CausalSmith · seminar slides

Separating Latent-Source Arrows with Cumulants

With a fixed number of middle source slots and an axis normalization, higher-order population cumulants generically separate the two bivariate latent linear non-Gaussian arrow conventions.

Overview

  • We study a bivariate latent linear non-Gaussian model with mm middle source slots.
  • The maintained input is an axis-normalized representation class for XYX\to Y and YXY\to X.
  • The observable object is the truncated joint cumulant vector through order LL.
  • The question is whether the same cumulants can arise from the opposite arrow convention.
  • The main answer: generically, order 2m+22m+2 recovers the loading support and excludes the opposite fiber.
  • For m3m\ge3, generic real opposite-fiber exclusion already holds at order 2m+12m+1.

Motivation

  • In Gaussian linear models, covariance information is symmetric in ways that often leave direction unresolved.
  • Non-Gaussianity supplies higher-order structure that can break that symmetry.
  • LiNGAM uses this asymmetry under linear non-Gaussian disturbances; latent sources make the direction problem harder.
  • We ask what the population cumulants alone establish once the number of source slots and axis convention are fixed.
  • This is a representation-level identification question for a structured econometric model class.

Running Example

  • Think of XX and YY as two economic outcomes linked by one direct channel and several latent shocks.
  • A forward representation assigns one direct XYX\to Y loading, one fixed vertical-axis source, and mm middle source slots.
  • A reverse representation assigns the symmetric YXY\to X convention.
  • The econometric question is whether the observed higher-order cumulants support both recursive specifications.
  • The source count mm is treated as a maintained modeling input.
Axis source fixed to X Middle sources m slots latent shocks Axis source fixed to Y X observed outcome Y observed outcome
illustrative Box-and-arrow schematic with observed boxes X and Y, a forward direct arrow X to Y, middle latent source boxes pointing to both X and Y, and fixed axis source boxes pointing to one observed variable each.

Related Literature

  • Pearl (2009), Spirtes et al. (2001), Peters et al. (2017), and Richardson and Spirtes (2002) give the structural and graphical background.
  • Shimizu et al. (2006) connect linear non-Gaussian structure to causal discovery through LiNGAM.
  • Hoyer et al. (2008) and Salehkaleybar et al. (2020) study latent-variable LiNGAM settings.
  • Brillinger (1969), McCullagh (1987), Comon and Mourrain (1996), and Landsberg (2012) supply the cumulant and binary-form tools.
  • Chen et al. (2025) use higher-order cumulants for direction under latent confounding.
  • We characterize the generic opposite-fiber geometry for the fixed-source, axis-normalized representation class.

Setup

  • There are m+2m+2 latent source slots S0,,Sm+1S_0,\ldots,S_{m+1}.
  • The mm middle slots have finite slopes and can load on both observed coordinates.
  • The two remaining slots are fixed by the axis normalization.
  • Sources are centered, mutually independent, non-Gaussian, and have finite moments through the retained order.
  • Independence makes cumulants add across source slots.
  • Non-Gaussianity makes higher-order cumulants informative beyond covariance.
Assumption A-1 (Independent latent sources)

The latent source variables S0,,Sm+1S_0,\ldots,S_{m+1} are mutually independent with respect to μ\mu: S0Sm+1. S_0\perp\cdots\perp S_{m+1}.

Axis Conventions

  • In the forward convention, the direct loading has slope γ\gamma, the middle slopes are ρ1,,ρm\rho_1,\ldots,\rho_m, and the vertical-axis source is fixed.
  • In the reverse convention, the direct loading has slope δ\delta, the middle slopes are σ1,,σm\sigma_1,\ldots,\sigma_m, and the horizontal-axis source is fixed.
  • Genericity means distinct finite slopes, nonzero direct slope, and nonzero retained source cumulants.
  • In the running example, this rules out coincident shock directions and a zero direct channel.
Assumption A-4

With forward loading vectors u0=(1,γ)u_0=(1,\gamma), uj=(1,ρj)u_j=(1,\rho_j) for 1jm1\le j\le m, and um+1=(0,1)u_{m+1}=(0,1), the bivariate observed outcome vector admits the forward latent linear representation (X,Y)=j=0m+1ujSj. (X,Y)^\top=\sum_{j=0}^{m+1} u_j S_j .

Assumption A-5

With reverse loading vectors v0=(1,0)v_0=(1,0), vj=(σj,1)v_j=(\sigma_j,1) for 1jm1\le j\le m, and vm+1=(δ,1)v_{m+1}=(\delta,1), the bivariate observed outcome vector admits the reverse latent linear representation (X,Y)=j=0m+1vjSj. (X,Y)^\top=\sum_{j=0}^{m+1} v_j S_j .

Cumulant Data

  • TL(P)T_L(P), the truncated cumulant vector, collects joint cumulants of XX and YY from orders 22 through LL.
  • Each source contributes its order-rr cumulant weight times a monomial in its loading vector.
  • The forward and reverse cumulant maps encode these contributions as polynomial maps.
  • A fiber is the set of parameters that produce the same truncated cumulant vector.
  • Direction separation means the opposite-arrow fiber is empty.
Latent sources cumulant weights Loadings loading vectors Forward map polynomial cumulants Reverse map polynomial cumulants T_L(P) orders 2 through L Forward fiber same T_L(P) Reverse fiber same T_L(P) Direction separation reverse fiber empty
illustrative Box-and-arrow schematic with boxes for latent sources and loadings, polynomial cumulant map, truncated cumulant vector, forward fiber, and reverse fiber.

Key Idea

  • The cumulant blocks behave like a simultaneous decomposition of binary forms.
  • At the apolar order 2m+22m+2, the equations recover the unordered set of loading directions.
  • The recovered support contains the fixed axis in a way determined by the arrow convention.
  • For generic parameters, the opposite convention would require the incompatible fixed-axis pattern.
  • This turns higher-order cumulants into a population separation device for the two representation fibers.

Main Result I

informal · Theorem T-1 At order 2m+22m+2, generic parameters have their unordered loading-slope support recovered and their full opposite-arrow cumulant fiber empty.

  • The result applies on relatively Zariski-open dense loci.
  • These loci contain real feasible neighborhoods, so the algebraic statement has real model points.
  • In the running example, the higher-order cumulants recover the shock-loading directions needed to rule out the competing recursive axis convention.
Theorem T-1

Let m1m\geq 1, put n=m+2n=m+2 and K=2m+2=2n2K=2m+2=2n-2. Use the support-annihilator and differential-operator conventions of Definition. There exist sets UmrightΘm,Kright,,UmleftΘm,Kleft, U_m^{\mathrm{right}}\subseteq \Theta^{\mathrm{right},\circ}_{m,K}, \qquad U_m^{\mathrm{left}}\subseteq \Theta^{\mathrm{left},\circ}_{m,K} such that:

  • (Generic loci.) Each of UmrightU_m^{\mathrm{right}} and UmleftU_m^{\mathrm{left}} is relatively Zariski open and relatively Zariski dense in the corresponding complex parameter space Θm,Kright\Theta^{\mathrm{right}}_{m,K} or Θm,Kleft\Theta^{\mathrm{left}}_{m,K} of Definition P-3, and is contained in the corresponding generic parameter locus of Definition P-7.
  • (Open neighborhoods meeting feasibility.) There are Euclidean-open sets Oright,OleftO^{\mathrm{right}},O^{\mathrm{left}} of real parameter space such that OrightFm,Kright,OleftFm,Kleft, O^{\mathrm{right}}\cap F^{\mathrm{right}}_{m,K}\neq\varnothing, \qquad O^{\mathrm{left}}\cap F^{\mathrm{left}}_{m,K}\neq\varnothing, where the feasible regions are those of Definition P-11; use the coordinatewise-complexification convention of Definition. Then (OrightFm,Kright)CUmright,(OleftFm,Kleft)CUmleft. \left(O^{\mathrm{right}}\cap F^{\mathrm{right}}_{m,K}\right)_{\mathbb C} \subseteq U_m^{\mathrm{right}}, \qquad \left(O^{\mathrm{left}}\cap F^{\mathrm{left}}_{m,K}\right)_{\mathbb C} \subseteq U_m^{\mathrm{left}}.
  • (Forward recovery and separation.) For every θUmright\theta\in U_m^{\mathrm{right}}, Rm,Kleft ⁣(Φm,Kright(θ))=. R^{\mathrm{left}}_{m,K}\!\left(\Phi^{\mathrm{right}}_{m,K}(\theta)\right)=\varnothing. Moreover, there is a nonzero squarefree polynomial QC[z]Q\in\mathbb C[z] such that Q(0)0Q(0)\neq0 and the multiset of roots of QQ is {γ,ρ1,,ρm}\{\gamma,\rho_1,\ldots,\rho_m\}. Every generic forward parameter θ\theta'' satisfying Φm,Kright(θ)=Φm,Kright(θ) \Phi^{\mathrm{right}}_{m,K}(\theta'')= \Phi^{\mathrm{right}}_{m,K}(\theta) has the same multiset {γ,ρ1,,ρm}\{\gamma,\rho_1,\ldots,\rho_m\}. If fr(x,y)=a=0r(ra)tr,axraya,t=Φm,Kright(θ), f_r(x,y)=\sum_{a=0}^{r}\binom{r}{a}t_{r,a}x^{r-a}y^a, \qquad t=\Phi^{\mathrm{right}}_{m,K}(\theta), let DD be the forward loading-direction support. Then QD0,xQD,yQD, Q_D\neq0,\qquad x\mid Q_D,\qquad y\nmid Q_D, and, for every homogeneous binary form qq of degree nn, [q()fn+k=0 for every 0km]cC: q=cQD. \bigl[q(\partial)f_{n+k}=0\ \text{for every }0\leq k\leq m\bigr] \quad\Longleftrightarrow\quad \exists c\in\mathbb C:\ q=cQ_D.
  • (Reverse recovery and separation.) For every ηUmleft\eta\in U_m^{\mathrm{left}}, Rm,Kright ⁣(Φm,Kleft(η))=. R^{\mathrm{right}}_{m,K}\!\left(\Phi^{\mathrm{left}}_{m,K}(\eta)\right)=\varnothing. Moreover, there is a nonzero squarefree polynomial QC[z]Q\in\mathbb C[z] such that Q(0)0Q(0)\neq0 and the multiset of roots of QQ is {δ,σ1,,σm}\{\delta,\sigma_1,\ldots,\sigma_m\}. Every generic reverse parameter η\eta'' satisfying Φm,Kleft(η)=Φm,Kleft(η) \Phi^{\mathrm{left}}_{m,K}(\eta'')= \Phi^{\mathrm{left}}_{m,K}(\eta) has the same multiset {δ,σ1,,σm}\{\delta,\sigma_1,\ldots,\sigma_m\}. If fr(x,y)=a=0r(ra)tr,axraya,t=Φm,Kleft(η), f_r(x,y)=\sum_{a=0}^{r}\binom{r}{a}t_{r,a}x^{r-a}y^a, \qquad t=\Phi^{\mathrm{left}}_{m,K}(\eta), let DD be the reverse loading-direction support. Then QD0,yQD,xQD, Q_D\neq0,\qquad y\mid Q_D,\qquad x\nmid Q_D, and, for every homogeneous binary form qq of degree nn, [q()fn+k=0 for every 0km]cC: q=cQD. \bigl[q(\partial)f_{n+k}=0\ \text{for every }0\leq k\leq m\bigr] \quad\Longleftrightarrow\quad \exists c\in\mathbb C:\ q=cQ_D.

Main Result II

informal · Theorem T-2 At order 2m+22m+2, generic opposite-arrow fibers are empty, while same-arrow fibers retain the stated direct-slot versus middle-slot ambiguity.

  • The arrow is separated at the representation level.
  • The unordered loading-slope multiset is pinned down.
  • Same-arrow parametrization still permits ambiguity beyond admissible middle-source relabelling.
  • For m2m\ge2, the same-arrow fiber has exact relative Zariski dimension m(m1)2\frac{m(m-1)}{2}.
Theorem T-2 (Generic Fiber Obstruction)

For every mNm\in\mathbb N with m1m\ge 1, set K=2m+2K=2m+2. There exist complex loci UmrightΘm,Kright,U^{\mathrm{right}}_m\subseteq \Theta^{\mathrm{right},\circ}_{m,K} and UmleftΘm,Kleft,U^{\mathrm{left}}_m\subseteq \Theta^{\mathrm{left},\circ}_{m,K}, with the generic loci as in Definition P-7, such that:

  • (Relative genericity.) Each of UmrightU^{\mathrm{right}}_m and UmleftU^{\mathrm{left}}_m is relatively Zariski open and relatively Zariski dense in the parameter spaces Θm,Kright\Theta^{\mathrm{right}}_{m,K} and Θm,Kleft\Theta^{\mathrm{left}}_{m,K} of Definition P-4 and Definition P-5, respectively.
  • (Real incidence.) There is a Euclidean-open set of real forward parameters whose intersection with Fm,KrightF^{\mathrm{right}}_{m,K} from Definition P-11 is nonempty, and every parameter in this intersection has coordinatewise complexification contained in UmrightU^{\mathrm{right}}_m. Likewise, there is a Euclidean-open set of real reverse parameters whose intersection with Fm,KleftF^{\mathrm{left}}_{m,K} is nonempty, and every parameter in this intersection has coordinatewise complexification contained in UmleftU^{\mathrm{left}}_m.
  • (Forward generic fibers.) For every θ=(γ,ρ,c)Umright\theta=(\gamma,\rho,c)\in U^{\mathrm{right}}_m, every forward parameter in the forward fiber Rm,Kright ⁣(Φm,Kright(θ)) R^{\mathrm{right}}_{m,K}\!\left(\Phi^{\mathrm{right}}_{m,K}(\theta)\right) of Definition P-8 and Definition P-4 has the same unordered loading-slope multiset {γ,ρ1,,ρm} \{\gamma,\rho_1,\ldots,\rho_m\} as θ\theta. The opposite-arrow fiber Rm,Kleft ⁣(Φm,Kright(θ)) R^{\mathrm{left}}_{m,K}\!\left(\Phi^{\mathrm{right}}_{m,K}(\theta)\right) of Definition P-8 (formed with the reverse map of Definition P-5) is empty. For every latent index i=1,,mi=1,\ldots,m, there is a generic forward parameter θΘm,Kright,\theta'\in\Theta^{\mathrm{right},\circ}_{m,K} obtained by swapping the full pairs (γ,(c0r)r=2K)(\gamma,(c_{0r})_{r=2}^K) and (ρi,(cir)r=2K)(\rho_i,(c_{ir})_{r=2}^K), leaving every other coordinate fixed, such that Φm,Kright(θ)=Φm,Kright(θ), \Phi^{\mathrm{right}}_{m,K}(\theta') = \Phi^{\mathrm{right}}_{m,K}(\theta), but θ\theta' is not in the admissible GmG_m-orbit of θ\theta from Definition P-9. If m2m\ge 2, this forward fiber has exact relative Zariski dimension m(m1)2. \frac{m(m-1)}{2}.
  • (Reverse generic fibers.) For every η=(δ,σ,d)Umleft\eta=(\delta,\sigma,d)\in U^{\mathrm{left}}_m, every reverse parameter in the reverse fiber Rm,Kleft ⁣(Φm,Kleft(η)) R^{\mathrm{left}}_{m,K}\!\left(\Phi^{\mathrm{left}}_{m,K}(\eta)\right) has the same unordered loading-slope multiset {δ,σ1,,σm} \{\delta,\sigma_1,\ldots,\sigma_m\} as η\eta. The opposite-arrow fiber Rm,Kright ⁣(Φm,Kleft(η)) R^{\mathrm{right}}_{m,K}\!\left(\Phi^{\mathrm{left}}_{m,K}(\eta)\right) is empty. For every latent index i=1,,mi=1,\ldots,m, there is a generic reverse parameter ηΘm,Kleft,\eta'\in\Theta^{\mathrm{left},\circ}_{m,K} obtained by swapping the full pairs (δ,(dm+1,r)r=2K)(\delta,(d_{m+1,r})_{r=2}^K) and (σi,(dir)r=2K)(\sigma_i,(d_{ir})_{r=2}^K), leaving every other coordinate fixed, such that Φm,Kleft(η)=Φm,Kleft(η), \Phi^{\mathrm{left}}_{m,K}(\eta') = \Phi^{\mathrm{left}}_{m,K}(\eta), but η\eta' is not in the admissible GmG_m-orbit of η\eta from Definition P-9. If m2m\ge 2, this reverse fiber has exact relative Zariski dimension m(m1)2. \frac{m(m-1)}{2}.

Exceptional Locus

informal · Theorem T-3 At order 2m+22m+2, the closure of cumulant vectors compatible with both arrow conventions has codimension exactly 11 in each arrow image variety.

  • The compatibility set is a property of the truncated cumulant vector.
  • Its preimages are exactly the generic same-arrow parameters whose cumulants retain a full opposite-arrow representation.
  • For m=1m=1 and m=2m=2, the theorem records explicit incidence systems.
  • For m2m\ge2, the compatibility geometry has codimension one rather than codimension mm.
Theorem T-3

For every m1m\geq 1, at K=2m+2K=2m+2, let Em\overline E_m, HmrightH^{\mathrm{right}}_m, and HmleftH^{\mathrm{left}}_m be as in Definition P-10, and let Cm,KrightC^{\mathrm{right}}_{m,K} and Cm,KleftC^{\mathrm{left}}_{m,K} be as in Definition P-6. Then:

  • (Codimension.) The exceptional closure Em\overline E_m has codimension exactly 11 in each of Cm,KrightC^{\mathrm{right}}_{m,K} and Cm,KleftC^{\mathrm{left}}_{m,K}. Moreover, if m2m\geq 2, then Em\overline E_m does not have codimension mm in either Cm,KrightC^{\mathrm{right}}_{m,K} or Cm,KleftC^{\mathrm{left}}_{m,K}.
  • (Forward preimage.) For every θΘm,Kright\theta\in\Theta^{\mathrm{right}}_{m,K}, θHmrightθΘm,Kright, and Rm,Kleft ⁣(Φm,Kright(θ)). \theta\in H^{\mathrm{right}}_m \quad\Longleftrightarrow\quad \theta\in\Theta^{\mathrm{right},\circ}_{m,K} \ \text{and}\ R^{\mathrm{left}}_{m,K}\!\left(\Phi^{\mathrm{right}}_{m,K}(\theta)\right)\neq\varnothing .
  • (Reverse preimage.) For every ηΘm,Kleft\eta\in\Theta^{\mathrm{left}}_{m,K}, ηHmleftηΘm,Kleft, and Rm,Kright ⁣(Φm,Kleft(η)). \eta\in H^{\mathrm{left}}_m \quad\Longleftrightarrow\quad \eta\in\Theta^{\mathrm{left},\circ}_{m,K} \ \text{and}\ R^{\mathrm{right}}_{m,K}\!\left(\Phi^{\mathrm{left}}_{m,K}(\eta)\right)\neq\varnothing .
  • (Low-dimensional specializations.) If m=1m=1, then E1={t: p satisfying the system A1, Φ1,4right(pright)=t}. E_1 = \left\{ t:\ \exists p\text{ satisfying the system }A_1,\ \Phi^{\mathrm{right}}_{1,4}(p^{\mathrm{right}})=t \right\}. If m=2m=2, then E2={t: p satisfying the system A2, Φ2,6right(pright)=t}, E_2 = \left\{ t:\ \exists p\text{ satisfying the system }A_2,\ \Phi^{\mathrm{right}}_{2,6}(p^{\mathrm{right}})=t \right\}, where A1,A2A_1,A_2 are the compatibility instances of Definition P-16.

Information Order

informal · Theorem T-4 For m3m\ge3, order 2m+12m+1 is sufficient for generic real exclusion of the opposite arrow, so K(m)2m+1K^\star(m)\le 2m+1.

  • The apolar theorem uses order 2m+22m+2 to recover loading support.
  • The real separation target is weaker: it asks only whether the opposite real feasible fiber is empty.
  • For m3m\ge3, generic real separation reaches that target one order earlier.
  • Thus 2m+12m+1 is enough for generic real arrow separation, while 2m+22m+2 supports loading-support recovery.
Theorem T-4 (Improved Information Order)

The theorem-local symbols ErightE^{\mathrm{right}} and EleftE^{\mathrm{left}} below are real parameter-space exceptional subsets and are unrelated to the cumulant-space compatibility locus EmE_m. For every mNm\in\mathbb N with 3m3\le m, at truncation order L=2m+1L=2m+1:

  • (Forward generic separation.) There is a proper real algebraic subset of the real forward parameter space whose complement meets Fm,LrightF^{\mathrm{right}}_{m,L}; choose one such subset and denote it locally in this theorem by ErightE^{\mathrm{right}}. For every θFm,LrightEright\theta\in F^{\mathrm{right}}_{m,L}\setminus E^{\mathrm{right}} there is no ηFm,Lleft\eta\in F^{\mathrm{left}}_{m,L} such that Φm,Lright(θ)=Φm,Lleft(η). \Phi^{\mathrm{right}}_{m,L}(\theta)=\Phi^{\mathrm{left}}_{m,L}(\eta).
  • (Reverse generic separation.) There is a proper real algebraic subset of the real reverse parameter space whose complement meets Fm,LleftF^{\mathrm{left}}_{m,L}; choose one such subset and denote it locally in this theorem by EleftE^{\mathrm{left}}. For every ηFm,LleftEleft\eta\in F^{\mathrm{left}}_{m,L}\setminus E^{\mathrm{left}} there is no θFm,Lright\theta\in F^{\mathrm{right}}_{m,L} such that Φm,Lleft(η)=Φm,Lright(θ). \Phi^{\mathrm{left}}_{m,L}(\eta)=\Phi^{\mathrm{right}}_{m,L}(\theta).

Consequently, for the information order K(m)K^\star(m) of Definition P-12, K(m)2m+1andK(m)2m+2. K^\star(m)\le 2m+1 \qquad\text{and}\qquad K^\star(m)\ne 2m+2 .

Proof Sketch

  • First, independence converts source contributions into additive cumulant blocks.
  • Second, the order-2m+22m+2 cumulants are organized as binary forms with m+2m+2 loading directions.
  • Third, apolar equations identify the unique support-annihilator on the generic loci.
  • Fourth, the recovered support reveals which fixed axis belongs to the representation.
  • Finally, the opposite-arrow convention would impose the other fixed-axis pattern, giving an empty opposite fiber generically.

Compatibility Geometry

  • The exceptional locus is studied by comparing simultaneous forward and reverse decompositions.
  • The comparison keeps the axis slots fixed and quotients only the middle-source label symmetry.
  • The resulting dimension calculation gives a single algebraic compatibility condition in each arrow image.
  • For small mm, we record explicit incidence systems that describe the same compatibility event.
  • This explains why generic separation coexists with a codimension-one family admitting both conventions.

Limitations and Future Work

informal · Remark T-5 A future direction is a finite computable semialgebraic atlas that exactly labels real forward and reverse feasible fibers for every fixed mm.

  • The current results are population statements about truncated cumulants.
  • The maintained inputs are the source count mm and the axis-normalized representation class.
  • Empirical use requires additional work on cumulant estimation, weak higher-order signal, and near-exceptional parameters.
  • The finite-sample and numerical decision problems are natural next steps.

Takeaways

  • Higher-order cumulants carry directional information in fixed-source latent linear non-Gaussian representations.
  • At order 2m+22m+2, the generic apolar equations recover unordered loading support and exclude the opposite arrow convention.
  • The remaining same-arrow ambiguity is characterized separately from arrow separation.
  • The opposite-arrow compatibility closure has codimension exactly one in each arrow image variety.
  • For m3m\ge3, generic real opposite-fiber exclusion already holds at order 2m+12m+1.