Principia Orthogona · Book IV · Higher Dimensions Contents · ← Ch 22 · Ch 24 →
Book IV · Chapter 23 · The Dimension Ladder, Continued

Duality and the Discriminant, or What the Projections Cost

Arc · Higher Dimensions, rung 22
Prerequisite · Ch 22 (Gauss map, conormal variety)
Status · classical, one correction, one row narrowed
Verification · ch23-verify.py · 12 blocks, all passing
Chapter 22 built the correspondence: the conormal variety Con(X) is the closure of the graph of the Gauss map, Legendrian in ℙT*ℙⁿ, and it projects two ways. This chapter asks what the two projections do. One recovers X. The other produces a discriminant, and the price of the passage is paid in degree — computed exactly by Plücker's formulas, and exactly refunded by the singularities. The same correspondence, read in an affine chart, is the Legendre transformation; and reading it that way settles a defect printed in Chapter 5, where a theorem was stated on a space where Bäcklund's theorem forbids it. The repair is not a retreat. It relocates the claim to the space the framework started in.

Contents

  1. Two projections, two questions
  2. The dual variety is a discriminant
  3. What the singularities cost
  4. When the dual is not a hypersurface
  5. Legendre, in an affine chart
  6. Bäcklund, and Chapter 5's Theorem 5.1
  7. The dm³ form is the standard one
  8. Honest inventory
§ 23.1

Two projections, two questions

The diagram of §22.6 is taken as given. Inside ℙT*ℙⁿ — pairs (point, hyperplane through it), of dimension 2n−1 — sits the conormal variety Con(X), Legendrian, of dimension n−1 whatever the dimension of X. It carries two projections, and the whole subject is the difference between them.

π₁ : Con(X) → X ⊂ ℙⁿ forget the hyperplane π₂ : Con(X) → X ⊂ (ℙⁿ) forget the point

π₁ is easy. Over a smooth point of a hypersurface its fibre is a single point, because the embedded tangent hyperplane is determined; over a smooth point of a variety of codimension c the fibre is a ℙc−1 of hyperplanes containing the tangent space. Either way the fibre dimension is constant on the smooth locus, and π₁ loses nothing.

π₂ is where the content is. Its fibre over a hyperplane H is the set of points of X at which H is tangent. Generically that is one point and π₂ is birational. Over special hyperplanes it is not: H may be tangent at two points, or tangent to higher order at one. Those are exactly the hyperplanes at which X fails to be smooth, and the book-keeping of that failure is the rest of this chapter.

Contact locus. For H ∈ X, the contact locus is π₁(π₂−1(H)) ⊆ X — the set of points where H is tangent. A general H has finite contact locus; a positive-dimensional contact locus for every H is the defective case of §23.4.

§ 23.2

The dual variety is a discriminant

The word discriminant arrives here rather than being imported. Consider the family of all binary forms of degree d, that is, all homogeneous polynomials f(u,v) = c₀ud + c₁ud−1v + ⋯ + cdvd. The coefficients (c₀ : ⋯ : cd) are a point of ℙd, and the forms with a repeated root make a hypersurface in that ℙd: the discriminant.

Separately, the rational normal curve Cd ⊂ ℙd is the image of [u:v] ↦ [ud : ud−1v : ⋯ : vd]. A hyperplane of ℙd is a linear condition on the coefficients, that is, a binary form; it is tangent to Cd exactly when that form has a repeated root. So:

The discriminant is a dual variety. The dual of the rational normal curve Cd ⊂ ℙd is the discriminant hypersurface of binary forms of degree d, and its degree is 2(d−1).

Checked

Computing Res(f, f′) in the coefficients and dividing out the leading coefficient — the resultant carries one extra factor of c₀ — gives degrees 2, 4, 6, 8 for d = 2, 3, 4, 5, which is 2(d−1) in each case. Block [7] of ch23-verify.py. For d = 2 this agrees with the smooth-plane-curve formula of §23.3, since C₂ is a conic and 2(2−1) = 2·1 = 2; the two formulas compute the same number by different routes, which is the only reason to trust either.

This is what the projection costs, stated positively. Passing from a variety to its dual converts a geometric object into the vanishing locus of an algebraic condition on a family. Every discriminant, resultant and determinant appearing anywhere in this corpus is π₂ applied to something.

§ 23.3

What the singularities cost

For a plane curve the accounting closes exactly. Let X ⊂ ℙ² be an irreducible curve of degree d whose only singularities are δ ordinary nodes and κ ordinary cusps. Then the dual curve — classically the class of X, the number of tangent lines through a general point — has degree

d = d(d−1) − 2δ − 3κ

Read that as a budget. A smooth curve of degree d has class d(d−1); every node refunds 2, every cusp refunds 3. The refunds are not conventions — they count how many of the d(d−1) tangent lines are absorbed into the singular point and are no longer available to pass through a general point elsewhere.

CurvedδκdCheck
smooth conic2002dual of a conic is a conic
smooth cubic3006a sextic, with 9 cusps at the 9 flexes
nodal cubic31046 − 2
cuspidal cubic30136 − 3 — computed independently below

The last row is the one worth doing by hand, because it can be checked without the formula. Parametrise the cuspidal cubic as (x, y) = (t², t³). The tangent line at parameter t is a x + b y + c = 0 subject to the two conditions that it meets the point and matches the direction; eliminating t between them gives the dual curve in line coordinates,

X : 4a³ + 27b²c = 0 degree 3, as the formula demands

Block [6]. Now read its singularities, because this is where the chart matters and where a careless reading goes wrong. The gradient (12a², 54bc, 27b²) vanishes on X at the single point [0 : 0 : 1]. That point is not visible in the affine chart b = 1, where the equation reads 4a³ + 27c = 0 — a smooth graph c = −4a³/27, with no singular point at all. It lives in the chart c = 1, where the same curve reads

4a³ + 27b² = 0 b² = −4a³/27 — a semicubical parabola, cusp at the origin

So the dual of a curve with a cusp is a curve with a cusp: the correspondence does not smooth anything out, it moves the singularity to the other side. But the cusp is only in view from the right chart, and a dual curve plotted in the wrong one looks smooth while being nothing of the kind. Block [11] locates the singular point projectively and then checks both charts, which is the only honest way to make the claim.

X : y² = x³ · d = 3, κ = 1 X∨ in the chart c = 1 : 4a³ + 27b² = 0 cusp at the origin · class 3 π₂
The cuspidal cubic and its dual, both plotted from their own equations rather than sketched, and both in a chart that contains the singular point. The cusp at the origin on the left reappears as the cusp at the origin on the right, opening the other way. Degree 3 on both sides: Plücker's refund of 3 for one cusp exactly cancels the rise from 3 to 6 that smoothness would have bought. Plotted in the chart b = 1 instead, the right-hand curve is the smooth graph c = −4a³/27 and the cusp is nowhere in the picture — the singular point of X∨ is the one line b = 0 that chart omits.
§ 23.4

When the dual is not a hypersurface

For a plane curve the dual is again a curve, and for most X the dual is a hypersurface. Not always. If every tangent hyperplane is tangent along a positive-dimensional locus, π₂ contracts, and X comes out too small.

Dual defect. def(X) = codim X − 1. It is zero exactly when the dual is a hypersurface. Biduality still holds; it is only the expectation of codimension one that fails.

The standard example is the Segre embedding ℙ¹ × ℙ² ⊂ ℙ⁵, which is the rank-one locus in the space of 2 × 3 matrices. Its dual is the locus where a general 2 × 3 matrix drops rank, and for a non-square matrix that locus has codimension 2, not 1. The defect is 1. Block [8].

Squareness is the whole of it: for m × m matrices the same construction gives the determinant, a genuine hypersurface with defect 0. The determinant is the dual variety of the Segre ℙm−1 × ℙm−1 — the cleanest statement in this chapter, and the one to keep if only one survives.

§ 23.5

Legendre, in an affine chart

Everything above is projective. Take an affine chart and the same correspondence turns into an object this book has used since Chapter 1 without naming it as duality.

On J¹(ℝ,ℝ) with coordinates (x, u, p) and contact form α = du − p dx, define

L(x, u, p) = (p, px − u, x)

The Legendre transformation is a contactomorphism, and an involution. L*α = −α, so L preserves ker α; and L ∘ L = id.

Checked

Both by direct computation, blocks [9] and [10]. The second is worth pausing on. Chapter 22 proved biduality (X) = X from the isomorphism ℙT*ℙⁿ ≅ ℙT*(ℙⁿ) exchanging the two projections. In an affine chart that isomorphism is L, and biduality is the identity L ∘ L = id, which takes one line to verify. The classical theorem and the undergraduate exercise are the same statement in two charts.

And L is not the prolongation of anything happening on the base. It sends the base coordinate x to the slope coordinate and the slope p to a base coordinate; a transformation of (x, u) alone could never do that. Hold that fact for one section.

§ 23.6

Bäcklund, and Chapter 5's Theorem 5.1

Chapter 5 identifies the Sator Square with J¹(ℝ, ℝ²) and states, as Theorem 5.1, that every contact symmetry of J¹(ℝ, ℝⁿ) is either a prolonged point transformation or a genuine contact transformation mixing x, yi and pi, the latter being the Galilean Contact Transformations of Chapter 6. As printed, the second alternative is empty on the space the chapter works in.

Bäcklund (1876). For n ≥ 2, every contact transformation of J¹(ℝ, ℝⁿ) — every diffeomorphism preserving the Cartan distribution — is the prolongation of a point transformation of the base. Genuine contact transformations exist only for n = 1.

The mechanism is dimensional, and visible before any theorem is invoked. A contact structure on a manifold of dimension 2m+1 is a maximally non-integrable field of hyperplanes — corank exactly one. On J¹(ℝ, ℝⁿ) the Cartan distribution is cut out by the n independent forms αi = dyi − pi dx, so its corank is n.

ndim J¹(ℝ,ℝⁿ)Cartan rankcorankcontact?
1321yes
2532no — a rank-2 Pfaffian system
3743no

So J¹(ℝ, ℝ²) is not a contact manifold at all, and "contact symmetry" there means symmetry of a Pfaffian system of rank 2. Bäcklund's theorem says such systems are rigid. The exceptional case n = 1 is exceptional precisely because corank 1 is the only case in which the distribution is a hyperplane field — and the Legendre transformation of §23.5 is the witness that the exception is not vacuous.

Theorem 5.1, as printed, is false for n ≥ 2. Its dichotomy offers a second alternative — genuine contact transformations of J¹(ℝ, ℝ²) — which Bäcklund's theorem shows is empty. Two repairs are available, and they are not equivalent.

(a) Keep the space. On J¹(ℝ, ℝ²) the correct statement is that every symmetry of the Cartan distribution is a prolonged point transformation. The Sator correspondence of §5.2 survives untouched; what does not survive is calling the resulting group "genuine contact transformations." It is the prolonged point group — a perfectly good group, and what Chapter 6 has in fact been computing.

(b) Move to the space where the claim is true. Genuine contact transformations exist on contact 3-manifolds. §23.7 shows the dm³ manifold is one, so the Chapter 6 programme is not lost — it is relocated to the space the framework started in.

Both repairs are honest, and (b) is the one this book takes, for the reason in the next section. Chapter 5 is amended to state Bäcklund and to name its own space correctly; the Sator material, which is a mnemonic and was never load-bearing, is unaffected.

§ 23.7

The dm³ form is the standard one

The framework's contact form is α = dz − r² dθ on ℝ³, with r² = x² + y². Three facts about it were assumed rather than checked, and one of them was assumed wrongly.

(i) r² dθ = x dy − y dx, so α = dz − x dy + y dx.

(ii) α ∧ dα = −2 dx ∧ dy ∧ dz. The form is contact on all of ℝ³, including the axis r = 0.

(iii) The shear ψ(x,y,z) = (x, y, z + xy) followed by y ↦ y/2 carries α to dz + y dx. Hence (ℝ³, ker α) is contactomorphic to the standard contact ℝ³ — that is, to J¹(ℝ, ℝ) with u = z and p = −y.

Checked

Blocks [1]–[3]. Point (ii) corrects a reading the corpus had carried: in cylindrical coordinates α ∧ dα = −2r dz ∧ dr ∧ dθ, which vanishes at r = 0, and this had been read as a degeneracy on the axis. The vanishing factor is the Jacobian of the polar chart. In Cartesian coordinates the coefficient is the constant −2, and there is no degeneracy anywhere.

Point (iii) is what makes repair (b) available. The dm³ manifold is J¹(ℝ, ℝ) up to contactomorphism; n = 1 is exactly Bäcklund's exceptional case; so genuine contact transformations — the Legendre transformation among them — do act on it. The Chapter 6 programme was being run on the one space where the theorem it needs is false, and it belongs on the space Chapter 3 already built.

What this does to the open question of §22.8

Chapter 22 left one row open: whether the dm³ contact structure is the real locus of a holomorphic contact structure on a projective variety. Point (iii) narrows it substantially. ℙT*ℙ² is a projective variety of dimension 3 carrying a canonical holomorphic contact structure, and in an affine chart that structure is the tautological form du − p dx. By (iii) the dm³ structure is contactomorphic to exactly that. So the structure algebraises.

What remains open, stated narrowly. Algebraising the structure is not algebraising the system. Half of the residual question has since been settled: the attractor is Legendrian. Its tangent is (0, 1, 1) in (r, θ, z), and α(γ̇) = ż − r²θ̇ = 1 − r², which vanishes at r = 1 — so γ* lies inside the contact distribution rather than transverse to it. (Chapter 3 had called it a closed Reeb orbit; the Reeb field of α is ∂z, which is the complementary condition, and Chapter 3 is corrected.) What is still open is algebraic: on the attractor ż = θ̇ = 1, so z = θ and the curve is a helix, whose parametrisation is transcendental in these coordinates even though the circle r = 1 is not. Whether the identification of (iii) carries it to an algebraic Legendrian in ℙT*ℙ², and whether the operator chain is induced by a morphism, is not settled here. ch24-verify.py, block [1].

§ 23.8

Honest inventory

ClaimStatusWhat it rests on
π₁ has constant fibre dimension on the smooth locusClassicaldefinition of Con(X)
dual of Cd is the discriminant, degree 2(d−1)ClassicalGKZ ch. 1; recomputed, block [7]
d = d(d−1) − 2δ − 3κClassicalPlücker; four cases checked, block [5]
dual of the cuspidal cubic is 4a³ + 27b²cComputedelimination, block [6]
that dual is itself cuspidal, at [0 : 0 : 1]Computedsingular locus, block [11]; invisible in the chart b = 1
4a³ + 27b² is dual, discriminant and bifurcation setComputedblock [12] — the bridge to Ch 24
def(ℙ¹ × ℙ²) = 1Classicalrank locus of 2 × 3 matrices, block [8]
L*α = −α and L ∘ L = idComputedblocks [9], [10]
J¹(ℝ,ℝⁿ) is contact iff n = 1Classicalcorank count, block [4]
Theorem 5.1 as printed is false for n ≥ 2CorrectionBäcklund — see §23.6
α = dz − r²dθ is contact on all of ℝ³Computedblock [2]; corrects the axis reading
dm³ form is contactomorphic to J¹(ℝ,ℝ)Computedblock [3], explicit shear
the dm³ structure algebraisesFollows(iii) plus the canonical form on ℙT*ℙ²
the dm³ attractor is LegendrianComputedα(γ̇) = 1 − r² = 0 at r = 1; ch24-verify.py [1]
the dm³ flow algebraisesOPENLegendrian yes; the helix z = θ is transcendental in these coordinates

Every row marked Computed is regenerated by ch23-verify.py, which sits beside this chapter and exits non-zero on any failure. It caught two errors while the chapter was being written. In §23.2: the resultant Res(f, f′) has degree 2d−1, not 2(d−1), and the discriminant is that divided by the leading coefficient. And in §23.3, the dual of the cuspidal cubic had been written in the chart b = 1 and plotted there, where it is a smooth graph — under a caption asserting that its cusp was visible. The singular point [0 : 0 : 1] is precisely the point that chart omits. Both figures printed above are the corrected ones. A dual curve plotted in a chart that misses its singular locus is the failure mode this chapter is about, arriving from the inside.

BäcklundA. V. Bäcklund, Ueber Flächentransformationen, Math. Ann. 9 (1876). The rigidity statement for n ≥ 2 is the one cited in §23.6.
DiscriminantsGelfand–Kapranov–Zelevinsky, Discriminants, Resultants and Multidimensional Determinants, ch. 1 for the identity of dual varieties and discriminants; ch. 9 for the determinant as a dual.
PlückerThe formulas are in every classical text; Griffiths–Harris, Principles of Algebraic Geometry, ch. 2 derives them from the same correspondence used here.
Dual defectEin's classification of smooth defective varieties; the Segre ℙ¹ × ℙ² is the smallest example.
Standard contact ℝ³That dz + r²dθ and dz + y dx define the same structure is standard — see Geiges, An Introduction to Contact Topology, §2.1. The explicit shear is written out here because §23.7 needs the map, not only its existence.
Where this goes nextBlock [12] is the hinge. 4a³ + 27b² is the dual variety computed in §23.3, the discriminant of x³ + ax + b, and the bifurcation set of the cusp catastrophe — one equation wearing three names. Ch 24 takes that identity back into dynamical systems: the bifurcation set of a family is the discriminant of that family, and π₂ is the projection from the equilibrium manifold to parameter space.
Principia Orthogona · Book IV · Chapter 23
Pablo Nogueira Grossi · G6 LLC, Newark, New Jersey · ORCID 0009-0000-6496-2186
ISBN 979-8-9954416-8-7 · doi 10.5281/zenodo.19117400
← Ch 22 · The Gauss Map · Contents · Ch 24 · The Bifurcation Set Is a Discriminant →
Proved · kernel-checked
discriminant book21/Spiral.lean:75 Each name above is declared in this repository at the line shown and appears in an axiom report with no sorryAx. A clean axiom report is not a reading of the statement: per R20, a theorem can assume its conclusion and still report clean. Follow the link before citing one as evidence.