Starting Point of Sigma: Support, Random Variables, and Truncation
Who decides where a summation begins: the mathematician or the intrinsic physical support of the problem?
How this lesson unfolds.
Begin with Why does a summation over a random variable start at r rather than 0 or 1, and who makes that decision? The sequence then connects Does the Starting Point Matter: Algebraic Equivalence vs. Support and builds toward Does the Starting Point Matter: Algebraic Equivalence vs. Support.
Why does a summation over a random variable start at r rather than 0 or 1, and who makes that decision?
Summation is not a mechanical ritual starting at 0 or 1. Every Sigma in probability is an accumulator over events. If an event cannot physically occur (e.g., getting 3 heads in fewer than 3 coin flips), its probability is identically zero. Summing from index r reflects the reality that the event space is empty below r.
- $$\operatorname{supp}(X) = \{x \in \mathbb{N} : x \ge r\} \implies \mathbb{E}[X] = \sum_{x=r}^{\infty} x P(X = x)$$
- Learning goalIdentify a random variable's physical support to determine the principled lower summation limit.
- Helpful foundationssigma notation · random variables · Bernoulli trials · Introduction to Mathematical Thinking (Keith Devlin, Stanford) · Essence of Linear Algebra (3Blue1Brown)
- Common trapDefaulting summations to 0 or 1 risks undefined combinations like C(x-1, r-1).
- Next connectionDetermine whether lowering the summation index alters sums or adds zeros.
More context & mechanics
When computing expectations or normalizing probabilities in discrete mathematics, we write:
∑_(x) P(X = x) = 1 and E[X] = ∑_(x) x P(X = x)
Consider the random variable X, defined as the number of independent Bernoulli trials required to observe the r-th success (a Negative Binomial / Pascal distribution). Suppose r = 3.
Can we observe 3 successes on trial 1? No (P(X = 1) = 0). Can we observe 3 successes on trial 2? No (P(X = 2) = 0). The absolute earliest moment the 3rd success can possibly occur is on trial 3 (x = r = 3).
For all x < r, the event is physically impossible, so P(X = x) = 0. Therefore:
∑_(x=r)^(∞) P(X = x) = 1 and E[X] = ∑_(x=r)^(∞) x P(X = x)
The mathematician does not pick the lower bound; the intrinsic physical boundary of the random variable's support Supp(X) = r, r+1, r+2, ṡ dictates it.
In reinforcement learning and telecommunication network routing, engineers calculate time-to-delivery or packet retransmission counts under lossy channels. If a cryptographic handshake requires r verified acknowledgments, the latency distribution begins at r packets. Setting lower bounds correctly prevents buffer underflow errors, division-by-zero anomalies in importance sampling ratios, and off-by-one bugs in Markov chain state transitions.
Leonhard Euler formalized capital Sigma (Σ) for summation in 1755 in Institutiones Calculi Differentialis. Later, in 1933, Andrey Kolmogorov established the axiomatic foundation of probability, defining random variables as measurable functions on a probability space (Ω, F, P), formalizing that summations exist only over the measurable support set where measure is positive.
The Zero-Cost Doorstep Rule:
Before knocking on a door, ask: Can anyone be inside? If the answer is impossible, the probability is 0. Do not start counting until the first possible real event stands at the door: First possible x = r.
Why do textbook formulas sometimes write the sum from 1 and other times from r, and does this change the numerical result?
Algebraic equality is indifferent to the presence of terms that multiply by zero: 0 + 0 + ṡ + v = v. Writing ∑_(x=1)^∞ x P(X=x) and ∑_(x=r)^∞ x P(X=x) gives the exact same total, but writing x=r honors domain integrity and protects combinatorial formulas (like factorials) from evaluating negative arguments.
- $$\sum_{x=1}^{\infty} x P(X=x) = \sum_{x=r}^{\infty} x P(X=x) \quad (P(X < r) = 0)$$
- Learning goalDifferentiate between formal algebraic equivalence and physical support boundaries in discrete series.
- Helpful foundationssigma notation · negative binomial PMF · binomial coefficients · Introduction to Mathematical Thinking (Keith Devlin, Stanford) · Essence of Linear Algebra (3Blue1Brown)
- Common trapEquating algebraic equivalence with domain validity; invalid indices cause undefined combinatorial evaluations.
- Next connectionApply indicator variables and set theory to count members of subsets systematically.
More context & mechanics
Compare the two expressions for the expected value of trials until the r-th success:
∑_(x=1)^(∞) x P(X = x) vs. ∑_(x=r)^(∞) x P(X = x)
Expanding the sum starting at x = 1:
∑_(x=1)^(∞) x P(X = x) = 1 · P(X=1) + 2 · P(X=2) + ṡ + (r-1) · P(X=r-1) + ∑_(x=r)^(∞) x P(X = x)
Because obtaining r successes in fewer than r trials has probability zero:
P(X = 1) = 0, P(X = 2) = 0, ṡ, P(X = r-1) = 0
Each skipped term is x · 0 = 0. Hence:
∑_(x=1)^(∞) x P(X = x) = 0 + 0 + ṡ + 0 + ∑_(x=r)^(∞) x P(X = x) = ∑_(x=r)^(∞) x P(X = x)
The numerical equality holds unconditionally. However, in analytical derivation, using x = r is critical because the PMF contains the binomial coefficient C(x-1, r-1), which is physically meaningless or requires analytic continuation for x < r.
In software implementations (Python, C++, NumPy, PyTorch), looping from 1 to r-1 wastes CPU cycles evaluating zeros. Worse, evaluating C(x-1, r-1) with naive factorial routines when x < r can cause integer underflow or invalid memory access. Vectorized tensor operations in deep learning libraries explicitly define index masks over the valid support to ensure numerical stability and speed.
The binomial coefficient convention C(n, k) = 0 for k > n or n < 0 was codified by Isaac Newton in 1665 when generalizing the Binomial Theorem to fractional and negative exponents, enabling series expansions that seamlessly handle boundary conditions.
The Ghost Seats Rule:
Empty seats in front of the theater don't change the ticket count (0 × \10 = 0), but you don't start the concert until you reach row r where the first ticket holder is sitting!