The QEC Conditions

A quantum code is a subspace $\mathcal{C} \subseteq (\mathbb{C}^2)^{\otimes n}$ — the codespace — together with an interpretation of certain orthonormal vectors $\{|\bar{i}\rangle\}$ in it as the encoded logical states. The central question of the theory is: given a set of physical errors that might strike the code, when can we recover the original logical information exactly? The Knill–Laflamme conditions answer this completely.

The error set

Model the noise as a set of operators $\{E_a\}$ acting on the $n$ physical qubits. These need not be unitary; in the most general setting each $E_a$ is a Kraus operator of some noise channel $\mathcal{N}(\rho) = \sum_a E_a \rho E_a^\dagger$ . We will see in the next lesson that for qubits it suffices to take the $E_a$ from the Pauli group, but for now keep them arbitrary.

The conditions

Let $P$ be the projector onto the codespace $\mathcal{C}$ . A recovery operation $\mathcal{R}$ that perfectly corrects every error in $\{E_a\}$ exists if and only if

P\, E_a^\dagger E_b\, P = \alpha_{ab}\, P \qquad \text{for all } a, b,

where $\alpha_{ab}$ is a Hermitian matrix of complex scalars independent of the codeword. Written on basis codewords $|\bar{i}\rangle, |\bar{j}\rangle$ this reads

\langle \bar{i} |\, E_a^\dagger E_b \, |\bar{j}\rangle = \alpha_{ab}\,\delta_{ij}.

Two requirements are packed into this single equation, and reading them separately is the key to intuition.

What each part means

Off-diagonal in the logical index ( $i \neq j$ ): $\langle \bar{i}| E_a^\dagger E_b |\bar{j}\rangle = 0$ . Distinct logical codewords must remain orthogonal even after any pair of errors acts. If two errors could carry $|\bar{0}\rangle$ and $|\bar{1}\rangle$ into overlapping states, no measurement could later tell them apart — information would be irretrievably scrambled.

Diagonal in the logical index ( $i = j$ ): $\langle \bar{i}| E_a^\dagger E_b |\bar{i}\rangle = \alpha_{ab}$ , the same number for every codeword. This is the subtle quantum requirement: the errors must not reveal which logical state we hold. If $\alpha_{ab}$ depended on $i$ , then measuring the environment (which effectively learns $E_a^\dagger E_b$ expectation values) would extract information about the logical qubit, and by the no-cloning logic that disturbance would destroy the encoded superposition.

Why "if and only if"

The "only if" direction is a consistency check: if recovery $\mathcal{R}$ exists with $\mathcal{R} \circ \mathcal{E} = \mathrm{id}$ on the codespace, expanding both sides in Kraus operators forces the conditions. The "if" direction is constructive: diagonalising the Hermitian matrix $\alpha$ gives a new error basis $\{F_k\}$ (linear combinations of the $E_a$ ) with $P F_k^\dagger F_l P = d_k \delta_{kl} P$ . Each $F_k$ maps the codespace isometrically (up to normalization) into a mutually orthogonal subspace $V_k$ , so measuring "which $V_k$ am I in" is a non-demolition syndrome measurement, after which a unitary returns each $V_k$ to $\mathcal{C}$ . This basis-diagonalisation is exactly why arbitrary errors reduce to a finite correctable set — the subject of the next lesson.

Degenerate vs. non-degenerate

If the matrix $\alpha_{ab}$ is full rank, the code is non-degenerate: distinct errors send the codespace to distinct orthogonal subspaces. If $\alpha$ is singular, the code is degenerate: different errors can act identically on the codespace (their difference annihilates it). Degeneracy is a uniquely quantum phenomenon with no classical analogue, and we devote a later lesson to it.

The takeaway

A code corrects an error set $\{E_a\}$ exactly when $P E_a^\dagger E_b P = \alpha_{ab} P$ . The off-diagonal part keeps logical codewords distinguishable after errors; the diagonal part — equal for all codewords — guarantees the errors leak no logical information. Everything else in this module is a corollary of these conditions.