When Can Errors Be Undone?
Juan Pablo Yamamoto Zazueta
R reference does not evolve Q quantum system E environment |0⁽ᴱ⁾⟩ initially |Ψ(RQ)⟩ entangled interaction After ℰ, can the listener apply 𝒟 to restore |Ψ(RQ)⟩ ?
Entanglement Fidelity
Fₑ = ⟨Ψ(RQ)| ρ″(RQ) |Ψ(RQ)⟩
0 completely lost 1 perfect recovery
Fₑ = 1  :  |Ψ(RQ)⟩ perfectly recovered
Fₑ < 1  :  Entanglement partially or fully lost
The Markov Chain & Data Processing
X channel Y processing H(X) ≥ H(X:Y) ≥ H(X:X̂)
Perfect error correction is possible when
H(X:Y) = H(X) ⟺ H(X|Y) = 0
Information lost cannot be recovered afterwards
Fano's Inequality
H(X|X̂) ≤ 1 + Pᴇ log N
H(X|Y) > 0
Pᴇ bounded away from 0

Imperfect recovery
H(X|Y) = 0
Bob can design decoder
with Pᴇ = 0
Perfect recovery
Pᴇ = 0  ⟺  H(X|Y) = 0  ⟺  H(X:Y) = H(X)
Coherent Information
The right measure for quantum information
I(Q;R) = S(R:Q) − S(R) = S(Q) − S(RQ)
Mutual information
S(R:Q) = 2S(Q)
double-counts
Coherent information
I(Q;R) = S(Q) ✓
for |Ψ(RQ)⟩ pure
Range:  −S(Q)  ≤  I(Q;R)  ≤  +S(Q)
Why Mutual Information Fails
S(R) = S(Q) S(Q) S(RQ) = 0 S(R:Q) = S(R) + S(Q) = 2S(Q)
Both circles count the same entanglement, once from each side
S(R:Q) = S(R) + S(Q) − S(RQ)
for a pure |Ψ(RQ)⟩
S(RQ) = 0
S(R) = S(Q)
∴ S(R:Q) = 2S(Q)  ✕
the problem
We are counting the information carried by R twice, while it is supposed to be just a bystander.
Subtract the Constant
I(Q;R) = S(R:Q) − S(R) = S(Q) − S(RQ)
value for pure |Ψ(RQ)⟩. Target is S(Q)
S(R:Q)
2S(Q)  ✕
S(Q)
target
I(Q;R)
= S(Q)  ✓
A Quantum Signature
Classical: −H(X|Y) ≤ 0 always Quantum: I(Q;R) can be positive −S(Q) 0 +S(Q) −S(Q) ≤ I(Q;R) ≤ +S(Q)
classical analog
−H(X|Y) ≤ 0   always
There is no notion of "order", information is either present or absent.
quantum setting: sign
I > 0: S(Q) > S(RQ)
Entanglement survives: the joint state is more ordered than Q alone.
I < 0: S(RQ) > S(Q)
Entanglement is lost: the joint state is more disordered.
Quantum Error Correction Condition
I(Q′;R) = S(Q)
No coherent information may be lost in the channel
A Concrete Code
Code for three-qubit phase-flip
ℰ(ρ) = (1−p)ρ + p ZρZ (per qubit)
Z|+⟩ = |−⟩    Z|−⟩ = |+⟩
codewords
|0𞁞⟩ = |+++⟩
|1𞁞⟩ = |−−−⟩
logical state
|ψʟ⟩ = α|0𞁞⟩ + β|1𞁞⟩
α and β are the relevant information
The Noise Channel & Codewords
the channel (per qubit)
ℰ(ρ) = (1−p)ρ + p ZρZ
Z in the two bases
standard basis
Z|0⟩ =  |0⟩
Z|1⟩ = −|1⟩
|0⟩, |1⟩ unaffected
conjugate basis
Z|+⟩ = |−⟩
Z|−⟩ = |+⟩
Acts as a bit flip
We reduce the problem of decoherence in the standard basis to detecting bit flip errors in another basis.
code subspace
|0𞁞⟩ = |+ + +⟩
|1𞁞⟩ = |− − −⟩
logical state
|ψʟ⟩ = α|0𞁞⟩ + β|1𞁞⟩
α and β are the amplitudes carrying the relevant quantum information
code subspace projector
Πc = |0𞁞⟩⟨0𞁞| + |1𞁞⟩⟨1𞁞|
2-dimensional subspace of ℋ⊗3 ≅ ℂ⁸
Four Orthogonal Subspaces
ℋ^⊗3 ≅ ℂ⁸ Π𞁞 code subspace α|0𞁞⟩+β|1𞁞⟩ Z⁽¹⁾Π𞁞Z⁽¹⁾ Qubit 1 error dim = 2 Z⁽²⁾Π𞁞Z⁽²⁾ Qubit 2 error dim = 2 Z⁽³⁾Π𞁞Z⁽³⁾ Qubit 3 error dim = 2 Z⁽¹⁾ Z⁽²⁾ Z⁽³⁾
Π𞁞 ⊥ Z⁽¹⁾Π𞁞Z⁽¹⁾ ⊥ Z⁽²⁾Π𞁞Z⁽²⁾ ⊥ Z⁽³⁾Π𞁞Z⁽³⁾
dimension count
4 subspaces × dim 2 = 8
= dim ℋ⊗3 ✓
direct sum decomposition
ℋ⊗ℋ⊗ℋ = Π𞁞
           ⊕ Z⁽¹⁾Π𞁞Z⁽¹⁾
           ⊕ Z⁽²⁾Π𞁞Z⁽²⁾
           ⊕ Z⁽³⁾Π𞁞Z⁽³⁾
The state after noise is in exactly one of the four subspaces, which are distinguishable by orthogonality.
Error Detection & Correction
error
operator
correction
None
𝟙
D₀ = Π𞁞
Qubit 1
Z⁽¹⁾
D₁ = Π𞁞Z⁽¹⁾
Qubit 2
Z⁽²⁾
D₂ = Π𞁞Z⁽²⁾
Qubit 3
Z⁽³⁾
D₃ = Π𞁞Z⁽³⁾
error identification projectors
{Π𞁞, Z⁽¹⁾Π𞁞Z⁽¹⁾, Z⁽²⁾Π𞁞Z⁽²⁾, Z⁽³⁾Π𞁞Z⁽³⁾}
key point
The error identification projectors tell us which subspace the state is in, not the amplitudes α and β that it carries
It can be implemented as a single measurement without the need for cloning (ancilla qubits)
Fₑ = 1 after 𝒟(ℰ(|ψʟ⟩⟨ψʟ|))
Information & Isolation
What does the environment learn?
ρ⁽ᴿᴱ⁾′ = ρ⁽ᴿ⁾ ⊗ σ⁽ᴱ⁾′
R and E have no correlation
The state of the environment is independent of Q's input
equivalent to
I(Q′;R) = S(Q)
The environment learns nothing if and only if the channel is invertible.
The Environment Learns Nothing
R Measure R α₁ → |ψ₁⟩ α₂ → |ψ₂⟩ α₃ → |ψ₃⟩ Interaction E σ⁽ᴱ⁾′ (same) Every conditional Q state leaves E in the same state
the conceptual trick
R can be measured first because it does not participate in the interaction. Each conditional Q state |ψα⟩ must have the same effect on E, making them independent.
Perfect quantum error correction is possible if and only if E is the same for every input.
why this works
If E cannot distinguish between Q states, it has learned nothing and the interaction can be undone.
A Wonderful Theorem
ρ⁽ᴿᴱ⁾′ = ρ⁽ᴿ⁾ ⊗ σ⁽ᴱ⁾′
Perfect Error Correction = Perfect Privacy
When can errors be undone?
Fₑ = 1
Perfect recovery
I(Q′;R) = S(Q)
No coherent info lost
ρ⁽ᴿᴱ⁾′ = ρ⁽ᴿ⁾ ⊗ σ⁽ᴱ⁾′
R and E are a product state
σα⁽ᴱ⁾′ independent of α
E learns nothing
"Quantum mechanics is what happens
when nobody is looking."