Energy-Flow Operators: Concepts & Applications

• Energy-flow operators are mathematical constructs that reveal pathways, budgets, and transfer laws governing energy dynamics across quantum fields, fluid systems, and computational networks.
• They are derived using diverse methodologies such as stress–tensor analysis, modal decomposition, and energy-constrained neural operator learning, each tailored to specific applications.
• Their practical applications include detector correlations in QFT, weak measurements in Schrödinger fields, reduced-order modeling in fluid dynamics, and sensitivity analysis in power systems.

Energy-flow operators are operator-valued or operator-theoretic constructions used to represent how energy is transported, redistributed, detected, or constrained. In the literature represented here, the term covers several non-identical formalisms: the energy detector operator at null infinity in quantum field theory, local energy-momentum-flow quantities derived from the Schrödinger field, modal transfer operators obtained from Galerkin projection of the Navier–Stokes equations, and operator-learning or optimization mappings whose defining property is energetic consistency or power-flow sensitivity (Chen et al., 2023, Hiley et al., 2014, Nakamura et al., 26 Mar 2025, Tanaka et al., 2024, Zhou et al., 2019). This suggests a common theme: energy-flow operators are not a single universal object, but a family of constructions that expose energy budgets, transfer pathways, and response structure in otherwise high-dimensional systems.

1. Scope and formal meanings

The main usages can be organized by the mathematical object on which the operator acts and by the notion of “flow” it resolves.

Domain	Operator or framework	Stated role
QFT and collider observables	$\mathcal E(\vec n)$	flux of the stress–energy tensor through null infinity
Schrödinger field theory	$T^{\mu\nu}$, especially $T^{00}$ and $T^{0i}$	energy density and energy-flux or momentum-density
Transient Navier–Stokes ROMs	$T_{ij\to k}$	energy moved from modes $i$ and $j$ into mode $k$
Operator learning for PDEs	$\mathcal S_\theta$ with $\mathcal H_\phi$	learned solution operator obeying energy conservation or dissipation
DC optimal power flow	$T^{\mu\nu}$0	mapping from user demand to generated power

In four-dimensional QFT, the energy–flow operator is defined directly from the stress–energy tensor. In the non-relativistic Schrödinger setting, the central objects are the components of the canonical energy–momentum tensor. In transient fluid dynamics, the operator language is attached to eigenmodes of the linearized Navier–Stokes operator and to the projected triadic coefficients governing modewise energy transfer. In machine learning for PDEs, the operator is a learned map between function spaces constrained by an energy law. In power networks, the operator viewpoint treats DC optimal power flow as a mapping whose Jacobian encodes sensitivity and robustness (Chen et al., 2023, Hiley et al., 2014, Nakamura et al., 26 Mar 2025, Tanaka et al., 2024, Zhou et al., 2019).

A common misconception is that the same phrase denotes the same object across these fields. The sources do not support that identification. Rather, they support several precise usages linked by the idea of extracting a structured energy budget or flow law from a more general dynamics.

2. Stress–tensor energy-flow operators in quantum field theory

In a four-dimensional QFT, for a null vector $T^{\mu\nu}$1, $T^{\mu\nu}$2, and its conjugate $T^{\mu\nu}$3, $T^{\mu\nu}$4, $T^{\mu\nu}$5, the energy–flow operator in direction $T^{\mu\nu}$6 is defined by the late-time limit of the flux of the stress–energy tensor through null infinity: $T^{\mu\nu}$7 Equivalently, in coordinates $T^{\mu\nu}$8,

$T^{\mu\nu}$9

This operator underlies detector energy flow correlations such as the Energy-Energy Correlator (EEC) (Chen et al., 2023).

When two energy detectors are nearly back-to-back, with

$T^{00}$0

the EEC obeys a transverse-momentum-dependent factorization formula at leading power: $T^{00}$1 with $T^{00}$2 and $T^{00}$3. Here $T^{00}$4 is the hard function, $T^{00}$5 and $T^{00}$6 are transverse-momentum-dependent jet functions, and $T^{00}$7 is the soft function. Solving the $T^{00}$8- and $T^{00}$9-RG equations resums the leading-power double logarithms $T^{0i}$0 to all orders (Chen et al., 2023).

The same work identifies the origin of logarithmically enhanced subleading-power corrections in the back-to-back limit. In the double-light-cone limit of a four-point Wightman correlator, the relevant singular structure is generated by the exchange of operators with low twists and large spins in the local operator product expansion. In conformal cross-ratios,

$T^{0i}$1

the back-to-back limit $T^{0i}$2 maps to $T^{0i}$3, $T^{0i}$4, and the correlator is expanded in superconformal blocks

$T^{0i}$5

The enhanced divergences beyond the usual single logarithm arise from the tail of large-spin operators at fixed low twist, explicitly the twist-2 tower with anomalous dimensions

$T^{0i}$6

Large-spin perturbation theory and twist conformal blocks then give an all-order resummation of the leading and next-to-leading logarithms beyond the leading power. The resulting closed form agrees with the fixed-order two- and three-loop EEC in $T^{0i}$7 SYM (Chen et al., 2023). In this setting, the energy–flow operator is not a reduced model or a learned map; it is a detector operator built from $T^{0i}$8, and its significance lies in converting asymptotic energy measurements into analytically tractable correlation functions.

3. Energy-momentum flow in the Schrödinger field

For a non-relativistic Schrödinger field with Lagrangian density

$T^{0i}$9

the canonical energy–momentum tensor is

$T_{ij\to k}$0

The components central to energy flow are

$T_{ij\to k}$1

and

$T_{ij\to k}$2

Writing $T_{ij\to k}$3 and $T_{ij\to k}$4, one obtains

$T_{ij\to k}$5

Local conservation is expressed by

$T_{ij\to k}$6

These relations identify $T_{ij\to k}$7 as energy density and $T_{ij\to k}$8 as energy-flux or momentum-density (Hiley et al., 2014).

The same structure appears from three derivations. From the Schrödinger equation,

$T_{ij\to k}$9

substitution of $i$0 yields the continuity equation

$i$1

and the quantum Hamilton–Jacobi equation

$i$2

The extra term

$i$3

is the quantum potential energy. From Lagrangian field theory, the same quantities appear directly in the canonical tensor. From the von Neumann–Moyal algebra, one introduces the Weyl operator $i$4, the Moyal star-product,

$i$5

and the Moyal and Baker brackets. The one-particle Wigner–Moyal distribution evolves according to

$i$6

whose projection onto configuration space reproduces the continuity equation and the quantum Hamilton–Jacobi equation (Hiley et al., 2014).

A distinctive feature of this framework is its connection to weak values. For the momentum operator at position $i$7,

$i$8

Hence

$i$9

while

$j$0

The sources emphasize that $j$1 and $j$2 are not operator eigenvalues and cannot be accessed by strong measurements; weak measurement techniques provide the proposed empirical route. The examples given—a free Gaussian wave packet, the infinite well, and the hydrogen $j$3 state—show how drift kinetic energy, quantum potential, and stationary states appear in the local energy-flow picture (Hiley et al., 2014).

4. Modal energy-transfer operators in transient Navier–Stokes dynamics

In the operator-driven reduced-order model of Nakamura et al., one begins with a prescribed base flow $j$4 and linearizes the incompressible Navier–Stokes equations about it: $j$5 Its formal adjoint is

$j$6

with eigenpairs

$j$7

Because $j$8 is non-self-adjoint, right and left eigenmodes satisfy the bi-orthogonality relation

$j$9

This bi-orthogonality is the key algebraic ingredient that makes a modal energy budget possible (Nakamura et al., 26 Mar 2025).

The full velocity field is expanded as

$k$0

Galerkin projection onto an adjoint mode $k$1 gives

$k$2

with

$k$3

$k$4

$k$5

For modal energy $k$6, differentiation yields

$k$7

In the stated interpretation, $k$8 is the net production or extraction of energy from the base flow, $k$9 is a net diffusion term, and $\mathcal S_\theta$0 is the nonlinear triadic transfer among modes (Nakamura et al., 26 Mar 2025).

The transfer operator is introduced in the simplest two-mode interaction as

$\mathcal S_\theta$1

which physically represents energy moved from modes $\mathcal S_\theta$2 and $\mathcal S_\theta$3 into mode $\mathcal S_\theta$4. Specializing to self-transfer gives

$\mathcal S_\theta$5

This furnishes a modewise decomposition of growth, damping, and cascade structure.

To treat strongly nonlinear transients, the same work introduces time-varying dynamic mode decomposition with a phase-control strategy for multiple time-series datasets from numerical simulations of the phase-controlled initial flow. The phase-controlled ensemble is defined by perturbations

$\mathcal S_\theta$6

followed by $\mathcal S_\theta$7 Navier–Stokes simulations and the instantaneous base flow

$\mathcal S_\theta$8

At each $\mathcal S_\theta$9, data matrices

$\mathcal H_\phi$0

are formed, exact DMD is applied via

$\mathcal H_\phi$1

global modes are reconstructed as

$\mathcal H_\phi$2

adjoint modes are recovered by suitable bi-orthonormalization, and amplitudes follow from

$\mathcal H_\phi$3

The time-varying modal-energy equation then takes the form

$\mathcal H_\phi$4

with time-dependent nonlinear triadics, production, and dissipation (Nakamura et al., 26 Mar 2025).

The illustrative examples are two-dimensional cylinder flow at $\mathcal H_\phi$5, and transients at $\mathcal H_\phi$6. Around the steady wake, the dominant eigenmode $\mathcal H_\phi$7 draws energy from the base flow through

$\mathcal H_\phi$8

is opposed by viscosity through

$\mathcal H_\phi$9

and experiences a small negative self-transfer

$T^{\mu\nu}$00

The spatial map of $T^{\mu\nu}$01 concentrates just upstream of the cylinder wake recirculation, and as $T^{\mu\nu}$02 increases, the location of peak transfer moves downstream linearly with the recirculation length. In the fully transient regime, the growth rate $T^{\mu\nu}$03 decreases from its linear value to zero as the wake saturates; nonlinear cascades $T^{\mu\nu}$04 remain negative and peak when $T^{\mu\nu}$05 is maximal; and at saturation $T^{\mu\nu}$06 of the energy pulled from the mean is eventually lost by viscous diffusion, with the remainder going into higher harmonics (Nakamura et al., 26 Mar 2025).

5. Energy-consistent neural operators

The Energy-consistent Neural Operator framework treats operator learning itself as an energy-constrained construction. Let $T^{\mu\nu}$07 denote the space of input functions and $T^{\mu\nu}$08 the space of solution functions on the space–time domain $T^{\mu\nu}$09. The true solution operator is

$T^{\mu\nu}$10

and ENO approximates it by a differentiable neural network $T^{\mu\nu}$11, so that

$T^{\mu\nu}$12

Architectures are agnostic—MLP, DeepONet, FNO, and others may be used—provided $T^{\mu\nu}$13 and $T^{\mu\nu}$14 are available through automatic differentiation (Tanaka et al., 2024).

The energetic structure is introduced through a total-energy functional

$T^{\mu\nu}$15

Examples given are the Korteweg–de Vries equation with

$T^{\mu\nu}$16

and the Cahn–Hilliard equation with

$T^{\mu\nu}$17

For a PDE of the form

$T^{\mu\nu}$18

Theorem 1 states that energy is conserved if $T^{\mu\nu}$19 is skew-symmetric and dissipated if $T^{\mu\nu}$20 is negative semidefinite (Tanaka et al., 2024).

ENO introduces the joint loss

$T^{\mu\nu}$21

where

$T^{\mu\nu}$22

and

$T^{\mu\nu}$23

The “energy net” $T^{\mu\nu}$24 takes as input the local state and its spatial derivatives and produces a scalar density $T^{\mu\nu}$25, giving

$T^{\mu\nu}$26

Its functional derivative is computed through automatic differentiation: $T^{\mu\nu}$27 By driving $T^{\mu\nu}$28, the learned operator satisfies in the continuous limit the correct gradient flow $T^{\mu\nu}$29, hence exactly conserving or dissipating $T^{\mu\nu}$30 (Tanaka et al., 2024).

The training algorithm initializes $T^{\mu\nu}$31 and $T^{\mu\nu}$32, samples minibatches of training pairs, samples $T^{\mu\nu}$33 query points, computes $T^{\mu\nu}$34, uses auto-diff to obtain temporal and spatial derivatives, computes $T^{\mu\nu}$35 and $T^{\mu\nu}$36, forms $T^{\mu\nu}$37 and $T^{\mu\nu}$38, and updates parameters by Adam, with optional early stopping on validation loss. Hyperparameters include $T^{\mu\nu}$39 and the maximum derivative order in $T^{\mu\nu}$40.

The reported experiments consider 1D KdV on $T^{\mu\nu}$41 and 1D Cahn–Hilliard on $T^{\mu\nu}$42, with data generated at $T^{\mu\nu}$43 by structure-preserving discretization and a Dormand–Prince ODE solver. In super-resolution tasks, training uses downsampled $T^{\mu\nu}$44, $T^{\mu\nu}$45, and $T^{\mu\nu}$46 data, while testing uses the full $T^{\mu\nu}$47 grid. Metrics are trajectory MSE, energy-MSE, and mass-MSE; baselines are Vanilla NO and DeepONet; and ENO achieves up to $T^{\mu\nu}$48 lower trajectory error, $T^{\mu\nu}$49 lower energy-drift, and correct mass conservation or dissipation, especially in low-resolution regimes (Tanaka et al., 2024). In this usage, the operator does not directly measure a flux. Instead, it is a learned solution map whose admissibility is tied to an energy law.

6. Operator viewpoints in power systems and cross-domain significance

A distinct operator-theoretic usage appears in DC optimal power flow. The framework treats the OPF problem as a mapping

$T^{\mu\nu}$50

from a load vector $T^{\mu\nu}$51 and network parameters

$T^{\mu\nu}$52

to the unique optimal generator output $T^{\mu\nu}$53, with network topology and line susceptances fixed (Zhou et al., 2019). The DC-OPF is written in generation–angle form, with generator injections $T^{\mu\nu}$54, loads $T^{\mu\nu}$55, bus voltage angles $T^{\mu\nu}$56, and line power flows $T^{\mu\nu}$57, and the associated KKT system provides the stationarity and complementary-slackness conditions.

The operator perspective is made precise by identifying parameter sets under which the mapping is single-valued and differentiable. The sets $T^{\mu\nu}$58 and $T^{\mu\nu}$59 encode, respectively, uniqueness of the OPF solution and independent binding constraints, and for

$T^{\mu\nu}$60

the optimal solution $T^{\mu\nu}$61 is well-defined, single-valued, and $T^{\mu\nu}$62 in $T^{\mu\nu}$63 away from a finite union of hyperplanes in load space. Fixing a load at which exactly $T^{\mu\nu}$64 inequalities are active, one forms the full-rank active-constraint Jacobian $T^{\mu\nu}$65, from which the Jacobian of the OPF operator is obtained in closed form: $T^{\mu\nu}$66 This Jacobian is constant on each “system-pattern region,” so $T^{\mu\nu}$67 is exactly affine in $T^{\mu\nu}$68 inside each such region (Zhou et al., 2019).

The same paper describes computation of derivatives through the homogeneous self-dual embedding. Writing the embedding operator as $T^{\mu\nu}$69, the sensitivity $T^{\mu\nu}$70 to perturbations $T^{\mu\nu}$71 obeys

$T^{\mu\nu}$72

after which the perturbation in generator output is recovered by projection. The resulting operator view supports sensitivity, worst-case robustness, privacy bounds through the Lipschitz constant, and derivatives of locational marginal prices (Zhou et al., 2019).

Taken together, these sources show that energy-flow operators occupy several technical niches. In QFT, they are detector operators built from $T^{\mu\nu}$73. In quantum mechanics, they are local energy-momentum-flow quantities tied to the quantum Hamilton–Jacobi equation and weak values. In transient fluid dynamics, they are modal transfer operators and energy budgets derived from bi-orthogonal projection. In neural operator learning, they are solution operators regularized to satisfy conservation or dissipation laws. In power systems, the operator language captures how demand perturbations propagate through constrained network optimization. This suggests that the unifying content of the term is structural rather than ontological: an energy-flow operator is a device for turning energetic behavior into a mathematically explicit object—an observable, a reduced transfer law, a differentiable map, or a learned dynamics—on which analysis can be performed.