The Legendre Transform as the Tropical Fourier Transform

Table of Contents

• 1. The Tropical Semiring
  • 1.1 Definition and Motivation
  • 1.2 Maslov’s Dequantization
• 2. Tropicalizing the Fourier Transform
  • 2.1 Classical Laplace Transform
  • 2.2 The Tropical Substitution
  • 2.3 The Legendre–Fenchel Conjugate
• 3. The Convolution Theorem
  • 3.1 Inf-Convolution as Tropical Convolution
  • 3.2 The Tropical Convolution Theorem
• 4. The Inversion Theorem
  • 4.1 Fourier Inversion and Its Tropical Analog
  • 4.2 The Biconjugate Theorem
• 5. The Full Dictionary
• 6. Connections and Applications
  • 6.1 Legendre Transform in Physics and Thermodynamics
  • 6.2 Tropical Characters and Affine Functions
  • 6.3 Connection to Neuroalgebraic Geometry
• 7. References

1. The Tropical Semiring 📐

1.1 Definition and Motivation

Definition (Tropical Semiring). The tropical semiring (or max-plus semiring) is the set \(\mathbb{T} = \mathbb{R} \cup \{-\infty\}\) equipped with: \[a \oplus b = \max(a, b), \qquad a \odot b = a + b.\]

The element \(-\infty\) is the additive identity (\(a \oplus (-\infty) = a\)) and \(0\) is the multiplicative identity (\(a \odot 0 = a\)). The min-plus semiring \((\mathbb{R} \cup \{+\infty\}, \min, +)\) is the dual; the two are related by negation.

A tropical polynomial in one variable is a max-plus expression:

\[p(x) = \max(a_0, a_1 + x, a_2 + 2x, \ldots, a_d + dx) = \bigoplus_{k=0}^d (a_k \odot x^{\odot k}),\]

where \(x^{\odot k} = kx\). This is a piecewise-linear convex function of \(x\). Its tropical roots are the breakpoints of this piecewise-linear function, counted with multiplicity.

1.2 Maslov’s Dequantization

The tropical semiring arises naturally as a classical limit. Define the deformed semiring \((\mathbb{R}, \oplus_\hbar, \odot_\hbar)\) by:

\[a \oplus_\hbar b = \hbar \log(e^{a/\hbar} + e^{b/\hbar}), \qquad a \odot_\hbar b = a + b.\]

At \(\hbar = 1\) (with \(e^a\) standing in for \(a\)), this recovers ordinary addition and multiplication. As \(\hbar \to 0\):

\[a \oplus_\hbar b = \hbar \log\bigl(e^{a/\hbar}(1 + e^{(b-a)/\hbar})\bigr) \xrightarrow{\hbar \to 0} \max(a, b).\]

The tropical semiring is the \(\hbar \to 0\) (classical) limit of the usual semiring under the logarithmic change of variables \(a \mapsto \hbar \log a\). This is Maslov dequantization: \(\hbar\) plays the role of Planck’s constant, and the tropical world is the “classical mechanics” limit of “quantum” arithmetic.

2. Tropicalizing the Fourier Transform 📐

2.1 Classical Laplace Transform

The Laplace transform (real-variable version of the Fourier transform) is:

\[\mathcal{L}[f](\xi) = \int_{\mathbb{R}^n} f(x)\, e^{x \cdot \xi}\, dx,\]

for functions \(f: \mathbb{R}^n \to \mathbb{R}_{\geq 0}\) (or more generally, \(f\) in the appropriate \(L^1\) class). The Fourier transform \(\hat{f}(\xi) = \int f(x) e^{ix\cdot\xi} dx\) is obtained by replacing \(\xi\) with \(i\xi\); the two are related by Wick rotation.

The key structural property: the Laplace transform converts convolution into pointwise multiplication:

\[\mathcal{L}[f * g] = \mathcal{L}[f] \cdot \mathcal{L}[g].\]

2.2 The Tropical Substitution

To tropicalize the Laplace transform, make the substitution dictated by Maslov dequantization: - Replace \(f(x)\) with \(e^{g(x)/\hbar}\) (writing \(f\) in logarithmic coordinates, \(g = \hbar \log f\)) - Replace \(e^{x \cdot \xi}\) with \(e^{x \cdot \xi / \hbar}\) (tropical multiplication \(\odot\)) - Replace \(\int\) (classical sum, \(\oplus_{\hbar=1}\)) with \(\sup\) (tropical sum \(\oplus_{\hbar=0}\))

The transform becomes:

\[\mathcal{L}_\hbar[g](\xi) = \hbar \log \int e^{g(x)/\hbar} e^{x \cdot \xi / \hbar}\, dx = \hbar \log \int e^{(g(x) + x \cdot \xi)/\hbar}\, dx.\]

By Laplace’s method (steepest descent), as \(\hbar \to 0\):

\[\mathcal{L}_\hbar[g](\xi) \xrightarrow{\hbar \to 0} \sup_{x \in \mathbb{R}^n}\bigl(g(x) + x \cdot \xi\bigr).\]

The Legendre transform is the \(\hbar \to 0\) limit of the Laplace transform in logarithmic coordinates.

2.3 The Legendre–Fenchel Conjugate

Definition (Legendre–Fenchel Conjugate). For \(f: \mathbb{R}^n \to \mathbb{R} \cup \{+\infty\}\), the convex conjugate (or Legendre–Fenchel transform) is:

\[f^*(\xi) = \sup_{x \in \mathbb{R}^n}\bigl(\langle x, \xi\rangle - f(x)\bigr).\]

This is the tropical Laplace transform of \(-f\), or equivalently the tropical Laplace transform of \(f\) in the min-plus convention. Both sign conventions appear in the literature; the key object is the supremum.

This Gaussian duality is the clearest finite-dimensional instance of the Fourier–Legendre analogy: \(\mathcal{F}[e^{-\|x\|^2/2\sigma^2}] \propto e^{-\sigma^2\|\xi\|^2/2}\) mirrors \(\left(\frac{1}{2\sigma^2}\|x\|^2\right)^* = \frac{\sigma^2}{2}\|\xi\|^2\).

3. The Convolution Theorem 🔑

3.1 Inf-Convolution as Tropical Convolution

Definition (Inf-Convolution). For \(f, g: \mathbb{R}^n \to \mathbb{R} \cup \{+\infty\}\), the inf-convolution is:

\[(f \,\square\, g)(x) = \inf_{y \in \mathbb{R}^n}\bigl(f(y) + g(x - y)\bigr).\]

This is the tropical analog of convolution: replace integration \(\int\) with \(\inf\) and multiplication \(f(y) g(x-y)\) with addition \(f(y) + g(x-y)\) — exactly the \((\min, +)\) substitution.

3.2 The Tropical Convolution Theorem

Theorem (Tropical Convolution Theorem). For any \(f, g: \mathbb{R}^n \to \mathbb{R} \cup \{+\infty\}\):

\[\boxed{(f \,\square\, g)^* = f^* + g^*.}\]

The Legendre–Fenchel transform converts inf-convolution into pointwise addition.

Proof. Compute directly: \[ (f\,\square\, g)^*(\xi) = \sup_x\Bigl(\langle x,\xi\rangle - \inf_y\bigl(f(y) + g(x-y)\bigr)\Bigr) = \sup_x \sup_y\Bigl(\langle x,\xi\rangle - f(y) - g(x-y)\Bigr). \] Substitute \(z = x - y\), so \(x = y + z\): \[ = \sup_{y,z}\Bigl(\langle y,\xi\rangle + \langle z,\xi\rangle - f(y) - g(z)\Bigr) = \sup_y\bigl(\langle y,\xi\rangle - f(y)\bigr) + \sup_z\bigl(\langle z,\xi\rangle - g(z)\bigr) = f^*(\xi) + g^*(\xi). \quad \square \]

Comparing with the classical convolution theorem \(\mathcal{F}[f * g] = \mathcal{F}[f] \cdot \mathcal{F}[g]\):

The proof is structurally identical to the proof of the classical convolution theorem: both use Fubini to separate variables after a substitution.

4. The Inversion Theorem 📐

4.1 Fourier Inversion and Its Tropical Analog

The classical Fourier inversion theorem states: if \(f \in L^1(\mathbb{R}^n)\) and \(\hat{f} \in L^1(\mathbb{R}^n)\), then \(f = \mathcal{F}^{-1}[\hat{f}]\). In the Laplace/Fourier context, the condition on \(f\) (square-integrability, \(L^1\), etc.) controls when inversion is valid.

The tropical analog asks: for which \(f\) does \(f^{**} = f\)? (Applying the Legendre transform twice.)

4.2 The Biconjugate Theorem

Theorem (Biconjugate / Fenchel–Moreau). For any \(f: \mathbb{R}^n \to \mathbb{R} \cup \{+\infty\}\) that is not identically \(+\infty\):

\[f^{**} = \overline{\mathrm{co}}\, f,\]

the lower semicontinuous convex hull of \(f\) (the largest closed convex function \(\leq f\)). In particular, \(f^{**} = f\) if and only if \(f\) is closed (lower semicontinuous) and convex.

This is the tropical Fourier inversion theorem: you can recover \(f\) from \(f^*\) exactly when \(f\) is convex, just as you can recover \(f\) from \(\hat{f}\) when \(f \in L^2\) (or \(L^1\)). The convexity condition is the tropical analog of the \(L^p\) condition.

The structure is exact:

5. The Full Dictionary 🔑

The complete correspondence between classical harmonic analysis and tropical convex analysis:

The tropical characters \(x \mapsto \langle x, \xi\rangle\) (linear/affine functions) play the same role in tropical harmonic analysis that complex exponentials \(x \mapsto e^{i\langle x, \xi\rangle}\) play in classical harmonic analysis: they are the “eigenfunctions” of tropical convolution operators.

6. Connections and Applications 💡

6.1 Legendre Transform in Physics and Thermodynamics

The Legendre transform is ubiquitous in physics precisely because physics operates in the \(\hbar \to 0\) limit. The correspondences:

• Thermodynamics: free energy \(F(T) = E - TS\) is the Legendre transform of entropy \(S(E)\) as a function of energy. The convolution theorem explains why free energies of independent subsystems add: their state densities convolve, so their free energies (Legendre transforms) add.
• Classical mechanics: the Hamiltonian \(H(q, p)\) is the Legendre transform of the Lagrangian \(L(q, \dot{q})\) in the velocity variable. This is the Maslov dequantization of the path integral (quantum mechanics, \(\hbar > 0\)) to Hamilton–Jacobi theory (classical mechanics, \(\hbar \to 0\)).
• Large deviations: the rate function \(I(\xi)\) of a large deviations principle is the Legendre transform of the log-moment generating function \(\Lambda(\lambda) = \log \mathbb{E}[e^{\lambda X}]\). This is exactly the tropical Laplace transform structure.

6.2 Tropical Characters and Affine Functions

In classical Fourier analysis, the characters \(\chi_\xi(x) = e^{i\langle x, \xi\rangle}\) satisfy: - They are group homomorphisms: \(\chi_\xi(x+y) = \chi_\xi(x)\chi_\xi(y)\) - They are eigenfunctions of translation: \(\tau_a \chi_\xi = e^{i\langle a,\xi\rangle} \chi_\xi\)

The tropical characters \(\ell_\xi(x) = \langle x, \xi\rangle\) (affine functions) satisfy: - Tropical homomorphism: \(\ell_\xi(x \odot y) = \ell_\xi(x) + \ell_\xi(y)\)… but \(x \odot y = x + y\) classically, so this just says linearity. - More precisely: \(\ell_\xi\) is a tropical monomial (monomial = linear function in the \(\log\)-coordinate world), and tropical monomials are the building blocks of tropical polynomials exactly as complex exponentials are the building blocks of classical functions via Fourier expansion.

The tropical Fourier expansion of a convex piecewise-linear function \(f\) decomposes it as a max of affine functions:

\[f(x) = \max_{\xi \in S}\bigl(\langle x, \xi\rangle + c(\xi)\bigr),\]

where \(S\) is the set of slopes (the subdifferential image of \(f\)) and \(c(\xi) = -f^*(\xi)\) is determined by \(f^*\). This is the tropical analog of the Fourier series.

6.3 Connection to Neuroalgebraic Geometry

ReLU neural networks compute tropical polynomials. A single ReLU unit computes \(\max(0, w \cdot x + b)\), which is a tropical polynomial in \((x, w, b)\)-space. A deep ReLU network computes iterated maxima and sums of linear functions — a composition of tropical polynomials — making its input-output map a piecewise-linear convex function on each activation region.

The consequence: the expressivity analysis of deep ReLU networks is a problem in tropical geometry. The number of linear regions of a depth-\(L\) width-\(N\) ReLU network equals (up to a polynomial factor) the number of vertices of a tropical variety — a tropical Grassmannian computation. This is why the expressivity-and-complexity note in Neuroalgebraic Geometry lists tropical geometry as a key tool.

The Legendre–Fourier analogy plays a role here too: the dual of a tropical polynomial (in the Legendre sense) encodes the Newton polytope of the corresponding classical polynomial, connecting tropical convexity to classical algebraic geometry via the Gelfand–Kapranov–Zelevinsky theory of \(A\)-discriminants.

7. References 📚