We analyze two recently-introduced flow measured that are based on a single trajectory only: trajectory stretching exponent (TSE) to detect hyperbolic (stretching) behavior, and trajectory angular velocity (TRA) to detect elliptic (rotation) behavior. Haller et al. [2021] and Haller et al. [2022] introduced TSE, TRA as well as the concept of quasi-objectivity, and formulated theorems about the objectivity and quasi-objectivity of TSE and TRA. In this paper, we present two counter-examples showing that all theorems in Haller et al. [2021] and Haller et al. [2022] are incorrect.
Introduction
Recently, Haller et al. [ [1] and Haller et al. [ [2] introduced measures based on a single trajectory only. For this, the concept of quasi-objectivity is introduced: Contrary to classical objectivity where a scalar value must be invariant under arbitrary time-dependent Euclidian transformations, for quasi-objectivity a condition (A) is introduced, and invariance is not demanded for all Euclidean transformations but only for those fulfilling (A). Then, Haller et al. [ [1] introduced several measures based on a single trajectory: extended trajectory stretching exponents TSE and TSE, and extended trajectory angular velocity TRA, TRA. Haller et al. [ [1] claimed that TSE and TSE are objective in the extended phase space, and that TRA is quasi-objective in the extended phase space under a certain condition put to the average vorticity in a certain neighborhood of the trajectory. The new measures have been applied in several follow-up papers: Encinas-Bartos et al. [ [3], Aksamit and Haller [ [4].
In this paper, we show that the claims by Haller et al. [2021] concerning objectivity of TSE, TSE, TRA are incorrect. In fact, we show by a counter-example that neither TSE nor TSE are objective in the extended phase space. Further, TRA is not quasi-objective in the extended phase space under an averaged-vorticity-based condition.
In an erratum, Haller et al. [ [2] reformulate claims about quasi-objectivity of TSE, TRA. We show by a second counter-example that the new claims in the erratum Haller et al. [ [2] are incorrect as well.
In summary, our two counter-examples show that all theorems in Haller et al. [2021] and Haller et al. [ [2] are incorrect.
We note that the first counter-example was already published in Theisel et al. [2022], and Haller et al. [2022] was published as a reaction on Theisel et al. [2022].
TSE, TRA, and objectivity
Objectivity, a concept from continuum mechanics, refers to the invariance of a measure under a moving reference system. Let s(x,t), w(x,t), T(x,t) be a time-dependent scalar field, vector field and tensor field, respectively. Further, let s(x,t), w(x,t), T(x,t) be their observations under the Euclidean frame change
x=Q(t)x+b(t)(1)
where Q=Q(t) is a time-dependent rotation tensor and b(t) is a time-dependent translation vector. Then s, w, T are objective if the following conditions hold, cf. Truesdell and Noll [1965]:
Haller et al. [2021] introduced measures for stretching and rotation that are based on single trajectories only: Extended trajectory stretching exponents TSE, TSE, and extended trajectory angular velocity TRA, TRA. Single-trajectory flow measures are attractive because they need minimal information to infer the flow behavior of an underlying field. Obviously, single-trajectory measures cannot be objective in the Euclidean observation space because one may think of a reference system moving with the trajectory, making each trajectory zero [1]. Because of this, Haller et al. [2021] considered objectivity in an extended phase space.
Given is a C2 continuous trajectory x(t) for t∈[t0,tN], its first and second derivatives x˙(t), x¨(t), and a positive constant v0 accounting for a certain ratio between space and time units to make them non-dimensionalized. Considering x(t) in an extended phase space gives for the first and second derivative of a trajectory x(t):
x˙(t)=(v01x˙(t)1),x¨(t)=(v01x¨(t)0).(3)
Then a local stretching measure can be defined as
tse=tsex(t),v0(t)=x˙Tx˙x˙Tx¨=xTx+v02x˙Tx¨(4)
from which the Lagrangian measures TSE and TSE are computed by integrating tse along the trajectory:
with Δt=tN−t0. The discretization in Eq. (6) samples x(t) at N+1 time steps t0<t1<…<tN.
For defining TRA, the (n+1)-dimensional matrix function
tra=trax(t),v0(t)=x˙Tx˙x˙x¨T−x¨x˙T(7)
can be introduced that describes the local angular velocity. Note that tra is an anti-symmetric matrix, from which one gets by integration along the trajectory Lagrangian measures
where Fr denotes the Frobenius norm of a matrix. Haller et al. [2021] claimed that TSE and TSE are objective in the extended phase space, and that TRA and TRA are quasi-objective in the extended phase space under a certain condition put to the average vorticity in a certain neighborhood of the trajectory.
TSE, TRA, and underlying velocity fields
We recapitulate the definition of TSE from Haller et al. [ [1], keeping their notation as much as possible. We start with a single observed trajectory x(t) in n-D (n=2,3) for t∈[t0,tN] running from x0=x(t0) to xN=x(tN). Further, we assume that x(t) is a trajectory (path line) of an underlying unsteady velocity field v(x,t), i.e., x˙(t)=dtdx=v(x(t),t) for all t∈[t0,tN]. Following Haller et al. [2021], v is transformed into a non-dimensionalized field u by
y=Lx,τ=τ0+Tt−t0,v0=TL(12)
where L, T, v0 are certain positive constants for a field that need to be determined by additional knowledge about the data. Generally, the scaling factor v0 is non-zero, i.e., v0=0. This transformation rephrases x(t) into the non-dimensionalized trajectory
y(τ)=L1x(t0+T(τ−τ0))(13)
running from y0=y(τ0)=L1x0 to yN=y(τN)=L1xN with τN=τ0+TtN−t0. Further, it gives the non-dimensionalized vector field
u(y,τ)=v01v(Ly,t0+T(τ−τ0)).(14)
Note that (14) contains a correction of a missing term v01 in formula (26) in [Haller et al., 2021]. The error in formula (26) in [1] can be seen in the following way: suppose v is a constant vector field, i.e., v(x,t)=vc. Then formula (26) in [1] would give u(y,τ)=vc no matter how v0 is chosen. This would contradict to the formula before (33) in [Haller et al., 2021].
Following Haller et al. [2021] further, an extended phase space Y=(yz) is introduced by adding the additional time dimension z. Transformation of y(τ) and u(y,τ) into this extended phase space gives
Y(τ)=(y(τ)τ),U(Y)=(u(y,z)1)(15)
where Y(τ) is the trajectory in the extended phase space running from Y0=Y(τ0)=(y0τ0) to YN=Y(τN)=(yNτN), and U(Y) is the underlying vector field. The tangent vector of Y(τ) is
Note that U(Y) is an autonomous dynamical system now: U is a steady velocity field in the extended phase space. Then Haller et al. Haller et al. [2021] defines TSE and TRA as
where Δt=tN−t0, (17) is identical to the right-hand side of (5), and (18) is identical to the right-hand side of (9). To show objectivity of TSE in the extended phase space, one has to prove that TSE is invariant under observation in any moving Euclidean reference system in the extended phase space. Analogous to Eq. (1), such moving reference system is defined by
Y=Q(τ)Y+B(τ),Q(τ)=(Q(τ)0T01),B(τ)=(b(τ)0)(20)
with Q(τ)∈SO(n) being a rotation matrix, and 0 being the zero-vector. The observed trajectory Y(τ) and the underlying velocity field U(Y,τ) in the new moving reference system are
Y(τ)=QT(τ)(Y(τ)−B(τ))(21)
U(Y,τ)=QT(τ)(U(Q(τ)Y+B(τ))−Q˙(τ)Y−B˙(τ))(22)
where the new trajectory Y(τ) runs from Y0=Y(τ0) to YN=Y(τN). Then, TSE in the moving reference system is
To show objectivity of TSE in the extended phase space, one has to prove TSE=TSE for any moving reference frame, as given by Eq. (20). To show quasi-objectivity of TRA under averaged-vorticity condition, one has to prove TRA=TRA for all reference frames (20) in which the averaged-vorticity condition is fulfilled.
The first counter-example
We show the non-objectivity of TSE in the extended phase space by a simple counter-example. We set the 2D observed trajectory x(t) and the underlying velocity field v(x,t) as
x(t)=(et−1t(t+1)),v(x,t)=(x+12t+1)(25)
for t∈[t0,tN]=[0,1] and x=(x,y)T. To calculate TSE as in Eq. (5), we only need information at time t0 and tN. This gives
Analogously, TRA follows by inserting (40)–(41) into (24). Since there is no positive constant v0, cf. (12), that makes TSE in (35) and TSE in (42) identical for any bc=(xc,yc)T, non-objectivity of TSE in the extended phase space is shown. Since in our example both U(Y) and U(Y,τ) have zero vorticity, the average-vorticity condition in Haller et al. [Haller et al., 2021] is trivially fulfilled. Thus, the difference of TRA and TRA gives that TRA is not quasi-objective in the extended phase under the average-vorticity condition.
Where is the error?
Haller et al. Haller et al. [2021] considered a non-zero vector ξ0 at (x0,t0) that is advected with v along x(t), resulting in
ξ˙(t)=∇v(x(t),t)ξ(t),ξ(t0)=ξ0.(43)
Then, ξ(t) is observed under a moving reference system (1). Objectivity of ξ is deduced from (43), (1):
ξ(t)=QT(t)ξ(t)(44)
where ξ is the observation of ξ under the moving reference system (1). From (44) follows the objectivity of Δt1ln∣ξ0∣∣ξ(tN)∣. We note that (44) follows from (43) and (1) only if another implicit assumption holds: objectivity of the seeding vector ξ0, i.e., ξ0=QT(t0)ξ0.
The approach of Haller et al. Haller et al. [2021] is to set ξ0=v0=v(x0,t0). With this, additional conditions are necessary to ensure
v˙(t)=∇v(x(t),t)v(x,t)(45)
v(x,t)=QT(t)v(x,t)(46)
where (45) corresponds to (43) and (46) corresponds to (44). To ensure (45), Haller et al. Haller et al. [2021] introduced the condition
∂tv(x,t)=0(47)
in the current observation frame. However, condition (A1) does not ensure (46) because ξ0=v0 is not objective. Since the observation of v under the moving reference system (1) is [Haller, 2021]
v(x,t)=QT(t)(v(x,t)−Q˙(t)x−b˙(t))(48)
Eq. (46) is in general only fulfilled for Q˙=0,b˙=0, i.e., the reference frame is not moving but static, resulting in demanding that v(x,t) is steady. This means that the condition for the quasi-objectivity of TSE is the steadiness of both v and v in all considered reference frames. We remark that this is a rather strong condition for quasi-objectivity: it excludes the consideration of all moving reference frames.
The transformation to the extended reference system transforms v to the steady vector field U, making the condition (A1) for (45) in the extended reference frame obsolete. However, the observation U of U under a moving reference system (20) is not a steady vector field anymore, as shown in (22). This means that
U(Y,τ)=Q(τ)U(Y)(49)
does not hold in general but only for particular steady reference frames. Because of this, TSE is in the extended phase space not objective but only quasi-objective under restriction to a static reference system.
Summary: The error was to assume that the observation of an autonomous system (steady vector field) in the extended phase space under a moving reference frame remains an autonomous system.
Remarks: A similar argumentation gives that TSE is not objective in the extended phase space, and and that TRA is not quasi-objective in the extended phase space under the averaged-vorticity-based condition. TSE, TSE, TRA and TRA are not even Galilean invariant because the moving reference system (36) in the counterexample was performing a Galilean transformation.
The second counter-example
In an erratum, Haller et al. Haller et al. [2022] give up the idea of considering extended phase space and non-dimensionalization, and introduce a new condition
∣∂tv(x(t),t)∣≪∣x¨(t)∣.(50)
Based on this, Theorem 3 in Haller et al. [2021] is reformulated. Unfortunately, the reformulated theorem is still incorrect. To show this, we consider a second counter-example. We consider the 3D trajectory and the underlying velocity field
x(t)=et−t0,v(x,t)=x−10(51)
with x=(x,y,z)T and t running from t0=0 to tN=1. Observing x,v under a reference system (1) with
Q(t)=I,b(t)=(0,t,0)T(52)
gives
x(t)=et−2t0,v(x,t)=x−20(53)
with x=(x,y,z)T. We note that
∂tv(x,t)=∂tv(x,t)=ω(x,t)=ω(x,t)=0(54)
x¨(t)=x¨(t)=(et,0,0)T(55)
where ω,ω are the vorticity of v,v, respectively. This means that all conditions (A1),(A2) (from Haller et al. [2021]) and (A3) (from Haller et al. [2022]) are fulfilled both in the original frame in (50) and in the particular moving reference frame in (52). Computing TSE, TSE, TRA, TRA from (50), and computing TSE,TSE,TRA,TRA from (52) gives
TSE=TSE=2ln(e2+1)−ln(2)≈0.71689(56)
TSE=TSE=2ln(e2+4)−ln(5)≈0.41161(57)
TRA=TRA=tan−1(e)−4π≈0.43288(58)
TRA=TRA=tan−1(e/2)−tan−1(1/2)≈0.47282(59)
For TSE being quasi-objective under condition (A3) following Haller et al. [2021], there must be an objective scalar P such that
TSE≈P,TSE≈P(60)
where P is the observation of P under reference frame (51) and “the accuracy of the approximation indicated by the symbol ≈ depends on the extend to which the conditions (A3) is satisfied” Haller et al. [2021]. Note that here assumption (A3) is completely satisfied, since the value ∣∂tv(x(t),t)∣/∣x¨(t)∣ as considered in Figure 1 of Haller et al. [2022] is constant zero in both reference frames, see Eq. (54)–(55). Since P=P due to the demanded objectivity of P and since assumption (A3) is completely satisfied, Eqs. (56) and (57) give that (59) cannot be fulfilled since TSE and TSE are significantly different. Hence, TSE is not quasi-objective under condition (A3).
Eqs. (56)–(59) give that similar statements hold for all measures TSE, TSE, TRA, TRA: none of them is quasi-objective under any combination of conditions (A1), (A2), (A3). This example and the example from § 3 show that all theorems in both Haller et al. [2021] and Haller et al. [2022] are incorrect.
References
[1]George Haller, Nikolas Aksamit, and Alex P. Encinas Bartos. Quasi-objective coherent structure diagnostics from single trajectories. Chaos: An Interdisciplinary Journal of Nonlinear Science, 31(4):043131, 2021. doi:10.1063/5.0044151. URL https://doi.org/10.1063/5.0044151.
[2]George Haller, Nikolas Aksamit, and Alex P. Encinas Bartos. Erratum: "Quasi-objective coherent structure diagnostics from single trajectories" [Chaos 31, 043131 (2021)]. Chaos: An Interdisciplinary Journal of Nonlinear Science, 32(5):059901, 2022. doi:10.1063/5.0090124. URL https://doi.org/10.1063/5.0090124.
[3]Alex P. Encinas-Bartos, Nikolas O. Aksamit, and George Haller. Quasi-objective eddy visualization from sparse drifter data. Chaos: An Interdisciplinary Journal of Nonlinear Science, 32(11):113143, 2022. doi:10.1063/5.0099859. URL https://doi.org/10.1063/5.0099859.
[4]Nikolas O. Aksamit and George Haller. Objective momentum barriers in wall turbulence. Journal of Fluid Mechanics, 941:A3, 2022. doi:10.1017/jfm.2022.316.DOI
Gerard M Leteba, Yi-Chi Wang, Thomas J A Slater, Rongsheng Cai, Conor Byrne, Christopher P Race, David R G Mitchell, Pieter B J Levecque, Neil P Young, Alex Walton et al. ·