Mathematics as the language of model design
Modified
Input
Mathematics makes model assumptions visible, comparable, and open to criticism. The essential skill is not symbolic complexity but translation between a real problem and a formal representation. AI students need mathematics to design and diagnose models, not merely to reproduce derivations.
Mathematics Is Not Decoration
AI discussions often treat mathematics as a barrier placed in front of an otherwise practical subject. Under that interpretation, students tolerate equations until they can reach the code. Mathematics becomes a credential, a source of difficult examinations, or a formal explanation added after an algorithm has already been chosen.
SIAI GSB takes the opposite view. Mathematics is the working language through which a modeller states what a system contains, how its components interact, what remains uncertain, and which conclusions follow from the assumptions.
Consider the statement:
Customer demand rises with income but eventually saturates.
The sentence already contains a model. It proposes a direction, nonlinearity, and a limiting behavior. One mathematical representation is
where $K$ is the saturation level, $a$ the rate of transition, and $b$ the income level around which demand changes most rapidly.
The equation does not make the statement true. It makes the statement precise enough to test, modify, and reject.
Model Design as Translation
A real system does not arrive with variables and equations attached. The modeller constructs them.
Table 1. How mathematical choices encode model assumptions
| Real-world question | Mathematical choice | Hidden commitment |
|---|---|---|
| What is being explained? | Outcome $Y$ | A measurement and unit |
| What information is available? | Inputs $X=(X_1,\ldots,X_p)$ | A boundary around the system |
| How are inputs related to outcomes? | Function $f$ | A functional form or model class |
| What remains unexplained? | Error $\varepsilon$ | A theory of uncertainty |
| What is learned from data? | Parameter $\theta$ or function $\widehat f$ | An estimation rule |
| What counts as success? | Loss $\ell$ or utility $U$ | A decision criterion |
The core translation can be written as
where $\mathcal F$ is the admissible model class, $P$ a probability structure, and $\mathcal A$ a method of calculation.
A student who can manipulate an equation but cannot construct this translation has mathematical technique without model-design ability. A student who can construct it can often learn the required technique as the problem demands.
Variables Define the World of the Model
The first mathematical decision is what to represent.
Suppose an institution wishes to estimate student learning. A convenient outcome is the final examination score:
Convenience does not establish validity. The score may combine memory, reasoning, language, time pressure, and prior exposure. If the intended object is independent model judgment, $Y_i$ may be a noisy proxy.
One can write
where $J_i$ is judgment, $K_i$ knowledge, and $M_i$ memory. The decomposition immediately creates questions:
- Can the three components be observed separately?
- Does the examination weight them as intended?
- Does time pressure increase $\gamma$?
- Would a different assessment change the meaning of $Y_i$?
Mathematics has exposed an educational-design problem that a raw score conceals.
Functions Encode Structure
The choice of a function is a claim about how the world behaves.
A linear model,
imposes a constant marginal effect:
If the effect varies with $X$, the linear representation may be inadequate. A quadratic model gives
The additional term is not simply greater complexity. It changes the substantive claim: the effect now depends on the state of the system.
Model design therefore requires students to read functions semantically. They should see a derivative as a marginal response, an interaction as conditional dependence, a constraint as institutional feasibility, and a probability distribution as a statement about uncertainty.
Linear Algebra Is the Language of Representation
AI models frequently transform a high-dimensional observation into a different coordinate system.
Let $x\in\mathbb{R}^{p}$ contain measured features. A linear representation is
where $W\in\mathbb{R}^{k\times p}$ maps the original variables into $k$ constructed features.
This simple expression supports several interpretations:
- $W$ may rotate the coordinate system;
- it may compress $p$ variables into $k<p$ dimensions;
- it may emphasize directions useful for prediction;
- it may approximate latent factors; or
- it may separate signal from numerical redundancy.
A neural-network layer extends the same idea:
where $\sigma$ permits nonlinear representation.
Students who view $W$ only as “weights learned by code” miss the model-design question. What information can this transformation preserve? Which distinctions become easier to express? Which information is discarded? Is the representation identifiable, stable, and relevant to the decision?
Probability Describes What the Model Does Not Know
Probability is sometimes taught as a catalogue of distributions. In model design, it is a language for incomplete information.
The expression
states that $f(X)$ captures the systematic conditional mean and that the remaining error has no conditional direction. If instead
the model has left systematic information in the error. This may result from omitted variables, selection, simultaneity, or an inappropriate functional form.
The assumption is not a technical footnote. It distinguishes a meaningful conditional model from a calculation whose coefficient may absorb uncontrolled structure.
Similarly, specifying
is a claim that the conditional mean and variance share a particular relationship. If the observed process is overdispersed or generated by heterogeneous subgroups, the distribution may be inappropriate even if software returns an estimate.
Optimization Defines What the Model Values
Every trained AI model answers an optimization question. The objective function states what the system is rewarded for doing.
The equation contains three institutional choices:
- $\ell$ defines the cost of error;
- $\Omega$ defines which forms of complexity are discouraged;
- $\lambda$ defines the exchange rate between fit and restraint.
Changing squared error to absolute error changes sensitivity to outliers. Weighting observations changes whose errors matter. Adding a fairness or capacity constraint changes which solutions are feasible. Optimization is therefore not a neutral computational step. It formalizes priorities.
Sensitivity Makes Assumptions Auditable
A model should not be judged only at its preferred specification. It should be perturbed.
Let an output be $q=g(\theta)$. A first-order approximation gives
Large derivatives indicate that small changes in assumptions or estimates can produce large changes in the conclusion.
For a linear system $Ax=b$, numerical sensitivity is summarized by the condition number
When $\kappa(A)$ is large, small changes in $A$ or $b$ may cause large changes in $x$. The calculation can be formally correct and institutionally unstable.
Sensitivity analysis turns the question “Is the model correct?” into the more useful question “Which parts of the conclusion survive plausible changes?”
Three Levels of Mathematical Competence
Read as a diagnostic rather than a second scorecard, the framework becomes clear: Symbolic execution — Student behavior: Substitutes values and reproduces derivations; Limitation: Cannot judge whether the equation belongs to the problem. Structural interpretation — Student behavior: Explains variables, assumptions, and functional relationships; Limitation: May not yet redesign the model. Model design — Student behavior: Translates a new problem, compares representations, and tests sensitivity; Limitation: Requires integrated mathematical and domain judgment.
Advanced AI education should move students through all three levels. Symbolic fluency matters because careless manipulation can invalidate a result. But the educational endpoint is the third level.
This is also why very advanced mathematics is not always the central difficulty. A model may use elementary linear algebra and probability while demanding sophisticated judgment about representation. Conversely, a technically difficult derivation may be educationally shallow if students merely reproduce it.
Constraints Are Part of the Model
Real decisions occur inside constraints. A portfolio may have a budget, a staffing policy may have capacity, and a medical allocation may have fairness and safety requirements.
A constrained decision problem is
subject to
Ignoring the constraints and optimizing prediction first can produce an answer that cannot be used. For example, ranking every candidate by expected benefit does not solve allocation when interventions differ in cost and only a fixed capacity is available.
The Lagrangian,
shows that each binding constraint has a shadow value. This gives institutional meaning to the mathematics: $\lambda_j$ describes how much the objective could improve if constraint $j$ were relaxed marginally.
Students who learn optimization only as an algorithm for finding minima miss this connection between formal constraints and organizational scarcity.
Applied Example: Designing a Retention Model
Suppose a school wants to identify students at risk of non-completion. A superficial design begins with a binary outcome:
But the institution must decide what it intends to do. If the action is academic support, the relevant target may be the probability of completion under support relative to completion without it:
A student with high predicted risk may have little treatment benefit. Another with moderate risk may respond strongly. Changing the decision changes the mathematical target from risk prediction to heterogeneous intervention effect.
Capacity adds a constraint:
If intervention costs vary, the constraint becomes
The model-design sequence is now visible:
- define completion and the intervention;
- distinguish risk from benefit;
- estimate uncertainty;
- encode capacity;
- select students to maximize expected educational value; and
- monitor whether intervention changes the future DGP.
No single algorithm determines this architecture. Mathematics connects the institutional question to a system whose assumptions can be debated.
Mathematics as Shared Institutional Memory
Formal notation also preserves reasoning across people and time. A statement such as “the model performed well” is difficult to audit later. The definition
records that validation left out entire campuses rather than random students. A future reviewer can understand which form of generalization was tested.
Mathematics does not eliminate ambiguity, but it reduces the amount that must be recovered from institutional memory. It provides a compact contract among researchers, engineers, instructors, and decision-makers.
The contemporary educational case for this language is becoming stronger. SIAI’s review of agentic machine science shows that machines already excel where objectives are explicit and answers mechanically checkable, but remain unreliable where the model must decide what counts as evidence. The Economy’s analysis of education beyond procedure therefore argues for assessment built around framing, assumptions, and defense rather than execution alone. SIAI’s work on the cognitive environment adds that habits of reasoning are shaped by repeated institutional practice. Mathematics matters because it makes those habits visible and contestable.
Conclusion
Mathematics is the language of model design because it forces choices that ordinary prose can conceal. It defines variables, relations, uncertainty, objectives, and constraints. It also permits a claim to be derived, challenged, and revised.
AI students do not need mathematics merely to understand how an existing algorithm works. They need it to decide which algorithm, representation, and evaluation belong to a problem in the first place.
The distinction changes teaching. Instead of asking students only to prove a result or calculate an answer, an advanced course should ask them to translate a situation, state the assumptions, compare formulations, and explain how the conclusion changes when the mathematical structure changes.
That is the point at which mathematics stops being a prerequisite and becomes the working language of AI.
References
George E. P. Box, “Science and Statistics”, Journal of the American Statistical Association 71, no. 356 (1976): 791-799.
SIAM Working Group on CSE Education, “Graduate Education in Computational Science and Engineering”, SIAM Review 43, no. 1 (2001): 163-177.
Gilbert Strang, Linear Algebra and Learning from Data, Wellesley-Cambridge Press, 2019.
Swiss Institute of Artificial Intelligence (2026) ‘What Agentic AI Can Prove and What It Still Fabricates: The Real Frontier of Machine Science’, SIAI Science Review, 15 August.
The Economy (2025) ‘Redesigning Education Beyond Procedure in the Age of AI’, The Economy Review, 17 September.
Swiss Institute of Artificial Intelligence (2026) ‘The Cognitive Environment: How Place and Education Shape the Habit of Reasoning’, SIAI Science Review, 12 June.
Similar Post
