Two-Dimensional Delta Method

• Two-dimensional delta method is a framework that generalizes one-dimensional techniques to analyze simultaneous integral equations and joint estimator distributions.
• It employs smoothing kernels and oscillatory integrals with a Kloosterman refinement to achieve precise detection in analytic number theory.
• In probability theory, the method provides non-asymptotic convergence rates for multivariate estimators, enhancing statistical inference.

The two-dimensional delta method encompasses two major analytic frameworks appearing in contemporary research: (A) the two-dimensional delta symbol method in analytic number theory, offering a refined mechanism for detecting simultaneity of two integral equations; and (B) the multivariate (here, specifically two-dimensional) delta method in probability theory, providing precise normal approximations for the joint distribution of estimators, most prominently maximum likelihood estimators (MLEs). These approaches, which both generalize one-dimensional “delta” techniques, play central roles in asymptotic analysis, equidistribution, and explicit bounding in their respective domains.

1. Explicit Definition of the Two-Dimensional Delta Symbol

The two-dimensional delta symbol $\delta_{\mathbb{Z}^2}(n)$ is defined as

$$\delta_{\mathbb{Z}^2}(n) = \begin{cases} 1, & n = (0,0), \ 0, & \text{otherwise}. \end{cases}$$

Li–Rydin Myerson–Vishe provide a smoothed analytic representation: $$\delta_{\mathbb{Z}^2}(n) = \sum_{q=1}^Q\sum_{\substack{a \bmod q \ \gcd(a, q)=1}} \int_{\mathbb{R}^2} P_Q(a, q; \omega)\, e\big((a/q + \omega)\cdot n\big)\, d\omega + O(Q^{-N})$$ where $a \in (\mathbb{Z}/q\mathbb{Z})^2$, $e(x) = e^{2\pi ix}$, and $P_Q(a,q;\omega)$ is a smooth kernel depending only on the lattice $$\mathcal{A}(a, q) = \{k a + q y: k \in \mathbb{Z}, y \in \mathbb{Z}^2\}$$. For small $q$, $P_Q(a, q; \omega) = 1 + O(Q^{-N})$ whenever $q < Q^{1/2-\varepsilon}$ and $|\omega| < q^{-1} Q^{-1/2-\varepsilon}$. In general, $P_Q(a, q;\omega)$ decays rapidly: $$P_Q(a, q; \omega) \ll_{N, \varepsilon} \frac{1}{q} (1+q|\omega| Q^{1/2})^{-N}(1+|\omega| Q^{3/2})^{-N}$$ This representation underpins a powerful “Kloosterman refinement” of the circle method, critical for the simultaneous detection of two integral constraints in analytic number theory (Li et al., 2024).

2. Smoothing, Derivation, and Kernel Construction

The derivation starts from a dual-divisor identity relating $n=0$ in $\mathbb{Z}^2$ to one-dimensional delta-detection on inner products. Any nonzero $n$ is uniquely written as $n = d d' c^\perp$ where $c \in \mathbb{Z}^2$ is primitive, $d, d' \in \mathbb{N}$, and $c^\perp$ denotes the orthogonal rotation. The representation of $\mathbf{1}_{n=0}$ is achieved through a sum over $d, d', c$ of the one-dimensional delta symbol $\mathbf{1}_{c\cdot n=0}$. Smoothing is introduced via a compactly supported “bump” function $w$ with normalization $\int w = 1$; an auxiliary smoothing $h(y, z)$ is constructed from $w$. This yields, for any $N$: $\mathbf{1}_{n=0} = \sum_{\substack{d, d' \leq Q^{1/2}\d|n}} \sum_{\substack{q \leq Q/d \ \gcd(q, a, c)=1}} \frac{d}{q Q^3} \sum_{\substack{a \bmod q \ \gcd(a,q)=1}} \int_{\mathbb{R}} h\left(\frac{d}{Q^{3/2}}, \frac{c\cdot n}{qQ^{3/2}}\right)\, e\left(\frac{a}{q} \, c\cdot n\right) d(c \cdot n) + O(Q^{-N})$ A final Poisson inversion in $\mathbb{Z}^2$ recovers the explicit sum stated above. The kernel $P_Q(a, q; \omega)$, which emerges, is built from oscillatory integrals associated with these smoothings and admits a decomposition in terms of $p_{1,q}(\omega)$ and $p_{2,c,k,q}(\omega)$ components, each possessing rapid decay with respect to $q$ and $\omega$ (Li et al., 2024).

3. Analytic Properties and Multivariate Kloosterman Refinement

The key analytic feature in the classical (one-dimensional) delta symbol is the smooth partition of the unit circle: $\int_0^1 e(n \alpha)\, d\alpha$ with orthogonality $\int e(\alpha n)\, d\alpha = \mathbf{1}_{n=0}$ and a Poisson-type expansion around rationals. The two-dimensional method replaces the “circle” by the unit square $[0,1]^2$, with orthogonality recovered through summation over $a \bmod q$ and integration over $\omega \in \mathbb{R}^2$. The Kloosterman refinement leverages averaging over $a \mapsto Aa$ to exploit square-root cancellation in exponential sums, producing a double Kloosterman sum which allows for two layers of savings: $q^{-1/2}$ from averaging and an additional $q^{-1/2}$ from the $q$-summation. Grouping all $a$ with the same lattice $\mathcal{A}(a, q)$ is essential, and $P_Q(A a, q; \omega) = P_Q(a, q; \omega)$ for $A \in (\mathbb{Z}/q\mathbb{Z})^\times$ (Li et al., 2024).

Compared to nested one-dimensional methods, this true two-dimensional method achieves the “optimal” modulus size $Q \sim M^{2/3}$ for detecting vectors $|n| \sim M$. Concrete applications include bounding the number of variables necessary in counting integral points on the intersection of two quadrics: reducing from $s \geq 11$ (unconditional, one-dimensional) to $s \geq 10$ (unconditional, two-dimensional) and down to $s \geq 9$ under the Generalized Lindelöf Hypothesis.

4. Convergence, Decay, and Key Structural Lemmas

The convergence and effectiveness of the two-dimensional delta symbol method rely on several technical results:

• Dual–divisor detection (Lemma 3.1): The indicator $\mathbf{1}_{n=0}$ is represented as a smooth sum over divisors and a one-dimensional delta method applied to $c\cdot n$.
• Definition of the p-functions (Lemma 3.2): After Fourier inversion, the kernel decomposes into $p_{1,q}(\omega)$ and $p_{2,c,k,q}(\omega)$, each an explicit oscillatory integral with rapid decay in $q$ and $\omega$.
• Decay bounds (Lemmas 3.3–3.6): $$p_{1,q}(\omega) \ll_N q(1+|\omega| Q^{3/2})^{-N} (1+|\omega| qQ^{1/2})^{-N}$$; $p_{2,c,k,q}(\omega)$ similarly decays unless certain smallness conditions on $\omega$ are met.
• Geometry-of-numbers duality (Lemma 4.2): For a $3 \times 2$ matrix $M$ and shortest vector length $\lambda$ in $\mathcal{A}(a,q)$, $|Mx| \geq \lambda$ for $x \in \mathcal{A}(a, q)$ and $\operatorname{covol} \mathcal{A}(a, q) = q$, ensuring a uniform truncation and the domain for which the main-term in $p_{2,c,k,q}(\omega)$ persists (Li et al., 2024).

5. Multivariate Delta Method in Probability: The Two-Dimensional Case

A distinct but key application of the two-dimensional delta method appears in the asymptotic normal approximation of estimators, notably the MLE for two-dimensional parameters (Anastasiou et al., 2016).

Given i.i.d. data $X_1, ..., X_n \in \mathbb{R}^t$ with density $f(x|\theta)$ for $\theta\in\Theta\subset\mathbb{R}^2$, assume $\ell(\theta; x)$ is $C^3$ and the MLE $\hat\theta_n$ exists uniquely and can be written as $q(\hat\theta_n) = \sum_{i=1}^n g(X_i)$ for smooth $q$ and $g$. The Fisher information $I(\theta)$ is positive-definite and finite.

Define the scaled estimator and normal comparison: $$Z \sim N_2(0, I_2),\qquad S_{ij} = \sum_{k=1}^2 [K]_{jk}(g_k(X_i)-\mathbb{E}[g_k(X_1)]),\qquad K = Dq(\theta_0) I(\theta_0)^{-1} Dq(\theta_0)^{T}$$ For $h\in C^2(\mathbb{R}^2)$ with finite semi-norms, an explicit Berry–Esseen-type bound is given: $$|\mathbb{E} h(\sqrt{n}(\hat\theta_n - \theta_0)) - \mathbb{E} h(Z)| \leq (6.83\|h\|_2 + 12\|h\|_1)n^{-1/2}$$ This bound, notable for its independence from $(\mu, \sigma^2)$, applies to the classical case of the joint MLE $(\bar{X}, \widehat{\sigma}^2)$ for normal mean and variance, where $S_{i1}$ and $S_{i2}$ are explicitly computable. The result exemplifies the multivariate delta method’s capacity to produce non-asymptotic, explicit convergence rates for the joint law of two parameter estimators (Anastasiou et al., 2016).

6. Comparative Analysis and Applications

The two-dimensional delta symbol method significantly outperforms prior approaches for detecting the intersection of two equations when the number of variables is large. Heuristically, setting $Q \sim P^{4/3}$ (with $P$ the “height” parameter) and exploiting two layers of Kloosterman savings, the error terms decay as $P^{2(s-1)/3}$, which is a substantial improvement over one-dimensional, nested, or Gaussian integer approaches. The method is optimally suited to count integer solutions to pairs of quadratic forms, especially as the ambient number of variables increases (Li et al., 2024).

In probability theory, the two-dimensional delta method delivers explicit, non-asymptotic convergence rates in the normal approximation regime, relying on smoothness assumptions and “sum-of-i.i.d.” representation. Its effectiveness is most pronounced for vector-valued parameters or where joint inference is required, providing dimension-dependent but otherwise uniform rates (Anastasiou et al., 2016).

7. Context, Future Perspectives, and Methodological Extensions

The two-dimensional delta method in analytic number theory represents a comprehensive blueprint for future $R$-dimensional generalizations. Its principal analytic features, flexible smoothing, and explicit decay conditions enable broad applications to counting, Diophantine approximation, and equidistribution. In the context of probabilistic asymptotics, the multivariate delta method extends to arbitrary finite dimension $d$, underpinning key developments in maximum likelihood theory and statistical inference for dependent or functionally transformed estimators.

A plausible implication is that as both analytic and statistical models become more complex, true $R$-dimensional delta methods—avoiding nested or iterative constructions—will continue to provide sharper error estimates and more efficient detection or approximation capabilities. The growing divergence in performance as the dimensionality of the problem increases is a significant motivation for further methodological innovation (Li et al., 2024).