Chapter 2: The Failure of Scalar Ethics

RUNNING EXAMPLE — Priya’s Model

Priya’s TrialMatch score is a textbook scalar: one number per patient, higher is better. She now sees the three failures playing out in her own system. First, no direction: the score cannot distinguish whether a patient scored 64 because of medical unsuitability or geographical isolation, yet these require opposite interventions. Second, uncertainty has shape: the model is highly confident about urban patients (ample training data) and deeply uncertain about rural ones, but the score encodes no shape of uncertainty. Third, paths cross boundaries: when a rural patient’s insurance status changes, her score can jump from 65 to 72 without any medical change, silently crossing the eligibility threshold. The scalar hides all of this.

The Parable of the Old Man and His Horse

There is an ancient Chinese parable known as 塞翁失马 (Sāi Wēng Shī Mǎ) — “The Old Man at the Border Loses His Horse.” It goes like this:

An old man living near the frontier lost his horse. His neighbors came to console him, but the old man said, “How do you know this isn’t good fortune?”

Some months later, the horse returned, bringing with it a herd of fine wild horses. The neighbors came to congratulate him, but the old man said, “How do you know this isn’t bad fortune?”

With so many horses, the old man’s son took to riding. One day he fell and broke his leg. The neighbors came to console the old man, but he said, “How do you know this isn’t good fortune?”

A year later, war came to the border. All the able-bodied young men were conscripted, and most died in battle. But the old man’s son, with his broken leg, was spared.

The parable is usually read as a lesson in epistemic humility: we cannot know whether present events are good or bad because we cannot foresee their consequences. “Maybe” is the only honest answer.

But there is a different reading — one that reveals something structural about moral evaluation itself. The old man’s “maybe” is not merely a confession of ignorance. It is a recognition that scalar evaluation is the wrong tool for the job.

When we say “losing the horse is bad,” we assign a number — call it S(x) = −1 — to the present state. When the horse returns with others, we revise: S(x) = +3. When the son breaks his leg, we revise again: S(x) = −2. And so on.

But notice what this scalar cannot represent:

Which directions matter. The loss of the horse is “bad” primarily along the wealth axis. It says nothing about the health axis, the family axis, or the political axis (the son’s eventual exemption from conscription). A scalar collapses all these dimensions into a single number, losing the information about where the badness lies.

Where uncertainty concentrates. The old man’s uncertainty is not uniform. He is quite certain the horse is gone. What he is uncertain about is whether events will unfold along axes where the loss matters. Will famine come (making the lost horse catastrophic)? Will the horse return (making the loss temporary)? Will war come (making his son’s presence at home decisive)? The uncertainty has shape — it lies along some directions more than others.

How evaluation changes along paths. The moral status of “son breaks leg” depends on whether war is coming. The trajectory matters. Crossing from peacetime into wartime changes the evaluative landscape discontinuously — what was unambiguously bad (broken leg) becomes ambiguously fortunate (exemption from death). A scalar at a point cannot represent these regime changes.

The parable points at a mathematical truth: moral evaluation requires geometric structure that scalars cannot provide.

The Insufficiency of Rank-0 Ethics

Let us be precise about what a scalar moral evaluation can and cannot do.

A scalar is a quantity fully specified by a single number. In ethics, scalar approaches assign a value — utility, welfare, goodness, rightness — to states of affairs, actions, or outcomes. The utilitarian calculus is scalar: sum the pleasures, subtract the pains, maximize the total. Cost-benefit analysis is scalar: all considerations reduce to a common currency. Even pluralistic theories that acknowledge multiple values typically seek, at the moment of decision, to collapse this plurality into a single ranking.

Formally, a scalar moral evaluation is a function:

S:M→R

assigning to each point x in the moral space M a real number S(x). The defining feature of a scalar is invariance: under any admissible coordinate transformation (any redescription of the situation), the value S(x) remains the same.

This invariance is both the strength and the weakness of scalar ethics. It is a strength because it promises objectivity: the goodness of a state should not depend on how we describe it. It is a weakness because it discards information: to achieve invariance, we must throw away everything that varies with perspective.

The parable of the old man reveals three specific structural limitations. We now develop each with the precision it requires.

Limitation 1: No Directional Information

A scalar S(x) tells us the magnitude of value at a point. It cannot tell us which directions in the moral space are responsible for that value, nor which directions would change the evaluation most dramatically.

Consider the moment the horse runs away. A scalar evaluation might say S = −1. But this number conceals the structure of the situation:

Along the wealth dimension: strongly negative (valuable asset lost)

Along the labor dimension: moderately negative (horse did farm work)

Along the health dimension: neutral (no one is sick or injured)

Along the family dimension: neutral (relationships unchanged)

Along the political dimension: indeterminate (depends on future events)

A vector can represent this structure. Let v=(-0.8,-0.4,0,0,?) be the “impact vector” of the horse’s departure, with components along each morally relevant dimension. The scalar S=-1 is some contraction of this vector — perhaps its magnitude, or a weighted sum — but the vector itself contains information the scalar discards.

Why does this matter? Because moral reasoning requires knowing which dimensions are engaged. If a proposed remedy addresses the wrong dimension — offering emotional support when the problem is financial — it will be ineffective despite targeting the “badness.” The vector structure tells us where to intervene; the scalar does not.

More precisely, if the moral state is described by a point x∈M in a nine-dimensional moral manifold (Chapter 5), then the impact of an event is a tangent vector δx∈T_xM describing the direction and magnitude of change. The scalar evaluation S(x) is a function on M; its gradient ∇S is a covector indicating the direction of steepest increase in moral value. But ∇S is determined by S — and if S was obtained by collapsing a richer structure, the gradient of the collapsed function need not point in the same direction as the gradient of any particular component.

Example. Let M have two dimensions: wealth ( w) and health ( h). Let the true moral state be the vector (w,h)=(-3,0) — significant financial loss, no health impact. Let the scalar collapse be S=w+h=-3. The gradient ∇S=(1,1) points equally toward improving wealth and health. But the actual deficit lies entirely along wealth. The scalar gradient misdirects remedial effort.

This is not a contrived example. Cost-benefit analyses that aggregate environmental damage and economic benefit into a single dollar figure routinely suffer from this pathology. The gradient of the aggregate points toward “more dollars,” indifferent to which dimension generates them.

Limitation 2: Uncertainty Has Shape

The old man’s “maybe” reflects uncertainty about the future. But his uncertainty is not uniform across all possibilities. It has shape: he is more uncertain along some dimensions than others, and — critically — his uncertainty is greatest along the dimensions that are most ethically decisive.

A scalar treatment of uncertainty adds error bars: S=-1±0.5. This says the true value lies somewhere in the interval [-1.5,-0.5], but nothing about why we are uncertain or where the uncertainty matters.

A geometric treatment represents uncertainty as a covariance tensor — a rank-2 object that encodes both the magnitude and the directional structure of our uncertainty:

Σ^ij=E[(δmⁱ)(δm^j)]

This tells us: uncertainty is large along axis i, small along axis j, and correlated between axes i and k.

In the parable, the old man’s uncertainty tensor might have the following structure:

Small variance along the current wealth axis (the horse is definitely gone): Σ^ww≈0

Large variance along the future wealth axis (will more horses come?): Σ^w'w'≫0

Large variance along the political axis (will there be war?): Σ^pp≫0

Strong covariance between political and son’s welfare axes (war affects conscription): Σ^ps≫0

The crucial insight is that uncertainty concentrated along ethically decisive directions matters more than uncertainty along irrelevant directions. If the old man were uncertain about the color of next year’s crops but certain about everything affecting his family’s survival, the first uncertainty would be ethically negligible. But if he is uncertain precisely about war and conscription — the dimensions that determine whether the broken leg is a tragedy or a salvation — then his uncertainty is ethically maximal.

Scalar uncertainty ( S±ε) cannot distinguish these cases. The covariance tensor Σ can.

The scalar quantity that captures this alignment is the variance of the moral evaluation under structured uncertainty:

σ_S²=Σ^ij(∂S)/(∂mi)(∂S)/(∂mj)

This is large when uncertainty concentrates along morally decisive directions (where ∇S is large), and small when uncertainty lies along morally irrelevant directions (where ∇S is small). The formula is a contraction of the uncertainty tensor with the moral gradient — a genuinely tensorial operation that no scalar framework can replicate.

Limitation 3: Paths Cross Boundaries

The most profound limitation of scalar evaluation is its inability to represent trajectory-dependent moral change, especially trajectories that cross regime boundaries.

In the parable, the evaluation of “son has a broken leg” depends on whether war comes. Before the declaration of war, a broken leg is unambiguously bad: pain, disability, inability to work. After war is declared, the evaluation bifurcates: for those without exemptions, conscription leads to probable death; for those with exemptions (including the son), survival is likely. The broken leg, unchanged in itself, has crossed a moral phase boundary.

This is not merely a matter of new information changing our estimate. It is a structural feature of the moral landscape: there exist strata (regimes, phases) within which smooth trade-offs apply, separated by boundaries where the rules change discontinuously.

A scalar function S:M→R, if it is continuous, cannot represent such discontinuities. It can represent gradual change — S increasing or decreasing smoothly — but not the sharp transitions that characterize moral thresholds: consent given vs. withheld, life vs. death, war vs. peace.

To represent regime boundaries, we need stratified spaces: spaces composed of smooth strata (within which differential calculus applies) joined along lower-dimensional boundaries (where discontinuities are permitted). And to represent how moral status evolves along paths that may cross these boundaries, we need path-dependent operations: parallel transport, holonomy, trajectory integrals.

Definition. A moral phase boundary is a codimension-1 submanifold B⊂M such that the moral evaluation function S (or more generally, the moral tensor field) is smooth on each side of B but may be discontinuous across B.

The parable contains a clear example. Let p∈[0,1] denote the probability of war. As p increases continuously, the evaluation of the son’s broken leg remains continuously negative — until p crosses some threshold p^* at which conscription becomes a near certainty. At that point, the evaluation jumps: the broken leg becomes a life-saving exemption. The function S(broken leg|p) is discontinuous at p=p^*.

This is not a pathology of the parable. Empirical evidence confirms that moral evaluation generically involves such discontinuities. Analysis of the Dear Abby corpus (Chapter 17) reveals that qualifying phrases — “you promised,” “he hit you,” “only if convenient” — function as semantic gates: discrete triggers that flip moral states between obligation and liberty, claim and no-claim. These transitions behave as step functions, not sigmoids. The geometry of moral space is not uniformly smooth.

What Geometric Structure Provides

The three limitations point toward three geometric structures beyond scalars:

Let us examine each, now as elements of a unified framework.

Gradients and the Direction of Moral Change

If moral evaluation were purely scalar, there would be no meaningful sense of “direction” in moral space. But our actual moral reasoning is saturated with directional concepts: obligations point toward required states; interests aim at objects; responsibility flows from agents to patients; improvement moves toward better configurations.

These are naturally modeled as vector quantities — objects with both magnitude and direction.

Consider an obligation. “You ought to help your neighbor” is not merely a magnitude of oughtness. It specifies a direction: from the current state (neighbor unhelped) toward a required state (neighbor helped). The obligation can be stronger or weaker (magnitude), but it also has an orientation in the space of possible actions.

Formally, we represent obligations as vector fields on the moral manifold:

O^μ(x):M→TM

At each point x in moral space, O^μ(x) is a tangent vector pointing in the direction of what is required. In the nine-dimensional framework of Chapter 5, this vector has components along each of the nine moral dimensions derived from the 3 × 3 matrix of ethical scopes and epistemic modes.

Interests, conversely, are represented as covector fields:

I_μ(x):M→T^*M

A covector (or 1-form) is a linear functional on vectors. The interest I_μ, applied to an obligation O^μ, yields a scalar: the satisfaction of interest I by obligation O.

S=I_μO^μ

This is the fundamental formula of geometric ethics (Chapter 6): satisfaction is the contraction of obligations with interests. It is coordinate-invariant (a scalar), but it arises from vector quantities that carry directional information. The scalar is not primary; it is a projection — a shadow cast by a richer object.

The gradient ∇S of the satisfaction function tells us: at this point in moral space, which direction increases satisfaction most rapidly? This is the direction of moral improvement — not a scalar claim (“things could be better”) but a vector claim (“things could be better in this specific way”).

The Metric Tensor and Moral Comparison

To speak of directions, we need a way to compare them. To speak of distances, we need a way to measure them. In differential geometry, both functions are performed by the metric tensor g_μν.

The metric tensor is a rank-2 object that defines the inner product between vectors:

⟨u,v⟩=g_μνu^μv^ν

This allows us to say when two directions are orthogonal (inner product zero), aligned (inner product positive), or opposed (inner product negative). It also defines the length of vectors and the distance between points.

In moral space, the metric encodes how we compare values. Two values are orthogonal if trading off one against the other is undefined — there is no exchange rate between them. They are aligned if improving one tends to improve the other. They are opposed if they conflict.

The claim that some values are incommensurable is, in geometric language, the claim that the moral metric is degenerate along certain directions — that there exist vectors v such that g_μνv^μw^ν=0 for all w, or equivalently that the metric has zero eigenvalues. This is a structural claim, not a mystical one. It says that the geometry of moral space is not Euclidean — not all directions are comparable in the way that spatial directions are.

Different ethical theories correspond to different metrics on the same underlying manifold:

Utilitarianism: g_μν=δ_μν (Euclidean — all dimensions equally weighted, all trade-offs permitted)

Rawlsian justice: g_μν structured so that the worst-off dimension dominates (supremum metric)

Lexicographic theories: g_μν with hierarchically ordered eigenvalues

Incommensurability claims: g_μν with zero eigenvalues or degenerate subspaces

The choice of metric is doing the real moral work — and scalar frameworks hide this choice. The geometric framework makes it explicit and debatable. Where the metric comes from is the subject of Chapter 9.

Stratification and Moral Phase Transitions

The parable’s most profound feature is the regime change brought by war. Before war is declared, the broken leg is bad. After war is declared, it becomes potentially good (exemption from death). This is not a gradual transition; it is a discontinuous jump at a boundary.

To represent such discontinuities, we need stratified spaces: spaces composed of smooth manifolds (strata) joined along boundaries where the smooth structure breaks down.

Definition. A stratified moral space M consists of:

Strata M_α: smooth manifolds of various dimensions, representing regimes within which ordinary calculus applies

Incidence relations: rules specifying how strata adjoin (Whitney conditions ensuring geometric regularity)

Stratum-dependent evaluation: moral functions that are smooth on each stratum but may jump across boundaries

The boundary between “peacetime” and “wartime” is a moral stratum boundary. On either side, smooth trade-offs apply (more wealth is better, less pain is better). But crossing the boundary changes which smooth trade-offs apply. The rules are different.

This is why the old man cannot assign a stable scalar to his son’s broken leg. The evaluation depends on which stratum the world occupies, and that is exactly what he is uncertain about.

The formal apparatus of stratification — including absorbing strata (nullifiers), semantic gates (discrete transition functions), and the Whitney conditions that ensure well-behaved boundary geometry — is developed in Chapter 8.

The Parable Revisited

Let us return to the old man with the full geometric apparatus.

Moment 1: The horse runs away.

The moral state is x₁∈M. The impact vector lies primarily along the wealth dimension:

δx₁≈(-0.8,-0.4,0,0,0,…)

The obligation vector O^μ points toward “recover the horse or find an alternative.” The uncertainty tensor Σ^ij is large along future-wealth and political axes.

A scalar evaluation says S(x₁)≈-1. But this discards the directional structure. The old man, implicitly recognizing the geometric structure, says “maybe.”

Moment 2: The horse returns with others.

The moral state is x₂. The impact vector is strongly positive along wealth. The obligation vector O^μ now reorients toward “maintain and steward this windfall.” The uncertainty tensor remains large along the political axis.

A scalar evaluation says S(x₂)≈+3. The neighbors celebrate. The old man, still tracking the full tensor, says “maybe.”

Moment 3: The son breaks his leg.

The moral state is x₃. The impact vector is negative along health and capability. The uncertainty tensor Σ^ij becomes crucial: there is high covariance between the political dimension (will there be war?) and the welfare dimension (will the son survive?).

Crucially, the son’s condition now sits near a stratum boundary. If war comes, the moral evaluation of the broken leg will discontinuously shift. The gradient ∇S is undefined at the boundary — it points one way in peacetime, another in wartime.

A scalar evaluation says S(x₂)≈+3. The neighbors celebrate. The old man, still tracking the full tensor, says “maybe.”

Moment 4: War comes; the son is spared.

The moral state crosses the boundary into the wartime stratum. The broken leg, unchanged in itself, is now on a different stratum. Its evaluation, relative to the counterfactual (able-bodied son conscripted and killed), is strongly positive.

The scalar is now S(x₄)≫0. But this conceals the path-dependence: the same physical state (broken leg) has different moral valence depending on which stratum it occupies. The tensor captures this; the scalar cannot.

Why “Maybe” Is Geometric, Not Merely Epistemic

The standard interpretation of the parable is epistemic: we should say “maybe” because we lack knowledge of the future. If only we knew whether war was coming, we could assign definite values.

The geometric interpretation suggests something deeper: “maybe” is the correct answer even with perfect knowledge when the evaluation structure is geometric rather than scalar.

Suppose the old man had an oracle who told him exactly what would happen. Would he then assign a definite scalar to each moment?

Only if he were willing to commit to:

A fixed metric: how much does wealth matter relative to health relative to family? (The choice of g_μν.)

A fixed treatment of path-dependence: does the broken leg’s value depend on the trajectory through the wartime stratum, or just the final state? (The choice of contraction.)

A specific contraction: which rank-0 summary of the rank- n moral tensor serves as the action guide? (The projection from tensor to scalar.)

These choices are not determined by the facts. They are perspective-dependent — different agents, with different interests and different metrics, will perform different contractions and arrive at different scalars. As Chapter 6 makes precise, the fundamental formula S=I_μO^μ already shows this: the scalar S depends on both the obligation vector (fixed by the situation) and the interest covector (varying with the stakeholder).

The tensor is the invariant structure. The scalar is a projection — a shadow — that loses information. “Maybe” is what you say when you recognize that no single scalar faithfully represents the tensor.

From Parable to Framework

The parable motivates the mathematical framework developed in subsequent chapters:

Moral space has dimension (Chapter 5). The space of morally relevant configurations is not a line (totally ordered by goodness) but a nine-dimensional manifold, with independent axes derived from the intersection of three ethical scopes (individual, relational, systemic) and three epistemic modes (empirical, deontic, aretaic). These dimensions are not arbitrary; they are the minimum set required to distinguish situations that manifestly differ in moral character.

Moral quantities are tensors (Chapter 6). Obligations, interests, evaluations, and uncertainty are tensors of various ranks, carrying directional and relational information that scalars discard. The fundamental formula S=I_μO^μ makes the passage from tensor to scalar explicit and auditable.

Moral space has a metric (Chapters 6, 9). The structure that allows comparison of values, measurement of moral distance, and identification of incommensurable dimensions is a metric tensor g_μν. Different ethical theories correspond to different metrics; the metric is not given by the mathematics but by substantive moral commitment.

Moral space is stratified (Chapter 8). The space is not uniformly smooth but divided into strata — regimes within which smooth trade-offs apply — separated by boundaries where rules change discontinuously. Phase transitions, absorbing strata, and semantic gates formalize the discrete jumps that punctuate moral life.

Moral transformation has structure (Chapters 5, 10). What happens to moral evaluations when we shift perspective, permute agents, or translate across contexts? Tensor transformation laws and parallel transport answer this question with precision.

The old man at the border, without the language of differential geometry, had the insight: moral reality is richer than any scalar can capture. What geometric ethics provides is the mathematical apparatus to make that insight rigorous.

Anticipating Objections

The skeptical reader will have objections. Let us address the two most immediate; deeper objections receive extended treatment later.

“Isn’t this just saying ethics is complicated?”

No. “Complicated” suggests more of the same — more variables, more factors, more considerations to weigh. Geometric structure is different in kind. A vector is not a complicated scalar; it has properties (direction, transformation behavior) that scalars lack categorically. The claim is not merely that ethics has many dimensions (though it does) but that moral quantities transform in specific ways under change of perspective, and this transformation behavior is precisely what geometric objects — tensors, connections, fiber bundles — are designed to capture.

An analogy: a photograph of a landscape is not a “complicated pixel.” It is a two-dimensional array with structure — spatial relationships, color gradients, edges — that no single number can represent. The move from scalar to geometric ethics is like the move from a pixel to a photograph: not more of the same, but a different kind of representation.

“We can’t measure moral quantities precisely, so what use is this formalism?”

The same objection was raised against utility theory, and against the use of calculus in economics. The response is twofold.

First, the framework’s value lies in structural insights, not numerical precision. Knowing that two values are orthogonal (incommensurable under the current metric) is useful even if we cannot assign magnitudes to either. Knowing that uncertainty concentrates along the health dimension rather than the wealth dimension changes how we deliberate, even without exact numbers. The qualitative geometry constrains reasoning before any quantity is measured.

Second, the framework identifies what would need to be measured to make ethical reasoning precise. Physics progressed from qualitative insights (“force causes acceleration”) to quantitative laws ( F=ma) as measurement improved. Chapter 17 presents evidence that geometric structures predicted by the framework — multi-dimensionality, correlative symmetry, context-dependent weighting, temporal stability — are empirically detectable in natural moral reasoning, using the Dear Abby corpus as ground truth. The measurements are coarse, but they exist, and the framework tells us how to refine them.

Conclusion

The old man’s “maybe” is not resignation. It is recognition.

A scalar S(x) can label the present state, but it cannot represent which directions are ethically decisive, whether uncertainty concentrates along those directions, or how evaluation changes along trajectories that cross stratification boundaries.

These three failures are not peripheral. They are structural — consequences of the fact that a rank-0 object (a scalar) cannot faithfully represent a rank- n structure (a tensor). The information lost in the projection from tensor to scalar is precisely the information that matters most in hard moral cases: which values are at stake, where the uncertainty bites, and whether we are near a phase transition where the rules change.

The geometric structures that resolve these failures are:

Vectors and covectors — capturing the direction of obligation and interest

The covariance tensor — encoding uncertainty in morally relevant coordinates

Stratified spaces — representing the discontinuous boundaries between moral regimes

Parallel transport — tracking how moral evaluations evolve along paths that may cross boundaries

Without these structures, “maybe” must be treated as an ad hoc heuristic — a confession of ignorance — rather than emerging from the model as a structural feature of moral reality.

The parable, read geometrically, is not about the limits of human knowledge. It is about the geometry of ethical evaluation. The old man sees that the tensor cannot be contracted to a scalar without loss. His “maybe” is a holding operation — a refusal to project — until the full structure of the situation is revealed.

The chapters that follow develop the apparatus: the historical precedents (Chapter 3), the mathematical preliminaries (Chapter 4), the moral manifold (Chapter 5), the tensor hierarchy (Chapter 6), and the pedagogical centerpiece (Chapter 7). But the core insight is here, in the parable: ethics is not a number. It is a geometric structure. And the first step in understanding that structure is recognizing what scalars cannot do.

❖

The horse runs away. Good? Bad? Maybe.

The geometry agrees: the tensor is not yet fully contracted. Hold the projection. Watch the structure unfold.