Ultrametric Geometry as Passive Quantum Error Correction: Bruhat–Tits Qudit Codes and the Non-Archimedean Bound

In the paper's words

We formalize a proposal for passive quantum error correction (QEC) in which logical qudits are encoded on the vertices of a ball of the Bruhat–Tits tree T_p+1, the incidence geometry of the p-adic numbers Q_p, with Hecke-type operators acting as the natural gate set. The central observation is that the non-Archimedean inequality \|x+y\|_p ≤ max(\|x\|_p,\|y\|_p) forces the tree's balls to be either nested or disjoint, so localized errors confined to one ball cannot accumulate coherently across distinct branches. We prove a nesting lemma, use it to show that every error set of weight ≤ t on a protected ball is correctable under the Knill–Laflamme conditions, and derive the code parameters n_t = 1 + (p+1)(p^t-1)/(p-1) and $d = 2t+1$.

The abstract, verbatim, and its opening sentence is the claim. Every line below is quoted from the text.

Numbers it reports

1. A QNFO entry benchmarks 17 qubit-based platforms against a qudit architecture on a joules-per-solution energy metric, noting all 17 use two-level systems [10].

   from 2. Background and Related Work

2. Under independent depolarizing noise at physical error rate $\varepsilon = 10^{-3}$ we compute, with full arithmetic, a passive logical failure probability $P_{\mathrm{pass}} \approx 9.98\times 10^{-6}$, a suppression factor of $\approx 100$ over an unencoded qudit, and, as an explicitly labeled projection, $\approx 4.08$ better than an active distance-3 baseline with noisy syndrome extraction.

   from Abstract

3. Suppression factor relative to the unencoded qudit (failure probability $\varepsilon = 10^{-3}$): $$\frac{\varepsilon}{P_{\mathrm{pass}}} = \frac{10^{-3}}{9.98 \times 10^{-6}} \approx 100.2.$$ Case $p=3$, $t=2$ ($n_2 = 17$, $d = 5$).

   from 4.3 Passive failure probability (exact computation)

4. Assumptions (stated explicitly; this is a projection, not a measurement): (A1) the active baseline is a generic distance-3 code on the same $n_1 = 5$ data qudits; (A2) each protection epoch requires one syndrome-extraction round in which every data qudit participates in at least one gate/measurement interaction with additional error rate $\varepsilon_m$; (A3) errors combine as $1 - (1-\varepsilon)(1-\varepsilon_m)$ per qudit; (A4) we take the pessimistic-for-us-but-standard case $\varepsilon_m = \varepsilon = 10^{-3}$.

   from 4.4 Active baseline comparison (labeled projection)

5. Step 5: $P_{\mathrm{act}} = 1 - 0.990045096 - 0.009914240 = 0.000040664$, i.e. $$P_{\mathrm{act}} \approx 4.07 \times 10^{-5}.$$ Ratio: $$\frac{P_{\mathrm{act}}}{P_{\mathrm{pass}}} = \frac{4.07 \times 10^{-5}}{9.98 \times 10^{-6}} \approx 4.08.$$ So, under assumptions (A1)–(A4), the passive scheme yields a factor $\approx 4.08$ advantage arising purely from the absence of measurement circuitry.

   from 4.4 Active baseline comparison (labeled projection)

6. Linear interpolation: $$\varepsilon_{\mathrm{th}} \approx 0.32 + 0.01 \times \frac{1.00618 - 1}{1.00618 - 0.99252} = 0.32 + 0.01 \times \frac{0.00618}{0.01366} = 0.32 + 0.01 \times 0.4524 = 0.3245.$$ The pseudo-threshold under the dominant-term approximation is therefore $\varepsilon_{\mathrm{th}} \approx 0.3245$; this is a projection, since the dominant-term approximation ignores higher-order terms of $P_{\mathrm{pass}}(\varepsilon)$.

   from 4.5 Pseudo-threshold (projection)

How it is done

1. We prove a nesting lemma, use it to show that every error set of weight $\leq t$ on a protected ball is correctable under the Knill–Laflamme conditions, and derive the code parameters $n_t = 1 + (p+1)(p^t-1)/(p-1)$ and $d = 2t+1$.

   from Abstract

2. First, we formalize "ultrametric geometric QEC": we define the code, the error model, and the correctable error sets, and we prove a nesting lemma from which correctability of all weight-$\leq t$ errors follows.

   from 1. Introduction

3. The continuous-time quantum error correction (CTEC) approach treats both noise and correction as processes continuous in time, based on continuous weak measurements and feedback, analyzed from the point of view of the subsystem principle [1].

   from 2. Background and Related Work

4. The defining combinatorial fact, used repeatedly, is the nesting lemma : Lemma 1 (nesting/disjointness).

   from 3.1 The geometry

What would falsify it

1. All performance figures are analytic projections under stated assumptions, not simulations; limitations and falsification criteria are discussed.

   from Abstract

2. The central claims would be falsified if: (a) an explicit encoder on $B_t(v_0)$ fails the Knill–Laflamme conditions for some weight-$\leq t$ Weyl error set, contradicting Proposition 1; (b) two distinct equal-radius balls of $T_{p+1}$ overlap without coinciding, contradicting Lemma 1; or (c) a fault-tolerant simulation of the code under the stated noise model yields logical failure rates inconsistent with the binomial computation of Section 4.3.

   from 6. Discussion

3. The scheme's promise is real but conditional on constructing an explicit symmetric encoder; we have stated the assumptions, the failure modes, and the falsification criteria plainly so that the proposal can be tested rather than taken on trust.

   from 7. Conclusion

Where the claim stops

1. Proposition 1, however, is conditional: the branch-permutation symmetry of the code states is a design requirement, and no explicit encoder (state vectors, stabilizer generators, or logical operators) is constructed in this paper.

   from 6. Discussion

2. If no symmetric encoder exists for a given ball, the weight-$\leq t$ correctability claim and the distance $d = 2t+1$ fail.

   from 6. Discussion

3. (iii) The active-baseline comparison rests on assumptions (A1)–(A4), in particular the pessimistic choice $\varepsilon_m = \varepsilon$ and the crude combination rule of (A3); a real baseline may have much smaller $\varepsilon_m$, which would shrink the $\approx 4.08$ factor toward $1$, or repeated measurement rounds per epoch, which would grow it.

   from 6. Discussion

Where to read

1. Abstract205 words · 1 min
2. 1. Introduction400 words · 2 min
3. 2. Background and Related Work785 words · 3 min
4. 3. Methods511 words · 2 min
5. 4. Analysis892 words · 4 min
6. 5. Results200 words · 1 min
7. 6. Discussion380 words · 2 min
8. 7. Conclusion145 words · 1 min