Оглавление: :: Library Catalog

Сохранить в:

Библиографические подробности
Главный автор:	McEvoy, Adam L
Формат:	Recurso digital
Язык:	английский
Опубликовано:	Zenodo 2026
Предметы:	Goldbach conjecture Hardy-Littlewood prime pairs number theory spectral analysis thedr
Online-ссылка:	https://doi.org/10.5281/zenodo.19155290
Метки:	Добавить метку Нет меток, Требуется 1-ая метка записи!

Оглавление:

# Fractal Geometric Analysis of the Goldbach Comet via the Fractal Correction Engine **Authors:** Adam L McEvoy **Date:** March 21, 2026 **Keywords:** Goldbach conjecture, fractal dimension, Hardy-Littlewood, prime pairs, number theory, spectral analysis, curvature, golden ratio, Hurst exponent --- ## Abstract I present the first fractal geometric analysis of the Goldbach comet --- the scatter plot of Goldbach pair counts $g(n)$ for even integers $n$ --- using the Fractal Correction Engine (FCE), a computational framework based on $\pi$-normalized curvature, multi-scale Gaussian decomposition, and Riemann zeta-inspired weighting. Applied to all even integers in $[4, 100{,}000]$, we characterize the comet's fractal structure (Higuchi $D_H = 2.0000$, box-counting $D_{BC} = 1.6614$), establish its near-random temporal dynamics (Hurst $H = 0.5236$), and identify statistically significant golden ratio dominance in prime pair spacings ($\varphi$-correlation $= 0.5396$, $p = 0.002$). We construct an enhanced ensemble predictor that achieves a mean absolute error (MAE) of 32.82, outperforming the classical Hardy-Littlewood asymptotic formula (MAE $= 158.84$) by a factor of $4.84\times$. The dominant prediction strategy exploits modular arithmetic structure ($n \bmod 6$), capturing multiplicative patterns invisible to the continuous Hardy-Littlewood approximation. Spectral analysis reveals a dominant period of 3 (reflecting the mod-6 structure) and near-zero correlation with Riemann zeta zeros at this scale. We do not claim to prove the Goldbach conjecture; rather, we demonstrate that the Goldbach comet possesses rich, predictable internal structure beyond what classical analytic number theory captures. --- ## 1. Introduction ### 1.1 The Goldbach Conjecture The Goldbach conjecture (1742) states that every even integer $n > 2$ can be expressed as the sum of two prime numbers. Despite nearly three centuries of effort, this remains one of the oldest unsolved problems in mathematics. The strongest partial result is Chen's theorem (1966), which proves every sufficiently large even integer is the sum of a prime and a semiprime. Computational verification has confirmed the conjecture up to $4 \times 10^{18}$ (Oliveira e Silva et al., 2014). For each even $n$, we define the Goldbach function: $$g(n) = \#\{(p, q) : p + q = n,\; p \leq q,\; p, q \text{ prime}\}$$ The plot of $(n, g(n))$ for even $n$ produces a distinctive scatter pattern known as the **Goldbach comet** --- a structured cloud whose density increases with $n$ in a manner governed by prime distribution. ### 1.2 The Hardy-Littlewood Prediction Hardy and Littlewood (1923) conjectured an asymptotic formula for $g(n)$ based on the circle method: $$g(n) \sim C_2 \cdot S(n) \cdot \frac{n}{\ln^2 n}$$ where $C_2 \approx 0.6602$ is the twin prime constant: $$C_2 = 2 \prod_{p > 2} \left(1 - \frac{1}{(p-1)^2}\right) \approx 0.6601618158$$ and $S(n)$ is the singular series: $$S(n) = \prod_{\substack{p \mid n \\ p > 2}} \frac{p - 1}{p - 2}$$ This formula provides a smooth approximation to the comet's overall trend but fails to capture its fine arithmetic structure, particularly the systematic residuals that depend on the modular properties of $n$. ### 1.3 Motivation While the Hardy-Littlewood formula captures the leading-order behavior of $g(n)$, the residuals $r(n) = g(n) - \text{HL}(n)$ display rich structure that has not been systematically characterized using modern tools from fractal geometry and signal processing. This work asks: 1. What is the fractal dimension of the Goldbach comet? 2. Does $g(n)$ exhibit long-range dependence or memory? 3. Can fractal-geometric methods predict $g(n)$ more accurately than Hardy-Littlewood? 4. What spectral structure underlies the comet? I address these questions using the Fractal Correction Engine (FCE), a computational framework originally developed for trajectory analysis in dynamical systems. This paper constitutes the third in a trilogy connecting Riemann zeta structure, computational complexity, and additive prime theory through the FCE framework. --- ## 2. The Fractal Correction Engine (FCE) ### 2.1 Overview The Fractal Correction Engine is a signal analysis and prediction framework that operates on discrete sequences through three core mechanisms: 1. **$\pi$-normalized curvature analysis** --- measuring local geometric complexity relative to a canonical semicircular arc 2. **Multi-scale Gaussian decomposition** --- separating structure at $\pi$-scaled resolution levels 3. **Zeta-weighted correction** --- combining detail levels with Riemann zeta-inspired weights The FCE treats any discrete sequence $f(x_1), f(x_2), \ldots, f(x_N)$ as a geometric trajectory and applies differential-geometric tools to characterize and predict its behavior. ### 2.2 Discrete Curvature For a discrete sequence $f$, the curvature at point $i$ is computed via finite differences: $$\kappa(i) = \frac{|f''(i)|}{(1 + (f'(i))^2)^{3/2}}$$ where $f'(i)$ and $f''(i)$ are the first and second discrete derivatives computed via `numpy.gradient` (central differences). The denominator normalizes for slope, ensuring curvature measures intrinsic bending rather than steepness. ### 2.3 $\pi$-Normalized Curvature The curvature is normalized by $\pi$ to produce a dimensionless measure: $$\kappa_{\text{norm}}(i) = \frac{\kappa(i)}{\pi}$$ This normalization has geometric meaning: - $\kappa_{\text{norm}} = 1$: the trajectory bends as tightly as a semicircular arc - $\kappa_{\text{norm}} \to 0$: the trajectory is locally straight - $\kappa_{\text{norm}} > 1$: the trajectory bends more sharply than a semicircle The radius of curvature is the reciprocal: $$r(i) = \frac{1}{\kappa(i) + \epsilon}$$ where $\epsilon = 10^{-30}$ prevents division by zero. To avoid numerical overflow when $\kappa \to 0$ (near-straight segments), the radius is clamped: $$r(i) = \min\left(\frac{1}{\kappa(i) + \epsilon},\; N\right)$$ where $N$ is the sequence length. ### 2.4 Multi-Scale Gaussian Decomposition The FCE decomposes the input sequence at $\pi$-scaled resolution levels using Gaussian convolution. For $n_s$ scales (default $n_s = 5$), the smoothing widths are: $$\sigma_k = \pi \cdot 2^k, \quad k = 0, 1, \ldots, n_s - 1$$ At each scale, the sequence is convolved with a normalized Gaussian kernel: $$G_{\sigma}(x) = \frac{1}{\sqrt{2\pi}\,\sigma} \exp\!\left(-\frac{x^2}{2\sigma^2}\right)$$ The smoothed versions are $S_k = G_{\sigma_k} * f$, and the detail levels capture the structure between consecutive scales: $$D_k = S_{k-1} - S_k, \quad k = 1, \ldots, n_s - 1$$ with $S_0 = f$ (the original sequence). Each $D_k$ isolates oscillatory structure at spatial frequencies corresponding to the scale $\sigma_k$. ### 2.5 The FCE Correction Formula The core FCE correction combines curvature, multi-scale details, and Riemann zeta-inspired weighting: $$f_{\text{corrected}}(i) = f(i) + \alpha \cdot \pi \cdot r(i) \cdot \sum_{k=1}^{n_s} \frac{D_k(i)}{k^{3/2}}$$ where: - $\alpha = 0.01$ is the correction strength - $\pi$ is the normalization constant - $r(i)$ is the clamped radius of curvature - $D_k(i)$ is the $k$-th detail level at point $i$ - $k^{-3/2}$ provides zeta-inspired weighting (cf. $\zeta(3/2) = \sum k^{-3/2}$), giving more weight to finer scales The correction is larger where the sequence is locally straight ($r$ large) and rich in multi-scale detail ($D_k$ large), and smaller where the sequence already curves sharply ($r$ small). ### 2.6 Fractal Path Extraction The FCE extracts a "fractal path" from an observed sequence through iterative refinement: $$\hat{f}^{(0)} = \text{FCE}(f)$$ $$\hat{f}^{(t+1)} = \hat{f}^{(t)} + \frac{1}{2}\,\text{FCE}\!\left(f - \hat{f}^{(t)}\right), \quad t = 0, 1, 2$$ After 3 iterations, the path match is evaluated via: - **Pearson correlation** $\rho(\hat{f}, f)$: measures shape similarity - **Hausdorff-like distance** $d_H = \max|f - \hat{f}| / \text{std}(f)$: measures maximum deviation ### 2.7 Configuration Parameters The FCE is governed by the following parameters: | Parameter | Symbol | Default | Description | |-----------|--------|---------|-------------| | Correction strength | $\alpha$ | 0.01 | Amplitude of correction term | | Decay timescale | $\tau$ | 50.0 | Exponential decay in extrapolation | | Oscillation frequency | $\omega$ | 0.1 | Frequency for oscillatory predictions | | Number of scales | $n_s$ | 5 | Multi-scale decomposition depth | | Prediction horizon | $H_p$ | 500 | Forward prediction range | | Higuchi max lag | $k_{\max}$ | 200 | Maximum lag for Higuchi FD | --- ## 3. Methods ### 3.1 Computational Pipeline The analysis follows a 7-stage pipeline executed with parallelism across $\lfloor C/2 \rfloor$ workers, where $C$ is the number of CPU cores: 1. **Prime Sieve** --- Sieve of Eratosthenes up to $n_{\max}$ 2. **Goldbach Comet Computation** --- parallel enumeration of $g(n)$ for all even $n \in [4, n_{\max}]$ 3. **FCE Fractal Analysis** --- curvature, fractal dimensions, predictions, waveform decomposition 4. **Spectral Analysis** --- FFT, Lomb-Scargle, power-law fitting, zeta-zero correlation 5. **Statistical Validation** --- bootstrap CIs, permutation tests, OEIS verification 6. **Visualization** --- 16 diagnostic plots 7. **Report Generation** --- JSON and text summaries System health is monitored continuously via CPU temperature and memory pressure sensors, with automatic worker throttling above 85$^\circ$C and dynamic reduction of bootstrap/permutation counts above 85% memory usage. ### 3.2 Goldbach Pair Counting For each even $n$, pairs are counted by testing all primes $p \leq n/2$: $$g(n) = \sum_{\substack{p \leq n/2 \\ p \text{ prime}}} \mathbb{1}[n - p \text{ is prime}]$$ where $\mathbb{1}[\cdot]$ is the indicator function. Prime testing uses $O(1)$ lookup via a boolean sieve array. The computation is parallelized by distributing chunks of even numbers across workers. The Hardy-Littlewood prediction for each $n$ is computed as: $$\text{HL}(n) = C_2 \cdot S(n) \cdot \frac{n}{\ln^2 n}$$ with $S(n)$ computed by trial division of $n$ by odd primes. ### 3.3 Fractal Dimension Estimation #### 3.3.1 Higuchi Fractal Dimension The Higuchi method (Higuchi, 1988) estimates fractal dimension from a time series without embedding in phase space. For a sequence $x_1, x_2, \ldots, x_N$ and lag $k$: 1. Construct subsequences for each offset $m = 1, \ldots, k$: $$x_m^{(k)} = x_m, x_{m+k}, x_{m+2k}, \ldots, x_{m + \lfloor(N-m)/k\rfloor \cdot k}$$ 2. Compute the normalized path length: $$L_m(k) = \frac{1}{k^2} \cdot \frac{N - 1}{\lfloor(N-m)/k\rfloor} \sum_{i=1}^{\lfloor(N-m)/k\rfloor} |x_{m+ik} - x_{m+(i-1)k}|$$ 3. Average over offsets: $$L(k) = \frac{1}{k} \sum_{m=1}^{k} L_m(k)$$ 4. The fractal dimension $D_H$ is the negative slope of the log-log regression: $$\log L(k) = -D_H \log k + c$$ We use $k = 1, 2, \ldots, k_{\max}$ with $k_{\max} = \min(200, N/4)$, and clamp $D_H \in [1.0, 2.0]$. #### 3.3.2 Box-Counting Fractal Dimension The box-counting method estimates the dimension of a point cloud $(n_i, g(n_i))$: 1. Normalize coordinates to the unit square $[0,1]^2$ 2. For each box size $\varepsilon$ from $10^{-0.3}$ to $10^{-2.5}$ (20 logarithmically spaced values):    - Overlay a grid with cells of side $\varepsilon$    - Count the number of non-empty cells $N(\varepsilon)$ 3. Fit the log-log regression: $$\log N(\varepsilon) = -D_{BC} \log \varepsilon + c$$ The slope $-D_{BC}$ gives the box-counting dimension, clamped to $[1.0, 2.0]$. ### 3.4 Hurst Exponent (R/S Analysis) The Hurst exponent $H$ measures long-range dependence via rescaled range analysis (Hurst, 1951): 1. Detrend the series by subtracting a fitted cubic polynomial 2. For window sizes $W$ from 10 to $N/2$ (geometric progression with factor 1.5):    - Partition the detrended series into non-overlapping windows of size $W$    - For each window, compute:      - Mean-adjusted cumulative deviation: $Y_t = \sum_{i=1}^{t}(x_i - \bar{x})$      - Range: $R = \max(Y) - \min(Y)$      - Standard deviation: $S = \text{std}(x)$      - Rescaled range: $R/S$ (if $S > 10^{-10}$)    - Average $R/S$ across all windows at this scale 3. Fit the log-log regression: $$\log(R/S) = H \log W + c$$ The exponent is clamped to $[0, 1]$ with interpretation: - $H > 0.5$: persistent (trending) behavior - $H = 0.5$: random walk - $H < 0.5$: anti-persistent (mean-reverting) behavior ### 3.5 Golden Ratio Analysis The FCE tests whether autocorrelation peak spacings in $g(n)$ preferentially align with the golden ratio $\varphi = (1 + \sqrt{5})/2 \approx 1.618$: 1. Compute the autocorrelation via FFT (O($N \log N$)): $$R(\tau) = \mathcal{F}^{-1}\!\left[|\mathcal{F}[x - \bar{x}]|^2\right]$$ normalized so $R(0) = 1$. 2. Identify peaks: local maxima of $R(\tau)$ with $R > 0.05$ 3. Compute successive peak-spacing ratios: $$\rho_i = \frac{\tau_{i+1}}{\tau_i}$$ 4. Measure proximity to $\varphi$ and $\pi$: $$\text{corr}_\varphi = \exp\!\left(-\frac{1}{M}\sum_{i=1}^{M}|\rho_i - \varphi|\right)$$ $$\text{corr}_\pi = \exp\!\left(-\frac{1}{M}\sum_{i=1}^{M}|\rho_i - \pi|\right)$$ 5. $\varphi$-dominance holds if $\text{corr}_\varphi > \text{corr}_\pi$ ### 3.6 Wave Interference Patterns For each even $n$, we define a binary interference pattern over candidate primes: $$I_n(p) = \begin{cases} 1 & \text{if both } p \text{ and } n - p \text{ are prime} \\ 0 & \text{otherwise} \end{cases}$$ for $p \in [2, n/2]$. These patterns are superposed across multiple values of $n$ by normalizing to a common phase space $\theta \in [0, 2\pi]$: $$\theta(p) = \frac{p - 2}{n/2 - 2} \cdot 2\pi$$ The superposition is the average over sampled $n$ values: $$\Psi(\theta) = \frac{1}{|\mathcal{N}|} \sum_{n \in \mathcal{N}} I_n(\theta)$$ **Constructive zones** are regions where $\Psi(\theta) > \bar{\Psi} + 2\sigma_\Psi$, indicating phase-space locations where prime pairing is enhanced. **Destructive zones** satisfy $\Psi(\theta) < \bar{\Psi} - \sigma_\Psi$. ### 3.7 Enhanced Prediction Strategies The FCE employs five prediction strategies, each predicting $g(n)$ for future even numbers beyond the training range. All strategies use the Hardy-Littlewood formula as a baseline and model the residual $r(n) = g(n) - \text{HL}(n)$. #### 3.7.1 Bias Correction Smooths the H-L residual with a Gaussian kernel of width $\sigma = 10\pi$ and extrapolates via curvature-based trajectory prediction: $$\hat{g}(n) = \text{HL}(n) + \text{extrapolate}(\text{smooth}(r_{\text{train}}))$$ #### 3.7.2 Modular Decomposition ($n \bmod 6$) Exploits the fact that all even $n > 2$ fall into residue classes $\{0, 2, 4\}$ modulo 6, each with distinct $g(n)$ statistics: $$\hat{g}(n) = \text{HL}(n) + \hat{r}_c(n), \quad c = n \bmod 6$$ where $\hat{r}_c$ is the curvature-extrapolated residual trained on the subsequence $\{g(m) : m \equiv c \pmod{6}\}$. #### 3.7.3 Multi-Scale $\pi$ Prediction Decomposes the residual at $\pi$-based scales and extrapolates each level independently with inverse-scale weighting: $$\hat{r}(n) = \frac{\sum_{k} w_k \cdot \hat{D}_k(n)}{\sum_{k} w_k}, \quad w_k = \frac{1}{k+1}$$ Coarser scales (larger $k$) receive less weight but provide smoother, more reliable extrapolations. #### 3.7.4 Curvature Template Matching Searches for historical segments whose $\pi$-normalized curvature profile matches the recent curvature pattern, then uses their continuations as predictions: 1. Query template: $\kappa_{\text{norm}}$ over the last 50 training points 2. Slide over all historical positions, computing correlation with the query 3. Select the top 20 matches with correlation $> 0.3$ 4. Weight by squared correlation: $$\hat{r}(n) = \frac{\sum_{j} \rho_j^2 \cdot r_j^{\text{cont}}(n)}{\sum_{j} \rho_j^2}$$ where $r_j^{\text{cont}}$ is the continuation following the $j$-th matched segment, amplitude-scaled to match recent levels. #### 3.7.5 Fractal Path Extrapolation Applies the FCE correction directly to the residual and extrapolates the resulting path: $$\hat{f} = \text{FCE}^{(3)}(r_{\text{train}})$$ $$\hat{r}(n) = \text{extrapolate}(\hat{f})$$ This is the "purest" FCE strategy but is susceptible to numerical instability when curvature approaches zero. #### 3.7.6 Validation-Weighted Ensemble Strategies are combined via validation-weighted averaging: 1. Hold out the last 20% of training data for validation 2. For each strategy $s$, compute $\text{MAE}_s$ on the validation set 3. Assign weights: $$w_s = \begin{cases} (\text{MAE}_s^2 + \epsilon)^{-1} & \text{if } \text{MAE}_s < \text{MAE}_{\text{HL}} \\ 0 & \text{otherwise} \end{cases}$$ 4. Normalize: $\hat{w}_s = w_s / \sum_s w_s$ 5. Final prediction: $$\hat{g}(n) = \sum_s \hat{w}_s \cdot \hat{g}_s(n)$$ Only strategies that beat Hardy-Littlewood on validation data receive nonzero weight. ### 3.8 Curvature-Based Trajectory Extrapolation The supporting extrapolation mechanism branches on curvature intensity at the sequence tail: **High curvature** ($\kappa_{\text{norm}} > 0.1$): damped oscillatory continuation: $$\hat{f}(t) = f_{\text{last}} + s \cdot t \cdot e^{-t/(3L)} + A \sin(\omega t + \phi) \cdot e^{-t/(2L)}$$ where $s$ is the trailing slope, $L$ is the tail length, $A$ is the recent oscillation amplitude, and $\phi$ is the terminal phase. **Low curvature** ($\kappa_{\text{norm}} \leq 0.1$): exponentially decaying linear trend: $$\hat{f}(t) = f_{\text{last}} + s \cdot t \cdot e^{-t/(5L)}$$ ### 3.9 Spectral Analysis #### 3.9.1 FFT Power Spectrum The centered, Hann-windowed sequence is transformed: $$X_k = \sum_{i=0}^{N-1} w_i \cdot (g_i - \bar{g}) \cdot e^{-2\pi i \cdot k \cdot i / N}$$ $$P(f_k) = |X_k|^2$$ where $w_i$ is the Hann window function and $f_k = k/N$. #### 3.9.2 Power-Law Fitting A power-law relationship $P(f) \propto f^{-\beta}$ is fit via log-log linear regression: $$\log P = -\beta \log f + c$$ The exponent $\beta$ characterizes the spectral color: $\beta \approx 0$ (white noise), $\beta \approx 1$ (pink/flicker noise), $\beta \approx 2$ (Brownian noise). #### 3.9.3 Lomb-Scargle Periodogram For potentially irregular sampling, the Lomb-Scargle periodogram (Lomb, 1976; Scargle, 1982) provides normalized spectral power at angular frequencies $\omega_j = 2\pi f_j$ across 2000 linearly spaced frequencies in $[0.001, 0.5]$. #### 3.9.4 Riemann Zeta-Zero Correlation The first 30 non-trivial zeros of the Riemann zeta function $\gamma_k$ (starting with $\gamma_1 \approx 14.135$) are mapped to normalized frequencies: $$f_k^{\zeta} = \frac{\gamma_k}{2\pi \cdot \gamma_{\max}}$$ The spectral power at these frequencies is compared to the overall mean: $$\text{corr}_\zeta = \frac{\overline{P(f^{\zeta})}}{\overline{P}}$$ A ratio significantly above 1 would indicate that zeta zeros are encoded in the comet's spectral structure. ### 3.10 Statistical Validation #### 3.10.1 Bootstrap Confidence Intervals For each statistic $\theta$ (Higuchi $D_H$, Hurst $H$): 1. Draw $B = 1000$ bootstrap samples with replacement 2. Compute $\hat{\theta}_b$ for each sample 3. Report the 95% percentile interval: $[\hat{\theta}_{(\alpha/2)}, \hat{\theta}_{(1-\alpha/2)}]$ To manage memory, series longer than 5000 points are subsampled before bootstrapping. #### 3.10.2 Permutation Tests **$\varphi$-dominance test** ($n_{\text{perm}} = 500$): Under $H_0$, the observed $\varphi$-correlation arises by chance. For each permutation, $g(n)$ is randomly shuffled and $\varphi$-correlation recomputed. The $p$-value is: $$p = \frac{\#\{\text{corr}_\varphi^{\text{perm}} \geq \text{corr}_\varphi^{\text{obs}}\} + 1}{n_{\text{perm}} + 1}$$ **FCE prediction test** ($n_{\text{perm}} = 500$): Under $H_0$, the FCE prediction MAE is no better than prediction from shuffled training data. #### 3.10.3 OEIS Verification The first 50 computed values of $g(n)$ are verified against the OEIS sequence A002375 (number of ways of writing $2n$ as a sum of two primes). #### 3.10.4 Information Horizon Test MAE is evaluated at increasing prediction horizons $h \in \{50, 100, 200, 500\}$. A sharp discontinuity in MAE growth would indicate a fundamental predictability limit (information horizon). This connects to the broader question of computational irreducibility in number-theoretic sequences. ### 3.11 Sub-Waveform Analysis The FCE analysis is applied to 23 derived waveforms, each isolating a different aspect of Goldbach structure: - **H-L Residual**: $r(n) = g(n) - \text{HL}(n)$ - **Prime Gaps**: $p_{k+1} - p_k$ for primes up to $n_{\max}$ - **Singular Series**: $S(n)$ values - **Pair Spacings**: differences between consecutive $g(n)$ values - **Cumulative Sum**: $\sum_{k=1}^{n} g(k)$ - **$g(n)/\text{HL}(n)$ Ratio**: deviation from unity - **Phase-Space Orbit**: curvature of the $(g(n), g(n+2))$ trajectory - **Smoothed Bias**: Gaussian-smoothed H-L residual - **Acceleration**: $\Delta^2 g(n)$ (second difference) - **Twin-Prime Gaps**: gaps between consecutive twin primes - **Log-Derivative**: $\frac{d}{dn}\ln g(n)$ - **Phase Orbit**: $(g(n), g(n+2))$ distance sequence - **Spectral Envelope**: $|\text{FFT}(g)|$ - **Modular classes**: $g(n)$ restricted to $n \equiv r \pmod{6}$ and $n \equiv r \pmod{30}$ Each waveform receives the full FCE treatment: Higuchi $D_H$, Hurst $H$, $\varphi$-correlation, and forward/backward prediction MAE. --- ## 4. Results All results are from a single run with $n_{\max} = 100{,}000$, testing 49,999 even integers. Computation time: 1,023.7 seconds on a 24-core system using 12 parallel workers. The Goldbach conjecture holds for all tested values ($g(n) \geq 1$ for all even $n \in [4, 100{,}000]$). ### 4.1 Goldbach Comet Statistics | Metric | Value | |--------|-------| | Even numbers tested | 49,999 | | $g(n)$ range | $[1, 2168]$ | | Mean $g(n)$ | 507.35 | | Conjecture holds | Yes (all $n$) | | OEIS A002375 match | Pass | ### 4.2 Fractal Characterization | Metric | Value | 95% CI | |--------|-------|--------| | Higuchi dimension $D_H$ | 2.0000 | [1.9996, 2.0000] | | Box-counting dimension $D_{BC}$ | 1.6614 | --- | | Hurst exponent $H$ | 0.5236 | [0.5076, 0.5578] | The Higuchi dimension $D_H = 2.0$ indicates maximal local roughness --- at fine scales, the Goldbach comet fills the plane as densely as a 2D surface. This is consistent with the pseudo-random nature of prime pair counts at small scales. The box-counting dimension $D_{BC} = 1.66$ captures the coarser self-similar structure of the comet's envelope. The gap $D_H - D_{BC} \approx 0.34$ reflects the difference between local complexity (noise-like) and global geometry (structured cloud). The Hurst exponent $H = 0.52 \approx 0.5$ confirms that $g(n)$ behaves as a near-random walk, with no significant long-range memory. Knowledge of $g(n)$ at small $n$ provides no predictive advantage for large $n$ beyond the Hardy-Littlewood trend. ### 4.3 Curvature Statistics | Metric | Value | |--------|-------| | Mean $\kappa / \pi$ | 0.6023 | | Std $\kappa / \pi$ | 5.5387 | The mean $\pi$-normalized curvature of 0.60 indicates that the Goldbach comet's trajectory bends, on average, at about 60% of the intensity of a semicircular arc. The high standard deviation (5.54) reflects extreme local variability, with some points exhibiting near-straight segments and others showing very sharp turns. ### 4.4 Golden Ratio Dominance | Metric | Value | |--------|-------| | $\varphi$-correlation | 0.5396 | | $\pi$-correlation | 0.1176 | | $\varphi$-dominant | Yes | | $p$-value (permutation) | 0.0020 | The autocorrelation peak spacings of $g(n)$ align significantly more closely with the golden ratio ($\varphi \approx 1.618$) than with $\pi$ ($\approx 3.14$). The permutation test confirms this is statistically significant at $p < 0.01$, with only 1 out of 500 shuffled sequences producing an equal or greater $\varphi$-correlation. ### 4.5 Hardy-Littlewood Comparison | Metric | Value | |--------|-------| | H-L $R^2$ | 0.8924 | | H-L RMSE | 109.09 | | H-L mean residual | 92.12 | The Hardy-Littlewood formula explains 89.2% of the variance in $g(n)$, with an RMSE of 109 pair counts. The positive mean residual indicates systematic underestimation --- the formula predicts fewer pairs than observed, on average. ### 4.6 Enhanced FCE Prediction Performance #### 4.6.1 Individual Strategy Results | Strategy | Forward MAE | Beats H-L? | Ensemble Weight | |----------|-------------|------------|-----------------| | Modular ($n \bmod 6$) | 29.06 | Yes | 0.787 | | Bias correction | 83.93 | Yes | 0.094 | | Template matching | 88.97 | Yes | 0.084 | | Multi-scale $\pi$ | 138.36 | Yes | 0.035 | | Fractal path | $3.14 \times 10^{10}$ | No | 0.000 | | **H-L baseline** | **158.84** | --- | --- | | **Ensemble** | **32.82** | **Yes** | --- | #### 4.6.2 Forward vs. Backward Prediction | Predictor | Forward MAE | Backward MAE | |-----------|-------------|--------------| | Original FCE | 383.17 | 21.51 | | Hardy-Littlewood | 158.84 | 4.02 | | Enhanced ensemble | 32.82 | 1.72 | The forward/backward asymmetry (31.10) reflects the fundamental difficulty of extrapolation versus interpolation. The backward MAE of 1.72 demonstrates that the ensemble accurately captures the sequence's internal structure when not extrapolating. #### 4.6.3 Summary $$\text{Improvement factor} = \frac{\text{MAE}_{\text{HL}}}{\text{MAE}_{\text{FCE}}} = \frac{158.84}{32.82} = 4.84\times$$ The enhanced FCE ensemble achieves a $4.84\times$ improvement over Hardy-Littlewood, with the modular $n \bmod 6$ strategy contributing 78.7% of the ensemble weight. This strategy captures the multiplicative structure of even numbers: integers divisible by 6 have systematically different pair count distributions than those $\equiv 2$ or $4 \pmod{6}$, because divisibility by both 2 and 3 opens additional combinatorial pathways for prime decomposition. The FCE prediction improvement is statistically significant ($p = 0.002$, permutation test with 500 shuffles). ### 4.7 Spectral Analysis | Metric | Value | |--------|-------| | Dominant frequency | 0.3333 | | Dominant period | 3.0 | | Power-law exponent $\beta$ | 0.031 | | Spectral entropy | 1.973 | | Zeta-zero correlation | 0.004 | The dominant spectral period of 3 corresponds directly to the modular structure of even numbers: every third even number is divisible by 6, creating a period-3 oscillation in $g(n)$. The near-zero power-law exponent ($\beta = 0.03$) indicates an approximately flat (white noise) spectrum, consistent with the Hurst exponent $H \approx 0.5$. The negligible zeta-zero correlation (0.004) indicates no detectable direct spectral link between the Goldbach comet and the Riemann zeta zeros at this scale ($n \leq 100{,}000$). The connection between primes and zeta zeros operates through the prime counting function $\pi(x)$, not directly through Goldbach pair counts. ### 4.8 Constructive Interference | Metric | Value | |--------|-------| | Constructive zones | 5,928 | | Destructive zones | 18 | | Twin-prime correlation | 0.1356 | | Superposition peaks | 10 | | Dominant superposition period | 500.0 | The overwhelming dominance of constructive zones (5,928 vs. 18 destructive) reflects the strong bias toward prime pairing across the phase space --- most regions of the interference pattern show enhanced prime pair density, consistent with $g(n)$ growing with $n$. ### 4.9 Sub-Waveform Analysis The 23 sub-waveforms reveal differentiated fractal and predictive characteristics: | Waveform | $D_H$ | $H$ | $\varphi$-dom | Fwd MAE | Bwd MAE | |----------|--------|------|---------------|---------|---------| | H-L Residual | 2.000 | 0.631 | Yes | 81.4 | 3.3 | | Prime Gaps | 2.000 | 0.479 | No | 10.1 | 4.9 | | Singular Series | 2.000 | 0.110 | Yes | 0.8 | 0.7 | | Pair Spacings | 2.000 | 0.566 | No | 3.7 | 2.1 | | Cumulative Sum | 1.508 | 1.000 | Yes | $1.5 \times 10^6$ | 2,684 | | $g/\text{HL}$ Ratio | 2.000 | 0.736 | Yes | 0.0 | 0.3 | | Phase-Space Orbit | 2.000 | 0.861 | Yes | 0.0 | 0.6 | | Smoothed Bias | 1.899 | 0.921 | Yes | 10.1 | 3.0 | | Acceleration $\Delta^2 g$ | 2.000 | 0.081 | Yes | 1,374.9 | 17.9 | | Twin-Prime Gaps | 1.994 | 0.534 | Yes | 85.3 | 38.9 | | Log-Derivative | 2.000 | 0.107 | Yes | 0.4 | 0.3 | | Phase Orbit | 2.000 | 0.410 | Yes | 368.8 | 10.9 | | Spectral Envelope | 1.980 | 0.546 | Yes | 1,012.5 | 37,169 | | $n \bmod 6 = 0$ | 2.000 | 0.552 | Yes | 138.6 | 33.8 | | $n \bmod 6 = 2$ | 2.000 | 0.561 | Yes | 73.4 | 16.0 | | $n \bmod 6 = 4$ | 2.000 | 0.575 | Yes | 72.0 | 17.4 | Notable findings: - **Cumulative Sum** ($D_H = 1.508$, $H = 1.0$): The integrated Goldbach function has the lowest fractal dimension and maximal persistence, reflecting its smooth, monotonically increasing nature. - **Smoothed Bias** ($D_H = 1.899$, $H = 0.921$): The Gaussian-smoothed residual is strongly persistent, confirming the H-L bias has long-range structure. - **Singular Series** ($H = 0.110$): Strongly anti-persistent, reflecting the rapidly oscillating multiplicative function $S(n)$. - **Prime Gaps** and **Pair Spacings**: Show no golden ratio dominance, consistent with these being difference sequences rather than structured waveforms. - **$\varphi$-dominance**: 21 of 23 waveforms show golden ratio dominance, with the two exceptions being Prime Gaps and Pair Spacings. ### 4.10 Fractal Path Validation | Metric | Value | |--------|-------| | Path match correlation | $-0.0434$ | | Path Hausdorff distance | $1.57 \times 10^{10}$ | The fractal path strategy produces poor results, with negative correlation and large Hausdorff distance. This reflects the fundamental challenge of iteratively applying curvature corrections to a noisy signal: each refinement step can amplify errors when the radius of curvature is large (near-straight segments). The strategy is correctly excluded from the ensemble (weight = 0). ### 4.11 Information Horizon No information horizon was detected within the tested range. Prediction accuracy degrades smoothly with increasing horizon rather than exhibiting a sharp cutoff, suggesting that the predictable structure of $g(n)$ does not terminate abruptly at any particular scale below $n = 100{,}000$. --- ## 5. Discussion ### 5.1 The Modular Structure of the Goldbach Comet The most striking result is the dominance of the modular $n \bmod 6$ strategy, which alone achieves MAE = 29.06 --- better than the full ensemble of other strategies. This reveals that the "noise" in the Hardy-Littlewood residual is largely **arithmetic**, not random. The three residue classes $\{0, 2, 4\} \pmod{6}$ have systematically different $g(n)$ distributions because: - Numbers $n \equiv 0 \pmod{6}$ are divisible by both 2 and 3, admitting representations like $n = 3 + (n-3)$ where $n-3$ is always odd and divisible by 3 only if $n \equiv 0$ - The singular series $S(n)$ already accounts for some of this structure, but the H-L formula's continuous approximation cannot capture the discrete jumps between residue classes The spectral analysis confirms this interpretation: the dominant period of 3 in the FFT corresponds to the period-6 oscillation of $g(n)$ across even numbers (since we index only even $n$, period 6 maps to period 3 in the index). ### 5.2 Fractal Geometry of the Comet The comet exhibits a dual fractal character: - **Locally** ($D_H = 2.0$): maximal roughness, space-filling at fine scales - **Globally** ($D_{BC} = 1.66$): structured self-similar envelope This combination is characteristic of point processes with smooth envelope functions contaminated by fine-scale noise --- exactly what one expects for prime pair counts, where the smooth trend is given by Hardy-Littlewood and the fine structure is determined by the specific distribution of primes near $n$. ### 5.3 The Role of the Golden Ratio The statistically significant $\varphi$-dominance ($p = 0.002$) in autocorrelation peak spacings is intriguing but should be interpreted cautiously. The golden ratio appears in many natural processes due to its connection with optimal packing and quasi-periodic dynamics. In the context of prime pair spacings, this may reflect the quasi-crystalline distribution of primes at certain scales, though further investigation at larger $n$ ranges would be needed to confirm whether this is a fundamental feature or a finite-size effect. ### 5.4 Limitations 1. **Finite range**: The analysis covers $n \leq 100{,}000$, far below the range where asymptotic predictions become reliable. Results may not extrapolate to arbitrarily large $n$. 2. **No proof implications**: Predicting $g(n)$ more accurately does not address whether $g(n) \geq 1$ for all even $n > 2$. Even a perfect predictor cannot constitute a proof because no finite computation covers infinitely many cases. 3. **Fractal path instability**: The FCE fractal path extraction (Strategy 5) suffers from numerical amplification in iterative refinement, producing MAE $\sim 10^{10}$. This strategy requires fundamental algorithmic revision (e.g., convergent fixed-point iteration) before it can contribute to the ensemble. 4. **Higuchi saturation**: Most waveforms show $D_H = 2.0$, which may reflect either genuine maximal roughness or a limitation of the Higuchi method for short, noisy sequences. Alternative methods (e.g., detrended fluctuation analysis) could provide complementary estimates. 5. **Ensemble extrapolation**: The $4.84\times$ improvement is measured on a 500-point prediction horizon within the training range. Performance at much larger prediction horizons or for $n \gg 100{,}000$ is unknown. ### 5.5 Connection to Related Work This analysis complements the Hardy-Littlewood circle method approach by providing a geometric characterization of the residual structure that the analytic formula misses. The modular decomposition finding parallels work on the distribution of primes in arithmetic progressions (Dirichlet, 1837; Green-Tao, 2008), while the spectral analysis connects to the broader program of understanding prime distribution through Fourier analysis (Montgomery, 1973). --- ## 6. Conclusion I have presented the first comprehensive fractal geometric analysis of the Goldbach comet using the Fractal Correction Engine. Our key findings are: 1. **Fractal characterization**: The Goldbach comet has Higuchi dimension $D_H = 2.000$ (local roughness) and box-counting dimension $D_{BC} = 1.661$ (global structure), with Hurst exponent $H = 0.524$ indicating near-random dynamics. 2. **Prediction improvement**: An ensemble of FCE-based strategies achieves MAE = 32.82, outperforming the Hardy-Littlewood formula (MAE = 158.84) by a factor of $4.84\times$ ($p = 0.002$). 3. **Modular structure**: The dominant prediction signal comes from the $n \bmod 6$ decomposition (78.7% ensemble weight), revealing that the H-L residual is primarily arithmetic rather than random. 4. **Golden ratio signature**: Autocorrelation peak spacings show statistically significant alignment with $\varphi = (1+\sqrt{5})/2$ ($p = 0.002$), present in 21 of 23 analyzed waveforms. 5. **Spectral structure**: The dominant period of 3 (reflecting mod-6 structure) and flat power spectrum ($\beta = 0.031$) characterize the comet as arithmetic noise on a smooth trend. I emphasize that these results constitute a computational number theory contribution --- characterizing and predicting the internal structure of the Goldbach comet --- and do not constitute or approach a proof of the Goldbach conjecture. --- ## 7. Data Availability All results, including the complete analysis pipeline (`goldbach_fce.py`), metrics (`metrics.json`), computed comet data (`goldbach_comet.csv`), and 16 diagnostic visualizations, are available in the accompanying Zenodo deposit. --- ## 8. References 1. Chen, J. R. (1966). On the representation of a large even integer as the sum of a prime and the product of at most two primes. *Scientia Sinica*, 16, 157--176. 2. Dirichlet, P. G. L. (1837). Beweis des Satzes, dass jede unbegrenzte arithmetische Progression, deren erstes Glied und Differenz ganze Zahlen ohne gemeinschaftlichen Factor sind, unendlich viele Primzahlen enthalt. *Abhandlungen der Koniglichen Preussischen Akademie der Wissenschaften*, 45--81. 3. Goldbach, C. (1742). Letter to Leonhard Euler, June 7, 1742. 4. Green, B., & Tao, T. (2008). The primes contain arbitrarily long arithmetic progressions. *Annals of Mathematics*, 167(2), 481--547. 5. Hardy, G. H., & Littlewood, J. E. (1923). Some problems of 'Partitio Numerorum'; III: On the expression of a number as a sum of primes. *Acta Mathematica*, 44, 1--70. 6. Higuchi, T. (1988). Approach to an irregular time series on the basis of the fractal theory. *Physica D*, 31(2), 277--283. 7. Hurst, H. E. (1951). Long-term storage capacity of reservoirs. *Transactions of the American Society of Civil Engineers*, 116, 770--799. 8. Lomb, N. R. (1976). Least-squares frequency analysis of unequally spaced data. *Astrophysics and Space Science*, 39, 447--462. 9. Montgomery, H. L. (1973). The pair correlation of zeros of the zeta function. *Proceedings of Symposia in Pure Mathematics*, 24, 181--193. 10. Oliveira e Silva, T., Herzog, S., & Pardi, S. (2014). Empirical verification of the even Goldbach conjecture and computation of prime gaps up to $4 \times 10^{18}$. *Mathematics of Computation*, 83(288), 2033--2060. 11. Scargle, J. D. (1982). Studies in astronomical time series analysis. II. Statistical aspects of spectral analysis of unevenly spaced data. *The Astrophysical Journal*, 263, 835--853.

Схожие документы