NMF-FFB: Non-negative matrix factorization with feedforward-feedback structure

Kenichi Satoh

arxiv: 2512.18250 · v2 · pith:ALGLLTAAnew · submitted 2025-12-20 · 📊 stat.ME

NMF-FFB: Non-negative matrix factorization with feedforward-feedback structure

Kenichi Satoh This is my paper

Pith reviewed 2026-05-21 16:15 UTC · model grok-4.3

classification 📊 stat.ME

keywords non-negative matrix factorizationstructural equation modelingLeontief inversefeedback effectslatent factorsexploratory analysismultiplicative updates

0 comments

The pith

NMF-FFB embeds simultaneous equations into non-negative matrix factorization to separate direct from cumulative feedback effects using a latent Leontief inverse.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper develops NMF-FFB as a way to incorporate feedback among endogenous variables into the standard non-negative matrix factorization setup. Standard NMF is feedforward only, depending solely on exogenous inputs, but here the coefficients include a term for latent feedback from the endogenous data itself. When the product of the factor matrix and feedback matrix has spectral radius below one, the model reduces to a form that uses the Leontief inverse to accumulate amplified effects over feedback loops. A reader might care because this offers an exploratory tool for non-negative data sets in small samples, avoiding the need for covariance-based fitting common in traditional structural models and allowing automatic grouping into latent factors. The method includes regularization for sparsity and orthogonality plus bootstrap procedures to assess uncertainty in the feedback strength and effects.

Core claim

NMF-FFB approximates the endogenous matrix Y1 as X B, where the coefficient matrix B satisfies the simultaneous equation B = Θ1 Y1 + Θ2 Y2 with non-negative latent feedback matrix Θ1 and exogenous pathway matrix Θ2. Under the condition that the spectral radius of X Θ1 is less than one, the reduced form becomes Y1 ≈ (I - X Θ1)^{-1} X Θ2 Y2, which defines a latent Leontief inverse that isolates direct effects from those amplified through cumulative feedback. The framework is positioned as data-fitting structural equation modeling suitable for non-negative additive data and exploratory analysis with user-specified latent rank Q.

What carries the argument

The embedded simultaneous equation B = Θ1 Y1 + Θ2 Y2 within the NMF approximation, which under stability yields the Leontief inverse (I - X Θ1)^{-1} to compute cumulative feedback-amplified effects.

If this is right

It allows interpretable parts-based factorizations for data such as health indicators or pollution measurements.
Users can obtain estimates of feedback spectral radius, amplification ratio, and path coefficients along with uncertainty measures from X-fixed bootstrap.
The approach suits small samples where covariance estimation in maximum-likelihood SEM would be ill-conditioned.
X groups endogenous indicators into latent factors automatically without confirmatory assumptions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Extending this to longitudinal data could model evolving feedback over time periods.
Linking the latent factors to economic sectors might adapt input-output analysis for exploratory settings.
Simulations with known feedback structures could test how well the method recovers true amplification ratios.
Policy applications might quantify total impacts in environmental or public health systems by tracing feedback paths.

Load-bearing premise

The data are non-negative and additive, and the chosen latent rank together with regularization penalties produce stable factors that keep the feedback spectral radius below one without further adjustments.

What would settle it

Generate synthetic non-negative data from a known linear system with specified feedback matrix whose spectral radius is below one, apply NMF-FFB, and check if the recovered reduced form and amplification ratios match the generating parameters within bootstrap error bounds.

read the original abstract

Non-negative matrix factorization (NMF) approximates a non-negative endogenous data matrix as $Y_1 \approx XB$, with non-negative latent components $X$ and coefficients $B$. Standard covariate-aware NMF is feedforward: $B$ depends only on exogenous variables $Y_2$, with no latent feedback among endogenous variables. We propose NMF-FFB (NMF with feedforward-feedback structure), an exploratory data-fitting framework that embeds the simultaneous equation $B = \Theta_1 Y_1 + \Theta_2 Y_2$ in NMF, where $\Theta_1$ is non-negative latent feedback and $\Theta_2$ non-negative exogenous pathways. NMF-FFB is positioned within data-fitting structural equation modeling (SEM): it fits $Y_1$ directly rather than a model-implied covariance, and is not a confirmatory measurement model or a replacement for maximum-likelihood SEM under standard confirmatory factor analysis assumptions. When $\rho(X\Theta_1)<1$, the reduced form $Y_1 \approx (I-X\Theta_1)^{-1} X\Theta_2 Y_2$ defines a latent Leontief inverse separating direct from cumulative feedback-amplified effects. Estimation uses regularized multiplicative updates with orthogonality and sparsity penalties; an $X$-fixed bootstrap summarizes uncertainty for the feedback spectral radius, the amplification ratio, and path coefficients. Unlike conventional SEM, NMF-FFB requires only the latent rank $Q$ and lets $X$ group endogenous indicators into latent factors. This suits non-negative additive data, automatic loading discovery, Leontief-type cumulative effects, and small samples where covariance-based maximum-likelihood fitting is ill-conditioned. Applications to Holzinger-Swineford, Los Angeles pollution-mortality, and Mississippi county-level health data demonstrate interpretable parts-based representations across distinct latent-feedback regimes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

NMF-FFB embeds a simultaneous equation into NMF to get a latent Leontief inverse for feedback effects, but the fitting procedure does not enforce the required spectral radius condition below 1.

read the letter

The core contribution is a straightforward algebraic extension: they fold the equation B = Θ1 Y1 + Θ2 Y2 into the NMF factorization Y1 ≈ X B, then rearrange to a reduced form that isolates cumulative feedback-amplified effects via the inverse when the radius condition holds. This is new relative to plain NMF or standard SEM, and it targets non-negative additive data where parts-based factors and small-sample stability matter. The paper does a clean job laying out the estimation via regularized multiplicative updates, the bootstrap for uncertainty on the radius and amplification ratio, and three real-data examples that illustrate different feedback regimes. Those examples show the method can produce readable latent groupings without forcing a confirmatory structure. The positioning as exploratory data-fitting rather than full ML SEM is also sensible for the intended use cases. The main limitation is that nothing in the updates or penalties guarantees ρ(X Θ1) stays below 1 after fitting. Because Q and the regularization weights are free choices that directly affect the scale of X and Θ1, the inverse can fail to exist or the amplification numbers can jump with modest hyperparameter shifts. The abstract and description give no post-fit verification step or sensitivity checks on this point, so the interpretive claim about separating direct versus cumulative effects rests on an assumption that may not hold in practice. The work is aimed at applied statisticians handling non-negative matrices in health or environmental settings who already use NMF and want a lightweight way to add feedback without full covariance modeling. It is coherent enough on its own terms to merit a serious referee, though the referee would need to press on empirical checks for the radius condition and direct comparisons to baseline NMF.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes NMF-FFB, which embeds the simultaneous equation B = Θ₁ Y₁ + Θ₂ Y₂ into non-negative matrix factorization for approximating the endogenous matrix as Y₁ ≈ X B. This introduces non-negative latent feedback Θ₁ among endogenous variables alongside exogenous pathways Θ₂. Under the condition ρ(X Θ₁) < 1, the reduced form Y₁ ≈ (I - X Θ₁)^{-1} X Θ₂ Y₂ yields a latent Leontief inverse that separates direct effects from cumulative feedback-amplified effects. Estimation proceeds via regularized multiplicative updates incorporating orthogonality and sparsity penalties, with an X-fixed bootstrap used to quantify uncertainty in the spectral radius, amplification ratio, and path coefficients. The framework is illustrated on the Holzinger-Swineford, Los Angeles pollution-mortality, and Mississippi county health datasets to produce interpretable parts-based latent factors across varying feedback regimes.

Significance. If the central claims hold, NMF-FFB offers a practical exploratory tool for non-negative additive data exhibiting feedback, particularly in small-sample regimes where covariance-based maximum-likelihood SEM becomes ill-conditioned. It combines NMF's automatic loading discovery with a structural reduced-form derivation, enabling Leontief-style cumulative effect analysis without requiring confirmatory factor assumptions. The explicit bootstrap procedure for the spectral radius and amplification ratio, together with the three real-data applications demonstrating distinct latent-feedback regimes, constitute concrete strengths that could support adoption in environmental health and social science settings.

major comments (2)

[Estimation section (regularized multiplicative updates)] Estimation section (regularized multiplicative updates): the updates minimize the NMF objective plus orthogonality/sparsity penalties but contain no explicit constraint, projection, or post-fit check to enforce ρ(X Θ₁) < 1. Because the reduced-form Leontief inverse and the separation of direct versus cumulative effects (stated in the abstract and derived from B = Θ₁ Y₁ + Θ₂ Y₂) are defined only when this spectral-radius condition holds, the absence of enforcement means that admissible interpretations can fail for some user-chosen Q or penalty values; the X-fixed bootstrap merely reports uncertainty and does not restore the condition when it is violated.
[Applications section] Applications section (Holzinger-Swineford, pollution-mortality, and Mississippi health examples): while the fitted factors are described as interpretable, the manuscript does not report the empirical frequency with which ρ(X Θ₁) < 1 holds across the reported Q values and penalty settings, nor does it include sensitivity tables showing how the amplification ratio changes with modest hyperparameter perturbations. This information is load-bearing for the claim that the method reliably separates direct from feedback-amplified effects in practice.

minor comments (2)

[Model section] Notation: the symbols Θ₁, Θ₂, Y₁, and Y₂ are introduced clearly in the abstract but should be restated with consistent subscripting in the first paragraph of the model section to prevent momentary ambiguity when the simultaneous equation is first written.
[Bootstrap subsection] Bootstrap description: the precise mechanism by which X is held fixed while resampling the remaining quantities could be expanded by one sentence or a short algorithm box to improve reproducibility of the uncertainty summaries for the spectral radius.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment below and have revised the manuscript to incorporate the suggested improvements where feasible.

read point-by-point responses

Referee: Estimation section (regularized multiplicative updates): the updates minimize the NMF objective plus orthogonality/sparsity penalties but contain no explicit constraint, projection, or post-fit check to enforce ρ(X Θ₁) < 1. Because the reduced-form Leontief inverse and the separation of direct versus cumulative effects (stated in the abstract and derived from B = Θ₁ Y₁ + Θ₂ Y₂) are defined only when this spectral-radius condition holds, the absence of enforcement means that admissible interpretations can fail for some user-chosen Q or penalty values; the X-fixed bootstrap merely reports uncertainty and does not restore the condition when it is violated.

Authors: We agree that the estimation procedure as currently described does not include an explicit constraint or post-fit safeguard to enforce ρ(X Θ₁) < 1. While the theoretical development conditions the reduced-form interpretation on this inequality, the regularized multiplicative updates can produce solutions that violate it for particular choices of Q or penalty strength. In the revised manuscript we will add a post-estimation verification step that computes the spectral radius after convergence; solutions violating the condition will be flagged as inadmissible and either discarded or re-estimated with adjusted regularization. The estimation section will be updated to document this procedure explicitly. revision: yes
Referee: Applications section (Holzinger-Swineford, pollution-mortality, and Mississippi health examples): while the fitted factors are described as interpretable, the manuscript does not report the empirical frequency with which ρ(X Θ₁) < 1 holds across the reported Q values and penalty settings, nor does it include sensitivity tables showing how the amplification ratio changes with modest hyperparameter perturbations. This information is load-bearing for the claim that the method reliably separates direct from feedback-amplified effects in practice.

Authors: We acknowledge the value of reporting the empirical stability of the spectral-radius condition. In the revised applications section we will include, for each dataset, the proportion of (Q, penalty) combinations that satisfy ρ(X Θ₁) < 1. We will also add sensitivity tables or supplementary figures that display how the amplification ratio responds to modest perturbations around the selected hyperparameter values. These additions will provide direct empirical support for the practical reliability of the direct-versus-cumulative effect separation. revision: yes

Circularity Check

1 steps flagged

Leontief reduced form and direct/cumulative separation are algebraic rearrangement of the embedded simultaneous equation

specific steps

self definitional [Abstract]
"When ρ(XΘ1)<1, the reduced form Y1 ≈ (I-XΘ1)^{-1} XΘ2 Y2 defines a latent Leontief inverse separating direct from cumulative feedback-amplified effects."

The quoted claim is obtained by substituting the paper's defining equation B = Θ1 Y1 + Θ2 Y2 into Y1 ≈ X B, rearranging to (I - XΘ1) Y1 ≈ X Θ2 Y2, and inverting. The separation of direct versus cumulative effects is the definitional property of the Leontief inverse in any simultaneous-equation system; once the simultaneous structure is imposed and parameters are fitted, the reduced-form interpretation holds by algebraic identity rather than by additional derivation or data-driven validation.

full rationale

The paper embeds the simultaneous equation B = Θ1 Y1 + Θ2 Y2 into the NMF factorization Y1 ≈ X B. The reduced form Y1 ≈ (I - XΘ1)^{-1} XΘ2 Y2 and its interpretation as separating direct from feedback-amplified effects then follow immediately by substitution and inversion under the stated spectral-radius condition. This step adds no new empirical content or independent test; it is the standard reduced-form transformation of the defining model. The NMF estimation and penalties are independent of this algebraic identity, but the central interpretive claim about latent Leontief effects reduces directly to the model structure by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The framework depends on user choice of latent rank Q and regularization strengths, assumes non-negativity of all matrices, and introduces fitted feedback parameters Θ1 whose stability condition is checked after fitting.

free parameters (2)

latent rank Q
User-specified number of latent components that determines the dimension of X and B.
regularization parameters for orthogonality and sparsity
Penalty weights in the multiplicative updates that control solution properties.

axioms (2)

domain assumption All entries in Y1 and Y2 are non-negative
Required for the NMF approximation and non-negative constraints on X, B, and Θ to remain valid.
ad hoc to paper Spectral radius ρ(XΘ1) < 1
Invoked to guarantee existence of the matrix inverse and interpretability of cumulative effects.

invented entities (1)

Latent feedback matrix Θ1 no independent evidence
purpose: Captures endogenous feedback pathways among the variables in Y1
New parameter matrix introduced by the simultaneous equation; no external falsifiable prediction supplied in the abstract.

pith-pipeline@v0.9.0 · 5868 in / 1561 out tokens · 47894 ms · 2026-05-21T16:15:38.989130+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

When ρ(XΘ₁)<1, the reduced form Y₁≈(I−XΘ₁)⁻¹XΘ₂Y₂ defines a latent Leontief inverse separating direct from cumulative feedback-amplified effects.
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean embed_add unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The equilibrium operator M_model=(I−XΘ₁)⁻¹XΘ₂ admits the Neumann expansion … each term corresponds to additional rounds of latent propagation.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

Boutsidis C, Gallopoulos E (2008) Svd based initialization: A head start for nonneg- ative matrix factorization. Pattern recognition 41(4):1350–1362 Brook RD, Rajagopalan S, Pope III CA, et al (2010) Particulate ma tter air pollution and cardiovascular disease: an update to the scientiﬁc statement from the american heart association. Circulation 121(21):2...

work page 2008
[2]

Nature 401(6755):788–791 15 Leontief WW (1936) Quantitative input and output relations in the ec onomic systems of the united states

MIT Press, URL https://proceedings.neurips.cc/paper ﬁles/paper/ 2000/ﬁle/f9d1152547c0bde01830b7e8bd60024c-Paper.pdf Lee DD, Seung HS (1999) Learning the parts of objects by non-ne gative matrix factorization. Nature 401(6755):788–791 15 Leontief WW (1936) Quantitative input and output relations in the ec onomic systems of the united states. The Review of ...

work page doi:10.1007/s42081-025-00314-0 2000

[1] [1]

Boutsidis C, Gallopoulos E (2008) Svd based initialization: A head start for nonneg- ative matrix factorization. Pattern recognition 41(4):1350–1362 Brook RD, Rajagopalan S, Pope III CA, et al (2010) Particulate ma tter air pollution and cardiovascular disease: an update to the scientiﬁc statement from the american heart association. Circulation 121(21):2...

work page 2008

[2] [2]

Nature 401(6755):788–791 15 Leontief WW (1936) Quantitative input and output relations in the ec onomic systems of the united states

MIT Press, URL https://proceedings.neurips.cc/paper ﬁles/paper/ 2000/ﬁle/f9d1152547c0bde01830b7e8bd60024c-Paper.pdf Lee DD, Seung HS (1999) Learning the parts of objects by non-ne gative matrix factorization. Nature 401(6755):788–791 15 Leontief WW (1936) Quantitative input and output relations in the ec onomic systems of the united states. The Review of ...

work page doi:10.1007/s42081-025-00314-0 2000