5.7. Error decomposition proof 125
call a nonempty subsetSofEacandidate level-0 error ifS consists of either phase flip errors or XXXX measurement errors, but not both, and if it is contained in a spacetime box [x, x+ 1]×[y, y+ 1]×[t, t]. S will be an actual level-0 error if S andE\S are (a, a, b)-separated, whereaand bare fixed positive integers. The union of all actual level-0 errors is called level-0 noise, which we denote by E0.
Now we proceed to define candidate and actual level-n errors inductively. For n >0, suppose that candidate level-kerrors, actual level-kerrors, and level-knoise Ek have been defined for k < n.
A nonempty set S⊆E\En−1 is a candidate level-n error if (i) S is contained in a spacetime box of sizeQn×Qn×Un, and (ii) S contains at least two disjoint candidate level-(n-1) errors that are
(aQn−1, aQn−1, bUn−1)-linked.
S will be an actual level-n error if, additionally,
(iii) S does not contain two candidate level-nerrors that are (4(a+ 2)Qn−1,4(a+ 2)Qn−1,4(b+ 2)Un−1)-separated, and (iv) S andE\(S∪En−1) are (aQn, aQn, bUn)-separated.
Level-nnoiseEn is defined as the union ofEn−1 and all actual level-nerrors inE.
We will further define a candidate level-n error S as minimal if S contains exactly 2n points. Now we will proceed to show by induction that every point in E\En−1 belongs to at least one minimal candidate level-nerrorS. The base case of n = 1 is true provided that Q ≥ a and U ≥ b (so that two points that are (a, a, b)-linked fit inside a box of sizeQ×Q×U).
For the inductive step, let us assume that the hypothesis holds true for (n−1).
Let ~x1 ≡ (x1, y1, t1) be a spacetime point that is a member of E\En−1; we will show that this point must belong to a minimal candidate level-nerror. The point must be a member ofE\En−2, so by the inductive hypothesis it belongs to some minimal candidate level-(n-1) errorS1. The minimality implies thatS1 consists of
5.7. Error decomposition proof 126
two disjoint minimal candidate level-(n-2) errors that are (aQn−2, aQn−2, bUn−2)- linked. Therefore,S1 must fit into a box of size (a+ 2)Qn−2×(a+ 2)Qn−2×(b+ 2)Un−2. Now, if any point inS1 were part of an actual level-(n-1) error, then all of S1 would be contained in that actual level-(n-1) error. But~x1∈S1and~x1 ∈/En−1, which means that S1 must not be contained in any actual level-(n-1) error. Since S1 satisfies condition (iii), it must not satisfy condition (iv). Then there exists some ~x2 ∈ E \En−1 such that ~x2 ∈/ S1 and ~x2, S1 are (aQn−1, aQn−1, bUn−1)- linked. Now, it is true that ~x2 must belong to the set of points in E\En−1 that are not in S1 such that~x2, ~x1 are ((a+ 1)Qn−1,(a+ 1)Qn−1,(b+ 1)Un−1)-linked.
The inductive hypothesis says that there must be a minimal candidate level-(n-1) errorS2 containing~x2.
We can also choose S2 such that S1 and S2 are disjoint. There are two cases to be considered. The first case is to suppose that there exists some~x2 such that
~
x2 is 2(a+ 2)Qn−2-separated from ~x1. Then the size restrictions on S1 and S2 force them to be disjoint. The second case is to suppose that the first case does not hold; i.e., there does not exist such an~x2. Then at worstS≡S1∪S2 fits in a box of size 4(a+ 2)Qn−2×4(a+ 2)Qn−2×4(b+ 2)Un−2 which is ((a+ 1)Qn−1− 4(a+ 2)Qn−2,(a+ 1)Qn−1−4(a+ 2)Qn−2,(b+ 1)Un−1−4(b+ 2)Un−2)-separated from the rest ofE\En−1. SinceQ≥4(a+ 2) andU ≥4(b+ 2),S satisfies (i) and (iv) in the definition of an actual level-(n-1) error. But S also satisfies (ii) (since it containsS1) and (iii). Then S must be an actual level-(n-1) error. However, we have already shown thatS1 cannot be in an actual level-(n-1) error, so the second case leads to a contradiction and cannot occur; then S1 and S2 can be chosen to be disjoint, and the set S is a minimal candidate level-nerror containing ~x1.
This argument actually allows us to place more restrictive bounds on level- n errors. By replacing (n−1) with n in the argument above, we can see that a minimal candidate level-n error S1 fits into a box of size (a+ 2)Qn−1 ×(a+ 2)Qn−1×(b+2)Un−1. Also, given a point~xinE\En−1that is (2(a+2)Qn−1,2(a+
2)Qn−1,2(b+2)Un−1)-separated from at least one member ofS1, the point~xlies in a second minimal candidate level-n errorS2 that is disjoint from S1. If S1, S2 are
5.7. Error decomposition proof 127
(4(a+ 2)Qn−1,4(a+ 2)Qn−1,4(b+ 2)Un−1)-separated and also (aQn, aQn, bUn)- linked, we can see from conditions (iii) and (iv) thatS1 andS2cannot be part of an actual level-nerror. Thus, ifS1 is part of a actual level-nerror, any point inS\S1 must be (5(a+2)Qn−1,5(a+2)Qn−1,5(b+2)Un−1)-linked withS1. This means that the entire error will fit into a box of size 7(a+2)Qn−1×7(a+2)Qn−1×7(b+2)Un−1).
We can take the intersection of this box with the bounding box of sizeQn×Qn×Un from condition (i). This reveals that an actual level-n error fits inside a box of size min{Q,7(a+ 2)}Qn−1×min{Q,7(a+ 2)}Qn−1×min{U,7(b+ 2)}Un−1. We will use this result in the proof demonstrating correction of higher level errors in Section 5.8.2.
Finally, let us define the level-n error raten as the probability that a box of sizeQn×Qn×Un has nonempty intersection with at least one candidate level-n error. A spacetime box of unit size cannot include more than four edges and four vertices of the lattice, so clearly 0 ≤ 4(p+q), where p and q are the phase flip and measurement error probabilities per time step. An upper bound on n can be found by considering the probability that a box of sizeQn×Qn×Unhas nonempty intersection with at least one candidate level-(n-1) error. This probability must be bounded above by Q2U n−1. Gray asserts that by the Kesten-Vandenberg inequality, since each candidate level-nerror is composed of at least two candidate level-(n-1) errors,
n≤(Q2U n−1)2. (5.1)
Then we can see that
n≤(Q2U)2n+1−2(4(p+q))2n <(Q4U24(p+q))2nforn >0 . (5.2) That is to say, level-nerrors get progressively more rare (in fact, double exponen- tially) as ngets larger, as long as
(p+q)< Q−4U−2/4. (5.3)