Spectral Kan-Do Flow Matching (SKFM)

• SKFM is a causal discovery method that leverages intervention-induced continuous-time vector fields and Lie bracket geometry to reveal latent confounders.
• It employs an amortized flow matching loss and spectral factorization via eigendecomposition of the curvature Gram matrix to estimate latent dimensionality.
• The method integrates calibration, acyclicity penalties, and latent contamination scoring to extract visible DAGs, with strong performance on structured synthetic and real-data scenarios.

Searching arXiv for the cited SKFM-related papers and closely related flow-matching work. Spectral Kan-Do Flow Matching (SKFM) is a causal discovery method for settings with latent confounding that recasts the gap between observation and intervention as a problem of learning, analyzing, and spectrally factorizing intervention-induced geometry. In the formulation introduced in “Latent Confounded Causal Discovery via Lie Bracket Geometry,” SKFM learns intervention-conditioned continuous-time vector fields, measures their non-closure under Lie brackets, uses a curvature Gram matrix to extract a low-rank latent confounding footprint, and then recovers a visible directed acyclic graph (DAG) under a soft Lie-algebraic acyclicity penalty (Mahadevan, 17 Jun 2026).

1. Conceptual setting in Kan-Do-Calculus

SKFM is developed inside the broader framework of Kan-Do-Calculus (KDC), where the distinction between passive observation and active intervention is expressed as the bi-adjunction

$Lan_K \dashv K^* \dashv Ran_K.$

Within this formulation, interventions are modeled by left Kan extensions $Lan_K$, conditioning by right Kan extensions $Ran_K$, and restriction to observational contexts by precomposition $K^*=(-\circ K)$. In the smooth statistical setting of the paper, this categorical distinction is pushed into differential geometry: differences between observational and interventional laws become Radon–Nikodym density ratios, those ratios induce local causal vector fields, and latent confounding appears as a failure of visible intervention fields to close under Lie brackets (Mahadevan, 17 Jun 2026).

The target problem is visible causal discovery when hidden common causes may be present. The paper places special emphasis on the mismatch between conditioning and intervention, written in Pearl-style notation as

$P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$

and, when a valid back-door adjustment set $Z$ exists,

$$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$

SKFM is positioned as the direct, end-to-end realization of this observation: rather than using learned geometry only as a pruning or screening stage, it attempts to estimate the intervention geometry itself, infer the latent subspace behind failures of visible closure, and then extract a visible DAG from that learned geometry (Mahadevan, 17 Jun 2026).

In the same paper, BRIDGE and SKFM are paired as complementary algorithms. BRIDGE performs high-recall screening using bracket residuals and passes a reduced arrow family to downstream score-based or differentiable discovery routines, whereas SKFM is the more ambitious construction that learns amortized intervention fields and factors latent curvature spectrally. This places SKFM at the “direct Lie-space endpoint” of the paper’s broader program (Mahadevan, 17 Jun 2026).

2. Amortized intervention-field learning

SKFM starts from an observational law $P$ and intervention-specific laws $P^{(i)}$. Their mismatch is encoded through the Radon–Nikodym ratio

$$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$

The associated calibration identity is

$Lan_K$0

From these quantities, the paper defines local causal vector fields. For candidate intervention $Lan_K$1,

$Lan_K$2

depending on whether one uses a static density-ratio model or a continuous flow (Mahadevan, 17 Jun 2026).

SKFM adopts the second route. Its learned object is an amortized intervention-conditioned flow

$Lan_K$3

where $Lan_K$4 is an intervention embedding. The training objective is a conditional flow matching loss,

$Lan_K$5

with $Lan_K$6 the conditional vector field transporting a base measure $Lan_K$7 to the intervention-specific target law $Lan_K$8. In this sense, SKFM is a genuine flow-matching method: it learns intervention-specific velocity fields directly rather than scores or adjacency weights (Mahadevan, 17 Jun 2026).

The paper combines this with additional causal objectives. Its total objective is

$Lan_K$9

Here $Ran_K$0 is a KDC-style calibration term, $Ran_K$1 is a Lie-algebraic acyclicity term, and $Ran_K$2 is introduced as an optimal-transport regularizer under weak overlap. The paper also states a Fisher-energy regularizer,

$Ran_K$3

although the implementation emphasis is on the total objective above (Mahadevan, 17 Jun 2026).

3. Lie brackets and spectral latent-curvature factorization

The central geometric observable in SKFM is the Lie bracket of two learned intervention fields,

$Ran_K$4

The bracket is interpreted as infinitesimal order sensitivity: if two intervention flows commute, their local geometry is compatible; if they do not, visible dynamics retain an order effect that may indicate hidden structure, confounding, regime mismatch, or missing coordinates. The paper ties this directly to Frobenius integrability: if the visible intervention fields span an involutive distribution, visible geometry is coherent; if not, the non-closing component is a witness of failed visible integrability (Mahadevan, 17 Jun 2026).

SKFM turns this into a spectral latent-variable construction by aggregating bracket energy into a curvature Gram matrix

$Ran_K$5

The paper also writes the pointwise bracket matrix as

$Ran_K$6

with

$Ran_K$7

By construction, $Ran_K$8 is real, symmetric, and positive semidefinite (Mahadevan, 17 Jun 2026).

SKFM then eigendecomposes this matrix,

$Ran_K$9

estimates a latent rank $K^*=(-\circ K)$0 through eigenvalue thresholding $K^*=(-\circ K)$1, and collects the top eigenvectors into

$K^*=(-\circ K)$2

The associated latent coordinates are

$K^*=(-\circ K)$3

with projector

$K^*=(-\circ K)$4

This is the sense in which the method is “spectral”: the spectral object is not a Fourier basis for signals, but an eigendecomposition of a Lie-bracket curvature matrix whose dominant eigenvectors encode a low-rank latent confounding footprint (Mahadevan, 17 Jun 2026).

The paper’s main identifiability statement for this step is Proposition 2, “Latent Dimensionality Consistency.” Under additive latent confounders, independence of latent and visible noise, linearly independent visible loading directions, finite second moments of bracket features, and an eigengap after the $K^*=(-\circ K)$5-th eigenvalue, the population curvature matrix factors as

$K^*=(-\circ K)$6

so that

$K^*=(-\circ K)$7

Under those assumptions, spectral thresholding consistently recovers the dimension of the visible latent footprint; if full-rank or cancellation assumptions fail, the recovered rank should be interpreted more conservatively as the rank of that footprint rather than the literal number of hidden variables (Mahadevan, 17 Jun 2026).

4. Lie-algebraic acyclicity and graph extraction

After learning intervention fields and latent curvature directions, SKFM imposes an acyclicity condition through a solvable-Lie representation of a DAG. The paper expands each bracket as

$K^*=(-\circ K)$8

where $K^*=(-\circ K)$9 are visible structure constants and $P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$0 is the latent residual. The corresponding soft acyclicity penalty is

$P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$1

This is explicitly order-dependent: in a fixed ordered basis, triangularity of the structure constants is used as the acyclicity criterion. Proposition 3 states that if $P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$2 in a fixed ordered basis, then the extracted adjacency is acyclic; conversely every DAG admits such a triangular representation in some topological order. The paper is explicit that this is not an order-free acyclicity theorem (Mahadevan, 17 Jun 2026).

The algorithmic extraction rule is correspondingly local and ordered. An edge $P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$3 is declared when

$P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$4

The influence score is also written more explicitly as

$P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$5

or, in another passage, with the Euclidean norm in place of the Frobenius norm. Because raw Jacobian influence conflates direct effects, indirect ancestry, and confounded dependence, the paper introduces a second calibration layer based on latent contamination (Mahadevan, 17 Jun 2026).

Two diagnostic quantities are central. The node contamination score is

$P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$6

and the latent bracket fraction is

$P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$7

Edges are then rescored as

$P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$8

The paper further adds root-pair masking and score-aware transitive pruning. Operationally, this means that SKFM does not rely on Jacobian magnitude alone: it also asks whether the target node is more contaminated than the source and whether the relevant bracket energy mostly lies in the latent subspace (Mahadevan, 17 Jun 2026).

The optimization loop in Algorithm 2 states that parameters are updated by natural gradient descent,

$P(Y\mid X)\neq P(Y\mid \mathrm{do}(X)),$9

with $Z$0 the Fisher metric. The manuscript, however, does not provide a concrete approximation scheme for $Z$1, minibatch Fisher estimation, or a full training pseudocode for the Wasserstein term. That omission is one of the method’s implementation-level ambiguities (Mahadevan, 17 Jun 2026).

5. Reported empirical behavior

The empirical record presented for SKFM is mixed in a way that is central to understanding the method. On synthetic settings with strong structural priors, the learned intervention fields and latent projector are highly accurate; on more general motifs and real data, field learning remains strong but direct graph extraction is less stable. The paper therefore supports SKFM most strongly as a geometry-learning and latent-structure factorization method, and more conditionally as a universal end-to-end extractor (Mahadevan, 17 Jun 2026).

Setting	Reported SKFM outcome
Endpoint-confounded latent chains, $Z$2, $Z$3	SHD $Z$4, $Z$5 across all four dimensions
Diamond, collider/fork, multi-branch six-node motifs	Field corr. $Z$6, $Z$7, $Z$8; default $Z$9 $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$0, $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$1, $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$2
Calibrated diamond and collider/fork extraction	Precision $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$3, recall $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$4 for both motifs
Ten-node nonlinear random DAGs	“Same ten-node seeds defeated direct SKFM extraction”; BRIDGE hybrid mean directed $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$5
Sachs pilot	Influence corr. $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$6, relative MSE $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$7, latent rank $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$8, directed $$P(Y\mid \mathrm{do}(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z).$$9
S9 pilot	Influence corr. $P$0, relative MSE $P$1, latent rank $P$2, compact 10-edge graph
PISA pilot	Influence corr. $P$3, relative MSE $P$4, latent rank $P$5, 4-edge graph

For the ordered latent-chain experiments, the paper reports exact visible chain recovery across $P$6, with field correlations $P$7, $P$8, $P$9, and $P^{(i)}$0; relative influence MSEs $P^{(i)}$1, $P^{(i)}$2, $P^{(i)}$3, and $P^{(i)}$4; latent dimensions $P^{(i)}$5; and SHD $P^{(i)}$6 with $P^{(i)}$7 in all cases. The manuscript immediately qualifies these as ordered-chain settings with a highly local extractor, not order-free general discovery (Mahadevan, 17 Jun 2026).

The non-chain motif experiments sharpen the distinction between field learning and graph extraction. On diamond, collider/fork, and multi-branch six-node nonlinear motifs, field quality remains high, with field correlations $P^{(i)}$8, $P^{(i)}$9, and $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$0, yet default direct extraction yields only $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$1, $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$2, and $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$3. After focused Step-9 calibration with continuity regularization, a rank-two spectral contamination subspace, contamination-gradient boosting, latent-bracket penalties, root-pair masking, and score-aware transitive pruning, the diamond and collider/fork motifs are both recovered exactly, with precision $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$4 and recall $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$5 (Mahadevan, 17 Jun 2026).

On larger random graphs, the paper treats BRIDGE as the practical fallback. It states that the same ten-node nonlinear DAG seeds defeated direct SKFM extraction, while BRIDGE plus local BIC over the geometry-pruned mask reached mean directed $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$6. On real-data pilots, SKFM’s learned fields again track kernel geometry well—on Sachs, influence correlation $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$7 and relative MSE $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$8; on S9, $$\rho_i(\omega)=\frac{dP^{(i)}}{dP}(\omega),\qquad \ell_i(\omega)=\log \rho_i(\omega).$$9 and $Lan_K$00; on PISA, $Lan_K$01 and $Lan_K$02—but the extracted visible graphs are much less decisive. This pattern is one of the method’s defining empirical conclusions (Mahadevan, 17 Jun 2026).

6. Position within spectral flow-matching research

SKFM belongs to a broader family of methods that inject structure into flow matching, but its notion of “spectral” is unusually specific. In image-generation work such as “SpectralDiT,” spectral structure is introduced as a timestep-conditioned low/high-frequency correction inside a Diffusion Transformer block, while “Low-Pass Flow Matching” changes the probability path itself through an operator-modulated interpolant $Lan_K$03 and analyzes the induced power spectral density along the path (Tian, 17 Jun 2026, Ruscio et al., 1 Jun 2026). SKFM does neither: its spectral object is the eigensystem of a Lie-bracket curvature Gram matrix, and its flow-matching target is an intervention-conditioned causal vector field rather than an image residual or a signal-domain transport map (Mahadevan, 17 Jun 2026).

Other recent work has placed conditional flow matching in explicit spectral latent spaces. “Conditional Flow Matching for Continuous Anomaly Detection in Autonomous Driving on a Manifold-Aware Spectral Space” uses whitened PCA coefficients as a low-rank spectral manifold and performs exact likelihood evaluation through a conditional continuous normalizing flow, while “Functional MRI Time Series Generation via Wavelet-Based Image Transform and Spectral Flow Matching for Brain Disorder Identification” performs flow matching in a DCT domain built on wavelet decomposition maps (Guillen-Perez, 19 Feb 2026, Tew et al., 28 May 2026). By contrast, SKFM does not transport PCA coefficients, DCT coefficients, or Fourier modes of observed data. Its spectral coordinates arise only after learning visible intervention fields and measuring their pairwise non-closure (Mahadevan, 17 Jun 2026).

The “Kan-Do” in SKFM also has a specific meaning. It refers to Kan-Do-Calculus, not to the Kolmogorov–Arnold Network literature. This is materially different from KAN-based flow-matching work in robotics, where KAN layers are used as lightweight function approximators inside visuomotor policies (Chen et al., 1 Feb 2026). A separate spectral flow line, “Flow Along the $Lan_K$04-Amplitude for Generative Modeling,” uses a scaling parameter $Lan_K$05 as a flow variable over grouped spectral coefficients, but there the spectral structure is attached to signal decomposition rather than Lie-space curvature (Du et al., 27 Apr 2025). SKFM therefore occupies a distinct niche: it is a causal-discovery method whose spectral step is a latent-curvature factorization rather than a signal-space spectral parameterization (Mahadevan, 17 Jun 2026).

The principal limitations follow directly from that design. SKFM assumes smooth statistical manifolds with $Lan_K$06 vector fields, sufficient support overlap to define Radon–Nikodym ratios, and, for spectral latent-rank consistency, additive latent structure, full-rank visible loading directions, finite bracket moments, and an eigengap. Its acyclicity guarantee is relative to a fixed or correctly learned order. Empirically, the method’s field-learning and latent-factorization components are consistently stronger than its universal graph extractor. This places SKFM, in the current evidence, as a geometric and spectral compression framework for latent-confounded discovery, with end-to-end graph extraction working best in ordered chains and calibrated small motifs rather than as a general replacement for downstream combinatorial search (Mahadevan, 17 Jun 2026).