ADM Momentum Components in Relativity & Optimization

• ADM momentum components are the canonical variables derived from the ADM formalism that quantify gravitational energy-momentum in asymptotically flat spacetimes.
• They are computed via surface integrals at spatial infinity under precise decay conditions, playing a crucial role in defining conserved charges like energy and angular momentum.
• The framework extends to alternative formulations such as teleparallel gravity and is analogously applied in optimization algorithms to stabilize and accelerate convergence.

The Arnowitt–Deser–Misner (ADM) formalism decomposes spacetime into a family of spatial hypersurfaces, allowing one to express the gravitational field in canonical variables. The ADM momentum components, as well as related angular momentum and center-of-mass quantities, play a key role in characterizing the dynamical and conserved quantities of asymptotically flat spacetimes. Their rigorous construction, asymptotic conditions, and generalized extensions underpin both mathematical relativity and modern approaches to gravitational energy-momentum, including teleparallel and pseudotensor methods.

1. Canonical Definition of ADM Momentum Components

In the ADM decomposition, the spacetime metric is split as:

$$ds^2 = -N^2 dt^2 + g_{ij}(dx^i + N^i dt)(dx^j + N^j dt)$$

where $g_{ij}$ is the induced spatial metric, $N$ the lapse, and $N^i$ the shift vector. The canonical conjugate to $g_{ij}$ is the momentum density:

$\pi^{ij} = \sqrt{g}\,(K^{ij} - g^{ij}K)$

with $K_{ij}$ the extrinsic curvature of the hypersurface and $K = g^{ij} K_{ij}$ its trace (Golovnev, 2013, Fatibene et al., 2010, Kim et al., 2024). The momentum constraint, expressing diffeomorphism invariance, reads:

$\mathcal{H}_i = -2 D_j \pi^j{}_i = 0$

On shell, $\pi^{ij}$ is covariantly divergence-free: $D_j \pi^j{}_i = 0$.

The ADM spatial momentum component is defined through a surface integral at spatial infinity:

$$P_i = \frac{1}{16\pi} \lim_{r \to \infty} \oint_{S_r} \left( \pi^j{}_i - \delta^j{}_i\,\pi^\ell{}_\ell \right) dS_j$$

where $dS_j$ denotes the area element of the sphere at infinity (Fatibene et al., 2010, Formiga et al., 2022).

2. Surface Integrals and Asymptotic Flatness

The ADM formalism requires spacetime to be asymptotically flat:

• Metric: $\gamma_{ij}(x) = \delta_{ij} + O(r^{-1})$, $\partial_k \gamma_{ij} = O(r^{-2})$
• Extrinsic curvature: $K_{ij}(x) = O(r^{-2})$, $\partial_k K_{ij} = O(r^{-3})$

For the momentum component, the integral converges under these conditions. In terms of extrinsic curvature:

$$P_i = \frac{1}{8\pi} \lim_{r \to \infty} \oint_{S_r} (K_{ij} - \gamma_{ij}K) n^j dS$$

with $n^j$ the unit normal at infinity. The gauge is typically fixed to asymptotically Cartesian coordinates (Fatibene et al., 2010, Formiga et al., 2022). For energy, the ADM formula is:

$$E_{\rm ADM} = k \lim_{r\to\infty} \oint_{S_r} dS_i (\partial_j \gamma_{ij} - \partial_i \gamma_{jj}),\quad k = \frac{1}{16\pi G}$$

and similarly for momentum (Formiga et al., 2022).

3. Generalization, Covariance, and Teleparallel Formulations

Formiga & Gonçalves (Formiga et al., 2022) demonstrated that teleparallel energy-momentum generalizes the ADM result, with the teleparallel 4-momentum coinciding with the ADM 4-momentum under the time gauge and suitable coordinate conditions. In the teleparallel equivalent of GR (TEGR), the superpotential yields:

$$P^a_{\rm TEGR} = 4k \oint_{S} dS_\nu\, e\, \Sigma^{a0\nu}$$

With the Schwinger time gauge $e^{(0)}{}_i=0$, one finds:

$$P^{(0)}_{\rm TEGR} = E_{\rm ADM}, \quad P^{(k)}_{\rm TEGR} = P^k_{\rm ADM}$$

This equivalence holds whenever ADM integrals are well-defined. The teleparallel approach extends to non-asymptotically flat or non-Cartesian coordinates, such as Kruskal or FLRW cases, where ADM expressions fail.

4. ADM Angular Momentum and Center of Mass

The ADM angular momentum, for rotation Killing field $Y = x^i\partial_j - x^j\partial_i$, is given by:

$$J(Y) = \frac{1}{8\pi} \lim_{r \to \infty} \int_{S_r} \pi_{jk} Y^j \frac{x^k}{r} d\sigma_0$$

No parity condition is strictly necessary; the finiteness of $J(Y)$ is guaranteed under sharp decay/exansion assumptions on $\pi_{ij}$ (Chen et al., 2014). The Beig–Ó Murchadha–Regge–Teitelboim center of mass reads:

$$C^i = \frac{1}{16\pi} \lim_{r\to\infty} \int_{|x|=r} \Big[ x^i (\partial_k g_{jk} - \partial_j g_{kk}) \frac{x^j}{r} - (g_{ki}-\delta_{ki}) \frac{x^k}{r} + (g_{kk}-\delta_{kk}) \frac{x^i}{r} \Big] d\sigma_0$$

Explicit examples in Minkowski and Schwarzschild backgrounds, even without parity, produce finite nonzero $J$, $C$ with vanishing energy-momentum (Chen et al., 2014). Parity conditions are essential only for strict coordinate invariance.

5. Covariant Noether Current Interpretation and Coordinate Dependence

The ADM momentum can be interpreted via a Noether current approach, with the canonical density $\pi^{ij}$ forming the charge for spatial diffeomorphisms. For the translation generator $\xi^i$, integrating the momentum constraint yields the ADM momentum as the associated conserved quantity (Fatibene et al., 2010). The formal construction is coordinate-independent up to the gauge-fixing of asymptotic flatness and Cartesian reference frames. Under large diffeomorphisms which destroy the asymptotic Euclidean structure, the value of $P_i$ changes analogously to Lorentz boosts in special relativity (Fatibene et al., 2010).

6. ADM Momentum in Machine Learning and Optimization

In the context of optimization algorithms such as Adam and AdaMomentum (Wang et al., 2021), "momentum components" refer to exponential moving averages (EMAs) of gradients and their squares, which are used to adaptively scale the update steps. Adam maintains:

$m_t = \beta_1 m_{t-1} + (1-\beta_1)g_t$

$v_t = \beta_2 v_{t-1} + (1-\beta_2)g_t^2$

While AdaMomentum replaces $g_t^2$ with $m_t^2$ in the second moment:

$v_t = \beta_2 v_{t-1} + (1-\beta_2)m_t^2$

This "twofold EMA" smooths both the direction and magnitude estimates, stabilizing updates and improving generalization. The logic of using momentum in the denominator aligns with geometric stabilization in the convergence trajectory and promotes residence in flatter minima (Wang et al., 2021).

In ADMM-based optimization, "momentum components" appear as inertial extrapolations such as $\bar z_k = z_k + \alpha_k v_k$, with $v_k = z_k - z_{k-1}$. Adaptive, trajectory-following variants further generalize these components to multi-step polynomial extrapolations, achieving uniform acceleration even in spiral convergence regimes (Poon et al., 2019). The underlying mechanism reflects a geometric learning of local trajectory curvature via recent increments.

7. Conserved Charges, Symmetry Generators, and Ward Identities

Recent work (Kim et al., 2024) connects the ADM canonical momentum to generalized global shift symmetries on ADM hypersurfaces. A constant shift $$\delta g_{ij} = \partial_i \xi_j + \partial_j \xi_i$$—generated by $\pi^{ij}$—leads to a conserved vector charge:

$Q^j = -2 \oint_{S^2_\infty} d^2\Sigma_k\, \pi^{kj}$

Exponentiating $Q^j$ yields a unitary operator $U[\xi]$ whose action on Wilson line observables produces nontrivial phases. The fundamental Poisson bracket:

$$\{g_{ij}(x), \pi^{kl}(y)\} = \frac{1}{2}(\delta_i^k \delta_j^l + \delta_i^l\delta_j^k)\delta^{(3)}(x-y)$$

confirms that $\pi^{ij}$ truly generates the shift symmetry. The ensuing Ward identities reveal gapless Goldstone-type modes living on the hypersurface.

---

In summary, ADM momentum components serve as the pivotal link between gravitational canonical structure, conserved quantities at infinity, and generalized symmetries in both relativity and optimization settings. Their formalism is robust with respect to coordinate choices under suitable conditions but admits generalizations through teleparallel and covariant perspectives. In both physical and algorithmic contexts, the structure and role of canonical momentum are dictated by geometric, analytic, and symmetry properties encoded at the surface and in the dynamics of the system.