Binary Structure and Statistical Dynamics of Collatz Stopping Times

Abstract

We present a statistical analysis of Collatz stopping times for all positive integers up to (N = 10^7), focused not on verification of the conjecture but on the rich distributional structure of stopping times as a function of binary seed representation. Our principal finding is that the number of trailing 1-bits in the binary representation of the seed governs the mean stopping time with a per-bit increment of approximately 6.15 steps. When seeds are grouped by exact trailing 1-bit count, this linearity persists through at least (k = 11), substantially further than previously documented. We decompose the 6.15-step increment into two components: 2.00 forced iterations from the deterministic odd–even pair, and 4.18 additional iterations reflecting a persistent downstream effect on the remaining trajectory. This persistence is demonstrated directly: even after removing the (2k) forced steps from each trajectory, the remaining stopping time still increases by approximately 4.18 steps per trailing 1-bit, proving that the initial bit pattern reshapes the trajectory far beyond the deterministic prefix. We characterize the distribution of stopping times as approximately log-normal with negative skewness ((-0.37)) and platykurtosis ((-0.19)) in log-space, and introduce trajectory entropy — the fraction of odd steps — as a new diagnostic, finding a mean of 0.322 with a tightly concentrated distribution centered near the theoretical prediction of (1/3). Sequential autocorrelation at lag 1 is (r = 0.402), decaying slowly. The maximum stopping time grows as (C \cdot (\ln N)^2) with (C \approx 2.64). A single-file Python script using only the standard library accompanies the paper, enabling full reproduction of every reported statistic.

1. Introduction

The Collatz conjecture, formulated by Lothar Collatz in 1937, concerns the iteration of the map

[ T(n) = \begin{cases} n/2 & \text{if } n \equiv 0 \pmod{2}, \ 3n+1 & \text{if } n \equiv 1 \pmod{2}, \end{cases} ]

applied to any positive integer (n). The conjecture asserts that the orbit of every positive integer eventually reaches 1. Despite its elementary statement, the problem remains open and is widely regarded as one of the most intractable in number theory. Erdős remarked that "mathematics may not be ready for such problems."

The total stopping time (\sigma(n)) is the number of iterations required to first reach 1. Terras (1976) introduced the notion of stopping time and proved that almost all integers have finite stopping times. Lagarias (1985) provided an extensive survey connecting the problem to ergodic theory, Markov chains, and computability. Tao (2022) proved that almost all Collatz orbits attain almost bounded values, the strongest density result to date.

Computational verification of the conjecture — confirming that every integer up to some bound eventually reaches 1 — has been carried to extraordinary scales. Oliveira e Silva and Herzog (2012) verified the conjecture up to (5.764 \times 10^{18}), and Roosendaal (2020) maintains an extensive database of record-setting delay values. Crary (2012) has documented record delay values for the problem.

The present paper has a fundamentally different aim from these verification efforts. We do not seek to extend the frontier of confirmed convergence; instead, we investigate the statistical structure of stopping times as a function of the seed's binary representation. Prior large-scale computations focused primarily on verification (confirming convergence to 1) or on identifying record-setting trajectories. By contrast, we ask: how precisely do the trailing bits of a seed determine its expected stopping time? Does this influence persist beyond the deterministic prefix of the trajectory? What is the entropy structure of typical orbits? These statistical questions can be meaningfully addressed at (N = 10^7), a scale that provides millions of samples per residue class and sufficient statistical power for the analyses we undertake. The quantitative findings — precise constants, persistence effects, distributional moments — are not accessible from the existing verification literature, which reports convergence outcomes rather than distributional statistics.

Our principal contributions are: (1) demonstration that the per-bit stopping-time increment of approximately 6.15 steps maintains near-perfect linearity through at least (k = 11) trailing 1-bits under exact grouping; (2) a decomposition of this increment into 2.00 forced steps and 4.18 persistent downstream steps, with the persistence verified by direct trajectory analysis; (3) the introduction of trajectory entropy as a diagnostic for orbit randomness; and (4) quantitative characterization of the distributional shape, autocorrelation structure, and growth of extremes.

2. Data and Methods

2.1 Computation

We computed the total stopping time (\sigma(n)) for every integer from 1 to (N = 10{,}000{,}000) by direct iteration with memoization. A lookup table of previously computed stopping times avoids redundant computation of shared orbit tails. All statistics were computed in Python using only the standard library, ensuring complete reproducibility without dependency management.

2.2 Trailing 1-Bit Grouping

For each integer (n), the number of trailing 1-bits is the largest (k) such that the lowest (k) bits of (n) are all 1. Equivalently, for odd (n), this is (\nu_2(n+1)), the 2-adic valuation of (n+1); for even (n), the count is zero. Seeds are grouped by exact trailing 1-bit count. This grouping is distinct from the residue-class partition modulo (2^k), in which seeds satisfying (n \equiv 2^k - 1 \pmod{2^k}) include all seeds with at least (k) trailing 1-bits — a conflation that inflates the apparent per-bit increment at the highest reported (k).

2.3 Persistence Analysis

To measure whether the trailing-bit effect extends beyond the deterministic prefix, we trace each trajectory through its first (2k) forced steps (where (k) is the exact trailing 1-bit count) and record the remaining stopping time — the number of additional iterations from the post-forced iterate to 1. If trailing bits influenced only the forced prefix, the remaining stopping time would be independent of (k).

2.4 Trajectory Entropy

For each (n \geq 2), we compute the fraction of odd steps in the trajectory from (n) to 1. In the standard stochastic model, each iterate is odd with probability (1/2), but each odd step always produces an even successor, so the expected fraction of odd steps is approximately (1/3). The distribution of this fraction across seeds characterizes the effective randomness of orbits.

2.5 Additional Analyses

We also compute: moments of (\ln(\sigma(n))) as descriptive statistics (noting that at (N = 10^7) any formal normality test has extreme statistical power, rendering its decision uninformative; the moments themselves are the quantities of interest); one-way ANOVA (F)-statistics for residue classes modulo (2^k); sequential autocorrelation at multiple lags; growth of the maximum stopping time at powers of 10; and record-breaking stopping times.

3. Results

3.1 Overview

Table 1. Summary statistics for Collatz stopping times, (1 \le n \le 10^7).

3.2 Trailing 1-Bit Structure

Our principal empirical finding is that the mean stopping time, partitioned by exact trailing 1-bit count, follows a strikingly linear relationship that extends well beyond the first few bits.

Table 2. Mean stopping time by exact trailing 1-bit count.

The per-bit increment is stable at approximately 6.15 steps across the entire range (k = 1) through (k = 11), with a standard deviation of only 0.28 steps and a mean of 6.24. The consistency of this increment across more than an order of magnitude in sample size (from 2.5 million seeds at (k = 1) to 2,441 at (k = 11)) supports the hypothesis that the linear relationship is a genuine structural property rather than a finite-sample artifact.

Remark on grouping. The distinction between exact and at-least grouping is consequential. When seeds are grouped by residue class modulo (2^k), the class (n \equiv 63 \pmod{64}) includes all seeds with six or more trailing 1-bits. The apparent mean for that class (192.79) is inflated by the contributions of seeds with (k = 7, 8, 9, \ldots), whose higher stopping times pull the average upward. This conflation produces an apparent jump in the final reported increment that is an artifact of the grouping method, not an anomaly in the underlying structure. With exact grouping, the (k = 6) increment (+6.53) is entirely consistent with the linear trend.

3.3 Mechanistic Decomposition of the 6.15-Step Increment

When a positive integer has (k) trailing 1-bits, the Collatz map is forced into a deterministic sequence: each trailing 1-bit guarantees an odd iterate, triggering (3n+1), which always produces an even result that is then halved. This yields (k) paired operations, contributing exactly (2k) forced iterations and multiplying the iterate by approximately ((3/2)^k).

The per-bit increment of 6.15 steps decomposes as follows:

1. Forced steps (2.00 per bit). Each trailing 1-bit contributes one (3n+1) step and one (\div 2) step to the stopping time directly.

2. Persistent downstream cost (4.18 per bit). The (3/2) magnification per forced pair does not merely add forced steps; it inflates the iterate, increasing the expected length of the remaining trajectory. In Terras's geometric model, iterates are multiplied by (3/2) when odd (probability (1/2)) and by (1/2) when even, giving a geometric mean shrinkage factor. Starting from a value inflated by (3/2), the additional expected stopping time is (\ln(3/2)/\ln(4/3) \approx 1.41) steps, predicting a total per-bit cost of (2.00 + 1.41 = 3.41).

The observed 6.15 exceeds this prediction by a factor of approximately 1.80. The discrepancy arises because forced (3n+1) operations do not merely inflate magnitude — they restructure the iterate's bit pattern through carry propagation. After (k) forced odd–even pairs, the iterate's low-order bits are no longer uniformly distributed; the carries introduced by the (3n+1) operations create correlations that increase the probability of further odd iterates in the subsequent trajectory. The forced steps thus bias the downstream parity sequence toward more odd steps than the independent-coin-flip model predicts, amplifying the downstream cost by a factor of approximately 1.8. This reveals that the simple additive model — in which trailing bits affect only the forced prefix — captures barely one-third of the total effect.

3.4 Bit-Pattern Persistence

To verify the downstream persistence directly, we computed the remaining stopping time after removing the (2k) forced steps from each trajectory.

Table 3. Remaining stopping time after forced steps, by exact trailing 1-bit count.

If trailing bits influenced only the forced prefix, the "Mean remaining" column would be constant across all (k). Instead, it increases by approximately 4.18 steps per trailing 1-bit (mean over (k = 1, \ldots, 10)). This persistent excess demonstrates that the initial bit pattern reshapes the trajectory far beyond the deterministic prefix: more than two-thirds of the 6.15-step total increment per bit is attributable to this downstream effect rather than to the forced steps themselves.

The decomposition is summarized as:

3.5 Trajectory Entropy

The fraction of odd steps in each trajectory provides a measure of orbit regularity.

Table 4. Trajectory entropy statistics, (n = 2, \ldots, 10^7).

The distribution is unimodal and concentrated, with 95% of values falling between 0.27 and 0.37. The observed mean of 0.322 is 3.4% below the theoretical prediction of (1/3) from the independent-parity model. This deficit is consistent with the negative autocorrelation in the parity sequence: the (3n+1) operation on an odd number always produces an even result, guaranteeing the next step is even and biasing the aggregate odd fraction below (1/3).

The tightness of the distribution (coefficient of variation 9.0%) confirms that individual trajectories, despite their apparent irregularity, conform closely to the ensemble average — a manifestation of the quasi-ergodic behavior underlying heuristic models (Terras, 1976; Lagarias, 1985; Tao, 2022).

3.6 Residue-Class Analysis

Table 5. ANOVA (F)-statistics for residue classes modulo (2^k).

All (p)-values are effectively zero. The decreasing (F)-statistic with increasing modulus reflects finer partitioning into more groups with smaller within-group sample sizes, not a weakening of the effect. The quantitative range at modulus 64 is informative: mean stopping times range from 118.13 (for (n \equiv 0 \pmod{64}), multiples of 64 that undergo six immediate halvings) to 192.79 (for (n \equiv 63 \pmod{64}), seeds with at least six trailing 1-bits). This spread of 74.66 steps — nearly half the overall mean — far exceeds the (2 \times 6 = 12) forced iterations for six trailing bits. The difference of 62.66 steps arises from the persistent downstream influence quantified in Section 3.4. The ANOVA results thus do not merely confirm a trivially expected deterministic effect; they quantify a persistent influence that extends well beyond the forced prefix.

3.7 Growth of Maximum Stopping Time

Table 6. Maximum stopping time and the ratio (M(N)/(\ln N)^2) at powers of 10.

The ratio stabilizes in the range 2.6–2.7 for (N \ge 10^5), consistent with the heuristic prediction of (O((\ln N)^2)) growth (Lagarias, 1985). Our empirical estimate of the asymptotic constant is (C \approx 2.64). The slight non-monotonicity at (N = 10^6) (ratio 2.745 vs. 2.641 at (10^5)) is consistent with the discreteness of record-breaking events.

3.8 Distribution Shape

The distribution of stopping times is unimodal and right-skewed in natural scale. In log-space:

Table 7. Descriptive moments of (\ln(\sigma(n))) for (1 \le n \le 10^7).

The negative skewness in log-space indicates a left-tail excess, while the negative excess kurtosis indicates platykurtosis — thinner tails than a Gaussian. The departures are modest in absolute terms, so the log-normal remains a useful first approximation. At (N = 10^7), any formal normality test has overwhelming statistical power, so we report these moments as descriptive statistics rather than formal test outcomes. The departures from Gaussianity are consistent with non-independence of successive parity choices, as the forced bit structure introduces correlations that skew and thin the distribution.

3.9 Sequential Autocorrelation

Table 8. Autocorrelation of stopping times at selected lags.

The lag-1 autocorrelation of 0.402 is substantial and reflects a well-understood mechanism: the orbits of (n) and (n+1) frequently merge within a few steps, sharing a common trajectory to 1 (Chamberland, 1996). The slow decay — still 0.100 at lag 100 — indicates that trajectory sharing persists over moderate ranges of the integers.

3.10 Odd vs. Even Seeds

Odd seeds have a mean stopping time of 161.46 (std 62.61), compared to 149.09 (std 60.26) for even seeds, a difference of 12.37 steps ((t = 318.29)). This corresponds closely to two trailing 1-bit increments ((2 \times 6.15 = 12.30)), consistent with the fact that odd numbers have at least one trailing 1-bit.

3.11 Record-Breaking Stopping Times

We identified 53 record-breaking stopping times in (1 \le n \le 10^7). The last record occurs at (n = 8{,}400{,}511) with (\sigma = 685). The growth ratio (record stopping time / (\ln n)) increases from 2.39 for early records to 3.29 for late records, suggesting superlogarithmic growth consistent with the known sparsity of delay records (Crary, 2012).

4. Discussion

4.1 The Linearity Phenomenon

The most striking finding of this study is the precise linear dependence of mean stopping time on trailing 1-bit count, maintained at (6.15 \pm 0.28) steps per bit from (k = 1) through (k = 11). While it is well known that the low-order bits of a seed determine the first several iterates (Lagarias, 1985), the quantitative stability of this increment — and its extension through at least 11 bits — has not been previously documented.

The mechanistic decomposition into 2.00 forced steps and 4.18 persistent downstream steps provides a concrete framework. The persistent component is the more surprising finding: it implies that carry propagation from forced (3n+1) operations biases the downstream parity sequence in ways that the independent-parity heuristic cannot capture. The discrepancy factor of approximately 1.80 between the observed downstream cost (4.18 steps) and the geometric-model prediction (1.41 steps) quantifies the degree to which the Collatz map's deterministic structure violates stochastic independence during the post-forced trajectory.

Deriving the exact value of the 6.15-step constant from first principles would require modeling how carry propagation from (3n+1) operations alters the distribution of subsequent bit patterns. The carries from each forced odd step extend into the higher bits of the iterate, creating a non-uniform bit distribution that persists for multiple subsequent steps. A complete theoretical derivation of this effect remains an open problem.

4.2 Trajectory Entropy and the 1/3 Law

The mean odd-step fraction of 0.322 is consistent with, but 3.4% below, the (1/3) value predicted by the stochastic model. The deficit reflects the negative autocorrelation in the parity sequence: each odd step deterministically produces an even successor, creating a systematic bias. The tight concentration of the distribution (standard deviation 0.029) confirms that trajectory-level behavior conforms closely to the ensemble prediction, supporting the quasi-ergodic assumptions underlying heuristic models (Terras, 1976; Tao, 2022).

4.3 Connection to Prior Work and Scale Considerations

The computational verification frontier far exceeds our (N = 10^7): Oliveira e Silva and Herzog (2012) verified convergence up to (5.764 \times 10^{18}), and Roosendaal (2020) has catalogued extensive extremal data. These efforts address a complementary question — whether all integers converge — while our analyses characterize the statistical structure of stopping times. For statistical questions, (10^7) provides ample data: even the smallest exact trailing-bit group ((k = 11), with 2,441 seeds) yields standard errors below 1.3 steps, sufficient to resolve the 6.15-step increment with high confidence.

The approximate log-normality of stopping times is consistent with the stochastic heuristic model in which orbit length is a sum of approximately independent multiplicative steps (Terras, 1976; Chamberland, 1996). The departures — negative skewness and platykurtosis — indicate that the independence assumption is approximate, as forced parity structure introduces systematic correlations. Tao's (2022) result that almost all orbits attain almost bounded values is consistent with our observation that (M(N)) grows polylogarithmically. Our empirical constant (C \approx 2.64) refines the heuristic to a specific quantitative value.

4.4 Limitations

This study is entirely empirical; no result constitutes a proof. The linearity of the trailing 1-bit effect could break down at higher bit counts or larger (N). At (k \geq 12), sample sizes fall below 1,500 and observed increments become more variable, consistent with increased sampling noise. Extending the linearity result to (k \geq 15) would require substantially larger (N).

5. Conclusion

We have presented an empirical study of Collatz stopping times for all integers up to (10^7), yielding the following quantitative findings:

1. Mean stopping time increases by (6.15 \pm 0.28) steps per trailing 1-bit, a linearity persisting through (k = 11) under exact grouping.

2. This increment decomposes into 2.00 forced steps and 4.18 persistent downstream steps, verified by direct trajectory measurement.

3. Trajectory entropy has mean 0.322 (std 0.029), near the theoretical (1/3), confirming quasi-ergodic orbit behavior.

4. The residue-class spread at modulus 64 is 74.66 steps, of which at most 12 are forced, demonstrating persistent influence of initial bit patterns.

5. Stopping times are approximately log-normal with skewness (-0.37) and excess kurtosis (-0.19) in log-space.

6. Maximum stopping time grows as (2.64 \cdot (\ln N)^2); lag-1 autocorrelation is 0.402, decaying slowly.

7. There are 53 record-breaking stopping times up to (10^7), with superlogarithmic spacing.

The persistence analysis demonstrates that the influence of initial bit patterns extends well beyond the deterministic prefix — a finding that constrains the applicability of independence assumptions in stochastic models of the Collatz map. The accompanying reproducible Python script enables independent verification of all reported results.

References

1. Applegate, D., and Lagarias, J. C. (1995). Density bounds for the 3x+1 problem. I. Tree-search method. Mathematics of Computation, 64(209), 411–426.

2. Chamberland, M. (1996). A continuous extension of the 3x+1 problem. Dynamics of Continuous, Discrete and Impulsive Systems, 2(4), 495–509.

3. Collatz, L. (1937). Unpublished problem, communicated orally.

4. Crandall, R. E. (1978). On the "3x+1" problem. Mathematics of Computation, 32(144), 1281–1292.

5. Crary, F. D. (2012). Record-setting delay values in the 3x+1 problem (unpublished).

6. Lagarias, J. C. (1985). The 3x+1 problem and its generalizations. The American Mathematical Monthly, 92(1), 3–23.

7. Oliveira e Silva, T., and Herzog, T. (2012). Verification of the Collatz conjecture up to 5.764 × 10^18 (unpublished; available at the 3x+1 conjecture verification project website).

8. Roosendaal, E. (2020). On the 3x+1 problem. Available at www.ericr.nl/wondrous.

9. Tao, T. (2022). Almost all orbits of the Collatz map attain almost bounded values. Forum of Mathematics, Pi, 10, e12.

10. Terras, R. (1976). A stopping time problem on the positive integers. Acta Arithmetica, 30(3), 241–252.