Predictive Complexity of Quantum Subsystems Authors: Curtis T. Asplund, Elisa Panciu Published: 2024 ABSTRACT We define predictive states and predictive complexity for quantum systems composed of distinct subsystems. This complexity is a generalization of entanglement entropy. It is inspired by the statistical or forecasting complexity of predictive state analysis of stochastic and complex systems theory but is intrinsically quantum. Predictive states of a subsystem are formed by equivalence classes of state vectors in the exterior Hilbert space that effectively predict the same future behavior of that subsystem for some time. As an illustrative example, we present calculations in the dynamics of an isotropic Heisenberg model spin chain and show that, in comparison to the entanglement entropy, the predictive complexity better signifies dynamically important events, such as magnon collisions. It can also serve as a local order parameter that can distinguish long and short range entanglement. 1. Introduction Quantum systems of many particles are capable of an astonishing variety of complicated behaviors. Characterizing these behaviors in a useful way can be challenging. Building on the foundation of entanglement entropy, and incorporating ideas from complex systems theory, we define new quantities for such systems: predictive states and predictive complexity. We then apply them to a prototypical example, a Heisenberg model spin chain, and find that they provide an improved measure of some dynamical processes and they can serve as promising new tools for analyzing general quantum dynamical systems. We are primarily guided by two lines of research. On one hand, entanglement entropy of lattice systems and field theories has been a highly successful tool in recent times, from the celebrated area law [1] to many applications in quantum dynamics [2,3]. On the other hand, we have the statistical or forecasting complexity of stochastic and complex systems [4,5,6], which has developed into a sophisticated formalism that has been applied to numerous systems [7,8,9]. We distinguish our definition of predictive complexity from the quantum Hamiltonian complexity [10,11], which classifies systems according to complexity classes, and so does not determine a definite complexity value. It is also distinct from the quantum computational or gate complexity [12,13], which is the basis for most work in high-energy physics and gauge/gravity (AdS/CFT) duality [14,15]. In particular, our construction is information-theoretic and has no dependence on a choice of gates or reference state. Our definitions can be seen as quantum extensions of predictive state analyses of classical spin systems [16,17,18,19], cellular automata [20], various spatio-temporal dynamical systems [21,22], and of optimal quantum models of stochastic systems [23,24,25,26,27,28]. It is similar in spirit to using the statistical complexity to analyze the distribution of measurement outcomes of a quantum system [29,30,31], but we do not involve measurement distributions of any particular observable. Several other quantities called “complexity” have been applied to quantum and spin systems [32,33,34,35,36,37]. These are defined differently to the complexity we define here, but relations between these would be interesting to investigate further. 2. Methods and Definitions 2.1. Predictive Equivalence Consider a quantum system composed of a subsystem A and its complement or exterior B. We have in mind a finite set of interacting particles, spins, or qubits, with A consisting of a subset. Accordingly, we write the finite-dimensional Hilbert space H of this system as (1)H=HA⊗HB. For any density operator of such a system, one can form the reduced density operators using partial traces ρA=trBρ and ρB=trAρ. In the case of a pure state, ρ=|ψ〉〈ψ|, the von Neumann entropies are equal and give the entanglement entropy between A and B:(2)SA=−trρAlogρA=−trρBlogρB=SB. From information theory (coding theorems), this entropy may be interpreted as the amount of information needed to describe the state of A, on average, using a suitable encoding of a basis of HA. We may also interpret this as the amount of information about B, relevant to A, that has been lost due to tracing out B. This measures how much A is entangled with B and vice versa. Grouping or coarse-graining outcomes of a stochastic process generally leads to a reduction in entropy since less precise information is needed to describe the outcomes. Suppose some states in HB were equivalent, in some sense, say, in terms of how they affect the evolution of ρA. Then, we would expect to need less information about what is going on in B to effectively predict the behavior in A. Below, we describe how to implement any such equivalence in terms of how it affects ρA and the entropy. We focus on an equivalence defined by identical evolution of a subsystem for some time, and the consequently reduced entropy is thus a measure of the average minimal information needed to sufficiently predict the state of A, or observables in A, for that time. This is the rough idea for what we call the predictive complexity. We are following the ideas of causal or predictive state analysis, sometimes also called computational mechanics [4,5,6]. In the context of classical and stochastic dynamical systems, this type of analysis defines an equivalence relation between states that yield equivalent probabilistic predictions for the system. In our quantum context, let us call two states |ϕ1〉,|ϕ2〉∈HB, equivalent if and only if they lead to the same dynamics in A. More precisely, for a given |ψ〉∈HA, we can form the states |Ψ1〉=|ψ〉|ϕ1〉,|Ψ2〉=|ψ〉|ϕ2〉∈H and the corresponding density operators ρ1=|ψ〉|ϕ1〉〈ϕ1|〈ψ| and ρ2=|ψ〉|ϕ2〉〈ϕ2|〈ψ|. Let us write ρ1(t) and ρ2(t) for the time-evolved density operators, i.e., if U(t)=Ut=e−iHt/ℏ is the time-evolution operator for a time-independent, whole-system Hamiltonian H, ρ1(t)=U(t)ρ1U†(t). Then, we can define |ϕ1〉 and |ϕ2〉 to be equivalent if and only if (3)trBρ1(t)=trBρ2(t), for all |ψ〉∈HA and for some interval of time, which we call the time horizon th. Reflexivity, symmetry, and transitivity of the relation follow quickly from Equation (3), so it is an equivalence relation. This also comes from it being defined by the equality of the images of a function on HB. One can express Equation (3) in terms of relative entropy, or Kullback–Leibler divergence, defined for any two density operators ρ and σ to be S(ρ∥σ)≡tr(ρlogρ−ρlogσ). Since the relative entropy of two density operators is zero if and only if they are equal, Equation (3) becomes (4)StrBρ1(t)∥trBρ2(t)=0. Equivalently, one could express this by saying the distance between the two density operators is zero for a given interval of time, under any metric on the space of density operators, such as the trace distance. In fact, all one needs to have well-defined equivalence classes is a partition of HB, and this could come from a generalization of Equation (4), where the right-hand side is not strictly zero but just sufficiently close to zero so that it leads to sufficient clustering of states in HB to be partitioned (in a way that respects the linear and inner product structure, as we describe further below). We plan to investigate such generalizations in future work. The relation clearly depends on the time horizon. In relativistic systems and in a case where A is taken to be a compact spatial region, this would translate into a spatial horizon distance rh=th/c. By relativistic causality, we expect all states of HB that only differ in their configurations at distances from A greater than rh to be equivalent to each other. We have similar expectations for quantum systems obeying a Lieb–Robinson bound since this similarly limits the speed of propagation of information [38]. In that case, though, exponentially small corrections are allowed. This is where small corrections to the right-hand side of Equation (4) could be important. We are interested in the states in HB that can be partitioned into equivalence classes in terms of their effect on the subsystem A. We will explicitly see how this works in the case of the Heisenberg model below. As we will see, the linear and inner product structure of H combined with the unitary evolution operator naturally gives rise to equivalence classes that can be thought of as parallel affine subspaces, akin to superselection sectors. They may also be interpreted as projective subspaces in projective Hilbert space, which can be a useful perspective since it quotients out un-normalized state vectors and we are ultimately interested in normalized states and density operators. (Projective Hilbert space, or ray space, is the space formed from a Hilbert space by taking the quotient by the relation |ψ〉∼z|ψ〉, where z is any non-zero complex number. It can be thought of as the space of physically distinct states in H [39,40,41]). The predictive equivalence defined by Equation (3) is natural from the point of view of reduced density operators, but to define an equivalence relation with the desired linearity properties on the Hilbert space HB, we need a further requirement on the dynamics of the equivalent states. To see this, let |ϕ1〉,|ϕ2〉∈HB be normalized, orthogonal state vectors satisfying Equation (3). Then, let |ϕ3〉=a1|ϕ1〉+a2|ϕ2〉≠0 with |a1|2+|a2|2=1, i.e., an arbitrary normalized linear combination. We can use a Schmidt decomposition to write the evolution of the states |Ψ1〉=|ψ〉|ϕ1〉,|Ψ2〉=|ψ〉|ϕ2〉 (in the Schrödinger picture of time evolution) as (5)|Ψ1(t)〉=Ut|Ψ1〉=c1ij(t)|ψi〉|ϕj〉|Ψ2(t)〉=Ut|Ψ2〉=c2ij(t)|ψi〉|ϕj〉, where i and j are summed over bases for HA and HB, respectively. The evolution of |ψ〉|ϕ3〉 can then be written (6)Ut|ψ〉(a1|ϕ1〉+a2|ϕ2〉)=a1c1ij(t)|ψi〉|ϕj〉+a2c2ij(t)|ψi〉|ϕj〉. A short calculation shows that to ensure the property in Equation (3) holds for ρ3, we need, for all i and k, (7)∑jc1ij(t)c2kj*(t)=0. This can be understood as requiring that Ut maintain orthogonality between equivalent states in HB, and it can also be rewritten (8)trB|Ψ1(t)〉〈Ψ2(t)|=0. We have written this as a strict operator equation, but as with our discussion of Equation (3), all we need for our construction to work is for the operator OA:=trB|Ψ1(t)〉〈Ψ2(t)| to be sufficiently small, in the sense of whatever metric one is using. For example, with respect to trace distance, Equation (8) would be replaced with a requirement that the trace norm ∥OA∥1=trOA†OA is small compared with the trace distance between any reduced density operators generated by inequivalent states, for times within the time horizon. Below, we argue that this holds for many systems and we numerically verify it for our main example system of a Heisenberg spin chain. It is straightforward to show that Equation (8) holds when A and B are decoupled since then we may write |Ψi(t)〉=|ψ(t)〉|ϕi(t)〉. Thus, (9)trB|Ψ1(t)〉〈Ψ2(t)|=〈ϕi||ψ(t)〉|ϕ1(t)〉〈ψ(t)|〈ϕ2(t)||ϕi〉=〈ϕ2(t)||ϕi〉〈ϕi||ϕ1(t)〉|ψ(t)〉〈ψ(t)|=〈ϕ2(t)|ϕ1(t)〉|ψ(t)〉〈ψ(t)|, which equals zero since |ϕ1〉 and |ϕ2〉 were assumed to be orthogonal, and this is preserved if the dynamics in A and B are decoupled. To see what this means in terms of the system Hamiltonian H that describes local interactions in a spatially extended system, let us assume that it can be written in the form H=HA+HB+HAB, where all these terms are assumed to be time independent and where HA and HB are defined in terms of operators that only act on HA and HB, respectively. In particular, [HA,HB]=0 and HAB contains all terms that couple the subsystems A and B. In general, [HA,HAB]≠0 and [HB,HAB]≠0, but we expect these commutators to be relatively small since they only involve operators on the boundary between A and B. More precisely, in a d-dimensional lattice system with O(Nd) sites, we expect expectation values of HA and of HB to be O(Nd), while those of HAB and its commutators to be O(Nd−1). This is important as we analyze the criteria in Equations (7) and (8). We first note that orthogonality is preserved under the evolution generated by HB alone, i.e., (10)〈ϕ1|ϕ2〉=〈ϕ1|eiHBte−iHBt|ϕ2〉=0. Now, we write the full evolution operator as Ut=exp[−i(HA+HB+HAB)t/ℏ]. By using the Zassenhaus formula (a relative of the Baker–Campbell–Hausdorff formula) twice, we can write (11)Ut≈e−iHBt/ℏe−iHAt/ℏe−iHABt/ℏ× e−12[HA,HAB](−it/ℏ)2e−12[HB,HAB](−it/ℏ)2, where higher order factors involve nested commutators of HB and HAB, higher powers of −it/ℏ, and increasingly small coefficients in the exponents, all of which generally tend to suppress their relative importance [42]. The factors leading to any violation of Equation (8) are thus systematically suppressed. This is in addition to the difference in scaling of HA, HB, and HAB already mentioned. An exact analysis of these factors will depend on the system. For example, in this paper, we look in detail at the the Heisenberg spin chain model with N sites and periodic boundary conditions, which has Hamiltonian (12)H=−J∑n=1NSn·Sn+1, where Sn=(Snx,Sny,Snz) is the spin operator at site n. We take A to be a block of adjacent spins going from site 1 to NA, and in this case, (13)HAB=−J(SN·S1+SNA·SNA+1). From this, we can see that [HA,HAB] and [HB,HAB] will involve terms of the form (14)[Sn−1·Sn,Sn·Sn+1]=[Sn−1iSni,SnjSn+1j]=Sn−1iSn+1j[Sni,Snj]=Sn−1iSn+1jiℏϵijkSnk=iℏϵijkSn−1iSn+1jSnk=iℏ(Sn−1×Sn+1)·Sn, where we have used the fact that spin operators at different sites commute. Applying this, we have (15)[HB,HAB]/iℏ=(SN−1×S1)·SN−(SNA×SNA+2)·SNA+1, and (16)[HA,HAB]/iℏ=(SNA−1×SNA+1)·SNA−(SN×S2)·S1. We see from Equations (13), (15) and (16) that HAB and its commutators scale like the boundary of A, which in this one dimensional case, is just fixed to be two lattice points or O(N0) (while HA and HB will generally have O(N) terms). When inserted into the Zassenhaus expansion in Equation (11), these factors may lead to some violation of Equation (8), but this will be suppressed in the limit of a large lattice. This is consistent with the Lieb–Robinson bound, which allows for small corrections [38]. As part of this work, we generated numerical evidence that these corrections do indeed remain small in the system we consider. In particular, we considered a set operators like those appearing in Equation (8), constructed from 14 randomly chosen pairs of equivalent states in HB, and verified that their trace norm remained small, O(10−3) or less, up until the time horizon. We plot a sample of these data in Figure 1. In comparison, the trace distance between reduced density operators coming from inequivalent states is O(1). 2.2. Predictive States Any equivalence relation on HB can be seen as a map from it to a quotient space HB′, which we will call the predictive state space. We will see below that the quotient map is a linear projection map and HB′ can be considered a Hilbert space itself. Here, we focus on how the quotient map affects the reduced density operator ρA and its entropy, and how to interpret the results information-theoretically. Let us consider a Hilbert space HB of arbitrary dimension N3. Suppose {|φi〉} is an orthonormal basis for a linear subspace of equivalent states, of dimension N1. Let N2 be the dimension of the orthogonal complement of the subspace, so that N1+N2=N3. We define a linear projection operator P that maps all difference vectors |φi〉−|φj〉 to zero. This projects the subspace onto the span of the normalized vector |γ〉=(1/N1)∑i=1N1|φi〉 and may be written as P=IdN2+|γ〉〈γ|. To have a probabilistic interpretation of the states after projection, they must be normalized. In line with the idea of grouping equivalent states, we require that the probability of the predictive state equal the sum of all the probabilities in the equivalence class. A general state |Ψ〉∈H=HA⊗HB may be written in a Schmidt decomposition as (17)|Ψ〉=∑i=1dimHA∑j=1N3aij|ψi〉|ϕj〉. The normalization of the equivalent states will have a non-trivial effect on terms of the form a1|ψ〉i|φ1〉+…+aN1|ψ〉i|φN1〉, i.e., terms that include equivalent states. In our prescription, these terms will be mapped to the following in the normalized state |Ψ′〉:(18)∑j|aj|2∑jaj∑jaj|ψi〉|γ〉, where the sums are from 1 to N1. The first factor, involving a square root, is the normalization necessary for a consistent probabilistic interpretation of the coefficients. The second factor is a unit complex number that comes from the projection by P. This normalizing of coefficients makes the overall map, from the normalized states in HB to the normalized states in HB′, non-linear. In the case of additional, distinct subspaces of equivalent states, the discussion above is iterated and more coefficients will be involved. The equivalent subspaces may be directly summed into an increasing sequence and so, in linear algebra terms, form a (partial) flag for HB. This is the same kind of structure that appears when decomposing a Hilbert space into superselection sectors [43]. The relevant projection operator P is then the identity on the subspace of inequivalent states plus a sum over projectors |γi〉〈γi| onto rays, one for each distinct subspace of equivalent states. We omit the explicit general expressions since we do not need them for our purposes here. In the end, we are interested in a normalized state |Ψ′〉 in the predictive state space HB′, and we have implicitly defined it to be P|Ψ〉 after its coefficients are normalized accordingly (18). We may then define the predictive state density operator ρ′=|Ψ′〉〈Ψ′| and its partial traces ρB′=trAρ′ and ρA′=trB′ρ′. Here, trB′ means a partial trace over the the predictive state space HB′. We work out explicit examples of these definitions in Section 2.4. The predictive states are used to define the predictive complexity, which we examine next. 2.3. Predictive Complexity We define the predictive complexity CA of A to be the von Neumann entropy of the predictive state density operator defined above, ρA′=trB′|Ψ′〉〈Ψ′|, (19)CA=−trρA′logρA′, where |Ψ′〉 is the state formed by implementing predictive equivalence, as defined by Equation (3), and normalized according to Equation (18). We interpret CA as the entropy of A due to tracing over A’s environment B, modulo predictively equivalent states in B. One could similarly define other measures of predictive complexity from other information measures of ρA′, such as the Renyi entropy, negativity, etc. We conjecture that CA≤SA=−trρAlogρA, i.e., that the predictive complexity is bounded above by the standard entanglement entropy, with equality if no states in B are predictively equivalent. This is true in all cases we have examined. Similarly, in practice, one may only know that a certain subspace of states in HB is predictively equivalent for some interval of time. For example, states supported outside of some horizon around A. The true set of equivalencies may be greater, and we expect this would only decrease the resulting complexity. In that case, one computes an upper bound on the predictive complexity. Many additional questions may be asked about the predictive complexity CA and its properties. We address some of these in the Discussion, but many are the subject of future research, and we turn now to an example calculation of this quantity in a model system. 2.4. Predictive States and Complexity in a Small System We present a low-dimensional example where we can work out the definitions of predictive state and predictive complexity explicitly. Let us consider HA≅C2 and HB≅C3. Let {|1〉A,|2〉A} be an orthonormal basis for HA and {|1〉B,|2〉B,|3〉B} be one for HB, and suppose that |1〉B∼|2〉B, i.e., they are found to be equivalent under Equation (3). Form the normalized sum and difference vectors |α〉:=(1/2)(|1〉B−|2〉B) and |β〉:=(1/2)(|1〉B+|2〉B). We form the projection operator P=|3〉B〈3|B+|β〉〈β| on HB that projects along |α〉 and onto its orthogonal complement, the image of P. Call that image HB′, in this case, a two (complex)-dimensional subspace, spanned by |3〉B and |β〉. Let us look first at a general state in the 1−2 subspace of HB, where we have P(a1|1〉B+a2|2〉B)=12(a1+a2)(|1〉B+|2〉B)=12(a1+a2)|β〉. If we normalized this state, we would end up with (20)a1+a2|a1+a2||β〉. We note that if a1+a2=0, the state is mapped to the zero vector and so lies in the kernel of P. Such states are effectively excluded from the final state space and are concomitant of having state vectors that are physically equivalent in some sense. The same phenomenon occurs in quantization of gauge theories, as in the BRST formalism, and we plan to elaborate on this connection in future work. Next, take a general normalized state in H, given by (21)|ψ〉=a1|11〉+a2|12〉+a3|13〉+a4|21〉+a5|22〉+a6|23〉, where |ij〉=|i〉A⊗|j〉B. Under P, this state is mapped to the unnormalized state (22)P|ψ〉=12(a1+a2)|1〉A|β〉+a3|1〉A|3〉B+12(a4+a5)|2〉A|β〉+a6|2〉A|3〉B. To normalize this state, we focus on the terms that have been affected by the projection. For example, the contribution of the first two terms in Equation (21) to the norm of |ψ〉 was |a1|2+|a2|2. For a consistent probabilistic interpretation of the coefficients, we need that to be the contribution of the first term in the projected state. Thus, we define the normalized state (23)|ψ′〉=|a1|2+|a2|2a1+a2|a1+a2||1〉A|β〉+a3|1〉A|3〉B+|a4|2+|a5|2a4+a5|a4+a5||2〉A|β〉+a6|2〉A|3〉B. We are interested in how this affects the reduced density operator ρA=trB|ψ〉〈ψ|, whose matrix elements are defined by ρA=∑ijρAij|i〉A〈j|A. Represented as a matrix in the given basis, the original reduced density operator is (24)ρA=|a1|2+|a2|2+|a3|2a1a4*+a2a5*+a3a6*a1*a4+a2*a5+a3*a6|a4|2+|a5|2+|a6|2. Under the action of the equivalence relation, we have a modified reduced density operator (25)ρA′=trB′|ψ′〉〈ψ′|=〈β|ψ′〉〈ψ′|β〉+〈3|ψ′〉〈ψ′|3〉. Represented as a matrix, the diagonal elements (and thus the trace) are unchanged, but the upper off-diagonal element is now (26)ρA12′=a3a6*+(|a1|2+|a2|2)(|a4|2+|a5|2)(a1+a2)(a4*+a5*)|a1+a2||a4+a5|. The other off-diagonal element is the complex conjugate since ρA′ is Hermitian. The modification of the off-diagonals affects the eigenvalues and hence the von Neumann entropy of ρA′, which we identify as the predictive complexity CA. The explicit expression for CA in terms of the ai is long and not illuminating, so we omit it. To give a concrete example of how the predictive complexity will differ from the entanglement entropy, suppose that we have a state that is a Bell state involving entanglement between subsystem A and the states |1〉B and |2〉B:(27)|ψ〉=12(|11〉+|22〉). This corresponds to a1=a5=1/2 with the rest of the ai equal to zero. The reduced density operator, Equation (24), is diagonal, and the entanglement entropy is clearly SA=log2. Under the equivalence relation, |1〉B and |2〉B are in the same predictive state, and according to Equation (26), the reduced density operator ρA′ picks up off-diagonal elements equal to a1a5*=1/2. This changes the eigenvalues to be 1 and 0, and so we obtain CA=0. This is consistent with the idea that in the case |1〉B∼|2〉B, the state |ψ′〉=(1/2)(|1〉A|β〉+|2〉|β〉) factorizes and the entanglement disappears. We have worked this out in a low-dimensional case, but the extension to an arbitrary finite-dimensional Hilbert space is relatively straightforward. In particular, in the following section, we will apply it to the N=32 site Heisenberg model. 3. Results: Predictive Complexity of the Heisenberg Model 3.1. The Heisenberg Model It is interesting to apply these concepts to a quantum system with highly non-trivial spacetime dynamics, but that is nonetheless exactly solvable. We consider a standard spin chain from condensed matter theory, the 1D isotropic (XXX) Heisenberg model on N=32 sites, with ferromagnetic, nearest neighbor interactions and periodic boundary conditions. The Hamiltonian was given above in Equation (12). The value of J does not affect our analysis, so we set J=1 (equivalently, we choose time t to be in units of ℏ/J). Analyzing the Hilbert space of this model is greatly facilitated by the Bethe ansatz for the energy eigenstates, and we will largely follow the very helpful review article [44]. Within the two-magnon sector, we have calculated the complete time-dependent wavefunction and so can compute the entanglement entropies and predictive complexities exactly. This could also be accomplished by direct diagonalization, but we found the Bethe ansatz helpful for distinguishing contributions from scattering states and bound states. The full Hilbert space has the dimension 232 and has an orthogonal basis of position states |σ1…σN〉 and a position basis, with σn=↑ or ↓ representing spin up or down, respectively, in the z direction at site n. These position states are generally not eigenstates of the Hamiltonian, and hence evolve non-trivially with time. We will restrict our attention to the two-magnon sector, a subspace of dimension 322=496, with an orthonormal positional basis consisting of states of the form (28)|n1,n2〉=|↑↑⋯↓⋯↓⋯↑↑〉,n1n2 where 1≤n1