---
title: Basic Concepts in Math and Computation
url: https://www.emergentmind.com/topics/basic
type: topic
---

# Basic Concepts in Math and Computation

The term “basic” arises in numerous mathematical and computational contexts, often denoting a canonical or irreducible object, a structural primitive, or a special level within a hierarchy. The following article surveys the rigorous technical meanings and frameworks where "basic" is formally defined, as exemplified by recent research on arXiv.

## 1. Basic-Level Categorization in Cognitive and Vision-Language Models

In psychology, Rosch (1976) introduced the “basic level” of categorization as an intermediate granularity in the taxonomic hierarchy: superordinate (e.g., “animal”), basic (“dog”), subordinate (“golden retriever”). Basic-level categories are favored by humans for visual identification and naming due to their optimal information density and distinctiveness.

Recent work investigates whether vision-language models (VLMs) acquire this humanlike preference. Using Llama 3.2 Vision Instruct (11B) and Molmo 7B-D, researchers prompted these models to label objects in images with both neutral and expert-style instructions, benchmarking model responses on the Ecoset dataset of 565 human-curated basic categories [2503.12530]. The basic-level usage rate (fraction of captions matching basic-level ground truth) significantly exceeded both the frequency expected by chance and minimal human estimates (e.g., 60.2% for Llama 3.2), with pronounced preference for biological categories. "Expert" prompting systematically shifted models toward subordinate labels, mirroring the “expert basic level shift” observed in human cognition.

These findings suggest that large VLMs not only reproduce basic-level labeling biases observed in human categorization, but also reflect nuanced effects such as greater basic-level salience for biological items and context-induced subordinate labeling. Such emergent cognitive organization in VLMs supports using these models as proxies for investigating organization principles of conceptual spaces [2503.12530].

## 2. Basic Questions in Visual Question Answering

In the context of visual question answering (VQA), “basic questions” (BQs) refer to minimal sub-questions that decompose a complex main question (MQ) into simpler components. The VQABQ framework addresses VQA using a two-stage system [1703.06492]:
- **BQ generation**: Given MQ, the system identifies up to three BQs by embedding MQ and a large candidate BQ pool into a shared vector space (using Skip-Thought encodings) and solving a LASSO problem
  $$
  \min_{x \in \mathbb{R}^N} \frac{1}{2}\|A x - b\|_2^2 + \lambda\|x\|_1
  $$
  where $A$ collects BQ embeddings, $b$ encodes MQ, and $\lambda$ controls sparsity.
- **Answering**: The main question and selected BQs are concatenated and processed using an alternating co-attention between image features and question text.

This BQ-driven approach achieves measurable improvements in open-ended VQA accuracy by supplying targeted sub-questions as auxiliary context for the model. The selection of BQs is governed by explicit threshold and ratio rules on LASSO weights, ensuring only informative BQs augment the MQ [1703.06492].

## 3. Basic Pregeometries and Coset Geometries

In incidence geometry, a **basic pregeometry** is a combinatorial structure that cannot be nontrivially decomposed by certain quotient operations. Formally, given a Buekenhout–Tits pregeometry $(\Gamma, G)$ with connected rank-2 truncations and $G$ acting transitively on elements of each type:
- $\Gamma$ is **$G$-primitive-basic** if every $G$-invariant type-refining partition yields only degenerate (trivial) quotients.
- $\Gamma$ is **$G$-normal-basic** if every proper normal quotient by a normal subgroup of $G$ is degenerate.

Reduction theorems [1009.0075, 1010.0481] show that any incidence-transitive pregeometry can be decomposed into direct sums of basic pregeometries where $G$ is primitive or quasiprimitive on each type. These structures serve as the atomic building blocks: all other geometries in the class arise through direct sums or successive quotients. Recent constructions demonstrate the existence of infinite families of thick, high-rank basic coset geometries for each O'Nan–Scott primitive group type, demonstrating their fundamental role in the landscape of permutation group actions [1010.0481].

## 4. Basic Properties of Algebraic Structures: Orders and Varieties

In algebra, "basic" can denote multiple structural primitives:

- **Basic quaternion orders**: An order $\mathcal{O}$ in a quaternion algebra over a Dedekind domain $R$ is basic if it contains an integrally closed quadratic $R$-order. The property of being basic is local: $\mathcal{O}$ is basic if and only if its localization at every nonzero prime is basic [1903.00560]. The main theorem establishes equivalence between being basic and being Bass (i.e., every $R$-superorder is Gorenstein), both locally (DVR case) and globally.
  
- **Basic superranks of varieties of algebras**: For a variety $V$ of algebras, the **basic superrank** is the minimal pair $(r, s)$ (even, odd generators) such that the variety generated by the free $V$-superalgebra on $(r, s)$ generators exhausts all identities of $V$. Classic results show that nearly associative varieties can have infinite basic rank, but the use of superalgebras often restores finite basic superrank. For example, the metabelian alternative algebra Alt(2) has basic spectrum $\{(1,1),(0,3)\}$ [1508.05956].

## 5. Basic Objects in Foliation Theory: Basic Forms, Cohomology, and Invariants

In differential geometry, the term "basic" designates differential forms and invariants that are transverse (i.e., invariant under leafwise flows) to a foliation:

- **Basic forms**: On a foliated manifold $(M, \mathcal{F})$, $\omega$ is basic if $i_X\omega=0$ and $L_X\omega=0$ for all leafwise vector fields $X$. The complex of basic forms provides the basic cohomology ring $H^*_\mathrm{basic}(M, \mathcal{F})$ [1004.1043].
- **Equivariant basic cohomology**: If the foliation is Riemannian (i.e., the normal bundle admits a transverse metric) and admits a global abelian structural Killing algebra $\mathfrak{a}$ (Molino’s theory), basic cohomology admits a Cartan model analogous to equivariant de Rham theory, with a resulting localization theory [1004.1043, 1011.4826].
- **Basic characteristic classes and numbers**: For a Riemannian foliation, basic characteristic classes are constructed using basic connections on associated principal bundles. The basic characteristic numbers, obtained by integrating top-degree basic characteristic classes, are global invariants determined solely by the infinitesimal structure at closed leaves. These numbers admit explicit formulas (ABBV-type localization) involving equivariant basic Euler classes at closed leaves [1011.4826].
- **Basic Dolbeault cohomology**: For transversely Kähler foliations (notably, Sasakian manifolds), one defines basic Dolbeault groups $H_B^{p,q}$ and associated Hodge numbers. Rigidity theorems show that basic Hodge numbers are invariant under certain deformations and vanish off the diagonal when the foliation has finitely many closed leaves, mirroring properties in classical complex geometry [1206.2803].

## 6. The “Basic” Principle in Model Distillation and Representation Learning

The notion of basic behaviors or competencies provides a template for improving the sample efficiency and generality of small models in machine learning. In Basic Reading Distillation (BRD), a small language model undergoes “basic reading education” by learning to emulate large LLMs at generic, non-task-specific behaviors: named entity recognition (NER) and question raising/answering about arbitrary text [2507.19741]. Distillation targets not only output alignment but the internal probability distributions associated with “basic” textual comprehension skills. BRD is provably orthogonal to knowledge or task distillation and leads to striking performance gains in zero- and few-shot settings, closing much of the gap between a 564M-parameter student and a 13B-parameter teacher model. These results position “basic” skill distillation as a formal mechanism for bootstrapping broad competencies in efficient neural architectures.

## 7. Concluding Remarks

Across mathematics, cognitive science, and artificial intelligence, the “basic” concept consistently identifies a privileged structural level—be it a category in human naming and visual grounding, an irreducible geometric or algebraic object, a formative behavioral competency for neural models, or canonical forms and invariants in foliated topology. In each domain, basic structures serve as the minimal, nonredundant units for decomposition, expressive power, or theoretical reduction. Ongoing research continues to elucidate how these basic layers are internalized, represented, and manipulated in both natural and artificial systems, serving as essential building blocks for higher-level organization and reasoning [2503.12530, 1703.06492, 1009.0075, 1903.00560, 1004.1043, 1508.05956, 1011.4826, 1206.2803, 2507.19741].

Source: https://www.emergentmind.com/topics/basic