Correctable Errors

2.1 Where the environment gets information

2.1.1 Inspect the entire environmental state

Let P be the projector onto a code subspace, and let a noise channel have Kraus operators Eₐ. A dilation writes its action as \(V|\psi\rangle=\sum_aE_a|\psi\rangle|a\rangle_E\), with \(\sum_aE_a^\dagger E_a=I\). The environment label is an orthonormal basis used to describe the dilation; it need not correspond to a detector that anyone actually reads.

After discarding the output system, the environment’s matrix entry in row a and column b is \(\langle\psi|E_b^\dagger E_a|\psi\rangle\). Notice the off-diagonal entries. It is not enough that the probabilities of the individual error labels are independent of the input. The environment can also carry input-dependent coherence between labels.

If exact recovery restores an arbitrary reference-entangled logical state, the recovered pure logical–reference pair must factor from the environment. The environment can therefore contain no record of the logical state. Requiring its matrix entries to be constant for every vector in the code gives

Why does constancy of expectation values imply an operator identity? Test basis vectors first, then their normalized sums and sums with relative phase i. These reveal the diagonal, real off-diagonal, and imaginary off-diagonal matrix entries. A quadratic form that is the same constant on every normalized code vector must be that constant times identity on the code.

This is the Knill–Laflamme condition. The necessity argument uses environmental independence; the sufficiency construction follows next. The original theorem is developed in Knill and Laflamme, A Theory of Quantum Error-Correcting Codes.

2.2 Constructing a recovery

2.2.1 Separate error spaces, then reverse their isometries

Suppose the condition holds. The matrix c is positive semidefinite: for coefficients vₐ, the quadratic form \(\sum_{a,b}v_a^*c_{ab}v_b\) equals the squared norm of \(\sum_bv_bE_b|\psi\rangle\) for any normalized code state. We can therefore diagonalize c and choose linear combinations Fᵣ of the errors such that \(PF_r^\dagger F_sP=d_r\delta_{rs}P\), with \(d_r\ge0\).

For each nonzero dᵣ, the map \(W_r=F_rP/\sqrt{d_r}\) is an isometry on the code: it preserves all inner products there. Different r have orthogonal images. This gives an explicit recovery strategy. Measure which image subspace contains the state, then apply the inverse isometry on that subspace. Each branch returns the same logical state. A completion on the unused orthogonal space makes the recovery a channel on the entire physical space.

For the repetition code and errors I,X₁,X₂,X₃, we already met these orthogonal images. They are the spans of {000,111}, {100,011}, {010,101}, and {001,110}. Cross products compress to zero, while each error’s squared product compresses to P. The abstract construction is exactly the syndrome measurement and conditional correction from Chapter 1.

It can happen that distinct physical errors have the same action on the code. Their error images then overlap before diagonalizing c; some eigenvalues dᵣ may be zero. This is called degeneracy. Recovery need not determine which microscopic error occurred. It only needs to undo its action on the information we stored. Trying to distinguish equivalent errors can add an unnecessary task.

2.3 From discrete errors to continuous noise

2.3.1 Linearity protects an error span

Real noise is not obliged to choose one Pauli operator from a menu. A small rotation may act coherently: \(e^{-i\theta X/2}=\cos(\theta/2)I-i\sin(\theta/2)X\). Does a code that corrects I and X also correct this rotation? Yes, if the correction condition holds for their entire error set. The channel’s Kraus operators can be linear combinations of the correctable errors.

To see why, write \(F_\mu=\sum_a f_{\mu a}E_a\). Then \(PF_\mu^\dagger F_\nu P=\sum_{a,b}f_{\mu a}^*f_{\nu b}c_{ab}P\), again a scalar times P. The code corrects every channel whose Kraus operators lie in that span. This is an exact algebraic statement; it does not require pretending the environment sampled a classical Pauli error.

Every operator on one qubit is a complex linear combination of I,X,Y,Z. Consequently, correcting these Pauli errors on each possible single position suffices to correct an arbitrary channel acting on one unknown position. The condition must include cross products between errors at different positions. Correcting X alone is not enough: on the repetition code, PZ₁P acts as logical Z and is not proportional to P.

There is also an important boundary. Independent noise on every qubit produces Kraus products of many weights. A code correcting all weight-one errors does not exactly correct every such product. It suppresses the leading contribution only under an appropriate noise model. Later, fault tolerance will use that distinction to organize which collections of faults are dangerous.