The Zcash ArboretumI–IX The Complete Arboretum
Contents
PDF
Theme

6 Confirmation requirements

Theorem 5.2 reduces the double-spend success probability to the function r(q,n) of the attacker’s share and the confirmation count, tabulated in Table 1 and plotted in Figure 3. This section tabulates the confirmation counts required for three risk targets, proves that no finite count has zero risk (Proposition 6.1) and that the required count diverges as q approaches 12 (Proposition 6.2), and converts confirmation counts into wall-clock time at the Zcash block spacing.

q r<10% r<1% r<0.1%
2% 1 2 3
5% 2 3 4
8% 2 3 5
10% 2 4 6
15% 3 6 9
20% 4 8 13
25% 5 12 20
30% 9 20 32
35% 15 36 58
40% 34 82 133
45% 135 331 539
Table 2: The least number of confirmations for which r(q,n) falls below each of three targets, computed by script. The count grows logarithmically in the inverse target risk and without bound as q approaches 50%: at the strictest target a 10% attacker requires 6 confirmations, a 45% attacker 539, and a 50% attacker no finite number.
Proposition 6.1 (No finite count is final).

For every 0<q<p and every n, r(q,n)≥2qn>0.

Proof.

By Theorem 5.2, r=2Pr[An], and one way for the attacker to win the first-to-n race is to find the first n blocks outright, an event of probability qn. □

Proposition 6.2 (Divergence at the majority line).

Fix a target 0<ε<1. For 0<q<12, let nε(q) be the least n with r(q,n)<ε. Then nε(q)→∞ as q↑12.

Proof.

First, nε(q) exists. The event An that the attacker reaches n blocks first is the event that at least n of the first 2n−1 block discoveries are the attacker’s. Since q<p, each such length-(2n−1) sequence has probability at most qnpn−1, and there are fewer than 22⁢n−1 sequences. Hence

r⁢(q,n)=2⁢P⁢r⁢[An]≤(4⁢p⁢q)np⟶0,

because 4pq<1.

For fixed n the function q↦r(q,n) is a polynomial on [0,12], hence continuous, and r(12,n)=1: at q=p each half of the race of Remark 5.3 has probability exactly 12 by Lemma 3.6 and symmetry. Given any N, for each 1≤n≤N there is a δn>0 with r(q,n)>ε whenever q>12−δn. Taking δ=min1≤n≤Nδn>0 excludes every n≤N throughout that neighbourhood, so nε(q)>N. Since N was arbitrary, the claimed divergence follows. □

Remark 6.3 (Nothing is special about six).

The paper (§5): “There is nothing special about the default, often-cited figure of 6 confirmations. It was chosen based on the assumption that an attacker is unlikely to amass more than 10% of the hashrate, and that a negligible risk of less than 0.1% is acceptable. Both these figures are arbitrary” — six confirmations are “overkill for casual attackers, and at the same time powerless against more dedicated attackers”. Table 2 replaces that figure: a policy fixes the attacker share to be withstood and the acceptable risk, and the table gives the corresponding n. The pair (q,ε) is chosen by policy; the model does not determine it.

6.1 Confirmation times at 25-second spacing

By Corollary 4.7, the success probability depends on the confirmation count, not on elapsed time. The wall-clock duration of a count is set by the target spacing: 25 seconds in ZIP-218. While the attacker mines in private, only the honest share of the hashrate extends the public chain, and the merchant’s wait lengthens accordingly.

Proposition 6.4 (Expected wait for n confirmations).

Under Assumptions 4.1 and 4.2, let the attacker withhold all his blocks from the public chain, and let the first public block found after a transaction’s broadcast include it. The time Sn from the broadcast to the transaction’s n-th confirmation has expectation

𝔼⁢[Sn]=n⁢T0p.
Proof.

Only honest blocks extend the public chain, and the honest network’s clock has rate λ=p/T0 (Proposition 3.4). At the broadcast, and after each attacker block, that clock has not fired, so by Lemma 3.2 its remaining wait has the law of a fresh clock. The restarts of Proposition 3.4 therefore leave the honest blocks, counted from the broadcast, the arrivals of an exponential clock of rate λ in the sense of the Math Guide, §“Exponential clocks, Poisson counts, and the Chernoff method”. By Definition 2.3 the n-th confirmation is the n-th arrival, whose time has density fn(s)=λnsn−1e−λ⁢s/(n−1)! for s≥0 by that section’s lemma “Arrival-time density”. Since sfn(s)=(n/λ)fn+1(s) and fn+1 integrates to 1,

𝔼⁢[Sn]=∫0∞s⁢fn⁢(s)⁢𝑑s=nλ⁢∫0∞fn+1⁢(s)⁢𝑑s=n⁢T0p.∎

Table 3 evaluates Proposition 6.4 for every count of Table 2, against the attacker share for which that count was computed, taking T0=25 seconds, the spacing produced when the full hashrate publishes. The nominal wait nT0 understates each entry by the factor p: the 539 confirmations required against a 45% attacker take a nominal 3.74 hours, but 6.81 hours in expectation while that attacker withholds. A wait of hours spans many difficulty-averaging windows (Section 8.2), over which retargeting would lower the public chain’s difficulty; the proposition rests on Assumption 4.2 and omits that effect. Three confirmations take a nominal 75 seconds and ten take 250 seconds; a 10% withholding attacker divides these waits by 0.9.

r<10% r<1% r<0.1%
q n min n min n min
2% 1 0.4 2 0.9 3 1.3
5% 2 0.9 3 1.3 4 1.8
8% 2 0.9 3 1.4 5 2.3
10% 2 0.9 4 1.9 6 2.8
15% 3 1.5 6 2.9 9 4.4
20% 4 2.1 8 4.2 13 6.8
25% 5 2.8 12 6.7 20 11.1
30% 9 5.4 20 11.9 32 19.0
35% 15 9.6 36 23.1 58 37.2
40% 34 23.6 82 56.9 133 92.4
45% 135 102.3 331 250.8 539 408.3
Table 3: The confirmation counts n of Table 2 and their expected waits in minutes, nT0/p with T0=25 seconds by Proposition 6.4, while an attacker of share q withholds his blocks; computed by script.

A nominal thirty-minute wait yields 72 confirmations at 25-second spacing and three at Bitcoin’s 600-second spacing. The corresponding strict-overtake probabilities against a 10% attacker are approximately 9.35×10−34 and 1.712%, respectively; either count takes 30/0.9≈33.3 minutes in expectation under withholding. These are model calculations, not guarantees about either network.

The comparison holds q fixed. The model contains no network, so it omits the honest work lost when honest blocks race one another during propagation; quantifying that loss requires a propagation distribution and a mining topology. A policy that waits for a fixed amount of accumulated miner reward measures a different quantity: if the reward per block decreases in proportion to the target spacing, reaching that amount requires proportionally more blocks, and, ignoring transaction fees (value paid for transaction inclusion) and rounding, the spacing cancels out of the mean wait. The monetary and fixed-hashrate models are therefore not interchangeable.

Remark 6.5 (Limits of the model).

Three assumptions of the model fail on the deployed chain and limit how far the calculations transfer to it. (i) Difficulty is not constant: Zcash retunes the target after every block (Section 8.2). The race is scored in work, not block counts; the share q approximates the attacker’s share of work production while the branches’ block weights remain close, and once their difficulty schedules diverge, the equal-step random-walk formula is no longer exact. (ii) Hashrate is not constant: the share q in the theorems is the share the attacker sustains for the duration of the attack; the paper’s caveat is that T0 re-enters only if the attacker cannot sustain his hashrate long enough, which it judges unlikely for a serious attacker. (iii) Propagation is not instantaneous: honest blocks found concurrently compete during propagation, and the work of the losing block is wasted, which raises the attacker’s effective share; the size of this effect is not quantified here. A fourth limitation lies outside the model: the attacker of this volume obeys the protocol’s validity rules and attacks only the choice between valid histories. Other volumes treat attacks on the rules themselves.