betaFeedback

WorksPhysics and Astronomy

Construction of approximate invariants for non-integrable Hamiltonian systems

Abstract

We present a method to construct high-order polynomial approximate invariants (AI) for non-integrable Hamiltonian dynamical systems, and apply it to modern ring-based particle accelerators. Taking advantage of a special property of one-turn transformation maps in the form of a square matrix, AIs can be constructed order-by-order iteratively. Evaluating AI with simulation data, we observe that AI's fluctuation is actually a measure of chaos. Through minimizing the fluctuations with control knobs in accelerators, the stable region of long-term motions could be enlarged.

Introduction

The construction of the invariant phase-space tori of Hamiltonian systems is important in many mathematical physics fields (see [3, 4, 5, 6, 7, 8] [1, 2] and references therein). For non-integrable Hamiltonian systems, exact solutions for motion are often impossible. They generally lack sufficient conserved quantities. However, Kolmogorov-Arnold-Moser theory [9] proves that approximate invariants (AI) could still exist in such systems. The theory addresses the stability of motion in near-integrable Hamiltonian systems that can be seen as small perturbations of integrable systems. If the perturbation is small enough, a significant portion of the invariant tori of the integrable system survive, though deformed. These surviving tori act as quasi-invariants or AIs, where the motion remains confined and quasi-periodic, preserving the structure of the integrable case. Even though resonances can disrupt some tori, many tori persist in a “cantorized” form, as long as certain non-resonance conditions are met. The remnants of the AI tori provide a structure that partially resists chaotic diffusion, preserving long-term stability in certain regions of phase space. Such region is also known as “dynamic aperture (DA)” in particle accelerator physics [10]. A sufficient DA is often required for beam manipulation and machine operation. It is important to explicitly obtain the AIs of given lattices and evaluate their fluctuations in designing and optimizing modern accelerator magnetic lattice [1, 2].

Method

The evolution of motion in a periodical Hamiltonian system can be approximately represented with polynomial maps. These maps can be extracted with the truncated power series algorithm [11, 12, 13] to any arbitrary order. For example, consider a 4-dimensional phase space vector [x,px,y,py]T[x,p_x,y,p_y]^{T}, which represents an on-momentum charged particle’s motion in the transverse plane in an accelerator. The superscript TT represents the transpose of matrices or vectors. First we construct a Ω\Omega-th order vector

Z=[x,px,y,py,x2,xpx,xy,xpy,px2,pxy,⋯ ,pyΩ]T.(1)Z=[x,p_x,y,p_y,x^2,xp_x,xy,xp_y,p_x^2,p_xy,\cdots,p_y^{\Omega}]^{T}. \tag*{(1)}

Let the one-turn map of the system be written in a matrix form as

Zfinal=MZinitial,(2)Z_{final}=MZ_{initial}, \tag*{(2)}

where MM is a square matrix. Such map is usually non-symplectic due to the truncation of high order polynomial terms. Following the method in Chao’s note [14], an Ω\Omega-order Taylor series approximate invariant WW can be written as

W=VTZ,(3)W=V^{T}Z, \tag*{(3)}

with VV as the coefficient vector yet to be found. Since WW is an invariant, for any ZZ after one-turn transformation through MM, we have

VTZ=VTMZ⟹V=MTV.(4)V^{T}Z=V^{T}MZ\Longrightarrow V=M^{T}V. \tag*{(4)}

It means that the vector VV needs to be MTM^{T}’s eigenvector with its eigenvalue as 1. However, when applying a direct eigen-analysis on the Ω(≥3)\Omega(\ge3)-order matrix MTM^{T}, it usually fails to yield proper AIs. It is because the coefficients of higher orders in MTM^{T} are significantly greater than the lower orders. More specifically, the magnitude incremental step at each nthn^{th} order is n!n!. Therefore, the highest order terms dominate the eigen-analysis, and the obtained eigen-vectors, usually have near to zero coefficients in the low order terms. It might yield multiple but independent eigen-vectors. To overcome this difficulty, we introduce a method to start with the well-known linear Wronskian invariants (aka Courant-Snyder invariants [15] for uncoupled linear motion) and then iteratively construct higher order AIs by taking advantage of the special property of the square matrix. In this paper, the National Synchrotron Light Source II (NSLS-II) storage ring [16] is used as a real-world example for illustration. Its 5th5^{th} one-turn square matrix MM is constructed by sequentially concatenated the magnetic elements as shown in Figure 1. MTM^{T} matrix is a sparse matrix with its elements in the top-right blocks are zeros, because any nthn^{th} order term in ZZ must be a function of the same and even-higher orders terms (≥nth\geq n^{th}), but not of the lower order terms (<nth< n^{th}), i.e.,

Image not extracted

Figure 1. Up to 5th-order blocked square matrix MTM^T. The first top-left 14×1414 \times14 block is used to construct linear AIs W1,2(2)W^{(2)}_{1,2}. Then the expanded 34×3434 \times34, 69×6969 \times69 and 125×125125 \times125, ⋯\cdots square matrices are used to obtain high order AIs W1,2(3,4,5,⋯ )W^{(3,4,5,\cdots)}_{1,2} respectively, and so on. Note that all elements in the top-right block’s 0(3−5)0^{(3-5)} are 0’s.

Zfinaln=∑i≥nmiZinitiali(5)Z^{n}_{final}=\sum_{i\geq n}m_i Z^{i}_{initial} \tag*{(5)}

This property actually enables us to construct AIs order-by-order iteratively.

We start with the linear invariants, then construct higher order AIs on top of them. Even in the linear case, ZZ must be kept up to the quadratic terms in order to find two quadratic invariants W1,2(2)W^{(2)}_{1,2}. Here superscript (2)(2) indicates the order of WW, and subscripts 1,2 are their indexes. First, a 14×1414 \times14 square matrix MTM^T truncated up to the quadratic terms is used. The eigen-analysis on MTM^T yields two eigen-vectors V1,2(2)V^{(2)}_{1,2} with their eigenvalues as 1:

V2_1                 | V2_2
---------------------|----------------------
2 0 0 0  4.883952e-02 | 0 0 2 0  2.969949e-01
0 2 0 0  2.047522e+01 | 0 0 0 2  3.367061e+00

Here each monomial is represented with the first four integers as the power indexes of xx, pxp_x, yy, pyp_y, and the last float as its coefficient. These two invariants can be easily recognized as the Courant-Snyder invariants in the horizontal and vertical planes respectively, Wu=x,y=γuu2+2αuupu+βupu2W_{u=x,y}=\gamma_u u^2+2\alpha_u u p_u+\beta_u p_u^2, with the Twiss parameters α\alpha, β\beta, γ\gamma. The disappearance of two crossing terms xpxx p_x and ypyy p_y terms is simply due to the local αx,y=0\alpha_{x,y}=0.

Next we construct cubic AIs by adding some additional cubic terms into linear invariants V=[V(2),V(3)]V=[V^{(2)},V^{(3)}], here V(3)V^{(3)} represents the cubic coefficients yet to be found. By observing the form of the cubic square matrix MTM^T (34×3434 \times34) in Figure 1, we notice that its upright blocks 0(3)0^{(3)} are composed of zeros, therefore, it can be divided into 4 submatrices

MT=[m00(3)m2m3].M^T=\begin{bmatrix}m_0 & 0^{(3)}\\m_2 & m_3\end{bmatrix}.

Here m0m_0 with a dimension (14×1414 \times14) is the quadratic matrix already being analyzed previously, 0(3)0^{(3)} (14×2014 \times20) is a zero matrix, m2m_2 (20×1420 \times14) and m3m_3 (20×2020 \times20) are two matrices with the cubic term coefficients. The new cubic AIs need to satisfy the following condition:

\begin{bmatrix}V^{(2)}\\V^{(3)}\end{bmatrix} = \begin{bmatrix}m_0 & 0^{(3)}\\m_2 & m_3\end{bmatrix} \begin{bmatrix}V^{(2)}\\V^{(3)}\end{bmatrix}. \tag*{(6)}\end{bmatrix}

Because the first row in Eq. (6) is already valid automatically, the cubic term coefficient V(3)V^{(3)} can be simply solved with:

V(3)=(I−m3)−1m2V(2).(7)V^{(3)}=(I-m_3)^{-1}m_2V^{(2)}. \tag*{(7)}

By solving Eq. (7), the cubic terms in two AIs are obtained as:

V3_1                         | V3_2
-----------------------------|----------------------
3 0 0 0  1.327349e-01        | 1 0 2 0  -4.703005e+00
2 1 0 0 -6.078392e-01        | 1 0 1 1  -1.778960e+00
...                           | ...
0 1 0 2  4.370392e+01        | 0 1 0 2  -5.597472e+01

The constructed cubic AI’s coefficients [V(2),V(3)]T[V^{(2)},V^{(3)}]^T can be compared with its one-turn transformation MT[V(2),V(3)]TM^T[V^{(2)},V^{(3)}]^T as shown in Figure 2. The maximum difference of coefficients is around 3×10−133 \times10^{-13}. It is worthwhile to mention that, the existence of the solution in Eq. (7) requires the square matrix (I−m3)(I-m_3) to be invertible. In some cases, this condition might not be satisfied, which means the appearance of broken order [14] due to a 3rd3^{rd} resonance. However, for most stable systems such as particle accelerators, their (I−m3)(I-m_3) usually invertible because their fundamental frequency is intentionally chosen to be off-resonant. Once the cubic AIs are obtained, we can iteratively apply Eq. (6) to obtain even higher order AIs.

Image not extracted

Figure 2. Comparison a 3rd3^{rd}-order AI’s coefficient [V(2)+V(3)][V^{(2)}+V^{(3)}] with its one-turn transformation MT[V(2)+V(3)]M^{T}[V^{(2)}+V^{(3)}]. Top: the values of all 34 monomial coefficients. Bottom: the difference between them represents the quality of the AI.

Application in lattice design

Once the AIs are obtained, we can verify them with simulation data. Here a simulated 128 turns trajectory starting with an initial condition x=6.4mmx=6.4\mathrm{mm}, y=2.6mmy=2.6\mathrm{mm}, px,y=0p_{x,y}=0 is obtained with a symplectic tracking algorithm [17]. After evaluating different order AIs W1,2(2,3,4,5)W^{(2,3,4,5)}_{1,2} with the simulated data, they are observed converging gradually with reduced fluctuations as illustrated in Figure 3, which indicates the accuracy of AIs is improved by including more higher order terms. However, the AIs’ quality becomes worse when their order Ω\Omega is above 5 in this specific example. Thus far, it is unclear if the higher-order AIs are spoiled by the accumulated numerical errors, or they simply don’t exist at all.

Image not extracted

Figure 3. Fluctuations of two AIs W1,2(2,3,4,5)W_{1,2}^{(2,3,4,5)} at different orders (2nd−5th2^{nd}-5^{th}) by evaluating with a simulated trajectory.

As seen in Figure 3, these AIs are only approximately constant, rather than exactly. After evaluating with simulated data, AI’s fluctuation increase gradually with oscillation amplitude as shown in Figure 4. The fluctuations of linear AIs (aka “the action smear”) are traditionally used as a chaos indicator in accelerator physics [18]. Once the fluctuations exceed a certain range, approximate tori are broken and the stable particle motion can not be maintained.

Image not extracted

Figure 4. Absolute fluctuations of the 5th5^{th}-order AIs with the horizontal oscillation amplitude (the initial vertical amplitude is 6 mm) in the NSLS-II ring.

There exist many other chaos indicators (some of them are compared in Ref. [19, 20]), which can be used for studying nonlinear beam dynamics. Among them, the frequency map analysis (FMA) [21, 22], Shannon entropy [23], forward-reversal integration [24, 25, 26], are selected to compare with the fluctuation of AIs for the NSLS-II ring in Figure 5. Here the relative AIs’ fluctuation is defined as σW1∣⟨W1⟩∣+σW2∣⟨W2⟩∣\frac{\sigma_{W_1}}{|\langle W_1\rangle|}+\frac{\sigma_{W_2}}{|\langle W_2\rangle|}. σW1,2\sigma_{W_{1,2}} represent the standard

Image not extracted

Figure 5. Comparison of different chaos indicators for the NSLS-II ring. Top-left: frequency map analysis; top-right: Shannon entropy; bottom-left: forward-reversal integration; bottom-right: fluctuations of approximate invariant. They display a common pattern, in which regular motions gradually dilute into chaos with the amplitude.

deviations of multiple-turn W’s, and ∣⟨W1,2⟩∣|\langle W_{1,2}\rangle| are their absolute mean values. In Figure 5, various chaos indicators display a common pattern, in which regular motions gradually dilute into chaos with amplitude. Note that the color scales used in these 4 cases are not directly comparable, because of the different chaos metrics used. From the comparison, we observe that the FMA excels at detecting resonances, which needs a spectral analysis, such as FFT or NAFF. Such spectral analysis is also adopted in ref. [7]. The forward-reversal integration mainly describes the unpredictability of long-term motions, relying on the randomness in computation. The Shannon entropy and AI’s fluctuations are actually similar. Both describe the fluctuations of predefined motion constants. While computing Shannon entropy, the phase space region needs to be discretized into a mesh grid, which makes the transition from regular motion to chaos not as smooth as AI. While AI has an explicit function form, the characterization of chaos is more accurate.

In designing accelerators, it is generally believed that, by suppressing the chaos of beam motions, their dynamic apertures can be enlarged. Since AI’s fluctuations is a measure of chaos, we explore the possibility of optimizing the NSLS-II magnetic lattice by suppressing the fluctuations of its 5th5^{th} order AIs. Technically, by tuning some harmonic sextupoles, we aim to minimize the AI’s fluctuations on a specific trajectory starting with a given initial condition x=15.7mmx=15.7\mathrm{mm}, y=3.7mmy=3.7\mathrm{mm}, px,y=0p_{x,y}=0. Once the optimal sextupole configuration is reached, the on- and off-momentum dynamic apertures are obtained by tracking particles for 2048 turns with the simulation code ELEGANT [27] and illustrated in Figure 6. More detailed simulations confirm that, this nonlinear lattice performance is similar to our baseline design [16]. However, the suppression of AI’s fluctuation is more efficient than other multi-particle tracking-based optimizations, such as [28, 29]. The optimization is accomplished within several hours even on a single core computer.

Image not extracted

Figure 6. On- and off-momentum dynamic apertures for the optimal solution obtained by minimizing the fluctuations of AIs. Here δ=ΔPP\delta= \frac{\Delta P}{P} represents charged particle’s relative momentum deviations.

It is worth to mention another method to construct approximate actions also from a square matrix in the ref. [7]. Its square matrix is constructed but with the monomials of the phase space variables z=xˉ−ipˉz = \bar{x} - i\bar{p} after applying the Courant-Snyder parameterization. Then a Jordan analysis is carried out to determine the amplitude-dependent-tune-shift and approximate actions W0W_0. Predicting the long-term stability of each initial condition in phase space requires an additional transformation from the approximate action-angle pair to a rigid rotation, achieved through an iterative process in the frequency domain. Mathematically, such Jordan decomposition is a block-diagonal matrix with Jordan blocks. Ref. [7] further explains a method of detecting resonances, including high-order weak ones, which demands substantial computational resources such as high-dimensional Fourier analysis. Alternatively, our proposed method utilizes the eigen-analysis on the raw square matrix (without the Courant-Snyder parameterization). It aims to finding a diagonal matrix (or as close to diagonal as possible) that is similar to the original matrix. The needed computational effort significantly reduces. However, the most valuable information, i.e., the approximate invariants, are constructed explicitly. With the obtained AIs, other tune related information, such as the tune-shift-with-amplitude, could be further derived using the methods explained in refs. [30, 31].

Conclusion

In summary, we present a method to construct the approximate invariants of non-integrable Hamiltonian systems and demonstrate its application on a modern accelerator. We also demonstrate the fluctuation of AIs when evaluating with simulation data is a chaos indicator, which can be potentially used as an efficient optimization objective in designing modern accelerators.

Acknowledgments

We would like to thank S. Nagaitsev (BNL), C. Mitchell (LBNL) for their valuable discussion and constructive suggestion, P. Belanger and G. Sterbini (CERN) for a detailed numerical comparison and confirmation, H. Zhang (JLab), Y. Hidaka and D. Hidas (BNL) for their consistent supports in computation. This research is supported by the U.S. Department of Energy (DOE) under Contract No. DE-SC0012704, Office of Basic Energy Sciences, Field Work Proposal 2025-BNL-PS040, and the DOE HEP award DE-SC0019403.

The data that support the findings of this article are not publicly available because of legal restrictions preventing unrestricted public distribution. The data are available from the authors upon reasonable request.

References

  1. [3]Mikko Kaasalainen and James Binney, “Construction of invariant tori and integrable hamiltonians,” Physical review letters 73, 2377 (1994).pubmed.ncbi.nlm.nih.gov/10057044
  2. [4]Mikko Kaasalainen, “Construction of invariant tori in chaotic regions,” Physical Review E 52, 1193 (1995).pubmed.ncbi.nlm.nih.gov/9963524
  3. [5]Alessandra Celletti, Antonio Giorgilli, and Ugo Locatelli, “Improved estimates on the existence of invariant tori for Hamiltonian systems,” Nonlinearity 13, 397 (2000).DOI
  4. [6]Teemu Laakso and Mikko Kaasalainen, “Canonical methods of constructing invariant tori by phase–space sampling,” Physica D: Nonlinear Phenomena 243, 14–19 (2013).DOI
  5. [7]Li Hua Yu, “Analysis of nonlinear dynamics by square matrix method,” Phys. Rev. Accel. Beams 20, 034001 (2017).
  6. [8]Philippe Belanger and Guido Sterbini, “Numerical evaluation of the integrals of motion in particle accelerator tracking codes,” arXiv preprint arXiv:2503.19122 (2025).arxiv.org/abs/2503.19122
  7. [9]Rafael De La Llave et al., “A tutorial on KAM theory,” in Proceedings of Symposia in Pure Mathematics, Vol. 69 (Citeseer, 2001) pp. 175–296.DOI
  8. [10]Alexander Wu Chao, Maury Tigner, Hans Weisse, and Frank Zimmermann, Handbook of accelerator physics and engineering (World scientific, 2023) Chap. 2.3.7.
  9. [11]Martin Berz, “Modern map methods for charged particle optics,” Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 363, 100–104 (1995).DOI
  10. [12]Lingyun Yang, “Array based truncature power series package,” Proc. ICAP’09 , 371–373 (2009).
  11. [13]He Zhang, “cppTPSA/pyTPSA: a C++/Python package for truncated power series algebra,” Journal of Open Source Software 9, 4818 (2024).
  12. [14]Alexander Wu Chao, Lectures on accelerator physics (World Scientific, 2020).
  13. [15]Ernest D Courant and Hartland S Snyder, “Theory of the alternating-gradient synchrotron,” Annals of physics 3, 1–48 (1958).
  14. [16]Steve Dierker, “NSLS-II preliminary design report,” Brookhaven National Laboratory (2007).DOI
  15. [17]Haruo Yoshida, “Construction of higher order symplectic integrators,” Physics letters A 150, 226–268 (1990).
  16. [18]A Chao, D Johnson, S Peggs, J Peterson, C Saltmarsh, L Schachinger, R Meller, R Siemann, R Talman, P Morton, et al., “Experimental investigation of nonlinear dynamics in the fermilab tevatrοn,” Physical review letters 61, 2752 (1988).pubmed.ncbi.nlm.nih.gov/10039213
  17. [21]Jacques Laskar, “Frequency map analysis and particle accelerators,” in Proceedings of the 2003 Particle Accelerator Conference, Vol. 1 (IEEE, 2003) pp. 378–382.DOI
  18. [22]Yannis Papaphilippou, “Detecting chaos in particle accelerators through the frequency map analysis method,” Chaos: An Interdisciplinary Journal of Nonlinear Science 24 (2014).DOI
  19. [23]Yongjun Li, Kelly Anderson, Derong Xu, Yue Hao, Kiman Ha, Yoshiteru Hidaka, Minghao Song, Robert Rainer, Victor Smaluk, and Timur Shaftan, “Online regularization of poincaré map of storage rings with shannon entropy,” Phys. Rev. Accel. Beams 28, 034001 (2025).arxiv.org/abs/2408.14333
  20. [24]Kilean Hwang, Chad Mitchell, and Robert Ryne, “Rapidly converging chaos indicator for studying dynamic aperture in a storage ring with space charge,” Phys. Rev. Accel. Beams 23, 084601 (2020).DOI
  21. [25]Yongjun Li, Yue Hao, Kilean Hwang, Robert Rainer, An He, and Ao Liu, “Fast dynamic aperture optimization with forward-reversal integration,” Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 988, 164936 (2021).DOI
  22. [26]Federico Panichi, Krzyszof Goździewski, and Giorgio Turchetti, “The reversibility error method (REM): a new, dynamical fast indicator for planetary dynamics,” Monthly Notices of the Royal Astronomical Society 468, 469–491 (2017).DOI
  23. [27]Michael Borland, Elegant: A flexible SDDS-compliant code for accelerator simulation, Tech. Rep. (Argonne National Lab., IL (US), 2000).DOI
  24. [28]Lingyun Yang, Yongjun Li, Weiming Guo, and Samuel Krinsky, “Multiobjective optimization of dynamic aperture,” Physical Review Special Topics—Accelerators and Beams 14, 054001 (2011).DOI
  25. [29]Yongjun Li, Weixing Cheng, Li Hua Yu, and Robert Rainer, “Genetic algorithm enhanced by machine learning in dynamic aperture optimization,” Physical Review Accelerators and Beams 21, 054601 (2018).
  26. [30]Sergei Nagaitsev and Timofey Zolkin, “Betatron frequency and the poincaré rotation number,” Phys. Rev. Accel. Beams 23, 054001 (2020).
  27. [31]Chad E. Mitchell, Robert D. Ryne, Kilean Hwang, Sergei Nagaitsev, and Timofey Zolkin, “Extracting dynamical frequencies from invariants of motion in finite-dimensional nonlinear integrable systems,” Phys. Rev. E 103, 062216 (2021).DOI

Paper details

vlnc:W_w9z9tn4k
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents