--- title: 'Influence Algebra: Operator Approaches' url: https://www.emergentmind.com/topics/influence-algebra type: topic --- # Influence Algebra: Operator Approaches “Influence Algebra” denotes, across several distinct research programs, an algebraic organization of influence phenomena in terms of operators, vector spaces, bilinear forms, semigroups, polynomial families, and subspace projections. In the Friedkin–Johnsen model, the algebraic elements are the vector space $\mathbb{R}^n$ of opinions, the linear operator $A:=SW$ for influence propagation, the diagonal anchoring operator $B:=I-S$, the resolvent $R_A:=(I-A)^{-1}$, and the influence operator $H:=R_A B$, with steady state $x^*=H x^0$ [2602.08704]. In machine learning, the same label is used for a calculus built from $H_{\hat\theta}^{-1}$, bilinear test-loss influence, dimensionality-reduced approximations on classifier parameters, scalar invariance under relative Hessians, and norm-thresholded predicates for memorization and generalization [2305.16094]. In Boolean analysis, it refers to an operator-theoretic and Fourier-analytic scaffold relating multi-bit derivatives, their energies, and Fourier tails through semigroup evolution and functional inequalities [2404.00084]. This suggests that the term does not identify a single canonical formalism; rather, it names a recurring style of analysis in which influence is represented by algebraic objects that support decomposition, perturbation analysis, and computation. ## 1. Conceptual scope and recurring algebraic motifs The recurring content of “Influence Algebra” is the use of explicit algebraic objects to encode how local inputs produce global effects. In social influence, the basic objects are $A:=SW$, $B:=I-S$, $R_A:=(I-SW)^{-1}$, and $H=(I-SW)^{-1}(I-S)$, together with an interior Green’s operator $G_S:=(I_I-S_IW_{II})^{-1}$ [2602.08704]. In large-scale machine learning, the corresponding objects are the Hessian inverse $H_{\hat\theta}^{-1}$, the linear map $I_{up}(\hat\theta,z_p)=-H_{\hat\theta}^{-1}\nabla_\theta L(z_p,\hat\theta)$, and the bilinear form $I_{loss}(\hat\theta,z_t,z_p)=-\nabla_\theta L(z_t,\hat\theta)^\top H_{\hat\theta}^{-1}\nabla_\theta L(z_p,\hat\theta)$ [2305.16094]. In Boolean function analysis, the basic operators are the commuting derivatives $D_i$ and $D_S$, the heat semigroup $P_t$, and the associated energies $\mathrm{Inf}_S(f)=\|D_Sf\|_2^2$ [2404.00084]. In symbolic influence diagrams, expected utilities correspond to families of polynomials, and chance elimination, decision maximization, and utility aggregation become algebraic transformations such as Schur products, block summations, and selector matrices [1607.08485]. A second recurring motif is that influence is often characterized by membership in a structured linear space. For separable Bayesian networks, separable conditional probability tables are exactly those whose columns lie in the column space $R(B)$ of an event matrix $B$, with separability tested by the fixed-point condition $P(C)=C$ for the orthogonal projection $P(C)=A(A^\prime A)^{-1}A^\prime C$ [1206.6827]. For influence-based network centrality, influence profiles are vectorized into $p(I)\in\mathbb{R}^M$, layered-graph instances form a basis of $\mathbb{R}^M$, and every Bayesian centrality is uniquely determined by its values on that basis [1810.09981]. This suggests that “influence algebra” is less a domain-specific definition than a common research strategy: represent influence by algebraic objects on which one can prove uniqueness, approximation, stability, and computational tractability. ## 2. Friedkin–Johnsen dynamics as a resolvent-enabled linear framework In the Friedkin–Johnsen model, the network is a directed graph $G=(V,E)$ of $n=|V|$ agents with row-stochastic influence matrix $W\in\mathbb{R}^{n\times n}$, susceptibilities $S=\mathrm{diag}(s_1,\dots,s_n)$ with $s_i\in[0,1]$, anchoring operator $B:=I-S$, opinion vector $x_t\in\mathbb{R}^n$, and exogenous opinion vector $x^0\in\mathbb{R}^n$. The core dynamics are $$ x_{t+1}=SWx_t+(I-S)x^0=Ax_t+Bx^0,\qquad A:=SW. $$ If $\rho(SW)<1$, then the global steady state is $$ x^*=(I-SW)^{-1}(I-S)x^0=Hx^0, $$ where $H:=(I-SW)^{-1}(I-S)$ is the influence operator mapping exogenous opinions to the steady state. The Neumann series $$ (I-SW)^{-1}=\sum_{k=0}^{\infty}(SW)^k $$ gives the path expansion $$ x^*=\sum_{k=0}^{\infty}(SW)^k(I-S)x^0, $$ so $(SW)^k$ encodes influence propagation along length-$k$ walks, attenuated by susceptibilities [2602.08704]. The boundary-value formulation partitions the agents into stubborn boundary agents $\mathcal{B}=\{i:s_i=0\}$ and susceptible interior agents $\mathcal{I}=\{i:s_i>0\}$. With $V=\mathcal{I}\cup\mathcal{B}$ and $$ W=\begin{bmatrix}W_{II}&W_{IB}\\W_{BI}&W_{BB}\end{bmatrix},\qquad S=\begin{bmatrix}S_I&0\\0&0\end{bmatrix}, $$ fully stubborn boundary nodes satisfy $x_B^*=x_B^0$, while the interior steady state solves $$ (I_I-S_IW_{II})x_I^*=S_IW_{IB}x_B^*+(I_I-S_I)x_I^0. $$ On the interior, the Green’s operator is $$ G_S:=(I_I-S_IW_{II})^{-1}=\sum_{k\ge 0}(S_IW_{II})^k, $$ and $$ x_I^*=G_S\big[S_IW_{IB}x_B^*+(I_I-S_I)x_I^0\big]. $$ Boundary agents act as Dirichlet boundary conditions; interior agents satisfy a linear system driven by a discounted random walk $S_IW_{II}$. The framework yields precise transient, sensitivity, and robustness results. The error recursion is $$ x_{t+1}-x^*=SW(x_t-x^*)=A(x_t-x^*), $$ hence $$ x_t-x^*=A^t(x_0-x^*),\qquad \|x_t-x^*\|\le \|A^t\|\,\|x_0-x^*\|. $$ For any submultiplicative norm and $\delta>0$, $\|A^t\|\le C_\delta(\rho(A)+\delta)^t$, so $$ \limsup_{t\to\infty}\|x_t-x^*\|^{1/t}\le \rho(SW). $$ The differential of the steady state under $dS$ and $dW$ is $$ dx^*=(I-SW)^{-1}[\,dS\cdot W\cdot x^*+S\cdot dW\cdot x^*-dS\cdot x^0\,], $$ and the resolvent identity $$ (I-(A+\Delta A))^{-1}-(I-A)^{-1} =(I-(A+\Delta A))^{-1}\Delta A(I-A)^{-1} $$ delivers first-order and non-asymptotic perturbation bounds. Under boundary reachability, $\rho(S_IW_{II})<1$ iff every directed cycle in the interior contains a node with $s_i<1$; undamped interior cycles with $s_i=1$ on a cycle preclude invertibility of $I-S_IW_{II}$. The same algebra supports influenceability measures. For the $i$th canonical basis vector $e_i$, define $H_i:=He_i$ and $$ b_i^{(p)}:=\|H_i\|_p,\qquad p\in\{1,2,\infty\}. $$ These column-wise broadcasting centralities measure how much agent $i$’s initial opinion affects the network at steady state. They satisfy $$ \|H_i\|_p\le \|(I-SW)^{-1}\|_p\cdot (1-s_i), $$ while broadcasting centralization is $$ C_b:=\sum_i[\max_k b_k-b_i]. $$ A Gini index over $\{b_i\}$ is an alternative inequality measure. On the Zachary karate club graph with $n=34$, the Monte Carlo illustration uses $W=D^{-1}A$, draws $s$ via zero-inflated Beta, scans sources, and computes steady-state broadcasting. The empirical findings are that broadcasting centralities align strongly with classical degree, closeness, betweenness, eigenvector, and PageRank, with Pearson $r$ roughly $0.90$–$0.98$, Spearman $0.68$–$0.96$, and high top-5 overlap; susceptibility heterogeneity attenuates magnitudes and induces variability in graph-level centralization, especially for path-dependent indices such as log-metric closeness and betweenness. ## 3. Influence functions, relative influence, and memorization in machine learning In the machine-learning setting, the starting point is a training set $Z=\{(x_i,y_i),i\in[1,n]\}$, empirical risk $$ R(\theta)=\frac{1}{n}\sum_{i=1}^n L(z_i,\theta), $$ ERM parameters $\hat\theta=\arg\min_\theta R(\theta)$, and Hessian $$ H_{\hat\theta}=\frac{1}{n}\sum_{i=1}^n \nabla_\theta^2L(z_i,\hat\theta), $$ assumed positive definite and invertible. Under negligible higher-order terms, the parameter influence of reweighting a point $z_p$ is $$ I_{up}(\hat\theta,z_p)=-H_{\hat\theta}^{-1}\nabla_\theta L(z_p,\hat\theta), $$ the removal perturbation is $$ \hat\theta_{-1/n,z_p}-\hat\theta=\frac{1}{n}H_{\hat\theta}^{-1}\nabla_\theta L(z_p,\hat\theta), $$ and the test-loss influence is $$ I_{loss}(\hat\theta,z_t,z_p) =-\nabla_\theta L(z_t,\hat\theta)^\top H_{\hat\theta}^{-1}\nabla_\theta L(z_p,\hat\theta). $$ The induced loss change satisfies $$ L(z_t,\hat\theta_{-1/n,z_p})-L(z_t,\hat\theta) =\frac{1}{n}\nabla_\theta L(z_t,\hat\theta)^\top H_{\hat\theta}^{-1}\nabla_\theta L(z_p,\hat\theta) =-\frac{1}{n}I_{loss}(-1/n,\hat\theta,z_t,z_p) $$ under negligible terms [2305.16094]. The paper formalizes a dimensionality-reduced version called classification influence. With prediction $\tilde y=C(F(x))=C(v)$, $v=F(x)$, and parameter split $\theta=\theta_F\cup\theta_C$, the featurized dataset is $$ \bar Z=\{\zeta_i=(F(\hat\theta_F,x_i),y_i):(x_i,y_i)\in Z,\ i\in[1,n]\}. $$ At the optimum, $$ L((x_i,y_i),\hat\theta)=L((F(x_i),y_i),\hat\theta_C)\Rightarrow L(z_i,\hat\theta)=L(\zeta_i,\hat\theta_C). $$ The classifier parameter influence and classifier test-loss influence are $$ I_{up}^{(C)}(\hat\theta_C,\zeta_p) =-H_{\hat\theta_C}^{-1}\nabla_{\theta_C}L(\zeta_p,\hat\theta_C), $$ $$ I_{loss}^{(C)}(\hat\theta_C,\zeta_t,\zeta_p) =-\nabla_{\theta_C}L(\zeta_t,\hat\theta_C)^\top H_{\hat\theta_C}^{-1}\nabla_{\theta_C}L(\zeta_p,\hat\theta_C). $$ Theorem 1 gives an approximation ratio involving $Q_C,\Lambda_C,\Lambda_F,\alpha,\beta,\gamma,\mathcal{T}_1,\mathcal{T}_2$. If the classifier inverse-Hessian eigenvalues in $\Lambda_C^{-1}$ and eigenvectors $Q_C$ dominate, then $\alpha,\beta,\gamma,\Lambda_F^{-1}$ are negligible, $\mathcal{T}_1$ and $\mathcal{T}_2$ are negligible, and $I_{loss}^{(C)}$ is a close approximation to $I_{loss}$. The paper states that this can reduce the parameter count from hundreds of millions to thousands. A second reduction is relative influence on a small subset $Z^{(m)}=\{(x_{i_j},y_{i_j}),j\in[1,m]\}$, using $$ H_{\hat\theta}^{(m)}=\frac{1}{m}\sum_{j=1}^m\nabla_\theta^2L((x_{i_j},y_{i_j}),\hat\theta), $$ $$ I_{up}^{(m)}(\hat\theta,z_p)=-(H_{\hat\theta}^{(m)})^{-1}\nabla_\theta L(z_p,\hat\theta), $$ $$ I_{loss}^{(m)}(\hat\theta,z_t,z_p) =-\nabla_\theta L(z_t,\hat\theta)^\top (H_{\hat\theta}^{(m)})^{-1}\nabla_\theta L(z_p,\hat\theta). $$ The Loss Estimate Preserving result fixes $z_t$, defines ratios $$ r_p=\frac{I_{loss}(\hat\theta,z_t,z_p)}{I_{loss}^{(m)}(\hat\theta,z_t,z_p)}, $$ and, under two negligibility conditions involving $G(\delta,z_t)$ and $\mathcal{N}(r,\delta,H_{\hat\theta},H_{\hat\theta}^{(m)},z_t)$, proves uniform scaling: $$ \frac{I_{loss}^{(m)}(\hat\theta,z_t,z_p)}{I_{loss}(\hat\theta,z_t,z_p)}=\rho^{-1},\qquad \forall z_p\in Z, $$ omitting negligible terms. The interpretation given is scalar invariance: sign, ordering, and relative comparisons are preserved up to a constant for fixed $z_t$. The framework also defines pointwise memorization and generalization. A point is $D$-memorizable if $$ \|H_{\hat\theta}^{-1}\nabla_\theta L(z_p,\hat\theta)\|_1>D, $$ and $D$-generalizable if $$ \|H_{\hat\theta}^{-1}\nabla_\theta L(z_p,\hat\theta)\|_1\le D, $$ with norms taken in the orthonormal eigenbasis of $H_{\hat\theta}$. The sign-to-memorization theorem introduces $T^+(z_p,z_t)$, $T^-(z_p,z_t)$, $\mathcal{T}^+$, $\mathcal{T}^-$, $\lambda_{\min}$, and $\max_\psi$, and states: $$ I_{loss}(\hat\theta,z_t,z_p)>0 \Rightarrow z_p\ \text{is } D^+\text{-memorizable with } D^+=\frac{\mathcal{T}^+\lambda_{\min}}{\max_\psi}, $$ $$ I_{loss}(\hat\theta,z_t,z_p)<0 \Rightarrow z_p\ \text{is } D^-\text{-memorizable with } D^-=\frac{\mathcal{T}^-\lambda_{\min}}{\max_\psi}. $$ For a single linear layer classifier with BCE, $$ C(u)=\mathrm{sigm}\!\left(\sum_{i=0}^{D-1}w_i u_i+b\right),\qquad \mathrm{sigm}(x)=\frac{1}{1+e^{-x}}, $$ the paper derives sign conditions under $u_i\ge 0$ for all $i$ and an inverse Hessian with almost positive elements. Under these assumptions and the classification approximation, the sign of test-loss influence satisfies $$ \mathrm{sign}(I_{loss}(\hat\theta,z_t,z_p)) =\mathbf{1}(y_t\ne y_p)-\mathbf{1}(y_t=y_p). $$ The corresponding loss-change statements are $$ L(z_t,\hat\theta_{-1/n,z_p})\ge L(z_t,\hat\theta)\ \text{if } y_t=0\text{ and }y_p=0, $$ $$ L(z_t,\hat\theta_{-1/n,z_p})0$ while $$ \max_{|S|=d}\mathrm{Inf}_S(H_n^d)\le C_d\left(\frac{\log n}{n}\right)^d. $$ A further theorem generalizes Oleszkiewicz: if $\mathrm{Inf}_S(f)\le \alpha((\ln n)/n)^{d+1}$ for every $(d+1)$-set $S$ and $\alpha\in(0,C_3)$, then there exists a degree-$d$ Boolean function $g$ such that for every $T$ with $|T|\le d$, $$ |\hat f(T)-\hat g(T)|\le C_4\alpha\left(\frac{\ln n}{n}\right)^{|T|}. $$ A distinct but related “Influence Algebra” concerns total $L_1$ influence of bounded functions on the cube. For $f:\{-1,1\}^n\to\mathbb{R}$, the discrete derivative is $$ \partial_i f(x)=\frac{f(x)-f(x^i)}{2}, $$ the $i$th $L_1$ influence is $\mathrm{Inf}_i(f)=\|\partial_i f\|_1$, and the total $L_1$ influence is $$ \mathrm{Inf}(f)=\sum_{i=1}^n\|\partial_i f\|_1. $$ The main theorem states that if $f$ is $[-1,1]$-valued with Fourier degree $d$, then $$ \mathrm{Inf}(f)=O(d^3\|f\|_\infty), $$ and if $f$ is homogeneous of degree $d$, then $$ \mathrm{Inf}(f)=O(d^2\|f\|_\infty). $$ The central new operator is $$ M_Sf(x):=\sum_{R:\,|R\cap S|=1}\hat f(R)\chi_R(x), $$ together with the proxy quantity $$ \mathcal{I}_p(f):= \mathbb{E}_{S\leftarrow_p[n]} \left[\sum_{i\in S}\mathrm{Inf}_i(M_Sf)\right]. $$ For homogeneous $f$, $$ \mathcal{I}_p(f)\ge p(1-p)^{d-1}\mathrm{Inf}(f), $$ while generally $$ p\,\mathrm{Inf}(f)\le \mathcal{I}_p(T_{(1-p)^{-1}}f). $$ The upper bound comes from $\|M_Sf\|_\infty=O(d\|f\|_\infty)$, proved via convolution kernels $P_S$ built from $d$-admissible measures $H$ with $\|H\|_1=O(d)$. This gives a specifically $L_1$ toolkit where no simple Parseval-type identity is available [1302.4625]. Taken together, these two lines of work show two distinct operator languages for Boolean influence: one centered on $D_S$, $P_t$, and Fourier tails, the other on $\partial_i$, $M_S$, $\mathcal{I}_p$, and noise operators. This suggests that “influence algebra” in Boolean settings names a family of compatible operator frameworks rather than a single definition. ## 5. Polynomial and subspace formulations in graphical decision and probabilistic models In multiplicative influence diagrams, all random variables and decision spaces are finite and discrete, and expected utilities are represented symbolically as families of polynomials. With utility nodes $U_1,\dots,U_m$, criterion weights $k_i\in(0,1)$, and interaction parameter $h$, the global utility is $$ U(\mathbf y_{[n]})= \begin{cases} \displaystyle \sum_{i\in[m]}k_iU_i(\mathbf y_{P_i}), & \text{if } h=0,\\[1ex] \displaystyle \sum_{I\in\mathcal{P}_0([m])} h^{\,n_I-1}\prod_{i\in I}k_iU_i(\mathbf y_{P_i}), & \text{otherwise}, \end{cases} $$ where $h$ is the unique non-zero solution $\ge -1$ to $$ 1+h=\prod_{i\in[m]}(1+hk_i). $$ Backward recursion yields stagewise expected utilities $\bar U_i(\mathbf y_{B_i})$, with different recursions for decision and chance nodes and for nodes that immediately precede a utility node. The symbolic propagation algorithm uses three operations: $$ \bar{\mathbf U}_{i+1}+^{EU}\boldsymbol{\psi}_j \triangleq hk_j(\bar{\mathbf U}_{i+1}^\prime\circ \boldsymbol{\psi}_j^\prime) +k_j\boldsymbol{\psi}_j^\prime+\bar{\mathbf U}_{i+1}^\prime, $$ $$ \bar{\mathbf U}_i\triangleq I_{i,\mathbb V}\times (\bar{\mathbf U}_{i+1}^\prime\circ \mathbf p_i^\prime), $$ $$ \bar{\mathbf U}_i\triangleq I_{i,\mathbb D}\times \bar{\mathbf U}_{i+1}, $$ namely EUMultiSum, EUMarginalization, and EUMaximization. Expected utilities therefore become polynomials in the ring $$ \mathbb{R}[\{p_{iy\pi}\},\{\psi_{j\sigma}\},\{k_j\},h], $$ and standard influence-diagram manipulations become algebraic transformations: arc reversal becomes a rational reparameterization, barren node removal becomes elimination of variables, sufficiency becomes substitution by marginalization and normalization, and asymmetries are modeled by eliminating monomials corresponding to incompatible configurations [1607.08485]. A different algebraic formulation appears in separable Bayesian networks, also called the Influence Model. For finite discrete variables $X,Y,Z$, the event matrix is $$ B_{XY}=[\,I_{m_x}\otimes 1_{m_y}\mid 1_{m_x}\otimes I_{m_y}\,], $$ with rank $$ \mathrm{rank}(B)=m_x+m_y-1. $$ For a conditional probability table $C$ of $P(Z\mid XY)$, the map from a joint PMF $q$ to the output marginal on $Z$ is $$ \Phi_P(q)=q^\prime C. $$ Sufficiency means that $q_1^\prime B=q_2^\prime B$ implies $q_1^\prime C=q_2^\prime C$. Separability means there exist $P_X(Z\mid X)$, $P_Y(Z\mid Y)$, and $\gamma\in[0,1]$ such that $$ P(Z\mid XY)=\gamma P_X(Z\mid X)+(1-\gamma)P_Y(Z\mid Y), $$ or equivalently, at the matrix level, $$ C=\gamma B_{XY,x}C_X+(1-\gamma)B_{XY,y}C_Y. $$ The key results are $$ C\text{ is sufficient } \Longleftrightarrow N(B^\prime)\subseteq N(C^\prime), $$ $$ C\text{ is sufficient } \Longleftrightarrow R(C)\subseteq R(B), $$ and therefore $$ C\text{ is sufficient } \Longleftrightarrow P(Z\mid XY)\text{ is separable}. $$ For multiple parents $X_1,\dots,X_n$, the general separable form is $$ C=\sum_{i=1}^n \gamma_i B_i C_i, $$ with $\sum_i\gamma_i=1$ [1206.6827]. The computational core is a basis-and-projection method. A full-rank basis $A$ for $R(B)$ is constructed recursively by $$ A_1=I_{m_1},\qquad A_i=[\,A_{i-1}\otimes 1_{m_i}\mid 1_{M_{i-1}}\otimes \bar I_{m_i}\,], $$ and the orthogonal projection onto $R(B)$ is $$ P(C)=A(A^\prime A)^{-1}A^\prime C. $$ Theorem 10 states that $C$ is separable iff $P(C)=C$. In the dynamic Influence Model, the next-state PMF at site $i$ is $$ p_i[k+1] =d_{i1}s_1[k]^\prime A_{i1}+\cdots+d_{in}s_n[k]^\prime A_{in}, $$ with $D$ an $n\times n$ stochastic influence matrix. A DBN in which all CPTs are separable is therefore an Influence Model. These two strands use different algebraic media—polynomial rings in one case, linear subspaces and orthogonal projections in the other—but both treat influence by replacing direct enumeration of large joint objects with structured algebraic representations. ## 6. Influence-based network centrality, layered bases, and Bayesian uniqueness In influence-based network centrality, an influence instance is a tuple $$ I=(V,E,P_I), $$ where $P_I$ assigns to each seed set $S_0$ a probability distribution over progressive cascading sequences $$ (S_0,S_1,\dots,S_{n-1}), $$ with monotonicity and $G$-continuity. For a cascade, the cascading distance of node $u$ is $d_u(S_0,S_1,\dots,S_{n-1})=t$ if $u\in\Delta_t$, and $\infty$ otherwise. Vectorizing all nonredundant valid sequences yields a profile vector $p(I)\in\mathbb{R}^M$. Addition and scalar multiplication are componentwise, and mixtures of influence instances correspond to convex combinations of profile vectors. An influence-based centrality is a map $\psi:I\mapsto (\psi_v(I))_{v\in V}$, and the Bayesian axiom requires linearity under convex mixtures: $$ P_I=\alpha P_{I_1}+(1-\alpha)P_{I_2} \Rightarrow \psi_v(I)=\alpha\psi_v(I_1)+(1-\alpha)\psi_v(I_2). $$ Given an anonymous function $f:(\mathbb{R}\cup\{\infty\})^n\to\mathbb{R}$, the induced graph centrality is $$ \mu[f]_v(G)=f(\vec d_G(\{v\})), $$ and the influence-based version is $$ \psi[f]_v(I)= \mathbb{E}_{(S_1,\dots,S_{n-1})\sim P_I(\{v\})} \big[f(\vec d(\{v\},S_1,\dots,S_{n-1}))\big]. $$ Special cases include degree, harmonic, reachability, and radius-$\delta$ sphere-of-influence centralities [1810.09981]. The decisive algebraic result is the layered basis theorem. For disjoint nonempty layers $R_0,\dots,R_t\subseteq V$, the layered graph $L_V(R_0,\dots,R_t)$ has all edges from $R_{i-1}$ to $R_i$, and its BFS influence instance is $I(R_0,\dots,R_t)$. The theorem states that the vectors $$ p(I(R_0,\dots,R_t)) $$ of all nontrivial layered-graph instances form a basis of $\mathbb{R}^M$. Consequently, every profile vector $p(I)$ is a unique linear combination of layered-graph basis vectors, and every Bayesian centrality is uniquely determined by its values on layered graphs. Conformance with the corresponding graph centrality on BFS instances therefore yields a characterization theorem: for anonymous $f$, $\psi[f]$ is the unique influence-based centrality that conforms with $\mu[f]$ and satisfies Anonymity and Bayesian. The same framework extends to groups and cooperative games. The group centrality is $$ \psi[f]_S(I)= \mathbb{E}_{(S_1,\dots,S_{n-1})\sim P_I(S)} \big[f(\vec d(S,S_1,\dots,S_{n-1}))\big], $$ and the Shapley centrality is the Shapley value of the cooperative game with characteristic function $\tau(S)=\psi[f]_S(I)$: $$ \phi_v(\tau) =\sum_{S\subseteq V\setminus\{v\}} \frac{|S|!(n-|S|-1)!}{n!}\bigl(\tau(S\cup\{v\})-\tau(S)\bigr). $$ The paper proves matching characterization theorems for group and Shapley centralities. For additive functions $f(\vec d)=\sum_{u\in V} g(d_u)$ with $g(\infty)=0$, the paper gives an RR-set estimator under the triggering model. The key identity is $$ \psi[f]_u(I)=n\cdot \mathbb{E}[g(d_{R_v}(u,v))], $$ and the estimator based on $\theta$ RR sets is unbiased. With probability at least $1-1/n^\ell$, $$ |\hat\psi_v-\psi_v|\le \epsilon \psi_v \quad\text{if } \psi_v>\psi^{(k)}, $$ and $$ |\hat\psi_v-\psi_v|\le \epsilon \psi^{(k)} \quad\text{otherwise}, $$ provided $\psi^{(k)}\ge 1$. The expected running time is $$ O\!\left(\frac{\ell (m+n)\log n\cdot \mathbb{E}[\sigma(\tilde v)]}{\psi^{(k)}\epsilon^2}\right). $$ This network-centrality line and the Friedkin–Johnsen line both define influence-based centralities, but they do so through different algebraic objects. One uses columns of the resolvent-based operator $H=(I-SW)^{-1}(I-S)$ to define broadcasting centralities $b_i^{(p)}=\|H_i\|_p$ and centralization $C_b$ [2602.08704]; the other uses linear functionals on the vector space of cascade profiles [1810.09981]. This suggests that centrality is not a single influence-algebraic primitive. Rather, different models induce different algebras of influence, and the associated centrality notions inherit their structure from those algebras. ## 7. Limitations, assumptions, and interpretive boundaries Across the literature, the algebraic representation is always tied to explicit assumptions. In Friedkin–Johnsen dynamics, the main resolvent formulas require $W$ row-stochastic and $\rho(SW)<1$ globally or $\rho(S_IW_{II})<1$ on the interior block, together with boundary reachability; edge cases with $s_i=1$ can create undamped cycles and violate invertibility, and reducible $W$ requires treatment by strongly connected components [2602.08704]. In machine learning, the Hessian and all relative Hessians must be positive definite, Taylor truncations require negligible higher-order terms, classifier-space approximation requires dominance of $\Lambda_C^{-1}$ and $Q_C$, LEP requires the constructed matrices $G(\delta,z_t)$ and $\mathcal{N}(\cdot)$ to satisfy negligibility, and the label–sign correlation requires a single-layer classifier with BCE, nonnegative features, and an inverse Hessian with almost positive elements [2305.16094]. In Boolean analysis, the KKL-type and FKN-type results are proved under the uniform product measure on $\{-1,1\}^n$ and Walsh normalization, while the constants depend on $d$ and extensions to biased product measures, Gaussian space, or multi-valued domains remain open [2404.00084]. The $L_1$-influence theory applies to bounded functions on the cube, and the cubic bound $\mathrm{Inf}(f)=O(d^3\|f\|_\infty)$ is the tightest bound proved in that paper via its methods, although an optimal $O(d)$ bound is conjectured there [1302.4625]. In multiplicative influence diagrams, variables are finite and discrete, utility parent sets are disjoint, and multiplicative interactions can cause polynomial blow-up, especially when $h\ne 0$ [1607.08485]. In separable Bayesian networks, the linear-subspace characterization covers finite discrete variables with row-stochastic CPTs; separability reduces complexity, but the representation is not unique [1206.6827]. In influence-based centrality, the algebra and characterizations assume progressive cascades and full distributions over cascading sequences, while the algorithmic guarantees are established under the triggering model and additive $f$ [1810.09981]. A common misconception would be to treat “Influence Algebra” as a universally standardized theory. The papers instead present several domain-specific algebras: resolvent calculus for social influence, Hessian-based bilinear forms for influence functions, derivative-semigroup identities for Boolean analysis, polynomial propagation for influence diagrams, subspace geometry for separable Bayesian networks, and basis decompositions for influence-based centrality. This suggests that the phrase is best read as a methodological designation for algebraic representations of influence, rather than as the name of a single formally unified discipline. Source: https://www.emergentmind.com/topics/influence-algebra