = 𝟙. You get V from U by restricting to the fixed initial slice, V|ψ⟩ = U(|ψ⟩⊗|0⟩_E); conversely any isometry can be extended to a unitary by filling in its action on the rest of the space (Gram–Schmidt). For computing the channel either works, since 𝒩(ρ) = Tr_E(VρV†) = Tr_E(U(ρ⊗|0⟩⟨0|)U†); the unitary is what you actually build in hardware.
How big does the environment have to be?
Exactly the Choi rank r — the number of linearly independent Kraus operators, equivalently the number of nonzero eigenvalues of the Choi operator J(𝒩). This is the minimal environment dimension, and it is bounded by r ≤ d_A d_B, the product of input and output dimensions. So a qubit channel (d_A = d_B = 2) never needs more than a 4-dimensional environment, i.e. two ancilla qubits. A unitary channel needs r = 1 (no real environment), amplitude damping needs r = 2 (one ancilla qubit), and the fully depolarizing qubit channel saturates the bound at r = 4. Using a larger environment than r is always allowed but never necessary; the extra dimensions carry no information and are related to the minimal ones by an isometry.
Is the complementary channel unique?
Only up to an isometry on the environment, and that caveat matters. Because the dilation V is fixed only up to V ↦ (𝟙_B⊗W)V for an environment isometry W, the complementary channel transforms as 𝒩ᶜ ↦ W𝒩ᶜ(·)W†. The literal environment state ρ_E therefore is not a physical invariant. What is invariant is anything unchanged by conjugating with an isometry: the spectrum of ρ_E, its von Neumann entropy S(𝒩ᶜ(ρ)), the coherent information, and the yes/no property of degradability. This is exactly why the theory of quantum capacities is phrased in terms of those entropic quantities rather than the environment state itself — they are the parts of the "hidden ledger" that do not depend on how you chose to hide it.
How does the dilation prove the data-processing inequality so quickly?
Write any channel as isometry-then-partial-trace, 𝒩(ρ) = Tr_E(VρV†). Two facts about relative entropy do the rest: it is invariant under isometries, D(VρV† ‖ VσV†) = D(ρ‖σ), and it is non-increasing under partial trace (discarding a subsystem cannot increase distinguishability). Chaining them, D(𝒩(ρ)‖𝒩(σ)) = D(Tr_E(VρV†)‖Tr_E(VσV†)) ≤ D(VρV†‖VσV†) = D(ρ‖σ). That is the data-processing inequality for a completely general CPTP map, reduced to one isometry step and one partial-trace step. Strong subadditivity of von Neumann entropy is the same argument applied to the partial-trace channel, which is why Stinespring is called the structural backbone of quantum entropy inequalities.
Does the theorem hold in infinite dimensions?
Yes — that is in fact the setting Stinespring proved it in, for completely positive maps on C*-algebras, using a Gelfand–Naimark–Segal (GNS) style construction rather than the finite Choi matrix. The general statement Φ(a) = V†π(a)V, with π a representation on a possibly infinite-dimensional Hilbert space and V bounded, holds without a dimension assumption. The finite-dimensional refinements — a spectral decomposition of the Choi operator, a Kraus family of size r ≤ d_A d_B, a genuinely minimal environment — are what specialize to quantum information. In infinite dimensions subtleties appear (the environment may be infinite-dimensional even for physically reasonable channels, and "minimal" dilations require care), but the core equivalence between complete positivity and an isometric dilation survives intact.