32. The Fundamental Problem

32.1Why classical redundancy works

Classical redundancy works because bits can be copied, measured, and processed by nonlinear logic. Store one bit three times and decode by majority vote: if each copy flips independently with probability p, the value fails only when two or more copies flip, so p_L = 3p^2 - 2p^3, which is below p for every p < ½. Shannon generalized this into a full theory of reliable communication over noisy channels. Notice the three ingredients: unlimited copying, direct measurement of the stored value, and a nonlinear decoding rule. Quantum mechanics will deny the first two and constrain the third. Everything in this part is a way to recover them indirectly.

32.2Why copying a qubit doesn't work

The obvious move — copy the unknown state — fails immediately. Apply CNOT with the unknown qubit as control and a fresh |0⟩ as target: α|0⟩ + β|1⟩ ⊗ |0⟩ → α|00⟩ + β|11⟩. That is an entangled state, not two copies (α|0⟩ + β|1⟩) ⊗ (α|0⟩ + β|1⟩). Measuring to learn α and β does not help either: the Born rule gives one sample per copy, the amplitudes are continuous numbers, and the measurement disturbs the state it reads. So redundancy cannot be added by inspection. It must be added by a unitary that never looks at the amplitudes at all — which is what encoding is.

32.3No-cloning theorem

The no-cloning theorem (Wootters, Zurek, and Dieks, 1982) makes the failure general: no unitary copies an arbitrary unknown quantum state. The proof is two lines of linearity. Suppose U|0⟩|0⟩ = |0⟩|0⟩ and U|1⟩|1⟩ = |1⟩|1⟩. Then on a superposition, U(α|0⟩ + β|1⟩)|0⟩ = α|00⟩ + β|11⟩, but a true clone would be (α|0⟩ + β|1⟩) ⊗ (α|0⟩ + β|1⟩), which contains cross terms αβ|01⟩ and αβ|10⟩. Since U is linear, both cannot hold. Consequences: no checkpointing a quantum process, no amplifying a quantum signal without destroying it, and — constructively — error correction must work on states it never learns.

32.4Encoding quantum information

The fix is to encode into a larger Hilbert space with a unitary that depends only on the basis, not the amplitudes:

encoding  |psi> = a|0> + b|1>  into the 3-qubit bit-flip code

q0: |psi> --●----●--
q1: |0>  ---⊕----|-        result:  a|000> + b|111>
q2: |0>  --------⊕--

The information now lives in the correlations between qubits; no single qubit carries α or β. If an X error hits qubit 1, the state becomes α|010⟩ + β|101⟩ — a different pair of orthogonal basis vectors, but the same two-dimensional structure. An operation exists that maps it back, without ever knowing α or β. Classically, information is in the bits; quantumly, it is in the pattern of entanglement. That reframing is the entire conceptual leap of quantum error correction.

32.5Detecting errors without measuring the state

How do you learn which error occurred without learning α and β? Measure an observable whose eigenvectors do not distinguish the codewords. For the encoded state above, the operator Z0Z1 works: |000⟩ and |111⟩ are both +1 eigenstates of Z0Z1 (each has an even number of 1s on qubits 0 and 1), so any superposition α|000⟩ + β|111⟩ is also a +1 eigenstate — the outcome is +1 regardless of the amplitudes. After an X error on qubit 0, both components move entirely into the −1 eigenspace, so the outcome is −1. The measurement reveals the error's fingerprint and nothing about the data. "Measure the parity, not the state" is the design principle behind every code in this part.

32.6Syndrome measurement

Measuring the two parities gives a pair of classical bits — the syndrome (syndrome):

syndrome (Z0Z1, Z1Z2)   outcome   implied error          correction
      (+1, +1)            00       none (or Z-type)       do nothing
      (-1, +1)            10       X on qubit 0           X on qubit 0
      (-1, -1)            11       X on qubit 1           X on qubit 1
      (+1, -1)            01       X on qubit 2           X on qubit 2

Correction is then a conditional Pauli. Two subtleties matter. First, the syndrome identifies the error only up to multiplication by a stabilizer (34.3): errors E and E·S produce identical syndromes and identical effects on the codespace, so "which error occurred" is only defined modulo equivalence. Second, the (+1, +1) row also covers uncorrectable errors: a triple flip acts as the logical operator X⊗X⊗X and a Z error is invisible to Z-parities. A distance-3 code corrects any single-qubit error; that is all it promises.