要旨
1. はじめに
The mathematical study of self-organizing systems has a rich history spanning statistical mechanics, evolutionary game theory, and complex adaptive systems theory. In each of these domains, a common pattern emerges: self-organization requires a parameter that controls the boundary between ordered and disordered phases. In statistical mechanics, this parameter is temperature. In evolutionary game theory, it is mutation rate. In complex adaptive systems, it is the balance between exploration and exploitation. This paper identifies governance density as the analogous parameter for multi-agent decision systems — the universal control variable that determines whether a system of autonomous agents self-organizes into productive specialization, stagnates in over-constrained rigidity, or dissolves into chaotic divergence.
The significance of this identification extends beyond theoretical elegance. If governance density is indeed the phase transition controller, then designing stable agentic companies reduces to a concrete optimization problem: tune D so that exact loop gain remains below 1 and, ideally, buffered specialization is preserved. This provides a measurable and actionable design principle for enterprise AI governance — a domain that currently lacks mathematical foundations and relies on ad hoc policy design.
Moreover, the governance density framework scales. An enterprise is a collection of agents governed by corporate policies. A market is a collection of enterprises governed by regulations. A civilization is a collection of markets governed by laws. At each scale, governance density plays the same structural role: constraining influence propagation to prevent chaos while permitting sufficient freedom for productive self-organization. The effective governance density D<sub>eff</sub> composes multiplicatively across scales, providing a unified framework for analyzing stability from the agent level to the civilization level.
2. The Mathematical Model
2.1 Formal Definition
We define the agentic company at time t as a dynamic graph-augmented constrained Markov decision process G<sub>t</sub> = (A<sub>t</sub>, E<sub>t</sub>, S<sub>t</sub>, Π<sub>t</sub>, R<sub>t</sub>, D<sub>t</sub>) where: A<sub>t</sub> = {a<sub>1</sub>, ..., a<sub>n</sub>} is the set of n autonomous agents, E<sub>t</sub> ∈ R<sup>n×n</sup> is the edge weight matrix encoding inter-agent dependencies (communication bandwidth, resource sharing, decision influence), S<sub>t</sub> = [F<sub>t</sub>, K<sub>t</sub>, H<sub>t</sub>, L<sub>t</sub>, C<sub>t</sub>] is the organizational state vector (financial, KPI, human capacity, risk, communication structure), Π<sub>t</sub> = {π<sub>1</sub>, ..., π<sub>n</sub>} is the set of agent policies mapping states to action distributions, R<sub>t</sub>: S × A<sup>n</sup> → R is the organizational reward function, and D<sub>t</sub> ∈ (0, 1) is the governance density.
2.2 State Dynamics
The state transition is governed by: S<sub>t+1</sub> = f(S<sub>t</sub>, a<sub>1,t</sub>, ..., a<sub>n,t</sub>, E<sub>t</sub>, D<sub>t</sub>) + ξ<sub>t</sub> where a<sub>i,t</sub> ~ π<sub>i</sub>(S<sub>t</sub>) is agent i's action sampled from its policy, and ξ<sub>t</sub> is exogenous noise (market shocks, regulatory changes, external events). The function f encodes how joint agent actions transform organizational state, mediated by the dependency structure E<sub>t</sub> and constrained by governance density D<sub>t</sub>. Critically, f is nonlinear and non-separable — agent actions interact through the dependency network, and the effect of agent i's action depends on what all other agents are simultaneously doing.
2.3 Influence Matrix
The agent influence matrix A<sub>t</sub> ∈ R<sup>n×n</sup> is defined as the Jacobian of the effective state transition with respect to individual agent actions: [A<sub>t</sub>]<sub>ij</sub> = ∂f<sub>j</sub> / ∂a<sub>i</sub> evaluated at the current operating point. Entry a<sub>ij</sub> measures the sensitivity of agent j's effective state to agent i's actions. This matrix is generally non-symmetric (influence is directional), time-varying (organizational dynamics shift influence patterns), and dense (in tightly coupled organizations, most agents influence most others to some degree).
3. Governance Density Theory
3.1 Formal Definition
Governance density D<sub>t</sub> = |C<sub>t</sub>| / |A<sub>t</sub>| is defined as the cardinality of the active constraint set divided by the cardinality of the available action set, both measured at time t. The constraint set C<sub>t</sub> includes: (1) approval gates — decisions requiring review before execution, (2) evidence requirements — mandatory documentation and justification, (3) risk thresholds — escalation triggers for high-risk actions, (4) responsibility boundaries — authority limits and scope constraints, and (5) compliance rules — regulatory and policy requirements. The action set A<sub>t</sub> includes all actions that any agent could potentially take, including those that are constrained.
3.2 境界条件
D = 0 represents the absence of governance — no constraints exist, and agents can take any action without review, documentation, or approval. This is the anarchic limit. D = 1 represents total governance — every possible action is constrained, and no autonomous execution is possible. This is the paralysis limit. Neither extreme is viable for a functioning organization. The practical range is D ∈ (0.1, 0.9), and the theoretical optimum depends on the spectral properties of the influence matrix and the organizational objectives.
3.3 Governance as Damping Operator
ガバナンスの密度は、影響の伝播に対する減衰作用素として機能します。有効影響行列 A<sub>eff</sub> = (1 − D) · A を考えてみましょう。各ガバナンス制約は、制約されたアクションとその下流の効果の間の影響経路を中断または減衰するため、ガバナンス密度によってエージェント間の有効影響が (1 − D) 倍減少します。有効行列のスペクトル半径は、 λ<sub>max</sub>(A<sub>eff</sub>) = (1 − D) · λ<sub>max</sub>(A) です。安定性のためには、 λ<sub>max</sub>(A<sub>eff</sub>) < 1 が必要です。これにより、(1 − D) · λ<sub>max</sub>(A) < 1、または小さい D の場合、同等の λ<sub>max</sub>(A) < 1 / (1 − D) ≈ 1 + D が得られます。実際の安定条件 λ<sub>max</sub>(A) < 1 − D は、より保守的です。より大きな値を提供する境界安定マージン。
4. Spectral Analysis and the Stability Law
4.1 The Fundamental Stability Condition
Theorem 1 (Exact Local Contraction and Buffered Envelope). In the scalar-damped approximation A<sub>eff,t</sub> = (1 − D<sub>t</sub>)A<sub>t</sub>, the exact local contraction condition is (1 − D<sub>t</sub>)λ<sub>max</sub>(A<sub>t</sub>) < 1 for all t sufficiently large. The stricter buffered operating envelope is λ<sub>max</sub>(A<sub>t</sub>) < 1 − D<sub>t</sub>.
Proof sketch. Define the Lyapunov function V(t) = E[||S<sub>t</sub> − S*||<sup>2</sup>]. Under scalar damping, the effective Jacobian is A<sub>eff,t</sub> = (1 − D<sub>t</sub>)A<sub>t</sub>, so the state dynamics give V(t+1) ≤ ρ<sup>2</sup> · V(t) + σ<sup>2</sup><sub>ξ</sub> where ρ = (1 − D<sub>t</sub>)λ<sub>max</sub>(A<sub>t</sub>) is the contraction factor and σ<sup>2</sup><sub>ξ</sub> is the noise variance. When ρ < 1, the Lyapunov function is a supermartingale with geometric decay: V(t) ≤ ρ<sup>2t</sup> · V(0) + σ<sup>2</sup><sub>ξ</sub> / (1 − ρ<sup>2</sup>). The first term vanishes exponentially, and the second term provides the noise floor of the equilibrium. The buffered envelope λ<sub>max</sub>(A) < 1 − D is stricter than required for contraction, but carves out the subset of the contractive region with positive reserve against perturbations and model error.
4.2 Spectral Radius Computation
Computing λ<sub>max</sub>(A<sub>t</sub>) from observed agent interactions requires estimating the influence matrix. We use the empirical Jacobian approach: perturb each agent's actions slightly and observe the response in other agents' behavior. Specifically, [A<sub>t</sub>]<sub>ij</sub> ≈ Δresponse<sub>j</sub> / Δaction<sub>i</sub> averaged over a window of W decision cycles. The spectral radius is then computed using power iteration, which converges in O(log(n) / log(λ<sub>max</sub> / λ<sub>2</sub>)) iterations where λ<sub>2</sub> is the second-largest eigenvalue.
5. Phase Diagram Derivation
5.1 Four Regimes
The organizational dynamics exhibit four distinct regimes governed by the parameters (C<sub>task</sub>, B<sub>comm</sub>, D):
Stagnation Regime (D > D<sub>crit,high</sub>). When governance density exceeds the upper critical value D<sub>crit,high</sub> ≈ 0.7 + 0.1 · B<sub>comm</sub>, the constraint cost in the utility function dominates, and agents converge to a minimal set of safe, low-impact roles. Role entropy H(r) approaches zero. Decision throughput drops to 10-20% of unconstrained capacity. The organization is stable but unproductive — it has maximized self-observation at the expense of self-action.
Buffered Specialization Regime. When the buffered envelope λ<sub>max</sub>(A) < 1 − D holds and governance remains below the stagnation ceiling, agents self-organize into specialized roles with positive operating reserve. Role entropy stabilizes at a moderate level, throughput remains high, and perturbations are absorbed without large cascades.
Fragile Specialization Regime. When exact contraction still holds but the buffered envelope fails — i.e., (1 − D)λ<sub>max</sub>(A) < 1 but λ<sub>max</sub>(A) ≥ 1 − D — the system continues to converge, but without meaningful reserve. Specialization still forms, yet convergence slows, anomaly sensitivity rises, and modest shocks can push the organization into cascade behavior.
カスケード レジーム。 正確な収縮が違反される場合、つまり (1 − D)λ<sub>max</sub>(A) ≥ 1 の場合、影響の伝播によって摂動が増幅されます。役割エントロピーは最大に近づくか、不規則に変動し、異常率は急上昇し、連鎖エラーがネットワークを通じて伝播するにつれて意思決定の品質が急速に低下します。
5.2 位相境界方程式
The buffered phase boundary is defined by λ<sub>max</sub>(A) = 1 − D, which we can express parametrically as D<sub>buffer</sub>(C, B) = 1 − λ<sub>max</sub>(A(C, B)) where the spectral radius depends on task complexity (higher C increases inter-agent coupling) and communication bandwidth (higher B enables more effective coordination, reducing effective coupling). The exact contraction boundary is defined by (1 − D)λ<sub>max</sub>(A) = 1, giving D<sub>exact</sub>(C, B) = 1 − 1 / λ<sub>max</sub>(A(C, B)) when λ<sub>max</sub>(A) > 1. The upper phase boundary is defined by the stagnation condition: D<sub>upper</sub>(B) = 1 − ε<sub>min</sub>(B) where ε<sub>min</sub>(B) is the minimum effective autonomy required for productive role specialization, which increases with communication bandwidth because coordination overhead requires more freedom.
6. 最適化としての役割の特化
6.1 Utility Function Decomposition
The agent utility function U<sub>i</sub>(r | C, B, D) decomposes into three terms: (1) Efficiency term: α · Eff<sub>i</sub>(r) = α · exp(−||c<sub>i</sub> − c<sub>r</sub>||<sup>2</sup> / 2σ<sup>2</sup>) measures the match between agent i's capability vector c<sub>i</sub> and role r's requirement vector c<sub>r</sub>. (2) Impact term: β · Impact(r) = β · d<sub>out</sub>(r) / max<sub>r'</sub> d<sub>out</sub>(r') measures the organizational influence of role r, normalized by the maximum influence across all roles, where d<sub>out</sub>(r) is the out-degree of role r in the organizational decision graph. (3) Constraint cost: γ · Cost(r, D) = γ · D · Impact(r) measures the governance burden imposed on role r, which increases with both governance density and role impact (high-impact roles bear more constraints).
6.2 Nash Equilibrium of Role Assignment
The role assignment game has a Nash equilibrium when no agent can improve its utility by unilaterally switching roles. The equilibrium distribution p*(r) satisfies: for all agents i and all roles r' ≠ r<sub>i</sub>, U<sub>i</sub>(r<sub>i</sub>) ≥ U<sub>i</sub>(r'). We show that this equilibrium exists and is unique in the buffered specialization regime, and remains locally contractive in the fragile specialization regime, by verifying that the best-response dynamics form a contraction mapping on the space of role distributions whenever exact loop gain remains below 1. The equilibrium concentrates probability mass on roles that balance efficiency (capability match) against constraint cost (governance burden), producing the characteristic pattern of moderate role specialization.
7. Civilization Extension
7.1 Two-Tier Governance
The civilization extension introduces a second governance layer above the enterprise. While D<sub>company</sub> captures corporate governance (gates, policies, role constraints), D<sub>civ</sub> captures civic governance (laws, regulations, constitutional constraints). The effective governance density compounds multiplicatively: D<sub>eff</sub> = 1 − (1 − D<sub>company</sub>)(1 − D<sub>civ</sub>). The multiplicative composition means that each governance layer provides independent constraint coverage. If D<sub>company</sub> = 0.4 and D<sub>civ</sub> = 0.3, then D<sub>eff</sub> = 1 − (0.6)(0.7) = 0.58. Neither layer alone provides sufficient governance, but together they provide adequate coverage.
7.2 多層的な影響
文明規模では、影響力マトリックスは層に分解されます。A<sup>(1)</sup> は企業レベルの影響力 (企業内のエージェント間)、A<sup>(2)</sup> は市場レベルの影響力 (市場相互作用を通じた企業間)、A<sup>(3)</sup> は政治レベルの影響力 (政策変更を通じた規制から企業) です。多層の正確な安定条件には、(1 − D<sub>eff</sub>) max<sub>k</sub> λ<sub>max</sub>(A<sup>(k)</sup>) < 1 が必要です。より厳密なバッファリングされたターゲットは、max<sub>k</sub> λ<sub>max</sub>(A<sup>(k)</sup>) < 1 − D<sub>eff</sub> です。最も弱い層、つまり最大のスペクトル半径を持つ層が、システムの安定性を決定します。これは、文明が企業レベルでは安定していても、市場の影響力の伝播が効果的なガバナンスの限界を超えた場合、市場レベルでは脆弱または不安定になる可能性があることを意味します。
8. Multi-Layer Stability Analysis
8.1 クロスレイヤーカップリング
3 つの影響層は独立しているわけではなく、層間の結合を通じて相互作用します。企業の決定は市場動向に影響を与えます (企業の行動は株価、サプライヤーとの関係、顧客の行動に影響を与えます)。市場の力学は政治的決定に影響を与えます（経済危機は規制対応の引き金となります）。政治的決定は企業運営に影響を与えます (新しい規制は制約の状況を変えます)。この層間結合は、層 k の状態を層 l のダイナミクスにリンクする結合行列 C<sup>(k,l)</sup> によって取得されます。
8.2 Composite Stability Condition
With cross-layer coupling, the exact stability condition becomes: (1 − D<sub>eff</sub>)ρ(A<sub>composite</sub>) < 1 where A<sub>composite</sub> is the block matrix incorporating all within-layer and cross-layer influence terms. The stricter buffered operating target is ρ(A<sub>composite</sub>) < 1 − D<sub>eff</sub>. This is a stronger requirement than the uncoupled per-layer condition because cross-layer coupling can create amplification pathways that do not exist within any single layer. In practice, the composite spectral radius is typically 10-30% higher than the maximum per-layer spectral radius, requiring correspondingly higher D<sub>eff</sub>.
9. Market Revaluation Model
9.1 Price Dynamics
Asset prices in the civilization model follow: P<sub>t+1</sub> = P<sub>t</sub> + κ(V<sub>t</sub> − P<sub>t</sub>) + ζ<sub>t</sub> where P<sub>t</sub> is the current market price, V<sub>t</sub> is the estimated intrinsic value, κ ∈ (0, 1) is the adjustment speed (how quickly prices converge to value), and ζ<sub>t</sub> is the revaluation shock. In the absence of shocks (ζ = 0), prices converge exponentially to intrinsic value with rate κ. Shocks perturb prices away from value, and the interplay between adjustment and shock determines price stability.
9.2 Periodic Revaluation Shocks
When revaluation occurs periodically with period T, the shock process takes the form: ζ<sub>t</sub> = σ · sin(2πt / T) + η<sub>t</sub> where σ is the revaluation amplitude and η<sub>t</sub> is random noise. Shorter periods (smaller T) produce more frequent shocks, increasing the effective volatility: Var[P<sub>t</sub>] ≈ σ<sup>2</sup> / (2κ) + σ<sup>2</sup><sub>η</sub> / (2κ). This has a critical implication for governance design: shorter revaluation cycles require higher D<sub>civ</sub> to compensate for the increased instability. Specifically, the governance density must satisfy: D<sub>civ</sub> ≥ D<sub>civ,min</sub>(σ, T, κ) = 1 − exp(−σ / (κ · T)).
10. Land Development Model
10.1 地価の動態
土地開発モデルは、文明の物理的インフラストラクチャーの次元を捉えます: L<sub>t+1</sub> = L<sub>t</sub> + α · Dev<sub>t</sub> − β · Risk<sub>t</sub> ここで、L<sub>t</sub> は地価、Dev<sub>t</sub> は開発投資、Risk<sub>t</sub> は地域のリスク レベル、α、β は感応度パラメーターです。開発投資は次のようになります: Dev<sub>t</sub> = min(Budget<sub>t</sub>, c<sub>0</sub> + c<sub>1</sub> · LandSize + c<sub>2</sub> · InfraGap) ここで、InfraGap は必要なインフラストラクチャと利用可能なインフラストラクチャの差を測定します。開発コストは土地の大きさとインフラ不足に応じて増加し、開発率には自然な上限が生じます。
10.2 Land-Market Coupling
地価と市場価格は、開発と投資のサイクルを通じて結びついています。地価の上昇はより多くの投資を呼び込み、より多くの開発資金が供給され、地価が上昇します。この正のフィードバック ループは、バブル (価格が本質的価値を超えた場合) や破綻 (突然の再評価ショックによって過大評価が明らかになった場合) を引き起こす可能性があります。ガバナンス密度 D<sub>civ</sub> は、特に急速な再評価の期間中に、このフィードバック ループを弱めるのに十分でなければなりません。結合係数 μ = ∂P / ∂L · ∂L / ∂P はフィードバック ループの強度を測定し、μ > 1 の場合、安定性には D<sub>civ</sub> > 1 − 1/μ が必要です。
11. Convergence Proofs
11.1 Contraction Mapping Argument
定理 2 (収束)。 厳密な収縮条件 (1 − D<sub>t</sub>)λ<sub>max</sub>(A<sub>t</sub>) < 1 の下では、状態ダイナミクス S<sub>t</sub> は、L<sup>2</sup> 収束の意味で一意の平衡 S に収束します: lim<sub>t→∞</sub> E[||S<sub>t</sub> − S||<sup>2</sup>] = σ<sup>2</sup><sub>ξ</sub> / (1 − ρ<sup>2</sup>) ここで、ρ = (1 − D)λ<sub>max</sub>(A) < 1 は収縮率です。
Proof. Define the operator T: S → S by T(S) = f(S, π(S), E, D) where π(S) is the equilibrium policy profile. Under scalar damping, the local Lipschitz factor is bounded by ρ = (1 − D)λ<sub>max</sub>(A). When ρ < 1, ||T(S) − T(S')|| ≤ ρ · ||S − S'|| for all S, S' in the state space. By the Banach fixed-point theorem, T has a unique fixed point S and the iterates S<sub>t+1</sub> = T(S<sub>t</sub>) + ξ<sub>t</sub> converge to a neighborhood of S with radius proportional to the noise magnitude.
11.2 Civilization-Scale Convergence
The civilization-scale convergence proof extends Theorem 2 to the multi-layer setting. The composite contraction factor is ρ<sub>civ</sub> = (1 − D<sub>eff</sub>) max<sub>k</sub> λ<sub>max</sub>(A<sup>(k)</sup>) in the uncoupled approximation, or (1 − D<sub>eff</sub>)ρ(A<sub>composite</sub>) when coupling is explicit, and convergence follows from the same Banach argument applied to the composite state space. The key additional requirement is that cross-layer coupling does not increase the effective contraction factor beyond 1, which is ensured when D<sub>eff</sub> accounts for cross-layer amplification.
11.3 Rate of Convergence
The convergence rate is governed by the exact contraction margin δ<sub>exact</sub> = 1 − (1 − D)λ<sub>max</sub>(A). The expected time to reach equilibrium (within tolerance ε of S) scales as: t<sub>conv</sub> = O(log(||S<sub>0</sub> − S|| / ε) / log(1/ρ)) = O(log(||S<sub>0</sub> − S*|| / ε) / δ<sub>exact</sub>) for small δ<sub>exact</sub>. The buffered reserve δ<sub>buffer</sub> = 1 − D − λ<sub>max</sub>(A) is stricter: it measures not whether the system contracts, but whether it contracts with usable reserve. Organizations operating with positive exact margin but negative buffer converge slowly and are vulnerable to perturbations, while organizations with both margins positive converge quickly and resist disturbances robustly.
12. 結論
This paper establishes governance density as a universal stability parameter for self-organizing multi-agent systems across scales. The exact local stability law (1 − D)λ<sub>max</sub>(A) < 1 determines whether the scalar-damped system contracts, while the buffered operating envelope λ<sub>max</sub>(A) < 1 − D identifies the portion of that region with usable reserve. The phase diagram therefore contains four organizational regimes — stagnation, buffered specialization, fragile specialization, and cascade — as functions of governance density, task complexity, and communication bandwidth. The civilization extension demonstrates that the same mathematical framework applies at enterprise, market, and political scales through the effective governance density composition D<sub>eff</sub> = 1 − (1 − D<sub>company</sub>)(1 − D<sub>civ</sub>).
The key insight is profound in its simplicity: governance is not a cost. It is the parameter that controls phase transitions. Too little governance allows cascade behavior — influence propagation goes unbounded and the system diverges. Too much governance causes stagnation — agents lose the freedom needed for productive specialization. The optimal governance density positions the organization in the buffered specialization regime where self-organization produces meaningful role differentiation, decisions are made efficiently, and the system maintains the metacognitive self-awareness needed to correct its own errors.
For MARIA OS, this theory provides concrete design principles. The Gate Engine implements D. The Doctor system monitors λ<sub>max</sub>(A), exact loop gain, and buffered reserve. The Evidence Layer provides the observation infrastructure. Together, they keep exact loop gain below 1 and, ideally, preserve positive operating buffer, keeping the organization in the buffered regime where autonomous AI operations produce value under governance guarantees. The mathematics are clear: the governance density law is the fundamental equation of agentic enterprise design.
参考文献
1.ニューマン、M.E.J. （2010年）。 ネットワーク: 概要。オックスフォード大学出版局。
2. ストロガッツ、S.H. （2015年）。 非線形ダイナミクスとカオス。ウエストビュープレス。
3. Hofbauer, J. & Sigmund, K. (1998)。 進化ゲームと人口動態。ケンブリッジ大学出版局。
4. Sutton, R.S. & Barto, A.G. (2018). Reinforcement learning: An introduction. MIT Press.
5. Acemoglu, D. & Robinson, J.A. (2012). Why nations fail. Crown Business.
6. Jackson, M.O. (2008). Social and economic networks. Princeton University Press.
7. O.J.ブランチャード& フィッシャー、S. (1989)。 マクロ経済学の講義。 MITプレス。
8. MARIA OS Documentation. (2026). Governance Density Framework. os.maria-code.ai/docs.