Chapter 6: Teaching as Geodesic Guidance

6.1.1 The Pedagogical Heuristic

Definition 6.1 (Pedagogical Heuristic). The pedagogical heuristic h_T: \mathcal{L} \to \mathbb{R}_{\geq 0} is the teacher’s estimate of the remaining educational cost from the student’s current state \mathbf{x} to the goal region G:

h_T(\mathbf{x}) = \text{teacher's estimate of } d_{\mathcal{L}}(\mathbf{x}, G)

where d_{\mathcal{L}}(\mathbf{x}, G) = \inf_{\mathbf{y} \in G} d_{\mathcal{L}}(\mathbf{x}, \mathbf{y}) is the geodesic distance from \mathbf{x} to the nearest point in G.

The pedagogical heuristic is not a number the teacher consciously computes. It is the formalization of what teaching intuition does: a skilled teacher looks at a student’s work and forms an estimate of how far the student is from understanding. “This student is almost there — they just need to see one more example.” That is h_T(\mathbf{x}) \approx \epsilon. “This student is fundamentally confused — we need to go back to prerequisites.” That is h_T(\mathbf{x}) \gg 0. The estimate may be accurate or inaccurate, but every teaching decision is implicitly a decision based on an estimate of the student’s distance from the goal.

6.1.2 Shulman’s Pedagogical Content Knowledge

Lee Shulman’s (1986) concept of pedagogical content knowledge (PCK) — the specialized knowledge of how to teach specific content to specific students — is precisely the calibration of the heuristic function for a specific domain and student population.

A teacher with strong PCK in calculus knows: - How far a typical student is from understanding limits after the first lecture (h_T for post-lecture states) - Which misconceptions create high curvature (e.g., the confusion between a limit and a function value — a high-curvature region where the metric is steep) - Which examples reduce the distance most efficiently (examples that transport the student along the geodesic) - Which common errors indicate which specific learning gaps (diagnostic: the error reveals the student’s position on the manifold)

PCK is domain-specific heuristic calibration. A teacher with strong PCK in calculus may have weak PCK in linear algebra — the heuristic is well-calibrated for one domain and poorly calibrated for another. This is a local property of the heuristic field: it is accurate in regions the teacher has mapped and inaccurate in unfamiliar territory.

6.2.1 Strictly Admissible: Master Teachers

Definition 6.2 (Admissible Pedagogical Heuristic). A pedagogical heuristic h_T is admissible if h_T(\mathbf{x}) \leq d_{\mathcal{L}}(\mathbf{x}, G) for all \mathbf{x} \in \mathcal{L}.

An admissible heuristic never overestimates the student’s distance from understanding. The teacher may think the student is further from understanding than they actually are (leading to extra practice, which is suboptimal but safe), but the teacher never thinks the student understands when they do not (which would cause the teacher to move on before the student is ready, potentially sending the student onto a path where the manifold is disconnected).

Theorem 6.1 (Pedagogical Optimality). If the pedagogical heuristic h_T is admissible, then A search on the learner manifold using h_T finds the geodesic \gamma^* — the minimum-cost path from the student’s current state to the goal region.*

This theorem is a direct instantiation of the A* optimality theorem from Geometric Reasoning (Ch. 3). Its educational content is: if the teacher’s estimate of student understanding is reliably conservative (never overestimates readiness), then the teaching path is optimal. Good teaching is admissible heuristic guidance.

Master teachers have strictly admissible heuristics. They reliably detect gaps in understanding. They are careful not to advance before the student is ready. They may occasionally provide more practice than strictly necessary (because the heuristic underestimates — the teacher thinks the student is further from understanding than they are), but they never skip essential steps. The result is a near-geodesic path through the manifold.

6.2.2 Epsilon-Admissible: Competent Teachers

An epsilon-admissible heuristic satisfies h_T(\mathbf{x}) \leq d_{\mathcal{L}}(\mathbf{x}, G) + \epsilon for some small \epsilon > 0. The teacher’s estimate may slightly overestimate the student’s distance from understanding. The resulting teaching path is within \epsilon of optimal: not perfect, but close.

Most competent teachers have epsilon-admissible heuristics. They occasionally misjudge a student’s readiness, moving on slightly too fast or providing slightly too little practice. The resulting teaching path is near-geodesic but not exactly geodesic. The epsilon is the gap between professional competence and pedagogical perfection.

6.2.3 Inadmissible: Miscalibrated Teachers

An inadmissible heuristic systematically overestimates the student’s readiness: h_T(\mathbf{x}) > d_{\mathcal{L}}(\mathbf{x}, G) + \epsilon for many states \mathbf{x}. The teacher thinks the student is closer to understanding than they actually are, and moves on before prerequisites are established.

Inadmissible heuristics produce non-geodesic paths. The student is pushed into regions of the manifold where the prerequisites have not been built, the metric is degenerate (the necessary tangent vectors are absent), and learning stalls. The student does not merely fail to learn; they arrive at a position on the manifold from which further progress requires backtracking to build the missing prerequisites — a longer and more costly path than the geodesic would have been.

Two common causes of inadmissible heuristics:

Pace pressure. Curriculum schedules that require covering a fixed amount of material in a fixed time create pressure to advance regardless of student readiness. The teacher’s heuristic becomes inadmissible not because the teacher misjudges the student but because the institution overrides the teacher’s judgment with a pace constraint. The schedule overestimates the student’s readiness by assumption: “Students should understand derivatives by week 6” is a heuristic value imposed externally, regardless of the actual student state.

Frustration deflation. Teachers who have experienced years of student underperformance may develop heuristics that underestimate student capacity — expecting less than students can achieve, providing scaffolding that is not needed, and accepting lower standards. This is the opposite of inadmissibility: the heuristic is too pessimistic, producing paths that are longer than the geodesic (too much review, too many prerequisites revisited). Frustration deflation is a suboptimal heuristic, but it is admissible and therefore safe — the student arrives at the goal, just more slowly than necessary.

6.2.4 Gauge-Variant: Biased Teachers

Definition 6.3 (Gauge-Variant Pedagogical Heuristic). A pedagogical heuristic h_T is gauge-variant if h_T(\mathbf{x}) depends on features of the student that are not coordinates on the learner manifold — features such as race, gender, accent, clothing, first language, family background, or physical appearance.

A gauge-variant heuristic assigns different distance estimates to students at the same point on the learner manifold, based on surface features that should not affect the assessment. If the teacher estimates that Alex (first-generation, working-class, Latinx) is further from understanding than Ethan (continuing-generation, upper-middle-class, white) when both are at the same learner state, the heuristic is gauge-variant.

Gauge variance in the pedagogical heuristic is not always conscious bias. It can arise from statistical conditioning: if the teacher has observed (correctly) that students from certain backgrounds tend to arrive with lower d_1, the teacher may apply this prior to new students from those backgrounds without checking whether the specific student fits the pattern. The prior shifts the heuristic estimate. The shift is gauge-variant because it depends on the student’s background (a non-manifold feature), not on the student’s current learner state (a manifold position).

The damage from gauge-variant heuristics is directional: they systematically underestimate the capacity of students from disadvantaged backgrounds and overestimate the capacity of students from advantaged backgrounds. The resulting teaching paths diverge: the underestimated student receives less challenge, less scaffolding into advanced material, and lower expectations — a path that stays in low-curvature regions of the manifold and reaches a goal region that is smaller and less ambitious. The overestimated student receives more challenge, more opportunity, and higher expectations — a path that explores more of the manifold and reaches a larger, more ambitious goal region.

6.3.1 When to Remove the Scaffold

The scaffold should be removed when the student can traverse the transition independently — that is, when the student’s position on the manifold has changed enough that the natural metric in the relevant neighborhood is low enough for independent traversal.

The timing of scaffold removal is a heuristic judgment: the teacher must estimate when the student’s natural metric has changed sufficiently. Removing the scaffold too early (the student is not yet ready for independent traversal) forces the student back into a high-curvature region without support. Removing the scaffold too late (the student has already mastered the transition) wastes time and can create dependence — the student relies on the scaffold instead of developing the natural metric.

6.4.1 The Cost of Absent Heuristics

The cost of navigating without a heuristic is well-quantified in search theory: A* with h = 0 degenerates to uniform-cost search (Dijkstra’s algorithm), which explores states in order of cost-from-start without any estimate of cost-to-goal. Uniform-cost search is complete (it finds the goal) but maximally inefficient (it explores every state up to the geodesic cost).

The educational analogue: a first-generation student without a heuristic field will eventually find the path to the degree (given sufficient time, resources, and persistence), but the path will be longer, more expensive, and involve more dead ends than the path of a student guided by a well-calibrated heuristic. The additional path cost is measurable: first-generation students take, on average, 1.5 to 2 semesters longer to complete a degree, change majors more frequently, and have higher rates of late withdrawal from courses — all indicators of exploratory search without heuristic guidance.

6.5.1 Heuristic Corruption

The teacher’s heuristic is corrupted by irrelevant information. In teaching, heuristic corruption occurs when the teacher’s estimate of student understanding is warped by factors unrelated to understanding: the student’s behavior (a quiet student is assumed not to understand), the student’s background (a student from a “low-performing” school is assumed to be behind), or the teacher’s own projections (a student who reminds the teacher of a previously unsuccessful student is assumed to be similarly at risk).

6.5.2 Objective Hijacking

The teacher optimizes for a proxy rather than the actual educational objective. The most common form: optimizing for test scores rather than for learning. A teacher who “teaches to the test” is following the gradient of the scalar projection (G) rather than the gradient of the geodesic. The teacher’s heuristic estimates distance-to-high-test-score rather than distance-to-understanding. The resulting path is \gamma_G^* rather than \gamma_\mathcal{L}^* — the GPA-optimal path rather than the geodesic.

6.5.3 Local Minima

The student gets stuck in a state from which the heuristic provides no productive direction. Educational local minima include: prerequisite traps (the student lacks a prerequisite but the system does not offer remediation), motivation valleys (the student has lost interest and the curriculum does not recover it), and skill plateaus (the student has automatized an incorrect procedure and the correct procedure requires unlearning).

6.5.4 Gauge Breaking

The teacher’s heuristic varies under gauge transformations — it gives different estimates for the same learner state depending on the student’s surface characteristics. This is gauge variance, discussed in Section 6.2.4.