WorksPhysics and Astronomy
Conservation, crossing symmetry, and completeness in diagrammatic theories
Abstract
The diagrammatic analysis of interacting particle assemblies harbors a fundamental mismatch between two of its main implementations: Phi-derivable (conserving) approximations and parquet (crossing symmetric) models. No termwise expansion, short of the exact theory itself, can be both conserving and crossing symmetric. This work applies the Kraichnan embedded-Hamiltonian formalism for strongly coupled systems to investigate consistency of the interplay between purely pair-mediated correlations and pair-irreducible ones. The approach sheds a different light on the issue of crossing symmetry versus conservation. In the process, the parquet equations acquire a different formulation.
Introduction
A puzzling characteristic of many-body perturbation expansions concerns an inherent constraint on their practical applicability to interacting systems [1, 2, 3]. It is not possible for any approximate series to incorporate both of the two fundamental attributes of the underlying exact description: microscopic conservation, and the dominance of particle statistics in the correlated state as implemented in the exchange, or crossing, symmetry of the exact two-body scattering amplitude.
Only the full diagrammatic expansion can satisfy both conservation and crossing symmetry. The impossibility for any subseries to do so was discussed initially by Becker and Grosser for nuclear scattering [1], then more generally by Jackson and Smith [2] and finally in great detail by Smith [3], who analyzed the parquet-like structure of -derivable two-body scattering amplitudes in the context of the failure of crossing symmetry.
That limitation leaves just one of two mutually exclusive options for calculation. Either one selects an interacting model that conserves microscopically and globally but lacks crossing symmetry; or else, say in the specific case of fermion scattering, one sets up a model that respects antisymmetry and is richer in correlations, but nonconserving.
The incompatibility of crossing symmetry and conservation in diagrammatics shows itself as a mismatch in the approximate treatment of the single-particle self-energy. If derived variationally from a model Luttinger-Ward (LW) functional [4, 5, 6], then even if the two-body scattering kernel defining the LW object is crossing symmetric, the two-body scattering kernel subsequently derived from the associated self-energy will carry additional noncrossing symmetric terms that must be retained for the model’s transport and dynamic response descriptions to be fully conserving. The origin of the mismatch lies in an ambiguity in defining the truncated LW kernel, unlike the exact object which is evidently unique.
On the other side, if one solves the appropriate parquet equations self-consistently [7, 8, 9, 10], a manifestly crossing symmetric two-body scattering kernel is obtained which then defines an associated self-energy. Even as it exhibits more structure than its closest -derivable analog, just as for the latter the parquet self-energy also generates, variationally, new non-crossing symmetric additions to its originating two-body scattering amplitude. These must be discarded to keep crossing symmetry, but since within the model they represent physically consistent scattering effects, removing them takes away the variational consistency needed for conservation.
Our goal is to make more transparent what it is about diagrammatics that underlies the seemingly odd fact, at least for fermions, by which conservation and the natural outcome of particle statistics cannot coexist in any rational approximation. We do not attempt to overcome a demonstrably intractable limitation. Rather, we offer a conceptually different insight into a many-particle system’s graphical constitution and a different outlook on the character of both derivability and parquet theory.
Our instrument is the Kraichnan formalism [11, 12], whose two-step construction leads to an extended Hamiltonian description for a basic class of -derivable approximations, performing all-order diagram summations abstracted in a strictly conserving fashion from the exact system expansion. By preserving the Hamiltonian’s unitarity, the approach also safeguards the structure of all sum rules based on causality. It has been applied to particular cases [13] to establish the general basis for their satisfaction of the full set of conservation relations: for example, the dynamical sum-rule constraints on the density-density and related correlation functions [14, 15]. Identities dependent upon the completeness of Fock space are not preserved and, as will be shown, the breakdown of crossing symmetry in derivability is intimately related to this feature.
To compare parquet and derivability we start by applying Kraichnan’s Hamiltonian embedding to a familiar subtype of selfconsistent correlations: those mediated only by the three possible avenues for purely two-body propagation between successive interactions. They are: particle-particle and hole-hole ladders ( channel) and particle-hole polarization bubbles ( channel) with the latter’s exchanges, the particle-hole ladders ( channel) [16]. This exclusively pair-wise “” description of excitations, here couched in very different terms, is identical computationally to the -derivable theory of Scalapino and co-authors known as the fluctuation exchange, or FLEX, model [17]. See also the review by Bickers [10].
Section II briefly revisits the theory of the Luttinger-Ward correlation energy functional [4]. Two standard representations of the LW functional are introduced, both parametrized by the interaction potential and, in one case, by the renormalized one-particle Green function [4] while the other uses the renormalized two-particle Green function [18, 19]. There follows an overview of Kraichnan’s construction for the stu-FLEX model, indicating significant features derived from the formalism. In Sec. III irreducibility in the exact expansion of the LW functional is discussed (here “irreducible” denotes a diagram not generated within the stu framework). The exact parquet equations are given an alternative derivation, from which an alternative interpretation to the standard one also follows, where Kraichnan again provides the tool. Section IV analyzes the loss of completeness and failure of crossing symmetry in terms of how Kraichnan stochastics operates. There follows a related phenomenon in the two-body description of the LW functional: the breakdown of the identity relating static and dynamic structure factors.
We close with Sec. V summarizing our conclusions and several implications for future study. Appendices A and B recall the accounting numerics, first for diagrams with less than maximum topological symmetry and second for the Kraichnan average of closed diagrams. Appendix C covers identification of the primitively irreducible two-body scattering kernel.
Essentials
Luttinger-Ward functional
We recapitulate the Kraichnan formalism’s objective: the structure of the interacting ground state. To simplify the discussion we take a spatially homogeneous system in the momentum-spin representation at zero temperature, although the formalism works for any two-body interaction for uniform and nonuniform cases, at finite temperature or not [11, 12].
The standard system Hamiltonian comprises the one-body kinetic-energy part determined by the reference basis states. The kinetic energy does not play a direct role in the correlation analysis and is not considered further. We focus on the two-body interaction operator [15]:
The sum over states has the momentum and spin conservation restriction . We absorb a factor of inverse system volume into , the interaction potential [15]. Indices , etc. may address a single-particle phase space in more than one dimension.
Central to the development of conserving approximations and certainly to -derivable models possessing an explicit Hamiltonian, is the Luttinger-Ward functional: the component of the ground state energy (generally, the free energy) manifesting, and in one sense generating, its full interacting structure [4]. The LW functional provides the correlation energy in the ground state , expressed as a Hellmann-Feynman coupling constant integral:
There are two distinct but equivalent ways to describe the correlations in the LW functional. The approach closer to the analysis of Luttinger and Ward, extended to conserving approximations by Kadanoff and Baym [5, 6], addresses in terms of the fully renormalized one-body propagator and self-energy. Closer to parquet analysis in stressing two-body processes [7, 8, 9] is the theory of the pair correlation function in its static and dynamic forms; more particularly, its Fourier transform, the structure factor.
Pair-correlation description of
The following is based on Pines and Nozières [14]. In a closed interacting system the static structure factor measures the instantaneous correlation between a pair of constituent particles. Its formal definition is
in which generalizes the particle number operator [14] and is the total particle number. The LW functional has expression as a two-body object:
here we emphasize the implicit functional dependence of on the interaction within the Hellmann-Feynman formula.
In the exact problem the static structure factor is also expressible as the inverse Fourier transform, at equal times, of the particle-number autocorrelation in the frequency domain. As the response to a weak perturbation term added to the Hamiltonian, the dynamic structure factor is determined by the departure from the unperturbed state:
satisfying the identity [14]
Insofar as it holds for the exact case, the importance of (6) as one nexus between the static form (3) and the dynamic form (5), will become apparent in the context of a -derivable approximation’s departure from the exact expansion. In conserving approximations the consistent interpretation of the two structure factors needs care, and we postpone this to Sec. IV. Their mismatch could act as a numerical fidelity check on such models, but its true significance is conceptual.
We go on to specify the LW functional in terms of the single-particle Green function, or propagator, and its associated self-energy . We follow Kadanoff and Baym [5, 6] and the original paper of Luttinger and Ward [4].
One-body description of
The original form of the exact Luttinger-Ward functional [4] is also a coupling-constant integral, expressed in terms of renormalized one-body quantities:
we explore the second right-hand-side expression shortly. Each dot on the right-hand side of Eq. (7) denotes an internal summation over momentum-energy and spin.
As usual the Dyson equation defines the one-body propagator determining the LW functional:
In the momentum-energy representation the noninteracting propagator is with the Fermi energy. The Dyson equation is inherently selfconsistent since the self-energy is the variation of with respect to :
Note that this variation is restricted. It treats Eq. (7) as a functional of , not of the full physical one-body propagator . Diagrammatically it is equivalent to the full variation for the dressed Luttinger-Ward functional introduced in Eq. (36) of Appendix A. Unless noted otherwise, variations with respect to in the main text following are understood to conform to its use in Eq. (9).
Although the abstract properties of the correlation energy functional have a nonperturbative development, either in terms of and [19] or of and [20, 21], Eqs. (7)–(9) would remain a computational tautology without some understanding of the LW functional’s diagrammatic provenance. For this, we recapitulate the graphical content of its exact structure [4, 22].
(A) The object is the collection of all closed Feynman skeleton diagrams beyond Hartree, to every order in . A skeleton diagram cannot be cut into two disjoint parts by severing any one pair of internal lines . The coupling-constant integral Eq. (7) automatically generates the required combinatorial factors at every order of the interaction while the diagrammatic topology is fixed by the integrand .
(B) is invariant under particle pair exchange.
(C) The kernel is unique and microscopically reversible: and every internal line is renormalized selfconsistently with the same where
The two-body scattering kernel , introduced in the second line of the definition of , Eq. (7), is the central quantity in comparing approximations to the correlation structure with the ideal specification of . By nature of the exact state and for the exact state alone, carries every physically possible mode of interaction within the system. It leads to the fourth basic principle underlying the full LW functional:
(D) the exact ground-state description is self-contained. No new interaction configuration, not already manifest in the kernel itself, can arise from the second variation of the LW functional with . That is,
Item (A) does not apply in approximation since only a subset of the LW functional’s complete diagrammatic content can ever be incorporated. While a -derivable model will satisfy Eq. (9) by construction [5] and is normally expected to satisfy (B)—a notable exception being the classic random-phase approximation [14]—it will not satisfy (D) other than trivially (Hartree-Fock).
The physical context for Kraichnan’s formalism has been set out. Now we recall its constitution.
Kraichnan theory
Basic conception
Figure 1 schematizes the two-step process arriving at an embedding of the physical Hamiltonian in an extended object that can be modified while preserving the Hermitian property of the original. Construction of this extended Hamiltonian is complicated. We set out its various constructive steps.

Figure 1. Construction of the Kraichnan Hamiltonian. (a) The exact Hamiltonian is embedded first in an arbitrarily large sum of identical but distinguishable copies, indexed by . A Fourier transform over the index generates a collective description. The interaction is augmented with a parameter transforming in its Fourier indices exactly as in its physical indices . (b) The collective Hamiltonian is next embedded in an arbitrarily large sum of topologically identical replicas, except that each now carries a unique set of factors . The extended -sized Hamiltonian remains Hermitian. Setting to unity recovers the exact physical expectations. When is tailored to be randomly assigned over the -fold ensemble of collective Hamiltonians, a selected subset of correlation diagrams is distinguished by their total product of coupling factors factoring out to unity. On averaging over the stochastic distribution, random phasing suppresses everything else. All canonical relationships valid for the underlying Hamiltonian remain valid in the reduced model.
Step 1A. Take a large number, , of identical but distinguishable copies of the Hamiltonian, Eq. (1). Label them with . The label could be thought of as a pseudospin, but we will consider it to satisfy periodicity just like the physical label on a real-space unit cell within a uniform lattice; that is, is identified with . We let tend to infinity in the limit.
Step 1B. A “collective” index is defined via a Fourier transform over , detailed in Eq. (13) following. This acts exactly like the momentum label that collectively encodes the individual unit cells in the real-space lattice.
Step 1C. The -fold collective Hamiltonian inherits the interaction . When this is not modified, expectations over the collective indices still recover the original properties of any member . However, a novel degree of freedom is introduced via and we exploit it by attaching to a new coupling factor . The symmetries of in its indices are identical to that of on its labels. The modified collective Hamiltonian stays Hermitian, with real eigenvalues (these must differ from the exact case except when is unity).
Step 2A. The -fold collective Hamiltonian, carrying its own , is a well formed entity in its own right. We now generate a new -fold set of such replicas for large, and sum them to create a far bigger Hamiltonian. Here, in contrast with the first step, each Hamiltonian member comes with a particular collection of which differs from one replica to the next in the superassembly. Nevertheless the total “Hamiltonian of Hamiltonians” is still Hermitian.
Step 2B. For each quadruple of collective indices the values for the coupling are then defined as functions of randomly distributed parameters. See Eq. (14) and Figure 2. We are at the heart of Kraichnan’s procedure. The functional form of the couplings is specifically tailored so that, taking a double expectation (or, Kraichnan average) over the indices and the stochastic distributions of the parameters, only certain subsets of terms are sure to survive while all else drops out by destructive interference among random coefficients when these do not cancel mutually to yield an overall

Figure 2. Ambiguity in interpreting the second-order scattering term in the self-energy. (a) Basic contribution; (b)–(d): each channel generates a formally diverse but physically identical representation which, uncompensated in Eq. (12), would introduce overcounting when the second variation is taken for the Kraichnan LW functional. In the Hamiltonian the coupling , in place of just , inhibits redundancy while allowing free recursive interplay of every pairing process. The diagrams above provide the simplest instance of structural degeneracy in a model where the nominal kernel may be equally defined to be -, or -, or -like.
product of unity. (Details of the process are in Sec. IIF and Appendix B below, notably how the constraint on the indices operates just like conservation of momentum.)
Since the Hermitian structure is preserved in taking expectations over the superassembly, the fundamental identities between expectation values that depend on hermiticity still hold after Kraichnan averaging. This guarantees, first and foremost, that the conservation laws are satisfied by the approximated dynamical quantities such as particle number, flux, momentum, and energy.
Formalism
We first (anti)symmetrize the elementary interaction:
where is the species flag, for fermions and for bosons. From now on we deal explicitly with fermions. Following the process outlined above in Step 2B, the interaction piece of a Kraichnan collective Hamiltonian, in which the generic operator from Eq. (1) is embedded, is distinguished by the assigned set of couplings in their -fold distribution.
These Kraichnan couplings (K couplings hereafter) may take a variety of configurations provided their symmetry on index permutation is identical to that of the microscopic potential in its physical indices, as in Step 1C. Note that if is an admissible K coupling then its complement is also admissible. We will develop this idea in Sec. III in a different analysis of the exact diagrammatic expansion.
(12) is interpreted as follows:
The restriction on the sum now connotes conservation of the Kraichnan collective indices, as well as the physical conservation constraint on the momenta and spins, . We have conflated index and momentum-spin labels into the single form , so [23].
The collective creation and annihilation operators and are defined by Fourier sums over the distinguishable system copies, each with its set of operators and :
It can be shown that these collective operators satisfy the anticommutation relations and .
(3) The Kraichnan coupling comprises the , and channels:
This prohibits overcounting of coincident closed graphs in . The need for it is shown in Figure 2 in the context of allowing free interplay of the three channels without duplication of physically indistinguishable terms. The channel has a label exchange relative to the definition of its physical exchange counterpart, the channel. Label exchange leads to while always remains exchange symmetric.
(4) A key identity in deriving the -derivable response within Kraichnan’s approach is
(5) Finally, for the second embedding in the Kraichnan prescription, Figure 1(b), the phase parameters , and
each become elements of a uniformly random distribution of size in the limit of large .
The outcome of averaging stochastically over the distribution of for a skeleton graph in the LW functional is described in Appendix B. Channels , and are the sole possibilities for particle-particle and particle-antiparticle pair excitations. In terms of a system described by pair interactions, this means that Kraichnan’s construct is the most general approximation, based on an explicit Hamiltonian, that encompasses all possible pairwise modes.
As mentioned, the K coupling provides a new degree of freedom that lets one preselect how the creation-annihilation operators bind to the elementary interaction. Effectively, this fixes the possible causal orderings of the propagator pairs at the level of the extended Hamiltonian. For conventional parquet, the types of ordering within the pairing channels (particle-particle or particle-hole) are specified at the level of the equations, as the interaction itself is blind to temporal sequencing of the Green functions. In Kraichnan’s formalism the dynamical flow across the interaction is predetermined before the equations are derived from the structure of the extended Hamiltonian.
Reduction of by Kraichnan averaging
The Luttinger-Ward functional obtained from the Hamiltonian is
Prior to taking Kraichnan expectations, the topological content of Eq. (16) remains that of the exact Eq. (2), with the addition of the collective-index degree of freedom. Since has label symmetry, is exchange invariant. Therefore so will its Kraichnan average.
Expectations on both sides of Eq. (2), over the stochastic distribution of K couplings of Eq. (14), reduce the set of LW correlation terms to those in Figure 3(b) with interaction kernels defined in 3(a). We will denote such expectations by the subscript , in which case we define

Figure 3. (a) Definition of the primary all-order and interactions. Dark ovals: antisymmetrized potential ; linking lines are one-body propagators. Kraichnan couplings from Eq. (14), each selecting for its channel, are shown. In the channel to leading order, the full Hartree term appears with its Fock exchange; for and it is their superposition as an exchange pair that generates the full Hartree-Fock term. (b) Symbolic definition of , the LW correlation energy functional (combinatorial weightings [15] are understood) following the Kraichnan average over all K couplings according to Eq. (14). Subtraction of two second-order skeletons corrects for threefold overcounting in the and channels. While the skeleton graphs for appear simple, their complexity lies in the selfconsistent nesting of self-energy insertions in the one-body propagators. The topology is fully revealed only when the response to an external probe is extracted. Universality of the LW functional [20, 21] means that the topology of its constitutive diagrams in (b) is unaltered in moving the interaction from to when the Hamiltonian is itself invariant with respect to exchange. This does not affect the overall accounting.
To trace the fate of the K couplings in the upcoming variational analysis of the model, we consider the object although to reintroduce in it is redundant.
where we have used the single-particle form, Eq. (7). The renormalized propagator remains defined self-consistently by Dyson’s equation, Eq. (8), now with the reduced two-body generating kernel of Figure 3(a) that includes all allowed
and pairwise-only correlations, whose K coupling phases cancel identically to survive averaging.
The dressed skeletons making up yield a kernel that satisfies Eq. (10) automatically since stochastic averaging leaves that identity intact. Nevertheless, even though is well defined and crossing symmetric, the second functional derivative cannot satisfy condition (D), Eq. (11). Unavoidably, it generates new structures beyond . Only in the exact case (besides Hartree-Fock, the simplest model possible) is condition (D) true [2, 3].
The model’s violation of (D), forced by conservation, clearly differs from parquet which does not have a counterpart to the K couplings’ structural constraints that spoil crossing symmetry. The parquet equations do not constrain the connectivity of the channels, as does, and that leads to a richer family of intermediate pair scattering processes; but parquet’s greater complexity is not derived canonically from a LW functional. Parquet is capable of inferring one by invoking Eqs. (9) and (10) for its associated self-energy.
Despite the implicit appeal to a generating functional, the parquet self-energy as such still cannot offer a way to ensure crossing symmetry when the kernel is derived, violating condition (D) as for . Noncrossing terms still arise, which in parquet have to be discarded by force.
Kraichnan’s Hamiltonian secures all the exact causal-analytic identities for the reduced structures resulting from stochastic averaging. Every identity depending on unitarity must hold for each effective Hamiltonian , as for the exact one. Provided the K coupling average is done consistently on each side of such an identity, it follows that . Relations that depend directly on the completeness of Fock space do not survive owing to the random-phase induced decoherence.
Bethe-Salpeter Equation
The broad difference between two-body processes in derivability and parquet is that the first yields a Bethe-Salpeter equation [15] for the particle-hole vertex of an excitation away from equilibrium in a strictly closed system, where particle emission cannot occur. In parquet, the adopted extension of Bethe-Salpeter does not make a particular distinction among particle-particle or particle-antiparticle scattering processes. The first description has physical constraints not necessarily applicable within the second account.
To arrive at the parquet equations’ -derivable analogs we start by probing the system with an external, formally non-local, potential coupling physically to each member of the Kraichnan ensemble. It does not couple to the abstract collective indices defined over the ensemble.
Two-body correlations unfold within their Green function like an expanding concertina. Like a concertina, their intrinsic topology does not change from its inner assignment in . While the physical momentum-energy flow from the outside brings out the internal dynamical structure, the characteristic set of indexed K couplings is unchanged. What now changes in every two-body contribution is that, in each of an infinity of recursions, a set number of G lines is singled out by the perturbation node that each carries. This is Baym and Kadanoff’s philosophy [5, 6] in Kraichnan terms.
In summary, the insertion of perturbation nodes has no effect on the assignment of collective indices and thus on the combinatorics of the K couplings. The physical effect is only on energy-momentum transfer (also spin etc.).
The one-body perturbation augments the interaction Hamiltonian, Eq. (12):
Response to a local field is generated by setting , dynamically linking (contracting) the particle-hole propagators that terminate and start at .
Next we take physical expectations only, retaining the explicit K couplings to keep track of all potential pair processes before Kraichnan averaging. We sketch the logic [16]. The two-body Green function is [5]; note, quite generally, that adds linearly to so so .
Varying both sides of the Dyson equation (7) produces
where we define via , being accompanied by an effective K coupling as for Figure 4. The functional equation remains completely general, applicable to any suitable choice of Kraichnan coupling including the exact case .

Figure 4. Index convention for a Kraichnan vertex, associating with it the nominal K coupling . A response term results when lines are attached left and right and join at a perturbation node. The diagram contributes if and only if the internal sum of coupling phases cancels on connection to the effective K coupling as in Eq. (15). This construct manifests the same closed topology previously implicit in the unitary structure of .
In shorthand, with the two-point identity matrix, the Neumann series for becomes
Recalling Figure 3(a), the form of the generating kernel for purely correlations, namely for the reduced LW functional , can be read off in terms of the subsidiary kernels and for each channel. To put all interactions on the same K coupling-free footing as we write for the latter’s coupling and factor it out; refer also to Figure 4. Any surviving chain of K couplings, whose phases cancel right across, finally merges constructively with as exemplified in Figure 3(b). In the response description the unpaired outermost indices will be contracted when the terminating lines link to the perturbation nodes, since only for equal indices across a node is there a nontrivial Kraichnan expectation. (Figures 6
and 7 below hold more details.) Then
where
The different accounting for , which applies likewise to the kernel equations to follow, is because the channel incorporates its own ladder exchange while and are each other’s distinct exchanges; only if summed would they carry the full Hartree-Fock interaction . The last three relations in Eq. (20) express the content of Figure 3(a).
Conservation in derivability, and its cost
We face the inevitable outcome of every -derivable model: unlike Item (D) for the exact theory, the response kernel carries new terms beyond . The process is shown graphically in Figure 5. These apparently extraneous vertices are essential to consistency in conservation for the two-body Green function , but are no longer crossing symmetric and therefore cannot contribute to the parquet scattering amplitude as conventionally understood [3, 10].

Figure 5. Systematic removal of a propagator internal to the self-energy after Kadanoff and Baym [5, 6]. This generates the primary scattering kernel . Removing , solid line, simply regenerates . Removing any internal lines (dotted) other than $G(32) yields the additional vertices required by microscopic conservation. (a) Beyond the -channel ladder the noncrossing symmetric -like term and -term are generated. (b) Generation of and the nonsymmetric . (c) Generation of with and . No diagrammatic structure emerges that is not already incorporated recursively in the propagators within . Note that starts at third order in .
From Eq. (19) the complete four-point kernel is defined:
Consequently the conserving two-body Green function is
Momentum transfer in the above is determined by all the K couplings attached to each interaction, not by imposing the proper analytic form on intermediate pairs . This differs from the expansion of the standard parquet vertex [10], in which the mode of transfer has to be specified explicitly for the each of the three possible pairs .
The task is to show the equivalence between the system of (19) to (22) on the one hand and, on the other, the coupled parquet-like equations emerging from the Kraichnan formalism.
derivability in the Kraichnan representation
The K coupling constraints make a radical change to the physical content of the parquet equations’ -derivable form, in contrast with their conventional presentation. Before deriving them we clarify the one-to-one correspondence between the Baym-Kadanoff and Kraichnan interpretations of response.
In Figure 6 we illustrate how the standard Baym-Kadanoff derivation implements the total system response as the sum

Figure 6. Recursive construction of response kernel within derivability. Dots: external perturbation nodes. (a) Two contributions A and B to the particle-hole response function combine into (b), a new contribution. Graphs (c) and (d) show a more complex combination with a third allowed contribution C. In the response description Eq. (18), the system cannot tell a direct perturbation from one that is induced, so a perturbation node may be freely replaced with an induced perturbation. Fusion of the response terms produces a new contribution to the total. In the process the internal topology of the resultant response, virtual within the renormalization of , becomes manifest. All the kernel parts in Figure 5 are recursively convolved in this way.
of the direct response to the external perturbation and all the nonequilibrium fluctuations induced by it within the correlated system. Since the component particles cannot distinguish between direct and induced disturbances, the response to every such dynamical stimulus is highly selfconsistent. It follows that the primitive components of the response become convolved, leading to the systematic cascade of contributions implemented in Eq. (18).
Construction of the response within Kraichnan’s canonical formulation is expressed differently but describes the very same processes as the more heuristic -derivable description. The rules, analogous to physical conservation, are (1) conservation of incoming and outgoing index sums across any pair-scattering amplitude (alongside conservation of momentum) and (ii) conservation of index across any perturbation node. Figure 7 replicates Figure 6 in these terms. An alternative criterion for a candidate response term is whether it can be recollapsed to a diagram of .

Figure 7. Recursion of the response kernel in the Kraichnan approach, functionally equivalent to Figure 6. Terms (a) and (c) go to (b) and (d) respectively; their fusion is now dictated by global index-sum conservation. In (a), terms A and B survive Kraichnan averaging separately. Since their internal phases are unaffected by their fusion to form (b), the new term trivially survives averaging; likewise the case of (c) going to (d). The outcome maps exactly onto the -derivable structure of Figure 6, replicating the topology of all contributions to the response .
A different way to state the above is that the phase structure of a Kraichnan-embedded diagram within the LW functional, not just skeletal but recursively defined, cannot change under an external stimulus. While the input of external energy and momentum reveals the distinctly dynamical correlations (otherwise implicit at equilibrium), only those correlations with the same K coupling topology, showing overall phase cancellation, will survive as response contributions in the form of Figure 6 and Figure 7.
With respect to Figure 7 in particular it is noteworthy that, for a convolution of two independently closed response diagrams within a composite term, the overall phase cancellation for the interactions of one component is completely disjoint from the corresponding configuration in the other component. Indeed, the disjoint parts could each have a completely different coupling scheme without affecting the accounting of its complement. This separability will be exploited in the following Section. Figure 7 shows how phase-sum decoupling is induced by the conservation rules for the collective indices.
Minimal parquet: version
Within the approximation, Eq. (22) contains all conserving dynamical processes involving two-body propagation between interactions. Its form, however, obscures the distinct action of the three channels and , as evidenced in classic parquet. To recast Eq. (22) and reveal its parquet-like architecture, we review the derivation by K coupling analysis.
To the bare potential as the primary irreducible starting term, the extra noncrossing symmetric term has to be added. This could not appear in normal parquet as it violates crossing symmetry for fermions. Were it crossing symmetric, then itself would be crossing symmetric as well as conserving and therefore exact [3]; but precisely because is conserving, is needed to construct the appropriate two-body Green function and system response. Let
Three auxiliary kernels are defined. Each is irreducible in the channel for which it is the input. Each carries only terms that do not propagate directly in the given channel. That is, the kernel’s terms can be split only by cutting pairs propagating in the other two channels.
The auxiliary kernels drive the equation for the complete vertex , fed into the two-body Green function in Eq. (22). When contracting this to obtain the response, stochastic averaging of product sequences of K couplings keeps those terms compatible with unitarity. As part of the seed kernel, will regenerate itself while , whose expansion starts at third order in , generates all the additional response terms need for conservation. The -, - and -irreducible auxiliary kernels are
which combine to yield the complete response kernel [24]:
Beyond partitioning the response kernel in terms of its (mutually) irreducible pairwise processes, there is a subtle procedural distinction between the Kraichnan perspective of Eqs. (24) and (25) and the Kadanoff-Baym one of Eqs. (21) and (22). The latter starts from an already approximate form for the LW functional and proceeds by tracking its self-consistently recursive structure, Figure 6. In the former approach everything stays exact prior to stochastic averaging whereas, in our derivation of the parquet equations, we have quietly neglected every contribution that is not pairwise linked in the pre-average collective Kraichnan LW functional. The a posteriori justification is that, in any case, stochastic averaging projects out only the purely pairwise correlations.
In -derivable models is not the main goal. The closure procedure in Figure 7, tied to the perturbation nodes, is regulated by the outermost K coupling . The presence of this overarching constraint says that the open links in the vertices of Eq. (25) have a different role in derivability from their unconstrained analogs in standard parquet, with its S-matrix view.
Crossing symmetry plays an indirect role in response, shaping the four-point kernel in the abstract. There is consistency with Pauli exclusion but no direct connection to the response obtained from the general two-body Green function. Probing a closed system involves directly exciting particle-hole pairs rather than particle-particle. A notable example of a noncrossing symmetric yet physically justified -derivable model is the random-phase approximation [14], with exchange having no part in an essentially long-wavelength description.
There is a significant point to make about the compensating function of the K couplings on the right-hand side of Eq. (25). Standard parquet has no counterpart to . As Figure 2 shows, however, in any reconstruction of the LW functional starting from a kernel, the three different pairing channels result in the same contribution to and thus . The K coupling within is defined to prevent such overcounting. Therefore recovering a model LW functional, by working up from the parquet equations, means introducing by hand a subtraction that would remove the threefold redundancy in .
Irreducibility and exact parquet
Kraichnan’s formalism provides a systematic procedure, a stochastic algorithm, to isolate every irreducible contribution to the exact Luttinger-Ward functional that has no description in purely pairwise terms. These will convolve naturally with the strictly correlations to recover the canonical LW description and lead to a different formulation of the exact parquet equations.
Recall that, prior to taking K coupling expectations, the extended Kraichnan Hamiltonian retains the functional structure of its original. The only distinction between the pair-only Luttinger-Ward and the irreducible complement, call it , lies in how the latter’s K couplings are defined.
The correlation structure beyond must be governed by a Kraichnan coupling complementary to , or
where , and .
Contributions to the corresponding LW functional include all the -irreducible terms to all orders in beyond the leading Hartree-Fock diagrams. With no loss of generality can be recast. To a K coupling associate the K coupling , which from Eq. (15) is identically unity. Writing lets one conveniently treat both K couplings and “anticouplings” uniformly. Manifestly, the indices of will pass unchanged across the associated interaction, as if the latter were absent from the Kraichnan point of view.
In constructing we note that its kernel, call it , plays the role of in but now selects all possible non-pairwise-connected skeletons from the exact LW functional; since retains exchange invariance, a crossing symmetric exists.
The complementary Hamiltonian is
On K-averaging this leads to the definition of the complementary LW functional
where, as in Eq. (17), the Kraichnan-averaged propagator is renormalized in keeping with the reduced diagrammatics of this particular model. The counterparts to Eqs. (21) and (22) apply, with the residual kernel no longer crossing symmetric, as for earlier.
Interplay of and complementary sectors
In Figure 8 we show representative diagrams of the full Luttinger-Ward functional, both -reducible and -

Figure 8. Comparison of reducible and irreducible skeletons in the expansion of the Luttinger-Ward functional. (a) Closed diagrams comprise those of Figure 3(b): every node, dark dots, is a K-coupled interaction . (b) Non- term at fourth order in the complementary node as light dots. (c) Fifth-order non- skeleton, (d) fifth-order primitive irreducible that, unlike (b), (c) and (e) cannot be split by cutting two line pairs (other than by trivial removal of a node). (e) Non- two-pair reducibles at sixth order; (f) primitively irreducible sixth-order graph.
irredudible. Here the three channels and from Figure 3 are conflated and characterized by the single node which freely selects all possible modes for exclusively pairwise transfer of conserved quantities. The -irreducibles, by definition, cannot be represented in this way; their extended coupling is set by the Kraichnan-modified interaction .
Unlike Figure 8(d) and (Figure 8f), not all non- diagrams appear to fulfill the requirement (C) for symmetry, as met both by the exact four-point kernel and the reduced series, recast as Figure 8(a). The lack of full symmetry among propagators is seen in Figure 8(b), (Figure 8c) and (Figure 8e). Resolution of the apparent violation of requirement (C) for some non- graphs is found in Appendix A, which discusses the appropriate numerical weighting in evaluating . The claim in [16] that all closed diagrams should display perfect equivalence of its single-particle lines, is too restrictive.
From Eq. (26) it follows that and are mutually exclusive. Take any skeleton of order chosen from those in Figure 8(a) but with replacing each . The product of complementary K couplings becomes
The Kraichnan average is taken for each product of factors on the right-hand side of Eq. (28). At every occurrence of , the indices do not change in crossing the interaction and the total number of free indices decreases by two. Therefore, as every freely indexed factor in Eq. (28) is replaced with its constrained partner , the same order- skeleton graph is in play but is averaged only for the analogous order- product of free while carrying the sign factor , where .
Combinatorially, the total of the identical skeletons contributing to the K-average over the assembly of terms on the right-hand side of Eq. (28) is
Hence excludes all skeletons. The result holds in the presence of an external perturbation , namely for terms as in Figure 7, so all diagrams for the two-body Green function are also excluded from its counterpart . The exclusion is mutual, since a complementary element could not be in the set without contradiction by the foregoing argument.
Exact expansion in the Kraichnan representation
The expectations over both types of K coupling, each allowed to act in its own right in the context of the structures of and , recover the exact form of because every closed nonpairwise graph suppressed by averaging in the original mode, is restored on averaging in the complementary mode but crucially now in the full presence of the pairing dynamics embodied in the generic Eq. (21). Write the exact Hamiltonian as
the exact LW functional will be independent of .
In the Kraichnan-based analysis of the exact correlation functional we retain the explicit structure of from Eq. (27) to act as the seed kernel for the expansion in terms of the pair channels. We then define the one-pair reducible complement
Since acts as a strong perturbation on , new hybrid terms are generated so the perturbative expansion of itself, while consisting of pairwise connected elements, will no longer be just from Eq. (17) and Figure 3. It will involve autonomous components convolved with autonomous non- ones, coupled via the same index conservation principle illustrated in Figure 7.
The exact Luttinger-Ward functional takes on the form
The propagator is determined by Eqs. (7) and (8) with the exact self-energy . The kernels and are coupled by their joint renormalization of .
Being exact, is complete to all orders in . It supports every permissible skeleton diagram, irreducible or not. Unlike a -derivable approximation, no new two-body vertex can be generated that does not already appear in
The sum is naturally conserving while would only be so autonomously with its own propagator fixed from Eq. (27). However, is not independent of and so cannot stand as an autonomous conserving kernel although is well defined; but now the exact nature of in Eq. (31) implies
meaning, with exact, that mutual coupling of the two species of vertex terms contrives to restructure the residuals such that what were previously distinct noncrossing symmetric components are now absorbed into in crossing symmetric partnership with their counterparts.
Kraichnan parquet for the exact ground state
Having recast the exact Hamiltonian in the Kraichnan embedding as a strictly pair-linked plus a complementary part, we have split the LW functional into the irreducible so must then exhibit, on an equal footing, all three possible modes for pair-reducible composite excitations. One can account explicitly for all permissible topologies, in all combinations occurring in the exact kernel , by returning to Eqs. (24) and (25). Enlarge Eq. (23) to
there are now no residuals. As with earlier, we retain the (otherwise redundant) anticoupling in to provide an audit trail for the action of both and in the following.
Equation (32) includes all terms in the complete kernel that are not pair reducible overall. Nevertheless, examination of Figure 8(b), (Figure 8c) and (Figure 8e) tells us that some diagrams in the -irreducible expansion of are similar to the purely series of Figure 8(a) in having components separable by cutting two pairs of propagators. Thus the expansion of , while excluding structures in the global sense, will still include convolutions that are pair reducible. One could reverse engineer the structure of the Kraichnan to identify its primitively irreducible terms, such as in Figure 8(d) and (Figure 8f). While not needed here, the procedure to do so is in Appendix C.
The strategy for obtaining the complete now follows the same methodology [10] of sorting out the individual pairing channels as in Eqs. (24) and (25). The extended Kraichnan parquet equations for the exact ground state are
so
This four-point representation is not for any two-body collision process whatsoever, but is tied to the particle-hole dynamics dictating the system response to an external, number-preserving probe. The overall outer coupling still has to be carried on the right-hand side of Eq. (33) as the essential bookkeeping device to that end. By the principle of Figure 7, the autonomous scattering elements linked by pairs in the equations above have cycles of K couplings or anticouplings that factor out to unity independently.
Standard parquet and the exact ground state
We now address the formal distinction between (a), the rather different specification of parquet via the exact Eqns. (33) and (34), and (b) the standard parquet version in which K couplings and anticouplings do not exist and the distinction among channels is made purely through the particular momentum-flow combinations of the three possible pairings in [12]; no further knowledge is adduced here to establish which irreducibles are represented in . Unlike in the embedded-Hamiltonian approach, there is no recipe to identify explicitly the content of the (conserving) subseries and, importantly, that of its complement.
For case (a) we recall how the (relative) causal ordering of the propagators is chosen automatically. (12) for the Hamiltonian fixes ab initio, through the definition of in Eq. (14), how the creation-annihilation operators are to couple. For (b) this is done by imposing the three possible dynamical flows on the form of the parquet equations. In Kraichnan the parquet structure emerges more naturally.
The exact Hamiltonian and are independent of the K couplings. There are two obvious choices for .
Choosing means . Then and vanish so . Since access to the innermost structure of is unavailable, Eqs. (33) and (34), though exact, hide the essential contributions of the irreducible versus the strictly pairwise correlations. No insight is gained.
The choice forces and now and are zero in Eq. (30). Then becomes , carrying everything for the exact problem while in Eq. (34) goes to which vanishes at order . Once more there is no gain.
It would be inconsistent to set in Eq. (33) ignoring the anticorrelation in actuated through the Kraichnan phase average. The invariance of the exact Hamiltonian, Eq. (29), would be violated. Choosing and as unity in Eq. (29) simply doubles the interaction strength, which is not equivalent to putting in Eqs. (33) and (34) since the interaction there no longer matches in the rescaled Hamiltonian.
As already pointed out, if an exact expansion of were to dispense with a classification such as Kraichnan coupling, an extra intervention would be needed to avoid redundant graphs in the progression from up to . Moreover, without a way to discriminate systematically between pairing and irreducible sectors, the irreducibles in must be picked out basically by inspection.
Rationale for the Kraichnan approach
In the exact case the question is: doesn’t the Kraichnan construction merely replicate, by rather more convoluted reasoning, what standard parquet already conveys? The answer is no, and it comes in two parts. The first concerns the different way in which each triad treatment couples the pairing channels in its analog to , Eq. (34). The second is the fate of the unavoidable truncations of the exact expansion within the respective parquet accounts.
Figure 9 illustrates the issue for pairing-channel diagrams typically met in conventional parquet, alongside their exchanges which are common to it and . Although legitimate Feynman terms [8], Figure 9(a) and (Figure 9c) are excluded from the -derivable model. For Figure 9(a) the total K coupling phase over a pair of channel interactions is subject to -like exchange indexing. Following Eq. (14) it becomes

Figure 9. Composite scattering terms presented within their response contributions. (a) -channel process, mediated by the screened interaction from Figure 3(a) and inadmissible as a term, unlike its direct counterpart (b), corresponding to in Figure 4(b). (c) -channel ladder again mediated by , also inadmissible, unlike (d). Figures (a) and (c) are admissible when replaces , as the first survives in pure -channel mode and the second in pure -channel mode. At higher orders in , the -coupling phases associated with cancel only in negligible ranges of their indices in the limit of infinite system copies. When mediated by their anticouplings, however, such non- diagrams survive and appear in the complementary series generated by .
and vanishes only for the asymptotically negligible ranges or . At the same time the phases for Figure 9(b), replicating from Figure 4(b), cancel identically so is admissible. A similar situation holds for Figure 9(c) vis à vis the term Figure 9(d): the latter is admissible, its exchange is not.
In contrast with the above, the standard parquet summation convention for pairing channels in the kernel takes the contents of Eq. (33) with no coupling factors. Consequently it is blind to the source and destination of the pairs on either side of every interaction .
Standard parquet assigns diagrams to the pairing channels of its analog that cannot appear in the corresponding Kraichnan object. The exact expansion executed in standard parquet analysis requires a version of the irreducible that is a subset of the skeletons incorporated in the Kraichnan associated with Eq. (27), because, as with the examples in Figure 9, structures that count as pair reducible in the standard formulation are excluded in the formulation. The irreducible seed kernel required by parquet analysis corresponds to the one identified in Appendix C.
As long as one considers the exact ground state, there seems little to choose between the two versions of parquet, but that is not the case. In the version proposed here, the equations emerge from a Hamiltonian, informed by the constraints of unitarity as well as from a more natural partitioning of correlations into and complementary sectors. In their conventional form the parquet equations are inferred bottom-up; for the exact case their structure presumes the existence of a well defined set of irreducible kernel components. Nevertheless it lacks a systematic way to construct the full set. Passage to Kraichnan anticouplings provides an explicit selection algorithm, at least in principle.
The conceptual difference in codifying correlations has practical consequences. Since it is necessary to truncate the exact expansion for a viable calculation, in Kraichnan’s version of derivability one can do so by selecting a physically motivated subset of the irreducible complement . While such an essentially intuitive choice means loss of contact with a formal Hamiltonian and with crossing symmetry, the generic Luttinger-Ward structure of Eq. (30) persists to sustain selfconsistency and conservation; the Kraichnan analysis falls back to Baym and Kadanoff [5, 6]. The price of retaining crossing symmetry in the standard parquet equations, on the other hand, is loss of contact with unitarity [3].
It is worth revisiting a related aspect with regard to response. There, the tight coordination between self-energy corrections carried in , or “out-of-the-beam” scattering, and two-body scattering vertices, or “into-the-beam” scattering, is the crucial element in conservation when multi-pair excitations are involved [25]. The functional definition underpins the approximate but conserving Bethe-Salpeter equation. It is then essential for the extra noncrossing symmetric vertex terms to be kept in . If the parquet kernel is to remain crossing symmetric by discarding the incompatible terms that would come from variation of its own associated self-energy , the conserving nexus between and breaks down, compromising the response properties.
The inference from this Section is that while the diagrammatic structure of the ground state uniquely conditions the exact conserving parquet equations, the traditional parquet equations, set out heuristically, cannot by themselves automatically reconstitute the ground state, nor the excitations defining the physics of response. Additional correlational knowledge is needed.
Completeness, Uniqueness, Consistency
Completeness and crossing symmetry
We are ready to advance a Kraichnan-based clarification of the results of Refs. [1, 2, 3] foreclosing all chance of incorporating both crossing symmetry and conservation in any truncated description of an interacting system. For this we return to the basic definition that embeds the physical interaction Hamiltonian in the Kraichnan collective form renormalized by its K couplings:
in which, for clarity, we resolve the index into its physical and collective components and we leave unsymmetrized the elementary interaction.
Recall that the Luttinger-Ward functional from Eq. (35) is exact for this specific Hamiltonian. The corresponding kernel is unique, conserving and crossing symmetric. What happens, then, in the average over the stochastically defined K couplings? Consider the collective Fock space of a typical member of the -sized superensemble, with distribution and Hamiltonian .
When we have a direct sum of physically identical but distinguishable Fock spaces arranged, as it were, as block diagonals for which the collective description indexed by has no operative role, and the expectation for over the discrete blocks is exact. As soon as the K coupling function takes a nontrivial form, there is cross-linking among the Fock-space copies making up the collective set; correlations are induced across copies, off the block-diagonal and mediated by the indexed .
Any closed cross-copy interaction graph has a topology identifiable within the exact physical expansion, with the addition of the spin-like indices . It must have an exchange counterpart also diagrammatically identical to the physical exchange. Thus crossing symmetry applies to the pair. The presence of does not in itself destroy formal crossing symmetry in , but the exchange of indices scrambles the overall K coupling phase in a very different fashion from how physical exchange acts on the momentum transfer for .
In , composite correlations in the form of Figure 6 or Figure 7 numerically survive the stochastic phase average but are no longer guaranteed to have crossing symmetric exchange partners that also survive. Their exchanges, as with Figure 9(a) and (Figure 9c), will have zero Kraichnan phase only in an asymptotically negligible range over the index representation. Meanwhile the total Kraichnan phase for the allowed terms, as with Figure 9(b) and (Figure 9d), will be identically zero over the entire range of collective indices. The forward-scattering sum rule [26, 27], that is , is violated for the residuals even if not for . This is one instance of the loss of completeness (coherence) of the extended Fock space; another is discussed below for the structure factors.
Stochastic averaging kills the off-diagonal exchange correlations in the collective index description of the cross-coupled Fock-space blocks, thus also the (still formally present) crossing symmetry. However, conservation, including for the -derivable dynamical response structure, depends only on the properties of the surviving index-diagonal correlations. They are fixed by the recursive topology of the originating unitary Luttinger-Ward functional, invariant under perturbations.
Prior antisymmetrization of at the Hamiltonian level ameliorates the loss of exchange symmetry by making it explicit at the individual interaction level. This is not sufficient to secure global crossing symmetry. In general it is not local topology but that of entire sets of diagrams in superposition, that determines crossing symmetry for the complete assembly; a principle that also applies to conservation within its sector.
In terms of the extended Fock-space scenario, one might think of parquet theory as the partial importation of off-diagonal correlation structures in index space, namely the exchange complements, to act side by side with diagonal conserving ones. When done in a way that is uncontrolled from the Kraichnan perspective at any rate, it is not surprising that conservation is compromised.
Uniqueness and structural ambiguity
There remains the inevitable ambiguity in defining the effective kernel for a Kraichnan model LW functional. Antisymmetrization of the elementary potential makes it easier to construct a physically reasonable crossing symmetric seed kernel, but one could more crudely restructure the closed graphs of by directly antisymmetrizing . We have seen that the Kraichnan average must decohere crossing symmetric superpositions while maintaining conservation. The variation is then bound to produce additional terms distinct from and that cannot be crossing symmetric, else would be crossing symmetric and conserving, so itself would be unique, being exact.
As shown in Appendix A, residual terms are able in their own right to reconstitute by reclosing with two propagators and integrating the Hellmann-Feynman identity in Eq. (36). This in turn means that need not be unique as a generator for the LW functional. A concrete example for the model would be to sum appropriate terms in Figure 5, for instance:
from the -channel, Figure 5(a),
from the -channel, Figure 5(b), and
from the -channel, Figure 5(c).
In passing we note that there are identities between the extra kernel terms in Figure 5, not explored here, that render the outcome of bosonic Kraichnan different from the present fermionic one.
Two-body Consistency
We have explored the model in its Kraichnan setting through the single-particle formalism of Kadanoff and Baym
[5, 6] in the language of Luttinger and Ward [4]. To probe derivability from a perspective closer to the object of parquet theory, we switch the topic to two-particle processes as they fix the structure factors and the status of Eq. (6) in § 2. There we reviewed the role of the two-body Green function in structuring the exact Luttinger-Ward functional; here we take it up in the setting of a conserving approximation. This leads to reevaluation of the relation between static (more properly, equal-times) and dynamic response functions [14].
The primary definition of is given by Eq. (3), not Eq. (6) which is a consequence of Fock-space completeness. Here we recapitulate its perturbative relation to the exact kernel . From Eq. (36) in Appendix A we obtain the variational derivative, with the full physical propagator in contrast with Eq. (9):
Cancellation of terms results in
and the right-hand expression can be inserted into Eq. (3) to obtain as schematized in Figure 10(a). For a -derivable model with an approximate replacing , the proof follows through. Moreover, the static structure factor is real-valued [14, 28] since analyticity (the closed diagrams’ causal structure) is unaffected.

Figure 10. Two aspects of two-body diagrammatics, contrasted. (a) Generation of the static structure factor through removal of one interaction line from . In the model the original coupling , concomitant with , is unaffected. It stays convolved with the internal couplings through the collective indices. A Kraichnan averaging yields the primary kernel of Figure 3(a), defining from Eqs. (3) and (17). (b) Generation of the dynamic structure factor under an external perturbation. The collective-index rules allow for a larger set of terms in the response of the total kernel , Eq. (21). In the exact expansion and is identical to the frequency integral of in Eq. (6). In a -derivable approximation (b) does not replicate (a). In parquet theory, is iterated heuristically to redefine [10], bootstrapping the self-energy until convergence. This still cannot result in a conserving even if its (a) and (b) analogs were compatible.
At least in principle, constructive parquet theory builds up its own irreducible by successively inserting an additional link within the self-energy to replace it with avoiding overcounting of terms already generated via the pairwise parquet equations; this generates new primitive irreducibles alongside the preceding input expression. Then is recalculated. Old and new crossing symmetric terms are retained to form a new . Noncrossing symmetric terms, as unavoidable here as in for derivability, are discarded. The parquet equations are solved once again for a new . Details can be found in Bickers [10].
In parquet, by the logic above, iteration leads to and , or more properly , converging to the same crossing symmetric object in both Figure 10(a) and 10(b). Without connection to a variationally consistent, conserving two-body Green function, it is unclear whether its would accord with in Eq. (6).
From an entirely different cause, violation of Eq. (6) is a known characteristic of derivability. In this case, the reason is the loss of correlated-state coherence in Kraichnan averaging [16]. A significant instance of this violation is the random-phase approximation, for which is always (trivially) positive while its real-space static pair correlation function, obtained from integration of , becomes negative in the short-range limit [28].
Time translation invariance of the -derivable ground-state description implies that if were to replace in Figure 10(b) to define an object that we may call , then trivially its inverse Fourier at equal times is
Similarly, if we replace in Figure 10(a) with and call this object , then
Consistency in a -derivable calculation precludes confusing with the proper defined as indicated in Figure 10(a), and confusing with the proper defined in accordance with Figure 10(b).
Summary
In this work we have returned to the basic makeup of diagrammatic expansions for the strongly interacting ground state, offering a different understanding of the puzzling incompatibility between two classic many-body methods: parquet and -derivable approximations. The first cannot sustain conservation for its response functions and the second cannot sustain crossing symmetry for its complete two-body scattering kernel. To cast another light on the established proofs of this incompatibility [1, 2, 3] we have advanced an interpretation via Kraichnan’s stochastic Hamiltonian embedding.
Resolution of the exact Hamiltonian with this machinery lets one draw the following distinction between the approaches in question. On the one hand the conventional analysis of the parquet equations accommodates, in a more intuitive way, a range of inter-channel pair correlations beyond those from the Hamiltonian-based formalism we have described. On the other hand, from the Kraichnan derivation of exact parquet, there emerges a systematic algorithm to isolate, in theory, all the pair-irreducible terms in a rational way faithful to the exact Hamiltonian template.
Tracking the functional interaction between pair and irreducible processes in fine detail seems beyond the current scope of either analysis. It is hard to avoid the suspicion that crossing symmetry in the exact Luttinger-Ward functional is under-written by just such interplay. Separating out pair processes, to study them in minimal configurations, has long accounted for much crucial physics [29], with great success. Beyond this remarkable record, clarifying further effects is likely to bring in more than two-body dynamics. While these remain to be explored beyond more refined mean-field pictures [26, 30], theoretical investigations of three-body parquet already exist [31].
The difference between parquet and Kraichnan philosophies can be viewed in analogy with two railroad switchyards for which, in the former case, decisions as to which units couple where are primarily made locally at each junction while, in the latter, they are primarily made globally and autonomously by a central algorithm. A microscopic model ideally respects the global topology that guarantees unitarity, or conservation, for the associated response behavior [5, 6, 25]. derivability ensures this, but only at the price of losing the crossing symmetry characteristic of fermionic systems.
Developments based on Kraichnan’s approach might go in several directions. In the first place, a broader study of the range of dynamical sum rules for idealized models, such as Hubbard, would put practical numbers on rates and kinds of violation by parquet and -FLEX treatments of the same example.
Kraichnan’s original papers [11] and especially [12] provide straightforward stability proofs for the bound states of the ladder ( channel only) and ring ( channel only) models. While he foreshadows multichannel extensions such as the present picture, a corresponding stability analysis for is not at hand. The fact that it involves a nontrivial superposition of the elementary channels complicates matters, offering a natural topic for further work.
An issue of relevance is the examination of the boundary conditions assumed in parquet over against -derivable theories which, at face value, are very different. Parquet, conceived for strong scattering in open systems, relies on exact particle-antiparticle correspondence. In adapting parquet to condensed matter, one has then to consider: (a) that (quasi)particle antisymmetry holds strictly only at the Fermi surface; (b) that the system is closed to particle entry and escape; thus (c) that ”incoming/outgoing” particles in the condensed state are not asymptotically free of the collective background. Conversely, conserving models address the ground state and relatively low-lying excitations (more generally the free energy). It does not then follow that they are appropriate to open-system scattering from/to the unbound vacuum. Reexamination and development of Kraichnan’s technique could motivate exploring other contexts in which the approach may be relevant, namely those reliant on a Hamiltonian or that can be referred back to one. In particular, for bosonic systems the additional kernel terms represented by no longer clash with particle symmetry. Then the distinction between derivability and parquet really comes down to each one’s relationship to conservation.
Acknowledgments
I thank T. L. Ainsworth, M. P. Das and R. A. Smith for many enlightening discussions, over a long period, on the nature of many-body diagrammatic expansions. Their reflections have stimulated my revisiting the topics discussed in this and earlier papers.
Self-energies with less than full symmetry
In this Appendix we recall general criteria for the status of closed diagrams contributing to the Luttinger-Ward functional at finite order in the interaction. The discussion is simplified by applying the equivalent expression for the LW correlation energy functional, due to Luttinger and Ward [4]:
Here the one-body propagators are everywhere renormalized with their internal interaction at full strength. Unlike the coupling-constant integral in Eq. (7), in the above covers only the multiplicity of the interactions within the skeleton for .
On the basis of the self-energy’s being given uniquely by , variation with of in Eq. (36) again returns the result of Eq. (9), but for tracing combinatorics within the LW skeletons, the diagrammatic expression in the form of is more straightforward than its correlation energy expression . One need only consider each graphical contribution to avoiding the implicit dependence of .
For a given contribution to at order in it is clear that, if it belongs to as in Figure 3(b) and Figure 8(a), or else if absolutely irreducible as, for example, in Figure 8(d) and (Figure 8f), every propagator in such a diagram is strictly equivalent to every other, yielding -fold symmetry. The combinatorial weighting , carried in the evaluation of the term, is duly undone in the variation which is the self-energy.
This is not so for composite irreducible diagrams such as those of Figure 8(b), (Figure 8c) and (Figure 8e), which have less than complete -fold symmetry. Since they carry inequivalent lines, the way that a unique term in is generated is less obvious. We consider the concrete example of the fourth-order diagram Figure 8(b). Call its kernel . After integrating out the
coupling constant, its contribution to is
In Figure 11 we display the outcome of taking the variation with respect to by opening up lines everywhere in . Given that its diagrammatic symmetry is lower, variation with on each side generates a set of dissimilar structures in the self-energy, but each comes weighted by its symmetry factor, defining the total self-energy as their weighted sum.

Figure 11. Leftmost graphs: two variants of the fourth-order non- contribution to the LW functional, after Figure 8(b), whose structural symmetry is lower than the eightfold maximum. (a) Right-hand diagrams: variation for the self-energy for the term with two embedded polarization bubbles, yields one fourfold and two twofold self-energy parts with appropriate weightings. (b) Right-hand diagrams: -type ladder replaces one bubble. Symmetry is lower than for (a) while the total self-energy retains proper overall weighting. In each case the replacement in Eq. (36) exactly recovers ; but so will one of the distinct self-energy components when closed up again by a line and again introduced, with unit weight, as the coupling-constant integrand in Eq. (36). More widely, although the self-energy from any set of closed skeleton diagrams is always derivable consistently and unambiguously, a unique kernel may not be definable.
The test for -derivable consistency is whether performing a coupling-constant integral with the functional replacing the form in the integrand on the right-hand side of Eq. (36) recovers the original fourth-order piece of the LW functional. With the assigned symmetry weightings, it does. So and hence are legitimate.
For a general closed graph of of order in the interaction, the procedure is straightforward. Let there be groups of topologically interchangeable one-body propagators, each with -fold symmetry so . Then should be even if every graph is at least microscopically reversible. We can define a accordingly as the corresponding component in Eq. (36). With representative of the th group,
Now any on the right-hand side of Eq. (37), given full unit weight, is enough on its own to reconstitute the full structure when closed up again with a propagator so . It follows that in constructing a general component to approximate the LW functional, the choice of an effective two-body scattering kernel may not be unique (apart from first order, namely Hartree-Fock). This is clear for closed graphs of less than maximal symmetry; but Figure 2 for the second-order self-energy and inspection of the diagrams making up itself, Figure 8(a), show this to apply as well to for at least a subclass of skeletons with maximal symmetry at all orders. This is demonstrated in Figure 5 for .
Kraichnan average of a skeleton graph
Take any closed skeleton graph in the series for at order in . Recalling that , the Kraichnan average of the factors will be
To show this, consider a typical product in the expansion of the left-hand side of Eq. (38). It has the form
where for each channel denotes that the product has factors with and the stochastically uncorrelated phases for each channel decouple in the overall average.
If an expectation for channel in the expression (B2) does not vanish, it must be identically unity. Then its graph may be closed by detaching it from the other channel products and thus represents a legitimate, autonomous diagram of order . However, if the constraints on index sums means that one, and only one, pair could have connected the parts, similarly to Figure 7; connection by multipairs could not force equality of the indices across every linking pair. Therefore the subgraph must constitute a self-energy insertion and the original diagram would be one-pair reducible, not a skeleton as assumed.
It follows that each , if not zero, must be equal to and the sole combinations of products allowed are those in Eq. (38). In practice the full result applies nontrivially only for ; see Figure 2. For higher order at most one of the expectations can survive.
Equation (B1) has the following consequence for expectations of anticouplings: the Kraichnan average of for any skeleton graph in the complete Luttinger-Ward functional is
If the left-hand side of Eq. (40) vanishes, it can only be when the expectation on the right is unity, since the skeleton must be of form. On the other hand, if the left-hand side does not vanish, it must be unity while, on the right-hand side, the expectation vanishes if and only if
or, in other words, when the skeleton is not in the set.
Extraction of primitively irreducible kernel
Our premise is that there exists a well defined kernel generating the complementary non- kernel , defined in Eq. (27), with which all its pair-reducible components can be obtained. We assume the parquet-like equations
with .
The last expression in Eq. (41) can also be cast as
leading to the formal solutions for the auxiliary kernels
Now from Eqs. (41)–(43) we can also rewrite as
to arrive at a final formulation for purely in terms of and the selective channel anticoupling factors:
rolling back, in each channel, all the pair-reducible iterations of within . The kernel thus obtained is closely related to the complete, primitively irreducible kernel invoked in parquet theory.
Although from the Kraichnan viewpoint there is no real relevance to Eq. (44), rewriting the exact Eqs. (33) and (34) in terms of results in a set of equations more closely resembling classical parquet, namely
for . (C5)
This version of exact parquet departs in two significant ways from the standard case. First, the K coupling compensates for overcounting when is reconstructed from Eq. (C5). This correction can also be applied, if by hand, in the context of normal parquet. Second, to represent systematically all the structural combinations in the exact , the K couplings
and their anticouplings operate independently of one another, despite the fact that they sum identically to unity. In that way the variationally defined kernel stays isomorphic with the structure of the generating Luttinger-Ward functional.
References
- [1]W. Becker and D. Grosser, Nuov. Cim. A 10, 343 (1972).
- [2]A. D. Jackson and R. A. Smith, Phys. Rev. A 36, 2517 (1987).
- [3]R. A. Smith, Phys. Rev. A 46, 4586 (1992).
- [4]J. M. Luttinger and J. C. Ward, Phys. Rev. 118, 1417 (1960).
- [5]G. Baym and L. P. Kadanoff, Phys. Rev. 124, 287 (1961).
- [6]G. Baym, Phys. Rev. 127, 1391 (1962).
- [7]R. W. Haymaker and R. Blankenbecler, Phys. Rev. 171, 1581 (1968).
- [8]A. D. Jackson, A. Lande, and R. A. Smith, Phys. Rep. 86, 55 (1982).
- [9]A. Lande, and R. A. Smith, Phys. Lett. B 131, 253 (1983).
- [10]N. E. Bickers, Int. J. Mod. Phys. B 5, 253 (1991).
- [11]R. H. Kraichnan, J. Math. Phys. 3, 475 (1962).
- [12]R. H. Kraichnan, J. Math. Phys. 3, 496 (1962).
- [13]F. Green, Phys. Rev. A 99, 062118 (2019).
- [14]D. Pines and P. Nozières, The Theory of Quantum Liquids, Volume I: Normal Fermi Liquids (Benjamin, New York, 1966).
- [15]P. Nozières, Theory of Interacting Fermi Systems (Benjamin, New York, 1964), Ch. 5.
- [16]F. Green and T. L. Ainsworth, Phys. Rev. A 106, 052208 (2022).
- [17]N. E. Bickers, D. J. Scalapino, and S. R. White, Phys. Rev. Lett. 62, 961 (1989).
- [18]C. De Dominicis and P. C. Martin, J. Math. Phys. 5, 14 (1964).
- [19]T. Kita, J. Phys. Soc. Jpn. 91, 114002 (2022).
- [20]M. Potthoff, Condens. Matter Physics 9, 557 (2006); https://doi.org/10.48550/arXiv.cond-mat/0406671.
- [21]L. Lin and M. Lindsey, Proc. Natl. Acad. Sci. 115, 2282 (2018); Arch. Rational Mech. Anal. 242, 581 (2021).
- [22]L. Lin and M. Lindsey, Arch. Rational Mech. Anal. 242, 527 (2021).
- [23]Conservation of the collective-index sum comes from invariance of the individual embedded Hamiltonians, analogous to translational invariance of the individual unit cells embedded in a lattice, leading to a conserved crystal momentum.
- [24]For the three auxiliary stu kernels, Λs is antisymmetric while Λt = −Λu on label exchange. Dropping Λ″ in the stu form of the parquet equations results in a total kernel that is crossing symmetric though no longer conserving. It is the analog of minimal standard parquet [10] where V̄ is the only irreducible input kernel.
- [25]G. Baym, in Progress in Nonequilibrium Green’s Functions, M. Bonitz ed. (World Scientific, Singapore, 2000), pp 17-32.
- [26]T. L. Ainsworth and K. S. Bedell, Phys. Rev. B 35, 8425 (1987).
- [27]V. Lislie, Concepts in Quantum Field Theory (Springer, Cham, 2016), Ch. 12.
- [28]G. D. Mahan, Many-Particle Physics (Plenum, New York, 1981), Ch. 5.
- [29]N. W. Ashcroft and N. D. Mermin, Solid state physics (Saunders, New York, 1976).
- [30]A. Georges, G. Kotliar, W. Krauth, and M. Rozenberg, Rev. Mod. Phys. 68, 13 (1996).
- [31]A. Lande and R. A. Smith, Phys. Rev. A 45, 913 (1992).