A new mathematical approach to finding global solutions of the magnetic structure determination problem

Determination of magnetic structure is an important analytical procedure utilized in various fields ranging from fundamental condensed-matter physics and chemistry to advanced manufacturing. It is typically performed using a neutron diffraction technique; however, finding global solutions of the magnetic structure optimization problem represents a significant challenge. Generally, it is not possible to mathematically prove that the obtained magnetic structure is a truly global solution and that no solution exists when no acceptable structure is found. In this study, the global optimization technique called semidefinite relaxation of quadratic optimization, which has attracted much interest in the field of applied mathematics, is proposed to use as a new analytical method for the determination of magnetic structure, followed by the application of polarized neutron diffraction data. This mathematical approach allows avoiding spurious local solutions, decreasing the amount of time required to find a tentative solution and finding multiple solutions when they exist.

In this study, a new analytical method for the determination of magnetic structure is proposed, which allows (i) judging whether the obtained solution is a truly global one; (ii) evaluating the uniqueness of the solution obtained for a given set of experimental data; and (iii) completing the global optimization procedure in a very short time (typically, less than several seconds). Furthermore, the applicability of the proposed method is demonstrated using polarized-neutron powder diffraction data.

Mathematical Formulation for Applications of Semidefinite Relaxation Method
Relaxation techniques have been previously developed in the field of mathematical programming. In particular, semidefinite relaxation (SDR) combined with semidefinite programming (SDP) has found many applications in applied mathematics and engineering (for example, see Chapter 2.2 in ref. 24 ). SDR is an efficient technique for solving nonlinear optimization problems, such as quadratic programs (QPs), namely the minimization of multivariate quadratic polynomials under constraints. It should be noted that SDR could provide both fast convergence and a numerical proof of the global optimality of the obtained solutions using the global convergence property and duality theorem of the convex programming methods, as schematically shown in Fig. 1(B,C).
In the field of optical imaging, Candès et al. published a pioneering work (called the PhaseLift method), in which a sparse modelling approach was adopted for general phase retrieval 25 . Furthermore, one of the authors (ROT) developed an SDR-based mathematical approach to investigate whether a crystal structure could be uniquely identified using only diffraction data and independent atomic model 26 .
In this study, this method was applied to magnetic structure analysis by replacing the optimized parameters with magnetic moments. The problem of magnetic structure determination can be mathematically described by the following set of N quadratic equations: denote the observed and calculated integrated magnetic diffraction intensities at the magnetic reflection = Q h k l k mag mag mag , respectively; x = (x 1 , …, x n ) T and S k represents the coefficient matrix numerically obtained using the scattering cross-section formula for each Q k , which consists of the absolute intensity scale factor, magnetic form factor, Lorentz factor, multiplicity and structural weight factor (the details of this formula are summarized in Supplementary Information).
The symbol ≈ indicates that each I Q ( ) k mag,obs value includes an experimental error. In order to incorporate all errors, the following optimization problem is solved: The objective function f is globally minimized to the global optimum value f (gl) . When the duality gap ∆f is close to zero (corresponding to the square root of the machine epsilon), the global optimization process is complete according to the duality theorem. In panels (B) and (C), the solid balls represent the paths toward feasible solutions chosen by the interior point method. The inequality ≤ x R i i can be also removed if the permitted range of x i (R i , for example, the upper limit of magnetic moment) is uncertain while maintaining the validity of the subsequent discussion of the SDR and SDP techniques.
Equation (2) is classified as a so-called l 1 -norm minimization problem, which is equivalent to the following form of the quadratic programming problem: The problems described by equations (2) and (3) are equivalent since they have the same set of solutions x and the minimized values of the objective functions (more details are provided in Supplementary Information). The basic strategy utilized in this work is to apply SDR to equation (3) in order to solve equation (2) and then provide a computational proof on the global convergence property of the solution. The least-squares minimization of the function is avoided because a small increase in the polynomial degree considerably magnifies the size of the SDR problem (it increases proportionally to the power of N). Thus, it is possible to locate the global optimum by performing l 1 -norm minimization, which can be subsequently used as the initial parameter of the normal least-squares method to calculate the refined parameters and then compare them with literature data.

Application to Experimental Data Analysis
Experimental results. The developed SDR method was verified by applying it to the experimental data obtained for pyrochlore Nd 2 Ir 2 O 7 , which served as a proximate material for a three-dimensional Weyl semimetal and a component of spintronic devices on the basis of its geometrically frustrated magnetism and 5d-electron configuration [27][28][29][30][31][32] . This state theoretically corresponds to the all-in all-out type of magnetic structure described by the magnetic propagation vector k mag = (0, 0, 0), in which all the magnetic moments are oriented either towards the centre of the participating tetrahedron or in the opposite direction 30,32 . This prediction was experimentally confirmed in previous unpolarized neutron diffraction studies 33,34 . However, the superposition of the nuclear and magnetic reflections observed for the magnetic structures with k mag = (0, 0, 0) produces ambiguous results during their separation. Furthermore, it is not possible to mathematically prove with 100% certainty the absence of other acceptable solutions to the problem of magnetic structure determination. Thus, in this work, polarized neutron diffraction studies were performed for this material, and the obtained magnetic structure was verified mathematically. Figure 2 shows the representative neutron diffraction data obtained for Nd 2 Ir 2 O 7 at the minimum temperature T = 1.4 K. The non-spin-flip and spin-flip parameters I OFF and I ON roughly correspond to the nuclear and magnetic reflection components I nuc and I mag , respectively. The magnitude of I OFF is systematically larger than I ON at all temperatures, thus confirming the necessity of conducting polarized neutron diffraction experiments. For I ON , the intensity of the 113 reflection increases with decreasing temperature, indicating the existence of strong temperature dependence for this reflection, which is not observed for the 222 reflection. Moreover, as the flipping ratio of the neutron beam in the actual experiments is finite, the magnitudes of I ON and I OFF can be expressed by the following formulas:  Γ Nd = Γ 3 + 2Γ 5 + 3Γ 7 + 6Γ 9 and Γ Ir = Γ 3 + 2Γ 5 + 3Γ 7 + 6Γ 9 , where Γ 3 corresponds to the all-in all-out type of the magnetic structure, and the coefficients denote the numbers of basis vectors summarized in Table 2 18,35 . Hence, the total numbers of variables are reduced to n Γ3 = 2, n Γ5 = 4, n Γ7 = 6 and n Γ9 = 12. Furthermore, representation Γ 9 includes six ferromagnetic basis vectors. However, the Nd 2 Ir 2 O 7 structure exhibits only extremely weak spontaneous magnetization (around 10 −4 μ B /formula unit) 29 , which is significantly below the detection limit of the neutron diffraction technique. Therefore, the ferromagnetic basis vectors are precluded and the value of n Γ9 is further reduced from 12 to 6 during the analysis of the neutron diffraction data. Thus, the variables used for the magnetic structure analysis in this work are defined as follows: . The goal is to determine the globally optimal solutions x for representations Γ 3 , Γ 5 , Γ 7 and Γ 9 using the I mag,obs magnitudes listed in Table 1. After that, the solution characterized by the best fit can be selected.
The output values produced by the SDP solver are as follows.
(1) The convergence procedure results in the following coefficients:    , where θ and φ are arbitrary. Without SDR, it is difficult to find the optimum solutions for this type of problems and prove that better solutions do not exist. Furthermore, the zero magnitude of x Γ7 (opt) is obtained for representation Γ 7 . Indeed, when x Γ7 values are finite (non-zero), the I mag,cal values for = I 0 mag,obs indices increase more rapidly than those for ≠ I 0 mag,obs indices. Thus, the SDR method automatically overcomes these issues.
All these facts suggest that Γ 3 represents the best solution, whereas representations Γ 5 , Γ 7 and Γ 9 are excluded for the first time. For clarity, the values of I mag,cal are listed in Table 2 and also shown in Fig. 3. The magnitude of I mag,cal-Γ3 matches I mag,obs very well, whereas the globally optimal I mag,cal-Γ5 , I mag,cal-Γ7 and I mag,cal-Γ9 parameters substantially differ from the corresponding I mag,obs values.
After proving mathematically the uniqueness of representation

Discussion
Representation Γ 9 (the only IR containing ferromagnetic basis vectors 35 ) is rejected as the optimal solution despite the existence of weak ferromagnetism 29 . The observed inconsistency suggests that this ferromagnetism is symmetrically decoupled from the bulk magnetic structure (for example, as a surface or interface-protected property). In fact, by performing careful measurements of the macroscopic magnetization and electrical resistivity of isomorphic Cd 2 Os 2 O 7 , it was found that the surface ferromagnetism coupled with novel spin-polarized conductivity emerged on the walls between the all-in all-out and all-out all-in antiferromagnetic domains 36 . The macroscopic observation is consistent with the findings of this study verified both microscopically and mathematically, indicating their potential applicability in domain wall spin electronics. We discuss the expected application scope of the SDR method. After determining the list of (h mag , k mag , l mag , I mag,obs ) values, the described SDR method can be used to find the corresponding magnetic structures as the global solutions. First, the configurations with k mag = (0, 0, 0) are considered, indicating their high potential applicability in various fields (including magnet materials). Second, this method is not restricted to polarized neutron diffraction experiments, but can be also used in studies involving unpolarized neutrons when the magnetic structure is characterized by k mag ≠ (0, 0, 0) or k mag = (0, 0, 0) with detectable magnetic moment. Both the magnetic and crystallographic structures can be simultaneously refined by the normal least-squares method, in which the obtained SDR solutions are utilized as the initial values of the magnetic structural parameters. Third, the SDR technique is able to easily process thousands of independent variables. Therefore, magnetic structures of arbitrary types can be theoretically determined using the advanced diffractometers that provide a relatively large number of reflection points (even for the target materials with complex compositions).
The diffraction intensities are not represented by quadratic functions of the atomic positions r j ; hence, the SDR technique seems to be inapplicable for determining the values of r j (crystal structure). However, the neutron and X-ray diffraction intensities are represented by those of the nuclear and electron densities (generalized crystal structure), respectively. Likewise, a magnetic structure is also generalized to the magnetic moment density. Thus, the SDR technique is expected to enable analysing the global solutions of the generalized structures together with the aforementioned high-volume processing ability, such as protonic/ionic distributions in the conductors and electronic spin-orbital distributions.

Materials and Methods
Calculations. The system of quadratic equations was solved using the SDP solver SDPA 37 . To find the global optima of SDP problems, interior point methods were efficiently used 38 . Experiments. The polarized neutron elastic scattering experiments were performed using the HB1 thermal neutron three-axis spectrometer located at the High Flux Isotope Reactor (HFIR) of the Oak Ridge National Laboratory (ORNL). Heusler alloy 111 reflection crystals were utilized as the monochromator and analyser. The flipping ratio R = 10 was obtained using the nuclear 222 reflection at a paramagnetic temperature of 40 K corresponding to the beam polarization P = 0.82. The polarization vector was set parallel to the scattering vector (P//Q). The incident energy of neutrons was E i = 13.5 meV. The horizontal collimator sequence was 48′(open)-80′-80′-240′. A pyrolytic graphite Bragg-reflection filter was used to efficiently eliminate the contamination caused by higher-order wavelengths. A powder Nd 2 Ir 2 O 7 sample was synthesized by a solid-state reaction method inside a quartz tube. About 4.8 g of the sample was wrapped in thin aluminium foil and shaped to a hollow cylinder with a thickness of 0.8 mm and diameter of 20 mm to mitigate the effect of the strong neutron absorption of Ir nuclei. The cylinder was stored in an aluminium container filled with He gas and placed under the cold heads of a He-closed-cycle (Displex) or liquid-He-type (Orange) cryostat.