The duality-based computability framework of Hoffman and Johnson (Encyclopedia of Computational Mechanics, 2004) shows empirically that time-averaged outputs of turbulent incompressible flow, such as mean drag on a bluff body, are computable to a few percent at feasible cost because the stability factors of the associated linearized dual (adjoint) problems are moderate, while instantaneous outputs appear non-computable. No theory explains this: exponential-in-time estimates predict astronomically large stability factors, the observed cancellation among the oscillatory dual coefficients is captured by no theoretical technique, the existence and mesh-tolerance theory rests on a boundedness assumption on the discrete velocity, and the class of flows and outputs with moderate stability factors is delimited only by computation. The open problem is to establish that missing rigorous theory — connecting Navier–Stokes turbulence statistics to dual stability and delineating the computable output class — which would answer which outputs of a turbulent flow can be computed to a given tolerance with guaranteed accuracy at feasible cost.

Background

The incompressible Navier–Stokes (NS) equations are the fundamental model for everyday fluid flow, from water pipes to aircraft. Their solutions are conventionally characterized by the Reynolds number Re=UL/νRe = UL/\nu (the ratio of inertial to viscous forces, with UU and LL a characteristic speed and length and ν\nu the kinematic viscosity); beyond ReRe of a few thousand most flows become turbulent — chaotic, multi-scale motion whose smallest active scales shrink rapidly as ReRe grows. Direct numerical simulation (DNS), which resolves every scale down to the viscous dissipation scale, then needs a number of space-time degrees of freedom that grows roughly like Re3Re^3 (the source chapter's estimate), far beyond available computing for engineering-scale turbulent flow. The practical alternatives — Reynolds-averaged (RANS) and large-eddy simulation (LES) closure models, which average or model the unresolved scales — reduce cost but carry modeling errors without a definitive theoretical justification, and none of these approaches says when a computed output is actually accurate. The source chapter calls this the principal open problem of computational fluid dynamics (CFD): whether turbulent flows at very high ReRe can be simulated with guaranteed accuracy in chosen outputs.

One route to guaranteed accuracy is output-oriented a posteriori error estimation, in which the error of a computed solution is bounded using the computation itself rather than a priori assumptions. In the dual-weighted residual framework (Becker and Rannacher, 2001), the error in a chosen output functional — e.g., the drag of a body — is bounded by the discrete residual weighted by the solution of a linearized dual (adjoint) problem: the sensitivity generator for the output, obtained by solving the linearized NS equations backward in time with coefficients frozen at the computed (primal) flow. The dual stability factor, denoted SεS_{\varepsilon} in the source, is the norm (gain) of this dual solution; it measures how strongly residual errors propagate into the output, and the a posteriori error bound it produces is only as good as the available bound on SεS_{\varepsilon}. Classical exponential-in-time (Grönwall-type) estimates for dual problems linearized at a turbulent solution predict astronomically large stability factors, which would make every output non-computable — a prediction contradicted by what is actually observed.

Hoffman and Johnson built a computability program on this duality. Their G2 method (general Galerkin) is a stabilized finite element method on space-time meshes whose residual can be made small in a weak norm; an ε\varepsilon-weak solution is a velocity–pressure field whose residual is at most ε\varepsilon in a weak dual norm (a relaxation of the classical weak-solution concept designed to remain meaningful for turbulent flow), and EG2 is G2 applied to the Euler equations with a slip-with-friction boundary condition (friction coefficient β\beta), used to study inviscid-limit phenomena such as the drag crisis (the abrupt drop in a bluff body's drag as the boundary layer becomes turbulent near Re∼105Re \sim 10^5; a bluff body is one with flow separation, such as a cylinder or sphere). Computationally they found that for time-averaged outputs of bluff-body turbulent flow the computed dual stability factors are moderate (order 1010–10210^2, growing slowly with resolution), so mean drag and lift are computable to a few percent on modest resources, whereas for instantaneous outputs the factors grow rapidly with resolution, indicating non-computability. The moderate size is attributed empirically to cancellation among the oscillatory coefficients of the dual problem; the chapter states that such effects seem impossible to account for in purely theoretical estimates, so the stability factors entering real error bounds are themselves obtained by computing the dual problem, without guarantees.

Two further gaps are acknowledged in the same framework. First, the existence proof for ε\varepsilon-weak discrete solutions and the ε\varepsilon–hh relations, which connect the weak-solution tolerance ε\varepsilon to the mesh size hh, assume the discrete maximum velocity M(U)M(U) stays bounded (or grows more slowly than h−1/2h^{-1/2}); removing this assumption is unresolved and borders on the open global existence/regularity problem for 3D NS. Second, the class of flows, output functionals, boundary conditions, linearization points and dual time-sampling strategies for which SεS_{\varepsilon} is moderate has been delineated only by systematic computation, not by theory.

A neighboring literature has grown up around the associated question of adjoint sensitivity in chaotic systems: ensemble-adjoint methods, rigorous linear-response results available only for uniformly hyperbolic systems (those with SRB — Sinai–Ruelle–Bowen — invariant measures), and least-squares shadowing techniques that compute long-time-average sensitivities without the exponential divergence of naive adjoints. A separate line certifies rigorous two-sided bounds on time-averaged quantities for specific flows via the background method (which bounds time-averaged turbulent transport by energy-stability arguments on fluctuations around a steady 'background' flow field) and semidefinite programming. None of these supplies the requested theory for Navier–Stokes turbulence or guaranteed output-error bounds for turbulent simulation.

Problem Statement

The open problem, posed by Hoffman and Johnson (2004), is the missing rigorous theory of dual (adjoint) stability for high-Reynolds-number incompressible turbulent flow within the ε\varepsilon-weak/G2 duality-based framework. Here an ε\varepsilon-weak solution is a velocity–pressure field whose Navier–Stokes residual is at most ε\varepsilon in a weak dual norm, and G2 (general Galerkin) is the stabilized finite-element discretization used to compute such solutions. The linearized dual (adjoint) problem is −ϕ˙−u⋅∇ϕ+(∇u)Tϕ−νΔϕ=ψ-\dot{\phi} - u \cdot \nabla\phi + (\nabla u)^T \phi - \nu \Delta \phi = \psi on [0,T^)[0, \hat{T}) with terminal condition ϕ(T^)=0\phi(\hat{T}) = 0, solved backward in time with the convection/reaction coefficients frozen at the computed ε\varepsilon-weak (G2) solution U^\hat{U} and with data ψ^\hat{\psi} encoding the chosen output functional M(⋅)M(\cdot). Following the source's companion book (Hoffman and Johnson, 2007), the dual stability factor is the norm ratio S(ψ^)=∥ϕ∥/∥ψ^∥S(\hat{\psi}) = \|\phi\|/\|\hat{\psi}\| of dual solution to dual data, with the dual data normalized so that ∥ψ^∥=1\|\hat{\psi}\| = 1, where the specific norms are determined by the output error representation and may differ between ϕ\phi and ψ^\hat{\psi}; on meshes it is computed as Sh(U^;ψ^)=∥ϕ^h∥S_h(\hat{U}; \hat{\psi}) = \|\hat{\phi}_h\| and extrapolated toward the continuum factor S0(ψ^)=S(ψ^)S_0(\hat{\psi}) = S(\hat{\psi}), the quantity denoted SεS_{\varepsilon} in the source chapter and below. It enters the a posteriori output estimate ∣M(u)−M(U^)∣≤(ε+CU hR(U^)) Sh(U^;ψ^)|M(u) - M(\hat{U})| \le (\varepsilon + C_U\, h R(\hat{U}))\, S_h(\hat{U}; \hat{\psi}) with stopping criterion hR(U^) S0(ψ^)≤TOLh R(\hat{U})\, S_0(\hat{\psi}) \le \mathrm{TOL}. Following the source's own criterion, S0(ψ^)S_0(\hat{\psi}) counts as large when S0(ψ^)≫ν−1/2S_0(\hat{\psi}) \gg \nu^{-1/2}, and as moderate when it does not — in particular when ∥ϕ^h∥\|\hat{\phi}_h\| grows slowly (near-constantly) under mesh refinement. The problem's scientific objective is to make dual-based a posteriori output-error bounds for turbulent flow guaranteed rather than empirically estimated — that is, to answer the chapter's core question: which outputs of a turbulent flow can be computed to a given tolerance with feasible computational cost? As in the source's framing, bounds are sought per flow and per output (estimating the size of SεS_{\varepsilon} for the data corresponding to different outputs); no uniformity in ReRe is imposed.

(1) Explain the moderate size of dual stability factors: prove bounds on SεS_{\varepsilon} for mean-value (time-averaged) outputs of high-Reynolds-number turbulent flow in terms of the flow's statistical properties, beating exponential-in-time (Grönwall-type) estimates.

(2) Delineate, with guarantees, the class of flows, output functionals, boundary conditions, linearization points and dual time-sampling strategies for which SεS_{\varepsilon} is moderate, separating computable outputs (such as long-time averages) from non-computable ones (such as instantaneous values).

The two deliverables serve one objective: with (1) the observed computability of mean outputs becomes a theorem instead of an empirical regularity, and with (2) the boundary of that computability is known in advance rather than discovered by computation.

Scope note: founding the ε\varepsilon-weak existence/ε\varepsilon–hh theory without the discrete-velocity boundedness assumption lies outside this problem's completion boundary: the program already claims existence of ε\varepsilon-weak G2 solutions conditionally on the assumption that the discrete maximum velocity M(U)M(U) remains bounded (or grows more slowly than h−1/2h^{-1/2}; Hoffman and Johnson, 2007, Ch. 16), and removing that assumption unconditionally borders on the open global existence/regularity problem for 3D NS. The same chapter also acknowledges as open the first-principles calibration of the wall-model friction coefficient β\beta, the validation of EG2's limiting behavior as Re→∞Re \to \infty for general geometries, and the resource-budgeting question of the minimal feasible ε\varepsilon; these are adjacent open issues with independent completion boundaries, outside this problem's completion boundary.

References

  1. Hoffman, J., Johnson, C. (2004). Computability and Adaptivity in CFD. In: Encyclopedia of Computational Mechanics, Wiley. DOI: 10.1002/0470091355.ecm057.pub2. https://doi.org/10.1002/0470091355.ecm057.pub2
  2. Hoffman, J., Johnson, C. (2006). A new approach to computational turbulence modeling. Computer Methods in Applied Mechanics and Engineering. DOI: 10.1016/j.cma.2004.09.015. https://doi.org/10.1016/j.cma.2004.09.015
  3. Hoffman, J., Johnson, C. (2006). Stability of the dual Navier–Stokes equations and efficient computation of mean output in turbulent flow using adaptive DNS/LES. Computer Methods in Applied Mechanics and Engineering. DOI: 10.1016/j.cma.2005.05.034. https://doi.org/10.1016/j.cma.2005.05.034
  4. Hoffman, J., Johnson, C. (2007). Computational Turbulent Incompressible Flow. Springer. DOI: 10.1007/978-3-540-46533-1. https://doi.org/10.1007/978-3-540-46533-1
  5. Hoffman, J. (2021). Energy stability analysis of turbulent incompressible flow based on the triple decomposition of the velocity gradient tensor. Physics of Fluids. DOI: 10.1063/5.0060584. https://doi.org/10.1063/5.0060584
  6. Becker, R., Rannacher, R. (2001). An optimal control approach to a posteriori error estimation in finite element methods. Acta Numerica. DOI: 10.1017/S0962492901000010. https://doi.org/10.1017/S0962492901000010
  7. Lea, D. J., Allen, M. R., Haine, T. W. N. (2000). Sensitivity analysis of the climate of a chaotic system. Tellus A. DOI: 10.3402/tellusa.v52i5.12283. https://doi.org/10.3402/tellusa.v52i5.12283
  8. Eyink, G. L., Haine, T. W. N., Lea, D. J. (2004). Ruelle's linear response formula, ensemble adjoint schemes and Lévy flights. Nonlinearity. DOI: 10.1088/0951-7715/17/5/016. https://doi.org/10.1088/0951-7715/17/5/016
  9. Wang, Q. (2013). Forward and adjoint sensitivity computation of chaotic dynamical systems. Journal of Computational Physics. DOI: 10.1016/j.jcp.2012.09.007. https://doi.org/10.1016/j.jcp.2012.09.007
  10. Wang, Q., Hu, R., Blonigan, P. (2014). Least Squares Shadowing sensitivity analysis of chaotic limit cycle oscillations. Journal of Computational Physics. DOI: 10.1016/j.jcp.2014.03.002. https://doi.org/10.1016/j.jcp.2014.03.002
  11. Ni, A., Wang, Q. (2017). Sensitivity analysis on chaotic dynamical systems by Non-Intrusive Least Squares Shadowing (NILSS). Journal of Computational Physics. DOI: 10.1016/j.jcp.2017.06.033. https://doi.org/10.1016/j.jcp.2017.06.033
  12. Ni, A., Talnikar, C. (2019). Adjoint sensitivity analysis on chaotic dynamical systems by Non-Intrusive Least Squares Adjoint Shadowing (NILSAS). Journal of Computational Physics. DOI: 10.1016/j.jcp.2019.06.035. https://doi.org/10.1016/j.jcp.2019.06.035
  13. Chandramoorthy, N., Fernandez, P., Talnikar, C., Wang, Q. (2017). An Analysis of the Ensemble Adjoint Approach to Sensitivity Analysis in Chaotic Systems. 23rd AIAA Computational Fluid Dynamics Conference. DOI: 10.2514/6.2017-3799. https://doi.org/10.2514/6.2017-3799
  14. Fantuzzi, G., Wynn, A. (2016). Optimal bounds with semidefinite programming: An application to stress-driven shear flows. Physical Review E. DOI: 10.1103/PhysRevE.93.043308. https://doi.org/10.1103/PhysRevE.93.043308
  15. Slotnick, J., Khodadoust, A., Alonso, J., Darmofal, D., Gropp, W., Lurie, E., Mavriplis, D. (2014). CFD Vision 2030 Study: A Path to Revolutionary Computational Aerosciences. NASA/CR-2014-218178. https://ntrs.nasa.gov/citations/20140003093
  16. Jansson, J., Johnson, C., Scott, R. (2022). Predictive Euler CFD - Resolution of NASA Vision 2030. AIAA AVIATION 2022 Forum. DOI: 10.2514/6.2022-3589. https://doi.org/10.2514/6.2022-3589