---
title: Conservative Matrix Field (CMF)
url: https://www.emergentmind.com/topics/conservative-matrix-field-cmf
type: topic
---

# Conservative Matrix Field (CMF)

Searching arXiv for recent papers on "Conservative Matrix Field" and closely related usages to ground the article.
Conservative Matrix Field (CMF) is a term used in several recent research programs to denote a structure constrained by a conservation, integrability, or path-independence principle, but the object itself depends strongly on context. In arithmetic and Ore-algebra settings, a CMF is a matrix-valued cocycle on the lattice $\mathbb{Z}^d$ whose discrete flatness yields path-independent transport, generalized Apéry limits, and higher-dimensional asymptotics [2303.09318; 2507.08138; 2604.09723]. In reduced stochastic modeling, the CMF is the symmetric matrix field $S$ in a drift decomposition $f(x)\approx \Phi s(x)$ with $\Phi=S+N$, where $S$ determines the conservative drift and the diffusion tensor [2505.01895]. In CNN feature processing, the term is not introduced explicitly, but it naturally describes feature tensors whose channel groups form a curl-free spatial vector field, operationalized through Green’s function convolution layers that project feature maps onto conservative fields [2003.05182].

## 1. Principal meanings of the term

The contemporary literature does not use a single universal definition of CMF. Instead, the phrase organizes several technical notions that share a conservation constraint.

| Context | CMF object | Conservative meaning |
|---|---|---|
| Arithmetic and Ore algebras | $M:\mathbb{Z}^d\to GL_r(K(\mathbf{x}))$ | cocycle, path-independence, discrete flatness |
| Reduced stochastic dynamics | symmetric part $S$ of $\Phi=S+N$ | reversible drift $f_c(x)=S\,s(x)$ |
| CNN feature processing | channel-grouped vector field tensor | curl-free, integrable field $F=\nabla\phi$ |

In the arithmetic line, the term was introduced as a new structure initially developed to elucidate the methodologies employed by Apéry in his proof of the irrationality of $\zeta(3)$, and then extended to higher-dimensional, matrix-valued, shift-invariant settings that connect to gauge transformations and finite-dimensional modules of Ore algebras [2303.09318; 2507.08138]. In the stochastic line, the CMF is explicitly the symmetric matrix field $S$ that maps the score $s(x)=\nabla\log\rho^*(x)$ to the conservative drift and simultaneously fixes the diffusion tensor of a reduced Langevin model [2505.01895]. In the CNN line, the 2020 paper does not introduce the term “Conservative Matrix Field” explicitly, but it operationalizes projection onto conservative feature fields via Poisson solves in Fourier space and describes the result as “regularizing the field by forcing it to be conservative and physically interpretable” [2003.05182].

This suggests that the adjective “conservative” is overloaded across subfields but consistently signals an integrability constraint: path-independent transport on a lattice, reversible drift on an invariant measure, or curl-free feature geometry.

## 2. Discrete flatness, potentials, and gauge structure

In the arithmetic formulation of 2023, a CMF is a pair of $2\times 2$ polynomial matrices
$$
M_X(x,y),\ M_Y(x,y)\in \mathrm{Mat}_{2\times 2}(\mathbb{Z}[x,y])
$$
satisfying
$$
M_X(x,y)\,M_Y(x+1,y)=M_Y(x,y)\,M_X(x,y+1)
$$
for all integers $x,y$. This is a discrete zero-curl condition: the product around each unit square is conserved. Fixing an origin, one obtains a matrix potential $S(n,m)$ from path products, and the conservative condition implies that $S(n,m)$ is independent of the monotone path chosen from the basepoint. When the matrices are invertible on a simply-connected grid region, the discrete Poincaré lemma holds, so there exists a potential $S$ such that
$$
M_X(x,y)=S(x,y)^{-1}S(x+1,y),\qquad M_Y(x,y)=S(x,y)^{-1}S(x,y+1),
$$
with uniqueness up to right multiplication by a constant matrix [2303.09318].

The 2025 extension recasts this in a dimension-$d$, rank-$r$ language. A CMF over a field $K$ is a map
$$
M:\mathbb{Z}^d\to GL_r(K(\mathbf{x})),\qquad v\mapsto M_v
$$
satisfying the cocycle equation
$$
M_{v+w}=M_v\cdot \sigma_v(M_w).
$$
If $A_i(x):=M_{e_i}(x)$, this is equivalent to the generator flatness relations
$$
M_i\cdot \sigma_i(M_j)=M_j\cdot \sigma_j(M_i),
$$
or, evaluated pointwise,
$$
A_i(x+e_j)A_j(x)=A_j(x+e_i)A_i(x).
$$
The same paper states that on any simply connected domain of regular points there exists an invertible matrix potential $F$ with
$$
A_i(x)=F(x+e_i)F(x)^{-1},\qquad M_v(x)=F(x+v)F(x)^{-1},
$$
and encodes path-independence as
$$
M_{v+w}(x)=M_v(x)\cdot M_w(x+v).
$$
Gauge or coboundary equivalence is given by
$$
\bar M_v=A\cdot M_v\cdot \sigma_v(A^{-1}),
$$
which preserves the cocycle condition and changes only the trivialization of the same discrete flat connection [2507.08138].

A central construction in the Ore-algebra framework starts from a D-finite function $f$ and a basis $B=(b_1.f,\dots,b_r.f)$ of the finite-dimensional module generated by shift and Euler operators. The basis-change matrices defined by
$$
(b_1.f,\dots,b_r.f)\cdot M_v^f=(S_v b_1.f,\dots,S_v b_r.f)=\sigma_v(B)
$$
form a CMF. This identifies CMFs with finite-dimensional representations of shift operators on D-finite modules and makes gauge transformations, contiguous relations, and companion reductions intrinsic rather than ad hoc [2507.08138].

## 3. Apéry limits, continued fractions, and irrationality theory

One-dimensional CMF ratios recover the classical ratio paradigm behind Apéry limits. For a CMF $M$, a trajectory $(x,v)$, and vectors $p,p',q,q'$, the ratio
$$
L_{x,v}^{p,p',q,q'}(n)=\frac{p^t M_{nv}(x) p'}{q^t M_{nv}(x) q'}
$$
specializes in the rank-one, one-dimensional case to an ordinary scalar ratio, and in companion-form rank-$r$ trajectories to ratios of D-finite sequences. The 2025 asymptotic paper states that classical Apéry limits arise as special cases, and that when a trajectory matrix is brought into companion form satisfying strict Poincaré–Perron hypotheses, the limit and convergence rate are controlled by the dominant characteristic roots exactly as in the ordinary recurrence setting [2507.08138].

The 2023 paper develops this mechanism concretely for $\zeta(3)$. It constructs a self-dual CMF with normalized generators
$$
M_X(x,y)=\begin{pmatrix}0&1\\ -\,(x+1)^6 & a(x,y)\end{pmatrix},\qquad
M_Y(x,y)=\begin{pmatrix}\bar f(x,y)&1\\ -\,x^6& f(x,y)\end{pmatrix},
$$
where
$$
a(x,y)=x^3+(x+1)^3+2y(y-1)(2x+1).
$$
On the bottom line $y=1$, the associated convergents satisfy
$$
\frac{p_n(1)}{q_n(1)}=\sum_{k=1}^{n}\frac{1}{k^3},\qquad q_n(1)=(n!)^3,
$$
so the CMF recovers the standard partial sums of $\zeta(3)$. Along the diagonal $(n,m)=(N,N+1)$, the system reduces to a generalized continued fraction
$$
\frac{6}{\zeta(3)}-5=K_1^{\infty}\frac{-k^6}{34k^3+51k^2+27k+5},
$$
whose continuants satisfy the Apéry-type recurrence
$$
U_{n+1}=F(n)\,U_n-n^6\,U_{n-1},\qquad
F(n)=34n^3+51n^2+27n+5,
$$
or equivalently
$$
(n+1)^3 U_{n+1}=(34n^3+51n^2+27n+5)\,U_n-n^3\,U_{n-1}.
$$
With initial conditions $A_0=1,A_1=5$ and $B_0=0,B_1=6$, this is Apéry’s recurrence, and $B_n/A_n\to \zeta(3)$ [2303.09318].

The same work derives the growth roots
$$
\lambda_{\pm}=17\pm 12\sqrt{2}=(1\pm \sqrt{2})^4
$$
and the factorial reduction estimate
$$
\gcd(P_n,Q_n)\ \ge\ \frac{(n!)^6}{\mathrm{lcm}[n]^3}.
$$
Combined with the bound $e^3<\lambda_+$ and the error estimate for $|Q_n\zeta(3)-P_n|$, this yields the irrationality criterion and recovers Apéry’s theorem within the CMF formalism [2303.09318].

The 2023 paper also presents CMF realizations for other constants. For $\ln 2$, it uses
$$
f(x,y)=x+y,\qquad \bar f(x,y)=x-y,
$$
giving
$$
\ln 2=\frac{1}{1+K_1^{\infty}\frac{n^2}{1}}.
$$
For $e$, it uses
$$
f(x,y)=x+y,\qquad \bar f(x,y)=1,
$$
yielding
$$
K_1^{\infty}\frac{n}{n}=\frac{1}{e-1}.
$$
For $\pi$, it notes the generalized continued fraction
$$
\pi=3+K_1^{\infty}\frac{(2n-1)^2}{6}.
$$
The paper explicitly states that no CMF-based factorial reduction is currently known that would imply new irrationality measures for $\pi$, and that for $\ln 2$, $e$, and $\zeta(2)$ the framework is primarily explanatory rather than stronger than the best known arithmetic methods [2303.09318].

## 4. Rank-2 CMFs, symmetric squares, and arithmetic functoriality

The 2026 arithmetic paper develops CMFs in a rank-$2$ framework adapted to hypergeometric recurrences, canonical polynomial recurrences, and Apéry-like kernels. Here a CMF of dimension $d$ and rank $r$ over a field $K$ is described as a $1$-cocycle of the lattice $\mathbb{Z}^d$ with values in $GL_r(K(\mathfrak{x}))$, again satisfying
$$
\mathcal M_{v+w}=\mathcal M_v\,{}_v(\mathcal M_w).
$$
For rank-$2$ objects, the $\operatorname{Sym}^2$ functor acts on
$$
M=\begin{pmatrix}a&b\\ c&d\end{pmatrix}
$$
by
$$
\operatorname{Sym}^2(M)=
\begin{pmatrix}
a^2 & ab & b^2\\
2ac & ad+bc & 2bd\\
c^2 & cd & d^2
\end{pmatrix},
$$
and the paper proves an explicit square-gauge matrix
$$
\Phi(a,b,c;z)=
\begin{pmatrix}
1 & 0 & \dfrac{2abz}{1-z}\\[2mm]
0 & 2 & \dfrac{2((a+b)z-c+1)}{1-z}\\[2mm]
0 & 0 & 2
\end{pmatrix},
\qquad \det\Phi=4,
$$
with
$$
M_u^{(g)}(a,b,c;z)=\Phi(a,b,c;z)^{-1}\,\operatorname{Sym}^2\!\bigl(M_u^{(f)}(a,b,c;z)\bigr)\,\sigma_u(\Phi)(a,b,c;z).
$$
This identifies the square of a Gauss hypergeometric rank-$2$ CMF with a rank-$3$ symmetric-square CMF by an explicit rational gauge [2604.09723].

A major result is the summation-lift classification of the order-$3$ canonical recurrences printed in Appendix B.6 of the cited work of Raz, Shalyt, Leibtag, Kalisch, Weinbaum, Hadad, and Kaminer. The paper proves that each such order-$3$ recurrence is a shifted summation lift of an explicit order-$2$ kernel. It identifies the three kernels as follows: the first $\pi$-kernel is an explicit rescaling of the sporadic Apéry-like sequence $A036917$; the second $\pi$-kernel is an explicit rescaling of the Domb numbers $A002895$; and the Catalan kernel is a hypergeometric twist of the Gauss-square coefficient sequence at $(a,b,c)=(\tfrac12,1,\tfrac32)$ [2604.09723].

The same paper places these constructions in a unified pullback–twist formalism. For a rational pullback $\phi$ and scalar twist $\rho$, a transported CMF basis satisfies
$$
\widetilde M_v(x,\lambda)=\frac{\sigma_v(\rho)}{\rho}\,M_v(\phi(x),\lambda),\qquad
\widetilde M_{\theta_x}(x,\lambda)=\frac{x\rho'}{\rho}\,I_r+\frac{x\phi'}{\phi}\,M_{\theta_z}(\phi(x),\lambda),
$$
and for rank-$2$ objects
$$
\operatorname{Sym}^2\bigl(\rho^{1/2} B\circ\phi\bigr)=\rho\,\operatorname{Sym}^2(B)\circ\phi.
$$
The Domb kernel is recovered by recasting the degree-$3$ Belyi pullback
$$
\phi(x)=\frac{108x^2}{(1-4x)^3}
$$
and its algebraic twist in CMF language [2604.09723].

The inverse classification theorem in that paper isolates the unique $\operatorname{Sym}^2(\mathrm{Gauss})$ point in a one-parameter family of Fuchsian operators by the accessory-parameter condition
$$
\lambda_0=2\gamma_1\gamma_2(1-2\alpha).
$$
It further reports a Belyi-pullback scan over $5040$ configurations, producing $11$ additional integer sequences of the form
$$
[x^n]\lambda^n\,{}_2F_1(a,b;c;\phi(x))^2,
$$
with integrality proved and all examples placed in the same $\operatorname{Sym}^2$-pullback framework [2604.09723].

## 5. Conservative matrix fields in reduced stochastic dynamics

In the multiscale stochastic literature, the CMF is not a cocycle on a lattice but the symmetric matrix field governing the reversible component of a reduced Langevin model. The starting point is the additive-noise SDE
$$
\mathrm{d}x_t = f(x_t)\,\mathrm{d}t + \sqrt{2}\,\Sigma\,\mathrm{d}W_t,
$$
with diffusion tensor
$$
D=\Sigma\Sigma^\top.
$$
For the stationary density $\rho^*(x)$, the score is
$$
s(x):=\nabla\log\rho^*(x),
$$
and the stationary probability current is
$$
j(x)=\bigl(f(x)-D\,s(x)\bigr)\rho^*(x).
$$
Under detailed balance, $j(x)\equiv 0$, and the conservative drift is
$$
f_c(x):=D\,s(x).
$$
In the non-reversible case the paper writes
$$
f(x)=f_c(x)+f_{\mathrm{irr}}(x),\qquad f_c(x)=D\,s(x),
$$
with
$$
j(x)=f_{\mathrm{irr}}(x)\rho^*(x),\qquad \nabla\cdot j(x)=0.
$$
It also introduces an antisymmetric tensor field $R(x)$ and uses an approximate constant antisymmetric matrix $\tilde R$ to represent “minimal irreversible circulation” [2505.01895].

The reduced drift is modeled as
$$
f(x)\approx \Phi\,s(x),\qquad \Phi\in\mathbb{R}^{D\times D}\ \text{constant},
$$
with
$$
\Phi=S+N,\qquad
S:=\tfrac12(\Phi+\Phi^\top),\qquad
N:=\tfrac12(\Phi-\Phi^\top).
$$
In this formulation, the conservative drift is
$$
f_c(x)=S\,s(x),
$$
and the paper refers to the symmetric matrix field $S$ as the Conservative Matrix Field. The same $S$ determines the diffusion tensor through
$$
\Sigma=\mathrm{chol}(S),\qquad D=S,
$$
so the CMF simultaneously fixes the reversible drift and the noise covariance of the surrogate model [2505.01895].

Identification proceeds in two stages. First, the score is estimated from stationary data with the k-means Gaussian-mixture method (KGMM), using the conditional-expectation identity
$$
s(x)=-\frac{1}{\sigma_G^2}\,\mathbb{E}[z\mid x]
$$
for an equally weighted isotropic-kernel mixture. Second, short-time transitions on a finite-volume partition define a rate matrix $Q$ with $Q\pi=0$, and short-time correlation matching yields
$$
M=\Phi\,V^\top,\qquad \Phi=M\,(V^\top)^{-1},
$$
with the Moore–Penrose pseudoinverse selecting the minimum-norm $\Phi$ and therefore the minimum-norm antisymmetric part $N$. The paper describes this as implementing the “minimal irreversible circulation” principle [2505.01895].

The framework is validated on three systems. In the one-dimensional nonlinear multiplicative-noise benchmark, there is no nontrivial antisymmetric part, so
$$
N\equiv 0,\qquad S=D>0,\qquad f(x)=S\,s(x),
$$
and the method reproduces the stationary density and autocorrelation function with high fidelity. In the two-dimensional asymmetric four-well potential with non-gradient drift, the reconstruction gives
$$
S\approx I,\qquad
N\approx
\begin{pmatrix}
0&-0.8\\
0.8&0
\end{pmatrix},
$$
so the CMF is identity while the antisymmetric part recovers the rotational component. In stochastic Lorenz-63, the constant-matrix approximation preserves the invariant measure and matches short-time correlations well for $x_1,x_2$, but for $x_3$ the ACF peak positions are captured while oscillations are smoothed [2505.01895].

## 6. Conservative feature fields in convolutional neural networks

The CNN usage begins from the classical definition of a conservative vector field. For a domain $\Omega\subset\mathbb{R}^d$, with $d\in\{2,3\}$, a vector field $F$ is conservative if there exists a scalar potential $\phi$ such that $F=\nabla\phi$; equivalently, on simply connected domains, $F$ is conservative if and only if it is curl-free, and the orthogonal $L^2$ projection of a general field onto conservative fields is obtained by solving
$$
\Delta\phi=\nabla\cdot F,\qquad F_{\mathrm{cons}}=\nabla\phi.
$$
In CNN feature maps of shape $H\times W\times C$ or $D\times H\times W\times C$, a “Conservative Matrix Field” in this context is a tensor whose channel groups represent a vector field that is conservative across spatial dimensions: if channels are split into $F_x,F_y$ or $F_x,F_y,F_z$, then there exists $\phi$ such that the grouped tensor equals $\nabla\phi$ [2003.05182].

The 2020 paper does not name this object explicitly, but implements its projection by Green’s function convolution (GFC) layers. Using the discrete Laplacian kernel
$$
L_\Delta=
\begin{bmatrix}
0&1&0\\
1&-4&1\\
0&1&0
\end{bmatrix},
$$
embedded into a padded domain, and a discrete Dirac delta, the numerical Green’s function in Fourier space is
$$
\hat G(k)=\frac{\mathcal F(\delta)(k)}{\mathcal F(L_\Delta)(k)}.
$$
The Poisson solve is then
$$
\phi=\Re\left(\mathcal F^{-1}\big(\mathcal F(L)\cdot \hat G\big)\right)+c,
$$
with padding, periodic boundary conditions on the padded array, and a fixed DC component $\hat G[0,0]=0$ to remove the Laplacian null space. Each GFC call has complexity $O(N\log N)$ per channel, where $N$ is the number of pixels or voxels [2003.05182].

Three Green’s-function-based layers are defined. Laplacian Integration (LI) interprets the input tensor as a discrete Laplacian and returns the potential $\phi$. Gradient Integration (GI) interprets half the channels as $F_x$ and half as $F_y$, forms the divergence
$$
L=\nabla\cdot F=\partial_x F_x+\partial_y F_y,
$$
solves the Poisson equation, and returns the least-error potential. Gradient Integration Derivative (GID) optionally applies a linear $1\times 1$ convolution, computes derivatives, integrates via GFC, differentiates again, and outputs the conservative field
$$
F_{\mathrm{cons}}=\nabla\phi=[\partial_x\phi,\partial_y\phi].
$$
The paper explicitly characterizes GI and GID as hard projection layers rather than soft penalties: the architecture regularizes the feature space by forcing it to be conservative and physically interpretable, while the integration step itself has no trainable parameters except for the optional linear $1\times 1$ recombination [2003.05182].

Empirically, the method is evaluated on MNIST classification with a reduced GoogLeNet containing two Inception modules, implemented in TensorFlow with Adam, learning rate $1\mathrm{e}{-4}$, and batch size $50$. One GID layer is inserted per Inception module. The reported results are that convergence to $97\%$ validation accuracy is $\approx 5.1\times$ faster in iterations with GID; after $20{,}000$ iterations, smoothed validation accuracy is $98.80\%$ with GID versus $98.35\%$ for the baseline, corresponding to a $\approx 27\%$ reduction in error rate; training curves are smoother with GID; per-iteration training time increases by $\approx 2.0\times$ due to FFTs; and time to $97\%$ is still $\approx 2.5\times$ faster in wall-clock. The paper describes these as early prototype results and states that broader benchmarks and deeper CNNs remain future work [2003.05182].

## 7. Limitations, conjectures, and unresolved directions

The arithmetic CMF literature emphasizes open asymptotic and algorithmic problems. The 2025 paper formulates continuity conjectures asserting that, away from resonant directions, CMF limits, normalized convergence rates $\rho(v)/\|v\|$, irrationality measures, and normalized log-eigenvalues vary continuously with the direction $v$ on the lattice. If proved, these statements would extend Poincaré–Perron asymptotics to higher dimensions and support optimization-based searches for new irrationality proofs. The same paper also points to resonant directions, where eigenvalue moduli cross and non-convergence or discontinuity may occur [2507.08138].

The 2023 CMF paper lists additional structural problems: classification of nondegenerate conjugate pairs $(f,\bar f)$ in higher degree, treatment of non-invertible regions, extension beyond $2\times 2$ matrices, stronger factorial reduction for constants such as $\pi$, $e$, and $\ln 2$, precise links to modularity, systematic algorithmic discovery, and analytic criteria ensuring that all paths with $\max(n_i,m_i)\to\infty$ converge to the same limit [2303.09318]. The 2026 rank-$2$ paper, by contrast, shows that functorial operations such as $\operatorname{Sym}^2$, pullback, and twist can be handled explicitly at the CMF level, but its inverse classification and integrality results also make clear that the structure is tightly constrained by accessory parameters, Belyi pullbacks, and the choice of gauge [2604.09723].

The stochastic and CNN literatures have different limitations. In reduced Langevin models, the constant-matrix approximation means that a constant antisymmetric part $N$ cannot capture strongly state-dependent circulation, as illustrated by Lorenz-63, and the use of additive noise may oversmooth multiplicative-noise effects even when PDFs and short-time ACFs are recovered well [2505.01895]. In CNNs, the FFT-based Green’s function convolution is more expensive than a standard convolution, the method depends on padding and DC handling, sensitivity at high resolution and in 3D remains to be studied, and broader benchmarking across deeper architectures and more complex images is still required [2003.05182].

Taken together, these strands indicate that CMF is best understood as a family of conservation-based formalisms rather than a single canonical object. In one strand, conservativity means cocycle path-independence on a lattice; in another, it means the reversible part of a drift compatible with an invariant measure; in another, it means curl-free and integrable feature geometry. The shared theme is not a common implementation, but the imposition of a nonlocal compatibility condition that constrains admissible dynamics, transports, or representations.

Source: https://www.emergentmind.com/topics/conservative-matrix-field-cmf