Glossary - UCF CAP-6318

Key terms from CAP-6318, organized by unit. Each term links to the lecture that introduces it.

Julia¶

abstract type: A type that can never be instantiated directly; it organizes related composite types into a family and lets you define shared methods for the whole family at once. Introduced in Julia Types and Methods.
composite type: A type with named fields holding related data, defined with struct Name ... end; an instance is treated and passed around as a single value. Introduced in Julia Types and Methods.
Dict: A collection that maps keys to values, created with Dict(key => value, ...); also called a dictionary. Introduced in Julia Foundations.
environment: An isolated, named set of installed packages and their versions, described by Project.toml and Manifest.toml, so different projects can depend on different package sets without conflict. Introduced in Julia Package Development.
export: Listing a name inside a module with the export keyword so it is added automatically to a caller’s namespace when they run using ModuleName. Introduced in Julia Code Organization.
include: A function that evaluates the code in a .jl file as if it were pasted directly into the current session — the basic tool for splitting code across files. Introduced in Julia Code Organization.
Julia: A general-purpose programming language built for numerical computation: as easy to read and write as Python, but compiled to run as fast as C. Introduced in Julia Foundations.
Manifest.toml: The companion file to Project.toml that pins the exact version of every package, direct and indirect, in an environment, so it can be reproduced exactly. Introduced in Julia Package Development.
method: One specific implementation of a function, specialized to a particular combination of argument types (e.g. isint(x::Integer) = true); a function can have many methods, chosen by multiple dispatch. Introduced in Julia Types and Methods.
module: A named namespace (module Name ... end) that bundles related code so its contents don’t automatically pollute the caller’s session; members are reached via Module.name or brought in with using. Introduced in Julia Code Organization.
multiple dispatch: Choosing which method of a function runs based on the types of all its arguments, not just the first. Julia’s core design idea. Introduced in Julia Types and Methods.
named tuple: A hybrid of a tuple and a Dict: an ordered, immutable collection whose elements have names, created with (name = value, ...). Introduced in Julia Foundations.
package: A unit of reusable Julia code published and installed through Julia’s package manager (Pkg), extending the language with new types and functions. Introduced in Julia Foundations.
package composability: The property that packages and built-in code get the same treatment in Julia, so a package can operate on types defined by a completely different package (e.g. plotting a DataFrame). Introduced in Julia Foundations.
Pkg: Julia’s built-in package manager, used to add, remove, and update packages — either as the Pkg module (Pkg.add(...)) or the ] REPL mode. Introduced in Julia Package Development.
primitive type: A built-in type representing a fixed collection of bits, e.g. Float64, Bool, Int8; defined by Julia itself, not typically created by users. Introduced in Julia Types and Methods.
Project.toml: The TOML file listing an environment’s direct package dependencies as NAME = UUID pairs; created and updated automatically by Pkg. Introduced in Julia Package Development.
REPL: Julia’s Read-Eval-Print-Loop, the interactive julia> prompt where you type an expression and immediately see its value. Introduced in Julia Setup.
standard library: The set of packages (e.g. Dates) that ship bundled with Julia itself, already installed and ready to using. Introduced in Julia Foundations.
TOML: “Tom’s Obvious Markup Language” — the plain-text configuration format used by Project.toml and Manifest.toml to describe a Julia environment. Introduced in Julia Package Development.
tuple: An immutable, ordered collection written with parentheses, e.g. (1, 2, "hello"); unlike an array, a tuple is not meant to be updated or used in linear algebra. Introduced in Julia Foundations.
type: What kind of object lives at a given memory address. Julia uses a value’s type to decide which syntax and which methods apply to it. Introduced in Julia Types and Methods.
type hierarchy: The tree of Julia types rooted at Any (every object) with Union{} at the bottom (no objects); every type has at most one parent. Introduced in Julia Types and Methods.

Networks¶

adjacency matrix: A matrix $A\in\{0,1\}^{N\times N}$ for a graph of $N$ nodes, with $A_{ij}=1$ if there is an edge between nodes $i$ and $j$ . A weighted graph generalizes the entries to real-valued weights. Introduced in Graphs.
Balance Theorem: If a labeled complete graph is balanced, then either all pairs of nodes are friends, or the nodes split into two groups X and Y such that every pair within X or within Y are friends and every pair across X and Y are enemies. Introduced in Structural Balance.
Bellman equation: The recursive restriction $J(v)=\min_{n\in F_v} w_{vn}+J(n)$ satisfied by the cost-to-go function $J$ , where $F_v$ is the set of nodes reachable from $v$ ; used to compute shortest paths by guess-and-iterate. Introduced in Weighted Graphs.
betweenness centrality: For node $v$ , $C_B(v)=\sum_{s\ne v\ne t}\sigma_{st}(v)/\sigma_{st}$ , where $\sigma_{st}$ is the number of shortest paths between $s$ and $t$ and $\sigma_{st}(v)$ is the number passing through $v$ ; captures how much flow passes through a node. Introduced in Strong and Weak Ties.
breadth-first search: An algorithm that computes distances from a starting node by visiting nodes layer by layer, in order of increasing distance. Introduced in Graphs.
bridge (graph): An edge that, if removed, would split the nodes it connects into different components of the graph. Contrast with the weaker local bridge. Introduced in Strong and Weak Ties.
clustering coefficient: The fraction of a node’s (or graph’s) possible triangles that are actually closed; equivalently, the probability that two random friends of a node are themselves friends. Introduced in Strong and Weak Ties.
component (graph): A self-contained subset of nodes such that every node in it has a path to every other, and the subset is not part of any larger such subset. Introduced in Graphs.
connected graph: A graph in which there is a path between every pair of nodes, so any node can be reached from any other. Introduced in Graphs.
cost-to-go: $J(v)$ , the minimum cumulative weight needed to travel from node $v$ to a fixed destination along a weighted graph. Introduced in Weighted Graphs.
cycle: A path that starts and ends at the same node. Introduced in Graphs.
degree: The number of edges connected to a node, $\sum_j A_{ij}$ ; a basic measure of node importance. High-degree nodes are often called hubs. Introduced in Graphs.
demand shock: A one-time change $\Delta d$ in the final-demand vector; the resulting change in equilibrium output is $\Delta x=L\Delta d$ , decomposable round-by-round via the Neumann series lemma as $\Delta d+A\Delta d+A^2\Delta d+\cdots$ . Introduced in Production Networks.
directed graph: A graph in which each edge has a direction, so a connection flows from one specific node to another rather than being symmetric. Introduced in Graphs.
distance (graph): The length of the shortest path between two nodes, where length is the number of edge-steps along the path. Introduced in Graphs.
edge: A relationship between two nodes in a graph; may be directed or undirected, weighted or unweighted, and in a signed graph labeled + or -. Introduced in Graphs.
eigenvalue: For $A\in\mathbb{R}^{n\times n}$ , a scalar $\lambda\in\mathbb{C}$ with $Ae=\lambda e$ for some nonzero vector $e$ (the eigenvector). The set of all eigenvalues is $A$ ’s spectrum, $\sigma(A)$ . Introduced in Production Networks.
eigenvector: A nonzero vector $e$ satisfying $Ae=\lambda e$ for an eigenvalue $\lambda$ ; a vector $\epsilon$ with $A'\epsilon=\lambda\epsilon$ is a left-eigenvector. Introduced in Production Networks.
eigenvector centrality: A node-importance measure equal to the dominant eigenvector of a graph’s (weighted) adjacency matrix — a node scores higher if it connects to other highly central nodes. Introduced in Production Networks.
giant component: A single component containing most of a network’s nodes, with the rest scattered in small components or isolated — a common structure in real-world networks. Introduced in Graphs.
Girvan-Newman algorithm: A divisive graph partitioning algorithm that repeatedly removes the edges with highest betweenness centrality and recomputes centrality on the resulting pieces. Introduced in Strong and Weak Ties.
graph: A structure that records relationships between a collection of items: each item is a node and each relationship is an edge. Introduced in Graphs.
graph partitioning: The algorithmic problem of breaking a graph into tightly-knit regions, typically by removing local bridges (divisive methods) or growing regions from a seed node (agglomerative methods). Introduced in Strong and Weak Ties.
graph theory: The branch of mathematics that studies graphs: collections of nodes and the edges that connect them. It is the language this course uses for networks. Introduced in Graphs.
homophily: The tendency for connected nodes to be more similar to each other than nodes further apart in the network — “we are similar to our friends.” Introduced in Homophily.
input-output coefficient: $a_{ij}=\dfrac{\text{value of sector }j\text{'s inputs from sector }i}{\text{total sales of sector }j}$ , the entry of the input-output matrix $A$ recording how much sector $j$ buys from sector $i$ per dollar of its own sales. Introduced in Production Networks.
inverse homophily: A pattern in which the observed cross-characteristic edge share is well above the random-formation benchmark, i.e. more cross-type edges than random mixing would predict. Introduced in Homophily.
Leontief inverse: The matrix $L\equiv(I-A)^{-1}$ that maps a final-demand vector $d$ to the equilibrium output vector $x^*=Ld$ , given input-output coefficient matrix $A$ with $r(A)<1$ . Introduced in Production Networks.
Leontief model: A family of production-network models (Leontief, 1941) built on an input-output table, used to analyze an economy as sectors linked by intermediate-good flows. Introduced in Production Networks.
local bridge: An edge A-B that is a local bridge if A and B have no neighbors in common — a looser, more common relative of a true bridge. Introduced in Strong and Weak Ties.
Neumann series lemma: If $A\in\mathbb{R}^{n\times n}$ and $r(A)<1$ , then $I-A$ is nonsingular and $(I-A)^{-1}=\sum_{m=0}^{\infty}A^m$ . Introduced in Production Networks.
node: A single item in a graph, connected to other nodes by edges. Graphs.jl calls these “vertices.” Introduced in Graphs.
path: A flow from one node to another across a sequence of edges. Introduced in Graphs.
random-formation benchmark: The cross-characteristic edge probability $2p_x(1-p_x)$ expected if each edge endpoint’s type were drawn independently from the population share $p_x$ of characteristic $X$ ; the comparison point used to test for homophily. Introduced in Homophily.
shortest path: In a weighted graph with edge weights $w_{ab}$ , the path between two nodes that minimizes the sum $\sum w_{ab}$ of weights along it, rather than the number of edges. Introduced in Weighted Graphs.
signed graph: An undirected graph where every edge is labeled + (friends) or - (enemies); structural balance analysis restricts attention to complete signed graphs. Introduced in Structural Balance.
spectral radius: $r(A)\equiv\max\{|\lambda|:\lambda\in\sigma(A)\}$ , the largest-magnitude eigenvalue of $A$ ; governs convergence of $x_{t+1}=Ax_t$ and of the Neumann series lemma. Introduced in Production Networks.
strong tie: A label on an edge marking a close friendship, as opposed to a weak tie. Introduced in Strong and Weak Ties.
Strong Triadic Closure Property: A node A violates this property if it has strong ties to both B and C but there is no edge at all, strong or weak, between B and C; a node satisfies it if it never violates it. Introduced in Strong and Weak Ties.
structural balance: A complete signed graph satisfies the Structural Balance Property if every set of three nodes has exactly one or three positive edges among them. Introduced in Structural Balance.
tree: A connected graph with no cycles; a tree on $N$ nodes has exactly $N-1$ edges, the minimum needed to stay connected. Introduced in Graphs.
triadic closure: The empirical tendency for an edge B-C to form when edges A-B and A-C already exist — the open side of a triangle tends to close. Introduced in Strong and Weak Ties.
triangle (graph): The smallest structure involving more than two nodes: three nodes joined by either two edges (open triangle) or three edges (closed triangle). Introduced in Strong and Weak Ties.
weak tie: A label on an edge marking an acquaintance rather than a close friendship. A local bridge must be a weak tie. Introduced in Strong and Weak Ties.
weighted graph: A graph in which each edge carries a numeric weight denoting strength, rather than just presence or absence of a connection. Introduced in Weighted Graphs.

Agent-Based Models¶

agent (ABM): An autonomous, reactive entity in an agent-based model (ABM) with its own agent state at each time step, aware of that state and able to communicate select properties to other agents. Introduced in Agent Based Models.
agent state: The current values of an agent's properties at a given time step — e.g. position, happiness, wealth, citation count. Introduced in Agent Based Models.
agent-based model (ABM): A model that specifies rules for how individual agents respond to their state and environment, rather than solving a global optimization problem. Built from three pieces: agents, an environment, and rules. Introduced in Agent Based Models.
Agents.jl: A Julia library for building and running agent-based models: agent macros, built-in environments like GridSpace, schedulers that set agent activation order, and visualization/exploration tools. Introduced in Schelling Model with Agents.jl.
calibration: Fitting a model’s parameters to match statistics observed in real data, rather than choosing them arbitrarily. Introduced in Academic Paper Virality: An Agent-Based Model.
citation network: A directed graph where nodes are papers and an edge from paper A to paper B means A cites B; since papers can only cite earlier work, the graph has no cycles in time. Introduced in Academic Paper Virality: An Agent-Based Model.
emergence: A complex, often surprising aggregate pattern — or, at large enough scale, a qualitatively new capability — that arises from simple local rules or from scale, without being intended or designed by any individual part. Segregation in the Schelling segregation model is the canonical example; new reasoning or tool-use abilities appearing once a language model crosses a size threshold are another. Introduced in Agent Based Models.
environment (ABM): The space an agent acts in — natural or artificial (a grid, a market, a network). Its properties can be fixed or varying, and agents can observe and respond to them. Introduced in Agent Based Models.
equation-based model: A model that describes each agent’s decision-making with equations, typically a constrained optimization problem, plus equations for how agents interact. Precise but hard to calibrate at scale. Contrast with agent-based model (ABM). Introduced in Agent Based Models.
GridSpace: A grid-based environment in Agents.jl where each agent occupies a cell and has nearby_agents it can interact with; the grid can be bounded or periodic (opposite edges wrap around, like Pac-Man). Introduced in Schelling Model with Agents.jl.
heterogeneity: Variation across agents in properties or behavior. Agent-based models capture heterogeneity directly, which is why individual differences, not just population averages, can drive social outcomes. Introduced in Agent Based Models.
model: A simplified mathematical object — equations, rules, or distributional assumptions — that approximates a real-world system well enough to reason about it. A model describes how likely different outcomes are, not one exact outcome. Introduced in Agent Based Models.
money model: A minimal agent-based model (ABM) with no environment: $N$ agents each start with 1 unit of wealth, and each step an agent with wealth gives 1 unit to a randomly chosen agent. Despite the symmetric rule, wealth concentrates into a power law after a few steps. Also called the wealth exchange model. Introduced in Agent Based Models Review/Practice.
path dependence: The property that history and timing shape outcomes: small early differences — who acts first, an early lead in adoption or citations — can compound into large differences later. Introduced in Agent Based Models.
power law: A distribution where a small number of agents hold a disproportionately large share of some quantity (wealth, citations) while most agents have very little — the “rich get richer” pattern. Introduced in Agent Based Models Review/Practice.
preferential attachment: A selection rule where agents, papers, or nodes with more existing connections or citations are more likely to gain new ones — “rich get richer.” Drives the citation counts in the paper-virality model. Introduced in Academic Paper Virality: An Agent-Based Model.
rule (ABM): The equations or functions that update an agent's state between time steps $t$ and $t+1$ . Shared across agents of a type but applied to each agent’s own state, often involving neighbors or randomness. Introduced in Agent Based Models.
Schelling segregation model: An agent-based model (ABM), proposed by Thomas Schelling in 1969, where agents of two types live on a grid and relocate if fewer than a threshold share of their neighbors share their type. Mild individual preferences produce strong neighborhood segregation, a classic case of emergence. Introduced in Agent Based Models.
sleeping beauty: In the paper-virality model, a paper that is cited rarely at first but accumulates citations later — the opposite pattern from a viral paper. Introduced in Academic Paper Virality: An Agent-Based Model.
step function: The function(s) that advance an agent-based model (ABM) by one time step: agent_step! defines how a single agent updates itself, while step!/run! apply it (and any model_step!) across all agents and collect data. Introduced in Schelling Model with Agents.jl.
viral paper: In the paper-virality model, a paper whose citation count grows quickly early in its life, in contrast to a sleeping beauty. Introduced in Academic Paper Virality: An Agent-Based Model.

Game Theory¶

all-pay auction: An auction format in which every bidder pays their bid regardless of outcome, but only the highest bidder wins the item. Also satisfies revenue equivalence. Introduced in Auctions.
auction: A market mechanism between one seller and many buyers in which rules determine how buyers signal willingness to pay, used especially when a good’s valuation is private. Introduced in Auctions.
backward induction: The method for solving a dynamic game’s extensive form by reasoning from the last decision nodes back to the first, determining each player’s optimal choice given what happens afterward. Introduced in More Game Theory.
best response: The action for a player that yields the highest payoff given the strategy, or belief about the strategy, of the other players. Introduced in Game Theory 1.
bid shading: Bidding below one’s true valuation, as is optimal in a first-price auction, trading off the risk of losing against the size of the payoff from winning. Introduced in Auctions.
bimatrix: For an $N$ -player game with $N_i$ strategies each, an array holding every player’s payoff directly in each cell, avoiding the separate-payoff matrix-per-player bookkeeping. Introduced in Game Theory 1.
Braess’ paradox: The counterintuitive result that adding capacity, a new link, to a network can make the Nash equilibrium outcome worse for everyone. Introduced in Network Traffic with Game Theory.
commitment: The ability of a player to credibly bind themselves to a stated strategy before others act, which can change what other players rationally do in response. Introduced in More Game Theory.
credible: A strategy or threat that a player would actually be willing to carry out if called upon, as opposed to one that is part of a Nash equilibrium on paper but irrational to execute. Introduced in More Game Theory.
dominant strategy: A strategy that is a best response no matter what the other players do; strictly dominant if it is always strictly better than every alternative. Introduced in Game Theory 1.
Dutch auction: A descending-price auction: price starts high and falls until a bidder accepts the current price, ending the auction there. Introduced in Auctions.
dynamic game: A game in which players make choices sequentially, at different known times, rather than simultaneously. Introduced in More Game Theory.
English auction: An ascending-price auction: the auctioneer calls increasingly higher prices, bidders indicate willingness to pay, and it ends when no one outbids the current highest bid. Introduced in Auctions.
expected payoff: A player’s payoff averaged over the probabilities implied by a mixed strategy, e.g. $p_1(H\mid q)=-1\cdot q+1\cdot(1-q)$ for player 1 facing opponent mix $q$ . Introduced in More Game Theory.
extensive form: A representation of a dynamic game as a tree, where nodes are decision points and branches are the actions available at each node. Introduced in More Game Theory.
first-price auction: A sealed-bid auction where all bidders submit bids simultaneously; the highest bidder wins and pays the amount of their own bid. Introduced in Auctions.
game: A description of a strategic environment: a finite set of $N$ players, a feasible action set $S_i$ for each, and a payoff function $p_i:\Sigma\to\mathbb{R}$ for each. Introduced in Game Theory 1.
game theory: The study of strategic interaction: how decision-makers choose when each one’s payoff depends on what the others do. A game writes such a situation down precisely. Introduced in Game Theory 1.
indifference: The requirement that, in a mixed strategy Nash equilibrium, every pure strategy a player mixes over yields the same expected payoff; otherwise the player would switch entirely to the better one. Introduced in More Game Theory.
iterated deletion of dominated strategies: An algorithm that repeatedly removes, for each player, any strategy dominated given the opponents’ remaining strategies, until a unique profile (a Nash equilibrium) or an empty set remains. Introduced in Game Theory 1.
mixed strategy: A probability distribution over a player’s set of pure strategies, $\sigma_i\in\triangle(S_i)\subseteq[0,1]^{M_i}$ ; a pure strategy is the degenerate special case. Introduced in Game Theory 1.
Nash equilibrium: A strategy profile $\sigma$ such that each player’s $\sigma_i$ is a best response to everyone else’s $\sigma_{-i}$ ; no player wants to unilaterally deviate. Introduced in Game Theory 1.
payoff: The numerical reward $p_i(\sigma)$ a player receives for a given outcome $\sigma$ of the game. Introduced in Game Theory 1.
payoff matrix: A matrix $P_i\in\mathbb{R}^{M_1\times M_2}$ giving player $i$ ’s payoff for each combination of the two players’ actions; row $i$ is player 1’s action, column $j$ is player 2’s. Introduced in Game Theory 1.
player: One of the $N$ decision-makers in a game, each with their own action set and payoff function. Introduced in Game Theory 1.
price of anarchy: The ratio $\text{PoA}=\dfrac{\text{Worst Nash Equilibrium Cost}}{\text{Social Optimum Cost}}$ , measuring how much self-interested behavior degrades outcomes compared to central coordination. Introduced in Network Traffic with Game Theory.
pure strategy: Choosing one action outright with certainty, as opposed to randomizing over actions with a mixed strategy. Introduced in Game Theory 1.
revelation principle: The principle that, to derive optimal bidding, attention can be restricted to strategies derived directly from a bidder’s true valuation, since such a strategy’s expected payoff is at least as good as one based on any other value. Introduced in Auctions.
revenue equivalence: The result that, under equilibrium bidding, a seller’s expected revenue is the same across first-price auction, second-price auction, and all-pay auction formats. Introduced in Auctions.
second-price auction: A sealed-bid auction, also called a Vickrey auction, where all bidders submit bids simultaneously; the highest bidder wins but pays the amount of the second-highest bid. Truth-telling is a weakly dominant strategy here. Introduced in Auctions.
social optimum: The outcome a social planner would choose — the one maximizing total social welfare — as opposed to the possibly worse outcome reached at a Nash equilibrium. Introduced in Network Traffic with Game Theory.
social planner: A hypothetical authority that chooses actions for all agents to maximize social welfare, rather than letting each agent act in their own self-interest. Introduced in Network Traffic with Game Theory.
social welfare: The aggregate, summed payoff across all players in a game. Introduced in Network Traffic with Game Theory.
strategy: A player’s choice from their feasible action set $S_i$ ; in a dynamic game, a complete plan of action at every decision point. Introduced in Game Theory 1.
support enumeration: An algorithm (used by GameTheory.jl’s support_enumeration) that computes every Nash equilibrium of a game, pure or mixed. Introduced in Game Theory 1.
truth-telling: Bidding exactly one’s true valuation, $b_i=v_i$ ; a weakly dominant strategy in a second-price auction. Introduced in Auctions.
valuation: A bidder’s private dollar value for winning the auctioned item, $v_i$ . Introduced in Auctions.
zero-sum game: A game in which the payoffs of all players always sum to zero — one player’s gain is exactly another’s loss. Introduced in More Game Theory.

Platforms and Marketplaces¶

blocking pair: A pair $(d,h)$ , from opposite sides of a matching market, such that $d$ prefers $h$ to its current match and $h$ prefers $d$ to its current match — both would rather abandon the matching and pair with each other. Introduced in Matching Markets: When Prices Can’t Do the Job.
chicken-and-egg problem: The platform-launch problem in which neither side will join without the other already present — diners won’t carry a card few restaurants accept, restaurants won’t accept a card few diners carry. Introduced in Platforms and Matchmakers.
competitive bottleneck: The pattern where platforms court the side that single-homes and tax the side that multihomes, because access to single-homers is what rival platforms compete over. Introduced in Platform Pricing: Who Pays, Who Gets Paid.
critical mass: The unstable middle fulfilled-expectations equilibrium: below it, adoption collapses toward zero; above it, feedback pushes adoption toward the high equilibrium. Introduced in Platforms and Matchmakers.
cross-side externality: The effect a price change on one side of a two-sided market has on the other side’s value through shared volume: raising $p_B$ loses buyer sales directly and reduces every seller’s completed transactions, even though no seller left. Introduced in Platform Pricing: Who Pays, Who Gets Paid.
cross-side network effect: A network effect running from one group to another: more drivers improve service for riders, more riders create earnings for drivers. The feedback can reinforce itself across sides. Introduced in Platforms and Matchmakers.
deferred acceptance: Gale and Shapley’s algorithm: unmatched proposers propose to their best remaining choice each round; receivers tentatively hold their best offer and reject the rest; the process always terminates at a stable matching. Introduced in Matching Markets: When Prices Can’t Do the Job.
equilibrium selection: The problem of which Nash equilibrium a coordination game settles into when more than one exists, e.g. choosing between everyone joining and everyone staying out in the chicken-and-egg problem. Introduced in Ignition: Chicken-and-Egg, Lock-In, and Tipping.
fulfilled-expectations equilibrium: An adoption level $n^*$ at which expected and realized network size agree: $n^*=g(n^*;p)$ . Expectations determine adoption; adoption then confirms or contradicts the expectation. Introduced in Platforms and Matchmakers.
increasing returns: A dynamic in which a platform’s attractiveness grows more than proportionally with its installed base, modeled as adoption probability $P(A\mid t)=n_A(t)^\gamma/(n_A(t)^\gamma+n_B(t)^\gamma)$ with $\gamma>1$ . Introduced in Ignition: Chicken-and-Egg, Lock-In, and Tipping.
lock-in: The outcome in which later adopters rationally follow a lead created by earlier, possibly random, accidents, so a market settles on one platform even absent a quality difference. Introduced in Ignition: Chicken-and-Egg, Lock-In, and Tipping.
matchmaker: What a platform is: an intermediary whose product is the connection itself, since each side joins partly because the other side is there. Introduced in Platforms and Matchmakers.
multihoming: Participating on more than one platform at once (e.g. riders install both Uber and Lyft). Multihoming weakens the feedback that locks a market into one winner and shapes who a platform subsidizes under a competitive bottleneck. Introduced in Platforms and Matchmakers.
network effect: One participant’s presence changing another participant’s value from joining a network. Splits into a same-side network effect (within a group) and a cross-side network effect (across groups). Introduced in Platforms and Matchmakers.
pipe: A business that creates value by transforming inputs into outputs along one production chain, so no participant’s value depends on who else shows up. Contrast with platform. Introduced in Platforms and Matchmakers.
platform: A business that creates value by connecting two or more distinct groups who need each other, rather than by transforming inputs into outputs along one chain. Contrast with pipe. Introduced in Platforms and Matchmakers.
price level: How much a platform collects in total across both sides of an interaction, $p_B+p_S$ . Contrast with price structure. Introduced in Platforms and Matchmakers.
price structure: How a platform’s total price is divided between sides. Holding the price level fixed, changing the split can still change participation because of a cross-side network effect. Introduced in Platforms and Matchmakers.
proposer-optimality: The property that the side that proposes in deferred acceptance ends up with its best outcome among all stable matchings — the choice of who proposes is itself a form of market power. Introduced in Matching Markets: When Prices Can’t Do the Job.
repugnant market: A market, such as organ transplants, where using prices to clear supply and demand is illegal or socially unacceptable, motivating stable matching-based market design instead of pricing. Introduced in Matching Markets: When Prices Can’t Do the Job.
same-side network effect: A network effect running among participants in the same group; can be positive (more gamers, more fun) or negative (more riders, more congestion). Introduced in Platforms and Matchmakers.
single-homing: Participating on only one platform. Contrast with multihoming; single-homers are the users platforms court with low or negative prices under a competitive bottleneck. Introduced in Platform Pricing: Who Pays, Who Gets Paid.
stable matching: A matching $\mu$ pairing each member of one side with one of the other side such that it contains no blocking pair. Introduced in Matching Markets: When Prices Can’t Do the Job.
strategy-proofness: The property that truthful preference reporting is a dominant strategy for the proposing side in deferred acceptance. Introduced in Matching Markets: When Prices Can’t Do the Job.
threshold adoption rule: A diffusion model where node $i$ adopts once at least a fraction $\tau$ of its neighbors have adopted: $\sum_{j\in N(i)}a_j(t)/|N(i)|\ge\tau$ . Shows how network topology, not just seed count, shapes critical mass. Introduced in Ignition: Chicken-and-Egg, Lock-In, and Tipping.
tipping: Self-reinforcing movement of adoption toward a high, or low, equilibrium once critical mass is crossed. Introduced in Platforms and Matchmakers.
two-sided market: A market in which a platform’s profit depends on participation from two distinct groups whose demands interact, formalized by the Rochet-Tirole model $\pi(p_B,p_S)=(p_B+p_S-c)D_B(p_B)D_S(p_S)$ . Introduced in Platform Pricing: Who Pays, Who Gets Paid.
winner-take-all: The prediction that strong network effects concentrate an entire market on a single platform. Not automatic: multihoming and local network effects can break it. Introduced in Ignition: Chicken-and-Egg, Lock-In, and Tipping.

AI Agents¶

AI agent: An LLM wrapped with a role, system prompt, memory, and often tools, that pursues a task rather than just answering one prompt. Distinct from the agent of the Agent-Based Models unit, which follows fixed rules rather than reasoning with an LLM. Introduced in From Rule-Based to Learning Agents.
chain-of-thought reasoning: Asking a model to show its reasoning steps before answering; each step becomes context for the next and improves accuracy on complex tasks. Introduced in From Rule-Based to Learning Agents.
chunking: Splitting a document into smaller, focused pieces before embedding, so retrieval can return precise passages instead of whole documents. Introduced in Retrieval-Augmented Generation: Augmenting AI with External Knowledge.
confused deputy: A security failure where an AI agent has legitimate tool access but is tricked, via injected instructions, into using it on the attacker’s behalf instead of the user’s. Introduced in Security for AI Agents: The Lethal Trifecta.
context window: The maximum number of tokens an LLM can see at once, its working memory. Long conversations eventually fall outside it. Introduced in From Rule-Based to Learning Agents.
cosine similarity: A similarity score between two vectors, $\cos(\mathbf{a},\mathbf{b})=\dfrac{\mathbf{a}\cdot\mathbf{b}}{\lVert\mathbf{a}\rVert\,\lVert\mathbf{b}\rVert}$ , ranging from -1 to 1; used to rank embeddings by meaning. Introduced in Retrieval-Augmented Generation: Augmenting AI with External Knowledge.
dependency injection: Passing an AI agent’s external resources — database, API keys, config — in explicitly at run time rather than hard-coding them, so tools stay testable and swappable. Introduced in Type-Safe Agent Development with PydanticAI Patterns.
deterministic evaluator: A code-based check with a clear pass/fail rule — exact match, type check, format validation — used in an evaluation when correctness is objective. Introduced in Evaluating AI Systems: Testing Agents at Scale.
dual LLM pattern: A defense that splits an agent into a privileged LLM, which has tools but never sees untrusted data, and a quarantined LLM, which sees untrusted data but has no tools, connected only by clean symbolic references. Introduced in Security for AI Agents: The Lethal Trifecta.
embedding: A function $f:\text{Text}\to\mathbb{R}^d$ mapping text to a vector so that texts with similar meaning land close together. Introduced in Retrieval-Augmented Generation: Augmenting AI with External Knowledge.
evaluation: Systematic testing of a non-deterministic AI system: a dataset of cases run through the system and scored by evaluators, analogous to a test suite. Also called evals. Introduced in Evaluating AI Systems: Testing Agents at Scale.
hallucination: An LLM generating confident, plausible-sounding text that is factually wrong — most dangerous when it invents citations or data. Introduced in Retrieval-Augmented Generation: Augmenting AI with External Knowledge.
in-context learning: Teaching an LLM a new task by showing examples directly in the prompt, with no retraining needed. Also called few-shot prompting. Introduced in From Rule-Based to Learning Agents.
indirect prompt injection: Prompt injection hidden inside external content an agent processes — an email, a web page, a GitHub issue — rather than typed directly by the user; the dangerous, hard-to-detect form. Introduced in Security for AI Agents: The Lethal Trifecta.
JSON Schema: A structured, machine-readable description of a function’s name, purpose, and parameters that tells an LLM how to call it correctly. Introduced in Function Calling and Tool Use: From Talk to Action.
knowledge cutoff: The date after which an LLM has seen no training data, so it cannot know later events, papers, or statistics. Introduced in Retrieval-Augmented Generation: Augmenting AI with External Knowledge.
large language model, LLM: A neural network trained to predict the next token in a sequence. At scale this simple objective produces a general-purpose reasoning engine. Introduced in From Rule-Based to Learning Agents.
lethal trifecta: The combination of access to private data, exposure to untrusted content, and the ability to exfiltrate data in one AI agent; each is fine alone, but all three together make data theft near-inevitable. Introduced in Security for AI Agents: The Lethal Trifecta.
LLM-as-judge: Using a second LLM to score an output against a rubric when correctness is subjective, such as tone or faithfulness, rather than exact-match checkable. Introduced in Evaluating AI Systems: Testing Agents at Scale.
MCP server: A process, built with e.g. FastMCP, that exposes a set of tools, resources, and prompts over the MCP for any compatible client to call. Introduced in Building Agent Tools with FastMCP.
Model Context Protocol, MCP: A standard protocol, “USB-C for AI,” for exposing tools, resources, and prompts to any AI client, so a tool is written once and used by many agents and apps. Introduced in Building Agent Tools with FastMCP.
multi-agent system: A collection of AI agents, each with a distinct role, that communicate to solve problems no single agent could solve alone. Introduced in Multi-Agent Conversations and Emergent Collaboration.
parametric knowledge: What an LLM “knows” from training, encoded in its weights — fast and broad, but frozen at training time and prone to hallucination. Introduced in Retrieval-Augmented Generation: Augmenting AI with External Knowledge.
principle of least privilege: Giving an AI agent only the minimum tools and data access it needs for its task — the core defensive design rule against excessive agency. Introduced in Security for AI Agents: The Lethal Trifecta.
prompt: Natural-language instructions that tell an LLM what to do. Prompting is programming in natural language with probabilistic outcomes. Introduced in From Rule-Based to Learning Agents.
prompt injection: An attack where malicious text manipulates an LLM into ignoring its instructions and following the attacker’s instead, because the model cannot distinguish trusted instructions from untrusted data. Introduced in Security for AI Agents: The Lethal Trifecta.
retrieval-augmented generation, RAG: Augmenting an LLM with a retrieval step: search a knowledge base for relevant text, inject it as context, then generate an answer grounded in that text. Introduced in Retrieval-Augmented Generation: Augmenting AI with External Knowledge.
sandboxing: Running agent-executed code or tools in an isolated environment with limited resources and no access to the host system, to contain damage if the agent is compromised. Introduced in Function Calling and Tool Use: From Talk to Action.
span-based evaluation: Checking an AI agent’s execution trace — which tools it called, in what order — rather than just its final output, to catch right answers reached the wrong way. Introduced in Evaluating AI Systems: Testing Agents at Scale.
spotlighting: A defense that wraps untrusted content in explicit markers and instructs the model never to follow instructions found inside them. Introduced in Security for AI Agents: The Lethal Trifecta.
structured output: Forcing an LLM's response to conform to a predefined schema, e.g. a Pydantic model, instead of free text, so the result is guaranteed parseable and valid. Introduced in Type-Safe Agent Development with PydanticAI Patterns.
system prompt: The instructions that set an AI agent’s persistent role and behavior, separate from the per-turn user message. Introduced in Multi-Agent Conversations and Emergent Collaboration.
token: The unit an LLM predicts and consumes, roughly a word piece. Context windows and API costs are measured in tokens. Introduced in From Rule-Based to Learning Agents.
tool use: Giving an LLM access to external functions it can invoke, described to it via a JSON Schema, so it can act instead of just generating text. Also called function calling. Introduced in Function Calling and Tool Use: From Talk to Action.
type safety: Validating every input and output against a defined schema before it is used, so invalid data is rejected immediately instead of corrupting state downstream. Also called validation-first development. Introduced in Type-Safe Agent Development with PydanticAI Patterns.
vector store: A system that stores embeddings alongside their source text and supports fast similarity search over them. Also called a vector database. Introduced in Retrieval-Augmented Generation: Augmenting AI with External Knowledge.