Dominance Doctrine: Theory & Practice

• Dominance Doctrine is a framework of universal comparison principles used to evaluate effectiveness across strategic, mathematical, and biological contexts.
• It formalizes performance by employing order relations, game-theoretic refinements, and probabilistic measures to compare competing entities.
• In AI forecasting, the doctrine underpins the theory that the first actor to achieve advanced AI may secure decisive strategic, military, and economic dominance.

Dominance doctrine denotes a recurrent family of comparison principles rather than a single unified theory. Across current technical literatures, the expression is used for frameworks in which one object dominates another because it is at least as effective, informative, positive, or complex under a universal comparison over contingencies, utilities, strategies, representations, or geometric states. In a narrower AI-forecasting sense, the term denotes the claim that the first actor to achieve sufficiently advanced AI, especially ASI, could obtain a decisive strategic advantage over all others (Fatima et al., 2 Mar 2026, Amadori et al., 24 Sep 2025).

1. Order-theoretic core

A common formal pattern is the replacement of informal superiority by a preorder or partial order. In representation theory and algebraic geometry, the relevant order is dominance or majorization on partitions: for partitions $a,b$ of the same weight,

$$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$

When the weights differ, comparison is normalized by weight, yielding the condition $|b|a \succeq |a|b$. In linear-inequality formulations, weak and strict dominance are distinguished as

$$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$

In nondeterministic process theory, strategic dominance is expressed as

$$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$

so that every deterministic refinement of one model is covered by some deterministic refinement of another (Fatima et al., 2 Mar 2026, Ball, 2023, Henzinger et al., 2024).

Taken together, these formulations suggest a stable doctrinal template: dominance is rarely mere pairwise preference. It is usually a universally quantified transfer principle. The dominated object must be no better under every relevant test family—partial sums, coordinates, utilities, words, or strategies—while the dominating object may vary only within the admissible comparison class fixed by the theory. This is why dominance doctrines often support monotonicity theorems, elimination principles, and classification results.

2. Strategic, decision-theoretic, and informational doctrines

In decision theory, dominance doctrine appears in especially austere form in the Monty Hall problem. Gnedin formulates the game on a finite door set $D$ with $|D|=n\ge 3$, and proves that the class of always-switching strategies is weakly dominant: for every strategy $(x,a_x)$, there exists an always-switching strategy $(x',\mathrm{switch})$ such that

$$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$

for all winning doors $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$0 and all admissible host mechanisms $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$1. The force of the result is pre-probabilistic: it eliminates all strategies that ever choose “match” without using priors on prize location or host behavior beyond admissibility (Gnedin, 2011).

In dynamic games, local dominance modifies the doctrine by weakening forward planning. The comparison of two actions is localized to the scenarios that may realize before the player next moves, and continuation is evaluated only under the requirement of “continuing in the same way” through a mimicking continuation. In static games, local dominance coincides with obvious dominance, but in dynamic games it is neither weaker nor stronger than obvious dominance. This permits implementation results unavailable under obvious strategy-proofness: the ascending-price auction is locally strategy-proof, and the top-trading-cycles rule admits a locally strategy-proof implementation even though it is not implementable in obviously dominant strategies (Catonini et al., 2020).

A different strategic doctrine appears in games with public information providers. One oracle dominates another, written $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$2, if for every signaling strategy $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$3 and every Bayesian game there exists an $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$4-measurable signaling strategy $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$5 such that the full sets of Nash-equilibrium outcome distributions coincide: $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$6 With a unique common-knowledge component, dominance reduces to refinement. With multiple components, refinement within each component remains necessary, but information loops become decisive: balancedness, loop covers, and irreducible loop structure govern whether one oracle can replicate another’s strategic effects (Lagziel et al., 6 Nov 2025).

3. Biological doctrines of dominance

In animal conflict models, dominance doctrine is not an abstract order on objects but a theory of how stable asymmetry emerges from repeated interaction. In the asymmetric Hawk-Dove model of social animals, a dominance relation is defined behaviorally as a persistent pattern in which one animal tends to play Hawk while the other tends to play Dove. The asymmetry variable is

$$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$7

and the expected payoffs from Hawk-Hawk escalation are

$$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$8

Animals begin with incomplete information about $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$9, update beliefs by Bayes’ rule after contests,

$|b|a \succeq |a|b$0

and then update Hawk probabilities according to the sign of $|b|a \succeq |a|b$1. In this framework, dominance is not a fixed attribute of the stronger animal; it is an emergent strategic relation generated by repeated contests, information acquisition, and strategy adjustment under uncertainty (Grewal et al., 2013).

In genetics, the doctrine is revised in a different direction. Dominance is classically the relative difference between a heterozygote and the average of the two corresponding homozygotes at a focal locus, summarized by a dominance coefficient $|b|a \succeq |a|b$2. The paper on multi-locus interaction argues that the observed dominance can vary with genetic background through cis-regulation, trans-acting modifiers, lineage, sex, mating type, or other epistatic structure. Wade’s three-locus model makes the point explicitly: with a focal locus $|b|a \succeq |a|b$3 modified by loci $|b|a \succeq |a|b$4 and $|b|a \succeq |a|b$5, the heterozygote phenotype is

$|b|a \succeq |a|b$6

while the homozygotes are

$|b|a \succeq |a|b$7

This allows incomplete dominance, under-dominance, and over-dominance to emerge from modifier frequencies rather than from an invariant property of the focal alleles. A plausible implication is that any single-locus dominance coefficient estimated in natural populations often averages over unresolved multilocus backgrounds rather than identifying a fixed intrinsic parameter (Li et al., 2023).

4. Algorithmic, optimization, and synthesis doctrines

In constraint programming, dominance doctrine is elevated to a general solution concept. A Constraint Dominance Problem is a tuple

$|b|a \succeq |a|b$8

where $|b|a \succeq |a|b$9 is a CSP and $$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$0 is a preorder on its solutions. Completeness, domination-freeness, and equivalence-freeness are defined directly on subsets of the solution space, and the operational device is the dominance nogood. After finding a solution $$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$1, one excludes future dominated solutions by adding

$$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$2

or the stronger $$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$3 for equivalence-free output. This framework captures single-objective optimization, lexicographic optimization, Pareto optimization, minimal models, minimum correction subsets, CP-net local dominance, and pattern dominance within one declarative scheme (Guns et al., 2018).

In stochastic control, dominance doctrine becomes a distributional risk constraint. For random variables $$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$4 and $$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$5, increasing concave stochastic dominance is

$$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$6

for all increasing concave $$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$7, equivalently

$$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$8

For average-reward MDPs, imposing such dominance on the empirical distribution of a reward signal $$u<v \iff u\le v \text{ and } u\ne v, \qquad u\ll v \iff u_j<v_j \ \forall j.$$9 yields linear constraints on occupation measures: $$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$0 The primal problem is a linear program in $$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$1, and the dual introduces a new pricing term

$$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$2

which enters the Bellman inequality as $$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$3. Dominance is therefore internalized as an endogenous concave revaluation of constrained performance (Haskell et al., 2012).

In reactive synthesis and quantitative graph games, ordinary admissibility fails because a dominated strategy need not be dominated by any admissible strategy. The remedy proposed in the chain-based doctrine is to compare increasing chains of strategies under cofinal dominance: $$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$4 For generalized safety/reachability games, the paper identifies maximal uniform chains—chains realized by parameterized automata with one counter—as the appropriate rationality objects. Every dominated finite-memory strategy is dominated by either an admissible finite-memory strategy or a maximal uniform chain, and key decision problems about whether a parameterized automaton realizes a chain, and whether one uniform chain dominates another, are decidable in polynomial time (Basset et al., 2018).

5. Dynamical-systems, process, and pursuit doctrines

In nonlinear control, dominance doctrine generalizes Lyapunov stability to low-dimensional asymptotic behavior. A linear system $$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$5 is $$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$6-dominant with rate $$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$7 if there exists a symmetric matrix $$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$8 with inertia $$\forall f \in R(A): \exists g \in R(B): \forall w \in \Sigma^\omega,\ A^f(w)\le B^g(w),$$9 such that

$D$0

For nonlinear systems $D$1, the differential version requires

$D$2

Here $D$3 recovers contraction and convergence to a unique equilibrium, $D$4 supports multistability, and $D$5 supports planar asymptotic behavior, including limit cycles. The associated differential dissipativity theory preserves the compositional flavor of classical passivity and small-gain arguments while replacing positive-definite storage by a quadratic form of fixed inertia (Forni et al., 2017).

The robustness extension of this doctrine is given by dominance margins. For a SISO LTI system with transfer function $D$6, the shifted transfer function is

$D$7

Gain, phase, and disk margins are then reinterpreted as certificates of robust $D$8-dominance rather than merely equilibrium stability. In Lure systems satisfying a differential sector condition, Nyquist encirclement and circle-criterion geometry determine whether closed-loop dynamics remain $D$9-dominant or $|D|=n\ge 3$0-dominant, hence whether multistability or oscillatory behavior persists under uncertainty (Padoan et al., 2019).

In Dirichlet-form theory, domination of semigroups is turned into a structural theorem. For two $|D|=n\ge 3$1-semigroups $|D|=n\ge 3$2 and $|D|=n\ge 3$3, domination means

$|D|=n\ge 3$4

The main result is that domination is equivalent to a killing transformation of the associated Markov process. If $|D|=n\ge 3$5 is the relevant multiplicative functional, the dominated form is represented as

$|D|=n\ge 3$6

where $|D|=n\ge 3$7 is the bivariate Revuz measure. This identifies analytic domination with subprocess structure and underlies the characterization of operators sandwiched between Dirichlet and Neumann Laplacians, including local and non-local Robin boundary conditions (Li et al., 2024).

In geometric pursuit-evasion, dominance becomes a reachability partition of the state space. With shortest-path metric $|D|=n\ge 3$8 in a closed region $|D|=n\ge 3$9, the obstacle-aware evader dominance region is

$(x,a_x)$0

while in obstacle-free environments with capture radius $(x,a_x)$1,

$(x,a_x)$2

The paper proves that the initial dominance region is exactly the evader’s open-loop reachable region. It also proves that, without obstacles, the pursuer has a non-anticipative strategy to capture the evader before it leaves the closure of its initial dominance region; with obstacles, such a strategy does not always exist, and only necessary or specialized sufficient conditions are available (Huang et al., 5 Feb 2025).

6. Algebraic, geometric, topological, and combinatorial doctrines

In algebraic geometry, the phrase names a monotonicity principle for positivity of Schur functors. Let $(x,a_x)$3 be a vector bundle and $(x,a_x)$4, $(x,a_x)$5 the Schur functors associated to partitions $(x,a_x)$6 and $(x,a_x)$7. If $(x,a_x)$8 is an algebraic property—additive, multiplicative, and satisfying exponent elimination—and

$(x,a_x)$9

then

$(x',\mathrm{switch})$0

Since ampleness, semiampleness, $(x',\mathrm{switch})$1-ampleness, and nefness are algebraic in this sense, positivity descends along the dominance order. The representation-theoretic engine is a Littlewood–Richardson saturation statement: for sufficiently large $(x',\mathrm{switch})$2, every partition $(x',\mathrm{switch})$3 appears in $(x',\mathrm{switch})$4 (Fatima et al., 2 Mar 2026).

In knot theory, 1-domination is defined by proper degree-one maps between knot exteriors: $(x',\mathrm{switch})$5 This relation is a partial order on knots in $(x',\mathrm{switch})$6. It supports the philosophy that $(x',\mathrm{switch})$7 means $(x',\mathrm{switch})$8 is more complicated, and the paper supplies monotonicity evidence: if $(x',\mathrm{switch})$9, then genus and Gromov volume do not increase, the Alexander module splits, $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$0, and $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$1. Satellites dominate their pattern knots, connected sums dominate their summands, and strict domination chains are bounded in several important classes (Boileau et al., 2015).

In graph theory, dominance doctrine concerns majorization of degree sequences. For degree sequences $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$2 and $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$3, $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$4 means equal total degree and dominance of all partial sums. A forbidden induced-subgraph family $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$5 is dominance monotone when forcible $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$6-freeness is upward-closed in the dominance order. The paper proves strong necessary conditions: in every dominance monotone family, the graph with the smallest number of edges has maximum degree at most $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$7. It then classifies all dominance monotone forbidden sets of cardinality $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$8, $$W(\theta,x',\mathrm{switch}) \ge W\bigl(\theta,x,a_x(d_\theta(x))\bigr)$$9, and $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$00. The irreducible triple examples include the classical threshold-graph and split-graph families $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$01 and $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$02, but also $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$03 and $$a \succeq b \iff \sum_{i=1}^j a_i \ge \sum_{i=1}^j b_i \quad \text{for all } j.$$04, which are not subclasses of split graphs (Barrus et al., 2019).

7. AI strategic dominance doctrine

In AI-forecasting and governance discourse, the dominance doctrine has a specific and much narrower meaning. It is the view that the first actor to develop sufficiently advanced AI, especially ASI, will gain overwhelming power over all others. The paper defines this as the expectation that the first actor to achieve ASI will obtain a decisive strategic advantage: “a position of strategic superiority sufficient to allow [the actor] to achieve unilateral military and economic dominance over the rest of the world,” or even “complete world domination” (Amadori et al., 24 Sep 2025).

The doctrine’s causal structure has three linked steps. First, automated AI R&D, recursive self-improvement, or intelligence recursion creates a compounding lead in AI capability itself. Second, that lead spills into military, cyber, industrial, and surveillance capabilities, including cyberwarfare, autonomous weapons, drone swarms, missile defense, industrial automation, and “fog-of-war machines.” Third, the leader may become able to “cheaply and quickly neutralize any adversaries, with little or no cost to itself,” halt rival AI programs, and stabilize a singleton-like world order. The doctrine is therefore stronger than the claim that AI leadership matters; it is a winner-take-all thesis about strategic monopoly (Amadori et al., 24 Sep 2025).

The same paper is explicit that this doctrine depends on a demanding package of assumptions: near-term ASI, some kind of phase change around automated AI R&D, strong first-mover advantage, concentrated control of advanced AI, rapid deployment, offense-favoring strategic effects, and, most sharply, the feasibility of maintaining control over ASI. This last condition marks the divide from the extinction doctrine. Dominance and extinction share many upstream assumptions about capability and takeoff, but differ on whether humans can steer and retain operational control of superintelligence. The paper also records objections internal to the dominance view itself: diffusion and balancing coalitions may prevent monopoly, deterrence may survive, internal “snap coups” may destabilize the winner, and democratic governance may be incompatible with such concentrated ASI power. In that sense, the AI usage of dominance doctrine is both the narrowest and the most geopolitically consequential use of the term in the current literature (Amadori et al., 24 Sep 2025).