Artificial Intelligence (AI) can be viewed not just as a computational tool, but as an emerging branch of mathematics in its own right. Modern AI, especially machine learning, relies on deep theoretical ideas from linear algebra, calculus, probability, and more – and it increasingly contributes back to mathematical research by discovering patterns and suggesting conjectures. In this extended discussion, we explore how AI connects with various mathematical disciplines and speculate on its potential role as a unifying force in mathematics. We draw parallels to historical unifications (like the Langlands Program, Fourier analysis, and Gödel’s encoding of logic in arithmetic) to illustrate how AI might play a similar unifying role.
AI and Number Theory
Pattern discovery in number theory: Number theory has a rich history of patterns (for example, in prime numbers or in values of L-functions) that often elude straightforward analysis. AI algorithms excel at detecting subtle structures in large data sets, and researchers are leveraging this to tackle number-theoretic problems. A notable example is the “Ramanujan Machine,” an AI system designed to discover new conjectures about fundamental constants and number relationships. This system automatically searches for formulas – for instance, new infinite series or continued fractions for π or e – that appear true but have no proof yet (New AI 'Ramanujan Machine' uncovers hidden patterns in numbers | Live Science). In essence, the Ramanujan Machine reveals hidden relationships between numbers and outputs conjectures that mathematicians can then attempt to prove (New AI 'Ramanujan Machine' uncovers hidden patterns in numbers | Live Science). Such AI-driven conjecture generation extends the experimental tradition in number theory, where patterns precede proofs.
Automorphic forms and L-functions: The Langlands Program famously links number theory and harmonic analysis, predicting deep connections between prime number-based $L$-functions and automorphic forms. AI could become a valuable ally in this domain by sifting through vast numerical data for these objects.
For example, the L-Functions and Modular Forms Database (LMFDB) contains millions of entries of modular forms, elliptic curves, and L-functions. Machine learning can be applied to this data to guess unknown correspondences or identify outliers. Indeed, recent work trained neural networks on data from elliptic curves and modular forms to predict properties like the vanishing order of L-functions (Machine learning the vanishing order of rational -functions) (Machine learning the vanishing order of rational -functions). This kind of analysis has led to the discovery of unexpected “murmurations” patterns in the statistics of elliptic curves: when plotting aggregate data related to elliptic curve invariants, researchers noticed flock-like shapes (hence the name “murmurations,” after flocking starlings) (Elliptic Curve ‘Murmurations’ Found With AI Take Flight | Quanta Magazine). Initially found through computational experiments, these patterns have now been proven to occur generally, not just in the examples where AI first spotted them (Elliptic Curve ‘Murmurations’ Found With AI Take Flight | Quanta Magazine). Even more impressively, similar patterns were found in other number theoretic objects – modular forms and Dirichlet characters (which are intimately connected to prime number distributions via L-functions) – and mathematicians proved these “murmurations” phenomena occur there as well (Elliptic Curve ‘Murmurations’ Found With AI Take Flight | Quanta Magazine). This demonstrates AI’s capacity to uncover structure in number theory, such as subtle regularities in the distribution of primes or zeros of L-functions, which might be beyond immediate human detection.
Primes and prediction: Although the distribution of prime numbers is famously irregular, AI has been used in exploratory ways to study primes. Researchers have experimented with neural networks to distinguish primes from composites or to guess the location of large primes, with mixed results. More promising is using AI to study statistical distributions related to primes (for instance, in the context of the Riemann zeta function’s zeros). While no AI has cracked the Riemann Hypothesis or revealed a new formula for primes yet, these efforts might yield conjectures about prime gaps or patterns in prime factorization. In probabilistic number theory, machine learning has even helped derive known theorems: one team used a learning approach with maximum entropy methods to re-derive results like the Hardy–Ramanujan theorem on the normal distribution of prime factors ([2403.12588] Machine Learning of the Prime Distribution). This suggests that AI can act as a kind of experimental mathematician in number theory – proposing phenomena that mathematicians can then attempt to explain or prove.
AI and Algebraic Geometry
Identifying patterns in geometry: Algebraic geometry deals with highly complex structures (varieties, sheaves, moduli spaces) where patterns are often subtle and high-dimensional. AI offers new ways to recognize and analyze these patterns. One active area is using machine learning to study Calabi–Yau manifolds – intricate geometric objects important in string theory. Calabi–Yau spaces have many continuous parameters (forming a moduli space), and computing certain properties (like Ricci-flat metrics) is notoriously hard. Researchers have successfully trained AI to predict properties of Calabi–Yau manifolds from limited data. For example, machine learning was applied to find numerical Calabi–Yau metrics much faster than traditional algorithms, achieving accurate approximations of the metric (and related geometric quantities) with far fewer sample points ([1910.08605] Machine learning Calabi-Yau metrics). By learning from a small set of exact calculations, the AI was able to generalize and estimate the solution across the entire manifold, speeding up computations by orders of magnitude ([1910.08605] Machine learning Calabi-Yau metrics). This is a striking case of AI discerning the underlying geometric pattern (the shape of the metric) in a complicated space of possibilities.
Sheaves and moduli spaces: The classification of algebraic varieties often boils down to understanding their moduli spaces and invariants (such as certain cohomology classes or sheaf counts). AI can assist by clustering or visualizing high-dimensional moduli data. For instance, there are millions of distinct Calabi–Yau threefolds known in databases; machine learning has been used to classify these manifolds by their properties (Elliptic Curve ‘Murmurations’ Found With AI Take Flight | Quanta Magazine). In one project, a neural network was trained to recognize if two Calabi–Yau manifolds were in the same family or to predict topological invariants of a given manifold. Yang-Hui He and collaborators demonstrated such techniques, effectively letting AI sort through the “landscape” of possible geometries (Elliptic Curve ‘Murmurations’ Found With AI Take Flight | Quanta Magazine). This kind of pattern recognition could extend to vector bundles or sheaves on varieties – for example, AI might conjecture relationships between certain bundle stability conditions and easy-to-compute invariants by observing enough examples.
Crossover with AI architectures: Interestingly, the flow of ideas goes both ways – concepts from algebraic geometry are illuminating the structure of AI models. A recent line of work represents neural network computations in geometric terms. For example, a deep neural network with piecewise-linear activations can be described using tropical geometry (a variant of algebraic geometry over the min-plus algebra). In fact, researchers have shown that the decision boundary of a neural network can be characterized as part of a tropical hypersurface of a corresponding polynomial (Machine Learning Needs a Langlands Programme) (Machine Learning Needs a Langlands Programme). In simpler terms, the complex partition of space created by a trained neural net has a rigorous geometric description. This connection has practical value: by translating neural networks into geometric objects, one can apply mathematical tools to compare and simplify networks (Machine Learning Needs a Langlands Programme). It’s a beautiful example of AI and pure mathematics enriching each other – AI inspires new questions in algebraic geometry (like understanding the space of tropical polynomials arising from network architectures), and algebraic geometry provides a new language to understand AI.
AI and Harmonic Analysis
Spectral perspectives on AI: Harmonic analysis – the field centered on representing functions as superpositions of basic waves (Fourier series, for example) – has classic roots but modern relevance for AI. Deep neural networks can be thought of as complicated functions, and researchers are using harmonic analysis to probe the behavior of these AI models. A 200-year-old tool like Fourier analysis can reveal what patterns a neural network has learned. In one study, scientists applied Fourier transforms to a trained deep network for fluid dynamics and discovered they could identify which scales and frequency components the network was focusing on (Scientific AI’s ‘black box’ is no match for 200-year-old method | Rice News | News and Media Relations | Rice University) (Scientific AI’s ‘black box’ is no match for 200-year-old method | Rice News | News and Media Relations | Rice University). This is significant because it connects the “black box” of a neural network with the classical language of spectra: essentially, they found that certain frequencies of the input (analogous to certain “notes” in a Fourier decomposition) were amplified by the network, corresponding to physical modes of the system (Scientific AI’s ‘black box’ is no match for 200-year-old method | Rice News | News and Media Relations | Rice University). Fourier analysis thus becomes a bridge between the empirical training of AI and a theoretical understanding, showing how AI might be viewed through the lens of functional analysis and signal processing.
Neural networks as universal approximators: One of the foundational theoretical results in AI is that neural networks are universal function approximators – much like Fourier series can approximate any reasonable periodic function. This hints at a unification: a neural network can be seen as a flexible basis expansion, not of sines and cosines, but of learned motifs. In practice, networks often rediscover Fourier-like representations when needed (for example, learning to detect edges or periodic patterns in data). This has led to hybrid techniques where neural networks incorporate Fourier features explicitly for better learning of frequency-rich patterns (used in audio, image processing, etc.). The synergy between neural nets and harmonic analysis goes deeper in representation theory as well. Many advanced neural architectures (like convolutional networks or transformers) exploit symmetry and invariance, concepts native to representation theory. One can view a convolution as averaging a function over a group action (a very harmonic analysis idea), and thus convolutional networks connect to the representation theory of translation groups.
AI in representation theory: Representation theory (a core part of modern harmonic analysis) studies how abstract algebraic groups manifest as symmetries of spaces. AI can assist by handling large combinatorial data from group representations that humans find cumbersome. In a recent breakthrough, an AI system guided mathematicians to a result in knot theory and also to progress on a long-standing problem in representation theory. Specifically, DeepMind’s collaboration led to a proposed algorithm related to the combinatorial invariance conjecture for symmetric groups (Advancing mathematics by guiding human intuition with AI - PubMed). The symmetric group (which permutes elements) has complex representation-theoretic invariants; the AI analyzed data and predicted a pattern – effectively a conjectured solution algorithm – for this representation theory problem (Advancing mathematics by guiding human intuition with AI - PubMed). This example shows AI acting almost like a theoretician in harmonic analysis, suggesting how an abstract algebraic structure might behave.
It’s conceivable that AI could similarly tackle problems like classifying automorphic representations or identifying spectral signatures of different types of mathematical objects. In the Langlands Program, automorphic forms on a Lie group correspond to Galois representations; an AI could potentially match these by comparing spectral data (eigenvalues of operators) with arithmetic data, automating parts of the discovery. While this is speculative, it aligns with AI’s strength in pattern matching: the “music” of arithmetic (to use Langlands philosopher-poet analogy) might be recognized by a machine and linked to the right “lyrics” in geometry.
AI and Logic & Formal Mathematics
Automated theorem proving: Logic and formal mathematics have long been intertwined with computer science and AI. Since Gödel, we know not all mathematical truths can be formally derived, but a great many can – and AI is increasingly powerful in this domain.
Automated Theorem Provers (ATPs) have achieved notable successes. One famous milestone was the proof of Robbins’ conjecture (an algebraic logic problem) by the EQP theorem prover in 1996, solving a conjecture about Boolean algebra identities that had been open for decades (Josef Urban on Machine Learning and Automated Reasoning - Machine Intelligence Research Institute). That proof, found after days of automated search, was something no human had achieved since the problem’s formulation in 1933. This success showed that computers could navigate enormous logical search spaces to find valid proofs. Modern AI builds on these ATP methods by adding learning.
For example, the system GPT-f (an AI model based on the GPT language model) was used as a proof assistant in the Metamath formal system. It managed to find new, shorter proofs of theorems, some of which were accepted into the Metamath library ([2009.03393] Generative Language Modeling for Automated Theorem Proving). This marked the first time a machine-learning-based system contributed original proofs that human mathematicians adopted as improvements ([2009.03393] Generative Language Modeling for Automated Theorem Proving). AI can thus assist in formal proof checking and even in discovering proofs, effectively automating parts of logical reasoning.
Assisting with undecidability and complex proofs: Certain statements in mathematics are independent of given axioms (Gödel’s incompleteness taught us that), but within a strong enough system, many deep statements are still either true or false – just extremely hard to prove. AI might help explore such statements by searching for proofs or counterexamples. For example, combinatorial problems that border on undecidability, like finding a certain combinatorial design or solving a large puzzle, can sometimes be translated into satisfiability problems and cracked by SAT solvers or AI search (as was done for the Boolean Pythagorean Triples problem and others). We’re also seeing AI integrate with interactive theorem provers like Lean, Coq, and Isabelle. In these systems, a human guides a proof but an AI can suggest the next step or fill in a straightforward gap.
Over time, as AI improves, it could handle more of the tedious logical bookkeeping, letting mathematicians focus on high-level ideas. This human-AI collaboration might tackle problems that are currently impractical, such as verifying every step of a classification theorem or searching huge game trees of possible lemmas. There’s even hope that AI could help with meta-mathematical questions: for instance, by experimenting with different axiom systems or searching for a proof that a certain proposition is independent of Zermelo–Fraenkel set theory, etc. While those are grand challenges, the trend is clear – AI is becoming a versatile tool in formal logic and proof, augmenting our ability to navigate complex logical deductions.
Neural-symbolic integration: A particularly exciting development is the blending of neural (learning-based) and symbolic (logic-based) approaches. Each has complementary strengths: neural networks excel at intuition and pattern recognition; symbolic systems excel at exact reasoning. Researchers like Josef Urban have pioneered systems that use a feedback loop between learning and proving – an AI will learn from past successful proofs to guide the search for new ones (Josef Urban on Machine Learning and Automated Reasoning - Machine Intelligence Research Institute). In one approach, the AI tries many possible proof steps (deductive search), learns from those that lead to progress, and then uses that experience to prune the search space next time.
This loop of “experience-guided reasoning” starts to resemble how a human mathematician accumulates intuition. Over large formal libraries (like the archive of thousands of proven theorems in mathematics), such a system can gradually improve, potentially automating algebraic logic tasks that currently require expert insight. In the long run, neural-symbolic AI might address problems that are at the boundary of human capability, either finding proofs that are too long or complex for us, or providing insight into why certain problems (like the halting problem or continuum hypothesis) are undecidable by exploring many models or interpretations.
AI in Computation and Deep Learning for Mathematics
Experimental mathematics 2.0: Mathematics has a tradition of conjecture and experiment – from numerical evidence for the prime number theorem to computer-generated insights like the Birch–Swinnerton-Dyer conjecture. AI supercharges this approach. With deep learning and other techniques, AI can trawl through huge computational datasets and highlight patterns that deserve attention.
We’ve already discussed several instances (elliptic curve murmurations, symmetric group invariants, Ramanujan Machine conjectures). What’s notable is how AI not only finds patterns but also suggests mechanisms or structures underlying them. In the work by DeepMind and collaborators, the process was: use machine learning to find a pattern, use AI’s interpretability tools (attribution of which features matter) to guess a conjecture, then have humans prove or refine it (Advancing mathematics by guiding human intuition with AI - PubMed).
This ML-guided discovery framework essentially makes AI a partner in research, where it proposes hypotheses that humans might not have considered (Advancing mathematics by guiding human intuition with AI - PubMed). As computational power grows, we can envision AI doing massive experiments in areas like combinatorics, topology, or even category theory – generating data far beyond what a lone researcher could, and sifting that data for promising conjectural relationships.
Neural networks as mathematical objects: Deep learning models themselves have become objects of study for mathematicians. The weights of a neural network, the structure of its layers, and the training dynamics all constitute a new kind of high-dimensional, non-linear mathematical system. Studying these has led to insights like the neural tangent kernel (which connects training dynamics of infinitely wide networks to kernel methods in analysis) and Gaussian process interpretations of networks.
These theoretical bridges show that AI has a foot in classical mathematics. For example, a neural network in the limit of infinite width can be seen as solving a kernel regression problem – linking it to well-understood analytic theory. Additionally, gradient descent (the workhorse algorithm for training networks) can be analyzed with dynamical systems and variational calculus tools.
By examining AI through a mathematical lens, researchers not only demystify AI (“opening the black box”) but also enrich mathematics with new questions – such as the geometry of loss surfaces (analyzing high-dimensional real functions for critical points) or the algebra of weight symmetries (identifying when two different networks are functionally the same). In this sense, AI becomes a field of mathematical inquiry: understanding the space of learnable functions and the algorithms that navigate that space might one day be as fundamental as understanding solvable equations in calculus.
Neural and symbolic integration for insight: Another promising direction is integrating neural networks with symbolic computation systems. We already see tools like deep learning models conjecturing formulas that are then verified by computer algebra, or neural nets guiding the simplification of algebraic expressions. Conversely, symbolic systems can enforce constraints or exactness in neural models (for example, training a network that always outputs a symmetric polynomial, by building symmetry into the architecture). This hybrid approach could extend mathematical insight by ensuring that AI’s guesses are consistent with known theory.
One concrete example is in integrals and differential equations: a neural network might guess the form of a solution, and a CAS (Computer Algebra System) can then check that guess or refine it to an exact solution. Such neural-symbolic loops might tackle problems that were previously intractable – for instance, conjecturing a formula for the coefficients of a series solution to an unsolved differential equation, or predicting the behavior of a chaotic system and then proving those predictions with rigorous math. Over time, this blurs the line between computation and theory, making AI an indispensable part of the mathematician’s toolkit.
AI as a Unifying Force in Mathematics
Beyond tackling specific problems, AI has the potential to unify disparate areas of mathematics by acting as a Rosetta stone or intermediary – much like historical frameworks that brought unity to math. It’s worth drawing comparisons to see how AI might play a similar role:
Langlands Program: The Langlands program creates dictionaries between number theory and geometry/analysis, so that truths in one domain translate to truths in another. AI could become a practical Langlands-like bridge by finding empirical correspondences. For instance, if a number theoretic pattern (say in prime distributions or $L$-function zeros) has an analog in a geometric setting (like spectra of a differential operator), an AI searching both spaces might notice the match. Langlands is often called a “Rosetta Stone” (Machine Learning Needs a Langlands Programme) because it connects different mathematical languages; AI, by handling huge datasets across domains, could expand this translation. We might one day see an AI that takes an unsolved problem in arithmetic and transforms it (via learned correspondences) into a solvable problem in, say, topology – effectively automating the insight that Langlands and others provided by hand. Even within machine learning theory itself, scholars have called for a “Langlands program for Machine Learning” to unify its many subfields (Machine Learning Needs a Langlands Programme) (Machine Learning Needs a Langlands Programme). This means AI folks recognize the need for unifying principles, and mathematics might supply them (e.g., category theory to relate architectures, or symmetries to relate different learning problems). In turn, AI’s unifying power in math could lie in its ability to detect when two problems from different branches are “the same” in structure.
Fourier Analysis: Fourier’s insight that any reasonable function can be decomposed into trigonometric series was revolutionary – it unified seemingly arbitrary functions under a common basis and unlocked solutions to differential equations. Neural networks have a similar universal approximation property, suggesting a unification of function spaces via learnable features. One could say that AI (deep learning) generalizes Fourier analysis: rather than a fixed set of basis functions, we let the network learn its own basis from data. In doing so, AI might unify the study of different function families by showing they are all reachable through one framework (a sufficiently deep network). Practically, we see Fourier analysis and AI blending to solve problems – like using convolutional neural nets (which implicitly use translation symmetry and thus a Fourier basis) for image and signal processing, traditionally the domain of Fourier transforms. The synergy is so deep that one can even implement a Fourier transform as a one-layer neural network (The Fourier transform is a neural network | sidsite), and conversely, use Fourier transforms inside a neural network to improve learning (Fourier Transform in Neural Networks ??!! | FNet – Weights & Biases). This interplay hints that the gap between “analytical solutions” and “AI solutions” to problems could be bridged – a unification where a problem might be attacked either by classic math (Fourier series, etc.) or by training a network, whichever is more convenient, without seeing them as fundamentally different approaches.
Gödel’s Arithmetic Encoding: Gödel astonished the world by showing how to encode statements about logic as numeric statements – an unexpected unification of logic and number theory. This encoding enabled proofs about the unprovability of certain claims. In a loose analogy, modern AI often encodes logical or discrete structures into continuous numeric form (think of word embeddings or graph neural networks encoding combinatorial graphs into vectors). By embedding structured mathematical objects into continuous spaces, AI provides a kind of universal coding scheme across mathematics. For example, a complex algebraic structure (like a group or a ring) might be represented by a learned vector in a neural network, allowing the network to reason about it. We see this in neural theorem provers that turn formal statements into sequences of symbols (numbers) that a language model can process. In effect, AI encodes semantics into arithmetic – not by a fixed Gödel numbering, but by learning an encoding that preserves the essence of the problem. This might unify mathematics by enabling cross-talk between discrete symbolic math and continuous analytic math. A topology problem could be turned into a data-driven task for a neural network, or a logical theory could be associated with a geometry in high-dimensional space (in the way word vectors place “king - man + woman ≈ queen”, one could imagine encoding mathematical concepts so that algebraic relationships correspond to geometric relations in the embedding). Such embeddings, if understood, could reveal deep connections – essentially, an AI might “discover” a Gödel-like encoding that links two fields, just as Gödel manually linked logic and arithmetic.
In all these ways, AI holds promise as a unifying force: it can translate problems between domains (like an oracle that recognizes the same pattern in two different guises), bring old theories under a common roof of computation, and connect the discrete with the continuous. It’s important to note that these are speculative and long-term potentials. However, the trajectory is reminiscent of how other unifying frameworks started – initially as bold conjectures or philosophical ideas, later becoming rigorous theory. For instance, category theory began as an abstract unifier of math structures and now underpins modern algebraic geometry and homotopy theory; one could envision a future “AI-theoretic” viewpoint that provides a high-level language for relating different mathematical disciplines.
Conclusion
AI’s integration into mathematics is still in its early stages, but it is accelerating. We have gone from using computers for brute-force calculation to using AI for insight and conjecture generation. This shifts AI from being an outsider to being part of the mathematical process – effectively, a new branch of mathematics that deals with algorithms that learn and reason.
In connecting with number theory, algebraic geometry, harmonic analysis, logic, and beyond, AI does more than solve individual problems; it creates bridges between areas. It enables a form of empirical metastructure for mathematics, where patterns in one field can be detected and related to patterns in another through data and learning.
Historical unification programs like Langlands or the use of Fourier analysis enriched mathematics by revealing hidden unity. AI has the potential to do the same on a grand scale, perhaps even faster. We can imagine a future where conjectures in topology are suggested by an analysis of geometric datasets, or where a number-theoretic hypothesis is reinforced by a machine’s pattern recognition that also finds an equivalent statement in a completely different field. AI might also unify the practice of mathematics: blending the experimental, the conjectural, and the formal into one workflow. In this sense, AI could serve as both a toolbox and a grand organizer, much like category theory or set theory provided common languages for disparate subfields.
The optimism should be tempered with realism – many deep mathematical insights require creativity and intuition that we don’t yet know how to replicate in machines. But as AI systems become more sophisticated, collaborative frameworks like the ones we discussed (human intuition guided by AI suggestions (Advancing mathematics by guiding human intuition with AI - PubMed)) could become commonplace. Mathematics has always advanced by finding connections and analogies; AI is a powerful new agent for discovering those connections.
By considering AI as a branch of mathematics, we invite a two-way exchange: using math to understand AI, and using AI to advance math. The outcome could be a more unified mathematical world view, where the boundaries between subfields fade and a network of interrelated truths emerges – with AI both navigating and strengthening that network, much as Euler or Fourier or Langlands did in earlier eras, but with the speed and scale of computation at its aid.
Comentarios