When Probability Distribution Graph Rises

To understand the geometry of a discrete PMF without plotting every single bar, examine the consecutive ratio R(x) = P(X = x + 1) / P(X = x):

• If R(x) < 1 everywhere, every successive bar is strictly shorter than the preceding one; the graph falls monotonically from x = 1 (e.g., Geometric distribution).
• If R(x) > 1 for initial values of x, the bars grow taller; the graph rises until R(x) crosses 1, reaches its peak (mode), and subsequently decays (e.g., Negative Binomial, Binomial, Poisson).

Two complementary mental models explain this:

• Model 1 (Analytical): Look at the algebraic ratio of consecutive terms.
• Model 2 (Combinatorial): Count the arrangements. More permutations of arranging intermediate successes temporarily overpower the penalty of additional trials.

Epidemic modeling (e.g., SEIR models of viral reproduction R₀) uses this ratio test to determine whether infection clusters will peak and explode or immediately burn out monotonically.

Abraham de Moivre first solved these discrete ratio inequalities in 1738 in 'The Doctrine of Chances', discovering the central maximum of binomial expansions that foreshadowed the normal curve.

'Ratio Above One, Climbing Has Begun; Ratio Below One, The Peak Is Done'.

Let X ~ Geometric(p) be the number of independent trials required to obtain the first success. The PMF is P(X = x) = (1 - p)^(x - 1) p for x = 1, 2, 3... Evaluating the consecutive ratio: P(X = x + 1) / P(X = x) = [(1 - p)^x p] / [(1 - p)^(x - 1) * p] = 1 - p. Since p in (0, 1], the failure probability (1 - p) is strictly less than 1. Because there is only one possible sequence of outcomes leading to success on trial x (namely x - 1 failures followed by 1 success), there are no alternative arrangements to counteract the compounding decay factor (1 - p). Hence, every bar is smaller than its predecessor by exactly (1 - p).

Site reliability engineers modeling mean time to recover (MTTR) under memoryless restart attempts observe exponential/geometric decay in crash-resolution times.

Jacob Bernoulli formulated the geometric progression of trials in 1713, noting its complete lack of internal memory.

'One Path, Pure Decay' — since there is only 1 path to the first success, the geometric ratio is always (1 - p) < 1.

Let X ~ NegativeBinomial(r, p) be the trial on which the r-th success occurs. The final trial must be a success, but the preceding r - 1 successes can be arranged anywhere among the first x - 1 trials in C(x - 1, r - 1) ways. For r = 3:

• At x = 3: SSS -> 1 arrangement.
• At x = 4: FSSS, SFSS, SSFS -> 3 arrangements.
• At x = 5: 6 arrangements.

Evaluating the consecutive ratio: P(X = x + 1) / P(X = x) = [x / (x - r + 1)] (1 - p). When x is small, the arrangement multiplier [x / (x - r + 1)] is substantially greater than 1! If [x / (x - r + 1)] (1 - p) > 1, the probability of requiring x + 1 trials is strictly greater than x trials, causing the graph to climb.

Sales pipeline analytics and clinical trial recruitment model customer acquisition as a Negative Binomial process: reaching a target quota of r closed enterprise contracts requires a variable number of meetings whose peak probability sits far beyond r.

Montmort studied the problem of points for multiple successes in 1713, establishing the negative binomial series as the multi-success generalization of Bernoulli trials.

'Combinations Lift, Penalties Sift' — the combination term lifts the graph up; compounding failures eventually pull it back down.

A single Geometric waiting period Gᵢ is heavily right-skewed with a peak at 1. However, waiting for r successes is identical to waiting for the 1st success (G₁), then starting fresh and waiting for the 2nd (G₂), up to G_r: X = sum_(i=1)^r Gᵢ. Mathematically, the probability distribution of a sum of independent random variables is the convolution of their individual distributions. As r increases, the Central Limit Theorem (CLT) governs the convolution: extreme delays in one stage are counterbalanced by rapid successes in another, producing an increasingly symmetric, bell-shaped distribution centered at E[X] = r/p.

Project management frameworks (PERT and agile sprint velocity models) sum individual task completion times using Negative Binomial convolutions to compute accurate critical-path completion deadlines.

Pierre-Simon Laplace proved the first general Central Limit Theorem in 1810, demonstrating that repeated sums of independent random causes invariably converge to the Gaussian bell curve.

'Sum the Skewed, Bell is Viewed' — adding independent waiting times washes away one-sided skewness into a symmetric bell.

Evaluate both propositions under n = 50 tosses with fair coin p = 0.5:

• Bet A: Let X be the toss of the 1st Head (Geometric). The probability that the first head occurs within 50 tosses is P(X <= 50) = 1 - P(all 50 tails) = 1 - (0.5)⁵0 approx 1.0 - 8.88 * 10^(-16). Bet A is virtually a guaranteed win.
• Bet B: Let Y be the toss of the 15th Head (Negative Binomial with r = 15). The expected number of tosses to see 15 heads is E[Y] = r/p = 15 / 0.5 = 30 tosses. While 30 < 50, the variance is Var(Y) = 15 * 0.5 / 0.25 = 30, meaning standard deviation is 5.48 tosses. The probability P(Y <= 50) = sum_(k=15)^(50) C(50, k) (0.5)⁵0 = 1 - sum_(k=0)^(14) C(50, k) (0.5)⁵0 = 0.9987.

While both are likely, Bet A's failure probability is 10^(-15) whereas Bet B's risk of failure is over a billion times larger!

Quantitative hedge funds use this multi-event thresholding to evaluate portfolio default risk: avoiding 1 catastrophic default is drastically different from surviving a series of 15 localized market drawdowns.

The Chevalier de Mere posed similar gambling comparison paradoxes to Blaise Pascal in 1654, sparking the formal mathematical calculation of compound betting odds.

'First is Fast, Multi Must Last' — waiting for the 1st success is exponentially more forgiving than accumulating multiple successes.

Inspect the two competing forces in the Negative Binomial ratio:

• Force 1 (Arrangements): [x / (x - r + 1)]. As x starts at r, this ratio is r/1 = r (very large). As x grows, this ratio decreases toward 1.
• Force 2 (Decay): (1 - p), a constant fraction strictly less than 1.

Initially, [x / (x - r + 1)] (1 - p) > 1, so the bars rise. However, as x increases, [x / (x - r + 1)] relentlessly approaches 1. Eventually, the product must drop below 1 because 1 (1 - p) = (1 - p) < 1. The exact peak occurs at the critical value x^ where [x^ / (x^ - r + 1)](1 - p) = 1. Solving for x^: x^(1 - p) = x^ - r + 1 => x^(1 - 1 + p) = r - 1 => x^ = (r - 1)/p. Past this peak, probability decay dominates permanently, pulling the distribution into an exponential tail.

Cloud server autoscaling policies set provisioning buffers at the mode x^* of connection-holding time distributions to ensure peak traffic loads are served with zero queue backlog.

Thomas Simpson established the analytical method for finding discrete modes using consecutive difference ratios in 1740.

'Polynomial Tries, Exponential Dies' — combinations lift the start, but exponential decay always wins the tail.

A universal three-step diagnostic predicts every discrete PMF:

• Step 1: Form the consecutive ratio R(x) = P(X = x + 1) / P(X = x).
• Step 2: Check R(x_min). If R(x_min) < 1, the distribution starts by falling and never rises (e.g., Geometric: R = 1 - p < 1).
• Step 3: If R(x_min) > 1, solve R(x) = 1 for the peak x^. The graph will rise for x < x^, peak at x^, and fall for x > x^ (e.g., Negative Binomial: R = [x/(x - r + 1)](1 - p); Poisson: R = lambda / (x + 1); Binomial: R = [(n - x)/(x + 1)][p / (1 - p)]).

This ratio method serves as the discrete analog of setting the first derivative f'(x) = 0 in continuous calculus.

Automated statistical software (e.g., R's fitdistrplus, Python's scipy.stats) uses ratio checks on empirical frequency data to automatically select candidate distribution families for maximum likelihood estimation.

Karl Pearson formalized the continuous analog of this ratio criterion in 1895 with the Pearson Distribution System, classifying all classical distributions by the differential equation df/dx = (x - a)f / (b₀ + b₁ x + b₂ x²).

'Ratios Rule the Realm' — R > 1 climbs up, R = 1 crowns the king (mode), R < 1 falls down.