🤖 AI

Lattice theory and algebraic models for deep convolutional learning based on mathematical morphology

This paper establishes a rigorous algebraic framework based on lattice theory and mathematical morphology to analyze deep convolutional networks, revealing that standard CNN layers form non-idempotent cross-lattice operators that explain the representational power of depth, while also proposing and characterizing three genuine idempotent morphological layer designs and unifying various pooling and pyramid techniques under a unified adjoint theory.

Original authors: Gustavo (Jesus), Angulo

Published 2026-05-26

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Gustavo (Jesus), Angulo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand how a deep learning computer "sees" an image. Usually, we think of these networks as a series of mathematical steps: a filter blurs the image, a function cuts off negative numbers, and a pooler shrinks the picture.

This paper, written by Gustavo Angulo, argues that we have been looking at these steps through the wrong lens. Instead of just seeing them as arithmetic, the author suggests we view them through the lens of Mathematical Morphology—a branch of math originally designed for analyzing shapes, like finding the outline of a rock or the edge of a cloud.

Here is the paper's story, broken down into simple concepts and analogies.

1. The Core Idea: The "Shape" of Math

The paper claims that deep learning networks (like CNNs, ResNets, and UNets) are actually built on a hidden structure called Lattice Theory.

Think of a lattice as a set of rules for comparing things. In a standard network, we compare numbers (is 5 bigger than 3?). In this "morphological" view, we compare shapes and structures.

Erosion: Imagine shrinking a shape by sanding off its edges. In the paper, this is like a "filter" that looks for specific patterns.
Dilation: Imagine a shape growing or expanding. This is like "pooling," where the network takes the biggest value in a neighborhood.
Opening: If you shrink a shape and then grow it back, you get a smoothed version of the original. This is called an "opening."

2. The Big Surprise: Standard Networks are "Broken"

The paper's most famous finding is that the standard way we build AI networks today is actually mathematically inconsistent.

The Analogy: Imagine you are building a machine. You have a part that works in "Metric System" (Centimeters) and another part that works in "Imperial System" (Inches). If you connect them directly without a converter, the machine doesn't work right.
The Paper's Claim:
- The Convolution step (the filter) lives in the "Fourier Lattice" (a world of frequencies and waves).
- The Max-Pooling step (shrinking the image) lives in the "Pointwise Lattice" (a world of individual pixel values).
- The Problem: When you connect them, you are jumping between two different mathematical worlds. Because of this "cross-lattice" jump, the network is not idempotent.
What is Idempotent? Imagine a coffee filter. If you pour coffee through it once, you get clean coffee. If you pour that clean coffee through the same filter again, it stays clean. It doesn't change anymore. That is "idempotent."
The Result: The paper proves that standard CNN layers are not like that coffee filter. If you run an image through a standard layer twice, you get a different result than running it once. The paper argues that this "instability" is actually why deep networks are so powerful—they keep changing the data, adding new layers of complexity. But it also means they are mathematically messy.

3. The Solution: Three "Perfect" Designs

The author doesn't just point out the mess; he designs three new types of layers that are mathematically perfect (idempotent). Think of these as three different ways to build a "perfect coffee filter."

Type I: The Pure Shape Filter.
- This uses the same "shape" to shrink and then grow the data. It stays in the same mathematical world the whole time.
- Result: It stabilizes instantly. If you run an image through it once, it's done. Running it again changes nothing.
Type II: The Frequency Filter.
- This stays in the "Fourier" world (the world of waves). It uses a special math trick (Wiener deconvolution) to clean up the signal.
- Result: It is perfect in the limit, acting like a precise spectral filter.
Type III: The Balanced (Self-Dual) Filter.
- Standard networks treat positive numbers (bright spots) and negative numbers (dark spots) very differently. They often just delete the negative ones (using ReLU).
- This new design treats positive and negative numbers as two sides of the same coin. It uses a "Median Lattice" where the rules are symmetrical.
- Result: It's perfect for data that has both positive and negative values (like the "residuals" in ResNets). It preserves the balance of the data.

4. New Architectures: The "U-ResNet"

Based on these findings, the author proposes a new network design called UResNet.

The Old Way (UNet): Imagine a pipeline where you compress a message (encoder) and then try to expand it back (decoder). To help the decoder, you send a copy of the original message down the side (skip connection). In standard networks, this copy is just a "concatenation" (gluing the data together).
The New Way (UResNet): The paper argues that the skip connection should carry the difference (the residue) between the original and the compressed version.
The Analogy: Instead of sending a photocopy of the whole document to the decoder, you send a "correction note" that says, "Here is what we lost when we compressed it." This allows the decoder to reconstruct the image exactly, without losing any detail.

5. Why ReLU is Weird

The paper also analyzes ReLU (the function that turns negative numbers to zero).

The Finding: ReLU is a "closing" operation (it expands the data to include zero), but its "partner" (the mathematical inverse) is a global operator.
The Metaphor: Imagine a local rule: "If you see a red car, stop." That's a local rule. ReLU's partner rule is: "If anywhere in the entire universe there is a red car, stop."
The Consequence: Because ReLU's partner is "global" (it looks at the whole image at once), it cannot form a perfect mathematical pair with local operations like max-pooling. This is another reason why standard networks are "cross-lattice" and messy.

Summary

This paper is a rigorous mathematical audit of deep learning. It says:

Current networks are messy: They jump between different mathematical worlds, which is why they are powerful but hard to analyze.
We can build "perfect" layers: By sticking to one mathematical world (using specific erosion/dilation pairs), we can create layers that stabilize instantly and are mathematically predictable.
We can fix the architecture: By changing how we handle skip connections (sending "residues" instead of raw data), we can build networks that reconstruct images perfectly.

The author isn't claiming these new networks are already better at winning image contests; rather, he is providing the algebraic blueprint for how to build them so they make mathematical sense. He is giving us the "physics" behind the "engineering" of deep learning.

Technical Summary: Lattice Theory and Algebraic Models for Deep Convolutional Learning Based on Mathematical Morphology

Problem Statement
The paper addresses the lack of a rigorous algebraic framework for understanding the internal structure of deep convolutional neural networks (CNNs), ResNets, and encoder–decoder architectures like UNet. While existing approaches utilize tropical geometry (for ReLU networks), harmonic analysis (for spectral bias), or category theory (for compositional structure), they often lack a unified operator-theoretic foundation grounded in the specific algebraic properties of the operations used. The central problem is to determine whether standard CNN pipelines (linear convolution + ReLU + max-pooling) possess the algebraic properties of morphological operators (specifically adjunctions and idempotency) and, if not, to identify the precise algebraic reasons for their behavior and propose theoretically consistent alternatives.

Methodology
The author employs Mathematical Morphology and Lattice Theory as the primary analytical tools. The core theoretical engine is the Matheron–Maragos–Banon–Barrera (MMBB) universal representation theory, which characterizes translation-invariant (TI) operators on complete lattices.

The methodology proceeds through the following steps:

Lattice Identification: The paper maps standard deep learning operations to specific complete lattices. It distinguishes between the pointwise lattice $(\text{Fun}(E, \mathbb{R}), \leq)$ , the Fourier inf-semilattice $(L_n, \leq_F)$ , and the median inf-semilattice $(\text{Fun}(E, \mathbb{R}), \preceq)$ .
Operator Classification: Standard operations (convolution, ReLU, max-pooling) are analyzed as morphological operators (erosions, dilations, openings, closings) within these lattices.
Adjunction Analysis: The paper investigates whether pairs of operations form adjunctions (Galois connections). It specifically examines the "cross-lattice" nature of standard CNNs, where convolution acts as an erosion in the Fourier lattice while pooling acts as a dilation in the pointwise lattice.
Fixed-Point and Idempotency Theory: The paper analyzes the convergence properties of layered architectures. It defines idempotency ( $\Phi \circ \Phi = \Phi$ ) as the defining characteristic of morphological openings and tests whether standard and proposed layers satisfy this.
Algebraic Construction: Based on the analysis, the paper constructs new layer designs (Type I, II, III) that strictly adhere to lattice-theoretic principles, ensuring idempotency and well-defined adjoints.

Key Contributions

C1. Max-Times Algebra Equivalence: The paper establishes the max-times adjunction (multiplicative structuring functions) as a canonical morphological adjunction for positive feature maps (e.g., post-softmax). It proves its algebraic equivalence to the classical max-plus adjunction via the log/exp isomorphism.
C2. Banon–Barrera Representation for Signed Operators: The author extends the MMBB framework to non-increasing operators (essential for signed CNN kernels) using the Banon–Barrera (1993) sup-generating decomposition. This decomposes any TI operator into a supremum of inf-combinations of erosions and anti-dilations.
C3. Morphological Basis of Convolution: The paper characterizes the morphological basis of a convolution kernel $k$ as isomorphic to the hyperplane orthogonal to $k$ in $\mathbb{R}^N$ . It introduces the characteristic matrix $A_k$ and the concept of a virtual basis for quantized signals, providing an exact algebraic representation for binary and low-bit networks.
C4. Cross-Lattice Structure of Standard CNNs: A principal finding is that the standard CNN pipeline (Convolution + ReLU + Max-Pooling) is a cross-lattice operator. Convolution is an erosion in the Fourier lattice, while max-pooling is a dilation in the pointwise lattice. Consequently, their composition is not a morphological opening in either lattice and is not idempotent. This non-idempotency is identified as the algebraic source of the representational power of depth in standard CNNs.
C5. Non-Locality of the ReLU Adjoint: The paper proves that the upper adjoint of ReLU in the pointwise lattice is a global (non-local) operator (identity on globally non-negative functions, $-\infty$ otherwise). This implies that no local morphological erosion can form an adjunction pair with ReLU, explaining why ReLU cannot be part of a standard morphological opening.
C6. Activation–Pooling Dilation (APD) Factorisation: The paper unifies ReLU and decimated max-pooling into a single dilation operator, $\Psi^{\text{APD}}_{R;\alpha}$ . It derives the correct global adjoint for this operator, which is a piecewise-constant upsampling followed by a bias shift.
C7. Self-Dual Operators and Signed Activations: The author introduces the median inf-semilattice and self-dual operators (erosions and openings that commute with negation). This provides a mathematically consistent framework for signed feature maps (e.g., ResNet residuals), where standard ReLU fails to preserve lattice duality.
C8. Three Idempotent Layer Designs: The paper identifies and characterizes three families of CNN-like layers that are genuine morphological openings:
- Type I: Pure max-plus morphological layer (pointwise lattice).
- Type II: Spectral Wiener layer (Fourier lattice, idempotent in the limit $\epsilon \to 0$ ).
- Type III: Self-dual opening (median lattice).

Results and Architectural Proposals

Convergence of Morphological ResNets: For Type I layers, the paper proves that iterating the layer converges to a fixed point in one step due to idempotency. In contrast, a naive residual iteration $\Phi(f) + f$ (without the specific approximation $F \approx \gamma - \text{id}$ ) diverges.
UResNet Architecture: The paper proposes UResNet, a modification of UNet where skip connections carry the top-hat residue ( $f - \gamma(f)$ ) rather than concatenated features. This design allows the decoder to achieve exact scale-by-scale reconstruction of the encoder input by algebraically compensating for the information lost during the morphological opening.
Fixed-Point Sets: The paper characterizes the fixed-point sets of these layers. For Type I, the set is the image of the opening. For Type III (self-dual), the fixed-point set is closed under negation, making it suitable for symmetric data distributions.
Pooling and Pyramids: Max-pooling, strided convolution, and the Laplacian pyramid are unified under the Goutsias–Heijmans adjoint pyramid theory, showing they are all cases of adjoint erosion-dilation pairs in appropriate lattices.

Significance and Claims

The paper claims to provide the lattice-specific operator-theoretic content that is currently missing from both categorical and computational approaches to deep learning. Its significance lies in:

Algebraic Explanation of Depth: It provides a precise algebraic reason why standard CNNs are powerful: their layers are not idempotent because they are cross-lattice compositions. Depth introduces genuine representational power precisely because the composition is not a simple projection.
Unified Language: It offers a unified language to analyze CNNs, ResNets, and UNets in terms of lattice operators, adjunctions, and morphological bases, independent of whether the practitioner explicitly uses morphological layers.
Theoretical Foundation for New Architectures: It proposes new architectures (APMO, UResNet) that are theoretically well-founded, offering alternatives to standard designs that guarantee idempotency and exact reconstruction properties.
Clarification of ReLU: It rigorously demonstrates that ReLU is a closing (not a dilation in the max-plus sense) with a non-local adjoint, distinguishing it from standard morphological dilations and explaining the limitations of forming adjunctions with it.

The paper explicitly states it does not design, train, or benchmark new layers computationally; rather, it provides the algebraic framework that underpins both morphological neural networks and standard deep architectures viewed through a morphological lens. It positions the learning of morphological bases as an implicit process in standard training, where gradient descent selects a basin of attraction in the space of structuring functions.