Magnetism I: from the atom to the solid state Preface The Lecture on “Magnetism I: from the Atom to the solid state” is an introduction to the fundamental concepts in magnetism. It consists of two parts: the first one (by D. Pescia) deals with magnetic effects in atoms (diamagnetism, paramagnetism, formation of magnetic moments in atoms) and with the occurrence of magnetic order in the ground state of a solid in virtue of the exchange interaction. The second part (by A. Vindigni) treats the occurrence of magnetism at finite temperatures, the role of small interactions such as the dipolar interaction, and presents the essential facts about the statistical physics of magnetism. A very extended introduction in modern magnetism can be found in the book by J. Stöhr and H.C. Siegmann ” Magnetism: from fundamentals to the Nanoscale dynamics”, Springer-Verlag, Berlin Heidelberg 2006. Zürich, September 2012 D. Pescia ii Contents Preface ii 1 Magnetism in Atoms 1.1 Magnetism in classical physics . . . . . . . . . . . . . . . . . 1.2 Magnetism in quantum mechanics . . . . . . . . . . . . . . . 1.2.1 Free electrons in a magnetic field . . . . . . . . . . . 1.2.2 Electron in a magnetic field and a central potential . 1.3 The formation of the magnetic moment in atoms . . . . . . . 1.3.1 Paramagnetism in Atoms . . . . . . . . . . . . . . . . 1.4 Exchange interaction and the Heisenberg-Dirac-Van Vleck Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Magnetism in solids 2.1 Stoner-Wohlfahrt model . . . . . . . . . . . . . . . 2.2 Friedel-Oscillations . . . . . . . . . . . . . . . . . . 2.2.1 The interatomic exchange interaction . . . . 2.2.2 RKKY oscillations . . . . . . . . . . . . . . 2.2.3 Anhang: mathematical details of the model 3 Magnetic order at finite temperature 3.1 Coupled effective spins: an N -body problem 3.2 Mean-field approximation (MFA) . . . . . . 3.3 Mean-field universality class . . . . . . . . . 3.4 The Landau approach . . . . . . . . . . . . 3.5 Classical spin models . . . . . . . . . . . . . 3.6 Correlation functions . . . . . . . . . . . . . 3.7 Landau theory of correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 5 5 7 9 16 . 19 . . . . . 23 23 31 31 32 34 . . . . . . . 39 40 42 45 48 53 61 66 4 Magnetic domains and domain walls 71 4.1 Magnetic anisotropy . . . . . . . . . . . . . . . . . . . . . . . 71 4.2 Domain walls in the classical Heisenberg model . . . . . . . . 74 iii iv CONTENTS 4.3 4.4 4.5 4.6 4.7 4.8 Continuum formalism . . . . . . . . . . . Beyond the Mean-Field Approximation . Finite size and superparamagnetic limit . Dipolar interaction . . . . . . . . . . . . Dipolar interaction in extended systems . Origin of magnetic domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 80 88 90 94 100 1 CONTENTS Part I D. Pescia Chapter 1 Magnetism in Atoms 1.1 Magnetism in classical physics The experimental facts about magnetism in solids are as old as the history of mankind: some materials have the property of producing a sizable magnetic field that attracts or repel other materials. One of the first references to the magnetic properties of what we know now to be magnetite F e3 O4 (lodestone) is by 6th century BCE Greek philosopher Thales of Miletus. The name ”magnet”may come from the lodestones found in Magnesia. In China, the earliest literary reference to magnetism lies in a 4th century BC book called Book of the Devil Valley Master: ”The lodestone makes iron come or it attracts it”. The lodestone based compass was used for navigation in medieval China by the 12th century. The main observation about the origin of the magnetic field originates with the experiments of Ampere and Oersted in the early decades of the 19th century, demonstrating that i a current is able to influence a magnetic needle (Oersted) and ii a mechanical force exists between two wires injected with current (Ampere). Later, Faraday completed our knowledge of magnetic field by discovering that time dependent magnetic fields can produce a magnetic current. J.C. Maxwell gave a complete description of electromagnetic fields that is still very precise (Maxwell equations). The origin of the magnetism in matter remained debated: Ampere postulated that magnetism in atoms originates from the existence of a closed atomic-sized current. Poisson and later Maxwell, instead, favored magnetic charges that appear always coupled into dipoles as the source of the magnetic field. Distinguishing between the two hypothesis is a subtle problem, as a paper in 1977 by J.D. Jackson show. Following Ampere hypothesis, the magnetic field produced by a current circulating within a small loop C is 2 ~ given in term of the magnetic moment vector m ~ = 2e ~x × ~ẋ = 2m L, which 2 CHAPTER 1. MAGNETISM IN ATOMS 3 Figure 1.1: Drawing of the magnetic field of the Earth by René Descartes, from his ”Principia Philosophiae”, 1644. This was one of the first drawings of a magnetic field. . R for a closed loop amounts to m ~ = I S ~nds, where S is the surface within the loop C and ~n is the vector normal to S. ~ ~ r ) = µ0 ∇ ~ × (∇ ~ × m B(~ ) 4π | ~r | µ0 1 ~ µ0 ~ ~ m )− m△ ~ ∇(∇ · = 4π | ~r | 4π | ~r | µ0 ~ ~ 1 ) + µ0 · m = ~ · δ(~r) ∇(m ~ ·∇ 4π | ~r | i h µ 3~r(~r · m) ~ − m~ ~ r 2 µ0 0 − m ~ · δ(~r) + µ0 m ~ · δ(~r) = 4π | ~r |5 3 The first component of the magnetic field produced by the small current ~ field produced by an electric dipole moment. loop is formally identical to E This lead Poisson and later Maxwell to state that ” ...we may regard the magnet,..., as made up of small particles, each of which has two equal and opposite poles” (magnetic charges). This idea, which was alternative to the one of Ampere, describes almost anywhere the magnetic field correctly. At the position of the loop, however, – the second term, given by the Dirac delta function – the picture of magnetic charges would produce a field amounting to − µ30 m ~ · δ(~r). Experiments aimed at measuring the ”contact” term – such as the hyperfine line of atomic hydrogen (relevant in astrophysics) – CHAPTER 1. MAGNETISM IN ATOMS 4 Figure 1.2: Pierre Pelerin de Maricourt (French), Petrus Peregrinus de Maricourt (Latin) or Peter Peregrinus of Maricourt was a 13th century French scholar who conducted experiments on magnetism and wrote the first extant treatise describing the properties of magnets. His work is particularly noted for containing the earliest detailed discussion of freely pivoting compass needles, a fundamental component of the dry compass soon to appear in medieval navigation. can ultimately discriminate between the two hypotheses: the hypothesis by Ampere that the magnetic moment is due to atomic current rather than atomic charges is now well accepted. While Maxwell equations and the postulate of Ampere are very exact in describing atomic magnetic moments and their magnetic fields, the very existence of magnetic moments in classical physics is challenged by a famous theorem, the Bohr-van Leuwen theorem. Theorem: given that the classical Hamilton function for an electron in an ~ writes H = 1 (~p − eA) ~ 2 + eφ it follows that the applied magnetic field B 2m canonical average of the atomic magnetic moment ≺ m ~ ≻ vanishes exactly and any finite temperature. Proof: 1. When one writes down explicitly the equations of motion in a uniform 5 CHAPTER 1. MAGNETISM IN ATOMS field one can construct an integral of the motion – the total energy – which reads 1 ˙2 ~ m~x − ∇φ(~r) (1.1) 2 ~ – indicating that the Remarkably, the total energy is independent of B Lorentz force is not doing any work, being perpendicular to the velocity. Accordingly, when the statistical average based on the canonical Gibbs distribution is computed, the energy of the various classical states over which one integrate does not contain the magnetic field and any partial derivative with respect to the magnetic field – such as the average magnetic moment – must vanish. 2. Considering that ∂H ≺m ~ ≻= − ≺ ≻ (1.2) ~ ∂B one obtains − ≺ mz ≻ = R 1 (~ ~ 2 −eφ(~ p−eA) x) − 2m kB T d~pd~x −(px − 2e Bz y)y + (py + 2e Bz x)x e (1.3) 1 (~ ~ 2 −eφ(~ p−eA) x) − 2m R kB T d~pd~xe Using the variable transformation p′x x p′y 1 0 x′ = 0 0 0 1 eB 2 0 0 0 1 0 px 0 x 0 py y 1 −eB 2 (1.4) one can easily convince oneself that the integral – and with it the average magnetic moment – is vanishing – a result known as Bohr van Leuwen theorem of classical statistical physics. 1.2 1.2.1 Magnetism in quantum mechanics Free electrons in a magnetic field It was first Landau (1930) who explicitly produced an average magnetic moment and thus determined the ”birth” of magnetism in matter. Landau solved the problem of a free electron moving in a uniform magnetic field using quantum mechanics to describe the motion, i.e. breaking away from 6 CHAPTER 1. MAGNETISM IN ATOMS Lorentz force. For a simple case of one free electron, the energy levels write (see textbooks in QM) En = ~ωc (n + 1/2) + ~2 kz2 2m (1.5) ), their degeneracy being (not including the spin of the (n = 0, 1, 2..., ωc = eB m L2 eB electron) 2π~ (L: linear dimension of the system). The partition function reads L2 eB X X −En Q(T, B, L) = exp 2π~ k n kB T z 3 X −n~ω L eB p c − ~ω 2 e kB T 2πk T e = B (2π~)2 n L3 eB p 1 = 2πkB T ~ωc 2(2π~)2 sinh( 2k ) bT (1.6) The magnetization (per particle) amounts to kB T kB T ~ωc ~ωc ∂lnQ = − coth( ) ∂B B 2B 2kb T (1.7) For small magnetic field we have the simple result mz = − e 2 ~2 B 12m2 kB T (1.8) mz being the average magnetic moment of the free electron. This result sanctions the appearance of a non-vanishing magnetic moment. Typically, one deals with N free electrons, each energy level being occupied by two electrons at the most, up to the Fermi energy. Considering that only electrons within an energy range kb T below the Fermi energy can be excited, the magnetization (magnetic moment per unit volume) induced by the magnetic field amount to (a= lattice constant) e2 ~2 B · ρ(EF )kB T e 2 ~2 . mz Mz = 3 = − = − ρ(EF ) a 12m2 kB T 12m2 (1.9) Here is ρ(EF ) the density of states at the Fermi level, which for a typical free electron metal amounts to 0.805 · 1022 eV −1 cm−3 . The final result is that the induced magnetization in a free electron gas (without considering the spin of the electron) is independent of the temperature. Defining the magnetic susceptibility as ∂Mz . χz = µ0 (1.10) ∂Bz CHAPTER 1. MAGNETISM IN ATOMS 7 as a way of comparing the various magnetic responses, we find a value for χz in the case of free electrons of about 10−6 , which means that an external field of the order of 1 Tesla produces a magnetic field µ0 Mz of the order of 1 microtesla. Notice that letting ~ go to zero produces the vanishing of M required by classical physics. This result shows that quantum mechanics is the key for the appearance of a magnetic moment in matter, albeit weak. 1.2.2 Electron in a magnetic field and a central potential In an atom, the Hamilton operator in the presence of a uniform magnetic ~ = − 1 (~r × B) ~ reads field A 2 2 2 ~ r) + V (~r) = [− ~ △ + V ] p~ − eA(~ 2m ~ − B·µ ~ → (a) 2 e ~2 + A → (b) 2m ~ . = ~ = −µB · L~ is the operator representing the magnetic moment (µB |e|~ µ 2m −23 J 0.927410 T is the Bohr magneton). The strength of the various terms can be computed e.g. on hydrogen orbitals. We obtain ~2 <s|− △ + V | s >= −E0 = −13.6eV 2m ~ · L) ~ | l = 1, ml = −1 >= −µB · B (a) →< l = 1, ml = −1 | −µB (B e2 ~ 2 A →| s > < s | 2m B 4 = 10−12 ( )2 (b) → E0 3 T In the last line B in supposed to be given in T . Term a. This term is called the Zeeman term, the corresponding energy is the Zeeman energy. Its contribution is non-vanishing when the atomic orbital has a finite z-component of the angular momentum. It shows that an atomic state with appropriate z-component of the angular momentum (in the present case −~ lowers the total energy of the atomic level by an amount µB · B, which, for a magnetic field of 1T corresponds to about −10−4 eV. Notice that, at finite temperatures, levels with Lz = 0, +1 are also partially occupied, decreasing the total average magnetic moment, which becomes strongly temperature dependent, as we will discuss at the end of this chapter. CHAPTER 1. MAGNETISM IN ATOMS 8 Term b. This term is much smaller than the Zeeman energy, almost temperature independent, and only observable if the total magnetic moment of the atoms is exactly vanishing: it builds the so called diamagnetic contribution to the energy. From the expression for the diamagnetic energy one can deD fine an effective diamagnetic moment per atom through µD = − ∂E and ∂B 2 ∂ ED −11 a diamagnetic susceptibility χD = −µ0 · a3 ∂B 2 amounting to ≈ −10 eV /T 0 and −10−6 , respectively, in a magnetic field of 1T . Thus, the diamagnetic moment per electron points antiparallel to the applied magnetic field. Figure 1.3: The top of the figure describes some alternative definitions of the magnetic susceptibility. The Table on the left shows values of the diamagnetic susceptibility for some substances (to find χD multiply the values by 10−5 ). The figure on the right illustrates e2 ~ 2 A →| s > to the diamagnetic a straightforward generalization of the expression < s | 2m µ0 e2 P 2 susceptibility of many in an atom, which writes χD = − 6m i < ri > the sum extending over all electrons building the magnetic response (including the core electrons). Therefore, the diamagnetic susceptibility should scale with Z· < ra2 >. This is confirmed by the plot bottom right, where the horizontal axis shows the total number of electrons Z multiplied by the square of the ionic radius < ra > in units of A2 , for noble gases and ions of atoms with filled atomic shells, for which no net magnetic moment results. CHAPTER 1. MAGNETISM IN ATOMS 1.3 9 The formation of the magnetic moment in atoms In the previous section, we have shown that the magnetic field couples with the operator of the (negative) orbital angular momentum by means of the coupling constant µB . We have now to take into account that • each electron have an intrinsic spin that adds up in some way to the orbital angular momentum to produce a total angular momentum quantum number J, which results from the addition of the orbital quantum number of the electron and its spin. • atoms have generally many interacting electrons: there spins and orbital angular momenta must be added in some way to find the total ground state angular momentum which results from the addition of the total angular momentum and the total spin of all electrons! We are looking for a set of rules that allow angular momenta to be added in quantum mechanics and to a set of rules that allow to select, among the many possible angular momenta, those proper to the atomic ground state of an atom. In other words: the key problem we have to solve is: given N electrons in the configuration (ni , li , lz,i , s = 1/2), (i = 1...N ), 1st what are the possible values of the total spin S, total orbital angular momentum L and the total angular Momentum J (Russel Sounders symbol for the electronic configuration: (2S+1) LJ ?) 2nd: which configuration has the lowest energy (ground state)? The first rule underlying all steps we are going to introduce to finding the electronic configuration of an atom, is a central theorem of quantum mechanics (known as Clebsch-Gordan series) about adding angular momentum operators: Theorem: given an angular momentum operator P~ with z components ~ with z-components [q, q − 1, ..., −q], the total angular [p, p − 1, ... − p] and Q ~ can assume the values momentum operator T~ = P~ + Q P + Q, P + Q − 1, P + Q − 2, ...., | P − Q | The total spin: The total spin of an assembly of spin 12 particles is given by the Clebsch-Gordan sum of all spins. For instance, for N = 2 we have the possible values for S of 0, 1, carrying singlet (χs ) respectively triplet (χt ) spin CHAPTER 1. MAGNETISM IN ATOMS 10 eigenfunctions: r 1 1 √ (u1/2 u−1/2 − u−1/2 u1/2 ) = (|↑>|↓> − |↓>|↑>) 2 2 u1/2 u1/2 = |↑↑> r 1 1 √ (u1/2 u−1/2 + u−1/2 u1/2 ) = (|↑>|↓> + |↓>|↑>) 2 2 u−1/2 u−1/2 = |↓>|↓> The total orbital angular momentum Again, it is given by the ClebschGordan sum of all orbital angular momenta. The total angular momentum. It is given by the Clebsch-Gordan sum of total spin and total angular momentum. Pauli principle (exchange interaction part I) There is a restriction on the possible configurations: the total wave function of Fermions must be antisymmetric (Pauli principle). A theorem by Weyl helps implementing the Pauli principle in the electronic structure of a many electron system. Theorem (H. Weyl): All eigenfunctions of Sz have the same symmetry with respect to particle permutation: the symmetry property of the eigenfunction of Sz under permutation are called ”Spinrasse”. This is e.g. clearly visible in the two-electron system, where the spin singlet is antisymmetric with respect to particle permutations and the spin triplet is symmetric. This theorem divides the electronic structure of many-electrons atoms into separated thermal schemes, according to their total spin (see e.g. ”para” and ”ortho” He): in fact, because of the requirement that the total wave function be antisymmetric, each spin rasse determines uniquely the symmetry properties of the orbital wave functions under permutation. For instance, the energy levels of the singlet thermal scheme can only have symmetric orbital wave functions: 1 √ (un1 ,l1 (~r1 )un2 ,l2 (~r2 ) + un1 ,l1 (~r2 )un2 ,l2 (~r1 )) 2 while the triplet spin state carries only antisymmetric wave functions: 1 √ (un1 ,l1 (~r1 )un2 ,l2 (~r2 ) − un1 ,l1 (~r2 )un2 ,l2 (~r1 )) 2 The Pauli principles implies that, among the possible configurations arising from the Clebsch-Gordan sum, some cannot be realized because orbital wave functions of the proper symmetry do not exist. For instance, starting from two inequivalent s electrons [n, s],[n′ , s], one can obtain the two configurations 3 S1 ,1 S0 . Starting e.g. from two equivalent s electrons [n, s],[n, s], CHAPTER 1. MAGNETISM IN ATOMS 11 the triplet state cannot be constructed as the requirement of antisymmetric orbital wave function makes it vanishing. In other words: The Pauli principle is responsible that two identical spin 1/2 particles can only be in the same orbital state (n1 , l1 , lz,1 = n2 , l2 , lz,2 ) if they form a singlet. In other words: more than two electrons cannot be in the same orbital state: if they are, they (sloppy) ”must have opposite spin”. Therefore, although the Hamilton operator does not contain any spin dependent term (and the two electrons might not be even interacting), the non distinguishibility postulate of QM removes the degeneracy of some states and even forbids a well defined spin state to be a possible configuration. The Pauli principle acts as some sort of effective interaction that removes some degeneracy and distinguishes between spin states. This interaction is know in the literature with the broad terminology of ”exchange interaction”. It is a purely QM effect and completely disappears together with the spin upon transition to classical mechanics. In the present case the exchange interaction dictates e.g. that the ground state of two identical spin 1/2-particles is a singlet. The Pauli principle and the exchange interaction are key results explaining the stability of matter and, for our purposes, are responsible for the formation of magnetic moments in atoms. Hund’s rules: exchange interaction part II The Pauli principle acts to remove some configurations among the set arising from the Clebsch-Gordan sum but, for instance, does not lift the degeneracy between triplet and singlet states of electrons with different orbital states. F. Hund (1925) and Russel and Saunders (Astrophysics Journal, 61, 38, 1925) formulated on the base of spectral data, a set of empirical rules that allows a further analysis of the various configurations. 1st Hund rule. Provided there is sufficient degeneracy that non-equivalent orbital wave functions can be constructed, the configuration realizing the lowest energy state corresponds to a state of maximum spin number. In other words: if the orbital states involved have different quantum numbers, the filling of the electronic states with parallel spins produces the lowest energy electronic configuration. Thus, provided orbitally degenerate states exist, the triplet state is energetically favored. If degeneracy is absent, then the singlet state is energetically favored. Accordingly, the formation of a finite total spin requires orbital degeneracy. There is an intuitive explanation for this result. The Coulomb energy is large when the two electrons are closer to each other. In a triplet spin state the antisymmetric orbital wave functions the two electrons takes care that the two electrons are as far as possible from each other, an this reduces the Coulomb repulsion with respect to the symmetric orbital wave function, where the two electrons are, on the average,allow to be closer to each other. CHAPTER 1. MAGNETISM IN ATOMS 12 2nd Hund rule. The second Hund rule states that, after having established the maximum S, the lowest energy state correspond to the configuration that maximizes L (again, the Coulomb energy is reduced for high values of L). 3rd Hund rule. The third rule states that the total angular momentum quantum number J minimizing the energy is | L + S | if the shell is more than half full, | L − S | if the shell is less than half full. This last rule minimizes the energy arising from spin-orbit interaction. Of course, other electronic configurations exist, which have empirically higher energies. These rules, together with the Pauli principle, determine completely the total values of J, L, S for the ground state and its electronic symbol 2S+1 LJ . We are now ready to discuss the magnetic moment operator arising from the electronic configuration 2S+1 LJ . As electrons do have both orbital and spin angular momentum, the magnetic coupling with a magnetic field must be extended to include the magnetic moment arising from the spin. We ~ with gS = 2 know from the Dirac theory of the electron that ~µS = −gS µB S, (in contrast to gL = 1). Quantum electrodynamic corrects this value to gS = 2.0023, which is extremely close to the experimental value. Accordingly, the operator describing the interaction of a magnetic moment with a magnetic field becomes ~ + g S) ~ ·B ~ = −µB (J~ + (g − 1)S) ~ ·B ~ HZ = −µB (L ~ S ~ (let us call gS g, for simplicity). HZ is called the Zeeman operator. J~ = L+ ~ S ~ and J~ are dimensionless is the operator of the total angular momentum. L, angular momenta. Let us now assume that we have determined (by solving the many body electron problem and using the Hund’s rules) the total orbital angular momentum quantum number L of the ground state and its total spin angular momentum quantum number S. Notice that the possible eigenvalues of J~2 also labels the eigenspaces of an Hamiltonian that contains the spin ~ ·L ~ interaction, as both J~2 and the spin orbit interaction are scalar orbit S under rotation and therefore have, according to the Wigner-Eckart theorem, the same eigenspaces. Notice that, again because of the W.-E. theorem, Jz , Lz and Sz have also common eigenspaces, as both are the z-components of a vector operator under rotations. We consider Hz to be a perturbation of the fine structure energy levels of an atom, i.e. those energy level resulting from Hund’s rules and carrying a well defined quantum number J. To find the first order eigenvalues of HZ we solve the eigenvalue problem of HZ CHAPTER 1. MAGNETISM IN ATOMS 13 within the 2J + 1 dimensional space containing the symmetry adapted wave functions to this J-value: un,l,j,mj . We seek the first order correction, i.e. the magnetic field is small enough so that the Zeeman splitting is smaller than the level splitting between the various multiplet components J arising from ~ = (0, 0, B): the spin-orbit coupling. We choose B HZ = µB · B · (Jz + (g − 1) · Sz ) The eigenvalue problem of Jz is simply solved, because (un,l,j,mj , Jz un,l,j,mj ) = mj (1.11) mj = J, J − 1, ..., −J. The eigenvalue problem of Sz is simplified by the Wigner-Eckart theorem, which states that (un,l,j,mj , Sz un,l,j,mj ) = τLSJ mj (1.12) τLSJ is common to all mj and QM shows that τLSJ = J(J + 1) + S(S + 1) − L(L + 1) 2J(J + 1) (1.13) Accordingly, we obtain (un,l,j,mj , Hz un,l,j,mj ) = µB · B · mj · gLSJ with the Lande factor gLSJ = 1 + (g − 1) · J(J + 1) + S(S + 1) − L(L + 1) 2J(J + 1) (1.14) Accordingly, a magnetic field lifts the 2J + 1-degeneracy of the fine structure level completely. The Zeeman splitting is symmetric around the unperturbed level Enj . The distance between two consecutive Zeeman levels is △EZ = µB · B · gLSJ (1.15) i.e. it is proportional to B. These results are known as the anomalous Zeeman effect, to be compared with the normal Zeeman effect, with g = 1 and gLSJ = 1. CHAPTER 1. MAGNETISM IN ATOMS 14 CHAPTER 1. MAGNETISM IN ATOMS 15 Figure 1.4: Top: Effective magnetic moment of ions of rare earths atoms as a function of the number of f -electrons and their electronic configurations. CHAPTER 1. MAGNETISM IN ATOMS 1.3.1 16 Paramagnetism in Atoms The statistical mechanics of an ensemble of non interacting N identical atoms each carrying the same quantum number J allows to compute the partition function as hX −gLSJ · µB · mj · B iN ZN = exp( (1.16) kB · T m j the total free energy f (T, B) per atom as hX −gLSJ · µB · mj · B ] −kB · T ln exp( kB · T m j and the mean magnetic moment per atom as −∂f (T, B) = gLSJ · µB · J · BJ (α) ∂B 2J + 1 2J + 1 1 α BJ (α) = · coth[ α] − coth 2J 2J 2J 2J . gLSJ · µB · J · B α= kB · T < µz >= (1.17) For the particular case J = 1/2 (a spin 1/2, s-state Atom) we obtain the simple equation µB · B < µz >= µB tanh[ ] (1.18) kB · T For small arguments α, BJ (α) ≈ ) < µ(P >≈ z J+1 α J 3 and we obtain the Curie law (gLSJ )2 · J(J + 1) · µ2B ·B 3kB · T (1.19) which contains the purely quantum mechanical quantity (gLSJ )2 · J(J + 1). The paramagnetic susceptibility in this limit amounts to C (1.20) χP ≈ µ0 T with the Curie constant N (gLSJ )2 · J(J + 1) · µ2B (1.21) C= V 3kB At room temperature is χP ≈ 10−3 . Notice that the determination of the Curie constant is a key experiment to access the quantities J and gLSM determining the ground state electronic configuration of atoms and thus provide a reliable test of our quantum mechanical approach to the ground state configuration in atoms. 17 CHAPTER 1. MAGNETISM IN ATOMS Figure 1.5: Temperature dependence of 1 χ for some Cupric salts CHAPTER 1. MAGNETISM IN ATOMS 18 19 CHAPTER 1. MAGNETISM IN ATOMS 1.4 Exchange interaction and the Heisenberg-Dirac-Van Vleck Hamiltonian An estimate of the triplet-singlet splitting in atoms is often obtained by explicitly computing the lowest energy states of the simplest many electrons atomic system: the He-atom. The full Hamiltonian in the BornOppenheimer approximation amounts to H0 + V (1, 2) with (e2 standing for e2 ) 4πǫ0 −~2 ~ 2 ~ 2 Ze2 Ze2 ( ∇1 + ∇2 ) − − 2m r1 r2 2 e V (1, 2) = r12 H0 (1, 2) = Let us first neglect V12 . The energy levels of H0 (i) are Eni = (1.22) −Z 2 e2 , 2·a·n2i a being Rni ,li r · Yli ,mI (ϑ, ϕ). the Bohr radius. The eigenfunctions of Eni are ϕni li mi = The ground state of H0 corresponds to the state in which both electrons are 2 2 in a 1s-orbital and has energy E0 = 2 · E1 = −Za e . Its wave function is ψ0 = ϕ1s (1)ϕ1s (2) ⊗ χ(1, 2) ϕ1s (1)ϕ1s (2) = 1 Z 3 − Z (r1 +r2 ) ( )e a π a and 1 χ(1, 2) = √ (u1/2 u−1/2 − u−1/2 u1/2 ) 2 The wave function ϕ1s (1)ϕ1s (2) is symmetric with respect to change of the coordinate vectors and the ground state of He is a singlet, as required by Pauli principle. The net magnetic moment is vanishing: S = 1 is prohibited in the ground state. Let us now introduce V12 as a perturbation. The energy of the ground state is modified to EG = E0 + Q with Z e2 Q = dV (1)dV (2)ϕ21s (1) ϕ21s (2) r12 In order to explicitly compute this last integral, one develops 1/r12 in spherical harmonics: 1 4π X 1 r2 ∗ = ( )l · Yl.m (ϑ1 , ϕ1 )Yl,m (ϑ2 , ϕ2 ) ⇐⇒ r1 > r2 | ~r1 − ~r2 | r1 2l + 1 r1 l,m 1 | ~r1 − ~r2 | = 4π X 1 r1 ∗ ( )l · Yl,m (ϑ1 , ϕ1 )Yl,m (ϑ2 , ϕ2 ) ⇐⇒ r1 < r2 r2 2l + 1 r2 l,m 20 CHAPTER 1. MAGNETISM IN ATOMS Inserting this development in Q and using the orthogonality of spherical harmonics we obtain Z r1 Z Z ∞ 2Zr2 2Zr 2Zr 4e2 Z 6 ∞ 2 − a1 1 2 − a2 dr2 r2 e− a ] dr1 r1 · e dr2 r2 e ( ) [ + Q= π a r 1 0 0 r1 Partial integration leads to Q = same order of magnitude as E0 . 5Ze2 8a and positive. Notice that the correction Q is of the Summarizing : the two He-electrons have a ground state configuration (1s)2 2 with energy EG = − Zea (Z − 58 ). The first excited state of the H0 -operator corresponds to the electronic configuration (1s)1 (2s)1 , with antisymmetric wave function 1 ψS=0 = √ [ϕ1s (1)ϕ2s (2) + ϕ1s (2)ϕ2s (1)] ⊗ χs 2 1 ψS=1 = √ [ϕ1s (1)ϕ2s (2) − ϕ1s (2)ϕ2s (1)] ⊗ χt 2 The S = 0 wave function belongs to the parahelium thermal scheme, the S = 1-wave functions belong to the orthohelium thermal scheme. Parahelium shows diamagnetism, orthohelium is a paramagnet. Without consideration of V12 para and ortho (1s)(2s) states are degenerate. In the absence of spin-orbit coupling, optical transitions between the two thermal schemes are absolutely forbidden. Should one be able to pump a He atom gas (by electrons excitation or other mechanism) into the triplet excited state, then it will practically stay forever (many months) in that state. Notice that because of spin orbit splitting the triplet states degeneracy is lifted and a fine structure appear in the excitation spectrum of ortho helium. We show now that the result of introducing the Coulomb interaction is the lifting of the degeneracy between triplet and singlet states, leading, in some circumstances, to the formation of a magnetic moment, at least in one excited state 1s2s with different orbital quantum numbers. We consider the four states eigenspace of H0 to (1s)(2s) and solve the eigenvalue problem of H0 + V12 within this space. The Hamiltonian matrix reads E1s + E2s + Q + J 0 0 0 0 E1s + E2s + Q − J 0 0 0 0 E1s + E2s + Q − J 0 0 0 0 E1s + E2s + Q − J mit Z e2 Q = r12 Z e2 ∗ ∗ J = dV1 dV2 ϕ1s (1)ϕ2s (2) ϕ1s (2)ϕ2s (1) r12 dV1 dV2 ϕ21s (1)ϕ22s (2) CHAPTER 1. MAGNETISM IN ATOMS 21 from which the sougth for eigenvalues can be read out immediately. The integral Q is the Coulomb energy. The integral J is the result of the exchange interaction and is called exchange integral. It provides the exchange energy contribution that arises from the correlation of the two electron as a consequence of symmetrizing the wave functions according to the Pauli principle. The Coulomb interaction produces, via Pauli principle, a splitting of the initial degeneracy of the (1s)(2s) configuration: the singlet state has the energy E1s + E2s + Q + J, the triplet state has the energy E1s + E2s + Q − J. In summary: there are two distinct thermal schemes for He, consisting of para (singlet) - and ortho (triplet) states. Levels with the same quantum numbers of the orbital wave functions (e.g. (1s)(1s)) are prohibited in ortho helium. Furthermore, the splitting between ortho and para states depends on the relative sign and strength of Q(n, l) and J(n, l). In the present example, both integrals Qn,s and Jn,s are positive, so that the t-states level is lower than the s-state. The first Hund’s rule states that the positivity of J when degenerate non-equivalent states are involved is a general feature of atoms. The strength of the splitting – the strength of the parameter J which we will call the Hund exchange parameter JHu – is typically of the order of few eV , i.e. the same order of magnitude as the Coulomb interaction, from which the exchange interatcion actually originates. One can formally obtain the t − s splitting by caricaturing the exchange interaction (which is actually acting in the orbital space) with an effective spin Hamiltonian: the Heisenberg-Dirac-Van Vleck operator. Dirac defined the operator acting in spin space HSpin = (E1s + E2s + Q) · E − J · P12 with the exchange operator P12 P12 | + >| + >=| + >| + > P12 | + >| − >=| − >| + > P12 | − >| + >=| + >| − > P12 | − >| − >=| − >| − > The eigenvalues of the operator Hspin when restricted to the 1s2s eigenspace, are identical with the eigenvalues of the physical operator. A useful way of writing P12 is E + ~σ1 · ~σ2 1 ~1 · S ~2 ] P12 = = [E + 4 · S 2 2 where ~σ are the two-by-two Pauli matrices and the product σi σj must be taken as a Kronecker product of matrices. The spin Hamiltonian simulating CHAPTER 1. MAGNETISM IN ATOMS 22 the original Hamiltonian acting in the orbital space reads 1 ~r · S ~s ] HSpin = (E1s + E2s + Q) · E − J · [ E + 2S 2 It can be formally generalized to many electrons: HSpin = X r Er + X 1X 1 ~ ~ Q(r, s) − J(r, s) · [ + S r · Ss ] 2 r,s 4 r6=s where r, s is a set of quantum numbers describing orbital wave functions and a factor 21 has been placed in front of the sums to avoid double counting. Dirac has shown with quite general arguments that the eigenvalues of this effective spin Hamilton operator are correct within first order perturbation theory. The difficulty is one of computing the various coupling constants forthcoming in the operator, which ultimately requires knowledge of the orbital wave functions. Chapter 2 Magnetism in solids In the first chapter we have shown how atomic magnetic moments are produced. On these grounds, diamagnetism and paramagnetism, which are the main manifestation of atomic magnetism at finite temperature, are well understood. However, we still need to answer the broad question about ” why Fe is ferromagnetic”. This means that we need to understand 1. what happens to an atomic magnetic moment when it is embedded into a ”see” of free electrons, 2. how do magnetic moments (if they ”survive” the contact with the free electrons) couple to align along the same direction (what makes the triplet state between spins on different lattice sites energetically preferred with respect to the single state) and what is the strength of the exchange energy distinguishing between triplet and single state on different lattice sites. interaction 2.1 Stoner-Wohlfahrt model Let us now introduce the Stoner-Wohlfahrt (SW) model of magnetism in solids. In a solid, electrons have, in reality, wave functions which, for some kind of inner shell electrons, like the d-electrons in transition metals, have a strong component localized in the vicinity of the ion core (remember the tight-binding model). So to speak, they spend a lot of time close to the ion core: during this time they are almost atomic like and can feel the Hund rule as we know it from atomic physics. We will see that this Hund rule provides ultimately the net persistence of the net magnetic moment at each atom in the solid state, although the contact with free electrons acts to destroy the magnetic moment, in general. Although the atomic magnetic moment might persist when the atom in embedded into a solid, there are two important experimental features that distinguish atomic magnetism from magnetism 23 CHAPTER 2. MAGNETISM IN SOLIDS 24 in the solid state. First, in atoms the formation of the magnetic moments is the result of the Hund rules that weight correctly, at least empirically, the orbital angular momentum and the spin of the electrons into forming the total angular momentum of the many electrons in one atom. The total angular momentum produces the net atomic magnetic moment, which is ultimately responsible for the occurrence of magnetism in matter. When the magnetic moment per atom is measured in the solid state, it appears that only the spin part of the total angular momentum is contributing to it: one speaks of ”quenching” of orbital angular momentum, in particular in transition metals, by the crystal field. Second, in angular momentum quenched metals, magnetic moments per atom should be an integer multiple of 2µB , i.e. their magneton number is an integer. This would lead, for example, to atomic magnetic moments of 2,3, respective 4 µB for Ni, Co and Fe. The experimentally measured value in bulk Fe is 0.616, 1.715 and 2.216 µB . The Stoner-Wohlfahrt-Slater model of magnetism provides the correct framework to explain the occurrence of 1. finite magnetic moments in solids and 2. the existence of non integer magneton numbers. While the formation of a magnetic moment, because of the first Hund rule, is almost the rule in atoms, it is a very rare event in solid, where the electrons can be considered as delocalized and are therefore better described by a band structure. The SW model in its simplest version considers free electrons where energy levels are filled up to the Fermi radius kF = (3π 2 N/V )1/3 , N/V being the electron density. In virtue of this filling ~2 k2 the electron gas has a total kinetic energy amounting to Ekin = N 35 2mF . The non magnetic ground state foresees that all states up to EF are filled with two electrons carrying opposite spins. Introducing an exchange interaction shifts the energy levels of minority electrons to higher energies, while the energy levels of majority spin electrons are shifted downwards. This produces an energy gain that actually favors the relative shift of energy bands and, ultimately, the formation of a magnetic moment. On the other side, the radius of the Fermi sphere must be increased to host all the electrons, as the double occupancy of each level is no longer possible. This produces an increase of the total kinetic energy that goes against the formation of a magnetic moment. Therefore, the formation of the magnetic moment is the result of a delicate energy balance and is subject to strong restrictions. Let us now work out the Stoner criterion for ferromagnetism, i.e. for the existence, in a Fermi gas subject to exchange interaction, of an imbalance between spin up and spin down electrons. We start from a density of CHAPTER 2. MAGNETISM IN SOLIDS 25 states n0 (E) common to both spin channels and change the energy levels ± according to E~k,ν = E~k,ν ∓ 21 I · M , where + refers to majority spins. I is the intra atomic exchange interaction responsible for the Hund’s rule in atoms, and P is the sought for spin imbalance, M = N↑ − N↓ . Accordingly, the density of states separates out for the two spin channels according to ( n± E) = n0 (E ∓ 21 I · M ). Integrating up to the (to be determined) Fermi Figure 2.1: DOS for majority spins and minority spins are shifted by an amount IM . energy EF gives the number of electrons per unit cell and the total moment per unit cell: Z EF (M ) IM IM N= [n0 (E + ) + n0 (E − )] 2 2 0 Z EF (M ) IM IM ) − n0 (E − )] (2.1) M= [n0 (E + 2 2 0 These are two equations for the sought for parameters EF (M ) and M . Solving the first one can obtain in principle EF (M ). Inserting this in the second one we obtain an implicit equation for M : M = F (M ); Z EF (M ) IM IM F (M ) = [n0 (E + ) − n0 (E − )] 2 2 (2.2) CHAPTER 2. MAGNETISM IN SOLIDS 26 The function F (M ) has following important properties: 1. F (0) = 0 2. F (−M ) = −F (M ) , i.e. EF (−M ) = EF (M ) 3. F (∞) = M∞ and −M∞ < F (M ) < +M∞ 4. dF dM |M =0 ≥ 0. Proof: From ′ F (M )M =0 IM IM I 0 0 n (E + ) + n (E − ) = 2 2 2 M =0 IM IM dEF + n0 (E + ) − n0 (E − ) (2.3) 2 2 dM M =0 we obtain F ′ (0) = I · n0 (EF ) ≥ 0 M∞ is the largest magnetic moment by complete spin polarization of the electron gas and corresponds to the first Hund’ rule magnetic moment. Under these conditions, the graphical solution of the implicit equation for M has two possible scenarios: either M = F (M ) has only the solution M = 0 for F ′ (0) < 1, or it has two solutions: M = 0 and M finite but not necessarily an integer for F ′ (0) > 1. In this case one can show that the M = 0 solution maximizes the total energy, while the two solutions with opposite sign are the sought for minima that establish a finite spin imbalance in the ground state. Accordingly, the Stoner criterion for ferro magnetism reads I · n0 (EF ) > 1 (2.4) I is essentially an atomic quantity of the order of 0.7 eV for 3d atoms. The tendency to ferro-magnetism therefore requires a high density of states of the non-spin polarized band structure at the Fermi level. This can only be achieved when the states close to the Fermi level are sufficiently localized, i.e. their bandwidth is small enough, as is the case for d metals with partially filled d shells. In the following figures we will illustrate the Stoner criterium on a set of examples, starting with single magnetic impurities in metals. CHAPTER 2. MAGNETISM IN SOLIDS 27 Figure 2.2: Graphical solution of the implicit equation for M . In A the only solution is M = 0, in B a finite magnetization minimizes the energy. Figure 2.3: Left: Local DOS of Mn in Ag according to LSDA computations by R. Podloucky et al. Phys. Rev. B33, 5777 (1980). The spin splitting is about 3 eV (experiment: 4 eV). Right: Computed and measured values of 3d- impurity atoms in Ag, Cu and Al. The highest moment appears always in the middle of the 3d-row. CHAPTER 2. MAGNETISM IN SOLIDS Figure 2.4: 28 CHAPTER 2. MAGNETISM IN SOLIDS 29 Figure 2.5: FeRh: dispersion curves for the two spin directions; (a) spin+, (b) spin(-). The broken line is the Fermi level. CHAPTER 2. MAGNETISM IN SOLIDS 30 Figure 2.6: top: Band structure of F e along the ∆ direction. The minority spin bands (broken lines) are shifted toward higher energies with respect to majority spin bands (continuous lines). The symmetry label of the bands is also shown. In contrast to the Stoner model, the shift is not exactly rigid but depends slightly on ~k. On the top of the figure, the ~k-points selected by the used photon energy are indicated. Bottom. On the left is an energy resolved photo emission spectrum taken at normal emission from a (100)-surface of Fe. In the middle, the same electrons are analyzed for their spin polarization. on the right, the photo-emission intensity for the two spin channels is plotted separately, by suitably compounding the total intensity and the polarization data. The two peaks are identified as due to electrons originating from the spin split bands close to the Γ-point, see top. CHAPTER 2. MAGNETISM IN SOLIDS 2.2 2.2.1 31 Friedel-Oscillations The interatomic exchange interaction The long sought explanation for the origin of ferromagnet was provided after years of research and illustrated e.g. in a review article by M.B. Stearn, Physics Today, April 1978, p.34. The first condition for ferromagnetism is that we have some localized magnetic moments, and this conditions is met in Fe by the localized d-electrons, which keep a part of their atomic magnetic moment produced by the Hund-rule intra atomic exchange. The second condition for ferromagnetism it that all these moments line up parallel to each other, i.e. triplet state coupling between neighboring spins. However we know that the chemical bonding between equivalent orbitals favors, in line with the Pauli principle, the singlet state, so that we need a mechanism that acts ”against” the Pauli principle in order to get triplet coupling. This alternative mechanism is provided, according to Stearns, by the indirect exchange between localized d-electrons through RKKY coupling with the delocalized part of the d-wave functions (or with the s-like electrons). This mechanism is therefore based on the existence of ”degenerate” states in the band structure of solids. The RKKY coupling mechanism is also a central one in modern research on magnetism and we want to illustrate its peculiarities with a simple, computable exact model which is also relevant in thin films coupling phenomena. Notice that the singlet coupling underlying the chemical bond and the absence of magnetism associated with it in the ground state is a ”robust” result, in the sense that there exists a very strong theorem by Lieb and Mattis that states that in a linear arrangement of atoms the non-magnetic state, i.e. the state with lowest total spin, is the ground state. One needs to go higher than one dimension to escape this theorem, because in higher dimensions electrons states with different symmetry – atomic orbitals with different quantum numbers – can hybridize: it is this degeneracy between orbital wave function with different symmetry that provide a route to escape the strong Pauli principle that favor antiparallel alignment between orbital states with the same quantum number. The situation is exactly the same as in atoms: only if the electronic states participating to the formation of the magnetic moment have different quantum number, the exchange interaction can act to lower the energy of the triplet state. The situation can be therefore summarized as follows. The chemical bond between same orbitals centered at different atoms favors the antiparallel ground state, in virtue of Pauli principle. The crystal potential, however, can act to mix different symmetries and different orbitals into the wave functions forming the valence bands in solids. As in CHAPTER 2. MAGNETISM IN SOLIDS 32 atoms, different symmetries might favor energetically the parallel coupling, thus producing ferromagnetic alignment between neighboring atoms. However, it depends on the crystal potential and on the orbitals involved, whether a total spin in the ground state is formed or not. We point also out that the strength of the effective exchange energy that favors the triplet coupling between two different sites (the interatomic exchange interaction) is one to two orders of magnitude smaller that the onesite (intra-atomic) exchange interaction (which amounts to about 3 − 5 eV). This means rotating one spin in the presence of the other ones needs much less energy that suppressing the magnetic moment. It is the interatomic exchange interaction which is relevant for determining the temperature scale at which collective ferromagnetic order vanishes (in the next chapter, the so called Curie temperature). 2.2.2 RKKY oscillations The presence of a more or less localized magnetic moment, made of d-wave functions creates a potential sink with the strength of the s − d exchange interaction at the location of the magnetic moment for majority (spin up) s-electrons, in virtue of the atomic Hund-rules. The minority spin down electrons can be considered as non-affected by the impurity. A local perturbation in one spin channel produces an oscillating density in the affected spin channel (see the Anhang at the end of this chapter), while the other spin remains uniformly distributed. This produces a local spin polarization of the electron gas surrounding the impurity P= 2· < Sz > ρ+ − ρ− κ cos 2kF x = + ≈ O( ) − ~ ρ +ρ kF x (2.5) that propagates far away from the perturbing magnetic moment. At some location x within the spin polarized s-electron gas a spin imbalance appears. This spin imbalance acts as an effective exchange field for d-waves functions and tends to align a d-derived magnetic moment at that location parallel to itself: A magnetic moment at the location x would lower his energy by aligning along the direction of P. In this way, the exchange interaction can propagate, oscillating between positive and negative depending on the position x and can couple spins which are quite distant from each other. CHAPTER 2. MAGNETISM IN SOLIDS 33 Figure 2.7: a): the left-hand side shows a typical hysteresis curve (M versus magnetic field H) recorded for exchange-coupled Co films. At the shift field H = Hj the magnetizations of the individual films are aligned to the direction specified by the external magnetic field. The critical field Hj is measured as a function of the Cu spacer thickness τ by scanning a focused laser beam over a wedge-like multi layered structure, shown schematically on the right-hand side. b): Hj versus τ for a room temperature grown wedge-like multi layered structure. A finite shift field means AFM coupling in the ground state. A vanishing shift field means FM coupling. The thickness of the Co films are 13.2 ML and 15.8 mono layers, respectively. Inset, the Fourier transform, the two peaks corresponding to the two periodicities 2.4 Ml and 5.4 ML. The long period dominates. c): as b) but with the Cu wedge and the final Co film deposited and measured at 160 K. The short period now dominates, see Fourier transform). 34 CHAPTER 2. MAGNETISM IN SOLIDS 2.2.3 Anhang: mathematical details of the model We introduce in a 1d free electron gas a perturbing potential localized at the origin and look for the total charge density produced by it at a location x – essentially a continuation of the Anderson impurity model to. In order to simplify the mathematics, we consider a Dirac-Delta like perturbation potential of strength λ located at the origin of a one dimensional solid filled with a free electron gas. As we are interested here at the behavior of the charge (or spin) density far away from the location of the perturbing potential, we do not use the more realistic but also more cumbersome traditional potential well with finite width Let us consider a segment extending from −L/2 to +L/2 along the x-axis and establish in it a potential V (x) = λ · aδ(x). a represents the width of an hypothetical well and λ its strength. We refer to the segment with x < 0 as the left-hand side l and to the segment with x > 0 as the right-hand side r. We have to solve the Schr”odinger equation [ −~2 + V (x)]ψ(x) = Eψ(x) 2m (2.6) under the boundary conditions for a δ like potential ψl′ (0) − ψr′ )0) − ψl (0) = ψr (0); 2mλa ψ(0) = 0 ~2 (2.7) In the two regions l and r the respective solutions have to fulfill the SE −~2 ψ(x) = Eψ(x) 2m (2.8) We distinguish two cases: E < 0 and E > 0. For E < 0 the SE away from the singularity has two solutions, one growing exponentially toward ±∞, the other decaying exponentially toward ±∞. This last solution is the only physical one as it has a finite norm and produces a bound state with energy amounting to Eb = − . (κ = λam ) ~2 ~2 κ 2 mλ2 a2 =− 2m 2~2 (2.9) and wave function ψlb (x) = √ κeκx ψrb (x) = √ κe−κx (2.10) CHAPTER 2. MAGNETISM IN SOLIDS 35 In the range E > 0 we have free electrons moving left and right, and the solutions in each range l, r, under periodic boundary conditions at ±L/2, are ~2 k 2 2π ; k= ·n 2m L r r r r 2 2 2 2 sin kx + Bl cos kx); ψr (x) = Ar sin kx + Br cos kx) ψl (x) = Al L L L L both basis functions being normalized to 1/2 in the range x ∈ [± − L/2, 0]. The boundary conditions at x = 0 read 2mλa Bl = Br ; k · (Al − Ar ) − Bl = 0 (2.11) ~2 E= There are two classes of wave functions fulfilling these conditions. One class has A = Ar ⇒ r rl 2 2 u u sin kx; ψr (x) = sin kx ψl (x) = L L where the total wave function is normalized to 1 over the segment with length L. Bl = 0; . The second class has Bl = Br = B 6= 0 and must be even under change of sign, so that Ar = −Al = A and B = Ak . This type of wave κ functions read r r 2 k 2 g sin kx + cos kx) ψl (x) = A[ L κ L r r 2 k 2 g ψr (x) = A[− sin kx + cos kx) L κ L A must be chosen so that the entire wave function is normalized to 1 in the range x ∈ [−L/2, L/2]. This means κ (2.12) A= √ κ2 + k 2 and r r 1 2 2 g ψl (x) = √ [ κ · sin kx + k · cos kx) 2 2 L L κ +k r r 1 2 2 g ψr (x) = √ [− κ · sin kx + k · cos kx) L L κ2 + k 2 36 CHAPTER 2. MAGNETISM IN SOLIDS Using these wave functions, we compute the total charge density at a point, e.g. x ≥ 0: Z kF h i (κ sin kx − k cos kx)2 2 L 2 −2κx · dk + sin kx ρ(x) = κe + · L 2π 0 κ2 + k 2 Z kF h 1 κ2 cos 2kx + κk sin 2kx i −2κx = κe + · dk 1 − π 0 κ2 + k 2 Z ∞ h 2 κ cos 2kx + κk sin 2kx i 1 kF dk − = κe−2κx + π π 0 κ2 + k 2 Z ∞ h 2 κ cos 2kx + κk sin 2kx i 1 dk + π kF κ2 + k 2 One can prove by complex integration that i kF κ·i −2κx h ρ(x) = + e · E1 (−2κx + 2ikF x) − E1 (−2κx − 2ikF x) π 2π E1 (z) being the exponential integral. Notice that the contribution of the bound state to the total charge density cancels out with part of the contribution of the free electron states. The following figure plots the charge density as a function of the variable x for some characteristic values of κ and k. Some particular limits are worked out now. (E1 (z) ≈ −γ − lnz) For small x we have kF κ kF kF κ + (π − arctan ) − O(x) ≈ + − O(x) π π κ π 2 For large z we have e−z E1 (z) ≈ z and accordingly ρ(x) ≈ e2κx+2ikF x e2κx−2ikF x − = −2κx + 2ikF x −2κx − 2ikF x h i 4ie2κx κx sin 2k x − k x cos 2k x F F F 4κ2 x2 + kF2 x2 and in the limit of large x (and small κ) we obtain ρ(x) ≈ κ cos 2kF x kF + O( ) π kF x (2.13) CHAPTER 2. MAGNETISM IN SOLIDS 37 Figure 2.8: This result underlines the formation of a Friedel-like oscillation of the charge density away from a localized perturbation. The wave-length of the oscillation is kπF . The oscillation decays as the inverse of the distance from the the perturbation (this is in 1d: in 2d we have a decay as the square, in 3d as the third power of the inverse distance. 38 CHAPTER 2. MAGNETISM IN SOLIDS Part II A. Vindigni Chapter 3 Magnetic order at finite temperature In the first two Chapters • we have shown how magnetic moments are created at the atomic level according to the Hund’s rules (intra-atomic exchange interaction); • we commented on how atomic magnetic moments, deduced assuming spherically symmetric surrounding (Hund’s rules), generally reduce when the atom is “put” in a crystal, which lowers the symmetry of its environment (Stoner-Wohlfahrt model); • we have shown how an interatomic exchange interaction can arise in a metal by means of the RKKY interaction; • we defined the conditions under which a metal may show ferromagnetic coupling between different magnetic moments. Already at that level, it was clear that ferromagnetism is not the rule but rather an exception, in the sense that many factors that are encountered in ordinary materials usually prevent the formation of magnetic moments or that of a ferromagnetic interatomic coupling. All the above-mentioned properties1 have been deduced neglecting the temperature or, in other words, they are ground-state properties. In this Chapter we discuss the consequences of introducing the temperature. The general trend is that thermal fluctuations destroy the ground-state ferromagnetism (when present). In the same line as before, we will define some conditions under which ferromagnetism can “survive” at finite temperatures as well. 1 apart from the Brillouin function. 39 CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 3.1 40 Coupled effective spins: an N -body problem In the previous Chapters we have considered the conditions under which an isolated atom possesses a finite magnetic moment. Hund’s rules allow computing the ground-state multiplet, characterized by the total angular momentum (orbital plus spin contribution) which results from all the unpaired electrons. When dealing with coupled magnetic moments, we will indicate the atomic total angular momentum with S i) in order to avoid confusion with the exchange interaction (J) and ii) because – in this context – people often speak about “spin” or “effective spin” to indicate the total single-atom angular momentum. As far as a single isolated atom is concerned, its magnetic moment at finite temperature is well described by the Brillouin function: gµB S B z m = −gµB hŜ i = gµB S BS , (3.1) kB T with α 1 2S + 1 α − coth 2S 2S 2S gµB S B . kB T (3.2) In the derivation of this function, we have implicitly used the knowledge of i) the eigenstates of the atom in an external applied field and ii) the way of performing thermal averages for a quantum system (see Appendix). Note that the intra-atomic and the Zeeman interaction have been treated on a different ground: we have considered only the ground-state multiplet (which minimizes the intra-atomic exchange interaction) but we have applied Boltzmann statistics to the levels of this multiplet in case they have been split by an external field (Zeeman interaction). The reason for such a different treatment reside in the characteristic energy scales of the two interactions in relationship with the thermal energy kB T . In fact, the intra-atomic exchange energy is of the order of 4 − 10 eV∼ 105 Kelvin, while the Zeeman splitting is roughly 0.1 meV ∼ 1 Kelvin for one-Tesla applied field. Further on, we have seen how a ferromagnetic interatomic exchange interaction is necessary for the occurrence of ferromagnetism in a solid. Under specific and relatively strict conditions, this goal is attained by means of the RKKY interaction2 . The order of magnitude of the RKKY interaction is 10 − 50 meV ∼ 100 − 500 Kelvin. Thus, depending on the material and the 2S + 1 coth BS (α) = 2S 2 and α = Other mechanisms are responsible for exchange interaction, e.g. super-exchange or direct exchange, in insulators. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 41 Typical exchange energies and magnetic moments Fe Co Ni Gd Dy EuO EuS Tc [K] 1043 1395 629 289 87 69.4 16.5 C[K] 2.22 2.24 0.588 4.68 3.06 µ[µB ] 2.22 1.71 0.605 7.1 J[eV] 0.012 0.015 0.013 0.00025 J[k] 139 174 151 2.9 Ms [Gauss] 1746 1446 0.510 2.060 2920 1930 1240 Table 3.1: Some typical values for the energy scales, Curie temperature, Curie constant and saturation magnetization Ms = gµB S/a3 (a lattice constant). Often it turns out useful to express the exchange interaction in Kelvin units: 1 eV ≃ 1.16 × 104 K. temperature range of interest, a statistical-mechanics treatment is required for the RKKY interaction as well. The competition between this interatomic exchange interaction and thermal fluctuations is indeed responsible for the loss of ferromagnetism above a certain temperature, called Curie temperature TC . Table 3.1 reports the values of the Curie temperature, the exchange interaction and other relevant parameters for few typical magnetic materials. Let us come back to the formal treatment of magnetism at finite temperatures, restricting ourselves to ferromagnetic exchange interactions. A system of coupled magnetic moments arranged in a lattice can then be described by the Hamiltonian X X 1 H=− J Ŝ(n) · Ŝ(n′ ) + gµB B Ŝ z (n) . (3.3) 2 ′ n |n−n |=1 The dimension of the Hilbert space associated with this quantum many-body problem scales as (2S + 1)N , N being the number of magnetic moments (spins) in the lattice. Due to such an exponential dependence on N , the exact treatment of a system of many coupled spins becomes intractable – even numerically – as far as the number of spins approaches that of realistic extended systems3 . In practice, one can try to circumvent this problem in several ways: 3 Some effective zero-dimensional structures (magnetic clusters or nanoparticles) are also studied in the context of nanomagnetism. For some of these systems, exact diagonalization of the associated quantum problem is still feasible numerically and makes it possible to describe their magnetic behavior at any temperature. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 42 1. Reduce the many-body problem to a single-particle problem. This corresponds to the mean-field approximation (MFA). 2. Simplify the problem replacing the quantum-spin operators by classical vectors. 3. Take advantage of specific symmetries in the problem under investigation and use a Hamiltonian which can easily be diagonalized. 4. Consider only a selected family of excitations of the ground state, which can have either local (domain walls) or non-local (spin waves) character. 3.2 Mean-field approximation (MFA) The goal of the Mean-Field Approximation (MFA) is to reduce the manybody problem (3.3) to the a single-particle problem. This means to get rid – somehow – of terms which directly involve two-spin operators such as Ŝ(n) · Ŝ(n′ ). In this context, we understand a paramagnet as the reference single-particle problem. We will first make use of the Brillouin function (3.1) Many-body problem Single-particle problem MFA effective field Figure 3.1: Sketch of the idea behind the mean-field approximation. to write down the MF equation of state heuristically and discuss its relevant implications. Further on, within the more rigorous Landau approach, we will prove that the MFA is actually the best approximation of the Hamiltonian (3.3) in terms of a single-particle Hamiltonian. Equation of state Referring to the sketch in Fig. 3.1, we may think that the actual field experienced by each spin in a ferromagnetic sample contains a contribution CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 43 arising from the interaction with its neighbors, besides the typical Zeeman term (due to the interaction with the external, applied field). More explicitly, we assume that the physics of each spin can be described by a single-particle Hamiltonian of the form Hsp (n) = gµB (B + BW ) Ŝ z (n) (3.4) where gµB BW = −z̄JhŜ z (n)iHsp . The origin of the Weiss field BW is the interatomic exchange interaction, whose effect is taken into account only as an average and not rigorously. Such an average is performed using the Hamiltonian (3.4) itself and z̄ indicates the number of nearest neighbors of each spin. As anticipated, the Hamiltonian Hsp is equivalent to the one of a paramagnetic atom in a magnetic field B t = B + BW so that the thermal average of the Ŝ z (n) projection is given by the Brillouin function: with hŜ z (n)iHsp = −SBS (α) (3.5) gµB S B − z̄JShŜ z (n)iHsp gµB S B t = . α= kB T kB T (3.6) Since the average hŜ z (n)iHsp is also contained in α, i.e. the argument of the Brillouin function, Eq. (3.5) is actually a self-consistent equation. To write Eq. (3.5) in a more transparent way, we exploit the relation between the average of the spin component along the field and the associated magnetic moment m = −gµB hŜ z (n)iHsp . The MF equation of state finally reads: z̄J S m gµB S B . (3.7) + m = gµB S BS kB T gµB kB T In order to visualize the solution of Eq. (3.7) graphically, it is convenient to set m = BS (α) gµB S gµB B kB T α− . σ= 2 z̄JS z̄JS ( σ= (3.8) Let us list the most remarkable facts arising from the graphical analysis of solutions depending on the external parameters T and B. 1. When B = 0, there exists a non-trivial solution with σ 6= 0 only if the slope of the Brillouin function exceeds that of the straight line (the CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 44 second one of Eqs. (3.8)). In particular, expanding the former around α ≃ 0 yields S+1 α + ... (3.9) BS (α) ≃ 3S so that a spontaneous magnetization (σ 6= 0) only arises for T < TC with S + 1 z̄JS 2 TC = . (3.10) 3 S kB One can show that if solutions with σ 6= 0 exist, they have a lower free energy than the solution corresponding to σ = 0. 2. For T < TC , the system of Eqs. (3.8) admit two graphical solutions of opposite sign in the region B ∈ [−Bc , Bc ]. Outside of this interval the solution is unique and with σ > 0 (σ < 0) for B > 0 (B < 0). 3. For T > TC and small α the Brillouin function can again be linearized and the system of Eqs. (3.8) takes the simplified form (σ = S + 1 α 3S kB T gµB B σ= α− ; 2 z̄JS z̄JS (3.11) by using the definition of TC given in Eq. (3.10), the solution of the previous set of equations can be written as T gµB 1− σ=− B (3.12) TC z̄JS or equivalently (using again Eq. (3.10)) 1 (gµB )2 TC (gµB )2 S(S + 1) B= B. z̄J T − TC 3kB T − TC (3.13) The pre-factor of B on the right-hand side is the susceptibility (computed in B = 0) C χ= (3.14) T − TC which is the well-known Curie-Weiss law with m = gµB Sσ = C= (gµB )2 S(S + 1) 3kB (3.15) being the Curie constant (already encountered when discussing paramagnetism). CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 45 Equation (4.49), together with other MF predictions, is not expected to hold true in the vicinity of TC . In fact, in this critical region the neglected terms (fluctuations) play a major role. This statement should sound clearer at the end of this chapter. Keeping in mind the limitations of the MFA, it is still interesting to investigate how different observables should behave according to the MFA as a reference framework for introducing critical phenomena. 3.3 Mean-field universality class Now we will discuss some consequences of this equation of state (3.7) in the vicinity of TC . Since the results presented in this section (3.3) do not depend on the value of the effective spin S, we will deduce some scaling relations for the simplest case: S = 1/2. For this particular case, BS (α) = tanh(α) and TC = z̄J/4kB . Additionally, we assume that S = 1/2 refers to the spin of an electron4 so that g = 2. Under these hypotheses, Eqs. (3.8) can be written in the compact form µB B T c σ = tanh + σ . (3.16) kB T T By using the fact that tanh(α) ≃ α − 31 α3 for α ≃ 0, Eq. (3.16) can be expanded for small σ and B as follows: µB B T c 1 σ= + σ − σ 3 + O(Bσ 2 ) (3.17) kB T T 3 which, for T ≃ TC and neglecting higher infinitesimal than σ 3 , becomes a polynomial of the reduced temperature τ = (T − TC )/TC : 1 1 µB B T − TC σ + σ3 = τ σ + σ3 . = (3.18) kB T TC 3 3 Equation (3.18) is suitable for deriving some critical exponents. Mean-field critical exponents 1. Setting B = 0, one has 4 T − Tc Tc 1 σ + σ3 = 0 3 In fact, any system whose ground state is two-fold degenerate with a degeneration that can be removed by the application of an external field can be thought of as possessing an effective spin 1/2, with g 6= 2 in general. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE whose solutions are σ=0 β MF √ 1 T σ ≃ 3 1− ∝ (−τ ) 2 T | {z c } 46 for T ≥ Tc for T < Tc . :=−τ This result provides the value of the critical exponent β within the mean-field approximation: β MF = 1/2. 2. Now we want to evaluate the behavior of the susceptibility around TC . First, let us recall the proportionality relation ∂M ∂σ χ(T, B = 0) = . ∼ ∂B B=0 ∂B Then the derivative ∂σ/∂B can easily be put in relationship with the reduced temperature and σ by differentiating both sides of Eq. (3.18): ∂σ ∂σ T µB · + σ2 ∼− 1− kB TC TC ∂B ∂B Since the relevant infinitesimal quantity is the reduced temperature τ , we have identified T = TC on the left-hand side of the equation above. For T > TC , we can further neglect the term containing σ 2 so that ∂σ 1 ∼ µB ∂B T − TC ⇒ χ(T ) ≃ C T − TC for T > TC . (3.19) This is nothing but the Curie-Weiss law deduced in an alternative way in Eq. (4.49). For T > TC , instead, we have to take into account that σ 2 ≃ −3τ . In this case one has ∂σ µB 1 ∼ ∂B 2 TC − T ⇒ χ(T ) ≃ C 1 2 TC − T for T < TC .(3.20) The Eqs. (3.19) and (3.20) give the mean-field prediction for another critical exponent: γ = 1. Critical exponents in general The fact that these observables behave like powers of the reduced temperature τ close to the transition point is not an artifact of the MFA. On the CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 47 contrary, this feature defines the condition of criticality. Other critical exponents can be deduced similarly to β and γ. Below we recall the definition of some of them. Letting τ = (T − TC )/TC be the reduced temperature, α, β, γ and δ critical exponents are defined as follows: C(τ, B = 0) ∼ |τ |−α M (τ, B = 0) ∼ (−τ )β , χ(B = 0, τ ) ∼ |τ |−γ τ <0 |M (τ = 0, B)| ∼ |B|1/δ . Another important feature captured by the MFA is that critical exponents do not depend on the details of the model, e.g., on J and the details of the lattice (dimensionality or z̄). However, this universality of mean-field critical exponents is somewhat exaggerated: for instance the correct critical exponents do depend on the dimensionality of the lattice. In the following table the critical exponents obtained within the MFA (classical values of the critical exponents) are compared with the numerical values obtained for the 3d Heisenberg model (see the following sections): α β γ δ MFA 3d-Heisenberg 0 (Jump) −0.11 ± 0.006 0.5 0.365 ± 0.002 1.0 1.386 ± 0.004 3.0 4.46 Experimentally, critical exponents are independent of the values of Ms , S, g, TC , J, etc. which are specific of a given magnetic material. They rather depend on more general symmetries of the experimental system under investigation. The concept of universality class is associated with this property. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 3.4 48 The Landau approach There is no univocal way of introducing it nor the literature is consistent in what it is meant by Landau approach to critical phenomena. The procedure proposed here aims at highlighting how the MFA is a drastic simplification of the more general theoretical-field approach based on functional integrals. We will try to avoid confusion between the Landau free-energy functional and the MF (Gibbs) free energy. Finally, we should be able to discuss some qualitative arguments whose validity goes beyond the mean-field approach. To avoid useless complications, we refer to S = 1/2 operators and assume that only their z component enters the spin Hamiltonian: 1 H=− J 2 X |n−n′ |=1 Ŝ z (n) Ŝ z (n′ ) + gµB B X Ŝ z (n) . (3.21) n Later on, we will discuss the Ising Hamiltonian (3.21) in some more details. For now we use only the fact that the corresponding energy levels can be written in terms of two-valued classical variables σi = ±1: H [{σ}] = − X 1X 1 σi , σi Ji,j σj + gµB B 2 i,j 2 i (3.22) where σ = (σ1 , σ2 , . . . σN ) and Ji,j is a symmetric matrix describing the (exchange) coupling among different spins in the lattice5 . The partition function associated with the Ising Hamiltonian reads ( " #) X 1 X (3.23) σi Ji,j σj − h σi Z (B, T ) = T r exp β {σ} 2 i,j i with β = 1/kB T and h = βgµB B/2. The summation over all the configurations {σ} can be performed analytically only for the one-dimensional lattice (Ising chain) and in 2d for B = 0. Before proceeding, it is useful to make a mathematical digression and recall the well-known Gaussian identity r Z 2π s2 /2κ 2 −sη η −κ dη = e 2 e (3.24) κ ℜ which can easily be obtained from the integral of a Gaussian by completing the square at the exponent. With some more efforts, the above result can be 5 Referring to Hamiltonian (3.21) the non-zero terms of Ji,j equal J/4. However, the matrix Ji,j can describe a more general coupling among spins. 49 CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE generalized to N variables η = (η1 , η2 , . . . ηN ) ∈ ℜN : s (2π)N 1 T 1 T −1 dη1 , dη2 . . . exp − η Mη − s · η dηN = exp s M s . 2 det(M ) 2 (3.25) Apart from a constant pre-factor, the right-hand side of Eq. (3.25) becomes equal to the term involving spin pairs in Eq. (3.23) when M−1 i,j = βJi,j and s = σ are chosen. The partition function of the Ising model (in any dimension) can, thus, be written as Z −η T J−1 η/2β −σ·η −σ·h e dηN (3.26) e dη1 , dη2 . . . e Z (B, T ) = N T r Z {σ} where6 h = (h, h, . . . h) and N is a constant irrelevant for magnetic observables. The trace over the variables {σ} appearing in Eq. (3.26) can now be performed analytically ) ( X − P (h+ηi )σi Tr e i e−(h+ηi )σi = 2N Π cosh(h + ηi ) . (3.27) =Π i {σ} σi =±1 i In order to write the partition function in (3.26) in a more transparent way, we make the linear change of variables η = βJφ: ! Z X 1 Z (B, T ) = N ′ dφ1 . . . exp − βφT Jφ 2N Π cosh h + β Ji,j φj dφN . i 2 j (3.28) The latter is usually expressed in a more compact form Z Z (B, T ) = D[φ] e−βL[{φ}] where the symbol R D[φ] ∝ R (3.29) Π dφi stands for the functional integral and i " !# X X 1X ln cosh h + β Ji,j φj φi Ji,j φj −β −1 −β −1 N ln(2) L {φ} = 2 i,j i j (3.30) for the Landau free-energy functional. In this representation Z has been rewritten as a Gaussian average over the auxiliary fields φ of the partition 6 The vector h has this simple form because a uniform B has been assumed but – in principle – different sites could experience different external fields. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 50 function of a paramagnet, experiencing the external field plus the auxiliary fields themselves. By doing this, we have somehow “traded” the original spin-spin interaction with the coupling (in principle site-dependent) of each spin with a set of auxiliary fields φj . Note that no assumption has been made on such fields which can, thus, span all over ℜN . From Eqs. (3.29) and (3.30) an implicit form for the averaged spin projections hσi i can be deduced. This is obtained straightforwardly if we let the field h be site dependent: ! X ∂ ln(Z) 1 ∂Z hσi i = − (3.31) =− = −htanh h + β Ji,j φj i{φ} , ∂hi Z ∂hi j where h. . . i{φ} stands for average over the auxiliary fields. We will come back to this result in the following. Generally, performing the functional integral in Eq. (3.29), i.e. tracing over the auxiliary fields {φ}, is far from being trivial. The simplest approximation which can be made to evaluate Z is replacing the functional integral by the maximum value of the integrand, namely Z n o −βL[{φ}] −βL[{φ}] MFA = exp −βMin L {φ} = Max e Z (B, T ) = D[φ] e {φ} {φ} (3.32) which is known as saddle-point approximation. This is equivalent to the mean-field approximation. In fact, by requiring ∂L/∂φi = 0 for φi = φ̄i , the following equation is obtained ! X φ̄i = tanh h + β Ji,j φ̄j . (3.33) j As we are considering nearest-neighbor ferromagnetic exchange coupling, the solution to the previous equation turns out to be independent of the site index i, meaning that the field which minimizes the Landau free-energy functional is spatially homogeneous. Consequently, Eq. (3.33) is equivalent to the MF equation of state (3.16). It is worth remarking that, within this framework, only when it is evaluated in its minimum the Landau free-energy functional L {φ} acquires the meaning of Gibbs free energy: MFA F (B, T ) = −β −1 ln (Z) = Min L {φ} = L {φ̄} . {φ} (3.34) The equivalence between the MFA and the saddle-point approximation of the functional integral (3.32) allows establishing that the average of the CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 51 auxiliary fields φ is proportional to local magnetic moments. In other words, the average φ̄ plays the role of the Weiss field BW (apart from constant factors). As far as the critical behavior (T ≃ TC ) is concerned, it makes sense to expand the Landau free-energy functional for small values of the fields φ. To this aim we make use of the Taylor expansion 1 1 ln [cosh(x)] = x2 − x4 + O(x6 ) for 2 12 x ≃ 0. (3.35) After some algebra and taking the continuum limit φi → φ(x) one obtains Z 1 b 2 λ 4 d 2 LGL [φ] = (3.36) J (∇φ) + φ + φ d x − β −1 N ln(2) 2 2 4 with ( b = z̄J 1 − 41 β z̄J = z̄J 1 − TTC ≃ z̄Jτ 4 TC 3 kB TC ≃ 16 k T λ = 13 β 3 12 z̄J = 16 3 T 3 B C (3.37) where we have used the fact that the MF Curie temperature is TC = z̄J/4kB for a system of spins one-half (see Eq. (3.10)). (The subscript in LGL stands for Ginzburg-Landau). Note that in both Eqs. (3.30) and (3.36) the paramagnetic limit L = −kB T N ln(2) is recovered at high temperature, when βJ → 0 and φ → 0 (kB ln(2) being the entropy of an isolated spin one-half). Limiting – for the time being – ourselves to homogeneous fields φ(x), we can set the gradient term to zero. First, we remark that λ appearing in Eq. (3.36) is always positive. On the contrary, τ can change its sign originating two different free-energy landscapes. For τ > 0, the Landau functional LGL has a minimum for φ = 0 only, which clearly corresponds to the magnetically disordered phase. For τ < 0, the Landau functional displays the typical Mexican-hat shape with two minima occurring at some finite φ = ±φ̄ (see Fig. 3.2). These minima are degenerate in the absence of an external field and correspond to the non-trivial solution of the MF equation of state for T < TC . When φ = φ̄, LGL {φ̄} acquires the meaning of Gibbs free energy. Then, from the knowledge of the Landau free-energy functional the MF critical exponent α related to the specific heat can be deduced. For T > TC (τ > 0), we have LGL {φ̄} = −kB T N ln(2) so that 2 2 the specific heat C = −T ∂ LGL /∂T2 = 0. For T < TC (τ < 0), instead, LGL {φ̄} = −kB T N ln(2) + z̄JO(τ ). Therefore the specific heat is finite when TC is approached from lower temperatures. This discontinuity implies that α = 0 within the MF theory. Note that only even powers of φ appear in the functional LGL in Eq. (3.36). This fact is not accidental and reflects the symmetry σn ↔ −σn intrinsic to CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 52 Figure 3.2: Sketch of the Landau free-energy functional: the landscape changes from a Mexican-hat shape to a single-minimum function when passing from the ferromagnetic to the paramagnetic phase. the problem. Landau developed his theory of phase transitions starting from the idea that the effective free energy should be an analytic function of the order parameter (which needs to be identified with φ in our approach), consistently with the requirements of symmetry of the considered problem. In fact, all the critical properties (critical exponents, etc.) derived in the previous section could have been obtained just postulating the form of Eq. (3.36) for the Landau free-energy functional. For the Ising model discussed here, postulating a form for LGL was not necessary since we could carry out the calculation from first principles (i.e., starting from Hamiltonian (3.22)). For problems characterized by less trivial symmetries and, e.g., vectorial order parameters, being able to write the Landau free-energy functional on the basis of symmetry arguments alone is often very useful. Then, performing a saddle-point approximation analogous to Eq. (3.32) one can normally deduce MF critical exponents with little mathematical efforts. Such exponents are named classical, or mean-field, critical exponents and depend only on the symmetry of the problem reflected in the functional LGL . However, a functional built with the same symmetry criteria can be used as a starting point CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 53 for more sophisticated mathematical treatments, like the renormalizationgroup approach. A successful example is the theory of critical phenomena for which Kenneth G. Wilson was awarded the Nobel Prize in 1982. 3.5 Classical spin models If the conditions to have magnetic moments coupled ferromagnetically among them are fulfilled in the ground state, MF theory predicts that magnetic order is retained up to some finite temperature TC . Above this temperature ferromagnetism is lost. This scenario describes a phase transition which is indeed observed in real ferromagnets. However, the Curie temperature given by formula (3.10) is almost always an overestimate compared to the values computed with more sophisticated methods or observed in experiment. This is because fluctuations, neglected in MF theory, tend to have a disordering effect, and therefore suppress the true TC value. In sufficiently low dimensions, this suppression can lead to total loss of magnetic order at any temperature. We will discuss this phenomenon for the simplest collective spin model: the Ising model. From what discussed about the MF critical exponents in the previous section, it should be clear that they only depend on the powers of the order parameter appearing in the Landau free-energy functional of Eq. (3.36). More generally, MF critical exponents depend on the symmetry of the considered problem but not, e.g., on the dimensionality of the lattice. The fact that such exponents are independent of the dimensionality is another artifact of the MF approximation. On the contrary, two facts remain true beyond the MF approximation: i) critical exponents do not depend on some details of the system such as the strength of interactions while ii) they do depend on the symmetry of the considered problem. In the following some of these issues will be clarified in the context of classical spin models. Spins with continuous symmetry The substitution of the quantum-spin operators in Hamiltonian (3.3) by classical spins is somewhat justified in the limit S → ∞, that is when the relative spacing between levels inside each multiplet S(n) becomes smaller and smaller. Moreover, when correlations among spins develop, cooperative effects create a sort of collective large spin which behaves classically. Then, CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE one has: ~ Ŝ(n) → S(n) ≡ S0 (sin θ cos ϕ, sin θ sin ϕ, cos θ) 54 (3.38) where S02 = S (S + 1) (more often S0 = 1 is assumed and the term S02 = S (S + 1) is re-absorbed into the definition of the other constants, J and g, in Eq. (3.3)). The Hamiltonian given in Eq. (3.3) is modified into X X 1 ~ ′ ) + gµB B ~ · S(n S z (n) . (3.39) H=− J S(n) 2 ′ n |n−n |=1 and, accordingly, the partition function becomes Z Z Z ~ Z = dΩ1 dΩ2 . . . dΩN e−βH({S(n)}) , (3.40) with dΩn = sin θn dθn dϕn being the solid-angle element of the spin located at the site n. In some cases, due to the symmetry of the problem, it is more realistic to describe each spin with a two-component vector (in-plane), which can thus be parameterized with just one angle ~ S(n) ≡ S0 (cos ϕ, sin ϕ) . (3.41) Vocabulary of classical models with continuous symmetry: ~ • three-component S(n) ≡ S0 (sin θ cos ϕ, sin θ sin ϕ, cos θ): classical Heisenberg model ~ • two-component S(n) ≡ S0 (cos ϕ, sin ϕ): classical planar or XY model. The Landau free-energy functionals associated with these two types of classical spins are generally different, between them and from the one given in Eq. (3.36) for the Ising model (see below). This means that the number of components of the order parameter determines the Landau free-energy functional and eventually the critical behavior of a system. This is one of the features characterizing a specific universality class. Spins with discrete symmetry: the Ising model When consistent with the symmetry of the problem, two-value classical spins, S z , can be assumed: X X 1 S z (n) . (3.42) S z (n) S z (n′ ) + gµB B H [{S z (n)}] = − J 2 ′ n |n−n |=1 CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 55 The configurations of the classical Hamiltonian (3.42) correspond to the spectrum of eigenvalues of the quantum Hamiltonian 1 H=− J 2 X Ŝ z (n) Ŝ z (n′ ) + gµB B X Ŝ z (n) , (3.43) n |n−n′ |=1 which is, indeed, diagonal on the basis |ϕi i ≡ |σ1 σ2 . . . σN i 1 with Ŝ z |σn i = σn |σn i and σn = ±1 . 2 (3.44) More often, the Ising model is introduced directly assuming the Hamiltonian X X 1 σn H [{σn }] = − J ′ σn σn′ − h′ 2 ′ n (3.45) |n−n |=1 with classical variables σn = ±1. Note that the Hamiltonian (3.45) is equivalent to the one given in Eq. (3.22) in which the nearest-neighbor coupling is expressed in a matrix form. The model (3.45) corresponds to the exact spectrum of eigenvalues relative to the Hamiltonian of coupled quantum spins one-half given in Eq. (3.43) provided that: ( J ′ = 14 J (3.46) h′ = − 21 gµB B . However, the Ising model is applied in many different contexts rather than magnetism, ranging from biophysics to social sciences. Magnetic order and lattice dimensionality Probably one of the most striking failure of MF theory is the prediction of a magnetic phase transition for d=1. In fact, for one-dimensional systems rigorous proofs exist which forbid the occurrence of a magnetically ordered phase at finite temperature in the sole presence of short-range coupling between spins. But let us clarify first what is meant by lattice dimension in this specific context. The dimension d corresponds to the number of directions along which the exchange coupling propagates indefinitely. In practice, this dimension may also be different from the actual dimensionality of the considered solid. If the latter is D, in general one has d≤D. The lattice dimensionality d is another fundamental feature characterizing a specific universality class. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 56 Limiting ourselves – for now – to the Ising model, we start considering the case d=1. d=1 An Ising chain composed of N spins can be represented schematically as in Fig. 3.3. Following an argument due to Landau, we evaluate the variation E 2, S 2 E 1, S 1 - J 2 J DF = F2 - F1 = - kBT ln N 2 DE = E2 - E1 = DS = S2 - S1 = kB ln N Figure 3.3: Sketch representing the free energy difference between a uniform state, with all the spins parallel to each other, and a configuration consisting of two domains with opposite spin alignment (one domain wall). of the free energy associated with the creation of a domain wall in a configuration with all the spins parallel to each other. Creating a domain wall increases the exchange energy by a factor J/2. However, such a domain wall may occupy N different positions in the spin chain, so that this set of configurations has an entropy ∼ kB ln(N ). The free-energy difference between the two configurations sketched in Fig. 3.3 is given by ∆F = J − kB T ln(N ) . 2 Thus, splitting the ground state into domains is ( convenient if ln(N ) > 2kJB T ⇒ N > eJ/2kB T inconvenient if ln(N ) < 2kJB T ⇒ N < eJ/2kB T . (3.47) (3.48) The inequalities written above suggest an estimate of how many consecutive aligned spins can be found at finite temperature in an Ising chain. In particular, when the thermodynamic limit N → ∞ is taken, one immediately realizes that it is always convenient to split the system into groups of parallel spins (magnetic domains), i.e., ferromagnetism is destroyed at any finite temperature. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 57 d=2 A similar, but more rough argument can be given for d=2 as well. In this case we should refer to the possibility of reversing a cluster of spins enclosed in a perimeter of l lattice sites and embedded in a region of spins all pointing in the same direction. We consider for simplicity a square lattice. The total cost in terms of exchange energy is of the order ∼ lJ/2. To estimate the entropy we can think of a self-avoiding random walk: at each step the walker has at most three choices of which way to go, since it has to avoid itself. Thus, we expect the number of closed loops corresponding to the perimeter l to be of the order pl , with p < 3. As a result, the free-energy variation associated with the flip of a cluster delimited by a perimeter l is roughly ∆F = lJ/2 − kB T l ln p. Therefore, for T < J/(2KB ln p) the ordered phase should be stable against the formation of large domains of reversed spins. This argument for the existence of a phase transition in the 2d Ising model was first given, in more precise terms, by Peierls. Rigorous results The Ising model represents a particularly lucky case in which the heuristic arguments given above can be checked by solving the problem analytically. Even if we will not derive these results, it is useful to recall which crucial steps should be followed to prove rigorously whether a model is consistent with a phase with spontaneous magnetization (finite magnetization in zero external field) for T 6= 0 or not. To this end, one has to compute: 1. the partition function z Z = T r e−βH[{S (n)}] (3.49) where the trace is obtained by letting each discrete variable take the two possible values S z (n) = ±1/2 (Z is a sum with 2N terms!) 2. the average magnetic moment m(T, B) = − 1 ∂F 1 1 ∂ ln Z = N ∂B N β ∂B (3.50) 3. the limit m(T, 0) = lim+ m(T, B) B→0 (3.51) and evaluate if there exists a temperature TC below which the limit (3.51) takes a non-zero value. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 58 This procedure can be carried out analytically for d=1 or d=2 only, producing different results: • for d=1, no spontaneous magnetization is possible at finite temperatures • for d=2, a spontaneous magnetization appears for T < TC ≃ 2.27 4kJB . Indeed these exact results show that the MF approximation overlooks some important features as it predicts the occurrence of a phase with spontaneous magnetization independently of the dimension d. d=1: spin chains with uniaxial anisotropy To fix the ideas, we take S = 1/2. As anticipated, in this case # " 1 X z m(T, 0) = lim+ m(T, B) = −gµB lim+ lim hS (n)i = 0 . (3.52) B→0 B→0 N →∞ N n For the 1d case, two-spin correlations can also be computed: r 1 βJ 1 z z hSi Si+r i = tanh = e−r/ξ 4 4 4 with ξ=− 1 ln tanh βJ 4 . (3.53) (3.54) ξ is called correlation length and it is a fundamental quantity in the study of critical phenomena. The correlation length of the 1d Ising model is characterized by an exponential divergence at low temperatures: ξ ∼ eJ/2kB T . (3.55) By comparing the inequalities in Eq. (3.48) with the formula for the correlation length it is clear that ξ gives the order of magnitude of the average size of groups of correlated spins. The existence of such a correlation, marks a major difference between a 1d system of coupled spins and a paramagnet. This is evidenced by the differential susceptibility at B = 0: χ(T, B = 0) = In practical cases ∂m ξ ∼ ∂B T (3.56) CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 59 Figure 3.4: Example of a molecular spin chain with uniaxial anisotropy which – in a proper range of temperature – behaves as a 1d Ising chian. C. Coulon et al., Physical Review B 69 p.132408 (2004). At low temperature, χ T saturates because of the presence of non-magnetic impurities and 3d interactions with the other spin chains in the crystal. • the plot of 1/χ versus T highlights deviations from the paramagnetic behavior (Curie-Weiss law) • the plot of ln [χ T ] versus 1/T highlights an 1d Ising-like behavior (when experimental points at low temperature lie on a line). The behavior of two-spin correlations for the 1d Ising model is plotted in Fig. 3.5. d=2: ultrathin magnetic films with uniaxial anisotropy The 2d Ising model was solved for the first time by Lars Onsager in 1944. Such a solution is a “veritable mathematical tour de force” (M. Le Bellac). To our purposes, it is enough to recall the formula which gives the spontaneous CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 60 magnetization7 for T < TC : # " 1 X z hS (n)i m(T, 0) = lim+ m(T, B) = −gµB lim+ lim B→0 B→0 N →∞ N n 81 βJ −4 = gµB 1 − sinh 2 and the definition of TC itself J =1 ⇒ sinh 2kB TC TC = J J 2 √ ≃ 2.27 . 4kB 1 + 2 4kB (3.57) (3.58) As anticipated at the beginning of this Chapter, MF theory typically overestimates the transition temperature. The specific value reported in Eq. (3.58) has to be compared with the MF Curie temperature given by Eq. (3.10) for a spin one-half and for z̄ = 4 (square lattice): TCMF = J/kB . Expanding the spontaneous magnetization m(0) close to TC yields 1 m(T, 0) ∼ (TC − T ) 8 . (3.59) Thus, for the 2d Ising model β = 1/8 at odds with the MF value β MF = 1/2. 1 4 MFA Ising Figure 3.5: Two-spin correlation function for the ferromagnetic Ising chain (adapted from Quantum and Statistical Field Theory, M. Le Bellac). 7 We identify the magnetization with the average magnetic moment per magnetic atom (or molecule), while in the SI (Système international d’unités) it is the average mangetic moment per unit volume. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 3.6 61 Correlation functions Figure 3.6: Group of people in the “paramagnetic” (a) and in the “ferromagnetic” (b) phase. (Taken from L. J. de Jongh and A. R. Miedema Adv. Phys. 50 p. 947-1170 (2001)). In the following sections we will try to render more quantitative the effect of neglecting fluctuations in the mean-field approximation. A pictorial idea of what happens when a system passes from the paramagnetic phase to the ferromagnetic phase is sketched in Fig. 3.6. In the picture on the left-hand side, people walk in the street without conditioning each other, like magnetic moments do in the paramagnetic phase. In the picture on the righthand side, instead, a strong feed-back mechanism is present so that if one of the individuals is attracted, e.g., by a window all the others are conditioned and end up staring at the same thing. This situation can be assimilated to spontaneous symmetry breaking occurring in a magnet below TC . However, everyday experience offers also intermediate degrees of correlation in which such a feed-back mechanism involves a limited number of people. Think, for instance, of a road artist playing music in a subway station: the majority of people will be more concerned of not missing the train rather than listening at his/her music. Nevertheless, there will still be a sort of shortrange correlation among the people whose train is not departing soon and whose attention is captured by the musician. This last situation resembles short-range correlations in magnetic systems. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 62 Figure 3.7: Magnetic specific heats of the S=1/2 Ising model for d=1,2,3 and within the MF approximation. d=2 is given by the Onsager solution for the square lattice while d=3 is obtained by high-temperature series expansion for a simple cubic lattice. All temperatures are expressed in units of the corresponding MF transition temperature, θ, with z̄ = 2, 4, 6 for the d=1,2,3 respectively. (Taken from L. J. de Jongh and A. R. Miedema Adv. Phys. 50 p. 947-1170 (2001)). Specific heat tail As already noticed, the mean-field theory predicts a finite discontinuity in the specific heat at the critical temperature. On the contrary, in experimental systems showing a magnetic phase transition the specific heat diverges at TC . This behavior is reproduced by more sophisticated models. In Fig. 3.7, the behavior of the specific heat is plotted for the Ising model. There the MF prediction is compared with exact results for d=1,2 and high-temperature series expansion for d=3 (no exact results solution is available in this case yet). First, we notice that the true TC is lower than the MF value in each case. In particular, TC shifts at lower temperature as the lattice dimensionality is reduced, down to TC = 0 for the Ising spin chain (d=1). This is an indication that the effect of thermal fluctuations becomes progressively more dramatic as the lattice dimensionality is reduced. Second, apart from the MF calculation, all the models show a hightemperature tail in the specific heat. Also this feature is more enhanced CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 63 the lower the lattice dimensionality is. Such tail is due to the presence of short-range correlations above TC that are not taken into account in the MF theory. Both calculations for d=2 and d=3 show the expected singularity at TC , while the specific heat does not diverge at any temperature for d=1, not even at T = 0. The most natural way to characterize short-range correlations is by studying the behavior of the correlation function. The fluctuation-response theorem Here we show how two-spin correlations are related to the susceptibility. From the definition of the magnetization itself, it follows that 1 ∂F 1 1 ∂ ln Z = N ∂B N β ∂B (" ) # X 1 X z 1 1 1 Tr −βgµB = S z (n) e−βH = −gµB hS (n)i NβZ N n n m(T, B) = − (3.60) where the trace is taken over all the possible values of the N variables S z (n). With the definition of Eq. (3.60), the magnetization equals the average magnetic moment. This value can be converted to any other unit to compare with experiments (Bohr magneton per atom, emu/mol, A/m, etc.). The susceptibility is the derivative of the magnetization with respect to the applied field X ∂m(T, B) 1 1 S z (n) · S z (n′ ) e−βH χ(T, B) = = T r β(gµB )2 ∂B NZ nn′ # ))2 ( (" X 1 1 S z (n) e−βH − β Tr gµB N Z n !2 2 X X β(gµB ) S z (n) i − h S z (n)i2 , = h N n n (3.61) 2 P P z S (n) where we have used the fact that = nn′ S z (n) S z (n′ ). n Defining the correlation function as Gnn′ = hS z (n)S z (n′ )i − hS z (n)i hS z (n′ )i , (3.62) 64 CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE a relation between Gnn′ and the magnetic susceptibility can be deduced χ(T, B) = β(gµB )2 X Gnn′ , N nn′ (3.63) also known as fluctuation-response theorem. Remarkably, according to the definition of the correlation function given in Figure 3.8: Qualitative behavior of the correlation function for T > TC (left) and T < TC (right) assuming an exponential decay Gij ∼ exp (−rij /ξ) (taken from Quantum and Statistical Field Theory, M. Le Bellac). Eq. (3.62), when hS z (n′ )i 6= 0 (i.e. for T < TC or for B 6= 0) the assertion that “two spins are uncorrelated” means hS z (n) S z (n′ )i = hS z (n)i hS z (n′ )i ∝ m2 . In other words, the correlation function Gnn′ only measures the degree of short-range correlation. The term relating to long-range order as been eliminated by subtracting hS z (n)i hS z (n′ )i (see Eq. (3.62)). Susceptibility for the different magnetic phases at B = 0 2d-3d systems T ≃ TC χ(T, 0) χ∼ Γ± |T −TC |γ 1d-system χ∼ ξ T Paramagnet χ= C T From the previous table and from the theorem (3.63), it is clear that the differential susceptibility is strictly related to the degree of correlation of fluctuations at the considered temperature: χ(T, B) = β(gµB )2 X Gnn′ . N nn′ (3.64) CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 65 Note that even for the ordered phase (occurring for d≥2) the susceptibility diverges only at T = TC and tends to zero as T → 0 (in a perfectly ordered system fluctuations are not allowed). Fourier transform of correlations The Fourier transform (G̃) of the correlation function can be accessed experimentally, for instance through neutron scattering. At T = TC , G̃(q) is found to behave as 1 G̃(q) ∼ 2−η . (3.65) q A simple argument based on dimensional analysis suggests that Z 1 at T = TC . G(r) ≃ G̃(q)dd q ⇒ G(r) ∼ d−2+η r (3.66) From Eq. (3.66) we learn that at the critical point spatial correlations decay with power law. In other words, the system is not characterized by any typical length scale and possesses the property of being self-similar at different spatial scales8 . Such a scale invariance is the key ingredient of the theory of critical phenomena and it is, indeed, strictly related to the divergence of the correlation length for T → TC . Notice that the fluctuation-response theorem may also be written as χ(T, B) ∼ G̃(0) . (3.67) Since in practice η < 2 always, Eq. (3.65) implies that G̃(q = 0) diverges for T → TC and so does the susceptibility. In the next section we will see that within the Landau theory of critical phenomena the Fourier transfrom of the correlation function is given by G̃(q) ∼ q2 1 . + ξ −2 (3.68) The inverse Fourier transform of Eq. (3.68) gives the asymptotic behavior G(r) ∼ e−r/ξ r(d−1)/2 for T far away from TC . (3.69) In this case the decay of the correlation function is characterized by the typical length scale ξ, the correlation length. The different behavior of the 8 See the paper C. H. Back, et al., Nature 378, p. 597. The authors report on the experimental check of the scaling hypothesis on a Fe film which behaves as the 2d Ising model. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 66 correlation function above and below TC and both in real and reciprocal space is sketched in Figs. 3.8 and 3.9. It is worth remarking that we are considering short-range correlations as the unique source of broadening of the quasi-elastic peaks. Of course, experimentally this is not true and the correlation length can be deduced only after removing the other sources of broadening such as experimental resolution, etc. Figure 3.9: Qualitative behavior of the correlation function in the real space (b) and in the Fourier space (a). (Taken from Quantum and Statistical Field Theory, M. Le Bellac). 3.7 Landau theory of correlations Before deducing correlations within the Laundau theory, let us recall some basic concepts that have been discussed some sections ago. First, we showed that the calculation of the partition function of the Ising model (for any dimension d) can be recasted into the following problem: Z Z (B, T ) = D[φ] e−βL[{φ}] (3.70) from which, in principle (but not always in practice!), the whole thermodynamics can be deduced. Two possible independent approximations can be made to tackle the problem stated by Eq. (3.70): 1. the saddle-point approximation, which consists in evaluating the partition function only in the minimum of the functional L [φ] (see Eq. (3.32)); CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 67 2. a Taylor expansion of L [φ] itself for small auxiliary fields φ, which holds in the critical region (i.e., for T ≃ TC ) and gives the Ginzburg-Landau functional LGL [φ]. The condition 1. requires to set to zero the functional derivative of L [φ] with respect to φ; this leads to a self-consistent equation for the average magnetic moment that turns out to be equivalent to the MF equation of state, for every T . As a consequence, by making both approximations 1. and 2. the MF critical behavior can be studied. More concretely, one can start from the Ginzburg-Landau functional in Eq. (3.36) (which we recall here for convenience) Z 1 b 2 λ 4 d 2 LGL [φ] = (3.71) J (∇φ) + φ + φ d x − β −1 N ln(2) , 2 2 4 set the gradient term to zero (because we seek for spatially homogeneous solutions) and minimize the integrand with respect to φ. This leads to the equation φ̄ bφ̄ + λφ̄3 = 0 (3.72) whose solutions are given by ( φ̄ = 0 φ̄2 = − λb for τ > 0 for τ < 0 , (3.73) with b ≃ z̄Jτ and λ ≃ 16kB TC /3. To evaluate correlation functions we need to go slightly beyond the crude saddle-point approximation (equivalent to the MFA). In practice, we allow the field φ to deviate slightly from the MF solution obtained for τ < 0 (the treatment for τ > 0 is analogous): φ = φ̄ + δφ , (3.74) δφ being a small, random field. The integrand (fGL ) of the Ginzburg-Landau functional in Eq. (3.71) takes the form b λ 1 fGL [φ̄ + δφ] = J (∇δφ)2 + (φ̄ + δφ)2 + (φ̄ + δφ)4 2 2 4 1 b 6λ 2 2 = J (∇δφ) + + φ̄ δφ2 + fGL [φ̄] + (. . . ) 2 2 4 1 = J (∇δφ)2 − b δφ2 + fGL [φ̄] + (. . . ) 2 (3.75) CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 68 where (. . . ) stands for constants, O(δφ4 ) and odds terms in δφ which vanish after spatial or thermal averaging. The first two terms in the last line of Eq. (3.75) are associated with fluctuations around the MF solution φ̄. Thus, at this level of approximation, the partition function reads: Z −βLGL [φ̄] Z (B, T ) = e + D[δφ] e−βLfl [δφ] (3.76) where the first term on the right-hand side corresponds to the saddle-point approximation while Z 1 2 2 (3.77) Lfl [δφ] = J (∇δφ) − b (δφ) dd x . 2 is the functional associated with the fluctuation field δφ(x). The underlying strategy of the mathematical passages described above aims at considering corrections to the saddle-point approximation which are described by a sort of quadratic Hamiltonian with respect to the fluctuation field. In fact, the functional (3.77) is formally equivalent to the potential energy of set of coupled harmonic oscillators, described in the continuum formalism (remember that b < 0 for τ < 0): Thermal averages of the fluctuation field δφ(x) can be computed similarly to average displacements in a system of harmonic oscillators. The gradient term in Lfl [δφ] effectively couples the fluctuation fields δφ(x) defined at different points in space, at different locations x. However, the functional (3.77) can be decoupled (diagonalized) passing to the Fourier space: Z 1 1 2 2 d ˜ Jq − 2b |δφ(q)| Lfl [δφ] = d q. (3.78) d (2π) 2 If we forget the parametric9 dependence on temperature of b, Eq. (3.78) has the form of a quadratic Hamiltonian with respect to the independent degrees ˜ Now, equipartition theorem can be applied to get of freedom δφ(q). 1 1 2 ˜ Jq 2 − 2b h|δφ(q)| ifl = kB T 2 2 ⇒ 2 ˜ h|δφ(q)| ifl = kB T , Jq 2 − 2b (3.79) where the subscript reminds that h. . . ifl represents an average over the fluctuation field δφ. Before proceeding, it is useful to establish a contact between the correlation 9 This implicit temperature dependence, which may look strange at first sight, comes from having expanded LGL [φ] around its minimum: it is reasonable that coefficients of the expansion contain information about the saddle point φ̄, corresponding to the minimum. 69 CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE function defined in Eq. (3.62) and the fluctuation field δφ. In the formalism of the present section we shall write the correlation function as G(x, x′ ) = hσ(x)σ(x′ )i − hσ(x)i hσ(x′ )i = h φ̄ + δφ(x) φ̄ + δφ(x′ ) ifl − φ̄2 = hδφ(x)δφ(x′ )ifl . (3.80) By comparing Eq. (3.79) with Eq. (3.80) we obtain the following result G̃(q) = kB T . Jq 2 − 2b (3.81) Note that this result corresponds to the case in which τ < 0 and hence b < 0. For τ > 0, the same calculation would yield G̃(q) = kB T . Jq 2 + b (3.82) Summarizing, within the Landau theory, the correlation function takes the Ornstein-Zernicke form: G̃(q) = 1 kB T , 2 J q + ξ −2 (3.83) with ξ ∼ |τ |−1/2 . As discussed at the end of Section 3.6, ξ has the meaning of correlation length. The corresponding classical critical exponent ξ ∼ |τ |−ν is ν cl = 1/2. CHAPTER 3. MAGNETIC ORDER AT FINITE TEMPERATURE 70 Literature • A. Aharoni, Introduction to the Theory of Ferromagnetism Oxford University Press (Chapter IV: Magnetization vs. Temperature) • J. Cardy, Scaling and Renormalization in Statistical Physics Cambridge University Press (Chapter VI: Low dimensional systems) • M. Le Bellac, Quantum and Statistical Field Theory Oxford University Press (Chapter I: Introduction to critical phenomena. Chapter II: Landau theory) • G. Morandi, F. Napoli, E. Ercolessi, Statistical Mechanics World Scientific Singapore (Chapter III: Spin Hamiltonians I: Classical) • L. D. Landau and E. M. Lifshitz, Statistical Physics Oxford Pergamon Press • D. C. Mattis, The Theory of Magnetism II Springer Series in Solid-State Science (Advanced) Chapter 4 Magnetic domains and domain walls 4.1 Magnetic anisotropy Let us go back to consider a single magnetic center. For the atom embedded in a spherically symmetric environment Hund’s rules generally succeed in predicting the observed magnetic moment. When this scenario holds, the spin “points” with the same probability along any spatial direction in the absence of an external magnetic field. Due to the reduced symmetry of the surrounding, the situation is generally different for an atom in a solid. As already seen in Part I, a first consequence is that magnetic moments are generally smaller in solids with respect to those predicted by Hund’s rules DS2 J 0 p 2 p Figure 4.1: Schematic representation of the energy landscape associated with a uniaxial-anisotropy term as a function of the polar coordinate θ of a classical spin. 71 CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 72 (remember the quenching of the angular momentum). Another implication is that magnetic moments (effective spins) prefer to lie along some crystallographic directions. This tendency is taken into account by introducing a magnetic anisotropy energy which is function of the effective-spin projections along the crystallographic axes. The simplest anisotropy term that can be considered in single-spin Hamiltonian is HA = −D(Ŝ z )2 . (4.1) Notice that the symmetry Ŝ z → −Ŝ z is not broken by such an anisotropy term. Additional terms which combine higher powers of the single-spin operators may arise according to the symmetry of the lattice in which the magnetic atom is embedded (or according to the symmetry of the substrate for adatoms). For example, in the case of a crystal lattice with cubic symmetry the first non-zero anisotropy term is a fourth-order combination of the spin operators; thus in this case the term in Eq. (4.1) vanishes. The physical mechanism which couples the spin degrees of freedom with the spatial degrees of freedom is the spin-orbit interaction. Magnetic anisotropy away from the bulk Figure 4.2: Ab-initio calculation of the magnetic anisotropy energy, DS 2 , and the magnetic moment per Co atom on Pt(111). Values in brackets have been computed with a different computational method. Remember that 1 meV ≃ 11.6 K (C. R. Physique 6 p. 75 (2005)). As stated above, a crucial ingredient for magnetic anisotropy to arise is the reduced symmetry of the surrounding, “seen” by a magnetic atom in a solid, with respect to the spherical symmetry (Hund’s rules). It is not surprising that a further increase of the anisotropy is observed when the symmetry of the environment is further reduced. This happens, e.g., when magnetic atoms are arranged in clusters (0d) or in 1d and 2d nanostructures. In Fig 4.2 CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 73 Figure 4.3: Experimental results: magnetic anisotropy energy DS 2 (right, (b)), and magnetic moment per Co atom on Pt(111) (left, (a)) C. R. Physique 6 p. 75 (2005). theoretical predictions (from ab initio calculations) for different aggregates of Co atoms on a Pt(111) surface are reported. Notice that when passing from a single atom to five atoms the value of the magnetic anisotropy per atom already decreases of one order of magnitude. The magnetic moment per atom also decreases with increasing the number of atoms. This fact is instead associated with the degree of hybridization of magnetic orbitals, which becomes more and more significant when approaching the bulk limit. The theoretical predictions of Fig 4.2 are in qualitative agreement with the experimental results reported in Fig 4.3. Indeed, the fact that the magnetic anisotropy increases up to a factor 103 when approaching the atomic scale is a good trend in view of magneto-storage applications. Classical approximation If the operator in Eq. (4.1) is substituted by a classical spin the anisotropy energy reads HA = −DS 2 cos2 θ . (4.2) Depending on the sign of D, the energy (4.2) has either one minimum for θ = π/2 (D < 0) or two minima for θ = 0, π (D > 0), which describes the two physical situations D<0 D>0 easy plane easy axis / uniaxial . (4.3) CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 74 For the easy-axis case, D > 0 (see Fig. 4.1), only few configurations around θ = 0 or θ = π will be statistically relevant for kB T ≪ DS 2 . In other words, the spin will spend about half of the time visiting configurations for which θ ≃ 0 and half of the time around θ ≃ π. For kB T ≪ DS 2 , the escape rate 2 from each one of the two wells is ν = ν0 e−DS /kB T , so that the relaxation time diverges exponentially as: DS 2 τA ∼ e k B T . (4.4) This time represents the average time it takes the system to jump from one minimum of Fig. 4.1 to the other. 4.2 Domain walls in the classical Heisenberg model In chapter 2 we gave a justification for the use of the Heisenberg exchange interaction which is isotropic. If we add to the classical Heisenberg Hamiltonian (Eq. (3.3)) an anisotropy term like the one in Eq. (4.1) we get 1 H=− J 2 X |n−n′ |=1 Ŝ(n) · Ŝ(n′ ) + gµB X n ~ · Ŝ(n) − D B X (Ŝ z (n))2 . (4.5) n When D becomes large with respect to |J|, the model described by Hamiltonian (4.5) can be replaced with the two following models D → +∞ |J| D → −∞ |J| Ising model (4.6) XY / planar model . Domain walls: discrete lattice In the following, we consider the Hamiltonian (4.5) taking: • D > 0, uniaxial anisotropy • J > 0, ferromagnetic exchange interaction (parallel alignment of nearest-neighboring spins is favored). CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 75 The study of the model described by the Hamiltonian (4.5) can be significantly simplified by substituting the quantum spin operators Ŝ(n) with clas~ sical vectors S(n). This simplification is justified by thinking that a sort of “collective” spin can be associated with a group of spins coupled ferromagnetically (J > 0). Such groups can emerge in a magnetic system due to either long-range order or short-range correlations. In the latter case, the correlation length needs to be large enough. In both situations the collective spin can be so large that its quantum-mechanical character becomes negligible1 . In the appropriate temperature regimes, the results are then the same if the classical approximation is assumed at the level of single effective spins. For many theoretical and applicative aspects of magnetism, domain walls, i.e. the boundaries between regions with opposite magnetization, play a crucial role. In particular, their structure and the energy associated with 1 Remarkably, in this sense the classical-spin approximation is more justified for low temperatures than for high ones. In the paramagnetic limit (kB T /J >> 1) one has to recover a behavior described by the Brillouin function, in which the quantum nature of each spin is relevant (S 2 → S(S + 1)). 2.0 1.5 Broad DW 1.0 Sharp DW 0.5 0.01 0.1 1 10 Figure 4.4: One-wall energy in J units vs D/J: minimum energy solution of the non-linear equation (4.10) computed numerically (solid line); continuum limit solution (dashed line). Inset: spin profile vs lattice distance: sharp wall (low-right) and broad wall for D/J = 10−2 (up-left). CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 76 the creation of a domain wall in a uniformly magnetized configuration are relevant. These features can be evaluated letting the spin direction vary only along one spatial direction. This effectively reduces the problem to a mono-dimensional one: Nx h i X ~i · S ~i+1 + D (S z )2 , HH = − (4.7) JS i i=1 ~i are classical spins and the constants J and D have to be thought of where S per unit length or per unit surface if the dimensionality of the original lattice was d=2 or d=3 respectively. With the Hamiltonian (4.7), the domain wall can be larger than one lattice spacing. In fact, spreading the wall over more than one lattice spacing reduces the global exchange-energy cost. On the other hand, the anisotropy term would favor configurations with as less spins misaligned to the easy axis, z, as possible. The domain-wall profile results from the competition between these two energies (two opposite limits are reported in the insets of Fig. 4.4). The lowest-energy deviations from the uniform state can be parameterized through the angle that each spin forms with the z axis, θ, as EH = Nx X i=1 J − J cos (θi+1 − θi ) + D sin2 θi . (4.8) The energy cost for creating a domain-wall in a uniformly magnetized configuration is given by the spin profile which fulfills the boundary conditions ( θ1 =π (4.9) θN x = 0 and minimizes the energy (4.8) with respect to θi : ∂EH D (4.10) = sin (θi − θi−1 ) − sin (θi+1 − θi ) + sin (2θi ) = 0 . ∂θi J Eq. (4.10) can be solved numerically and the solution provides the spin profile with respect to which the energy (4.8) is stationary. The true lowest-energy profile can be obtained comparing different solutions, among which the sharpwall profile (see lower-right inset of Fig. 4.4): ( θi = π for 1 ≤ i < N2x (4.11) θi = 0 for N2x ≤ i ≤ Nx which is also a solution of (4.10). In Fig. 4.4 the resulting energy (solid line) is compared with that obtained from a continuum limit calculation (dashed line) – that we are going to present in the next paragraph – as a function of the ratio D/J. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 4.3 77 Continuum formalism Referring to the classical version of Hamiltonian (3.3), we rewrite in a different way the exchange-interaction term: XX X 1 ~ ′ ) = −J ~ + eµ ) ~ ~ (4.12) Hexch = − J · S(n · S(n S(n) S(n) 2 ′ n µ |n−n |=1 where µ = x, y, z (spatial directions) and eµ is the unit vector along µ. Notice that 2 2 ~ ~ 2 ~ ~ ~ ~ + eµ ) . (4.13) · S(n S(n) − S(n + eµ ) = S(n) + S(n + eµ ) − 2S(n) ~ With the hypothesis that the direction along which each classical spin S(n) is pointing varies smoothly from one lattice site to the other (index n), one ~ can describe S(n) as a vector field which is a smooth function of a continuum spatial variable r = an, a being the lattice spacing. This approximation is justified • In the classical isotropic Heisenberg chain (D = 0) at low temperatures. In fact, the lowest lying excitations – which actually destroy ferromagnetism for d≤2 – are spin-waves with very long wavelength (in the following we will show that for small wave vectors q → 0 the spectrum of fluctuations is gapless). • When the walls separating domains with opposite spin directions are broad enough. Further on, we will render this statement quantitative. In the presence of ferromagnetic (J > 0) exchange interaction and uniaxial anisotropy (D > 0), such a requirement is fulfilled for J ≫ D. Thus one has, ~ ~ ~ + eµ ) − S(n) ~ + aeµ ) − S(r) ~ S(n ≃ S(r ≃ a∂µ S(r) (4.14) where in the first passage we have taken the continuum limit and in the second one we have performed a Taylor expansion. Combining Eq. (4.13) with Eq. (4.14), the exchange interaction between the spin located in r and half of its nearest neighbors is obtained X ~ + eµ ) ~ · S(n −J S(n) µ 2 X 1 2 X ~ 2 ~ ≃ Ja ∂µ S(r) − J S(r) 2 µ µ X z̄ X ν 1 (∂µ S ν (r) · ∂µ S ν (r)) − J (S (r) · S ν (r)) = Ja2 2 2 µ,ν ν (4.15) CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 78 with z̄ number of nearest neighbors and ν = x, y, z label of the spin components. Normally, the first term is written as ~ 2 X ∇ S(r) (∂µ S ν (r) · ∂µ S ν (r)) = = µ,ν 2 = (∂x S (r)) + (∂y S x (r))2 + (∂z S x (r))2 x (4.16) + (∂x S y (r))2 + (∂y S y (r))2 + (∂z S y (r))2 + (∂x S z (r))2 + (∂y S z (r))2 + (∂z S z (r))2 . Taking the usual continuum limit for the sum Z X 1 · · · ≃ d . . . dd x , a (4.17) the classical and continuum version of Hamiltonian Eq. (4.5) is finally obtained Z Z z̄ 1 1 2−d ~ 2 d ~ 2 d d x − J S(r) ∇ S(r) H = Ja d x 2 2 ad Z Z (4.18) 1 1 2 d d z ~ ~ − D d |S (r)| d x + µB g d B · S(r)d x . a a ~ Within the continuum model the field S(r) can be simplified as a twocomponent vector field or as a scalar field (see the two limits (4.6)). However, some additional constraints or effective energy terms are normally introduced ~ 2 in place of the stringent constraint on the spin modulus S(r) = S 2 . Of course, the latter condition is automatically fulfilled if each spin is parameterized with polar coordinates x S (r) = S sin(θ(r)) cos(ϕ(r)) (4.19) S y (r) = S sin(θ(r)) sin(ϕ(r)) S z (r) = S cos(θ(r)) . Broad domain walls: continuum limit To the aim of computing the domain-wall energy in the continuum limit, we let the polar angles (4.19) be a function of one spatial variable only, say x. For B = 0, the Hamiltonian (4.18) can then be written as 2 # Z " 2 dθ dϕ 1 + sin2 (θ(x)) dx H = JNy Nz aS 2 2 dx dx (4.20) Z 2 21 cos (θ(x))dx + const − DNy Nz S a CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 79 where we have implicitly assumed the integration domain to be a parallelepiped Nx Ny Nz a3 . The functional Eq. (4.20) can be minimized with respect to the functions θ(x) and ϕ(x). The corresponding Euler-Lagrange equation is ( dϕ 2 dθ Ja sin2 (θ(x)) ddxϕ2 + 2Ja sin(θ(x)) cos(θ(x)) dx =0 dx dϕ 2 d2 θ D Ja dx2 − Ja sin(θ(x)) cos(θ(x)) dx − 2 a sin(θ(x)) cos(θ(x)) = 0 (4.21) The solution to Eq. (4.21) with boundary conditions ( limx→−∞ θ(x) = π (4.22) limx→+∞ θ(x) = 0 and corresponding to the minimum energy is ( cos(θ(x)) = tanh xδ ϕ(x) = const. (4.23) q J with δ = a 2D . Such a solution was proposed by Landau and Lifshitz in 1935. √ √ The energy density associated with the spin profile (4.23) is Ew = 2 2S 2 DJ (per unit length for d=2 and per unit surface for d=3). In Fig. 4.4 the domain-wall energy obtained numerically for the discrete-lattice calculation (solid line) is compared with that obtained in the continuum limit (dashed line) as a function of the ratio D/J. The agreement is already good for ratios D/J < 0.3. In the opposite limit, the discrete lattice calculation recovers the domain-wall energy of the Ising model Ew = 2JS 2 (sharp domain wall defined by Eqs. (4.11)). In those relevant limits one has ( J ≪D⇒δ=a and Ew = 2S 2 J q √ √ (4.24) J J ≫ D ⇒ δ = a 2D > 1 and Ew = 2 2S 2 DJ . For J ≪ D, the wall-energy cost equals the Ising case and follows from having δ = a. Concerning J ≫ D, Ew is one-soliton energy 2 . As one can appreciate in Fig. 4.4, the two regimes are very well recovered and the transition region, where none of the two limits (4.24) is expected to hold, is surprisingly narrow. 2 Sometimes the domain-wall q q width is defined with some numerical factors of difference J with respect to δ: π 2D or 2J D for instance. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 80 The crossover between the sharp-wall (δ = a) and the broad-wall (δ > a) regime3 occurs at D/J = 2/3. In summary, the conditions (4.24) provide a criterion for the simplification of the classical Heisenberg model in terms of ( Ising model D/J ≥ 2/3 (4.25) continuum limit D/J ≤ 0.3 . Typically, metallic nanowires of technological relevance fall in the broad-wall regime. In fact, materials like Co, Ni, Fe or Permalloy are characterized by D ≃ 1 − 10 K (∼ 0.1 − 1 meV) and J ≃ 100 − 500 K (∼ 10 − 50 meV) corresponding to a domain-wall width of the order 10 − 100 nm. 4.4 Beyond the Mean-Field Approximation The argument used in chapter 3 to state that the Ising model does not show a magnetically ordered phase at finite temperature for d=1 holds also for the Heisenberg chain with uniaxial anisotropy, provided that the appropriate domain-wall energy is considered (remember Fig. 3.3). Similarly, one concludes that the same model can sustain ferromagnetism at finite temperatures in d=2. Different arguments are, instead, needed to provide a conclusive statement about the existence or not of magnetic order at finite temperature in systems with continuous symmetry. For the last ones, it will turn out that linear excitations are able to destroy ferromagnetism both in d=1 and d=2. For the Heisenberg model, these linear excitations can be identified with spin waves. Spin waves are usually introduced as linear solutions to the Landau-Lifshitz equation of motion. However, the capability of these type of excitations to destroy magnetic ordering for d≤2 in systems with continuous symmetry can be evidenced without the need of introducing dynamics. We prefer to follow this way because it is straightforward to apply a unique argument to both the XY and the Heisenberg model. 3 The crossover ratio D/J = 2/3 can be obtained analytically by analyzing the stability of the sharp-wall profile, Eqs. (4.11), against small deviations between successive angles θi (B. Barbara, Journal de Physique 34, p. 139 (1973)). CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 81 Linear excitations in models with continuous symmetry We rewrite for convenience the classical spin Hamiltonian introduced in chapter 3, Eq. (3.39): 1 H=− J 2 X |n−n′ |=1 ~ ′ ) + gµB B ~ · S(n S(n) X S z (n) . (4.26) n The minimal energy is obtained by letting all the magnetic moments be aligned along the direction of the applied field (spins along negative z direction). We consider how the energy increases due to small deviations from this configuration. Our goal is to simplify the original problem by means of an effective Hamiltonian that is formally equivalent to the one describing a system of coupled harmonic oscillators. To this end, we may write s X 1 X α z S (n) = − 1 − (S α (n))2 ≃ −1 + (S (n))2 2 (4.27) α=x,y α=x,y with the hypothesis (S α (n))2 ≪ (S z (n))2 . Note that for the planar (or XY) model α takes just one value and two values for the Heisenberg model. From now on, we will not specify the number of extra components but z represented by the index α; while doing so, we are going to derive results that apply to both models. The approximation in Eq. (4.27) reflects in the Hamiltonian as follows: " # " # X X ′ 1X α 1 1 (S α (n′ ))2 1− H≃− J (S (n))2 × 1 − 2 2 2 α α′ |n−n′ |=1 # " X X X X 1 1 1− S α (n)S α (n′ ) − gµB B (S α (n))2 − J 2 2 α n |n−n′ |=1 α X X X X 1 1 1 ′ (S α (n))2 + (S α (n′ ))2 = − z̄N J − gµB BN + z̄J 2 2 2 n α n ′ α′ 1 − J 2 X X |n−n′ |=1 α ≃Eg.s. + Hh.o. XX 1 S α (n)S α (n′ ) + gµB B (S α (n))2 + O (S α )4 2 n α (4.28) CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 82 P P P P where, by nothing that the double summations α′ are n α and n′ actually the same, we have defined X X 1 XX α 1 Hh.o. = z̄J S α (n)S α (n′ ) (S (n))2 − J 2 2 α n |n−n′ |=1 α (4.29) XX 1 + gµB B (S α (n))2 2 n α and the constant ground-state energy 1 Eg.s. = − z̄N J − gµB BN . (4.30) 2 The Hamiltonian Hh.o. , written in Eq. (4.29), is equivalent to the Hamiltonian of N coupled harmonic oscillators which can be decoupled by the usual Fourier transform in the discrete space: ( P S α (n) = √1N q S̃ α (q) e−iq·n (4.31) P S̃ α (q) = √1N n S α (n) eiq·n with orthogonality relation X ′ ei(q−q )·n = N δq,q′ . (4.32) n For simplicity we assume unitary lattice constant. It is convenient to evaluate the two relevant summations appearing in the Hamiltonian of Eq. (4.29) separately. The first summation reads X X 1 XX α ′ S̃ (q)S̃ α (q ′ ) e−i(q+q )·n = (S α (n))2 = |S̃ α (q)|2 . (4.33) N n q,q′ n q This is nothing but the Parseval’s formula for the discrete-lattice Fourier transform. For what concerns the second summation on the right-hand side of Eq. (4.29), we first rewrite it as X XX S α (n)S α (n′ ) = S α (n)S α (n + δ) (4.34) n |n−n′ |=1 δ where δ is a vector connecting the site n with its nearest neighbors. For simplicity, we will consider just a linear, square and simple-cubic lattice for d=1, 2 and 3, respectively. Passing to the Fourier space one finds XX XX 1 X ′ ′ S̃ α (q)S̃ α (q ′ ) e−i(q+q )·n e−iq ·δ S α (n)S α (n + δ) = N q,q′ n n δ δ (4.35) X XX X α 2 −iq·δ α 2 = |S̃ (q)| e = |S̃ (q)| 2 cos(q · δ) ; δ q q {δ>0} CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 83 the notation {δ > 0} means that the summation extends over half of the nearest neighbors of the spin located at site n: it consists of z̄/2 terms. Eqs. (4.33) and (4.35) enable us to decouple the elastic Hamiltonian given in Eq. (4.29), which then reads X X X 1 z̄ − Hh.o. = J 2 cos(q · δ) |S̃ α (q)|2 2 q α 1 + gµB B 2 = {δ>0} XX q α |S̃ α (q)|2 (4.36) 1 XX Γ(q)|S̃ α (q)|2 , 2 α q with Γ(q) = J[z̄ − X {δ>0} 2 cos(q · δ)] + gµB B . (4.37) Figure 4.5: Sketch of a spin-wave excitation in a Heisenberg ferromagnetic spin chain. Indeed, for the Heisenberg model, the linear excitations associated with the quadratic Hamiltonian in Eq. (4.29) are spin waves with dispersion relation ~ω(q) = Γ(q). Spin waves are collective excitations analogous to phonons. Similarly to phonons, spin waves are also quantized and the specific dependence of Γ(q) on the wave vector (especially for q ≃ 0) determines the behavior of the magnetization at low temperature (in the absence of anisotropy). The dispersion curve Γ(q) can be measured, e.g., by inelastic neutron scattering. Coming back to our goal, we proceed by evaluating the average of fluctuations, namely those terms in Eq. (4.27) that we have assumed to be CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 84 small for linearizing the Hamiltonian (4.26). The approximated Hamiltonian, Eq. (4.36), consists of N independent quadratic degrees of freedom so that equipartition theorem applies: 1 1 Γ(q) h|S̃ α (q)|2 ith = kB T 2 2 ⇒ h|S̃ α (q)|2 ith = kB T ; Γ(q) (4.38) h. . . ith denotes thermal average performed using the Hamiltonian Hh.o. in Eq. (4.36). Thermal averages of the squared transverse components in real space read 1 X α 1 X α ′ h(S α (n))2 ith = hS̃ (q)S̃ α (q ′ )ith e−i(q+q )·n = h|S̃ (q)|2 ith , N q,q′ N q (4.39) where we have used the fact that transverse components fluctuate randomly so that hS̃ α (q)S̃ α (q ′ )ith = δq,q′ h|S̃ α (q)|2 ith . Note that the right-hand side of Eq. (4.39) is independent of the lattice site, thus the label n will be dropped henceforth from h(S α (n))2 ith . In order to evaluate whether the considered linear excitations are able or not to destroy ferromagnetism, we shall let the field B → 0+ . First, we approximate the summation on the right-hand side of Eq. (4.39) with an integral Z d kB T d q α 2 . (4.40) h(S ) i ≃ d (2π) Γ(q) Since what matters is the behavior for small values of q (i.e., the effect of fluctuations at large spatial scales), the denominator of the integral can be linearized as X 1 z̄ 1 (1− qµ2 )+gµB B = J z̄−2J( − q 2 )+gµB B = Jq 2 +gµB B Γ(q) ≃ J z̄−2J 2 2 2 µ 2 with µ=1. . . d and q = P (4.41) 2 µ qµ , which yields Z dd q kB T α 2 . h(S ) i ≃ (2π)d Jq 2 + gµB B (4.42) When taking the limit B → 0+ , the integral in Eq. (4.42) has an infrared divergence4 for d≤2. The consequences of such a divergence can be appreciated more effectively by setting a lower bond to the integral: qmin = π/Nα , 4 A possible ultraviolet divergence does not matter i) because the lattice unit sets a physical upper limit to large values of q ii) because we are interested in fluctuations acting on large spatial scales corresponding to q ∼ 0. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 85 with Nα being of the order of the linear size of the system in lattice units. Depending on the dimensionality of the lattice we have h(S α )2 i ∼ kBJ T Nα Z d=1 d−1 k T q dq B h(S α )2 i ∼ ⇒ d=2 h(S α )2 i ∼ kBJ T ln(Nα ) (4.43) 2 J q qmin d=3 h(S α )2 i < ∞ In order to understand what a divergence with increasing Nα means, it is convenient to rephrase the mathematical steps that we followed according to their physical sense: • We assumed the system to be in a ferromagnetic state at T = 0, namely, with all the spins aligned along the same direction. • We let each spin deviate by a small amount from its direction of alignment, z. • We built an effective linear Hamiltonian, describing these family of small excitations, which can easily be decoupled passing to the Fourier space. • We calculated thermal averages of such small excitations (transverse spin components) in the Fourier space. • We transformed those averages back to the real space. • We evaluated if the initial hypothesis stated in Eq. (4.27) remains valid at finite temperature. The set of Eqs. (4.43) allows stating that in the thermodynamic limit, Nα → ∞, the hypothesis of small deviations fails for d=1, 2 at any finite T . This fact suggests that spontaneous magnetization is not stable against thermal fluctuations for d≤2. On the contrary, according to Eqs. (4.43), it seems possible to have ferromagnetism up to some finite temperature for d=3. This scenario is indeed confirmed by more rigorous proofs such as the MerminWagner theorem. At this point we are in the position to state that for both the isotropic (D = 0) Heisenberg and XY classical model with short-range interactions the lower critical dimension is d=2 (the highest dimensionality for which magnetic order cannot occur at any finite temperature). This result marks a major difference between the universality class of classical spin models with continuous or discrete symmetry (Ising). As already noticed, in systems with continuous symmetry the effects of thermal fluctuations are more severe and manage to destroy ferromagnetism more easily. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 86 The effect of uniaxial anisotropy The presence of uniaxial anisotropy stabilizes a system against the linear excitations considered above. Due to this interaction an additional term like X Hm.a. = −D (S z (n))2 (4.44) n appears in the Hamiltonian in Eq. (4.26). The anisotropy energy indeed favors configurations in which spins lie along a specific axis regardless of the sign of their projections. In this case we choose theP same axis as the one along which the field is applied. Because of the equality α (S α (n))2 +(S z (n))2 = 1, the Hamiltonian (4.44) may also be written as XX (S α (n))2 . Hm.a. = −D + D (4.45) n α The summation appearing above transforms into the Fourier space according to the Parseval’s formula in Eq. (4.33). Finally, the uniaxial anisotropy provides a term into the energy spectrum Γ(q) formally equivalent to the magnetic field: X Γ(q) = J[z̄ − 2 cos(q · δ)] + gµB B + 2D . (4.46) {δ>0} For small values of q and B = 0, we get Γ(q) ≃ Jq 2 + 2D . The average of transverse fluctuations is modified as follows Z Z kB T q d−1 dq dd q α 2 h(S ) i ≃ ∼ k T . B (2π)d Jq 2 + 2D Jq 2 + 2D (4.47) (4.48) Henceforth, let us refer only to the thermodynamic limit Nα → ∞, consistent with qmin = 0. Clearly, the introduction of uniaxial anisotropy removes the infrared divergence4 from the average of fluctuations independently of the lattice dimensionality. The consequences of this result have to be understood as follows: “The considered linear excitations alone are not able to destroy ferromagnetism at any finite temperature”. This statement does not exclude: 1. that ferromagnetism may be destroyed by some other type of excitations 2. that these linear excitations play any role in the “suppression” of ferromagnetism at finite temperature. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 87 An obvious counterexample, supporting the comment 1., is represented by the 1d Ising model. The latter can be considered as a limit case of the Heisenberg model with uniaxial anisotropy (the one we are discussing about here) for D ≫ J. For d=1, as for any d, the integral in Eq. (4.48) is convergent so that linear excitations are not able to destroy ferromagnetism for every T 6= 0. However, we have seen that domain walls manage to destroy ferromagnetism at any temperature in the 1d Ising model. In conclusion, even if the absence of an infrared divergence in the integral on the right-hand side of Eq. (4.48) would allow for ferromagnetism at finite T , we know that a phase with spontaneous magnetization does not occur. Limitations of the mean-field approximation To conclude this part about the critical aspects of magnetism at finite temperature we summarize the artifacts produced by the MFA around the critical region, T ≃ TC . 1. The MFA predicts a the occurrence of a magnetic phase transition at finite temperature independently of the lattice dimensionality, d. However, a phase with spontaneous magnetization is not encountered in the 1d Ising model. The Mermin-Wagner theorem forbids the occurrence of spontaneous magnetization (spontaneous symmetry breaking) in classical models with short-range interactions and with continuous symmetry for d≤ 2. This fundamental theorem applies to both the Heisenberg and the XY model. 2. The transition temperature is generally overestimated within the MFA. 3. The classical values of the critical exponents, i.e., those given by the MFA, are generally not correct even when a phase with spontaneous magnetization exists at finite temperature. Depending on the model, classical critical exponents are wrong for d larger than the lower critical dimension, dl , and smaller that the upper critical dimension (du =4 for systems with short-range interactions). For the Ising model dl =1, while it is dl =2 for the Heisenberg and XY models. 4. The classical critical exponents turn our to be exact for d≥4 in systems with short-range interactions (Ginzburg criterion). Strictly speaking, only the critical exponents are exact for d≥4. However, the MFA is expected to give a more appropriate description of finite-temperature properties for a given model when the number of spins with which each spin interacts increases. This number increases with increasing d or 88 CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS with increasing the range of interactions. The best realization of the MFA is an ideal case in which every spin interacts with any other one with the same intensity. This is – of course – unrealistic when magnetic systems are considered. For the model H=− J X z Ŝ (n) · Ŝ z (n′ ) , N n6=n′ (4.49) with the summation extended over all the different couples, the Mean Field Approximation is exact. The model described by the Hamiltonian in Eq. (4.49) is called Curie-Weiss model but sometimes also “meanfield” model. However, one should not confuse this model with the Mean Field Approximation which does not assume from the very beginning an all-to-all interaction, like in the Hamiltonian of Eq. (4.49)5 . 4.5 Finite size and superparamagnetic limit Figure 3.3 and the following discussion about the absence of ferromagnetism at finite temperature in the 1d Ising model represent the analogous of the arguments given in the present chapter for systems with continuous symmetry, Eqs. (4.43). For the 1d Ising model, through the inequalities in Eq. (3.48), we commented that for small enough system sizes ferromagnetism – possibly present at T = 0 – is stable against thermal fluctuations. Indeed, Eqs. (4.43) allow drawing similar conclusions for system with continuous symmetry: both for d=1 and d=2 ferromagnetism is not destroyed at finite temperature if the system is small enough. Under this condition, the averages of transverse spin components do not necessarily diverge and the inequality h(S α )2 i ≪ h(S z )2 i may be fulfilled. Bistability is a crucial property for most of the applications of nanosized magnets (nanomagnetism). Thus an important question to be addressed is: “what do we understand for bistability when dealing with a real nanomagnet?” Rephrasing what we have just stated about small enough systems, we can answer that when a magnetic lattice does not extend indefinitely correlations – either of short- or long-range nature – may always develop; the system as a whole then behaves like a giant classical spin. In the presence of uniaxial anisotropy, similar arguments as for a single classical spin (macrospin), described by Eq. (4.2), then apply. In particular, the most relevant quantity is the average escape rate from the minima of the total anisotropy energy, located at θ = 0, π (see Fig. 4.1). One possible way to pass from one of the two 5 The exchange interaction is divided by N to guaranty extensivity of the energy. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 89 configurations to the other one is a rigid (coherent) rotation of all the spins6 . The relaxation time for such a mechanism is nothing but a generalization of Eq. (4.4): vDS 2 τ ∼ e kB T , (4.50) with v = Nx Ny Nz . Often the magnetic anisotropy is defined per unit volume Kv = D/a3 so that the usual volume V = a3Nx Ny Nz can be used in Eq. (4.50). Due to the exponential dependence on the system size, the characteristic time given in Eq. (4.50) can become very large even for nanomagnets. Referring again to Fig. 4.1, assume to magnetize the system by means of an external field, thus lowering the energy of one of the two minima corresponding to opposite magnetization. This essentially allows preparing the system in a chosen state. Then remove the external field. Now, due to the exponential divergence of Eq. (4.50), the system may behave as if it had undergone a magnetic phase transition, i.e. it may show a remanent magnetization. But such a situation corresponds to • a metastable state • which is not an equilibrium state (one cannot associate to it a free energy F in the same meaning as, e.g., in chapter 3). • If one could wait long enough, t ≫ τ , an average zero magnetization would be obtained (in the absence of an external field). • A similar scenario is recovered irrespectively of the dimensionality of the lattice d as far as ξ ≫ Nν for all ν = x, y, z. Superparamagnetic limit For magnetic memory manufacturing, the quest to increase the density of data storage calls for reducing the linear dimensions of nanomagents. Even if reducing the linear dimensions of a magnetic unit prevents the occurrence of a magnetic phase transition, one can just require that bistability holds for “long enough”. The required time over which one can reasonably assume that a nanomagnet remains in the desired metastable state depends on the practical application it is supposed to be used for. But, according to Eq. (4.50) and more general approaches, the relaxation time decreases when reducing the linear size of a nanomagnet. As a consequence, when the total volume becomes too small bistability is lost. This intrinsic constraint to the linear dimensions of a bistable magnetic unit is called superparamagnetic limit. 6 For mesoscopic systems, processes which involve non-uniform magnetization reversal may be more convenient. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 4.6 90 Dipolar interaction Assume that we are considering a material which fulfills all the groundstate requirements to give rise to ferromagnetism, as defined in the first two chapters and summarized at the beginning of chapter 3. Assume that ferromagnetism has also “survived” at finite temperatures, meaning that there is a phase with spontaneous magnetization, which requires: • the dimensionality of the magnetic lattice be d≥2 for systems with uniaxial anisotropy • the dimensionality of the magnetic lattice be d=3 for systems with continuous symmetry (Heisenberg or XY). In the former case, ferromagnetism is destroyed by thermally excited domain walls for T > TC (with TC = 0 for d=1). In the latter case, ferromagnetism is destroyed by thermally excited spin waves (or linear excitations in general) for T > TC (with TC = 0 for d=1, 2). Is there another mechanism which can destroy or frustrate the “surviving” ferromagnetism? The answer is “yes”. In fact, the dipolar interaction of magnetostatic origin, neglected so far but always present, may play such a role. Magnetostatic dipole-dipole interaction So far we have not considered the contribution of due to the magnetostatic dipole-dipole interaction, which arises directly from Maxwell equations: µ0 ~µ1 · ~µ2 (~µ1 · r12 ) (~µ2 · r12 ) Hdd = −3 . (4.51) 3 5 4π r12 r12 For our purposes, the pointlike dipoles are the magnetic moments of each ~i (i = 1, 2)7 and magnetic atom in the classical approximation ~µi = −gµB S r12 = r1 − r2 . Since thermal effects may be considered at different levels of approximation, we prefer to distinguish the ground-state magnetic moment ~i from its thermal average m ~µi = −gµB S ~ i = h~µi ith introduced in the previous chapter. The typical strength of the dipole-dipole energy is generally small compared to the exchange energy. However, its characteristics impose to handle the dipolar interaction with extreme caution. 7 For other applications µ ~ i could be, e.g., the electron or the nucleus magnetic moments. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 91 1. Consider just two pointlike magnetic moments which interact via the dipole-dipole interaction (4.51): The sign and the intensity of the dipolar interaction strongly depend on the relative orientation of the two interacting magnetic moments and on their relative spatial position r12 . 2. In contrast to the exchange interaction, the dipolar interaction couples spins located indefinitely far from each other and the decay of its strength with the distance is relatively slow: 1/r3 (long-ranged). 3. In a 3d solid, the dipolar interaction introduces a dependence of the total energy on the shape of the sample itself (shape anisotropy). Equally spaced dipoles E =+ 1 a) E =- 1 b) r12 E =- 2 c) E =+ 2 d) r12 r12 r12 Figure 4.6: Different sample configurations of two dipoles ~µi (i = 1, 2), placed at a fixed distance r12 . The different values of the interaction energy, E, are µ0 µ 2 . given in units Edd = 4π r3 12 To fix the ideas about point 1), let us consider two dipoles at a fixed µ 0 µ2 distance so that the relevant energy scale is given by Edd = 4π 3 . Referring r12 to the configurations in Fig. 4.6, it is clear that when for some reasons (other energies or geometrical constraints) • two interacting magnetic moments are forced to lie perpendicularly to the direction of ~r12 , then the antiparallel alignment is favored by the dipolar interaction (cases a) and b) in Fig. 4.6); • two interacting magnetic moments are forced to lie along the direction of ~r12 , then the dipolar interaction favors their parallel alignment (cases c) and d) in Fig. 4.6). CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 92 In an absolute sense, the configuration c) has the lowest energy of those reported in Fig. 4.6. Parallel dipoles In order to evaluate the dependence of the energy (4.51) on the relative orientation in space of the two point-like dipoles, ~µ1 and ~µ2 , it can be useful to set the origin of the spatial coordinates in ~µ1 , with the z axis parallel to the direction of ~µ1 itself. Then choose ~µ1 k ~µ2 . The resulting energy only 1.0 0.5 E 0.0 -0.5 ferro antiferro ferro -1.0 -1.5 -2.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 q Figure 4.7: Plot of the interaction energy (4.52) as a function of the polar µ0 µ1 µ2 angle θ (rad). The energy values, E, are given in units Edd = 4π . Regions 3 r12 with E < 0 correspond to ferromagnetic coupling, while regions with E > 0 correspond to antiferromagnetic coupling. depends on the polar angle θ defined as the angle that the vector r12 forms with the z axis of our reference frame or, equivalently, with any of the two parallel magnetic moments, ~µ1 and ~µ2 . Within this geometry, the interaction energy (4.51) reduces to µ0 µ1 µ2 2 1 − 3 cos θ . (4.52) Hdd = 3 4π r12 Eq. (4.52) is very interesting because it shows that the dipole-dipole interaction is • ferromagnetic for θ ∈ [0, θM ] and θ ∈ [π − θM , π] 93 CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS • antiferromagnetic for θ ∈ [θM , π − θM ] where θM is the magic angle such that cos2 (θM ) = 1/3 (see Fig. 4.7). Exactly at the magic angle the dipolar interaction vanishes, meaning that in this geometrical configuration the two magnetic moments don’t “feel” each other, for what concerns the dipolar interaction. ? a) r13 r23 r12 F b) S S F S c) S S F S S Figure 4.8: Dipolar-frustrated configurations. a) The vectors ~r12 , ~r23 and ~r13 lie all on the same plane and all the magnetic moments are constrained to be aligned along the indicated direction: up (green) or down (violet). b) and c) All the four dipoles are assumed to lie onto the same plane; the red “F” indicates the frustrated bonds and black “S” satisfied bonds. Fig. 4.8 evidences how the dipolar interaction easily introduces frustration as far as more than two magnetic moments are considered. The configuration a) in Fig. 4.8 represents three magnetic moments magnetized out of plane lying at the vertices of an equilateral triangle. The spins at the bottom of the sketch (up-green and down-violet) minimize their interaction energy by aligning antiparallelly to each other, as in the case of Fig. 4.6 b. Then, according to Fig. 4.7, the two bonds r13 or r23 correspond to ferromagnetic coupling because for these specific cases θ = π/6 < θM . Thus both states up or down of the third spin – the one located at the upper vertex – will produce the frustration of one of the two bonds r13 or r23 . Fig. 4.8 b) and c) refer to a situation in which the four magnetic moments lie onto the same plane. Due to the fact that the configuration c) of Fig. 4.8 is to the global minimum of the two-dipole interaction, the vertical bonds are first fulfilled (ferromagnetic). The horizontal bonds (antiferromagnetic) and the diagonal bonds (ferromagnetic) cannot be satisfied at the same time, so that some frustration is introduced anyway: frustrated bonds are highlighted with red. A detailed calculation shows that, eventually, the configuration in Fig. 4.8 c has a lower energy for a square lattice. The triangle (a) and the square (b) of Fig. 4.8, can be thought of as unit cells of a 2d triangular and square lattice respectively: one can easily imagine that, CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 94 when passing to the thermodynamic limit, many different configurations will have the same, or nearly the same, energy. This fact typically gives rise to very complex behaviors, such as glassiness, metastability, order induced by disorder, spin-ice behavior, etc. 4.7 Dipolar interaction in extended systems In extended systems, the dipolar interaction (4.51) is always present and it involves all the magnetic moments. The resulting contribution to the total energy is " # ~µn · rnn′ ~µn′ · rnn′ ~ n · ~µn′ 1 µ0 X µ Hdip = (4.53) −3 3 5 2 4π n6=n′ rnn rnn ′ ′ where the sum is extended over all the different couples (note the factor 12 !) labeled by n and n′ , and rnn′ = a(n − n′ ) (of course the modulus is rnn′ = a |n − n′ |). Evaluating the term (4.53) is usually complicated analytically and computationally expensive in numerical calculations (due to the longrange character of the dipolar interaction). Thus, in practice, one tries to neglect the dipolar contribution or simplify it taking advantage from the fact that the dipole-dipole interaction normally has a much smaller strength than the exchange interaction. With the help of Table 3.1 one can easily get convinced of this. The strength of the nearest-neighbor dipolar interaction µ0 (gµB S)2 is 4π = M02 a3 (if the saturation magnetization M0 is expressed in the a3 Gauss system, M02 has the units erg/cm3 ). Putting the proper numbers one finds that this energy is of the order of few Kelvins or smaller. However, there are cases in which the dipolar interaction may affect crucially the macroscopic behavior of a magnetic system. In the following we will give some examples. 2d systems with uniaxial anisotropy Here we start again from the 2d Ising model, in which spontaneous magnetization is encountered at finite T , and consider the qualitative effect of the dipolar interaction. Typically, real systems with uniaxial anisotropy can possibly be experimental counterparts of the Ising model. We should also keep in mind that the exchange interaction has usually a much larger strength than the dipolar interaction (a factor 102 − 103 ). From what stated in the previous sections, it is clear that in such systems the dipolar interaction will play a different role depending on the direction of the easy axis: CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 95 • magnetization in plane: the dipolar interaction is globally satisfied (see Fig. 4.8 b) • magnetization out of plane: the dipolar interaction is frustrated and tends to “destroy” ferromagnetism (see Fig. 4.6 b). Figure 4.9: Experimental check of the 2d-Ising scaling behavior, observed in Fe/W(110) films magnetized in plane. C. H. Back et al., Nature 378 p. 597. The first case is realized, for instance, in Fe/W(110) ultrathin films. These films are model realizations of the 2d Ising model as they obey the predicted scaling behavior for T ∼ TC over eighteen orders of magnitude (see Fig 4.9). As pointed out in Nature 378 p. 597, neither the dipolar interaction nor other effects which are neglected in the ideal model seem to affect the 2dIsing critical behavior observed in Fe/W(110). An example of ultrathin film magnetized out of plane is represented by Fe/Cu(001). Here the competition between the ferromagnetic exchange interaction – originating ferromagnetism for T < TC – and the dipolar interaction – frustrating ferromagnetism on a larger scale – produces a sort of phase separation between regions of positive and negative magnetization perpendicular to the film plane (see the scheme in Fig. 4.10). In other words, ferromagnetism is limited to some spatial regions in which all the magnetic moments point along the same direction. Such regions are called magnetic domains. The whole scenario holds below the Curie temperature, thus magnetic domains need not be confused with the spatial regions defined by the correlation length ξ. In Fig. 4.11 some images of different magnetic-domain patterns observed in Fe/Cu(001) films are shown. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 96 Figure 4.10: The competition between exchange interaction and the dipolar interaction originates a modulated phase in magnetic films magnetized out of plane. Other dimensions Since a spontaneous magnetization does not exist at any finite temperature, the effect of dipolar interaction is usually less dramatic in 1d. However, the dipolar interaction may still affect the elementary excitations which destroy ferromagnetism • spin-waves in isotropic spin chains (delocalized excitations) • domain wall in spin chains with uniaxial anisotropy (localized excitations) and finally modify the behavior of ξ(T ) (e.g. how it diverges at low T ). The calculation of the dipolar interaction energy of one magnetic moment with all the others involves an integral (see next section) like Z dd r (4.54) 3 Nd r which is convergent for d<3. For d=3 the total magnetostatic energy is conditionally convergent and, as anticipated, depends on the shape of the CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 97 Figure 4.11: SEMPA images of magnetic domains in Fe/Cu(001) films: dark and light gray regions correspond to domains with opposite out-of-plane magnetization. sample (shape anisotropy). In 3d, both frustration effects and magneticdomain formation occur in a similar way to what discussed for d=2. Continuum limit The continuum version of the dipolar energy given in Eq. (4.53) can be ~n followed by the usual substitution obtained by setting ~µn = −gµB S Z X 1 · · · ≃ d . . . dd r , (4.55) a which yields Z ~ ~ ′) S(r) · S(r Hdip dd r ′ − d r |r − r′ |3 ih i h ′ Z Z S(r) ~ ′ ) · (r − r′ ) ~ 2 ) S(r · (r − r µ0 (gµB S) 1 dd r ′ + −3 dd r 5 3 2d−3 ′ 8π a a |r − r | Z Z µ0 (gµB S)2 1 d ~ ′ )δ (r − r′ ) dd r′ ~ + · S(r d r S(r) 6 a3 a2d−3 (4.56) µ0 (gµB S)2 1 = 8π a3 a2d−3 Z d The unit lengths a have been grouped in such a way that the characteristic µ0 (gµB S)2 is separated from the energy scale of the dipolar interaction Ω = 4π a3 geometrical terms. The term δ (r − r′ ) arises from a detailed magnetostatic calculation. It essentially accounts for the fact that the pointlike-dipole approximation has to include this term to compensate the divergence at short distances, r ≃ r′ . Thus the total magnetostatic energy is not ill-defined. CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 98 It is also useful to recall that the magnetization is introduced in elemen~ (r) tary courses as a coarse-grained quantity. Ideally, the magnetization M already represents an average over an elementary volume whose magnetic ~ (r)d3 r. The magnetostatic energy then reads moment is8 ~µ = M Z Z µ0 µ0 3 Edip = ρm (r)φm (r)d r + σm (r)φm (r)d2 r , (4.57) 2 V 2 ΣV where ~ ·M ~ ρm = − ∇ and ~ · n̂ σm = M (4.58) are the volume and surface charge density respectively (n̂ normal to the surface ΣV ), while Z Z 1 1 ρm (r′ ) 3 ′ σm (r′ ) ′ φm (r) = d r + dΣ (4.59) 4π V |r − r′ | 4π ΣV |r − r′ | V is the scalar magnetic potential (see, e.g., “Classical Electrodynamics”, J. D. Jackson). Note that, within this formalism, the total magnetostatic energy is the Coulomb energy of an effective Coulomb charge distribution ρm (r). Such an analogy provides a very helpful “rule of thumb” for analyzing realistic situations: the magnetostatic energy is minimized by configurations for which there are as less magnetic charges as possible. Even if derived in a coarse-grained context, the validity of this “rule of thumb” is quite general, i.e., it provides the correct hints also in the discretelattice formalism. Bloch and Néel domain walls In section 4.3 of the present chapter we have investigated how domain walls with a finite width δ emerge from the competition between the anisotropy and the exchange energy. The “compromise” which minimizes the domainwall energy is represented by the solution (4.23) which we recall here for convenience: ( cos(θ(x)) = tanh xδ (4.60) ϕ(x) = const. All the solutions with constant ϕ(x) give the same energy if inserted in the Hamiltonian (4.20). Now we ask ourselves whether the introduction of the ~ (r)d3 r, but for the present purposes it is convenient More precisely, it is m ~ = h~ µith = M to keep considering the dipolar interaction at T = 0, for which m ~ =µ ~. 8 CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 99 dipolar energy term (4.56) may favor one specific value of ϕ(x). The two extreme cases are named ( ϕ(x) = π2 ∀x Bloch domain wall (4.61) ϕ(x) = 0 ∀x Néel domain wall . More specifically the magnetization for these two cases will be ( ~ = M0 (0, sin(θ(x)), cos(θ(x))) M Bloch domain wall ~ M = M0 (sin(θ(x)), 0, cos(θ(x))) Néel domain wall . (4.62) with θ(x) given by Eq. (4.60). In the bulk case (3d), it is evident that the Figure 4.12: From Introduction to the Theory of Ferromagnetism by A. Aharoni. Energy per unit wall area, γ, (solid curves) as a function of the thickness for a permalloy film magnetized in plane. Dashed curves display the domain-wall width (q ∝ δ in our notation). Bloch wall has always a lower energy since ( ~ ·M ~ = ∂ x M x + ∂ y My + ∂ z Mz = 0 ∇ 6= 0 Bloch domain wall Néel domain wall , (4.63) CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 100 so that only the Néel domain wall produces some magnetic charges ρm = ~ ·M ~ . This rule-of-thumb prediction is confirmed by detailed calculations −∇ and experiment. The situation is analogue for thin films (2d) with the easy axis pointing out of plane. For thin films magnetized in the xz plane (with easy axis parallel to z), the surface charges σm produced by a Bloch wall (My needs to point out of plane) can be so large that the Néel wall becomes energetically more convenient. In this case, surface charges (Bloch wall) are replaced by volume charges ρm . The solid curves in Fig. 4.12 display the domain-wall energy corresponding to a Bloch and a Néel wall, computed with the phenomenological parameters of permalloy (alloy of ≃ 20% Fe and ≃ 80% Ni), as a function of the film thickness. Indeed, for this specific material with this specific geometry, up to thicknesses of the order of 60 nm the Néel wall has a lower energy. 4.8 Origin of magnetic domains The scenario described schematically in Fig. 4.10 is just a particular case in which the competition between the exchange and dipolar energy gives rise to a configuration with zero global magnetization. Such a configuration is the compromise which produces as less magnetic charges as possible with the minimum frustration of the exchange interaction. What results from the competition between these two energies is generally different depending • on the easy-axis direction • on the geometry of the sample so that each case needs to be evaluated on its own. As a simple example, let us consider – again – a 2d system magnetized out of plane and evaluate the energy variation associated with the creation of a domain wall from a uniformly magnetized state (see Fig 4.13). Both the exchange energy and uniaxial anisotropy contribute to the domain-wall energy Ew , thoroughly discussed in the previous sections and whose values are summarized schematically in Eq. (4.24). When deriving those results we assumed that the spin profile was a function of one spatial variable only. Now we consider a film of finite thickness t and a domain wall developing indefinitely along the y direction, so that the total increase of the exchange and anisotropy energy is Ny Nz Ew , with Nz = t/a. For a film magnetized out of plane and with thickness t of few monolayers the dipolar energy (4.56) can CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 101 be approximated as Hdip 1 t2 = Ω 3 2 a Z 2 dr Z S z (r)S z (r′ ) 2 ′ dr . |r − r′ |3 (4.64) Splitting the xy plane into two half-planes and reversing the magnetization (along z) of one of the two half-planes produces a decrease of the dipolar Figure 4.13: Sketch of the two configurations corresponding to the energy difference evaluated in Eq. (4.65). Arrows represent magnetic moments pointing out of plane. energy by a factor ∆Edip t2 = 2Ny Ω 2 a Z Nx a/2 dx δ Z 0 dx −Nx a/2 ′ Z +∞ −∞ 1 [(x − x′ ) 2 + y 2 ]3/2 dy ; (4.65) the integral over dx is performed starting from a length scale equal to the domain-wall width δ in order to avoid an unphysical divergence. In other words, spins located inside the domain wall have been ideally “removed” from the calculation of the dipolar energy. Note that ∆Edip in Eq. (4.65) represents the variation of the interaction energy between the two half-planes; the magnetostatic self-energy of the two half-planes remains the same. The integral in Eq. (4.65) can be performed analytically and gives t2 Nx a + 2δ ∆Edip = 4Ny Ω 2 ln . (4.66) a 4δ The condition ∆Edip = Ny Nz Ew gives the minimum linear dimension Nx that a slab should have in order that splitting the uniform state into domains becomes favorable. In the realistic limit of Nx a ≫ δ, one finds that for Nx larger than the threshold value 4δ Ew a N̄x ≃ exp (4.67) a 4Ωt CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 102 it is convenient for the system to split into domains of opposite out-of-plane magnetization. From this rough calculation one expects the typical size of domains to be Leq ≃ N̄x a/2. The exponential dependence on the ratio Ew /Ω is typical of d=2. For the 3d case, a further integration would be involved in the evaluation of ∆Edip which, eventually, would result in a much weaker dependence of the domain size Leq on the ratio Ew /Ω. Striped pattern Figure 4.14: Schematic view of the striped ground state of a ferromagnetic film magnetized out of plane in the presence of the dipolar interaction and B = 0. Stripes of different colors represent regions of opposite out-of-plane magnetization. The typical stripe width is Leq . In spite of the crude approximations that have been performed to obtain Eq. (4.67), the predicted scaling with Ew /Ω and with the film thickness t match with the optimal stripe width for an ideal stripe pattern at T = 0 (see Fig. 4.14). This pattern corresponds to the ground state of a film magnetized out of plane in the presence of dipolar interaction and zero external field. Detailed calculations9 yield for Ω ≪ D ≪ J 10 Ew a (4.68) δ exp Leq = 3π 4Ωt and for Ω ≪ J ≪ D (Ising domain wall) Leq = L0 exp 9 Ja 2Ωt (4.69) Case Ω ≪ D ≪ J adapted from S. A. Pighı́n et al., Journal of Magnetism and Magnetic Materials 322 p. 3889 (2010). Case Ω ≪ J ≪ D (Ising) adapted from A. B. MacIsaac et al., Physical Review B 51 p. 16033 (1995). CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 103 with L0 = 0.871a. Remembering that in the second case Ew = 2J and δ = a, both Eqs. (4.68) and (4.69) confirm the scaling predicted by Eq. (4.67) on the basis of heuristic arguments. Characteristic length scales emerge from competing interactions The typical domain size Leq is the second characteristic length scale that we have encountered arising from the competition between different energies ( J versus D ⇒ δ domain-wall width (4.70) J versus Ω ⇒ Leq domain width . Notice that calculations to obtain both δ and Leq have been performed at T = 0. Since the exchange interaction, the anisotropy energy and the dipolar interaction involve different number of spins and different components of each spin (e.g., uniaxial anisotropy contains only a single-site term (S z )2 ), in general, they will be affected differently by thermal fluctuations. This fact, may finally favor one of two competing interactions, thus introducing an effective dependence on temperature in the characteristic length scales δ and Leq . CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS Literature • A. Aharoni, Introduction to the Theory of Ferromagnetism Oxford University Press • L. D. Landau and E. M. Lifshitz, Statistical Physics Oxford Pergamon Press • J. D. Jackson, Classical Electrodynamics John Wiley and Sons • C. Kittel, Introduction to Solid State Physics John Wiley and Sons Dr. Alessandro Vindigni Laboratorium für Festkörperphysik Wolfgang-Pauli-Str.16 ETH Hönggerberg, HPT C 2.2 Tel. +41 44 633 2077 Fax. +41 44 633 1096 e-mail: vindigni@phys.ethz.ch 104 CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 105 Appendix Averages at finite temperatures Classical models In the canonical ensemble, the partition function is given by Z 3N 3N d qd p −βH(p,q) Z= e , (2π~)3N H being the Hamiltonian of the system and β = free energy, F , via the general relation 1 . kB T Z is related to the 1 F = − ln Z . β The average of any observable O (p, q) can be computed as Z 3N 3N 1 d qd p hOi = O (p, q) e−βH(p,q) . Z (2π~)3N Classically, the trace operator is defined as Z d3N qd3N p , T r = ... (2π~)3N which allows defining Z = T r e−βH(p,q) and hOi = (4.71) 1 T r O (p, q) e−βH(p,q) . Z (4.72) (4.73) (4.74) (4.75) CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 106 Quantum models Assume that |ψα i be a complete basis of the Hilbert space on which the Hamiltonian of the model is defined. Quantum-mechanically, the trace is then given by X Tr= hψα | . . . |ψα i . (4.76) α By analogy with (4.75), the partition function and thermal averages are accordingly defined X hψα |e−βH |ψα i Z = T r e−βH = α 1 1 X hOi = T r Oe−βH = hψα |Oe−βH |ψα i . Z Z α (4.77) In few advanced computations one stops at this level. Generally, the trace is evaluated on a complete basis of eigenstates of H: H|ϕi i = E i |ϕi i . (4.78) The computation of (4.77) is, consequently, simplified: X X i Z= hϕi |e−βH |ϕi i = e−βE i i 1 hOi = T r Oe Z −βH 1 X i i = hϕ |O|ϕi ie−βE . Z i (4.79) Spin models Limiting ourself to a Hamiltonian of the type 1 H=− J 2 X |n−n′ |=1 Ŝ(n) · Ŝ(n′ ) + gµB B X Ŝ z (n) (4.80) n one possible choice for the basis of the Hilbert space is the following one: |ψα i=|M1 , M2 , . . . MN i=|M1 i ⊗ |M2 i · · · ⊗ |MN i with Ŝ z (n)|Mn i=Mn |Mn i and n label for the lattice site. Note that the Hamiltonian in Eq. (4.80) is not diagonal on this basis. Thermal averages are computed according to Eqs. (4.79). For many problems in magnetism, substituting the quantum-mechanical operators Ŝ(n) by classical vectors is legitimate: ~ ≡ S0 (sin θ cos ϕ, sin θ sin ϕ, cos θ) Ŝ(n) → S(n) (4.81) CHAPTER 4. MAGNETIC DOMAINS AND DOMAIN WALLS 107 where S02 = S (S + 1) (more often S0 = 1). The partition function then reads Z Z Z ~ Z = dΩ1 dΩ2 . . . dΩN e−βH({S(n)}) , (4.82) with dΩn = sin θn dθn dϕn being the solid-angle element of the spin located at the site n. Both in the quantum and the classical case Z depends on T and B. Therefore, the free energy associated with the canonical averages is the Gibbs free energy of macroscopic thermodynamics: 1 G(B, T ) = − ln [Z(B, T )] . β (4.83) Often in the literature the letter F is used also when dealing with spin models to stress the fact that this type of averages are performed in the canonical ensemble (constant number of particles). We will follow this convention.