Analytical Treatise: Cauchy–Lorentz Pathologies & Robust Inference

First-Principles Derivations Accompanying the Cauchy–Lorentz vs. Gaussian Drylab

1. Physical Lighthouse Geometry & Normal-Ratio Duality

MATH-PROB 99.8 · HTML5-ARCH 99.6
THEOREM 1.1 · GULL–CAUCHY LIGHTHOUSE (1988)

Uniform Azimuthal Emission Produces 1/x² Wings

Place an isotropic optical beacon at coordinates (x₀, γ) with altitude γ > 0 above a linear detector along the x-axis. A ray emitted at uniform angle Θ ~ Uniform(−π/2, π/2) has density f_Θ(θ) = 1/π and strikes the detector at X = x₀ + γ tan Θ. Inverting via θ(x) = arctan((x − x₀)/γ) and applying the change-of-variables Jacobian yields:

fX(x) = fΘ(θ(x)) |dθdx| = 1π ⋅ γγ2+(x−x0)2 = 1πγ[1+(x−x0γ)2]

In RF resonators and atomic transitions, the same Lorentzian line shape arises as the power spectrum |H(ω)|² ∝ [(ω − ω₀)² + γ²]⁻¹ of an exponentially damped oscillation e^(−γt + iω₀t) u(t), where γ is the half-width at half-maximum (HWHM).

THEOREM 1.2 · RATIO OF INDEPENDENT GAUSSIANS

Why Dividing Two Bell Curves Breaks Classical Statistics

Let Z₁, Z₂ ~ N(0, 1) be independent standard Gaussian variables. Their joint density f(z₁, z₂) = (2π)⁻¹ exp(−½(z₁² + z₂²)) is rotationally invariant. Passing to polar coordinates (Z₁, Z₂) = (R sin Θ, R cos Θ) makes the polar angle uniform on (−π, π], so the quotient cancels radius R identically:

Q=Z1Z2 = tanΘ ∼ Cauchy(0,1) ≡ tν=1

Near-zero realizations of denominator Z₂ have non-vanishing density f_{Z₂}(0) = 1/√(2π), propelling |Q| into extreme algebraic excursions whenever noisy ratios or transfer-function divisions are formed experimentally.

2. Algebraic Tail Survival vs. Lebesgue Non-Integrability

MATH-PROB 99.8 · STAT-NUM 99.8
PROPOSITION 2.1 · MILL'S RATIO VS. ARCTANGENT DEFECT

Exponential Core Compression vs. Polynomial Wings

Integrating f_X(x) gives exact cumulative distribution F_X(x) = ½ + π⁻¹ arctan((x − x₀)/γ). Expanding arctan(k) = π/2 − k⁻¹ + O(k⁻³) for k ≫ 1 contrasts two-sided tail probabilities against Gaussian Mill's ratio:

ℙ(|X−x0|>kγ) ∼ 2πk , ℙ(|G−μ|>kσ) ∼ 2π e−k2/2k

At k = 10 scale units, a Gaussian exceeds 10σ with probability 1.52 × 10⁻²³, whereas a Cauchy process exceeds 10γ on 6.35% of all observations—roughly once every sixteen draws.

THEOREM 2.2 · LEBESGUE EXPECTATION VS. PRINCIPAL VALUE

Why Symmetry Cannot Rescue an Undefined Mean

For 𝔼[X] to exist in the Lebesgue sense, 𝔼[|X|] must be finite. Setting x₀ = 0 and integrating over independent lower and upper cutoffs R₁, R₂ > 0 reveals logarithmic divergence:

∫−R1R2 xfX(x)dx = γ2π ln (γ2+R22γ2+R12) ⟶R2=cR1→∞ γπlnc

Choosing ratio c > 0 steers the improper integral to any real number; only the symmetric Cauchy Principal Value (c = 1) gives zero. Meanwhile 𝔼[|X|] = lim_{R→∞} (γ/π) ln(1 + R²/γ²) = +∞, and every higher moment 𝔼[|X|^r] = +∞ for all r ≥ 1.

3. Complex Residue Derivation of φ_X(t) & Sample-Mean Invariance

MATH-PROB 99.8 · DRYLAB-QA 99.8
THEOREM 3.1 · UPPER-HALF-PLANE CONTOUR INTEGRATION

Simple Poles at z_± = x₀ ± iγ Generate the |t| Cusp

Extend the integrand of φ_X(t) = ∫ e^(itx) f_X(x) dx to ℂ: its analytic continuation has simple poles at z_± = x₀ ± iγ. For t > 0, close a semicircle C_R⁺ in Im(z) ≥ 0; Jordan's Lemma annihilates the arc as R → ∞, leaving the residue at z_+:

φX(t) = 2πi Resz=x0+iγ γeitzπ(z−z+)(z−z−) = eix0t−γt (t>0)

Closing in Im(z) ≤ 0 for t < 0 encloses z_− = x₀ − iγ, yielding φ_X(t) = exp(ix₀t − γ|t|) for all t ∈ ℝ. At t = 0, one-sided derivatives jump from ix₀ + γ to ix₀ − γ, confirming non-existence of φ′_X(0) = i𝔼[X].

THEOREM 3.2 · EXACT INVARIANCE AT LÉVY INDEX α = 1

Why Averaging n Independent Cauchy Draws Changes Nothing

For independent X₁, …, Xₙ ~ Cauchy(x₀, γ), homogeneity of degree one in |t/n| = |t|/n cancels sample size n inside the product of characteristic functions for X̄ₙ = n⁻¹ ∑ X_k:

φX¯n(t) = [φX(tn)]n = exp(n⋅ix0tn−nγ|t|n) = φX(t)

In general symmetric α-stable laws exp(−γ^α |t|^α), sample averages scale as n^(1/α − 1): Gaussian (α = 2) narrows as n^(−1/2), whereas Cauchy sits squarely on the critical boundary α = 1 where n⁰ = 1 forever.

4. Voigt Convolution Profile & Faddeeva Wing Dominance

MATH-PROB 99.9 · JS-SIM 99.7
THEOREM 4.1 · FOURIER MULTIPLICATION & FADDEEVA FUNCTION

Convolving Doppler Gaussian Core with Lorentzian Damping

When independent thermal Doppler broadening G ~ N(0, σ²) and collisional/natural lifetime broadening L ~ Cauchy(x₀, γ) act simultaneously, the observed line shape V = G + L has characteristic function φ_V(t) = exp(ix₀t − γ|t| − ½σ²t²). Inverting over ℝ expresses the Voigt density through the complex Faddeeva function w(z) = e^(−z²) erfc(−iz):

V(x;x0,σ,γ) = Re[w(z)]σ2π , z= (x−x0)+iγσ2

In the Drylab engine (drylab.core.js), Humlíček's four-region rational approximation evaluates Re[w(z)] in Im(z) ≥ 0 to better than 10⁻⁴ relative accuracy across both pure Gaussian (γ → 0) and pure Lorentzian (σ → 0) limits.

PROPOSITION 4.2 · ALGEBRAIC INHERITANCE IN THE FAR WINGS

Why Any Nonzero γ Destroys Variance of the Convolution

Expanding w(z) asymptotically for |z| → ∞ in the upper half-plane yields w(z) = i / (√π z) + O(|z|⁻³). Extracting the real part shows that the Gaussian factor vanishes from the leading far-wing term:

V(x;x0,σ,γ) = γπ(x−x0)2 + γ(3σ2−γ2)π(x−x0)4 + O(|x−x0|−6)

Even when γ ≪ σ so the peak appears visually indistinguishable from a Gaussian, the |t| cusp remains present in φ_V(t) at t = 0, forcing Var(V) = +∞.

5. Orthogonal Fisher Information & Dual 8/π² Quantile Efficiency

STAT-NUM 99.8 · MATH-PROB 99.9
THEOREM 5.1 · EXACT 2×2 FISHER INFORMATION MATRIX

Finite Parametric Precision Despite Infinite Moments

Writing z = (x − x₀)/γ, the single-observation log-density ℓ = −ln(πγ) − ln(1 + z²) has bounded score components ∂ℓ/∂x₀ = (2/γ) z/(1 + z²) and ∂ℓ/∂γ = γ⁻¹(z² − 1)/(z² + 1). Substituting z = tan θ evaluates all three trigonometric expectations in closed form:

ℐ(x0,γ) = ( 12γ2 0 0 12γ2 ) , CRB(x^0) = CRB(γ^) = 2γ2n

Vanishing off-diagonal Fisher information proves that location x̂₀ and scale γ̂ are asymptotically orthogonal: estimating γ incurs zero asymptotic variance penalty on x̂₀.

THEOREM 5.2 · PITMAN EFFICIENCY OF MEDIAN & HALF-IQR

Exact Covariance Symmetry of Order-Statistic Estimators

At p = ½, density f_X(x₀) = (πγ)⁻¹ gives Bahadur median variance Var(X̃ₙ) ∼ [4n f_X(x₀)²]⁻¹ = π²γ²/(4n). At quartiles p₁ = ¼, p₂ = ¾ (x₀ ± γ), f_X(x₀ ± γ) = (2πγ)⁻¹, so the joint quantile covariance matrix yields for γ̂_IQR = ½(Q₀.₇₅ − Q₀.₂₅):

Var(γ^IQR) ∼ π2γ24n , ARE(X˜n) = ARE(γ^IQR) = 8π2 ≈ 81.06%

Both robust estimators share the exact same 8/π² relative efficiency, providing deterministic O(n log n) starting points within O(n⁻¹/²) of the global maximum likelihood peak.

6. Conditional Likelihood Traps vs. Copas (1975) Joint Unimodality

STAT-NUM 99.8 · DRYLAB-QA 99.9
THEOREM 6.1 · BARNETT (1966) POLYNOMIAL MULTIMODALITY

Up to n Local Maxima When Scale γ Is Held Fixed

Fixing γ reduces the location score ∂ℓ(x₀ | γ)/∂x₀ = 0 to a rational sum whose common denominator clears to a polynomial of degree 2n − 1:

∑i=1n 2(xi−x0)γ2+(xi−x0)2 =0

Because each Cauchy kernel has negative curvature only inside |x − x_i| < γ and positive curvature outside, observations spaced farther apart than 2γ generate isolated score roots—producing the 3 distinct peaks seen in Section 3's n = 5 preset [-8.2, -7.6, 0.0, 7.5, 8.3].

THEOREM 6.2 · COPAS (1975) & PROFILE ENVELOPE REDUCTION

Why Joint Scale Profiling Eliminates Spurious Peaks

For any candidate x₀ (assuming fewer than n/2 observations coincide), the scale score equation defines a unique smooth profile width γ̂(x₀) > 0 via monotonic bisection:

∑i=1n 11+(xi−x0γ^(x0))2 = n2

By the Envelope Theorem, dℓ(x₀, γ̂(x₀))/dx₀ = ∂ℓ/∂x₀, so every local extremum of the 1D profile curve is a joint stationary point of ℓ(x₀, γ). Copas (Biometrika, 1975) proved that the Hessian of ℓ(x₀, γ) is strictly negative-definite at every stationary point in γ > 0 (up to symmetric even-n ties), guaranteeing a unimodal profile envelope.