Bernstein–Sato Polynomials and Zeta Functions of Singularities

Table of Contents

• 1. Overview
• 2. The Gelfand Zeta Function
  • 2.1 Definition and Motivating Integral
  • 2.2 Meromorphic Continuation via Resolution
  • 2.3 The State Density Function and Mellin Transform
• 3. The Bernstein–Sato Polynomial
  • 3.1 The Functional Equation
  • 3.2 Existence Theorem (Bernstein 1972)
  • 3.3 Kashiwara’s Rationality Theorem
  • 3.4 Worked Examples
• 4. The D-Module Perspective
  • [[#4.1 The Module D[s] · f^s|4.1 The Module D[s] · f^s]]
  • 4.2 The Kashiwara–Malgrange V-Filtration
• 5. Zeta Function Poles and the b-Function
  • 5.1 Poles of the Gelfand Zeta Function
  • 5.2 The RLCT as the Largest Pole
  • 5.3 Saito’s Theorem
• 6. Monodromy and the b-Function
  • 6.1 Vanishing Cycles and the Milnor Fiber
  • 6.2 Monodromy Eigenvalues and Roots of b_f
• 7. Connection to Singular Learning Theory
  • 7.1 K(w) as the Defining Function
  • 7.2 Reading Off the RLCT from the b-Function
  • 7.3 Worked Example: Rank-1 Matrix Factorization
• 8. The Asymptotic Pipeline
• 9. References

1. Overview 📐

The Gelfand zeta function and the Bernstein–Sato polynomial (also called the b-function) are two faces of the same algebraic-analytic object: the local singularity theory of an analytic function \(f\) near its zero set. Their interplay with the real log-canonical threshold (RLCT) is the mathematical engine driving Watanabe’s singular learning theory (SLT).

The logical chain is:

flowchart TD
    A["Gelfand zeta function
ζ(z) = ∫|f|^{2z}φ dw"]
    B["Meromorphic continuation
via Hironaka resolution"]
    C["Poles of ζ(z)
at negative rationals"]
    D["Bernstein-Sato polynomial
b_f(s) f^s = P f^{s+1}"]
    E["Largest root of b_f
= minus RLCT λ"]
    F["Bayes free energy
F_n ~ nS_n + λ log n"]

    A --> B
    B --> C
    D --> E
    C --> E
    E --> F

The bridge from line 1 (analysis: poles of \(\zeta\)) to line 2 (algebra: roots of \(b_f\)) is Kashiwara’s theorem (1976). The bridge to statistics is Watanabe’s free energy theorem, treated in Singular Learning Theory. This note develops the algebraic geometry side.

2. The Gelfand Zeta Function 📐

2.1 Definition and Motivating Integral

Let \(f: (\mathbb{R}^d, 0) \to (\mathbb{R}, 0)\) be a real-analytic function defined near the origin, and let \(\varphi \in C_c^\infty(\mathbb{R}^d)\) be a smooth compactly supported test function with \(\varphi \geq 0\).

Definition (Gelfand Zeta Function). The Gelfand zeta function associated to \(f\) and \(\varphi\) is

\[\zeta_f(z) = \int_{\mathbb{R}^d} |f(w)|^{2z}\, \varphi(w)\, dw, \qquad \mathrm{Re}(z) > 0.\]

For \(\mathrm{Re}(z) > 0\) the integral converges absolutely (since \(|f|^{2z}\) is locally integrable for \(\mathrm{Re}(z) > 0\)). The factor \(2z\) rather than \(z\) is conventional: with \(f\) real, \(|f|^{2z} = (f^2)^z\), making contact with the complex-variable zeta function \(\int (f^2)^z \varphi\,dw\).

The poles of \(\zeta_f\) encode the singularity structure of \(f\) near its zero set \(\{f = 0\}\). In the regular case (\(df \neq 0\) on \(\{f=0\}\)), \(\zeta_f\) has no poles at all in \(\{-1 < \mathrm{Re}(z) \leq 0\}\). Poles appear precisely when \(f\) has singularities.

2.2 Meromorphic Continuation via Resolution

The cleanest proof of meromorphic continuation uses Hironaka’s resolution of singularities (established in full generality in 1964). Let \(g: \tilde{U} \to U\) be a resolution of singularities of \(\{f = 0\}\): a proper birational map such that

\[f \circ g = u \cdot \prod_{i=1}^r y_i^{k_i}\]

where \(u\) is a unit (nowhere zero) and \((y_1, \ldots, y_r, y_{r+1}, \ldots, y_d)\) are local coordinates on \(\tilde{U}\) forming a normal crossing divisor — the exceptional locus \(\bigcup_i \{y_i = 0\}\) has only transverse intersections.

After the change of variables \(w = g(y)\), the integral becomes

\[\zeta_f(z) = \int_{\tilde{U}} \left|u \prod_{i=1}^r y_i^{k_i}\right|^{2z} (\varphi \circ g)\, |\det Dg|\, dy.\]

Since \(u\) is a unit, \(|u|^{2z}\) is smooth and nonzero. The Jacobian \(|\det Dg|\) is a smooth function times a monomial \(\prod_i |y_i|^{a_i - 1}\) (where \(a_i\) counts the multiplicity of the exceptional divisor \(E_i\) in the jacobian). So the integrand is

\[\text{(smooth, nonzero)} \times \prod_i |y_i|^{2k_i z + a_i - 1}.\]

By Fubini (integrating out one coordinate at a time), each factor \(\int |y_i|^{2k_i z + a_i - 1} d y_i\) is a beta-type integral that meromorphically continues to all of \(\mathbb{C}\) with simple poles at

\[2k_i z + a_i - 1 = -1, -2, -3, \ldots \quad \Longrightarrow \quad z = -\frac{a_i + j}{2k_i},\quad j = 0, 1, 2, \ldots\]

Thus the poles of \(\zeta_f\) are contained in \(\left\{-\dfrac{a_i + j}{2k_i} : i = 1,\ldots,r,\ j \in \mathbb{Z}_{\geq 0}\right\}\), a discrete set of negative rationals.

2.3 The State Density Function and Mellin Transform

In Watanabe’s SLT, \(f\) is replaced by the KL divergence \(K(w) = \int q(x) \log \frac{q(x)}{p(x|w)}\, dx \geq 0\), and \(\varphi = \varphi(w)\) is the prior density. The zeta function of the model is

\[\zeta(z) = \int_W K(w)^z\, \varphi(w)\, dw.\]

This is the Mellin transform of the state density function

\[\nu(t) = \frac{d}{dt} \int_{\{K(w) \leq t\}} \varphi(w)\, dw,\]

the density of prior mass at KL level \(t\):

\[\zeta(z) = \int_0^\infty t^{z-1}\, \nu(t)\, dt = \mathcal{M}[\nu](z) \cdot \Gamma(z)^{-1} \cdot \Gamma(z).\]

The asymptotic expansion of \(\nu(t)\) near \(t = 0\) (governing generalization) is read directly from the poles of \(\zeta(z)\) via the inverse Mellin transform. This is the analytic pathway from algebraic geometry to Bayesian asymptotics.

3. The Bernstein–Sato Polynomial 📐

3.1 The Functional Equation

The Bernstein–Sato polynomial (or b-function) provides an algebraic route to the meromorphic continuation of \(|f|^{2z}\) that bypasses the geometric resolution.

Definition (Bernstein–Sato Polynomial). Let \(f \in \mathbb{C}[x_1, \ldots, x_n]\) (or a convergent power series). The Bernstein–Sato polynomial \(b_f(s) \in \mathbb{C}[s]\) is the monic polynomial of minimal degree for which there exists a differential operator \(P(s, x, \partial_x) \in \mathbb{C}[s]\langle x, \partial_x \rangle\) satisfying the functional equation

\[\boxed{b_f(s)\, f(x)^s = P(s, x, \partial_x)\cdot f(x)^{s+1}.}\]

Here \(f^s\) is treated formally as a symbol: the equation is an identity of distributions on \(\mathbb{C}^n\), or equivalently a relation in the \(\mathcal{D}[s]\)-module \(\mathcal{D}[s] \cdot f^s\) (see §4). The operator \(P\) is allowed to depend polynomially on \(s\).

3.2 Existence Theorem (Bernstein 1972)

Theorem (Bernstein 1972). For every polynomial \(f \in \mathbb{C}[x_1, \ldots, x_n]\), the Bernstein–Sato polynomial \(b_f(s)\) exists and is non-zero.

The proof is algebraic: it works entirely within the Weyl algebra \(D_n = \mathbb{C}\langle x_1, \ldots, x_n, \partial_1, \ldots, \partial_n\rangle\) and exploits the fact that \(D_n[s] \cdot f^s\) is a finitely generated \(D_n[s]\)-module (a consequence of the Noetherianness of \(D_n\)). The functional equation then follows from the annihilator being non-trivial.

3.3 Kashiwara’s Rationality Theorem

Theorem (Kashiwara 1976). All roots of \(b_f(s)\) are negative rational numbers.

Kashiwara’s proof uses the resolution of singularities and the theory of \(\mathcal{D}\)-modules (specifically, the structure of nearby cycles). After pulling back to normal-crossing coordinates, the functional equation reduces to monomial cases, for which the roots are explicitly rational (as in the monomial example above).

Key consequence: The poles of \(\zeta_f(z) = \int |f|^{2z}\varphi\, dw\) are contained in

\[\left\{ -\frac{\alpha}{2} : b_f(-\alpha) = 0,\ \alpha \in \mathbb{Q}_{>0} \right\} \cup \text{(poles from jacobian of resolution)}.\]

In fact, the poles of \(\zeta_f\) are precisely of the form \(-\alpha/2\) where \(e^{2\pi i \alpha}\) is an eigenvalue of the monodromy on nearby cycles — this is the Monodromy Theorem.

3.4 Worked Examples

The monomial case is the most computationally useful. For \(K(w) = w_1^{2h_1} \cdots w_r^{2h_r}\) (even exponents, as arises from squared-residual KL terms):

\[b_K(s) = \prod_{j=1}^r \prod_{i=1}^{2h_j} \left(s + \frac{i}{2h_j}\right),\]

and the RLCT is

\[\lambda = \min_j \frac{1}{2h_j}.\]

For example, \(K(w) = w_1^2 w_2^2\) (a rank-1 matrix factorization residual) has \(h_1 = h_2 = 1\), so \(b_K(s) = (s+\frac{1}{2})(s+1)(s+\frac{1}{2})(s+1) = (s+\frac{1}{2})^2(s+1)^2\), and \(\lambda = 1/2\).

4. The D-Module Perspective 📐

4.1 The Module D[s] · f^s

The cleanest algebraic home for the b-function is the theory of \(\mathcal{D}\)-modules (modules over the sheaf of differential operators). Fix the Weyl algebra \(D = D_n = \mathbb{C}\langle x_1, \ldots, x_n, \partial_1, \ldots, \partial_n\rangle\) with the fundamental commutation relation \([\partial_i, x_j] = \delta_{ij}\).

Definition (D[s]-module of \(f^s\)). Let \(D[s] = D \otimes_\mathbb{C} \mathbb{C}[s]\). Define

\[M_f = D[s] \cdot f^s,\]

the cyclic \(D[s]\)-module generated by the formal symbol \(f^s\). The action is: \(x_i\) acts by multiplication, \(\partial_i\) acts by \(\partial_i(f^s) = s\, (\partial_i f)\, f^{s-1}\), and \(s\) acts by shifting the exponent.

The Bernstein–Sato polynomial \(b_f(s)\) is then characterized as:

Definition (D-module characterization). \(b_f(s)\) is the minimal polynomial of \(s\) acting on the quotient

\[M_f \big/ f \cdot M_f = D[s] \cdot f^s \big/ D[s] \cdot f^{s+1}.\]

The functional equation \(b_f(s) f^s = P \cdot f^{s+1}\) says exactly that \(b_f(s)\) annihilates the class \([f^s]\) in this quotient.

4.2 The Kashiwara–Malgrange V-Filtration

The V-filtration (Kashiwara 1983, Malgrange 1983) is a canonical filtration of \(D\)-modules adapted to a hypersurface \(\{f = 0\}\) that simultaneously encodes: - the nearby cycles functor \(\psi_f\) - the vanishing cycles functor \(\phi_f\) - the roots of \(b_f\) as eigenvalues of \(s\) on associated graded pieces

For our purposes, the key fact is: the roots of \(b_f(s)\) in the interval \((-1, 0]\) are in bijection with the eigenvalues of the monodromy acting on the Milnor fiber of \(f\) (up to the exponential map \(\alpha \mapsto e^{2\pi i \alpha}\)). This is the content of the next section.

5. Zeta Function Poles and the b-Function 📐

5.1 Poles of the Gelfand Zeta Function

The functional equation \(b_f(s) f^s = P \cdot f^{s+1}\) can be iterated: applying it \(k\) times yields

\[b_f(s) b_f(s-1) \cdots b_f(s-k+1)\, f^s = Q_k(s, x, \partial)\cdot f^{s+k},\]

for some differential operator \(Q_k\). This shows that \(|f|^{2s}\) (as a distribution-valued function of \(s\)) can be meromorphically continued past each pole of \(\prod_{j=0}^{k-1} b_f(s-j)^{-1}\), with potential poles at the roots of \(b_f(s-j) = 0\), i.e., at \(s = \alpha + j\) where \(b_f(\alpha) = 0\) and \(j \in \mathbb{Z}_{\geq 0}\).

Conclusion: The poles of \(\zeta_f(z) = \int |f|^{2z} \varphi\, dw\) are contained in

\[\bigcup_{\substack{\alpha \in \mathbb{Q}_{<0} \\ b_f(\alpha)=0}} \left\{\frac{\alpha}{2} - j : j \in \mathbb{Z}_{\geq 0}\right\}.\]

5.2 The RLCT as the Largest Pole

Theorem (Kashiwara–Saito). The largest pole of \(\zeta_f(z)\) — i.e., \(-\lambda\) where \(\lambda = \mathrm{RLCT}(f)\) — equals

\[-\lambda = \frac{\alpha_0}{2},\]

where \(\alpha_0\) is the largest root of \(b_f(s)\) (the root with smallest absolute value). Equivalently,

\[\boxed{\lambda = \mathrm{RLCT}(f) = -\frac{\alpha_0}{2} = -\frac{1}{2}\max\{\alpha \in \mathbb{Q}_{<0} : b_f(\alpha) = 0\}.}\]

5.3 Saito’s Theorem

More precisely, Saito (2007) established:

Theorem (Saito 2007). Let \(\alpha_0 = \max\{\alpha \in \mathbb{Q}_{<0} : b_f(\alpha) = 0\}\) be the largest root of \(b_f\). Then the real log-canonical threshold satisfies

\[\lambda = \mathrm{RLCT}(f;\varphi) = -\alpha_0\]

(in the convention \(\zeta_f(z) = \int |f|^z \varphi\, dw\)).

This is the key algebraic criterion: to compute the RLCT, one computes the b-function and reads off its largest root. For many architecturally natural functions \(K(w)\), the b-function can be computed via Gröbner basis methods (implemented in Macaulay2 or Singular).

6. Monodromy and the b-Function 📐

6.1 Vanishing Cycles and the Milnor Fiber

Near an isolated singularity of \(f\) at the origin, the Milnor fiber \(F_t = f^{-1}(t) \cap B_\varepsilon\) (for small \(|t| > 0\), \(\varepsilon > 0\)) is a smooth manifold homotopy equivalent to a wedge of \((n-1)\)-spheres. The number of such spheres is the Milnor number \(\mu\).

As \(t\) circles the origin in \(\mathbb{C}^*\), the fiber \(F_t\) undergoes a continuous deformation that returns to itself: this defines the monodromy operator

\[T: H_{n-1}(F_t; \mathbb{C}) \to H_{n-1}(F_t; \mathbb{C}).\]

Since \(T\) is quasi-unipotent (the Monodromy Theorem), its eigenvalues are roots of unity: \(e^{2\pi i \alpha}\) for \(\alpha \in \mathbb{Q}\).

6.2 Monodromy Eigenvalues and Roots of b_f

Theorem (Monodromy Theorem / Malgrange 1975). The roots of \(b_f(s)\) in the interval \((-n, 0)\) are exactly the numbers \(\alpha \in (-n, 0) \cap \mathbb{Q}\) such that \(e^{2\pi i \alpha}\) is an eigenvalue of the monodromy \(T\) on \(H_{n-1}(F_t; \mathbb{C})\).

Concretely: if \(e^{2\pi i \alpha}\) is a monodromy eigenvalue, then \(\alpha\) is a root of \(b_f\). The spectrum of the singularity \(\sigma(f)\) records all these \(\alpha\) (with multiplicity), and \(b_f\) is divisible by \(\prod_{\alpha \in \sigma(f)} (s - \alpha)\).

The Picard–Lefschetz formula gives the explicit action of monodromy on vanishing cycles: if \(\delta \in H_{n-1}(F_t; \mathbb{Z})\) is a vanishing cycle (a class that collapses to a point as \(t \to 0\)), then

\[T(\gamma) = \gamma \pm (\gamma \cdot \delta)\, \delta\]

for any class \(\gamma\), where \(\gamma \cdot \delta\) is the intersection form. The eigenvalues of \(T\) on all of \(H_{n-1}\) follow from this formula.

7. Connection to Singular Learning Theory 🔑

7.1 K(w) as the Defining Function

In Watanabe’s framework, the function playing the role of \(f\) is the KL divergence from the true distribution:

\[K(w) = \int_\mathcal{X} q(x) \log \frac{q(x)}{p(x \mid w)}\, dx \geq 0,\]

where \(q\) is the true data-generating distribution and \(p(\cdot \mid w)\) is the parametric model. The crucial properties are: - \(K(w) \geq 0\) everywhere - \(K(w) = 0\) if and only if \(p(\cdot \mid w) = q\) - The optimal parameter set \(W_0 = K^{-1}(0) = \{w : K(w) = 0\}\) is an analytic variety (in particular, not necessarily a single point)

For neural networks, \(K\) is a polynomial (or Nash) function in \(w\), making it amenable to b-function computation.

The RLCT of the model is defined as \(\lambda = \mathrm{RLCT}(K; \varphi)\), the RLCT of \(K\) with respect to the prior \(\varphi\). By the b-function–RLCT correspondence (Saito’s theorem):

\[\lambda = -\alpha_0 = -\max\{\alpha \in \mathbb{Q}_{<0} : b_K(\alpha) = 0\}.\]

7.2 Reading Off the RLCT from the b-Function

The practical procedure:

1. Write \(K(w)\) as a polynomial in the parameters \(w = (w_1, \ldots, w_d)\).
2. Compute \(b_K(s)\) — either algebraically (Gröbner basis methods in Macaulay2/Singular) or via Newton polyhedra (for monomial or near-monomial \(K\)).
3. Find the largest root \(\alpha_0 = \max\{\alpha : b_K(\alpha) = 0\}\) (most negative root in absolute value, but closest to 0).
4. Set \(\lambda = -\alpha_0\).
5. The multiplicity \(m\) of \(\alpha_0\) as a root of \(b_K\) gives the power of \((\log n)\) in the free energy asymptotics:

\[F_n = nS_n + \lambda \log n - (m - 1) \log \log n + O_p(1).\]

7.3 Worked Example: Rank-1 Matrix Factorization

Consider approximating a rank-\(r_0\) matrix \(A \in \mathbb{R}^{M \times N}\) with a rank-1 product \(w_1 w_2^\top\) where \(w_1 \in \mathbb{R}^M\), \(w_2 \in \mathbb{R}^N\). The squared Frobenius loss is

\[L(w_1, w_2) = \|A - w_1 w_2^\top\|_F^2.\]

In the realizable case (\(r_0 = 0\), i.e., \(A = 0\)), the KL divergence (for a Gaussian observation model with variance \(\sigma^2\)) is proportional to

\[K(w_1, w_2) = \frac{1}{2\sigma^2}\|w_1 w_2^\top\|_F^2 = \frac{1}{2\sigma^2} \|w_1\|^2 \|w_2\|^2.\]

Near any point where \(\|w_1\|\) and \(\|w_2\|\) are small, \(K\) locally behaves like a monomial in the norms. For the simplest case \(M = N = 1\) (scalar factorization \(K = w_1^2 w_2^2 / 2\sigma^2\), up to a constant):

\[K(w_1, w_2) \sim w_1^2 w_2^2.\]

By the monomial b-function formula:

\[b_K(s) = \prod_{i=1}^2\left(s + \frac{i}{2}\right)^2 = \left(s + \frac{1}{2}\right)^2 (s + 1)^2.\]

The largest root is \(\alpha_0 = -1/2\), so:

\[\boxed{\lambda = \mathrm{RLCT}(w_1^2 w_2^2) = \frac{1}{2}, \qquad m = 2.}\]

The free energy asymptotics is therefore

\[F_n = nS_n + \frac{1}{2}\log n - \log\log n + O_p(1),\]

a much smaller complexity penalty than BIC (\(= \frac{2}{2}\log n = \log n\), since we have 2 parameters). The singularity at \(w_1 = w_2 = 0\) — the origin of the non-identifiability — literally halves the effective model complexity.

For general rank \(r\) approximation of a rank-\(r_0\) matrix with \(r_0 < r\), Watanabe and Aoyagi computed

\[\lambda = \frac{r(M + N) - r^2 + \min(r_0, M + N - r) \cdot (2r - M - N + r_0)}{2} \cdot \frac{1}{\text{(appropriate normalization)}}.\]

This formula follows from a careful resolution of singularities of the determinantal variety \(\{\mathrm{rank}(W) \leq r_0\} \subset \mathbb{R}^{M \times N}\).

8. The Asymptotic Pipeline 🔑

This section collects the full chain from singularity data to statistical asymptotics.

Let \(\mathcal{M} = \{p(\cdot \mid w)\}\) be a statistical model with parameter space \(W \subset \mathbb{R}^d\), prior \(\varphi(w)\), and true distribution \(q\). Assume \(q \in \mathcal{M}\) (realizable case). Define:

• \(K(w) = \mathrm{KL}(q \| p(\cdot | w)) \geq 0\) — the KL divergence
• \(W_0 = K^{-1}(0)\) — the optimal parameter set (a real-analytic variety)
• \(\lambda = \mathrm{RLCT}(K; \varphi)\) — the real log-canonical threshold
• $m = $ the order of the pole of \(\zeta_K\) at \(-\lambda\) — the singular fluctuation

Watanabe’s Free Energy Theorem: As \(n \to \infty\),

\[F_n = nS_n + \lambda \log n - (m-1)\log\log n + O_p(1),\]

where \(S_n = -\frac{1}{n}\sum_{i=1}^n \log q(X_i)\) is the empirical entropy of the true distribution.

The b-function contribution to this pipeline:

flowchart LR
    A["Architecture
(neural network, matrix factorization)"]
    B["KL divergence K(w)
as analytic function"]
    C["b-function b_K(s)
(algebraic computation)"]
    D["Largest root alpha_0
RLCT lambda = -alpha_0"]
    E["Free energy
F_n ~ n S_n + lambda log n"]

    A --> B
    B --> C
    C --> D
    D --> E

9. References 📚